arch/ia64/kernel/mca.c

   1 /*
   2  * File:        mca.c
   3  * Purpose:     Generic MCA handling layer
   4  *
   5  * Copyright (C) 2003 Hewlett-Packard Co
   6  *      David Mosberger-Tang <davidm@hpl.hp.com>
   7  *
   8  * Copyright (C) 2002 Dell Inc.
   9  * Copyright (C) Matt Domsch <Matt_Domsch@dell.com>
  10  *
  11  * Copyright (C) 2002 Intel
  12  * Copyright (C) Jenna Hall <jenna.s.hall@intel.com>
  13  *
  14  * Copyright (C) 2001 Intel
  15  * Copyright (C) Fred Lewis <frederick.v.lewis@intel.com>
  16  *
  17  * Copyright (C) 2000 Intel
  18  * Copyright (C) Chuck Fleckenstein <cfleck@co.intel.com>
  19  *
  20  * Copyright (C) 1999, 2004-2008 Silicon Graphics, Inc.
  21  * Copyright (C) Vijay Chander <vijay@engr.sgi.com>
  22  *
  23  * Copyright (C) 2006 FUJITSU LIMITED
  24  * Copyright (C) Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
  25  *
  26  * 2000-03-29 Chuck Fleckenstein <cfleck@co.intel.com>
  27  *            Fixed PAL/SAL update issues, began MCA bug fixes, logging issues,
  28  *            added min save state dump, added INIT handler.
  29  *
  30  * 2001-01-03 Fred Lewis <frederick.v.lewis@intel.com>
  31  *            Added setup of CMCI and CPEI IRQs, logging of corrected platform
  32  *            errors, completed code for logging of corrected & uncorrected
  33  *            machine check errors, and updated for conformance with Nov. 2000
  34  *            revision of the SAL 3.0 spec.
  35  *
  36  * 2002-01-04 Jenna Hall <jenna.s.hall@intel.com>
  37  *            Aligned MCA stack to 16 bytes, added platform vs. CPU error flag,
  38  *            set SAL default return values, changed error record structure to
  39  *            linked list, added init call to sal_get_state_info_size().
  40  *
  41  * 2002-03-25 Matt Domsch <Matt_Domsch@dell.com>
  42  *            GUID cleanups.
  43  *
  44  * 2003-04-15 David Mosberger-Tang <davidm@hpl.hp.com>
  45  *            Added INIT backtrace support.
  46  *
  47  * 2003-12-08 Keith Owens <kaos@sgi.com>
  48  *            smp_call_function() must not be called from interrupt context
  49  *            (can deadlock on tasklist_lock).
  50  *            Use keventd to call smp_call_function().
  51  *
  52  * 2004-02-01 Keith Owens <kaos@sgi.com>
  53  *            Avoid deadlock when using printk() for MCA and INIT records.
  54  *            Delete all record printing code, moved to salinfo_decode in user
  55  *            space.  Mark variables and functions static where possible.
  56  *            Delete dead variables and functions.  Reorder to remove the need
  57  *            for forward declarations and to consolidate related code.
  58  *
  59  * 2005-08-12 Keith Owens <kaos@sgi.com>
  60  *            Convert MCA/INIT handlers to use per event stacks and SAL/OS
  61  *            state.
  62  *
  63  * 2005-10-07 Keith Owens <kaos@sgi.com>
  64  *            Add notify_die() hooks.
  65  *
  66  * 2006-09-15 Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
  67  *            Add printing support for MCA/INIT.
  68  *
  69  * 2007-04-27 Russ Anderson <rja@sgi.com>
  70  *            Support multiple cpus going through OS_MCA in the same event.
  71  */
  72 #include <linux/jiffies.h>
  73 #include <linux/types.h>
  74 #include <linux/init.h>
  75 #include <linux/sched.h>
  76 #include <linux/interrupt.h>
  77 #include <linux/irq.h>
  78 #include <linux/bootmem.h>
  79 #include <linux/acpi.h>
  80 #include <linux/timer.h>
  81 #include <linux/module.h>
  82 #include <linux/kernel.h>
  83 #include <linux/smp.h>
  84 #include <linux/workqueue.h>
  85 #include <linux/cpumask.h>
  86 #include <linux/kdebug.h>
  87 #include <linux/cpu.h>
  88 #include <linux/gfp.h>
  89
  90 #include <asm/delay.h>
  91 #include <asm/machvec.h>
  92 #include <asm/meminit.h>
  93 #include <asm/page.h>
  94 #include <asm/ptrace.h>
  95 #include <asm/system.h>
  96 #include <asm/sal.h>
  97 #include <asm/mca.h>
  98 #include <asm/kexec.h>
  99
 100 #include <asm/irq.h>
 101 #include <asm/hw_irq.h>
 102 #include <asm/tlb.h>
 103
 104 #include "mca_drv.h"
 105 #include "entry.h"
 106
 107 #if defined(IA64_MCA_DEBUG_INFO)
 108 # define IA64_MCA_DEBUG(fmt...) printk(fmt)
 109 #else
 110 # define IA64_MCA_DEBUG(fmt...)
 111 #endif
 112
 113 #define NOTIFY_INIT(event, regs, arg, spin)                             \
 114 do {                                                                    \
 115         if ((notify_die((event), "INIT", (regs), (arg), 0, 0)           \
 116                         == NOTIFY_STOP) && ((spin) == 1))               \
 117                 ia64_mca_spin(__func__);                                \
 118 } while (0)
 119
 120 #define NOTIFY_MCA(event, regs, arg, spin)                              \
 121 do {                                                                    \
 122         if ((notify_die((event), "MCA", (regs), (arg), 0, 0)            \
 123                         == NOTIFY_STOP) && ((spin) == 1))               \
 124                 ia64_mca_spin(__func__);                                \
 125 } while (0)
 126
 127 /* Used by mca_asm.S */
 128 DEFINE_PER_CPU(u64, ia64_mca_data); /* == __per_cpu_mca[smp_processor_id()] */
 129 DEFINE_PER_CPU(u64, ia64_mca_per_cpu_pte); /* PTE to map per-CPU area */
 130 DEFINE_PER_CPU(u64, ia64_mca_pal_pte);      /* PTE to map PAL code */
 131 DEFINE_PER_CPU(u64, ia64_mca_pal_base);    /* vaddr PAL code granule */
 132 DEFINE_PER_CPU(u64, ia64_mca_tr_reload);   /* Flag for TR reload */
 133
 134 unsigned long __per_cpu_mca[NR_CPUS];
 135
 136 /* In mca_asm.S */
 137 extern void                     ia64_os_init_dispatch_monarch (void);
 138 extern void                     ia64_os_init_dispatch_slave (void);
 139
 140 static int monarch_cpu = -1;
 141
 142 static ia64_mc_info_t           ia64_mc_info;
 143
 144 #define MAX_CPE_POLL_INTERVAL (15*60*HZ) /* 15 minutes */
 145 #define MIN_CPE_POLL_INTERVAL (2*60*HZ)  /* 2 minutes */
 146 #define CMC_POLL_INTERVAL     (1*60*HZ)  /* 1 minute */
 147 #define CPE_HISTORY_LENGTH    5
 148 #define CMC_HISTORY_LENGTH    5
 149
 150 #ifdef CONFIG_ACPI
 151 static struct timer_list cpe_poll_timer;
 152 #endif
 153 static struct timer_list cmc_poll_timer;
 154 /*
 155  * This variable tells whether we are currently in polling mode.
 156  * Start with this in the wrong state so we won't play w/ timers
 157  * before the system is ready.
 158  */
 159 static int cmc_polling_enabled = 1;
 160
 161 /*
 162  * Clearing this variable prevents CPE polling from getting activated
 163  * in mca_late_init.  Use it if your system doesn't provide a CPEI,
 164  * but encounters problems retrieving CPE logs.  This should only be
 165  * necessary for debugging.
 166  */
 167 static int cpe_poll_enabled = 1;
 168
 169 extern void salinfo_log_wakeup(int type, u8 *buffer, u64 size, int irqsafe);
 170
 171 static int mca_init __initdata;
 172
 173 /*
 174  * limited & delayed printing support for MCA/INIT handler
 175  */
 176
 177 #define mprintk(fmt...) ia64_mca_printk(fmt)
 178
 179 #define MLOGBUF_SIZE (512+256*NR_CPUS)
 180 #define MLOGBUF_MSGMAX 256
 181 static char mlogbuf[MLOGBUF_SIZE];
 182 static DEFINE_SPINLOCK(mlogbuf_wlock);  /* mca context only */
 183 static DEFINE_SPINLOCK(mlogbuf_rlock);  /* normal context only */
 184 static unsigned long mlogbuf_start;
 185 static unsigned long mlogbuf_end;
 186 static unsigned int mlogbuf_finished = 0;
 187 static unsigned long mlogbuf_timestamp = 0;
 188
 189 static int loglevel_save = -1;
 190 #define BREAK_LOGLEVEL(__console_loglevel)              \
 191         oops_in_progress = 1;                           \
 192         if (loglevel_save < 0)                          \
 193                 loglevel_save = __console_loglevel;     \
 194         __console_loglevel = 15;
 195
 196 #define RESTORE_LOGLEVEL(__console_loglevel)            \
 197         if (loglevel_save >= 0) {                       \
 198                 __console_loglevel = loglevel_save;     \
 199                 loglevel_save = -1;                     \
 200         }                                               \
 201         mlogbuf_finished = 0;                           \
 202         oops_in_progress = 0;
 203
 204 /*
 205  * Push messages into buffer, print them later if not urgent.
 206  */
 207 void ia64_mca_printk(const char *fmt, ...)
 208 {
 209         va_list args;
 210         int printed_len;
 211         char temp_buf[MLOGBUF_MSGMAX];
 212         char *p;
 213
 214         va_start(args, fmt);
 215         printed_len = vscnprintf(temp_buf, sizeof(temp_buf), fmt, args);
 216         va_end(args);
 217
 218         /* Copy the output into mlogbuf */
 219         if (oops_in_progress) {
 220                 /* mlogbuf was abandoned, use printk directly instead. */
 221                 printk(temp_buf);
 222         } else {
 223                 spin_lock(&mlogbuf_wlock);
 224                 for (p = temp_buf; *p; p++) {
 225                         unsigned long next = (mlogbuf_end + 1) % MLOGBUF_SIZE;
 226                         if (next != mlogbuf_start) {
 227                                 mlogbuf[mlogbuf_end] = *p;
 228                                 mlogbuf_end = next;
 229                         } else {
 230                                 /* buffer full */
 231                                 break;
 232                         }
 233                 }
 234                 mlogbuf[mlogbuf_end] = '\0';
 235                 spin_unlock(&mlogbuf_wlock);
 236         }
 237 }
 238 EXPORT_SYMBOL(ia64_mca_printk);
 239
 240 /*
 241  * Print buffered messages.
 242  *  NOTE: call this after returning normal context. (ex. from salinfod)
 243  */
 244 void ia64_mlogbuf_dump(void)
 245 {
 246         char temp_buf[MLOGBUF_MSGMAX];
 247         char *p;
 248         unsigned long index;
 249         unsigned long flags;
 250         unsigned int printed_len;
 251
 252         /* Get output from mlogbuf */
 253         while (mlogbuf_start != mlogbuf_end) {
 254                 temp_buf[0] = '\0';
 255                 p = temp_buf;
 256                 printed_len = 0;
 257
 258                 spin_lock_irqsave(&mlogbuf_rlock, flags);
 259
 260                 index = mlogbuf_start;
 261                 while (index != mlogbuf_end) {
 262                         *p = mlogbuf[index];
 263                         index = (index + 1) % MLOGBUF_SIZE;
 264                         if (!*p)
 265                                 break;
 266                         p++;
 267                         if (++printed_len >= MLOGBUF_MSGMAX - 1)
 268                                 break;
 269                 }
 270                 *p = '\0';
 271                 if (temp_buf[0])
 272                         printk(temp_buf);
 273                 mlogbuf_start = index;
 274
 275                 mlogbuf_timestamp = 0;
 276                 spin_unlock_irqrestore(&mlogbuf_rlock, flags);
 277         }
 278 }
 279 EXPORT_SYMBOL(ia64_mlogbuf_dump);
 280
 281 /*
 282  * Call this if system is going to down or if immediate flushing messages to
 283  * console is required. (ex. recovery was failed, crash dump is going to be
 284  * invoked, long-wait rendezvous etc.)
