Linux Audio

Check our new training course

Loading...
v5.4
   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#ifdef CONFIG_PERFMON
  61/*
  62 * perfmon context state
  63 */
  64#define PFM_CTX_UNLOADED	1	/* context is not loaded onto any task */
  65#define PFM_CTX_LOADED		2	/* context is loaded onto a task */
  66#define PFM_CTX_MASKED		3	/* context is loaded but monitoring is masked due to overflow */
  67#define PFM_CTX_ZOMBIE		4	/* owner of the context is closing it */
  68
  69#define PFM_INVALID_ACTIVATION	(~0UL)
  70
  71#define PFM_NUM_PMC_REGS	64	/* PMC save area for ctxsw */
  72#define PFM_NUM_PMD_REGS	64	/* PMD save area for ctxsw */
  73
  74/*
  75 * depth of message queue
  76 */
  77#define PFM_MAX_MSGS		32
  78#define PFM_CTXQ_EMPTY(g)	((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
  79
  80/*
  81 * type of a PMU register (bitmask).
  82 * bitmask structure:
  83 * 	bit0   : register implemented
  84 * 	bit1   : end marker
  85 * 	bit2-3 : reserved
  86 * 	bit4   : pmc has pmc.pm
  87 * 	bit5   : pmc controls a counter (has pmc.oi), pmd is used as counter
  88 * 	bit6-7 : register type
  89 * 	bit8-31: reserved
  90 */
  91#define PFM_REG_NOTIMPL		0x0 /* not implemented at all */
  92#define PFM_REG_IMPL		0x1 /* register implemented */
  93#define PFM_REG_END		0x2 /* end marker */
  94#define PFM_REG_MONITOR		(0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
  95#define PFM_REG_COUNTING	(0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
  96#define PFM_REG_CONTROL		(0x4<<4|PFM_REG_IMPL) /* PMU control register */
  97#define	PFM_REG_CONFIG		(0x8<<4|PFM_REG_IMPL) /* configuration register */
  98#define PFM_REG_BUFFER	 	(0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
  99
 100#define PMC_IS_LAST(i)	(pmu_conf->pmc_desc[i].type & PFM_REG_END)
 101#define PMD_IS_LAST(i)	(pmu_conf->pmd_desc[i].type & PFM_REG_END)
 102
 103#define PMC_OVFL_NOTIFY(ctx, i)	((ctx)->ctx_pmds[i].flags &  PFM_REGFL_OVFL_NOTIFY)
 104
 105/* i assumed unsigned */
 106#define PMC_IS_IMPL(i)	  (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
 107#define PMD_IS_IMPL(i)	  (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
 108
 109/* XXX: these assume that register i is implemented */
 110#define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
 111#define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
 112#define PMC_IS_MONITOR(i)  ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR)  == PFM_REG_MONITOR)
 113#define PMC_IS_CONTROL(i)  ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL)  == PFM_REG_CONTROL)
 114
 115#define PMC_DFL_VAL(i)     pmu_conf->pmc_desc[i].default_value
 116#define PMC_RSVD_MASK(i)   pmu_conf->pmc_desc[i].reserved_mask
 117#define PMD_PMD_DEP(i)	   pmu_conf->pmd_desc[i].dep_pmd[0]
 118#define PMC_PMD_DEP(i)	   pmu_conf->pmc_desc[i].dep_pmd[0]
 119
 120#define PFM_NUM_IBRS	  IA64_NUM_DBG_REGS
 121#define PFM_NUM_DBRS	  IA64_NUM_DBG_REGS
 122
 123#define CTX_OVFL_NOBLOCK(c)	((c)->ctx_fl_block == 0)
 124#define CTX_HAS_SMPL(c)		((c)->ctx_fl_is_sampling)
 125#define PFM_CTX_TASK(h)		(h)->ctx_task
 126
 127#define PMU_PMC_OI		5 /* position of pmc.oi bit */
 128
 129/* XXX: does not support more than 64 PMDs */
 130#define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
 131#define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
 132
 133#define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
 134
 135#define CTX_USED_IBR(ctx,n) 	(ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
 136#define CTX_USED_DBR(ctx,n) 	(ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
 137#define CTX_USES_DBREGS(ctx)	(((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
 138#define PFM_CODE_RR	0	/* requesting code range restriction */
 139#define PFM_DATA_RR	1	/* requestion data range restriction */
 140
 141#define PFM_CPUINFO_CLEAR(v)	pfm_get_cpu_var(pfm_syst_info) &= ~(v)
 142#define PFM_CPUINFO_SET(v)	pfm_get_cpu_var(pfm_syst_info) |= (v)
 143#define PFM_CPUINFO_GET()	pfm_get_cpu_var(pfm_syst_info)
 144
 145#define RDEP(x)	(1UL<<(x))
 146
 147/*
 148 * context protection macros
 149 * in SMP:
 150 * 	- we need to protect against CPU concurrency (spin_lock)
 151 * 	- we need to protect against PMU overflow interrupts (local_irq_disable)
 152 * in UP:
 153 * 	- we need to protect against PMU overflow interrupts (local_irq_disable)
 154 *
 155 * spin_lock_irqsave()/spin_unlock_irqrestore():
 156 * 	in SMP: local_irq_disable + spin_lock
 157 * 	in UP : local_irq_disable
 158 *
 159 * spin_lock()/spin_lock():
 160 * 	in UP : removed automatically
 161 * 	in SMP: protect against context accesses from other CPU. interrupts
 162 * 	        are not masked. This is useful for the PMU interrupt handler
 163 * 	        because we know we will not get PMU concurrency in that code.
 164 */
 165#define PROTECT_CTX(c, f) \
 166	do {  \
 167		DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, task_pid_nr(current))); \
 168		spin_lock_irqsave(&(c)->ctx_lock, f); \
 169		DPRINT(("spinlocked ctx %p  by [%d]\n", c, task_pid_nr(current))); \
 170	} while(0)
 171
 172#define UNPROTECT_CTX(c, f) \
 173	do { \
 174		DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, task_pid_nr(current))); \
 175		spin_unlock_irqrestore(&(c)->ctx_lock, f); \
 176	} while(0)
 177
 178#define PROTECT_CTX_NOPRINT(c, f) \
 179	do {  \
 180		spin_lock_irqsave(&(c)->ctx_lock, f); \
 181	} while(0)
 182
 183
 184#define UNPROTECT_CTX_NOPRINT(c, f) \
 185	do { \
 186		spin_unlock_irqrestore(&(c)->ctx_lock, f); \
 187	} while(0)
 188
 189
 190#define PROTECT_CTX_NOIRQ(c) \
 191	do {  \
 192		spin_lock(&(c)->ctx_lock); \
 193	} while(0)
 194
 195#define UNPROTECT_CTX_NOIRQ(c) \
 196	do { \
 197		spin_unlock(&(c)->ctx_lock); \
 198	} while(0)
 199
 200
 201#ifdef CONFIG_SMP
 202
 203#define GET_ACTIVATION()	pfm_get_cpu_var(pmu_activation_number)
 204#define INC_ACTIVATION()	pfm_get_cpu_var(pmu_activation_number)++
 205#define SET_ACTIVATION(c)	(c)->ctx_last_activation = GET_ACTIVATION()
 206
 207#else /* !CONFIG_SMP */
 208#define SET_ACTIVATION(t) 	do {} while(0)
 209#define GET_ACTIVATION(t) 	do {} while(0)
 210#define INC_ACTIVATION(t) 	do {} while(0)
 211#endif /* CONFIG_SMP */
 212
 213#define SET_PMU_OWNER(t, c)	do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
 214#define GET_PMU_OWNER()		pfm_get_cpu_var(pmu_owner)
 215#define GET_PMU_CTX()		pfm_get_cpu_var(pmu_ctx)
 216
 217#define LOCK_PFS(g)	    	spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
 218#define UNLOCK_PFS(g)	    	spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
 219
 220#define PFM_REG_RETFLAG_SET(flags, val)	do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
 221
 222/*
 223 * cmp0 must be the value of pmc0
 224 */
 225#define PMC0_HAS_OVFL(cmp0)  (cmp0 & ~0x1UL)
 226
 227#define PFMFS_MAGIC 0xa0b4d889
 228
 229/*
 230 * debugging
 231 */
 232#define PFM_DEBUGGING 1
 233#ifdef PFM_DEBUGGING
 234#define DPRINT(a) \
 235	do { \
 236		if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
 237	} while (0)
 238
 239#define DPRINT_ovfl(a) \
 240	do { \
 241		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; } \
 242	} while (0)
 243#endif
 244
 245/*
 246 * 64-bit software counter structure
 247 *
 248 * the next_reset_type is applied to the next call to pfm_reset_regs()
 249 */
 250typedef struct {
 251	unsigned long	val;		/* virtual 64bit counter value */
 252	unsigned long	lval;		/* last reset value */
 253	unsigned long	long_reset;	/* reset value on sampling overflow */
 254	unsigned long	short_reset;    /* reset value on overflow */
 255	unsigned long	reset_pmds[4];  /* which other pmds to reset when this counter overflows */
 256	unsigned long	smpl_pmds[4];   /* which pmds are accessed when counter overflow */
 257	unsigned long	seed;		/* seed for random-number generator */
 258	unsigned long	mask;		/* mask for random-number generator */
 259	unsigned int 	flags;		/* notify/do not notify */
 260	unsigned long	eventid;	/* overflow event identifier */
 261} pfm_counter_t;
 262
 263/*
 264 * context flags
 265 */
 266typedef struct {
 267	unsigned int block:1;		/* when 1, task will blocked on user notifications */
 268	unsigned int system:1;		/* do system wide monitoring */
 269	unsigned int using_dbreg:1;	/* using range restrictions (debug registers) */
 270	unsigned int is_sampling:1;	/* true if using a custom format */
 271	unsigned int excl_idle:1;	/* exclude idle task in system wide session */
 272	unsigned int going_zombie:1;	/* context is zombie (MASKED+blocking) */
 273	unsigned int trap_reason:2;	/* reason for going into pfm_handle_work() */
 274	unsigned int no_msg:1;		/* no message sent on overflow */
 275	unsigned int can_restart:1;	/* allowed to issue a PFM_RESTART */
 276	unsigned int reserved:22;
 277} pfm_context_flags_t;
 278
 279#define PFM_TRAP_REASON_NONE		0x0	/* default value */
 280#define PFM_TRAP_REASON_BLOCK		0x1	/* we need to block on overflow */
 281#define PFM_TRAP_REASON_RESET		0x2	/* we need to reset PMDs */
 282
 283
 284/*
 285 * perfmon context: encapsulates all the state of a monitoring session
 286 */
 287
 288typedef struct pfm_context {
 289	spinlock_t		ctx_lock;		/* context protection */
 290
 291	pfm_context_flags_t	ctx_flags;		/* bitmask of flags  (block reason incl.) */
 292	unsigned int		ctx_state;		/* state: active/inactive (no bitfield) */
 293
 294	struct task_struct 	*ctx_task;		/* task to which context is attached */
 295
 296	unsigned long		ctx_ovfl_regs[4];	/* which registers overflowed (notification) */
 297
 298	struct completion	ctx_restart_done;  	/* use for blocking notification mode */
 299
 300	unsigned long		ctx_used_pmds[4];	/* bitmask of PMD used            */
 301	unsigned long		ctx_all_pmds[4];	/* bitmask of all accessible PMDs */
 302	unsigned long		ctx_reload_pmds[4];	/* bitmask of force reload PMD on ctxsw in */
 303
 304	unsigned long		ctx_all_pmcs[4];	/* bitmask of all accessible PMCs */
 305	unsigned long		ctx_reload_pmcs[4];	/* bitmask of force reload PMC on ctxsw in */
 306	unsigned long		ctx_used_monitors[4];	/* bitmask of monitor PMC being used */
 307
 308	unsigned long		ctx_pmcs[PFM_NUM_PMC_REGS];	/*  saved copies of PMC values */
 309
 310	unsigned int		ctx_used_ibrs[1];		/* bitmask of used IBR (speedup ctxsw in) */
 311	unsigned int		ctx_used_dbrs[1];		/* bitmask of used DBR (speedup ctxsw in) */
 312	unsigned long		ctx_dbrs[IA64_NUM_DBG_REGS];	/* DBR values (cache) when not loaded */
 313	unsigned long		ctx_ibrs[IA64_NUM_DBG_REGS];	/* IBR values (cache) when not loaded */
 314
 315	pfm_counter_t		ctx_pmds[PFM_NUM_PMD_REGS]; /* software state for PMDS */
 316
 317	unsigned long		th_pmcs[PFM_NUM_PMC_REGS];	/* PMC thread save state */
 318	unsigned long		th_pmds[PFM_NUM_PMD_REGS];	/* PMD thread save state */
 319
 320	unsigned long		ctx_saved_psr_up;	/* only contains psr.up value */
 321
 322	unsigned long		ctx_last_activation;	/* context last activation number for last_cpu */
 323	unsigned int		ctx_last_cpu;		/* CPU id of current or last CPU used (SMP only) */
 324	unsigned int		ctx_cpu;		/* cpu to which perfmon is applied (system wide) */
 325
 326	int			ctx_fd;			/* file descriptor used my this context */
 327	pfm_ovfl_arg_t		ctx_ovfl_arg;		/* argument to custom buffer format handler */
 328
 329	pfm_buffer_fmt_t	*ctx_buf_fmt;		/* buffer format callbacks */
 330	void			*ctx_smpl_hdr;		/* points to sampling buffer header kernel vaddr */
 331	unsigned long		ctx_smpl_size;		/* size of sampling buffer */
 332	void			*ctx_smpl_vaddr;	/* user level virtual address of smpl buffer */
 333
 334	wait_queue_head_t 	ctx_msgq_wait;
 335	pfm_msg_t		ctx_msgq[PFM_MAX_MSGS];
 336	int			ctx_msgq_head;
 337	int			ctx_msgq_tail;
 338	struct fasync_struct	*ctx_async_queue;
 339
 340	wait_queue_head_t 	ctx_zombieq;		/* termination cleanup wait queue */
 341} pfm_context_t;
 342
 343/*
 344 * magic number used to verify that structure is really
 345 * a perfmon context
 346 */
 347#define PFM_IS_FILE(f)		((f)->f_op == &pfm_file_ops)
 348
 349#define PFM_GET_CTX(t)	 	((pfm_context_t *)(t)->thread.pfm_context)
 350
 351#ifdef CONFIG_SMP
 352#define SET_LAST_CPU(ctx, v)	(ctx)->ctx_last_cpu = (v)
 353#define GET_LAST_CPU(ctx)	(ctx)->ctx_last_cpu
 354#else
 355#define SET_LAST_CPU(ctx, v)	do {} while(0)
 356#define GET_LAST_CPU(ctx)	do {} while(0)
 357#endif
 358
 359
 360#define ctx_fl_block		ctx_flags.block
 361#define ctx_fl_system		ctx_flags.system
 362#define ctx_fl_using_dbreg	ctx_flags.using_dbreg
 363#define ctx_fl_is_sampling	ctx_flags.is_sampling
 364#define ctx_fl_excl_idle	ctx_flags.excl_idle
 365#define ctx_fl_going_zombie	ctx_flags.going_zombie
 366#define ctx_fl_trap_reason	ctx_flags.trap_reason
 367#define ctx_fl_no_msg		ctx_flags.no_msg
 368#define ctx_fl_can_restart	ctx_flags.can_restart
 369
 370#define PFM_SET_WORK_PENDING(t, v)	do { (t)->thread.pfm_needs_checking = v; } while(0);
 371#define PFM_GET_WORK_PENDING(t)		(t)->thread.pfm_needs_checking
 372
 373/*
 374 * global information about all sessions
 375 * mostly used to synchronize between system wide and per-process
 376 */
 377typedef struct {
 378	spinlock_t		pfs_lock;		   /* lock the structure */
 379
 380	unsigned int		pfs_task_sessions;	   /* number of per task sessions */
 381	unsigned int		pfs_sys_sessions;	   /* number of per system wide sessions */
 382	unsigned int		pfs_sys_use_dbregs;	   /* incremented when a system wide session uses debug regs */
 383	unsigned int		pfs_ptrace_use_dbregs;	   /* incremented when a process uses debug regs */
 384	struct task_struct	*pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
 385} pfm_session_t;
 386
 387/*
 388 * information about a PMC or PMD.
 389 * dep_pmd[]: a bitmask of dependent PMD registers
 390 * dep_pmc[]: a bitmask of dependent PMC registers
 391 */
 392typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
 393typedef struct {
 394	unsigned int		type;
 395	int			pm_pos;
 396	unsigned long		default_value;	/* power-on default value */
 397	unsigned long		reserved_mask;	/* bitmask of reserved bits */
 398	pfm_reg_check_t		read_check;
 399	pfm_reg_check_t		write_check;
 400	unsigned long		dep_pmd[4];
 401	unsigned long		dep_pmc[4];
 402} pfm_reg_desc_t;
 403
 404/* assume cnum is a valid monitor */
 405#define PMC_PM(cnum, val)	(((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
 406
 407/*
 408 * This structure is initialized at boot time and contains
 409 * a description of the PMU main characteristics.
 410 *
 411 * If the probe function is defined, detection is based
 412 * on its return value: 
 413 * 	- 0 means recognized PMU
 414 * 	- anything else means not supported
 415 * When the probe function is not defined, then the pmu_family field
 416 * is used and it must match the host CPU family such that:
 417 * 	- cpu->family & config->pmu_family != 0
 418 */
 419typedef struct {
 420	unsigned long  ovfl_val;	/* overflow value for counters */
 421
 422	pfm_reg_desc_t *pmc_desc;	/* detailed PMC register dependencies descriptions */
 423	pfm_reg_desc_t *pmd_desc;	/* detailed PMD register dependencies descriptions */
 424
 425	unsigned int   num_pmcs;	/* number of PMCS: computed at init time */
 426	unsigned int   num_pmds;	/* number of PMDS: computed at init time */
 427	unsigned long  impl_pmcs[4];	/* bitmask of implemented PMCS */
 428	unsigned long  impl_pmds[4];	/* bitmask of implemented PMDS */
 429
 430	char	      *pmu_name;	/* PMU family name */
 431	unsigned int  pmu_family;	/* cpuid family pattern used to identify pmu */
 432	unsigned int  flags;		/* pmu specific flags */
 433	unsigned int  num_ibrs;		/* number of IBRS: computed at init time */
 434	unsigned int  num_dbrs;		/* number of DBRS: computed at init time */
 435	unsigned int  num_counters;	/* PMC/PMD counting pairs : computed at init time */
 436	int           (*probe)(void);   /* customized probe routine */
 437	unsigned int  use_rr_dbregs:1;	/* set if debug registers used for range restriction */
 438} pmu_config_t;
 439/*
 440 * PMU specific flags
 441 */
 442#define PFM_PMU_IRQ_RESEND	1	/* PMU needs explicit IRQ resend */
 443
 444/*
 445 * debug register related type definitions
 446 */
 447typedef struct {
 448	unsigned long ibr_mask:56;
 449	unsigned long ibr_plm:4;
 450	unsigned long ibr_ig:3;
 451	unsigned long ibr_x:1;
 452} ibr_mask_reg_t;
 453
 454typedef struct {
 455	unsigned long dbr_mask:56;
 456	unsigned long dbr_plm:4;
 457	unsigned long dbr_ig:2;
 458	unsigned long dbr_w:1;
 459	unsigned long dbr_r:1;
 460} dbr_mask_reg_t;
 461
 462typedef union {
 463	unsigned long  val;
 464	ibr_mask_reg_t ibr;
 465	dbr_mask_reg_t dbr;
 466} dbreg_t;
 467
 468
 469/*
 470 * perfmon command descriptions
 471 */
 472typedef struct {
 473	int		(*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
 474	char		*cmd_name;
 475	int		cmd_flags;
 476	unsigned int	cmd_narg;
 477	size_t		cmd_argsize;
 478	int		(*cmd_getsize)(void *arg, size_t *sz);
 479} pfm_cmd_desc_t;
 480
 481#define PFM_CMD_FD		0x01	/* command requires a file descriptor */
 482#define PFM_CMD_ARG_READ	0x02	/* command must read argument(s) */
 483#define PFM_CMD_ARG_RW		0x04	/* command must read/write argument(s) */
 484#define PFM_CMD_STOP		0x08	/* command does not work on zombie context */
 485
 486
 487#define PFM_CMD_NAME(cmd)	pfm_cmd_tab[(cmd)].cmd_name
 488#define PFM_CMD_READ_ARG(cmd)	(pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
 489#define PFM_CMD_RW_ARG(cmd)	(pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
 490#define PFM_CMD_USE_FD(cmd)	(pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
 491#define PFM_CMD_STOPPED(cmd)	(pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
 492
 493#define PFM_CMD_ARG_MANY	-1 /* cannot be zero */
 494
 495typedef struct {
 496	unsigned long pfm_spurious_ovfl_intr_count;	/* keep track of spurious ovfl interrupts */
 497	unsigned long pfm_replay_ovfl_intr_count;	/* keep track of replayed ovfl interrupts */
 498	unsigned long pfm_ovfl_intr_count; 		/* keep track of ovfl interrupts */
 499	unsigned long pfm_ovfl_intr_cycles;		/* cycles spent processing ovfl interrupts */
 500	unsigned long pfm_ovfl_intr_cycles_min;		/* min cycles spent processing ovfl interrupts */
 501	unsigned long pfm_ovfl_intr_cycles_max;		/* max cycles spent processing ovfl interrupts */
 502	unsigned long pfm_smpl_handler_calls;
 503	unsigned long pfm_smpl_handler_cycles;
 504	char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
 505} pfm_stats_t;
 506
 507/*
 508 * perfmon internal variables
 509 */
 510static pfm_stats_t		pfm_stats[NR_CPUS];
 511static pfm_session_t		pfm_sessions;	/* global sessions information */
 512
 513static DEFINE_SPINLOCK(pfm_alt_install_check);
 514static pfm_intr_handler_desc_t  *pfm_alt_intr_handler;
 515
 516static struct proc_dir_entry 	*perfmon_dir;
 517static pfm_uuid_t		pfm_null_uuid = {0,};
 518
 519static spinlock_t		pfm_buffer_fmt_lock;
 520static LIST_HEAD(pfm_buffer_fmt_list);
 521
 522static pmu_config_t		*pmu_conf;
 523
 524/* sysctl() controls */
 525pfm_sysctl_t pfm_sysctl;
 526EXPORT_SYMBOL(pfm_sysctl);
 527
 528static struct ctl_table pfm_ctl_table[] = {
 529	{
 530		.procname	= "debug",
 531		.data		= &pfm_sysctl.debug,
 532		.maxlen		= sizeof(int),
 533		.mode		= 0666,
 534		.proc_handler	= proc_dointvec,
 535	},
 536	{
 537		.procname	= "debug_ovfl",
 538		.data		= &pfm_sysctl.debug_ovfl,
 539		.maxlen		= sizeof(int),
 540		.mode		= 0666,
 541		.proc_handler	= proc_dointvec,
 542	},
 543	{
 544		.procname	= "fastctxsw",
 545		.data		= &pfm_sysctl.fastctxsw,
 546		.maxlen		= sizeof(int),
 547		.mode		= 0600,
 548		.proc_handler	= proc_dointvec,
 549	},
 550	{
 551		.procname	= "expert_mode",
 552		.data		= &pfm_sysctl.expert_mode,
 553		.maxlen		= sizeof(int),
 554		.mode		= 0600,
 555		.proc_handler	= proc_dointvec,
 556	},
 557	{}
 558};
 559static struct ctl_table pfm_sysctl_dir[] = {
 560	{
 561		.procname	= "perfmon",
 562		.mode		= 0555,
 563		.child		= pfm_ctl_table,
 564	},
 565 	{}
 566};
 567static struct ctl_table pfm_sysctl_root[] = {
 568	{
 569		.procname	= "kernel",
 570		.mode		= 0555,
 571		.child		= pfm_sysctl_dir,
 572	},
 573 	{}
 574};
 575static struct ctl_table_header *pfm_sysctl_header;
 576
 577static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
 578
 579#define pfm_get_cpu_var(v)		__ia64_per_cpu_var(v)
 580#define pfm_get_cpu_data(a,b)		per_cpu(a, b)
 581
 582static inline void
 583pfm_put_task(struct task_struct *task)
 584{
 585	if (task != current) put_task_struct(task);
 586}
 587
 
 
 
 
 
 
 
 
 
 
 
 588static inline unsigned long
 589pfm_protect_ctx_ctxsw(pfm_context_t *x)
 590{
 591	spin_lock(&(x)->ctx_lock);
 592	return 0UL;
 593}
 594
 595static inline void
 596pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
 597{
 598	spin_unlock(&(x)->ctx_lock);
 599}
 600
 601/* forward declaration */
 602static const struct dentry_operations pfmfs_dentry_operations;
 603
 604static int pfmfs_init_fs_context(struct fs_context *fc)
 
 605{
 606	struct pseudo_fs_context *ctx = init_pseudo(fc, PFMFS_MAGIC);
 607	if (!ctx)
 608		return -ENOMEM;
 609	ctx->dops = &pfmfs_dentry_operations;
 610	return 0;
 611}
 612
 613static struct file_system_type pfm_fs_type = {
 614	.name			= "pfmfs",
 615	.init_fs_context	= pfmfs_init_fs_context,
 616	.kill_sb		= kill_anon_super,
 617};
 618MODULE_ALIAS_FS("pfmfs");
 619
 620DEFINE_PER_CPU(unsigned long, pfm_syst_info);
 621DEFINE_PER_CPU(struct task_struct *, pmu_owner);
 622DEFINE_PER_CPU(pfm_context_t  *, pmu_ctx);
 623DEFINE_PER_CPU(unsigned long, pmu_activation_number);
 624EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info);
 625
 626
 627/* forward declaration */
 628static const struct file_operations pfm_file_ops;
 629
 630/*
 631 * forward declarations
 632 */
 633#ifndef CONFIG_SMP
 634static void pfm_lazy_save_regs (struct task_struct *ta);
 635#endif
 636
 637void dump_pmu_state(const char *);
 638static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
 639
 640#include "perfmon_itanium.h"
 641#include "perfmon_mckinley.h"
 642#include "perfmon_montecito.h"
 643#include "perfmon_generic.h"
 644
 645static pmu_config_t *pmu_confs[]={
 646	&pmu_conf_mont,
 647	&pmu_conf_mck,
 648	&pmu_conf_ita,
 649	&pmu_conf_gen, /* must be last */
 650	NULL
 651};
 652
 653
 654static int pfm_end_notify_user(pfm_context_t *ctx);
 655
 656static inline void
 657pfm_clear_psr_pp(void)
 658{
 659	ia64_rsm(IA64_PSR_PP);
 660	ia64_srlz_i();
 661}
 662
 663static inline void
 664pfm_set_psr_pp(void)
 665{
 666	ia64_ssm(IA64_PSR_PP);
 667	ia64_srlz_i();
 668}
 669
 670static inline void
 671pfm_clear_psr_up(void)
 672{
 673	ia64_rsm(IA64_PSR_UP);
 674	ia64_srlz_i();
 675}
 676
 677static inline void
 678pfm_set_psr_up(void)
 679{
 680	ia64_ssm(IA64_PSR_UP);
 681	ia64_srlz_i();
 682}
 683
 684static inline unsigned long
 685pfm_get_psr(void)
 686{
 687	unsigned long tmp;
 688	tmp = ia64_getreg(_IA64_REG_PSR);
 689	ia64_srlz_i();
 690	return tmp;
 691}
 692
 693static inline void
 694pfm_set_psr_l(unsigned long val)
 695{
 696	ia64_setreg(_IA64_REG_PSR_L, val);
 697	ia64_srlz_i();
 698}
 699
 700static inline void
 701pfm_freeze_pmu(void)
 702{
 703	ia64_set_pmc(0,1UL);
 704	ia64_srlz_d();
 705}
 706
 707static inline void
 708pfm_unfreeze_pmu(void)
 709{
 710	ia64_set_pmc(0,0UL);
 711	ia64_srlz_d();
 712}
 713
 714static inline void
 715pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
 716{
 717	int i;
 718
 719	for (i=0; i < nibrs; i++) {
 720		ia64_set_ibr(i, ibrs[i]);
 721		ia64_dv_serialize_instruction();
 722	}
 723	ia64_srlz_i();
 724}
 725
 726static inline void
 727pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
 728{
 729	int i;
 730
 731	for (i=0; i < ndbrs; i++) {
 732		ia64_set_dbr(i, dbrs[i]);
 733		ia64_dv_serialize_data();
 734	}
 735	ia64_srlz_d();
 736}
 737
 738/*
 739 * PMD[i] must be a counter. no check is made
 740 */
 741static inline unsigned long
 742pfm_read_soft_counter(pfm_context_t *ctx, int i)
 743{
 744	return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
 745}
 746
 747/*
 748 * PMD[i] must be a counter. no check is made
 749 */
 750static inline void
 751pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
 752{
 753	unsigned long ovfl_val = pmu_conf->ovfl_val;
 754
 755	ctx->ctx_pmds[i].val = val  & ~ovfl_val;
 756	/*
 757	 * writing to unimplemented part is ignore, so we do not need to
 758	 * mask off top part
 759	 */
 760	ia64_set_pmd(i, val & ovfl_val);
 761}
 762
 763static pfm_msg_t *
 764pfm_get_new_msg(pfm_context_t *ctx)
 765{
 766	int idx, next;
 767
 768	next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
 769
 770	DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
 771	if (next == ctx->ctx_msgq_head) return NULL;
 772
 773 	idx = 	ctx->ctx_msgq_tail;
 774	ctx->ctx_msgq_tail = next;
 775
 776	DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
 777
 778	return ctx->ctx_msgq+idx;
 779}
 780
 781static pfm_msg_t *
 782pfm_get_next_msg(pfm_context_t *ctx)
 783{
 784	pfm_msg_t *msg;
 785
 786	DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
 787
 788	if (PFM_CTXQ_EMPTY(ctx)) return NULL;
 789
 790	/*
 791	 * get oldest message
 792	 */
 793	msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
 794
 795	/*
 796	 * and move forward
 797	 */
 798	ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
 799
 800	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));
 801
 802	return msg;
 803}
 804
 805static void
 806pfm_reset_msgq(pfm_context_t *ctx)
 807{
 808	ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
 809	DPRINT(("ctx=%p msgq reset\n", ctx));
 810}
 811
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 812static pfm_context_t *
 813pfm_context_alloc(int ctx_flags)
 814{
 815	pfm_context_t *ctx;
 816
 817	/* 
 818	 * allocate context descriptor 
 819	 * must be able to free with interrupts disabled
 820	 */
 821	ctx = kzalloc(sizeof(pfm_context_t), GFP_KERNEL);
 822	if (ctx) {
 823		DPRINT(("alloc ctx @%p\n", ctx));
 824
 825		/*
 826		 * init context protection lock
 827		 */
 828		spin_lock_init(&ctx->ctx_lock);
 829
 830		/*
 831		 * context is unloaded
 832		 */
 833		ctx->ctx_state = PFM_CTX_UNLOADED;
 834
 835		/*
 836		 * initialization of context's flags
 837		 */
 838		ctx->ctx_fl_block       = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
 839		ctx->ctx_fl_system      = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
 840		ctx->ctx_fl_no_msg      = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
 841		/*
 842		 * will move to set properties
 843		 * ctx->ctx_fl_excl_idle   = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
 844		 */
 845
 846		/*
 847		 * init restart semaphore to locked
 848		 */
 849		init_completion(&ctx->ctx_restart_done);
 850
 851		/*
 852		 * activation is used in SMP only
 853		 */
 854		ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
 855		SET_LAST_CPU(ctx, -1);
 856
 857		/*
 858		 * initialize notification message queue
 859		 */
 860		ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
 861		init_waitqueue_head(&ctx->ctx_msgq_wait);
 862		init_waitqueue_head(&ctx->ctx_zombieq);
 863
 864	}
 865	return ctx;
 866}
 867
 868static void
 869pfm_context_free(pfm_context_t *ctx)
 870{
 871	if (ctx) {
 872		DPRINT(("free ctx @%p\n", ctx));
 873		kfree(ctx);
 874	}
 875}
 876
 877static void
 878pfm_mask_monitoring(struct task_struct *task)
 879{
 880	pfm_context_t *ctx = PFM_GET_CTX(task);
 881	unsigned long mask, val, ovfl_mask;
 882	int i;
 883
 884	DPRINT_ovfl(("masking monitoring for [%d]\n", task_pid_nr(task)));
 885
 886	ovfl_mask = pmu_conf->ovfl_val;
 887	/*
 888	 * monitoring can only be masked as a result of a valid
 889	 * counter overflow. In UP, it means that the PMU still
 890	 * has an owner. Note that the owner can be different
 891	 * from the current task. However the PMU state belongs
 892	 * to the owner.
 893	 * In SMP, a valid overflow only happens when task is
 894	 * current. Therefore if we come here, we know that
 895	 * the PMU state belongs to the current task, therefore
 896	 * we can access the live registers.
 897	 *
 898	 * So in both cases, the live register contains the owner's
 899	 * state. We can ONLY touch the PMU registers and NOT the PSR.
 900	 *
 901	 * As a consequence to this call, the ctx->th_pmds[] array
 902	 * contains stale information which must be ignored
 903	 * when context is reloaded AND monitoring is active (see
 904	 * pfm_restart).
 905	 */
 906	mask = ctx->ctx_used_pmds[0];
 907	for (i = 0; mask; i++, mask>>=1) {
 908		/* skip non used pmds */
 909		if ((mask & 0x1) == 0) continue;
 910		val = ia64_get_pmd(i);
 911
 912		if (PMD_IS_COUNTING(i)) {
 913			/*
 914		 	 * we rebuild the full 64 bit value of the counter
 915		 	 */
 916			ctx->ctx_pmds[i].val += (val & ovfl_mask);
 917		} else {
 918			ctx->ctx_pmds[i].val = val;
 919		}
 920		DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
 921			i,
 922			ctx->ctx_pmds[i].val,
 923			val & ovfl_mask));
 924	}
 925	/*
 926	 * mask monitoring by setting the privilege level to 0
 927	 * we cannot use psr.pp/psr.up for this, it is controlled by
 928	 * the user
 929	 *
 930	 * if task is current, modify actual registers, otherwise modify
 931	 * thread save state, i.e., what will be restored in pfm_load_regs()
 932	 */
 933	mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
 934	for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
 935		if ((mask & 0x1) == 0UL) continue;
 936		ia64_set_pmc(i, ctx->th_pmcs[i] & ~0xfUL);
 937		ctx->th_pmcs[i] &= ~0xfUL;
 938		DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
 939	}
 940	/*
 941	 * make all of this visible
 942	 */
 943	ia64_srlz_d();
 944}
 945
 946/*
 947 * must always be done with task == current
 948 *
 949 * context must be in MASKED state when calling
 950 */
 951static void
 952pfm_restore_monitoring(struct task_struct *task)
 953{
 954	pfm_context_t *ctx = PFM_GET_CTX(task);
 955	unsigned long mask, ovfl_mask;
 956	unsigned long psr, val;
 957	int i, is_system;
 958
 959	is_system = ctx->ctx_fl_system;
 960	ovfl_mask = pmu_conf->ovfl_val;
 961
 962	if (task != current) {
 963		printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task_pid_nr(task), task_pid_nr(current));
 964		return;
 965	}
 966	if (ctx->ctx_state != PFM_CTX_MASKED) {
 967		printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
 968			task_pid_nr(task), task_pid_nr(current), ctx->ctx_state);
 969		return;
 970	}
 971	psr = pfm_get_psr();
 972	/*
 973	 * monitoring is masked via the PMC.
 974	 * As we restore their value, we do not want each counter to
 975	 * restart right away. We stop monitoring using the PSR,
 976	 * restore the PMC (and PMD) and then re-establish the psr
 977	 * as it was. Note that there can be no pending overflow at
 978	 * this point, because monitoring was MASKED.
 979	 *
 980	 * system-wide session are pinned and self-monitoring
 981	 */
 982	if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
 983		/* disable dcr pp */
 984		ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
 985		pfm_clear_psr_pp();
 986	} else {
 987		pfm_clear_psr_up();
 988	}
 989	/*
 990	 * first, we restore the PMD
 991	 */
 992	mask = ctx->ctx_used_pmds[0];
 993	for (i = 0; mask; i++, mask>>=1) {
 994		/* skip non used pmds */
 995		if ((mask & 0x1) == 0) continue;
 996
 997		if (PMD_IS_COUNTING(i)) {
 998			/*
 999			 * we split the 64bit value according to
1000			 * counter width
1001			 */
1002			val = ctx->ctx_pmds[i].val & ovfl_mask;
1003			ctx->ctx_pmds[i].val &= ~ovfl_mask;
1004		} else {
1005			val = ctx->ctx_pmds[i].val;
1006		}
1007		ia64_set_pmd(i, val);
1008
1009		DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1010			i,
1011			ctx->ctx_pmds[i].val,
1012			val));
1013	}
1014	/*
1015	 * restore the PMCs
1016	 */
1017	mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1018	for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1019		if ((mask & 0x1) == 0UL) continue;
1020		ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1021		ia64_set_pmc(i, ctx->th_pmcs[i]);
1022		DPRINT(("[%d] pmc[%d]=0x%lx\n",
1023					task_pid_nr(task), i, ctx->th_pmcs[i]));
1024	}
1025	ia64_srlz_d();
1026
1027	/*
1028	 * must restore DBR/IBR because could be modified while masked
1029	 * XXX: need to optimize 
1030	 */
1031	if (ctx->ctx_fl_using_dbreg) {
1032		pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1033		pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1034	}
1035
1036	/*
1037	 * now restore PSR
1038	 */
1039	if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1040		/* enable dcr pp */
1041		ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1042		ia64_srlz_i();
1043	}
1044	pfm_set_psr_l(psr);
1045}
1046
1047static inline void
1048pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1049{
1050	int i;
1051
1052	ia64_srlz_d();
1053
1054	for (i=0; mask; i++, mask>>=1) {
1055		if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1056	}
1057}
1058
1059/*
1060 * reload from thread state (used for ctxw only)
1061 */
1062static inline void
1063pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1064{
1065	int i;
1066	unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1067
1068	for (i=0; mask; i++, mask>>=1) {
1069		if ((mask & 0x1) == 0) continue;
1070		val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1071		ia64_set_pmd(i, val);
1072	}
1073	ia64_srlz_d();
1074}
1075
1076/*
1077 * propagate PMD from context to thread-state
1078 */
1079static inline void
1080pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1081{
1082	unsigned long ovfl_val = pmu_conf->ovfl_val;
1083	unsigned long mask = ctx->ctx_all_pmds[0];
1084	unsigned long val;
1085	int i;
1086
1087	DPRINT(("mask=0x%lx\n", mask));
1088
1089	for (i=0; mask; i++, mask>>=1) {
1090
1091		val = ctx->ctx_pmds[i].val;
1092
1093		/*
1094		 * We break up the 64 bit value into 2 pieces
1095		 * the lower bits go to the machine state in the
1096		 * thread (will be reloaded on ctxsw in).
1097		 * The upper part stays in the soft-counter.
1098		 */
1099		if (PMD_IS_COUNTING(i)) {
1100			ctx->ctx_pmds[i].val = val & ~ovfl_val;
1101			 val &= ovfl_val;
1102		}
1103		ctx->th_pmds[i] = val;
1104
1105		DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1106			i,
1107			ctx->th_pmds[i],
1108			ctx->ctx_pmds[i].val));
1109	}
1110}
1111
1112/*
1113 * propagate PMC from context to thread-state
1114 */
1115static inline void
1116pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1117{
1118	unsigned long mask = ctx->ctx_all_pmcs[0];
1119	int i;
1120
1121	DPRINT(("mask=0x%lx\n", mask));
1122
1123	for (i=0; mask; i++, mask>>=1) {
1124		/* masking 0 with ovfl_val yields 0 */
1125		ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1126		DPRINT(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
1127	}
1128}
1129
1130
1131
1132static inline void
1133pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1134{
1135	int i;
1136
1137	for (i=0; mask; i++, mask>>=1) {
1138		if ((mask & 0x1) == 0) continue;
1139		ia64_set_pmc(i, pmcs[i]);
1140	}
1141	ia64_srlz_d();
1142}
1143
1144static inline int
1145pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1146{
1147	return memcmp(a, b, sizeof(pfm_uuid_t));
1148}
1149
1150static inline int
1151pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1152{
1153	int ret = 0;
1154	if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1155	return ret;
1156}
1157
1158static inline int
1159pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1160{
1161	int ret = 0;
1162	if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1163	return ret;
1164}
1165
1166
1167static inline int
1168pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1169		     int cpu, void *arg)
1170{
1171	int ret = 0;
1172	if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1173	return ret;
1174}
1175
1176static inline int
1177pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1178		     int cpu, void *arg)
1179{
1180	int ret = 0;
1181	if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1182	return ret;
1183}
1184
1185static inline int
1186pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1187{
1188	int ret = 0;
1189	if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1190	return ret;
1191}
1192
1193static inline int
1194pfm_buf_fmt_restart_active(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1195{
1196	int ret = 0;
1197	if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1198	return ret;
1199}
1200
1201static pfm_buffer_fmt_t *
1202__pfm_find_buffer_fmt(pfm_uuid_t uuid)
1203{
1204	struct list_head * pos;
1205	pfm_buffer_fmt_t * entry;
1206
1207	list_for_each(pos, &pfm_buffer_fmt_list) {
1208		entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1209		if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1210			return entry;
1211	}
1212	return NULL;
1213}
1214 
1215/*
1216 * find a buffer format based on its uuid
1217 */
1218static pfm_buffer_fmt_t *
1219pfm_find_buffer_fmt(pfm_uuid_t uuid)
1220{
1221	pfm_buffer_fmt_t * fmt;
1222	spin_lock(&pfm_buffer_fmt_lock);
1223	fmt = __pfm_find_buffer_fmt(uuid);
1224	spin_unlock(&pfm_buffer_fmt_lock);
1225	return fmt;
1226}
1227 
1228int
1229pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1230{
1231	int ret = 0;
1232
1233	/* some sanity checks */
1234	if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1235
1236	/* we need at least a handler */
1237	if (fmt->fmt_handler == NULL) return -EINVAL;
1238
1239	/*
1240	 * XXX: need check validity of fmt_arg_size
1241	 */
1242
1243	spin_lock(&pfm_buffer_fmt_lock);
1244
1245	if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1246		printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1247		ret = -EBUSY;
1248		goto out;
1249	} 
1250	list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1251	printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1252
1253out:
1254	spin_unlock(&pfm_buffer_fmt_lock);
1255 	return ret;
1256}
1257EXPORT_SYMBOL(pfm_register_buffer_fmt);
1258
1259int
1260pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1261{
1262	pfm_buffer_fmt_t *fmt;
1263	int ret = 0;
1264
1265	spin_lock(&pfm_buffer_fmt_lock);
1266
1267	fmt = __pfm_find_buffer_fmt(uuid);
1268	if (!fmt) {
1269		printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1270		ret = -EINVAL;
1271		goto out;
1272	}
1273	list_del_init(&fmt->fmt_list);
1274	printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1275
1276out:
1277	spin_unlock(&pfm_buffer_fmt_lock);
1278	return ret;
1279
1280}
1281EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1282
 
 
1283static int
1284pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1285{
1286	unsigned long flags;
1287	/*
1288	 * validity checks on cpu_mask have been done upstream
1289	 */
1290	LOCK_PFS(flags);
1291
1292	DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1293		pfm_sessions.pfs_sys_sessions,
1294		pfm_sessions.pfs_task_sessions,
1295		pfm_sessions.pfs_sys_use_dbregs,
1296		is_syswide,
1297		cpu));
1298
1299	if (is_syswide) {
1300		/*
1301		 * cannot mix system wide and per-task sessions
1302		 */
1303		if (pfm_sessions.pfs_task_sessions > 0UL) {
1304			DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1305			  	pfm_sessions.pfs_task_sessions));
1306			goto abort;
1307		}
1308
1309		if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1310
1311		DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1312
1313		pfm_sessions.pfs_sys_session[cpu] = task;
1314
1315		pfm_sessions.pfs_sys_sessions++ ;
1316
1317	} else {
1318		if (pfm_sessions.pfs_sys_sessions) goto abort;
1319		pfm_sessions.pfs_task_sessions++;
1320	}
1321
1322	DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1323		pfm_sessions.pfs_sys_sessions,
1324		pfm_sessions.pfs_task_sessions,
1325		pfm_sessions.pfs_sys_use_dbregs,
1326		is_syswide,
1327		cpu));
1328
1329	/*
1330	 * Force idle() into poll mode
1331	 */
1332	cpu_idle_poll_ctrl(true);
1333
1334	UNLOCK_PFS(flags);
1335
1336	return 0;
1337
1338error_conflict:
1339	DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1340  		task_pid_nr(pfm_sessions.pfs_sys_session[cpu]),
1341		cpu));
1342abort:
1343	UNLOCK_PFS(flags);
1344
1345	return -EBUSY;
1346
1347}
1348
1349static int
1350pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1351{
1352	unsigned long flags;
1353	/*
1354	 * validity checks on cpu_mask have been done upstream
1355	 */
1356	LOCK_PFS(flags);
1357
1358	DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1359		pfm_sessions.pfs_sys_sessions,
1360		pfm_sessions.pfs_task_sessions,
1361		pfm_sessions.pfs_sys_use_dbregs,
1362		is_syswide,
1363		cpu));
1364
1365
1366	if (is_syswide) {
1367		pfm_sessions.pfs_sys_session[cpu] = NULL;
1368		/*
1369		 * would not work with perfmon+more than one bit in cpu_mask
1370		 */
1371		if (ctx && ctx->ctx_fl_using_dbreg) {
1372			if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1373				printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1374			} else {
1375				pfm_sessions.pfs_sys_use_dbregs--;
1376			}
1377		}
1378		pfm_sessions.pfs_sys_sessions--;
1379	} else {
1380		pfm_sessions.pfs_task_sessions--;
1381	}
1382	DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1383		pfm_sessions.pfs_sys_sessions,
1384		pfm_sessions.pfs_task_sessions,
1385		pfm_sessions.pfs_sys_use_dbregs,
1386		is_syswide,
1387		cpu));
1388
1389	/* Undo forced polling. Last session reenables pal_halt */
1390	cpu_idle_poll_ctrl(false);
 
 
 
