xref: /openbmc/linux/arch/ia64/kernel/mca.c (revision a09d2831)
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 
89 #include <asm/delay.h>
90 #include <asm/machvec.h>
91 #include <asm/meminit.h>
92 #include <asm/page.h>
93 #include <asm/ptrace.h>
94 #include <asm/system.h>
95 #include <asm/sal.h>
96 #include <asm/mca.h>
97 #include <asm/kexec.h>
98 
99 #include <asm/irq.h>
100 #include <asm/hw_irq.h>
101 #include <asm/tlb.h>
102 
103 #include "mca_drv.h"
104 #include "entry.h"
105 
106 #if defined(IA64_MCA_DEBUG_INFO)
107 # define IA64_MCA_DEBUG(fmt...)	printk(fmt)
108 #else
109 # define IA64_MCA_DEBUG(fmt...)
110 #endif
111 
112 #define NOTIFY_INIT(event, regs, arg, spin)				\
113 do {									\
114 	if ((notify_die((event), "INIT", (regs), (arg), 0, 0)		\
115 			== NOTIFY_STOP) && ((spin) == 1))		\
116 		ia64_mca_spin(__func__);				\
117 } while (0)
118 
119 #define NOTIFY_MCA(event, regs, arg, spin)				\
120 do {									\
121 	if ((notify_die((event), "MCA", (regs), (arg), 0, 0)		\
122 			== NOTIFY_STOP) && ((spin) == 1))		\
123 		ia64_mca_spin(__func__);				\
124 } while (0)
125 
126 /* Used by mca_asm.S */
127 DEFINE_PER_CPU(u64, ia64_mca_data); /* == __per_cpu_mca[smp_processor_id()] */
128 DEFINE_PER_CPU(u64, ia64_mca_per_cpu_pte); /* PTE to map per-CPU area */
129 DEFINE_PER_CPU(u64, ia64_mca_pal_pte);	    /* PTE to map PAL code */
130 DEFINE_PER_CPU(u64, ia64_mca_pal_base);    /* vaddr PAL code granule */
131 DEFINE_PER_CPU(u64, ia64_mca_tr_reload);   /* Flag for TR reload */
132 
133 unsigned long __per_cpu_mca[NR_CPUS];
134 
135 /* In mca_asm.S */
136 extern void			ia64_os_init_dispatch_monarch (void);
137 extern void			ia64_os_init_dispatch_slave (void);
138 
139 static int monarch_cpu = -1;
140 
141 static ia64_mc_info_t		ia64_mc_info;
142 
143 #define MAX_CPE_POLL_INTERVAL (15*60*HZ) /* 15 minutes */
144 #define MIN_CPE_POLL_INTERVAL (2*60*HZ)  /* 2 minutes */
145 #define CMC_POLL_INTERVAL     (1*60*HZ)  /* 1 minute */
146 #define CPE_HISTORY_LENGTH    5
147 #define CMC_HISTORY_LENGTH    5
148 
149 #ifdef CONFIG_ACPI
150 static struct timer_list cpe_poll_timer;
151 #endif
152 static struct timer_list cmc_poll_timer;
153 /*
154  * This variable tells whether we are currently in polling mode.
155  * Start with this in the wrong state so we won't play w/ timers
156  * before the system is ready.
157  */
158 static int cmc_polling_enabled = 1;
159 
160 /*
161  * Clearing this variable prevents CPE polling from getting activated
162  * in mca_late_init.  Use it if your system doesn't provide a CPEI,
163  * but encounters problems retrieving CPE logs.  This should only be
164  * necessary for debugging.
165  */
166 static int cpe_poll_enabled = 1;
167 
168 extern void salinfo_log_wakeup(int type, u8 *buffer, u64 size, int irqsafe);
169 
170 static int mca_init __initdata;
171 
172 /*
173  * limited & delayed printing support for MCA/INIT handler
174  */
175 
176 #define mprintk(fmt...) ia64_mca_printk(fmt)
177 
178 #define MLOGBUF_SIZE (512+256*NR_CPUS)
179 #define MLOGBUF_MSGMAX 256
180 static char mlogbuf[MLOGBUF_SIZE];
181 static DEFINE_SPINLOCK(mlogbuf_wlock);	/* mca context only */
182 static DEFINE_SPINLOCK(mlogbuf_rlock);	/* normal context only */
183 static unsigned long mlogbuf_start;
184 static unsigned long mlogbuf_end;
185 static unsigned int mlogbuf_finished = 0;
186 static unsigned long mlogbuf_timestamp = 0;
187 
188 static int loglevel_save = -1;
189 #define BREAK_LOGLEVEL(__console_loglevel)		\
190 	oops_in_progress = 1;				\
191 	if (loglevel_save < 0)				\
192 		loglevel_save = __console_loglevel;	\
193 	__console_loglevel = 15;
194 
195 #define RESTORE_LOGLEVEL(__console_loglevel)		\
196 	if (loglevel_save >= 0) {			\
197 		__console_loglevel = loglevel_save;	\
198 		loglevel_save = -1;			\
199 	}						\
200 	mlogbuf_finished = 0;				\
201 	oops_in_progress = 0;
202 
203 /*
204  * Push messages into buffer, print them later if not urgent.
205  */
206 void ia64_mca_printk(const char *fmt, ...)
207 {
208 	va_list args;
209 	int printed_len;
210 	char temp_buf[MLOGBUF_MSGMAX];
211 	char *p;
212 
213 	va_start(args, fmt);
214 	printed_len = vscnprintf(temp_buf, sizeof(temp_buf), fmt, args);
215 	va_end(args);
216 
217 	/* Copy the output into mlogbuf */
218 	if (oops_in_progress) {
219 		/* mlogbuf was abandoned, use printk directly instead. */
220 		printk(temp_buf);
221 	} else {
222 		spin_lock(&mlogbuf_wlock);
223 		for (p = temp_buf; *p; p++) {
224 			unsigned long next = (mlogbuf_end + 1) % MLOGBUF_SIZE;
225 			if (next != mlogbuf_start) {
226 				mlogbuf[mlogbuf_end] = *p;
227 				mlogbuf_end = next;
228 			} else {
229 				/* buffer full */
230 				break;
231 			}
232 		}
233 		mlogbuf[mlogbuf_end] = '\0';
234 		spin_unlock(&mlogbuf_wlock);
235 	}
236 }
237 EXPORT_SYMBOL(ia64_mca_printk);
238 
239 /*
240  * Print buffered messages.
241  *  NOTE: call this after returning normal context. (ex. from salinfod)
242  */
243 void ia64_mlogbuf_dump(void)
244 {
245 	char temp_buf[MLOGBUF_MSGMAX];
246 	char *p;
247 	unsigned long index;
248 	unsigned long flags;
249 	unsigned int printed_len;
250 
251 	/* Get output from mlogbuf */
252 	while (mlogbuf_start != mlogbuf_end) {
253 		temp_buf[0] = '\0';
254 		p = temp_buf;
255 		printed_len = 0;
256 
257 		spin_lock_irqsave(&mlogbuf_rlock, flags);
258 
259 		index = mlogbuf_start;
260 		while (index != mlogbuf_end) {
261 			*p = mlogbuf[index];
262 			index = (index + 1) % MLOGBUF_SIZE;
263 			if (!*p)
264 				break;
265 			p++;
266 			if (++printed_len >= MLOGBUF_MSGMAX - 1)
267 				break;
268 		}
269 		*p = '\0';
270 		if (temp_buf[0])
271 			printk(temp_buf);
272 		mlogbuf_start = index;
273 
274 		mlogbuf_timestamp = 0;
275 		spin_unlock_irqrestore(&mlogbuf_rlock, flags);
276 	}
277 }
278 EXPORT_SYMBOL(ia64_mlogbuf_dump);
279 
280 /*
281  * Call this if system is going to down or if immediate flushing messages to
282  * console is required. (ex. recovery was failed, crash dump is going to be
283  * invoked, long-wait rendezvous etc.)
284  *  NOTE: this should be called from monarch.
285  */
286 static void ia64_mlogbuf_finish(int wait)
287 {
288 	BREAK_LOGLEVEL(console_loglevel);
289 
290 	spin_lock_init(&mlogbuf_rlock);
291 	ia64_mlogbuf_dump();
292 	printk(KERN_EMERG "mlogbuf_finish: printing switched to urgent mode, "
293 		"MCA/INIT might be dodgy or fail.\n");
294 
295 	if (!wait)
296 		return;
297 
298 	/* wait for console */
299 	printk("Delaying for 5 seconds...\n");
300 	udelay(5*1000000);
301 
302 	mlogbuf_finished = 1;
303 }
304 
305 /*
306  * Print buffered messages from INIT context.
307  */
308 static void ia64_mlogbuf_dump_from_init(void)
309 {
310 	if (mlogbuf_finished)
311 		return;
312 
313 	if (mlogbuf_timestamp &&
314 			time_before(jiffies, mlogbuf_timestamp + 30 * HZ)) {
315 		printk(KERN_ERR "INIT: mlogbuf_dump is interrupted by INIT "
316 			" and the system seems to be messed up.\n");
317 		ia64_mlogbuf_finish(0);
318 		return;
319 	}
320 
321 	if (!spin_trylock(&mlogbuf_rlock)) {
322 		printk(KERN_ERR "INIT: mlogbuf_dump is interrupted by INIT. "
323 			"Generated messages other than stack dump will be "
324 			"buffered to mlogbuf and will be printed later.\n");
325 		printk(KERN_ERR "INIT: If messages would not printed after "
326 			"this INIT, wait 30sec and assert INIT again.\n");
327 		if (!mlogbuf_timestamp)
328 			mlogbuf_timestamp = jiffies;
329 		return;
330 	}
331 	spin_unlock(&mlogbuf_rlock);
332 	ia64_mlogbuf_dump();
333 }
334 
335 static void inline
336 ia64_mca_spin(const char *func)
337 {
338 	if (monarch_cpu == smp_processor_id())
339 		ia64_mlogbuf_finish(0);
340 	mprintk(KERN_EMERG "%s: spinning here, not returning to SAL\n", func);
341 	while (1)
342 		cpu_relax();
343 }
344 /*
345  * IA64_MCA log support
346  */
347 #define IA64_MAX_LOGS		2	/* Double-buffering for nested MCAs */
348 #define IA64_MAX_LOG_TYPES      4   /* MCA, INIT, CMC, CPE */
349 
350 typedef struct ia64_state_log_s
351 {
352 	spinlock_t	isl_lock;
353 	int		isl_index;
354 	unsigned long	isl_count;
355 	ia64_err_rec_t  *isl_log[IA64_MAX_LOGS]; /* need space to store header + error log */
356 } ia64_state_log_t;
357 
358 static ia64_state_log_t ia64_state_log[IA64_MAX_LOG_TYPES];
359 
360 #define IA64_LOG_ALLOCATE(it, size) \
361 	{ia64_state_log[it].isl_log[IA64_LOG_CURR_INDEX(it)] = \
362 		(ia64_err_rec_t *)alloc_bootmem(size); \
363 	ia64_state_log[it].isl_log[IA64_LOG_NEXT_INDEX(it)] = \
364 		(ia64_err_rec_t *)alloc_bootmem(size);}
365 #define IA64_LOG_LOCK_INIT(it) spin_lock_init(&ia64_state_log[it].isl_lock)
366 #define IA64_LOG_LOCK(it)      spin_lock_irqsave(&ia64_state_log[it].isl_lock, s)
367 #define IA64_LOG_UNLOCK(it)    spin_unlock_irqrestore(&ia64_state_log[it].isl_lock,s)
368 #define IA64_LOG_NEXT_INDEX(it)    ia64_state_log[it].isl_index
369 #define IA64_LOG_CURR_INDEX(it)    1 - ia64_state_log[it].isl_index
370 #define IA64_LOG_INDEX_INC(it) \
371     {ia64_state_log[it].isl_index = 1 - ia64_state_log[it].isl_index; \
372     ia64_state_log[it].isl_count++;}
373 #define IA64_LOG_INDEX_DEC(it) \
374     ia64_state_log[it].isl_index = 1 - ia64_state_log[it].isl_index
375 #define IA64_LOG_NEXT_BUFFER(it)   (void *)((ia64_state_log[it].isl_log[IA64_LOG_NEXT_INDEX(it)]))
376 #define IA64_LOG_CURR_BUFFER(it)   (void *)((ia64_state_log[it].isl_log[IA64_LOG_CURR_INDEX(it)]))
377 #define IA64_LOG_COUNT(it)         ia64_state_log[it].isl_count
378 
379 /*
380  * ia64_log_init
381  *	Reset the OS ia64 log buffer
382  * Inputs   :   info_type   (SAL_INFO_TYPE_{MCA,INIT,CMC,CPE})
383  * Outputs	:	None
384  */
385 static void __init
386 ia64_log_init(int sal_info_type)
387 {
388 	u64	max_size = 0;
389 
390 	IA64_LOG_NEXT_INDEX(sal_info_type) = 0;
391 	IA64_LOG_LOCK_INIT(sal_info_type);
392 
393 	// SAL will tell us the maximum size of any error record of this type
394 	max_size = ia64_sal_get_state_info_size(sal_info_type);
395 	if (!max_size)
396 		/* alloc_bootmem() doesn't like zero-sized allocations! */
397 		return;
398 
399 	// set up OS data structures to hold error info
400 	IA64_LOG_ALLOCATE(sal_info_type, max_size);
401 	memset(IA64_LOG_CURR_BUFFER(sal_info_type), 0, max_size);
402 	memset(IA64_LOG_NEXT_BUFFER(sal_info_type), 0, max_size);
403 }
404 
405 /*
406  * ia64_log_get
407  *
408  *	Get the current MCA log from SAL and copy it into the OS log buffer.
