git.oblomov.eu Git - linux-2.6/blob - 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/types.h>
  73 #include <linux/init.h>
  74 #include <linux/sched.h>
  75 #include <linux/interrupt.h>
  76 #include <linux/irq.h>
  77 #include <linux/bootmem.h>
  78 #include <linux/acpi.h>
  79 #include <linux/timer.h>
  80 #include <linux/module.h>
  81 #include <linux/kernel.h>
  82 #include <linux/smp.h>
  83 #include <linux/workqueue.h>
  84 #include <linux/cpumask.h>
  85 #include <linux/kdebug.h>
  86 #include <linux/cpu.h>
  87
  88 #include <asm/delay.h>
  89 #include <asm/machvec.h>
  90 #include <asm/meminit.h>
  91 #include <asm/page.h>
  92 #include <asm/ptrace.h>
  93 #include <asm/system.h>
  94 #include <asm/sal.h>
  95 #include <asm/mca.h>
  96 #include <asm/kexec.h>
  97
  98 #include <asm/irq.h>
  99 #include <asm/hw_irq.h>
 100
 101 #include "mca_drv.h"
 102 #include "entry.h"
 103
 104 #if defined(IA64_MCA_DEBUG_INFO)
 105 # define IA64_MCA_DEBUG(fmt...) printk(fmt)
 106 #else
 107 # define IA64_MCA_DEBUG(fmt...)
 108 #endif
 109
 110 /* Used by mca_asm.S */
 111 DEFINE_PER_CPU(u64, ia64_mca_data); /* == __per_cpu_mca[smp_processor_id()] */
 112 DEFINE_PER_CPU(u64, ia64_mca_per_cpu_pte); /* PTE to map per-CPU area */
 113 DEFINE_PER_CPU(u64, ia64_mca_pal_pte);      /* PTE to map PAL code */
 114 DEFINE_PER_CPU(u64, ia64_mca_pal_base);    /* vaddr PAL code granule */
 115
 116 unsigned long __per_cpu_mca[NR_CPUS];
 117
 118 /* In mca_asm.S */
 119 extern void                     ia64_os_init_dispatch_monarch (void);
 120 extern void                     ia64_os_init_dispatch_slave (void);
 121
 122 static int monarch_cpu = -1;
 123
 124 static ia64_mc_info_t           ia64_mc_info;
 125
 126 #define MAX_CPE_POLL_INTERVAL (15*60*HZ) /* 15 minutes */
 127 #define MIN_CPE_POLL_INTERVAL (2*60*HZ)  /* 2 minutes */
 128 #define CMC_POLL_INTERVAL     (1*60*HZ)  /* 1 minute */
 129 #define CPE_HISTORY_LENGTH    5
 130 #define CMC_HISTORY_LENGTH    5
 131
 132 #ifdef CONFIG_ACPI
 133 static struct timer_list cpe_poll_timer;
 134 #endif
 135 static struct timer_list cmc_poll_timer;
 136 /*
 137  * This variable tells whether we are currently in polling mode.
 138  * Start with this in the wrong state so we won't play w/ timers
 139  * before the system is ready.
 140  */
 141 static int cmc_polling_enabled = 1;
 142
 143 /*
 144  * Clearing this variable prevents CPE polling from getting activated
 145  * in mca_late_init.  Use it if your system doesn't provide a CPEI,
 146  * but encounters problems retrieving CPE logs.  This should only be
 147  * necessary for debugging.
 148  */
 149 static int cpe_poll_enabled = 1;
 150
 151 extern void salinfo_log_wakeup(int type, u8 *buffer, u64 size, int irqsafe);
 152
 153 static int mca_init __initdata;
 154
 155 /*
 156  * limited & delayed printing support for MCA/INIT handler
 157  */
 158
 159 #define mprintk(fmt...) ia64_mca_printk(fmt)
 160
 161 #define MLOGBUF_SIZE (512+256*NR_CPUS)
 162 #define MLOGBUF_MSGMAX 256
 163 static char mlogbuf[MLOGBUF_SIZE];
 164 static DEFINE_SPINLOCK(mlogbuf_wlock);  /* mca context only */
 165 static DEFINE_SPINLOCK(mlogbuf_rlock);  /* normal context only */
 166 static unsigned long mlogbuf_start;
 167 static unsigned long mlogbuf_end;
 168 static unsigned int mlogbuf_finished = 0;
 169 static unsigned long mlogbuf_timestamp = 0;
 170
 171 static int loglevel_save = -1;
 172 #define BREAK_LOGLEVEL(__console_loglevel)              \
 173         oops_in_progress = 1;                           \
 174         if (loglevel_save < 0)                          \
 175                 loglevel_save = __console_loglevel;     \
 176         __console_loglevel = 15;
 177
 178 #define RESTORE_LOGLEVEL(__console_loglevel)            \
 179         if (loglevel_save >= 0) {                       \
 180                 __console_loglevel = loglevel_save;     \
 181                 loglevel_save = -1;                     \
 182         }                                               \
 183         mlogbuf_finished = 0;                           \
 184         oops_in_progress = 0;
 185
 186 /*
 187  * Push messages into buffer, print them later if not urgent.
 188  */
 189 void ia64_mca_printk(const char *fmt, ...)
 190 {
 191         va_list args;
 192         int printed_len;
 193         char temp_buf[MLOGBUF_MSGMAX];
 194         char *p;
 195
 196         va_start(args, fmt);
 197         printed_len = vscnprintf(temp_buf, sizeof(temp_buf), fmt, args);
 198         va_end(args);
 199
 200         /* Copy the output into mlogbuf */
 201         if (oops_in_progress) {
 202                 /* mlogbuf was abandoned, use printk directly instead. */
 203                 printk(temp_buf);
 204         } else {
 205                 spin_lock(&mlogbuf_wlock);
 206                 for (p = temp_buf; *p; p++) {
 207                         unsigned long next = (mlogbuf_end + 1) % MLOGBUF_SIZE;
 208                         if (next != mlogbuf_start) {
 209                                 mlogbuf[mlogbuf_end] = *p;
 210                                 mlogbuf_end = next;
 211                         } else {
 212                                 /* buffer full */
 213                                 break;
 214                         }
 215                 }
 216                 mlogbuf[mlogbuf_end] = '\0';
 217                 spin_unlock(&mlogbuf_wlock);
 218         }
 219 }
 220 EXPORT_SYMBOL(ia64_mca_printk);
 221
 222 /*
 223  * Print buffered messages.
 224  *  NOTE: call this after returning normal context. (ex. from salinfod)
 225  */
 226 void ia64_mlogbuf_dump(void)
 227 {
 228         char temp_buf[MLOGBUF_MSGMAX];
 229         char *p;
 230         unsigned long index;
 231         unsigned long flags;
 232         unsigned int printed_len;
 233
 234         /* Get output from mlogbuf */
 235         while (mlogbuf_start != mlogbuf_end) {
 236                 temp_buf[0] = '\0';
 237                 p = temp_buf;
 238                 printed_len = 0;
 239
 240                 spin_lock_irqsave(&mlogbuf_rlock, flags);
 241
 242                 index = mlogbuf_start;
 243                 while (index != mlogbuf_end) {
 244                         *p = mlogbuf[index];
 245                         index = (index + 1) % MLOGBUF_SIZE;
 246                         if (!*p)
 247                                 break;
 248                         p++;
 249                         if (++printed_len >= MLOGBUF_MSGMAX - 1)
 250                                 break;
 251                 }
 252                 *p = '\0';
 253                 if (temp_buf[0])
 254                         printk(temp_buf);
 255                 mlogbuf_start = index;
 256
 257                 mlogbuf_timestamp = 0;
 258                 spin_unlock_irqrestore(&mlogbuf_rlock, flags);
 259         }
 260 }
 261 EXPORT_SYMBOL(ia64_mlogbuf_dump);
 262
 263 /*
 264  * Call this if system is going to down or if immediate flushing messages to
 265  * console is required. (ex. recovery was failed, crash dump is going to be
 266  * invoked, long-wait rendezvous etc.)
 267  *  NOTE: this should be called from monarch.
 268  */
 269 static void ia64_mlogbuf_finish(int wait)
 270 {
 271         BREAK_LOGLEVEL(console_loglevel);
 272
 273         spin_lock_init(&mlogbuf_rlock);
 274         ia64_mlogbuf_dump();
 275         printk(KERN_EMERG "mlogbuf_finish: printing switched to urgent mode, "
 276                 "MCA/INIT might be dodgy or fail.\n");
 277
 278         if (!wait)
 279                 return;
 280
 281         /* wait for console */
 282         printk("Delaying for 5 seconds...\n");
 283         udelay(5*1000000);
 284
 285         mlogbuf_finished = 1;
 286 }
 287
 288 /*
 289  * Print buffered messages from INIT context.
