xref: /linux/arch/x86/platform/uv/uv_nmi.c (revision 3932b9ca55b0be314a36d3e84faff3e823c081f5)
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 
32 #include <asm/apic.h>
33 #include <asm/current.h>
34 #include <asm/kdebug.h>
35 #include <asm/local64.h>
36 #include <asm/nmi.h>
37 #include <asm/traps.h>
38 #include <asm/uv/uv.h>
39 #include <asm/uv/uv_hub.h>
40 #include <asm/uv/uv_mmrs.h>
41 
42 /*
43  * UV handler for NMI
44  *
45  * Handle system-wide NMI events generated by the global 'power nmi' command.
46  *
47  * Basic operation is to field the NMI interrupt on each cpu and wait
48  * until all cpus have arrived into the nmi handler.  If some cpus do not
49  * make it into the handler, try and force them in with the IPI(NMI) signal.
50  *
51  * We also have to lessen UV Hub MMR accesses as much as possible as this
52  * disrupts the UV Hub's primary mission of directing NumaLink traffic and
53  * can cause system problems to occur.
54  *
55  * To do this we register our primary NMI notifier on the NMI_UNKNOWN
56  * chain.  This reduces the number of false NMI calls when the perf
57  * tools are running which generate an enormous number of NMIs per
58  * second (~4M/s for 1024 cpu threads).  Our secondary NMI handler is
59  * very short as it only checks that if it has been "pinged" with the
60  * IPI(NMI) signal as mentioned above, and does not read the UV Hub's MMR.
61  *
62  */
63 
64 static struct uv_hub_nmi_s **uv_hub_nmi_list;
65 
66 DEFINE_PER_CPU(struct uv_cpu_nmi_s, __uv_cpu_nmi);
67 EXPORT_PER_CPU_SYMBOL_GPL(__uv_cpu_nmi);
68 
69 static unsigned long nmi_mmr;
70 static unsigned long nmi_mmr_clear;
71 static unsigned long nmi_mmr_pending;
72 
73 static atomic_t	uv_in_nmi;
74 static atomic_t uv_nmi_cpu = ATOMIC_INIT(-1);
75 static atomic_t uv_nmi_cpus_in_nmi = ATOMIC_INIT(-1);
76 static atomic_t uv_nmi_slave_continue;
77 static cpumask_var_t uv_nmi_cpu_mask;
78 
79 /* Values for uv_nmi_slave_continue */
80 #define SLAVE_CLEAR	0
81 #define SLAVE_CONTINUE	1
82 #define SLAVE_EXIT	2
83 
84 /*
85  * Default is all stack dumps go to the console and buffer.
86  * Lower level to send to log buffer only.
87  */
88 static int uv_nmi_loglevel = CONSOLE_LOGLEVEL_DEFAULT;
89 module_param_named(dump_loglevel, uv_nmi_loglevel, int, 0644);
90 
91 /*
92  * The following values show statistics on how perf events are affecting
93  * this system.
