perf_event_intel_ds.c 14 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642
  1. #ifdef CONFIG_CPU_SUP_INTEL
  2. /* The maximal number of PEBS events: */
  3. #define MAX_PEBS_EVENTS 4
  4. /* The size of a BTS record in bytes: */
  5. #define BTS_RECORD_SIZE 24
  6. #define BTS_BUFFER_SIZE (PAGE_SIZE << 4)
  7. #define PEBS_BUFFER_SIZE PAGE_SIZE
  8. /*
  9. * pebs_record_32 for p4 and core not supported
  10. struct pebs_record_32 {
  11. u32 flags, ip;
  12. u32 ax, bc, cx, dx;
  13. u32 si, di, bp, sp;
  14. };
  15. */
  16. struct pebs_record_core {
  17. u64 flags, ip;
  18. u64 ax, bx, cx, dx;
  19. u64 si, di, bp, sp;
  20. u64 r8, r9, r10, r11;
  21. u64 r12, r13, r14, r15;
  22. };
  23. struct pebs_record_nhm {
  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;
  29. u64 status, dla, dse, lat;
  30. };
  31. /*
  32. * A debug store configuration.
  33. *
  34. * We only support architectures that use 64bit fields.
  35. */
  36. struct debug_store {
  37. u64 bts_buffer_base;
  38. u64 bts_index;
  39. u64 bts_absolute_maximum;
  40. u64 bts_interrupt_threshold;
  41. u64 pebs_buffer_base;
  42. u64 pebs_index;
  43. u64 pebs_absolute_maximum;
  44. u64 pebs_interrupt_threshold;
  45. u64 pebs_event_reset[MAX_PEBS_EVENTS];
  46. };
  47. static void init_debug_store_on_cpu(int cpu)
  48. {
  49. struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
  50. if (!ds)
  51. return;
  52. wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
  53. (u32)((u64)(unsigned long)ds),
  54. (u32)((u64)(unsigned long)ds >> 32));
  55. }
  56. static void fini_debug_store_on_cpu(int cpu)
  57. {
  58. if (!per_cpu(cpu_hw_events, cpu).ds)
  59. return;
  60. wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
  61. }
  62. static void release_ds_buffers(void)
  63. {
  64. int cpu;
  65. if (!x86_pmu.bts && !x86_pmu.pebs)
  66. return;
  67. get_online_cpus();
  68. for_each_online_cpu(cpu)
  69. fini_debug_store_on_cpu(cpu);
  70. for_each_possible_cpu(cpu) {
  71. struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
  72. if (!ds)
  73. continue;
  74. per_cpu(cpu_hw_events, cpu).ds = NULL;
  75. kfree((void *)(unsigned long)ds->pebs_buffer_base);
  76. kfree((void *)(unsigned long)ds->bts_buffer_base);
  77. kfree(ds);
  78. }
  79. put_online_cpus();
  80. }
  81. static int reserve_ds_buffers(void)
  82. {
  83. int cpu, err = 0;
  84. if (!x86_pmu.bts && !x86_pmu.pebs)
  85. return 0;
  86. get_online_cpus();
  87. for_each_possible_cpu(cpu) {
  88. struct debug_store *ds;
  89. void *buffer;
  90. int max, thresh;
  91. err = -ENOMEM;
  92. ds = kzalloc(sizeof(*ds), GFP_KERNEL);
  93. if (unlikely(!ds))
  94. break;
  95. per_cpu(cpu_hw_events, cpu).ds = ds;
  96. if (x86_pmu.bts) {
  97. buffer = kzalloc(BTS_BUFFER_SIZE, GFP_KERNEL);
  98. if (unlikely(!buffer))
  99. break;
  100. max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
  101. thresh = max / 16;
  102. ds->bts_buffer_base = (u64)(unsigned long)buffer;
  103. ds->bts_index = ds->bts_buffer_base;
  104. ds->bts_absolute_maximum = ds->bts_buffer_base +
