xref: /openbmc/linux/arch/x86/platform/uv/uv_nmi.c (revision 56a0eccd)
1 /*
2  * SGI NMI support routines
3  *
4  *  This program is free software; you can redistribute it and/or modify
5  *  it under the terms of the GNU General Public License as published by
6  *  the Free Software Foundation; either version 2 of the License, or
7  *  (at your option) any later version.
8  *
9  *  This program is distributed in the hope that it will be useful,
10  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
11  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  *  GNU General Public License for more details.
13  *
14  *  You should have received a copy of the GNU General Public License
15  *  along with this program; if not, write to the Free Software
16  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
17  *
18  *  Copyright (c) 2009-2013 Silicon Graphics, Inc.  All Rights Reserved.
19  *  Copyright (c) Mike Travis
20  */
21 
22 #include <linux/cpu.h>
23 #include <linux/delay.h>
24 #include <linux/kdb.h>
25 #include <linux/kexec.h>
26 #include <linux/kgdb.h>
27 #include <linux/module.h>
28 #include <linux/nmi.h>
29 #include <linux/sched.h>
30 #include <linux/slab.h>
31 #include <linux/clocksource.h>
32 
33 #include <asm/apic.h>
34 #include <asm/current.h>
35 #include <asm/kdebug.h>
36 #include <asm/local64.h>
37 #include <asm/nmi.h>
38 #include <asm/traps.h>
39 #include <asm/uv/uv.h>
40 #include <asm/uv/uv_hub.h>
41 #include <asm/uv/uv_mmrs.h>
42 
43 /*
44  * UV handler for NMI
45  *
46  * Handle system-wide NMI events generated by the global 'power nmi' command.
47  *
48  * Basic operation is to field the NMI interrupt on each cpu and wait
49  * until all cpus have arrived into the nmi handler.  If some cpus do not
50  * make it into the handler, try and force them in with the IPI(NMI) signal.
51  *
52  * We also have to lessen UV Hub MMR accesses as much as possible as this
53  * disrupts the UV Hub's primary mission of directing NumaLink traffic and
54  * can cause system problems to occur.
55  *
56  * To do this we register our primary NMI notifier on the NMI_UNKNOWN
57  * chain.  This reduces the number of false NMI calls when the perf
58  * tools are running which generate an enormous number of NMIs per
59  * second (~4M/s for 1024 cpu threads).  Our secondary NMI handler is
60  * very short as it only checks that if it has been "pinged" with the
61  * IPI(NMI) signal as mentioned above, and does not read the UV Hub's MMR.
62  *
63  */
64 
65 static struct uv_hub_nmi_s **uv_hub_nmi_list;
66 
67 DEFINE_PER_CPU(struct uv_cpu_nmi_s, uv_cpu_nmi);
68 EXPORT_PER_CPU_SYMBOL_GPL(uv_cpu_nmi);
69 
70 static unsigned long nmi_mmr;
71 static unsigned long nmi_mmr_clear;
72 static unsigned long nmi_mmr_pending;
73 
74 static atomic_t	uv_in_nmi;
75 static atomic_t uv_nmi_cpu = ATOMIC_INIT(-1);
76 static atomic_t uv_nmi_cpus_in_nmi = ATOMIC_INIT(-1);
77 static atomic_t uv_nmi_slave_continue;
78 static cpumask_var_t uv_nmi_cpu_mask;
79 
80 /* Values for uv_nmi_slave_continue */
81 #define SLAVE_CLEAR	0
82 #define SLAVE_CONTINUE	1
83 #define SLAVE_EXIT	2
84 
85 /*
86  * Default is all stack dumps go to the console and buffer.
87  * Lower level to send to log buffer only.
88  */
89 static int uv_nmi_loglevel = CONSOLE_LOGLEVEL_DEFAULT;
90 module_param_named(dump_loglevel, uv_nmi_loglevel, int, 0644);
91 
92 /*
93  * The following values show statistics on how perf events are affecting
94  * this system.
