2 * Performance events - AMD IBS
4 * Copyright (C) 2011 Advanced Micro Devices, Inc., Robert Richter
6 * For licencing details see kernel-base/COPYING
9 #include <linux/perf_event.h>
10 #include <linux/module.h>
11 #include <linux/pci.h>
12 #include <linux/ptrace.h>
16 #include "perf_event.h"
20 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_AMD)
22 #include <linux/kprobes.h>
23 #include <linux/hardirq.h>
27 #define IBS_FETCH_CONFIG_MASK (IBS_FETCH_RAND_EN | IBS_FETCH_MAX_CNT)
28 #define IBS_OP_CONFIG_MASK IBS_OP_MAX_CNT
39 struct perf_event *event;
40 unsigned long state[BITS_TO_LONGS(IBS_MAX_STATES)];
51 unsigned long offset_mask[1];
53 struct cpu_perf_ibs __percpu *pcpu;
54 u64 (*get_count)(u64 config);
57 struct perf_ibs_data {
60 u32 data[0]; /* data buffer starts here */
63 u64 regs[MSR_AMD64_IBS_REG_COUNT_MAX];
67 perf_event_set_period(struct hw_perf_event *hwc, u64 min, u64 max, u64 *hw_period)
69 s64 left = local64_read(&hwc->period_left);
70 s64 period = hwc->sample_period;
74 * If we are way outside a reasonable range then just skip forward:
76 if (unlikely(left <= -period)) {
78 local64_set(&hwc->period_left, left);
79 hwc->last_period = period;
83 if (unlikely(left < (s64)min)) {
85 local64_set(&hwc->period_left, left);
86 hwc->last_period = period;
91 * If the hw period that triggers the sw overflow is too short
92 * we might hit the irq handler. This biases the results.
93 * Thus we shorten the next-to-last period and set the last
94 * period to the max period.
104 *hw_period = (u64)left;
110 perf_event_try_update(struct perf_event *event, u64 new_raw_count, int width)
112 struct hw_perf_event *hwc = &event->hw;
113 int shift = 64 - width;
118 * Careful: an NMI might modify the previous event value.
120 * Our tactic to handle this is to first atomically read and
121 * exchange a new raw count - then add that new-prev delta
122 * count to the generic event atomically:
124 prev_raw_count = local64_read(&hwc->prev_count);
125 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
126 new_raw_count) != prev_raw_count)
130 * Now we have the new raw value and have updated the prev
131 * timestamp already. We can now calculate the elapsed delta
132 * (event-)time and add that to the generic event.
134 * Careful, not all hw sign-extends above the physical width
137 delta = (new_raw_count << shift) - (prev_raw_count << shift);
140 local64_add(delta, &event->count);
141 local64_sub(delta, &hwc->period_left);
146 static struct perf_ibs perf_ibs_fetch;
147 static struct perf_ibs perf_ibs_op;
149 static struct perf_ibs *get_ibs_pmu(int type)
151 if (perf_ibs_fetch.pmu.type == type)
152 return &perf_ibs_fetch;
153 if (perf_ibs_op.pmu.type == type)
159 * Use IBS for precise event sampling:
161 * perf record -a -e cpu-cycles:p ... # use ibs op counting cycle count
162 * perf record -a -e r076:p ... # same as -e cpu-cycles:p
163 * perf record -a -e r0C1:p ... # use ibs op counting micro-ops
165 * IbsOpCntCtl (bit 19) of IBS Execution Control Register (IbsOpCtl,
166 * MSRC001_1033) is used to select either cycle or micro-ops counting
169 * The rip of IBS samples has skid 0. Thus, IBS supports precise
170 * levels 1 and 2 and the PERF_EFLAGS_EXACT is set. In rare cases the
171 * rip is invalid when IBS was not able to record the rip correctly.
172 * We clear PERF_EFLAGS_EXACT and take the rip from pt_regs then.
