1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * File:	mca.c
   4 * Purpose:	Generic MCA handling layer
   5 *
   6 * Copyright (C) 2003 Hewlett-Packard Co
   7 *	David Mosberger-Tang <davidm@hpl.hp.com>
   8 *
   9 * Copyright (C) 2002 Dell Inc.
  10 * Copyright (C) Matt Domsch <Matt_Domsch@dell.com>
  11 *
  12 * Copyright (C) 2002 Intel
  13 * Copyright (C) Jenna Hall <jenna.s.hall@intel.com>
  14 *
  15 * Copyright (C) 2001 Intel
  16 * Copyright (C) Fred Lewis <frederick.v.lewis@intel.com>
  17 *
  18 * Copyright (C) 2000 Intel
  19 * Copyright (C) Chuck Fleckenstein <cfleck@co.intel.com>
  20 *
  21 * Copyright (C) 1999, 2004-2008 Silicon Graphics, Inc.
  22 * Copyright (C) Vijay Chander <vijay@engr.sgi.com>
  23 *
  24 * Copyright (C) 2006 FUJITSU LIMITED
  25 * Copyright (C) Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
  26 *
  27 * 2000-03-29 Chuck Fleckenstein <cfleck@co.intel.com>
  28 *	      Fixed PAL/SAL update issues, began MCA bug fixes, logging issues,
  29 *	      added min save state dump, added INIT handler.
  30 *
  31 * 2001-01-03 Fred Lewis <frederick.v.lewis@intel.com>
  32 *	      Added setup of CMCI and CPEI IRQs, logging of corrected platform
  33 *	      errors, completed code for logging of corrected & uncorrected
  34 *	      machine check errors, and updated for conformance with Nov. 2000
  35 *	      revision of the SAL 3.0 spec.
  36 *
  37 * 2002-01-04 Jenna Hall <jenna.s.hall@intel.com>
  38 *	      Aligned MCA stack to 16 bytes, added platform vs. CPU error flag,
  39 *	      set SAL default return values, changed error record structure to
  40 *	      linked list, added init call to sal_get_state_info_size().
  41 *
  42 * 2002-03-25 Matt Domsch <Matt_Domsch@dell.com>
  43 *	      GUID cleanups.
  44 *
  45 * 2003-04-15 David Mosberger-Tang <davidm@hpl.hp.com>
  46 *	      Added INIT backtrace support.
  47 *
  48 * 2003-12-08 Keith Owens <kaos@sgi.com>
  49 *	      smp_call_function() must not be called from interrupt context
  50 *	      (can deadlock on tasklist_lock).
  51 *	      Use keventd to call smp_call_function().
  52 *
  53 * 2004-02-01 Keith Owens <kaos@sgi.com>
  54 *	      Avoid deadlock when using printk() for MCA and INIT records.
  55 *	      Delete all record printing code, moved to salinfo_decode in user
  56 *	      space.  Mark variables and functions static where possible.
  57 *	      Delete dead variables and functions.  Reorder to remove the need
  58 *	      for forward declarations and to consolidate related code.
  59 *
  60 * 2005-08-12 Keith Owens <kaos@sgi.com>
  61 *	      Convert MCA/INIT handlers to use per event stacks and SAL/OS
  62 *	      state.
  63 *
  64 * 2005-10-07 Keith Owens <kaos@sgi.com>
  65 *	      Add notify_die() hooks.
  66 *
  67 * 2006-09-15 Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
  68 *	      Add printing support for MCA/INIT.
  69 *
  70 * 2007-04-27 Russ Anderson <rja@sgi.com>
  71 *	      Support multiple cpus going through OS_MCA in the same event.
  72 */
  73#include <linux/jiffies.h>
  74#include <linux/types.h>
  75#include <linux/init.h>
  76#include <linux/sched/signal.h>
  77#include <linux/sched/debug.h>
  78#include <linux/sched/task.h>
  79#include <linux/interrupt.h>
  80#include <linux/irq.h>
  81#include <linux/memblock.h>
  82#include <linux/acpi.h>
  83#include <linux/timer.h>
  84#include <linux/module.h>
  85#include <linux/kernel.h>
  86#include <linux/smp.h>
  87#include <linux/workqueue.h>
  88#include <linux/cpumask.h>
  89#include <linux/kdebug.h>
  90#include <linux/cpu.h>
  91#include <linux/gfp.h>
  92
  93#include <asm/delay.h>
  94#include <asm/meminit.h>
  95#include <asm/page.h>
  96#include <asm/ptrace.h>
  97#include <asm/sal.h>
  98#include <asm/mca.h>
  99#include <asm/kexec.h>
 100
 101#include <asm/irq.h>
 102#include <asm/hw_irq.h>
 103#include <asm/tlb.h>
 104
 105#include "mca_drv.h"
 106#include "entry.h"
 
 107
 108#if defined(IA64_MCA_DEBUG_INFO)
 109# define IA64_MCA_DEBUG(fmt...)	printk(fmt)
 110#else
 111# define IA64_MCA_DEBUG(fmt...)
 112#endif
 113
 114#define NOTIFY_INIT(event, regs, arg, spin)				\
 115do {									\
 116	if ((notify_die((event), "INIT", (regs), (arg), 0, 0)		\
 117			== NOTIFY_STOP) && ((spin) == 1))		\
 118		ia64_mca_spin(__func__);				\
 119} while (0)
 120
 121#define NOTIFY_MCA(event, regs, arg, spin)				\
 122do {									\
 123	if ((notify_die((event), "MCA", (regs), (arg), 0, 0)		\
 124			== NOTIFY_STOP) && ((spin) == 1))		\
 125		ia64_mca_spin(__func__);				\
 126} while (0)
 127
 128/* Used by mca_asm.S */
 129DEFINE_PER_CPU(u64, ia64_mca_data); /* == __per_cpu_mca[smp_processor_id()] */
 130DEFINE_PER_CPU(u64, ia64_mca_per_cpu_pte); /* PTE to map per-CPU area */
 131DEFINE_PER_CPU(u64, ia64_mca_pal_pte);	    /* PTE to map PAL code */
 132DEFINE_PER_CPU(u64, ia64_mca_pal_base);    /* vaddr PAL code granule */
 133DEFINE_PER_CPU(u64, ia64_mca_tr_reload);   /* Flag for TR reload */
 134
 135unsigned long __per_cpu_mca[NR_CPUS];
 136
 137/* In mca_asm.S */
 138extern void			ia64_os_init_dispatch_monarch (void);
 139extern void			ia64_os_init_dispatch_slave (void);
 140
 141static int monarch_cpu = -1;
 142
 143static ia64_mc_info_t		ia64_mc_info;
 144
 145#define MAX_CPE_POLL_INTERVAL (15*60*HZ) /* 15 minutes */
 146#define MIN_CPE_POLL_INTERVAL (2*60*HZ)  /* 2 minutes */
 147#define CMC_POLL_INTERVAL     (1*60*HZ)  /* 1 minute */
 148#define CPE_HISTORY_LENGTH    5
 149#define CMC_HISTORY_LENGTH    5
 150
 151static struct timer_list cpe_poll_timer;
 152static struct timer_list cmc_poll_timer;
 153/*
 154 * This variable tells whether we are currently in polling mode.
 155 * Start with this in the wrong state so we won't play w/ timers
 156 * before the system is ready.
 157 */
 158static int cmc_polling_enabled = 1;
 159
 160/*
 161 * Clearing this variable prevents CPE polling from getting activated
 162 * in mca_late_init.  Use it if your system doesn't provide a CPEI,
 163 * but encounters problems retrieving CPE logs.  This should only be
 164 * necessary for debugging.
 165 */
 166static int cpe_poll_enabled = 1;
 167
 168extern void salinfo_log_wakeup(int type, u8 *buffer, u64 size, int irqsafe);
 169
 170static int mca_init __initdata;
 171
 172/*
 173 * limited & delayed printing support for MCA/INIT handler
 174 */
 175
 176#define mprintk(fmt...) ia64_mca_printk(fmt)
 177
 178#define MLOGBUF_SIZE (512+256*NR_CPUS)
 179#define MLOGBUF_MSGMAX 256
 180static char mlogbuf[MLOGBUF_SIZE];
 181static DEFINE_SPINLOCK(mlogbuf_wlock);	/* mca context only */
 182static DEFINE_SPINLOCK(mlogbuf_rlock);	/* normal context only */
 183static unsigned long mlogbuf_start;
 184static unsigned long mlogbuf_end;
 185static unsigned int mlogbuf_finished = 0;
 186static unsigned long mlogbuf_timestamp = 0;
 187
 188static int loglevel_save = -1;
 189#define BREAK_LOGLEVEL(__console_loglevel)		\
 190	oops_in_progress = 1;				\
 191	if (loglevel_save < 0)				\
 192		loglevel_save = __console_loglevel;	\
 193	__console_loglevel = 15;
 194
 195#define RESTORE_LOGLEVEL(__console_loglevel)		\
 196	if (loglevel_save >= 0) {			\
 197		__console_loglevel = loglevel_save;	\
 198		loglevel_save = -1;			\
 199	}						\
 200	mlogbuf_finished = 0;				\
 201	oops_in_progress = 0;
 202
 203/*
 204 * Push messages into buffer, print them later if not urgent.
 205 */
 206void ia64_mca_printk(const char *fmt, ...)
 207{
 208	va_list args;
 209	int printed_len;
 210	char temp_buf[MLOGBUF_MSGMAX];
 211	char *p;
 212
 213	va_start(args, fmt);
 214	printed_len = vscnprintf(temp_buf, sizeof(temp_buf), fmt, args);
 215	va_end(args);
 216
 217	/* Copy the output into mlogbuf */
 218	if (oops_in_progress) {
 219		/* mlogbuf was abandoned, use printk directly instead. */
 220		printk("%s", temp_buf);
 221	} else {
 222		spin_lock(&mlogbuf_wlock);
 223		for (p = temp_buf; *p; p++) {
 224			unsigned long next = (mlogbuf_end + 1) % MLOGBUF_SIZE;
 225			if (next != mlogbuf_start) {
 226				mlogbuf[mlogbuf_end] = *p;
 227				mlogbuf_end = next;
 228			} else {
 229				/* buffer full */
 230				break;
 231			}
 232		}
 233		mlogbuf[mlogbuf_end] = '\0';
 234		spin_unlock(&mlogbuf_wlock);
 235	}
 236}
 237EXPORT_SYMBOL(ia64_mca_printk);
 238
 239/*
 240 * Print buffered messages.
 241 *  NOTE: call this after returning normal context. (ex. from salinfod)
 242 */
 243void ia64_mlogbuf_dump(void)
 244{
 245	char temp_buf[MLOGBUF_MSGMAX];
 246	char *p;
 247	unsigned long index;
 248	unsigned long flags;
 249	unsigned int printed_len;
 250
 251	/* Get output from mlogbuf */
 252	while (mlogbuf_start != mlogbuf_end) {
 253		temp_buf[0] = '\0';
 254		p = temp_buf;
 255		printed_len = 0;
 256
 257		spin_lock_irqsave(&mlogbuf_rlock, flags);
 258
 259		index = mlogbuf_start;
 260		while (index != mlogbuf_end) {
 261			*p = mlogbuf[index];
 262			index = (index + 1) % MLOGBUF_SIZE;
 263			if (!*p)
 264				break;
 265			p++;
 266			if (++printed_len >= MLOGBUF_MSGMAX - 1)
 267				break;
 268		}
 269		*p = '\0';
 270		if (temp_buf[0])
 271			printk("%s", temp_buf);
 272		mlogbuf_start = index;
 273
 274		mlogbuf_timestamp = 0;
 275		spin_unlock_irqrestore(&mlogbuf_rlock, flags);
 276	}
 277}
 278EXPORT_SYMBOL(ia64_mlogbuf_dump);
 279
 280/*
 281 * Call this if system is going to down or if immediate flushing messages to
 282 * console is required. (ex. recovery was failed, crash dump is going to be
 283 * invoked, long-wait rendezvous etc.)
