xref: /openbmc/linux/arch/ia64/kernel/mca.c (revision 5b628549)
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/signal.h>
76 #include <linux/sched/debug.h>
77 #include <linux/sched/task.h>
78 #include <linux/interrupt.h>
79 #include <linux/irq.h>
80 #include <linux/memblock.h>
81 #include <linux/acpi.h>
82 #include <linux/timer.h>
83 #include <linux/module.h>
84 #include <linux/kernel.h>
85 #include <linux/smp.h>
86 #include <linux/workqueue.h>
87 #include <linux/cpumask.h>
88 #include <linux/kdebug.h>
89 #include <linux/cpu.h>
90 #include <linux/gfp.h>
91 
92 #include <asm/delay.h>
93 #include <asm/machvec.h>
94 #include <asm/meminit.h>
95 #include <asm/page.h>
96 #include <asm/ptrace.h>
97 #include <asm/sal.h>
98 #include <asm/mca.h>
99 #include <asm/kexec.h>
100 
101 #include <asm/irq.h>
102 #include <asm/hw_irq.h>
103 #include <asm/tlb.h>
104 
105 #include "mca_drv.h"
106 #include "entry.h"
107 
108 #if defined(IA64_MCA_DEBUG_INFO)
109 # define IA64_MCA_DEBUG(fmt...)	printk(fmt)
110 #else
111 # define IA64_MCA_DEBUG(fmt...)
112 #endif
113 
114 #define NOTIFY_INIT(event, regs, arg, spin)				\
115 do {									\
116 	if ((notify_die((event), "INIT", (regs), (arg), 0, 0)		\
117 			== NOTIFY_STOP) && ((spin) == 1))		\
118 		ia64_mca_spin(__func__);				\
119 } while (0)
120 
121 #define NOTIFY_MCA(event, regs, arg, spin)				\
122 do {									\
123 	if ((notify_die((event), "MCA", (regs), (arg), 0, 0)		\
124 			== NOTIFY_STOP) && ((spin) == 1))		\
125 		ia64_mca_spin(__func__);				\
126 } while (0)
127 
128 /* Used by mca_asm.S */
129 DEFINE_PER_CPU(u64, ia64_mca_data); /* == __per_cpu_mca[smp_processor_id()] */
130 DEFINE_PER_CPU(u64, ia64_mca_per_cpu_pte); /* PTE to map per-CPU area */
131 DEFINE_PER_CPU(u64, ia64_mca_pal_pte);	    /* PTE to map PAL code */
132 DEFINE_PER_CPU(u64, ia64_mca_pal_base);    /* vaddr PAL code granule */
133 DEFINE_PER_CPU(u64, ia64_mca_tr_reload);   /* Flag for TR reload */
134 
135 unsigned long __per_cpu_mca[NR_CPUS];
136 
137 /* In mca_asm.S */
138 extern void			ia64_os_init_dispatch_monarch (void);
139 extern void			ia64_os_init_dispatch_slave (void);
140 
141 static int monarch_cpu = -1;
142 
143 static ia64_mc_info_t		ia64_mc_info;
144 
145 #define MAX_CPE_POLL_INTERVAL (15*60*HZ) /* 15 minutes */
146 #define MIN_CPE_POLL_INTERVAL (2*60*HZ)  /* 2 minutes */
147 #define CMC_POLL_INTERVAL     (1*60*HZ)  /* 1 minute */
148 #define CPE_HISTORY_LENGTH    5
149 #define CMC_HISTORY_LENGTH    5
150 
151 #ifdef CONFIG_ACPI
152 static struct timer_list cpe_poll_timer;
153 #endif
154 static struct timer_list cmc_poll_timer;
155 /*
156  * This variable tells whether we are currently in polling mode.
157  * Start with this in the wrong state so we won't play w/ timers
158  * before the system is ready.
159  */
160 static int cmc_polling_enabled = 1;
161 
162 /*
163  * Clearing this variable prevents CPE polling from getting activated
164  * in mca_late_init.  Use it if your system doesn't provide a CPEI,
165  * but encounters problems retrieving CPE logs.  This should only be
166  * necessary for debugging.
167  */
168 static int cpe_poll_enabled = 1;
169 
170 extern void salinfo_log_wakeup(int type, u8 *buffer, u64 size, int irqsafe);
171 
172 static int mca_init __initdata;
173 
174 /*
175  * limited & delayed printing support for MCA/INIT handler
176  */
177 
178 #define mprintk(fmt...) ia64_mca_printk(fmt)
179 
180 #define MLOGBUF_SIZE (512+256*NR_CPUS)
181 #define MLOGBUF_MSGMAX 256
182 static char mlogbuf[MLOGBUF_SIZE];
183 static DEFINE_SPINLOCK(mlogbuf_wlock);	/* mca context only */
184 static DEFINE_SPINLOCK(mlogbuf_rlock);	/* normal context only */
185 static unsigned long mlogbuf_start;
186 static unsigned long mlogbuf_end;
187 static unsigned int mlogbuf_finished = 0;
188 static unsigned long mlogbuf_timestamp = 0;
189 
190 static int loglevel_save = -1;
191 #define BREAK_LOGLEVEL(__console_loglevel)		\
192 	oops_in_progress = 1;				\
193 	if (loglevel_save < 0)				\
194 		loglevel_save = __console_loglevel;	\
195 	__console_loglevel = 15;
196 
197 #define RESTORE_LOGLEVEL(__console_loglevel)		\
198 	if (loglevel_save >= 0) {			\
199 		__console_loglevel = loglevel_save;	\
200 		loglevel_save = -1;			\
201 	}						\
202 	mlogbuf_finished = 0;				\
203 	oops_in_progress = 0;
204 
205 /*
206  * Push messages into buffer, print them later if not urgent.
207  */
208 void ia64_mca_printk(const char *fmt, ...)
209 {
210 	va_list args;
211 	int printed_len;
212 	char temp_buf[MLOGBUF_MSGMAX];
213 	char *p;
214 
215 	va_start(args, fmt);
216 	printed_len = vscnprintf(temp_buf, sizeof(temp_buf), fmt, args);
217 	va_end(args);
218 
219 	/* Copy the output into mlogbuf */
220 	if (oops_in_progress) {
221 		/* mlogbuf was abandoned, use printk directly instead. */
222 		printk("%s", temp_buf);
223 	} else {
224 		spin_lock(&mlogbuf_wlock);
225 		for (p = temp_buf; *p; p++) {
226 			unsigned long next = (mlogbuf_end + 1) % MLOGBUF_SIZE;
227 			if (next != mlogbuf_start) {
228 				mlogbuf[mlogbuf_end] = *p;
229 				mlogbuf_end = next;
230 			} else {
231 				/* buffer full */
232 				break;
233 			}
234 		}
235 		mlogbuf[mlogbuf_end] = '\0';
236 		spin_unlock(&mlogbuf_wlock);
237 	}
238 }
239 EXPORT_SYMBOL(ia64_mca_printk);
240 
241 /*
242  * Print buffered messages.
243  *  NOTE: call this after returning normal context. (ex. from salinfod)
244  */
245 void ia64_mlogbuf_dump(void)
246 {
247 	char temp_buf[MLOGBUF_MSGMAX];
248 	char *p;
249 	unsigned long index;
250 	unsigned long flags;
251 	unsigned int printed_len;
252 
253 	/* Get output from mlogbuf */
254 	while (mlogbuf_start != mlogbuf_end) {
255 		temp_buf[0] = '\0';
256 		p = temp_buf;
257 		printed_len = 0;
258 
259 		spin_lock_irqsave(&mlogbuf_rlock, flags);
260 
261 		index = mlogbuf_start;
262 		while (index != mlogbuf_end) {
263 			*p = mlogbuf[index];
264 			index = (index + 1) % MLOGBUF_SIZE;
265 			if (!*p)
266 				break;
267 			p++;
268 			if (++printed_len >= MLOGBUF_MSGMAX - 1)
269 				break;
270 		}
271 		*p = '\0';
272 		if (temp_buf[0])
273 			printk("%s", temp_buf);
274 		mlogbuf_start = index;
275 
276 		mlogbuf_timestamp = 0;
277 		spin_unlock_irqrestore(&mlogbuf_rlock, flags);
278 	}
279 }
280 EXPORT_SYMBOL(ia64_mlogbuf_dump);
281 
282 /*
283  * Call this if system is going to down or if immediate flushing messages to
284  * console is required. (ex. recovery was failed, crash dump is going to be
285  * invoked, long-wait rendezvous etc.)
286  *  NOTE: this should be called from monarch.
287  */
288 static void ia64_mlogbuf_finish(int wait)
289 {
290 	BREAK_LOGLEVEL(console_loglevel);
291 
292 	spin_lock_init(&mlogbuf_rlock);
293 	ia64_mlogbuf_dump();
294 	printk(KERN_EMERG "mlogbuf_finish: printing switched to urgent mode, "
295 		"MCA/INIT might be dodgy or fail.\n");
296 
297 	if (!wait)
298 		return;
299 
300 	/* wait for console */
301 	printk("Delaying for 5 seconds...\n");
302 	udelay(5*1000000);
303 
304 	mlogbuf_finished = 1;
305 }
306 
307 /*
308  * Print buffered messages from INIT context.