175 static int perf_ibs_precise_event(struct perf_event *event, u64 *config)
177 switch (event->attr.precise_ip) {
187 switch (event->attr.type) {
188 case PERF_TYPE_HARDWARE:
189 switch (event->attr.config) {
190 case PERF_COUNT_HW_CPU_CYCLES:
196 switch (event->attr.config) {
201 *config = IBS_OP_CNT_CTL;
212 static const struct perf_event_attr ibs_notsupp = {
221 static int perf_ibs_init(struct perf_event *event)
223 struct hw_perf_event *hwc = &event->hw;
224 struct perf_ibs *perf_ibs;
228 perf_ibs = get_ibs_pmu(event->attr.type);
230 config = event->attr.config;
232 perf_ibs = &perf_ibs_op;
233 ret = perf_ibs_precise_event(event, &config);
238 if (event->pmu != &perf_ibs->pmu)
241 if (perf_flags(&event->attr) & perf_flags(&ibs_notsupp))
244 if (config & ~perf_ibs->config_mask)
247 if (hwc->sample_period) {
248 if (config & perf_ibs->cnt_mask)
249 /* raw max_cnt may not be set */
251 if (!event->attr.sample_freq && hwc->sample_period & 0x0f)
253 * lower 4 bits can not be set in ibs max cnt,
254 * but allowing it in case we adjust the
255 * sample period to set a frequency.
258 hwc->sample_period &= ~0x0FULL;
259 if (!hwc->sample_period)
260 hwc->sample_period = 0x10;
262 max_cnt = config & perf_ibs->cnt_mask;
263 config &= ~perf_ibs->cnt_mask;
264 event->attr.sample_period = max_cnt << 4;
265 hwc->sample_period = event->attr.sample_period;
268 if (!hwc->sample_period)
272 * If we modify hwc->sample_period, we also need to update
273 * hwc->last_period and hwc->period_left.
275 hwc->last_period = hwc->sample_period;
276 local64_set(&hwc->period_left, hwc->sample_period);
278 hwc->config_base = perf_ibs->msr;
279 hwc->config = config;
284 static int perf_ibs_set_period(struct perf_ibs *perf_ibs,
285 struct hw_perf_event *hwc, u64 *period)
289 /* ignore lower 4 bits in min count: */
290 overflow = perf_event_set_period(hwc, 1<<4, perf_ibs->max_period, period);
291 local64_set(&hwc->prev_count, 0);
296 static u64 get_ibs_fetch_count(u64 config)
298 return (config & IBS_FETCH_CNT) >> 12;
301 static u64 get_ibs_op_count(u64 config)
305 if (config & IBS_OP_VAL)
306 count += (config & IBS_OP_MAX_CNT) << 4; /* cnt rolled over */
308 if (ibs_caps & IBS_CAPS_RDWROPCNT)
309 count += (config & IBS_OP_CUR_CNT) >> 32;
315 perf_ibs_event_update(struct perf_ibs *perf_ibs, struct perf_event *event,
318 u64 count = perf_ibs->get_count(*config);
321 * Set width to 64 since we do not overflow on max width but
322 * instead on max count. In perf_ibs_set_period() we clear
323 * prev count manually on overflow.
325 while (!perf_event_try_update(event, count, 64)) {
326 rdmsrl(event->hw.config_base, *config);
327 count = perf_ibs->get_count(*config);
331 static inline void perf_ibs_enable_event(struct perf_ibs *perf_ibs,
332 struct hw_perf_event *hwc, u64 config)
334 wrmsrl(hwc->config_base, hwc->config | config | perf_ibs->enable_mask);
338 * Erratum #420 Instruction-Based Sampling Engine May Generate
339 * Interrupt that Cannot Be Cleared:
341 * Must clear counter mask first, then clear the enable bit. See
342 * Revision Guide for AMD Family 10h Processors, Publication #41322.
344 static inline void perf_ibs_disable_event(struct perf_ibs *perf_ibs,
345 struct hw_perf_event *hwc, u64 config)
347 config &= ~perf_ibs->cnt_mask;
348 wrmsrl(hwc->config_base, config);
349 config &= ~perf_ibs->enable_mask;
350 wrmsrl(hwc->config_base, config);
354 * We cannot restore the ibs pmu state, so we always needs to update
355 * the event while stopping it and then reset the state when starting
356 * again. Thus, ignoring PERF_EF_RELOAD and PERF_EF_UPDATE flags in
357 * perf_ibs_start()/perf_ibs_stop() and instead always do it.
