1/*
2 * File:	mca.c
3 * Purpose:	Generic MCA handling layer
4 *
5 * Copyright (C) 2003 Hewlett-Packard Co
6 *	David Mosberger-Tang <davidm@hpl.hp.com>
7 *
8 * Copyright (C) 2002 Dell Inc.
9 * Copyright (C) Matt Domsch <Matt_Domsch@dell.com>
10 *
11 * Copyright (C) 2002 Intel
12 * Copyright (C) Jenna Hall <jenna.s.hall@intel.com>
13 *
14 * Copyright (C) 2001 Intel
15 * Copyright (C) Fred Lewis <frederick.v.lewis@intel.com>
16 *
17 * Copyright (C) 2000 Intel
18 * Copyright (C) Chuck Fleckenstein <cfleck@co.intel.com>
19 *
20 * Copyright (C) 1999, 2004-2008 Silicon Graphics, Inc.
21 * Copyright (C) Vijay Chander <vijay@engr.sgi.com>
22 *
23 * Copyright (C) 2006 FUJITSU LIMITED
24 * Copyright (C) Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
25 *
26 * 2000-03-29 Chuck Fleckenstein <cfleck@co.intel.com>
27 *	      Fixed PAL/SAL update issues, began MCA bug fixes, logging issues,
28 *	      added min save state dump, added INIT handler.
29 *
30 * 2001-01-03 Fred Lewis <frederick.v.lewis@intel.com>
31 *	      Added setup of CMCI and CPEI IRQs, logging of corrected platform
32 *	      errors, completed code for logging of corrected & uncorrected
33 *	      machine check errors, and updated for conformance with Nov. 2000
34 *	      revision of the SAL 3.0 spec.
35 *
36 * 2002-01-04 Jenna Hall <jenna.s.hall@intel.com>
37 *	      Aligned MCA stack to 16 bytes, added platform vs. CPU error flag,
38 *	      set SAL default return values, changed error record structure to
39 *	      linked list, added init call to sal_get_state_info_size().
40 *
41 * 2002-03-25 Matt Domsch <Matt_Domsch@dell.com>
42 *	      GUID cleanups.
43 *
44 * 2003-04-15 David Mosberger-Tang <davidm@hpl.hp.com>
45 *	      Added INIT backtrace support.
46 *
47 * 2003-12-08 Keith Owens <kaos@sgi.com>
48 *	      smp_call_function() must not be called from interrupt context
49 *	      (can deadlock on tasklist_lock).
50 *	      Use keventd to call smp_call_function().
51 *
52 * 2004-02-01 Keith Owens <kaos@sgi.com>
53 *	      Avoid deadlock when using printk() for MCA and INIT records.
54 *	      Delete all record printing code, moved to salinfo_decode in user
55 *	      space.  Mark variables and functions static where possible.
56 *	      Delete dead variables and functions.  Reorder to remove the need
57 *	      for forward declarations and to consolidate related code.
58 *
59 * 2005-08-12 Keith Owens <kaos@sgi.com>
60 *	      Convert MCA/INIT handlers to use per event stacks and SAL/OS
61 *	      state.
62 *
63 * 2005-10-07 Keith Owens <kaos@sgi.com>
64 *	      Add notify_die() hooks.
65 *
66 * 2006-09-15 Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
67 *	      Add printing support for MCA/INIT.
68 *
69 * 2007-04-27 Russ Anderson <rja@sgi.com>
70 *	      Support multiple cpus going through OS_MCA in the same event.
71 */
72#include <linux/jiffies.h>
73#include <linux/types.h>
74#include <linux/init.h>
75#include <linux/sched.h>
76#include <linux/interrupt.h>
77#include <linux/irq.h>
78#include <linux/bootmem.h>
79#include <linux/acpi.h>
80#include <linux/timer.h>
81#include <linux/module.h>
82#include <linux/kernel.h>
83#include <linux/smp.h>
84#include <linux/workqueue.h>
85#include <linux/cpumask.h>
86#include <linux/kdebug.h>
87#include <linux/cpu.h>
88#include <linux/gfp.h>
89
90#include <asm/delay.h>
91#include <asm/machvec.h>
92#include <asm/meminit.h>
93#include <asm/page.h>
94#include <asm/ptrace.h>
95#include <asm/sal.h>
96#include <asm/mca.h>
97#include <asm/kexec.h>
98
99#include <asm/irq.h>
100#include <asm/hw_irq.h>
101#include <asm/tlb.h>
102
103#include "mca_drv.h"
104#include "entry.h"
105
106#if defined(IA64_MCA_DEBUG_INFO)
107# define IA64_MCA_DEBUG(fmt...)	printk(fmt)
108#else
109# define IA64_MCA_DEBUG(fmt...)
110#endif
111
112#define NOTIFY_INIT(event, regs, arg, spin)				\
113do {									\
114	if ((notify_die((event), "INIT", (regs), (arg), 0, 0)		\
115			== NOTIFY_STOP) && ((spin) == 1))		\
116		ia64_mca_spin(__func__);				\
117} while (0)
118
119#define NOTIFY_MCA(event, regs, arg, spin)				\
120do {									\
121	if ((notify_die((event), "MCA", (regs), (arg), 0, 0)		\
122			== NOTIFY_STOP) && ((spin) == 1))		\
123		ia64_mca_spin(__func__);				\
124} while (0)
125
126/* Used by mca_asm.S */
127DEFINE_PER_CPU(u64, ia64_mca_data); /* == __per_cpu_mca[smp_processor_id()] */
128DEFINE_PER_CPU(u64, ia64_mca_per_cpu_pte); /* PTE to map per-CPU area */
129DEFINE_PER_CPU(u64, ia64_mca_pal_pte);	    /* PTE to map PAL code */
130DEFINE_PER_CPU(u64, ia64_mca_pal_base);    /* vaddr PAL code granule */
131DEFINE_PER_CPU(u64, ia64_mca_tr_reload);   /* Flag for TR reload */
132
133unsigned long __per_cpu_mca[NR_CPUS];
134
135/* In mca_asm.S */
136extern void			ia64_os_init_dispatch_monarch (void);
137extern void			ia64_os_init_dispatch_slave (void);
138
139static int monarch_cpu = -1;
140
141static ia64_mc_info_t		ia64_mc_info;
142
143#define MAX_CPE_POLL_INTERVAL (15*60*HZ) /* 15 minutes */
144#define MIN_CPE_POLL_INTERVAL (2*60*HZ)  /* 2 minutes */
145#define CMC_POLL_INTERVAL     (1*60*HZ)  /* 1 minute */
146#define CPE_HISTORY_LENGTH    5
147#define CMC_HISTORY_LENGTH    5
148
149#ifdef CONFIG_ACPI
150static struct timer_list cpe_poll_timer;
151#endif
152static struct timer_list cmc_poll_timer;
153/*
154 * This variable tells whether we are currently in polling mode.
155 * Start with this in the wrong state so we won't play w/ timers
156 * before the system is ready.
157 */
158static int cmc_polling_enabled = 1;
159
160/*
161 * Clearing this variable prevents CPE polling from getting activated
162 * in mca_late_init.  Use it if your system doesn't provide a CPEI,
163 * but encounters problems retrieving CPE logs.  This should only be
164 * necessary for debugging.
165 */
166static int cpe_poll_enabled = 1;
167
168extern void salinfo_log_wakeup(int type, u8 *buffer, u64 size, int irqsafe);
169
170static int mca_init __initdata;
171
172/*
173 * limited & delayed printing support for MCA/INIT handler
174 */
175
176#define mprintk(fmt...) ia64_mca_printk(fmt)
177
178#define MLOGBUF_SIZE (512+256*NR_CPUS)
179#define MLOGBUF_MSGMAX 256
180static char mlogbuf[MLOGBUF_SIZE];
181static DEFINE_SPINLOCK(mlogbuf_wlock);	/* mca context only */
182static DEFINE_SPINLOCK(mlogbuf_rlock);	/* normal context only */
183static unsigned long mlogbuf_start;
184static unsigned long mlogbuf_end;
185static unsigned int mlogbuf_finished = 0;
186static unsigned long mlogbuf_timestamp = 0;
187
188static int loglevel_save = -1;
189#define BREAK_LOGLEVEL(__console_loglevel)		\
190	oops_in_progress = 1;				\
191	if (loglevel_save < 0)				\
192		loglevel_save = __console_loglevel;	\
193	__console_loglevel = 15;
194
195#define RESTORE_LOGLEVEL(__console_loglevel)		\
196	if (loglevel_save >= 0) {			\
197		__console_loglevel = loglevel_save;	\
198		loglevel_save = -1;			\
199	}						\
200	mlogbuf_finished = 0;				\
201	oops_in_progress = 0;
202
203/*
204 * Push messages into buffer, print them later if not urgent.
205 */
206void ia64_mca_printk(const char *fmt, ...)
207{
208	va_list args;
209	int printed_len;
210	char temp_buf[MLOGBUF_MSGMAX];
211	char *p;
212
213	va_start(args, fmt);
214	printed_len = vscnprintf(temp_buf, sizeof(temp_buf), fmt, args);
215	va_end(args);
216
217	/* Copy the output into mlogbuf */
218	if (oops_in_progress) {
219		/* mlogbuf was abandoned, use printk directly instead. */
220		printk("%s", temp_buf);
221	} else {
222		spin_lock(&mlogbuf_wlock);
223		for (p = temp_buf; *p; p++) {
224			unsigned long next = (mlogbuf_end + 1) % MLOGBUF_SIZE;
225			if (next != mlogbuf_start) {
226				mlogbuf[mlogbuf_end] = *p;
227				mlogbuf_end = next;
228			} else {
229				/* buffer full */
230				break;
231			}
232		}
233		mlogbuf[mlogbuf_end] = '\0';
234		spin_unlock(&mlogbuf_wlock);
235	}
236}
237EXPORT_SYMBOL(ia64_mca_printk);
238
239/*
240 * Print buffered messages.
241 *  NOTE: call this after returning normal context. (ex. from salinfod)
242 */
243void ia64_mlogbuf_dump(void)
244{
245	char temp_buf[MLOGBUF_MSGMAX];
246	char *p;
247	unsigned long index;
248	unsigned long flags;
249	unsigned int printed_len;
250
251	/* Get output from mlogbuf */
252	while (mlogbuf_start != mlogbuf_end) {
253		temp_buf[0] = '\0';
254		p = temp_buf;
255		printed_len = 0;
256
257		spin_lock_irqsave(&mlogbuf_rlock, flags);
258
259		index = mlogbuf_start;
260		while (index != mlogbuf_end) {
261			*p = mlogbuf[index];
262			index = (index + 1) % MLOGBUF_SIZE;
263			if (!*p)
264				break;
265			p++;
266			if (++printed_len >= MLOGBUF_MSGMAX - 1)
267				break;
268		}
269		*p = '\0';
270		if (temp_buf[0])
271			printk("%s", temp_buf);
272		mlogbuf_start = index;
273
274		mlogbuf_timestamp = 0;
275		spin_unlock_irqrestore(&mlogbuf_rlock, flags);
276	}
277}
278EXPORT_SYMBOL(ia64_mlogbuf_dump);
279
280/*
281 * Call this if system is going to down or if immediate flushing messages to
282 * console is required. (ex. recovery was failed, crash dump is going to be
283 * invoked, long-wait rendezvous etc.)