309  */
310 static void ia64_mlogbuf_dump_from_init(void)
311 {
312 	if (mlogbuf_finished)
313 		return;
314 
315 	if (mlogbuf_timestamp &&
316 			time_before(jiffies, mlogbuf_timestamp + 30 * HZ)) {
317 		printk(KERN_ERR "INIT: mlogbuf_dump is interrupted by INIT "
318 			" and the system seems to be messed up.\n");
319 		ia64_mlogbuf_finish(0);
320 		return;
321 	}
322 
323 	if (!spin_trylock(&mlogbuf_rlock)) {
324 		printk(KERN_ERR "INIT: mlogbuf_dump is interrupted by INIT. "
325 			"Generated messages other than stack dump will be "
326 			"buffered to mlogbuf and will be printed later.\n");
327 		printk(KERN_ERR "INIT: If messages would not printed after "
328 			"this INIT, wait 30sec and assert INIT again.\n");
329 		if (!mlogbuf_timestamp)
330 			mlogbuf_timestamp = jiffies;
331 		return;
332 	}
333 	spin_unlock(&mlogbuf_rlock);
334 	ia64_mlogbuf_dump();
335 }
336 
337 static inline void
338 ia64_mca_spin(const char *func)
339 {
340 	if (monarch_cpu == smp_processor_id())
341 		ia64_mlogbuf_finish(0);
342 	mprintk(KERN_EMERG "%s: spinning here, not returning to SAL\n", func);
343 	while (1)
344 		cpu_relax();
345 }
346 /*
347  * IA64_MCA log support
348  */
349 #define IA64_MAX_LOGS		2	/* Double-buffering for nested MCAs */
350 #define IA64_MAX_LOG_TYPES      4   /* MCA, INIT, CMC, CPE */
351 
352 typedef struct ia64_state_log_s
353 {
354 	spinlock_t	isl_lock;
355 	int		isl_index;
356 	unsigned long	isl_count;
357 	ia64_err_rec_t  *isl_log[IA64_MAX_LOGS]; /* need space to store header + error log */
358 } ia64_state_log_t;
359 
360 static ia64_state_log_t ia64_state_log[IA64_MAX_LOG_TYPES];
361 
362 #define IA64_LOG_LOCK_INIT(it) spin_lock_init(&ia64_state_log[it].isl_lock)
363 #define IA64_LOG_LOCK(it)      spin_lock_irqsave(&ia64_state_log[it].isl_lock, s)
364 #define IA64_LOG_UNLOCK(it)    spin_unlock_irqrestore(&ia64_state_log[it].isl_lock,s)
365 #define IA64_LOG_NEXT_INDEX(it)    ia64_state_log[it].isl_index
366 #define IA64_LOG_CURR_INDEX(it)    1 - ia64_state_log[it].isl_index
367 #define IA64_LOG_INDEX_INC(it) \
368     {ia64_state_log[it].isl_index = 1 - ia64_state_log[it].isl_index; \
369     ia64_state_log[it].isl_count++;}
370 #define IA64_LOG_INDEX_DEC(it) \
371     ia64_state_log[it].isl_index = 1 - ia64_state_log[it].isl_index
372 #define IA64_LOG_NEXT_BUFFER(it)   (void *)((ia64_state_log[it].isl_log[IA64_LOG_NEXT_INDEX(it)]))
373 #define IA64_LOG_CURR_BUFFER(it)   (void *)((ia64_state_log[it].isl_log[IA64_LOG_CURR_INDEX(it)]))
374 #define IA64_LOG_COUNT(it)         ia64_state_log[it].isl_count
375 
376 static inline void ia64_log_allocate(int it, u64 size)
377 {
378 	ia64_state_log[it].isl_log[IA64_LOG_CURR_INDEX(it)] =
379 		(ia64_err_rec_t *)memblock_alloc(size, SMP_CACHE_BYTES);
380 	if (!ia64_state_log[it].isl_log[IA64_LOG_CURR_INDEX(it)])
381 		panic("%s: Failed to allocate %llu bytes\n", __func__, size);
382 
383 	ia64_state_log[it].isl_log[IA64_LOG_NEXT_INDEX(it)] =
384 		(ia64_err_rec_t *)memblock_alloc(size, SMP_CACHE_BYTES);
385 	if (!ia64_state_log[it].isl_log[IA64_LOG_NEXT_INDEX(it)])
386 		panic("%s: Failed to allocate %llu bytes\n", __func__, size);
387 }
388 
389 /*
390  * ia64_log_init
391  *	Reset the OS ia64 log buffer
392  * Inputs   :   info_type   (SAL_INFO_TYPE_{MCA,INIT,CMC,CPE})
393  * Outputs	:	None
394  */
395 static void __init
396 ia64_log_init(int sal_info_type)
397 {
398 	u64	max_size = 0;
399 
400 	IA64_LOG_NEXT_INDEX(sal_info_type) = 0;
401 	IA64_LOG_LOCK_INIT(sal_info_type);
402 
403 	// SAL will tell us the maximum size of any error record of this type
404 	max_size = ia64_sal_get_state_info_size(sal_info_type);
405 	if (!max_size)
406 		/* alloc_bootmem() doesn't like zero-sized allocations! */
407 		return;
408 
409 	// set up OS data structures to hold error info
410 	ia64_log_allocate(sal_info_type, max_size);
411 }
412 
413 /*
414  * ia64_log_get
415  *
416  *	Get the current MCA log from SAL and copy it into the OS log buffer.
417  *
418  *  Inputs  :   info_type   (SAL_INFO_TYPE_{MCA,INIT,CMC,CPE})
419  *              irq_safe    whether you can use printk at this point
420  *  Outputs :   size        (total record length)
421  *              *buffer     (ptr to error record)
422  *
423  */
424 static u64
425 ia64_log_get(int sal_info_type, u8 **buffer, int irq_safe)
426 {
427 	sal_log_record_header_t     *log_buffer;
428 	u64                         total_len = 0;
429 	unsigned long               s;
430 
431 	IA64_LOG_LOCK(sal_info_type);
432 
433 	/* Get the process state information */
434 	log_buffer = IA64_LOG_NEXT_BUFFER(sal_info_type);
435 
436 	total_len = ia64_sal_get_state_info(sal_info_type, (u64 *)log_buffer);
437 
438 	if (total_len) {
439 		IA64_LOG_INDEX_INC(sal_info_type);
440 		IA64_LOG_UNLOCK(sal_info_type);
441 		if (irq_safe) {
442 			IA64_MCA_DEBUG("%s: SAL error record type %d retrieved. Record length = %ld\n",
443 				       __func__, sal_info_type, total_len);
444 		}
445 		*buffer = (u8 *) log_buffer;
446 		return total_len;
447 	} else {
448 		IA64_LOG_UNLOCK(sal_info_type);
449 		return 0;
450 	}
451 }
452 
453 /*
454  *  ia64_mca_log_sal_error_record
455  *
456  *  This function retrieves a specified error record type from SAL
457  *  and wakes up any processes waiting for error records.
458  *
459  *  Inputs  :   sal_info_type   (Type of error record MCA/CMC/CPE)
460  *              FIXME: remove MCA and irq_safe.
461  */
462 static void
463 ia64_mca_log_sal_error_record(int sal_info_type)
464 {
465 	u8 *buffer;
466 	sal_log_record_header_t *rh;
467 	u64 size;
468 	int irq_safe = sal_info_type != SAL_INFO_TYPE_MCA;
469 #ifdef IA64_MCA_DEBUG_INFO
470 	static const char * const rec_name[] = { "MCA", "INIT", "CMC", "CPE" };
471 #endif
472 
473 	size = ia64_log_get(sal_info_type, &buffer, irq_safe);
474 	if (!size)
475 		return;
476 
477 	salinfo_log_wakeup(sal_info_type, buffer, size, irq_safe);
478 
479 	if (irq_safe)
480 		IA64_MCA_DEBUG("CPU %d: SAL log contains %s error record\n",
481 			smp_processor_id(),
482 			sal_info_type < ARRAY_SIZE(rec_name) ? rec_name[sal_info_type] : "UNKNOWN");
483 
484 	/* Clear logs from corrected errors in case there's no user-level logger */
485 	rh = (sal_log_record_header_t *)buffer;
486 	if (rh->severity == sal_log_severity_corrected)
487 		ia64_sal_clear_state_info(sal_info_type);
488 }
489 
490 /*
491  * search_mca_table
492  *  See if the MCA surfaced in an instruction range
493  *  that has been tagged as recoverable.
494  *
495  *  Inputs
496  *	first	First address range to check
497  *	last	Last address range to check
498  *	ip	Instruction pointer, address we are looking for
499  *
500  * Return value:
501  *      1 on Success (in the table)/ 0 on Failure (not in the  table)
502  */
503 int
504 search_mca_table (const struct mca_table_entry *first,
505                 const struct mca_table_entry *last,
506                 unsigned long ip)
507 {
508         const struct mca_table_entry *curr;
509         u64 curr_start, curr_end;
510 
511         curr = first;
512         while (curr <= last) {
513                 curr_start = (u64) &curr->start_addr + curr->start_addr;
514                 curr_end = (u64) &curr->end_addr + curr->end_addr;
515 
516                 if ((ip >= curr_start) && (ip <= curr_end)) {
517                         return 1;
518                 }
519                 curr++;
520         }
521         return 0;
522 }
523 
524 /* Given an address, look for it in the mca tables. */
525 int mca_recover_range(unsigned long addr)
526 {
527 	extern struct mca_table_entry __start___mca_table[];
528 	extern struct mca_table_entry __stop___mca_table[];
529 
530 	return search_mca_table(__start___mca_table, __stop___mca_table-1, addr);
531 }
532 EXPORT_SYMBOL_GPL(mca_recover_range);
533 
534 #ifdef CONFIG_ACPI
535 
536 int cpe_vector = -1;
537 int ia64_cpe_irq = -1;
538 
539 static irqreturn_t
540 ia64_mca_cpe_int_handler (int cpe_irq, void *arg)
541 {
542 	static unsigned long	cpe_history[CPE_HISTORY_LENGTH];
543 	static int		index;
544 	static DEFINE_SPINLOCK(cpe_history_lock);
545 
546 	IA64_MCA_DEBUG("%s: received interrupt vector = %#x on CPU %d\n",
547 		       __func__, cpe_irq, smp_processor_id());
548 
549 	/* SAL spec states this should run w/ interrupts enabled */
550 	local_irq_enable();
551 
552 	spin_lock(&cpe_history_lock);
553 	if (!cpe_poll_enabled && cpe_vector >= 0) {
554 
555 		int i, count = 1; /* we know 1 happened now */
556 		unsigned long now = jiffies;
557 
558 		for (i = 0; i < CPE_HISTORY_LENGTH; i++) {
559 			if (now - cpe_history[i] <= HZ)
560 				count++;
561 		}
562 
563 		IA64_MCA_DEBUG(KERN_INFO "CPE threshold %d/%d\n", count, CPE_HISTORY_LENGTH);
564 		if (count >= CPE_HISTORY_LENGTH) {
565 
566 			cpe_poll_enabled = 1;
567 			spin_unlock(&cpe_history_lock);
568 			disable_irq_nosync(local_vector_to_irq(IA64_CPE_VECTOR));
569 
570 			/*
571 			 * Corrected errors will still be corrected, but
572 			 * make sure there's a log somewhere that indicates
573 			 * something is generating more than we can handle.
574 			 */
575 			printk(KERN_WARNING "WARNING: Switching to polling CPE handler; error records may be lost\n");
576 
577 			mod_timer(&cpe_poll_timer, jiffies + MIN_CPE_POLL_INTERVAL);
578 
579 			/* lock already released, get out now */
580 			goto out;
581 		} else {
582 			cpe_history[index++] = now;
583 			if (index == CPE_HISTORY_LENGTH)
584 				index = 0;
585 		}
586 	}
587 	spin_unlock(&cpe_history_lock);
588 out:
589 	/* Get the CPE error record and log it */
590 	ia64_mca_log_sal_error_record(SAL_INFO_TYPE_CPE);
591 
592 	local_irq_disable();
593 
594 	return IRQ_HANDLED;
595 }
596 
597 #endif /* CONFIG_ACPI */
598 
599 #ifdef CONFIG_ACPI
600 /*
601  * ia64_mca_register_cpev
602  *
603  *  Register the corrected platform error vector with SAL.
604  *
605  *  Inputs
606  *      cpev        Corrected Platform Error Vector number
607  *
608  *  Outputs
609  *      None
610  */
611 void
612 ia64_mca_register_cpev (int cpev)
613 {
614 	/* Register the CPE interrupt vector with SAL */
615 	struct ia64_sal_retval isrv;
616 
617 	isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_CPE_INT, SAL_MC_PARAM_MECHANISM_INT, cpev, 0, 0);
618 	if (isrv.status) {
619 		printk(KERN_ERR "Failed to register Corrected Platform "
620 		       "Error interrupt vector with SAL (status %ld)\n", isrv.status);
621 		return;
622 	}
623 
624 	IA64_MCA_DEBUG("%s: corrected platform error "
625 		       "vector %#x registered\n", __func__, cpev);
626 }
627 #endif /* CONFIG_ACPI */
628 
629 /*
630  * ia64_mca_cmc_vector_setup
631  *
632  *  Setup the corrected machine check vector register in the processor.
