xref: /openbmc/linux/arch/ia64/kernel/mca.c (revision 7a010c3c)
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/efi.h>
95 #include <asm/meminit.h>
96 #include <asm/page.h>
97 #include <asm/ptrace.h>
98 #include <asm/sal.h>
99 #include <asm/mca.h>
100 #include <asm/mca_asm.h>
101 #include <asm/kexec.h>
102 
103 #include <asm/irq.h>
104 #include <asm/hw_irq.h>
105 #include <asm/tlb.h>
106 
107 #include "mca_drv.h"
108 #include "entry.h"
109 #include "irq.h"
110 
111 #if defined(IA64_MCA_DEBUG_INFO)
112 # define IA64_MCA_DEBUG(fmt...) printk(fmt)
113 #else
114 # define IA64_MCA_DEBUG(fmt...) do {} while (0)
115 #endif
116 
117 #define NOTIFY_INIT(event, regs, arg, spin)				\
118 do {									\
119 	if ((notify_die((event), "INIT", (regs), (arg), 0, 0)		\
120 			== NOTIFY_STOP) && ((spin) == 1))		\
121 		ia64_mca_spin(__func__);				\
122 } while (0)
123 
124 #define NOTIFY_MCA(event, regs, arg, spin)				\
125 do {									\
126 	if ((notify_die((event), "MCA", (regs), (arg), 0, 0)		\
127 			== NOTIFY_STOP) && ((spin) == 1))		\
128 		ia64_mca_spin(__func__);				\
129 } while (0)
130 
131 /* Used by mca_asm.S */
132 DEFINE_PER_CPU(u64, ia64_mca_data); /* == __per_cpu_mca[smp_processor_id()] */
133 DEFINE_PER_CPU(u64, ia64_mca_per_cpu_pte); /* PTE to map per-CPU area */
134 DEFINE_PER_CPU(u64, ia64_mca_pal_pte);	    /* PTE to map PAL code */
135 DEFINE_PER_CPU(u64, ia64_mca_pal_base);    /* vaddr PAL code granule */
136 DEFINE_PER_CPU(u64, ia64_mca_tr_reload);   /* Flag for TR reload */
137 
138 unsigned long __per_cpu_mca[NR_CPUS];
139 
140 /* In mca_asm.S */
141 extern void			ia64_os_init_dispatch_monarch (void);
142 extern void			ia64_os_init_dispatch_slave (void);
143 
144 static int monarch_cpu = -1;
145 
146 static ia64_mc_info_t		ia64_mc_info;
147 
148 #define MAX_CPE_POLL_INTERVAL (15*60*HZ) /* 15 minutes */
149 #define MIN_CPE_POLL_INTERVAL (2*60*HZ)  /* 2 minutes */
150 #define CMC_POLL_INTERVAL     (1*60*HZ)  /* 1 minute */
151 #define CPE_HISTORY_LENGTH    5
152 #define CMC_HISTORY_LENGTH    5
153 
154 static struct timer_list cpe_poll_timer;
155 static struct timer_list cmc_poll_timer;
156 /*
157  * This variable tells whether we are currently in polling mode.
158  * Start with this in the wrong state so we won't play w/ timers
159  * before the system is ready.
160  */
161 static int cmc_polling_enabled = 1;
162 
163 /*
164  * Clearing this variable prevents CPE polling from getting activated
165  * in mca_late_init.  Use it if your system doesn't provide a CPEI,
166  * but encounters problems retrieving CPE logs.  This should only be
167  * necessary for debugging.
168  */
169 static int cpe_poll_enabled = 1;
170 
171 extern void salinfo_log_wakeup(int type, u8 *buffer, u64 size, int irqsafe);
172 
173 static int mca_init __initdata;
174 
175 /*
176  * limited & delayed printing support for MCA/INIT handler
177  */
178 
179 #define mprintk(fmt...) ia64_mca_printk(fmt)
180 
181 #define MLOGBUF_SIZE (512+256*NR_CPUS)
182 #define MLOGBUF_MSGMAX 256
183 static char mlogbuf[MLOGBUF_SIZE];
184 static DEFINE_SPINLOCK(mlogbuf_wlock);	/* mca context only */
185 static DEFINE_SPINLOCK(mlogbuf_rlock);	/* normal context only */
186 static unsigned long mlogbuf_start;
187 static unsigned long mlogbuf_end;
188 static unsigned int mlogbuf_finished = 0;
189 static unsigned long mlogbuf_timestamp = 0;
190 
191 static int loglevel_save = -1;
192 #define BREAK_LOGLEVEL(__console_loglevel)		\
193 	oops_in_progress = 1;				\
194 	if (loglevel_save < 0)				\
195 		loglevel_save = __console_loglevel;	\
196 	__console_loglevel = 15;
197 
198 #define RESTORE_LOGLEVEL(__console_loglevel)		\
199 	if (loglevel_save >= 0) {			\
200 		__console_loglevel = loglevel_save;	\
201 		loglevel_save = -1;			\
202 	}						\
203 	mlogbuf_finished = 0;				\
204 	oops_in_progress = 0;
205 
206 /*
207  * Push messages into buffer, print them later if not urgent.
208  */
209 void ia64_mca_printk(const char *fmt, ...)
210 {
211 	va_list args;
212 	int printed_len;
213 	char temp_buf[MLOGBUF_MSGMAX];
214 	char *p;
215 
216 	va_start(args, fmt);
217 	printed_len = vscnprintf(temp_buf, sizeof(temp_buf), fmt, args);
218 	va_end(args);
219 
220 	/* Copy the output into mlogbuf */
221 	if (oops_in_progress) {
222 		/* mlogbuf was abandoned, use printk directly instead. */
223 		printk("%s", temp_buf);
224 	} else {
225 		spin_lock(&mlogbuf_wlock);
226 		for (p = temp_buf; *p; p++) {
227 			unsigned long next = (mlogbuf_end + 1) % MLOGBUF_SIZE;
228 			if (next != mlogbuf_start) {
229 				mlogbuf[mlogbuf_end] = *p;
230 				mlogbuf_end = next;
231 			} else {
232 				/* buffer full */
233 				break;
234 			}
235 		}
236 		mlogbuf[mlogbuf_end] = '\0';
237 		spin_unlock(&mlogbuf_wlock);
238 	}
239 }
240 EXPORT_SYMBOL(ia64_mca_printk);
241 
242 /*
243  * Print buffered messages.
244  *  NOTE: call this after returning normal context. (ex. from salinfod)
245  */
246 void ia64_mlogbuf_dump(void)
247 {
248 	char temp_buf[MLOGBUF_MSGMAX];
249 	char *p;
250 	unsigned long index;
251 	unsigned long flags;
252 	unsigned int printed_len;
253 
254 	/* Get output from mlogbuf */
255 	while (mlogbuf_start != mlogbuf_end) {
256 		temp_buf[0] = '\0';
257 		p = temp_buf;
258 		printed_len = 0;
259 
260 		spin_lock_irqsave(&mlogbuf_rlock, flags);
261 
262 		index = mlogbuf_start;
263 		while (index != mlogbuf_end) {
264 			*p = mlogbuf[index];
265 			index = (index + 1) % MLOGBUF_SIZE;
266 			if (!*p)
267 				break;
268 			p++;
269 			if (++printed_len >= MLOGBUF_MSGMAX - 1)
270 				break;
271 		}
272 		*p = '\0';
273 		if (temp_buf[0])
274 			printk("%s", temp_buf);
275 		mlogbuf_start = index;
276 
277 		mlogbuf_timestamp = 0;
278 		spin_unlock_irqrestore(&mlogbuf_rlock, flags);
279 	}
280 }
281 EXPORT_SYMBOL(ia64_mlogbuf_dump);
282 
283 /*
284  * Call this if system is going to down or if immediate flushing messages to
285  * console is required. (ex. recovery was failed, crash dump is going to be
286  * invoked, long-wait rendezvous etc.)
287  *  NOTE: this should be called from monarch.
288  */
289 static void ia64_mlogbuf_finish(int wait)
290 {
291 	BREAK_LOGLEVEL(console_loglevel);
292 
293 	spin_lock_init(&mlogbuf_rlock);
294 	ia64_mlogbuf_dump();
295 	printk(KERN_EMERG "mlogbuf_finish: printing switched to urgent mode, "
296 		"MCA/INIT might be dodgy or fail.\n");
297 
298 	if (!wait)
299 		return;
300 
301 	/* wait for console */
302 	printk("Delaying for 5 seconds...\n");
303 	udelay(5*1000000);
304 
305 	mlogbuf_finished = 1;
306 }
307 
308 /*
309  * Print buffered messages from INIT context.