 284 *  NOTE: this should be called from monarch.
 285 */
 286static void ia64_mlogbuf_finish(int wait)
 287{
 288	BREAK_LOGLEVEL(console_loglevel);
 289
 290	spin_lock_init(&mlogbuf_rlock);
 291	ia64_mlogbuf_dump();
 292	printk(KERN_EMERG "mlogbuf_finish: printing switched to urgent mode, "
 293		"MCA/INIT might be dodgy or fail.\n");
 294
 295	if (!wait)
 296		return;
 297
 298	/* wait for console */
 299	printk("Delaying for 5 seconds...\n");
 300	udelay(5*1000000);
 301
 302	mlogbuf_finished = 1;
 303}
 304
 305/*
 306 * Print buffered messages from INIT context.
 307 */
 308static void ia64_mlogbuf_dump_from_init(void)
 309{
 310	if (mlogbuf_finished)
 311		return;
 312
 313	if (mlogbuf_timestamp &&
 314			time_before(jiffies, mlogbuf_timestamp + 30 * HZ)) {
 315		printk(KERN_ERR "INIT: mlogbuf_dump is interrupted by INIT "
 316			" and the system seems to be messed up.\n");
 317		ia64_mlogbuf_finish(0);
 318		return;
 319	}
 320
 321	if (!spin_trylock(&mlogbuf_rlock)) {
 322		printk(KERN_ERR "INIT: mlogbuf_dump is interrupted by INIT. "
 323			"Generated messages other than stack dump will be "
 324			"buffered to mlogbuf and will be printed later.\n");
 325		printk(KERN_ERR "INIT: If messages would not printed after "
 326			"this INIT, wait 30sec and assert INIT again.\n");
 327		if (!mlogbuf_timestamp)
 328			mlogbuf_timestamp = jiffies;
 329		return;
 330	}
 331	spin_unlock(&mlogbuf_rlock);
 332	ia64_mlogbuf_dump();
 333}
 334
 335static inline void
 336ia64_mca_spin(const char *func)
 337{
 338	if (monarch_cpu == smp_processor_id())
 339		ia64_mlogbuf_finish(0);
 340	mprintk(KERN_EMERG "%s: spinning here, not returning to SAL\n", func);
 341	while (1)
 342		cpu_relax();
 343}
 344/*
 345 * IA64_MCA log support
 346 */
 347#define IA64_MAX_LOGS		2	/* Double-buffering for nested MCAs */
 348#define IA64_MAX_LOG_TYPES      4   /* MCA, INIT, CMC, CPE */
 349
 350typedef struct ia64_state_log_s
 351{
 352	spinlock_t	isl_lock;
 353	int		isl_index;
 354	unsigned long	isl_count;
 355	ia64_err_rec_t  *isl_log[IA64_MAX_LOGS]; /* need space to store header + error log */
 356} ia64_state_log_t;
 357
 358static ia64_state_log_t ia64_state_log[IA64_MAX_LOG_TYPES];
 359
 360#define IA64_LOG_LOCK_INIT(it) spin_lock_init(&ia64_state_log[it].isl_lock)
 361#define IA64_LOG_LOCK(it)      spin_lock_irqsave(&ia64_state_log[it].isl_lock, s)
 362#define IA64_LOG_UNLOCK(it)    spin_unlock_irqrestore(&ia64_state_log[it].isl_lock,s)
 363#define IA64_LOG_NEXT_INDEX(it)    ia64_state_log[it].isl_index
 364#define IA64_LOG_CURR_INDEX(it)    1 - ia64_state_log[it].isl_index
 365#define IA64_LOG_INDEX_INC(it) \
 366    {ia64_state_log[it].isl_index = 1 - ia64_state_log[it].isl_index; \
 367    ia64_state_log[it].isl_count++;}
 368#define IA64_LOG_INDEX_DEC(it) \
 369    ia64_state_log[it].isl_index = 1 - ia64_state_log[it].isl_index
 370#define IA64_LOG_NEXT_BUFFER(it)   (void *)((ia64_state_log[it].isl_log[IA64_LOG_NEXT_INDEX(it)]))
 371#define IA64_LOG_CURR_BUFFER(it)   (void *)((ia64_state_log[it].isl_log[IA64_LOG_CURR_INDEX(it)]))
 372#define IA64_LOG_COUNT(it)         ia64_state_log[it].isl_count
 373
 374static inline void ia64_log_allocate(int it, u64 size)
 375{
 376	ia64_state_log[it].isl_log[IA64_LOG_CURR_INDEX(it)] =
 377		(ia64_err_rec_t *)memblock_alloc(size, SMP_CACHE_BYTES);
 378	if (!ia64_state_log[it].isl_log[IA64_LOG_CURR_INDEX(it)])
 379		panic("%s: Failed to allocate %llu bytes\n", __func__, size);
 380
 381	ia64_state_log[it].isl_log[IA64_LOG_NEXT_INDEX(it)] =
 382		(ia64_err_rec_t *)memblock_alloc(size, SMP_CACHE_BYTES);
 383	if (!ia64_state_log[it].isl_log[IA64_LOG_NEXT_INDEX(it)])
 384		panic("%s: Failed to allocate %llu bytes\n", __func__, size);
 385}
 386
 387/*
 388 * ia64_log_init
 389 *	Reset the OS ia64 log buffer
 390 * Inputs   :   info_type   (SAL_INFO_TYPE_{MCA,INIT,CMC,CPE})
 391 * Outputs	:	None
 392 */
 393static void __init
 394ia64_log_init(int sal_info_type)
 395{
 396	u64	max_size = 0;
 397
 398	IA64_LOG_NEXT_INDEX(sal_info_type) = 0;
 399	IA64_LOG_LOCK_INIT(sal_info_type);
 400
 401	// SAL will tell us the maximum size of any error record of this type
 402	max_size = ia64_sal_get_state_info_size(sal_info_type);
 403	if (!max_size)
 404		/* alloc_bootmem() doesn't like zero-sized allocations! */
 405		return;
 406
 407	// set up OS data structures to hold error info
 408	ia64_log_allocate(sal_info_type, max_size);
 409}
 410
 411/*
 412 * ia64_log_get
 413 *
 414 *	Get the current MCA log from SAL and copy it into the OS log buffer.
 415 *
 416 *  Inputs  :   info_type   (SAL_INFO_TYPE_{MCA,INIT,CMC,CPE})
 417 *              irq_safe    whether you can use printk at this point
 418 *  Outputs :   size        (total record length)
 419 *              *buffer     (ptr to error record)
 420 *
 421 */
 422static u64
 423ia64_log_get(int sal_info_type, u8 **buffer, int irq_safe)
 424{
 425	sal_log_record_header_t     *log_buffer;
 426	u64                         total_len = 0;
 427	unsigned long               s;
 428
 429	IA64_LOG_LOCK(sal_info_type);
 430
 431	/* Get the process state information */
 432	log_buffer = IA64_LOG_NEXT_BUFFER(sal_info_type);
 433
 434	total_len = ia64_sal_get_state_info(sal_info_type, (u64 *)log_buffer);
 435
 436	if (total_len) {
 437		IA64_LOG_INDEX_INC(sal_info_type);
 438		IA64_LOG_UNLOCK(sal_info_type);
 439		if (irq_safe) {
 440			IA64_MCA_DEBUG("%s: SAL error record type %d retrieved. Record length = %ld\n",
 441				       __func__, sal_info_type, total_len);
 442		}
 443		*buffer = (u8 *) log_buffer;
 444		return total_len;
 445	} else {
 446		IA64_LOG_UNLOCK(sal_info_type);
 447		return 0;
 448	}
 449}
 450
 451/*
 452 *  ia64_mca_log_sal_error_record
 453 *
 454 *  This function retrieves a specified error record type from SAL
 455 *  and wakes up any processes waiting for error records.
 456 *
 457 *  Inputs  :   sal_info_type   (Type of error record MCA/CMC/CPE)
 458 *              FIXME: remove MCA and irq_safe.
 459 */
 460static void
 461ia64_mca_log_sal_error_record(int sal_info_type)
 462{
 463	u8 *buffer;
 464	sal_log_record_header_t *rh;
 465	u64 size;
 466	int irq_safe = sal_info_type != SAL_INFO_TYPE_MCA;
 467#ifdef IA64_MCA_DEBUG_INFO
 468	static const char * const rec_name[] = { "MCA", "INIT", "CMC", "CPE" };
 469#endif
 470
 471	size = ia64_log_get(sal_info_type, &buffer, irq_safe);
 472	if (!size)
 473		return;
 474
 475	salinfo_log_wakeup(sal_info_type, buffer, size, irq_safe);
 476
 477	if (irq_safe)
 478		IA64_MCA_DEBUG("CPU %d: SAL log contains %s error record\n",
 479			smp_processor_id(),
 480			sal_info_type < ARRAY_SIZE(rec_name) ? rec_name[sal_info_type] : "UNKNOWN");
 481
 482	/* Clear logs from corrected errors in case there's no user-level logger */
 483	rh = (sal_log_record_header_t *)buffer;
 484	if (rh->severity == sal_log_severity_corrected)
 485		ia64_sal_clear_state_info(sal_info_type);
 486}
 487
 488/*
 489 * search_mca_table
 490 *  See if the MCA surfaced in an instruction range
 491 *  that has been tagged as recoverable.
 492 *
 493 *  Inputs
 494 *	first	First address range to check
 495 *	last	Last address range to check
 496 *	ip	Instruction pointer, address we are looking for
 497 *
 498 * Return value:
 499 *      1 on Success (in the table)/ 0 on Failure (not in the  table)
 500 */
 501int
 502search_mca_table (const struct mca_table_entry *first,
 503                const struct mca_table_entry *last,
 504                unsigned long ip)
 505{
 506        const struct mca_table_entry *curr;
 507        u64 curr_start, curr_end;
 508
 509        curr = first;
 510        while (curr <= last) {
 511                curr_start = (u64) &curr->start_addr + curr->start_addr;
 512                curr_end = (u64) &curr->end_addr + curr->end_addr;
 513
 514                if ((ip >= curr_start) && (ip <= curr_end)) {
 515                        return 1;
 516                }
 517                curr++;
 518        }
 519        return 0;
 520}
 521
 522/* Given an address, look for it in the mca tables. */
 523int mca_recover_range(unsigned long addr)
 524{
 525	extern struct mca_table_entry __start___mca_table[];
 526	extern struct mca_table_entry __stop___mca_table[];
 527
 528	return search_mca_table(__start___mca_table, __stop___mca_table-1, addr);
 529}
 530EXPORT_SYMBOL_GPL(mca_recover_range);
 531
 532int cpe_vector = -1;
 533int ia64_cpe_irq = -1;
 534
 535static irqreturn_t
 536ia64_mca_cpe_int_handler (int cpe_irq, void *arg)
 537{
 538	static unsigned long	cpe_history[CPE_HISTORY_LENGTH];
 539	static int		index;
 540	static DEFINE_SPINLOCK(cpe_history_lock);
 541
 542	IA64_MCA_DEBUG("%s: received interrupt vector = %#x on CPU %d\n",
 543		       __func__, cpe_irq, smp_processor_id());
 544
 545	/* SAL spec states this should run w/ interrupts enabled */
 546	local_irq_enable();
 547
 548	spin_lock(&cpe_history_lock);
 549	if (!cpe_poll_enabled && cpe_vector >= 0) {
 550
 551		int i, count = 1; /* we know 1 happened now */
 552		unsigned long now = jiffies;
 553
 554		for (i = 0; i < CPE_HISTORY_LENGTH; i++) {
 555			if (now - cpe_history[i] <= HZ)
 556				count++;
 557		}
 558
 559		IA64_MCA_DEBUG(KERN_INFO "CPE threshold %d/%d\n", count, CPE_HISTORY_LENGTH);
 560		if (count >= CPE_HISTORY_LENGTH) {
 561
 562			cpe_poll_enabled = 1;
 563			spin_unlock(&cpe_history_lock);
 564			disable_irq_nosync(local_vector_to_irq(IA64_CPE_VECTOR));
 565
 566			/*
 567			 * Corrected errors will still be corrected, but
 568			 * make sure there's a log somewhere that indicates
 569			 * something is generating more than we can handle.
