xref: /linux/arch/x86/events/intel/ds.c (revision 7fc2cd2e4b398c57c9cf961cfea05eadbf34c05c)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/bitops.h>
3 #include <linux/types.h>
4 #include <linux/slab.h>
5 #include <linux/sched/clock.h>
6 
7 #include <asm/cpu_entry_area.h>
8 #include <asm/debugreg.h>
9 #include <asm/perf_event.h>
10 #include <asm/tlbflush.h>
11 #include <asm/insn.h>
12 #include <asm/io.h>
13 #include <asm/msr.h>
14 #include <asm/timer.h>
15 
16 #include "../perf_event.h"
17 
18 /* Waste a full page so it can be mapped into the cpu_entry_area */
19 DEFINE_PER_CPU_PAGE_ALIGNED(struct debug_store, cpu_debug_store);
20 
21 /* The size of a BTS record in bytes: */
22 #define BTS_RECORD_SIZE		24
23 
24 #define PEBS_FIXUP_SIZE		PAGE_SIZE
25 
26 /*
27  * pebs_record_32 for p4 and core not supported
28 
29 struct pebs_record_32 {
30 	u32 flags, ip;
31 	u32 ax, bc, cx, dx;
32 	u32 si, di, bp, sp;
33 };
34 
35  */
36 
37 union intel_x86_pebs_dse {
38 	u64 val;
39 	struct {
40 		unsigned int ld_dse:4;
41 		unsigned int ld_stlb_miss:1;
42 		unsigned int ld_locked:1;
43 		unsigned int ld_data_blk:1;
44 		unsigned int ld_addr_blk:1;
45 		unsigned int ld_reserved:24;
46 	};
47 	struct {
48 		unsigned int st_l1d_hit:1;
49 		unsigned int st_reserved1:3;
50 		unsigned int st_stlb_miss:1;
51 		unsigned int st_locked:1;
52 		unsigned int st_reserved2:26;
53 	};
54 	struct {
55 		unsigned int st_lat_dse:4;
56 		unsigned int st_lat_stlb_miss:1;
57 		unsigned int st_lat_locked:1;
58 		unsigned int ld_reserved3:26;
59 	};
60 	struct {
61 		unsigned int mtl_dse:5;
62 		unsigned int mtl_locked:1;
63 		unsigned int mtl_stlb_miss:1;
64 		unsigned int mtl_fwd_blk:1;
65 		unsigned int ld_reserved4:24;
66 	};
67 	struct {
68 		unsigned int lnc_dse:8;
69 		unsigned int ld_reserved5:2;
70 		unsigned int lnc_stlb_miss:1;
71 		unsigned int lnc_locked:1;
72 		unsigned int lnc_data_blk:1;
73 		unsigned int lnc_addr_blk:1;
74 		unsigned int ld_reserved6:18;
75 	};
76 };
77 
78 
79 /*
80  * Map PEBS Load Latency Data Source encodings to generic
81  * memory data source information
82  */
83 #define P(a, b) PERF_MEM_S(a, b)
84 #define OP_LH (P(OP, LOAD) | P(LVL, HIT))
85 #define LEVEL(x) P(LVLNUM, x)
86 #define REM P(REMOTE, REMOTE)
87 #define SNOOP_NONE_MISS (P(SNOOP, NONE) | P(SNOOP, MISS))
88 
89 /* Version for Sandy Bridge and later */
90 static u64 pebs_data_source[PERF_PEBS_DATA_SOURCE_MAX] = {
91 	P(OP, LOAD) | P(LVL, MISS) | LEVEL(L3) | P(SNOOP, NA),/* 0x00:ukn L3 */
92 	OP_LH | P(LVL, L1)  | LEVEL(L1) | P(SNOOP, NONE),  /* 0x01: L1 local */
93 	OP_LH | P(LVL, LFB) | LEVEL(LFB) | P(SNOOP, NONE), /* 0x02: LFB hit */
94 	OP_LH | P(LVL, L2)  | LEVEL(L2) | P(SNOOP, NONE),  /* 0x03: L2 hit */
95 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, NONE),  /* 0x04: L3 hit */
96 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, MISS),  /* 0x05: L3 hit, snoop miss */
97 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HIT),   /* 0x06: L3 hit, snoop hit */
98 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HITM),  /* 0x07: L3 hit, snoop hitm */
99 	OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HIT),  /* 0x08: L3 miss snoop hit */
100 	OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HITM), /* 0x09: L3 miss snoop hitm*/
101 	OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | P(SNOOP, HIT),       /* 0x0a: L3 miss, shared */
102 	OP_LH | P(LVL, REM_RAM1) | REM | LEVEL(L3) | P(SNOOP, HIT),  /* 0x0b: L3 miss, shared */
103 	OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | SNOOP_NONE_MISS,     /* 0x0c: L3 miss, excl */
104 	OP_LH | P(LVL, REM_RAM1) | LEVEL(RAM) | REM | SNOOP_NONE_MISS, /* 0x0d: L3 miss, excl */
105 	OP_LH | P(LVL, IO)  | LEVEL(NA) | P(SNOOP, NONE), /* 0x0e: I/O */
106 	OP_LH | P(LVL, UNC) | LEVEL(NA) | P(SNOOP, NONE), /* 0x0f: uncached */
107 };
108 
109 /* Patch up minor differences in the bits */
110 void __init intel_pmu_pebs_data_source_nhm(void)
111 {
112 	pebs_data_source[0x05] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HIT);
113 	pebs_data_source[0x06] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
114 	pebs_data_source[0x07] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
115 }
116 
117 static void __init __intel_pmu_pebs_data_source_skl(bool pmem, u64 *data_source)
118 {
119 	u64 pmem_or_l4 = pmem ? LEVEL(PMEM) : LEVEL(L4);
120 
121 	data_source[0x08] = OP_LH | pmem_or_l4 | P(SNOOP, HIT);
122 	data_source[0x09] = OP_LH | pmem_or_l4 | REM | P(SNOOP, HIT);
123 	data_source[0x0b] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, NONE);
124 	data_source[0x0c] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOPX, FWD);
125 	data_source[0x0d] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOP, HITM);
126 }
127 
128 void __init intel_pmu_pebs_data_source_skl(bool pmem)
129 {
130 	__intel_pmu_pebs_data_source_skl(pmem, pebs_data_source);
131 }
132 
133 static void __init __intel_pmu_pebs_data_source_grt(u64 *data_source)
134 {
135 	data_source[0x05] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HIT);
136 	data_source[0x06] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
137 	data_source[0x08] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOPX, FWD);
138 }
139 
140 void __init intel_pmu_pebs_data_source_grt(void)
141 {
142 	__intel_pmu_pebs_data_source_grt(pebs_data_source);
143 }
144 
145 void __init intel_pmu_pebs_data_source_adl(void)
146 {
147 	u64 *data_source;
148 
149 	data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX].pebs_data_source;
150 	memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
151 	__intel_pmu_pebs_data_source_skl(false, data_source);
152 
153 	data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX].pebs_data_source;
154 	memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
155 	__intel_pmu_pebs_data_source_grt(data_source);
156 }
157 
158 static void __init __intel_pmu_pebs_data_source_cmt(u64 *data_source)
159 {
160 	data_source[0x07] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOPX, FWD);
161 	data_source[0x08] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
162 	data_source[0x0a] = OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | P(SNOOP, NONE);
163 	data_source[0x0b] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, NONE);
164 	data_source[0x0c] = OP_LH | LEVEL(RAM) | REM | P(SNOOPX, FWD);
165 	data_source[0x0d] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, HITM);
166 }
167 
168 void __init intel_pmu_pebs_data_source_mtl(void)
169 {
170 	u64 *data_source;
171 
172 	data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX].pebs_data_source;
173 	memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
174 	__intel_pmu_pebs_data_source_skl(false, data_source);
175 
176 	data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX].pebs_data_source;
177 	memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
178 	__intel_pmu_pebs_data_source_cmt(data_source);
179 }
180 
181 void __init intel_pmu_pebs_data_source_arl_h(void)
182 {
183 	u64 *data_source;
184 
185 	intel_pmu_pebs_data_source_lnl();
186 
187 	data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_TINY_IDX].pebs_data_source;
188 	memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
189 	__intel_pmu_pebs_data_source_cmt(data_source);
190 }
191 
192 void __init intel_pmu_pebs_data_source_cmt(void)
193 {
194 	__intel_pmu_pebs_data_source_cmt(pebs_data_source);
195 }
196 
197 /* Version for Lion Cove and later */
198 static u64 lnc_pebs_data_source[PERF_PEBS_DATA_SOURCE_MAX] = {
199 	P(OP, LOAD) | P(LVL, MISS) | LEVEL(L3) | P(SNOOP, NA),	/* 0x00: ukn L3 */
200 	OP_LH | P(LVL, L1)  | LEVEL(L1) | P(SNOOP, NONE),	/* 0x01: L1 hit */
201 	OP_LH | P(LVL, L1)  | LEVEL(L1) | P(SNOOP, NONE),	/* 0x02: L1 hit */
202 	OP_LH | P(LVL, LFB) | LEVEL(LFB) | P(SNOOP, NONE),	/* 0x03: LFB/L1 Miss Handling Buffer hit */
203 	0,							/* 0x04: Reserved */
204 	OP_LH | P(LVL, L2)  | LEVEL(L2) | P(SNOOP, NONE),	/* 0x05: L2 Hit */
205 	OP_LH | LEVEL(L2_MHB) | P(SNOOP, NONE),			/* 0x06: L2 Miss Handling Buffer Hit */
206 	0,							/* 0x07: Reserved */
207 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, NONE),	/* 0x08: L3 Hit */
208 	0,							/* 0x09: Reserved */
209 	0,							/* 0x0a: Reserved */
210 	0,							/* 0x0b: Reserved */
211 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOPX, FWD),	/* 0x0c: L3 Hit Snoop Fwd */
212 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HITM),	/* 0x0d: L3 Hit Snoop HitM */
213 	0,							/* 0x0e: Reserved */
214 	P(OP, LOAD) | P(LVL, MISS) | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HITM),	/* 0x0f: L3 Miss Snoop HitM */
215 	OP_LH | LEVEL(MSC) | P(SNOOP, NONE),			/* 0x10: Memory-side Cache Hit */
216 	OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | P(SNOOP, NONE), /* 0x11: Local Memory Hit */
217 };
218 
219 void __init intel_pmu_pebs_data_source_lnl(void)
220 {
221 	u64 *data_source;
222 
223 	data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX].pebs_data_source;
224 	memcpy(data_source, lnc_pebs_data_source, sizeof(lnc_pebs_data_source));
225 
226 	data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX].pebs_data_source;
227 	memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
228 	__intel_pmu_pebs_data_source_cmt(data_source);
229 }
230 
231 static u64 precise_store_data(u64 status)
232 {
233 	union intel_x86_pebs_dse dse;
234 	u64 val = P(OP, STORE) | P(SNOOP, NA) | P(LVL, L1) | P(TLB, L2);
235 
236 	dse.val = status;
237 
238 	/*
239 	 * bit 4: TLB access
240 	 * 1 = stored missed 2nd level TLB
241 	 *
242 	 * so it either hit the walker or the OS
243 	 * otherwise hit 2nd level TLB
244 	 */
245 	if (dse.st_stlb_miss)
246 		val |= P(TLB, MISS);
247 	else
248 		val |= P(TLB, HIT);
249 
250 	/*
251 	 * bit 0: hit L1 data cache
252 	 * if not set, then all we know is that
253 	 * it missed L1D
254 	 */
255 	if (dse.st_l1d_hit)
256 		val |= P(LVL, HIT);
257 	else
258 		val |= P(LVL, MISS);
259 
260 	/*
261 	 * bit 5: Locked prefix
262 	 */
263 	if (dse.st_locked)
264 		val |= P(LOCK, LOCKED);
265 
266 	return val;
267 }
268 
269 static u64 precise_datala_hsw(struct perf_event *event, u64 status)
270 {
271 	union perf_mem_data_src dse;
272 
273 	dse.val = PERF_MEM_NA;
274 
275 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
276 		dse.mem_op = PERF_MEM_OP_STORE;
277 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_LD_HSW)
278 		dse.mem_op = PERF_MEM_OP_LOAD;
279 
280 	/*
281 	 * L1 info only valid for following events:
282 	 *
283 	 * MEM_UOPS_RETIRED.STLB_MISS_STORES
284 	 * MEM_UOPS_RETIRED.LOCK_STORES
285 	 * MEM_UOPS_RETIRED.SPLIT_STORES
286 	 * MEM_UOPS_RETIRED.ALL_STORES
287 	 */
288 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW) {
289 		if (status & 1)
290 			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_HIT;
291 		else
292 			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_MISS;
293 	}
294 	return dse.val;
295 }
296 
297 static inline void pebs_set_tlb_lock(u64 *val, bool tlb, bool lock)
298 {
299 	/*
300 	 * TLB access
301 	 * 0 = did not miss 2nd level TLB
302 	 * 1 = missed 2nd level TLB
303 	 */
304 	if (tlb)
305 		*val |= P(TLB, MISS) | P(TLB, L2);
306 	else
307 		*val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
308 
309 	/* locked prefix */
310 	if (lock)
311 		*val |= P(LOCK, LOCKED);
312 }
313 
314 /* Retrieve the latency data for e-core of ADL */
315 static u64 __grt_latency_data(struct perf_event *event, u64 status,
316 			       u8 dse, bool tlb, bool lock, bool blk)
317 {
318 	u64 val;
319 
320 	WARN_ON_ONCE(is_hybrid() &&
321 		     hybrid_pmu(event->pmu)->pmu_type == hybrid_big);
322 
323 	dse &= PERF_PEBS_DATA_SOURCE_GRT_MASK;
324 	val = hybrid_var(event->pmu, pebs_data_source)[dse];
325 
326 	pebs_set_tlb_lock(&val, tlb, lock);
327 
328 	if (blk)
329 		val |= P(BLK, DATA);
330 	else
331 		val |= P(BLK, NA);
332 
333 	return val;
334 }
335 
336 u64 grt_latency_data(struct perf_event *event, u64 status)
337 {
338 	union intel_x86_pebs_dse dse;
339 
340 	dse.val = status;
341 
342 	return __grt_latency_data(event, status, dse.ld_dse,
343 				  dse.ld_locked, dse.ld_stlb_miss,
344 				  dse.ld_data_blk);
345 }
346 
347 /* Retrieve the latency data for e-core of MTL */
348 u64 cmt_latency_data(struct perf_event *event, u64 status)
349 {
350 	union intel_x86_pebs_dse dse;
351 
352 	dse.val = status;
353 
354 	return __grt_latency_data(event, status, dse.mtl_dse,
355 				  dse.mtl_stlb_miss, dse.mtl_locked,
356 				  dse.mtl_fwd_blk);
357 }
358 
359 static u64 lnc_latency_data(struct perf_event *event, u64 status)
360 {
361 	union intel_x86_pebs_dse dse;
362 	union perf_mem_data_src src;
363 	u64 val;
364 
365 	dse.val = status;
366 
367 	/* LNC core latency data */
368 	val = hybrid_var(event->pmu, pebs_data_source)[status & PERF_PEBS_DATA_SOURCE_MASK];
369 	if (!val)
370 		val = P(OP, LOAD) | LEVEL(NA) | P(SNOOP, NA);
371 
372 	if (dse.lnc_stlb_miss)
373 		val |= P(TLB, MISS) | P(TLB, L2);
374 	else
375 		val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
376 
377 	if (dse.lnc_locked)
378 		val |= P(LOCK, LOCKED);
379 
380 	if (dse.lnc_data_blk)
381 		val |= P(BLK, DATA);
382 	if (dse.lnc_addr_blk)
383 		val |= P(BLK, ADDR);
384 	if (!dse.lnc_data_blk && !dse.lnc_addr_blk)
385 		val |= P(BLK, NA);
386 
387 	src.val = val;
388 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
389 		src.mem_op = P(OP, STORE);
390 
391 	return src.val;
392 }
393 
394 u64 lnl_latency_data(struct perf_event *event, u64 status)
395 {
396 	struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu);
397 
398 	if (pmu->pmu_type == hybrid_small)
399 		return cmt_latency_data(event, status);
400 
401 	return lnc_latency_data(event, status);
402 }
403 
404 u64 arl_h_latency_data(struct perf_event *event, u64 status)
405 {
406 	struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu);
407 
408 	if (pmu->pmu_type == hybrid_tiny)
409 		return cmt_latency_data(event, status);
410 
411 	return lnl_latency_data(event, status);
412 }
413 
414 static u64 load_latency_data(struct perf_event *event, u64 status)
415 {
416 	union intel_x86_pebs_dse dse;
417 	u64 val;
418 
419 	dse.val = status;
420 
421 	/*
422 	 * use the mapping table for bit 0-3
423 	 */
424 	val = hybrid_var(event->pmu, pebs_data_source)[dse.ld_dse];
425 
426 	/*
427 	 * Nehalem models do not support TLB, Lock infos
428 	 */
429 	if (x86_pmu.pebs_no_tlb) {
430 		val |= P(TLB, NA) | P(LOCK, NA);
431 		return val;
432 	}
433 
434 	pebs_set_tlb_lock(&val, dse.ld_stlb_miss, dse.ld_locked);
435 
436 	/*
437 	 * Ice Lake and earlier models do not support block infos.
