aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
Diffstat
-rw-r--r--arch/i386/kernel/cpu/Makefile2
-rw-r--r--arch/i386/kernel/cpu/perfctr-watchdog.c658
-rw-r--r--arch/i386/kernel/nmi.c829
-rw-r--r--include/asm-i386/nmi.h8
4 files changed, 721 insertions, 776 deletions
diff --git a/arch/i386/kernel/cpu/Makefile b/arch/i386/kernel/cpu/Makefile
index 5fb1a7560438..74f27a463db0 100644
--- a/arch/i386/kernel/cpu/Makefile
+++ b/arch/i386/kernel/cpu/Makefile
@@ -17,3 +17,5 @@ obj-$(CONFIG_X86_MCE) += mcheck/
17 17
18obj-$(CONFIG_MTRR) += mtrr/ 18obj-$(CONFIG_MTRR) += mtrr/
19obj-$(CONFIG_CPU_FREQ) += cpufreq/ 19obj-$(CONFIG_CPU_FREQ) += cpufreq/
20
21obj-$(CONFIG_X86_LOCAL_APIC) += perfctr-watchdog.o
diff --git a/arch/i386/kernel/cpu/perfctr-watchdog.c b/arch/i386/kernel/cpu/perfctr-watchdog.c
new file mode 100644
index 000000000000..2b04c8f1db62
--- /dev/null
+++ b/arch/i386/kernel/cpu/perfctr-watchdog.c
@@ -0,0 +1,658 @@
1/* local apic based NMI watchdog for various CPUs.
2 This file also handles reservation of performance counters for coordination
3 with other users (like oprofile).
4
5 Note that these events normally don't tick when the CPU idles. This means
6 the frequency varies with CPU load.
7
8 Original code for K7/P6 written by Keith Owens */
9
10#include <linux/percpu.h>
11#include <linux/module.h>
12#include <linux/kernel.h>
13#include <linux/bitops.h>
14#include <linux/smp.h>
15#include <linux/nmi.h>
16#include <asm/apic.h>
17#include <asm/intel_arch_perfmon.h>
18
19struct nmi_watchdog_ctlblk {
20 unsigned int cccr_msr;
21 unsigned int perfctr_msr; /* the MSR to reset in NMI handler */
22 unsigned int evntsel_msr; /* the MSR to select the events to handle */
23};
24
25/* Interface defining a CPU specific perfctr watchdog */
26struct wd_ops {
27 int (*reserve)(void);
28 void (*unreserve)(void);
29 int (*setup)(unsigned nmi_hz);
30 void (*rearm)(struct nmi_watchdog_ctlblk *wd, unsigned nmi_hz);
31 void (*stop)(void *);
32 unsigned perfctr;
33 unsigned evntsel;
34 u64 checkbit;
35};
36
37static struct wd_ops *wd_ops;
38
39/* this number is calculated from Intel's MSR_P4_CRU_ESCR5 register and it's
40 * offset from MSR_P4_BSU_ESCR0. It will be the max for all platforms (for now)
41 */
42#define NMI_MAX_COUNTER_BITS 66
43
44/* perfctr_nmi_owner tracks the ownership of the perfctr registers:
45 * evtsel_nmi_owner tracks the ownership of the event selection
46 * - different performance counters/ event selection may be reserved for
47 * different subsystems this reservation system just tries to coordinate
48 * things a little
49 */
50static DECLARE_BITMAP(perfctr_nmi_owner, NMI_MAX_COUNTER_BITS);
51static DECLARE_BITMAP(evntsel_nmi_owner, NMI_MAX_COUNTER_BITS);
52
53static DEFINE_PER_CPU(struct nmi_watchdog_ctlblk, nmi_watchdog_ctlblk);
54
55/* converts an msr to an appropriate reservation bit */
56static inline unsigned int nmi_perfctr_msr_to_bit(unsigned int msr)
57{
58 return wd_ops ? msr - wd_ops->perfctr : 0;
59}
60
61/* converts an msr to an appropriate reservation bit */
62/* returns the bit offset of the event selection register */
63static inline unsigned int nmi_evntsel_msr_to_bit(unsigned int msr)
64{
65 return wd_ops ? msr - wd_ops->evntsel : 0;
66}
67
68/* checks for a bit availability (hack for oprofile) */
69int avail_to_resrv_perfctr_nmi_bit(unsigned int counter)
70{
71 BUG_ON(counter > NMI_MAX_COUNTER_BITS);
72
73 return (!test_bit(counter, perfctr_nmi_owner));
74}
75
76/* checks the an msr for availability */
77int avail_to_resrv_perfctr_nmi(unsigned int msr)
78{
79 unsigned int counter;
80
81 counter = nmi_perfctr_msr_to_bit(msr);
82 BUG_ON(counter > NMI_MAX_COUNTER_BITS);
83
84 return (!test_bit(counter, perfctr_nmi_owner));
85}
86
87int reserve_perfctr_nmi(unsigned int msr)
88{
89 unsigned int counter;
90
91 counter = nmi_perfctr_msr_to_bit(msr);
92 BUG_ON(counter > NMI_MAX_COUNTER_BITS);
93
94 if (!test_and_set_bit(counter, perfctr_nmi_owner))
95 return 1;
96 return 0;
97}
98
99void release_perfctr_nmi(unsigned int msr)
100{
101 unsigned int counter;
102
103 counter = nmi_perfctr_msr_to_bit(msr);
104 BUG_ON(counter > NMI_MAX_COUNTER_BITS);
105
106 clear_bit(counter, perfctr_nmi_owner);
107}
108
109int reserve_evntsel_nmi(unsigned int msr)
110{
111 unsigned int counter;
112
113 counter = nmi_evntsel_msr_to_bit(msr);
114 BUG_ON(counter > NMI_MAX_COUNTER_BITS);
115
116 if (!test_and_set_bit(counter, evntsel_nmi_owner))
117 return 1;
118 return 0;
119}
120
121void release_evntsel_nmi(unsigned int msr)
122{
123 unsigned int counter;
124
125 counter = nmi_evntsel_msr_to_bit(msr);
126 BUG_ON(counter > NMI_MAX_COUNTER_BITS);
127
128 clear_bit(counter, evntsel_nmi_owner);
129}
130
131EXPORT_SYMBOL(avail_to_resrv_perfctr_nmi);
132EXPORT_SYMBOL(avail_to_resrv_perfctr_nmi_bit);
133EXPORT_SYMBOL(reserve_perfctr_nmi);
134EXPORT_SYMBOL(release_perfctr_nmi);
135EXPORT_SYMBOL(reserve_evntsel_nmi);
136EXPORT_SYMBOL(release_evntsel_nmi);
137
138void disable_lapic_nmi_watchdog(void)
139{
140 BUG_ON(nmi_watchdog != NMI_LOCAL_APIC);
141
142 if (atomic_read(&nmi_active) <= 0)
143 return;
144
145 on_each_cpu(wd_ops->stop, NULL, 0, 1);
146 wd_ops->unreserve();
147
148 BUG_ON(atomic_read(&nmi_active) != 0);
149}
150
151void enable_lapic_nmi_watchdog(void)
152{
153 BUG_ON(nmi_watchdog != NMI_LOCAL_APIC);
154
155 /* are we already enabled */
156 if (atomic_read(&nmi_active) != 0)
157 return;
158
159 /* are we lapic aware */
160 if (!wd_ops)
161 return;
162 if (!wd_ops->reserve()) {
163 printk(KERN_ERR "NMI watchdog: cannot reserve perfctrs\n");
164 return;
165 }
166
167 on_each_cpu(setup_apic_nmi_watchdog, NULL, 0, 1);
168 touch_nmi_watchdog();
169}
170
171/*
172 * Activate the NMI watchdog via the local APIC.
173 */
174
175static unsigned int adjust_for_32bit_ctr(unsigned int hz)
176{
177 u64 counter_val;
178 unsigned int retval = hz;
179
180 /*
181 * On Intel CPUs with P6/ARCH_PERFMON only 32 bits in the counter
182 * are writable, with higher bits sign extending from bit 31.
183 * So, we can only program the counter with 31 bit values and
184 * 32nd bit should be 1, for 33.. to be 1.
