xref: /openbmc/linux/arch/x86/platform/uv/uv_nmi.c (revision 9b68f30b)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * SGI NMI support routines
4  *
5  * (C) Copyright 2020 Hewlett Packard Enterprise Development LP
6  * Copyright (C) 2007-2017 Silicon Graphics, Inc. All rights reserved.
7  * Copyright (c) Mike Travis
8  */
9 
10 #include <linux/cpu.h>
11 #include <linux/delay.h>
12 #include <linux/kdb.h>
13 #include <linux/kexec.h>
14 #include <linux/kgdb.h>
15 #include <linux/moduleparam.h>
16 #include <linux/nmi.h>
17 #include <linux/sched.h>
18 #include <linux/sched/debug.h>
19 #include <linux/slab.h>
20 #include <linux/clocksource.h>
21 
22 #include <asm/apic.h>
23 #include <asm/current.h>
24 #include <asm/kdebug.h>
25 #include <asm/local64.h>
26 #include <asm/nmi.h>
27 #include <asm/reboot.h>
28 #include <asm/traps.h>
29 #include <asm/uv/uv.h>
30 #include <asm/uv/uv_hub.h>
31 #include <asm/uv/uv_mmrs.h>
32 
33 /*
34  * UV handler for NMI
35  *
36  * Handle system-wide NMI events generated by the global 'power nmi' command.
37  *
38  * Basic operation is to field the NMI interrupt on each CPU and wait
39  * until all CPU's have arrived into the nmi handler.  If some CPU's do not
40  * make it into the handler, try and force them in with the IPI(NMI) signal.
41  *
42  * We also have to lessen UV Hub MMR accesses as much as possible as this
43  * disrupts the UV Hub's primary mission of directing NumaLink traffic and
44  * can cause system problems to occur.
45  *
46  * To do this we register our primary NMI notifier on the NMI_UNKNOWN
47  * chain.  This reduces the number of false NMI calls when the perf
48  * tools are running which generate an enormous number of NMIs per
49  * second (~4M/s for 1024 CPU threads).  Our secondary NMI handler is
50  * very short as it only checks that if it has been "pinged" with the
51  * IPI(NMI) signal as mentioned above, and does not read the UV Hub's MMR.
52  *
53  */
54 
55 static struct uv_hub_nmi_s **uv_hub_nmi_list;
56 
57 DEFINE_PER_CPU(struct uv_cpu_nmi_s, uv_cpu_nmi);
58 
59 /* Newer SMM NMI handler, not present in all systems */
60 static unsigned long uvh_nmi_mmrx;		/* UVH_EVENT_OCCURRED0/1 */
61 static unsigned long uvh_nmi_mmrx_clear;	/* UVH_EVENT_OCCURRED0/1_ALIAS */
62 static int uvh_nmi_mmrx_shift;			/* UVH_EVENT_OCCURRED0/1_EXTIO_INT0_SHFT */
63 static char *uvh_nmi_mmrx_type;			/* "EXTIO_INT0" */
64 
65 /* Non-zero indicates newer SMM NMI handler present */
66 static unsigned long uvh_nmi_mmrx_supported;	/* UVH_EXTIO_INT0_BROADCAST */
67 
68 /* Indicates to BIOS that we want to use the newer SMM NMI handler */
69 static unsigned long uvh_nmi_mmrx_req;		/* UVH_BIOS_KERNEL_MMR_ALIAS_2 */
70 static int uvh_nmi_mmrx_req_shift;		/* 62 */
71 
72 /* UV hubless values */
73 #define NMI_CONTROL_PORT	0x70
74 #define NMI_DUMMY_PORT		0x71
75 #define PAD_OWN_GPP_D_0		0x2c
76 #define GPI_NMI_STS_GPP_D_0	0x164
77 #define GPI_NMI_ENA_GPP_D_0	0x174
78 #define STS_GPP_D_0_MASK	0x1
79 #define PAD_CFG_DW0_GPP_D_0	0x4c0
80 #define GPIROUTNMI		(1ul << 17)
81 #define PCH_PCR_GPIO_1_BASE	0xfdae0000ul
82 #define PCH_PCR_GPIO_ADDRESS(offset) (int *)((u64)(pch_base) | (u64)(offset))
83 
84 static u64 *pch_base;
85 static unsigned long nmi_mmr;
86 static unsigned long nmi_mmr_clear;
87 static unsigned long nmi_mmr_pending;
88 
89 static atomic_t	uv_in_nmi;
90 static atomic_t uv_nmi_cpu = ATOMIC_INIT(-1);
91 static atomic_t uv_nmi_cpus_in_nmi = ATOMIC_INIT(-1);
92 static atomic_t uv_nmi_slave_continue;
93 static cpumask_var_t uv_nmi_cpu_mask;
94 
95 static atomic_t uv_nmi_kexec_failed;
96 
97 /* Values for uv_nmi_slave_continue */
98 #define SLAVE_CLEAR	0
99 #define SLAVE_CONTINUE	1
100 #define SLAVE_EXIT	2
101 
102 /*
103  * Default is all stack dumps go to the console and buffer.
104  * Lower level to send to log buffer only.
105  */
106 static int uv_nmi_loglevel = CONSOLE_LOGLEVEL_DEFAULT;
107 module_param_named(dump_loglevel, uv_nmi_loglevel, int, 0644);
108 
109 /*
110  * The following values show statistics on how perf events are affecting
111  * this system.
