1#include <linux/bitops.h>
2#include <linux/types.h>
3#include <linux/slab.h>
4
5#include <asm/perf_event.h>
6#include <asm/insn.h>
7
8#include "perf_event.h"
9
10/* The size of a BTS record in bytes: */
11#define BTS_RECORD_SIZE		24
12
13#define BTS_BUFFER_SIZE		(PAGE_SIZE << 4)
14#define PEBS_BUFFER_SIZE	(PAGE_SIZE << 4)
15#define PEBS_FIXUP_SIZE		PAGE_SIZE
16
17/*
18 * pebs_record_32 for p4 and core not supported
19
20struct pebs_record_32 {
21	u32 flags, ip;
22	u32 ax, bc, cx, dx;
23	u32 si, di, bp, sp;
24};
25
26 */
27
28union intel_x86_pebs_dse {
29	u64 val;
30	struct {
31		unsigned int ld_dse:4;
32		unsigned int ld_stlb_miss:1;
33		unsigned int ld_locked:1;
34		unsigned int ld_reserved:26;
35	};
36	struct {
37		unsigned int st_l1d_hit:1;
38		unsigned int st_reserved1:3;
39		unsigned int st_stlb_miss:1;
40		unsigned int st_locked:1;
41		unsigned int st_reserved2:26;
42	};
43};
44
45
46/*
47 * Map PEBS Load Latency Data Source encodings to generic
48 * memory data source information
49 */
50#define P(a, b) PERF_MEM_S(a, b)
51#define OP_LH (P(OP, LOAD) | P(LVL, HIT))
52#define SNOOP_NONE_MISS (P(SNOOP, NONE) | P(SNOOP, MISS))
53
54/* Version for Sandy Bridge and later */
55static u64 pebs_data_source[] = {
56	P(OP, LOAD) | P(LVL, MISS) | P(LVL, L3) | P(SNOOP, NA),/* 0x00:ukn L3 */
57	OP_LH | P(LVL, L1)  | P(SNOOP, NONE),	/* 0x01: L1 local */
58	OP_LH | P(LVL, LFB) | P(SNOOP, NONE),	/* 0x02: LFB hit */
59	OP_LH | P(LVL, L2)  | P(SNOOP, NONE),	/* 0x03: L2 hit */
60	OP_LH | P(LVL, L3)  | P(SNOOP, NONE),	/* 0x04: L3 hit */
61	OP_LH | P(LVL, L3)  | P(SNOOP, MISS),	/* 0x05: L3 hit, snoop miss */
62	OP_LH | P(LVL, L3)  | P(SNOOP, HIT),	/* 0x06: L3 hit, snoop hit */
63	OP_LH | P(LVL, L3)  | P(SNOOP, HITM),	/* 0x07: L3 hit, snoop hitm */
64	OP_LH | P(LVL, REM_CCE1) | P(SNOOP, HIT),  /* 0x08: L3 miss snoop hit */
65	OP_LH | P(LVL, REM_CCE1) | P(SNOOP, HITM), /* 0x09: L3 miss snoop hitm*/
66	OP_LH | P(LVL, LOC_RAM)  | P(SNOOP, HIT),  /* 0x0a: L3 miss, shared */
67	OP_LH | P(LVL, REM_RAM1) | P(SNOOP, HIT),  /* 0x0b: L3 miss, shared */
68	OP_LH | P(LVL, LOC_RAM)  | SNOOP_NONE_MISS,/* 0x0c: L3 miss, excl */
69	OP_LH | P(LVL, REM_RAM1) | SNOOP_NONE_MISS,/* 0x0d: L3 miss, excl */
70	OP_LH | P(LVL, IO)  | P(SNOOP, NONE), /* 0x0e: I/O */
71	OP_LH | P(LVL, UNC) | P(SNOOP, NONE), /* 0x0f: uncached */
72};
73
74/* Patch up minor differences in the bits */
75void __init intel_pmu_pebs_data_source_nhm(void)
76{
77	pebs_data_source[0x05] = OP_LH | P(LVL, L3)  | P(SNOOP, HIT);
78	pebs_data_source[0x06] = OP_LH | P(LVL, L3)  | P(SNOOP, HITM);
79	pebs_data_source[0x07] = OP_LH | P(LVL, L3)  | P(SNOOP, HITM);
80}
81
82static u64 precise_store_data(u64 status)
83{
84	union intel_x86_pebs_dse dse;
85	u64 val = P(OP, STORE) | P(SNOOP, NA) | P(LVL, L1) | P(TLB, L2);
86
87	dse.val = status;
88
89	/*
90	 * bit 4: TLB access
91	 * 1 = stored missed 2nd level TLB
92	 *
93	 * so it either hit the walker or the OS
94	 * otherwise hit 2nd level TLB
95	 */
96	if (dse.st_stlb_miss)
97		val |= P(TLB, MISS);
98	else
99		val |= P(TLB, HIT);
100
101	/*
102	 * bit 0: hit L1 data cache
103	 * if not set, then all we know is that
104	 * it missed L1D
105	 */
106	if (dse.st_l1d_hit)
107		val |= P(LVL, HIT);
108	else
109		val |= P(LVL, MISS);
110
111	/*
112	 * bit 5: Locked prefix
113	 */
114	if (dse.st_locked)
115		val |= P(LOCK, LOCKED);
116
117	return val;
118}
119
120static u64 precise_datala_hsw(struct perf_event *event, u64 status)
121{
122	union perf_mem_data_src dse;
123
124	dse.val = PERF_MEM_NA;
125
126	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
127		dse.mem_op = PERF_MEM_OP_STORE;
128	else if (event->hw.flags & PERF_X86_EVENT_PEBS_LD_HSW)