185 * Find the appropriate nmi_hz
186 */
187 counter_val = (u64)cpu_khz * 1000;
188 do_div(counter_val, retval);
189 if (counter_val > 0x7fffffffULL) {
190 u64 count = (u64)cpu_khz * 1000;
191 do_div(count, 0x7fffffffUL);
192 retval = count + 1;
193 }
194 return retval;
195}
196
197static void
198write_watchdog_counter(unsigned int perfctr_msr, const char *descr, unsigned nmi_hz)
199{
200 u64 count = (u64)cpu_khz * 1000;
201
202 do_div(count, nmi_hz);
203 if(descr)
204 Dprintk("setting %s to -0x%08Lx\n", descr, count);
205 wrmsrl(perfctr_msr, 0 - count);
206}
207
208static void write_watchdog_counter32(unsigned int perfctr_msr,
209 const char *descr, unsigned nmi_hz)
210{
211 u64 count = (u64)cpu_khz * 1000;
212
213 do_div(count, nmi_hz);
214 if(descr)
215 Dprintk("setting %s to -0x%08Lx\n", descr, count);
216 wrmsr(perfctr_msr, (u32)(-count), 0);
217}
218
219/* AMD K7/K8/Family10h/Family11h support. AMD keeps this interface
220 nicely stable so there is not much variety */
221
222#define K7_EVNTSEL_ENABLE (1 << 22)
223#define K7_EVNTSEL_INT (1 << 20)
224#define K7_EVNTSEL_OS (1 << 17)
225#define K7_EVNTSEL_USR (1 << 16)
226#define K7_EVENT_CYCLES_PROCESSOR_IS_RUNNING 0x76
227#define K7_NMI_EVENT K7_EVENT_CYCLES_PROCESSOR_IS_RUNNING
228
229static int setup_k7_watchdog(unsigned nmi_hz)
230{
231 unsigned int perfctr_msr, evntsel_msr;
232 unsigned int evntsel;
233 struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
234
235 perfctr_msr = MSR_K7_PERFCTR0;
236 evntsel_msr = MSR_K7_EVNTSEL0;
237
238 wrmsrl(perfctr_msr, 0UL);
239
240 evntsel = K7_EVNTSEL_INT
241 | K7_EVNTSEL_OS
242 | K7_EVNTSEL_USR
243 | K7_NMI_EVENT;
244
245 /* setup the timer */
246 wrmsr(evntsel_msr, evntsel, 0);
247 write_watchdog_counter(perfctr_msr, "K7_PERFCTR0",nmi_hz);
248 apic_write(APIC_LVTPC, APIC_DM_NMI);
249 evntsel |= K7_EVNTSEL_ENABLE;
250 wrmsr(evntsel_msr, evntsel, 0);
251
252 wd->perfctr_msr = perfctr_msr;
253 wd->evntsel_msr = evntsel_msr;
254 wd->cccr_msr = 0; //unused
255 return 1;
256}
257
258static void single_msr_stop_watchdog(void *arg)
259{
260 struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
261
262 wrmsr(wd->evntsel_msr, 0, 0);
263}
264
265static int single_msr_reserve(void)
266{
267 if (!reserve_perfctr_nmi(wd_ops->perfctr))
268 return 0;
269
270 if (!reserve_evntsel_nmi(wd_ops->evntsel)) {
271 release_perfctr_nmi(wd_ops->perfctr);
272 return 0;
273 }
274 return 1;
275}
276
277static void single_msr_unreserve(void)
278{
279 release_evntsel_nmi(wd_ops->perfctr);
280 release_perfctr_nmi(wd_ops->evntsel);
281}
282
283static void single_msr_rearm(struct nmi_watchdog_ctlblk *wd, unsigned nmi_hz)
284{
285 /* start the cycle over again */
286 write_watchdog_counter(wd->perfctr_msr, NULL, nmi_hz);
287}
288
289static struct wd_ops k7_wd_ops = {
290 .reserve = single_msr_reserve,
291 .unreserve = single_msr_unreserve,
292 .setup = setup_k7_watchdog,
293 .rearm = single_msr_rearm,
294 .stop = single_msr_stop_watchdog,
295 .perfctr = MSR_K7_PERFCTR0,
296 .evntsel = MSR_K7_EVNTSEL0,
297 .checkbit = 1ULL<<63,
298};
299
300/* Intel Model 6 (PPro+,P2,P3,P-M,Core1) */
301
302#define P6_EVNTSEL0_ENABLE (1 << 22)
303#define P6_EVNTSEL_INT (1 << 20)
304#define P6_EVNTSEL_OS (1 << 17)
305#define P6_EVNTSEL_USR (1 << 16)
306#define P6_EVENT_CPU_CLOCKS_NOT_HALTED 0x79
307#define P6_NMI_EVENT P6_EVENT_CPU_CLOCKS_NOT_HALTED
308
309static int setup_p6_watchdog(unsigned nmi_hz)
310{
311 unsigned int perfctr_msr, evntsel_msr;
312 unsigned int evntsel;
313 struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
314
315 perfctr_msr = MSR_P6_PERFCTR0;
316 evntsel_msr = MSR_P6_EVNTSEL0;
317
318 wrmsrl(perfctr_msr, 0UL);
319
320 evntsel = P6_EVNTSEL_INT
321 | P6_EVNTSEL_OS
322 | P6_EVNTSEL_USR
323 | P6_NMI_EVENT;
324
325 /* setup the timer */
326 wrmsr(evntsel_msr, evntsel, 0);
327 nmi_hz = adjust_for_32bit_ctr(nmi_hz);
328 write_watchdog_counter32(perfctr_msr, "P6_PERFCTR0",nmi_hz);
329 apic_write(APIC_LVTPC, APIC_DM_NMI);
330 evntsel |= P6_EVNTSEL0_ENABLE;
331 wrmsr(evntsel_msr, evntsel, 0);
332
333 wd->perfctr_msr = perfctr_msr;
334 wd->evntsel_msr = evntsel_msr;
335 wd->cccr_msr = 0; //unused
336 return 1;
337}
338
339static void p6_rearm(struct nmi_watchdog_ctlblk *wd, unsigned nmi_hz)
340{
341 /* P6 based Pentium M need to re-unmask
342 * the apic vector but it doesn't hurt
343 * other P6 variant.
344 * ArchPerfom/Core Duo also needs this */
345 apic_write(APIC_LVTPC, APIC_DM_NMI);
346 /* P6/ARCH_PERFMON has 32 bit counter write */
347 write_watchdog_counter32(wd->perfctr_msr, NULL,nmi_hz);
348}
349
350static struct wd_ops p6_wd_ops = {
351 .reserve = single_msr_reserve,
352 .unreserve = single_msr_unreserve,
353 .setup = setup_p6_watchdog,
354 .rearm = p6_rearm,
355 .stop = single_msr_stop_watchdog,
356 .perfctr = MSR_P6_PERFCTR0,
357 .evntsel = MSR_P6_EVNTSEL0,
358 .checkbit = 1ULL<<39,
359};
360
361/* Intel P4 performance counters. By far the most complicated of all. */
362
363#define MSR_P4_MISC_ENABLE_PERF_AVAIL (1<<7)
364#define P4_ESCR_EVENT_SELECT(N) ((N)<<25)
365#define P4_ESCR_OS (1<<3)
366#define P4_ESCR_USR (1<<2)
367#define P4_CCCR_OVF_PMI0 (1<<26)
368#define P4_CCCR_OVF_PMI1 (1<<27)
369#define P4_CCCR_THRESHOLD(N) ((N)<<20)
370#define P4_CCCR_COMPLEMENT (1<<19)
371#define P4_CCCR_COMPARE (1<<18)
372#define P4_CCCR_REQUIRED (3<<16)
373#define P4_CCCR_ESCR_SELECT(N) ((N)<<13)
374#define P4_CCCR_ENABLE (1<<12)
375#define P4_CCCR_OVF (1<<31)
376
377/* Set up IQ_COUNTER0 to behave like a clock, by having IQ_CCCR0 filter
378 CRU_ESCR0 (with any non-null event selector) through a complemented
379 max threshold. [IA32-Vol3, Section 14.9.9] */
380
381static int setup_p4_watchdog(unsigned nmi_hz)
382{
383 unsigned int perfctr_msr, evntsel_msr, cccr_msr;
384 unsigned int evntsel, cccr_val;
385 unsigned int misc_enable, dummy;
386 unsigned int ht_num;
387 struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
388
389 rdmsr(MSR_IA32_MISC_ENABLE, misc_enable, dummy);
390 if (!(misc_enable & MSR_P4_MISC_ENABLE_PERF_AVAIL))
391 return 0;
392
393#ifdef CONFIG_SMP
394 /* detect which hyperthread we are on */
395 if (smp_num_siblings == 2) {
396 unsigned int ebx, apicid;
397
398 ebx = cpuid_ebx(1);
399 apicid = (ebx >> 24) & 0xff;
400 ht_num = apicid & 1;
401 } else
402#endif
403 ht_num = 0;
404
405 /* performance counters are shared resources
406 * assign each hyperthread its own set
407 * (re-use the ESCR0 register, seems safe
408 * and keeps the cccr_val the same)