438 	 */
439 	if (!x86_pmu.pebs_block) {
440 		val |= P(BLK, NA);
441 		return val;
442 	}
443 	/*
444 	 * bit 6: load was blocked since its data could not be forwarded
445 	 *        from a preceding store
446 	 */
447 	if (dse.ld_data_blk)
448 		val |= P(BLK, DATA);
449 
450 	/*
451 	 * bit 7: load was blocked due to potential address conflict with
452 	 *        a preceding store
453 	 */
454 	if (dse.ld_addr_blk)
455 		val |= P(BLK, ADDR);
456 
457 	if (!dse.ld_data_blk && !dse.ld_addr_blk)
458 		val |= P(BLK, NA);
459 
460 	return val;
461 }
462 
463 static u64 store_latency_data(struct perf_event *event, u64 status)
464 {
465 	union intel_x86_pebs_dse dse;
466 	union perf_mem_data_src src;
467 	u64 val;
468 
469 	dse.val = status;
470 
471 	/*
472 	 * use the mapping table for bit 0-3
473 	 */
474 	val = hybrid_var(event->pmu, pebs_data_source)[dse.st_lat_dse];
475 
476 	pebs_set_tlb_lock(&val, dse.st_lat_stlb_miss, dse.st_lat_locked);
477 
478 	val |= P(BLK, NA);
479 
480 	/*
481 	 * the pebs_data_source table is only for loads
482 	 * so override the mem_op to say STORE instead
483 	 */
484 	src.val = val;
485 	src.mem_op = P(OP,STORE);
486 
487 	return src.val;
488 }
489 
490 struct pebs_record_core {
491 	u64 flags, ip;
492 	u64 ax, bx, cx, dx;
493 	u64 si, di, bp, sp;
494 	u64 r8,  r9,  r10, r11;
495 	u64 r12, r13, r14, r15;
496 };
497 
498 struct pebs_record_nhm {
499 	u64 flags, ip;
500 	u64 ax, bx, cx, dx;
501 	u64 si, di, bp, sp;
502 	u64 r8,  r9,  r10, r11;
503 	u64 r12, r13, r14, r15;
504 	u64 status, dla, dse, lat;
505 };
506 
507 /*
508  * Same as pebs_record_nhm, with two additional fields.
509  */
510 struct pebs_record_hsw {
511 	u64 flags, ip;
512 	u64 ax, bx, cx, dx;
513 	u64 si, di, bp, sp;
514 	u64 r8,  r9,  r10, r11;
515 	u64 r12, r13, r14, r15;
516 	u64 status, dla, dse, lat;
517 	u64 real_ip, tsx_tuning;
518 };
519 
520 union hsw_tsx_tuning {
521 	struct {
522 		u32 cycles_last_block     : 32,
523 		    hle_abort		  : 1,
524 		    rtm_abort		  : 1,
525 		    instruction_abort     : 1,
526 		    non_instruction_abort : 1,
527 		    retry		  : 1,
528 		    data_conflict	  : 1,
529 		    capacity_writes	  : 1,
530 		    capacity_reads	  : 1;
531 	};
532 	u64	    value;
533 };
534 
535 #define PEBS_HSW_TSX_FLAGS	0xff00000000ULL
536 
537 /* Same as HSW, plus TSC */
538 
539 struct pebs_record_skl {
540 	u64 flags, ip;
541 	u64 ax, bx, cx, dx;
542 	u64 si, di, bp, sp;
543 	u64 r8,  r9,  r10, r11;
544 	u64 r12, r13, r14, r15;
545 	u64 status, dla, dse, lat;
546 	u64 real_ip, tsx_tuning;
547 	u64 tsc;
548 };
549 
550 void init_debug_store_on_cpu(int cpu)
551 {
552 	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
553 
554 	if (!ds)
555 		return;
556 
557 	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
558 		     (u32)((u64)(unsigned long)ds),
559 		     (u32)((u64)(unsigned long)ds >> 32));
560 }
561 
562 void fini_debug_store_on_cpu(int cpu)
563 {
564 	if (!per_cpu(cpu_hw_events, cpu).ds)
565 		return;
566 
567 	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
568 }
569 
570 static DEFINE_PER_CPU(void *, insn_buffer);
571 
572 static void ds_update_cea(void *cea, void *addr, size_t size, pgprot_t prot)
573 {
574 	unsigned long start = (unsigned long)cea;
575 	phys_addr_t pa;
576 	size_t msz = 0;
577 
578 	pa = virt_to_phys(addr);
579 
580 	preempt_disable();
581 	for (; msz < size; msz += PAGE_SIZE, pa += PAGE_SIZE, cea += PAGE_SIZE)
582 		cea_set_pte(cea, pa, prot);
583 
584 	/*
585 	 * This is a cross-CPU update of the cpu_entry_area, we must shoot down
586 	 * all TLB entries for it.
587 	 */
588 	flush_tlb_kernel_range(start, start + size);
589 	preempt_enable();
590 }
591 
592 static void ds_clear_cea(void *cea, size_t size)
593 {
594 	unsigned long start = (unsigned long)cea;
595 	size_t msz = 0;
596 
597 	preempt_disable();
598 	for (; msz < size; msz += PAGE_SIZE, cea += PAGE_SIZE)
599 		cea_set_pte(cea, 0, PAGE_NONE);
600 
601 	flush_tlb_kernel_range(start, start + size);
602 	preempt_enable();
603 }
604 
605 static void *dsalloc_pages(size_t size, gfp_t flags, int cpu)
606 {
607 	unsigned int order = get_order(size);
608 	int node = cpu_to_node(cpu);
609 	struct page *page;
610 
611 	page = __alloc_pages_node(node, flags | __GFP_ZERO, order);
612 	return page ? page_address(page) : NULL;
613 }
614 
615 static void dsfree_pages(const void *buffer, size_t size)
616 {
617 	if (buffer)
618 		free_pages((unsigned long)buffer, get_order(size));
619 }
620 
621 static int alloc_pebs_buffer(int cpu)
622 {
623 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
624 	struct debug_store *ds = hwev->ds;
625 	size_t bsiz = x86_pmu.pebs_buffer_size;
626 	int max, node = cpu_to_node(cpu);
627 	void *buffer, *insn_buff, *cea;
628 
629 	if (!intel_pmu_has_pebs())
630 		return 0;
631 
632 	buffer = dsalloc_pages(bsiz, GFP_KERNEL, cpu);
633 	if (unlikely(!buffer))
634 		return -ENOMEM;
635 
636 	if (x86_pmu.arch_pebs) {
637 		hwev->pebs_vaddr = buffer;
638 		return 0;
639 	}
640 
641 	/*
642 	 * HSW+ already provides us the eventing ip; no need to allocate this
643 	 * buffer then.
644 	 */
645 	if (x86_pmu.intel_cap.pebs_format < 2) {
646 		insn_buff = kzalloc_node(PEBS_FIXUP_SIZE, GFP_KERNEL, node);
647 		if (!insn_buff) {
648 			dsfree_pages(buffer, bsiz);
649 			return -ENOMEM;
650 		}
651 		per_cpu(insn_buffer, cpu) = insn_buff;
652 	}
653 	hwev->pebs_vaddr = buffer;
654 	/* Update the cpu entry area mapping */
655 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
656 	ds->pebs_buffer_base = (unsigned long) cea;
657 	ds_update_cea(cea, buffer, bsiz, PAGE_KERNEL);
658 	ds->pebs_index = ds->pebs_buffer_base;
659 	max = x86_pmu.pebs_record_size * (bsiz / x86_pmu.pebs_record_size);
660 	ds->pebs_absolute_maximum = ds->pebs_buffer_base + max;
661 	return 0;
662 }
663 
664 static void release_pebs_buffer(int cpu)
665 {
666 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
667 	void *cea;
668 
669 	if (!intel_pmu_has_pebs())
670 		return;
671 
672 	if (x86_pmu.ds_pebs) {
673 		kfree(per_cpu(insn_buffer, cpu));
674 		per_cpu(insn_buffer, cpu) = NULL;
675 
676 		/* Clear the fixmap */
677 		cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
678 		ds_clear_cea(cea, x86_pmu.pebs_buffer_size);
679 	}
680 
681 	dsfree_pages(hwev->pebs_vaddr, x86_pmu.pebs_buffer_size);
682 	hwev->pebs_vaddr = NULL;
683 }
684 
685 static int alloc_bts_buffer(int cpu)
686 {
687 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
688 	struct debug_store *ds = hwev->ds;
689 	void *buffer, *cea;
690 	int max;
691 
692 	if (!x86_pmu.bts)
693 		return 0;
694 
695 	buffer = dsalloc_pages(BTS_BUFFER_SIZE, GFP_KERNEL | __GFP_NOWARN, cpu);
696 	if (unlikely(!buffer)) {
697 		WARN_ONCE(1, "%s: BTS buffer allocation failure\n", __func__);
698 		return -ENOMEM;
699 	}
700 	hwev->ds_bts_vaddr = buffer;
701 	/* Update the fixmap */
702 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
703 	ds->bts_buffer_base = (unsigned long) cea;
704 	ds_update_cea(cea, buffer, BTS_BUFFER_SIZE, PAGE_KERNEL);
705 	ds->bts_index = ds->bts_buffer_base;
706 	max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
707 	ds->bts_absolute_maximum = ds->bts_buffer_base +
708 					max * BTS_RECORD_SIZE;
709 	ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
710 					(max / 16) * BTS_RECORD_SIZE;
711 	return 0;
712 }
713 
714 static void release_bts_buffer(int cpu)
715 {
716 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
717 	void *cea;
718 
719 	if (!x86_pmu.bts)
720 		return;
721 
722 	/* Clear the fixmap */
723 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
724 	ds_clear_cea(cea, BTS_BUFFER_SIZE);
725 	dsfree_pages(hwev->ds_bts_vaddr, BTS_BUFFER_SIZE);
726 	hwev->ds_bts_vaddr = NULL;
727 }
728 
729 static int alloc_ds_buffer(int cpu)
730 {
731 	struct debug_store *ds = &get_cpu_entry_area(cpu)->cpu_debug_store;
732 
733 	memset(ds, 0, sizeof(*ds));
734 	per_cpu(cpu_hw_events, cpu).ds = ds;
735 	return 0;
736 }
737 
738 static void release_ds_buffer(int cpu)
739 {
740 	per_cpu(cpu_hw_events, cpu).ds = NULL;
741 }
742 
743 void release_ds_buffers(void)
744 {
745 	int cpu;
746 
747 	if (!x86_pmu.bts && !x86_pmu.ds_pebs)
748 		return;
749 
750 	for_each_possible_cpu(cpu)
751 		release_ds_buffer(cpu);
752 
753 	for_each_possible_cpu(cpu) {
754 		/*
755 		 * Again, ignore errors from offline CPUs, they will no longer
756 		 * observe cpu_hw_events.ds and not program the DS_AREA when
757 		 * they come up.
758 		 */
759 		fini_debug_store_on_cpu(cpu);
760 	}
761 
762 	for_each_possible_cpu(cpu) {
763 		if (x86_pmu.ds_pebs)
764 			release_pebs_buffer(cpu);
765 		release_bts_buffer(cpu);
766 	}
767 }
768 
769 void reserve_ds_buffers(void)
770 {
771 	int bts_err = 0, pebs_err = 0;
772 	int cpu;
773 
774 	x86_pmu.bts_active = 0;
775 
776 	if (x86_pmu.ds_pebs)
777 		x86_pmu.pebs_active = 0;
778 
779 	if (!x86_pmu.bts && !x86_pmu.ds_pebs)
780 		return;
781 
782 	if (!x86_pmu.bts)
783 		bts_err = 1;
784 
785 	if (!x86_pmu.ds_pebs)
786 		pebs_err = 1;
787 
788 	for_each_possible_cpu(cpu) {
789 		if (alloc_ds_buffer(cpu)) {
790 			bts_err = 1;
791 			pebs_err = 1;
792 		}
793 
794 		if (!bts_err && alloc_bts_buffer(cpu))
795 			bts_err = 1;
796 
797 		if (x86_pmu.ds_pebs && !pebs_err &&
798 		    alloc_pebs_buffer(cpu))
799 			pebs_err = 1;
800 
801 		if (bts_err && pebs_err)
802 			break;
803 	}
804 
805 	if (bts_err) {
806 		for_each_possible_cpu(cpu)
807 			release_bts_buffer(cpu);
808 	}
809 
810 	if (x86_pmu.ds_pebs && pebs_err) {
811 		for_each_possible_cpu(cpu)
812 			release_pebs_buffer(cpu);
813 	}
814 
815 	if (bts_err && pebs_err) {
816 		for_each_possible_cpu(cpu)
817 			release_ds_buffer(cpu);
818 	} else {
819 		if (x86_pmu.bts && !bts_err)
820 			x86_pmu.bts_active = 1;
821 
822 		if (x86_pmu.ds_pebs && !pebs_err)
823 			x86_pmu.pebs_active = 1;
824 
825 		for_each_possible_cpu(cpu) {
826 			/*
827 			 * Ignores wrmsr_on_cpu() errors for offline CPUs they
828 			 * will get this call through intel_pmu_cpu_starting().