  105. max * BTS_RECORD_SIZE;
  106. ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
  107. thresh * BTS_RECORD_SIZE;
  108. }
  109. if (x86_pmu.pebs) {
  110. buffer = kzalloc(PEBS_BUFFER_SIZE, GFP_KERNEL);
  111. if (unlikely(!buffer))
  112. break;
  113. max = PEBS_BUFFER_SIZE / x86_pmu.pebs_record_size;
  114. ds->pebs_buffer_base = (u64)(unsigned long)buffer;
  115. ds->pebs_index = ds->pebs_buffer_base;
  116. ds->pebs_absolute_maximum = ds->pebs_buffer_base +
  117. max * x86_pmu.pebs_record_size;
  118. /*
  119. * Always use single record PEBS
  120. */
  121. ds->pebs_interrupt_threshold = ds->pebs_buffer_base +
  122. x86_pmu.pebs_record_size;
  123. }
  124. err = 0;
  125. }
  126. if (err)
  127. release_ds_buffers();
  128. else {
  129. for_each_online_cpu(cpu)
  130. init_debug_store_on_cpu(cpu);
  131. }
  132. put_online_cpus();
  133. return err;
  134. }
  135. /*
  136. * BTS
  137. */
  138. static struct event_constraint bts_constraint =
  139. EVENT_CONSTRAINT(0, 1ULL << X86_PMC_IDX_FIXED_BTS, 0);
  140. static void intel_pmu_enable_bts(u64 config)
  141. {
  142. unsigned long debugctlmsr;
  143. debugctlmsr = get_debugctlmsr();
  144. debugctlmsr |= DEBUGCTLMSR_TR;
  145. debugctlmsr |= DEBUGCTLMSR_BTS;
  146. debugctlmsr |= DEBUGCTLMSR_BTINT;
  147. if (!(config & ARCH_PERFMON_EVENTSEL_OS))
  148. debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS;
  149. if (!(config & ARCH_PERFMON_EVENTSEL_USR))
  150. debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR;
  151. update_debugctlmsr(debugctlmsr);
  152. }
  153. static void intel_pmu_disable_bts(void)
  154. {
  155. struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
  156. unsigned long debugctlmsr;
  157. if (!cpuc->ds)
  158. return;
  159. debugctlmsr = get_debugctlmsr();
  160. debugctlmsr &=
  161. ~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT |
  162. DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR);
  163. update_debugctlmsr(debugctlmsr);
  164. }
  165. static int intel_pmu_drain_bts_buffer(void)
  166. {
  167. struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
  168. struct debug_store *ds = cpuc->ds;
  169. struct bts_record {
  170. u64 from;
  171. u64 to;
  172. u64 flags;
  173. };
  174. struct perf_event *event = cpuc->events[X86_PMC_IDX_FIXED_BTS];
  175. struct bts_record *at, *top;
  176. struct perf_output_handle handle;
  177. struct perf_event_header header;
  178. struct perf_sample_data data;
  179. struct pt_regs regs;
  180. if (!event)
  181. return 0;
  182. if (!ds)
  183. return 0;
  184. at = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
  185. top = (struct bts_record *)(unsigned long)ds->bts_index;
  186. if (top <= at)
  187. return 0;
  188. ds->bts_index = ds->bts_buffer_base;
  189. perf_sample_data_init(&data, 0);
  190. data.period = event->hw.last_period;
  191. regs.ip = 0;
  192. /*
  193. * Prepare a generic sample, i.e. fill in the invariant fields.
  194. * We will overwrite the from and to address before we output
  195. * the sample.
