xref: /openbmc/linux/arch/x86/platform/uv/uv_nmi.c (revision 2cf1c348)
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 	/* First determine arch specific MMRs to handshake with BIOS */
248 	if (UVH_EVENT_OCCURRED0_EXTIO_INT0_MASK) {
249 		uvh_nmi_mmrx = UVH_EVENT_OCCURRED0;
250 		uvh_nmi_mmrx_clear = UVH_EVENT_OCCURRED0_ALIAS;
251 		uvh_nmi_mmrx_shift = UVH_EVENT_OCCURRED0_EXTIO_INT0_SHFT;
252 		uvh_nmi_mmrx_type = "OCRD0-EXTIO_INT0";
253 
254 		uvh_nmi_mmrx_supported = UVH_EXTIO_INT0_BROADCAST;
255 		uvh_nmi_mmrx_req = UVH_BIOS_KERNEL_MMR_ALIAS_2;
256 		uvh_nmi_mmrx_req_shift = 62;
257 
258 	} else if (UVH_EVENT_OCCURRED1_EXTIO_INT0_MASK) {
259 		uvh_nmi_mmrx = UVH_EVENT_OCCURRED1;
260 		uvh_nmi_mmrx_clear = UVH_EVENT_OCCURRED1_ALIAS;
261 		uvh_nmi_mmrx_shift = UVH_EVENT_OCCURRED1_EXTIO_INT0_SHFT;
262 		uvh_nmi_mmrx_type = "OCRD1-EXTIO_INT0";
263 
264 		uvh_nmi_mmrx_supported = UVH_EXTIO_INT0_BROADCAST;
265 		uvh_nmi_mmrx_req = UVH_BIOS_KERNEL_MMR_ALIAS_2;
266 		uvh_nmi_mmrx_req_shift = 62;
267 
268 	} else {
269 		pr_err("UV:%s:cannot find EVENT_OCCURRED*_EXTIO_INT0\n",
270 			__func__);
271 		return;
272 	}
273 
274 	/* Then find out if new NMI is supported */
275 	if (likely(uv_read_local_mmr(uvh_nmi_mmrx_supported))) {
276 		uv_write_local_mmr(uvh_nmi_mmrx_req,
277 					1UL << uvh_nmi_mmrx_req_shift);
278 		nmi_mmr = uvh_nmi_mmrx;
279 		nmi_mmr_clear = uvh_nmi_mmrx_clear;
280 		nmi_mmr_pending = 1UL << uvh_nmi_mmrx_shift;
281 		pr_info("UV: SMI NMI support: %s\n", uvh_nmi_mmrx_type);
282 	} else {
283 		nmi_mmr = UVH_NMI_MMR;
284 		nmi_mmr_clear = UVH_NMI_MMR_CLEAR;
285 		nmi_mmr_pending = 1UL << UVH_NMI_MMR_SHIFT;
286 		pr_info("UV: SMI NMI support: %s\n", UVH_NMI_MMR_TYPE);
287 	}
288 }
289 
290 /* Read NMI MMR and check if NMI flag was set by BMC. */
291 static inline int uv_nmi_test_mmr(struct uv_hub_nmi_s *hub_nmi)
292 {
293 	hub_nmi->nmi_value = uv_read_local_mmr(nmi_mmr);
294 	atomic_inc(&hub_nmi->read_mmr_count);
295 	return !!(hub_nmi->nmi_value & nmi_mmr_pending);
296 }
297 
298 static inline void uv_local_mmr_clear_nmi(void)
299 {
300 	uv_write_local_mmr(nmi_mmr_clear, nmi_mmr_pending);
301 }
302 
303 /*
304  * UV hubless NMI handler functions
305  */
306 static inline void uv_reassert_nmi(void)
307 {
308 	/* (from arch/x86/include/asm/mach_traps.h) */
309 	outb(0x8f, NMI_CONTROL_PORT);
310 	inb(NMI_DUMMY_PORT);		/* dummy read */
311 	outb(0x0f, NMI_CONTROL_PORT);
312 	inb(NMI_DUMMY_PORT);		/* dummy read */
313 }
314 
315 static void uv_init_hubless_pch_io(int offset, int mask, int data)
316 {
317 	int *addr = PCH_PCR_GPIO_ADDRESS(offset);
318 	int readd = readl(addr);
319 
320 	if (mask) {			/* OR in new data */
321 		int writed = (readd & ~mask) | data;
322 
323 		nmi_debug("UV:PCH: %p = %x & %x | %x (%x)\n",
324 			addr, readd, ~mask, data, writed);
325 		writel(writed, addr);
326 	} else if (readd & data) {	/* clear status bit */
