xref: /openbmc/linux/arch/x86/platform/uv/uv_nmi.c (revision 3ddc8b84)
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];
206 
207 	/* (remove possible '\n') */
208 	strscpy(arg, val, strnchrnul(val, sizeof(arg)-1, '\n') - val + 1);
209 
210 	for (i = 0; i < n; i++)
211 		if (!strcmp(arg, valid_acts[i].action))
212 			break;
213 
214 	if (i < n) {
215 		strscpy(uv_nmi_action, arg, sizeof(uv_nmi_action));
216 		pr_info("UV: New NMI action:%s\n", uv_nmi_action);
217 		return 0;
218 	}
219 
220 	pr_err("UV: Invalid NMI action:%s, valid actions are:\n", arg);
221 	for (i = 0; i < n; i++)
222 		pr_err("UV: %-8s - %s\n",
223 			valid_acts[i].action, valid_acts[i].desc);
224 	return -EINVAL;
225 }
226 
227 static const struct kernel_param_ops param_ops_action = {
228 	.get = param_get_action,
229 	.set = param_set_action,
230 };
231 #define param_check_action(name, p) __param_check(name, p, action_t)
232 
233 module_param_named(action, uv_nmi_action, action, 0644);
234 
235 static inline bool uv_nmi_action_is(const char *action)
236 {
237 	return (strncmp(uv_nmi_action, action, strlen(action)) == 0);
238 }
239 
240 /* Setup which NMI support is present in system */
241 static void uv_nmi_setup_mmrs(void)
242 {
243 	bool new_nmi_method_only = false;
244 
245 	/* First determine arch specific MMRs to handshake with BIOS */
246 	if (UVH_EVENT_OCCURRED0_EXTIO_INT0_MASK) {	/* UV2,3,4 setup */
247 		uvh_nmi_mmrx = UVH_EVENT_OCCURRED0;
248 		uvh_nmi_mmrx_clear = UVH_EVENT_OCCURRED0_ALIAS;
249 		uvh_nmi_mmrx_shift = UVH_EVENT_OCCURRED0_EXTIO_INT0_SHFT;
250 		uvh_nmi_mmrx_type = "OCRD0-EXTIO_INT0";
251 
252 		uvh_nmi_mmrx_supported = UVH_EXTIO_INT0_BROADCAST;
253 		uvh_nmi_mmrx_req = UVH_BIOS_KERNEL_MMR_ALIAS_2;
254 		uvh_nmi_mmrx_req_shift = 62;
255 
256 	} else if (UVH_EVENT_OCCURRED1_EXTIO_INT0_MASK) { /* UV5+ setup */
257 		uvh_nmi_mmrx = UVH_EVENT_OCCURRED1;
258 		uvh_nmi_mmrx_clear = UVH_EVENT_OCCURRED1_ALIAS;
259 		uvh_nmi_mmrx_shift = UVH_EVENT_OCCURRED1_EXTIO_INT0_SHFT;
260 		uvh_nmi_mmrx_type = "OCRD1-EXTIO_INT0";
261 
262 		new_nmi_method_only = true;		/* Newer nmi always valid on UV5+ */
263 		uvh_nmi_mmrx_req = 0;			/* no request bit to clear */
264 
265 	} else {
266 		pr_err("UV:%s:NMI support not available on this system\n", __func__);
267 		return;
268 	}
269 
270 	/* Then find out if new NMI is supported */
271 	if (new_nmi_method_only || uv_read_local_mmr(uvh_nmi_mmrx_supported)) {
272 		if (uvh_nmi_mmrx_req)
273 			uv_write_local_mmr(uvh_nmi_mmrx_req,
274 						1UL << uvh_nmi_mmrx_req_shift);
275 		nmi_mmr = uvh_nmi_mmrx;
276 		nmi_mmr_clear = uvh_nmi_mmrx_clear;
277 		nmi_mmr_pending = 1UL << uvh_nmi_mmrx_shift;
278 		pr_info("UV: SMI NMI support: %s\n", uvh_nmi_mmrx_type);
279 	} else {
280 		nmi_mmr = UVH_NMI_MMR;
281 		nmi_mmr_clear = UVH_NMI_MMR_CLEAR;
282 		nmi_mmr_pending = 1UL << UVH_NMI_MMR_SHIFT;
283 		pr_info("UV: SMI NMI support: %s\n", UVH_NMI_MMR_TYPE);
284 	}
285 }
286 
287 /* Read NMI MMR and check if NMI flag was set by BMC. */
288 static inline int uv_nmi_test_mmr(struct uv_hub_nmi_s *hub_nmi)
289 {
290 	hub_nmi->nmi_value = uv_read_local_mmr(nmi_mmr);
291 	atomic_inc(&hub_nmi->read_mmr_count);
292 	return !!(hub_nmi->nmi_value & nmi_mmr_pending);
293 }
294 
295 static inline void uv_local_mmr_clear_nmi(void)
296 {
297 	uv_write_local_mmr(nmi_mmr_clear, nmi_mmr_pending);