  196. */
  197. perf_prepare_sample(&header, &data, event, &regs);
  198. if (perf_output_begin(&handle, event, header.size * (top - at), 1, 1))
  199. return 1;
  200. for (; at < top; at++) {
  201. data.ip = at->from;
  202. data.addr = at->to;
  203. perf_output_sample(&handle, &header, &data, event);
  204. }
  205. perf_output_end(&handle);
  206. /* There's new data available. */
  207. event->hw.interrupts++;
  208. event->pending_kill = POLL_IN;
  209. return 1;
  210. }
  211. /*
  212. * PEBS
  213. */
  214. static struct event_constraint intel_core_pebs_events[] = {
  215. PEBS_EVENT_CONSTRAINT(0x00c0, 0x1), /* INSTR_RETIRED.ANY */
  216. PEBS_EVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
  217. PEBS_EVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
  218. PEBS_EVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
  219. PEBS_EVENT_CONSTRAINT(0x01cb, 0x1), /* MEM_LOAD_RETIRED.L1D_MISS */
  220. PEBS_EVENT_CONSTRAINT(0x02cb, 0x1), /* MEM_LOAD_RETIRED.L1D_LINE_MISS */
  221. PEBS_EVENT_CONSTRAINT(0x04cb, 0x1), /* MEM_LOAD_RETIRED.L2_MISS */
  222. PEBS_EVENT_CONSTRAINT(0x08cb, 0x1), /* MEM_LOAD_RETIRED.L2_LINE_MISS */
  223. PEBS_EVENT_CONSTRAINT(0x10cb, 0x1), /* MEM_LOAD_RETIRED.DTLB_MISS */
  224. EVENT_CONSTRAINT_END
  225. };
  226. static struct event_constraint intel_nehalem_pebs_events[] = {
  227. PEBS_EVENT_CONSTRAINT(0x00c0, 0xf), /* INSTR_RETIRED.ANY */
  228. PEBS_EVENT_CONSTRAINT(0xfec1, 0xf), /* X87_OPS_RETIRED.ANY */
  229. PEBS_EVENT_CONSTRAINT(0x00c5, 0xf), /* BR_INST_RETIRED.MISPRED */
  230. PEBS_EVENT_CONSTRAINT(0x1fc7, 0xf), /* SIMD_INST_RETURED.ANY */
  231. PEBS_EVENT_CONSTRAINT(0x01cb, 0xf), /* MEM_LOAD_RETIRED.L1D_MISS */
  232. PEBS_EVENT_CONSTRAINT(0x02cb, 0xf), /* MEM_LOAD_RETIRED.L1D_LINE_MISS */
  233. PEBS_EVENT_CONSTRAINT(0x04cb, 0xf), /* MEM_LOAD_RETIRED.L2_MISS */
  234. PEBS_EVENT_CONSTRAINT(0x08cb, 0xf), /* MEM_LOAD_RETIRED.L2_LINE_MISS */
  235. PEBS_EVENT_CONSTRAINT(0x10cb, 0xf), /* MEM_LOAD_RETIRED.DTLB_MISS */
  236. EVENT_CONSTRAINT_END
  237. };
  238. static struct event_constraint *
  239. intel_pebs_constraints(struct perf_event *event)
  240. {
  241. struct event_constraint *c;
  242. if (!event->attr.precise_ip)
  243. return NULL;
  244. if (x86_pmu.pebs_constraints) {
  245. for_each_event_constraint(c, x86_pmu.pebs_constraints) {
  246. if ((event->hw.config & c->cmask) == c->code)
  247. return c;
  248. }
  249. }
  250. return &emptyconstraint;
  251. }
  252. static void intel_pmu_pebs_enable(struct perf_event *event)
  253. {
  254. struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
  255. struct hw_perf_event *hwc = &event->hw;
  256. hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
  257. cpuc->pebs_enabled |= 1ULL << hwc->idx;
  258. WARN_ON_ONCE(cpuc->enabled);
  259. if (x86_pmu.intel_cap.pebs_trap && event->attr.precise_ip > 1)
  260. intel_pmu_lbr_enable(event);
  261. }
  262. static void intel_pmu_pebs_disable(struct perf_event *event)
  263. {
  264. struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
  265. struct hw_perf_event *hwc = &event->hw;
  266. cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
  267. if (cpuc->enabled)
  268. wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
  269. hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
  270. if (x86_pmu.intel_cap.pebs_trap && event->attr.precise_ip > 1)
  271. intel_pmu_lbr_disable(event);
  272. }
  273. static void intel_pmu_pebs_enable_all(void)