 570			 */
 571			printk(KERN_WARNING "WARNING: Switching to polling CPE handler; error records may be lost\n");
 572
 573			mod_timer(&cpe_poll_timer, jiffies + MIN_CPE_POLL_INTERVAL);
 574
 575			/* lock already released, get out now */
 576			goto out;
 577		} else {
 578			cpe_history[index++] = now;
 579			if (index == CPE_HISTORY_LENGTH)
 580				index = 0;
 581		}
 582	}
 583	spin_unlock(&cpe_history_lock);
 584out:
 585	/* Get the CPE error record and log it */
 586	ia64_mca_log_sal_error_record(SAL_INFO_TYPE_CPE);
 587
 588	local_irq_disable();
 589
 590	return IRQ_HANDLED;
 591}
 592
 593/*
 594 * ia64_mca_register_cpev
 595 *
 596 *  Register the corrected platform error vector with SAL.
 597 *
 598 *  Inputs
 599 *      cpev        Corrected Platform Error Vector number
 600 *
 601 *  Outputs
 602 *      None
 603 */
 604void
 605ia64_mca_register_cpev (int cpev)
 606{
 607	/* Register the CPE interrupt vector with SAL */
 608	struct ia64_sal_retval isrv;
 609
 610	isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_CPE_INT, SAL_MC_PARAM_MECHANISM_INT, cpev, 0, 0);
 611	if (isrv.status) {
 612		printk(KERN_ERR "Failed to register Corrected Platform "
 613		       "Error interrupt vector with SAL (status %ld)\n", isrv.status);
 614		return;
 615	}
 616
 617	IA64_MCA_DEBUG("%s: corrected platform error "
 618		       "vector %#x registered\n", __func__, cpev);
 619}
 620
 621/*
 622 * ia64_mca_cmc_vector_setup
 623 *
 624 *  Setup the corrected machine check vector register in the processor.
 625 *  (The interrupt is masked on boot. ia64_mca_late_init unmask this.)
 626 *  This function is invoked on a per-processor basis.
 627 *
 628 * Inputs
 629 *      None
 630 *
 631 * Outputs
 632 *	None
 633 */
 634void
 635ia64_mca_cmc_vector_setup (void)
 636{
 637	cmcv_reg_t	cmcv;
 638
 639	cmcv.cmcv_regval	= 0;
 640	cmcv.cmcv_mask		= 1;        /* Mask/disable interrupt at first */
 641	cmcv.cmcv_vector	= IA64_CMC_VECTOR;
 642	ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
 643
 644	IA64_MCA_DEBUG("%s: CPU %d corrected machine check vector %#x registered.\n",
 645		       __func__, smp_processor_id(), IA64_CMC_VECTOR);
 646
 647	IA64_MCA_DEBUG("%s: CPU %d CMCV = %#016lx\n",
 648		       __func__, smp_processor_id(), ia64_getreg(_IA64_REG_CR_CMCV));
 649}
 650
 651/*
 652 * ia64_mca_cmc_vector_disable
 653 *
 654 *  Mask the corrected machine check vector register in the processor.
 655 *  This function is invoked on a per-processor basis.
 656 *
 657 * Inputs
 658 *      dummy(unused)
 659 *
 660 * Outputs
 661 *	None
 662 */
 663static void
 664ia64_mca_cmc_vector_disable (void *dummy)
 665{
 666	cmcv_reg_t	cmcv;
 667
 668	cmcv.cmcv_regval = ia64_getreg(_IA64_REG_CR_CMCV);
 669
 670	cmcv.cmcv_mask = 1; /* Mask/disable interrupt */
 671	ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
 672
 673	IA64_MCA_DEBUG("%s: CPU %d corrected machine check vector %#x disabled.\n",
 674		       __func__, smp_processor_id(), cmcv.cmcv_vector);
 675}
 676
 677/*
 678 * ia64_mca_cmc_vector_enable
 679 *
 680 *  Unmask the corrected machine check vector register in the processor.
 681 *  This function is invoked on a per-processor basis.
 682 *
 683 * Inputs
 684 *      dummy(unused)
 685 *
 686 * Outputs
 687 *	None
 688 */
 689static void
 690ia64_mca_cmc_vector_enable (void *dummy)
 691{
 692	cmcv_reg_t	cmcv;
 693
 694	cmcv.cmcv_regval = ia64_getreg(_IA64_REG_CR_CMCV);
 695
 696	cmcv.cmcv_mask = 0; /* Unmask/enable interrupt */
 697	ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
 698
 699	IA64_MCA_DEBUG("%s: CPU %d corrected machine check vector %#x enabled.\n",
 700		       __func__, smp_processor_id(), cmcv.cmcv_vector);
 701}
 702
 703/*
 704 * ia64_mca_cmc_vector_disable_keventd
 705 *
 706 * Called via keventd (smp_call_function() is not safe in interrupt context) to
 707 * disable the cmc interrupt vector.
 708 */
 709static void
 710ia64_mca_cmc_vector_disable_keventd(struct work_struct *unused)
 711{
 712	on_each_cpu(ia64_mca_cmc_vector_disable, NULL, 0);
 713}
 714
 715/*
 716 * ia64_mca_cmc_vector_enable_keventd
 717 *
 718 * Called via keventd (smp_call_function() is not safe in interrupt context) to
 719 * enable the cmc interrupt vector.
 720 */
 721static void
 722ia64_mca_cmc_vector_enable_keventd(struct work_struct *unused)
 723{
 724	on_each_cpu(ia64_mca_cmc_vector_enable, NULL, 0);
 725}
 726
 727/*
 728 * ia64_mca_wakeup
 729 *
 730 *	Send an inter-cpu interrupt to wake-up a particular cpu.
 731 *
 732 *  Inputs  :   cpuid
 733 *  Outputs :   None
 734 */
 735static void
 736ia64_mca_wakeup(int cpu)
 737{
 738	ia64_send_ipi(cpu, IA64_MCA_WAKEUP_VECTOR, IA64_IPI_DM_INT, 0);
 739}
 740
 741/*
 742 * ia64_mca_wakeup_all
 743 *
 744 *	Wakeup all the slave cpus which have rendez'ed previously.
 745 *
 746 *  Inputs  :   None
 747 *  Outputs :   None
 748 */
 749static void
 750ia64_mca_wakeup_all(void)
 751{
 752	int cpu;
 753
 754	/* Clear the Rendez checkin flag for all cpus */
 755	for_each_online_cpu(cpu) {
 756		if (ia64_mc_info.imi_rendez_checkin[cpu] == IA64_MCA_RENDEZ_CHECKIN_DONE)
 757			ia64_mca_wakeup(cpu);
 758	}
 759
 760}
 761
 762/*
 763 * ia64_mca_rendez_interrupt_handler
 764 *
 765 *	This is handler used to put slave processors into spinloop
 766 *	while the monarch processor does the mca handling and later
 767 *	wake each slave up once the monarch is done.  The state
 768 *	IA64_MCA_RENDEZ_CHECKIN_DONE indicates the cpu is rendez'ed
 769 *	in SAL.  The state IA64_MCA_RENDEZ_CHECKIN_NOTDONE indicates
 770 *	the cpu has come out of OS rendezvous.
 771 *
 772 *  Inputs  :   None
 773 *  Outputs :   None
 774 */
 775static irqreturn_t
 776ia64_mca_rendez_int_handler(int rendez_irq, void *arg)
 777{
 778	unsigned long flags;
 779	int cpu = smp_processor_id();
 780	struct ia64_mca_notify_die nd =
 781		{ .sos = NULL, .monarch_cpu = &monarch_cpu };
 782
 783	/* Mask all interrupts */
 784	local_irq_save(flags);
 785
 786	NOTIFY_MCA(DIE_MCA_RENDZVOUS_ENTER, get_irq_regs(), (long)&nd, 1);
 787
 788	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_DONE;
 789	/* Register with the SAL monarch that the slave has
 790	 * reached SAL
 791	 */
 792	ia64_sal_mc_rendez();
 793
 794	NOTIFY_MCA(DIE_MCA_RENDZVOUS_PROCESS, get_irq_regs(), (long)&nd, 1);
 795
 796	/* Wait for the monarch cpu to exit. */
 797	while (monarch_cpu != -1)
 798	       cpu_relax();	/* spin until monarch leaves */
 799
 800	NOTIFY_MCA(DIE_MCA_RENDZVOUS_LEAVE, get_irq_regs(), (long)&nd, 1);
 801
 802	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
 803	/* Enable all interrupts */
 804	local_irq_restore(flags);
 805	return IRQ_HANDLED;
 806}
 807
 808/*
 809 * ia64_mca_wakeup_int_handler
 810 *
 811 *	The interrupt handler for processing the inter-cpu interrupt to the
 812 *	slave cpu which was spinning in the rendez loop.
 813 *	Since this spinning is done by turning off the interrupts and
 814 *	polling on the wakeup-interrupt bit in the IRR, there is
 815 *	nothing useful to be done in the handler.
 816 *
 817 *  Inputs  :   wakeup_irq  (Wakeup-interrupt bit)
 818 *	arg		(Interrupt handler specific argument)
 819 *  Outputs :   None
 820 *
 821 */
 822static irqreturn_t
 823ia64_mca_wakeup_int_handler(int wakeup_irq, void *arg)
 824{
 825	return IRQ_HANDLED;
 826}
 827
 828/* Function pointer for extra MCA recovery */
 829int (*ia64_mca_ucmc_extension)
 830	(void*,struct ia64_sal_os_state*)
 831	= NULL;
 832
 833int
 834ia64_reg_MCA_extension(int (*fn)(void *, struct ia64_sal_os_state *))
 835{
 836	if (ia64_mca_ucmc_extension)
 837		return 1;
 838
 839	ia64_mca_ucmc_extension = fn;
 840	return 0;
 841}
 842
 843void
 844ia64_unreg_MCA_extension(void)
 845{
 846	if (ia64_mca_ucmc_extension)
 847		ia64_mca_ucmc_extension = NULL;
 848}
 849
 850EXPORT_SYMBOL(ia64_reg_MCA_extension);
 851EXPORT_SYMBOL(ia64_unreg_MCA_extension);
 852
 853
 854static inline void
 855copy_reg(const u64 *fr, u64 fnat, unsigned long *tr, unsigned long *tnat)
 856{
 857	u64 fslot, tslot, nat;
 858	*tr = *fr;
 859	fslot = ((unsigned long)fr >> 3) & 63;
 860	tslot = ((unsigned long)tr >> 3) & 63;
 861	*tnat &= ~(1UL << tslot);
 862	nat = (fnat >> fslot) & 1;
 863	*tnat |= (nat << tslot);
 864}
 865
 866/* Change the comm field on the MCA/INT task to include the pid that
 867 * was interrupted, it makes for easier debugging.  If that pid was 0
 868 * (swapper or nested MCA/INIT) then use the start of the previous comm
 869 * field suffixed with its cpu.