829 			 */
830 			init_debug_store_on_cpu(cpu);
831 		}
832 	}
833 }
834 
835 inline int alloc_arch_pebs_buf_on_cpu(int cpu)
836 {
837 	if (!x86_pmu.arch_pebs)
838 		return 0;
839 
840 	return alloc_pebs_buffer(cpu);
841 }
842 
843 inline void release_arch_pebs_buf_on_cpu(int cpu)
844 {
845 	if (!x86_pmu.arch_pebs)
846 		return;
847 
848 	release_pebs_buffer(cpu);
849 }
850 
851 void init_arch_pebs_on_cpu(int cpu)
852 {
853 	struct cpu_hw_events *cpuc = per_cpu_ptr(&cpu_hw_events, cpu);
854 	u64 arch_pebs_base;
855 
856 	if (!x86_pmu.arch_pebs)
857 		return;
858 
859 	if (!cpuc->pebs_vaddr) {
860 		WARN(1, "Fail to allocate PEBS buffer on CPU %d\n", cpu);
861 		x86_pmu.pebs_active = 0;
862 		return;
863 	}
864 
865 	/*
866 	 * 4KB-aligned pointer of the output buffer
867 	 * (__alloc_pages_node() return page aligned address)
868 	 * Buffer Size = 4KB * 2^SIZE
869 	 * contiguous physical buffer (__alloc_pages_node() with order)
870 	 */
871 	arch_pebs_base = virt_to_phys(cpuc->pebs_vaddr) | PEBS_BUFFER_SHIFT;
872 	wrmsr_on_cpu(cpu, MSR_IA32_PEBS_BASE, (u32)arch_pebs_base,
873 		     (u32)(arch_pebs_base >> 32));
874 	x86_pmu.pebs_active = 1;
875 }
876 
877 inline void fini_arch_pebs_on_cpu(int cpu)
878 {
879 	if (!x86_pmu.arch_pebs)
880 		return;
881 
882 	wrmsr_on_cpu(cpu, MSR_IA32_PEBS_BASE, 0, 0);
883 }
884 
885 /*
886  * BTS
887  */
888 
889 struct event_constraint bts_constraint =
890 	EVENT_CONSTRAINT(0, 1ULL << INTEL_PMC_IDX_FIXED_BTS, 0);
891 
892 void intel_pmu_enable_bts(u64 config)
893 {
894 	unsigned long debugctlmsr;
895 
896 	debugctlmsr = get_debugctlmsr();
897 
898 	debugctlmsr |= DEBUGCTLMSR_TR;
899 	debugctlmsr |= DEBUGCTLMSR_BTS;
900 	if (config & ARCH_PERFMON_EVENTSEL_INT)
901 		debugctlmsr |= DEBUGCTLMSR_BTINT;
902 
903 	if (!(config & ARCH_PERFMON_EVENTSEL_OS))
904 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS;
905 
906 	if (!(config & ARCH_PERFMON_EVENTSEL_USR))
907 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR;
908 
909 	update_debugctlmsr(debugctlmsr);
910 }
911 
912 void intel_pmu_disable_bts(void)
913 {
914 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
915 	unsigned long debugctlmsr;
916 
917 	if (!cpuc->ds)
918 		return;
919 
920 	debugctlmsr = get_debugctlmsr();
921 
922 	debugctlmsr &=
923 		~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT |
924 		  DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR);
925 
926 	update_debugctlmsr(debugctlmsr);
927 }
928 
929 int intel_pmu_drain_bts_buffer(void)
930 {
931 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
932 	struct debug_store *ds = cpuc->ds;
933 	struct bts_record {
934 		u64	from;
935 		u64	to;
936 		u64	flags;
937 	};
938 	struct perf_event *event = cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
939 	struct bts_record *at, *base, *top;
940 	struct perf_output_handle handle;
941 	struct perf_event_header header;
942 	struct perf_sample_data data;
943 	unsigned long skip = 0;
944 	struct pt_regs regs;
945 
946 	if (!event)
947 		return 0;
948 
949 	if (!x86_pmu.bts_active)
950 		return 0;
951 
952 	base = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
953 	top  = (struct bts_record *)(unsigned long)ds->bts_index;
954 
955 	if (top <= base)
956 		return 0;
957 
958 	memset(&regs, 0, sizeof(regs));
959 
960 	ds->bts_index = ds->bts_buffer_base;
961 
962 	perf_sample_data_init(&data, 0, event->hw.last_period);
963 
964 	/*
965 	 * BTS leaks kernel addresses in branches across the cpl boundary,
966 	 * such as traps or system calls, so unless the user is asking for
967 	 * kernel tracing (and right now it's not possible), we'd need to
968 	 * filter them out. But first we need to count how many of those we
969 	 * have in the current batch. This is an extra O(n) pass, however,
970 	 * it's much faster than the other one especially considering that
971 	 * n <= 2560 (BTS_BUFFER_SIZE / BTS_RECORD_SIZE * 15/16; see the
972 	 * alloc_bts_buffer()).
973 	 */
974 	for (at = base; at < top; at++) {
975 		/*
976 		 * Note that right now *this* BTS code only works if
977 		 * attr::exclude_kernel is set, but let's keep this extra
978 		 * check here in case that changes.
979 		 */
980 		if (event->attr.exclude_kernel &&
981 		    (kernel_ip(at->from) || kernel_ip(at->to)))
982 			skip++;
983 	}
984 
985 	/*
986 	 * Prepare a generic sample, i.e. fill in the invariant fields.
987 	 * We will overwrite the from and to address before we output
988 	 * the sample.
989 	 */
990 	rcu_read_lock();
991 	perf_prepare_sample(&data, event, &regs);
992 	perf_prepare_header(&header, &data, event, &regs);
993 
994 	if (perf_output_begin(&handle, &data, event,
995 			      header.size * (top - base - skip)))
996 		goto unlock;
997 
998 	for (at = base; at < top; at++) {
999 		/* Filter out any records that contain kernel addresses. */
1000 		if (event->attr.exclude_kernel &&
1001 		    (kernel_ip(at->from) || kernel_ip(at->to)))
1002 			continue;
1003 
1004 		data.ip		= at->from;
1005 		data.addr	= at->to;
1006 
1007 		perf_output_sample(&handle, &header, &data, event);
1008 	}
1009 
1010 	perf_output_end(&handle);
1011 
1012 	/* There's new data available. */
1013 	event->hw.interrupts++;
1014 	event->pending_kill = POLL_IN;
1015 unlock:
1016 	rcu_read_unlock();
1017 	return 1;
1018 }
1019 
1020 void intel_pmu_drain_pebs_buffer(void)
1021 {
1022 	struct perf_sample_data data;
1023 
1024 	static_call(x86_pmu_drain_pebs)(NULL, &data);
1025 }
1026 
1027 /*
1028  * PEBS
1029  */
1030 struct event_constraint intel_core2_pebs_event_constraints[] = {
1031 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
1032 	INTEL_FLAGS_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
1033 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
1034 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
1035 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
1036 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
1037 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01),
1038 	EVENT_CONSTRAINT_END
1039 };
1040 
1041 struct event_constraint intel_atom_pebs_event_constraints[] = {
1042 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
1043 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
1044 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
1045 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
1046 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01),
1047 	/* Allow all events as PEBS with no flags */
1048 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
1049 	EVENT_CONSTRAINT_END
1050 };
1051 
1052 struct event_constraint intel_slm_pebs_event_constraints[] = {
1053 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
1054 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x1),
1055 	/* Allow all events as PEBS with no flags */
1056 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
1057 	EVENT_CONSTRAINT_END
1058 };
1059 
1060 struct event_constraint intel_glm_pebs_event_constraints[] = {
1061 	/* Allow all events as PEBS with no flags */
1062 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
1063 	EVENT_CONSTRAINT_END
1064 };
1065 
1066 struct event_constraint intel_grt_pebs_event_constraints[] = {
1067 	/* Allow all events as PEBS with no flags */
1068 	INTEL_HYBRID_LAT_CONSTRAINT(0x5d0, 0x3),
1069 	INTEL_HYBRID_LAT_CONSTRAINT(0x6d0, 0xf),
1070 	EVENT_CONSTRAINT_END
1071 };
1072 
1073 struct event_constraint intel_nehalem_pebs_event_constraints[] = {
1074 	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
1075 	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
1076 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
1077 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INST_RETIRED.ANY */
1078 	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
1079 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
1080 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
1081 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
1082 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
1083 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
1084 	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
1085 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
1086 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
1087 	EVENT_CONSTRAINT_END
1088 };
1089 
1090 struct event_constraint intel_westmere_pebs_event_constraints[] = {
1091 	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
1092 	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
1093 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
1094 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INSTR_RETIRED.* */
1095 	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
1096 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
1097 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc5, 0xf),    /* BR_MISP_RETIRED.* */
1098 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
1099 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
1100 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
1101 	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
1102 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
1103 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
1104 	EVENT_CONSTRAINT_END
1105 };
1106 
1107 struct event_constraint intel_snb_pebs_event_constraints[] = {
1108 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
1109 	INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
1110 	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
1111 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
1112 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
1113         INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
1114         INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
1115         INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
1116         INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
1117 	/* Allow all events as PEBS with no flags */
1118 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
1119 	EVENT_CONSTRAINT_END
1120 };
1121 
1122 struct event_constraint intel_ivb_pebs_event_constraints[] = {
1123         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
1124         INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
1125 	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
1126 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
1127 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
1128 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
1129 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
1130 	INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
1131 	INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
1132 	INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
1133 	INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
1134 	/* Allow all events as PEBS with no flags */
1135 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
1136         EVENT_CONSTRAINT_END
1137 };
1138 
1139 struct event_constraint intel_hsw_pebs_event_constraints[] = {
1140 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
1141 	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
1142 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
1143 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
1144 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
1145 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
1146 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
1147 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
1148 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
1149 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
1150 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
1151 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
1152 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
1153 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
1154 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
1155 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
1156 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
1157 	/* Allow all events as PEBS with no flags */
1158 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
1159 	EVENT_CONSTRAINT_END
1160 };
1161 
1162 struct event_constraint intel_bdw_pebs_event_constraints[] = {
1163 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
1164 	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
1165 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
1166 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
1167 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
1168 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
1169 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
1170 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
1171 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
1172 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
1173 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
1174 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
1175 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
1176 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
1177 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
1178 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
1179 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
1180 	/* Allow all events as PEBS with no flags */
1181 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
1182 	EVENT_CONSTRAINT_END
1183 };
1184 
1185 
1186 struct event_constraint intel_skl_pebs_event_constraints[] = {
1187 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1c0, 0x2),	/* INST_RETIRED.PREC_DIST */
1188 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
1189 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
1190 	/* INST_RETIRED.TOTAL_CYCLES_PS (inv=1, cmask=16) (cycles:p). */
1191 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
1192 	INTEL_PLD_CONSTRAINT(0x1cd, 0xf),		      /* MEM_TRANS_RETIRED.* */
1193 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */
1194 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */
1195 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */
1196 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x22d0, 0xf), /* MEM_INST_RETIRED.LOCK_STORES */
1197 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */
1198 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */
1199 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */
1200 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */
1201 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_RETIRED.* */
1202 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_L3_HIT_RETIRED.* */
1203 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_L3_MISS_RETIRED.* */
1204 	/* Allow all events as PEBS with no flags */
1205 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
1206 	EVENT_CONSTRAINT_END
1207 };
1208 
1209 struct event_constraint intel_icl_pebs_event_constraints[] = {
1210 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x100000000ULL),	/* old INST_RETIRED.PREC_DIST */
1211 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x0100, 0x100000000ULL),	/* INST_RETIRED.PREC_DIST */
1212 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x0400, 0x800000000ULL),	/* SLOTS */
1213 
1214 	INTEL_PLD_CONSTRAINT(0x1cd, 0xff),			/* MEM_TRANS_RETIRED.LOAD_LATENCY */
1215 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf),	/* MEM_INST_RETIRED.STLB_MISS_LOADS */
1216 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf),	/* MEM_INST_RETIRED.STLB_MISS_STORES */
1217 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf),	/* MEM_INST_RETIRED.LOCK_LOADS */
1218 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf),	/* MEM_INST_RETIRED.SPLIT_LOADS */
1219 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf),	/* MEM_INST_RETIRED.SPLIT_STORES */
1220 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf),	/* MEM_INST_RETIRED.ALL_LOADS */
1221 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf),	/* MEM_INST_RETIRED.ALL_STORES */
1222 
1223 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(0xd1, 0xd4, 0xf), /* MEM_LOAD_*_RETIRED.* */
1224 
1225 	INTEL_FLAGS_EVENT_CONSTRAINT(0xd0, 0xf),		/* MEM_INST_RETIRED.* */
1226 
1227 	/*
1228 	 * Everything else is handled by PMU_FL_PEBS_ALL, because we
1229 	 * need the full constraints from the main table.
1230 	 */
1231 
1232 	EVENT_CONSTRAINT_END
1233 };
1234 
1235 struct event_constraint intel_glc_pebs_event_constraints[] = {
1236 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x100, 0x100000000ULL),	/* INST_RETIRED.PREC_DIST */
1237 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x0400, 0x800000000ULL),
1238 
1239 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xfe),
1240 	INTEL_PLD_CONSTRAINT(0x1cd, 0xfe),
1241 	INTEL_PSD_CONSTRAINT(0x2cd, 0x1),
1242 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf),	/* MEM_INST_RETIRED.STLB_MISS_LOADS */
1243 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf),	/* MEM_INST_RETIRED.STLB_MISS_STORES */
1244 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf),	/* MEM_INST_RETIRED.LOCK_LOADS */
1245 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf),	/* MEM_INST_RETIRED.SPLIT_LOADS */
1246 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf),	/* MEM_INST_RETIRED.SPLIT_STORES */
1247 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf),	/* MEM_INST_RETIRED.ALL_LOADS */
1248 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf),	/* MEM_INST_RETIRED.ALL_STORES */
1249 
1250 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(0xd1, 0xd4, 0xf),
1251 
1252 	INTEL_FLAGS_EVENT_CONSTRAINT(0xd0, 0xf),
1253 
1254 	/*
1255 	 * Everything else is handled by PMU_FL_PEBS_ALL, because we
1256 	 * need the full constraints from the main table.
