2 * SGI NMI support routines
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
18 * Copyright (c) 2009-2013 Silicon Graphics, Inc. All Rights Reserved.
19 * Copyright (c) Mike Travis
22 #include <linux/cpu.h>
23 #include <linux/delay.h>
24 #include <linux/kdb.h>
25 #include <linux/kexec.h>
26 #include <linux/kgdb.h>
27 #include <linux/module.h>
28 #include <linux/nmi.h>
29 #include <linux/sched.h>
30 #include <linux/slab.h>
33 #include <asm/current.h>
34 #include <asm/kdebug.h>
35 #include <asm/local64.h>
37 #include <asm/traps.h>
38 #include <asm/uv/uv.h>
39 #include <asm/uv/uv_hub.h>
40 #include <asm/uv/uv_mmrs.h>
45 * Handle system-wide NMI events generated by the global 'power nmi' command.
47 * Basic operation is to field the NMI interrupt on each cpu and wait
48 * until all cpus have arrived into the nmi handler. If some cpus do not
49 * make it into the handler, try and force them in with the IPI(NMI) signal.
51 * We also have to lessen UV Hub MMR accesses as much as possible as this
52 * disrupts the UV Hub's primary mission of directing NumaLink traffic and
53 * can cause system problems to occur.
55 * To do this we register our primary NMI notifier on the NMI_UNKNOWN
56 * chain. This reduces the number of false NMI calls when the perf
57 * tools are running which generate an enormous number of NMIs per
58 * second (~4M/s for 1024 cpu threads). Our secondary NMI handler is
59 * very short as it only checks that if it has been "pinged" with the
60 * IPI(NMI) signal as mentioned above, and does not read the UV Hub's MMR.
64 static struct uv_hub_nmi_s
**uv_hub_nmi_list
;
66 DEFINE_PER_CPU(struct uv_cpu_nmi_s
, __uv_cpu_nmi
);
67 EXPORT_PER_CPU_SYMBOL_GPL(__uv_cpu_nmi
);
69 static unsigned long nmi_mmr
;
70 static unsigned long nmi_mmr_clear
;
71 static unsigned long nmi_mmr_pending
;
73 static atomic_t uv_in_nmi
;
74 static atomic_t uv_nmi_cpu
= ATOMIC_INIT(-1);
75 static atomic_t uv_nmi_cpus_in_nmi
= ATOMIC_INIT(-1);
76 static atomic_t uv_nmi_slave_continue
;
77 static atomic_t uv_nmi_kexec_failed
;
78 static cpumask_var_t uv_nmi_cpu_mask
;
80 /* Values for uv_nmi_slave_continue */
82 #define SLAVE_CONTINUE 1
86 * Default is all stack dumps go to the console and buffer.
87 * Lower level to send to log buffer only.
89 static int uv_nmi_loglevel
= 7;
90 module_param_named(dump_loglevel
, uv_nmi_loglevel
, int, 0644);
93 * The following values show statistics on how perf events are affecting
96 static int param_get_local64(char *buffer
, const struct kernel_param
*kp
)
98 return sprintf(buffer
, "%lu\n", local64_read((local64_t
*)kp
->arg
));
101 static int param_set_local64(const char *val
, const struct kernel_param
*kp
)
103 /* clear on any write */
104 local64_set((local64_t
*)kp
->arg
, 0);
108 static struct kernel_param_ops param_ops_local64
= {
109 .get
= param_get_local64
,