359 static void perf_ibs_start(struct perf_event *event, int flags)
361 struct hw_perf_event *hwc = &event->hw;
362 struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
363 struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
366 if (WARN_ON_ONCE(!(hwc->state & PERF_HES_STOPPED)))
369 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
372 perf_ibs_set_period(perf_ibs, hwc, &period);
373 set_bit(IBS_STARTED, pcpu->state);
374 perf_ibs_enable_event(perf_ibs, hwc, period >> 4);
376 perf_event_update_userpage(event);
379 static void perf_ibs_stop(struct perf_event *event, int flags)
381 struct hw_perf_event *hwc = &event->hw;
382 struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
383 struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
387 stopping = test_and_clear_bit(IBS_STARTED, pcpu->state);
389 if (!stopping && (hwc->state & PERF_HES_UPTODATE))
392 rdmsrl(hwc->config_base, config);
395 set_bit(IBS_STOPPING, pcpu->state);
396 perf_ibs_disable_event(perf_ibs, hwc, config);
397 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
398 hwc->state |= PERF_HES_STOPPED;
401 if (hwc->state & PERF_HES_UPTODATE)
405 * Clear valid bit to not count rollovers on update, rollovers
406 * are only updated in the irq handler.
408 config &= ~perf_ibs->valid_mask;
410 perf_ibs_event_update(perf_ibs, event, &config);
411 hwc->state |= PERF_HES_UPTODATE;
414 static int perf_ibs_add(struct perf_event *event, int flags)
416 struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
417 struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
419 if (test_and_set_bit(IBS_ENABLED, pcpu->state))
422 event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
426 if (flags & PERF_EF_START)
427 perf_ibs_start(event, PERF_EF_RELOAD);
432 static void perf_ibs_del(struct perf_event *event, int flags)
434 struct perf_ibs *perf_ibs = container_of(event->pmu, struct perf_ibs, pmu);
435 struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
437 if (!test_and_clear_bit(IBS_ENABLED, pcpu->state))
440 perf_ibs_stop(event, PERF_EF_UPDATE);
444 perf_event_update_userpage(event);
447 static void perf_ibs_read(struct perf_event *event) { }
449 static struct perf_ibs perf_ibs_fetch = {
451 .task_ctx_nr = perf_invalid_context,
453 .event_init = perf_ibs_init,
456 .start = perf_ibs_start,
457 .stop = perf_ibs_stop,
458 .read = perf_ibs_read,
460 .msr = MSR_AMD64_IBSFETCHCTL,
461 .config_mask = IBS_FETCH_CONFIG_MASK,
462 .cnt_mask = IBS_FETCH_MAX_CNT,
463 .enable_mask = IBS_FETCH_ENABLE,
464 .valid_mask = IBS_FETCH_VAL,
465 .max_period = IBS_FETCH_MAX_CNT << 4,
466 .offset_mask = { MSR_AMD64_IBSFETCH_REG_MASK },
467 .offset_max = MSR_AMD64_IBSFETCH_REG_COUNT,
469 .get_count = get_ibs_fetch_count,
472 static struct perf_ibs perf_ibs_op = {
474 .task_ctx_nr = perf_invalid_context,
476 .event_init = perf_ibs_init,
479 .start = perf_ibs_start,
480 .stop = perf_ibs_stop,
481 .read = perf_ibs_read,
483 .msr = MSR_AMD64_IBSOPCTL,
484 .config_mask = IBS_OP_CONFIG_MASK,
485 .cnt_mask = IBS_OP_MAX_CNT,
486 .enable_mask = IBS_OP_ENABLE,
487 .valid_mask = IBS_OP_VAL,
488 .max_period = IBS_OP_MAX_CNT << 4,
489 .offset_mask = { MSR_AMD64_IBSOP_REG_MASK },
490 .offset_max = MSR_AMD64_IBSOP_REG_COUNT,
492 .get_count = get_ibs_op_count,
495 static int perf_ibs_handle_irq(struct perf_ibs *perf_ibs, struct pt_regs *iregs)
497 struct cpu_perf_ibs *pcpu = this_cpu_ptr(perf_ibs->pcpu);
498 struct perf_event *event = pcpu->event;
499 struct hw_perf_event *hwc = &event->hw;
500 struct perf_sample_data data;
501 struct perf_raw_record raw;
503 struct perf_ibs_data ibs_data;
504 int offset, size, check_rip, offset_max, throttle = 0;
506 u64 *buf, *config, period;
508 if (!test_bit(IBS_STARTED, pcpu->state)) {
510 * Catch spurious interrupts after stopping IBS: After
511 * disabling IBS there could be still incomming NMIs
512 * with samples that even have the valid bit cleared.
513 * Mark all this NMIs as handled.