284 *  NOTE: this should be called from monarch.
285 */
286static void ia64_mlogbuf_finish(int wait)
287{
288	BREAK_LOGLEVEL(console_loglevel);
289
290	spin_lock_init(&mlogbuf_rlock);
291	ia64_mlogbuf_dump();
292	printk(KERN_EMERG "mlogbuf_finish: printing switched to urgent mode, "
293		"MCA/INIT might be dodgy or fail.\n");
294
295	if (!wait)
296		return;
297
298	/* wait for console */
299	printk("Delaying for 5 seconds...\n");
300	udelay(5*1000000);
301
302	mlogbuf_finished = 1;
303}
304
305/*
306 * Print buffered messages from INIT context.
307 */
308static void ia64_mlogbuf_dump_from_init(void)
309{
310	if (mlogbuf_finished)
311		return;
312
313	if (mlogbuf_timestamp &&
314			time_before(jiffies, mlogbuf_timestamp + 30 * HZ)) {
315		printk(KERN_ERR "INIT: mlogbuf_dump is interrupted by INIT "
316			" and the system seems to be messed up.\n");
317		ia64_mlogbuf_finish(0);
318		return;
319	}
320
321	if (!spin_trylock(&mlogbuf_rlock)) {
322		printk(KERN_ERR "INIT: mlogbuf_dump is interrupted by INIT. "
323			"Generated messages other than stack dump will be "
324			"buffered to mlogbuf and will be printed later.\n");
325		printk(KERN_ERR "INIT: If messages would not printed after "
326			"this INIT, wait 30sec and assert INIT again.\n");
327		if (!mlogbuf_timestamp)
328			mlogbuf_timestamp = jiffies;
329		return;
330	}
331	spin_unlock(&mlogbuf_rlock);
332	ia64_mlogbuf_dump();
333}
334
335static void inline
336ia64_mca_spin(const char *func)
337{
338	if (monarch_cpu == smp_processor_id())
339		ia64_mlogbuf_finish(0);
340	mprintk(KERN_EMERG "%s: spinning here, not returning to SAL\n", func);
341	while (1)
342		cpu_relax();
343}
344/*
345 * IA64_MCA log support
346 */
347#define IA64_MAX_LOGS		2	/* Double-buffering for nested MCAs */
348#define IA64_MAX_LOG_TYPES      4   /* MCA, INIT, CMC, CPE */
349
350typedef struct ia64_state_log_s
351{
352	spinlock_t	isl_lock;
353	int		isl_index;
354	unsigned long	isl_count;
355	ia64_err_rec_t  *isl_log[IA64_MAX_LOGS]; /* need space to store header + error log */
356} ia64_state_log_t;
357
358static ia64_state_log_t ia64_state_log[IA64_MAX_LOG_TYPES];
359
360#define IA64_LOG_ALLOCATE(it, size) \
361	{ia64_state_log[it].isl_log[IA64_LOG_CURR_INDEX(it)] = \
362		(ia64_err_rec_t *)alloc_bootmem(size); \
363	ia64_state_log[it].isl_log[IA64_LOG_NEXT_INDEX(it)] = \
364		(ia64_err_rec_t *)alloc_bootmem(size);}
365#define IA64_LOG_LOCK_INIT(it) spin_lock_init(&ia64_state_log[it].isl_lock)
366#define IA64_LOG_LOCK(it)      spin_lock_irqsave(&ia64_state_log[it].isl_lock, s)
367#define IA64_LOG_UNLOCK(it)    spin_unlock_irqrestore(&ia64_state_log[it].isl_lock,s)
368#define IA64_LOG_NEXT_INDEX(it)    ia64_state_log[it].isl_index
369#define IA64_LOG_CURR_INDEX(it)    1 - ia64_state_log[it].isl_index
370#define IA64_LOG_INDEX_INC(it) \
371    {ia64_state_log[it].isl_index = 1 - ia64_state_log[it].isl_index; \
372    ia64_state_log[it].isl_count++;}
373#define IA64_LOG_INDEX_DEC(it) \
374    ia64_state_log[it].isl_index = 1 - ia64_state_log[it].isl_index
375#define IA64_LOG_NEXT_BUFFER(it)   (void *)((ia64_state_log[it].isl_log[IA64_LOG_NEXT_INDEX(it)]))
376#define IA64_LOG_CURR_BUFFER(it)   (void *)((ia64_state_log[it].isl_log[IA64_LOG_CURR_INDEX(it)]))
377#define IA64_LOG_COUNT(it)         ia64_state_log[it].isl_count
378
379/*
380 * ia64_log_init
381 *	Reset the OS ia64 log buffer
382 * Inputs   :   info_type   (SAL_INFO_TYPE_{MCA,INIT,CMC,CPE})
383 * Outputs	:	None
384 */
385static void __init
386ia64_log_init(int sal_info_type)
387{
388	u64	max_size = 0;
389
390	IA64_LOG_NEXT_INDEX(sal_info_type) = 0;
391	IA64_LOG_LOCK_INIT(sal_info_type);
392
393	// SAL will tell us the maximum size of any error record of this type
394	max_size = ia64_sal_get_state_info_size(sal_info_type);
395	if (!max_size)
396		/* alloc_bootmem() doesn't like zero-sized allocations! */
397		return;
398
399	// set up OS data structures to hold error info
400	IA64_LOG_ALLOCATE(sal_info_type, max_size);
401	memset(IA64_LOG_CURR_BUFFER(sal_info_type), 0, max_size);
402	memset(IA64_LOG_NEXT_BUFFER(sal_info_type), 0, max_size);
403}
404
405/*
406 * ia64_log_get
407 *
408 *	Get the current MCA log from SAL and copy it into the OS log buffer.
409 *
410 *  Inputs  :   info_type   (SAL_INFO_TYPE_{MCA,INIT,CMC,CPE})
411 *              irq_safe    whether you can use printk at this point
412 *  Outputs :   size        (total record length)
413 *              *buffer     (ptr to error record)
414 *
415 */
416static u64
417ia64_log_get(int sal_info_type, u8 **buffer, int irq_safe)
418{
419	sal_log_record_header_t     *log_buffer;
420	u64                         total_len = 0;
421	unsigned long               s;
422
423	IA64_LOG_LOCK(sal_info_type);
424
425	/* Get the process state information */
426	log_buffer = IA64_LOG_NEXT_BUFFER(sal_info_type);
427
428	total_len = ia64_sal_get_state_info(sal_info_type, (u64 *)log_buffer);
429
430	if (total_len) {
431		IA64_LOG_INDEX_INC(sal_info_type);
432		IA64_LOG_UNLOCK(sal_info_type);
433		if (irq_safe) {
434			IA64_MCA_DEBUG("%s: SAL error record type %d retrieved. Record length = %ld\n",
435				       __func__, sal_info_type, total_len);
436		}
437		*buffer = (u8 *) log_buffer;
438		return total_len;
439	} else {
440		IA64_LOG_UNLOCK(sal_info_type);
441		return 0;
442	}
443}
444
445/*
446 *  ia64_mca_log_sal_error_record
447 *
448 *  This function retrieves a specified error record type from SAL
449 *  and wakes up any processes waiting for error records.
450 *
451 *  Inputs  :   sal_info_type   (Type of error record MCA/CMC/CPE)
452 *              FIXME: remove MCA and irq_safe.
453 */
454static void
455ia64_mca_log_sal_error_record(int sal_info_type)
456{
457	u8 *buffer;
458	sal_log_record_header_t *rh;
459	u64 size;
460	int irq_safe = sal_info_type != SAL_INFO_TYPE_MCA;
461#ifdef IA64_MCA_DEBUG_INFO
462	static const char * const rec_name[] = { "MCA", "INIT", "CMC", "CPE" };
463#endif
464
465	size = ia64_log_get(sal_info_type, &buffer, irq_safe);
466	if (!size)
467		return;
468
469	salinfo_log_wakeup(sal_info_type, buffer, size, irq_safe);
470
471	if (irq_safe)
472		IA64_MCA_DEBUG("CPU %d: SAL log contains %s error record\n",
473			smp_processor_id(),
474			sal_info_type < ARRAY_SIZE(rec_name) ? rec_name[sal_info_type] : "UNKNOWN");
475
476	/* Clear logs from corrected errors in case there's no user-level logger */
477	rh = (sal_log_record_header_t *)buffer;
478	if (rh->severity == sal_log_severity_corrected)
479		ia64_sal_clear_state_info(sal_info_type);
480}
481
482/*
483 * search_mca_table
484 *  See if the MCA surfaced in an instruction range
485 *  that has been tagged as recoverable.
486 *
487 *  Inputs
488 *	first	First address range to check
489 *	last	Last address range to check
490 *	ip	Instruction pointer, address we are looking for
491 *
492 * Return value:
493 *      1 on Success (in the table)/ 0 on Failure (not in the  table)
494 */
495int
496search_mca_table (const struct mca_table_entry *first,
497                const struct mca_table_entry *last,
498                unsigned long ip)
499{
500        const struct mca_table_entry *curr;
501        u64 curr_start, curr_end;
502
503        curr = first;
504        while (curr <= last) {
505                curr_start = (u64) &curr->start_addr + curr->start_addr;
506                curr_end = (u64) &curr->end_addr + curr->end_addr;
507
508                if ((ip >= curr_start) && (ip <= curr_end)) {
509                        return 1;
510                }
511                curr++;
512        }
513        return 0;
514}
515
516/* Given an address, look for it in the mca tables. */
517int mca_recover_range(unsigned long addr)
518{
519	extern struct mca_table_entry __start___mca_table[];
520	extern struct mca_table_entry __stop___mca_table[];
521
522	return search_mca_table(__start___mca_table, __stop___mca_table-1, addr);
523}
524EXPORT_SYMBOL_GPL(mca_recover_range);
525
526#ifdef CONFIG_ACPI
527
528int cpe_vector = -1;
529int ia64_cpe_irq = -1;
530
531static irqreturn_t
532ia64_mca_cpe_int_handler (int cpe_irq, void *arg)
533{
534	static unsigned long	cpe_history[CPE_HISTORY_LENGTH];
535	static int		index;
536	static DEFINE_SPINLOCK(cpe_history_lock);
537
538	IA64_MCA_DEBUG("%s: received interrupt vector = %#x on CPU %d\n",
539		       __func__, cpe_irq, smp_processor_id());
540
541	/* SAL spec states this should run w/ interrupts enabled */
542	local_irq_enable();
543
544	spin_lock(&cpe_history_lock);
545	if (!cpe_poll_enabled && cpe_vector >= 0) {
546
547		int i, count = 1; /* we know 1 happened now */
548		unsigned long now = jiffies;
549
550		for (i = 0; i < CPE_HISTORY_LENGTH; i++) {
551			if (now - cpe_history[i] <= HZ)
552				count++;
553		}
554
555		IA64_MCA_DEBUG(KERN_INFO "CPE threshold %d/%d\n", count, CPE_HISTORY_LENGTH);
556		if (count >= CPE_HISTORY_LENGTH) {
557
558			cpe_poll_enabled = 1;
559			spin_unlock(&cpe_history_lock);
560			disable_irq_nosync(local_vector_to_irq(IA64_CPE_VECTOR));
561
562			/*
563			 * Corrected errors will still be corrected, but
564			 * make sure there's a log somewhere that indicates
565			 * something is generating more than we can handle.