1257 	 */
1258 
1259 	EVENT_CONSTRAINT_END
1260 };
1261 
1262 struct event_constraint intel_lnc_pebs_event_constraints[] = {
1263 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x100, 0x100000000ULL),	/* INST_RETIRED.PREC_DIST */
1264 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x0400, 0x800000000ULL),
1265 
1266 	INTEL_HYBRID_LDLAT_CONSTRAINT(0x1cd, 0x3fc),
1267 	INTEL_HYBRID_STLAT_CONSTRAINT(0x2cd, 0x3),
1268 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf),	/* MEM_INST_RETIRED.STLB_MISS_LOADS */
1269 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf),	/* MEM_INST_RETIRED.STLB_MISS_STORES */
1270 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf),	/* MEM_INST_RETIRED.LOCK_LOADS */
1271 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf),	/* MEM_INST_RETIRED.SPLIT_LOADS */
1272 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf),	/* MEM_INST_RETIRED.SPLIT_STORES */
1273 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf),	/* MEM_INST_RETIRED.ALL_LOADS */
1274 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf),	/* MEM_INST_RETIRED.ALL_STORES */
1275 
1276 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(0xd1, 0xd4, 0xf),
1277 
1278 	INTEL_FLAGS_EVENT_CONSTRAINT(0xd0, 0xf),
1279 
1280 	/*
1281 	 * Everything else is handled by PMU_FL_PEBS_ALL, because we
1282 	 * need the full constraints from the main table.
1283 	 */
1284 
1285 	EVENT_CONSTRAINT_END
1286 };
1287 
1288 struct event_constraint *intel_pebs_constraints(struct perf_event *event)
1289 {
1290 	struct event_constraint *pebs_constraints = hybrid(event->pmu, pebs_constraints);
1291 	struct event_constraint *c;
1292 
1293 	if (!event->attr.precise_ip)
1294 		return NULL;
1295 
1296 	if (pebs_constraints) {
1297 		for_each_event_constraint(c, pebs_constraints) {
1298 			if (constraint_match(c, event->hw.config)) {
1299 				event->hw.flags |= c->flags;
1300 				return c;
1301 			}
1302 		}
1303 	}
1304 
1305 	/*
1306 	 * Extended PEBS support
1307 	 * Makes the PEBS code search the normal constraints.
1308 	 */
1309 	if (x86_pmu.flags & PMU_FL_PEBS_ALL)
1310 		return NULL;
1311 
1312 	return &emptyconstraint;
1313 }
1314 
1315 /*
1316  * We need the sched_task callback even for per-cpu events when we use
1317  * the large interrupt threshold, such that we can provide PID and TID
1318  * to PEBS samples.
1319  */
1320 static inline bool pebs_needs_sched_cb(struct cpu_hw_events *cpuc)
1321 {
1322 	if (cpuc->n_pebs == cpuc->n_pebs_via_pt)
1323 		return false;
1324 
1325 	return cpuc->n_pebs && (cpuc->n_pebs == cpuc->n_large_pebs);
1326 }
1327 
1328 void intel_pmu_pebs_sched_task(struct perf_event_pmu_context *pmu_ctx, bool sched_in)
1329 {
1330 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1331 
1332 	if (!sched_in && pebs_needs_sched_cb(cpuc))
1333 		intel_pmu_drain_pebs_buffer();
1334 }
1335 
1336 static inline void pebs_update_threshold(struct cpu_hw_events *cpuc)
1337 {
1338 	struct debug_store *ds = cpuc->ds;
1339 	int max_pebs_events = intel_pmu_max_num_pebs(cpuc->pmu);
1340 	u64 threshold;
1341 	int reserved;
1342 
1343 	if (cpuc->n_pebs_via_pt)
1344 		return;
1345 
1346 	if (x86_pmu.flags & PMU_FL_PEBS_ALL)
1347 		reserved = max_pebs_events + x86_pmu_max_num_counters_fixed(cpuc->pmu);
1348 	else
1349 		reserved = max_pebs_events;
1350 
1351 	if (cpuc->n_pebs == cpuc->n_large_pebs) {
1352 		threshold = ds->pebs_absolute_maximum -
1353 			reserved * cpuc->pebs_record_size;
1354 	} else {
1355 		threshold = ds->pebs_buffer_base + cpuc->pebs_record_size;
1356 	}
1357 
1358 	ds->pebs_interrupt_threshold = threshold;
1359 }
1360 
1361 #define PEBS_DATACFG_CNTRS(x)						\
1362 	((x >> PEBS_DATACFG_CNTR_SHIFT) & PEBS_DATACFG_CNTR_MASK)
1363 
1364 #define PEBS_DATACFG_CNTR_BIT(x)					\
1365 	(((1ULL << x) & PEBS_DATACFG_CNTR_MASK) << PEBS_DATACFG_CNTR_SHIFT)
1366 
1367 #define PEBS_DATACFG_FIX(x)						\
1368 	((x >> PEBS_DATACFG_FIX_SHIFT) & PEBS_DATACFG_FIX_MASK)
1369 
1370 #define PEBS_DATACFG_FIX_BIT(x)						\
1371 	(((1ULL << (x)) & PEBS_DATACFG_FIX_MASK)			\
1372 	 << PEBS_DATACFG_FIX_SHIFT)
1373 
1374 static void adaptive_pebs_record_size_update(void)
1375 {
1376 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1377 	u64 pebs_data_cfg = cpuc->pebs_data_cfg;
1378 	int sz = sizeof(struct pebs_basic);
1379 
1380 	if (pebs_data_cfg & PEBS_DATACFG_MEMINFO)
1381 		sz += sizeof(struct pebs_meminfo);
1382 	if (pebs_data_cfg & PEBS_DATACFG_GP)
1383 		sz += sizeof(struct pebs_gprs);
1384 	if (pebs_data_cfg & PEBS_DATACFG_XMMS)
1385 		sz += sizeof(struct pebs_xmm);
1386 	if (pebs_data_cfg & PEBS_DATACFG_LBRS)
1387 		sz += x86_pmu.lbr_nr * sizeof(struct lbr_entry);
1388 	if (pebs_data_cfg & (PEBS_DATACFG_METRICS | PEBS_DATACFG_CNTR)) {
1389 		sz += sizeof(struct pebs_cntr_header);
1390 
1391 		/* Metrics base and Metrics Data */
1392 		if (pebs_data_cfg & PEBS_DATACFG_METRICS)
1393 			sz += 2 * sizeof(u64);
1394 
1395 		if (pebs_data_cfg & PEBS_DATACFG_CNTR) {
1396 			sz += (hweight64(PEBS_DATACFG_CNTRS(pebs_data_cfg)) +
1397 			       hweight64(PEBS_DATACFG_FIX(pebs_data_cfg))) *
1398 			      sizeof(u64);
1399 		}
1400 	}
1401 
1402 	cpuc->pebs_record_size = sz;
1403 }
1404 
1405 static void __intel_pmu_pebs_update_cfg(struct perf_event *event,
1406 					int idx, u64 *pebs_data_cfg)
1407 {
1408 	if (is_metric_event(event)) {
1409 		*pebs_data_cfg |= PEBS_DATACFG_METRICS;
1410 		return;
1411 	}
1412 
1413 	*pebs_data_cfg |= PEBS_DATACFG_CNTR;
1414 
1415 	if (idx >= INTEL_PMC_IDX_FIXED)
1416 		*pebs_data_cfg |= PEBS_DATACFG_FIX_BIT(idx - INTEL_PMC_IDX_FIXED);
1417 	else
1418 		*pebs_data_cfg |= PEBS_DATACFG_CNTR_BIT(idx);
1419 }
1420 
1421 
1422 void intel_pmu_pebs_late_setup(struct cpu_hw_events *cpuc)
1423 {
1424 	struct perf_event *event;
1425 	u64 pebs_data_cfg = 0;
1426 	int i;
1427 
1428 	for (i = 0; i < cpuc->n_events; i++) {
1429 		event = cpuc->event_list[i];
1430 		if (!is_pebs_counter_event_group(event))
1431 			continue;
1432 		__intel_pmu_pebs_update_cfg(event, cpuc->assign[i], &pebs_data_cfg);
1433 	}
1434 
1435 	if (pebs_data_cfg & ~cpuc->pebs_data_cfg)
1436 		cpuc->pebs_data_cfg |= pebs_data_cfg | PEBS_UPDATE_DS_SW;
1437 }
1438 
1439 #define PERF_PEBS_MEMINFO_TYPE	(PERF_SAMPLE_ADDR | PERF_SAMPLE_DATA_SRC |   \
1440 				PERF_SAMPLE_PHYS_ADDR |			     \
1441 				PERF_SAMPLE_WEIGHT_TYPE |		     \
1442 				PERF_SAMPLE_TRANSACTION |		     \
1443 				PERF_SAMPLE_DATA_PAGE_SIZE)
1444 
1445 static u64 pebs_update_adaptive_cfg(struct perf_event *event)
1446 {
1447 	struct perf_event_attr *attr = &event->attr;
1448 	u64 sample_type = attr->sample_type;
1449 	u64 pebs_data_cfg = 0;
1450 	bool gprs, tsx_weight;
1451 
1452 	if (!(sample_type & ~(PERF_SAMPLE_IP|PERF_SAMPLE_TIME)) &&
1453 	    attr->precise_ip > 1)
1454 		return pebs_data_cfg;
1455 
1456 	if (sample_type & PERF_PEBS_MEMINFO_TYPE)
1457 		pebs_data_cfg |= PEBS_DATACFG_MEMINFO;
1458 
1459 	/*
1460 	 * We need GPRs when:
1461 	 * + user requested them
1462 	 * + precise_ip < 2 for the non event IP
1463 	 * + For RTM TSX weight we need GPRs for the abort code.
1464 	 */
1465 	gprs = ((sample_type & PERF_SAMPLE_REGS_INTR) &&
1466 		(attr->sample_regs_intr & PEBS_GP_REGS)) ||
1467 	       ((sample_type & PERF_SAMPLE_REGS_USER) &&
1468 		(attr->sample_regs_user & PEBS_GP_REGS));
1469 
1470 	tsx_weight = (sample_type & PERF_SAMPLE_WEIGHT_TYPE) &&
1471 		     ((attr->config & INTEL_ARCH_EVENT_MASK) ==
1472 		      x86_pmu.rtm_abort_event);
1473 
1474 	if (gprs || (attr->precise_ip < 2) || tsx_weight)
1475 		pebs_data_cfg |= PEBS_DATACFG_GP;
1476 
1477 	if ((sample_type & PERF_SAMPLE_REGS_INTR) &&
1478 	    (attr->sample_regs_intr & PERF_REG_EXTENDED_MASK))
1479 		pebs_data_cfg |= PEBS_DATACFG_XMMS;
1480 
1481 	if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
1482 		/*
1483 		 * For now always log all LBRs. Could configure this
1484 		 * later.
1485 		 */
1486 		pebs_data_cfg |= PEBS_DATACFG_LBRS |
1487 			((x86_pmu.lbr_nr-1) << PEBS_DATACFG_LBR_SHIFT);
1488 	}
1489 
1490 	return pebs_data_cfg;
1491 }
1492 
1493 static void
1494 pebs_update_state(bool needed_cb, struct cpu_hw_events *cpuc,
1495 		  struct perf_event *event, bool add)
1496 {
1497 	struct pmu *pmu = event->pmu;
1498 
1499 	/*
1500 	 * Make sure we get updated with the first PEBS event.
1501 	 * During removal, ->pebs_data_cfg is still valid for
1502 	 * the last PEBS event. Don't clear it.
1503 	 */
1504 	if ((cpuc->n_pebs == 1) && add)
1505 		cpuc->pebs_data_cfg = PEBS_UPDATE_DS_SW;
1506 
1507 	if (needed_cb != pebs_needs_sched_cb(cpuc)) {
1508 		if (!needed_cb)
1509 			perf_sched_cb_inc(pmu);
1510 		else
1511 			perf_sched_cb_dec(pmu);
1512 
1513 		cpuc->pebs_data_cfg |= PEBS_UPDATE_DS_SW;
1514 	}
1515 
1516 	/*
1517 	 * The PEBS record doesn't shrink on pmu::del(). Doing so would require
1518 	 * iterating all remaining PEBS events to reconstruct the config.
1519 	 */
1520 	if (x86_pmu.intel_cap.pebs_baseline && add) {
1521 		u64 pebs_data_cfg;
1522 
1523 		pebs_data_cfg = pebs_update_adaptive_cfg(event);
1524 		/*
1525 		 * Be sure to update the thresholds when we change the record.