515 return test_and_clear_bit(IBS_STOPPING, pcpu->state) ? 1 : 0;
518 msr = hwc->config_base;
521 if (!(*buf++ & perf_ibs->valid_mask))
524 config = &ibs_data.regs[0];
525 perf_ibs_event_update(perf_ibs, event, config);
526 perf_sample_data_init(&data, 0, hwc->last_period);
527 if (!perf_ibs_set_period(perf_ibs, hwc, &period))
528 goto out; /* no sw counter overflow */
530 ibs_data.caps = ibs_caps;
533 check_rip = (perf_ibs == &perf_ibs_op && (ibs_caps & IBS_CAPS_RIPINVALIDCHK));
534 if (event->attr.sample_type & PERF_SAMPLE_RAW)
535 offset_max = perf_ibs->offset_max;
541 rdmsrl(msr + offset, *buf++);
543 offset = find_next_bit(perf_ibs->offset_mask,
544 perf_ibs->offset_max,
546 } while (offset < offset_max);
547 ibs_data.size = sizeof(u64) * size;
550 if (check_rip && (ibs_data.regs[2] & IBS_RIP_INVALID)) {
551 regs.flags &= ~PERF_EFLAGS_EXACT;
553 set_linear_ip(®s, ibs_data.regs[1]);
554 regs.flags |= PERF_EFLAGS_EXACT;
557 if (event->attr.sample_type & PERF_SAMPLE_RAW) {
558 raw.size = sizeof(u32) + ibs_data.size;
559 raw.data = ibs_data.data;
563 throttle = perf_event_overflow(event, &data, ®s);
566 perf_ibs_disable_event(perf_ibs, hwc, *config);
568 perf_ibs_enable_event(perf_ibs, hwc, period >> 4);
570 perf_event_update_userpage(event);
576 perf_ibs_nmi_handler(unsigned int cmd, struct pt_regs *regs)
580 handled += perf_ibs_handle_irq(&perf_ibs_fetch, regs);
581 handled += perf_ibs_handle_irq(&perf_ibs_op, regs);
584 inc_irq_stat(apic_perf_irqs);
589 static __init int perf_ibs_pmu_init(struct perf_ibs *perf_ibs, char *name)
591 struct cpu_perf_ibs __percpu *pcpu;
594 pcpu = alloc_percpu(struct cpu_perf_ibs);
598 perf_ibs->pcpu = pcpu;
600 ret = perf_pmu_register(&perf_ibs->pmu, name, -1);
602 perf_ibs->pcpu = NULL;
609 static __init int perf_event_ibs_init(void)
612 return -ENODEV; /* ibs not supported by the cpu */
614 perf_ibs_pmu_init(&perf_ibs_fetch, "ibs_fetch");
615 if (ibs_caps & IBS_CAPS_OPCNT)
616 perf_ibs_op.config_mask |= IBS_OP_CNT_CTL;
617 perf_ibs_pmu_init(&perf_ibs_op, "ibs_op");
618 register_nmi_handler(NMI_LOCAL, perf_ibs_nmi_handler, 0, "perf_ibs");
619 printk(KERN_INFO "perf: AMD IBS detected (0x%08x)\n", ibs_caps);
624 #else /* defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_AMD) */
626 static __init int perf_event_ibs_init(void) { return 0; }
630 /* IBS - apic initialization, for perf and oprofile */
632 static __init u32 __get_ibs_caps(void)
635 unsigned int max_level;
637 if (!boot_cpu_has(X86_FEATURE_IBS))
640 /* check IBS cpuid feature flags */
641 max_level = cpuid_eax(0x80000000);
642 if (max_level < IBS_CPUID_FEATURES)
643 return IBS_CAPS_DEFAULT;
645 caps = cpuid_eax(IBS_CPUID_FEATURES);
646 if (!(caps & IBS_CAPS_AVAIL))
647 /* cpuid flags not valid */
648 return IBS_CAPS_DEFAULT;
653 u32 get_ibs_caps(void)
658 EXPORT_SYMBOL(get_ibs_caps);
660 static inline int get_eilvt(int offset)
662 return !setup_APIC_eilvt(offset, 0, APIC_EILVT_MSG_NMI, 1);
665 static inline int put_eilvt(int offset)
667 return !setup_APIC_eilvt(offset, 0, 0, 1);
671 * Check and reserve APIC extended interrupt LVT offset for IBS if available.