1391
1392	UNLOCK_PFS(flags);
1393
1394	return 0;
1395}
1396
1397/*
1398 * removes virtual mapping of the sampling buffer.
1399 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1400 * a PROTECT_CTX() section.
1401 */
1402static int
1403pfm_remove_smpl_mapping(void *vaddr, unsigned long size)
1404{
1405	struct task_struct *task = current;
1406	int r;
1407
1408	/* sanity checks */
1409	if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1410		printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task_pid_nr(task), task->mm);
1411		return -EINVAL;
1412	}
1413
1414	DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1415
1416	/*
1417	 * does the actual unmapping
1418	 */
1419	r = vm_munmap((unsigned long)vaddr, size);
1420
1421	if (r !=0) {
1422		printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task_pid_nr(task), vaddr, size);
1423	}
1424
1425	DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1426
1427	return 0;
1428}
1429
1430/*
1431 * free actual physical storage used by sampling buffer
1432 */
1433#if 0
1434static int
1435pfm_free_smpl_buffer(pfm_context_t *ctx)
1436{
1437	pfm_buffer_fmt_t *fmt;
1438
1439	if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1440
1441	/*
1442	 * we won't use the buffer format anymore
1443	 */
1444	fmt = ctx->ctx_buf_fmt;
1445
1446	DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1447		ctx->ctx_smpl_hdr,
1448		ctx->ctx_smpl_size,
1449		ctx->ctx_smpl_vaddr));
1450
1451	pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1452
1453	/*
1454	 * free the buffer
1455	 */
1456	vfree(ctx->ctx_smpl_hdr);
1457
1458	ctx->ctx_smpl_hdr  = NULL;
1459	ctx->ctx_smpl_size = 0UL;
1460
1461	return 0;
1462
1463invalid_free:
1464	printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", task_pid_nr(current));
1465	return -EINVAL;
1466}
1467#endif
1468
1469static inline void
1470pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1471{
1472	if (fmt == NULL) return;
1473
1474	pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1475
1476}
1477
1478/*
1479 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1480 * no real gain from having the whole whorehouse mounted. So we don't need
1481 * any operations on the root directory. However, we need a non-trivial
1482 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1483 */
1484static struct vfsmount *pfmfs_mnt __read_mostly;
1485
1486static int __init
1487init_pfm_fs(void)
1488{
1489	int err = register_filesystem(&pfm_fs_type);
1490	if (!err) {
1491		pfmfs_mnt = kern_mount(&pfm_fs_type);
1492		err = PTR_ERR(pfmfs_mnt);
1493		if (IS_ERR(pfmfs_mnt))
1494			unregister_filesystem(&pfm_fs_type);
1495		else
1496			err = 0;
1497	}
1498	return err;
1499}
1500
1501static ssize_t
1502pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
1503{
1504	pfm_context_t *ctx;
1505	pfm_msg_t *msg;
1506	ssize_t ret;
1507	unsigned long flags;
1508  	DECLARE_WAITQUEUE(wait, current);
1509	if (PFM_IS_FILE(filp) == 0) {
1510		printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1511		return -EINVAL;
1512	}
1513
1514	ctx = filp->private_data;
1515	if (ctx == NULL) {
1516		printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", task_pid_nr(current));
1517		return -EINVAL;
1518	}
1519
1520	/*
1521	 * check even when there is no message
1522	 */
1523	if (size < sizeof(pfm_msg_t)) {
1524		DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1525		return -EINVAL;
1526	}
1527
1528	PROTECT_CTX(ctx, flags);
1529
1530  	/*
1531	 * put ourselves on the wait queue
1532	 */
1533  	add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1534
1535
1536  	for(;;) {
1537		/*
1538		 * check wait queue
1539		 */
1540
1541  		set_current_state(TASK_INTERRUPTIBLE);
1542
1543		DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1544
1545		ret = 0;
1546		if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1547
1548		UNPROTECT_CTX(ctx, flags);
1549
1550		/*
1551		 * check non-blocking read
1552		 */
1553      		ret = -EAGAIN;
1554		if(filp->f_flags & O_NONBLOCK) break;
1555
1556		/*
1557		 * check pending signals
1558		 */
1559		if(signal_pending(current)) {
1560			ret = -EINTR;
1561			break;
1562		}
1563      		/*
1564		 * no message, so wait
1565		 */
1566      		schedule();
1567
1568		PROTECT_CTX(ctx, flags);
1569	}
1570	DPRINT(("[%d] back to running ret=%ld\n", task_pid_nr(current), ret));
1571  	set_current_state(TASK_RUNNING);
1572	remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1573
1574	if (ret < 0) goto abort;
1575
1576	ret = -EINVAL;
1577	msg = pfm_get_next_msg(ctx);
1578	if (msg == NULL) {
1579		printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, task_pid_nr(current));
1580		goto abort_locked;
1581	}
1582
1583	DPRINT(("fd=%d type=%d\n", msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1584
1585	ret = -EFAULT;
1586  	if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1587
1588abort_locked:
1589	UNPROTECT_CTX(ctx, flags);
1590abort:
1591	return ret;
1592}
1593
1594static ssize_t
1595pfm_write(struct file *file, const char __user *ubuf,
1596			  size_t size, loff_t *ppos)
1597{
1598	DPRINT(("pfm_write called\n"));
1599	return -EINVAL;
1600}
1601
1602static __poll_t
1603pfm_poll(struct file *filp, poll_table * wait)
1604{
1605	pfm_context_t *ctx;
1606	unsigned long flags;
1607	__poll_t mask = 0;
1608
1609	if (PFM_IS_FILE(filp) == 0) {
1610		printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1611		return 0;
1612	}
1613
1614	ctx = filp->private_data;
1615	if (ctx == NULL) {
1616		printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", task_pid_nr(current));
1617		return 0;
1618	}
1619
1620
1621	DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1622
1623	poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1624
1625	PROTECT_CTX(ctx, flags);
1626
1627	if (PFM_CTXQ_EMPTY(ctx) == 0)
1628		mask =  EPOLLIN | EPOLLRDNORM;
1629
1630	UNPROTECT_CTX(ctx, flags);
1631
1632	DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1633
1634	return mask;
1635}
1636
1637static long
1638pfm_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1639{
1640	DPRINT(("pfm_ioctl called\n"));
1641	return -EINVAL;
1642}
1643
1644/*
1645 * interrupt cannot be masked when coming here
1646 */
1647static inline int
1648pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1649{
1650	int ret;
1651
1652	ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1653
1654	DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1655		task_pid_nr(current),
1656		fd,
1657		on,
1658		ctx->ctx_async_queue, ret));
1659
1660	return ret;
1661}
1662
1663static int
1664pfm_fasync(int fd, struct file *filp, int on)
1665{
1666	pfm_context_t *ctx;
1667	int ret;
1668
1669	if (PFM_IS_FILE(filp) == 0) {
1670		printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", task_pid_nr(current));
1671		return -EBADF;
1672	}
1673
1674	ctx = filp->private_data;
1675	if (ctx == NULL) {
1676		printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", task_pid_nr(current));
1677		return -EBADF;
1678	}
1679	/*
1680	 * we cannot mask interrupts during this call because this may
1681	 * may go to sleep if memory is not readily avalaible.
1682	 *
1683	 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1684	 * done in caller. Serialization of this function is ensured by caller.
1685	 */
1686	ret = pfm_do_fasync(fd, filp, ctx, on);
1687
1688
1689	DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1690		fd,
1691		on,
1692		ctx->ctx_async_queue, ret));
1693
1694	return ret;
1695}
1696
1697#ifdef CONFIG_SMP
1698/*
1699 * this function is exclusively called from pfm_close().
1700 * The context is not protected at that time, nor are interrupts
1701 * on the remote CPU. That's necessary to avoid deadlocks.
1702 */
1703static void
1704pfm_syswide_force_stop(void *info)
1705{
1706	pfm_context_t   *ctx = (pfm_context_t *)info;
1707	struct pt_regs *regs = task_pt_regs(current);
1708	struct task_struct *owner;
1709	unsigned long flags;
1710	int ret;
1711
1712	if (ctx->ctx_cpu != smp_processor_id()) {
1713		printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d  but on CPU%d\n",
1714			ctx->ctx_cpu,
1715			smp_processor_id());
1716		return;
1717	}
1718	owner = GET_PMU_OWNER();
1719	if (owner != ctx->ctx_task) {
1720		printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1721			smp_processor_id(),
1722			task_pid_nr(owner), task_pid_nr(ctx->ctx_task));
1723		return;
1724	}
1725	if (GET_PMU_CTX() != ctx) {
1726		printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1727			smp_processor_id(),
1728			GET_PMU_CTX(), ctx);
1729		return;
1730	}
1731
1732	DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), task_pid_nr(ctx->ctx_task)));
1733	/*
1734	 * the context is already protected in pfm_close(), we simply
1735	 * need to mask interrupts to avoid a PMU interrupt race on
1736	 * this CPU
1737	 */
1738	local_irq_save(flags);
1739
1740	ret = pfm_context_unload(ctx, NULL, 0, regs);
1741	if (ret) {
1742		DPRINT(("context_unload returned %d\n", ret));
1743	}
1744
1745	/*
1746	 * unmask interrupts, PMU interrupts are now spurious here
1747	 */
1748	local_irq_restore(flags);
1749}
1750
1751static void
1752pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1753{
1754	int ret;
1755
1756	DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1757	ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 1);
1758	DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1759}
1760#endif /* CONFIG_SMP */
1761
1762/*
1763 * called for each close(). Partially free resources.
1764 * When caller is self-monitoring, the context is unloaded.
1765 */
1766static int
1767pfm_flush(struct file *filp, fl_owner_t id)
1768{
1769	pfm_context_t *ctx;
1770	struct task_struct *task;
1771	struct pt_regs *regs;
1772	unsigned long flags;
1773	unsigned long smpl_buf_size = 0UL;
1774	void *smpl_buf_vaddr = NULL;
1775	int state, is_system;
1776
1777	if (PFM_IS_FILE(filp) == 0) {
1778		DPRINT(("bad magic for\n"));
1779		return -EBADF;
1780	}
1781
1782	ctx = filp->private_data;
1783	if (ctx == NULL) {
1784		printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", task_pid_nr(current));
1785		return -EBADF;
1786	}
1787
1788	/*
1789	 * remove our file from the async queue, if we use this mode.
1790	 * This can be done without the context being protected. We come
1791	 * here when the context has become unreachable by other tasks.
1792	 *
1793	 * We may still have active monitoring at this point and we may
1794	 * end up in pfm_overflow_handler(). However, fasync_helper()
1795	 * operates with interrupts disabled and it cleans up the
1796	 * queue. If the PMU handler is called prior to entering
1797	 * fasync_helper() then it will send a signal. If it is
1798	 * invoked after, it will find an empty queue and no
1799	 * signal will be sent. In both case, we are safe
1800	 */
1801	PROTECT_CTX(ctx, flags);
1802
1803	state     = ctx->ctx_state;
1804	is_system = ctx->ctx_fl_system;
1805
1806	task = PFM_CTX_TASK(ctx);
1807	regs = task_pt_regs(task);
1808
1809	DPRINT(("ctx_state=%d is_current=%d\n",
1810		state,
1811		task == current ? 1 : 0));
1812
1813	/*
1814	 * if state == UNLOADED, then task is NULL
1815	 */
1816
1817	/*
1818	 * we must stop and unload because we are losing access to the context.
1819	 */
1820	if (task == current) {
1821#ifdef CONFIG_SMP
1822		/*
1823		 * the task IS the owner but it migrated to another CPU: that's bad
1824		 * but we must handle this cleanly. Unfortunately, the kernel does
1825		 * not provide a mechanism to block migration (while the context is loaded).
1826		 *
1827		 * We need to release the resource on the ORIGINAL cpu.
1828		 */
1829		if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1830
1831			DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1832			/*
1833			 * keep context protected but unmask interrupt for IPI
1834			 */
1835			local_irq_restore(flags);
1836
1837			pfm_syswide_cleanup_other_cpu(ctx);
1838
1839			/*
1840			 * restore interrupt masking
1841			 */
1842			local_irq_save(flags);
1843
1844			/*
1845			 * context is unloaded at this point
1846			 */
1847		} else
1848#endif /* CONFIG_SMP */
1849		{
1850
1851			DPRINT(("forcing unload\n"));
1852			/*
1853		 	* stop and unload, returning with state UNLOADED
1854		 	* and session unreserved.
1855		 	*/
1856			pfm_context_unload(ctx, NULL, 0, regs);
1857
1858			DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1859		}
1860	}
1861
1862	/*
1863	 * remove virtual mapping, if any, for the calling task.
1864	 * cannot reset ctx field until last user is calling close().
1865	 *
1866	 * ctx_smpl_vaddr must never be cleared because it is needed
1867	 * by every task with access to the context
1868	 *
1869	 * When called from do_exit(), the mm context is gone already, therefore
1870	 * mm is NULL, i.e., the VMA is already gone  and we do not have to
1871	 * do anything here
1872	 */
1873	if (ctx->ctx_smpl_vaddr && current->mm) {
1874		smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1875		smpl_buf_size  = ctx->ctx_smpl_size;
1876	}
1877
1878	UNPROTECT_CTX(ctx, flags);
1879
1880	/*
1881	 * if there was a mapping, then we systematically remove it
1882	 * at this point. Cannot be done inside critical section
1883	 * because some VM function reenables interrupts.
1884	 *
1885	 */
1886	if (smpl_buf_vaddr) pfm_remove_smpl_mapping(smpl_buf_vaddr, smpl_buf_size);
1887
1888	return 0;
1889}
1890/*
1891 * called either on explicit close() or from exit_files(). 
1892 * Only the LAST user of the file gets to this point, i.e., it is
1893 * called only ONCE.
1894 *
1895 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero 
1896 * (fput()),i.e, last task to access the file. Nobody else can access the 
1897 * file at this point.
1898 *
1899 * When called from exit_files(), the VMA has been freed because exit_mm()
1900 * is executed before exit_files().
1901 *
1902 * When called from exit_files(), the current task is not yet ZOMBIE but we
1903 * flush the PMU state to the context. 
1904 */
1905static int
1906pfm_close(struct inode *inode, struct file *filp)
1907{
1908	pfm_context_t *ctx;
1909	struct task_struct *task;
1910	struct pt_regs *regs;
1911  	DECLARE_WAITQUEUE(wait, current);
1912	unsigned long flags;
1913	unsigned long smpl_buf_size = 0UL;
1914	void *smpl_buf_addr = NULL;
1915	int free_possible = 1;
1916	int state, is_system;
1917
1918	DPRINT(("pfm_close called private=%p\n", filp->private_data));
1919
1920	if (PFM_IS_FILE(filp) == 0) {
1921		DPRINT(("bad magic\n"));
1922		return -EBADF;
1923	}
1924	
1925	ctx = filp->private_data;
1926	if (ctx == NULL) {
1927		printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", task_pid_nr(current));
1928		return -EBADF;
1929	}
1930
1931	PROTECT_CTX(ctx, flags);
1932
1933	state     = ctx->ctx_state;
1934	is_system = ctx->ctx_fl_system;
1935
1936	task = PFM_CTX_TASK(ctx);
1937	regs = task_pt_regs(task);
1938
1939	DPRINT(("ctx_state=%d is_current=%d\n", 
1940		state,
1941		task == current ? 1 : 0));
1942
1943	/*
1944	 * if task == current, then pfm_flush() unloaded the context
1945	 */
1946	if (state == PFM_CTX_UNLOADED) goto doit;
1947
1948	/*
1949	 * context is loaded/masked and task != current, we need to
1950	 * either force an unload or go zombie
1951	 */
1952
1953	/*
1954	 * The task is currently blocked or will block after an overflow.
1955	 * we must force it to wakeup to get out of the
1956	 * MASKED state and transition to the unloaded state by itself.
1957	 *
1958	 * This situation is only possible for per-task mode
1959	 */
1960	if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
1961
1962		/*
1963		 * set a "partial" zombie state to be checked
1964		 * upon return from down() in pfm_handle_work().
1965		 *
1966		 * We cannot use the ZOMBIE state, because it is checked
1967		 * by pfm_load_regs() which is called upon wakeup from down().
1968		 * In such case, it would free the context and then we would
1969		 * return to pfm_handle_work() which would access the
1970		 * stale context. Instead, we set a flag invisible to pfm_load_regs()
1971		 * but visible to pfm_handle_work().
1972		 *
1973		 * For some window of time, we have a zombie context with
1974		 * ctx_state = MASKED  and not ZOMBIE
1975		 */
1976		ctx->ctx_fl_going_zombie = 1;
1977
1978		/*
1979		 * force task to wake up from MASKED state
1980		 */
1981		complete(&ctx->ctx_restart_done);
1982
1983		DPRINT(("waking up ctx_state=%d\n", state));
1984
1985		/*
1986		 * put ourself to sleep waiting for the other
1987		 * task to report completion
1988		 *
1989		 * the context is protected by mutex, therefore there
1990		 * is no risk of being notified of completion before
1991		 * begin actually on the waitq.
1992		 */
1993  		set_current_state(TASK_INTERRUPTIBLE);
1994  		add_wait_queue(&ctx->ctx_zombieq, &wait);
1995
1996		UNPROTECT_CTX(ctx, flags);
1997
1998		/*
1999		 * XXX: check for signals :
2000		 * 	- ok for explicit close
2001		 * 	- not ok when coming from exit_files()
2002		 */
2003      		schedule();
2004
2005
2006		PROTECT_CTX(ctx, flags);
2007
2008
2009		remove_wait_queue(&ctx->ctx_zombieq, &wait);
2010  		set_current_state(TASK_RUNNING);
2011
2012		/*
2013		 * context is unloaded at this point
2014		 */
2015		DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2016	}
2017	else if (task != current) {
2018#ifdef CONFIG_SMP
2019		/*
2020	 	 * switch context to zombie state
2021	 	 */
2022		ctx->ctx_state = PFM_CTX_ZOMBIE;
2023
2024		DPRINT(("zombie ctx for [%d]\n", task_pid_nr(task)));
2025		/*
2026		 * cannot free the context on the spot. deferred until
2027		 * the task notices the ZOMBIE state
2028		 */
2029		free_possible = 0;
2030#else
2031		pfm_context_unload(ctx, NULL, 0, regs);
2032#endif
2033	}
2034
2035doit:
2036	/* reload state, may have changed during  opening of critical section */
2037	state = ctx->ctx_state;
2038
2039	/*
2040	 * the context is still attached to a task (possibly current)
2041	 * we cannot destroy it right now
2042	 */
2043
2044	/*
2045	 * we must free the sampling buffer right here because
2046	 * we cannot rely on it being cleaned up later by the
2047	 * monitored task. It is not possible to free vmalloc'ed
2048	 * memory in pfm_load_regs(). Instead, we remove the buffer
2049	 * now. should there be subsequent PMU overflow originally
2050	 * meant for sampling, the will be converted to spurious
2051	 * and that's fine because the monitoring tools is gone anyway.
2052	 */
2053	if (ctx->ctx_smpl_hdr) {
2054		smpl_buf_addr = ctx->ctx_smpl_hdr;
2055		smpl_buf_size = ctx->ctx_smpl_size;
2056		/* no more sampling */
2057		ctx->ctx_smpl_hdr = NULL;
2058		ctx->ctx_fl_is_sampling = 0;
2059	}
2060
2061	DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2062		state,
2063		free_possible,
2064		smpl_buf_addr,
2065		smpl_buf_size));
2066
2067	if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2068
2069	/*
2070	 * UNLOADED that the session has already been unreserved.
2071	 */
2072	if (state == PFM_CTX_ZOMBIE) {
2073		pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2074	}
2075
2076	/*
2077	 * disconnect file descriptor from context must be done
2078	 * before we unlock.
2079	 */
2080	filp->private_data = NULL;
2081
2082	/*
2083	 * if we free on the spot, the context is now completely unreachable
2084	 * from the callers side. The monitored task side is also cut, so we
2085	 * can freely cut.
2086	 *
2087	 * If we have a deferred free, only the caller side is disconnected.
2088	 */
2089	UNPROTECT_CTX(ctx, flags);
2090
2091	/*
2092	 * All memory free operations (especially for vmalloc'ed memory)
2093	 * MUST be done with interrupts ENABLED.
2094	 */
2095	vfree(smpl_buf_addr);
2096
2097	/*
2098	 * return the memory used by the context
2099	 */
2100	if (free_possible) pfm_context_free(ctx);
2101
2102	return 0;
2103}
2104
 
 
 
 
 
 
 
 
 
2105static const struct file_operations pfm_file_ops = {
2106	.llseek		= no_llseek,
2107	.read		= pfm_read,
2108	.write		= pfm_write,
2109	.poll		= pfm_poll,
2110	.unlocked_ioctl = pfm_ioctl,
 
2111	.fasync		= pfm_fasync,
2112	.release	= pfm_close,
2113	.flush		= pfm_flush
2114};
2115
 
 
 
 
 
 
2116static char *pfmfs_dname(struct dentry *dentry, char *buffer, int buflen)
2117{
2118	return dynamic_dname(dentry, buffer, buflen, "pfm:[%lu]",
2119			     d_inode(dentry)->i_ino);
2120}
2121
2122static const struct dentry_operations pfmfs_dentry_operations = {
2123	.d_delete = always_delete_dentry,
2124	.d_dname = pfmfs_dname,
2125};
2126
2127
2128static struct file *
2129pfm_alloc_file(pfm_context_t *ctx)
2130{
2131	struct file *file;
2132	struct inode *inode;
2133	struct path path;
2134	struct qstr this = { .name = "" };
2135
2136	/*
2137	 * allocate a new inode
2138	 */
2139	inode = new_inode(pfmfs_mnt->mnt_sb);
2140	if (!inode)
2141		return ERR_PTR(-ENOMEM);
2142
2143	DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2144
2145	inode->i_mode = S_IFCHR|S_IRUGO;
2146	inode->i_uid  = current_fsuid();
2147	inode->i_gid  = current_fsgid();
2148
2149	/*
2150	 * allocate a new dcache entry
2151	 */
2152	path.dentry = d_alloc(pfmfs_mnt->mnt_root, &this);
2153	if (!path.dentry) {
2154		iput(inode);
2155		return ERR_PTR(-ENOMEM);
2156	}
2157	path.mnt = mntget(pfmfs_mnt);
2158
2159	d_add(path.dentry, inode);
2160
2161	file = alloc_file(&path, FMODE_READ, &pfm_file_ops);
2162	if (IS_ERR(file)) {
2163		path_put(&path);
2164		return file;
2165	}
2166
2167	file->f_flags = O_RDONLY;
2168	file->private_data = ctx;
2169
2170	return file;
2171}
2172
2173static int
2174pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2175{
2176	DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2177
2178	while (size > 0) {
2179		unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2180
2181
2182		if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2183			return -ENOMEM;
2184
2185		addr  += PAGE_SIZE;
2186		buf   += PAGE_SIZE;
2187		size  -= PAGE_SIZE;
2188	}
2189	return 0;
2190}
2191
2192/*
2193 * allocate a sampling buffer and remaps it into the user address space of the task
2194 */
2195static int
2196pfm_smpl_buffer_alloc(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2197{
2198	struct mm_struct *mm = task->mm;
2199	struct vm_area_struct *vma = NULL;
2200	unsigned long size;
2201	void *smpl_buf;
2202
2203
2204	/*
2205	 * the fixed header + requested size and align to page boundary
2206	 */
2207	size = PAGE_ALIGN(rsize);
2208
2209	DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2210
2211	/*
2212	 * check requested size to avoid Denial-of-service attacks
2213	 * XXX: may have to refine this test
2214	 * Check against address space limit.
2215	 *
2216	 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2217	 * 	return -ENOMEM;
2218	 */
2219	if (size > task_rlimit(task, RLIMIT_MEMLOCK))
2220		return -ENOMEM;
2221
2222	/*
2223	 * We do the easy to undo allocations first.
 
 
2224	 */
2225	smpl_buf = vzalloc(size);
2226	if (smpl_buf == NULL) {
2227		DPRINT(("Can't allocate sampling buffer\n"));
2228		return -ENOMEM;
2229	}
2230
2231	DPRINT(("smpl_buf @%p\n", smpl_buf));
2232
2233	/* allocate vma */
2234	vma = vm_area_alloc(mm);
2235	if (!vma) {
2236		DPRINT(("Cannot allocate vma\n"));
2237		goto error_kmem;
2238	}
 
2239
2240	/*
2241	 * partially initialize the vma for the sampling buffer
2242	 */
2243	vma->vm_file	     = get_file(filp);
2244	vma->vm_flags	     = VM_READ|VM_MAYREAD|VM_DONTEXPAND|VM_DONTDUMP;
 
2245	vma->vm_page_prot    = PAGE_READONLY; /* XXX may need to change */
2246
2247	/*
2248	 * Now we have everything we need and we can initialize
2249	 * and connect all the data structures
2250	 */
2251
2252	ctx->ctx_smpl_hdr   = smpl_buf;
2253	ctx->ctx_smpl_size  = size; /* aligned size */
2254
2255	/*
2256	 * Let's do the difficult operations next.
2257	 *
2258	 * now we atomically find some area in the address space and
2259	 * remap the buffer in it.
2260	 */
2261	down_write(&task->mm->mmap_sem);
2262
2263	/* find some free area in address space, must have mmap sem held */
2264	vma->vm_start = get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS);
2265	if (IS_ERR_VALUE(vma->vm_start)) {
2266		DPRINT(("Cannot find unmapped area for size %ld\n", size));
2267		up_write(&task->mm->mmap_sem);
2268		goto error;
2269	}
2270	vma->vm_end = vma->vm_start + size;
2271	vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2272
2273	DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2274
2275	/* can only be applied to current task, need to have the mm semaphore held when called */
2276	if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2277		DPRINT(("Can't remap buffer\n"));
2278		up_write(&task->mm->mmap_sem);
2279		goto error;
2280	}
2281
 
 
2282	/*
2283	 * now insert the vma in the vm list for the process, must be
2284	 * done with mmap lock held
2285	 */
2286	insert_vm_struct(mm, vma);
2287
2288	vm_stat_account(vma->vm_mm, vma->vm_flags, vma_pages(vma));
 
 
2289	up_write(&task->mm->mmap_sem);
2290
2291	/*
2292	 * keep track of user level virtual address
2293	 */
2294	ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2295	*(unsigned long *)user_vaddr = vma->vm_start;
2296
2297	return 0;
2298
2299error:
2300	vm_area_free(vma);
2301error_kmem:
2302	vfree(smpl_buf);
2303
2304	return -ENOMEM;
2305}
2306
2307/*
2308 * XXX: do something better here
2309 */
2310static int
2311pfm_bad_permissions(struct task_struct *task)
2312{
2313	const struct cred *tcred;
2314	kuid_t uid = current_uid();
2315	kgid_t gid = current_gid();
2316	int ret;
2317
2318	rcu_read_lock();
2319	tcred = __task_cred(task);
2320
2321	/* inspired by ptrace_attach() */
2322	DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2323		from_kuid(&init_user_ns, uid),
2324		from_kgid(&init_user_ns, gid),
2325		from_kuid(&init_user_ns, tcred->euid),
2326		from_kuid(&init_user_ns, tcred->suid),
2327		from_kuid(&init_user_ns, tcred->uid),
2328		from_kgid(&init_user_ns, tcred->egid),
2329		from_kgid(&init_user_ns, tcred->sgid)));
2330
2331	ret = ((!uid_eq(uid, tcred->euid))
2332	       || (!uid_eq(uid, tcred->suid))
2333	       || (!uid_eq(uid, tcred->uid))
2334	       || (!gid_eq(gid, tcred->egid))
2335	       || (!gid_eq(gid, tcred->sgid))
2336	       || (!gid_eq(gid, tcred->gid))) && !capable(CAP_SYS_PTRACE);
2337
2338	rcu_read_unlock();
2339	return ret;
2340}
2341
2342static int
2343pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2344{
2345	int ctx_flags;
2346
2347	/* valid signal */
2348
2349	ctx_flags = pfx->ctx_flags;
2350
2351	if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2352
2353		/*
2354		 * cannot block in this mode
2355		 */
2356		if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2357			DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2358			return -EINVAL;
2359		}
2360	} else {
2361	}
2362	/* probably more to add here */
2363
2364	return 0;
2365}
2366
2367static int
2368pfm_setup_buffer_fmt(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned int ctx_flags,
2369		     unsigned int cpu, pfarg_context_t *arg)
2370{
2371	pfm_buffer_fmt_t *fmt = NULL;
2372	unsigned long size = 0UL;
2373	void *uaddr = NULL;
2374	void *fmt_arg = NULL;
2375	int ret = 0;
2376#define PFM_CTXARG_BUF_ARG(a)	(pfm_buffer_fmt_t *)(a+1)
2377
2378	/* invoke and lock buffer format, if found */
2379	fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2380	if (fmt == NULL) {
2381		DPRINT(("[%d] cannot find buffer format\n", task_pid_nr(task)));
2382		return -EINVAL;
2383	}
2384
2385	/*
2386	 * buffer argument MUST be contiguous to pfarg_context_t
2387	 */
2388	if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2389
2390	ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2391
2392	DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task_pid_nr(task), ctx_flags, cpu, fmt_arg, ret));
2393
2394	if (ret) goto error;
2395
2396	/* link buffer format and context */
2397	ctx->ctx_buf_fmt = fmt;
2398	ctx->ctx_fl_is_sampling = 1; /* assume record() is defined */
2399
2400	/*
2401	 * check if buffer format wants to use perfmon buffer allocation/mapping service
2402	 */
2403	ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2404	if (ret) goto error;
2405
2406	if (size) {
2407		/*
2408		 * buffer is always remapped into the caller's address space
2409		 */
2410		ret = pfm_smpl_buffer_alloc(current, filp, ctx, size, &uaddr);
2411		if (ret) goto error;
2412
2413		/* keep track of user address of buffer */
2414		arg->ctx_smpl_vaddr = uaddr;
2415	}
2416	ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2417
2418error:
2419	return ret;
2420}
2421
2422static void
2423pfm_reset_pmu_state(pfm_context_t *ctx)
2424{
2425	int i;
2426
2427	/*
2428	 * install reset values for PMC.
2429	 */
2430	for (i=1; PMC_IS_LAST(i) == 0; i++) {
2431		if (PMC_IS_IMPL(i) == 0) continue;
2432		ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2433		DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2434	}
2435	/*
2436	 * PMD registers are set to 0UL when the context in memset()
2437	 */
2438
2439	/*
2440	 * On context switched restore, we must restore ALL pmc and ALL pmd even
2441	 * when they are not actively used by the task. In UP, the incoming process
2442	 * may otherwise pick up left over PMC, PMD state from the previous process.
2443	 * As opposed to PMD, stale PMC can cause harm to the incoming
2444	 * process because they may change what is being measured.
2445	 * Therefore, we must systematically reinstall the entire
2446	 * PMC state. In SMP, the same thing is possible on the
2447	 * same CPU but also on between 2 CPUs.
2448	 *
2449	 * The problem with PMD is information leaking especially
2450	 * to user level when psr.sp=0
2451	 *
2452	 * There is unfortunately no easy way to avoid this problem
2453	 * on either UP or SMP. This definitively slows down the
2454	 * pfm_load_regs() function.
2455	 */
2456
2457	 /*
2458	  * bitmask of all PMCs accessible to this context
2459	  *
2460	  * PMC0 is treated differently.
2461	  */
2462	ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2463
2464	/*
2465	 * bitmask of all PMDs that are accessible to this context
2466	 */
2467	ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2468
2469	DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2470
2471	/*
2472	 * useful in case of re-enable after disable
2473	 */
2474	ctx->ctx_used_ibrs[0] = 0UL;
2475	ctx->ctx_used_dbrs[0] = 0UL;
2476}
2477
2478static int
2479pfm_ctx_getsize(void *arg, size_t *sz)
2480{
2481	pfarg_context_t *req = (pfarg_context_t *)arg;
2482	pfm_buffer_fmt_t *fmt;
2483
2484	*sz = 0;
2485
2486	if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2487
2488	fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2489	if (fmt == NULL) {
2490		DPRINT(("cannot find buffer format\n"));
2491		return -EINVAL;
2492	}
2493	/* get just enough to copy in user parameters */
2494	*sz = fmt->fmt_arg_size;
2495	DPRINT(("arg_size=%lu\n", *sz));
2496
2497	return 0;
2498}
2499
2500
2501
2502/*
2503 * cannot attach if :
2504 * 	- kernel task
2505 * 	- task not owned by caller
2506 * 	- task incompatible with context mode
2507 */
2508static int
2509pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2510{
2511	/*
2512	 * no kernel task or task not owner by caller
2513	 */
2514	if (task->mm == NULL) {
2515		DPRINT(("task [%d] has not memory context (kernel thread)\n", task_pid_nr(task)));
2516		return -EPERM;
2517	}
2518	if (pfm_bad_permissions(task)) {
2519		DPRINT(("no permission to attach to  [%d]\n", task_pid_nr(task)));
2520		return -EPERM;
2521	}
2522	/*
2523	 * cannot block in self-monitoring mode
2524	 */
2525	if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2526		DPRINT(("cannot load a blocking context on self for [%d]\n", task_pid_nr(task)));
2527		return -EINVAL;
2528	}
2529
2530	if (task->exit_state == EXIT_ZOMBIE) {
2531		DPRINT(("cannot attach to  zombie task [%d]\n", task_pid_nr(task)));
2532		return -EBUSY;
2533	}
2534
2535	/*
2536	 * always ok for self
2537	 */
2538	if (task == current) return 0;
2539
2540	if (!task_is_stopped_or_traced(task)) {
2541		DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task_pid_nr(task), task->state));
2542		return -EBUSY;
2543	}
2544	/*
2545	 * make sure the task is off any CPU
2546	 */
2547	wait_task_inactive(task, 0);
2548
2549	/* more to come... */
2550
2551	return 0;
2552}
2553
2554static int
2555pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2556{
2557	struct task_struct *p = current;
2558	int ret;
2559
2560	/* XXX: need to add more checks here */
2561	if (pid < 2) return -EPERM;
2562
2563	if (pid != task_pid_vnr(current)) {
 
 
 
 
 
2564		/* make sure task cannot go away while we operate on it */
2565		p = find_get_task_by_vpid(pid);
2566		if (!p)
2567			return -ESRCH;
 