298 }
299 
300 /*
301  * UV hubless NMI handler functions
302  */
303 static inline void uv_reassert_nmi(void)
304 {
305 	/* (from arch/x86/include/asm/mach_traps.h) */
306 	outb(0x8f, NMI_CONTROL_PORT);
307 	inb(NMI_DUMMY_PORT);		/* dummy read */
308 	outb(0x0f, NMI_CONTROL_PORT);
309 	inb(NMI_DUMMY_PORT);		/* dummy read */
310 }
311 
312 static void uv_init_hubless_pch_io(int offset, int mask, int data)
313 {
314 	int *addr = PCH_PCR_GPIO_ADDRESS(offset);
315 	int readd = readl(addr);
316 
317 	if (mask) {			/* OR in new data */
318 		int writed = (readd & ~mask) | data;
319 
320 		nmi_debug("UV:PCH: %p = %x & %x | %x (%x)\n",
321 			addr, readd, ~mask, data, writed);
322 		writel(writed, addr);
323 	} else if (readd & data) {	/* clear status bit */
324 		nmi_debug("UV:PCH: %p = %x\n", addr, data);
325 		writel(data, addr);
326 	}
327 
328 	(void)readl(addr);		/* flush write data */
329 }
330 
331 static void uv_nmi_setup_hubless_intr(void)
332 {
333 	uv_pch_intr_now_enabled = uv_pch_intr_enable;
334 
335 	uv_init_hubless_pch_io(
336 		PAD_CFG_DW0_GPP_D_0, GPIROUTNMI,
337 		uv_pch_intr_now_enabled ? GPIROUTNMI : 0);
338 
339 	nmi_debug("UV:NMI: GPP_D_0 interrupt %s\n",
340 		uv_pch_intr_now_enabled ? "enabled" : "disabled");
341 }
342 
343 static struct init_nmi {
344 	unsigned int	offset;
345 	unsigned int	mask;
346 	unsigned int	data;
347 } init_nmi[] = {
348 	{	/* HOSTSW_OWN_GPP_D_0 */
349 	.offset = 0x84,
350 	.mask = 0x1,
351 	.data = 0x0,	/* ACPI Mode */
352 	},
353 
354 /* Clear status: */
355 	{	/* GPI_INT_STS_GPP_D_0 */
356 	.offset = 0x104,
357 	.mask = 0x0,
358 	.data = 0x1,	/* Clear Status */
359 	},
360 	{	/* GPI_GPE_STS_GPP_D_0 */
361 	.offset = 0x124,
362 	.mask = 0x0,
363 	.data = 0x1,	/* Clear Status */
364 	},
365 	{	/* GPI_SMI_STS_GPP_D_0 */
366 	.offset = 0x144,
367 	.mask = 0x0,
368 	.data = 0x1,	/* Clear Status */
369 	},
370 	{	/* GPI_NMI_STS_GPP_D_0 */
371 	.offset = 0x164,
372 	.mask = 0x0,
373 	.data = 0x1,	/* Clear Status */
374 	},
375 
376 /* Disable interrupts: */
377 	{	/* GPI_INT_EN_GPP_D_0 */
378 	.offset = 0x114,
379 	.mask = 0x1,
380 	.data = 0x0,	/* Disable interrupt generation */
381 	},
382 	{	/* GPI_GPE_EN_GPP_D_0 */
383 	.offset = 0x134,
384 	.mask = 0x1,
385 	.data = 0x0,	/* Disable interrupt generation */
386 	},
387 	{	/* GPI_SMI_EN_GPP_D_0 */
388 	.offset = 0x154,
389 	.mask = 0x1,
390 	.data = 0x0,	/* Disable interrupt generation */
391 	},
392 	{	/* GPI_NMI_EN_GPP_D_0 */
393 	.offset = 0x174,
394 	.mask = 0x1,
395 	.data = 0x0,	/* Disable interrupt generation */
396 	},
397 
398 /* Setup GPP_D_0 Pad Config: */
399 	{	/* PAD_CFG_DW0_GPP_D_0 */
400 	.offset = 0x4c0,
401 	.mask = 0xffffffff,
402 	.data = 0x82020100,
403 /*
404  *  31:30 Pad Reset Config (PADRSTCFG): = 2h  # PLTRST# (default)
405  *
406  *  29    RX Pad State Select (RXPADSTSEL): = 0 # Raw RX pad state directly
407  *                                                from RX buffer (default)
408  *
409  *  28    RX Raw Override to '1' (RXRAW1): = 0 # No Override
410  *
411  *  26:25 RX Level/Edge Configuration (RXEVCFG):
412  *      = 0h # Level
413  *      = 1h # Edge
414  *
415  *  23    RX Invert (RXINV): = 0 # No Inversion (signal active high)
416  *
417  *  20    GPIO Input Route IOxAPIC (GPIROUTIOXAPIC):
418  * = 0 # Routing does not cause peripheral IRQ...
