xref: /openbmc/linux/arch/ia64/kernel/mca.c (revision 15e3ae36)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * File:	mca.c
4  * Purpose:	Generic MCA handling layer
5  *
6  * Copyright (C) 2003 Hewlett-Packard Co
7  *	David Mosberger-Tang <davidm@hpl.hp.com>
8  *
9  * Copyright (C) 2002 Dell Inc.
10  * Copyright (C) Matt Domsch <Matt_Domsch@dell.com>
11  *
12  * Copyright (C) 2002 Intel
13  * Copyright (C) Jenna Hall <jenna.s.hall@intel.com>
14  *
15  * Copyright (C) 2001 Intel
16  * Copyright (C) Fred Lewis <frederick.v.lewis@intel.com>
17  *
18  * Copyright (C) 2000 Intel
19  * Copyright (C) Chuck Fleckenstein <cfleck@co.intel.com>
20  *
21  * Copyright (C) 1999, 2004-2008 Silicon Graphics, Inc.
22  * Copyright (C) Vijay Chander <vijay@engr.sgi.com>
23  *
24  * Copyright (C) 2006 FUJITSU LIMITED
25  * Copyright (C) Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
26  *
27  * 2000-03-29 Chuck Fleckenstein <cfleck@co.intel.com>
28  *	      Fixed PAL/SAL update issues, began MCA bug fixes, logging issues,
29  *	      added min save state dump, added INIT handler.
30  *
31  * 2001-01-03 Fred Lewis <frederick.v.lewis@intel.com>
32  *	      Added setup of CMCI and CPEI IRQs, logging of corrected platform
33  *	      errors, completed code for logging of corrected & uncorrected
34  *	      machine check errors, and updated for conformance with Nov. 2000
35  *	      revision of the SAL 3.0 spec.
36  *
37  * 2002-01-04 Jenna Hall <jenna.s.hall@intel.com>
38  *	      Aligned MCA stack to 16 bytes, added platform vs. CPU error flag,
39  *	      set SAL default return values, changed error record structure to
40  *	      linked list, added init call to sal_get_state_info_size().
41  *
42  * 2002-03-25 Matt Domsch <Matt_Domsch@dell.com>
43  *	      GUID cleanups.
44  *
45  * 2003-04-15 David Mosberger-Tang <davidm@hpl.hp.com>
46  *	      Added INIT backtrace support.
47  *
48  * 2003-12-08 Keith Owens <kaos@sgi.com>
49  *	      smp_call_function() must not be called from interrupt context
50  *	      (can deadlock on tasklist_lock).
51  *	      Use keventd to call smp_call_function().
52  *
53  * 2004-02-01 Keith Owens <kaos@sgi.com>
54  *	      Avoid deadlock when using printk() for MCA and INIT records.
55  *	      Delete all record printing code, moved to salinfo_decode in user
56  *	      space.  Mark variables and functions static where possible.
57  *	      Delete dead variables and functions.  Reorder to remove the need
58  *	      for forward declarations and to consolidate related code.
59  *
60  * 2005-08-12 Keith Owens <kaos@sgi.com>
61  *	      Convert MCA/INIT handlers to use per event stacks and SAL/OS
62  *	      state.
63  *
64  * 2005-10-07 Keith Owens <kaos@sgi.com>
65  *	      Add notify_die() hooks.
66  *
67  * 2006-09-15 Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
68  *	      Add printing support for MCA/INIT.
69  *
70  * 2007-04-27 Russ Anderson <rja@sgi.com>
71  *	      Support multiple cpus going through OS_MCA in the same event.
72  */
73 #include <linux/jiffies.h>
74 #include <linux/types.h>
75 #include <linux/init.h>
76 #include <linux/sched/signal.h>
77 #include <linux/sched/debug.h>
78 #include <linux/sched/task.h>
79 #include <linux/interrupt.h>
80 #include <linux/irq.h>
81 #include <linux/memblock.h>
82 #include <linux/acpi.h>
83 #include <linux/timer.h>
84 #include <linux/module.h>
85 #include <linux/kernel.h>
86 #include <linux/smp.h>
87 #include <linux/workqueue.h>
88 #include <linux/cpumask.h>
89 #include <linux/kdebug.h>
90 #include <linux/cpu.h>
91 #include <linux/gfp.h>
92 
93 #include <asm/delay.h>
94 #include <asm/meminit.h>
95 #include <asm/page.h>
96 #include <asm/ptrace.h>
97 #include <asm/sal.h>
98 #include <asm/mca.h>
99 #include <asm/kexec.h>
100 
101 #include <asm/irq.h>
102 #include <asm/hw_irq.h>
103 #include <asm/tlb.h>
104 
105 #include "mca_drv.h"
106 #include "entry.h"
107 #include "irq.h"
108 
109 #if defined(IA64_MCA_DEBUG_INFO)
110 # define IA64_MCA_DEBUG(fmt...)	printk(fmt)
111 #else
112 # define IA64_MCA_DEBUG(fmt...)
113 #endif
114 
115 #define NOTIFY_INIT(event, regs, arg, spin)				\
116 do {									\
117 	if ((notify_die((event), "INIT", (regs), (arg), 0, 0)		\
118 			== NOTIFY_STOP) && ((spin) == 1))		\
119 		ia64_mca_spin(__func__);				\
120 } while (0)
121 
122 #define NOTIFY_MCA(event, regs, arg, spin)				\
123 do {									\
124 	if ((notify_die((event), "MCA", (regs), (arg), 0, 0)		\
125 			== NOTIFY_STOP) && ((spin) == 1))		\
126 		ia64_mca_spin(__func__);				\
127 } while (0)
128 
129 /* Used by mca_asm.S */
130 DEFINE_PER_CPU(u64, ia64_mca_data); /* == __per_cpu_mca[smp_processor_id()] */
131 DEFINE_PER_CPU(u64, ia64_mca_per_cpu_pte); /* PTE to map per-CPU area */
132 DEFINE_PER_CPU(u64, ia64_mca_pal_pte);	    /* PTE to map PAL code */
133 DEFINE_PER_CPU(u64, ia64_mca_pal_base);    /* vaddr PAL code granule */
134 DEFINE_PER_CPU(u64, ia64_mca_tr_reload);   /* Flag for TR reload */
135 
136 unsigned long __per_cpu_mca[NR_CPUS];
137 
138 /* In mca_asm.S */
139 extern void			ia64_os_init_dispatch_monarch (void);
140 extern void			ia64_os_init_dispatch_slave (void);
141 
142 static int monarch_cpu = -1;
143 
144 static ia64_mc_info_t		ia64_mc_info;
145 
146 #define MAX_CPE_POLL_INTERVAL (15*60*HZ) /* 15 minutes */
147 #define MIN_CPE_POLL_INTERVAL (2*60*HZ)  /* 2 minutes */
148 #define CMC_POLL_INTERVAL     (1*60*HZ)  /* 1 minute */
149 #define CPE_HISTORY_LENGTH    5
150 #define CMC_HISTORY_LENGTH    5
151 
152 static struct timer_list cpe_poll_timer;
153 static struct timer_list cmc_poll_timer;
154 /*
155  * This variable tells whether we are currently in polling mode.
156  * Start with this in the wrong state so we won't play w/ timers
157  * before the system is ready.
158  */
159 static int cmc_polling_enabled = 1;
160 
161 /*
162  * Clearing this variable prevents CPE polling from getting activated
163  * in mca_late_init.  Use it if your system doesn't provide a CPEI,
164  * but encounters problems retrieving CPE logs.  This should only be
165  * necessary for debugging.
166  */
167 static int cpe_poll_enabled = 1;
168 
169 extern void salinfo_log_wakeup(int type, u8 *buffer, u64 size, int irqsafe);
170 
171 static int mca_init __initdata;
172 
173 /*
174  * limited & delayed printing support for MCA/INIT handler
175  */
176 
177 #define mprintk(fmt...) ia64_mca_printk(fmt)
178 
179 #define MLOGBUF_SIZE (512+256*NR_CPUS)
180 #define MLOGBUF_MSGMAX 256
181 static char mlogbuf[MLOGBUF_SIZE];
182 static DEFINE_SPINLOCK(mlogbuf_wlock);	/* mca context only */
183 static DEFINE_SPINLOCK(mlogbuf_rlock);	/* normal context only */
184 static unsigned long mlogbuf_start;
185 static unsigned long mlogbuf_end;
186 static unsigned int mlogbuf_finished = 0;
187 static unsigned long mlogbuf_timestamp = 0;
188 
189 static int loglevel_save = -1;
190 #define BREAK_LOGLEVEL(__console_loglevel)		\
191 	oops_in_progress = 1;				\
192 	if (loglevel_save < 0)				\
193 		loglevel_save = __console_loglevel;	\
194 	__console_loglevel = 15;
195 
196 #define RESTORE_LOGLEVEL(__console_loglevel)		\
197 	if (loglevel_save >= 0) {			\
198 		__console_loglevel = loglevel_save;	\
199 		loglevel_save = -1;			\
200 	}						\
201 	mlogbuf_finished = 0;				\
202 	oops_in_progress = 0;
203 
204 /*
205  * Push messages into buffer, print them later if not urgent.
206  */
207 void ia64_mca_printk(const char *fmt, ...)
208 {
209 	va_list args;
210 	int printed_len;
211 	char temp_buf[MLOGBUF_MSGMAX];
212 	char *p;
213 
214 	va_start(args, fmt);
215 	printed_len = vscnprintf(temp_buf, sizeof(temp_buf), fmt, args);
216 	va_end(args);
217 
218 	/* Copy the output into mlogbuf */
219 	if (oops_in_progress) {
220 		/* mlogbuf was abandoned, use printk directly instead. */
221 		printk("%s", temp_buf);
222 	} else {
223 		spin_lock(&mlogbuf_wlock);
224 		for (p = temp_buf; *p; p++) {
225 			unsigned long next = (mlogbuf_end + 1) % MLOGBUF_SIZE;
226 			if (next != mlogbuf_start) {
227 				mlogbuf[mlogbuf_end] = *p;
228 				mlogbuf_end = next;
229 			} else {
230 				/* buffer full */
231 				break;
232 			}
233 		}
234 		mlogbuf[mlogbuf_end] = '\0';
235 		spin_unlock(&mlogbuf_wlock);
236 	}
237 }
238 EXPORT_SYMBOL(ia64_mca_printk);
239 
240 /*
241  * Print buffered messages.
242  *  NOTE: call this after returning normal context. (ex. from salinfod)
243  */
244 void ia64_mlogbuf_dump(void)
245 {
246 	char temp_buf[MLOGBUF_MSGMAX];
247 	char *p;
248 	unsigned long index;
249 	unsigned long flags;
250 	unsigned int printed_len;
251 
252 	/* Get output from mlogbuf */
253 	while (mlogbuf_start != mlogbuf_end) {
254 		temp_buf[0] = '\0';
255 		p = temp_buf;
256 		printed_len = 0;
257 
258 		spin_lock_irqsave(&mlogbuf_rlock, flags);
259 
260 		index = mlogbuf_start;
261 		while (index != mlogbuf_end) {
262 			*p = mlogbuf[index];
263 			index = (index + 1) % MLOGBUF_SIZE;
264 			if (!*p)
265 				break;
266 			p++;
267 			if (++printed_len >= MLOGBUF_MSGMAX - 1)
268 				break;
269 		}
270 		*p = '\0';
271 		if (temp_buf[0])
272 			printk("%s", temp_buf);
273 		mlogbuf_start = index;
274 
275 		mlogbuf_timestamp = 0;
276 		spin_unlock_irqrestore(&mlogbuf_rlock, flags);
277 	}
278 }
279 EXPORT_SYMBOL(ia64_mlogbuf_dump);
280 
281 /*
282  * Call this if system is going to down or if immediate flushing messages to
283  * console is required. (ex. recovery was failed, crash dump is going to be
284  * invoked, long-wait rendezvous etc.)