1526 		 */
1527 		if (pebs_data_cfg & ~cpuc->pebs_data_cfg)
1528 			cpuc->pebs_data_cfg |= pebs_data_cfg | PEBS_UPDATE_DS_SW;
1529 	}
1530 }
1531 
1532 u64 intel_get_arch_pebs_data_config(struct perf_event *event)
1533 {
1534 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1535 	u64 pebs_data_cfg = 0;
1536 	u64 cntr_mask;
1537 
1538 	if (WARN_ON(event->hw.idx < 0 || event->hw.idx >= X86_PMC_IDX_MAX))
1539 		return 0;
1540 
1541 	pebs_data_cfg |= pebs_update_adaptive_cfg(event);
1542 
1543 	cntr_mask = (PEBS_DATACFG_CNTR_MASK << PEBS_DATACFG_CNTR_SHIFT) |
1544 		    (PEBS_DATACFG_FIX_MASK << PEBS_DATACFG_FIX_SHIFT) |
1545 		    PEBS_DATACFG_CNTR | PEBS_DATACFG_METRICS;
1546 	pebs_data_cfg |= cpuc->pebs_data_cfg & cntr_mask;
1547 
1548 	return pebs_data_cfg;
1549 }
1550 
1551 void intel_pmu_pebs_add(struct perf_event *event)
1552 {
1553 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1554 	struct hw_perf_event *hwc = &event->hw;
1555 	bool needed_cb = pebs_needs_sched_cb(cpuc);
1556 
1557 	cpuc->n_pebs++;
1558 	if (hwc->flags & PERF_X86_EVENT_LARGE_PEBS)
1559 		cpuc->n_large_pebs++;
1560 	if (hwc->flags & PERF_X86_EVENT_PEBS_VIA_PT)
1561 		cpuc->n_pebs_via_pt++;
1562 
1563 	pebs_update_state(needed_cb, cpuc, event, true);
1564 }
1565 
1566 static void intel_pmu_pebs_via_pt_disable(struct perf_event *event)
1567 {
1568 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1569 
1570 	if (!is_pebs_pt(event))
1571 		return;
1572 
1573 	if (!(cpuc->pebs_enabled & ~PEBS_VIA_PT_MASK))
1574 		cpuc->pebs_enabled &= ~PEBS_VIA_PT_MASK;
1575 }
1576 
1577 static void intel_pmu_pebs_via_pt_enable(struct perf_event *event)
1578 {
1579 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1580 	struct hw_perf_event *hwc = &event->hw;
1581 	struct debug_store *ds = cpuc->ds;
1582 	u64 value = ds->pebs_event_reset[hwc->idx];
1583 	u32 base = MSR_RELOAD_PMC0;
1584 	unsigned int idx = hwc->idx;
1585 
1586 	if (!is_pebs_pt(event))
1587 		return;
1588 
1589 	if (!(event->hw.flags & PERF_X86_EVENT_LARGE_PEBS))
1590 		cpuc->pebs_enabled |= PEBS_PMI_AFTER_EACH_RECORD;
1591 
1592 	cpuc->pebs_enabled |= PEBS_OUTPUT_PT;
1593 
1594 	if (hwc->idx >= INTEL_PMC_IDX_FIXED) {
1595 		base = MSR_RELOAD_FIXED_CTR0;
1596 		idx = hwc->idx - INTEL_PMC_IDX_FIXED;
1597 		if (x86_pmu.intel_cap.pebs_format < 5)
1598 			value = ds->pebs_event_reset[MAX_PEBS_EVENTS_FMT4 + idx];
1599 		else
1600 			value = ds->pebs_event_reset[MAX_PEBS_EVENTS + idx];
1601 	}
1602 	wrmsrq(base + idx, value);
1603 }
1604 
1605 static inline void intel_pmu_drain_large_pebs(struct cpu_hw_events *cpuc)
1606 {
1607 	if (cpuc->n_pebs == cpuc->n_large_pebs &&
1608 	    cpuc->n_pebs != cpuc->n_pebs_via_pt)
1609 		intel_pmu_drain_pebs_buffer();
1610 }
1611 
1612 static void __intel_pmu_pebs_enable(struct perf_event *event)
1613 {
1614 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1615 	struct hw_perf_event *hwc = &event->hw;
1616 
1617 	hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
1618 	cpuc->pebs_enabled |= 1ULL << hwc->idx;
1619 }
1620 
1621 void intel_pmu_pebs_enable(struct perf_event *event)
1622 {
1623 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1624 	u64 pebs_data_cfg = cpuc->pebs_data_cfg & ~PEBS_UPDATE_DS_SW;
1625 	struct hw_perf_event *hwc = &event->hw;
1626 	struct debug_store *ds = cpuc->ds;
1627 	unsigned int idx = hwc->idx;
1628 
1629 	__intel_pmu_pebs_enable(event);
1630 
1631 	if ((event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT) && (x86_pmu.version < 5))
1632 		cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32);
1633 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
1634 		cpuc->pebs_enabled |= 1ULL << 63;
1635 
1636 	if (x86_pmu.intel_cap.pebs_baseline) {
1637 		hwc->config |= ICL_EVENTSEL_ADAPTIVE;
1638 		if (pebs_data_cfg != cpuc->active_pebs_data_cfg) {
1639 			/*
1640 			 * drain_pebs() assumes uniform record size;
1641 			 * hence we need to drain when changing said
1642 			 * size.
1643 			 */
1644 			intel_pmu_drain_pebs_buffer();
1645 			adaptive_pebs_record_size_update();
1646 			wrmsrq(MSR_PEBS_DATA_CFG, pebs_data_cfg);
1647 			cpuc->active_pebs_data_cfg = pebs_data_cfg;
1648 		}
1649 	}
1650 	if (cpuc->pebs_data_cfg & PEBS_UPDATE_DS_SW) {
1651 		cpuc->pebs_data_cfg = pebs_data_cfg;
1652 		pebs_update_threshold(cpuc);
1653 	}
1654 
1655 	if (idx >= INTEL_PMC_IDX_FIXED) {
1656 		if (x86_pmu.intel_cap.pebs_format < 5)
1657 			idx = MAX_PEBS_EVENTS_FMT4 + (idx - INTEL_PMC_IDX_FIXED);
1658 		else
1659 			idx = MAX_PEBS_EVENTS + (idx - INTEL_PMC_IDX_FIXED);
1660 	}
1661 
1662 	/*
1663 	 * Use auto-reload if possible to save a MSR write in the PMI.
1664 	 * This must be done in pmu::start(), because PERF_EVENT_IOC_PERIOD.
1665 	 */
1666 	if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
1667 		ds->pebs_event_reset[idx] =
1668 			(u64)(-hwc->sample_period) & x86_pmu.cntval_mask;
1669 	} else {
1670 		ds->pebs_event_reset[idx] = 0;
1671 	}
1672 
1673 	intel_pmu_pebs_via_pt_enable(event);
1674 }
1675 
1676 void intel_pmu_pebs_del(struct perf_event *event)
1677 {
1678 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1679 	struct hw_perf_event *hwc = &event->hw;
1680 	bool needed_cb = pebs_needs_sched_cb(cpuc);
1681 
1682 	cpuc->n_pebs--;
1683 	if (hwc->flags & PERF_X86_EVENT_LARGE_PEBS)
1684 		cpuc->n_large_pebs--;
1685 	if (hwc->flags & PERF_X86_EVENT_PEBS_VIA_PT)
1686 		cpuc->n_pebs_via_pt--;
1687 
1688 	pebs_update_state(needed_cb, cpuc, event, false);
1689 }
1690 
1691 static void __intel_pmu_pebs_disable(struct perf_event *event)
1692 {
1693 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1694 	struct hw_perf_event *hwc = &event->hw;
1695 
1696 	intel_pmu_drain_large_pebs(cpuc);
1697 	cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
1698 	hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
1699 }
1700 
1701 void intel_pmu_pebs_disable(struct perf_event *event)
1702 {
1703 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1704 	struct hw_perf_event *hwc = &event->hw;
1705 
1706 	__intel_pmu_pebs_disable(event);
1707 
1708 	if ((event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT) &&
1709 	    (x86_pmu.version < 5))
1710 		cpuc->pebs_enabled &= ~(1ULL << (hwc->idx + 32));
1711 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
1712 		cpuc->pebs_enabled &= ~(1ULL << 63);
1713 
1714 	intel_pmu_pebs_via_pt_disable(event);
1715 
1716 	if (cpuc->enabled)
1717 		wrmsrq(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
1718 }
1719 
1720 void intel_pmu_pebs_enable_all(void)
1721 {
1722 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1723 
1724 	if (cpuc->pebs_enabled)
1725 		wrmsrq(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
1726 }
1727 
1728 void intel_pmu_pebs_disable_all(void)
1729 {
1730 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1731 
1732 	if (cpuc->pebs_enabled)
1733 		__intel_pmu_pebs_disable_all();
1734 }
1735 
1736 static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
1737 {
1738 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1739 	unsigned long from = cpuc->lbr_entries[0].from;
1740 	unsigned long old_to, to = cpuc->lbr_entries[0].to;
1741 	unsigned long ip = regs->ip;
1742 	int is_64bit = 0;
1743 	void *kaddr;
1744 	int size;
1745 
1746 	/*
1747 	 * We don't need to fixup if the PEBS assist is fault like
1748 	 */
1749 	if (!x86_pmu.intel_cap.pebs_trap)
1750 		return 1;
1751 
1752 	/*
1753 	 * No LBR entry, no basic block, no rewinding
1754 	 */
1755 	if (!cpuc->lbr_stack.nr || !from || !to)
1756 		return 0;
1757 
1758 	/*
1759 	 * Basic blocks should never cross user/kernel boundaries
1760 	 */
1761 	if (kernel_ip(ip) != kernel_ip(to))
1762 		return 0;
1763 
1764 	/*
1765 	 * unsigned math, either ip is before the start (impossible) or
1766 	 * the basic block is larger than 1 page (sanity)
1767 	 */
1768 	if ((ip - to) > PEBS_FIXUP_SIZE)
1769 		return 0;
1770 
1771 	/*
1772 	 * We sampled a branch insn, rewind using the LBR stack
1773 	 */
1774 	if (ip == to) {
1775 		set_linear_ip(regs, from);
1776 		return 1;
1777 	}
1778 
1779 	size = ip - to;
1780 	if (!kernel_ip(ip)) {
1781 		int bytes;
1782 		u8 *buf = this_cpu_read(insn_buffer);
1783 
1784 		/* 'size' must fit our buffer, see above */
1785 		bytes = copy_from_user_nmi(buf, (void __user *)to, size);
1786 		if (bytes != 0)
1787 			return 0;
1788 
1789 		kaddr = buf;
1790 	} else {
1791 		kaddr = (void *)to;
1792 	}
1793 
1794 	do {
1795 		struct insn insn;
1796 
1797 		old_to = to;
1798 
1799 #ifdef CONFIG_X86_64
1800 		is_64bit = kernel_ip(to) || any_64bit_mode(regs);
1801 #endif
1802 		insn_init(&insn, kaddr, size, is_64bit);
1803 
1804 		/*
1805 		 * Make sure there was not a problem decoding the instruction.
1806 		 * This is doubly important because we have an infinite loop if
1807 		 * insn.length=0.
1808 		 */
1809 		if (insn_get_length(&insn))
1810 			break;
1811 
1812 		to += insn.length;
1813 		kaddr += insn.length;
1814 		size -= insn.length;
1815 	} while (to < ip);
1816 
1817 	if (to == ip) {
1818 		set_linear_ip(regs, old_to);
1819 		return 1;
1820 	}
1821 
1822 	/*
1823 	 * Even though we decoded the basic block, the instruction stream
1824 	 * never matched the given IP, either the TO or the IP got corrupted.
1825 	 */
1826 	return 0;
1827 }
1828 
1829 static inline u64 intel_get_tsx_weight(u64 tsx_tuning)
1830 {
1831 	if (tsx_tuning) {
1832 		union hsw_tsx_tuning tsx = { .value = tsx_tuning };
1833 		return tsx.cycles_last_block;
1834 	}
1835 	return 0;
1836 }
1837 
1838 static inline u64 intel_get_tsx_transaction(u64 tsx_tuning, u64 ax)
1839 {
1840 	u64 txn = (tsx_tuning & PEBS_HSW_TSX_FLAGS) >> 32;
1841 
1842 	/* For RTM XABORTs also log the abort code from AX */
1843 	if ((txn & PERF_TXN_TRANSACTION) && (ax & 1))
1844 		txn |= ((ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT;
1845 	return txn;
1846 }
1847 
1848 static inline u64 get_pebs_status(void *n)
1849 {
1850 	if (x86_pmu.intel_cap.pebs_format < 4)
1851 		return ((struct pebs_record_nhm *)n)->status;
1852 	return ((struct pebs_basic *)n)->applicable_counters;
1853 }
1854 
1855 #define PERF_X86_EVENT_PEBS_HSW_PREC \
1856 		(PERF_X86_EVENT_PEBS_ST_HSW | \
1857 		 PERF_X86_EVENT_PEBS_LD_HSW | \
1858 		 PERF_X86_EVENT_PEBS_NA_HSW)
1859 
1860 static u64 get_data_src(struct perf_event *event, u64 aux)
1861 {
1862 	u64 val = PERF_MEM_NA;
1863 	int fl = event->hw.flags;
1864 	bool fst = fl & (PERF_X86_EVENT_PEBS_ST | PERF_X86_EVENT_PEBS_HSW_PREC);
1865 
1866 	if (fl & PERF_X86_EVENT_PEBS_LDLAT)
1867 		val = load_latency_data(event, aux);
1868 	else if (fl & PERF_X86_EVENT_PEBS_STLAT)
1869 		val = store_latency_data(event, aux);
1870 	else if (fl & PERF_X86_EVENT_PEBS_LAT_HYBRID)
1871 		val = x86_pmu.pebs_latency_data(event, aux);
1872 	else if (fst && (fl & PERF_X86_EVENT_PEBS_HSW_PREC))
1873 		val = precise_datala_hsw(event, aux);
1874 	else if (fst)
1875 		val = precise_store_data(aux);
1876 	return val;
1877 }
1878 
1879 static void setup_pebs_time(struct perf_event *event,
1880 			    struct perf_sample_data *data,
1881 			    u64 tsc)
1882 {
1883 	/* Converting to a user-defined clock is not supported yet. */
1884 	if (event->attr.use_clockid != 0)
1885 		return;
1886 
1887 	/*
1888 	 * Doesn't support the conversion when the TSC is unstable.
1889 	 * The TSC unstable case is a corner case and very unlikely to
1890 	 * happen. If it happens, the TSC in a PEBS record will be
1891 	 * dropped and fall back to perf_event_clock().
1892 	 */
1893 	if (!using_native_sched_clock() || !sched_clock_stable())
1894 		return;
1895 
1896 	data->time = native_sched_clock_from_tsc(tsc) + __sched_clock_offset;
1897 	data->sample_flags |= PERF_SAMPLE_TIME;
1898 }
1899 
1900 #define PERF_SAMPLE_ADDR_TYPE	(PERF_SAMPLE_ADDR |		\
1901 				 PERF_SAMPLE_PHYS_ADDR |	\
1902 				 PERF_SAMPLE_DATA_PAGE_SIZE)
1903 
1904 static void setup_pebs_fixed_sample_data(struct perf_event *event,
1905 				   struct pt_regs *iregs, void *__pebs,
1906 				   struct perf_sample_data *data,
1907 				   struct pt_regs *regs)
1908 {
1909 	/*
1910 	 * We cast to the biggest pebs_record but are careful not to
1911 	 * unconditionally access the 'extra' entries.
1912 	 */
1913 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1914 	struct pebs_record_skl *pebs = __pebs;
1915 	u64 sample_type;
1916 	int fll;
1917 
1918 	if (pebs == NULL)
1919 		return;
1920 
1921 	sample_type = event->attr.sample_type;
1922 	fll = event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT;
1923 
1924 	perf_sample_data_init(data, 0, event->hw.last_period);
1925 
1926 	/*
1927 	 * Use latency for weight (only avail with PEBS-LL)
1928 	 */
1929 	if (fll && (sample_type & PERF_SAMPLE_WEIGHT_TYPE)) {
1930 		data->weight.full = pebs->lat;
1931 		data->sample_flags |= PERF_SAMPLE_WEIGHT_TYPE;
1932 	}
1933 
1934 	/*
1935 	 * data.data_src encodes the data source
1936 	 */
1937 	if (sample_type & PERF_SAMPLE_DATA_SRC) {
1938 		data->data_src.val = get_data_src(event, pebs->dse);
1939 		data->sample_flags |= PERF_SAMPLE_DATA_SRC;
1940 	}
1941 
1942 	/*
1943 	 * We must however always use iregs for the unwinder to stay sane; the
1944 	 * record BP,SP,IP can point into thin air when the record is from a
1945 	 * previous PMI context or an (I)RET happened between the record and
1946 	 * PMI.
1947 	 */
1948 	perf_sample_save_callchain(data, event, iregs);
1949 
1950 	/*
1951 	 * We use the interrupt regs as a base because the PEBS record does not
1952 	 * contain a full regs set, specifically it seems to lack segment
1953 	 * descriptors, which get used by things like user_mode().