112  */
113 static int param_get_local64(char *buffer, const struct kernel_param *kp)
114 {
115 	return sprintf(buffer, "%lu\n", local64_read((local64_t *)kp->arg));
116 }
117 
118 static int param_set_local64(const char *val, const struct kernel_param *kp)
119 {
120 	/* Clear on any write */
121 	local64_set((local64_t *)kp->arg, 0);
122 	return 0;
123 }
124 
125 static const struct kernel_param_ops param_ops_local64 = {
126 	.get = param_get_local64,
127 	.set = param_set_local64,
128 };
129 #define param_check_local64(name, p) __param_check(name, p, local64_t)
130 
131 static local64_t uv_nmi_count;
132 module_param_named(nmi_count, uv_nmi_count, local64, 0644);
133 
134 static local64_t uv_nmi_misses;
135 module_param_named(nmi_misses, uv_nmi_misses, local64, 0644);
136 
137 static local64_t uv_nmi_ping_count;
138 module_param_named(ping_count, uv_nmi_ping_count, local64, 0644);
139 
140 static local64_t uv_nmi_ping_misses;
141 module_param_named(ping_misses, uv_nmi_ping_misses, local64, 0644);
142 
143 /*
144  * Following values allow tuning for large systems under heavy loading
145  */
146 static int uv_nmi_initial_delay = 100;
147 module_param_named(initial_delay, uv_nmi_initial_delay, int, 0644);
148 
149 static int uv_nmi_slave_delay = 100;
150 module_param_named(slave_delay, uv_nmi_slave_delay, int, 0644);
151 
152 static int uv_nmi_loop_delay = 100;
153 module_param_named(loop_delay, uv_nmi_loop_delay, int, 0644);
154 
155 static int uv_nmi_trigger_delay = 10000;
156 module_param_named(trigger_delay, uv_nmi_trigger_delay, int, 0644);
157 
158 static int uv_nmi_wait_count = 100;
159 module_param_named(wait_count, uv_nmi_wait_count, int, 0644);
160 
161 static int uv_nmi_retry_count = 500;
162 module_param_named(retry_count, uv_nmi_retry_count, int, 0644);
163 
164 static bool uv_pch_intr_enable = true;
165 static bool uv_pch_intr_now_enabled;
166 module_param_named(pch_intr_enable, uv_pch_intr_enable, bool, 0644);
167 
168 static bool uv_pch_init_enable = true;
169 module_param_named(pch_init_enable, uv_pch_init_enable, bool, 0644);
170 
171 static int uv_nmi_debug;
172 module_param_named(debug, uv_nmi_debug, int, 0644);
173 
174 #define nmi_debug(fmt, ...)				\
175 	do {						\
176 		if (uv_nmi_debug)			\
177 			pr_info(fmt, ##__VA_ARGS__);	\
178 	} while (0)
179 
180 /* Valid NMI Actions */
181 #define	ACTION_LEN	16
182 static struct nmi_action {
183 	char	*action;
184 	char	*desc;
185 } valid_acts[] = {
186 	{	"kdump",	"do kernel crash dump"			},
187 	{	"dump",		"dump process stack for each cpu"	},
188 	{	"ips",		"dump Inst Ptr info for each cpu"	},
189 	{	"kdb",		"enter KDB (needs kgdboc= assignment)"	},
190 	{	"kgdb",		"enter KGDB (needs gdb target remote)"	},
191 	{	"health",	"check if CPUs respond to NMI"		},
192 };
193 typedef char action_t[ACTION_LEN];
194 static action_t uv_nmi_action = { "dump" };
195 
196 static int param_get_action(char *buffer, const struct kernel_param *kp)
197 {
198 	return sprintf(buffer, "%s\n", uv_nmi_action);
199 }
200 
201 static int param_set_action(const char *val, const struct kernel_param *kp)
202 {
203 	int i;
204 	int n = ARRAY_SIZE(valid_acts);
205 	char arg[ACTION_LEN], *p;
206 
207 	/* (remove possible '\n') */
208 	strncpy(arg, val, ACTION_LEN - 1);
209 	arg[ACTION_LEN - 1] = '\0';
210 	p = strchr(arg, '\n');
211 	if (p)
212 		*p = '\0';
213 
214 	for (i = 0; i < n; i++)
215 		if (!strcmp(arg, valid_acts[i].action))
216 			break;
217 
218 	if (i < n) {
219 		strcpy(uv_nmi_action, arg);
220 		pr_info("UV: New NMI action:%s\n", uv_nmi_action);
221 		return 0;
222 	}
223 
224 	pr_err("UV: Invalid NMI action:%s, valid actions are:\n", arg);
225 	for (i = 0; i < n; i++)
226 		pr_err("UV: %-8s - %s\n",
227 			valid_acts[i].action, valid_acts[i].desc);
228 	return -EINVAL;
229 }
230 
231 static const struct kernel_param_ops param_ops_action = {
232 	.get = param_get_action,
233 	.set = param_set_action,
234 };
235 #define param_check_action(name, p) __param_check(name, p, action_t)
236 
237 module_param_named(action, uv_nmi_action, action, 0644);
238 
239 static inline bool uv_nmi_action_is(const char *action)
240 {
241 	return (strncmp(uv_nmi_action, action, strlen(action)) == 0);
242 }
243 
244 /* Setup which NMI support is present in system */
245 static void uv_nmi_setup_mmrs(void)
246 {
247 	bool new_nmi_method_only = false;
248 
249 	/* First determine arch specific MMRs to handshake with BIOS */
250 	if (UVH_EVENT_OCCURRED0_EXTIO_INT0_MASK) {	/* UV2,3,4 setup */
251 		uvh_nmi_mmrx = UVH_EVENT_OCCURRED0;
252 		uvh_nmi_mmrx_clear = UVH_EVENT_OCCURRED0_ALIAS;
253 		uvh_nmi_mmrx_shift = UVH_EVENT_OCCURRED0_EXTIO_INT0_SHFT;
254 		uvh_nmi_mmrx_type = "OCRD0-EXTIO_INT0";
255 
256 		uvh_nmi_mmrx_supported = UVH_EXTIO_INT0_BROADCAST;
257 		uvh_nmi_mmrx_req = UVH_BIOS_KERNEL_MMR_ALIAS_2;
258 		uvh_nmi_mmrx_req_shift = 62;
259 