285  *  NOTE: this should be called from monarch.
286  */
287 static void ia64_mlogbuf_finish(int wait)
288 {
289 	BREAK_LOGLEVEL(console_loglevel);
290 
291 	spin_lock_init(&mlogbuf_rlock);
292 	ia64_mlogbuf_dump();
293 	printk(KERN_EMERG "mlogbuf_finish: printing switched to urgent mode, "
294 		"MCA/INIT might be dodgy or fail.\n");
295 
296 	if (!wait)
297 		return;
298 
299 	/* wait for console */
300 	printk("Delaying for 5 seconds...\n");
301 	udelay(5*1000000);
302 
303 	mlogbuf_finished = 1;
304 }
305 
306 /*
307  * Print buffered messages from INIT context.
308  */
309 static void ia64_mlogbuf_dump_from_init(void)
310 {
311 	if (mlogbuf_finished)
312 		return;
313 
314 	if (mlogbuf_timestamp &&
315 			time_before(jiffies, mlogbuf_timestamp + 30 * HZ)) {
316 		printk(KERN_ERR "INIT: mlogbuf_dump is interrupted by INIT "
317 			" and the system seems to be messed up.\n");
318 		ia64_mlogbuf_finish(0);
319 		return;
320 	}
321 
322 	if (!spin_trylock(&mlogbuf_rlock)) {
323 		printk(KERN_ERR "INIT: mlogbuf_dump is interrupted by INIT. "
324 			"Generated messages other than stack dump will be "
325 			"buffered to mlogbuf and will be printed later.\n");
326 		printk(KERN_ERR "INIT: If messages would not printed after "
327 			"this INIT, wait 30sec and assert INIT again.\n");
328 		if (!mlogbuf_timestamp)
329 			mlogbuf_timestamp = jiffies;
330 		return;
331 	}
332 	spin_unlock(&mlogbuf_rlock);
333 	ia64_mlogbuf_dump();
334 }
335 
336 static inline void
337 ia64_mca_spin(const char *func)
338 {
339 	if (monarch_cpu == smp_processor_id())
340 		ia64_mlogbuf_finish(0);
341 	mprintk(KERN_EMERG "%s: spinning here, not returning to SAL\n", func);
342 	while (1)
343 		cpu_relax();
344 }
345 /*
346  * IA64_MCA log support
347  */
348 #define IA64_MAX_LOGS		2	/* Double-buffering for nested MCAs */
349 #define IA64_MAX_LOG_TYPES      4   /* MCA, INIT, CMC, CPE */
350 
351 typedef struct ia64_state_log_s
352 {
353 	spinlock_t	isl_lock;
354 	int		isl_index;
355 	unsigned long	isl_count;
356 	ia64_err_rec_t  *isl_log[IA64_MAX_LOGS]; /* need space to store header + error log */
357 } ia64_state_log_t;
358 
359 static ia64_state_log_t ia64_state_log[IA64_MAX_LOG_TYPES];
360 
361 #define IA64_LOG_LOCK_INIT(it) spin_lock_init(&ia64_state_log[it].isl_lock)
362 #define IA64_LOG_LOCK(it)      spin_lock_irqsave(&ia64_state_log[it].isl_lock, s)
363 #define IA64_LOG_UNLOCK(it)    spin_unlock_irqrestore(&ia64_state_log[it].isl_lock,s)
364 #define IA64_LOG_NEXT_INDEX(it)    ia64_state_log[it].isl_index
365 #define IA64_LOG_CURR_INDEX(it)    1 - ia64_state_log[it].isl_index
366 #define IA64_LOG_INDEX_INC(it) \
367     {ia64_state_log[it].isl_index = 1 - ia64_state_log[it].isl_index; \
368     ia64_state_log[it].isl_count++;}
369 #define IA64_LOG_INDEX_DEC(it) \
370     ia64_state_log[it].isl_index = 1 - ia64_state_log[it].isl_index
371 #define IA64_LOG_NEXT_BUFFER(it)   (void *)((ia64_state_log[it].isl_log[IA64_LOG_NEXT_INDEX(it)]))
372 #define IA64_LOG_CURR_BUFFER(it)   (void *)((ia64_state_log[it].isl_log[IA64_LOG_CURR_INDEX(it)]))
373 #define IA64_LOG_COUNT(it)         ia64_state_log[it].isl_count
374 
375 static inline void ia64_log_allocate(int it, u64 size)
376 {
377 	ia64_state_log[it].isl_log[IA64_LOG_CURR_INDEX(it)] =
378 		(ia64_err_rec_t *)memblock_alloc(size, SMP_CACHE_BYTES);
379 	if (!ia64_state_log[it].isl_log[IA64_LOG_CURR_INDEX(it)])
380 		panic("%s: Failed to allocate %llu bytes\n", __func__, size);
381 
382 	ia64_state_log[it].isl_log[IA64_LOG_NEXT_INDEX(it)] =
383 		(ia64_err_rec_t *)memblock_alloc(size, SMP_CACHE_BYTES);
384 	if (!ia64_state_log[it].isl_log[IA64_LOG_NEXT_INDEX(it)])
385 		panic("%s: Failed to allocate %llu bytes\n", __func__, size);
386 }
387 
388 /*
389  * ia64_log_init
390  *	Reset the OS ia64 log buffer
391  * Inputs   :   info_type   (SAL_INFO_TYPE_{MCA,INIT,CMC,CPE})
392  * Outputs	:	None
393  */
394 static void __init
395 ia64_log_init(int sal_info_type)
396 {
397 	u64	max_size = 0;
398 
399 	IA64_LOG_NEXT_INDEX(sal_info_type) = 0;
400 	IA64_LOG_LOCK_INIT(sal_info_type);
401 
402 	// SAL will tell us the maximum size of any error record of this type
403 	max_size = ia64_sal_get_state_info_size(sal_info_type);
404 	if (!max_size)
405 		/* alloc_bootmem() doesn't like zero-sized allocations! */
406 		return;
407 
408 	// set up OS data structures to hold error info
409 	ia64_log_allocate(sal_info_type, max_size);
410 }
411 
412 /*
413  * ia64_log_get
414  *
415  *	Get the current MCA log from SAL and copy it into the OS log buffer.
416  *
417  *  Inputs  :   info_type   (SAL_INFO_TYPE_{MCA,INIT,CMC,CPE})
418  *              irq_safe    whether you can use printk at this point
419  *  Outputs :   size        (total record length)
420  *              *buffer     (ptr to error record)
421  *
422  */
423 static u64
424 ia64_log_get(int sal_info_type, u8 **buffer, int irq_safe)
425 {
426 	sal_log_record_header_t     *log_buffer;
427 	u64                         total_len = 0;
428 	unsigned long               s;
429 
430 	IA64_LOG_LOCK(sal_info_type);
431 
432 	/* Get the process state information */
433 	log_buffer = IA64_LOG_NEXT_BUFFER(sal_info_type);
434 
435 	total_len = ia64_sal_get_state_info(sal_info_type, (u64 *)log_buffer);
436 
437 	if (total_len) {
438 		IA64_LOG_INDEX_INC(sal_info_type);
439 		IA64_LOG_UNLOCK(sal_info_type);
440 		if (irq_safe) {
441 			IA64_MCA_DEBUG("%s: SAL error record type %d retrieved. Record length = %ld\n",
442 				       __func__, sal_info_type, total_len);
443 		}
444 		*buffer = (u8 *) log_buffer;
445 		return total_len;
446 	} else {
447 		IA64_LOG_UNLOCK(sal_info_type);
448 		return 0;
449 	}
450 }
451 
452 /*
453  *  ia64_mca_log_sal_error_record
454  *
455  *  This function retrieves a specified error record type from SAL
456  *  and wakes up any processes waiting for error records.
457  *
458  *  Inputs  :   sal_info_type   (Type of error record MCA/CMC/CPE)
459  *              FIXME: remove MCA and irq_safe.
460  */
461 static void
462 ia64_mca_log_sal_error_record(int sal_info_type)
463 {
464 	u8 *buffer;
465 	sal_log_record_header_t *rh;
466 	u64 size;
467 	int irq_safe = sal_info_type != SAL_INFO_TYPE_MCA;
468 #ifdef IA64_MCA_DEBUG_INFO
469 	static const char * const rec_name[] = { "MCA", "INIT", "CMC", "CPE" };
470 #endif
471 
472 	size = ia64_log_get(sal_info_type, &buffer, irq_safe);
473 	if (!size)
474 		return;
475 
476 	salinfo_log_wakeup(sal_info_type, buffer, size, irq_safe);
477 
478 	if (irq_safe)
479 		IA64_MCA_DEBUG("CPU %d: SAL log contains %s error record\n",
480 			smp_processor_id(),
481 			sal_info_type < ARRAY_SIZE(rec_name) ? rec_name[sal_info_type] : "UNKNOWN");
482 
483 	/* Clear logs from corrected errors in case there's no user-level logger */
484 	rh = (sal_log_record_header_t *)buffer;
485 	if (rh->severity == sal_log_severity_corrected)
486 		ia64_sal_clear_state_info(sal_info_type);
487 }
488 
489 /*
490  * search_mca_table
491  *  See if the MCA surfaced in an instruction range
492  *  that has been tagged as recoverable.
493  *
494  *  Inputs
495  *	first	First address range to check
496  *	last	Last address range to check
497  *	ip	Instruction pointer, address we are looking for
498  *
499  * Return value:
500  *      1 on Success (in the table)/ 0 on Failure (not in the  table)
501  */
502 int
503 search_mca_table (const struct mca_table_entry *first,
504                 const struct mca_table_entry *last,
505                 unsigned long ip)
506 {
507         const struct mca_table_entry *curr;
508         u64 curr_start, curr_end;
509 
510         curr = first;
511         while (curr <= last) {
512                 curr_start = (u64) &curr->start_addr + curr->start_addr;
513                 curr_end = (u64) &curr->end_addr + curr->end_addr;
514 
515                 if ((ip >= curr_start) && (ip <= curr_end)) {
516                         return 1;
517                 }
518                 curr++;
519         }
520         return 0;
521 }
522 
523 /* Given an address, look for it in the mca tables. */
524 int mca_recover_range(unsigned long addr)
525 {
526 	extern struct mca_table_entry __start___mca_table[];
527 	extern struct mca_table_entry __stop___mca_table[];
528 
529 	return search_mca_table(__start___mca_table, __stop___mca_table-1, addr);
530 }
531 EXPORT_SYMBOL_GPL(mca_recover_range);
532 
533 int cpe_vector = -1;
534 int ia64_cpe_irq = -1;
535 
536 static irqreturn_t
537 ia64_mca_cpe_int_handler (int cpe_irq, void *arg)
538 {
539 	static unsigned long	cpe_history[CPE_HISTORY_LENGTH];
540 	static int		index;
541 	static DEFINE_SPINLOCK(cpe_history_lock);
542 
543 	IA64_MCA_DEBUG("%s: received interrupt vector = %#x on CPU %d\n",
544 		       __func__, cpe_irq, smp_processor_id());
545 
546 	/* SAL spec states this should run w/ interrupts enabled */
547 	local_irq_enable();
548 
549 	spin_lock(&cpe_history_lock);
550 	if (!cpe_poll_enabled && cpe_vector >= 0) {
551 
552 		int i, count = 1; /* we know 1 happened now */
553 		unsigned long now = jiffies;
554 
555 		for (i = 0; i < CPE_HISTORY_LENGTH; i++) {
556 			if (now - cpe_history[i] <= HZ)
557 				count++;
558 		}
559 
560 		IA64_MCA_DEBUG(KERN_INFO "CPE threshold %d/%d\n", count, CPE_HISTORY_LENGTH);
561 		if (count >= CPE_HISTORY_LENGTH) {
562 
563 			cpe_poll_enabled = 1;
564 			spin_unlock(&cpe_history_lock);
565 			disable_irq_nosync(local_vector_to_irq(IA64_CPE_VECTOR));
566 
567 			/*
568 			 * Corrected errors will still be corrected, but
569 			 * make sure there's a log somewhere that indicates
570 			 * something is generating more than we can handle.