 290  */
 291 static void ia64_mlogbuf_dump_from_init(void)
 292 {
 293         if (mlogbuf_finished)
 294                 return;
 295
 296         if (mlogbuf_timestamp && (mlogbuf_timestamp + 30*HZ > jiffies)) {
 297                 printk(KERN_ERR "INIT: mlogbuf_dump is interrupted by INIT "
 298                         " and the system seems to be messed up.\n");
 299                 ia64_mlogbuf_finish(0);
 300                 return;
 301         }
 302
 303         if (!spin_trylock(&mlogbuf_rlock)) {
 304                 printk(KERN_ERR "INIT: mlogbuf_dump is interrupted by INIT. "
 305                         "Generated messages other than stack dump will be "
 306                         "buffered to mlogbuf and will be printed later.\n");
 307                 printk(KERN_ERR "INIT: If messages would not printed after "
 308                         "this INIT, wait 30sec and assert INIT again.\n");
 309                 if (!mlogbuf_timestamp)
 310                         mlogbuf_timestamp = jiffies;
 311                 return;
 312         }
 313         spin_unlock(&mlogbuf_rlock);
 314         ia64_mlogbuf_dump();
 315 }
 316
 317 static void inline
 318 ia64_mca_spin(const char *func)
 319 {
 320         if (monarch_cpu == smp_processor_id())
 321                 ia64_mlogbuf_finish(0);
 322         mprintk(KERN_EMERG "%s: spinning here, not returning to SAL\n", func);
 323         while (1)
 324                 cpu_relax();
 325 }
 326 /*
 327  * IA64_MCA log support
 328  */
 329 #define IA64_MAX_LOGS           2       /* Double-buffering for nested MCAs */
 330 #define IA64_MAX_LOG_TYPES      4   /* MCA, INIT, CMC, CPE */
 331
 332 typedef struct ia64_state_log_s
 333 {
 334         spinlock_t      isl_lock;
 335         int             isl_index;
 336         unsigned long   isl_count;
 337         ia64_err_rec_t  *isl_log[IA64_MAX_LOGS]; /* need space to store header + error log */
 338 } ia64_state_log_t;
 339
 340 static ia64_state_log_t ia64_state_log[IA64_MAX_LOG_TYPES];
 341
 342 #define IA64_LOG_ALLOCATE(it, size) \
 343         {ia64_state_log[it].isl_log[IA64_LOG_CURR_INDEX(it)] = \
 344                 (ia64_err_rec_t *)alloc_bootmem(size); \
 345         ia64_state_log[it].isl_log[IA64_LOG_NEXT_INDEX(it)] = \
 346                 (ia64_err_rec_t *)alloc_bootmem(size);}
 347 #define IA64_LOG_LOCK_INIT(it) spin_lock_init(&ia64_state_log[it].isl_lock)
 348 #define IA64_LOG_LOCK(it)      spin_lock_irqsave(&ia64_state_log[it].isl_lock, s)
 349 #define IA64_LOG_UNLOCK(it)    spin_unlock_irqrestore(&ia64_state_log[it].isl_lock,s)
 350 #define IA64_LOG_NEXT_INDEX(it)    ia64_state_log[it].isl_index
 351 #define IA64_LOG_CURR_INDEX(it)    1 - ia64_state_log[it].isl_index
 352 #define IA64_LOG_INDEX_INC(it) \
 353     {ia64_state_log[it].isl_index = 1 - ia64_state_log[it].isl_index; \
 354     ia64_state_log[it].isl_count++;}
 355 #define IA64_LOG_INDEX_DEC(it) \
 356     ia64_state_log[it].isl_index = 1 - ia64_state_log[it].isl_index
 357 #define IA64_LOG_NEXT_BUFFER(it)   (void *)((ia64_state_log[it].isl_log[IA64_LOG_NEXT_INDEX(it)]))
 358 #define IA64_LOG_CURR_BUFFER(it)   (void *)((ia64_state_log[it].isl_log[IA64_LOG_CURR_INDEX(it)]))
 359 #define IA64_LOG_COUNT(it)         ia64_state_log[it].isl_count
 360
 361 /*
 362  * ia64_log_init
 363  *      Reset the OS ia64 log buffer
 364  * Inputs   :   info_type   (SAL_INFO_TYPE_{MCA,INIT,CMC,CPE})
 365  * Outputs      :       None
 366  */
 367 static void __init
 368 ia64_log_init(int sal_info_type)
 369 {
 370         u64     max_size = 0;
 371
 372         IA64_LOG_NEXT_INDEX(sal_info_type) = 0;
 373         IA64_LOG_LOCK_INIT(sal_info_type);
 374
 375         // SAL will tell us the maximum size of any error record of this type
 376         max_size = ia64_sal_get_state_info_size(sal_info_type);
 377         if (!max_size)
 378                 /* alloc_bootmem() doesn't like zero-sized allocations! */
 379                 return;
 380
 381         // set up OS data structures to hold error info
 382         IA64_LOG_ALLOCATE(sal_info_type, max_size);
 383         memset(IA64_LOG_CURR_BUFFER(sal_info_type), 0, max_size);
 384         memset(IA64_LOG_NEXT_BUFFER(sal_info_type), 0, max_size);
 385 }
 386
 387 /*
 388  * ia64_log_get
 389  *
 390  *      Get the current MCA log from SAL and copy it into the OS log buffer.
 391  *
 392  *  Inputs  :   info_type   (SAL_INFO_TYPE_{MCA,INIT,CMC,CPE})
 393  *              irq_safe    whether you can use printk at this point
 394  *  Outputs :   size        (total record length)
 395  *              *buffer     (ptr to error record)
 396  *
 397  */
 398 static u64
 399 ia64_log_get(int sal_info_type, u8 **buffer, int irq_safe)
 400 {
 401         sal_log_record_header_t     *log_buffer;
 402         u64                         total_len = 0;
 403         unsigned long               s;
 404
 405         IA64_LOG_LOCK(sal_info_type);
 406
 407         /* Get the process state information */
 408         log_buffer = IA64_LOG_NEXT_BUFFER(sal_info_type);
 409
 410         total_len = ia64_sal_get_state_info(sal_info_type, (u64 *)log_buffer);
 411
 412         if (total_len) {
 413                 IA64_LOG_INDEX_INC(sal_info_type);
 414                 IA64_LOG_UNLOCK(sal_info_type);
 415                 if (irq_safe) {
 416                         IA64_MCA_DEBUG("%s: SAL error record type %d retrieved. "
 417                                        "Record length = %ld\n", __FUNCTION__, sal_info_type, total_len);
 418                 }
 419                 *buffer = (u8 *) log_buffer;
 420                 return total_len;
 421         } else {
 422                 IA64_LOG_UNLOCK(sal_info_type);
 423                 return 0;
 424         }
 425 }
 426
 427 /*
 428  *  ia64_mca_log_sal_error_record
 429  *
 430  *  This function retrieves a specified error record type from SAL
 431  *  and wakes up any processes waiting for error records.
 432  *
 433  *  Inputs  :   sal_info_type   (Type of error record MCA/CMC/CPE)
 434  *              FIXME: remove MCA and irq_safe.
 435  */
 436 static void
 437 ia64_mca_log_sal_error_record(int sal_info_type)
 438 {
 439         u8 *buffer;
 440         sal_log_record_header_t *rh;
 441         u64 size;
 442         int irq_safe = sal_info_type != SAL_INFO_TYPE_MCA;
 443 #ifdef IA64_MCA_DEBUG_INFO
 444         static const char * const rec_name[] = { "MCA", "INIT", "CMC", "CPE" };
 445 #endif
 446
 447         size = ia64_log_get(sal_info_type, &buffer, irq_safe);
 448         if (!size)
 449                 return;
 450
 451         salinfo_log_wakeup(sal_info_type, buffer, size, irq_safe);
 452
 453         if (irq_safe)
 454                 IA64_MCA_DEBUG("CPU %d: SAL log contains %s error record\n",
 455                         smp_processor_id(),
 456                         sal_info_type < ARRAY_SIZE(rec_name) ? rec_name[sal_info_type] : "UNKNOWN");
 457
 458         /* Clear logs from corrected errors in case there's no user-level logger */
 459         rh = (sal_log_record_header_t *)buffer;
 460         if (rh->severity == sal_log_severity_corrected)
 461                 ia64_sal_clear_state_info(sal_info_type);
 462 }
 463
 464 /*
 465  * search_mca_table
 466  *  See if the MCA surfaced in an instruction range
 467  *  that has been tagged as recoverable.
 468  *
 469  *  Inputs
 470  *      first   First address range to check
 471  *      last    Last address range to check
 472  *      ip      Instruction pointer, address we are looking for
 473  *
 474  * Return value:
 475  *      1 on Success (in the table)/ 0 on Failure (not in the  table)
 476  */
 477 int
 478 search_mca_table (const struct mca_table_entry *first,
 479                 const struct mca_table_entry *last,
 480                 unsigned long ip)
 481 {
 482         const struct mca_table_entry *curr;
 483         u64 curr_start, curr_end;
 484
 485         curr = first;
 486         while (curr <= last) {
 487                 curr_start = (u64) &curr->start_addr + curr->start_addr;
 488                 curr_end = (u64) &curr->end_addr + curr->end_addr;
 489
 490                 if ((ip >= curr_start) && (ip <= curr_end)) {
 491                         return 1;
 492                 }
 493                 curr++;
 494         }
 495         return 0;
 496 }
 497
 498 /* Given an address, look for it in the mca tables. */
 499 int mca_recover_range(unsigned long addr)
 500 {
 501         extern struct mca_table_entry __start___mca_table[];
 502         extern struct mca_table_entry __stop___mca_table[];
 503
 504         return search_mca_table(__start___mca_table, __stop___mca_table-1, addr);
 505 }
 506 EXPORT_SYMBOL_GPL(mca_recover_range);
 507
 508 #ifdef CONFIG_ACPI
 509
 510 int cpe_vector = -1;
 511 int ia64_cpe_irq = -1;
 512
 513 static irqreturn_t
 514 ia64_mca_cpe_int_handler (int cpe_irq, void *arg)
 515 {
 516         static unsigned long    cpe_history[CPE_HISTORY_LENGTH];
 517         static int              index;
 518         static DEFINE_SPINLOCK(cpe_history_lock);
 519
 520         IA64_MCA_DEBUG("%s: received interrupt vector = %#x on CPU %d\n",
 521                        __FUNCTION__, cpe_irq, smp_processor_id());
 522
 523         /* SAL spec states this should run w/ interrupts enabled */
 524         local_irq_enable();
 525
 526         spin_lock(&cpe_history_lock);
 527         if (!cpe_poll_enabled && cpe_vector >= 0) {
 528
 529                 int i, count = 1; /* we know 1 happened now */
 530                 unsigned long now = jiffies;
 531
 532                 for (i = 0; i < CPE_HISTORY_LENGTH; i++) {
 533                         if (now - cpe_history[i] <= HZ)
 534                                 count++;
 535                 }
 536
 537                 IA64_MCA_DEBUG(KERN_INFO "CPE threshold %d/%d\n", count, CPE_HISTORY_LENGTH);
 538                 if (count >= CPE_HISTORY_LENGTH) {
 539
 540                         cpe_poll_enabled = 1;
 541                         spin_unlock(&cpe_history_lock);
 542                         disable_irq_nosync(local_vector_to_irq(IA64_CPE_VECTOR));
 543
 544                         /*
 545                          * Corrected errors will still be corrected, but
 546                          * make sure there's a log somewhere that indicates
 547                          * something is generating more than we can handle.
 548                          */
 549                         printk(KERN_WARNING "WARNING: Switching to polling CPE handler; error records may be lost\n");
 550
 551                         mod_timer(&cpe_poll_timer, jiffies + MIN_CPE_POLL_INTERVAL);
 552
 553                         /* lock already released, get out now */
 554                         goto out;
 555                 } else {
 556                         cpe_history[index++] = now;
 557                         if (index == CPE_HISTORY_LENGTH)
 558                                 index = 0;
 559                 }
 560         }
 561         spin_unlock(&cpe_history_lock);
 562 out:
 563         /* Get the CPE error record and log it */
 564         ia64_mca_log_sal_error_record(SAL_INFO_TYPE_CPE);
 565
 566         return IRQ_HANDLED;
 567 }
 568
 569 #endif /* CONFIG_ACPI */
 570
 571 #ifdef CONFIG_ACPI
 572 /*
 573  * ia64_mca_register_cpev
 574  *
 575  *  Register the corrected platform error vector with SAL.
 576  *
 577  *  Inputs
 578  *      cpev        Corrected Platform Error Vector number
 579  *
 580  *  Outputs
 581  *      None
 582  */
 583 void
 584 ia64_mca_register_cpev (int cpev)
 585 {
 586         /* Register the CPE interrupt vector with SAL */
 587         struct ia64_sal_retval isrv;
 588
 589         isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_CPE_INT, SAL_MC_PARAM_MECHANISM_INT, cpev, 0, 0);
 590         if (isrv.status) {
 591                 printk(KERN_ERR "Failed to register Corrected Platform "
 592                        "Error interrupt vector with SAL (status %ld)\n", isrv.status);
 593                 return;
 594         }
 595
 596         IA64_MCA_DEBUG("%s: corrected platform error "
 597                        "vector %#x registered\n", __FUNCTION__, cpev);
 598 }
 599 #endif /* CONFIG_ACPI */
 600
 601 /*
 602  * ia64_mca_cmc_vector_setup
 603  *
 604  *  Setup the corrected machine check vector register in the processor.