327 		nmi_debug("UV:PCH: %p = %x\n", addr, data);
328 		writel(data, addr);
329 	}
330 
331 	(void)readl(addr);		/* flush write data */
332 }
333 
334 static void uv_nmi_setup_hubless_intr(void)
335 {
336 	uv_pch_intr_now_enabled = uv_pch_intr_enable;
337 
338 	uv_init_hubless_pch_io(
339 		PAD_CFG_DW0_GPP_D_0, GPIROUTNMI,
340 		uv_pch_intr_now_enabled ? GPIROUTNMI : 0);
341 
342 	nmi_debug("UV:NMI: GPP_D_0 interrupt %s\n",
343 		uv_pch_intr_now_enabled ? "enabled" : "disabled");
344 }
345 
346 static struct init_nmi {
347 	unsigned int	offset;
348 	unsigned int	mask;
349 	unsigned int	data;
350 } init_nmi[] = {
351 	{	/* HOSTSW_OWN_GPP_D_0 */
352 	.offset = 0x84,
353 	.mask = 0x1,
354 	.data = 0x0,	/* ACPI Mode */
355 	},
356 
357 /* Clear status: */
358 	{	/* GPI_INT_STS_GPP_D_0 */
359 	.offset = 0x104,
360 	.mask = 0x0,
361 	.data = 0x1,	/* Clear Status */
362 	},
363 	{	/* GPI_GPE_STS_GPP_D_0 */
364 	.offset = 0x124,
365 	.mask = 0x0,
366 	.data = 0x1,	/* Clear Status */
367 	},
368 	{	/* GPI_SMI_STS_GPP_D_0 */
369 	.offset = 0x144,
370 	.mask = 0x0,
371 	.data = 0x1,	/* Clear Status */
372 	},
373 	{	/* GPI_NMI_STS_GPP_D_0 */
374 	.offset = 0x164,
375 	.mask = 0x0,
376 	.data = 0x1,	/* Clear Status */
377 	},
378 
379 /* Disable interrupts: */
380 	{	/* GPI_INT_EN_GPP_D_0 */
381 	.offset = 0x114,
382 	.mask = 0x1,
383 	.data = 0x0,	/* Disable interrupt generation */
384 	},
385 	{	/* GPI_GPE_EN_GPP_D_0 */
386 	.offset = 0x134,
387 	.mask = 0x1,
388 	.data = 0x0,	/* Disable interrupt generation */
389 	},
390 	{	/* GPI_SMI_EN_GPP_D_0 */
391 	.offset = 0x154,
392 	.mask = 0x1,
393 	.data = 0x0,	/* Disable interrupt generation */
394 	},
395 	{	/* GPI_NMI_EN_GPP_D_0 */
396 	.offset = 0x174,
397 	.mask = 0x1,
398 	.data = 0x0,	/* Disable interrupt generation */
399 	},
400 
401 /* Setup GPP_D_0 Pad Config: */
402 	{	/* PAD_CFG_DW0_GPP_D_0 */
403 	.offset = 0x4c0,
404 	.mask = 0xffffffff,
405 	.data = 0x82020100,
406 /*
407  *  31:30 Pad Reset Config (PADRSTCFG): = 2h  # PLTRST# (default)
408  *
409  *  29    RX Pad State Select (RXPADSTSEL): = 0 # Raw RX pad state directly
410  *                                                from RX buffer (default)
411  *
412  *  28    RX Raw Override to '1' (RXRAW1): = 0 # No Override
413  *
414  *  26:25 RX Level/Edge Configuration (RXEVCFG):
415  *      = 0h # Level
416  *      = 1h # Edge
417  *
418  *  23    RX Invert (RXINV): = 0 # No Inversion (signal active high)
419  *
420  *  20    GPIO Input Route IOxAPIC (GPIROUTIOXAPIC):
421  * = 0 # Routing does not cause peripheral IRQ...
422  *     # (we want an NMI not an IRQ)
423  *
424  *  19    GPIO Input Route SCI (GPIROUTSCI): = 0 # Routing does not cause SCI.
425  *  18    GPIO Input Route SMI (GPIROUTSMI): = 0 # Routing does not cause SMI.
426  *  17    GPIO Input Route NMI (GPIROUTNMI): = 1 # Routing can cause NMI.
427  *
428  *  11:10 Pad Mode (PMODE1/0): = 0h = GPIO control the Pad.