95  */
96 static int param_get_local64(char *buffer, const struct kernel_param *kp)
97 {
98 	return sprintf(buffer, "%lu\n", local64_read((local64_t *)kp->arg));
99 }
100 
101 static int param_set_local64(const char *val, const struct kernel_param *kp)
102 {
103 	/* clear on any write */
104 	local64_set((local64_t *)kp->arg, 0);
105 	return 0;
106 }
107 
108 static const struct kernel_param_ops param_ops_local64 = {
109 	.get = param_get_local64,
110 	.set = param_set_local64,
111 };
112 #define param_check_local64(name, p) __param_check(name, p, local64_t)
113 
114 static local64_t uv_nmi_count;
115 module_param_named(nmi_count, uv_nmi_count, local64, 0644);
116 
117 static local64_t uv_nmi_misses;
118 module_param_named(nmi_misses, uv_nmi_misses, local64, 0644);
119 
120 static local64_t uv_nmi_ping_count;
121 module_param_named(ping_count, uv_nmi_ping_count, local64, 0644);
122 
123 static local64_t uv_nmi_ping_misses;
124 module_param_named(ping_misses, uv_nmi_ping_misses, local64, 0644);
125 
126 /*
127  * Following values allow tuning for large systems under heavy loading
128  */
129 static int uv_nmi_initial_delay = 100;
130 module_param_named(initial_delay, uv_nmi_initial_delay, int, 0644);
131 
132 static int uv_nmi_slave_delay = 100;
133 module_param_named(slave_delay, uv_nmi_slave_delay, int, 0644);
134 
135 static int uv_nmi_loop_delay = 100;
136 module_param_named(loop_delay, uv_nmi_loop_delay, int, 0644);
137 
138 static int uv_nmi_trigger_delay = 10000;
139 module_param_named(trigger_delay, uv_nmi_trigger_delay, int, 0644);
140 
141 static int uv_nmi_wait_count = 100;
142 module_param_named(wait_count, uv_nmi_wait_count, int, 0644);
143 
144 static int uv_nmi_retry_count = 500;
145 module_param_named(retry_count, uv_nmi_retry_count, int, 0644);
146 
147 /*
148  * Valid NMI Actions:
149  *  "dump"	- dump process stack for each cpu
150  *  "ips"	- dump IP info for each cpu
151  *  "kdump"	- do crash dump
152  *  "kdb"	- enter KDB (default)
153  *  "kgdb"	- enter KGDB
154  */
155 static char uv_nmi_action[8] = "kdb";
156 module_param_string(action, uv_nmi_action, sizeof(uv_nmi_action), 0644);
157 
158 static inline bool uv_nmi_action_is(const char *action)
159 {
160 	return (strncmp(uv_nmi_action, action, strlen(action)) == 0);
161 }
162 
163 /* Setup which NMI support is present in system */
164 static void uv_nmi_setup_mmrs(void)
165 {
166 	if (uv_read_local_mmr(UVH_NMI_MMRX_SUPPORTED)) {
167 		uv_write_local_mmr(UVH_NMI_MMRX_REQ,
168 					1UL << UVH_NMI_MMRX_REQ_SHIFT);
169 		nmi_mmr = UVH_NMI_MMRX;
170 		nmi_mmr_clear = UVH_NMI_MMRX_CLEAR;
171 		nmi_mmr_pending = 1UL << UVH_NMI_MMRX_SHIFT;
172 		pr_info("UV: SMI NMI support: %s\n", UVH_NMI_MMRX_TYPE);
173 	} else {
174 		nmi_mmr = UVH_NMI_MMR;
175 		nmi_mmr_clear = UVH_NMI_MMR_CLEAR;
176 		nmi_mmr_pending = 1UL << UVH_NMI_MMR_SHIFT;
177 		pr_info("UV: SMI NMI support: %s\n", UVH_NMI_MMR_TYPE);
178 	}
179 }
180 
181 /* Read NMI MMR and check if NMI flag was set by BMC. */
182 static inline int uv_nmi_test_mmr(struct uv_hub_nmi_s *hub_nmi)
183 {
184 	hub_nmi->nmi_value = uv_read_local_mmr(nmi_mmr);
185 	atomic_inc(&hub_nmi->read_mmr_count);
186 	return !!(hub_nmi->nmi_value & nmi_mmr_pending);
187 }
188 
189 static inline void uv_local_mmr_clear_nmi(void)
190 {
191 	uv_write_local_mmr(nmi_mmr_clear, nmi_mmr_pending);
192 }
193 
194 /*
195  * If first cpu in on this hub, set hub_nmi "in_nmi" and "owner" values and
196  * return true.  If first cpu in on the system, set global "in_nmi" flag.
