Linux Audio

Check our new training course

Loading...
v3.1
 
   1/*
   2 * This file implements the perfmon-2 subsystem which is used
   3 * to program the IA-64 Performance Monitoring Unit (PMU).
   4 *
   5 * The initial version of perfmon.c was written by
   6 * Ganesh Venkitachalam, IBM Corp.
   7 *
   8 * Then it was modified for perfmon-1.x by Stephane Eranian and
   9 * David Mosberger, Hewlett Packard Co.
  10 *
  11 * Version Perfmon-2.x is a rewrite of perfmon-1.x
  12 * by Stephane Eranian, Hewlett Packard Co.
  13 *
  14 * Copyright (C) 1999-2005  Hewlett Packard Co
  15 *               Stephane Eranian <eranian@hpl.hp.com>
  16 *               David Mosberger-Tang <davidm@hpl.hp.com>
  17 *
  18 * More information about perfmon available at:
  19 * 	http://www.hpl.hp.com/research/linux/perfmon
  20 */
  21
  22#include <linux/module.h>
  23#include <linux/kernel.h>
  24#include <linux/sched.h>
 
 
  25#include <linux/interrupt.h>
  26#include <linux/proc_fs.h>
  27#include <linux/seq_file.h>
  28#include <linux/init.h>
  29#include <linux/vmalloc.h>
  30#include <linux/mm.h>
  31#include <linux/sysctl.h>
  32#include <linux/list.h>
  33#include <linux/file.h>
  34#include <linux/poll.h>
  35#include <linux/vfs.h>
  36#include <linux/smp.h>
  37#include <linux/pagemap.h>
  38#include <linux/mount.h>
 
  39#include <linux/bitops.h>
  40#include <linux/capability.h>
  41#include <linux/rcupdate.h>
  42#include <linux/completion.h>
  43#include <linux/tracehook.h>
  44#include <linux/slab.h>
 
  45
  46#include <asm/errno.h>
  47#include <asm/intrinsics.h>
  48#include <asm/page.h>
  49#include <asm/perfmon.h>
  50#include <asm/processor.h>
  51#include <asm/signal.h>
  52#include <asm/system.h>
  53#include <asm/uaccess.h>
  54#include <asm/delay.h>
  55
 
 
  56#ifdef CONFIG_PERFMON
  57/*
  58 * perfmon context state
  59 */
  60#define PFM_CTX_UNLOADED	1	/* context is not loaded onto any task */
  61#define PFM_CTX_LOADED		2	/* context is loaded onto a task */
  62#define PFM_CTX_MASKED		3	/* context is loaded but monitoring is masked due to overflow */
  63#define PFM_CTX_ZOMBIE		4	/* owner of the context is closing it */
  64
  65#define PFM_INVALID_ACTIVATION	(~0UL)
  66
  67#define PFM_NUM_PMC_REGS	64	/* PMC save area for ctxsw */
  68#define PFM_NUM_PMD_REGS	64	/* PMD save area for ctxsw */
  69
  70/*
  71 * depth of message queue
  72 */
  73#define PFM_MAX_MSGS		32
  74#define PFM_CTXQ_EMPTY(g)	((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
  75
  76/*
  77 * type of a PMU register (bitmask).
  78 * bitmask structure:
  79 * 	bit0   : register implemented
  80 * 	bit1   : end marker
  81 * 	bit2-3 : reserved
  82 * 	bit4   : pmc has pmc.pm
  83 * 	bit5   : pmc controls a counter (has pmc.oi), pmd is used as counter
  84 * 	bit6-7 : register type
  85 * 	bit8-31: reserved
  86 */
  87#define PFM_REG_NOTIMPL		0x0 /* not implemented at all */
  88#define PFM_REG_IMPL		0x1 /* register implemented */
  89#define PFM_REG_END		0x2 /* end marker */
  90#define PFM_REG_MONITOR		(0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
  91#define PFM_REG_COUNTING	(0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
  92#define PFM_REG_CONTROL		(0x4<<4|PFM_REG_IMPL) /* PMU control register */
  93#define	PFM_REG_CONFIG		(0x8<<4|PFM_REG_IMPL) /* configuration register */
  94#define PFM_REG_BUFFER	 	(0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
  95
  96#define PMC_IS_LAST(i)	(pmu_conf->pmc_desc[i].type & PFM_REG_END)
  97#define PMD_IS_LAST(i)	(pmu_conf->pmd_desc[i].type & PFM_REG_END)
  98
  99#define PMC_OVFL_NOTIFY(ctx, i)	((ctx)->ctx_pmds[i].flags &  PFM_REGFL_OVFL_NOTIFY)
 100
 101/* i assumed unsigned */
 102#define PMC_IS_IMPL(i)	  (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
 103#define PMD_IS_IMPL(i)	  (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
 104
 105/* XXX: these assume that register i is implemented */
 106#define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
 107#define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
 108#define PMC_IS_MONITOR(i)  ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR)  == PFM_REG_MONITOR)
 109#define PMC_IS_CONTROL(i)  ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL)  == PFM_REG_CONTROL)
 110
 111#define PMC_DFL_VAL(i)     pmu_conf->pmc_desc[i].default_value
 112#define PMC_RSVD_MASK(i)   pmu_conf->pmc_desc[i].reserved_mask
 113#define PMD_PMD_DEP(i)	   pmu_conf->pmd_desc[i].dep_pmd[0]
 114#define PMC_PMD_DEP(i)	   pmu_conf->pmc_desc[i].dep_pmd[0]
 115
 116#define PFM_NUM_IBRS	  IA64_NUM_DBG_REGS
 117#define PFM_NUM_DBRS	  IA64_NUM_DBG_REGS
 118
 119#define CTX_OVFL_NOBLOCK(c)	((c)->ctx_fl_block == 0)
 120#define CTX_HAS_SMPL(c)		((c)->ctx_fl_is_sampling)
 121#define PFM_CTX_TASK(h)		(h)->ctx_task
 122
 123#define PMU_PMC_OI		5 /* position of pmc.oi bit */
 124
 125/* XXX: does not support more than 64 PMDs */
 126#define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
 127#define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
 128
 129#define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
 130
 131#define CTX_USED_IBR(ctx,n) 	(ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
 132#define CTX_USED_DBR(ctx,n) 	(ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
 133#define CTX_USES_DBREGS(ctx)	(((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
 134#define PFM_CODE_RR	0	/* requesting code range restriction */
 135#define PFM_DATA_RR	1	/* requestion data range restriction */
 136
 137#define PFM_CPUINFO_CLEAR(v)	pfm_get_cpu_var(pfm_syst_info) &= ~(v)
 138#define PFM_CPUINFO_SET(v)	pfm_get_cpu_var(pfm_syst_info) |= (v)
 139#define PFM_CPUINFO_GET()	pfm_get_cpu_var(pfm_syst_info)
 140
 141#define RDEP(x)	(1UL<<(x))
 142
 143/*
 144 * context protection macros
 145 * in SMP:
 146 * 	- we need to protect against CPU concurrency (spin_lock)
 147 * 	- we need to protect against PMU overflow interrupts (local_irq_disable)
 148 * in UP:
 149 * 	- we need to protect against PMU overflow interrupts (local_irq_disable)
 150 *
 151 * spin_lock_irqsave()/spin_unlock_irqrestore():
 152 * 	in SMP: local_irq_disable + spin_lock
 153 * 	in UP : local_irq_disable
 154 *
 155 * spin_lock()/spin_lock():
 156 * 	in UP : removed automatically
 157 * 	in SMP: protect against context accesses from other CPU. interrupts
 158 * 	        are not masked. This is useful for the PMU interrupt handler
 159 * 	        because we know we will not get PMU concurrency in that code.
 160 */
 161#define PROTECT_CTX(c, f) \
 162	do {  \
 163		DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, task_pid_nr(current))); \
 164		spin_lock_irqsave(&(c)->ctx_lock, f); \
 165		DPRINT(("spinlocked ctx %p  by [%d]\n", c, task_pid_nr(current))); \
 166	} while(0)
 167
 168#define UNPROTECT_CTX(c, f) \
 169	do { \
 170		DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, task_pid_nr(current))); \
 171		spin_unlock_irqrestore(&(c)->ctx_lock, f); \
 172	} while(0)
 173
 174#define PROTECT_CTX_NOPRINT(c, f) \
 175	do {  \
 176		spin_lock_irqsave(&(c)->ctx_lock, f); \
 177	} while(0)
 178
 179
 180#define UNPROTECT_CTX_NOPRINT(c, f) \
 181	do { \
 182		spin_unlock_irqrestore(&(c)->ctx_lock, f); \
 183	} while(0)
 184
 185
 186#define PROTECT_CTX_NOIRQ(c) \
 187	do {  \
 188		spin_lock(&(c)->ctx_lock); \
 189	} while(0)
 190
 191#define UNPROTECT_CTX_NOIRQ(c) \
 192	do { \
 193		spin_unlock(&(c)->ctx_lock); \
 194	} while(0)
 195
 196
 197#ifdef CONFIG_SMP
 198
 199#define GET_ACTIVATION()	pfm_get_cpu_var(pmu_activation_number)
 200#define INC_ACTIVATION()	pfm_get_cpu_var(pmu_activation_number)++
 201#define SET_ACTIVATION(c)	(c)->ctx_last_activation = GET_ACTIVATION()
 202
 203#else /* !CONFIG_SMP */
 204#define SET_ACTIVATION(t) 	do {} while(0)
 205#define GET_ACTIVATION(t) 	do {} while(0)
 206#define INC_ACTIVATION(t) 	do {} while(0)
 207#endif /* CONFIG_SMP */
 208
 209#define SET_PMU_OWNER(t, c)	do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
 210#define GET_PMU_OWNER()		pfm_get_cpu_var(pmu_owner)
 211#define GET_PMU_CTX()		pfm_get_cpu_var(pmu_ctx)
 212
 213#define LOCK_PFS(g)	    	spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
 214#define UNLOCK_PFS(g)	    	spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
 215
 216#define PFM_REG_RETFLAG_SET(flags, val)	do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
 217
 218/*
 219 * cmp0 must be the value of pmc0
 220 */
 221#define PMC0_HAS_OVFL(cmp0)  (cmp0 & ~0x1UL)
 222
 223#define PFMFS_MAGIC 0xa0b4d889
 224
 225/*
 226 * debugging
 227 */
 228#define PFM_DEBUGGING 1
 229#ifdef PFM_DEBUGGING
 230#define DPRINT(a) \
 231	do { \
 232		if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
 233	} while (0)
 234
 235#define DPRINT_ovfl(a) \
 236	do { \
 237		if (unlikely(pfm_sysctl.debug > 0 && pfm_sysctl.debug_ovfl >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
 238	} while (0)
 239#endif
 240
 241/*
 242 * 64-bit software counter structure
 243 *
 244 * the next_reset_type is applied to the next call to pfm_reset_regs()
 245 */
 246typedef struct {
 247	unsigned long	val;		/* virtual 64bit counter value */
 248	unsigned long	lval;		/* last reset value */
 249	unsigned long	long_reset;	/* reset value on sampling overflow */
 250	unsigned long	short_reset;    /* reset value on overflow */
 251	unsigned long	reset_pmds[4];  /* which other pmds to reset when this counter overflows */
 252	unsigned long	smpl_pmds[4];   /* which pmds are accessed when counter overflow */
 253	unsigned long	seed;		/* seed for random-number generator */
 254	unsigned long	mask;		/* mask for random-number generator */
 255	unsigned int 	flags;		/* notify/do not notify */
 256	unsigned long	eventid;	/* overflow event identifier */
 257} pfm_counter_t;
 258
 259/*
 260 * context flags
 261 */
 262typedef struct {
 263	unsigned int block:1;		/* when 1, task will blocked on user notifications */
 264	unsigned int system:1;		/* do system wide monitoring */
 265	unsigned int using_dbreg:1;	/* using range restrictions (debug registers) */
 266	unsigned int is_sampling:1;	/* true if using a custom format */
 267	unsigned int excl_idle:1;	/* exclude idle task in system wide session */
 268	unsigned int going_zombie:1;	/* context is zombie (MASKED+blocking) */
 269	unsigned int trap_reason:2;	/* reason for going into pfm_handle_work() */
 270	unsigned int no_msg:1;		/* no message sent on overflow */
 271	unsigned int can_restart:1;	/* allowed to issue a PFM_RESTART */
 272	unsigned int reserved:22;
 273} pfm_context_flags_t;
 274
 275#define PFM_TRAP_REASON_NONE		0x0	/* default value */
 276#define PFM_TRAP_REASON_BLOCK		0x1	/* we need to block on overflow */
 277#define PFM_TRAP_REASON_RESET		0x2	/* we need to reset PMDs */
 278
 279
 280/*
 281 * perfmon context: encapsulates all the state of a monitoring session
 282 */
 283
 284typedef struct pfm_context {
 285	spinlock_t		ctx_lock;		/* context protection */
 286
 287	pfm_context_flags_t	ctx_flags;		/* bitmask of flags  (block reason incl.) */
 288	unsigned int		ctx_state;		/* state: active/inactive (no bitfield) */
 289
 290	struct task_struct 	*ctx_task;		/* task to which context is attached */
 291
 292	unsigned long		ctx_ovfl_regs[4];	/* which registers overflowed (notification) */
 293
 294	struct completion	ctx_restart_done;  	/* use for blocking notification mode */
 295
 296	unsigned long		ctx_used_pmds[4];	/* bitmask of PMD used            */
 297	unsigned long		ctx_all_pmds[4];	/* bitmask of all accessible PMDs */
 298	unsigned long		ctx_reload_pmds[4];	/* bitmask of force reload PMD on ctxsw in */
 299
 300	unsigned long		ctx_all_pmcs[4];	/* bitmask of all accessible PMCs */
 301	unsigned long		ctx_reload_pmcs[4];	/* bitmask of force reload PMC on ctxsw in */
 302	unsigned long		ctx_used_monitors[4];	/* bitmask of monitor PMC being used */
 303
 304	unsigned long		ctx_pmcs[PFM_NUM_PMC_REGS];	/*  saved copies of PMC values */
 305
 306	unsigned int		ctx_used_ibrs[1];		/* bitmask of used IBR (speedup ctxsw in) */
 307	unsigned int		ctx_used_dbrs[1];		/* bitmask of used DBR (speedup ctxsw in) */
 308	unsigned long		ctx_dbrs[IA64_NUM_DBG_REGS];	/* DBR values (cache) when not loaded */
 309	unsigned long		ctx_ibrs[IA64_NUM_DBG_REGS];	/* IBR values (cache) when not loaded */
 310
 311	pfm_counter_t		ctx_pmds[PFM_NUM_PMD_REGS]; /* software state for PMDS */
 312
 313	unsigned long		th_pmcs[PFM_NUM_PMC_REGS];	/* PMC thread save state */
 314	unsigned long		th_pmds[PFM_NUM_PMD_REGS];	/* PMD thread save state */
 315
 316	unsigned long		ctx_saved_psr_up;	/* only contains psr.up value */
 317
 318	unsigned long		ctx_last_activation;	/* context last activation number for last_cpu */
 319	unsigned int		ctx_last_cpu;		/* CPU id of current or last CPU used (SMP only) */
 320	unsigned int		ctx_cpu;		/* cpu to which perfmon is applied (system wide) */
 321
 322	int			ctx_fd;			/* file descriptor used my this context */
 323	pfm_ovfl_arg_t		ctx_ovfl_arg;		/* argument to custom buffer format handler */
 324
 325	pfm_buffer_fmt_t	*ctx_buf_fmt;		/* buffer format callbacks */
 326	void			*ctx_smpl_hdr;		/* points to sampling buffer header kernel vaddr */
 327	unsigned long		ctx_smpl_size;		/* size of sampling buffer */
 328	void			*ctx_smpl_vaddr;	/* user level virtual address of smpl buffer */
 329
 330	wait_queue_head_t 	ctx_msgq_wait;
 331	pfm_msg_t		ctx_msgq[PFM_MAX_MSGS];
 332	int			ctx_msgq_head;
 333	int			ctx_msgq_tail;
 334	struct fasync_struct	*ctx_async_queue;
 335
 336	wait_queue_head_t 	ctx_zombieq;		/* termination cleanup wait queue */
 337} pfm_context_t;
 338
 339/*
 340 * magic number used to verify that structure is really
 341 * a perfmon context
 342 */
 343#define PFM_IS_FILE(f)		((f)->f_op == &pfm_file_ops)
 344
 345#define PFM_GET_CTX(t)	 	((pfm_context_t *)(t)->thread.pfm_context)
 346
 347#ifdef CONFIG_SMP
 348#define SET_LAST_CPU(ctx, v)	(ctx)->ctx_last_cpu = (v)
 349#define GET_LAST_CPU(ctx)	(ctx)->ctx_last_cpu
 350#else
 351#define SET_LAST_CPU(ctx, v)	do {} while(0)
 352#define GET_LAST_CPU(ctx)	do {} while(0)
 353#endif
 354
 355
 356#define ctx_fl_block		ctx_flags.block
 357#define ctx_fl_system		ctx_flags.system
 358#define ctx_fl_using_dbreg	ctx_flags.using_dbreg
 359#define ctx_fl_is_sampling	ctx_flags.is_sampling
 360#define ctx_fl_excl_idle	ctx_flags.excl_idle
 361#define ctx_fl_going_zombie	ctx_flags.going_zombie
 362#define ctx_fl_trap_reason	ctx_flags.trap_reason
 363#define ctx_fl_no_msg		ctx_flags.no_msg
 364#define ctx_fl_can_restart	ctx_flags.can_restart
 365
 366#define PFM_SET_WORK_PENDING(t, v)	do { (t)->thread.pfm_needs_checking = v; } while(0);
 367#define PFM_GET_WORK_PENDING(t)		(t)->thread.pfm_needs_checking
 368
 369/*
 370 * global information about all sessions
 371 * mostly used to synchronize between system wide and per-process
 372 */
 373typedef struct {
 374	spinlock_t		pfs_lock;		   /* lock the structure */
 375
 376	unsigned int		pfs_task_sessions;	   /* number of per task sessions */
 377	unsigned int		pfs_sys_sessions;	   /* number of per system wide sessions */
 378	unsigned int		pfs_sys_use_dbregs;	   /* incremented when a system wide session uses debug regs */
 379	unsigned int		pfs_ptrace_use_dbregs;	   /* incremented when a process uses debug regs */
 380	struct task_struct	*pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
 381} pfm_session_t;
 382
 383/*
 384 * information about a PMC or PMD.
 385 * dep_pmd[]: a bitmask of dependent PMD registers
 386 * dep_pmc[]: a bitmask of dependent PMC registers
 387 */
 388typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
 389typedef struct {
 390	unsigned int		type;
 391	int			pm_pos;
 392	unsigned long		default_value;	/* power-on default value */
 393	unsigned long		reserved_mask;	/* bitmask of reserved bits */
 394	pfm_reg_check_t		read_check;
 395	pfm_reg_check_t		write_check;
 396	unsigned long		dep_pmd[4];
 397	unsigned long		dep_pmc[4];
 398} pfm_reg_desc_t;
 399
 400/* assume cnum is a valid monitor */
 401#define PMC_PM(cnum, val)	(((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
 402
 403/*
 404 * This structure is initialized at boot time and contains
 405 * a description of the PMU main characteristics.
 406 *
 407 * If the probe function is defined, detection is based
 408 * on its return value: 
 409 * 	- 0 means recognized PMU
 410 * 	- anything else means not supported
 411 * When the probe function is not defined, then the pmu_family field
 412 * is used and it must match the host CPU family such that:
 413 * 	- cpu->family & config->pmu_family != 0
 414 */
 415typedef struct {
 416	unsigned long  ovfl_val;	/* overflow value for counters */
 417
 418	pfm_reg_desc_t *pmc_desc;	/* detailed PMC register dependencies descriptions */
 419	pfm_reg_desc_t *pmd_desc;	/* detailed PMD register dependencies descriptions */
 420
 421	unsigned int   num_pmcs;	/* number of PMCS: computed at init time */
 422	unsigned int   num_pmds;	/* number of PMDS: computed at init time */
 423	unsigned long  impl_pmcs[4];	/* bitmask of implemented PMCS */
 424	unsigned long  impl_pmds[4];	/* bitmask of implemented PMDS */
 425
 426	char	      *pmu_name;	/* PMU family name */
 427	unsigned int  pmu_family;	/* cpuid family pattern used to identify pmu */
 428	unsigned int  flags;		/* pmu specific flags */
 429	unsigned int  num_ibrs;		/* number of IBRS: computed at init time */
 430	unsigned int  num_dbrs;		/* number of DBRS: computed at init time */
 431	unsigned int  num_counters;	/* PMC/PMD counting pairs : computed at init time */
 432	int           (*probe)(void);   /* customized probe routine */
 433	unsigned int  use_rr_dbregs:1;	/* set if debug registers used for range restriction */
 434} pmu_config_t;
 435/*
 436 * PMU specific flags
 437 */
 438#define PFM_PMU_IRQ_RESEND	1	/* PMU needs explicit IRQ resend */
 439
 440/*
 441 * debug register related type definitions
 442 */
 443typedef struct {
 444	unsigned long ibr_mask:56;
 445	unsigned long ibr_plm:4;
 446	unsigned long ibr_ig:3;
 447	unsigned long ibr_x:1;
 448} ibr_mask_reg_t;
 449
 450typedef struct {
 451	unsigned long dbr_mask:56;
 452	unsigned long dbr_plm:4;
 453	unsigned long dbr_ig:2;
 454	unsigned long dbr_w:1;
 455	unsigned long dbr_r:1;
 456} dbr_mask_reg_t;
 457
 458typedef union {
 459	unsigned long  val;
 460	ibr_mask_reg_t ibr;
 461	dbr_mask_reg_t dbr;
 462} dbreg_t;
 463
 464
 465/*
 466 * perfmon command descriptions
 467 */
 468typedef struct {
 469	int		(*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
 470	char		*cmd_name;
 471	int		cmd_flags;
 472	unsigned int	cmd_narg;
 473	size_t		cmd_argsize;
 474	int		(*cmd_getsize)(void *arg, size_t *sz);
 475} pfm_cmd_desc_t;
 476
 477#define PFM_CMD_FD		0x01	/* command requires a file descriptor */
 478#define PFM_CMD_ARG_READ	0x02	/* command must read argument(s) */
 479#define PFM_CMD_ARG_RW		0x04	/* command must read/write argument(s) */
 480#define PFM_CMD_STOP		0x08	/* command does not work on zombie context */
 481
 482
 483#define PFM_CMD_NAME(cmd)	pfm_cmd_tab[(cmd)].cmd_name
 484#define PFM_CMD_READ_ARG(cmd)	(pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
 485#define PFM_CMD_RW_ARG(cmd)	(pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
 486#define PFM_CMD_USE_FD(cmd)	(pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
 487#define PFM_CMD_STOPPED(cmd)	(pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
 488
 489#define PFM_CMD_ARG_MANY	-1 /* cannot be zero */
 490
 491typedef struct {
 492	unsigned long pfm_spurious_ovfl_intr_count;	/* keep track of spurious ovfl interrupts */
 493	unsigned long pfm_replay_ovfl_intr_count;	/* keep track of replayed ovfl interrupts */
 494	unsigned long pfm_ovfl_intr_count; 		/* keep track of ovfl interrupts */
 495	unsigned long pfm_ovfl_intr_cycles;		/* cycles spent processing ovfl interrupts */
 496	unsigned long pfm_ovfl_intr_cycles_min;		/* min cycles spent processing ovfl interrupts */
 497	unsigned long pfm_ovfl_intr_cycles_max;		/* max cycles spent processing ovfl interrupts */
 498	unsigned long pfm_smpl_handler_calls;
 499	unsigned long pfm_smpl_handler_cycles;
 500	char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
 501} pfm_stats_t;
 502
 503/*
 504 * perfmon internal variables
 505 */
 506static pfm_stats_t		pfm_stats[NR_CPUS];
 507static pfm_session_t		pfm_sessions;	/* global sessions information */
 508
 509static DEFINE_SPINLOCK(pfm_alt_install_check);
 510static pfm_intr_handler_desc_t  *pfm_alt_intr_handler;
 511
 512static struct proc_dir_entry 	*perfmon_dir;
 513static pfm_uuid_t		pfm_null_uuid = {0,};
 514
 515static spinlock_t		pfm_buffer_fmt_lock;
 516static LIST_HEAD(pfm_buffer_fmt_list);
 517
 518static pmu_config_t		*pmu_conf;
 519
 520/* sysctl() controls */
 521pfm_sysctl_t pfm_sysctl;
 522EXPORT_SYMBOL(pfm_sysctl);
 523
 524static ctl_table pfm_ctl_table[]={
 525	{
 526		.procname	= "debug",
 527		.data		= &pfm_sysctl.debug,
 528		.maxlen		= sizeof(int),
 529		.mode		= 0666,
 530		.proc_handler	= proc_dointvec,
 531	},
 532	{
 533		.procname	= "debug_ovfl",
 534		.data		= &pfm_sysctl.debug_ovfl,
 535		.maxlen		= sizeof(int),
 536		.mode		= 0666,
 537		.proc_handler	= proc_dointvec,
 538	},
 539	{
 540		.procname	= "fastctxsw",
 541		.data		= &pfm_sysctl.fastctxsw,
 542		.maxlen		= sizeof(int),
 543		.mode		= 0600,
 544		.proc_handler	= proc_dointvec,
 545	},
 546	{
 547		.procname	= "expert_mode",
 548		.data		= &pfm_sysctl.expert_mode,
 549		.maxlen		= sizeof(int),
 550		.mode		= 0600,
 551		.proc_handler	= proc_dointvec,
 552	},
 553	{}
 554};
 555static ctl_table pfm_sysctl_dir[] = {
 556	{
 557		.procname	= "perfmon",
 558		.mode		= 0555,
 559		.child		= pfm_ctl_table,
 560	},
 561 	{}
 562};
 563static ctl_table pfm_sysctl_root[] = {
 564	{
 565		.procname	= "kernel",
 566		.mode		= 0555,
 567		.child		= pfm_sysctl_dir,
 568	},
 569 	{}
 570};
 571static struct ctl_table_header *pfm_sysctl_header;
 572
 573static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
 574
 575#define pfm_get_cpu_var(v)		__ia64_per_cpu_var(v)
 576#define pfm_get_cpu_data(a,b)		per_cpu(a, b)
 577
 578static inline void
 579pfm_put_task(struct task_struct *task)
 580{
 581	if (task != current) put_task_struct(task);
 582}
 583
 584static inline void
 585pfm_reserve_page(unsigned long a)
 586{
 587	SetPageReserved(vmalloc_to_page((void *)a));
 588}
 589static inline void
 590pfm_unreserve_page(unsigned long a)
 591{
 592	ClearPageReserved(vmalloc_to_page((void*)a));
 593}
 594
 595static inline unsigned long
 596pfm_protect_ctx_ctxsw(pfm_context_t *x)
 597{
 598	spin_lock(&(x)->ctx_lock);
 599	return 0UL;
 600}
 601
 602static inline void
 603pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
 604{
 605	spin_unlock(&(x)->ctx_lock);
 606}
 607
 608static inline unsigned int
 609pfm_do_munmap(struct mm_struct *mm, unsigned long addr, size_t len, int acct)
 610{
 611	return do_munmap(mm, addr, len);
 612}
 613
 614static inline unsigned long 
 615pfm_get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, unsigned long pgoff, unsigned long flags, unsigned long exec)
 616{
 617	return get_unmapped_area(file, addr, len, pgoff, flags);
 618}
 619
 620/* forward declaration */
 621static const struct dentry_operations pfmfs_dentry_operations;
 622
 623static struct dentry *
 624pfmfs_mount(struct file_system_type *fs_type, int flags, const char *dev_name, void *data)
 625{
 626	return mount_pseudo(fs_type, "pfm:", NULL, &pfmfs_dentry_operations,
 627			PFMFS_MAGIC);
 
 
 
 628}
 629
 630static struct file_system_type pfm_fs_type = {
 631	.name     = "pfmfs",
 632	.mount    = pfmfs_mount,
 633	.kill_sb  = kill_anon_super,
 634};
 
 635
 636DEFINE_PER_CPU(unsigned long, pfm_syst_info);
 637DEFINE_PER_CPU(struct task_struct *, pmu_owner);
 638DEFINE_PER_CPU(pfm_context_t  *, pmu_ctx);
 639DEFINE_PER_CPU(unsigned long, pmu_activation_number);
 640EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info);
 641
 642
 643/* forward declaration */
 644static const struct file_operations pfm_file_ops;
 645
 646/*
 647 * forward declarations
 648 */
 649#ifndef CONFIG_SMP
 650static void pfm_lazy_save_regs (struct task_struct *ta);
 651#endif
 652
 653void dump_pmu_state(const char *);
 654static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
 655
 656#include "perfmon_itanium.h"
 657#include "perfmon_mckinley.h"
 658#include "perfmon_montecito.h"
 659#include "perfmon_generic.h"
 660
 661static pmu_config_t *pmu_confs[]={
 662	&pmu_conf_mont,
 663	&pmu_conf_mck,
 664	&pmu_conf_ita,
 665	&pmu_conf_gen, /* must be last */
 666	NULL
 667};
 668
 669
 670static int pfm_end_notify_user(pfm_context_t *ctx);
 671
 672static inline void
 673pfm_clear_psr_pp(void)
 674{
 675	ia64_rsm(IA64_PSR_PP);
 676	ia64_srlz_i();
 677}
 678
 679static inline void
 680pfm_set_psr_pp(void)
 681{
 682	ia64_ssm(IA64_PSR_PP);
 683	ia64_srlz_i();
 684}
 685
 686static inline void
 687pfm_clear_psr_up(void)
 688{
 689	ia64_rsm(IA64_PSR_UP);
 690	ia64_srlz_i();
 691}
 692
 693static inline void
 694pfm_set_psr_up(void)
 695{
 696	ia64_ssm(IA64_PSR_UP);
 697	ia64_srlz_i();
 698}
 699
 700static inline unsigned long
 701pfm_get_psr(void)
 702{
 703	unsigned long tmp;
 704	tmp = ia64_getreg(_IA64_REG_PSR);
 705	ia64_srlz_i();
 706	return tmp;
 707}
 708
 709static inline void
 710pfm_set_psr_l(unsigned long val)
 711{
 712	ia64_setreg(_IA64_REG_PSR_L, val);
 713	ia64_srlz_i();
 714}
 715
 716static inline void
 717pfm_freeze_pmu(void)
 718{
 719	ia64_set_pmc(0,1UL);
 720	ia64_srlz_d();
 721}
 722
 723static inline void
 724pfm_unfreeze_pmu(void)
 725{
 726	ia64_set_pmc(0,0UL);
 727	ia64_srlz_d();
 728}
 729
 730static inline void
 731pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
 732{
 733	int i;
 734
 735	for (i=0; i < nibrs; i++) {
 736		ia64_set_ibr(i, ibrs[i]);
 737		ia64_dv_serialize_instruction();
 738	}
 739	ia64_srlz_i();
 740}
 741
 742static inline void
 743pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
 744{
 745	int i;
 746
 747	for (i=0; i < ndbrs; i++) {
 748		ia64_set_dbr(i, dbrs[i]);
 749		ia64_dv_serialize_data();
 750	}
 751	ia64_srlz_d();
 752}
 753
 754/*
 755 * PMD[i] must be a counter. no check is made
 756 */
 757static inline unsigned long
 758pfm_read_soft_counter(pfm_context_t *ctx, int i)
 759{
 760	return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
 761}
 762
 763/*
 764 * PMD[i] must be a counter. no check is made
 765 */
 766static inline void
 767pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
 768{
 769	unsigned long ovfl_val = pmu_conf->ovfl_val;
 770
 771	ctx->ctx_pmds[i].val = val  & ~ovfl_val;
 772	/*
 773	 * writing to unimplemented part is ignore, so we do not need to
 774	 * mask off top part
 775	 */
 776	ia64_set_pmd(i, val & ovfl_val);
 777}
 778
 779static pfm_msg_t *
 780pfm_get_new_msg(pfm_context_t *ctx)
 781{
 782	int idx, next;
 783
 784	next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
 785
 786	DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
 787	if (next == ctx->ctx_msgq_head) return NULL;
 788
 789 	idx = 	ctx->ctx_msgq_tail;
 790	ctx->ctx_msgq_tail = next;
 791
 792	DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
 793
 794	return ctx->ctx_msgq+idx;
 795}
 796
 797static pfm_msg_t *
 798pfm_get_next_msg(pfm_context_t *ctx)
 799{
 800	pfm_msg_t *msg;
 801
 802	DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
 803
 804	if (PFM_CTXQ_EMPTY(ctx)) return NULL;
 805
 806	/*
 807	 * get oldest message
 808	 */
 809	msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
 810
 811	/*
 812	 * and move forward
 813	 */
 814	ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
 815
 816	DPRINT(("ctx=%p head=%d tail=%d type=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, msg->pfm_gen_msg.msg_type));
 817
 818	return msg;
 819}
 820
 821static void
 822pfm_reset_msgq(pfm_context_t *ctx)
 823{
 824	ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
 825	DPRINT(("ctx=%p msgq reset\n", ctx));
 826}
 827
 828static void *
 829pfm_rvmalloc(unsigned long size)
 830{
 831	void *mem;
 832	unsigned long addr;
 833
 834	size = PAGE_ALIGN(size);
 835	mem  = vzalloc(size);
 836	if (mem) {
 837		//printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
 838		addr = (unsigned long)mem;
 839		while (size > 0) {
 840			pfm_reserve_page(addr);
 841			addr+=PAGE_SIZE;
 842			size-=PAGE_SIZE;
 843		}
 844	}
 845	return mem;
 846}
 847
 848static void
 849pfm_rvfree(void *mem, unsigned long size)
 850{
 851	unsigned long addr;
 852
 853	if (mem) {
 854		DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size));
 855		addr = (unsigned long) mem;
 856		while ((long) size > 0) {
 857			pfm_unreserve_page(addr);
 858			addr+=PAGE_SIZE;
 859			size-=PAGE_SIZE;
 860		}
 861		vfree(mem);
 862	}
 863	return;
 864}
 865
 866static pfm_context_t *
 867pfm_context_alloc(int ctx_flags)
 868{
 869	pfm_context_t *ctx;
 870
 871	/* 
 872	 * allocate context descriptor 
 873	 * must be able to free with interrupts disabled
 874	 */
 875	ctx = kzalloc(sizeof(pfm_context_t), GFP_KERNEL);
 876	if (ctx) {
 877		DPRINT(("alloc ctx @%p\n", ctx));
 878
 879		/*
 880		 * init context protection lock
 881		 */
 882		spin_lock_init(&ctx->ctx_lock);
 883
 884		/*
 885		 * context is unloaded
 886		 */
 887		ctx->ctx_state = PFM_CTX_UNLOADED;
 888
 889		/*
 890		 * initialization of context's flags
 891		 */
 892		ctx->ctx_fl_block       = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
 893		ctx->ctx_fl_system      = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
 894		ctx->ctx_fl_no_msg      = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
 895		/*
 896		 * will move to set properties
 897		 * ctx->ctx_fl_excl_idle   = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
 898		 */
 899
 900		/*
 901		 * init restart semaphore to locked
 902		 */
 903		init_completion(&ctx->ctx_restart_done);
 904
 905		/*
 906		 * activation is used in SMP only
 907		 */
 908		ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
 909		SET_LAST_CPU(ctx, -1);
 910
 911		/*
 912		 * initialize notification message queue
 913		 */
 914		ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
 915		init_waitqueue_head(&ctx->ctx_msgq_wait);
 916		init_waitqueue_head(&ctx->ctx_zombieq);
 917
 918	}
 919	return ctx;
 920}
 921
 922static void
 923pfm_context_free(pfm_context_t *ctx)
 924{
 925	if (ctx) {
 926		DPRINT(("free ctx @%p\n", ctx));
 927		kfree(ctx);
 928	}
 929}
 930
 931static void
 932pfm_mask_monitoring(struct task_struct *task)
 933{
 934	pfm_context_t *ctx = PFM_GET_CTX(task);
 935	unsigned long mask, val, ovfl_mask;
 936	int i;
 937
 938	DPRINT_ovfl(("masking monitoring for [%d]\n", task_pid_nr(task)));
 939
 940	ovfl_mask = pmu_conf->ovfl_val;
 941	/*
 942	 * monitoring can only be masked as a result of a valid
 943	 * counter overflow. In UP, it means that the PMU still
 944	 * has an owner. Note that the owner can be different
 945	 * from the current task. However the PMU state belongs
 946	 * to the owner.
 947	 * In SMP, a valid overflow only happens when task is
 948	 * current. Therefore if we come here, we know that
 949	 * the PMU state belongs to the current task, therefore
 950	 * we can access the live registers.
 951	 *
 952	 * So in both cases, the live register contains the owner's
 953	 * state. We can ONLY touch the PMU registers and NOT the PSR.
 954	 *
 955	 * As a consequence to this call, the ctx->th_pmds[] array
 956	 * contains stale information which must be ignored
 957	 * when context is reloaded AND monitoring is active (see
 958	 * pfm_restart).
 959	 */
 960	mask = ctx->ctx_used_pmds[0];
 961	for (i = 0; mask; i++, mask>>=1) {
 962		/* skip non used pmds */
 963		if ((mask & 0x1) == 0) continue;
 964		val = ia64_get_pmd(i);
 965
 966		if (PMD_IS_COUNTING(i)) {
 967			/*
 968		 	 * we rebuild the full 64 bit value of the counter
 969		 	 */
 970			ctx->ctx_pmds[i].val += (val & ovfl_mask);
 971		} else {
 972			ctx->ctx_pmds[i].val = val;
 973		}
 974		DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
 975			i,
 976			ctx->ctx_pmds[i].val,
 977			val & ovfl_mask));
 978	}
 979	/*
 980	 * mask monitoring by setting the privilege level to 0
 981	 * we cannot use psr.pp/psr.up for this, it is controlled by
 982	 * the user
 983	 *
 984	 * if task is current, modify actual registers, otherwise modify
 985	 * thread save state, i.e., what will be restored in pfm_load_regs()
 986	 */
 987	mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
 988	for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
 989		if ((mask & 0x1) == 0UL) continue;
 990		ia64_set_pmc(i, ctx->th_pmcs[i] & ~0xfUL);
 991		ctx->th_pmcs[i] &= ~0xfUL;
 992		DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
 993	}
 994	/*
 995	 * make all of this visible
 996	 */
 997	ia64_srlz_d();
 998}
 999
1000/*
1001 * must always be done with task == current
1002 *
1003 * context must be in MASKED state when calling
1004 */
1005static void
1006pfm_restore_monitoring(struct task_struct *task)
1007{
1008	pfm_context_t *ctx = PFM_GET_CTX(task);
1009	unsigned long mask, ovfl_mask;
1010	unsigned long psr, val;
1011	int i, is_system;
1012
1013	is_system = ctx->ctx_fl_system;
1014	ovfl_mask = pmu_conf->ovfl_val;
1015
1016	if (task != current) {
1017		printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task_pid_nr(task), task_pid_nr(current));
1018		return;
1019	}
1020	if (ctx->ctx_state != PFM_CTX_MASKED) {
1021		printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
1022			task_pid_nr(task), task_pid_nr(current), ctx->ctx_state);
1023		return;
1024	}
1025	psr = pfm_get_psr();
1026	/*
1027	 * monitoring is masked via the PMC.
1028	 * As we restore their value, we do not want each counter to
1029	 * restart right away. We stop monitoring using the PSR,
1030	 * restore the PMC (and PMD) and then re-establish the psr
1031	 * as it was. Note that there can be no pending overflow at
1032	 * this point, because monitoring was MASKED.
1033	 *
1034	 * system-wide session are pinned and self-monitoring
1035	 */
1036	if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1037		/* disable dcr pp */
1038		ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
1039		pfm_clear_psr_pp();
1040	} else {
1041		pfm_clear_psr_up();
1042	}
1043	/*
1044	 * first, we restore the PMD
1045	 */
1046	mask = ctx->ctx_used_pmds[0];
1047	for (i = 0; mask; i++, mask>>=1) {
1048		/* skip non used pmds */
1049		if ((mask & 0x1) == 0) continue;
1050
1051		if (PMD_IS_COUNTING(i)) {
1052			/*
1053			 * we split the 64bit value according to
1054			 * counter width
1055			 */
1056			val = ctx->ctx_pmds[i].val & ovfl_mask;
1057			ctx->ctx_pmds[i].val &= ~ovfl_mask;
1058		} else {
1059			val = ctx->ctx_pmds[i].val;
1060		}
1061		ia64_set_pmd(i, val);
1062
1063		DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1064			i,
1065			ctx->ctx_pmds[i].val,
1066			val));
1067	}
1068	/*
1069	 * restore the PMCs
1070	 */
1071	mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1072	for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1073		if ((mask & 0x1) == 0UL) continue;
1074		ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1075		ia64_set_pmc(i, ctx->th_pmcs[i]);
1076		DPRINT(("[%d] pmc[%d]=0x%lx\n",
1077					task_pid_nr(task), i, ctx->th_pmcs[i]));
1078	}
1079	ia64_srlz_d();
1080
1081	/*
1082	 * must restore DBR/IBR because could be modified while masked
1083	 * XXX: need to optimize 
1084	 */
1085	if (ctx->ctx_fl_using_dbreg) {
1086		pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1087		pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1088	}
1089
1090	/*
1091	 * now restore PSR
1092	 */
1093	if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1094		/* enable dcr pp */
1095		ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1096		ia64_srlz_i();
1097	}
1098	pfm_set_psr_l(psr);
1099}
1100
1101static inline void
1102pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1103{
1104	int i;
1105
1106	ia64_srlz_d();
1107
1108	for (i=0; mask; i++, mask>>=1) {
1109		if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1110	}
1111}
1112
1113/*
1114 * reload from thread state (used for ctxw only)
1115 */
1116static inline void
1117pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1118{
1119	int i;
1120	unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1121
1122	for (i=0; mask; i++, mask>>=1) {
1123		if ((mask & 0x1) == 0) continue;
1124		val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1125		ia64_set_pmd(i, val);
1126	}
1127	ia64_srlz_d();
1128}
1129
1130/*
1131 * propagate PMD from context to thread-state
1132 */
1133static inline void
1134pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1135{
1136	unsigned long ovfl_val = pmu_conf->ovfl_val;
1137	unsigned long mask = ctx->ctx_all_pmds[0];
1138	unsigned long val;
1139	int i;
1140
1141	DPRINT(("mask=0x%lx\n", mask));
1142
1143	for (i=0; mask; i++, mask>>=1) {
1144
1145		val = ctx->ctx_pmds[i].val;
1146
1147		/*
1148		 * We break up the 64 bit value into 2 pieces
1149		 * the lower bits go to the machine state in the
1150		 * thread (will be reloaded on ctxsw in).
1151		 * The upper part stays in the soft-counter.
1152		 */
1153		if (PMD_IS_COUNTING(i)) {
1154			ctx->ctx_pmds[i].val = val & ~ovfl_val;
1155			 val &= ovfl_val;
1156		}
1157		ctx->th_pmds[i] = val;
1158
1159		DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1160			i,
1161			ctx->th_pmds[i],
1162			ctx->ctx_pmds[i].val));
1163	}
1164}
1165
1166/*
1167 * propagate PMC from context to thread-state
1168 */
1169static inline void
1170pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1171{
1172	unsigned long mask = ctx->ctx_all_pmcs[0];
1173	int i;
1174
1175	DPRINT(("mask=0x%lx\n", mask));
1176
1177	for (i=0; mask; i++, mask>>=1) {
1178		/* masking 0 with ovfl_val yields 0 */
1179		ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1180		DPRINT(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
1181	}
1182}
1183
1184
1185
1186static inline void
1187pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1188{
1189	int i;
1190
1191	for (i=0; mask; i++, mask>>=1) {
1192		if ((mask & 0x1) == 0) continue;
1193		ia64_set_pmc(i, pmcs[i]);
1194	}
1195	ia64_srlz_d();
1196}
1197
1198static inline int
1199pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1200{
1201	return memcmp(a, b, sizeof(pfm_uuid_t));
1202}
1203
1204static inline int
1205pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1206{
1207	int ret = 0;
1208	if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1209	return ret;
1210}
1211
1212static inline int
1213pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1214{
1215	int ret = 0;
1216	if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1217	return ret;
1218}
1219
1220
1221static inline int
1222pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1223		     int cpu, void *arg)
1224{
1225	int ret = 0;
1226	if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1227	return ret;
1228}
1229
1230static inline int
1231pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1232		     int cpu, void *arg)
1233{
1234	int ret = 0;
1235	if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1236	return ret;
1237}
1238
1239static inline int
1240pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1241{
1242	int ret = 0;
1243	if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1244	return ret;
1245}
1246
1247static inline int
1248pfm_buf_fmt_restart_active(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1249{
1250	int ret = 0;
1251	if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1252	return ret;
1253}
1254
1255static pfm_buffer_fmt_t *
1256__pfm_find_buffer_fmt(pfm_uuid_t uuid)
1257{
1258	struct list_head * pos;
1259	pfm_buffer_fmt_t * entry;
1260
1261	list_for_each(pos, &pfm_buffer_fmt_list) {
1262		entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1263		if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1264			return entry;
1265	}
1266	return NULL;
1267}
1268 
1269/*
1270 * find a buffer format based on its uuid
1271 */
1272static pfm_buffer_fmt_t *
1273pfm_find_buffer_fmt(pfm_uuid_t uuid)
1274{
1275	pfm_buffer_fmt_t * fmt;
1276	spin_lock(&pfm_buffer_fmt_lock);
1277	fmt = __pfm_find_buffer_fmt(uuid);
1278	spin_unlock(&pfm_buffer_fmt_lock);
1279	return fmt;
1280}
1281 
1282int
1283pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1284{
1285	int ret = 0;
1286
1287	/* some sanity checks */
1288	if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1289
1290	/* we need at least a handler */
1291	if (fmt->fmt_handler == NULL) return -EINVAL;
1292
1293	/*
1294	 * XXX: need check validity of fmt_arg_size
1295	 */
1296
1297	spin_lock(&pfm_buffer_fmt_lock);
1298
1299	if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1300		printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1301		ret = -EBUSY;
1302		goto out;
1303	} 
1304	list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1305	printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1306
1307out:
1308	spin_unlock(&pfm_buffer_fmt_lock);
1309 	return ret;
1310}
1311EXPORT_SYMBOL(pfm_register_buffer_fmt);
1312
1313int
1314pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1315{
1316	pfm_buffer_fmt_t *fmt;
1317	int ret = 0;
1318
1319	spin_lock(&pfm_buffer_fmt_lock);
1320
1321	fmt = __pfm_find_buffer_fmt(uuid);
1322	if (!fmt) {
1323		printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1324		ret = -EINVAL;
1325		goto out;
1326	}
1327	list_del_init(&fmt->fmt_list);
1328	printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1329
1330out:
1331	spin_unlock(&pfm_buffer_fmt_lock);
1332	return ret;
1333
1334}
1335EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1336
1337extern void update_pal_halt_status(int);
1338
1339static int
1340pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1341{
1342	unsigned long flags;
1343	/*
1344	 * validity checks on cpu_mask have been done upstream
1345	 */
1346	LOCK_PFS(flags);
1347
1348	DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1349		pfm_sessions.pfs_sys_sessions,
1350		pfm_sessions.pfs_task_sessions,
1351		pfm_sessions.pfs_sys_use_dbregs,
1352		is_syswide,
1353		cpu));
1354
1355	if (is_syswide) {
1356		/*
1357		 * cannot mix system wide and per-task sessions
1358		 */
1359		if (pfm_sessions.pfs_task_sessions > 0UL) {
1360			DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1361			  	pfm_sessions.pfs_task_sessions));
1362			goto abort;
1363		}
1364
1365		if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1366
1367		DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1368
1369		pfm_sessions.pfs_sys_session[cpu] = task;
1370
1371		pfm_sessions.pfs_sys_sessions++ ;
1372
1373	} else {
1374		if (pfm_sessions.pfs_sys_sessions) goto abort;
1375		pfm_sessions.pfs_task_sessions++;
1376	}
1377
1378	DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1379		pfm_sessions.pfs_sys_sessions,
1380		pfm_sessions.pfs_task_sessions,
1381		pfm_sessions.pfs_sys_use_dbregs,
1382		is_syswide,
1383		cpu));
1384
1385	/*
1386	 * disable default_idle() to go to PAL_HALT
1387	 */
1388	update_pal_halt_status(0);
1389
1390	UNLOCK_PFS(flags);
1391
1392	return 0;
1393
1394error_conflict:
1395	DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1396  		task_pid_nr(pfm_sessions.pfs_sys_session[cpu]),
1397		cpu));
1398abort:
1399	UNLOCK_PFS(flags);
1400
1401	return -EBUSY;
1402
1403}
1404
1405static int
1406pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1407{
1408	unsigned long flags;
1409	/*
1410	 * validity checks on cpu_mask have been done upstream
1411	 */
1412	LOCK_PFS(flags);
1413
1414	DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1415		pfm_sessions.pfs_sys_sessions,
1416		pfm_sessions.pfs_task_sessions,
1417		pfm_sessions.pfs_sys_use_dbregs,
1418		is_syswide,
1419		cpu));
1420
1421
1422	if (is_syswide) {
1423		pfm_sessions.pfs_sys_session[cpu] = NULL;
1424		/*
1425		 * would not work with perfmon+more than one bit in cpu_mask
1426		 */
1427		if (ctx && ctx->ctx_fl_using_dbreg) {
1428			if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1429				printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1430			} else {
1431				pfm_sessions.pfs_sys_use_dbregs--;
1432			}
1433		}
1434		pfm_sessions.pfs_sys_sessions--;
1435	} else {
1436		pfm_sessions.pfs_task_sessions--;
1437	}
1438	DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1439		pfm_sessions.pfs_sys_sessions,
1440		pfm_sessions.pfs_task_sessions,
1441		pfm_sessions.pfs_sys_use_dbregs,
1442		is_syswide,
1443		cpu));
1444
1445	/*
1446	 * if possible, enable default_idle() to go into PAL_HALT
1447	 */
1448	if (pfm_sessions.pfs_task_sessions == 0 && pfm_sessions.pfs_sys_sessions == 0)
1449		update_pal_halt_status(1);
1450
1451	UNLOCK_PFS(flags);
1452
1453	return 0;
1454}
1455
1456/*
1457 * removes virtual mapping of the sampling buffer.
1458 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1459 * a PROTECT_CTX() section.
1460 */
1461static int
1462pfm_remove_smpl_mapping(struct task_struct *task, void *vaddr, unsigned long size)
1463{
 