 
2568	}
2569
2570	ret = pfm_task_incompatible(ctx, p);
2571	if (ret == 0) {
2572		*task = p;
2573	} else if (p != current) {
2574		pfm_put_task(p);
2575	}
2576	return ret;
2577}
2578
2579
2580
2581static int
2582pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2583{
2584	pfarg_context_t *req = (pfarg_context_t *)arg;
2585	struct file *filp;
2586	struct path path;
2587	int ctx_flags;
2588	int fd;
2589	int ret;
2590
2591	/* let's check the arguments first */
2592	ret = pfarg_is_sane(current, req);
2593	if (ret < 0)
2594		return ret;
2595
2596	ctx_flags = req->ctx_flags;
2597
2598	ret = -ENOMEM;
2599
2600	fd = get_unused_fd_flags(0);
2601	if (fd < 0)
2602		return fd;
2603
2604	ctx = pfm_context_alloc(ctx_flags);
2605	if (!ctx)
2606		goto error;
2607
2608	filp = pfm_alloc_file(ctx);
2609	if (IS_ERR(filp)) {
2610		ret = PTR_ERR(filp);
2611		goto error_file;
2612	}
2613
2614	req->ctx_fd = ctx->ctx_fd = fd;
2615
2616	/*
2617	 * does the user want to sample?
2618	 */
2619	if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2620		ret = pfm_setup_buffer_fmt(current, filp, ctx, ctx_flags, 0, req);
2621		if (ret)
2622			goto buffer_error;
2623	}
2624
2625	DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d\n",
2626		ctx,
2627		ctx_flags,
2628		ctx->ctx_fl_system,
2629		ctx->ctx_fl_block,
2630		ctx->ctx_fl_excl_idle,
2631		ctx->ctx_fl_no_msg,
2632		ctx->ctx_fd));
2633
2634	/*
2635	 * initialize soft PMU state
2636	 */
2637	pfm_reset_pmu_state(ctx);
2638
2639	fd_install(fd, filp);
2640
2641	return 0;
2642
2643buffer_error:
2644	path = filp->f_path;
2645	put_filp(filp);
2646	path_put(&path);
2647
2648	if (ctx->ctx_buf_fmt) {
2649		pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2650	}
2651error_file:
2652	pfm_context_free(ctx);
2653
2654error:
2655	put_unused_fd(fd);
2656	return ret;
2657}
2658
2659static inline unsigned long
2660pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2661{
2662	unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2663	unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2664	extern unsigned long carta_random32 (unsigned long seed);
2665
2666	if (reg->flags & PFM_REGFL_RANDOM) {
2667		new_seed = carta_random32(old_seed);
2668		val -= (old_seed & mask);	/* counter values are negative numbers! */
2669		if ((mask >> 32) != 0)
2670			/* construct a full 64-bit random value: */
2671			new_seed |= carta_random32(old_seed >> 32) << 32;
2672		reg->seed = new_seed;
2673	}
2674	reg->lval = val;
2675	return val;
2676}
2677
2678static void
2679pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2680{
2681	unsigned long mask = ovfl_regs[0];
2682	unsigned long reset_others = 0UL;
2683	unsigned long val;
2684	int i;
2685
2686	/*
2687	 * now restore reset value on sampling overflowed counters
2688	 */
2689	mask >>= PMU_FIRST_COUNTER;
2690	for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2691
2692		if ((mask & 0x1UL) == 0UL) continue;
2693
2694		ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2695		reset_others        |= ctx->ctx_pmds[i].reset_pmds[0];
2696
2697		DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2698	}
2699
2700	/*
2701	 * Now take care of resetting the other registers
2702	 */
2703	for(i = 0; reset_others; i++, reset_others >>= 1) {
2704
2705		if ((reset_others & 0x1) == 0) continue;
2706
2707		ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2708
2709		DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2710			  is_long_reset ? "long" : "short", i, val));
2711	}
2712}
2713
2714static void
2715pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2716{
2717	unsigned long mask = ovfl_regs[0];
2718	unsigned long reset_others = 0UL;
2719	unsigned long val;
2720	int i;
2721
2722	DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2723
2724	if (ctx->ctx_state == PFM_CTX_MASKED) {
2725		pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2726		return;
2727	}
2728
2729	/*
2730	 * now restore reset value on sampling overflowed counters
2731	 */
2732	mask >>= PMU_FIRST_COUNTER;
2733	for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2734
2735		if ((mask & 0x1UL) == 0UL) continue;
2736
2737		val           = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2738		reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2739
2740		DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2741
2742		pfm_write_soft_counter(ctx, i, val);
2743	}
2744
2745	/*
2746	 * Now take care of resetting the other registers
2747	 */
2748	for(i = 0; reset_others; i++, reset_others >>= 1) {
2749
2750		if ((reset_others & 0x1) == 0) continue;
2751
2752		val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2753
2754		if (PMD_IS_COUNTING(i)) {
2755			pfm_write_soft_counter(ctx, i, val);
2756		} else {
2757			ia64_set_pmd(i, val);
2758		}
2759		DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2760			  is_long_reset ? "long" : "short", i, val));
2761	}
2762	ia64_srlz_d();
2763}
2764
2765static int
2766pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2767{
2768	struct task_struct *task;
2769	pfarg_reg_t *req = (pfarg_reg_t *)arg;
2770	unsigned long value, pmc_pm;
2771	unsigned long smpl_pmds, reset_pmds, impl_pmds;
2772	unsigned int cnum, reg_flags, flags, pmc_type;
2773	int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2774	int is_monitor, is_counting, state;
2775	int ret = -EINVAL;
2776	pfm_reg_check_t	wr_func;
2777#define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2778
2779	state     = ctx->ctx_state;
2780	is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2781	is_system = ctx->ctx_fl_system;
2782	task      = ctx->ctx_task;
2783	impl_pmds = pmu_conf->impl_pmds[0];
2784
2785	if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2786
2787	if (is_loaded) {
2788		/*
2789		 * In system wide and when the context is loaded, access can only happen
2790		 * when the caller is running on the CPU being monitored by the session.
2791		 * It does not have to be the owner (ctx_task) of the context per se.
2792		 */
2793		if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2794			DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2795			return -EBUSY;
2796		}
2797		can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2798	}
2799	expert_mode = pfm_sysctl.expert_mode; 
2800
2801	for (i = 0; i < count; i++, req++) {
2802
2803		cnum       = req->reg_num;
2804		reg_flags  = req->reg_flags;
2805		value      = req->reg_value;
2806		smpl_pmds  = req->reg_smpl_pmds[0];
2807		reset_pmds = req->reg_reset_pmds[0];
2808		flags      = 0;
2809
2810
2811		if (cnum >= PMU_MAX_PMCS) {
2812			DPRINT(("pmc%u is invalid\n", cnum));
2813			goto error;
2814		}
2815
2816		pmc_type   = pmu_conf->pmc_desc[cnum].type;
2817		pmc_pm     = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2818		is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2819		is_monitor  = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2820
2821		/*
2822		 * we reject all non implemented PMC as well
2823		 * as attempts to modify PMC[0-3] which are used
2824		 * as status registers by the PMU
2825		 */
2826		if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2827			DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2828			goto error;
2829		}
2830		wr_func = pmu_conf->pmc_desc[cnum].write_check;
2831		/*
2832		 * If the PMC is a monitor, then if the value is not the default:
2833		 * 	- system-wide session: PMCx.pm=1 (privileged monitor)
2834		 * 	- per-task           : PMCx.pm=0 (user monitor)
2835		 */
2836		if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2837			DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2838				cnum,
2839				pmc_pm,
2840				is_system));
2841			goto error;
2842		}
2843
2844		if (is_counting) {
2845			/*
2846		 	 * enforce generation of overflow interrupt. Necessary on all
2847		 	 * CPUs.
2848		 	 */
2849			value |= 1 << PMU_PMC_OI;
2850
2851			if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2852				flags |= PFM_REGFL_OVFL_NOTIFY;
2853			}
2854
2855			if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2856
2857			/* verify validity of smpl_pmds */
2858			if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2859				DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2860				goto error;
2861			}
2862
2863			/* verify validity of reset_pmds */
2864			if ((reset_pmds & impl_pmds) != reset_pmds) {
2865				DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2866				goto error;
2867			}
2868		} else {
2869			if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2870				DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2871				goto error;
2872			}
2873			/* eventid on non-counting monitors are ignored */
2874		}
2875
2876		/*
2877		 * execute write checker, if any
2878		 */
2879		if (likely(expert_mode == 0 && wr_func)) {
2880			ret = (*wr_func)(task, ctx, cnum, &value, regs);
2881			if (ret) goto error;
2882			ret = -EINVAL;
2883		}
2884
2885		/*
2886		 * no error on this register
2887		 */
2888		PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2889
2890		/*
2891		 * Now we commit the changes to the software state
2892		 */
2893
2894		/*
2895		 * update overflow information
2896		 */
2897		if (is_counting) {
2898			/*
2899		 	 * full flag update each time a register is programmed
2900		 	 */
2901			ctx->ctx_pmds[cnum].flags = flags;
2902
2903			ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2904			ctx->ctx_pmds[cnum].smpl_pmds[0]  = smpl_pmds;
2905			ctx->ctx_pmds[cnum].eventid       = req->reg_smpl_eventid;
2906
2907			/*
2908			 * Mark all PMDS to be accessed as used.
2909			 *
2910			 * We do not keep track of PMC because we have to
2911			 * systematically restore ALL of them.
2912			 *
2913			 * We do not update the used_monitors mask, because
2914			 * if we have not programmed them, then will be in
2915			 * a quiescent state, therefore we will not need to
2916			 * mask/restore then when context is MASKED.
2917			 */
2918			CTX_USED_PMD(ctx, reset_pmds);
2919			CTX_USED_PMD(ctx, smpl_pmds);
2920			/*
2921		 	 * make sure we do not try to reset on
2922		 	 * restart because we have established new values
2923		 	 */
2924			if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
2925		}
2926		/*
2927		 * Needed in case the user does not initialize the equivalent
2928		 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
2929		 * possible leak here.
2930		 */
2931		CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
2932
2933		/*
2934		 * keep track of the monitor PMC that we are using.
2935		 * we save the value of the pmc in ctx_pmcs[] and if
2936		 * the monitoring is not stopped for the context we also
2937		 * place it in the saved state area so that it will be
2938		 * picked up later by the context switch code.
2939		 *
2940		 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
2941		 *
2942		 * The value in th_pmcs[] may be modified on overflow, i.e.,  when
2943		 * monitoring needs to be stopped.
2944		 */
2945		if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
2946
2947		/*
2948		 * update context state
2949		 */
2950		ctx->ctx_pmcs[cnum] = value;
2951
2952		if (is_loaded) {
2953			/*
2954			 * write thread state
2955			 */
2956			if (is_system == 0) ctx->th_pmcs[cnum] = value;
2957
2958			/*
2959			 * write hardware register if we can
2960			 */
2961			if (can_access_pmu) {
2962				ia64_set_pmc(cnum, value);
2963			}
2964#ifdef CONFIG_SMP
2965			else {
2966				/*
2967				 * per-task SMP only here
2968				 *
2969			 	 * we are guaranteed that the task is not running on the other CPU,
2970			 	 * we indicate that this PMD will need to be reloaded if the task
2971			 	 * is rescheduled on the CPU it ran last on.
2972			 	 */
2973				ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
2974			}
2975#endif
2976		}
2977
2978		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",
2979			  cnum,
2980			  value,
2981			  is_loaded,
2982			  can_access_pmu,
2983			  flags,
2984			  ctx->ctx_all_pmcs[0],
2985			  ctx->ctx_used_pmds[0],
2986			  ctx->ctx_pmds[cnum].eventid,
2987			  smpl_pmds,
2988			  reset_pmds,
2989			  ctx->ctx_reload_pmcs[0],
2990			  ctx->ctx_used_monitors[0],
2991			  ctx->ctx_ovfl_regs[0]));
2992	}
2993
2994	/*
2995	 * make sure the changes are visible
2996	 */
2997	if (can_access_pmu) ia64_srlz_d();
2998
2999	return 0;
3000error:
3001	PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3002	return ret;
3003}
3004
3005static int
3006pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3007{
3008	struct task_struct *task;
3009	pfarg_reg_t *req = (pfarg_reg_t *)arg;
3010	unsigned long value, hw_value, ovfl_mask;
3011	unsigned int cnum;
3012	int i, can_access_pmu = 0, state;
3013	int is_counting, is_loaded, is_system, expert_mode;
3014	int ret = -EINVAL;
3015	pfm_reg_check_t wr_func;
3016
3017
3018	state     = ctx->ctx_state;
3019	is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3020	is_system = ctx->ctx_fl_system;
3021	ovfl_mask = pmu_conf->ovfl_val;
3022	task      = ctx->ctx_task;
3023
3024	if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3025
3026	/*
3027	 * on both UP and SMP, we can only write to the PMC when the task is
3028	 * the owner of the local PMU.
3029	 */
3030	if (likely(is_loaded)) {
3031		/*
3032		 * In system wide and when the context is loaded, access can only happen
3033		 * when the caller is running on the CPU being monitored by the session.
3034		 * It does not have to be the owner (ctx_task) of the context per se.
3035		 */
3036		if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3037			DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3038			return -EBUSY;
3039		}
3040		can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3041	}
3042	expert_mode = pfm_sysctl.expert_mode; 
3043
3044	for (i = 0; i < count; i++, req++) {
3045
3046		cnum  = req->reg_num;
3047		value = req->reg_value;
3048
3049		if (!PMD_IS_IMPL(cnum)) {
3050			DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3051			goto abort_mission;
3052		}
3053		is_counting = PMD_IS_COUNTING(cnum);
3054		wr_func     = pmu_conf->pmd_desc[cnum].write_check;
3055
3056		/*
3057		 * execute write checker, if any
3058		 */
3059		if (unlikely(expert_mode == 0 && wr_func)) {
3060			unsigned long v = value;
3061
3062			ret = (*wr_func)(task, ctx, cnum, &v, regs);
3063			if (ret) goto abort_mission;
3064
3065			value = v;
3066			ret   = -EINVAL;
3067		}
3068
3069		/*
3070		 * no error on this register
3071		 */
3072		PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3073
3074		/*
3075		 * now commit changes to software state
3076		 */
3077		hw_value = value;
3078
3079		/*
3080		 * update virtualized (64bits) counter
3081		 */
3082		if (is_counting) {
3083			/*
3084			 * write context state
3085			 */
3086			ctx->ctx_pmds[cnum].lval = value;
3087
3088			/*
3089			 * when context is load we use the split value
3090			 */
3091			if (is_loaded) {
3092				hw_value = value &  ovfl_mask;
3093				value    = value & ~ovfl_mask;
3094			}
3095		}
3096		/*
3097		 * update reset values (not just for counters)
3098		 */
3099		ctx->ctx_pmds[cnum].long_reset  = req->reg_long_reset;
3100		ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3101
3102		/*
3103		 * update randomization parameters (not just for counters)
3104		 */
3105		ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3106		ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3107
3108		/*
3109		 * update context value
3110		 */
3111		ctx->ctx_pmds[cnum].val  = value;
3112
3113		/*
3114		 * Keep track of what we use
3115		 *
3116		 * We do not keep track of PMC because we have to
3117		 * systematically restore ALL of them.
3118		 */
3119		CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3120
3121		/*
3122		 * mark this PMD register used as well
3123		 */
3124		CTX_USED_PMD(ctx, RDEP(cnum));
3125
3126		/*
3127		 * make sure we do not try to reset on
3128		 * restart because we have established new values
3129		 */
3130		if (is_counting && state == PFM_CTX_MASKED) {
3131			ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3132		}
3133
3134		if (is_loaded) {
3135			/*
3136		 	 * write thread state
3137		 	 */
3138			if (is_system == 0) ctx->th_pmds[cnum] = hw_value;
3139
3140			/*
3141			 * write hardware register if we can
3142			 */
3143			if (can_access_pmu) {
3144				ia64_set_pmd(cnum, hw_value);
3145			} else {
3146#ifdef CONFIG_SMP
3147				/*
3148			 	 * we are guaranteed that the task is not running on the other CPU,
3149			 	 * we indicate that this PMD will need to be reloaded if the task
3150			 	 * is rescheduled on the CPU it ran last on.
3151			 	 */
3152				ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3153#endif
3154			}
3155		}
3156
3157		DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx  short_reset=0x%lx "
3158			  "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",
3159			cnum,
3160			value,
3161			is_loaded,
3162			can_access_pmu,
3163			hw_value,
3164			ctx->ctx_pmds[cnum].val,
3165			ctx->ctx_pmds[cnum].short_reset,
3166			ctx->ctx_pmds[cnum].long_reset,
3167			PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3168			ctx->ctx_pmds[cnum].seed,
3169			ctx->ctx_pmds[cnum].mask,
3170			ctx->ctx_used_pmds[0],
3171			ctx->ctx_pmds[cnum].reset_pmds[0],
3172			ctx->ctx_reload_pmds[0],
3173			ctx->ctx_all_pmds[0],
3174			ctx->ctx_ovfl_regs[0]));
3175	}
3176
3177	/*
3178	 * make changes visible
3179	 */
3180	if (can_access_pmu) ia64_srlz_d();
3181
3182	return 0;
3183
3184abort_mission:
3185	/*
3186	 * for now, we have only one possibility for error
3187	 */
3188	PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3189	return ret;
3190}
3191
3192/*
3193 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3194 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3195 * interrupt is delivered during the call, it will be kept pending until we leave, making
3196 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3197 * guaranteed to return consistent data to the user, it may simply be old. It is not
3198 * trivial to treat the overflow while inside the call because you may end up in
3199 * some module sampling buffer code causing deadlocks.
3200 */
3201static int
3202pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3203{
3204	struct task_struct *task;
3205	unsigned long val = 0UL, lval, ovfl_mask, sval;
3206	pfarg_reg_t *req = (pfarg_reg_t *)arg;
3207	unsigned int cnum, reg_flags = 0;
3208	int i, can_access_pmu = 0, state;
3209	int is_loaded, is_system, is_counting, expert_mode;
3210	int ret = -EINVAL;
3211	pfm_reg_check_t rd_func;
3212
3213	/*
3214	 * access is possible when loaded only for
3215	 * self-monitoring tasks or in UP mode
3216	 */
3217
3218	state     = ctx->ctx_state;
3219	is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3220	is_system = ctx->ctx_fl_system;
3221	ovfl_mask = pmu_conf->ovfl_val;
3222	task      = ctx->ctx_task;
3223
3224	if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3225
3226	if (likely(is_loaded)) {
3227		/*
3228		 * In system wide and when the context is loaded, access can only happen
3229		 * when the caller is running on the CPU being monitored by the session.
3230		 * It does not have to be the owner (ctx_task) of the context per se.
3231		 */
3232		if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3233			DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3234			return -EBUSY;
3235		}
3236		/*
3237		 * this can be true when not self-monitoring only in UP
3238		 */
3239		can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3240
3241		if (can_access_pmu) ia64_srlz_d();
3242	}
3243	expert_mode = pfm_sysctl.expert_mode; 
3244
3245	DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3246		is_loaded,
3247		can_access_pmu,
3248		state));
3249
3250	/*
3251	 * on both UP and SMP, we can only read the PMD from the hardware register when
3252	 * the task is the owner of the local PMU.
3253	 */
3254
3255	for (i = 0; i < count; i++, req++) {
3256
3257		cnum        = req->reg_num;
3258		reg_flags   = req->reg_flags;
3259
3260		if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3261		/*
3262		 * we can only read the register that we use. That includes
3263		 * the one we explicitly initialize AND the one we want included
3264		 * in the sampling buffer (smpl_regs).
3265		 *
3266		 * Having this restriction allows optimization in the ctxsw routine
3267		 * without compromising security (leaks)
3268		 */
3269		if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3270
3271		sval        = ctx->ctx_pmds[cnum].val;
3272		lval        = ctx->ctx_pmds[cnum].lval;
3273		is_counting = PMD_IS_COUNTING(cnum);
3274
3275		/*
3276		 * If the task is not the current one, then we check if the
3277		 * PMU state is still in the local live register due to lazy ctxsw.
3278		 * If true, then we read directly from the registers.
3279		 */
3280		if (can_access_pmu){
3281			val = ia64_get_pmd(cnum);
3282		} else {
3283			/*
3284			 * context has been saved
3285			 * if context is zombie, then task does not exist anymore.
3286			 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3287			 */
3288			val = is_loaded ? ctx->th_pmds[cnum] : 0UL;
3289		}
3290		rd_func = pmu_conf->pmd_desc[cnum].read_check;
3291
3292		if (is_counting) {
3293			/*
3294			 * XXX: need to check for overflow when loaded
3295			 */
3296			val &= ovfl_mask;
3297			val += sval;
3298		}
3299
3300		/*
3301		 * execute read checker, if any
3302		 */
3303		if (unlikely(expert_mode == 0 && rd_func)) {
3304			unsigned long v = val;
3305			ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3306			if (ret) goto error;
3307			val = v;
3308			ret = -EINVAL;
3309		}
3310
3311		PFM_REG_RETFLAG_SET(reg_flags, 0);
3312
3313		DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3314
3315		/*
3316		 * update register return value, abort all if problem during copy.
3317		 * we only modify the reg_flags field. no check mode is fine because
3318		 * access has been verified upfront in sys_perfmonctl().
3319		 */
3320		req->reg_value            = val;
3321		req->reg_flags            = reg_flags;
3322		req->reg_last_reset_val   = lval;
3323	}
3324
3325	return 0;
3326
3327error:
3328	PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3329	return ret;
3330}
3331
3332int
3333pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3334{
3335	pfm_context_t *ctx;
3336
3337	if (req == NULL) return -EINVAL;
3338
3339 	ctx = GET_PMU_CTX();
3340
3341	if (ctx == NULL) return -EINVAL;
3342
3343	/*
3344	 * for now limit to current task, which is enough when calling
3345	 * from overflow handler
3346	 */
3347	if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3348
3349	return pfm_write_pmcs(ctx, req, nreq, regs);
3350}
3351EXPORT_SYMBOL(pfm_mod_write_pmcs);
3352
3353int
3354pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3355{
3356	pfm_context_t *ctx;
3357
3358	if (req == NULL) return -EINVAL;
3359
3360 	ctx = GET_PMU_CTX();
3361
3362	if (ctx == NULL) return -EINVAL;
3363
3364	/*
3365	 * for now limit to current task, which is enough when calling
3366	 * from overflow handler
3367	 */
3368	if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3369
3370	return pfm_read_pmds(ctx, req, nreq, regs);
3371}
3372EXPORT_SYMBOL(pfm_mod_read_pmds);
3373
3374/*
3375 * Only call this function when a process it trying to
3376 * write the debug registers (reading is always allowed)
3377 */
3378int
3379pfm_use_debug_registers(struct task_struct *task)
3380{
3381	pfm_context_t *ctx = task->thread.pfm_context;
3382	unsigned long flags;
3383	int ret = 0;
3384
3385	if (pmu_conf->use_rr_dbregs == 0) return 0;
3386
3387	DPRINT(("called for [%d]\n", task_pid_nr(task)));
3388
3389	/*
3390	 * do it only once
3391	 */
3392	if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3393
3394	/*
3395	 * Even on SMP, we do not need to use an atomic here because
3396	 * the only way in is via ptrace() and this is possible only when the
3397	 * process is stopped. Even in the case where the ctxsw out is not totally
3398	 * completed by the time we come here, there is no way the 'stopped' process
3399	 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3400	 * So this is always safe.
3401	 */
3402	if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3403
3404	LOCK_PFS(flags);
3405
3406	/*
3407	 * We cannot allow setting breakpoints when system wide monitoring
3408	 * sessions are using the debug registers.
3409	 */
3410	if (pfm_sessions.pfs_sys_use_dbregs> 0)
3411		ret = -1;
3412	else
3413		pfm_sessions.pfs_ptrace_use_dbregs++;
3414
3415	DPRINT(("ptrace_use_dbregs=%u  sys_use_dbregs=%u by [%d] ret = %d\n",
3416		  pfm_sessions.pfs_ptrace_use_dbregs,
3417		  pfm_sessions.pfs_sys_use_dbregs,
3418		  task_pid_nr(task), ret));
3419
3420	UNLOCK_PFS(flags);
3421
3422	return ret;
3423}
3424
3425/*
3426 * This function is called for every task that exits with the
3427 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3428 * able to use the debug registers for debugging purposes via
3429 * ptrace(). Therefore we know it was not using them for
3430 * performance monitoring, so we only decrement the number
3431 * of "ptraced" debug register users to keep the count up to date
3432 */
3433int
3434pfm_release_debug_registers(struct task_struct *task)
3435{
3436	unsigned long flags;
3437	int ret;
3438
3439	if (pmu_conf->use_rr_dbregs == 0) return 0;
3440
3441	LOCK_PFS(flags);
3442	if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3443		printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task_pid_nr(task));
3444		ret = -1;
3445	}  else {
3446		pfm_sessions.pfs_ptrace_use_dbregs--;
3447		ret = 0;
3448	}
3449	UNLOCK_PFS(flags);
3450
3451	return ret;
3452}
3453
3454static int
3455pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3456{
3457	struct task_struct *task;
3458	pfm_buffer_fmt_t *fmt;
3459	pfm_ovfl_ctrl_t rst_ctrl;
3460	int state, is_system;
3461	int ret = 0;
3462
3463	state     = ctx->ctx_state;
3464	fmt       = ctx->ctx_buf_fmt;
3465	is_system = ctx->ctx_fl_system;
3466	task      = PFM_CTX_TASK(ctx);
3467
3468	switch(state) {
3469		case PFM_CTX_MASKED:
3470			break;
3471		case PFM_CTX_LOADED: 
3472			if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3473			/* fall through */
3474		case PFM_CTX_UNLOADED:
3475		case PFM_CTX_ZOMBIE:
3476			DPRINT(("invalid state=%d\n", state));
3477			return -EBUSY;
3478		default:
3479			DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3480			return -EINVAL;
3481	}
3482
3483	/*
3484 	 * In system wide and when the context is loaded, access can only happen
3485 	 * when the caller is running on the CPU being monitored by the session.
3486 	 * It does not have to be the owner (ctx_task) of the context per se.
3487 	 */
3488	if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3489		DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3490		return -EBUSY;
3491	}
3492
3493	/* sanity check */
3494	if (unlikely(task == NULL)) {
3495		printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", task_pid_nr(current));
3496		return -EINVAL;
3497	}
3498
3499	if (task == current || is_system) {
3500
3501		fmt = ctx->ctx_buf_fmt;
3502
3503		DPRINT(("restarting self %d ovfl=0x%lx\n",
3504			task_pid_nr(task),
3505			ctx->ctx_ovfl_regs[0]));
3506
3507		if (CTX_HAS_SMPL(ctx)) {
3508
3509			prefetch(ctx->ctx_smpl_hdr);
3510
3511			rst_ctrl.bits.mask_monitoring = 0;
3512			rst_ctrl.bits.reset_ovfl_pmds = 0;
3513
3514			if (state == PFM_CTX_LOADED)
3515				ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3516			else
3517				ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3518		} else {
3519			rst_ctrl.bits.mask_monitoring = 0;
3520			rst_ctrl.bits.reset_ovfl_pmds = 1;
3521		}
3522
3523		if (ret == 0) {
3524			if (rst_ctrl.bits.reset_ovfl_pmds)
3525				pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3526
3527			if (rst_ctrl.bits.mask_monitoring == 0) {
3528				DPRINT(("resuming monitoring for [%d]\n", task_pid_nr(task)));
3529
3530				if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3531			} else {
3532				DPRINT(("keeping monitoring stopped for [%d]\n", task_pid_nr(task)));
3533
3534				// cannot use pfm_stop_monitoring(task, regs);
3535			}
3536		}
3537		/*
3538		 * clear overflowed PMD mask to remove any stale information
3539		 */
3540		ctx->ctx_ovfl_regs[0] = 0UL;
3541
3542		/*
3543		 * back to LOADED state
3544		 */
3545		ctx->ctx_state = PFM_CTX_LOADED;
3546
3547		/*
3548		 * XXX: not really useful for self monitoring
3549		 */
3550		ctx->ctx_fl_can_restart = 0;
3551
3552		return 0;
3553	}
3554
3555	/* 
3556	 * restart another task
3557	 */
3558
3559	/*
3560	 * When PFM_CTX_MASKED, we cannot issue a restart before the previous 
3561	 * one is seen by the task.
3562	 */
3563	if (state == PFM_CTX_MASKED) {
3564		if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3565		/*
3566		 * will prevent subsequent restart before this one is
3567		 * seen by other task
3568		 */
3569		ctx->ctx_fl_can_restart = 0;
3570	}
3571
3572	/*
3573	 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3574	 * the task is blocked or on its way to block. That's the normal
3575	 * restart path. If the monitoring is not masked, then the task
3576	 * can be actively monitoring and we cannot directly intervene.
3577	 * Therefore we use the trap mechanism to catch the task and
3578	 * force it to reset the buffer/reset PMDs.
3579	 *
3580	 * if non-blocking, then we ensure that the task will go into
3581	 * pfm_handle_work() before returning to user mode.
3582	 *
3583	 * We cannot explicitly reset another task, it MUST always
3584	 * be done by the task itself. This works for system wide because
3585	 * the tool that is controlling the session is logically doing 
3586	 * "self-monitoring".
3587	 */
3588	if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3589		DPRINT(("unblocking [%d]\n", task_pid_nr(task)));
3590		complete(&ctx->ctx_restart_done);
3591	} else {
3592		DPRINT(("[%d] armed exit trap\n", task_pid_nr(task)));
3593
3594		ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3595
3596		PFM_SET_WORK_PENDING(task, 1);
3597
3598		set_notify_resume(task);
3599
3600		/*
3601		 * XXX: send reschedule if task runs on another CPU
3602		 */
3603	}
3604	return 0;
3605}
3606
3607static int
3608pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3609{
3610	unsigned int m = *(unsigned int *)arg;
3611
3612	pfm_sysctl.debug = m == 0 ? 0 : 1;
3613
3614	printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3615
3616	if (m == 0) {
3617		memset(pfm_stats, 0, sizeof(pfm_stats));
3618		for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3619	}
3620	return 0;
3621}
3622
3623/*
3624 * arg can be NULL and count can be zero for this function
3625 */
3626static int
3627pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3628{
3629	struct thread_struct *thread = NULL;
3630	struct task_struct *task;
3631	pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3632	unsigned long flags;
3633	dbreg_t dbreg;
3634	unsigned int rnum;
3635	int first_time;
3636	int ret = 0, state;
3637	int i, can_access_pmu = 0;
3638	int is_system, is_loaded;
3639
3640	if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3641
3642	state     = ctx->ctx_state;
3643	is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3644	is_system = ctx->ctx_fl_system;
3645	task      = ctx->ctx_task;
3646
3647	if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3648
3649	/*
3650	 * on both UP and SMP, we can only write to the PMC when the task is
3651	 * the owner of the local PMU.
3652	 */
3653	if (is_loaded) {
3654		thread = &task->thread;
3655		/*
3656		 * In system wide and when the context is loaded, access can only happen
3657		 * when the caller is running on the CPU being monitored by the session.
3658		 * It does not have to be the owner (ctx_task) of the context per se.
3659		 */
3660		if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3661			DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3662			return -EBUSY;
3663		}
3664		can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3665	}
3666
3667	/*
3668	 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3669	 * ensuring that no real breakpoint can be installed via this call.
3670	 *
3671	 * IMPORTANT: regs can be NULL in this function
3672	 */
3673
3674	first_time = ctx->ctx_fl_using_dbreg == 0;
3675
3676	/*
3677	 * don't bother if we are loaded and task is being debugged
3678	 */
3679	if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3680		DPRINT(("debug registers already in use for [%d]\n", task_pid_nr(task)));
3681		return -EBUSY;
3682	}
3683
3684	/*
3685	 * check for debug registers in system wide mode
3686	 *
3687	 * If though a check is done in pfm_context_load(),
3688	 * we must repeat it here, in case the registers are
3689	 * written after the context is loaded
3690	 */
3691	if (is_loaded) {
3692		LOCK_PFS(flags);
3693
3694		if (first_time && is_system) {
3695			if (pfm_sessions.pfs_ptrace_use_dbregs)
3696				ret = -EBUSY;
3697			else
3698				pfm_sessions.pfs_sys_use_dbregs++;
3699		}
3700		UNLOCK_PFS(flags);
3701	}
3702
3703	if (ret != 0) return ret;
3704
3705	/*
3706	 * mark ourself as user of the debug registers for
3707	 * perfmon purposes.
3708	 */
3709	ctx->ctx_fl_using_dbreg = 1;
3710
3711	/*
3712 	 * clear hardware registers to make sure we don't
3713 	 * pick up stale state.
3714	 *
3715	 * for a system wide session, we do not use
3716	 * thread.dbr, thread.ibr because this process
3717	 * never leaves the current CPU and the state
3718	 * is shared by all processes running on it
3719 	 */
3720	if (first_time && can_access_pmu) {
3721		DPRINT(("[%d] clearing ibrs, dbrs\n", task_pid_nr(task)));
3722		for (i=0; i < pmu_conf->num_ibrs; i++) {
3723			ia64_set_ibr(i, 0UL);
3724			ia64_dv_serialize_instruction();
3725		}
3726		ia64_srlz_i();
3727		for (i=0; i < pmu_conf->num_dbrs; i++) {
3728			ia64_set_dbr(i, 0UL);
3729			ia64_dv_serialize_data();
3730		}
3731		ia64_srlz_d();
3732	}
3733
3734	/*
3735	 * Now install the values into the registers
3736	 */
3737	for (i = 0; i < count; i++, req++) {
3738
3739		rnum      = req->dbreg_num;
3740		dbreg.val = req->dbreg_value;
3741
3742		ret = -EINVAL;
3743
3744		if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3745			DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3746				  rnum, dbreg.val, mode, i, count));
3747
3748			goto abort_mission;
3749		}
3750
3751		/*
3752		 * make sure we do not install enabled breakpoint
3753		 */
3754		if (rnum & 0x1) {
3755			if (mode == PFM_CODE_RR)
3756				dbreg.ibr.ibr_x = 0;
3757			else
3758				dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3759		}
3760
3761		PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3762
3763		/*
3764		 * Debug registers, just like PMC, can only be modified
3765		 * by a kernel call. Moreover, perfmon() access to those
3766		 * registers are centralized in this routine. The hardware
3767		 * does not modify the value of these registers, therefore,
3768		 * if we save them as they are written, we can avoid having
3769		 * to save them on context switch out. This is made possible
3770		 * by the fact that when perfmon uses debug registers, ptrace()
3771		 * won't be able to modify them concurrently.
3772		 */
3773		if (mode == PFM_CODE_RR) {
3774			CTX_USED_IBR(ctx, rnum);
3775
3776			if (can_access_pmu) {
3777				ia64_set_ibr(rnum, dbreg.val);
3778				ia64_dv_serialize_instruction();
3779			}
3780
3781			ctx->ctx_ibrs[rnum] = dbreg.val;
3782
3783			DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3784				rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3785		} else {
3786			CTX_USED_DBR(ctx, rnum);
3787
3788			if (can_access_pmu) {
3789				ia64_set_dbr(rnum, dbreg.val);
3790				ia64_dv_serialize_data();
3791			}
3792			ctx->ctx_dbrs[rnum] = dbreg.val;
3793
3794			DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3795				rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3796		}
3797	}
3798
3799	return 0;
3800
3801abort_mission:
3802	/*
3803	 * in case it was our first attempt, we undo the global modifications
3804	 */
3805	if (first_time) {
3806		LOCK_PFS(flags);
3807		if (ctx->ctx_fl_system) {
3808			pfm_sessions.pfs_sys_use_dbregs--;
3809		}
3810		UNLOCK_PFS(flags);
3811		ctx->ctx_fl_using_dbreg = 0;
3812	}
3813	/*
3814	 * install error return flag
3815	 */
3816	PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3817
3818	return ret;
3819}
3820
3821static int
3822pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3823{
3824	return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3825}
3826
3827static int
3828pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3829{
3830	return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3831}
3832
3833int
3834pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3835{
3836	pfm_context_t *ctx;
3837
3838	if (req == NULL) return -EINVAL;
3839
3840 	ctx = GET_PMU_CTX();
3841
3842	if (ctx == NULL) return -EINVAL;
3843
3844	/*
3845	 * for now limit to current task, which is enough when calling
3846	 * from overflow handler
3847	 */
3848	if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3849
3850	return pfm_write_ibrs(ctx, req, nreq, regs);
3851}
3852EXPORT_SYMBOL(pfm_mod_write_ibrs);
3853
3854int
3855pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3856{
3857	pfm_context_t *ctx;
3858
3859	if (req == NULL) return -EINVAL;
3860
3861 	ctx = GET_PMU_CTX();
3862
3863	if (ctx == NULL) return -EINVAL;
3864
3865	/*
3866	 * for now limit to current task, which is enough when calling
3867	 * from overflow handler
3868	 */
3869	if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3870
3871	return pfm_write_dbrs(ctx, req, nreq, regs);
3872}
3873EXPORT_SYMBOL(pfm_mod_write_dbrs);
3874
3875
3876static int
3877pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3878{
3879	pfarg_features_t *req = (pfarg_features_t *)arg;
3880
3881	req->ft_version = PFM_VERSION;
3882	return 0;
3883}
3884
3885static int
3886pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3887{
3888	struct pt_regs *tregs;
3889	struct task_struct *task = PFM_CTX_TASK(ctx);
3890	int state, is_system;
3891
3892	state     = ctx->ctx_state;
3893	is_system = ctx->ctx_fl_system;
3894
3895	/*
3896	 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3897	 */
3898	if (state == PFM_CTX_UNLOADED) return -EINVAL;
3899
3900	/*
3901 	 * In system wide and when the context is loaded, access can only happen
3902 	 * when the caller is running on the CPU being monitored by the session.
3903 	 * It does not have to be the owner (ctx_task) of the context per se.
3904 	 */
3905	if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3906		DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3907		return -EBUSY;
3908	}
3909	DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
3910		task_pid_nr(PFM_CTX_TASK(ctx)),
3911		state,
3912		is_system));
3913	/*
3914	 * in system mode, we need to update the PMU directly
3915	 * and the user level state of the caller, which may not
3916	 * necessarily be the creator of the context.
3917	 */
3918	if (is_system) {
3919		/*
3920		 * Update local PMU first
3921		 *
3922		 * disable dcr pp
3923		 */
3924		ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
3925		ia64_srlz_i();
3926
3927		/*
3928		 * update local cpuinfo
3929		 */
3930		PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
3931
3932		/*
3933		 * stop monitoring, does srlz.i
3934		 */
3935		pfm_clear_psr_pp();
3936
3937		/*
3938		 * stop monitoring in the caller
3939		 */
3940		ia64_psr(regs)->pp = 0;
3941
3942		return 0;
3943	}
3944	/*
3945	 * per-task mode
3946	 */
3947
3948	if (task == current) {
3949		/* stop monitoring  at kernel level */
3950		pfm_clear_psr_up();
3951
3952		/*
3953	 	 * stop monitoring at the user level
3954	 	 */
3955		ia64_psr(regs)->up = 0;
3956	} else {
3957		tregs = task_pt_regs(task);
3958
3959		/*
3960	 	 * stop monitoring at the user level
3961	 	 */
3962		ia64_psr(tregs)->up = 0;
3963
3964		/*
3965		 * monitoring disabled in kernel at next reschedule
3966		 */
3967		ctx->ctx_saved_psr_up = 0;
3968		DPRINT(("task=[%d]\n", task_pid_nr(task)));
3969	}
3970	return 0;
3971}
3972
3973
3974static int
3975pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3976{
3977	struct pt_regs *tregs;
3978	int state, is_system;
3979
3980	state     = ctx->ctx_state;
3981	is_system = ctx->ctx_fl_system;
3982
3983	if (state != PFM_CTX_LOADED) return -EINVAL;
3984
3985	/*
3986 	 * In system wide and when the context is loaded, access can only happen
3987 	 * when the caller is running on the CPU being monitored by the session.
3988 	 * It does not have to be the owner (ctx_task) of the context per se.
3989 	 */
3990	if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3991		DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3992		return -EBUSY;
3993	}
3994
3995	/*
3996	 * in system mode, we need to update the PMU directly
3997	 * and the user level state of the caller, which may not
3998	 * necessarily be the creator of the context.
3999	 */
4000	if (is_system) {
4001
4002		/*
4003		 * set user level psr.pp for the caller
4004		 */
4005		ia64_psr(regs)->pp = 1;
4006
4007		/*
4008		 * now update the local PMU and cpuinfo
4009		 */
4010		PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4011
4012		/*
4013		 * start monitoring at kernel level
4014		 */
4015		pfm_set_psr_pp();
4016
4017		/* enable dcr pp */
4018		ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4019		ia64_srlz_i();
4020
4021		return 0;
4022	}
4023
4024	/*
4025	 * per-process mode
4026	 */
4027
4028	if (ctx->ctx_task == current) {
4029
4030		/* start monitoring at kernel level */
4031		pfm_set_psr_up();
4032
4033		/*
4034		 * activate monitoring at user level
4035		 */
4036		ia64_psr(regs)->up = 1;
4037
4038	} else {
4039		tregs = task_pt_regs(ctx->ctx_task);
4040
4041		/*
4042		 * start monitoring at the kernel level the next
4043		 * time the task is scheduled
4044		 */
4045		ctx->ctx_saved_psr_up = IA64_PSR_UP;
4046
4047		/*
4048		 * activate monitoring at user level
4049		 */
4050		ia64_psr(tregs)->up = 1;
4051	}
4052	return 0;
4053}
4054
4055static int
4056pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4057{
4058	pfarg_reg_t *req = (pfarg_reg_t *)arg;
4059	unsigned int cnum;
4060	int i;
4061	int ret = -EINVAL;
4062
4063	for (i = 0; i < count; i++, req++) {
4064
4065		cnum = req->reg_num;
4066
4067		if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4068
4069		req->reg_value = PMC_DFL_VAL(cnum);
4070
4071		PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4072
4073		DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4074	}
4075	return 0;
4076
4077abort_mission:
4078	PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4079	return ret;
4080}
4081
4082static int
4083pfm_check_task_exist(pfm_context_t *ctx)
4084{
4085	struct task_struct *g, *t;
4086	int ret = -ESRCH;
4087
4088	read_lock(&tasklist_lock);
4089
4090	do_each_thread (g, t) {
4091		if (t->thread.pfm_context == ctx) {
4092			ret = 0;
4093			goto out;
4094		}
4095	} while_each_thread (g, t);
4096out:
4097	read_unlock(&tasklist_lock);
4098
4099	DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4100
4101	return ret;
4102}
4103
4104static int
4105pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4106{
4107	struct task_struct *task;
4108	struct thread_struct *thread;
4109	struct pfm_context_t *old;
4110	unsigned long flags;
4111#ifndef CONFIG_SMP
4112	struct task_struct *owner_task = NULL;
4113#endif
4114	pfarg_load_t *req = (pfarg_load_t *)arg;
4115	unsigned long *pmcs_source, *pmds_source;
4116	int the_cpu;
4117	int ret = 0;
4118	int state, is_system, set_dbregs = 0;
4119
4120	state     = ctx->ctx_state;
4121	is_system = ctx->ctx_fl_system;
4122	/*
4123	 * can only load from unloaded or terminated state
4124	 */
4125	if (state != PFM_CTX_UNLOADED) {
4126		DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4127			req->load_pid,
4128			ctx->ctx_state));
4129		return -EBUSY;
4130	}
4131
4132	DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4133
4134	if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4135		DPRINT(("cannot use blocking mode on self\n"));
4136		return -EINVAL;
4137	}
4138
4139	ret = pfm_get_task(ctx, req->load_pid, &task);
4140	if (ret) {
4141		DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4142		return ret;
4143	}
4144
4145	ret = -EINVAL;
4146
4147	/*
4148	 * system wide is self monitoring only
4149	 */
4150	if (is_system && task != current) {
4151		DPRINT(("system wide is self monitoring only load_pid=%d\n",
4152			req->load_pid));
4153		goto error;
4154	}
4155
4156	thread = &task->thread;
4157
4158	ret = 0;
4159	/*
4160	 * cannot load a context which is using range restrictions,
4161	 * into a task that is being debugged.
4162	 */
4163	if (ctx->ctx_fl_using_dbreg) {
4164		if (thread->flags & IA64_THREAD_DBG_VALID) {
4165			ret = -EBUSY;
4166			DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4167			goto error;
4168		}
4169		LOCK_PFS(flags);
4170
4171		if (is_system) {
4172			if (pfm_sessions.pfs_ptrace_use_dbregs) {
4173				DPRINT(("cannot load [%d] dbregs in use\n",
4174							task_pid_nr(task)));
4175				ret = -EBUSY;
4176			} else {
4177				pfm_sessions.pfs_sys_use_dbregs++;
4178				DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task_pid_nr(task), pfm_sessions.pfs_sys_use_dbregs));
4179				set_dbregs = 1;
4180			}
4181		}
4182
4183		UNLOCK_PFS(flags);
4184
4185		if (ret) goto error;
4186	}
4187
4188	/*
4189	 * SMP system-wide monitoring implies self-monitoring.
4190	 *
4191	 * The programming model expects the task to
4192	 * be pinned on a CPU throughout the session.
4193	 * Here we take note of the current CPU at the
4194	 * time the context is loaded. No call from
4195	 * another CPU will be allowed.
4196	 *
4197	 * The pinning via shed_setaffinity()
4198	 * must be done by the calling task prior
4199	 * to this call.
4200	 *
4201	 * systemwide: keep track of CPU this session is supposed to run on
4202	 */
4203	the_cpu = ctx->ctx_cpu = smp_processor_id();
4204
4205	ret = -EBUSY;
4206	/*
4207	 * now reserve the session
4208	 */
4209	ret = pfm_reserve_session(current, is_system, the_cpu);
4210	if (ret) goto error;
4211
4212	/*
4213	 * task is necessarily stopped at this point.
4214	 *
4215	 * If the previous context was zombie, then it got removed in
4216	 * pfm_save_regs(). Therefore we should not see it here.
4217	 * If we see a context, then this is an active context
4218	 *
4219	 * XXX: needs to be atomic
4220	 */
4221	DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4222		thread->pfm_context, ctx));
4223
4224	ret = -EBUSY;
4225	old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4226	if (old != NULL) {
4227		DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4228		goto error_unres;
4229	}
4230
4231	pfm_reset_msgq(ctx);
4232
4233	ctx->ctx_state = PFM_CTX_LOADED;
4234
4235	/*
4236	 * link context to task
4237	 */
4238	ctx->ctx_task = task;
4239
4240	if (is_system) {
4241		/*
4242		 * we load as stopped
4243		 */
4244		PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4245		PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4246
4247		if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4248	} else {
4249		thread->flags |= IA64_THREAD_PM_VALID;
4250	}
4251
4252	/*
4253	 * propagate into thread-state
4254	 */
4255	pfm_copy_pmds(task, ctx);
4256	pfm_copy_pmcs(task, ctx);
4257
4258	pmcs_source = ctx->th_pmcs;
4259	pmds_source = ctx->th_pmds;
4260
4261	/*
4262	 * always the case for system-wide
4263	 */
4264	if (task == current) {
4265
4266		if (is_system == 0) {
4267
4268			/* allow user level control */
4269			ia64_psr(regs)->sp = 0;
4270			DPRINT(("clearing psr.sp for [%d]\n", task_pid_nr(task)));
4271
4272			SET_LAST_CPU(ctx, smp_processor_id());
4273			INC_ACTIVATION();
4274			SET_ACTIVATION(ctx);
4275#ifndef CONFIG_SMP
4276			/*
4277			 * push the other task out, if any
4278			 */
4279			owner_task = GET_PMU_OWNER();
4280			if (owner_task) pfm_lazy_save_regs(owner_task);
4281#endif
4282		}
4283		/*
4284		 * load all PMD from ctx to PMU (as opposed to thread state)
4285		 * restore all PMC from ctx to PMU
4286		 */
4287		pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4288		pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4289
4290		ctx->ctx_reload_pmcs[0] = 0UL;
4291		ctx->ctx_reload_pmds[0] = 0UL;
4292
4293		/*
4294		 * guaranteed safe by earlier check against DBG_VALID
4295		 */
4296		if (ctx->ctx_fl_using_dbreg) {
4297			pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4298			pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4299		}
4300		/*
4301		 * set new ownership
4302		 */
4303		SET_PMU_OWNER(task, ctx);
4304
4305		DPRINT(("context loaded on PMU for [%d]\n", task_pid_nr(task)));
4306	} else {
4307		/*
4308		 * when not current, task MUST be stopped, so this is safe
4309		 */
4310		regs = task_pt_regs(task);
4311
4312		/* force a full reload */
4313		ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4314		SET_LAST_CPU(ctx, -1);
4315
4316		/* initial saved psr (stopped) */
4317		ctx->ctx_saved_psr_up = 0UL;
4318		ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4319	}
4320
4321	ret = 0;
4322
4323error_unres:
4324	if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4325error:
4326	/*
4327	 * we must undo the dbregs setting (for system-wide)
4328	 */
4329	if (ret && set_dbregs) {
4330		LOCK_PFS(flags);
4331		pfm_sessions.pfs_sys_use_dbregs--;
4332		UNLOCK_PFS(flags);
4333	}
4334	/*
4335	 * release task, there is now a link with the context
4336	 */
4337	if (is_system == 0 && task != current) {
4338		pfm_put_task(task);
4339
4340		if (ret == 0) {
4341			ret = pfm_check_task_exist(ctx);
4342			if (ret) {
4343				ctx->ctx_state = PFM_CTX_UNLOADED;
4344				ctx->ctx_task  = NULL;
4345			}
4346		}
4347	}
4348	return ret;
4349}
4350
4351/*
4352 * in this function, we do not need to increase the use count
4353 * for the task via get_task_struct(), because we hold the
4354 * context lock. If the task were to disappear while having
4355 * a context attached, it would go through pfm_exit_thread()
4356 * which also grabs the context lock  and would therefore be blocked
4357 * until we are here.
4358 */
4359static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4360
4361static int
4362pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4363{
4364	struct task_struct *task = PFM_CTX_TASK(ctx);
4365	struct pt_regs *tregs;
4366	int prev_state, is_system;
4367	int ret;
4368
4369	DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task_pid_nr(task) : -1));
4370
4371	prev_state = ctx->ctx_state;
4372	is_system  = ctx->ctx_fl_system;
4373
4374	/*
4375	 * unload only when necessary
4376	 */
4377	if (prev_state == PFM_CTX_UNLOADED) {
4378		DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4379		return 0;
4380	}
4381
4382	/*
4383	 * clear psr and dcr bits
4384	 */
4385	ret = pfm_stop(ctx, NULL, 0, regs);
4386	if (ret) return ret;
4387
4388	ctx->ctx_state = PFM_CTX_UNLOADED;
4389
4390	/*
4391	 * in system mode, we need to update the PMU directly
4392	 * and the user level state of the caller, which may not
4393	 * necessarily be the creator of the context.
4394	 */
4395	if (is_system) {
4396
4397		/*
4398		 * Update cpuinfo
4399		 *
4400		 * local PMU is taken care of in pfm_stop()
4401		 */
4402		PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4403		PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4404
4405		/*
4406		 * save PMDs in context
4407		 * release ownership
4408		 */
4409		pfm_flush_pmds(current, ctx);
4410
4411		/*
4412		 * at this point we are done with the PMU
4413		 * so we can unreserve the resource.
4414		 */
4415		if (prev_state != PFM_CTX_ZOMBIE) 
4416			pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4417
4418		/*
4419		 * disconnect context from task
4420		 */
4421		task->thread.pfm_context = NULL;
4422		/*
4423		 * disconnect task from context
4424		 */
4425		ctx->ctx_task = NULL;
4426
4427		/*
4428		 * There is nothing more to cleanup here.
4429		 */
4430		return 0;
4431	}
4432
4433	/*
4434	 * per-task mode
4435	 */
4436	tregs = task == current ? regs : task_pt_regs(task);
4437
4438	if (task == current) {
4439		/*
4440		 * cancel user level control
4441		 */
4442		ia64_psr(regs)->sp = 1;
4443
4444		DPRINT(("setting psr.sp for [%d]\n", task_pid_nr(task)));
4445	}
4446	/*
4447	 * save PMDs to context
4448	 * release ownership
4449	 */
4450	pfm_flush_pmds(task, ctx);
4451
4452	/*
4453	 * at this point we are done with the PMU
4454	 * so we can unreserve the resource.
4455	 *
4456	 * when state was ZOMBIE, we have already unreserved.
4457	 */
4458	if (prev_state != PFM_CTX_ZOMBIE) 
4459		pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4460
4461	/*
4462	 * reset activation counter and psr
4463	 */
4464	ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4465	SET_LAST_CPU(ctx, -1);
4466
4467	/*
4468	 * PMU state will not be restored
4469	 */
4470	task->thread.flags &= ~IA64_THREAD_PM_VALID;
4471
4472	/*
4473	 * break links between context and task
4474	 */
4475	task->thread.pfm_context  = NULL;
4476	ctx->ctx_task             = NULL;
4477
4478	PFM_SET_WORK_PENDING(task, 0);
4479
4480	ctx->ctx_fl_trap_reason  = PFM_TRAP_REASON_NONE;
4481	ctx->ctx_fl_can_restart  = 0;
4482	ctx->ctx_fl_going_zombie = 0;
4483
4484	DPRINT(("disconnected [%d] from context\n", task_pid_nr(task)));
4485
4486	return 0;
4487}
4488
4489
4490/*
4491 * called only from exit_thread()
4492 * we come here only if the task has a context attached (loaded or masked)
4493 */
4494void
4495pfm_exit_thread(struct task_struct *task)
4496{
4497	pfm_context_t *ctx;
4498	unsigned long flags;
4499	struct pt_regs *regs = task_pt_regs(task);
4500	int ret, state;
4501	int free_ok = 0;
4502
4503	ctx = PFM_GET_CTX(task);
4504
4505	PROTECT_CTX(ctx, flags);
4506
4507	DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task_pid_nr(task)));
4508
4509	state = ctx->ctx_state;
4510	switch(state) {
4511		case PFM_CTX_UNLOADED:
4512			/*
4513	 		 * only comes to this function if pfm_context is not NULL, i.e., cannot
4514			 * be in unloaded state
4515	 		 */
4516			printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task_pid_nr(task));
4517			break;
4518		case PFM_CTX_LOADED:
4519		case PFM_CTX_MASKED:
4520			ret = pfm_context_unload(ctx, NULL, 0, regs);
4521			if (ret) {
4522				printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4523			}
4524			DPRINT(("ctx unloaded for current state was %d\n", state));
4525
4526			pfm_end_notify_user(ctx);
4527			break;
4528		case PFM_CTX_ZOMBIE:
4529			ret = pfm_context_unload(ctx, NULL, 0, regs);
4530			if (ret) {
4531				printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4532			}
4533			free_ok = 1;
4534			break;
4535		default:
4536			printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task_pid_nr(task), state);
4537			break;
4538	}
4539	UNPROTECT_CTX(ctx, flags);
4540
4541	{ u64 psr = pfm_get_psr();
4542	  BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4543	  BUG_ON(GET_PMU_OWNER());
4544	  BUG_ON(ia64_psr(regs)->up);
4545	  BUG_ON(ia64_psr(regs)->pp);
4546	}
4547
4548	/*
4549	 * All memory free operations (especially for vmalloc'ed memory)
4550	 * MUST be done with interrupts ENABLED.
4551	 */
4552	if (free_ok) pfm_context_free(ctx);
4553}
4554
4555/*
4556 * functions MUST be listed in the increasing order of their index (see permfon.h)
4557 */
4558#define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4559#define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4560#define PFM_CMD_PCLRWS	(PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4561#define PFM_CMD_PCLRW	(PFM_CMD_FD|PFM_CMD_ARG_RW)
4562#define PFM_CMD_NONE	{ NULL, "no-cmd", 0, 0, 0, NULL}
4563
4564static pfm_cmd_desc_t pfm_cmd_tab[]={
4565/* 0  */PFM_CMD_NONE,
4566/* 1  */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4567/* 2  */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4568/* 3  */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4569/* 4  */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4570/* 5  */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4571/* 6  */PFM_CMD_NONE,
4572/* 7  */PFM_CMD_NONE,
4573/* 8  */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4574/* 9  */PFM_CMD_NONE,
4575/* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4576/* 11 */PFM_CMD_NONE,
4577/* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4578/* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4579/* 14 */PFM_CMD_NONE,
4580/* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4581/* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4582/* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4583/* 18 */PFM_CMD_NONE,
4584/* 19 */PFM_CMD_NONE,
4585/* 20 */PFM_CMD_NONE,
4586/* 21 */PFM_CMD_NONE,
4587/* 22 */PFM_CMD_NONE,
4588/* 23 */PFM_CMD_NONE,
4589/* 24 */PFM_CMD_NONE,
4590/* 25 */PFM_CMD_NONE,
4591/* 26 */PFM_CMD_NONE,
4592/* 27 */PFM_CMD_NONE,
4593/* 28 */PFM_CMD_NONE,
4594/* 29 */PFM_CMD_NONE,
4595/* 30 */PFM_CMD_NONE,
4596/* 31 */PFM_CMD_NONE,
4597/* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4598/* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4599};
4600#define PFM_CMD_COUNT	(sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4601
4602static int
4603pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4604{
4605	struct task_struct *task;
4606	int state, old_state;
4607
4608recheck:
4609	state = ctx->ctx_state;
4610	task  = ctx->ctx_task;
4611
4612	if (task == NULL) {
4613		DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4614		return 0;
4615	}
4616
4617	DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4618		ctx->ctx_fd,
4619		state,
4620		task_pid_nr(task),
4621		task->state, PFM_CMD_STOPPED(cmd)));
4622
4623	/*
4624	 * self-monitoring always ok.
4625	 *
4626	 * for system-wide the caller can either be the creator of the
4627	 * context (to one to which the context is attached to) OR
4628	 * a task running on the same CPU as the session.
4629	 */
4630	if (task == current || ctx->ctx_fl_system) return 0;
4631
4632	/*
4633	 * we are monitoring another thread
4634	 */
4635	switch(state) {
4636		case PFM_CTX_UNLOADED:
4637			/*
4638			 * if context is UNLOADED we are safe to go
4639			 */
4640			return 0;
4641		case PFM_CTX_ZOMBIE:
4642			/*
4643			 * no command can operate on a zombie context
4644			 */
4645			DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4646			return -EINVAL;
4647		case PFM_CTX_MASKED:
4648			/*
4649			 * PMU state has been saved to software even though
4650			 * the thread may still be running.
4651			 */
4652			if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4653	}
4654
4655	/*
4656	 * context is LOADED or MASKED. Some commands may need to have 
4657	 * the task stopped.
4658	 *
4659	 * We could lift this restriction for UP but it would mean that
4660	 * the user has no guarantee the task would not run between
4661	 * two successive calls to perfmonctl(). That's probably OK.
4662	 * If this user wants to ensure the task does not run, then
4663	 * the task must be stopped.
4664	 */
4665	if (PFM_CMD_STOPPED(cmd)) {
4666		if (!task_is_stopped_or_traced(task)) {
4667			DPRINT(("[%d] task not in stopped state\n", task_pid_nr(task)));
4668			return -EBUSY;
4669		}
4670		/*
4671		 * task is now stopped, wait for ctxsw out
4672		 *
4673		 * This is an interesting point in the code.
4674		 * We need to unprotect the context because
4675		 * the pfm_save_regs() routines needs to grab
4676		 * the same lock. There are danger in doing
4677		 * this because it leaves a window open for
4678		 * another task to get access to the context
4679		 * and possibly change its state. The one thing
4680		 * that is not possible is for the context to disappear
4681		 * because we are protected by the VFS layer, i.e.,
4682		 * get_fd()/put_fd().
4683		 */
4684		old_state = state;
4685
4686		UNPROTECT_CTX(ctx, flags);
4687
4688		wait_task_inactive(task, 0);
4689
4690		PROTECT_CTX(ctx, flags);
4691
4692		/*
4693		 * we must recheck to verify if state has changed
4694		 */
4695		if (ctx->ctx_state != old_state) {
4696			DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4697			goto recheck;
4698		}
4699	}
4700	return 0;
4701}
4702
4703/*
4704 * system-call entry point (must return long)
4705 */
4706asmlinkage long
4707sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4708{
4709	struct fd f = {NULL, 0};
4710	pfm_context_t *ctx = NULL;
4711	unsigned long flags = 0UL;
4712	void *args_k = NULL;
4713	long ret; /* will expand int return types */
4714	size_t base_sz, sz, xtra_sz = 0;
4715	int narg, completed_args = 0, call_made = 0, cmd_flags;
4716	int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4717	int (*getsize)(void *arg, size_t *sz);
4718#define PFM_MAX_ARGSIZE	4096
4719
4720	/*
4721	 * reject any call if perfmon was disabled at initialization
4722	 */
4723	if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4724
4725	if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4726		DPRINT(("invalid cmd=%d\n", cmd));
4727		return -EINVAL;
4728	}
4729
4730	func      = pfm_cmd_tab[cmd].cmd_func;
4731	narg      = pfm_cmd_tab[cmd].cmd_narg;
4732	base_sz   = pfm_cmd_tab[cmd].cmd_argsize;
4733	getsize   = pfm_cmd_tab[cmd].cmd_getsize;
4734	cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4735
4736	if (unlikely(func == NULL)) {
4737		DPRINT(("invalid cmd=%d\n", cmd));
4738		return -EINVAL;
4739	}
4740
4741	DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4742		PFM_CMD_NAME(cmd),
4743		cmd,
4744		narg,
4745		base_sz,
4746		count));
4747
4748	/*
4749	 * check if number of arguments matches what the command expects
4750	 */
4751	if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4752		return -EINVAL;
4753
4754restart_args:
4755	sz = xtra_sz + base_sz*count;
4756	/*
4757	 * limit abuse to min page size
4758	 */
4759	if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4760		printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", task_pid_nr(current), sz);
4761		return -E2BIG;
4762	}
4763
4764	/*
4765	 * allocate default-sized argument buffer
4766	 */
4767	if (likely(count && args_k == NULL)) {
4768		args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4769		if (args_k == NULL) return -ENOMEM;
4770	}
4771
4772	ret = -EFAULT;
4773
4774	/*
4775	 * copy arguments
4776	 *
4777	 * assume sz = 0 for command without parameters
4778	 */
4779	if (sz && copy_from_user(args_k, arg, sz)) {
4780		DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4781		goto error_args;
4782	}
4783
4784	/*
4785	 * check if command supports extra parameters
4786	 */
4787	if (completed_args == 0 && getsize) {
4788		/*
4789		 * get extra parameters size (based on main argument)
4790		 */
4791		ret = (*getsize)(args_k, &xtra_sz);
4792		if (ret) goto error_args;
4793
4794		completed_args = 1;
4795
4796		DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4797
4798		/* retry if necessary */
4799		if (likely(xtra_sz)) goto restart_args;
4800	}
4801
4802	if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4803
4804	ret = -EBADF;
4805
4806	f = fdget(fd);
4807	if (unlikely(f.file == NULL)) {
4808		DPRINT(("invalid fd %d\n", fd));
4809		goto error_args;
4810	}
4811	if (unlikely(PFM_IS_FILE(f.file) == 0)) {
4812		DPRINT(("fd %d not related to perfmon\n", fd));
4813		goto error_args;
4814	}
4815
4816	ctx = f.file->private_data;
4817	if (unlikely(ctx == NULL)) {
4818		DPRINT(("no context for fd %d\n", fd));
4819		goto error_args;
4820	}
4821	prefetch(&ctx->ctx_state);
4822
4823	PROTECT_CTX(ctx, flags);
4824
4825	/*
4826	 * check task is stopped
4827	 */
4828	ret = pfm_check_task_state(ctx, cmd, flags);
4829	if (unlikely(ret)) goto abort_locked;
4830
4831skip_fd:
4832	ret = (*func)(ctx, args_k, count, task_pt_regs(current));
4833
4834	call_made = 1;
4835
4836abort_locked:
4837	if (likely(ctx)) {
4838		DPRINT(("context unlocked\n"));
4839		UNPROTECT_CTX(ctx, flags);
4840	}
4841
4842	/* copy argument back to user, if needed */
4843	if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4844
4845error_args:
4846	if (f.file)
4847		fdput(f);
4848
4849	kfree(args_k);
4850
4851	DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4852
4853	return ret;
4854}
4855
4856static void
4857pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4858{
4859	pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4860	pfm_ovfl_ctrl_t rst_ctrl;
4861	int state;
4862	int ret = 0;
4863
4864	state = ctx->ctx_state;
4865	/*
4866	 * Unlock sampling buffer and reset index atomically
4867	 * XXX: not really needed when blocking
4868	 */
4869	if (CTX_HAS_SMPL(ctx)) {
4870
4871		rst_ctrl.bits.mask_monitoring = 0;
4872		rst_ctrl.bits.reset_ovfl_pmds = 0;
4873
4874		if (state == PFM_CTX_LOADED)
4875			ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4876		else
4877			ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4878	} else {
4879		rst_ctrl.bits.mask_monitoring = 0;
4880		rst_ctrl.bits.reset_ovfl_pmds = 1;
4881	}
4882
4883	if (ret == 0) {
4884		if (rst_ctrl.bits.reset_ovfl_pmds) {
4885			pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4886		}
4887		if (rst_ctrl.bits.mask_monitoring == 0) {
4888			DPRINT(("resuming monitoring\n"));
4889			if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4890		} else {
4891			DPRINT(("stopping monitoring\n"));
4892			//pfm_stop_monitoring(current, regs);
4893		}
4894		ctx->ctx_state = PFM_CTX_LOADED;
4895	}
4896}
4897
4898/*
4899 * context MUST BE LOCKED when calling
4900 * can only be called for current
4901 */
4902static void
4903pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4904{
4905	int ret;
4906
4907	DPRINT(("entering for [%d]\n", task_pid_nr(current)));
4908
4909	ret = pfm_context_unload(ctx, NULL, 0, regs);
4910	if (ret) {
4911		printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", task_pid_nr(current), ret);
4912	}
4913
4914	/*
4915	 * and wakeup controlling task, indicating we are now disconnected
4916	 */
4917	wake_up_interruptible(&ctx->ctx_zombieq);
4918
4919	/*
4920	 * given that context is still locked, the controlling
4921	 * task will only get access when we return from
4922	 * pfm_handle_work().
4923	 */
4924}
4925
4926static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
4927
4928 /*
4929  * pfm_handle_work() can be called with interrupts enabled
4930  * (TIF_NEED_RESCHED) or disabled. The down_interruptible
4931  * call may sleep, therefore we must re-enable interrupts
4932  * to avoid deadlocks. It is safe to do so because this function
4933  * is called ONLY when returning to user level (pUStk=1), in which case
4934  * there is no risk of kernel stack overflow due to deep
4935  * interrupt nesting.
4936  */
4937void
4938pfm_handle_work(void)
4939{
4940	pfm_context_t *ctx;
4941	struct pt_regs *regs;
4942	unsigned long flags, dummy_flags;
4943	unsigned long ovfl_regs;
4944	unsigned int reason;
4945	int ret;
4946
4947	ctx = PFM_GET_CTX(current);
4948	if (ctx == NULL) {
4949		printk(KERN_ERR "perfmon: [%d] has no PFM context\n",
4950			task_pid_nr(current));
4951		return;
4952	}
4953
4954	PROTECT_CTX(ctx, flags);
4955
4956	PFM_SET_WORK_PENDING(current, 0);
4957
4958	regs = task_pt_regs(current);
4959
4960	/*
4961	 * extract reason for being here and clear
4962	 */
4963	reason = ctx->ctx_fl_trap_reason;
4964	ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
4965	ovfl_regs = ctx->ctx_ovfl_regs[0];
4966
4967	DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
4968
4969	/*
4970	 * must be done before we check for simple-reset mode
4971	 */
4972	if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE)
4973		goto do_zombie;
4974
4975	//if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
4976	if (reason == PFM_TRAP_REASON_RESET)
4977		goto skip_blocking;
4978
4979	/*
4980	 * restore interrupt mask to what it was on entry.
4981	 * Could be enabled/diasbled.
4982	 */
4983	UNPROTECT_CTX(ctx, flags);
4984
4985	/*
4986	 * force interrupt enable because of down_interruptible()
4987	 */
4988	local_irq_enable();
4989
4990	DPRINT(("before block sleeping\n"));
4991
4992	/*
4993	 * may go through without blocking on SMP systems
4994	 * if restart has been received already by the time we call down()
4995	 */
4996	ret = wait_for_completion_interruptible(&ctx->ctx_restart_done);
4997
4998	DPRINT(("after block sleeping ret=%d\n", ret));
4999
5000	/*
5001	 * lock context and mask interrupts again
5002	 * We save flags into a dummy because we may have
5003	 * altered interrupts mask compared to entry in this
5004	 * function.
5005	 */
5006	PROTECT_CTX(ctx, dummy_flags);
5007
5008	/*
5009	 * we need to read the ovfl_regs only after wake-up
5010	 * because we may have had pfm_write_pmds() in between
5011	 * and that can changed PMD values and therefore 
5012	 * ovfl_regs is reset for these new PMD values.
5013	 */
5014	ovfl_regs = ctx->ctx_ovfl_regs[0];
5015
5016	if (ctx->ctx_fl_going_zombie) {
5017do_zombie:
5018		DPRINT(("context is zombie, bailing out\n"));
5019		pfm_context_force_terminate(ctx, regs);
5020		goto nothing_to_do;
5021	}
5022	/*
5023	 * in case of interruption of down() we don't restart anything
5024	 */
5025	if (ret < 0)
5026		goto nothing_to_do;
5027
5028skip_blocking:
5029	pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5030	ctx->ctx_ovfl_regs[0] = 0UL;
5031
5032nothing_to_do:
5033	/*
5034	 * restore flags as they were upon entry
5035	 */
5036	UNPROTECT_CTX(ctx, flags);
5037}
5038
5039static int
5040pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5041{
5042	if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5043		DPRINT(("ignoring overflow notification, owner is zombie\n"));
5044		return 0;
5045	}
5046
5047	DPRINT(("waking up somebody\n"));
5048
5049	if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5050
5051	/*
5052	 * safe, we are not in intr handler, nor in ctxsw when
5053	 * we come here
5054	 */
5055	kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5056
5057	return 0;
5058}
5059
5060static int
5061pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5062{
5063	pfm_msg_t *msg = NULL;
5064
5065	if (ctx->ctx_fl_no_msg == 0) {
5066		msg = pfm_get_new_msg(ctx);
5067		if (msg == NULL) {
5068			printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5069			return -1;
5070		}
5071
5072		msg->pfm_ovfl_msg.msg_type         = PFM_MSG_OVFL;
5073		msg->pfm_ovfl_msg.msg_ctx_fd       = ctx->ctx_fd;
5074		msg->pfm_ovfl_msg.msg_active_set   = 0;
5075		msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5076		msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5077		msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5078		msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5079		msg->pfm_ovfl_msg.msg_tstamp       = 0UL;
5080	}
5081
5082	DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5083		msg,
5084		ctx->ctx_fl_no_msg,
5085		ctx->ctx_fd,
5086		ovfl_pmds));
5087
5088	return pfm_notify_user(ctx, msg);
5089}
5090
5091static int
5092pfm_end_notify_user(pfm_context_t *ctx)
5093{
5094	pfm_msg_t *msg;
5095
5096	msg = pfm_get_new_msg(ctx);
5097	if (msg == NULL) {
5098		printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5099		return -1;
5100	}
5101	/* no leak */
5102	memset(msg, 0, sizeof(*msg));
5103
5104	msg->pfm_end_msg.msg_type    = PFM_MSG_END;
5105	msg->pfm_end_msg.msg_ctx_fd  = ctx->ctx_fd;
5106	msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5107
5108	DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5109		msg,
5110		ctx->ctx_fl_no_msg,
5111		ctx->ctx_fd));
5112
5113	return pfm_notify_user(ctx, msg);
5114}
5115
5116/*
5117 * main overflow processing routine.
5118 * it can be called from the interrupt path or explicitly during the context switch code
5119 */
5120static void pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx,
5121				unsigned long pmc0, struct pt_regs *regs)
5122{
5123	pfm_ovfl_arg_t *ovfl_arg;
5124	unsigned long mask;
5125	unsigned long old_val, ovfl_val, new_val;
5126	unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5127	unsigned long tstamp;
5128	pfm_ovfl_ctrl_t	ovfl_ctrl;
5129	unsigned int i, has_smpl;
5130	int must_notify = 0;
5131
5132	if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5133
5134	/*
5135	 * sanity test. Should never happen
5136	 */
5137	if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5138
5139	tstamp   = ia64_get_itc();
5140	mask     = pmc0 >> PMU_FIRST_COUNTER;
5141	ovfl_val = pmu_conf->ovfl_val;
5142	has_smpl = CTX_HAS_SMPL(ctx);
5143
5144	DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5145		     "used_pmds=0x%lx\n",
5146			pmc0,
5147			task ? task_pid_nr(task): -1,
5148			(regs ? regs->cr_iip : 0),
5149			CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5150			ctx->ctx_used_pmds[0]));
5151
5152
5153	/*
5154	 * first we update the virtual counters
5155	 * assume there was a prior ia64_srlz_d() issued
5156	 */
5157	for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5158
5159		/* skip pmd which did not overflow */
5160		if ((mask & 0x1) == 0) continue;
5161
5162		/*
5163		 * Note that the pmd is not necessarily 0 at this point as qualified events
5164		 * may have happened before the PMU was frozen. The residual count is not
5165		 * taken into consideration here but will be with any read of the pmd via
5166		 * pfm_read_pmds().
5167		 */
5168		old_val              = new_val = ctx->ctx_pmds[i].val;
5169		new_val             += 1 + ovfl_val;
5170		ctx->ctx_pmds[i].val = new_val;
5171
5172		/*
5173		 * check for overflow condition
5174		 */
5175		if (likely(old_val > new_val)) {
5176			ovfl_pmds |= 1UL << i;
5177			if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5178		}
5179
5180		DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5181			i,
5182			new_val,
5183			old_val,
5184			ia64_get_pmd(i) & ovfl_val,
5185			ovfl_pmds,
5186			ovfl_notify));
5187	}
5188
5189	/*
5190	 * there was no 64-bit overflow, nothing else to do
5191	 */
5192	if (ovfl_pmds == 0UL) return;
5193
5194	/* 
5195	 * reset all control bits
5196	 */
5197	ovfl_ctrl.val = 0;
5198	reset_pmds    = 0UL;
5199
5200	/*
5201	 * if a sampling format module exists, then we "cache" the overflow by 
5202	 * calling the module's handler() routine.
5203	 */
5204	if (has_smpl) {
5205		unsigned long start_cycles, end_cycles;
5206		unsigned long pmd_mask;
5207		int j, k, ret = 0;
5208		int this_cpu = smp_processor_id();
5209
5210		pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5211		ovfl_arg = &ctx->ctx_ovfl_arg;
5212
5213		prefetch(ctx->ctx_smpl_hdr);
5214
5215		for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5216
5217			mask = 1UL << i;
5218
5219			if ((pmd_mask & 0x1) == 0) continue;
5220
5221			ovfl_arg->ovfl_pmd      = (unsigned char )i;
5222			ovfl_arg->ovfl_notify   = ovfl_notify & mask ? 1 : 0;
5223			ovfl_arg->active_set    = 0;
5224			ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5225			ovfl_arg->smpl_pmds[0]  = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5226
5227			ovfl_arg->pmd_value      = ctx->ctx_pmds[i].val;
5228			ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5229			ovfl_arg->pmd_eventid    = ctx->ctx_pmds[i].eventid;
5230
5231			/*
5232		 	 * copy values of pmds of interest. Sampling format may copy them
5233		 	 * into sampling buffer.
5234		 	 */
5235			if (smpl_pmds) {
5236				for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5237					if ((smpl_pmds & 0x1) == 0) continue;
5238					ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ?  pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5239					DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5240				}
5241			}
5242
5243			pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5244
5245			start_cycles = ia64_get_itc();
5246
5247			/*
5248		 	 * call custom buffer format record (handler) routine
5249		 	 */
5250			ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5251
5252			end_cycles = ia64_get_itc();
5253
5254			/*
5255			 * For those controls, we take the union because they have
5256			 * an all or nothing behavior.
5257			 */
5258			ovfl_ctrl.bits.notify_user     |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5259			ovfl_ctrl.bits.block_task      |= ovfl_arg->ovfl_ctrl.bits.block_task;
5260			ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5261			/*
5262			 * build the bitmask of pmds to reset now
5263			 */
5264			if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5265
5266			pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5267		}
5268		/*
5269		 * when the module cannot handle the rest of the overflows, we abort right here
5270		 */
5271		if (ret && pmd_mask) {
5272			DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5273				pmd_mask<<PMU_FIRST_COUNTER));
5274		}
5275		/*
5276		 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5277		 */
5278		ovfl_pmds &= ~reset_pmds;
5279	} else {
5280		/*
5281		 * when no sampling module is used, then the default
5282		 * is to notify on overflow if requested by user
5283		 */
5284		ovfl_ctrl.bits.notify_user     = ovfl_notify ? 1 : 0;
5285		ovfl_ctrl.bits.block_task      = ovfl_notify ? 1 : 0;
5286		ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5287		ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5288		/*
5289		 * if needed, we reset all overflowed pmds
5290		 */
5291		if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5292	}
5293
5294	DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5295
5296	/*
5297	 * reset the requested PMD registers using the short reset values
5298	 */
5299	if (reset_pmds) {
5300		unsigned long bm = reset_pmds;
5301		pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5302	}
5303
5304	if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5305		/*
5306		 * keep track of what to reset when unblocking
5307		 */
5308		ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5309
5310		/*
5311		 * check for blocking context 
5312		 */
5313		if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5314
5315			ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5316
5317			/*
5318			 * set the perfmon specific checking pending work for the task
5319			 */
5320			PFM_SET_WORK_PENDING(task, 1);
5321
5322			/*
5323			 * when coming from ctxsw, current still points to the
5324			 * previous task, therefore we must work with task and not current.
5325			 */
5326			set_notify_resume(task);
5327		}
5328		/*
5329		 * defer until state is changed (shorten spin window). the context is locked
5330		 * anyway, so the signal receiver would come spin for nothing.
5331		 */
5332		must_notify = 1;
5333	}
5334
5335	DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5336			GET_PMU_OWNER() ? task_pid_nr(GET_PMU_OWNER()) : -1,
5337			PFM_GET_WORK_PENDING(task),
5338			ctx->ctx_fl_trap_reason,
5339			ovfl_pmds,
5340			ovfl_notify,
5341			ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5342	/*
5343	 * in case monitoring must be stopped, we toggle the psr bits
5344	 */
5345	if (ovfl_ctrl.bits.mask_monitoring) {
5346		pfm_mask_monitoring(task);
5347		ctx->ctx_state = PFM_CTX_MASKED;
5348		ctx->ctx_fl_can_restart = 1;
5349	}
5350
5351	/*
5352	 * send notification now
5353	 */
5354	if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5355
5356	return;
5357
5358sanity_check:
5359	printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5360			smp_processor_id(),
5361			task ? task_pid_nr(task) : -1,
5362			pmc0);
5363	return;
5364
5365stop_monitoring:
5366	/*
5367	 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5368	 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5369	 * come here as zombie only if the task is the current task. In which case, we
5370	 * can access the PMU  hardware directly.
5371	 *
5372	 * Note that zombies do have PM_VALID set. So here we do the minimal.
5373	 *
5374	 * In case the context was zombified it could not be reclaimed at the time
5375	 * the monitoring program exited. At this point, the PMU reservation has been
5376	 * returned, the sampiing buffer has been freed. We must convert this call
5377	 * into a spurious interrupt. However, we must also avoid infinite overflows
5378	 * by stopping monitoring for this task. We can only come here for a per-task
5379	 * context. All we need to do is to stop monitoring using the psr bits which
5380	 * are always task private. By re-enabling secure montioring, we ensure that
5381	 * the monitored task will not be able to re-activate monitoring.
5382	 * The task will eventually be context switched out, at which point the context
5383	 * will be reclaimed (that includes releasing ownership of the PMU).
5384	 *
5385	 * So there might be a window of time where the number of per-task session is zero
5386	 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5387	 * context. This is safe because if a per-task session comes in, it will push this one
5388	 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5389	 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5390	 * also push our zombie context out.
5391	 *
5392	 * Overall pretty hairy stuff....
5393	 */
5394	DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task_pid_nr(task): -1));
5395	pfm_clear_psr_up();
5396	ia64_psr(regs)->up = 0;
5397	ia64_psr(regs)->sp = 1;
5398	return;
5399}
5400
5401static int
5402pfm_do_interrupt_handler(void *arg, struct pt_regs *regs)
5403{
5404	struct task_struct *task;
5405	pfm_context_t *ctx;
5406	unsigned long flags;
5407	u64 pmc0;
5408	int this_cpu = smp_processor_id();
5409	int retval = 0;
5410
5411	pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5412
5413	/*
5414	 * srlz.d done before arriving here
5415	 */
5416	pmc0 = ia64_get_pmc(0);
5417
5418	task = GET_PMU_OWNER();
5419	ctx  = GET_PMU_CTX();
5420
5421	/*
5422	 * if we have some pending bits set
5423	 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5424	 */
5425	if (PMC0_HAS_OVFL(pmc0) && task) {
5426		/*
5427		 * we assume that pmc0.fr is always set here
5428		 */
5429
5430		/* sanity check */
5431		if (!ctx) goto report_spurious1;
5432
5433		if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0) 
5434			goto report_spurious2;
5435
5436		PROTECT_CTX_NOPRINT(ctx, flags);
5437
5438		pfm_overflow_handler(task, ctx, pmc0, regs);
5439
5440		UNPROTECT_CTX_NOPRINT(ctx, flags);
5441
5442	} else {
5443		pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5444		retval = -1;
5445	}
5446	/*
5447	 * keep it unfrozen at all times
5448	 */
5449	pfm_unfreeze_pmu();
5450
5451	return retval;
5452
5453report_spurious1:
5454	printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5455		this_cpu, task_pid_nr(task));
5456	pfm_unfreeze_pmu();
5457	return -1;
5458report_spurious2:
5459	printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n", 
5460		this_cpu, 
5461		task_pid_nr(task));
5462	pfm_unfreeze_pmu();
5463	return -1;
5464}
5465
5466static irqreturn_t
5467pfm_interrupt_handler(int irq, void *arg)
5468{
5469	unsigned long start_cycles, total_cycles;
5470	unsigned long min, max;
5471	int this_cpu;
5472	int ret;
5473	struct pt_regs *regs = get_irq_regs();
5474
5475	this_cpu = get_cpu();
5476	if (likely(!pfm_alt_intr_handler)) {
5477		min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5478		max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5479
5480		start_cycles = ia64_get_itc();
5481
5482		ret = pfm_do_interrupt_handler(arg, regs);
5483
5484		total_cycles = ia64_get_itc();
5485
5486		/*
5487		 * don't measure spurious interrupts
5488		 */
5489		if (likely(ret == 0)) {
5490			total_cycles -= start_cycles;
5491
5492			if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5493			if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5494
5495			pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5496		}
5497	}
5498	else {
5499		(*pfm_alt_intr_handler->handler)(irq, arg, regs);
5500	}
5501
5502	put_cpu();
5503	return IRQ_HANDLED;
5504}
5505
5506/*
5507 * /proc/perfmon interface, for debug only
5508 */
5509
5510#define PFM_PROC_SHOW_HEADER	((void *)(long)nr_cpu_ids+1)
5511
5512static void *
5513pfm_proc_start(struct seq_file *m, loff_t *pos)
5514{
5515	if (*pos == 0) {
5516		return PFM_PROC_SHOW_HEADER;
5517	}
5518
5519	while (*pos <= nr_cpu_ids) {
5520		if (cpu_online(*pos - 1)) {
5521			return (void *)*pos;
5522		}
5523		++*pos;
5524	}
5525	return NULL;
5526}
5527
5528static void *
5529pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5530{
5531	++*pos;
5532	return pfm_proc_start(m, pos);
5533}
5534
5535static void
5536pfm_proc_stop(struct seq_file *m, void *v)
5537{
5538}
5539
5540static void
5541pfm_proc_show_header(struct seq_file *m)
5542{
5543	struct list_head * pos;
5544	pfm_buffer_fmt_t * entry;
5545	unsigned long flags;
5546
5547 	seq_printf(m,
5548		"perfmon version           : %u.%u\n"
5549		"model                     : %s\n"
5550		"fastctxsw                 : %s\n"
5551		"expert mode               : %s\n"
5552		"ovfl_mask                 : 0x%lx\n"
5553		"PMU flags                 : 0x%x\n",
5554		PFM_VERSION_MAJ, PFM_VERSION_MIN,
5555		pmu_conf->pmu_name,
5556		pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5557		pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5558		pmu_conf->ovfl_val,
5559		pmu_conf->flags);
5560
5561  	LOCK_PFS(flags);
5562
5563 	seq_printf(m,
5564 		"proc_sessions             : %u\n"
5565 		"sys_sessions              : %u\n"
5566 		"sys_use_dbregs            : %u\n"
5567 		"ptrace_use_dbregs         : %u\n",
5568 		pfm_sessions.pfs_task_sessions,
5569 		pfm_sessions.pfs_sys_sessions,
5570 		pfm_sessions.pfs_sys_use_dbregs,
5571 		pfm_sessions.pfs_ptrace_use_dbregs);
5572
5573  	UNLOCK_PFS(flags);
5574
5575	spin_lock(&pfm_buffer_fmt_lock);
5576
5577	list_for_each(pos, &pfm_buffer_fmt_list) {
5578		entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5579		seq_printf(m, "format                    : %16phD %s\n",
5580			   entry->fmt_uuid, entry->fmt_name);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5581	}
5582	spin_unlock(&pfm_buffer_fmt_lock);
5583
5584}
5585
5586static int
5587pfm_proc_show(struct seq_file *m, void *v)
5588{
5589	unsigned long psr;
5590	unsigned int i;
5591	int cpu;
5592
5593	if (v == PFM_PROC_SHOW_HEADER) {
5594		pfm_proc_show_header(m);
5595		return 0;
5596	}
5597
5598	/* show info for CPU (v - 1) */
5599
5600	cpu = (long)v - 1;
5601	seq_printf(m,
5602		"CPU%-2d overflow intrs      : %lu\n"
5603		"CPU%-2d overflow cycles     : %lu\n"
5604		"CPU%-2d overflow min        : %lu\n"
5605		"CPU%-2d overflow max        : %lu\n"
5606		"CPU%-2d smpl handler calls  : %lu\n"
5607		"CPU%-2d smpl handler cycles : %lu\n"
5608		"CPU%-2d spurious intrs      : %lu\n"
5609		"CPU%-2d replay   intrs      : %lu\n"
5610		"CPU%-2d syst_wide           : %d\n"
5611		"CPU%-2d dcr_pp              : %d\n"
5612		"CPU%-2d exclude idle        : %d\n"
5613		"CPU%-2d owner               : %d\n"
5614		"CPU%-2d context             : %p\n"
5615		"CPU%-2d activations         : %lu\n",
5616		cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5617		cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5618		cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5619		cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5620		cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5621		cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5622		cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5623		cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5624		cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5625		cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5626		cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5627		cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5628		cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5629		cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5630
5631	if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5632
5633		psr = pfm_get_psr();
5634
5635		ia64_srlz_d();
5636
5637		seq_printf(m, 
5638			"CPU%-2d psr                 : 0x%lx\n"
5639			"CPU%-2d pmc0                : 0x%lx\n", 
5640			cpu, psr,
5641			cpu, ia64_get_pmc(0));
5642
5643		for (i=0; PMC_IS_LAST(i) == 0;  i++) {
5644			if (PMC_IS_COUNTING(i) == 0) continue;
5645   			seq_printf(m, 
5646				"CPU%-2d pmc%u                : 0x%lx\n"
5647   				"CPU%-2d pmd%u                : 0x%lx\n", 
5648				cpu, i, ia64_get_pmc(i),
5649				cpu, i, ia64_get_pmd(i));
5650  		}
5651	}
5652	return 0;
5653}
5654
5655const struct seq_operations pfm_seq_ops = {
5656	.start =	pfm_proc_start,
5657 	.next =		pfm_proc_next,
5658 	.stop =		pfm_proc_stop,
5659 	.show =		pfm_proc_show
5660};
5661
 