409 */
410 if (!ht_num) {
411 /* logical cpu 0 */
412 perfctr_msr = MSR_P4_IQ_PERFCTR0;
413 evntsel_msr = MSR_P4_CRU_ESCR0;
414 cccr_msr = MSR_P4_IQ_CCCR0;
415 cccr_val = P4_CCCR_OVF_PMI0 | P4_CCCR_ESCR_SELECT(4);
416 } else {
417 /* logical cpu 1 */
418 perfctr_msr = MSR_P4_IQ_PERFCTR1;
419 evntsel_msr = MSR_P4_CRU_ESCR0;
420 cccr_msr = MSR_P4_IQ_CCCR1;
421 cccr_val = P4_CCCR_OVF_PMI1 | P4_CCCR_ESCR_SELECT(4);
422 }
423
424 evntsel = P4_ESCR_EVENT_SELECT(0x3F)
425 | P4_ESCR_OS
426 | P4_ESCR_USR;
427
428 cccr_val |= P4_CCCR_THRESHOLD(15)
429 | P4_CCCR_COMPLEMENT
430 | P4_CCCR_COMPARE
431 | P4_CCCR_REQUIRED;
432
433 wrmsr(evntsel_msr, evntsel, 0);
434 wrmsr(cccr_msr, cccr_val, 0);
435 write_watchdog_counter(perfctr_msr, "P4_IQ_COUNTER0", nmi_hz);
436 apic_write(APIC_LVTPC, APIC_DM_NMI);
437 cccr_val |= P4_CCCR_ENABLE;
438 wrmsr(cccr_msr, cccr_val, 0);
439 wd->perfctr_msr = perfctr_msr;
440 wd->evntsel_msr = evntsel_msr;
441 wd->cccr_msr = cccr_msr;
442 return 1;
443}
444
445static void stop_p4_watchdog(void *arg)
446{
447 struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
448 wrmsr(wd->cccr_msr, 0, 0);
449 wrmsr(wd->evntsel_msr, 0, 0);
450}
451
452static int p4_reserve(void)
453{
454 if (!reserve_perfctr_nmi(MSR_P4_IQ_PERFCTR0))
455 return 0;
456#ifdef CONFIG_SMP
457 if (smp_num_siblings > 1 && !reserve_perfctr_nmi(MSR_P4_IQ_PERFCTR1))
458 goto fail1;
459#endif
460 if (!reserve_evntsel_nmi(MSR_P4_CRU_ESCR0))
461 goto fail2;
462 /* RED-PEN why is ESCR1 not reserved here? */
463 return 1;
464 fail2:
465#ifdef CONFIG_SMP
466 if (smp_num_siblings > 1)
467 release_perfctr_nmi(MSR_P4_IQ_PERFCTR1);
468 fail1:
469#endif
470 release_perfctr_nmi(MSR_P4_IQ_PERFCTR0);
471 return 0;
472}
473
474static void p4_unreserve(void)
475{
476#ifdef CONFIG_SMP
477 if (smp_num_siblings > 1)
478 release_evntsel_nmi(MSR_P4_IQ_PERFCTR1);
479#endif
480 release_evntsel_nmi(MSR_P4_IQ_PERFCTR0);
481 release_perfctr_nmi(MSR_P4_CRU_ESCR0);
482}
483
484static void p4_rearm(struct nmi_watchdog_ctlblk *wd, unsigned nmi_hz)
485{
486 unsigned dummy;
487 /*
488 * P4 quirks:
489 * - An overflown perfctr will assert its interrupt
490 * until the OVF flag in its CCCR is cleared.
491 * - LVTPC is masked on interrupt and must be
492 * unmasked by the LVTPC handler.
493 */
494 rdmsrl(wd->cccr_msr, dummy);
495 dummy &= ~P4_CCCR_OVF;
496 wrmsrl(wd->cccr_msr, dummy);
497 apic_write(APIC_LVTPC, APIC_DM_NMI);
498 /* start the cycle over again */
499 write_watchdog_counter(wd->perfctr_msr, NULL, nmi_hz);
500}
501
502static struct wd_ops p4_wd_ops = {
503 .reserve = p4_reserve,
504 .unreserve = p4_unreserve,
505 .setup = setup_p4_watchdog,
506 .rearm = p4_rearm,
507 .stop = stop_p4_watchdog,
508 /* RED-PEN this is wrong for the other sibling */
509 .perfctr = MSR_P4_BPU_PERFCTR0,
510 .evntsel = MSR_P4_BSU_ESCR0,
511 .checkbit = 1ULL<<39,
512};
513
514/* Watchdog using the Intel architected PerfMon. Used for Core2 and hopefully
515 all future Intel CPUs. */
516
517#define ARCH_PERFMON_NMI_EVENT_SEL ARCH_PERFMON_UNHALTED_CORE_CYCLES_SEL
518#define ARCH_PERFMON_NMI_EVENT_UMASK ARCH_PERFMON_UNHALTED_CORE_CYCLES_UMASK
519
520static int setup_intel_arch_watchdog(unsigned nmi_hz)
521{
522 unsigned int ebx;
523 union cpuid10_eax eax;
524 unsigned int unused;
525 unsigned int perfctr_msr, evntsel_msr;
526 unsigned int evntsel;
527 struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
528
529 /*
530 * Check whether the Architectural PerfMon supports
531 * Unhalted Core Cycles Event or not.
532 * NOTE: Corresponding bit = 0 in ebx indicates event present.
533 */
534 cpuid(10, &(eax.full), &ebx, &unused, &unused);
535 if ((eax.split.mask_length < (ARCH_PERFMON_UNHALTED_CORE_CYCLES_INDEX+1)) ||
536 (ebx & ARCH_PERFMON_UNHALTED_CORE_CYCLES_PRESENT))
537 return 0;
538
539 perfctr_msr = MSR_ARCH_PERFMON_PERFCTR1;
540 evntsel_msr = MSR_ARCH_PERFMON_EVENTSEL1;
541
542 wrmsrl(perfctr_msr, 0UL);
543
544 evntsel = ARCH_PERFMON_EVENTSEL_INT
545 | ARCH_PERFMON_EVENTSEL_OS
546 | ARCH_PERFMON_EVENTSEL_USR
547 | ARCH_PERFMON_NMI_EVENT_SEL
548 | ARCH_PERFMON_NMI_EVENT_UMASK;
549
550 /* setup the timer */
551 wrmsr(evntsel_msr, evntsel, 0);
552 nmi_hz = adjust_for_32bit_ctr(nmi_hz);
553 write_watchdog_counter32(perfctr_msr, "INTEL_ARCH_PERFCTR0", nmi_hz);
554 apic_write(APIC_LVTPC, APIC_DM_NMI);
555 evntsel |= ARCH_PERFMON_EVENTSEL0_ENABLE;
556 wrmsr(evntsel_msr, evntsel, 0);
557
558 wd->perfctr_msr = perfctr_msr;
559 wd->evntsel_msr = evntsel_msr;
560 wd->cccr_msr = 0; //unused
561 wd_ops->checkbit = 1ULL << (eax.split.bit_width - 1);
562 return 1;
563}
564
565static struct wd_ops intel_arch_wd_ops = {
566 .reserve = single_msr_reserve,
567 .unreserve = single_msr_unreserve,
568 .setup = setup_intel_arch_watchdog,
569 .rearm = p6_rearm,
570 .stop = single_msr_stop_watchdog,
571 .perfctr = MSR_ARCH_PERFMON_PERFCTR0,
572 .evntsel = MSR_ARCH_PERFMON_EVENTSEL0,
573};
574
575static void probe_nmi_watchdog(void)
576{
577 switch (boot_cpu_data.x86_vendor) {
578 case X86_VENDOR_AMD:
579 if (boot_cpu_data.x86 != 6 && boot_cpu_data.x86 != 15 &&
580 boot_cpu_data.x86 != 16)
581 return;
582 wd_ops = &k7_wd_ops;
583 break;
584 case X86_VENDOR_INTEL:
585 if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
586 wd_ops = &intel_arch_wd_ops;
587 break;
588 }
589 switch (boot_cpu_data.x86) {
590 case 6:
591 if (boot_cpu_data.x86_model > 0xd)
592 return;
593
594 wd_ops = &p6_wd_ops;
595 break;
596 case 15:
597 if (boot_cpu_data.x86_model > 0x4)
598 return;
599
600 wd_ops = &p4_wd_ops;
601 break;
602 default:
603 return;
604 }
605 break;
606 }
607}
608
609/* Interface to nmi.c */
610
611int lapic_watchdog_init(unsigned nmi_hz)
612{
613 if (!wd_ops) {
614 probe_nmi_watchdog();
615 if (!wd_ops)
616 return -1;
617 }
618
619 if (!(wd_ops->setup(nmi_hz))) {
620 printk(KERN_ERR "Cannot setup NMI watchdog on CPU %d\n",
621 raw_smp_processor_id());
622 return -1;
623 }
624
625 return 0;
626}
627
628void lapic_watchdog_stop(void)
629{
630 if (wd_ops)
631 wd_ops->stop(NULL);
632}
633
634unsigned lapic_adjust_nmi_hz(unsigned hz)
635{
636 struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
637 if (wd->perfctr_msr == MSR_P6_PERFCTR0 ||
638 wd->perfctr_msr == MSR_ARCH_PERFMON_PERFCTR1)
639 hz = adjust_for_32bit_ctr(hz);
640 return hz;
641}
642
643int lapic_wd_event(unsigned nmi_hz)
644{
645 struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
646 u64 ctr;
647 rdmsrl(wd->perfctr_msr, ctr);