  274. {
  275. struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
  276. if (cpuc->pebs_enabled)
  277. wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
  278. }
  279. static void intel_pmu_pebs_disable_all(void)
  280. {
  281. struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
  282. if (cpuc->pebs_enabled)
  283. wrmsrl(MSR_IA32_PEBS_ENABLE, 0);
  284. }
  285. #include <asm/insn.h>
  286. static inline bool kernel_ip(unsigned long ip)
  287. {
  288. #ifdef CONFIG_X86_32
  289. return ip > PAGE_OFFSET;
  290. #else
  291. return (long)ip < 0;
  292. #endif
  293. }
  294. static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
  295. {
  296. struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
  297. unsigned long from = cpuc->lbr_entries[0].from;
  298. unsigned long old_to, to = cpuc->lbr_entries[0].to;
  299. unsigned long ip = regs->ip;
  300. /*
  301. * We don't need to fixup if the PEBS assist is fault like
  302. */
  303. if (!x86_pmu.intel_cap.pebs_trap)
  304. return 1;
  305. /*
  306. * No LBR entry, no basic block, no rewinding
  307. */
  308. if (!cpuc->lbr_stack.nr || !from || !to)
  309. return 0;
  310. /*
  311. * Basic blocks should never cross user/kernel boundaries
  312. */
  313. if (kernel_ip(ip) != kernel_ip(to))
  314. return 0;
  315. /*
  316. * unsigned math, either ip is before the start (impossible) or
  317. * the basic block is larger than 1 page (sanity)
  318. */
  319. if ((ip - to) > PAGE_SIZE)
  320. return 0;
  321. /*
  322. * We sampled a branch insn, rewind using the LBR stack
  323. */
  324. if (ip == to) {
  325. regs->ip = from;
  326. return 1;
  327. }
  328. do {
  329. struct insn insn;
  330. u8 buf[MAX_INSN_SIZE];
  331. void *kaddr;
  332. old_to = to;
  333. if (!kernel_ip(ip)) {
  334. int bytes, size = MAX_INSN_SIZE;
  335. bytes = copy_from_user_nmi(buf, (void __user *)to, size);
  336. if (bytes != size)
  337. return 0;
  338. kaddr = buf;
  339. } else
  340. kaddr = (void *)to;
  341. kernel_insn_init(&insn, kaddr);
  342. insn_get_length(&insn);
  343. to += insn.length;
  344. } while (to < ip);
  345. if (to == ip) {
  346. regs->ip = old_to;
  347. return 1;
  348. }
  349. /*
  350. * Even though we decoded the basic block, the instruction stream
  351. * never matched the given IP, either the TO or the IP got corrupted.
  352. */
  353. return 0;
  354. }
  355. static int intel_pmu_save_and_restart(struct perf_event *event);
  356. static void __intel_pmu_pebs_event(struct perf_event *event,
  357. struct pt_regs *iregs, void *__pebs)
  358. {
  359. /*
  360. * We cast to pebs_record_core since that is a subset of
  361. * both formats and we don't use the other fields in this
  362. * routine.
  363. */
  364. struct pebs_record_core *pebs = __pebs;
  365. struct perf_sample_data data;
  366. struct pt_regs regs;
  367. if (!intel_pmu_save_and_restart(event))
  368. return;
  369. perf_sample_data_init(&data, 0);
  370. data.period = event->hw.last_period;
  371. /*
  372. * We use the interrupt regs as a base because the PEBS record
  373. * does not contain a full regs set, specifically it seems to
  374. * lack segment descriptors, which get used by things like
  375. * user_mode().
  376. *
  377. * In the simple case fix up only the IP and BP,SP regs, for
  378. * PERF_SAMPLE_IP and PERF_SAMPLE_CALLCHAIN to function properly.
  379. * A possible PERF_SAMPLE_REGS will have to transfer all regs.