419  *     # (we want an NMI not an IRQ)
420  *
421  *  19    GPIO Input Route SCI (GPIROUTSCI): = 0 # Routing does not cause SCI.
422  *  18    GPIO Input Route SMI (GPIROUTSMI): = 0 # Routing does not cause SMI.
423  *  17    GPIO Input Route NMI (GPIROUTNMI): = 1 # Routing can cause NMI.
424  *
425  *  11:10 Pad Mode (PMODE1/0): = 0h = GPIO control the Pad.
426  *   9    GPIO RX Disable (GPIORXDIS):
427  * = 0 # Enable the input buffer (active low enable)
428  *
429  *   8    GPIO TX Disable (GPIOTXDIS):
430  * = 1 # Disable the output buffer; i.e. Hi-Z
431  *
432  *   1 GPIO RX State (GPIORXSTATE): This is the current internal RX pad state..
433  *   0 GPIO TX State (GPIOTXSTATE):
434  * = 0 # (Leave at default)
435  */
436 	},
437 
438 /* Pad Config DW1 */
439 	{	/* PAD_CFG_DW1_GPP_D_0 */
440 	.offset = 0x4c4,
441 	.mask = 0x3c00,
442 	.data = 0,	/* Termination = none (default) */
443 	},
444 };
445 
446 static void uv_init_hubless_pch_d0(void)
447 {
448 	int i, read;
449 
450 	read = *PCH_PCR_GPIO_ADDRESS(PAD_OWN_GPP_D_0);
451 	if (read != 0) {
452 		pr_info("UV: Hubless NMI already configured\n");
453 		return;
454 	}
455 
456 	nmi_debug("UV: Initializing UV Hubless NMI on PCH\n");
457 	for (i = 0; i < ARRAY_SIZE(init_nmi); i++) {
458 		uv_init_hubless_pch_io(init_nmi[i].offset,
459 					init_nmi[i].mask,
460 					init_nmi[i].data);
461 	}
462 }
463 
464 static int uv_nmi_test_hubless(struct uv_hub_nmi_s *hub_nmi)
465 {
466 	int *pstat = PCH_PCR_GPIO_ADDRESS(GPI_NMI_STS_GPP_D_0);
467 	int status = *pstat;
468 
469 	hub_nmi->nmi_value = status;
470 	atomic_inc(&hub_nmi->read_mmr_count);
471 
472 	if (!(status & STS_GPP_D_0_MASK))	/* Not a UV external NMI */
473 		return 0;
474 
475 	*pstat = STS_GPP_D_0_MASK;	/* Is a UV NMI: clear GPP_D_0 status */
476 	(void)*pstat;			/* Flush write */
477 
478 	return 1;
479 }
480 
481 static int uv_test_nmi(struct uv_hub_nmi_s *hub_nmi)
482 {
483 	if (hub_nmi->hub_present)
484 		return uv_nmi_test_mmr(hub_nmi);
485 
486 	if (hub_nmi->pch_owner)		/* Only PCH owner can check status */
487 		return uv_nmi_test_hubless(hub_nmi);
488 
489 	return -1;
490 }
491 
492 /*
493  * If first CPU in on this hub, set hub_nmi "in_nmi" and "owner" values and
494  * return true.  If first CPU in on the system, set global "in_nmi" flag.