648 if (ctr & wd_ops->checkbit) { /* perfctr still running? */
649 return 0;
650 }
651 wd_ops->rearm(wd, nmi_hz);
652 return 1;
653}
654
655int lapic_watchdog_ok(void)
656{
657 return wd_ops != NULL;
658}
diff --git a/arch/i386/kernel/nmi.c b/arch/i386/kernel/nmi.c
index 2dec1b105737..33cf2f3c444f 100644
--- a/arch/i386/kernel/nmi.c
+++ b/arch/i386/kernel/nmi.c
@@ -20,7 +20,6 @@
20#include <linux/sysdev.h> 20#include <linux/sysdev.h>
21#include <linux/sysctl.h> 21#include <linux/sysctl.h>
22#include <linux/percpu.h> 22#include <linux/percpu.h>
23#include <linux/dmi.h>
24#include <linux/kprobes.h> 23#include <linux/kprobes.h>
25#include <linux/cpumask.h> 24#include <linux/cpumask.h>
26#include <linux/kernel_stat.h> 25#include <linux/kernel_stat.h>
@@ -28,30 +27,14 @@
28#include <asm/smp.h> 27#include <asm/smp.h>
29#include <asm/nmi.h> 28#include <asm/nmi.h>
30#include <asm/kdebug.h> 29#include <asm/kdebug.h>
31#include <asm/intel_arch_perfmon.h>
32 30
33#include "mach_traps.h" 31#include "mach_traps.h"
34 32
35int unknown_nmi_panic; 33int unknown_nmi_panic;
36int nmi_watchdog_enabled; 34int nmi_watchdog_enabled;
37 35
38/* perfctr_nmi_owner tracks the ownership of the perfctr registers:
39 * evtsel_nmi_owner tracks the ownership of the event selection
40 * - different performance counters/ event selection may be reserved for
41 * different subsystems this reservation system just tries to coordinate
42 * things a little
43 */
44
45/* this number is calculated from Intel's MSR_P4_CRU_ESCR5 register and it's
46 * offset from MSR_P4_BSU_ESCR0. It will be the max for all platforms (for now)
47 */
48#define NMI_MAX_COUNTER_BITS 66
49#define NMI_MAX_COUNTER_LONGS BITS_TO_LONGS(NMI_MAX_COUNTER_BITS)
50
51static DEFINE_PER_CPU(unsigned long, perfctr_nmi_owner[NMI_MAX_COUNTER_LONGS]);
52static DEFINE_PER_CPU(unsigned long, evntsel_nmi_owner[NMI_MAX_COUNTER_LONGS]);
53
54static cpumask_t backtrace_mask = CPU_MASK_NONE; 36static cpumask_t backtrace_mask = CPU_MASK_NONE;
37
55/* nmi_active: 38/* nmi_active:
56 * >0: the lapic NMI watchdog is active, but can be disabled 39 * >0: the lapic NMI watchdog is active, but can be disabled
57 * <0: the lapic NMI watchdog has not been set up, and cannot 40 * <0: the lapic NMI watchdog has not been set up, and cannot
@@ -63,203 +46,11 @@ atomic_t nmi_active = ATOMIC_INIT(0); /* oprofile uses this */
63unsigned int nmi_watchdog = NMI_DEFAULT; 46unsigned int nmi_watchdog = NMI_DEFAULT;
64static unsigned int nmi_hz = HZ; 47static unsigned int nmi_hz = HZ;
65 48
66struct nmi_watchdog_ctlblk { 49static DEFINE_PER_CPU(short, wd_enabled);
67 int enabled;
68 u64 check_bit;
69 unsigned int cccr_msr;
70 unsigned int perfctr_msr; /* the MSR to reset in NMI handler */
71 unsigned int evntsel_msr; /* the MSR to select the events to handle */
72};
73static DEFINE_PER_CPU(struct nmi_watchdog_ctlblk, nmi_watchdog_ctlblk);
74 50
75/* local prototypes */ 51/* local prototypes */
76static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu); 52static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu);
77 53
78/* converts an msr to an appropriate reservation bit */
79static inline unsigned int nmi_perfctr_msr_to_bit(unsigned int msr)
80{
81 /* returns the bit offset of the performance counter register */
82 switch (boot_cpu_data.x86_vendor) {
83 case X86_VENDOR_AMD:
84 return (msr - MSR_K7_PERFCTR0);
85 case X86_VENDOR_INTEL:
86 if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
87 return (msr - MSR_ARCH_PERFMON_PERFCTR0);
88
89 switch (boot_cpu_data.x86) {
90 case 6:
91 return (msr - MSR_P6_PERFCTR0);
92 case 15:
93 return (msr - MSR_P4_BPU_PERFCTR0);
94 }
95 }
96 return 0;
97}
98
99/* converts an msr to an appropriate reservation bit */
100static inline unsigned int nmi_evntsel_msr_to_bit(unsigned int msr)
101{
102 /* returns the bit offset of the event selection register */
103 switch (boot_cpu_data.x86_vendor) {
104 case X86_VENDOR_AMD:
105 return (msr - MSR_K7_EVNTSEL0);
106 case X86_VENDOR_INTEL:
107 if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
108 return (msr - MSR_ARCH_PERFMON_EVENTSEL0);
109
110 switch (boot_cpu_data.x86) {
111 case 6:
112 return (msr - MSR_P6_EVNTSEL0);
113 case 15:
114 return (msr - MSR_P4_BSU_ESCR0);
115 }
116 }
117 return 0;
118}
119
120/* checks for a bit availability (hack for oprofile) */
121int avail_to_resrv_perfctr_nmi_bit(unsigned int counter)
122{
123 int cpu;
124 BUG_ON(counter > NMI_MAX_COUNTER_BITS);
125 for_each_possible_cpu (cpu) {
126 if (test_bit(counter, &per_cpu(perfctr_nmi_owner, cpu)[0]))
127 return 0;
128 }
129 return 1;
130}
131
132/* checks the an msr for availability */
133int avail_to_resrv_perfctr_nmi(unsigned int msr)
134{
135 unsigned int counter;
136 int cpu;
137
138 counter = nmi_perfctr_msr_to_bit(msr);
139 BUG_ON(counter > NMI_MAX_COUNTER_BITS);
140
141 for_each_possible_cpu (cpu) {
142 if (test_bit(counter, &per_cpu(perfctr_nmi_owner, cpu)[0]))
143 return 0;
144 }
145 return 1;
146}
147
148static int __reserve_perfctr_nmi(int cpu, unsigned int msr)
149{
150 unsigned int counter;
151 if (cpu < 0)
152 cpu = smp_processor_id();
153
154 counter = nmi_perfctr_msr_to_bit(msr);
155 BUG_ON(counter > NMI_MAX_COUNTER_BITS);
156
157 if (!test_and_set_bit(counter, &per_cpu(perfctr_nmi_owner, cpu)[0]))
158 return 1;
159 return 0;
160}
161
162static void __release_perfctr_nmi(int cpu, unsigned int msr)
163{
164 unsigned int counter;
165 if (cpu < 0)
166 cpu = smp_processor_id();
167
168 counter = nmi_perfctr_msr_to_bit(msr);
169 BUG_ON(counter > NMI_MAX_COUNTER_BITS);
170
171 clear_bit(counter, &per_cpu(perfctr_nmi_owner, cpu)[0]);
172}
173
174int reserve_perfctr_nmi(unsigned int msr)
175{
176 int cpu, i;
177 for_each_possible_cpu (cpu) {
178 if (!__reserve_perfctr_nmi(cpu, msr)) {
179 for_each_possible_cpu (i) {
180 if (i >= cpu)
181 break;
182 __release_perfctr_nmi(i, msr);
183 }
184 return 0;
185 }
186 }
187 return 1;
188}
189
190void release_perfctr_nmi(unsigned int msr)
191{
192 int cpu;
193 for_each_possible_cpu (cpu) {
194 __release_perfctr_nmi(cpu, msr);
195 }
196}
197
198int __reserve_evntsel_nmi(int cpu, unsigned int msr)
199{
200 unsigned int counter;
201 if (cpu < 0)
202 cpu = smp_processor_id();
203
204 counter = nmi_evntsel_msr_to_bit(msr);
205 BUG_ON(counter > NMI_MAX_COUNTER_BITS);
206
207 if (!test_and_set_bit(counter, &per_cpu(evntsel_nmi_owner, cpu)[0]))
208 return 1;
209 return 0;
210}
211
212static void __release_evntsel_nmi(int cpu, unsigned int msr)
213{
214 unsigned int counter;