94  */
95 static int param_get_local64(char *buffer, const struct kernel_param *kp)
96 {
97 	return sprintf(buffer, "%lu\n", local64_read((local64_t *)kp->arg));
98 }
99 
100 static int param_set_local64(const char *val, const struct kernel_param *kp)
101 {
102 	/* clear on any write */
103 	local64_set((local64_t *)kp->arg, 0);
104 	return 0;
105 }
106 
107 static struct kernel_param_ops param_ops_local64 = {
108 	.get = param_get_local64,
109 	.set = param_set_local64,
110 };
111 #define param_check_local64(name, p) __param_check(name, p, local64_t)
112 
113 static local64_t uv_nmi_count;
114 module_param_named(nmi_count, uv_nmi_count, local64, 0644);
115 
116 static local64_t uv_nmi_misses;
117 module_param_named(nmi_misses, uv_nmi_misses, local64, 0644);
118 
119 static local64_t uv_nmi_ping_count;
120 module_param_named(ping_count, uv_nmi_ping_count, local64, 0644);
121 
122 static local64_t uv_nmi_ping_misses;
123 module_param_named(ping_misses, uv_nmi_ping_misses, local64, 0644);
124 
125 /*
126  * Following values allow tuning for large systems under heavy loading
127  */
128 static int uv_nmi_initial_delay = 100;
129 module_param_named(initial_delay, uv_nmi_initial_delay, int, 0644);
130 
131 static int uv_nmi_slave_delay = 100;
132 module_param_named(slave_delay, uv_nmi_slave_delay, int, 0644);
133 
134 static int uv_nmi_loop_delay = 100;
135 module_param_named(loop_delay, uv_nmi_loop_delay, int, 0644);
136 
137 static int uv_nmi_trigger_delay = 10000;
138 module_param_named(trigger_delay, uv_nmi_trigger_delay, int, 0644);
139 
140 static int uv_nmi_wait_count = 100;
141 module_param_named(wait_count, uv_nmi_wait_count, int, 0644);
142 
143 static int uv_nmi_retry_count = 500;
144 module_param_named(retry_count, uv_nmi_retry_count, int, 0644);
145 
146 /*
147  * Valid NMI Actions:
148  *  "dump"	- dump process stack for each cpu
149  *  "ips"	- dump IP info for each cpu
150  *  "kdump"	- do crash dump
151  *  "kdb"	- enter KDB (default)
152  *  "kgdb"	- enter KGDB
153  */
154 static char uv_nmi_action[8] = "kdb";
155 module_param_string(action, uv_nmi_action, sizeof(uv_nmi_action), 0644);
156 
157 static inline bool uv_nmi_action_is(const char *action)
158 {
159 	return (strncmp(uv_nmi_action, action, strlen(action)) == 0);
160 }
161 
162 /* Setup which NMI support is present in system */
163 static void uv_nmi_setup_mmrs(void)
164 {
165 	if (uv_read_local_mmr(UVH_NMI_MMRX_SUPPORTED)) {
166 		uv_write_local_mmr(UVH_NMI_MMRX_REQ,
167 					1UL << UVH_NMI_MMRX_REQ_SHIFT);
168 		nmi_mmr = UVH_NMI_MMRX;
169 		nmi_mmr_clear = UVH_NMI_MMRX_CLEAR;
170 		nmi_mmr_pending = 1UL << UVH_NMI_MMRX_SHIFT;
171 		pr_info("UV: SMI NMI support: %s\n", UVH_NMI_MMRX_TYPE);
172 	} else {
173 		nmi_mmr = UVH_NMI_MMR;
174 		nmi_mmr_clear = UVH_NMI_MMR_CLEAR;
175 		nmi_mmr_pending = 1UL << UVH_NMI_MMR_SHIFT;
176 		pr_info("UV: SMI NMI support: %s\n", UVH_NMI_MMR_TYPE);
177 	}
178 }
179 
180 /* Read NMI MMR and check if NMI flag was set by BMC. */
181 static inline int uv_nmi_test_mmr(struct uv_hub_nmi_s *hub_nmi)
182 {
183 	hub_nmi->nmi_value = uv_read_local_mmr(nmi_mmr);
184 	atomic_inc(&hub_nmi->read_mmr_count);
185 	return !!(hub_nmi->nmi_value & nmi_mmr_pending);
186 }
187 
188 static inline void uv_local_mmr_clear_nmi(void)
189 {
190 	uv_write_local_mmr(nmi_mmr_clear, nmi_mmr_pending);
191 }
192 
193 /*
194  * If first cpu in on this hub, set hub_nmi "in_nmi" and "owner" values and
195  * return true.  If first cpu in on the system, set global "in_nmi" flag.