 
 
 
 
 
 
5662/*
5663 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5664 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5665 * is active or inactive based on mode. We must rely on the value in
5666 * local_cpu_data->pfm_syst_info
5667 */
5668void
5669pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5670{
5671	struct pt_regs *regs;
5672	unsigned long dcr;
5673	unsigned long dcr_pp;
5674
5675	dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5676
5677	/*
5678	 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5679	 * on every CPU, so we can rely on the pid to identify the idle task.
5680	 */
5681	if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5682		regs = task_pt_regs(task);
5683		ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5684		return;
5685	}
5686	/*
5687	 * if monitoring has started
5688	 */
5689	if (dcr_pp) {
5690		dcr = ia64_getreg(_IA64_REG_CR_DCR);
5691		/*
5692		 * context switching in?
5693		 */
5694		if (is_ctxswin) {
5695			/* mask monitoring for the idle task */
5696			ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5697			pfm_clear_psr_pp();
5698			ia64_srlz_i();
5699			return;
5700		}
5701		/*
5702		 * context switching out
5703		 * restore monitoring for next task
5704		 *
5705		 * Due to inlining this odd if-then-else construction generates
5706		 * better code.
5707		 */
5708		ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5709		pfm_set_psr_pp();
5710		ia64_srlz_i();
5711	}
5712}
5713
5714#ifdef CONFIG_SMP
5715
5716static void
5717pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5718{
5719	struct task_struct *task = ctx->ctx_task;
5720
5721	ia64_psr(regs)->up = 0;
5722	ia64_psr(regs)->sp = 1;
5723
5724	if (GET_PMU_OWNER() == task) {
5725		DPRINT(("cleared ownership for [%d]\n",
5726					task_pid_nr(ctx->ctx_task)));
5727		SET_PMU_OWNER(NULL, NULL);
5728	}
5729
5730	/*
5731	 * disconnect the task from the context and vice-versa
5732	 */
5733	PFM_SET_WORK_PENDING(task, 0);
5734
5735	task->thread.pfm_context  = NULL;
5736	task->thread.flags       &= ~IA64_THREAD_PM_VALID;
5737
5738	DPRINT(("force cleanup for [%d]\n",  task_pid_nr(task)));
5739}
5740
5741
5742/*
5743 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5744 */
5745void
5746pfm_save_regs(struct task_struct *task)
5747{
5748	pfm_context_t *ctx;
5749	unsigned long flags;
5750	u64 psr;
5751
5752
5753	ctx = PFM_GET_CTX(task);
5754	if (ctx == NULL) return;
5755
5756	/*
5757 	 * we always come here with interrupts ALREADY disabled by
5758 	 * the scheduler. So we simply need to protect against concurrent
5759	 * access, not CPU concurrency.
5760	 */
5761	flags = pfm_protect_ctx_ctxsw(ctx);
5762
5763	if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5764		struct pt_regs *regs = task_pt_regs(task);
5765
5766		pfm_clear_psr_up();
5767
5768		pfm_force_cleanup(ctx, regs);
5769
5770		BUG_ON(ctx->ctx_smpl_hdr);
5771
5772		pfm_unprotect_ctx_ctxsw(ctx, flags);
5773
5774		pfm_context_free(ctx);
5775		return;
5776	}
5777
5778	/*
5779	 * save current PSR: needed because we modify it
5780	 */
5781	ia64_srlz_d();
5782	psr = pfm_get_psr();
5783
5784	BUG_ON(psr & (IA64_PSR_I));
5785
5786	/*
5787	 * stop monitoring:
5788	 * This is the last instruction which may generate an overflow
5789	 *
5790	 * We do not need to set psr.sp because, it is irrelevant in kernel.
5791	 * It will be restored from ipsr when going back to user level
5792	 */
5793	pfm_clear_psr_up();
5794
5795	/*
5796	 * keep a copy of psr.up (for reload)
5797	 */
5798	ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5799
5800	/*
5801	 * release ownership of this PMU.
5802	 * PM interrupts are masked, so nothing
5803	 * can happen.
5804	 */
5805	SET_PMU_OWNER(NULL, NULL);
5806
5807	/*
5808	 * we systematically save the PMD as we have no
5809	 * guarantee we will be schedule at that same
5810	 * CPU again.
5811	 */
5812	pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5813
5814	/*
5815	 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5816	 * we will need it on the restore path to check
5817	 * for pending overflow.
5818	 */
5819	ctx->th_pmcs[0] = ia64_get_pmc(0);
5820
5821	/*
5822	 * unfreeze PMU if had pending overflows
5823	 */
5824	if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5825
5826	/*
5827	 * finally, allow context access.
5828	 * interrupts will still be masked after this call.
5829	 */
5830	pfm_unprotect_ctx_ctxsw(ctx, flags);
5831}
5832
5833#else /* !CONFIG_SMP */
5834void
5835pfm_save_regs(struct task_struct *task)
5836{
5837	pfm_context_t *ctx;
5838	u64 psr;
5839
5840	ctx = PFM_GET_CTX(task);
5841	if (ctx == NULL) return;
5842
5843	/*
5844	 * save current PSR: needed because we modify it
5845	 */
5846	psr = pfm_get_psr();
5847
5848	BUG_ON(psr & (IA64_PSR_I));
5849
5850	/*
5851	 * stop monitoring:
5852	 * This is the last instruction which may generate an overflow
5853	 *
5854	 * We do not need to set psr.sp because, it is irrelevant in kernel.
5855	 * It will be restored from ipsr when going back to user level
5856	 */
5857	pfm_clear_psr_up();
5858
5859	/*
5860	 * keep a copy of psr.up (for reload)
5861	 */
5862	ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5863}
5864
5865static void
5866pfm_lazy_save_regs (struct task_struct *task)
5867{
5868	pfm_context_t *ctx;
5869	unsigned long flags;
5870
5871	{ u64 psr  = pfm_get_psr();
5872	  BUG_ON(psr & IA64_PSR_UP);
5873	}
5874
5875	ctx = PFM_GET_CTX(task);
5876
5877	/*
5878	 * we need to mask PMU overflow here to
5879	 * make sure that we maintain pmc0 until
5880	 * we save it. overflow interrupts are
5881	 * treated as spurious if there is no
5882	 * owner.
5883	 *
5884	 * XXX: I don't think this is necessary
5885	 */
5886	PROTECT_CTX(ctx,flags);
5887
5888	/*
5889	 * release ownership of this PMU.
5890	 * must be done before we save the registers.
5891	 *
5892	 * after this call any PMU interrupt is treated
5893	 * as spurious.
5894	 */
5895	SET_PMU_OWNER(NULL, NULL);
5896
5897	/*
5898	 * save all the pmds we use
5899	 */
5900	pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5901
5902	/*
5903	 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5904	 * it is needed to check for pended overflow
5905	 * on the restore path
5906	 */
5907	ctx->th_pmcs[0] = ia64_get_pmc(0);
5908
5909	/*
5910	 * unfreeze PMU if had pending overflows
5911	 */
5912	if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5913
5914	/*
5915	 * now get can unmask PMU interrupts, they will
5916	 * be treated as purely spurious and we will not
5917	 * lose any information
5918	 */
5919	UNPROTECT_CTX(ctx,flags);
5920}
5921#endif /* CONFIG_SMP */
5922
5923#ifdef CONFIG_SMP
5924/*
5925 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5926 */
5927void
5928pfm_load_regs (struct task_struct *task)
5929{
5930	pfm_context_t *ctx;
5931	unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
5932	unsigned long flags;
5933	u64 psr, psr_up;
5934	int need_irq_resend;
5935
5936	ctx = PFM_GET_CTX(task);
5937	if (unlikely(ctx == NULL)) return;
5938
5939	BUG_ON(GET_PMU_OWNER());
5940
5941	/*
5942	 * possible on unload
5943	 */
5944	if (unlikely((task->thread.flags & IA64_THREAD_PM_VALID) == 0)) return;
5945
5946	/*
5947 	 * we always come here with interrupts ALREADY disabled by
5948 	 * the scheduler. So we simply need to protect against concurrent
5949	 * access, not CPU concurrency.
5950	 */
5951	flags = pfm_protect_ctx_ctxsw(ctx);
5952	psr   = pfm_get_psr();
5953
5954	need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
5955
5956	BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
5957	BUG_ON(psr & IA64_PSR_I);
5958
5959	if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
5960		struct pt_regs *regs = task_pt_regs(task);
5961
5962		BUG_ON(ctx->ctx_smpl_hdr);
5963
5964		pfm_force_cleanup(ctx, regs);
5965
5966		pfm_unprotect_ctx_ctxsw(ctx, flags);
5967
5968		/*
5969		 * this one (kmalloc'ed) is fine with interrupts disabled
5970		 */
5971		pfm_context_free(ctx);
5972
5973		return;
5974	}
5975
5976	/*
5977	 * we restore ALL the debug registers to avoid picking up
5978	 * stale state.
5979	 */
5980	if (ctx->ctx_fl_using_dbreg) {
5981		pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
5982		pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
5983	}
5984	/*
5985	 * retrieve saved psr.up
5986	 */
5987	psr_up = ctx->ctx_saved_psr_up;
5988
5989	/*
5990	 * if we were the last user of the PMU on that CPU,
5991	 * then nothing to do except restore psr
5992	 */
5993	if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
5994
5995		/*
5996		 * retrieve partial reload masks (due to user modifications)
5997		 */
5998		pmc_mask = ctx->ctx_reload_pmcs[0];
5999		pmd_mask = ctx->ctx_reload_pmds[0];
6000
6001	} else {
6002		/*
6003	 	 * To avoid leaking information to the user level when psr.sp=0,
6004	 	 * we must reload ALL implemented pmds (even the ones we don't use).
6005	 	 * In the kernel we only allow PFM_READ_PMDS on registers which
6006	 	 * we initialized or requested (sampling) so there is no risk there.
6007	 	 */
6008		pmd_mask = pfm_sysctl.fastctxsw ?  ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6009
6010		/*
6011	 	 * ALL accessible PMCs are systematically reloaded, unused registers
6012	 	 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6013	 	 * up stale configuration.
6014	 	 *
6015	 	 * PMC0 is never in the mask. It is always restored separately.
6016	 	 */
6017		pmc_mask = ctx->ctx_all_pmcs[0];
6018	}
6019	/*
6020	 * when context is MASKED, we will restore PMC with plm=0
6021	 * and PMD with stale information, but that's ok, nothing
6022	 * will be captured.
6023	 *
6024	 * XXX: optimize here
6025	 */
6026	if (pmd_mask) pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6027	if (pmc_mask) pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6028
6029	/*
6030	 * check for pending overflow at the time the state
6031	 * was saved.
6032	 */
6033	if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6034		/*
6035		 * reload pmc0 with the overflow information
6036		 * On McKinley PMU, this will trigger a PMU interrupt
6037		 */
6038		ia64_set_pmc(0, ctx->th_pmcs[0]);
6039		ia64_srlz_d();
6040		ctx->th_pmcs[0] = 0UL;
6041
6042		/*
6043		 * will replay the PMU interrupt
6044		 */
6045		if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6046
6047		pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6048	}
6049
6050	/*
6051	 * we just did a reload, so we reset the partial reload fields
6052	 */
6053	ctx->ctx_reload_pmcs[0] = 0UL;
6054	ctx->ctx_reload_pmds[0] = 0UL;
6055
6056	SET_LAST_CPU(ctx, smp_processor_id());
6057
6058	/*
6059	 * dump activation value for this PMU
6060	 */
6061	INC_ACTIVATION();
6062	/*
6063	 * record current activation for this context
6064	 */
6065	SET_ACTIVATION(ctx);
6066
6067	/*
6068	 * establish new ownership. 
6069	 */
6070	SET_PMU_OWNER(task, ctx);
6071
6072	/*
6073	 * restore the psr.up bit. measurement
6074	 * is active again.
6075	 * no PMU interrupt can happen at this point
6076	 * because we still have interrupts disabled.
6077	 */
6078	if (likely(psr_up)) pfm_set_psr_up();
6079
6080	/*
6081	 * allow concurrent access to context
6082	 */
6083	pfm_unprotect_ctx_ctxsw(ctx, flags);
6084}
6085#else /*  !CONFIG_SMP */
6086/*
6087 * reload PMU state for UP kernels
6088 * in 2.5 we come here with interrupts disabled
6089 */
6090void
6091pfm_load_regs (struct task_struct *task)
6092{
6093	pfm_context_t *ctx;
6094	struct task_struct *owner;
6095	unsigned long pmd_mask, pmc_mask;
6096	u64 psr, psr_up;
6097	int need_irq_resend;
6098
6099	owner = GET_PMU_OWNER();
6100	ctx   = PFM_GET_CTX(task);
6101	psr   = pfm_get_psr();
6102
6103	BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6104	BUG_ON(psr & IA64_PSR_I);
6105
6106	/*
6107	 * we restore ALL the debug registers to avoid picking up
6108	 * stale state.
6109	 *
6110	 * This must be done even when the task is still the owner
6111	 * as the registers may have been modified via ptrace()
6112	 * (not perfmon) by the previous task.
6113	 */
6114	if (ctx->ctx_fl_using_dbreg) {
6115		pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6116		pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6117	}
6118
6119	/*
6120	 * retrieved saved psr.up
6121	 */
6122	psr_up = ctx->ctx_saved_psr_up;
6123	need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6124
6125	/*
6126	 * short path, our state is still there, just
6127	 * need to restore psr and we go
6128	 *
6129	 * we do not touch either PMC nor PMD. the psr is not touched
6130	 * by the overflow_handler. So we are safe w.r.t. to interrupt
6131	 * concurrency even without interrupt masking.
6132	 */
6133	if (likely(owner == task)) {
6134		if (likely(psr_up)) pfm_set_psr_up();
6135		return;
6136	}
6137
6138	/*
6139	 * someone else is still using the PMU, first push it out and
6140	 * then we'll be able to install our stuff !
6141	 *
6142	 * Upon return, there will be no owner for the current PMU
6143	 */
6144	if (owner) pfm_lazy_save_regs(owner);
6145
6146	/*
6147	 * To avoid leaking information to the user level when psr.sp=0,
6148	 * we must reload ALL implemented pmds (even the ones we don't use).
6149	 * In the kernel we only allow PFM_READ_PMDS on registers which
6150	 * we initialized or requested (sampling) so there is no risk there.
6151	 */
6152	pmd_mask = pfm_sysctl.fastctxsw ?  ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6153
6154	/*
6155	 * ALL accessible PMCs are systematically reloaded, unused registers
6156	 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6157	 * up stale configuration.
6158	 *
6159	 * PMC0 is never in the mask. It is always restored separately
6160	 */
6161	pmc_mask = ctx->ctx_all_pmcs[0];
6162
6163	pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6164	pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6165
6166	/*
6167	 * check for pending overflow at the time the state
6168	 * was saved.
6169	 */
6170	if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6171		/*
6172		 * reload pmc0 with the overflow information
6173		 * On McKinley PMU, this will trigger a PMU interrupt
6174		 */
6175		ia64_set_pmc(0, ctx->th_pmcs[0]);
6176		ia64_srlz_d();
6177
6178		ctx->th_pmcs[0] = 0UL;
6179
6180		/*
6181		 * will replay the PMU interrupt
6182		 */
6183		if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6184
6185		pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6186	}
6187
6188	/*
6189	 * establish new ownership. 
6190	 */
6191	SET_PMU_OWNER(task, ctx);
6192
6193	/*
6194	 * restore the psr.up bit. measurement
6195	 * is active again.
6196	 * no PMU interrupt can happen at this point
6197	 * because we still have interrupts disabled.
6198	 */
6199	if (likely(psr_up)) pfm_set_psr_up();
6200}
6201#endif /* CONFIG_SMP */
6202
6203/*
6204 * this function assumes monitoring is stopped
6205 */
6206static void
6207pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6208{
6209	u64 pmc0;
6210	unsigned long mask2, val, pmd_val, ovfl_val;
6211	int i, can_access_pmu = 0;
6212	int is_self;
6213
6214	/*
6215	 * is the caller the task being monitored (or which initiated the
6216	 * session for system wide measurements)
6217	 */
6218	is_self = ctx->ctx_task == task ? 1 : 0;
6219
6220	/*
6221	 * can access PMU is task is the owner of the PMU state on the current CPU
6222	 * or if we are running on the CPU bound to the context in system-wide mode
6223	 * (that is not necessarily the task the context is attached to in this mode).
6224	 * In system-wide we always have can_access_pmu true because a task running on an
6225	 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6226	 */
6227	can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6228	if (can_access_pmu) {
6229		/*
6230		 * Mark the PMU as not owned
6231		 * This will cause the interrupt handler to do nothing in case an overflow
6232		 * interrupt was in-flight
6233		 * This also guarantees that pmc0 will contain the final state
6234		 * It virtually gives us full control on overflow processing from that point
6235		 * on.
6236		 */
6237		SET_PMU_OWNER(NULL, NULL);
6238		DPRINT(("releasing ownership\n"));
6239
6240		/*
6241		 * read current overflow status:
6242		 *
6243		 * we are guaranteed to read the final stable state
6244		 */
6245		ia64_srlz_d();
6246		pmc0 = ia64_get_pmc(0); /* slow */
6247
6248		/*
6249		 * reset freeze bit, overflow status information destroyed
6250		 */
6251		pfm_unfreeze_pmu();
6252	} else {
6253		pmc0 = ctx->th_pmcs[0];
6254		/*
6255		 * clear whatever overflow status bits there were
6256		 */
6257		ctx->th_pmcs[0] = 0;
6258	}
6259	ovfl_val = pmu_conf->ovfl_val;
6260	/*
6261	 * we save all the used pmds
6262	 * we take care of overflows for counting PMDs
6263	 *
6264	 * XXX: sampling situation is not taken into account here
6265	 */
6266	mask2 = ctx->ctx_used_pmds[0];
6267
6268	DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6269
6270	for (i = 0; mask2; i++, mask2>>=1) {
6271
6272		/* skip non used pmds */
6273		if ((mask2 & 0x1) == 0) continue;
6274
6275		/*
6276		 * can access PMU always true in system wide mode
6277		 */
6278		val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : ctx->th_pmds[i];
6279
6280		if (PMD_IS_COUNTING(i)) {
6281			DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6282				task_pid_nr(task),
6283				i,
6284				ctx->ctx_pmds[i].val,
6285				val & ovfl_val));
6286
6287			/*
6288			 * we rebuild the full 64 bit value of the counter
6289			 */
6290			val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6291
6292			/*
6293			 * now everything is in ctx_pmds[] and we need
6294			 * to clear the saved context from save_regs() such that
6295			 * pfm_read_pmds() gets the correct value
6296			 */
6297			pmd_val = 0UL;
6298
6299			/*
6300			 * take care of overflow inline
6301			 */
6302			if (pmc0 & (1UL << i)) {
6303				val += 1 + ovfl_val;
6304				DPRINT(("[%d] pmd[%d] overflowed\n", task_pid_nr(task), i));
6305			}
6306		}
6307
6308		DPRINT(("[%d] ctx_pmd[%d]=0x%lx  pmd_val=0x%lx\n", task_pid_nr(task), i, val, pmd_val));
6309
6310		if (is_self) ctx->th_pmds[i] = pmd_val;
6311
6312		ctx->ctx_pmds[i].val = val;
6313	}
6314}
6315
6316static struct irqaction perfmon_irqaction = {
6317	.handler = pfm_interrupt_handler,
 
6318	.name    = "perfmon"
6319};
6320
6321static void
6322pfm_alt_save_pmu_state(void *data)
6323{
6324	struct pt_regs *regs;
6325
6326	regs = task_pt_regs(current);
6327
6328	DPRINT(("called\n"));
6329
6330	/*
6331	 * should not be necessary but
6332	 * let's take not risk
6333	 */
6334	pfm_clear_psr_up();
6335	pfm_clear_psr_pp();
6336	ia64_psr(regs)->pp = 0;
6337
6338	/*
6339	 * This call is required
6340	 * May cause a spurious interrupt on some processors
6341	 */
6342	pfm_freeze_pmu();
6343
6344	ia64_srlz_d();
6345}
6346
6347void
6348pfm_alt_restore_pmu_state(void *data)
6349{
6350	struct pt_regs *regs;
6351
6352	regs = task_pt_regs(current);
6353
6354	DPRINT(("called\n"));
6355
6356	/*
6357	 * put PMU back in state expected
6358	 * by perfmon
6359	 */
6360	pfm_clear_psr_up();
6361	pfm_clear_psr_pp();
6362	ia64_psr(regs)->pp = 0;
6363
6364	/*
6365	 * perfmon runs with PMU unfrozen at all times
6366	 */
6367	pfm_unfreeze_pmu();
6368
6369	ia64_srlz_d();
6370}
6371
6372int
6373pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6374{
6375	int ret, i;
6376	int reserve_cpu;
6377
6378	/* some sanity checks */
6379	if (hdl == NULL || hdl->handler == NULL) return -EINVAL;
6380
6381	/* do the easy test first */
6382	if (pfm_alt_intr_handler) return -EBUSY;
6383
6384	/* one at a time in the install or remove, just fail the others */
6385	if (!spin_trylock(&pfm_alt_install_check)) {
6386		return -EBUSY;
6387	}
6388
6389	/* reserve our session */
6390	for_each_online_cpu(reserve_cpu) {
6391		ret = pfm_reserve_session(NULL, 1, reserve_cpu);
6392		if (ret) goto cleanup_reserve;
6393	}
6394
6395	/* save the current system wide pmu states */
6396	on_each_cpu(pfm_alt_save_pmu_state, NULL, 1);
 
 
 
 
6397
6398	/* officially change to the alternate interrupt handler */
6399	pfm_alt_intr_handler = hdl;
6400
6401	spin_unlock(&pfm_alt_install_check);
6402
6403	return 0;
6404
6405cleanup_reserve:
6406	for_each_online_cpu(i) {
6407		/* don't unreserve more than we reserved */
6408		if (i >= reserve_cpu) break;
6409
6410		pfm_unreserve_session(NULL, 1, i);
6411	}
6412
6413	spin_unlock(&pfm_alt_install_check);
6414
6415	return ret;
6416}
6417EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt);
6418
6419int
6420pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6421{
6422	int i;
 
6423
6424	if (hdl == NULL) return -EINVAL;
6425
6426	/* cannot remove someone else's handler! */
6427	if (pfm_alt_intr_handler != hdl) return -EINVAL;
6428
6429	/* one at a time in the install or remove, just fail the others */
6430	if (!spin_trylock(&pfm_alt_install_check)) {
6431		return -EBUSY;
6432	}
6433
6434	pfm_alt_intr_handler = NULL;
6435
6436	on_each_cpu(pfm_alt_restore_pmu_state, NULL, 1);
 
 
 
6437
6438	for_each_online_cpu(i) {
6439		pfm_unreserve_session(NULL, 1, i);
6440	}
6441
6442	spin_unlock(&pfm_alt_install_check);
6443
6444	return 0;
6445}
6446EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt);
6447
6448/*
6449 * perfmon initialization routine, called from the initcall() table
6450 */
6451static int init_pfm_fs(void);
6452
6453static int __init
6454pfm_probe_pmu(void)
6455{
6456	pmu_config_t **p;
6457	int family;
6458
6459	family = local_cpu_data->family;
6460	p      = pmu_confs;
6461
6462	while(*p) {
6463		if ((*p)->probe) {
6464			if ((*p)->probe() == 0) goto found;
6465		} else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6466			goto found;
6467		}
6468		p++;
6469	}
6470	return -1;
6471found:
6472	pmu_conf = *p;
6473	return 0;
6474}
6475
 
 
 
 
 
 
 
6476int __init
6477pfm_init(void)
6478{
6479	unsigned int n, n_counters, i;
6480
6481	printk("perfmon: version %u.%u IRQ %u\n",
6482		PFM_VERSION_MAJ,
6483		PFM_VERSION_MIN,
6484		IA64_PERFMON_VECTOR);
6485
6486	if (pfm_probe_pmu()) {
6487		printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n", 
6488				local_cpu_data->family);
6489		return -ENODEV;
6490	}
6491
6492	/*
6493	 * compute the number of implemented PMD/PMC from the
6494	 * description tables
6495	 */
6496	n = 0;
6497	for (i=0; PMC_IS_LAST(i) == 0;  i++) {
6498		if (PMC_IS_IMPL(i) == 0) continue;
6499		pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6500		n++;
6501	}
6502	pmu_conf->num_pmcs = n;
6503
6504	n = 0; n_counters = 0;
6505	for (i=0; PMD_IS_LAST(i) == 0;  i++) {
6506		if (PMD_IS_IMPL(i) == 0) continue;
6507		pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6508		n++;
6509		if (PMD_IS_COUNTING(i)) n_counters++;
6510	}
6511	pmu_conf->num_pmds      = n;
6512	pmu_conf->num_counters  = n_counters;
6513
6514	/*
6515	 * sanity checks on the number of debug registers
6516	 */
6517	if (pmu_conf->use_rr_dbregs) {
6518		if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6519			printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6520			pmu_conf = NULL;
6521			return -1;
6522		}
6523		if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6524			printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6525			pmu_conf = NULL;
6526			return -1;
6527		}
6528	}
6529
6530	printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6531	       pmu_conf->pmu_name,
6532	       pmu_conf->num_pmcs,
6533	       pmu_conf->num_pmds,
6534	       pmu_conf->num_counters,
6535	       ffz(pmu_conf->ovfl_val));
6536
6537	/* sanity check */
6538	if (pmu_conf->num_pmds >= PFM_NUM_PMD_REGS || pmu_conf->num_pmcs >= PFM_NUM_PMC_REGS) {
6539		printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6540		pmu_conf = NULL;
6541		return -1;
6542	}
6543
6544	/*
6545	 * create /proc/perfmon (mostly for debugging purposes)
6546	 */
6547	perfmon_dir = proc_create_seq("perfmon", S_IRUGO, NULL, &pfm_seq_ops);
6548	if (perfmon_dir == NULL) {
6549		printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6550		pmu_conf = NULL;
6551		return -1;
6552	}
6553
6554	/*
6555	 * create /proc/sys/kernel/perfmon (for debugging purposes)
6556	 */
6557	pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root);
6558
6559	/*
6560	 * initialize all our spinlocks
6561	 */
6562	spin_lock_init(&pfm_sessions.pfs_lock);
6563	spin_lock_init(&pfm_buffer_fmt_lock);
6564
6565	init_pfm_fs();
6566
6567	for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6568
6569	return 0;
6570}
6571
6572__initcall(pfm_init);
6573
6574/*
6575 * this function is called before pfm_init()
6576 */
6577void
6578pfm_init_percpu (void)
6579{
6580	static int first_time=1;
6581	/*
6582	 * make sure no measurement is active
6583	 * (may inherit programmed PMCs from EFI).
6584	 */
6585	pfm_clear_psr_pp();
6586	pfm_clear_psr_up();
6587
6588	/*
6589	 * we run with the PMU not frozen at all times
6590	 */
6591	pfm_unfreeze_pmu();
6592
6593	if (first_time) {
6594		register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6595		first_time=0;
6596	}
6597
6598	ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6599	ia64_srlz_d();
6600}
6601
6602/*
6603 * used for debug purposes only
6604 */
6605void
6606dump_pmu_state(const char *from)
6607{
6608	struct task_struct *task;
6609	struct pt_regs *regs;
6610	pfm_context_t *ctx;
6611	unsigned long psr, dcr, info, flags;
6612	int i, this_cpu;
6613
6614	local_irq_save(flags);
6615
6616	this_cpu = smp_processor_id();
6617	regs     = task_pt_regs(current);
6618	info     = PFM_CPUINFO_GET();
6619	dcr      = ia64_getreg(_IA64_REG_CR_DCR);
6620
6621	if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6622		local_irq_restore(flags);
6623		return;
6624	}
6625
6626	printk("CPU%d from %s() current [%d] iip=0x%lx %s\n", 
6627		this_cpu, 
6628		from, 
6629		task_pid_nr(current),
6630		regs->cr_iip,
6631		current->comm);
6632
6633	task = GET_PMU_OWNER();
6634	ctx  = GET_PMU_CTX();
6635
6636	printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task_pid_nr(task) : -1, ctx);
6637
6638	psr = pfm_get_psr();
6639
6640	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", 
6641		this_cpu,
6642		ia64_get_pmc(0),
6643		psr & IA64_PSR_PP ? 1 : 0,
6644		psr & IA64_PSR_UP ? 1 : 0,
6645		dcr & IA64_DCR_PP ? 1 : 0,
6646		info,
6647		ia64_psr(regs)->up,
6648		ia64_psr(regs)->pp);
6649
6650	ia64_psr(regs)->up = 0;
6651	ia64_psr(regs)->pp = 0;
6652
6653	for (i=1; PMC_IS_LAST(i) == 0; i++) {
6654		if (PMC_IS_IMPL(i) == 0) continue;
6655		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]);
6656	}
6657
6658	for (i=1; PMD_IS_LAST(i) == 0; i++) {
6659		if (PMD_IS_IMPL(i) == 0) continue;
6660		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]);
6661	}
6662
6663	if (ctx) {
6664		printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6665				this_cpu,
6666				ctx->ctx_state,
6667				ctx->ctx_smpl_vaddr,
6668				ctx->ctx_smpl_hdr,
6669				ctx->ctx_msgq_head,
6670				ctx->ctx_msgq_tail,
6671				ctx->ctx_saved_psr_up);
6672	}
6673	local_irq_restore(flags);
6674}
6675
6676/*
6677 * called from process.c:copy_thread(). task is new child.
6678 */
6679void
6680pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6681{
6682	struct thread_struct *thread;
6683
6684	DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task_pid_nr(task)));
6685
6686	thread = &task->thread;
6687
6688	/*
6689	 * cut links inherited from parent (current)
6690	 */
6691	thread->pfm_context = NULL;
6692
6693	PFM_SET_WORK_PENDING(task, 0);
6694
6695	/*
6696	 * the psr bits are already set properly in copy_threads()
6697	 */
6698}
6699#else  /* !CONFIG_PERFMON */
6700asmlinkage long
6701sys_perfmonctl (int fd, int cmd, void *arg, int count)
6702{
6703	return -ENOSYS;
6704}
6705#endif /* CONFIG_PERFMON */
v3.5.6
 