197  */
198 static int uv_set_in_nmi(int cpu, struct uv_hub_nmi_s *hub_nmi)
199 {
200 	int first = atomic_add_unless(&hub_nmi->in_nmi, 1, 1);
201 
202 	if (first) {
203 		atomic_set(&hub_nmi->cpu_owner, cpu);
204 		if (atomic_add_unless(&uv_in_nmi, 1, 1))
205 			atomic_set(&uv_nmi_cpu, cpu);
206 
207 		atomic_inc(&hub_nmi->nmi_count);
208 	}
209 	return first;
210 }
211 
212 /* Check if this is a system NMI event */
213 static int uv_check_nmi(struct uv_hub_nmi_s *hub_nmi)
214 {
215 	int cpu = smp_processor_id();
216 	int nmi = 0;
217 
218 	local64_inc(&uv_nmi_count);
219 	this_cpu_inc(uv_cpu_nmi.queries);
220 
221 	do {
222 		nmi = atomic_read(&hub_nmi->in_nmi);
223 		if (nmi)
224 			break;
225 
226 		if (raw_spin_trylock(&hub_nmi->nmi_lock)) {
227 
228 			/* check hub MMR NMI flag */
229 			if (uv_nmi_test_mmr(hub_nmi)) {
230 				uv_set_in_nmi(cpu, hub_nmi);
231 				nmi = 1;
232 				break;
233 			}
234 
235 			/* MMR NMI flag is clear */
236 			raw_spin_unlock(&hub_nmi->nmi_lock);
237 
238 		} else {
239 			/* wait a moment for the hub nmi locker to set flag */
240 			cpu_relax();
241 			udelay(uv_nmi_slave_delay);
242 
243 			/* re-check hub in_nmi flag */
244 			nmi = atomic_read(&hub_nmi->in_nmi);
245 			if (nmi)
246 				break;
247 		}
248 
249 		/* check if this BMC missed setting the MMR NMI flag */
250 		if (!nmi) {
251 			nmi = atomic_read(&uv_in_nmi);
252 			if (nmi)
253 				uv_set_in_nmi(cpu, hub_nmi);
254 		}
255 
256 	} while (0);
257 
258 	if (!nmi)
259 		local64_inc(&uv_nmi_misses);
260 
261 	return nmi;
262 }
263 
264 /* Need to reset the NMI MMR register, but only once per hub. */
265 static inline void uv_clear_nmi(int cpu)
266 {
267 	struct uv_hub_nmi_s *hub_nmi = uv_hub_nmi;
268 
269 	if (cpu == atomic_read(&hub_nmi->cpu_owner)) {
270 		atomic_set(&hub_nmi->cpu_owner, -1);
271 		atomic_set(&hub_nmi->in_nmi, 0);
272 		uv_local_mmr_clear_nmi();
273 		raw_spin_unlock(&hub_nmi->nmi_lock);
274 	}
275 }
276 
277 /* Ping non-responding cpus attemping to force them into the NMI handler */
278 static void uv_nmi_nr_cpus_ping(void)
279 {
280 	int cpu;
281 
282 	for_each_cpu(cpu, uv_nmi_cpu_mask)
283 		uv_cpu_nmi_per(cpu).pinging = 1;
284 
285 	apic->send_IPI_mask(uv_nmi_cpu_mask, APIC_DM_NMI);
286 }
287 
288 /* Clean up flags for cpus that ignored both NMI and ping */
289 static void uv_nmi_cleanup_mask(void)
290 {
291 	int cpu;
292 
293 	for_each_cpu(cpu, uv_nmi_cpu_mask) {
294 		uv_cpu_nmi_per(cpu).pinging =  0;
295 		uv_cpu_nmi_per(cpu).state = UV_NMI_STATE_OUT;
296 		cpumask_clear_cpu(cpu, uv_nmi_cpu_mask);
297 	}
298 }
299 
300 /* Loop waiting as cpus enter nmi handler */
301 static int uv_nmi_wait_cpus(int first)
302 {
303 	int i, j, k, n = num_online_cpus();
304 	int last_k = 0, waiting = 0;
305 
306 	if (first) {