633  *  (The interrupt is masked on boot. ia64_mca_late_init unmask this.)
634  *  This function is invoked on a per-processor basis.
635  *
636  * Inputs
637  *      None
638  *
639  * Outputs
640  *	None
641  */
642 void
643 ia64_mca_cmc_vector_setup (void)
644 {
645 	cmcv_reg_t	cmcv;
646 
647 	cmcv.cmcv_regval	= 0;
648 	cmcv.cmcv_mask		= 1;        /* Mask/disable interrupt at first */
649 	cmcv.cmcv_vector	= IA64_CMC_VECTOR;
650 	ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
651 
652 	IA64_MCA_DEBUG("%s: CPU %d corrected machine check vector %#x registered.\n",
653 		       __func__, smp_processor_id(), IA64_CMC_VECTOR);
654 
655 	IA64_MCA_DEBUG("%s: CPU %d CMCV = %#016lx\n",
656 		       __func__, smp_processor_id(), ia64_getreg(_IA64_REG_CR_CMCV));
657 }
658 
659 /*
660  * ia64_mca_cmc_vector_disable
661  *
662  *  Mask the corrected machine check vector register in the processor.
663  *  This function is invoked on a per-processor basis.
664  *
665  * Inputs
666  *      dummy(unused)
667  *
668  * Outputs
669  *	None
670  */
671 static void
672 ia64_mca_cmc_vector_disable (void *dummy)
673 {
674 	cmcv_reg_t	cmcv;
675 
676 	cmcv.cmcv_regval = ia64_getreg(_IA64_REG_CR_CMCV);
677 
678 	cmcv.cmcv_mask = 1; /* Mask/disable interrupt */
679 	ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
680 
681 	IA64_MCA_DEBUG("%s: CPU %d corrected machine check vector %#x disabled.\n",
682 		       __func__, smp_processor_id(), cmcv.cmcv_vector);
683 }
684 
685 /*
686  * ia64_mca_cmc_vector_enable
687  *
688  *  Unmask the corrected machine check vector register in the processor.
689  *  This function is invoked on a per-processor basis.
690  *
691  * Inputs
692  *      dummy(unused)
693  *
694  * Outputs
695  *	None
696  */
697 static void
698 ia64_mca_cmc_vector_enable (void *dummy)
699 {
700 	cmcv_reg_t	cmcv;
701 
702 	cmcv.cmcv_regval = ia64_getreg(_IA64_REG_CR_CMCV);
703 
704 	cmcv.cmcv_mask = 0; /* Unmask/enable interrupt */
705 	ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
706 
707 	IA64_MCA_DEBUG("%s: CPU %d corrected machine check vector %#x enabled.\n",
708 		       __func__, smp_processor_id(), cmcv.cmcv_vector);
709 }
710 
711 /*
712  * ia64_mca_cmc_vector_disable_keventd
713  *
714  * Called via keventd (smp_call_function() is not safe in interrupt context) to
715  * disable the cmc interrupt vector.
716  */
717 static void
718 ia64_mca_cmc_vector_disable_keventd(struct work_struct *unused)
719 {
720 	on_each_cpu(ia64_mca_cmc_vector_disable, NULL, 0);
721 }
722 
723 /*
724  * ia64_mca_cmc_vector_enable_keventd
725  *
726  * Called via keventd (smp_call_function() is not safe in interrupt context) to
727  * enable the cmc interrupt vector.
728  */
729 static void
730 ia64_mca_cmc_vector_enable_keventd(struct work_struct *unused)
731 {
732 	on_each_cpu(ia64_mca_cmc_vector_enable, NULL, 0);
733 }
734 
735 /*
736  * ia64_mca_wakeup
737  *
738  *	Send an inter-cpu interrupt to wake-up a particular cpu.
739  *
740  *  Inputs  :   cpuid
741  *  Outputs :   None
742  */
743 static void
744 ia64_mca_wakeup(int cpu)
745 {
746 	platform_send_ipi(cpu, IA64_MCA_WAKEUP_VECTOR, IA64_IPI_DM_INT, 0);
747 }
748 
749 /*
750  * ia64_mca_wakeup_all
751  *
752  *	Wakeup all the slave cpus which have rendez'ed previously.
753  *
754  *  Inputs  :   None
755  *  Outputs :   None
756  */
757 static void
758 ia64_mca_wakeup_all(void)
759 {
760 	int cpu;
761 
762 	/* Clear the Rendez checkin flag for all cpus */
763 	for_each_online_cpu(cpu) {
764 		if (ia64_mc_info.imi_rendez_checkin[cpu] == IA64_MCA_RENDEZ_CHECKIN_DONE)
765 			ia64_mca_wakeup(cpu);
766 	}
767 
768 }
769 
770 /*
771  * ia64_mca_rendez_interrupt_handler
772  *
773  *	This is handler used to put slave processors into spinloop
774  *	while the monarch processor does the mca handling and later
775  *	wake each slave up once the monarch is done.  The state
776  *	IA64_MCA_RENDEZ_CHECKIN_DONE indicates the cpu is rendez'ed
777  *	in SAL.  The state IA64_MCA_RENDEZ_CHECKIN_NOTDONE indicates
778  *	the cpu has come out of OS rendezvous.
779  *
780  *  Inputs  :   None
781  *  Outputs :   None
782  */
783 static irqreturn_t
784 ia64_mca_rendez_int_handler(int rendez_irq, void *arg)
785 {
786 	unsigned long flags;
787 	int cpu = smp_processor_id();
788 	struct ia64_mca_notify_die nd =
789 		{ .sos = NULL, .monarch_cpu = &monarch_cpu };
790 
791 	/* Mask all interrupts */
792 	local_irq_save(flags);
793 
794 	NOTIFY_MCA(DIE_MCA_RENDZVOUS_ENTER, get_irq_regs(), (long)&nd, 1);
795 
796 	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_DONE;
797 	/* Register with the SAL monarch that the slave has
798 	 * reached SAL
799 	 */
800 	ia64_sal_mc_rendez();
801 
802 	NOTIFY_MCA(DIE_MCA_RENDZVOUS_PROCESS, get_irq_regs(), (long)&nd, 1);
803 
804 	/* Wait for the monarch cpu to exit. */
805 	while (monarch_cpu != -1)
806 	       cpu_relax();	/* spin until monarch leaves */
807 
808 	NOTIFY_MCA(DIE_MCA_RENDZVOUS_LEAVE, get_irq_regs(), (long)&nd, 1);
809 
810 	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
811 	/* Enable all interrupts */
812 	local_irq_restore(flags);
813 	return IRQ_HANDLED;
814 }
815 
816 /*
817  * ia64_mca_wakeup_int_handler
818  *
819  *	The interrupt handler for processing the inter-cpu interrupt to the
820  *	slave cpu which was spinning in the rendez loop.
821  *	Since this spinning is done by turning off the interrupts and
822  *	polling on the wakeup-interrupt bit in the IRR, there is
823  *	nothing useful to be done in the handler.
824  *
825  *  Inputs  :   wakeup_irq  (Wakeup-interrupt bit)
826  *	arg		(Interrupt handler specific argument)
827  *  Outputs :   None
828  *
829  */
830 static irqreturn_t
831 ia64_mca_wakeup_int_handler(int wakeup_irq, void *arg)
832 {
833 	return IRQ_HANDLED;
834 }
835 
836 /* Function pointer for extra MCA recovery */
837 int (*ia64_mca_ucmc_extension)
838 	(void*,struct ia64_sal_os_state*)
839 	= NULL;
840 
841 int
842 ia64_reg_MCA_extension(int (*fn)(void *, struct ia64_sal_os_state *))
843 {
844 	if (ia64_mca_ucmc_extension)
845 		return 1;
846 
847 	ia64_mca_ucmc_extension = fn;
848 	return 0;
849 }
850 
851 void
852 ia64_unreg_MCA_extension(void)
853 {
854 	if (ia64_mca_ucmc_extension)
855 		ia64_mca_ucmc_extension = NULL;
856 }
857 
858 EXPORT_SYMBOL(ia64_reg_MCA_extension);
859 EXPORT_SYMBOL(ia64_unreg_MCA_extension);
860 
861 
862 static inline void
863 copy_reg(const u64 *fr, u64 fnat, unsigned long *tr, unsigned long *tnat)
864 {
865 	u64 fslot, tslot, nat;
866 	*tr = *fr;
867 	fslot = ((unsigned long)fr >> 3) & 63;
868 	tslot = ((unsigned long)tr >> 3) & 63;
869 	*tnat &= ~(1UL << tslot);
870 	nat = (fnat >> fslot) & 1;
871 	*tnat |= (nat << tslot);
872 }
873 
874 /* Change the comm field on the MCA/INT task to include the pid that
875  * was interrupted, it makes for easier debugging.  If that pid was 0
876  * (swapper or nested MCA/INIT) then use the start of the previous comm
877  * field suffixed with its cpu.
