KVM: VMX: enable VMXON check with SMX enabled (Intel TXT)
[linux-2.6/x86.git] / arch / x86 / kernel / cpu / perf_event_intel_ds.c
blobec8b2e12e10452b2a67ffd88bb02067672c285e4
1 #ifdef CONFIG_CPU_SUP_INTEL
3 /* The maximal number of PEBS events: */
4 #define MAX_PEBS_EVENTS 4
6 /* The size of a BTS record in bytes: */
7 #define BTS_RECORD_SIZE 24
9 #define BTS_BUFFER_SIZE (PAGE_SIZE << 4)
10 #define PEBS_BUFFER_SIZE PAGE_SIZE
13 * pebs_record_32 for p4 and core not supported
15 struct pebs_record_32 {
16 u32 flags, ip;
17 u32 ax, bc, cx, dx;
18 u32 si, di, bp, sp;
23 struct pebs_record_core {
24 u64 flags, ip;
25 u64 ax, bx, cx, dx;
26 u64 si, di, bp, sp;
27 u64 r8, r9, r10, r11;
28 u64 r12, r13, r14, r15;
31 struct pebs_record_nhm {
32 u64 flags, ip;
33 u64 ax, bx, cx, dx;
34 u64 si, di, bp, sp;
35 u64 r8, r9, r10, r11;
36 u64 r12, r13, r14, r15;
37 u64 status, dla, dse, lat;
41 * A debug store configuration.
43 * We only support architectures that use 64bit fields.
45 struct debug_store {
46 u64 bts_buffer_base;
47 u64 bts_index;
48 u64 bts_absolute_maximum;
49 u64 bts_interrupt_threshold;
50 u64 pebs_buffer_base;
51 u64 pebs_index;
52 u64 pebs_absolute_maximum;
53 u64 pebs_interrupt_threshold;
54 u64 pebs_event_reset[MAX_PEBS_EVENTS];
57 static void init_debug_store_on_cpu(int cpu)
59 struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
61 if (!ds)
62 return;
64 wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
65 (u32)((u64)(unsigned long)ds),
66 (u32)((u64)(unsigned long)ds >> 32));
69 static void fini_debug_store_on_cpu(int cpu)
71 if (!per_cpu(cpu_hw_events, cpu).ds)
72 return;
74 wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
77 static void release_ds_buffers(void)
79 int cpu;
81 if (!x86_pmu.bts && !x86_pmu.pebs)
82 return;
84 get_online_cpus();
86 for_each_online_cpu(cpu)
87 fini_debug_store_on_cpu(cpu);
89 for_each_possible_cpu(cpu) {
90 struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
92 if (!ds)
93 continue;
95 per_cpu(cpu_hw_events, cpu).ds = NULL;
97 kfree((void *)(unsigned long)ds->pebs_buffer_base);
98 kfree((void *)(unsigned long)ds->bts_buffer_base);
99 kfree(ds);
102 put_online_cpus();
105 static int reserve_ds_buffers(void)
107 int cpu, err = 0;
109 if (!x86_pmu.bts && !x86_pmu.pebs)
110 return 0;
112 get_online_cpus();
114 for_each_possible_cpu(cpu) {
115 struct debug_store *ds;
116 void *buffer;
117 int max, thresh;
119 err = -ENOMEM;
120 ds = kzalloc(sizeof(*ds), GFP_KERNEL);
121 if (unlikely(!ds))
122 break;
123 per_cpu(cpu_hw_events, cpu).ds = ds;
125 if (x86_pmu.bts) {
126 buffer = kzalloc(BTS_BUFFER_SIZE, GFP_KERNEL);
127 if (unlikely(!buffer))
128 break;
130 max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
131 thresh = max / 16;
133 ds->bts_buffer_base = (u64)(unsigned long)buffer;
134 ds->bts_index = ds->bts_buffer_base;
135 ds->bts_absolute_maximum = ds->bts_buffer_base +
136 max * BTS_RECORD_SIZE;
137 ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
138 thresh * BTS_RECORD_SIZE;
141 if (x86_pmu.pebs) {
142 buffer = kzalloc(PEBS_BUFFER_SIZE, GFP_KERNEL);
143 if (unlikely(!buffer))
144 break;
146 max = PEBS_BUFFER_SIZE / x86_pmu.pebs_record_size;
148 ds->pebs_buffer_base = (u64)(unsigned long)buffer;
149 ds->pebs_index = ds->pebs_buffer_base;
150 ds->pebs_absolute_maximum = ds->pebs_buffer_base +
151 max * x86_pmu.pebs_record_size;
153 * Always use single record PEBS
155 ds->pebs_interrupt_threshold = ds->pebs_buffer_base +
156 x86_pmu.pebs_record_size;
159 err = 0;
162 if (err)
163 release_ds_buffers();
164 else {
165 for_each_online_cpu(cpu)
166 init_debug_store_on_cpu(cpu);
169 put_online_cpus();
171 return err;
175 * BTS
178 static struct event_constraint bts_constraint =
179 EVENT_CONSTRAINT(0, 1ULL << X86_PMC_IDX_FIXED_BTS, 0);
181 static void intel_pmu_enable_bts(u64 config)
183 unsigned long debugctlmsr;
185 debugctlmsr = get_debugctlmsr();