196  */
197 static int uv_set_in_nmi(int cpu, struct uv_hub_nmi_s *hub_nmi)
198 {
199 	int first = atomic_add_unless(&hub_nmi->in_nmi, 1, 1);
200 
201 	if (first) {
202 		atomic_set(&hub_nmi->cpu_owner, cpu);
203 		if (atomic_add_unless(&uv_in_nmi, 1, 1))
204 			atomic_set(&uv_nmi_cpu, cpu);
205 
206 		atomic_inc(&hub_nmi->nmi_count);
207 	}
208 	return first;
209 }
210 
211 /* Check if this is a system NMI event */
212 static int uv_check_nmi(struct uv_hub_nmi_s *hub_nmi)
213 {
214 	int cpu = smp_processor_id();
215 	int nmi = 0;
216 
217 	local64_inc(&uv_nmi_count);
218 	uv_cpu_nmi.queries++;
219 
220 	do {
221 		nmi = atomic_read(&hub_nmi->in_nmi);
222 		if (nmi)
223 			break;
224 
225 		if (raw_spin_trylock(&hub_nmi->nmi_lock)) {
226 
227 			/* check hub MMR NMI flag */
228 			if (uv_nmi_test_mmr(hub_nmi)) {
229 				uv_set_in_nmi(cpu, hub_nmi);
230 				nmi = 1;
231 				break;
232 			}
233 
234 			/* MMR NMI flag is clear */
235 			raw_spin_unlock(&hub_nmi->nmi_lock);
236 
237 		} else {
238 			/* wait a moment for the hub nmi locker to set flag */
239 			cpu_relax();
240 			udelay(uv_nmi_slave_delay);
241 
242 			/* re-check hub in_nmi flag */
243 			nmi = atomic_read(&hub_nmi->in_nmi);
244 			if (nmi)
245 				break;
246 		}
247 
248 		/* check if this BMC missed setting the MMR NMI flag */
249 		if (!nmi) {
250 			nmi = atomic_read(&uv_in_nmi);
251 			if (nmi)
252 				uv_set_in_nmi(cpu, hub_nmi);
253 		}
254 
255 	} while (0);
256 
257 	if (!nmi)
258 		local64_inc(&uv_nmi_misses);
259 
260 	return nmi;
261 }
262 
263 /* Need to reset the NMI MMR register, but only once per hub. */
264 static inline void uv_clear_nmi(int cpu)
265 {
266 	struct uv_hub_nmi_s *hub_nmi = uv_hub_nmi;
267 
268 	if (cpu == atomic_read(&hub_nmi->cpu_owner)) {
269 		atomic_set(&hub_nmi->cpu_owner, -1);
270 		atomic_set(&hub_nmi->in_nmi, 0);
271 		uv_local_mmr_clear_nmi();
272 		raw_spin_unlock(&hub_nmi->nmi_lock);
273 	}
274 }
275 
276 /* Print non-responding cpus */
277 static void uv_nmi_nr_cpus_pr(char *fmt)
278 {
279 	static char cpu_list[1024];
280 	int len = sizeof(cpu_list);
281 	int c = cpumask_weight(uv_nmi_cpu_mask);
282 	int n = cpulist_scnprintf(cpu_list, len, uv_nmi_cpu_mask);
283 
284 	if (n >= len-1)
285 		strcpy(&cpu_list[len - 6], "...\n");
286 
287 	printk(fmt, c, cpu_list);
288 }
289 
290 /* Ping non-responding cpus attemping to force them into the NMI handler */
291 static void uv_nmi_nr_cpus_ping(void)
292 {
293 	int cpu;
294 
295 	for_each_cpu(cpu, uv_nmi_cpu_mask)
296 		atomic_set(&uv_cpu_nmi_per(cpu).pinging, 1);
297 
298 	apic->send_IPI_mask(uv_nmi_cpu_mask, APIC_DM_NMI);
299 }
300 
301 /* Clean up flags for cpus that ignored both NMI and ping */
302 static void uv_nmi_cleanup_mask(void)
303 {
304 	int cpu;
305 
306 	for_each_cpu(cpu, uv_nmi_cpu_mask) {
307 		atomic_set(&uv_cpu_nmi_per(cpu).pinging, 0);
308 		atomic_set(&uv_cpu_nmi_per(cpu).state, UV_NMI_STATE_OUT);
309 		cpumask_clear_cpu(cpu, uv_nmi_cpu_mask);
310 	}