 870 */
 871
 872static void
 873ia64_mca_modify_comm(const struct task_struct *previous_current)
 874{
 875	char *p, comm[sizeof(current->comm)];
 876	if (previous_current->pid)
 877		snprintf(comm, sizeof(comm), "%s %d",
 878			current->comm, previous_current->pid);
 879	else {
 880		int l;
 881		if ((p = strchr(previous_current->comm, ' ')))
 882			l = p - previous_current->comm;
 883		else
 884			l = strlen(previous_current->comm);
 885		snprintf(comm, sizeof(comm), "%s %*s %d",
 886			current->comm, l, previous_current->comm,
 887			task_thread_info(previous_current)->cpu);
 888	}
 889	memcpy(current->comm, comm, sizeof(current->comm));
 890}
 891
 892static void
 893finish_pt_regs(struct pt_regs *regs, struct ia64_sal_os_state *sos,
 894		unsigned long *nat)
 895{
 896	const pal_min_state_area_t *ms = sos->pal_min_state;
 897	const u64 *bank;
 898
 899	/* If ipsr.ic then use pmsa_{iip,ipsr,ifs}, else use
 900	 * pmsa_{xip,xpsr,xfs}
 901	 */
 902	if (ia64_psr(regs)->ic) {
 903		regs->cr_iip = ms->pmsa_iip;
 904		regs->cr_ipsr = ms->pmsa_ipsr;
 905		regs->cr_ifs = ms->pmsa_ifs;
 906	} else {
 907		regs->cr_iip = ms->pmsa_xip;
 908		regs->cr_ipsr = ms->pmsa_xpsr;
 909		regs->cr_ifs = ms->pmsa_xfs;
 910
 911		sos->iip = ms->pmsa_iip;
 912		sos->ipsr = ms->pmsa_ipsr;
 913		sos->ifs = ms->pmsa_ifs;
 914	}
 915	regs->pr = ms->pmsa_pr;
 916	regs->b0 = ms->pmsa_br0;
 917	regs->ar_rsc = ms->pmsa_rsc;
 918	copy_reg(&ms->pmsa_gr[1-1], ms->pmsa_nat_bits, &regs->r1, nat);
 919	copy_reg(&ms->pmsa_gr[2-1], ms->pmsa_nat_bits, &regs->r2, nat);
 920	copy_reg(&ms->pmsa_gr[3-1], ms->pmsa_nat_bits, &regs->r3, nat);
 921	copy_reg(&ms->pmsa_gr[8-1], ms->pmsa_nat_bits, &regs->r8, nat);
 922	copy_reg(&ms->pmsa_gr[9-1], ms->pmsa_nat_bits, &regs->r9, nat);
 923	copy_reg(&ms->pmsa_gr[10-1], ms->pmsa_nat_bits, &regs->r10, nat);
 924	copy_reg(&ms->pmsa_gr[11-1], ms->pmsa_nat_bits, &regs->r11, nat);
 925	copy_reg(&ms->pmsa_gr[12-1], ms->pmsa_nat_bits, &regs->r12, nat);
 926	copy_reg(&ms->pmsa_gr[13-1], ms->pmsa_nat_bits, &regs->r13, nat);
 927	copy_reg(&ms->pmsa_gr[14-1], ms->pmsa_nat_bits, &regs->r14, nat);
 928	copy_reg(&ms->pmsa_gr[15-1], ms->pmsa_nat_bits, &regs->r15, nat);
 929	if (ia64_psr(regs)->bn)
 930		bank = ms->pmsa_bank1_gr;
 931	else
 932		bank = ms->pmsa_bank0_gr;
 933	copy_reg(&bank[16-16], ms->pmsa_nat_bits, &regs->r16, nat);
 934	copy_reg(&bank[17-16], ms->pmsa_nat_bits, &regs->r17, nat);
 935	copy_reg(&bank[18-16], ms->pmsa_nat_bits, &regs->r18, nat);
 936	copy_reg(&bank[19-16], ms->pmsa_nat_bits, &regs->r19, nat);
 937	copy_reg(&bank[20-16], ms->pmsa_nat_bits, &regs->r20, nat);
 938	copy_reg(&bank[21-16], ms->pmsa_nat_bits, &regs->r21, nat);
 939	copy_reg(&bank[22-16], ms->pmsa_nat_bits, &regs->r22, nat);
 940	copy_reg(&bank[23-16], ms->pmsa_nat_bits, &regs->r23, nat);
 941	copy_reg(&bank[24-16], ms->pmsa_nat_bits, &regs->r24, nat);
 942	copy_reg(&bank[25-16], ms->pmsa_nat_bits, &regs->r25, nat);
 943	copy_reg(&bank[26-16], ms->pmsa_nat_bits, &regs->r26, nat);
 944	copy_reg(&bank[27-16], ms->pmsa_nat_bits, &regs->r27, nat);
 945	copy_reg(&bank[28-16], ms->pmsa_nat_bits, &regs->r28, nat);
 946	copy_reg(&bank[29-16], ms->pmsa_nat_bits, &regs->r29, nat);
 947	copy_reg(&bank[30-16], ms->pmsa_nat_bits, &regs->r30, nat);
 948	copy_reg(&bank[31-16], ms->pmsa_nat_bits, &regs->r31, nat);
 949}
 950
 951/* On entry to this routine, we are running on the per cpu stack, see
 952 * mca_asm.h.  The original stack has not been touched by this event.  Some of
 953 * the original stack's registers will be in the RBS on this stack.  This stack
 954 * also contains a partial pt_regs and switch_stack, the rest of the data is in
 955 * PAL minstate.
 956 *
 957 * The first thing to do is modify the original stack to look like a blocked
 958 * task so we can run backtrace on the original task.  Also mark the per cpu
 959 * stack as current to ensure that we use the correct task state, it also means
 960 * that we can do backtrace on the MCA/INIT handler code itself.
 961 */
 962
 963static struct task_struct *
 964ia64_mca_modify_original_stack(struct pt_regs *regs,
 965		const struct switch_stack *sw,
 966		struct ia64_sal_os_state *sos,
 967		const char *type)
 968{
 969	char *p;
 970	ia64_va va;
 971	extern char ia64_leave_kernel[];	/* Need asm address, not function descriptor */
 972	const pal_min_state_area_t *ms = sos->pal_min_state;
 973	struct task_struct *previous_current;
 974	struct pt_regs *old_regs;
 975	struct switch_stack *old_sw;
 976	unsigned size = sizeof(struct pt_regs) +
 977			sizeof(struct switch_stack) + 16;
 978	unsigned long *old_bspstore, *old_bsp;
 979	unsigned long *new_bspstore, *new_bsp;
 980	unsigned long old_unat, old_rnat, new_rnat, nat;
 981	u64 slots, loadrs = regs->loadrs;
 982	u64 r12 = ms->pmsa_gr[12-1], r13 = ms->pmsa_gr[13-1];
 983	u64 ar_bspstore = regs->ar_bspstore;
 984	u64 ar_bsp = regs->ar_bspstore + (loadrs >> 16);
 985	const char *msg;
 986	int cpu = smp_processor_id();
 987
 988	previous_current = curr_task(cpu);
 989	ia64_set_curr_task(cpu, current);
 990	if ((p = strchr(current->comm, ' ')))
 991		*p = '\0';
 992
 993	/* Best effort attempt to cope with MCA/INIT delivered while in
 994	 * physical mode.
 995	 */
 996	regs->cr_ipsr = ms->pmsa_ipsr;
 997	if (ia64_psr(regs)->dt == 0) {
 998		va.l = r12;
 999		if (va.f.reg == 0) {
1000			va.f.reg = 7;
1001			r12 = va.l;
1002		}
1003		va.l = r13;
1004		if (va.f.reg == 0) {
1005			va.f.reg = 7;
1006			r13 = va.l;
1007		}
1008	}
1009	if (ia64_psr(regs)->rt == 0) {
1010		va.l = ar_bspstore;
1011		if (va.f.reg == 0) {
1012			va.f.reg = 7;
1013			ar_bspstore = va.l;
1014		}
1015		va.l = ar_bsp;
1016		if (va.f.reg == 0) {
1017			va.f.reg = 7;
1018			ar_bsp = va.l;
1019		}
1020	}
1021
1022	/* mca_asm.S ia64_old_stack() cannot assume that the dirty registers
1023	 * have been copied to the old stack, the old stack may fail the
1024	 * validation tests below.  So ia64_old_stack() must restore the dirty
1025	 * registers from the new stack.  The old and new bspstore probably
1026	 * have different alignments, so loadrs calculated on the old bsp
1027	 * cannot be used to restore from the new bsp.  Calculate a suitable
1028	 * loadrs for the new stack and save it in the new pt_regs, where
1029	 * ia64_old_stack() can get it.
1030	 */
1031	old_bspstore = (unsigned long *)ar_bspstore;
1032	old_bsp = (unsigned long *)ar_bsp;
1033	slots = ia64_rse_num_regs(old_bspstore, old_bsp);
1034	new_bspstore = (unsigned long *)((u64)current + IA64_RBS_OFFSET);
1035	new_bsp = ia64_rse_skip_regs(new_bspstore, slots);
1036	regs->loadrs = (new_bsp - new_bspstore) * 8 << 16;
1037
1038	/* Verify the previous stack state before we change it */
1039	if (user_mode(regs)) {
1040		msg = "occurred in user space";
1041		/* previous_current is guaranteed to be valid when the task was
1042		 * in user space, so ...
1043		 */
1044		ia64_mca_modify_comm(previous_current);
1045		goto no_mod;
1046	}
1047
1048	if (r13 != sos->prev_IA64_KR_CURRENT) {
1049		msg = "inconsistent previous current and r13";
1050		goto no_mod;
1051	}
1052
1053	if (!mca_recover_range(ms->pmsa_iip)) {
1054		if ((r12 - r13) >= KERNEL_STACK_SIZE) {
1055			msg = "inconsistent r12 and r13";
1056			goto no_mod;
1057		}
1058		if ((ar_bspstore - r13) >= KERNEL_STACK_SIZE) {
1059			msg = "inconsistent ar.bspstore and r13";
1060			goto no_mod;
1061		}
1062		va.p = old_bspstore;
1063		if (va.f.reg < 5) {
1064			msg = "old_bspstore is in the wrong region";
1065			goto no_mod;
1066		}
1067		if ((ar_bsp - r13) >= KERNEL_STACK_SIZE) {
1068			msg = "inconsistent ar.bsp and r13";
1069			goto no_mod;
1070		}
1071		size += (ia64_rse_skip_regs(old_bspstore, slots) - old_bspstore) * 8;
1072		if (ar_bspstore + size > r12) {
1073			msg = "no room for blocked state";
1074			goto no_mod;
1075		}
1076	}
1077
1078	ia64_mca_modify_comm(previous_current);
1079
1080	/* Make the original task look blocked.  First stack a struct pt_regs,
1081	 * describing the state at the time of interrupt.  mca_asm.S built a
1082	 * partial pt_regs, copy it and fill in the blanks using minstate.