310  */
311 static void ia64_mlogbuf_dump_from_init(void)
312 {
313 	if (mlogbuf_finished)
314 		return;
315 
316 	if (mlogbuf_timestamp &&
317 			time_before(jiffies, mlogbuf_timestamp + 30 * HZ)) {
318 		printk(KERN_ERR "INIT: mlogbuf_dump is interrupted by INIT "
319 			" and the system seems to be messed up.\n");
320 		ia64_mlogbuf_finish(0);
321 		return;
322 	}
323 
324 	if (!spin_trylock(&mlogbuf_rlock)) {
325 		printk(KERN_ERR "INIT: mlogbuf_dump is interrupted by INIT. "
326 			"Generated messages other than stack dump will be "
327 			"buffered to mlogbuf and will be printed later.\n");
328 		printk(KERN_ERR "INIT: If messages would not printed after "
329 			"this INIT, wait 30sec and assert INIT again.\n");
330 		if (!mlogbuf_timestamp)
331 			mlogbuf_timestamp = jiffies;
332 		return;
333 	}
334 	spin_unlock(&mlogbuf_rlock);
335 	ia64_mlogbuf_dump();
336 }
337 
338 static inline void
339 ia64_mca_spin(const char *func)
340 {
341 	if (monarch_cpu == smp_processor_id())
342 		ia64_mlogbuf_finish(0);
343 	mprintk(KERN_EMERG "%s: spinning here, not returning to SAL\n", func);
344 	while (1)
345 		cpu_relax();
346 }
347 /*
348  * IA64_MCA log support
349  */
350 #define IA64_MAX_LOGS		2	/* Double-buffering for nested MCAs */
351 #define IA64_MAX_LOG_TYPES      4   /* MCA, INIT, CMC, CPE */
352 
353 typedef struct ia64_state_log_s
354 {
355 	spinlock_t	isl_lock;
356 	int		isl_index;
357 	unsigned long	isl_count;
358 	ia64_err_rec_t  *isl_log[IA64_MAX_LOGS]; /* need space to store header + error log */
359 } ia64_state_log_t;
360 
361 static ia64_state_log_t ia64_state_log[IA64_MAX_LOG_TYPES];
362 
363 #define IA64_LOG_LOCK_INIT(it) spin_lock_init(&ia64_state_log[it].isl_lock)
364 #define IA64_LOG_LOCK(it)      spin_lock_irqsave(&ia64_state_log[it].isl_lock, s)
365 #define IA64_LOG_UNLOCK(it)    spin_unlock_irqrestore(&ia64_state_log[it].isl_lock,s)
366 #define IA64_LOG_NEXT_INDEX(it)    ia64_state_log[it].isl_index
367 #define IA64_LOG_CURR_INDEX(it)    1 - ia64_state_log[it].isl_index
368 #define IA64_LOG_INDEX_INC(it) \
369     {ia64_state_log[it].isl_index = 1 - ia64_state_log[it].isl_index; \
370     ia64_state_log[it].isl_count++;}
371 #define IA64_LOG_INDEX_DEC(it) \
372     ia64_state_log[it].isl_index = 1 - ia64_state_log[it].isl_index
373 #define IA64_LOG_NEXT_BUFFER(it)   (void *)((ia64_state_log[it].isl_log[IA64_LOG_NEXT_INDEX(it)]))
374 #define IA64_LOG_CURR_BUFFER(it)   (void *)((ia64_state_log[it].isl_log[IA64_LOG_CURR_INDEX(it)]))
375 #define IA64_LOG_COUNT(it)         ia64_state_log[it].isl_count
376 
377 static inline void ia64_log_allocate(int it, u64 size)
378 {
379 	ia64_state_log[it].isl_log[IA64_LOG_CURR_INDEX(it)] =
380 		(ia64_err_rec_t *)memblock_alloc(size, SMP_CACHE_BYTES);
381 	if (!ia64_state_log[it].isl_log[IA64_LOG_CURR_INDEX(it)])
382 		panic("%s: Failed to allocate %llu bytes\n", __func__, size);
383 
384 	ia64_state_log[it].isl_log[IA64_LOG_NEXT_INDEX(it)] =
385 		(ia64_err_rec_t *)memblock_alloc(size, SMP_CACHE_BYTES);
386 	if (!ia64_state_log[it].isl_log[IA64_LOG_NEXT_INDEX(it)])
387 		panic("%s: Failed to allocate %llu bytes\n", __func__, size);
388 }
389 
390 /*
391  * ia64_log_init
392  *	Reset the OS ia64 log buffer
393  * Inputs   :   info_type   (SAL_INFO_TYPE_{MCA,INIT,CMC,CPE})
394  * Outputs	:	None
395  */
396 static void __init
397 ia64_log_init(int sal_info_type)
398 {
399 	u64	max_size = 0;
400 
401 	IA64_LOG_NEXT_INDEX(sal_info_type) = 0;
402 	IA64_LOG_LOCK_INIT(sal_info_type);
403 
404 	// SAL will tell us the maximum size of any error record of this type
405 	max_size = ia64_sal_get_state_info_size(sal_info_type);
406 	if (!max_size)
407 		/* alloc_bootmem() doesn't like zero-sized allocations! */
408 		return;
409 
410 	// set up OS data structures to hold error info
411 	ia64_log_allocate(sal_info_type, max_size);
412 }
413 
414 /*
415  * ia64_log_get
416  *
417  *	Get the current MCA log from SAL and copy it into the OS log buffer.
418  *
419  *  Inputs  :   info_type   (SAL_INFO_TYPE_{MCA,INIT,CMC,CPE})
420  *              irq_safe    whether you can use printk at this point
421  *  Outputs :   size        (total record length)
422  *              *buffer     (ptr to error record)
423  *
424  */
425 static u64
426 ia64_log_get(int sal_info_type, u8 **buffer, int irq_safe)
427 {
428 	sal_log_record_header_t     *log_buffer;
429 	u64                         total_len = 0;
430 	unsigned long               s;
431 
432 	IA64_LOG_LOCK(sal_info_type);
433 
434 	/* Get the process state information */
435 	log_buffer = IA64_LOG_NEXT_BUFFER(sal_info_type);
436 
437 	total_len = ia64_sal_get_state_info(sal_info_type, (u64 *)log_buffer);
438 
439 	if (total_len) {
440 		IA64_LOG_INDEX_INC(sal_info_type);
441 		IA64_LOG_UNLOCK(sal_info_type);
442 		if (irq_safe) {
443 			IA64_MCA_DEBUG("%s: SAL error record type %d retrieved. Record length = %ld\n",
444 				       __func__, sal_info_type, total_len);
445 		}
446 		*buffer = (u8 *) log_buffer;
447 		return total_len;
448 	} else {
449 		IA64_LOG_UNLOCK(sal_info_type);
450 		return 0;
451 	}
452 }
453 
454 /*
455  *  ia64_mca_log_sal_error_record
456  *
457  *  This function retrieves a specified error record type from SAL
458  *  and wakes up any processes waiting for error records.
459  *
460  *  Inputs  :   sal_info_type   (Type of error record MCA/CMC/CPE)
461  *              FIXME: remove MCA and irq_safe.
462  */
463 static void
464 ia64_mca_log_sal_error_record(int sal_info_type)
465 {
466 	u8 *buffer;
467 	sal_log_record_header_t *rh;
468 	u64 size;
469 	int irq_safe = sal_info_type != SAL_INFO_TYPE_MCA;
470 #ifdef IA64_MCA_DEBUG_INFO
471 	static const char * const rec_name[] = { "MCA", "INIT", "CMC", "CPE" };
472 #endif
473 
474 	size = ia64_log_get(sal_info_type, &buffer, irq_safe);
475 	if (!size)
476 		return;
477 
478 	salinfo_log_wakeup(sal_info_type, buffer, size, irq_safe);
479 
480 	if (irq_safe)
481 		IA64_MCA_DEBUG("CPU %d: SAL log contains %s error record\n",
482 			smp_processor_id(),
483 			sal_info_type < ARRAY_SIZE(rec_name) ? rec_name[sal_info_type] : "UNKNOWN");
484 
485 	/* Clear logs from corrected errors in case there's no user-level logger */
486 	rh = (sal_log_record_header_t *)buffer;
487 	if (rh->severity == sal_log_severity_corrected)
488 		ia64_sal_clear_state_info(sal_info_type);
489 }
490 
491 /*
492  * search_mca_table
493  *  See if the MCA surfaced in an instruction range
494  *  that has been tagged as recoverable.
495  *
496  *  Inputs
497  *	first	First address range to check
498  *	last	Last address range to check
499  *	ip	Instruction pointer, address we are looking for
500  *
501  * Return value:
502  *      1 on Success (in the table)/ 0 on Failure (not in the  table)
503  */
504 int
505 search_mca_table (const struct mca_table_entry *first,
506                 const struct mca_table_entry *last,
507                 unsigned long ip)
508 {
509         const struct mca_table_entry *curr;
510         u64 curr_start, curr_end;
511 
512         curr = first;
513         while (curr <= last) {
514                 curr_start = (u64) &curr->start_addr + curr->start_addr;
515                 curr_end = (u64) &curr->end_addr + curr->end_addr;
516 
517                 if ((ip >= curr_start) && (ip <= curr_end)) {
518                         return 1;
519                 }
520                 curr++;
521         }
522         return 0;
523 }
524 
525 /* Given an address, look for it in the mca tables. */
526 int mca_recover_range(unsigned long addr)
527 {
528 	extern struct mca_table_entry __start___mca_table[];
529 	extern struct mca_table_entry __stop___mca_table[];
530 
531 	return search_mca_table(__start___mca_table, __stop___mca_table-1, addr);
532 }
533 EXPORT_SYMBOL_GPL(mca_recover_range);
534 
535 int cpe_vector = -1;
536 int ia64_cpe_irq = -1;
537 
538 static irqreturn_t
539 ia64_mca_cpe_int_handler (int cpe_irq, void *arg)
540 {
541 	static unsigned long	cpe_history[CPE_HISTORY_LENGTH];
542 	static int		index;
543 	static DEFINE_SPINLOCK(cpe_history_lock);
544 
545 	IA64_MCA_DEBUG("%s: received interrupt vector = %#x on CPU %d\n",
546 		       __func__, cpe_irq, smp_processor_id());
547 
548 	/* SAL spec states this should run w/ interrupts enabled */
549 	local_irq_enable();
550 
551 	spin_lock(&cpe_history_lock);
552 	if (!cpe_poll_enabled && cpe_vector >= 0) {
553 
554 		int i, count = 1; /* we know 1 happened now */
555 		unsigned long now = jiffies;
556 
557 		for (i = 0; i < CPE_HISTORY_LENGTH; i++) {
558 			if (now - cpe_history[i] <= HZ)
559 				count++;
560 		}
561 
562 		IA64_MCA_DEBUG(KERN_INFO "CPE threshold %d/%d\n", count, CPE_HISTORY_LENGTH);
563 		if (count >= CPE_HISTORY_LENGTH) {
564 
565 			cpe_poll_enabled = 1;
566 			spin_unlock(&cpe_history_lock);
567 			disable_irq_nosync(local_vector_to_irq(IA64_CPE_VECTOR));
568 
569 			/*
570 			 * Corrected errors will still be corrected, but
571 			 * make sure there's a log somewhere that indicates
572 			 * something is generating more than we can handle.
573 			 */
574 			printk(KERN_WARNING "WARNING: Switching to polling CPE handler; error records may be lost\n");
575 
576 			mod_timer(&cpe_poll_timer, jiffies + MIN_CPE_POLL_INTERVAL);
577 
578 			/* lock already released, get out now */
579 			goto out;
580 		} else {
581 			cpe_history[index++] = now;
582 			if (index == CPE_HISTORY_LENGTH)
583 				index = 0;
584 		}
585 	}
586 	spin_unlock(&cpe_history_lock);
587 out:
588 	/* Get the CPE error record and log it */
589 	ia64_mca_log_sal_error_record(SAL_INFO_TYPE_CPE);
590 
591 	local_irq_disable();
592 
593 	return IRQ_HANDLED;
594 }
595 
596 /*
597  * ia64_mca_register_cpev
598  *
599  *  Register the corrected platform error vector with SAL.
600  *
601  *  Inputs
602  *      cpev        Corrected Platform Error Vector number
603  *
604  *  Outputs
605  *      None
606  */
607 void
608 ia64_mca_register_cpev (int cpev)
609 {
610 	/* Register the CPE interrupt vector with SAL */
611 	struct ia64_sal_retval isrv;
612 
613 	isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_CPE_INT, SAL_MC_PARAM_MECHANISM_INT, cpev, 0, 0);
614 	if (isrv.status) {
615 		printk(KERN_ERR "Failed to register Corrected Platform "
616 		       "Error interrupt vector with SAL (status %ld)\n", isrv.status);
617 		return;
618 	}
619 
620 	IA64_MCA_DEBUG("%s: corrected platform error "
621 		       "vector %#x registered\n", __func__, cpev);
622 }
623 
624 /*
625  * ia64_mca_cmc_vector_setup
626  *
627  *  Setup the corrected machine check vector register in the processor.
628  *  (The interrupt is masked on boot. ia64_mca_late_init unmask this.)
629  *  This function is invoked on a per-processor basis.