409  *
410  *  Inputs  :   info_type   (SAL_INFO_TYPE_{MCA,INIT,CMC,CPE})
411  *              irq_safe    whether you can use printk at this point
412  *  Outputs :   size        (total record length)
413  *              *buffer     (ptr to error record)
414  *
415  */
416 static u64
417 ia64_log_get(int sal_info_type, u8 **buffer, int irq_safe)
418 {
419 	sal_log_record_header_t     *log_buffer;
420 	u64                         total_len = 0;
421 	unsigned long               s;
422 
423 	IA64_LOG_LOCK(sal_info_type);
424 
425 	/* Get the process state information */
426 	log_buffer = IA64_LOG_NEXT_BUFFER(sal_info_type);
427 
428 	total_len = ia64_sal_get_state_info(sal_info_type, (u64 *)log_buffer);
429 
430 	if (total_len) {
431 		IA64_LOG_INDEX_INC(sal_info_type);
432 		IA64_LOG_UNLOCK(sal_info_type);
433 		if (irq_safe) {
434 			IA64_MCA_DEBUG("%s: SAL error record type %d retrieved. Record length = %ld\n",
435 				       __func__, sal_info_type, total_len);
436 		}
437 		*buffer = (u8 *) log_buffer;
438 		return total_len;
439 	} else {
440 		IA64_LOG_UNLOCK(sal_info_type);
441 		return 0;
442 	}
443 }
444 
445 /*
446  *  ia64_mca_log_sal_error_record
447  *
448  *  This function retrieves a specified error record type from SAL
449  *  and wakes up any processes waiting for error records.
450  *
451  *  Inputs  :   sal_info_type   (Type of error record MCA/CMC/CPE)
452  *              FIXME: remove MCA and irq_safe.
453  */
454 static void
455 ia64_mca_log_sal_error_record(int sal_info_type)
456 {
457 	u8 *buffer;
458 	sal_log_record_header_t *rh;
459 	u64 size;
460 	int irq_safe = sal_info_type != SAL_INFO_TYPE_MCA;
461 #ifdef IA64_MCA_DEBUG_INFO
462 	static const char * const rec_name[] = { "MCA", "INIT", "CMC", "CPE" };
463 #endif
464 
465 	size = ia64_log_get(sal_info_type, &buffer, irq_safe);
466 	if (!size)
467 		return;
468 
469 	salinfo_log_wakeup(sal_info_type, buffer, size, irq_safe);
470 
471 	if (irq_safe)
472 		IA64_MCA_DEBUG("CPU %d: SAL log contains %s error record\n",
473 			smp_processor_id(),
474 			sal_info_type < ARRAY_SIZE(rec_name) ? rec_name[sal_info_type] : "UNKNOWN");
475 
476 	/* Clear logs from corrected errors in case there's no user-level logger */
477 	rh = (sal_log_record_header_t *)buffer;
478 	if (rh->severity == sal_log_severity_corrected)
479 		ia64_sal_clear_state_info(sal_info_type);
480 }
481 
482 /*
483  * search_mca_table
484  *  See if the MCA surfaced in an instruction range
485  *  that has been tagged as recoverable.
486  *
487  *  Inputs
488  *	first	First address range to check
489  *	last	Last address range to check
490  *	ip	Instruction pointer, address we are looking for
491  *
492  * Return value:
493  *      1 on Success (in the table)/ 0 on Failure (not in the  table)
494  */
495 int
496 search_mca_table (const struct mca_table_entry *first,
497                 const struct mca_table_entry *last,
498                 unsigned long ip)
499 {
500         const struct mca_table_entry *curr;
501         u64 curr_start, curr_end;
502 
503         curr = first;
504         while (curr <= last) {
505                 curr_start = (u64) &curr->start_addr + curr->start_addr;
506                 curr_end = (u64) &curr->end_addr + curr->end_addr;
507 
508                 if ((ip >= curr_start) && (ip <= curr_end)) {
509                         return 1;
510                 }
511                 curr++;
512         }
513         return 0;
514 }
515 
516 /* Given an address, look for it in the mca tables. */
517 int mca_recover_range(unsigned long addr)
518 {
519 	extern struct mca_table_entry __start___mca_table[];
520 	extern struct mca_table_entry __stop___mca_table[];
521 
522 	return search_mca_table(__start___mca_table, __stop___mca_table-1, addr);
523 }
524 EXPORT_SYMBOL_GPL(mca_recover_range);
525 
526 #ifdef CONFIG_ACPI
527 
528 int cpe_vector = -1;
529 int ia64_cpe_irq = -1;
530 
531 static irqreturn_t
532 ia64_mca_cpe_int_handler (int cpe_irq, void *arg)
533 {
534 	static unsigned long	cpe_history[CPE_HISTORY_LENGTH];
535 	static int		index;
536 	static DEFINE_SPINLOCK(cpe_history_lock);
537 
538 	IA64_MCA_DEBUG("%s: received interrupt vector = %#x on CPU %d\n",
539 		       __func__, cpe_irq, smp_processor_id());
540 
541 	/* SAL spec states this should run w/ interrupts enabled */
542 	local_irq_enable();
543 
544 	spin_lock(&cpe_history_lock);
545 	if (!cpe_poll_enabled && cpe_vector >= 0) {
546 
547 		int i, count = 1; /* we know 1 happened now */
548 		unsigned long now = jiffies;
549 
550 		for (i = 0; i < CPE_HISTORY_LENGTH; i++) {
551 			if (now - cpe_history[i] <= HZ)
552 				count++;
553 		}
554 
555 		IA64_MCA_DEBUG(KERN_INFO "CPE threshold %d/%d\n", count, CPE_HISTORY_LENGTH);
556 		if (count >= CPE_HISTORY_LENGTH) {
557 
558 			cpe_poll_enabled = 1;
559 			spin_unlock(&cpe_history_lock);
560 			disable_irq_nosync(local_vector_to_irq(IA64_CPE_VECTOR));
561 
562 			/*
563 			 * Corrected errors will still be corrected, but
564 			 * make sure there's a log somewhere that indicates
565 			 * something is generating more than we can handle.
566 			 */
567 			printk(KERN_WARNING "WARNING: Switching to polling CPE handler; error records may be lost\n");
568 
569 			mod_timer(&cpe_poll_timer, jiffies + MIN_CPE_POLL_INTERVAL);
570 
571 			/* lock already released, get out now */
572 			goto out;
573 		} else {
574 			cpe_history[index++] = now;
575 			if (index == CPE_HISTORY_LENGTH)
576 				index = 0;
577 		}
578 	}
579 	spin_unlock(&cpe_history_lock);
580 out:
581 	/* Get the CPE error record and log it */
582 	ia64_mca_log_sal_error_record(SAL_INFO_TYPE_CPE);
583 
584 	return IRQ_HANDLED;
585 }
586 
587 #endif /* CONFIG_ACPI */
588 
589 #ifdef CONFIG_ACPI
590 /*
591  * ia64_mca_register_cpev
592  *
593  *  Register the corrected platform error vector with SAL.
594  *
595  *  Inputs
596  *      cpev        Corrected Platform Error Vector number
597  *
598  *  Outputs
599  *      None
600  */
601 void
602 ia64_mca_register_cpev (int cpev)
603 {
604 	/* Register the CPE interrupt vector with SAL */
605 	struct ia64_sal_retval isrv;
606 
607 	isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_CPE_INT, SAL_MC_PARAM_MECHANISM_INT, cpev, 0, 0);
608 	if (isrv.status) {
609 		printk(KERN_ERR "Failed to register Corrected Platform "
610 		       "Error interrupt vector with SAL (status %ld)\n", isrv.status);
611 		return;
612 	}
613 
614 	IA64_MCA_DEBUG("%s: corrected platform error "
615 		       "vector %#x registered\n", __func__, cpev);
616 }
617 #endif /* CONFIG_ACPI */
618 
619 /*
620  * ia64_mca_cmc_vector_setup
621  *
622  *  Setup the corrected machine check vector register in the processor.
623  *  (The interrupt is masked on boot. ia64_mca_late_init unmask this.)
624  *  This function is invoked on a per-processor basis.
625  *
626  * Inputs
627  *      None
628  *
629  * Outputs
630  *	None
631  */
632 void __cpuinit
633 ia64_mca_cmc_vector_setup (void)
634 {
635 	cmcv_reg_t	cmcv;
636 
637 	cmcv.cmcv_regval	= 0;
638 	cmcv.cmcv_mask		= 1;        /* Mask/disable interrupt at first */
639 	cmcv.cmcv_vector	= IA64_CMC_VECTOR;
640 	ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
641 
642 	IA64_MCA_DEBUG("%s: CPU %d corrected machine check vector %#x registered.\n",
643 		       __func__, smp_processor_id(), IA64_CMC_VECTOR);
644 
645 	IA64_MCA_DEBUG("%s: CPU %d CMCV = %#016lx\n",
646 		       __func__, smp_processor_id(), ia64_getreg(_IA64_REG_CR_CMCV));
647 }
648 
649 /*
650  * ia64_mca_cmc_vector_disable
651  *
652  *  Mask the corrected machine check vector register in the processor.