  380. */
  381. regs = *iregs;
  382. regs.ip = pebs->ip;
  383. regs.bp = pebs->bp;
  384. regs.sp = pebs->sp;
  385. if (event->attr.precise_ip > 1 && intel_pmu_pebs_fixup_ip(&regs))
  386. regs.flags |= PERF_EFLAGS_EXACT;
  387. else
  388. regs.flags &= ~PERF_EFLAGS_EXACT;
  389. if (perf_event_overflow(event, 1, &data, &regs))
  390. x86_pmu_stop(event, 0);
  391. }
  392. static void intel_pmu_drain_pebs_core(struct pt_regs *iregs)
  393. {
  394. struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
  395. struct debug_store *ds = cpuc->ds;
  396. struct perf_event *event = cpuc->events[0]; /* PMC0 only */
  397. struct pebs_record_core *at, *top;
  398. int n;
  399. if (!ds || !x86_pmu.pebs)
  400. return;
  401. at = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
  402. top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
  403. /*
  404. * Whatever else happens, drain the thing
  405. */
  406. ds->pebs_index = ds->pebs_buffer_base;
  407. if (!test_bit(0, cpuc->active_mask))
  408. return;
  409. WARN_ON_ONCE(!event);
  410. if (!event->attr.precise_ip)
  411. return;
  412. n = top - at;
  413. if (n <= 0)
  414. return;
  415. /*
  416. * Should not happen, we program the threshold at 1 and do not
  417. * set a reset value.
  418. */
  419. WARN_ON_ONCE(n > 1);
  420. at += n - 1;
  421. __intel_pmu_pebs_event(event, iregs, at);
  422. }
  423. static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs)
  424. {
  425. struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
  426. struct debug_store *ds = cpuc->ds;
  427. struct pebs_record_nhm *at, *top;
  428. struct perf_event *event = NULL;
  429. u64 status = 0;
  430. int bit, n;
  431. if (!ds || !x86_pmu.pebs)
  432. return;
  433. at = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
  434. top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
  435. ds->pebs_index = ds->pebs_buffer_base;
  436. n = top - at;
  437. if (n <= 0)
  438. return;
  439. /*
  440. * Should not happen, we program the threshold at 1 and do not
  441. * set a reset value.
  442. */
  443. WARN_ON_ONCE(n > MAX_PEBS_EVENTS);
  444. for ( ; at < top; at++) {
  445. for_each_set_bit(bit, (unsigned long *)&at->status, MAX_PEBS_EVENTS) {
  446. event = cpuc->events[bit];
  447. if (!test_bit(bit, cpuc->active_mask))
  448. continue;
  449. WARN_ON_ONCE(!event);
  450. if (!event->attr.precise_ip)
  451. continue;
  452. if (__test_and_set_bit(bit, (unsigned long *)&status))
  453. continue;
  454. break;
  455. }
  456. if (!event || bit >= MAX_PEBS_EVENTS)
  457. continue;
  458. __intel_pmu_pebs_event(event, iregs, at);
  459. }
  460. }
  461. /*
  462. * BTS, PEBS probe and setup
  463. */
  464. static void intel_ds_init(void)
  465. {
  466. /*
  467. * No support for 32bit formats
  468. */
  469. if (!boot_cpu_has(X86_FEATURE_DTES64))
  470. return;
  471. x86_pmu.bts = boot_cpu_has(X86_FEATURE_BTS);
  472. x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
  473. if (x86_pmu.pebs) {
  474. char pebs_type = x86_pmu.intel_cap.pebs_trap ? '+' : '-';
  475. int format = x86_pmu.intel_cap.pebs_format;
  476. switch (format) {
  477. case 0:
  478. printk(KERN_CONT "PEBS fmt0%c, ", pebs_type);
  479. x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
  480. x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
  481. x86_pmu.pebs_constraints = intel_core_pebs_events;
  482. break;
  483. case 1:
  484. printk(KERN_CONT "PEBS fmt1%c, ", pebs_type);
  485. x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
  486. x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
  487. x86_pmu.pebs_constraints = intel_nehalem_pebs_events;
  488. break;
  489. default:
  490. printk(KERN_CONT "no PEBS fmt%d%c, ", format, pebs_type);
  491. x86_pmu.pebs = 0;
  492. break;
  493. }
  494. }
  495. }
  496. #else /* CONFIG_CPU_SUP_INTEL */
  497. static int reserve_ds_buffers(void)
  498. {
  499. return 0;
  500. }
  501. static void release_ds_buffers(void)
  502. {
  503. }
  504. #endif /* CONFIG_CPU_SUP_INTEL */