187 debugctlmsr |= DEBUGCTLMSR_TR;
188 debugctlmsr |= DEBUGCTLMSR_BTS;
189 debugctlmsr |= DEBUGCTLMSR_BTINT;
191 if (!(config & ARCH_PERFMON_EVENTSEL_OS))
192 debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS;
194 if (!(config & ARCH_PERFMON_EVENTSEL_USR))
195 debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR;
197 update_debugctlmsr(debugctlmsr);
200 static void intel_pmu_disable_bts(void)
202 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
203 unsigned long debugctlmsr;
205 if (!cpuc->ds)
206 return;
208 debugctlmsr = get_debugctlmsr();
210 debugctlmsr &=
211 ~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT |
212 DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR);
214 update_debugctlmsr(debugctlmsr);
217 static void intel_pmu_drain_bts_buffer(void)
219 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
220 struct debug_store *ds = cpuc->ds;
221 struct bts_record {
222 u64 from;
223 u64 to;
224 u64 flags;
226 struct perf_event *event = cpuc->events[X86_PMC_IDX_FIXED_BTS];
227 struct bts_record *at, *top;
228 struct perf_output_handle handle;
229 struct perf_event_header header;
230 struct perf_sample_data data;
231 struct pt_regs regs;
233 if (!event)
234 return;
236 if (!ds)
237 return;
239 at = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
240 top = (struct bts_record *)(unsigned long)ds->bts_index;
242 if (top <= at)
243 return;
245 ds->bts_index = ds->bts_buffer_base;
247 perf_sample_data_init(&data, 0);
248 data.period = event->hw.last_period;
249 regs.ip = 0;
252 * Prepare a generic sample, i.e. fill in the invariant fields.
253 * We will overwrite the from and to address before we output
254 * the sample.
256 perf_prepare_sample(&header, &data, event, &regs);
258 if (perf_output_begin(&handle, event, header.size * (top - at), 1, 1))
259 return;
261 for (; at < top; at++) {
262 data.ip = at->from;
263 data.addr = at->to;
265 perf_output_sample(&handle, &header, &data, event);
268 perf_output_end(&handle);
270 /* There's new data available. */
271 event->hw.interrupts++;
272 event->pending_kill = POLL_IN;
276 * PEBS
279 static struct event_constraint intel_core_pebs_events[] = {
280 PEBS_EVENT_CONSTRAINT(0x00c0, 0x1), /* INSTR_RETIRED.ANY */
281 PEBS_EVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
282 PEBS_EVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
283 PEBS_EVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
284 PEBS_EVENT_CONSTRAINT(0x01cb, 0x1), /* MEM_LOAD_RETIRED.L1D_MISS */
285 PEBS_EVENT_CONSTRAINT(0x02cb, 0x1), /* MEM_LOAD_RETIRED.L1D_LINE_MISS */
286 PEBS_EVENT_CONSTRAINT(0x04cb, 0x1), /* MEM_LOAD_RETIRED.L2_MISS */
287 PEBS_EVENT_CONSTRAINT(0x08cb, 0x1), /* MEM_LOAD_RETIRED.L2_LINE_MISS */
288 PEBS_EVENT_CONSTRAINT(0x10cb, 0x1), /* MEM_LOAD_RETIRED.DTLB_MISS */
289 EVENT_CONSTRAINT_END
292 static struct event_constraint intel_nehalem_pebs_events[] = {
293 PEBS_EVENT_CONSTRAINT(0x00c0, 0xf), /* INSTR_RETIRED.ANY */
294 PEBS_EVENT_CONSTRAINT(0xfec1, 0xf), /* X87_OPS_RETIRED.ANY */
295 PEBS_EVENT_CONSTRAINT(0x00c5, 0xf), /* BR_INST_RETIRED.MISPRED */
296 PEBS_EVENT_CONSTRAINT(0x1fc7, 0xf), /* SIMD_INST_RETURED.ANY */
297 PEBS_EVENT_CONSTRAINT(0x01cb, 0xf), /* MEM_LOAD_RETIRED.L1D_MISS */
298 PEBS_EVENT_CONSTRAINT(0x02cb, 0xf), /* MEM_LOAD_RETIRED.L1D_LINE_MISS */
299 PEBS_EVENT_CONSTRAINT(0x04cb, 0xf), /* MEM_LOAD_RETIRED.L2_MISS */
300 PEBS_EVENT_CONSTRAINT(0x08cb, 0xf), /* MEM_LOAD_RETIRED.L2_LINE_MISS */
301 PEBS_EVENT_CONSTRAINT(0x10cb, 0xf), /* MEM_LOAD_RETIRED.DTLB_MISS */
302 EVENT_CONSTRAINT_END
305 static struct event_constraint *
306 intel_pebs_constraints(struct perf_event *event)
308 struct event_constraint *c;
310 if (!event->attr.precise)
311 return NULL;
313 if (x86_pmu.pebs_constraints) {
314 for_each_event_constraint(c, x86_pmu.pebs_constraints) {
315 if ((event->hw.config & c->cmask) == c->code)
316 return c;