311 }
312 
313 /* Loop waiting as cpus enter nmi handler */
314 static int uv_nmi_wait_cpus(int first)
315 {
316 	int i, j, k, n = num_online_cpus();
317 	int last_k = 0, waiting = 0;
318 
319 	if (first) {
320 		cpumask_copy(uv_nmi_cpu_mask, cpu_online_mask);
321 		k = 0;
322 	} else {
323 		k = n - cpumask_weight(uv_nmi_cpu_mask);
324 	}
325 
326 	udelay(uv_nmi_initial_delay);
327 	for (i = 0; i < uv_nmi_retry_count; i++) {
328 		int loop_delay = uv_nmi_loop_delay;
329 
330 		for_each_cpu(j, uv_nmi_cpu_mask) {
331 			if (atomic_read(&uv_cpu_nmi_per(j).state)) {
332 				cpumask_clear_cpu(j, uv_nmi_cpu_mask);
333 				if (++k >= n)
334 					break;
335 			}
336 		}
337 		if (k >= n) {		/* all in? */
338 			k = n;
339 			break;
340 		}
341 		if (last_k != k) {	/* abort if no new cpus coming in */
342 			last_k = k;
343 			waiting = 0;
344 		} else if (++waiting > uv_nmi_wait_count)
345 			break;
346 
347 		/* extend delay if waiting only for cpu 0 */
348 		if (waiting && (n - k) == 1 &&
349 		    cpumask_test_cpu(0, uv_nmi_cpu_mask))
350 			loop_delay *= 100;
351 
352 		udelay(loop_delay);
353 	}
354 	atomic_set(&uv_nmi_cpus_in_nmi, k);
355 	return n - k;
356 }
357 
358 /* Wait until all slave cpus have entered UV NMI handler */
359 static void uv_nmi_wait(int master)
360 {
361 	/* indicate this cpu is in */
362 	atomic_set(&uv_cpu_nmi.state, UV_NMI_STATE_IN);
363 
364 	/* if not the first cpu in (the master), then we are a slave cpu */
365 	if (!master)
366 		return;
367 
368 	do {
369 		/* wait for all other cpus to gather here */
370 		if (!uv_nmi_wait_cpus(1))
371 			break;
372 
373 		/* if not all made it in, send IPI NMI to them */
374 		uv_nmi_nr_cpus_pr(KERN_ALERT
375 			"UV: Sending NMI IPI to %d non-responding CPUs: %s\n");
376 		uv_nmi_nr_cpus_ping();
377 
378 		/* if all cpus are in, then done */
379 		if (!uv_nmi_wait_cpus(0))
380 			break;
381 
382 		uv_nmi_nr_cpus_pr(KERN_ALERT
383 			"UV: %d CPUs not in NMI loop: %s\n");
384 	} while (0);
385 
386 	pr_alert("UV: %d of %d CPUs in NMI\n",
387 		atomic_read(&uv_nmi_cpus_in_nmi), num_online_cpus());
388 }
389 
390 static void uv_nmi_dump_cpu_ip_hdr(void)
391 {
392 	printk(KERN_DEFAULT
393 		"\nUV: %4s %6s %-32s %s   (Note: PID 0 not listed)\n",
394 		"CPU", "PID", "COMMAND", "IP");
395 }
396 
397 static void uv_nmi_dump_cpu_ip(int cpu, struct pt_regs *regs)
398 {
399 	printk(KERN_DEFAULT "UV: %4d %6d %-32.32s ",
400 		cpu, current->pid, current->comm);
401 
402 	printk_address(regs->ip);
403 }
404 
405 /* Dump this cpu's state */
406 static void uv_nmi_dump_state_cpu(int cpu, struct pt_regs *regs)
407 {
408 	const char *dots = " ................................. ";
409 
410 	if (uv_nmi_action_is("ips")) {
411 		if (cpu == 0)
412 			uv_nmi_dump_cpu_ip_hdr();
413 
414 		if (current->pid != 0)
415 			uv_nmi_dump_cpu_ip(cpu, regs);
416 
417 	} else if (uv_nmi_action_is("dump")) {
418 		printk(KERN_DEFAULT
419 			"UV:%sNMI process trace for CPU %d\n", dots, cpu);