571 			 */
572 			printk(KERN_WARNING "WARNING: Switching to polling CPE handler; error records may be lost\n");
573 
574 			mod_timer(&cpe_poll_timer, jiffies + MIN_CPE_POLL_INTERVAL);
575 
576 			/* lock already released, get out now */
577 			goto out;
578 		} else {
579 			cpe_history[index++] = now;
580 			if (index == CPE_HISTORY_LENGTH)
581 				index = 0;
582 		}
583 	}
584 	spin_unlock(&cpe_history_lock);
585 out:
586 	/* Get the CPE error record and log it */
587 	ia64_mca_log_sal_error_record(SAL_INFO_TYPE_CPE);
588 
589 	local_irq_disable();
590 
591 	return IRQ_HANDLED;
592 }
593 
594 /*
595  * ia64_mca_register_cpev
596  *
597  *  Register the corrected platform error vector with SAL.
598  *
599  *  Inputs
600  *      cpev        Corrected Platform Error Vector number
601  *
602  *  Outputs
603  *      None
604  */
605 void
606 ia64_mca_register_cpev (int cpev)
607 {
608 	/* Register the CPE interrupt vector with SAL */
609 	struct ia64_sal_retval isrv;
610 
611 	isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_CPE_INT, SAL_MC_PARAM_MECHANISM_INT, cpev, 0, 0);
612 	if (isrv.status) {
613 		printk(KERN_ERR "Failed to register Corrected Platform "
614 		       "Error interrupt vector with SAL (status %ld)\n", isrv.status);
615 		return;
616 	}
617 
618 	IA64_MCA_DEBUG("%s: corrected platform error "
619 		       "vector %#x registered\n", __func__, cpev);
620 }
621 
622 /*
623  * ia64_mca_cmc_vector_setup
624  *
625  *  Setup the corrected machine check vector register in the processor.
626  *  (The interrupt is masked on boot. ia64_mca_late_init unmask this.)
627  *  This function is invoked on a per-processor basis.
628  *
629  * Inputs
630  *      None
631  *
632  * Outputs
633  *	None
634  */
635 void
636 ia64_mca_cmc_vector_setup (void)
637 {
638 	cmcv_reg_t	cmcv;
639 
640 	cmcv.cmcv_regval	= 0;
641 	cmcv.cmcv_mask		= 1;        /* Mask/disable interrupt at first */
642 	cmcv.cmcv_vector	= IA64_CMC_VECTOR;
643 	ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
644 
645 	IA64_MCA_DEBUG("%s: CPU %d corrected machine check vector %#x registered.\n",
646 		       __func__, smp_processor_id(), IA64_CMC_VECTOR);
647 
648 	IA64_MCA_DEBUG("%s: CPU %d CMCV = %#016lx\n",
649 		       __func__, smp_processor_id(), ia64_getreg(_IA64_REG_CR_CMCV));
650 }
651 
652 /*
653  * ia64_mca_cmc_vector_disable
654  *
655  *  Mask the corrected machine check vector register in the processor.
656  *  This function is invoked on a per-processor basis.
657  *
658  * Inputs
659  *      dummy(unused)
660  *
661  * Outputs
662  *	None
663  */
664 static void
665 ia64_mca_cmc_vector_disable (void *dummy)
666 {
667 	cmcv_reg_t	cmcv;
668 
669 	cmcv.cmcv_regval = ia64_getreg(_IA64_REG_CR_CMCV);
670 
671 	cmcv.cmcv_mask = 1; /* Mask/disable interrupt */
672 	ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
673 
674 	IA64_MCA_DEBUG("%s: CPU %d corrected machine check vector %#x disabled.\n",
675 		       __func__, smp_processor_id(), cmcv.cmcv_vector);
676 }
677 
678 /*
679  * ia64_mca_cmc_vector_enable
680  *
681  *  Unmask the corrected machine check vector register in the processor.
682  *  This function is invoked on a per-processor basis.
683  *
684  * Inputs
685  *      dummy(unused)
686  *
687  * Outputs
688  *	None
689  */
690 static void
691 ia64_mca_cmc_vector_enable (void *dummy)
692 {
693 	cmcv_reg_t	cmcv;
694 
695 	cmcv.cmcv_regval = ia64_getreg(_IA64_REG_CR_CMCV);
696 
697 	cmcv.cmcv_mask = 0; /* Unmask/enable interrupt */
698 	ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
699 
700 	IA64_MCA_DEBUG("%s: CPU %d corrected machine check vector %#x enabled.\n",
701 		       __func__, smp_processor_id(), cmcv.cmcv_vector);
702 }
703 
704 /*
705  * ia64_mca_cmc_vector_disable_keventd
706  *
707  * Called via keventd (smp_call_function() is not safe in interrupt context) to
708  * disable the cmc interrupt vector.
709  */
710 static void
711 ia64_mca_cmc_vector_disable_keventd(struct work_struct *unused)
712 {
713 	on_each_cpu(ia64_mca_cmc_vector_disable, NULL, 0);
714 }
715 
716 /*
717  * ia64_mca_cmc_vector_enable_keventd
718  *
719  * Called via keventd (smp_call_function() is not safe in interrupt context) to
720  * enable the cmc interrupt vector.
721  */
722 static void
723 ia64_mca_cmc_vector_enable_keventd(struct work_struct *unused)
724 {
725 	on_each_cpu(ia64_mca_cmc_vector_enable, NULL, 0);
726 }
727 
728 /*
729  * ia64_mca_wakeup
730  *
731  *	Send an inter-cpu interrupt to wake-up a particular cpu.
732  *
733  *  Inputs  :   cpuid
734  *  Outputs :   None
735  */
736 static void
737 ia64_mca_wakeup(int cpu)
738 {
739 	ia64_send_ipi(cpu, IA64_MCA_WAKEUP_VECTOR, IA64_IPI_DM_INT, 0);
740 }
741 
742 /*
743  * ia64_mca_wakeup_all
744  *
745  *	Wakeup all the slave cpus which have rendez'ed previously.
746  *
747  *  Inputs  :   None
748  *  Outputs :   None
749  */
750 static void
751 ia64_mca_wakeup_all(void)
752 {
753 	int cpu;
754 
755 	/* Clear the Rendez checkin flag for all cpus */
756 	for_each_online_cpu(cpu) {
757 		if (ia64_mc_info.imi_rendez_checkin[cpu] == IA64_MCA_RENDEZ_CHECKIN_DONE)
758 			ia64_mca_wakeup(cpu);
759 	}
760 
761 }
762 
763 /*
764  * ia64_mca_rendez_interrupt_handler
765  *
766  *	This is handler used to put slave processors into spinloop
767  *	while the monarch processor does the mca handling and later
768  *	wake each slave up once the monarch is done.  The state
769  *	IA64_MCA_RENDEZ_CHECKIN_DONE indicates the cpu is rendez'ed
770  *	in SAL.  The state IA64_MCA_RENDEZ_CHECKIN_NOTDONE indicates
771  *	the cpu has come out of OS rendezvous.
772  *
773  *  Inputs  :   None
774  *  Outputs :   None
775  */
776 static irqreturn_t
777 ia64_mca_rendez_int_handler(int rendez_irq, void *arg)
778 {
779 	unsigned long flags;
780 	int cpu = smp_processor_id();
781 	struct ia64_mca_notify_die nd =
782 		{ .sos = NULL, .monarch_cpu = &monarch_cpu };
783 
784 	/* Mask all interrupts */
785 	local_irq_save(flags);
786 
787 	NOTIFY_MCA(DIE_MCA_RENDZVOUS_ENTER, get_irq_regs(), (long)&nd, 1);
788 
789 	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_DONE;
790 	/* Register with the SAL monarch that the slave has
791 	 * reached SAL
792 	 */
793 	ia64_sal_mc_rendez();
794 
795 	NOTIFY_MCA(DIE_MCA_RENDZVOUS_PROCESS, get_irq_regs(), (long)&nd, 1);
796 
797 	/* Wait for the monarch cpu to exit. */
798 	while (monarch_cpu != -1)
799 	       cpu_relax();	/* spin until monarch leaves */
800 
801 	NOTIFY_MCA(DIE_MCA_RENDZVOUS_LEAVE, get_irq_regs(), (long)&nd, 1);
802 
803 	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
804 	/* Enable all interrupts */
805 	local_irq_restore(flags);
806 	return IRQ_HANDLED;
807 }
808 
809 /*
810  * ia64_mca_wakeup_int_handler
811  *
812  *	The interrupt handler for processing the inter-cpu interrupt to the
813  *	slave cpu which was spinning in the rendez loop.
814  *	Since this spinning is done by turning off the interrupts and
815  *	polling on the wakeup-interrupt bit in the IRR, there is
816  *	nothing useful to be done in the handler.
817  *
818  *  Inputs  :   wakeup_irq  (Wakeup-interrupt bit)
819  *	arg		(Interrupt handler specific argument)
820  *  Outputs :   None
821  *
822  */
823 static irqreturn_t
824 ia64_mca_wakeup_int_handler(int wakeup_irq, void *arg)
825 {
826 	return IRQ_HANDLED;
827 }
828 
829 /* Function pointer for extra MCA recovery */
830 int (*ia64_mca_ucmc_extension)
831 	(void*,struct ia64_sal_os_state*)
832 	= NULL;
833 
834 int
835 ia64_reg_MCA_extension(int (*fn)(void *, struct ia64_sal_os_state *))
836 {
837 	if (ia64_mca_ucmc_extension)
838 		return 1;
839 
840 	ia64_mca_ucmc_extension = fn;
841 	return 0;
842 }
843 
844 void
845 ia64_unreg_MCA_extension(void)
846 {
847 	if (ia64_mca_ucmc_extension)
848 		ia64_mca_ucmc_extension = NULL;
849 }
850 
851 EXPORT_SYMBOL(ia64_reg_MCA_extension);
852 EXPORT_SYMBOL(ia64_unreg_MCA_extension);
853 
854 
855 static inline void
856 copy_reg(const u64 *fr, u64 fnat, unsigned long *tr, unsigned long *tnat)
857 {
858 	u64 fslot, tslot, nat;
859 	*tr = *fr;
860 	fslot = ((unsigned long)fr >> 3) & 63;
861 	tslot = ((unsigned long)tr >> 3) & 63;
862 	*tnat &= ~(1UL << tslot);
863 	nat = (fnat >> fslot) & 1;
864 	*tnat |= (nat << tslot);
865 }
866 
867 /* Change the comm field on the MCA/INT task to include the pid that
868  * was interrupted, it makes for easier debugging.  If that pid was 0
869  * (swapper or nested MCA/INIT) then use the start of the previous comm
870  * field suffixed with its cpu.