1464	int r;
1465
1466	/* sanity checks */
1467	if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1468		printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task_pid_nr(task), task->mm);
1469		return -EINVAL;
1470	}
1471
1472	DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1473
1474	/*
1475	 * does the actual unmapping
1476	 */
1477	down_write(&task->mm->mmap_sem);
1478
1479	DPRINT(("down_write done smpl_vaddr=%p size=%lu\n", vaddr, size));
1480
1481	r = pfm_do_munmap(task->mm, (unsigned long)vaddr, size, 0);
1482
1483	up_write(&task->mm->mmap_sem);
1484	if (r !=0) {
1485		printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task_pid_nr(task), vaddr, size);
1486	}
1487
1488	DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1489
1490	return 0;
1491}
1492
1493/*
1494 * free actual physical storage used by sampling buffer
1495 */
1496#if 0
1497static int
1498pfm_free_smpl_buffer(pfm_context_t *ctx)
1499{
1500	pfm_buffer_fmt_t *fmt;
1501
1502	if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1503
1504	/*
1505	 * we won't use the buffer format anymore
1506	 */
1507	fmt = ctx->ctx_buf_fmt;
1508
1509	DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1510		ctx->ctx_smpl_hdr,
1511		ctx->ctx_smpl_size,
1512		ctx->ctx_smpl_vaddr));
1513
1514	pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1515
1516	/*
1517	 * free the buffer
1518	 */
1519	pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);
1520
1521	ctx->ctx_smpl_hdr  = NULL;
1522	ctx->ctx_smpl_size = 0UL;
1523
1524	return 0;
1525
1526invalid_free:
1527	printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", task_pid_nr(current));
1528	return -EINVAL;
1529}
1530#endif
1531
1532static inline void
1533pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1534{
1535	if (fmt == NULL) return;
1536
1537	pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1538
1539}
1540
1541/*
1542 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1543 * no real gain from having the whole whorehouse mounted. So we don't need
1544 * any operations on the root directory. However, we need a non-trivial
1545 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1546 */
1547static struct vfsmount *pfmfs_mnt __read_mostly;
1548
1549static int __init
1550init_pfm_fs(void)
1551{
1552	int err = register_filesystem(&pfm_fs_type);
1553	if (!err) {
1554		pfmfs_mnt = kern_mount(&pfm_fs_type);
1555		err = PTR_ERR(pfmfs_mnt);
1556		if (IS_ERR(pfmfs_mnt))
1557			unregister_filesystem(&pfm_fs_type);
1558		else
1559			err = 0;
1560	}
1561	return err;
1562}
1563
1564static ssize_t
1565pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
1566{
1567	pfm_context_t *ctx;
1568	pfm_msg_t *msg;
1569	ssize_t ret;
1570	unsigned long flags;
1571  	DECLARE_WAITQUEUE(wait, current);
1572	if (PFM_IS_FILE(filp) == 0) {
1573		printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1574		return -EINVAL;
1575	}
1576
1577	ctx = filp->private_data;
1578	if (ctx == NULL) {
1579		printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", task_pid_nr(current));
1580		return -EINVAL;
1581	}
1582
1583	/*
1584	 * check even when there is no message
1585	 */
1586	if (size < sizeof(pfm_msg_t)) {
1587		DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1588		return -EINVAL;
1589	}
1590
1591	PROTECT_CTX(ctx, flags);
1592
1593  	/*
1594	 * put ourselves on the wait queue
1595	 */
1596  	add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1597
1598
1599  	for(;;) {
1600		/*
1601		 * check wait queue
1602		 */
1603
1604  		set_current_state(TASK_INTERRUPTIBLE);
1605
1606		DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1607
1608		ret = 0;
1609		if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1610
1611		UNPROTECT_CTX(ctx, flags);
1612
1613		/*
1614		 * check non-blocking read
1615		 */
1616      		ret = -EAGAIN;
1617		if(filp->f_flags & O_NONBLOCK) break;
1618
1619		/*
1620		 * check pending signals
1621		 */
1622		if(signal_pending(current)) {
1623			ret = -EINTR;
1624			break;
1625		}
1626      		/*
1627		 * no message, so wait
1628		 */
1629      		schedule();
1630
1631		PROTECT_CTX(ctx, flags);
1632	}
1633	DPRINT(("[%d] back to running ret=%ld\n", task_pid_nr(current), ret));
1634  	set_current_state(TASK_RUNNING);
1635	remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1636
1637	if (ret < 0) goto abort;
1638
1639	ret = -EINVAL;
1640	msg = pfm_get_next_msg(ctx);
1641	if (msg == NULL) {
1642		printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, task_pid_nr(current));
1643		goto abort_locked;
1644	}
1645
1646	DPRINT(("fd=%d type=%d\n", msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1647
1648	ret = -EFAULT;
1649  	if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1650
1651abort_locked:
1652	UNPROTECT_CTX(ctx, flags);
1653abort:
1654	return ret;
1655}
1656
1657static ssize_t
1658pfm_write(struct file *file, const char __user *ubuf,
1659			  size_t size, loff_t *ppos)
1660{
1661	DPRINT(("pfm_write called\n"));
1662	return -EINVAL;
1663}
1664
1665static unsigned int
1666pfm_poll(struct file *filp, poll_table * wait)
1667{
1668	pfm_context_t *ctx;
1669	unsigned long flags;
1670	unsigned int mask = 0;
1671
1672	if (PFM_IS_FILE(filp) == 0) {
1673		printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1674		return 0;
1675	}
1676
1677	ctx = filp->private_data;
1678	if (ctx == NULL) {
1679		printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", task_pid_nr(current));
1680		return 0;
1681	}
1682
1683
1684	DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1685
1686	poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1687
1688	PROTECT_CTX(ctx, flags);
1689
1690	if (PFM_CTXQ_EMPTY(ctx) == 0)
1691		mask =  POLLIN | POLLRDNORM;
1692
1693	UNPROTECT_CTX(ctx, flags);
1694
1695	DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1696
1697	return mask;
1698}
1699
1700static long
1701pfm_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1702{
1703	DPRINT(("pfm_ioctl called\n"));
1704	return -EINVAL;
1705}
1706
1707/*
1708 * interrupt cannot be masked when coming here
1709 */
1710static inline int
1711pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1712{
1713	int ret;
1714
1715	ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1716
1717	DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1718		task_pid_nr(current),
1719		fd,
1720		on,
1721		ctx->ctx_async_queue, ret));
1722
1723	return ret;
1724}
1725
1726static int
1727pfm_fasync(int fd, struct file *filp, int on)
1728{
1729	pfm_context_t *ctx;
1730	int ret;
1731
1732	if (PFM_IS_FILE(filp) == 0) {
1733		printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", task_pid_nr(current));
1734		return -EBADF;
1735	}
1736
1737	ctx = filp->private_data;
1738	if (ctx == NULL) {
1739		printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", task_pid_nr(current));
1740		return -EBADF;
1741	}
1742	/*
1743	 * we cannot mask interrupts during this call because this may
1744	 * may go to sleep if memory is not readily avalaible.
1745	 *
1746	 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1747	 * done in caller. Serialization of this function is ensured by caller.
1748	 */
1749	ret = pfm_do_fasync(fd, filp, ctx, on);
1750
1751
1752	DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1753		fd,
1754		on,
1755		ctx->ctx_async_queue, ret));
1756
1757	return ret;
1758}
1759
1760#ifdef CONFIG_SMP
1761/*
1762 * this function is exclusively called from pfm_close().
1763 * The context is not protected at that time, nor are interrupts
1764 * on the remote CPU. That's necessary to avoid deadlocks.
1765 */
1766static void
1767pfm_syswide_force_stop(void *info)
1768{
1769	pfm_context_t   *ctx = (pfm_context_t *)info;
1770	struct pt_regs *regs = task_pt_regs(current);
1771	struct task_struct *owner;
1772	unsigned long flags;
1773	int ret;
1774
1775	if (ctx->ctx_cpu != smp_processor_id()) {
1776		printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d  but on CPU%d\n",
1777			ctx->ctx_cpu,
1778			smp_processor_id());
1779		return;
1780	}
1781	owner = GET_PMU_OWNER();
1782	if (owner != ctx->ctx_task) {
1783		printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1784			smp_processor_id(),
1785			task_pid_nr(owner), task_pid_nr(ctx->ctx_task));
1786		return;
1787	}
1788	if (GET_PMU_CTX() != ctx) {
1789		printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1790			smp_processor_id(),
1791			GET_PMU_CTX(), ctx);
1792		return;
1793	}
1794
1795	DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), task_pid_nr(ctx->ctx_task)));
1796	/*
1797	 * the context is already protected in pfm_close(), we simply
1798	 * need to mask interrupts to avoid a PMU interrupt race on
1799	 * this CPU
1800	 */
1801	local_irq_save(flags);
1802
1803	ret = pfm_context_unload(ctx, NULL, 0, regs);
1804	if (ret) {
1805		DPRINT(("context_unload returned %d\n", ret));
1806	}
1807
1808	/*
1809	 * unmask interrupts, PMU interrupts are now spurious here
1810	 */
1811	local_irq_restore(flags);
1812}
1813
1814static void
1815pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1816{
1817	int ret;
1818
1819	DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1820	ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 1);
1821	DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1822}
1823#endif /* CONFIG_SMP */
1824
1825/*
1826 * called for each close(). Partially free resources.
1827 * When caller is self-monitoring, the context is unloaded.
1828 */
1829static int
1830pfm_flush(struct file *filp, fl_owner_t id)
1831{
1832	pfm_context_t *ctx;
1833	struct task_struct *task;
1834	struct pt_regs *regs;
1835	unsigned long flags;
1836	unsigned long smpl_buf_size = 0UL;
1837	void *smpl_buf_vaddr = NULL;
1838	int state, is_system;
1839
1840	if (PFM_IS_FILE(filp) == 0) {
1841		DPRINT(("bad magic for\n"));
1842		return -EBADF;
1843	}
1844
1845	ctx = filp->private_data;
1846	if (ctx == NULL) {
1847		printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", task_pid_nr(current));
1848		return -EBADF;
1849	}
1850
1851	/*
1852	 * remove our file from the async queue, if we use this mode.
1853	 * This can be done without the context being protected. We come
1854	 * here when the context has become unreachable by other tasks.
1855	 *
1856	 * We may still have active monitoring at this point and we may
1857	 * end up in pfm_overflow_handler(). However, fasync_helper()
1858	 * operates with interrupts disabled and it cleans up the
1859	 * queue. If the PMU handler is called prior to entering
1860	 * fasync_helper() then it will send a signal. If it is
1861	 * invoked after, it will find an empty queue and no
1862	 * signal will be sent. In both case, we are safe
1863	 */
1864	PROTECT_CTX(ctx, flags);
1865
1866	state     = ctx->ctx_state;
1867	is_system = ctx->ctx_fl_system;
1868
1869	task = PFM_CTX_TASK(ctx);
1870	regs = task_pt_regs(task);
1871
1872	DPRINT(("ctx_state=%d is_current=%d\n",
1873		state,
1874		task == current ? 1 : 0));
1875
1876	/*
1877	 * if state == UNLOADED, then task is NULL
1878	 */
1879
1880	/*
1881	 * we must stop and unload because we are losing access to the context.
1882	 */
1883	if (task == current) {
1884#ifdef CONFIG_SMP
1885		/*
1886		 * the task IS the owner but it migrated to another CPU: that's bad
1887		 * but we must handle this cleanly. Unfortunately, the kernel does
1888		 * not provide a mechanism to block migration (while the context is loaded).
1889		 *
1890		 * We need to release the resource on the ORIGINAL cpu.
1891		 */
1892		if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1893
1894			DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1895			/*
1896			 * keep context protected but unmask interrupt for IPI
1897			 */
1898			local_irq_restore(flags);
1899
1900			pfm_syswide_cleanup_other_cpu(ctx);
1901
1902			/*
1903			 * restore interrupt masking
1904			 */
1905			local_irq_save(flags);
1906
1907			/*
1908			 * context is unloaded at this point
1909			 */
1910		} else
1911#endif /* CONFIG_SMP */
1912		{
1913
1914			DPRINT(("forcing unload\n"));
1915			/*
1916		 	* stop and unload, returning with state UNLOADED
1917		 	* and session unreserved.
1918		 	*/
1919			pfm_context_unload(ctx, NULL, 0, regs);
1920
1921			DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1922		}
1923	}
1924
1925	/*
1926	 * remove virtual mapping, if any, for the calling task.
1927	 * cannot reset ctx field until last user is calling close().
1928	 *
1929	 * ctx_smpl_vaddr must never be cleared because it is needed
1930	 * by every task with access to the context
1931	 *
1932	 * When called from do_exit(), the mm context is gone already, therefore
1933	 * mm is NULL, i.e., the VMA is already gone  and we do not have to
1934	 * do anything here
1935	 */
1936	if (ctx->ctx_smpl_vaddr && current->mm) {
1937		smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1938		smpl_buf_size  = ctx->ctx_smpl_size;
1939	}
1940
1941	UNPROTECT_CTX(ctx, flags);
1942
1943	/*
1944	 * if there was a mapping, then we systematically remove it
1945	 * at this point. Cannot be done inside critical section
1946	 * because some VM function reenables interrupts.
1947	 *
1948	 */
1949	if (smpl_buf_vaddr) pfm_remove_smpl_mapping(current, smpl_buf_vaddr, smpl_buf_size);
1950
1951	return 0;
1952}
1953/*
1954 * called either on explicit close() or from exit_files(). 
1955 * Only the LAST user of the file gets to this point, i.e., it is
1956 * called only ONCE.
1957 *
1958 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero 
1959 * (fput()),i.e, last task to access the file. Nobody else can access the 
1960 * file at this point.
1961 *
1962 * When called from exit_files(), the VMA has been freed because exit_mm()
1963 * is executed before exit_files().
1964 *
1965 * When called from exit_files(), the current task is not yet ZOMBIE but we
1966 * flush the PMU state to the context. 
1967 */
1968static int
1969pfm_close(struct inode *inode, struct file *filp)
1970{
1971	pfm_context_t *ctx;
1972	struct task_struct *task;
1973	struct pt_regs *regs;
1974  	DECLARE_WAITQUEUE(wait, current);
1975	unsigned long flags;
1976	unsigned long smpl_buf_size = 0UL;
1977	void *smpl_buf_addr = NULL;
1978	int free_possible = 1;
1979	int state, is_system;
1980
1981	DPRINT(("pfm_close called private=%p\n", filp->private_data));
1982
1983	if (PFM_IS_FILE(filp) == 0) {
1984		DPRINT(("bad magic\n"));
1985		return -EBADF;
1986	}
1987	
1988	ctx = filp->private_data;
1989	if (ctx == NULL) {
1990		printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", task_pid_nr(current));
1991		return -EBADF;
1992	}
1993
1994	PROTECT_CTX(ctx, flags);
1995
1996	state     = ctx->ctx_state;
1997	is_system = ctx->ctx_fl_system;
1998
1999	task = PFM_CTX_TASK(ctx);
2000	regs = task_pt_regs(task);
2001
2002	DPRINT(("ctx_state=%d is_current=%d\n", 
2003		state,
2004		task == current ? 1 : 0));
2005
2006	/*
2007	 * if task == current, then pfm_flush() unloaded the context
2008	 */
2009	if (state == PFM_CTX_UNLOADED) goto doit;
2010
2011	/*
2012	 * context is loaded/masked and task != current, we need to
2013	 * either force an unload or go zombie
2014	 */
2015
2016	/*
2017	 * The task is currently blocked or will block after an overflow.
2018	 * we must force it to wakeup to get out of the
2019	 * MASKED state and transition to the unloaded state by itself.
2020	 *
2021	 * This situation is only possible for per-task mode
2022	 */
2023	if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
2024
2025		/*
2026		 * set a "partial" zombie state to be checked
2027		 * upon return from down() in pfm_handle_work().
2028		 *
2029		 * We cannot use the ZOMBIE state, because it is checked
2030		 * by pfm_load_regs() which is called upon wakeup from down().
2031		 * In such case, it would free the context and then we would
2032		 * return to pfm_handle_work() which would access the
2033		 * stale context. Instead, we set a flag invisible to pfm_load_regs()
2034		 * but visible to pfm_handle_work().
2035		 *
2036		 * For some window of time, we have a zombie context with
2037		 * ctx_state = MASKED  and not ZOMBIE
2038		 */
2039		ctx->ctx_fl_going_zombie = 1;
2040
2041		/*
2042		 * force task to wake up from MASKED state
2043		 */
2044		complete(&ctx->ctx_restart_done);
2045
2046		DPRINT(("waking up ctx_state=%d\n", state));
2047
2048		/*
2049		 * put ourself to sleep waiting for the other
2050		 * task to report completion
2051		 *
2052		 * the context is protected by mutex, therefore there
2053		 * is no risk of being notified of completion before
2054		 * begin actually on the waitq.
2055		 */
2056  		set_current_state(TASK_INTERRUPTIBLE);
2057  		add_wait_queue(&ctx->ctx_zombieq, &wait);
2058
2059		UNPROTECT_CTX(ctx, flags);
2060
2061		/*
2062		 * XXX: check for signals :
2063		 * 	- ok for explicit close
2064		 * 	- not ok when coming from exit_files()
2065		 */
2066      		schedule();
2067
2068
2069		PROTECT_CTX(ctx, flags);
2070
2071
2072		remove_wait_queue(&ctx->ctx_zombieq, &wait);
2073  		set_current_state(TASK_RUNNING);
2074
2075		/*
2076		 * context is unloaded at this point
2077		 */
2078		DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2079	}
2080	else if (task != current) {
2081#ifdef CONFIG_SMP
2082		/*
2083	 	 * switch context to zombie state
2084	 	 */
2085		ctx->ctx_state = PFM_CTX_ZOMBIE;
2086
2087		DPRINT(("zombie ctx for [%d]\n", task_pid_nr(task)));
2088		/*
2089		 * cannot free the context on the spot. deferred until
2090		 * the task notices the ZOMBIE state
2091		 */
2092		free_possible = 0;
2093#else
2094		pfm_context_unload(ctx, NULL, 0, regs);
2095#endif
2096	}
2097
2098doit:
2099	/* reload state, may have changed during  opening of critical section */
2100	state = ctx->ctx_state;
2101
2102	/*
2103	 * the context is still attached to a task (possibly current)
2104	 * we cannot destroy it right now
2105	 */
2106
2107	/*
2108	 * we must free the sampling buffer right here because
2109	 * we cannot rely on it being cleaned up later by the
2110	 * monitored task. It is not possible to free vmalloc'ed
2111	 * memory in pfm_load_regs(). Instead, we remove the buffer
2112	 * now. should there be subsequent PMU overflow originally
2113	 * meant for sampling, the will be converted to spurious
2114	 * and that's fine because the monitoring tools is gone anyway.
2115	 */
2116	if (ctx->ctx_smpl_hdr) {
2117		smpl_buf_addr = ctx->ctx_smpl_hdr;
2118		smpl_buf_size = ctx->ctx_smpl_size;
2119		/* no more sampling */
2120		ctx->ctx_smpl_hdr = NULL;
2121		ctx->ctx_fl_is_sampling = 0;
2122	}
2123
2124	DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2125		state,
2126		free_possible,
2127		smpl_buf_addr,
2128		smpl_buf_size));
2129
2130	if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2131
2132	/*
2133	 * UNLOADED that the session has already been unreserved.
2134	 */
2135	if (state == PFM_CTX_ZOMBIE) {
2136		pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2137	}
2138
2139	/*
2140	 * disconnect file descriptor from context must be done
2141	 * before we unlock.
2142	 */
2143	filp->private_data = NULL;
2144
2145	/*
2146	 * if we free on the spot, the context is now completely unreachable
2147	 * from the callers side. The monitored task side is also cut, so we
2148	 * can freely cut.
2149	 *
2150	 * If we have a deferred free, only the caller side is disconnected.
2151	 */
2152	UNPROTECT_CTX(ctx, flags);
2153
2154	/*
2155	 * All memory free operations (especially for vmalloc'ed memory)
2156	 * MUST be done with interrupts ENABLED.
2157	 */
2158	if (smpl_buf_addr)  pfm_rvfree(smpl_buf_addr, smpl_buf_size);
2159
2160	/*
2161	 * return the memory used by the context
2162	 */
2163	if (free_possible) pfm_context_free(ctx);
2164
2165	return 0;
2166}
2167
2168static int
2169pfm_no_open(struct inode *irrelevant, struct file *dontcare)
2170{
2171	DPRINT(("pfm_no_open called\n"));
2172	return -ENXIO;
2173}
2174
2175
2176
2177static const struct file_operations pfm_file_ops = {
2178	.llseek		= no_llseek,
2179	.read		= pfm_read,
2180	.write		= pfm_write,
2181	.poll		= pfm_poll,
2182	.unlocked_ioctl = pfm_ioctl,
2183	.open		= pfm_no_open,	/* special open code to disallow open via /proc */
2184	.fasync		= pfm_fasync,
2185	.release	= pfm_close,
2186	.flush		= pfm_flush
2187};
2188
2189static int
2190pfmfs_delete_dentry(const struct dentry *dentry)
2191{
2192	return 1;
2193}
2194
2195static char *pfmfs_dname(struct dentry *dentry, char *buffer, int buflen)
2196{
2197	return dynamic_dname(dentry, buffer, buflen, "pfm:[%lu]",
2198			     dentry->d_inode->i_ino);
2199}
2200
2201static const struct dentry_operations pfmfs_dentry_operations = {
2202	.d_delete = pfmfs_delete_dentry,
2203	.d_dname = pfmfs_dname,
2204};
2205
2206
2207static struct file *
2208pfm_alloc_file(pfm_context_t *ctx)
2209{
2210	struct file *file;
2211	struct inode *inode;
2212	struct path path;
2213	struct qstr this = { .name = "" };
2214
2215	/*
2216	 * allocate a new inode
2217	 */
2218	inode = new_inode(pfmfs_mnt->mnt_sb);
2219	if (!inode)
2220		return ERR_PTR(-ENOMEM);
2221
2222	DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2223
2224	inode->i_mode = S_IFCHR|S_IRUGO;
2225	inode->i_uid  = current_fsuid();
2226	inode->i_gid  = current_fsgid();
2227
2228	/*
2229	 * allocate a new dcache entry
2230	 */
2231	path.dentry = d_alloc(pfmfs_mnt->mnt_sb->s_root, &this);
2232	if (!path.dentry) {
2233		iput(inode);
2234		return ERR_PTR(-ENOMEM);
2235	}
2236	path.mnt = mntget(pfmfs_mnt);
2237
2238	d_add(path.dentry, inode);
2239
2240	file = alloc_file(&path, FMODE_READ, &pfm_file_ops);
2241	if (!file) {
2242		path_put(&path);
2243		return ERR_PTR(-ENFILE);
2244	}
2245
2246	file->f_flags = O_RDONLY;
2247	file->private_data = ctx;
2248
2249	return file;
2250}
2251
2252static int
2253pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2254{
2255	DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2256
2257	while (size > 0) {
2258		unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2259
2260
2261		if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2262			return -ENOMEM;
2263
2264		addr  += PAGE_SIZE;
2265		buf   += PAGE_SIZE;
2266		size  -= PAGE_SIZE;
2267	}
2268	return 0;
2269}
2270
2271/*
2272 * allocate a sampling buffer and remaps it into the user address space of the task
2273 */
2274static int
2275pfm_smpl_buffer_alloc(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2276{
2277	struct mm_struct *mm = task->mm;
2278	struct vm_area_struct *vma = NULL;
2279	unsigned long size;
2280	void *smpl_buf;
2281
2282
2283	/*
2284	 * the fixed header + requested size and align to page boundary
2285	 */
2286	size = PAGE_ALIGN(rsize);
2287
2288	DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2289
2290	/*
2291	 * check requested size to avoid Denial-of-service attacks
2292	 * XXX: may have to refine this test
2293	 * Check against address space limit.
2294	 *
2295	 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2296	 * 	return -ENOMEM;
2297	 */
2298	if (size > task_rlimit(task, RLIMIT_MEMLOCK))
2299		return -ENOMEM;
2300
2301	/*
2302	 * We do the easy to undo allocations first.
2303 	 *
2304	 * pfm_rvmalloc(), clears the buffer, so there is no leak
2305	 */
2306	smpl_buf = pfm_rvmalloc(size);
2307	if (smpl_buf == NULL) {
2308		DPRINT(("Can't allocate sampling buffer\n"));
2309		return -ENOMEM;
2310	}
2311
2312	DPRINT(("smpl_buf @%p\n", smpl_buf));
2313
2314	/* allocate vma */
2315	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2316	if (!vma) {
2317		DPRINT(("Cannot allocate vma\n"));
2318		goto error_kmem;
2319	}
2320	INIT_LIST_HEAD(&vma->anon_vma_chain);
2321
2322	/*
2323	 * partially initialize the vma for the sampling buffer
2324	 */
2325	vma->vm_mm	     = mm;
2326	vma->vm_file	     = filp;
2327	vma->vm_flags	     = VM_READ| VM_MAYREAD |VM_RESERVED;
2328	vma->vm_page_prot    = PAGE_READONLY; /* XXX may need to change */
2329
2330	/*
2331	 * Now we have everything we need and we can initialize
2332	 * and connect all the data structures
2333	 */
2334
2335	ctx->ctx_smpl_hdr   = smpl_buf;
2336	ctx->ctx_smpl_size  = size; /* aligned size */
2337
2338	/*
2339	 * Let's do the difficult operations next.
2340	 *
2341	 * now we atomically find some area in the address space and
2342	 * remap the buffer in it.
2343	 */
2344	down_write(&task->mm->mmap_sem);
2345
2346	/* find some free area in address space, must have mmap sem held */
2347	vma->vm_start = pfm_get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS, 0);
2348	if (vma->vm_start == 0UL) {
2349		DPRINT(("Cannot find unmapped area for size %ld\n", size));
2350		up_write(&task->mm->mmap_sem);
2351		goto error;
2352	}
2353	vma->vm_end = vma->vm_start + size;
2354	vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2355
2356	DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2357
2358	/* can only be applied to current task, need to have the mm semaphore held when called */
2359	if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2360		DPRINT(("Can't remap buffer\n"));
2361		up_write(&task->mm->mmap_sem);
2362		goto error;
2363	}
2364
2365	get_file(filp);
2366
2367	/*
2368	 * now insert the vma in the vm list for the process, must be
2369	 * done with mmap lock held
2370	 */
2371	insert_vm_struct(mm, vma);
2372
2373	mm->total_vm  += size >> PAGE_SHIFT;
2374	vm_stat_account(vma->vm_mm, vma->vm_flags, vma->vm_file,
2375							vma_pages(vma));
2376	up_write(&task->mm->mmap_sem);
2377
2378	/*
2379	 * keep track of user level virtual address
2380	 */
2381	ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2382	*(unsigned long *)user_vaddr = vma->vm_start;
2383
2384	return 0;
2385
2386error:
2387	kmem_cache_free(vm_area_cachep, vma);
2388error_kmem:
2389	pfm_rvfree(smpl_buf, size);
2390
2391	return -ENOMEM;
2392}
2393
2394/*
2395 * XXX: do something better here
2396 */
2397static int
2398pfm_bad_permissions(struct task_struct *task)
2399{
2400	const struct cred *tcred;
2401	uid_t uid = current_uid();
2402	gid_t gid = current_gid();
2403	int ret;
2404
2405	rcu_read_lock();
2406	tcred = __task_cred(task);
2407
2408	/* inspired by ptrace_attach() */
2409	DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2410		uid,
2411		gid,
2412		tcred->euid,
2413		tcred->suid,
2414		tcred->uid,
2415		tcred->egid,
2416		tcred->sgid));
2417
2418	ret = ((uid != tcred->euid)
2419	       || (uid != tcred->suid)
2420	       || (uid != tcred->uid)
2421	       || (gid != tcred->egid)
2422	       || (gid != tcred->sgid)
2423	       || (gid != tcred->gid)) && !capable(CAP_SYS_PTRACE);
2424
2425	rcu_read_unlock();
2426	return ret;
2427}
2428
2429static int
2430pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2431{
2432	int ctx_flags;
2433
2434	/* valid signal */
2435
2436	ctx_flags = pfx->ctx_flags;
2437
2438	if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2439
2440		/*
2441		 * cannot block in this mode
2442		 */
2443		if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2444			DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2445			return -EINVAL;
2446		}
2447	} else {
2448	}
2449	/* probably more to add here */
2450
2451	return 0;
2452}
2453
2454static int
2455pfm_setup_buffer_fmt(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned int ctx_flags,
2456		     unsigned int cpu, pfarg_context_t *arg)
2457{
2458	pfm_buffer_fmt_t *fmt = NULL;
2459	unsigned long size = 0UL;
2460	void *uaddr = NULL;
2461	void *fmt_arg = NULL;
2462	int ret = 0;
2463#define PFM_CTXARG_BUF_ARG(a)	(pfm_buffer_fmt_t *)(a+1)
2464
2465	/* invoke and lock buffer format, if found */
2466	fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2467	if (fmt == NULL) {
2468		DPRINT(("[%d] cannot find buffer format\n", task_pid_nr(task)));
2469		return -EINVAL;
2470	}
2471
2472	/*
2473	 * buffer argument MUST be contiguous to pfarg_context_t
2474	 */
2475	if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2476
2477	ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2478
2479	DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task_pid_nr(task), ctx_flags, cpu, fmt_arg, ret));
2480
2481	if (ret) goto error;
2482
2483	/* link buffer format and context */
2484	ctx->ctx_buf_fmt = fmt;
2485	ctx->ctx_fl_is_sampling = 1; /* assume record() is defined */
2486
2487	/*
2488	 * check if buffer format wants to use perfmon buffer allocation/mapping service
2489	 */
2490	ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2491	if (ret) goto error;
2492
2493	if (size) {
2494		/*
2495		 * buffer is always remapped into the caller's address space
2496		 */
2497		ret = pfm_smpl_buffer_alloc(current, filp, ctx, size, &uaddr);
2498		if (ret) goto error;
2499
2500		/* keep track of user address of buffer */
2501		arg->ctx_smpl_vaddr = uaddr;
2502	}
2503	ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2504
2505error:
2506	return ret;
2507}
2508
2509static void
2510pfm_reset_pmu_state(pfm_context_t *ctx)
2511{
2512	int i;
2513
2514	/*
2515	 * install reset values for PMC.
2516	 */
2517	for (i=1; PMC_IS_LAST(i) == 0; i++) {
2518		if (PMC_IS_IMPL(i) == 0) continue;
2519		ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2520		DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2521	}
2522	/*
2523	 * PMD registers are set to 0UL when the context in memset()
2524	 */
2525
2526	/*
2527	 * On context switched restore, we must restore ALL pmc and ALL pmd even
2528	 * when they are not actively used by the task. In UP, the incoming process
2529	 * may otherwise pick up left over PMC, PMD state from the previous process.
2530	 * As opposed to PMD, stale PMC can cause harm to the incoming
2531	 * process because they may change what is being measured.
2532	 * Therefore, we must systematically reinstall the entire
2533	 * PMC state. In SMP, the same thing is possible on the
2534	 * same CPU but also on between 2 CPUs.
2535	 *
2536	 * The problem with PMD is information leaking especially
2537	 * to user level when psr.sp=0
2538	 *
2539	 * There is unfortunately no easy way to avoid this problem
2540	 * on either UP or SMP. This definitively slows down the
2541	 * pfm_load_regs() function.
2542	 */
2543
2544	 /*
2545	  * bitmask of all PMCs accessible to this context
2546	  *
2547	  * PMC0 is treated differently.
2548	  */
2549	ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2550
2551	/*
2552	 * bitmask of all PMDs that are accessible to this context
2553	 */
2554	ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2555
2556	DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2557
2558	/*
2559	 * useful in case of re-enable after disable
2560	 */
2561	ctx->ctx_used_ibrs[0] = 0UL;
2562	ctx->ctx_used_dbrs[0] = 0UL;
2563}
2564
2565static int
2566pfm_ctx_getsize(void *arg, size_t *sz)
2567{
2568	pfarg_context_t *req = (pfarg_context_t *)arg;
2569	pfm_buffer_fmt_t *fmt;
2570
2571	*sz = 0;
2572
2573	if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2574
2575	fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2576	if (fmt == NULL) {
2577		DPRINT(("cannot find buffer format\n"));
2578		return -EINVAL;
2579	}
2580	/* get just enough to copy in user parameters */
2581	*sz = fmt->fmt_arg_size;
2582	DPRINT(("arg_size=%lu\n", *sz));
2583
2584	return 0;
2585}
2586
2587
2588
2589/*
2590 * cannot attach if :
2591 * 	- kernel task
2592 * 	- task not owned by caller
2593 * 	- task incompatible with context mode
2594 */
2595static int
2596pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2597{
2598	/*
2599	 * no kernel task or task not owner by caller
2600	 */
2601	if (task->mm == NULL) {
2602		DPRINT(("task [%d] has not memory context (kernel thread)\n", task_pid_nr(task)));
2603		return -EPERM;
2604	}
2605	if (pfm_bad_permissions(task)) {
2606		DPRINT(("no permission to attach to  [%d]\n", task_pid_nr(task)));
2607		return -EPERM;
2608	}
2609	/*
2610	 * cannot block in self-monitoring mode
2611	 */
2612	if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2613		DPRINT(("cannot load a blocking context on self for [%d]\n", task_pid_nr(task)));
2614		return -EINVAL;
2615	}
2616
2617	if (task->exit_state == EXIT_ZOMBIE) {
2618		DPRINT(("cannot attach to  zombie task [%d]\n", task_pid_nr(task)));
2619		return -EBUSY;
2620	}
2621
2622	/*
2623	 * always ok for self
2624	 */
2625	if (task == current) return 0;
2626
2627	if (!task_is_stopped_or_traced(task)) {
2628		DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task_pid_nr(task), task->state));
2629		return -EBUSY;
2630	}
2631	/*
2632	 * make sure the task is off any CPU
2633	 */
2634	wait_task_inactive(task, 0);
2635
2636	/* more to come... */
2637
2638	return 0;
2639}
2640
2641static int
2642pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2643{
2644	struct task_struct *p = current;
2645	int ret;
2646
2647	/* XXX: need to add more checks here */
2648	if (pid < 2) return -EPERM;
2649
2650	if (pid != task_pid_vnr(current)) {
2651
2652		read_lock(&tasklist_lock);
2653
2654		p = find_task_by_vpid(pid);
2655
2656		/* make sure task cannot go away while we operate on it */
2657		if (p) get_task_struct(p);
2658
2659		read_unlock(&tasklist_lock);
2660
2661		if (p == NULL) return -ESRCH;
2662	}
2663
2664	ret = pfm_task_incompatible(ctx, p);
2665	if (ret == 0) {
2666		*task = p;
2667	} else if (p != current) {
2668		pfm_put_task(p);
2669	}
2670	return ret;
2671}
2672
2673
2674
2675static int
2676pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2677{
2678	pfarg_context_t *req = (pfarg_context_t *)arg;
2679	struct file *filp;
2680	struct path path;
2681	int ctx_flags;
2682	int fd;
2683	int ret;
2684
2685	/* let's check the arguments first */
2686	ret = pfarg_is_sane(current, req);
2687	if (ret < 0)
2688		return ret;
2689
2690	ctx_flags = req->ctx_flags;
2691
2692	ret = -ENOMEM;
2693
2694	fd = get_unused_fd();
2695	if (fd < 0)
2696		return fd;
2697
2698	ctx = pfm_context_alloc(ctx_flags);
2699	if (!ctx)
2700		goto error;
2701
2702	filp = pfm_alloc_file(ctx);
2703	if (IS_ERR(filp)) {
2704		ret = PTR_ERR(filp);
2705		goto error_file;
2706	}
2707
2708	req->ctx_fd = ctx->ctx_fd = fd;
2709
2710	/*
2711	 * does the user want to sample?
2712	 */
2713	if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2714		ret = pfm_setup_buffer_fmt(current, filp, ctx, ctx_flags, 0, req);
2715		if (ret)
2716			goto buffer_error;
2717	}
2718
2719	DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d\n",
2720		ctx,
2721		ctx_flags,
2722		ctx->ctx_fl_system,
2723		ctx->ctx_fl_block,
2724		ctx->ctx_fl_excl_idle,
2725		ctx->ctx_fl_no_msg,
2726		ctx->ctx_fd));
2727
2728	/*
2729	 * initialize soft PMU state
2730	 */
2731	pfm_reset_pmu_state(ctx);
2732
2733	fd_install(fd, filp);
2734
2735	return 0;
2736
2737buffer_error:
2738	path = filp->f_path;
2739	put_filp(filp);
2740	path_put(&path);
2741
2742	if (ctx->ctx_buf_fmt) {
2743		pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2744	}
2745error_file:
2746	pfm_context_free(ctx);
2747
2748error:
2749	put_unused_fd(fd);
2750	return ret;
2751}
2752
2753static inline unsigned long
2754pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2755{
2756	unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2757	unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2758	extern unsigned long carta_random32 (unsigned long seed);
2759
2760	if (reg->flags & PFM_REGFL_RANDOM) {
2761		new_seed = carta_random32(old_seed);
2762		val -= (old_seed & mask);	/* counter values are negative numbers! */
2763		if ((mask >> 32) != 0)
2764			/* construct a full 64-bit random value: */
2765			new_seed |= carta_random32(old_seed >> 32) << 32;
2766		reg->seed = new_seed;
2767	}
2768	reg->lval = val;
2769	return val;
2770}
2771
2772static void
2773pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2774{
2775	unsigned long mask = ovfl_regs[0];
2776	unsigned long reset_others = 0UL;
2777	unsigned long val;
2778	int i;
2779
2780	/*
2781	 * now restore reset value on sampling overflowed counters
2782	 */
2783	mask >>= PMU_FIRST_COUNTER;
2784	for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2785
2786		if ((mask & 0x1UL) == 0UL) continue;
2787
2788		ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2789		reset_others        |= ctx->ctx_pmds[i].reset_pmds[0];
2790
2791		DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2792	}
2793
2794	/*
2795	 * Now take care of resetting the other registers
2796	 */
2797	for(i = 0; reset_others; i++, reset_others >>= 1) {
2798
2799		if ((reset_others & 0x1) == 0) continue;
2800
2801		ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2802
2803		DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2804			  is_long_reset ? "long" : "short", i, val));
2805	}
2806}
2807
2808static void
2809pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2810{
2811	unsigned long mask = ovfl_regs[0];
2812	unsigned long reset_others = 0UL;
2813	unsigned long val;
2814	int i;
2815
2816	DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2817
2818	if (ctx->ctx_state == PFM_CTX_MASKED) {
2819		pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2820		return;
2821	}
2822
2823	/*
2824	 * now restore reset value on sampling overflowed counters
2825	 */
2826	mask >>= PMU_FIRST_COUNTER;
2827	for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2828
2829		if ((mask & 0x1UL) == 0UL) continue;
2830
2831		val           = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2832		reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2833
2834		DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2835
2836		pfm_write_soft_counter(ctx, i, val);
2837	}
2838
2839	/*
2840	 * Now take care of resetting the other registers
2841	 */
2842	for(i = 0; reset_others; i++, reset_others >>= 1) {
2843
2844		if ((reset_others & 0x1) == 0) continue;
2845
2846		val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2847
2848		if (PMD_IS_COUNTING(i)) {
2849			pfm_write_soft_counter(ctx, i, val);
2850		} else {
2851			ia64_set_pmd(i, val);
2852		}
2853		DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2854			  is_long_reset ? "long" : "short", i, val));
2855	}
2856	ia64_srlz_d();
2857}
2858
2859static int
2860pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2861{
2862	struct task_struct *task;
2863	pfarg_reg_t *req = (pfarg_reg_t *)arg;
2864	unsigned long value, pmc_pm;
2865	unsigned long smpl_pmds, reset_pmds, impl_pmds;
2866	unsigned int cnum, reg_flags, flags, pmc_type;
2867	int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2868	int is_monitor, is_counting, state;
2869	int ret = -EINVAL;
2870	pfm_reg_check_t	wr_func;
2871#define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2872
2873	state     = ctx->ctx_state;
2874	is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2875	is_system = ctx->ctx_fl_system;
2876	task      = ctx->ctx_task;
2877	impl_pmds = pmu_conf->impl_pmds[0];
2878
2879	if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2880
2881	if (is_loaded) {
2882		/*
2883		 * In system wide and when the context is loaded, access can only happen
2884		 * when the caller is running on the CPU being monitored by the session.
2885		 * It does not have to be the owner (ctx_task) of the context per se.
2886		 */
2887		if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2888			DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2889			return -EBUSY;
2890		}
2891		can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2892	}
2893	expert_mode = pfm_sysctl.expert_mode; 
2894
2895	for (i = 0; i < count; i++, req++) {
2896
2897		cnum       = req->reg_num;
2898		reg_flags  = req->reg_flags;
2899		value      = req->reg_value;
2900		smpl_pmds  = req->reg_smpl_pmds[0];
2901		reset_pmds = req->reg_reset_pmds[0];
2902		flags      = 0;
2903
2904
2905		if (cnum >= PMU_MAX_PMCS) {
2906			DPRINT(("pmc%u is invalid\n", cnum));
2907			goto error;
2908		}
2909
2910		pmc_type   = pmu_conf->pmc_desc[cnum].type;
2911		pmc_pm     = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2912		is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2913		is_monitor  = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2914
2915		/*
2916		 * we reject all non implemented PMC as well
2917		 * as attempts to modify PMC[0-3] which are used
2918		 * as status registers by the PMU
2919		 */
2920		if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2921			DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2922			goto error;
2923		}
2924		wr_func = pmu_conf->pmc_desc[cnum].write_check;
2925		/*
2926		 * If the PMC is a monitor, then if the value is not the default:
2927		 * 	- system-wide session: PMCx.pm=1 (privileged monitor)
2928		 * 	- per-task           : PMCx.pm=0 (user monitor)
2929		 */
2930		if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2931			DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2932				cnum,
2933				pmc_pm,
2934				is_system));
2935			goto error;
2936		}
2937
2938		if (is_counting) {
2939			/*
2940		 	 * enforce generation of overflow interrupt. Necessary on all
2941		 	 * CPUs.
2942		 	 */
2943			value |= 1 << PMU_PMC_OI;
2944
2945			if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2946				flags |= PFM_REGFL_OVFL_NOTIFY;
2947			}
2948
2949			if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2950
2951			/* verify validity of smpl_pmds */
2952			if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2953				DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2954				goto error;
2955			}
2956
2957			/* verify validity of reset_pmds */
2958			if ((reset_pmds & impl_pmds) != reset_pmds) {
2959				DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2960				goto error;
2961			}
2962		} else {
2963			if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2964				DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2965				goto error;
2966			}
2967			/* eventid on non-counting monitors are ignored */
2968		}
2969
2970		/*
2971		 * execute write checker, if any
2972		 */
2973		if (likely(expert_mode == 0 && wr_func)) {
2974			ret = (*wr_func)(task, ctx, cnum, &value, regs);
2975			if (ret) goto error;
2976			ret = -EINVAL;
2977		}
2978
2979		/*
2980		 * no error on this register
2981		 */
2982		PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2983
2984		/*
2985		 * Now we commit the changes to the software state
2986		 */
2987
2988		/*
2989		 * update overflow information
2990		 */
2991		if (is_counting) {
2992			/*
2993		 	 * full flag update each time a register is programmed
2994		 	 */
2995			ctx->ctx_pmds[cnum].flags = flags;
2996
2997			ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2998			ctx->ctx_pmds[cnum].smpl_pmds[0]  = smpl_pmds;
2999			ctx->ctx_pmds[cnum].eventid       = req->reg_smpl_eventid;
3000
3001			/*
3002			 * Mark all PMDS to be accessed as used.
3003			 *
3004			 * We do not keep track of PMC because we have to
3005			 * systematically restore ALL of them.
3006			 *
3007			 * We do not update the used_monitors mask, because
3008			 * if we have not programmed them, then will be in
3009			 * a quiescent state, therefore we will not need to
3010			 * mask/restore then when context is MASKED.
3011			 */
3012			CTX_USED_PMD(ctx, reset_pmds);
3013			CTX_USED_PMD(ctx, smpl_pmds);
3014			/*
3015		 	 * make sure we do not try to reset on
3016		 	 * restart because we have established new values
3017		 	 */
3018			if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3019		}
3020		/*
3021		 * Needed in case the user does not initialize the equivalent
3022		 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3023		 * possible leak here.
3024		 */
3025		CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
3026
3027		/*
3028		 * keep track of the monitor PMC that we are using.
3029		 * we save the value of the pmc in ctx_pmcs[] and if
3030		 * the monitoring is not stopped for the context we also
3031		 * place it in the saved state area so that it will be
3032		 * picked up later by the context switch code.
3033		 *
3034		 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3035		 *
3036		 * The value in th_pmcs[] may be modified on overflow, i.e.,  when
3037		 * monitoring needs to be stopped.
3038		 */
3039		if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3040
3041		/*
3042		 * update context state
3043		 */
3044		ctx->ctx_pmcs[cnum] = value;
3045
3046		if (is_loaded) {
3047			/*
3048			 * write thread state
3049			 */
3050			if (is_system == 0) ctx->th_pmcs[cnum] = value;
3051
3052			/*
3053			 * write hardware register if we can
3054			 */
3055			if (can_access_pmu) {
3056				ia64_set_pmc(cnum, value);
3057			}
3058#ifdef CONFIG_SMP
3059			else {
3060				/*
3061				 * per-task SMP only here
3062				 *
3063			 	 * we are guaranteed that the task is not running on the other CPU,
3064			 	 * we indicate that this PMD will need to be reloaded if the task
3065			 	 * is rescheduled on the CPU it ran last on.
3066			 	 */
3067				ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3068			}
3069#endif
3070		}
3071
3072		DPRINT(("pmc[%u]=0x%lx ld=%d apmu=%d flags=0x%x all_pmcs=0x%lx used_pmds=0x%lx eventid=%ld smpl_pmds=0x%lx reset_pmds=0x%lx reloads_pmcs=0x%lx used_monitors=0x%lx ovfl_regs=0x%lx\n",
3073			  cnum,
3074			  value,
3075			  is_loaded,
3076			  can_access_pmu,
3077			  flags,
3078			  ctx->ctx_all_pmcs[0],
3079			  ctx->ctx_used_pmds[0],
3080			  ctx->ctx_pmds[cnum].eventid,
3081			  smpl_pmds,
3082			  reset_pmds,
3083			  ctx->ctx_reload_pmcs[0],
3084			  ctx->ctx_used_monitors[0],
3085			  ctx->ctx_ovfl_regs[0]));
3086	}
3087
3088	/*
3089	 * make sure the changes are visible
3090	 */
3091	if (can_access_pmu) ia64_srlz_d();
3092
3093	return 0;
3094error:
3095	PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3096	return ret;
3097}
3098
3099static int
3100pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3101{
3102	struct task_struct *task;
3103	pfarg_reg_t *req = (pfarg_reg_t *)arg;
3104	unsigned long value, hw_value, ovfl_mask;
3105	unsigned int cnum;
3106	int i, can_access_pmu = 0, state;
3107	int is_counting, is_loaded, is_system, expert_mode;
3108	int ret = -EINVAL;
3109	pfm_reg_check_t wr_func;
3110
3111
3112	state     = ctx->ctx_state;
3113	is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3114	is_system = ctx->ctx_fl_system;
3115	ovfl_mask = pmu_conf->ovfl_val;
3116	task      = ctx->ctx_task;
3117
3118	if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3119
3120	/*
3121	 * on both UP and SMP, we can only write to the PMC when the task is
3122	 * the owner of the local PMU.
3123	 */
3124	if (likely(is_loaded)) {
3125		/*
3126		 * In system wide and when the context is loaded, access can only happen
3127		 * when the caller is running on the CPU being monitored by the session.
3128		 * It does not have to be the owner (ctx_task) of the context per se.
3129		 */
3130		if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3131			DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3132			return -EBUSY;
3133		}
3134		can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3135	}
3136	expert_mode = pfm_sysctl.expert_mode; 
3137
3138	for (i = 0; i < count; i++, req++) {
3139
3140		cnum  = req->reg_num;
3141		value = req->reg_value;
3142
3143		if (!PMD_IS_IMPL(cnum)) {
3144			DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3145			goto abort_mission;
3146		}
3147		is_counting = PMD_IS_COUNTING(cnum);
3148		wr_func     = pmu_conf->pmd_desc[cnum].write_check;
3149
3150		/*
3151		 * execute write checker, if any
3152		 */
3153		if (unlikely(expert_mode == 0 && wr_func)) {
3154			unsigned long v = value;
3155
3156			ret = (*wr_func)(task, ctx, cnum, &v, regs);
3157			if (ret) goto abort_mission;
3158
3159			value = v;
3160			ret   = -EINVAL;
3161		}
3162
3163		/*
3164		 * no error on this register
3165		 */
3166		PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3167
3168		/*
3169		 * now commit changes to software state
3170		 */
3171		hw_value = value;
3172
3173		/*
3174		 * update virtualized (64bits) counter
3175		 */
3176		if (is_counting) {
3177			/*
3178			 * write context state
3179			 */
3180			ctx->ctx_pmds[cnum].lval = value;
3181
3182			/*
3183			 * when context is load we use the split value
3184			 */
3185			if (is_loaded) {
3186				hw_value = value &  ovfl_mask;
3187				value    = value & ~ovfl_mask;
3188			}
3189		}
3190		/*
3191		 * update reset values (not just for counters)
3192		 */
3193		ctx->ctx_pmds[cnum].long_reset  = req->reg_long_reset;
3194		ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3195
3196		/*
3197		 * update randomization parameters (not just for counters)
3198		 */
3199		ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3200		ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3201
3202		/*
3203		 * update context value
3204		 */
3205		ctx->ctx_pmds[cnum].val  = value;
3206
3207		/*
3208		 * Keep track of what we use
3209		 *
3210		 * We do not keep track of PMC because we have to
3211		 * systematically restore ALL of them.
3212		 */
3213		CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3214
3215		/*
3216		 * mark this PMD register used as well
3217		 */
3218		CTX_USED_PMD(ctx, RDEP(cnum));
3219
3220		/*
3221		 * make sure we do not try to reset on
3222		 * restart because we have established new values
3223		 */
3224		if (is_counting && state == PFM_CTX_MASKED) {
3225			ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3226		}
3227
3228		if (is_loaded) {
3229			/*
3230		 	 * write thread state
3231		 	 */
3232			if (is_system == 0) ctx->th_pmds[cnum] = hw_value;
3233
3234			/*
3235			 * write hardware register if we can
3236			 */
3237			if (can_access_pmu) {
3238				ia64_set_pmd(cnum, hw_value);
3239			} else {
3240#ifdef CONFIG_SMP
3241				/*
3242			 	 * we are guaranteed that the task is not running on the other CPU,
3243			 	 * we indicate that this PMD will need to be reloaded if the task
3244			 	 * is rescheduled on the CPU it ran last on.
3245			 	 */
3246				ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3247#endif
3248			}
3249		}
3250
3251		DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx  short_reset=0x%lx "
3252			  "long_reset=0x%lx notify=%c seed=0x%lx mask=0x%lx used_pmds=0x%lx reset_pmds=0x%lx reload_pmds=0x%lx all_pmds=0x%lx ovfl_regs=0x%lx\n",
3253			cnum,
3254			value,
3255			is_loaded,
3256			can_access_pmu,
3257			hw_value,
3258			ctx->ctx_pmds[cnum].val,
3259			ctx->ctx_pmds[cnum].short_reset,
3260			ctx->ctx_pmds[cnum].long_reset,
3261			PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3262			ctx->ctx_pmds[cnum].seed,
3263			ctx->ctx_pmds[cnum].mask,
3264			ctx->ctx_used_pmds[0],
3265			ctx->ctx_pmds[cnum].reset_pmds[0],
3266			ctx->ctx_reload_pmds[0],
3267			ctx->ctx_all_pmds[0],
3268			ctx->ctx_ovfl_regs[0]));
3269	}
3270
3271	/*
3272	 * make changes visible
3273	 */
3274	if (can_access_pmu) ia64_srlz_d();
3275
3276	return 0;
3277
3278abort_mission:
3279	/*
3280	 * for now, we have only one possibility for error
3281	 */
3282	PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3283	return ret;
3284}
3285
3286/*
3287 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3288 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3289 * interrupt is delivered during the call, it will be kept pending until we leave, making
3290 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3291 * guaranteed to return consistent data to the user, it may simply be old. It is not
3292 * trivial to treat the overflow while inside the call because you may end up in
3293 * some module sampling buffer code causing deadlocks.
3294 */
3295static int
3296pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3297{
3298	struct task_struct *task;
3299	unsigned long val = 0UL, lval, ovfl_mask, sval;
3300	pfarg_reg_t *req = (pfarg_reg_t *)arg;
3301	unsigned int cnum, reg_flags = 0;
3302	int i, can_access_pmu = 0, state;
3303	int is_loaded, is_system, is_counting, expert_mode;
3304	int ret = -EINVAL;
3305	pfm_reg_check_t rd_func;
3306
3307	/*
3308	 * access is possible when loaded only for
3309	 * self-monitoring tasks or in UP mode
3310	 */
3311
3312	state     = ctx->ctx_state;
3313	is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3314	is_system = ctx->ctx_fl_system;
3315	ovfl_mask = pmu_conf->ovfl_val;
3316	task      = ctx->ctx_task;
3317
3318	if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3319
3320	if (likely(is_loaded)) {
3321		/*
3322		 * In system wide and when the context is loaded, access can only happen
3323		 * when the caller is running on the CPU being monitored by the session.
3324		 * It does not have to be the owner (ctx_task) of the context per se.
3325		 */
3326		if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3327			DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3328			return -EBUSY;
3329		}
3330		/*
3331		 * this can be true when not self-monitoring only in UP
3332		 */
3333		can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3334
3335		if (can_access_pmu) ia64_srlz_d();
3336	}
3337	expert_mode = pfm_sysctl.expert_mode; 
3338
3339	DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3340		is_loaded,
3341		can_access_pmu,
3342		state));
3343
3344	/*
3345	 * on both UP and SMP, we can only read the PMD from the hardware register when
3346	 * the task is the owner of the local PMU.
3347	 */
3348
3349	for (i = 0; i < count; i++, req++) {
3350
3351		cnum        = req->reg_num;
3352		reg_flags   = req->reg_flags;
3353
3354		if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3355		/*
3356		 * we can only read the register that we use. That includes
3357		 * the one we explicitly initialize AND the one we want included
3358		 * in the sampling buffer (smpl_regs).
3359		 *
3360		 * Having this restriction allows optimization in the ctxsw routine
3361		 * without compromising security (leaks)
3362		 */
3363		if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3364
3365		sval        = ctx->ctx_pmds[cnum].val;
3366		lval        = ctx->ctx_pmds[cnum].lval;
3367		is_counting = PMD_IS_COUNTING(cnum);
3368
3369		/*
3370		 * If the task is not the current one, then we check if the
3371		 * PMU state is still in the local live register due to lazy ctxsw.
3372		 * If true, then we read directly from the registers.
3373		 */
3374		if (can_access_pmu){
3375			val = ia64_get_pmd(cnum);
3376		} else {
3377			/*
3378			 * context has been saved
3379			 * if context is zombie, then task does not exist anymore.
3380			 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3381			 */
3382			val = is_loaded ? ctx->th_pmds[cnum] : 0UL;
3383		}
3384		rd_func = pmu_conf->pmd_desc[cnum].read_check;
3385
3386		if (is_counting) {
3387			/*
3388			 * XXX: need to check for overflow when loaded
3389			 */
3390			val &= ovfl_mask;
3391			val += sval;
3392		}
3393
3394		/*
3395		 * execute read checker, if any
3396		 */
3397		if (unlikely(expert_mode == 0 && rd_func)) {
3398			unsigned long v = val;
3399			ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3400			if (ret) goto error;
3401			val = v;
3402			ret = -EINVAL;
3403		}
3404
3405		PFM_REG_RETFLAG_SET(reg_flags, 0);
3406
3407		DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3408
3409		/*
3410		 * update register return value, abort all if problem during copy.
3411		 * we only modify the reg_flags field. no check mode is fine because
3412		 * access has been verified upfront in sys_perfmonctl().
3413		 */
3414		req->reg_value            = val;
3415		req->reg_flags            = reg_flags;
3416		req->reg_last_reset_val   = lval;
3417	}
3418
3419	return 0;
3420
3421error:
3422	PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3423	return ret;
3424}
3425
3426int
3427pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3428{
3429	pfm_context_t *ctx;
3430
3431	if (req == NULL) return -EINVAL;
3432
3433 	ctx = GET_PMU_CTX();
3434
3435	if (ctx == NULL) return -EINVAL;
3436
3437	/*
3438	 * for now limit to current task, which is enough when calling
3439	 * from overflow handler
3440	 */
3441	if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3442
3443	return pfm_write_pmcs(ctx, req, nreq, regs);
3444}
3445EXPORT_SYMBOL(pfm_mod_write_pmcs);
3446
3447int
3448pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3449{
3450	pfm_context_t *ctx;
3451
3452	if (req == NULL) return -EINVAL;
3453
3454 	ctx = GET_PMU_CTX();
3455
3456	if (ctx == NULL) return -EINVAL;
3457
3458	/*
3459	 * for now limit to current task, which is enough when calling
3460	 * from overflow handler
3461	 */
3462	if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3463
3464	return pfm_read_pmds(ctx, req, nreq, regs);
3465}
3466EXPORT_SYMBOL(pfm_mod_read_pmds);
3467
3468/*
3469 * Only call this function when a process it trying to
3470 * write the debug registers (reading is always allowed)
3471 */
3472int
3473pfm_use_debug_registers(struct task_struct *task)
3474{
3475	pfm_context_t *ctx = task->thread.pfm_context;
3476	unsigned long flags;
3477	int ret = 0;
3478
3479	if (pmu_conf->use_rr_dbregs == 0) return 0;
3480
3481	DPRINT(("called for [%d]\n", task_pid_nr(task)));
3482
3483	/*
3484	 * do it only once
3485	 */
3486	if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3487
3488	/*
3489	 * Even on SMP, we do not need to use an atomic here because
3490	 * the only way in is via ptrace() and this is possible only when the
3491	 * process is stopped. Even in the case where the ctxsw out is not totally
3492	 * completed by the time we come here, there is no way the 'stopped' process
3493	 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3494	 * So this is always safe.
3495	 */
3496	if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3497
3498	LOCK_PFS(flags);
3499
3500	/*
3501	 * We cannot allow setting breakpoints when system wide monitoring
3502	 * sessions are using the debug registers.
3503	 */
3504	if (pfm_sessions.pfs_sys_use_dbregs> 0)
3505		ret = -1;
3506	else
3507		pfm_sessions.pfs_ptrace_use_dbregs++;
3508
3509	DPRINT(("ptrace_use_dbregs=%u  sys_use_dbregs=%u by [%d] ret = %d\n",
3510		  pfm_sessions.pfs_ptrace_use_dbregs,
3511		  pfm_sessions.pfs_sys_use_dbregs,
3512		  task_pid_nr(task), ret));
3513
3514	UNLOCK_PFS(flags);
3515
3516	return ret;
3517}
3518
3519/*
3520 * This function is called for every task that exits with the
3521 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3522 * able to use the debug registers for debugging purposes via
3523 * ptrace(). Therefore we know it was not using them for
3524 * performance monitoring, so we only decrement the number
3525 * of "ptraced" debug register users to keep the count up to date
3526 */
3527int
3528pfm_release_debug_registers(struct task_struct *task)
3529{
3530	unsigned long flags;
3531	int ret;
3532
3533	if (pmu_conf->use_rr_dbregs == 0) return 0;
3534
3535	LOCK_PFS(flags);
3536	if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3537		printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task_pid_nr(task));
3538		ret = -1;
3539	}  else {
3540		pfm_sessions.pfs_ptrace_use_dbregs--;
3541		ret = 0;
3542	}
3543	UNLOCK_PFS(flags);
3544
3545	return ret;
3546}
3547
3548static int
3549pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3550{
3551	struct task_struct *task;
3552	pfm_buffer_fmt_t *fmt;
3553	pfm_ovfl_ctrl_t rst_ctrl;
3554	int state, is_system;
3555	int ret = 0;
3556
3557	state     = ctx->ctx_state;
3558	fmt       = ctx->ctx_buf_fmt;
3559	is_system = ctx->ctx_fl_system;
3560	task      = PFM_CTX_TASK(ctx);
3561
3562	switch(state) {
3563		case PFM_CTX_MASKED:
3564			break;
3565		case PFM_CTX_LOADED: 
3566			if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3567			/* fall through */
3568		case PFM_CTX_UNLOADED:
3569		case PFM_CTX_ZOMBIE:
3570			DPRINT(("invalid state=%d\n", state));
3571			return -EBUSY;
3572		default:
3573			DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3574			return -EINVAL;
3575	}
3576
3577	/*
3578 	 * In system wide and when the context is loaded, access can only happen
3579 	 * when the caller is running on the CPU being monitored by the session.
3580 	 * It does not have to be the owner (ctx_task) of the context per se.
3581 	 */
3582	if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3583		DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3584		return -EBUSY;
3585	}
3586
3587	/* sanity check */
3588	if (unlikely(task == NULL)) {
3589		printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", task_pid_nr(current));
3590		return -EINVAL;
3591	}
3592
3593	if (task == current || is_system) {
3594
3595		fmt = ctx->ctx_buf_fmt;
3596
3597		DPRINT(("restarting self %d ovfl=0x%lx\n",
3598			task_pid_nr(task),
3599			ctx->ctx_ovfl_regs[0]));
3600
3601		if (CTX_HAS_SMPL(ctx)) {
3602
3603			prefetch(ctx->ctx_smpl_hdr);
3604
3605			rst_ctrl.bits.mask_monitoring = 0;
3606			rst_ctrl.bits.reset_ovfl_pmds = 0;
3607
3608			if (state == PFM_CTX_LOADED)
3609				ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3610			else
3611				ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3612		} else {
3613			rst_ctrl.bits.mask_monitoring = 0;
3614			rst_ctrl.bits.reset_ovfl_pmds = 1;
3615		}
3616
3617		if (ret == 0) {
3618			if (rst_ctrl.bits.reset_ovfl_pmds)
3619				pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3620
3621			if (rst_ctrl.bits.mask_monitoring == 0) {
3622				DPRINT(("resuming monitoring for [%d]\n", task_pid_nr(task)));
3623
3624				if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3625			} else {
3626				DPRINT(("keeping monitoring stopped for [%d]\n", task_pid_nr(task)));
3627
3628				// cannot use pfm_stop_monitoring(task, regs);
3629			}
3630		}
3631		/*
3632		 * clear overflowed PMD mask to remove any stale information
3633		 */
3634		ctx->ctx_ovfl_regs[0] = 0UL;
3635
3636		/*
3637		 * back to LOADED state
3638		 */
3639		ctx->ctx_state = PFM_CTX_LOADED;
3640
3641		/*
3642		 * XXX: not really useful for self monitoring
3643		 */
3644		ctx->ctx_fl_can_restart = 0;
3645
3646		return 0;
3647	}
3648
3649	/* 
3650	 * restart another task
3651	 */
3652
3653	/*
3654	 * When PFM_CTX_MASKED, we cannot issue a restart before the previous 
3655	 * one is seen by the task.
3656	 */
3657	if (state == PFM_CTX_MASKED) {
3658		if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3659		/*
3660		 * will prevent subsequent restart before this one is
3661		 * seen by other task
3662		 */
3663		ctx->ctx_fl_can_restart = 0;
3664	}
3665
3666	/*
3667	 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3668	 * the task is blocked or on its way to block. That's the normal
3669	 * restart path. If the monitoring is not masked, then the task
3670	 * can be actively monitoring and we cannot directly intervene.
3671	 * Therefore we use the trap mechanism to catch the task and
3672	 * force it to reset the buffer/reset PMDs.
3673	 *
3674	 * if non-blocking, then we ensure that the task will go into
3675	 * pfm_handle_work() before returning to user mode.
3676	 *
3677	 * We cannot explicitly reset another task, it MUST always
3678	 * be done by the task itself. This works for system wide because
3679	 * the tool that is controlling the session is logically doing 
3680	 * "self-monitoring".
3681	 */
3682	if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3683		DPRINT(("unblocking [%d]\n", task_pid_nr(task)));
3684		complete(&ctx->ctx_restart_done);
3685	} else {
3686		DPRINT(("[%d] armed exit trap\n", task_pid_nr(task)));
3687
3688		ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3689
3690		PFM_SET_WORK_PENDING(task, 1);
3691
3692		set_notify_resume(task);
3693
3694		/*
3695		 * XXX: send reschedule if task runs on another CPU
3696		 */
3697	}
3698	return 0;
3699}
3700
3701static int
3702pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3703{
3704	unsigned int m = *(unsigned int *)arg;
3705
3706	pfm_sysctl.debug = m == 0 ? 0 : 1;
3707
3708	printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3709
3710	if (m == 0) {
3711		memset(pfm_stats, 0, sizeof(pfm_stats));
3712		for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3713	}
3714	return 0;
3715}
3716
3717/*
3718 * arg can be NULL and count can be zero for this function
3719 */
3720static int
3721pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3722{
3723	struct thread_struct *thread = NULL;
3724	struct task_struct *task;
3725	pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3726	unsigned long flags;
3727	dbreg_t dbreg;
3728	unsigned int rnum;
3729	int first_time;
3730	int ret = 0, state;
3731	int i, can_access_pmu = 0;
3732	int is_system, is_loaded;
3733
3734	if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3735
3736	state     = ctx->ctx_state;
3737	is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3738	is_system = ctx->ctx_fl_system;
3739	task      = ctx->ctx_task;
3740
3741	if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3742
3743	/*
3744	 * on both UP and SMP, we can only write to the PMC when the task is
3745	 * the owner of the local PMU.
3746	 */
3747	if (is_loaded) {
3748		thread = &task->thread;
3749		/*
3750		 * In system wide and when the context is loaded, access can only happen
3751		 * when the caller is running on the CPU being monitored by the session.
3752		 * It does not have to be the owner (ctx_task) of the context per se.
3753		 */
3754		if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3755			DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3756			return -EBUSY;
3757		}
3758		can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3759	}
3760
3761	/*
3762	 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3763	 * ensuring that no real breakpoint can be installed via this call.
3764	 *
3765	 * IMPORTANT: regs can be NULL in this function
3766	 */
3767
3768	first_time = ctx->ctx_fl_using_dbreg == 0;
3769
3770	/*
3771	 * don't bother if we are loaded and task is being debugged
3772	 */
3773	if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3774		DPRINT(("debug registers already in use for [%d]\n", task_pid_nr(task)));
3775		return -EBUSY;
3776	}
3777
3778	/*
3779	 * check for debug registers in system wide mode
3780	 *
3781	 * If though a check is done in pfm_context_load(),
3782	 * we must repeat it here, in case the registers are
3783	 * written after the context is loaded
3784	 */
3785	if (is_loaded) {
3786		LOCK_PFS(flags);
3787
3788		if (first_time && is_system) {
3789			if (pfm_sessions.pfs_ptrace_use_dbregs)
3790				ret = -EBUSY;
3791			else
3792				pfm_sessions.pfs_sys_use_dbregs++;
3793		}
3794		UNLOCK_PFS(flags);
3795	}
3796
3797	if (ret != 0) return ret;
3798
3799	/*
3800	 * mark ourself as user of the debug registers for
3801	 * perfmon purposes.
3802	 */
3803	ctx->ctx_fl_using_dbreg = 1;
3804
3805	/*
3806 	 * clear hardware registers to make sure we don't
3807 	 * pick up stale state.
3808	 *
3809	 * for a system wide session, we do not use
3810	 * thread.dbr, thread.ibr because this process
3811	 * never leaves the current CPU and the state
3812	 * is shared by all processes running on it
3813 	 */
3814	if (first_time && can_access_pmu) {
3815		DPRINT(("[%d] clearing ibrs, dbrs\n", task_pid_nr(task)));
3816		for (i=0; i < pmu_conf->num_ibrs; i++) {
3817			ia64_set_ibr(i, 0UL);
3818			ia64_dv_serialize_instruction();
3819		}
3820		ia64_srlz_i();
3821		for (i=0; i < pmu_conf->num_dbrs; i++) {
3822			ia64_set_dbr(i, 0UL);
3823			ia64_dv_serialize_data();
3824		}
3825		ia64_srlz_d();
3826	}
3827
3828	/*
3829	 * Now install the values into the registers
3830	 */
3831	for (i = 0; i < count; i++, req++) {
3832
3833		rnum      = req->dbreg_num;
3834		dbreg.val = req->dbreg_value;
3835
3836		ret = -EINVAL;
3837
3838		if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3839			DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3840				  rnum, dbreg.val, mode, i, count));
3841
3842			goto abort_mission;
3843		}
3844
3845		/*
3846		 * make sure we do not install enabled breakpoint
3847		 */
3848		if (rnum & 0x1) {
3849			if (mode == PFM_CODE_RR)
3850				dbreg.ibr.ibr_x = 0;
3851			else
3852				dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3853		}
3854
3855		PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3856
3857		/*
3858		 * Debug registers, just like PMC, can only be modified
3859		 * by a kernel call. Moreover, perfmon() access to those
3860		 * registers are centralized in this routine. The hardware
3861		 * does not modify the value of these registers, therefore,
3862		 * if we save them as they are written, we can avoid having
3863		 * to save them on context switch out. This is made possible
3864		 * by the fact that when perfmon uses debug registers, ptrace()
3865		 * won't be able to modify them concurrently.
3866		 */
3867		if (mode == PFM_CODE_RR) {
3868			CTX_USED_IBR(ctx, rnum);
3869
3870			if (can_access_pmu) {
3871				ia64_set_ibr(rnum, dbreg.val);
3872				ia64_dv_serialize_instruction();
3873			}
3874
3875			ctx->ctx_ibrs[rnum] = dbreg.val;
3876
3877			DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3878				rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3879		} else {
3880			CTX_USED_DBR(ctx, rnum);
3881
3882			if (can_access_pmu) {
3883				ia64_set_dbr(rnum, dbreg.val);
3884				ia64_dv_serialize_data();
3885			}
3886			ctx->ctx_dbrs[rnum] = dbreg.val;
3887
3888			DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3889				rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3890		}
3891	}
3892
3893	return 0;
3894
3895abort_mission:
3896	/*
3897	 * in case it was our first attempt, we undo the global modifications
3898	 */
3899	if (first_time) {
3900		LOCK_PFS(flags);
3901		if (ctx->ctx_fl_system) {
3902			pfm_sessions.pfs_sys_use_dbregs--;
3903		}
3904		UNLOCK_PFS(flags);
3905		ctx->ctx_fl_using_dbreg = 0;
3906	}
3907	/*
3908	 * install error return flag
3909	 */
3910	PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3911
3912	return ret;
3913}
3914
3915static int
3916pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3917{
3918	return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3919}
3920
3921static int
3922pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3923{
3924	return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3925}
3926
3927int
3928pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3929{
3930	pfm_context_t *ctx;
3931
3932	if (req == NULL) return -EINVAL;
3933
3934 	ctx = GET_PMU_CTX();
3935
3936	if (ctx == NULL) return -EINVAL;
3937
3938	/*
3939	 * for now limit to current task, which is enough when calling
3940	 * from overflow handler
3941	 */
3942	if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3943
3944	return pfm_write_ibrs(ctx, req, nreq, regs);
3945}
3946EXPORT_SYMBOL(pfm_mod_write_ibrs);
3947
3948int
3949pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3950{
3951	pfm_context_t *ctx;
3952
3953	if (req == NULL) return -EINVAL;
3954
3955 	ctx = GET_PMU_CTX();
3956
3957	if (ctx == NULL) return -EINVAL;
3958
3959	/*
3960	 * for now limit to current task, which is enough when calling
3961	 * from overflow handler
3962	 */
3963	if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3964
3965	return pfm_write_dbrs(ctx, req, nreq, regs);
3966}
3967EXPORT_SYMBOL(pfm_mod_write_dbrs);
3968
3969
3970static int
3971pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3972{
3973	pfarg_features_t *req = (pfarg_features_t *)arg;
3974
3975	req->ft_version = PFM_VERSION;
3976	return 0;
3977}
3978
3979static int
3980pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3981{
3982	struct pt_regs *tregs;
3983	struct task_struct *task = PFM_CTX_TASK(ctx);
3984	int state, is_system;
3985
3986	state     = ctx->ctx_state;
3987	is_system = ctx->ctx_fl_system;
3988
3989	/*
3990	 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3991	 */
3992	if (state == PFM_CTX_UNLOADED) return -EINVAL;
3993
3994	/*
3995 	 * In system wide and when the context is loaded, access can only happen
3996 	 * when the caller is running on the CPU being monitored by the session.
3997 	 * It does not have to be the owner (ctx_task) of the context per se.
3998 	 */
3999	if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4000		DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4001		return -EBUSY;
4002	}
4003	DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
4004		task_pid_nr(PFM_CTX_TASK(ctx)),
4005		state,
4006		is_system));
4007	/*
4008	 * in system mode, we need to update the PMU directly
4009	 * and the user level state of the caller, which may not
4010	 * necessarily be the creator of the context.
4011	 */
4012	if (is_system) {
4013		/*
4014		 * Update local PMU first
4015		 *
4016		 * disable dcr pp
4017		 */
4018		ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
4019		ia64_srlz_i();
4020
4021		/*
4022		 * update local cpuinfo
4023		 */
4024		PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4025
4026		/*
4027		 * stop monitoring, does srlz.i
4028		 */
4029		pfm_clear_psr_pp();
4030
4031		/*
4032		 * stop monitoring in the caller
4033		 */
4034		ia64_psr(regs)->pp = 0;
4035
4036		return 0;
4037	}
4038	/*
4039	 * per-task mode
4040	 */
4041
4042	if (task == current) {
4043		/* stop monitoring  at kernel level */
4044		pfm_clear_psr_up();
4045
4046		/*
4047	 	 * stop monitoring at the user level
4048	 	 */
4049		ia64_psr(regs)->up = 0;
4050	} else {
4051		tregs = task_pt_regs(task);
4052
4053		/*
4054	 	 * stop monitoring at the user level
4055	 	 */
4056		ia64_psr(tregs)->up = 0;
4057
4058		/*
4059		 * monitoring disabled in kernel at next reschedule
4060		 */
4061		ctx->ctx_saved_psr_up = 0;
4062		DPRINT(("task=[%d]\n", task_pid_nr(task)));
4063	}
4064	return 0;
4065}
4066
4067
4068static int
4069pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4070{
4071	struct pt_regs *tregs;
4072	int state, is_system;
4073
4074	state     = ctx->ctx_state;
4075	is_system = ctx->ctx_fl_system;
4076
4077	if (state != PFM_CTX_LOADED) return -EINVAL;
4078
4079	/*
4080 	 * In system wide and when the context is loaded, access can only happen
4081 	 * when the caller is running on the CPU being monitored by the session.
4082 	 * It does not have to be the owner (ctx_task) of the context per se.
4083 	 */
4084	if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4085		DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4086		return -EBUSY;
4087	}
4088
4089	/*
4090	 * in system mode, we need to update the PMU directly
4091	 * and the user level state of the caller, which may not
4092	 * necessarily be the creator of the context.
4093	 */
4094	if (is_system) {
4095
4096		/*
4097		 * set user level psr.pp for the caller
4098		 */
4099		ia64_psr(regs)->pp = 1;
4100
4101		/*
4102		 * now update the local PMU and cpuinfo
4103		 */
4104		PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4105
4106		/*
4107		 * start monitoring at kernel level
4108		 */
4109		pfm_set_psr_pp();
4110
4111		/* enable dcr pp */
4112		ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4113		ia64_srlz_i();
4114
4115		return 0;
4116	}
4117
4118	/*
4119	 * per-process mode
4120	 */
4121
4122	if (ctx->ctx_task == current) {
4123
4124		/* start monitoring at kernel level */
4125		pfm_set_psr_up();
4126
4127		/*
4128		 * activate monitoring at user level
4129		 */
4130		ia64_psr(regs)->up = 1;
4131
4132	} else {
4133		tregs = task_pt_regs(ctx->ctx_task);
4134
4135		/*
4136		 * start monitoring at the kernel level the next
4137		 * time the task is scheduled
4138		 */
4139		ctx->ctx_saved_psr_up = IA64_PSR_UP;
4140
4141		/*
4142		 * activate monitoring at user level
4143		 */
4144		ia64_psr(tregs)->up = 1;
4145	}
4146	return 0;
4147}
4148
4149static int
4150pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4151{
4152	pfarg_reg_t *req = (pfarg_reg_t *)arg;
4153	unsigned int cnum;
4154	int i;
4155	int ret = -EINVAL;
4156
4157	for (i = 0; i < count; i++, req++) {
4158
4159		cnum = req->reg_num;
4160
4161		if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4162
4163		req->reg_value = PMC_DFL_VAL(cnum);
4164
4165		PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4166
4167		DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4168	}
4169	return 0;
4170
4171abort_mission:
4172	PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4173	return ret;
4174}
4175
4176static int
4177pfm_check_task_exist(pfm_context_t *ctx)
4178{
4179	struct task_struct *g, *t;
4180	int ret = -ESRCH;
4181
4182	read_lock(&tasklist_lock);
4183
4184	do_each_thread (g, t) {
4185		if (t->thread.pfm_context == ctx) {
4186			ret = 0;
4187			goto out;
4188		}
4189	} while_each_thread (g, t);
4190out:
4191	read_unlock(&tasklist_lock);
4192
4193	DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4194
4195	return ret;
4196}
4197
4198static int
4199pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4200{
4201	struct task_struct *task;
4202	struct thread_struct *thread;
4203	struct pfm_context_t *old;
4204	unsigned long flags;
4205#ifndef CONFIG_SMP
4206	struct task_struct *owner_task = NULL;
4207#endif
4208	pfarg_load_t *req = (pfarg_load_t *)arg;
4209	unsigned long *pmcs_source, *pmds_source;
4210	int the_cpu;
4211	int ret = 0;
4212	int state, is_system, set_dbregs = 0;
4213
4214	state     = ctx->ctx_state;
4215	is_system = ctx->ctx_fl_system;
4216	/*
4217	 * can only load from unloaded or terminated state
4218	 */
4219	if (state != PFM_CTX_UNLOADED) {
4220		DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4221			req->load_pid,
4222			ctx->ctx_state));
4223		return -EBUSY;
4224	}
4225
4226	DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4227
4228	if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4229		DPRINT(("cannot use blocking mode on self\n"));
4230		return -EINVAL;
4231	}
4232
4233	ret = pfm_get_task(ctx, req->load_pid, &task);
4234	if (ret) {
4235		DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4236		return ret;
4237	}
4238
4239	ret = -EINVAL;
4240
4241	/*
4242	 * system wide is self monitoring only
4243	 */
4244	if (is_system && task != current) {
4245		DPRINT(("system wide is self monitoring only load_pid=%d\n",
4246			req->load_pid));
4247		goto error;
4248	}
4249
4250	thread = &task->thread;
4251
4252	ret = 0;
4253	/*
4254	 * cannot load a context which is using range restrictions,
4255	 * into a task that is being debugged.
4256	 */
4257	if (ctx->ctx_fl_using_dbreg) {
4258		if (thread->flags & IA64_THREAD_DBG_VALID) {
4259			ret = -EBUSY;
4260			DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4261			goto error;
4262		}
4263		LOCK_PFS(flags);
4264
4265		if (is_system) {
4266			if (pfm_sessions.pfs_ptrace_use_dbregs) {
4267				DPRINT(("cannot load [%d] dbregs in use\n",
4268							task_pid_nr(task)));
4269				ret = -EBUSY;
4270			} else {
4271				pfm_sessions.pfs_sys_use_dbregs++;
4272				DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task_pid_nr(task), pfm_sessions.pfs_sys_use_dbregs));
4273				set_dbregs = 1;
4274			}
4275		}
4276
4277		UNLOCK_PFS(flags);
4278
4279		if (ret) goto error;
4280	}
4281
4282	/*
4283	 * SMP system-wide monitoring implies self-monitoring.
4284	 *
4285	 * The programming model expects the task to
4286	 * be pinned on a CPU throughout the session.
4287	 * Here we take note of the current CPU at the
4288	 * time the context is loaded. No call from
4289	 * another CPU will be allowed.
4290	 *
4291	 * The pinning via shed_setaffinity()
4292	 * must be done by the calling task prior
4293	 * to this call.
4294	 *
4295	 * systemwide: keep track of CPU this session is supposed to run on
4296	 */
4297	the_cpu = ctx->ctx_cpu = smp_processor_id();
4298
4299	ret = -EBUSY;
4300	/*
4301	 * now reserve the session
4302	 */
4303	ret = pfm_reserve_session(current, is_system, the_cpu);
4304	if (ret) goto error;
4305
4306	/*
4307	 * task is necessarily stopped at this point.
4308	 *
4309	 * If the previous context was zombie, then it got removed in
4310	 * pfm_save_regs(). Therefore we should not see it here.
4311	 * If we see a context, then this is an active context
4312	 *
4313	 * XXX: needs to be atomic
4314	 */
4315	DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4316		thread->pfm_context, ctx));
4317
4318	ret = -EBUSY;
4319	old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4320	if (old != NULL) {
4321		DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4322		goto error_unres;
4323	}
4324
4325	pfm_reset_msgq(ctx);
4326
4327	ctx->ctx_state = PFM_CTX_LOADED;
4328
4329	/*
4330	 * link context to task
4331	 */
4332	ctx->ctx_task = task;
4333
4334	if (is_system) {
4335		/*
4336		 * we load as stopped
4337		 */
4338		PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4339		PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4340
4341		if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4342	} else {
4343		thread->flags |= IA64_THREAD_PM_VALID;
4344	}
4345
4346	/*
4347	 * propagate into thread-state
4348	 */
4349	pfm_copy_pmds(task, ctx);
4350	pfm_copy_pmcs(task, ctx);
4351
4352	pmcs_source = ctx->th_pmcs;
4353	pmds_source = ctx->th_pmds;
4354
4355	/*
4356	 * always the case for system-wide
4357	 */
4358	if (task == current) {
4359
4360		if (is_system == 0) {
4361
4362			/* allow user level control */
4363			ia64_psr(regs)->sp = 0;
4364			DPRINT(("clearing psr.sp for [%d]\n", task_pid_nr(task)));
4365
4366			SET_LAST_CPU(ctx, smp_processor_id());
4367			INC_ACTIVATION();
4368			SET_ACTIVATION(ctx);
4369#ifndef CONFIG_SMP
4370			/*
4371			 * push the other task out, if any
4372			 */
4373			owner_task = GET_PMU_OWNER();
4374			if (owner_task) pfm_lazy_save_regs(owner_task);
4375#endif
4376		}
4377		/*
4378		 * load all PMD from ctx to PMU (as opposed to thread state)
4379		 * restore all PMC from ctx to PMU
4380		 */
4381		pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4382		pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4383
4384		ctx->ctx_reload_pmcs[0] = 0UL;
4385		ctx->ctx_reload_pmds[0] = 0UL;
4386
4387		/*
4388		 * guaranteed safe by earlier check against DBG_VALID
4389		 */
4390		if (ctx->ctx_fl_using_dbreg) {
4391			pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4392			pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4393		}
4394		/*
4395		 * set new ownership
4396		 */
4397		SET_PMU_OWNER(task, ctx);
4398
4399		DPRINT(("context loaded on PMU for [%d]\n", task_pid_nr(task)));
4400	} else {
4401		/*
4402		 * when not current, task MUST be stopped, so this is safe
4403		 */
4404		regs = task_pt_regs(task);
4405
4406		/* force a full reload */
4407		ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4408		SET_LAST_CPU(ctx, -1);
4409
4410		/* initial saved psr (stopped) */
4411		ctx->ctx_saved_psr_up = 0UL;
4412		ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4413	}
4414
4415	ret = 0;
4416
4417error_unres:
4418	if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4419error:
4420	/*
4421	 * we must undo the dbregs setting (for system-wide)
4422	 */
4423	if (ret && set_dbregs) {
4424		LOCK_PFS(flags);
4425		pfm_sessions.pfs_sys_use_dbregs--;
4426		UNLOCK_PFS(flags);
4427	}
4428	/*
4429	 * release task, there is now a link with the context
4430	 */
4431	if (is_system == 0 && task != current) {
4432		pfm_put_task(task);
4433
4434		if (ret == 0) {
4435			ret = pfm_check_task_exist(ctx);
4436			if (ret) {
4437				ctx->ctx_state = PFM_CTX_UNLOADED;
4438				ctx->ctx_task  = NULL;
4439			}
4440		}
4441	}
4442	return ret;
4443}
4444
4445/*
4446 * in this function, we do not need to increase the use count
4447 * for the task via get_task_struct(), because we hold the
4448 * context lock. If the task were to disappear while having
4449 * a context attached, it would go through pfm_exit_thread()
4450 * which also grabs the context lock  and would therefore be blocked
4451 * until we are here.
4452 */
4453static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4454
4455static int
4456pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4457{
4458	struct task_struct *task = PFM_CTX_TASK(ctx);
4459	struct pt_regs *tregs;
4460	int prev_state, is_system;
4461	int ret;
4462
4463	DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task_pid_nr(task) : -1));
4464
4465	prev_state = ctx->ctx_state;
4466	is_system  = ctx->ctx_fl_system;
4467
4468	/*
4469	 * unload only when necessary
4470	 */
4471	if (prev_state == PFM_CTX_UNLOADED) {
4472		DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4473		return 0;
4474	}
4475
4476	/*
4477	 * clear psr and dcr bits
4478	 */
4479	ret = pfm_stop(ctx, NULL, 0, regs);
4480	if (ret) return ret;
4481
4482	ctx->ctx_state = PFM_CTX_UNLOADED;
4483
4484	/*
4485	 * in system mode, we need to update the PMU directly
4486	 * and the user level state of the caller, which may not
4487	 * necessarily be the creator of the context.
4488	 */
4489	if (is_system) {
4490
4491		/*
4492		 * Update cpuinfo
4493		 *
4494		 * local PMU is taken care of in pfm_stop()
4495		 */
4496		PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4497		PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4498
4499		/*
4500		 * save PMDs in context
4501		 * release ownership
4502		 */
4503		pfm_flush_pmds(current, ctx);
4504
4505		/*
4506		 * at this point we are done with the PMU
4507		 * so we can unreserve the resource.
4508		 */
4509		if (prev_state != PFM_CTX_ZOMBIE) 
4510			pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4511
4512		/*
4513		 * disconnect context from task
4514		 */
4515		task->thread.pfm_context = NULL;
4516		/*
4517		 * disconnect task from context
4518		 */
4519		ctx->ctx_task = NULL;
4520
4521		/*
4522		 * There is nothing more to cleanup here.
4523		 */
4524		return 0;
4525	}
4526
4527	/*
4528	 * per-task mode
4529	 */
4530	tregs = task == current ? regs : task_pt_regs(task);
4531
4532	if (task == current) {
4533		/*
4534		 * cancel user level control
4535		 */
4536		ia64_psr(regs)->sp = 1;
4537
4538		DPRINT(("setting psr.sp for [%d]\n", task_pid_nr(task)));
4539	}
4540	/*
4541	 * save PMDs to context
4542	 * release ownership
4543	 */
4544	pfm_flush_pmds(task, ctx);
4545
4546	/*
4547	 * at this point we are done with the PMU
4548	 * so we can unreserve the resource.
4549	 *
4550	 * when state was ZOMBIE, we have already unreserved.
4551	 */
4552	if (prev_state != PFM_CTX_ZOMBIE) 
4553		pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4554
4555	/*
4556	 * reset activation counter and psr
4557	 */
4558	ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4559	SET_LAST_CPU(ctx, -1);
4560
4561	/*
4562	 * PMU state will not be restored
4563	 */
4564	task->thread.flags &= ~IA64_THREAD_PM_VALID;
4565
4566	/*
4567	 * break links between context and task
4568	 */
4569	task->thread.pfm_context  = NULL;
4570	ctx->ctx_task             = NULL;
4571
4572	PFM_SET_WORK_PENDING(task, 0);
4573
4574	ctx->ctx_fl_trap_reason  = PFM_TRAP_REASON_NONE;
4575	ctx->ctx_fl_can_restart  = 0;
4576	ctx->ctx_fl_going_zombie = 0;
4577
4578	DPRINT(("disconnected [%d] from context\n", task_pid_nr(task)));
4579
4580	return 0;
4581}
4582
4583
4584/*
4585 * called only from exit_thread(): task == current
4586 * we come here only if current has a context attached (loaded or masked)
4587 */
4588void
4589pfm_exit_thread(struct task_struct *task)
4590{
4591	pfm_context_t *ctx;
4592	unsigned long flags;
4593	struct pt_regs *regs = task_pt_regs(task);
4594	int ret, state;
4595	int free_ok = 0;
4596
4597	ctx = PFM_GET_CTX(task);
4598
4599	PROTECT_CTX(ctx, flags);
4600
4601	DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task_pid_nr(task)));
4602
4603	state = ctx->ctx_state;
4604	switch(state) {
4605		case PFM_CTX_UNLOADED:
4606			/*
4607	 		 * only comes to this function if pfm_context is not NULL, i.e., cannot
4608			 * be in unloaded state
4609	 		 */
4610			printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task_pid_nr(task));
4611			break;
4612		case PFM_CTX_LOADED:
4613		case PFM_CTX_MASKED:
4614			ret = pfm_context_unload(ctx, NULL, 0, regs);
4615			if (ret) {
4616				printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4617			}
4618			DPRINT(("ctx unloaded for current state was %d\n", state));
4619
4620			pfm_end_notify_user(ctx);
4621			break;
4622		case PFM_CTX_ZOMBIE:
4623			ret = pfm_context_unload(ctx, NULL, 0, regs);
4624			if (ret) {
4625				printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4626			}
4627			free_ok = 1;
4628			break;
4629		default:
4630			printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task_pid_nr(task), state);
4631			break;
4632	}
4633	UNPROTECT_CTX(ctx, flags);
4634
4635	{ u64 psr = pfm_get_psr();
4636	  BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4637	  BUG_ON(GET_PMU_OWNER());
4638	  BUG_ON(ia64_psr(regs)->up);
4639	  BUG_ON(ia64_psr(regs)->pp);
4640	}
4641
4642	/*
4643	 * All memory free operations (especially for vmalloc'ed memory)
4644	 * MUST be done with interrupts ENABLED.
4645	 */
4646	if (free_ok) pfm_context_free(ctx);
4647}
4648
4649/*
4650 * functions MUST be listed in the increasing order of their index (see permfon.h)
4651 */
4652#define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4653#define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4654#define PFM_CMD_PCLRWS	(PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4655#define PFM_CMD_PCLRW	(PFM_CMD_FD|PFM_CMD_ARG_RW)
4656#define PFM_CMD_NONE	{ NULL, "no-cmd", 0, 0, 0, NULL}
4657
4658static pfm_cmd_desc_t pfm_cmd_tab[]={
4659/* 0  */PFM_CMD_NONE,
4660/* 1  */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4661/* 2  */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4662/* 3  */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4663/* 4  */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4664/* 5  */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4665/* 6  */PFM_CMD_NONE,
4666/* 7  */PFM_CMD_NONE,
4667/* 8  */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4668/* 9  */PFM_CMD_NONE,
4669/* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4670/* 11 */PFM_CMD_NONE,
4671/* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4672/* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4673/* 14 */PFM_CMD_NONE,
4674/* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4675/* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4676/* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4677/* 18 */PFM_CMD_NONE,
4678/* 19 */PFM_CMD_NONE,
4679/* 20 */PFM_CMD_NONE,
4680/* 21 */PFM_CMD_NONE,
4681/* 22 */PFM_CMD_NONE,
4682/* 23 */PFM_CMD_NONE,
4683/* 24 */PFM_CMD_NONE,
4684/* 25 */PFM_CMD_NONE,
4685/* 26 */PFM_CMD_NONE,
4686/* 27 */PFM_CMD_NONE,
4687/* 28 */PFM_CMD_NONE,
4688/* 29 */PFM_CMD_NONE,
4689/* 30 */PFM_CMD_NONE,
4690/* 31 */PFM_CMD_NONE,
4691/* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4692/* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4693};
4694#define PFM_CMD_COUNT	(sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4695
4696static int
4697pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4698{
4699	struct task_struct *task;
4700	int state, old_state;
4701
4702recheck:
4703	state = ctx->ctx_state;
4704	task  = ctx->ctx_task;
4705
4706	if (task == NULL) {
4707		DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4708		return 0;
4709	}
4710
4711	DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4712		ctx->ctx_fd,
4713		state,
4714		task_pid_nr(task),
4715		task->state, PFM_CMD_STOPPED(cmd)));
4716
4717	/*
4718	 * self-monitoring always ok.
4719	 *
4720	 * for system-wide the caller can either be the creator of the
4721	 * context (to one to which the context is attached to) OR
4722	 * a task running on the same CPU as the session.
4723	 */
4724	if (task == current || ctx->ctx_fl_system) return 0;
4725
4726	/*
4727	 * we are monitoring another thread
4728	 */
4729	switch(state) {
4730		case PFM_CTX_UNLOADED:
4731			/*
4732			 * if context is UNLOADED we are safe to go
4733			 */
4734			return 0;
4735		case PFM_CTX_ZOMBIE:
4736			/*
4737			 * no command can operate on a zombie context
4738			 */
4739			DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4740			return -EINVAL;
4741		case PFM_CTX_MASKED:
4742			/*
4743			 * PMU state has been saved to software even though
4744			 * the thread may still be running.
4745			 */
4746			if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4747	}
4748
4749	/*
4750	 * context is LOADED or MASKED. Some commands may need to have 
4751	 * the task stopped.
4752	 *
4753	 * We could lift this restriction for UP but it would mean that
4754	 * the user has no guarantee the task would not run between
4755	 * two successive calls to perfmonctl(). That's probably OK.
4756	 * If this user wants to ensure the task does not run, then
4757	 * the task must be stopped.
4758	 */
4759	if (PFM_CMD_STOPPED(cmd)) {
4760		if (!task_is_stopped_or_traced(task)) {
4761			DPRINT(("[%d] task not in stopped state\n", task_pid_nr(task)));
4762			return -EBUSY;
4763		}
4764		/*
4765		 * task is now stopped, wait for ctxsw out
4766		 *
4767		 * This is an interesting point in the code.
4768		 * We need to unprotect the context because
4769		 * the pfm_save_regs() routines needs to grab
4770		 * the same lock. There are danger in doing
4771		 * this because it leaves a window open for
4772		 * another task to get access to the context
4773		 * and possibly change its state. The one thing
4774		 * that is not possible is for the context to disappear
4775		 * because we are protected by the VFS layer, i.e.,
4776		 * get_fd()/put_fd().
4777		 */
4778		old_state = state;
4779
4780		UNPROTECT_CTX(ctx, flags);
4781
4782		wait_task_inactive(task, 0);
4783
4784		PROTECT_CTX(ctx, flags);
4785
4786		/*
4787		 * we must recheck to verify if state has changed
4788		 */
4789		if (ctx->ctx_state != old_state) {
4790			DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4791			goto recheck;
4792		}
4793	}
4794	return 0;
4795}
4796
4797/*
4798 * system-call entry point (must return long)
4799 */
4800asmlinkage long
4801sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4802{
4803	struct file *file = NULL;
4804	pfm_context_t *ctx = NULL;
4805	unsigned long flags = 0UL;
4806	void *args_k = NULL;
4807	long ret; /* will expand int return types */
4808	size_t base_sz, sz, xtra_sz = 0;
4809	int narg, completed_args = 0, call_made = 0, cmd_flags;
4810	int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4811	int (*getsize)(void *arg, size_t *sz);
4812#define PFM_MAX_ARGSIZE	4096
4813
4814	/*
4815	 * reject any call if perfmon was disabled at initialization
4816	 */
4817	if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4818
4819	if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4820		DPRINT(("invalid cmd=%d\n", cmd));
4821		return -EINVAL;
4822	}
4823
4824	func      = pfm_cmd_tab[cmd].cmd_func;
4825	narg      = pfm_cmd_tab[cmd].cmd_narg;
4826	base_sz   = pfm_cmd_tab[cmd].cmd_argsize;
4827	getsize   = pfm_cmd_tab[cmd].cmd_getsize;
4828	cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4829
4830	if (unlikely(func == NULL)) {
4831		DPRINT(("invalid cmd=%d\n", cmd));
4832		return -EINVAL;
4833	}
4834
4835	DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4836		PFM_CMD_NAME(cmd),
4837		cmd,
4838		narg,
4839		base_sz,
4840		count));
4841
4842	/*
4843	 * check if number of arguments matches what the command expects
4844	 */
4845	if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4846		return -EINVAL;
4847
4848restart_args:
4849	sz = xtra_sz + base_sz*count;
4850	/*
4851	 * limit abuse to min page size
4852	 */
4853	if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4854		printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", task_pid_nr(current), sz);
4855		return -E2BIG;
4856	}
4857
4858	/*
4859	 * allocate default-sized argument buffer
4860	 */
4861	if (likely(count && args_k == NULL)) {
4862		args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4863		if (args_k == NULL) return -ENOMEM;
4864	}
4865
4866	ret = -EFAULT;
4867
4868	/*
4869	 * copy arguments
4870	 *
4871	 * assume sz = 0 for command without parameters
4872	 */
4873	if (sz && copy_from_user(args_k, arg, sz)) {
4874		DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4875		goto error_args;
4876	}
4877
4878	/*
4879	 * check if command supports extra parameters
4880	 */
4881	if (completed_args == 0 && getsize) {
4882		/*
4883		 * get extra parameters size (based on main argument)
4884		 */
4885		ret = (*getsize)(args_k, &xtra_sz);
4886		if (ret) goto error_args;
4887
4888		completed_args = 1;
4889
4890		DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4891
4892		/* retry if necessary */
4893		if (likely(xtra_sz)) goto restart_args;
4894	}
4895
4896	if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4897
4898	ret = -EBADF;
4899
4900	file = fget(fd);
4901	if (unlikely(file == NULL)) {
4902		DPRINT(("invalid fd %d\n", fd));
4903		goto error_args;
4904	}
4905	if (unlikely(PFM_IS_FILE(file) == 0)) {
4906		DPRINT(("fd %d not related to perfmon\n", fd));
4907		goto error_args;
4908	}
4909
4910	ctx = file->private_data;
4911	if (unlikely(ctx == NULL)) {
4912		DPRINT(("no context for fd %d\n", fd));
4913		goto error_args;
4914	}
4915	prefetch(&ctx->ctx_state);
4916
4917	PROTECT_CTX(ctx, flags);
4918
4919	/*
4920	 * check task is stopped
4921	 */
4922	ret = pfm_check_task_state(ctx, cmd, flags);
4923	if (unlikely(ret)) goto abort_locked;
4924
4925skip_fd:
4926	ret = (*func)(ctx, args_k, count, task_pt_regs(current));
4927
4928	call_made = 1;
4929
4930abort_locked:
4931	if (likely(ctx)) {
4932		DPRINT(("context unlocked\n"));
4933		UNPROTECT_CTX(ctx, flags);
4934	}
4935
4936	/* copy argument back to user, if needed */
4937	if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4938
4939error_args:
4940	if (file)
4941		fput(file);
4942
4943	kfree(args_k);
4944
4945	DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4946
4947	return ret;
4948}
4949
4950static void
4951pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4952{
4953	pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4954	pfm_ovfl_ctrl_t rst_ctrl;
4955	int state;
4956	int ret = 0;
4957
4958	state = ctx->ctx_state;
4959	/*
4960	 * Unlock sampling buffer and reset index atomically
4961	 * XXX: not really needed when blocking
4962	 */
4963	if (CTX_HAS_SMPL(ctx)) {
4964
4965		rst_ctrl.bits.mask_monitoring = 0;
4966		rst_ctrl.bits.reset_ovfl_pmds = 0;
4967
4968		if (state == PFM_CTX_LOADED)
4969			ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4970		else
4971			ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4972	} else {
4973		rst_ctrl.bits.mask_monitoring = 0;
4974		rst_ctrl.bits.reset_ovfl_pmds = 1;
4975	}
4976
4977	if (ret == 0) {
4978		if (rst_ctrl.bits.reset_ovfl_pmds) {
4979			pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4980		}
4981		if (rst_ctrl.bits.mask_monitoring == 0) {
4982			DPRINT(("resuming monitoring\n"));
4983			if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4984		} else {
4985			DPRINT(("stopping monitoring\n"));
4986			//pfm_stop_monitoring(current, regs);
4987		}
4988		ctx->ctx_state = PFM_CTX_LOADED;
4989	}
4990}
4991
4992/*
4993 * context MUST BE LOCKED when calling
4994 * can only be called for current
4995 */
4996static void
4997pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4998{
4999	int ret;
5000
5001	DPRINT(("entering for [%d]\n", task_pid_nr(current)));
5002
5003	ret = pfm_context_unload(ctx, NULL, 0, regs);
5004	if (ret) {
5005		printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", task_pid_nr(current), ret);
5006	}
5007
5008	/*
5009	 * and wakeup controlling task, indicating we are now disconnected
5010	 */
5011	wake_up_interruptible(&ctx->ctx_zombieq);
5012
5013	/*
5014	 * given that context is still locked, the controlling
5015	 * task will only get access when we return from
5016	 * pfm_handle_work().
5017	 */
5018}
5019
5020static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
5021
5022 /*
5023  * pfm_handle_work() can be called with interrupts enabled
5024  * (TIF_NEED_RESCHED) or disabled. The down_interruptible
5025  * call may sleep, therefore we must re-enable interrupts
5026  * to avoid deadlocks. It is safe to do so because this function
5027  * is called ONLY when returning to user level (pUStk=1), in which case
5028  * there is no risk of kernel stack overflow due to deep
5029  * interrupt nesting.
5030  */
5031void
5032pfm_handle_work(void)
5033{
5034	pfm_context_t *ctx;
5035	struct pt_regs *regs;
5036	unsigned long flags, dummy_flags;
5037	unsigned long ovfl_regs;
5038	unsigned int reason;
5039	int ret;
5040
5041	ctx = PFM_GET_CTX(current);
5042	if (ctx == NULL) {
5043		printk(KERN_ERR "perfmon: [%d] has no PFM context\n",
5044			task_pid_nr(current));
5045		return;
5046	}
5047
5048	PROTECT_CTX(ctx, flags);
5049
5050	PFM_SET_WORK_PENDING(current, 0);
5051
5052	regs = task_pt_regs(current);
5053
5054	/*
5055	 * extract reason for being here and clear
5056	 */
5057	reason = ctx->ctx_fl_trap_reason;
5058	ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5059	ovfl_regs = ctx->ctx_ovfl_regs[0];
5060
5061	DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5062
5063	/*
5064	 * must be done before we check for simple-reset mode
5065	 */
5066	if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE)
5067		goto do_zombie;
5068
5069	//if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5070	if (reason == PFM_TRAP_REASON_RESET)
5071		goto skip_blocking;
5072
5073	/*
5074	 * restore interrupt mask to what it was on entry.
5075	 * Could be enabled/diasbled.
5076	 */
5077	UNPROTECT_CTX(ctx, flags);
5078
5079	/*
5080	 * force interrupt enable because of down_interruptible()
5081	 */
5082	local_irq_enable();
5083
5084	DPRINT(("before block sleeping\n"));
5085
5086	/*
5087	 * may go through without blocking on SMP systems
5088	 * if restart has been received already by the time we call down()
5089	 */
5090	ret = wait_for_completion_interruptible(&ctx->ctx_restart_done);
5091
5092	DPRINT(("after block sleeping ret=%d\n", ret));
5093
5094	/*
5095	 * lock context and mask interrupts again
5096	 * We save flags into a dummy because we may have
5097	 * altered interrupts mask compared to entry in this
5098	 * function.
5099	 */
5100	PROTECT_CTX(ctx, dummy_flags);
5101
5102	/*
5103	 * we need to read the ovfl_regs only after wake-up
5104	 * because we may have had pfm_write_pmds() in between
5105	 * and that can changed PMD values and therefore 
5106	 * ovfl_regs is reset for these new PMD values.
5107	 */
5108	ovfl_regs = ctx->ctx_ovfl_regs[0];
5109
5110	if (ctx->ctx_fl_going_zombie) {
5111do_zombie:
5112		DPRINT(("context is zombie, bailing out\n"));
5113		pfm_context_force_terminate(ctx, regs);
5114		goto nothing_to_do;
5115	}
5116	/*
5117	 * in case of interruption of down() we don't restart anything
5118	 */
5119	if (ret < 0)
5120		goto nothing_to_do;
5121
5122skip_blocking:
5123	pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5124	ctx->ctx_ovfl_regs[0] = 0UL;
5125
5126nothing_to_do:
5127	/*
5128	 * restore flags as they were upon entry
5129	 */
5130	UNPROTECT_CTX(ctx, flags);
5131}
5132
5133static int
5134pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5135{
5136	if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5137		DPRINT(("ignoring overflow notification, owner is zombie\n"));
5138		return 0;
5139	}
5140
5141	DPRINT(("waking up somebody\n"));
5142
5143	if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5144
5145	/*
5146	 * safe, we are not in intr handler, nor in ctxsw when
5147	 * we come here
5148	 */
5149	kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5150
5151	return 0;
5152}
5153
5154static int
5155pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5156{
5157	pfm_msg_t *msg = NULL;
5158
5159	if (ctx->ctx_fl_no_msg == 0) {
5160		msg = pfm_get_new_msg(ctx);
5161		if (msg == NULL) {
5162			printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5163			return -1;
5164		}
5165
5166		msg->pfm_ovfl_msg.msg_type         = PFM_MSG_OVFL;
5167		msg->pfm_ovfl_msg.msg_ctx_fd       = ctx->ctx_fd;
5168		msg->pfm_ovfl_msg.msg_active_set   = 0;
5169		msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5170		msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5171		msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5172		msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5173		msg->pfm_ovfl_msg.msg_tstamp       = 0UL;
5174	}
5175
5176	DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5177		msg,
5178		ctx->ctx_fl_no_msg,
5179		ctx->ctx_fd,
5180		ovfl_pmds));
5181
5182	return pfm_notify_user(ctx, msg);
5183}
5184
5185static int
5186pfm_end_notify_user(pfm_context_t *ctx)
5187{
5188	pfm_msg_t *msg;
5189
5190	msg = pfm_get_new_msg(ctx);
5191	if (msg == NULL) {
5192		printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5193		return -1;
5194	}
5195	/* no leak */
5196	memset(msg, 0, sizeof(*msg));
5197
5198	msg->pfm_end_msg.msg_type    = PFM_MSG_END;
5199	msg->pfm_end_msg.msg_ctx_fd  = ctx->ctx_fd;
5200	msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5201
5202	DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5203		msg,
5204		ctx->ctx_fl_no_msg,
5205		ctx->ctx_fd));
5206
5207	return pfm_notify_user(ctx, msg);
5208}
5209
5210/*
5211 * main overflow processing routine.
5212 * it can be called from the interrupt path or explicitly during the context switch code
5213 */
5214static void pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx,
5215				unsigned long pmc0, struct pt_regs *regs)
5216{
5217	pfm_ovfl_arg_t *ovfl_arg;
5218	unsigned long mask;
5219	unsigned long old_val, ovfl_val, new_val;
5220	unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5221	unsigned long tstamp;
5222	pfm_ovfl_ctrl_t	ovfl_ctrl;
5223	unsigned int i, has_smpl;
5224	int must_notify = 0;
5225
5226	if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5227
5228	/*
5229	 * sanity test. Should never happen
5230	 */
5231	if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5232
5233	tstamp   = ia64_get_itc();
5234	mask     = pmc0 >> PMU_FIRST_COUNTER;
5235	ovfl_val = pmu_conf->ovfl_val;
5236	has_smpl = CTX_HAS_SMPL(ctx);
5237
5238	DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5239		     "used_pmds=0x%lx\n",
5240			pmc0,
5241			task ? task_pid_nr(task): -1,
5242			(regs ? regs->cr_iip : 0),
5243			CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5244			ctx->ctx_used_pmds[0]));
5245
5246
5247	/*
5248	 * first we update the virtual counters
5249	 * assume there was a prior ia64_srlz_d() issued
5250	 */
5251	for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5252
5253		/* skip pmd which did not overflow */
5254		if ((mask & 0x1) == 0) continue;
5255
5256		/*
5257		 * Note that the pmd is not necessarily 0 at this point as qualified events
5258		 * may have happened before the PMU was frozen. The residual count is not
5259		 * taken into consideration here but will be with any read of the pmd via
5260		 * pfm_read_pmds().
5261		 */
5262		old_val              = new_val = ctx->ctx_pmds[i].val;
5263		new_val             += 1 + ovfl_val;
5264		ctx->ctx_pmds[i].val = new_val;
5265
5266		/*
5267		 * check for overflow condition
5268		 */
5269		if (likely(old_val > new_val)) {
5270			ovfl_pmds |= 1UL << i;
5271			if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5272		}
5273
5274		DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5275			i,
5276			new_val,
5277			old_val,
5278			ia64_get_pmd(i) & ovfl_val,
5279			ovfl_pmds,
5280			ovfl_notify));
5281	}
5282
5283	/*
5284	 * there was no 64-bit overflow, nothing else to do
5285	 */
5286	if (ovfl_pmds == 0UL) return;
5287
5288	/* 
5289	 * reset all control bits
5290	 */
5291	ovfl_ctrl.val = 0;
5292	reset_pmds    = 0UL;
5293
5294	/*
5295	 * if a sampling format module exists, then we "cache" the overflow by 
5296	 * calling the module's handler() routine.
5297	 */
5298	if (has_smpl) {
5299		unsigned long start_cycles, end_cycles;
5300		unsigned long pmd_mask;
5301		int j, k, ret = 0;
5302		int this_cpu = smp_processor_id();
5303
5304		pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5305		ovfl_arg = &ctx->ctx_ovfl_arg;
5306
5307		prefetch(ctx->ctx_smpl_hdr);
5308
5309		for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5310
5311			mask = 1UL << i;
5312
5313			if ((pmd_mask & 0x1) == 0) continue;
5314
5315			ovfl_arg->ovfl_pmd      = (unsigned char )i;
5316			ovfl_arg->ovfl_notify   = ovfl_notify & mask ? 1 : 0;
5317			ovfl_arg->active_set    = 0;
5318			ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5319			ovfl_arg->smpl_pmds[0]  = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5320
5321			ovfl_arg->pmd_value      = ctx->ctx_pmds[i].val;
5322			ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5323			ovfl_arg->pmd_eventid    = ctx->ctx_pmds[i].eventid;
5324
5325			/*
5326		 	 * copy values of pmds of interest. Sampling format may copy them
5327		 	 * into sampling buffer.
5328		 	 */
5329			if (smpl_pmds) {
5330				for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5331					if ((smpl_pmds & 0x1) == 0) continue;
5332					ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ?  pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5333					DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5334				}
5335			}
5336
5337			pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5338
5339			start_cycles = ia64_get_itc();
5340
5341			/*
5342		 	 * call custom buffer format record (handler) routine
5343		 	 */
5344			ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5345
5346			end_cycles = ia64_get_itc();
5347
5348			/*
5349			 * For those controls, we take the union because they have
5350			 * an all or nothing behavior.
5351			 */
5352			ovfl_ctrl.bits.notify_user     |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5353			ovfl_ctrl.bits.block_task      |= ovfl_arg->ovfl_ctrl.bits.block_task;
5354			ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5355			/*
5356			 * build the bitmask of pmds to reset now
5357			 */
5358			if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5359
5360			pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5361		}
5362		/*
5363		 * when the module cannot handle the rest of the overflows, we abort right here
5364		 */
5365		if (ret && pmd_mask) {
5366			DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5367				pmd_mask<<PMU_FIRST_COUNTER));
5368		}
5369		/*
5370		 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5371		 */
5372		ovfl_pmds &= ~reset_pmds;
5373	} else {
5374		/*
5375		 * when no sampling module is used, then the default
5376		 * is to notify on overflow if requested by user
5377		 */
5378		ovfl_ctrl.bits.notify_user     = ovfl_notify ? 1 : 0;
5379		ovfl_ctrl.bits.block_task      = ovfl_notify ? 1 : 0;
5380		ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5381		ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5382		/*
5383		 * if needed, we reset all overflowed pmds
5384		 */
5385		if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5386	}
5387
5388	DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5389
5390	/*
5391	 * reset the requested PMD registers using the short reset values
5392	 */
5393	if (reset_pmds) {
5394		unsigned long bm = reset_pmds;
5395		pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5396	}
5397
5398	if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5399		/*
5400		 * keep track of what to reset when unblocking
5401		 */
5402		ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5403
5404		/*
5405		 * check for blocking context 
5406		 */
5407		if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5408
5409			ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5410
5411			/*
5412			 * set the perfmon specific checking pending work for the task
5413			 */
5414			PFM_SET_WORK_PENDING(task, 1);
5415
5416			/*
5417			 * when coming from ctxsw, current still points to the
5418			 * previous task, therefore we must work with task and not current.
5419			 */
5420			set_notify_resume(task);
5421		}
5422		/*
5423		 * defer until state is changed (shorten spin window). the context is locked
5424		 * anyway, so the signal receiver would come spin for nothing.
5425		 */
5426		must_notify = 1;
5427	}
5428
5429	DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5430			GET_PMU_OWNER() ? task_pid_nr(GET_PMU_OWNER()) : -1,
5431			PFM_GET_WORK_PENDING(task),
5432			ctx->ctx_fl_trap_reason,
5433			ovfl_pmds,
5434			ovfl_notify,
5435			ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5436	/*
5437	 * in case monitoring must be stopped, we toggle the psr bits
5438	 */
5439	if (ovfl_ctrl.bits.mask_monitoring) {
5440		pfm_mask_monitoring(task);
5441		ctx->ctx_state = PFM_CTX_MASKED;
5442		ctx->ctx_fl_can_restart = 1;
5443	}
5444
5445	/*
5446	 * send notification now
5447	 */
5448	if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5449
5450	return;
5451
5452sanity_check:
5453	printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5454			smp_processor_id(),
5455			task ? task_pid_nr(task) : -1,
5456			pmc0);
5457	return;
5458
5459stop_monitoring:
5460	/*
5461	 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5462	 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5463	 * come here as zombie only if the task is the current task. In which case, we
5464	 * can access the PMU  hardware directly.
5465	 *
5466	 * Note that zombies do have PM_VALID set. So here we do the minimal.
5467	 *
5468	 * In case the context was zombified it could not be reclaimed at the time
5469	 * the monitoring program exited. At this point, the PMU reservation has been
5470	 * returned, the sampiing buffer has been freed. We must convert this call
5471	 * into a spurious interrupt. However, we must also avoid infinite overflows
5472	 * by stopping monitoring for this task. We can only come here for a per-task
5473	 * context. All we need to do is to stop monitoring using the psr bits which
5474	 * are always task private. By re-enabling secure montioring, we ensure that
5475	 * the monitored task will not be able to re-activate monitoring.
5476	 * The task will eventually be context switched out, at which point the context
5477	 * will be reclaimed (that includes releasing ownership of the PMU).
5478	 *
5479	 * So there might be a window of time where the number of per-task session is zero
5480	 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5481	 * context. This is safe because if a per-task session comes in, it will push this one
5482	 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5483	 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5484	 * also push our zombie context out.
5485	 *
5486	 * Overall pretty hairy stuff....
5487	 */
5488	DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task_pid_nr(task): -1));
5489	pfm_clear_psr_up();
5490	ia64_psr(regs)->up = 0;
5491	ia64_psr(regs)->sp = 1;
5492	return;
5493}
5494
5495static int
5496pfm_do_interrupt_handler(void *arg, struct pt_regs *regs)
5497{
5498	struct task_struct *task;
5499	pfm_context_t *ctx;
5500	unsigned long flags;
5501	u64 pmc0;
5502	int this_cpu = smp_processor_id();
5503	int retval = 0;
5504
5505	pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5506
5507	/*
5508	 * srlz.d done before arriving here
5509	 */
5510	pmc0 = ia64_get_pmc(0);
5511
5512	task = GET_PMU_OWNER();
5513	ctx  = GET_PMU_CTX();
5514
5515	/*
5516	 * if we have some pending bits set
5517	 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5518	 */
5519	if (PMC0_HAS_OVFL(pmc0) && task) {
5520		/*
5521		 * we assume that pmc0.fr is always set here
5522		 */
5523
5524		/* sanity check */
5525		if (!ctx) goto report_spurious1;
5526
5527		if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0) 
5528			goto report_spurious2;
5529
5530		PROTECT_CTX_NOPRINT(ctx, flags);
5531
5532		pfm_overflow_handler(task, ctx, pmc0, regs);
5533
5534		UNPROTECT_CTX_NOPRINT(ctx, flags);
5535
5536	} else {
5537		pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5538		retval = -1;
5539	}
5540	/*
5541	 * keep it unfrozen at all times
5542	 */
5543	pfm_unfreeze_pmu();
5544
5545	return retval;
5546
5547report_spurious1:
5548	printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5549		this_cpu, task_pid_nr(task));
5550	pfm_unfreeze_pmu();
5551	return -1;
5552report_spurious2:
5553	printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n", 
5554		this_cpu, 
5555		task_pid_nr(task));
5556	pfm_unfreeze_pmu();
5557	return -1;
5558}
5559
5560static irqreturn_t
5561pfm_interrupt_handler(int irq, void *arg)
5562{
5563	unsigned long start_cycles, total_cycles;
5564	unsigned long min, max;
5565	int this_cpu;
5566	int ret;
5567	struct pt_regs *regs = get_irq_regs();
5568
5569	this_cpu = get_cpu();
5570	if (likely(!pfm_alt_intr_handler)) {
5571		min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5572		max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5573
5574		start_cycles = ia64_get_itc();
5575
5576		ret = pfm_do_interrupt_handler(arg, regs);
5577
5578		total_cycles = ia64_get_itc();
5579
5580		/*
5581		 * don't measure spurious interrupts
5582		 */
5583		if (likely(ret == 0)) {
5584			total_cycles -= start_cycles;
5585
5586			if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5587			if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5588
5589			pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5590		}
5591	}
5592	else {
5593		(*pfm_alt_intr_handler->handler)(irq, arg, regs);
5594	}
5595
5596	put_cpu();
5597	return IRQ_HANDLED;
5598}
5599
5600/*
5601 * /proc/perfmon interface, for debug only
5602 */
5603
5604#define PFM_PROC_SHOW_HEADER	((void *)(long)nr_cpu_ids+1)
5605
5606static void *
5607pfm_proc_start(struct seq_file *m, loff_t *pos)
5608{
5609	if (*pos == 0) {
5610		return PFM_PROC_SHOW_HEADER;
5611	}
5612
5613	while (*pos <= nr_cpu_ids) {
5614		if (cpu_online(*pos - 1)) {
5615			return (void *)*pos;
5616		}
5617		++*pos;
5618	}
5619	return NULL;
5620}
5621
5622static void *
5623pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5624{
5625	++*pos;
5626	return pfm_proc_start(m, pos);
5627}
5628
5629static void
5630pfm_proc_stop(struct seq_file *m, void *v)
5631{
5632}
5633
5634static void
5635pfm_proc_show_header(struct seq_file *m)
5636{
5637	struct list_head * pos;
5638	pfm_buffer_fmt_t * entry;
5639	unsigned long flags;
5640
5641 	seq_printf(m,
5642		"perfmon version           : %u.%u\n"
5643		"model                     : %s\n"
5644		"fastctxsw                 : %s\n"
5645		"expert mode               : %s\n"
5646		"ovfl_mask                 : 0x%lx\n"
5647		"PMU flags                 : 0x%x\n",
5648		PFM_VERSION_MAJ, PFM_VERSION_MIN,
5649		pmu_conf->pmu_name,
5650		pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5651		pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5652		pmu_conf->ovfl_val,
5653		pmu_conf->flags);
5654
5655  	LOCK_PFS(flags);
5656
5657 	seq_printf(m,
5658 		"proc_sessions             : %u\n"
5659 		"sys_sessions              : %u\n"
5660 		"sys_use_dbregs            : %u\n"
5661 		"ptrace_use_dbregs         : %u\n",
5662 		pfm_sessions.pfs_task_sessions,
5663 		pfm_sessions.pfs_sys_sessions,
5664 		pfm_sessions.pfs_sys_use_dbregs,
5665 		pfm_sessions.pfs_ptrace_use_dbregs);
5666
5667  	UNLOCK_PFS(flags);
5668
5669	spin_lock(&pfm_buffer_fmt_lock);
5670
5671	list_for_each(pos, &pfm_buffer_fmt_list) {
5672		entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5673		seq_printf(m, "format                    : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5674			entry->fmt_uuid[0],
5675			entry->fmt_uuid[1],
5676			entry->fmt_uuid[2],
5677			entry->fmt_uuid[3],
5678			entry->fmt_uuid[4],
5679			entry->fmt_uuid[5],
5680			entry->fmt_uuid[6],
5681			entry->fmt_uuid[7],
5682			entry->fmt_uuid[8],
5683			entry->fmt_uuid[9],
5684			entry->fmt_uuid[10],
5685			entry->fmt_uuid[11],
5686			entry->fmt_uuid[12],
5687			entry->fmt_uuid[13],
5688			entry->fmt_uuid[14],
5689			entry->fmt_uuid[15],
5690			entry->fmt_name);
5691	}
5692	spin_unlock(&pfm_buffer_fmt_lock);
5693
5694}
5695
5696static int
5697pfm_proc_show(struct seq_file *m, void *v)
5698{
5699	unsigned long psr;
5700	unsigned int i;
5701	int cpu;
5702
5703	if (v == PFM_PROC_SHOW_HEADER) {
5704		pfm_proc_show_header(m);
5705		return 0;
5706	}
5707
5708	/* show info for CPU (v - 1) */
5709
5710	cpu = (long)v - 1;
5711	seq_printf(m,
5712		"CPU%-2d overflow intrs      : %lu\n"
5713		"CPU%-2d overflow cycles     : %lu\n"
5714		"CPU%-2d overflow min        : %lu\n"
5715		"CPU%-2d overflow max        : %lu\n"
5716		"CPU%-2d smpl handler calls  : %lu\n"
5717		"CPU%-2d smpl handler cycles : %lu\n"
5718		"CPU%-2d spurious intrs      : %lu\n"
5719		"CPU%-2d replay   intrs      : %lu\n"
5720		"CPU%-2d syst_wide           : %d\n"
5721		"CPU%-2d dcr_pp              : %d\n"
5722		"CPU%-2d exclude idle        : %d\n"
5723		"CPU%-2d owner               : %d\n"
5724		"CPU%-2d context             : %p\n"
5725		"CPU%-2d activations         : %lu\n",
5726		cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5727		cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5728		cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5729		cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5730		cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5731		cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5732		cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5733		cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5734		cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5735		cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5736		cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5737		cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5738		cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5739		cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5740
5741	if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5742
5743		psr = pfm_get_psr();
5744
5745		ia64_srlz_d();
5746
5747		seq_printf(m, 
5748			"CPU%-2d psr                 : 0x%lx\n"
5749			"CPU%-2d pmc0                : 0x%lx\n", 
5750			cpu, psr,
5751			cpu, ia64_get_pmc(0));
5752
5753		for (i=0; PMC_IS_LAST(i) == 0;  i++) {
5754			if (PMC_IS_COUNTING(i) == 0) continue;
5755   			seq_printf(m, 
5756				"CPU%-2d pmc%u                : 0x%lx\n"
5757   				"CPU%-2d pmd%u                : 0x%lx\n", 
5758				cpu, i, ia64_get_pmc(i),
5759				cpu, i, ia64_get_pmd(i));
5760  		}
5761	}
5762	return 0;
5763}
5764
5765const struct seq_operations pfm_seq_ops = {
5766	.start =	pfm_proc_start,
5767 	.next =		pfm_proc_next,
5768 	.stop =		pfm_proc_stop,
5769 	.show =		pfm_proc_show
5770};
5771
5772static int
5773pfm_proc_open(struct inode *inode, struct file *file)
5774{
5775	return seq_open(file, &pfm_seq_ops);
5776}
5777
5778
5779/*
5780 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5781 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5782 * is active or inactive based on mode. We must rely on the value in
5783 * local_cpu_data->pfm_syst_info
5784 */
5785void
5786pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5787{
5788	struct pt_regs *regs;
5789	unsigned long dcr;
5790	unsigned long dcr_pp;
5791
5792	dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5793
5794	/*
5795	 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5796	 * on every CPU, so we can rely on the pid to identify the idle task.
5797	 */
5798	if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5799		regs = task_pt_regs(task);
5800		ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5801		return;
5802	}
5803	/*
5804	 * if monitoring has started
5805	 */
5806	if (dcr_pp) {
5807		dcr = ia64_getreg(_IA64_REG_CR_DCR);
5808		/*
5809		 * context switching in?
5810		 */
5811		if (is_ctxswin) {
5812			/* mask monitoring for the idle task */
5813			ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5814			pfm_clear_psr_pp();
5815			ia64_srlz_i();
5816			return;
5817		}
5818		/*
5819		 * context switching out
5820		 * restore monitoring for next task
5821		 *
5822		 * Due to inlining this odd if-then-else construction generates
5823		 * better code.
5824		 */
5825		ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5826		pfm_set_psr_pp();
5827		ia64_srlz_i();
5828	}
5829}
5830
5831#ifdef CONFIG_SMP
5832
5833static void
5834pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5835{
5836	struct task_struct *task = ctx->ctx_task;
5837
5838	ia64_psr(regs)->up = 0;
5839	ia64_psr(regs)->sp = 1;
5840
5841	if (GET_PMU_OWNER() == task) {
5842		DPRINT(("cleared ownership for [%d]\n",
5843					task_pid_nr(ctx->ctx_task)));
5844		SET_PMU_OWNER(NULL, NULL);
5845	}
5846
5847	/*
5848	 * disconnect the task from the context and vice-versa
5849	 */
5850	PFM_SET_WORK_PENDING(task, 0);
5851
5852	task->thread.pfm_context  = NULL;
5853	task->thread.flags       &= ~IA64_THREAD_PM_VALID;
5854
5855	DPRINT(("force cleanup for [%d]\n",  task_pid_nr(task)));
5856}
5857
5858
5859/*
5860 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5861 */
5862void
5863pfm_save_regs(struct task_struct *task)
5864{
5865	pfm_context_t *ctx;
5866	unsigned long flags;
5867	u64 psr;
5868
5869
5870	ctx = PFM_GET_CTX(task);
5871	if (ctx == NULL) return;
5872
5873	/*
5874 	 * we always come here with interrupts ALREADY disabled by
5875 	 * the scheduler. So we simply need to protect against concurrent
5876	 * access, not CPU concurrency.
5877	 */
5878	flags = pfm_protect_ctx_ctxsw(ctx);
5879
5880	if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5881		struct pt_regs *regs = task_pt_regs(task);
5882
5883		pfm_clear_psr_up();
5884
5885		pfm_force_cleanup(ctx, regs);
5886
5887		BUG_ON(ctx->ctx_smpl_hdr);
5888
5889		pfm_unprotect_ctx_ctxsw(ctx, flags);
5890
5891		pfm_context_free(ctx);
5892		return;
5893	}
5894
5895	/*
5896	 * save current PSR: needed because we modify it
5897	 */
5898	ia64_srlz_d();
5899	psr = pfm_get_psr();
5900
5901	BUG_ON(psr & (IA64_PSR_I));
5902
5903	/*
5904	 * stop monitoring:
5905	 * This is the last instruction which may generate an overflow
5906	 *
5907	 * We do not need to set psr.sp because, it is irrelevant in kernel.
5908	 * It will be restored from ipsr when going back to user level
5909	 */
5910	pfm_clear_psr_up();
5911
5912	/*
5913	 * keep a copy of psr.up (for reload)
5914	 */
5915	ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5916
5917	/*
5918	 * release ownership of this PMU.
5919	 * PM interrupts are masked, so nothing
5920	 * can happen.
5921	 */
5922	SET_PMU_OWNER(NULL, NULL);
5923
5924	/*
5925	 * we systematically save the PMD as we have no
5926	 * guarantee we will be schedule at that same
5927	 * CPU again.
5928	 */
5929	pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5930
5931	/*
5932	 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5933	 * we will need it on the restore path to check
5934	 * for pending overflow.
5935	 */
5936	ctx->th_pmcs[0] = ia64_get_pmc(0);
5937
5938	/*
5939	 * unfreeze PMU if had pending overflows
5940	 */
5941	if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5942
5943	/*
5944	 * finally, allow context access.
5945	 * interrupts will still be masked after this call.
5946	 */
5947	pfm_unprotect_ctx_ctxsw(ctx, flags);
5948}
5949
5950#else /* !CONFIG_SMP */
5951void
5952pfm_save_regs(struct task_struct *task)
5953{
5954	pfm_context_t *ctx;
5955	u64 psr;
5956
5957	ctx = PFM_GET_CTX(task);
5958	if (ctx == NULL) return;
5959
5960	/*
5961	 * save current PSR: needed because we modify it
5962	 */
5963	psr = pfm_get_psr();
5964
5965	BUG_ON(psr & (IA64_PSR_I));
5966
5967	/*
5968	 * stop monitoring:
5969	 * This is the last instruction which may generate an overflow
5970	 *
5971	 * We do not need to set psr.sp because, it is irrelevant in kernel.
5972	 * It will be restored from ipsr when going back to user level
5973	 */
5974	pfm_clear_psr_up();
5975
5976	/*
5977	 * keep a copy of psr.up (for reload)
5978	 */
5979	ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5980}
5981
5982static void
5983pfm_lazy_save_regs (struct task_struct *task)
5984{
5985	pfm_context_t *ctx;
5986	unsigned long flags;
5987
5988	{ u64 psr  = pfm_get_psr();
5989	  BUG_ON(psr & IA64_PSR_UP);
5990	}
5991
5992	ctx = PFM_GET_CTX(task);
5993
5994	/*
5995	 * we need to mask PMU overflow here to
5996	 * make sure that we maintain pmc0 until
5997	 * we save it. overflow interrupts are
5998	 * treated as spurious if there is no
5999	 * owner.
6000	 *
6001	 * XXX: I don't think this is necessary
6002	 */
6003	PROTECT_CTX(ctx,flags);
6004
6005	/*
6006	 * release ownership of this PMU.
6007	 * must be done before we save the registers.
6008	 *
6009	 * after this call any PMU interrupt is treated
6010	 * as spurious.
6011	 */
6012	SET_PMU_OWNER(NULL, NULL);
6013
6014	/*
6015	 * save all the pmds we use
6016	 */
6017	pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
6018
6019	/*
6020	 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
6021	 * it is needed to check for pended overflow
6022	 * on the restore path
6023	 */
6024	ctx->th_pmcs[0] = ia64_get_pmc(0);
6025
6026	/*
6027	 * unfreeze PMU if had pending overflows
6028	 */
6029	if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
6030
6031	/*
6032	 * now get can unmask PMU interrupts, they will
6033	 * be treated as purely spurious and we will not
6034	 * lose any information
6035	 */
6036	UNPROTECT_CTX(ctx,flags);
6037}
6038#endif /* CONFIG_SMP */
6039
6040#ifdef CONFIG_SMP
6041/*
6042 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
6043 */
6044void
6045pfm_load_regs (struct task_struct *task)
6046{
6047	pfm_context_t *ctx;
6048	unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
6049	unsigned long flags;
6050	u64 psr, psr_up;
6051	int need_irq_resend;
6052
6053	ctx = PFM_GET_CTX(task);
6054	if (unlikely(ctx == NULL)) return;
6055
6056	BUG_ON(GET_PMU_OWNER());
6057
6058	/*
6059	 * possible on unload
6060	 */
6061	if (unlikely((task->thread.flags & IA64_THREAD_PM_VALID) == 0)) return;
6062
6063	/*
6064 	 * we always come here with interrupts ALREADY disabled by
6065 	 * the scheduler. So we simply need to protect against concurrent
6066	 * access, not CPU concurrency.
6067	 */
6068	flags = pfm_protect_ctx_ctxsw(ctx);
6069	psr   = pfm_get_psr();
6070
6071	need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6072
6073	BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6074	BUG_ON(psr & IA64_PSR_I);
6075
6076	if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6077		struct pt_regs *regs = task_pt_regs(task);
6078
6079		BUG_ON(ctx->ctx_smpl_hdr);
6080
6081		pfm_force_cleanup(ctx, regs);
6082
6083		pfm_unprotect_ctx_ctxsw(ctx, flags);
6084
6085		/*
6086		 * this one (kmalloc'ed) is fine with interrupts disabled
6087		 */
6088		pfm_context_free(ctx);
6089
6090		return;
6091	}
6092
6093	/*
6094	 * we restore ALL the debug registers to avoid picking up
6095	 * stale state.
6096	 */
6097	if (ctx->ctx_fl_using_dbreg) {
6098		pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6099		pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6100	}
6101	/*
6102	 * retrieve saved psr.up
6103	 */
6104	psr_up = ctx->ctx_saved_psr_up;
6105
6106	/*
6107	 * if we were the last user of the PMU on that CPU,
6108	 * then nothing to do except restore psr
6109	 */
6110	if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6111
6112		/*
6113		 * retrieve partial reload masks (due to user modifications)
6114		 */
6115		pmc_mask = ctx->ctx_reload_pmcs[0];
6116		pmd_mask = ctx->ctx_reload_pmds[0];
6117
6118	} else {
6119		/*
6120	 	 * To avoid leaking information to the user level when psr.sp=0,
6121	 	 * we must reload ALL implemented pmds (even the ones we don't use).
6122	 	 * In the kernel we only allow PFM_READ_PMDS on registers which
6123	 	 * we initialized or requested (sampling) so there is no risk there.
6124	 	 */
6125		pmd_mask = pfm_sysctl.fastctxsw ?  ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6126
6127		/*
6128	 	 * ALL accessible PMCs are systematically reloaded, unused registers
6129	 	 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6130	 	 * up stale configuration.
6131	 	 *
6132	 	 * PMC0 is never in the mask. It is always restored separately.
6133	 	 */
6134		pmc_mask = ctx->ctx_all_pmcs[0];
6135	}
6136	/*
6137	 * when context is MASKED, we will restore PMC with plm=0
6138	 * and PMD with stale information, but that's ok, nothing
6139	 * will be captured.
6140	 *
6141	 * XXX: optimize here
6142	 */
6143	if (pmd_mask) pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6144	if (pmc_mask) pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6145
6146	/*
6147	 * check for pending overflow at the time the state
6148	 * was saved.
6149	 */
6150	if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6151		/*
6152		 * reload pmc0 with the overflow information
6153		 * On McKinley PMU, this will trigger a PMU interrupt
6154		 */
6155		ia64_set_pmc(0, ctx->th_pmcs[0]);
6156		ia64_srlz_d();
6157		ctx->th_pmcs[0] = 0UL;
6158
6159		/*
6160		 * will replay the PMU interrupt
6161		 */
6162		if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6163
6164		pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6165	}
6166
6167	/*
6168	 * we just did a reload, so we reset the partial reload fields
6169	 */
6170	ctx->ctx_reload_pmcs[0] = 0UL;
6171	ctx->ctx_reload_pmds[0] = 0UL;
6172
6173	SET_LAST_CPU(ctx, smp_processor_id());
6174
6175	/*
6176	 * dump activation value for this PMU
6177	 */
6178	INC_ACTIVATION();
6179	/*
6180	 * record current activation for this context
6181	 */
6182	SET_ACTIVATION(ctx);
6183
6184	/*
6185	 * establish new ownership. 
6186	 */
6187	SET_PMU_OWNER(task, ctx);
6188
6189	/*
6190	 * restore the psr.up bit. measurement
6191	 * is active again.
6192	 * no PMU interrupt can happen at this point
6193	 * because we still have interrupts disabled.
6194	 */
6195	if (likely(psr_up)) pfm_set_psr_up();
6196
6197	/*
6198	 * allow concurrent access to context
6199	 */
6200	pfm_unprotect_ctx_ctxsw(ctx, flags);
6201}
6202#else /*  !CONFIG_SMP */
6203/*
6204 * reload PMU state for UP kernels
6205 * in 2.5 we come here with interrupts disabled
6206 */
6207void
6208pfm_load_regs (struct task_struct *task)
6209{
6210	pfm_context_t *ctx;
6211	struct task_struct *owner;
6212	unsigned long pmd_mask, pmc_mask;
6213	u64 psr, psr_up;
6214	int need_irq_resend;
6215
6216	owner = GET_PMU_OWNER();
6217	ctx   = PFM_GET_CTX(task);
6218	psr   = pfm_get_psr();
6219
6220	BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6221	BUG_ON(psr & IA64_PSR_I);
6222
6223	/*
6224	 * we restore ALL the debug registers to avoid picking up
6225	 * stale state.
6226	 *
6227	 * This must be done even when the task is still the owner
6228	 * as the registers may have been modified via ptrace()
6229	 * (not perfmon) by the previous task.
6230	 */
6231	if (ctx->ctx_fl_using_dbreg) {
6232		pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6233		pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6234	}
6235
6236	/*
6237	 * retrieved saved psr.up
6238	 */
6239	psr_up = ctx->ctx_saved_psr_up;
6240	need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6241
6242	/*
6243	 * short path, our state is still there, just
6244	 * need to restore psr and we go
6245	 *
6246	 * we do not touch either PMC nor PMD. the psr is not touched
6247	 * by the overflow_handler. So we are safe w.r.t. to interrupt
6248	 * concurrency even without interrupt masking.
6249	 */
6250	if (likely(owner == task)) {
6251		if (likely(psr_up)) pfm_set_psr_up();
6252		return;
6253	}
6254
6255	/*
6256	 * someone else is still using the PMU, first push it out and
6257	 * then we'll be able to install our stuff !
6258	 *
6259	 * Upon return, there will be no owner for the current PMU
6260	 */
6261	if (owner) pfm_lazy_save_regs(owner);
6262
6263	/*
6264	 * To avoid leaking information to the user level when psr.sp=0,
6265	 * we must reload ALL implemented pmds (even the ones we don't use).
6266	 * In the kernel we only allow PFM_READ_PMDS on registers which
6267	 * we initialized or requested (sampling) so there is no risk there.
6268	 */
6269	pmd_mask = pfm_sysctl.fastctxsw ?  ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6270
6271	/*
6272	 * ALL accessible PMCs are systematically reloaded, unused registers
6273	 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6274	 * up stale configuration.
6275	 *
6276	 * PMC0 is never in the mask. It is always restored separately
6277	 */
6278	pmc_mask = ctx->ctx_all_pmcs[0];
6279
6280	pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6281	pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6282
6283	/*
6284	 * check for pending overflow at the time the state
6285	 * was saved.
6286	 */
6287	if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6288		/*
6289		 * reload pmc0 with the overflow information
6290		 * On McKinley PMU, this will trigger a PMU interrupt
6291		 */
6292		ia64_set_pmc(0, ctx->th_pmcs[0]);
6293		ia64_srlz_d();
6294
6295		ctx->th_pmcs[0] = 0UL;
6296
6297		/*
6298		 * will replay the PMU interrupt
6299		 */
6300		if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6301
6302		pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6303	}
6304
6305	/*
6306	 * establish new ownership. 
6307	 */
6308	SET_PMU_OWNER(task, ctx);
6309
6310	/*
6311	 * restore the psr.up bit. measurement
6312	 * is active again.
6313	 * no PMU interrupt can happen at this point
6314	 * because we still have interrupts disabled.
6315	 */
6316	if (likely(psr_up)) pfm_set_psr_up();
6317}
6318#endif /* CONFIG_SMP */
6319
6320/*
6321 * this function assumes monitoring is stopped
6322 */
6323static void
6324pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6325{
6326	u64 pmc0;
6327	unsigned long mask2, val, pmd_val, ovfl_val;
6328	int i, can_access_pmu = 0;
6329	int is_self;
6330
6331	/*
6332	 * is the caller the task being monitored (or which initiated the
6333	 * session for system wide measurements)
6334	 */
6335	is_self = ctx->ctx_task == task ? 1 : 0;
6336
6337	/*
6338	 * can access PMU is task is the owner of the PMU state on the current CPU
6339	 * or if we are running on the CPU bound to the context in system-wide mode
6340	 * (that is not necessarily the task the context is attached to in this mode).
6341	 * In system-wide we always have can_access_pmu true because a task running on an
6342	 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6343	 */
6344	can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6345	if (can_access_pmu) {
6346		/*
6347		 * Mark the PMU as not owned
6348		 * This will cause the interrupt handler to do nothing in case an overflow
6349		 * interrupt was in-flight
6350		 * This also guarantees that pmc0 will contain the final state
6351		 * It virtually gives us full control on overflow processing from that point
6352		 * on.
6353		 */
6354		SET_PMU_OWNER(NULL, NULL);
6355		DPRINT(("releasing ownership\n"));
6356
6357		/*
6358		 * read current overflow status:
6359		 *
6360		 * we are guaranteed to read the final stable state
6361		 */
6362		ia64_srlz_d();
6363		pmc0 = ia64_get_pmc(0); /* slow */
6364
6365		/*
6366		 * reset freeze bit, overflow status information destroyed
6367		 */
6368		pfm_unfreeze_pmu();
6369	} else {
6370		pmc0 = ctx->th_pmcs[0];
6371		/*
6372		 * clear whatever overflow status bits there were
6373		 */
6374		ctx->th_pmcs[0] = 0;
6375	}
6376	ovfl_val = pmu_conf->ovfl_val;
6377	/*
6378	 * we save all the used pmds
6379	 * we take care of overflows for counting PMDs
6380	 *
6381	 * XXX: sampling situation is not taken into account here
6382	 */
6383	mask2 = ctx->ctx_used_pmds[0];
6384
6385	DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6386
6387	for (i = 0; mask2; i++, mask2>>=1) {
6388
6389		/* skip non used pmds */
6390		if ((mask2 & 0x1) == 0) continue;
6391
6392		/*
6393		 * can access PMU always true in system wide mode
6394		 */
6395		val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : ctx->th_pmds[i];
6396
6397		if (PMD_IS_COUNTING(i)) {
6398			DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6399				task_pid_nr(task),
6400				i,
6401				ctx->ctx_pmds[i].val,
6402				val & ovfl_val));
6403
6404			/*
6405			 * we rebuild the full 64 bit value of the counter
6406			 */
6407			val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6408
6409			/*
6410			 * now everything is in ctx_pmds[] and we need
6411			 * to clear the saved context from save_regs() such that
6412			 * pfm_read_pmds() gets the correct value
6413			 */
6414			pmd_val = 0UL;
6415
6416			/*
6417			 * take care of overflow inline
6418			 */
6419			if (pmc0 & (1UL << i)) {
6420				val += 1 + ovfl_val;
6421				DPRINT(("[%d] pmd[%d] overflowed\n", task_pid_nr(task), i));
6422			}
6423		}
6424
6425		DPRINT(("[%d] ctx_pmd[%d]=0x%lx  pmd_val=0x%lx\n", task_pid_nr(task), i, val, pmd_val));
6426
6427		if (is_self) ctx->th_pmds[i] = pmd_val;
6428
6429		ctx->ctx_pmds[i].val = val;
6430	}
6431}
6432
6433static struct irqaction perfmon_irqaction = {
6434	.handler = pfm_interrupt_handler,
6435	.flags   = IRQF_DISABLED,
6436	.name    = "perfmon"
6437};
6438
6439static void
6440pfm_alt_save_pmu_state(void *data)
6441{
6442	struct pt_regs *regs;
6443
6444	regs = task_pt_regs(current);
6445
6446	DPRINT(("called\n"));
6447
6448	/*
6449	 * should not be necessary but
6450	 * let's take not risk
6451	 */
6452	pfm_clear_psr_up();
6453	pfm_clear_psr_pp();
6454	ia64_psr(regs)->pp = 0;
6455
6456	/*
6457	 * This call is required
6458	 * May cause a spurious interrupt on some processors
6459	 */
6460	pfm_freeze_pmu();
6461
6462	ia64_srlz_d();
6463}
6464
6465void
6466pfm_alt_restore_pmu_state(void *data)
6467{
6468	struct pt_regs *regs;
6469
6470	regs = task_pt_regs(current);
6471
6472	DPRINT(("called\n"));
6473
6474	/*
6475	 * put PMU back in state expected
6476	 * by perfmon
6477	 */
6478	pfm_clear_psr_up();
6479	pfm_clear_psr_pp();
6480	ia64_psr(regs)->pp = 0;
6481
6482	/*
6483	 * perfmon runs with PMU unfrozen at all times
6484	 */
6485	pfm_unfreeze_pmu();
6486
6487	ia64_srlz_d();
6488}
6489
6490int
6491pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6492{
6493	int ret, i;
6494	int reserve_cpu;
6495
6496	/* some sanity checks */
6497	if (hdl == NULL || hdl->handler == NULL) return -EINVAL;
6498
6499	/* do the easy test first */
6500	if (pfm_alt_intr_handler) return -EBUSY;
6501
6502	/* one at a time in the install or remove, just fail the others */
6503	if (!spin_trylock(&pfm_alt_install_check)) {
6504		return -EBUSY;
6505	}
6506
6507	/* reserve our session */
6508	for_each_online_cpu(reserve_cpu) {
6509		ret = pfm_reserve_session(NULL, 1, reserve_cpu);
6510		if (ret) goto cleanup_reserve;
6511	}
6512
6513	/* save the current system wide pmu states */
6514	ret = on_each_cpu(pfm_alt_save_pmu_state, NULL, 1);
6515	if (ret) {
6516		DPRINT(("on_each_cpu() failed: %d\n", ret));
6517		goto cleanup_reserve;
6518	}
6519
6520	/* officially change to the alternate interrupt handler */
6521	pfm_alt_intr_handler = hdl;
6522
6523	spin_unlock(&pfm_alt_install_check);
6524
6525	return 0;
6526
6527cleanup_reserve:
6528	for_each_online_cpu(i) {
6529		/* don't unreserve more than we reserved */
6530		if (i >= reserve_cpu) break;
6531
6532		pfm_unreserve_session(NULL, 1, i);
6533	}
6534
6535	spin_unlock(&pfm_alt_install_check);
6536
6537	return ret;
6538}
6539EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt);
6540
6541int
6542pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6543{
6544	int i;
6545	int ret;
6546
6547	if (hdl == NULL) return -EINVAL;
6548
6549	/* cannot remove someone else's handler! */
6550	if (pfm_alt_intr_handler != hdl) return -EINVAL;
6551
6552	/* one at a time in the install or remove, just fail the others */
6553	if (!spin_trylock(&pfm_alt_install_check)) {
6554		return -EBUSY;
6555	}
6556
6557	pfm_alt_intr_handler = NULL;
6558
6559	ret = on_each_cpu(pfm_alt_restore_pmu_state, NULL, 1);
6560	if (ret) {
6561		DPRINT(("on_each_cpu() failed: %d\n", ret));
6562	}
6563
6564	for_each_online_cpu(i) {
6565		pfm_unreserve_session(NULL, 1, i);
6566	}
6567
6568	spin_unlock(&pfm_alt_install_check);
6569
6570	return 0;
6571}
6572EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt);
6573
6574/*
6575 * perfmon initialization routine, called from the initcall() table
6576 */
6577static int init_pfm_fs(void);
6578
6579static int __init
6580pfm_probe_pmu(void)
6581{
6582	pmu_config_t **p;
6583	int family;
6584
6585	family = local_cpu_data->family;
6586	p      = pmu_confs;
6587
6588	while(*p) {
6589		if ((*p)->probe) {
6590			if ((*p)->probe() == 0) goto found;
6591		} else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6592			goto found;
6593		}
6594		p++;
6595	}
6596	return -1;
6597found:
6598	pmu_conf = *p;
6599	return 0;
6600}
6601
6602static const struct file_operations pfm_proc_fops = {
6603	.open		= pfm_proc_open,
6604	.read		= seq_read,
6605	.llseek		= seq_lseek,
6606	.release	= seq_release,
6607};
6608
6609int __init
6610pfm_init(void)
6611{
6612	unsigned int n, n_counters, i;
6613
6614	printk("perfmon: version %u.%u IRQ %u\n",
6615		PFM_VERSION_MAJ,
6616		PFM_VERSION_MIN,
6617		IA64_PERFMON_VECTOR);
6618
6619	if (pfm_probe_pmu()) {
6620		printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n", 
6621				local_cpu_data->family);
6622		return -ENODEV;
6623	}
6624
6625	/*
6626	 * compute the number of implemented PMD/PMC from the
6627	 * description tables
6628	 */
6629	n = 0;
6630	for (i=0; PMC_IS_LAST(i) == 0;  i++) {
6631		if (PMC_IS_IMPL(i) == 0) continue;
6632		pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6633		n++;
6634	}
6635	pmu_conf->num_pmcs = n;
6636
6637	n = 0; n_counters = 0;
6638	for (i=0; PMD_IS_LAST(i) == 0;  i++) {
6639		if (PMD_IS_IMPL(i) == 0) continue;
6640		pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6641		n++;
6642		if (PMD_IS_COUNTING(i)) n_counters++;
6643	}
6644	pmu_conf->num_pmds      = n;
6645	pmu_conf->num_counters  = n_counters;
6646
6647	/*
6648	 * sanity checks on the number of debug registers
6649	 */
6650	if (pmu_conf->use_rr_dbregs) {
6651		if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6652			printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6653			pmu_conf = NULL;
6654			return -1;
6655		}
6656		if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6657			printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6658			pmu_conf = NULL;
6659			return -1;
6660		}
6661	}
6662
6663	printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6664	       pmu_conf->pmu_name,
6665	       pmu_conf->num_pmcs,
6666	       pmu_conf->num_pmds,
6667	       pmu_conf->num_counters,
6668	       ffz(pmu_conf->ovfl_val));
6669
6670	/* sanity check */
6671	if (pmu_conf->num_pmds >= PFM_NUM_PMD_REGS || pmu_conf->num_pmcs >= PFM_NUM_PMC_REGS) {
6672		printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6673		pmu_conf = NULL;
6674		return -1;
6675	}
6676
6677	/*
6678	 * create /proc/perfmon (mostly for debugging purposes)
6679	 */
6680	perfmon_dir = proc_create("perfmon", S_IRUGO, NULL, &pfm_proc_fops);
6681	if (perfmon_dir == NULL) {
6682		printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6683		pmu_conf = NULL;
6684		return -1;
6685	}
6686
6687	/*
6688	 * create /proc/sys/kernel/perfmon (for debugging purposes)
6689	 */
6690	pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root);
6691
6692	/*
6693	 * initialize all our spinlocks
6694	 */
6695	spin_lock_init(&pfm_sessions.pfs_lock);
6696	spin_lock_init(&pfm_buffer_fmt_lock);
6697
6698	init_pfm_fs();
6699
6700	for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6701
6702	return 0;
6703}
6704
6705__initcall(pfm_init);
6706
6707/*
6708 * this function is called before pfm_init()
6709 */
6710void
6711pfm_init_percpu (void)
6712{
6713	static int first_time=1;
6714	/*
6715	 * make sure no measurement is active
6716	 * (may inherit programmed PMCs from EFI).
6717	 */
6718	pfm_clear_psr_pp();
6719	pfm_clear_psr_up();
6720
6721	/*
6722	 * we run with the PMU not frozen at all times
6723	 */
6724	pfm_unfreeze_pmu();
6725
6726	if (first_time) {
6727		register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
 