260 	} else if (UVH_EVENT_OCCURRED1_EXTIO_INT0_MASK) { /* UV5+ setup */
261 		uvh_nmi_mmrx = UVH_EVENT_OCCURRED1;
262 		uvh_nmi_mmrx_clear = UVH_EVENT_OCCURRED1_ALIAS;
263 		uvh_nmi_mmrx_shift = UVH_EVENT_OCCURRED1_EXTIO_INT0_SHFT;
264 		uvh_nmi_mmrx_type = "OCRD1-EXTIO_INT0";
265 
266 		new_nmi_method_only = true;		/* Newer nmi always valid on UV5+ */
267 		uvh_nmi_mmrx_req = 0;			/* no request bit to clear */
268 
269 	} else {
270 		pr_err("UV:%s:NMI support not available on this system\n", __func__);
271 		return;
272 	}
273 
274 	/* Then find out if new NMI is supported */
275 	if (new_nmi_method_only || uv_read_local_mmr(uvh_nmi_mmrx_supported)) {
276 		if (uvh_nmi_mmrx_req)
277 			uv_write_local_mmr(uvh_nmi_mmrx_req,
278 						1UL << uvh_nmi_mmrx_req_shift);
279 		nmi_mmr = uvh_nmi_mmrx;
280 		nmi_mmr_clear = uvh_nmi_mmrx_clear;
281 		nmi_mmr_pending = 1UL << uvh_nmi_mmrx_shift;
282 		pr_info("UV: SMI NMI support: %s\n", uvh_nmi_mmrx_type);
283 	} else {
284 		nmi_mmr = UVH_NMI_MMR;
285 		nmi_mmr_clear = UVH_NMI_MMR_CLEAR;
286 		nmi_mmr_pending = 1UL << UVH_NMI_MMR_SHIFT;
287 		pr_info("UV: SMI NMI support: %s\n", UVH_NMI_MMR_TYPE);
288 	}
289 }
290 
291 /* Read NMI MMR and check if NMI flag was set by BMC. */
292 static inline int uv_nmi_test_mmr(struct uv_hub_nmi_s *hub_nmi)
293 {
294 	hub_nmi->nmi_value = uv_read_local_mmr(nmi_mmr);
295 	atomic_inc(&hub_nmi->read_mmr_count);
296 	return !!(hub_nmi->nmi_value & nmi_mmr_pending);
297 }
298 
299 static inline void uv_local_mmr_clear_nmi(void)
300 {
301 	uv_write_local_mmr(nmi_mmr_clear, nmi_mmr_pending);
302 }
303 
304 /*
305  * UV hubless NMI handler functions
306  */
307 static inline void uv_reassert_nmi(void)
308 {
309 	/* (from arch/x86/include/asm/mach_traps.h) */
310 	outb(0x8f, NMI_CONTROL_PORT);
311 	inb(NMI_DUMMY_PORT);		/* dummy read */
312 	outb(0x0f, NMI_CONTROL_PORT);
313 	inb(NMI_DUMMY_PORT);		/* dummy read */
314 }
315 
316 static void uv_init_hubless_pch_io(int offset, int mask, int data)
317 {
318 	int *addr = PCH_PCR_GPIO_ADDRESS(offset);
319 	int readd = readl(addr);
320 
321 	if (mask) {			/* OR in new data */
322 		int writed = (readd & ~mask) | data;
323 
324 		nmi_debug("UV:PCH: %p = %x & %x | %x (%x)\n",
325 			addr, readd, ~mask, data, writed);
326 		writel(writed, addr);
327 	} else if (readd & data) {	/* clear status bit */
328 		nmi_debug("UV:PCH: %p = %x\n", addr, data);
329 		writel(data, addr);
330 	}
331 
332 	(void)readl(addr);		/* flush write data */
333 }
334 
335 static void uv_nmi_setup_hubless_intr(void)
336 {
337 	uv_pch_intr_now_enabled = uv_pch_intr_enable;
338 
339 	uv_init_hubless_pch_io(
340 		PAD_CFG_DW0_GPP_D_0, GPIROUTNMI,
341 		uv_pch_intr_now_enabled ? GPIROUTNMI : 0);
342 
343 	nmi_debug("UV:NMI: GPP_D_0 interrupt %s\n",
344 		uv_pch_intr_now_enabled ? "enabled" : "disabled");
345 }
346 
347 static struct init_nmi {
348 	unsigned int	offset;
349 	unsigned int	mask;
350 	unsigned int	data;
351 } init_nmi[] = {
352 	{	/* HOSTSW_OWN_GPP_D_0 */
353 	.offset = 0x84,
354 	.mask = 0x1,
355 	.data = 0x0,	/* ACPI Mode */
356 	},
357 
358 /* Clear status: */
359 	{	/* GPI_INT_STS_GPP_D_0 */
360 	.offset = 0x104,
361 	.mask = 0x0,
362 	.data = 0x1,	/* Clear Status */
363 	},
364 	{	/* GPI_GPE_STS_GPP_D_0 */
365 	.offset = 0x124,
366 	.mask = 0x0,
367 	.data = 0x1,	/* Clear Status */
368 	},
369 	{	/* GPI_SMI_STS_GPP_D_0 */
370 	.offset = 0x144,
371 	.mask = 0x0,
372 	.data = 0x1,	/* Clear Status */
373 	},
374 	{	/* GPI_NMI_STS_GPP_D_0 */
375 	.offset = 0x164,
376 	.mask = 0x0,
377 	.data = 0x1,	/* Clear Status */
378 	},
379 
380 /* Disable interrupts: */
381 	{	/* GPI_INT_EN_GPP_D_0 */
382 	.offset = 0x114,
383 	.mask = 0x1,
384 	.data = 0x0,	/* Disable interrupt generation */
385 	},
386 	{	/* GPI_GPE_EN_GPP_D_0 */
387 	.offset = 0x134,
388 	.mask = 0x1,
389 	.data = 0x0,	/* Disable interrupt generation */
390 	},
391 	{	/* GPI_SMI_EN_GPP_D_0 */
392 	.offset = 0x154,
393 	.mask = 0x1,
394 	.data = 0x0,	/* Disable interrupt generation */
395 	},
396 	{	/* GPI_NMI_EN_GPP_D_0 */
397 	.offset = 0x174,
398 	.mask = 0x1,
399 	.data = 0x0,	/* Disable interrupt generation */
400 	},
401 
402 /* Setup GPP_D_0 Pad Config: */
403 	{	/* PAD_CFG_DW0_GPP_D_0 */
404 	.offset = 0x4c0,
405 	.mask = 0xffffffff,
406 	.data = 0x82020100,
407 /*
408  *  31:30 Pad Reset Config (PADRSTCFG): = 2h  # PLTRST# (default)
409  *
410  *  29    RX Pad State Select (RXPADSTSEL): = 0 # Raw RX pad state directly
411  *                                                from RX buffer (default)
412  *
413  *  28    RX Raw Override to '1' (RXRAW1): = 0 # No Override
414  *
415  *  26:25 RX Level/Edge Configuration (RXEVCFG):
416  *      = 0h # Level
417  *      = 1h # Edge
418  *
419  *  23    RX Invert (RXINV): = 0 # No Inversion (signal active high)
420  *
421  *  20    GPIO Input Route IOxAPIC (GPIROUTIOXAPIC):
422  * = 0 # Routing does not cause peripheral IRQ...