1083	 */
1084	p = (char *)r12 - sizeof(*regs);
1085	old_regs = (struct pt_regs *)p;
1086	memcpy(old_regs, regs, sizeof(*regs));
1087	old_regs->loadrs = loadrs;
1088	old_unat = old_regs->ar_unat;
1089	finish_pt_regs(old_regs, sos, &old_unat);
1090
1091	/* Next stack a struct switch_stack.  mca_asm.S built a partial
1092	 * switch_stack, copy it and fill in the blanks using pt_regs and
1093	 * minstate.
1094	 *
1095	 * In the synthesized switch_stack, b0 points to ia64_leave_kernel,
1096	 * ar.pfs is set to 0.
1097	 *
1098	 * unwind.c::unw_unwind() does special processing for interrupt frames.
1099	 * It checks if the PRED_NON_SYSCALL predicate is set, if the predicate
1100	 * is clear then unw_unwind() does _not_ adjust bsp over pt_regs.  Not
1101	 * that this is documented, of course.  Set PRED_NON_SYSCALL in the
1102	 * switch_stack on the original stack so it will unwind correctly when
1103	 * unwind.c reads pt_regs.
1104	 *
1105	 * thread.ksp is updated to point to the synthesized switch_stack.
1106	 */
1107	p -= sizeof(struct switch_stack);
1108	old_sw = (struct switch_stack *)p;
1109	memcpy(old_sw, sw, sizeof(*sw));
1110	old_sw->caller_unat = old_unat;
1111	old_sw->ar_fpsr = old_regs->ar_fpsr;
1112	copy_reg(&ms->pmsa_gr[4-1], ms->pmsa_nat_bits, &old_sw->r4, &old_unat);
1113	copy_reg(&ms->pmsa_gr[5-1], ms->pmsa_nat_bits, &old_sw->r5, &old_unat);
1114	copy_reg(&ms->pmsa_gr[6-1], ms->pmsa_nat_bits, &old_sw->r6, &old_unat);
1115	copy_reg(&ms->pmsa_gr[7-1], ms->pmsa_nat_bits, &old_sw->r7, &old_unat);
1116	old_sw->b0 = (u64)ia64_leave_kernel;
1117	old_sw->b1 = ms->pmsa_br1;
1118	old_sw->ar_pfs = 0;
1119	old_sw->ar_unat = old_unat;
1120	old_sw->pr = old_regs->pr | (1UL << PRED_NON_SYSCALL);
1121	previous_current->thread.ksp = (u64)p - 16;
1122
1123	/* Finally copy the original stack's registers back to its RBS.
1124	 * Registers from ar.bspstore through ar.bsp at the time of the event
1125	 * are in the current RBS, copy them back to the original stack.  The
1126	 * copy must be done register by register because the original bspstore
1127	 * and the current one have different alignments, so the saved RNAT
1128	 * data occurs at different places.
1129	 *
1130	 * mca_asm does cover, so the old_bsp already includes all registers at
1131	 * the time of MCA/INIT.  It also does flushrs, so all registers before
1132	 * this function have been written to backing store on the MCA/INIT
1133	 * stack.
1134	 */
1135	new_rnat = ia64_get_rnat(ia64_rse_rnat_addr(new_bspstore));
1136	old_rnat = regs->ar_rnat;
1137	while (slots--) {
1138		if (ia64_rse_is_rnat_slot(new_bspstore)) {
1139			new_rnat = ia64_get_rnat(new_bspstore++);
1140		}
1141		if (ia64_rse_is_rnat_slot(old_bspstore)) {
1142			*old_bspstore++ = old_rnat;
1143			old_rnat = 0;
1144		}
1145		nat = (new_rnat >> ia64_rse_slot_num(new_bspstore)) & 1UL;
1146		old_rnat &= ~(1UL << ia64_rse_slot_num(old_bspstore));
1147		old_rnat |= (nat << ia64_rse_slot_num(old_bspstore));
1148		*old_bspstore++ = *new_bspstore++;
1149	}
1150	old_sw->ar_bspstore = (unsigned long)old_bspstore;
1151	old_sw->ar_rnat = old_rnat;
1152
1153	sos->prev_task = previous_current;
1154	return previous_current;
1155
1156no_mod:
1157	mprintk(KERN_INFO "cpu %d, %s %s, original stack not modified\n",
1158			smp_processor_id(), type, msg);
1159	old_unat = regs->ar_unat;
1160	finish_pt_regs(regs, sos, &old_unat);
1161	return previous_current;
1162}
1163
1164/* The monarch/slave interaction is based on monarch_cpu and requires that all
1165 * slaves have entered rendezvous before the monarch leaves.  If any cpu has
1166 * not entered rendezvous yet then wait a bit.  The assumption is that any
1167 * slave that has not rendezvoused after a reasonable time is never going to do
1168 * so.  In this context, slave includes cpus that respond to the MCA rendezvous
1169 * interrupt, as well as cpus that receive the INIT slave event.
1170 */
1171
1172static void
1173ia64_wait_for_slaves(int monarch, const char *type)
1174{
1175	int c, i , wait;
1176
1177	/*
1178	 * wait 5 seconds total for slaves (arbitrary)
1179	 */
1180	for (i = 0; i < 5000; i++) {
1181		wait = 0;
1182		for_each_online_cpu(c) {
1183			if (c == monarch)
1184				continue;
1185			if (ia64_mc_info.imi_rendez_checkin[c]
1186					== IA64_MCA_RENDEZ_CHECKIN_NOTDONE) {
1187				udelay(1000);		/* short wait */
1188				wait = 1;
1189				break;
1190			}
1191		}
1192		if (!wait)
1193			goto all_in;
1194	}
1195
1196	/*
1197	 * Maybe slave(s) dead. Print buffered messages immediately.
1198	 */
1199	ia64_mlogbuf_finish(0);
1200	mprintk(KERN_INFO "OS %s slave did not rendezvous on cpu", type);
1201	for_each_online_cpu(c) {
1202		if (c == monarch)
1203			continue;
1204		if (ia64_mc_info.imi_rendez_checkin[c] == IA64_MCA_RENDEZ_CHECKIN_NOTDONE)
1205			mprintk(" %d", c);
1206	}
1207	mprintk("\n");
1208	return;
1209
1210all_in:
1211	mprintk(KERN_INFO "All OS %s slaves have reached rendezvous\n", type);
1212	return;
1213}
1214
1215/*  mca_insert_tr
1216 *
1217 *  Switch rid when TR reload and needed!
1218 *  iord: 1: itr, 2: itr;
1219 *
1220*/
1221static void mca_insert_tr(u64 iord)
1222{
1223
1224	int i;
1225	u64 old_rr;
1226	struct ia64_tr_entry *p;
1227	unsigned long psr;
1228	int cpu = smp_processor_id();
1229
1230	if (!ia64_idtrs[cpu])
1231		return;
1232
1233	psr = ia64_clear_ic();
1234	for (i = IA64_TR_ALLOC_BASE; i < IA64_TR_ALLOC_MAX; i++) {
1235		p = ia64_idtrs[cpu] + (iord - 1) * IA64_TR_ALLOC_MAX;
1236		if (p->pte & 0x1) {
1237			old_rr = ia64_get_rr(p->ifa);
1238			if (old_rr != p->rr) {
1239				ia64_set_rr(p->ifa, p->rr);
1240				ia64_srlz_d();
1241			}
1242			ia64_ptr(iord, p->ifa, p->itir >> 2);
1243			ia64_srlz_i();
1244			if (iord & 0x1) {
1245				ia64_itr(0x1, i, p->ifa, p->pte, p->itir >> 2);
1246				ia64_srlz_i();
1247			}
1248			if (iord & 0x2) {
1249				ia64_itr(0x2, i, p->ifa, p->pte, p->itir >> 2);
1250				ia64_srlz_i();
1251			}
1252			if (old_rr != p->rr) {
1253				ia64_set_rr(p->ifa, old_rr);
1254				ia64_srlz_d();
1255			}
1256		}
1257	}
1258	ia64_set_psr(psr);
1259}
1260
1261/*
1262 * ia64_mca_handler
1263 *
1264 *	This is uncorrectable machine check handler called from OS_MCA
1265 *	dispatch code which is in turn called from SAL_CHECK().
1266 *	This is the place where the core of OS MCA handling is done.
1267 *	Right now the logs are extracted and displayed in a well-defined
1268 *	format. This handler code is supposed to be run only on the
1269 *	monarch processor. Once the monarch is done with MCA handling
1270 *	further MCA logging is enabled by clearing logs.
1271 *	Monarch also has the duty of sending wakeup-IPIs to pull the
1272 *	slave processors out of rendezvous spinloop.
1273 *
1274 *	If multiple processors call into OS_MCA, the first will become
1275 *	the monarch.  Subsequent cpus will be recorded in the mca_cpu
1276 *	bitmask.  After the first monarch has processed its MCA, it
1277 *	will wake up the next cpu in the mca_cpu bitmask and then go
1278 *	into the rendezvous loop.  When all processors have serviced
1279 *	their MCA, the last monarch frees up the rest of the processors.
1280 */
1281void
1282ia64_mca_handler(struct pt_regs *regs, struct switch_stack *sw,
1283		 struct ia64_sal_os_state *sos)
1284{
1285	int recover, cpu = smp_processor_id();
1286	struct task_struct *previous_current;
1287	struct ia64_mca_notify_die nd =
1288		{ .sos = sos, .monarch_cpu = &monarch_cpu, .data = &recover };
1289	static atomic_t mca_count;
1290	static cpumask_t mca_cpu;
1291
1292	if (atomic_add_return(1, &mca_count) == 1) {
1293		monarch_cpu = cpu;
1294		sos->monarch = 1;
1295	} else {
1296		cpumask_set_cpu(cpu, &mca_cpu);
1297		sos->monarch = 0;
1298	}
1299	mprintk(KERN_INFO "Entered OS MCA handler. PSP=%lx cpu=%d "
1300		"monarch=%ld\n", sos->proc_state_param, cpu, sos->monarch);
1301
1302	previous_current = ia64_mca_modify_original_stack(regs, sw, sos, "MCA");
1303
1304	NOTIFY_MCA(DIE_MCA_MONARCH_ENTER, regs, (long)&nd, 1);
1305
1306	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_CONCURRENT_MCA;
1307	if (sos->monarch) {
1308		ia64_wait_for_slaves(cpu, "MCA");
1309
1310		/* Wakeup all the processors which are spinning in the
1311		 * rendezvous loop.  They will leave SAL, then spin in the OS
1312		 * with interrupts disabled until this monarch cpu leaves the
1313		 * MCA handler.  That gets control back to the OS so we can
1314		 * backtrace the other cpus, backtrace when spinning in SAL
1315		 * does not work.
