xref: /openbmc/linux/arch/ia64/kernel/mca.c (revision 64c70b1c)
1 /*
2  * File:	mca.c
3  * Purpose:	Generic MCA handling layer
4  *
5  * Updated for latest kernel
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 Silicon Graphics, Inc.
22  * Copyright (C) Vijay Chander(vijay@engr.sgi.com)
23  *
24  * 03/04/15 D. Mosberger Added INIT backtrace support.
25  * 02/03/25 M. Domsch	GUID cleanups
26  *
27  * 02/01/04 J. Hall	Aligned MCA stack to 16 bytes, added platform vs. CPU
28  *			error flag, set SAL default return values, changed
29  *			error record structure to linked list, added init call
30  *			to sal_get_state_info_size().
31  *
32  * 01/01/03 F. Lewis    Added setup of CMCI and CPEI IRQs, logging of corrected
33  *                      platform errors, completed code for logging of
34  *                      corrected & uncorrected machine check errors, and
35  *                      updated for conformance with Nov. 2000 revision of the
36  *                      SAL 3.0 spec.
37  * 00/03/29 C. Fleckenstein  Fixed PAL/SAL update issues, began MCA bug fixes, logging issues,
38  *                           added min save state dump, added INIT handler.
39  *
40  * 2003-12-08 Keith Owens <kaos@sgi.com>
41  *            smp_call_function() must not be called from interrupt context (can
42  *            deadlock on tasklist_lock).  Use keventd to call smp_call_function().
43  *
44  * 2004-02-01 Keith Owens <kaos@sgi.com>
45  *            Avoid deadlock when using printk() for MCA and INIT records.
46  *            Delete all record printing code, moved to salinfo_decode in user space.
47  *            Mark variables and functions static where possible.
48  *            Delete dead variables and functions.
49  *            Reorder to remove the need for forward declarations and to consolidate
50  *            related code.
51  *
52  * 2005-08-12 Keith Owens <kaos@sgi.com>
53  *	      Convert MCA/INIT handlers to use per event stacks and SAL/OS state.
54  *
55  * 2005-10-07 Keith Owens <kaos@sgi.com>
56  *	      Add notify_die() hooks.
57  *
58  * 2006-09-15 Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
59  * 	      Add printing support for MCA/INIT.
60  */
61 #include <linux/types.h>
62 #include <linux/init.h>
63 #include <linux/sched.h>
64 #include <linux/interrupt.h>
65 #include <linux/irq.h>
66 #include <linux/bootmem.h>
67 #include <linux/acpi.h>
68 #include <linux/timer.h>
69 #include <linux/module.h>
70 #include <linux/kernel.h>
71 #include <linux/smp.h>
72 #include <linux/workqueue.h>
73 #include <linux/cpumask.h>
74 #include <linux/kdebug.h>
75 
76 #include <asm/delay.h>
77 #include <asm/machvec.h>
78 #include <asm/meminit.h>
79 #include <asm/page.h>
80 #include <asm/ptrace.h>
81 #include <asm/system.h>
82 #include <asm/sal.h>
83 #include <asm/mca.h>
84 #include <asm/kexec.h>
85 
86 #include <asm/irq.h>
87 #include <asm/hw_irq.h>
88 
89 #include "mca_drv.h"
90 #include "entry.h"
91 
92 #if defined(IA64_MCA_DEBUG_INFO)
93 # define IA64_MCA_DEBUG(fmt...)	printk(fmt)
94 #else
95 # define IA64_MCA_DEBUG(fmt...)
96 #endif
97 
98 /* Used by mca_asm.S */
99 u32				ia64_mca_serialize;
100 DEFINE_PER_CPU(u64, ia64_mca_data); /* == __per_cpu_mca[smp_processor_id()] */
101 DEFINE_PER_CPU(u64, ia64_mca_per_cpu_pte); /* PTE to map per-CPU area */
102 DEFINE_PER_CPU(u64, ia64_mca_pal_pte);	    /* PTE to map PAL code */
103 DEFINE_PER_CPU(u64, ia64_mca_pal_base);    /* vaddr PAL code granule */
104 
105 unsigned long __per_cpu_mca[NR_CPUS];
106 
107 /* In mca_asm.S */
108 extern void			ia64_os_init_dispatch_monarch (void);
109 extern void			ia64_os_init_dispatch_slave (void);
110 
111 static int monarch_cpu = -1;
112 
113 static ia64_mc_info_t		ia64_mc_info;
114 
115 #define MAX_CPE_POLL_INTERVAL (15*60*HZ) /* 15 minutes */
116 #define MIN_CPE_POLL_INTERVAL (2*60*HZ)  /* 2 minutes */
117 #define CMC_POLL_INTERVAL     (1*60*HZ)  /* 1 minute */
118 #define CPE_HISTORY_LENGTH    5
119 #define CMC_HISTORY_LENGTH    5
120 
121 #ifdef CONFIG_ACPI
122 static struct timer_list cpe_poll_timer;
123 #endif
124 static struct timer_list cmc_poll_timer;
125 /*
126  * This variable tells whether we are currently in polling mode.
127  * Start with this in the wrong state so we won't play w/ timers
128  * before the system is ready.
129  */
130 static int cmc_polling_enabled = 1;
131 
132 /*
133  * Clearing this variable prevents CPE polling from getting activated
134  * in mca_late_init.  Use it if your system doesn't provide a CPEI,
135  * but encounters problems retrieving CPE logs.  This should only be
136  * necessary for debugging.
137  */
138 static int cpe_poll_enabled = 1;
139 
140 extern void salinfo_log_wakeup(int type, u8 *buffer, u64 size, int irqsafe);
141 
142 static int mca_init __initdata;
143 
144 /*
145  * limited & delayed printing support for MCA/INIT handler
146  */
147 
148 #define mprintk(fmt...) ia64_mca_printk(fmt)
149 
150 #define MLOGBUF_SIZE (512+256*NR_CPUS)
151 #define MLOGBUF_MSGMAX 256
152 static char mlogbuf[MLOGBUF_SIZE];
153 static DEFINE_SPINLOCK(mlogbuf_wlock);	/* mca context only */
154 static DEFINE_SPINLOCK(mlogbuf_rlock);	/* normal context only */
155 static unsigned long mlogbuf_start;
156 static unsigned long mlogbuf_end;
157 static unsigned int mlogbuf_finished = 0;
158 static unsigned long mlogbuf_timestamp = 0;
159 
160 static int loglevel_save = -1;
161 #define BREAK_LOGLEVEL(__console_loglevel)		\
162 	oops_in_progress = 1;				\
163 	if (loglevel_save < 0)				\
164 		loglevel_save = __console_loglevel;	\
165 	__console_loglevel = 15;
166 
167 #define RESTORE_LOGLEVEL(__console_loglevel)		\
168 	if (loglevel_save >= 0) {			\
169 		__console_loglevel = loglevel_save;	\
170 		loglevel_save = -1;			\
171 	}						\
172 	mlogbuf_finished = 0;				\
173 	oops_in_progress = 0;
174 
175 /*
176  * Push messages into buffer, print them later if not urgent.
177  */
178 void ia64_mca_printk(const char *fmt, ...)
179 {
180 	va_list args;
181 	int printed_len;
182 	char temp_buf[MLOGBUF_MSGMAX];
183 	char *p;
184 
185 	va_start(args, fmt);
186 	printed_len = vscnprintf(temp_buf, sizeof(temp_buf), fmt, args);
187 	va_end(args);
188 
189 	/* Copy the output into mlogbuf */
190 	if (oops_in_progress) {
191 		/* mlogbuf was abandoned, use printk directly instead. */
192 		printk(temp_buf);
193 	} else {
194 		spin_lock(&mlogbuf_wlock);
195 		for (p = temp_buf; *p; p++) {
196 			unsigned long next = (mlogbuf_end + 1) % MLOGBUF_SIZE;
197 			if (next != mlogbuf_start) {
198 				mlogbuf[mlogbuf_end] = *p;
199 				mlogbuf_end = next;
200 			} else {
201 				/* buffer full */
202 				break;
203 			}
204 		}
205 		mlogbuf[mlogbuf_end] = '\0';
206 		spin_unlock(&mlogbuf_wlock);
207 	}
208 }
209 EXPORT_SYMBOL(ia64_mca_printk);
210 
211 /*
212  * Print buffered messages.
213  *  NOTE: call this after returning normal context. (ex. from salinfod)
214  */
215 void ia64_mlogbuf_dump(void)
216 {
217 	char temp_buf[MLOGBUF_MSGMAX];
218 	char *p;
219 	unsigned long index;
220 	unsigned long flags;
221 	unsigned int printed_len;
222 
223 	/* Get output from mlogbuf */
224 	while (mlogbuf_start != mlogbuf_end) {
225 		temp_buf[0] = '\0';
226 		p = temp_buf;
227 		printed_len = 0;
228 
229 		spin_lock_irqsave(&mlogbuf_rlock, flags);
230 
231 		index = mlogbuf_start;
232 		while (index != mlogbuf_end) {
233 			*p = mlogbuf[index];
234 			index = (index + 1) % MLOGBUF_SIZE;
235 			if (!*p)
236 				break;
237 			p++;
238 			if (++printed_len >= MLOGBUF_MSGMAX - 1)
239 				break;
240 		}
241 		*p = '\0';
242 		if (temp_buf[0])
243 			printk(temp_buf);
244 		mlogbuf_start = index;
245 
246 		mlogbuf_timestamp = 0;
247 		spin_unlock_irqrestore(&mlogbuf_rlock, flags);
248 	}
249 }
250 EXPORT_SYMBOL(ia64_mlogbuf_dump);
251 
252 /*
253  * Call this if system is going to down or if immediate flushing messages to
254  * console is required. (ex. recovery was failed, crash dump is going to be
255  * invoked, long-wait rendezvous etc.)
256  *  NOTE: this should be called from monarch.
257  */
258 static void ia64_mlogbuf_finish(int wait)
259 {
260 	BREAK_LOGLEVEL(console_loglevel);
261 
262 	spin_lock_init(&mlogbuf_rlock);
263 	ia64_mlogbuf_dump();
264 	printk(KERN_EMERG "mlogbuf_finish: printing switched to urgent mode, "
265 		"MCA/INIT might be dodgy or fail.\n");
266 
267 	if (!wait)
268 		return;
269 
270 	/* wait for console */
271 	printk("Delaying for 5 seconds...\n");
272 	udelay(5*1000000);
273 
274 	mlogbuf_finished = 1;
275 }
276 
277 /*
278  * Print buffered messages from INIT context.