129		dse.mem_op = PERF_MEM_OP_LOAD;
130
131	/*
132	 * L1 info only valid for following events:
133	 *
134	 * MEM_UOPS_RETIRED.STLB_MISS_STORES
135	 * MEM_UOPS_RETIRED.LOCK_STORES
136	 * MEM_UOPS_RETIRED.SPLIT_STORES
137	 * MEM_UOPS_RETIRED.ALL_STORES
138	 */
139	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW) {
140		if (status & 1)
141			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_HIT;
142		else
143			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_MISS;
144	}
145	return dse.val;
146}
147
148static u64 load_latency_data(u64 status)
149{
150	union intel_x86_pebs_dse dse;
151	u64 val;
152	int model = boot_cpu_data.x86_model;
153	int fam = boot_cpu_data.x86;
154
155	dse.val = status;
156
157	/*
158	 * use the mapping table for bit 0-3
159	 */
160	val = pebs_data_source[dse.ld_dse];
161
162	/*
163	 * Nehalem models do not support TLB, Lock infos
164	 */
165	if (fam == 0x6 && (model == 26 || model == 30
166	    || model == 31 || model == 46)) {
167		val |= P(TLB, NA) | P(LOCK, NA);
168		return val;
169	}
170	/*
171	 * bit 4: TLB access
172	 * 0 = did not miss 2nd level TLB
173	 * 1 = missed 2nd level TLB
174	 */
175	if (dse.ld_stlb_miss)
176		val |= P(TLB, MISS) | P(TLB, L2);
177	else
178		val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
179
180	/*
181	 * bit 5: locked prefix
182	 */
183	if (dse.ld_locked)
184		val |= P(LOCK, LOCKED);
185
186	return val;
187}
188
189struct pebs_record_core {
190	u64 flags, ip;
191	u64 ax, bx, cx, dx;
192	u64 si, di, bp, sp;
193	u64 r8,  r9,  r10, r11;
194	u64 r12, r13, r14, r15;
195};
196
197struct pebs_record_nhm {
198	u64 flags, ip;
199	u64 ax, bx, cx, dx;
200	u64 si, di, bp, sp;
201	u64 r8,  r9,  r10, r11;
202	u64 r12, r13, r14, r15;
203	u64 status, dla, dse, lat;
204};
205
206/*
207 * Same as pebs_record_nhm, with two additional fields.
208 */
209struct pebs_record_hsw {
210	u64 flags, ip;
211	u64 ax, bx, cx, dx;
212	u64 si, di, bp, sp;
213	u64 r8,  r9,  r10, r11;
214	u64 r12, r13, r14, r15;
215	u64 status, dla, dse, lat;
216	u64 real_ip, tsx_tuning;
217};
218
219union hsw_tsx_tuning {
220	struct {
221		u32 cycles_last_block     : 32,
222		    hle_abort		  : 1,
223		    rtm_abort		  : 1,
224		    instruction_abort     : 1,
225		    non_instruction_abort : 1,
226		    retry		  : 1,
227		    data_conflict	  : 1,
228		    capacity_writes	  : 1,
229		    capacity_reads	  : 1;
230	};
231	u64	    value;
232};
233
234#define PEBS_HSW_TSX_FLAGS	0xff00000000ULL
235
236/* Same as HSW, plus TSC */
237
238struct pebs_record_skl {
239	u64 flags, ip;
240	u64 ax, bx, cx, dx;
241	u64 si, di, bp, sp;
242	u64 r8,  r9,  r10, r11;
243	u64 r12, r13, r14, r15;
244	u64 status, dla, dse, lat;
245	u64 real_ip, tsx_tuning;
246	u64 tsc;
247};
248
249void init_debug_store_on_cpu(int cpu)
250{
251	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
252
253	if (!ds)
254		return;
255
256	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
257		     (u32)((u64)(unsigned long)ds),
258		     (u32)((u64)(unsigned long)ds >> 32));
259}
260
261void fini_debug_store_on_cpu(int cpu)
262{
263	if (!per_cpu(cpu_hw_events, cpu).ds)
264		return;
265
266	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
267}
268
269static DEFINE_PER_CPU(void *, insn_buffer);
270
271static int alloc_pebs_buffer(int cpu)
272{
273	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
274	int node = cpu_to_node(cpu);
275	int max;
276	void *buffer, *ibuffer;
277
278	if (!x86_pmu.pebs)
279		return 0;
280
281	buffer = kzalloc_node(x86_pmu.pebs_buffer_size, GFP_KERNEL, node);
282	if (unlikely(!buffer))
283		return -ENOMEM;
284
285	/*
286	 * HSW+ already provides us the eventing ip; no need to allocate this
287	 * buffer then.