653  *  This function is invoked on a per-processor basis.
654  *
655  * Inputs
656  *      dummy(unused)
657  *
658  * Outputs
659  *	None
660  */
661 static void
662 ia64_mca_cmc_vector_disable (void *dummy)
663 {
664 	cmcv_reg_t	cmcv;
665 
666 	cmcv.cmcv_regval = ia64_getreg(_IA64_REG_CR_CMCV);
667 
668 	cmcv.cmcv_mask = 1; /* Mask/disable interrupt */
669 	ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
670 
671 	IA64_MCA_DEBUG("%s: CPU %d corrected machine check vector %#x disabled.\n",
672 		       __func__, smp_processor_id(), cmcv.cmcv_vector);
673 }
674 
675 /*
676  * ia64_mca_cmc_vector_enable
677  *
678  *  Unmask the corrected machine check vector register in the processor.
679  *  This function is invoked on a per-processor basis.
680  *
681  * Inputs
682  *      dummy(unused)
683  *
684  * Outputs
685  *	None
686  */
687 static void
688 ia64_mca_cmc_vector_enable (void *dummy)
689 {
690 	cmcv_reg_t	cmcv;
691 
692 	cmcv.cmcv_regval = ia64_getreg(_IA64_REG_CR_CMCV);
693 
694 	cmcv.cmcv_mask = 0; /* Unmask/enable interrupt */
695 	ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
696 
697 	IA64_MCA_DEBUG("%s: CPU %d corrected machine check vector %#x enabled.\n",
698 		       __func__, smp_processor_id(), cmcv.cmcv_vector);
699 }
700 
701 /*
702  * ia64_mca_cmc_vector_disable_keventd
703  *
704  * Called via keventd (smp_call_function() is not safe in interrupt context) to
705  * disable the cmc interrupt vector.
706  */
707 static void
708 ia64_mca_cmc_vector_disable_keventd(struct work_struct *unused)
709 {
710 	on_each_cpu(ia64_mca_cmc_vector_disable, NULL, 0);
711 }
712 
713 /*
714  * ia64_mca_cmc_vector_enable_keventd
715  *
716  * Called via keventd (smp_call_function() is not safe in interrupt context) to
717  * enable the cmc interrupt vector.
718  */
719 static void
720 ia64_mca_cmc_vector_enable_keventd(struct work_struct *unused)
721 {
722 	on_each_cpu(ia64_mca_cmc_vector_enable, NULL, 0);
723 }
724 
725 /*
726  * ia64_mca_wakeup
727  *
728  *	Send an inter-cpu interrupt to wake-up a particular cpu.
729  *
730  *  Inputs  :   cpuid
731  *  Outputs :   None
732  */
733 static void
734 ia64_mca_wakeup(int cpu)
735 {
736 	platform_send_ipi(cpu, IA64_MCA_WAKEUP_VECTOR, IA64_IPI_DM_INT, 0);
737 }
738 
739 /*
740  * ia64_mca_wakeup_all
741  *
742  *	Wakeup all the slave cpus which have rendez'ed previously.
743  *
744  *  Inputs  :   None
745  *  Outputs :   None
746  */
747 static void
748 ia64_mca_wakeup_all(void)
749 {
750 	int cpu;
751 
752 	/* Clear the Rendez checkin flag for all cpus */
753 	for_each_online_cpu(cpu) {
754 		if (ia64_mc_info.imi_rendez_checkin[cpu] == IA64_MCA_RENDEZ_CHECKIN_DONE)
755 			ia64_mca_wakeup(cpu);
756 	}
757 
758 }
759 
760 /*
761  * ia64_mca_rendez_interrupt_handler
762  *
763  *	This is handler used to put slave processors into spinloop
764  *	while the monarch processor does the mca handling and later
765  *	wake each slave up once the monarch is done.  The state
766  *	IA64_MCA_RENDEZ_CHECKIN_DONE indicates the cpu is rendez'ed
767  *	in SAL.  The state IA64_MCA_RENDEZ_CHECKIN_NOTDONE indicates
768  *	the cpu has come out of OS rendezvous.
769  *
770  *  Inputs  :   None
771  *  Outputs :   None
772  */
773 static irqreturn_t
774 ia64_mca_rendez_int_handler(int rendez_irq, void *arg)
775 {
776 	unsigned long flags;
777 	int cpu = smp_processor_id();
778 	struct ia64_mca_notify_die nd =
779 		{ .sos = NULL, .monarch_cpu = &monarch_cpu };
780 
781 	/* Mask all interrupts */
782 	local_irq_save(flags);
783 
784 	NOTIFY_MCA(DIE_MCA_RENDZVOUS_ENTER, get_irq_regs(), (long)&nd, 1);
785 
786 	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_DONE;
787 	/* Register with the SAL monarch that the slave has
788 	 * reached SAL
789 	 */
790 	ia64_sal_mc_rendez();
791 
792 	NOTIFY_MCA(DIE_MCA_RENDZVOUS_PROCESS, get_irq_regs(), (long)&nd, 1);
793 
794 	/* Wait for the monarch cpu to exit. */
795 	while (monarch_cpu != -1)
796 	       cpu_relax();	/* spin until monarch leaves */
797 
798 	NOTIFY_MCA(DIE_MCA_RENDZVOUS_LEAVE, get_irq_regs(), (long)&nd, 1);
799 
800 	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
801 	/* Enable all interrupts */
802 	local_irq_restore(flags);
803 	return IRQ_HANDLED;
804 }
805 
806 /*
807  * ia64_mca_wakeup_int_handler
808  *
809  *	The interrupt handler for processing the inter-cpu interrupt to the
810  *	slave cpu which was spinning in the rendez loop.
811  *	Since this spinning is done by turning off the interrupts and
812  *	polling on the wakeup-interrupt bit in the IRR, there is
813  *	nothing useful to be done in the handler.
814  *
815  *  Inputs  :   wakeup_irq  (Wakeup-interrupt bit)
816  *	arg		(Interrupt handler specific argument)
817  *  Outputs :   None
818  *
819  */
820 static irqreturn_t
821 ia64_mca_wakeup_int_handler(int wakeup_irq, void *arg)
822 {
823 	return IRQ_HANDLED;
824 }
825 
826 /* Function pointer for extra MCA recovery */
827 int (*ia64_mca_ucmc_extension)
828 	(void*,struct ia64_sal_os_state*)
829 	= NULL;
830 
831 int
832 ia64_reg_MCA_extension(int (*fn)(void *, struct ia64_sal_os_state *))
833 {
834 	if (ia64_mca_ucmc_extension)
835 		return 1;
836 
837 	ia64_mca_ucmc_extension = fn;
838 	return 0;
839 }
840 
841 void
842 ia64_unreg_MCA_extension(void)
843 {
844 	if (ia64_mca_ucmc_extension)
845 		ia64_mca_ucmc_extension = NULL;
846 }
847 
848 EXPORT_SYMBOL(ia64_reg_MCA_extension);
849 EXPORT_SYMBOL(ia64_unreg_MCA_extension);
850 
851 
852 static inline void
853 copy_reg(const u64 *fr, u64 fnat, unsigned long *tr, unsigned long *tnat)
854 {
855 	u64 fslot, tslot, nat;
856 	*tr = *fr;
857 	fslot = ((unsigned long)fr >> 3) & 63;
858 	tslot = ((unsigned long)tr >> 3) & 63;
859 	*tnat &= ~(1UL << tslot);
860 	nat = (fnat >> fslot) & 1;
861 	*tnat |= (nat << tslot);
862 }
863 
864 /* Change the comm field on the MCA/INT task to include the pid that
865  * was interrupted, it makes for easier debugging.  If that pid was 0
866  * (swapper or nested MCA/INIT) then use the start of the previous comm
867  * field suffixed with its cpu.
868  */
869 
870 static void
871 ia64_mca_modify_comm(const struct task_struct *previous_current)
872 {
873 	char *p, comm[sizeof(current->comm)];
874 	if (previous_current->pid)
875 		snprintf(comm, sizeof(comm), "%s %d",
876 			current->comm, previous_current->pid);
877 	else {
878 		int l;
879 		if ((p = strchr(previous_current->comm, ' ')))
880 			l = p - previous_current->comm;
881 		else
882 			l = strlen(previous_current->comm);
883 		snprintf(comm, sizeof(comm), "%s %*s %d",
884 			current->comm, l, previous_current->comm,
885 			task_thread_info(previous_current)->cpu);
886 	}
887 	memcpy(current->comm, comm, sizeof(current->comm));
888 }
889 
890 static void
891 finish_pt_regs(struct pt_regs *regs, struct ia64_sal_os_state *sos,
892 		unsigned long *nat)
893 {
894 	const pal_min_state_area_t *ms = sos->pal_min_state;
895 	const u64 *bank;
896 
897 	/* If ipsr.ic then use pmsa_{iip,ipsr,ifs}, else use
898 	 * pmsa_{xip,xpsr,xfs}
899 	 */
900 	if (ia64_psr(regs)->ic) {
901 		regs->cr_iip = ms->pmsa_iip;
902 		regs->cr_ipsr = ms->pmsa_ipsr;
903 		regs->cr_ifs = ms->pmsa_ifs;
904 	} else {
905 		regs->cr_iip = ms->pmsa_xip;
906 		regs->cr_ipsr = ms->pmsa_xpsr;
907 		regs->cr_ifs = ms->pmsa_xfs;
908 
909 		sos->iip = ms->pmsa_iip;
910 		sos->ipsr = ms->pmsa_ipsr;
911 		sos->ifs = ms->pmsa_ifs;
912 	}
913 	regs->pr = ms->pmsa_pr;
914 	regs->b0 = ms->pmsa_br0;
915 	regs->ar_rsc = ms->pmsa_rsc;
916 	copy_reg(&ms->pmsa_gr[1-1], ms->pmsa_nat_bits, &regs->r1, nat);
917 	copy_reg(&ms->pmsa_gr[2-1], ms->pmsa_nat_bits, &regs->r2, nat);
918 	copy_reg(&ms->pmsa_gr[3-1], ms->pmsa_nat_bits, &regs->r3, nat);
919 	copy_reg(&ms->pmsa_gr[8-1], ms->pmsa_nat_bits, &regs->r8, nat);
920 	copy_reg(&ms->pmsa_gr[9-1], ms->pmsa_nat_bits, &regs->r9, nat);
921 	copy_reg(&ms->pmsa_gr[10-1], ms->pmsa_nat_bits, &regs->r10, nat);
922 	copy_reg(&ms->pmsa_gr[11-1], ms->pmsa_nat_bits, &regs->r11, nat);
923 	copy_reg(&ms->pmsa_gr[12-1], ms->pmsa_nat_bits, &regs->r12, nat);
924 	copy_reg(&ms->pmsa_gr[13-1], ms->pmsa_nat_bits, &regs->r13, nat);
925 	copy_reg(&ms->pmsa_gr[14-1], ms->pmsa_nat_bits, &regs->r14, nat);
926 	copy_reg(&ms->pmsa_gr[15-1], ms->pmsa_nat_bits, &regs->r15, nat);
927 	if (ia64_psr(regs)->bn)
928 		bank = ms->pmsa_bank1_gr;
929 	else
930 		bank = ms->pmsa_bank0_gr;
931 	copy_reg(&bank[16-16], ms->pmsa_nat_bits, &regs->r16, nat);
932 	copy_reg(&bank[17-16], ms->pmsa_nat_bits, &regs->r17, nat);
933 	copy_reg(&bank[18-16], ms->pmsa_nat_bits, &regs->r18, nat);
934 	copy_reg(&bank[19-16], ms->pmsa_nat_bits, &regs->r19, nat);
935 	copy_reg(&bank[20-16], ms->pmsa_nat_bits, &regs->r20, nat);
936 	copy_reg(&bank[21-16], ms->pmsa_nat_bits, &regs->r21, nat);
937 	copy_reg(&bank[22-16], ms->pmsa_nat_bits, &regs->r22, nat);
938 	copy_reg(&bank[23-16], ms->pmsa_nat_bits, &regs->r23, nat);
939 	copy_reg(&bank[24-16], ms->pmsa_nat_bits, &regs->r24, nat);
940 	copy_reg(&bank[25-16], ms->pmsa_nat_bits, &regs->r25, nat);
941 	copy_reg(&bank[26-16], ms->pmsa_nat_bits, &regs->r26, nat);
942 	copy_reg(&bank[27-16], ms->pmsa_nat_bits, &regs->r27, nat);
943 	copy_reg(&bank[28-16], ms->pmsa_nat_bits, &regs->r28, nat);
944 	copy_reg(&bank[29-16], ms->pmsa_nat_bits, &regs->r29, nat);
945 	copy_reg(&bank[30-16], ms->pmsa_nat_bits, &regs->r30, nat);
946 	copy_reg(&bank[31-16], ms->pmsa_nat_bits, &regs->r31, nat);
947 }
948 
949 /* On entry to this routine, we are running on the per cpu stack, see
950  * mca_asm.h.  The original stack has not been touched by this event.  Some of
951  * the original stack's registers will be in the RBS on this stack.  This stack
952  * also contains a partial pt_regs and switch_stack, the rest of the data is in
953  * PAL minstate.