1954 	 *
1955 	 * In the simple case fix up only the IP for PERF_SAMPLE_IP.
1956 	 */
1957 	*regs = *iregs;
1958 
1959 	/*
1960 	 * Initialize regs_>flags from PEBS,
1961 	 * Clear exact bit (which uses x86 EFLAGS Reserved bit 3),
1962 	 * i.e., do not rely on it being zero:
1963 	 */
1964 	regs->flags = pebs->flags & ~PERF_EFLAGS_EXACT;
1965 
1966 	if (sample_type & PERF_SAMPLE_REGS_INTR) {
1967 		regs->ax = pebs->ax;
1968 		regs->bx = pebs->bx;
1969 		regs->cx = pebs->cx;
1970 		regs->dx = pebs->dx;
1971 		regs->si = pebs->si;
1972 		regs->di = pebs->di;
1973 
1974 		regs->bp = pebs->bp;
1975 		regs->sp = pebs->sp;
1976 
1977 #ifndef CONFIG_X86_32
1978 		regs->r8 = pebs->r8;
1979 		regs->r9 = pebs->r9;
1980 		regs->r10 = pebs->r10;
1981 		regs->r11 = pebs->r11;
1982 		regs->r12 = pebs->r12;
1983 		regs->r13 = pebs->r13;
1984 		regs->r14 = pebs->r14;
1985 		regs->r15 = pebs->r15;
1986 #endif
1987 	}
1988 
1989 	if (event->attr.precise_ip > 1) {
1990 		/*
1991 		 * Haswell and later processors have an 'eventing IP'
1992 		 * (real IP) which fixes the off-by-1 skid in hardware.
1993 		 * Use it when precise_ip >= 2 :
1994 		 */
1995 		if (x86_pmu.intel_cap.pebs_format >= 2) {
1996 			set_linear_ip(regs, pebs->real_ip);
1997 			regs->flags |= PERF_EFLAGS_EXACT;
1998 		} else {
1999 			/* Otherwise, use PEBS off-by-1 IP: */
2000 			set_linear_ip(regs, pebs->ip);
2001 
2002 			/*
2003 			 * With precise_ip >= 2, try to fix up the off-by-1 IP
2004 			 * using the LBR. If successful, the fixup function
2005 			 * corrects regs->ip and calls set_linear_ip() on regs:
2006 			 */
2007 			if (intel_pmu_pebs_fixup_ip(regs))
2008 				regs->flags |= PERF_EFLAGS_EXACT;
2009 		}
2010 	} else {
2011 		/*
2012 		 * When precise_ip == 1, return the PEBS off-by-1 IP,
2013 		 * no fixup attempted:
2014 		 */
2015 		set_linear_ip(regs, pebs->ip);
2016 	}
2017 
2018 
2019 	if ((sample_type & PERF_SAMPLE_ADDR_TYPE) &&
2020 	    x86_pmu.intel_cap.pebs_format >= 1) {
2021 		data->addr = pebs->dla;
2022 		data->sample_flags |= PERF_SAMPLE_ADDR;
2023 	}
2024 
2025 	if (x86_pmu.intel_cap.pebs_format >= 2) {
2026 		/* Only set the TSX weight when no memory weight. */
2027 		if ((sample_type & PERF_SAMPLE_WEIGHT_TYPE) && !fll) {
2028 			data->weight.full = intel_get_tsx_weight(pebs->tsx_tuning);
2029 			data->sample_flags |= PERF_SAMPLE_WEIGHT_TYPE;
2030 		}
2031 		if (sample_type & PERF_SAMPLE_TRANSACTION) {
2032 			data->txn = intel_get_tsx_transaction(pebs->tsx_tuning,
2033 							      pebs->ax);
2034 			data->sample_flags |= PERF_SAMPLE_TRANSACTION;
2035 		}
2036 	}
2037 
2038 	/*
2039 	 * v3 supplies an accurate time stamp, so we use that
2040 	 * for the time stamp.
2041 	 *
2042 	 * We can only do this for the default trace clock.
2043 	 */
2044 	if (x86_pmu.intel_cap.pebs_format >= 3)
2045 		setup_pebs_time(event, data, pebs->tsc);
2046 
2047 	perf_sample_save_brstack(data, event, &cpuc->lbr_stack, NULL);
2048 }
2049 
2050 static void adaptive_pebs_save_regs(struct pt_regs *regs,
2051 				    struct pebs_gprs *gprs)
2052 {
2053 	regs->ax = gprs->ax;
2054 	regs->bx = gprs->bx;
2055 	regs->cx = gprs->cx;
2056 	regs->dx = gprs->dx;
2057 	regs->si = gprs->si;
2058 	regs->di = gprs->di;
2059 	regs->bp = gprs->bp;
2060 	regs->sp = gprs->sp;
2061 #ifndef CONFIG_X86_32
2062 	regs->r8 = gprs->r8;
2063 	regs->r9 = gprs->r9;
2064 	regs->r10 = gprs->r10;
2065 	regs->r11 = gprs->r11;
2066 	regs->r12 = gprs->r12;
2067 	regs->r13 = gprs->r13;
2068 	regs->r14 = gprs->r14;
2069 	regs->r15 = gprs->r15;
2070 #endif
2071 }
2072 
2073 static void intel_perf_event_update_pmc(struct perf_event *event, u64 pmc)
2074 {
2075 	int shift = 64 - x86_pmu.cntval_bits;
2076 	struct hw_perf_event *hwc;
2077 	u64 delta, prev_pmc;
2078 
2079 	/*
2080 	 * A recorded counter may not have an assigned event in the
2081 	 * following cases. The value should be dropped.
2082 	 * - An event is deleted. There is still an active PEBS event.
2083 	 *   The PEBS record doesn't shrink on pmu::del().
2084 	 *   If the counter of the deleted event once occurred in a PEBS
2085 	 *   record, PEBS still records the counter until the counter is
2086 	 *   reassigned.
2087 	 * - An event is stopped for some reason, e.g., throttled.
2088 	 *   During this period, another event is added and takes the
2089 	 *   counter of the stopped event. The stopped event is assigned
2090 	 *   to another new and uninitialized counter, since the
2091 	 *   x86_pmu_start(RELOAD) is not invoked for a stopped event.
2092 	 *   The PEBS__DATA_CFG is updated regardless of the event state.
2093 	 *   The uninitialized counter can be recorded in a PEBS record.
2094 	 *   But the cpuc->events[uninitialized_counter] is always NULL,
2095 	 *   because the event is stopped. The uninitialized value is
2096 	 *   safely dropped.
2097 	 */
2098 	if (!event)
2099 		return;
2100 
2101 	hwc = &event->hw;
2102 	prev_pmc = local64_read(&hwc->prev_count);
2103 
2104 	/* Only update the count when the PMU is disabled */
2105 	WARN_ON(this_cpu_read(cpu_hw_events.enabled));
2106 	local64_set(&hwc->prev_count, pmc);
2107 
2108 	delta = (pmc << shift) - (prev_pmc << shift);
2109 	delta >>= shift;
2110 
2111 	local64_add(delta, &event->count);
2112 	local64_sub(delta, &hwc->period_left);
2113 }
2114 
2115 static inline void __setup_pebs_counter_group(struct cpu_hw_events *cpuc,
2116 					      struct perf_event *event,
2117 					      struct pebs_cntr_header *cntr,
2118 					      void *next_record)
2119 {
2120 	int bit;
2121 
2122 	for_each_set_bit(bit, (unsigned long *)&cntr->cntr, INTEL_PMC_MAX_GENERIC) {
2123 		intel_perf_event_update_pmc(cpuc->events[bit], *(u64 *)next_record);
2124 		next_record += sizeof(u64);
2125 	}
2126 
2127 	for_each_set_bit(bit, (unsigned long *)&cntr->fixed, INTEL_PMC_MAX_FIXED) {
2128 		/* The slots event will be handled with perf_metric later */
2129 		if ((cntr->metrics == INTEL_CNTR_METRICS) &&
2130 		    (bit + INTEL_PMC_IDX_FIXED == INTEL_PMC_IDX_FIXED_SLOTS)) {
2131 			next_record += sizeof(u64);
2132 			continue;
2133 		}
2134 		intel_perf_event_update_pmc(cpuc->events[bit + INTEL_PMC_IDX_FIXED],
2135 					    *(u64 *)next_record);
2136 		next_record += sizeof(u64);
2137 	}
2138 
2139 	/* HW will reload the value right after the overflow. */
2140 	if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
2141 		local64_set(&event->hw.prev_count, (u64)-event->hw.sample_period);
2142 
2143 	if (cntr->metrics == INTEL_CNTR_METRICS) {
2144 		static_call(intel_pmu_update_topdown_event)
2145 			   (cpuc->events[INTEL_PMC_IDX_FIXED_SLOTS],
2146 			    (u64 *)next_record);
2147 		next_record += 2 * sizeof(u64);
2148 	}
2149 }
2150 
2151 #define PEBS_LATENCY_MASK			0xffff
2152 
2153 static inline void __setup_perf_sample_data(struct perf_event *event,
2154 					    struct pt_regs *iregs,
2155 					    struct perf_sample_data *data)
2156 {
2157 	perf_sample_data_init(data, 0, event->hw.last_period);
2158 
2159 	/*
2160 	 * We must however always use iregs for the unwinder to stay sane; the
2161 	 * record BP,SP,IP can point into thin air when the record is from a
2162 	 * previous PMI context or an (I)RET happened between the record and
2163 	 * PMI.
2164 	 */
2165 	perf_sample_save_callchain(data, event, iregs);
2166 }
2167 
2168 static inline void __setup_pebs_basic_group(struct perf_event *event,
2169 					    struct pt_regs *regs,
2170 					    struct perf_sample_data *data,
2171 					    u64 sample_type, u64 ip,
2172 					    u64 tsc, u16 retire)
2173 {
2174 	/* The ip in basic is EventingIP */
2175 	set_linear_ip(regs, ip);
2176 	regs->flags = PERF_EFLAGS_EXACT;
2177 	setup_pebs_time(event, data, tsc);
2178 
2179 	if (sample_type & PERF_SAMPLE_WEIGHT_STRUCT)
2180 		data->weight.var3_w = retire;
2181 }
2182 
2183 static inline void __setup_pebs_gpr_group(struct perf_event *event,
2184 					  struct pt_regs *regs,
2185 					  struct pebs_gprs *gprs,
2186 					  u64 sample_type)
2187 {
2188 	if (event->attr.precise_ip < 2) {
2189 		set_linear_ip(regs, gprs->ip);
2190 		regs->flags &= ~PERF_EFLAGS_EXACT;
2191 	}
2192 
2193 	if (sample_type & (PERF_SAMPLE_REGS_INTR | PERF_SAMPLE_REGS_USER))
2194 		adaptive_pebs_save_regs(regs, gprs);
2195 }
2196 
2197 static inline void __setup_pebs_meminfo_group(struct perf_event *event,
2198 					      struct perf_sample_data *data,
2199 					      u64 sample_type, u64 latency,
2200 					      u16 instr_latency, u64 address,
2201 					      u64 aux, u64 tsx_tuning, u64 ax)
2202 {
2203 	if (sample_type & PERF_SAMPLE_WEIGHT_TYPE) {
2204 		u64 tsx_latency = intel_get_tsx_weight(tsx_tuning);
2205 
2206 		data->weight.var2_w = instr_latency;
2207 
2208 		/*
2209 		 * Although meminfo::latency is defined as a u64,
2210 		 * only the lower 32 bits include the valid data
2211 		 * in practice on Ice Lake and earlier platforms.
2212 		 */
2213 		if (sample_type & PERF_SAMPLE_WEIGHT)
2214 			data->weight.full = latency ?: tsx_latency;
2215 		else
2216 			data->weight.var1_dw = (u32)latency ?: tsx_latency;
2217 
2218 		data->sample_flags |= PERF_SAMPLE_WEIGHT_TYPE;
2219 	}
2220 
2221 	if (sample_type & PERF_SAMPLE_DATA_SRC) {
2222 		data->data_src.val = get_data_src(event, aux);
2223 		data->sample_flags |= PERF_SAMPLE_DATA_SRC;
2224 	}
2225 
2226 	if (sample_type & PERF_SAMPLE_ADDR_TYPE) {
2227 		data->addr = address;
2228 		data->sample_flags |= PERF_SAMPLE_ADDR;
2229 	}
2230 
2231 	if (sample_type & PERF_SAMPLE_TRANSACTION) {
2232 		data->txn = intel_get_tsx_transaction(tsx_tuning, ax);
2233 		data->sample_flags |= PERF_SAMPLE_TRANSACTION;
2234 	}
2235 }
2236 
2237 /*
2238  * With adaptive PEBS the layout depends on what fields are configured.
2239  */
2240 static void setup_pebs_adaptive_sample_data(struct perf_event *event,
2241 					    struct pt_regs *iregs, void *__pebs,
2242 					    struct perf_sample_data *data,
2243 					    struct pt_regs *regs)
2244 {
2245 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2246 	u64 sample_type = event->attr.sample_type;
2247 	struct pebs_basic *basic = __pebs;
2248 	void *next_record = basic + 1;
2249 	struct pebs_meminfo *meminfo = NULL;
2250 	struct pebs_gprs *gprs = NULL;
2251 	struct x86_perf_regs *perf_regs;
2252 	u64 format_group;
2253 	u16 retire;
2254 
2255 	if (basic == NULL)
2256 		return;
2257 
2258 	perf_regs = container_of(regs, struct x86_perf_regs, regs);
2259 	perf_regs->xmm_regs = NULL;
2260 
2261 	format_group = basic->format_group;
2262 
2263 	__setup_perf_sample_data(event, iregs, data);
2264 
2265 	*regs = *iregs;
2266 
2267 	/* basic group */
2268 	retire = x86_pmu.flags & PMU_FL_RETIRE_LATENCY ?
2269 			basic->retire_latency : 0;
2270 	__setup_pebs_basic_group(event, regs, data, sample_type,
2271 				 basic->ip, basic->tsc, retire);
2272 
2273 	/*
2274 	 * The record for MEMINFO is in front of GP
2275 	 * But PERF_SAMPLE_TRANSACTION needs gprs->ax.
2276 	 * Save the pointer here but process later.
2277 	 */
2278 	if (format_group & PEBS_DATACFG_MEMINFO) {
2279 		meminfo = next_record;
2280 		next_record = meminfo + 1;
2281 	}
2282 
2283 	if (format_group & PEBS_DATACFG_GP) {
2284 		gprs = next_record;
2285 		next_record = gprs + 1;
2286 
2287 		__setup_pebs_gpr_group(event, regs, gprs, sample_type);
2288 	}
2289 
2290 	if (format_group & PEBS_DATACFG_MEMINFO) {
2291 		u64 latency = x86_pmu.flags & PMU_FL_INSTR_LATENCY ?
2292 				meminfo->cache_latency : meminfo->mem_latency;
2293 		u64 instr_latency = x86_pmu.flags & PMU_FL_INSTR_LATENCY ?