6728		first_time=0;
6729	}
6730
6731	ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6732	ia64_srlz_d();
6733}
6734
6735/*
6736 * used for debug purposes only
6737 */
6738void
6739dump_pmu_state(const char *from)
6740{
6741	struct task_struct *task;
6742	struct pt_regs *regs;
6743	pfm_context_t *ctx;
6744	unsigned long psr, dcr, info, flags;
6745	int i, this_cpu;
6746
6747	local_irq_save(flags);
6748
6749	this_cpu = smp_processor_id();
6750	regs     = task_pt_regs(current);
6751	info     = PFM_CPUINFO_GET();
6752	dcr      = ia64_getreg(_IA64_REG_CR_DCR);
6753
6754	if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6755		local_irq_restore(flags);
6756		return;
6757	}
6758
6759	printk("CPU%d from %s() current [%d] iip=0x%lx %s\n", 
6760		this_cpu, 
6761		from, 
6762		task_pid_nr(current),
6763		regs->cr_iip,
6764		current->comm);
6765
6766	task = GET_PMU_OWNER();
6767	ctx  = GET_PMU_CTX();
6768
6769	printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task_pid_nr(task) : -1, ctx);
6770
6771	psr = pfm_get_psr();
6772
6773	printk("->CPU%d pmc0=0x%lx psr.pp=%d psr.up=%d dcr.pp=%d syst_info=0x%lx user_psr.up=%d user_psr.pp=%d\n", 
6774		this_cpu,
6775		ia64_get_pmc(0),
6776		psr & IA64_PSR_PP ? 1 : 0,
6777		psr & IA64_PSR_UP ? 1 : 0,
6778		dcr & IA64_DCR_PP ? 1 : 0,
6779		info,
6780		ia64_psr(regs)->up,
6781		ia64_psr(regs)->pp);
6782
6783	ia64_psr(regs)->up = 0;
6784	ia64_psr(regs)->pp = 0;
6785
6786	for (i=1; PMC_IS_LAST(i) == 0; i++) {
6787		if (PMC_IS_IMPL(i) == 0) continue;
6788		printk("->CPU%d pmc[%d]=0x%lx thread_pmc[%d]=0x%lx\n", this_cpu, i, ia64_get_pmc(i), i, ctx->th_pmcs[i]);
6789	}
6790
6791	for (i=1; PMD_IS_LAST(i) == 0; i++) {
6792		if (PMD_IS_IMPL(i) == 0) continue;
6793		printk("->CPU%d pmd[%d]=0x%lx thread_pmd[%d]=0x%lx\n", this_cpu, i, ia64_get_pmd(i), i, ctx->th_pmds[i]);
6794	}
6795
6796	if (ctx) {
6797		printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6798				this_cpu,
6799				ctx->ctx_state,
6800				ctx->ctx_smpl_vaddr,
6801				ctx->ctx_smpl_hdr,
6802				ctx->ctx_msgq_head,
6803				ctx->ctx_msgq_tail,
6804				ctx->ctx_saved_psr_up);
6805	}
6806	local_irq_restore(flags);
6807}
6808
6809/*
6810 * called from process.c:copy_thread(). task is new child.
6811 */
6812void
6813pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6814{
6815	struct thread_struct *thread;
6816
6817	DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task_pid_nr(task)));
6818
6819	thread = &task->thread;
6820
6821	/*
6822	 * cut links inherited from parent (current)
6823	 */
6824	thread->pfm_context = NULL;
6825
6826	PFM_SET_WORK_PENDING(task, 0);
6827
6828	/*
6829	 * the psr bits are already set properly in copy_threads()
6830	 */
6831}
6832#else  /* !CONFIG_PERFMON */
6833asmlinkage long
6834sys_perfmonctl (int fd, int cmd, void *arg, int count)
6835{
6836	return -ENOSYS;
6837}
6838#endif /* CONFIG_PERFMON */
v5.9
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * This file implements the perfmon-2 subsystem which is used
   4 * to program the IA-64 Performance Monitoring Unit (PMU).
   5 *
   6 * The initial version of perfmon.c was written by
   7 * Ganesh Venkitachalam, IBM Corp.
   8 *
   9 * Then it was modified for perfmon-1.x by Stephane Eranian and
  10 * David Mosberger, Hewlett Packard Co.
  11 *
  12 * Version Perfmon-2.x is a rewrite of perfmon-1.x
  13 * by Stephane Eranian, Hewlett Packard Co.
  14 *
  15 * Copyright (C) 1999-2005  Hewlett Packard Co
  16 *               Stephane Eranian <eranian@hpl.hp.com>
  17 *               David Mosberger-Tang <davidm@hpl.hp.com>
  18 *
  19 * More information about perfmon available at:
  20 * 	http://www.hpl.hp.com/research/linux/perfmon
  21 */
  22
  23#include <linux/module.h>
  24#include <linux/kernel.h>
  25#include <linux/sched.h>
  26#include <linux/sched/task.h>
  27#include <linux/sched/task_stack.h>
  28#include <linux/interrupt.h>
  29#include <linux/proc_fs.h>
  30#include <linux/seq_file.h>
  31#include <linux/init.h>
  32#include <linux/vmalloc.h>
  33#include <linux/mm.h>
  34#include <linux/sysctl.h>
  35#include <linux/list.h>
  36#include <linux/file.h>
  37#include <linux/poll.h>
  38#include <linux/vfs.h>
  39#include <linux/smp.h>
  40#include <linux/pagemap.h>
  41#include <linux/mount.h>
  42#include <linux/pseudo_fs.h>
  43#include <linux/bitops.h>
  44#include <linux/capability.h>
  45#include <linux/rcupdate.h>
  46#include <linux/completion.h>
  47#include <linux/tracehook.h>
  48#include <linux/slab.h>
  49#include <linux/cpu.h>
  50
  51#include <asm/errno.h>
  52#include <asm/intrinsics.h>
  53#include <asm/page.h>
  54#include <asm/perfmon.h>
  55#include <asm/processor.h>
  56#include <asm/signal.h>
  57#include <linux/uaccess.h>
 