1316		 */
1317		ia64_mca_wakeup_all();
1318	} else {
1319		while (cpumask_test_cpu(cpu, &mca_cpu))
1320			cpu_relax();	/* spin until monarch wakes us */
1321	}
1322
1323	NOTIFY_MCA(DIE_MCA_MONARCH_PROCESS, regs, (long)&nd, 1);
1324
1325	/* Get the MCA error record and log it */
1326	ia64_mca_log_sal_error_record(SAL_INFO_TYPE_MCA);
1327
1328	/* MCA error recovery */
1329	recover = (ia64_mca_ucmc_extension
1330		&& ia64_mca_ucmc_extension(
1331			IA64_LOG_CURR_BUFFER(SAL_INFO_TYPE_MCA),
1332			sos));
1333
1334	if (recover) {
1335		sal_log_record_header_t *rh = IA64_LOG_CURR_BUFFER(SAL_INFO_TYPE_MCA);
1336		rh->severity = sal_log_severity_corrected;
1337		ia64_sal_clear_state_info(SAL_INFO_TYPE_MCA);
1338		sos->os_status = IA64_MCA_CORRECTED;
1339	} else {
1340		/* Dump buffered message to console */
1341		ia64_mlogbuf_finish(1);
1342	}
1343
1344	if (__this_cpu_read(ia64_mca_tr_reload)) {
1345		mca_insert_tr(0x1); /*Reload dynamic itrs*/
1346		mca_insert_tr(0x2); /*Reload dynamic itrs*/
1347	}
1348
1349	NOTIFY_MCA(DIE_MCA_MONARCH_LEAVE, regs, (long)&nd, 1);
1350
1351	if (atomic_dec_return(&mca_count) > 0) {
1352		int i;
1353
1354		/* wake up the next monarch cpu,
1355		 * and put this cpu in the rendez loop.
1356		 */
1357		for_each_online_cpu(i) {
1358			if (cpumask_test_cpu(i, &mca_cpu)) {
1359				monarch_cpu = i;
1360				cpumask_clear_cpu(i, &mca_cpu);	/* wake next cpu */
1361				while (monarch_cpu != -1)
1362					cpu_relax();	/* spin until last cpu leaves */
1363				ia64_set_curr_task(cpu, previous_current);
1364				ia64_mc_info.imi_rendez_checkin[cpu]
1365						= IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1366				return;
1367			}
1368		}
1369	}
1370	ia64_set_curr_task(cpu, previous_current);
1371	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1372	monarch_cpu = -1;	/* This frees the slaves and previous monarchs */
1373}
1374
1375static DECLARE_WORK(cmc_disable_work, ia64_mca_cmc_vector_disable_keventd);
1376static DECLARE_WORK(cmc_enable_work, ia64_mca_cmc_vector_enable_keventd);
1377
1378/*
1379 * ia64_mca_cmc_int_handler
1380 *
1381 *  This is corrected machine check interrupt handler.
1382 *	Right now the logs are extracted and displayed in a well-defined
1383 *	format.
1384 *
1385 * Inputs
1386 *      interrupt number
1387 *      client data arg ptr
1388 *
1389 * Outputs
1390 *	None
1391 */
1392static irqreturn_t
1393ia64_mca_cmc_int_handler(int cmc_irq, void *arg)
1394{
1395	static unsigned long	cmc_history[CMC_HISTORY_LENGTH];
1396	static int		index;
1397	static DEFINE_SPINLOCK(cmc_history_lock);
1398
1399	IA64_MCA_DEBUG("%s: received interrupt vector = %#x on CPU %d\n",
1400		       __func__, cmc_irq, smp_processor_id());
1401
1402	/* SAL spec states this should run w/ interrupts enabled */
1403	local_irq_enable();
1404
1405	spin_lock(&cmc_history_lock);
1406	if (!cmc_polling_enabled) {
1407		int i, count = 1; /* we know 1 happened now */
1408		unsigned long now = jiffies;
1409
1410		for (i = 0; i < CMC_HISTORY_LENGTH; i++) {
1411			if (now - cmc_history[i] <= HZ)
1412				count++;
1413		}
1414
1415		IA64_MCA_DEBUG(KERN_INFO "CMC threshold %d/%d\n", count, CMC_HISTORY_LENGTH);
1416		if (count >= CMC_HISTORY_LENGTH) {
1417
1418			cmc_polling_enabled = 1;
1419			spin_unlock(&cmc_history_lock);
1420			/* If we're being hit with CMC interrupts, we won't
1421			 * ever execute the schedule_work() below.  Need to
1422			 * disable CMC interrupts on this processor now.
1423			 */
1424			ia64_mca_cmc_vector_disable(NULL);
1425			schedule_work(&cmc_disable_work);
1426
1427			/*
1428			 * Corrected errors will still be corrected, but
1429			 * make sure there's a log somewhere that indicates
1430			 * something is generating more than we can handle.
1431			 */
1432			printk(KERN_WARNING "WARNING: Switching to polling CMC handler; error records may be lost\n");
1433
1434			mod_timer(&cmc_poll_timer, jiffies + CMC_POLL_INTERVAL);
1435
1436			/* lock already released, get out now */
1437			goto out;
1438		} else {
1439			cmc_history[index++] = now;
1440			if (index == CMC_HISTORY_LENGTH)
1441				index = 0;
1442		}
1443	}
1444	spin_unlock(&cmc_history_lock);
1445out:
1446	/* Get the CMC error record and log it */
1447	ia64_mca_log_sal_error_record(SAL_INFO_TYPE_CMC);
1448
1449	local_irq_disable();
1450
1451	return IRQ_HANDLED;
1452}
1453
1454/*
1455 *  ia64_mca_cmc_int_caller
1456 *
1457 * 	Triggered by sw interrupt from CMC polling routine.  Calls
1458 * 	real interrupt handler and either triggers a sw interrupt
1459 * 	on the next cpu or does cleanup at the end.
1460 *
1461 * Inputs
1462 *	interrupt number
1463 *	client data arg ptr
1464 * Outputs
1465 * 	handled
1466 */
1467static irqreturn_t
1468ia64_mca_cmc_int_caller(int cmc_irq, void *arg)
1469{
1470	static int start_count = -1;
1471	unsigned int cpuid;
1472
1473	cpuid = smp_processor_id();
1474
1475	/* If first cpu, update count */
1476	if (start_count == -1)
1477		start_count = IA64_LOG_COUNT(SAL_INFO_TYPE_CMC);
1478
1479	ia64_mca_cmc_int_handler(cmc_irq, arg);
1480
1481	cpuid = cpumask_next(cpuid+1, cpu_online_mask);
1482
1483	if (cpuid < nr_cpu_ids) {
1484		ia64_send_ipi(cpuid, IA64_CMCP_VECTOR, IA64_IPI_DM_INT, 0);
1485	} else {
1486		/* If no log record, switch out of polling mode */
1487		if (start_count == IA64_LOG_COUNT(SAL_INFO_TYPE_CMC)) {
1488
1489			printk(KERN_WARNING "Returning to interrupt driven CMC handler\n");
1490			schedule_work(&cmc_enable_work);
1491			cmc_polling_enabled = 0;
1492
1493		} else {
1494
1495			mod_timer(&cmc_poll_timer, jiffies + CMC_POLL_INTERVAL);
1496		}
1497
1498		start_count = -1;
1499	}
1500
1501	return IRQ_HANDLED;
1502}
1503
1504/*
1505 *  ia64_mca_cmc_poll
1506 *
1507 *	Poll for Corrected Machine Checks (CMCs)
1508 *
1509 * Inputs   :   dummy(unused)
1510 * Outputs  :   None
1511 *
1512 */
1513static void
1514ia64_mca_cmc_poll (struct timer_list *unused)
1515{
1516	/* Trigger a CMC interrupt cascade  */
1517	ia64_send_ipi(cpumask_first(cpu_online_mask), IA64_CMCP_VECTOR,
1518							IA64_IPI_DM_INT, 0);
1519}
1520
1521/*
1522 *  ia64_mca_cpe_int_caller
1523 *
1524 * 	Triggered by sw interrupt from CPE polling routine.  Calls
1525 * 	real interrupt handler and either triggers a sw interrupt
1526 * 	on the next cpu or does cleanup at the end.
1527 *
1528 * Inputs
1529 *	interrupt number
1530 *	client data arg ptr
1531 * Outputs
1532 * 	handled
1533 */
1534static irqreturn_t
1535ia64_mca_cpe_int_caller(int cpe_irq, void *arg)
1536{
1537	static int start_count = -1;
1538	static int poll_time = MIN_CPE_POLL_INTERVAL;
1539	unsigned int cpuid;
1540
1541	cpuid = smp_processor_id();
1542
1543	/* If first cpu, update count */
1544	if (start_count == -1)
1545		start_count = IA64_LOG_COUNT(SAL_INFO_TYPE_CPE);
1546
1547	ia64_mca_cpe_int_handler(cpe_irq, arg);
1548
1549	cpuid = cpumask_next(cpuid+1, cpu_online_mask);
1550
1551	if (cpuid < NR_CPUS) {
1552		ia64_send_ipi(cpuid, IA64_CPEP_VECTOR, IA64_IPI_DM_INT, 0);
1553	} else {
1554		/*
1555		 * If a log was recorded, increase our polling frequency,
1556		 * otherwise, backoff or return to interrupt mode.
1557		 */
1558		if (start_count != IA64_LOG_COUNT(SAL_INFO_TYPE_CPE)) {
1559			poll_time = max(MIN_CPE_POLL_INTERVAL, poll_time / 2);
1560		} else if (cpe_vector < 0) {
1561			poll_time = min(MAX_CPE_POLL_INTERVAL, poll_time * 2);
1562		} else {
1563			poll_time = MIN_CPE_POLL_INTERVAL;
1564
1565			printk(KERN_WARNING "Returning to interrupt driven CPE handler\n");
1566			enable_irq(local_vector_to_irq(IA64_CPE_VECTOR));
1567			cpe_poll_enabled = 0;
1568		}
1569
1570		if (cpe_poll_enabled)
1571			mod_timer(&cpe_poll_timer, jiffies + poll_time);
1572		start_count = -1;
1573	}
1574
1575	return IRQ_HANDLED;
1576}
1577
1578/*
1579 *  ia64_mca_cpe_poll
1580 *
1581 *	Poll for Corrected Platform Errors (CPEs), trigger interrupt
1582 *	on first cpu, from there it will trickle through all the cpus.
1583 *
1584 * Inputs   :   dummy(unused)
1585 * Outputs  :   None
1586 *
1587 */
1588static void
1589ia64_mca_cpe_poll (struct timer_list *unused)
1590{
1591	/* Trigger a CPE interrupt cascade  */
1592	ia64_send_ipi(cpumask_first(cpu_online_mask), IA64_CPEP_VECTOR,
1593							IA64_IPI_DM_INT, 0);
1594}
1595
1596static int
1597default_monarch_init_process(struct notifier_block *self, unsigned long val, void *data)
1598{
1599	int c;
1600	struct task_struct *g, *t;
1601	if (val != DIE_INIT_MONARCH_PROCESS)
1602		return NOTIFY_DONE;
1603#ifdef CONFIG_KEXEC
1604	if (atomic_read(&kdump_in_progress))
1605		return NOTIFY_DONE;
1606#endif
1607
1608	/*
1609	 * FIXME: mlogbuf will brim over with INIT stack dumps.
1610	 * To enable show_stack from INIT, we use oops_in_progress which should
1611	 * be used in real oops. This would cause something wrong after INIT.