878  */
879 
880 static void
881 ia64_mca_modify_comm(const struct task_struct *previous_current)
882 {
883 	char *p, comm[sizeof(current->comm)];
884 	if (previous_current->pid)
885 		snprintf(comm, sizeof(comm), "%s %d",
886 			current->comm, previous_current->pid);
887 	else {
888 		int l;
889 		if ((p = strchr(previous_current->comm, ' ')))
890 			l = p - previous_current->comm;
891 		else
892 			l = strlen(previous_current->comm);
893 		snprintf(comm, sizeof(comm), "%s %*s %d",
894 			current->comm, l, previous_current->comm,
895 			task_thread_info(previous_current)->cpu);
896 	}
897 	memcpy(current->comm, comm, sizeof(current->comm));
898 }
899 
900 static void
901 finish_pt_regs(struct pt_regs *regs, struct ia64_sal_os_state *sos,
902 		unsigned long *nat)
903 {
904 	const pal_min_state_area_t *ms = sos->pal_min_state;
905 	const u64 *bank;
906 
907 	/* If ipsr.ic then use pmsa_{iip,ipsr,ifs}, else use
908 	 * pmsa_{xip,xpsr,xfs}
909 	 */
910 	if (ia64_psr(regs)->ic) {
911 		regs->cr_iip = ms->pmsa_iip;
912 		regs->cr_ipsr = ms->pmsa_ipsr;
913 		regs->cr_ifs = ms->pmsa_ifs;
914 	} else {
915 		regs->cr_iip = ms->pmsa_xip;
916 		regs->cr_ipsr = ms->pmsa_xpsr;
917 		regs->cr_ifs = ms->pmsa_xfs;
918 
919 		sos->iip = ms->pmsa_iip;
920 		sos->ipsr = ms->pmsa_ipsr;
921 		sos->ifs = ms->pmsa_ifs;
922 	}
923 	regs->pr = ms->pmsa_pr;
924 	regs->b0 = ms->pmsa_br0;
925 	regs->ar_rsc = ms->pmsa_rsc;
926 	copy_reg(&ms->pmsa_gr[1-1], ms->pmsa_nat_bits, &regs->r1, nat);
927 	copy_reg(&ms->pmsa_gr[2-1], ms->pmsa_nat_bits, &regs->r2, nat);
928 	copy_reg(&ms->pmsa_gr[3-1], ms->pmsa_nat_bits, &regs->r3, nat);
929 	copy_reg(&ms->pmsa_gr[8-1], ms->pmsa_nat_bits, &regs->r8, nat);
930 	copy_reg(&ms->pmsa_gr[9-1], ms->pmsa_nat_bits, &regs->r9, nat);
931 	copy_reg(&ms->pmsa_gr[10-1], ms->pmsa_nat_bits, &regs->r10, nat);
932 	copy_reg(&ms->pmsa_gr[11-1], ms->pmsa_nat_bits, &regs->r11, nat);
933 	copy_reg(&ms->pmsa_gr[12-1], ms->pmsa_nat_bits, &regs->r12, nat);
934 	copy_reg(&ms->pmsa_gr[13-1], ms->pmsa_nat_bits, &regs->r13, nat);
935 	copy_reg(&ms->pmsa_gr[14-1], ms->pmsa_nat_bits, &regs->r14, nat);
936 	copy_reg(&ms->pmsa_gr[15-1], ms->pmsa_nat_bits, &regs->r15, nat);
937 	if (ia64_psr(regs)->bn)
938 		bank = ms->pmsa_bank1_gr;
939 	else
940 		bank = ms->pmsa_bank0_gr;
941 	copy_reg(&bank[16-16], ms->pmsa_nat_bits, &regs->r16, nat);
942 	copy_reg(&bank[17-16], ms->pmsa_nat_bits, &regs->r17, nat);
943 	copy_reg(&bank[18-16], ms->pmsa_nat_bits, &regs->r18, nat);
944 	copy_reg(&bank[19-16], ms->pmsa_nat_bits, &regs->r19, nat);
945 	copy_reg(&bank[20-16], ms->pmsa_nat_bits, &regs->r20, nat);
946 	copy_reg(&bank[21-16], ms->pmsa_nat_bits, &regs->r21, nat);
947 	copy_reg(&bank[22-16], ms->pmsa_nat_bits, &regs->r22, nat);
948 	copy_reg(&bank[23-16], ms->pmsa_nat_bits, &regs->r23, nat);
949 	copy_reg(&bank[24-16], ms->pmsa_nat_bits, &regs->r24, nat);
950 	copy_reg(&bank[25-16], ms->pmsa_nat_bits, &regs->r25, nat);
951 	copy_reg(&bank[26-16], ms->pmsa_nat_bits, &regs->r26, nat);
952 	copy_reg(&bank[27-16], ms->pmsa_nat_bits, &regs->r27, nat);
953 	copy_reg(&bank[28-16], ms->pmsa_nat_bits, &regs->r28, nat);
954 	copy_reg(&bank[29-16], ms->pmsa_nat_bits, &regs->r29, nat);
955 	copy_reg(&bank[30-16], ms->pmsa_nat_bits, &regs->r30, nat);
956 	copy_reg(&bank[31-16], ms->pmsa_nat_bits, &regs->r31, nat);
957 }
958 
959 /* On entry to this routine, we are running on the per cpu stack, see
960  * mca_asm.h.  The original stack has not been touched by this event.  Some of
961  * the original stack's registers will be in the RBS on this stack.  This stack
962  * also contains a partial pt_regs and switch_stack, the rest of the data is in
963  * PAL minstate.
964  *
965  * The first thing to do is modify the original stack to look like a blocked
966  * task so we can run backtrace on the original task.  Also mark the per cpu
967  * stack as current to ensure that we use the correct task state, it also means
968  * that we can do backtrace on the MCA/INIT handler code itself.
969  */
970 
971 static struct task_struct *
972 ia64_mca_modify_original_stack(struct pt_regs *regs,
973 		const struct switch_stack *sw,
974 		struct ia64_sal_os_state *sos,
975 		const char *type)
976 {
977 	char *p;
978 	ia64_va va;
979 	extern char ia64_leave_kernel[];	/* Need asm address, not function descriptor */
980 	const pal_min_state_area_t *ms = sos->pal_min_state;
981 	struct task_struct *previous_current;
982 	struct pt_regs *old_regs;
983 	struct switch_stack *old_sw;
984 	unsigned size = sizeof(struct pt_regs) +
985 			sizeof(struct switch_stack) + 16;
986 	unsigned long *old_bspstore, *old_bsp;
987 	unsigned long *new_bspstore, *new_bsp;
988 	unsigned long old_unat, old_rnat, new_rnat, nat;
989 	u64 slots, loadrs = regs->loadrs;
990 	u64 r12 = ms->pmsa_gr[12-1], r13 = ms->pmsa_gr[13-1];
991 	u64 ar_bspstore = regs->ar_bspstore;
992 	u64 ar_bsp = regs->ar_bspstore + (loadrs >> 16);
993 	const char *msg;
994 	int cpu = smp_processor_id();
995 
996 	previous_current = curr_task(cpu);
997 	ia64_set_curr_task(cpu, current);
998 	if ((p = strchr(current->comm, ' ')))
999 		*p = '\0';
1000 
1001 	/* Best effort attempt to cope with MCA/INIT delivered while in
1002 	 * physical mode.
1003 	 */
1004 	regs->cr_ipsr = ms->pmsa_ipsr;
1005 	if (ia64_psr(regs)->dt == 0) {
1006 		va.l = r12;
1007 		if (va.f.reg == 0) {
1008 			va.f.reg = 7;
1009 			r12 = va.l;
1010 		}
1011 		va.l = r13;
1012 		if (va.f.reg == 0) {
1013 			va.f.reg = 7;
1014 			r13 = va.l;
1015 		}
1016 	}
1017 	if (ia64_psr(regs)->rt == 0) {
1018 		va.l = ar_bspstore;
1019 		if (va.f.reg == 0) {
1020 			va.f.reg = 7;
1021 			ar_bspstore = va.l;
1022 		}
1023 		va.l = ar_bsp;
1024 		if (va.f.reg == 0) {
1025 			va.f.reg = 7;
1026 			ar_bsp = va.l;
1027 		}
1028 	}
1029 
1030 	/* mca_asm.S ia64_old_stack() cannot assume that the dirty registers
1031 	 * have been copied to the old stack, the old stack may fail the
1032 	 * validation tests below.  So ia64_old_stack() must restore the dirty
1033 	 * registers from the new stack.  The old and new bspstore probably
1034 	 * have different alignments, so loadrs calculated on the old bsp
1035 	 * cannot be used to restore from the new bsp.  Calculate a suitable
1036 	 * loadrs for the new stack and save it in the new pt_regs, where
1037 	 * ia64_old_stack() can get it.
1038 	 */
1039 	old_bspstore = (unsigned long *)ar_bspstore;
1040 	old_bsp = (unsigned long *)ar_bsp;
1041 	slots = ia64_rse_num_regs(old_bspstore, old_bsp);
1042 	new_bspstore = (unsigned long *)((u64)current + IA64_RBS_OFFSET);
1043 	new_bsp = ia64_rse_skip_regs(new_bspstore, slots);
1044 	regs->loadrs = (new_bsp - new_bspstore) * 8 << 16;
1045 
1046 	/* Verify the previous stack state before we change it */
1047 	if (user_mode(regs)) {
1048 		msg = "occurred in user space";
1049 		/* previous_current is guaranteed to be valid when the task was
1050 		 * in user space, so ...
1051 		 */
1052 		ia64_mca_modify_comm(previous_current);
1053 		goto no_mod;
1054 	}
1055 
1056 	if (r13 != sos->prev_IA64_KR_CURRENT) {
1057 		msg = "inconsistent previous current and r13";
1058 		goto no_mod;
1059 	}
1060 
1061 	if (!mca_recover_range(ms->pmsa_iip)) {
1062 		if ((r12 - r13) >= KERNEL_STACK_SIZE) {
1063 			msg = "inconsistent r12 and r13";
1064 			goto no_mod;
1065 		}
1066 		if ((ar_bspstore - r13) >= KERNEL_STACK_SIZE) {
1067 			msg = "inconsistent ar.bspstore and r13";
1068 			goto no_mod;
1069 		}
1070 		va.p = old_bspstore;
1071 		if (va.f.reg < 5) {
1072 			msg = "old_bspstore is in the wrong region";
1073 			goto no_mod;
1074 		}
1075 		if ((ar_bsp - r13) >= KERNEL_STACK_SIZE) {
1076 			msg = "inconsistent ar.bsp and r13";
1077 			goto no_mod;
1078 		}
1079 		size += (ia64_rse_skip_regs(old_bspstore, slots) - old_bspstore) * 8;
1080 		if (ar_bspstore + size > r12) {
1081 			msg = "no room for blocked state";
1082 			goto no_mod;
1083 		}
1084 	}
1085 
1086 	ia64_mca_modify_comm(previous_current);
1087 
1088 	/* Make the original task look blocked.  First stack a struct pt_regs,
1089 	 * describing the state at the time of interrupt.  mca_asm.S built a
1090 	 * partial pt_regs, copy it and fill in the blanks using minstate.
1091 	 */
1092 	p = (char *)r12 - sizeof(*regs);
1093 	old_regs = (struct pt_regs *)p;
1094 	memcpy(old_regs, regs, sizeof(*regs));
1095 	old_regs->loadrs = loadrs;
1096 	old_unat = old_regs->ar_unat;
1097 	finish_pt_regs(old_regs, sos, &old_unat);
1098 
1099 	/* Next stack a struct switch_stack.  mca_asm.S built a partial
1100 	 * switch_stack, copy it and fill in the blanks using pt_regs and
1101 	 * minstate.
1102 	 *
1103 	 * In the synthesized switch_stack, b0 points to ia64_leave_kernel,
1104 	 * ar.pfs is set to 0.
1105 	 *
1106 	 * unwind.c::unw_unwind() does special processing for interrupt frames.
1107 	 * It checks if the PRED_NON_SYSCALL predicate is set, if the predicate
1108 	 * is clear then unw_unwind() does _not_ adjust bsp over pt_regs.  Not
1109 	 * that this is documented, of course.  Set PRED_NON_SYSCALL in the
1110 	 * switch_stack on the original stack so it will unwind correctly when
1111 	 * unwind.c reads pt_regs.
1112 	 *
1113 	 * thread.ksp is updated to point to the synthesized switch_stack.
1114 	 */
1115 	p -= sizeof(struct switch_stack);
1116 	old_sw = (struct switch_stack *)p;
1117 	memcpy(old_sw, sw, sizeof(*sw));
1118 	old_sw->caller_unat = old_unat;
1119 	old_sw->ar_fpsr = old_regs->ar_fpsr;
1120 	copy_reg(&ms->pmsa_gr[4-1], ms->pmsa_nat_bits, &old_sw->r4, &old_unat);
1121 	copy_reg(&ms->pmsa_gr[5-1], ms->pmsa_nat_bits, &old_sw->r5, &old_unat);
1122 	copy_reg(&ms->pmsa_gr[6-1], ms->pmsa_nat_bits, &old_sw->r6, &old_unat);
1123 	copy_reg(&ms->pmsa_gr[7-1], ms->pmsa_nat_bits, &old_sw->r7, &old_unat);
1124 	old_sw->b0 = (u64)ia64_leave_kernel;
1125 	old_sw->b1 = ms->pmsa_br1;
1126 	old_sw->ar_pfs = 0;
1127 	old_sw->ar_unat = old_unat;
1128 	old_sw->pr = old_regs->pr | (1UL << PRED_NON_SYSCALL);
1129 	previous_current->thread.ksp = (u64)p - 16;
1130 
1131 	/* Finally copy the original stack's registers back to its RBS.