871  */
872 
873 static void
874 ia64_mca_modify_comm(const struct task_struct *previous_current)
875 {
876 	char *p, comm[sizeof(current->comm)];
877 	if (previous_current->pid)
878 		snprintf(comm, sizeof(comm), "%s %d",
879 			current->comm, previous_current->pid);
880 	else {
881 		int l;
882 		if ((p = strchr(previous_current->comm, ' ')))
883 			l = p - previous_current->comm;
884 		else
885 			l = strlen(previous_current->comm);
886 		snprintf(comm, sizeof(comm), "%s %*s %d",
887 			current->comm, l, previous_current->comm,
888 			task_thread_info(previous_current)->cpu);
889 	}
890 	memcpy(current->comm, comm, sizeof(current->comm));
891 }
892 
893 static void
894 finish_pt_regs(struct pt_regs *regs, struct ia64_sal_os_state *sos,
895 		unsigned long *nat)
896 {
897 	const pal_min_state_area_t *ms = sos->pal_min_state;
898 	const u64 *bank;
899 
900 	/* If ipsr.ic then use pmsa_{iip,ipsr,ifs}, else use
901 	 * pmsa_{xip,xpsr,xfs}
902 	 */
903 	if (ia64_psr(regs)->ic) {
904 		regs->cr_iip = ms->pmsa_iip;
905 		regs->cr_ipsr = ms->pmsa_ipsr;
906 		regs->cr_ifs = ms->pmsa_ifs;
907 	} else {
908 		regs->cr_iip = ms->pmsa_xip;
909 		regs->cr_ipsr = ms->pmsa_xpsr;
910 		regs->cr_ifs = ms->pmsa_xfs;
911 
912 		sos->iip = ms->pmsa_iip;
913 		sos->ipsr = ms->pmsa_ipsr;
914 		sos->ifs = ms->pmsa_ifs;
915 	}
916 	regs->pr = ms->pmsa_pr;
917 	regs->b0 = ms->pmsa_br0;
918 	regs->ar_rsc = ms->pmsa_rsc;
919 	copy_reg(&ms->pmsa_gr[1-1], ms->pmsa_nat_bits, &regs->r1, nat);
920 	copy_reg(&ms->pmsa_gr[2-1], ms->pmsa_nat_bits, &regs->r2, nat);
921 	copy_reg(&ms->pmsa_gr[3-1], ms->pmsa_nat_bits, &regs->r3, nat);
922 	copy_reg(&ms->pmsa_gr[8-1], ms->pmsa_nat_bits, &regs->r8, nat);
923 	copy_reg(&ms->pmsa_gr[9-1], ms->pmsa_nat_bits, &regs->r9, nat);
924 	copy_reg(&ms->pmsa_gr[10-1], ms->pmsa_nat_bits, &regs->r10, nat);
925 	copy_reg(&ms->pmsa_gr[11-1], ms->pmsa_nat_bits, &regs->r11, nat);
926 	copy_reg(&ms->pmsa_gr[12-1], ms->pmsa_nat_bits, &regs->r12, nat);
927 	copy_reg(&ms->pmsa_gr[13-1], ms->pmsa_nat_bits, &regs->r13, nat);
928 	copy_reg(&ms->pmsa_gr[14-1], ms->pmsa_nat_bits, &regs->r14, nat);
929 	copy_reg(&ms->pmsa_gr[15-1], ms->pmsa_nat_bits, &regs->r15, nat);
930 	if (ia64_psr(regs)->bn)
931 		bank = ms->pmsa_bank1_gr;
932 	else
933 		bank = ms->pmsa_bank0_gr;
934 	copy_reg(&bank[16-16], ms->pmsa_nat_bits, &regs->r16, nat);
935 	copy_reg(&bank[17-16], ms->pmsa_nat_bits, &regs->r17, nat);
936 	copy_reg(&bank[18-16], ms->pmsa_nat_bits, &regs->r18, nat);
937 	copy_reg(&bank[19-16], ms->pmsa_nat_bits, &regs->r19, nat);
938 	copy_reg(&bank[20-16], ms->pmsa_nat_bits, &regs->r20, nat);
939 	copy_reg(&bank[21-16], ms->pmsa_nat_bits, &regs->r21, nat);
940 	copy_reg(&bank[22-16], ms->pmsa_nat_bits, &regs->r22, nat);
941 	copy_reg(&bank[23-16], ms->pmsa_nat_bits, &regs->r23, nat);
942 	copy_reg(&bank[24-16], ms->pmsa_nat_bits, &regs->r24, nat);
943 	copy_reg(&bank[25-16], ms->pmsa_nat_bits, &regs->r25, nat);
944 	copy_reg(&bank[26-16], ms->pmsa_nat_bits, &regs->r26, nat);
945 	copy_reg(&bank[27-16], ms->pmsa_nat_bits, &regs->r27, nat);
946 	copy_reg(&bank[28-16], ms->pmsa_nat_bits, &regs->r28, nat);
947 	copy_reg(&bank[29-16], ms->pmsa_nat_bits, &regs->r29, nat);
948 	copy_reg(&bank[30-16], ms->pmsa_nat_bits, &regs->r30, nat);
949 	copy_reg(&bank[31-16], ms->pmsa_nat_bits, &regs->r31, nat);
950 }
951 
952 /* On entry to this routine, we are running on the per cpu stack, see
953  * mca_asm.h.  The original stack has not been touched by this event.  Some of
954  * the original stack's registers will be in the RBS on this stack.  This stack
955  * also contains a partial pt_regs and switch_stack, the rest of the data is in
956  * PAL minstate.
957  *
958  * The first thing to do is modify the original stack to look like a blocked
959  * task so we can run backtrace on the original task.  Also mark the per cpu
960  * stack as current to ensure that we use the correct task state, it also means
961  * that we can do backtrace on the MCA/INIT handler code itself.
962  */
963 
964 static struct task_struct *
965 ia64_mca_modify_original_stack(struct pt_regs *regs,
966 		const struct switch_stack *sw,
967 		struct ia64_sal_os_state *sos,
968 		const char *type)
969 {
970 	char *p;
971 	ia64_va va;
972 	extern char ia64_leave_kernel[];	/* Need asm address, not function descriptor */
973 	const pal_min_state_area_t *ms = sos->pal_min_state;
974 	struct task_struct *previous_current;
975 	struct pt_regs *old_regs;
976 	struct switch_stack *old_sw;
977 	unsigned size = sizeof(struct pt_regs) +
978 			sizeof(struct switch_stack) + 16;
979 	unsigned long *old_bspstore, *old_bsp;
980 	unsigned long *new_bspstore, *new_bsp;
981 	unsigned long old_unat, old_rnat, new_rnat, nat;
982 	u64 slots, loadrs = regs->loadrs;
983 	u64 r12 = ms->pmsa_gr[12-1], r13 = ms->pmsa_gr[13-1];
984 	u64 ar_bspstore = regs->ar_bspstore;
985 	u64 ar_bsp = regs->ar_bspstore + (loadrs >> 16);
986 	const char *msg;
987 	int cpu = smp_processor_id();
988 
989 	previous_current = curr_task(cpu);
990 	ia64_set_curr_task(cpu, current);
991 	if ((p = strchr(current->comm, ' ')))
992 		*p = '\0';
993 
994 	/* Best effort attempt to cope with MCA/INIT delivered while in
995 	 * physical mode.
996 	 */
997 	regs->cr_ipsr = ms->pmsa_ipsr;
998 	if (ia64_psr(regs)->dt == 0) {
999 		va.l = r12;
1000 		if (va.f.reg == 0) {
1001 			va.f.reg = 7;
1002 			r12 = va.l;
1003 		}
1004 		va.l = r13;
1005 		if (va.f.reg == 0) {
1006 			va.f.reg = 7;
1007 			r13 = va.l;
1008 		}
1009 	}
1010 	if (ia64_psr(regs)->rt == 0) {
1011 		va.l = ar_bspstore;
1012 		if (va.f.reg == 0) {
1013 			va.f.reg = 7;
1014 			ar_bspstore = va.l;
1015 		}
1016 		va.l = ar_bsp;
1017 		if (va.f.reg == 0) {
1018 			va.f.reg = 7;
1019 			ar_bsp = va.l;
1020 		}
1021 	}
1022 
1023 	/* mca_asm.S ia64_old_stack() cannot assume that the dirty registers
1024 	 * have been copied to the old stack, the old stack may fail the
1025 	 * validation tests below.  So ia64_old_stack() must restore the dirty
1026 	 * registers from the new stack.  The old and new bspstore probably
1027 	 * have different alignments, so loadrs calculated on the old bsp
1028 	 * cannot be used to restore from the new bsp.  Calculate a suitable
1029 	 * loadrs for the new stack and save it in the new pt_regs, where
1030 	 * ia64_old_stack() can get it.
1031 	 */
1032 	old_bspstore = (unsigned long *)ar_bspstore;
1033 	old_bsp = (unsigned long *)ar_bsp;
1034 	slots = ia64_rse_num_regs(old_bspstore, old_bsp);
1035 	new_bspstore = (unsigned long *)((u64)current + IA64_RBS_OFFSET);
1036 	new_bsp = ia64_rse_skip_regs(new_bspstore, slots);
1037 	regs->loadrs = (new_bsp - new_bspstore) * 8 << 16;
1038 
1039 	/* Verify the previous stack state before we change it */
1040 	if (user_mode(regs)) {
1041 		msg = "occurred in user space";
1042 		/* previous_current is guaranteed to be valid when the task was
1043 		 * in user space, so ...
1044 		 */
1045 		ia64_mca_modify_comm(previous_current);
1046 		goto no_mod;
1047 	}
1048 
1049 	if (r13 != sos->prev_IA64_KR_CURRENT) {
1050 		msg = "inconsistent previous current and r13";
1051 		goto no_mod;
1052 	}
1053 
1054 	if (!mca_recover_range(ms->pmsa_iip)) {
1055 		if ((r12 - r13) >= KERNEL_STACK_SIZE) {
1056 			msg = "inconsistent r12 and r13";
1057 			goto no_mod;
1058 		}
1059 		if ((ar_bspstore - r13) >= KERNEL_STACK_SIZE) {
1060 			msg = "inconsistent ar.bspstore and r13";
1061 			goto no_mod;
1062 		}
1063 		va.p = old_bspstore;
1064 		if (va.f.reg < 5) {
1065 			msg = "old_bspstore is in the wrong region";
1066 			goto no_mod;
1067 		}
1068 		if ((ar_bsp - r13) >= KERNEL_STACK_SIZE) {
1069 			msg = "inconsistent ar.bsp and r13";
1070 			goto no_mod;
1071 		}
1072 		size += (ia64_rse_skip_regs(old_bspstore, slots) - old_bspstore) * 8;
1073 		if (ar_bspstore + size > r12) {
1074 			msg = "no room for blocked state";
1075 			goto no_mod;
1076 		}
1077 	}
1078 
1079 	ia64_mca_modify_comm(previous_current);
1080 
1081 	/* Make the original task look blocked.  First stack a struct pt_regs,
1082 	 * describing the state at the time of interrupt.  mca_asm.S built a
1083 	 * partial pt_regs, copy it and fill in the blanks using minstate.
1084 	 */
1085 	p = (char *)r12 - sizeof(*regs);
1086 	old_regs = (struct pt_regs *)p;
1087 	memcpy(old_regs, regs, sizeof(*regs));
1088 	old_regs->loadrs = loadrs;
1089 	old_unat = old_regs->ar_unat;
1090 	finish_pt_regs(old_regs, sos, &old_unat);
1091 
1092 	/* Next stack a struct switch_stack.  mca_asm.S built a partial
1093 	 * switch_stack, copy it and fill in the blanks using pt_regs and
1094 	 * minstate.