110 .set
= param_set_local64
,
112 #define param_check_local64(name, p) __param_check(name, p, local64_t)
114 static local64_t uv_nmi_count
;
115 module_param_named(nmi_count
, uv_nmi_count
, local64
, 0644);
117 static local64_t uv_nmi_misses
;
118 module_param_named(nmi_misses
, uv_nmi_misses
, local64
, 0644);
120 static local64_t uv_nmi_ping_count
;
121 module_param_named(ping_count
, uv_nmi_ping_count
, local64
, 0644);
123 static local64_t uv_nmi_ping_misses
;
124 module_param_named(ping_misses
, uv_nmi_ping_misses
, local64
, 0644);
127 * Following values allow tuning for large systems under heavy loading
129 static int uv_nmi_initial_delay
= 100;
130 module_param_named(initial_delay
, uv_nmi_initial_delay
, int, 0644);
132 static int uv_nmi_slave_delay
= 100;
133 module_param_named(slave_delay
, uv_nmi_slave_delay
, int, 0644);
135 static int uv_nmi_loop_delay
= 100;
136 module_param_named(loop_delay
, uv_nmi_loop_delay
, int, 0644);
138 static int uv_nmi_trigger_delay
= 10000;
139 module_param_named(trigger_delay
, uv_nmi_trigger_delay
, int, 0644);
141 static int uv_nmi_wait_count
= 100;
142 module_param_named(wait_count
, uv_nmi_wait_count
, int, 0644);
144 static int uv_nmi_retry_count
= 500;
145 module_param_named(retry_count
, uv_nmi_retry_count
, int, 0644);
149 * "dump" - dump process stack for each cpu
150 * "ips" - dump IP info for each cpu
151 * "kdump" - do crash dump
152 * "kdb" - enter KDB/KGDB (default)
154 static char uv_nmi_action
[8] = "kdb";
155 module_param_string(action
, uv_nmi_action
, sizeof(uv_nmi_action
), 0644);
157 static inline bool uv_nmi_action_is(const char *action
)
159 return (strncmp(uv_nmi_action
, action
, strlen(action
)) == 0);
162 /* Setup which NMI support is present in system */
163 static void uv_nmi_setup_mmrs(void)
165 if (uv_read_local_mmr(UVH_NMI_MMRX_SUPPORTED
)) {
166 uv_write_local_mmr(UVH_NMI_MMRX_REQ
,
167 1UL << UVH_NMI_MMRX_REQ_SHIFT
);
168 nmi_mmr
= UVH_NMI_MMRX
;
169 nmi_mmr_clear
= UVH_NMI_MMRX_CLEAR
;
170 nmi_mmr_pending
= 1UL << UVH_NMI_MMRX_SHIFT
;
171 pr_info("UV: SMI NMI support: %s\n", UVH_NMI_MMRX_TYPE
);
173 nmi_mmr
= UVH_NMI_MMR
;
174 nmi_mmr_clear
= UVH_NMI_MMR_CLEAR
;
175 nmi_mmr_pending
= 1UL << UVH_NMI_MMR_SHIFT
;
176 pr_info("UV: SMI NMI support: %s\n", UVH_NMI_MMR_TYPE
);
180 /* Read NMI MMR and check if NMI flag was set by BMC. */
181 static inline int uv_nmi_test_mmr(struct uv_hub_nmi_s
*hub_nmi
)
183 hub_nmi
->nmi_value
= uv_read_local_mmr(nmi_mmr
);
184 atomic_inc(&hub_nmi
->read_mmr_count
);
185 return !!(hub_nmi
->nmi_value
& nmi_mmr_pending
);
188 static inline void uv_local_mmr_clear_nmi(void)
190 uv_write_local_mmr(nmi_mmr_clear
, nmi_mmr_pending
);
194 * If first cpu in on this hub, set hub_nmi "in_nmi" and "owner" values and
195 * return true. If first cpu in on the system, set global "in_nmi" flag.