2294 				meminfo->instr_latency : 0;
2295 		u64 ax = gprs ? gprs->ax : 0;
2296 
2297 		__setup_pebs_meminfo_group(event, data, sample_type, latency,
2298 					   instr_latency, meminfo->address,
2299 					   meminfo->aux, meminfo->tsx_tuning,
2300 					   ax);
2301 	}
2302 
2303 	if (format_group & PEBS_DATACFG_XMMS) {
2304 		struct pebs_xmm *xmm = next_record;
2305 
2306 		next_record = xmm + 1;
2307 		perf_regs->xmm_regs = xmm->xmm;
2308 	}
2309 
2310 	if (format_group & PEBS_DATACFG_LBRS) {
2311 		struct lbr_entry *lbr = next_record;
2312 		int num_lbr = ((format_group >> PEBS_DATACFG_LBR_SHIFT)
2313 					& 0xff) + 1;
2314 		next_record = next_record + num_lbr * sizeof(struct lbr_entry);
2315 
2316 		if (has_branch_stack(event)) {
2317 			intel_pmu_store_pebs_lbrs(lbr);
2318 			intel_pmu_lbr_save_brstack(data, cpuc, event);
2319 		}
2320 	}
2321 
2322 	if (format_group & (PEBS_DATACFG_CNTR | PEBS_DATACFG_METRICS)) {
2323 		struct pebs_cntr_header *cntr = next_record;
2324 		unsigned int nr;
2325 
2326 		next_record += sizeof(struct pebs_cntr_header);
2327 		/*
2328 		 * The PEBS_DATA_CFG is a global register, which is the
2329 		 * superset configuration for all PEBS events.
2330 		 * For the PEBS record of non-sample-read group, ignore
2331 		 * the counter snapshot fields.
2332 		 */
2333 		if (is_pebs_counter_event_group(event)) {
2334 			__setup_pebs_counter_group(cpuc, event, cntr, next_record);
2335 			data->sample_flags |= PERF_SAMPLE_READ;
2336 		}
2337 
2338 		nr = hweight32(cntr->cntr) + hweight32(cntr->fixed);
2339 		if (cntr->metrics == INTEL_CNTR_METRICS)
2340 			nr += 2;
2341 		next_record += nr * sizeof(u64);
2342 	}
2343 
2344 	WARN_ONCE(next_record != __pebs + basic->format_size,
2345 			"PEBS record size %u, expected %llu, config %llx\n",
2346 			basic->format_size,
2347 			(u64)(next_record - __pebs),
2348 			format_group);
2349 }
2350 
2351 static inline bool arch_pebs_record_continued(struct arch_pebs_header *header)
2352 {
2353 	/* Continue bit or null PEBS record indicates fragment follows. */
2354 	return header->cont || !(header->format & GENMASK_ULL(63, 16));
2355 }
2356 
2357 static void setup_arch_pebs_sample_data(struct perf_event *event,
2358 					struct pt_regs *iregs,
2359 					void *__pebs,
2360 					struct perf_sample_data *data,
2361 					struct pt_regs *regs)
2362 {
2363 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2364 	u64 sample_type = event->attr.sample_type;
2365 	struct arch_pebs_header *header = NULL;
2366 	struct arch_pebs_aux *meminfo = NULL;
2367 	struct arch_pebs_gprs *gprs = NULL;
2368 	struct x86_perf_regs *perf_regs;
2369 	void *next_record;
2370 	void *at = __pebs;
2371 
2372 	if (at == NULL)
2373 		return;
2374 
2375 	perf_regs = container_of(regs, struct x86_perf_regs, regs);
2376 	perf_regs->xmm_regs = NULL;
2377 
2378 	__setup_perf_sample_data(event, iregs, data);
2379 
2380 	*regs = *iregs;
2381 
2382 again:
2383 	header = at;
2384 	next_record = at + sizeof(struct arch_pebs_header);
2385 	if (header->basic) {
2386 		struct arch_pebs_basic *basic = next_record;
2387 		u16 retire = 0;
2388 
2389 		next_record = basic + 1;
2390 
2391 		if (sample_type & PERF_SAMPLE_WEIGHT_STRUCT)
2392 			retire = basic->valid ? basic->retire : 0;
2393 		__setup_pebs_basic_group(event, regs, data, sample_type,
2394 				 basic->ip, basic->tsc, retire);
2395 	}
2396 
2397 	/*
2398 	 * The record for MEMINFO is in front of GP
2399 	 * But PERF_SAMPLE_TRANSACTION needs gprs->ax.
2400 	 * Save the pointer here but process later.
2401 	 */
2402 	if (header->aux) {
2403 		meminfo = next_record;
2404 		next_record = meminfo + 1;
2405 	}
2406 
2407 	if (header->gpr) {
2408 		gprs = next_record;
2409 		next_record = gprs + 1;
2410 
2411 		__setup_pebs_gpr_group(event, regs,
2412 				       (struct pebs_gprs *)gprs,
2413 				       sample_type);
2414 	}
2415 
2416 	if (header->aux) {
2417 		u64 ax = gprs ? gprs->ax : 0;
2418 
2419 		__setup_pebs_meminfo_group(event, data, sample_type,
2420 					   meminfo->cache_latency,
2421 					   meminfo->instr_latency,
2422 					   meminfo->address, meminfo->aux,
2423 					   meminfo->tsx_tuning, ax);
2424 	}
2425 
2426 	if (header->xmm) {
2427 		struct pebs_xmm *xmm;
2428 
2429 		next_record += sizeof(struct arch_pebs_xer_header);
2430 
2431 		xmm = next_record;
2432 		perf_regs->xmm_regs = xmm->xmm;
2433 		next_record = xmm + 1;
2434 	}
2435 
2436 	if (header->lbr) {
2437 		struct arch_pebs_lbr_header *lbr_header = next_record;
2438 		struct lbr_entry *lbr;
2439 		int num_lbr;
2440 
2441 		next_record = lbr_header + 1;
2442 		lbr = next_record;
2443 
2444 		num_lbr = header->lbr == ARCH_PEBS_LBR_NUM_VAR ?
2445 				lbr_header->depth :
2446 				header->lbr * ARCH_PEBS_BASE_LBR_ENTRIES;
2447 		next_record += num_lbr * sizeof(struct lbr_entry);
2448 
2449 		if (has_branch_stack(event)) {
2450 			intel_pmu_store_pebs_lbrs(lbr);
2451 			intel_pmu_lbr_save_brstack(data, cpuc, event);
2452 		}
2453 	}
2454 
2455 	if (header->cntr) {
2456 		struct arch_pebs_cntr_header *cntr = next_record;
2457 		unsigned int nr;
2458 
2459 		next_record += sizeof(struct arch_pebs_cntr_header);
2460 
2461 		if (is_pebs_counter_event_group(event)) {
2462 			__setup_pebs_counter_group(cpuc, event,
2463 				(struct pebs_cntr_header *)cntr, next_record);
2464 			data->sample_flags |= PERF_SAMPLE_READ;
2465 		}
2466 
2467 		nr = hweight32(cntr->cntr) + hweight32(cntr->fixed);
2468 		if (cntr->metrics == INTEL_CNTR_METRICS)
2469 			nr += 2;
2470 		next_record += nr * sizeof(u64);
2471 	}
2472 
2473 	/* Parse followed fragments if there are. */
2474 	if (arch_pebs_record_continued(header)) {
2475 		at = at + header->size;
2476 		goto again;
2477 	}
2478 }
2479 
2480 static inline void *
2481 get_next_pebs_record_by_bit(void *base, void *top, int bit)
2482 {
2483 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2484 	void *at;
2485 	u64 pebs_status;
2486 
2487 	/*
2488 	 * fmt0 does not have a status bitfield (does not use
2489 	 * perf_record_nhm format)
2490 	 */
2491 	if (x86_pmu.intel_cap.pebs_format < 1)
2492 		return base;
2493 
2494 	if (base == NULL)
2495 		return NULL;
2496 
2497 	for (at = base; at < top; at += cpuc->pebs_record_size) {
2498 		unsigned long status = get_pebs_status(at);
2499 
2500 		if (test_bit(bit, (unsigned long *)&status)) {
2501 			/* PEBS v3 has accurate status bits */
2502 			if (x86_pmu.intel_cap.pebs_format >= 3)
2503 				return at;
2504 
2505 			if (status == (1 << bit))
2506 				return at;
2507 
2508 			/* clear non-PEBS bit and re-check */
2509 			pebs_status = status & cpuc->pebs_enabled;
2510 			pebs_status &= PEBS_COUNTER_MASK;
2511 			if (pebs_status == (1 << bit))
2512 				return at;
2513 		}
2514 	}
2515 	return NULL;
2516 }
2517 
2518 /*
2519  * Special variant of intel_pmu_save_and_restart() for auto-reload.
2520  */
2521 static int
2522 intel_pmu_save_and_restart_reload(struct perf_event *event, int count)
2523 {
2524 	struct hw_perf_event *hwc = &event->hw;
2525 	int shift = 64 - x86_pmu.cntval_bits;
2526 	u64 period = hwc->sample_period;
2527 	u64 prev_raw_count, new_raw_count;
2528 	s64 new, old;
2529 
2530 	WARN_ON(!period);
2531 
2532 	/*
2533 	 * drain_pebs() only happens when the PMU is disabled.
2534 	 */
2535 	WARN_ON(this_cpu_read(cpu_hw_events.enabled));
2536 
2537 	prev_raw_count = local64_read(&hwc->prev_count);
2538 	new_raw_count = rdpmc(hwc->event_base_rdpmc);
2539 	local64_set(&hwc->prev_count, new_raw_count);
2540 
2541 	/*
2542 	 * Since the counter increments a negative counter value and
2543 	 * overflows on the sign switch, giving the interval:
2544 	 *
2545 	 *   [-period, 0]
2546 	 *
2547 	 * the difference between two consecutive reads is:
2548 	 *
2549 	 *   A) value2 - value1;
2550 	 *      when no overflows have happened in between,
2551 	 *
2552 	 *   B) (0 - value1) + (value2 - (-period));
2553 	 *      when one overflow happened in between,
2554 	 *
2555 	 *   C) (0 - value1) + (n - 1) * (period) + (value2 - (-period));
2556 	 *      when @n overflows happened in between.
2557 	 *
2558 	 * Here A) is the obvious difference, B) is the extension to the
2559 	 * discrete interval, where the first term is to the top of the
2560 	 * interval and the second term is from the bottom of the next
2561 	 * interval and C) the extension to multiple intervals, where the
2562 	 * middle term is the whole intervals covered.
2563 	 *
2564 	 * An equivalent of C, by reduction, is:
2565 	 *
2566 	 *   value2 - value1 + n * period
2567 	 */
2568 	new = ((s64)(new_raw_count << shift) >> shift);
2569 	old = ((s64)(prev_raw_count << shift) >> shift);
2570 	local64_add(new - old + count * period, &event->count);
2571 
2572 	local64_set(&hwc->period_left, -new);
2573 
2574 	perf_event_update_userpage(event);
2575 
2576 	return 0;
2577 }
2578 
2579 typedef void (*setup_fn)(struct perf_event *, struct pt_regs *, void *,
2580 			 struct perf_sample_data *, struct pt_regs *);
2581 
2582 static struct pt_regs dummy_iregs;
2583 
2584 static __always_inline void
2585 __intel_pmu_pebs_event(struct perf_event *event,
2586 		       struct pt_regs *iregs,
2587 		       struct pt_regs *regs,
2588 		       struct perf_sample_data *data,
2589 		       void *at,
2590 		       setup_fn setup_sample)
2591 {
2592 	setup_sample(event, iregs, at, data, regs);
2593 	perf_event_output(event, data, regs);
2594 }
2595 
2596 static __always_inline void
2597 __intel_pmu_pebs_last_event(struct perf_event *event,
2598 			    struct pt_regs *iregs,
2599 			    struct pt_regs *regs,
2600 			    struct perf_sample_data *data,
2601 			    void *at,
2602 			    int count,
2603 			    setup_fn setup_sample)
2604 {
2605 	struct hw_perf_event *hwc = &event->hw;
2606 
2607 	setup_sample(event, iregs, at, data, regs);
2608 	if (iregs == &dummy_iregs) {
2609 		/*
2610 		 * The PEBS records may be drained in the non-overflow context,
2611 		 * e.g., large PEBS + context switch. Perf should treat the
2612 		 * last record the same as other PEBS records, and doesn't
2613 		 * invoke the generic overflow handler.
2614 		 */
2615 		perf_event_output(event, data, regs);
2616 	} else {
2617 		/*
2618 		 * All but the last records are processed.
2619 		 * The last one is left to be able to call the overflow handler.
2620 		 */
2621 		perf_event_overflow(event, data, regs);
2622 	}
2623 
2624 	if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
2625 		if ((is_pebs_counter_event_group(event))) {
2626 			/*
2627 			 * The value of each sample has been updated when setup
2628 			 * the corresponding sample data.
2629 			 */
2630 			perf_event_update_userpage(event);
2631 		} else {
2632 			/*
2633 			 * Now, auto-reload is only enabled in fixed period mode.
2634 			 * The reload value is always hwc->sample_period.
2635 			 * May need to change it, if auto-reload is enabled in
2636 			 * freq mode later.
2637 			 */
2638 			intel_pmu_save_and_restart_reload(event, count);
2639 		}
2640 	} else {
2641 		/*
2642 		 * For a non-precise event, it's possible the
2643 		 * counters-snapshotting records a positive value for the
2644 		 * overflowed event. Then the HW auto-reload mechanism
2645 		 * reset the counter to 0 immediately, because the
2646 		 * pebs_event_reset is cleared if the PERF_X86_EVENT_AUTO_RELOAD
2647 		 * is not set. The counter backwards may be observed in a
2648 		 * PMI handler.
2649 		 *
2650 		 * Since the event value has been updated when processing the
2651 		 * counters-snapshotting record, only needs to set the new
2652 		 * period for the counter.