630  *
631  * Inputs
632  *      None
633  *
634  * Outputs
635  *	None
636  */
637 void
638 ia64_mca_cmc_vector_setup (void)
639 {
640 	cmcv_reg_t	cmcv;
641 
642 	cmcv.cmcv_regval	= 0;
643 	cmcv.cmcv_mask		= 1;        /* Mask/disable interrupt at first */
644 	cmcv.cmcv_vector	= IA64_CMC_VECTOR;
645 	ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
646 
647 	IA64_MCA_DEBUG("%s: CPU %d corrected machine check vector %#x registered.\n",
648 		       __func__, smp_processor_id(), IA64_CMC_VECTOR);
649 
650 	IA64_MCA_DEBUG("%s: CPU %d CMCV = %#016lx\n",
651 		       __func__, smp_processor_id(), ia64_getreg(_IA64_REG_CR_CMCV));
652 }
653 
654 /*
655  * ia64_mca_cmc_vector_disable
656  *
657  *  Mask the corrected machine check vector register in the processor.
658  *  This function is invoked on a per-processor basis.
659  *
660  * Inputs
661  *      dummy(unused)
662  *
663  * Outputs
664  *	None
665  */
666 static void
667 ia64_mca_cmc_vector_disable (void *dummy)
668 {
669 	cmcv_reg_t	cmcv;
670 
671 	cmcv.cmcv_regval = ia64_getreg(_IA64_REG_CR_CMCV);
672 
673 	cmcv.cmcv_mask = 1; /* Mask/disable interrupt */
674 	ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
675 
676 	IA64_MCA_DEBUG("%s: CPU %d corrected machine check vector %#x disabled.\n",
677 		       __func__, smp_processor_id(), cmcv.cmcv_vector);
678 }
679 
680 /*
681  * ia64_mca_cmc_vector_enable
682  *
683  *  Unmask the corrected machine check vector register in the processor.
684  *  This function is invoked on a per-processor basis.
685  *
686  * Inputs
687  *      dummy(unused)
688  *
689  * Outputs
690  *	None
691  */
692 static void
693 ia64_mca_cmc_vector_enable (void *dummy)
694 {
695 	cmcv_reg_t	cmcv;
696 
697 	cmcv.cmcv_regval = ia64_getreg(_IA64_REG_CR_CMCV);
698 
699 	cmcv.cmcv_mask = 0; /* Unmask/enable interrupt */
700 	ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
701 
702 	IA64_MCA_DEBUG("%s: CPU %d corrected machine check vector %#x enabled.\n",
703 		       __func__, smp_processor_id(), cmcv.cmcv_vector);
704 }
705 
706 /*
707  * ia64_mca_cmc_vector_disable_keventd
708  *
709  * Called via keventd (smp_call_function() is not safe in interrupt context) to
710  * disable the cmc interrupt vector.
711  */
712 static void
713 ia64_mca_cmc_vector_disable_keventd(struct work_struct *unused)
714 {
715 	on_each_cpu(ia64_mca_cmc_vector_disable, NULL, 0);
716 }
717 
718 /*
719  * ia64_mca_cmc_vector_enable_keventd
720  *
721  * Called via keventd (smp_call_function() is not safe in interrupt context) to
722  * enable the cmc interrupt vector.
723  */
724 static void
725 ia64_mca_cmc_vector_enable_keventd(struct work_struct *unused)
726 {
727 	on_each_cpu(ia64_mca_cmc_vector_enable, NULL, 0);
728 }
729 
730 /*
731  * ia64_mca_wakeup
732  *
733  *	Send an inter-cpu interrupt to wake-up a particular cpu.
734  *
735  *  Inputs  :   cpuid
736  *  Outputs :   None
737  */
738 static void
739 ia64_mca_wakeup(int cpu)
740 {
741 	ia64_send_ipi(cpu, IA64_MCA_WAKEUP_VECTOR, IA64_IPI_DM_INT, 0);
742 }
743 
744 /*
745  * ia64_mca_wakeup_all
746  *
747  *	Wakeup all the slave cpus which have rendez'ed previously.
748  *
749  *  Inputs  :   None
750  *  Outputs :   None
751  */
752 static void
753 ia64_mca_wakeup_all(void)
754 {
755 	int cpu;
756 
757 	/* Clear the Rendez checkin flag for all cpus */
758 	for_each_online_cpu(cpu) {
759 		if (ia64_mc_info.imi_rendez_checkin[cpu] == IA64_MCA_RENDEZ_CHECKIN_DONE)
760 			ia64_mca_wakeup(cpu);
761 	}
762 
763 }
764 
765 /*
766  * ia64_mca_rendez_interrupt_handler
767  *
768  *	This is handler used to put slave processors into spinloop
769  *	while the monarch processor does the mca handling and later
770  *	wake each slave up once the monarch is done.  The state
771  *	IA64_MCA_RENDEZ_CHECKIN_DONE indicates the cpu is rendez'ed
772  *	in SAL.  The state IA64_MCA_RENDEZ_CHECKIN_NOTDONE indicates
773  *	the cpu has come out of OS rendezvous.
774  *
775  *  Inputs  :   None
776  *  Outputs :   None
777  */
778 static irqreturn_t
779 ia64_mca_rendez_int_handler(int rendez_irq, void *arg)
780 {
781 	unsigned long flags;
782 	int cpu = smp_processor_id();
783 	struct ia64_mca_notify_die nd =
784 		{ .sos = NULL, .monarch_cpu = &monarch_cpu };
785 
786 	/* Mask all interrupts */
787 	local_irq_save(flags);
788 
789 	NOTIFY_MCA(DIE_MCA_RENDZVOUS_ENTER, get_irq_regs(), (long)&nd, 1);
790 
791 	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_DONE;
792 	/* Register with the SAL monarch that the slave has
793 	 * reached SAL
794 	 */
795 	ia64_sal_mc_rendez();
796 
797 	NOTIFY_MCA(DIE_MCA_RENDZVOUS_PROCESS, get_irq_regs(), (long)&nd, 1);
798 
799 	/* Wait for the monarch cpu to exit. */
800 	while (monarch_cpu != -1)
801 	       cpu_relax();	/* spin until monarch leaves */
802 
803 	NOTIFY_MCA(DIE_MCA_RENDZVOUS_LEAVE, get_irq_regs(), (long)&nd, 1);
804 
805 	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
806 	/* Enable all interrupts */
807 	local_irq_restore(flags);
808 	return IRQ_HANDLED;
809 }
810 
811 /*
812  * ia64_mca_wakeup_int_handler
813  *
814  *	The interrupt handler for processing the inter-cpu interrupt to the
815  *	slave cpu which was spinning in the rendez loop.
816  *	Since this spinning is done by turning off the interrupts and
817  *	polling on the wakeup-interrupt bit in the IRR, there is
818  *	nothing useful to be done in the handler.
819  *
820  *  Inputs  :   wakeup_irq  (Wakeup-interrupt bit)
821  *	arg		(Interrupt handler specific argument)
822  *  Outputs :   None
823  *
824  */
825 static irqreturn_t
826 ia64_mca_wakeup_int_handler(int wakeup_irq, void *arg)
827 {
828 	return IRQ_HANDLED;
829 }
830 
831 /* Function pointer for extra MCA recovery */
832 int (*ia64_mca_ucmc_extension)
833 	(void*,struct ia64_sal_os_state*)
834 	= NULL;
835 
836 int
837 ia64_reg_MCA_extension(int (*fn)(void *, struct ia64_sal_os_state *))
838 {
839 	if (ia64_mca_ucmc_extension)
840 		return 1;
841 
842 	ia64_mca_ucmc_extension = fn;
843 	return 0;
844 }
845 
846 void
847 ia64_unreg_MCA_extension(void)
848 {
849 	if (ia64_mca_ucmc_extension)
850 		ia64_mca_ucmc_extension = NULL;
851 }
852 
853 EXPORT_SYMBOL(ia64_reg_MCA_extension);
854 EXPORT_SYMBOL(ia64_unreg_MCA_extension);
855 
856 
857 static inline void
858 copy_reg(const u64 *fr, u64 fnat, unsigned long *tr, unsigned long *tnat)
859 {
860 	u64 fslot, tslot, nat;
861 	*tr = *fr;
862 	fslot = ((unsigned long)fr >> 3) & 63;
863 	tslot = ((unsigned long)tr >> 3) & 63;
864 	*tnat &= ~(1UL << tslot);
865 	nat = (fnat >> fslot) & 1;
866 	*tnat |= (nat << tslot);
867 }
868 
869 /* Change the comm field on the MCA/INT task to include the pid that
870  * was interrupted, it makes for easier debugging.  If that pid was 0
871  * (swapper or nested MCA/INIT) then use the start of the previous comm
872  * field suffixed with its cpu.