420 		show_regs(regs);
421 	}
422 	atomic_set(&uv_cpu_nmi.state, UV_NMI_STATE_DUMP_DONE);
423 }
424 
425 /* Trigger a slave cpu to dump it's state */
426 static void uv_nmi_trigger_dump(int cpu)
427 {
428 	int retry = uv_nmi_trigger_delay;
429 
430 	if (atomic_read(&uv_cpu_nmi_per(cpu).state) != UV_NMI_STATE_IN)
431 		return;
432 
433 	atomic_set(&uv_cpu_nmi_per(cpu).state, UV_NMI_STATE_DUMP);
434 	do {
435 		cpu_relax();
436 		udelay(10);
437 		if (atomic_read(&uv_cpu_nmi_per(cpu).state)
438 				!= UV_NMI_STATE_DUMP)
439 			return;
440 	} while (--retry > 0);
441 
442 	pr_crit("UV: CPU %d stuck in process dump function\n", cpu);
443 	atomic_set(&uv_cpu_nmi_per(cpu).state, UV_NMI_STATE_DUMP_DONE);
444 }
445 
446 /* Wait until all cpus ready to exit */
447 static void uv_nmi_sync_exit(int master)
448 {
449 	atomic_dec(&uv_nmi_cpus_in_nmi);
450 	if (master) {
451 		while (atomic_read(&uv_nmi_cpus_in_nmi) > 0)
452 			cpu_relax();
453 		atomic_set(&uv_nmi_slave_continue, SLAVE_CLEAR);
454 	} else {
455 		while (atomic_read(&uv_nmi_slave_continue))
456 			cpu_relax();
457 	}
458 }
459 
460 /* Walk through cpu list and dump state of each */
461 static void uv_nmi_dump_state(int cpu, struct pt_regs *regs, int master)
462 {
463 	if (master) {
464 		int tcpu;
465 		int ignored = 0;
466 		int saved_console_loglevel = console_loglevel;
467 
468 		pr_alert("UV: tracing %s for %d CPUs from CPU %d\n",
469 			uv_nmi_action_is("ips") ? "IPs" : "processes",
470 			atomic_read(&uv_nmi_cpus_in_nmi), cpu);
471 
472 		console_loglevel = uv_nmi_loglevel;
473 		atomic_set(&uv_nmi_slave_continue, SLAVE_EXIT);
474 		for_each_online_cpu(tcpu) {
475 			if (cpumask_test_cpu(tcpu, uv_nmi_cpu_mask))
476 				ignored++;
477 			else if (tcpu == cpu)
478 				uv_nmi_dump_state_cpu(tcpu, regs);
479 			else
480 				uv_nmi_trigger_dump(tcpu);
481 		}
482 		if (ignored)
483 			printk(KERN_DEFAULT "UV: %d CPUs ignored NMI\n",
484 				ignored);
485 
486 		console_loglevel = saved_console_loglevel;
487 		pr_alert("UV: process trace complete\n");
488 	} else {
489 		while (!atomic_read(&uv_nmi_slave_continue))
490 			cpu_relax();
491 		while (atomic_read(&uv_cpu_nmi.state) != UV_NMI_STATE_DUMP)
492 			cpu_relax();
493 		uv_nmi_dump_state_cpu(cpu, regs);
494 	}
495 	uv_nmi_sync_exit(master);
496 }
497 
498 static void uv_nmi_touch_watchdogs(void)
499 {
500 	touch_softlockup_watchdog_sync();
501 	clocksource_touch_watchdog();
502 	rcu_cpu_stall_reset();
503 	touch_nmi_watchdog();
504 }
505 
506 #if defined(CONFIG_KEXEC)
507 static atomic_t uv_nmi_kexec_failed;
508 static void uv_nmi_kdump(int cpu, int master, struct pt_regs *regs)
509 {
510 	/* Call crash to dump system state */
511 	if (master) {
512 		pr_emerg("UV: NMI executing crash_kexec on CPU%d\n", cpu);
513 		crash_kexec(regs);
514 
515 		pr_emerg("UV: crash_kexec unexpectedly returned, ");
516 		if (!kexec_crash_image) {
517 			pr_cont("crash kernel not loaded\n");
518 			atomic_set(&uv_nmi_kexec_failed, 1);