197 static int uv_set_in_nmi(int cpu
, struct uv_hub_nmi_s
*hub_nmi
)
199 int first
= atomic_add_unless(&hub_nmi
->in_nmi
, 1, 1);
202 atomic_set(&hub_nmi
->cpu_owner
, cpu
);
203 if (atomic_add_unless(&uv_in_nmi
, 1, 1))
204 atomic_set(&uv_nmi_cpu
, cpu
);
206 atomic_inc(&hub_nmi
->nmi_count
);
211 /* Check if this is a system NMI event */
212 static int uv_check_nmi(struct uv_hub_nmi_s
*hub_nmi
)
214 int cpu
= smp_processor_id();
217 local64_inc(&uv_nmi_count
);
218 uv_cpu_nmi
.queries
++;
221 nmi
= atomic_read(&hub_nmi
->in_nmi
);
225 if (raw_spin_trylock(&hub_nmi
->nmi_lock
)) {
227 /* check hub MMR NMI flag */
228 if (uv_nmi_test_mmr(hub_nmi
)) {
229 uv_set_in_nmi(cpu
, hub_nmi
);
234 /* MMR NMI flag is clear */
235 raw_spin_unlock(&hub_nmi
->nmi_lock
);
238 /* wait a moment for the hub nmi locker to set flag */
240 udelay(uv_nmi_slave_delay
);
242 /* re-check hub in_nmi flag */
243 nmi
= atomic_read(&hub_nmi
->in_nmi
);
248 /* check if this BMC missed setting the MMR NMI flag */
250 nmi
= atomic_read(&uv_in_nmi
);
252 uv_set_in_nmi(cpu
, hub_nmi
);
258 local64_inc(&uv_nmi_misses
);
263 /* Need to reset the NMI MMR register, but only once per hub. */
264 static inline void uv_clear_nmi(int cpu
)
266 struct uv_hub_nmi_s
*hub_nmi
= uv_hub_nmi
;
268 if (cpu
== atomic_read(&hub_nmi
->cpu_owner
)) {
269 atomic_set(&hub_nmi
->cpu_owner
, -1);
270 atomic_set(&hub_nmi
->in_nmi
, 0);
271 uv_local_mmr_clear_nmi();
272 raw_spin_unlock(&hub_nmi
->nmi_lock
);
276 /* Print non-responding cpus */
277 static void uv_nmi_nr_cpus_pr(char *fmt
)
279 static char cpu_list
[1024];
280 int len
= sizeof(cpu_list
);
281 int c
= cpumask_weight(uv_nmi_cpu_mask
);
282 int n
= cpulist_scnprintf(cpu_list
, len
, uv_nmi_cpu_mask
);
285 strcpy(&cpu_list
[len
- 6], "...\n");
287 printk(fmt
, c
, cpu_list
);
290 /* Ping non-responding cpus attemping to force them into the NMI handler */
291 static void uv_nmi_nr_cpus_ping(void)
295 for_each_cpu(cpu
, uv_nmi_cpu_mask
)
296 atomic_set(&uv_cpu_nmi_per(cpu
).pinging
, 1);
298 apic
->send_IPI_mask(uv_nmi_cpu_mask
, APIC_DM_NMI
);
301 /* Clean up flags for cpus that ignored both NMI and ping */
302 static void uv_nmi_cleanup_mask(void)
306 for_each_cpu(cpu
, uv_nmi_cpu_mask
) {
307 atomic_set(&uv_cpu_nmi_per(cpu
).pinging
, 0);
308 atomic_set(&uv_cpu_nmi_per(cpu
).state
, UV_NMI_STATE_OUT
);
309 cpumask_clear_cpu(cpu
, uv_nmi_cpu_mask
);
313 /* Loop waiting as cpus enter nmi handler */
314 static int uv_nmi_wait_cpus(int first
)
316 int i
, j
, k
, n
= num_online_cpus();
317 int last_k
= 0, waiting
= 0;
320 cpumask_copy(uv_nmi_cpu_mask
, cpu_online_mask
);
323 k
= n
- cpumask_weight(uv_nmi_cpu_mask
);
326 udelay(uv_nmi_initial_delay
);
327 for (i
= 0; i
< uv_nmi_retry_count
; i
++) {
328 int loop_delay
= uv_nmi_loop_delay
;
330 for_each_cpu(j