873  */
874 
875 static void
876 ia64_mca_modify_comm(const struct task_struct *previous_current)
877 {
878 	char *p, comm[sizeof(current->comm)];
879 	if (previous_current->pid)
880 		snprintf(comm, sizeof(comm), "%s %d",
881 			current->comm, previous_current->pid);
882 	else {
883 		int l;
884 		if ((p = strchr(previous_current->comm, ' ')))
885 			l = p - previous_current->comm;
886 		else
887 			l = strlen(previous_current->comm);
888 		snprintf(comm, sizeof(comm), "%s %*s %d",
889 			current->comm, l, previous_current->comm,
890 			task_thread_info(previous_current)->cpu);
891 	}
892 	memcpy(current->comm, comm, sizeof(current->comm));
893 }
894 
895 static void
896 finish_pt_regs(struct pt_regs *regs, struct ia64_sal_os_state *sos,
897 		unsigned long *nat)
898 {
899 	const struct pal_min_state_area *ms = sos->pal_min_state;
900 	const u64 *bank;
901 
902 	/* If ipsr.ic then use pmsa_{iip,ipsr,ifs}, else use
903 	 * pmsa_{xip,xpsr,xfs}
904 	 */
905 	if (ia64_psr(regs)->ic) {
906 		regs->cr_iip = ms->pmsa_iip;
907 		regs->cr_ipsr = ms->pmsa_ipsr;
908 		regs->cr_ifs = ms->pmsa_ifs;
909 	} else {
910 		regs->cr_iip = ms->pmsa_xip;
911 		regs->cr_ipsr = ms->pmsa_xpsr;
912 		regs->cr_ifs = ms->pmsa_xfs;
913 
914 		sos->iip = ms->pmsa_iip;
915 		sos->ipsr = ms->pmsa_ipsr;
916 		sos->ifs = ms->pmsa_ifs;
917 	}
918 	regs->pr = ms->pmsa_pr;
919 	regs->b0 = ms->pmsa_br0;
920 	regs->ar_rsc = ms->pmsa_rsc;
921 	copy_reg(&ms->pmsa_gr[1-1], ms->pmsa_nat_bits, &regs->r1, nat);
922 	copy_reg(&ms->pmsa_gr[2-1], ms->pmsa_nat_bits, &regs->r2, nat);
923 	copy_reg(&ms->pmsa_gr[3-1], ms->pmsa_nat_bits, &regs->r3, nat);
924 	copy_reg(&ms->pmsa_gr[8-1], ms->pmsa_nat_bits, &regs->r8, nat);
925 	copy_reg(&ms->pmsa_gr[9-1], ms->pmsa_nat_bits, &regs->r9, nat);
926 	copy_reg(&ms->pmsa_gr[10-1], ms->pmsa_nat_bits, &regs->r10, nat);
927 	copy_reg(&ms->pmsa_gr[11-1], ms->pmsa_nat_bits, &regs->r11, nat);
928 	copy_reg(&ms->pmsa_gr[12-1], ms->pmsa_nat_bits, &regs->r12, nat);
929 	copy_reg(&ms->pmsa_gr[13-1], ms->pmsa_nat_bits, &regs->r13, nat);
930 	copy_reg(&ms->pmsa_gr[14-1], ms->pmsa_nat_bits, &regs->r14, nat);
931 	copy_reg(&ms->pmsa_gr[15-1], ms->pmsa_nat_bits, &regs->r15, nat);
932 	if (ia64_psr(regs)->bn)
933 		bank = ms->pmsa_bank1_gr;
934 	else
935 		bank = ms->pmsa_bank0_gr;
936 	copy_reg(&bank[16-16], ms->pmsa_nat_bits, &regs->r16, nat);
937 	copy_reg(&bank[17-16], ms->pmsa_nat_bits, &regs->r17, nat);
938 	copy_reg(&bank[18-16], ms->pmsa_nat_bits, &regs->r18, nat);
939 	copy_reg(&bank[19-16], ms->pmsa_nat_bits, &regs->r19, nat);
940 	copy_reg(&bank[20-16], ms->pmsa_nat_bits, &regs->r20, nat);
941 	copy_reg(&bank[21-16], ms->pmsa_nat_bits, &regs->r21, nat);
942 	copy_reg(&bank[22-16], ms->pmsa_nat_bits, &regs->r22, nat);
943 	copy_reg(&bank[23-16], ms->pmsa_nat_bits, &regs->r23, nat);
944 	copy_reg(&bank[24-16], ms->pmsa_nat_bits, &regs->r24, nat);
945 	copy_reg(&bank[25-16], ms->pmsa_nat_bits, &regs->r25, nat);
946 	copy_reg(&bank[26-16], ms->pmsa_nat_bits, &regs->r26, nat);
947 	copy_reg(&bank[27-16], ms->pmsa_nat_bits, &regs->r27, nat);
948 	copy_reg(&bank[28-16], ms->pmsa_nat_bits, &regs->r28, nat);
949 	copy_reg(&bank[29-16], ms->pmsa_nat_bits, &regs->r29, nat);
950 	copy_reg(&bank[30-16], ms->pmsa_nat_bits, &regs->r30, nat);
951 	copy_reg(&bank[31-16], ms->pmsa_nat_bits, &regs->r31, nat);
952 }
953 
954 /* On entry to this routine, we are running on the per cpu stack, see
955  * mca_asm.h.  The original stack has not been touched by this event.  Some of
956  * the original stack's registers will be in the RBS on this stack.  This stack
957  * also contains a partial pt_regs and switch_stack, the rest of the data is in
958  * PAL minstate.
959  *
960  * The first thing to do is modify the original stack to look like a blocked
961  * task so we can run backtrace on the original task.  Also mark the per cpu
962  * stack as current to ensure that we use the correct task state, it also means
963  * that we can do backtrace on the MCA/INIT handler code itself.
964  */
965 
966 static struct task_struct *
967 ia64_mca_modify_original_stack(struct pt_regs *regs,
968 		const struct switch_stack *sw,
969 		struct ia64_sal_os_state *sos,
970 		const char *type)
971 {
972 	char *p;
973 	ia64_va va;
974 	extern char ia64_leave_kernel[];	/* Need asm address, not function descriptor */
975 	const struct pal_min_state_area *ms = sos->pal_min_state;
976 	struct task_struct *previous_current;
977 	struct pt_regs *old_regs;
978 	struct switch_stack *old_sw;
979 	unsigned size = sizeof(struct pt_regs) +
980 			sizeof(struct switch_stack) + 16;
981 	unsigned long *old_bspstore, *old_bsp;
982 	unsigned long *new_bspstore, *new_bsp;
983 	unsigned long old_unat, old_rnat, new_rnat, nat;
984 	u64 slots, loadrs = regs->loadrs;
985 	u64 r12 = ms->pmsa_gr[12-1], r13 = ms->pmsa_gr[13-1];
986 	u64 ar_bspstore = regs->ar_bspstore;
987 	u64 ar_bsp = regs->ar_bspstore + (loadrs >> 16);
988 	const char *msg;
989 	int cpu = smp_processor_id();
990 
991 	previous_current = curr_task(cpu);
992 	ia64_set_curr_task(cpu, current);
993 	if ((p = strchr(current->comm, ' ')))
994 		*p = '\0';
995 
996 	/* Best effort attempt to cope with MCA/INIT delivered while in
997 	 * physical mode.
998 	 */
999 	regs->cr_ipsr = ms->pmsa_ipsr;
1000 	if (ia64_psr(regs)->dt == 0) {
1001 		va.l = r12;
1002 		if (va.f.reg == 0) {
1003 			va.f.reg = 7;
1004 			r12 = va.l;
1005 		}
1006 		va.l = r13;
1007 		if (va.f.reg == 0) {
1008 			va.f.reg = 7;
1009 			r13 = va.l;
1010 		}
1011 	}
1012 	if (ia64_psr(regs)->rt == 0) {
1013 		va.l = ar_bspstore;
1014 		if (va.f.reg == 0) {
1015 			va.f.reg = 7;
1016 			ar_bspstore = va.l;
1017 		}
1018 		va.l = ar_bsp;
1019 		if (va.f.reg == 0) {
1020 			va.f.reg = 7;
1021 			ar_bsp = va.l;
1022 		}
1023 	}
1024 
1025 	/* mca_asm.S ia64_old_stack() cannot assume that the dirty registers
1026 	 * have been copied to the old stack, the old stack may fail the
1027 	 * validation tests below.  So ia64_old_stack() must restore the dirty
1028 	 * registers from the new stack.  The old and new bspstore probably
1029 	 * have different alignments, so loadrs calculated on the old bsp
1030 	 * cannot be used to restore from the new bsp.  Calculate a suitable
1031 	 * loadrs for the new stack and save it in the new pt_regs, where
1032 	 * ia64_old_stack() can get it.
1033 	 */
1034 	old_bspstore = (unsigned long *)ar_bspstore;
1035 	old_bsp = (unsigned long *)ar_bsp;
1036 	slots = ia64_rse_num_regs(old_bspstore, old_bsp);
1037 	new_bspstore = (unsigned long *)((u64)current + IA64_RBS_OFFSET);
1038 	new_bsp = ia64_rse_skip_regs(new_bspstore, slots);
1039 	regs->loadrs = (new_bsp - new_bspstore) * 8 << 16;
1040 
1041 	/* Verify the previous stack state before we change it */
1042 	if (user_mode(regs)) {
1043 		msg = "occurred in user space";
1044 		/* previous_current is guaranteed to be valid when the task was
1045 		 * in user space, so ...
1046 		 */
1047 		ia64_mca_modify_comm(previous_current);
1048 		goto no_mod;
1049 	}
1050 
1051 	if (r13 != sos->prev_IA64_KR_CURRENT) {
1052 		msg = "inconsistent previous current and r13";
1053 		goto no_mod;
1054 	}
1055 
1056 	if (!mca_recover_range(ms->pmsa_iip)) {
1057 		if ((r12 - r13) >= KERNEL_STACK_SIZE) {
1058 			msg = "inconsistent r12 and r13";
1059 			goto no_mod;
1060 		}
1061 		if ((ar_bspstore - r13) >= KERNEL_STACK_SIZE) {
1062 			msg = "inconsistent ar.bspstore and r13";
1063 			goto no_mod;
1064 		}
1065 		va.p = old_bspstore;
1066 		if (va.f.reg < 5) {
1067 			msg = "old_bspstore is in the wrong region";
1068 			goto no_mod;
1069 		}
1070 		if ((ar_bsp - r13) >= KERNEL_STACK_SIZE) {
1071 			msg = "inconsistent ar.bsp and r13";
1072 			goto no_mod;
1073 		}
1074 		size += (ia64_rse_skip_regs(old_bspstore, slots) - old_bspstore) * 8;
1075 		if (ar_bspstore + size > r12) {
1076 			msg = "no room for blocked state";
1077 			goto no_mod;
1078 		}
1079 	}
1080 
1081 	ia64_mca_modify_comm(previous_current);
1082 
1083 	/* Make the original task look blocked.  First stack a struct pt_regs,
1084 	 * describing the state at the time of interrupt.  mca_asm.S built a
1085 	 * partial pt_regs, copy it and fill in the blanks using minstate.
1086 	 */
1087 	p = (char *)r12 - sizeof(*regs);
1088 	old_regs = (struct pt_regs *)p;
1089 	memcpy(old_regs, regs, sizeof(*regs));
1090 	old_regs->loadrs = loadrs;
1091 	old_unat = old_regs->ar_unat;
1092 	finish_pt_regs(old_regs, sos, &old_unat);
1093 
1094 	/* Next stack a struct switch_stack.  mca_asm.S built a partial
1095 	 * switch_stack, copy it and fill in the blanks using pt_regs and
1096 	 * minstate.
1097 	 *
1098 	 * In the synthesized switch_stack, b0 points to ia64_leave_kernel,
1099 	 * ar.pfs is set to 0.
1100 	 *
1101 	 * unwind.c::unw_unwind() does special processing for interrupt frames.
1102 	 * It checks if the PRED_NON_SYSCALL predicate is set, if the predicate
1103 	 * is clear then unw_unwind() does _not_ adjust bsp over pt_regs.  Not
1104 	 * that this is documented, of course.  Set PRED_NON_SYSCALL in the
1105 	 * switch_stack on the original stack so it will unwind correctly when
1106 	 * unwind.c reads pt_regs.
1107 	 *
1108 	 * thread.ksp is updated to point to the synthesized switch_stack.
1109 	 */
1110 	p -= sizeof(struct switch_stack);
1111 	old_sw = (struct switch_stack *)p;
1112 	memcpy(old_sw, sw, sizeof(*sw));
1113 	old_sw->caller_unat = old_unat;
1114 	old_sw->ar_fpsr = old_regs->ar_fpsr;
1115 	copy_reg(&ms->pmsa_gr[4-1], ms->pmsa_nat_bits, &old_sw->r4, &old_unat);
1116 	copy_reg(&ms->pmsa_gr[5-1], ms->pmsa_nat_bits, &old_sw->r5, &old_unat);
1117 	copy_reg(&ms->pmsa_gr[6-1], ms->pmsa_nat_bits, &old_sw->r6, &old_unat);
1118 	copy_reg(&ms->pmsa_gr[7-1], ms->pmsa_nat_bits, &old_sw->r7, &old_unat);
1119 	old_sw->b0 = (u64)ia64_leave_kernel;
1120 	old_sw->b1 = ms->pmsa_br1;
1121 	old_sw->ar_pfs = 0;
1122 	old_sw->ar_unat = old_unat;
1123 	old_sw->pr = old_regs->pr | (1UL << PRED_NON_SYSCALL);
1124 	previous_current->thread.ksp = (u64)p - 16;
1125 
1126 	/* Finally copy the original stack's registers back to its RBS.