307 		cpumask_copy(uv_nmi_cpu_mask, cpu_online_mask);
308 		k = 0;
309 	} else {
310 		k = n - cpumask_weight(uv_nmi_cpu_mask);
311 	}
312 
313 	udelay(uv_nmi_initial_delay);
314 	for (i = 0; i < uv_nmi_retry_count; i++) {
315 		int loop_delay = uv_nmi_loop_delay;
316 
317 		for_each_cpu(j, uv_nmi_cpu_mask) {
318 			if (uv_cpu_nmi_per(j).state) {
319 				cpumask_clear_cpu(j, uv_nmi_cpu_mask);
320 				if (++k >= n)
321 					break;
322 			}
323 		}
324 		if (k >= n) {		/* all in? */
325 			k = n;
326 			break;
327 		}
328 		if (last_k != k) {	/* abort if no new cpus coming in */
329 			last_k = k;
330 			waiting = 0;
331 		} else if (++waiting > uv_nmi_wait_count)
332 			break;
333 
334 		/* extend delay if waiting only for cpu 0 */
335 		if (waiting && (n - k) == 1 &&
336 		    cpumask_test_cpu(0, uv_nmi_cpu_mask))
337 			loop_delay *= 100;
338 
339 		udelay(loop_delay);
340 	}
341 	atomic_set(&uv_nmi_cpus_in_nmi, k);
342 	return n - k;
343 }
344 
345 /* Wait until all slave cpus have entered UV NMI handler */
346 static void uv_nmi_wait(int master)
347 {
348 	/* indicate this cpu is in */
349 	this_cpu_write(uv_cpu_nmi.state, UV_NMI_STATE_IN);
350 
351 	/* if not the first cpu in (the master), then we are a slave cpu */
352 	if (!master)
353 		return;
354 
355 	do {
356 		/* wait for all other cpus to gather here */
357 		if (!uv_nmi_wait_cpus(1))
358 			break;
359 
360 		/* if not all made it in, send IPI NMI to them */
361 		pr_alert("UV: Sending NMI IPI to %d non-responding CPUs: %*pbl\n",
362 			 cpumask_weight(uv_nmi_cpu_mask),
363 			 cpumask_pr_args(uv_nmi_cpu_mask));
364 
365 		uv_nmi_nr_cpus_ping();
366 
367 		/* if all cpus are in, then done */
368 		if (!uv_nmi_wait_cpus(0))
369 			break;
370 
371 		pr_alert("UV: %d CPUs not in NMI loop: %*pbl\n",
372 			 cpumask_weight(uv_nmi_cpu_mask),
373 			 cpumask_pr_args(uv_nmi_cpu_mask));
374 	} while (0);
375 
376 	pr_alert("UV: %d of %d CPUs in NMI\n",
377 		atomic_read(&uv_nmi_cpus_in_nmi), num_online_cpus());
378 }
379 
380 /* Dump Instruction Pointer header */
381 static void uv_nmi_dump_cpu_ip_hdr(void)
382 {
383 	pr_info("\nUV: %4s %6s %-32s %s   (Note: PID 0 not listed)\n",
384 		"CPU", "PID", "COMMAND", "IP");
385 }
386 
387 /* Dump Instruction Pointer info */
388 static void uv_nmi_dump_cpu_ip(int cpu, struct pt_regs *regs)
389 {
390 	pr_info("UV: %4d %6d %-32.32s ", cpu, current->pid, current->comm);
391 	printk_address(regs->ip);
392 }
393 
394 /*
395  * Dump this CPU's state.  If action was set to "kdump" and the crash_kexec
396  * failed, then we provide "dump" as an alternate action.  Action "dump" now
397  * also includes the show "ips" (instruction pointers) action whereas the
398  * action "ips" only displays instruction pointers for the non-idle CPU's.
399  * This is an abbreviated form of the "ps" command.