 285  *  NOTE: this should be called from monarch.
 286  */
 287 static void ia64_mlogbuf_finish(int wait)
 288 {
 289         BREAK_LOGLEVEL(console_loglevel);
 290
 291         spin_lock_init(&mlogbuf_rlock);
 292         ia64_mlogbuf_dump();
 293         printk(KERN_EMERG "mlogbuf_finish: printing switched to urgent mode, "
 294                 "MCA/INIT might be dodgy or fail.\n");
 295
 296         if (!wait)
 297                 return;
 298
 299         /* wait for console */
 300         printk("Delaying for 5 seconds...\n");
 301         udelay(5*1000000);
 302
 303         mlogbuf_finished = 1;
 304 }
 305
 306 /*
 307  * Print buffered messages from INIT context.
 308  */
 309 static void ia64_mlogbuf_dump_from_init(void)
 310 {
 311         if (mlogbuf_finished)
 312                 return;
 313
 314         if (mlogbuf_timestamp &&
 315                         time_before(jiffies, mlogbuf_timestamp + 30 * HZ)) {
 316                 printk(KERN_ERR "INIT: mlogbuf_dump is interrupted by INIT "
 317                         " and the system seems to be messed up.\n");
 318                 ia64_mlogbuf_finish(0);
 319                 return;
 320         }
 321
 322         if (!spin_trylock(&mlogbuf_rlock)) {
 323                 printk(KERN_ERR "INIT: mlogbuf_dump is interrupted by INIT. "
 324                         "Generated messages other than stack dump will be "
 325                         "buffered to mlogbuf and will be printed later.\n");
 326                 printk(KERN_ERR "INIT: If messages would not printed after "
 327                         "this INIT, wait 30sec and assert INIT again.\n");
 328                 if (!mlogbuf_timestamp)
 329                         mlogbuf_timestamp = jiffies;
 330                 return;
 331         }
 332         spin_unlock(&mlogbuf_rlock);
 333         ia64_mlogbuf_dump();
 334 }
 335
 336 static void inline
 337 ia64_mca_spin(const char *func)
 338 {
 339         if (monarch_cpu == smp_processor_id())
 340                 ia64_mlogbuf_finish(0);
 341         mprintk(KERN_EMERG "%s: spinning here, not returning to SAL\n", func);
 342         while (1)
 343                 cpu_relax();
 344 }
 345 /*
 346  * IA64_MCA log support
 347  */
 348 #define IA64_MAX_LOGS           2       /* Double-buffering for nested MCAs */
 349 #define IA64_MAX_LOG_TYPES      4   /* MCA, INIT, CMC, CPE */
 350
 351 typedef struct ia64_state_log_s
 352 {
 353         spinlock_t      isl_lock;
 354         int             isl_index;
 355         unsigned long   isl_count;
 356         ia64_err_rec_t  *isl_log[IA64_MAX_LOGS]; /* need space to store header + error log */
 357 } ia64_state_log_t;
 358
 359 static ia64_state_log_t ia64_state_log[IA64_MAX_LOG_TYPES];
 360
 361 #define IA64_LOG_ALLOCATE(it, size) \
 362         {ia64_state_log[it].isl_log[IA64_LOG_CURR_INDEX(it)] = \
 363                 (ia64_err_rec_t *)alloc_bootmem(size); \
 364         ia64_state_log[it].isl_log[IA64_LOG_NEXT_INDEX(it)] = \
 365                 (ia64_err_rec_t *)alloc_bootmem(size);}
 366 #define IA64_LOG_LOCK_INIT(it) spin_lock_init(&ia64_state_log[it].isl_lock)
 367 #define IA64_LOG_LOCK(it)      spin_lock_irqsave(&ia64_state_log[it].isl_lock, s)
 368 #define IA64_LOG_UNLOCK(it)    spin_unlock_irqrestore(&ia64_state_log[it].isl_lock,s)
 369 #define IA64_LOG_NEXT_INDEX(it)    ia64_state_log[it].isl_index
 370 #define IA64_LOG_CURR_INDEX(it)    1 - ia64_state_log[it].isl_index
 371 #define IA64_LOG_INDEX_INC(it) \
 372     {ia64_state_log[it].isl_index = 1 - ia64_state_log[it].isl_index; \
 373     ia64_state_log[it].isl_count++;}
 374 #define IA64_LOG_INDEX_DEC(it) \
 375     ia64_state_log[it].isl_index = 1 - ia64_state_log[it].isl_index
 376 #define IA64_LOG_NEXT_BUFFER(it)   (void *)((ia64_state_log[it].isl_log[IA64_LOG_NEXT_INDEX(it)]))
 377 #define IA64_LOG_CURR_BUFFER(it)   (void *)((ia64_state_log[it].isl_log[IA64_LOG_CURR_INDEX(it)]))
 378 #define IA64_LOG_COUNT(it)         ia64_state_log[it].isl_count
 379
 380 /*
 381  * ia64_log_init
 382  *      Reset the OS ia64 log buffer
 383  * Inputs   :   info_type   (SAL_INFO_TYPE_{MCA,INIT,CMC,CPE})
 384  * Outputs      :       None
 385  */
 386 static void __init
 387 ia64_log_init(int sal_info_type)
 388 {
 389         u64     max_size = 0;
 390
 391         IA64_LOG_NEXT_INDEX(sal_info_type) = 0;
 392         IA64_LOG_LOCK_INIT(sal_info_type);
 393
 394         // SAL will tell us the maximum size of any error record of this type
 395         max_size = ia64_sal_get_state_info_size(sal_info_type);
 396         if (!max_size)
 397                 /* alloc_bootmem() doesn't like zero-sized allocations! */
 398                 return;
 399
 400         // set up OS data structures to hold error info
 401         IA64_LOG_ALLOCATE(sal_info_type, max_size);
 402         memset(IA64_LOG_CURR_BUFFER(sal_info_type), 0, max_size);
 403         memset(IA64_LOG_NEXT_BUFFER(sal_info_type), 0, max_size);
 404 }
 405
 406 /*
 407  * ia64_log_get
 408  *
 409  *      Get the current MCA log from SAL and copy it into the OS log buffer.
 410  *
 411  *  Inputs  :   info_type   (SAL_INFO_TYPE_{MCA,INIT,CMC,CPE})
 412  *              irq_safe    whether you can use printk at this point
 413  *  Outputs :   size        (total record length)
 414  *              *buffer     (ptr to error record)
 415  *
 416  */
 417 static u64
 418 ia64_log_get(int sal_info_type, u8 **buffer, int irq_safe)
 419 {
 420         sal_log_record_header_t     *log_buffer;
 421         u64                         total_len = 0;
 422         unsigned long               s;
 423
 424         IA64_LOG_LOCK(sal_info_type);
 425
 426         /* Get the process state information */
 427         log_buffer = IA64_LOG_NEXT_BUFFER(sal_info_type);
 428
 429         total_len = ia64_sal_get_state_info(sal_info_type, (u64 *)log_buffer);
 430
 431         if (total_len) {
 432                 IA64_LOG_INDEX_INC(sal_info_type);
 433                 IA64_LOG_UNLOCK(sal_info_type);
 434                 if (irq_safe) {
 435                         IA64_MCA_DEBUG("%s: SAL error record type %d retrieved. Record length = %ld\n",
 436                                        __func__, sal_info_type, total_len);
 437                 }
 438                 *buffer = (u8 *) log_buffer;
 439                 return total_len;
 440         } else {
 441                 IA64_LOG_UNLOCK(sal_info_type);
 442                 return 0;
 443         }
 444 }
 445
 446 /*
 447  *  ia64_mca_log_sal_error_record
 448  *
 449  *  This function retrieves a specified error record type from SAL
 450  *  and wakes up any processes waiting for error records.
 451  *
 452  *  Inputs  :   sal_info_type   (Type of error record MCA/CMC/CPE)
 453  *              FIXME: remove MCA and irq_safe.
 454  */
 455 static void
 456 ia64_mca_log_sal_error_record(int sal_info_type)
 457 {
 458         u8 *buffer;
 459         sal_log_record_header_t *rh;
 460         u64 size;
 461         int irq_safe = sal_info_type != SAL_INFO_TYPE_MCA;
 462 #ifdef IA64_MCA_DEBUG_INFO
 463         static const char * const rec_name[] = { "MCA", "INIT", "CMC", "CPE" };
 464 #endif
 465
 466         size = ia64_log_get(sal_info_type, &buffer, irq_safe);
 467         if (!size)
 468                 return;
 469
 470         salinfo_log_wakeup(sal_info_type, buffer, size, irq_safe);
 471
 472         if (irq_safe)
 473                 IA64_MCA_DEBUG("CPU %d: SAL log contains %s error record\n",
 474                         smp_processor_id(),
 475                         sal_info_type < ARRAY_SIZE(rec_name) ? rec_name[sal_info_type] : "UNKNOWN");
 476
 477         /* Clear logs from corrected errors in case there's no user-level logger */
 478         rh = (sal_log_record_header_t *)buffer;
 479         if (rh->severity == sal_log_severity_corrected)
 480                 ia64_sal_clear_state_info(sal_info_type);
 481 }
 482
 483 /*
 484  * search_mca_table
 485  *  See if the MCA surfaced in an instruction range
 486  *  that has been tagged as recoverable.
 487  *
 488  *  Inputs
 489  *      first   First address range to check
 490  *      last    Last address range to check
 491  *      ip      Instruction pointer, address we are looking for
 492  *
 493  * Return value:
 494  *      1 on Success (in the table)/ 0 on Failure (not in the  table)
 495  */
 496 int
 497 search_mca_table (const struct mca_table_entry *first,
 498                 const struct mca_table_entry *last,
 499                 unsigned long ip)
 500 {
 501         const struct mca_table_entry *curr;
 502         u64 curr_start, curr_end;
 503
 504         curr = first;
 505         while (curr <= last) {
 506                 curr_start = (u64) &curr->start_addr + curr->start_addr;
 507                 curr_end = (u64) &curr->end_addr + curr->end_addr;
 508
 509                 if ((ip >= curr_start) && (ip <= curr_end)) {
 510                         return 1;
 511                 }
 512                 curr++;
 513         }
 514         return 0;
 515 }
 516
 517 /* Given an address, look for it in the mca tables. */
 518 int mca_recover_range(unsigned long addr)
 519 {
 520         extern struct mca_table_entry __start___mca_table[];
 521         extern struct mca_table_entry __stop___mca_table[];
 522
 523         return search_mca_table(__start___mca_table, __stop___mca_table-1, addr);
 524 }
 525 EXPORT_SYMBOL_GPL(mca_recover_range);
 526
 527 #ifdef CONFIG_ACPI
 528
 529 int cpe_vector = -1;
 530 int ia64_cpe_irq = -1;
 531
 532 static irqreturn_t
 533 ia64_mca_cpe_int_handler (int cpe_irq, void *arg)
 534 {
 535         static unsigned long    cpe_history[CPE_HISTORY_LENGTH];
 536         static int              index;
 537         static DEFINE_SPINLOCK(cpe_history_lock);
 538
 539         IA64_MCA_DEBUG("%s: received interrupt vector = %#x on CPU %d\n",
 540                        __func__, cpe_irq, smp_processor_id());
 541
 542         /* SAL spec states this should run w/ interrupts enabled */
 543         local_irq_enable();
 544
 545         spin_lock(&cpe_history_lock);
 546         if (!cpe_poll_enabled && cpe_vector >= 0) {
 547
 548                 int i, count = 1; /* we know 1 happened now */
 549                 unsigned long now = jiffies;
 550
 551                 for (i = 0; i < CPE_HISTORY_LENGTH; i++) {
 552                         if (now - cpe_history[i] <= HZ)
 553                                 count++;
 554                 }
 555
 556                 IA64_MCA_DEBUG(KERN_INFO "CPE threshold %d/%d\n", count, CPE_HISTORY_LENGTH);
 557                 if (count >= CPE_HISTORY_LENGTH) {
 558
 559                         cpe_poll_enabled = 1;
 560                         spin_unlock(&cpe_history_lock);
 561                         disable_irq_nosync(local_vector_to_irq(IA64_CPE_VECTOR));
 562
 563                         /*
 564                          * Corrected errors will still be corrected, but
 565                          * make sure there's a log somewhere that indicates
 566                          * something is generating more than we can handle.