429  *   9    GPIO RX Disable (GPIORXDIS):
430  * = 0 # Enable the input buffer (active low enable)
431  *
432  *   8    GPIO TX Disable (GPIOTXDIS):
433  * = 1 # Disable the output buffer; i.e. Hi-Z
434  *
435  *   1 GPIO RX State (GPIORXSTATE): This is the current internal RX pad state..
436  *   0 GPIO TX State (GPIOTXSTATE):
437  * = 0 # (Leave at default)
438  */
439 	},
440 
441 /* Pad Config DW1 */
442 	{	/* PAD_CFG_DW1_GPP_D_0 */
443 	.offset = 0x4c4,
444 	.mask = 0x3c00,
445 	.data = 0,	/* Termination = none (default) */
446 	},
447 };
448 
449 static void uv_init_hubless_pch_d0(void)
450 {
451 	int i, read;
452 
453 	read = *PCH_PCR_GPIO_ADDRESS(PAD_OWN_GPP_D_0);
454 	if (read != 0) {
455 		pr_info("UV: Hubless NMI already configured\n");
456 		return;
457 	}
458 
459 	nmi_debug("UV: Initializing UV Hubless NMI on PCH\n");
460 	for (i = 0; i < ARRAY_SIZE(init_nmi); i++) {
461 		uv_init_hubless_pch_io(init_nmi[i].offset,
462 					init_nmi[i].mask,
463 					init_nmi[i].data);
464 	}
465 }
466 
467 static int uv_nmi_test_hubless(struct uv_hub_nmi_s *hub_nmi)
468 {
469 	int *pstat = PCH_PCR_GPIO_ADDRESS(GPI_NMI_STS_GPP_D_0);
470 	int status = *pstat;
471 
472 	hub_nmi->nmi_value = status;
473 	atomic_inc(&hub_nmi->read_mmr_count);
474 
475 	if (!(status & STS_GPP_D_0_MASK))	/* Not a UV external NMI */
476 		return 0;
477 
478 	*pstat = STS_GPP_D_0_MASK;	/* Is a UV NMI: clear GPP_D_0 status */
479 	(void)*pstat;			/* Flush write */
480 
481 	return 1;
482 }
483 
484 static int uv_test_nmi(struct uv_hub_nmi_s *hub_nmi)
485 {
486 	if (hub_nmi->hub_present)
487 		return uv_nmi_test_mmr(hub_nmi);
488 
489 	if (hub_nmi->pch_owner)		/* Only PCH owner can check status */
490 		return uv_nmi_test_hubless(hub_nmi);
491 
492 	return -1;
493 }
494 
495 /*
496  * If first CPU in on this hub, set hub_nmi "in_nmi" and "owner" values and
497  * return true.  If first CPU in on the system, set global "in_nmi" flag.
498  */
499 static int uv_set_in_nmi(int cpu, struct uv_hub_nmi_s *hub_nmi)
500 {
501 	int first = atomic_add_unless(&hub_nmi->in_nmi, 1, 1);
502 
503 	if (first) {
504 		atomic_set(&hub_nmi->cpu_owner, cpu);
505 		if (atomic_add_unless(&uv_in_nmi, 1, 1))
506 			atomic_set(&uv_nmi_cpu, cpu);
507 
508 		atomic_inc(&hub_nmi->nmi_count);
509 	}
510 	return first;
511 }
512 
513 /* Check if this is a system NMI event */
514 static int uv_check_nmi(struct uv_hub_nmi_s *hub_nmi)
515 {
516 	int cpu = smp_processor_id();
517 	int nmi = 0;
518 	int nmi_detected = 0;
519 
520 	local64_inc(&uv_nmi_count);
521 	this_cpu_inc(uv_cpu_nmi.queries);
522 
523 	do {
524 		nmi = atomic_read(&hub_nmi->in_nmi);
525 		if (nmi)
526 			break;
527 
528 		if (raw_spin_trylock(&hub_nmi->nmi_lock)) {
529 			nmi_detected = uv_test_nmi(hub_nmi);
530 
531 			/* Check flag for UV external NMI */
532 			if (nmi_detected > 0) {
533 				uv_set_in_nmi(cpu, hub_nmi);
534 				nmi = 1;
535 				break;
536 			}
537 
538 			/* A non-PCH node in a hubless system waits for NMI */
539 			else if (nmi_detected < 0)
540 				goto slave_wait;
541 
542 			/* MMR/PCH NMI flag is clear */
543 			raw_spin_unlock(&hub_nmi->nmi_lock);
544 
545 		} else {
546 
547 			/* Wait a moment for the HUB NMI locker to set flag */
548 slave_wait:		cpu_relax();
549 			udelay(uv_nmi_slave_delay);
550 