1127 	 * Registers from ar.bspstore through ar.bsp at the time of the event
1128 	 * are in the current RBS, copy them back to the original stack.  The
1129 	 * copy must be done register by register because the original bspstore
1130 	 * and the current one have different alignments, so the saved RNAT
1131 	 * data occurs at different places.
1132 	 *
1133 	 * mca_asm does cover, so the old_bsp already includes all registers at
1134 	 * the time of MCA/INIT.  It also does flushrs, so all registers before
1135 	 * this function have been written to backing store on the MCA/INIT
1136 	 * stack.
1137 	 */
1138 	new_rnat = ia64_get_rnat(ia64_rse_rnat_addr(new_bspstore));
1139 	old_rnat = regs->ar_rnat;
1140 	while (slots--) {
1141 		if (ia64_rse_is_rnat_slot(new_bspstore)) {
1142 			new_rnat = ia64_get_rnat(new_bspstore++);
1143 		}
1144 		if (ia64_rse_is_rnat_slot(old_bspstore)) {
1145 			*old_bspstore++ = old_rnat;
1146 			old_rnat = 0;
1147 		}
1148 		nat = (new_rnat >> ia64_rse_slot_num(new_bspstore)) & 1UL;
1149 		old_rnat &= ~(1UL << ia64_rse_slot_num(old_bspstore));
1150 		old_rnat |= (nat << ia64_rse_slot_num(old_bspstore));
1151 		*old_bspstore++ = *new_bspstore++;
1152 	}
1153 	old_sw->ar_bspstore = (unsigned long)old_bspstore;
1154 	old_sw->ar_rnat = old_rnat;
1155 
1156 	sos->prev_task = previous_current;
1157 	return previous_current;
1158 
1159 no_mod:
1160 	mprintk(KERN_INFO "cpu %d, %s %s, original stack not modified\n",
1161 			smp_processor_id(), type, msg);
1162 	old_unat = regs->ar_unat;
1163 	finish_pt_regs(regs, sos, &old_unat);
1164 	return previous_current;
1165 }
1166 
1167 /* The monarch/slave interaction is based on monarch_cpu and requires that all
1168  * slaves have entered rendezvous before the monarch leaves.  If any cpu has
1169  * not entered rendezvous yet then wait a bit.  The assumption is that any
1170  * slave that has not rendezvoused after a reasonable time is never going to do
1171  * so.  In this context, slave includes cpus that respond to the MCA rendezvous
1172  * interrupt, as well as cpus that receive the INIT slave event.
1173  */
1174 
1175 static void
1176 ia64_wait_for_slaves(int monarch, const char *type)
1177 {
1178 	int c, i , wait;
1179 
1180 	/*
1181 	 * wait 5 seconds total for slaves (arbitrary)
1182 	 */
1183 	for (i = 0; i < 5000; i++) {
1184 		wait = 0;
1185 		for_each_online_cpu(c) {
1186 			if (c == monarch)
1187 				continue;
1188 			if (ia64_mc_info.imi_rendez_checkin[c]
1189 					== IA64_MCA_RENDEZ_CHECKIN_NOTDONE) {
1190 				udelay(1000);		/* short wait */
1191 				wait = 1;
1192 				break;
1193 			}
1194 		}
1195 		if (!wait)
1196 			goto all_in;
1197 	}
1198 
1199 	/*
1200 	 * Maybe slave(s) dead. Print buffered messages immediately.
1201 	 */
1202 	ia64_mlogbuf_finish(0);
1203 	mprintk(KERN_INFO "OS %s slave did not rendezvous on cpu", type);
1204 	for_each_online_cpu(c) {
1205 		if (c == monarch)
1206 			continue;
1207 		if (ia64_mc_info.imi_rendez_checkin[c] == IA64_MCA_RENDEZ_CHECKIN_NOTDONE)
1208 			mprintk(" %d", c);
1209 	}
1210 	mprintk("\n");
1211 	return;
1212 
1213 all_in:
1214 	mprintk(KERN_INFO "All OS %s slaves have reached rendezvous\n", type);
1215 	return;
1216 }
1217 
1218 /*  mca_insert_tr
1219  *
1220  *  Switch rid when TR reload and needed!
1221  *  iord: 1: itr, 2: itr;
1222  *
1223 */
1224 static void mca_insert_tr(u64 iord)
1225 {
1226 
1227 	int i;
1228 	u64 old_rr;
1229 	struct ia64_tr_entry *p;
1230 	unsigned long psr;
1231 	int cpu = smp_processor_id();
1232 
1233 	if (!ia64_idtrs[cpu])
1234 		return;
1235 
1236 	psr = ia64_clear_ic();
1237 	for (i = IA64_TR_ALLOC_BASE; i < IA64_TR_ALLOC_MAX; i++) {
1238 		p = ia64_idtrs[cpu] + (iord - 1) * IA64_TR_ALLOC_MAX;
1239 		if (p->pte & 0x1) {
1240 			old_rr = ia64_get_rr(p->ifa);
1241 			if (old_rr != p->rr) {
1242 				ia64_set_rr(p->ifa, p->rr);
1243 				ia64_srlz_d();
1244 			}
1245 			ia64_ptr(iord, p->ifa, p->itir >> 2);
1246 			ia64_srlz_i();
1247 			if (iord & 0x1) {
1248 				ia64_itr(0x1, i, p->ifa, p->pte, p->itir >> 2);
1249 				ia64_srlz_i();
1250 			}
1251 			if (iord & 0x2) {
1252 				ia64_itr(0x2, i, p->ifa, p->pte, p->itir >> 2);
1253 				ia64_srlz_i();
1254 			}
1255 			if (old_rr != p->rr) {
1256 				ia64_set_rr(p->ifa, old_rr);
1257 				ia64_srlz_d();
1258 			}
1259 		}
1260 	}
1261 	ia64_set_psr(psr);
1262 }
1263 
1264 /*
1265  * ia64_mca_handler
1266  *
1267  *	This is uncorrectable machine check handler called from OS_MCA
1268  *	dispatch code which is in turn called from SAL_CHECK().
1269  *	This is the place where the core of OS MCA handling is done.
1270  *	Right now the logs are extracted and displayed in a well-defined
1271  *	format. This handler code is supposed to be run only on the
1272  *	monarch processor. Once the monarch is done with MCA handling
1273  *	further MCA logging is enabled by clearing logs.
1274  *	Monarch also has the duty of sending wakeup-IPIs to pull the
1275  *	slave processors out of rendezvous spinloop.
1276  *
1277  *	If multiple processors call into OS_MCA, the first will become
1278  *	the monarch.  Subsequent cpus will be recorded in the mca_cpu
1279  *	bitmask.  After the first monarch has processed its MCA, it
1280  *	will wake up the next cpu in the mca_cpu bitmask and then go
1281  *	into the rendezvous loop.  When all processors have serviced
1282  *	their MCA, the last monarch frees up the rest of the processors.
1283  */
1284 void
1285 ia64_mca_handler(struct pt_regs *regs, struct switch_stack *sw,
1286 		 struct ia64_sal_os_state *sos)
1287 {
1288 	int recover, cpu = smp_processor_id();
1289 	struct task_struct *previous_current;
1290 	struct ia64_mca_notify_die nd =
1291 		{ .sos = sos, .monarch_cpu = &monarch_cpu, .data = &recover };
1292 	static atomic_t mca_count;
1293 	static cpumask_t mca_cpu;
1294 
1295 	if (atomic_add_return(1, &mca_count) == 1) {
1296 		monarch_cpu = cpu;
1297 		sos->monarch = 1;
1298 	} else {
1299 		cpumask_set_cpu(cpu, &mca_cpu);
1300 		sos->monarch = 0;
1301 	}
1302 	mprintk(KERN_INFO "Entered OS MCA handler. PSP=%lx cpu=%d "
1303 		"monarch=%ld\n", sos->proc_state_param, cpu, sos->monarch);
1304 
1305 	previous_current = ia64_mca_modify_original_stack(regs, sw, sos, "MCA");
1306 
1307 	NOTIFY_MCA(DIE_MCA_MONARCH_ENTER, regs, (long)&nd, 1);
1308 
1309 	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_CONCURRENT_MCA;
1310 	if (sos->monarch) {
1311 		ia64_wait_for_slaves(cpu, "MCA");
1312 
1313 		/* Wakeup all the processors which are spinning in the
1314 		 * rendezvous loop.  They will leave SAL, then spin in the OS
1315 		 * with interrupts disabled until this monarch cpu leaves the
1316 		 * MCA handler.  That gets control back to the OS so we can
1317 		 * backtrace the other cpus, backtrace when spinning in SAL
1318 		 * does not work.
1319 		 */
1320 		ia64_mca_wakeup_all();
1321 	} else {
1322 		while (cpumask_test_cpu(cpu, &mca_cpu))
1323 			cpu_relax();	/* spin until monarch wakes us */
1324 	}
1325 
1326 	NOTIFY_MCA(DIE_MCA_MONARCH_PROCESS, regs, (long)&nd, 1);
1327 
1328 	/* Get the MCA error record and log it */
1329 	ia64_mca_log_sal_error_record(SAL_INFO_TYPE_MCA);
1330 
1331 	/* MCA error recovery */
1332 	recover = (ia64_mca_ucmc_extension
1333 		&& ia64_mca_ucmc_extension(
1334 			IA64_LOG_CURR_BUFFER(SAL_INFO_TYPE_MCA),
1335 			sos));
1336 
1337 	if (recover) {
1338 		sal_log_record_header_t *rh = IA64_LOG_CURR_BUFFER(SAL_INFO_TYPE_MCA);
1339 		rh->severity = sal_log_severity_corrected;
1340 		ia64_sal_clear_state_info(SAL_INFO_TYPE_MCA);
1341 		sos->os_status = IA64_MCA_CORRECTED;
1342 	} else {
1343 		/* Dump buffered message to console */
1344 		ia64_mlogbuf_finish(1);
1345 	}
1346 
1347 	if (__this_cpu_read(ia64_mca_tr_reload)) {
1348 		mca_insert_tr(0x1); /*Reload dynamic itrs*/
1349 		mca_insert_tr(0x2); /*Reload dynamic itrs*/
1350 	}
1351 
1352 	NOTIFY_MCA(DIE_MCA_MONARCH_LEAVE, regs, (long)&nd, 1);
1353 
1354 	if (atomic_dec_return(&mca_count) > 0) {
1355 		int i;
1356 
1357 		/* wake up the next monarch cpu,
1358 		 * and put this cpu in the rendez loop.