 567                          */
 568                         printk(KERN_WARNING "WARNING: Switching to polling CPE handler; error records may be lost\n");
 569
 570                         mod_timer(&cpe_poll_timer, jiffies + MIN_CPE_POLL_INTERVAL);
 571
 572                         /* lock already released, get out now */
 573                         goto out;
 574                 } else {
 575                         cpe_history[index++] = now;
 576                         if (index == CPE_HISTORY_LENGTH)
 577                                 index = 0;
 578                 }
 579         }
 580         spin_unlock(&cpe_history_lock);
 581 out:
 582         /* Get the CPE error record and log it */
 583         ia64_mca_log_sal_error_record(SAL_INFO_TYPE_CPE);
 584
 585         return IRQ_HANDLED;
 586 }
 587
 588 #endif /* CONFIG_ACPI */
 589
 590 #ifdef CONFIG_ACPI
 591 /*
 592  * ia64_mca_register_cpev
 593  *
 594  *  Register the corrected platform error vector with SAL.
 595  *
 596  *  Inputs
 597  *      cpev        Corrected Platform Error Vector number
 598  *
 599  *  Outputs
 600  *      None
 601  */
 602 void
 603 ia64_mca_register_cpev (int cpev)
 604 {
 605         /* Register the CPE interrupt vector with SAL */
 606         struct ia64_sal_retval isrv;
 607
 608         isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_CPE_INT, SAL_MC_PARAM_MECHANISM_INT, cpev, 0, 0);
 609         if (isrv.status) {
 610                 printk(KERN_ERR "Failed to register Corrected Platform "
 611                        "Error interrupt vector with SAL (status %ld)\n", isrv.status);
 612                 return;
 613         }
 614
 615         IA64_MCA_DEBUG("%s: corrected platform error "
 616                        "vector %#x registered\n", __func__, cpev);
 617 }
 618 #endif /* CONFIG_ACPI */
 619
 620 /*
 621  * ia64_mca_cmc_vector_setup
 622  *
 623  *  Setup the corrected machine check vector register in the processor.
 624  *  (The interrupt is masked on boot. ia64_mca_late_init unmask this.)
 625  *  This function is invoked on a per-processor basis.
 626  *
 627  * Inputs
 628  *      None
 629  *
 630  * Outputs
 631  *      None
 632  */
 633 void __cpuinit
 634 ia64_mca_cmc_vector_setup (void)
 635 {
 636         cmcv_reg_t      cmcv;
 637
 638         cmcv.cmcv_regval        = 0;
 639         cmcv.cmcv_mask          = 1;        /* Mask/disable interrupt at first */
 640         cmcv.cmcv_vector        = IA64_CMC_VECTOR;
 641         ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
 642
 643         IA64_MCA_DEBUG("%s: CPU %d corrected machine check vector %#x registered.\n",
 644                        __func__, smp_processor_id(), IA64_CMC_VECTOR);
 645
 646         IA64_MCA_DEBUG("%s: CPU %d CMCV = %#016lx\n",
 647                        __func__, smp_processor_id(), ia64_getreg(_IA64_REG_CR_CMCV));
 648 }
 649
 650 /*
 651  * ia64_mca_cmc_vector_disable
 652  *
 653  *  Mask the corrected machine check vector register in the processor.
 654  *  This function is invoked on a per-processor basis.
 655  *
 656  * Inputs
 657  *      dummy(unused)
 658  *
 659  * Outputs
 660  *      None
 661  */
 662 static void
 663 ia64_mca_cmc_vector_disable (void *dummy)
 664 {
 665         cmcv_reg_t      cmcv;
 666
 667         cmcv.cmcv_regval = ia64_getreg(_IA64_REG_CR_CMCV);
 668
 669         cmcv.cmcv_mask = 1; /* Mask/disable interrupt */
 670         ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
 671
 672         IA64_MCA_DEBUG("%s: CPU %d corrected machine check vector %#x disabled.\n",
 673                        __func__, smp_processor_id(), cmcv.cmcv_vector);
 674 }
 675
 676 /*
 677  * ia64_mca_cmc_vector_enable
 678  *
 679  *  Unmask the corrected machine check vector register in the processor.
 680  *  This function is invoked on a per-processor basis.
 681  *
 682  * Inputs
 683  *      dummy(unused)
 684  *
 685  * Outputs
 686  *      None
 687  */
 688 static void
 689 ia64_mca_cmc_vector_enable (void *dummy)
 690 {
 691         cmcv_reg_t      cmcv;
 692
 693         cmcv.cmcv_regval = ia64_getreg(_IA64_REG_CR_CMCV);
 694
 695         cmcv.cmcv_mask = 0; /* Unmask/enable interrupt */
 696         ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
 697
 698         IA64_MCA_DEBUG("%s: CPU %d corrected machine check vector %#x enabled.\n",
 699                        __func__, smp_processor_id(), cmcv.cmcv_vector);
 700 }
 701
 702 /*
 703  * ia64_mca_cmc_vector_disable_keventd
 704  *
 705  * Called via keventd (smp_call_function() is not safe in interrupt context) to
 706  * disable the cmc interrupt vector.
 707  */
 708 static void
 709 ia64_mca_cmc_vector_disable_keventd(struct work_struct *unused)
 710 {
 711         on_each_cpu(ia64_mca_cmc_vector_disable, NULL, 0);
 712 }
 713
 714 /*
 715  * ia64_mca_cmc_vector_enable_keventd
 716  *
 717  * Called via keventd (smp_call_function() is not safe in interrupt context) to
 718  * enable the cmc interrupt vector.
 719  */
 720 static void
 721 ia64_mca_cmc_vector_enable_keventd(struct work_struct *unused)
 722 {
 723         on_each_cpu(ia64_mca_cmc_vector_enable, NULL, 0);
 724 }
 725
 726 /*
 727  * ia64_mca_wakeup
 728  *
 729  *      Send an inter-cpu interrupt to wake-up a particular cpu.
 730  *
 731  *  Inputs  :   cpuid
 732  *  Outputs :   None
 733  */
 734 static void
 735 ia64_mca_wakeup(int cpu)
 736 {
 737         platform_send_ipi(cpu, IA64_MCA_WAKEUP_VECTOR, IA64_IPI_DM_INT, 0);
 738 }
 739
 740 /*
 741  * ia64_mca_wakeup_all
 742  *
 743  *      Wakeup all the slave cpus which have rendez'ed previously.
 744  *
 745  *  Inputs  :   None
 746  *  Outputs :   None
 747  */
 748 static void
 749 ia64_mca_wakeup_all(void)
 750 {
 751         int cpu;
 752
 753         /* Clear the Rendez checkin flag for all cpus */
 754         for_each_online_cpu(cpu) {
 755                 if (ia64_mc_info.imi_rendez_checkin[cpu] == IA64_MCA_RENDEZ_CHECKIN_DONE)
 756                         ia64_mca_wakeup(cpu);
 757         }
 758
 759 }
 760
 761 /*
 762  * ia64_mca_rendez_interrupt_handler
 763  *
 764  *      This is handler used to put slave processors into spinloop
 765  *      while the monarch processor does the mca handling and later
 766  *      wake each slave up once the monarch is done.  The state
 767  *      IA64_MCA_RENDEZ_CHECKIN_DONE indicates the cpu is rendez'ed
 768  *      in SAL.  The state IA64_MCA_RENDEZ_CHECKIN_NOTDONE indicates
 769  *      the cpu has come out of OS rendezvous.
 770  *
 771  *  Inputs  :   None
 772  *  Outputs :   None
 773  */
 774 static irqreturn_t
 775 ia64_mca_rendez_int_handler(int rendez_irq, void *arg)
 776 {
 777         unsigned long flags;
 778         int cpu = smp_processor_id();
 779         struct ia64_mca_notify_die nd =
 780                 { .sos = NULL, .monarch_cpu = &monarch_cpu };
 781
 782         /* Mask all interrupts */
 783         local_irq_save(flags);
 784
 785         NOTIFY_MCA(DIE_MCA_RENDZVOUS_ENTER, get_irq_regs(), (long)&nd, 1);
 786
 787         ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_DONE;
 788         /* Register with the SAL monarch that the slave has
 789          * reached SAL
 790          */
 791         ia64_sal_mc_rendez();
 792
 793         NOTIFY_MCA(DIE_MCA_RENDZVOUS_PROCESS, get_irq_regs(), (long)&nd, 1);
 794
 795         /* Wait for the monarch cpu to exit. */
 796         while (monarch_cpu != -1)
 797                cpu_relax();     /* spin until monarch leaves */
 798
 799         NOTIFY_MCA(DIE_MCA_RENDZVOUS_LEAVE, get_irq_regs(), (long)&nd, 1);
 800
 801         ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
 802         /* Enable all interrupts */
 803         local_irq_restore(flags);
 804         return IRQ_HANDLED;
 805 }
 806
 807 /*
 808  * ia64_mca_wakeup_int_handler
 809  *
 810  *      The interrupt handler for processing the inter-cpu interrupt to the
 811  *      slave cpu which was spinning in the rendez loop.
 812  *      Since this spinning is done by turning off the interrupts and
 813  *      polling on the wakeup-interrupt bit in the IRR, there is
 814  *      nothing useful to be done in the handler.
 815  *
 816  *  Inputs  :   wakeup_irq  (Wakeup-interrupt bit)
 817  *      arg             (Interrupt handler specific argument)
 818  *  Outputs :   None
 819  *
 820  */
 821 static irqreturn_t
 822 ia64_mca_wakeup_int_handler(int wakeup_irq, void *arg)
 823 {
 824         return IRQ_HANDLED;
 825 }
 826
 827 /* Function pointer for extra MCA recovery */
 828 int (*ia64_mca_ucmc_extension)
 829         (void*,struct ia64_sal_os_state*)
 830         = NULL;
 831
 832 int
 833 ia64_reg_MCA_extension(int (*fn)(void *, struct ia64_sal_os_state *))
 834 {
 835         if (ia64_mca_ucmc_extension)
 836                 return 1;
 837
 838         ia64_mca_ucmc_extension = fn;
 839         return 0;
 840 }
 841
 842 void
 843 ia64_unreg_MCA_extension(void)
 844 {
 845         if (ia64_mca_ucmc_extension)
 846                 ia64_mca_ucmc_extension = NULL;
 847 }
 848
 849 EXPORT_SYMBOL(ia64_reg_MCA_extension);
 850 EXPORT_SYMBOL(ia64_unreg_MCA_extension);
 851
 852
 853 static inline void
 854 copy_reg(const u64 *fr, u64 fnat, unsigned long *tr, unsigned long *tnat)
 855 {
 856         u64 fslot, tslot, nat;
 857         *tr = *fr;
 858         fslot = ((unsigned long)fr >> 3) & 63;
 859         tslot = ((unsigned long)tr >> 3) & 63;
 860         *tnat &= ~(1UL << tslot);
 861         nat = (fnat >> fslot) & 1;
 862         *tnat |= (nat << tslot);
 863 }
 864
 865 /* Change the comm field on the MCA/INT task to include the pid that
 866  * was interrupted, it makes for easier debugging.  If that pid was 0
 867  * (swapper or nested MCA/INIT) then use the start of the previous comm
 868  * field suffixed with its cpu.