279  */
280 static void ia64_mlogbuf_dump_from_init(void)
281 {
282 	if (mlogbuf_finished)
283 		return;
284 
285 	if (mlogbuf_timestamp && (mlogbuf_timestamp + 30*HZ > jiffies)) {
286 		printk(KERN_ERR "INIT: mlogbuf_dump is interrupted by INIT "
287 			" and the system seems to be messed up.\n");
288 		ia64_mlogbuf_finish(0);
289 		return;
290 	}
291 
292 	if (!spin_trylock(&mlogbuf_rlock)) {
293 		printk(KERN_ERR "INIT: mlogbuf_dump is interrupted by INIT. "
294 			"Generated messages other than stack dump will be "
295 			"buffered to mlogbuf and will be printed later.\n");
296 		printk(KERN_ERR "INIT: If messages would not printed after "
297 			"this INIT, wait 30sec and assert INIT again.\n");
298 		if (!mlogbuf_timestamp)
299 			mlogbuf_timestamp = jiffies;
300 		return;
301 	}
302 	spin_unlock(&mlogbuf_rlock);
303 	ia64_mlogbuf_dump();
304 }
305 
306 static void inline
307 ia64_mca_spin(const char *func)
308 {
309 	if (monarch_cpu == smp_processor_id())
310 		ia64_mlogbuf_finish(0);
311 	mprintk(KERN_EMERG "%s: spinning here, not returning to SAL\n", func);
312 	while (1)
313 		cpu_relax();
314 }
315 /*
316  * IA64_MCA log support
317  */
318 #define IA64_MAX_LOGS		2	/* Double-buffering for nested MCAs */
319 #define IA64_MAX_LOG_TYPES      4   /* MCA, INIT, CMC, CPE */
320 
321 typedef struct ia64_state_log_s
322 {
323 	spinlock_t	isl_lock;
324 	int		isl_index;
325 	unsigned long	isl_count;
326 	ia64_err_rec_t  *isl_log[IA64_MAX_LOGS]; /* need space to store header + error log */
327 } ia64_state_log_t;
328 
329 static ia64_state_log_t ia64_state_log[IA64_MAX_LOG_TYPES];
330 
331 #define IA64_LOG_ALLOCATE(it, size) \
332 	{ia64_state_log[it].isl_log[IA64_LOG_CURR_INDEX(it)] = \
333 		(ia64_err_rec_t *)alloc_bootmem(size); \
334 	ia64_state_log[it].isl_log[IA64_LOG_NEXT_INDEX(it)] = \
335 		(ia64_err_rec_t *)alloc_bootmem(size);}
336 #define IA64_LOG_LOCK_INIT(it) spin_lock_init(&ia64_state_log[it].isl_lock)
337 #define IA64_LOG_LOCK(it)      spin_lock_irqsave(&ia64_state_log[it].isl_lock, s)
338 #define IA64_LOG_UNLOCK(it)    spin_unlock_irqrestore(&ia64_state_log[it].isl_lock,s)
339 #define IA64_LOG_NEXT_INDEX(it)    ia64_state_log[it].isl_index
340 #define IA64_LOG_CURR_INDEX(it)    1 - ia64_state_log[it].isl_index
341 #define IA64_LOG_INDEX_INC(it) \
342     {ia64_state_log[it].isl_index = 1 - ia64_state_log[it].isl_index; \
343     ia64_state_log[it].isl_count++;}
344 #define IA64_LOG_INDEX_DEC(it) \
345     ia64_state_log[it].isl_index = 1 - ia64_state_log[it].isl_index
346 #define IA64_LOG_NEXT_BUFFER(it)   (void *)((ia64_state_log[it].isl_log[IA64_LOG_NEXT_INDEX(it)]))
347 #define IA64_LOG_CURR_BUFFER(it)   (void *)((ia64_state_log[it].isl_log[IA64_LOG_CURR_INDEX(it)]))
348 #define IA64_LOG_COUNT(it)         ia64_state_log[it].isl_count
349 
350 /*
351  * ia64_log_init
352  *	Reset the OS ia64 log buffer
353  * Inputs   :   info_type   (SAL_INFO_TYPE_{MCA,INIT,CMC,CPE})
354  * Outputs	:	None
355  */
356 static void __init
357 ia64_log_init(int sal_info_type)
358 {
359 	u64	max_size = 0;
360 
361 	IA64_LOG_NEXT_INDEX(sal_info_type) = 0;
362 	IA64_LOG_LOCK_INIT(sal_info_type);
363 
364 	// SAL will tell us the maximum size of any error record of this type
365 	max_size = ia64_sal_get_state_info_size(sal_info_type);
366 	if (!max_size)
367 		/* alloc_bootmem() doesn't like zero-sized allocations! */
368 		return;
369 
370 	// set up OS data structures to hold error info
371 	IA64_LOG_ALLOCATE(sal_info_type, max_size);
372 	memset(IA64_LOG_CURR_BUFFER(sal_info_type), 0, max_size);
373 	memset(IA64_LOG_NEXT_BUFFER(sal_info_type), 0, max_size);
374 }
375 
376 /*
377  * ia64_log_get
378  *
379  *	Get the current MCA log from SAL and copy it into the OS log buffer.
380  *
381  *  Inputs  :   info_type   (SAL_INFO_TYPE_{MCA,INIT,CMC,CPE})
382  *              irq_safe    whether you can use printk at this point
383  *  Outputs :   size        (total record length)
384  *              *buffer     (ptr to error record)
385  *
386  */
387 static u64
388 ia64_log_get(int sal_info_type, u8 **buffer, int irq_safe)
389 {
390 	sal_log_record_header_t     *log_buffer;
391 	u64                         total_len = 0;
392 	unsigned long               s;
393 
394 	IA64_LOG_LOCK(sal_info_type);
395 
396 	/* Get the process state information */
397 	log_buffer = IA64_LOG_NEXT_BUFFER(sal_info_type);
398 
399 	total_len = ia64_sal_get_state_info(sal_info_type, (u64 *)log_buffer);
400 
401 	if (total_len) {
402 		IA64_LOG_INDEX_INC(sal_info_type);
403 		IA64_LOG_UNLOCK(sal_info_type);
404 		if (irq_safe) {
405 			IA64_MCA_DEBUG("%s: SAL error record type %d retrieved. "
406 				       "Record length = %ld\n", __FUNCTION__, sal_info_type, total_len);
407 		}
408 		*buffer = (u8 *) log_buffer;
409 		return total_len;
410 	} else {
411 		IA64_LOG_UNLOCK(sal_info_type);
412 		return 0;
413 	}
414 }
415 
416 /*
417  *  ia64_mca_log_sal_error_record
418  *
419  *  This function retrieves a specified error record type from SAL
420  *  and wakes up any processes waiting for error records.
421  *
422  *  Inputs  :   sal_info_type   (Type of error record MCA/CMC/CPE)
423  *              FIXME: remove MCA and irq_safe.
424  */
425 static void
426 ia64_mca_log_sal_error_record(int sal_info_type)
427 {
428 	u8 *buffer;
429 	sal_log_record_header_t *rh;
430 	u64 size;
431 	int irq_safe = sal_info_type != SAL_INFO_TYPE_MCA;
432 #ifdef IA64_MCA_DEBUG_INFO
433 	static const char * const rec_name[] = { "MCA", "INIT", "CMC", "CPE" };
434 #endif
435 
436 	size = ia64_log_get(sal_info_type, &buffer, irq_safe);
437 	if (!size)
438 		return;
439 
440 	salinfo_log_wakeup(sal_info_type, buffer, size, irq_safe);
441 
442 	if (irq_safe)
443 		IA64_MCA_DEBUG("CPU %d: SAL log contains %s error record\n",
444 			smp_processor_id(),
445 			sal_info_type < ARRAY_SIZE(rec_name) ? rec_name[sal_info_type] : "UNKNOWN");
446 
447 	/* Clear logs from corrected errors in case there's no user-level logger */
448 	rh = (sal_log_record_header_t *)buffer;
449 	if (rh->severity == sal_log_severity_corrected)
450 		ia64_sal_clear_state_info(sal_info_type);
451 }
452 
453 /*
454  * search_mca_table
455  *  See if the MCA surfaced in an instruction range
456  *  that has been tagged as recoverable.
457  *
458  *  Inputs
459  *	first	First address range to check
460  *	last	Last address range to check
461  *	ip	Instruction pointer, address we are looking for
462  *
463  * Return value:
464  *      1 on Success (in the table)/ 0 on Failure (not in the  table)
465  */
466 int
467 search_mca_table (const struct mca_table_entry *first,
468                 const struct mca_table_entry *last,
469                 unsigned long ip)
470 {
471         const struct mca_table_entry *curr;
472         u64 curr_start, curr_end;
473 
474         curr = first;
475         while (curr <= last) {
476                 curr_start = (u64) &curr->start_addr + curr->start_addr;
477                 curr_end = (u64) &curr->end_addr + curr->end_addr;
478 
479                 if ((ip >= curr_start) && (ip <= curr_end)) {
480                         return 1;
481                 }
482                 curr++;
483         }
484         return 0;
485 }
486 
487 /* Given an address, look for it in the mca tables. */
488 int mca_recover_range(unsigned long addr)
489 {
490 	extern struct mca_table_entry __start___mca_table[];
491 	extern struct mca_table_entry __stop___mca_table[];
492 
493 	return search_mca_table(__start___mca_table, __stop___mca_table-1, addr);
494 }
495 EXPORT_SYMBOL_GPL(mca_recover_range);
496 
497 #ifdef CONFIG_ACPI
498 
499 int cpe_vector = -1;
500 int ia64_cpe_irq = -1;
501 
502 static irqreturn_t
503 ia64_mca_cpe_int_handler (int cpe_irq, void *arg)
504 {
505 	static unsigned long	cpe_history[CPE_HISTORY_LENGTH];
506 	static int		index;
507 	static DEFINE_SPINLOCK(cpe_history_lock);
508 
509 	IA64_MCA_DEBUG("%s: received interrupt vector = %#x on CPU %d\n",
510 		       __FUNCTION__, cpe_irq, smp_processor_id());
511 
512 	/* SAL spec states this should run w/ interrupts enabled */
513 	local_irq_enable();
514 
515 	spin_lock(&cpe_history_lock);
516 	if (!cpe_poll_enabled && cpe_vector >= 0) {
517 
518 		int i, count = 1; /* we know 1 happened now */
519 		unsigned long now = jiffies;
520 
521 		for (i = 0; i < CPE_HISTORY_LENGTH; i++) {
522 			if (now - cpe_history[i] <= HZ)
523 				count++;
524 		}
525 
526 		IA64_MCA_DEBUG(KERN_INFO "CPE threshold %d/%d\n", count, CPE_HISTORY_LENGTH);
527 		if (count >= CPE_HISTORY_LENGTH) {
528 
529 			cpe_poll_enabled = 1;
530 			spin_unlock(&cpe_history_lock);
531 			disable_irq_nosync(local_vector_to_irq(IA64_CPE_VECTOR));
532 
533 			/*
534 			 * Corrected errors will still be corrected, but
535 			 * make sure there's a log somewhere that indicates
536 			 * something is generating more than we can handle.