215 if (cpu < 0)
216 cpu = smp_processor_id();
217
218 counter = nmi_evntsel_msr_to_bit(msr);
219 BUG_ON(counter > NMI_MAX_COUNTER_BITS);
220
221 clear_bit(counter, &per_cpu(evntsel_nmi_owner, cpu)[0]);
222}
223
224int reserve_evntsel_nmi(unsigned int msr)
225{
226 int cpu, i;
227 for_each_possible_cpu (cpu) {
228 if (!__reserve_evntsel_nmi(cpu, msr)) {
229 for_each_possible_cpu (i) {
230 if (i >= cpu)
231 break;
232 __release_evntsel_nmi(i, msr);
233 }
234 return 0;
235 }
236 }
237 return 1;
238}
239
240void release_evntsel_nmi(unsigned int msr)
241{
242 int cpu;
243 for_each_possible_cpu (cpu) {
244 __release_evntsel_nmi(cpu, msr);
245 }
246}
247
248static __cpuinit inline int nmi_known_cpu(void)
249{
250 switch (boot_cpu_data.x86_vendor) {
251 case X86_VENDOR_AMD:
252 return ((boot_cpu_data.x86 == 15) || (boot_cpu_data.x86 == 6)
253 || (boot_cpu_data.x86 == 16));
254 case X86_VENDOR_INTEL:
255 if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
256 return 1;
257 else
258 return ((boot_cpu_data.x86 == 15) || (boot_cpu_data.x86 == 6));
259 }
260 return 0;
261}
262
263static int endflag __initdata = 0; 54static int endflag __initdata = 0;
264 55
265#ifdef CONFIG_SMP 56#ifdef CONFIG_SMP
@@ -281,28 +72,6 @@ static __init void nmi_cpu_busy(void *data)
281} 72}
282#endif 73#endif
283 74
284static unsigned int adjust_for_32bit_ctr(unsigned int hz)
285{
286 u64 counter_val;
287 unsigned int retval = hz;
288
289 /*
290 * On Intel CPUs with P6/ARCH_PERFMON only 32 bits in the counter
291 * are writable, with higher bits sign extending from bit 31.
292 * So, we can only program the counter with 31 bit values and
293 * 32nd bit should be 1, for 33.. to be 1.
294 * Find the appropriate nmi_hz
295 */
296 counter_val = (u64)cpu_khz * 1000;
297 do_div(counter_val, retval);
298 if (counter_val > 0x7fffffffULL) {
299 u64 count = (u64)cpu_khz * 1000;
300 do_div(count, 0x7fffffffUL);
301 retval = count + 1;
302 }
303 return retval;
304}
305
306static int __init check_nmi_watchdog(void) 75static int __init check_nmi_watchdog(void)
307{ 76{
308 unsigned int *prev_nmi_count; 77 unsigned int *prev_nmi_count;
@@ -335,14 +104,14 @@ static int __init check_nmi_watchdog(void)
335 if (!cpu_isset(cpu, cpu_callin_map)) 104 if (!cpu_isset(cpu, cpu_callin_map))
336 continue; 105 continue;
337#endif 106#endif
338 if (!per_cpu(nmi_watchdog_ctlblk, cpu).enabled) 107 if (!per_cpu(wd_enabled, cpu))
339 continue; 108 continue;
340 if (nmi_count(cpu) - prev_nmi_count[cpu] <= 5) { 109 if (nmi_count(cpu) - prev_nmi_count[cpu] <= 5) {
341 printk("CPU#%d: NMI appears to be stuck (%d->%d)!\n", 110 printk("CPU#%d: NMI appears to be stuck (%d->%d)!\n",
342 cpu, 111 cpu,
343 prev_nmi_count[cpu], 112 prev_nmi_count[cpu],
344 nmi_count(cpu)); 113 nmi_count(cpu));
345 per_cpu(nmi_watchdog_ctlblk, cpu).enabled = 0; 114 per_cpu(wd_enabled, cpu) = 0;
346 atomic_dec(&nmi_active); 115 atomic_dec(&nmi_active);
347 } 116 }
348 } 117 }
@@ -356,16 +125,8 @@ static int __init check_nmi_watchdog(void)
356 125
357 /* now that we know it works we can reduce NMI frequency to 126 /* now that we know it works we can reduce NMI frequency to
358 something more reasonable; makes a difference in some configs */ 127 something more reasonable; makes a difference in some configs */
359 if (nmi_watchdog == NMI_LOCAL_APIC) { 128 if (nmi_watchdog == NMI_LOCAL_APIC)
360 struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk); 129 nmi_hz = lapic_adjust_nmi_hz(1);
361
362 nmi_hz = 1;
363
364 if (wd->perfctr_msr == MSR_P6_PERFCTR0 ||
365 wd->perfctr_msr == MSR_ARCH_PERFMON_PERFCTR1) {
366 nmi_hz = adjust_for_32bit_ctr(nmi_hz);
367 }
368 }
369 130
370 kfree(prev_nmi_count); 131 kfree(prev_nmi_count);
371 return 0; 132 return 0;
@@ -388,85 +149,8 @@ static int __init setup_nmi_watchdog(char *str)
388 149
389__setup("nmi_watchdog=", setup_nmi_watchdog); 150__setup("nmi_watchdog=", setup_nmi_watchdog);
390 151
391static void disable_lapic_nmi_watchdog(void)
392{
393 BUG_ON(nmi_watchdog != NMI_LOCAL_APIC);
394
395 if (atomic_read(&nmi_active) <= 0)
396 return;
397
398 on_each_cpu(stop_apic_nmi_watchdog, NULL, 0, 1);
399
400 BUG_ON(atomic_read(&nmi_active) != 0);
401}
402
403static void enable_lapic_nmi_watchdog(void)
404{
405 BUG_ON(nmi_watchdog != NMI_LOCAL_APIC);
406
407 /* are we already enabled */
408 if (atomic_read(&nmi_active) != 0)
409 return;
410
411 /* are we lapic aware */
412 if (nmi_known_cpu() <= 0)
413 return;
414
415 on_each_cpu(setup_apic_nmi_watchdog, NULL, 0, 1);
416 touch_nmi_watchdog();
417}
418
419void disable_timer_nmi_watchdog(void)
420{
421 BUG_ON(nmi_watchdog != NMI_IO_APIC);
422
423 if (atomic_read(&nmi_active) <= 0)
424 return;
425
426 disable_irq(0);
427 on_each_cpu(stop_apic_nmi_watchdog, NULL, 0, 1);
428
429 BUG_ON(atomic_read(&nmi_active) != 0);
430}
431
432void enable_timer_nmi_watchdog(void)
433{
434 BUG_ON(nmi_watchdog != NMI_IO_APIC);
435
436 if (atomic_read(&nmi_active) == 0) {
437 touch_nmi_watchdog();
438 on_each_cpu(setup_apic_nmi_watchdog, NULL, 0, 1);
439 enable_irq(0);
440 }
441}
442
443static void __acpi_nmi_disable(void *__unused)
444{
445 apic_write_around(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
446}
447 152
448/* 153/* Suspend/resume support */
449 * Disable timer based NMIs on all CPUs:
450 */
451void acpi_nmi_disable(void)
452{
453 if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
454 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
455}
456
457static void __acpi_nmi_enable(void *__unused)
458{
459 apic_write_around(APIC_LVT0, APIC_DM_NMI);
460}
461
462/*
463 * Enable timer based NMIs on all CPUs:
464 */
465void acpi_nmi_enable(void)
466{
467 if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
468 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
469}
470 154
471#ifdef CONFIG_PM 155#ifdef CONFIG_PM
472 156
@@ -513,7 +197,7 @@ static int __init init_lapic_nmi_sysfs(void)
513 if (nmi_watchdog != NMI_LOCAL_APIC) 197 if (nmi_watchdog != NMI_LOCAL_APIC)
514 return 0; 198 return 0;
515 199
516 if ( atomic_read(&nmi_active) < 0 ) 200 if (atomic_read(&nmi_active) < 0)
517 return 0; 201 return 0;
518 202
519 error = sysdev_class_register(&nmi_sysclass); 203 error = sysdev_class_register(&nmi_sysclass);
@@ -526,433 +210,69 @@ late_initcall(init_lapic_nmi_sysfs);
526 210
527#endif /* CONFIG_PM */ 211#endif /* CONFIG_PM */
528 212
529/* 213static void __acpi_nmi_enable(void *__unused)
530 * Activate the NMI watchdog via the local APIC.