  58#include <asm/delay.h>
  59
  60#include "irq.h"
  61
  62#ifdef CONFIG_PERFMON
  63/*
  64 * perfmon context state
  65 */
  66#define PFM_CTX_UNLOADED	1	/* context is not loaded onto any task */
  67#define PFM_CTX_LOADED		2	/* context is loaded onto a task */
  68#define PFM_CTX_MASKED		3	/* context is loaded but monitoring is masked due to overflow */
  69#define PFM_CTX_ZOMBIE		4	/* owner of the context is closing it */
  70
  71#define PFM_INVALID_ACTIVATION	(~0UL)
  72
  73#define PFM_NUM_PMC_REGS	64	/* PMC save area for ctxsw */
  74#define PFM_NUM_PMD_REGS	64	/* PMD save area for ctxsw */
  75
  76/*
  77 * depth of message queue
  78 */
  79#define PFM_MAX_MSGS		32
  80#define PFM_CTXQ_EMPTY(g)	((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
  81
  82/*
  83 * type of a PMU register (bitmask).
  84 * bitmask structure:
  85 * 	bit0   : register implemented
  86 * 	bit1   : end marker
  87 * 	bit2-3 : reserved
  88 * 	bit4   : pmc has pmc.pm
  89 * 	bit5   : pmc controls a counter (has pmc.oi), pmd is used as counter
  90 * 	bit6-7 : register type
  91 * 	bit8-31: reserved
  92 */
  93#define PFM_REG_NOTIMPL		0x0 /* not implemented at all */
  94#define PFM_REG_IMPL		0x1 /* register implemented */
  95#define PFM_REG_END		0x2 /* end marker */
  96#define PFM_REG_MONITOR		(0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
  97#define PFM_REG_COUNTING	(0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
  98#define PFM_REG_CONTROL		(0x4<<4|PFM_REG_IMPL) /* PMU control register */
  99#define	PFM_REG_CONFIG		(0x8<<4|PFM_REG_IMPL) /* configuration register */
 100#define PFM_REG_BUFFER	 	(0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
 101
 102#define PMC_IS_LAST(i)	(pmu_conf->pmc_desc[i].type & PFM_REG_END)
 103#define PMD_IS_LAST(i)	(pmu_conf->pmd_desc[i].type & PFM_REG_END)
 104
 105#define PMC_OVFL_NOTIFY(ctx, i)	((ctx)->ctx_pmds[i].flags &  PFM_REGFL_OVFL_NOTIFY)
 106
 107/* i assumed unsigned */
 108#define PMC_IS_IMPL(i)	  (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
 109#define PMD_IS_IMPL(i)	  (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
 110
 111/* XXX: these assume that register i is implemented */
 112#define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
 113#define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
 114#define PMC_IS_MONITOR(i)  ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR)  == PFM_REG_MONITOR)
 115#define PMC_IS_CONTROL(i)  ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL)  == PFM_REG_CONTROL)
 116
 117#define PMC_DFL_VAL(i)     pmu_conf->pmc_desc[i].default_value
 118#define PMC_RSVD_MASK(i)   pmu_conf->pmc_desc[i].reserved_mask
 119#define PMD_PMD_DEP(i)	   pmu_conf->pmd_desc[i].dep_pmd[0]
 120#define PMC_PMD_DEP(i)	   pmu_conf->pmc_desc[i].dep_pmd[0]
 121
 122#define PFM_NUM_IBRS	  IA64_NUM_DBG_REGS
 123#define PFM_NUM_DBRS	  IA64_NUM_DBG_REGS
 124
 125#define CTX_OVFL_NOBLOCK(c)	((c)->ctx_fl_block == 0)
 126#define CTX_HAS_SMPL(c)		((c)->ctx_fl_is_sampling)
 127#define PFM_CTX_TASK(h)		(h)->ctx_task
 128
 129#define PMU_PMC_OI		5 /* position of pmc.oi bit */
 130
 131/* XXX: does not support more than 64 PMDs */
 132#define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
 133#define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
 134
 135#define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
 136
 137#define CTX_USED_IBR(ctx,n) 	(ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
 138#define CTX_USED_DBR(ctx,n) 	(ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
 139#define CTX_USES_DBREGS(ctx)	(((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
 140#define PFM_CODE_RR	0	/* requesting code range restriction */
 141#define PFM_DATA_RR	1	/* requestion data range restriction */
 142
 143#define PFM_CPUINFO_CLEAR(v)	pfm_get_cpu_var(pfm_syst_info) &= ~(v)
 144#define PFM_CPUINFO_SET(v)	pfm_get_cpu_var(pfm_syst_info) |= (v)
 145#define PFM_CPUINFO_GET()	pfm_get_cpu_var(pfm_syst_info)
 146
 147#define RDEP(x)	(1UL<<(x))
 148
 149/*
 150 * context protection macros
 151 * in SMP:
 152 * 	- we need to protect against CPU concurrency (spin_lock)
 153 * 	- we need to protect against PMU overflow interrupts (local_irq_disable)
 154 * in UP:
 155 * 	- we need to protect against PMU overflow interrupts (local_irq_disable)
 156 *
 157 * spin_lock_irqsave()/spin_unlock_irqrestore():
 158 * 	in SMP: local_irq_disable + spin_lock
 159 * 	in UP : local_irq_disable
 160 *
 161 * spin_lock()/spin_lock():
 162 * 	in UP : removed automatically
 163 * 	in SMP: protect against context accesses from other CPU. interrupts
 164 * 	        are not masked. This is useful for the PMU interrupt handler
 165 * 	        because we know we will not get PMU concurrency in that code.
 166 */
 167#define PROTECT_CTX(c, f) \
 168	do {  \
 169		DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, task_pid_nr(current))); \
 170		spin_lock_irqsave(&(c)->ctx_lock, f); \
 171		DPRINT(("spinlocked ctx %p  by [%d]\n", c, task_pid_nr(current))); \
 172	} while(0)
 173
 174#define UNPROTECT_CTX(c, f) \
 175	do { \
 176		DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, task_pid_nr(current))); \
 177		spin_unlock_irqrestore(&(c)->ctx_lock, f); \
 178	} while(0)
 179
 180#define PROTECT_CTX_NOPRINT(c, f) \
 181	do {  \
 182		spin_lock_irqsave(&(c)->ctx_lock, f); \
 183	} while(0)
 184
 185
 186#define UNPROTECT_CTX_NOPRINT(c, f) \
 187	do { \
 188		spin_unlock_irqrestore(&(c)->ctx_lock, f); \
 189	} while(0)
 190
 191
 192#define PROTECT_CTX_NOIRQ(c) \
 193	do {  \
 194		spin_lock(&(c)->ctx_lock); \
 195	} while(0)
 196
 197#define UNPROTECT_CTX_NOIRQ(c) \
 198	do { \
 199		spin_unlock(&(c)->ctx_lock); \
 200	} while(0)
 201
 202
 203#ifdef CONFIG_SMP
 204
 205#define GET_ACTIVATION()	pfm_get_cpu_var(pmu_activation_number)
 206#define INC_ACTIVATION()	pfm_get_cpu_var(pmu_activation_number)++
 207#define SET_ACTIVATION(c)	(c)->ctx_last_activation = GET_ACTIVATION()
 208
 209#else /* !CONFIG_SMP */
 210#define SET_ACTIVATION(t) 	do {} while(0)
 211#define GET_ACTIVATION(t) 	do {} while(0)
 212#define INC_ACTIVATION(t) 	do {} while(0)
 213#endif /* CONFIG_SMP */
 214
 215#define SET_PMU_OWNER(t, c)	do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
 216#define GET_PMU_OWNER()		pfm_get_cpu_var(pmu_owner)
 217#define GET_PMU_CTX()		pfm_get_cpu_var(pmu_ctx)
 218
 219#define LOCK_PFS(g)	    	spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
 220#define UNLOCK_PFS(g)	    	spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
 221
 222#define PFM_REG_RETFLAG_SET(flags, val)	do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
 223
 224/*
 225 * cmp0 must be the value of pmc0
 226 */
 227#define PMC0_HAS_OVFL(cmp0)  (cmp0 & ~0x1UL)
 228
 229#define PFMFS_MAGIC 0xa0b4d889
 230
 231/*
 232 * debugging
 233 */
 234#define PFM_DEBUGGING 1
 235#ifdef PFM_DEBUGGING
 236#define DPRINT(a) \
 237	do { \
 238		if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
 239	} while (0)
 240
 241#define DPRINT_ovfl(a) \
 242	do { \
 243		if (unlikely(pfm_sysctl.debug > 0 && pfm_sysctl.debug_ovfl >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
 244	} while (0)
 245#endif
 246
 247/*
 248 * 64-bit software counter structure
 249 *
 250 * the next_reset_type is applied to the next call to pfm_reset_regs()
 251 */
 252typedef struct {
 253	unsigned long	val;		/* virtual 64bit counter value */
 254	unsigned long	lval;		/* last reset value */
 255	unsigned long	long_reset;	/* reset value on sampling overflow */
 256	unsigned long	short_reset;    /* reset value on overflow */
 257	unsigned long	reset_pmds[4];  /* which other pmds to reset when this counter overflows */
 258	unsigned long	smpl_pmds[4];   /* which pmds are accessed when counter overflow */
 259	unsigned long	seed;		/* seed for random-number generator */
 260	unsigned long	mask;		/* mask for random-number generator */
 261	unsigned int 	flags;		/* notify/do not notify */
 262	unsigned long	eventid;	/* overflow event identifier */
 263} pfm_counter_t;
 264
 265/*
 266 * context flags
 267 */
 268typedef struct {
 269	unsigned int block:1;		/* when 1, task will blocked on user notifications */
 270	unsigned int system:1;		/* do system wide monitoring */
 271	unsigned int using_dbreg:1;	/* using range restrictions (debug registers) */
 272	unsigned int is_sampling:1;	/* true if using a custom format */
 273	unsigned int excl_idle:1;	/* exclude idle task in system wide session */
 274	unsigned int going_zombie:1;	/* context is zombie (MASKED+blocking) */
 275	unsigned int trap_reason:2;	/* reason for going into pfm_handle_work() */
 276	unsigned int no_msg:1;		/* no message sent on overflow */
 277	unsigned int can_restart:1;	/* allowed to issue a PFM_RESTART */
 278	unsigned int reserved:22;
 279} pfm_context_flags_t;
 280
 281#define PFM_TRAP_REASON_NONE		0x0	/* default value */
 282#define PFM_TRAP_REASON_BLOCK		0x1	/* we need to block on overflow */
 283#define PFM_TRAP_REASON_RESET		0x2	/* we need to reset PMDs */
 284
 285
 286/*
 287 * perfmon context: encapsulates all the state of a monitoring session
 288 */
 289
 290typedef struct pfm_context {
 291	spinlock_t		ctx_lock;		/* context protection */
 292
 293	pfm_context_flags_t	ctx_flags;		/* bitmask of flags  (block reason incl.) */
 294	unsigned int		ctx_state;		/* state: active/inactive (no bitfield) */
 295
 296	struct task_struct 	*ctx_task;		/* task to which context is attached */
 297
 298	unsigned long		ctx_ovfl_regs[4];	/* which registers overflowed (notification) */
 299
 300	struct completion	ctx_restart_done;  	/* use for blocking notification mode */
 301
 302	unsigned long		ctx_used_pmds[4];	/* bitmask of PMD used            */
 303	unsigned long		ctx_all_pmds[4];	/* bitmask of all accessible PMDs */
 304	unsigned long		ctx_reload_pmds[4];	/* bitmask of force reload PMD on ctxsw in */
 305
 306	unsigned long		ctx_all_pmcs[4];	/* bitmask of all accessible PMCs */
 307	unsigned long		ctx_reload_pmcs[4];	/* bitmask of force reload PMC on ctxsw in */
 308	unsigned long		ctx_used_monitors[4];	/* bitmask of monitor PMC being used */
 309
 310	unsigned long		ctx_pmcs[PFM_NUM_PMC_REGS];	/*  saved copies of PMC values */
 311
 312	unsigned int		ctx_used_ibrs[1];		/* bitmask of used IBR (speedup ctxsw in) */
 313	unsigned int		ctx_used_dbrs[1];		/* bitmask of used DBR (speedup ctxsw in) */
 314	unsigned long		ctx_dbrs[IA64_NUM_DBG_REGS];	/* DBR values (cache) when not loaded */
 315	unsigned long		ctx_ibrs[IA64_NUM_DBG_REGS];	/* IBR values (cache) when not loaded */
 316
 317	pfm_counter_t		ctx_pmds[PFM_NUM_PMD_REGS]; /* software state for PMDS */
 318
 319	unsigned long		th_pmcs[PFM_NUM_PMC_REGS];	/* PMC thread save state */
 320	unsigned long		th_pmds[PFM_NUM_PMD_REGS];	/* PMD thread save state */
 321
 322	unsigned long		ctx_saved_psr_up;	/* only contains psr.up value */
 323
 324	unsigned long		ctx_last_activation;	/* context last activation number for last_cpu */
 325	unsigned int		ctx_last_cpu;		/* CPU id of current or last CPU used (SMP only) */
 326	unsigned int		ctx_cpu;		/* cpu to which perfmon is applied (system wide) */
 327
 328	int			ctx_fd;			/* file descriptor used my this context */
 329	pfm_ovfl_arg_t		ctx_ovfl_arg;		/* argument to custom buffer format handler */
 330
 331	pfm_buffer_fmt_t	*ctx_buf_fmt;		/* buffer format callbacks */
 332	void			*ctx_smpl_hdr;		/* points to sampling buffer header kernel vaddr */
 333	unsigned long		ctx_smpl_size;		/* size of sampling buffer */
 334	void			*ctx_smpl_vaddr;	/* user level virtual address of smpl buffer */
 335
 336	wait_queue_head_t 	ctx_msgq_wait;
 337	pfm_msg_t		ctx_msgq[PFM_MAX_MSGS];
 338	int			ctx_msgq_head;
 339	int			ctx_msgq_tail;
 340	struct fasync_struct	*ctx_async_queue;
 341
 342	wait_queue_head_t 	ctx_zombieq;		/* termination cleanup wait queue */
 343} pfm_context_t;
 344
 345/*
 346 * magic number used to verify that structure is really
 347 * a perfmon context
 348 */
 349#define PFM_IS_FILE(f)		((f)->f_op == &pfm_file_ops)
 350
 351#define PFM_GET_CTX(t)	 	((pfm_context_t *)(t)->thread.pfm_context)
 352
 353#ifdef CONFIG_SMP
 354#define SET_LAST_CPU(ctx, v)	(ctx)->ctx_last_cpu = (v)
 355#define GET_LAST_CPU(ctx)	(ctx)->ctx_last_cpu
 356#else
 357#define SET_LAST_CPU(ctx, v)	do {} while(0)
 358#define GET_LAST_CPU(ctx)	do {} while(0)
 359#endif
 360
 361
 362#define ctx_fl_block		ctx_flags.block
 363#define ctx_fl_system		ctx_flags.system
 364#define ctx_fl_using_dbreg	ctx_flags.using_dbreg
 365#define ctx_fl_is_sampling	ctx_flags.is_sampling
 366#define ctx_fl_excl_idle	ctx_flags.excl_idle
 367#define ctx_fl_going_zombie	ctx_flags.going_zombie
 368#define ctx_fl_trap_reason	ctx_flags.trap_reason
 369#define ctx_fl_no_msg		ctx_flags.no_msg
 370#define ctx_fl_can_restart	ctx_flags.can_restart
 371
 372#define PFM_SET_WORK_PENDING(t, v)	do { (t)->thread.pfm_needs_checking = v; } while(0);
 373#define PFM_GET_WORK_PENDING(t)		(t)->thread.pfm_needs_checking
 374
 375/*
 376 * global information about all sessions
 377 * mostly used to synchronize between system wide and per-process
 378 */
 379typedef struct {
 380	spinlock_t		pfs_lock;		   /* lock the structure */
 381
 382	unsigned int		pfs_task_sessions;	   /* number of per task sessions */
 383	unsigned int		pfs_sys_sessions;	   /* number of per system wide sessions */
 384	unsigned int		pfs_sys_use_dbregs;	   /* incremented when a system wide session uses debug regs */
 385	unsigned int		pfs_ptrace_use_dbregs;	   /* incremented when a process uses debug regs */
 386	struct task_struct	*pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
 387} pfm_session_t;
 388
 389/*
 390 * information about a PMC or PMD.
 391 * dep_pmd[]: a bitmask of dependent PMD registers
 392 * dep_pmc[]: a bitmask of dependent PMC registers
 393 */
 394typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
 395typedef struct {
 396	unsigned int		type;
 397	int			pm_pos;
 398	unsigned long		default_value;	/* power-on default value */
 399	unsigned long		reserved_mask;	/* bitmask of reserved bits */
 400	pfm_reg_check_t		read_check;
 401	pfm_reg_check_t		write_check;
 402	unsigned long		dep_pmd[4];
 403	unsigned long		dep_pmc[4];
 404} pfm_reg_desc_t;
 405
 406/* assume cnum is a valid monitor */
 407#define PMC_PM(cnum, val)	(((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
 408
 409/*
 410 * This structure is initialized at boot time and contains
 411 * a description of the PMU main characteristics.
 412 *
 413 * If the probe function is defined, detection is based
 414 * on its return value: 
 415 * 	- 0 means recognized PMU
 416 * 	- anything else means not supported
 417 * When the probe function is not defined, then the pmu_family field
 418 * is used and it must match the host CPU family such that:
 419 * 	- cpu->family & config->pmu_family != 0
 420 */
 421typedef struct {
 422	unsigned long  ovfl_val;	/* overflow value for counters */
 423
 424	pfm_reg_desc_t *pmc_desc;	/* detailed PMC register dependencies descriptions */
 425	pfm_reg_desc_t *pmd_desc;	/* detailed PMD register dependencies descriptions */
 426
 427	unsigned int   num_pmcs;	/* number of PMCS: computed at init time */
 428	unsigned int   num_pmds;	/* number of PMDS: computed at init time */
 429	unsigned long  impl_pmcs[4];	/* bitmask of implemented PMCS */
 430	unsigned long  impl_pmds[4];	/* bitmask of implemented PMDS */
 431
 432	char	      *pmu_name;	/* PMU family name */
 433	unsigned int  pmu_family;	/* cpuid family pattern used to identify pmu */
 434	unsigned int  flags;		/* pmu specific flags */
 435	unsigned int  num_ibrs;		/* number of IBRS: computed at init time */
 436	unsigned int  num_dbrs;		/* number of DBRS: computed at init time */
 437	unsigned int  num_counters;	/* PMC/PMD counting pairs : computed at init time */
 438	int           (*probe)(void);   /* customized probe routine */
 439	unsigned int  use_rr_dbregs:1;	/* set if debug registers used for range restriction */
 440} pmu_config_t;
 441/*
 442 * PMU specific flags
 443 */
 444#define PFM_PMU_IRQ_RESEND	1	/* PMU needs explicit IRQ resend */
 445
 446/*
 447 * debug register related type definitions
 448 */
 449typedef struct {
 450	unsigned long ibr_mask:56;
 451	unsigned long ibr_plm:4;
 452	unsigned long ibr_ig:3;
 453	unsigned long ibr_x:1;
 454} ibr_mask_reg_t;
 455
 456typedef struct {
 457	unsigned long dbr_mask:56;
 458	unsigned long dbr_plm:4;
 459	unsigned long dbr_ig:2;
 460	unsigned long dbr_w:1;
 461	unsigned long dbr_r:1;
 462} dbr_mask_reg_t;
 463
 464typedef union {
 465	unsigned long  val;
 466	ibr_mask_reg_t ibr;
 467	dbr_mask_reg_t dbr;
 468} dbreg_t;
 469
 470
 471/*
 472 * perfmon command descriptions
 473 */
 474typedef struct {
 475	int		(*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
 476	char		*cmd_name;
 477	int		cmd_flags;
 478	unsigned int	cmd_narg;
 479	size_t		cmd_argsize;
 480	int		(*cmd_getsize)(void *arg, size_t *sz);
 481} pfm_cmd_desc_t;
 482
 483#define PFM_CMD_FD		0x01	/* command requires a file descriptor */
 484#define PFM_CMD_ARG_READ	0x02	/* command must read argument(s) */
 485#define PFM_CMD_ARG_RW		0x04	/* command must read/write argument(s) */
 486#define PFM_CMD_STOP		0x08	/* command does not work on zombie context */
 487
 488
 489#define PFM_CMD_NAME(cmd)	pfm_cmd_tab[(cmd)].cmd_name
 490#define PFM_CMD_READ_ARG(cmd)	(pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
 491#define PFM_CMD_RW_ARG(cmd)	(pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
 492#define PFM_CMD_USE_FD(cmd)	(pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
 493#define PFM_CMD_STOPPED(cmd)	(pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
 494
 495#define PFM_CMD_ARG_MANY	-1 /* cannot be zero */
 496
 497typedef struct {
 498	unsigned long pfm_spurious_ovfl_intr_count;	/* keep track of spurious ovfl interrupts */
 499	unsigned long pfm_replay_ovfl_intr_count;	/* keep track of replayed ovfl interrupts */
 500	unsigned long pfm_ovfl_intr_count; 		/* keep track of ovfl interrupts */
 501	unsigned long pfm_ovfl_intr_cycles;		/* cycles spent processing ovfl interrupts */
 502	unsigned long pfm_ovfl_intr_cycles_min;		/* min cycles spent processing ovfl interrupts */
 503	unsigned long pfm_ovfl_intr_cycles_max;		/* max cycles spent processing ovfl interrupts */
 504	unsigned long pfm_smpl_handler_calls;
 505	unsigned long pfm_smpl_handler_cycles;
 506	char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
 507} pfm_stats_t;
 508
 509/*
 510 * perfmon internal variables
 511 */
 512static pfm_stats_t		pfm_stats[NR_CPUS];
 513static pfm_session_t		pfm_sessions;	/* global sessions information */
 514
 515static DEFINE_SPINLOCK(pfm_alt_install_check);
 516static pfm_intr_handler_desc_t  *pfm_alt_intr_handler;
 517
 518static struct proc_dir_entry 	*perfmon_dir;
 519static pfm_uuid_t		pfm_null_uuid = {0,};
 520
 521static spinlock_t		pfm_buffer_fmt_lock;
 522static LIST_HEAD(pfm_buffer_fmt_list);
 523
 524static pmu_config_t		*pmu_conf;
 525
 526/* sysctl() controls */
 527pfm_sysctl_t pfm_sysctl;
 528EXPORT_SYMBOL(pfm_sysctl);
 529
 530static struct ctl_table pfm_ctl_table[] = {
 531	{
 532		.procname	= "debug",
 533		.data		= &pfm_sysctl.debug,
 534		.maxlen		= sizeof(int),
 535		.mode		= 0666,
 536		.proc_handler	= proc_dointvec,
 537	},
 538	{
 539		.procname	= "debug_ovfl",
 540		.data		= &pfm_sysctl.debug_ovfl,
 541		.maxlen		= sizeof(int),
 542		.mode		= 0666,
 543		.proc_handler	= proc_dointvec,
 544	},
 545	{
 546		.procname	= "fastctxsw",
 547		.data		= &pfm_sysctl.fastctxsw,
 548		.maxlen		= sizeof(int),
 549		.mode		= 0600,
 550		.proc_handler	= proc_dointvec,
 551	},
 552	{
 553		.procname	= "expert_mode",
 554		.data		= &pfm_sysctl.expert_mode,
 555		.maxlen		= sizeof(int),
 556		.mode		= 0600,
 557		.proc_handler	= proc_dointvec,
 558	},
 559	{}
 560};
 561static struct ctl_table pfm_sysctl_dir[] = {
 562	{
 563		.procname	= "perfmon",
 564		.mode		= 0555,
 565		.child		= pfm_ctl_table,
 566	},
 567 	{}
 568};
 569static struct ctl_table pfm_sysctl_root[] = {
 570	{
 571		.procname	= "kernel",
 572		.mode		= 0555,
 573		.child		= pfm_sysctl_dir,
 574	},
 575 	{}
 576};
 577static struct ctl_table_header *pfm_sysctl_header;
 578
 579static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
 580
 581#define pfm_get_cpu_var(v)		__ia64_per_cpu_var(v)
 582#define pfm_get_cpu_data(a,b)		per_cpu(a, b)
 583
 584static inline void
 585pfm_put_task(struct task_struct *task)
 586{
 587	if (task != current) put_task_struct(task);
 588}
 589
 