519 			uv_nmi_sync_exit(1);
520 			return;
521 		}
522 		pr_cont("kexec busy, stalling cpus while waiting\n");
523 	}
524 
525 	/* If crash exec fails the slaves should return, otherwise stall */
526 	while (atomic_read(&uv_nmi_kexec_failed) == 0)
527 		mdelay(10);
528 
529 	/* Crash kernel most likely not loaded, return in an orderly fashion */
530 	uv_nmi_sync_exit(0);
531 }
532 
533 #else /* !CONFIG_KEXEC */
534 static inline void uv_nmi_kdump(int cpu, int master, struct pt_regs *regs)
535 {
536 	if (master)
537 		pr_err("UV: NMI kdump: KEXEC not supported in this kernel\n");
538 }
539 #endif /* !CONFIG_KEXEC */
540 
541 #ifdef CONFIG_KGDB
542 #ifdef CONFIG_KGDB_KDB
543 static inline int uv_nmi_kdb_reason(void)
544 {
545 	return KDB_REASON_SYSTEM_NMI;
546 }
547 #else /* !CONFIG_KGDB_KDB */
548 static inline int uv_nmi_kdb_reason(void)
549 {
550 	/* Insure user is expecting to attach gdb remote */
551 	if (uv_nmi_action_is("kgdb"))
552 		return 0;
553 
554 	pr_err("UV: NMI error: KDB is not enabled in this kernel\n");
555 	return -1;
556 }
557 #endif /* CONFIG_KGDB_KDB */
558 
559 /*
560  * Call KGDB/KDB from NMI handler
561  *
562  * Note that if both KGDB and KDB are configured, then the action of 'kgdb' or
563  * 'kdb' has no affect on which is used.  See the KGDB documention for further
564  * information.
565  */
566 static void uv_call_kgdb_kdb(int cpu, struct pt_regs *regs, int master)
567 {
568 	if (master) {
569 		int reason = uv_nmi_kdb_reason();
570 		int ret;
571 
572 		if (reason < 0)
573 			return;
574 
575 		/* call KGDB NMI handler as MASTER */
576 		ret = kgdb_nmicallin(cpu, X86_TRAP_NMI, regs, reason,
577 				&uv_nmi_slave_continue);
578 		if (ret) {
579 			pr_alert("KGDB returned error, is kgdboc set?\n");
580 			atomic_set(&uv_nmi_slave_continue, SLAVE_EXIT);
581 		}
582 	} else {
583 		/* wait for KGDB signal that it's ready for slaves to enter */
584 		int sig;
585 
586 		do {
587 			cpu_relax();
588 			sig = atomic_read(&uv_nmi_slave_continue);
589 		} while (!sig);
590 
591 		/* call KGDB as slave */
592 		if (sig == SLAVE_CONTINUE)
593 			kgdb_nmicallback(cpu, regs);
594 	}
595 	uv_nmi_sync_exit(master);
596 }
597 
598 #else /* !CONFIG_KGDB */
599 static inline void uv_call_kgdb_kdb(int cpu, struct pt_regs *regs, int master)
600 {
601 	pr_err("UV: NMI error: KGDB is not enabled in this kernel\n");
602 }
603 #endif /* !CONFIG_KGDB */
604 
605 /*
606  * UV NMI handler
607  */
608 int uv_handle_nmi(unsigned int reason, struct pt_regs *regs)
609 {
610 	struct uv_hub_nmi_s *hub_nmi = uv_hub_nmi;
611 	int cpu = smp_processor_id();
612 	int master = 0;
613 	unsigned long flags;
614 
615 	local_irq_save(flags);
616 
617 	/* If not a UV System NMI, ignore */
618 	if (!atomic_read(&uv_cpu_nmi.pinging) && !uv_check_nmi(hub_nmi)) {
619 		local_irq_restore(flags);
620 		return NMI_DONE;
621 	}
622 
623 	/* Indicate we are the first CPU into the NMI handler */
624 	master = (atomic_read(&uv_nmi_cpu) == cpu);
625 