320 return &emptyconstraint;
323 static void intel_pmu_pebs_enable(struct perf_event *event)
325 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
326 struct hw_perf_event *hwc = &event->hw;
328 hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
330 cpuc->pebs_enabled |= 1ULL << hwc->idx;
331 WARN_ON_ONCE(cpuc->enabled);
333 if (x86_pmu.intel_cap.pebs_trap)
334 intel_pmu_lbr_enable(event);
337 static void intel_pmu_pebs_disable(struct perf_event *event)
339 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
340 struct hw_perf_event *hwc = &event->hw;
342 cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
343 if (cpuc->enabled)
344 wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
346 hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
348 if (x86_pmu.intel_cap.pebs_trap)
349 intel_pmu_lbr_disable(event);
352 static void intel_pmu_pebs_enable_all(void)
354 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
356 if (cpuc->pebs_enabled)
357 wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
360 static void intel_pmu_pebs_disable_all(void)
362 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
364 if (cpuc->pebs_enabled)
365 wrmsrl(MSR_IA32_PEBS_ENABLE, 0);
368 #include <asm/insn.h>
370 static inline bool kernel_ip(unsigned long ip)
372 #ifdef CONFIG_X86_32
373 return ip > PAGE_OFFSET;
374 #else
375 return (long)ip < 0;
376 #endif
379 static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
381 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
382 unsigned long from = cpuc->lbr_entries[0].from;
383 unsigned long old_to, to = cpuc->lbr_entries[0].to;
384 unsigned long ip = regs->ip;
387 * We don't need to fixup if the PEBS assist is fault like
389 if (!x86_pmu.intel_cap.pebs_trap)
390 return 1;
393 * No LBR entry, no basic block, no rewinding
395 if (!cpuc->lbr_stack.nr || !from || !to)
396 return 0;
399 * Basic blocks should never cross user/kernel boundaries
401 if (kernel_ip(ip) != kernel_ip(to))
402 return 0;
405 * unsigned math, either ip is before the start (impossible) or
406 * the basic block is larger than 1 page (sanity)
408 if ((ip - to) > PAGE_SIZE)
409 return 0;
412 * We sampled a branch insn, rewind using the LBR stack
414 if (ip == to) {
415 regs->ip = from;
416 return 1;
419 do {
420 struct insn insn;
421 u8 buf[MAX_INSN_SIZE];
422 void *kaddr;
424 old_to = to;
425 if (!kernel_ip(ip)) {
426 int bytes, size = MAX_INSN_SIZE;
428 bytes = copy_from_user_nmi(buf, (void __user *)to, size);
429 if (bytes != size)
430 return 0;
432 kaddr = buf;
433 } else
434 kaddr = (void *)to;
436 kernel_insn_init(&insn, kaddr);
437 insn_get_length(&insn);
438 to += insn.length;
439 } while (to < ip);
441 if (to == ip) {
442 regs->ip = old_to;
443 return 1;
447 * Even though we decoded the basic block, the instruction stream
448 * never matched the given IP, either the TO or the IP got corrupted.
450 return 0;
453 static int intel_pmu_save_and_restart(struct perf_event *event);
455 static void intel_pmu_drain_pebs_core(struct pt_regs *iregs)
457 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
458 struct debug_store *ds = cpuc->ds;
459 struct perf_event *event = cpuc->events[0]; /* PMC0 only */
460 struct pebs_record_core *at, *top;
461 struct perf_sample_data data;
462 struct perf_raw_record raw;
463 struct pt_regs regs;
464 int n;
466 if (!ds || !x86_pmu.pebs)
467 return;
469 at = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
470 top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
473 * Whatever else happens, drain the thing
475 ds->pebs_index = ds->pebs_buffer_base;
477 if (!test_bit(0, cpuc->active_mask))
478 return;
480 WARN_ON_ONCE(!event);
482 if (!event->attr.precise)
483 return;
485 n = top - at;
486 if (n <= 0)
487 return;
489 if (!intel_pmu_save_and_restart(event))
490 return;
493 * Should not happen, we program the threshold at 1 and do not
494 * set a reset value.