495  */
496 static int uv_set_in_nmi(int cpu, struct uv_hub_nmi_s *hub_nmi)
497 {
498 	int first = atomic_add_unless(&hub_nmi->in_nmi, 1, 1);
499 
500 	if (first) {
501 		atomic_set(&hub_nmi->cpu_owner, cpu);
502 		if (atomic_add_unless(&uv_in_nmi, 1, 1))
503 			atomic_set(&uv_nmi_cpu, cpu);
504 
505 		atomic_inc(&hub_nmi->nmi_count);
506 	}
507 	return first;
508 }
509 
510 /* Check if this is a system NMI event */
511 static int uv_check_nmi(struct uv_hub_nmi_s *hub_nmi)
512 {
513 	int cpu = smp_processor_id();
514 	int nmi = 0;
515 	int nmi_detected = 0;
516 
517 	local64_inc(&uv_nmi_count);
518 	this_cpu_inc(uv_cpu_nmi.queries);
519 
520 	do {
521 		nmi = atomic_read(&hub_nmi->in_nmi);
522 		if (nmi)
523 			break;
524 
525 		if (raw_spin_trylock(&hub_nmi->nmi_lock)) {
526 			nmi_detected = uv_test_nmi(hub_nmi);
527 
528 			/* Check flag for UV external NMI */
529 			if (nmi_detected > 0) {
530 				uv_set_in_nmi(cpu, hub_nmi);
531 				nmi = 1;
532 				break;
533 			}
534 
535 			/* A non-PCH node in a hubless system waits for NMI */
536 			else if (nmi_detected < 0)
537 				goto slave_wait;
538 
539 			/* MMR/PCH NMI flag is clear */
540 			raw_spin_unlock(&hub_nmi->nmi_lock);
541 
542 		} else {
543 
544 			/* Wait a moment for the HUB NMI locker to set flag */
545 slave_wait:		cpu_relax();
546 			udelay(uv_nmi_slave_delay);
547 
548 			/* Re-check hub in_nmi flag */
549 			nmi = atomic_read(&hub_nmi->in_nmi);
550 			if (nmi)
551 				break;
552 		}
553 
554 		/*
555 		 * Check if this BMC missed setting the MMR NMI flag (or)
556 		 * UV hubless system where only PCH owner can check flag
557 		 */
558 		if (!nmi) {
559 			nmi = atomic_read(&uv_in_nmi);
560 			if (nmi)
561 				uv_set_in_nmi(cpu, hub_nmi);
562 		}
563 
564 		/* If we're holding the hub lock, release it now */
565 		if (nmi_detected < 0)
566 			raw_spin_unlock(&hub_nmi->nmi_lock);
567 
568 	} while (0);
569 
570 	if (!nmi)
571 		local64_inc(&uv_nmi_misses);
572 
573 	return nmi;
574 }
575 
576 /* Need to reset the NMI MMR register, but only once per hub. */
577 static inline void uv_clear_nmi(int cpu)
578 {
579 	struct uv_hub_nmi_s *hub_nmi = uv_hub_nmi;
580 
581 	if (cpu == atomic_read(&hub_nmi->cpu_owner)) {
582 		atomic_set(&hub_nmi->cpu_owner, -1);
583 		atomic_set(&hub_nmi->in_nmi, 0);
584 		if (hub_nmi->hub_present)
585 			uv_local_mmr_clear_nmi();
586 		else
587 			uv_reassert_nmi();
588 		raw_spin_unlock(&hub_nmi->nmi_lock);
589 	}
590 }
591 
592 /* Ping non-responding CPU's attempting to force them into the NMI handler */
593 static void uv_nmi_nr_cpus_ping(void)
594 {
595 	int cpu;
596 
597 	for_each_cpu(cpu, uv_nmi_cpu_mask)
598 		uv_cpu_nmi_per(cpu).pinging = 1;
599 
600 	__apic_send_IPI_mask(uv_nmi_cpu_mask, APIC_DM_NMI);
601 }
602 
603 /* Clean up flags for CPU's that ignored both NMI and ping */
604 static void uv_nmi_cleanup_mask(void)
605 {
606 	int cpu;
607 
608 	for_each_cpu(cpu, uv_nmi_cpu_mask) {
609 		uv_cpu_nmi_per(cpu).pinging =  0;
610 		uv_cpu_nmi_per(cpu).state = UV_NMI_STATE_OUT;
611 		cpumask_clear_cpu(cpu, uv_nmi_cpu_mask);
612 	}
613 }
614 
615 /* Loop waiting as CPU's enter NMI handler */
616 static int uv_nmi_wait_cpus(int first)
617 {
618 	int i, j, k, n = num_online_cpus();
619 	int last_k = 0, waiting = 0;
620 	int cpu = smp_processor_id();
621 
622 	if (first) {
623 		cpumask_copy(uv_nmi_cpu_mask, cpu_online_mask);
624 		k = 0;
625 	} else {
626 		k = n - cpumask_weight(uv_nmi_cpu_mask);
627 	}
628 
629 	/* PCH NMI causes only one CPU to respond */
630 	if (first && uv_pch_intr_now_enabled) {
631 		cpumask_clear_cpu(cpu, uv_nmi_cpu_mask);
632 		return n - k - 1;
633 	}
634 
635 	udelay(uv_nmi_initial_delay);
636 	for (i = 0; i < uv_nmi_retry_count; i++) {
637 		int loop_delay = uv_nmi_loop_delay;