 605  *  (The interrupt is masked on boot. ia64_mca_late_init unmask this.)
 606  *  This function is invoked on a per-processor basis.
 607  *
 608  * Inputs
 609  *      None
 610  *
 611  * Outputs
 612  *      None
 613  */
 614 void __cpuinit
 615 ia64_mca_cmc_vector_setup (void)
 616 {
 617         cmcv_reg_t      cmcv;
 618
 619         cmcv.cmcv_regval        = 0;
 620         cmcv.cmcv_mask          = 1;        /* Mask/disable interrupt at first */
 621         cmcv.cmcv_vector        = IA64_CMC_VECTOR;
 622         ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
 623
 624         IA64_MCA_DEBUG("%s: CPU %d corrected "
 625                        "machine check vector %#x registered.\n",
 626                        __FUNCTION__, smp_processor_id(), IA64_CMC_VECTOR);
 627
 628         IA64_MCA_DEBUG("%s: CPU %d CMCV = %#016lx\n",
 629                        __FUNCTION__, smp_processor_id(), ia64_getreg(_IA64_REG_CR_CMCV));
 630 }
 631
 632 /*
 633  * ia64_mca_cmc_vector_disable
 634  *
 635  *  Mask the corrected machine check vector register in the processor.
 636  *  This function is invoked on a per-processor basis.
 637  *
 638  * Inputs
 639  *      dummy(unused)
 640  *
 641  * Outputs
 642  *      None
 643  */
 644 static void
 645 ia64_mca_cmc_vector_disable (void *dummy)
 646 {
 647         cmcv_reg_t      cmcv;
 648
 649         cmcv.cmcv_regval = ia64_getreg(_IA64_REG_CR_CMCV);
 650
 651         cmcv.cmcv_mask = 1; /* Mask/disable interrupt */
 652         ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
 653
 654         IA64_MCA_DEBUG("%s: CPU %d corrected "
 655                        "machine check vector %#x disabled.\n",
 656                        __FUNCTION__, smp_processor_id(), cmcv.cmcv_vector);
 657 }
 658
 659 /*
 660  * ia64_mca_cmc_vector_enable
 661  *
 662  *  Unmask the corrected machine check vector register in the processor.
 663  *  This function is invoked on a per-processor basis.
 664  *
 665  * Inputs
 666  *      dummy(unused)
 667  *
 668  * Outputs
 669  *      None
 670  */
 671 static void
 672 ia64_mca_cmc_vector_enable (void *dummy)
 673 {
 674         cmcv_reg_t      cmcv;
 675
 676         cmcv.cmcv_regval = ia64_getreg(_IA64_REG_CR_CMCV);
 677
 678         cmcv.cmcv_mask = 0; /* Unmask/enable interrupt */
 679         ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
 680
 681         IA64_MCA_DEBUG("%s: CPU %d corrected "
 682                        "machine check vector %#x enabled.\n",
 683                        __FUNCTION__, smp_processor_id(), cmcv.cmcv_vector);
 684 }
 685
 686 /*
 687  * ia64_mca_cmc_vector_disable_keventd
 688  *
 689  * Called via keventd (smp_call_function() is not safe in interrupt context) to
 690  * disable the cmc interrupt vector.
 691  */
 692 static void
 693 ia64_mca_cmc_vector_disable_keventd(struct work_struct *unused)
 694 {
 695         on_each_cpu(ia64_mca_cmc_vector_disable, NULL, 1, 0);
 696 }
 697
 698 /*
 699  * ia64_mca_cmc_vector_enable_keventd
 700  *
 701  * Called via keventd (smp_call_function() is not safe in interrupt context) to
 702  * enable the cmc interrupt vector.
 703  */
 704 static void
 705 ia64_mca_cmc_vector_enable_keventd(struct work_struct *unused)
 706 {
 707         on_each_cpu(ia64_mca_cmc_vector_enable, NULL, 1, 0);
 708 }
 709
 710 /*
 711  * ia64_mca_wakeup
 712  *
 713  *      Send an inter-cpu interrupt to wake-up a particular cpu.
 714  *
 715  *  Inputs  :   cpuid
 716  *  Outputs :   None
 717  */
 718 static void
 719 ia64_mca_wakeup(int cpu)
 720 {
 721         platform_send_ipi(cpu, IA64_MCA_WAKEUP_VECTOR, IA64_IPI_DM_INT, 0);
 722 }
 723
 724 /*
 725  * ia64_mca_wakeup_all
 726  *
 727  *      Wakeup all the slave cpus which have rendez'ed previously.
 728  *
 729  *  Inputs  :   None
 730  *  Outputs :   None
 731  */
 732 static void
 733 ia64_mca_wakeup_all(void)
 734 {
 735         int cpu;
 736
 737         /* Clear the Rendez checkin flag for all cpus */
 738         for_each_online_cpu(cpu) {
 739                 if (ia64_mc_info.imi_rendez_checkin[cpu] == IA64_MCA_RENDEZ_CHECKIN_DONE)
 740                         ia64_mca_wakeup(cpu);
 741         }
 742
 743 }
 744
 745 /*
 746  * ia64_mca_rendez_interrupt_handler
 747  *
 748  *      This is handler used to put slave processors into spinloop
 749  *      while the monarch processor does the mca handling and later
 750  *      wake each slave up once the monarch is done.  The state
 751  *      IA64_MCA_RENDEZ_CHECKIN_DONE indicates the cpu is rendez'ed
 752  *      in SAL.  The state IA64_MCA_RENDEZ_CHECKIN_NOTDONE indicates
 753  *      the cpu has come out of OS rendezvous.
 754  *
 755  *  Inputs  :   None
 756  *  Outputs :   None
 757  */
 758 static irqreturn_t
 759 ia64_mca_rendez_int_handler(int rendez_irq, void *arg)
 760 {
 761         unsigned long flags;
 762         int cpu = smp_processor_id();
 763         struct ia64_mca_notify_die nd =
 764                 { .sos = NULL, .monarch_cpu = &monarch_cpu };
 765
 766         /* Mask all interrupts */
 767         local_irq_save(flags);
 768         if (notify_die(DIE_MCA_RENDZVOUS_ENTER, "MCA", get_irq_regs(),
 769                        (long)&nd, 0, 0) == NOTIFY_STOP)
 770                 ia64_mca_spin(__FUNCTION__);
 771
 772         ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_DONE;
 773         /* Register with the SAL monarch that the slave has
 774          * reached SAL
 775          */
 776         ia64_sal_mc_rendez();
 777
 778         if (notify_die(DIE_MCA_RENDZVOUS_PROCESS, "MCA", get_irq_regs(),
 779                        (long)&nd, 0, 0) == NOTIFY_STOP)
 780                 ia64_mca_spin(__FUNCTION__);
 781
 782         /* Wait for the monarch cpu to exit. */
 783         while (monarch_cpu != -1)
 784                cpu_relax();     /* spin until monarch leaves */
 785
 786         if (notify_die(DIE_MCA_RENDZVOUS_LEAVE, "MCA", get_irq_regs(),
 787                        (long)&nd, 0, 0) == NOTIFY_STOP)
 788                 ia64_mca_spin(__FUNCTION__);
 789
 790         ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
 791         /* Enable all interrupts */
 792         local_irq_restore(flags);
 793         return IRQ_HANDLED;
 794 }
 795
 796 /*
 797  * ia64_mca_wakeup_int_handler
 798  *
 799  *      The interrupt handler for processing the inter-cpu interrupt to the
 800  *      slave cpu which was spinning in the rendez loop.
 801  *      Since this spinning is done by turning off the interrupts and
 802  *      polling on the wakeup-interrupt bit in the IRR, there is
 803  *      nothing useful to be done in the handler.
 804  *
 805  *  Inputs  :   wakeup_irq  (Wakeup-interrupt bit)
 806  *      arg             (Interrupt handler specific argument)
 807  *  Outputs :   None
 808  *
 809  */
 810 static irqreturn_t
 811 ia64_mca_wakeup_int_handler(int wakeup_irq, void *arg)
 812 {
 813         return IRQ_HANDLED;
 814 }
 815
 816 /* Function pointer for extra MCA recovery */
 817 int (*ia64_mca_ucmc_extension)
 818         (void*,struct ia64_sal_os_state*)
 819         = NULL;
 820
 821 int
 822 ia64_reg_MCA_extension(int (*fn)(void *, struct ia64_sal_os_state *))
 823 {
 824         if (ia64_mca_ucmc_extension)
 825                 return 1;
 826
 827         ia64_mca_ucmc_extension = fn;
 828         return 0;
 829 }
 830
 831 void
 832 ia64_unreg_MCA_extension(void)
 833 {
 834         if (ia64_mca_ucmc_extension)
 835                 ia64_mca_ucmc_extension = NULL;
 836 }
 837
 838 EXPORT_SYMBOL(ia64_reg_MCA_extension);
 839 EXPORT_SYMBOL(ia64_unreg_MCA_extension);
 840
 841
 842 static inline void
 843 copy_reg(const u64 *fr, u64 fnat, u64 *tr, u64 *tnat)
 844 {
 845         u64 fslot, tslot, nat;
 846         *tr = *fr;
 847         fslot = ((unsigned long)fr >> 3) & 63;
 848         tslot = ((unsigned long)tr >> 3) & 63;
 849         *tnat &= ~(1UL << tslot);
 850         nat = (fnat >> fslot) & 1;
 851         *tnat |= (nat << tslot);
 852 }
 853
 854 /* Change the comm field on the MCA/INT task to include the pid that
 855  * was interrupted, it makes for easier debugging.  If that pid was 0
 856  * (swapper or nested MCA/INIT) then use the start of the previous comm
 857  * field suffixed with its cpu.
 858  */
 859
 860 static void
 861 ia64_mca_modify_comm(const struct task_struct *previous_current)
 862 {
 863         char *p, comm[sizeof(current->comm)];
 864         if (previous_current->pid)
 865                 snprintf(comm, sizeof(comm), "%s %d",
 866                         current->comm, previous_current->pid);
 867         else {
 868                 int l;
 869                 if ((p = strchr(previous_current->comm, ' ')))
 870                         l = p - previous_current->comm;
 871                 else
 872                         l = strlen(previous_current->comm);
 873                 snprintf(comm, sizeof(comm), "%s %*s %d",
 874                         current->comm, l, previous_current->comm,
 875                         task_thread_info(previous_current)->cpu);
 876         }
 877         memcpy(current->comm, comm, sizeof(current->comm));
 878 }
 879
 880 /* On entry to this routine, we are running on the per cpu stack, see
 881  * mca_asm.h.  The original stack has not been touched by this event.  Some of
 882  * the original stack's registers will be in the RBS on this stack.  This stack
 883  * also contains a partial pt_regs and switch_stack, the rest of the data is in
 884  * PAL minstate.
 885  *
 886  * The first thing to do is modify the original stack to look like a blocked
 887  * task so we can run backtrace on the original task.  Also mark the per cpu
 888  * stack as current to ensure that we use the correct task state, it also means
 889  * that we can do backtrace on the MCA/INIT handler code itself.