537 			 */
538 			printk(KERN_WARNING "WARNING: Switching to polling CPE handler; error records may be lost\n");
539 
540 			mod_timer(&cpe_poll_timer, jiffies + MIN_CPE_POLL_INTERVAL);
541 
542 			/* lock already released, get out now */
543 			goto out;
544 		} else {
545 			cpe_history[index++] = now;
546 			if (index == CPE_HISTORY_LENGTH)
547 				index = 0;
548 		}
549 	}
550 	spin_unlock(&cpe_history_lock);
551 out:
552 	/* Get the CPE error record and log it */
553 	ia64_mca_log_sal_error_record(SAL_INFO_TYPE_CPE);
554 
555 	return IRQ_HANDLED;
556 }
557 
558 #endif /* CONFIG_ACPI */
559 
560 #ifdef CONFIG_ACPI
561 /*
562  * ia64_mca_register_cpev
563  *
564  *  Register the corrected platform error vector with SAL.
565  *
566  *  Inputs
567  *      cpev        Corrected Platform Error Vector number
568  *
569  *  Outputs
570  *      None
571  */
572 static void __init
573 ia64_mca_register_cpev (int cpev)
574 {
575 	/* Register the CPE interrupt vector with SAL */
576 	struct ia64_sal_retval isrv;
577 
578 	isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_CPE_INT, SAL_MC_PARAM_MECHANISM_INT, cpev, 0, 0);
579 	if (isrv.status) {
580 		printk(KERN_ERR "Failed to register Corrected Platform "
581 		       "Error interrupt vector with SAL (status %ld)\n", isrv.status);
582 		return;
583 	}
584 
585 	IA64_MCA_DEBUG("%s: corrected platform error "
586 		       "vector %#x registered\n", __FUNCTION__, cpev);
587 }
588 #endif /* CONFIG_ACPI */
589 
590 /*
591  * ia64_mca_cmc_vector_setup
592  *
593  *  Setup the corrected machine check vector register in the processor.
594  *  (The interrupt is masked on boot. ia64_mca_late_init unmask this.)
595  *  This function is invoked on a per-processor basis.
596  *
597  * Inputs
598  *      None
599  *
600  * Outputs
601  *	None
602  */
603 void __cpuinit
604 ia64_mca_cmc_vector_setup (void)
605 {
606 	cmcv_reg_t	cmcv;
607 
608 	cmcv.cmcv_regval	= 0;
609 	cmcv.cmcv_mask		= 1;        /* Mask/disable interrupt at first */
610 	cmcv.cmcv_vector	= IA64_CMC_VECTOR;
611 	ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
612 
613 	IA64_MCA_DEBUG("%s: CPU %d corrected "
614 		       "machine check vector %#x registered.\n",
615 		       __FUNCTION__, smp_processor_id(), IA64_CMC_VECTOR);
616 
617 	IA64_MCA_DEBUG("%s: CPU %d CMCV = %#016lx\n",
618 		       __FUNCTION__, smp_processor_id(), ia64_getreg(_IA64_REG_CR_CMCV));
619 }
620 
621 /*
622  * ia64_mca_cmc_vector_disable
623  *
624  *  Mask the corrected machine check vector register in the processor.
625  *  This function is invoked on a per-processor basis.
626  *
627  * Inputs
628  *      dummy(unused)
629  *
630  * Outputs
631  *	None
632  */
633 static void
634 ia64_mca_cmc_vector_disable (void *dummy)
635 {
636 	cmcv_reg_t	cmcv;
637 
638 	cmcv.cmcv_regval = ia64_getreg(_IA64_REG_CR_CMCV);
639 
640 	cmcv.cmcv_mask = 1; /* Mask/disable interrupt */
641 	ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
642 
643 	IA64_MCA_DEBUG("%s: CPU %d corrected "
644 		       "machine check vector %#x disabled.\n",
645 		       __FUNCTION__, smp_processor_id(), cmcv.cmcv_vector);
646 }
647 
648 /*
649  * ia64_mca_cmc_vector_enable
650  *
651  *  Unmask the corrected machine check vector register in the processor.
652  *  This function is invoked on a per-processor basis.
653  *
654  * Inputs
655  *      dummy(unused)
656  *
657  * Outputs
658  *	None
659  */
660 static void
661 ia64_mca_cmc_vector_enable (void *dummy)
662 {
663 	cmcv_reg_t	cmcv;
664 
665 	cmcv.cmcv_regval = ia64_getreg(_IA64_REG_CR_CMCV);
666 
667 	cmcv.cmcv_mask = 0; /* Unmask/enable interrupt */
668 	ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
669 
670 	IA64_MCA_DEBUG("%s: CPU %d corrected "
671 		       "machine check vector %#x enabled.\n",
672 		       __FUNCTION__, smp_processor_id(), cmcv.cmcv_vector);
673 }
674 
675 /*
676  * ia64_mca_cmc_vector_disable_keventd
677  *
678  * Called via keventd (smp_call_function() is not safe in interrupt context) to
679  * disable the cmc interrupt vector.
680  */
681 static void
682 ia64_mca_cmc_vector_disable_keventd(struct work_struct *unused)
683 {
684 	on_each_cpu(ia64_mca_cmc_vector_disable, NULL, 1, 0);
685 }
686 
687 /*
688  * ia64_mca_cmc_vector_enable_keventd
689  *
690  * Called via keventd (smp_call_function() is not safe in interrupt context) to
691  * enable the cmc interrupt vector.
692  */
693 static void
694 ia64_mca_cmc_vector_enable_keventd(struct work_struct *unused)
695 {
696 	on_each_cpu(ia64_mca_cmc_vector_enable, NULL, 1, 0);
697 }
698 
699 /*
700  * ia64_mca_wakeup
701  *
702  *	Send an inter-cpu interrupt to wake-up a particular cpu
703  *	and mark that cpu to be out of rendez.
704  *
705  *  Inputs  :   cpuid
706  *  Outputs :   None
707  */
708 static void
709 ia64_mca_wakeup(int cpu)
710 {
711 	platform_send_ipi(cpu, IA64_MCA_WAKEUP_VECTOR, IA64_IPI_DM_INT, 0);
712 	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
713 
714 }
715 
716 /*
717  * ia64_mca_wakeup_all
718  *
719  *	Wakeup all the cpus which have rendez'ed previously.
720  *
721  *  Inputs  :   None
722  *  Outputs :   None
723  */
724 static void
725 ia64_mca_wakeup_all(void)
726 {
727 	int cpu;
728 
729 	/* Clear the Rendez checkin flag for all cpus */
730 	for_each_online_cpu(cpu) {
731 		if (ia64_mc_info.imi_rendez_checkin[cpu] == IA64_MCA_RENDEZ_CHECKIN_DONE)
732 			ia64_mca_wakeup(cpu);
733 	}
734 
735 }
736 
737 /*
738  * ia64_mca_rendez_interrupt_handler
739  *
740  *	This is handler used to put slave processors into spinloop
741  *	while the monarch processor does the mca handling and later
742  *	wake each slave up once the monarch is done.
743  *
744  *  Inputs  :   None
745  *  Outputs :   None
746  */
747 static irqreturn_t
748 ia64_mca_rendez_int_handler(int rendez_irq, void *arg)
749 {
750 	unsigned long flags;
751 	int cpu = smp_processor_id();
752 	struct ia64_mca_notify_die nd =
753 		{ .sos = NULL, .monarch_cpu = &monarch_cpu };
754 
755 	/* Mask all interrupts */
756 	local_irq_save(flags);
757 	if (notify_die(DIE_MCA_RENDZVOUS_ENTER, "MCA", get_irq_regs(),
758 		       (long)&nd, 0, 0) == NOTIFY_STOP)
759 		ia64_mca_spin(__FUNCTION__);
760 
761 	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_DONE;
762 	/* Register with the SAL monarch that the slave has
763 	 * reached SAL
764 	 */
765 	ia64_sal_mc_rendez();
766 
767 	if (notify_die(DIE_MCA_RENDZVOUS_PROCESS, "MCA", get_irq_regs(),
768 		       (long)&nd, 0, 0) == NOTIFY_STOP)
769 		ia64_mca_spin(__FUNCTION__);
770 
771 	/* Wait for the monarch cpu to exit. */
772 	while (monarch_cpu != -1)
773 	       cpu_relax();	/* spin until monarch leaves */
774 
775 	if (notify_die(DIE_MCA_RENDZVOUS_LEAVE, "MCA", get_irq_regs(),
776 		       (long)&nd, 0, 0) == NOTIFY_STOP)
777 		ia64_mca_spin(__FUNCTION__);
778 
779 	/* Enable all interrupts */
780 	local_irq_restore(flags);
781 	return IRQ_HANDLED;
782 }
783 
784 /*
785  * ia64_mca_wakeup_int_handler
786  *
787  *	The interrupt handler for processing the inter-cpu interrupt to the
788  *	slave cpu which was spinning in the rendez loop.
789  *	Since this spinning is done by turning off the interrupts and
790  *	polling on the wakeup-interrupt bit in the IRR, there is
791  *	nothing useful to be done in the handler.
792  *
793  *  Inputs  :   wakeup_irq  (Wakeup-interrupt bit)
794  *	arg		(Interrupt handler specific argument)
795  *  Outputs :   None
796  *
797  */
798 static irqreturn_t
799 ia64_mca_wakeup_int_handler(int wakeup_irq, void *arg)
800 {
801 	return IRQ_HANDLED;
802 }
803 
804 /* Function pointer for extra MCA recovery */
805 int (*ia64_mca_ucmc_extension)
806 	(void*,struct ia64_sal_os_state*)
807 	= NULL;
808 
809 int
810 ia64_reg_MCA_extension(int (*fn)(void *, struct ia64_sal_os_state *))
811 {
812 	if (ia64_mca_ucmc_extension)
813 		return 1;
814 
815 	ia64_mca_ucmc_extension = fn;
816 	return 0;
817 }
818 
819 void
820 ia64_unreg_MCA_extension(void)
821 {
822 	if (ia64_mca_ucmc_extension)
823 		ia64_mca_ucmc_extension = NULL;
824 }
825 
826 EXPORT_SYMBOL(ia64_reg_MCA_extension);
827 EXPORT_SYMBOL(ia64_unreg_MCA_extension);
828 
829 
830 static inline void
831 copy_reg(const u64 *fr, u64 fnat, u64 *tr, u64 *tnat)
832 {
833 	u64 fslot, tslot, nat;
834 	*tr = *fr;
835 	fslot = ((unsigned long)fr >> 3) & 63;
836 	tslot = ((unsigned long)tr >> 3) & 63;
837 	*tnat &= ~(1UL << tslot);
838 	nat = (fnat >> fslot) & 1;
839 	*tnat |= (nat << tslot);
840 }
841 
842 /* Change the comm field on the MCA/INT task to include the pid that
843  * was interrupted, it makes for easier debugging.  If that pid was 0
844  * (swapper or nested MCA/INIT) then use the start of the previous comm
845  * field suffixed with its cpu.