400  */
401 static void uv_nmi_dump_state_cpu(int cpu, struct pt_regs *regs)
402 {
403 	const char *dots = " ................................. ";
404 
405 	if (cpu == 0)
406 		uv_nmi_dump_cpu_ip_hdr();
407 
408 	if (current->pid != 0 || !uv_nmi_action_is("ips"))
409 		uv_nmi_dump_cpu_ip(cpu, regs);
410 
411 	if (uv_nmi_action_is("dump")) {
412 		pr_info("UV:%sNMI process trace for CPU %d\n", dots, cpu);
413 		show_regs(regs);
414 	}
415 
416 	this_cpu_write(uv_cpu_nmi.state, UV_NMI_STATE_DUMP_DONE);
417 }
418 
419 /* Trigger a slave cpu to dump it's state */
420 static void uv_nmi_trigger_dump(int cpu)
421 {
422 	int retry = uv_nmi_trigger_delay;
423 
424 	if (uv_cpu_nmi_per(cpu).state != UV_NMI_STATE_IN)
425 		return;
426 
427 	uv_cpu_nmi_per(cpu).state = UV_NMI_STATE_DUMP;
428 	do {
429 		cpu_relax();
430 		udelay(10);
431 		if (uv_cpu_nmi_per(cpu).state
432 				!= UV_NMI_STATE_DUMP)
433 			return;
434 	} while (--retry > 0);
435 
436 	pr_crit("UV: CPU %d stuck in process dump function\n", cpu);
437 	uv_cpu_nmi_per(cpu).state = UV_NMI_STATE_DUMP_DONE;
438 }
439 
440 /* Wait until all cpus ready to exit */
441 static void uv_nmi_sync_exit(int master)
442 {
443 	atomic_dec(&uv_nmi_cpus_in_nmi);
444 	if (master) {
445 		while (atomic_read(&uv_nmi_cpus_in_nmi) > 0)
446 			cpu_relax();
447 		atomic_set(&uv_nmi_slave_continue, SLAVE_CLEAR);
448 	} else {
449 		while (atomic_read(&uv_nmi_slave_continue))
450 			cpu_relax();
451 	}
452 }
453 
454 /* Walk through cpu list and dump state of each */
455 static void uv_nmi_dump_state(int cpu, struct pt_regs *regs, int master)
456 {
457 	if (master) {
458 		int tcpu;
459 		int ignored = 0;
460 		int saved_console_loglevel = console_loglevel;
461 
462 		pr_alert("UV: tracing %s for %d CPUs from CPU %d\n",
463 			uv_nmi_action_is("ips") ? "IPs" : "processes",
464 			atomic_read(&uv_nmi_cpus_in_nmi), cpu);
465 
466 		console_loglevel = uv_nmi_loglevel;
467 		atomic_set(&uv_nmi_slave_continue, SLAVE_EXIT);
468 		for_each_online_cpu(tcpu) {
469 			if (cpumask_test_cpu(tcpu, uv_nmi_cpu_mask))
470 				ignored++;
471 			else if (tcpu == cpu)
472 				uv_nmi_dump_state_cpu(tcpu, regs);
473 			else
474 				uv_nmi_trigger_dump(tcpu);
475 		}
476 		if (ignored)
477 			pr_alert("UV: %d CPUs ignored NMI\n", ignored);
478 
479 		console_loglevel = saved_console_loglevel;
480 		pr_alert("UV: process trace complete\n");
481 	} else {
482 		while (!atomic_read(&uv_nmi_slave_continue))
483 			cpu_relax();
484 		while (this_cpu_read(uv_cpu_nmi.state) != UV_NMI_STATE_DUMP)
485 			cpu_relax();
486 		uv_nmi_dump_state_cpu(cpu, regs);
487 	}
488 	uv_nmi_sync_exit(master);
489 }
490 
491 static void uv_nmi_touch_watchdogs(void)
492 {
493 	touch_softlockup_watchdog_sync();
494 	clocksource_touch_watchdog();
495 	rcu_cpu_stall_reset();
496 	touch_nmi_watchdog();
497 }
498 
499 static atomic_t uv_nmi_kexec_failed;
500 
501 #if defined(CONFIG_KEXEC_CORE)
502 static void uv_nmi_kdump(int cpu, int master, struct pt_regs *regs)
503 {
504 	/* Call crash to dump system state */
505 	if (master) {
506 		pr_emerg("UV: NMI executing crash_kexec on CPU%d\n", cpu);
507 		crash_kexec(regs);