566			 */
567			printk(KERN_WARNING "WARNING: Switching to polling CPE handler; error records may be lost\n");
568
569			mod_timer(&cpe_poll_timer, jiffies + MIN_CPE_POLL_INTERVAL);
570
571			/* lock already released, get out now */
572			goto out;
573		} else {
574			cpe_history[index++] = now;
575			if (index == CPE_HISTORY_LENGTH)
576				index = 0;
577		}
578	}
579	spin_unlock(&cpe_history_lock);
580out:
581	/* Get the CPE error record and log it */
582	ia64_mca_log_sal_error_record(SAL_INFO_TYPE_CPE);
583
584	local_irq_disable();
585
586	return IRQ_HANDLED;
587}
588
589#endif /* CONFIG_ACPI */
590
591#ifdef CONFIG_ACPI
592/*
593 * ia64_mca_register_cpev
594 *
595 *  Register the corrected platform error vector with SAL.
596 *
597 *  Inputs
598 *      cpev        Corrected Platform Error Vector number
599 *
600 *  Outputs
601 *      None
602 */
603void
604ia64_mca_register_cpev (int cpev)
605{
606	/* Register the CPE interrupt vector with SAL */
607	struct ia64_sal_retval isrv;
608
609	isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_CPE_INT, SAL_MC_PARAM_MECHANISM_INT, cpev, 0, 0);
610	if (isrv.status) {
611		printk(KERN_ERR "Failed to register Corrected Platform "
612		       "Error interrupt vector with SAL (status %ld)\n", isrv.status);
613		return;
614	}
615
616	IA64_MCA_DEBUG("%s: corrected platform error "
617		       "vector %#x registered\n", __func__, cpev);
618}
619#endif /* CONFIG_ACPI */
620
621/*
622 * ia64_mca_cmc_vector_setup
623 *
624 *  Setup the corrected machine check vector register in the processor.
625 *  (The interrupt is masked on boot. ia64_mca_late_init unmask this.)
626 *  This function is invoked on a per-processor basis.
627 *
628 * Inputs
629 *      None
630 *
631 * Outputs
632 *	None
633 */
634void
635ia64_mca_cmc_vector_setup (void)
636{
637	cmcv_reg_t	cmcv;
638
639	cmcv.cmcv_regval	= 0;
640	cmcv.cmcv_mask		= 1;        /* Mask/disable interrupt at first */
641	cmcv.cmcv_vector	= IA64_CMC_VECTOR;
642	ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
643
644	IA64_MCA_DEBUG("%s: CPU %d corrected machine check vector %#x registered.\n",
645		       __func__, smp_processor_id(), IA64_CMC_VECTOR);
646
647	IA64_MCA_DEBUG("%s: CPU %d CMCV = %#016lx\n",
648		       __func__, smp_processor_id(), ia64_getreg(_IA64_REG_CR_CMCV));
649}
650
651/*
652 * ia64_mca_cmc_vector_disable
653 *
654 *  Mask the corrected machine check vector register in the processor.
655 *  This function is invoked on a per-processor basis.
656 *
657 * Inputs
658 *      dummy(unused)
659 *
660 * Outputs
661 *	None
662 */
663static void
664ia64_mca_cmc_vector_disable (void *dummy)
665{
666	cmcv_reg_t	cmcv;
667
668	cmcv.cmcv_regval = ia64_getreg(_IA64_REG_CR_CMCV);
669
670	cmcv.cmcv_mask = 1; /* Mask/disable interrupt */
671	ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
672
673	IA64_MCA_DEBUG("%s: CPU %d corrected machine check vector %#x disabled.\n",
674		       __func__, smp_processor_id(), cmcv.cmcv_vector);
675}
676
677/*
678 * ia64_mca_cmc_vector_enable
679 *
680 *  Unmask the corrected machine check vector register in the processor.
681 *  This function is invoked on a per-processor basis.
682 *
683 * Inputs
684 *      dummy(unused)
685 *
686 * Outputs
687 *	None
688 */
689static void
690ia64_mca_cmc_vector_enable (void *dummy)
691{
692	cmcv_reg_t	cmcv;
693
694	cmcv.cmcv_regval = ia64_getreg(_IA64_REG_CR_CMCV);
695
696	cmcv.cmcv_mask = 0; /* Unmask/enable interrupt */
697	ia64_setreg(_IA64_REG_CR_CMCV, cmcv.cmcv_regval);
698
699	IA64_MCA_DEBUG("%s: CPU %d corrected machine check vector %#x enabled.\n",
700		       __func__, smp_processor_id(), cmcv.cmcv_vector);
701}
702
703/*
704 * ia64_mca_cmc_vector_disable_keventd
705 *
706 * Called via keventd (smp_call_function() is not safe in interrupt context) to
707 * disable the cmc interrupt vector.
708 */
709static void
710ia64_mca_cmc_vector_disable_keventd(struct work_struct *unused)
711{
712	on_each_cpu(ia64_mca_cmc_vector_disable, NULL, 0);
713}
714
715/*
716 * ia64_mca_cmc_vector_enable_keventd
717 *
718 * Called via keventd (smp_call_function() is not safe in interrupt context) to
719 * enable the cmc interrupt vector.
720 */
721static void
722ia64_mca_cmc_vector_enable_keventd(struct work_struct *unused)
723{
724	on_each_cpu(ia64_mca_cmc_vector_enable, NULL, 0);
725}
726
727/*
728 * ia64_mca_wakeup
729 *
730 *	Send an inter-cpu interrupt to wake-up a particular cpu.
731 *
732 *  Inputs  :   cpuid
733 *  Outputs :   None
734 */
735static void
736ia64_mca_wakeup(int cpu)
737{
738	platform_send_ipi(cpu, IA64_MCA_WAKEUP_VECTOR, IA64_IPI_DM_INT, 0);
739}
740
741/*
742 * ia64_mca_wakeup_all
743 *
744 *	Wakeup all the slave cpus which have rendez'ed previously.
745 *
746 *  Inputs  :   None
747 *  Outputs :   None
748 */
749static void
750ia64_mca_wakeup_all(void)
751{
752	int cpu;
753
754	/* Clear the Rendez checkin flag for all cpus */
755	for_each_online_cpu(cpu) {
756		if (ia64_mc_info.imi_rendez_checkin[cpu] == IA64_MCA_RENDEZ_CHECKIN_DONE)
757			ia64_mca_wakeup(cpu);
758	}
759
760}
761
762/*
763 * ia64_mca_rendez_interrupt_handler
764 *
765 *	This is handler used to put slave processors into spinloop
766 *	while the monarch processor does the mca handling and later
767 *	wake each slave up once the monarch is done.  The state
768 *	IA64_MCA_RENDEZ_CHECKIN_DONE indicates the cpu is rendez'ed
769 *	in SAL.  The state IA64_MCA_RENDEZ_CHECKIN_NOTDONE indicates
770 *	the cpu has come out of OS rendezvous.
771 *
772 *  Inputs  :   None
773 *  Outputs :   None
774 */
775static irqreturn_t
776ia64_mca_rendez_int_handler(int rendez_irq, void *arg)
777{
778	unsigned long flags;
779	int cpu = smp_processor_id();
780	struct ia64_mca_notify_die nd =
781		{ .sos = NULL, .monarch_cpu = &monarch_cpu };
782
783	/* Mask all interrupts */
784	local_irq_save(flags);
785
786	NOTIFY_MCA(DIE_MCA_RENDZVOUS_ENTER, get_irq_regs(), (long)&nd, 1);
787
788	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_DONE;
789	/* Register with the SAL monarch that the slave has
790	 * reached SAL
791	 */
792	ia64_sal_mc_rendez();
793
794	NOTIFY_MCA(DIE_MCA_RENDZVOUS_PROCESS, get_irq_regs(), (long)&nd, 1);
795
796	/* Wait for the monarch cpu to exit. */
797	while (monarch_cpu != -1)
798	       cpu_relax();	/* spin until monarch leaves */
799
800	NOTIFY_MCA(DIE_MCA_RENDZVOUS_LEAVE, get_irq_regs(), (long)&nd, 1);
801
802	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
803	/* Enable all interrupts */
804	local_irq_restore(flags);
805	return IRQ_HANDLED;
806}
807
808/*
809 * ia64_mca_wakeup_int_handler
810 *
811 *	The interrupt handler for processing the inter-cpu interrupt to the
812 *	slave cpu which was spinning in the rendez loop.
813 *	Since this spinning is done by turning off the interrupts and
814 *	polling on the wakeup-interrupt bit in the IRR, there is
815 *	nothing useful to be done in the handler.
816 *
817 *  Inputs  :   wakeup_irq  (Wakeup-interrupt bit)
818 *	arg		(Interrupt handler specific argument)
819 *  Outputs :   None
820 *
821 */
822static irqreturn_t
823ia64_mca_wakeup_int_handler(int wakeup_irq, void *arg)
824{
825	return IRQ_HANDLED;
826}
827
828/* Function pointer for extra MCA recovery */
829int (*ia64_mca_ucmc_extension)
830	(void*,struct ia64_sal_os_state*)
831	= NULL;
832
833int
834ia64_reg_MCA_extension(int (*fn)(void *, struct ia64_sal_os_state *))
835{
836	if (ia64_mca_ucmc_extension)
837		return 1;
838
839	ia64_mca_ucmc_extension = fn;
840	return 0;
841}
842
843void
844ia64_unreg_MCA_extension(void)
845{
846	if (ia64_mca_ucmc_extension)
847		ia64_mca_ucmc_extension = NULL;
848}
849
850EXPORT_SYMBOL(ia64_reg_MCA_extension);
851EXPORT_SYMBOL(ia64_unreg_MCA_extension);
852
853
854static inline void
855copy_reg(const u64 *fr, u64 fnat, unsigned long *tr, unsigned long *tnat)
856{
857	u64 fslot, tslot, nat;
858	*tr = *fr;
859	fslot = ((unsigned long)fr >> 3) & 63;
860	tslot = ((unsigned long)tr >> 3) & 63;
861	*tnat &= ~(1UL << tslot);
862	nat = (fnat >> fslot) & 1;
863	*tnat |= (nat << tslot);
864}
865
866/* Change the comm field on the MCA/INT task to include the pid that
867 * was interrupted, it makes for easier debugging.  If that pid was 0
868 * (swapper or nested MCA/INIT) then use the start of the previous comm
869 * field suffixed with its cpu.