   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/uaccess.h>
  53#include <asm/delay.h>
  54
  55#ifdef CONFIG_PERFMON
  56/*
  57 * perfmon context state
  58 */
  59#define PFM_CTX_UNLOADED	1	/* context is not loaded onto any task */
  60#define PFM_CTX_LOADED		2	/* context is loaded onto a task */
  61#define PFM_CTX_MASKED		3	/* context is loaded but monitoring is masked due to overflow */
  62#define PFM_CTX_ZOMBIE		4	/* owner of the context is closing it */
  63
  64#define PFM_INVALID_ACTIVATION	(~0UL)
  65
  66#define PFM_NUM_PMC_REGS	64	/* PMC save area for ctxsw */
  67#define PFM_NUM_PMD_REGS	64	/* PMD save area for ctxsw */
  68
  69/*
  70 * depth of message queue
  71 */
  72#define PFM_MAX_MSGS		32
  73#define PFM_CTXQ_EMPTY(g)	((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
  74
  75/*
  76 * type of a PMU register (bitmask).
  77 * bitmask structure:
  78 * 	bit0   : register implemented
  79 * 	bit1   : end marker
  80 * 	bit2-3 : reserved
  81 * 	bit4   : pmc has pmc.pm
  82 * 	bit5   : pmc controls a counter (has pmc.oi), pmd is used as counter
  83 * 	bit6-7 : register type
  84 * 	bit8-31: reserved
  85 */
  86#define PFM_REG_NOTIMPL		0x0 /* not implemented at all */
  87#define PFM_REG_IMPL		0x1 /* register implemented */
  88#define PFM_REG_END		0x2 /* end marker */
  89#define PFM_REG_MONITOR		(0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
  90#define PFM_REG_COUNTING	(0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
  91#define PFM_REG_CONTROL		(0x4<<4|PFM_REG_IMPL) /* PMU control register */
  92#define	PFM_REG_CONFIG		(0x8<<4|PFM_REG_IMPL) /* configuration register */
  93#define PFM_REG_BUFFER	 	(0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
  94
  95#define PMC_IS_LAST(i)	(pmu_conf->pmc_desc[i].type & PFM_REG_END)
  96#define PMD_IS_LAST(i)	(pmu_conf->pmd_desc[i].type & PFM_REG_END)
  97
  98#define PMC_OVFL_NOTIFY(ctx, i)	((ctx)->ctx_pmds[i].flags &  PFM_REGFL_OVFL_NOTIFY)
  99
 100/* i assumed unsigned */
 101#define PMC_IS_IMPL(i)	  (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
 102#define PMD_IS_IMPL(i)	  (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
 103
 104/* XXX: these assume that register i is implemented */
 105#define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
 106#define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
 107#define PMC_IS_MONITOR(i)  ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR)  == PFM_REG_MONITOR)
 108#define PMC_IS_CONTROL(i)  ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL)  == PFM_REG_CONTROL)
 109
 110#define PMC_DFL_VAL(i)     pmu_conf->pmc_desc[i].default_value
 111#define PMC_RSVD_MASK(i)   pmu_conf->pmc_desc[i].reserved_mask
 112#define PMD_PMD_DEP(i)	   pmu_conf->pmd_desc[i].dep_pmd[0]
 113#define PMC_PMD_DEP(i)	   pmu_conf->pmc_desc[i].dep_pmd[0]
 114
 115#define PFM_NUM_IBRS	  IA64_NUM_DBG_REGS
 116#define PFM_NUM_DBRS	  IA64_NUM_DBG_REGS
 117
 118#define CTX_OVFL_NOBLOCK(c)	((c)->ctx_fl_block == 0)
 119#define CTX_HAS_SMPL(c)		((c)->ctx_fl_is_sampling)
 120#define PFM_CTX_TASK(h)		(h)->ctx_task
 121
 122#define PMU_PMC_OI		5 /* position of pmc.oi bit */
 123
 124/* XXX: does not support more than 64 PMDs */
 125#define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
 126#define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
 127
 128#define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
 129
 130#define CTX_USED_IBR(ctx,n) 	(ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
 131#define CTX_USED_DBR(ctx,n) 	(ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
 132#define CTX_USES_DBREGS(ctx)	(((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
 133#define PFM_CODE_RR	0	/* requesting code range restriction */
 134#define PFM_DATA_RR	1	/* requestion data range restriction */
 135
 136#define PFM_CPUINFO_CLEAR(v)	pfm_get_cpu_var(pfm_syst_info) &= ~(v)
 137#define PFM_CPUINFO_SET(v)	pfm_get_cpu_var(pfm_syst_info) |= (v)
 138#define PFM_CPUINFO_GET()	pfm_get_cpu_var(pfm_syst_info)
 139
 140#define RDEP(x)	(1UL<<(x))
 141
 142/*
 143 * context protection macros
 144 * in SMP:
 145 * 	- we need to protect against CPU concurrency (spin_lock)
 146 * 	- we need to protect against PMU overflow interrupts (local_irq_disable)
 147 * in UP:
 148 * 	- we need to protect against PMU overflow interrupts (local_irq_disable)
 149 *
 150 * spin_lock_irqsave()/spin_unlock_irqrestore():
 151 * 	in SMP: local_irq_disable + spin_lock
 152 * 	in UP : local_irq_disable
 153 *
 154 * spin_lock()/spin_lock():
 155 * 	in UP : removed automatically
 156 * 	in SMP: protect against context accesses from other CPU. interrupts
 157 * 	        are not masked. This is useful for the PMU interrupt handler
 158 * 	        because we know we will not get PMU concurrency in that code.
 159 */
 160#define PROTECT_CTX(c, f) \
 161	do {  \
 162		DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, task_pid_nr(current))); \
 163		spin_lock_irqsave(&(c)->ctx_lock, f); \
 164		DPRINT(("spinlocked ctx %p  by [%d]\n", c, task_pid_nr(current))); \
 165	} while(0)
 166
 167#define UNPROTECT_CTX(c, f) \
 168	do { \
 169		DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, task_pid_nr(current))); \
 170		spin_unlock_irqrestore(&(c)->ctx_lock, f); \
 171	} while(0)
 172
 173#define PROTECT_CTX_NOPRINT(c, f) \
 174	do {  \
 175		spin_lock_irqsave(&(c)->ctx_lock, f); \
 176	} while(0)
 177
 178
 179#define UNPROTECT_CTX_NOPRINT(c, f) \
 180	do { \
 181		spin_unlock_irqrestore(&(c)->ctx_lock, f); \
 182	} while(0)
 183
 184
 185#define PROTECT_CTX_NOIRQ(c) \
 186	do {  \
 187		spin_lock(&(c)->ctx_lock); \
 188	} while(0)
 189
 190#define UNPROTECT_CTX_NOIRQ(c) \
 191	do { \
 192		spin_unlock(&(c)->ctx_lock); \
 193	} while(0)
 194
 195
 196#ifdef CONFIG_SMP
 197
 198#define GET_ACTIVATION()	pfm_get_cpu_var(pmu_activation_number)
 199#define INC_ACTIVATION()	pfm_get_cpu_var(pmu_activation_number)++
 200#define SET_ACTIVATION(c)	(c)->ctx_last_activation = GET_ACTIVATION()
 201
 202#else /* !CONFIG_SMP */
 203#define SET_ACTIVATION(t) 	do {} while(0)
 204#define GET_ACTIVATION(t) 	do {} while(0)
 205#define INC_ACTIVATION(t) 	do {} while(0)
 206#endif /* CONFIG_SMP */
 207
 208#define SET_PMU_OWNER(t, c)	do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
 209#define GET_PMU_OWNER()		pfm_get_cpu_var(pmu_owner)
 210#define GET_PMU_CTX()		pfm_get_cpu_var(pmu_ctx)
 211
 212#define LOCK_PFS(g)	    	spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
 213#define UNLOCK_PFS(g)	    	spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
 214
 215#define PFM_REG_RETFLAG_SET(flags, val)	do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
 216
 217/*
 218 * cmp0 must be the value of pmc0
 219 */
 220#define PMC0_HAS_OVFL(cmp0)  (cmp0 & ~0x1UL)
 221
 222#define PFMFS_MAGIC 0xa0b4d889
 223
 224/*
 225 * debugging
 226 */
 227#define PFM_DEBUGGING 1
 228#ifdef PFM_DEBUGGING
 229#define DPRINT(a) \
 230	do { \
 231		if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
 232	} while (0)
 233
 234#define DPRINT_ovfl(a) \
 235	do { \
 236		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; } \
 237	} while (0)
 238#endif
 239
 240/*
 241 * 64-bit software counter structure
 242 *
 243 * the next_reset_type is applied to the next call to pfm_reset_regs()
 244 */
 245typedef struct {
 246	unsigned long	val;		/* virtual 64bit counter value */
 247	unsigned long	lval;		/* last reset value */
 248	unsigned long	long_reset;	/* reset value on sampling overflow */
 249	unsigned long	short_reset;    /* reset value on overflow */
 250	unsigned long	reset_pmds[4];  /* which other pmds to reset when this counter overflows */
 251	unsigned long	smpl_pmds[4];   /* which pmds are accessed when counter overflow */
 252	unsigned long	seed;		/* seed for random-number generator */
 253	unsigned long	mask;		/* mask for random-number generator */
 254	unsigned int 	flags;		/* notify/do not notify */
 255	unsigned long	eventid;	/* overflow event identifier */
 256} pfm_counter_t;
 257
 258/*
 259 * context flags
 260 */
 261typedef struct {
 262	unsigned int block:1;		/* when 1, task will blocked on user notifications */
 263	unsigned int system:1;		/* do system wide monitoring */
 264	unsigned int using_dbreg:1;	/* using range restrictions (debug registers) */
 265	unsigned int is_sampling:1;	/* true if using a custom format */
 266	unsigned int excl_idle:1;	/* exclude idle task in system wide session */
 267	unsigned int going_zombie:1;	/* context is zombie (MASKED+blocking) */
 268	unsigned int trap_reason:2;	/* reason for going into pfm_handle_work() */
 269	unsigned int no_msg:1;		/* no message sent on overflow */
 270	unsigned int can_restart:1;	/* allowed to issue a PFM_RESTART */
 271	unsigned int reserved:22;
 272} pfm_context_flags_t;
 273
 274#define PFM_TRAP_REASON_NONE		0x0	/* default value */
 275#define PFM_TRAP_REASON_BLOCK		0x1	/* we need to block on overflow */
 276#define PFM_TRAP_REASON_RESET		0x2	/* we need to reset PMDs */
 277
 278
 279/*
 280 * perfmon context: encapsulates all the state of a monitoring session
 281 */
 282
 283typedef struct pfm_context {
 284	spinlock_t		ctx_lock;		/* context protection */
 285
 286	pfm_context_flags_t	ctx_flags;		/* bitmask of flags  (block reason incl.) */
 287	unsigned int		ctx_state;		/* state: active/inactive (no bitfield) */
 288
 289	struct task_struct 	*ctx_task;		/* task to which context is attached */
 290
 291	unsigned long		ctx_ovfl_regs[4];	/* which registers overflowed (notification) */
 292
 293	struct completion	ctx_restart_done;  	/* use for blocking notification mode */
 294
 295	unsigned long		ctx_used_pmds[4];	/* bitmask of PMD used            */
 296	unsigned long		ctx_all_pmds[4];	/* bitmask of all accessible PMDs */
 297	unsigned long		ctx_reload_pmds[4];	/* bitmask of force reload PMD on ctxsw in */
 298
 299	unsigned long		ctx_all_pmcs[4];	/* bitmask of all accessible PMCs */
 300	unsigned long		ctx_reload_pmcs[4];	/* bitmask of force reload PMC on ctxsw in */
 301	unsigned long		ctx_used_monitors[4];	/* bitmask of monitor PMC being used */
 302
 303	unsigned long		ctx_pmcs[PFM_NUM_PMC_REGS];	/*  saved copies of PMC values */
 304
 305	unsigned int		ctx_used_ibrs[1];		/* bitmask of used IBR (speedup ctxsw in) */
 306	unsigned int		ctx_used_dbrs[1];		/* bitmask of used DBR (speedup ctxsw in) */
 307	unsigned long		ctx_dbrs[IA64_NUM_DBG_REGS];	/* DBR values (cache) when not loaded */
 308	unsigned long		ctx_ibrs[IA64_NUM_DBG_REGS];	/* IBR values (cache) when not loaded */
 309
 310	pfm_counter_t		ctx_pmds[PFM_NUM_PMD_REGS]; /* software state for PMDS */
 311
 312	unsigned long		th_pmcs[PFM_NUM_PMC_REGS];	/* PMC thread save state */
 313	unsigned long		th_pmds[PFM_NUM_PMD_REGS];	/* PMD thread save state */
 314
 315	unsigned long		ctx_saved_psr_up;	/* only contains psr.up value */
 316
 317	unsigned long		ctx_last_activation;	/* context last activation number for last_cpu */
 318	unsigned int		ctx_last_cpu;		/* CPU id of current or last CPU used (SMP only) */
 319	unsigned int		ctx_cpu;		/* cpu to which perfmon is applied (system wide) */
 320
 321	int			ctx_fd;			/* file descriptor used my this context */
 322	pfm_ovfl_arg_t		ctx_ovfl_arg;		/* argument to custom buffer format handler */
 323
 324	pfm_buffer_fmt_t	*ctx_buf_fmt;		/* buffer format callbacks */
 325	void			*ctx_smpl_hdr;		/* points to sampling buffer header kernel vaddr */
 326	unsigned long		ctx_smpl_size;		/* size of sampling buffer */
 327	void			*ctx_smpl_vaddr;	/* user level virtual address of smpl buffer */
 328
 329	wait_queue_head_t 	ctx_msgq_wait;
 330	pfm_msg_t		ctx_msgq[PFM_MAX_MSGS];
 331	int			ctx_msgq_head;
 332	int			ctx_msgq_tail;
 333	struct fasync_struct	*ctx_async_queue;
 334
 335	wait_queue_head_t 	ctx_zombieq;		/* termination cleanup wait queue */
 336} pfm_context_t;
 337
 338/*
 339 * magic number used to verify that structure is really
 340 * a perfmon context
 341 */
 342#define PFM_IS_FILE(f)		((f)->f_op == &pfm_file_ops)
 343
 344#define PFM_GET_CTX(t)	 	((pfm_context_t *)(t)->thread.pfm_context)
 345
 346#ifdef CONFIG_SMP
 347#define SET_LAST_CPU(ctx, v)	(ctx)->ctx_last_cpu = (v)
 348#define GET_LAST_CPU(ctx)	(ctx)->ctx_last_cpu
 349#else
 350#define SET_LAST_CPU(ctx, v)	do {} while(0)
 351#define GET_LAST_CPU(ctx)	do {} while(0)
 352#endif
 353
 354
 355#define ctx_fl_block		ctx_flags.block
 356#define ctx_fl_system		ctx_flags.system
 357#define ctx_fl_using_dbreg	ctx_flags.using_dbreg
 358#define ctx_fl_is_sampling	ctx_flags.is_sampling
 359#define ctx_fl_excl_idle	ctx_flags.excl_idle
 360#define ctx_fl_going_zombie	ctx_flags.going_zombie
 361#define ctx_fl_trap_reason	ctx_flags.trap_reason
 362#define ctx_fl_no_msg		ctx_flags.no_msg
 363#define ctx_fl_can_restart	ctx_flags.can_restart
 364
 365#define PFM_SET_WORK_PENDING(t, v)	do { (t)->thread.pfm_needs_checking = v; } while(0);
 366#define PFM_GET_WORK_PENDING(t)		(t)->thread.pfm_needs_checking
 367
 368/*
 369 * global information about all sessions
 370 * mostly used to synchronize between system wide and per-process
 371 */
 372typedef struct {
 373	spinlock_t		pfs_lock;		   /* lock the structure */
 374
 375	unsigned int		pfs_task_sessions;	   /* number of per task sessions */
 376	unsigned int		pfs_sys_sessions;	   /* number of per system wide sessions */
 377	unsigned int		pfs_sys_use_dbregs;	   /* incremented when a system wide session uses debug regs */
 378	unsigned int		pfs_ptrace_use_dbregs;	   /* incremented when a process uses debug regs */
 379	struct task_struct	*pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
 380} pfm_session_t;
 381
 382/*
 383 * information about a PMC or PMD.
 384 * dep_pmd[]: a bitmask of dependent PMD registers
 385 * dep_pmc[]: a bitmask of dependent PMC registers
 386 */
 387typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
 388typedef struct {
 389	unsigned int		type;
 390	int			pm_pos;
 391	unsigned long		default_value;	/* power-on default value */
 392	unsigned long		reserved_mask;	/* bitmask of reserved bits */
 393	pfm_reg_check_t		read_check;
 394	pfm_reg_check_t		write_check;
 395	unsigned long		dep_pmd[4];
 396	unsigned long		dep_pmc[4];
 397} pfm_reg_desc_t;
 398
 399/* assume cnum is a valid monitor */
 400#define PMC_PM(cnum, val)	(((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
 401
 402/*
 403 * This structure is initialized at boot time and contains
 404 * a description of the PMU main characteristics.
 405 *
 406 * If the probe function is defined, detection is based
 407 * on its return value: 
 408 * 	- 0 means recognized PMU
 409 * 	- anything else means not supported
 410 * When the probe function is not defined, then the pmu_family field
 411 * is used and it must match the host CPU family such that:
 412 * 	- cpu->family & config->pmu_family != 0
 413 */
 414typedef struct {
 415	unsigned long  ovfl_val;	/* overflow value for counters */
 416
 417	pfm_reg_desc_t *pmc_desc;	/* detailed PMC register dependencies descriptions */
 418	pfm_reg_desc_t *pmd_desc;	/* detailed PMD register dependencies descriptions */
 419
 420	unsigned int   num_pmcs;	/* number of PMCS: computed at init time */
 421	unsigned int   num_pmds;	/* number of PMDS: computed at init time */
 422	unsigned long  impl_pmcs[4];	/* bitmask of implemented PMCS */
 423	unsigned long  impl_pmds[4];	/* bitmask of implemented PMDS */
 424
 425	char	      *pmu_name;	/* PMU family name */
 426	unsigned int  pmu_family;	/* cpuid family pattern used to identify pmu */
 427	unsigned int  flags;		/* pmu specific flags */
 428	unsigned int  num_ibrs;		/* number of IBRS: computed at init time */
 429	unsigned int  num_dbrs;		/* number of DBRS: computed at init time */
 430	unsigned int  num_counters;	/* PMC/PMD counting pairs : computed at init time */
 431	int           (*probe)(void);   /* customized probe routine */
 432	unsigned int  use_rr_dbregs:1;	/* set if debug registers used for range restriction */
 433} pmu_config_t;
 434/*
 435 * PMU specific flags
 436 */
 437#define PFM_PMU_IRQ_RESEND	1	/* PMU needs explicit IRQ resend */
 438
 439/*
 440 * debug register related type definitions
 441 */
 442typedef struct {
 443	unsigned long ibr_mask:56;
 444	unsigned long ibr_plm:4;
 445	unsigned long ibr_ig:3;
 446	unsigned long ibr_x:1;
 447} ibr_mask_reg_t;
 448
 449typedef struct {
 450	unsigned long dbr_mask:56;
 451	unsigned long dbr_plm:4;
 452	unsigned long dbr_ig:2;
 453	unsigned long dbr_w:1;
 454	unsigned long dbr_r:1;
 455} dbr_mask_reg_t;
 456
 457typedef union {
 458	unsigned long  val;
 459	ibr_mask_reg_t ibr;
 460	dbr_mask_reg_t dbr;
 461} dbreg_t;
 462
 463
 464/*
 465 * perfmon command descriptions
 466 */
 467typedef struct {
 468	int		(*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
 469	char		*cmd_name;
 470	int		cmd_flags;
 471	unsigned int	cmd_narg;
 472	size_t		cmd_argsize;
 473	int		(*cmd_getsize)(void *arg, size_t *sz);
 474} pfm_cmd_desc_t;
 475
 476#define PFM_CMD_FD		0x01	/* command requires a file descriptor */
 477#define PFM_CMD_ARG_READ	0x02	/* command must read argument(s) */
 478#define PFM_CMD_ARG_RW		0x04	/* command must read/write argument(s) */
 479#define PFM_CMD_STOP		0x08	/* command does not work on zombie context */
 480
 481
 482#define PFM_CMD_NAME(cmd)	pfm_cmd_tab[(cmd)].cmd_name
 483#define PFM_CMD_READ_ARG(cmd)	(pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
 484#define PFM_CMD_RW_ARG(cmd)	(pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
 485#define PFM_CMD_USE_FD(cmd)	(pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
 486#define PFM_CMD_STOPPED(cmd)	(pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
 487
 488#define PFM_CMD_ARG_MANY	-1 /* cannot be zero */
 489
 490typedef struct {
 491	unsigned long pfm_spurious_ovfl_intr_count;	/* keep track of spurious ovfl interrupts */
 492	unsigned long pfm_replay_ovfl_intr_count;	/* keep track of replayed ovfl interrupts */
 493	unsigned long pfm_ovfl_intr_count; 		/* keep track of ovfl interrupts */
 494	unsigned long pfm_ovfl_intr_cycles;		/* cycles spent processing ovfl interrupts */
 495	unsigned long pfm_ovfl_intr_cycles_min;		/* min cycles spent processing ovfl interrupts */
 496	unsigned long pfm_ovfl_intr_cycles_max;		/* max cycles spent processing ovfl interrupts */
 497	unsigned long pfm_smpl_handler_calls;
 498	unsigned long pfm_smpl_handler_cycles;
 499	char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
 500} pfm_stats_t;
 501
 502/*
 503 * perfmon internal variables
 504 */
 505static pfm_stats_t		pfm_stats[NR_CPUS];
 506static pfm_session_t		pfm_sessions;	/* global sessions information */
 507
 508static DEFINE_SPINLOCK(pfm_alt_install_check);
 509static pfm_intr_handler_desc_t  *pfm_alt_intr_handler;
 510
 511static struct proc_dir_entry 	*perfmon_dir;
 512static pfm_uuid_t		pfm_null_uuid = {0,};
 513
 514static spinlock_t		pfm_buffer_fmt_lock;
 515static LIST_HEAD(pfm_buffer_fmt_list);
 516
 517static pmu_config_t		*pmu_conf;
 518
 519/* sysctl() controls */
 520pfm_sysctl_t pfm_sysctl;
 521EXPORT_SYMBOL(pfm_sysctl);
 522
 523static ctl_table pfm_ctl_table[]={
 524	{
 525		.procname	= "debug",
 526		.data		= &pfm_sysctl.debug,
 527		.maxlen		= sizeof(int),
 528		.mode		= 0666,
 529		.proc_handler	= proc_dointvec,
 530	},
 531	{
 532		.procname	= "debug_ovfl",
 533		.data		= &pfm_sysctl.debug_ovfl,
 534		.maxlen		= sizeof(int),
 535		.mode		= 0666,
 536		.proc_handler	= proc_dointvec,
 537	},
 538	{
 539		.procname	= "fastctxsw",
 540		.data		= &pfm_sysctl.fastctxsw,
 541		.maxlen		= sizeof(int),
 542		.mode		= 0600,
 543		.proc_handler	= proc_dointvec,
 544	},
 545	{
 546		.procname	= "expert_mode",
 547		.data		= &pfm_sysctl.expert_mode,
 548		.maxlen		= sizeof(int),
 549		.mode		= 0600,
 550		.proc_handler	= proc_dointvec,
 551	},
 552	{}
 553};
 554static ctl_table pfm_sysctl_dir[] = {
 555	{
 556		.procname	= "perfmon",
 557		.mode		= 0555,
 558		.child		= pfm_ctl_table,
 559	},
 560 	{}
 561};
 562static ctl_table pfm_sysctl_root[] = {
 563	{
 564		.procname	= "kernel",
 565		.mode		= 0555,
 566		.child		= pfm_sysctl_dir,
 567	},
 568 	{}
 569};
 570static struct ctl_table_header *pfm_sysctl_header;
 571
 572static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
 573
 574#define pfm_get_cpu_var(v)		__ia64_per_cpu_var(v)
 575#define pfm_get_cpu_data(a,b)		per_cpu(a, b)
 576
 577static inline void
 578pfm_put_task(struct task_struct *task)
 579{
 580	if (task != current) put_task_struct(task);
 581}
 582
 583static inline void
 584pfm_reserve_page(unsigned long a)
 585{
 586	SetPageReserved(vmalloc_to_page((void *)a));
 587}
 588static inline void
 589pfm_unreserve_page(unsigned long a)
 590{
 591	ClearPageReserved(vmalloc_to_page((void*)a));
 592}
 593
 594static inline unsigned long
 595pfm_protect_ctx_ctxsw(pfm_context_t *x)
 596{
 597	spin_lock(&(x)->ctx_lock);
 598	return 0UL;
 599}
 600
 601static inline void
 602pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
 603{
 604	spin_unlock(&(x)->ctx_lock);
 605}
 606
 607/* forward declaration */
 608static const struct dentry_operations pfmfs_dentry_operations;
 609
 610static struct dentry *
 611pfmfs_mount(struct file_system_type *fs_type, int flags, const char *dev_name, void *data)
 612{
 613	return mount_pseudo(fs_type, "pfm:", NULL, &pfmfs_dentry_operations,
 614			PFMFS_MAGIC);
 
 
 
 615}
 616
 617static struct file_system_type pfm_fs_type = {
 618	.name     = "pfmfs",
 619	.mount    = pfmfs_mount,
 620	.kill_sb  = kill_anon_super,
 621};
 