 
 
 
 
 
 
 
 
 
 
 590static inline unsigned long
 591pfm_protect_ctx_ctxsw(pfm_context_t *x)
 592{
 593	spin_lock(&(x)->ctx_lock);
 594	return 0UL;
 595}
 596
 597static inline void
 598pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
 599{
 600	spin_unlock(&(x)->ctx_lock);
 601}
 602
 
 
 
 
 
 
 
 
 
 
 
 
 603/* forward declaration */
 604static const struct dentry_operations pfmfs_dentry_operations;
 605
 606static int pfmfs_init_fs_context(struct fs_context *fc)
 
 607{
 608	struct pseudo_fs_context *ctx = init_pseudo(fc, PFMFS_MAGIC);
 609	if (!ctx)
 610		return -ENOMEM;
 611	ctx->dops = &pfmfs_dentry_operations;
 612	return 0;
 613}
 614
 615static struct file_system_type pfm_fs_type = {
 616	.name			= "pfmfs",
 617	.init_fs_context	= pfmfs_init_fs_context,
 618	.kill_sb		= kill_anon_super,
 619};
 620MODULE_ALIAS_FS("pfmfs");
 621
 622DEFINE_PER_CPU(unsigned long, pfm_syst_info);
 623DEFINE_PER_CPU(struct task_struct *, pmu_owner);
 624DEFINE_PER_CPU(pfm_context_t  *, pmu_ctx);
 625DEFINE_PER_CPU(unsigned long, pmu_activation_number);
 626EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info);
 627
 628
 629/* forward declaration */
 630static const struct file_operations pfm_file_ops;
 631
 632/*
 633 * forward declarations
 634 */
 635#ifndef CONFIG_SMP
 636static void pfm_lazy_save_regs (struct task_struct *ta);
 637#endif
 638
 639void dump_pmu_state(const char *);
 640static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
 641
 642#include "perfmon_itanium.h"
 643#include "perfmon_mckinley.h"
 644#include "perfmon_montecito.h"
 645#include "perfmon_generic.h"
 646
 647static pmu_config_t *pmu_confs[]={
 648	&pmu_conf_mont,
 649	&pmu_conf_mck,
 650	&pmu_conf_ita,
 651	&pmu_conf_gen, /* must be last */
 652	NULL
 653};
 654
 655
 656static int pfm_end_notify_user(pfm_context_t *ctx);
 657
 658static inline void
 659pfm_clear_psr_pp(void)
 660{
 661	ia64_rsm(IA64_PSR_PP);
 662	ia64_srlz_i();
 663}
 664
 665static inline void
 666pfm_set_psr_pp(void)
 667{
 668	ia64_ssm(IA64_PSR_PP);
 669	ia64_srlz_i();
 670}
 671
 672static inline void
 673pfm_clear_psr_up(void)
 674{
 675	ia64_rsm(IA64_PSR_UP);
 676	ia64_srlz_i();
 677}
 678
 679static inline void
 680pfm_set_psr_up(void)
 681{
 682	ia64_ssm(IA64_PSR_UP);
 683	ia64_srlz_i();
 684}
 685
 686static inline unsigned long
 687pfm_get_psr(void)
 688{
 689	unsigned long tmp;
 690	tmp = ia64_getreg(_IA64_REG_PSR);
 691	ia64_srlz_i();
 692	return tmp;
 693}
 694
 695static inline void
 696pfm_set_psr_l(unsigned long val)
 697{
 698	ia64_setreg(_IA64_REG_PSR_L, val);
 699	ia64_srlz_i();
 700}
 701
 702static inline void
 703pfm_freeze_pmu(void)
 704{
 705	ia64_set_pmc(0,1UL);
 706	ia64_srlz_d();
 707}
 708
 709static inline void
 710pfm_unfreeze_pmu(void)
 711{
 712	ia64_set_pmc(0,0UL);
 713	ia64_srlz_d();
 714}
 715
 716static inline void
 717pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
 718{
 719	int i;
 720
 721	for (i=0; i < nibrs; i++) {
 722		ia64_set_ibr(i, ibrs[i]);
 723		ia64_dv_serialize_instruction();
 724	}
 725	ia64_srlz_i();
 726}
 727
 728static inline void
 729pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
 730{
 731	int i;
 732
 733	for (i=0; i < ndbrs; i++) {
 734		ia64_set_dbr(i, dbrs[i]);
 735		ia64_dv_serialize_data();
 736	}
 737	ia64_srlz_d();
 738}
 739
 740/*
 741 * PMD[i] must be a counter. no check is made
 742 */
 743static inline unsigned long
 744pfm_read_soft_counter(pfm_context_t *ctx, int i)
 745{
 746	return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
 747}
 748
 749/*
 750 * PMD[i] must be a counter. no check is made
 751 */
 752static inline void
 753pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
 754{
 755	unsigned long ovfl_val = pmu_conf->ovfl_val;
 756
 757	ctx->ctx_pmds[i].val = val  & ~ovfl_val;
 758	/*
 759	 * writing to unimplemented part is ignore, so we do not need to
 760	 * mask off top part
 761	 */
 762	ia64_set_pmd(i, val & ovfl_val);
 763}
 764
 765static pfm_msg_t *
 766pfm_get_new_msg(pfm_context_t *ctx)
 767{
 768	int idx, next;
 769
 770	next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
 771
 772	DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
 773	if (next == ctx->ctx_msgq_head) return NULL;
 774
 775 	idx = 	ctx->ctx_msgq_tail;
 776	ctx->ctx_msgq_tail = next;
 777
 778	DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
 779
 780	return ctx->ctx_msgq+idx;
 781}
 782
 783static pfm_msg_t *
 784pfm_get_next_msg(pfm_context_t *ctx)
 785{
 786	pfm_msg_t *msg;
 787
 788	DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
 789
 790	if (PFM_CTXQ_EMPTY(ctx)) return NULL;
 791
 792	/*
 793	 * get oldest message
 794	 */
 795	msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
 796
 797	/*
 798	 * and move forward
 799	 */
 800	ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
 801
 802	DPRINT(("ctx=%p head=%d tail=%d type=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, msg->pfm_gen_msg.msg_type));
 803
 804	return msg;
 805}
 806
 807static void
 808pfm_reset_msgq(pfm_context_t *ctx)
 809{
 810	ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
 811	DPRINT(("ctx=%p msgq reset\n", ctx));
 812}
 813
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 814static pfm_context_t *
 815pfm_context_alloc(int ctx_flags)
 816{
 817	pfm_context_t *ctx;
 818
 819	/* 
 820	 * allocate context descriptor 
 821	 * must be able to free with interrupts disabled
 822	 */
 823	ctx = kzalloc(sizeof(pfm_context_t), GFP_KERNEL);
 824	if (ctx) {
 825		DPRINT(("alloc ctx @%p\n", ctx));
 826
 827		/*
 828		 * init context protection lock
 829		 */
 830		spin_lock_init(&ctx->ctx_lock);
 831
 832		/*
 833		 * context is unloaded
 834		 */
 835		ctx->ctx_state = PFM_CTX_UNLOADED;
 836
 837		/*
 838		 * initialization of context's flags
 839		 */
 840		ctx->ctx_fl_block       = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
 841		ctx->ctx_fl_system      = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
 842		ctx->ctx_fl_no_msg      = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
 843		/*
 844		 * will move to set properties
 845		 * ctx->ctx_fl_excl_idle   = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
 846		 */
 847
 848		/*
 849		 * init restart semaphore to locked
 850		 */
 851		init_completion(&ctx->ctx_restart_done);
 852
 853		/*
 854		 * activation is used in SMP only
 855		 */
 856		ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
 857		SET_LAST_CPU(ctx, -1);
 858
 859		/*
 860		 * initialize notification message queue
 861		 */
 862		ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
 863		init_waitqueue_head(&ctx->ctx_msgq_wait);
 864		init_waitqueue_head(&ctx->ctx_zombieq);
 865
 866	}
 867	return ctx;
 868}
 869
 870static void
 871pfm_context_free(pfm_context_t *ctx)
 872{
 873	if (ctx) {
 874		DPRINT(("free ctx @%p\n", ctx));
 875		kfree(ctx);
 876	}
 877}
 878
 879static void
 880pfm_mask_monitoring(struct task_struct *task)
 881{
 882	pfm_context_t *ctx = PFM_GET_CTX(task);
 883	unsigned long mask, val, ovfl_mask;
 884	int i;
 885
 886	DPRINT_ovfl(("masking monitoring for [%d]\n", task_pid_nr(task)));
 887
 888	ovfl_mask = pmu_conf->ovfl_val;
 889	/*
 890	 * monitoring can only be masked as a result of a valid
 891	 * counter overflow. In UP, it means that the PMU still
 892	 * has an owner. Note that the owner can be different
 893	 * from the current task. However the PMU state belongs
 894	 * to the owner.
 895	 * In SMP, a valid overflow only happens when task is
 896	 * current. Therefore if we come here, we know that
 897	 * the PMU state belongs to the current task, therefore
 898	 * we can access the live registers.
 899	 *
 900	 * So in both cases, the live register contains the owner's
 901	 * state. We can ONLY touch the PMU registers and NOT the PSR.
 902	 *
 903	 * As a consequence to this call, the ctx->th_pmds[] array
 904	 * contains stale information which must be ignored
 905	 * when context is reloaded AND monitoring is active (see
 906	 * pfm_restart).
 907	 */
 908	mask = ctx->ctx_used_pmds[0];
 909	for (i = 0; mask; i++, mask>>=1) {
 910		/* skip non used pmds */
 911		if ((mask & 0x1) == 0) continue;
 912		val = ia64_get_pmd(i);
 913
 914		if (PMD_IS_COUNTING(i)) {
 915			/*
 916		 	 * we rebuild the full 64 bit value of the counter
 917		 	 */
 918			ctx->ctx_pmds[i].val += (val & ovfl_mask);
 919		} else {
 920			ctx->ctx_pmds[i].val = val;
 921		}
 922		DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
 923			i,
 924			ctx->ctx_pmds[i].val,
 925			val & ovfl_mask));
 926	}
 927	/*
 928	 * mask monitoring by setting the privilege level to 0
 929	 * we cannot use psr.pp/psr.up for this, it is controlled by
 930	 * the user
 931	 *
 932	 * if task is current, modify actual registers, otherwise modify
 933	 * thread save state, i.e., what will be restored in pfm_load_regs()
 934	 */
 935	mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
 936	for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
 937		if ((mask & 0x1) == 0UL) continue;
 938		ia64_set_pmc(i, ctx->th_pmcs[i] & ~0xfUL);
 939		ctx->th_pmcs[i] &= ~0xfUL;
 940		DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
 941	}
 942	/*
 943	 * make all of this visible
 944	 */
 945	ia64_srlz_d();
 946}
 947
 948/*
 949 * must always be done with task == current
 950 *
 951 * context must be in MASKED state when calling
 952 */
 953static void
 954pfm_restore_monitoring(struct task_struct *task)
 955{
 956	pfm_context_t *ctx = PFM_GET_CTX(task);
 957	unsigned long mask, ovfl_mask;
 958	unsigned long psr, val;
 959	int i, is_system;
 960
 961	is_system = ctx->ctx_fl_system;
 962	ovfl_mask = pmu_conf->ovfl_val;
 963
 964	if (task != current) {
 965		printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task_pid_nr(task), task_pid_nr(current));
 966		return;
 967	}
 968	if (ctx->ctx_state != PFM_CTX_MASKED) {
 969		printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
 970			task_pid_nr(task), task_pid_nr(current), ctx->ctx_state);
 971		return;
 972	}
 973	psr = pfm_get_psr();
 974	/*
 975	 * monitoring is masked via the PMC.
 976	 * As we restore their value, we do not want each counter to
 977	 * restart right away. We stop monitoring using the PSR,
 978	 * restore the PMC (and PMD) and then re-establish the psr
 979	 * as it was. Note that there can be no pending overflow at
 980	 * this point, because monitoring was MASKED.
 981	 *
 982	 * system-wide session are pinned and self-monitoring
 983	 */
 984	if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
 985		/* disable dcr pp */
 986		ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
 987		pfm_clear_psr_pp();
 988	} else {
 989		pfm_clear_psr_up();
 990	}
 991	/*
 992	 * first, we restore the PMD
 993	 */
 994	mask = ctx->ctx_used_pmds[0];
 995	for (i = 0; mask; i++, mask>>=1) {
 996		/* skip non used pmds */
 997		if ((mask & 0x1) == 0) continue;
 998
 999		if (PMD_IS_COUNTING(i)) {
1000			/*
1001			 * we split the 64bit value according to
1002			 * counter width
1003			 */
1004			val = ctx->ctx_pmds[i].val & ovfl_mask;
1005			ctx->ctx_pmds[i].val &= ~ovfl_mask;
1006		} else {
1007			val = ctx->ctx_pmds[i].val;
1008		}
1009		ia64_set_pmd(i, val);
1010
1011		DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1012			i,
1013			ctx->ctx_pmds[i].val,
1014			val));
1015	}
1016	/*
1017	 * restore the PMCs
1018	 */
1019	mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1020	for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1021		if ((mask & 0x1) == 0UL) continue;
1022		ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1023		ia64_set_pmc(i, ctx->th_pmcs[i]);
1024		DPRINT(("[%d] pmc[%d]=0x%lx\n",
1025					task_pid_nr(task), i, ctx->th_pmcs[i]));
1026	}
1027	ia64_srlz_d();
1028
1029	/*
1030	 * must restore DBR/IBR because could be modified while masked
1031	 * XXX: need to optimize 
1032	 */
1033	if (ctx->ctx_fl_using_dbreg) {
1034		pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1035		pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1036	}
1037
1038	/*
1039	 * now restore PSR
1040	 */
1041	if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1042		/* enable dcr pp */
1043		ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1044		ia64_srlz_i();
1045	}
1046	pfm_set_psr_l(psr);
1047}
1048
1049static inline void
1050pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1051{
1052	int i;
1053
1054	ia64_srlz_d();
1055
1056	for (i=0; mask; i++, mask>>=1) {
1057		if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1058	}
1059}
1060
1061/*
1062 * reload from thread state (used for ctxw only)
1063 */
1064static inline void
1065pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1066{
1067	int i;
1068	unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1069
1070	for (i=0; mask; i++, mask>>=1) {
1071		if ((mask & 0x1) == 0) continue;
1072		val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1073		ia64_set_pmd(i, val);
1074	}
1075	ia64_srlz_d();
1076}
1077
1078/*
1079 * propagate PMD from context to thread-state
1080 */
1081static inline void
1082pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1083{
1084	unsigned long ovfl_val = pmu_conf->ovfl_val;
1085	unsigned long mask = ctx->ctx_all_pmds[0];
1086	unsigned long val;
1087	int i;
1088
1089	DPRINT(("mask=0x%lx\n", mask));
1090
1091	for (i=0; mask; i++, mask>>=1) {
1092
1093		val = ctx->ctx_pmds[i].val;
1094
1095		/*
1096		 * We break up the 64 bit value into 2 pieces
1097		 * the lower bits go to the machine state in the
1098		 * thread (will be reloaded on ctxsw in).
1099		 * The upper part stays in the soft-counter.
1100		 */
1101		if (PMD_IS_COUNTING(i)) {
1102			ctx->ctx_pmds[i].val = val & ~ovfl_val;
1103			 val &= ovfl_val;
1104		}
1105		ctx->th_pmds[i] = val;
1106
1107		DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1108			i,
1109			ctx->th_pmds[i],
1110			ctx->ctx_pmds[i].val));
1111	}
1112}
1113
1114/*
1115 * propagate PMC from context to thread-state
1116 */
1117static inline void
1118pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1119{
1120	unsigned long mask = ctx->ctx_all_pmcs[0];
1121	int i;
1122
1123	DPRINT(("mask=0x%lx\n", mask));
1124
1125	for (i=0; mask; i++, mask>>=1) {
1126		/* masking 0 with ovfl_val yields 0 */
1127		ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1128		DPRINT(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
1129	}
1130}
1131
1132
1133
1134static inline void
1135pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1136{
1137	int i;
1138
1139	for (i=0; mask; i++, mask>>=1) {
1140		if ((mask & 0x1) == 0) continue;
1141		ia64_set_pmc(i, pmcs[i]);
1142	}
1143	ia64_srlz_d();
1144}
1145
1146static inline int
1147pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1148{
1149	return memcmp(a, b, sizeof(pfm_uuid_t));
1150}
1151
1152static inline int
1153pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1154{
1155	int ret = 0;
1156	if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1157	return ret;
1158}
1159
1160static inline int
1161pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1162{
1163	int ret = 0;
1164	if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1165	return ret;
1166}
1167
1168
1169static inline int
1170pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1171		     int cpu, void *arg)
1172{
1173	int ret = 0;
1174	if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1175	return ret;
1176}
1177
1178static inline int
1179pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1180		     int cpu, void *arg)
1181{
1182	int ret = 0;
1183	if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1184	return ret;
1185}
1186
1187static inline int
1188pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1189{
1190	int ret = 0;
1191	if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1192	return ret;
1193}
1194
1195static inline int
1196pfm_buf_fmt_restart_active(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1197{
1198	int ret = 0;
1199	if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1200	return ret;
1201}
1202
1203static pfm_buffer_fmt_t *
1204__pfm_find_buffer_fmt(pfm_uuid_t uuid)
1205{
1206	struct list_head * pos;
1207	pfm_buffer_fmt_t * entry;
1208
1209	list_for_each(pos, &pfm_buffer_fmt_list) {
1210		entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1211		if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1212			return entry;
1213	}
1214	return NULL;
1215}
1216 
1217/*
1218 * find a buffer format based on its uuid
1219 */
1220static pfm_buffer_fmt_t *
1221pfm_find_buffer_fmt(pfm_uuid_t uuid)
1222{
1223	pfm_buffer_fmt_t * fmt;
1224	spin_lock(&pfm_buffer_fmt_lock);
1225	fmt = __pfm_find_buffer_fmt(uuid);
1226	spin_unlock(&pfm_buffer_fmt_lock);
1227	return fmt;
1228}
1229 
1230int
1231pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1232{
1233	int ret = 0;
1234
1235	/* some sanity checks */
1236	if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1237
1238	/* we need at least a handler */
1239	if (fmt->fmt_handler == NULL) return -EINVAL;
1240
1241	/*
1242	 * XXX: need check validity of fmt_arg_size
1243	 */
1244
1245	spin_lock(&pfm_buffer_fmt_lock);
1246
1247	if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1248		printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1249		ret = -EBUSY;
1250		goto out;
1251	} 
1252	list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1253	printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1254
1255out:
1256	spin_unlock(&pfm_buffer_fmt_lock);
1257 	return ret;
1258}
1259EXPORT_SYMBOL(pfm_register_buffer_fmt);
1260
1261int
1262pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1263{
1264	pfm_buffer_fmt_t *fmt;
1265	int ret = 0;
1266
1267	spin_lock(&pfm_buffer_fmt_lock);
1268
1269	fmt = __pfm_find_buffer_fmt(uuid);
1270	if (!fmt) {
1271		printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1272		ret = -EINVAL;
1273		goto out;
1274	}
1275	list_del_init(&fmt->fmt_list);
1276	printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1277
1278out:
1279	spin_unlock(&pfm_buffer_fmt_lock);
1280	return ret;
1281
1282}
1283EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1284
 
 
1285static int
1286pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1287{
1288	unsigned long flags;
1289	/*
1290	 * validity checks on cpu_mask have been done upstream
1291	 */
1292	LOCK_PFS(flags);
1293
1294	DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1295		pfm_sessions.pfs_sys_sessions,
1296		pfm_sessions.pfs_task_sessions,
1297		pfm_sessions.pfs_sys_use_dbregs,
1298		is_syswide,
1299		cpu));
1300
1301	if (is_syswide) {
1302		/*
1303		 * cannot mix system wide and per-task sessions
1304		 */
1305		if (pfm_sessions.pfs_task_sessions > 0UL) {
1306			DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1307			  	pfm_sessions.pfs_task_sessions));
1308			goto abort;
1309		}
1310
1311		if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1312
1313		DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1314
1315		pfm_sessions.pfs_sys_session[cpu] = task;
1316
1317		pfm_sessions.pfs_sys_sessions++ ;
1318
1319	} else {
1320		if (pfm_sessions.pfs_sys_sessions) goto abort;
1321		pfm_sessions.pfs_task_sessions++;
1322	}
1323
1324	DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1325		pfm_sessions.pfs_sys_sessions,
1326		pfm_sessions.pfs_task_sessions,
1327		pfm_sessions.pfs_sys_use_dbregs,
1328		is_syswide,
1329		cpu));
1330
1331	/*
1332	 * Force idle() into poll mode
1333	 */
1334	cpu_idle_poll_ctrl(true);
1335
1336	UNLOCK_PFS(flags);
1337
1338	return 0;
1339
1340error_conflict:
1341	DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1342  		task_pid_nr(pfm_sessions.pfs_sys_session[cpu]),
1343		cpu));
1344abort:
1345	UNLOCK_PFS(flags);
1346
1347	return -EBUSY;
1348
1349}
1350
1351static int
1352pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1353{
1354	unsigned long flags;
1355	/*
1356	 * validity checks on cpu_mask have been done upstream
1357	 */
1358	LOCK_PFS(flags);
1359
1360	DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1361		pfm_sessions.pfs_sys_sessions,
1362		pfm_sessions.pfs_task_sessions,
1363		pfm_sessions.pfs_sys_use_dbregs,
1364		is_syswide,
1365		cpu));
1366
1367
1368	if (is_syswide) {
1369		pfm_sessions.pfs_sys_session[cpu] = NULL;
1370		/*
1371		 * would not work with perfmon+more than one bit in cpu_mask
1372		 */
1373		if (ctx && ctx->ctx_fl_using_dbreg) {
1374			if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1375				printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1376			} else {
1377				pfm_sessions.pfs_sys_use_dbregs--;
1378			}
1379		}
1380		pfm_sessions.pfs_sys_sessions--;
1381	} else {
1382		pfm_sessions.pfs_task_sessions--;
1383	}
1384	DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1385		pfm_sessions.pfs_sys_sessions,
1386		pfm_sessions.pfs_task_sessions,
1387		pfm_sessions.pfs_sys_use_dbregs,
1388		is_syswide,
1389		cpu));
1390
1391	/* Undo forced polling. Last session reenables pal_halt */
1392	cpu_idle_poll_ctrl(false);
 
 
 
1393
1394	UNLOCK_PFS(flags);
1395
1396	return 0;
1397}
1398
1399/*
1400 * removes virtual mapping of the sampling buffer.
1401 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1402 * a PROTECT_CTX() section.
1403 */
1404static int
1405pfm_remove_smpl_mapping(void *vaddr, unsigned long size)
1406{
1407	struct task_struct *task = current;
1408	int r;
1409
1410	/* sanity checks */
1411	if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1412		printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task_pid_nr(task), task->mm);
1413		return -EINVAL;
1414	}
1415
1416	DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1417
1418	/*
1419	 * does the actual unmapping
1420	 */
1421	r = vm_munmap((unsigned long)vaddr, size);
 
 
 