551 			/* Re-check hub in_nmi flag */
552 			nmi = atomic_read(&hub_nmi->in_nmi);
553 			if (nmi)
554 				break;
555 		}
556 
557 		/*
558 		 * Check if this BMC missed setting the MMR NMI flag (or)
559 		 * UV hubless system where only PCH owner can check flag
560 		 */
561 		if (!nmi) {
562 			nmi = atomic_read(&uv_in_nmi);
563 			if (nmi)
564 				uv_set_in_nmi(cpu, hub_nmi);
565 		}
566 
567 		/* If we're holding the hub lock, release it now */
568 		if (nmi_detected < 0)
569 			raw_spin_unlock(&hub_nmi->nmi_lock);
570 
571 	} while (0);
572 
573 	if (!nmi)
574 		local64_inc(&uv_nmi_misses);
575 
576 	return nmi;
577 }
578 
579 /* Need to reset the NMI MMR register, but only once per hub. */
580 static inline void uv_clear_nmi(int cpu)
581 {
582 	struct uv_hub_nmi_s *hub_nmi = uv_hub_nmi;
583 
584 	if (cpu == atomic_read(&hub_nmi->cpu_owner)) {
585 		atomic_set(&hub_nmi->cpu_owner, -1);
586 		atomic_set(&hub_nmi->in_nmi, 0);
587 		if (hub_nmi->hub_present)
588 			uv_local_mmr_clear_nmi();
589 		else
590 			uv_reassert_nmi();
591 		raw_spin_unlock(&hub_nmi->nmi_lock);
592 	}
593 }
594 
595 /* Ping non-responding CPU's attempting to force them into the NMI handler */
596 static void uv_nmi_nr_cpus_ping(void)
597 {
598 	int cpu;
599 
600 	for_each_cpu(cpu, uv_nmi_cpu_mask)
601 		uv_cpu_nmi_per(cpu).pinging = 1;
602 
603 	apic->send_IPI_mask(uv_nmi_cpu_mask, APIC_DM_NMI);
604 }
605 
606 /* Clean up flags for CPU's that ignored both NMI and ping */
607 static void uv_nmi_cleanup_mask(void)
608 {
609 	int cpu;
610 
611 	for_each_cpu(cpu, uv_nmi_cpu_mask) {
612 		uv_cpu_nmi_per(cpu).pinging =  0;
613 		uv_cpu_nmi_per(cpu).state = UV_NMI_STATE_OUT;
614 		cpumask_clear_cpu(cpu, uv_nmi_cpu_mask);
615 	}
616 }
617 
618 /* Loop waiting as CPU's enter NMI handler */
619 static int uv_nmi_wait_cpus(int first)
620 {
621 	int i, j, k, n = num_online_cpus();
622 	int last_k = 0, waiting = 0;
623 	int cpu = smp_processor_id();
624 
625 	if (first) {
626 		cpumask_copy(uv_nmi_cpu_mask, cpu_online_mask);
627 		k = 0;
628 	} else {
629 		k = n - cpumask_weight(uv_nmi_cpu_mask);
630 	}
631 
632 	/* PCH NMI causes only one CPU to respond */
633 	if (first && uv_pch_intr_now_enabled) {
634 		cpumask_clear_cpu(cpu, uv_nmi_cpu_mask);
635 		return n - k - 1;
636 	}
637 
638 	udelay(uv_nmi_initial_delay);
639 	for (i = 0; i < uv_nmi_retry_count; i++) {
640 		int loop_delay = uv_nmi_loop_delay;
641 
642 		for_each_cpu(j, uv_nmi_cpu_mask) {
643 			if (uv_cpu_nmi_per(j).state) {
644 				cpumask_clear_cpu(j, uv_nmi_cpu_mask);
645 				if (++k >= n)
646 					break;
647 			}
648 		}
649 		if (k >= n) {		/* all in? */
650 			k = n;
651 			break;
652 		}
653 		if (last_k != k) {	/* abort if no new CPU's coming in */
654 			last_k = k;
655 			waiting = 0;
656 		} else if (++waiting > uv_nmi_wait_count)
657 			break;
658 
659 		/* Extend delay if waiting only for CPU 0: */
660 		if (waiting && (n - k) == 1 &&
661 		    cpumask_test_cpu(0, uv_nmi_cpu_mask))
662 			loop_delay *= 100;
663 
664 		udelay(loop_delay);
665 	}
666 	atomic_set(&uv_nmi_cpus_in_nmi, k);
667 	return n - k;
668 }
669 
670 /* Wait until all slave CPU's have entered UV NMI handler */
671 static void uv_nmi_wait(int master)
672 {
673 	/* Indicate this CPU is in: */
674 	this_cpu_write(uv_cpu_nmi.state, UV_NMI_STATE_IN);