1095 	 *
1096 	 * In the synthesized switch_stack, b0 points to ia64_leave_kernel,
1097 	 * ar.pfs is set to 0.
1098 	 *
1099 	 * unwind.c::unw_unwind() does special processing for interrupt frames.
1100 	 * It checks if the PRED_NON_SYSCALL predicate is set, if the predicate
1101 	 * is clear then unw_unwind() does _not_ adjust bsp over pt_regs.  Not
1102 	 * that this is documented, of course.  Set PRED_NON_SYSCALL in the
1103 	 * switch_stack on the original stack so it will unwind correctly when
1104 	 * unwind.c reads pt_regs.
1105 	 *
1106 	 * thread.ksp is updated to point to the synthesized switch_stack.
1107 	 */
1108 	p -= sizeof(struct switch_stack);
1109 	old_sw = (struct switch_stack *)p;
1110 	memcpy(old_sw, sw, sizeof(*sw));
1111 	old_sw->caller_unat = old_unat;
1112 	old_sw->ar_fpsr = old_regs->ar_fpsr;
1113 	copy_reg(&ms->pmsa_gr[4-1], ms->pmsa_nat_bits, &old_sw->r4, &old_unat);
1114 	copy_reg(&ms->pmsa_gr[5-1], ms->pmsa_nat_bits, &old_sw->r5, &old_unat);
1115 	copy_reg(&ms->pmsa_gr[6-1], ms->pmsa_nat_bits, &old_sw->r6, &old_unat);
1116 	copy_reg(&ms->pmsa_gr[7-1], ms->pmsa_nat_bits, &old_sw->r7, &old_unat);
1117 	old_sw->b0 = (u64)ia64_leave_kernel;
1118 	old_sw->b1 = ms->pmsa_br1;
1119 	old_sw->ar_pfs = 0;
1120 	old_sw->ar_unat = old_unat;
1121 	old_sw->pr = old_regs->pr | (1UL << PRED_NON_SYSCALL);
1122 	previous_current->thread.ksp = (u64)p - 16;
1123 
1124 	/* Finally copy the original stack's registers back to its RBS.
1125 	 * Registers from ar.bspstore through ar.bsp at the time of the event
1126 	 * are in the current RBS, copy them back to the original stack.  The
1127 	 * copy must be done register by register because the original bspstore
1128 	 * and the current one have different alignments, so the saved RNAT
1129 	 * data occurs at different places.
1130 	 *
1131 	 * mca_asm does cover, so the old_bsp already includes all registers at
1132 	 * the time of MCA/INIT.  It also does flushrs, so all registers before
1133 	 * this function have been written to backing store on the MCA/INIT
1134 	 * stack.
1135 	 */
1136 	new_rnat = ia64_get_rnat(ia64_rse_rnat_addr(new_bspstore));
1137 	old_rnat = regs->ar_rnat;
1138 	while (slots--) {
1139 		if (ia64_rse_is_rnat_slot(new_bspstore)) {
1140 			new_rnat = ia64_get_rnat(new_bspstore++);
1141 		}
1142 		if (ia64_rse_is_rnat_slot(old_bspstore)) {
1143 			*old_bspstore++ = old_rnat;
1144 			old_rnat = 0;
1145 		}
1146 		nat = (new_rnat >> ia64_rse_slot_num(new_bspstore)) & 1UL;
1147 		old_rnat &= ~(1UL << ia64_rse_slot_num(old_bspstore));
1148 		old_rnat |= (nat << ia64_rse_slot_num(old_bspstore));
1149 		*old_bspstore++ = *new_bspstore++;
1150 	}
1151 	old_sw->ar_bspstore = (unsigned long)old_bspstore;
1152 	old_sw->ar_rnat = old_rnat;
1153 
1154 	sos->prev_task = previous_current;
1155 	return previous_current;
1156 
1157 no_mod:
1158 	mprintk(KERN_INFO "cpu %d, %s %s, original stack not modified\n",
1159 			smp_processor_id(), type, msg);
1160 	old_unat = regs->ar_unat;
1161 	finish_pt_regs(regs, sos, &old_unat);
1162 	return previous_current;
1163 }
1164 
1165 /* The monarch/slave interaction is based on monarch_cpu and requires that all
1166  * slaves have entered rendezvous before the monarch leaves.  If any cpu has
1167  * not entered rendezvous yet then wait a bit.  The assumption is that any
1168  * slave that has not rendezvoused after a reasonable time is never going to do
1169  * so.  In this context, slave includes cpus that respond to the MCA rendezvous
1170  * interrupt, as well as cpus that receive the INIT slave event.
1171  */
1172 
1173 static void
1174 ia64_wait_for_slaves(int monarch, const char *type)
1175 {
1176 	int c, i , wait;
1177 
1178 	/*
1179 	 * wait 5 seconds total for slaves (arbitrary)
1180 	 */
1181 	for (i = 0; i < 5000; i++) {
1182 		wait = 0;
1183 		for_each_online_cpu(c) {
1184 			if (c == monarch)
1185 				continue;
1186 			if (ia64_mc_info.imi_rendez_checkin[c]
1187 					== IA64_MCA_RENDEZ_CHECKIN_NOTDONE) {
1188 				udelay(1000);		/* short wait */
1189 				wait = 1;
1190 				break;
1191 			}
1192 		}
1193 		if (!wait)
1194 			goto all_in;
1195 	}
1196 
1197 	/*
1198 	 * Maybe slave(s) dead. Print buffered messages immediately.
1199 	 */
1200 	ia64_mlogbuf_finish(0);
1201 	mprintk(KERN_INFO "OS %s slave did not rendezvous on cpu", type);
1202 	for_each_online_cpu(c) {
1203 		if (c == monarch)
1204 			continue;
1205 		if (ia64_mc_info.imi_rendez_checkin[c] == IA64_MCA_RENDEZ_CHECKIN_NOTDONE)
1206 			mprintk(" %d", c);
1207 	}
1208 	mprintk("\n");
1209 	return;
1210 
1211 all_in:
1212 	mprintk(KERN_INFO "All OS %s slaves have reached rendezvous\n", type);
1213 	return;
1214 }
1215 
1216 /*  mca_insert_tr
1217  *
1218  *  Switch rid when TR reload and needed!
1219  *  iord: 1: itr, 2: itr;
1220  *
1221 */
1222 static void mca_insert_tr(u64 iord)
1223 {
1224 
1225 	int i;
1226 	u64 old_rr;
1227 	struct ia64_tr_entry *p;
1228 	unsigned long psr;
1229 	int cpu = smp_processor_id();
1230 
1231 	if (!ia64_idtrs[cpu])
1232 		return;
1233 
1234 	psr = ia64_clear_ic();
1235 	for (i = IA64_TR_ALLOC_BASE; i < IA64_TR_ALLOC_MAX; i++) {
1236 		p = ia64_idtrs[cpu] + (iord - 1) * IA64_TR_ALLOC_MAX;
1237 		if (p->pte & 0x1) {
1238 			old_rr = ia64_get_rr(p->ifa);
1239 			if (old_rr != p->rr) {
1240 				ia64_set_rr(p->ifa, p->rr);
1241 				ia64_srlz_d();
1242 			}
1243 			ia64_ptr(iord, p->ifa, p->itir >> 2);
1244 			ia64_srlz_i();
1245 			if (iord & 0x1) {
1246 				ia64_itr(0x1, i, p->ifa, p->pte, p->itir >> 2);
1247 				ia64_srlz_i();
1248 			}
1249 			if (iord & 0x2) {
1250 				ia64_itr(0x2, i, p->ifa, p->pte, p->itir >> 2);
1251 				ia64_srlz_i();
1252 			}
1253 			if (old_rr != p->rr) {
1254 				ia64_set_rr(p->ifa, old_rr);
1255 				ia64_srlz_d();
1256 			}
1257 		}
1258 	}
1259 	ia64_set_psr(psr);
1260 }
1261 
1262 /*
1263  * ia64_mca_handler
1264  *
1265  *	This is uncorrectable machine check handler called from OS_MCA
1266  *	dispatch code which is in turn called from SAL_CHECK().
1267  *	This is the place where the core of OS MCA handling is done.
1268  *	Right now the logs are extracted and displayed in a well-defined
1269  *	format. This handler code is supposed to be run only on the
1270  *	monarch processor. Once the monarch is done with MCA handling
1271  *	further MCA logging is enabled by clearing logs.
1272  *	Monarch also has the duty of sending wakeup-IPIs to pull the
1273  *	slave processors out of rendezvous spinloop.
1274  *
1275  *	If multiple processors call into OS_MCA, the first will become
1276  *	the monarch.  Subsequent cpus will be recorded in the mca_cpu
1277  *	bitmask.  After the first monarch has processed its MCA, it
1278  *	will wake up the next cpu in the mca_cpu bitmask and then go
1279  *	into the rendezvous loop.  When all processors have serviced
1280  *	their MCA, the last monarch frees up the rest of the processors.
1281  */
1282 void
1283 ia64_mca_handler(struct pt_regs *regs, struct switch_stack *sw,
1284 		 struct ia64_sal_os_state *sos)
1285 {
1286 	int recover, cpu = smp_processor_id();
1287 	struct task_struct *previous_current;
1288 	struct ia64_mca_notify_die nd =
1289 		{ .sos = sos, .monarch_cpu = &monarch_cpu, .data = &recover };
1290 	static atomic_t mca_count;
1291 	static cpumask_t mca_cpu;
1292 
1293 	if (atomic_add_return(1, &mca_count) == 1) {
1294 		monarch_cpu = cpu;
1295 		sos->monarch = 1;
1296 	} else {
1297 		cpumask_set_cpu(cpu, &mca_cpu);
1298 		sos->monarch = 0;
1299 	}
1300 	mprintk(KERN_INFO "Entered OS MCA handler. PSP=%lx cpu=%d "
1301 		"monarch=%ld\n", sos->proc_state_param, cpu, sos->monarch);
1302 
1303 	previous_current = ia64_mca_modify_original_stack(regs, sw, sos, "MCA");
1304 
1305 	NOTIFY_MCA(DIE_MCA_MONARCH_ENTER, regs, (long)&nd, 1);
1306 
1307 	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_CONCURRENT_MCA;
1308 	if (sos->monarch) {
1309 		ia64_wait_for_slaves(cpu, "MCA");
1310 
1311 		/* Wakeup all the processors which are spinning in the
1312 		 * rendezvous loop.  They will leave SAL, then spin in the OS
1313 		 * with interrupts disabled until this monarch cpu leaves the
1314 		 * MCA handler.  That gets control back to the OS so we can
1315 		 * backtrace the other cpus, backtrace when spinning in SAL
1316 		 * does not work.