638 
639 		for_each_cpu(j, uv_nmi_cpu_mask) {
640 			if (uv_cpu_nmi_per(j).state) {
641 				cpumask_clear_cpu(j, uv_nmi_cpu_mask);
642 				if (++k >= n)
643 					break;
644 			}
645 		}
646 		if (k >= n) {		/* all in? */
647 			k = n;
648 			break;
649 		}
650 		if (last_k != k) {	/* abort if no new CPU's coming in */
651 			last_k = k;
652 			waiting = 0;
653 		} else if (++waiting > uv_nmi_wait_count)
654 			break;
655 
656 		/* Extend delay if waiting only for CPU 0: */
657 		if (waiting && (n - k) == 1 &&
658 		    cpumask_test_cpu(0, uv_nmi_cpu_mask))
659 			loop_delay *= 100;
660 
661 		udelay(loop_delay);
662 	}
663 	atomic_set(&uv_nmi_cpus_in_nmi, k);
664 	return n - k;
665 }
666 
667 /* Wait until all slave CPU's have entered UV NMI handler */
668 static void uv_nmi_wait(int master)
669 {
670 	/* Indicate this CPU is in: */
671 	this_cpu_write(uv_cpu_nmi.state, UV_NMI_STATE_IN);
672 
673 	/* If not the first CPU in (the master), then we are a slave CPU */
674 	if (!master)
675 		return;
676 
677 	do {
678 		/* Wait for all other CPU's to gather here */
679 		if (!uv_nmi_wait_cpus(1))
680 			break;
681 
682 		/* If not all made it in, send IPI NMI to them */
683 		pr_alert("UV: Sending NMI IPI to %d CPUs: %*pbl\n",
684 			 cpumask_weight(uv_nmi_cpu_mask),
685 			 cpumask_pr_args(uv_nmi_cpu_mask));
686 
687 		uv_nmi_nr_cpus_ping();
688 
689 		/* If all CPU's are in, then done */
690 		if (!uv_nmi_wait_cpus(0))
691 			break;
692 
693 		pr_alert("UV: %d CPUs not in NMI loop: %*pbl\n",
694 			 cpumask_weight(uv_nmi_cpu_mask),
695 			 cpumask_pr_args(uv_nmi_cpu_mask));
696 	} while (0);
697 
698 	pr_alert("UV: %d of %d CPUs in NMI\n",
699 		atomic_read(&uv_nmi_cpus_in_nmi), num_online_cpus());
700 }
701 
702 /* Dump Instruction Pointer header */
703 static void uv_nmi_dump_cpu_ip_hdr(void)
704 {
705 	pr_info("\nUV: %4s %6s %-32s %s   (Note: PID 0 not listed)\n",
706 		"CPU", "PID", "COMMAND", "IP");
707 }
708 
709 /* Dump Instruction Pointer info */
710 static void uv_nmi_dump_cpu_ip(int cpu, struct pt_regs *regs)
711 {
712 	pr_info("UV: %4d %6d %-32.32s %pS",
713 		cpu, current->pid, current->comm, (void *)regs->ip);
714 }
715 
716 /*
717  * Dump this CPU's state.  If action was set to "kdump" and the crash_kexec
718  * failed, then we provide "dump" as an alternate action.  Action "dump" now
719  * also includes the show "ips" (instruction pointers) action whereas the
720  * action "ips" only displays instruction pointers for the non-idle CPU's.
721  * This is an abbreviated form of the "ps" command.
722  */
723 static void uv_nmi_dump_state_cpu(int cpu, struct pt_regs *regs)
724 {
725 	const char *dots = " ................................. ";
726 
727 	if (cpu == 0)
728 		uv_nmi_dump_cpu_ip_hdr();
729 
730 	if (current->pid != 0 || !uv_nmi_action_is("ips"))
731 		uv_nmi_dump_cpu_ip(cpu, regs);
732 
733 	if (uv_nmi_action_is("dump")) {
734 		pr_info("UV:%sNMI process trace for CPU %d\n", dots, cpu);
735 		show_regs(regs);
736 	}
737 
738 	this_cpu_write(uv_cpu_nmi.state, UV_NMI_STATE_DUMP_DONE);
739 }
740 
741 /* Trigger a slave CPU to dump it's state */
742 static void uv_nmi_trigger_dump(int cpu)
743 {
744 	int retry = uv_nmi_trigger_delay;
745 
746 	if (uv_cpu_nmi_per(cpu).state != UV_NMI_STATE_IN)
747 		return;
748 
749 	uv_cpu_nmi_per(cpu).state = UV_NMI_STATE_DUMP;
750 	do {
751 		cpu_relax();
752 		udelay(10);
753 		if (uv_cpu_nmi_per(cpu).state
754 				!= UV_NMI_STATE_DUMP)
755 			return;
756 	} while (--retry > 0);
757 
758 	pr_crit("UV: CPU %d stuck in process dump function\n", cpu);
759 	uv_cpu_nmi_per(cpu).state = UV_NMI_STATE_DUMP_DONE;
760 }
761 
762 /* Wait until all CPU's ready to exit */
763 static void uv_nmi_sync_exit(int master)
764 {