1612	 */
1613	BREAK_LOGLEVEL(console_loglevel);
1614	ia64_mlogbuf_dump_from_init();
1615
1616	printk(KERN_ERR "Processes interrupted by INIT -");
1617	for_each_online_cpu(c) {
1618		struct ia64_sal_os_state *s;
1619		t = __va(__per_cpu_mca[c] + IA64_MCA_CPU_INIT_STACK_OFFSET);
1620		s = (struct ia64_sal_os_state *)((char *)t + MCA_SOS_OFFSET);
1621		g = s->prev_task;
1622		if (g) {
1623			if (g->pid)
1624				printk(" %d", g->pid);
1625			else
1626				printk(" %d (cpu %d task 0x%p)", g->pid, task_cpu(g), g);
1627		}
1628	}
1629	printk("\n\n");
1630	if (read_trylock(&tasklist_lock)) {
1631		do_each_thread (g, t) {
1632			printk("\nBacktrace of pid %d (%s)\n", t->pid, t->comm);
1633			show_stack(t, NULL);
1634		} while_each_thread (g, t);
1635		read_unlock(&tasklist_lock);
1636	}
1637	/* FIXME: This will not restore zapped printk locks. */
1638	RESTORE_LOGLEVEL(console_loglevel);
1639	return NOTIFY_DONE;
1640}
1641
1642/*
1643 * C portion of the OS INIT handler
1644 *
1645 * Called from ia64_os_init_dispatch
1646 *
1647 * Inputs: pointer to pt_regs where processor info was saved.  SAL/OS state for
1648 * this event.  This code is used for both monarch and slave INIT events, see
1649 * sos->monarch.
1650 *
1651 * All INIT events switch to the INIT stack and change the previous process to
1652 * blocked status.  If one of the INIT events is the monarch then we are
1653 * probably processing the nmi button/command.  Use the monarch cpu to dump all
1654 * the processes.  The slave INIT events all spin until the monarch cpu
1655 * returns.  We can also get INIT slave events for MCA, in which case the MCA
1656 * process is the monarch.
1657 */
1658
1659void
1660ia64_init_handler(struct pt_regs *regs, struct switch_stack *sw,
1661		  struct ia64_sal_os_state *sos)
1662{
1663	static atomic_t slaves;
1664	static atomic_t monarchs;
1665	struct task_struct *previous_current;
1666	int cpu = smp_processor_id();
1667	struct ia64_mca_notify_die nd =
1668		{ .sos = sos, .monarch_cpu = &monarch_cpu };
1669
1670	NOTIFY_INIT(DIE_INIT_ENTER, regs, (long)&nd, 0);
1671
1672	mprintk(KERN_INFO "Entered OS INIT handler. PSP=%lx cpu=%d monarch=%ld\n",
1673		sos->proc_state_param, cpu, sos->monarch);
1674	salinfo_log_wakeup(SAL_INFO_TYPE_INIT, NULL, 0, 0);
1675
1676	previous_current = ia64_mca_modify_original_stack(regs, sw, sos, "INIT");
1677	sos->os_status = IA64_INIT_RESUME;
1678
1679	/* FIXME: Workaround for broken proms that drive all INIT events as
1680	 * slaves.  The last slave that enters is promoted to be a monarch.
1681	 * Remove this code in September 2006, that gives platforms a year to
1682	 * fix their proms and get their customers updated.
1683	 */
1684	if (!sos->monarch && atomic_add_return(1, &slaves) == num_online_cpus()) {
1685		mprintk(KERN_WARNING "%s: Promoting cpu %d to monarch.\n",
1686		        __func__, cpu);
1687		atomic_dec(&slaves);
1688		sos->monarch = 1;
1689	}
1690
1691	/* FIXME: Workaround for broken proms that drive all INIT events as
1692	 * monarchs.  Second and subsequent monarchs are demoted to slaves.
1693	 * Remove this code in September 2006, that gives platforms a year to
1694	 * fix their proms and get their customers updated.
1695	 */
1696	if (sos->monarch && atomic_add_return(1, &monarchs) > 1) {
1697		mprintk(KERN_WARNING "%s: Demoting cpu %d to slave.\n",
1698			       __func__, cpu);
1699		atomic_dec(&monarchs);
1700		sos->monarch = 0;
1701	}
1702
1703	if (!sos->monarch) {
1704		ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_INIT;
1705
1706#ifdef CONFIG_KEXEC
1707		while (monarch_cpu == -1 && !atomic_read(&kdump_in_progress))
1708			udelay(1000);
1709#else
1710		while (monarch_cpu == -1)
1711			cpu_relax();	/* spin until monarch enters */
1712#endif
1713
1714		NOTIFY_INIT(DIE_INIT_SLAVE_ENTER, regs, (long)&nd, 1);
1715		NOTIFY_INIT(DIE_INIT_SLAVE_PROCESS, regs, (long)&nd, 1);
1716
1717#ifdef CONFIG_KEXEC
1718		while (monarch_cpu != -1 && !atomic_read(&kdump_in_progress))
1719			udelay(1000);
1720#else
1721		while (monarch_cpu != -1)
1722			cpu_relax();	/* spin until monarch leaves */
1723#endif
1724
1725		NOTIFY_INIT(DIE_INIT_SLAVE_LEAVE, regs, (long)&nd, 1);
1726
1727		mprintk("Slave on cpu %d returning to normal service.\n", cpu);
1728		ia64_set_curr_task(cpu, previous_current);
1729		ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1730		atomic_dec(&slaves);
1731		return;
1732	}
1733
1734	monarch_cpu = cpu;
1735	NOTIFY_INIT(DIE_INIT_MONARCH_ENTER, regs, (long)&nd, 1);
1736
1737	/*
1738	 * Wait for a bit.  On some machines (e.g., HP's zx2000 and zx6000, INIT can be
1739	 * generated via the BMC's command-line interface, but since the console is on the
1740	 * same serial line, the user will need some time to switch out of the BMC before
1741	 * the dump begins.
1742	 */
1743	mprintk("Delaying for 5 seconds...\n");
1744	udelay(5*1000000);
1745	ia64_wait_for_slaves(cpu, "INIT");
1746	/* If nobody intercepts DIE_INIT_MONARCH_PROCESS then we drop through
1747	 * to default_monarch_init_process() above and just print all the
1748	 * tasks.
1749	 */
1750	NOTIFY_INIT(DIE_INIT_MONARCH_PROCESS, regs, (long)&nd, 1);
1751	NOTIFY_INIT(DIE_INIT_MONARCH_LEAVE, regs, (long)&nd, 1);
1752
1753	mprintk("\nINIT dump complete.  Monarch on cpu %d returning to normal service.\n", cpu);
1754	atomic_dec(&monarchs);
1755	ia64_set_curr_task(cpu, previous_current);
1756	monarch_cpu = -1;
1757	return;
1758}
1759
1760static int __init
1761ia64_mca_disable_cpe_polling(char *str)
1762{
1763	cpe_poll_enabled = 0;
1764	return 1;
1765}
1766
1767__setup("disable_cpe_poll", ia64_mca_disable_cpe_polling);
1768
1769static struct irqaction cmci_irqaction = {
1770	.handler =	ia64_mca_cmc_int_handler,
1771	.name =		"cmc_hndlr"
1772};
1773
1774static struct irqaction cmcp_irqaction = {
1775	.handler =	ia64_mca_cmc_int_caller,
1776	.name =		"cmc_poll"
1777};
1778
1779static struct irqaction mca_rdzv_irqaction = {
1780	.handler =	ia64_mca_rendez_int_handler,
1781	.name =		"mca_rdzv"
1782};
1783
1784static struct irqaction mca_wkup_irqaction = {
1785	.handler =	ia64_mca_wakeup_int_handler,
1786	.name =		"mca_wkup"
1787};
1788
1789static struct irqaction mca_cpe_irqaction = {
1790	.handler =	ia64_mca_cpe_int_handler,
1791	.name =		"cpe_hndlr"
1792};
1793
1794static struct irqaction mca_cpep_irqaction = {
1795	.handler =	ia64_mca_cpe_int_caller,
1796	.name =		"cpe_poll"
1797};
1798
1799/* Minimal format of the MCA/INIT stacks.  The pseudo processes that run on
1800 * these stacks can never sleep, they cannot return from the kernel to user
1801 * space, they do not appear in a normal ps listing.  So there is no need to
1802 * format most of the fields.
1803 */
1804
1805static void
1806format_mca_init_stack(void *mca_data, unsigned long offset,
1807		const char *type, int cpu)
1808{
1809	struct task_struct *p = (struct task_struct *)((char *)mca_data + offset);
1810	struct thread_info *ti;
1811	memset(p, 0, KERNEL_STACK_SIZE);
1812	ti = task_thread_info(p);
1813	ti->flags = _TIF_MCA_INIT;
1814	ti->preempt_count = 1;
1815	ti->task = p;
1816	ti->cpu = cpu;
1817	p->stack = ti;
1818	p->state = TASK_UNINTERRUPTIBLE;
1819	cpumask_set_cpu(cpu, &p->cpus_mask);
1820	INIT_LIST_HEAD(&p->tasks);
1821	p->parent = p->real_parent = p->group_leader = p;
1822	INIT_LIST_HEAD(&p->children);
1823	INIT_LIST_HEAD(&p->sibling);
1824	strncpy(p->comm, type, sizeof(p->comm)-1);
1825}
1826
1827/* Caller prevents this from being called after init */
1828static void * __ref mca_bootmem(void)
1829{
1830	return memblock_alloc(sizeof(struct ia64_mca_cpu), KERNEL_STACK_SIZE);
1831}
1832
1833/* Do per-CPU MCA-related initialization.  */
1834void
1835ia64_mca_cpu_init(void *cpu_data)
1836{
1837	void *pal_vaddr;
1838	void *data;
1839	long sz = sizeof(struct ia64_mca_cpu);
1840	int cpu = smp_processor_id();
1841	static int first_time = 1;
1842
1843	/*
1844	 * Structure will already be allocated if cpu has been online,
1845	 * then offlined.
1846	 */
1847	if (__per_cpu_mca[cpu]) {
1848		data = __va(__per_cpu_mca[cpu]);
1849	} else {
1850		if (first_time) {
1851			data = mca_bootmem();
1852			first_time = 0;
1853		} else
1854			data = (void *)__get_free_pages(GFP_KERNEL,
1855							get_order(sz));
1856		if (!data)
1857			panic("Could not allocate MCA memory for cpu %d\n",
1858					cpu);
1859	}
1860	format_mca_init_stack(data, offsetof(struct ia64_mca_cpu, mca_stack),
1861		"MCA", cpu);
1862	format_mca_init_stack(data, offsetof(struct ia64_mca_cpu, init_stack),
1863		"INIT", cpu);
1864	__this_cpu_write(ia64_mca_data, (__per_cpu_mca[cpu] = __pa(data)));
1865
1866	/*
1867	 * Stash away a copy of the PTE needed to map the per-CPU page.
1868	 * We may need it during MCA recovery.
1869	 */
1870	__this_cpu_write(ia64_mca_per_cpu_pte,
1871		pte_val(mk_pte_phys(__pa(cpu_data), PAGE_KERNEL)));
1872
1873	/*
1874	 * Also, stash away a copy of the PAL address and the PTE
1875	 * needed to map it.