675 
676 	/* If not the first CPU in (the master), then we are a slave CPU */
677 	if (!master)
678 		return;
679 
680 	do {
681 		/* Wait for all other CPU's to gather here */
682 		if (!uv_nmi_wait_cpus(1))
683 			break;
684 
685 		/* If not all made it in, send IPI NMI to them */
686 		pr_alert("UV: Sending NMI IPI to %d CPUs: %*pbl\n",
687 			 cpumask_weight(uv_nmi_cpu_mask),
688 			 cpumask_pr_args(uv_nmi_cpu_mask));
689 
690 		uv_nmi_nr_cpus_ping();
691 
692 		/* If all CPU's are in, then done */
693 		if (!uv_nmi_wait_cpus(0))
694 			break;
695 
696 		pr_alert("UV: %d CPUs not in NMI loop: %*pbl\n",
697 			 cpumask_weight(uv_nmi_cpu_mask),
698 			 cpumask_pr_args(uv_nmi_cpu_mask));
699 	} while (0);
700 
701 	pr_alert("UV: %d of %d CPUs in NMI\n",
702 		atomic_read(&uv_nmi_cpus_in_nmi), num_online_cpus());
703 }
704 
705 /* Dump Instruction Pointer header */
706 static void uv_nmi_dump_cpu_ip_hdr(void)
707 {
708 	pr_info("\nUV: %4s %6s %-32s %s   (Note: PID 0 not listed)\n",
709 		"CPU", "PID", "COMMAND", "IP");
710 }
711 
712 /* Dump Instruction Pointer info */
713 static void uv_nmi_dump_cpu_ip(int cpu, struct pt_regs *regs)
714 {
715 	pr_info("UV: %4d %6d %-32.32s %pS",
716 		cpu, current->pid, current->comm, (void *)regs->ip);
717 }
718 
719 /*
720  * Dump this CPU's state.  If action was set to "kdump" and the crash_kexec
721  * failed, then we provide "dump" as an alternate action.  Action "dump" now
722  * also includes the show "ips" (instruction pointers) action whereas the
723  * action "ips" only displays instruction pointers for the non-idle CPU's.
724  * This is an abbreviated form of the "ps" command.
725  */
726 static void uv_nmi_dump_state_cpu(int cpu, struct pt_regs *regs)
727 {
728 	const char *dots = " ................................. ";
729 
730 	if (cpu == 0)
731 		uv_nmi_dump_cpu_ip_hdr();
732 
733 	if (current->pid != 0 || !uv_nmi_action_is("ips"))
734 		uv_nmi_dump_cpu_ip(cpu, regs);
735 
736 	if (uv_nmi_action_is("dump")) {
737 		pr_info("UV:%sNMI process trace for CPU %d\n", dots, cpu);
738 		show_regs(regs);
739 	}
740 
741 	this_cpu_write(uv_cpu_nmi.state, UV_NMI_STATE_DUMP_DONE);
742 }
743 
744 /* Trigger a slave CPU to dump it's state */
745 static void uv_nmi_trigger_dump(int cpu)
746 {
747 	int retry = uv_nmi_trigger_delay;
748 
749 	if (uv_cpu_nmi_per(cpu).state != UV_NMI_STATE_IN)
750 		return;
751 
752 	uv_cpu_nmi_per(cpu).state = UV_NMI_STATE_DUMP;
753 	do {
754 		cpu_relax();
755 		udelay(10);
756 		if (uv_cpu_nmi_per(cpu).state
757 				!= UV_NMI_STATE_DUMP)
758 			return;
759 	} while (--retry > 0);
760 
761 	pr_crit("UV: CPU %d stuck in process dump function\n", cpu);
762 	uv_cpu_nmi_per(cpu).state = UV_NMI_STATE_DUMP_DONE;
763 }
764 
765 /* Wait until all CPU's ready to exit */
766 static void uv_nmi_sync_exit(int master)
767 {
768 	atomic_dec(&uv_nmi_cpus_in_nmi);
769 	if (master) {
770 		while (atomic_read(&uv_nmi_cpus_in_nmi) > 0)
771 			cpu_relax();
772 		atomic_set(&uv_nmi_slave_continue, SLAVE_CLEAR);
773 	} else {
774 		while (atomic_read(&uv_nmi_slave_continue))
775 			cpu_relax();
776 	}
777 }
778 
779 /* Current "health" check is to check which CPU's are responsive */
780 static void uv_nmi_action_health(int cpu, struct pt_regs *regs, int master)
781 {
782 	if (master) {