531 * Original code written by Keith Owens.
532 */
533
534static void write_watchdog_counter(unsigned int perfctr_msr, const char *descr)
535{
536 u64 count = (u64)cpu_khz * 1000;
537
538 do_div(count, nmi_hz);
539 if(descr)
540 Dprintk("setting %s to -0x%08Lx\n", descr, count);
541 wrmsrl(perfctr_msr, 0 - count);
542}
543
544static void write_watchdog_counter32(unsigned int perfctr_msr,
545 const char *descr)
546{
547 u64 count = (u64)cpu_khz * 1000;
548
549 do_div(count, nmi_hz);
550 if(descr)
551 Dprintk("setting %s to -0x%08Lx\n", descr, count);
552 wrmsr(perfctr_msr, (u32)(-count), 0);
553}
554
555/* Note that these events don't tick when the CPU idles. This means
556 the frequency varies with CPU load. */
557
558#define K7_EVNTSEL_ENABLE (1 << 22)
559#define K7_EVNTSEL_INT (1 << 20)
560#define K7_EVNTSEL_OS (1 << 17)
561#define K7_EVNTSEL_USR (1 << 16)
562#define K7_EVENT_CYCLES_PROCESSOR_IS_RUNNING 0x76
563#define K7_NMI_EVENT K7_EVENT_CYCLES_PROCESSOR_IS_RUNNING
564
565static int setup_k7_watchdog(void)
566{
567 unsigned int perfctr_msr, evntsel_msr;
568 unsigned int evntsel;
569 struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
570
571 perfctr_msr = MSR_K7_PERFCTR0;
572 evntsel_msr = MSR_K7_EVNTSEL0;
573 if (!__reserve_perfctr_nmi(-1, perfctr_msr))
574 goto fail;
575
576 if (!__reserve_evntsel_nmi(-1, evntsel_msr))
577 goto fail1;
578
579 wrmsrl(perfctr_msr, 0UL);
580
581 evntsel = K7_EVNTSEL_INT
582 | K7_EVNTSEL_OS
583 | K7_EVNTSEL_USR
584 | K7_NMI_EVENT;
585
586 /* setup the timer */
587 wrmsr(evntsel_msr, evntsel, 0);
588 write_watchdog_counter(perfctr_msr, "K7_PERFCTR0");
589 apic_write(APIC_LVTPC, APIC_DM_NMI);
590 evntsel |= K7_EVNTSEL_ENABLE;
591 wrmsr(evntsel_msr, evntsel, 0);
592
593 wd->perfctr_msr = perfctr_msr;
594 wd->evntsel_msr = evntsel_msr;
595 wd->cccr_msr = 0; //unused
596 wd->check_bit = 1ULL<<63;
597 return 1;
598fail1:
599 __release_perfctr_nmi(-1, perfctr_msr);
600fail:
601 return 0;
602}
603
604static void stop_k7_watchdog(void)
605{
606 struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
607
608 wrmsr(wd->evntsel_msr, 0, 0);
609
610 __release_evntsel_nmi(-1, wd->evntsel_msr);
611 __release_perfctr_nmi(-1, wd->perfctr_msr);
612}
613
614#define P6_EVNTSEL0_ENABLE (1 << 22)
615#define P6_EVNTSEL_INT (1 << 20)
616#define P6_EVNTSEL_OS (1 << 17)
617#define P6_EVNTSEL_USR (1 << 16)
618#define P6_EVENT_CPU_CLOCKS_NOT_HALTED 0x79
619#define P6_NMI_EVENT P6_EVENT_CPU_CLOCKS_NOT_HALTED
620
621static int setup_p6_watchdog(void)
622{
623 unsigned int perfctr_msr, evntsel_msr;
624 unsigned int evntsel;
625 struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
626
627 perfctr_msr = MSR_P6_PERFCTR0;
628 evntsel_msr = MSR_P6_EVNTSEL0;
629 if (!__reserve_perfctr_nmi(-1, perfctr_msr))
630 goto fail;
631
632 if (!__reserve_evntsel_nmi(-1, evntsel_msr))
633 goto fail1;
634
635 wrmsrl(perfctr_msr, 0UL);
636
637 evntsel = P6_EVNTSEL_INT
638 | P6_EVNTSEL_OS
639 | P6_EVNTSEL_USR
640 | P6_NMI_EVENT;
641
642 /* setup the timer */
643 wrmsr(evntsel_msr, evntsel, 0);
644 nmi_hz = adjust_for_32bit_ctr(nmi_hz);
645 write_watchdog_counter32(perfctr_msr, "P6_PERFCTR0");
646 apic_write(APIC_LVTPC, APIC_DM_NMI);
647 evntsel |= P6_EVNTSEL0_ENABLE;
648 wrmsr(evntsel_msr, evntsel, 0);
649
650 wd->perfctr_msr = perfctr_msr;
651 wd->evntsel_msr = evntsel_msr;
652 wd->cccr_msr = 0; //unused
653 wd->check_bit = 1ULL<<39;
654 return 1;
655fail1:
656 __release_perfctr_nmi(-1, perfctr_msr);
657fail:
658 return 0;
659}
660
661static void stop_p6_watchdog(void)
662{
663 struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
664
665 wrmsr(wd->evntsel_msr, 0, 0);
666
667 __release_evntsel_nmi(-1, wd->evntsel_msr);
668 __release_perfctr_nmi(-1, wd->perfctr_msr);
669}
670
671/* Note that these events don't tick when the CPU idles. This means
672 the frequency varies with CPU load. */
673
674#define MSR_P4_MISC_ENABLE_PERF_AVAIL (1<<7)
675#define P4_ESCR_EVENT_SELECT(N) ((N)<<25)
676#define P4_ESCR_OS (1<<3)
677#define P4_ESCR_USR (1<<2)
678#define P4_CCCR_OVF_PMI0 (1<<26)
679#define P4_CCCR_OVF_PMI1 (1<<27)
680#define P4_CCCR_THRESHOLD(N) ((N)<<20)
681#define P4_CCCR_COMPLEMENT (1<<19)
682#define P4_CCCR_COMPARE (1<<18)
683#define P4_CCCR_REQUIRED (3<<16)
684#define P4_CCCR_ESCR_SELECT(N) ((N)<<13)
685#define P4_CCCR_ENABLE (1<<12)
686#define P4_CCCR_OVF (1<<31)
687/* Set up IQ_COUNTER0 to behave like a clock, by having IQ_CCCR0 filter
688 CRU_ESCR0 (with any non-null event selector) through a complemented
689 max threshold. [IA32-Vol3, Section 14.9.9] */
690
691static int setup_p4_watchdog(void)
692{ 214{
693 unsigned int perfctr_msr, evntsel_msr, cccr_msr; 215 apic_write_around(APIC_LVT0, APIC_DM_NMI);
694 unsigned int evntsel, cccr_val;
695 unsigned int misc_enable, dummy;
696 unsigned int ht_num;