626 	/* If NMI action is "kdump", then attempt to do it */
627 	if (uv_nmi_action_is("kdump"))
628 		uv_nmi_kdump(cpu, master, regs);
629 
630 	/* Pause as all cpus enter the NMI handler */
631 	uv_nmi_wait(master);
632 
633 	/* Dump state of each cpu */
634 	if (uv_nmi_action_is("ips") || uv_nmi_action_is("dump"))
635 		uv_nmi_dump_state(cpu, regs, master);
636 
637 	/* Call KGDB/KDB if enabled */
638 	else if (uv_nmi_action_is("kdb") || uv_nmi_action_is("kgdb"))
639 		uv_call_kgdb_kdb(cpu, regs, master);
640 
641 	/* Clear per_cpu "in nmi" flag */
642 	atomic_set(&uv_cpu_nmi.state, UV_NMI_STATE_OUT);
643 
644 	/* Clear MMR NMI flag on each hub */
645 	uv_clear_nmi(cpu);
646 
647 	/* Clear global flags */
648 	if (master) {
649 		if (cpumask_weight(uv_nmi_cpu_mask))
650 			uv_nmi_cleanup_mask();
651 		atomic_set(&uv_nmi_cpus_in_nmi, -1);
652 		atomic_set(&uv_nmi_cpu, -1);
653 		atomic_set(&uv_in_nmi, 0);
654 	}
655 
656 	uv_nmi_touch_watchdogs();
657 	local_irq_restore(flags);
658 
659 	return NMI_HANDLED;
660 }
661 
662 /*
663  * NMI handler for pulling in CPUs when perf events are grabbing our NMI
664  */
665 static int uv_handle_nmi_ping(unsigned int reason, struct pt_regs *regs)
666 {
667 	int ret;
668 
669 	uv_cpu_nmi.queries++;
670 	if (!atomic_read(&uv_cpu_nmi.pinging)) {
671 		local64_inc(&uv_nmi_ping_misses);
672 		return NMI_DONE;
673 	}
674 
675 	uv_cpu_nmi.pings++;
676 	local64_inc(&uv_nmi_ping_count);
677 	ret = uv_handle_nmi(reason, regs);
678 	atomic_set(&uv_cpu_nmi.pinging, 0);
679 	return ret;
680 }
681 
682 static void uv_register_nmi_notifier(void)
683 {
684 	if (register_nmi_handler(NMI_UNKNOWN, uv_handle_nmi, 0, "uv"))
685 		pr_warn("UV: NMI handler failed to register\n");
686 
687 	if (register_nmi_handler(NMI_LOCAL, uv_handle_nmi_ping, 0, "uvping"))
688 		pr_warn("UV: PING NMI handler failed to register\n");
689 }
690 
691 void uv_nmi_init(void)
692 {
693 	unsigned int value;
694 
695 	/*
696 	 * Unmask NMI on all cpus
697 	 */
698 	value = apic_read(APIC_LVT1) | APIC_DM_NMI;
699 	value &= ~APIC_LVT_MASKED;
700 	apic_write(APIC_LVT1, value);
701 }
702 
703 void uv_nmi_setup(void)
704 {
705 	int size = sizeof(void *) * (1 << NODES_SHIFT);
706 	int cpu, nid;
707 
708 	/* Setup hub nmi info */
709 	uv_nmi_setup_mmrs();
710 	uv_hub_nmi_list = kzalloc(size, GFP_KERNEL);
711 	pr_info("UV: NMI hub list @ 0x%p (%d)\n", uv_hub_nmi_list, size);
712 	BUG_ON(!uv_hub_nmi_list);
713 	size = sizeof(struct uv_hub_nmi_s);
714 	for_each_present_cpu(cpu) {
715 		nid = cpu_to_node(cpu);
716 		if (uv_hub_nmi_list[nid] == NULL) {
717 			uv_hub_nmi_list[nid] = kzalloc_node(size,
718 							    GFP_KERNEL, nid);
719 			BUG_ON(!uv_hub_nmi_list[nid]);
720 			raw_spin_lock_init(&(uv_hub_nmi_list[nid]->nmi_lock));
721 			atomic_set(&uv_hub_nmi_list[nid]->cpu_owner, -1);
722 		}
723 		uv_hub_nmi_per(cpu) = uv_hub_nmi_list[nid];
724 	}
725 	BUG_ON(!alloc_cpumask_var(&uv_nmi_cpu_mask, GFP_KERNEL));
726 	uv_register_nmi_notifier();
727 }
728