954  *
955  * The first thing to do is modify the original stack to look like a blocked
956  * task so we can run backtrace on the original task.  Also mark the per cpu
957  * stack as current to ensure that we use the correct task state, it also means
958  * that we can do backtrace on the MCA/INIT handler code itself.
959  */
960 
961 static struct task_struct *
962 ia64_mca_modify_original_stack(struct pt_regs *regs,
963 		const struct switch_stack *sw,
964 		struct ia64_sal_os_state *sos,
965 		const char *type)
966 {
967 	char *p;
968 	ia64_va va;
969 	extern char ia64_leave_kernel[];	/* Need asm address, not function descriptor */
970 	const pal_min_state_area_t *ms = sos->pal_min_state;
971 	struct task_struct *previous_current;
972 	struct pt_regs *old_regs;
973 	struct switch_stack *old_sw;
974 	unsigned size = sizeof(struct pt_regs) +
975 			sizeof(struct switch_stack) + 16;
976 	unsigned long *old_bspstore, *old_bsp;
977 	unsigned long *new_bspstore, *new_bsp;
978 	unsigned long old_unat, old_rnat, new_rnat, nat;
979 	u64 slots, loadrs = regs->loadrs;
980 	u64 r12 = ms->pmsa_gr[12-1], r13 = ms->pmsa_gr[13-1];
981 	u64 ar_bspstore = regs->ar_bspstore;
982 	u64 ar_bsp = regs->ar_bspstore + (loadrs >> 16);
983 	const char *msg;
984 	int cpu = smp_processor_id();
985 
986 	previous_current = curr_task(cpu);
987 	set_curr_task(cpu, current);
988 	if ((p = strchr(current->comm, ' ')))
989 		*p = '\0';
990 
991 	/* Best effort attempt to cope with MCA/INIT delivered while in
992 	 * physical mode.
993 	 */
994 	regs->cr_ipsr = ms->pmsa_ipsr;
995 	if (ia64_psr(regs)->dt == 0) {
996 		va.l = r12;
997 		if (va.f.reg == 0) {
998 			va.f.reg = 7;
999 			r12 = va.l;
1000 		}
1001 		va.l = r13;
1002 		if (va.f.reg == 0) {
1003 			va.f.reg = 7;
1004 			r13 = va.l;
1005 		}
1006 	}
1007 	if (ia64_psr(regs)->rt == 0) {
1008 		va.l = ar_bspstore;
1009 		if (va.f.reg == 0) {
1010 			va.f.reg = 7;
1011 			ar_bspstore = va.l;
1012 		}
1013 		va.l = ar_bsp;
1014 		if (va.f.reg == 0) {
1015 			va.f.reg = 7;
1016 			ar_bsp = va.l;
1017 		}
1018 	}
1019 
1020 	/* mca_asm.S ia64_old_stack() cannot assume that the dirty registers
1021 	 * have been copied to the old stack, the old stack may fail the
1022 	 * validation tests below.  So ia64_old_stack() must restore the dirty
1023 	 * registers from the new stack.  The old and new bspstore probably
1024 	 * have different alignments, so loadrs calculated on the old bsp
1025 	 * cannot be used to restore from the new bsp.  Calculate a suitable
1026 	 * loadrs for the new stack and save it in the new pt_regs, where
1027 	 * ia64_old_stack() can get it.
1028 	 */
1029 	old_bspstore = (unsigned long *)ar_bspstore;
1030 	old_bsp = (unsigned long *)ar_bsp;
1031 	slots = ia64_rse_num_regs(old_bspstore, old_bsp);
1032 	new_bspstore = (unsigned long *)((u64)current + IA64_RBS_OFFSET);
1033 	new_bsp = ia64_rse_skip_regs(new_bspstore, slots);
1034 	regs->loadrs = (new_bsp - new_bspstore) * 8 << 16;
1035 
1036 	/* Verify the previous stack state before we change it */
1037 	if (user_mode(regs)) {
1038 		msg = "occurred in user space";
1039 		/* previous_current is guaranteed to be valid when the task was
1040 		 * in user space, so ...
1041 		 */
1042 		ia64_mca_modify_comm(previous_current);
1043 		goto no_mod;
1044 	}
1045 
1046 	if (r13 != sos->prev_IA64_KR_CURRENT) {
1047 		msg = "inconsistent previous current and r13";
1048 		goto no_mod;
1049 	}
1050 
1051 	if (!mca_recover_range(ms->pmsa_iip)) {
1052 		if ((r12 - r13) >= KERNEL_STACK_SIZE) {
1053 			msg = "inconsistent r12 and r13";
1054 			goto no_mod;
1055 		}
1056 		if ((ar_bspstore - r13) >= KERNEL_STACK_SIZE) {
1057 			msg = "inconsistent ar.bspstore and r13";
1058 			goto no_mod;
1059 		}
1060 		va.p = old_bspstore;
1061 		if (va.f.reg < 5) {
1062 			msg = "old_bspstore is in the wrong region";
1063 			goto no_mod;
1064 		}
1065 		if ((ar_bsp - r13) >= KERNEL_STACK_SIZE) {
1066 			msg = "inconsistent ar.bsp and r13";
1067 			goto no_mod;
1068 		}
1069 		size += (ia64_rse_skip_regs(old_bspstore, slots) - old_bspstore) * 8;
1070 		if (ar_bspstore + size > r12) {
1071 			msg = "no room for blocked state";
1072 			goto no_mod;
1073 		}
1074 	}
1075 
1076 	ia64_mca_modify_comm(previous_current);
1077 
1078 	/* Make the original task look blocked.  First stack a struct pt_regs,
1079 	 * describing the state at the time of interrupt.  mca_asm.S built a
1080 	 * partial pt_regs, copy it and fill in the blanks using minstate.
1081 	 */
1082 	p = (char *)r12 - sizeof(*regs);
1083 	old_regs = (struct pt_regs *)p;
1084 	memcpy(old_regs, regs, sizeof(*regs));
1085 	old_regs->loadrs = loadrs;
1086 	old_unat = old_regs->ar_unat;
1087 	finish_pt_regs(old_regs, sos, &old_unat);
1088 
1089 	/* Next stack a struct switch_stack.  mca_asm.S built a partial
1090 	 * switch_stack, copy it and fill in the blanks using pt_regs and
1091 	 * minstate.
1092 	 *
1093 	 * In the synthesized switch_stack, b0 points to ia64_leave_kernel,
1094 	 * ar.pfs is set to 0.
1095 	 *
1096 	 * unwind.c::unw_unwind() does special processing for interrupt frames.
1097 	 * It checks if the PRED_NON_SYSCALL predicate is set, if the predicate
1098 	 * is clear then unw_unwind() does _not_ adjust bsp over pt_regs.  Not
1099 	 * that this is documented, of course.  Set PRED_NON_SYSCALL in the
1100 	 * switch_stack on the original stack so it will unwind correctly when
1101 	 * unwind.c reads pt_regs.
1102 	 *
1103 	 * thread.ksp is updated to point to the synthesized switch_stack.
1104 	 */
1105 	p -= sizeof(struct switch_stack);
1106 	old_sw = (struct switch_stack *)p;
1107 	memcpy(old_sw, sw, sizeof(*sw));
1108 	old_sw->caller_unat = old_unat;
1109 	old_sw->ar_fpsr = old_regs->ar_fpsr;
1110 	copy_reg(&ms->pmsa_gr[4-1], ms->pmsa_nat_bits, &old_sw->r4, &old_unat);
1111 	copy_reg(&ms->pmsa_gr[5-1], ms->pmsa_nat_bits, &old_sw->r5, &old_unat);
1112 	copy_reg(&ms->pmsa_gr[6-1], ms->pmsa_nat_bits, &old_sw->r6, &old_unat);
1113 	copy_reg(&ms->pmsa_gr[7-1], ms->pmsa_nat_bits, &old_sw->r7, &old_unat);
1114 	old_sw->b0 = (u64)ia64_leave_kernel;
1115 	old_sw->b1 = ms->pmsa_br1;
1116 	old_sw->ar_pfs = 0;
1117 	old_sw->ar_unat = old_unat;
1118 	old_sw->pr = old_regs->pr | (1UL << PRED_NON_SYSCALL);
1119 	previous_current->thread.ksp = (u64)p - 16;
1120 
1121 	/* Finally copy the original stack's registers back to its RBS.
1122 	 * Registers from ar.bspstore through ar.bsp at the time of the event
1123 	 * are in the current RBS, copy them back to the original stack.  The
1124 	 * copy must be done register by register because the original bspstore
1125 	 * and the current one have different alignments, so the saved RNAT
1126 	 * data occurs at different places.
1127 	 *
1128 	 * mca_asm does cover, so the old_bsp already includes all registers at
1129 	 * the time of MCA/INIT.  It also does flushrs, so all registers before
1130 	 * this function have been written to backing store on the MCA/INIT
1131 	 * stack.