846  */
847 
848 static void
849 ia64_mca_modify_comm(const struct task_struct *previous_current)
850 {
851 	char *p, comm[sizeof(current->comm)];
852 	if (previous_current->pid)
853 		snprintf(comm, sizeof(comm), "%s %d",
854 			current->comm, previous_current->pid);
855 	else {
856 		int l;
857 		if ((p = strchr(previous_current->comm, ' ')))
858 			l = p - previous_current->comm;
859 		else
860 			l = strlen(previous_current->comm);
861 		snprintf(comm, sizeof(comm), "%s %*s %d",
862 			current->comm, l, previous_current->comm,
863 			task_thread_info(previous_current)->cpu);
864 	}
865 	memcpy(current->comm, comm, sizeof(current->comm));
866 }
867 
868 /* On entry to this routine, we are running on the per cpu stack, see
869  * mca_asm.h.  The original stack has not been touched by this event.  Some of
870  * the original stack's registers will be in the RBS on this stack.  This stack
871  * also contains a partial pt_regs and switch_stack, the rest of the data is in
872  * PAL minstate.
873  *
874  * The first thing to do is modify the original stack to look like a blocked
875  * task so we can run backtrace on the original task.  Also mark the per cpu
876  * stack as current to ensure that we use the correct task state, it also means
877  * that we can do backtrace on the MCA/INIT handler code itself.
878  */
879 
880 static struct task_struct *
881 ia64_mca_modify_original_stack(struct pt_regs *regs,
882 		const struct switch_stack *sw,
883 		struct ia64_sal_os_state *sos,
884 		const char *type)
885 {
886 	char *p;
887 	ia64_va va;
888 	extern char ia64_leave_kernel[];	/* Need asm address, not function descriptor */
889 	const pal_min_state_area_t *ms = sos->pal_min_state;
890 	struct task_struct *previous_current;
891 	struct pt_regs *old_regs;
892 	struct switch_stack *old_sw;
893 	unsigned size = sizeof(struct pt_regs) +
894 			sizeof(struct switch_stack) + 16;
895 	u64 *old_bspstore, *old_bsp;
896 	u64 *new_bspstore, *new_bsp;
897 	u64 old_unat, old_rnat, new_rnat, nat;
898 	u64 slots, loadrs = regs->loadrs;
899 	u64 r12 = ms->pmsa_gr[12-1], r13 = ms->pmsa_gr[13-1];
900 	u64 ar_bspstore = regs->ar_bspstore;
901 	u64 ar_bsp = regs->ar_bspstore + (loadrs >> 16);
902 	const u64 *bank;
903 	const char *msg;
904 	int cpu = smp_processor_id();
905 
906 	previous_current = curr_task(cpu);
907 	set_curr_task(cpu, current);
908 	if ((p = strchr(current->comm, ' ')))
909 		*p = '\0';
910 
911 	/* Best effort attempt to cope with MCA/INIT delivered while in
912 	 * physical mode.
913 	 */
914 	regs->cr_ipsr = ms->pmsa_ipsr;
915 	if (ia64_psr(regs)->dt == 0) {
916 		va.l = r12;
917 		if (va.f.reg == 0) {
918 			va.f.reg = 7;
919 			r12 = va.l;
920 		}
921 		va.l = r13;
922 		if (va.f.reg == 0) {
923 			va.f.reg = 7;
924 			r13 = va.l;
925 		}
926 	}
927 	if (ia64_psr(regs)->rt == 0) {
928 		va.l = ar_bspstore;
929 		if (va.f.reg == 0) {
930 			va.f.reg = 7;
931 			ar_bspstore = va.l;
932 		}
933 		va.l = ar_bsp;
934 		if (va.f.reg == 0) {
935 			va.f.reg = 7;
936 			ar_bsp = va.l;
937 		}
938 	}
939 
940 	/* mca_asm.S ia64_old_stack() cannot assume that the dirty registers
941 	 * have been copied to the old stack, the old stack may fail the
942 	 * validation tests below.  So ia64_old_stack() must restore the dirty
943 	 * registers from the new stack.  The old and new bspstore probably
944 	 * have different alignments, so loadrs calculated on the old bsp
945 	 * cannot be used to restore from the new bsp.  Calculate a suitable
946 	 * loadrs for the new stack and save it in the new pt_regs, where
947 	 * ia64_old_stack() can get it.
948 	 */
949 	old_bspstore = (u64 *)ar_bspstore;
950 	old_bsp = (u64 *)ar_bsp;
951 	slots = ia64_rse_num_regs(old_bspstore, old_bsp);
952 	new_bspstore = (u64 *)((u64)current + IA64_RBS_OFFSET);
953 	new_bsp = ia64_rse_skip_regs(new_bspstore, slots);
954 	regs->loadrs = (new_bsp - new_bspstore) * 8 << 16;
955 
956 	/* Verify the previous stack state before we change it */
957 	if (user_mode(regs)) {
958 		msg = "occurred in user space";
959 		/* previous_current is guaranteed to be valid when the task was
960 		 * in user space, so ...
961 		 */
962 		ia64_mca_modify_comm(previous_current);
963 		goto no_mod;
964 	}
965 
966 	if (!mca_recover_range(ms->pmsa_iip)) {
967 		if (r13 != sos->prev_IA64_KR_CURRENT) {
968 			msg = "inconsistent previous current and r13";
969 			goto no_mod;
970 		}
971 		if ((r12 - r13) >= KERNEL_STACK_SIZE) {
972 			msg = "inconsistent r12 and r13";
973 			goto no_mod;
974 		}
975 		if ((ar_bspstore - r13) >= KERNEL_STACK_SIZE) {
976 			msg = "inconsistent ar.bspstore and r13";
977 			goto no_mod;
978 		}
979 		va.p = old_bspstore;
980 		if (va.f.reg < 5) {
981 			msg = "old_bspstore is in the wrong region";
982 			goto no_mod;
983 		}
984 		if ((ar_bsp - r13) >= KERNEL_STACK_SIZE) {
985 			msg = "inconsistent ar.bsp and r13";
986 			goto no_mod;
987 		}
988 		size += (ia64_rse_skip_regs(old_bspstore, slots) - old_bspstore) * 8;
989 		if (ar_bspstore + size > r12) {
990 			msg = "no room for blocked state";
991 			goto no_mod;
992 		}
993 	}
994 
995 	ia64_mca_modify_comm(previous_current);
996 
997 	/* Make the original task look blocked.  First stack a struct pt_regs,
998 	 * describing the state at the time of interrupt.  mca_asm.S built a
999 	 * partial pt_regs, copy it and fill in the blanks using minstate.
1000 	 */
1001 	p = (char *)r12 - sizeof(*regs);
1002 	old_regs = (struct pt_regs *)p;
1003 	memcpy(old_regs, regs, sizeof(*regs));
1004 	/* If ipsr.ic then use pmsa_{iip,ipsr,ifs}, else use
1005 	 * pmsa_{xip,xpsr,xfs}
1006 	 */
1007 	if (ia64_psr(regs)->ic) {
1008 		old_regs->cr_iip = ms->pmsa_iip;
1009 		old_regs->cr_ipsr = ms->pmsa_ipsr;
1010 		old_regs->cr_ifs = ms->pmsa_ifs;
1011 	} else {
1012 		old_regs->cr_iip = ms->pmsa_xip;
1013 		old_regs->cr_ipsr = ms->pmsa_xpsr;
1014 		old_regs->cr_ifs = ms->pmsa_xfs;
1015 	}
1016 	old_regs->pr = ms->pmsa_pr;
1017 	old_regs->b0 = ms->pmsa_br0;
1018 	old_regs->loadrs = loadrs;
1019 	old_regs->ar_rsc = ms->pmsa_rsc;
1020 	old_unat = old_regs->ar_unat;
1021 	copy_reg(&ms->pmsa_gr[1-1], ms->pmsa_nat_bits, &old_regs->r1, &old_unat);
1022 	copy_reg(&ms->pmsa_gr[2-1], ms->pmsa_nat_bits, &old_regs->r2, &old_unat);
1023 	copy_reg(&ms->pmsa_gr[3-1], ms->pmsa_nat_bits, &old_regs->r3, &old_unat);
1024 	copy_reg(&ms->pmsa_gr[8-1], ms->pmsa_nat_bits, &old_regs->r8, &old_unat);
1025 	copy_reg(&ms->pmsa_gr[9-1], ms->pmsa_nat_bits, &old_regs->r9, &old_unat);
1026 	copy_reg(&ms->pmsa_gr[10-1], ms->pmsa_nat_bits, &old_regs->r10, &old_unat);
1027 	copy_reg(&ms->pmsa_gr[11-1], ms->pmsa_nat_bits, &old_regs->r11, &old_unat);
1028 	copy_reg(&ms->pmsa_gr[12-1], ms->pmsa_nat_bits, &old_regs->r12, &old_unat);
1029 	copy_reg(&ms->pmsa_gr[13-1], ms->pmsa_nat_bits, &old_regs->r13, &old_unat);
1030 	copy_reg(&ms->pmsa_gr[14-1], ms->pmsa_nat_bits, &old_regs->r14, &old_unat);
1031 	copy_reg(&ms->pmsa_gr[15-1], ms->pmsa_nat_bits, &old_regs->r15, &old_unat);
1032 	if (ia64_psr(old_regs)->bn)
1033 		bank = ms->pmsa_bank1_gr;
1034 	else
1035 		bank = ms->pmsa_bank0_gr;
1036 	copy_reg(&bank[16-16], ms->pmsa_nat_bits, &old_regs->r16, &old_unat);
1037 	copy_reg(&bank[17-16], ms->pmsa_nat_bits, &old_regs->r17, &old_unat);
1038 	copy_reg(&bank[18-16], ms->pmsa_nat_bits, &old_regs->r18, &old_unat);
1039 	copy_reg(&bank[19-16], ms->pmsa_nat_bits, &old_regs->r19, &old_unat);
1040 	copy_reg(&bank[20-16], ms->pmsa_nat_bits, &old_regs->r20, &old_unat);
1041 	copy_reg(&bank[21-16], ms->pmsa_nat_bits, &old_regs->r21, &old_unat);
1042 	copy_reg(&bank[22-16], ms->pmsa_nat_bits, &old_regs->r22, &old_unat);
1043 	copy_reg(&bank[23-16], ms->pmsa_nat_bits, &old_regs->r23, &old_unat);
1044 	copy_reg(&bank[24-16], ms->pmsa_nat_bits, &old_regs->r24, &old_unat);
1045 	copy_reg(&bank[25-16], ms->pmsa_nat_bits, &old_regs->r25, &old_unat);
1046 	copy_reg(&bank[26-16], ms->pmsa_nat_bits, &old_regs->r26, &old_unat);
1047 	copy_reg(&bank[27-16], ms->pmsa_nat_bits, &old_regs->r27, &old_unat);
1048 	copy_reg(&bank[28-16], ms->pmsa_nat_bits, &old_regs->r28, &old_unat);
1049 	copy_reg(&bank[29-16], ms->pmsa_nat_bits, &old_regs->r29, &old_unat);
1050 	copy_reg(&bank[30-16], ms->pmsa_nat_bits, &old_regs->r30, &old_unat);
1051 	copy_reg(&bank[31-16], ms->pmsa_nat_bits, &old_regs->r31, &old_unat);
1052 
1053 	/* Next stack a struct switch_stack.  mca_asm.S built a partial
1054 	 * switch_stack, copy it and fill in the blanks using pt_regs and
1055 	 * minstate.