870 */
871
872static void
873ia64_mca_modify_comm(const struct task_struct *previous_current)
874{
875	char *p, comm[sizeof(current->comm)];
876	if (previous_current->pid)
877		snprintf(comm, sizeof(comm), "%s %d",
878			current->comm, previous_current->pid);
879	else {
880		int l;
881		if ((p = strchr(previous_current->comm, ' ')))
882			l = p - previous_current->comm;
883		else
884			l = strlen(previous_current->comm);
885		snprintf(comm, sizeof(comm), "%s %*s %d",
886			current->comm, l, previous_current->comm,
887			task_thread_info(previous_current)->cpu);
888	}
889	memcpy(current->comm, comm, sizeof(current->comm));
890}
891
892static void
893finish_pt_regs(struct pt_regs *regs, struct ia64_sal_os_state *sos,
894		unsigned long *nat)
895{
896	const pal_min_state_area_t *ms = sos->pal_min_state;
897	const u64 *bank;
898
899	/* If ipsr.ic then use pmsa_{iip,ipsr,ifs}, else use
900	 * pmsa_{xip,xpsr,xfs}
901	 */
902	if (ia64_psr(regs)->ic) {
903		regs->cr_iip = ms->pmsa_iip;
904		regs->cr_ipsr = ms->pmsa_ipsr;
905		regs->cr_ifs = ms->pmsa_ifs;
906	} else {
907		regs->cr_iip = ms->pmsa_xip;
908		regs->cr_ipsr = ms->pmsa_xpsr;
909		regs->cr_ifs = ms->pmsa_xfs;
910
911		sos->iip = ms->pmsa_iip;
912		sos->ipsr = ms->pmsa_ipsr;
913		sos->ifs = ms->pmsa_ifs;
914	}
915	regs->pr = ms->pmsa_pr;
916	regs->b0 = ms->pmsa_br0;
917	regs->ar_rsc = ms->pmsa_rsc;
918	copy_reg(&ms->pmsa_gr[1-1], ms->pmsa_nat_bits, &regs->r1, nat);
919	copy_reg(&ms->pmsa_gr[2-1], ms->pmsa_nat_bits, &regs->r2, nat);
920	copy_reg(&ms->pmsa_gr[3-1], ms->pmsa_nat_bits, &regs->r3, nat);
921	copy_reg(&ms->pmsa_gr[8-1], ms->pmsa_nat_bits, &regs->r8, nat);
922	copy_reg(&ms->pmsa_gr[9-1], ms->pmsa_nat_bits, &regs->r9, nat);
923	copy_reg(&ms->pmsa_gr[10-1], ms->pmsa_nat_bits, &regs->r10, nat);
924	copy_reg(&ms->pmsa_gr[11-1], ms->pmsa_nat_bits, &regs->r11, nat);
925	copy_reg(&ms->pmsa_gr[12-1], ms->pmsa_nat_bits, &regs->r12, nat);
926	copy_reg(&ms->pmsa_gr[13-1], ms->pmsa_nat_bits, &regs->r13, nat);
927	copy_reg(&ms->pmsa_gr[14-1], ms->pmsa_nat_bits, &regs->r14, nat);
928	copy_reg(&ms->pmsa_gr[15-1], ms->pmsa_nat_bits, &regs->r15, nat);
929	if (ia64_psr(regs)->bn)
930		bank = ms->pmsa_bank1_gr;
931	else
932		bank = ms->pmsa_bank0_gr;
933	copy_reg(&bank[16-16], ms->pmsa_nat_bits, &regs->r16, nat);
934	copy_reg(&bank[17-16], ms->pmsa_nat_bits, &regs->r17, nat);
935	copy_reg(&bank[18-16], ms->pmsa_nat_bits, &regs->r18, nat);
936	copy_reg(&bank[19-16], ms->pmsa_nat_bits, &regs->r19, nat);
937	copy_reg(&bank[20-16], ms->pmsa_nat_bits, &regs->r20, nat);
938	copy_reg(&bank[21-16], ms->pmsa_nat_bits, &regs->r21, nat);
939	copy_reg(&bank[22-16], ms->pmsa_nat_bits, &regs->r22, nat);
940	copy_reg(&bank[23-16], ms->pmsa_nat_bits, &regs->r23, nat);
941	copy_reg(&bank[24-16], ms->pmsa_nat_bits, &regs->r24, nat);
942	copy_reg(&bank[25-16], ms->pmsa_nat_bits, &regs->r25, nat);
943	copy_reg(&bank[26-16], ms->pmsa_nat_bits, &regs->r26, nat);
944	copy_reg(&bank[27-16], ms->pmsa_nat_bits, &regs->r27, nat);
945	copy_reg(&bank[28-16], ms->pmsa_nat_bits, &regs->r28, nat);
946	copy_reg(&bank[29-16], ms->pmsa_nat_bits, &regs->r29, nat);
947	copy_reg(&bank[30-16], ms->pmsa_nat_bits, &regs->r30, nat);
948	copy_reg(&bank[31-16], ms->pmsa_nat_bits, &regs->r31, nat);
949}
950
951/* On entry to this routine, we are running on the per cpu stack, see
952 * mca_asm.h.  The original stack has not been touched by this event.  Some of
953 * the original stack's registers will be in the RBS on this stack.  This stack
954 * also contains a partial pt_regs and switch_stack, the rest of the data is in
955 * PAL minstate.
956 *
957 * The first thing to do is modify the original stack to look like a blocked
958 * task so we can run backtrace on the original task.  Also mark the per cpu
959 * stack as current to ensure that we use the correct task state, it also means
960 * that we can do backtrace on the MCA/INIT handler code itself.
961 */
962
963static struct task_struct *
964ia64_mca_modify_original_stack(struct pt_regs *regs,
965		const struct switch_stack *sw,
966		struct ia64_sal_os_state *sos,
967		const char *type)
968{
969	char *p;
970	ia64_va va;
971	extern char ia64_leave_kernel[];	/* Need asm address, not function descriptor */
972	const pal_min_state_area_t *ms = sos->pal_min_state;
973	struct task_struct *previous_current;
974	struct pt_regs *old_regs;
975	struct switch_stack *old_sw;
976	unsigned size = sizeof(struct pt_regs) +
977			sizeof(struct switch_stack) + 16;
978	unsigned long *old_bspstore, *old_bsp;
979	unsigned long *new_bspstore, *new_bsp;
980	unsigned long old_unat, old_rnat, new_rnat, nat;
981	u64 slots, loadrs = regs->loadrs;
982	u64 r12 = ms->pmsa_gr[12-1], r13 = ms->pmsa_gr[13-1];
983	u64 ar_bspstore = regs->ar_bspstore;
984	u64 ar_bsp = regs->ar_bspstore + (loadrs >> 16);
985	const char *msg;
986	int cpu = smp_processor_id();
987
988	previous_current = curr_task(cpu);
989	set_curr_task(cpu, current);
990	if ((p = strchr(current->comm, ' ')))
991		*p = '\0';
992
993	/* Best effort attempt to cope with MCA/INIT delivered while in
994	 * physical mode.
995	 */
996	regs->cr_ipsr = ms->pmsa_ipsr;
997	if (ia64_psr(regs)->dt == 0) {
998		va.l = r12;
999		if (va.f.reg == 0) {
1000			va.f.reg = 7;
1001			r12 = va.l;
1002		}
1003		va.l = r13;
1004		if (va.f.reg == 0) {
1005			va.f.reg = 7;
1006			r13 = va.l;
1007		}
1008	}
1009	if (ia64_psr(regs)->rt == 0) {
1010		va.l = ar_bspstore;
1011		if (va.f.reg == 0) {
1012			va.f.reg = 7;
1013			ar_bspstore = va.l;
1014		}
1015		va.l = ar_bsp;
1016		if (va.f.reg == 0) {
1017			va.f.reg = 7;
1018			ar_bsp = va.l;
1019		}
1020	}
1021
1022	/* mca_asm.S ia64_old_stack() cannot assume that the dirty registers
1023	 * have been copied to the old stack, the old stack may fail the
1024	 * validation tests below.  So ia64_old_stack() must restore the dirty
1025	 * registers from the new stack.  The old and new bspstore probably
1026	 * have different alignments, so loadrs calculated on the old bsp
1027	 * cannot be used to restore from the new bsp.  Calculate a suitable
1028	 * loadrs for the new stack and save it in the new pt_regs, where
1029	 * ia64_old_stack() can get it.
1030	 */
1031	old_bspstore = (unsigned long *)ar_bspstore;
1032	old_bsp = (unsigned long *)ar_bsp;
1033	slots = ia64_rse_num_regs(old_bspstore, old_bsp);
1034	new_bspstore = (unsigned long *)((u64)current + IA64_RBS_OFFSET);
1035	new_bsp = ia64_rse_skip_regs(new_bspstore, slots);
1036	regs->loadrs = (new_bsp - new_bspstore) * 8 << 16;
1037
1038	/* Verify the previous stack state before we change it */
1039	if (user_mode(regs)) {
1040		msg = "occurred in user space";
1041		/* previous_current is guaranteed to be valid when the task was
1042		 * in user space, so ...
1043		 */
1044		ia64_mca_modify_comm(previous_current);
1045		goto no_mod;
1046	}
1047
1048	if (r13 != sos->prev_IA64_KR_CURRENT) {
1049		msg = "inconsistent previous current and r13";
1050		goto no_mod;
1051	}
1052
1053	if (!mca_recover_range(ms->pmsa_iip)) {
1054		if ((r12 - r13) >= KERNEL_STACK_SIZE) {
1055			msg = "inconsistent r12 and r13";
1056			goto no_mod;
1057		}
1058		if ((ar_bspstore - r13) >= KERNEL_STACK_SIZE) {
1059			msg = "inconsistent ar.bspstore and r13";
1060			goto no_mod;
1061		}
1062		va.p = old_bspstore;
1063		if (va.f.reg < 5) {
1064			msg = "old_bspstore is in the wrong region";
1065			goto no_mod;
1066		}
1067		if ((ar_bsp - r13) >= KERNEL_STACK_SIZE) {
1068			msg = "inconsistent ar.bsp and r13";
1069			goto no_mod;
1070		}
1071		size += (ia64_rse_skip_regs(old_bspstore, slots) - old_bspstore) * 8;
1072		if (ar_bspstore + size > r12) {
1073			msg = "no room for blocked state";
1074			goto no_mod;
1075		}
1076	}
1077
1078	ia64_mca_modify_comm(previous_current);
1079
1080	/* Make the original task look blocked.  First stack a struct pt_regs,
1081	 * describing the state at the time of interrupt.  mca_asm.S built a
1082	 * partial pt_regs, copy it and fill in the blanks using minstate.
1083	 */
1084	p = (char *)r12 - sizeof(*regs);
1085	old_regs = (struct pt_regs *)p;
1086	memcpy(old_regs, regs, sizeof(*regs));
1087	old_regs->loadrs = loadrs;
1088	old_unat = old_regs->ar_unat;
1089	finish_pt_regs(old_regs, sos, &old_unat);
1090
1091	/* Next stack a struct switch_stack.  mca_asm.S built a partial
1092	 * switch_stack, copy it and fill in the blanks using pt_regs and
1093	 * minstate.
1094	 *
1095	 * In the synthesized switch_stack, b0 points to ia64_leave_kernel,
1096	 * ar.pfs is set to 0.
1097	 *
1098	 * unwind.c::unw_unwind() does special processing for interrupt frames.
1099	 * It checks if the PRED_NON_SYSCALL predicate is set, if the predicate
1100	 * is clear then unw_unwind() does _not_ adjust bsp over pt_regs.  Not
1101	 * that this is documented, of course.  Set PRED_NON_SYSCALL in the
1102	 * switch_stack on the original stack so it will unwind correctly when
1103	 * unwind.c reads pt_regs.
1104	 *
1105	 * thread.ksp is updated to point to the synthesized switch_stack.