1132 	 */
1133 	new_rnat = ia64_get_rnat(ia64_rse_rnat_addr(new_bspstore));
1134 	old_rnat = regs->ar_rnat;
1135 	while (slots--) {
1136 		if (ia64_rse_is_rnat_slot(new_bspstore)) {
1137 			new_rnat = ia64_get_rnat(new_bspstore++);
1138 		}
1139 		if (ia64_rse_is_rnat_slot(old_bspstore)) {
1140 			*old_bspstore++ = old_rnat;
1141 			old_rnat = 0;
1142 		}
1143 		nat = (new_rnat >> ia64_rse_slot_num(new_bspstore)) & 1UL;
1144 		old_rnat &= ~(1UL << ia64_rse_slot_num(old_bspstore));
1145 		old_rnat |= (nat << ia64_rse_slot_num(old_bspstore));
1146 		*old_bspstore++ = *new_bspstore++;
1147 	}
1148 	old_sw->ar_bspstore = (unsigned long)old_bspstore;
1149 	old_sw->ar_rnat = old_rnat;
1150 
1151 	sos->prev_task = previous_current;
1152 	return previous_current;
1153 
1154 no_mod:
1155 	mprintk(KERN_INFO "cpu %d, %s %s, original stack not modified\n",
1156 			smp_processor_id(), type, msg);
1157 	old_unat = regs->ar_unat;
1158 	finish_pt_regs(regs, sos, &old_unat);
1159 	return previous_current;
1160 }
1161 
1162 /* The monarch/slave interaction is based on monarch_cpu and requires that all
1163  * slaves have entered rendezvous before the monarch leaves.  If any cpu has
1164  * not entered rendezvous yet then wait a bit.  The assumption is that any
1165  * slave that has not rendezvoused after a reasonable time is never going to do
1166  * so.  In this context, slave includes cpus that respond to the MCA rendezvous
1167  * interrupt, as well as cpus that receive the INIT slave event.
1168  */
1169 
1170 static void
1171 ia64_wait_for_slaves(int monarch, const char *type)
1172 {
1173 	int c, i , wait;
1174 
1175 	/*
1176 	 * wait 5 seconds total for slaves (arbitrary)
1177 	 */
1178 	for (i = 0; i < 5000; i++) {
1179 		wait = 0;
1180 		for_each_online_cpu(c) {
1181 			if (c == monarch)
1182 				continue;
1183 			if (ia64_mc_info.imi_rendez_checkin[c]
1184 					== IA64_MCA_RENDEZ_CHECKIN_NOTDONE) {
1185 				udelay(1000);		/* short wait */
1186 				wait = 1;
1187 				break;
1188 			}
1189 		}
1190 		if (!wait)
1191 			goto all_in;
1192 	}
1193 
1194 	/*
1195 	 * Maybe slave(s) dead. Print buffered messages immediately.
1196 	 */
1197 	ia64_mlogbuf_finish(0);
1198 	mprintk(KERN_INFO "OS %s slave did not rendezvous on cpu", type);
1199 	for_each_online_cpu(c) {
1200 		if (c == monarch)
1201 			continue;
1202 		if (ia64_mc_info.imi_rendez_checkin[c] == IA64_MCA_RENDEZ_CHECKIN_NOTDONE)
1203 			mprintk(" %d", c);
1204 	}
1205 	mprintk("\n");
1206 	return;
1207 
1208 all_in:
1209 	mprintk(KERN_INFO "All OS %s slaves have reached rendezvous\n", type);
1210 	return;
1211 }
1212 
1213 /*  mca_insert_tr
1214  *
1215  *  Switch rid when TR reload and needed!
1216  *  iord: 1: itr, 2: itr;
1217  *
1218 */
1219 static void mca_insert_tr(u64 iord)
1220 {
1221 
1222 	int i;
1223 	u64 old_rr;
1224 	struct ia64_tr_entry *p;
1225 	unsigned long psr;
1226 	int cpu = smp_processor_id();
1227 
1228 	if (!ia64_idtrs[cpu])
1229 		return;
1230 
1231 	psr = ia64_clear_ic();
1232 	for (i = IA64_TR_ALLOC_BASE; i < IA64_TR_ALLOC_MAX; i++) {
1233 		p = ia64_idtrs[cpu] + (iord - 1) * IA64_TR_ALLOC_MAX;
1234 		if (p->pte & 0x1) {
1235 			old_rr = ia64_get_rr(p->ifa);
1236 			if (old_rr != p->rr) {
1237 				ia64_set_rr(p->ifa, p->rr);
1238 				ia64_srlz_d();
1239 			}
1240 			ia64_ptr(iord, p->ifa, p->itir >> 2);
1241 			ia64_srlz_i();
1242 			if (iord & 0x1) {
1243 				ia64_itr(0x1, i, p->ifa, p->pte, p->itir >> 2);
1244 				ia64_srlz_i();
1245 			}
1246 			if (iord & 0x2) {
1247 				ia64_itr(0x2, i, p->ifa, p->pte, p->itir >> 2);
1248 				ia64_srlz_i();
1249 			}
1250 			if (old_rr != p->rr) {
1251 				ia64_set_rr(p->ifa, old_rr);
1252 				ia64_srlz_d();
1253 			}
1254 		}
1255 	}
1256 	ia64_set_psr(psr);
1257 }
1258 
1259 /*
1260  * ia64_mca_handler
1261  *
1262  *	This is uncorrectable machine check handler called from OS_MCA
1263  *	dispatch code which is in turn called from SAL_CHECK().
1264  *	This is the place where the core of OS MCA handling is done.
1265  *	Right now the logs are extracted and displayed in a well-defined
1266  *	format. This handler code is supposed to be run only on the
1267  *	monarch processor. Once the monarch is done with MCA handling
1268  *	further MCA logging is enabled by clearing logs.
1269  *	Monarch also has the duty of sending wakeup-IPIs to pull the
1270  *	slave processors out of rendezvous spinloop.
1271  *
1272  *	If multiple processors call into OS_MCA, the first will become
1273  *	the monarch.  Subsequent cpus will be recorded in the mca_cpu
1274  *	bitmask.  After the first monarch has processed its MCA, it
1275  *	will wake up the next cpu in the mca_cpu bitmask and then go
1276  *	into the rendezvous loop.  When all processors have serviced
1277  *	their MCA, the last monarch frees up the rest of the processors.
1278  */
1279 void
1280 ia64_mca_handler(struct pt_regs *regs, struct switch_stack *sw,
1281 		 struct ia64_sal_os_state *sos)
1282 {
1283 	int recover, cpu = smp_processor_id();
1284 	struct task_struct *previous_current;
1285 	struct ia64_mca_notify_die nd =
1286 		{ .sos = sos, .monarch_cpu = &monarch_cpu, .data = &recover };
1287 	static atomic_t mca_count;
1288 	static cpumask_t mca_cpu;
1289 
1290 	if (atomic_add_return(1, &mca_count) == 1) {
1291 		monarch_cpu = cpu;
1292 		sos->monarch = 1;
1293 	} else {
1294 		cpu_set(cpu, mca_cpu);
1295 		sos->monarch = 0;
1296 	}
1297 	mprintk(KERN_INFO "Entered OS MCA handler. PSP=%lx cpu=%d "
1298 		"monarch=%ld\n", sos->proc_state_param, cpu, sos->monarch);
1299 
1300 	previous_current = ia64_mca_modify_original_stack(regs, sw, sos, "MCA");
1301 
1302 	NOTIFY_MCA(DIE_MCA_MONARCH_ENTER, regs, (long)&nd, 1);
1303 
1304 	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_CONCURRENT_MCA;
1305 	if (sos->monarch) {
1306 		ia64_wait_for_slaves(cpu, "MCA");
1307 
1308 		/* Wakeup all the processors which are spinning in the
1309 		 * rendezvous loop.  They will leave SAL, then spin in the OS
1310 		 * with interrupts disabled until this monarch cpu leaves the
1311 		 * MCA handler.  That gets control back to the OS so we can
1312 		 * backtrace the other cpus, backtrace when spinning in SAL
1313 		 * does not work.
1314 		 */
1315 		ia64_mca_wakeup_all();
1316 	} else {
1317 		while (cpu_isset(cpu, mca_cpu))
1318 			cpu_relax();	/* spin until monarch wakes us */
1319 	}
1320 
1321 	NOTIFY_MCA(DIE_MCA_MONARCH_PROCESS, regs, (long)&nd, 1);
1322 
1323 	/* Get the MCA error record and log it */
1324 	ia64_mca_log_sal_error_record(SAL_INFO_TYPE_MCA);
1325 
1326 	/* MCA error recovery */
1327 	recover = (ia64_mca_ucmc_extension
1328 		&& ia64_mca_ucmc_extension(
1329 			IA64_LOG_CURR_BUFFER(SAL_INFO_TYPE_MCA),
1330 			sos));
1331 
1332 	if (recover) {
1333 		sal_log_record_header_t *rh = IA64_LOG_CURR_BUFFER(SAL_INFO_TYPE_MCA);
1334 		rh->severity = sal_log_severity_corrected;
1335 		ia64_sal_clear_state_info(SAL_INFO_TYPE_MCA);
1336 		sos->os_status = IA64_MCA_CORRECTED;
1337 	} else {
1338 		/* Dump buffered message to console */
1339 		ia64_mlogbuf_finish(1);
1340 	}
1341 
1342 	if (__get_cpu_var(ia64_mca_tr_reload)) {
1343 		mca_insert_tr(0x1); /*Reload dynamic itrs*/
1344 		mca_insert_tr(0x2); /*Reload dynamic itrs*/
1345 	}
1346 
1347 	NOTIFY_MCA(DIE_MCA_MONARCH_LEAVE, regs, (long)&nd, 1);
1348 
1349 	if (atomic_dec_return(&mca_count) > 0) {
1350 		int i;
1351 
1352 		/* wake up the next monarch cpu,
1353 		 * and put this cpu in the rendez loop.
1354 		 */
1355 		for_each_online_cpu(i) {
1356 			if (cpu_isset(i, mca_cpu)) {
1357 				monarch_cpu = i;
1358 				cpu_clear(i, mca_cpu);	/* wake next cpu */
1359 				while (monarch_cpu != -1)
1360 					cpu_relax();	/* spin until last cpu leaves */
1361 				set_curr_task(cpu, previous_current);
1362 				ia64_mc_info.imi_rendez_checkin[cpu]
1363 						= IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1364 				return;
1365 			}
1366 		}
1367 	}
1368 	set_curr_task(cpu, previous_current);
1369 	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1370 	monarch_cpu = -1;	/* This frees the slaves and previous monarchs */
1371 }
1372 
1373 static DECLARE_WORK(cmc_disable_work, ia64_mca_cmc_vector_disable_keventd);
1374 static DECLARE_WORK(cmc_enable_work, ia64_mca_cmc_vector_enable_keventd);
1375 
1376 /*
1377  * ia64_mca_cmc_int_handler
1378  *
1379  *  This is corrected machine check interrupt handler.
1380  *	Right now the logs are extracted and displayed in a well-defined
1381  *	format.