673 static inline int ibs_eilvt_valid(void)
681 rdmsrl(MSR_AMD64_IBSCTL, val);
682 offset = val & IBSCTL_LVT_OFFSET_MASK;
684 if (!(val & IBSCTL_LVT_OFFSET_VALID)) {
685 pr_err(FW_BUG "cpu %d, invalid IBS interrupt offset %d (MSR%08X=0x%016llx)\n",
686 smp_processor_id(), offset, MSR_AMD64_IBSCTL, val);
690 if (!get_eilvt(offset)) {
691 pr_err(FW_BUG "cpu %d, IBS interrupt offset %d not available (MSR%08X=0x%016llx)\n",
692 smp_processor_id(), offset, MSR_AMD64_IBSCTL, val);
703 static int setup_ibs_ctl(int ibs_eilvt_off)
705 struct pci_dev *cpu_cfg;
712 cpu_cfg = pci_get_device(PCI_VENDOR_ID_AMD,
713 PCI_DEVICE_ID_AMD_10H_NB_MISC,
718 pci_write_config_dword(cpu_cfg, IBSCTL, ibs_eilvt_off
719 | IBSCTL_LVT_OFFSET_VALID);
720 pci_read_config_dword(cpu_cfg, IBSCTL, &value);
721 if (value != (ibs_eilvt_off | IBSCTL_LVT_OFFSET_VALID)) {
722 pci_dev_put(cpu_cfg);
723 printk(KERN_DEBUG "Failed to setup IBS LVT offset, "
724 "IBSCTL = 0x%08x\n", value);
730 printk(KERN_DEBUG "No CPU node configured for IBS\n");
738 * This runs only on the current cpu. We try to find an LVT offset and
739 * setup the local APIC. For this we must disable preemption. On
740 * success we initialize all nodes with this offset. This updates then
741 * the offset in the IBS_CTL per-node msr. The per-core APIC setup of
742 * the IBS interrupt vector is handled by perf_ibs_cpu_notifier that
743 * is using the new offset.
745 static int force_ibs_eilvt_setup(void)
751 /* find the next free available EILVT entry, skip offset 0 */
752 for (offset = 1; offset < APIC_EILVT_NR_MAX; offset++) {
753 if (get_eilvt(offset))
758 if (offset == APIC_EILVT_NR_MAX) {
759 printk(KERN_DEBUG "No EILVT entry available\n");
763 ret = setup_ibs_ctl(offset);
767 if (!ibs_eilvt_valid()) {
772 pr_info("IBS: LVT offset %d assigned\n", offset);
782 static inline int get_ibs_lvt_offset(void)
786 rdmsrl(MSR_AMD64_IBSCTL, val);
787 if (!(val & IBSCTL_LVT_OFFSET_VALID))
790 return val & IBSCTL_LVT_OFFSET_MASK;
793 static void setup_APIC_ibs(void *dummy)
797 offset = get_ibs_lvt_offset();
801 if (!setup_APIC_eilvt(offset, 0, APIC_EILVT_MSG_NMI, 0))
804 pr_warn("perf: IBS APIC setup failed on cpu #%d\n",
808 static void clear_APIC_ibs(void *dummy)
812 offset = get_ibs_lvt_offset();
814 setup_APIC_eilvt(offset, 0, APIC_EILVT_MSG_FIX, 1);
818 perf_ibs_cpu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
820 switch (action & ~CPU_TASKS_FROZEN) {
822 setup_APIC_ibs(NULL);
825 clear_APIC_ibs(NULL);
834 static __init int amd_ibs_init(void)
839 caps = __get_ibs_caps();
841 return -ENODEV; /* ibs not supported by the cpu */
844 * Force LVT offset assignment for family 10h: The offsets are
845 * not assigned by the BIOS for this family, so the OS is
846 * responsible for doing it. If the OS assignment fails, fall
847 * back to BIOS settings and try to setup this.
849 if (boot_cpu_data.x86 == 0x10)
850 force_ibs_eilvt_setup();
852 if (!ibs_eilvt_valid())
857 /* make ibs_caps visible to other cpus: */
859 perf_cpu_notifier(perf_ibs_cpu_notifier);
860 smp_call_function(setup_APIC_ibs, NULL, 1);
863 ret = perf_event_ibs_init();
866 pr_err("Failed to setup IBS, %d\n", ret);
870 /* Since we need the pci subsystem to init ibs we can't do this earlier: */
871 device_initcall(amd_ibs_init);