288	 */
289	if (x86_pmu.intel_cap.pebs_format < 2) {
290		ibuffer = kzalloc_node(PEBS_FIXUP_SIZE, GFP_KERNEL, node);
291		if (!ibuffer) {
292			kfree(buffer);
293			return -ENOMEM;
294		}
295		per_cpu(insn_buffer, cpu) = ibuffer;
296	}
297
298	max = x86_pmu.pebs_buffer_size / x86_pmu.pebs_record_size;
299
300	ds->pebs_buffer_base = (u64)(unsigned long)buffer;
301	ds->pebs_index = ds->pebs_buffer_base;
302	ds->pebs_absolute_maximum = ds->pebs_buffer_base +
303		max * x86_pmu.pebs_record_size;
304
305	return 0;
306}
307
308static void release_pebs_buffer(int cpu)
309{
310	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
311
312	if (!ds || !x86_pmu.pebs)
313		return;
314
315	kfree(per_cpu(insn_buffer, cpu));
316	per_cpu(insn_buffer, cpu) = NULL;
317
318	kfree((void *)(unsigned long)ds->pebs_buffer_base);
319	ds->pebs_buffer_base = 0;
320}
321
322static int alloc_bts_buffer(int cpu)
323{
324	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
325	int node = cpu_to_node(cpu);
326	int max, thresh;
327	void *buffer;
328
329	if (!x86_pmu.bts)
330		return 0;
331
332	buffer = kzalloc_node(BTS_BUFFER_SIZE, GFP_KERNEL | __GFP_NOWARN, node);
333	if (unlikely(!buffer)) {
334		WARN_ONCE(1, "%s: BTS buffer allocation failure\n", __func__);
335		return -ENOMEM;
336	}
337
338	max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
339	thresh = max / 16;
340
341	ds->bts_buffer_base = (u64)(unsigned long)buffer;
342	ds->bts_index = ds->bts_buffer_base;
343	ds->bts_absolute_maximum = ds->bts_buffer_base +
344		max * BTS_RECORD_SIZE;
345	ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
346		thresh * BTS_RECORD_SIZE;
347
348	return 0;
349}
350
351static void release_bts_buffer(int cpu)
352{
353	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
354
355	if (!ds || !x86_pmu.bts)
356		return;
357
358	kfree((void *)(unsigned long)ds->bts_buffer_base);
359	ds->bts_buffer_base = 0;
360}
361
362static int alloc_ds_buffer(int cpu)
363{
364	int node = cpu_to_node(cpu);
365	struct debug_store *ds;
366
367	ds = kzalloc_node(sizeof(*ds), GFP_KERNEL, node);
368	if (unlikely(!ds))
369		return -ENOMEM;
370
371	per_cpu(cpu_hw_events, cpu).ds = ds;
372
373	return 0;
374}
375
376static void release_ds_buffer(int cpu)
377{
378	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
379
380	if (!ds)
381		return;
382
383	per_cpu(cpu_hw_events, cpu).ds = NULL;
384	kfree(ds);
385}
386
387void release_ds_buffers(void)
388{
389	int cpu;
390
391	if (!x86_pmu.bts && !x86_pmu.pebs)
392		return;
393
394	get_online_cpus();
395	for_each_online_cpu(cpu)
396		fini_debug_store_on_cpu(cpu);
397
398	for_each_possible_cpu(cpu) {
399		release_pebs_buffer(cpu);
400		release_bts_buffer(cpu);
401		release_ds_buffer(cpu);
402	}
403	put_online_cpus();
404}
405
406void reserve_ds_buffers(void)
407{
408	int bts_err = 0, pebs_err = 0;
409	int cpu;
410
411	x86_pmu.bts_active = 0;
412	x86_pmu.pebs_active = 0;
413
414	if (!x86_pmu.bts && !x86_pmu.pebs)
415		return;
416
417	if (!x86_pmu.bts)
418		bts_err = 1;
419
420	if (!x86_pmu.pebs)
421		pebs_err = 1;
422
423	get_online_cpus();
424
425	for_each_possible_cpu(cpu) {
426		if (alloc_ds_buffer(cpu)) {
427			bts_err = 1;
428			pebs_err = 1;
429		}
430
431		if (!bts_err && alloc_bts_buffer(cpu))
432			bts_err = 1;
433
434		if (!pebs_err && alloc_pebs_buffer(cpu))
435			pebs_err = 1;
436
437		if (bts_err && pebs_err)
438			break;
439	}
440
441	if (bts_err) {
442		for_each_possible_cpu(cpu)
443			release_bts_buffer(cpu);
444	}
445
446	if (pebs_err) {
447		for_each_possible_cpu(cpu)
448			release_pebs_buffer(cpu);
449	}
450
451	if (bts_err && pebs_err) {
452		for_each_possible_cpu(cpu)
453			release_ds_buffer(cpu);
454	} else {
455		if (x86_pmu.bts && !bts_err)
456			x86_pmu.bts_active = 1;
457
458		if (x86_pmu.pebs && !pebs_err)
459			x86_pmu.pebs_active = 1;
460
461		for_each_online_cpu(cpu)
462			init_debug_store_on_cpu(cpu);
463	}
464
465	put_online_cpus();
466}
467
468/*
469 * BTS
470 */
471
472struct event_constraint bts_constraint =
473	EVENT_CONSTRAINT(0, 1ULL << INTEL_PMC_IDX_FIXED_BTS, 0);
474
475void intel_pmu_enable_bts(u64 config)
476{
477	unsigned long debugctlmsr;
478
479	debugctlmsr = get_debugctlmsr();
480
481	debugctlmsr |= DEBUGCTLMSR_TR;
482	debugctlmsr |= DEBUGCTLMSR_BTS;
483	if (config & ARCH_PERFMON_EVENTSEL_INT)
484		debugctlmsr |= DEBUGCTLMSR_BTINT;
485
486	if (!(config & ARCH_PERFMON_EVENTSEL_OS))
487		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS;
488
489	if (!(config & ARCH_PERFMON_EVENTSEL_USR))
490		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR;
491
492	update_debugctlmsr(debugctlmsr);
493}
494
495void intel_pmu_disable_bts(void)
496{
497	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
498	unsigned long debugctlmsr;
499
500	if (!cpuc->ds)
501		return;
502
503	debugctlmsr = get_debugctlmsr();
504
505	debugctlmsr &=
506		~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT |
507		  DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR);
508
509	update_debugctlmsr(debugctlmsr);
510}
511
512int intel_pmu_drain_bts_buffer(void)
513{
514	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
515	struct debug_store *ds = cpuc->ds;
516	struct bts_record {
517		u64	from;
518		u64	to;
519		u64	flags;
520	};
521	struct perf_event *event = cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
522	struct bts_record *at, *base, *top;
523	struct perf_output_handle handle;
524	struct perf_event_header header;
525	struct perf_sample_data data;
526	unsigned long skip = 0;
527	struct pt_regs regs;
528
529	if (!event)
530		return 0;
531
532	if (!x86_pmu.bts_active)
533		return 0;
534
535	base = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
536	top  = (struct bts_record *)(unsigned long)ds->bts_index;
537
538	if (top <= base)
539		return 0;
540
541	memset(&regs, 0, sizeof(regs));
542
543	ds->bts_index = ds->bts_buffer_base;
544
545	perf_sample_data_init(&data, 0, event->hw.last_period);
546
547	/*
548	 * BTS leaks kernel addresses in branches across the cpl boundary,
549	 * such as traps or system calls, so unless the user is asking for
550	 * kernel tracing (and right now it's not possible), we'd need to
551	 * filter them out. But first we need to count how many of those we
552	 * have in the current batch. This is an extra O(n) pass, however,
553	 * it's much faster than the other one especially considering that
554	 * n <= 2560 (BTS_BUFFER_SIZE / BTS_RECORD_SIZE * 15/16; see the
555	 * alloc_bts_buffer()).
556	 */
557	for (at = base; at < top; at++) {
558		/*
559		 * Note that right now *this* BTS code only works if
560		 * attr::exclude_kernel is set, but let's keep this extra
561		 * check here in case that changes.
562		 */
563		if (event->attr.exclude_kernel &&
564		    (kernel_ip(at->from) || kernel_ip(at->to)))