697 struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
698
699 rdmsr(MSR_IA32_MISC_ENABLE, misc_enable, dummy);
700 if (!(misc_enable & MSR_P4_MISC_ENABLE_PERF_AVAIL))
701 return 0;
702
703#ifdef CONFIG_SMP
704 /* detect which hyperthread we are on */
705 if (smp_num_siblings == 2) {
706 unsigned int ebx, apicid;
707
708 ebx = cpuid_ebx(1);
709 apicid = (ebx >> 24) & 0xff;
710 ht_num = apicid & 1;
711 } else
712#endif
713 ht_num = 0;
714
715 /* performance counters are shared resources
716 * assign each hyperthread its own set
717 * (re-use the ESCR0 register, seems safe
718 * and keeps the cccr_val the same)
719 */
720 if (!ht_num) {
721 /* logical cpu 0 */
722 perfctr_msr = MSR_P4_IQ_PERFCTR0;
723 evntsel_msr = MSR_P4_CRU_ESCR0;
724 cccr_msr = MSR_P4_IQ_CCCR0;
725 cccr_val = P4_CCCR_OVF_PMI0 | P4_CCCR_ESCR_SELECT(4);
726 } else {
727 /* logical cpu 1 */
728 perfctr_msr = MSR_P4_IQ_PERFCTR1;
729 evntsel_msr = MSR_P4_CRU_ESCR0;
730 cccr_msr = MSR_P4_IQ_CCCR1;
731 cccr_val = P4_CCCR_OVF_PMI1 | P4_CCCR_ESCR_SELECT(4);
732 }
733
734 if (!__reserve_perfctr_nmi(-1, perfctr_msr))
735 goto fail;
736
737 if (!__reserve_evntsel_nmi(-1, evntsel_msr))
738 goto fail1;
739
740 evntsel = P4_ESCR_EVENT_SELECT(0x3F)
741 | P4_ESCR_OS
742 | P4_ESCR_USR;
743
744 cccr_val |= P4_CCCR_THRESHOLD(15)
745 | P4_CCCR_COMPLEMENT
746 | P4_CCCR_COMPARE
747 | P4_CCCR_REQUIRED;
748
749 wrmsr(evntsel_msr, evntsel, 0);
750 wrmsr(cccr_msr, cccr_val, 0);
751 write_watchdog_counter(perfctr_msr, "P4_IQ_COUNTER0");
752 apic_write(APIC_LVTPC, APIC_DM_NMI);
753 cccr_val |= P4_CCCR_ENABLE;
754 wrmsr(cccr_msr, cccr_val, 0);
755 wd->perfctr_msr = perfctr_msr;
756 wd->evntsel_msr = evntsel_msr;
757 wd->cccr_msr = cccr_msr;
758 wd->check_bit = 1ULL<<39;
759 return 1;
760fail1:
761 __release_perfctr_nmi(-1, perfctr_msr);
762fail:
763 return 0;
764} 216}
765 217
766static void stop_p4_watchdog(void) 218/*
219 * Enable timer based NMIs on all CPUs:
220 */
221void acpi_nmi_enable(void)
767{ 222{
768 struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk); 223 if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
769 224 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
770 wrmsr(wd->cccr_msr, 0, 0);
771 wrmsr(wd->evntsel_msr, 0, 0);
772
773 __release_evntsel_nmi(-1, wd->evntsel_msr);
774 __release_perfctr_nmi(-1, wd->perfctr_msr);
775} 225}
776 226
777#define ARCH_PERFMON_NMI_EVENT_SEL ARCH_PERFMON_UNHALTED_CORE_CYCLES_SEL 227static void __acpi_nmi_disable(void *__unused)
778#define ARCH_PERFMON_NMI_EVENT_UMASK ARCH_PERFMON_UNHALTED_CORE_CYCLES_UMASK
779
780static int setup_intel_arch_watchdog(void)
781{ 228{
782 unsigned int ebx; 229 apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
783 union cpuid10_eax eax;
784 unsigned int unused;
785 unsigned int perfctr_msr, evntsel_msr;
786 unsigned int evntsel;
787 struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
788
789 /*
790 * Check whether the Architectural PerfMon supports
791 * Unhalted Core Cycles Event or not.
792 * NOTE: Corresponding bit = 0 in ebx indicates event present.
793 */
794 cpuid(10, &(eax.full), &ebx, &unused, &unused);
795 if ((eax.split.mask_length < (ARCH_PERFMON_UNHALTED_CORE_CYCLES_INDEX+1)) ||
796 (ebx & ARCH_PERFMON_UNHALTED_CORE_CYCLES_PRESENT))
797 goto fail;
798
799 perfctr_msr = MSR_ARCH_PERFMON_PERFCTR1;
800 evntsel_msr = MSR_ARCH_PERFMON_EVENTSEL1;
801
802 if (!__reserve_perfctr_nmi(-1, perfctr_msr))
803 goto fail;
804
805 if (!__reserve_evntsel_nmi(-1, evntsel_msr))
806 goto fail1;
807
808 wrmsrl(perfctr_msr, 0UL);
809
810 evntsel = ARCH_PERFMON_EVENTSEL_INT
811 | ARCH_PERFMON_EVENTSEL_OS
812 | ARCH_PERFMON_EVENTSEL_USR
813 | ARCH_PERFMON_NMI_EVENT_SEL
814 | ARCH_PERFMON_NMI_EVENT_UMASK;
815
816 /* setup the timer */
817 wrmsr(evntsel_msr, evntsel, 0);
818 nmi_hz = adjust_for_32bit_ctr(nmi_hz);
819 write_watchdog_counter32(perfctr_msr, "INTEL_ARCH_PERFCTR0");
820 apic_write(APIC_LVTPC, APIC_DM_NMI);
821 evntsel |= ARCH_PERFMON_EVENTSEL0_ENABLE;
822 wrmsr(evntsel_msr, evntsel, 0);
823
824 wd->perfctr_msr = perfctr_msr;
825 wd->evntsel_msr = evntsel_msr;
826 wd->cccr_msr = 0; //unused
827 wd->check_bit = 1ULL << (eax.split.bit_width - 1);
828 return 1;
829fail1:
830 __release_perfctr_nmi(-1, perfctr_msr);
831fail:
832 return 0;
833} 230}
834 231
835static void stop_intel_arch_watchdog(void) 232/*
233 * Disable timer based NMIs on all CPUs:
234 */
235void acpi_nmi_disable(void)
836{ 236{
837 unsigned int ebx; 237 if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
838 union cpuid10_eax eax; 238 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
839 unsigned int unused;
840 struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
841
842 /*
843 * Check whether the Architectural PerfMon supports
844 * Unhalted Core Cycles Event or not.
845 * NOTE: Corresponding bit = 0 in ebx indicates event present.