496 WARN_ON_ONCE(n > 1);
497 at += n - 1;
499 perf_sample_data_init(&data, 0);
500 data.period = event->hw.last_period;
502 if (event->attr.sample_type & PERF_SAMPLE_RAW) {
503 raw.size = x86_pmu.pebs_record_size;
504 raw.data = at;
505 data.raw = &raw;
509 * We use the interrupt regs as a base because the PEBS record
510 * does not contain a full regs set, specifically it seems to
511 * lack segment descriptors, which get used by things like
512 * user_mode().
514 * In the simple case fix up only the IP and BP,SP regs, for
515 * PERF_SAMPLE_IP and PERF_SAMPLE_CALLCHAIN to function properly.
516 * A possible PERF_SAMPLE_REGS will have to transfer all regs.
518 regs = *iregs;
519 regs.ip = at->ip;
520 regs.bp = at->bp;
521 regs.sp = at->sp;
523 if (intel_pmu_pebs_fixup_ip(&regs))
524 regs.flags |= PERF_EFLAGS_EXACT;
525 else
526 regs.flags &= ~PERF_EFLAGS_EXACT;
528 if (perf_event_overflow(event, 1, &data, &regs))
529 x86_pmu_stop(event);
532 static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs)
534 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
535 struct debug_store *ds = cpuc->ds;
536 struct pebs_record_nhm *at, *top;
537 struct perf_sample_data data;
538 struct perf_event *event = NULL;
539 struct perf_raw_record raw;
540 struct pt_regs regs;
541 u64 status = 0;
542 int bit, n;
544 if (!ds || !x86_pmu.pebs)
545 return;
547 at = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
548 top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
550 ds->pebs_index = ds->pebs_buffer_base;
552 n = top - at;
553 if (n <= 0)
554 return;
557 * Should not happen, we program the threshold at 1 and do not
558 * set a reset value.
560 WARN_ON_ONCE(n > MAX_PEBS_EVENTS);
562 for ( ; at < top; at++) {
563 for_each_set_bit(bit, (unsigned long *)&at->status, MAX_PEBS_EVENTS) {
564 event = cpuc->events[bit];
565 if (!test_bit(bit, cpuc->active_mask))
566 continue;
568 WARN_ON_ONCE(!event);
570 if (!event->attr.precise)
571 continue;
573 if (__test_and_set_bit(bit, (unsigned long *)&status))
574 continue;
576 break;
579 if (!event || bit >= MAX_PEBS_EVENTS)
580 continue;
582 if (!intel_pmu_save_and_restart(event))
583 continue;
585 perf_sample_data_init(&data, 0);
586 data.period = event->hw.last_period;
588 if (event->attr.sample_type & PERF_SAMPLE_RAW) {
589 raw.size = x86_pmu.pebs_record_size;
590 raw.data = at;
591 data.raw = &raw;
595 * See the comment in intel_pmu_drain_pebs_core()
597 regs = *iregs;
598 regs.ip = at->ip;
599 regs.bp = at->bp;
600 regs.sp = at->sp;
602 if (intel_pmu_pebs_fixup_ip(&regs))
603 regs.flags |= PERF_EFLAGS_EXACT;
604 else
605 regs.flags &= ~PERF_EFLAGS_EXACT;
607 if (perf_event_overflow(event, 1, &data, &regs))
608 x86_pmu_stop(event);
613 * BTS, PEBS probe and setup
616 static void intel_ds_init(void)
619 * No support for 32bit formats
621 if (!boot_cpu_has(X86_FEATURE_DTES64))
622 return;
624 x86_pmu.bts = boot_cpu_has(X86_FEATURE_BTS);
625 x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
626 if (x86_pmu.pebs) {
627 char pebs_type = x86_pmu.intel_cap.pebs_trap ? '+' : '-';
628 int format = x86_pmu.intel_cap.pebs_format;
630 switch (format) {
631 case 0:
632 printk(KERN_CONT "PEBS fmt0%c, ", pebs_type);
633 x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
634 x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
635 x86_pmu.pebs_constraints = intel_core_pebs_events;
636 break;
638 case 1:
639 printk(KERN_CONT "PEBS fmt1%c, ", pebs_type);
640 x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
641 x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
642 x86_pmu.pebs_constraints = intel_nehalem_pebs_events;
643 break;
645 default:
646 printk(KERN_CONT "no PEBS fmt%d%c, ", format, pebs_type);
647 x86_pmu.pebs = 0;
648 break;
653 #else /* CONFIG_CPU_SUP_INTEL */
655 static int reserve_ds_buffers(void)
657 return 0;
660 static void release_ds_buffers(void)
664 #endif /* CONFIG_CPU_SUP_INTEL */