565			skip++;
566	}
567
568	/*
569	 * Prepare a generic sample, i.e. fill in the invariant fields.
570	 * We will overwrite the from and to address before we output
571	 * the sample.
572	 */
573	perf_prepare_sample(&header, &data, event, &regs);
574
575	if (perf_output_begin(&handle, event, header.size *
576			      (top - base - skip)))
577		return 1;
578
579	for (at = base; at < top; at++) {
580		/* Filter out any records that contain kernel addresses. */
581		if (event->attr.exclude_kernel &&
582		    (kernel_ip(at->from) || kernel_ip(at->to)))
583			continue;
584
585		data.ip		= at->from;
586		data.addr	= at->to;
587
588		perf_output_sample(&handle, &header, &data, event);
589	}
590
591	perf_output_end(&handle);
592
593	/* There's new data available. */
594	event->hw.interrupts++;
595	event->pending_kill = POLL_IN;
596	return 1;
597}
598
599static inline void intel_pmu_drain_pebs_buffer(void)
600{
601	struct pt_regs regs;
602
603	x86_pmu.drain_pebs(&regs);
604}
605
606void intel_pmu_pebs_sched_task(struct perf_event_context *ctx, bool sched_in)
607{
608	if (!sched_in)
609		intel_pmu_drain_pebs_buffer();
610}
611
612/*
613 * PEBS
614 */
615struct event_constraint intel_core2_pebs_event_constraints[] = {
616	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
617	INTEL_FLAGS_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
618	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
619	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
620	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
621	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
622	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x01),
623	EVENT_CONSTRAINT_END
624};
625
626struct event_constraint intel_atom_pebs_event_constraints[] = {
627	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
628	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
629	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
630	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
631	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x01),
632	EVENT_CONSTRAINT_END
633};
634
635struct event_constraint intel_slm_pebs_event_constraints[] = {
636	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
637	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x1),
638	/* Allow all events as PEBS with no flags */
639	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
640	EVENT_CONSTRAINT_END
641};
642
643struct event_constraint intel_nehalem_pebs_event_constraints[] = {
644	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
645	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
646	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
647	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INST_RETIRED.ANY */
648	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
649	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
650	INTEL_FLAGS_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
651	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
652	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
653	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
654	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
655	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
656	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x0f),
657	EVENT_CONSTRAINT_END
658};
659
660struct event_constraint intel_westmere_pebs_event_constraints[] = {
661	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
662	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
663	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
664	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INSTR_RETIRED.* */
665	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
666	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
667	INTEL_FLAGS_EVENT_CONSTRAINT(0xc5, 0xf),    /* BR_MISP_RETIRED.* */
668	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
669	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
670	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
671	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
672	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
673	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x0f),
674	EVENT_CONSTRAINT_END
675};
676
677struct event_constraint intel_snb_pebs_event_constraints[] = {
678	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
679	INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
680	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
681	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
682	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
683        INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
684        INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
685        INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
686        INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
687	/* Allow all events as PEBS with no flags */
688	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
689	EVENT_CONSTRAINT_END
690};
691
692struct event_constraint intel_ivb_pebs_event_constraints[] = {
693        INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
694        INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
695	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
696	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
697	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
698	INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
699	INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
700	INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
701	INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
702	/* Allow all events as PEBS with no flags */
703	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
704        EVENT_CONSTRAINT_END
705};
706
707struct event_constraint intel_hsw_pebs_event_constraints[] = {
708	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
709	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
710	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
711	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
712	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
713	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
714	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
715	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
716	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
717	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
718	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
719	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
720	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
721	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
722	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
723	/* Allow all events as PEBS with no flags */