 622
 623DEFINE_PER_CPU(unsigned long, pfm_syst_info);
 624DEFINE_PER_CPU(struct task_struct *, pmu_owner);
 625DEFINE_PER_CPU(pfm_context_t  *, pmu_ctx);
 626DEFINE_PER_CPU(unsigned long, pmu_activation_number);
 627EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info);
 628
 629
 630/* forward declaration */
 631static const struct file_operations pfm_file_ops;
 632
 633/*
 634 * forward declarations
 635 */
 636#ifndef CONFIG_SMP
 637static void pfm_lazy_save_regs (struct task_struct *ta);
 638#endif
 639
 640void dump_pmu_state(const char *);
 641static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
 642
 643#include "perfmon_itanium.h"
 644#include "perfmon_mckinley.h"
 645#include "perfmon_montecito.h"
 646#include "perfmon_generic.h"
 647
 648static pmu_config_t *pmu_confs[]={
 649	&pmu_conf_mont,
 650	&pmu_conf_mck,
 651	&pmu_conf_ita,
 652	&pmu_conf_gen, /* must be last */
 653	NULL
 654};
 655
 656
 657static int pfm_end_notify_user(pfm_context_t *ctx);
 658
 659static inline void
 660pfm_clear_psr_pp(void)
 661{
 662	ia64_rsm(IA64_PSR_PP);
 663	ia64_srlz_i();
 664}
 665
 666static inline void
 667pfm_set_psr_pp(void)
 668{
 669	ia64_ssm(IA64_PSR_PP);
 670	ia64_srlz_i();
 671}
 672
 673static inline void
 674pfm_clear_psr_up(void)
 675{
 676	ia64_rsm(IA64_PSR_UP);
 677	ia64_srlz_i();
 678}
 679
 680static inline void
 681pfm_set_psr_up(void)
 682{
 683	ia64_ssm(IA64_PSR_UP);
 684	ia64_srlz_i();
 685}
 686
 687static inline unsigned long
 688pfm_get_psr(void)
 689{
 690	unsigned long tmp;
 691	tmp = ia64_getreg(_IA64_REG_PSR);
 692	ia64_srlz_i();
 693	return tmp;
 694}
 695
 696static inline void
 697pfm_set_psr_l(unsigned long val)
 698{
 699	ia64_setreg(_IA64_REG_PSR_L, val);
 700	ia64_srlz_i();
 701}
 702
 703static inline void
 704pfm_freeze_pmu(void)
 705{
 706	ia64_set_pmc(0,1UL);
 707	ia64_srlz_d();
 708}
 709
 710static inline void
 711pfm_unfreeze_pmu(void)
 712{
 713	ia64_set_pmc(0,0UL);
 714	ia64_srlz_d();
 715}
 716
 717static inline void
 718pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
 719{
 720	int i;
 721
 722	for (i=0; i < nibrs; i++) {
 723		ia64_set_ibr(i, ibrs[i]);
 724		ia64_dv_serialize_instruction();
 725	}
 726	ia64_srlz_i();
 727}
 728
 729static inline void
 730pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
 731{
 732	int i;
 733
 734	for (i=0; i < ndbrs; i++) {
 735		ia64_set_dbr(i, dbrs[i]);
 736		ia64_dv_serialize_data();
 737	}
 738	ia64_srlz_d();
 739}
 740
 741/*
 742 * PMD[i] must be a counter. no check is made
 743 */
 744static inline unsigned long
 745pfm_read_soft_counter(pfm_context_t *ctx, int i)
 746{
 747	return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
 748}
 749
 750/*
 751 * PMD[i] must be a counter. no check is made
 752 */
 753static inline void
 754pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
 755{
 756	unsigned long ovfl_val = pmu_conf->ovfl_val;
 757
 758	ctx->ctx_pmds[i].val = val  & ~ovfl_val;
 759	/*
 760	 * writing to unimplemented part is ignore, so we do not need to
 761	 * mask off top part
 762	 */
 763	ia64_set_pmd(i, val & ovfl_val);
 764}
 765
 766static pfm_msg_t *
 767pfm_get_new_msg(pfm_context_t *ctx)
 768{
 769	int idx, next;
 770
 771	next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
 772
 773	DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
 774	if (next == ctx->ctx_msgq_head) return NULL;
 775
 776 	idx = 	ctx->ctx_msgq_tail;
 777	ctx->ctx_msgq_tail = next;
 778
 779	DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
 780
 781	return ctx->ctx_msgq+idx;
 782}
 783
 784static pfm_msg_t *
 785pfm_get_next_msg(pfm_context_t *ctx)
 786{
 787	pfm_msg_t *msg;
 788
 789	DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
 790
 791	if (PFM_CTXQ_EMPTY(ctx)) return NULL;
 792
 793	/*
 794	 * get oldest message
 795	 */
 796	msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
 797
 798	/*
 799	 * and move forward
 800	 */
 801	ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
 802
 803	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));
 804
 805	return msg;
 806}
 807
 808static void
 809pfm_reset_msgq(pfm_context_t *ctx)
 810{
 811	ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
 812	DPRINT(("ctx=%p msgq reset\n", ctx));
 813}
 814
 815static void *
 816pfm_rvmalloc(unsigned long size)
 817{
 818	void *mem;
 819	unsigned long addr;
 820
 821	size = PAGE_ALIGN(size);
 822	mem  = vzalloc(size);
 823	if (mem) {
 824		//printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
 825		addr = (unsigned long)mem;
 826		while (size > 0) {
 827			pfm_reserve_page(addr);
 828			addr+=PAGE_SIZE;
 829			size-=PAGE_SIZE;
 830		}
 831	}
 832	return mem;
 833}
 834
 835static void
 836pfm_rvfree(void *mem, unsigned long size)
 837{
 838	unsigned long addr;
 839
 840	if (mem) {
 841		DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size));
 842		addr = (unsigned long) mem;
 843		while ((long) size > 0) {
 844			pfm_unreserve_page(addr);
 845			addr+=PAGE_SIZE;
 846			size-=PAGE_SIZE;
 847		}
 848		vfree(mem);
 849	}
 850	return;
 851}
 852
 853static pfm_context_t *
 854pfm_context_alloc(int ctx_flags)
 855{
 856	pfm_context_t *ctx;
 857
 858	/* 
 859	 * allocate context descriptor 
 860	 * must be able to free with interrupts disabled
 861	 */
 862	ctx = kzalloc(sizeof(pfm_context_t), GFP_KERNEL);
 863	if (ctx) {
 864		DPRINT(("alloc ctx @%p\n", ctx));
 865
 866		/*
 867		 * init context protection lock
 868		 */
 869		spin_lock_init(&ctx->ctx_lock);
 870
 871		/*
 872		 * context is unloaded
 873		 */
 874		ctx->ctx_state = PFM_CTX_UNLOADED;
 875
 876		/*
 877		 * initialization of context's flags
 878		 */
 879		ctx->ctx_fl_block       = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
 880		ctx->ctx_fl_system      = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
 881		ctx->ctx_fl_no_msg      = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
 882		/*
 883		 * will move to set properties
 884		 * ctx->ctx_fl_excl_idle   = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
 885		 */
 886
 887		/*
 888		 * init restart semaphore to locked
 889		 */
 890		init_completion(&ctx->ctx_restart_done);
 891
 892		/*
 893		 * activation is used in SMP only
 894		 */
 895		ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
 896		SET_LAST_CPU(ctx, -1);
 897
 898		/*
 899		 * initialize notification message queue
 900		 */
 901		ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
 902		init_waitqueue_head(&ctx->ctx_msgq_wait);
 903		init_waitqueue_head(&ctx->ctx_zombieq);
 904
 905	}
 906	return ctx;
 907}
 908
 909static void
 910pfm_context_free(pfm_context_t *ctx)
 911{
 912	if (ctx) {
 913		DPRINT(("free ctx @%p\n", ctx));
 914		kfree(ctx);
 915	}
 916}
 917
 918static void
 919pfm_mask_monitoring(struct task_struct *task)
 920{
 921	pfm_context_t *ctx = PFM_GET_CTX(task);
 922	unsigned long mask, val, ovfl_mask;
 923	int i;
 924
 925	DPRINT_ovfl(("masking monitoring for [%d]\n", task_pid_nr(task)));
 926
 927	ovfl_mask = pmu_conf->ovfl_val;
 928	/*
 929	 * monitoring can only be masked as a result of a valid
 930	 * counter overflow. In UP, it means that the PMU still
 931	 * has an owner. Note that the owner can be different
 932	 * from the current task. However the PMU state belongs
 933	 * to the owner.
 934	 * In SMP, a valid overflow only happens when task is
 935	 * current. Therefore if we come here, we know that
 936	 * the PMU state belongs to the current task, therefore
 937	 * we can access the live registers.
 938	 *
 939	 * So in both cases, the live register contains the owner's
 940	 * state. We can ONLY touch the PMU registers and NOT the PSR.
 941	 *
 942	 * As a consequence to this call, the ctx->th_pmds[] array
 943	 * contains stale information which must be ignored
 944	 * when context is reloaded AND monitoring is active (see
 945	 * pfm_restart).
 946	 */
 947	mask = ctx->ctx_used_pmds[0];
 948	for (i = 0; mask; i++, mask>>=1) {
 949		/* skip non used pmds */
 950		if ((mask & 0x1) == 0) continue;
 951		val = ia64_get_pmd(i);
 952
 953		if (PMD_IS_COUNTING(i)) {
 954			/*
 955		 	 * we rebuild the full 64 bit value of the counter
 956		 	 */
 957			ctx->ctx_pmds[i].val += (val & ovfl_mask);
 958		} else {
 959			ctx->ctx_pmds[i].val = val;
 960		}
 961		DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
 962			i,
 963			ctx->ctx_pmds[i].val,
 964			val & ovfl_mask));
 965	}
 966	/*
 967	 * mask monitoring by setting the privilege level to 0
 968	 * we cannot use psr.pp/psr.up for this, it is controlled by
 969	 * the user
 970	 *
 971	 * if task is current, modify actual registers, otherwise modify
 972	 * thread save state, i.e., what will be restored in pfm_load_regs()
 973	 */
 974	mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
 975	for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
 976		if ((mask & 0x1) == 0UL) continue;
 977		ia64_set_pmc(i, ctx->th_pmcs[i] & ~0xfUL);
 978		ctx->th_pmcs[i] &= ~0xfUL;
 979		DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
 980	}
 981	/*
 982	 * make all of this visible
 983	 */
 984	ia64_srlz_d();
 985}
 986
 987/*
 988 * must always be done with task == current
 989 *
 990 * context must be in MASKED state when calling
 991 */
 992static void
 993pfm_restore_monitoring(struct task_struct *task)
 994{
 995	pfm_context_t *ctx = PFM_GET_CTX(task);
 996	unsigned long mask, ovfl_mask;
 997	unsigned long psr, val;
 998	int i, is_system;
 999
1000	is_system = ctx->ctx_fl_system;
1001	ovfl_mask = pmu_conf->ovfl_val;
1002
1003	if (task != current) {
1004		printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task_pid_nr(task), task_pid_nr(current));
1005		return;
1006	}
1007	if (ctx->ctx_state != PFM_CTX_MASKED) {
1008		printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
1009			task_pid_nr(task), task_pid_nr(current), ctx->ctx_state);
1010		return;
1011	}
1012	psr = pfm_get_psr();
1013	/*
1014	 * monitoring is masked via the PMC.
1015	 * As we restore their value, we do not want each counter to
1016	 * restart right away. We stop monitoring using the PSR,
1017	 * restore the PMC (and PMD) and then re-establish the psr
1018	 * as it was. Note that there can be no pending overflow at
1019	 * this point, because monitoring was MASKED.
1020	 *
1021	 * system-wide session are pinned and self-monitoring
1022	 */
1023	if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1024		/* disable dcr pp */
1025		ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
1026		pfm_clear_psr_pp();
1027	} else {
1028		pfm_clear_psr_up();
1029	}
1030	/*
1031	 * first, we restore the PMD
1032	 */
1033	mask = ctx->ctx_used_pmds[0];
1034	for (i = 0; mask; i++, mask>>=1) {
1035		/* skip non used pmds */
1036		if ((mask & 0x1) == 0) continue;
1037
1038		if (PMD_IS_COUNTING(i)) {
1039			/*
1040			 * we split the 64bit value according to
1041			 * counter width
1042			 */
1043			val = ctx->ctx_pmds[i].val & ovfl_mask;
1044			ctx->ctx_pmds[i].val &= ~ovfl_mask;
1045		} else {
1046			val = ctx->ctx_pmds[i].val;
1047		}
1048		ia64_set_pmd(i, val);
1049
1050		DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1051			i,
1052			ctx->ctx_pmds[i].val,
1053			val));
1054	}
1055	/*
1056	 * restore the PMCs
1057	 */
1058	mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1059	for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1060		if ((mask & 0x1) == 0UL) continue;
1061		ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1062		ia64_set_pmc(i, ctx->th_pmcs[i]);
1063		DPRINT(("[%d] pmc[%d]=0x%lx\n",
1064					task_pid_nr(task), i, ctx->th_pmcs[i]));
1065	}
1066	ia64_srlz_d();
1067
1068	/*
1069	 * must restore DBR/IBR because could be modified while masked
1070	 * XXX: need to optimize 
1071	 */
1072	if (ctx->ctx_fl_using_dbreg) {
1073		pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1074		pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1075	}
1076
1077	/*
1078	 * now restore PSR
1079	 */
1080	if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1081		/* enable dcr pp */
1082		ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1083		ia64_srlz_i();
1084	}
1085	pfm_set_psr_l(psr);
1086}
1087
1088static inline void
1089pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1090{
1091	int i;
1092
1093	ia64_srlz_d();
1094
1095	for (i=0; mask; i++, mask>>=1) {
1096		if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1097	}
1098}
1099
1100/*
1101 * reload from thread state (used for ctxw only)
1102 */
1103static inline void
1104pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1105{
1106	int i;
1107	unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1108
1109	for (i=0; mask; i++, mask>>=1) {
1110		if ((mask & 0x1) == 0) continue;
1111		val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1112		ia64_set_pmd(i, val);
1113	}
1114	ia64_srlz_d();
1115}
1116
1117/*
1118 * propagate PMD from context to thread-state
1119 */
1120static inline void
1121pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1122{
1123	unsigned long ovfl_val = pmu_conf->ovfl_val;
1124	unsigned long mask = ctx->ctx_all_pmds[0];
1125	unsigned long val;
1126	int i;
1127
1128	DPRINT(("mask=0x%lx\n", mask));
1129
1130	for (i=0; mask; i++, mask>>=1) {
1131
1132		val = ctx->ctx_pmds[i].val;
1133
1134		/*
1135		 * We break up the 64 bit value into 2 pieces
1136		 * the lower bits go to the machine state in the
1137		 * thread (will be reloaded on ctxsw in).
1138		 * The upper part stays in the soft-counter.
1139		 */
1140		if (PMD_IS_COUNTING(i)) {
1141			ctx->ctx_pmds[i].val = val & ~ovfl_val;
1142			 val &= ovfl_val;
1143		}
1144		ctx->th_pmds[i] = val;
1145
1146		DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1147			i,
1148			ctx->th_pmds[i],
1149			ctx->ctx_pmds[i].val));
1150	}
1151}
1152
1153/*
1154 * propagate PMC from context to thread-state
1155 */
1156static inline void
1157pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1158{
1159	unsigned long mask = ctx->ctx_all_pmcs[0];
1160	int i;
1161
1162	DPRINT(("mask=0x%lx\n", mask));
1163
1164	for (i=0; mask; i++, mask>>=1) {
1165		/* masking 0 with ovfl_val yields 0 */
1166		ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1167		DPRINT(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
1168	}
1169}
1170
1171
1172
1173static inline void
1174pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1175{
1176	int i;
1177
1178	for (i=0; mask; i++, mask>>=1) {
1179		if ((mask & 0x1) == 0) continue;
1180		ia64_set_pmc(i, pmcs[i]);
1181	}
1182	ia64_srlz_d();
1183}
1184
1185static inline int
1186pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1187{
1188	return memcmp(a, b, sizeof(pfm_uuid_t));
1189}
1190
1191static inline int
1192pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1193{
1194	int ret = 0;
1195	if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1196	return ret;
1197}
1198
1199static inline int
1200pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1201{
1202	int ret = 0;
1203	if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1204	return ret;
1205}
1206
1207
1208static inline int
1209pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1210		     int cpu, void *arg)
1211{
1212	int ret = 0;
1213	if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1214	return ret;
1215}
1216
1217static inline int
1218pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1219		     int cpu, void *arg)
1220{
1221	int ret = 0;
1222	if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1223	return ret;
1224}
1225
1226static inline int
1227pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1228{
1229	int ret = 0;
1230	if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1231	return ret;
1232}
1233
1234static inline int
1235pfm_buf_fmt_restart_active(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1236{
1237	int ret = 0;
1238	if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1239	return ret;
1240}
1241
1242static pfm_buffer_fmt_t *
1243__pfm_find_buffer_fmt(pfm_uuid_t uuid)
1244{
1245	struct list_head * pos;
1246	pfm_buffer_fmt_t * entry;
1247
1248	list_for_each(pos, &pfm_buffer_fmt_list) {
1249		entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1250		if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1251			return entry;
1252	}
1253	return NULL;
1254}
1255 
1256/*
1257 * find a buffer format based on its uuid
1258 */
1259static pfm_buffer_fmt_t *
1260pfm_find_buffer_fmt(pfm_uuid_t uuid)
1261{
1262	pfm_buffer_fmt_t * fmt;
1263	spin_lock(&pfm_buffer_fmt_lock);
1264	fmt = __pfm_find_buffer_fmt(uuid);
1265	spin_unlock(&pfm_buffer_fmt_lock);
1266	return fmt;
1267}
1268 
1269int
1270pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1271{
1272	int ret = 0;
1273
1274	/* some sanity checks */
1275	if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1276
1277	/* we need at least a handler */
1278	if (fmt->fmt_handler == NULL) return -EINVAL;
1279
1280	/*
1281	 * XXX: need check validity of fmt_arg_size
1282	 */
1283
1284	spin_lock(&pfm_buffer_fmt_lock);
1285
1286	if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1287		printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1288		ret = -EBUSY;
1289		goto out;
1290	} 
1291	list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1292	printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1293
1294out:
1295	spin_unlock(&pfm_buffer_fmt_lock);
1296 	return ret;
1297}
1298EXPORT_SYMBOL(pfm_register_buffer_fmt);
1299
1300int
1301pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1302{
1303	pfm_buffer_fmt_t *fmt;
1304	int ret = 0;
1305
1306	spin_lock(&pfm_buffer_fmt_lock);
1307
1308	fmt = __pfm_find_buffer_fmt(uuid);
1309	if (!fmt) {
1310		printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1311		ret = -EINVAL;
1312		goto out;
1313	}
1314	list_del_init(&fmt->fmt_list);
1315	printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1316
1317out:
1318	spin_unlock(&pfm_buffer_fmt_lock);
1319	return ret;
1320
1321}
1322EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1323
1324extern void update_pal_halt_status(int);
1325
1326static int
1327pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1328{
1329	unsigned long flags;
1330	/*
1331	 * validity checks on cpu_mask have been done upstream
1332	 */
1333	LOCK_PFS(flags);
1334
1335	DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1336		pfm_sessions.pfs_sys_sessions,
1337		pfm_sessions.pfs_task_sessions,
1338		pfm_sessions.pfs_sys_use_dbregs,
1339		is_syswide,
1340		cpu));
1341
1342	if (is_syswide) {
1343		/*
1344		 * cannot mix system wide and per-task sessions
1345		 */
1346		if (pfm_sessions.pfs_task_sessions > 0UL) {
1347			DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1348			  	pfm_sessions.pfs_task_sessions));
1349			goto abort;
1350		}
1351
1352		if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1353
1354		DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1355
1356		pfm_sessions.pfs_sys_session[cpu] = task;
1357
1358		pfm_sessions.pfs_sys_sessions++ ;
1359
1360	} else {
1361		if (pfm_sessions.pfs_sys_sessions) goto abort;
1362		pfm_sessions.pfs_task_sessions++;
1363	}
1364
1365	DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1366		pfm_sessions.pfs_sys_sessions,
1367		pfm_sessions.pfs_task_sessions,
1368		pfm_sessions.pfs_sys_use_dbregs,
1369		is_syswide,
1370		cpu));
1371
1372	/*
1373	 * disable default_idle() to go to PAL_HALT
1374	 */
1375	update_pal_halt_status(0);
1376
1377	UNLOCK_PFS(flags);
1378
1379	return 0;
1380
1381error_conflict:
1382	DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1383  		task_pid_nr(pfm_sessions.pfs_sys_session[cpu]),
1384		cpu));
1385abort:
1386	UNLOCK_PFS(flags);
1387
1388	return -EBUSY;
1389
1390}
1391
1392static int
1393pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1394{
1395	unsigned long flags;
1396	/*
1397	 * validity checks on cpu_mask have been done upstream
1398	 */
1399	LOCK_PFS(flags);
1400
1401	DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1402		pfm_sessions.pfs_sys_sessions,
1403		pfm_sessions.pfs_task_sessions,
1404		pfm_sessions.pfs_sys_use_dbregs,
1405		is_syswide,
1406		cpu));
1407
1408
1409	if (is_syswide) {
1410		pfm_sessions.pfs_sys_session[cpu] = NULL;
1411		/*
1412		 * would not work with perfmon+more than one bit in cpu_mask
1413		 */
1414		if (ctx && ctx->ctx_fl_using_dbreg) {
1415			if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1416				printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1417			} else {
1418				pfm_sessions.pfs_sys_use_dbregs--;
1419			}
1420		}
1421		pfm_sessions.pfs_sys_sessions--;
1422	} else {
1423		pfm_sessions.pfs_task_sessions--;
1424	}
1425	DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1426		pfm_sessions.pfs_sys_sessions,
1427		pfm_sessions.pfs_task_sessions,
1428		pfm_sessions.pfs_sys_use_dbregs,
1429		is_syswide,
1430		cpu));
1431
1432	/*
1433	 * if possible, enable default_idle() to go into PAL_HALT
1434	 */
1435	if (pfm_sessions.pfs_task_sessions == 0 && pfm_sessions.pfs_sys_sessions == 0)
1436		update_pal_halt_status(1);
1437
1438	UNLOCK_PFS(flags);
1439
1440	return 0;
1441}
1442
1443/*
1444 * removes virtual mapping of the sampling buffer.
1445 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1446 * a PROTECT_CTX() section.
1447 */
1448static int
1449pfm_remove_smpl_mapping(void *vaddr, unsigned long size)
1450{
1451	struct task_struct *task = current;
1452	int r;
1453
1454	/* sanity checks */
1455	if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1456		printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task_pid_nr(task), task->mm);
1457		return -EINVAL;
1458	}
1459
1460	DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1461
1462	/*
1463	 * does the actual unmapping
1464	 */
1465	r = vm_munmap((unsigned long)vaddr, size);
1466
1467	if (r !=0) {
1468		printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task_pid_nr(task), vaddr, size);
1469	}
1470
1471	DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1472
1473	return 0;
1474}
1475
1476/*
1477 * free actual physical storage used by sampling buffer
1478 */
1479#if 0
1480static int
1481pfm_free_smpl_buffer(pfm_context_t *ctx)
1482{
1483	pfm_buffer_fmt_t *fmt;
1484
1485	if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1486
1487	/*
1488	 * we won't use the buffer format anymore
1489	 */
1490	fmt = ctx->ctx_buf_fmt;
1491
1492	DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1493		ctx->ctx_smpl_hdr,
1494		ctx->ctx_smpl_size,
1495		ctx->ctx_smpl_vaddr));
1496
1497	pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1498
1499	/*
1500	 * free the buffer
1501	 */
1502	pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);
1503
1504	ctx->ctx_smpl_hdr  = NULL;
1505	ctx->ctx_smpl_size = 0UL;
1506
1507	return 0;
1508
1509invalid_free:
1510	printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", task_pid_nr(current));
1511	return -EINVAL;
1512}
1513#endif
1514
1515static inline void
1516pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1517{
1518	if (fmt == NULL) return;
1519
1520	pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1521
1522}
1523
1524/*
1525 * pfmfs should _never_ be mounted by userland - too much of security hassle,
1526 * no real gain from having the whole whorehouse mounted. So we don't need
1527 * any operations on the root directory. However, we need a non-trivial
1528 * d_name - pfm: will go nicely and kill the special-casing in procfs.
1529 */
1530static struct vfsmount *pfmfs_mnt __read_mostly;
1531
1532static int __init
1533init_pfm_fs(void)
1534{
1535	int err = register_filesystem(&pfm_fs_type);
1536	if (!err) {
1537		pfmfs_mnt = kern_mount(&pfm_fs_type);
1538		err = PTR_ERR(pfmfs_mnt);
1539		if (IS_ERR(pfmfs_mnt))
1540			unregister_filesystem(&pfm_fs_type);
1541		else
1542			err = 0;
1543	}
1544	return err;
1545}
1546
1547static ssize_t
1548pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
1549{
1550	pfm_context_t *ctx;
1551	pfm_msg_t *msg;
1552	ssize_t ret;
1553	unsigned long flags;
1554  	DECLARE_WAITQUEUE(wait, current);
1555	if (PFM_IS_FILE(filp) == 0) {
1556		printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1557		return -EINVAL;
1558	}
1559
1560	ctx = filp->private_data;
1561	if (ctx == NULL) {
1562		printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", task_pid_nr(current));
1563		return -EINVAL;
1564	}
1565
1566	/*
1567	 * check even when there is no message
1568	 */
1569	if (size < sizeof(pfm_msg_t)) {
1570		DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1571		return -EINVAL;
1572	}
1573
1574	PROTECT_CTX(ctx, flags);
1575
1576  	/*
1577	 * put ourselves on the wait queue
1578	 */
1579  	add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1580
1581
1582  	for(;;) {
1583		/*
1584		 * check wait queue
1585		 */
1586
1587  		set_current_state(TASK_INTERRUPTIBLE);
1588
1589		DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1590
1591		ret = 0;
1592		if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1593
1594		UNPROTECT_CTX(ctx, flags);
1595
1596		/*
1597		 * check non-blocking read
1598		 */
1599      		ret = -EAGAIN;
1600		if(filp->f_flags & O_NONBLOCK) break;
1601
1602		/*
1603		 * check pending signals
1604		 */
1605		if(signal_pending(current)) {
1606			ret = -EINTR;
1607			break;
1608		}
1609      		/*
1610		 * no message, so wait
1611		 */
1612      		schedule();
1613
1614		PROTECT_CTX(ctx, flags);
1615	}
1616	DPRINT(("[%d] back to running ret=%ld\n", task_pid_nr(current), ret));
1617  	set_current_state(TASK_RUNNING);
1618	remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1619
1620	if (ret < 0) goto abort;
1621
1622	ret = -EINVAL;
1623	msg = pfm_get_next_msg(ctx);
1624	if (msg == NULL) {
1625		printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, task_pid_nr(current));
1626		goto abort_locked;
1627	}
1628
1629	DPRINT(("fd=%d type=%d\n", msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1630
1631	ret = -EFAULT;
1632  	if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1633
1634abort_locked:
1635	UNPROTECT_CTX(ctx, flags);
1636abort:
1637	return ret;
1638}
1639
1640static ssize_t
1641pfm_write(struct file *file, const char __user *ubuf,
1642			  size_t size, loff_t *ppos)
1643{
1644	DPRINT(("pfm_write called\n"));
1645	return -EINVAL;
1646}
1647
1648static unsigned int
1649pfm_poll(struct file *filp, poll_table * wait)
1650{
1651	pfm_context_t *ctx;
1652	unsigned long flags;
1653	unsigned int mask = 0;
1654
1655	if (PFM_IS_FILE(filp) == 0) {
1656		printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1657		return 0;
1658	}
1659
1660	ctx = filp->private_data;
1661	if (ctx == NULL) {
1662		printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", task_pid_nr(current));
1663		return 0;
1664	}
1665
1666
1667	DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1668
1669	poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1670
1671	PROTECT_CTX(ctx, flags);
1672
1673	if (PFM_CTXQ_EMPTY(ctx) == 0)
1674		mask =  POLLIN | POLLRDNORM;
1675
1676	UNPROTECT_CTX(ctx, flags);
1677
1678	DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1679
1680	return mask;
1681}
1682
1683static long
1684pfm_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1685{
1686	DPRINT(("pfm_ioctl called\n"));
1687	return -EINVAL;
1688}
1689
1690/*
1691 * interrupt cannot be masked when coming here
1692 */
1693static inline int
1694pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1695{
1696	int ret;
1697
1698	ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1699
1700	DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1701		task_pid_nr(current),
1702		fd,
1703		on,
1704		ctx->ctx_async_queue, ret));
1705
1706	return ret;
1707}
1708
1709static int
1710pfm_fasync(int fd, struct file *filp, int on)
1711{
1712	pfm_context_t *ctx;
1713	int ret;
1714
1715	if (PFM_IS_FILE(filp) == 0) {
1716		printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", task_pid_nr(current));
1717		return -EBADF;
1718	}
1719
1720	ctx = filp->private_data;
1721	if (ctx == NULL) {
1722		printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", task_pid_nr(current));
1723		return -EBADF;
1724	}
1725	/*
1726	 * we cannot mask interrupts during this call because this may
1727	 * may go to sleep if memory is not readily avalaible.
1728	 *
1729	 * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1730	 * done in caller. Serialization of this function is ensured by caller.
1731	 */
1732	ret = pfm_do_fasync(fd, filp, ctx, on);
1733
1734
1735	DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1736		fd,
1737		on,
1738		ctx->ctx_async_queue, ret));
1739
1740	return ret;
1741}
1742
1743#ifdef CONFIG_SMP
1744/*
1745 * this function is exclusively called from pfm_close().
1746 * The context is not protected at that time, nor are interrupts
1747 * on the remote CPU. That's necessary to avoid deadlocks.
1748 */
1749static void
1750pfm_syswide_force_stop(void *info)
1751{
1752	pfm_context_t   *ctx = (pfm_context_t *)info;
1753	struct pt_regs *regs = task_pt_regs(current);
1754	struct task_struct *owner;
1755	unsigned long flags;
1756	int ret;
1757
1758	if (ctx->ctx_cpu != smp_processor_id()) {
1759		printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d  but on CPU%d\n",
1760			ctx->ctx_cpu,
1761			smp_processor_id());
1762		return;
1763	}
1764	owner = GET_PMU_OWNER();
1765	if (owner != ctx->ctx_task) {
1766		printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1767			smp_processor_id(),
1768			task_pid_nr(owner), task_pid_nr(ctx->ctx_task));
1769		return;
1770	}
1771	if (GET_PMU_CTX() != ctx) {
1772		printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1773			smp_processor_id(),
1774			GET_PMU_CTX(), ctx);
1775		return;
1776	}
1777
1778	DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), task_pid_nr(ctx->ctx_task)));
1779	/*
1780	 * the context is already protected in pfm_close(), we simply
1781	 * need to mask interrupts to avoid a PMU interrupt race on
1782	 * this CPU
1783	 */
1784	local_irq_save(flags);
1785
1786	ret = pfm_context_unload(ctx, NULL, 0, regs);
1787	if (ret) {
1788		DPRINT(("context_unload returned %d\n", ret));
1789	}
1790
1791	/*
1792	 * unmask interrupts, PMU interrupts are now spurious here
1793	 */
1794	local_irq_restore(flags);
1795}
1796
1797static void
1798pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1799{
1800	int ret;
1801
1802	DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1803	ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 1);
1804	DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1805}
1806#endif /* CONFIG_SMP */
1807
1808/*
1809 * called for each close(). Partially free resources.
1810 * When caller is self-monitoring, the context is unloaded.
1811 */
1812static int
1813pfm_flush(struct file *filp, fl_owner_t id)
1814{
1815	pfm_context_t *ctx;
1816	struct task_struct *task;
1817	struct pt_regs *regs;
1818	unsigned long flags;
1819	unsigned long smpl_buf_size = 0UL;
1820	void *smpl_buf_vaddr = NULL;
1821	int state, is_system;
1822
1823	if (PFM_IS_FILE(filp) == 0) {
1824		DPRINT(("bad magic for\n"));
1825		return -EBADF;
1826	}
1827
1828	ctx = filp->private_data;
1829	if (ctx == NULL) {
1830		printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", task_pid_nr(current));
1831		return -EBADF;
1832	}
1833
1834	/*
1835	 * remove our file from the async queue, if we use this mode.
1836	 * This can be done without the context being protected. We come
1837	 * here when the context has become unreachable by other tasks.
1838	 *
1839	 * We may still have active monitoring at this point and we may
1840	 * end up in pfm_overflow_handler(). However, fasync_helper()
1841	 * operates with interrupts disabled and it cleans up the
1842	 * queue. If the PMU handler is called prior to entering
1843	 * fasync_helper() then it will send a signal. If it is
1844	 * invoked after, it will find an empty queue and no
1845	 * signal will be sent. In both case, we are safe
1846	 */
1847	PROTECT_CTX(ctx, flags);
1848
1849	state     = ctx->ctx_state;
1850	is_system = ctx->ctx_fl_system;
1851
1852	task = PFM_CTX_TASK(ctx);
1853	regs = task_pt_regs(task);
1854
1855	DPRINT(("ctx_state=%d is_current=%d\n",
1856		state,
1857		task == current ? 1 : 0));
1858
1859	/*
1860	 * if state == UNLOADED, then task is NULL
1861	 */
1862
1863	/*
1864	 * we must stop and unload because we are losing access to the context.
1865	 */
1866	if (task == current) {
1867#ifdef CONFIG_SMP
1868		/*
1869		 * the task IS the owner but it migrated to another CPU: that's bad
1870		 * but we must handle this cleanly. Unfortunately, the kernel does
1871		 * not provide a mechanism to block migration (while the context is loaded).
1872		 *
1873		 * We need to release the resource on the ORIGINAL cpu.
1874		 */
1875		if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1876
1877			DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1878			/*
1879			 * keep context protected but unmask interrupt for IPI
1880			 */
1881			local_irq_restore(flags);
1882
1883			pfm_syswide_cleanup_other_cpu(ctx);
1884
1885			/*
1886			 * restore interrupt masking
1887			 */
1888			local_irq_save(flags);
1889
1890			/*
1891			 * context is unloaded at this point
1892			 */
1893		} else
1894#endif /* CONFIG_SMP */
1895		{
1896
1897			DPRINT(("forcing unload\n"));
1898			/*
1899		 	* stop and unload, returning with state UNLOADED
1900		 	* and session unreserved.
1901		 	*/
1902			pfm_context_unload(ctx, NULL, 0, regs);
1903
1904			DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1905		}
1906	}
1907
1908	/*
1909	 * remove virtual mapping, if any, for the calling task.
1910	 * cannot reset ctx field until last user is calling close().
1911	 *
1912	 * ctx_smpl_vaddr must never be cleared because it is needed
1913	 * by every task with access to the context
1914	 *
1915	 * When called from do_exit(), the mm context is gone already, therefore
1916	 * mm is NULL, i.e., the VMA is already gone  and we do not have to
1917	 * do anything here
1918	 */
1919	if (ctx->ctx_smpl_vaddr && current->mm) {
1920		smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1921		smpl_buf_size  = ctx->ctx_smpl_size;
1922	}
1923
1924	UNPROTECT_CTX(ctx, flags);
1925
1926	/*
1927	 * if there was a mapping, then we systematically remove it
1928	 * at this point. Cannot be done inside critical section
1929	 * because some VM function reenables interrupts.
1930	 *
1931	 */
1932	if (smpl_buf_vaddr) pfm_remove_smpl_mapping(smpl_buf_vaddr, smpl_buf_size);
1933
1934	return 0;
1935}
1936/*
1937 * called either on explicit close() or from exit_files(). 
1938 * Only the LAST user of the file gets to this point, i.e., it is
1939 * called only ONCE.
1940 *
1941 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero 
1942 * (fput()),i.e, last task to access the file. Nobody else can access the 
1943 * file at this point.
1944 *
1945 * When called from exit_files(), the VMA has been freed because exit_mm()
1946 * is executed before exit_files().
1947 *
1948 * When called from exit_files(), the current task is not yet ZOMBIE but we
1949 * flush the PMU state to the context. 
1950 */
1951static int
1952pfm_close(struct inode *inode, struct file *filp)
1953{
1954	pfm_context_t *ctx;
1955	struct task_struct *task;
1956	struct pt_regs *regs;
1957  	DECLARE_WAITQUEUE(wait, current);
1958	unsigned long flags;
1959	unsigned long smpl_buf_size = 0UL;
1960	void *smpl_buf_addr = NULL;
1961	int free_possible = 1;
1962	int state, is_system;
1963
1964	DPRINT(("pfm_close called private=%p\n", filp->private_data));
1965
1966	if (PFM_IS_FILE(filp) == 0) {
1967		DPRINT(("bad magic\n"));
1968		return -EBADF;
1969	}
1970	
1971	ctx = filp->private_data;
1972	if (ctx == NULL) {
1973		printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", task_pid_nr(current));
1974		return -EBADF;
1975	}
1976
1977	PROTECT_CTX(ctx, flags);
1978
1979	state     = ctx->ctx_state;
1980	is_system = ctx->ctx_fl_system;
1981
1982	task = PFM_CTX_TASK(ctx);
1983	regs = task_pt_regs(task);
1984
1985	DPRINT(("ctx_state=%d is_current=%d\n", 
1986		state,
1987		task == current ? 1 : 0));
1988
1989	/*
1990	 * if task == current, then pfm_flush() unloaded the context
1991	 */
1992	if (state == PFM_CTX_UNLOADED) goto doit;
1993
1994	/*
1995	 * context is loaded/masked and task != current, we need to
1996	 * either force an unload or go zombie
1997	 */
1998
1999	/*
2000	 * The task is currently blocked or will block after an overflow.
2001	 * we must force it to wakeup to get out of the
2002	 * MASKED state and transition to the unloaded state by itself.
2003	 *
2004	 * This situation is only possible for per-task mode
2005	 */
2006	if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
2007
2008		/*
2009		 * set a "partial" zombie state to be checked
2010		 * upon return from down() in pfm_handle_work().
2011		 *
2012		 * We cannot use the ZOMBIE state, because it is checked
2013		 * by pfm_load_regs() which is called upon wakeup from down().
2014		 * In such case, it would free the context and then we would
2015		 * return to pfm_handle_work() which would access the
2016		 * stale context. Instead, we set a flag invisible to pfm_load_regs()
2017		 * but visible to pfm_handle_work().
2018		 *
2019		 * For some window of time, we have a zombie context with
2020		 * ctx_state = MASKED  and not ZOMBIE
2021		 */
2022		ctx->ctx_fl_going_zombie = 1;
2023
2024		/*
2025		 * force task to wake up from MASKED state
2026		 */
2027		complete(&ctx->ctx_restart_done);
2028
2029		DPRINT(("waking up ctx_state=%d\n", state));
2030
2031		/*
2032		 * put ourself to sleep waiting for the other
2033		 * task to report completion
2034		 *
2035		 * the context is protected by mutex, therefore there
2036		 * is no risk of being notified of completion before
2037		 * begin actually on the waitq.
2038		 */
2039  		set_current_state(TASK_INTERRUPTIBLE);
2040  		add_wait_queue(&ctx->ctx_zombieq, &wait);
2041
2042		UNPROTECT_CTX(ctx, flags);
2043
2044		/*
2045		 * XXX: check for signals :
2046		 * 	- ok for explicit close
2047		 * 	- not ok when coming from exit_files()
2048		 */
2049      		schedule();
2050
2051
2052		PROTECT_CTX(ctx, flags);
2053
2054
2055		remove_wait_queue(&ctx->ctx_zombieq, &wait);
2056  		set_current_state(TASK_RUNNING);
2057
2058		/*
2059		 * context is unloaded at this point
2060		 */
2061		DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2062	}
2063	else if (task != current) {
2064#ifdef CONFIG_SMP
2065		/*
2066	 	 * switch context to zombie state
2067	 	 */
2068		ctx->ctx_state = PFM_CTX_ZOMBIE;
2069
2070		DPRINT(("zombie ctx for [%d]\n", task_pid_nr(task)));
2071		/*
2072		 * cannot free the context on the spot. deferred until
2073		 * the task notices the ZOMBIE state
2074		 */
2075		free_possible = 0;
2076#else
2077		pfm_context_unload(ctx, NULL, 0, regs);
2078#endif
2079	}
2080
2081doit:
2082	/* reload state, may have changed during  opening of critical section */
2083	state = ctx->ctx_state;
2084
2085	/*
2086	 * the context is still attached to a task (possibly current)
2087	 * we cannot destroy it right now
2088	 */
2089
2090	/*
2091	 * we must free the sampling buffer right here because
2092	 * we cannot rely on it being cleaned up later by the
2093	 * monitored task. It is not possible to free vmalloc'ed
2094	 * memory in pfm_load_regs(). Instead, we remove the buffer
2095	 * now. should there be subsequent PMU overflow originally
2096	 * meant for sampling, the will be converted to spurious
2097	 * and that's fine because the monitoring tools is gone anyway.
2098	 */
2099	if (ctx->ctx_smpl_hdr) {
2100		smpl_buf_addr = ctx->ctx_smpl_hdr;
2101		smpl_buf_size = ctx->ctx_smpl_size;
2102		/* no more sampling */
2103		ctx->ctx_smpl_hdr = NULL;
2104		ctx->ctx_fl_is_sampling = 0;
2105	}
2106
2107	DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2108		state,
2109		free_possible,
2110		smpl_buf_addr,
2111		smpl_buf_size));
2112
2113	if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2114
2115	/*
2116	 * UNLOADED that the session has already been unreserved.
2117	 */
2118	if (state == PFM_CTX_ZOMBIE) {
2119		pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2120	}
2121
2122	/*
2123	 * disconnect file descriptor from context must be done
2124	 * before we unlock.
2125	 */
2126	filp->private_data = NULL;
2127
2128	/*
2129	 * if we free on the spot, the context is now completely unreachable
2130	 * from the callers side. The monitored task side is also cut, so we
2131	 * can freely cut.
2132	 *
2133	 * If we have a deferred free, only the caller side is disconnected.
2134	 */
2135	UNPROTECT_CTX(ctx, flags);
2136
2137	/*
2138	 * All memory free operations (especially for vmalloc'ed memory)
2139	 * MUST be done with interrupts ENABLED.
2140	 */
2141	if (smpl_buf_addr)  pfm_rvfree(smpl_buf_addr, smpl_buf_size);
2142
2143	/*
2144	 * return the memory used by the context
2145	 */
2146	if (free_possible) pfm_context_free(ctx);
2147
2148	return 0;
2149}
2150
2151static int
2152pfm_no_open(struct inode *irrelevant, struct file *dontcare)
2153{
2154	DPRINT(("pfm_no_open called\n"));
2155	return -ENXIO;
2156}
2157
2158
2159
2160static const struct file_operations pfm_file_ops = {
2161	.llseek		= no_llseek,
2162	.read		= pfm_read,
2163	.write		= pfm_write,
2164	.poll		= pfm_poll,
2165	.unlocked_ioctl = pfm_ioctl,
2166	.open		= pfm_no_open,	/* special open code to disallow open via /proc */
2167	.fasync		= pfm_fasync,
2168	.release	= pfm_close,
2169	.flush		= pfm_flush
2170};
2171
2172static int
2173pfmfs_delete_dentry(const struct dentry *dentry)
2174{
2175	return 1;
2176}
2177
2178static char *pfmfs_dname(struct dentry *dentry, char *buffer, int buflen)
2179{
2180	return dynamic_dname(dentry, buffer, buflen, "pfm:[%lu]",
2181			     dentry->d_inode->i_ino);
2182}
2183
2184static const struct dentry_operations pfmfs_dentry_operations = {
2185	.d_delete = pfmfs_delete_dentry,
2186	.d_dname = pfmfs_dname,
2187};
2188
2189
2190static struct file *
2191pfm_alloc_file(pfm_context_t *ctx)
2192{
2193	struct file *file;
2194	struct inode *inode;
2195	struct path path;
2196	struct qstr this = { .name = "" };
2197
2198	/*
2199	 * allocate a new inode
2200	 */
2201	inode = new_inode(pfmfs_mnt->mnt_sb);
2202	if (!inode)
2203		return ERR_PTR(-ENOMEM);
2204
2205	DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2206
2207	inode->i_mode = S_IFCHR|S_IRUGO;
2208	inode->i_uid  = current_fsuid();
2209	inode->i_gid  = current_fsgid();
2210
2211	/*
2212	 * allocate a new dcache entry
2213	 */
2214	path.dentry = d_alloc(pfmfs_mnt->mnt_root, &this);
2215	if (!path.dentry) {
2216		iput(inode);
2217		return ERR_PTR(-ENOMEM);
2218	}
2219	path.mnt = mntget(pfmfs_mnt);
2220
2221	d_add(path.dentry, inode);
2222
2223	file = alloc_file(&path, FMODE_READ, &pfm_file_ops);
2224	if (!file) {
2225		path_put(&path);
2226		return ERR_PTR(-ENFILE);
2227	}
2228
2229	file->f_flags = O_RDONLY;
2230	file->private_data = ctx;
2231
2232	return file;
2233}
2234
2235static int
2236pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2237{
2238	DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2239
2240	while (size > 0) {
2241		unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2242
2243
2244		if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2245			return -ENOMEM;
2246
2247		addr  += PAGE_SIZE;
2248		buf   += PAGE_SIZE;
2249		size  -= PAGE_SIZE;
2250	}
2251	return 0;
2252}
2253
2254/*
2255 * allocate a sampling buffer and remaps it into the user address space of the task
2256 */
2257static int
2258pfm_smpl_buffer_alloc(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2259{
2260	struct mm_struct *mm = task->mm;
2261	struct vm_area_struct *vma = NULL;
2262	unsigned long size;
2263	void *smpl_buf;
2264
2265
2266	/*
2267	 * the fixed header + requested size and align to page boundary
2268	 */
2269	size = PAGE_ALIGN(rsize);
2270
2271	DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2272
2273	/*
2274	 * check requested size to avoid Denial-of-service attacks
2275	 * XXX: may have to refine this test
2276	 * Check against address space limit.
2277	 *
2278	 * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2279	 * 	return -ENOMEM;
2280	 */
2281	if (size > task_rlimit(task, RLIMIT_MEMLOCK))
2282		return -ENOMEM;
2283
2284	/*
2285	 * We do the easy to undo allocations first.
2286 	 *
2287	 * pfm_rvmalloc(), clears the buffer, so there is no leak
2288	 */
2289	smpl_buf = pfm_rvmalloc(size);
2290	if (smpl_buf == NULL) {
2291		DPRINT(("Can't allocate sampling buffer\n"));
2292		return -ENOMEM;
2293	}
2294
2295	DPRINT(("smpl_buf @%p\n", smpl_buf));
2296
2297	/* allocate vma */
2298	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2299	if (!vma) {
2300		DPRINT(("Cannot allocate vma\n"));
2301		goto error_kmem;
2302	}
2303	INIT_LIST_HEAD(&vma->anon_vma_chain);
2304
2305	/*
2306	 * partially initialize the vma for the sampling buffer
2307	 */
2308	vma->vm_mm	     = mm;
2309	vma->vm_file	     = filp;
2310	vma->vm_flags	     = VM_READ| VM_MAYREAD |VM_RESERVED;
2311	vma->vm_page_prot    = PAGE_READONLY; /* XXX may need to change */
2312
2313	/*
2314	 * Now we have everything we need and we can initialize
2315	 * and connect all the data structures
2316	 */
2317
2318	ctx->ctx_smpl_hdr   = smpl_buf;
2319	ctx->ctx_smpl_size  = size; /* aligned size */
2320
2321	/*
2322	 * Let's do the difficult operations next.
2323	 *
2324	 * now we atomically find some area in the address space and
2325	 * remap the buffer in it.
2326	 */
2327	down_write(&task->mm->mmap_sem);
2328
2329	/* find some free area in address space, must have mmap sem held */
2330	vma->vm_start = get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS);
2331	if (IS_ERR_VALUE(vma->vm_start)) {
2332		DPRINT(("Cannot find unmapped area for size %ld\n", size));
2333		up_write(&task->mm->mmap_sem);
2334		goto error;
2335	}
2336	vma->vm_end = vma->vm_start + size;
2337	vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2338
2339	DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2340
2341	/* can only be applied to current task, need to have the mm semaphore held when called */
2342	if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2343		DPRINT(("Can't remap buffer\n"));
2344		up_write(&task->mm->mmap_sem);
2345		goto error;
2346	}
2347
2348	get_file(filp);
2349
2350	/*
2351	 * now insert the vma in the vm list for the process, must be
2352	 * done with mmap lock held
2353	 */
2354	insert_vm_struct(mm, vma);
2355
2356	mm->total_vm  += size >> PAGE_SHIFT;
2357	vm_stat_account(vma->vm_mm, vma->vm_flags, vma->vm_file,
2358							vma_pages(vma));
2359	up_write(&task->mm->mmap_sem);
2360
2361	/*
2362	 * keep track of user level virtual address
2363	 */
2364	ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2365	*(unsigned long *)user_vaddr = vma->vm_start;
2366
2367	return 0;
2368
2369error:
2370	kmem_cache_free(vm_area_cachep, vma);
2371error_kmem:
2372	pfm_rvfree(smpl_buf, size);
2373
2374	return -ENOMEM;
2375}
2376
2377/*
2378 * XXX: do something better here
2379 */
2380static int
2381pfm_bad_permissions(struct task_struct *task)
2382{
2383	const struct cred *tcred;
2384	uid_t uid = current_uid();
2385	gid_t gid = current_gid();
2386	int ret;
2387
2388	rcu_read_lock();
2389	tcred = __task_cred(task);
2390
2391	/* inspired by ptrace_attach() */
2392	DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2393		uid,
2394		gid,
2395		tcred->euid,
2396		tcred->suid,
2397		tcred->uid,
2398		tcred->egid,
2399		tcred->sgid));
2400
2401	ret = ((uid != tcred->euid)
2402	       || (uid != tcred->suid)
2403	       || (uid != tcred->uid)
2404	       || (gid != tcred->egid)
2405	       || (gid != tcred->sgid)
2406	       || (gid != tcred->gid)) && !capable(CAP_SYS_PTRACE);
2407
2408	rcu_read_unlock();
2409	return ret;
2410}
2411
2412static int
2413pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2414{
2415	int ctx_flags;
2416
2417	/* valid signal */
2418
2419	ctx_flags = pfx->ctx_flags;
2420
2421	if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2422
2423		/*
2424		 * cannot block in this mode
2425		 */
2426		if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2427			DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2428			return -EINVAL;
2429		}
2430	} else {
2431	}
2432	/* probably more to add here */
2433
2434	return 0;
2435}
2436
2437static int
2438pfm_setup_buffer_fmt(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned int ctx_flags,
2439		     unsigned int cpu, pfarg_context_t *arg)
2440{
2441	pfm_buffer_fmt_t *fmt = NULL;
2442	unsigned long size = 0UL;
2443	void *uaddr = NULL;
2444	void *fmt_arg = NULL;
2445	int ret = 0;
2446#define PFM_CTXARG_BUF_ARG(a)	(pfm_buffer_fmt_t *)(a+1)
2447
2448	/* invoke and lock buffer format, if found */
2449	fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2450	if (fmt == NULL) {
2451		DPRINT(("[%d] cannot find buffer format\n", task_pid_nr(task)));
2452		return -EINVAL;
2453	}
2454
2455	/*
2456	 * buffer argument MUST be contiguous to pfarg_context_t
2457	 */
2458	if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2459
2460	ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2461
2462	DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task_pid_nr(task), ctx_flags, cpu, fmt_arg, ret));
2463
2464	if (ret) goto error;
2465
2466	/* link buffer format and context */
2467	ctx->ctx_buf_fmt = fmt;
2468	ctx->ctx_fl_is_sampling = 1; /* assume record() is defined */
2469
2470	/*
2471	 * check if buffer format wants to use perfmon buffer allocation/mapping service
2472	 */
2473	ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2474	if (ret) goto error;
2475
2476	if (size) {
2477		/*
2478		 * buffer is always remapped into the caller's address space
2479		 */
2480		ret = pfm_smpl_buffer_alloc(current, filp, ctx, size, &uaddr);
2481		if (ret) goto error;
2482
2483		/* keep track of user address of buffer */
2484		arg->ctx_smpl_vaddr = uaddr;
2485	}
2486	ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2487
2488error:
2489	return ret;
2490}
2491
2492static void
2493pfm_reset_pmu_state(pfm_context_t *ctx)
2494{
2495	int i;
2496
2497	/*
2498	 * install reset values for PMC.
2499	 */
2500	for (i=1; PMC_IS_LAST(i) == 0; i++) {
2501		if (PMC_IS_IMPL(i) == 0) continue;
2502		ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2503		DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2504	}
2505	/*
2506	 * PMD registers are set to 0UL when the context in memset()
2507	 */
2508
2509	/*
2510	 * On context switched restore, we must restore ALL pmc and ALL pmd even
2511	 * when they are not actively used by the task. In UP, the incoming process
2512	 * may otherwise pick up left over PMC, PMD state from the previous process.
2513	 * As opposed to PMD, stale PMC can cause harm to the incoming
2514	 * process because they may change what is being measured.
2515	 * Therefore, we must systematically reinstall the entire
2516	 * PMC state. In SMP, the same thing is possible on the
2517	 * same CPU but also on between 2 CPUs.
2518	 *
2519	 * The problem with PMD is information leaking especially
2520	 * to user level when psr.sp=0
2521	 *
2522	 * There is unfortunately no easy way to avoid this problem
2523	 * on either UP or SMP. This definitively slows down the
2524	 * pfm_load_regs() function.
2525	 */
2526
2527	 /*
2528	  * bitmask of all PMCs accessible to this context
2529	  *
2530	  * PMC0 is treated differently.
2531	  */
2532	ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2533
2534	/*
2535	 * bitmask of all PMDs that are accessible to this context
2536	 */
2537	ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2538
2539	DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2540
2541	/*
2542	 * useful in case of re-enable after disable
2543	 */
2544	ctx->ctx_used_ibrs[0] = 0UL;
2545	ctx->ctx_used_dbrs[0] = 0UL;
2546}
2547
2548static int
2549pfm_ctx_getsize(void *arg, size_t *sz)
2550{
2551	pfarg_context_t *req = (pfarg_context_t *)arg;
2552	pfm_buffer_fmt_t *fmt;
2553
2554	*sz = 0;
2555
2556	if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2557
2558	fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2559	if (fmt == NULL) {
2560		DPRINT(("cannot find buffer format\n"));
2561		return -EINVAL;
2562	}
2563	/* get just enough to copy in user parameters */
2564	*sz = fmt->fmt_arg_size;
2565	DPRINT(("arg_size=%lu\n", *sz));
2566
2567	return 0;
2568}
2569
2570
2571
2572/*
2573 * cannot attach if :
2574 * 	- kernel task
2575 * 	- task not owned by caller
2576 * 	- task incompatible with context mode
2577 */
2578static int
2579pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2580{
2581	/*
2582	 * no kernel task or task not owner by caller
2583	 */
2584	if (task->mm == NULL) {
2585		DPRINT(("task [%d] has not memory context (kernel thread)\n", task_pid_nr(task)));
2586		return -EPERM;
2587	}
2588	if (pfm_bad_permissions(task)) {
2589		DPRINT(("no permission to attach to  [%d]\n", task_pid_nr(task)));
2590		return -EPERM;
2591	}
2592	/*
2593	 * cannot block in self-monitoring mode
2594	 */
2595	if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2596		DPRINT(("cannot load a blocking context on self for [%d]\n", task_pid_nr(task)));
2597		return -EINVAL;
2598	}
2599
2600	if (task->exit_state == EXIT_ZOMBIE) {
2601		DPRINT(("cannot attach to  zombie task [%d]\n", task_pid_nr(task)));
2602		return -EBUSY;
2603	}
2604
2605	/*
2606	 * always ok for self
2607	 */
2608	if (task == current) return 0;
2609
2610	if (!task_is_stopped_or_traced(task)) {
2611		DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task_pid_nr(task), task->state));
2612		return -EBUSY;
2613	}
2614	/*
2615	 * make sure the task is off any CPU
2616	 */
2617	wait_task_inactive(task, 0);
2618
2619	/* more to come... */
2620
2621	return 0;
2622}
2623
2624static int
2625pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2626{
2627	struct task_struct *p = current;
2628	int ret;
2629
2630	/* XXX: need to add more checks here */
2631	if (pid < 2) return -EPERM;
2632
2633	if (pid != task_pid_vnr(current)) {
2634
2635		read_lock(&tasklist_lock);
2636
2637		p = find_task_by_vpid(pid);
2638
2639		/* make sure task cannot go away while we operate on it */
2640		if (p) get_task_struct(p);
2641
2642		read_unlock(&tasklist_lock);
2643
2644		if (p == NULL) return -ESRCH;
2645	}
2646
2647	ret = pfm_task_incompatible(ctx, p);
2648	if (ret == 0) {
2649		*task = p;
2650	} else if (p != current) {
2651		pfm_put_task(p);
2652	}
2653	return ret;
2654}
2655
2656
2657
2658static int
2659pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2660{
2661	pfarg_context_t *req = (pfarg_context_t *)arg;
2662	struct file *filp;
2663	struct path path;
2664	int ctx_flags;
2665	int fd;
2666	int ret;
2667
2668	/* let's check the arguments first */
2669	ret = pfarg_is_sane(current, req);
2670	if (ret < 0)
2671		return ret;
2672
2673	ctx_flags = req->ctx_flags;
2674
2675	ret = -ENOMEM;
2676
2677	fd = get_unused_fd();
2678	if (fd < 0)
2679		return fd;
2680
2681	ctx = pfm_context_alloc(ctx_flags);
2682	if (!ctx)
2683		goto error;
2684
2685	filp = pfm_alloc_file(ctx);
2686	if (IS_ERR(filp)) {
2687		ret = PTR_ERR(filp);
2688		goto error_file;
2689	}
2690
2691	req->ctx_fd = ctx->ctx_fd = fd;
2692
2693	/*
2694	 * does the user want to sample?
2695	 */
2696	if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2697		ret = pfm_setup_buffer_fmt(current, filp, ctx, ctx_flags, 0, req);
2698		if (ret)
2699			goto buffer_error;
2700	}
2701
2702	DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d\n",
2703		ctx,
2704		ctx_flags,
2705		ctx->ctx_fl_system,
2706		ctx->ctx_fl_block,
2707		ctx->ctx_fl_excl_idle,
2708		ctx->ctx_fl_no_msg,
2709		ctx->ctx_fd));
2710
2711	/*
2712	 * initialize soft PMU state
2713	 */
2714	pfm_reset_pmu_state(ctx);
2715
2716	fd_install(fd, filp);
2717
2718	return 0;
2719
2720buffer_error:
2721	path = filp->f_path;
2722	put_filp(filp);
2723	path_put(&path);
2724
2725	if (ctx->ctx_buf_fmt) {
2726		pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2727	}
2728error_file:
2729	pfm_context_free(ctx);
2730
2731error:
2732	put_unused_fd(fd);
2733	return ret;
2734}
2735
2736static inline unsigned long
2737pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2738{
2739	unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2740	unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2741	extern unsigned long carta_random32 (unsigned long seed);
2742
2743	if (reg->flags & PFM_REGFL_RANDOM) {
2744		new_seed = carta_random32(old_seed);
2745		val -= (old_seed & mask);	/* counter values are negative numbers! */
2746		if ((mask >> 32) != 0)
2747			/* construct a full 64-bit random value: */
2748			new_seed |= carta_random32(old_seed >> 32) << 32;
2749		reg->seed = new_seed;
2750	}
2751	reg->lval = val;
2752	return val;
2753}
2754
2755static void
2756pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2757{
2758	unsigned long mask = ovfl_regs[0];
2759	unsigned long reset_others = 0UL;
2760	unsigned long val;
2761	int i;
2762
2763	/*
2764	 * now restore reset value on sampling overflowed counters
2765	 */
2766	mask >>= PMU_FIRST_COUNTER;
2767	for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2768
2769		if ((mask & 0x1UL) == 0UL) continue;
2770