1132 	 * Registers from ar.bspstore through ar.bsp at the time of the event
1133 	 * are in the current RBS, copy them back to the original stack.  The
1134 	 * copy must be done register by register because the original bspstore
1135 	 * and the current one have different alignments, so the saved RNAT
1136 	 * data occurs at different places.
1137 	 *
1138 	 * mca_asm does cover, so the old_bsp already includes all registers at
1139 	 * the time of MCA/INIT.  It also does flushrs, so all registers before
1140 	 * this function have been written to backing store on the MCA/INIT
1141 	 * stack.
1142 	 */
1143 	new_rnat = ia64_get_rnat(ia64_rse_rnat_addr(new_bspstore));
1144 	old_rnat = regs->ar_rnat;
1145 	while (slots--) {
1146 		if (ia64_rse_is_rnat_slot(new_bspstore)) {
1147 			new_rnat = ia64_get_rnat(new_bspstore++);
1148 		}
1149 		if (ia64_rse_is_rnat_slot(old_bspstore)) {
1150 			*old_bspstore++ = old_rnat;
1151 			old_rnat = 0;
1152 		}
1153 		nat = (new_rnat >> ia64_rse_slot_num(new_bspstore)) & 1UL;
1154 		old_rnat &= ~(1UL << ia64_rse_slot_num(old_bspstore));
1155 		old_rnat |= (nat << ia64_rse_slot_num(old_bspstore));
1156 		*old_bspstore++ = *new_bspstore++;
1157 	}
1158 	old_sw->ar_bspstore = (unsigned long)old_bspstore;
1159 	old_sw->ar_rnat = old_rnat;
1160 
1161 	sos->prev_task = previous_current;
1162 	return previous_current;
1163 
1164 no_mod:
1165 	mprintk(KERN_INFO "cpu %d, %s %s, original stack not modified\n",
1166 			smp_processor_id(), type, msg);
1167 	old_unat = regs->ar_unat;
1168 	finish_pt_regs(regs, sos, &old_unat);
1169 	return previous_current;
1170 }
1171 
1172 /* The monarch/slave interaction is based on monarch_cpu and requires that all
1173  * slaves have entered rendezvous before the monarch leaves.  If any cpu has
1174  * not entered rendezvous yet then wait a bit.  The assumption is that any
1175  * slave that has not rendezvoused after a reasonable time is never going to do
1176  * so.  In this context, slave includes cpus that respond to the MCA rendezvous
1177  * interrupt, as well as cpus that receive the INIT slave event.
1178  */
1179 
1180 static void
1181 ia64_wait_for_slaves(int monarch, const char *type)
1182 {
1183 	int c, i , wait;
1184 
1185 	/*
1186 	 * wait 5 seconds total for slaves (arbitrary)
1187 	 */
1188 	for (i = 0; i < 5000; i++) {
1189 		wait = 0;
1190 		for_each_online_cpu(c) {
1191 			if (c == monarch)
1192 				continue;
1193 			if (ia64_mc_info.imi_rendez_checkin[c]
1194 					== IA64_MCA_RENDEZ_CHECKIN_NOTDONE) {
1195 				udelay(1000);		/* short wait */
1196 				wait = 1;
1197 				break;
1198 			}
1199 		}
1200 		if (!wait)
1201 			goto all_in;
1202 	}
1203 
1204 	/*
1205 	 * Maybe slave(s) dead. Print buffered messages immediately.
1206 	 */
1207 	ia64_mlogbuf_finish(0);
1208 	mprintk(KERN_INFO "OS %s slave did not rendezvous on cpu", type);
1209 	for_each_online_cpu(c) {
1210 		if (c == monarch)
1211 			continue;
1212 		if (ia64_mc_info.imi_rendez_checkin[c] == IA64_MCA_RENDEZ_CHECKIN_NOTDONE)
1213 			mprintk(" %d", c);
1214 	}
1215 	mprintk("\n");
1216 	return;
1217 
1218 all_in:
1219 	mprintk(KERN_INFO "All OS %s slaves have reached rendezvous\n", type);
1220 	return;
1221 }
1222 
1223 /*  mca_insert_tr
1224  *
1225  *  Switch rid when TR reload and needed!
1226  *  iord: 1: itr, 2: itr;
1227  *
1228 */
1229 static void mca_insert_tr(u64 iord)
1230 {
1231 
1232 	int i;
1233 	u64 old_rr;
1234 	struct ia64_tr_entry *p;
1235 	unsigned long psr;
1236 	int cpu = smp_processor_id();
1237 
1238 	if (!ia64_idtrs[cpu])
1239 		return;
1240 
1241 	psr = ia64_clear_ic();
1242 	for (i = IA64_TR_ALLOC_BASE; i < IA64_TR_ALLOC_MAX; i++) {
1243 		p = ia64_idtrs[cpu] + (iord - 1) * IA64_TR_ALLOC_MAX;
1244 		if (p->pte & 0x1) {
1245 			old_rr = ia64_get_rr(p->ifa);
1246 			if (old_rr != p->rr) {
1247 				ia64_set_rr(p->ifa, p->rr);
1248 				ia64_srlz_d();
1249 			}
1250 			ia64_ptr(iord, p->ifa, p->itir >> 2);
1251 			ia64_srlz_i();
1252 			if (iord & 0x1) {
1253 				ia64_itr(0x1, i, p->ifa, p->pte, p->itir >> 2);
1254 				ia64_srlz_i();
1255 			}
1256 			if (iord & 0x2) {
1257 				ia64_itr(0x2, i, p->ifa, p->pte, p->itir >> 2);
1258 				ia64_srlz_i();
1259 			}
1260 			if (old_rr != p->rr) {
1261 				ia64_set_rr(p->ifa, old_rr);
1262 				ia64_srlz_d();
1263 			}
1264 		}
1265 	}
1266 	ia64_set_psr(psr);
1267 }
1268 
1269 /*
1270  * ia64_mca_handler
1271  *
1272  *	This is uncorrectable machine check handler called from OS_MCA
1273  *	dispatch code which is in turn called from SAL_CHECK().
1274  *	This is the place where the core of OS MCA handling is done.
1275  *	Right now the logs are extracted and displayed in a well-defined
1276  *	format. This handler code is supposed to be run only on the
1277  *	monarch processor. Once the monarch is done with MCA handling
1278  *	further MCA logging is enabled by clearing logs.
1279  *	Monarch also has the duty of sending wakeup-IPIs to pull the
1280  *	slave processors out of rendezvous spinloop.
1281  *
1282  *	If multiple processors call into OS_MCA, the first will become
1283  *	the monarch.  Subsequent cpus will be recorded in the mca_cpu
1284  *	bitmask.  After the first monarch has processed its MCA, it
1285  *	will wake up the next cpu in the mca_cpu bitmask and then go
1286  *	into the rendezvous loop.  When all processors have serviced
1287  *	their MCA, the last monarch frees up the rest of the processors.
1288  */
1289 void
1290 ia64_mca_handler(struct pt_regs *regs, struct switch_stack *sw,
1291 		 struct ia64_sal_os_state *sos)
1292 {
1293 	int recover, cpu = smp_processor_id();
1294 	struct task_struct *previous_current;
1295 	struct ia64_mca_notify_die nd =
1296 		{ .sos = sos, .monarch_cpu = &monarch_cpu, .data = &recover };
1297 	static atomic_t mca_count;
1298 	static cpumask_t mca_cpu;
1299 
1300 	if (atomic_add_return(1, &mca_count) == 1) {
1301 		monarch_cpu = cpu;
1302 		sos->monarch = 1;
1303 	} else {
1304 		cpumask_set_cpu(cpu, &mca_cpu);
1305 		sos->monarch = 0;
1306 	}
1307 	mprintk(KERN_INFO "Entered OS MCA handler. PSP=%lx cpu=%d "
1308 		"monarch=%ld\n", sos->proc_state_param, cpu, sos->monarch);
1309 
1310 	previous_current = ia64_mca_modify_original_stack(regs, sw, sos, "MCA");
1311 
1312 	NOTIFY_MCA(DIE_MCA_MONARCH_ENTER, regs, (long)&nd, 1);
1313 
1314 	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_CONCURRENT_MCA;
1315 	if (sos->monarch) {
1316 		ia64_wait_for_slaves(cpu, "MCA");
1317 
1318 		/* Wakeup all the processors which are spinning in the
1319 		 * rendezvous loop.  They will leave SAL, then spin in the OS
1320 		 * with interrupts disabled until this monarch cpu leaves the
1321 		 * MCA handler.  That gets control back to the OS so we can
1322 		 * backtrace the other cpus, backtrace when spinning in SAL
1323 		 * does not work.
1324 		 */
1325 		ia64_mca_wakeup_all();
1326 	} else {
1327 		while (cpumask_test_cpu(cpu, &mca_cpu))
1328 			cpu_relax();	/* spin until monarch wakes us */
1329 	}
1330 
1331 	NOTIFY_MCA(DIE_MCA_MONARCH_PROCESS, regs, (long)&nd, 1);
1332 
1333 	/* Get the MCA error record and log it */
1334 	ia64_mca_log_sal_error_record(SAL_INFO_TYPE_MCA);
1335 
1336 	/* MCA error recovery */
1337 	recover = (ia64_mca_ucmc_extension
1338 		&& ia64_mca_ucmc_extension(
1339 			IA64_LOG_CURR_BUFFER(SAL_INFO_TYPE_MCA),
1340 			sos));
1341 
1342 	if (recover) {
1343 		sal_log_record_header_t *rh = IA64_LOG_CURR_BUFFER(SAL_INFO_TYPE_MCA);
1344 		rh->severity = sal_log_severity_corrected;
1345 		ia64_sal_clear_state_info(SAL_INFO_TYPE_MCA);
1346 		sos->os_status = IA64_MCA_CORRECTED;
1347 	} else {
1348 		/* Dump buffered message to console */
1349 		ia64_mlogbuf_finish(1);
1350 	}
1351 
1352 	if (__this_cpu_read(ia64_mca_tr_reload)) {
1353 		mca_insert_tr(0x1); /*Reload dynamic itrs*/
1354 		mca_insert_tr(0x2); /*Reload dynamic itrs*/
1355 	}
1356 
1357 	NOTIFY_MCA(DIE_MCA_MONARCH_LEAVE, regs, (long)&nd, 1);
1358 
1359 	if (atomic_dec_return(&mca_count) > 0) {
1360 		int i;
1361 
1362 		/* wake up the next monarch cpu,
1363 		 * and put this cpu in the rendez loop.