1382  *
1383  * Inputs
1384  *      interrupt number
1385  *      client data arg ptr
1386  *
1387  * Outputs
1388  *	None
1389  */
1390 static irqreturn_t
1391 ia64_mca_cmc_int_handler(int cmc_irq, void *arg)
1392 {
1393 	static unsigned long	cmc_history[CMC_HISTORY_LENGTH];
1394 	static int		index;
1395 	static DEFINE_SPINLOCK(cmc_history_lock);
1396 
1397 	IA64_MCA_DEBUG("%s: received interrupt vector = %#x on CPU %d\n",
1398 		       __func__, cmc_irq, smp_processor_id());
1399 
1400 	/* SAL spec states this should run w/ interrupts enabled */
1401 	local_irq_enable();
1402 
1403 	spin_lock(&cmc_history_lock);
1404 	if (!cmc_polling_enabled) {
1405 		int i, count = 1; /* we know 1 happened now */
1406 		unsigned long now = jiffies;
1407 
1408 		for (i = 0; i < CMC_HISTORY_LENGTH; i++) {
1409 			if (now - cmc_history[i] <= HZ)
1410 				count++;
1411 		}
1412 
1413 		IA64_MCA_DEBUG(KERN_INFO "CMC threshold %d/%d\n", count, CMC_HISTORY_LENGTH);
1414 		if (count >= CMC_HISTORY_LENGTH) {
1415 
1416 			cmc_polling_enabled = 1;
1417 			spin_unlock(&cmc_history_lock);
1418 			/* If we're being hit with CMC interrupts, we won't
1419 			 * ever execute the schedule_work() below.  Need to
1420 			 * disable CMC interrupts on this processor now.
1421 			 */
1422 			ia64_mca_cmc_vector_disable(NULL);
1423 			schedule_work(&cmc_disable_work);
1424 
1425 			/*
1426 			 * Corrected errors will still be corrected, but
1427 			 * make sure there's a log somewhere that indicates
1428 			 * something is generating more than we can handle.
1429 			 */
1430 			printk(KERN_WARNING "WARNING: Switching to polling CMC handler; error records may be lost\n");
1431 
1432 			mod_timer(&cmc_poll_timer, jiffies + CMC_POLL_INTERVAL);
1433 
1434 			/* lock already released, get out now */
1435 			goto out;
1436 		} else {
1437 			cmc_history[index++] = now;
1438 			if (index == CMC_HISTORY_LENGTH)
1439 				index = 0;
1440 		}
1441 	}
1442 	spin_unlock(&cmc_history_lock);
1443 out:
1444 	/* Get the CMC error record and log it */
1445 	ia64_mca_log_sal_error_record(SAL_INFO_TYPE_CMC);
1446 
1447 	return IRQ_HANDLED;
1448 }
1449 
1450 /*
1451  *  ia64_mca_cmc_int_caller
1452  *
1453  * 	Triggered by sw interrupt from CMC polling routine.  Calls
1454  * 	real interrupt handler and either triggers a sw interrupt
1455  * 	on the next cpu or does cleanup at the end.
1456  *
1457  * Inputs
1458  *	interrupt number
1459  *	client data arg ptr
1460  * Outputs
1461  * 	handled
1462  */
1463 static irqreturn_t
1464 ia64_mca_cmc_int_caller(int cmc_irq, void *arg)
1465 {
1466 	static int start_count = -1;
1467 	unsigned int cpuid;
1468 
1469 	cpuid = smp_processor_id();
1470 
1471 	/* If first cpu, update count */
1472 	if (start_count == -1)
1473 		start_count = IA64_LOG_COUNT(SAL_INFO_TYPE_CMC);
1474 
1475 	ia64_mca_cmc_int_handler(cmc_irq, arg);
1476 
1477 	cpuid = cpumask_next(cpuid+1, cpu_online_mask);
1478 
1479 	if (cpuid < nr_cpu_ids) {
1480 		platform_send_ipi(cpuid, IA64_CMCP_VECTOR, IA64_IPI_DM_INT, 0);
1481 	} else {
1482 		/* If no log record, switch out of polling mode */
1483 		if (start_count == IA64_LOG_COUNT(SAL_INFO_TYPE_CMC)) {
1484 
1485 			printk(KERN_WARNING "Returning to interrupt driven CMC handler\n");
1486 			schedule_work(&cmc_enable_work);
1487 			cmc_polling_enabled = 0;
1488 
1489 		} else {
1490 
1491 			mod_timer(&cmc_poll_timer, jiffies + CMC_POLL_INTERVAL);
1492 		}
1493 
1494 		start_count = -1;
1495 	}
1496 
1497 	return IRQ_HANDLED;
1498 }
1499 
1500 /*
1501  *  ia64_mca_cmc_poll
1502  *
1503  *	Poll for Corrected Machine Checks (CMCs)
1504  *
1505  * Inputs   :   dummy(unused)
1506  * Outputs  :   None
1507  *
1508  */
1509 static void
1510 ia64_mca_cmc_poll (unsigned long dummy)
1511 {
1512 	/* Trigger a CMC interrupt cascade  */
1513 	platform_send_ipi(first_cpu(cpu_online_map), IA64_CMCP_VECTOR, IA64_IPI_DM_INT, 0);
1514 }
1515 
1516 /*
1517  *  ia64_mca_cpe_int_caller
1518  *
1519  * 	Triggered by sw interrupt from CPE polling routine.  Calls
1520  * 	real interrupt handler and either triggers a sw interrupt
1521  * 	on the next cpu or does cleanup at the end.
1522  *
1523  * Inputs
1524  *	interrupt number
1525  *	client data arg ptr
1526  * Outputs
1527  * 	handled
1528  */
1529 #ifdef CONFIG_ACPI
1530 
1531 static irqreturn_t
1532 ia64_mca_cpe_int_caller(int cpe_irq, void *arg)
1533 {
1534 	static int start_count = -1;
1535 	static int poll_time = MIN_CPE_POLL_INTERVAL;
1536 	unsigned int cpuid;
1537 
1538 	cpuid = smp_processor_id();
1539 
1540 	/* If first cpu, update count */
1541 	if (start_count == -1)
1542 		start_count = IA64_LOG_COUNT(SAL_INFO_TYPE_CPE);
1543 
1544 	ia64_mca_cpe_int_handler(cpe_irq, arg);
1545 
1546 	cpuid = cpumask_next(cpuid+1, cpu_online_mask);
1547 
1548 	if (cpuid < NR_CPUS) {
1549 		platform_send_ipi(cpuid, IA64_CPEP_VECTOR, IA64_IPI_DM_INT, 0);
1550 	} else {
1551 		/*
1552 		 * If a log was recorded, increase our polling frequency,
1553 		 * otherwise, backoff or return to interrupt mode.
1554 		 */
1555 		if (start_count != IA64_LOG_COUNT(SAL_INFO_TYPE_CPE)) {
1556 			poll_time = max(MIN_CPE_POLL_INTERVAL, poll_time / 2);
1557 		} else if (cpe_vector < 0) {
1558 			poll_time = min(MAX_CPE_POLL_INTERVAL, poll_time * 2);
1559 		} else {
1560 			poll_time = MIN_CPE_POLL_INTERVAL;
1561 
1562 			printk(KERN_WARNING "Returning to interrupt driven CPE handler\n");
1563 			enable_irq(local_vector_to_irq(IA64_CPE_VECTOR));
1564 			cpe_poll_enabled = 0;
1565 		}
1566 
1567 		if (cpe_poll_enabled)
1568 			mod_timer(&cpe_poll_timer, jiffies + poll_time);
1569 		start_count = -1;
1570 	}
1571 
1572 	return IRQ_HANDLED;
1573 }
1574 
1575 /*
1576  *  ia64_mca_cpe_poll
1577  *
1578  *	Poll for Corrected Platform Errors (CPEs), trigger interrupt
1579  *	on first cpu, from there it will trickle through all the cpus.
1580  *
1581  * Inputs   :   dummy(unused)
1582  * Outputs  :   None
1583  *
1584  */
1585 static void
1586 ia64_mca_cpe_poll (unsigned long dummy)
1587 {
1588 	/* Trigger a CPE interrupt cascade  */
1589 	platform_send_ipi(first_cpu(cpu_online_map), IA64_CPEP_VECTOR, IA64_IPI_DM_INT, 0);
1590 }
1591 
1592 #endif /* CONFIG_ACPI */
1593 
1594 static int
1595 default_monarch_init_process(struct notifier_block *self, unsigned long val, void *data)
1596 {
1597 	int c;
1598 	struct task_struct *g, *t;
1599 	if (val != DIE_INIT_MONARCH_PROCESS)
1600 		return NOTIFY_DONE;
1601 #ifdef CONFIG_KEXEC
1602 	if (atomic_read(&kdump_in_progress))
1603 		return NOTIFY_DONE;
1604 #endif
1605 
1606 	/*
1607 	 * FIXME: mlogbuf will brim over with INIT stack dumps.
1608 	 * To enable show_stack from INIT, we use oops_in_progress which should
1609 	 * be used in real oops. This would cause something wrong after INIT.
1610 	 */
1611 	BREAK_LOGLEVEL(console_loglevel);
1612 	ia64_mlogbuf_dump_from_init();
1613 
1614 	printk(KERN_ERR "Processes interrupted by INIT -");
1615 	for_each_online_cpu(c) {
1616 		struct ia64_sal_os_state *s;
1617 		t = __va(__per_cpu_mca[c] + IA64_MCA_CPU_INIT_STACK_OFFSET);
1618 		s = (struct ia64_sal_os_state *)((char *)t + MCA_SOS_OFFSET);
1619 		g = s->prev_task;
1620 		if (g) {
1621 			if (g->pid)
1622 				printk(" %d", g->pid);
1623 			else
1624 				printk(" %d (cpu %d task 0x%p)", g->pid, task_cpu(g), g);
1625 		}
1626 	}
1627 	printk("\n\n");
1628 	if (read_trylock(&tasklist_lock)) {
1629 		do_each_thread (g, t) {
1630 			printk("\nBacktrace of pid %d (%s)\n", t->pid, t->comm);
1631 			show_stack(t, NULL);
1632 		} while_each_thread (g, t);
1633 		read_unlock(&tasklist_lock);
1634 	}
1635 	/* FIXME: This will not restore zapped printk locks. */
1636 	RESTORE_LOGLEVEL(console_loglevel);
1637 	return NOTIFY_DONE;
1638 }
1639 
1640 /*
1641  * C portion of the OS INIT handler
1642  *
1643  * Called from ia64_os_init_dispatch
1644  *
1645  * Inputs: pointer to pt_regs where processor info was saved.  SAL/OS state for
1646  * this event.  This code is used for both monarch and slave INIT events, see
1647  * sos->monarch.
1648  *
1649  * All INIT events switch to the INIT stack and change the previous process to
1650  * blocked status.  If one of the INIT events is the monarch then we are
1651  * probably processing the nmi button/command.  Use the monarch cpu to dump all
1652  * the processes.  The slave INIT events all spin until the monarch cpu
1653  * returns.  We can also get INIT slave events for MCA, in which case the MCA
1654  * process is the monarch.