 890  */
 891
 892 static struct task_struct *
 893 ia64_mca_modify_original_stack(struct pt_regs *regs,
 894                 const struct switch_stack *sw,
 895                 struct ia64_sal_os_state *sos,
 896                 const char *type)
 897 {
 898         char *p;
 899         ia64_va va;
 900         extern char ia64_leave_kernel[];        /* Need asm address, not function descriptor */
 901         const pal_min_state_area_t *ms = sos->pal_min_state;
 902         struct task_struct *previous_current;
 903         struct pt_regs *old_regs;
 904         struct switch_stack *old_sw;
 905         unsigned size = sizeof(struct pt_regs) +
 906                         sizeof(struct switch_stack) + 16;
 907         u64 *old_bspstore, *old_bsp;
 908         u64 *new_bspstore, *new_bsp;
 909         u64 old_unat, old_rnat, new_rnat, nat;
 910         u64 slots, loadrs = regs->loadrs;
 911         u64 r12 = ms->pmsa_gr[12-1], r13 = ms->pmsa_gr[13-1];
 912         u64 ar_bspstore = regs->ar_bspstore;
 913         u64 ar_bsp = regs->ar_bspstore + (loadrs >> 16);
 914         const u64 *bank;
 915         const char *msg;
 916         int cpu = smp_processor_id();
 917
 918         previous_current = curr_task(cpu);
 919         set_curr_task(cpu, current);
 920         if ((p = strchr(current->comm, ' ')))
 921                 *p = '\0';
 922
 923         /* Best effort attempt to cope with MCA/INIT delivered while in
 924          * physical mode.
 925          */
 926         regs->cr_ipsr = ms->pmsa_ipsr;
 927         if (ia64_psr(regs)->dt == 0) {
 928                 va.l = r12;
 929                 if (va.f.reg == 0) {
 930                         va.f.reg = 7;
 931                         r12 = va.l;
 932                 }
 933                 va.l = r13;
 934                 if (va.f.reg == 0) {
 935                         va.f.reg = 7;
 936                         r13 = va.l;
 937                 }
 938         }
 939         if (ia64_psr(regs)->rt == 0) {
 940                 va.l = ar_bspstore;
 941                 if (va.f.reg == 0) {
 942                         va.f.reg = 7;
 943                         ar_bspstore = va.l;
 944                 }
 945                 va.l = ar_bsp;
 946                 if (va.f.reg == 0) {
 947                         va.f.reg = 7;
 948                         ar_bsp = va.l;
 949                 }
 950         }
 951
 952         /* mca_asm.S ia64_old_stack() cannot assume that the dirty registers
 953          * have been copied to the old stack, the old stack may fail the
 954          * validation tests below.  So ia64_old_stack() must restore the dirty
 955          * registers from the new stack.  The old and new bspstore probably
 956          * have different alignments, so loadrs calculated on the old bsp
 957          * cannot be used to restore from the new bsp.  Calculate a suitable
 958          * loadrs for the new stack and save it in the new pt_regs, where
 959          * ia64_old_stack() can get it.
 960          */
 961         old_bspstore = (u64 *)ar_bspstore;
 962         old_bsp = (u64 *)ar_bsp;
 963         slots = ia64_rse_num_regs(old_bspstore, old_bsp);
 964         new_bspstore = (u64 *)((u64)current + IA64_RBS_OFFSET);
 965         new_bsp = ia64_rse_skip_regs(new_bspstore, slots);
 966         regs->loadrs = (new_bsp - new_bspstore) * 8 << 16;
 967
 968         /* Verify the previous stack state before we change it */
 969         if (user_mode(regs)) {
 970                 msg = "occurred in user space";
 971                 /* previous_current is guaranteed to be valid when the task was
 972                  * in user space, so ...
 973                  */
 974                 ia64_mca_modify_comm(previous_current);
 975                 goto no_mod;
 976         }
 977
 978         if (r13 != sos->prev_IA64_KR_CURRENT) {
 979                 msg = "inconsistent previous current and r13";
 980                 goto no_mod;
 981         }
 982
 983         if (!mca_recover_range(ms->pmsa_iip)) {
 984                 if ((r12 - r13) >= KERNEL_STACK_SIZE) {
 985                         msg = "inconsistent r12 and r13";
 986                         goto no_mod;
 987                 }
 988                 if ((ar_bspstore - r13) >= KERNEL_STACK_SIZE) {
 989                         msg = "inconsistent ar.bspstore and r13";
 990                         goto no_mod;
 991                 }
 992                 va.p = old_bspstore;
 993                 if (va.f.reg < 5) {
 994                         msg = "old_bspstore is in the wrong region";
 995                         goto no_mod;
 996                 }
 997                 if ((ar_bsp - r13) >= KERNEL_STACK_SIZE) {
 998                         msg = "inconsistent ar.bsp and r13";
 999                         goto no_mod;
1000                 }
1001                 size += (ia64_rse_skip_regs(old_bspstore, slots) - old_bspstore) * 8;
1002                 if (ar_bspstore + size > r12) {
1003                         msg = "no room for blocked state";
1004                         goto no_mod;
1005                 }
1006         }
1007
1008         ia64_mca_modify_comm(previous_current);
1009
1010         /* Make the original task look blocked.  First stack a struct pt_regs,
1011          * describing the state at the time of interrupt.  mca_asm.S built a
1012          * partial pt_regs, copy it and fill in the blanks using minstate.
1013          */
1014         p = (char *)r12 - sizeof(*regs);
1015         old_regs = (struct pt_regs *)p;
1016         memcpy(old_regs, regs, sizeof(*regs));
1017         /* If ipsr.ic then use pmsa_{iip,ipsr,ifs}, else use
1018          * pmsa_{xip,xpsr,xfs}
1019          */
1020         if (ia64_psr(regs)->ic) {
1021                 old_regs->cr_iip = ms->pmsa_iip;
1022                 old_regs->cr_ipsr = ms->pmsa_ipsr;
1023                 old_regs->cr_ifs = ms->pmsa_ifs;
1024         } else {
1025                 old_regs->cr_iip = ms->pmsa_xip;
1026                 old_regs->cr_ipsr = ms->pmsa_xpsr;
1027                 old_regs->cr_ifs = ms->pmsa_xfs;
1028         }
1029         old_regs->pr = ms->pmsa_pr;
1030         old_regs->b0 = ms->pmsa_br0;
1031         old_regs->loadrs = loadrs;
1032         old_regs->ar_rsc = ms->pmsa_rsc;
1033         old_unat = old_regs->ar_unat;
1034         copy_reg(&ms->pmsa_gr[1-1], ms->pmsa_nat_bits, &old_regs->r1, &old_unat);
1035         copy_reg(&ms->pmsa_gr[2-1], ms->pmsa_nat_bits, &old_regs->r2, &old_unat);
1036         copy_reg(&ms->pmsa_gr[3-1], ms->pmsa_nat_bits, &old_regs->r3, &old_unat);
1037         copy_reg(&ms->pmsa_gr[8-1], ms->pmsa_nat_bits, &old_regs->r8, &old_unat);
1038         copy_reg(&ms->pmsa_gr[9-1], ms->pmsa_nat_bits, &old_regs->r9, &old_unat);
1039         copy_reg(&ms->pmsa_gr[10-1], ms->pmsa_nat_bits, &old_regs->r10, &old_unat);
1040         copy_reg(&ms->pmsa_gr[11-1], ms->pmsa_nat_bits, &old_regs->r11, &old_unat);
1041         copy_reg(&ms->pmsa_gr[12-1], ms->pmsa_nat_bits, &old_regs->r12, &old_unat);
1042         copy_reg(&ms->pmsa_gr[13-1], ms->pmsa_nat_bits, &old_regs->r13, &old_unat);
1043         copy_reg(&ms->pmsa_gr[14-1], ms->pmsa_nat_bits, &old_regs->r14, &old_unat);
1044         copy_reg(&ms->pmsa_gr[15-1], ms->pmsa_nat_bits, &old_regs->r15, &old_unat);
1045         if (ia64_psr(old_regs)->bn)
1046                 bank = ms->pmsa_bank1_gr;
1047         else
1048                 bank = ms->pmsa_bank0_gr;
1049         copy_reg(&bank[16-16], ms->pmsa_nat_bits, &old_regs->r16, &old_unat);
1050         copy_reg(&bank[17-16], ms->pmsa_nat_bits, &old_regs->r17, &old_unat);
1051         copy_reg(&bank[18-16], ms->pmsa_nat_bits, &old_regs->r18, &old_unat);
1052         copy_reg(&bank[19-16], ms->pmsa_nat_bits, &old_regs->r19, &old_unat);
1053         copy_reg(&bank[20-16], ms->pmsa_nat_bits, &old_regs->r20, &old_unat);
1054         copy_reg(&bank[21-16], ms->pmsa_nat_bits, &old_regs->r21, &old_unat);
1055         copy_reg(&bank[22-16], ms->pmsa_nat_bits, &old_regs->r22, &old_unat);
1056         copy_reg(&bank[23-16], ms->pmsa_nat_bits, &old_regs->r23, &old_unat);
1057         copy_reg(&bank[24-16], ms->pmsa_nat_bits, &old_regs->r24, &old_unat);
1058         copy_reg(&bank[25-16], ms->pmsa_nat_bits, &old_regs->r25, &old_unat);
1059         copy_reg(&bank[26-16], ms->pmsa_nat_bits, &old_regs->r26, &old_unat);
1060         copy_reg(&bank[27-16], ms->pmsa_nat_bits, &old_regs->r27, &old_unat);
1061         copy_reg(&bank[28-16], ms->pmsa_nat_bits, &old_regs->r28, &old_unat);
1062         copy_reg(&bank[29-16], ms->pmsa_nat_bits, &old_regs->r29, &old_unat);
1063         copy_reg(&bank[30-16], ms->pmsa_nat_bits, &old_regs->r30, &old_unat);
1064         copy_reg(&bank[31-16], ms->pmsa_nat_bits, &old_regs->r31, &old_unat);
1065
1066         /* Next stack a struct switch_stack.  mca_asm.S built a partial
1067          * switch_stack, copy it and fill in the blanks using pt_regs and
1068          * minstate.
1069          *
1070          * In the synthesized switch_stack, b0 points to ia64_leave_kernel,
1071          * ar.pfs is set to 0.
1072          *
1073          * unwind.c::unw_unwind() does special processing for interrupt frames.
1074          * It checks if the PRED_NON_SYSCALL predicate is set, if the predicate
1075          * is clear then unw_unwind() does _not_ adjust bsp over pt_regs.  Not
1076          * that this is documented, of course.  Set PRED_NON_SYSCALL in the
1077          * switch_stack on the original stack so it will unwind correctly when
1078          * unwind.c reads pt_regs.
1079          *
1080          * thread.ksp is updated to point to the synthesized switch_stack.