783 		int in = atomic_read(&uv_nmi_cpus_in_nmi);
784 		int out = num_online_cpus() - in;
785 
786 		pr_alert("UV: NMI CPU health check (non-responding:%d)\n", out);
787 		atomic_set(&uv_nmi_slave_continue, SLAVE_EXIT);
788 	} else {
789 		while (!atomic_read(&uv_nmi_slave_continue))
790 			cpu_relax();
791 	}
792 	uv_nmi_sync_exit(master);
793 }
794 
795 /* Walk through CPU list and dump state of each */
796 static void uv_nmi_dump_state(int cpu, struct pt_regs *regs, int master)
797 {
798 	if (master) {
799 		int tcpu;
800 		int ignored = 0;
801 		int saved_console_loglevel = console_loglevel;
802 
803 		pr_alert("UV: tracing %s for %d CPUs from CPU %d\n",
804 			uv_nmi_action_is("ips") ? "IPs" : "processes",
805 			atomic_read(&uv_nmi_cpus_in_nmi), cpu);
806 
807 		console_loglevel = uv_nmi_loglevel;
808 		atomic_set(&uv_nmi_slave_continue, SLAVE_EXIT);
809 		for_each_online_cpu(tcpu) {
810 			if (cpumask_test_cpu(tcpu, uv_nmi_cpu_mask))
811 				ignored++;
812 			else if (tcpu == cpu)
813 				uv_nmi_dump_state_cpu(tcpu, regs);
814 			else
815 				uv_nmi_trigger_dump(tcpu);
816 		}
817 		if (ignored)
818 			pr_alert("UV: %d CPUs ignored NMI\n", ignored);
819 
820 		console_loglevel = saved_console_loglevel;
821 		pr_alert("UV: process trace complete\n");
822 	} else {
823 		while (!atomic_read(&uv_nmi_slave_continue))
824 			cpu_relax();
825 		while (this_cpu_read(uv_cpu_nmi.state) != UV_NMI_STATE_DUMP)
826 			cpu_relax();
827 		uv_nmi_dump_state_cpu(cpu, regs);
828 	}
829 	uv_nmi_sync_exit(master);
830 }
831 
832 static void uv_nmi_touch_watchdogs(void)
833 {
834 	touch_softlockup_watchdog_sync();
835 	clocksource_touch_watchdog();
836 	rcu_cpu_stall_reset();
837 	touch_nmi_watchdog();
838 }
839 
840 static void uv_nmi_kdump(int cpu, int main, struct pt_regs *regs)
841 {
842 	/* Check if kdump kernel loaded for both main and secondary CPUs */
843 	if (!kexec_crash_image) {
844 		if (main)
845 			pr_err("UV: NMI error: kdump kernel not loaded\n");
846 		return;
847 	}
848 
849 	/* Call crash to dump system state */
850 	if (main) {
851 		pr_emerg("UV: NMI executing crash_kexec on CPU%d\n", cpu);
852 		crash_kexec(regs);
853 
854 		pr_emerg("UV: crash_kexec unexpectedly returned\n");
855 		atomic_set(&uv_nmi_kexec_failed, 1);
856 
857 	} else { /* secondary */
858 
859 		/* If kdump kernel fails, secondaries will exit this loop */
860 		while (atomic_read(&uv_nmi_kexec_failed) == 0) {
861 
862 			/* Once shootdown cpus starts, they do not return */
863 			run_crash_ipi_callback(regs);
864 
865 			mdelay(10);
866 		}
867 	}
868 }
869 
870 #ifdef CONFIG_KGDB
871 #ifdef CONFIG_KGDB_KDB
872 static inline int uv_nmi_kdb_reason(void)
873 {
874 	return KDB_REASON_SYSTEM_NMI;
875 }
876 #else /* !CONFIG_KGDB_KDB */
877 static inline int uv_nmi_kdb_reason(void)
878 {
879 	/* Ensure user is expecting to attach gdb remote */
880 	if (uv_nmi_action_is("kgdb"))
881 		return 0;
882 
883 	pr_err("UV: NMI error: KDB is not enabled in this kernel\n");
884 	return -1;
885 }
886 #endif /* CONFIG_KGDB_KDB */
887 
888 /*
889  * Call KGDB/KDB from NMI handler
890  *
891  * Note that if both KGDB and KDB are configured, then the action of 'kgdb' or
892  * 'kdb' has no affect on which is used.  See the KGDB documentation for further
893  * information.