1106	 */
1107	p -= sizeof(struct switch_stack);
1108	old_sw = (struct switch_stack *)p;
1109	memcpy(old_sw, sw, sizeof(*sw));
1110	old_sw->caller_unat = old_unat;
1111	old_sw->ar_fpsr = old_regs->ar_fpsr;
1112	copy_reg(&ms->pmsa_gr[4-1], ms->pmsa_nat_bits, &old_sw->r4, &old_unat);
1113	copy_reg(&ms->pmsa_gr[5-1], ms->pmsa_nat_bits, &old_sw->r5, &old_unat);
1114	copy_reg(&ms->pmsa_gr[6-1], ms->pmsa_nat_bits, &old_sw->r6, &old_unat);
1115	copy_reg(&ms->pmsa_gr[7-1], ms->pmsa_nat_bits, &old_sw->r7, &old_unat);
1116	old_sw->b0 = (u64)ia64_leave_kernel;
1117	old_sw->b1 = ms->pmsa_br1;
1118	old_sw->ar_pfs = 0;
1119	old_sw->ar_unat = old_unat;
1120	old_sw->pr = old_regs->pr | (1UL << PRED_NON_SYSCALL);
1121	previous_current->thread.ksp = (u64)p - 16;
1122
1123	/* Finally copy the original stack's registers back to its RBS.
1124	 * Registers from ar.bspstore through ar.bsp at the time of the event
1125	 * are in the current RBS, copy them back to the original stack.  The
1126	 * copy must be done register by register because the original bspstore
1127	 * and the current one have different alignments, so the saved RNAT
1128	 * data occurs at different places.
1129	 *
1130	 * mca_asm does cover, so the old_bsp already includes all registers at
1131	 * the time of MCA/INIT.  It also does flushrs, so all registers before
1132	 * this function have been written to backing store on the MCA/INIT
1133	 * stack.
1134	 */
1135	new_rnat = ia64_get_rnat(ia64_rse_rnat_addr(new_bspstore));
1136	old_rnat = regs->ar_rnat;
1137	while (slots--) {
1138		if (ia64_rse_is_rnat_slot(new_bspstore)) {
1139			new_rnat = ia64_get_rnat(new_bspstore++);
1140		}
1141		if (ia64_rse_is_rnat_slot(old_bspstore)) {
1142			*old_bspstore++ = old_rnat;
1143			old_rnat = 0;
1144		}
1145		nat = (new_rnat >> ia64_rse_slot_num(new_bspstore)) & 1UL;
1146		old_rnat &= ~(1UL << ia64_rse_slot_num(old_bspstore));
1147		old_rnat |= (nat << ia64_rse_slot_num(old_bspstore));
1148		*old_bspstore++ = *new_bspstore++;
1149	}
1150	old_sw->ar_bspstore = (unsigned long)old_bspstore;
1151	old_sw->ar_rnat = old_rnat;
1152
1153	sos->prev_task = previous_current;
1154	return previous_current;
1155
1156no_mod:
1157	mprintk(KERN_INFO "cpu %d, %s %s, original stack not modified\n",
1158			smp_processor_id(), type, msg);
1159	old_unat = regs->ar_unat;
1160	finish_pt_regs(regs, sos, &old_unat);
1161	return previous_current;
1162}
1163
1164/* The monarch/slave interaction is based on monarch_cpu and requires that all
1165 * slaves have entered rendezvous before the monarch leaves.  If any cpu has
1166 * not entered rendezvous yet then wait a bit.  The assumption is that any
1167 * slave that has not rendezvoused after a reasonable time is never going to do
1168 * so.  In this context, slave includes cpus that respond to the MCA rendezvous
1169 * interrupt, as well as cpus that receive the INIT slave event.
1170 */
1171
1172static void
1173ia64_wait_for_slaves(int monarch, const char *type)
1174{
1175	int c, i , wait;
1176
1177	/*
1178	 * wait 5 seconds total for slaves (arbitrary)
1179	 */
1180	for (i = 0; i < 5000; i++) {
1181		wait = 0;
1182		for_each_online_cpu(c) {
1183			if (c == monarch)
1184				continue;
1185			if (ia64_mc_info.imi_rendez_checkin[c]
1186					== IA64_MCA_RENDEZ_CHECKIN_NOTDONE) {
1187				udelay(1000);		/* short wait */
1188				wait = 1;
1189				break;
1190			}
1191		}
1192		if (!wait)
1193			goto all_in;
1194	}
1195
1196	/*
1197	 * Maybe slave(s) dead. Print buffered messages immediately.
1198	 */
1199	ia64_mlogbuf_finish(0);
1200	mprintk(KERN_INFO "OS %s slave did not rendezvous on cpu", type);
1201	for_each_online_cpu(c) {
1202		if (c == monarch)
1203			continue;
1204		if (ia64_mc_info.imi_rendez_checkin[c] == IA64_MCA_RENDEZ_CHECKIN_NOTDONE)
1205			mprintk(" %d", c);
1206	}
1207	mprintk("\n");
1208	return;
1209
1210all_in:
1211	mprintk(KERN_INFO "All OS %s slaves have reached rendezvous\n", type);
1212	return;
1213}
1214
1215/*  mca_insert_tr
1216 *
1217 *  Switch rid when TR reload and needed!
1218 *  iord: 1: itr, 2: itr;
1219 *
1220*/
1221static void mca_insert_tr(u64 iord)
1222{
1223
1224	int i;
1225	u64 old_rr;
1226	struct ia64_tr_entry *p;
1227	unsigned long psr;
1228	int cpu = smp_processor_id();
1229
1230	if (!ia64_idtrs[cpu])
1231		return;
1232
1233	psr = ia64_clear_ic();
1234	for (i = IA64_TR_ALLOC_BASE; i < IA64_TR_ALLOC_MAX; i++) {
1235		p = ia64_idtrs[cpu] + (iord - 1) * IA64_TR_ALLOC_MAX;
1236		if (p->pte & 0x1) {
1237			old_rr = ia64_get_rr(p->ifa);
1238			if (old_rr != p->rr) {
1239				ia64_set_rr(p->ifa, p->rr);
1240				ia64_srlz_d();
1241			}
1242			ia64_ptr(iord, p->ifa, p->itir >> 2);
1243			ia64_srlz_i();
1244			if (iord & 0x1) {
1245				ia64_itr(0x1, i, p->ifa, p->pte, p->itir >> 2);
1246				ia64_srlz_i();
1247			}
1248			if (iord & 0x2) {
1249				ia64_itr(0x2, i, p->ifa, p->pte, p->itir >> 2);
1250				ia64_srlz_i();
1251			}
1252			if (old_rr != p->rr) {
1253				ia64_set_rr(p->ifa, old_rr);
1254				ia64_srlz_d();
1255			}
1256		}
1257	}
1258	ia64_set_psr(psr);
1259}
1260
1261/*
1262 * ia64_mca_handler
1263 *
1264 *	This is uncorrectable machine check handler called from OS_MCA
1265 *	dispatch code which is in turn called from SAL_CHECK().
1266 *	This is the place where the core of OS MCA handling is done.
1267 *	Right now the logs are extracted and displayed in a well-defined
1268 *	format. This handler code is supposed to be run only on the
1269 *	monarch processor. Once the monarch is done with MCA handling
1270 *	further MCA logging is enabled by clearing logs.
1271 *	Monarch also has the duty of sending wakeup-IPIs to pull the
1272 *	slave processors out of rendezvous spinloop.
1273 *
1274 *	If multiple processors call into OS_MCA, the first will become
1275 *	the monarch.  Subsequent cpus will be recorded in the mca_cpu
1276 *	bitmask.  After the first monarch has processed its MCA, it
1277 *	will wake up the next cpu in the mca_cpu bitmask and then go
1278 *	into the rendezvous loop.  When all processors have serviced
1279 *	their MCA, the last monarch frees up the rest of the processors.
1280 */
1281void
1282ia64_mca_handler(struct pt_regs *regs, struct switch_stack *sw,
1283		 struct ia64_sal_os_state *sos)
1284{
1285	int recover, cpu = smp_processor_id();
1286	struct task_struct *previous_current;
1287	struct ia64_mca_notify_die nd =
1288		{ .sos = sos, .monarch_cpu = &monarch_cpu, .data = &recover };
1289	static atomic_t mca_count;
1290	static cpumask_t mca_cpu;
1291
1292	if (atomic_add_return(1, &mca_count) == 1) {
1293		monarch_cpu = cpu;
1294		sos->monarch = 1;
1295	} else {
1296		cpumask_set_cpu(cpu, &mca_cpu);
1297		sos->monarch = 0;
1298	}
1299	mprintk(KERN_INFO "Entered OS MCA handler. PSP=%lx cpu=%d "
1300		"monarch=%ld\n", sos->proc_state_param, cpu, sos->monarch);
1301
1302	previous_current = ia64_mca_modify_original_stack(regs, sw, sos, "MCA");
1303
1304	NOTIFY_MCA(DIE_MCA_MONARCH_ENTER, regs, (long)&nd, 1);
1305
1306	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_CONCURRENT_MCA;
1307	if (sos->monarch) {
1308		ia64_wait_for_slaves(cpu, "MCA");
1309
1310		/* Wakeup all the processors which are spinning in the
1311		 * rendezvous loop.  They will leave SAL, then spin in the OS
1312		 * with interrupts disabled until this monarch cpu leaves the
1313		 * MCA handler.  That gets control back to the OS so we can
1314		 * backtrace the other cpus, backtrace when spinning in SAL
1315		 * does not work.
1316		 */
1317		ia64_mca_wakeup_all();
1318	} else {
1319		while (cpumask_test_cpu(cpu, &mca_cpu))
1320			cpu_relax();	/* spin until monarch wakes us */
1321	}
1322
1323	NOTIFY_MCA(DIE_MCA_MONARCH_PROCESS, regs, (long)&nd, 1);
1324
1325	/* Get the MCA error record and log it */
1326	ia64_mca_log_sal_error_record(SAL_INFO_TYPE_MCA);
1327
1328	/* MCA error recovery */
1329	recover = (ia64_mca_ucmc_extension
1330		&& ia64_mca_ucmc_extension(
1331			IA64_LOG_CURR_BUFFER(SAL_INFO_TYPE_MCA),
1332			sos));
1333
1334	if (recover) {
1335		sal_log_record_header_t *rh = IA64_LOG_CURR_BUFFER(SAL_INFO_TYPE_MCA);
1336		rh->severity = sal_log_severity_corrected;
1337		ia64_sal_clear_state_info(SAL_INFO_TYPE_MCA);
1338		sos->os_status = IA64_MCA_CORRECTED;
1339	} else {
1340		/* Dump buffered message to console */
1341		ia64_mlogbuf_finish(1);
1342	}
1343
1344	if (__this_cpu_read(ia64_mca_tr_reload)) {
1345		mca_insert_tr(0x1); /*Reload dynamic itrs*/
1346		mca_insert_tr(0x2); /*Reload dynamic itrs*/
1347	}
1348
1349	NOTIFY_MCA(DIE_MCA_MONARCH_LEAVE, regs, (long)&nd, 1);
1350
1351	if (atomic_dec_return(&mca_count) > 0) {
1352		int i;
1353
1354		/* wake up the next monarch cpu,
1355		 * and put this cpu in the rendez loop.