1876	 */
1877	pal_vaddr = efi_get_pal_addr();
1878	if (!pal_vaddr)
1879		return;
1880	__this_cpu_write(ia64_mca_pal_base,
1881		GRANULEROUNDDOWN((unsigned long) pal_vaddr));
1882	__this_cpu_write(ia64_mca_pal_pte, pte_val(mk_pte_phys(__pa(pal_vaddr),
1883							      PAGE_KERNEL)));
1884}
1885
1886static int ia64_mca_cpu_online(unsigned int cpu)
1887{
1888	unsigned long flags;
1889
1890	local_irq_save(flags);
1891	if (!cmc_polling_enabled)
1892		ia64_mca_cmc_vector_enable(NULL);
1893	local_irq_restore(flags);
1894	return 0;
1895}
1896
1897/*
1898 * ia64_mca_init
1899 *
1900 *  Do all the system level mca specific initialization.
1901 *
1902 *	1. Register spinloop and wakeup request interrupt vectors
1903 *
1904 *	2. Register OS_MCA handler entry point
1905 *
1906 *	3. Register OS_INIT handler entry point
1907 *
1908 *  4. Initialize MCA/CMC/INIT related log buffers maintained by the OS.
1909 *
1910 *  Note that this initialization is done very early before some kernel
1911 *  services are available.
1912 *
1913 *  Inputs  :   None
1914 *
1915 *  Outputs :   None
1916 */
1917void __init
1918ia64_mca_init(void)
1919{
1920	ia64_fptr_t *init_hldlr_ptr_monarch = (ia64_fptr_t *)ia64_os_init_dispatch_monarch;
1921	ia64_fptr_t *init_hldlr_ptr_slave = (ia64_fptr_t *)ia64_os_init_dispatch_slave;
1922	ia64_fptr_t *mca_hldlr_ptr = (ia64_fptr_t *)ia64_os_mca_dispatch;
1923	int i;
1924	long rc;
1925	struct ia64_sal_retval isrv;
1926	unsigned long timeout = IA64_MCA_RENDEZ_TIMEOUT; /* platform specific */
1927	static struct notifier_block default_init_monarch_nb = {
1928		.notifier_call = default_monarch_init_process,
1929		.priority = 0/* we need to notified last */
1930	};
1931
1932	IA64_MCA_DEBUG("%s: begin\n", __func__);
1933
1934	/* Clear the Rendez checkin flag for all cpus */
1935	for(i = 0 ; i < NR_CPUS; i++)
1936		ia64_mc_info.imi_rendez_checkin[i] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1937
1938	/*
1939	 * Register the rendezvous spinloop and wakeup mechanism with SAL
1940	 */
1941
1942	/* Register the rendezvous interrupt vector with SAL */
1943	while (1) {
1944		isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_RENDEZ_INT,
1945					      SAL_MC_PARAM_MECHANISM_INT,
1946					      IA64_MCA_RENDEZ_VECTOR,
1947					      timeout,
1948					      SAL_MC_PARAM_RZ_ALWAYS);
1949		rc = isrv.status;
1950		if (rc == 0)
1951			break;
1952		if (rc == -2) {
1953			printk(KERN_INFO "Increasing MCA rendezvous timeout from "
1954				"%ld to %ld milliseconds\n", timeout, isrv.v0);
1955			timeout = isrv.v0;
1956			NOTIFY_MCA(DIE_MCA_NEW_TIMEOUT, NULL, timeout, 0);
1957			continue;
1958		}
1959		printk(KERN_ERR "Failed to register rendezvous interrupt "
1960		       "with SAL (status %ld)\n", rc);
1961		return;
1962	}
1963
1964	/* Register the wakeup interrupt vector with SAL */
1965	isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_RENDEZ_WAKEUP,
1966				      SAL_MC_PARAM_MECHANISM_INT,
1967				      IA64_MCA_WAKEUP_VECTOR,
1968				      0, 0);
1969	rc = isrv.status;
1970	if (rc) {
1971		printk(KERN_ERR "Failed to register wakeup interrupt with SAL "
1972		       "(status %ld)\n", rc);
1973		return;
1974	}
1975
1976	IA64_MCA_DEBUG("%s: registered MCA rendezvous spinloop and wakeup mech.\n", __func__);
1977
1978	ia64_mc_info.imi_mca_handler        = ia64_tpa(mca_hldlr_ptr->fp);
1979	/*
1980	 * XXX - disable SAL checksum by setting size to 0; should be
1981	 *	ia64_tpa(ia64_os_mca_dispatch_end) - ia64_tpa(ia64_os_mca_dispatch);
1982	 */
1983	ia64_mc_info.imi_mca_handler_size	= 0;
1984
1985	/* Register the os mca handler with SAL */
1986	if ((rc = ia64_sal_set_vectors(SAL_VECTOR_OS_MCA,
1987				       ia64_mc_info.imi_mca_handler,
1988				       ia64_tpa(mca_hldlr_ptr->gp),
1989				       ia64_mc_info.imi_mca_handler_size,
1990				       0, 0, 0)))
1991	{
1992		printk(KERN_ERR "Failed to register OS MCA handler with SAL "
1993		       "(status %ld)\n", rc);
1994		return;
1995	}
1996
1997	IA64_MCA_DEBUG("%s: registered OS MCA handler with SAL at 0x%lx, gp = 0x%lx\n", __func__,
1998		       ia64_mc_info.imi_mca_handler, ia64_tpa(mca_hldlr_ptr->gp));
1999
2000	/*
2001	 * XXX - disable SAL checksum by setting size to 0, should be
2002	 * size of the actual init handler in mca_asm.S.
2003	 */
2004	ia64_mc_info.imi_monarch_init_handler		= ia64_tpa(init_hldlr_ptr_monarch->fp);
2005	ia64_mc_info.imi_monarch_init_handler_size	= 0;
2006	ia64_mc_info.imi_slave_init_handler		= ia64_tpa(init_hldlr_ptr_slave->fp);
2007	ia64_mc_info.imi_slave_init_handler_size	= 0;
2008
2009	IA64_MCA_DEBUG("%s: OS INIT handler at %lx\n", __func__,
2010		       ia64_mc_info.imi_monarch_init_handler);
2011
2012	/* Register the os init handler with SAL */
2013	if ((rc = ia64_sal_set_vectors(SAL_VECTOR_OS_INIT,
2014				       ia64_mc_info.imi_monarch_init_handler,
2015				       ia64_tpa(ia64_getreg(_IA64_REG_GP)),
2016				       ia64_mc_info.imi_monarch_init_handler_size,
2017				       ia64_mc_info.imi_slave_init_handler,
2018				       ia64_tpa(ia64_getreg(_IA64_REG_GP)),
2019				       ia64_mc_info.imi_slave_init_handler_size)))
2020	{
2021		printk(KERN_ERR "Failed to register m/s INIT handlers with SAL "
2022		       "(status %ld)\n", rc);
2023		return;
2024	}
2025	if (register_die_notifier(&default_init_monarch_nb)) {
2026		printk(KERN_ERR "Failed to register default monarch INIT process\n");
2027		return;
2028	}
2029
2030	IA64_MCA_DEBUG("%s: registered OS INIT handler with SAL\n", __func__);
2031
2032	/* Initialize the areas set aside by the OS to buffer the
2033	 * platform/processor error states for MCA/INIT/CMC
2034	 * handling.
2035	 */
2036	ia64_log_init(SAL_INFO_TYPE_MCA);
2037	ia64_log_init(SAL_INFO_TYPE_INIT);
2038	ia64_log_init(SAL_INFO_TYPE_CMC);
2039	ia64_log_init(SAL_INFO_TYPE_CPE);
2040
2041	mca_init = 1;
2042	printk(KERN_INFO "MCA related initialization done\n");
2043}
2044
2045
2046/*
2047 * These pieces cannot be done in ia64_mca_init() because it is called before
2048 * early_irq_init() which would wipe out our percpu irq registrations. But we
2049 * cannot leave them until ia64_mca_late_init() because by then all the other
2050 * processors have been brought online and have set their own CMC vectors to
2051 * point at a non-existant action. Called from arch_early_irq_init().
2052 */
2053void __init ia64_mca_irq_init(void)
2054{
2055	/*
2056	 *  Configure the CMCI/P vector and handler. Interrupts for CMC are
2057	 *  per-processor, so AP CMC interrupts are setup in smp_callin() (smpboot.c).
2058	 */
2059	register_percpu_irq(IA64_CMC_VECTOR, &cmci_irqaction);
2060	register_percpu_irq(IA64_CMCP_VECTOR, &cmcp_irqaction);
 
 
2061	ia64_mca_cmc_vector_setup();       /* Setup vector on BSP */
2062
2063	/* Setup the MCA rendezvous interrupt vector */
2064	register_percpu_irq(IA64_MCA_RENDEZ_VECTOR, &mca_rdzv_irqaction);
 
2065
2066	/* Setup the MCA wakeup interrupt vector */
2067	register_percpu_irq(IA64_MCA_WAKEUP_VECTOR, &mca_wkup_irqaction);
 
2068
2069	/* Setup the CPEI/P handler */
2070	register_percpu_irq(IA64_CPEP_VECTOR, &mca_cpep_irqaction);
 
2071}
2072
2073/*
2074 * ia64_mca_late_init
2075 *
2076 *	Opportunity to setup things that require initialization later
2077 *	than ia64_mca_init.  Setup a timer to poll for CPEs if the
2078 *	platform doesn't support an interrupt driven mechanism.
2079 *
2080 *  Inputs  :   None
2081 *  Outputs :   Status
2082 */
2083static int __init
2084ia64_mca_late_init(void)
2085{
2086	if (!mca_init)
2087		return 0;
2088
2089	/* Setup the CMCI/P vector and handler */
2090	timer_setup(&cmc_poll_timer, ia64_mca_cmc_poll, 0);
2091
2092	/* Unmask/enable the vector */
2093	cmc_polling_enabled = 0;
2094	cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "ia64/mca:online",
2095			  ia64_mca_cpu_online, NULL);
2096	IA64_MCA_DEBUG("%s: CMCI/P setup and enabled.\n", __func__);
2097
2098	/* Setup the CPEI/P vector and handler */
2099	cpe_vector = acpi_request_vector(ACPI_INTERRUPT_CPEI);
2100	timer_setup(&cpe_poll_timer, ia64_mca_cpe_poll, 0);
2101
2102	{
2103		unsigned int irq;
2104
2105		if (cpe_vector >= 0) {
2106			/* If platform supports CPEI, enable the irq. */
2107			irq = local_vector_to_irq(cpe_vector);
2108			if (irq > 0) {
2109				cpe_poll_enabled = 0;
2110				irq_set_status_flags(irq, IRQ_PER_CPU);
2111				setup_irq(irq, &mca_cpe_irqaction);
 
 
2112				ia64_cpe_irq = irq;
2113				ia64_mca_register_cpev(cpe_vector);
2114				IA64_MCA_DEBUG("%s: CPEI/P setup and enabled.\n",
2115					__func__);
2116				return 0;
2117			}
2118			printk(KERN_ERR "%s: Failed to find irq for CPE "
2119					"interrupt handler, vector %d\n",
2120					__func__, cpe_vector);
2121		}
2122		/* If platform doesn't support CPEI, get the timer going. */
2123		if (cpe_poll_enabled) {
2124			ia64_mca_cpe_poll(0UL);
2125			IA64_MCA_DEBUG("%s: CPEP setup and enabled.\n", __func__);
2126		}
2127	}
2128
2129	return 0;
2130}
2131
2132device_initcall(ia64_mca_late_init);