423  *     # (we want an NMI not an IRQ)
424  *
425  *  19    GPIO Input Route SCI (GPIROUTSCI): = 0 # Routing does not cause SCI.
426  *  18    GPIO Input Route SMI (GPIROUTSMI): = 0 # Routing does not cause SMI.
427  *  17    GPIO Input Route NMI (GPIROUTNMI): = 1 # Routing can cause NMI.
428  *
429  *  11:10 Pad Mode (PMODE1/0): = 0h = GPIO control the Pad.
430  *   9    GPIO RX Disable (GPIORXDIS):
431  * = 0 # Enable the input buffer (active low enable)
432  *
433  *   8    GPIO TX Disable (GPIOTXDIS):
434  * = 1 # Disable the output buffer; i.e. Hi-Z
435  *
436  *   1 GPIO RX State (GPIORXSTATE): This is the current internal RX pad state..
437  *   0 GPIO TX State (GPIOTXSTATE):
438  * = 0 # (Leave at default)
439  */
440 	},
441 
442 /* Pad Config DW1 */
443 	{	/* PAD_CFG_DW1_GPP_D_0 */
444 	.offset = 0x4c4,
445 	.mask = 0x3c00,
446 	.data = 0,	/* Termination = none (default) */
447 	},
448 };
449 
450 static void uv_init_hubless_pch_d0(void)
451 {
452 	int i, read;
453 
454 	read = *PCH_PCR_GPIO_ADDRESS(PAD_OWN_GPP_D_0);
455 	if (read != 0) {
456 		pr_info("UV: Hubless NMI already configured\n");
457 		return;
458 	}
459 
460 	nmi_debug("UV: Initializing UV Hubless NMI on PCH\n");
461 	for (i = 0; i < ARRAY_SIZE(init_nmi); i++) {
462 		uv_init_hubless_pch_io(init_nmi[i].offset,
463 					init_nmi[i].mask,
464 					init_nmi[i].data);
465 	}
466 }
467 
468 static int uv_nmi_test_hubless(struct uv_hub_nmi_s *hub_nmi)
469 {
470 	int *pstat = PCH_PCR_GPIO_ADDRESS(GPI_NMI_STS_GPP_D_0);
471 	int status = *pstat;
472 
473 	hub_nmi->nmi_value = status;
474 	atomic_inc(&hub_nmi->read_mmr_count);
475 
476 	if (!(status & STS_GPP_D_0_MASK))	/* Not a UV external NMI */
477 		return 0;
478 
479 	*pstat = STS_GPP_D_0_MASK;	/* Is a UV NMI: clear GPP_D_0 status */
480 	(void)*pstat;			/* Flush write */
481 
482 	return 1;
483 }
484 
485 static int uv_test_nmi(struct uv_hub_nmi_s *hub_nmi)
486 {
487 	if (hub_nmi->hub_present)
488 		return uv_nmi_test_mmr(hub_nmi);
489 
490 	if (hub_nmi->pch_owner)		/* Only PCH owner can check status */
491 		return uv_nmi_test_hubless(hub_nmi);
492 
493 	return -1;
494 }
495 
496 /*
497  * If first CPU in on this hub, set hub_nmi "in_nmi" and "owner" values and
498  * return true.  If first CPU in on the system, set global "in_nmi" flag.