1356		 */
1357		for_each_online_cpu(i) {
1358			if (cpumask_test_cpu(i, &mca_cpu)) {
1359				monarch_cpu = i;
1360				cpumask_clear_cpu(i, &mca_cpu);	/* wake next cpu */
1361				while (monarch_cpu != -1)
1362					cpu_relax();	/* spin until last cpu leaves */
1363				set_curr_task(cpu, previous_current);
1364				ia64_mc_info.imi_rendez_checkin[cpu]
1365						= IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1366				return;
1367			}
1368		}
1369	}
1370	set_curr_task(cpu, previous_current);
1371	ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1372	monarch_cpu = -1;	/* This frees the slaves and previous monarchs */
1373}
1374
1375static DECLARE_WORK(cmc_disable_work, ia64_mca_cmc_vector_disable_keventd);
1376static DECLARE_WORK(cmc_enable_work, ia64_mca_cmc_vector_enable_keventd);
1377
1378/*
1379 * ia64_mca_cmc_int_handler
1380 *
1381 *  This is corrected machine check interrupt handler.
1382 *	Right now the logs are extracted and displayed in a well-defined
1383 *	format.
1384 *
1385 * Inputs
1386 *      interrupt number
1387 *      client data arg ptr
1388 *
1389 * Outputs
1390 *	None
1391 */
1392static irqreturn_t
1393ia64_mca_cmc_int_handler(int cmc_irq, void *arg)
1394{
1395	static unsigned long	cmc_history[CMC_HISTORY_LENGTH];
1396	static int		index;
1397	static DEFINE_SPINLOCK(cmc_history_lock);
1398
1399	IA64_MCA_DEBUG("%s: received interrupt vector = %#x on CPU %d\n",
1400		       __func__, cmc_irq, smp_processor_id());
1401
1402	/* SAL spec states this should run w/ interrupts enabled */
1403	local_irq_enable();
1404
1405	spin_lock(&cmc_history_lock);
1406	if (!cmc_polling_enabled) {
1407		int i, count = 1; /* we know 1 happened now */
1408		unsigned long now = jiffies;
1409
1410		for (i = 0; i < CMC_HISTORY_LENGTH; i++) {
1411			if (now - cmc_history[i] <= HZ)
1412				count++;
1413		}
1414
1415		IA64_MCA_DEBUG(KERN_INFO "CMC threshold %d/%d\n", count, CMC_HISTORY_LENGTH);
1416		if (count >= CMC_HISTORY_LENGTH) {
1417
1418			cmc_polling_enabled = 1;
1419			spin_unlock(&cmc_history_lock);
1420			/* If we're being hit with CMC interrupts, we won't
1421			 * ever execute the schedule_work() below.  Need to
1422			 * disable CMC interrupts on this processor now.
1423			 */
1424			ia64_mca_cmc_vector_disable(NULL);
1425			schedule_work(&cmc_disable_work);
1426
1427			/*
1428			 * Corrected errors will still be corrected, but
1429			 * make sure there's a log somewhere that indicates
1430			 * something is generating more than we can handle.
1431			 */
1432			printk(KERN_WARNING "WARNING: Switching to polling CMC handler; error records may be lost\n");
1433
1434			mod_timer(&cmc_poll_timer, jiffies + CMC_POLL_INTERVAL);
1435
1436			/* lock already released, get out now */
1437			goto out;
1438		} else {
1439			cmc_history[index++] = now;
1440			if (index == CMC_HISTORY_LENGTH)
1441				index = 0;
1442		}
1443	}
1444	spin_unlock(&cmc_history_lock);
1445out:
1446	/* Get the CMC error record and log it */
1447	ia64_mca_log_sal_error_record(SAL_INFO_TYPE_CMC);
1448
1449	local_irq_disable();
1450
1451	return IRQ_HANDLED;
1452}
1453
1454/*
1455 *  ia64_mca_cmc_int_caller
1456 *
1457 * 	Triggered by sw interrupt from CMC polling routine.  Calls
1458 * 	real interrupt handler and either triggers a sw interrupt
1459 * 	on the next cpu or does cleanup at the end.
1460 *
1461 * Inputs
1462 *	interrupt number
1463 *	client data arg ptr
1464 * Outputs
1465 * 	handled
1466 */
1467static irqreturn_t
1468ia64_mca_cmc_int_caller(int cmc_irq, void *arg)
1469{
1470	static int start_count = -1;
1471	unsigned int cpuid;
1472
1473	cpuid = smp_processor_id();
1474
1475	/* If first cpu, update count */
1476	if (start_count == -1)
1477		start_count = IA64_LOG_COUNT(SAL_INFO_TYPE_CMC);
1478
1479	ia64_mca_cmc_int_handler(cmc_irq, arg);
1480
1481	cpuid = cpumask_next(cpuid+1, cpu_online_mask);
1482
1483	if (cpuid < nr_cpu_ids) {
1484		platform_send_ipi(cpuid, IA64_CMCP_VECTOR, IA64_IPI_DM_INT, 0);
1485	} else {
1486		/* If no log record, switch out of polling mode */
1487		if (start_count == IA64_LOG_COUNT(SAL_INFO_TYPE_CMC)) {
1488
1489			printk(KERN_WARNING "Returning to interrupt driven CMC handler\n");
1490			schedule_work(&cmc_enable_work);
1491			cmc_polling_enabled = 0;
1492
1493		} else {
1494
1495			mod_timer(&cmc_poll_timer, jiffies + CMC_POLL_INTERVAL);
1496		}
1497
1498		start_count = -1;
1499	}
1500
1501	return IRQ_HANDLED;
1502}
1503
1504/*
1505 *  ia64_mca_cmc_poll
1506 *
1507 *	Poll for Corrected Machine Checks (CMCs)
1508 *
1509 * Inputs   :   dummy(unused)
1510 * Outputs  :   None
1511 *
1512 */
1513static void
1514ia64_mca_cmc_poll (unsigned long dummy)
1515{
1516	/* Trigger a CMC interrupt cascade  */
1517	platform_send_ipi(cpumask_first(cpu_online_mask), IA64_CMCP_VECTOR,
1518							IA64_IPI_DM_INT, 0);
1519}
1520
1521/*
1522 *  ia64_mca_cpe_int_caller
1523 *
1524 * 	Triggered by sw interrupt from CPE polling routine.  Calls
1525 * 	real interrupt handler and either triggers a sw interrupt
1526 * 	on the next cpu or does cleanup at the end.
1527 *
1528 * Inputs
1529 *	interrupt number
1530 *	client data arg ptr
1531 * Outputs
1532 * 	handled
1533 */
1534#ifdef CONFIG_ACPI
1535
1536static irqreturn_t
1537ia64_mca_cpe_int_caller(int cpe_irq, void *arg)
1538{
1539	static int start_count = -1;
1540	static int poll_time = MIN_CPE_POLL_INTERVAL;
1541	unsigned int cpuid;
1542
1543	cpuid = smp_processor_id();
1544
1545	/* If first cpu, update count */
1546	if (start_count == -1)
1547		start_count = IA64_LOG_COUNT(SAL_INFO_TYPE_CPE);
1548
1549	ia64_mca_cpe_int_handler(cpe_irq, arg);
1550
1551	cpuid = cpumask_next(cpuid+1, cpu_online_mask);
1552
1553	if (cpuid < NR_CPUS) {
1554		platform_send_ipi(cpuid, IA64_CPEP_VECTOR, IA64_IPI_DM_INT, 0);
1555	} else {
1556		/*
1557		 * If a log was recorded, increase our polling frequency,
1558		 * otherwise, backoff or return to interrupt mode.
1559		 */
1560		if (start_count != IA64_LOG_COUNT(SAL_INFO_TYPE_CPE)) {
1561			poll_time = max(MIN_CPE_POLL_INTERVAL, poll_time / 2);
1562		} else if (cpe_vector < 0) {
1563			poll_time = min(MAX_CPE_POLL_INTERVAL, poll_time * 2);
1564		} else {
1565			poll_time = MIN_CPE_POLL_INTERVAL;
1566
1567			printk(KERN_WARNING "Returning to interrupt driven CPE handler\n");
1568			enable_irq(local_vector_to_irq(IA64_CPE_VECTOR));
1569			cpe_poll_enabled = 0;
1570		}
1571
1572		if (cpe_poll_enabled)
1573			mod_timer(&cpe_poll_timer, jiffies + poll_time);
1574		start_count = -1;
1575	}
1576
1577	return IRQ_HANDLED;
1578}
1579
1580/*
1581 *  ia64_mca_cpe_poll
1582 *
1583 *	Poll for Corrected Platform Errors (CPEs), trigger interrupt
1584 *	on first cpu, from there it will trickle through all the cpus.
1585 *
1586 * Inputs   :   dummy(unused)
1587 * Outputs  :   None
1588 *
1589 */
1590static void
1591ia64_mca_cpe_poll (unsigned long dummy)
1592{
1593	/* Trigger a CPE interrupt cascade  */
1594	platform_send_ipi(cpumask_first(cpu_online_mask), IA64_CPEP_VECTOR,
1595							IA64_IPI_DM_INT, 0);
1596}
1597
1598#endif /* CONFIG_ACPI */
1599
1600static int
1601default_monarch_init_process(struct notifier_block *self, unsigned long val, void *data)
1602{
1603	int c;
1604	struct task_struct *g, *t;
1605	if (val != DIE_INIT_MONARCH_PROCESS)
1606		return NOTIFY_DONE;
1607#ifdef CONFIG_KEXEC
1608	if (atomic_read(&kdump_in_progress))
1609		return NOTIFY_DONE;
1610#endif
1611
1612	/*
1613	 * FIXME: mlogbuf will brim over with INIT stack dumps.
1614	 * To enable show_stack from INIT, we use oops_in_progress which should
1615	 * be used in real oops. This would cause something wrong after INIT.
1616	 */
1617	BREAK_LOGLEVEL(console_loglevel);
1618	ia64_mlogbuf_dump_from_init();
1619
1620	printk(KERN_ERR "Processes interrupted by INIT -");
1621	for_each_online_cpu(c) {
1622		struct ia64_sal_os_state *s;
1623		t = __va(__per_cpu_mca[c] + IA64_MCA_CPU_INIT_STACK_OFFSET);
1624		s = (struct ia64_sal_os_state *)((char *)t + MCA_SOS_OFFSET);
1625		g = s->prev_task;
1626		if (g) {
1627			if (g->pid)
1628				printk(" %d", g->pid);
1629			else
1630				printk(" %d (cpu %d task 0x%p)", g->pid, task_cpu(g), g);
1631		}
1632	}
1633	printk("\n\n");
1634	if (read_trylock(&tasklist_lock)) {
1635		do_each_thread (g, t) {
1636			printk("\nBacktrace of pid %d (%s)\n", t->pid, t->comm);
1637			show_stack(t, NULL);
1638		} while_each_thread (g, t);
1639		read_unlock(&tasklist_lock);
1640	}
1641	/* FIXME: This will not restore zapped printk locks. */
1642	RESTORE_LOGLEVEL(console_loglevel);
1643	return NOTIFY_DONE;
1644}
1645
1646/*
1647 * C portion of the OS INIT handler
1648 *
1649 * Called from ia64_os_init_dispatch
1650 *
1651 * Inputs: pointer to pt_regs where processor info was saved.  SAL/OS state for
1652 * this event.  This code is used for both monarch and slave INIT events, see
1653 * sos->monarch.