894  */
895 static void uv_call_kgdb_kdb(int cpu, struct pt_regs *regs, int master)
896 {
897 	if (master) {
898 		int reason = uv_nmi_kdb_reason();
899 		int ret;
900 
901 		if (reason < 0)
902 			return;
903 
904 		/* Call KGDB NMI handler as MASTER */
905 		ret = kgdb_nmicallin(cpu, X86_TRAP_NMI, regs, reason,
906 				&uv_nmi_slave_continue);
907 		if (ret) {
908 			pr_alert("KGDB returned error, is kgdboc set?\n");
909 			atomic_set(&uv_nmi_slave_continue, SLAVE_EXIT);
910 		}
911 	} else {
912 		/* Wait for KGDB signal that it's ready for slaves to enter */
913 		int sig;
914 
915 		do {
916 			cpu_relax();
917 			sig = atomic_read(&uv_nmi_slave_continue);
918 		} while (!sig);
919 
920 		/* Call KGDB as slave */
921 		if (sig == SLAVE_CONTINUE)
922 			kgdb_nmicallback(cpu, regs);
923 	}
924 	uv_nmi_sync_exit(master);
925 }
926 
927 #else /* !CONFIG_KGDB */
928 static inline void uv_call_kgdb_kdb(int cpu, struct pt_regs *regs, int master)
929 {
930 	pr_err("UV: NMI error: KGDB is not enabled in this kernel\n");
931 }
932 #endif /* !CONFIG_KGDB */
933 
934 /*
935  * UV NMI handler
936  */
937 static int uv_handle_nmi(unsigned int reason, struct pt_regs *regs)
938 {
939 	struct uv_hub_nmi_s *hub_nmi = uv_hub_nmi;
940 	int cpu = smp_processor_id();
941 	int master = 0;
942 	unsigned long flags;
943 
944 	local_irq_save(flags);
945 
946 	/* If not a UV System NMI, ignore */
947 	if (!this_cpu_read(uv_cpu_nmi.pinging) && !uv_check_nmi(hub_nmi)) {
948 		local_irq_restore(flags);
949 		return NMI_DONE;
950 	}
951 
952 	/* Indicate we are the first CPU into the NMI handler */
953 	master = (atomic_read(&uv_nmi_cpu) == cpu);
954 
955 	/* If NMI action is "kdump", then attempt to do it */
956 	if (uv_nmi_action_is("kdump")) {
957 		uv_nmi_kdump(cpu, master, regs);
958 
959 		/* Unexpected return, revert action to "dump" */
960 		if (master)
961 			strncpy(uv_nmi_action, "dump", strlen(uv_nmi_action));
962 	}
963 
964 	/* Pause as all CPU's enter the NMI handler */
965 	uv_nmi_wait(master);
966 
967 	/* Process actions other than "kdump": */
968 	if (uv_nmi_action_is("health")) {
969 		uv_nmi_action_health(cpu, regs, master);
970 	} else if (uv_nmi_action_is("ips") || uv_nmi_action_is("dump")) {
971 		uv_nmi_dump_state(cpu, regs, master);
972 	} else if (uv_nmi_action_is("kdb") || uv_nmi_action_is("kgdb")) {
973 		uv_call_kgdb_kdb(cpu, regs, master);
974 	} else {
975 		if (master)
976 			pr_alert("UV: unknown NMI action: %s\n", uv_nmi_action);
977 		uv_nmi_sync_exit(master);
978 	}
979 
980 	/* Clear per_cpu "in_nmi" flag */
981 	this_cpu_write(uv_cpu_nmi.state, UV_NMI_STATE_OUT);
982 
983 	/* Clear MMR NMI flag on each hub */
984 	uv_clear_nmi(cpu);
985 
986 	/* Clear global flags */
987 	if (master) {
988 		if (cpumask_weight(uv_nmi_cpu_mask))
989 			uv_nmi_cleanup_mask();
990 		atomic_set(&uv_nmi_cpus_in_nmi, -1);
991 		atomic_set(&uv_nmi_cpu, -1);
992 		atomic_set(&uv_in_nmi, 0);
993 		atomic_set(&uv_nmi_kexec_failed, 0);
994 		atomic_set(&uv_nmi_slave_continue, SLAVE_CLEAR);
995 	}
996 
997 	uv_nmi_touch_watchdogs();