 869  */
 870
 871 static void
 872 ia64_mca_modify_comm(const struct task_struct *previous_current)
 873 {
 874         char *p, comm[sizeof(current->comm)];
 875         if (previous_current->pid)
 876                 snprintf(comm, sizeof(comm), "%s %d",
 877                         current->comm, previous_current->pid);
 878         else {
 879                 int l;
 880                 if ((p = strchr(previous_current->comm, ' ')))
 881                         l = p - previous_current->comm;
 882                 else
 883                         l = strlen(previous_current->comm);
 884                 snprintf(comm, sizeof(comm), "%s %*s %d",
 885                         current->comm, l, previous_current->comm,
 886                         task_thread_info(previous_current)->cpu);
 887         }
 888         memcpy(current->comm, comm, sizeof(current->comm));
 889 }
 890
 891 static void
 892 finish_pt_regs(struct pt_regs *regs, struct ia64_sal_os_state *sos,
 893                 unsigned long *nat)
 894 {
 895         const pal_min_state_area_t *ms = sos->pal_min_state;
 896         const u64 *bank;
 897
 898         /* If ipsr.ic then use pmsa_{iip,ipsr,ifs}, else use
 899          * pmsa_{xip,xpsr,xfs}
 900          */
 901         if (ia64_psr(regs)->ic) {
 902                 regs->cr_iip = ms->pmsa_iip;
 903                 regs->cr_ipsr = ms->pmsa_ipsr;
 904                 regs->cr_ifs = ms->pmsa_ifs;
 905         } else {
 906                 regs->cr_iip = ms->pmsa_xip;
 907                 regs->cr_ipsr = ms->pmsa_xpsr;
 908                 regs->cr_ifs = ms->pmsa_xfs;
 909
 910                 sos->iip = ms->pmsa_iip;
 911                 sos->ipsr = ms->pmsa_ipsr;
 912                 sos->ifs = ms->pmsa_ifs;
 913         }
 914         regs->pr = ms->pmsa_pr;
 915         regs->b0 = ms->pmsa_br0;
 916         regs->ar_rsc = ms->pmsa_rsc;
 917         copy_reg(&ms->pmsa_gr[1-1], ms->pmsa_nat_bits, &regs->r1, nat);
 918         copy_reg(&ms->pmsa_gr[2-1], ms->pmsa_nat_bits, &regs->r2, nat);
 919         copy_reg(&ms->pmsa_gr[3-1], ms->pmsa_nat_bits, &regs->r3, nat);
 920         copy_reg(&ms->pmsa_gr[8-1], ms->pmsa_nat_bits, &regs->r8, nat);
 921         copy_reg(&ms->pmsa_gr[9-1], ms->pmsa_nat_bits, &regs->r9, nat);
 922         copy_reg(&ms->pmsa_gr[10-1], ms->pmsa_nat_bits, &regs->r10, nat);
 923         copy_reg(&ms->pmsa_gr[11-1], ms->pmsa_nat_bits, &regs->r11, nat);
 924         copy_reg(&ms->pmsa_gr[12-1], ms->pmsa_nat_bits, &regs->r12, nat);
 925         copy_reg(&ms->pmsa_gr[13-1], ms->pmsa_nat_bits, &regs->r13, nat);
 926         copy_reg(&ms->pmsa_gr[14-1], ms->pmsa_nat_bits, &regs->r14, nat);
 927         copy_reg(&ms->pmsa_gr[15-1], ms->pmsa_nat_bits, &regs->r15, nat);
 928         if (ia64_psr(regs)->bn)
 929                 bank = ms->pmsa_bank1_gr;
 930         else
 931                 bank = ms->pmsa_bank0_gr;
 932         copy_reg(&bank[16-16], ms->pmsa_nat_bits, &regs->r16, nat);
 933         copy_reg(&bank[17-16], ms->pmsa_nat_bits, &regs->r17, nat);
 934         copy_reg(&bank[18-16], ms->pmsa_nat_bits, &regs->r18, nat);
 935         copy_reg(&bank[19-16], ms->pmsa_nat_bits, &regs->r19, nat);
 936         copy_reg(&bank[20-16], ms->pmsa_nat_bits, &regs->r20, nat);
 937         copy_reg(&bank[21-16], ms->pmsa_nat_bits, &regs->r21, nat);
 938         copy_reg(&bank[22-16], ms->pmsa_nat_bits, &regs->r22, nat);
 939         copy_reg(&bank[23-16], ms->pmsa_nat_bits, &regs->r23, nat);
 940         copy_reg(&bank[24-16], ms->pmsa_nat_bits, &regs->r24, nat);
 941         copy_reg(&bank[25-16], ms->pmsa_nat_bits, &regs->r25, nat);
 942         copy_reg(&bank[26-16], ms->pmsa_nat_bits, &regs->r26, nat);
 943         copy_reg(&bank[27-16], ms->pmsa_nat_bits, &regs->r27, nat);
 944         copy_reg(&bank[28-16], ms->pmsa_nat_bits, &regs->r28, nat);
 945         copy_reg(&bank[29-16], ms->pmsa_nat_bits, &regs->r29, nat);
 946         copy_reg(&bank[30-16], ms->pmsa_nat_bits, &regs->r30, nat);
 947         copy_reg(&bank[31-16], ms->pmsa_nat_bits, &regs->r31, nat);
 948 }
 949
 950 /* On entry to this routine, we are running on the per cpu stack, see
 951  * mca_asm.h.  The original stack has not been touched by this event.  Some of
 952  * the original stack's registers will be in the RBS on this stack.  This stack
 953  * also contains a partial pt_regs and switch_stack, the rest of the data is in
 954  * PAL minstate.
 955  *
 956  * The first thing to do is modify the original stack to look like a blocked
 957  * task so we can run backtrace on the original task.  Also mark the per cpu
 958  * stack as current to ensure that we use the correct task state, it also means
 959  * that we can do backtrace on the MCA/INIT handler code itself.
 960  */
 961
 962 static struct task_struct *
 963 ia64_mca_modify_original_stack(struct pt_regs *regs,
 964                 const struct switch_stack *sw,
 965                 struct ia64_sal_os_state *sos,
 966                 const char *type)
 967 {
 968         char *p;
 969         ia64_va va;
 970         extern char ia64_leave_kernel[];        /* Need asm address, not function descriptor */
 971         const pal_min_state_area_t *ms = sos->pal_min_state;
 972         struct task_struct *previous_current;
 973         struct pt_regs *old_regs;
 974         struct switch_stack *old_sw;
 975         unsigned size = sizeof(struct pt_regs) +
 976                         sizeof(struct switch_stack) + 16;
 977         unsigned long *old_bspstore, *old_bsp;
 978         unsigned long *new_bspstore, *new_bsp;
 979         unsigned long old_unat, old_rnat, new_rnat, nat;
 980         u64 slots, loadrs = regs->loadrs;
 981         u64 r12 = ms->pmsa_gr[12-1], r13 = ms->pmsa_gr[13-1];
 982         u64 ar_bspstore = regs->ar_bspstore;
 983         u64 ar_bsp = regs->ar_bspstore + (loadrs >> 16);
 984         const char *msg;
 985         int cpu = smp_processor_id();
 986
 987         previous_current = curr_task(cpu);
 988         set_curr_task(cpu, current);
 989         if ((p = strchr(current->comm, ' ')))
 990                 *p = '\0';
 991
 992         /* Best effort attempt to cope with MCA/INIT delivered while in
 993          * physical mode.
 994          */
 995         regs->cr_ipsr = ms->pmsa_ipsr;
 996         if (ia64_psr(regs)->dt == 0) {
 997                 va.l = r12;
 998                 if (va.f.reg == 0) {
 999                         va.f.reg = 7;
1000                         r12 = va.l;
1001                 }
1002                 va.l = r13;
1003                 if (va.f.reg == 0) {
1004                         va.f.reg = 7;
1005                         r13 = va.l;
1006                 }
1007         }
1008         if (ia64_psr(regs)->rt == 0) {
1009                 va.l = ar_bspstore;
1010                 if (va.f.reg == 0) {
1011                         va.f.reg = 7;
1012                         ar_bspstore = va.l;
1013                 }
1014                 va.l = ar_bsp;
1015                 if (va.f.reg == 0) {
1016                         va.f.reg = 7;
1017                         ar_bsp = va.l;
1018                 }
1019         }
1020
1021         /* mca_asm.S ia64_old_stack() cannot assume that the dirty registers
1022          * have been copied to the old stack, the old stack may fail the
1023          * validation tests below.  So ia64_old_stack() must restore the dirty
1024          * registers from the new stack.  The old and new bspstore probably
1025          * have different alignments, so loadrs calculated on the old bsp
1026          * cannot be used to restore from the new bsp.  Calculate a suitable
1027          * loadrs for the new stack and save it in the new pt_regs, where
1028          * ia64_old_stack() can get it.
1029          */
1030         old_bspstore = (unsigned long *)ar_bspstore;
1031         old_bsp = (unsigned long *)ar_bsp;
1032         slots = ia64_rse_num_regs(old_bspstore, old_bsp);
1033         new_bspstore = (unsigned long *)((u64)current + IA64_RBS_OFFSET);
1034         new_bsp = ia64_rse_skip_regs(new_bspstore, slots);
1035         regs->loadrs = (new_bsp - new_bspstore) * 8 << 16;
1036
1037         /* Verify the previous stack state before we change it */
1038         if (user_mode(regs)) {
1039                 msg = "occurred in user space";
1040                 /* previous_current is guaranteed to be valid when the task was
1041                  * in user space, so ...
1042                  */
1043                 ia64_mca_modify_comm(previous_current);
1044                 goto no_mod;
1045         }
1046
1047         if (r13 != sos->prev_IA64_KR_CURRENT) {
1048                 msg = "inconsistent previous current and r13";
1049                 goto no_mod;
1050         }
1051
1052         if (!mca_recover_range(ms->pmsa_iip)) {
1053                 if ((r12 - r13) >= KERNEL_STACK_SIZE) {
1054                         msg = "inconsistent r12 and r13";
1055                         goto no_mod;
1056                 }
1057                 if ((ar_bspstore - r13) >= KERNEL_STACK_SIZE) {
1058                         msg = "inconsistent ar.bspstore and r13";
1059                         goto no_mod;
1060                 }
1061                 va.p = old_bspstore;
1062                 if (va.f.reg < 5) {
1063                         msg = "old_bspstore is in the wrong region";
1064                         goto no_mod;
1065                 }
1066                 if ((ar_bsp - r13) >= KERNEL_STACK_SIZE) {
1067                         msg = "inconsistent ar.bsp and r13";
1068                         goto no_mod;
1069                 }
1070                 size += (ia64_rse_skip_regs(old_bspstore, slots) - old_bspstore) * 8;
1071                 if (ar_bspstore + size > r12) {
1072                         msg = "no room for blocked state";
1073                         goto no_mod;
1074                 }
1075         }
1076
1077         ia64_mca_modify_comm(previous_current);
1078
1079         /* Make the original task look blocked.  First stack a struct pt_regs,
1080          * describing the state at the time of interrupt.  mca_asm.S built a
1081          * partial pt_regs, copy it and fill in the blanks using minstate.