, uv_nmi_cpu_mask
) {
331 if (atomic_read(&uv_cpu_nmi_per(j
).state
)) {
332 cpumask_clear_cpu(j
, uv_nmi_cpu_mask
);
337 if (k
>= n
) { /* all in? */
341 if (last_k
!= k
) { /* abort if no new cpus coming in */
344 } else if (++waiting
> uv_nmi_wait_count
)
347 /* extend delay if waiting only for cpu 0 */
348 if (waiting
&& (n
- k
) == 1 &&
349 cpumask_test_cpu(0, uv_nmi_cpu_mask
))
354 atomic_set(&uv_nmi_cpus_in_nmi
, k
);
358 /* Wait until all slave cpus have entered UV NMI handler */
359 static void uv_nmi_wait(int master
)
361 /* indicate this cpu is in */
362 atomic_set(&uv_cpu_nmi
.state
, UV_NMI_STATE_IN
);
364 /* if not the first cpu in (the master), then we are a slave cpu */
369 /* wait for all other cpus to gather here */
370 if (!uv_nmi_wait_cpus(1))
373 /* if not all made it in, send IPI NMI to them */
374 uv_nmi_nr_cpus_pr(KERN_ALERT
375 "UV: Sending NMI IPI to %d non-responding CPUs: %s\n");
376 uv_nmi_nr_cpus_ping();
378 /* if all cpus are in, then done */
379 if (!uv_nmi_wait_cpus(0))
382 uv_nmi_nr_cpus_pr(KERN_ALERT
383 "UV: %d CPUs not in NMI loop: %s\n");
386 pr_alert("UV: %d of %d CPUs in NMI\n",
387 atomic_read(&uv_nmi_cpus_in_nmi
), num_online_cpus());
390 static void uv_nmi_dump_cpu_ip_hdr(void)
393 "\nUV: %4s %6s %-32s %s (Note: PID 0 not listed)\n",
394 "CPU", "PID", "COMMAND", "IP");
397 static void uv_nmi_dump_cpu_ip(int cpu
, struct pt_regs
*regs
)
399 printk(KERN_DEFAULT
"UV: %4d %6d %-32.32s ",
400 cpu
, current
->pid
, current
->comm
);
402 printk_address(regs
->ip
);
405 /* Dump this cpu's state */
406 static void uv_nmi_dump_state_cpu(int cpu
, struct pt_regs
*regs
)
408 const char *dots
= " ................................. ";
410 if (uv_nmi_action_is("ips")) {
412 uv_nmi_dump_cpu_ip_hdr();
414 if (current
->pid
!= 0)
415 uv_nmi_dump_cpu_ip(cpu
, regs
);
417 } else if (uv_nmi_action_is("dump")) {
419 "UV:%sNMI process trace for CPU %d\n", dots
, cpu
);
422 atomic_set(&uv_cpu_nmi
.state
, UV_NMI_STATE_DUMP_DONE
);
425 /* Trigger a slave cpu to dump it's state */
426 static void uv_nmi_trigger_dump(int cpu
)
428 int retry
= uv_nmi_trigger_delay
;
430 if (atomic_read(&uv_cpu_nmi_per(cpu
).state
) != UV_NMI_STATE_IN
)
433 atomic_set(&uv_cpu_nmi_per(cpu
).state
, UV_NMI_STATE_DUMP
);
437 if (atomic_read(&uv_cpu_nmi_per(cpu
).state
)
438 != UV_NMI_STATE_DUMP
)
440 } while (--retry
> 0);
442 pr_crit("UV: CPU %d stuck in process dump function\n", cpu
);
443 atomic_set(&uv_cpu_nmi_per(cpu
).state
, UV_NMI_STATE_DUMP_DONE
);
446 /* Wait until all cpus ready to exit */
447 static void uv_nmi_sync_exit(int master
)
449 atomic_dec(&uv_nmi_cpus_in_nmi
);
451 while (atomic_read(&uv_nmi_cpus_in_nmi
) > 0)
453 atomic_set(&uv_nmi_slave_continue
, SLAVE_CLEAR
);
455 while (atomic_read(&uv_nmi_slave_continue
))
460 /* Walk through cpu list and dump state of each */