1655  */
1656 
1657 void
1658 ia64_init_handler(struct pt_regs *regs, struct switch_stack *sw,
1659 		  struct ia64_sal_os_state *sos)
1660 {
1661 	static atomic_t slaves;
1662 	static atomic_t monarchs;
1663 	struct task_struct *previous_current;
1664 	int cpu = smp_processor_id();
1665 	struct ia64_mca_notify_die nd =
1666 		{ .sos = sos, .monarch_cpu = &monarch_cpu };
1667 
1668 	NOTIFY_INIT(DIE_INIT_ENTER, regs, (long)&nd, 0);
1669 
1670 	mprintk(KERN_INFO "Entered OS INIT handler. PSP=%lx cpu=%d monarch=%ld\n",
1671 		sos->proc_state_param, cpu, sos->monarch);
1672 	salinfo_log_wakeup(SAL_INFO_TYPE_INIT, NULL, 0, 0);
1673 
1674 	previous_current = ia64_mca_modify_original_stack(regs, sw, sos, "INIT");
1675 	sos->os_status = IA64_INIT_RESUME;
1676 
1677 	/* FIXME: Workaround for broken proms that drive all INIT events as
1678 	 * slaves.  The last slave that enters is promoted to be a monarch.
1679 	 * Remove this code in September 2006, that gives platforms a year to
1680 	 * fix their proms and get their customers updated.
1681 	 */
1682 	if (!sos->monarch && atomic_add_return(1, &slaves) == num_online_cpus()) {
1683 		mprintk(KERN_WARNING "%s: Promoting cpu %d to monarch.\n",
1684 		        __func__, cpu);
1685 		atomic_dec(&slaves);
1686 		sos->monarch = 1;
1687 	}
1688 
1689 	/* FIXME: Workaround for broken proms that drive all INIT events as
1690 	 * monarchs.  Second and subsequent monarchs are demoted to slaves.
1691 	 * Remove this code in September 2006, that gives platforms a year to
1692 	 * fix their proms and get their customers updated.
1693 	 */
1694 	if (sos->monarch && atomic_add_return(1, &monarchs) > 1) {
1695 		mprintk(KERN_WARNING "%s: Demoting cpu %d to slave.\n",
1696 			       __func__, cpu);
1697 		atomic_dec(&monarchs);
1698 		sos->monarch = 0;
1699 	}
1700 
1701 	if (!sos->monarch) {
1702 		ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_INIT;
1703 
1704 #ifdef CONFIG_KEXEC
1705 		while (monarch_cpu == -1 && !atomic_read(&kdump_in_progress))
1706 			udelay(1000);
1707 #else
1708 		while (monarch_cpu == -1)
1709 			cpu_relax();	/* spin until monarch enters */
1710 #endif
1711 
1712 		NOTIFY_INIT(DIE_INIT_SLAVE_ENTER, regs, (long)&nd, 1);
1713 		NOTIFY_INIT(DIE_INIT_SLAVE_PROCESS, regs, (long)&nd, 1);
1714 
1715 #ifdef CONFIG_KEXEC
1716 		while (monarch_cpu != -1 && !atomic_read(&kdump_in_progress))
1717 			udelay(1000);
1718 #else
1719 		while (monarch_cpu != -1)
1720 			cpu_relax();	/* spin until monarch leaves */
1721 #endif
1722 
1723 		NOTIFY_INIT(DIE_INIT_SLAVE_LEAVE, regs, (long)&nd, 1);
1724 
1725 		mprintk("Slave on cpu %d returning to normal service.\n", cpu);
1726 		set_curr_task(cpu, previous_current);
1727 		ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1728 		atomic_dec(&slaves);
1729 		return;
1730 	}
1731 
1732 	monarch_cpu = cpu;
1733 	NOTIFY_INIT(DIE_INIT_MONARCH_ENTER, regs, (long)&nd, 1);
1734 
1735 	/*
1736 	 * Wait for a bit.  On some machines (e.g., HP's zx2000 and zx6000, INIT can be
1737 	 * generated via the BMC's command-line interface, but since the console is on the
1738 	 * same serial line, the user will need some time to switch out of the BMC before
1739 	 * the dump begins.
1740 	 */
1741 	mprintk("Delaying for 5 seconds...\n");
1742 	udelay(5*1000000);
1743 	ia64_wait_for_slaves(cpu, "INIT");
1744 	/* If nobody intercepts DIE_INIT_MONARCH_PROCESS then we drop through
1745 	 * to default_monarch_init_process() above and just print all the
1746 	 * tasks.
1747 	 */
1748 	NOTIFY_INIT(DIE_INIT_MONARCH_PROCESS, regs, (long)&nd, 1);
1749 	NOTIFY_INIT(DIE_INIT_MONARCH_LEAVE, regs, (long)&nd, 1);
1750 
1751 	mprintk("\nINIT dump complete.  Monarch on cpu %d returning to normal service.\n", cpu);
1752 	atomic_dec(&monarchs);
1753 	set_curr_task(cpu, previous_current);
1754 	monarch_cpu = -1;
1755 	return;
1756 }
1757 
1758 static int __init
1759 ia64_mca_disable_cpe_polling(char *str)
1760 {
1761 	cpe_poll_enabled = 0;
1762 	return 1;
1763 }
1764 
1765 __setup("disable_cpe_poll", ia64_mca_disable_cpe_polling);
1766 
1767 static struct irqaction cmci_irqaction = {
1768 	.handler =	ia64_mca_cmc_int_handler,
1769 	.flags =	IRQF_DISABLED,
1770 	.name =		"cmc_hndlr"
1771 };
1772 
1773 static struct irqaction cmcp_irqaction = {
1774 	.handler =	ia64_mca_cmc_int_caller,
1775 	.flags =	IRQF_DISABLED,
1776 	.name =		"cmc_poll"
1777 };
1778 
1779 static struct irqaction mca_rdzv_irqaction = {
1780 	.handler =	ia64_mca_rendez_int_handler,
1781 	.flags =	IRQF_DISABLED,
1782 	.name =		"mca_rdzv"
1783 };
1784 
1785 static struct irqaction mca_wkup_irqaction = {
1786 	.handler =	ia64_mca_wakeup_int_handler,
1787 	.flags =	IRQF_DISABLED,
1788 	.name =		"mca_wkup"
1789 };
1790 
1791 #ifdef CONFIG_ACPI
1792 static struct irqaction mca_cpe_irqaction = {
1793 	.handler =	ia64_mca_cpe_int_handler,
1794 	.flags =	IRQF_DISABLED,
1795 	.name =		"cpe_hndlr"
1796 };
1797 
1798 static struct irqaction mca_cpep_irqaction = {
1799 	.handler =	ia64_mca_cpe_int_caller,
1800 	.flags =	IRQF_DISABLED,
1801 	.name =		"cpe_poll"
1802 };
1803 #endif /* CONFIG_ACPI */
1804 
1805 /* Minimal format of the MCA/INIT stacks.  The pseudo processes that run on
1806  * these stacks can never sleep, they cannot return from the kernel to user
1807  * space, they do not appear in a normal ps listing.  So there is no need to
1808  * format most of the fields.
1809  */
1810 
1811 static void __cpuinit
1812 format_mca_init_stack(void *mca_data, unsigned long offset,
1813 		const char *type, int cpu)
1814 {
1815 	struct task_struct *p = (struct task_struct *)((char *)mca_data + offset);
1816 	struct thread_info *ti;
1817 	memset(p, 0, KERNEL_STACK_SIZE);
1818 	ti = task_thread_info(p);
1819 	ti->flags = _TIF_MCA_INIT;
1820 	ti->preempt_count = 1;
1821 	ti->task = p;
1822 	ti->cpu = cpu;
1823 	p->stack = ti;
1824 	p->state = TASK_UNINTERRUPTIBLE;
1825 	cpu_set(cpu, p->cpus_allowed);
1826 	INIT_LIST_HEAD(&p->tasks);
1827 	p->parent = p->real_parent = p->group_leader = p;
1828 	INIT_LIST_HEAD(&p->children);
1829 	INIT_LIST_HEAD(&p->sibling);
1830 	strncpy(p->comm, type, sizeof(p->comm)-1);
1831 }
1832 
1833 /* Caller prevents this from being called after init */
1834 static void * __init_refok mca_bootmem(void)
1835 {
1836 	return __alloc_bootmem(sizeof(struct ia64_mca_cpu),
1837 	                    KERNEL_STACK_SIZE, 0);
1838 }
1839 
1840 /* Do per-CPU MCA-related initialization.  */
1841 void __cpuinit
1842 ia64_mca_cpu_init(void *cpu_data)
1843 {
1844 	void *pal_vaddr;
1845 	void *data;
1846 	long sz = sizeof(struct ia64_mca_cpu);
1847 	int cpu = smp_processor_id();
1848 	static int first_time = 1;
1849 
1850 	/*
1851 	 * Structure will already be allocated if cpu has been online,
1852 	 * then offlined.
1853 	 */
1854 	if (__per_cpu_mca[cpu]) {
1855 		data = __va(__per_cpu_mca[cpu]);
1856 	} else {
1857 		if (first_time) {
1858 			data = mca_bootmem();
1859 			first_time = 0;
1860 		} else
1861 			data = __get_free_pages(GFP_KERNEL, get_order(sz));
1862 		if (!data)
1863 			panic("Could not allocate MCA memory for cpu %d\n",
1864 					cpu);
1865 	}
1866 	format_mca_init_stack(data, offsetof(struct ia64_mca_cpu, mca_stack),
1867 		"MCA", cpu);
1868 	format_mca_init_stack(data, offsetof(struct ia64_mca_cpu, init_stack),
1869 		"INIT", cpu);
1870 	__get_cpu_var(ia64_mca_data) = __per_cpu_mca[cpu] = __pa(data);
1871 
1872 	/*
1873 	 * Stash away a copy of the PTE needed to map the per-CPU page.
1874 	 * We may need it during MCA recovery.
1875 	 */
1876 	__get_cpu_var(ia64_mca_per_cpu_pte) =
1877 		pte_val(mk_pte_phys(__pa(cpu_data), PAGE_KERNEL));
1878 
1879 	/*
1880 	 * Also, stash away a copy of the PAL address and the PTE
1881 	 * needed to map it.
1882 	 */
1883 	pal_vaddr = efi_get_pal_addr();
1884 	if (!pal_vaddr)
1885 		return;
1886 	__get_cpu_var(ia64_mca_pal_base) =
1887 		GRANULEROUNDDOWN((unsigned long) pal_vaddr);
1888 	__get_cpu_var(ia64_mca_pal_pte) = pte_val(mk_pte_phys(__pa(pal_vaddr),
1889 							      PAGE_KERNEL));
1890 }
1891 
1892 static void __cpuinit ia64_mca_cmc_vector_adjust(void *dummy)
1893 {
1894 	unsigned long flags;
1895 
1896 	local_irq_save(flags);
1897 	if (!cmc_polling_enabled)
1898 		ia64_mca_cmc_vector_enable(NULL);
1899 	local_irq_restore(flags);
1900 }
1901 
1902 static int __cpuinit mca_cpu_callback(struct notifier_block *nfb,
1903 				      unsigned long action,
1904 				      void *hcpu)
1905 {
1906 	int hotcpu = (unsigned long) hcpu;
1907 
1908 	switch (action) {
1909 	case CPU_ONLINE:
1910 	case CPU_ONLINE_FROZEN:
1911 		smp_call_function_single(hotcpu, ia64_mca_cmc_vector_adjust,
1912 					 NULL, 0);
1913 		break;
1914 	}
1915 	return NOTIFY_OK;
1916 }
1917 
1918 static struct notifier_block mca_cpu_notifier __cpuinitdata = {
1919 	.notifier_call = mca_cpu_callback
1920 };
1921 
1922 /*
1923  * ia64_mca_init
1924  *
1925  *  Do all the system level mca specific initialization.