1654 *
1655 * All INIT events switch to the INIT stack and change the previous process to
1656 * blocked status.  If one of the INIT events is the monarch then we are
1657 * probably processing the nmi button/command.  Use the monarch cpu to dump all
1658 * the processes.  The slave INIT events all spin until the monarch cpu
1659 * returns.  We can also get INIT slave events for MCA, in which case the MCA
1660 * process is the monarch.
1661 */
1662
1663void
1664ia64_init_handler(struct pt_regs *regs, struct switch_stack *sw,
1665		  struct ia64_sal_os_state *sos)
1666{
1667	static atomic_t slaves;
1668	static atomic_t monarchs;
1669	struct task_struct *previous_current;
1670	int cpu = smp_processor_id();
1671	struct ia64_mca_notify_die nd =
1672		{ .sos = sos, .monarch_cpu = &monarch_cpu };
1673
1674	NOTIFY_INIT(DIE_INIT_ENTER, regs, (long)&nd, 0);
1675
1676	mprintk(KERN_INFO "Entered OS INIT handler. PSP=%lx cpu=%d monarch=%ld\n",
1677		sos->proc_state_param, cpu, sos->monarch);
1678	salinfo_log_wakeup(SAL_INFO_TYPE_INIT, NULL, 0, 0);
1679
1680	previous_current = ia64_mca_modify_original_stack(regs, sw, sos, "INIT");
1681	sos->os_status = IA64_INIT_RESUME;
1682
1683	/* FIXME: Workaround for broken proms that drive all INIT events as
1684	 * slaves.  The last slave that enters is promoted to be a monarch.
1685	 * Remove this code in September 2006, that gives platforms a year to
1686	 * fix their proms and get their customers updated.
1687	 */
1688	if (!sos->monarch && atomic_add_return(1, &slaves) == num_online_cpus()) {
1689		mprintk(KERN_WARNING "%s: Promoting cpu %d to monarch.\n",
1690		        __func__, cpu);
1691		atomic_dec(&slaves);
1692		sos->monarch = 1;
1693	}
1694
1695	/* FIXME: Workaround for broken proms that drive all INIT events as
1696	 * monarchs.  Second and subsequent monarchs are demoted to slaves.
1697	 * Remove this code in September 2006, that gives platforms a year to
1698	 * fix their proms and get their customers updated.
1699	 */
1700	if (sos->monarch && atomic_add_return(1, &monarchs) > 1) {
1701		mprintk(KERN_WARNING "%s: Demoting cpu %d to slave.\n",
1702			       __func__, cpu);
1703		atomic_dec(&monarchs);
1704		sos->monarch = 0;
1705	}
1706
1707	if (!sos->monarch) {
1708		ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_INIT;
1709
1710#ifdef CONFIG_KEXEC
1711		while (monarch_cpu == -1 && !atomic_read(&kdump_in_progress))
1712			udelay(1000);
1713#else
1714		while (monarch_cpu == -1)
1715			cpu_relax();	/* spin until monarch enters */
1716#endif
1717
1718		NOTIFY_INIT(DIE_INIT_SLAVE_ENTER, regs, (long)&nd, 1);
1719		NOTIFY_INIT(DIE_INIT_SLAVE_PROCESS, regs, (long)&nd, 1);
1720
1721#ifdef CONFIG_KEXEC
1722		while (monarch_cpu != -1 && !atomic_read(&kdump_in_progress))
1723			udelay(1000);
1724#else
1725		while (monarch_cpu != -1)
1726			cpu_relax();	/* spin until monarch leaves */
1727#endif
1728
1729		NOTIFY_INIT(DIE_INIT_SLAVE_LEAVE, regs, (long)&nd, 1);
1730
1731		mprintk("Slave on cpu %d returning to normal service.\n", cpu);
1732		set_curr_task(cpu, previous_current);
1733		ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1734		atomic_dec(&slaves);
1735		return;
1736	}
1737
1738	monarch_cpu = cpu;
1739	NOTIFY_INIT(DIE_INIT_MONARCH_ENTER, regs, (long)&nd, 1);
1740
1741	/*
1742	 * Wait for a bit.  On some machines (e.g., HP's zx2000 and zx6000, INIT can be
1743	 * generated via the BMC's command-line interface, but since the console is on the
1744	 * same serial line, the user will need some time to switch out of the BMC before
1745	 * the dump begins.
1746	 */
1747	mprintk("Delaying for 5 seconds...\n");
1748	udelay(5*1000000);
1749	ia64_wait_for_slaves(cpu, "INIT");
1750	/* If nobody intercepts DIE_INIT_MONARCH_PROCESS then we drop through
1751	 * to default_monarch_init_process() above and just print all the
1752	 * tasks.
1753	 */
1754	NOTIFY_INIT(DIE_INIT_MONARCH_PROCESS, regs, (long)&nd, 1);
1755	NOTIFY_INIT(DIE_INIT_MONARCH_LEAVE, regs, (long)&nd, 1);
1756
1757	mprintk("\nINIT dump complete.  Monarch on cpu %d returning to normal service.\n", cpu);
1758	atomic_dec(&monarchs);
1759	set_curr_task(cpu, previous_current);
1760	monarch_cpu = -1;
1761	return;
1762}
1763
1764static int __init
1765ia64_mca_disable_cpe_polling(char *str)
1766{
1767	cpe_poll_enabled = 0;
1768	return 1;
1769}
1770
1771__setup("disable_cpe_poll", ia64_mca_disable_cpe_polling);
1772
1773static struct irqaction cmci_irqaction = {
1774	.handler =	ia64_mca_cmc_int_handler,
1775	.name =		"cmc_hndlr"
1776};
1777
1778static struct irqaction cmcp_irqaction = {
1779	.handler =	ia64_mca_cmc_int_caller,
1780	.name =		"cmc_poll"
1781};
1782
1783static struct irqaction mca_rdzv_irqaction = {
1784	.handler =	ia64_mca_rendez_int_handler,
1785	.name =		"mca_rdzv"
1786};
1787
1788static struct irqaction mca_wkup_irqaction = {
1789	.handler =	ia64_mca_wakeup_int_handler,
1790	.name =		"mca_wkup"
1791};
1792
1793#ifdef CONFIG_ACPI
1794static struct irqaction mca_cpe_irqaction = {
1795	.handler =	ia64_mca_cpe_int_handler,
1796	.name =		"cpe_hndlr"
1797};
1798
1799static struct irqaction mca_cpep_irqaction = {
1800	.handler =	ia64_mca_cpe_int_caller,
1801	.name =		"cpe_poll"
1802};
1803#endif /* CONFIG_ACPI */
1804
1805/* Minimal format of the MCA/INIT stacks.  The pseudo processes that run on
1806 * these stacks can never sleep, they cannot return from the kernel to user
1807 * space, they do not appear in a normal ps listing.  So there is no need to
1808 * format most of the fields.
1809 */
1810
1811static void
1812format_mca_init_stack(void *mca_data, unsigned long offset,
1813		const char *type, int cpu)
1814{
1815	struct task_struct *p = (struct task_struct *)((char *)mca_data + offset);
1816	struct thread_info *ti;
1817	memset(p, 0, KERNEL_STACK_SIZE);
1818	ti = task_thread_info(p);
1819	ti->flags = _TIF_MCA_INIT;
1820	ti->preempt_count = 1;
1821	ti->task = p;
1822	ti->cpu = cpu;
1823	p->stack = ti;
1824	p->state = TASK_UNINTERRUPTIBLE;
1825	cpumask_set_cpu(cpu, &p->cpus_allowed);
1826	INIT_LIST_HEAD(&p->tasks);
1827	p->parent = p->real_parent = p->group_leader = p;
1828	INIT_LIST_HEAD(&p->children);
1829	INIT_LIST_HEAD(&p->sibling);
1830	strncpy(p->comm, type, sizeof(p->comm)-1);
1831}
1832
1833/* Caller prevents this from being called after init */
1834static void * __init_refok mca_bootmem(void)
1835{
1836	return __alloc_bootmem(sizeof(struct ia64_mca_cpu),
1837	                    KERNEL_STACK_SIZE, 0);
1838}
1839
1840/* Do per-CPU MCA-related initialization.  */
1841void
1842ia64_mca_cpu_init(void *cpu_data)
1843{
1844	void *pal_vaddr;
1845	void *data;
1846	long sz = sizeof(struct ia64_mca_cpu);
1847	int cpu = smp_processor_id();
1848	static int first_time = 1;
1849
1850	/*
1851	 * Structure will already be allocated if cpu has been online,
1852	 * then offlined.
1853	 */
1854	if (__per_cpu_mca[cpu]) {
1855		data = __va(__per_cpu_mca[cpu]);
1856	} else {
1857		if (first_time) {
1858			data = mca_bootmem();
1859			first_time = 0;
1860		} else
1861			data = (void *)__get_free_pages(GFP_KERNEL,
1862							get_order(sz));
1863		if (!data)
1864			panic("Could not allocate MCA memory for cpu %d\n",
1865					cpu);
1866	}
1867	format_mca_init_stack(data, offsetof(struct ia64_mca_cpu, mca_stack),
1868		"MCA", cpu);
1869	format_mca_init_stack(data, offsetof(struct ia64_mca_cpu, init_stack),
1870		"INIT", cpu);
1871	__this_cpu_write(ia64_mca_data, (__per_cpu_mca[cpu] = __pa(data)));
1872
1873	/*
1874	 * Stash away a copy of the PTE needed to map the per-CPU page.
1875	 * We may need it during MCA recovery.
1876	 */
1877	__this_cpu_write(ia64_mca_per_cpu_pte,
1878		pte_val(mk_pte_phys(__pa(cpu_data), PAGE_KERNEL)));
1879
1880	/*
1881	 * Also, stash away a copy of the PAL address and the PTE
1882	 * needed to map it.
1883	 */
1884	pal_vaddr = efi_get_pal_addr();
1885	if (!pal_vaddr)
1886		return;
1887	__this_cpu_write(ia64_mca_pal_base,
1888		GRANULEROUNDDOWN((unsigned long) pal_vaddr));
1889	__this_cpu_write(ia64_mca_pal_pte, pte_val(mk_pte_phys(__pa(pal_vaddr),
1890							      PAGE_KERNEL)));
1891}
1892
1893static void ia64_mca_cmc_vector_adjust(void *dummy)
1894{
1895	unsigned long flags;
1896
1897	local_irq_save(flags);
1898	if (!cmc_polling_enabled)
1899		ia64_mca_cmc_vector_enable(NULL);
1900	local_irq_restore(flags);
1901}
1902
1903static int mca_cpu_callback(struct notifier_block *nfb,
1904				      unsigned long action,
1905				      void *hcpu)
1906{
1907	int hotcpu = (unsigned long) hcpu;
1908
1909	switch (action) {
1910	case CPU_ONLINE:
1911	case CPU_ONLINE_FROZEN:
1912		smp_call_function_single(hotcpu, ia64_mca_cmc_vector_adjust,
1913					 NULL, 0);
1914		break;
1915	}
1916	return NOTIFY_OK;
1917}
1918
1919static struct notifier_block mca_cpu_notifier = {
1920	.notifier_call = mca_cpu_callback
1921};
1922
1923/*
1924 * ia64_mca_init
1925 *
1926 *  Do all the system level mca specific initialization.