1056 	 *
1057 	 * In the synthesized switch_stack, b0 points to ia64_leave_kernel,
1058 	 * ar.pfs is set to 0.
1059 	 *
1060 	 * unwind.c::unw_unwind() does special processing for interrupt frames.
1061 	 * It checks if the PRED_NON_SYSCALL predicate is set, if the predicate
1062 	 * is clear then unw_unwind() does _not_ adjust bsp over pt_regs.  Not
1063 	 * that this is documented, of course.  Set PRED_NON_SYSCALL in the
1064 	 * switch_stack on the original stack so it will unwind correctly when
1065 	 * unwind.c reads pt_regs.
1066 	 *
1067 	 * thread.ksp is updated to point to the synthesized switch_stack.
1068 	 */
1069 	p -= sizeof(struct switch_stack);
1070 	old_sw = (struct switch_stack *)p;
1071 	memcpy(old_sw, sw, sizeof(*sw));
1072 	old_sw->caller_unat = old_unat;
1073 	old_sw->ar_fpsr = old_regs->ar_fpsr;
1074 	copy_reg(&ms->pmsa_gr[4-1], ms->pmsa_nat_bits, &old_sw->r4, &old_unat);
1075 	copy_reg(&ms->pmsa_gr[5-1], ms->pmsa_nat_bits, &old_sw->r5, &old_unat);
1076 	copy_reg(&ms->pmsa_gr[6-1], ms->pmsa_nat_bits, &old_sw->r6, &old_unat);
1077 	copy_reg(&ms->pmsa_gr[7-1], ms->pmsa_nat_bits, &old_sw->r7, &old_unat);
1078 	old_sw->b0 = (u64)ia64_leave_kernel;
1079 	old_sw->b1 = ms->pmsa_br1;
1080 	old_sw->ar_pfs = 0;
1081 	old_sw->ar_unat = old_unat;
1082 	old_sw->pr = old_regs->pr | (1UL << PRED_NON_SYSCALL);
1083 	previous_current->thread.ksp = (u64)p - 16;
1084 
1085 	/* Finally copy the original stack's registers back to its RBS.
1086 	 * Registers from ar.bspstore through ar.bsp at the time of the event
1087 	 * are in the current RBS, copy them back to the original stack.  The
1088 	 * copy must be done register by register because the original bspstore
1089 	 * and the current one have different alignments, so the saved RNAT
1090 	 * data occurs at different places.
1091 	 *
1092 	 * mca_asm does cover, so the old_bsp already includes all registers at
1093 	 * the time of MCA/INIT.  It also does flushrs, so all registers before
1094 	 * this function have been written to backing store on the MCA/INIT
1095 	 * stack.
1096 	 */
1097 	new_rnat = ia64_get_rnat(ia64_rse_rnat_addr(new_bspstore));
1098 	old_rnat = regs->ar_rnat;
1099 	while (slots--) {
1100 		if (ia64_rse_is_rnat_slot(new_bspstore)) {
1101 			new_rnat = ia64_get_rnat(new_bspstore++);
1102 		}
1103 		if (ia64_rse_is_rnat_slot(old_bspstore)) {
1104 			*old_bspstore++ = old_rnat;
1105 			old_rnat = 0;
1106 		}
1107 		nat = (new_rnat >> ia64_rse_slot_num(new_bspstore)) & 1UL;
1108 		old_rnat &= ~(1UL << ia64_rse_slot_num(old_bspstore));
1109 		old_rnat |= (nat << ia64_rse_slot_num(old_bspstore));
1110 		*old_bspstore++ = *new_bspstore++;
1111 	}
1112 	old_sw->ar_bspstore = (unsigned long)old_bspstore;
1113 	old_sw->ar_rnat = old_rnat;
1114 
1115 	sos->prev_task = previous_current;
1116 	return previous_current;
1117 
1118 no_mod:
1119 	printk(KERN_INFO "cpu %d, %s %s, original stack not modified\n",
1120 			smp_processor_id(), type, msg);
1121 	return previous_current;
1122 }
1123 
1124 /* The monarch/slave interaction is based on monarch_cpu and requires that all
1125  * slaves have entered rendezvous before the monarch leaves.  If any cpu has
1126  * not entered rendezvous yet then wait a bit.  The assumption is that any
1127  * slave that has not rendezvoused after a reasonable time is never going to do
1128  * so.  In this context, slave includes cpus that respond to the MCA rendezvous
1129  * interrupt, as well as cpus that receive the INIT slave event.
1130  */
1131 
1132 static void
1133 ia64_wait_for_slaves(int monarch, const char *type)
1134 {
1135 	int c, wait = 0, missing = 0;
1136 	for_each_online_cpu(c) {
1137 		if (c == monarch)
1138 			continue;
1139 		if (ia64_mc_info.imi_rendez_checkin[c] == IA64_MCA_RENDEZ_CHECKIN_NOTDONE) {
1140 			udelay(1000);		/* short wait first */
1141 			wait = 1;
1142 			break;
1143 		}
1144 	}
1145 	if (!wait)
1146 		goto all_in;
1147 	for_each_online_cpu(c) {
1148 		if (c == monarch)
1149 			continue;
1150 		if (ia64_mc_info.imi_rendez_checkin[c] == IA64_MCA_RENDEZ_CHECKIN_NOTDONE) {
1151 			udelay(5*1000000);	/* wait 5 seconds for slaves (arbitrary) */
1152 			if (ia64_mc_info.imi_rendez_checkin[c] == IA64_MCA_RENDEZ_CHECKIN_NOTDONE)
1153 				missing = 1;
1154 			break;
1155 		}
1156 	}
1157 	if (!missing)
1158 		goto all_in;
1159 	/*
1160 	 * Maybe slave(s) dead. Print buffered messages immediately.
1161 	 */
1162 	ia64_mlogbuf_finish(0);
1163 	mprintk(KERN_INFO "OS %s slave did not rendezvous on cpu", type);
1164 	for_each_online_cpu(c) {
1165 		if (c == monarch)
1166 			continue;
1167 		if (ia64_mc_info.imi_rendez_checkin[c] == IA64_MCA_RENDEZ_CHECKIN_NOTDONE)
1168 			mprintk(" %d", c);
1169 	}
1170 	mprintk("\n");
1171 	return;
1172 
1173 all_in:
1174 	mprintk(KERN_INFO "All OS %s slaves have reached rendezvous\n", type);
1175 	return;
1176 }
1177 
1178 /*
1179  * ia64_mca_handler
1180  *
1181  *	This is uncorrectable machine check handler called from OS_MCA
1182  *	dispatch code which is in turn called from SAL_CHECK().
1183  *	This is the place where the core of OS MCA handling is done.
1184  *	Right now the logs are extracted and displayed in a well-defined
1185  *	format. This handler code is supposed to be run only on the
1186  *	monarch processor. Once the monarch is done with MCA handling
1187  *	further MCA logging is enabled by clearing logs.
1188  *	Monarch also has the duty of sending wakeup-IPIs to pull the
1189  *	slave processors out of rendezvous spinloop.
1190  */
1191 void
1192 ia64_mca_handler(struct pt_regs *regs, struct switch_stack *sw,
1193 		 struct ia64_sal_os_state *sos)
1194 {
1195 	int recover, cpu = smp_processor_id();
1196 	struct task_struct *previous_current;
1197 	struct ia64_mca_notify_die nd =
1198 		{ .sos = sos, .monarch_cpu = &monarch_cpu };
1199 
1200 	mprintk(KERN_INFO "Entered OS MCA handler. PSP=%lx cpu=%d "
1201 		"monarch=%ld\n", sos->proc_state_param, cpu, sos->monarch);
1202 
1203 	previous_current = ia64_mca_modify_original_stack(regs, sw, sos, "MCA");
1204 	monarch_cpu = cpu;
1205 	if (notify_die(DIE_MCA_MONARCH_ENTER, "MCA", regs, (long)&nd, 0, 0)
1206 			== NOTIFY_STOP)
1207 		ia64_mca_spin(__FUNCTION__);
1208 	ia64_wait_for_slaves(cpu, "MCA");
1209 
1210 	/* Wakeup all the processors which are spinning in the rendezvous loop.
1211 	 * They will leave SAL, then spin in the OS with interrupts disabled
1212 	 * until this monarch cpu leaves the MCA handler.  That gets control
1213 	 * back to the OS so we can backtrace the other cpus, backtrace when
1214 	 * spinning in SAL does not work.
1215 	 */
1216 	ia64_mca_wakeup_all();
1217 	if (notify_die(DIE_MCA_MONARCH_PROCESS, "MCA", regs, (long)&nd, 0, 0)
1218 			== NOTIFY_STOP)
1219 		ia64_mca_spin(__FUNCTION__);
1220 
1221 	/* Get the MCA error record and log it */
1222 	ia64_mca_log_sal_error_record(SAL_INFO_TYPE_MCA);
1223 
1224 	/* MCA error recovery */
1225 	recover = (ia64_mca_ucmc_extension
1226 		&& ia64_mca_ucmc_extension(
1227 			IA64_LOG_CURR_BUFFER(SAL_INFO_TYPE_MCA),
1228 			sos));
1229 
1230 	if (recover) {
1231 		sal_log_record_header_t *rh = IA64_LOG_CURR_BUFFER(SAL_INFO_TYPE_MCA);
1232 		rh->severity = sal_log_severity_corrected;
1233 		ia64_sal_clear_state_info(SAL_INFO_TYPE_MCA);
1234 		sos->os_status = IA64_MCA_CORRECTED;
1235 	} else {
1236 		/* Dump buffered message to console */
1237 		ia64_mlogbuf_finish(1);
1238 #ifdef CONFIG_KEXEC
1239 		atomic_set(&kdump_in_progress, 1);
1240 		monarch_cpu = -1;
1241 #endif
1242 	}
1243 	if (notify_die(DIE_MCA_MONARCH_LEAVE, "MCA", regs, (long)&nd, 0, recover)
1244 			== NOTIFY_STOP)
1245 		ia64_mca_spin(__FUNCTION__);
1246 
1247 	set_curr_task(cpu, previous_current);
1248 	monarch_cpu = -1;
1249 }
1250 
1251 static DECLARE_WORK(cmc_disable_work, ia64_mca_cmc_vector_disable_keventd);
1252 static DECLARE_WORK(cmc_enable_work, ia64_mca_cmc_vector_enable_keventd);
1253 
1254 /*
1255  * ia64_mca_cmc_int_handler
1256  *
1257  *  This is corrected machine check interrupt handler.
1258  *	Right now the logs are extracted and displayed in a well-defined
1259  *	format.