 
1422
 
1423	if (r !=0) {
1424		printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task_pid_nr(task), vaddr, size);
1425	}
1426
1427	DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1428
1429	return 0;
1430}
1431
1432/*
1433 * free actual physical storage used by sampling buffer
1434 */
1435#if 0
1436static int
1437pfm_free_smpl_buffer(pfm_context_t *ctx)
1438{
1439	pfm_buffer_fmt_t *fmt;
1440
1441	if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1442
1443	/*
1444	 * we won't use the buffer format anymore
1445	 */
1446	fmt = ctx->ctx_buf_fmt;
1447
1448	DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1449		ctx->ctx_smpl_hdr,
1450		ctx->ctx_smpl_size,
1451		ctx->ctx_smpl_vaddr));
1452
1453	pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1454
1455	/*
1456	 * free the buffer
1457	 */
1458	vfree(ctx->ctx_smpl_hdr);
1459
1460	ctx->ctx_smpl_hdr  = NULL;
1461	ctx->ctx_smpl_size = 0UL;
1462
1463	return 0;
1464
1465invalid_free:
1466	printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", task_pid_nr(current));
1467	return -EINVAL;
1468}
1469#endif
1470
1471static inline void
1472pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1473{
1474	if (fmt == NULL) return;
1475
1476	pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1477
1478}
1479
1480/*
1481 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1482 * no real gain from having the whole whorehouse mounted. So we don't need
1483 * any operations on the root directory. However, we need a non-trivial
1484 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1485 */
1486static struct vfsmount *pfmfs_mnt __read_mostly;
1487
1488static int __init
1489init_pfm_fs(void)
1490{
1491	int err = register_filesystem(&pfm_fs_type);
1492	if (!err) {
1493		pfmfs_mnt = kern_mount(&pfm_fs_type);
1494		err = PTR_ERR(pfmfs_mnt);
1495		if (IS_ERR(pfmfs_mnt))
1496			unregister_filesystem(&pfm_fs_type);
1497		else
1498			err = 0;
1499	}
1500	return err;
1501}
1502
1503static ssize_t
1504pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
1505{
1506	pfm_context_t *ctx;
1507	pfm_msg_t *msg;
1508	ssize_t ret;
1509	unsigned long flags;
1510  	DECLARE_WAITQUEUE(wait, current);
1511	if (PFM_IS_FILE(filp) == 0) {
1512		printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1513		return -EINVAL;
1514	}
1515
1516	ctx = filp->private_data;
1517	if (ctx == NULL) {
1518		printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", task_pid_nr(current));
1519		return -EINVAL;
1520	}
1521
1522	/*
1523	 * check even when there is no message
1524	 */
1525	if (size < sizeof(pfm_msg_t)) {
1526		DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1527		return -EINVAL;
1528	}
1529
1530	PROTECT_CTX(ctx, flags);
1531
1532  	/*
1533	 * put ourselves on the wait queue
1534	 */
1535  	add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1536
1537
1538  	for(;;) {
1539		/*
1540		 * check wait queue
1541		 */
1542
1543  		set_current_state(TASK_INTERRUPTIBLE);
1544
1545		DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1546
1547		ret = 0;
1548		if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1549
1550		UNPROTECT_CTX(ctx, flags);
1551
1552		/*
1553		 * check non-blocking read
1554		 */
1555      		ret = -EAGAIN;
1556		if(filp->f_flags & O_NONBLOCK) break;
1557
1558		/*
1559		 * check pending signals
1560		 */
1561		if(signal_pending(current)) {
1562			ret = -EINTR;
1563			break;
1564		}
1565      		/*
1566		 * no message, so wait
1567		 */
1568      		schedule();
1569
1570		PROTECT_CTX(ctx, flags);
1571	}
1572	DPRINT(("[%d] back to running ret=%ld\n", task_pid_nr(current), ret));
1573  	set_current_state(TASK_RUNNING);
1574	remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1575
1576	if (ret < 0) goto abort;
1577
1578	ret = -EINVAL;
1579	msg = pfm_get_next_msg(ctx);
1580	if (msg == NULL) {
1581		printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, task_pid_nr(current));
1582		goto abort_locked;
1583	}
1584
1585	DPRINT(("fd=%d type=%d\n", msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1586
1587	ret = -EFAULT;
1588  	if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1589
1590abort_locked:
1591	UNPROTECT_CTX(ctx, flags);
1592abort:
1593	return ret;
1594}
1595
1596static ssize_t
1597pfm_write(struct file *file, const char __user *ubuf,
1598			  size_t size, loff_t *ppos)
1599{
1600	DPRINT(("pfm_write called\n"));
1601	return -EINVAL;
1602}
1603
1604static __poll_t
1605pfm_poll(struct file *filp, poll_table * wait)
1606{
1607	pfm_context_t *ctx;
1608	unsigned long flags;
1609	__poll_t mask = 0;
1610
1611	if (PFM_IS_FILE(filp) == 0) {
1612		printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1613		return 0;
1614	}
1615
1616	ctx = filp->private_data;
1617	if (ctx == NULL) {
1618		printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", task_pid_nr(current));
1619		return 0;
1620	}
1621
1622
1623	DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1624
1625	poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1626
1627	PROTECT_CTX(ctx, flags);
1628
1629	if (PFM_CTXQ_EMPTY(ctx) == 0)
1630		mask =  EPOLLIN | EPOLLRDNORM;
1631
1632	UNPROTECT_CTX(ctx, flags);
1633
1634	DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1635
1636	return mask;
1637}
1638
1639static long
1640pfm_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1641{
1642	DPRINT(("pfm_ioctl called\n"));
1643	return -EINVAL;
1644}
1645
1646/*
1647 * interrupt cannot be masked when coming here
1648 */
1649static inline int
1650pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1651{
1652	int ret;
1653
1654	ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1655
1656	DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1657		task_pid_nr(current),
1658		fd,
1659		on,
1660		ctx->ctx_async_queue, ret));
1661
1662	return ret;
1663}
1664
1665static int
1666pfm_fasync(int fd, struct file *filp, int on)
1667{
1668	pfm_context_t *ctx;
1669	int ret;
1670
1671	if (PFM_IS_FILE(filp) == 0) {
1672		printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", task_pid_nr(current));
1673		return -EBADF;
1674	}
1675
1676	ctx = filp->private_data;
1677	if (ctx == NULL) {
1678		printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", task_pid_nr(current));
1679		return -EBADF;
1680	}
1681	/*
1682	 * we cannot mask interrupts during this call because this may
1683	 * may go to sleep if memory is not readily avalaible.
1684	 *
1685	 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1686	 * done in caller. Serialization of this function is ensured by caller.
1687	 */
1688	ret = pfm_do_fasync(fd, filp, ctx, on);
1689
1690
1691	DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1692		fd,
1693		on,
1694		ctx->ctx_async_queue, ret));
1695
1696	return ret;
1697}
1698
1699#ifdef CONFIG_SMP
1700/*
1701 * this function is exclusively called from pfm_close().
1702 * The context is not protected at that time, nor are interrupts
1703 * on the remote CPU. That's necessary to avoid deadlocks.
1704 */
1705static void
1706pfm_syswide_force_stop(void *info)
1707{
1708	pfm_context_t   *ctx = (pfm_context_t *)info;
1709	struct pt_regs *regs = task_pt_regs(current);
1710	struct task_struct *owner;
1711	unsigned long flags;
1712	int ret;
1713
1714	if (ctx->ctx_cpu != smp_processor_id()) {
1715		printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d  but on CPU%d\n",
1716			ctx->ctx_cpu,
1717			smp_processor_id());
1718		return;
1719	}
1720	owner = GET_PMU_OWNER();
1721	if (owner != ctx->ctx_task) {
1722		printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1723			smp_processor_id(),
1724			task_pid_nr(owner), task_pid_nr(ctx->ctx_task));
1725		return;
1726	}
1727	if (GET_PMU_CTX() != ctx) {
1728		printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1729			smp_processor_id(),
1730			GET_PMU_CTX(), ctx);
1731		return;
1732	}
1733
1734	DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), task_pid_nr(ctx->ctx_task)));
1735	/*
1736	 * the context is already protected in pfm_close(), we simply
1737	 * need to mask interrupts to avoid a PMU interrupt race on
1738	 * this CPU
1739	 */
1740	local_irq_save(flags);
1741
1742	ret = pfm_context_unload(ctx, NULL, 0, regs);
1743	if (ret) {
1744		DPRINT(("context_unload returned %d\n", ret));
1745	}
1746
1747	/*
1748	 * unmask interrupts, PMU interrupts are now spurious here
1749	 */
1750	local_irq_restore(flags);
1751}
1752
1753static void
1754pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1755{
1756	int ret;
1757
1758	DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1759	ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 1);
1760	DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1761}
1762#endif /* CONFIG_SMP */
1763
1764/*
1765 * called for each close(). Partially free resources.
1766 * When caller is self-monitoring, the context is unloaded.
1767 */
1768static int
1769pfm_flush(struct file *filp, fl_owner_t id)
1770{
1771	pfm_context_t *ctx;
1772	struct task_struct *task;
1773	struct pt_regs *regs;
1774	unsigned long flags;
1775	unsigned long smpl_buf_size = 0UL;
1776	void *smpl_buf_vaddr = NULL;
1777	int state, is_system;
1778
1779	if (PFM_IS_FILE(filp) == 0) {
1780		DPRINT(("bad magic for\n"));
1781		return -EBADF;
1782	}
1783
1784	ctx = filp->private_data;
1785	if (ctx == NULL) {
1786		printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", task_pid_nr(current));
1787		return -EBADF;
1788	}
1789
1790	/*
1791	 * remove our file from the async queue, if we use this mode.
1792	 * This can be done without the context being protected. We come
1793	 * here when the context has become unreachable by other tasks.
1794	 *
1795	 * We may still have active monitoring at this point and we may
1796	 * end up in pfm_overflow_handler(). However, fasync_helper()
1797	 * operates with interrupts disabled and it cleans up the
1798	 * queue. If the PMU handler is called prior to entering
1799	 * fasync_helper() then it will send a signal. If it is
1800	 * invoked after, it will find an empty queue and no
1801	 * signal will be sent. In both case, we are safe
1802	 */
1803	PROTECT_CTX(ctx, flags);
1804
1805	state     = ctx->ctx_state;
1806	is_system = ctx->ctx_fl_system;
1807
1808	task = PFM_CTX_TASK(ctx);
1809	regs = task_pt_regs(task);
1810
1811	DPRINT(("ctx_state=%d is_current=%d\n",
1812		state,
1813		task == current ? 1 : 0));
1814
1815	/*
1816	 * if state == UNLOADED, then task is NULL
1817	 */
1818
1819	/*
1820	 * we must stop and unload because we are losing access to the context.
1821	 */
1822	if (task == current) {
1823#ifdef CONFIG_SMP
1824		/*
1825		 * the task IS the owner but it migrated to another CPU: that's bad
1826		 * but we must handle this cleanly. Unfortunately, the kernel does
1827		 * not provide a mechanism to block migration (while the context is loaded).
1828		 *
1829		 * We need to release the resource on the ORIGINAL cpu.
1830		 */
1831		if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1832
1833			DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1834			/*
1835			 * keep context protected but unmask interrupt for IPI
1836			 */
1837			local_irq_restore(flags);
1838
1839			pfm_syswide_cleanup_other_cpu(ctx);
1840
1841			/*
1842			 * restore interrupt masking
1843			 */
1844			local_irq_save(flags);
1845
1846			/*
1847			 * context is unloaded at this point
1848			 */
1849		} else
1850#endif /* CONFIG_SMP */
1851		{
1852
1853			DPRINT(("forcing unload\n"));
1854			/*
1855		 	* stop and unload, returning with state UNLOADED
1856		 	* and session unreserved.
1857		 	*/
1858			pfm_context_unload(ctx, NULL, 0, regs);
1859
1860			DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1861		}
1862	}
1863
1864	/*
1865	 * remove virtual mapping, if any, for the calling task.
1866	 * cannot reset ctx field until last user is calling close().
1867	 *
1868	 * ctx_smpl_vaddr must never be cleared because it is needed
1869	 * by every task with access to the context
1870	 *
1871	 * When called from do_exit(), the mm context is gone already, therefore
1872	 * mm is NULL, i.e., the VMA is already gone  and we do not have to
1873	 * do anything here
1874	 */
1875	if (ctx->ctx_smpl_vaddr && current->mm) {
1876		smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1877		smpl_buf_size  = ctx->ctx_smpl_size;
1878	}
1879
1880	UNPROTECT_CTX(ctx, flags);
1881
1882	/*
1883	 * if there was a mapping, then we systematically remove it
1884	 * at this point. Cannot be done inside critical section
1885	 * because some VM function reenables interrupts.
1886	 *
1887	 */
1888	if (smpl_buf_vaddr) pfm_remove_smpl_mapping(smpl_buf_vaddr, smpl_buf_size);
1889
1890	return 0;
1891}
1892/*
1893 * called either on explicit close() or from exit_files(). 
1894 * Only the LAST user of the file gets to this point, i.e., it is
1895 * called only ONCE.
1896 *
1897 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero 
1898 * (fput()),i.e, last task to access the file. Nobody else can access the 
1899 * file at this point.
1900 *
1901 * When called from exit_files(), the VMA has been freed because exit_mm()
1902 * is executed before exit_files().
1903 *
1904 * When called from exit_files(), the current task is not yet ZOMBIE but we
1905 * flush the PMU state to the context. 
1906 */
1907static int
1908pfm_close(struct inode *inode, struct file *filp)
1909{
1910	pfm_context_t *ctx;
1911	struct task_struct *task;
1912	struct pt_regs *regs;
1913  	DECLARE_WAITQUEUE(wait, current);
1914	unsigned long flags;
1915	unsigned long smpl_buf_size = 0UL;
1916	void *smpl_buf_addr = NULL;
1917	int free_possible = 1;
1918	int state, is_system;
1919
1920	DPRINT(("pfm_close called private=%p\n", filp->private_data));
1921
1922	if (PFM_IS_FILE(filp) == 0) {
1923		DPRINT(("bad magic\n"));
1924		return -EBADF;
1925	}
1926	
1927	ctx = filp->private_data;
1928	if (ctx == NULL) {
1929		printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", task_pid_nr(current));
1930		return -EBADF;
1931	}
1932
1933	PROTECT_CTX(ctx, flags);
1934
1935	state     = ctx->ctx_state;
1936	is_system = ctx->ctx_fl_system;
1937
1938	task = PFM_CTX_TASK(ctx);
1939	regs = task_pt_regs(task);
1940
1941	DPRINT(("ctx_state=%d is_current=%d\n", 
1942		state,
1943		task == current ? 1 : 0));
1944
1945	/*
1946	 * if task == current, then pfm_flush() unloaded the context
1947	 */
1948	if (state == PFM_CTX_UNLOADED) goto doit;
1949
1950	/*
1951	 * context is loaded/masked and task != current, we need to
1952	 * either force an unload or go zombie
1953	 */
1954
1955	/*
1956	 * The task is currently blocked or will block after an overflow.
1957	 * we must force it to wakeup to get out of the
1958	 * MASKED state and transition to the unloaded state by itself.
1959	 *
1960	 * This situation is only possible for per-task mode
1961	 */
1962	if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
1963
1964		/*
1965		 * set a "partial" zombie state to be checked
1966		 * upon return from down() in pfm_handle_work().
1967		 *
1968		 * We cannot use the ZOMBIE state, because it is checked
1969		 * by pfm_load_regs() which is called upon wakeup from down().
1970		 * In such case, it would free the context and then we would
1971		 * return to pfm_handle_work() which would access the
1972		 * stale context. Instead, we set a flag invisible to pfm_load_regs()
1973		 * but visible to pfm_handle_work().
1974		 *
1975		 * For some window of time, we have a zombie context with
1976		 * ctx_state = MASKED  and not ZOMBIE
1977		 */
1978		ctx->ctx_fl_going_zombie = 1;
1979
1980		/*
1981		 * force task to wake up from MASKED state
1982		 */
1983		complete(&ctx->ctx_restart_done);
1984
1985		DPRINT(("waking up ctx_state=%d\n", state));
1986
1987		/*
1988		 * put ourself to sleep waiting for the other
1989		 * task to report completion
1990		 *
1991		 * the context is protected by mutex, therefore there
1992		 * is no risk of being notified of completion before
1993		 * begin actually on the waitq.
1994		 */
1995  		set_current_state(TASK_INTERRUPTIBLE);
1996  		add_wait_queue(&ctx->ctx_zombieq, &wait);
1997
1998		UNPROTECT_CTX(ctx, flags);
1999
2000		/*
2001		 * XXX: check for signals :
2002		 * 	- ok for explicit close
2003		 * 	- not ok when coming from exit_files()
2004		 */
2005      		schedule();
2006
2007
2008		PROTECT_CTX(ctx, flags);
2009
2010
2011		remove_wait_queue(&ctx->ctx_zombieq, &wait);
2012  		set_current_state(TASK_RUNNING);
2013
2014		/*
2015		 * context is unloaded at this point
2016		 */
2017		DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2018	}
2019	else if (task != current) {
2020#ifdef CONFIG_SMP
2021		/*
2022	 	 * switch context to zombie state
2023	 	 */
2024		ctx->ctx_state = PFM_CTX_ZOMBIE;
2025
2026		DPRINT(("zombie ctx for [%d]\n", task_pid_nr(task)));
2027		/*
2028		 * cannot free the context on the spot. deferred until
2029		 * the task notices the ZOMBIE state
2030		 */
2031		free_possible = 0;
2032#else
2033		pfm_context_unload(ctx, NULL, 0, regs);
2034#endif
2035	}
2036
2037doit:
2038	/* reload state, may have changed during  opening of critical section */
2039	state = ctx->ctx_state;
2040
2041	/*
2042	 * the context is still attached to a task (possibly current)
2043	 * we cannot destroy it right now
2044	 */
2045
2046	/*
2047	 * we must free the sampling buffer right here because
2048	 * we cannot rely on it being cleaned up later by the
2049	 * monitored task. It is not possible to free vmalloc'ed
2050	 * memory in pfm_load_regs(). Instead, we remove the buffer
2051	 * now. should there be subsequent PMU overflow originally
2052	 * meant for sampling, the will be converted to spurious
2053	 * and that's fine because the monitoring tools is gone anyway.
2054	 */
2055	if (ctx->ctx_smpl_hdr) {
2056		smpl_buf_addr = ctx->ctx_smpl_hdr;
2057		smpl_buf_size = ctx->ctx_smpl_size;
2058		/* no more sampling */
2059		ctx->ctx_smpl_hdr = NULL;
2060		ctx->ctx_fl_is_sampling = 0;
2061	}
2062
2063	DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2064		state,
2065		free_possible,
2066		smpl_buf_addr,
2067		smpl_buf_size));
2068
2069	if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2070
2071	/*
2072	 * UNLOADED that the session has already been unreserved.
2073	 */
2074	if (state == PFM_CTX_ZOMBIE) {
2075		pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2076	}
2077
2078	/*
2079	 * disconnect file descriptor from context must be done
2080	 * before we unlock.
2081	 */
2082	filp->private_data = NULL;
2083
2084	/*
2085	 * if we free on the spot, the context is now completely unreachable
2086	 * from the callers side. The monitored task side is also cut, so we
2087	 * can freely cut.
2088	 *
2089	 * If we have a deferred free, only the caller side is disconnected.
2090	 */
2091	UNPROTECT_CTX(ctx, flags);
2092
2093	/*
2094	 * All memory free operations (especially for vmalloc'ed memory)
2095	 * MUST be done with interrupts ENABLED.
2096	 */
2097	vfree(smpl_buf_addr);
2098
2099	/*
2100	 * return the memory used by the context
2101	 */
2102	if (free_possible) pfm_context_free(ctx);
2103
2104	return 0;
2105}
2106
 
 
 
 
 
 
 
 
 
2107static const struct file_operations pfm_file_ops = {
2108	.llseek		= no_llseek,
2109	.read		= pfm_read,
2110	.write		= pfm_write,
2111	.poll		= pfm_poll,
2112	.unlocked_ioctl = pfm_ioctl,
 
2113	.fasync		= pfm_fasync,
2114	.release	= pfm_close,
2115	.flush		= pfm_flush
2116};
2117
 
 
 
 
 
 
2118static char *pfmfs_dname(struct dentry *dentry, char *buffer, int buflen)
2119{
2120	return dynamic_dname(dentry, buffer, buflen, "pfm:[%lu]",
2121			     d_inode(dentry)->i_ino);
2122}
2123
2124static const struct dentry_operations pfmfs_dentry_operations = {
2125	.d_delete = always_delete_dentry,
2126	.d_dname = pfmfs_dname,
2127};
2128
2129
2130static struct file *
2131pfm_alloc_file(pfm_context_t *ctx)
2132{
2133	struct file *file;
2134	struct inode *inode;
2135	struct path path;
2136	struct qstr this = { .name = "" };
2137
2138	/*
2139	 * allocate a new inode
2140	 */
2141	inode = new_inode(pfmfs_mnt->mnt_sb);
2142	if (!inode)
2143		return ERR_PTR(-ENOMEM);
2144
2145	DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2146
2147	inode->i_mode = S_IFCHR|S_IRUGO;
2148	inode->i_uid  = current_fsuid();
2149	inode->i_gid  = current_fsgid();
2150
2151	/*
2152	 * allocate a new dcache entry
2153	 */
2154	path.dentry = d_alloc(pfmfs_mnt->mnt_root, &this);
2155	if (!path.dentry) {
2156		iput(inode);
2157		return ERR_PTR(-ENOMEM);
2158	}
2159	path.mnt = mntget(pfmfs_mnt);
2160
2161	d_add(path.dentry, inode);
2162
2163	file = alloc_file(&path, FMODE_READ, &pfm_file_ops);
2164	if (IS_ERR(file)) {
2165		path_put(&path);
2166		return file;
2167	}
2168
2169	file->f_flags = O_RDONLY;
2170	file->private_data = ctx;
2171
2172	return file;
2173}
2174
2175static int
2176pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2177{
2178	DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2179
2180	while (size > 0) {
2181		unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2182
2183
2184		if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2185			return -ENOMEM;
2186
2187		addr  += PAGE_SIZE;
2188		buf   += PAGE_SIZE;
2189		size  -= PAGE_SIZE;
2190	}
2191	return 0;
2192}
2193
2194/*
2195 * allocate a sampling buffer and remaps it into the user address space of the task
2196 */
2197static int
2198pfm_smpl_buffer_alloc(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2199{
2200	struct mm_struct *mm = task->mm;
2201	struct vm_area_struct *vma = NULL;
2202	unsigned long size;
2203	void *smpl_buf;
2204
2205
2206	/*
2207	 * the fixed header + requested size and align to page boundary
2208	 */
2209	size = PAGE_ALIGN(rsize);
2210
2211	DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2212
2213	/*
2214	 * check requested size to avoid Denial-of-service attacks
2215	 * XXX: may have to refine this test
2216	 * Check against address space limit.
2217	 *
2218	 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2219	 * 	return -ENOMEM;
2220	 */
2221	if (size > task_rlimit(task, RLIMIT_MEMLOCK))
2222		return -ENOMEM;
2223
2224	/*
2225	 * We do the easy to undo allocations first.
 
 
2226	 */
2227	smpl_buf = vzalloc(size);
2228	if (smpl_buf == NULL) {
2229		DPRINT(("Can't allocate sampling buffer\n"));
2230		return -ENOMEM;
2231	}
2232
2233	DPRINT(("smpl_buf @%p\n", smpl_buf));
2234
2235	/* allocate vma */
2236	vma = vm_area_alloc(mm);
2237	if (!vma) {
2238		DPRINT(("Cannot allocate vma\n"));
2239		goto error_kmem;
2240	}
 
2241
2242	/*
2243	 * partially initialize the vma for the sampling buffer
2244	 */
2245	vma->vm_file	     = get_file(filp);
2246	vma->vm_flags	     = VM_READ|VM_MAYREAD|VM_DONTEXPAND|VM_DONTDUMP;
 
2247	vma->vm_page_prot    = PAGE_READONLY; /* XXX may need to change */
2248
2249	/*
2250	 * Now we have everything we need and we can initialize
2251	 * and connect all the data structures
2252	 */
2253
2254	ctx->ctx_smpl_hdr   = smpl_buf;
2255	ctx->ctx_smpl_size  = size; /* aligned size */
2256
2257	/*
2258	 * Let's do the difficult operations next.
2259	 *
2260	 * now we atomically find some area in the address space and
2261	 * remap the buffer in it.
2262	 */
2263	mmap_write_lock(task->mm);
2264
2265	/* find some free area in address space, must have mmap sem held */
2266	vma->vm_start = get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS);
2267	if (IS_ERR_VALUE(vma->vm_start)) {
2268		DPRINT(("Cannot find unmapped area for size %ld\n", size));
2269		mmap_write_unlock(task->mm);
2270		goto error;
2271	}
2272	vma->vm_end = vma->vm_start + size;
2273	vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2274
2275	DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2276
2277	/* can only be applied to current task, need to have the mm semaphore held when called */
2278	if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2279		DPRINT(("Can't remap buffer\n"));
2280		mmap_write_unlock(task->mm);
2281		goto error;
2282	}
2283
 
 
2284	/*
2285	 * now insert the vma in the vm list for the process, must be
2286	 * done with mmap lock held
2287	 */
2288	insert_vm_struct(mm, vma);
2289
2290	vm_stat_account(vma->vm_mm, vma->vm_flags, vma_pages(vma));
2291	mmap_write_unlock(task->mm);
 
 
2292
2293	/*
2294	 * keep track of user level virtual address
2295	 */
2296	ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2297	*(unsigned long *)user_vaddr = vma->vm_start;
2298
2299	return 0;
2300
2301error:
2302	vm_area_free(vma);
2303error_kmem:
2304	vfree(smpl_buf);
2305
2306	return -ENOMEM;
2307}
2308
2309/*
2310 * XXX: do something better here
2311 */
2312static int
2313pfm_bad_permissions(struct task_struct *task)
2314{
2315	const struct cred *tcred;
2316	kuid_t uid = current_uid();
2317	kgid_t gid = current_gid();
2318	int ret;
2319
2320	rcu_read_lock();
2321	tcred = __task_cred(task);
2322
2323	/* inspired by ptrace_attach() */
2324	DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2325		from_kuid(&init_user_ns, uid),
2326		from_kgid(&init_user_ns, gid),
2327		from_kuid(&init_user_ns, tcred->euid),
2328		from_kuid(&init_user_ns, tcred->suid),
2329		from_kuid(&init_user_ns, tcred->uid),
2330		from_kgid(&init_user_ns, tcred->egid),
2331		from_kgid(&init_user_ns, tcred->sgid)));
2332
2333	ret = ((!uid_eq(uid, tcred->euid))
2334	       || (!uid_eq(uid, tcred->suid))
2335	       || (!uid_eq(uid, tcred->uid))
2336	       || (!gid_eq(gid, tcred->egid))
2337	       || (!gid_eq(gid, tcred->sgid))
2338	       || (!gid_eq(gid, tcred->gid))) && !capable(CAP_SYS_PTRACE);
2339
2340	rcu_read_unlock();
2341	return ret;
2342}
2343
2344static int
2345pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2346{
2347	int ctx_flags;
2348
2349	/* valid signal */
2350
2351	ctx_flags = pfx->ctx_flags;
2352
2353	if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2354
2355		/*
2356		 * cannot block in this mode
2357		 */
2358		if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2359			DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2360			return -EINVAL;
2361		}
2362	} else {
2363	}
2364	/* probably more to add here */
2365
2366	return 0;
2367}
2368
2369static int
2370pfm_setup_buffer_fmt(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned int ctx_flags,
2371		     unsigned int cpu, pfarg_context_t *arg)
2372{
2373	pfm_buffer_fmt_t *fmt = NULL;
2374	unsigned long size = 0UL;
2375	void *uaddr = NULL;
2376	void *fmt_arg = NULL;
2377	int ret = 0;
2378#define PFM_CTXARG_BUF_ARG(a)	(pfm_buffer_fmt_t *)(a+1)
2379
2380	/* invoke and lock buffer format, if found */
2381	fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2382	if (fmt == NULL) {
2383		DPRINT(("[%d] cannot find buffer format\n", task_pid_nr(task)));
2384		return -EINVAL;
2385	}
2386
2387	/*
2388	 * buffer argument MUST be contiguous to pfarg_context_t
2389	 */
2390	if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2391
2392	ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2393
2394	DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task_pid_nr(task), ctx_flags, cpu, fmt_arg, ret));
2395
2396	if (ret) goto error;
2397
2398	/* link buffer format and context */
2399	ctx->ctx_buf_fmt = fmt;
2400	ctx->ctx_fl_is_sampling = 1; /* assume record() is defined */
2401
2402	/*
2403	 * check if buffer format wants to use perfmon buffer allocation/mapping service
2404	 */
2405	ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2406	if (ret) goto error;
2407
2408	if (size) {
2409		/*
2410		 * buffer is always remapped into the caller's address space
2411		 */
2412		ret = pfm_smpl_buffer_alloc(current, filp, ctx, size, &uaddr);
2413		if (ret) goto error;
2414
2415		/* keep track of user address of buffer */
2416		arg->ctx_smpl_vaddr = uaddr;
2417	}
2418	ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2419
2420error:
2421	return ret;
2422}
2423
2424static void
2425pfm_reset_pmu_state(pfm_context_t *ctx)
2426{
2427	int i;
2428
2429	/*
2430	 * install reset values for PMC.
2431	 */
2432	for (i=1; PMC_IS_LAST(i) == 0; i++) {
2433		if (PMC_IS_IMPL(i) == 0) continue;
2434		ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2435		DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2436	}
2437	/*
2438	 * PMD registers are set to 0UL when the context in memset()
2439	 */
2440
2441	/*
2442	 * On context switched restore, we must restore ALL pmc and ALL pmd even
2443	 * when they are not actively used by the task. In UP, the incoming process
2444	 * may otherwise pick up left over PMC, PMD state from the previous process.
2445	 * As opposed to PMD, stale PMC can cause harm to the incoming
2446	 * process because they may change what is being measured.
2447	 * Therefore, we must systematically reinstall the entire
2448	 * PMC state. In SMP, the same thing is possible on the
2449	 * same CPU but also on between 2 CPUs.
2450	 *
2451	 * The problem with PMD is information leaking especially
2452	 * to user level when psr.sp=0
2453	 *
2454	 * There is unfortunately no easy way to avoid this problem
2455	 * on either UP or SMP. This definitively slows down the
2456	 * pfm_load_regs() function.
2457	 */
2458
2459	 /*
2460	  * bitmask of all PMCs accessible to this context
2461	  *
2462	  * PMC0 is treated differently.
2463	  */
2464	ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2465
2466	/*
2467	 * bitmask of all PMDs that are accessible to this context
2468	 */
2469	ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2470
2471	DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2472
2473	/*
2474	 * useful in case of re-enable after disable
2475	 */
2476	ctx->ctx_used_ibrs[0] = 0UL;
2477	ctx->ctx_used_dbrs[0] = 0UL;
2478}
2479
2480static int
2481pfm_ctx_getsize(void *arg, size_t *sz)
2482{
2483	pfarg_context_t *req = (pfarg_context_t *)arg;
2484	pfm_buffer_fmt_t *fmt;
2485
2486	*sz = 0;
2487
2488	if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2489
2490	fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2491	if (fmt == NULL) {
2492		DPRINT(("cannot find buffer format\n"));
2493		return -EINVAL;
2494	}
2495	/* get just enough to copy in user parameters */
2496	*sz = fmt->fmt_arg_size;
2497	DPRINT(("arg_size=%lu\n", *sz));
2498
2499	return 0;
2500}
2501
2502
2503
2504/*
2505 * cannot attach if :
2506 * 	- kernel task
2507 * 	- task not owned by caller
2508 * 	- task incompatible with context mode
2509 */
2510static int
2511pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2512{
2513	/*
2514	 * no kernel task or task not owner by caller
2515	 */
2516	if (task->mm == NULL) {
2517		DPRINT(("task [%d] has not memory context (kernel thread)\n", task_pid_nr(task)));
2518		return -EPERM;
2519	}
2520	if (pfm_bad_permissions(task)) {
2521		DPRINT(("no permission to attach to  [%d]\n", task_pid_nr(task)));
2522		return -EPERM;
2523	}
2524	/*
2525	 * cannot block in self-monitoring mode
2526	 */
2527	if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2528		DPRINT(("cannot load a blocking context on self for [%d]\n", task_pid_nr(task)));
2529		return -EINVAL;
2530	}
2531
2532	if (task->exit_state == EXIT_ZOMBIE) {
2533		DPRINT(("cannot attach to  zombie task [%d]\n", task_pid_nr(task)));
2534		return -EBUSY;
2535	}
2536
2537	/*
2538	 * always ok for self
2539	 */
2540	if (task == current) return 0;
2541
2542	if (!task_is_stopped_or_traced(task)) {
2543		DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task_pid_nr(task), task->state));
2544		return -EBUSY;
2545	}
2546	/*
2547	 * make sure the task is off any CPU
2548	 */
2549	wait_task_inactive(task, 0);
2550
2551	/* more to come... */
2552
2553	return 0;
2554}
2555
2556static int
2557pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2558{
2559	struct task_struct *p = current;
2560	int ret;
2561
2562	/* XXX: need to add more checks here */
2563	if (pid < 2) return -EPERM;
2564
2565	if (pid != task_pid_vnr(current)) {
 
 
 
 
 
2566		/* make sure task cannot go away while we operate on it */
2567		p = find_get_task_by_vpid(pid);
2568		if (!p)
2569			return -ESRCH;
 