461 static void uv_nmi_dump_state(int cpu
, struct pt_regs
*regs
, int master
)
466 int saved_console_loglevel
= console_loglevel
;
468 pr_alert("UV: tracing %s for %d CPUs from CPU %d\n",
469 uv_nmi_action_is("ips") ? "IPs" : "processes",
470 atomic_read(&uv_nmi_cpus_in_nmi
), cpu
);
472 console_loglevel
= uv_nmi_loglevel
;
473 atomic_set(&uv_nmi_slave_continue
, SLAVE_EXIT
);
474 for_each_online_cpu(tcpu
) {
475 if (cpumask_test_cpu(tcpu
, uv_nmi_cpu_mask
))
477 else if (tcpu
== cpu
)
478 uv_nmi_dump_state_cpu(tcpu
, regs
);
480 uv_nmi_trigger_dump(tcpu
);
483 printk(KERN_DEFAULT
"UV: %d CPUs ignored NMI\n",
486 console_loglevel
= saved_console_loglevel
;
487 pr_alert("UV: process trace complete\n");
489 while (!atomic_read(&uv_nmi_slave_continue
))
491 while (atomic_read(&uv_cpu_nmi
.state
) != UV_NMI_STATE_DUMP
)
493 uv_nmi_dump_state_cpu(cpu
, regs
);
495 uv_nmi_sync_exit(master
);
498 static void uv_nmi_touch_watchdogs(void)
500 touch_softlockup_watchdog_sync();
501 clocksource_touch_watchdog();
502 rcu_cpu_stall_reset();
503 touch_nmi_watchdog();
506 #if defined(CONFIG_KEXEC)
507 static void uv_nmi_kdump(int cpu
, int master
, struct pt_regs
*regs
)
509 /* Call crash to dump system state */
511 pr_emerg("UV: NMI executing crash_kexec on CPU%d\n", cpu
);
514 pr_emerg("UV: crash_kexec unexpectedly returned, ");
515 if (!kexec_crash_image
) {
516 pr_cont("crash kernel not loaded\n");
517 atomic_set(&uv_nmi_kexec_failed
, 1);
521 pr_cont("kexec busy, stalling cpus while waiting\n");
524 /* If crash exec fails the slaves should return, otherwise stall */
525 while (atomic_read(&uv_nmi_kexec_failed
) == 0)
528 /* Crash kernel most likely not loaded, return in an orderly fashion */
532 #else /* !CONFIG_KEXEC */
533 static inline void uv_nmi_kdump(int cpu
, int master
, struct pt_regs
*regs
)
536 pr_err("UV: NMI kdump: KEXEC not supported in this kernel\n");
538 #endif /* !CONFIG_KEXEC */
540 #ifdef CONFIG_KGDB_KDB
541 /* Call KDB from NMI handler */
542 static void uv_call_kdb(int cpu
, struct pt_regs
*regs
, int master
)
547 /* call KGDB NMI handler as MASTER */
548 ret
= kgdb_nmicallin(cpu
, X86_TRAP_NMI
, regs
,
549 &uv_nmi_slave_continue
);
551 pr_alert("KDB returned error, is kgdboc set?\n");
552 atomic_set(&uv_nmi_slave_continue
, SLAVE_EXIT
);
555 /* wait for KGDB signal that it's ready for slaves to enter */
560 sig
= atomic_read(&uv_nmi_slave_continue
);
563 /* call KGDB as slave */
564 if (sig
== SLAVE_CONTINUE
)
565 kgdb_nmicallback(cpu
, regs
);
567 uv_nmi_sync_exit(master
);
570 #else /* !CONFIG_KGDB_KDB */
571 static inline void uv_call_kdb(int cpu
, struct pt_regs
*regs
, int master
)
573 pr_err("UV: NMI error: KGDB/KDB is not enabled in this kernel\n");
575 #endif /* !CONFIG_KGDB_KDB */
580 int uv_handle_nmi(unsigned int reason