1260  *
1261  * Inputs
1262  *      interrupt number
1263  *      client data arg ptr
1264  *
1265  * Outputs
1266  *	None
1267  */
1268 static irqreturn_t
1269 ia64_mca_cmc_int_handler(int cmc_irq, void *arg)
1270 {
1271 	static unsigned long	cmc_history[CMC_HISTORY_LENGTH];
1272 	static int		index;
1273 	static DEFINE_SPINLOCK(cmc_history_lock);
1274 
1275 	IA64_MCA_DEBUG("%s: received interrupt vector = %#x on CPU %d\n",
1276 		       __FUNCTION__, cmc_irq, smp_processor_id());
1277 
1278 	/* SAL spec states this should run w/ interrupts enabled */
1279 	local_irq_enable();
1280 
1281 	spin_lock(&cmc_history_lock);
1282 	if (!cmc_polling_enabled) {
1283 		int i, count = 1; /* we know 1 happened now */
1284 		unsigned long now = jiffies;
1285 
1286 		for (i = 0; i < CMC_HISTORY_LENGTH; i++) {
1287 			if (now - cmc_history[i] <= HZ)
1288 				count++;
1289 		}
1290 
1291 		IA64_MCA_DEBUG(KERN_INFO "CMC threshold %d/%d\n", count, CMC_HISTORY_LENGTH);
1292 		if (count >= CMC_HISTORY_LENGTH) {
1293 
1294 			cmc_polling_enabled = 1;
1295 			spin_unlock(&cmc_history_lock);
1296 			/* If we're being hit with CMC interrupts, we won't
1297 			 * ever execute the schedule_work() below.  Need to
1298 			 * disable CMC interrupts on this processor now.
1299 			 */
1300 			ia64_mca_cmc_vector_disable(NULL);
1301 			schedule_work(&cmc_disable_work);
1302 
1303 			/*
1304 			 * Corrected errors will still be corrected, but
1305 			 * make sure there's a log somewhere that indicates
1306 			 * something is generating more than we can handle.
1307 			 */
1308 			printk(KERN_WARNING "WARNING: Switching to polling CMC handler; error records may be lost\n");
1309 
1310 			mod_timer(&cmc_poll_timer, jiffies + CMC_POLL_INTERVAL);
1311 
1312 			/* lock already released, get out now */
1313 			goto out;
1314 		} else {
1315 			cmc_history[index++] = now;
1316 			if (index == CMC_HISTORY_LENGTH)
1317 				index = 0;
1318 		}
1319 	}
1320 	spin_unlock(&cmc_history_lock);
1321 out:
1322 	/* Get the CMC error record and log it */
1323 	ia64_mca_log_sal_error_record(SAL_INFO_TYPE_CMC);
1324 
1325 	return IRQ_HANDLED;
1326 }
1327 
1328 /*
1329  *  ia64_mca_cmc_int_caller
1330  *
1331  * 	Triggered by sw interrupt from CMC polling routine.  Calls
1332  * 	real interrupt handler and either triggers a sw interrupt
1333  * 	on the next cpu or does cleanup at the end.
1334  *
1335  * Inputs
1336  *	interrupt number
1337  *	client data arg ptr
1338  * Outputs
1339  * 	handled
1340  */
1341 static irqreturn_t
1342 ia64_mca_cmc_int_caller(int cmc_irq, void *arg)
1343 {
1344 	static int start_count = -1;
1345 	unsigned int cpuid;
1346 
1347 	cpuid = smp_processor_id();
1348 
1349 	/* If first cpu, update count */
1350 	if (start_count == -1)
1351 		start_count = IA64_LOG_COUNT(SAL_INFO_TYPE_CMC);
1352 
1353 	ia64_mca_cmc_int_handler(cmc_irq, arg);
1354 
1355 	for (++cpuid ; cpuid < NR_CPUS && !cpu_online(cpuid) ; cpuid++);
1356 
1357 	if (cpuid < NR_CPUS) {
1358 		platform_send_ipi(cpuid, IA64_CMCP_VECTOR, IA64_IPI_DM_INT, 0);
1359 	} else {
1360 		/* If no log record, switch out of polling mode */
1361 		if (start_count == IA64_LOG_COUNT(SAL_INFO_TYPE_CMC)) {
1362 
1363 			printk(KERN_WARNING "Returning to interrupt driven CMC handler\n");
1364 			schedule_work(&cmc_enable_work);
1365 			cmc_polling_enabled = 0;
1366 
1367 		} else {
1368 
1369 			mod_timer(&cmc_poll_timer, jiffies + CMC_POLL_INTERVAL);
1370 		}
1371 
1372 		start_count = -1;
1373 	}
1374 
1375 	return IRQ_HANDLED;
1376 }
1377 
1378 /*
1379  *  ia64_mca_cmc_poll
1380  *
1381  *	Poll for Corrected Machine Checks (CMCs)
1382  *
1383  * Inputs   :   dummy(unused)
1384  * Outputs  :   None
1385  *
1386  */
1387 static void
1388 ia64_mca_cmc_poll (unsigned long dummy)
1389 {
1390 	/* Trigger a CMC interrupt cascade  */
1391 	platform_send_ipi(first_cpu(cpu_online_map), IA64_CMCP_VECTOR, IA64_IPI_DM_INT, 0);
1392 }
1393 
1394 /*
1395  *  ia64_mca_cpe_int_caller
1396  *
1397  * 	Triggered by sw interrupt from CPE polling routine.  Calls
1398  * 	real interrupt handler and either triggers a sw interrupt
1399  * 	on the next cpu or does cleanup at the end.
1400  *
1401  * Inputs
1402  *	interrupt number
1403  *	client data arg ptr
1404  * Outputs
1405  * 	handled
1406  */
1407 #ifdef CONFIG_ACPI
1408 
1409 static irqreturn_t
1410 ia64_mca_cpe_int_caller(int cpe_irq, void *arg)
1411 {
1412 	static int start_count = -1;
1413 	static int poll_time = MIN_CPE_POLL_INTERVAL;
1414 	unsigned int cpuid;
1415 
1416 	cpuid = smp_processor_id();
1417 
1418 	/* If first cpu, update count */
1419 	if (start_count == -1)
1420 		start_count = IA64_LOG_COUNT(SAL_INFO_TYPE_CPE);
1421 
1422 	ia64_mca_cpe_int_handler(cpe_irq, arg);
1423 
1424 	for (++cpuid ; cpuid < NR_CPUS && !cpu_online(cpuid) ; cpuid++);
1425 
1426 	if (cpuid < NR_CPUS) {
1427 		platform_send_ipi(cpuid, IA64_CPEP_VECTOR, IA64_IPI_DM_INT, 0);
1428 	} else {
1429 		/*
1430 		 * If a log was recorded, increase our polling frequency,
1431 		 * otherwise, backoff or return to interrupt mode.
1432 		 */
1433 		if (start_count != IA64_LOG_COUNT(SAL_INFO_TYPE_CPE)) {
1434 			poll_time = max(MIN_CPE_POLL_INTERVAL, poll_time / 2);
1435 		} else if (cpe_vector < 0) {
1436 			poll_time = min(MAX_CPE_POLL_INTERVAL, poll_time * 2);
1437 		} else {
1438 			poll_time = MIN_CPE_POLL_INTERVAL;
1439 
1440 			printk(KERN_WARNING "Returning to interrupt driven CPE handler\n");
1441 			enable_irq(local_vector_to_irq(IA64_CPE_VECTOR));
1442 			cpe_poll_enabled = 0;
1443 		}
1444 
1445 		if (cpe_poll_enabled)
1446 			mod_timer(&cpe_poll_timer, jiffies + poll_time);
1447 		start_count = -1;
1448 	}
1449 
1450 	return IRQ_HANDLED;
1451 }
1452 
1453 /*
1454  *  ia64_mca_cpe_poll
1455  *
1456  *	Poll for Corrected Platform Errors (CPEs), trigger interrupt
1457  *	on first cpu, from there it will trickle through all the cpus.
1458  *
1459  * Inputs   :   dummy(unused)
1460  * Outputs  :   None
1461  *
1462  */
1463 static void
1464 ia64_mca_cpe_poll (unsigned long dummy)
1465 {
1466 	/* Trigger a CPE interrupt cascade  */
1467 	platform_send_ipi(first_cpu(cpu_online_map), IA64_CPEP_VECTOR, IA64_IPI_DM_INT, 0);
1468 }
1469 
1470 #endif /* CONFIG_ACPI */
1471 
1472 static int
1473 default_monarch_init_process(struct notifier_block *self, unsigned long val, void *data)
1474 {
1475 	int c;
1476 	struct task_struct *g, *t;
1477 	if (val != DIE_INIT_MONARCH_PROCESS)
1478 		return NOTIFY_DONE;
1479 #ifdef CONFIG_KEXEC
1480 	if (atomic_read(&kdump_in_progress))
1481 		return NOTIFY_DONE;
1482 #endif
1483 
1484 	/*
1485 	 * FIXME: mlogbuf will brim over with INIT stack dumps.
1486 	 * To enable show_stack from INIT, we use oops_in_progress which should
1487 	 * be used in real oops. This would cause something wrong after INIT.
1488 	 */
1489 	BREAK_LOGLEVEL(console_loglevel);
1490 	ia64_mlogbuf_dump_from_init();
1491 
1492 	printk(KERN_ERR "Processes interrupted by INIT -");
1493 	for_each_online_cpu(c) {
1494 		struct ia64_sal_os_state *s;
1495 		t = __va(__per_cpu_mca[c] + IA64_MCA_CPU_INIT_STACK_OFFSET);
1496 		s = (struct ia64_sal_os_state *)((char *)t + MCA_SOS_OFFSET);
1497 		g = s->prev_task;
1498 		if (g) {
1499 			if (g->pid)
1500 				printk(" %d", g->pid);
1501 			else
1502 				printk(" %d (cpu %d task 0x%p)", g->pid, task_cpu(g), g);
1503 		}
1504 	}
1505 	printk("\n\n");
1506 	if (read_trylock(&tasklist_lock)) {
1507 		do_each_thread (g, t) {
1508 			printk("\nBacktrace of pid %d (%s)\n", t->pid, t->comm);
1509 			show_stack(t, NULL);
1510 		} while_each_thread (g, t);
1511 		read_unlock(&tasklist_lock);
1512 	}
1513 	/* FIXME: This will not restore zapped printk locks. */
1514 	RESTORE_LOGLEVEL(console_loglevel);
1515 	return NOTIFY_DONE;
1516 }
1517 
1518 /*
1519  * C portion of the OS INIT handler
1520  *
1521  * Called from ia64_os_init_dispatch
1522  *
1523  * Inputs: pointer to pt_regs where processor info was saved.  SAL/OS state for
1524  * this event.  This code is used for both monarch and slave INIT events, see
1525  * sos->monarch.
1526  *
1527  * All INIT events switch to the INIT stack and change the previous process to
1528  * blocked status.  If one of the INIT events is the monarch then we are
1529  * probably processing the nmi button/command.  Use the monarch cpu to dump all
1530  * the processes.  The slave INIT events all spin until the monarch cpu
1531  * returns.  We can also get INIT slave events for MCA, in which case the MCA
1532  * process is the monarch.