508 
509 		pr_emerg("UV: crash_kexec unexpectedly returned, ");
510 		atomic_set(&uv_nmi_kexec_failed, 1);
511 		if (!kexec_crash_image) {
512 			pr_cont("crash kernel not loaded\n");
513 			return;
514 		}
515 		pr_cont("kexec busy, stalling cpus while waiting\n");
516 	}
517 
518 	/* If crash exec fails the slaves should return, otherwise stall */
519 	while (atomic_read(&uv_nmi_kexec_failed) == 0)
520 		mdelay(10);
521 }
522 
523 #else /* !CONFIG_KEXEC_CORE */
524 static inline void uv_nmi_kdump(int cpu, int master, struct pt_regs *regs)
525 {
526 	if (master)
527 		pr_err("UV: NMI kdump: KEXEC not supported in this kernel\n");
528 	atomic_set(&uv_nmi_kexec_failed, 1);
529 }
530 #endif /* !CONFIG_KEXEC_CORE */
531 
532 #ifdef CONFIG_KGDB
533 #ifdef CONFIG_KGDB_KDB
534 static inline int uv_nmi_kdb_reason(void)
535 {
536 	return KDB_REASON_SYSTEM_NMI;
537 }
538 #else /* !CONFIG_KGDB_KDB */
539 static inline int uv_nmi_kdb_reason(void)
540 {
541 	/* Insure user is expecting to attach gdb remote */
542 	if (uv_nmi_action_is("kgdb"))
543 		return 0;
544 
545 	pr_err("UV: NMI error: KDB is not enabled in this kernel\n");
546 	return -1;
547 }
548 #endif /* CONFIG_KGDB_KDB */
549 
550 /*
551  * Call KGDB/KDB from NMI handler
552  *
553  * Note that if both KGDB and KDB are configured, then the action of 'kgdb' or
554  * 'kdb' has no affect on which is used.  See the KGDB documention for further
555  * information.
556  */
557 static void uv_call_kgdb_kdb(int cpu, struct pt_regs *regs, int master)
558 {
559 	if (master) {
560 		int reason = uv_nmi_kdb_reason();
561 		int ret;
562 
563 		if (reason < 0)
564 			return;
565 
566 		/* call KGDB NMI handler as MASTER */
567 		ret = kgdb_nmicallin(cpu, X86_TRAP_NMI, regs, reason,
568 				&uv_nmi_slave_continue);
569 		if (ret) {
570 			pr_alert("KGDB returned error, is kgdboc set?\n");
571 			atomic_set(&uv_nmi_slave_continue, SLAVE_EXIT);
572 		}
573 	} else {
574 		/* wait for KGDB signal that it's ready for slaves to enter */
575 		int sig;
576 
577 		do {
578 			cpu_relax();
579 			sig = atomic_read(&uv_nmi_slave_continue);
580 		} while (!sig);
581 
582 		/* call KGDB as slave */
583 		if (sig == SLAVE_CONTINUE)
584 			kgdb_nmicallback(cpu, regs);
585 	}
586 	uv_nmi_sync_exit(master);
587 }
588 
589 #else /* !CONFIG_KGDB */
590 static inline void uv_call_kgdb_kdb(int cpu, struct pt_regs *regs, int master)
591 {
592 	pr_err("UV: NMI error: KGDB is not enabled in this kernel\n");
593 }
594 #endif /* !CONFIG_KGDB */
595 
596 /*
597  * UV NMI handler
598  */
599 int uv_handle_nmi(unsigned int reason, struct pt_regs *regs)
600 {
601 	struct uv_hub_nmi_s *hub_nmi = uv_hub_nmi;
602 	int cpu = smp_processor_id();
603 	int master = 0;
604 	unsigned long flags;
605 
606 	local_irq_save(flags);
607 
608 	/* If not a UV System NMI, ignore */
609 	if (!this_cpu_read(uv_cpu_nmi.pinging) && !uv_check_nmi(hub_nmi)) {
610 		local_irq_restore(flags);
611 		return NMI_DONE;
612 	}
613 
614 	/* Indicate we are the first CPU into the NMI handler */
615 	master = (atomic_read(&uv_nmi_cpu) == cpu);
616 
617 	/* If NMI action is "kdump", then attempt to do it */