2771		ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2772		reset_others        |= ctx->ctx_pmds[i].reset_pmds[0];
2773
2774		DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2775	}
2776
2777	/*
2778	 * Now take care of resetting the other registers
2779	 */
2780	for(i = 0; reset_others; i++, reset_others >>= 1) {
2781
2782		if ((reset_others & 0x1) == 0) continue;
2783
2784		ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2785
2786		DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2787			  is_long_reset ? "long" : "short", i, val));
2788	}
2789}
2790
2791static void
2792pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2793{
2794	unsigned long mask = ovfl_regs[0];
2795	unsigned long reset_others = 0UL;
2796	unsigned long val;
2797	int i;
2798
2799	DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2800
2801	if (ctx->ctx_state == PFM_CTX_MASKED) {
2802		pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2803		return;
2804	}
2805
2806	/*
2807	 * now restore reset value on sampling overflowed counters
2808	 */
2809	mask >>= PMU_FIRST_COUNTER;
2810	for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2811
2812		if ((mask & 0x1UL) == 0UL) continue;
2813
2814		val           = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2815		reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2816
2817		DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2818
2819		pfm_write_soft_counter(ctx, i, val);
2820	}
2821
2822	/*
2823	 * Now take care of resetting the other registers
2824	 */
2825	for(i = 0; reset_others; i++, reset_others >>= 1) {
2826
2827		if ((reset_others & 0x1) == 0) continue;
2828
2829		val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2830
2831		if (PMD_IS_COUNTING(i)) {
2832			pfm_write_soft_counter(ctx, i, val);
2833		} else {
2834			ia64_set_pmd(i, val);
2835		}
2836		DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2837			  is_long_reset ? "long" : "short", i, val));
2838	}
2839	ia64_srlz_d();
2840}
2841
2842static int
2843pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2844{
2845	struct task_struct *task;
2846	pfarg_reg_t *req = (pfarg_reg_t *)arg;
2847	unsigned long value, pmc_pm;
2848	unsigned long smpl_pmds, reset_pmds, impl_pmds;
2849	unsigned int cnum, reg_flags, flags, pmc_type;
2850	int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2851	int is_monitor, is_counting, state;
2852	int ret = -EINVAL;
2853	pfm_reg_check_t	wr_func;
2854#define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2855
2856	state     = ctx->ctx_state;
2857	is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2858	is_system = ctx->ctx_fl_system;
2859	task      = ctx->ctx_task;
2860	impl_pmds = pmu_conf->impl_pmds[0];
2861
2862	if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2863
2864	if (is_loaded) {
2865		/*
2866		 * In system wide and when the context is loaded, access can only happen
2867		 * when the caller is running on the CPU being monitored by the session.
2868		 * It does not have to be the owner (ctx_task) of the context per se.
2869		 */
2870		if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2871			DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2872			return -EBUSY;
2873		}
2874		can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2875	}
2876	expert_mode = pfm_sysctl.expert_mode; 
2877
2878	for (i = 0; i < count; i++, req++) {
2879
2880		cnum       = req->reg_num;
2881		reg_flags  = req->reg_flags;
2882		value      = req->reg_value;
2883		smpl_pmds  = req->reg_smpl_pmds[0];
2884		reset_pmds = req->reg_reset_pmds[0];
2885		flags      = 0;
2886
2887
2888		if (cnum >= PMU_MAX_PMCS) {
2889			DPRINT(("pmc%u is invalid\n", cnum));
2890			goto error;
2891		}
2892
2893		pmc_type   = pmu_conf->pmc_desc[cnum].type;
2894		pmc_pm     = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2895		is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2896		is_monitor  = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2897
2898		/*
2899		 * we reject all non implemented PMC as well
2900		 * as attempts to modify PMC[0-3] which are used
2901		 * as status registers by the PMU
2902		 */
2903		if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2904			DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2905			goto error;
2906		}
2907		wr_func = pmu_conf->pmc_desc[cnum].write_check;
2908		/*
2909		 * If the PMC is a monitor, then if the value is not the default:
2910		 * 	- system-wide session: PMCx.pm=1 (privileged monitor)
2911		 * 	- per-task           : PMCx.pm=0 (user monitor)
2912		 */
2913		if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2914			DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2915				cnum,
2916				pmc_pm,
2917				is_system));
2918			goto error;
2919		}
2920
2921		if (is_counting) {
2922			/*
2923		 	 * enforce generation of overflow interrupt. Necessary on all
2924		 	 * CPUs.
2925		 	 */
2926			value |= 1 << PMU_PMC_OI;
2927
2928			if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2929				flags |= PFM_REGFL_OVFL_NOTIFY;
2930			}
2931
2932			if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2933
2934			/* verify validity of smpl_pmds */
2935			if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2936				DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2937				goto error;
2938			}
2939
2940			/* verify validity of reset_pmds */
2941			if ((reset_pmds & impl_pmds) != reset_pmds) {
2942				DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2943				goto error;
2944			}
2945		} else {
2946			if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2947				DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2948				goto error;
2949			}
2950			/* eventid on non-counting monitors are ignored */
2951		}
2952
2953		/*
2954		 * execute write checker, if any
2955		 */
2956		if (likely(expert_mode == 0 && wr_func)) {
2957			ret = (*wr_func)(task, ctx, cnum, &value, regs);
2958			if (ret) goto error;
2959			ret = -EINVAL;
2960		}
2961
2962		/*
2963		 * no error on this register
2964		 */
2965		PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2966
2967		/*
2968		 * Now we commit the changes to the software state
2969		 */
2970
2971		/*
2972		 * update overflow information
2973		 */
2974		if (is_counting) {
2975			/*
2976		 	 * full flag update each time a register is programmed
2977		 	 */
2978			ctx->ctx_pmds[cnum].flags = flags;
2979
2980			ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2981			ctx->ctx_pmds[cnum].smpl_pmds[0]  = smpl_pmds;
2982			ctx->ctx_pmds[cnum].eventid       = req->reg_smpl_eventid;
2983
2984			/*
2985			 * Mark all PMDS to be accessed as used.
2986			 *
2987			 * We do not keep track of PMC because we have to
2988			 * systematically restore ALL of them.
2989			 *
2990			 * We do not update the used_monitors mask, because
2991			 * if we have not programmed them, then will be in
2992			 * a quiescent state, therefore we will not need to
2993			 * mask/restore then when context is MASKED.
2994			 */
2995			CTX_USED_PMD(ctx, reset_pmds);
2996			CTX_USED_PMD(ctx, smpl_pmds);
2997			/*
2998		 	 * make sure we do not try to reset on
2999		 	 * restart because we have established new values
3000		 	 */
3001			if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3002		}
3003		/*
3004		 * Needed in case the user does not initialize the equivalent
3005		 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
3006		 * possible leak here.
3007		 */
3008		CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
3009
3010		/*
3011		 * keep track of the monitor PMC that we are using.
3012		 * we save the value of the pmc in ctx_pmcs[] and if
3013		 * the monitoring is not stopped for the context we also
3014		 * place it in the saved state area so that it will be
3015		 * picked up later by the context switch code.
3016		 *
3017		 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3018		 *
3019		 * The value in th_pmcs[] may be modified on overflow, i.e.,  when
3020		 * monitoring needs to be stopped.
3021		 */
3022		if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3023
3024		/*
3025		 * update context state
3026		 */
3027		ctx->ctx_pmcs[cnum] = value;
3028
3029		if (is_loaded) {
3030			/*
3031			 * write thread state
3032			 */
3033			if (is_system == 0) ctx->th_pmcs[cnum] = value;
3034
3035			/*
3036			 * write hardware register if we can
3037			 */
3038			if (can_access_pmu) {
3039				ia64_set_pmc(cnum, value);
3040			}
3041#ifdef CONFIG_SMP
3042			else {
3043				/*
3044				 * per-task SMP only here
3045				 *
3046			 	 * we are guaranteed that the task is not running on the other CPU,
3047			 	 * we indicate that this PMD will need to be reloaded if the task
3048			 	 * is rescheduled on the CPU it ran last on.
3049			 	 */
3050				ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3051			}
3052#endif
3053		}
3054
3055		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",
3056			  cnum,
3057			  value,
3058			  is_loaded,
3059			  can_access_pmu,
3060			  flags,
3061			  ctx->ctx_all_pmcs[0],
3062			  ctx->ctx_used_pmds[0],
3063			  ctx->ctx_pmds[cnum].eventid,
3064			  smpl_pmds,
3065			  reset_pmds,
3066			  ctx->ctx_reload_pmcs[0],
3067			  ctx->ctx_used_monitors[0],
3068			  ctx->ctx_ovfl_regs[0]));
3069	}
3070
3071	/*
3072	 * make sure the changes are visible
3073	 */
3074	if (can_access_pmu) ia64_srlz_d();
3075
3076	return 0;
3077error:
3078	PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3079	return ret;
3080}
3081
3082static int
3083pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3084{
3085	struct task_struct *task;
3086	pfarg_reg_t *req = (pfarg_reg_t *)arg;
3087	unsigned long value, hw_value, ovfl_mask;
3088	unsigned int cnum;
3089	int i, can_access_pmu = 0, state;
3090	int is_counting, is_loaded, is_system, expert_mode;
3091	int ret = -EINVAL;
3092	pfm_reg_check_t wr_func;
3093
3094
3095	state     = ctx->ctx_state;
3096	is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3097	is_system = ctx->ctx_fl_system;
3098	ovfl_mask = pmu_conf->ovfl_val;
3099	task      = ctx->ctx_task;
3100
3101	if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3102
3103	/*
3104	 * on both UP and SMP, we can only write to the PMC when the task is
3105	 * the owner of the local PMU.
3106	 */
3107	if (likely(is_loaded)) {
3108		/*
3109		 * In system wide and when the context is loaded, access can only happen
3110		 * when the caller is running on the CPU being monitored by the session.
3111		 * It does not have to be the owner (ctx_task) of the context per se.
3112		 */
3113		if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3114			DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3115			return -EBUSY;
3116		}
3117		can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3118	}
3119	expert_mode = pfm_sysctl.expert_mode; 
3120
3121	for (i = 0; i < count; i++, req++) {
3122
3123		cnum  = req->reg_num;
3124		value = req->reg_value;
3125
3126		if (!PMD_IS_IMPL(cnum)) {
3127			DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3128			goto abort_mission;
3129		}
3130		is_counting = PMD_IS_COUNTING(cnum);
3131		wr_func     = pmu_conf->pmd_desc[cnum].write_check;
3132
3133		/*
3134		 * execute write checker, if any
3135		 */
3136		if (unlikely(expert_mode == 0 && wr_func)) {
3137			unsigned long v = value;
3138
3139			ret = (*wr_func)(task, ctx, cnum, &v, regs);
3140			if (ret) goto abort_mission;
3141
3142			value = v;
3143			ret   = -EINVAL;
3144		}
3145
3146		/*
3147		 * no error on this register
3148		 */
3149		PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3150
3151		/*
3152		 * now commit changes to software state
3153		 */
3154		hw_value = value;
3155
3156		/*
3157		 * update virtualized (64bits) counter
3158		 */
3159		if (is_counting) {
3160			/*
3161			 * write context state
3162			 */
3163			ctx->ctx_pmds[cnum].lval = value;
3164
3165			/*
3166			 * when context is load we use the split value
3167			 */
3168			if (is_loaded) {
3169				hw_value = value &  ovfl_mask;
3170				value    = value & ~ovfl_mask;
3171			}
3172		}
3173		/*
3174		 * update reset values (not just for counters)
3175		 */
3176		ctx->ctx_pmds[cnum].long_reset  = req->reg_long_reset;
3177		ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3178
3179		/*
3180		 * update randomization parameters (not just for counters)
3181		 */
3182		ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3183		ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3184
3185		/*
3186		 * update context value
3187		 */
3188		ctx->ctx_pmds[cnum].val  = value;
3189
3190		/*
3191		 * Keep track of what we use
3192		 *
3193		 * We do not keep track of PMC because we have to
3194		 * systematically restore ALL of them.
3195		 */
3196		CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3197
3198		/*
3199		 * mark this PMD register used as well
3200		 */
3201		CTX_USED_PMD(ctx, RDEP(cnum));
3202
3203		/*
3204		 * make sure we do not try to reset on
3205		 * restart because we have established new values
3206		 */
3207		if (is_counting && state == PFM_CTX_MASKED) {
3208			ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3209		}
3210
3211		if (is_loaded) {
3212			/*
3213		 	 * write thread state
3214		 	 */
3215			if (is_system == 0) ctx->th_pmds[cnum] = hw_value;
3216
3217			/*
3218			 * write hardware register if we can
3219			 */
3220			if (can_access_pmu) {
3221				ia64_set_pmd(cnum, hw_value);
3222			} else {
3223#ifdef CONFIG_SMP
3224				/*
3225			 	 * we are guaranteed that the task is not running on the other CPU,
3226			 	 * we indicate that this PMD will need to be reloaded if the task
3227			 	 * is rescheduled on the CPU it ran last on.
3228			 	 */
3229				ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3230#endif
3231			}
3232		}
3233
3234		DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx  short_reset=0x%lx "
3235			  "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",
3236			cnum,
3237			value,
3238			is_loaded,
3239			can_access_pmu,
3240			hw_value,
3241			ctx->ctx_pmds[cnum].val,
3242			ctx->ctx_pmds[cnum].short_reset,
3243			ctx->ctx_pmds[cnum].long_reset,
3244			PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3245			ctx->ctx_pmds[cnum].seed,
3246			ctx->ctx_pmds[cnum].mask,
3247			ctx->ctx_used_pmds[0],
3248			ctx->ctx_pmds[cnum].reset_pmds[0],
3249			ctx->ctx_reload_pmds[0],
3250			ctx->ctx_all_pmds[0],
3251			ctx->ctx_ovfl_regs[0]));
3252	}
3253
3254	/*
3255	 * make changes visible
3256	 */
3257	if (can_access_pmu) ia64_srlz_d();
3258
3259	return 0;
3260
3261abort_mission:
3262	/*
3263	 * for now, we have only one possibility for error
3264	 */
3265	PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3266	return ret;
3267}
3268
3269/*
3270 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3271 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3272 * interrupt is delivered during the call, it will be kept pending until we leave, making
3273 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3274 * guaranteed to return consistent data to the user, it may simply be old. It is not
3275 * trivial to treat the overflow while inside the call because you may end up in
3276 * some module sampling buffer code causing deadlocks.
3277 */
3278static int
3279pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3280{
3281	struct task_struct *task;
3282	unsigned long val = 0UL, lval, ovfl_mask, sval;
3283	pfarg_reg_t *req = (pfarg_reg_t *)arg;
3284	unsigned int cnum, reg_flags = 0;
3285	int i, can_access_pmu = 0, state;
3286	int is_loaded, is_system, is_counting, expert_mode;
3287	int ret = -EINVAL;
3288	pfm_reg_check_t rd_func;
3289
3290	/*
3291	 * access is possible when loaded only for
3292	 * self-monitoring tasks or in UP mode
3293	 */
3294
3295	state     = ctx->ctx_state;
3296	is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3297	is_system = ctx->ctx_fl_system;
3298	ovfl_mask = pmu_conf->ovfl_val;
3299	task      = ctx->ctx_task;
3300
3301	if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3302
3303	if (likely(is_loaded)) {
3304		/*
3305		 * In system wide and when the context is loaded, access can only happen
3306		 * when the caller is running on the CPU being monitored by the session.
3307		 * It does not have to be the owner (ctx_task) of the context per se.
3308		 */
3309		if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3310			DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3311			return -EBUSY;
3312		}
3313		/*
3314		 * this can be true when not self-monitoring only in UP
3315		 */
3316		can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3317
3318		if (can_access_pmu) ia64_srlz_d();
3319	}
3320	expert_mode = pfm_sysctl.expert_mode; 
3321
3322	DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3323		is_loaded,
3324		can_access_pmu,
3325		state));
3326
3327	/*
3328	 * on both UP and SMP, we can only read the PMD from the hardware register when
3329	 * the task is the owner of the local PMU.
3330	 */
3331
3332	for (i = 0; i < count; i++, req++) {
3333
3334		cnum        = req->reg_num;
3335		reg_flags   = req->reg_flags;
3336
3337		if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3338		/*
3339		 * we can only read the register that we use. That includes
3340		 * the one we explicitly initialize AND the one we want included
3341		 * in the sampling buffer (smpl_regs).
3342		 *
3343		 * Having this restriction allows optimization in the ctxsw routine
3344		 * without compromising security (leaks)
3345		 */
3346		if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3347
3348		sval        = ctx->ctx_pmds[cnum].val;
3349		lval        = ctx->ctx_pmds[cnum].lval;
3350		is_counting = PMD_IS_COUNTING(cnum);
3351
3352		/*
3353		 * If the task is not the current one, then we check if the
3354		 * PMU state is still in the local live register due to lazy ctxsw.
3355		 * If true, then we read directly from the registers.
3356		 */
3357		if (can_access_pmu){
3358			val = ia64_get_pmd(cnum);
3359		} else {
3360			/*
3361			 * context has been saved
3362			 * if context is zombie, then task does not exist anymore.
3363			 * In this case, we use the full value saved in the context (pfm_flush_regs()).
3364			 */
3365			val = is_loaded ? ctx->th_pmds[cnum] : 0UL;
3366		}
3367		rd_func = pmu_conf->pmd_desc[cnum].read_check;
3368
3369		if (is_counting) {
3370			/*
3371			 * XXX: need to check for overflow when loaded
3372			 */
3373			val &= ovfl_mask;
3374			val += sval;
3375		}
3376
3377		/*
3378		 * execute read checker, if any
3379		 */
3380		if (unlikely(expert_mode == 0 && rd_func)) {
3381			unsigned long v = val;
3382			ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3383			if (ret) goto error;
3384			val = v;
3385			ret = -EINVAL;
3386		}
3387
3388		PFM_REG_RETFLAG_SET(reg_flags, 0);
3389
3390		DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3391
3392		/*
3393		 * update register return value, abort all if problem during copy.
3394		 * we only modify the reg_flags field. no check mode is fine because
3395		 * access has been verified upfront in sys_perfmonctl().
3396		 */
3397		req->reg_value            = val;
3398		req->reg_flags            = reg_flags;
3399		req->reg_last_reset_val   = lval;
3400	}
3401
3402	return 0;
3403
3404error:
3405	PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3406	return ret;
3407}
3408
3409int
3410pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3411{
3412	pfm_context_t *ctx;
3413
3414	if (req == NULL) return -EINVAL;
3415
3416 	ctx = GET_PMU_CTX();
3417
3418	if (ctx == NULL) return -EINVAL;
3419
3420	/*
3421	 * for now limit to current task, which is enough when calling
3422	 * from overflow handler
3423	 */
3424	if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3425
3426	return pfm_write_pmcs(ctx, req, nreq, regs);
3427}
3428EXPORT_SYMBOL(pfm_mod_write_pmcs);
3429
3430int
3431pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3432{
3433	pfm_context_t *ctx;
3434
3435	if (req == NULL) return -EINVAL;
3436
3437 	ctx = GET_PMU_CTX();
3438
3439	if (ctx == NULL) return -EINVAL;
3440
3441	/*
3442	 * for now limit to current task, which is enough when calling
3443	 * from overflow handler
3444	 */
3445	if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3446
3447	return pfm_read_pmds(ctx, req, nreq, regs);
3448}
3449EXPORT_SYMBOL(pfm_mod_read_pmds);
3450
3451/*
3452 * Only call this function when a process it trying to
3453 * write the debug registers (reading is always allowed)
3454 */
3455int
3456pfm_use_debug_registers(struct task_struct *task)
3457{
3458	pfm_context_t *ctx = task->thread.pfm_context;
3459	unsigned long flags;
3460	int ret = 0;
3461
3462	if (pmu_conf->use_rr_dbregs == 0) return 0;
3463
3464	DPRINT(("called for [%d]\n", task_pid_nr(task)));
3465
3466	/*
3467	 * do it only once
3468	 */
3469	if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3470
3471	/*
3472	 * Even on SMP, we do not need to use an atomic here because
3473	 * the only way in is via ptrace() and this is possible only when the
3474	 * process is stopped. Even in the case where the ctxsw out is not totally
3475	 * completed by the time we come here, there is no way the 'stopped' process
3476	 * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3477	 * So this is always safe.
3478	 */
3479	if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3480
3481	LOCK_PFS(flags);
3482
3483	/*
3484	 * We cannot allow setting breakpoints when system wide monitoring
3485	 * sessions are using the debug registers.
3486	 */
3487	if (pfm_sessions.pfs_sys_use_dbregs> 0)
3488		ret = -1;
3489	else
3490		pfm_sessions.pfs_ptrace_use_dbregs++;
3491
3492	DPRINT(("ptrace_use_dbregs=%u  sys_use_dbregs=%u by [%d] ret = %d\n",
3493		  pfm_sessions.pfs_ptrace_use_dbregs,
3494		  pfm_sessions.pfs_sys_use_dbregs,
3495		  task_pid_nr(task), ret));
3496
3497	UNLOCK_PFS(flags);
3498
3499	return ret;
3500}
3501
3502/*
3503 * This function is called for every task that exits with the
3504 * IA64_THREAD_DBG_VALID set. This indicates a task which was
3505 * able to use the debug registers for debugging purposes via
3506 * ptrace(). Therefore we know it was not using them for
3507 * performance monitoring, so we only decrement the number
3508 * of "ptraced" debug register users to keep the count up to date
3509 */
3510int
3511pfm_release_debug_registers(struct task_struct *task)
3512{
3513	unsigned long flags;
3514	int ret;
3515
3516	if (pmu_conf->use_rr_dbregs == 0) return 0;
3517
3518	LOCK_PFS(flags);
3519	if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3520		printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task_pid_nr(task));
3521		ret = -1;
3522	}  else {
3523		pfm_sessions.pfs_ptrace_use_dbregs--;
3524		ret = 0;
3525	}
3526	UNLOCK_PFS(flags);
3527
3528	return ret;
3529}
3530
3531static int
3532pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3533{
3534	struct task_struct *task;
3535	pfm_buffer_fmt_t *fmt;
3536	pfm_ovfl_ctrl_t rst_ctrl;
3537	int state, is_system;
3538	int ret = 0;
3539
3540	state     = ctx->ctx_state;
3541	fmt       = ctx->ctx_buf_fmt;
3542	is_system = ctx->ctx_fl_system;
3543	task      = PFM_CTX_TASK(ctx);
3544
3545	switch(state) {
3546		case PFM_CTX_MASKED:
3547			break;
3548		case PFM_CTX_LOADED: 
3549			if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3550			/* fall through */
3551		case PFM_CTX_UNLOADED:
3552		case PFM_CTX_ZOMBIE:
3553			DPRINT(("invalid state=%d\n", state));
3554			return -EBUSY;
3555		default:
3556			DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3557			return -EINVAL;
3558	}
3559
3560	/*
3561 	 * In system wide and when the context is loaded, access can only happen
3562 	 * when the caller is running on the CPU being monitored by the session.
3563 	 * It does not have to be the owner (ctx_task) of the context per se.
3564 	 */
3565	if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3566		DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3567		return -EBUSY;
3568	}
3569
3570	/* sanity check */
3571	if (unlikely(task == NULL)) {
3572		printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", task_pid_nr(current));
3573		return -EINVAL;
3574	}
3575
3576	if (task == current || is_system) {
3577
3578		fmt = ctx->ctx_buf_fmt;
3579
3580		DPRINT(("restarting self %d ovfl=0x%lx\n",
3581			task_pid_nr(task),
3582			ctx->ctx_ovfl_regs[0]));
3583
3584		if (CTX_HAS_SMPL(ctx)) {
3585
3586			prefetch(ctx->ctx_smpl_hdr);
3587
3588			rst_ctrl.bits.mask_monitoring = 0;
3589			rst_ctrl.bits.reset_ovfl_pmds = 0;
3590
3591			if (state == PFM_CTX_LOADED)
3592				ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3593			else
3594				ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3595		} else {
3596			rst_ctrl.bits.mask_monitoring = 0;
3597			rst_ctrl.bits.reset_ovfl_pmds = 1;
3598		}
3599
3600		if (ret == 0) {
3601			if (rst_ctrl.bits.reset_ovfl_pmds)
3602				pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3603
3604			if (rst_ctrl.bits.mask_monitoring == 0) {
3605				DPRINT(("resuming monitoring for [%d]\n", task_pid_nr(task)));
3606
3607				if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3608			} else {
3609				DPRINT(("keeping monitoring stopped for [%d]\n", task_pid_nr(task)));
3610
3611				// cannot use pfm_stop_monitoring(task, regs);
3612			}
3613		}
3614		/*
3615		 * clear overflowed PMD mask to remove any stale information
3616		 */
3617		ctx->ctx_ovfl_regs[0] = 0UL;
3618
3619		/*
3620		 * back to LOADED state
3621		 */
3622		ctx->ctx_state = PFM_CTX_LOADED;
3623
3624		/*
3625		 * XXX: not really useful for self monitoring
3626		 */
3627		ctx->ctx_fl_can_restart = 0;
3628
3629		return 0;
3630	}
3631
3632	/* 
3633	 * restart another task
3634	 */
3635
3636	/*
3637	 * When PFM_CTX_MASKED, we cannot issue a restart before the previous 
3638	 * one is seen by the task.
3639	 */
3640	if (state == PFM_CTX_MASKED) {
3641		if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3642		/*
3643		 * will prevent subsequent restart before this one is
3644		 * seen by other task
3645		 */
3646		ctx->ctx_fl_can_restart = 0;
3647	}
3648
3649	/*
3650	 * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3651	 * the task is blocked or on its way to block. That's the normal
3652	 * restart path. If the monitoring is not masked, then the task
3653	 * can be actively monitoring and we cannot directly intervene.
3654	 * Therefore we use the trap mechanism to catch the task and
3655	 * force it to reset the buffer/reset PMDs.
3656	 *
3657	 * if non-blocking, then we ensure that the task will go into
3658	 * pfm_handle_work() before returning to user mode.
3659	 *
3660	 * We cannot explicitly reset another task, it MUST always
3661	 * be done by the task itself. This works for system wide because
3662	 * the tool that is controlling the session is logically doing 
3663	 * "self-monitoring".
3664	 */
3665	if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3666		DPRINT(("unblocking [%d]\n", task_pid_nr(task)));
3667		complete(&ctx->ctx_restart_done);
3668	} else {
3669		DPRINT(("[%d] armed exit trap\n", task_pid_nr(task)));
3670
3671		ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3672
3673		PFM_SET_WORK_PENDING(task, 1);
3674
3675		set_notify_resume(task);
3676
3677		/*
3678		 * XXX: send reschedule if task runs on another CPU
3679		 */
3680	}
3681	return 0;
3682}
3683
3684static int
3685pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3686{
3687	unsigned int m = *(unsigned int *)arg;
3688
3689	pfm_sysctl.debug = m == 0 ? 0 : 1;
3690
3691	printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3692
3693	if (m == 0) {
3694		memset(pfm_stats, 0, sizeof(pfm_stats));
3695		for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3696	}
3697	return 0;
3698}
3699
3700/*
3701 * arg can be NULL and count can be zero for this function
3702 */
3703static int
3704pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3705{
3706	struct thread_struct *thread = NULL;
3707	struct task_struct *task;
3708	pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3709	unsigned long flags;
3710	dbreg_t dbreg;
3711	unsigned int rnum;
3712	int first_time;
3713	int ret = 0, state;
3714	int i, can_access_pmu = 0;
3715	int is_system, is_loaded;
3716
3717	if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3718
3719	state     = ctx->ctx_state;
3720	is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3721	is_system = ctx->ctx_fl_system;
3722	task      = ctx->ctx_task;
3723
3724	if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3725
3726	/*
3727	 * on both UP and SMP, we can only write to the PMC when the task is
3728	 * the owner of the local PMU.
3729	 */
3730	if (is_loaded) {
3731		thread = &task->thread;
3732		/*
3733		 * In system wide and when the context is loaded, access can only happen
3734		 * when the caller is running on the CPU being monitored by the session.
3735		 * It does not have to be the owner (ctx_task) of the context per se.
3736		 */
3737		if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3738			DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3739			return -EBUSY;
3740		}
3741		can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3742	}
3743
3744	/*
3745	 * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3746	 * ensuring that no real breakpoint can be installed via this call.
3747	 *
3748	 * IMPORTANT: regs can be NULL in this function
3749	 */
3750
3751	first_time = ctx->ctx_fl_using_dbreg == 0;
3752
3753	/*
3754	 * don't bother if we are loaded and task is being debugged
3755	 */
3756	if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3757		DPRINT(("debug registers already in use for [%d]\n", task_pid_nr(task)));
3758		return -EBUSY;
3759	}
3760
3761	/*
3762	 * check for debug registers in system wide mode
3763	 *
3764	 * If though a check is done in pfm_context_load(),
3765	 * we must repeat it here, in case the registers are
3766	 * written after the context is loaded
3767	 */
3768	if (is_loaded) {
3769		LOCK_PFS(flags);
3770
3771		if (first_time && is_system) {
3772			if (pfm_sessions.pfs_ptrace_use_dbregs)
3773				ret = -EBUSY;
3774			else
3775				pfm_sessions.pfs_sys_use_dbregs++;
3776		}
3777		UNLOCK_PFS(flags);
3778	}
3779
3780	if (ret != 0) return ret;
3781
3782	/*
3783	 * mark ourself as user of the debug registers for
3784	 * perfmon purposes.
3785	 */
3786	ctx->ctx_fl_using_dbreg = 1;
3787
3788	/*
3789 	 * clear hardware registers to make sure we don't
3790 	 * pick up stale state.
3791	 *
3792	 * for a system wide session, we do not use
3793	 * thread.dbr, thread.ibr because this process
3794	 * never leaves the current CPU and the state
3795	 * is shared by all processes running on it
3796 	 */
3797	if (first_time && can_access_pmu) {
3798		DPRINT(("[%d] clearing ibrs, dbrs\n", task_pid_nr(task)));
3799		for (i=0; i < pmu_conf->num_ibrs; i++) {
3800			ia64_set_ibr(i, 0UL);
3801			ia64_dv_serialize_instruction();
3802		}
3803		ia64_srlz_i();
3804		for (i=0; i < pmu_conf->num_dbrs; i++) {
3805			ia64_set_dbr(i, 0UL);
3806			ia64_dv_serialize_data();
3807		}
3808		ia64_srlz_d();
3809	}
3810
3811	/*
3812	 * Now install the values into the registers
3813	 */
3814	for (i = 0; i < count; i++, req++) {
3815
3816		rnum      = req->dbreg_num;
3817		dbreg.val = req->dbreg_value;
3818
3819		ret = -EINVAL;
3820
3821		if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3822			DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3823				  rnum, dbreg.val, mode, i, count));
3824
3825			goto abort_mission;
3826		}
3827
3828		/*
3829		 * make sure we do not install enabled breakpoint
3830		 */
3831		if (rnum & 0x1) {
3832			if (mode == PFM_CODE_RR)
3833				dbreg.ibr.ibr_x = 0;
3834			else
3835				dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3836		}
3837
3838		PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3839
3840		/*
3841		 * Debug registers, just like PMC, can only be modified
3842		 * by a kernel call. Moreover, perfmon() access to those
3843		 * registers are centralized in this routine. The hardware
3844		 * does not modify the value of these registers, therefore,
3845		 * if we save them as they are written, we can avoid having
3846		 * to save them on context switch out. This is made possible
3847		 * by the fact that when perfmon uses debug registers, ptrace()
3848		 * won't be able to modify them concurrently.
3849		 */
3850		if (mode == PFM_CODE_RR) {
3851			CTX_USED_IBR(ctx, rnum);
3852
3853			if (can_access_pmu) {
3854				ia64_set_ibr(rnum, dbreg.val);
3855				ia64_dv_serialize_instruction();
3856			}
3857
3858			ctx->ctx_ibrs[rnum] = dbreg.val;
3859
3860			DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3861				rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3862		} else {
3863			CTX_USED_DBR(ctx, rnum);
3864
3865			if (can_access_pmu) {
3866				ia64_set_dbr(rnum, dbreg.val);
3867				ia64_dv_serialize_data();
3868			}
3869			ctx->ctx_dbrs[rnum] = dbreg.val;
3870
3871			DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3872				rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3873		}
3874	}
3875
3876	return 0;
3877
3878abort_mission:
3879	/*
3880	 * in case it was our first attempt, we undo the global modifications
3881	 */
3882	if (first_time) {
3883		LOCK_PFS(flags);
3884		if (ctx->ctx_fl_system) {
3885			pfm_sessions.pfs_sys_use_dbregs--;
3886		}
3887		UNLOCK_PFS(flags);
3888		ctx->ctx_fl_using_dbreg = 0;
3889	}
3890	/*
3891	 * install error return flag
3892	 */
3893	PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3894
3895	return ret;
3896}
3897
3898static int
3899pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3900{
3901	return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3902}
3903
3904static int
3905pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3906{
3907	return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3908}
3909
3910int
3911pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3912{
3913	pfm_context_t *ctx;
3914
3915	if (req == NULL) return -EINVAL;
3916
3917 	ctx = GET_PMU_CTX();
3918
3919	if (ctx == NULL) return -EINVAL;
3920
3921	/*
3922	 * for now limit to current task, which is enough when calling
3923	 * from overflow handler
3924	 */
3925	if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3926
3927	return pfm_write_ibrs(ctx, req, nreq, regs);
3928}
3929EXPORT_SYMBOL(pfm_mod_write_ibrs);
3930
3931int
3932pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3933{
3934	pfm_context_t *ctx;
3935
3936	if (req == NULL) return -EINVAL;
3937
3938 	ctx = GET_PMU_CTX();
3939
3940	if (ctx == NULL) return -EINVAL;
3941
3942	/*
3943	 * for now limit to current task, which is enough when calling
3944	 * from overflow handler
3945	 */
3946	if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3947
3948	return pfm_write_dbrs(ctx, req, nreq, regs);
3949}
3950EXPORT_SYMBOL(pfm_mod_write_dbrs);
3951
3952
3953static int
3954pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3955{
3956	pfarg_features_t *req = (pfarg_features_t *)arg;
3957
3958	req->ft_version = PFM_VERSION;
3959	return 0;
3960}
3961
3962static int
3963pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3964{
3965	struct pt_regs *tregs;
3966	struct task_struct *task = PFM_CTX_TASK(ctx);
3967	int state, is_system;
3968
3969	state     = ctx->ctx_state;
3970	is_system = ctx->ctx_fl_system;
3971
3972	/*
3973	 * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3974	 */
3975	if (state == PFM_CTX_UNLOADED) return -EINVAL;
3976
3977	/*
3978 	 * In system wide and when the context is loaded, access can only happen
3979 	 * when the caller is running on the CPU being monitored by the session.
3980 	 * It does not have to be the owner (ctx_task) of the context per se.
3981 	 */
3982	if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3983		DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3984		return -EBUSY;
3985	}
3986	DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
3987		task_pid_nr(PFM_CTX_TASK(ctx)),
3988		state,
3989		is_system));
3990	/*
3991	 * in system mode, we need to update the PMU directly
3992	 * and the user level state of the caller, which may not
3993	 * necessarily be the creator of the context.
3994	 */
3995	if (is_system) {
3996		/*
3997		 * Update local PMU first
3998		 *
3999		 * disable dcr pp
4000		 */
4001		ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
4002		ia64_srlz_i();
4003
4004		/*
4005		 * update local cpuinfo
4006		 */
4007		PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4008
4009		/*
4010		 * stop monitoring, does srlz.i
4011		 */
4012		pfm_clear_psr_pp();
4013
4014		/*
4015		 * stop monitoring in the caller
4016		 */
4017		ia64_psr(regs)->pp = 0;
4018
4019		return 0;
4020	}
4021	/*
4022	 * per-task mode
4023	 */
4024
4025	if (task == current) {
4026		/* stop monitoring  at kernel level */
4027		pfm_clear_psr_up();
4028
4029		/*
4030	 	 * stop monitoring at the user level
4031	 	 */
4032		ia64_psr(regs)->up = 0;
4033	} else {
4034		tregs = task_pt_regs(task);
4035
4036		/*
4037	 	 * stop monitoring at the user level
4038	 	 */
4039		ia64_psr(tregs)->up = 0;
4040
4041		/*
4042		 * monitoring disabled in kernel at next reschedule
4043		 */
4044		ctx->ctx_saved_psr_up = 0;
4045		DPRINT(("task=[%d]\n", task_pid_nr(task)));
4046	}
4047	return 0;
4048}
4049
4050
4051static int
4052pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4053{
4054	struct pt_regs *tregs;
4055	int state, is_system;
4056
4057	state     = ctx->ctx_state;
4058	is_system = ctx->ctx_fl_system;
4059
4060	if (state != PFM_CTX_LOADED) return -EINVAL;
4061
4062	/*
4063 	 * In system wide and when the context is loaded, access can only happen
4064 	 * when the caller is running on the CPU being monitored by the session.
4065 	 * It does not have to be the owner (ctx_task) of the context per se.
4066 	 */
4067	if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4068		DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4069		return -EBUSY;
4070	}
4071
4072	/*
4073	 * in system mode, we need to update the PMU directly
4074	 * and the user level state of the caller, which may not
4075	 * necessarily be the creator of the context.
4076	 */
4077	if (is_system) {
4078
4079		/*
4080		 * set user level psr.pp for the caller
4081		 */
4082		ia64_psr(regs)->pp = 1;
4083
4084		/*
4085		 * now update the local PMU and cpuinfo
4086		 */
4087		PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4088
4089		/*
4090		 * start monitoring at kernel level
4091		 */
4092		pfm_set_psr_pp();
4093
4094		/* enable dcr pp */
4095		ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4096		ia64_srlz_i();
4097
4098		return 0;
4099	}
4100
4101	/*
4102	 * per-process mode
4103	 */
4104
4105	if (ctx->ctx_task == current) {
4106
4107		/* start monitoring at kernel level */
4108		pfm_set_psr_up();
4109
4110		/*
4111		 * activate monitoring at user level
4112		 */
4113		ia64_psr(regs)->up = 1;
4114
4115	} else {
4116		tregs = task_pt_regs(ctx->ctx_task);
4117
4118		/*
4119		 * start monitoring at the kernel level the next
4120		 * time the task is scheduled
4121		 */
4122		ctx->ctx_saved_psr_up = IA64_PSR_UP;
4123
4124		/*
4125		 * activate monitoring at user level
4126		 */
4127		ia64_psr(tregs)->up = 1;
4128	}
4129	return 0;
4130}
4131
4132static int
4133pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4134{
4135	pfarg_reg_t *req = (pfarg_reg_t *)arg;
4136	unsigned int cnum;
4137	int i;
4138	int ret = -EINVAL;
4139
4140	for (i = 0; i < count; i++, req++) {
4141
4142		cnum = req->reg_num;
4143
4144		if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4145
4146		req->reg_value = PMC_DFL_VAL(cnum);
4147
4148		PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4149
4150		DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4151	}
4152	return 0;
4153
4154abort_mission:
4155	PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4156	return ret;
4157}
4158
4159static int
4160pfm_check_task_exist(pfm_context_t *ctx)
4161{
4162	struct task_struct *g, *t;
4163	int ret = -ESRCH;
4164
4165	read_lock(&tasklist_lock);
4166
4167	do_each_thread (g, t) {
4168		if (t->thread.pfm_context == ctx) {
4169			ret = 0;
4170			goto out;
4171		}
4172	} while_each_thread (g, t);
4173out:
4174	read_unlock(&tasklist_lock);
4175
4176	DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4177
4178	return ret;
4179}
4180
4181static int
4182pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4183{
4184	struct task_struct *task;
4185	struct thread_struct *thread;
4186	struct pfm_context_t *old;
4187	unsigned long flags;
4188#ifndef CONFIG_SMP
4189	struct task_struct *owner_task = NULL;
4190#endif
4191	pfarg_load_t *req = (pfarg_load_t *)arg;
4192	unsigned long *pmcs_source, *pmds_source;
4193	int the_cpu;
4194	int ret = 0;
4195	int state, is_system, set_dbregs = 0;
4196
4197	state     = ctx->ctx_state;
4198	is_system = ctx->ctx_fl_system;
4199	/*
4200	 * can only load from unloaded or terminated state
4201	 */
4202	if (state != PFM_CTX_UNLOADED) {
4203		DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4204			req->load_pid,
4205			ctx->ctx_state));
4206		return -EBUSY;
4207	}
4208
4209	DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4210
4211	if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4212		DPRINT(("cannot use blocking mode on self\n"));
4213		return -EINVAL;
4214	}
4215
4216	ret = pfm_get_task(ctx, req->load_pid, &task);
4217	if (ret) {
4218		DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4219		return ret;
4220	}
4221
4222	ret = -EINVAL;
4223
4224	/*
4225	 * system wide is self monitoring only
4226	 */
4227	if (is_system && task != current) {
4228		DPRINT(("system wide is self monitoring only load_pid=%d\n",
4229			req->load_pid));
4230		goto error;
4231	}
4232
4233	thread = &task->thread;
4234
4235	ret = 0;
4236	/*
4237	 * cannot load a context which is using range restrictions,
4238	 * into a task that is being debugged.
4239	 */
4240	if (ctx->ctx_fl_using_dbreg) {
4241		if (thread->flags & IA64_THREAD_DBG_VALID) {
4242			ret = -EBUSY;
4243			DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4244			goto error;
4245		}
4246		LOCK_PFS(flags);
4247
4248		if (is_system) {
4249			if (pfm_sessions.pfs_ptrace_use_dbregs) {
4250				DPRINT(("cannot load [%d] dbregs in use\n",
4251							task_pid_nr(task)));
4252				ret = -EBUSY;
4253			} else {
4254				pfm_sessions.pfs_sys_use_dbregs++;
4255				DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task_pid_nr(task), pfm_sessions.pfs_sys_use_dbregs));
4256				set_dbregs = 1;
4257			}
4258		}
4259
4260		UNLOCK_PFS(flags);
4261
4262		if (ret) goto error;
4263	}
4264
4265	/*
4266	 * SMP system-wide monitoring implies self-monitoring.
4267	 *
4268	 * The programming model expects the task to
4269	 * be pinned on a CPU throughout the session.
4270	 * Here we take note of the current CPU at the
4271	 * time the context is loaded. No call from
4272	 * another CPU will be allowed.
4273	 *
4274	 * The pinning via shed_setaffinity()
4275	 * must be done by the calling task prior
4276	 * to this call.
4277	 *
4278	 * systemwide: keep track of CPU this session is supposed to run on
4279	 */
4280	the_cpu = ctx->ctx_cpu = smp_processor_id();
4281
4282	ret = -EBUSY;
4283	/*
4284	 * now reserve the session
4285	 */
4286	ret = pfm_reserve_session(current, is_system, the_cpu);
4287	if (ret) goto error;
4288
4289	/*
4290	 * task is necessarily stopped at this point.
4291	 *
4292	 * If the previous context was zombie, then it got removed in
4293	 * pfm_save_regs(). Therefore we should not see it here.
4294	 * If we see a context, then this is an active context
4295	 *
4296	 * XXX: needs to be atomic
4297	 */
4298	DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4299		thread->pfm_context, ctx));
4300
4301	ret = -EBUSY;
4302	old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4303	if (old != NULL) {
4304		DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4305		goto error_unres;
4306	}
4307
4308	pfm_reset_msgq(ctx);
4309
4310	ctx->ctx_state = PFM_CTX_LOADED;
4311
4312	/*
4313	 * link context to task
4314	 */
4315	ctx->ctx_task = task;
4316
4317	if (is_system) {
4318		/*
4319		 * we load as stopped
4320		 */
4321		PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4322		PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4323
4324		if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4325	} else {
4326		thread->flags |= IA64_THREAD_PM_VALID;
4327	}
4328
4329	/*
4330	 * propagate into thread-state
4331	 */
4332	pfm_copy_pmds(task, ctx);
4333	pfm_copy_pmcs(task, ctx);
4334
4335	pmcs_source = ctx->th_pmcs;
4336	pmds_source = ctx->th_pmds;
4337
4338	/*
4339	 * always the case for system-wide
4340	 */
4341	if (task == current) {
4342
4343		if (is_system == 0) {
4344
4345			/* allow user level control */
4346			ia64_psr(regs)->sp = 0;
4347			DPRINT(("clearing psr.sp for [%d]\n", task_pid_nr(task)));
4348
4349			SET_LAST_CPU(ctx, smp_processor_id());
4350			INC_ACTIVATION();
4351			SET_ACTIVATION(ctx);
4352#ifndef CONFIG_SMP
4353			/*
4354			 * push the other task out, if any
4355			 */
4356			owner_task = GET_PMU_OWNER();
4357			if (owner_task) pfm_lazy_save_regs(owner_task);
4358#endif
4359		}
4360		/*
4361		 * load all PMD from ctx to PMU (as opposed to thread state)
4362		 * restore all PMC from ctx to PMU
4363		 */
4364		pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4365		pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4366
4367		ctx->ctx_reload_pmcs[0] = 0UL;
4368		ctx->ctx_reload_pmds[0] = 0UL;
4369
4370		/*
4371		 * guaranteed safe by earlier check against DBG_VALID
4372		 */
4373		if (ctx->ctx_fl_using_dbreg) {
4374			pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4375			pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4376		}
4377		/*
4378		 * set new ownership
4379		 */
4380		SET_PMU_OWNER(task, ctx);
4381
4382		DPRINT(("context loaded on PMU for [%d]\n", task_pid_nr(task)));
4383	} else {
4384		/*
4385		 * when not current, task MUST be stopped, so this is safe
4386		 */
4387		regs = task_pt_regs(task);
4388
4389		/* force a full reload */
4390		ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4391		SET_LAST_CPU(ctx, -1);
4392
4393		/* initial saved psr (stopped) */
4394		ctx->ctx_saved_psr_up = 0UL;
4395		ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4396	}
4397
4398	ret = 0;
4399
4400error_unres:
4401	if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4402error:
4403	/*
4404	 * we must undo the dbregs setting (for system-wide)
4405	 */
4406	if (ret && set_dbregs) {
4407		LOCK_PFS(flags);
4408		pfm_sessions.pfs_sys_use_dbregs--;
4409		UNLOCK_PFS(flags);
4410	}
4411	/*
4412	 * release task, there is now a link with the context
4413	 */
4414	if (is_system == 0 && task != current) {
4415		pfm_put_task(task);
4416
4417		if (ret == 0) {
4418			ret = pfm_check_task_exist(ctx);
4419			if (ret) {
4420				ctx->ctx_state = PFM_CTX_UNLOADED;
4421				ctx->ctx_task  = NULL;
4422			}
4423		}
4424	}
4425	return ret;
4426}
4427
4428/*
4429 * in this function, we do not need to increase the use count
4430 * for the task via get_task_struct(), because we hold the
4431 * context lock. If the task were to disappear while having
4432 * a context attached, it would go through pfm_exit_thread()
4433 * which also grabs the context lock  and would therefore be blocked
4434 * until we are here.
4435 */
4436static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4437
4438static int
4439pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4440{
4441	struct task_struct *task = PFM_CTX_TASK(ctx);
4442	struct pt_regs *tregs;
4443	int prev_state, is_system;
4444	int ret;
4445
4446	DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task_pid_nr(task) : -1));
4447
4448	prev_state = ctx->ctx_state;
4449	is_system  = ctx->ctx_fl_system;
4450
4451	/*
4452	 * unload only when necessary
4453	 */
4454	if (prev_state == PFM_CTX_UNLOADED) {
4455		DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4456		return 0;
4457	}
4458
4459	/*
4460	 * clear psr and dcr bits
4461	 */
4462	ret = pfm_stop(ctx, NULL, 0, regs);
4463	if (ret) return ret;
4464
4465	ctx->ctx_state = PFM_CTX_UNLOADED;
4466
4467	/*
4468	 * in system mode, we need to update the PMU directly
4469	 * and the user level state of the caller, which may not
4470	 * necessarily be the creator of the context.
4471	 */
4472	if (is_system) {
4473
4474		/*
4475		 * Update cpuinfo
4476		 *
4477		 * local PMU is taken care of in pfm_stop()
4478		 */
4479		PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4480		PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4481
4482		/*
4483		 * save PMDs in context
4484		 * release ownership
4485		 */
4486		pfm_flush_pmds(current, ctx);
4487
4488		/*
4489		 * at this point we are done with the PMU
4490		 * so we can unreserve the resource.
4491		 */
4492		if (prev_state != PFM_CTX_ZOMBIE) 
4493			pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4494
4495		/*
4496		 * disconnect context from task
4497		 */
4498		task->thread.pfm_context = NULL;
4499		/*
4500		 * disconnect task from context
4501		 */
4502		ctx->ctx_task = NULL;
4503
4504		/*
4505		 * There is nothing more to cleanup here.
4506		 */
4507		return 0;
4508	}
4509
4510	/*
4511	 * per-task mode
4512	 */
4513	tregs = task == current ? regs : task_pt_regs(task);
4514
4515	if (task == current) {
4516		/*
4517		 * cancel user level control
4518		 */
4519		ia64_psr(regs)->sp = 1;
4520
4521		DPRINT(("setting psr.sp for [%d]\n", task_pid_nr(task)));
4522	}
4523	/*
4524	 * save PMDs to context
4525	 * release ownership
4526	 */
4527	pfm_flush_pmds(task, ctx);
4528
4529	/*
4530	 * at this point we are done with the PMU
4531	 * so we can unreserve the resource.
4532	 *
4533	 * when state was ZOMBIE, we have already unreserved.
4534	 */
4535	if (prev_state != PFM_CTX_ZOMBIE) 
4536		pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4537
4538	/*
4539	 * reset activation counter and psr
4540	 */
4541	ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4542	SET_LAST_CPU(ctx, -1);
4543
4544	/*
4545	 * PMU state will not be restored
4546	 */
4547	task->thread.flags &= ~IA64_THREAD_PM_VALID;
4548
4549	/*
4550	 * break links between context and task
4551	 */
4552	task->thread.pfm_context  = NULL;
4553	ctx->ctx_task             = NULL;
4554
4555	PFM_SET_WORK_PENDING(task, 0);
4556
4557	ctx->ctx_fl_trap_reason  = PFM_TRAP_REASON_NONE;
4558	ctx->ctx_fl_can_restart  = 0;
4559	ctx->ctx_fl_going_zombie = 0;
4560
4561	DPRINT(("disconnected [%d] from context\n", task_pid_nr(task)));
4562
4563	return 0;
4564}
4565
4566
4567/*
4568 * called only from exit_thread(): task == current
4569 * we come here only if current has a context attached (loaded or masked)
4570 */
4571void
4572pfm_exit_thread(struct task_struct *task)
4573{
4574	pfm_context_t *ctx;
4575	unsigned long flags;
4576	struct pt_regs *regs = task_pt_regs(task);
4577	int ret, state;
4578	int free_ok = 0;
4579
4580	ctx = PFM_GET_CTX(task);
4581
4582	PROTECT_CTX(ctx, flags);
4583
4584	DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task_pid_nr(task)));
4585
4586	state = ctx->ctx_state;
4587	switch(state) {
4588		case PFM_CTX_UNLOADED:
4589			/*
4590	 		 * only comes to this function if pfm_context is not NULL, i.e., cannot
4591			 * be in unloaded state
4592	 		 */
4593			printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task_pid_nr(task));
4594			break;
4595		case PFM_CTX_LOADED:
4596		case PFM_CTX_MASKED:
4597			ret = pfm_context_unload(ctx, NULL, 0, regs);
4598			if (ret) {
4599				printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4600			}
4601			DPRINT(("ctx unloaded for current state was %d\n", state));
4602
4603			pfm_end_notify_user(ctx);
4604			break;
4605		case PFM_CTX_ZOMBIE:
4606			ret = pfm_context_unload(ctx, NULL, 0, regs);
4607			if (ret) {
4608				printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4609			}
4610			free_ok = 1;
4611			break;
4612		default:
4613			printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task_pid_nr(task), state);
4614			break;
4615	}
4616	UNPROTECT_CTX(ctx, flags);
4617
4618	{ u64 psr = pfm_get_psr();
4619	  BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4620	  BUG_ON(GET_PMU_OWNER());
4621	  BUG_ON(ia64_psr(regs)->up);
4622	  BUG_ON(ia64_psr(regs)->pp);
4623	}
4624
4625	/*
4626	 * All memory free operations (especially for vmalloc'ed memory)
4627	 * MUST be done with interrupts ENABLED.
4628	 */
4629	if (free_ok) pfm_context_free(ctx);
4630}
4631
4632/*
4633 * functions MUST be listed in the increasing order of their index (see permfon.h)
4634 */
4635#define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4636#define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4637#define PFM_CMD_PCLRWS	(PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4638#define PFM_CMD_PCLRW	(PFM_CMD_FD|PFM_CMD_ARG_RW)
4639#define PFM_CMD_NONE	{ NULL, "no-cmd", 0, 0, 0, NULL}
4640
4641static pfm_cmd_desc_t pfm_cmd_tab[]={
4642/* 0  */PFM_CMD_NONE,
4643/* 1  */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4644/* 2  */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4645/* 3  */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4646/* 4  */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4647/* 5  */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4648/* 6  */PFM_CMD_NONE,
4649/* 7  */PFM_CMD_NONE,
4650/* 8  */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4651/* 9  */PFM_CMD_NONE,
4652/* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4653/* 11 */PFM_CMD_NONE,
4654/* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4655/* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4656/* 14 */PFM_CMD_NONE,
4657/* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4658/* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4659/* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4660/* 18 */PFM_CMD_NONE,
4661/* 19 */PFM_CMD_NONE,
4662/* 20 */PFM_CMD_NONE,
4663/* 21 */PFM_CMD_NONE,
4664/* 22 */PFM_CMD_NONE,
4665/* 23 */PFM_CMD_NONE,
4666/* 24 */PFM_CMD_NONE,
4667/* 25 */PFM_CMD_NONE,
4668/* 26 */PFM_CMD_NONE,
4669/* 27 */PFM_CMD_NONE,
4670/* 28 */PFM_CMD_NONE,
4671/* 29 */PFM_CMD_NONE,
4672/* 30 */PFM_CMD_NONE,
4673/* 31 */PFM_CMD_NONE,
4674/* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4675/* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4676};
4677#define PFM_CMD_COUNT	(sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4678
4679static int
4680pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4681{
4682	struct task_struct *task;
4683	int state, old_state;
4684
4685recheck:
4686	state = ctx->ctx_state;
4687	task  = ctx->ctx_task;
4688
4689	if (task == NULL) {
4690		DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4691		return 0;
4692	}
4693
4694	DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4695		ctx->ctx_fd,
4696		state,
4697		task_pid_nr(task),
4698		task->state, PFM_CMD_STOPPED(cmd)));
4699
4700	/*
4701	 * self-monitoring always ok.
4702	 *
4703	 * for system-wide the caller can either be the creator of the
4704	 * context (to one to which the context is attached to) OR
4705	 * a task running on the same CPU as the session.
4706	 */
4707	if (task == current || ctx->ctx_fl_system) return 0;
4708
4709	/*
4710	 * we are monitoring another thread
4711	 */
4712	switch(state) {
4713		case PFM_CTX_UNLOADED:
4714			/*
4715			 * if context is UNLOADED we are safe to go
4716			 */
4717			return 0;
4718		case PFM_CTX_ZOMBIE:
4719			/*
4720			 * no command can operate on a zombie context
4721			 */
4722			DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4723			return -EINVAL;
4724		case PFM_CTX_MASKED:
4725			/*
4726			 * PMU state has been saved to software even though
4727			 * the thread may still be running.
4728			 */
4729			if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4730	}
4731
4732	/*
4733	 * context is LOADED or MASKED. Some commands may need to have 
4734	 * the task stopped.
4735	 *
4736	 * We could lift this restriction for UP but it would mean that
4737	 * the user has no guarantee the task would not run between
4738	 * two successive calls to perfmonctl(). That's probably OK.
4739	 * If this user wants to ensure the task does not run, then
4740	 * the task must be stopped.
4741	 */
4742	if (PFM_CMD_STOPPED(cmd)) {
4743		if (!task_is_stopped_or_traced(task)) {
4744			DPRINT(("[%d] task not in stopped state\n", task_pid_nr(task)));
4745			return -EBUSY;
4746		}
4747		/*
4748		 * task is now stopped, wait for ctxsw out
4749		 *
4750		 * This is an interesting point in the code.
4751		 * We need to unprotect the context because
4752		 * the pfm_save_regs() routines needs to grab
4753		 * the same lock. There are danger in doing
4754		 * this because it leaves a window open for
4755		 * another task to get access to the context
4756		 * and possibly change its state. The one thing
4757		 * that is not possible is for the context to disappear
4758		 * because we are protected by the VFS layer, i.e.,
4759		 * get_fd()/put_fd().
4760		 */
4761		old_state = state;
4762
4763		UNPROTECT_CTX(ctx, flags);
4764
4765		wait_task_inactive(task, 0);
4766
4767		PROTECT_CTX(ctx, flags);
4768
4769		/*
4770		 * we must recheck to verify if state has changed
4771		 */
4772		if (ctx->ctx_state != old_state) {