499  */
500 static int uv_set_in_nmi(int cpu, struct uv_hub_nmi_s *hub_nmi)
501 {
502 	int first = atomic_add_unless(&hub_nmi->in_nmi, 1, 1);
503 
504 	if (first) {
505 		atomic_set(&hub_nmi->cpu_owner, cpu);
506 		if (atomic_add_unless(&uv_in_nmi, 1, 1))
507 			atomic_set(&uv_nmi_cpu, cpu);
508 
509 		atomic_inc(&hub_nmi->nmi_count);
510 	}
511 	return first;
512 }
513 
514 /* Check if this is a system NMI event */
515 static int uv_check_nmi(struct uv_hub_nmi_s *hub_nmi)
516 {
517 	int cpu = smp_processor_id();
518 	int nmi = 0;
519 	int nmi_detected = 0;
520 
521 	local64_inc(&uv_nmi_count);
522 	this_cpu_inc(uv_cpu_nmi.queries);
523 
524 	do {
525 		nmi = atomic_read(&hub_nmi->in_nmi);
526 		if (nmi)
527 			break;
528 
529 		if (raw_spin_trylock(&hub_nmi->nmi_lock)) {
530 			nmi_detected = uv_test_nmi(hub_nmi);
531 
532 			/* Check flag for UV external NMI */
533 			if (nmi_detected > 0) {
534 				uv_set_in_nmi(cpu, hub_nmi);
535 				nmi = 1;
536 				break;
537 			}
538 
539 			/* A non-PCH node in a hubless system waits for NMI */
540 			else if (nmi_detected < 0)
541 				goto slave_wait;
542 
543 			/* MMR/PCH NMI flag is clear */
544 			raw_spin_unlock(&hub_nmi->nmi_lock);
545 
546 		} else {
547 
548 			/* Wait a moment for the HUB NMI locker to set flag */
549 slave_wait:		cpu_relax();
550 			udelay(uv_nmi_slave_delay);
551 
552 			/* Re-check hub in_nmi flag */
553 			nmi = atomic_read(&hub_nmi->in_nmi);
554 			if (nmi)
555 				break;
556 		}
557 
558 		/*
559 		 * Check if this BMC missed setting the MMR NMI flag (or)
560 		 * UV hubless system where only PCH owner can check flag
561 		 */
562 		if (!nmi) {
563 			nmi = atomic_read(&uv_in_nmi);
564 			if (nmi)
565 				uv_set_in_nmi(cpu, hub_nmi);
566 		}
567 
568 		/* If we're holding the hub lock, release it now */
569 		if (nmi_detected < 0)
570 			raw_spin_unlock(&hub_nmi->nmi_lock);
571 
572 	} while (0);
573 
574 	if (!nmi)
575 		local64_inc(&uv_nmi_misses);
576 
577 	return nmi;
578 }
579 
580 /* Need to reset the NMI MMR register, but only once per hub. */
581 static inline void uv_clear_nmi(int cpu)
582 {
583 	struct uv_hub_nmi_s *hub_nmi = uv_hub_nmi;
584 
585 	if (cpu == atomic_read(&hub_nmi->cpu_owner)) {
586 		atomic_set(&hub_nmi->cpu_owner, -1);
587 		atomic_set(&hub_nmi->in_nmi, 0);
588 		if (hub_nmi->hub_present)
589 			uv_local_mmr_clear_nmi();
590 		else
591 			uv_reassert_nmi();
592 		raw_spin_unlock(&hub_nmi->nmi_lock);
593 	}
594 }
595 
596 /* Ping non-responding CPU's attempting to force them into the NMI handler */
597 static void uv_nmi_nr_cpus_ping(void)
598 {
599 	int cpu;
600 
601 	for_each_cpu(cpu, uv_nmi_cpu_mask)
602 		uv_cpu_nmi_per(cpu).pinging = 1;
603 
604 	apic->send_IPI_mask(uv_nmi_cpu_mask, APIC_DM_NMI);
605 }
606 
607 /* Clean up flags for CPU's that ignored both NMI and ping */
608 static void uv_nmi_cleanup_mask(void)
609 {
610 	int cpu;
611 
612 	for_each_cpu(cpu, uv_nmi_cpu_mask) {
613 		uv_cpu_nmi_per(cpu).pinging =  0;
614 		uv_cpu_nmi_per(cpu).state = UV_NMI_STATE_OUT;
615 		cpumask_clear_cpu(cpu, uv_nmi_cpu_mask);
616 	}
617 }
618 
619 /* Loop waiting as CPU's enter NMI handler */
620 static int uv_nmi_wait_cpus(int first)
621 {
622 	int i, j, k, n = num_online_cpus();
623 	int last_k = 0, waiting = 0;
624 	int cpu = smp_processor_id();
625 
626 	if (first) {
627 		cpumask_copy(uv_nmi_cpu_mask, cpu_online_mask);
628 		k = 0;
629 	} else {
630 		k = n - cpumask_weight(uv_nmi_cpu_mask);
631 	}
632 
633 	/* PCH NMI causes only one CPU to respond */
634 	if (first && uv_pch_intr_now_enabled) {
635 		cpumask_clear_cpu(cpu, uv_nmi_cpu_mask);
636 		return n - k - 1;
637 	}
638 
639 	udelay(uv_nmi_initial_delay);
640 	for (i = 0; i < uv_nmi_retry_count; i++) {
641 		int loop_delay = uv_nmi_loop_delay;
642 
643 		for_each_cpu(j, uv_nmi_cpu_mask) {
644 			if (uv_cpu_nmi_per(j).state) {
645 				cpumask_clear_cpu(j, uv_nmi_cpu_mask);
646 				if (++k >= n)
647 					break;
648 			}
649 		}
650 		if (k >= n) {		/* all in? */
651 			k = n;
652 			break;
653 		}
654 		if (last_k != k) {	/* abort if no new CPU's coming in */
655 			last_k = k;
656 			waiting = 0;
657 		} else if (++waiting > uv_nmi_wait_count)
658 			break;
659 
660 		/* Extend delay if waiting only for CPU 0: */
661 		if (waiting && (n - k) == 1 &&
662 		    cpumask_test_cpu(0, uv_nmi_cpu_mask))
663 			loop_delay *= 100;