618 	if (uv_nmi_action_is("kdump")) {
619 		uv_nmi_kdump(cpu, master, regs);
620 
621 		/* Unexpected return, revert action to "dump" */
622 		if (master)
623 			strncpy(uv_nmi_action, "dump", strlen(uv_nmi_action));
624 	}
625 
626 	/* Pause as all cpus enter the NMI handler */
627 	uv_nmi_wait(master);
628 
629 	/* Dump state of each cpu */
630 	if (uv_nmi_action_is("ips") || uv_nmi_action_is("dump"))
631 		uv_nmi_dump_state(cpu, regs, master);
632 
633 	/* Call KGDB/KDB if enabled */
634 	else if (uv_nmi_action_is("kdb") || uv_nmi_action_is("kgdb"))
635 		uv_call_kgdb_kdb(cpu, regs, master);
636 
637 	/* Clear per_cpu "in nmi" flag */
638 	this_cpu_write(uv_cpu_nmi.state, UV_NMI_STATE_OUT);
639 
640 	/* Clear MMR NMI flag on each hub */
641 	uv_clear_nmi(cpu);
642 
643 	/* Clear global flags */
644 	if (master) {
645 		if (cpumask_weight(uv_nmi_cpu_mask))
646 			uv_nmi_cleanup_mask();
647 		atomic_set(&uv_nmi_cpus_in_nmi, -1);
648 		atomic_set(&uv_nmi_cpu, -1);
649 		atomic_set(&uv_in_nmi, 0);
650 		atomic_set(&uv_nmi_kexec_failed, 0);
651 	}
652 
653 	uv_nmi_touch_watchdogs();
654 	local_irq_restore(flags);
655 
656 	return NMI_HANDLED;
657 }
658 
659 /*
660  * NMI handler for pulling in CPUs when perf events are grabbing our NMI
661  */
662 static int uv_handle_nmi_ping(unsigned int reason, struct pt_regs *regs)
663 {
664 	int ret;
665 
666 	this_cpu_inc(uv_cpu_nmi.queries);
667 	if (!this_cpu_read(uv_cpu_nmi.pinging)) {
668 		local64_inc(&uv_nmi_ping_misses);
669 		return NMI_DONE;
670 	}
671 
672 	this_cpu_inc(uv_cpu_nmi.pings);
673 	local64_inc(&uv_nmi_ping_count);
674 	ret = uv_handle_nmi(reason, regs);
675 	this_cpu_write(uv_cpu_nmi.pinging, 0);
676 	return ret;
677 }
678 
679 static void uv_register_nmi_notifier(void)
680 {
681 	if (register_nmi_handler(NMI_UNKNOWN, uv_handle_nmi, 0, "uv"))
682 		pr_warn("UV: NMI handler failed to register\n");
683 
684 	if (register_nmi_handler(NMI_LOCAL, uv_handle_nmi_ping, 0, "uvping"))
685 		pr_warn("UV: PING NMI handler failed to register\n");
686 }
687 
688 void uv_nmi_init(void)
689 {
690 	unsigned int value;
691 
692 	/*
693 	 * Unmask NMI on all cpus
694 	 */
695 	value = apic_read(APIC_LVT1) | APIC_DM_NMI;
696 	value &= ~APIC_LVT_MASKED;
697 	apic_write(APIC_LVT1, value);
698 }
699 
700 void uv_nmi_setup(void)
701 {
702 	int size = sizeof(void *) * (1 << NODES_SHIFT);
703 	int cpu, nid;
704 
705 	/* Setup hub nmi info */
706 	uv_nmi_setup_mmrs();
707 	uv_hub_nmi_list = kzalloc(size, GFP_KERNEL);
708 	pr_info("UV: NMI hub list @ 0x%p (%d)\n", uv_hub_nmi_list, size);
709 	BUG_ON(!uv_hub_nmi_list);
710 	size = sizeof(struct uv_hub_nmi_s);
711 	for_each_present_cpu(cpu) {
712 		nid = cpu_to_node(cpu);
713 		if (uv_hub_nmi_list[nid] == NULL) {
714 			uv_hub_nmi_list[nid] = kzalloc_node(size,
715 							    GFP_KERNEL, nid);
716 			BUG_ON(!uv_hub_nmi_list[nid]);
717 			raw_spin_lock_init(&(uv_hub_nmi_list[nid]->nmi_lock));
718 			atomic_set(&uv_hub_nmi_list[nid]->cpu_owner, -1);
719 		}
720 		uv_hub_nmi_per(cpu) = uv_hub_nmi_list[nid];
721 	}
722 	BUG_ON(!alloc_cpumask_var(&uv_nmi_cpu_mask, GFP_KERNEL));
723 	uv_register_nmi_notifier();
724 }
725