724	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
725	EVENT_CONSTRAINT_END
726};
727
728struct event_constraint intel_skl_pebs_event_constraints[] = {
729	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1c0, 0x2),	/* INST_RETIRED.PREC_DIST */
730	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
731	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
732	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
733	INTEL_PLD_CONSTRAINT(0x1cd, 0xf),		      /* MEM_TRANS_RETIRED.* */
734	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */
735	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */
736	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */
737	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x22d0, 0xf), /* MEM_INST_RETIRED.LOCK_STORES */
738	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */
739	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */
740	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */
741	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */
742	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_RETIRED.* */
743	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_L3_HIT_RETIRED.* */
744	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_L3_MISS_RETIRED.* */
745	/* Allow all events as PEBS with no flags */
746	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
747	EVENT_CONSTRAINT_END
748};
749
750struct event_constraint *intel_pebs_constraints(struct perf_event *event)
751{
752	struct event_constraint *c;
753
754	if (!event->attr.precise_ip)
755		return NULL;
756
757	if (x86_pmu.pebs_constraints) {
758		for_each_event_constraint(c, x86_pmu.pebs_constraints) {
759			if ((event->hw.config & c->cmask) == c->code) {
760				event->hw.flags |= c->flags;
761				return c;
762			}
763		}
764	}
765
766	return &emptyconstraint;
767}
768
769static inline bool pebs_is_enabled(struct cpu_hw_events *cpuc)
770{
771	return (cpuc->pebs_enabled & ((1ULL << MAX_PEBS_EVENTS) - 1));
772}
773
774void intel_pmu_pebs_enable(struct perf_event *event)
775{
776	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
777	struct hw_perf_event *hwc = &event->hw;
778	struct debug_store *ds = cpuc->ds;
779	bool first_pebs;
780	u64 threshold;
781
782	hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
783
784	first_pebs = !pebs_is_enabled(cpuc);
785	cpuc->pebs_enabled |= 1ULL << hwc->idx;
786
787	if (event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT)
788		cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32);
789	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
790		cpuc->pebs_enabled |= 1ULL << 63;
791
792	/*
793	 * When the event is constrained enough we can use a larger
794	 * threshold and run the event with less frequent PMI.
795	 */
796	if (hwc->flags & PERF_X86_EVENT_FREERUNNING) {
797		threshold = ds->pebs_absolute_maximum -
798			x86_pmu.max_pebs_events * x86_pmu.pebs_record_size;
799
800		if (first_pebs)
801			perf_sched_cb_inc(event->ctx->pmu);
802	} else {
803		threshold = ds->pebs_buffer_base + x86_pmu.pebs_record_size;
804
805		/*
806		 * If not all events can use larger buffer,
807		 * roll back to threshold = 1
808		 */
809		if (!first_pebs &&
810		    (ds->pebs_interrupt_threshold > threshold))
811			perf_sched_cb_dec(event->ctx->pmu);
812	}
813
814	/* Use auto-reload if possible to save a MSR write in the PMI */
815	if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
816		ds->pebs_event_reset[hwc->idx] =
817			(u64)(-hwc->sample_period) & x86_pmu.cntval_mask;
818	}
819
820	if (first_pebs || ds->pebs_interrupt_threshold > threshold)
821		ds->pebs_interrupt_threshold = threshold;
822}
823
824void intel_pmu_pebs_disable(struct perf_event *event)
825{
826	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
827	struct hw_perf_event *hwc = &event->hw;
828	struct debug_store *ds = cpuc->ds;
829	bool large_pebs = ds->pebs_interrupt_threshold >
830		ds->pebs_buffer_base + x86_pmu.pebs_record_size;
831
832	if (large_pebs)
833		intel_pmu_drain_pebs_buffer();
834
835	cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
836
837	if (event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT)
838		cpuc->pebs_enabled &= ~(1ULL << (hwc->idx + 32));
839	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
840		cpuc->pebs_enabled &= ~(1ULL << 63);
841
842	if (large_pebs && !pebs_is_enabled(cpuc))
843		perf_sched_cb_dec(event->ctx->pmu);
844
845	if (cpuc->enabled)
846		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
847
848	hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
849}
850
851void intel_pmu_pebs_enable_all(void)
852{
853	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
854
855	if (cpuc->pebs_enabled)
856		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
857}
858
859void intel_pmu_pebs_disable_all(void)
860{
861	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
862
863	if (cpuc->pebs_enabled)
864		wrmsrl(MSR_IA32_PEBS_ENABLE, 0);
865}
866
867static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
868{
869	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
870	unsigned long from = cpuc->lbr_entries[0].from;
871	unsigned long old_to, to = cpuc->lbr_entries[0].to;
872	unsigned long ip = regs->ip;
873	int is_64bit = 0;
874	void *kaddr;
875	int size;
876
877	/*
878	 * We don't need to fixup if the PEBS assist is fault like
879	 */
880	if (!x86_pmu.intel_cap.pebs_trap)
881		return 1;
882
883	/*
884	 * No LBR entry, no basic block, no rewinding
885	 */
886	if (!cpuc->lbr_stack.nr || !from || !to)
887		return 0;
888
889	/*
890	 * Basic blocks should never cross user/kernel boundaries
891	 */
892	if (kernel_ip(ip) != kernel_ip(to))
893		return 0;
894
895	/*
896	 * unsigned math, either ip is before the start (impossible) or
897	 * the basic block is larger than 1 page (sanity)
898	 */
899	if ((ip - to) > PEBS_FIXUP_SIZE)
900		return 0;
901
902	/*
903	 * We sampled a branch insn, rewind using the LBR stack
904	 */
905	if (ip == to) {
906		set_linear_ip(regs, from);
907		return 1;
908	}
909
910	size = ip - to;
911	if (!kernel_ip(ip)) {
912		int bytes;
913		u8 *buf = this_cpu_read(insn_buffer);
914
915		/* 'size' must fit our buffer, see above */
916		bytes = copy_from_user_nmi(buf, (void __user *)to, size);
917		if (bytes != 0)
918			return 0;
919
920		kaddr = buf;
921	} else {
922		kaddr = (void *)to;
923	}
924
925	do {
926		struct insn insn;
927
928		old_to = to;
929
930#ifdef CONFIG_X86_64
931		is_64bit = kernel_ip(to) || !test_thread_flag(TIF_IA32);
932#endif
933		insn_init(&insn, kaddr, size, is_64bit);
934		insn_get_length(&insn);
935		/*
936		 * Make sure there was not a problem decoding the
937		 * instruction and getting the length.  This is
938		 * doubly important because we have an infinite
939		 * loop if insn.length=0.