765 	atomic_dec(&uv_nmi_cpus_in_nmi);
766 	if (master) {
767 		while (atomic_read(&uv_nmi_cpus_in_nmi) > 0)
768 			cpu_relax();
769 		atomic_set(&uv_nmi_slave_continue, SLAVE_CLEAR);
770 	} else {
771 		while (atomic_read(&uv_nmi_slave_continue))
772 			cpu_relax();
773 	}
774 }
775 
776 /* Current "health" check is to check which CPU's are responsive */
777 static void uv_nmi_action_health(int cpu, struct pt_regs *regs, int master)
778 {
779 	if (master) {
780 		int in = atomic_read(&uv_nmi_cpus_in_nmi);
781 		int out = num_online_cpus() - in;
782 
783 		pr_alert("UV: NMI CPU health check (non-responding:%d)\n", out);
784 		atomic_set(&uv_nmi_slave_continue, SLAVE_EXIT);
785 	} else {
786 		while (!atomic_read(&uv_nmi_slave_continue))
787 			cpu_relax();
788 	}
789 	uv_nmi_sync_exit(master);
790 }
791 
792 /* Walk through CPU list and dump state of each */
793 static void uv_nmi_dump_state(int cpu, struct pt_regs *regs, int master)
794 {
795 	if (master) {
796 		int tcpu;
797 		int ignored = 0;
798 		int saved_console_loglevel = console_loglevel;
799 
800 		pr_alert("UV: tracing %s for %d CPUs from CPU %d\n",
801 			uv_nmi_action_is("ips") ? "IPs" : "processes",
802 			atomic_read(&uv_nmi_cpus_in_nmi), cpu);
803 
804 		console_loglevel = uv_nmi_loglevel;
805 		atomic_set(&uv_nmi_slave_continue, SLAVE_EXIT);
806 		for_each_online_cpu(tcpu) {
807 			if (cpumask_test_cpu(tcpu, uv_nmi_cpu_mask))
808 				ignored++;
809 			else if (tcpu == cpu)
810 				uv_nmi_dump_state_cpu(tcpu, regs);
811 			else
812 				uv_nmi_trigger_dump(tcpu);
813 		}
814 		if (ignored)
815 			pr_alert("UV: %d CPUs ignored NMI\n", ignored);
816 
817 		console_loglevel = saved_console_loglevel;
818 		pr_alert("UV: process trace complete\n");
819 	} else {
820 		while (!atomic_read(&uv_nmi_slave_continue))
821 			cpu_relax();
822 		while (this_cpu_read(uv_cpu_nmi.state) != UV_NMI_STATE_DUMP)
823 			cpu_relax();
824 		uv_nmi_dump_state_cpu(cpu, regs);
825 	}
826 	uv_nmi_sync_exit(master);
827 }
828 
829 static void uv_nmi_touch_watchdogs(void)
830 {
831 	touch_softlockup_watchdog_sync();
832 	clocksource_touch_watchdog();
833 	rcu_cpu_stall_reset();
834 	touch_nmi_watchdog();
835 }
836 
837 static void uv_nmi_kdump(int cpu, int main, struct pt_regs *regs)
838 {
839 	/* Check if kdump kernel loaded for both main and secondary CPUs */
840 	if (!kexec_crash_image) {
841 		if (main)
842 			pr_err("UV: NMI error: kdump kernel not loaded\n");
843 		return;
844 	}
845 
846 	/* Call crash to dump system state */
847 	if (main) {
848 		pr_emerg("UV: NMI executing crash_kexec on CPU%d\n", cpu);
849 		crash_kexec(regs);
850 
851 		pr_emerg("UV: crash_kexec unexpectedly returned\n");
852 		atomic_set(&uv_nmi_kexec_failed, 1);
853 
854 	} else { /* secondary */
855 
856 		/* If kdump kernel fails, secondaries will exit this loop */
857 		while (atomic_read(&uv_nmi_kexec_failed) == 0) {
858 
859 			/* Once shootdown cpus starts, they do not return */
860 			run_crash_ipi_callback(regs);
861 
862 			mdelay(10);
863 		}
864 	}
865 }
866 
867 #ifdef CONFIG_KGDB
868 #ifdef CONFIG_KGDB_KDB
869 static inline int uv_nmi_kdb_reason(void)
870 {
871 	return KDB_REASON_SYSTEM_NMI;
872 }
873 #else /* !CONFIG_KGDB_KDB */
874 static inline int uv_nmi_kdb_reason(void)
875 {
876 	/* Ensure user is expecting to attach gdb remote */
877 	if (uv_nmi_action_is("kgdb"))
878 		return 0;
879 
880 	pr_err("UV: NMI error: KDB is not enabled in this kernel\n");
881 	return -1;
882 }
883 #endif /* CONFIG_KGDB_KDB */
884 
885 /*
886  * Call KGDB/KDB from NMI handler
887  *
888  * Note that if both KGDB and KDB are configured, then the action of 'kgdb' or
889  * 'kdb' has no affect on which is used.  See the KGDB documentation for further
890  * information.