1364 		 */
1365 		for_each_online_cpu(i) {
1366 			if (cpumask_test_cpu(i, &mca_cpu)) {
1367 				monarch_cpu = i;
1368 				cpumask_clear_cpu(i, &mca_cpu);	/* wake next cpu */
1369 				while (monarch_cpu != -1)
1370 					cpu_relax();	/* spin until last cpu leaves */
1371 				ia64_set_curr_task(cpu, previous_current);
1372 				ia64_mc_info.imi_rendez_checkin[cpu]
1373 						= IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1374 				return;
1375 			}
1376 		}
1377 	}
1378 	ia64_set_curr_task(cpu, previous_current);
1379 	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1380 	monarch_cpu = -1;	/* This frees the slaves and previous monarchs */
1381 }
1382 
1383 static DECLARE_WORK(cmc_disable_work, ia64_mca_cmc_vector_disable_keventd);
1384 static DECLARE_WORK(cmc_enable_work, ia64_mca_cmc_vector_enable_keventd);
1385 
1386 /*
1387  * ia64_mca_cmc_int_handler
1388  *
1389  *  This is corrected machine check interrupt handler.
1390  *	Right now the logs are extracted and displayed in a well-defined
1391  *	format.
1392  *
1393  * Inputs
1394  *      interrupt number
1395  *      client data arg ptr
1396  *
1397  * Outputs
1398  *	None
1399  */
1400 static irqreturn_t
1401 ia64_mca_cmc_int_handler(int cmc_irq, void *arg)
1402 {
1403 	static unsigned long	cmc_history[CMC_HISTORY_LENGTH];
1404 	static int		index;
1405 	static DEFINE_SPINLOCK(cmc_history_lock);
1406 
1407 	IA64_MCA_DEBUG("%s: received interrupt vector = %#x on CPU %d\n",
1408 		       __func__, cmc_irq, smp_processor_id());
1409 
1410 	/* SAL spec states this should run w/ interrupts enabled */
1411 	local_irq_enable();
1412 
1413 	spin_lock(&cmc_history_lock);
1414 	if (!cmc_polling_enabled) {
1415 		int i, count = 1; /* we know 1 happened now */
1416 		unsigned long now = jiffies;
1417 
1418 		for (i = 0; i < CMC_HISTORY_LENGTH; i++) {
1419 			if (now - cmc_history[i] <= HZ)
1420 				count++;
1421 		}
1422 
1423 		IA64_MCA_DEBUG(KERN_INFO "CMC threshold %d/%d\n", count, CMC_HISTORY_LENGTH);
1424 		if (count >= CMC_HISTORY_LENGTH) {
1425 
1426 			cmc_polling_enabled = 1;
1427 			spin_unlock(&cmc_history_lock);
1428 			/* If we're being hit with CMC interrupts, we won't
1429 			 * ever execute the schedule_work() below.  Need to
1430 			 * disable CMC interrupts on this processor now.
1431 			 */
1432 			ia64_mca_cmc_vector_disable(NULL);
1433 			schedule_work(&cmc_disable_work);
1434 
1435 			/*
1436 			 * Corrected errors will still be corrected, but
1437 			 * make sure there's a log somewhere that indicates
1438 			 * something is generating more than we can handle.
1439 			 */
1440 			printk(KERN_WARNING "WARNING: Switching to polling CMC handler; error records may be lost\n");
1441 
1442 			mod_timer(&cmc_poll_timer, jiffies + CMC_POLL_INTERVAL);
1443 
1444 			/* lock already released, get out now */
1445 			goto out;
1446 		} else {
1447 			cmc_history[index++] = now;
1448 			if (index == CMC_HISTORY_LENGTH)
1449 				index = 0;
1450 		}
1451 	}
1452 	spin_unlock(&cmc_history_lock);
1453 out:
1454 	/* Get the CMC error record and log it */
1455 	ia64_mca_log_sal_error_record(SAL_INFO_TYPE_CMC);
1456 
1457 	local_irq_disable();
1458 
1459 	return IRQ_HANDLED;
1460 }
1461 
1462 /*
1463  *  ia64_mca_cmc_int_caller
1464  *
1465  * 	Triggered by sw interrupt from CMC polling routine.  Calls
1466  * 	real interrupt handler and either triggers a sw interrupt
1467  * 	on the next cpu or does cleanup at the end.
1468  *
1469  * Inputs
1470  *	interrupt number
1471  *	client data arg ptr
1472  * Outputs
1473  * 	handled
1474  */
1475 static irqreturn_t
1476 ia64_mca_cmc_int_caller(int cmc_irq, void *arg)
1477 {
1478 	static int start_count = -1;
1479 	unsigned int cpuid;
1480 
1481 	cpuid = smp_processor_id();
1482 
1483 	/* If first cpu, update count */
1484 	if (start_count == -1)
1485 		start_count = IA64_LOG_COUNT(SAL_INFO_TYPE_CMC);
1486 
1487 	ia64_mca_cmc_int_handler(cmc_irq, arg);
1488 
1489 	cpuid = cpumask_next(cpuid+1, cpu_online_mask);
1490 
1491 	if (cpuid < nr_cpu_ids) {
1492 		platform_send_ipi(cpuid, IA64_CMCP_VECTOR, IA64_IPI_DM_INT, 0);
1493 	} else {
1494 		/* If no log record, switch out of polling mode */
1495 		if (start_count == IA64_LOG_COUNT(SAL_INFO_TYPE_CMC)) {
1496 
1497 			printk(KERN_WARNING "Returning to interrupt driven CMC handler\n");
1498 			schedule_work(&cmc_enable_work);
1499 			cmc_polling_enabled = 0;
1500 
1501 		} else {
1502 
1503 			mod_timer(&cmc_poll_timer, jiffies + CMC_POLL_INTERVAL);
1504 		}
1505 
1506 		start_count = -1;
1507 	}
1508 
1509 	return IRQ_HANDLED;
1510 }
1511 
1512 /*
1513  *  ia64_mca_cmc_poll
1514  *
1515  *	Poll for Corrected Machine Checks (CMCs)
1516  *
1517  * Inputs   :   dummy(unused)
1518  * Outputs  :   None
1519  *
1520  */
1521 static void
1522 ia64_mca_cmc_poll (struct timer_list *unused)
1523 {
1524 	/* Trigger a CMC interrupt cascade  */
1525 	platform_send_ipi(cpumask_first(cpu_online_mask), IA64_CMCP_VECTOR,
1526 							IA64_IPI_DM_INT, 0);
1527 }
1528 
1529 /*
1530  *  ia64_mca_cpe_int_caller
1531  *
1532  * 	Triggered by sw interrupt from CPE polling routine.  Calls
1533  * 	real interrupt handler and either triggers a sw interrupt
1534  * 	on the next cpu or does cleanup at the end.
1535  *
1536  * Inputs
1537  *	interrupt number
1538  *	client data arg ptr
1539  * Outputs
1540  * 	handled
1541  */
1542 #ifdef CONFIG_ACPI
1543 
1544 static irqreturn_t
1545 ia64_mca_cpe_int_caller(int cpe_irq, void *arg)
1546 {
1547 	static int start_count = -1;
1548 	static int poll_time = MIN_CPE_POLL_INTERVAL;
1549 	unsigned int cpuid;
1550 
1551 	cpuid = smp_processor_id();
1552 
1553 	/* If first cpu, update count */
1554 	if (start_count == -1)
1555 		start_count = IA64_LOG_COUNT(SAL_INFO_TYPE_CPE);
1556 
1557 	ia64_mca_cpe_int_handler(cpe_irq, arg);
1558 
1559 	cpuid = cpumask_next(cpuid+1, cpu_online_mask);
1560 
1561 	if (cpuid < NR_CPUS) {
1562 		platform_send_ipi(cpuid, IA64_CPEP_VECTOR, IA64_IPI_DM_INT, 0);
1563 	} else {
1564 		/*
1565 		 * If a log was recorded, increase our polling frequency,
1566 		 * otherwise, backoff or return to interrupt mode.
1567 		 */
1568 		if (start_count != IA64_LOG_COUNT(SAL_INFO_TYPE_CPE)) {
1569 			poll_time = max(MIN_CPE_POLL_INTERVAL, poll_time / 2);
1570 		} else if (cpe_vector < 0) {
1571 			poll_time = min(MAX_CPE_POLL_INTERVAL, poll_time * 2);
1572 		} else {
1573 			poll_time = MIN_CPE_POLL_INTERVAL;
1574 
1575 			printk(KERN_WARNING "Returning to interrupt driven CPE handler\n");
1576 			enable_irq(local_vector_to_irq(IA64_CPE_VECTOR));
1577 			cpe_poll_enabled = 0;
1578 		}
1579 
1580 		if (cpe_poll_enabled)
1581 			mod_timer(&cpe_poll_timer, jiffies + poll_time);
1582 		start_count = -1;
1583 	}
1584 
1585 	return IRQ_HANDLED;
1586 }
1587 
1588 /*
1589  *  ia64_mca_cpe_poll
1590  *
1591  *	Poll for Corrected Platform Errors (CPEs), trigger interrupt
1592  *	on first cpu, from there it will trickle through all the cpus.
1593  *
1594  * Inputs   :   dummy(unused)
1595  * Outputs  :   None
1596  *
1597  */
1598 static void
1599 ia64_mca_cpe_poll (struct timer_list *unused)
1600 {
1601 	/* Trigger a CPE interrupt cascade  */
1602 	platform_send_ipi(cpumask_first(cpu_online_mask), IA64_CPEP_VECTOR,
1603 							IA64_IPI_DM_INT, 0);
1604 }
1605 
1606 #endif /* CONFIG_ACPI */
1607 
1608 static int
1609 default_monarch_init_process(struct notifier_block *self, unsigned long val, void *data)
1610 {
1611 	int c;
1612 	struct task_struct *g, *t;
1613 	if (val != DIE_INIT_MONARCH_PROCESS)
1614 		return NOTIFY_DONE;
1615 #ifdef CONFIG_KEXEC
1616 	if (atomic_read(&kdump_in_progress))
1617 		return NOTIFY_DONE;
1618 #endif
1619 
1620 	/*
1621 	 * FIXME: mlogbuf will brim over with INIT stack dumps.
1622 	 * To enable show_stack from INIT, we use oops_in_progress which should
1623 	 * be used in real oops. This would cause something wrong after INIT.
1624 	 */
1625 	BREAK_LOGLEVEL(console_loglevel);
1626 	ia64_mlogbuf_dump_from_init();
1627 
1628 	printk(KERN_ERR "Processes interrupted by INIT -");
1629 	for_each_online_cpu(c) {
1630 		struct ia64_sal_os_state *s;
1631 		t = __va(__per_cpu_mca[c] + IA64_MCA_CPU_INIT_STACK_OFFSET);
1632 		s = (struct ia64_sal_os_state *)((char *)t + MCA_SOS_OFFSET);
1633 		g = s->prev_task;
1634 		if (g) {
1635 			if (g->pid)
1636 				printk(" %d", g->pid);
1637 			else
1638 				printk(" %d (cpu %d task 0x%p)", g->pid, task_cpu(g), g);
1639 		}
1640 	}
1641 	printk("\n\n");
1642 	if (read_trylock(&tasklist_lock)) {
1643 		do_each_thread (g, t) {
1644 			printk("\nBacktrace of pid %d (%s)\n", t->pid, t->comm);
1645 			show_stack(t, NULL);
1646 		} while_each_thread (g, t);
1647 		read_unlock(&tasklist_lock);
1648 	}
1649 	/* FIXME: This will not restore zapped printk locks. */
1650 	RESTORE_LOGLEVEL(console_loglevel);
1651 	return NOTIFY_DONE;
1652 }
1653 
1654 /*
1655  * C portion of the OS INIT handler
1656  *
1657  * Called from ia64_os_init_dispatch
1658  *
1659  * Inputs: pointer to pt_regs where processor info was saved.  SAL/OS state for
1660  * this event.  This code is used for both monarch and slave INIT events, see
1661  * sos->monarch.