940		 */
941		if (!insn.length)
942			break;
943
944		to += insn.length;
945		kaddr += insn.length;
946		size -= insn.length;
947	} while (to < ip);
948
949	if (to == ip) {
950		set_linear_ip(regs, old_to);
951		return 1;
952	}
953
954	/*
955	 * Even though we decoded the basic block, the instruction stream
956	 * never matched the given IP, either the TO or the IP got corrupted.
957	 */
958	return 0;
959}
960
961static inline u64 intel_hsw_weight(struct pebs_record_skl *pebs)
962{
963	if (pebs->tsx_tuning) {
964		union hsw_tsx_tuning tsx = { .value = pebs->tsx_tuning };
965		return tsx.cycles_last_block;
966	}
967	return 0;
968}
969
970static inline u64 intel_hsw_transaction(struct pebs_record_skl *pebs)
971{
972	u64 txn = (pebs->tsx_tuning & PEBS_HSW_TSX_FLAGS) >> 32;
973
974	/* For RTM XABORTs also log the abort code from AX */
975	if ((txn & PERF_TXN_TRANSACTION) && (pebs->ax & 1))
976		txn |= ((pebs->ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT;
977	return txn;
978}
979
980static void setup_pebs_sample_data(struct perf_event *event,
981				   struct pt_regs *iregs, void *__pebs,
982				   struct perf_sample_data *data,
983				   struct pt_regs *regs)
984{
985#define PERF_X86_EVENT_PEBS_HSW_PREC \
986		(PERF_X86_EVENT_PEBS_ST_HSW | \
987		 PERF_X86_EVENT_PEBS_LD_HSW | \
988		 PERF_X86_EVENT_PEBS_NA_HSW)
989	/*
990	 * We cast to the biggest pebs_record but are careful not to
991	 * unconditionally access the 'extra' entries.
992	 */
993	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
994	struct pebs_record_skl *pebs = __pebs;
995	u64 sample_type;
996	int fll, fst, dsrc;
997	int fl = event->hw.flags;
998
999	if (pebs == NULL)
1000		return;
1001
1002	sample_type = event->attr.sample_type;
1003	dsrc = sample_type & PERF_SAMPLE_DATA_SRC;
1004
1005	fll = fl & PERF_X86_EVENT_PEBS_LDLAT;
1006	fst = fl & (PERF_X86_EVENT_PEBS_ST | PERF_X86_EVENT_PEBS_HSW_PREC);
1007
1008	perf_sample_data_init(data, 0, event->hw.last_period);
1009
1010	data->period = event->hw.last_period;
1011
1012	/*
1013	 * Use latency for weight (only avail with PEBS-LL)
1014	 */
1015	if (fll && (sample_type & PERF_SAMPLE_WEIGHT))
1016		data->weight = pebs->lat;
1017
1018	/*
1019	 * data.data_src encodes the data source
1020	 */
1021	if (dsrc) {
1022		u64 val = PERF_MEM_NA;
1023		if (fll)
1024			val = load_latency_data(pebs->dse);
1025		else if (fst && (fl & PERF_X86_EVENT_PEBS_HSW_PREC))
1026			val = precise_datala_hsw(event, pebs->dse);
1027		else if (fst)
1028			val = precise_store_data(pebs->dse);
1029		data->data_src.val = val;
1030	}
1031
1032	/*
1033	 * We use the interrupt regs as a base because the PEBS record
1034	 * does not contain a full regs set, specifically it seems to
1035	 * lack segment descriptors, which get used by things like
1036	 * user_mode().
1037	 *
1038	 * In the simple case fix up only the IP and BP,SP regs, for
1039	 * PERF_SAMPLE_IP and PERF_SAMPLE_CALLCHAIN to function properly.
1040	 * A possible PERF_SAMPLE_REGS will have to transfer all regs.