1533  */
1534 
1535 void
1536 ia64_init_handler(struct pt_regs *regs, struct switch_stack *sw,
1537 		  struct ia64_sal_os_state *sos)
1538 {
1539 	static atomic_t slaves;
1540 	static atomic_t monarchs;
1541 	struct task_struct *previous_current;
1542 	int cpu = smp_processor_id();
1543 	struct ia64_mca_notify_die nd =
1544 		{ .sos = sos, .monarch_cpu = &monarch_cpu };
1545 
1546 	(void) notify_die(DIE_INIT_ENTER, "INIT", regs, (long)&nd, 0, 0);
1547 
1548 	mprintk(KERN_INFO "Entered OS INIT handler. PSP=%lx cpu=%d monarch=%ld\n",
1549 		sos->proc_state_param, cpu, sos->monarch);
1550 	salinfo_log_wakeup(SAL_INFO_TYPE_INIT, NULL, 0, 0);
1551 
1552 	previous_current = ia64_mca_modify_original_stack(regs, sw, sos, "INIT");
1553 	sos->os_status = IA64_INIT_RESUME;
1554 
1555 	/* FIXME: Workaround for broken proms that drive all INIT events as
1556 	 * slaves.  The last slave that enters is promoted to be a monarch.
1557 	 * Remove this code in September 2006, that gives platforms a year to
1558 	 * fix their proms and get their customers updated.
1559 	 */
1560 	if (!sos->monarch && atomic_add_return(1, &slaves) == num_online_cpus()) {
1561 		mprintk(KERN_WARNING "%s: Promoting cpu %d to monarch.\n",
1562 		       __FUNCTION__, cpu);
1563 		atomic_dec(&slaves);
1564 		sos->monarch = 1;
1565 	}
1566 
1567 	/* FIXME: Workaround for broken proms that drive all INIT events as
1568 	 * monarchs.  Second and subsequent monarchs are demoted to slaves.
1569 	 * Remove this code in September 2006, that gives platforms a year to
1570 	 * fix their proms and get their customers updated.
1571 	 */
1572 	if (sos->monarch && atomic_add_return(1, &monarchs) > 1) {
1573 		mprintk(KERN_WARNING "%s: Demoting cpu %d to slave.\n",
1574 			       __FUNCTION__, cpu);
1575 		atomic_dec(&monarchs);
1576 		sos->monarch = 0;
1577 	}
1578 
1579 	if (!sos->monarch) {
1580 		ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_INIT;
1581 		while (monarch_cpu == -1)
1582 		       cpu_relax();	/* spin until monarch enters */
1583 		if (notify_die(DIE_INIT_SLAVE_ENTER, "INIT", regs, (long)&nd, 0, 0)
1584 				== NOTIFY_STOP)
1585 			ia64_mca_spin(__FUNCTION__);
1586 		if (notify_die(DIE_INIT_SLAVE_PROCESS, "INIT", regs, (long)&nd, 0, 0)
1587 				== NOTIFY_STOP)
1588 			ia64_mca_spin(__FUNCTION__);
1589 		while (monarch_cpu != -1)
1590 		       cpu_relax();	/* spin until monarch leaves */
1591 		if (notify_die(DIE_INIT_SLAVE_LEAVE, "INIT", regs, (long)&nd, 0, 0)
1592 				== NOTIFY_STOP)
1593 			ia64_mca_spin(__FUNCTION__);
1594 		mprintk("Slave on cpu %d returning to normal service.\n", cpu);
1595 		set_curr_task(cpu, previous_current);
1596 		ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1597 		atomic_dec(&slaves);
1598 		return;
1599 	}
1600 
1601 	monarch_cpu = cpu;
1602 	if (notify_die(DIE_INIT_MONARCH_ENTER, "INIT", regs, (long)&nd, 0, 0)
1603 			== NOTIFY_STOP)
1604 		ia64_mca_spin(__FUNCTION__);
1605 
1606 	/*
1607 	 * Wait for a bit.  On some machines (e.g., HP's zx2000 and zx6000, INIT can be
1608 	 * generated via the BMC's command-line interface, but since the console is on the
1609 	 * same serial line, the user will need some time to switch out of the BMC before
1610 	 * the dump begins.
1611 	 */
1612 	mprintk("Delaying for 5 seconds...\n");
1613 	udelay(5*1000000);
1614 	ia64_wait_for_slaves(cpu, "INIT");
1615 	/* If nobody intercepts DIE_INIT_MONARCH_PROCESS then we drop through
1616 	 * to default_monarch_init_process() above and just print all the
1617 	 * tasks.
1618 	 */
1619 	if (notify_die(DIE_INIT_MONARCH_PROCESS, "INIT", regs, (long)&nd, 0, 0)
1620 			== NOTIFY_STOP)
1621 		ia64_mca_spin(__FUNCTION__);
1622 	if (notify_die(DIE_INIT_MONARCH_LEAVE, "INIT", regs, (long)&nd, 0, 0)
1623 			== NOTIFY_STOP)
1624 		ia64_mca_spin(__FUNCTION__);
1625 	mprintk("\nINIT dump complete.  Monarch on cpu %d returning to normal service.\n", cpu);
1626 	atomic_dec(&monarchs);
1627 	set_curr_task(cpu, previous_current);
1628 	monarch_cpu = -1;
1629 	return;
1630 }
1631 
1632 static int __init
1633 ia64_mca_disable_cpe_polling(char *str)
1634 {
1635 	cpe_poll_enabled = 0;
1636 	return 1;
1637 }
1638 
1639 __setup("disable_cpe_poll", ia64_mca_disable_cpe_polling);
1640 
1641 static struct irqaction cmci_irqaction = {
1642 	.handler =	ia64_mca_cmc_int_handler,
1643 	.flags =	IRQF_DISABLED,
1644 	.name =		"cmc_hndlr"
1645 };
1646 
1647 static struct irqaction cmcp_irqaction = {
1648 	.handler =	ia64_mca_cmc_int_caller,
1649 	.flags =	IRQF_DISABLED,
1650 	.name =		"cmc_poll"
1651 };
1652 
1653 static struct irqaction mca_rdzv_irqaction = {
1654 	.handler =	ia64_mca_rendez_int_handler,
1655 	.flags =	IRQF_DISABLED,
1656 	.name =		"mca_rdzv"
1657 };
1658 
1659 static struct irqaction mca_wkup_irqaction = {
1660 	.handler =	ia64_mca_wakeup_int_handler,
1661 	.flags =	IRQF_DISABLED,
1662 	.name =		"mca_wkup"
1663 };
1664 
1665 #ifdef CONFIG_ACPI
1666 static struct irqaction mca_cpe_irqaction = {
1667 	.handler =	ia64_mca_cpe_int_handler,
1668 	.flags =	IRQF_DISABLED,
1669 	.name =		"cpe_hndlr"
1670 };
1671 
1672 static struct irqaction mca_cpep_irqaction = {
1673 	.handler =	ia64_mca_cpe_int_caller,
1674 	.flags =	IRQF_DISABLED,
1675 	.name =		"cpe_poll"
1676 };
1677 #endif /* CONFIG_ACPI */
1678 
1679 /* Minimal format of the MCA/INIT stacks.  The pseudo processes that run on
1680  * these stacks can never sleep, they cannot return from the kernel to user
1681  * space, they do not appear in a normal ps listing.  So there is no need to
1682  * format most of the fields.
1683  */
1684 
1685 static void __cpuinit
1686 format_mca_init_stack(void *mca_data, unsigned long offset,
1687 		const char *type, int cpu)
1688 {
1689 	struct task_struct *p = (struct task_struct *)((char *)mca_data + offset);
1690 	struct thread_info *ti;
1691 	memset(p, 0, KERNEL_STACK_SIZE);
1692 	ti = task_thread_info(p);
1693 	ti->flags = _TIF_MCA_INIT;
1694 	ti->preempt_count = 1;
1695 	ti->task = p;
1696 	ti->cpu = cpu;
1697 	p->stack = ti;
1698 	p->state = TASK_UNINTERRUPTIBLE;
1699 	cpu_set(cpu, p->cpus_allowed);
1700 	INIT_LIST_HEAD(&p->tasks);
1701 	p->parent = p->real_parent = p->group_leader = p;
1702 	INIT_LIST_HEAD(&p->children);
1703 	INIT_LIST_HEAD(&p->sibling);
1704 	strncpy(p->comm, type, sizeof(p->comm)-1);
1705 }
1706 
1707 /* Do per-CPU MCA-related initialization.  */
1708 
1709 void __cpuinit
1710 ia64_mca_cpu_init(void *cpu_data)
1711 {
1712 	void *pal_vaddr;
1713 	static int first_time = 1;
1714 
1715 	if (first_time) {
1716 		void *mca_data;
1717 		int cpu;
1718 
1719 		first_time = 0;
1720 		mca_data = alloc_bootmem(sizeof(struct ia64_mca_cpu)
1721 					 * NR_CPUS + KERNEL_STACK_SIZE);
1722 		mca_data = (void *)(((unsigned long)mca_data +
1723 					KERNEL_STACK_SIZE - 1) &
1724 				(-KERNEL_STACK_SIZE));
1725 		for (cpu = 0; cpu < NR_CPUS; cpu++) {
1726 			format_mca_init_stack(mca_data,
1727 					offsetof(struct ia64_mca_cpu, mca_stack),
1728 					"MCA", cpu);
1729 			format_mca_init_stack(mca_data,
1730 					offsetof(struct ia64_mca_cpu, init_stack),
1731 					"INIT", cpu);
1732 			__per_cpu_mca[cpu] = __pa(mca_data);
1733 			mca_data += sizeof(struct ia64_mca_cpu);
1734 		}
1735 	}
1736 
1737 	/*
1738 	 * The MCA info structure was allocated earlier and its
1739 	 * physical address saved in __per_cpu_mca[cpu].  Copy that
1740 	 * address * to ia64_mca_data so we can access it as a per-CPU
1741 	 * variable.
1742 	 */
1743 	__get_cpu_var(ia64_mca_data) = __per_cpu_mca[smp_processor_id()];
1744 
1745 	/*
1746 	 * Stash away a copy of the PTE needed to map the per-CPU page.
1747 	 * We may need it during MCA recovery.
1748 	 */
1749 	__get_cpu_var(ia64_mca_per_cpu_pte) =
1750 		pte_val(mk_pte_phys(__pa(cpu_data), PAGE_KERNEL));
1751 
1752 	/*
1753 	 * Also, stash away a copy of the PAL address and the PTE
1754 	 * needed to map it.
1755 	 */
1756 	pal_vaddr = efi_get_pal_addr();
1757 	if (!pal_vaddr)
1758 		return;
1759 	__get_cpu_var(ia64_mca_pal_base) =
1760 		GRANULEROUNDDOWN((unsigned long) pal_vaddr);
1761 	__get_cpu_var(ia64_mca_pal_pte) = pte_val(mk_pte_phys(__pa(pal_vaddr),
1762 							      PAGE_KERNEL));
1763 }
1764 
1765 /*
1766  * ia64_mca_init
1767  *
1768  *  Do all the system level mca specific initialization.