4773			DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4774			goto recheck;
4775		}
4776	}
4777	return 0;
4778}
4779
4780/*
4781 * system-call entry point (must return long)
4782 */
4783asmlinkage long
4784sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4785{
4786	struct file *file = NULL;
4787	pfm_context_t *ctx = NULL;
4788	unsigned long flags = 0UL;
4789	void *args_k = NULL;
4790	long ret; /* will expand int return types */
4791	size_t base_sz, sz, xtra_sz = 0;
4792	int narg, completed_args = 0, call_made = 0, cmd_flags;
4793	int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4794	int (*getsize)(void *arg, size_t *sz);
4795#define PFM_MAX_ARGSIZE	4096
4796
4797	/*
4798	 * reject any call if perfmon was disabled at initialization
4799	 */
4800	if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4801
4802	if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4803		DPRINT(("invalid cmd=%d\n", cmd));
4804		return -EINVAL;
4805	}
4806
4807	func      = pfm_cmd_tab[cmd].cmd_func;
4808	narg      = pfm_cmd_tab[cmd].cmd_narg;
4809	base_sz   = pfm_cmd_tab[cmd].cmd_argsize;
4810	getsize   = pfm_cmd_tab[cmd].cmd_getsize;
4811	cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4812
4813	if (unlikely(func == NULL)) {
4814		DPRINT(("invalid cmd=%d\n", cmd));
4815		return -EINVAL;
4816	}
4817
4818	DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4819		PFM_CMD_NAME(cmd),
4820		cmd,
4821		narg,
4822		base_sz,
4823		count));
4824
4825	/*
4826	 * check if number of arguments matches what the command expects
4827	 */
4828	if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4829		return -EINVAL;
4830
4831restart_args:
4832	sz = xtra_sz + base_sz*count;
4833	/*
4834	 * limit abuse to min page size
4835	 */
4836	if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4837		printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", task_pid_nr(current), sz);
4838		return -E2BIG;
4839	}
4840
4841	/*
4842	 * allocate default-sized argument buffer
4843	 */
4844	if (likely(count && args_k == NULL)) {
4845		args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4846		if (args_k == NULL) return -ENOMEM;
4847	}
4848
4849	ret = -EFAULT;
4850
4851	/*
4852	 * copy arguments
4853	 *
4854	 * assume sz = 0 for command without parameters
4855	 */
4856	if (sz && copy_from_user(args_k, arg, sz)) {
4857		DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4858		goto error_args;
4859	}
4860
4861	/*
4862	 * check if command supports extra parameters
4863	 */
4864	if (completed_args == 0 && getsize) {
4865		/*
4866		 * get extra parameters size (based on main argument)
4867		 */
4868		ret = (*getsize)(args_k, &xtra_sz);
4869		if (ret) goto error_args;
4870
4871		completed_args = 1;
4872
4873		DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4874
4875		/* retry if necessary */
4876		if (likely(xtra_sz)) goto restart_args;
4877	}
4878
4879	if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4880
4881	ret = -EBADF;
4882
4883	file = fget(fd);
4884	if (unlikely(file == NULL)) {
4885		DPRINT(("invalid fd %d\n", fd));
4886		goto error_args;
4887	}
4888	if (unlikely(PFM_IS_FILE(file) == 0)) {
4889		DPRINT(("fd %d not related to perfmon\n", fd));
4890		goto error_args;
4891	}
4892
4893	ctx = file->private_data;
4894	if (unlikely(ctx == NULL)) {
4895		DPRINT(("no context for fd %d\n", fd));
4896		goto error_args;
4897	}
4898	prefetch(&ctx->ctx_state);
4899
4900	PROTECT_CTX(ctx, flags);
4901
4902	/*
4903	 * check task is stopped
4904	 */
4905	ret = pfm_check_task_state(ctx, cmd, flags);
4906	if (unlikely(ret)) goto abort_locked;
4907
4908skip_fd:
4909	ret = (*func)(ctx, args_k, count, task_pt_regs(current));
4910
4911	call_made = 1;
4912
4913abort_locked:
4914	if (likely(ctx)) {
4915		DPRINT(("context unlocked\n"));
4916		UNPROTECT_CTX(ctx, flags);
4917	}
4918
4919	/* copy argument back to user, if needed */
4920	if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4921
4922error_args:
4923	if (file)
4924		fput(file);
4925
4926	kfree(args_k);
4927
4928	DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4929
4930	return ret;
4931}
4932
4933static void
4934pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4935{
4936	pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4937	pfm_ovfl_ctrl_t rst_ctrl;
4938	int state;
4939	int ret = 0;
4940
4941	state = ctx->ctx_state;
4942	/*
4943	 * Unlock sampling buffer and reset index atomically
4944	 * XXX: not really needed when blocking
4945	 */
4946	if (CTX_HAS_SMPL(ctx)) {
4947
4948		rst_ctrl.bits.mask_monitoring = 0;
4949		rst_ctrl.bits.reset_ovfl_pmds = 0;
4950
4951		if (state == PFM_CTX_LOADED)
4952			ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4953		else
4954			ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4955	} else {
4956		rst_ctrl.bits.mask_monitoring = 0;
4957		rst_ctrl.bits.reset_ovfl_pmds = 1;
4958	}
4959
4960	if (ret == 0) {
4961		if (rst_ctrl.bits.reset_ovfl_pmds) {
4962			pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4963		}
4964		if (rst_ctrl.bits.mask_monitoring == 0) {
4965			DPRINT(("resuming monitoring\n"));
4966			if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4967		} else {
4968			DPRINT(("stopping monitoring\n"));
4969			//pfm_stop_monitoring(current, regs);
4970		}
4971		ctx->ctx_state = PFM_CTX_LOADED;
4972	}
4973}
4974
4975/*
4976 * context MUST BE LOCKED when calling
4977 * can only be called for current
4978 */
4979static void
4980pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4981{
4982	int ret;
4983
4984	DPRINT(("entering for [%d]\n", task_pid_nr(current)));
4985
4986	ret = pfm_context_unload(ctx, NULL, 0, regs);
4987	if (ret) {
4988		printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", task_pid_nr(current), ret);
4989	}
4990
4991	/*
4992	 * and wakeup controlling task, indicating we are now disconnected
4993	 */
4994	wake_up_interruptible(&ctx->ctx_zombieq);
4995
4996	/*
4997	 * given that context is still locked, the controlling
4998	 * task will only get access when we return from
4999	 * pfm_handle_work().
5000	 */
5001}
5002
5003static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
5004
5005 /*
5006  * pfm_handle_work() can be called with interrupts enabled
5007  * (TIF_NEED_RESCHED) or disabled. The down_interruptible
5008  * call may sleep, therefore we must re-enable interrupts
5009  * to avoid deadlocks. It is safe to do so because this function
5010  * is called ONLY when returning to user level (pUStk=1), in which case
5011  * there is no risk of kernel stack overflow due to deep
5012  * interrupt nesting.
5013  */
5014void
5015pfm_handle_work(void)
5016{
5017	pfm_context_t *ctx;
5018	struct pt_regs *regs;
5019	unsigned long flags, dummy_flags;
5020	unsigned long ovfl_regs;
5021	unsigned int reason;
5022	int ret;
5023
5024	ctx = PFM_GET_CTX(current);
5025	if (ctx == NULL) {
5026		printk(KERN_ERR "perfmon: [%d] has no PFM context\n",
5027			task_pid_nr(current));
5028		return;
5029	}
5030
5031	PROTECT_CTX(ctx, flags);
5032
5033	PFM_SET_WORK_PENDING(current, 0);
5034
5035	regs = task_pt_regs(current);
5036
5037	/*
5038	 * extract reason for being here and clear
5039	 */
5040	reason = ctx->ctx_fl_trap_reason;
5041	ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5042	ovfl_regs = ctx->ctx_ovfl_regs[0];
5043
5044	DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5045
5046	/*
5047	 * must be done before we check for simple-reset mode
5048	 */
5049	if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE)
5050		goto do_zombie;
5051
5052	//if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5053	if (reason == PFM_TRAP_REASON_RESET)
5054		goto skip_blocking;
5055
5056	/*
5057	 * restore interrupt mask to what it was on entry.
5058	 * Could be enabled/diasbled.
5059	 */
5060	UNPROTECT_CTX(ctx, flags);
5061
5062	/*
5063	 * force interrupt enable because of down_interruptible()
5064	 */
5065	local_irq_enable();
5066
5067	DPRINT(("before block sleeping\n"));
5068
5069	/*
5070	 * may go through without blocking on SMP systems
5071	 * if restart has been received already by the time we call down()
5072	 */
5073	ret = wait_for_completion_interruptible(&ctx->ctx_restart_done);
5074
5075	DPRINT(("after block sleeping ret=%d\n", ret));
5076
5077	/*
5078	 * lock context and mask interrupts again
5079	 * We save flags into a dummy because we may have
5080	 * altered interrupts mask compared to entry in this
5081	 * function.
5082	 */
5083	PROTECT_CTX(ctx, dummy_flags);
5084
5085	/*
5086	 * we need to read the ovfl_regs only after wake-up
5087	 * because we may have had pfm_write_pmds() in between
5088	 * and that can changed PMD values and therefore 
5089	 * ovfl_regs is reset for these new PMD values.
5090	 */
5091	ovfl_regs = ctx->ctx_ovfl_regs[0];
5092
5093	if (ctx->ctx_fl_going_zombie) {
5094do_zombie:
5095		DPRINT(("context is zombie, bailing out\n"));
5096		pfm_context_force_terminate(ctx, regs);
5097		goto nothing_to_do;
5098	}
5099	/*
5100	 * in case of interruption of down() we don't restart anything
5101	 */
5102	if (ret < 0)
5103		goto nothing_to_do;
5104
5105skip_blocking:
5106	pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5107	ctx->ctx_ovfl_regs[0] = 0UL;
5108
5109nothing_to_do:
5110	/*
5111	 * restore flags as they were upon entry
5112	 */
5113	UNPROTECT_CTX(ctx, flags);
5114}
5115
5116static int
5117pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5118{
5119	if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5120		DPRINT(("ignoring overflow notification, owner is zombie\n"));
5121		return 0;
5122	}
5123
5124	DPRINT(("waking up somebody\n"));
5125
5126	if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5127
5128	/*
5129	 * safe, we are not in intr handler, nor in ctxsw when
5130	 * we come here
5131	 */
5132	kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5133
5134	return 0;
5135}
5136
5137static int
5138pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5139{
5140	pfm_msg_t *msg = NULL;
5141
5142	if (ctx->ctx_fl_no_msg == 0) {
5143		msg = pfm_get_new_msg(ctx);
5144		if (msg == NULL) {
5145			printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5146			return -1;
5147		}
5148
5149		msg->pfm_ovfl_msg.msg_type         = PFM_MSG_OVFL;
5150		msg->pfm_ovfl_msg.msg_ctx_fd       = ctx->ctx_fd;
5151		msg->pfm_ovfl_msg.msg_active_set   = 0;
5152		msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5153		msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5154		msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5155		msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5156		msg->pfm_ovfl_msg.msg_tstamp       = 0UL;
5157	}
5158
5159	DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5160		msg,
5161		ctx->ctx_fl_no_msg,
5162		ctx->ctx_fd,
5163		ovfl_pmds));
5164
5165	return pfm_notify_user(ctx, msg);
5166}
5167
5168static int
5169pfm_end_notify_user(pfm_context_t *ctx)
5170{
5171	pfm_msg_t *msg;
5172
5173	msg = pfm_get_new_msg(ctx);
5174	if (msg == NULL) {
5175		printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5176		return -1;
5177	}
5178	/* no leak */
5179	memset(msg, 0, sizeof(*msg));
5180
5181	msg->pfm_end_msg.msg_type    = PFM_MSG_END;
5182	msg->pfm_end_msg.msg_ctx_fd  = ctx->ctx_fd;
5183	msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5184
5185	DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5186		msg,
5187		ctx->ctx_fl_no_msg,
5188		ctx->ctx_fd));
5189
5190	return pfm_notify_user(ctx, msg);
5191}
5192
5193/*
5194 * main overflow processing routine.
5195 * it can be called from the interrupt path or explicitly during the context switch code
5196 */
5197static void pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx,
5198				unsigned long pmc0, struct pt_regs *regs)
5199{
5200	pfm_ovfl_arg_t *ovfl_arg;
5201	unsigned long mask;
5202	unsigned long old_val, ovfl_val, new_val;
5203	unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5204	unsigned long tstamp;
5205	pfm_ovfl_ctrl_t	ovfl_ctrl;
5206	unsigned int i, has_smpl;
5207	int must_notify = 0;
5208
5209	if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5210
5211	/*
5212	 * sanity test. Should never happen
5213	 */
5214	if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5215
5216	tstamp   = ia64_get_itc();
5217	mask     = pmc0 >> PMU_FIRST_COUNTER;
5218	ovfl_val = pmu_conf->ovfl_val;
5219	has_smpl = CTX_HAS_SMPL(ctx);
5220
5221	DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5222		     "used_pmds=0x%lx\n",
5223			pmc0,
5224			task ? task_pid_nr(task): -1,
5225			(regs ? regs->cr_iip : 0),
5226			CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5227			ctx->ctx_used_pmds[0]));
5228
5229
5230	/*
5231	 * first we update the virtual counters
5232	 * assume there was a prior ia64_srlz_d() issued
5233	 */
5234	for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5235
5236		/* skip pmd which did not overflow */
5237		if ((mask & 0x1) == 0) continue;
5238
5239		/*
5240		 * Note that the pmd is not necessarily 0 at this point as qualified events
5241		 * may have happened before the PMU was frozen. The residual count is not
5242		 * taken into consideration here but will be with any read of the pmd via
5243		 * pfm_read_pmds().
5244		 */
5245		old_val              = new_val = ctx->ctx_pmds[i].val;
5246		new_val             += 1 + ovfl_val;
5247		ctx->ctx_pmds[i].val = new_val;
5248
5249		/*
5250		 * check for overflow condition
5251		 */
5252		if (likely(old_val > new_val)) {
5253			ovfl_pmds |= 1UL << i;
5254			if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5255		}
5256
5257		DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5258			i,
5259			new_val,
5260			old_val,
5261			ia64_get_pmd(i) & ovfl_val,
5262			ovfl_pmds,
5263			ovfl_notify));
5264	}
5265
5266	/*
5267	 * there was no 64-bit overflow, nothing else to do
5268	 */
5269	if (ovfl_pmds == 0UL) return;
5270
5271	/* 
5272	 * reset all control bits
5273	 */
5274	ovfl_ctrl.val = 0;
5275	reset_pmds    = 0UL;
5276
5277	/*
5278	 * if a sampling format module exists, then we "cache" the overflow by 
5279	 * calling the module's handler() routine.
5280	 */
5281	if (has_smpl) {
5282		unsigned long start_cycles, end_cycles;
5283		unsigned long pmd_mask;
5284		int j, k, ret = 0;
5285		int this_cpu = smp_processor_id();
5286
5287		pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5288		ovfl_arg = &ctx->ctx_ovfl_arg;
5289
5290		prefetch(ctx->ctx_smpl_hdr);
5291
5292		for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5293
5294			mask = 1UL << i;
5295
5296			if ((pmd_mask & 0x1) == 0) continue;
5297
5298			ovfl_arg->ovfl_pmd      = (unsigned char )i;
5299			ovfl_arg->ovfl_notify   = ovfl_notify & mask ? 1 : 0;
5300			ovfl_arg->active_set    = 0;
5301			ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5302			ovfl_arg->smpl_pmds[0]  = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5303
5304			ovfl_arg->pmd_value      = ctx->ctx_pmds[i].val;
5305			ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5306			ovfl_arg->pmd_eventid    = ctx->ctx_pmds[i].eventid;
5307
5308			/*
5309		 	 * copy values of pmds of interest. Sampling format may copy them
5310		 	 * into sampling buffer.
5311		 	 */
5312			if (smpl_pmds) {
5313				for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5314					if ((smpl_pmds & 0x1) == 0) continue;
5315					ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ?  pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5316					DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5317				}
5318			}
5319
5320			pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5321
5322			start_cycles = ia64_get_itc();
5323
5324			/*
5325		 	 * call custom buffer format record (handler) routine
5326		 	 */
5327			ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5328
5329			end_cycles = ia64_get_itc();
5330
5331			/*
5332			 * For those controls, we take the union because they have
5333			 * an all or nothing behavior.
5334			 */
5335			ovfl_ctrl.bits.notify_user     |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5336			ovfl_ctrl.bits.block_task      |= ovfl_arg->ovfl_ctrl.bits.block_task;
5337			ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5338			/*
5339			 * build the bitmask of pmds to reset now
5340			 */
5341			if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5342
5343			pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5344		}
5345		/*
5346		 * when the module cannot handle the rest of the overflows, we abort right here
5347		 */
5348		if (ret && pmd_mask) {
5349			DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5350				pmd_mask<<PMU_FIRST_COUNTER));
5351		}
5352		/*
5353		 * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5354		 */
5355		ovfl_pmds &= ~reset_pmds;
5356	} else {
5357		/*
5358		 * when no sampling module is used, then the default
5359		 * is to notify on overflow if requested by user
5360		 */
5361		ovfl_ctrl.bits.notify_user     = ovfl_notify ? 1 : 0;
5362		ovfl_ctrl.bits.block_task      = ovfl_notify ? 1 : 0;
5363		ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5364		ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5365		/*
5366		 * if needed, we reset all overflowed pmds
5367		 */
5368		if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5369	}
5370
5371	DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5372
5373	/*
5374	 * reset the requested PMD registers using the short reset values
5375	 */
5376	if (reset_pmds) {
5377		unsigned long bm = reset_pmds;
5378		pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5379	}
5380
5381	if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5382		/*
5383		 * keep track of what to reset when unblocking
5384		 */
5385		ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5386
5387		/*
5388		 * check for blocking context 
5389		 */
5390		if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5391
5392			ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5393
5394			/*
5395			 * set the perfmon specific checking pending work for the task
5396			 */
5397			PFM_SET_WORK_PENDING(task, 1);
5398
5399			/*
5400			 * when coming from ctxsw, current still points to the
5401			 * previous task, therefore we must work with task and not current.
5402			 */
5403			set_notify_resume(task);
5404		}
5405		/*
5406		 * defer until state is changed (shorten spin window). the context is locked
5407		 * anyway, so the signal receiver would come spin for nothing.
5408		 */
5409		must_notify = 1;
5410	}
5411
5412	DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5413			GET_PMU_OWNER() ? task_pid_nr(GET_PMU_OWNER()) : -1,
5414			PFM_GET_WORK_PENDING(task),
5415			ctx->ctx_fl_trap_reason,
5416			ovfl_pmds,
5417			ovfl_notify,
5418			ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5419	/*
5420	 * in case monitoring must be stopped, we toggle the psr bits
5421	 */
5422	if (ovfl_ctrl.bits.mask_monitoring) {
5423		pfm_mask_monitoring(task);
5424		ctx->ctx_state = PFM_CTX_MASKED;
5425		ctx->ctx_fl_can_restart = 1;
5426	}
5427
5428	/*
5429	 * send notification now
5430	 */
5431	if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5432
5433	return;
5434
5435sanity_check:
5436	printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5437			smp_processor_id(),
5438			task ? task_pid_nr(task) : -1,
5439			pmc0);
5440	return;
5441
5442stop_monitoring:
5443	/*
5444	 * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5445	 * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5446	 * come here as zombie only if the task is the current task. In which case, we
5447	 * can access the PMU  hardware directly.
5448	 *
5449	 * Note that zombies do have PM_VALID set. So here we do the minimal.
5450	 *
5451	 * In case the context was zombified it could not be reclaimed at the time
5452	 * the monitoring program exited. At this point, the PMU reservation has been
5453	 * returned, the sampiing buffer has been freed. We must convert this call
5454	 * into a spurious interrupt. However, we must also avoid infinite overflows
5455	 * by stopping monitoring for this task. We can only come here for a per-task
5456	 * context. All we need to do is to stop monitoring using the psr bits which
5457	 * are always task private. By re-enabling secure montioring, we ensure that
5458	 * the monitored task will not be able to re-activate monitoring.
5459	 * The task will eventually be context switched out, at which point the context
5460	 * will be reclaimed (that includes releasing ownership of the PMU).
5461	 *
5462	 * So there might be a window of time where the number of per-task session is zero
5463	 * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5464	 * context. This is safe because if a per-task session comes in, it will push this one
5465	 * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5466	 * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5467	 * also push our zombie context out.
5468	 *
5469	 * Overall pretty hairy stuff....
5470	 */
5471	DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task_pid_nr(task): -1));
5472	pfm_clear_psr_up();
5473	ia64_psr(regs)->up = 0;
5474	ia64_psr(regs)->sp = 1;
5475	return;
5476}
5477
5478static int
5479pfm_do_interrupt_handler(void *arg, struct pt_regs *regs)
5480{
5481	struct task_struct *task;
5482	pfm_context_t *ctx;
5483	unsigned long flags;
5484	u64 pmc0;
5485	int this_cpu = smp_processor_id();
5486	int retval = 0;
5487
5488	pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5489
5490	/*
5491	 * srlz.d done before arriving here
5492	 */
5493	pmc0 = ia64_get_pmc(0);
5494
5495	task = GET_PMU_OWNER();
5496	ctx  = GET_PMU_CTX();
5497
5498	/*
5499	 * if we have some pending bits set
5500	 * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5501	 */
5502	if (PMC0_HAS_OVFL(pmc0) && task) {
5503		/*
5504		 * we assume that pmc0.fr is always set here
5505		 */
5506
5507		/* sanity check */
5508		if (!ctx) goto report_spurious1;
5509
5510		if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0) 
5511			goto report_spurious2;
5512
5513		PROTECT_CTX_NOPRINT(ctx, flags);
5514
5515		pfm_overflow_handler(task, ctx, pmc0, regs);
5516
5517		UNPROTECT_CTX_NOPRINT(ctx, flags);
5518
5519	} else {
5520		pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5521		retval = -1;
5522	}
5523	/*
5524	 * keep it unfrozen at all times
5525	 */
5526	pfm_unfreeze_pmu();
5527
5528	return retval;
5529
5530report_spurious1:
5531	printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5532		this_cpu, task_pid_nr(task));
5533	pfm_unfreeze_pmu();
5534	return -1;
5535report_spurious2:
5536	printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n", 
5537		this_cpu, 
5538		task_pid_nr(task));
5539	pfm_unfreeze_pmu();
5540	return -1;
5541}
5542
5543static irqreturn_t
5544pfm_interrupt_handler(int irq, void *arg)
5545{
5546	unsigned long start_cycles, total_cycles;
5547	unsigned long min, max;
5548	int this_cpu;
5549	int ret;
5550	struct pt_regs *regs = get_irq_regs();
5551
5552	this_cpu = get_cpu();
5553	if (likely(!pfm_alt_intr_handler)) {
5554		min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5555		max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5556
5557		start_cycles = ia64_get_itc();
5558
5559		ret = pfm_do_interrupt_handler(arg, regs);
5560
5561		total_cycles = ia64_get_itc();
5562
5563		/*
5564		 * don't measure spurious interrupts
5565		 */
5566		if (likely(ret == 0)) {
5567			total_cycles -= start_cycles;
5568
5569			if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5570			if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5571
5572			pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5573		}
5574	}
5575	else {
5576		(*pfm_alt_intr_handler->handler)(irq, arg, regs);
5577	}
5578
5579	put_cpu();
5580	return IRQ_HANDLED;
5581}
5582
5583/*
5584 * /proc/perfmon interface, for debug only
5585 */
5586
5587#define PFM_PROC_SHOW_HEADER	((void *)(long)nr_cpu_ids+1)
5588
5589static void *
5590pfm_proc_start(struct seq_file *m, loff_t *pos)
5591{
5592	if (*pos == 0) {
5593		return PFM_PROC_SHOW_HEADER;
5594	}
5595
5596	while (*pos <= nr_cpu_ids) {
5597		if (cpu_online(*pos - 1)) {
5598			return (void *)*pos;
5599		}
5600		++*pos;
5601	}
5602	return NULL;
5603}
5604
5605static void *
5606pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5607{
5608	++*pos;
5609	return pfm_proc_start(m, pos);
5610}
5611
5612static void
5613pfm_proc_stop(struct seq_file *m, void *v)
5614{
5615}
5616
5617static void
5618pfm_proc_show_header(struct seq_file *m)
5619{
5620	struct list_head * pos;
5621	pfm_buffer_fmt_t * entry;
5622	unsigned long flags;
5623
5624 	seq_printf(m,
5625		"perfmon version           : %u.%u\n"
5626		"model                     : %s\n"
5627		"fastctxsw                 : %s\n"
5628		"expert mode               : %s\n"
5629		"ovfl_mask                 : 0x%lx\n"
5630		"PMU flags                 : 0x%x\n",
5631		PFM_VERSION_MAJ, PFM_VERSION_MIN,
5632		pmu_conf->pmu_name,
5633		pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5634		pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5635		pmu_conf->ovfl_val,
5636		pmu_conf->flags);
5637
5638  	LOCK_PFS(flags);
5639
5640 	seq_printf(m,
5641 		"proc_sessions             : %u\n"
5642 		"sys_sessions              : %u\n"
5643 		"sys_use_dbregs            : %u\n"
5644 		"ptrace_use_dbregs         : %u\n",
5645 		pfm_sessions.pfs_task_sessions,
5646 		pfm_sessions.pfs_sys_sessions,
5647 		pfm_sessions.pfs_sys_use_dbregs,
5648 		pfm_sessions.pfs_ptrace_use_dbregs);
5649
5650  	UNLOCK_PFS(flags);
5651
5652	spin_lock(&pfm_buffer_fmt_lock);
5653
5654	list_for_each(pos, &pfm_buffer_fmt_list) {
5655		entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5656		seq_printf(m, "format                    : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n",
5657			entry->fmt_uuid[0],
5658			entry->fmt_uuid[1],
5659			entry->fmt_uuid[2],
5660			entry->fmt_uuid[3],
5661			entry->fmt_uuid[4],
5662			entry->fmt_uuid[5],
5663			entry->fmt_uuid[6],
5664			entry->fmt_uuid[7],
5665			entry->fmt_uuid[8],
5666			entry->fmt_uuid[9],
5667			entry->fmt_uuid[10],
5668			entry->fmt_uuid[11],
5669			entry->fmt_uuid[12],
5670			entry->fmt_uuid[13],
5671			entry->fmt_uuid[14],
5672			entry->fmt_uuid[15],
5673			entry->fmt_name);
5674	}
5675	spin_unlock(&pfm_buffer_fmt_lock);
5676
5677}
5678
5679static int
5680pfm_proc_show(struct seq_file *m, void *v)
5681{
5682	unsigned long psr;
5683	unsigned int i;
5684	int cpu;
5685
5686	if (v == PFM_PROC_SHOW_HEADER) {
5687		pfm_proc_show_header(m);
5688		return 0;
5689	}
5690
5691	/* show info for CPU (v - 1) */
5692
5693	cpu = (long)v - 1;
5694	seq_printf(m,
5695		"CPU%-2d overflow intrs      : %lu\n"
5696		"CPU%-2d overflow cycles     : %lu\n"
5697		"CPU%-2d overflow min        : %lu\n"
5698		"CPU%-2d overflow max        : %lu\n"
5699		"CPU%-2d smpl handler calls  : %lu\n"
5700		"CPU%-2d smpl handler cycles : %lu\n"
5701		"CPU%-2d spurious intrs      : %lu\n"
5702		"CPU%-2d replay   intrs      : %lu\n"
5703		"CPU%-2d syst_wide           : %d\n"
5704		"CPU%-2d dcr_pp              : %d\n"
5705		"CPU%-2d exclude idle        : %d\n"
5706		"CPU%-2d owner               : %d\n"
5707		"CPU%-2d context             : %p\n"
5708		"CPU%-2d activations         : %lu\n",
5709		cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5710		cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5711		cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5712		cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5713		cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5714		cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5715		cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5716		cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5717		cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5718		cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5719		cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5720		cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5721		cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5722		cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5723
5724	if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5725
5726		psr = pfm_get_psr();
5727
5728		ia64_srlz_d();
5729
5730		seq_printf(m, 
5731			"CPU%-2d psr                 : 0x%lx\n"
5732			"CPU%-2d pmc0                : 0x%lx\n", 
5733			cpu, psr,
5734			cpu, ia64_get_pmc(0));
5735
5736		for (i=0; PMC_IS_LAST(i) == 0;  i++) {
5737			if (PMC_IS_COUNTING(i) == 0) continue;
5738   			seq_printf(m, 
5739				"CPU%-2d pmc%u                : 0x%lx\n"
5740   				"CPU%-2d pmd%u                : 0x%lx\n", 
5741				cpu, i, ia64_get_pmc(i),
5742				cpu, i, ia64_get_pmd(i));
5743  		}
5744	}
5745	return 0;
5746}
5747
5748const struct seq_operations pfm_seq_ops = {
5749	.start =	pfm_proc_start,
5750 	.next =		pfm_proc_next,
5751 	.stop =		pfm_proc_stop,
5752 	.show =		pfm_proc_show
5753};
5754
5755static int
5756pfm_proc_open(struct inode *inode, struct file *file)
5757{
5758	return seq_open(file, &pfm_seq_ops);
5759}
5760
5761
5762/*
5763 * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5764 * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5765 * is active or inactive based on mode. We must rely on the value in
5766 * local_cpu_data->pfm_syst_info
5767 */
5768void
5769pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5770{
5771	struct pt_regs *regs;
5772	unsigned long dcr;
5773	unsigned long dcr_pp;
5774
5775	dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5776
5777	/*
5778	 * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5779	 * on every CPU, so we can rely on the pid to identify the idle task.
5780	 */
5781	if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5782		regs = task_pt_regs(task);
5783		ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5784		return;
5785	}
5786	/*
5787	 * if monitoring has started
5788	 */
5789	if (dcr_pp) {
5790		dcr = ia64_getreg(_IA64_REG_CR_DCR);
5791		/*
5792		 * context switching in?
5793		 */
5794		if (is_ctxswin) {
5795			/* mask monitoring for the idle task */
5796			ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5797			pfm_clear_psr_pp();
5798			ia64_srlz_i();
5799			return;
5800		}
5801		/*
5802		 * context switching out
5803		 * restore monitoring for next task
5804		 *
5805		 * Due to inlining this odd if-then-else construction generates
5806		 * better code.
5807		 */
5808		ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5809		pfm_set_psr_pp();
5810		ia64_srlz_i();
5811	}
5812}
5813
5814#ifdef CONFIG_SMP
5815
5816static void
5817pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5818{
5819	struct task_struct *task = ctx->ctx_task;
5820
5821	ia64_psr(regs)->up = 0;
5822	ia64_psr(regs)->sp = 1;
5823
5824	if (GET_PMU_OWNER() == task) {
5825		DPRINT(("cleared ownership for [%d]\n",
5826					task_pid_nr(ctx->ctx_task)));
5827		SET_PMU_OWNER(NULL, NULL);
5828	}
5829
5830	/*
5831	 * disconnect the task from the context and vice-versa
5832	 */
5833	PFM_SET_WORK_PENDING(task, 0);
5834
5835	task->thread.pfm_context  = NULL;
5836	task->thread.flags       &= ~IA64_THREAD_PM_VALID;
5837
5838	DPRINT(("force cleanup for [%d]\n",  task_pid_nr(task)));
5839}
5840
5841
5842/*
5843 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5844 */
5845void
5846pfm_save_regs(struct task_struct *task)
5847{
5848	pfm_context_t *ctx;
5849	unsigned long flags;
5850	u64 psr;
5851
5852
5853	ctx = PFM_GET_CTX(task);
5854	if (ctx == NULL) return;
5855
5856	/*
5857 	 * we always come here with interrupts ALREADY disabled by
5858 	 * the scheduler. So we simply need to protect against concurrent
5859	 * access, not CPU concurrency.
5860	 */
5861	flags = pfm_protect_ctx_ctxsw(ctx);
5862
5863	if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5864		struct pt_regs *regs = task_pt_regs(task);
5865
5866		pfm_clear_psr_up();
5867
5868		pfm_force_cleanup(ctx, regs);
5869
5870		BUG_ON(ctx->ctx_smpl_hdr);
5871
5872		pfm_unprotect_ctx_ctxsw(ctx, flags);
5873
5874		pfm_context_free(ctx);
5875		return;
5876	}
5877
5878	/*
5879	 * save current PSR: needed because we modify it
5880	 */
5881	ia64_srlz_d();
5882	psr = pfm_get_psr();
5883
5884	BUG_ON(psr & (IA64_PSR_I));
5885
5886	/*
5887	 * stop monitoring:
5888	 * This is the last instruction which may generate an overflow
5889	 *
5890	 * We do not need to set psr.sp because, it is irrelevant in kernel.
5891	 * It will be restored from ipsr when going back to user level
5892	 */
5893	pfm_clear_psr_up();
5894
5895	/*
5896	 * keep a copy of psr.up (for reload)
5897	 */
5898	ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5899
5900	/*
5901	 * release ownership of this PMU.
5902	 * PM interrupts are masked, so nothing
5903	 * can happen.
5904	 */
5905	SET_PMU_OWNER(NULL, NULL);
5906
5907	/*
5908	 * we systematically save the PMD as we have no
5909	 * guarantee we will be schedule at that same
5910	 * CPU again.
5911	 */
5912	pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5913
5914	/*
5915	 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5916	 * we will need it on the restore path to check
5917	 * for pending overflow.
5918	 */
5919	ctx->th_pmcs[0] = ia64_get_pmc(0);
5920
5921	/*
5922	 * unfreeze PMU if had pending overflows
5923	 */
5924	if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5925
5926	/*
5927	 * finally, allow context access.
5928	 * interrupts will still be masked after this call.
5929	 */
5930	pfm_unprotect_ctx_ctxsw(ctx, flags);
5931}
5932
5933#else /* !CONFIG_SMP */
5934void
5935pfm_save_regs(struct task_struct *task)
5936{
5937	pfm_context_t *ctx;
5938	u64 psr;
5939
5940	ctx = PFM_GET_CTX(task);
5941	if (ctx == NULL) return;
5942
5943	/*
5944	 * save current PSR: needed because we modify it
5945	 */
5946	psr = pfm_get_psr();
5947
5948	BUG_ON(psr & (IA64_PSR_I));
5949
5950	/*
5951	 * stop monitoring:
5952	 * This is the last instruction which may generate an overflow
5953	 *
5954	 * We do not need to set psr.sp because, it is irrelevant in kernel.
5955	 * It will be restored from ipsr when going back to user level
5956	 */
5957	pfm_clear_psr_up();
5958
5959	/*
5960	 * keep a copy of psr.up (for reload)
5961	 */
5962	ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5963}
5964
5965static void
5966pfm_lazy_save_regs (struct task_struct *task)
5967{
5968	pfm_context_t *ctx;
5969	unsigned long flags;
5970
5971	{ u64 psr  = pfm_get_psr();
5972	  BUG_ON(psr & IA64_PSR_UP);
5973	}
5974
5975	ctx = PFM_GET_CTX(task);
5976
5977	/*
5978	 * we need to mask PMU overflow here to
5979	 * make sure that we maintain pmc0 until
5980	 * we save it. overflow interrupts are
5981	 * treated as spurious if there is no
5982	 * owner.
5983	 *
5984	 * XXX: I don't think this is necessary
5985	 */
5986	PROTECT_CTX(ctx,flags);
5987
5988	/*
5989	 * release ownership of this PMU.
5990	 * must be done before we save the registers.
5991	 *
5992	 * after this call any PMU interrupt is treated
5993	 * as spurious.
5994	 */
5995	SET_PMU_OWNER(NULL, NULL);
5996
5997	/*
5998	 * save all the pmds we use
5999	 */
6000	pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
6001
6002	/*
6003	 * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
6004	 * it is needed to check for pended overflow
6005	 * on the restore path
6006	 */
6007	ctx->th_pmcs[0] = ia64_get_pmc(0);
6008
6009	/*
6010	 * unfreeze PMU if had pending overflows
6011	 */
6012	if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
6013
6014	/*
6015	 * now get can unmask PMU interrupts, they will
6016	 * be treated as purely spurious and we will not
6017	 * lose any information
6018	 */
6019	UNPROTECT_CTX(ctx,flags);
6020}
6021#endif /* CONFIG_SMP */
6022
6023#ifdef CONFIG_SMP
6024/*
6025 * in 2.6, interrupts are masked when we come here and the runqueue lock is held
6026 */
6027void
6028pfm_load_regs (struct task_struct *task)
6029{
6030	pfm_context_t *ctx;
6031	unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
6032	unsigned long flags;
6033	u64 psr, psr_up;
6034	int need_irq_resend;
6035
6036	ctx = PFM_GET_CTX(task);
6037	if (unlikely(ctx == NULL)) return;
6038
6039	BUG_ON(GET_PMU_OWNER());
6040
6041	/*
6042	 * possible on unload
6043	 */
6044	if (unlikely((task->thread.flags & IA64_THREAD_PM_VALID) == 0)) return;
6045
6046	/*
6047 	 * we always come here with interrupts ALREADY disabled by
6048 	 * the scheduler. So we simply need to protect against concurrent
6049	 * access, not CPU concurrency.
6050	 */
6051	flags = pfm_protect_ctx_ctxsw(ctx);
6052	psr   = pfm_get_psr();
6053
6054	need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6055
6056	BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6057	BUG_ON(psr & IA64_PSR_I);
6058
6059	if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6060		struct pt_regs *regs = task_pt_regs(task);
6061
6062		BUG_ON(ctx->ctx_smpl_hdr);
6063
6064		pfm_force_cleanup(ctx, regs);
6065
6066		pfm_unprotect_ctx_ctxsw(ctx, flags);
6067
6068		/*
6069		 * this one (kmalloc'ed) is fine with interrupts disabled
6070		 */
6071		pfm_context_free(ctx);
6072
6073		return;
6074	}
6075
6076	/*
6077	 * we restore ALL the debug registers to avoid picking up
6078	 * stale state.
6079	 */
6080	if (ctx->ctx_fl_using_dbreg) {
6081		pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6082		pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6083	}
6084	/*
6085	 * retrieve saved psr.up
6086	 */
6087	psr_up = ctx->ctx_saved_psr_up;
6088
6089	/*
6090	 * if we were the last user of the PMU on that CPU,
6091	 * then nothing to do except restore psr
6092	 */
6093	if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6094
6095		/*
6096		 * retrieve partial reload masks (due to user modifications)
6097		 */
6098		pmc_mask = ctx->ctx_reload_pmcs[0];
6099		pmd_mask = ctx->ctx_reload_pmds[0];
6100
6101	} else {
6102		/*
6103	 	 * To avoid leaking information to the user level when psr.sp=0,
6104	 	 * we must reload ALL implemented pmds (even the ones we don't use).
6105	 	 * In the kernel we only allow PFM_READ_PMDS on registers which
6106	 	 * we initialized or requested (sampling) so there is no risk there.
6107	 	 */
6108		pmd_mask = pfm_sysctl.fastctxsw ?  ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6109
6110		/*
6111	 	 * ALL accessible PMCs are systematically reloaded, unused registers
6112	 	 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6113	 	 * up stale configuration.
6114	 	 *
6115	 	 * PMC0 is never in the mask. It is always restored separately.
6116	 	 */
6117		pmc_mask = ctx->ctx_all_pmcs[0];
6118	}
6119	/*
6120	 * when context is MASKED, we will restore PMC with plm=0
6121	 * and PMD with stale information, but that's ok, nothing
6122	 * will be captured.
6123	 *
6124	 * XXX: optimize here
6125	 */
6126	if (pmd_mask) pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6127	if (pmc_mask) pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6128
6129	/*
6130	 * check for pending overflow at the time the state
6131	 * was saved.
6132	 */
6133	if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6134		/*
6135		 * reload pmc0 with the overflow information
6136		 * On McKinley PMU, this will trigger a PMU interrupt
6137		 */
6138		ia64_set_pmc(0, ctx->th_pmcs[0]);
6139		ia64_srlz_d();
6140		ctx->th_pmcs[0] = 0UL;
6141
6142		/*
6143		 * will replay the PMU interrupt
6144		 */
6145		if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6146
6147		pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6148	}
6149
6150	/*
6151	 * we just did a reload, so we reset the partial reload fields
6152	 */
6153	ctx->ctx_reload_pmcs[0] = 0UL;
6154	ctx->ctx_reload_pmds[0] = 0UL;
6155
6156	SET_LAST_CPU(ctx, smp_processor_id());
6157
6158	/*
6159	 * dump activation value for this PMU
6160	 */
6161	INC_ACTIVATION();
6162	/*
6163	 * record current activation for this context
6164	 */
6165	SET_ACTIVATION(ctx);
6166
6167	/*
6168	 * establish new ownership. 
6169	 */
6170	SET_PMU_OWNER(task, ctx);
6171
6172	/*
6173	 * restore the psr.up bit. measurement
6174	 * is active again.
6175	 * no PMU interrupt can happen at this point
6176	 * because we still have interrupts disabled.
6177	 */
6178	if (likely(psr_up)) pfm_set_psr_up();
6179
6180	/*
6181	 * allow concurrent access to context
6182	 */
6183	pfm_unprotect_ctx_ctxsw(ctx, flags);
6184}
6185#else /*  !CONFIG_SMP */
6186/*
6187 * reload PMU state for UP kernels
6188 * in 2.5 we come here with interrupts disabled
6189 */
6190void
6191pfm_load_regs (struct task_struct *task)
6192{
6193	pfm_context_t *ctx;
6194	struct task_struct *owner;
6195	unsigned long pmd_mask, pmc_mask;
6196	u64 psr, psr_up;
6197	int need_irq_resend;
6198
6199	owner = GET_PMU_OWNER();
6200	ctx   = PFM_GET_CTX(task);
6201	psr   = pfm_get_psr();
6202
6203	BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6204	BUG_ON(psr & IA64_PSR_I);
6205
6206	/*
6207	 * we restore ALL the debug registers to avoid picking up
6208	 * stale state.
6209	 *
6210	 * This must be done even when the task is still the owner
6211	 * as the registers may have been modified via ptrace()
6212	 * (not perfmon) by the previous task.
6213	 */
6214	if (ctx->ctx_fl_using_dbreg) {
6215		pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6216		pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6217	}
6218
6219	/*
6220	 * retrieved saved psr.up
6221	 */
6222	psr_up = ctx->ctx_saved_psr_up;
6223	need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6224
6225	/*
6226	 * short path, our state is still there, just
6227	 * need to restore psr and we go
6228	 *
6229	 * we do not touch either PMC nor PMD. the psr is not touched
6230	 * by the overflow_handler. So we are safe w.r.t. to interrupt
6231	 * concurrency even without interrupt masking.
6232	 */
6233	if (likely(owner == task)) {
6234		if (likely(psr_up)) pfm_set_psr_up();
6235		return;
6236	}
6237
6238	/*
6239	 * someone else is still using the PMU, first push it out and
6240	 * then we'll be able to install our stuff !
6241	 *
6242	 * Upon return, there will be no owner for the current PMU
6243	 */
6244	if (owner) pfm_lazy_save_regs(owner);
6245
6246	/*
6247	 * To avoid leaking information to the user level when psr.sp=0,
6248	 * we must reload ALL implemented pmds (even the ones we don't use).
6249	 * In the kernel we only allow PFM_READ_PMDS on registers which
6250	 * we initialized or requested (sampling) so there is no risk there.
6251	 */
6252	pmd_mask = pfm_sysctl.fastctxsw ?  ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6253
6254	/*
6255	 * ALL accessible PMCs are systematically reloaded, unused registers
6256	 * get their default (from pfm_reset_pmu_state()) values to avoid picking
6257	 * up stale configuration.
6258	 *
6259	 * PMC0 is never in the mask. It is always restored separately
6260	 */
6261	pmc_mask = ctx->ctx_all_pmcs[0];
6262
6263	pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6264	pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6265
6266	/*
6267	 * check for pending overflow at the time the state
6268	 * was saved.
6269	 */
6270	if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6271		/*
6272		 * reload pmc0 with the overflow information
6273		 * On McKinley PMU, this will trigger a PMU interrupt
6274		 */
6275		ia64_set_pmc(0, ctx->th_pmcs[0]);
6276		ia64_srlz_d();
6277
6278		ctx->th_pmcs[0] = 0UL;
6279
6280		/*
6281		 * will replay the PMU interrupt
6282		 */
6283		if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6284
6285		pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6286	}
6287
6288	/*
6289	 * establish new ownership. 
6290	 */
6291	SET_PMU_OWNER(task, ctx);
6292
6293	/*
6294	 * restore the psr.up bit. measurement
6295	 * is active again.
6296	 * no PMU interrupt can happen at this point
6297	 * because we still have interrupts disabled.
6298	 */
6299	if (likely(psr_up)) pfm_set_psr_up();
6300}
6301#endif /* CONFIG_SMP */
6302
6303/*
6304 * this function assumes monitoring is stopped
6305 */
6306static void
6307pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6308{
6309	u64 pmc0;
6310	unsigned long mask2, val, pmd_val, ovfl_val;
6311	int i, can_access_pmu = 0;
6312	int is_self;
6313
6314	/*
6315	 * is the caller the task being monitored (or which initiated the
6316	 * session for system wide measurements)
6317	 */
6318	is_self = ctx->ctx_task == task ? 1 : 0;
6319
6320	/*
6321	 * can access PMU is task is the owner of the PMU state on the current CPU
6322	 * or if we are running on the CPU bound to the context in system-wide mode
6323	 * (that is not necessarily the task the context is attached to in this mode).
6324	 * In system-wide we always have can_access_pmu true because a task running on an
6325	 * invalid processor is flagged earlier in the call stack (see pfm_stop).
6326	 */
6327	can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6328	if (can_access_pmu) {
6329		/*
6330		 * Mark the PMU as not owned
6331		 * This will cause the interrupt handler to do nothing in case an overflow
6332		 * interrupt was in-flight
6333		 * This also guarantees that pmc0 will contain the final state
6334		 * It virtually gives us full control on overflow processing from that point
6335		 * on.
6336		 */
6337		SET_PMU_OWNER(NULL, NULL);
6338		DPRINT(("releasing ownership\n"));
6339
6340		/*
6341		 * read current overflow status:
6342		 *
6343		 * we are guaranteed to read the final stable state
6344		 */
6345		ia64_srlz_d();
6346		pmc0 = ia64_get_pmc(0); /* slow */
6347
6348		/*
6349		 * reset freeze bit, overflow status information destroyed
6350		 */
6351		pfm_unfreeze_pmu();
6352	} else {
6353		pmc0 = ctx->th_pmcs[0];
6354		/*
6355		 * clear whatever overflow status bits there were
6356		 */
6357		ctx->th_pmcs[0] = 0;
6358	}
6359	ovfl_val = pmu_conf->ovfl_val;
6360	/*
6361	 * we save all the used pmds
6362	 * we take care of overflows for counting PMDs
6363	 *
6364	 * XXX: sampling situation is not taken into account here
6365	 */
6366	mask2 = ctx->ctx_used_pmds[0];
6367
6368	DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6369
6370	for (i = 0; mask2; i++, mask2>>=1) {
6371
6372		/* skip non used pmds */
6373		if ((mask2 & 0x1) == 0) continue;
6374
6375		/*
6376		 * can access PMU always true in system wide mode
6377		 */
6378		val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : ctx->th_pmds[i];
6379
6380		if (PMD_IS_COUNTING(i)) {
6381			DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6382				task_pid_nr(task),
6383				i,
6384				ctx->ctx_pmds[i].val,
6385				val & ovfl_val));
6386
6387			/*
6388			 * we rebuild the full 64 bit value of the counter
6389			 */
6390			val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6391
6392			/*
6393			 * now everything is in ctx_pmds[] and we need
6394			 * to clear the saved context from save_regs() such that
6395			 * pfm_read_pmds() gets the correct value
6396			 */
6397			pmd_val = 0UL;
6398
6399			/*
6400			 * take care of overflow inline
6401			 */
6402			if (pmc0 & (1UL << i)) {
6403				val += 1 + ovfl_val;
6404				DPRINT(("[%d] pmd[%d] overflowed\n", task_pid_nr(task), i));
6405			}
6406		}
6407
6408		DPRINT(("[%d] ctx_pmd[%d]=0x%lx  pmd_val=0x%lx\n", task_pid_nr(task), i, val, pmd_val));
6409
6410		if (is_self) ctx->th_pmds[i] = pmd_val;
6411
6412		ctx->ctx_pmds[i].val = val;
6413	}
6414}
6415
6416static struct irqaction perfmon_irqaction = {
6417	.handler = pfm_interrupt_handler,
6418	.flags   = IRQF_DISABLED,
6419	.name    = "perfmon"
6420};
6421
6422static void
6423pfm_alt_save_pmu_state(void *data)
6424{
6425	struct pt_regs *regs;
6426
6427	regs = task_pt_regs(current);
6428
6429	DPRINT(("called\n"));
6430
6431	/*
6432	 * should not be necessary but
6433	 * let's take not risk
6434	 */
6435	pfm_clear_psr_up();
6436	pfm_clear_psr_pp();
6437	ia64_psr(regs)->pp = 0;
6438
6439	/*
6440	 * This call is required
6441	 * May cause a spurious interrupt on some processors
6442	 */
6443	pfm_freeze_pmu();
6444
6445	ia64_srlz_d();
6446}
6447
6448void
6449pfm_alt_restore_pmu_state(void *data)
6450{
6451	struct pt_regs *regs;
6452
6453	regs = task_pt_regs(current);
6454
6455	DPRINT(("called\n"));
6456
6457	/*
6458	 * put PMU back in state expected
6459	 * by perfmon
6460	 */
6461	pfm_clear_psr_up();
6462	pfm_clear_psr_pp();
6463	ia64_psr(regs)->pp = 0;
6464
6465	/*
6466	 * perfmon runs with PMU unfrozen at all times
6467	 */
6468	pfm_unfreeze_pmu();
6469
6470	ia64_srlz_d();
6471}
6472
6473int
6474pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6475{
6476	int ret, i;
6477	int reserve_cpu;
6478
6479	/* some sanity checks */
6480	if (hdl == NULL || hdl->handler == NULL) return -EINVAL;
6481
6482	/* do the easy test first */
6483	if (pfm_alt_intr_handler) return -EBUSY;
6484
6485	/* one at a time in the install or remove, just fail the others */
6486	if (!spin_trylock(&pfm_alt_install_check)) {
6487		return -EBUSY;
6488	}
6489
6490	/* reserve our session */
6491	for_each_online_cpu(reserve_cpu) {
6492		ret = pfm_reserve_session(NULL, 1, reserve_cpu);
6493		if (ret) goto cleanup_reserve;
6494	}
6495
6496	/* save the current system wide pmu states */
6497	ret = on_each_cpu(pfm_alt_save_pmu_state, NULL, 1);
6498	if (ret) {
6499		DPRINT(("on_each_cpu() failed: %d\n", ret));
6500		goto cleanup_reserve;
6501	}
6502
6503	/* officially change to the alternate interrupt handler */
6504	pfm_alt_intr_handler = hdl;
6505
6506	spin_unlock(&pfm_alt_install_check);
6507
6508	return 0;
6509
6510cleanup_reserve:
6511	for_each_online_cpu(i) {
6512		/* don't unreserve more than we reserved */
6513		if (i >= reserve_cpu) break;
6514
6515		pfm_unreserve_session(NULL, 1, i);
6516	}
6517
6518	spin_unlock(&pfm_alt_install_check);
6519
6520	return ret;
6521}
6522EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt);
6523
6524int
6525pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6526{
6527	int i;
6528	int ret;
6529
6530	if (hdl == NULL) return -EINVAL;
6531
6532	/* cannot remove someone else's handler! */
6533	if (pfm_alt_intr_handler != hdl) return -EINVAL;
6534
6535	/* one at a time in the install or remove, just fail the others */
6536	if (!spin_trylock(&pfm_alt_install_check)) {
6537		return -EBUSY;
6538	}
6539
6540	pfm_alt_intr_handler = NULL;
6541
6542	ret = on_each_cpu(pfm_alt_restore_pmu_state, NULL, 1);
6543	if (ret) {
6544		DPRINT(("on_each_cpu() failed: %d\n", ret));
6545	}
6546
6547	for_each_online_cpu(i) {
6548		pfm_unreserve_session(NULL, 1, i);
6549	}
6550
6551	spin_unlock(&pfm_alt_install_check);
6552
6553	return 0;
6554}
6555EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt);
6556
6557/*
6558 * perfmon initialization routine, called from the initcall() table
6559 */
6560static int init_pfm_fs(void);
6561
6562static int __init
6563pfm_probe_pmu(void)
6564{
6565	pmu_config_t **p;
6566	int family;
6567
6568	family = local_cpu_data->family;
6569	p      = pmu_confs;
6570
6571	while(*p) {
6572		if ((*p)->probe) {
6573			if ((*p)->probe() == 0) goto found;
6574		} else if ((*p)->pmu_family == family || (*p)->pmu_family == 0xff) {
6575			goto found;
6576		}
6577		p++;
6578	}
6579	return -1;
6580found:
6581	pmu_conf = *p;
6582	return 0;
6583}
6584
6585static const struct file_operations pfm_proc_fops = {
6586	.open		= pfm_proc_open,
6587	.read		= seq_read,
6588	.llseek		= seq_lseek,
6589	.release	= seq_release,
6590};
6591
6592int __init
6593pfm_init(void)
6594{
6595	unsigned int n, n_counters, i;
6596
6597	printk("perfmon: version %u.%u IRQ %u\n",
6598		PFM_VERSION_MAJ,
6599		PFM_VERSION_MIN,
6600		IA64_PERFMON_VECTOR);
6601
6602	if (pfm_probe_pmu()) {
6603		printk(KERN_INFO "perfmon: disabled, there is no support for processor family %d\n", 
6604				local_cpu_data->family);
6605		return -ENODEV;
6606	}
6607
6608	/*
6609	 * compute the number of implemented PMD/PMC from the
6610	 * description tables
6611	 */
6612	n = 0;
6613	for (i=0; PMC_IS_LAST(i) == 0;  i++) {
6614		if (PMC_IS_IMPL(i) == 0) continue;
6615		pmu_conf->impl_pmcs[i>>6] |= 1UL << (i&63);
6616		n++;
6617	}
6618	pmu_conf->num_pmcs = n;
6619
6620	n = 0; n_counters = 0;
6621	for (i=0; PMD_IS_LAST(i) == 0;  i++) {
6622		if (PMD_IS_IMPL(i) == 0) continue;
6623		pmu_conf->impl_pmds[i>>6] |= 1UL << (i&63);
6624		n++;
6625		if (PMD_IS_COUNTING(i)) n_counters++;
6626	}
6627	pmu_conf->num_pmds      = n;
6628	pmu_conf->num_counters  = n_counters;
6629
6630	/*
6631	 * sanity checks on the number of debug registers
6632	 */
6633	if (pmu_conf->use_rr_dbregs) {
6634		if (pmu_conf->num_ibrs > IA64_NUM_DBG_REGS) {
6635			printk(KERN_INFO "perfmon: unsupported number of code debug registers (%u)\n", pmu_conf->num_ibrs);
6636			pmu_conf = NULL;
6637			return -1;
6638		}
6639		if (pmu_conf->num_dbrs > IA64_NUM_DBG_REGS) {
6640			printk(KERN_INFO "perfmon: unsupported number of data debug registers (%u)\n", pmu_conf->num_ibrs);
6641			pmu_conf = NULL;
6642			return -1;
6643		}
6644	}
6645
6646	printk("perfmon: %s PMU detected, %u PMCs, %u PMDs, %u counters (%lu bits)\n",
6647	       pmu_conf->pmu_name,
6648	       pmu_conf->num_pmcs,
6649	       pmu_conf->num_pmds,
6650	       pmu_conf->num_counters,
6651	       ffz(pmu_conf->ovfl_val));
6652
6653	/* sanity check */
6654	if (pmu_conf->num_pmds >= PFM_NUM_PMD_REGS || pmu_conf->num_pmcs >= PFM_NUM_PMC_REGS) {
6655		printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
6656		pmu_conf = NULL;
6657		return -1;
6658	}
6659
6660	/*
6661	 * create /proc/perfmon (mostly for debugging purposes)
6662	 */
6663	perfmon_dir = proc_create("perfmon", S_IRUGO, NULL, &pfm_proc_fops);
6664	if (perfmon_dir == NULL) {
6665		printk(KERN_ERR "perfmon: cannot create /proc entry, perfmon disabled\n");
6666		pmu_conf = NULL;
6667		return -1;
6668	}
6669
6670	/*
6671	 * create /proc/sys/kernel/perfmon (for debugging purposes)
6672	 */
6673	pfm_sysctl_header = register_sysctl_table(pfm_sysctl_root);
6674
6675	/*
6676	 * initialize all our spinlocks
6677	 */
6678	spin_lock_init(&pfm_sessions.pfs_lock);
6679	spin_lock_init(&pfm_buffer_fmt_lock);
6680
6681	init_pfm_fs();
6682
6683	for(i=0; i < NR_CPUS; i++) pfm_stats[i].pfm_ovfl_intr_cycles_min = ~0UL;
6684
6685	return 0;
6686}
6687
6688__initcall(pfm_init);
6689
6690/*
6691 * this function is called before pfm_init()
6692 */
6693void
6694pfm_init_percpu (void)
6695{
6696	static int first_time=1;
6697	/*
6698	 * make sure no measurement is active
6699	 * (may inherit programmed PMCs from EFI).
6700	 */
6701	pfm_clear_psr_pp();
6702	pfm_clear_psr_up();
6703
6704	/*
6705	 * we run with the PMU not frozen at all times
6706	 */
6707	pfm_unfreeze_pmu();
6708
6709	if (first_time) {
6710		register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
6711		first_time=0;
6712	}
6713
6714	ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
6715	ia64_srlz_d();
6716}
6717
6718/*
6719 * used for debug purposes only
6720 */
6721void
6722dump_pmu_state(const char *from)
6723{
6724	struct task_struct *task;
6725	struct pt_regs *regs;
6726	pfm_context_t *ctx;
6727	unsigned long psr, dcr, info, flags;
6728	int i, this_cpu;
6729
6730	local_irq_save(flags);
6731
6732	this_cpu = smp_processor_id();
6733	regs     = task_pt_regs(current);
6734	info     = PFM_CPUINFO_GET();
6735	dcr      = ia64_getreg(_IA64_REG_CR_DCR);
6736
6737	if (info == 0 && ia64_psr(regs)->pp == 0 && (dcr & IA64_DCR_PP) == 0) {
6738		local_irq_restore(flags);
6739		return;
6740	}
6741
6742	printk("CPU%d from %s() current [%d] iip=0x%lx %s\n", 
6743		this_cpu, 
6744		from, 
6745		task_pid_nr(current),
6746		regs->cr_iip,
6747		current->comm);
6748
6749	task = GET_PMU_OWNER();
6750	ctx  = GET_PMU_CTX();
6751
6752	printk("->CPU%d owner [%d] ctx=%p\n", this_cpu, task ? task_pid_nr(task) : -1, ctx);
6753
6754	psr = pfm_get_psr();
6755
6756	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", 
6757		this_cpu,
6758		ia64_get_pmc(0),
6759		psr & IA64_PSR_PP ? 1 : 0,
6760		psr & IA64_PSR_UP ? 1 : 0,
6761		dcr & IA64_DCR_PP ? 1 : 0,
6762		info,
6763		ia64_psr(regs)->up,
6764		ia64_psr(regs)->pp);
6765
6766	ia64_psr(regs)->up = 0;
6767	ia64_psr(regs)->pp = 0;
6768
6769	for (i=1; PMC_IS_LAST(i) == 0; i++) {
6770		if (PMC_IS_IMPL(i) == 0) continue;
6771		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]);
6772	}
6773
6774	for (i=1; PMD_IS_LAST(i) == 0; i++) {
6775		if (PMD_IS_IMPL(i) == 0) continue;
6776		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]);
6777	}
6778
6779	if (ctx) {
6780		printk("->CPU%d ctx_state=%d vaddr=%p addr=%p fd=%d ctx_task=[%d] saved_psr_up=0x%lx\n",
6781				this_cpu,
6782				ctx->ctx_state,
6783				ctx->ctx_smpl_vaddr,
6784				ctx->ctx_smpl_hdr,
6785				ctx->ctx_msgq_head,
6786				ctx->ctx_msgq_tail,
6787				ctx->ctx_saved_psr_up);
6788	}
6789	local_irq_restore(flags);
6790}
6791
6792/*
6793 * called from process.c:copy_thread(). task is new child.
6794 */
6795void
6796pfm_inherit(struct task_struct *task, struct pt_regs *regs)
6797{
6798	struct thread_struct *thread;
6799
6800	DPRINT(("perfmon: pfm_inherit clearing state for [%d]\n", task_pid_nr(task)));
6801
6802	thread = &task->thread;
6803
6804	/*
6805	 * cut links inherited from parent (current)
6806	 */
6807	thread->pfm_context = NULL;
6808
6809	PFM_SET_WORK_PENDING(task, 0);
6810
6811	/*
6812	 * the psr bits are already set properly in copy_threads()
6813	 */
6814}
6815#else  /* !CONFIG_PERFMON */
6816asmlinkage long
6817sys_perfmonctl (int fd, int cmd, void *arg, int count)
6818{
6819	return -ENOSYS;
6820}
6821#endif /* CONFIG_PERFMON */