1041	 */
1042	*regs = *iregs;
1043	regs->flags = pebs->flags;
1044	set_linear_ip(regs, pebs->ip);
1045	regs->bp = pebs->bp;
1046	regs->sp = pebs->sp;
1047
1048	if (sample_type & PERF_SAMPLE_REGS_INTR) {
1049		regs->ax = pebs->ax;
1050		regs->bx = pebs->bx;
1051		regs->cx = pebs->cx;
1052		regs->dx = pebs->dx;
1053		regs->si = pebs->si;
1054		regs->di = pebs->di;
1055		regs->bp = pebs->bp;
1056		regs->sp = pebs->sp;
1057
1058		regs->flags = pebs->flags;
1059#ifndef CONFIG_X86_32
1060		regs->r8 = pebs->r8;
1061		regs->r9 = pebs->r9;
1062		regs->r10 = pebs->r10;
1063		regs->r11 = pebs->r11;
1064		regs->r12 = pebs->r12;
1065		regs->r13 = pebs->r13;
1066		regs->r14 = pebs->r14;
1067		regs->r15 = pebs->r15;
1068#endif
1069	}
1070
1071	if (event->attr.precise_ip > 1 && x86_pmu.intel_cap.pebs_format >= 2) {
1072		regs->ip = pebs->real_ip;
1073		regs->flags |= PERF_EFLAGS_EXACT;
1074	} else if (event->attr.precise_ip > 1 && intel_pmu_pebs_fixup_ip(regs))
1075		regs->flags |= PERF_EFLAGS_EXACT;
1076	else
1077		regs->flags &= ~PERF_EFLAGS_EXACT;
1078
1079	if ((sample_type & PERF_SAMPLE_ADDR) &&
1080	    x86_pmu.intel_cap.pebs_format >= 1)
1081		data->addr = pebs->dla;
1082
1083	if (x86_pmu.intel_cap.pebs_format >= 2) {
1084		/* Only set the TSX weight when no memory weight. */
1085		if ((sample_type & PERF_SAMPLE_WEIGHT) && !fll)
1086			data->weight = intel_hsw_weight(pebs);
1087
1088		if (sample_type & PERF_SAMPLE_TRANSACTION)
1089			data->txn = intel_hsw_transaction(pebs);
1090	}
1091
1092	/*
1093	 * v3 supplies an accurate time stamp, so we use that
1094	 * for the time stamp.
1095	 *
1096	 * We can only do this for the default trace clock.
1097	 */
1098	if (x86_pmu.intel_cap.pebs_format >= 3 &&
1099		event->attr.use_clockid == 0)
1100		data->time = native_sched_clock_from_tsc(pebs->tsc);
1101
1102	if (has_branch_stack(event))
1103		data->br_stack = &cpuc->lbr_stack;
1104}
1105
1106static inline void *
1107get_next_pebs_record_by_bit(void *base, void *top, int bit)
1108{
1109	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1110	void *at;
1111	u64 pebs_status;
1112
1113	if (base == NULL)
1114		return NULL;
1115
1116	for (at = base; at < top; at += x86_pmu.pebs_record_size) {
1117		struct pebs_record_nhm *p = at;
1118
1119		if (test_bit(bit, (unsigned long *)&p->status)) {
1120			/* PEBS v3 has accurate status bits */
1121			if (x86_pmu.intel_cap.pebs_format >= 3)
1122				return at;
1123
1124			if (p->status == (1 << bit))
1125				return at;
1126
1127			/* clear non-PEBS bit and re-check */
1128			pebs_status = p->status & cpuc->pebs_enabled;
1129			pebs_status &= (1ULL << MAX_PEBS_EVENTS) - 1;
1130			if (pebs_status == (1 << bit))
1131				return at;
1132		}
1133	}
1134	return NULL;
1135}
1136
1137static void __intel_pmu_pebs_event(struct perf_event *event,
1138				   struct pt_regs *iregs,
1139				   void *base, void *top,
1140				   int bit, int count)
1141{
1142	struct perf_sample_data data;
1143	struct pt_regs regs;
1144	void *at = get_next_pebs_record_by_bit(base, top, bit);
1145
1146	if (!intel_pmu_save_and_restart(event) &&
1147	    !(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD))
1148		return;
1149
1150	while (count > 1) {
1151		setup_pebs_sample_data(event, iregs, at, &data, &regs);
1152		perf_event_output(event, &data, &regs);
1153		at += x86_pmu.pebs_record_size;
1154		at = get_next_pebs_record_by_bit(at, top, bit);
1155		count--;
1156	}
1157
1158	setup_pebs_sample_data(event, iregs, at, &data, &regs);
1159
1160	/*
1161	 * All but the last records are processed.
1162	 * The last one is left to be able to call the overflow handler.