664 
665 		udelay(loop_delay);
666 	}
667 	atomic_set(&uv_nmi_cpus_in_nmi, k);
668 	return n - k;
669 }
670 
671 /* Wait until all slave CPU's have entered UV NMI handler */
672 static void uv_nmi_wait(int master)
673 {
674 	/* Indicate this CPU is in: */
675 	this_cpu_write(uv_cpu_nmi.state, UV_NMI_STATE_IN);
676 
677 	/* If not the first CPU in (the master), then we are a slave CPU */
678 	if (!master)
679 		return;
680 
681 	do {
682 		/* Wait for all other CPU's to gather here */
683 		if (!uv_nmi_wait_cpus(1))
684 			break;
685 
686 		/* If not all made it in, send IPI NMI to them */
687 		pr_alert("UV: Sending NMI IPI to %d CPUs: %*pbl\n",
688 			 cpumask_weight(uv_nmi_cpu_mask),
689 			 cpumask_pr_args(uv_nmi_cpu_mask));
690 
691 		uv_nmi_nr_cpus_ping();
692 
693 		/* If all CPU's are in, then done */
694 		if (!uv_nmi_wait_cpus(0))
695 			break;
696 
697 		pr_alert("UV: %d CPUs not in NMI loop: %*pbl\n",
698 			 cpumask_weight(uv_nmi_cpu_mask),
699 			 cpumask_pr_args(uv_nmi_cpu_mask));
700 	} while (0);
701 
702 	pr_alert("UV: %d of %d CPUs in NMI\n",
703 		atomic_read(&uv_nmi_cpus_in_nmi), num_online_cpus());
704 }
705 
706 /* Dump Instruction Pointer header */
707 static void uv_nmi_dump_cpu_ip_hdr(void)
708 {
709 	pr_info("\nUV: %4s %6s %-32s %s   (Note: PID 0 not listed)\n",
710 		"CPU", "PID", "COMMAND", "IP");
711 }
712 
713 /* Dump Instruction Pointer info */
714 static void uv_nmi_dump_cpu_ip(int cpu, struct pt_regs *regs)
715 {
716 	pr_info("UV: %4d %6d %-32.32s %pS",
717 		cpu, current->pid, current->comm, (void *)regs->ip);
718 }
719 
720 /*
721  * Dump this CPU's state.  If action was set to "kdump" and the crash_kexec
722  * failed, then we provide "dump" as an alternate action.  Action "dump" now
723  * also includes the show "ips" (instruction pointers) action whereas the
724  * action "ips" only displays instruction pointers for the non-idle CPU's.
725  * This is an abbreviated form of the "ps" command.
726  */
727 static void uv_nmi_dump_state_cpu(int cpu, struct pt_regs *regs)
728 {
729 	const char *dots = " ................................. ";
730 
731 	if (cpu == 0)
732 		uv_nmi_dump_cpu_ip_hdr();
733 
734 	if (current->pid != 0 || !uv_nmi_action_is("ips"))
735 		uv_nmi_dump_cpu_ip(cpu, regs);
736 
737 	if (uv_nmi_action_is("dump")) {
738 		pr_info("UV:%sNMI process trace for CPU %d\n", dots, cpu);
739 		show_regs(regs);
740 	}
741 
742 	this_cpu_write(uv_cpu_nmi.state, UV_NMI_STATE_DUMP_DONE);
743 }
744 
745 /* Trigger a slave CPU to dump it's state */
746 static void uv_nmi_trigger_dump(int cpu)
747 {
748 	int retry = uv_nmi_trigger_delay;
749 
750 	if (uv_cpu_nmi_per(cpu).state != UV_NMI_STATE_IN)
751 		return;
752 
753 	uv_cpu_nmi_per(cpu).state = UV_NMI_STATE_DUMP;
754 	do {
755 		cpu_relax();
756 		udelay(10);
757 		if (uv_cpu_nmi_per(cpu).state
758 				!= UV_NMI_STATE_DUMP)
759 			return;
760 	} while (--retry > 0);
761 
762 	pr_crit("UV: CPU %d stuck in process dump function\n", cpu);
763 	uv_cpu_nmi_per(cpu).state = UV_NMI_STATE_DUMP_DONE;
764 }
765 
766 /* Wait until all CPU's ready to exit */
767 static void uv_nmi_sync_exit(int master)
768 {
769 	atomic_dec(&uv_nmi_cpus_in_nmi);
770 	if (master) {
771 		while (atomic_read(&uv_nmi_cpus_in_nmi) > 0)
772 			cpu_relax();
773 		atomic_set(&uv_nmi_slave_continue, SLAVE_CLEAR);
774 	} else {
775 		while (atomic_read(&uv_nmi_slave_continue))
776 			cpu_relax();
777 	}
778 }
779 
780 /* Current "health" check is to check which CPU's are responsive */
781 static void uv_nmi_action_health(int cpu, struct pt_regs *regs, int master)
782 {
783 	if (master) {
784 		int in = atomic_read(&uv_nmi_cpus_in_nmi);
785 		int out = num_online_cpus() - in;
786 
787 		pr_alert("UV: NMI CPU health check (non-responding:%d)\n", out);
788 		atomic_set(&uv_nmi_slave_continue, SLAVE_EXIT);
789 	} else {
790 		while (!atomic_read(&uv_nmi_slave_continue))
791 			cpu_relax();
792 	}
793 	uv_nmi_sync_exit(master);
794 }
795 
796 /* Walk through CPU list and dump state of each */
797 static void uv_nmi_dump_state(int cpu, struct pt_regs *regs, int master)
798 {
799 	if (master) {
800 		int tcpu;
801 		int ignored = 0;
802 		int saved_console_loglevel = console_loglevel;
803 
804 		pr_alert("UV: tracing %s for %d CPUs from CPU %d\n",
805 			uv_nmi_action_is("ips") ? "IPs" : "processes",