 
2570	}
2571
2572	ret = pfm_task_incompatible(ctx, p);
2573	if (ret == 0) {
2574		*task = p;
2575	} else if (p != current) {
2576		pfm_put_task(p);
2577	}
2578	return ret;
2579}
2580
2581
2582
2583static int
2584pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2585{
2586	pfarg_context_t *req = (pfarg_context_t *)arg;
2587	struct file *filp;
2588	struct path path;
2589	int ctx_flags;
2590	int fd;
2591	int ret;
2592
2593	/* let's check the arguments first */
2594	ret = pfarg_is_sane(current, req);
2595	if (ret < 0)
2596		return ret;
2597
2598	ctx_flags = req->ctx_flags;
2599
2600	ret = -ENOMEM;
2601
2602	fd = get_unused_fd_flags(0);
2603	if (fd < 0)
2604		return fd;
2605
2606	ctx = pfm_context_alloc(ctx_flags);
2607	if (!ctx)
2608		goto error;
2609
2610	filp = pfm_alloc_file(ctx);
2611	if (IS_ERR(filp)) {
2612		ret = PTR_ERR(filp);
2613		goto error_file;
2614	}
2615
2616	req->ctx_fd = ctx->ctx_fd = fd;
2617
2618	/*
2619	 * does the user want to sample?
2620	 */
2621	if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2622		ret = pfm_setup_buffer_fmt(current, filp, ctx, ctx_flags, 0, req);
2623		if (ret)
2624			goto buffer_error;
2625	}
2626
2627	DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d\n",
2628		ctx,
2629		ctx_flags,
2630		ctx->ctx_fl_system,
2631		ctx->ctx_fl_block,
2632		ctx->ctx_fl_excl_idle,
2633		ctx->ctx_fl_no_msg,
2634		ctx->ctx_fd));
2635
2636	/*
2637	 * initialize soft PMU state
2638	 */
2639	pfm_reset_pmu_state(ctx);
2640
2641	fd_install(fd, filp);
2642
2643	return 0;
2644
2645buffer_error:
2646	path = filp->f_path;
2647	put_filp(filp);
2648	path_put(&path);
2649
2650	if (ctx->ctx_buf_fmt) {
2651		pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2652	}
2653error_file:
2654	pfm_context_free(ctx);
2655
2656error:
2657	put_unused_fd(fd);
2658	return ret;
2659}
2660
2661static inline unsigned long
2662pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2663{
2664	unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2665	unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2666	extern unsigned long carta_random32 (unsigned long seed);
2667
2668	if (reg->flags & PFM_REGFL_RANDOM) {
2669		new_seed = carta_random32(old_seed);
2670		val -= (old_seed & mask);	/* counter values are negative numbers! */
2671		if ((mask >> 32) != 0)
2672			/* construct a full 64-bit random value: */
2673			new_seed |= carta_random32(old_seed >> 32) << 32;
2674		reg->seed = new_seed;
2675	}
2676	reg->lval = val;
2677	return val;
2678}
2679
2680static void
2681pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2682{
2683	unsigned long mask = ovfl_regs[0];
2684	unsigned long reset_others = 0UL;
2685	unsigned long val;
2686	int i;
2687
2688	/*
2689	 * now restore reset value on sampling overflowed counters
2690	 */
2691	mask >>= PMU_FIRST_COUNTER;
2692	for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2693
2694		if ((mask & 0x1UL) == 0UL) continue;
2695
2696		ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2697		reset_others        |= ctx->ctx_pmds[i].reset_pmds[0];
2698
2699		DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2700	}
2701
2702	/*
2703	 * Now take care of resetting the other registers
2704	 */
2705	for(i = 0; reset_others; i++, reset_others >>= 1) {
2706
2707		if ((reset_others & 0x1) == 0) continue;
2708
2709		ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2710
2711		DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2712			  is_long_reset ? "long" : "short", i, val));
2713	}
2714}
2715
2716static void
2717pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2718{
2719	unsigned long mask = ovfl_regs[0];
2720	unsigned long reset_others = 0UL;
2721	unsigned long val;
2722	int i;
2723
2724	DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2725
2726	if (ctx->ctx_state == PFM_CTX_MASKED) {
2727		pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2728		return;
2729	}
2730
2731	/*
2732	 * now restore reset value on sampling overflowed counters
2733	 */
2734	mask >>= PMU_FIRST_COUNTER;
2735	for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2736
2737		if ((mask & 0x1UL) == 0UL) continue;
2738
2739		val           = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2740		reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2741
2742		DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2743
2744		pfm_write_soft_counter(ctx, i, val);
2745	}
2746
2747	/*
2748	 * Now take care of resetting the other registers
2749	 */
2750	for(i = 0; reset_others; i++, reset_others >>= 1) {
2751
2752		if ((reset_others & 0x1) == 0) continue;
2753
2754		val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2755
2756		if (PMD_IS_COUNTING(i)) {
2757			pfm_write_soft_counter(ctx, i, val);
2758		} else {
2759			ia64_set_pmd(i, val);
2760		}
2761		DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2762			  is_long_reset ? "long" : "short", i, val));
2763	}
2764	ia64_srlz_d();
2765}
2766
2767static int
2768pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2769{
2770	struct task_struct *task;
2771	pfarg_reg_t *req = (pfarg_reg_t *)arg;
2772	unsigned long value, pmc_pm;
2773	unsigned long smpl_pmds, reset_pmds, impl_pmds;
2774	unsigned int cnum, reg_flags, flags, pmc_type;
2775	int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2776	int is_monitor, is_counting, state;
2777	int ret = -EINVAL;
2778	pfm_reg_check_t	wr_func;
2779#define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2780
2781	state     = ctx->ctx_state;
2782	is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2783	is_system = ctx->ctx_fl_system;
2784	task      = ctx->ctx_task;
2785	impl_pmds = pmu_conf->impl_pmds[0];
2786
2787	if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2788
2789	if (is_loaded) {
2790		/*
2791		 * In system wide and when the context is loaded, access can only happen
2792		 * when the caller is running on the CPU being monitored by the session.
2793		 * It does not have to be the owner (ctx_task) of the context per se.
2794		 */
2795		if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2796			DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2797			return -EBUSY;
2798		}
2799		can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2800	}
2801	expert_mode = pfm_sysctl.expert_mode; 
2802
2803	for (i = 0; i < count; i++, req++) {
2804
2805		cnum       = req->reg_num;
2806		reg_flags  = req->reg_flags;
2807		value      = req->reg_value;
2808		smpl_pmds  = req->reg_smpl_pmds[0];
2809		reset_pmds = req->reg_reset_pmds[0];
2810		flags      = 0;
2811
2812
2813		if (cnum >= PMU_MAX_PMCS) {
2814			DPRINT(("pmc%u is invalid\n", cnum));
2815			goto error;
2816		}
2817
2818		pmc_type   = pmu_conf->pmc_desc[cnum].type;
2819		pmc_pm     = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2820		is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2821		is_monitor  = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2822
2823		/*
2824		 * we reject all non implemented PMC as well
2825		 * as attempts to modify PMC[0-3] which are used
2826		 * as status registers by the PMU
2827		 */
2828		if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2829			DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2830			goto error;
2831		}
2832		wr_func = pmu_conf->pmc_desc[cnum].write_check;
2833		/*
2834		 * If the PMC is a monitor, then if the value is not the default:
2835		 * 	- system-wide session: PMCx.pm=1 (privileged monitor)
2836		 * 	- per-task           : PMCx.pm=0 (user monitor)
2837		 */
2838		if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2839			DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2840				cnum,
2841				pmc_pm,
2842				is_system));
2843			goto error;
2844		}
2845
2846		if (is_counting) {
2847			/*
2848		 	 * enforce generation of overflow interrupt. Necessary on all
2849		 	 * CPUs.
2850		 	 */
2851			value |= 1 << PMU_PMC_OI;
2852
2853			if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2854				flags |= PFM_REGFL_OVFL_NOTIFY;
2855			}
2856
2857			if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2858
2859			/* verify validity of smpl_pmds */
2860			if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2861				DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2862				goto error;
2863			}
2864
2865			/* verify validity of reset_pmds */
2866			if ((reset_pmds & impl_pmds) != reset_pmds) {
2867				DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2868				goto error;
2869			}
2870		} else {
2871			if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2872				DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2873				goto error;
2874			}
2875			/* eventid on non-counting monitors are ignored */
2876		}
2877
2878		/*
2879		 * execute write checker, if any
2880		 */
2881		if (likely(expert_mode == 0 && wr_func)) {
2882			ret = (*wr_func)(task, ctx, cnum, &value, regs);
2883			if (ret) goto error;
2884			ret = -EINVAL;
2885		}
2886
2887		/*
2888		 * no error on this register
2889		 */
2890		PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2891
2892		/*
2893		 * Now we commit the changes to the software state
2894		 */
2895
2896		/*
2897		 * update overflow information
2898		 */
2899		if (is_counting) {
2900			/*
2901		 	 * full flag update each time a register is programmed
2902		 	 */
2903			ctx->ctx_pmds[cnum].flags = flags;
2904
2905			ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2906			ctx->ctx_pmds[cnum].smpl_pmds[0]  = smpl_pmds;
2907			ctx->ctx_pmds[cnum].eventid       = req->reg_smpl_eventid;
2908
2909			/*
2910			 * Mark all PMDS to be accessed as used.
2911			 *
2912			 * We do not keep track of PMC because we have to
2913			 * systematically restore ALL of them.
2914			 *
2915			 * We do not update the used_monitors mask, because
2916			 * if we have not programmed them, then will be in
2917			 * a quiescent state, therefore we will not need to
2918			 * mask/restore then when context is MASKED.
2919			 */
2920			CTX_USED_PMD(ctx, reset_pmds);
2921			CTX_USED_PMD(ctx, smpl_pmds);
2922			/*
2923		 	 * make sure we do not try to reset on
2924		 	 * restart because we have established new values
2925		 	 */
2926			if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
2927		}
2928		/*
2929		 * Needed in case the user does not initialize the equivalent
2930		 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
2931		 * possible leak here.
2932		 */
2933		CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
2934
2935		/*
2936		 * keep track of the monitor PMC that we are using.
2937		 * we save the value of the pmc in ctx_pmcs[] and if
2938		 * the monitoring is not stopped for the context we also
2939		 * place it in the saved state area so that it will be
2940		 * picked up later by the context switch code.
2941		 *
2942		 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
2943		 *
2944		 * The value in th_pmcs[] may be modified on overflow, i.e.,  when
2945		 * monitoring needs to be stopped.
2946		 */
2947		if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
2948
2949		/*
2950		 * update context state
2951		 */
2952		ctx->ctx_pmcs[cnum] = value;
2953
2954		if (is_loaded) {
2955			/*
2956			 * write thread state
2957			 */
2958			if (is_system == 0) ctx->th_pmcs[cnum] = value;
2959
2960			/*
2961			 * write hardware register if we can
2962			 */
2963			if (can_access_pmu) {
2964				ia64_set_pmc(cnum, value);
2965			}
2966#ifdef CONFIG_SMP
2967			else {
2968				/*
2969				 * per-task SMP only here
2970				 *
2971			 	 * we are guaranteed that the task is not running on the other CPU,
2972			 	 * we indicate that this PMD will need to be reloaded if the task
2973			 	 * is rescheduled on the CPU it ran last on.
2974			 	 */
2975				ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
2976			}
2977#endif
2978		}
2979
2980		DPRINT(("pmc[%u]=0x%lx ld=%d apmu=%d flags=0x%x all_pmcs=0x%lx used_pmds=0x%lx eventid=%ld smpl_pmds=0x%lx reset_pmds=0x%lx reloads_pmcs=0x%lx used_monitors=0x%lx ovfl_regs=0x%lx\n",
2981			  cnum,
2982			  value,
2983			  is_loaded,
2984			  can_access_pmu,
2985			  flags,
2986			  ctx->ctx_all_pmcs[0],
2987			  ctx->ctx_used_pmds[0],
2988			  ctx->ctx_pmds[cnum].eventid,
2989			  smpl_pmds,
2990			  reset_pmds,
2991			  ctx->ctx_reload_pmcs[0],
2992			  ctx->ctx_used_monitors[0],
2993			  ctx->ctx_ovfl_regs[0]));
2994	}
2995
2996	/*
2997	 * make sure the changes are visible
2998	 */
2999	if (can_access_pmu) ia64_srlz_d();
3000
3001	return 0;
3002error:
3003	PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3004	return ret;
3005}
3006
3007static int
3008pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3009{
3010	struct task_struct *task;
3011	pfarg_reg_t *req = (pfarg_reg_t *)arg;
3012	unsigned long value, hw_value, ovfl_mask;
3013	unsigned int cnum;
3014	int i, can_access_pmu = 0, state;
3015	int is_counting, is_loaded, is_system, expert_mode;
3016	int ret = -EINVAL;
3017	pfm_reg_check_t wr_func;
3018
3019
3020	state     = ctx->ctx_state;
3021	is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3022	is_system = ctx->ctx_fl_system;
3023	ovfl_mask = pmu_conf->ovfl_val;
3024	task      = ctx->ctx_task;
3025
3026	if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3027
3028	/*
3029	 * on both UP and SMP, we can only write to the PMC when the task is
3030	 * the owner of the local PMU.
3031	 */
3032	if (likely(is_loaded)) {
3033		/*
3034		 * In system wide and when the context is loaded, access can only happen
3035		 * when the caller is running on the CPU being monitored by the session.
3036		 * It does not have to be the owner (ctx_task) of the context per se.
3037		 */
3038		if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3039			DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3040			return -EBUSY;
3041		}
3042		can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3043	}
3044	expert_mode = pfm_sysctl.expert_mode; 
3045
3046	for (i = 0; i < count; i++, req++) {
3047
3048		cnum  = req->reg_num;
3049		value = req->reg_value;
3050
3051		if (!PMD_IS_IMPL(cnum)) {
3052			DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3053			goto abort_mission;
3054		}
3055		is_counting = PMD_IS_COUNTING(cnum);
3056		wr_func     = pmu_conf->pmd_desc[cnum].write_check;
3057
3058		/*
3059		 * execute write checker, if any
3060		 */
3061		if (unlikely(expert_mode == 0 && wr_func)) {
3062			unsigned long v = value;
3063
3064			ret = (*wr_func)(task, ctx, cnum, &v, regs);
3065			if (ret) goto abort_mission;
3066
3067			value = v;
3068			ret   = -EINVAL;
3069		}
3070
3071		/*
3072		 * no error on this register
3073		 */
3074		PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3075
3076		/*
3077		 * now commit changes to software state
3078		 */
3079		hw_value = value;
3080
3081		/*
3082		 * update virtualized (64bits) counter
3083		 */
3084		if (is_counting) {
3085			/*
3086			 * write context state
3087			 */
3088			ctx->ctx_pmds[cnum].lval = value;
3089
3090			/*
3091			 * when context is load we use the split value
3092			 */
3093			if (is_loaded) {
3094				hw_value = value &  ovfl_mask;
3095				value    = value & ~ovfl_mask;
3096			}
3097		}
3098		/*
3099		 * update reset values (not just for counters)
3100		 */
3101		ctx->ctx_pmds[cnum].long_reset  = req->reg_long_reset;
3102		ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3103
3104		/*
3105		 * update randomization parameters (not just for counters)
3106		 */
3107		ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3108		ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3109
3110		/*
3111		 * update context value
3112		 */
3113		ctx->ctx_pmds[cnum].val  = value;
3114
3115		/*
3116		 * Keep track of what we use
3117		 *
3118		 * We do not keep track of PMC because we have to
3119		 * systematically restore ALL of them.
3120		 */
3121		CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3122
3123		/*
3124		 * mark this PMD register used as well
3125		 */
3126		CTX_USED_PMD(ctx, RDEP(cnum));
3127
3128		/*
3129		 * make sure we do not try to reset on
3130		 * restart because we have established new values
3131		 */
3132		if (is_counting && state == PFM_CTX_MASKED) {
3133			ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3134		}
3135
3136		if (is_loaded) {
3137			/*
3138		 	 * write thread state
3139		 	 */
3140			if (is_system == 0) ctx->th_pmds[cnum] = hw_value;
3141
3142			/*
3143			 * write hardware register if we can
3144			 */
3145			if (can_access_pmu) {
3146				ia64_set_pmd(cnum, hw_value);
3147			} else {
3148#ifdef CONFIG_SMP
3149				/*
3150			 	 * we are guaranteed that the task is not running on the other CPU,
3151			 	 * we indicate that this PMD will need to be reloaded if the task
3152			 	 * is rescheduled on the CPU it ran last on.
3153			 	 */
3154				ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3155#endif
3156			}
3157		}
3158
3159		DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx  short_reset=0x%lx "
3160			  "long_reset=0x%lx notify=%c seed=0x%lx mask=0x%lx used_pmds=0x%lx reset_pmds=0x%lx reload_pmds=0x%lx all_pmds=0x%lx ovfl_regs=0x%lx\n",
3161			cnum,
3162			value,
3163			is_loaded,
3164			can_access_pmu,
3165			hw_value,
3166			ctx->ctx_pmds[cnum].val,
3167			ctx->ctx_pmds[cnum].short_reset,
3168			ctx->ctx_pmds[cnum].long_reset,
3169			PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3170			ctx->ctx_pmds[cnum].seed,
3171			ctx->ctx_pmds[cnum].mask,
3172			ctx->ctx_used_pmds[0],
3173			ctx->ctx_pmds[cnum].reset_pmds[0],
3174			ctx->ctx_reload_pmds[0],
3175			ctx->ctx_all_pmds[0],
3176			ctx->ctx_ovfl_regs[0]));
3177	}
3178
3179	/*
3180	 * make changes visible
3181	 */
3182	if (can_access_pmu) ia64_srlz_d();
3183
3184	return 0;
3185
3186abort_mission:
3187	/*
3188	 * for now, we have only one possibility for error
3189	 */
3190	PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3191	return ret;
3192}
3193
3194/*
3195 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3196 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3197 * interrupt is delivered during the call, it will be kept pending until we leave, making
3198 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3199 * guaranteed to return consistent data to the user, it may simply be old. It is not
3200 * trivial to treat the overflow while inside the call because you may end up in
3201 * some module sampling buffer code causing deadlocks.
3202 */
3203static int
3204pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3205{
3206	struct task_struct *task;
3207	unsigned long val = 0UL, lval, ovfl_mask, sval;
3208	pfarg_reg_t *req = (pfarg_reg_t *)arg;
3209	unsigned int cnum, reg_flags = 0;
3210	int i, can_access_pmu = 0, state;
3211	int is_loaded, is_system, is_counting, expert_mode;
3212	int ret = -EINVAL;
3213	pfm_reg_check_t rd_func;
3214
3215	/*
3216	 * access is possible when loaded only for
3217	 * self-monitoring tasks or in UP mode
3218	 */
3219
3220	state     = ctx->ctx_state;
3221	is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3222	is_system = ctx->ctx_fl_system;
3223	ovfl_mask = pmu_conf->ovfl_val;
3224	task      = ctx->ctx_task;
3225
3226	if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3227
3228	if (likely(is_loaded)) {
3229		/*
3230		 * In system wide and when the context is loaded, access can only happen
3231		 * when the caller is running on the CPU being monitored by the session.
3232		 * It does not have to be the owner (ctx_task) of the context per se.
3233		 */
3234		if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3235			DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3236			return -EBUSY;
3237		}
3238		/*
3239		 * this can be true when not self-monitoring only in UP
3240		 */
3241		can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3242
3243		if (can_access_pmu) ia64_srlz_d();
3244	}
3245	expert_mode = pfm_sysctl.expert_mode; 
3246
3247	DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3248		is_loaded,
3249		can_access_pmu,
3250		state));
3251
3252	/*
3253	 * on both UP and SMP, we can only read the PMD from the hardware register when
3254	 * the task is the owner of the local PMU.
3255	 */
3256
3257	for (i = 0; i < count; i++, req++) {
3258
3259		cnum        = req->reg_num;
3260		reg_flags   = req->reg_flags;
3261
3262		if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3263		/*
3264		 * we can only read the register that we use. That includes
3265		 * the one we explicitly initialize AND the one we want included
3266		 * in the sampling buffer (smpl_regs).
3267		 *
3268		 * Having this restriction allows optimization in the ctxsw routine
3269		 * without compromising security (leaks)
3270		 */
3271		if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3272
3273		sval        = ctx->ctx_pmds[cnum].val;
3274		lval        = ctx->ctx_pmds[cnum].lval;
3275		is_counting = PMD_IS_COUNTING(cnum);
3276
3277		/*
3278		 * If the task is not the current one, then we check if the
3279		 * PMU state is still in the local live register due to lazy ctxsw.
3280		 * If true, then we read directly from the registers.
3281		 */
3282		if (can_access_pmu){
3283			val = ia64_get_pmd(cnum);
3284		} else {
3285			/*
3286			 * context has been saved
3287			 * if context is zombie, then task does not exist anymore.
3288			 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3289			 */
3290			val = is_loaded ? ctx->th_pmds[cnum] : 0UL;
3291		}
3292		rd_func = pmu_conf->pmd_desc[cnum].read_check;
3293
3294		if (is_counting) {
3295			/*
3296			 * XXX: need to check for overflow when loaded
3297			 */
3298			val &= ovfl_mask;
3299			val += sval;
3300		}
3301
3302		/*
3303		 * execute read checker, if any
3304		 */
3305		if (unlikely(expert_mode == 0 && rd_func)) {
3306			unsigned long v = val;
3307			ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3308			if (ret) goto error;
3309			val = v;
3310			ret = -EINVAL;
3311		}
3312
3313		PFM_REG_RETFLAG_SET(reg_flags, 0);
3314
3315		DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3316
3317		/*
3318		 * update register return value, abort all if problem during copy.
3319		 * we only modify the reg_flags field. no check mode is fine because
3320		 * access has been verified upfront in sys_perfmonctl().
3321		 */
3322		req->reg_value            = val;
3323		req->reg_flags            = reg_flags;
3324		req->reg_last_reset_val   = lval;
3325	}
3326
3327	return 0;
3328
3329error:
3330	PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3331	return ret;
3332}
3333
3334int
3335pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3336{
3337	pfm_context_t *ctx;
3338
3339	if (req == NULL) return -EINVAL;
3340
3341 	ctx = GET_PMU_CTX();
3342
3343	if (ctx == NULL) return -EINVAL;
3344
3345	/*
3346	 * for now limit to current task, which is enough when calling
3347	 * from overflow handler
3348	 */
3349	if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3350
3351	return pfm_write_pmcs(ctx, req, nreq, regs);
3352}
3353EXPORT_SYMBOL(pfm_mod_write_pmcs);
3354
3355int
3356pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3357{
3358	pfm_context_t *ctx;
3359
3360	if (req == NULL) return -EINVAL;
3361
3362 	ctx = GET_PMU_CTX();
3363
3364	if (ctx == NULL) return -EINVAL;
3365
3366	/*
3367	 * for now limit to current task, which is enough when calling
3368	 * from overflow handler
3369	 */
3370	if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3371
3372	return pfm_read_pmds(ctx, req, nreq, regs);
3373}
3374EXPORT_SYMBOL(pfm_mod_read_pmds);
3375
3376/*
3377 * Only call this function when a process it trying to
3378 * write the debug registers (reading is always allowed)
3379 */
3380int
3381pfm_use_debug_registers(struct task_struct *task)
3382{
3383	pfm_context_t *ctx = task->thread.pfm_context;
3384	unsigned long flags;
3385	int ret = 0;
3386
3387	if (pmu_conf->use_rr_dbregs == 0) return 0;
3388
3389	DPRINT(("called for [%d]\n", task_pid_nr(task)));
3390
3391	/*
3392	 * do it only once
3393	 */
3394	if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3395
3396	/*
3397	 * Even on SMP, we do not need to use an atomic here because
3398	 * the only way in is via ptrace() and this is possible only when the
3399	 * process is stopped. Even in the case where the ctxsw out is not totally
3400	 * completed by the time we come here, there is no way the 'stopped' process
3401	 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3402	 * So this is always safe.
3403	 */
3404	if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3405
3406	LOCK_PFS(flags);
3407
3408	/*
3409	 * We cannot allow setting breakpoints when system wide monitoring
3410	 * sessions are using the debug registers.
3411	 */
3412	if (pfm_sessions.pfs_sys_use_dbregs> 0)
3413		ret = -1;
3414	else
3415		pfm_sessions.pfs_ptrace_use_dbregs++;
3416
3417	DPRINT(("ptrace_use_dbregs=%u  sys_use_dbregs=%u by [%d] ret = %d\n",
3418		  pfm_sessions.pfs_ptrace_use_dbregs,
3419		  pfm_sessions.pfs_sys_use_dbregs,
3420		  task_pid_nr(task), ret));
3421
3422	UNLOCK_PFS(flags);
3423
3424	return ret;
3425}
3426
3427/*
3428 * This function is called for every task that exits with the
3429 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3430 * able to use the debug registers for debugging purposes via
3431 * ptrace(). Therefore we know it was not using them for
3432 * performance monitoring, so we only decrement the number
3433 * of "ptraced" debug register users to keep the count up to date
3434 */
3435int
3436pfm_release_debug_registers(struct task_struct *task)
3437{
3438	unsigned long flags;
3439	int ret;
3440
3441	if (pmu_conf->use_rr_dbregs == 0) return 0;
3442
3443	LOCK_PFS(flags);
3444	if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3445		printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task_pid_nr(task));
3446		ret = -1;
3447	}  else {
3448		pfm_sessions.pfs_ptrace_use_dbregs--;
3449		ret = 0;
3450	}
3451	UNLOCK_PFS(flags);
3452
3453	return ret;
3454}
3455
3456static int
3457pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3458{
3459	struct task_struct *task;
3460	pfm_buffer_fmt_t *fmt;
3461	pfm_ovfl_ctrl_t rst_ctrl;
3462	int state, is_system;
3463	int ret = 0;
3464
3465	state     = ctx->ctx_state;
3466	fmt       = ctx->ctx_buf_fmt;
3467	is_system = ctx->ctx_fl_system;
3468	task      = PFM_CTX_TASK(ctx);
3469
3470	switch(state) {
3471		case PFM_CTX_MASKED:
3472			break;
3473		case PFM_CTX_LOADED: 
3474			if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3475			fallthrough;
3476		case PFM_CTX_UNLOADED:
3477		case PFM_CTX_ZOMBIE:
3478			DPRINT(("invalid state=%d\n", state));
3479			return -EBUSY;
3480		default:
3481			DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3482			return -EINVAL;
3483	}
3484
3485	/*
3486 	 * In system wide and when the context is loaded, access can only happen
3487 	 * when the caller is running on the CPU being monitored by the session.
3488 	 * It does not have to be the owner (ctx_task) of the context per se.
3489 	 */
3490	if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3491		DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3492		return -EBUSY;
3493	}
3494
3495	/* sanity check */
3496	if (unlikely(task == NULL)) {
3497		printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", task_pid_nr(current));
3498		return -EINVAL;
3499	}
3500
3501	if (task == current || is_system) {
3502
3503		fmt = ctx->ctx_buf_fmt;
3504
3505		DPRINT(("restarting self %d ovfl=0x%lx\n",
3506			task_pid_nr(task),
3507			ctx->ctx_ovfl_regs[0]));
3508
3509		if (CTX_HAS_SMPL(ctx)) {
3510
3511			prefetch(ctx->ctx_smpl_hdr);
3512
3513			rst_ctrl.bits.mask_monitoring = 0;
3514			rst_ctrl.bits.reset_ovfl_pmds = 0;
3515
3516			if (state == PFM_CTX_LOADED)
3517				ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3518			else
3519				ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3520		} else {
3521			rst_ctrl.bits.mask_monitoring = 0;
3522			rst_ctrl.bits.reset_ovfl_pmds = 1;
3523		}
3524
3525		if (ret == 0) {
3526			if (rst_ctrl.bits.reset_ovfl_pmds)
3527				pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3528
3529			if (rst_ctrl.bits.mask_monitoring == 0) {
3530				DPRINT(("resuming monitoring for [%d]\n", task_pid_nr(task)));
3531
3532				if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3533			} else {
3534				DPRINT(("keeping monitoring stopped for [%d]\n", task_pid_nr(task)));
3535
3536				// cannot use pfm_stop_monitoring(task, regs);
3537			}
3538		}
3539		/*
3540		 * clear overflowed PMD mask to remove any stale information
3541		 */
3542		ctx->ctx_ovfl_regs[0] = 0UL;
3543
3544		/*
3545		 * back to LOADED state
3546		 */
3547		ctx->ctx_state = PFM_CTX_LOADED;
3548
3549		/*
3550		 * XXX: not really useful for self monitoring
3551		 */
3552		ctx->ctx_fl_can_restart = 0;
3553
3554		return 0;
3555	}
3556
3557	/* 
3558	 * restart another task
3559	 */
3560
3561	/*
3562	 * When PFM_CTX_MASKED, we cannot issue a restart before the previous 
3563	 * one is seen by the task.
3564	 */
3565	if (state == PFM_CTX_MASKED) {
3566		if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3567		/*
3568		 * will prevent subsequent restart before this one is
3569		 * seen by other task
3570		 */
3571		ctx->ctx_fl_can_restart = 0;
3572	}
3573
3574	/*
3575	 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3576	 * the task is blocked or on its way to block. That's the normal
3577	 * restart path. If the monitoring is not masked, then the task
3578	 * can be actively monitoring and we cannot directly intervene.
3579	 * Therefore we use the trap mechanism to catch the task and
3580	 * force it to reset the buffer/reset PMDs.
3581	 *
3582	 * if non-blocking, then we ensure that the task will go into
3583	 * pfm_handle_work() before returning to user mode.
3584	 *
3585	 * We cannot explicitly reset another task, it MUST always
3586	 * be done by the task itself. This works for system wide because
3587	 * the tool that is controlling the session is logically doing 
3588	 * "self-monitoring".
3589	 */
3590	if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3591		DPRINT(("unblocking [%d]\n", task_pid_nr(task)));
3592		complete(&ctx->ctx_restart_done);
3593	} else {
3594		DPRINT(("[%d] armed exit trap\n", task_pid_nr(task)));
3595
3596		ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3597
3598		PFM_SET_WORK_PENDING(task, 1);
3599
3600		set_notify_resume(task);
3601
3602		/*
3603		 * XXX: send reschedule if task runs on another CPU
3604		 */
3605	}
3606	return 0;
3607}
3608
3609static int
3610pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3611{
3612	unsigned int m = *(unsigned int *)arg;
3613
3614	pfm_sysctl.debug = m == 0 ? 0 : 1;
3615
3616	printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3617
3618	if (m == 0) {
3619		memset(pfm_stats, 0, sizeof(pfm_stats));
3620		for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3621	}
3622	return 0;
3623}
3624
3625/*
3626 * arg can be NULL and count can be zero for this function
3627 */
3628static int
3629pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3630{
3631	struct thread_struct *thread = NULL;
3632	struct task_struct *task;
3633	pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3634	unsigned long flags;
3635	dbreg_t dbreg;
3636	unsigned int rnum;
3637	int first_time;
3638	int ret = 0, state;
3639	int i, can_access_pmu = 0;
3640	int is_system, is_loaded;
3641
3642	if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3643
3644	state     = ctx->ctx_state;
3645	is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3646	is_system = ctx->ctx_fl_system;
3647	task      = ctx->ctx_task;
3648
3649	if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3650
3651	/*
3652	 * on both UP and SMP, we can only write to the PMC when the task is
3653	 * the owner of the local PMU.
3654	 */
3655	if (is_loaded) {
3656		thread = &task->thread;
3657		/*
3658		 * In system wide and when the context is loaded, access can only happen
3659		 * when the caller is running on the CPU being monitored by the session.
3660		 * It does not have to be the owner (ctx_task) of the context per se.
3661		 */
3662		if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3663			DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3664			return -EBUSY;
3665		}
3666		can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3667	}
3668
3669	/*
3670	 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3671	 * ensuring that no real breakpoint can be installed via this call.
3672	 *
3673	 * IMPORTANT: regs can be NULL in this function
3674	 */
3675
3676	first_time = ctx->ctx_fl_using_dbreg == 0;
3677
3678	/*
3679	 * don't bother if we are loaded and task is being debugged
3680	 */
3681	if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3682		DPRINT(("debug registers already in use for [%d]\n", task_pid_nr(task)));
3683		return -EBUSY;
3684	}
3685
3686	/*
3687	 * check for debug registers in system wide mode
3688	 *
3689	 * If though a check is done in pfm_context_load(),
3690	 * we must repeat it here, in case the registers are
3691	 * written after the context is loaded
3692	 */
3693	if (is_loaded) {
3694		LOCK_PFS(flags);
3695
3696		if (first_time && is_system) {
3697			if (pfm_sessions.pfs_ptrace_use_dbregs)
3698				ret = -EBUSY;
3699			else
3700				pfm_sessions.pfs_sys_use_dbregs++;
3701		}
3702		UNLOCK_PFS(flags);
3703	}
3704
3705	if (ret != 0) return ret;
3706
3707	/*
3708	 * mark ourself as user of the debug registers for
3709	 * perfmon purposes.
3710	 */
3711	ctx->ctx_fl_using_dbreg = 1;
3712
3713	/*
3714 	 * clear hardware registers to make sure we don't
3715 	 * pick up stale state.
3716	 *
3717	 * for a system wide session, we do not use
3718	 * thread.dbr, thread.ibr because this process
3719	 * never leaves the current CPU and the state
3720	 * is shared by all processes running on it
3721 	 */
3722	if (first_time && can_access_pmu) {
3723		DPRINT(("[%d] clearing ibrs, dbrs\n", task_pid_nr(task)));
3724		for (i=0; i < pmu_conf->num_ibrs; i++) {
3725			ia64_set_ibr(i, 0UL);
3726			ia64_dv_serialize_instruction();
3727		}
3728		ia64_srlz_i();
3729		for (i=0; i < pmu_conf->num_dbrs; i++) {
3730			ia64_set_dbr(i, 0UL);
3731			ia64_dv_serialize_data();
3732		}
3733		ia64_srlz_d();
3734	}
3735
3736	/*
3737	 * Now install the values into the registers
3738	 */
3739	for (i = 0; i < count; i++, req++) {
3740
3741		rnum      = req->dbreg_num;
3742		dbreg.val = req->dbreg_value;
3743
3744		ret = -EINVAL;
3745
3746		if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3747			DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3748				  rnum, dbreg.val, mode, i, count));
3749
3750			goto abort_mission;
3751		}
3752
3753		/*
3754		 * make sure we do not install enabled breakpoint
3755		 */
3756		if (rnum & 0x1) {
3757			if (mode == PFM_CODE_RR)
3758				dbreg.ibr.ibr_x = 0;
3759			else
3760				dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3761		}
3762
3763		PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3764
3765		/*
3766		 * Debug registers, just like PMC, can only be modified
3767		 * by a kernel call. Moreover, perfmon() access to those
3768		 * registers are centralized in this routine. The hardware
3769		 * does not modify the value of these registers, therefore,
3770		 * if we save them as they are written, we can avoid having
3771		 * to save them on context switch out. This is made possible
3772		 * by the fact that when perfmon uses debug registers, ptrace()
3773		 * won't be able to modify them concurrently.
3774		 */
3775		if (mode == PFM_CODE_RR) {
3776			CTX_USED_IBR(ctx, rnum);
3777
3778			if (can_access_pmu) {
3779				ia64_set_ibr(rnum, dbreg.val);
3780				ia64_dv_serialize_instruction();
3781			}
3782
3783			ctx->ctx_ibrs[rnum] = dbreg.val;
3784
3785			DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3786				rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3787		} else {
3788			CTX_USED_DBR(ctx, rnum);
3789
3790			if (can_access_pmu) {
3791				ia64_set_dbr(rnum, dbreg.val);
3792				ia64_dv_serialize_data();
3793			}
3794			ctx->ctx_dbrs[rnum] = dbreg.val;
3795
3796			DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3797				rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3798		}
3799	}
3800
3801	return 0;
3802
3803abort_mission:
3804	/*
3805	 * in case it was our first attempt, we undo the global modifications
3806	 */
3807	if (first_time) {
3808		LOCK_PFS(flags);
3809		if (ctx->ctx_fl_system) {
3810			pfm_sessions.pfs_sys_use_dbregs--;
3811		}
3812		UNLOCK_PFS(flags);
3813		ctx->ctx_fl_using_dbreg = 0;
3814	}
3815	/*
3816	 * install error return flag
3817	 */
3818	PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3819
3820	return ret;
3821}
3822
3823static int
3824pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3825{
3826	return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3827}
3828
3829static int
3830pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3831{
3832	return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3833}
3834
3835int
3836pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3837{
3838	pfm_context_t *ctx;
3839
3840	if (req == NULL) return -EINVAL;
3841
3842 	ctx = GET_PMU_CTX();
3843
3844	if (ctx == NULL) return -EINVAL;
3845
3846	/*
3847	 * for now limit to current task, which is enough when calling
3848	 * from overflow handler
3849	 */
3850	if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3851
3852	return pfm_write_ibrs(ctx, req, nreq, regs);
3853}
3854EXPORT_SYMBOL(pfm_mod_write_ibrs);
3855
3856int
3857pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3858{
3859	pfm_context_t *ctx;
3860
3861	if (req == NULL) return -EINVAL;
3862
3863 	ctx = GET_PMU_CTX();
3864
3865	if (ctx == NULL) return -EINVAL;
3866
3867	/*
3868	 * for now limit to current task, which is enough when calling
3869	 * from overflow handler
3870	 */
3871	if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3872
3873	return pfm_write_dbrs(ctx, req, nreq, regs);
3874}
3875EXPORT_SYMBOL(pfm_mod_write_dbrs);
3876
3877
3878static int
3879pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3880{
3881	pfarg_features_t *req = (pfarg_features_t *)arg;
3882
3883	req->ft_version = PFM_VERSION;
3884	return 0;
3885}
3886
3887static int
3888pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3889{
3890	struct pt_regs *tregs;
3891	struct task_struct *task = PFM_CTX_TASK(ctx);
3892	int state, is_system;
3893
3894	state     = ctx->ctx_state;
3895	is_system = ctx->ctx_fl_system;
3896
3897	/*
3898	 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3899	 */
3900	if (state == PFM_CTX_UNLOADED) return -EINVAL;
3901
3902	/*
3903 	 * In system wide and when the context is loaded, access can only happen
3904 	 * when the caller is running on the CPU being monitored by the session.
3905 	 * It does not have to be the owner (ctx_task) of the context per se.
3906 	 */
3907	if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3908		DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3909		return -EBUSY;
3910	}
3911	DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
3912		task_pid_nr(PFM_CTX_TASK(ctx)),
3913		state,
3914		is_system));
3915	/*
3916	 * in system mode, we need to update the PMU directly
3917	 * and the user level state of the caller, which may not
3918	 * necessarily be the creator of the context.
3919	 */
3920	if (is_system) {
3921		/*
3922		 * Update local PMU first
3923		 *
3924		 * disable dcr pp
3925		 */
3926		ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
3927		ia64_srlz_i();
3928
3929		/*
3930		 * update local cpuinfo
3931		 */
3932		PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
3933
3934		/*
3935		 * stop monitoring, does srlz.i
3936		 */
3937		pfm_clear_psr_pp();
3938
3939		/*
3940		 * stop monitoring in the caller
3941		 */
3942		ia64_psr(regs)->pp = 0;
3943
3944		return 0;
3945	}
3946	/*
3947	 * per-task mode
3948	 */
3949
3950	if (task == current) {
3951		/* stop monitoring  at kernel level */
3952		pfm_clear_psr_up();
3953
3954		/*
3955	 	 * stop monitoring at the user level
3956	 	 */
3957		ia64_psr(regs)->up = 0;
3958	} else {
3959		tregs = task_pt_regs(task);
3960
3961		/*
3962	 	 * stop monitoring at the user level
3963	 	 */
3964		ia64_psr(tregs)->up = 0;
3965
3966		/*
3967		 * monitoring disabled in kernel at next reschedule
3968		 */
3969		ctx->ctx_saved_psr_up = 0;
3970		DPRINT(("task=[%d]\n", task_pid_nr(task)));
3971	}
3972	return 0;
3973}
3974
3975
3976static int
3977pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3978{
3979	struct pt_regs *tregs;
3980	int state, is_system;
3981
3982	state     = ctx->ctx_state;
3983	is_system = ctx->ctx_fl_system;
3984
3985	if (state != PFM_CTX_LOADED) return -EINVAL;
3986
3987	/*
3988 	 * In system wide and when the context is loaded, access can only happen
3989 	 * when the caller is running on the CPU being monitored by the session.
3990 	 * It does not have to be the owner (ctx_task) of the context per se.
3991 	 */
3992	if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3993		DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3994		return -EBUSY;
3995	}
3996
3997	/*
3998	 * in system mode, we need to update the PMU directly
3999	 * and the user level state of the caller, which may not
4000	 * necessarily be the creator of the context.
4001	 */
4002	if (is_system) {
4003
4004		/*
4005		 * set user level psr.pp for the caller
4006		 */
4007		ia64_psr(regs)->pp = 1;
4008
4009		/*
4010		 * now update the local PMU and cpuinfo
4011		 */
4012		PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4013
4014		/*
4015		 * start monitoring at kernel level
4016		 */
4017		pfm_set_psr_pp();
4018
4019		/* enable dcr pp */
4020		ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4021		ia64_srlz_i();
4022
4023		return 0;
4024	}
4025
4026	/*
4027	 * per-process mode
4028	 */
4029
4030	if (ctx->ctx_task == current) {
4031
4032		/* start monitoring at kernel level */
4033		pfm_set_psr_up();
4034
4035		/*
4036		 * activate monitoring at user level
4037		 */
4038		ia64_psr(regs)->up = 1;
4039
4040	} else {
4041		tregs = task_pt_regs(ctx->ctx_task);
4042
4043		/*
4044		 * start monitoring at the kernel level the next
4045		 * time the task is scheduled
4046		 */
4047		ctx->ctx_saved_psr_up = IA64_PSR_UP;
4048
4049		/*
4050		 * activate monitoring at user level
4051		 */
4052		ia64_psr(tregs)->up = 1;
4053	}
4054	return 0;
4055}
4056
4057static int
4058pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4059{
4060	pfarg_reg_t *req = (pfarg_reg_t *)arg;
4061	unsigned int cnum;
4062	int i;
4063	int ret = -EINVAL;
4064
4065	for (i = 0; i < count; i++, req++) {
4066
4067		cnum = req->reg_num;
4068
4069		if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4070
4071		req->reg_value = PMC_DFL_VAL(cnum);
4072
4073		PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4074
4075		DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4076	}
4077	return 0;
4078
4079abort_mission:
4080	PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4081	return ret;
4082}
4083
4084static int
4085pfm_check_task_exist(pfm_context_t *ctx)
4086{
4087	struct task_struct *g, *t;
4088	int ret = -ESRCH;
4089
4090	read_lock(&tasklist_lock);
4091
4092	do_each_thread (g, t) {
4093		if (t->thread.pfm_context == ctx) {
4094			ret = 0;
4095			goto out;
4096		}
4097	} while_each_thread (g, t);
4098out:
4099	read_unlock(&tasklist_lock);
4100
4101	DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4102
4103	return ret;
4104}
4105
4106static int
4107pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4108{
4109	struct task_struct *task;
4110	struct thread_struct *thread;
4111	struct pfm_context_t *old;
4112	unsigned long flags;
4113#ifndef CONFIG_SMP
4114	struct task_struct *owner_task = NULL;
4115#endif
4116	pfarg_load_t *req = (pfarg_load_t *)arg;
4117	unsigned long *pmcs_source, *pmds_source;
4118	int the_cpu;
4119	int ret = 0;
4120	int state, is_system, set_dbregs = 0;
4121
4122	state     = ctx->ctx_state;
4123	is_system = ctx->ctx_fl_system;
4124	/*
4125	 * can only load from unloaded or terminated state
4126	 */
4127	if (state != PFM_CTX_UNLOADED) {
4128		DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4129			req->load_pid,
4130			ctx->ctx_state));
4131		return -EBUSY;
4132	}
4133
4134	DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4135
4136	if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4137		DPRINT(("cannot use blocking mode on self\n"));
4138		return -EINVAL;
4139	}
4140
4141	ret = pfm_get_task(ctx, req->load_pid, &task);
4142	if (ret) {
4143		DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4144		return ret;
4145	}
4146
4147	ret = -EINVAL;
4148
4149	/*
4150	 * system wide is self monitoring only
4151	 */
4152	if (is_system && task != current) {
4153		DPRINT(("system wide is self monitoring only load_pid=%d\n",
4154			req->load_pid));
4155		goto error;
4156	}
4157
4158	thread = &task->thread;
4159
4160	ret = 0;
4161	/*
4162	 * cannot load a context which is using range restrictions,
4163	 * into a task that is being debugged.
4164	 */
4165	if (ctx->ctx_fl_using_dbreg) {
4166		if (thread->flags & IA64_THREAD_DBG_VALID) {
4167			ret = -EBUSY;
4168			DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4169			goto error;
4170		}
4171		LOCK_PFS(flags);
4172
4173		if (is_system) {
4174			if (pfm_sessions.pfs_ptrace_use_dbregs) {
4175				DPRINT(("cannot load [%d] dbregs in use\n",
4176							task_pid_nr(task)));
4177				ret = -EBUSY;
4178			} else {
4179				pfm_sessions.pfs_sys_use_dbregs++;
4180				DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task_pid_nr(task), pfm_sessions.pfs_sys_use_dbregs));
4181				set_dbregs = 1;
4182			}
4183		}
4184
4185		UNLOCK_PFS(flags);
4186
4187		if (ret) goto error;
4188	}
4189
4190	/*
4191	 * SMP system-wide monitoring implies self-monitoring.
4192	 *
4193	 * The programming model expects the task to
4194	 * be pinned on a CPU throughout the session.
4195	 * Here we take note of the current CPU at the
4196	 * time the context is loaded. No call from
4197	 * another CPU will be allowed.
4198	 *
4199	 * The pinning via shed_setaffinity()
4200	 * must be done by the calling task prior
4201	 * to this call.
4202	 *
4203	 * systemwide: keep track of CPU this session is supposed to run on
4204	 */
4205	the_cpu = ctx->ctx_cpu = smp_processor_id();
4206
4207	ret = -EBUSY;
4208	/*
4209	 * now reserve the session
4210	 */
4211	ret = pfm_reserve_session(current, is_system, the_cpu);
4212	if (ret) goto error;
4213
4214	/*
4215	 * task is necessarily stopped at this point.
4216	 *
4217	 * If the previous context was zombie, then it got removed in
4218	 * pfm_save_regs(). Therefore we should not see it here.
4219	 * If we see a context, then this is an active context
4220	 *
4221	 * XXX: needs to be atomic
4222	 */
4223	DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4224		thread->pfm_context, ctx));
4225
4226	ret = -EBUSY;
4227	old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4228	if (old != NULL) {
4229		DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4230		goto error_unres;
4231	}
4232
4233	pfm_reset_msgq(ctx);
4234
4235	ctx->ctx_state = PFM_CTX_LOADED;
4236
4237	/*
4238	 * link context to task
4239	 */
4240	ctx->ctx_task = task;
4241
4242	if (is_system) {
4243		/*
4244		 * we load as stopped
4245		 */
4246		PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4247		PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4248
4249		if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4250	} else {
4251		thread->flags |= IA64_THREAD_PM_VALID;
4252	}
4253
4254	/*
4255	 * propagate into thread-state
4256	 */
4257	pfm_copy_pmds(task, ctx);
4258	pfm_copy_pmcs(task, ctx);
4259
4260	pmcs_source = ctx->th_pmcs;
4261	pmds_source = ctx->th_pmds;
4262
4263	/*
4264	 * always the case for system-wide
4265	 */
4266	if (task == current) {
4267
4268		if (is_system == 0) {
4269
4270			/* allow user level control */
4271			ia64_psr(regs)->sp = 0;
4272			DPRINT(("clearing psr.sp for [%d]\n", task_pid_nr(task)));
4273
4274			SET_LAST_CPU(ctx, smp_processor_id());
4275			INC_ACTIVATION();
4276			SET_ACTIVATION(ctx);
4277#ifndef CONFIG_SMP
4278			/*
4279			 * push the other task out, if any
4280			 */
4281			owner_task = GET_PMU_OWNER();
4282			if (owner_task) pfm_lazy_save_regs(owner_task);
4283#endif
4284		}
4285		/*
4286		 * load all PMD from ctx to PMU (as opposed to thread state)
4287		 * restore all PMC from ctx to PMU
4288		 */
4289		pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4290		pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4291
4292		ctx->ctx_reload_pmcs[0] = 0UL;
4293		ctx->ctx_reload_pmds[0] = 0UL;
4294
4295		/*
4296		 * guaranteed safe by earlier check against DBG_VALID
4297		 */
4298		if (ctx->ctx_fl_using_dbreg) {
4299			pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4300			pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4301		}
4302		/*
4303		 * set new ownership
4304		 */
4305		SET_PMU_OWNER(task, ctx);
4306
4307		DPRINT(("context loaded on PMU for [%d]\n", task_pid_nr(task)));
4308	} else {
4309		/*
4310		 * when not current, task MUST be stopped, so this is safe
4311		 */
4312		regs = task_pt_regs(task);
4313
4314		/* force a full reload */
4315		ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4316		SET_LAST_CPU(ctx, -1);
4317
4318		/* initial saved psr (stopped) */
4319		ctx->ctx_saved_psr_up = 0UL;
4320		ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4321	}
4322
4323	ret = 0;
4324
4325error_unres:
4326	if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4327error:
4328	/*
4329	 * we must undo the dbregs setting (for system-wide)
4330	 */
4331	if (ret && set_dbregs) {
4332		LOCK_PFS(flags);
4333		pfm_sessions.pfs_sys_use_dbregs--;
4334		UNLOCK_PFS(flags);
4335	}
4336	/*
4337	 * release task, there is now a link with the context
4338	 */
4339	if (is_system == 0 && task != current) {
4340		pfm_put_task(task);
4341
4342		if (ret == 0) {
4343			ret = pfm_check_task_exist(ctx);
4344			if (ret) {
4345				ctx->ctx_state = PFM_CTX_UNLOADED;
4346				ctx->ctx_task  = NULL;
4347			}
4348		}
4349	}
4350	return ret;
4351}
4352
4353/*
4354 * in this function, we do not need to increase the use count
4355 * for the task via get_task_struct(), because we hold the
4356 * context lock. If the task were to disappear while having
4357 * a context attached, it would go through pfm_exit_thread()
4358 * which also grabs the context lock  and would therefore be blocked
4359 * until we are here.
4360 */
4361static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4362
4363static int
4364pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4365{
4366	struct task_struct *task = PFM_CTX_TASK(ctx);
4367	struct pt_regs *tregs;
4368	int prev_state, is_system;
4369	int ret;
4370
4371	DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task_pid_nr(task) : -1));
4372
4373	prev_state = ctx->ctx_state;
4374	is_system  = ctx->ctx_fl_system;
4375
4376	/*
4377	 * unload only when necessary
4378	 */
4379	if (prev_state == PFM_CTX_UNLOADED) {
4380		DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4381		return 0;
4382	}
4383
4384	/*
4385	 * clear psr and dcr bits
4386	 */
4387	ret = pfm_stop(ctx, NULL, 0, regs);
4388	if (ret) return ret;
4389
4390	ctx->ctx_state = PFM_CTX_UNLOADED;
4391
4392	/*
4393	 * in system mode, we need to update the PMU directly
4394	 * and the user level state of the caller, which may not
4395	 * necessarily be the creator of the context.
4396	 */
4397	if (is_system) {
4398
4399		/*
4400		 * Update cpuinfo
4401		 *
4402		 * local PMU is taken care of in pfm_stop()
4403		 */
4404		PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4405		PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4406
4407		/*
4408		 * save PMDs in context
4409		 * release ownership
4410		 */
4411		pfm_flush_pmds(current, ctx);
4412
4413		/*
4414		 * at this point we are done with the PMU
4415		 * so we can unreserve the resource.
4416		 */
4417		if (prev_state != PFM_CTX_ZOMBIE) 
4418			pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4419
4420		/*
4421		 * disconnect context from task
4422		 */
4423		task->thread.pfm_context = NULL;
4424		/*
4425		 * disconnect task from context
4426		 */
4427		ctx->ctx_task = NULL;
4428
4429		/*
4430		 * There is nothing more to cleanup here.
4431		 */
4432		return 0;
4433	}
4434
4435	/*
4436	 * per-task mode
4437	 */
4438	tregs = task == current ? regs : task_pt_regs(task);
4439
4440	if (task == current) {
4441		/*
4442		 * cancel user level control
4443		 */
4444		ia64_psr(regs)->sp = 1;
4445
4446		DPRINT(("setting psr.sp for [%d]\n", task_pid_nr(task)));
4447	}
4448	/*
4449	 * save PMDs to context
4450	 * release ownership
4451	 */
4452	pfm_flush_pmds(task, ctx);
4453
4454	/*
4455	 * at this point we are done with the PMU
4456	 * so we can unreserve the resource.
4457	 *
4458	 * when state was ZOMBIE, we have already unreserved.
4459	 */
4460	if (prev_state != PFM_CTX_ZOMBIE) 
4461		pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4462
4463	/*
4464	 * reset activation counter and psr
4465	 */
4466	ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4467	SET_LAST_CPU(ctx, -1);
4468
4469	/*
4470	 * PMU state will not be restored
4471	 */
4472	task->thread.flags &= ~IA64_THREAD_PM_VALID;
4473
4474	/*
4475	 * break links between context and task
4476	 */
4477	task->thread.pfm_context  = NULL;
4478	ctx->ctx_task             = NULL;
4479
4480	PFM_SET_WORK_PENDING(task, 0);
4481
4482	ctx->ctx_fl_trap_reason  = PFM_TRAP_REASON_NONE;
4483	ctx->ctx_fl_can_restart  = 0;
4484	ctx->ctx_fl_going_zombie = 0;
4485
4486	DPRINT(("disconnected [%d] from context\n", task_pid_nr(task)));
4487
4488	return 0;
4489}
4490
4491
4492/*
4493 * called only from exit_thread()
4494 * we come here only if the task has a context attached (loaded or masked)
4495 */
4496void
4497pfm_exit_thread(struct task_struct *task)
4498{
4499	pfm_context_t *ctx;
4500	unsigned long flags;
4501	struct pt_regs *regs = task_pt_regs(task);
4502	int ret, state;
4503	int free_ok = 0;
4504
4505	ctx = PFM_GET_CTX(task);
4506
4507	PROTECT_CTX(ctx, flags);
4508
4509	DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task_pid_nr(task)));
4510
4511	state = ctx->ctx_state;
4512	switch(state) {
4513		case PFM_CTX_UNLOADED:
4514			/*
4515	 		 * only comes to this function if pfm_context is not NULL, i.e., cannot
4516			 * be in unloaded state
4517	 		 */
4518			printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task_pid_nr(task));
4519			break;
4520		case PFM_CTX_LOADED:
4521		case PFM_CTX_MASKED:
4522			ret = pfm_context_unload(ctx, NULL, 0, regs);
4523			if (ret) {
4524				printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4525			}
4526			DPRINT(("ctx unloaded for current state was %d\n", state));
4527
4528			pfm_end_notify_user(ctx);
4529			break;
4530		case PFM_CTX_ZOMBIE:
4531			ret = pfm_context_unload(ctx, NULL, 0, regs);
4532			if (ret) {
4533				printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4534			}
4535			free_ok = 1;
4536			break;
4537		default:
4538			printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task_pid_nr(task), state);
4539			break;
4540	}
4541	UNPROTECT_CTX(ctx, flags);
4542
4543	{ u64 psr = pfm_get_psr();
4544	  BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4545	  BUG_ON(GET_PMU_OWNER());
4546	  BUG_ON(ia64_psr(regs)->up);
4547	  BUG_ON(ia64_psr(regs)->pp);
4548	}
4549
4550	/*
4551	 * All memory free operations (especially for vmalloc'ed memory)
4552	 * MUST be done with interrupts ENABLED.
4553	 */
4554	if (free_ok) pfm_context_free(ctx);
4555}
4556
4557/*
4558 * functions MUST be listed in the increasing order of their index (see permfon.h)
4559 */
4560#define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4561#define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4562#define PFM_CMD_PCLRWS	(PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4563#define PFM_CMD_PCLRW	(PFM_CMD_FD|PFM_CMD_ARG_RW)
4564#define PFM_CMD_NONE	{ NULL, "no-cmd", 0, 0, 0, NULL}
4565
4566static pfm_cmd_desc_t pfm_cmd_tab[]={
4567/* 0  */PFM_CMD_NONE,
4568/* 1  */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4569/* 2  */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4570/* 3  */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4571/* 4  */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4572/* 5  */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4573/* 6  */PFM_CMD_NONE,
4574/* 7  */PFM_CMD_NONE,
4575/* 8  */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4576/* 9  */PFM_CMD_NONE,
4577/* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4578/* 11 */PFM_CMD_NONE,
4579/* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4580/* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4581/* 14 */PFM_CMD_NONE,
4582/* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4583/* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4584/* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4585/* 18 */PFM_CMD_NONE,
4586/* 19 */PFM_CMD_NONE,
4587/* 20 */PFM_CMD_NONE,
4588/* 21 */PFM_CMD_NONE,
4589/* 22 */PFM_CMD_NONE,
4590/* 23 */PFM_CMD_NONE,
4591/* 24 */PFM_CMD_NONE,
4592/* 25 */PFM_CMD_NONE,
4593/* 26 */PFM_CMD_NONE,
4594/* 27 */PFM_CMD_NONE,
4595/* 28 */PFM_CMD_NONE,
4596/* 29 */PFM_CMD_NONE,
4597/* 30 */PFM_CMD_NONE,
4598/* 31 */PFM_CMD_NONE,
4599/* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4600/* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4601};
4602#define PFM_CMD_COUNT	(sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4603
4604static int
4605pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4606{
4607	struct task_struct *task;
4608	int state, old_state;
4609
4610recheck:
4611	state = ctx->ctx_state;
4612	task  = ctx->ctx_task;
4613
4614	if (task == NULL) {
4615		DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4616		return 0;
4617	}
4618
4619	DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4620		ctx->ctx_fd,
4621		state,
4622		task_pid_nr(task),
4623		task->state, PFM_CMD_STOPPED(cmd)));
4624
4625	/*
4626	 * self-monitoring always ok.
4627	 *
4628	 * for system-wide the caller can either be the creator of the
4629	 * context (to one to which the context is attached to) OR
4630	 * a task running on the same CPU as the session.
4631	 */
4632	if (task == current || ctx->ctx_fl_system) return 0;
4633
4634	/*
4635	 * we are monitoring another thread
4636	 */
4637	switch(state) {
4638		case PFM_CTX_UNLOADED:
4639			/*
4640			 * if context is UNLOADED we are safe to go
4641			 */
4642			return 0;
4643		case PFM_CTX_ZOMBIE:
4644			/*
4645			 * no command can operate on a zombie context
4646			 */
4647			DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4648			return -EINVAL;
4649		case PFM_CTX_MASKED:
4650			/*
4651			 * PMU state has been saved to software even though
4652			 * the thread may still be running.
4653			 */
4654			if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4655	}
4656
4657	/*
4658	 * context is LOADED or MASKED. Some commands may need to have 
4659	 * the task stopped.
4660	 *
4661	 * We could lift this restriction for UP but it would mean that
4662	 * the user has no guarantee the task would not run between
4663	 * two successive calls to perfmonctl(). That's probably OK.
4664	 * If this user wants to ensure the task does not run, then
4665	 * the task must be stopped.
4666	 */
4667	if (PFM_CMD_STOPPED(cmd)) {
4668		if (!task_is_stopped_or_traced(task)) {
4669			DPRINT(("[%d] task not in stopped state\n", task_pid_nr(task)));
4670			return -EBUSY;
4671		}
4672		/*
4673		 * task is now stopped, wait for ctxsw out
4674		 *
4675		 * This is an interesting point in the code.
4676		 * We need to unprotect the context because
4677		 * the pfm_save_regs() routines needs to grab
4678		 * the same lock. There are danger in doing
4679		 * this because it leaves a window open for
4680		 * another task to get access to the context
4681		 * and possibly change its state. The one thing
4682		 * that is not possible is for the context to disappear
4683		 * because we are protected by the VFS layer, i.e.,
4684		 * get_fd()/put_fd().
4685		 */
4686		old_state = state;
4687
4688		UNPROTECT_CTX(ctx, flags);
4689
4690		wait_task_inactive(task, 0);
4691
4692		PROTECT_CTX(ctx, flags);
4693
4694		/*
4695		 * we must recheck to verify if state has changed
4696		 */
4697		if (ctx->ctx_state != old_state) {
4698			DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4699			goto recheck;
4700		}
4701	}
4702	return 0;
4703}
4704
4705/*
4706 * system-call entry point (must return long)
4707 */
4708asmlinkage long
4709sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4710{
4711	struct fd f = {NULL, 0};
4712	pfm_context_t *ctx = NULL;
4713	unsigned long flags = 0UL;
4714	void *args_k = NULL;
4715	long ret; /* will expand int return types */
4716	size_t base_sz, sz, xtra_sz = 0;
4717	int narg, completed_args = 0, call_made = 0, cmd_flags;
4718	int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4719	int (*getsize)(void *arg, size_t *sz);
4720#define PFM_MAX_ARGSIZE	4096
4721
4722	/*
4723	 * reject any call if perfmon was disabled at initialization
4724	 */
4725	if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4726
4727	if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4728		DPRINT(("invalid cmd=%d\n", cmd));
4729		return -EINVAL;
4730	}
4731
4732	func      = pfm_cmd_tab[cmd].cmd_func;
4733	narg      = pfm_cmd_tab[cmd].cmd_narg;
4734	base_sz   = pfm_cmd_tab[cmd].cmd_argsize;
4735	getsize   = pfm_cmd_tab[cmd].cmd_getsize;
4736	cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4737
4738	if (unlikely(func == NULL)) {
4739		DPRINT(("invalid cmd=%d\n", cmd));
4740		return -EINVAL;
4741	}
4742
4743	DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4744		PFM_CMD_NAME(cmd),
4745		cmd,
4746		narg,
4747		base_sz,
4748		count));
4749
4750	/*
4751	 * check if number of arguments matches what the command expects
4752	 */
4753	if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4754		return -EINVAL;
4755
4756restart_args:
4757	sz = xtra_sz + base_sz*count;
4758	/*
4759	 * limit abuse to min page size
4760	 */
4761	if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4762		printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", task_pid_nr(current), sz);
4763		return -E2BIG;
4764	}
4765
4766	/*
4767	 * allocate default-sized argument buffer
4768	 */
4769	if (likely(count && args_k == NULL)) {
4770		args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4771		if (args_k == NULL) return -ENOMEM;
4772	}
4773
4774	ret = -EFAULT;
4775
4776	/*
4777	 * copy arguments
4778	 *
4779	 * assume sz = 0 for command without parameters
4780	 */
4781	if (sz && copy_from_user(args_k, arg, sz)) {
4782		DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4783		goto error_args;
4784	}
4785
4786	/*
4787	 * check if command supports extra parameters
4788	 */
4789	if (completed_args == 0 && getsize) {
4790		/*
4791		 * get extra parameters size (based on main argument)
4792		 */
4793		ret = (*getsize)(args_k, &xtra_sz);
4794		if (ret) goto error_args;
4795
4796		completed_args = 1;
4797
4798		DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4799
4800		/* retry if necessary */
4801		if (likely(xtra_sz)) goto restart_args;
4802	}
4803
4804	if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4805
4806	ret = -EBADF;
4807
4808	f = fdget(fd);
4809	if (unlikely(f.file == NULL)) {
4810		DPRINT(("invalid fd %d\n", fd));
4811		goto error_args;
4812	}
4813	if (unlikely(PFM_IS_FILE(f.file) == 0)) {
4814		DPRINT(("fd %d not related to perfmon\n", fd));
4815		goto error_args;
4816	}
4817
4818	ctx = f.file->private_data;
4819	if (unlikely(ctx == NULL)) {
4820		DPRINT(("no context for fd %d\n", fd));
4821		goto error_args;
4822	}
4823	prefetch(&ctx->ctx_state);
4824
4825	PROTECT_CTX(ctx, flags);
4826
4827	/*
4828	 * check task is stopped
4829	 */
4830	ret = pfm_check_task_state(ctx, cmd, flags);
4831	if (unlikely(ret)) goto abort_locked;
4832
4833skip_fd:
4834	ret = (*func)(ctx, args_k, count, task_pt_regs(current));
4835
4836	call_made = 1;
4837
4838abort_locked:
4839	if (likely(ctx)) {
4840		DPRINT(("context unlocked\n"));
4841		UNPROTECT_CTX(ctx, flags);
4842	}
4843
4844	/* copy argument back to user, if needed */
4845	if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4846
4847error_args:
4848	if (f.file)
4849		fdput(f);
4850
4851	kfree(args_k);
4852
4853	DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4854
4855	return ret;
4856}
4857
4858static void
4859pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4860{
4861	pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4862	pfm_ovfl_ctrl_t rst_ctrl;
4863	int state;
4864	int ret = 0;
4865
4866	state = ctx->ctx_state;
4867	/*
4868	 * Unlock sampling buffer and reset index atomically
4869	 * XXX: not really needed when blocking
4870	 */
4871	if (CTX_HAS_SMPL(ctx)) {
4872
4873		rst_ctrl.bits.mask_monitoring = 0;
4874		rst_ctrl.bits.reset_ovfl_pmds = 0;
4875
4876		if (state == PFM_CTX_LOADED)
4877			ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4878		else
4879			ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4880	} else {
4881		rst_ctrl.bits.mask_monitoring = 0;
4882		rst_ctrl.bits.reset_ovfl_pmds = 1;
4883	}
4884
4885	if (ret == 0) {
4886		if (rst_ctrl.bits.reset_ovfl_pmds) {
4887			pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4888		}
4889		if (rst_ctrl.bits.mask_monitoring == 0) {
4890			DPRINT(("resuming monitoring\n"));
4891			if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4892		} else {
4893			DPRINT(("stopping monitoring\n"));
4894			//pfm_stop_monitoring(current, regs);
4895		}
4896		ctx->ctx_state = PFM_CTX_LOADED;
4897	}
4898}
4899
4900/*
4901 * context MUST BE LOCKED when calling
4902 * can only be called for current
4903 */
4904static void
4905pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4906{
4907	int ret;
4908
4909	DPRINT(("entering for [%d]\n", task_pid_nr(current)));
4910
4911	ret = pfm_context_unload(ctx, NULL, 0, regs);
4912	if (ret) {
4913		printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", task_pid_nr(current), ret);
4914	}
4915
4916	/*
4917	 * and wakeup controlling task, indicating we are now disconnected
4918	 */
4919	wake_up_interruptible(&ctx->ctx_zombieq);
4920
4921	/*
4922	 * given that context is still locked, the controlling
4923	 * task will only get access when we return from
4924	 * pfm_handle_work().
4925	 */
4926}
4927
4928static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
4929
4930 /*
4931  * pfm_handle_work() can be called with interrupts enabled
4932  * (TIF_NEED_RESCHED) or disabled. The down_interruptible
4933  * call may sleep, therefore we must re-enable interrupts
4934  * to avoid deadlocks. It is safe to do so because this function
4935  * is called ONLY when returning to user level (pUStk=1), in which case
4936  * there is no risk of kernel stack overflow due to deep
4937  * interrupt nesting.
4938  */
4939void
4940pfm_handle_work(void)
4941{
4942	pfm_context_t *ctx;
4943	struct pt_regs *regs;
4944	unsigned long flags, dummy_flags;
4945	unsigned long ovfl_regs;
4946	unsigned int reason;
4947	int ret;
4948
4949	ctx = PFM_GET_CTX(current);
4950	if (ctx == NULL) {
4951		printk(KERN_ERR "perfmon: [%d] has no PFM context\n",
4952			task_pid_nr(current));
4953		return;
4954	}
4955
4956	PROTECT_CTX(ctx, flags);
4957
4958	PFM_SET_WORK_PENDING(current, 0);
4959
4960	regs = task_pt_regs(current);
4961
4962	/*
4963	 * extract reason for being here and clear
4964	 */
4965	reason = ctx->ctx_fl_trap_reason;
4966	ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
4967	ovfl_regs = ctx->ctx_ovfl_regs[0];
4968
4969	DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
4970
4971	/*
4972	 * must be done before we check for simple-reset mode
4973	 */
4974	if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE)
4975		goto do_zombie;
4976
4977	//if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
4978	if (reason == PFM_TRAP_REASON_RESET)
4979		goto skip_blocking;
4980
4981	/*
4982	 * restore interrupt mask to what it was on entry.
4983	 * Could be enabled/diasbled.
4984	 */
4985	UNPROTECT_CTX(ctx, flags);
4986
4987	/*
4988	 * force interrupt enable because of down_interruptible()
4989	 */
4990	local_irq_enable();
4991
4992	DPRINT(("before block sleeping\n"));
4993
4994	/*
4995	 * may go through without blocking on SMP systems
4996	 * if restart has been received already by the time we call down()
4997	 */
4998	ret = wait_for_completion_interruptible(&ctx->ctx_restart_done);
4999
5000	DPRINT(("after block sleeping ret=%d\n", ret));
5001
5002	/*
5003	 * lock context and mask interrupts again
5004	 * We save flags into a dummy because we may have
5005	 * altered interrupts mask compared to entry in this
5006	 * function.
5007	 */
5008	PROTECT_CTX(ctx, dummy_flags);
5009
5010	/*
5011	 * we need to read the ovfl_regs only after wake-up
5012	 * because we may have had pfm_write_pmds() in between
5013	 * and that can changed PMD values and therefore 
5014	 * ovfl_regs is reset for these new PMD values.
5015	 */
5016	ovfl_regs = ctx->ctx_ovfl_regs[0];
5017
5018	if (ctx->ctx_fl_going_zombie) {
5019do_zombie:
5020		DPRINT(("context is zombie, bailing out\n"));
5021		pfm_context_force_terminate(ctx, regs);
5022		goto nothing_to_do;
5023	}
5024	/*
5025	 * in case of interruption of down() we don't restart anything
5026	 */
5027	if (ret < 0)
5028		goto nothing_to_do;
5029
5030skip_blocking:
5031	pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5032	ctx->ctx_ovfl_regs[0] = 0UL;
5033
5034nothing_to_do:
5035	/*
5036	 * restore flags as they were upon entry
5037	 */
5038	UNPROTECT_CTX(ctx, flags);
5039}
5040
5041static int
5042pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5043{
5044	if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5045		DPRINT(("ignoring overflow notification, owner is zombie\n"));
5046		return 0;
5047	}
5048
5049	DPRINT(("waking up somebody\n"));
5050
5051	if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5052
5053	/*
5054	 * safe, we are not in intr handler, nor in ctxsw when
5055	 * we come here
5056	 */
5057	kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5058
5059	return 0;
5060}
5061
5062static int
5063pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5064{
5065	pfm_msg_t *msg = NULL;
5066
5067	if (ctx->ctx_fl_no_msg == 0) {
5068		msg = pfm_get_new_msg(ctx);
5069		if (msg == NULL) {
5070			printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5071			return -1;
5072		}
5073
5074		msg->pfm_ovfl_msg.msg_type         = PFM_MSG_OVFL;
5075		msg->pfm_ovfl_msg.msg_ctx_fd       = ctx->ctx_fd;
5076		msg->pfm_ovfl_msg.msg_active_set   = 0;
5077		msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5078		msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5079		msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5080		msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5081		msg->pfm_ovfl_msg.msg_tstamp       = 0UL;
5082	}
5083
5084	DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5085		msg,
5086		ctx->ctx_fl_no_msg,
5087		ctx->ctx_fd,
5088		ovfl_pmds));
5089
5090	return pfm_notify_user(ctx, msg);
5091}
5092
5093static int
5094pfm_end_notify_user(pfm_context_t *ctx)
5095{
5096	pfm_msg_t *msg;
5097
5098	msg = pfm_get_new_msg(ctx);
5099	if (msg == NULL) {
5100		printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5101		return -1;
5102	}
5103	/* no leak */
5104	memset(msg, 0, sizeof(*msg));
5105
5106	msg->pfm_end_msg.msg_type    = PFM_MSG_END;
5107	msg->pfm_end_msg.msg_ctx_fd  = ctx->ctx_fd;
5108	msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5109
5110	DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5111		msg,
5112		ctx->ctx_fl_no_msg,
5113		ctx->ctx_fd));
5114
5115	return pfm_notify_user(ctx, msg);
5116}
5117
5118/*
5119 * main overflow processing routine.
5120 * it can be called from the interrupt path or explicitly during the context switch code
5121 */
5122static void pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx,
5123				unsigned long pmc0, struct pt_regs *regs)
5124{
5125	pfm_ovfl_arg_t *ovfl_arg;
5126	unsigned long mask;
5127	unsigned long old_val, ovfl_val, new_val;
5128	unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5129	unsigned long tstamp;
5130	pfm_ovfl_ctrl_t	ovfl_ctrl;
5131	unsigned int i, has_smpl;
5132	int must_notify = 0;
5133
5134	if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5135
5136	/*
5137	 * sanity test. Should never happen
5138	 */
5139	if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5140
5141	tstamp   = ia64_get_itc();
5142	mask     = pmc0 >> PMU_FIRST_COUNTER;
5143	ovfl_val = pmu_conf->ovfl_val;
5144	has_smpl = CTX_HAS_SMPL(ctx);
5145
5146	DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5147		     "used_pmds=0x%lx\n",
5148			pmc0,
5149			task ? task_pid_nr(task): -1,
5150			(regs ? regs->cr_iip : 0),
5151			CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5152			ctx->ctx_used_pmds[0]));
5153
5154
5155	/*
5156	 * first we update the virtual counters
5157	 * assume there was a prior ia64_srlz_d() issued
5158	 */
5159	for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5160
5161		/* skip pmd which did not overflow */
5162		if ((mask & 0x1) == 0) continue;
5163
5164		/*
5165		 * Note that the pmd is not necessarily 0 at this point as qualified events
5166		 * may have happened before the PMU was frozen. The residual count is not
5167		 * taken into consideration here but will be with any read of the pmd via
5168		 * pfm_read_pmds().
5169		 */
5170		old_val              = new_val = ctx->ctx_pmds[i].val;
5171		new_val             += 1 + ovfl_val;
5172		ctx->ctx_pmds[i].val = new_val;
5173
5174		/*
5175		 * check for overflow condition
5176		 */
5177		if (likely(old_val > new_val)) {
5178			ovfl_pmds |= 1UL << i;
5179			if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5180		}
5181
5182		DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5183			i,
5184			new_val,
5185			old_val,
5186			ia64_get_pmd(i) & ovfl_val,
5187			ovfl_pmds,
5188			ovfl_notify));
5189	}
5190
5191	/*
5192	 * there was no 64-bit overflow, nothing else to do
5193	 */
5194	if (ovfl_pmds == 0UL) return;
5195
5196	/* 
5197	 * reset all control bits
5198	 */
5199	ovfl_ctrl.val = 0;
5200	reset_pmds    = 0UL;
5201
5202	/*
5203	 * if a sampling format module exists, then we "cache" the overflow by 
5204	 * calling the module's handler() routine.
5205	 */
5206	if (has_smpl) {
5207		unsigned long start_cycles, end_cycles;
5208		unsigned long pmd_mask;
5209		int j, k, ret = 0;
5210		int this_cpu = smp_processor_id();
5211
5212		pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5213		ovfl_arg = &ctx->ctx_ovfl_arg;
5214
5215		prefetch(ctx->ctx_smpl_hdr);
5216
5217		for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5218
5219			mask = 1UL << i;
5220
5221			if ((pmd_mask & 0x1) == 0) continue;
5222
5223			ovfl_arg->ovfl_pmd      = (unsigned char )i;
5224			ovfl_arg->ovfl_notify   = ovfl_notify & mask ? 1 : 0;
5225			ovfl_arg->active_set    = 0;
5226			ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5227			ovfl_arg->smpl_pmds[0]  = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5228
5229			ovfl_arg->pmd_value      = ctx->ctx_pmds[i].val;
5230			ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5231			ovfl_arg->pmd_eventid    = ctx->ctx_pmds[i].eventid;
5232
5233			/*
5234		 	 * copy values of pmds of interest. Sampling format may copy them
5235		 	 * into sampling buffer.
5236		 	 */
5237			if (smpl_pmds) {
5238				for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5239					if ((smpl_pmds & 0x1) == 0) continue;
5240					ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ?  pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5241					DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5242				}
5243			}
5244
5245			pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5246
5247			start_cycles = ia64_get_itc();
5248
5249			/*
5250		 	 * call custom buffer format record (handler) routine
5251		 	 */
5252			ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5253
5254			end_cycles = ia64_get_itc();
5255
5256			/*
5257			 * For those controls, we take the union because they have
5258			 * an all or nothing behavior.
5259			 */
5260			ovfl_ctrl.bits.notify_user     |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5261			ovfl_ctrl.bits.block_task      |= ovfl_arg->ovfl_ctrl.bits.block_task;
5262			ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5263			/*
5264			 * build the bitmask of pmds to reset now
5265			 */
5266			if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5267
5268			pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5269		}
5270		/*
5271		 * when the module cannot handle the rest of the overflows, we abort right here
5272		 */
5273		if (ret && pmd_mask) {
5274			DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5275				pmd_mask<<PMU_FIRST_COUNTER));
5276		}
5277		/*
5278		 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5279		 */
5280		ovfl_pmds &= ~reset_pmds;
5281	} else {
5282		/*
5283		 * when no sampling module is used, then the default
5284		 * is to notify on overflow if requested by user
5285		 */
5286		ovfl_ctrl.bits.notify_user     = ovfl_notify ? 1 : 0;
5287		ovfl_ctrl.bits.block_task      = ovfl_notify ? 1 : 0;
5288		ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5289		ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5290		/*
5291		 * if needed, we reset all overflowed pmds
5292		 */
5293		if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5294	}
5295
5296	DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5297
5298	/*
5299	 * reset the requested PMD registers using the short reset values
5300	 */
5301	if (reset_pmds) {
5302		unsigned long bm = reset_pmds;
5303		pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5304	}
5305
5306	if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5307		/*
5308		 * keep track of what to reset when unblocking
5309		 */
5310		ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5311
5312		/*
5313		 * check for blocking context 
5314		 */
5315		if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5316
5317			ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5318
5319			/*
5320			 * set the perfmon specific checking pending work for the task
5321			 */
5322			PFM_SET_WORK_PENDING(task, 1);
5323
5324			/*
5325			 * when coming from ctxsw, current still points to the
5326			 * previous task, therefore we must work with task and not current.
5327			 */
5328			set_notify_resume(task);
5329		}
5330		/*
5331		 * defer until state is changed (shorten spin window). the context is locked
5332		 * anyway, so the signal receiver would come spin for nothing.
5333		 */
5334		must_notify = 1;
5335	}
5336
5337	DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5338			GET_PMU_OWNER() ? task_pid_nr(GET_PMU_OWNER()) : -1,
5339			PFM_GET_WORK_PENDING(task),
5340			ctx->ctx_fl_trap_reason,
5341			ovfl_pmds,
5342			ovfl_notify,
5343			ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5344	/*
5345	 * in case monitoring must be stopped, we toggle the psr bits
5346	 */
5347	if (ovfl_ctrl.bits.mask_monitoring) {
5348		pfm_mask_monitoring(task);
5349		ctx->ctx_state = PFM_CTX_MASKED;
5350		ctx->ctx_fl_can_restart = 1;
5351	}
5352
5353	/*
5354	 * send notification now
5355	 */
5356	if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5357
5358	return;
5359
5360sanity_check:
5361	printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5362			smp_processor_id(),
5363			task ? task_pid_nr(task) : -1,
5364			pmc0);
5365	return;
5366
5367stop_monitoring:
5368	/*
5369	 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5370	 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5371	 * come here as zombie only if the task is the current task. In which case, we
5372	 * can access the PMU  hardware directly.
5373	 *
5374	 * Note that zombies do have PM_VALID set. So here we do the minimal.
5375	 *
5376	 * In case the context was zombified it could not be reclaimed at the time
5377	 * the monitoring program exited. At this point, the PMU reservation has been
5378	 * returned, the sampiing buffer has been freed. We must convert this call
5379	 * into a spurious interrupt. However, we must also avoid infinite overflows
5380	 * by stopping monitoring for this task. We can only come here for a per-task
5381	 * context. All we need to do is to stop monitoring using the psr bits which
5382	 * are always task private. By re-enabling secure montioring, we ensure that
5383	 * the monitored task will not be able to re-activate monitoring.
5384	 * The task will eventually be context switched out, at which point the context
5385	 * will be reclaimed (that includes releasing ownership of the PMU).
5386	 *
5387	 * So there might be a window of time where the number of per-task session is zero
5388	 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5389	 * context. This is safe because if a per-task session comes in, it will push this one
5390	 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5391	 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5392	 * also push our zombie context out.
5393	 *
5394	 * Overall pretty hairy stuff....
5395	 */
5396	DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task_pid_nr(task): -1));
5397	pfm_clear_psr_up();
5398	ia64_psr(regs)->up = 0;
5399	ia64_psr(regs)->sp = 1;
5400	return;
5401}
5402
5403static int
5404pfm_do_interrupt_handler(void *arg, struct pt_regs *regs)
5405{
5406	struct task_struct *task;
5407	pfm_context_t *ctx;
5408	unsigned long flags;
5409	u64 pmc0;
5410	int this_cpu = smp_processor_id();
5411	int retval = 0;
5412
5413	pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5414
5415	/*
5416	 * srlz.d done before arriving here
5417	 */
5418	pmc0 = ia64_get_pmc(0);
5419
5420	task = GET_PMU_OWNER();
5421	ctx  = GET_PMU_CTX();
5422
5423	/*
5424	 * if we have some pending bits set
5425	 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5426	 */
5427	if (PMC0_HAS_OVFL(pmc0) && task) {
5428		/*
5429		 * we assume that pmc0.fr is always set here
5430		 */
5431
5432		/* sanity check */
5433		if (!ctx) goto report_spurious1;
5434
5435		if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0) 
5436			goto report_spurious2;
5437
5438		PROTECT_CTX_NOPRINT(ctx, flags);
5439
5440		pfm_overflow_handler(task, ctx, pmc0, regs);
5441
5442		UNPROTECT_CTX_NOPRINT(ctx, flags);
5443
5444	} else {
5445		pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5446		retval = -1;
5447	}
5448	/*
5449	 * keep it unfrozen at all times
5450	 */
5451	pfm_unfreeze_pmu();
5452
5453	return retval;
5454
5455report_spurious1:
5456	printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5457		this_cpu, task_pid_nr(task));
5458	pfm_unfreeze_pmu();
5459	return -1;
5460report_spurious2:
5461	printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n", 
5462		this_cpu, 
5463		task_pid_nr(task));
5464	pfm_unfreeze_pmu();
5465	return -1;
5466}
5467
5468static irqreturn_t
5469pfm_interrupt_handler(int irq, void *arg)
5470{
5471	unsigned long start_cycles, total_cycles;
5472	unsigned long min, max;
5473	int this_cpu;
5474	int ret;
5475	struct pt_regs *regs = get_irq_regs();
5476
5477	this_cpu = get_cpu();
5478	if (likely(!pfm_alt_intr_handler)) {
5479		min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5480		max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5481
5482		start_cycles = ia64_get_itc();
5483
5484		ret = pfm_do_interrupt_handler(arg, regs);
5485
5486		total_cycles = ia64_get_itc();
5487
5488		/*
5489		 * don't measure spurious interrupts
5490		 */
5491		if (likely(ret == 0)) {
5492			total_cycles -= start_cycles;
5493
5494			if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5495			if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5496
5497			pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5498		}
5499	}
5500	else {
5501		(*pfm_alt_intr_handler->handler)(irq, arg, regs);
5502	}
5503
5504	put_cpu();
5505	return IRQ_HANDLED;
5506}
5507
5508/*
5509 * /proc/perfmon interface, for debug only
5510 */
5511
5512#define PFM_PROC_SHOW_HEADER	((void *)(long)nr_cpu_ids+1)
5513
5514static void *
5515pfm_proc_start(struct seq_file *m, loff_t *pos)
5516{
5517	if (*pos == 0) {
5518		return PFM_PROC_SHOW_HEADER;
5519	}
5520
5521	while (*pos <= nr_cpu_ids) {
5522		if (cpu_online(*pos - 1)) {
5523			return (void *)*pos;
5524		}
5525		++*pos;
5526	}
5527	return NULL;
5528}
5529
5530static void *
5531pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5532{
5533	++*pos;
5534	return pfm_proc_start(m, pos);
5535}
5536
5537static void
5538pfm_proc_stop(struct seq_file *m, void *v)
5539{
5540}
5541
5542static void
5543pfm_proc_show_header(struct seq_file *m)
5544{
5545	struct list_head * pos;
5546	pfm_buffer_fmt_t * entry;
5547	unsigned long flags;
5548
5549 	seq_printf(m,
5550		"perfmon version           : %u.%u\n"
5551		"model                     : %s\n"
5552		"fastctxsw                 : %s\n"
5553		"expert mode               : %s\n"
5554		"ovfl_mask                 : 0x%lx\n"
5555		"PMU flags                 : 0x%x\n",
5556		PFM_VERSION_MAJ, PFM_VERSION_MIN,
5557		pmu_conf->pmu_name,
5558		pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5559		pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5560		pmu_conf->ovfl_val,
5561		pmu_conf->flags);
5562
5563  	LOCK_PFS(flags);
5564
5565 	seq_printf(m,
5566 		"proc_sessions             : %u\n"
5567 		"sys_sessions              : %u\n"
5568 		"sys_use_dbregs            : %u\n"
5569 		"ptrace_use_dbregs         : %u\n",
5570 		pfm_sessions.pfs_task_sessions,
5571 		pfm_sessions.pfs_sys_sessions,
5572 		pfm_sessions.pfs_sys_use_dbregs,
5573 		pfm_sessions.pfs_ptrace_use_dbregs);
5574
5575  	UNLOCK_PFS(flags);
5576
5577	spin_lock(&pfm_buffer_fmt_lock);
5578
5579	list_for_each(pos, &pfm_buffer_fmt_list) {
5580		entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5581		seq_printf(m, "format                    : %16phD %s\n",
5582			   entry->fmt_uuid, entry->fmt_name);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5583	}
5584	spin_unlock(&pfm_buffer_fmt_lock);
5585
5586}
5587
5588static int
5589pfm_proc_show(struct seq_file *m, void *v)
5590{
5591	unsigned long psr;
5592	unsigned int i;
5593	int cpu;
5594
5595	if (v == PFM_PROC_SHOW_HEADER) {
5596		pfm_proc_show_header(m);
5597		return 0;
5598	}
5599
5600	/* show info for CPU (v - 1) */
5601
5602	cpu = (long)v - 1;
5603	seq_printf(m,
5604		"CPU%-2d overflow intrs      : %lu\n"
5605		"CPU%-2d overflow cycles     : %lu\n"
5606		"CPU%-2d overflow min        : %lu\n"
5607		"CPU%-2d overflow max        : %lu\n"
5608		"CPU%-2d smpl handler calls  : %lu\n"
5609		"CPU%-2d smpl handler cycles : %lu\n"
5610		"CPU%-2d spurious intrs      : %lu\n"
5611		"CPU%-2d replay   intrs      : %lu\n"
5612		"CPU%-2d syst_wide           : %d\n"
5613		"CPU%-2d dcr_pp              : %d\n"
5614		"CPU%-2d exclude idle        : %d\n"
5615		"CPU%-2d owner               : %d\n"
5616		"CPU%-2d context             : %p\n"
5617		"CPU%-2d activations         : %lu\n",
5618		cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5619		cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5620		cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5621		cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5622		cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5623		cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5624		cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5625		cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5626		cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5627		cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5628		cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5629		cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5630		cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5631		cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5632
5633	if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5634
5635		psr = pfm_get_psr();
5636
5637		ia64_srlz_d();
5638
5639		seq_printf(m, 
5640			"CPU%-2d psr                 : 0x%lx\n"
5641			"CPU%-2d pmc0                : 0x%lx\n", 
5642			cpu, psr,
5643			cpu, ia64_get_pmc(0));
5644
5645		for (i=0; PMC_IS_LAST(i) == 0;  i++) {
5646			if (PMC_IS_COUNTING(i) == 0) continue;
5647   			seq_printf(m, 
5648				"CPU%-2d pmc%u                : 0x%lx\n"
5649   				"CPU%-2d pmd%u                : 0x%lx\n", 
5650				cpu, i, ia64_get_pmc(i),
5651				cpu, i, ia64_get_pmd(i));
5652  		}
5653	}
5654	return 0;
5655}
5656
5657const struct seq_operations pfm_seq_ops = {
5658	.start =	pfm_proc_start,
5659 	.next =		pfm_proc_next,
5660 	.stop =		pfm_proc_stop,
5661 	.show =		pfm_proc_show
5662};
5663
 