1163	 */
1164	if (perf_event_overflow(event, &data, &regs)) {
1165		x86_pmu_stop(event, 0);
1166		return;
1167	}
1168
1169}
1170
1171static void intel_pmu_drain_pebs_core(struct pt_regs *iregs)
1172{
1173	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1174	struct debug_store *ds = cpuc->ds;
1175	struct perf_event *event = cpuc->events[0]; /* PMC0 only */
1176	struct pebs_record_core *at, *top;
1177	int n;
1178
1179	if (!x86_pmu.pebs_active)
1180		return;
1181
1182	at  = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
1183	top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
1184
1185	/*
1186	 * Whatever else happens, drain the thing
1187	 */
1188	ds->pebs_index = ds->pebs_buffer_base;
1189
1190	if (!test_bit(0, cpuc->active_mask))
1191		return;
1192
1193	WARN_ON_ONCE(!event);
1194
1195	if (!event->attr.precise_ip)
1196		return;
1197
1198	n = (top - at) / x86_pmu.pebs_record_size;
1199	if (n <= 0)
1200		return;
1201
1202	__intel_pmu_pebs_event(event, iregs, at, top, 0, n);
1203}
1204
1205static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs)
1206{
1207	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1208	struct debug_store *ds = cpuc->ds;
1209	struct perf_event *event;
1210	void *base, *at, *top;
1211	short counts[MAX_PEBS_EVENTS] = {};
1212	short error[MAX_PEBS_EVENTS] = {};
1213	int bit, i;
1214
1215	if (!x86_pmu.pebs_active)
1216		return;
1217
1218	base = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
1219	top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
1220
1221	ds->pebs_index = ds->pebs_buffer_base;
1222
1223	if (unlikely(base >= top))
1224		return;
1225
1226	for (at = base; at < top; at += x86_pmu.pebs_record_size) {
1227		struct pebs_record_nhm *p = at;
1228		u64 pebs_status;
1229
1230		/* PEBS v3 has accurate status bits */
1231		if (x86_pmu.intel_cap.pebs_format >= 3) {
1232			for_each_set_bit(bit, (unsigned long *)&p->status,
1233					 MAX_PEBS_EVENTS)
1234				counts[bit]++;
1235
1236			continue;
1237		}
1238
1239		pebs_status = p->status & cpuc->pebs_enabled;
1240		pebs_status &= (1ULL << x86_pmu.max_pebs_events) - 1;
1241
1242		bit = find_first_bit((unsigned long *)&pebs_status,
1243					x86_pmu.max_pebs_events);
1244		if (WARN(bit >= x86_pmu.max_pebs_events,
1245			 "PEBS record without PEBS event! status=%Lx pebs_enabled=%Lx active_mask=%Lx",
1246			 (unsigned long long)p->status, (unsigned long long)cpuc->pebs_enabled,
1247			 *(unsigned long long *)cpuc->active_mask))
1248			continue;
1249
1250		/*
1251		 * The PEBS hardware does not deal well with the situation
1252		 * when events happen near to each other and multiple bits
1253		 * are set. But it should happen rarely.
1254		 *
1255		 * If these events include one PEBS and multiple non-PEBS
1256		 * events, it doesn't impact PEBS record. The record will
1257		 * be handled normally. (slow path)
1258		 *
1259		 * If these events include two or more PEBS events, the
1260		 * records for the events can be collapsed into a single
1261		 * one, and it's not possible to reconstruct all events
1262		 * that caused the PEBS record. It's called collision.
1263		 * If collision happened, the record will be dropped.
1264		 */
1265		if (p->status != (1ULL << bit)) {
1266			for_each_set_bit(i, (unsigned long *)&pebs_status,
1267					 x86_pmu.max_pebs_events)
1268				error[i]++;
1269			continue;
1270		}
1271
1272		counts[bit]++;
1273	}
1274
1275	for (bit = 0; bit < x86_pmu.max_pebs_events; bit++) {
1276		if ((counts[bit] == 0) && (error[bit] == 0))
1277			continue;
1278
1279		event = cpuc->events[bit];
1280		WARN_ON_ONCE(!event);
1281		WARN_ON_ONCE(!event->attr.precise_ip);
1282
1283		/* log dropped samples number */
1284		if (error[bit])
1285			perf_log_lost_samples(event, error[bit]);
1286
1287		if (counts[bit]) {
1288			__intel_pmu_pebs_event(event, iregs, base,
1289					       top, bit, counts[bit]);
1290		}
1291	}
1292}
1293
1294/*
1295 * BTS, PEBS probe and setup
1296 */
1297
1298void __init intel_ds_init(void)
1299{
1300	/*
1301	 * No support for 32bit formats
1302	 */
1303	if (!boot_cpu_has(X86_FEATURE_DTES64))
1304		return;
1305
1306	x86_pmu.bts  = boot_cpu_has(X86_FEATURE_BTS);
1307	x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
1308	x86_pmu.pebs_buffer_size = PEBS_BUFFER_SIZE;
1309	if (x86_pmu.pebs) {
1310		char pebs_type = x86_pmu.intel_cap.pebs_trap ?  '+' : '-';
1311		int format = x86_pmu.intel_cap.pebs_format;
1312
1313		switch (format) {
1314		case 0:
1315			printk(KERN_CONT "PEBS fmt0%c, ", pebs_type);
1316			x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
1317			/*
1318			 * Using >PAGE_SIZE buffers makes the WRMSR to
1319			 * PERF_GLOBAL_CTRL in intel_pmu_enable_all()
1320			 * mysteriously hang on Core2.
1321			 *
1322			 * As a workaround, we don't do this.
1323			 */
1324			x86_pmu.pebs_buffer_size = PAGE_SIZE;
1325			x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
1326			break;
1327
1328		case 1:
1329			printk(KERN_CONT "PEBS fmt1%c, ", pebs_type);
1330			x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
1331			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1332			break;
1333
1334		case 2:
1335			pr_cont("PEBS fmt2%c, ", pebs_type);
1336			x86_pmu.pebs_record_size = sizeof(struct pebs_record_hsw);
1337			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1338			break;
1339
1340		case 3:
1341			pr_cont("PEBS fmt3%c, ", pebs_type);
1342			x86_pmu.pebs_record_size =
1343						sizeof(struct pebs_record_skl);
1344			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1345			x86_pmu.free_running_flags |= PERF_SAMPLE_TIME;
1346			break;
1347
1348		default:
1349			printk(KERN_CONT "no PEBS fmt%d%c, ", format, pebs_type);
1350			x86_pmu.pebs = 0;
1351		}
1352	}
1353}
1354
1355void perf_restore_debug_store(void)
1356{
1357	struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
1358
1359	if (!x86_pmu.bts && !x86_pmu.pebs)
1360		return;
1361
1362	wrmsrl(MSR_IA32_DS_AREA, (unsigned long)ds);
1363}
1364