1769  *
1770  *	1. Register spinloop and wakeup request interrupt vectors
1771  *
1772  *	2. Register OS_MCA handler entry point
1773  *
1774  *	3. Register OS_INIT handler entry point
1775  *
1776  *  4. Initialize MCA/CMC/INIT related log buffers maintained by the OS.
1777  *
1778  *  Note that this initialization is done very early before some kernel
1779  *  services are available.
1780  *
1781  *  Inputs  :   None
1782  *
1783  *  Outputs :   None
1784  */
1785 void __init
1786 ia64_mca_init(void)
1787 {
1788 	ia64_fptr_t *init_hldlr_ptr_monarch = (ia64_fptr_t *)ia64_os_init_dispatch_monarch;
1789 	ia64_fptr_t *init_hldlr_ptr_slave = (ia64_fptr_t *)ia64_os_init_dispatch_slave;
1790 	ia64_fptr_t *mca_hldlr_ptr = (ia64_fptr_t *)ia64_os_mca_dispatch;
1791 	int i;
1792 	s64 rc;
1793 	struct ia64_sal_retval isrv;
1794 	u64 timeout = IA64_MCA_RENDEZ_TIMEOUT;	/* platform specific */
1795 	static struct notifier_block default_init_monarch_nb = {
1796 		.notifier_call = default_monarch_init_process,
1797 		.priority = 0/* we need to notified last */
1798 	};
1799 
1800 	IA64_MCA_DEBUG("%s: begin\n", __FUNCTION__);
1801 
1802 	/* Clear the Rendez checkin flag for all cpus */
1803 	for(i = 0 ; i < NR_CPUS; i++)
1804 		ia64_mc_info.imi_rendez_checkin[i] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1805 
1806 	/*
1807 	 * Register the rendezvous spinloop and wakeup mechanism with SAL
1808 	 */
1809 
1810 	/* Register the rendezvous interrupt vector with SAL */
1811 	while (1) {
1812 		isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_RENDEZ_INT,
1813 					      SAL_MC_PARAM_MECHANISM_INT,
1814 					      IA64_MCA_RENDEZ_VECTOR,
1815 					      timeout,
1816 					      SAL_MC_PARAM_RZ_ALWAYS);
1817 		rc = isrv.status;
1818 		if (rc == 0)
1819 			break;
1820 		if (rc == -2) {
1821 			printk(KERN_INFO "Increasing MCA rendezvous timeout from "
1822 				"%ld to %ld milliseconds\n", timeout, isrv.v0);
1823 			timeout = isrv.v0;
1824 			(void) notify_die(DIE_MCA_NEW_TIMEOUT, "MCA", NULL, timeout, 0, 0);
1825 			continue;
1826 		}
1827 		printk(KERN_ERR "Failed to register rendezvous interrupt "
1828 		       "with SAL (status %ld)\n", rc);
1829 		return;
1830 	}
1831 
1832 	/* Register the wakeup interrupt vector with SAL */
1833 	isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_RENDEZ_WAKEUP,
1834 				      SAL_MC_PARAM_MECHANISM_INT,
1835 				      IA64_MCA_WAKEUP_VECTOR,
1836 				      0, 0);
1837 	rc = isrv.status;
1838 	if (rc) {
1839 		printk(KERN_ERR "Failed to register wakeup interrupt with SAL "
1840 		       "(status %ld)\n", rc);
1841 		return;
1842 	}
1843 
1844 	IA64_MCA_DEBUG("%s: registered MCA rendezvous spinloop and wakeup mech.\n", __FUNCTION__);
1845 
1846 	ia64_mc_info.imi_mca_handler        = ia64_tpa(mca_hldlr_ptr->fp);
1847 	/*
1848 	 * XXX - disable SAL checksum by setting size to 0; should be
1849 	 *	ia64_tpa(ia64_os_mca_dispatch_end) - ia64_tpa(ia64_os_mca_dispatch);
1850 	 */
1851 	ia64_mc_info.imi_mca_handler_size	= 0;
1852 
1853 	/* Register the os mca handler with SAL */
1854 	if ((rc = ia64_sal_set_vectors(SAL_VECTOR_OS_MCA,
1855 				       ia64_mc_info.imi_mca_handler,
1856 				       ia64_tpa(mca_hldlr_ptr->gp),
1857 				       ia64_mc_info.imi_mca_handler_size,
1858 				       0, 0, 0)))
1859 	{
1860 		printk(KERN_ERR "Failed to register OS MCA handler with SAL "
1861 		       "(status %ld)\n", rc);
1862 		return;
1863 	}
1864 
1865 	IA64_MCA_DEBUG("%s: registered OS MCA handler with SAL at 0x%lx, gp = 0x%lx\n", __FUNCTION__,
1866 		       ia64_mc_info.imi_mca_handler, ia64_tpa(mca_hldlr_ptr->gp));
1867 
1868 	/*
1869 	 * XXX - disable SAL checksum by setting size to 0, should be
1870 	 * size of the actual init handler in mca_asm.S.
1871 	 */
1872 	ia64_mc_info.imi_monarch_init_handler		= ia64_tpa(init_hldlr_ptr_monarch->fp);
1873 	ia64_mc_info.imi_monarch_init_handler_size	= 0;
1874 	ia64_mc_info.imi_slave_init_handler		= ia64_tpa(init_hldlr_ptr_slave->fp);
1875 	ia64_mc_info.imi_slave_init_handler_size	= 0;
1876 
1877 	IA64_MCA_DEBUG("%s: OS INIT handler at %lx\n", __FUNCTION__,
1878 		       ia64_mc_info.imi_monarch_init_handler);
1879 
1880 	/* Register the os init handler with SAL */
1881 	if ((rc = ia64_sal_set_vectors(SAL_VECTOR_OS_INIT,
1882 				       ia64_mc_info.imi_monarch_init_handler,
1883 				       ia64_tpa(ia64_getreg(_IA64_REG_GP)),
1884 				       ia64_mc_info.imi_monarch_init_handler_size,
1885 				       ia64_mc_info.imi_slave_init_handler,
1886 				       ia64_tpa(ia64_getreg(_IA64_REG_GP)),
1887 				       ia64_mc_info.imi_slave_init_handler_size)))
1888 	{
1889 		printk(KERN_ERR "Failed to register m/s INIT handlers with SAL "
1890 		       "(status %ld)\n", rc);
1891 		return;
1892 	}
1893 	if (register_die_notifier(&default_init_monarch_nb)) {
1894 		printk(KERN_ERR "Failed to register default monarch INIT process\n");
1895 		return;
1896 	}
1897 
1898 	IA64_MCA_DEBUG("%s: registered OS INIT handler with SAL\n", __FUNCTION__);
1899 
1900 	/*
1901 	 *  Configure the CMCI/P vector and handler. Interrupts for CMC are
1902 	 *  per-processor, so AP CMC interrupts are setup in smp_callin() (smpboot.c).
1903 	 */
1904 	register_percpu_irq(IA64_CMC_VECTOR, &cmci_irqaction);
1905 	register_percpu_irq(IA64_CMCP_VECTOR, &cmcp_irqaction);
1906 	ia64_mca_cmc_vector_setup();       /* Setup vector on BSP */
1907 
1908 	/* Setup the MCA rendezvous interrupt vector */
1909 	register_percpu_irq(IA64_MCA_RENDEZ_VECTOR, &mca_rdzv_irqaction);
1910 
1911 	/* Setup the MCA wakeup interrupt vector */
1912 	register_percpu_irq(IA64_MCA_WAKEUP_VECTOR, &mca_wkup_irqaction);
1913 
1914 #ifdef CONFIG_ACPI
1915 	/* Setup the CPEI/P handler */
1916 	register_percpu_irq(IA64_CPEP_VECTOR, &mca_cpep_irqaction);
1917 #endif
1918 
1919 	/* Initialize the areas set aside by the OS to buffer the
1920 	 * platform/processor error states for MCA/INIT/CMC
1921 	 * handling.
1922 	 */
1923 	ia64_log_init(SAL_INFO_TYPE_MCA);
1924 	ia64_log_init(SAL_INFO_TYPE_INIT);
1925 	ia64_log_init(SAL_INFO_TYPE_CMC);
1926 	ia64_log_init(SAL_INFO_TYPE_CPE);
1927 
1928 	mca_init = 1;
1929 	printk(KERN_INFO "MCA related initialization done\n");
1930 }
1931 
1932 /*
1933  * ia64_mca_late_init
1934  *
1935  *	Opportunity to setup things that require initialization later
1936  *	than ia64_mca_init.  Setup a timer to poll for CPEs if the
1937  *	platform doesn't support an interrupt driven mechanism.
1938  *
1939  *  Inputs  :   None
1940  *  Outputs :   Status
1941  */
1942 static int __init
1943 ia64_mca_late_init(void)
1944 {
1945 	if (!mca_init)
1946 		return 0;
1947 
1948 	/* Setup the CMCI/P vector and handler */
1949 	init_timer(&cmc_poll_timer);
1950 	cmc_poll_timer.function = ia64_mca_cmc_poll;
1951 
1952 	/* Unmask/enable the vector */
1953 	cmc_polling_enabled = 0;
1954 	schedule_work(&cmc_enable_work);
1955 
1956 	IA64_MCA_DEBUG("%s: CMCI/P setup and enabled.\n", __FUNCTION__);
1957 
1958 #ifdef CONFIG_ACPI
1959 	/* Setup the CPEI/P vector and handler */
1960 	cpe_vector = acpi_request_vector(ACPI_INTERRUPT_CPEI);
1961 	init_timer(&cpe_poll_timer);
1962 	cpe_poll_timer.function = ia64_mca_cpe_poll;
1963 
1964 	{
1965 		irq_desc_t *desc;
1966 		unsigned int irq;
1967 
1968 		if (cpe_vector >= 0) {
1969 			/* If platform supports CPEI, enable the irq. */
1970 			cpe_poll_enabled = 0;
1971 			for (irq = 0; irq < NR_IRQS; ++irq)
1972 				if (irq_to_vector(irq) == cpe_vector) {
1973 					desc = irq_desc + irq;
1974 					desc->status |= IRQ_PER_CPU;
1975 					setup_irq(irq, &mca_cpe_irqaction);
1976 					ia64_cpe_irq = irq;
1977 				}
1978 			ia64_mca_register_cpev(cpe_vector);
1979 			IA64_MCA_DEBUG("%s: CPEI/P setup and enabled.\n", __FUNCTION__);
1980 		} else {
1981 			/* If platform doesn't support CPEI, get the timer going. */
1982 			if (cpe_poll_enabled) {
1983 				ia64_mca_cpe_poll(0UL);
1984 				IA64_MCA_DEBUG("%s: CPEP setup and enabled.\n", __FUNCTION__);
1985 			}
1986 		}
1987 	}
1988 #endif
1989 
1990 	return 0;
1991 }
1992 
1993 device_initcall(ia64_mca_late_init);
1994