 
 
 
 
 
 
5664/*
5665 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5666 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5667 * is active or inactive based on mode. We must rely on the value in
5668 * local_cpu_data->pfm_syst_info
5669 */
5670void
5671pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5672{
5673	struct pt_regs *regs;
5674	unsigned long dcr;
5675	unsigned long dcr_pp;
5676
5677	dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5678
5679	/*
5680	 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5681	 * on every CPU, so we can rely on the pid to identify the idle task.
5682	 */
5683	if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5684		regs = task_pt_regs(task);
5685		ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5686		return;
5687	}
5688	/*
5689	 * if monitoring has started
5690	 */
5691	if (dcr_pp) {
5692		dcr = ia64_getreg(_IA64_REG_CR_DCR);
5693		/*
5694		 * context switching in?
5695		 */
5696		if (is_ctxswin) {
5697			/* mask monitoring for the idle task */
5698			ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5699			pfm_clear_psr_pp();
5700			ia64_srlz_i();
5701			return;
5702		}
5703		/*
5704		 * context switching out
5705		 * restore monitoring for next task
5706		 *
5707		 * Due to inlining this odd if-then-else construction generates
5708		 * better code.
5709		 */
5710		ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5711		pfm_set_psr_pp();
5712		ia64_srlz_i();
5713	}
5714}
5715
5716#ifdef CONFIG_SMP
5717
5718static void
5719pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5720{
5721	struct task_struct *task = ctx->ctx_task;
5722
5723	ia64_psr(regs)->up = 0;
5724	ia64_psr(regs)->sp = 1;
5725
5726	if (GET_PMU_OWNER() == task) {
5727		DPRINT(("cleared ownership for [%d]\n",
5728					task_pid_nr(ctx->ctx_task)));
5729		SET_PMU_OWNER(NULL, NULL);
5730	}
5731
5732	/*
5733	 * disconnect the task from the context and vice-versa
5734	 */
5735	PFM_SET_WORK_PENDING(task, 0);
5736
5737	task->thread.pfm_context  = NULL;
5738	task->thread.flags       &= ~IA64_THREAD_PM_VALID;
5739
5740	DPRINT(("force cleanup for [%d]\n",  task_pid_nr(task)));
5741}
5742
5743
5744/*
5745 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5746 */
5747void
5748pfm_save_regs(struct task_struct *task)
5749{
5750	pfm_context_t *ctx;
5751	unsigned long flags;
5752	u64 psr;
5753
5754
5755	ctx = PFM_GET_CTX(task);
5756	if (ctx == NULL) return;
5757
5758	/*
5759 	 * we always come here with interrupts ALREADY disabled by
5760 	 * the scheduler. So we simply need to protect against concurrent
5761	 * access, not CPU concurrency.
5762	 */
5763	flags = pfm_protect_ctx_ctxsw(ctx);
5764
5765	if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5766		struct pt_regs *regs = task_pt_regs(task);
5767
5768		pfm_clear_psr_up();
5769
5770		pfm_force_cleanup(ctx, regs);
5771
5772		BUG_ON(ctx->ctx_smpl_hdr);
5773
5774		pfm_unprotect_ctx_ctxsw(ctx, flags);
5775
5776		pfm_context_free(ctx);
5777		return;
5778	}
5779
5780	/*
5781	 * save current PSR: needed because we modify it
5782	 */
5783	ia64_srlz_d();
5784	psr = pfm_get_psr();
5785
5786	BUG_ON(psr & (IA64_PSR_I));
5787
5788	/*
5789	 * stop monitoring:
5790	 * This is the last instruction which may generate an overflow
5791	 *
5792	 * We do not need to set psr.sp because, it is irrelevant in kernel.
5793	 * It will be restored from ipsr when going back to user level
5794	 */
5795	pfm_clear_psr_up();
5796
5797	/*
5798	 * keep a copy of psr.up (for reload)
5799	 */
5800	ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5801
5802	/*
5803	 * release ownership of this PMU.
5804	 * PM interrupts are masked, so nothing
5805	 * can happen.
5806	 */
5807	SET_PMU_OWNER(NULL, NULL);
5808
5809	/*
5810	 * we systematically save the PMD as we have no
5811	 * guarantee we will be schedule at that same
5812	 * CPU again.
5813	 */
5814	pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5815
5816	/*
5817	 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5818	 * we will need it on the restore path to check
5819	 * for pending overflow.
5820	 */
5821	ctx->th_pmcs[0] = ia64_get_pmc(0);
5822
5823	/*
5824	 * unfreeze PMU if had pending overflows
5825	 */
5826	if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5827
5828	/*
5829	 * finally, allow context access.
5830	 * interrupts will still be masked after this call.
5831	 */
5832	pfm_unprotect_ctx_ctxsw(ctx, flags);
5833}
5834
5835#else /* !CONFIG_SMP */
5836void
5837pfm_save_regs(struct task_struct *task)
5838{
5839	pfm_context_t *ctx;
5840	u64 psr;
5841
5842	ctx = PFM_GET_CTX(task);
5843	if (ctx == NULL) return;
5844
5845	/*
5846	 * save current PSR: needed because we modify it
5847	 */
5848	psr = pfm_get_psr();
5849
5850	BUG_ON(psr & (IA64_PSR_I));
5851
5852	/*
5853	 * stop monitoring:
5854	 * This is the last instruction which may generate an overflow
5855	 *
5856	 * We do not need to set psr.sp because, it is irrelevant in kernel.
5857	 * It will be restored from ipsr when going back to user level
5858	 */
5859	pfm_clear_psr_up();
5860
5861	/*
5862	 * keep a copy of psr.up (for reload)
5863	 */
5864	ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5865}
5866
5867static void
5868pfm_lazy_save_regs (struct task_struct *task)
5869{
5870	pfm_context_t *ctx;
5871	unsigned long flags;
5872
5873	{ u64 psr  = pfm_get_psr();
5874	  BUG_ON(psr & IA64_PSR_UP);
5875	}
5876
5877	ctx = PFM_GET_CTX(task);
5878
5879	/*
5880	 * we need to mask PMU overflow here to
5881	 * make sure that we maintain pmc0 until
5882	 * we save it. overflow interrupts are
5883	 * treated as spurious if there is no
5884	 * owner.
5885	 *
5886	 * XXX: I don't think this is necessary
5887	 */
5888	PROTECT_CTX(ctx,flags);
5889
5890	/*
5891	 * release ownership of this PMU.
5892	 * must be done before we save the registers.
5893	 *
5894	 * after this call any PMU interrupt is treated
5895	 * as spurious.
5896	 */
5897	SET_PMU_OWNER(NULL, NULL);
5898
5899	/*
5900	 * save all the pmds we use
5901	 */
5902	pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5903
5904	/*
5905	 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5906	 * it is needed to check for pended overflow
5907	 * on the restore path
5908	 */
5909	ctx->th_pmcs[0] = ia64_get_pmc(0);
5910
5911	/*
5912	 * unfreeze PMU if had pending overflows
5913	 */
5914	if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5915
5916	/*
5917	 * now get can unmask PMU interrupts, they will
5918	 * be treated as purely spurious and we will not
5919	 * lose any information
5920	 */
5921	UNPROTECT_CTX(ctx,flags);
5922}
5923#endif /* CONFIG_SMP */
5924
5925#ifdef CONFIG_SMP
5926/*
5927 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5928 */
5929void
5930pfm_load_regs (struct task_struct *task)
5931{
5932	pfm_context_t *ctx;
5933	unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
5934	unsigned long flags;
5935	u64 psr, psr_up;
5936	int need_irq_resend;
5937
5938	ctx = PFM_GET_CTX(task);
5939	if (unlikely(ctx == NULL)) return;
5940
5941	BUG_ON(GET_PMU_OWNER());
5942
5943	/*
5944	 * possible on unload
5945	 */
5946	if (unlikely((task->thread.flags & IA64_THREAD_PM_VALID) == 0)) return;
5947
5948	/*
5949 	 * we always come here with interrupts ALREADY disabled by
5950 	 * the scheduler. So we simply need to protect against concurrent
5951	 * access, not CPU concurrency.
5952	 */
5953	flags = pfm_protect_ctx_ctxsw(ctx);
5954	psr   = pfm_get_psr();
5955
5956	need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
5957
5958	BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
5959	BUG_ON(psr & IA64_PSR_I);
5960
5961	if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
5962		struct pt_regs *regs = task_pt_regs(task);
5963
5964		BUG_ON(ctx->ctx_smpl_hdr);
5965
5966		pfm_force_cleanup(ctx, regs);
5967
5968		pfm_unprotect_ctx_ctxsw(ctx, flags);
5969
5970		/*
5971		 * this one (kmalloc'ed) is fine with interrupts disabled
5972		 */
5973		pfm_context_free(ctx);
5974
5975		return;
5976	}
5977
5978	/*
5979	 * we restore ALL the debug registers to avoid picking up
5980	 * stale state.
5981	 */
5982	if (ctx->ctx_fl_using_dbreg) {
5983		pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
5984		pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
5985	}
5986	/*
5987	 * retrieve saved psr.up
5988	 */
5989	psr_up = ctx->ctx_saved_psr_up;
5990
5991	/*
5992	 * if we were the last user of the PMU on that CPU,
5993	 * then nothing to do except restore psr
5994	 */
5995	if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
5996
5997		/*
5998		 * retrieve partial reload masks (due to user modifications)
5999		 */
6000		pmc_mask = ctx->ctx_reload_pmcs[0];
6001		pmd_mask = ctx->ctx_reload_pmds[0];
6002
6003	} else {
6004		/*
6005	 	 * To avoid leaking information to the user level when psr.sp=0,
6006	 	 * we must reload ALL implemented pmds (even the ones we don't use).
6007	 	 * In the kernel we only allow PFM_READ_PMDS on registers which
6008	 	 * we initialized or requested (sampling) so there is no risk there.
6009	 	 */
6010		pmd_mask = pfm_sysctl.fastctxsw ?  ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6011
6012		/*
6013	 	 * ALL accessible PMCs are systematically reloaded, unused registers
6014	 	 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6015	 	 * up stale configuration.
6016	 	 *
6017	 	 * PMC0 is never in the mask. It is always restored separately.
6018	 	 */
6019		pmc_mask = ctx->ctx_all_pmcs[0];
6020	}
6021	/*
6022	 * when context is MASKED, we will restore PMC with plm=0
6023	 * and PMD with stale information, but that's ok, nothing
6024	 * will be captured.
6025	 *
6026	 * XXX: optimize here
6027	 */
6028	if (pmd_mask) pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6029	if (pmc_mask) pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6030
6031	/*
6032	 * check for pending overflow at the time the state
6033	 * was saved.
6034	 */
6035	if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6036		/*
6037		 * reload pmc0 with the overflow information
6038		 * On McKinley PMU, this will trigger a PMU interrupt
6039		 */
6040		ia64_set_pmc(0, ctx->th_pmcs[0]);
6041		ia64_srlz_d();
6042		ctx->th_pmcs[0] = 0UL;
6043
6044		/*
6045		 * will replay the PMU interrupt
6046		 */
6047		if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6048
6049		pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6050	}
6051
6052	/*
6053	 * we just did a reload, so we reset the partial reload fields
6054	 */
6055	ctx->ctx_reload_pmcs[0] = 0UL;
6056	ctx->ctx_reload_pmds[0] = 0UL;
6057
6058	SET_LAST_CPU(ctx, smp_processor_id());
6059
6060	/*
6061	 * dump activation value for this PMU
6062	 */
6063	INC_ACTIVATION();
6064	/*
6065	 * record current activation for this context
6066	 */
6067	SET_ACTIVATION(ctx);
6068
6069	/*
6070	 * establish new ownership. 
6071	 */
6072	SET_PMU_OWNER(task, ctx);
6073
6074	/*
6075	 * restore the psr.up bit. measurement
6076	 * is active again.
6077	 * no PMU interrupt can happen at this point
6078	 * because we still have interrupts disabled.
6079	 */
6080	if (likely(psr_up)) pfm_set_psr_up();
6081
6082	/*
6083	 * allow concurrent access to context
6084	 */
6085	pfm_unprotect_ctx_ctxsw(ctx, flags);
6086}
6087#else /*  !CONFIG_SMP */
6088/*
6089 * reload PMU state for UP kernels
6090 * in 2.5 we come here with interrupts disabled
6091 */
6092void
6093pfm_load_regs (struct task_struct *task)
6094{
6095	pfm_context_t *ctx;
6096	struct task_struct *owner;
6097	unsigned long pmd_mask, pmc_mask;
6098	u64 psr, psr_up;
6099	int need_irq_resend;
6100
6101	owner = GET_PMU_OWNER();
6102	ctx   = PFM_GET_CTX(task);
6103	psr   = pfm_get_psr();
6104
6105	BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6106	BUG_ON(psr & IA64_PSR_I);
6107
6108	/*
6109	 * we restore ALL the debug registers to avoid picking up
6110	 * stale state.
6111	 *
6112	 * This must be done even when the task is still the owner
6113	 * as the registers may have been modified via ptrace()
6114	 * (not perfmon) by the previous task.
6115	 */
6116	if (ctx->ctx_fl_using_dbreg) {
6117		pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6118		pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6119	}
6120
6121	/*
6122	 * retrieved saved psr.up
6123	 */
6124	psr_up = ctx->ctx_saved_psr_up;
6125	need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6126
6127	/*
6128	 * short path, our state is still there, just
6129	 * need to restore psr and we go
6130	 *
6131	 * we do not touch either PMC nor PMD. the psr is not touched
6132	 * by the overflow_handler. So we are safe w.r.t. to interrupt
6133	 * concurrency even without interrupt masking.
6134	 */
6135	if (likely(owner == task)) {
6136		if (likely(psr_up)) pfm_set_psr_up();
6137		return;
6138	}
6139
6140	/*
6141	 * someone else is still using the PMU, first push it out and
6142	 * then we'll be able to install our stuff !
6143	 *
6144	 * Upon return, there will be no owner for the current PMU
6145	 */
6146	if (owner) pfm_lazy_save_regs(owner);
6147
6148	/*
6149	 * To avoid leaking information to the user level when psr.sp=0,
6150	 * we must reload ALL implemented pmds (even the ones we don't use).
6151	 * In the kernel we only allow PFM_READ_PMDS on registers which
6152	 * we initialized or requested (sampling) so there is no risk there.
6153	 */
6154	pmd_mask = pfm_sysctl.fastctxsw ?  ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6155
6156	/*
6157	 * ALL accessible PMCs are systematically reloaded, unused registers
6158	 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6159	 * up stale configuration.
6160	 *
6161	 * PMC0 is never in the mask. It is always restored separately
6162	 */
6163	pmc_mask = ctx->ctx_all_pmcs[0];
6164
6165	pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6166	pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6167
6168	/*
6169	 * check for pending overflow at the time the state
6170	 * was saved.
6171	 */
6172	if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6173		/*
6174		 * reload pmc0 with the overflow information
6175		 * On McKinley PMU, this will trigger a PMU interrupt
6176		 */
6177		ia64_set_pmc(0, ctx->th_pmcs[0]);
6178		ia64_srlz_d();
6179
6180		ctx->th_pmcs[0] = 0UL;
6181
6182		/*
6183		 * will replay the PMU interrupt
6184		 */
6185		if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6186
6187		pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6188	}
6189
6190	/*
6191	 * establish new ownership. 
6192	 */
6193	SET_PMU_OWNER(task, ctx);
6194
6195	/*
6196	 * restore the psr.up bit. measurement
6197	 * is active again.
6198	 * no PMU interrupt can happen at this point
6199	 * because we still have interrupts disabled.
6200	 */
6201	if (likely(psr_up)) pfm_set_psr_up();
6202}
6203#endif /* CONFIG_SMP */
6204
6205/*
6206 * this function assumes monitoring is stopped
6207 */
6208static void
6209pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6210{
6211	u64 pmc0;
6212	unsigned long mask2, val, pmd_val, ovfl_val;
6213	int i, can_access_pmu = 0;
6214	int is_self;
6215
6216	/*
6217	 * is the caller the task being monitored (or which initiated the
6218	 * session for system wide measurements)
6219	 */
6220	is_self = ctx->ctx_task == task ? 1 : 0;
6221
6222	/*
6223	 * can access PMU is task is the owner of the PMU state on the current CPU
6224	 * or if we are running on the CPU bound to the context in system-wide mode
6225	 * (that is not necessarily the task the context is attached to in this mode).
6226	 * In system-wide we always have can_access_pmu true because a task running on an
6227	 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6228	 */
6229	can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6230	if (can_access_pmu) {
6231		/*
6232		 * Mark the PMU as not owned
6233		 * This will cause the interrupt handler to do nothing in case an overflow
6234		 * interrupt was in-flight
6235		 * This also guarantees that pmc0 will contain the final state
6236		 * It virtually gives us full control on overflow processing from that point
6237		 * on.
6238		 */
6239		SET_PMU_OWNER(NULL, NULL);
6240		DPRINT(("releasing ownership\n"));
6241
6242		/*
6243		 * read current overflow status:
6244		 *
6245		 * we are guaranteed to read the final stable state
6246		 */
6247		ia64_srlz_d();
6248		pmc0 = ia64_get_pmc(0); /* slow */
6249
6250		/*
6251		 * reset freeze bit, overflow status information destroyed
6252		 */
6253		pfm_unfreeze_pmu();
6254	} else {
6255		pmc0 = ctx->th_pmcs[0];
6256		/*
6257		 * clear whatever overflow status bits there were
6258		 */
6259		ctx->th_pmcs[0] = 0;
6260	}
6261	ovfl_val = pmu_conf->ovfl_val;
6262	/*
6263	 * we save all the used pmds
6264	 * we take care of overflows for counting PMDs
6265	 *
6266	 * XXX: sampling situation is not taken into account here
6267	 */
6268	mask2 = ctx->ctx_used_pmds[0];
6269
6270	DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6271
6272	for (i = 0; mask2; i++, mask2>>=1) {
6273
6274		/* skip non used pmds */
6275		if ((mask2 & 0x1) == 0) continue;
6276
6277		/*
6278		 * can access PMU always true in system wide mode
6279		 */
6280		val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : ctx->th_pmds[i];
6281
6282		if (PMD_IS_COUNTING(i)) {
6283			DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6284				task_pid_nr(task),
6285				i,
6286				ctx->ctx_pmds[i].val,
6287				val & ovfl_val));
6288
6289			/*
6290			 * we rebuild the full 64 bit value of the counter
6291			 */
6292			val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6293
6294			/*
6295			 * now everything is in ctx_pmds[] and we need
6296			 * to clear the saved context from save_regs() such that
6297			 * pfm_read_pmds() gets the correct value
6298			 */
6299			pmd_val = 0UL;
6300
6301			/*
6302			 * take care of overflow inline
6303			 */
6304			if (pmc0 & (1UL << i)) {
6305				val += 1 + ovfl_val;
6306				DPRINT(("[%d] pmd[%d] overflowed\n", task_pid_nr(task), i));
6307			}
6308		}
6309
6310		DPRINT(("[%d] ctx_pmd[%d]=0x%lx  pmd_val=0x%lx\n", task_pid_nr(task), i, val, pmd_val));
6311
6312		if (is_self) ctx->th_pmds[i] = pmd_val;
6313
6314		ctx->ctx_pmds[i].val = val;
6315	}
6316}
6317
 
 
 
 
 
 
6318static void
6319pfm_alt_save_pmu_state(void *data)
6320{
6321	struct pt_regs *regs;
6322
6323	regs = task_pt_regs(current);
6324
6325	DPRINT(("called\n"));
6326
6327	/*
6328	 * should not be necessary but
6329	 * let's take not risk
6330	 */
6331	pfm_clear_psr_up();
6332	pfm_clear_psr_pp();
6333	ia64_psr(regs)->pp = 0;
6334
6335	/*
6336	 * This call is required
6337	 * May cause a spurious interrupt on some processors
6338	 */
6339	pfm_freeze_pmu();
6340
6341	ia64_srlz_d();
6342}
6343
6344void
6345pfm_alt_restore_pmu_state(void *data)
6346{
6347	struct pt_regs *regs;
6348
6349	regs = task_pt_regs(current);
6350
6351	DPRINT(("called\n"));
6352
6353	/*
6354	 * put PMU back in state expected
6355	 * by perfmon
6356	 */
6357	pfm_clear_psr_up();
6358	pfm_clear_psr_pp();
6359	ia64_psr(regs)->pp = 0;
6360
6361	/*
6362	 * perfmon runs with PMU unfrozen at all times
6363	 */
6364	pfm_unfreeze_pmu();
6365
6366	ia64_srlz_d();
6367}
6368
6369int
6370pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6371{
6372	int ret, i;
6373	int reserve_cpu;
6374
6375	/* some sanity checks */
6376	if (hdl == NULL || hdl->handler == NULL) return -EINVAL;
6377
6378	/* do the easy test first */
6379	if (pfm_alt_intr_handler) return -EBUSY;
6380
6381	/* one at a time in the install or remove, just fail the others */
6382	if (!spin_trylock(&pfm_alt_install_check)) {
6383		return -EBUSY;
6384	}
6385
6386	/* reserve our session */
6387	for_each_online_cpu(reserve_cpu) {
6388		ret = pfm_reserve_session(NULL, 1, reserve_cpu);
6389		if (ret) goto cleanup_reserve;
6390	}
6391
6392	/* save the current system wide pmu states */
6393	on_each_cpu(pfm_alt_save_pmu_state, NULL, 1);
 
 
 
 
6394
6395	/* officially change to the alternate interrupt handler */
6396	pfm_alt_intr_handler = hdl;
6397
6398	spin_unlock(&pfm_alt_install_check);
6399
6400	return 0;
6401
6402cleanup_reserve:
6403	for_each_online_cpu(i) {
6404		/* don't unreserve more than we reserved */
6405		if (i >= reserve_cpu) break;
6406
6407		pfm_unreserve_session(NULL, 1, i);
6408	}
6409
6410	spin_unlock(&pfm_alt_install_check);
6411
6412	return ret;
6413}
6414EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt);
6415
6416int
6417pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6418{
6419	int i;
 
6420
6421	if (hdl == NULL) return -EINVAL;
6422
6423	/* cannot remove someone else's handler! */
6424	if (pfm_alt_intr_handler != hdl) return -EINVAL;
6425
6426	/* one at a time in the install or remove, just fail the others */
6427	if (!spin_trylock(&pfm_alt_install_check)) {
6428		return -EBUSY;
6429	}
6430
6431	pfm_alt_intr_handler = NULL;
6432
6433	on_each_cpu(pfm_alt_restore_pmu_state, NULL, 1);
 
 
 
6434
6435	for_each_online_cpu(i) {
6436		pfm_unreserve_session(NULL, 1, i);
6437	}
6438
6439	spin_unlock(&pfm_alt_install_check);
6440
6441	return 0;
6442}
6443EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt);
6444
6445/*
6446 * perfmon initialization routine, called from the initcall() table
6447 */
6448static int init_pfm_fs(void);
6449
6450static int __init
6451pfm_probe_pmu(void)
6452{
6453	pmu_config_t **p;
6454	int family;
6455
6456	family = local_cpu_data->family;
6457	p      = pmu_confs;
6458
6459	while(*p) {
6460		if ((*p)->probe) {
6461			if ((*p)->probe() == 0) goto found;
6462		} else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6463			goto found;
6464		}
6465		p++;
6466	}
6467	return -1;
6468found:
6469	pmu_conf = *p;
6470	return 0;
6471}
6472
 
 
 
 
 
 
 
6473int __init
6474pfm_init(void)
6475{
6476	unsigned int n, n_counters, i;
6477
6478	printk("perfmon: version %u.%u IRQ %u\n",
6479		PFM_VERSION_MAJ,
6480		PFM_VERSION_MIN,
6481		IA64_PERFMON_VECTOR);
6482
6483	if (pfm_probe_pmu()) {
6484		printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n", 
6485				local_cpu_data->family);
6486		return -ENODEV;
6487	}
6488
6489	/*
6490	 * compute the number of implemented PMD/PMC from the
6491	 * description tables
6492	 */
6493	n = 0;
6494	for (i=0; PMC_IS_LAST(i) == 0;  i++) {
6495		if (PMC_IS_IMPL(i) == 0) continue;
6496		pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6497		n++;
6498	}
6499	pmu_conf->num_pmcs = n;
6500
6501	n = 0; n_counters = 0;
6502	for (i=0; PMD_IS_LAST(i) == 0;  i++) {
6503		if (PMD_IS_IMPL(i) == 0) continue;
6504		pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6505		n++;
6506		if (PMD_IS_COUNTING(i)) n_counters++;
6507	}
6508	pmu_conf->num_pmds      = n;
6509	pmu_conf->num_counters  = n_counters;
6510
6511	/*
6512	 * sanity checks on the number of debug registers
6513	 */
6514	if (pmu_conf->use_rr_dbregs) {
6515		if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6516			printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6517			pmu_conf = NULL;
6518			return -1;
6519		}
6520		if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6521			printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6522			pmu_conf = NULL;
6523			return -1;
6524		}
6525	}
6526
6527	printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6528	       pmu_conf->pmu_name,
6529	       pmu_conf->num_pmcs,
6530	       pmu_conf->num_pmds,
6531	       pmu_conf->num_counters,
6532	       ffz(pmu_conf->ovfl_val));
6533
6534	/* sanity check */
6535	if (pmu_conf->num_pmds >= PFM_NUM_PMD_REGS || pmu_conf->num_pmcs >= PFM_NUM_PMC_REGS) {
6536		printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6537		pmu_conf = NULL;
6538		return -1;
6539	}
6540
6541	/*
6542	 * create /proc/perfmon (mostly for debugging purposes)
6543	 */
6544	perfmon_dir = proc_create_seq("perfmon", S_IRUGO, NULL, &pfm_seq_ops);
6545	if (perfmon_dir == NULL) {
6546		printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6547		pmu_conf = NULL;
6548		return -1;
6549	}
6550
6551	/*
6552	 * create /proc/sys/kernel/perfmon (for debugging purposes)
6553	 */
6554	pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root);
6555
6556	/*
6557	 * initialize all our spinlocks
6558	 */
6559	spin_lock_init(&pfm_sessions.pfs_lock);
6560	spin_lock_init(&pfm_buffer_fmt_lock);
6561
6562	init_pfm_fs();
6563
6564	for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6565
6566	return 0;
6567}
6568
6569__initcall(pfm_init);
6570
6571/*
6572 * this function is called before pfm_init()
6573 */
6574void
6575pfm_init_percpu (void)
6576{
6577	static int first_time=1;
6578	/*
6579	 * make sure no measurement is active
6580	 * (may inherit programmed PMCs from EFI).
6581	 */
6582	pfm_clear_psr_pp();
6583	pfm_clear_psr_up();
6584
6585	/*
6586	 * we run with the PMU not frozen at all times
6587	 */
6588	pfm_unfreeze_pmu();
6589
6590	if (first_time) {
6591		register_percpu_irq(IA64_PERFMON_VECTOR, pfm_interrupt_handler,
6592				    0, "perfmon");
6593		first_time=0;
6594	}
6595
6596	ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6597	ia64_srlz_d();
6598}
6599
6600/*
6601 * used for debug purposes only
6602 */
6603void
6604dump_pmu_state(const char *from)
6605{
6606	struct task_struct *task;
6607	struct pt_regs *regs;
6608	pfm_context_t *ctx;
6609	unsigned long psr, dcr, info, flags;
6610	int i, this_cpu;
6611
6612	local_irq_save(flags);
6613
6614	this_cpu = smp_processor_id();
6615	regs     = task_pt_regs(current);
6616	info     = PFM_CPUINFO_GET();
6617	dcr      = ia64_getreg(_IA64_REG_CR_DCR);
6618
6619	if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6620		local_irq_restore(flags);
6621		return;
6622	}
6623
6624	printk("CPU%d from %s() current [%d] iip=0x%lx %s\n", 
6625		this_cpu, 
6626		from, 
6627		task_pid_nr(current),
6628		regs->cr_iip,
6629		current->comm);
6630
6631	task = GET_PMU_OWNER();
6632	ctx  = GET_PMU_CTX();
6633
6634	printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task_pid_nr(task) : -1, ctx);
6635
6636	psr = pfm_get_psr();
6637
6638	printk("->CPU%d pmc0=0x%lx psr.pp=%d psr.up=%d dcr.pp=%d syst_info=0x%lx user_psr.up=%d user_psr.pp=%d\n", 
6639		this_cpu,
6640		ia64_get_pmc(0),
6641		psr & IA64_PSR_PP ? 1 : 0,
6642		psr & IA64_PSR_UP ? 1 : 0,
6643		dcr & IA64_DCR_PP ? 1 : 0,
6644		info,
6645		ia64_psr(regs)->up,
6646		ia64_psr(regs)->pp);
6647
6648	ia64_psr(regs)->up = 0;
6649	ia64_psr(regs)->pp = 0;
6650
6651	for (i=1; PMC_IS_LAST(i) == 0; i++) {
6652		if (PMC_IS_IMPL(i) == 0) continue;
6653		printk("->CPU%d pmc[%d]=0x%lx thread_pmc[%d]=0x%lx\n", this_cpu, i, ia64_get_pmc(i), i, ctx->th_pmcs[i]);
6654	}
6655
6656	for (i=1; PMD_IS_LAST(i) == 0; i++) {
6657		if (PMD_IS_IMPL(i) == 0) continue;
6658		printk("->CPU%d pmd[%d]=0x%lx thread_pmd[%d]=0x%lx\n", this_cpu, i, ia64_get_pmd(i), i, ctx->th_pmds[i]);
6659	}
6660
6661	if (ctx) {
6662		printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6663				this_cpu,
6664				ctx->ctx_state,
6665				ctx->ctx_smpl_vaddr,
6666				ctx->ctx_smpl_hdr,
6667				ctx->ctx_msgq_head,
6668				ctx->ctx_msgq_tail,
6669				ctx->ctx_saved_psr_up);
6670	}
6671	local_irq_restore(flags);
6672}
6673
6674/*
6675 * called from process.c:copy_thread(). task is new child.
6676 */
6677void
6678pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6679{
6680	struct thread_struct *thread;
6681
6682	DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task_pid_nr(task)));
6683
6684	thread = &task->thread;
6685
6686	/*
6687	 * cut links inherited from parent (current)
6688	 */
6689	thread->pfm_context = NULL;
6690
6691	PFM_SET_WORK_PENDING(task, 0);
6692
6693	/*
6694	 * the psr bits are already set properly in copy_threads()
6695	 */
6696}
6697#else  /* !CONFIG_PERFMON */
6698asmlinkage long
6699sys_perfmonctl (int fd, int cmd, void *arg, int count)
6700{
6701	return -ENOSYS;
6702}
6703#endif /* CONFIG_PERFMON */