1662  *
1663  * All INIT events switch to the INIT stack and change the previous process to
1664  * blocked status.  If one of the INIT events is the monarch then we are
1665  * probably processing the nmi button/command.  Use the monarch cpu to dump all
1666  * the processes.  The slave INIT events all spin until the monarch cpu
1667  * returns.  We can also get INIT slave events for MCA, in which case the MCA
1668  * process is the monarch.
1669  */
1670 
1671 void
1672 ia64_init_handler(struct pt_regs *regs, struct switch_stack *sw,
1673 		  struct ia64_sal_os_state *sos)
1674 {
1675 	static atomic_t slaves;
1676 	static atomic_t monarchs;
1677 	struct task_struct *previous_current;
1678 	int cpu = smp_processor_id();
1679 	struct ia64_mca_notify_die nd =
1680 		{ .sos = sos, .monarch_cpu = &monarch_cpu };
1681 
1682 	NOTIFY_INIT(DIE_INIT_ENTER, regs, (long)&nd, 0);
1683 
1684 	mprintk(KERN_INFO "Entered OS INIT handler. PSP=%lx cpu=%d monarch=%ld\n",
1685 		sos->proc_state_param, cpu, sos->monarch);
1686 	salinfo_log_wakeup(SAL_INFO_TYPE_INIT, NULL, 0, 0);
1687 
1688 	previous_current = ia64_mca_modify_original_stack(regs, sw, sos, "INIT");
1689 	sos->os_status = IA64_INIT_RESUME;
1690 
1691 	/* FIXME: Workaround for broken proms that drive all INIT events as
1692 	 * slaves.  The last slave that enters is promoted to be a monarch.
1693 	 * Remove this code in September 2006, that gives platforms a year to
1694 	 * fix their proms and get their customers updated.
1695 	 */
1696 	if (!sos->monarch && atomic_add_return(1, &slaves) == num_online_cpus()) {
1697 		mprintk(KERN_WARNING "%s: Promoting cpu %d to monarch.\n",
1698 		        __func__, cpu);
1699 		atomic_dec(&slaves);
1700 		sos->monarch = 1;
1701 	}
1702 
1703 	/* FIXME: Workaround for broken proms that drive all INIT events as
1704 	 * monarchs.  Second and subsequent monarchs are demoted to slaves.
1705 	 * Remove this code in September 2006, that gives platforms a year to
1706 	 * fix their proms and get their customers updated.
1707 	 */
1708 	if (sos->monarch && atomic_add_return(1, &monarchs) > 1) {
1709 		mprintk(KERN_WARNING "%s: Demoting cpu %d to slave.\n",
1710 			       __func__, cpu);
1711 		atomic_dec(&monarchs);
1712 		sos->monarch = 0;
1713 	}
1714 
1715 	if (!sos->monarch) {
1716 		ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_INIT;
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 enters */
1724 #endif
1725 
1726 		NOTIFY_INIT(DIE_INIT_SLAVE_ENTER, regs, (long)&nd, 1);
1727 		NOTIFY_INIT(DIE_INIT_SLAVE_PROCESS, regs, (long)&nd, 1);
1728 
1729 #ifdef CONFIG_KEXEC
1730 		while (monarch_cpu != -1 && !atomic_read(&kdump_in_progress))
1731 			udelay(1000);
1732 #else
1733 		while (monarch_cpu != -1)
1734 			cpu_relax();	/* spin until monarch leaves */
1735 #endif
1736 
1737 		NOTIFY_INIT(DIE_INIT_SLAVE_LEAVE, regs, (long)&nd, 1);
1738 
1739 		mprintk("Slave on cpu %d returning to normal service.\n", cpu);
1740 		ia64_set_curr_task(cpu, previous_current);
1741 		ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1742 		atomic_dec(&slaves);
1743 		return;
1744 	}
1745 
1746 	monarch_cpu = cpu;
1747 	NOTIFY_INIT(DIE_INIT_MONARCH_ENTER, regs, (long)&nd, 1);
1748 
1749 	/*
1750 	 * Wait for a bit.  On some machines (e.g., HP's zx2000 and zx6000, INIT can be
1751 	 * generated via the BMC's command-line interface, but since the console is on the
1752 	 * same serial line, the user will need some time to switch out of the BMC before
1753 	 * the dump begins.
1754 	 */
1755 	mprintk("Delaying for 5 seconds...\n");
1756 	udelay(5*1000000);
1757 	ia64_wait_for_slaves(cpu, "INIT");
1758 	/* If nobody intercepts DIE_INIT_MONARCH_PROCESS then we drop through
1759 	 * to default_monarch_init_process() above and just print all the
1760 	 * tasks.
1761 	 */
1762 	NOTIFY_INIT(DIE_INIT_MONARCH_PROCESS, regs, (long)&nd, 1);
1763 	NOTIFY_INIT(DIE_INIT_MONARCH_LEAVE, regs, (long)&nd, 1);
1764 
1765 	mprintk("\nINIT dump complete.  Monarch on cpu %d returning to normal service.\n", cpu);
1766 	atomic_dec(&monarchs);
1767 	ia64_set_curr_task(cpu, previous_current);
1768 	monarch_cpu = -1;
1769 	return;
1770 }
1771 
1772 static int __init
1773 ia64_mca_disable_cpe_polling(char *str)
1774 {
1775 	cpe_poll_enabled = 0;
1776 	return 1;
1777 }
1778 
1779 __setup("disable_cpe_poll", ia64_mca_disable_cpe_polling);
1780 
1781 static struct irqaction cmci_irqaction = {
1782 	.handler =	ia64_mca_cmc_int_handler,
1783 	.name =		"cmc_hndlr"
1784 };
1785 
1786 static struct irqaction cmcp_irqaction = {
1787 	.handler =	ia64_mca_cmc_int_caller,
1788 	.name =		"cmc_poll"
1789 };
1790 
1791 static struct irqaction mca_rdzv_irqaction = {
1792 	.handler =	ia64_mca_rendez_int_handler,
1793 	.name =		"mca_rdzv"
1794 };
1795 
1796 static struct irqaction mca_wkup_irqaction = {
1797 	.handler =	ia64_mca_wakeup_int_handler,
1798 	.name =		"mca_wkup"
1799 };
1800 
1801 #ifdef CONFIG_ACPI
1802 static struct irqaction mca_cpe_irqaction = {
1803 	.handler =	ia64_mca_cpe_int_handler,
1804 	.name =		"cpe_hndlr"
1805 };
1806 
1807 static struct irqaction mca_cpep_irqaction = {
1808 	.handler =	ia64_mca_cpe_int_caller,
1809 	.name =		"cpe_poll"
1810 };
1811 #endif /* CONFIG_ACPI */
1812 
1813 /* Minimal format of the MCA/INIT stacks.  The pseudo processes that run on
1814  * these stacks can never sleep, they cannot return from the kernel to user
1815  * space, they do not appear in a normal ps listing.  So there is no need to
1816  * format most of the fields.
1817  */
1818 
1819 static void
1820 format_mca_init_stack(void *mca_data, unsigned long offset,
1821 		const char *type, int cpu)
1822 {
1823 	struct task_struct *p = (struct task_struct *)((char *)mca_data + offset);
1824 	struct thread_info *ti;
1825 	memset(p, 0, KERNEL_STACK_SIZE);
1826 	ti = task_thread_info(p);
1827 	ti->flags = _TIF_MCA_INIT;
1828 	ti->preempt_count = 1;
1829 	ti->task = p;
1830 	ti->cpu = cpu;
1831 	p->stack = ti;
1832 	p->state = TASK_UNINTERRUPTIBLE;
1833 	cpumask_set_cpu(cpu, &p->cpus_allowed);
1834 	INIT_LIST_HEAD(&p->tasks);
1835 	p->parent = p->real_parent = p->group_leader = p;
1836 	INIT_LIST_HEAD(&p->children);
1837 	INIT_LIST_HEAD(&p->sibling);
1838 	strncpy(p->comm, type, sizeof(p->comm)-1);
1839 }
1840 
1841 /* Caller prevents this from being called after init */
1842 static void * __ref mca_bootmem(void)
1843 {
1844 	return memblock_alloc(sizeof(struct ia64_mca_cpu), KERNEL_STACK_SIZE);
1845 }
1846 
1847 /* Do per-CPU MCA-related initialization.  */
1848 void
1849 ia64_mca_cpu_init(void *cpu_data)
1850 {
1851 	void *pal_vaddr;
1852 	void *data;
1853 	long sz = sizeof(struct ia64_mca_cpu);
1854 	int cpu = smp_processor_id();
1855 	static int first_time = 1;
1856 
1857 	/*
1858 	 * Structure will already be allocated if cpu has been online,
1859 	 * then offlined.
1860 	 */
1861 	if (__per_cpu_mca[cpu]) {
1862 		data = __va(__per_cpu_mca[cpu]);
1863 	} else {
1864 		if (first_time) {
1865 			data = mca_bootmem();
1866 			first_time = 0;
1867 		} else
1868 			data = (void *)__get_free_pages(GFP_KERNEL,
1869 							get_order(sz));
1870 		if (!data)
1871 			panic("Could not allocate MCA memory for cpu %d\n",
1872 					cpu);
1873 	}
1874 	format_mca_init_stack(data, offsetof(struct ia64_mca_cpu, mca_stack),
1875 		"MCA", cpu);
1876 	format_mca_init_stack(data, offsetof(struct ia64_mca_cpu, init_stack),
1877 		"INIT", cpu);
1878 	__this_cpu_write(ia64_mca_data, (__per_cpu_mca[cpu] = __pa(data)));
1879 
1880 	/*
1881 	 * Stash away a copy of the PTE needed to map the per-CPU page.
1882 	 * We may need it during MCA recovery.
1883 	 */
1884 	__this_cpu_write(ia64_mca_per_cpu_pte,
1885 		pte_val(mk_pte_phys(__pa(cpu_data), PAGE_KERNEL)));
1886 
1887 	/*
1888 	 * Also, stash away a copy of the PAL address and the PTE
1889 	 * needed to map it.
1890 	 */
1891 	pal_vaddr = efi_get_pal_addr();
1892 	if (!pal_vaddr)
1893 		return;
1894 	__this_cpu_write(ia64_mca_pal_base,
1895 		GRANULEROUNDDOWN((unsigned long) pal_vaddr));
1896 	__this_cpu_write(ia64_mca_pal_pte, pte_val(mk_pte_phys(__pa(pal_vaddr),
1897 							      PAGE_KERNEL)));
1898 }
1899 
1900 static int ia64_mca_cpu_online(unsigned int cpu)
1901 {
1902 	unsigned long flags;
1903 
1904 	local_irq_save(flags);
1905 	if (!cmc_polling_enabled)
1906 		ia64_mca_cmc_vector_enable(NULL);
1907 	local_irq_restore(flags);
1908 	return 0;
1909 }
1910 
1911 /*
1912  * ia64_mca_init
1913  *
1914  *  Do all the system level mca specific initialization.