1317 		 */
1318 		ia64_mca_wakeup_all();
1319 	} else {
1320 		while (cpumask_test_cpu(cpu, &mca_cpu))
1321 			cpu_relax();	/* spin until monarch wakes us */
1322 	}
1323 
1324 	NOTIFY_MCA(DIE_MCA_MONARCH_PROCESS, regs, (long)&nd, 1);
1325 
1326 	/* Get the MCA error record and log it */
1327 	ia64_mca_log_sal_error_record(SAL_INFO_TYPE_MCA);
1328 
1329 	/* MCA error recovery */
1330 	recover = (ia64_mca_ucmc_extension
1331 		&& ia64_mca_ucmc_extension(
1332 			IA64_LOG_CURR_BUFFER(SAL_INFO_TYPE_MCA),
1333 			sos));
1334 
1335 	if (recover) {
1336 		sal_log_record_header_t *rh = IA64_LOG_CURR_BUFFER(SAL_INFO_TYPE_MCA);
1337 		rh->severity = sal_log_severity_corrected;
1338 		ia64_sal_clear_state_info(SAL_INFO_TYPE_MCA);
1339 		sos->os_status = IA64_MCA_CORRECTED;
1340 	} else {
1341 		/* Dump buffered message to console */
1342 		ia64_mlogbuf_finish(1);
1343 	}
1344 
1345 	if (__this_cpu_read(ia64_mca_tr_reload)) {
1346 		mca_insert_tr(0x1); /*Reload dynamic itrs*/
1347 		mca_insert_tr(0x2); /*Reload dynamic itrs*/
1348 	}
1349 
1350 	NOTIFY_MCA(DIE_MCA_MONARCH_LEAVE, regs, (long)&nd, 1);
1351 
1352 	if (atomic_dec_return(&mca_count) > 0) {
1353 		int i;
1354 
1355 		/* wake up the next monarch cpu,
1356 		 * and put this cpu in the rendez loop.
1357 		 */
1358 		for_each_online_cpu(i) {
1359 			if (cpumask_test_cpu(i, &mca_cpu)) {
1360 				monarch_cpu = i;
1361 				cpumask_clear_cpu(i, &mca_cpu);	/* wake next cpu */
1362 				while (monarch_cpu != -1)
1363 					cpu_relax();	/* spin until last cpu leaves */
1364 				ia64_set_curr_task(cpu, previous_current);
1365 				ia64_mc_info.imi_rendez_checkin[cpu]
1366 						= IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1367 				return;
1368 			}
1369 		}
1370 	}
1371 	ia64_set_curr_task(cpu, previous_current);
1372 	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1373 	monarch_cpu = -1;	/* This frees the slaves and previous monarchs */
1374 }
1375 
1376 static DECLARE_WORK(cmc_disable_work, ia64_mca_cmc_vector_disable_keventd);
1377 static DECLARE_WORK(cmc_enable_work, ia64_mca_cmc_vector_enable_keventd);
1378 
1379 /*
1380  * ia64_mca_cmc_int_handler
1381  *
1382  *  This is corrected machine check interrupt handler.
1383  *	Right now the logs are extracted and displayed in a well-defined
1384  *	format.
1385  *
1386  * Inputs
1387  *      interrupt number
1388  *      client data arg ptr
1389  *
1390  * Outputs
1391  *	None
1392  */
1393 static irqreturn_t
1394 ia64_mca_cmc_int_handler(int cmc_irq, void *arg)
1395 {
1396 	static unsigned long	cmc_history[CMC_HISTORY_LENGTH];
1397 	static int		index;
1398 	static DEFINE_SPINLOCK(cmc_history_lock);
1399 
1400 	IA64_MCA_DEBUG("%s: received interrupt vector = %#x on CPU %d\n",
1401 		       __func__, cmc_irq, smp_processor_id());
1402 
1403 	/* SAL spec states this should run w/ interrupts enabled */
1404 	local_irq_enable();
1405 
1406 	spin_lock(&cmc_history_lock);
1407 	if (!cmc_polling_enabled) {
1408 		int i, count = 1; /* we know 1 happened now */
1409 		unsigned long now = jiffies;
1410 
1411 		for (i = 0; i < CMC_HISTORY_LENGTH; i++) {
1412 			if (now - cmc_history[i] <= HZ)
1413 				count++;
1414 		}
1415 
1416 		IA64_MCA_DEBUG(KERN_INFO "CMC threshold %d/%d\n", count, CMC_HISTORY_LENGTH);
1417 		if (count >= CMC_HISTORY_LENGTH) {
1418 
1419 			cmc_polling_enabled = 1;
1420 			spin_unlock(&cmc_history_lock);
1421 			/* If we're being hit with CMC interrupts, we won't
1422 			 * ever execute the schedule_work() below.  Need to
1423 			 * disable CMC interrupts on this processor now.
1424 			 */
1425 			ia64_mca_cmc_vector_disable(NULL);
1426 			schedule_work(&cmc_disable_work);
1427 
1428 			/*
1429 			 * Corrected errors will still be corrected, but
1430 			 * make sure there's a log somewhere that indicates
1431 			 * something is generating more than we can handle.
1432 			 */
1433 			printk(KERN_WARNING "WARNING: Switching to polling CMC handler; error records may be lost\n");
1434 
1435 			mod_timer(&cmc_poll_timer, jiffies + CMC_POLL_INTERVAL);
1436 
1437 			/* lock already released, get out now */
1438 			goto out;
1439 		} else {
1440 			cmc_history[index++] = now;
1441 			if (index == CMC_HISTORY_LENGTH)
1442 				index = 0;
1443 		}
1444 	}
1445 	spin_unlock(&cmc_history_lock);
1446 out:
1447 	/* Get the CMC error record and log it */
1448 	ia64_mca_log_sal_error_record(SAL_INFO_TYPE_CMC);
1449 
1450 	local_irq_disable();
1451 
1452 	return IRQ_HANDLED;
1453 }
1454 
1455 /*
1456  *  ia64_mca_cmc_int_caller
1457  *
1458  * 	Triggered by sw interrupt from CMC polling routine.  Calls
1459  * 	real interrupt handler and either triggers a sw interrupt
1460  * 	on the next cpu or does cleanup at the end.
1461  *
1462  * Inputs
1463  *	interrupt number
1464  *	client data arg ptr
1465  * Outputs
1466  * 	handled
1467  */
1468 static irqreturn_t
1469 ia64_mca_cmc_int_caller(int cmc_irq, void *arg)
1470 {
1471 	static int start_count = -1;
1472 	unsigned int cpuid;
1473 
1474 	cpuid = smp_processor_id();
1475 
1476 	/* If first cpu, update count */
1477 	if (start_count == -1)
1478 		start_count = IA64_LOG_COUNT(SAL_INFO_TYPE_CMC);
1479 
1480 	ia64_mca_cmc_int_handler(cmc_irq, arg);
1481 
1482 	cpuid = cpumask_next(cpuid+1, cpu_online_mask);
1483 
1484 	if (cpuid < nr_cpu_ids) {
1485 		ia64_send_ipi(cpuid, IA64_CMCP_VECTOR, IA64_IPI_DM_INT, 0);
1486 	} else {
1487 		/* If no log record, switch out of polling mode */
1488 		if (start_count == IA64_LOG_COUNT(SAL_INFO_TYPE_CMC)) {
1489 
1490 			printk(KERN_WARNING "Returning to interrupt driven CMC handler\n");
1491 			schedule_work(&cmc_enable_work);
1492 			cmc_polling_enabled = 0;
1493 
1494 		} else {
1495 
1496 			mod_timer(&cmc_poll_timer, jiffies + CMC_POLL_INTERVAL);
1497 		}
1498 
1499 		start_count = -1;
1500 	}
1501 
1502 	return IRQ_HANDLED;
1503 }
1504 
1505 /*
1506  *  ia64_mca_cmc_poll
1507  *
1508  *	Poll for Corrected Machine Checks (CMCs)
1509  *
1510  * Inputs   :   dummy(unused)
1511  * Outputs  :   None
1512  *
1513  */
1514 static void
1515 ia64_mca_cmc_poll (struct timer_list *unused)
1516 {
1517 	/* Trigger a CMC interrupt cascade  */
1518 	ia64_send_ipi(cpumask_first(cpu_online_mask), IA64_CMCP_VECTOR,
1519 							IA64_IPI_DM_INT, 0);
1520 }
1521 
1522 /*
1523  *  ia64_mca_cpe_int_caller
1524  *
1525  * 	Triggered by sw interrupt from CPE polling routine.  Calls
1526  * 	real interrupt handler and either triggers a sw interrupt
1527  * 	on the next cpu or does cleanup at the end.
1528  *
1529  * Inputs
1530  *	interrupt number
1531  *	client data arg ptr
1532  * Outputs
1533  * 	handled
1534  */
1535 static irqreturn_t
1536 ia64_mca_cpe_int_caller(int cpe_irq, void *arg)
1537 {
1538 	static int start_count = -1;
1539 	static int poll_time = MIN_CPE_POLL_INTERVAL;
1540 	unsigned int cpuid;
1541 
1542 	cpuid = smp_processor_id();
1543 
1544 	/* If first cpu, update count */
1545 	if (start_count == -1)
1546 		start_count = IA64_LOG_COUNT(SAL_INFO_TYPE_CPE);
1547 
1548 	ia64_mca_cpe_int_handler(cpe_irq, arg);
1549 
1550 	cpuid = cpumask_next(cpuid+1, cpu_online_mask);
1551 
1552 	if (cpuid < NR_CPUS) {
1553 		ia64_send_ipi(cpuid, IA64_CPEP_VECTOR, IA64_IPI_DM_INT, 0);
1554 	} else {
1555 		/*
1556 		 * If a log was recorded, increase our polling frequency,
1557 		 * otherwise, backoff or return to interrupt mode.
1558 		 */
1559 		if (start_count != IA64_LOG_COUNT(SAL_INFO_TYPE_CPE)) {
1560 			poll_time = max(MIN_CPE_POLL_INTERVAL, poll_time / 2);
1561 		} else if (cpe_vector < 0) {
1562 			poll_time = min(MAX_CPE_POLL_INTERVAL, poll_time * 2);
1563 		} else {
1564 			poll_time = MIN_CPE_POLL_INTERVAL;
1565 
1566 			printk(KERN_WARNING "Returning to interrupt driven CPE handler\n");
1567 			enable_irq(local_vector_to_irq(IA64_CPE_VECTOR));
1568 			cpe_poll_enabled = 0;
1569 		}
1570 
1571 		if (cpe_poll_enabled)
1572 			mod_timer(&cpe_poll_timer, jiffies + poll_time);
1573 		start_count = -1;
1574 	}
1575 
1576 	return IRQ_HANDLED;
1577 }
1578 
1579 /*
1580  *  ia64_mca_cpe_poll
1581  *
1582  *	Poll for Corrected Platform Errors (CPEs), trigger interrupt
1583  *	on first cpu, from there it will trickle through all the cpus.
1584  *
1585  * Inputs   :   dummy(unused)
1586  * Outputs  :   None
1587  *
1588  */
1589 static void
1590 ia64_mca_cpe_poll (struct timer_list *unused)
1591 {
1592 	/* Trigger a CPE interrupt cascade  */
1593 	ia64_send_ipi(cpumask_first(cpu_online_mask), IA64_CPEP_VECTOR,
1594 							IA64_IPI_DM_INT, 0);
1595 }
1596 
1597 static int
1598 default_monarch_init_process(struct notifier_block *self, unsigned long val, void *data)
1599 {
1600 	int c;
1601 	struct task_struct *g, *t;
1602 	if (val != DIE_INIT_MONARCH_PROCESS)
1603 		return NOTIFY_DONE;
1604 #ifdef CONFIG_KEXEC
1605 	if (atomic_read(&kdump_in_progress))
1606 		return NOTIFY_DONE;
1607 #endif
1608 
1609 	/*
1610 	 * FIXME: mlogbuf will brim over with INIT stack dumps.
1611 	 * To enable show_stack from INIT, we use oops_in_progress which should
1612 	 * be used in real oops. This would cause something wrong after INIT.