1082          */
1083         p = (char *)r12 - sizeof(*regs);
1084         old_regs = (struct pt_regs *)p;
1085         memcpy(old_regs, regs, sizeof(*regs));
1086         old_regs->loadrs = loadrs;
1087         old_unat = old_regs->ar_unat;
1088         finish_pt_regs(old_regs, sos, &old_unat);
1089
1090         /* Next stack a struct switch_stack.  mca_asm.S built a partial
1091          * switch_stack, copy it and fill in the blanks using pt_regs and
1092          * minstate.
1093          *
1094          * In the synthesized switch_stack, b0 points to ia64_leave_kernel,
1095          * ar.pfs is set to 0.
1096          *
1097          * unwind.c::unw_unwind() does special processing for interrupt frames.
1098          * It checks if the PRED_NON_SYSCALL predicate is set, if the predicate
1099          * is clear then unw_unwind() does _not_ adjust bsp over pt_regs.  Not
1100          * that this is documented, of course.  Set PRED_NON_SYSCALL in the
1101          * switch_stack on the original stack so it will unwind correctly when
1102          * unwind.c reads pt_regs.
1103          *
1104          * thread.ksp is updated to point to the synthesized switch_stack.
1105          */
1106         p -= sizeof(struct switch_stack);
1107         old_sw = (struct switch_stack *)p;
1108         memcpy(old_sw, sw, sizeof(*sw));
1109         old_sw->caller_unat = old_unat;
1110         old_sw->ar_fpsr = old_regs->ar_fpsr;
1111         copy_reg(&ms->pmsa_gr[4-1], ms->pmsa_nat_bits, &old_sw->r4, &old_unat);
1112         copy_reg(&ms->pmsa_gr[5-1], ms->pmsa_nat_bits, &old_sw->r5, &old_unat);
1113         copy_reg(&ms->pmsa_gr[6-1], ms->pmsa_nat_bits, &old_sw->r6, &old_unat);
1114         copy_reg(&ms->pmsa_gr[7-1], ms->pmsa_nat_bits, &old_sw->r7, &old_unat);
1115         old_sw->b0 = (u64)ia64_leave_kernel;
1116         old_sw->b1 = ms->pmsa_br1;
1117         old_sw->ar_pfs = 0;
1118         old_sw->ar_unat = old_unat;
1119         old_sw->pr = old_regs->pr | (1UL << PRED_NON_SYSCALL);
1120         previous_current->thread.ksp = (u64)p - 16;
1121
1122         /* Finally copy the original stack's registers back to its RBS.
1123          * Registers from ar.bspstore through ar.bsp at the time of the event
1124          * are in the current RBS, copy them back to the original stack.  The
1125          * copy must be done register by register because the original bspstore
1126          * and the current one have different alignments, so the saved RNAT
1127          * data occurs at different places.
1128          *
1129          * mca_asm does cover, so the old_bsp already includes all registers at
1130          * the time of MCA/INIT.  It also does flushrs, so all registers before
1131          * this function have been written to backing store on the MCA/INIT
1132          * stack.
1133          */
1134         new_rnat = ia64_get_rnat(ia64_rse_rnat_addr(new_bspstore));
1135         old_rnat = regs->ar_rnat;
1136         while (slots--) {
1137                 if (ia64_rse_is_rnat_slot(new_bspstore)) {
1138                         new_rnat = ia64_get_rnat(new_bspstore++);
1139                 }
1140                 if (ia64_rse_is_rnat_slot(old_bspstore)) {
1141                         *old_bspstore++ = old_rnat;
1142                         old_rnat = 0;
1143                 }
1144                 nat = (new_rnat >> ia64_rse_slot_num(new_bspstore)) & 1UL;
1145                 old_rnat &= ~(1UL << ia64_rse_slot_num(old_bspstore));
1146                 old_rnat |= (nat << ia64_rse_slot_num(old_bspstore));
1147                 *old_bspstore++ = *new_bspstore++;
1148         }
1149         old_sw->ar_bspstore = (unsigned long)old_bspstore;
1150         old_sw->ar_rnat = old_rnat;
1151
1152         sos->prev_task = previous_current;
1153         return previous_current;
1154
1155 no_mod:
1156         mprintk(KERN_INFO "cpu %d, %s %s, original stack not modified\n",
1157                         smp_processor_id(), type, msg);
1158         old_unat = regs->ar_unat;
1159         finish_pt_regs(regs, sos, &old_unat);
1160         return previous_current;
1161 }
1162
1163 /* The monarch/slave interaction is based on monarch_cpu and requires that all
1164  * slaves have entered rendezvous before the monarch leaves.  If any cpu has
1165  * not entered rendezvous yet then wait a bit.  The assumption is that any
1166  * slave that has not rendezvoused after a reasonable time is never going to do
1167  * so.  In this context, slave includes cpus that respond to the MCA rendezvous
1168  * interrupt, as well as cpus that receive the INIT slave event.
1169  */
1170
1171 static void
1172 ia64_wait_for_slaves(int monarch, const char *type)
1173 {
1174         int c, i , wait;
1175
1176         /*
1177          * wait 5 seconds total for slaves (arbitrary)
1178          */
1179         for (i = 0; i < 5000; i++) {
1180                 wait = 0;
1181                 for_each_online_cpu(c) {
1182                         if (c == monarch)
1183                                 continue;
1184                         if (ia64_mc_info.imi_rendez_checkin[c]
1185                                         == IA64_MCA_RENDEZ_CHECKIN_NOTDONE) {
1186                                 udelay(1000);           /* short wait */
1187                                 wait = 1;
1188                                 break;
1189                         }
1190                 }
1191                 if (!wait)
1192                         goto all_in;
1193         }
1194
1195         /*
1196          * Maybe slave(s) dead. Print buffered messages immediately.
1197          */
1198         ia64_mlogbuf_finish(0);
1199         mprintk(KERN_INFO "OS %s slave did not rendezvous on cpu", type);
1200         for_each_online_cpu(c) {
1201                 if (c == monarch)
1202                         continue;
1203                 if (ia64_mc_info.imi_rendez_checkin[c] == IA64_MCA_RENDEZ_CHECKIN_NOTDONE)
1204                         mprintk(" %d", c);
1205         }
1206         mprintk("\n");
1207         return;
1208
1209 all_in:
1210         mprintk(KERN_INFO "All OS %s slaves have reached rendezvous\n", type);
1211         return;
1212 }
1213
1214 /*  mca_insert_tr
1215  *
1216  *  Switch rid when TR reload and needed!
1217  *  iord: 1: itr, 2: itr;
1218  *
1219 */
1220 static void mca_insert_tr(u64 iord)
1221 {
1222
1223         int i;
1224         u64 old_rr;
1225         struct ia64_tr_entry *p;
1226         unsigned long psr;
1227         int cpu = smp_processor_id();
1228
1229         if (!ia64_idtrs[cpu])
1230                 return;
1231
1232         psr = ia64_clear_ic();
1233         for (i = IA64_TR_ALLOC_BASE; i < IA64_TR_ALLOC_MAX; i++) {
1234                 p = ia64_idtrs[cpu] + (iord - 1) * IA64_TR_ALLOC_MAX;
1235                 if (p->pte & 0x1) {
1236                         old_rr = ia64_get_rr(p->ifa);
1237                         if (old_rr != p->rr) {
1238                                 ia64_set_rr(p->ifa, p->rr);
1239                                 ia64_srlz_d();
1240                         }
1241                         ia64_ptr(iord, p->ifa, p->itir >> 2);
1242                         ia64_srlz_i();
1243                         if (iord & 0x1) {
1244                                 ia64_itr(0x1, i, p->ifa, p->pte, p->itir >> 2);
1245                                 ia64_srlz_i();
1246                         }
1247                         if (iord & 0x2) {
1248                                 ia64_itr(0x2, i, p->ifa, p->pte, p->itir >> 2);
1249                                 ia64_srlz_i();
1250                         }
1251                         if (old_rr != p->rr) {
1252                                 ia64_set_rr(p->ifa, old_rr);
1253                                 ia64_srlz_d();
1254                         }
1255                 }
1256         }
1257         ia64_set_psr(psr);
1258 }
1259
1260 /*
1261  * ia64_mca_handler
1262  *
1263  *      This is uncorrectable machine check handler called from OS_MCA
1264  *      dispatch code which is in turn called from SAL_CHECK().
1265  *      This is the place where the core of OS MCA handling is done.
1266  *      Right now the logs are extracted and displayed in a well-defined
1267  *      format. This handler code is supposed to be run only on the
1268  *      monarch processor. Once the monarch is done with MCA handling
1269  *      further MCA logging is enabled by clearing logs.
1270  *      Monarch also has the duty of sending wakeup-IPIs to pull the
1271  *      slave processors out of rendezvous spinloop.
1272  *
1273  *      If multiple processors call into OS_MCA, the first will become
1274  *      the monarch.  Subsequent cpus will be recorded in the mca_cpu
1275  *      bitmask.  After the first monarch has processed its MCA, it
1276  *      will wake up the next cpu in the mca_cpu bitmask and then go
1277  *      into the rendezvous loop.  When all processors have serviced
1278  *      their MCA, the last monarch frees up the rest of the processors.
1279  */
1280 void
1281 ia64_mca_handler(struct pt_regs *regs, struct switch_stack *sw,
1282                  struct ia64_sal_os_state *sos)
1283 {
1284         int recover, cpu = smp_processor_id();
1285         struct task_struct *previous_current;
1286         struct ia64_mca_notify_die nd =
1287                 { .sos = sos, .monarch_cpu = &monarch_cpu, .data = &recover };
1288         static atomic_t mca_count;
1289         static cpumask_t mca_cpu;
1290
1291         if (atomic_add_return(1, &mca_count) == 1) {
1292                 monarch_cpu = cpu;
1293                 sos->monarch = 1;
1294         } else {
1295                 cpu_set(cpu, mca_cpu);
1296                 sos->monarch = 0;
1297         }
1298         mprintk(KERN_INFO "Entered OS MCA handler. PSP=%lx cpu=%d "
1299                 "monarch=%ld\n", sos->proc_state_param, cpu, sos->monarch);
1300
1301         previous_current = ia64_mca_modify_original_stack(regs, sw, sos, "MCA");
1302
1303         NOTIFY_MCA(DIE_MCA_MONARCH_ENTER, regs, (long)&nd, 1);
1304
1305         ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_CONCURRENT_MCA;
1306         if (sos->monarch) {
1307                 ia64_wait_for_slaves(cpu, "MCA");
1308
1309                 /* Wakeup all the processors which are spinning in the
1310                  * rendezvous loop.  They will leave SAL, then spin in the OS
1311                  * with interrupts disabled until this monarch cpu leaves the
1312                  * MCA handler.  That gets control back to the OS so we can
1313                  * backtrace the other cpus, backtrace when spinning in SAL
1314                  * does not work.