1359 		 */
1360 		for_each_online_cpu(i) {
1361 			if (cpumask_test_cpu(i, &mca_cpu)) {
1362 				monarch_cpu = i;
1363 				cpumask_clear_cpu(i, &mca_cpu);	/* wake next cpu */
1364 				while (monarch_cpu != -1)
1365 					cpu_relax();	/* spin until last cpu leaves */
1366 				ia64_set_curr_task(cpu, previous_current);
1367 				ia64_mc_info.imi_rendez_checkin[cpu]
1368 						= IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1369 				return;
1370 			}
1371 		}
1372 	}
1373 	ia64_set_curr_task(cpu, previous_current);
1374 	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1375 	monarch_cpu = -1;	/* This frees the slaves and previous monarchs */
1376 }
1377 
1378 static DECLARE_WORK(cmc_disable_work, ia64_mca_cmc_vector_disable_keventd);
1379 static DECLARE_WORK(cmc_enable_work, ia64_mca_cmc_vector_enable_keventd);
1380 
1381 /*
1382  * ia64_mca_cmc_int_handler
1383  *
1384  *  This is corrected machine check interrupt handler.
1385  *	Right now the logs are extracted and displayed in a well-defined
1386  *	format.
1387  *
1388  * Inputs
1389  *      interrupt number
1390  *      client data arg ptr
1391  *
1392  * Outputs
1393  *	None
1394  */
1395 static irqreturn_t
1396 ia64_mca_cmc_int_handler(int cmc_irq, void *arg)
1397 {
1398 	static unsigned long	cmc_history[CMC_HISTORY_LENGTH];
1399 	static int		index;
1400 	static DEFINE_SPINLOCK(cmc_history_lock);
1401 
1402 	IA64_MCA_DEBUG("%s: received interrupt vector = %#x on CPU %d\n",
1403 		       __func__, cmc_irq, smp_processor_id());
1404 
1405 	/* SAL spec states this should run w/ interrupts enabled */
1406 	local_irq_enable();
1407 
1408 	spin_lock(&cmc_history_lock);
1409 	if (!cmc_polling_enabled) {
1410 		int i, count = 1; /* we know 1 happened now */
1411 		unsigned long now = jiffies;
1412 
1413 		for (i = 0; i < CMC_HISTORY_LENGTH; i++) {
1414 			if (now - cmc_history[i] <= HZ)
1415 				count++;
1416 		}
1417 
1418 		IA64_MCA_DEBUG(KERN_INFO "CMC threshold %d/%d\n", count, CMC_HISTORY_LENGTH);
1419 		if (count >= CMC_HISTORY_LENGTH) {
1420 
1421 			cmc_polling_enabled = 1;
1422 			spin_unlock(&cmc_history_lock);
1423 			/* If we're being hit with CMC interrupts, we won't
1424 			 * ever execute the schedule_work() below.  Need to
1425 			 * disable CMC interrupts on this processor now.
1426 			 */
1427 			ia64_mca_cmc_vector_disable(NULL);
1428 			schedule_work(&cmc_disable_work);
1429 
1430 			/*
1431 			 * Corrected errors will still be corrected, but
1432 			 * make sure there's a log somewhere that indicates
1433 			 * something is generating more than we can handle.
1434 			 */
1435 			printk(KERN_WARNING "WARNING: Switching to polling CMC handler; error records may be lost\n");
1436 
1437 			mod_timer(&cmc_poll_timer, jiffies + CMC_POLL_INTERVAL);
1438 
1439 			/* lock already released, get out now */
1440 			goto out;
1441 		} else {
1442 			cmc_history[index++] = now;
1443 			if (index == CMC_HISTORY_LENGTH)
1444 				index = 0;
1445 		}
1446 	}
1447 	spin_unlock(&cmc_history_lock);
1448 out:
1449 	/* Get the CMC error record and log it */
1450 	ia64_mca_log_sal_error_record(SAL_INFO_TYPE_CMC);
1451 
1452 	local_irq_disable();
1453 
1454 	return IRQ_HANDLED;
1455 }
1456 
1457 /*
1458  *  ia64_mca_cmc_int_caller
1459  *
1460  * 	Triggered by sw interrupt from CMC polling routine.  Calls
1461  * 	real interrupt handler and either triggers a sw interrupt
1462  * 	on the next cpu or does cleanup at the end.
1463  *
1464  * Inputs
1465  *	interrupt number
1466  *	client data arg ptr
1467  * Outputs
1468  * 	handled
1469  */
1470 static irqreturn_t
1471 ia64_mca_cmc_int_caller(int cmc_irq, void *arg)
1472 {
1473 	static int start_count = -1;
1474 	unsigned int cpuid;
1475 
1476 	cpuid = smp_processor_id();
1477 
1478 	/* If first cpu, update count */
1479 	if (start_count == -1)
1480 		start_count = IA64_LOG_COUNT(SAL_INFO_TYPE_CMC);
1481 
1482 	ia64_mca_cmc_int_handler(cmc_irq, arg);
1483 
1484 	cpuid = cpumask_next(cpuid+1, cpu_online_mask);
1485 
1486 	if (cpuid < nr_cpu_ids) {
1487 		ia64_send_ipi(cpuid, IA64_CMCP_VECTOR, IA64_IPI_DM_INT, 0);
1488 	} else {
1489 		/* If no log record, switch out of polling mode */
1490 		if (start_count == IA64_LOG_COUNT(SAL_INFO_TYPE_CMC)) {
1491 
1492 			printk(KERN_WARNING "Returning to interrupt driven CMC handler\n");
1493 			schedule_work(&cmc_enable_work);
1494 			cmc_polling_enabled = 0;
1495 
1496 		} else {
1497 
1498 			mod_timer(&cmc_poll_timer, jiffies + CMC_POLL_INTERVAL);
1499 		}
1500 
1501 		start_count = -1;
1502 	}
1503 
1504 	return IRQ_HANDLED;
1505 }
1506 
1507 /*
1508  *  ia64_mca_cmc_poll
1509  *
1510  *	Poll for Corrected Machine Checks (CMCs)
1511  *
1512  * Inputs   :   dummy(unused)
1513  * Outputs  :   None
1514  *
1515  */
1516 static void
1517 ia64_mca_cmc_poll (struct timer_list *unused)
1518 {
1519 	/* Trigger a CMC interrupt cascade  */
1520 	ia64_send_ipi(cpumask_first(cpu_online_mask), IA64_CMCP_VECTOR,
1521 							IA64_IPI_DM_INT, 0);
1522 }
1523 
1524 /*
1525  *  ia64_mca_cpe_int_caller
1526  *
1527  * 	Triggered by sw interrupt from CPE polling routine.  Calls
1528  * 	real interrupt handler and either triggers a sw interrupt
1529  * 	on the next cpu or does cleanup at the end.
1530  *
1531  * Inputs
1532  *	interrupt number
1533  *	client data arg ptr
1534  * Outputs
1535  * 	handled
1536  */
1537 static irqreturn_t
1538 ia64_mca_cpe_int_caller(int cpe_irq, void *arg)
1539 {
1540 	static int start_count = -1;
1541 	static int poll_time = MIN_CPE_POLL_INTERVAL;
1542 	unsigned int cpuid;
1543 
1544 	cpuid = smp_processor_id();
1545 
1546 	/* If first cpu, update count */
1547 	if (start_count == -1)
1548 		start_count = IA64_LOG_COUNT(SAL_INFO_TYPE_CPE);
1549 
1550 	ia64_mca_cpe_int_handler(cpe_irq, arg);
1551 
1552 	cpuid = cpumask_next(cpuid+1, cpu_online_mask);
1553 
1554 	if (cpuid < NR_CPUS) {
1555 		ia64_send_ipi(cpuid, IA64_CPEP_VECTOR, IA64_IPI_DM_INT, 0);
1556 	} else {
1557 		/*
1558 		 * If a log was recorded, increase our polling frequency,
1559 		 * otherwise, backoff or return to interrupt mode.
1560 		 */
1561 		if (start_count != IA64_LOG_COUNT(SAL_INFO_TYPE_CPE)) {
1562 			poll_time = max(MIN_CPE_POLL_INTERVAL, poll_time / 2);
1563 		} else if (cpe_vector < 0) {
1564 			poll_time = min(MAX_CPE_POLL_INTERVAL, poll_time * 2);
1565 		} else {
1566 			poll_time = MIN_CPE_POLL_INTERVAL;
1567 
1568 			printk(KERN_WARNING "Returning to interrupt driven CPE handler\n");
1569 			enable_irq(local_vector_to_irq(IA64_CPE_VECTOR));
1570 			cpe_poll_enabled = 0;
1571 		}
1572 
1573 		if (cpe_poll_enabled)
1574 			mod_timer(&cpe_poll_timer, jiffies + poll_time);
1575 		start_count = -1;
1576 	}
1577 
1578 	return IRQ_HANDLED;
1579 }
1580 
1581 /*
1582  *  ia64_mca_cpe_poll
1583  *
1584  *	Poll for Corrected Platform Errors (CPEs), trigger interrupt
1585  *	on first cpu, from there it will trickle through all the cpus.
1586  *
1587  * Inputs   :   dummy(unused)
1588  * Outputs  :   None
1589  *
1590  */
1591 static void
1592 ia64_mca_cpe_poll (struct timer_list *unused)
1593 {
1594 	/* Trigger a CPE interrupt cascade  */
1595 	ia64_send_ipi(cpumask_first(cpu_online_mask), IA64_CPEP_VECTOR,
1596 							IA64_IPI_DM_INT, 0);
1597 }
1598 
1599 static int
1600 default_monarch_init_process(struct notifier_block *self, unsigned long val, void *data)
1601 {
1602 	int c;
1603 	struct task_struct *g, *t;
1604 	if (val != DIE_INIT_MONARCH_PROCESS)
1605 		return NOTIFY_DONE;
1606 #ifdef CONFIG_KEXEC
1607 	if (atomic_read(&kdump_in_progress))
1608 		return NOTIFY_DONE;
1609 #endif
1610 
1611 	/*
1612 	 * FIXME: mlogbuf will brim over with INIT stack dumps.
1613 	 * To enable show_stack from INIT, we use oops_in_progress which should
1614 	 * be used in real oops. This would cause something wrong after INIT.
1615 	 */
1616 	BREAK_LOGLEVEL(console_loglevel);
1617 	ia64_mlogbuf_dump_from_init();
1618 
1619 	printk(KERN_ERR "Processes interrupted by INIT -");
1620 	for_each_online_cpu(c) {
1621 		struct ia64_sal_os_state *s;
1622 		t = __va(__per_cpu_mca[c] + IA64_MCA_CPU_INIT_STACK_OFFSET);
1623 		s = (struct ia64_sal_os_state *)((char *)t + MCA_SOS_OFFSET);
1624 		g = s->prev_task;
1625 		if (g) {
1626 			if (g->pid)
1627 				printk(" %d", g->pid);
1628 			else
1629 				printk(" %d (cpu %d task 0x%p)", g->pid, task_cpu(g), g);
1630 		}
1631 	}
1632 	printk("\n\n");
1633 	if (read_trylock(&tasklist_lock)) {
1634 		do_each_thread (g, t) {
1635 			printk("\nBacktrace of pid %d (%s)\n", t->pid, t->comm);
1636 			show_stack(t, NULL, KERN_DEFAULT);
1637 		} while_each_thread (g, t);
1638 		read_unlock(&tasklist_lock);
1639 	}
1640 	/* FIXME: This will not restore zapped printk locks. */
1641 	RESTORE_LOGLEVEL(console_loglevel);
1642 	return NOTIFY_DONE;
1643 }
1644 
1645 /*
1646  * C portion of the OS INIT handler
1647  *
1648  * Called from ia64_os_init_dispatch
1649  *
1650  * Inputs: pointer to pt_regs where processor info was saved.  SAL/OS state for
1651  * this event.  This code is used for both monarch and slave INIT events, see
1652  * sos->monarch.