806 			atomic_read(&uv_nmi_cpus_in_nmi), cpu);
807 
808 		console_loglevel = uv_nmi_loglevel;
809 		atomic_set(&uv_nmi_slave_continue, SLAVE_EXIT);
810 		for_each_online_cpu(tcpu) {
811 			if (cpumask_test_cpu(tcpu, uv_nmi_cpu_mask))
812 				ignored++;
813 			else if (tcpu == cpu)
814 				uv_nmi_dump_state_cpu(tcpu, regs);
815 			else
816 				uv_nmi_trigger_dump(tcpu);
817 		}
818 		if (ignored)
819 			pr_alert("UV: %d CPUs ignored NMI\n", ignored);
820 
821 		console_loglevel = saved_console_loglevel;
822 		pr_alert("UV: process trace complete\n");
823 	} else {
824 		while (!atomic_read(&uv_nmi_slave_continue))
825 			cpu_relax();
826 		while (this_cpu_read(uv_cpu_nmi.state) != UV_NMI_STATE_DUMP)
827 			cpu_relax();
828 		uv_nmi_dump_state_cpu(cpu, regs);
829 	}
830 	uv_nmi_sync_exit(master);
831 }
832 
833 static void uv_nmi_touch_watchdogs(void)
834 {
835 	touch_softlockup_watchdog_sync();
836 	clocksource_touch_watchdog();
837 	rcu_cpu_stall_reset();
838 	touch_nmi_watchdog();
839 }
840 
841 static void uv_nmi_kdump(int cpu, int main, struct pt_regs *regs)
842 {
843 	/* Check if kdump kernel loaded for both main and secondary CPUs */
844 	if (!kexec_crash_image) {
845 		if (main)
846 			pr_err("UV: NMI error: kdump kernel not loaded\n");
847 		return;
848 	}
849 
850 	/* Call crash to dump system state */
851 	if (main) {
852 		pr_emerg("UV: NMI executing crash_kexec on CPU%d\n", cpu);
853 		crash_kexec(regs);
854 
855 		pr_emerg("UV: crash_kexec unexpectedly returned\n");
856 		atomic_set(&uv_nmi_kexec_failed, 1);
857 
858 	} else { /* secondary */
859 
860 		/* If kdump kernel fails, secondaries will exit this loop */
861 		while (atomic_read(&uv_nmi_kexec_failed) == 0) {
862 
863 			/* Once shootdown cpus starts, they do not return */
864 			run_crash_ipi_callback(regs);
865 
866 			mdelay(10);
867 		}
868 	}
869 }
870 
871 #ifdef CONFIG_KGDB
872 #ifdef CONFIG_KGDB_KDB
873 static inline int uv_nmi_kdb_reason(void)
874 {
875 	return KDB_REASON_SYSTEM_NMI;
876 }
877 #else /* !CONFIG_KGDB_KDB */
878 static inline int uv_nmi_kdb_reason(void)
879 {
880 	/* Ensure user is expecting to attach gdb remote */
881 	if (uv_nmi_action_is("kgdb"))
882 		return 0;
883 
884 	pr_err("UV: NMI error: KDB is not enabled in this kernel\n");
885 	return -1;
886 }
887 #endif /* CONFIG_KGDB_KDB */
888 
889 /*
890  * Call KGDB/KDB from NMI handler
891  *
892  * Note that if both KGDB and KDB are configured, then the action of 'kgdb' or
893  * 'kdb' has no affect on which is used.  See the KGDB documentation for further
894  * information.
895  */
896 static void uv_call_kgdb_kdb(int cpu, struct pt_regs *regs, int master)
897 {
898 	if (master) {
899 		int reason = uv_nmi_kdb_reason();
900 		int ret;
901 
902 		if (reason < 0)
903 			return;
904 
905 		/* Call KGDB NMI handler as MASTER */
906 		ret = kgdb_nmicallin(cpu, X86_TRAP_NMI, regs, reason,
907 				&uv_nmi_slave_continue);
908 		if (ret) {
909 			pr_alert("KGDB returned error, is kgdboc set?\n");
910 			atomic_set(&uv_nmi_slave_continue, SLAVE_EXIT);
911 		}
912 	} else {
913 		/* Wait for KGDB signal that it's ready for slaves to enter */
914 		int sig;
915 
916 		do {
917 			cpu_relax();
918 			sig = atomic_read(&uv_nmi_slave_continue);
919 		} while (!sig);
920 
921 		/* Call KGDB as slave */
922 		if (sig == SLAVE_CONTINUE)
923 			kgdb_nmicallback(cpu, regs);
924 	}
925 	uv_nmi_sync_exit(master);
926 }
927 
928 #else /* !CONFIG_KGDB */
929 static inline void uv_call_kgdb_kdb(int cpu, struct pt_regs *regs, int master)
930 {
931 	pr_err("UV: NMI error: KGDB is not enabled in this kernel\n");
932 }
933 #endif /* !CONFIG_KGDB */
934 
935 /*
936  * UV NMI handler
937  */
938 static int uv_handle_nmi(unsigned int reason, struct pt_regs *regs)
939 {
940 	struct uv_hub_nmi_s *hub_nmi = uv_hub_nmi;
941 	int cpu = smp_processor_id();
942 	int master = 0;
943 	unsigned long flags;
944 
945 	local_irq_save(flags);
946 
947 	/* If not a UV System NMI, ignore */
948 	if (!this_cpu_read(uv_cpu_nmi.pinging) && !uv_check_nmi(hub_nmi)) {
949 		local_irq_restore(flags);
950 		return NMI_DONE;
951 	}
952 
953 	/* Indicate we are the first CPU into the NMI handler */
954 	master = (atomic_read(&uv_nmi_cpu) == cpu);
955 
956 	/* If NMI action is "kdump", then attempt to do it */
957 	if (uv_nmi_action_is("kdump")) {
958 		uv_nmi_kdump(cpu, master, regs);
959 