1315                  */
1316                 ia64_mca_wakeup_all();
1317         } else {
1318                 while (cpu_isset(cpu, mca_cpu))
1319                         cpu_relax();    /* spin until monarch wakes us */
1320         }
1321
1322         NOTIFY_MCA(DIE_MCA_MONARCH_PROCESS, regs, (long)&nd, 1);
1323
1324         /* Get the MCA error record and log it */
1325         ia64_mca_log_sal_error_record(SAL_INFO_TYPE_MCA);
1326
1327         /* MCA error recovery */
1328         recover = (ia64_mca_ucmc_extension
1329                 && ia64_mca_ucmc_extension(
1330                         IA64_LOG_CURR_BUFFER(SAL_INFO_TYPE_MCA),
1331                         sos));
1332
1333         if (recover) {
1334                 sal_log_record_header_t *rh = IA64_LOG_CURR_BUFFER(SAL_INFO_TYPE_MCA);
1335                 rh->severity = sal_log_severity_corrected;
1336                 ia64_sal_clear_state_info(SAL_INFO_TYPE_MCA);
1337                 sos->os_status = IA64_MCA_CORRECTED;
1338         } else {
1339                 /* Dump buffered message to console */
1340                 ia64_mlogbuf_finish(1);
1341         }
1342
1343         if (__get_cpu_var(ia64_mca_tr_reload)) {
1344                 mca_insert_tr(0x1); /*Reload dynamic itrs*/
1345                 mca_insert_tr(0x2); /*Reload dynamic itrs*/
1346         }
1347
1348         NOTIFY_MCA(DIE_MCA_MONARCH_LEAVE, regs, (long)&nd, 1);
1349
1350         if (atomic_dec_return(&mca_count) > 0) {
1351                 int i;
1352
1353                 /* wake up the next monarch cpu,
1354                  * and put this cpu in the rendez loop.
1355                  */
1356                 for_each_online_cpu(i) {
1357                         if (cpu_isset(i, mca_cpu)) {
1358                                 monarch_cpu = i;
1359                                 cpu_clear(i, mca_cpu);  /* wake next cpu */
1360                                 while (monarch_cpu != -1)
1361                                         cpu_relax();    /* spin until last cpu leaves */
1362                                 set_curr_task(cpu, previous_current);
1363                                 ia64_mc_info.imi_rendez_checkin[cpu]
1364                                                 = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1365                                 return;
1366                         }
1367                 }
1368         }
1369         set_curr_task(cpu, previous_current);
1370         ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1371         monarch_cpu = -1;       /* This frees the slaves and previous monarchs */
1372 }
1373
1374 static DECLARE_WORK(cmc_disable_work, ia64_mca_cmc_vector_disable_keventd);
1375 static DECLARE_WORK(cmc_enable_work, ia64_mca_cmc_vector_enable_keventd);
1376
1377 /*
1378  * ia64_mca_cmc_int_handler
1379  *
1380  *  This is corrected machine check interrupt handler.
1381  *      Right now the logs are extracted and displayed in a well-defined
1382  *      format.
1383  *
1384  * Inputs
1385  *      interrupt number
1386  *      client data arg ptr
1387  *
1388  * Outputs
1389  *      None
1390  */
1391 static irqreturn_t
1392 ia64_mca_cmc_int_handler(int cmc_irq, void *arg)
1393 {
1394         static unsigned long    cmc_history[CMC_HISTORY_LENGTH];
1395         static int              index;
1396         static DEFINE_SPINLOCK(cmc_history_lock);
1397
1398         IA64_MCA_DEBUG("%s: received interrupt vector = %#x on CPU %d\n",
1399                        __func__, cmc_irq, smp_processor_id());
1400
1401         /* SAL spec states this should run w/ interrupts enabled */
1402         local_irq_enable();
1403
1404         spin_lock(&cmc_history_lock);
1405         if (!cmc_polling_enabled) {
1406                 int i, count = 1; /* we know 1 happened now */
1407                 unsigned long now = jiffies;
1408
1409                 for (i = 0; i < CMC_HISTORY_LENGTH; i++) {
1410                         if (now - cmc_history[i] <= HZ)
1411                                 count++;
1412                 }
1413
1414                 IA64_MCA_DEBUG(KERN_INFO "CMC threshold %d/%d\n", count, CMC_HISTORY_LENGTH);
1415                 if (count >= CMC_HISTORY_LENGTH) {
1416
1417                         cmc_polling_enabled = 1;
1418                         spin_unlock(&cmc_history_lock);
1419                         /* If we're being hit with CMC interrupts, we won't
1420                          * ever execute the schedule_work() below.  Need to
1421                          * disable CMC interrupts on this processor now.
1422                          */
1423                         ia64_mca_cmc_vector_disable(NULL);
1424                         schedule_work(&cmc_disable_work);
1425
1426                         /*
1427                          * Corrected errors will still be corrected, but
1428                          * make sure there's a log somewhere that indicates
1429                          * something is generating more than we can handle.
1430                          */
1431                         printk(KERN_WARNING "WARNING: Switching to polling CMC handler; error records may be lost\n");
1432
1433                         mod_timer(&cmc_poll_timer, jiffies + CMC_POLL_INTERVAL);
1434
1435                         /* lock already released, get out now */
1436                         goto out;
1437                 } else {
1438                         cmc_history[index++] = now;
1439                         if (index == CMC_HISTORY_LENGTH)
1440                                 index = 0;
1441                 }
1442         }
1443         spin_unlock(&cmc_history_lock);
1444 out:
1445         /* Get the CMC error record and log it */
1446         ia64_mca_log_sal_error_record(SAL_INFO_TYPE_CMC);
1447
1448         return IRQ_HANDLED;
1449 }
1450
1451 /*
1452  *  ia64_mca_cmc_int_caller
1453  *
1454  *      Triggered by sw interrupt from CMC polling routine.  Calls
1455  *      real interrupt handler and either triggers a sw interrupt
1456  *      on the next cpu or does cleanup at the end.
1457  *
1458  * Inputs
1459  *      interrupt number
1460  *      client data arg ptr
1461  * Outputs
1462  *      handled
1463  */
1464 static irqreturn_t
1465 ia64_mca_cmc_int_caller(int cmc_irq, void *arg)
1466 {
1467         static int start_count = -1;
1468         unsigned int cpuid;
1469
1470         cpuid = smp_processor_id();
1471
1472         /* If first cpu, update count */
1473         if (start_count == -1)
1474                 start_count = IA64_LOG_COUNT(SAL_INFO_TYPE_CMC);
1475
1476         ia64_mca_cmc_int_handler(cmc_irq, arg);
1477
1478         cpuid = cpumask_next(cpuid+1, cpu_online_mask);
1479
1480         if (cpuid < nr_cpu_ids) {
1481                 platform_send_ipi(cpuid, IA64_CMCP_VECTOR, IA64_IPI_DM_INT, 0);
1482         } else {
1483                 /* If no log record, switch out of polling mode */
1484                 if (start_count == IA64_LOG_COUNT(SAL_INFO_TYPE_CMC)) {
1485
1486                         printk(KERN_WARNING "Returning to interrupt driven CMC handler\n");
1487                         schedule_work(&cmc_enable_work);
1488                         cmc_polling_enabled = 0;
1489
1490                 } else {
1491
1492                         mod_timer(&cmc_poll_timer, jiffies + CMC_POLL_INTERVAL);
1493                 }
1494
1495                 start_count = -1;
1496         }
1497
1498         return IRQ_HANDLED;
1499 }
1500
1501 /*
1502  *  ia64_mca_cmc_poll
1503  *
1504  *      Poll for Corrected Machine Checks (CMCs)
1505  *
1506  * Inputs   :   dummy(unused)
1507  * Outputs  :   None
1508  *
1509  */
1510 static void
1511 ia64_mca_cmc_poll (unsigned long dummy)
1512 {
1513         /* Trigger a CMC interrupt cascade  */
1514         platform_send_ipi(first_cpu(cpu_online_map), IA64_CMCP_VECTOR, IA64_IPI_DM_INT, 0);
1515 }
1516
1517 /*
1518  *  ia64_mca_cpe_int_caller
1519  *
1520  *      Triggered by sw interrupt from CPE polling routine.  Calls
1521  *      real interrupt handler and either triggers a sw interrupt
1522  *      on the next cpu or does cleanup at the end.
1523  *
1524  * Inputs
1525  *      interrupt number
1526  *      client data arg ptr
1527  * Outputs
1528  *      handled
1529  */
1530 #ifdef CONFIG_ACPI
1531
1532 static irqreturn_t
1533 ia64_mca_cpe_int_caller(int cpe_irq, void *arg)
1534 {
1535         static int start_count = -1;
1536         static int poll_time = MIN_CPE_POLL_INTERVAL;
1537         unsigned int cpuid;
1538
1539         cpuid = smp_processor_id();
1540
1541         /* If first cpu, update count */
1542         if (start_count == -1)
1543                 start_count = IA64_LOG_COUNT(SAL_INFO_TYPE_CPE);
1544
1545         ia64_mca_cpe_int_handler(cpe_irq, arg);
1546
1547         cpuid = cpumask_next(cpuid+1, cpu_online_mask);
1548
1549         if (cpuid < NR_CPUS) {
1550                 platform_send_ipi(cpuid, IA64_CPEP_VECTOR, IA64_IPI_DM_INT, 0);
1551         } else {
1552                 /*
1553                  * If a log was recorded, increase our polling frequency,
1554                  * otherwise, backoff or return to interrupt mode.
1555                  */
1556                 if (start_count != IA64_LOG_COUNT(SAL_INFO_TYPE_CPE)) {
1557                         poll_time = max(MIN_CPE_POLL_INTERVAL, poll_time / 2);
1558                 } else if (cpe_vector < 0) {
1559                         poll_time = min(MAX_CPE_POLL_INTERVAL, poll_time * 2);
1560                 } else {
1561                         poll_time = MIN_CPE_POLL_INTERVAL;
1562
1563                         printk(KERN_WARNING "Returning to interrupt driven CPE handler\n");
1564                         enable_irq(local_vector_to_irq(IA64_CPE_VECTOR));
1565                         cpe_poll_enabled = 0;
1566                 }
1567
1568                 if (cpe_poll_enabled)
1569                         mod_timer(&cpe_poll_timer, jiffies + poll_time);
1570                 start_count = -1;
1571         }
1572
1573         return IRQ_HANDLED;
1574 }
1575
1576 /*
1577  *  ia64_mca_cpe_poll
1578  *
1579  *      Poll for Corrected Platform Errors (CPEs), trigger interrupt
1580  *      on first cpu, from there it will trickle through all the cpus.
1581  *
1582  * Inputs   :   dummy(unused)
1583  * Outputs  :   None
1584  *
1585  */
1586 static void
1587 ia64_mca_cpe_poll (unsigned long dummy)
1588 {
1589         /* Trigger a CPE interrupt cascade  */
1590         platform_send_ipi(first_cpu(cpu_online_map), IA64_CPEP_VECTOR, IA64_IPI_DM_INT, 0);
1591 }
1592
1593 #endif /* CONFIG_ACPI */
1594
1595 static int
1596 default_monarch_init_process(struct notifier_block *self, unsigned long val, void *data)
1597 {
1598         int c;
1599         struct task_struct *g, *t;
1600         if (val != DIE_INIT_MONARCH_PROCESS)
1601                 return NOTIFY_DONE;
1602 #ifdef CONFIG_KEXEC
1603         if (atomic_read(&kdump_in_progress))
1604                 return NOTIFY_DONE;
1605 #endif
1606
1607         /*
1608          * FIXME: mlogbuf will brim over with INIT stack dumps.
1609          * To enable show_stack from INIT, we use oops_in_progress which should
1610          * be used in real oops. This would cause something wrong after INIT.