1081          */
1082         p -= sizeof(struct switch_stack);
1083         old_sw = (struct switch_stack *)p;
1084         memcpy(old_sw, sw, sizeof(*sw));
1085         old_sw->caller_unat = old_unat;
1086         old_sw->ar_fpsr = old_regs->ar_fpsr;
1087         copy_reg(&ms->pmsa_gr[4-1], ms->pmsa_nat_bits, &old_sw->r4, &old_unat);
1088         copy_reg(&ms->pmsa_gr[5-1], ms->pmsa_nat_bits, &old_sw->r5, &old_unat);
1089         copy_reg(&ms->pmsa_gr[6-1], ms->pmsa_nat_bits, &old_sw->r6, &old_unat);
1090         copy_reg(&ms->pmsa_gr[7-1], ms->pmsa_nat_bits, &old_sw->r7, &old_unat);
1091         old_sw->b0 = (u64)ia64_leave_kernel;
1092         old_sw->b1 = ms->pmsa_br1;
1093         old_sw->ar_pfs = 0;
1094         old_sw->ar_unat = old_unat;
1095         old_sw->pr = old_regs->pr | (1UL << PRED_NON_SYSCALL);
1096         previous_current->thread.ksp = (u64)p - 16;
1097
1098         /* Finally copy the original stack's registers back to its RBS.
1099          * Registers from ar.bspstore through ar.bsp at the time of the event
1100          * are in the current RBS, copy them back to the original stack.  The
1101          * copy must be done register by register because the original bspstore
1102          * and the current one have different alignments, so the saved RNAT
1103          * data occurs at different places.
1104          *
1105          * mca_asm does cover, so the old_bsp already includes all registers at
1106          * the time of MCA/INIT.  It also does flushrs, so all registers before
1107          * this function have been written to backing store on the MCA/INIT
1108          * stack.
1109          */
1110         new_rnat = ia64_get_rnat(ia64_rse_rnat_addr(new_bspstore));
1111         old_rnat = regs->ar_rnat;
1112         while (slots--) {
1113                 if (ia64_rse_is_rnat_slot(new_bspstore)) {
1114                         new_rnat = ia64_get_rnat(new_bspstore++);
1115                 }
1116                 if (ia64_rse_is_rnat_slot(old_bspstore)) {
1117                         *old_bspstore++ = old_rnat;
1118                         old_rnat = 0;
1119                 }
1120                 nat = (new_rnat >> ia64_rse_slot_num(new_bspstore)) & 1UL;
1121                 old_rnat &= ~(1UL << ia64_rse_slot_num(old_bspstore));
1122                 old_rnat |= (nat << ia64_rse_slot_num(old_bspstore));
1123                 *old_bspstore++ = *new_bspstore++;
1124         }
1125         old_sw->ar_bspstore = (unsigned long)old_bspstore;
1126         old_sw->ar_rnat = old_rnat;
1127
1128         sos->prev_task = previous_current;
1129         return previous_current;
1130
1131 no_mod:
1132         printk(KERN_INFO "cpu %d, %s %s, original stack not modified\n",
1133                         smp_processor_id(), type, msg);
1134         return previous_current;
1135 }
1136
1137 /* The monarch/slave interaction is based on monarch_cpu and requires that all
1138  * slaves have entered rendezvous before the monarch leaves.  If any cpu has
1139  * not entered rendezvous yet then wait a bit.  The assumption is that any
1140  * slave that has not rendezvoused after a reasonable time is never going to do
1141  * so.  In this context, slave includes cpus that respond to the MCA rendezvous
1142  * interrupt, as well as cpus that receive the INIT slave event.
1143  */
1144
1145 static void
1146 ia64_wait_for_slaves(int monarch, const char *type)
1147 {
1148         int c, i , wait;
1149
1150         /*
1151          * wait 5 seconds total for slaves (arbitrary)
1152          */
1153         for (i = 0; i < 5000; i++) {
1154                 wait = 0;
1155                 for_each_online_cpu(c) {
1156                         if (c == monarch)
1157                                 continue;
1158                         if (ia64_mc_info.imi_rendez_checkin[c]
1159                                         == IA64_MCA_RENDEZ_CHECKIN_NOTDONE) {
1160                                 udelay(1000);           /* short wait */
1161                                 wait = 1;
1162                                 break;
1163                         }
1164                 }
1165                 if (!wait)
1166                         goto all_in;
1167         }
1168
1169         /*
1170          * Maybe slave(s) dead. Print buffered messages immediately.
1171          */
1172         ia64_mlogbuf_finish(0);
1173         mprintk(KERN_INFO "OS %s slave did not rendezvous on cpu", type);
1174         for_each_online_cpu(c) {
1175                 if (c == monarch)
1176                         continue;
1177                 if (ia64_mc_info.imi_rendez_checkin[c] == IA64_MCA_RENDEZ_CHECKIN_NOTDONE)
1178                         mprintk(" %d", c);
1179         }
1180         mprintk("\n");
1181         return;
1182
1183 all_in:
1184         mprintk(KERN_INFO "All OS %s slaves have reached rendezvous\n", type);
1185         return;
1186 }
1187
1188 /*
1189  * ia64_mca_handler
1190  *
1191  *      This is uncorrectable machine check handler called from OS_MCA
1192  *      dispatch code which is in turn called from SAL_CHECK().
1193  *      This is the place where the core of OS MCA handling is done.
1194  *      Right now the logs are extracted and displayed in a well-defined
1195  *      format. This handler code is supposed to be run only on the
1196  *      monarch processor. Once the monarch is done with MCA handling
1197  *      further MCA logging is enabled by clearing logs.
1198  *      Monarch also has the duty of sending wakeup-IPIs to pull the
1199  *      slave processors out of rendezvous spinloop.
1200  *
1201  *      If multiple processors call into OS_MCA, the first will become
1202  *      the monarch.  Subsequent cpus will be recorded in the mca_cpu
1203  *      bitmask.  After the first monarch has processed its MCA, it
1204  *      will wake up the next cpu in the mca_cpu bitmask and then go
1205  *      into the rendezvous loop.  When all processors have serviced
1206  *      their MCA, the last monarch frees up the rest of the processors.
1207  */
1208 void
1209 ia64_mca_handler(struct pt_regs *regs, struct switch_stack *sw,
1210                  struct ia64_sal_os_state *sos)
1211 {
1212         int recover, cpu = smp_processor_id();
1213         struct task_struct *previous_current;
1214         struct ia64_mca_notify_die nd =
1215                 { .sos = sos, .monarch_cpu = &monarch_cpu };
1216         static atomic_t mca_count;
1217         static cpumask_t mca_cpu;
1218
1219         if (atomic_add_return(1, &mca_count) == 1) {
1220                 monarch_cpu = cpu;
1221                 sos->monarch = 1;
1222         } else {
1223                 cpu_set(cpu, mca_cpu);
1224                 sos->monarch = 0;
1225         }
1226         mprintk(KERN_INFO "Entered OS MCA handler. PSP=%lx cpu=%d "
1227                 "monarch=%ld\n", sos->proc_state_param, cpu, sos->monarch);
1228
1229         previous_current = ia64_mca_modify_original_stack(regs, sw, sos, "MCA");
1230
1231         if (notify_die(DIE_MCA_MONARCH_ENTER, "MCA", regs, (long)&nd, 0, 0)
1232                         == NOTIFY_STOP)
1233                 ia64_mca_spin(__FUNCTION__);
1234
1235         ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_CONCURRENT_MCA;
1236         if (sos->monarch) {
1237                 ia64_wait_for_slaves(cpu, "MCA");
1238
1239                 /* Wakeup all the processors which are spinning in the
1240                  * rendezvous loop.  They will leave SAL, then spin in the OS
1241                  * with interrupts disabled until this monarch cpu leaves the
1242                  * MCA handler.  That gets control back to the OS so we can
1243                  * backtrace the other cpus, backtrace when spinning in SAL
1244                  * does not work.
1245                  */
1246                 ia64_mca_wakeup_all();
1247                 if (notify_die(DIE_MCA_MONARCH_PROCESS, "MCA", regs, (long)&nd, 0, 0)
1248                                 == NOTIFY_STOP)
1249                         ia64_mca_spin(__FUNCTION__);
1250         } else {
1251                 while (cpu_isset(cpu, mca_cpu))
1252                         cpu_relax();    /* spin until monarch wakes us */
1253         }
1254
1255         /* Get the MCA error record and log it */
1256         ia64_mca_log_sal_error_record(SAL_INFO_TYPE_MCA);
1257
1258         /* MCA error recovery */
1259         recover = (ia64_mca_ucmc_extension
1260                 && ia64_mca_ucmc_extension(
1261                         IA64_LOG_CURR_BUFFER(SAL_INFO_TYPE_MCA),
1262                         sos));
1263
1264         if (recover) {
1265                 sal_log_record_header_t *rh = IA64_LOG_CURR_BUFFER(SAL_INFO_TYPE_MCA);
1266                 rh->severity = sal_log_severity_corrected;
1267                 ia64_sal_clear_state_info(SAL_INFO_TYPE_MCA);
1268                 sos->os_status = IA64_MCA_CORRECTED;
1269         } else {
1270                 /* Dump buffered message to console */
1271                 ia64_mlogbuf_finish(1);
1272 #ifdef CONFIG_KEXEC
1273                 atomic_set(&kdump_in_progress, 1);
1274                 monarch_cpu = -1;
1275 #endif
1276         }
1277         if (notify_die(DIE_MCA_MONARCH_LEAVE, "MCA", regs, (long)&nd, 0, recover)
1278                         == NOTIFY_STOP)
1279                 ia64_mca_spin(__FUNCTION__);
1280
1281
1282         if (atomic_dec_return(&mca_count) > 0) {
1283                 int i;
1284
1285                 /* wake up the next monarch cpu,
1286                  * and put this cpu in the rendez loop.
1287                  */
1288                 for_each_online_cpu(i) {
1289                         if (cpu_isset(i, mca_cpu)) {
1290                                 monarch_cpu = i;
1291                                 cpu_clear(i, mca_cpu);  /* wake next cpu */
1292                                 while (monarch_cpu != -1)
1293                                         cpu_relax();    /* spin until last cpu leaves */
1294                                 set_curr_task(cpu, previous_current);
1295                                 ia64_mc_info.imi_rendez_checkin[cpu]
1296                                                 = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1297                                 return;
1298                         }
1299                 }
1300         }
1301         set_curr_task(cpu, previous_current);
1302         ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1303         monarch_cpu = -1;       /* This frees the slaves and previous monarchs */
1304 }
1305
1306 static DECLARE_WORK(cmc_disable_work, ia64_mca_cmc_vector_disable_keventd);
1307 static DECLARE_WORK(cmc_enable_work, ia64_mca_cmc_vector_enable_keventd);
1308
1309 /*
1310  * ia64_mca_cmc_int_handler
1311  *
1312  *  This is corrected machine check interrupt handler.
1313  *      Right now the logs are extracted and displayed in a well-defined
1314  *      format.