1653  *
1654  * All INIT events switch to the INIT stack and change the previous process to
1655  * blocked status.  If one of the INIT events is the monarch then we are
1656  * probably processing the nmi button/command.  Use the monarch cpu to dump all
1657  * the processes.  The slave INIT events all spin until the monarch cpu
1658  * returns.  We can also get INIT slave events for MCA, in which case the MCA
1659  * process is the monarch.
1660  */
1661 
1662 void
1663 ia64_init_handler(struct pt_regs *regs, struct switch_stack *sw,
1664 		  struct ia64_sal_os_state *sos)
1665 {
1666 	static atomic_t slaves;
1667 	static atomic_t monarchs;
1668 	struct task_struct *previous_current;
1669 	int cpu = smp_processor_id();
1670 	struct ia64_mca_notify_die nd =
1671 		{ .sos = sos, .monarch_cpu = &monarch_cpu };
1672 
1673 	NOTIFY_INIT(DIE_INIT_ENTER, regs, (long)&nd, 0);
1674 
1675 	mprintk(KERN_INFO "Entered OS INIT handler. PSP=%lx cpu=%d monarch=%ld\n",
1676 		sos->proc_state_param, cpu, sos->monarch);
1677 	salinfo_log_wakeup(SAL_INFO_TYPE_INIT, NULL, 0, 0);
1678 
1679 	previous_current = ia64_mca_modify_original_stack(regs, sw, sos, "INIT");
1680 	sos->os_status = IA64_INIT_RESUME;
1681 
1682 	/* FIXME: Workaround for broken proms that drive all INIT events as
1683 	 * slaves.  The last slave that enters is promoted to be a monarch.
1684 	 * Remove this code in September 2006, that gives platforms a year to
1685 	 * fix their proms and get their customers updated.
1686 	 */
1687 	if (!sos->monarch && atomic_add_return(1, &slaves) == num_online_cpus()) {
1688 		mprintk(KERN_WARNING "%s: Promoting cpu %d to monarch.\n",
1689 		        __func__, cpu);
1690 		atomic_dec(&slaves);
1691 		sos->monarch = 1;
1692 	}
1693 
1694 	/* FIXME: Workaround for broken proms that drive all INIT events as
1695 	 * monarchs.  Second and subsequent monarchs are demoted to slaves.
1696 	 * Remove this code in September 2006, that gives platforms a year to
1697 	 * fix their proms and get their customers updated.
1698 	 */
1699 	if (sos->monarch && atomic_add_return(1, &monarchs) > 1) {
1700 		mprintk(KERN_WARNING "%s: Demoting cpu %d to slave.\n",
1701 			       __func__, cpu);
1702 		atomic_dec(&monarchs);
1703 		sos->monarch = 0;
1704 	}
1705 
1706 	if (!sos->monarch) {
1707 		ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_INIT;
1708 
1709 #ifdef CONFIG_KEXEC
1710 		while (monarch_cpu == -1 && !atomic_read(&kdump_in_progress))
1711 			udelay(1000);
1712 #else
1713 		while (monarch_cpu == -1)
1714 			cpu_relax();	/* spin until monarch enters */
1715 #endif
1716 
1717 		NOTIFY_INIT(DIE_INIT_SLAVE_ENTER, regs, (long)&nd, 1);
1718 		NOTIFY_INIT(DIE_INIT_SLAVE_PROCESS, regs, (long)&nd, 1);
1719 
1720 #ifdef CONFIG_KEXEC
1721 		while (monarch_cpu != -1 && !atomic_read(&kdump_in_progress))
1722 			udelay(1000);
1723 #else
1724 		while (monarch_cpu != -1)
1725 			cpu_relax();	/* spin until monarch leaves */
1726 #endif
1727 
1728 		NOTIFY_INIT(DIE_INIT_SLAVE_LEAVE, regs, (long)&nd, 1);
1729 
1730 		mprintk("Slave on cpu %d returning to normal service.\n", cpu);
1731 		ia64_set_curr_task(cpu, previous_current);
1732 		ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1733 		atomic_dec(&slaves);
1734 		return;
1735 	}
1736 
1737 	monarch_cpu = cpu;
1738 	NOTIFY_INIT(DIE_INIT_MONARCH_ENTER, regs, (long)&nd, 1);
1739 
1740 	/*
1741 	 * Wait for a bit.  On some machines (e.g., HP's zx2000 and zx6000, INIT can be
1742 	 * generated via the BMC's command-line interface, but since the console is on the
1743 	 * same serial line, the user will need some time to switch out of the BMC before
1744 	 * the dump begins.
1745 	 */
1746 	mprintk("Delaying for 5 seconds...\n");
1747 	udelay(5*1000000);
1748 	ia64_wait_for_slaves(cpu, "INIT");
1749 	/* If nobody intercepts DIE_INIT_MONARCH_PROCESS then we drop through
1750 	 * to default_monarch_init_process() above and just print all the
1751 	 * tasks.
1752 	 */
1753 	NOTIFY_INIT(DIE_INIT_MONARCH_PROCESS, regs, (long)&nd, 1);
1754 	NOTIFY_INIT(DIE_INIT_MONARCH_LEAVE, regs, (long)&nd, 1);
1755 
1756 	mprintk("\nINIT dump complete.  Monarch on cpu %d returning to normal service.\n", cpu);
1757 	atomic_dec(&monarchs);
1758 	ia64_set_curr_task(cpu, previous_current);
1759 	monarch_cpu = -1;
1760 	return;
1761 }
1762 
1763 static int __init
1764 ia64_mca_disable_cpe_polling(char *str)
1765 {
1766 	cpe_poll_enabled = 0;
1767 	return 1;
1768 }
1769 
1770 __setup("disable_cpe_poll", ia64_mca_disable_cpe_polling);
1771 
1772 /* Minimal format of the MCA/INIT stacks.  The pseudo processes that run on
1773  * these stacks can never sleep, they cannot return from the kernel to user
1774  * space, they do not appear in a normal ps listing.  So there is no need to
1775  * format most of the fields.
1776  */
1777 
1778 static void
1779 format_mca_init_stack(void *mca_data, unsigned long offset,
1780 		const char *type, int cpu)
1781 {
1782 	struct task_struct *p = (struct task_struct *)((char *)mca_data + offset);
1783 	struct thread_info *ti;
1784 	memset(p, 0, KERNEL_STACK_SIZE);
1785 	ti = task_thread_info(p);
1786 	ti->flags = _TIF_MCA_INIT;
1787 	ti->preempt_count = 1;
1788 	ti->task = p;
1789 	ti->cpu = cpu;
1790 	p->stack = ti;
1791 	p->__state = TASK_UNINTERRUPTIBLE;
1792 	cpumask_set_cpu(cpu, &p->cpus_mask);
1793 	INIT_LIST_HEAD(&p->tasks);
1794 	p->parent = p->real_parent = p->group_leader = p;
1795 	INIT_LIST_HEAD(&p->children);
1796 	INIT_LIST_HEAD(&p->sibling);
1797 	strncpy(p->comm, type, sizeof(p->comm)-1);
1798 }
1799 
1800 /* Caller prevents this from being called after init */
1801 static void * __ref mca_bootmem(void)
1802 {
1803 	return memblock_alloc(sizeof(struct ia64_mca_cpu), KERNEL_STACK_SIZE);
1804 }
1805 
1806 /* Do per-CPU MCA-related initialization.  */
1807 void
1808 ia64_mca_cpu_init(void *cpu_data)
1809 {
1810 	void *pal_vaddr;
1811 	void *data;
1812 	long sz = sizeof(struct ia64_mca_cpu);
1813 	int cpu = smp_processor_id();
1814 	static int first_time = 1;
1815 
1816 	/*
1817 	 * Structure will already be allocated if cpu has been online,
1818 	 * then offlined.
1819 	 */
1820 	if (__per_cpu_mca[cpu]) {
1821 		data = __va(__per_cpu_mca[cpu]);
1822 	} else {
1823 		if (first_time) {
1824 			data = mca_bootmem();
1825 			first_time = 0;
1826 		} else
1827 			data = (void *)__get_free_pages(GFP_ATOMIC,
1828 							get_order(sz));
1829 		if (!data)
1830 			panic("Could not allocate MCA memory for cpu %d\n",
1831 					cpu);
1832 	}
1833 	format_mca_init_stack(data, offsetof(struct ia64_mca_cpu, mca_stack),
1834 		"MCA", cpu);
1835 	format_mca_init_stack(data, offsetof(struct ia64_mca_cpu, init_stack),
1836 		"INIT", cpu);
1837 	__this_cpu_write(ia64_mca_data, (__per_cpu_mca[cpu] = __pa(data)));
1838 
1839 	/*
1840 	 * Stash away a copy of the PTE needed to map the per-CPU page.
1841 	 * We may need it during MCA recovery.
1842 	 */
1843 	__this_cpu_write(ia64_mca_per_cpu_pte,
1844 		pte_val(mk_pte_phys(__pa(cpu_data), PAGE_KERNEL)));
1845 
1846 	/*
1847 	 * Also, stash away a copy of the PAL address and the PTE
1848 	 * needed to map it.
1849 	 */
1850 	pal_vaddr = efi_get_pal_addr();
1851 	if (!pal_vaddr)
1852 		return;
1853 	__this_cpu_write(ia64_mca_pal_base,
1854 		GRANULEROUNDDOWN((unsigned long) pal_vaddr));
1855 	__this_cpu_write(ia64_mca_pal_pte, pte_val(mk_pte_phys(__pa(pal_vaddr),
1856 							      PAGE_KERNEL)));
1857 }
1858 
1859 static int ia64_mca_cpu_online(unsigned int cpu)
1860 {
1861 	unsigned long flags;
1862 
1863 	local_irq_save(flags);
1864 	if (!cmc_polling_enabled)
1865 		ia64_mca_cmc_vector_enable(NULL);
1866 	local_irq_restore(flags);
1867 	return 0;
1868 }
1869 
1870 /*
1871  * ia64_mca_init
1872  *
1873  *  Do all the system level mca specific initialization.
1874  *
1875  *	1. Register spinloop and wakeup request interrupt vectors
1876  *
1877  *	2. Register OS_MCA handler entry point
1878  *
1879  *	3. Register OS_INIT handler entry point
1880  *
1881  *  4. Initialize MCA/CMC/INIT related log buffers maintained by the OS.
1882  *
1883  *  Note that this initialization is done very early before some kernel
1884  *  services are available.