891  */
892 static void uv_call_kgdb_kdb(int cpu, struct pt_regs *regs, int master)
893 {
894 	if (master) {
895 		int reason = uv_nmi_kdb_reason();
896 		int ret;
897 
898 		if (reason < 0)
899 			return;
900 
901 		/* Call KGDB NMI handler as MASTER */
902 		ret = kgdb_nmicallin(cpu, X86_TRAP_NMI, regs, reason,
903 				&uv_nmi_slave_continue);
904 		if (ret) {
905 			pr_alert("KGDB returned error, is kgdboc set?\n");
906 			atomic_set(&uv_nmi_slave_continue, SLAVE_EXIT);
907 		}
908 	} else {
909 		/* Wait for KGDB signal that it's ready for slaves to enter */
910 		int sig;
911 
912 		do {
913 			cpu_relax();
914 			sig = atomic_read(&uv_nmi_slave_continue);
915 		} while (!sig);
916 
917 		/* Call KGDB as slave */
918 		if (sig == SLAVE_CONTINUE)
919 			kgdb_nmicallback(cpu, regs);
920 	}
921 	uv_nmi_sync_exit(master);
922 }
923 
924 #else /* !CONFIG_KGDB */
925 static inline void uv_call_kgdb_kdb(int cpu, struct pt_regs *regs, int master)
926 {
927 	pr_err("UV: NMI error: KGDB is not enabled in this kernel\n");
928 }
929 #endif /* !CONFIG_KGDB */
930 
931 /*
932  * UV NMI handler
933  */
934 static int uv_handle_nmi(unsigned int reason, struct pt_regs *regs)
935 {
936 	struct uv_hub_nmi_s *hub_nmi = uv_hub_nmi;
937 	int cpu = smp_processor_id();
938 	int master = 0;
939 	unsigned long flags;
940 
941 	local_irq_save(flags);
942 
943 	/* If not a UV System NMI, ignore */
944 	if (!this_cpu_read(uv_cpu_nmi.pinging) && !uv_check_nmi(hub_nmi)) {
945 		local_irq_restore(flags);
946 		return NMI_DONE;
947 	}
948 
949 	/* Indicate we are the first CPU into the NMI handler */
950 	master = (atomic_read(&uv_nmi_cpu) == cpu);
951 
952 	/* If NMI action is "kdump", then attempt to do it */
953 	if (uv_nmi_action_is("kdump")) {
954 		uv_nmi_kdump(cpu, master, regs);
955 
956 		/* Unexpected return, revert action to "dump" */
957 		if (master)
958 			strscpy(uv_nmi_action, "dump", sizeof(uv_nmi_action));
959 	}
960 
961 	/* Pause as all CPU's enter the NMI handler */
962 	uv_nmi_wait(master);
963 
964 	/* Process actions other than "kdump": */
965 	if (uv_nmi_action_is("health")) {
966 		uv_nmi_action_health(cpu, regs, master);
967 	} else if (uv_nmi_action_is("ips") || uv_nmi_action_is("dump")) {
968 		uv_nmi_dump_state(cpu, regs, master);
969 	} else if (uv_nmi_action_is("kdb") || uv_nmi_action_is("kgdb")) {
970 		uv_call_kgdb_kdb(cpu, regs, master);
971 	} else {
972 		if (master)
973 			pr_alert("UV: unknown NMI action: %s\n", uv_nmi_action);
974 		uv_nmi_sync_exit(master);
975 	}
976 
977 	/* Clear per_cpu "in_nmi" flag */
978 	this_cpu_write(uv_cpu_nmi.state, UV_NMI_STATE_OUT);
979 
980 	/* Clear MMR NMI flag on each hub */
981 	uv_clear_nmi(cpu);
982 
983 	/* Clear global flags */
984 	if (master) {
985 		if (!cpumask_empty(uv_nmi_cpu_mask))
986 			uv_nmi_cleanup_mask();
987 		atomic_set(&uv_nmi_cpus_in_nmi, -1);
988 		atomic_set(&uv_nmi_cpu, -1);
989 		atomic_set(&uv_in_nmi, 0);
990 		atomic_set(&uv_nmi_kexec_failed, 0);
991 		atomic_set(&uv_nmi_slave_continue, SLAVE_CLEAR);
992 	}
993 
994 	uv_nmi_touch_watchdogs();