960 		/* Unexpected return, revert action to "dump" */
961 		if (master)
962 			strncpy(uv_nmi_action, "dump", strlen(uv_nmi_action));
963 	}
964 
965 	/* Pause as all CPU's enter the NMI handler */
966 	uv_nmi_wait(master);
967 
968 	/* Process actions other than "kdump": */
969 	if (uv_nmi_action_is("health")) {
970 		uv_nmi_action_health(cpu, regs, master);
971 	} else if (uv_nmi_action_is("ips") || uv_nmi_action_is("dump")) {
972 		uv_nmi_dump_state(cpu, regs, master);
973 	} else if (uv_nmi_action_is("kdb") || uv_nmi_action_is("kgdb")) {
974 		uv_call_kgdb_kdb(cpu, regs, master);
975 	} else {
976 		if (master)
977 			pr_alert("UV: unknown NMI action: %s\n", uv_nmi_action);
978 		uv_nmi_sync_exit(master);
979 	}
980 
981 	/* Clear per_cpu "in_nmi" flag */
982 	this_cpu_write(uv_cpu_nmi.state, UV_NMI_STATE_OUT);
983 
984 	/* Clear MMR NMI flag on each hub */
985 	uv_clear_nmi(cpu);
986 
987 	/* Clear global flags */
988 	if (master) {
989 		if (!cpumask_empty(uv_nmi_cpu_mask))
990 			uv_nmi_cleanup_mask();
991 		atomic_set(&uv_nmi_cpus_in_nmi, -1);
992 		atomic_set(&uv_nmi_cpu, -1);
993 		atomic_set(&uv_in_nmi, 0);
994 		atomic_set(&uv_nmi_kexec_failed, 0);
995 		atomic_set(&uv_nmi_slave_continue, SLAVE_CLEAR);
996 	}
997 
998 	uv_nmi_touch_watchdogs();
999 	local_irq_restore(flags);
1000 
1001 	return NMI_HANDLED;
1002 }
1003 
1004 /*
1005  * NMI handler for pulling in CPU's when perf events are grabbing our NMI
1006  */
1007 static int uv_handle_nmi_ping(unsigned int reason, struct pt_regs *regs)
1008 {
1009 	int ret;
1010 
1011 	this_cpu_inc(uv_cpu_nmi.queries);
1012 	if (!this_cpu_read(uv_cpu_nmi.pinging)) {
1013 		local64_inc(&uv_nmi_ping_misses);
1014 		return NMI_DONE;
1015 	}
1016 
1017 	this_cpu_inc(uv_cpu_nmi.pings);
1018 	local64_inc(&uv_nmi_ping_count);
1019 	ret = uv_handle_nmi(reason, regs);
1020 	this_cpu_write(uv_cpu_nmi.pinging, 0);
1021 	return ret;
1022 }
1023 
1024 static void uv_register_nmi_notifier(void)
1025 {
1026 	if (register_nmi_handler(NMI_UNKNOWN, uv_handle_nmi, 0, "uv"))
1027 		pr_warn("UV: NMI handler failed to register\n");
1028 
1029 	if (register_nmi_handler(NMI_LOCAL, uv_handle_nmi_ping, 0, "uvping"))
1030 		pr_warn("UV: PING NMI handler failed to register\n");
1031 }
1032 
1033 void uv_nmi_init(void)
1034 {
1035 	unsigned int value;
1036 
1037 	/*
1038 	 * Unmask NMI on all CPU's
1039 	 */
1040 	value = apic_read(APIC_LVT1) | APIC_DM_NMI;
1041 	value &= ~APIC_LVT_MASKED;
1042 	apic_write(APIC_LVT1, value);
1043 }
1044 
1045 /* Setup HUB NMI info */
1046 static void __init uv_nmi_setup_common(bool hubbed)
1047 {
1048 	int size = sizeof(void *) * (1 << NODES_SHIFT);
1049 	int cpu;
1050 
1051 	uv_hub_nmi_list = kzalloc(size, GFP_KERNEL);
1052 	nmi_debug("UV: NMI hub list @ 0x%p (%d)\n", uv_hub_nmi_list, size);
1053 	BUG_ON(!uv_hub_nmi_list);
1054 	size = sizeof(struct uv_hub_nmi_s);
1055 	for_each_present_cpu(cpu) {
1056 		int nid = cpu_to_node(cpu);
1057 		if (uv_hub_nmi_list[nid] == NULL) {
1058 			uv_hub_nmi_list[nid] = kzalloc_node(size,
1059 							    GFP_KERNEL, nid);
1060 			BUG_ON(!uv_hub_nmi_list[nid]);
1061 			raw_spin_lock_init(&(uv_hub_nmi_list[nid]->nmi_lock));
1062 			atomic_set(&uv_hub_nmi_list[nid]->cpu_owner, -1);
1063 			uv_hub_nmi_list[nid]->hub_present = hubbed;
1064 			uv_hub_nmi_list[nid]->pch_owner = (nid == 0);
1065 		}
1066 		uv_hub_nmi_per(cpu) = uv_hub_nmi_list[nid];
1067 	}
1068 	BUG_ON(!alloc_cpumask_var(&uv_nmi_cpu_mask, GFP_KERNEL));
1069 }
1070 
1071 /* Setup for UV Hub systems */
1072 void __init uv_nmi_setup(void)
1073 {
1074 	uv_nmi_setup_mmrs();
1075 	uv_nmi_setup_common(true);
1076 	uv_register_nmi_notifier();
1077 	pr_info("UV: Hub NMI enabled\n");
1078 }
1079 
1080 /* Setup for UV Hubless systems */
1081 void __init uv_nmi_setup_hubless(void)
1082 {
1083 	uv_nmi_setup_common(false);
1084 	pch_base = xlate_dev_mem_ptr(PCH_PCR_GPIO_1_BASE);
1085 	nmi_debug("UV: PCH base:%p from 0x%lx, GPP_D_0\n",
1086 		pch_base, PCH_PCR_GPIO_1_BASE);
1087 	if (uv_pch_init_enable)
1088 		uv_init_hubless_pch_d0();
1089 	uv_init_hubless_pch_io(GPI_NMI_ENA_GPP_D_0,
1090 				STS_GPP_D_0_MASK, STS_GPP_D_0_MASK);
1091 	uv_nmi_setup_hubless_intr();
1092 	/* Ensure NMI enabled in Processor Interface Reg: */
1093 	uv_reassert_nmi();
1094 	uv_register_nmi_notifier();
1095 	pr_info("UV: PCH NMI enabled\n");
1096 }
1097