1315  *
1316  * Inputs
1317  *      interrupt number
1318  *      client data arg ptr
1319  *
1320  * Outputs
1321  *      None
1322  */
1323 static irqreturn_t
1324 ia64_mca_cmc_int_handler(int cmc_irq, void *arg)
1325 {
1326         static unsigned long    cmc_history[CMC_HISTORY_LENGTH];
1327         static int              index;
1328         static DEFINE_SPINLOCK(cmc_history_lock);
1329
1330         IA64_MCA_DEBUG("%s: received interrupt vector = %#x on CPU %d\n",
1331                        __FUNCTION__, cmc_irq, smp_processor_id());
1332
1333         /* SAL spec states this should run w/ interrupts enabled */
1334         local_irq_enable();
1335
1336         spin_lock(&cmc_history_lock);
1337         if (!cmc_polling_enabled) {
1338                 int i, count = 1; /* we know 1 happened now */
1339                 unsigned long now = jiffies;
1340
1341                 for (i = 0; i < CMC_HISTORY_LENGTH; i++) {
1342                         if (now - cmc_history[i] <= HZ)
1343                                 count++;
1344                 }
1345
1346                 IA64_MCA_DEBUG(KERN_INFO "CMC threshold %d/%d\n", count, CMC_HISTORY_LENGTH);
1347                 if (count >= CMC_HISTORY_LENGTH) {
1348
1349                         cmc_polling_enabled = 1;
1350                         spin_unlock(&cmc_history_lock);
1351                         /* If we're being hit with CMC interrupts, we won't
1352                          * ever execute the schedule_work() below.  Need to
1353                          * disable CMC interrupts on this processor now.
1354                          */
1355                         ia64_mca_cmc_vector_disable(NULL);
1356                         schedule_work(&cmc_disable_work);
1357
1358                         /*
1359                          * Corrected errors will still be corrected, but
1360                          * make sure there's a log somewhere that indicates
1361                          * something is generating more than we can handle.
1362                          */
1363                         printk(KERN_WARNING "WARNING: Switching to polling CMC handler; error records may be lost\n");
1364
1365                         mod_timer(&cmc_poll_timer, jiffies + CMC_POLL_INTERVAL);
1366
1367                         /* lock already released, get out now */
1368                         goto out;
1369                 } else {
1370                         cmc_history[index++] = now;
1371                         if (index == CMC_HISTORY_LENGTH)
1372                                 index = 0;
1373                 }
1374         }
1375         spin_unlock(&cmc_history_lock);
1376 out:
1377         /* Get the CMC error record and log it */
1378         ia64_mca_log_sal_error_record(SAL_INFO_TYPE_CMC);
1379
1380         return IRQ_HANDLED;
1381 }
1382
1383 /*
1384  *  ia64_mca_cmc_int_caller
1385  *
1386  *      Triggered by sw interrupt from CMC polling routine.  Calls
1387  *      real interrupt handler and either triggers a sw interrupt
1388  *      on the next cpu or does cleanup at the end.
1389  *
1390  * Inputs
1391  *      interrupt number
1392  *      client data arg ptr
1393  * Outputs
1394  *      handled
1395  */
1396 static irqreturn_t
1397 ia64_mca_cmc_int_caller(int cmc_irq, void *arg)
1398 {
1399         static int start_count = -1;
1400         unsigned int cpuid;
1401
1402         cpuid = smp_processor_id();
1403
1404         /* If first cpu, update count */
1405         if (start_count == -1)
1406                 start_count = IA64_LOG_COUNT(SAL_INFO_TYPE_CMC);
1407
1408         ia64_mca_cmc_int_handler(cmc_irq, arg);
1409
1410         for (++cpuid ; cpuid < NR_CPUS && !cpu_online(cpuid) ; cpuid++);
1411
1412         if (cpuid < NR_CPUS) {
1413                 platform_send_ipi(cpuid, IA64_CMCP_VECTOR, IA64_IPI_DM_INT, 0);
1414         } else {
1415                 /* If no log record, switch out of polling mode */
1416                 if (start_count == IA64_LOG_COUNT(SAL_INFO_TYPE_CMC)) {
1417
1418                         printk(KERN_WARNING "Returning to interrupt driven CMC handler\n");
1419                         schedule_work(&cmc_enable_work);
1420                         cmc_polling_enabled = 0;
1421
1422                 } else {
1423
1424                         mod_timer(&cmc_poll_timer, jiffies + CMC_POLL_INTERVAL);
1425                 }
1426
1427                 start_count = -1;
1428         }
1429
1430         return IRQ_HANDLED;
1431 }
1432
1433 /*
1434  *  ia64_mca_cmc_poll
1435  *
1436  *      Poll for Corrected Machine Checks (CMCs)
1437  *
1438  * Inputs   :   dummy(unused)
1439  * Outputs  :   None
1440  *
1441  */
1442 static void
1443 ia64_mca_cmc_poll (unsigned long dummy)
1444 {
1445         /* Trigger a CMC interrupt cascade  */
1446         platform_send_ipi(first_cpu(cpu_online_map), IA64_CMCP_VECTOR, IA64_IPI_DM_INT, 0);
1447 }
1448
1449 /*
1450  *  ia64_mca_cpe_int_caller
1451  *
1452  *      Triggered by sw interrupt from CPE polling routine.  Calls
1453  *      real interrupt handler and either triggers a sw interrupt
1454  *      on the next cpu or does cleanup at the end.
1455  *
1456  * Inputs
1457  *      interrupt number
1458  *      client data arg ptr
1459  * Outputs
1460  *      handled
1461  */
1462 #ifdef CONFIG_ACPI
1463
1464 static irqreturn_t
1465 ia64_mca_cpe_int_caller(int cpe_irq, void *arg)
1466 {
1467         static int start_count = -1;
1468         static int poll_time = MIN_CPE_POLL_INTERVAL;
1469         unsigned int cpuid;
1470
1471         cpuid = smp_processor_id();
1472
1473         /* If first cpu, update count */
1474         if (start_count == -1)
1475                 start_count = IA64_LOG_COUNT(SAL_INFO_TYPE_CPE);
1476
1477         ia64_mca_cpe_int_handler(cpe_irq, arg);
1478
1479         for (++cpuid ; cpuid < NR_CPUS && !cpu_online(cpuid) ; cpuid++);
1480
1481         if (cpuid < NR_CPUS) {
1482                 platform_send_ipi(cpuid, IA64_CPEP_VECTOR, IA64_IPI_DM_INT, 0);
1483         } else {
1484                 /*
1485                  * If a log was recorded, increase our polling frequency,
1486                  * otherwise, backoff or return to interrupt mode.
1487                  */
1488                 if (start_count != IA64_LOG_COUNT(SAL_INFO_TYPE_CPE)) {
1489                         poll_time = max(MIN_CPE_POLL_INTERVAL, poll_time / 2);
1490                 } else if (cpe_vector < 0) {
1491                         poll_time = min(MAX_CPE_POLL_INTERVAL, poll_time * 2);
1492                 } else {
1493                         poll_time = MIN_CPE_POLL_INTERVAL;
1494
1495                         printk(KERN_WARNING "Returning to interrupt driven CPE handler\n");
1496                         enable_irq(local_vector_to_irq(IA64_CPE_VECTOR));
1497                         cpe_poll_enabled = 0;
1498                 }
1499
1500                 if (cpe_poll_enabled)
1501                         mod_timer(&cpe_poll_timer, jiffies + poll_time);
1502                 start_count = -1;
1503         }
1504
1505         return IRQ_HANDLED;
1506 }
1507
1508 /*
1509  *  ia64_mca_cpe_poll
1510  *
1511  *      Poll for Corrected Platform Errors (CPEs), trigger interrupt
1512  *      on first cpu, from there it will trickle through all the cpus.
1513  *
1514  * Inputs   :   dummy(unused)
1515  * Outputs  :   None
1516  *
1517  */
1518 static void
1519 ia64_mca_cpe_poll (unsigned long dummy)
1520 {
1521         /* Trigger a CPE interrupt cascade  */
1522         platform_send_ipi(first_cpu(cpu_online_map), IA64_CPEP_VECTOR, IA64_IPI_DM_INT, 0);
1523 }
1524
1525 #endif /* CONFIG_ACPI */
1526
1527 static int
1528 default_monarch_init_process(struct notifier_block *self, unsigned long val, void *data)
1529 {
1530         int c;
1531         struct task_struct *g, *t;
1532         if (val != DIE_INIT_MONARCH_PROCESS)
1533                 return NOTIFY_DONE;
1534 #ifdef CONFIG_KEXEC
1535         if (atomic_read(&kdump_in_progress))
1536                 return NOTIFY_DONE;
1537 #endif
1538
1539         /*
1540          * FIXME: mlogbuf will brim over with INIT stack dumps.
1541          * To enable show_stack from INIT, we use oops_in_progress which should
1542          * be used in real oops. This would cause something wrong after INIT.
1543          */
1544         BREAK_LOGLEVEL(console_loglevel);
1545         ia64_mlogbuf_dump_from_init();
1546
1547         printk(KERN_ERR "Processes interrupted by INIT -");
1548         for_each_online_cpu(c) {
1549                 struct ia64_sal_os_state *s;
1550                 t = __va(__per_cpu_mca[c] + IA64_MCA_CPU_INIT_STACK_OFFSET);
1551                 s = (struct ia64_sal_os_state *)((char *)t + MCA_SOS_OFFSET);
1552                 g = s->prev_task;
1553                 if (g) {
1554                         if (g->pid)
1555                                 printk(" %d", g->pid);
1556                         else
1557                                 printk(" %d (cpu %d task 0x%p)", g->pid, task_cpu(g), g);
1558                 }
1559         }
1560         printk("\n\n");
1561         if (read_trylock(&tasklist_lock)) {
1562                 do_each_thread (g, t) {
1563                         printk("\nBacktrace of pid %d (%s)\n", t->pid, t->comm);
1564                         show_stack(t, NULL);
1565                 } while_each_thread (g, t);
1566                 read_unlock(&tasklist_lock);
1567         }
1568         /* FIXME: This will not restore zapped printk locks. */
1569         RESTORE_LOGLEVEL(console_loglevel);
1570         return NOTIFY_DONE;
1571 }
1572
1573 /*
1574  * C portion of the OS INIT handler
1575  *
1576  * Called from ia64_os_init_dispatch
1577  *
1578  * Inputs: pointer to pt_regs where processor info was saved.  SAL/OS state for
1579  * this event.  This code is used for both monarch and slave INIT events, see
1580  * sos->monarch.
1581  *
1582  * All INIT events switch to the INIT stack and change the previous process to
1583  * blocked status.  If one of the INIT events is the monarch then we are
1584  * probably processing the nmi button/command.  Use the monarch cpu to dump all
1585  * the processes.  The slave INIT events all spin until the monarch cpu
1586  * returns.  We can also get INIT slave events for MCA, in which case the MCA
1587  * process is the monarch.