1613 	 */
1614 	BREAK_LOGLEVEL(console_loglevel);
1615 	ia64_mlogbuf_dump_from_init();
1616 
1617 	printk(KERN_ERR "Processes interrupted by INIT -");
1618 	for_each_online_cpu(c) {
1619 		struct ia64_sal_os_state *s;
1620 		t = __va(__per_cpu_mca[c] + IA64_MCA_CPU_INIT_STACK_OFFSET);
1621 		s = (struct ia64_sal_os_state *)((char *)t + MCA_SOS_OFFSET);
1622 		g = s->prev_task;
1623 		if (g) {
1624 			if (g->pid)
1625 				printk(" %d", g->pid);
1626 			else
1627 				printk(" %d (cpu %d task 0x%p)", g->pid, task_cpu(g), g);
1628 		}
1629 	}
1630 	printk("\n\n");
1631 	if (read_trylock(&tasklist_lock)) {
1632 		do_each_thread (g, t) {
1633 			printk("\nBacktrace of pid %d (%s)\n", t->pid, t->comm);
1634 			show_stack(t, NULL);
1635 		} while_each_thread (g, t);
1636 		read_unlock(&tasklist_lock);
1637 	}
1638 	/* FIXME: This will not restore zapped printk locks. */
1639 	RESTORE_LOGLEVEL(console_loglevel);
1640 	return NOTIFY_DONE;
1641 }
1642 
1643 /*
1644  * C portion of the OS INIT handler
1645  *
1646  * Called from ia64_os_init_dispatch
1647  *
1648  * Inputs: pointer to pt_regs where processor info was saved.  SAL/OS state for
1649  * this event.  This code is used for both monarch and slave INIT events, see
1650  * sos->monarch.
1651  *
1652  * All INIT events switch to the INIT stack and change the previous process to
1653  * blocked status.  If one of the INIT events is the monarch then we are
1654  * probably processing the nmi button/command.  Use the monarch cpu to dump all
1655  * the processes.  The slave INIT events all spin until the monarch cpu
1656  * returns.  We can also get INIT slave events for MCA, in which case the MCA
1657  * process is the monarch.
1658  */
1659 
1660 void
1661 ia64_init_handler(struct pt_regs *regs, struct switch_stack *sw,
1662 		  struct ia64_sal_os_state *sos)
1663 {
1664 	static atomic_t slaves;
1665 	static atomic_t monarchs;
1666 	struct task_struct *previous_current;
1667 	int cpu = smp_processor_id();
1668 	struct ia64_mca_notify_die nd =
1669 		{ .sos = sos, .monarch_cpu = &monarch_cpu };
1670 
1671 	NOTIFY_INIT(DIE_INIT_ENTER, regs, (long)&nd, 0);
1672 
1673 	mprintk(KERN_INFO "Entered OS INIT handler. PSP=%lx cpu=%d monarch=%ld\n",
1674 		sos->proc_state_param, cpu, sos->monarch);
1675 	salinfo_log_wakeup(SAL_INFO_TYPE_INIT, NULL, 0, 0);
1676 
1677 	previous_current = ia64_mca_modify_original_stack(regs, sw, sos, "INIT");
1678 	sos->os_status = IA64_INIT_RESUME;
1679 
1680 	/* FIXME: Workaround for broken proms that drive all INIT events as
1681 	 * slaves.  The last slave that enters is promoted to be a monarch.
1682 	 * Remove this code in September 2006, that gives platforms a year to
1683 	 * fix their proms and get their customers updated.
1684 	 */
1685 	if (!sos->monarch && atomic_add_return(1, &slaves) == num_online_cpus()) {
1686 		mprintk(KERN_WARNING "%s: Promoting cpu %d to monarch.\n",
1687 		        __func__, cpu);
1688 		atomic_dec(&slaves);
1689 		sos->monarch = 1;
1690 	}
1691 
1692 	/* FIXME: Workaround for broken proms that drive all INIT events as
1693 	 * monarchs.  Second and subsequent monarchs are demoted to slaves.
1694 	 * Remove this code in September 2006, that gives platforms a year to
1695 	 * fix their proms and get their customers updated.
1696 	 */
1697 	if (sos->monarch && atomic_add_return(1, &monarchs) > 1) {
1698 		mprintk(KERN_WARNING "%s: Demoting cpu %d to slave.\n",
1699 			       __func__, cpu);
1700 		atomic_dec(&monarchs);
1701 		sos->monarch = 0;
1702 	}
1703 
1704 	if (!sos->monarch) {
1705 		ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_INIT;
1706 
1707 #ifdef CONFIG_KEXEC
1708 		while (monarch_cpu == -1 && !atomic_read(&kdump_in_progress))
1709 			udelay(1000);
1710 #else
1711 		while (monarch_cpu == -1)
1712 			cpu_relax();	/* spin until monarch enters */
1713 #endif
1714 
1715 		NOTIFY_INIT(DIE_INIT_SLAVE_ENTER, regs, (long)&nd, 1);
1716 		NOTIFY_INIT(DIE_INIT_SLAVE_PROCESS, regs, (long)&nd, 1);
1717 
1718 #ifdef CONFIG_KEXEC
1719 		while (monarch_cpu != -1 && !atomic_read(&kdump_in_progress))
1720 			udelay(1000);
1721 #else
1722 		while (monarch_cpu != -1)
1723 			cpu_relax();	/* spin until monarch leaves */
1724 #endif
1725 
1726 		NOTIFY_INIT(DIE_INIT_SLAVE_LEAVE, regs, (long)&nd, 1);
1727 
1728 		mprintk("Slave on cpu %d returning to normal service.\n", cpu);
1729 		ia64_set_curr_task(cpu, previous_current);
1730 		ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1731 		atomic_dec(&slaves);
1732 		return;
1733 	}
1734 
1735 	monarch_cpu = cpu;
1736 	NOTIFY_INIT(DIE_INIT_MONARCH_ENTER, regs, (long)&nd, 1);
1737 
1738 	/*
1739 	 * Wait for a bit.  On some machines (e.g., HP's zx2000 and zx6000, INIT can be
1740 	 * generated via the BMC's command-line interface, but since the console is on the
1741 	 * same serial line, the user will need some time to switch out of the BMC before
1742 	 * the dump begins.
1743 	 */
1744 	mprintk("Delaying for 5 seconds...\n");
1745 	udelay(5*1000000);
1746 	ia64_wait_for_slaves(cpu, "INIT");
1747 	/* If nobody intercepts DIE_INIT_MONARCH_PROCESS then we drop through
1748 	 * to default_monarch_init_process() above and just print all the
1749 	 * tasks.
1750 	 */
1751 	NOTIFY_INIT(DIE_INIT_MONARCH_PROCESS, regs, (long)&nd, 1);
1752 	NOTIFY_INIT(DIE_INIT_MONARCH_LEAVE, regs, (long)&nd, 1);
1753 
1754 	mprintk("\nINIT dump complete.  Monarch on cpu %d returning to normal service.\n", cpu);
1755 	atomic_dec(&monarchs);
1756 	ia64_set_curr_task(cpu, previous_current);
1757 	monarch_cpu = -1;
1758 	return;
1759 }
1760 
1761 static int __init
1762 ia64_mca_disable_cpe_polling(char *str)
1763 {
1764 	cpe_poll_enabled = 0;
1765 	return 1;
1766 }
1767 
1768 __setup("disable_cpe_poll", ia64_mca_disable_cpe_polling);
1769 
1770 /* Minimal format of the MCA/INIT stacks.  The pseudo processes that run on
1771  * these stacks can never sleep, they cannot return from the kernel to user
1772  * space, they do not appear in a normal ps listing.  So there is no need to
1773  * format most of the fields.
1774  */
1775 
1776 static void
1777 format_mca_init_stack(void *mca_data, unsigned long offset,
1778 		const char *type, int cpu)
1779 {
1780 	struct task_struct *p = (struct task_struct *)((char *)mca_data + offset);
1781 	struct thread_info *ti;
1782 	memset(p, 0, KERNEL_STACK_SIZE);
1783 	ti = task_thread_info(p);
1784 	ti->flags = _TIF_MCA_INIT;
1785 	ti->preempt_count = 1;
1786 	ti->task = p;
1787 	ti->cpu = cpu;
1788 	p->stack = ti;
1789 	p->state = TASK_UNINTERRUPTIBLE;
1790 	cpumask_set_cpu(cpu, &p->cpus_mask);
1791 	INIT_LIST_HEAD(&p->tasks);
1792 	p->parent = p->real_parent = p->group_leader = p;
1793 	INIT_LIST_HEAD(&p->children);
1794 	INIT_LIST_HEAD(&p->sibling);
1795 	strncpy(p->comm, type, sizeof(p->comm)-1);
1796 }
1797 
1798 /* Caller prevents this from being called after init */
1799 static void * __ref mca_bootmem(void)
1800 {
1801 	return memblock_alloc(sizeof(struct ia64_mca_cpu), KERNEL_STACK_SIZE);
1802 }
1803 
1804 /* Do per-CPU MCA-related initialization.  */
1805 void
1806 ia64_mca_cpu_init(void *cpu_data)
1807 {
1808 	void *pal_vaddr;
1809 	void *data;
1810 	long sz = sizeof(struct ia64_mca_cpu);
1811 	int cpu = smp_processor_id();
1812 	static int first_time = 1;
1813 
1814 	/*
1815 	 * Structure will already be allocated if cpu has been online,
1816 	 * then offlined.
1817 	 */
1818 	if (__per_cpu_mca[cpu]) {
1819 		data = __va(__per_cpu_mca[cpu]);
1820 	} else {
1821 		if (first_time) {
1822 			data = mca_bootmem();
1823 			first_time = 0;
1824 		} else
1825 			data = (void *)__get_free_pages(GFP_KERNEL,
1826 							get_order(sz));
1827 		if (!data)
1828 			panic("Could not allocate MCA memory for cpu %d\n",
1829 					cpu);
1830 	}
1831 	format_mca_init_stack(data, offsetof(struct ia64_mca_cpu, mca_stack),
1832 		"MCA", cpu);
1833 	format_mca_init_stack(data, offsetof(struct ia64_mca_cpu, init_stack),
1834 		"INIT", cpu);
1835 	__this_cpu_write(ia64_mca_data, (__per_cpu_mca[cpu] = __pa(data)));
1836 
1837 	/*
1838 	 * Stash away a copy of the PTE needed to map the per-CPU page.
1839 	 * We may need it during MCA recovery.
1840 	 */
1841 	__this_cpu_write(ia64_mca_per_cpu_pte,
1842 		pte_val(mk_pte_phys(__pa(cpu_data), PAGE_KERNEL)));
1843 
1844 	/*
1845 	 * Also, stash away a copy of the PAL address and the PTE
1846 	 * needed to map it.
1847 	 */
1848 	pal_vaddr = efi_get_pal_addr();
1849 	if (!pal_vaddr)
1850 		return;
1851 	__this_cpu_write(ia64_mca_pal_base,
1852 		GRANULEROUNDDOWN((unsigned long) pal_vaddr));
1853 	__this_cpu_write(ia64_mca_pal_pte, pte_val(mk_pte_phys(__pa(pal_vaddr),
1854 							      PAGE_KERNEL)));
1855 }
1856 
1857 static int ia64_mca_cpu_online(unsigned int cpu)
1858 {
1859 	unsigned long flags;
1860 
1861 	local_irq_save(flags);
1862 	if (!cmc_polling_enabled)
1863 		ia64_mca_cmc_vector_enable(NULL);
1864 	local_irq_restore(flags);
1865 	return 0;
1866 }
1867 
1868 /*
1869  * ia64_mca_init
1870  *
1871  *  Do all the system level mca specific initialization.