1926  *
1927  *	1. Register spinloop and wakeup request interrupt vectors
1928  *
1929  *	2. Register OS_MCA handler entry point
1930  *
1931  *	3. Register OS_INIT handler entry point
1932  *
1933  *  4. Initialize MCA/CMC/INIT related log buffers maintained by the OS.
1934  *
1935  *  Note that this initialization is done very early before some kernel
1936  *  services are available.
1937  *
1938  *  Inputs  :   None
1939  *
1940  *  Outputs :   None
1941  */
1942 void __init
1943 ia64_mca_init(void)
1944 {
1945 	ia64_fptr_t *init_hldlr_ptr_monarch = (ia64_fptr_t *)ia64_os_init_dispatch_monarch;
1946 	ia64_fptr_t *init_hldlr_ptr_slave = (ia64_fptr_t *)ia64_os_init_dispatch_slave;
1947 	ia64_fptr_t *mca_hldlr_ptr = (ia64_fptr_t *)ia64_os_mca_dispatch;
1948 	int i;
1949 	long rc;
1950 	struct ia64_sal_retval isrv;
1951 	unsigned long timeout = IA64_MCA_RENDEZ_TIMEOUT; /* platform specific */
1952 	static struct notifier_block default_init_monarch_nb = {
1953 		.notifier_call = default_monarch_init_process,
1954 		.priority = 0/* we need to notified last */
1955 	};
1956 
1957 	IA64_MCA_DEBUG("%s: begin\n", __func__);
1958 
1959 	/* Clear the Rendez checkin flag for all cpus */
1960 	for(i = 0 ; i < NR_CPUS; i++)
1961 		ia64_mc_info.imi_rendez_checkin[i] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1962 
1963 	/*
1964 	 * Register the rendezvous spinloop and wakeup mechanism with SAL
1965 	 */
1966 
1967 	/* Register the rendezvous interrupt vector with SAL */
1968 	while (1) {
1969 		isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_RENDEZ_INT,
1970 					      SAL_MC_PARAM_MECHANISM_INT,
1971 					      IA64_MCA_RENDEZ_VECTOR,
1972 					      timeout,
1973 					      SAL_MC_PARAM_RZ_ALWAYS);
1974 		rc = isrv.status;
1975 		if (rc == 0)
1976 			break;
1977 		if (rc == -2) {
1978 			printk(KERN_INFO "Increasing MCA rendezvous timeout from "
1979 				"%ld to %ld milliseconds\n", timeout, isrv.v0);
1980 			timeout = isrv.v0;
1981 			NOTIFY_MCA(DIE_MCA_NEW_TIMEOUT, NULL, timeout, 0);
1982 			continue;
1983 		}
1984 		printk(KERN_ERR "Failed to register rendezvous interrupt "
1985 		       "with SAL (status %ld)\n", rc);
1986 		return;
1987 	}
1988 
1989 	/* Register the wakeup interrupt vector with SAL */
1990 	isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_RENDEZ_WAKEUP,
1991 				      SAL_MC_PARAM_MECHANISM_INT,
1992 				      IA64_MCA_WAKEUP_VECTOR,
1993 				      0, 0);
1994 	rc = isrv.status;
1995 	if (rc) {
1996 		printk(KERN_ERR "Failed to register wakeup interrupt with SAL "
1997 		       "(status %ld)\n", rc);
1998 		return;
1999 	}
2000 
2001 	IA64_MCA_DEBUG("%s: registered MCA rendezvous spinloop and wakeup mech.\n", __func__);
2002 
2003 	ia64_mc_info.imi_mca_handler        = ia64_tpa(mca_hldlr_ptr->fp);
2004 	/*
2005 	 * XXX - disable SAL checksum by setting size to 0; should be
2006 	 *	ia64_tpa(ia64_os_mca_dispatch_end) - ia64_tpa(ia64_os_mca_dispatch);
2007 	 */
2008 	ia64_mc_info.imi_mca_handler_size	= 0;
2009 
2010 	/* Register the os mca handler with SAL */
2011 	if ((rc = ia64_sal_set_vectors(SAL_VECTOR_OS_MCA,
2012 				       ia64_mc_info.imi_mca_handler,
2013 				       ia64_tpa(mca_hldlr_ptr->gp),
2014 				       ia64_mc_info.imi_mca_handler_size,
2015 				       0, 0, 0)))
2016 	{
2017 		printk(KERN_ERR "Failed to register OS MCA handler with SAL "
2018 		       "(status %ld)\n", rc);
2019 		return;
2020 	}
2021 
2022 	IA64_MCA_DEBUG("%s: registered OS MCA handler with SAL at 0x%lx, gp = 0x%lx\n", __func__,
2023 		       ia64_mc_info.imi_mca_handler, ia64_tpa(mca_hldlr_ptr->gp));
2024 
2025 	/*
2026 	 * XXX - disable SAL checksum by setting size to 0, should be
2027 	 * size of the actual init handler in mca_asm.S.
2028 	 */
2029 	ia64_mc_info.imi_monarch_init_handler		= ia64_tpa(init_hldlr_ptr_monarch->fp);
2030 	ia64_mc_info.imi_monarch_init_handler_size	= 0;
2031 	ia64_mc_info.imi_slave_init_handler		= ia64_tpa(init_hldlr_ptr_slave->fp);
2032 	ia64_mc_info.imi_slave_init_handler_size	= 0;
2033 
2034 	IA64_MCA_DEBUG("%s: OS INIT handler at %lx\n", __func__,
2035 		       ia64_mc_info.imi_monarch_init_handler);
2036 
2037 	/* Register the os init handler with SAL */
2038 	if ((rc = ia64_sal_set_vectors(SAL_VECTOR_OS_INIT,
2039 				       ia64_mc_info.imi_monarch_init_handler,
2040 				       ia64_tpa(ia64_getreg(_IA64_REG_GP)),
2041 				       ia64_mc_info.imi_monarch_init_handler_size,
2042 				       ia64_mc_info.imi_slave_init_handler,
2043 				       ia64_tpa(ia64_getreg(_IA64_REG_GP)),
2044 				       ia64_mc_info.imi_slave_init_handler_size)))
2045 	{
2046 		printk(KERN_ERR "Failed to register m/s INIT handlers with SAL "
2047 		       "(status %ld)\n", rc);
2048 		return;
2049 	}
2050 	if (register_die_notifier(&default_init_monarch_nb)) {
2051 		printk(KERN_ERR "Failed to register default monarch INIT process\n");
2052 		return;
2053 	}
2054 
2055 	IA64_MCA_DEBUG("%s: registered OS INIT handler with SAL\n", __func__);
2056 
2057 	/*
2058 	 *  Configure the CMCI/P vector and handler. Interrupts for CMC are
2059 	 *  per-processor, so AP CMC interrupts are setup in smp_callin() (smpboot.c).
2060 	 */
2061 	register_percpu_irq(IA64_CMC_VECTOR, &cmci_irqaction);
2062 	register_percpu_irq(IA64_CMCP_VECTOR, &cmcp_irqaction);
2063 	ia64_mca_cmc_vector_setup();       /* Setup vector on BSP */
2064 
2065 	/* Setup the MCA rendezvous interrupt vector */
2066 	register_percpu_irq(IA64_MCA_RENDEZ_VECTOR, &mca_rdzv_irqaction);
2067 
2068 	/* Setup the MCA wakeup interrupt vector */
2069 	register_percpu_irq(IA64_MCA_WAKEUP_VECTOR, &mca_wkup_irqaction);
2070 
2071 #ifdef CONFIG_ACPI
2072 	/* Setup the CPEI/P handler */
2073 	register_percpu_irq(IA64_CPEP_VECTOR, &mca_cpep_irqaction);
2074 #endif
2075 
2076 	/* Initialize the areas set aside by the OS to buffer the
2077 	 * platform/processor error states for MCA/INIT/CMC
2078 	 * handling.
2079 	 */
2080 	ia64_log_init(SAL_INFO_TYPE_MCA);
2081 	ia64_log_init(SAL_INFO_TYPE_INIT);
2082 	ia64_log_init(SAL_INFO_TYPE_CMC);
2083 	ia64_log_init(SAL_INFO_TYPE_CPE);
2084 
2085 	mca_init = 1;
2086 	printk(KERN_INFO "MCA related initialization done\n");
2087 }
2088 
2089 /*
2090  * ia64_mca_late_init
2091  *
2092  *	Opportunity to setup things that require initialization later
2093  *	than ia64_mca_init.  Setup a timer to poll for CPEs if the
2094  *	platform doesn't support an interrupt driven mechanism.
2095  *
2096  *  Inputs  :   None
2097  *  Outputs :   Status
2098  */
2099 static int __init
2100 ia64_mca_late_init(void)
2101 {
2102 	if (!mca_init)
2103 		return 0;
2104 
2105 	register_hotcpu_notifier(&mca_cpu_notifier);
2106 
2107 	/* Setup the CMCI/P vector and handler */
2108 	init_timer(&cmc_poll_timer);
2109 	cmc_poll_timer.function = ia64_mca_cmc_poll;
2110 
2111 	/* Unmask/enable the vector */
2112 	cmc_polling_enabled = 0;
2113 	schedule_work(&cmc_enable_work);
2114 
2115 	IA64_MCA_DEBUG("%s: CMCI/P setup and enabled.\n", __func__);
2116 
2117 #ifdef CONFIG_ACPI
2118 	/* Setup the CPEI/P vector and handler */
2119 	cpe_vector = acpi_request_vector(ACPI_INTERRUPT_CPEI);
2120 	init_timer(&cpe_poll_timer);
2121 	cpe_poll_timer.function = ia64_mca_cpe_poll;
2122 
2123 	{
2124 		struct irq_desc *desc;
2125 		unsigned int irq;
2126 
2127 		if (cpe_vector >= 0) {
2128 			/* If platform supports CPEI, enable the irq. */
2129 			irq = local_vector_to_irq(cpe_vector);
2130 			if (irq > 0) {
2131 				cpe_poll_enabled = 0;
2132 				desc = irq_desc + irq;
2133 				desc->status |= IRQ_PER_CPU;
2134 				setup_irq(irq, &mca_cpe_irqaction);
2135 				ia64_cpe_irq = irq;
2136 				ia64_mca_register_cpev(cpe_vector);
2137 				IA64_MCA_DEBUG("%s: CPEI/P setup and enabled.\n",
2138 					__func__);
2139 				return 0;
2140 			}
2141 			printk(KERN_ERR "%s: Failed to find irq for CPE "
2142 					"interrupt handler, vector %d\n",
2143 					__func__, cpe_vector);
2144 		}
2145 		/* If platform doesn't support CPEI, get the timer going. */
2146 		if (cpe_poll_enabled) {
2147 			ia64_mca_cpe_poll(0UL);
2148 			IA64_MCA_DEBUG("%s: CPEP setup and enabled.\n", __func__);
2149 		}
2150 	}
2151 #endif
2152 
2153 	return 0;
2154 }
2155 
2156 device_initcall(ia64_mca_late_init);
2157