846 */
847 cpuid(10, &(eax.full), &ebx, &unused, &unused);
848 if ((eax.split.mask_length < (ARCH_PERFMON_UNHALTED_CORE_CYCLES_INDEX+1)) ||
849 (ebx & ARCH_PERFMON_UNHALTED_CORE_CYCLES_PRESENT))
850 return;
851
852 wrmsr(wd->evntsel_msr, 0, 0);
853 __release_evntsel_nmi(-1, wd->evntsel_msr);
854 __release_perfctr_nmi(-1, wd->perfctr_msr);
855} 239}
856 240
857void setup_apic_nmi_watchdog (void *unused) 241void setup_apic_nmi_watchdog (void *unused)
858{ 242{
859 struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk); 243 if (__get_cpu_var(wd_enabled))
860 244 return;
861 /* only support LOCAL and IO APICs for now */
862 if ((nmi_watchdog != NMI_LOCAL_APIC) &&
863 (nmi_watchdog != NMI_IO_APIC))
864 return;
865
866 if (wd->enabled == 1)
867 return;
868 245
869 /* cheap hack to support suspend/resume */ 246 /* cheap hack to support suspend/resume */
870 /* if cpu0 is not active neither should the other cpus */ 247 /* if cpu0 is not active neither should the other cpus */
871 if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0)) 248 if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0))
872 return; 249 return;
873 250
874 if (nmi_watchdog == NMI_LOCAL_APIC) { 251 switch (nmi_watchdog) {
875 switch (boot_cpu_data.x86_vendor) { 252 case NMI_LOCAL_APIC:
876 case X86_VENDOR_AMD: 253 __get_cpu_var(wd_enabled) = 1; /* enable it before to avoid race with handler */
877 if (boot_cpu_data.x86 != 6 && boot_cpu_data.x86 != 15 && 254 if (lapic_watchdog_init(nmi_hz) < 0) {
878 boot_cpu_data.x86 != 16) 255 __get_cpu_var(wd_enabled) = 0;
879 return;
880 if (!setup_k7_watchdog())
881 return;
882 break;
883 case X86_VENDOR_INTEL:
884 if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
885 if (!setup_intel_arch_watchdog())
886 return;
887 break;
888 }
889 switch (boot_cpu_data.x86) {
890 case 6:
891 if (boot_cpu_data.x86_model > 0xd)
892 return;
893
894 if (!setup_p6_watchdog())
895 return;
896 break;
897 case 15:
898 if (boot_cpu_data.x86_model > 0x4)
899 return;
900
901 if (!setup_p4_watchdog())
902 return;
903 break;
904 default:
905 return;
906 }
907 break;
908 default:
909 return; 256 return;
910 } 257 }
258 /* FALL THROUGH */
259 case NMI_IO_APIC:
260 __get_cpu_var(wd_enabled) = 1;
261 atomic_inc(&nmi_active);
911 } 262 }
912 wd->enabled = 1;
913 atomic_inc(&nmi_active);
914} 263}
915 264
916void stop_apic_nmi_watchdog(void *unused) 265void stop_apic_nmi_watchdog(void *unused)
917{ 266{
918 struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
919
920 /* only support LOCAL and IO APICs for now */ 267 /* only support LOCAL and IO APICs for now */
921 if ((nmi_watchdog != NMI_LOCAL_APIC) && 268 if ((nmi_watchdog != NMI_LOCAL_APIC) &&
922 (nmi_watchdog != NMI_IO_APIC)) 269 (nmi_watchdog != NMI_IO_APIC))
923 return; 270 return;
924 271 if (__get_cpu_var(wd_enabled) == 0)
925 if (wd->enabled == 0)
926 return; 272 return;
927 273 if (nmi_watchdog == NMI_LOCAL_APIC)
928 if (nmi_watchdog == NMI_LOCAL_APIC) { 274 lapic_watchdog_stop();
929 switch (boot_cpu_data.x86_vendor) { 275 __get_cpu_var(wd_enabled) = 0;
930 case X86_VENDOR_AMD:
931 stop_k7_watchdog();
932 break;
933 case X86_VENDOR_INTEL:
934 if (cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
935 stop_intel_arch_watchdog();
936 break;
937 }
938 switch (boot_cpu_data.x86) {
939 case 6:
940 if (boot_cpu_data.x86_model > 0xd)
941 break;
942 stop_p6_watchdog();
943 break;
944 case 15:
945 if (boot_cpu_data.x86_model > 0x4)
946 break;
947 stop_p4_watchdog();
948 break;
949 }
950 break;
951 default:
952 return;
953 }
954 }
955 wd->enabled = 0;
956 atomic_dec(&nmi_active); 276 atomic_dec(&nmi_active);
957} 277}
958 278
@@ -1008,8 +328,6 @@ __kprobes int nmi_watchdog_tick(struct pt_regs * regs, unsigned reason)
1008 unsigned int sum; 328 unsigned int sum;
1009 int touched = 0; 329 int touched = 0;
1010 int cpu = smp_processor_id(); 330 int cpu = smp_processor_id();
1011 struct nmi_watchdog_ctlblk *wd = &__get_cpu_var(nmi_watchdog_ctlblk);
1012 u64 dummy;
1013 int rc=0; 331 int rc=0;
1014 332
1015 /* check for other users first */ 333 /* check for other users first */
@@ -1052,53 +370,20 @@ __kprobes int nmi_watchdog_tick(struct pt_regs * regs, unsigned reason)
1052 alert_counter[cpu] = 0; 370 alert_counter[cpu] = 0;
1053 } 371 }
1054 /* see if the nmi watchdog went off */ 372 /* see if the nmi watchdog went off */
1055 if (wd->enabled) { 373 if (!__get_cpu_var(wd_enabled))
1056 if (nmi_watchdog == NMI_LOCAL_APIC) { 374 return rc;
1057 rdmsrl(wd->perfctr_msr, dummy); 375 switch (nmi_watchdog) {
1058 if (dummy & wd->check_bit){ 376 case NMI_LOCAL_APIC:
1059 /* this wasn't a watchdog timer interrupt */ 377 rc |= lapic_wd_event(nmi_hz);
1060 goto done; 378 break;
1061 } 379 case NMI_IO_APIC:
1062 380 /* don't know how to accurately check for this.
1063 /* only Intel P4 uses the cccr msr */ 381 * just assume it was a watchdog timer interrupt
1064 if (wd->cccr_msr != 0) { 382 * This matches the old behaviour.
1065 /* 383 */
1066 * P4 quirks: 384 rc = 1;
1067 * - An overflown perfctr will assert its interrupt 385 break;
1068 * until the OVF flag in its CCCR is cleared.
1069 * - LVTPC is masked on interrupt and must be
1070 * unmasked by the LVTPC handler.
1071 */
1072 rdmsrl(wd->cccr_msr, dummy);
1073 dummy &= ~P4_CCCR_OVF;
1074 wrmsrl(wd->cccr_msr, dummy);
1075 apic_write(APIC_LVTPC, APIC_DM_NMI);
1076 /* start the cycle over again */
1077 write_watchdog_counter(wd->perfctr_msr, NULL);
1078 }
1079 else if (wd->perfctr_msr == MSR_P6_PERFCTR0 ||
1080 wd->perfctr_msr == MSR_ARCH_PERFMON_PERFCTR1) {
1081 /* P6 based Pentium M need to re-unmask
1082 * the apic vector but it doesn't hurt
1083 * other P6 variant.
1084 * ArchPerfom/Core Duo also needs this */
1085 apic_write(APIC_LVTPC, APIC_DM_NMI);
1086 /* P6/ARCH_PERFMON has 32 bit counter write */
1087 write_watchdog_counter32(wd->perfctr_msr, NULL);
1088 } else {
1089 /* start the cycle over again */
1090 write_watchdog_counter(wd->perfctr_msr, NULL);
1091 }
1092 rc = 1;
1093 } else if (nmi_watchdog == NMI_IO_APIC) {
1094 /* don't know how to accurately check for this.
1095 * just assume it was a watchdog timer interrupt
1096 * This matches the old behaviour.
1097 */
1098 rc = 1;
1099 }
1100 } 386 }
1101done:
1102 return rc; 387 return rc;
1103} 388}
1104 389
@@ -1143,7 +428,7 @@ int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
1143 } 428 }
1144 429
1145 if (nmi_watchdog == NMI_DEFAULT) { 430 if (nmi_watchdog == NMI_DEFAULT) {
1146 if (nmi_known_cpu() > 0) 431 if (lapic_watchdog_ok())
1147 nmi_watchdog = NMI_LOCAL_APIC; 432 nmi_watchdog = NMI_LOCAL_APIC;
1148 else 433 else
1149 nmi_watchdog = NMI_IO_APIC; 434 nmi_watchdog = NMI_IO_APIC;
@@ -1179,11 +464,3 @@ void __trigger_all_cpu_backtrace(void)
1179 464
1180EXPORT_SYMBOL(nmi_active); 465EXPORT_SYMBOL(nmi_active);
1181EXPORT_SYMBOL(nmi_watchdog); 466EXPORT_SYMBOL(nmi_watchdog);
1182EXPORT_SYMBOL(avail_to_resrv_perfctr_nmi);
1183EXPORT_SYMBOL(avail_to_resrv_perfctr_nmi_bit);
1184EXPORT_SYMBOL(reserve_perfctr_nmi);
1185EXPORT_SYMBOL(release_perfctr_nmi);
1186EXPORT_SYMBOL(reserve_evntsel_nmi);
1187EXPORT_SYMBOL(release_evntsel_nmi);
1188EXPORT_SYMBOL(disable_timer_nmi_watchdog);
1189EXPORT_SYMBOL(enable_timer_nmi_watchdog);
diff --git a/include/asm-i386/nmi.h b/include/asm-i386/nmi.h
index b04333ea6f31..fb1e133efd9f 100644
--- a/include/asm-i386/nmi.h
+++ b/include/asm-i386/nmi.h
@@ -50,4 +50,12 @@ void __trigger_all_cpu_backtrace(void);
50 50
51#endif 51#endif
52 52
53void lapic_watchdog_stop(void);
54int lapic_watchdog_init(unsigned nmi_hz);
55int lapic_wd_event(unsigned nmi_hz);
56unsigned lapic_adjust_nmi_hz(unsigned hz);
57int lapic_watchdog_ok(void);
58void disable_lapic_nmi_watchdog(void);
59void enable_lapic_nmi_watchdog(void);
60
53#endif /* ASM_NMI_H */ 61#endif /* ASM_NMI_H */
generated by cgit v1.2.2 (git 2.25.0) at 2024-12-01 16:21:04 -0500