995 	local_irq_restore(flags);
996 
997 	return NMI_HANDLED;
998 }
999 
1000 /*
1001  * NMI handler for pulling in CPU's when perf events are grabbing our NMI
1002  */
1003 static int uv_handle_nmi_ping(unsigned int reason, struct pt_regs *regs)
1004 {
1005 	int ret;
1006 
1007 	this_cpu_inc(uv_cpu_nmi.queries);
1008 	if (!this_cpu_read(uv_cpu_nmi.pinging)) {
1009 		local64_inc(&uv_nmi_ping_misses);
1010 		return NMI_DONE;
1011 	}
1012 
1013 	this_cpu_inc(uv_cpu_nmi.pings);
1014 	local64_inc(&uv_nmi_ping_count);
1015 	ret = uv_handle_nmi(reason, regs);
1016 	this_cpu_write(uv_cpu_nmi.pinging, 0);
1017 	return ret;
1018 }
1019 
1020 static void uv_register_nmi_notifier(void)
1021 {
1022 	if (register_nmi_handler(NMI_UNKNOWN, uv_handle_nmi, 0, "uv"))
1023 		pr_warn("UV: NMI handler failed to register\n");
1024 
1025 	if (register_nmi_handler(NMI_LOCAL, uv_handle_nmi_ping, 0, "uvping"))
1026 		pr_warn("UV: PING NMI handler failed to register\n");
1027 }
1028 
1029 void uv_nmi_init(void)
1030 {
1031 	unsigned int value;
1032 
1033 	/*
1034 	 * Unmask NMI on all CPU's
1035 	 */
1036 	value = apic_read(APIC_LVT1) | APIC_DM_NMI;
1037 	value &= ~APIC_LVT_MASKED;
1038 	apic_write(APIC_LVT1, value);
1039 }
1040 
1041 /* Setup HUB NMI info */
1042 static void __init uv_nmi_setup_common(bool hubbed)
1043 {
1044 	int size = sizeof(void *) * (1 << NODES_SHIFT);
1045 	int cpu;
1046 
1047 	uv_hub_nmi_list = kzalloc(size, GFP_KERNEL);
1048 	nmi_debug("UV: NMI hub list @ 0x%p (%d)\n", uv_hub_nmi_list, size);
1049 	BUG_ON(!uv_hub_nmi_list);
1050 	size = sizeof(struct uv_hub_nmi_s);
1051 	for_each_present_cpu(cpu) {
1052 		int nid = cpu_to_node(cpu);
1053 		if (uv_hub_nmi_list[nid] == NULL) {
1054 			uv_hub_nmi_list[nid] = kzalloc_node(size,
1055 							    GFP_KERNEL, nid);
1056 			BUG_ON(!uv_hub_nmi_list[nid]);
1057 			raw_spin_lock_init(&(uv_hub_nmi_list[nid]->nmi_lock));
1058 			atomic_set(&uv_hub_nmi_list[nid]->cpu_owner, -1);
1059 			uv_hub_nmi_list[nid]->hub_present = hubbed;
1060 			uv_hub_nmi_list[nid]->pch_owner = (nid == 0);
1061 		}
1062 		uv_hub_nmi_per(cpu) = uv_hub_nmi_list[nid];
1063 	}
1064 	BUG_ON(!alloc_cpumask_var(&uv_nmi_cpu_mask, GFP_KERNEL));
1065 }
1066 
1067 /* Setup for UV Hub systems */
1068 void __init uv_nmi_setup(void)
1069 {
1070 	uv_nmi_setup_mmrs();
1071 	uv_nmi_setup_common(true);
1072 	uv_register_nmi_notifier();
1073 	pr_info("UV: Hub NMI enabled\n");
1074 }
1075 
1076 /* Setup for UV Hubless systems */
1077 void __init uv_nmi_setup_hubless(void)
1078 {
1079 	uv_nmi_setup_common(false);
1080 	pch_base = xlate_dev_mem_ptr(PCH_PCR_GPIO_1_BASE);
1081 	nmi_debug("UV: PCH base:%p from 0x%lx, GPP_D_0\n",
1082 		pch_base, PCH_PCR_GPIO_1_BASE);
1083 	if (uv_pch_init_enable)
1084 		uv_init_hubless_pch_d0();
1085 	uv_init_hubless_pch_io(GPI_NMI_ENA_GPP_D_0,
1086 				STS_GPP_D_0_MASK, STS_GPP_D_0_MASK);
1087 	uv_nmi_setup_hubless_intr();
1088 	/* Ensure NMI enabled in Processor Interface Reg: */
1089 	uv_reassert_nmi();
1090 	uv_register_nmi_notifier();
1091 	pr_info("UV: PCH NMI enabled\n");
1092 }
1093