1611          */
1612         BREAK_LOGLEVEL(console_loglevel);
1613         ia64_mlogbuf_dump_from_init();
1614
1615         printk(KERN_ERR "Processes interrupted by INIT -");
1616         for_each_online_cpu(c) {
1617                 struct ia64_sal_os_state *s;
1618                 t = __va(__per_cpu_mca[c] + IA64_MCA_CPU_INIT_STACK_OFFSET);
1619                 s = (struct ia64_sal_os_state *)((char *)t + MCA_SOS_OFFSET);
1620                 g = s->prev_task;
1621                 if (g) {
1622                         if (g->pid)
1623                                 printk(" %d", g->pid);
1624                         else
1625                                 printk(" %d (cpu %d task 0x%p)", g->pid, task_cpu(g), g);
1626                 }
1627         }
1628         printk("\n\n");
1629         if (read_trylock(&tasklist_lock)) {
1630                 do_each_thread (g, t) {
1631                         printk("\nBacktrace of pid %d (%s)\n", t->pid, t->comm);
1632                         show_stack(t, NULL);
1633                 } while_each_thread (g, t);
1634                 read_unlock(&tasklist_lock);
1635         }
1636         /* FIXME: This will not restore zapped printk locks. */
1637         RESTORE_LOGLEVEL(console_loglevel);
1638         return NOTIFY_DONE;
1639 }
1640
1641 /*
1642  * C portion of the OS INIT handler
1643  *
1644  * Called from ia64_os_init_dispatch
1645  *
1646  * Inputs: pointer to pt_regs where processor info was saved.  SAL/OS state for
1647  * this event.  This code is used for both monarch and slave INIT events, see
1648  * sos->monarch.
1649  *
1650  * All INIT events switch to the INIT stack and change the previous process to
1651  * blocked status.  If one of the INIT events is the monarch then we are
1652  * probably processing the nmi button/command.  Use the monarch cpu to dump all
1653  * the processes.  The slave INIT events all spin until the monarch cpu
1654  * returns.  We can also get INIT slave events for MCA, in which case the MCA
1655  * process is the monarch.
1656  */
1657
1658 void
1659 ia64_init_handler(struct pt_regs *regs, struct switch_stack *sw,
1660                   struct ia64_sal_os_state *sos)
1661 {
1662         static atomic_t slaves;
1663         static atomic_t monarchs;
1664         struct task_struct *previous_current;
1665         int cpu = smp_processor_id();
1666         struct ia64_mca_notify_die nd =
1667                 { .sos = sos, .monarch_cpu = &monarch_cpu };
1668
1669         NOTIFY_INIT(DIE_INIT_ENTER, regs, (long)&nd, 0);
1670
1671         mprintk(KERN_INFO "Entered OS INIT handler. PSP=%lx cpu=%d monarch=%ld\n",
1672                 sos->proc_state_param, cpu, sos->monarch);
1673         salinfo_log_wakeup(SAL_INFO_TYPE_INIT, NULL, 0, 0);
1674
1675         previous_current = ia64_mca_modify_original_stack(regs, sw, sos, "INIT");
1676         sos->os_status = IA64_INIT_RESUME;
1677
1678         /* FIXME: Workaround for broken proms that drive all INIT events as
1679          * slaves.  The last slave that enters is promoted to be a monarch.
1680          * Remove this code in September 2006, that gives platforms a year to
1681          * fix their proms and get their customers updated.
1682          */
1683         if (!sos->monarch && atomic_add_return(1, &slaves) == num_online_cpus()) {
1684                 mprintk(KERN_WARNING "%s: Promoting cpu %d to monarch.\n",
1685                         __func__, cpu);
1686                 atomic_dec(&slaves);
1687                 sos->monarch = 1;
1688         }
1689
1690         /* FIXME: Workaround for broken proms that drive all INIT events as
1691          * monarchs.  Second and subsequent monarchs are demoted to slaves.
1692          * Remove this code in September 2006, that gives platforms a year to
1693          * fix their proms and get their customers updated.
1694          */
1695         if (sos->monarch && atomic_add_return(1, &monarchs) > 1) {
1696                 mprintk(KERN_WARNING "%s: Demoting cpu %d to slave.\n",
1697                                __func__, cpu);
1698                 atomic_dec(&monarchs);
1699                 sos->monarch = 0;
1700         }
1701
1702         if (!sos->monarch) {
1703                 ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_INIT;
1704
1705 #ifdef CONFIG_KEXEC
1706                 while (monarch_cpu == -1 && !atomic_read(&kdump_in_progress))
1707                         udelay(1000);
1708 #else
1709                 while (monarch_cpu == -1)
1710                         cpu_relax();    /* spin until monarch enters */
1711 #endif
1712
1713                 NOTIFY_INIT(DIE_INIT_SLAVE_ENTER, regs, (long)&nd, 1);
1714                 NOTIFY_INIT(DIE_INIT_SLAVE_PROCESS, regs, (long)&nd, 1);
1715
1716 #ifdef CONFIG_KEXEC
1717                 while (monarch_cpu != -1 && !atomic_read(&kdump_in_progress))
1718                         udelay(1000);
1719 #else
1720                 while (monarch_cpu != -1)
1721                         cpu_relax();    /* spin until monarch leaves */
1722 #endif
1723
1724                 NOTIFY_INIT(DIE_INIT_SLAVE_LEAVE, regs, (long)&nd, 1);
1725
1726                 mprintk("Slave on cpu %d returning to normal service.\n", cpu);
1727                 set_curr_task(cpu, previous_current);
1728                 ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1729                 atomic_dec(&slaves);
1730                 return;
1731         }
1732
1733         monarch_cpu = cpu;
1734         NOTIFY_INIT(DIE_INIT_MONARCH_ENTER, regs, (long)&nd, 1);
1735
1736         /*
1737          * Wait for a bit.  On some machines (e.g., HP's zx2000 and zx6000, INIT can be
1738          * generated via the BMC's command-line interface, but since the console is on the
1739          * same serial line, the user will need some time to switch out of the BMC before
1740          * the dump begins.
1741          */
1742         mprintk("Delaying for 5 seconds...\n");
1743         udelay(5*1000000);
1744         ia64_wait_for_slaves(cpu, "INIT");
1745         /* If nobody intercepts DIE_INIT_MONARCH_PROCESS then we drop through
1746          * to default_monarch_init_process() above and just print all the
1747          * tasks.
1748          */
1749         NOTIFY_INIT(DIE_INIT_MONARCH_PROCESS, regs, (long)&nd, 1);
1750         NOTIFY_INIT(DIE_INIT_MONARCH_LEAVE, regs, (long)&nd, 1);
1751
1752         mprintk("\nINIT dump complete.  Monarch on cpu %d returning to normal service.\n", cpu);
1753         atomic_dec(&monarchs);
1754         set_curr_task(cpu, previous_current);
1755         monarch_cpu = -1;
1756         return;
1757 }
1758
1759 static int __init
1760 ia64_mca_disable_cpe_polling(char *str)
1761 {
1762         cpe_poll_enabled = 0;
1763         return 1;
1764 }
1765
1766 __setup("disable_cpe_poll", ia64_mca_disable_cpe_polling);
1767
1768 static struct irqaction cmci_irqaction = {
1769         .handler =      ia64_mca_cmc_int_handler,
1770         .flags =        IRQF_DISABLED,
1771         .name =         "cmc_hndlr"
1772 };
1773
1774 static struct irqaction cmcp_irqaction = {
1775         .handler =      ia64_mca_cmc_int_caller,
1776         .flags =        IRQF_DISABLED,
1777         .name =         "cmc_poll"
1778 };
1779
1780 static struct irqaction mca_rdzv_irqaction = {
1781         .handler =      ia64_mca_rendez_int_handler,
1782         .flags =        IRQF_DISABLED,
1783         .name =         "mca_rdzv"
1784 };
1785
1786 static struct irqaction mca_wkup_irqaction = {
1787         .handler =      ia64_mca_wakeup_int_handler,
1788         .flags =        IRQF_DISABLED,
1789         .name =         "mca_wkup"
1790 };
1791
1792 #ifdef CONFIG_ACPI
1793 static struct irqaction mca_cpe_irqaction = {
1794         .handler =      ia64_mca_cpe_int_handler,
1795         .flags =        IRQF_DISABLED,
1796         .name =         "cpe_hndlr"
1797 };
1798
1799 static struct irqaction mca_cpep_irqaction = {
1800         .handler =      ia64_mca_cpe_int_caller,
1801         .flags =        IRQF_DISABLED,
1802         .name =         "cpe_poll"
1803 };
1804 #endif /* CONFIG_ACPI */
1805
1806 /* Minimal format of the MCA/INIT stacks.  The pseudo processes that run on
1807  * these stacks can never sleep, they cannot return from the kernel to user
1808  * space, they do not appear in a normal ps listing.  So there is no need to
1809  * format most of the fields.
1810  */
1811
1812 static void __cpuinit
1813 format_mca_init_stack(void *mca_data, unsigned long offset,
1814                 const char *type, int cpu)
1815 {
1816         struct task_struct *p = (struct task_struct *)((char *)mca_data + offset);
1817         struct thread_info *ti;
1818         memset(p, 0, KERNEL_STACK_SIZE);
1819         ti = task_thread_info(p);
1820         ti->flags = _TIF_MCA_INIT;
1821         ti->preempt_count = 1;
1822         ti->task = p;
1823         ti->cpu = cpu;
1824         p->stack = ti;
1825         p->state = TASK_UNINTERRUPTIBLE;
1826         cpu_set(cpu, p->cpus_allowed);
1827         INIT_LIST_HEAD(&p->tasks);
1828         p->parent = p->real_parent = p->group_leader = p;
1829         INIT_LIST_HEAD(&p->children);
1830         INIT_LIST_HEAD(&p->sibling);
1831         strncpy(p->comm, type, sizeof(p->comm)-1);
1832 }
1833
1834 /* Caller prevents this from being called after init */
1835 static void * __init_refok mca_bootmem(void)
1836 {
1837         return __alloc_bootmem(sizeof(struct ia64_mca_cpu),
1838                             KERNEL_STACK_SIZE, 0);
1839 }
1840
1841 /* Do per-CPU MCA-related initialization.  */
1842 void __cpuinit
1843 ia64_mca_cpu_init(void *cpu_data)
1844 {
1845         void *pal_vaddr;
1846         void *data;
1847         long sz = sizeof(struct ia64_mca_cpu);
1848         int cpu = smp_processor_id();
1849         static int first_time = 1;
1850
1851         /*
1852          * Structure will already be allocated if cpu has been online,
1853          * then offlined.
1854          */
1855         if (__per_cpu_mca[cpu]) {
1856                 data = __va(__per_cpu_mca[cpu]);
1857         } else {
1858                 if (first_time) {
1859                         data = mca_bootmem();
1860                         first_time = 0;
1861                 } else
1862                         data = __get_free_pages(GFP_KERNEL, get_order(sz));
1863                 if (!data)
1864                         panic("Could not allocate MCA memory for cpu %d\n",
1865                                         cpu);
1866         }
1867         format_mca_init_stack(data, offsetof(struct ia64_mca_cpu, mca_stack),
1868                 "MCA", cpu);
1869         format_mca_init_stack(data, offsetof(struct ia64_mca_cpu, init_stack),
1870                 "INIT", cpu);
1871         __get_cpu_var(ia64_mca_data) = __per_cpu_mca[cpu] = __pa(data);
1872
1873         /*
1874          * Stash away a copy of the PTE needed to map the per-CPU page.
1875          * We may need it during MCA recovery.
1876          */
1877         __get_cpu_var(ia64_mca_per_cpu_pte) =
1878                 pte_val(mk_pte_phys(__pa(cpu_data), PAGE_KERNEL));
1879
1880         /*
1881          * Also, stash away a copy of the PAL address and the PTE
1882          * needed to map it.