998 	local_irq_restore(flags);
999 
1000 	return NMI_HANDLED;
1001 }
1002 
1003 /*
1004  * NMI handler for pulling in CPU's when perf events are grabbing our NMI
1005  */
1006 static int uv_handle_nmi_ping(unsigned int reason, struct pt_regs *regs)
1007 {
1008 	int ret;
1009 
1010 	this_cpu_inc(uv_cpu_nmi.queries);
1011 	if (!this_cpu_read(uv_cpu_nmi.pinging)) {
1012 		local64_inc(&uv_nmi_ping_misses);
1013 		return NMI_DONE;
1014 	}
1015 
1016 	this_cpu_inc(uv_cpu_nmi.pings);
1017 	local64_inc(&uv_nmi_ping_count);
1018 	ret = uv_handle_nmi(reason, regs);
1019 	this_cpu_write(uv_cpu_nmi.pinging, 0);
1020 	return ret;
1021 }
1022 
1023 static void uv_register_nmi_notifier(void)
1024 {
1025 	if (register_nmi_handler(NMI_UNKNOWN, uv_handle_nmi, 0, "uv"))
1026 		pr_warn("UV: NMI handler failed to register\n");
1027 
1028 	if (register_nmi_handler(NMI_LOCAL, uv_handle_nmi_ping, 0, "uvping"))
1029 		pr_warn("UV: PING NMI handler failed to register\n");
1030 }
1031 
1032 void uv_nmi_init(void)
1033 {
1034 	unsigned int value;
1035 
1036 	/*
1037 	 * Unmask NMI on all CPU's
1038 	 */
1039 	value = apic_read(APIC_LVT1) | APIC_DM_NMI;
1040 	value &= ~APIC_LVT_MASKED;
1041 	apic_write(APIC_LVT1, value);
1042 }
1043 
1044 /* Setup HUB NMI info */
1045 static void __init uv_nmi_setup_common(bool hubbed)
1046 {
1047 	int size = sizeof(void *) * (1 << NODES_SHIFT);
1048 	int cpu;
1049 
1050 	uv_hub_nmi_list = kzalloc(size, GFP_KERNEL);
1051 	nmi_debug("UV: NMI hub list @ 0x%p (%d)\n", uv_hub_nmi_list, size);
1052 	BUG_ON(!uv_hub_nmi_list);
1053 	size = sizeof(struct uv_hub_nmi_s);
1054 	for_each_present_cpu(cpu) {
1055 		int nid = cpu_to_node(cpu);
1056 		if (uv_hub_nmi_list[nid] == NULL) {
1057 			uv_hub_nmi_list[nid] = kzalloc_node(size,
1058 							    GFP_KERNEL, nid);
1059 			BUG_ON(!uv_hub_nmi_list[nid]);
1060 			raw_spin_lock_init(&(uv_hub_nmi_list[nid]->nmi_lock));
1061 			atomic_set(&uv_hub_nmi_list[nid]->cpu_owner, -1);
1062 			uv_hub_nmi_list[nid]->hub_present = hubbed;
1063 			uv_hub_nmi_list[nid]->pch_owner = (nid == 0);
1064 		}
1065 		uv_hub_nmi_per(cpu) = uv_hub_nmi_list[nid];
1066 	}
1067 	BUG_ON(!alloc_cpumask_var(&uv_nmi_cpu_mask, GFP_KERNEL));
1068 }
1069 
1070 /* Setup for UV Hub systems */
1071 void __init uv_nmi_setup(void)
1072 {
1073 	uv_nmi_setup_mmrs();
1074 	uv_nmi_setup_common(true);
1075 	uv_register_nmi_notifier();
1076 	pr_info("UV: Hub NMI enabled\n");
1077 }
1078 
1079 /* Setup for UV Hubless systems */
1080 void __init uv_nmi_setup_hubless(void)
1081 {
1082 	uv_nmi_setup_common(false);
1083 	pch_base = xlate_dev_mem_ptr(PCH_PCR_GPIO_1_BASE);
1084 	nmi_debug("UV: PCH base:%p from 0x%lx, GPP_D_0\n",
1085 		pch_base, PCH_PCR_GPIO_1_BASE);
1086 	if (uv_pch_init_enable)
1087 		uv_init_hubless_pch_d0();
1088 	uv_init_hubless_pch_io(GPI_NMI_ENA_GPP_D_0,
1089 				STS_GPP_D_0_MASK, STS_GPP_D_0_MASK);
1090 	uv_nmi_setup_hubless_intr();
1091 	/* Ensure NMI enabled in Processor Interface Reg: */
1092 	uv_reassert_nmi();
1093 	uv_register_nmi_notifier();
1094 	pr_info("UV: PCH NMI enabled\n");
1095 }
1096