1872  *
1873  *	1. Register spinloop and wakeup request interrupt vectors
1874  *
1875  *	2. Register OS_MCA handler entry point
1876  *
1877  *	3. Register OS_INIT handler entry point
1878  *
1879  *  4. Initialize MCA/CMC/INIT related log buffers maintained by the OS.
1880  *
1881  *  Note that this initialization is done very early before some kernel
1882  *  services are available.
1883  *
1884  *  Inputs  :   None
1885  *
1886  *  Outputs :   None
1887  */
1888 void __init
1889 ia64_mca_init(void)
1890 {
1891 	ia64_fptr_t *init_hldlr_ptr_monarch = (ia64_fptr_t *)ia64_os_init_dispatch_monarch;
1892 	ia64_fptr_t *init_hldlr_ptr_slave = (ia64_fptr_t *)ia64_os_init_dispatch_slave;
1893 	ia64_fptr_t *mca_hldlr_ptr = (ia64_fptr_t *)ia64_os_mca_dispatch;
1894 	int i;
1895 	long rc;
1896 	struct ia64_sal_retval isrv;
1897 	unsigned long timeout = IA64_MCA_RENDEZ_TIMEOUT; /* platform specific */
1898 	static struct notifier_block default_init_monarch_nb = {
1899 		.notifier_call = default_monarch_init_process,
1900 		.priority = 0/* we need to notified last */
1901 	};
1902 
1903 	IA64_MCA_DEBUG("%s: begin\n", __func__);
1904 
1905 	/* Clear the Rendez checkin flag for all cpus */
1906 	for(i = 0 ; i < NR_CPUS; i++)
1907 		ia64_mc_info.imi_rendez_checkin[i] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1908 
1909 	/*
1910 	 * Register the rendezvous spinloop and wakeup mechanism with SAL
1911 	 */
1912 
1913 	/* Register the rendezvous interrupt vector with SAL */
1914 	while (1) {
1915 		isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_RENDEZ_INT,
1916 					      SAL_MC_PARAM_MECHANISM_INT,
1917 					      IA64_MCA_RENDEZ_VECTOR,
1918 					      timeout,
1919 					      SAL_MC_PARAM_RZ_ALWAYS);
1920 		rc = isrv.status;
1921 		if (rc == 0)
1922 			break;
1923 		if (rc == -2) {
1924 			printk(KERN_INFO "Increasing MCA rendezvous timeout from "
1925 				"%ld to %ld milliseconds\n", timeout, isrv.v0);
1926 			timeout = isrv.v0;
1927 			NOTIFY_MCA(DIE_MCA_NEW_TIMEOUT, NULL, timeout, 0);
1928 			continue;
1929 		}
1930 		printk(KERN_ERR "Failed to register rendezvous interrupt "
1931 		       "with SAL (status %ld)\n", rc);
1932 		return;
1933 	}
1934 
1935 	/* Register the wakeup interrupt vector with SAL */
1936 	isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_RENDEZ_WAKEUP,
1937 				      SAL_MC_PARAM_MECHANISM_INT,
1938 				      IA64_MCA_WAKEUP_VECTOR,
1939 				      0, 0);
1940 	rc = isrv.status;
1941 	if (rc) {
1942 		printk(KERN_ERR "Failed to register wakeup interrupt with SAL "
1943 		       "(status %ld)\n", rc);
1944 		return;
1945 	}
1946 
1947 	IA64_MCA_DEBUG("%s: registered MCA rendezvous spinloop and wakeup mech.\n", __func__);
1948 
1949 	ia64_mc_info.imi_mca_handler        = ia64_tpa(mca_hldlr_ptr->fp);
1950 	/*
1951 	 * XXX - disable SAL checksum by setting size to 0; should be
1952 	 *	ia64_tpa(ia64_os_mca_dispatch_end) - ia64_tpa(ia64_os_mca_dispatch);
1953 	 */
1954 	ia64_mc_info.imi_mca_handler_size	= 0;
1955 
1956 	/* Register the os mca handler with SAL */
1957 	if ((rc = ia64_sal_set_vectors(SAL_VECTOR_OS_MCA,
1958 				       ia64_mc_info.imi_mca_handler,
1959 				       ia64_tpa(mca_hldlr_ptr->gp),
1960 				       ia64_mc_info.imi_mca_handler_size,
1961 				       0, 0, 0)))
1962 	{
1963 		printk(KERN_ERR "Failed to register OS MCA handler with SAL "
1964 		       "(status %ld)\n", rc);
1965 		return;
1966 	}
1967 
1968 	IA64_MCA_DEBUG("%s: registered OS MCA handler with SAL at 0x%lx, gp = 0x%lx\n", __func__,
1969 		       ia64_mc_info.imi_mca_handler, ia64_tpa(mca_hldlr_ptr->gp));
1970 
1971 	/*
1972 	 * XXX - disable SAL checksum by setting size to 0, should be
1973 	 * size of the actual init handler in mca_asm.S.
1974 	 */
1975 	ia64_mc_info.imi_monarch_init_handler		= ia64_tpa(init_hldlr_ptr_monarch->fp);
1976 	ia64_mc_info.imi_monarch_init_handler_size	= 0;
1977 	ia64_mc_info.imi_slave_init_handler		= ia64_tpa(init_hldlr_ptr_slave->fp);
1978 	ia64_mc_info.imi_slave_init_handler_size	= 0;
1979 
1980 	IA64_MCA_DEBUG("%s: OS INIT handler at %lx\n", __func__,
1981 		       ia64_mc_info.imi_monarch_init_handler);
1982 
1983 	/* Register the os init handler with SAL */
1984 	if ((rc = ia64_sal_set_vectors(SAL_VECTOR_OS_INIT,
1985 				       ia64_mc_info.imi_monarch_init_handler,
1986 				       ia64_tpa(ia64_getreg(_IA64_REG_GP)),
1987 				       ia64_mc_info.imi_monarch_init_handler_size,
1988 				       ia64_mc_info.imi_slave_init_handler,
1989 				       ia64_tpa(ia64_getreg(_IA64_REG_GP)),
1990 				       ia64_mc_info.imi_slave_init_handler_size)))
1991 	{
1992 		printk(KERN_ERR "Failed to register m/s INIT handlers with SAL "
1993 		       "(status %ld)\n", rc);
1994 		return;
1995 	}
1996 	if (register_die_notifier(&default_init_monarch_nb)) {
1997 		printk(KERN_ERR "Failed to register default monarch INIT process\n");
1998 		return;
1999 	}
2000 
2001 	IA64_MCA_DEBUG("%s: registered OS INIT handler with SAL\n", __func__);
2002 
2003 	/* Initialize the areas set aside by the OS to buffer the
2004 	 * platform/processor error states for MCA/INIT/CMC
2005 	 * handling.
2006 	 */
2007 	ia64_log_init(SAL_INFO_TYPE_MCA);
2008 	ia64_log_init(SAL_INFO_TYPE_INIT);
2009 	ia64_log_init(SAL_INFO_TYPE_CMC);
2010 	ia64_log_init(SAL_INFO_TYPE_CPE);
2011 
2012 	mca_init = 1;
2013 	printk(KERN_INFO "MCA related initialization done\n");
2014 }
2015 
2016 
2017 /*
2018  * These pieces cannot be done in ia64_mca_init() because it is called before
2019  * early_irq_init() which would wipe out our percpu irq registrations. But we
2020  * cannot leave them until ia64_mca_late_init() because by then all the other
2021  * processors have been brought online and have set their own CMC vectors to
2022  * point at a non-existant action. Called from arch_early_irq_init().
2023  */
2024 void __init ia64_mca_irq_init(void)
2025 {
2026 	/*
2027 	 *  Configure the CMCI/P vector and handler. Interrupts for CMC are
2028 	 *  per-processor, so AP CMC interrupts are setup in smp_callin() (smpboot.c).
2029 	 */
2030 	register_percpu_irq(IA64_CMC_VECTOR, ia64_mca_cmc_int_handler, 0,
2031 			    "cmc_hndlr");
2032 	register_percpu_irq(IA64_CMCP_VECTOR, ia64_mca_cmc_int_caller, 0,
2033 			    "cmc_poll");
2034 	ia64_mca_cmc_vector_setup();       /* Setup vector on BSP */
2035 
2036 	/* Setup the MCA rendezvous interrupt vector */
2037 	register_percpu_irq(IA64_MCA_RENDEZ_VECTOR, ia64_mca_rendez_int_handler,
2038 			    0, "mca_rdzv");
2039 
2040 	/* Setup the MCA wakeup interrupt vector */
2041 	register_percpu_irq(IA64_MCA_WAKEUP_VECTOR, ia64_mca_wakeup_int_handler,
2042 			    0, "mca_wkup");
2043 
2044 	/* Setup the CPEI/P handler */
2045 	register_percpu_irq(IA64_CPEP_VECTOR, ia64_mca_cpe_int_caller, 0,
2046 			    "cpe_poll");
2047 }
2048 
2049 /*
2050  * ia64_mca_late_init
2051  *
2052  *	Opportunity to setup things that require initialization later
2053  *	than ia64_mca_init.  Setup a timer to poll for CPEs if the
2054  *	platform doesn't support an interrupt driven mechanism.
2055  *
2056  *  Inputs  :   None
2057  *  Outputs :   Status
2058  */
2059 static int __init
2060 ia64_mca_late_init(void)
2061 {
2062 	if (!mca_init)
2063 		return 0;
2064 
2065 	/* Setup the CMCI/P vector and handler */
2066 	timer_setup(&cmc_poll_timer, ia64_mca_cmc_poll, 0);
2067 
2068 	/* Unmask/enable the vector */
2069 	cmc_polling_enabled = 0;
2070 	cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "ia64/mca:online",
2071 			  ia64_mca_cpu_online, NULL);
2072 	IA64_MCA_DEBUG("%s: CMCI/P setup and enabled.\n", __func__);
2073 
2074 	/* Setup the CPEI/P vector and handler */
2075 	cpe_vector = acpi_request_vector(ACPI_INTERRUPT_CPEI);
2076 	timer_setup(&cpe_poll_timer, ia64_mca_cpe_poll, 0);
2077 
2078 	{
2079 		unsigned int irq;
2080 
2081 		if (cpe_vector >= 0) {
2082 			/* If platform supports CPEI, enable the irq. */
2083 			irq = local_vector_to_irq(cpe_vector);
2084 			if (irq > 0) {
2085 				cpe_poll_enabled = 0;
2086 				irq_set_status_flags(irq, IRQ_PER_CPU);
2087 				if (request_irq(irq, ia64_mca_cpe_int_handler,
2088 						0, "cpe_hndlr", NULL))
2089 					pr_err("Failed to register cpe_hndlr interrupt\n");
2090 				ia64_cpe_irq = irq;
2091 				ia64_mca_register_cpev(cpe_vector);
2092 				IA64_MCA_DEBUG("%s: CPEI/P setup and enabled.\n",
2093 					__func__);
2094 				return 0;
2095 			}
2096 			printk(KERN_ERR "%s: Failed to find irq for CPE "
2097 					"interrupt handler, vector %d\n",
2098 					__func__, cpe_vector);
2099 		}
2100 		/* If platform doesn't support CPEI, get the timer going. */
2101 		if (cpe_poll_enabled) {
2102 			ia64_mca_cpe_poll(0UL);
2103 			IA64_MCA_DEBUG("%s: CPEP setup and enabled.\n", __func__);
2104 		}
2105 	}
2106 
2107 	return 0;
2108 }
2109 
2110 device_initcall(ia64_mca_late_init);
2111