1927 *
1928 *	1. Register spinloop and wakeup request interrupt vectors
1929 *
1930 *	2. Register OS_MCA handler entry point
1931 *
1932 *	3. Register OS_INIT handler entry point
1933 *
1934 *  4. Initialize MCA/CMC/INIT related log buffers maintained by the OS.
1935 *
1936 *  Note that this initialization is done very early before some kernel
1937 *  services are available.
1938 *
1939 *  Inputs  :   None
1940 *
1941 *  Outputs :   None
1942 */
1943void __init
1944ia64_mca_init(void)
1945{
1946	ia64_fptr_t *init_hldlr_ptr_monarch = (ia64_fptr_t *)ia64_os_init_dispatch_monarch;
1947	ia64_fptr_t *init_hldlr_ptr_slave = (ia64_fptr_t *)ia64_os_init_dispatch_slave;
1948	ia64_fptr_t *mca_hldlr_ptr = (ia64_fptr_t *)ia64_os_mca_dispatch;
1949	int i;
1950	long rc;
1951	struct ia64_sal_retval isrv;
1952	unsigned long timeout = IA64_MCA_RENDEZ_TIMEOUT; /* platform specific */
1953	static struct notifier_block default_init_monarch_nb = {
1954		.notifier_call = default_monarch_init_process,
1955		.priority = 0/* we need to notified last */
1956	};
1957
1958	IA64_MCA_DEBUG("%s: begin\n", __func__);
1959
1960	/* Clear the Rendez checkin flag for all cpus */
1961	for(i = 0 ; i < NR_CPUS; i++)
1962		ia64_mc_info.imi_rendez_checkin[i] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
1963
1964	/*
1965	 * Register the rendezvous spinloop and wakeup mechanism with SAL
1966	 */
1967
1968	/* Register the rendezvous interrupt vector with SAL */
1969	while (1) {
1970		isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_RENDEZ_INT,
1971					      SAL_MC_PARAM_MECHANISM_INT,
1972					      IA64_MCA_RENDEZ_VECTOR,
1973					      timeout,
1974					      SAL_MC_PARAM_RZ_ALWAYS);
1975		rc = isrv.status;
1976		if (rc == 0)
1977			break;
1978		if (rc == -2) {
1979			printk(KERN_INFO "Increasing MCA rendezvous timeout from "
1980				"%ld to %ld milliseconds\n", timeout, isrv.v0);
1981			timeout = isrv.v0;
1982			NOTIFY_MCA(DIE_MCA_NEW_TIMEOUT, NULL, timeout, 0);
1983			continue;
1984		}
1985		printk(KERN_ERR "Failed to register rendezvous interrupt "
1986		       "with SAL (status %ld)\n", rc);
1987		return;
1988	}
1989
1990	/* Register the wakeup interrupt vector with SAL */
1991	isrv = ia64_sal_mc_set_params(SAL_MC_PARAM_RENDEZ_WAKEUP,
1992				      SAL_MC_PARAM_MECHANISM_INT,
1993				      IA64_MCA_WAKEUP_VECTOR,
1994				      0, 0);
1995	rc = isrv.status;
1996	if (rc) {
1997		printk(KERN_ERR "Failed to register wakeup interrupt with SAL "
1998		       "(status %ld)\n", rc);
1999		return;
2000	}
2001
2002	IA64_MCA_DEBUG("%s: registered MCA rendezvous spinloop and wakeup mech.\n", __func__);
2003
2004	ia64_mc_info.imi_mca_handler        = ia64_tpa(mca_hldlr_ptr->fp);
2005	/*
2006	 * XXX - disable SAL checksum by setting size to 0; should be
2007	 *	ia64_tpa(ia64_os_mca_dispatch_end) - ia64_tpa(ia64_os_mca_dispatch);
2008	 */
2009	ia64_mc_info.imi_mca_handler_size	= 0;
2010
2011	/* Register the os mca handler with SAL */
2012	if ((rc = ia64_sal_set_vectors(SAL_VECTOR_OS_MCA,
2013				       ia64_mc_info.imi_mca_handler,
2014				       ia64_tpa(mca_hldlr_ptr->gp),
2015				       ia64_mc_info.imi_mca_handler_size,
2016				       0, 0, 0)))
2017	{
2018		printk(KERN_ERR "Failed to register OS MCA handler with SAL "
2019		       "(status %ld)\n", rc);
2020		return;
2021	}
2022
2023	IA64_MCA_DEBUG("%s: registered OS MCA handler with SAL at 0x%lx, gp = 0x%lx\n", __func__,
2024		       ia64_mc_info.imi_mca_handler, ia64_tpa(mca_hldlr_ptr->gp));
2025
2026	/*
2027	 * XXX - disable SAL checksum by setting size to 0, should be
2028	 * size of the actual init handler in mca_asm.S.
2029	 */
2030	ia64_mc_info.imi_monarch_init_handler		= ia64_tpa(init_hldlr_ptr_monarch->fp);
2031	ia64_mc_info.imi_monarch_init_handler_size	= 0;
2032	ia64_mc_info.imi_slave_init_handler		= ia64_tpa(init_hldlr_ptr_slave->fp);
2033	ia64_mc_info.imi_slave_init_handler_size	= 0;
2034
2035	IA64_MCA_DEBUG("%s: OS INIT handler at %lx\n", __func__,
2036		       ia64_mc_info.imi_monarch_init_handler);
2037
2038	/* Register the os init handler with SAL */
2039	if ((rc = ia64_sal_set_vectors(SAL_VECTOR_OS_INIT,
2040				       ia64_mc_info.imi_monarch_init_handler,
2041				       ia64_tpa(ia64_getreg(_IA64_REG_GP)),
2042				       ia64_mc_info.imi_monarch_init_handler_size,
2043				       ia64_mc_info.imi_slave_init_handler,
2044				       ia64_tpa(ia64_getreg(_IA64_REG_GP)),
2045				       ia64_mc_info.imi_slave_init_handler_size)))
2046	{
2047		printk(KERN_ERR "Failed to register m/s INIT handlers with SAL "
2048		       "(status %ld)\n", rc);
2049		return;
2050	}
2051	if (register_die_notifier(&default_init_monarch_nb)) {
2052		printk(KERN_ERR "Failed to register default monarch INIT process\n");
2053		return;
2054	}
2055
2056	IA64_MCA_DEBUG("%s: registered OS INIT handler with SAL\n", __func__);
2057
2058	/* Initialize the areas set aside by the OS to buffer the
2059	 * platform/processor error states for MCA/INIT/CMC
2060	 * handling.
2061	 */
2062	ia64_log_init(SAL_INFO_TYPE_MCA);
2063	ia64_log_init(SAL_INFO_TYPE_INIT);
2064	ia64_log_init(SAL_INFO_TYPE_CMC);
2065	ia64_log_init(SAL_INFO_TYPE_CPE);
2066
2067	mca_init = 1;
2068	printk(KERN_INFO "MCA related initialization done\n");
2069}
2070
2071
2072/*
2073 * These pieces cannot be done in ia64_mca_init() because it is called before
2074 * early_irq_init() which would wipe out our percpu irq registrations. But we
2075 * cannot leave them until ia64_mca_late_init() because by then all the other
2076 * processors have been brought online and have set their own CMC vectors to
2077 * point at a non-existant action. Called from arch_early_irq_init().
2078 */
2079void __init ia64_mca_irq_init(void)
2080{
2081	/*
2082	 *  Configure the CMCI/P vector and handler. Interrupts for CMC are
2083	 *  per-processor, so AP CMC interrupts are setup in smp_callin() (smpboot.c).
2084	 */
2085	register_percpu_irq(IA64_CMC_VECTOR, &cmci_irqaction);
2086	register_percpu_irq(IA64_CMCP_VECTOR, &cmcp_irqaction);
2087	ia64_mca_cmc_vector_setup();       /* Setup vector on BSP */
2088
2089	/* Setup the MCA rendezvous interrupt vector */
2090	register_percpu_irq(IA64_MCA_RENDEZ_VECTOR, &mca_rdzv_irqaction);
2091
2092	/* Setup the MCA wakeup interrupt vector */
2093	register_percpu_irq(IA64_MCA_WAKEUP_VECTOR, &mca_wkup_irqaction);
2094
2095#ifdef CONFIG_ACPI
2096	/* Setup the CPEI/P handler */
2097	register_percpu_irq(IA64_CPEP_VECTOR, &mca_cpep_irqaction);
2098#endif
2099}
2100
2101/*
2102 * ia64_mca_late_init
2103 *
2104 *	Opportunity to setup things that require initialization later
2105 *	than ia64_mca_init.  Setup a timer to poll for CPEs if the
2106 *	platform doesn't support an interrupt driven mechanism.
2107 *
2108 *  Inputs  :   None
2109 *  Outputs :   Status
2110 */
2111static int __init
2112ia64_mca_late_init(void)
2113{
2114	if (!mca_init)
2115		return 0;
2116
2117	register_hotcpu_notifier(&mca_cpu_notifier);
2118
2119	/* Setup the CMCI/P vector and handler */
2120	init_timer(&cmc_poll_timer);
2121	cmc_poll_timer.function = ia64_mca_cmc_poll;
2122
2123	/* Unmask/enable the vector */
2124	cmc_polling_enabled = 0;
2125	schedule_work(&cmc_enable_work);
2126
2127	IA64_MCA_DEBUG("%s: CMCI/P setup and enabled.\n", __func__);
2128
2129#ifdef CONFIG_ACPI
2130	/* Setup the CPEI/P vector and handler */
2131	cpe_vector = acpi_request_vector(ACPI_INTERRUPT_CPEI);
2132	init_timer(&cpe_poll_timer);
2133	cpe_poll_timer.function = ia64_mca_cpe_poll;
2134
2135	{
2136		unsigned int irq;
2137
2138		if (cpe_vector >= 0) {
2139			/* If platform supports CPEI, enable the irq. */
2140			irq = local_vector_to_irq(cpe_vector);
2141			if (irq > 0) {
2142				cpe_poll_enabled = 0;
2143				irq_set_status_flags(irq, IRQ_PER_CPU);
2144				setup_irq(irq, &mca_cpe_irqaction);
2145				ia64_cpe_irq = irq;
2146				ia64_mca_register_cpev(cpe_vector);
2147				IA64_MCA_DEBUG("%s: CPEI/P setup and enabled.\n",
2148					__func__);
2149				return 0;
2150			}
2151			printk(KERN_ERR "%s: Failed to find irq for CPE "
2152					"interrupt handler, vector %d\n",
2153					__func__, cpe_vector);
2154		}
2155		/* If platform doesn't support CPEI, get the timer going. */
2156		if (cpe_poll_enabled) {
2157			ia64_mca_cpe_poll(0UL);
2158			IA64_MCA_DEBUG("%s: CPEP setup and enabled.\n", __func__);
2159		}
2160	}
2161#endif
2162
2163	return 0;
2164}
2165
2166device_initcall(ia64_mca_late_init);
2167