, struct pt_regs
*regs
)
582 struct uv_hub_nmi_s
*hub_nmi
= uv_hub_nmi
;
583 int cpu
= smp_processor_id();
587 local_irq_save(flags
);
589 /* If not a UV System NMI, ignore */
590 if (!atomic_read(&uv_cpu_nmi
.pinging
) && !uv_check_nmi(hub_nmi
)) {
591 local_irq_restore(flags
);
595 /* Indicate we are the first CPU into the NMI handler */
596 master
= (atomic_read(&uv_nmi_cpu
) == cpu
);
598 /* If NMI action is "kdump", then attempt to do it */
599 if (uv_nmi_action_is("kdump"))
600 uv_nmi_kdump(cpu
, master
, regs
);
602 /* Pause as all cpus enter the NMI handler */
605 /* Dump state of each cpu */
606 if (uv_nmi_action_is("ips") || uv_nmi_action_is("dump"))
607 uv_nmi_dump_state(cpu
, regs
, master
);
609 /* Call KDB if enabled */
610 else if (uv_nmi_action_is("kdb"))
611 uv_call_kdb(cpu
, regs
, master
);
613 /* Clear per_cpu "in nmi" flag */
614 atomic_set(&uv_cpu_nmi
.state
, UV_NMI_STATE_OUT
);
616 /* Clear MMR NMI flag on each hub */
619 /* Clear global flags */
621 if (cpumask_weight(uv_nmi_cpu_mask
))
622 uv_nmi_cleanup_mask();
623 atomic_set(&uv_nmi_cpus_in_nmi
, -1);
624 atomic_set(&uv_nmi_cpu
, -1);
625 atomic_set(&uv_in_nmi
, 0);
628 uv_nmi_touch_watchdogs();
629 local_irq_restore(flags
);
635 * NMI handler for pulling in CPUs when perf events are grabbing our NMI
637 int uv_handle_nmi_ping(unsigned int reason
, struct pt_regs
*regs
)
641 uv_cpu_nmi
.queries
++;
642 if (!atomic_read(&uv_cpu_nmi
.pinging
)) {
643 local64_inc(&uv_nmi_ping_misses
);
648 local64_inc(&uv_nmi_ping_count
);
649 ret
= uv_handle_nmi(reason
, regs
);
650 atomic_set(&uv_cpu_nmi
.pinging
, 0);
654 void uv_register_nmi_notifier(void)
656 if (register_nmi_handler(NMI_UNKNOWN
, uv_handle_nmi
, 0, "uv"))
657 pr_warn("UV: NMI handler failed to register\n");
659 if (register_nmi_handler(NMI_LOCAL
, uv_handle_nmi_ping
, 0, "uvping"))
660 pr_warn("UV: PING NMI handler failed to register\n");
663 void uv_nmi_init(void)
668 * Unmask NMI on all cpus
670 value
= apic_read(APIC_LVT1
) | APIC_DM_NMI
;
671 value
&= ~APIC_LVT_MASKED
;
672 apic_write(APIC_LVT1
, value
);
675 void uv_nmi_setup(void)
677 int size
= sizeof(void *) * (1 << NODES_SHIFT
);
680 /* Setup hub nmi info */
682 uv_hub_nmi_list
= kzalloc(size
, GFP_KERNEL
);
683 pr_info("UV: NMI hub list @ 0x%p (%d)\n", uv_hub_nmi_list
, size
);
684 BUG_ON(!uv_hub_nmi_list
);
685 size
= sizeof(struct uv_hub_nmi_s
);
686 for_each_present_cpu(cpu
) {
687 nid
= cpu_to_node(cpu
);
688 if (uv_hub_nmi_list
[nid
] == NULL
) {
689 uv_hub_nmi_list
[nid
] = kzalloc_node(size
,
691 BUG_ON(!uv_hub_nmi_list
[nid
]);
692 raw_spin_lock_init(&(uv_hub_nmi_list
[nid
]->nmi_lock
));
693 atomic_set(&uv_hub_nmi_list
[nid
]->cpu_owner
, -1);
695 uv_hub_nmi_per(cpu
) = uv_hub_nmi_list
[nid
];
697 BUG_ON(!alloc_cpumask_var(&uv_nmi_cpu_mask
, GFP_KERNEL
));