1915  *
1916  *	1. Register spinloop and wakeup request interrupt vectors
1917  *
1918  *	2. Register OS_MCA handler entry point
1919  *
1920  *	3. Register OS_INIT handler entry point
1921  *
1922  *  4. Initialize MCA/CMC/INIT related log buffers maintained by the OS.
1923  *
1924  *  Note that this initialization is done very early before some kernel
1925  *  services are available.
1926  *
1927  *  Inputs  :   None
1928  *
1929  *  Outputs :   None
1930  */
1931 void __init
1932 ia64_mca_init(void)
1933 {
1934 	ia64_fptr_t *init_hldlr_ptr_monarch = (ia64_fptr_t *)ia64_os_init_dispatch_monarch;
1935 	ia64_fptr_t *init_hldlr_ptr_slave = (ia64_fptr_t *)ia64_os_init_dispatch_slave;
1936 	ia64_fptr_t *mca_hldlr_ptr = (ia64_fptr_t *)ia64_os_mca_dispatch;
1937 	int i;
1938 	long rc;
1939 	struct ia64_sal_retval isrv;
1940 	unsigned long timeout = IA64_MCA_RENDEZ_TIMEOUT; /* platform specific */
1941 	static struct notifier_block default_init_monarch_nb = {
1942 		.notifier_call = default_monarch_init_process,
1943 		.priority = 0/* we need to notified last */
1944 	};
1945 
1946 	IA64_MCA_DEBUG("%s: begin\n", __func__);
1947 
1948 	/* Clear the Rendez checkin flag for all cpus */
1949 	for(i = 0 ; i < NR_CPUS; i++)
1950 		ia64_mc_info.imi_rendez_checkin[i] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1951 
1952 	/*
1953 	 * Register the rendezvous spinloop and wakeup mechanism with SAL
1954 	 */
1955 
1956 	/* Register the rendezvous interrupt vector with SAL */
1957 	while (1) {
1958 		isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_RENDEZ_INT,
1959 					      SAL_MC_PARAM_MECHANISM_INT,
1960 					      IA64_MCA_RENDEZ_VECTOR,
1961 					      timeout,
1962 					      SAL_MC_PARAM_RZ_ALWAYS);
1963 		rc = isrv.status;
1964 		if (rc == 0)
1965 			break;
1966 		if (rc == -2) {
1967 			printk(KERN_INFO "Increasing MCA rendezvous timeout from "
1968 				"%ld to %ld milliseconds\n", timeout, isrv.v0);
1969 			timeout = isrv.v0;
1970 			NOTIFY_MCA(DIE_MCA_NEW_TIMEOUT, NULL, timeout, 0);
1971 			continue;
1972 		}
1973 		printk(KERN_ERR "Failed to register rendezvous interrupt "
1974 		       "with SAL (status %ld)\n", rc);
1975 		return;
1976 	}
1977 
1978 	/* Register the wakeup interrupt vector with SAL */
1979 	isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_RENDEZ_WAKEUP,
1980 				      SAL_MC_PARAM_MECHANISM_INT,
1981 				      IA64_MCA_WAKEUP_VECTOR,
1982 				      0, 0);
1983 	rc = isrv.status;
1984 	if (rc) {
1985 		printk(KERN_ERR "Failed to register wakeup interrupt with SAL "
1986 		       "(status %ld)\n", rc);
1987 		return;
1988 	}
1989 
1990 	IA64_MCA_DEBUG("%s: registered MCA rendezvous spinloop and wakeup mech.\n", __func__);
1991 
1992 	ia64_mc_info.imi_mca_handler        = ia64_tpa(mca_hldlr_ptr->fp);
1993 	/*
1994 	 * XXX - disable SAL checksum by setting size to 0; should be
1995 	 *	ia64_tpa(ia64_os_mca_dispatch_end) - ia64_tpa(ia64_os_mca_dispatch);
1996 	 */
1997 	ia64_mc_info.imi_mca_handler_size	= 0;
1998 
1999 	/* Register the os mca handler with SAL */
2000 	if ((rc = ia64_sal_set_vectors(SAL_VECTOR_OS_MCA,
2001 				       ia64_mc_info.imi_mca_handler,
2002 				       ia64_tpa(mca_hldlr_ptr->gp),
2003 				       ia64_mc_info.imi_mca_handler_size,
2004 				       0, 0, 0)))
2005 	{
2006 		printk(KERN_ERR "Failed to register OS MCA handler with SAL "
2007 		       "(status %ld)\n", rc);
2008 		return;
2009 	}
2010 
2011 	IA64_MCA_DEBUG("%s: registered OS MCA handler with SAL at 0x%lx, gp = 0x%lx\n", __func__,
2012 		       ia64_mc_info.imi_mca_handler, ia64_tpa(mca_hldlr_ptr->gp));
2013 
2014 	/*
2015 	 * XXX - disable SAL checksum by setting size to 0, should be
2016 	 * size of the actual init handler in mca_asm.S.
2017 	 */
2018 	ia64_mc_info.imi_monarch_init_handler		= ia64_tpa(init_hldlr_ptr_monarch->fp);
2019 	ia64_mc_info.imi_monarch_init_handler_size	= 0;
2020 	ia64_mc_info.imi_slave_init_handler		= ia64_tpa(init_hldlr_ptr_slave->fp);
2021 	ia64_mc_info.imi_slave_init_handler_size	= 0;
2022 
2023 	IA64_MCA_DEBUG("%s: OS INIT handler at %lx\n", __func__,
2024 		       ia64_mc_info.imi_monarch_init_handler);
2025 
2026 	/* Register the os init handler with SAL */
2027 	if ((rc = ia64_sal_set_vectors(SAL_VECTOR_OS_INIT,
2028 				       ia64_mc_info.imi_monarch_init_handler,
2029 				       ia64_tpa(ia64_getreg(_IA64_REG_GP)),
2030 				       ia64_mc_info.imi_monarch_init_handler_size,
2031 				       ia64_mc_info.imi_slave_init_handler,
2032 				       ia64_tpa(ia64_getreg(_IA64_REG_GP)),
2033 				       ia64_mc_info.imi_slave_init_handler_size)))
2034 	{
2035 		printk(KERN_ERR "Failed to register m/s INIT handlers with SAL "
2036 		       "(status %ld)\n", rc);
2037 		return;
2038 	}
2039 	if (register_die_notifier(&default_init_monarch_nb)) {
2040 		printk(KERN_ERR "Failed to register default monarch INIT process\n");
2041 		return;
2042 	}
2043 
2044 	IA64_MCA_DEBUG("%s: registered OS INIT handler with SAL\n", __func__);
2045 
2046 	/* Initialize the areas set aside by the OS to buffer the
2047 	 * platform/processor error states for MCA/INIT/CMC
2048 	 * handling.
2049 	 */
2050 	ia64_log_init(SAL_INFO_TYPE_MCA);
2051 	ia64_log_init(SAL_INFO_TYPE_INIT);
2052 	ia64_log_init(SAL_INFO_TYPE_CMC);
2053 	ia64_log_init(SAL_INFO_TYPE_CPE);
2054 
2055 	mca_init = 1;
2056 	printk(KERN_INFO "MCA related initialization done\n");
2057 }
2058 
2059 
2060 /*
2061  * These pieces cannot be done in ia64_mca_init() because it is called before
2062  * early_irq_init() which would wipe out our percpu irq registrations. But we
2063  * cannot leave them until ia64_mca_late_init() because by then all the other
2064  * processors have been brought online and have set their own CMC vectors to
2065  * point at a non-existant action. Called from arch_early_irq_init().
2066  */
2067 void __init ia64_mca_irq_init(void)
2068 {
2069 	/*
2070 	 *  Configure the CMCI/P vector and handler. Interrupts for CMC are
2071 	 *  per-processor, so AP CMC interrupts are setup in smp_callin() (smpboot.c).
2072 	 */
2073 	register_percpu_irq(IA64_CMC_VECTOR, &cmci_irqaction);
2074 	register_percpu_irq(IA64_CMCP_VECTOR, &cmcp_irqaction);
2075 	ia64_mca_cmc_vector_setup();       /* Setup vector on BSP */
2076 
2077 	/* Setup the MCA rendezvous interrupt vector */
2078 	register_percpu_irq(IA64_MCA_RENDEZ_VECTOR, &mca_rdzv_irqaction);
2079 
2080 	/* Setup the MCA wakeup interrupt vector */
2081 	register_percpu_irq(IA64_MCA_WAKEUP_VECTOR, &mca_wkup_irqaction);
2082 
2083 #ifdef CONFIG_ACPI
2084 	/* Setup the CPEI/P handler */
2085 	register_percpu_irq(IA64_CPEP_VECTOR, &mca_cpep_irqaction);
2086 #endif
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 	/* Setup the CMCI/P vector and handler */
2106 	timer_setup(&cmc_poll_timer, ia64_mca_cmc_poll, 0);
2107 
2108 	/* Unmask/enable the vector */
2109 	cmc_polling_enabled = 0;
2110 	cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "ia64/mca:online",
2111 			  ia64_mca_cpu_online, NULL);
2112 	IA64_MCA_DEBUG("%s: CMCI/P setup and enabled.\n", __func__);
2113 
2114 #ifdef CONFIG_ACPI
2115 	/* Setup the CPEI/P vector and handler */
2116 	cpe_vector = acpi_request_vector(ACPI_INTERRUPT_CPEI);
2117 	timer_setup(&cpe_poll_timer, ia64_mca_cpe_poll, 0);
2118 
2119 	{
2120 		unsigned int irq;
2121 
2122 		if (cpe_vector >= 0) {
2123 			/* If platform supports CPEI, enable the irq. */
2124 			irq = local_vector_to_irq(cpe_vector);
2125 			if (irq > 0) {
2126 				cpe_poll_enabled = 0;
2127 				irq_set_status_flags(irq, IRQ_PER_CPU);
2128 				setup_irq(irq, &mca_cpe_irqaction);
2129 				ia64_cpe_irq = irq;
2130 				ia64_mca_register_cpev(cpe_vector);
2131 				IA64_MCA_DEBUG("%s: CPEI/P setup and enabled.\n",
2132 					__func__);
2133 				return 0;
2134 			}
2135 			printk(KERN_ERR "%s: Failed to find irq for CPE "
2136 					"interrupt handler, vector %d\n",
2137 					__func__, cpe_vector);
2138 		}
2139 		/* If platform doesn't support CPEI, get the timer going. */
2140 		if (cpe_poll_enabled) {
2141 			ia64_mca_cpe_poll(0UL);
2142 			IA64_MCA_DEBUG("%s: CPEP setup and enabled.\n", __func__);
2143 		}
2144 	}
2145 #endif
2146 
2147 	return 0;
2148 }
2149 
2150 device_initcall(ia64_mca_late_init);
2151