1883          */
1884         pal_vaddr = efi_get_pal_addr();
1885         if (!pal_vaddr)
1886                 return;
1887         __get_cpu_var(ia64_mca_pal_base) =
1888                 GRANULEROUNDDOWN((unsigned long) pal_vaddr);
1889         __get_cpu_var(ia64_mca_pal_pte) = pte_val(mk_pte_phys(__pa(pal_vaddr),
1890                                                               PAGE_KERNEL));
1891 }
1892
1893 static void __cpuinit ia64_mca_cmc_vector_adjust(void *dummy)
1894 {
1895         unsigned long flags;
1896
1897         local_irq_save(flags);
1898         if (!cmc_polling_enabled)
1899                 ia64_mca_cmc_vector_enable(NULL);
1900         local_irq_restore(flags);
1901 }
1902
1903 static int __cpuinit mca_cpu_callback(struct notifier_block *nfb,
1904                                       unsigned long action,
1905                                       void *hcpu)
1906 {
1907         int hotcpu = (unsigned long) hcpu;
1908
1909         switch (action) {
1910         case CPU_ONLINE:
1911         case CPU_ONLINE_FROZEN:
1912                 smp_call_function_single(hotcpu, ia64_mca_cmc_vector_adjust,
1913                                          NULL, 0);
1914                 break;
1915         }
1916         return NOTIFY_OK;
1917 }
1918
1919 static struct notifier_block mca_cpu_notifier __cpuinitdata = {
1920         .notifier_call = mca_cpu_callback
1921 };
1922
1923 /*
1924  * ia64_mca_init
1925  *
1926  *  Do all the system level mca specific initialization.
1927  *
1928  *      1. Register spinloop and wakeup request interrupt vectors
1929  *
1930  *      2. Register OS_MCA handler entry point
1931  *
1932  *      3. Register OS_INIT handler entry point
1933  *
1934  *  4. Initialize MCA/CMC/INIT related log buffers maintained by the OS.
1935  *
1936  *  Note that this initialization is done very early before some kernel
1937  *  services are available.
1938  *
1939  *  Inputs  :   None
1940  *
1941  *  Outputs :   None
1942  */
1943 void __init
1944 ia64_mca_init(void)
1945 {
1946         ia64_fptr_t *init_hldlr_ptr_monarch = (ia64_fptr_t *)ia64_os_init_dispatch_monarch;
1947         ia64_fptr_t *init_hldlr_ptr_slave = (ia64_fptr_t *)ia64_os_init_dispatch_slave;
1948         ia64_fptr_t *mca_hldlr_ptr = (ia64_fptr_t *)ia64_os_mca_dispatch;
1949         int i;
1950         long rc;
1951         struct ia64_sal_retval isrv;
1952         unsigned long timeout = IA64_MCA_RENDEZ_TIMEOUT; /* platform specific */
1953         static struct notifier_block default_init_monarch_nb = {
1954                 .notifier_call = default_monarch_init_process,
1955                 .priority = 0/* we need to notified last */
1956         };
1957
1958         IA64_MCA_DEBUG("%s: begin\n", __func__);
1959
1960         /* Clear the Rendez checkin flag for all cpus */
1961         for(i = 0 ; i < NR_CPUS; i++)
1962                 ia64_mc_info.imi_rendez_checkin[i] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1963
1964         /*
1965          * Register the rendezvous spinloop and wakeup mechanism with SAL
1966          */
1967
1968         /* Register the rendezvous interrupt vector with SAL */
1969         while (1) {
1970                 isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_RENDEZ_INT,
1971                                               SAL_MC_PARAM_MECHANISM_INT,
1972                                               IA64_MCA_RENDEZ_VECTOR,
1973                                               timeout,
1974                                               SAL_MC_PARAM_RZ_ALWAYS);
1975                 rc = isrv.status;
1976                 if (rc == 0)
1977                         break;
1978                 if (rc == -2) {
1979                         printk(KERN_INFO "Increasing MCA rendezvous timeout from "
1980                                 "%ld to %ld milliseconds\n", timeout, isrv.v0);
1981                         timeout = isrv.v0;
1982                         NOTIFY_MCA(DIE_MCA_NEW_TIMEOUT, NULL, timeout, 0);
1983                         continue;
1984                 }
1985                 printk(KERN_ERR "Failed to register rendezvous interrupt "
1986                        "with SAL (status %ld)\n", rc);
1987                 return;
1988         }
1989
1990         /* Register the wakeup interrupt vector with SAL */
1991         isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_RENDEZ_WAKEUP,
1992                                       SAL_MC_PARAM_MECHANISM_INT,
1993                                       IA64_MCA_WAKEUP_VECTOR,
1994                                       0, 0);
1995         rc = isrv.status;
1996         if (rc) {
1997                 printk(KERN_ERR "Failed to register wakeup interrupt with SAL "
1998                        "(status %ld)\n", rc);
1999                 return;
2000         }
2001
2002         IA64_MCA_DEBUG("%s: registered MCA rendezvous spinloop and wakeup mech.\n", __func__);
2003
2004         ia64_mc_info.imi_mca_handler        = ia64_tpa(mca_hldlr_ptr->fp);
2005         /*
2006          * XXX - disable SAL checksum by setting size to 0; should be
2007          *      ia64_tpa(ia64_os_mca_dispatch_end) - ia64_tpa(ia64_os_mca_dispatch);
2008          */
2009         ia64_mc_info.imi_mca_handler_size       = 0;
2010
2011         /* Register the os mca handler with SAL */
2012         if ((rc = ia64_sal_set_vectors(SAL_VECTOR_OS_MCA,
2013                                        ia64_mc_info.imi_mca_handler,
2014                                        ia64_tpa(mca_hldlr_ptr->gp),
2015                                        ia64_mc_info.imi_mca_handler_size,
2016                                        0, 0, 0)))
2017         {
2018                 printk(KERN_ERR "Failed to register OS MCA handler with SAL "
2019                        "(status %ld)\n", rc);
2020                 return;
2021         }
2022
2023         IA64_MCA_DEBUG("%s: registered OS MCA handler with SAL at 0x%lx, gp = 0x%lx\n", __func__,
2024                        ia64_mc_info.imi_mca_handler, ia64_tpa(mca_hldlr_ptr->gp));
2025
2026         /*
2027          * XXX - disable SAL checksum by setting size to 0, should be
2028          * size of the actual init handler in mca_asm.S.
2029          */
2030         ia64_mc_info.imi_monarch_init_handler           = ia64_tpa(init_hldlr_ptr_monarch->fp);
2031         ia64_mc_info.imi_monarch_init_handler_size      = 0;
2032         ia64_mc_info.imi_slave_init_handler             = ia64_tpa(init_hldlr_ptr_slave->fp);
2033         ia64_mc_info.imi_slave_init_handler_size        = 0;
2034
2035         IA64_MCA_DEBUG("%s: OS INIT handler at %lx\n", __func__,
2036                        ia64_mc_info.imi_monarch_init_handler);
2037
2038         /* Register the os init handler with SAL */
2039         if ((rc = ia64_sal_set_vectors(SAL_VECTOR_OS_INIT,
2040                                        ia64_mc_info.imi_monarch_init_handler,
2041                                        ia64_tpa(ia64_getreg(_IA64_REG_GP)),
2042                                        ia64_mc_info.imi_monarch_init_handler_size,
2043                                        ia64_mc_info.imi_slave_init_handler,
2044                                        ia64_tpa(ia64_getreg(_IA64_REG_GP)),
2045                                        ia64_mc_info.imi_slave_init_handler_size)))
2046         {
2047                 printk(KERN_ERR "Failed to register m/s INIT handlers with SAL "
2048                        "(status %ld)\n", rc);
2049                 return;
2050         }
2051         if (register_die_notifier(&default_init_monarch_nb)) {
2052                 printk(KERN_ERR "Failed to register default monarch INIT process\n");
2053                 return;
2054         }
2055
2056         IA64_MCA_DEBUG("%s: registered OS INIT handler with SAL\n", __func__);
2057
2058         /*
2059          *  Configure the CMCI/P vector and handler. Interrupts for CMC are
2060          *  per-processor, so AP CMC interrupts are setup in smp_callin() (smpboot.c).
2061          */
2062         register_percpu_irq(IA64_CMC_VECTOR, &cmci_irqaction);
2063         register_percpu_irq(IA64_CMCP_VECTOR, &cmcp_irqaction);
2064         ia64_mca_cmc_vector_setup();       /* Setup vector on BSP */
2065
2066         /* Setup the MCA rendezvous interrupt vector */
2067         register_percpu_irq(IA64_MCA_RENDEZ_VECTOR, &mca_rdzv_irqaction);
2068
2069         /* Setup the MCA wakeup interrupt vector */
2070         register_percpu_irq(IA64_MCA_WAKEUP_VECTOR, &mca_wkup_irqaction);
2071
2072 #ifdef CONFIG_ACPI
2073         /* Setup the CPEI/P handler */
2074         register_percpu_irq(IA64_CPEP_VECTOR, &mca_cpep_irqaction);
2075 #endif
2076
2077         /* Initialize the areas set aside by the OS to buffer the
2078          * platform/processor error states for MCA/INIT/CMC
2079          * handling.
2080          */
2081         ia64_log_init(SAL_INFO_TYPE_MCA);
2082         ia64_log_init(SAL_INFO_TYPE_INIT);
2083         ia64_log_init(SAL_INFO_TYPE_CMC);
2084         ia64_log_init(SAL_INFO_TYPE_CPE);
2085
2086         mca_init = 1;
2087         printk(KERN_INFO "MCA related initialization done\n");
2088 }
2089
2090 /*
2091  * ia64_mca_late_init
2092  *
2093  *      Opportunity to setup things that require initialization later
2094  *      than ia64_mca_init.  Setup a timer to poll for CPEs if the
2095  *      platform doesn't support an interrupt driven mechanism.
2096  *
2097  *  Inputs  :   None
2098  *  Outputs :   Status
2099  */
2100 static int __init
2101 ia64_mca_late_init(void)
2102 {
2103         if (!mca_init)
2104                 return 0;
2105
2106         register_hotcpu_notifier(&mca_cpu_notifier);
2107
2108         /* Setup the CMCI/P vector and handler */
2109         init_timer(&cmc_poll_timer);
2110         cmc_poll_timer.function = ia64_mca_cmc_poll;
2111
2112         /* Unmask/enable the vector */
2113         cmc_polling_enabled = 0;
2114         schedule_work(&cmc_enable_work);
2115
2116         IA64_MCA_DEBUG("%s: CMCI/P setup and enabled.\n", __func__);
2117
2118 #ifdef CONFIG_ACPI
2119         /* Setup the CPEI/P vector and handler */
2120         cpe_vector = acpi_request_vector(ACPI_INTERRUPT_CPEI);
2121         init_timer(&cpe_poll_timer);
2122         cpe_poll_timer.function = ia64_mca_cpe_poll;
2123
2124         {
2125                 struct irq_desc *desc;
2126                 unsigned int irq;
2127
2128                 if (cpe_vector >= 0) {
2129                         /* If platform supports CPEI, enable the irq. */
2130                         irq = local_vector_to_irq(cpe_vector);
2131                         if (irq > 0) {
2132                                 cpe_poll_enabled = 0;
2133                                 desc = irq_desc + irq;
2134                                 desc->status |= IRQ_PER_CPU;
2135                                 setup_irq(irq, &mca_cpe_irqaction);
2136                                 ia64_cpe_irq = irq;
2137                                 ia64_mca_register_cpev(cpe_vector);
2138                                 IA64_MCA_DEBUG("%s: CPEI/P setup and enabled.\n",
2139                                         __func__);
2140                                 return 0;
2141                         }
2142                         printk(KERN_ERR "%s: Failed to find irq for CPE "
2143                                         "interrupt handler, vector %d\n",
2144                                         __func__, cpe_vector);
2145                 }
2146                 /* If platform doesn't support CPEI, get the timer going. */
2147                 if (cpe_poll_enabled) {
2148                         ia64_mca_cpe_poll(0UL);
2149                         IA64_MCA_DEBUG("%s: CPEP setup and enabled.\n", __func__);
2150                 }
2151         }
2152 #endif
2153
2154         return 0;
2155 }
2156
2157 device_initcall(ia64_mca_late_init);