1588  */
1589
1590 void
1591 ia64_init_handler(struct pt_regs *regs, struct switch_stack *sw,
1592                   struct ia64_sal_os_state *sos)
1593 {
1594         static atomic_t slaves;
1595         static atomic_t monarchs;
1596         struct task_struct *previous_current;
1597         int cpu = smp_processor_id();
1598         struct ia64_mca_notify_die nd =
1599                 { .sos = sos, .monarch_cpu = &monarch_cpu };
1600
1601         (void) notify_die(DIE_INIT_ENTER, "INIT", regs, (long)&nd, 0, 0);
1602
1603         mprintk(KERN_INFO "Entered OS INIT handler. PSP=%lx cpu=%d monarch=%ld\n",
1604                 sos->proc_state_param, cpu, sos->monarch);
1605         salinfo_log_wakeup(SAL_INFO_TYPE_INIT, NULL, 0, 0);
1606
1607         previous_current = ia64_mca_modify_original_stack(regs, sw, sos, "INIT");
1608         sos->os_status = IA64_INIT_RESUME;
1609
1610         /* FIXME: Workaround for broken proms that drive all INIT events as
1611          * slaves.  The last slave that enters is promoted to be a monarch.
1612          * Remove this code in September 2006, that gives platforms a year to
1613          * fix their proms and get their customers updated.
1614          */
1615         if (!sos->monarch && atomic_add_return(1, &slaves) == num_online_cpus()) {
1616                 mprintk(KERN_WARNING "%s: Promoting cpu %d to monarch.\n",
1617                        __FUNCTION__, cpu);
1618                 atomic_dec(&slaves);
1619                 sos->monarch = 1;
1620         }
1621
1622         /* FIXME: Workaround for broken proms that drive all INIT events as
1623          * monarchs.  Second and subsequent monarchs are demoted to slaves.
1624          * Remove this code in September 2006, that gives platforms a year to
1625          * fix their proms and get their customers updated.
1626          */
1627         if (sos->monarch && atomic_add_return(1, &monarchs) > 1) {
1628                 mprintk(KERN_WARNING "%s: Demoting cpu %d to slave.\n",
1629                                __FUNCTION__, cpu);
1630                 atomic_dec(&monarchs);
1631                 sos->monarch = 0;
1632         }
1633
1634         if (!sos->monarch) {
1635                 ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_INIT;
1636                 while (monarch_cpu == -1)
1637                        cpu_relax();     /* spin until monarch enters */
1638                 if (notify_die(DIE_INIT_SLAVE_ENTER, "INIT", regs, (long)&nd, 0, 0)
1639                                 == NOTIFY_STOP)
1640                         ia64_mca_spin(__FUNCTION__);
1641                 if (notify_die(DIE_INIT_SLAVE_PROCESS, "INIT", regs, (long)&nd, 0, 0)
1642                                 == NOTIFY_STOP)
1643                         ia64_mca_spin(__FUNCTION__);
1644                 while (monarch_cpu != -1)
1645                        cpu_relax();     /* spin until monarch leaves */
1646                 if (notify_die(DIE_INIT_SLAVE_LEAVE, "INIT", regs, (long)&nd, 0, 0)
1647                                 == NOTIFY_STOP)
1648                         ia64_mca_spin(__FUNCTION__);
1649                 mprintk("Slave on cpu %d returning to normal service.\n", cpu);
1650                 set_curr_task(cpu, previous_current);
1651                 ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1652                 atomic_dec(&slaves);
1653                 return;
1654         }
1655
1656         monarch_cpu = cpu;
1657         if (notify_die(DIE_INIT_MONARCH_ENTER, "INIT", regs, (long)&nd, 0, 0)
1658                         == NOTIFY_STOP)
1659                 ia64_mca_spin(__FUNCTION__);
1660
1661         /*
1662          * Wait for a bit.  On some machines (e.g., HP's zx2000 and zx6000, INIT can be
1663          * generated via the BMC's command-line interface, but since the console is on the
1664          * same serial line, the user will need some time to switch out of the BMC before
1665          * the dump begins.
1666          */
1667         mprintk("Delaying for 5 seconds...\n");
1668         udelay(5*1000000);
1669         ia64_wait_for_slaves(cpu, "INIT");
1670         /* If nobody intercepts DIE_INIT_MONARCH_PROCESS then we drop through
1671          * to default_monarch_init_process() above and just print all the
1672          * tasks.
1673          */
1674         if (notify_die(DIE_INIT_MONARCH_PROCESS, "INIT", regs, (long)&nd, 0, 0)
1675                         == NOTIFY_STOP)
1676                 ia64_mca_spin(__FUNCTION__);
1677         if (notify_die(DIE_INIT_MONARCH_LEAVE, "INIT", regs, (long)&nd, 0, 0)
1678                         == NOTIFY_STOP)
1679                 ia64_mca_spin(__FUNCTION__);
1680         mprintk("\nINIT dump complete.  Monarch on cpu %d returning to normal service.\n", cpu);
1681         atomic_dec(&monarchs);
1682         set_curr_task(cpu, previous_current);
1683         monarch_cpu = -1;
1684         return;
1685 }
1686
1687 static int __init
1688 ia64_mca_disable_cpe_polling(char *str)
1689 {
1690         cpe_poll_enabled = 0;
1691         return 1;
1692 }
1693
1694 __setup("disable_cpe_poll", ia64_mca_disable_cpe_polling);
1695
1696 static struct irqaction cmci_irqaction = {
1697         .handler =      ia64_mca_cmc_int_handler,
1698         .flags =        IRQF_DISABLED,
1699         .name =         "cmc_hndlr"
1700 };
1701
1702 static struct irqaction cmcp_irqaction = {
1703         .handler =      ia64_mca_cmc_int_caller,
1704         .flags =        IRQF_DISABLED,
1705         .name =         "cmc_poll"
1706 };
1707
1708 static struct irqaction mca_rdzv_irqaction = {
1709         .handler =      ia64_mca_rendez_int_handler,
1710         .flags =        IRQF_DISABLED,
1711         .name =         "mca_rdzv"
1712 };
1713
1714 static struct irqaction mca_wkup_irqaction = {
1715         .handler =      ia64_mca_wakeup_int_handler,
1716         .flags =        IRQF_DISABLED,
1717         .name =         "mca_wkup"
1718 };
1719
1720 #ifdef CONFIG_ACPI
1721 static struct irqaction mca_cpe_irqaction = {
1722         .handler =      ia64_mca_cpe_int_handler,
1723         .flags =        IRQF_DISABLED,
1724         .name =         "cpe_hndlr"
1725 };
1726
1727 static struct irqaction mca_cpep_irqaction = {
1728         .handler =      ia64_mca_cpe_int_caller,
1729         .flags =        IRQF_DISABLED,
1730         .name =         "cpe_poll"
1731 };
1732 #endif /* CONFIG_ACPI */
1733
1734 /* Minimal format of the MCA/INIT stacks.  The pseudo processes that run on
1735  * these stacks can never sleep, they cannot return from the kernel to user
1736  * space, they do not appear in a normal ps listing.  So there is no need to
1737  * format most of the fields.
1738  */
1739
1740 static void __cpuinit
1741 format_mca_init_stack(void *mca_data, unsigned long offset,
1742                 const char *type, int cpu)
1743 {
1744         struct task_struct *p = (struct task_struct *)((char *)mca_data + offset);
1745         struct thread_info *ti;
1746         memset(p, 0, KERNEL_STACK_SIZE);
1747         ti = task_thread_info(p);
1748         ti->flags = _TIF_MCA_INIT;
1749         ti->preempt_count = 1;
1750         ti->task = p;
1751         ti->cpu = cpu;
1752         p->stack = ti;
1753         p->state = TASK_UNINTERRUPTIBLE;
1754         cpu_set(cpu, p->cpus_allowed);
1755         INIT_LIST_HEAD(&p->tasks);
1756         p->parent = p->real_parent = p->group_leader = p;
1757         INIT_LIST_HEAD(&p->children);
1758         INIT_LIST_HEAD(&p->sibling);
1759         strncpy(p->comm, type, sizeof(p->comm)-1);
1760 }
1761
1762 /* Caller prevents this from being called after init */
1763 static void * __init_refok mca_bootmem(void)
1764 {
1765         return __alloc_bootmem(sizeof(struct ia64_mca_cpu),
1766                             KERNEL_STACK_SIZE, 0);
1767 }
1768
1769 /* Do per-CPU MCA-related initialization.  */
1770 void __cpuinit
1771 ia64_mca_cpu_init(void *cpu_data)
1772 {
1773         void *pal_vaddr;
1774         void *data;
1775         long sz = sizeof(struct ia64_mca_cpu);
1776         int cpu = smp_processor_id();
1777         static int first_time = 1;
1778
1779         /*
1780          * Structure will already be allocated if cpu has been online,
1781          * then offlined.
1782          */
1783         if (__per_cpu_mca[cpu]) {
1784                 data = __va(__per_cpu_mca[cpu]);
1785         } else {
1786                 if (first_time) {
1787                         data = mca_bootmem();
1788                         first_time = 0;
1789                 } else
1790                         data = page_address(alloc_pages_node(numa_node_id(),
1791                                         GFP_KERNEL, get_order(sz)));
1792                 if (!data)
1793                         panic("Could not allocate MCA memory for cpu %d\n",
1794                                         cpu);
1795         }
1796         format_mca_init_stack(data, offsetof(struct ia64_mca_cpu, mca_stack),
1797                 "MCA", cpu);
1798         format_mca_init_stack(data, offsetof(struct ia64_mca_cpu, init_stack),
1799                 "INIT", cpu);
1800         __get_cpu_var(ia64_mca_data) = __per_cpu_mca[cpu] = __pa(data);
1801
1802         /*
1803          * Stash away a copy of the PTE needed to map the per-CPU page.
1804          * We may need it during MCA recovery.
1805          */
1806         __get_cpu_var(ia64_mca_per_cpu_pte) =
1807                 pte_val(mk_pte_phys(__pa(cpu_data), PAGE_KERNEL));
1808
1809         /*
1810          * Also, stash away a copy of the PAL address and the PTE
1811          * needed to map it.
1812          */
1813         pal_vaddr = efi_get_pal_addr();
1814         if (!pal_vaddr)
1815                 return;
1816         __get_cpu_var(ia64_mca_pal_base) =
1817                 GRANULEROUNDDOWN((unsigned long) pal_vaddr);
1818         __get_cpu_var(ia64_mca_pal_pte) = pte_val(mk_pte_phys(__pa(pal_vaddr),
1819                                                               PAGE_KERNEL));
1820 }
1821
1822 static void __cpuinit ia64_mca_cmc_vector_adjust(void *dummy)
1823 {
1824         unsigned long flags;
1825
1826         local_irq_save(flags);
1827         if (!cmc_polling_enabled)
1828                 ia64_mca_cmc_vector_enable(NULL);
1829         local_irq_restore(flags);
1830 }
1831
1832 static int __cpuinit mca_cpu_callback(struct notifier_block *nfb,
1833                                       unsigned long action,
1834                                       void *hcpu)
1835 {
1836         int hotcpu = (unsigned long) hcpu;
1837
1838         switch (action) {
1839         case CPU_ONLINE:
1840         case CPU_ONLINE_FROZEN:
1841                 smp_call_function_single(hotcpu, ia64_mca_cmc_vector_adjust,
1842                                          NULL, 1, 0);
1843                 break;
1844         }
1845         return NOTIFY_OK;
1846 }
1847
1848 static struct notifier_block mca_cpu_notifier __cpuinitdata = {
1849         .notifier_call = mca_cpu_callback
1850 };
1851
1852 /*
1853  * ia64_mca_init
1854  *
1855  *  Do all the system level mca specific initialization.