1885  *
1886  *  Inputs  :   None
1887  *
1888  *  Outputs :   None
1889  */
1890 void __init
1891 ia64_mca_init(void)
1892 {
1893 	ia64_fptr_t *init_hldlr_ptr_monarch = (ia64_fptr_t *)ia64_os_init_dispatch_monarch;
1894 	ia64_fptr_t *init_hldlr_ptr_slave = (ia64_fptr_t *)ia64_os_init_dispatch_slave;
1895 	ia64_fptr_t *mca_hldlr_ptr = (ia64_fptr_t *)ia64_os_mca_dispatch;
1896 	int i;
1897 	long rc;
1898 	struct ia64_sal_retval isrv;
1899 	unsigned long timeout = IA64_MCA_RENDEZ_TIMEOUT; /* platform specific */
1900 	static struct notifier_block default_init_monarch_nb = {
1901 		.notifier_call = default_monarch_init_process,
1902 		.priority = 0/* we need to notified last */
1903 	};
1904 
1905 	IA64_MCA_DEBUG("%s: begin\n", __func__);
1906 
1907 	/* Clear the Rendez checkin flag for all cpus */
1908 	for(i = 0 ; i < NR_CPUS; i++)
1909 		ia64_mc_info.imi_rendez_checkin[i] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1910 
1911 	/*
1912 	 * Register the rendezvous spinloop and wakeup mechanism with SAL
1913 	 */
1914 
1915 	/* Register the rendezvous interrupt vector with SAL */
1916 	while (1) {
1917 		isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_RENDEZ_INT,
1918 					      SAL_MC_PARAM_MECHANISM_INT,
1919 					      IA64_MCA_RENDEZ_VECTOR,
1920 					      timeout,
1921 					      SAL_MC_PARAM_RZ_ALWAYS);
1922 		rc = isrv.status;
1923 		if (rc == 0)
1924 			break;
1925 		if (rc == -2) {
1926 			printk(KERN_INFO "Increasing MCA rendezvous timeout from "
1927 				"%ld to %ld milliseconds\n", timeout, isrv.v0);
1928 			timeout = isrv.v0;
1929 			NOTIFY_MCA(DIE_MCA_NEW_TIMEOUT, NULL, timeout, 0);
1930 			continue;
1931 		}
1932 		printk(KERN_ERR "Failed to register rendezvous interrupt "
1933 		       "with SAL (status %ld)\n", rc);
1934 		return;
1935 	}
1936 
1937 	/* Register the wakeup interrupt vector with SAL */
1938 	isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_RENDEZ_WAKEUP,
1939 				      SAL_MC_PARAM_MECHANISM_INT,
1940 				      IA64_MCA_WAKEUP_VECTOR,
1941 				      0, 0);
1942 	rc = isrv.status;
1943 	if (rc) {
1944 		printk(KERN_ERR "Failed to register wakeup interrupt with SAL "
1945 		       "(status %ld)\n", rc);
1946 		return;
1947 	}
1948 
1949 	IA64_MCA_DEBUG("%s: registered MCA rendezvous spinloop and wakeup mech.\n", __func__);
1950 
1951 	ia64_mc_info.imi_mca_handler        = ia64_tpa(mca_hldlr_ptr->fp);
1952 	/*
1953 	 * XXX - disable SAL checksum by setting size to 0; should be
1954 	 *	ia64_tpa(ia64_os_mca_dispatch_end) - ia64_tpa(ia64_os_mca_dispatch);
1955 	 */
1956 	ia64_mc_info.imi_mca_handler_size	= 0;
1957 
1958 	/* Register the os mca handler with SAL */
1959 	if ((rc = ia64_sal_set_vectors(SAL_VECTOR_OS_MCA,
1960 				       ia64_mc_info.imi_mca_handler,
1961 				       ia64_tpa(mca_hldlr_ptr->gp),
1962 				       ia64_mc_info.imi_mca_handler_size,
1963 				       0, 0, 0)))
1964 	{
1965 		printk(KERN_ERR "Failed to register OS MCA handler with SAL "
1966 		       "(status %ld)\n", rc);
1967 		return;
1968 	}
1969 
1970 	IA64_MCA_DEBUG("%s: registered OS MCA handler with SAL at 0x%lx, gp = 0x%lx\n", __func__,
1971 		       ia64_mc_info.imi_mca_handler, ia64_tpa(mca_hldlr_ptr->gp));
1972 
1973 	/*
1974 	 * XXX - disable SAL checksum by setting size to 0, should be
1975 	 * size of the actual init handler in mca_asm.S.
1976 	 */
1977 	ia64_mc_info.imi_monarch_init_handler		= ia64_tpa(init_hldlr_ptr_monarch->fp);
1978 	ia64_mc_info.imi_monarch_init_handler_size	= 0;
1979 	ia64_mc_info.imi_slave_init_handler		= ia64_tpa(init_hldlr_ptr_slave->fp);
1980 	ia64_mc_info.imi_slave_init_handler_size	= 0;
1981 
1982 	IA64_MCA_DEBUG("%s: OS INIT handler at %lx\n", __func__,
1983 		       ia64_mc_info.imi_monarch_init_handler);
1984 
1985 	/* Register the os init handler with SAL */
1986 	if ((rc = ia64_sal_set_vectors(SAL_VECTOR_OS_INIT,
1987 				       ia64_mc_info.imi_monarch_init_handler,
1988 				       ia64_tpa(ia64_getreg(_IA64_REG_GP)),
1989 				       ia64_mc_info.imi_monarch_init_handler_size,
1990 				       ia64_mc_info.imi_slave_init_handler,
1991 				       ia64_tpa(ia64_getreg(_IA64_REG_GP)),
1992 				       ia64_mc_info.imi_slave_init_handler_size)))
1993 	{
1994 		printk(KERN_ERR "Failed to register m/s INIT handlers with SAL "
1995 		       "(status %ld)\n", rc);
1996 		return;
1997 	}
1998 	if (register_die_notifier(&default_init_monarch_nb)) {
1999 		printk(KERN_ERR "Failed to register default monarch INIT process\n");
2000 		return;
2001 	}
2002 
2003 	IA64_MCA_DEBUG("%s: registered OS INIT handler with SAL\n", __func__);
2004 
2005 	/* Initialize the areas set aside by the OS to buffer the
2006 	 * platform/processor error states for MCA/INIT/CMC
2007 	 * handling.
2008 	 */
2009 	ia64_log_init(SAL_INFO_TYPE_MCA);
2010 	ia64_log_init(SAL_INFO_TYPE_INIT);
2011 	ia64_log_init(SAL_INFO_TYPE_CMC);
2012 	ia64_log_init(SAL_INFO_TYPE_CPE);
2013 
2014 	mca_init = 1;
2015 	printk(KERN_INFO "MCA related initialization done\n");
2016 }
2017 
2018 
2019 /*
2020  * These pieces cannot be done in ia64_mca_init() because it is called before
2021  * early_irq_init() which would wipe out our percpu irq registrations. But we
2022  * cannot leave them until ia64_mca_late_init() because by then all the other
2023  * processors have been brought online and have set their own CMC vectors to
2024  * point at a non-existant action. Called from arch_early_irq_init().
2025  */
2026 void __init ia64_mca_irq_init(void)
2027 {
2028 	/*
2029 	 *  Configure the CMCI/P vector and handler. Interrupts for CMC are
2030 	 *  per-processor, so AP CMC interrupts are setup in smp_callin() (smpboot.c).
2031 	 */
2032 	register_percpu_irq(IA64_CMC_VECTOR, ia64_mca_cmc_int_handler, 0,
2033 			    "cmc_hndlr");
2034 	register_percpu_irq(IA64_CMCP_VECTOR, ia64_mca_cmc_int_caller, 0,
2035 			    "cmc_poll");
2036 	ia64_mca_cmc_vector_setup();       /* Setup vector on BSP */
2037 
2038 	/* Setup the MCA rendezvous interrupt vector */
2039 	register_percpu_irq(IA64_MCA_RENDEZ_VECTOR, ia64_mca_rendez_int_handler,
2040 			    0, "mca_rdzv");
2041 
2042 	/* Setup the MCA wakeup interrupt vector */
2043 	register_percpu_irq(IA64_MCA_WAKEUP_VECTOR, ia64_mca_wakeup_int_handler,
2044 			    0, "mca_wkup");
2045 
2046 	/* Setup the CPEI/P handler */
2047 	register_percpu_irq(IA64_CPEP_VECTOR, ia64_mca_cpe_int_caller, 0,
2048 			    "cpe_poll");
2049 }
2050 
2051 /*
2052  * ia64_mca_late_init
2053  *
2054  *	Opportunity to setup things that require initialization later
2055  *	than ia64_mca_init.  Setup a timer to poll for CPEs if the
2056  *	platform doesn't support an interrupt driven mechanism.
2057  *
2058  *  Inputs  :   None
2059  *  Outputs :   Status
2060  */
2061 static int __init
2062 ia64_mca_late_init(void)
2063 {
2064 	if (!mca_init)
2065 		return 0;
2066 
2067 	/* Setup the CMCI/P vector and handler */
2068 	timer_setup(&cmc_poll_timer, ia64_mca_cmc_poll, 0);
2069 
2070 	/* Unmask/enable the vector */
2071 	cmc_polling_enabled = 0;
2072 	cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "ia64/mca:online",
2073 			  ia64_mca_cpu_online, NULL);
2074 	IA64_MCA_DEBUG("%s: CMCI/P setup and enabled.\n", __func__);
2075 
2076 	/* Setup the CPEI/P vector and handler */
2077 	cpe_vector = acpi_request_vector(ACPI_INTERRUPT_CPEI);
2078 	timer_setup(&cpe_poll_timer, ia64_mca_cpe_poll, 0);
2079 
2080 	{
2081 		unsigned int irq;
2082 
2083 		if (cpe_vector >= 0) {
2084 			/* If platform supports CPEI, enable the irq. */
2085 			irq = local_vector_to_irq(cpe_vector);
2086 			if (irq > 0) {
2087 				cpe_poll_enabled = 0;
2088 				irq_set_status_flags(irq, IRQ_PER_CPU);
2089 				if (request_irq(irq, ia64_mca_cpe_int_handler,
2090 						0, "cpe_hndlr", NULL))
2091 					pr_err("Failed to register cpe_hndlr interrupt\n");
2092 				ia64_cpe_irq = irq;
2093 				ia64_mca_register_cpev(cpe_vector);
2094 				IA64_MCA_DEBUG("%s: CPEI/P setup and enabled.\n",
2095 					__func__);
2096 				return 0;
2097 			}
2098 			printk(KERN_ERR "%s: Failed to find irq for CPE "
2099 					"interrupt handler, vector %d\n",
2100 					__func__, cpe_vector);
2101 		}
2102 		/* If platform doesn't support CPEI, get the timer going. */
2103 		if (cpe_poll_enabled) {
2104 			ia64_mca_cpe_poll(0UL);
2105 			IA64_MCA_DEBUG("%s: CPEP setup and enabled.\n", __func__);
2106 		}
2107 	}
2108 
2109 	return 0;
2110 }
2111 
2112 device_initcall(ia64_mca_late_init);
2113