1856  *
1857  *      1. Register spinloop and wakeup request interrupt vectors
1858  *
1859  *      2. Register OS_MCA handler entry point
1860  *
1861  *      3. Register OS_INIT handler entry point
1862  *
1863  *  4. Initialize MCA/CMC/INIT related log buffers maintained by the OS.
1864  *
1865  *  Note that this initialization is done very early before some kernel
1866  *  services are available.
1867  *
1868  *  Inputs  :   None
1869  *
1870  *  Outputs :   None
1871  */
1872 void __init
1873 ia64_mca_init(void)
1874 {
1875         ia64_fptr_t *init_hldlr_ptr_monarch = (ia64_fptr_t *)ia64_os_init_dispatch_monarch;
1876         ia64_fptr_t *init_hldlr_ptr_slave = (ia64_fptr_t *)ia64_os_init_dispatch_slave;
1877         ia64_fptr_t *mca_hldlr_ptr = (ia64_fptr_t *)ia64_os_mca_dispatch;
1878         int i;
1879         s64 rc;
1880         struct ia64_sal_retval isrv;
1881         u64 timeout = IA64_MCA_RENDEZ_TIMEOUT;  /* platform specific */
1882         static struct notifier_block default_init_monarch_nb = {
1883                 .notifier_call = default_monarch_init_process,
1884                 .priority = 0/* we need to notified last */
1885         };
1886
1887         IA64_MCA_DEBUG("%s: begin\n", __FUNCTION__);
1888
1889         /* Clear the Rendez checkin flag for all cpus */
1890         for(i = 0 ; i < NR_CPUS; i++)
1891                 ia64_mc_info.imi_rendez_checkin[i] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1892
1893         /*
1894          * Register the rendezvous spinloop and wakeup mechanism with SAL
1895          */
1896
1897         /* Register the rendezvous interrupt vector with SAL */
1898         while (1) {
1899                 isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_RENDEZ_INT,
1900                                               SAL_MC_PARAM_MECHANISM_INT,
1901                                               IA64_MCA_RENDEZ_VECTOR,
1902                                               timeout,
1903                                               SAL_MC_PARAM_RZ_ALWAYS);
1904                 rc = isrv.status;
1905                 if (rc == 0)
1906                         break;
1907                 if (rc == -2) {
1908                         printk(KERN_INFO "Increasing MCA rendezvous timeout from "
1909                                 "%ld to %ld milliseconds\n", timeout, isrv.v0);
1910                         timeout = isrv.v0;
1911                         (void) notify_die(DIE_MCA_NEW_TIMEOUT, "MCA", NULL, timeout, 0, 0);
1912                         continue;
1913                 }
1914                 printk(KERN_ERR "Failed to register rendezvous interrupt "
1915                        "with SAL (status %ld)\n", rc);
1916                 return;
1917         }
1918
1919         /* Register the wakeup interrupt vector with SAL */
1920         isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_RENDEZ_WAKEUP,
1921                                       SAL_MC_PARAM_MECHANISM_INT,
1922                                       IA64_MCA_WAKEUP_VECTOR,
1923                                       0, 0);
1924         rc = isrv.status;
1925         if (rc) {
1926                 printk(KERN_ERR "Failed to register wakeup interrupt with SAL "
1927                        "(status %ld)\n", rc);
1928                 return;
1929         }
1930
1931         IA64_MCA_DEBUG("%s: registered MCA rendezvous spinloop and wakeup mech.\n", __FUNCTION__);
1932
1933         ia64_mc_info.imi_mca_handler        = ia64_tpa(mca_hldlr_ptr->fp);
1934         /*
1935          * XXX - disable SAL checksum by setting size to 0; should be
1936          *      ia64_tpa(ia64_os_mca_dispatch_end) - ia64_tpa(ia64_os_mca_dispatch);
1937          */
1938         ia64_mc_info.imi_mca_handler_size       = 0;
1939
1940         /* Register the os mca handler with SAL */
1941         if ((rc = ia64_sal_set_vectors(SAL_VECTOR_OS_MCA,
1942                                        ia64_mc_info.imi_mca_handler,
1943                                        ia64_tpa(mca_hldlr_ptr->gp),
1944                                        ia64_mc_info.imi_mca_handler_size,
1945                                        0, 0, 0)))
1946         {
1947                 printk(KERN_ERR "Failed to register OS MCA handler with SAL "
1948                        "(status %ld)\n", rc);
1949                 return;
1950         }
1951
1952         IA64_MCA_DEBUG("%s: registered OS MCA handler with SAL at 0x%lx, gp = 0x%lx\n", __FUNCTION__,
1953                        ia64_mc_info.imi_mca_handler, ia64_tpa(mca_hldlr_ptr->gp));
1954
1955         /*
1956          * XXX - disable SAL checksum by setting size to 0, should be
1957          * size of the actual init handler in mca_asm.S.
1958          */
1959         ia64_mc_info.imi_monarch_init_handler           = ia64_tpa(init_hldlr_ptr_monarch->fp);
1960         ia64_mc_info.imi_monarch_init_handler_size      = 0;
1961         ia64_mc_info.imi_slave_init_handler             = ia64_tpa(init_hldlr_ptr_slave->fp);
1962         ia64_mc_info.imi_slave_init_handler_size        = 0;
1963
1964         IA64_MCA_DEBUG("%s: OS INIT handler at %lx\n", __FUNCTION__,
1965                        ia64_mc_info.imi_monarch_init_handler);
1966
1967         /* Register the os init handler with SAL */
1968         if ((rc = ia64_sal_set_vectors(SAL_VECTOR_OS_INIT,
1969                                        ia64_mc_info.imi_monarch_init_handler,
1970                                        ia64_tpa(ia64_getreg(_IA64_REG_GP)),
1971                                        ia64_mc_info.imi_monarch_init_handler_size,
1972                                        ia64_mc_info.imi_slave_init_handler,
1973                                        ia64_tpa(ia64_getreg(_IA64_REG_GP)),
1974                                        ia64_mc_info.imi_slave_init_handler_size)))
1975         {
1976                 printk(KERN_ERR "Failed to register m/s INIT handlers with SAL "
1977                        "(status %ld)\n", rc);
1978                 return;
1979         }
1980         if (register_die_notifier(&default_init_monarch_nb)) {
1981                 printk(KERN_ERR "Failed to register default monarch INIT process\n");
1982                 return;
1983         }
1984
1985         IA64_MCA_DEBUG("%s: registered OS INIT handler with SAL\n", __FUNCTION__);
1986
1987         /*
1988          *  Configure the CMCI/P vector and handler. Interrupts for CMC are
1989          *  per-processor, so AP CMC interrupts are setup in smp_callin() (smpboot.c).
1990          */
1991         register_percpu_irq(IA64_CMC_VECTOR, &cmci_irqaction);
1992         register_percpu_irq(IA64_CMCP_VECTOR, &cmcp_irqaction);
1993         ia64_mca_cmc_vector_setup();       /* Setup vector on BSP */
1994
1995         /* Setup the MCA rendezvous interrupt vector */
1996         register_percpu_irq(IA64_MCA_RENDEZ_VECTOR, &mca_rdzv_irqaction);
1997
1998         /* Setup the MCA wakeup interrupt vector */
1999         register_percpu_irq(IA64_MCA_WAKEUP_VECTOR, &mca_wkup_irqaction);
2000
2001 #ifdef CONFIG_ACPI
2002         /* Setup the CPEI/P handler */
2003         register_percpu_irq(IA64_CPEP_VECTOR, &mca_cpep_irqaction);
2004 #endif
2005
2006         /* Initialize the areas set aside by the OS to buffer the
2007          * platform/processor error states for MCA/INIT/CMC
2008          * handling.
2009          */
2010         ia64_log_init(SAL_INFO_TYPE_MCA);
2011         ia64_log_init(SAL_INFO_TYPE_INIT);
2012         ia64_log_init(SAL_INFO_TYPE_CMC);
2013         ia64_log_init(SAL_INFO_TYPE_CPE);
2014
2015         mca_init = 1;
2016         printk(KERN_INFO "MCA related initialization done\n");
2017 }
2018
2019 /*
2020  * ia64_mca_late_init
2021  *
2022  *      Opportunity to setup things that require initialization later
2023  *      than ia64_mca_init.  Setup a timer to poll for CPEs if the
2024  *      platform doesn't support an interrupt driven mechanism.
2025  *
2026  *  Inputs  :   None
2027  *  Outputs :   Status
2028  */
2029 static int __init
2030 ia64_mca_late_init(void)
2031 {
2032         if (!mca_init)
2033                 return 0;
2034
2035         register_hotcpu_notifier(&mca_cpu_notifier);
2036
2037         /* Setup the CMCI/P vector and handler */
2038         init_timer(&cmc_poll_timer);
2039         cmc_poll_timer.function = ia64_mca_cmc_poll;
2040
2041         /* Unmask/enable the vector */
2042         cmc_polling_enabled = 0;
2043         schedule_work(&cmc_enable_work);
2044
2045         IA64_MCA_DEBUG("%s: CMCI/P setup and enabled.\n", __FUNCTION__);
2046
2047 #ifdef CONFIG_ACPI
2048         /* Setup the CPEI/P vector and handler */
2049         cpe_vector = acpi_request_vector(ACPI_INTERRUPT_CPEI);
2050         init_timer(&cpe_poll_timer);
2051         cpe_poll_timer.function = ia64_mca_cpe_poll;
2052
2053         {
2054                 irq_desc_t *desc;
2055                 unsigned int irq;
2056
2057                 if (cpe_vector >= 0) {
2058                         /* If platform supports CPEI, enable the irq. */
2059                         irq = local_vector_to_irq(cpe_vector);
2060                         if (irq > 0) {
2061                                 cpe_poll_enabled = 0;
2062                                 desc = irq_desc + irq;
2063                                 desc->status |= IRQ_PER_CPU;
2064                                 setup_irq(irq, &mca_cpe_irqaction);
2065                                 ia64_cpe_irq = irq;
2066                                 ia64_mca_register_cpev(cpe_vector);
2067                                 IA64_MCA_DEBUG("%s: CPEI/P setup and enabled.\n",
2068                                         __FUNCTION__);
2069                                 return 0;
2070                         }
2071                         printk(KERN_ERR "%s: Failed to find irq for CPE "
2072                                         "interrupt handler, vector %d\n",
2073                                         __FUNCTION__, cpe_vector);
2074                 }
2075                 /* If platform doesn't support CPEI, get the timer going. */
2076                 if (cpe_poll_enabled) {
2077                         ia64_mca_cpe_poll(0UL);
2078                         IA64_MCA_DEBUG("%s: CPEP setup and enabled.\n", __FUNCTION__);
2079                 }
2080         }
2081 #endif
2082
2083         return 0;
2084 }
2085
2086 device_initcall(ia64_mca_late_init);