2653 		 */
2654 		if (is_pebs_counter_event_group(event))
2655 			static_call(x86_pmu_set_period)(event);
2656 		else
2657 			intel_pmu_save_and_restart(event);
2658 	}
2659 }
2660 
2661 static __always_inline void
2662 __intel_pmu_pebs_events(struct perf_event *event,
2663 			struct pt_regs *iregs,
2664 			struct perf_sample_data *data,
2665 			void *base, void *top,
2666 			int bit, int count,
2667 			setup_fn setup_sample)
2668 {
2669 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2670 	struct x86_perf_regs perf_regs;
2671 	struct pt_regs *regs = &perf_regs.regs;
2672 	void *at = get_next_pebs_record_by_bit(base, top, bit);
2673 	int cnt = count;
2674 
2675 	if (!iregs)
2676 		iregs = &dummy_iregs;
2677 
2678 	while (cnt > 1) {
2679 		__intel_pmu_pebs_event(event, iregs, regs, data, at, setup_sample);
2680 		at += cpuc->pebs_record_size;
2681 		at = get_next_pebs_record_by_bit(at, top, bit);
2682 		cnt--;
2683 	}
2684 
2685 	__intel_pmu_pebs_last_event(event, iregs, regs, data, at, count, setup_sample);
2686 }
2687 
2688 static void intel_pmu_drain_pebs_core(struct pt_regs *iregs, struct perf_sample_data *data)
2689 {
2690 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2691 	struct debug_store *ds = cpuc->ds;
2692 	struct perf_event *event = cpuc->events[0]; /* PMC0 only */
2693 	struct pebs_record_core *at, *top;
2694 	int n;
2695 
2696 	if (!x86_pmu.pebs_active)
2697 		return;
2698 
2699 	at  = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
2700 	top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
2701 
2702 	/*
2703 	 * Whatever else happens, drain the thing
2704 	 */
2705 	ds->pebs_index = ds->pebs_buffer_base;
2706 
2707 	if (!test_bit(0, cpuc->active_mask))
2708 		return;
2709 
2710 	WARN_ON_ONCE(!event);
2711 
2712 	if (!event->attr.precise_ip)
2713 		return;
2714 
2715 	n = top - at;
2716 	if (n <= 0) {
2717 		if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
2718 			intel_pmu_save_and_restart_reload(event, 0);
2719 		return;
2720 	}
2721 
2722 	__intel_pmu_pebs_events(event, iregs, data, at, top, 0, n,
2723 				setup_pebs_fixed_sample_data);
2724 }
2725 
2726 static void intel_pmu_pebs_event_update_no_drain(struct cpu_hw_events *cpuc, u64 mask)
2727 {
2728 	u64 pebs_enabled = cpuc->pebs_enabled & mask;
2729 	struct perf_event *event;
2730 	int bit;
2731 
2732 	/*
2733 	 * The drain_pebs() could be called twice in a short period
2734 	 * for auto-reload event in pmu::read(). There are no
2735 	 * overflows have happened in between.
2736 	 * It needs to call intel_pmu_save_and_restart_reload() to
2737 	 * update the event->count for this case.
2738 	 */
2739 	for_each_set_bit(bit, (unsigned long *)&pebs_enabled, X86_PMC_IDX_MAX) {
2740 		event = cpuc->events[bit];
2741 		if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
2742 			intel_pmu_save_and_restart_reload(event, 0);
2743 	}
2744 }
2745 
2746 static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs, struct perf_sample_data *data)
2747 {
2748 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2749 	struct debug_store *ds = cpuc->ds;
2750 	struct perf_event *event;
2751 	void *base, *at, *top;
2752 	short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
2753 	short error[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
2754 	int max_pebs_events = intel_pmu_max_num_pebs(NULL);
2755 	int bit, i, size;
2756 	u64 mask;
2757 
2758 	if (!x86_pmu.pebs_active)
2759 		return;
2760 
2761 	base = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
2762 	top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
2763 
2764 	ds->pebs_index = ds->pebs_buffer_base;
2765 
2766 	mask = x86_pmu.pebs_events_mask;
2767 	size = max_pebs_events;
2768 	if (x86_pmu.flags & PMU_FL_PEBS_ALL) {
2769 		mask |= x86_pmu.fixed_cntr_mask64 << INTEL_PMC_IDX_FIXED;
2770 		size = INTEL_PMC_IDX_FIXED + x86_pmu_max_num_counters_fixed(NULL);
2771 	}
2772 
2773 	if (unlikely(base >= top)) {
2774 		intel_pmu_pebs_event_update_no_drain(cpuc, mask);
2775 		return;
2776 	}
2777 
2778 	for (at = base; at < top; at += x86_pmu.pebs_record_size) {
2779 		struct pebs_record_nhm *p = at;
2780 		u64 pebs_status;
2781 
2782 		pebs_status = p->status & cpuc->pebs_enabled;
2783 		pebs_status &= mask;
2784 
2785 		/* PEBS v3 has more accurate status bits */
2786 		if (x86_pmu.intel_cap.pebs_format >= 3) {
2787 			for_each_set_bit(bit, (unsigned long *)&pebs_status, size)
2788 				counts[bit]++;
2789 
2790 			continue;
2791 		}
2792 
2793 		/*
2794 		 * On some CPUs the PEBS status can be zero when PEBS is
2795 		 * racing with clearing of GLOBAL_STATUS.
2796 		 *
2797 		 * Normally we would drop that record, but in the
2798 		 * case when there is only a single active PEBS event
2799 		 * we can assume it's for that event.
2800 		 */
2801 		if (!pebs_status && cpuc->pebs_enabled &&
2802 			!(cpuc->pebs_enabled & (cpuc->pebs_enabled-1)))
2803 			pebs_status = p->status = cpuc->pebs_enabled;
2804 
2805 		bit = find_first_bit((unsigned long *)&pebs_status,
2806 				     max_pebs_events);
2807 
2808 		if (!(x86_pmu.pebs_events_mask & (1 << bit)))
2809 			continue;
2810 
2811 		/*
2812 		 * The PEBS hardware does not deal well with the situation
2813 		 * when events happen near to each other and multiple bits
2814 		 * are set. But it should happen rarely.
2815 		 *
2816 		 * If these events include one PEBS and multiple non-PEBS
2817 		 * events, it doesn't impact PEBS record. The record will
2818 		 * be handled normally. (slow path)
2819 		 *
2820 		 * If these events include two or more PEBS events, the
2821 		 * records for the events can be collapsed into a single
2822 		 * one, and it's not possible to reconstruct all events
2823 		 * that caused the PEBS record. It's called collision.
2824 		 * If collision happened, the record will be dropped.
2825 		 */
2826 		if (pebs_status != (1ULL << bit)) {
2827 			for_each_set_bit(i, (unsigned long *)&pebs_status, size)
2828 				error[i]++;
2829 			continue;
2830 		}
2831 
2832 		counts[bit]++;
2833 	}
2834 
2835 	for_each_set_bit(bit, (unsigned long *)&mask, size) {
2836 		if ((counts[bit] == 0) && (error[bit] == 0))
2837 			continue;
2838 
2839 		event = cpuc->events[bit];
2840 		if (WARN_ON_ONCE(!event))
2841 			continue;
2842 
2843 		if (WARN_ON_ONCE(!event->attr.precise_ip))
2844 			continue;
2845 
2846 		/* log dropped samples number */
2847 		if (error[bit]) {
2848 			perf_log_lost_samples(event, error[bit]);
2849 
2850 			if (iregs)
2851 				perf_event_account_interrupt(event);
2852 		}
2853 
2854 		if (counts[bit]) {
2855 			__intel_pmu_pebs_events(event, iregs, data, base,
2856 						top, bit, counts[bit],
2857 						setup_pebs_fixed_sample_data);
2858 		}
2859 	}
2860 }
2861 
2862 static __always_inline void
2863 __intel_pmu_handle_pebs_record(struct pt_regs *iregs,
2864 			       struct pt_regs *regs,
2865 			       struct perf_sample_data *data,
2866 			       void *at, u64 pebs_status,
2867 			       short *counts, void **last,
2868 			       setup_fn setup_sample)
2869 {
2870 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2871 	struct perf_event *event;
2872 	int bit;
2873 
2874 	for_each_set_bit(bit, (unsigned long *)&pebs_status, X86_PMC_IDX_MAX) {
2875 		event = cpuc->events[bit];
2876 
2877 		if (WARN_ON_ONCE(!event) ||
2878 		    WARN_ON_ONCE(!event->attr.precise_ip))
2879 			continue;
2880 
2881 		if (counts[bit]++) {
2882 			__intel_pmu_pebs_event(event, iregs, regs, data,
2883 					       last[bit], setup_sample);
2884 		}
2885 
2886 		last[bit] = at;
2887 	}
2888 }
2889 
2890 static __always_inline void
2891 __intel_pmu_handle_last_pebs_record(struct pt_regs *iregs,
2892 				    struct pt_regs *regs,
2893 				    struct perf_sample_data *data,
2894 				    u64 mask, short *counts, void **last,
2895 				    setup_fn setup_sample)
2896 {
2897 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2898 	struct perf_event *event;
2899 	int bit;
2900 
2901 	for_each_set_bit(bit, (unsigned long *)&mask, X86_PMC_IDX_MAX) {
2902 		if (!counts[bit])
2903 			continue;
2904 
2905 		event = cpuc->events[bit];
2906 
2907 		__intel_pmu_pebs_last_event(event, iregs, regs, data, last[bit],
2908 					    counts[bit], setup_sample);
2909 	}
2910 
2911 }
2912 
2913 static void intel_pmu_drain_pebs_icl(struct pt_regs *iregs, struct perf_sample_data *data)
2914 {
2915 	short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
2916 	void *last[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS];
2917 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2918 	struct debug_store *ds = cpuc->ds;
2919 	struct x86_perf_regs perf_regs;
2920 	struct pt_regs *regs = &perf_regs.regs;
2921 	struct pebs_basic *basic;
2922 	void *base, *at, *top;
2923 	u64 mask;
2924 
2925 	if (!x86_pmu.pebs_active)
2926 		return;
2927 
2928 	base = (struct pebs_basic *)(unsigned long)ds->pebs_buffer_base;
2929 	top = (struct pebs_basic *)(unsigned long)ds->pebs_index;
2930 
2931 	ds->pebs_index = ds->pebs_buffer_base;
2932 
2933 	mask = hybrid(cpuc->pmu, pebs_events_mask) |
2934 	       (hybrid(cpuc->pmu, fixed_cntr_mask64) << INTEL_PMC_IDX_FIXED);
2935 	mask &= cpuc->pebs_enabled;
2936 
2937 	if (unlikely(base >= top)) {
2938 		intel_pmu_pebs_event_update_no_drain(cpuc, mask);
2939 		return;
2940 	}
2941 
2942 	if (!iregs)
2943 		iregs = &dummy_iregs;
2944 
2945 	/* Process all but the last event for each counter. */
2946 	for (at = base; at < top; at += basic->format_size) {
2947 		u64 pebs_status;
2948 
2949 		basic = at;
2950 		if (basic->format_size != cpuc->pebs_record_size)
2951 			continue;
2952 
2953 		pebs_status = mask & basic->applicable_counters;
2954 		__intel_pmu_handle_pebs_record(iregs, regs, data, at,
2955 					       pebs_status, counts, last,
2956 					       setup_pebs_adaptive_sample_data);
2957 	}
2958 
2959 	__intel_pmu_handle_last_pebs_record(iregs, regs, data, mask, counts, last,
2960 					    setup_pebs_adaptive_sample_data);
2961 }
2962 
2963 static void intel_pmu_drain_arch_pebs(struct pt_regs *iregs,
2964 				      struct perf_sample_data *data)
2965 {
2966 	short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
2967 	void *last[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS];
2968 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2969 	union arch_pebs_index index;
2970 	struct x86_perf_regs perf_regs;
2971 	struct pt_regs *regs = &perf_regs.regs;
2972 	void *base, *at, *top;
2973 	u64 mask;
2974 
2975 	rdmsrq(MSR_IA32_PEBS_INDEX, index.whole);
2976 
2977 	if (unlikely(!index.wr)) {
2978 		intel_pmu_pebs_event_update_no_drain(cpuc, X86_PMC_IDX_MAX);
2979 		return;
2980 	}
2981 
2982 	base = cpuc->pebs_vaddr;
2983 	top = cpuc->pebs_vaddr + (index.wr << ARCH_PEBS_INDEX_WR_SHIFT);
2984 
2985 	index.wr = 0;
2986 	index.full = 0;
2987 	index.en = 1;
2988 	if (cpuc->n_pebs == cpuc->n_large_pebs)
2989 		index.thresh = ARCH_PEBS_THRESH_MULTI;
2990 	else
2991 		index.thresh = ARCH_PEBS_THRESH_SINGLE;
2992 	wrmsrq(MSR_IA32_PEBS_INDEX, index.whole);
2993 
2994 	mask = hybrid(cpuc->pmu, arch_pebs_cap).counters & cpuc->pebs_enabled;
2995 
2996 	if (!iregs)
2997 		iregs = &dummy_iregs;
2998 
2999 	/* Process all but the last event for each counter. */
3000 	for (at = base; at < top;) {
3001 		struct arch_pebs_header *header;
3002 		struct arch_pebs_basic *basic;
3003 		u64 pebs_status;
3004 
3005 		header = at;
3006 
3007 		if (WARN_ON_ONCE(!header->size))
3008 			break;
3009 
3010 		/* 1st fragment or single record must have basic group */
3011 		if (!header->basic) {
3012 			at += header->size;
3013 			continue;
3014 		}
3015 
3016 		basic = at + sizeof(struct arch_pebs_header);
3017 		pebs_status = mask & basic->applicable_counters;
3018 		__intel_pmu_handle_pebs_record(iregs, regs, data, at,
3019 					       pebs_status, counts, last,
3020 					       setup_arch_pebs_sample_data);
3021 
3022 		/* Skip non-last fragments */
3023 		while (arch_pebs_record_continued(header)) {
3024 			if (!header->size)
3025 				break;
3026 			at += header->size;
3027 			header = at;
3028 		}
3029 
3030 		/* Skip last fragment or the single record */
3031 		at += header->size;
3032 	}
3033 
3034 	__intel_pmu_handle_last_pebs_record(iregs, regs, data, mask,
3035 					    counts, last,
3036 					    setup_arch_pebs_sample_data);
3037 }
3038 
3039 static void __init intel_arch_pebs_init(void)
3040 {
3041 	/*
3042 	 * Current hybrid platforms always both support arch-PEBS or not
3043 	 * on all kinds of cores. So directly set x86_pmu.arch_pebs flag
3044 	 * if boot cpu supports arch-PEBS.
3045 	 */
3046 	x86_pmu.arch_pebs = 1;
3047 	x86_pmu.pebs_buffer_size = PEBS_BUFFER_SIZE;
3048 	x86_pmu.drain_pebs = intel_pmu_drain_arch_pebs;
3049 	x86_pmu.pebs_capable = ~0ULL;
3050 	x86_pmu.flags |= PMU_FL_PEBS_ALL;
3051 
3052 	x86_pmu.pebs_enable = __intel_pmu_pebs_enable;
3053 	x86_pmu.pebs_disable = __intel_pmu_pebs_disable;
3054 }
3055 
3056 /*
3057  * PEBS probe and setup
3058  */
3059 
3060 static void __init intel_ds_pebs_init(void)
3061 {
3062 	/*
3063 	 * No support for 32bit formats
3064 	 */
3065 	if (!boot_cpu_has(X86_FEATURE_DTES64))
3066 		return;
3067 
3068 	x86_pmu.ds_pebs = boot_cpu_has(X86_FEATURE_PEBS);
3069 	x86_pmu.pebs_buffer_size = PEBS_BUFFER_SIZE;
3070 	if (x86_pmu.version <= 4)
3071 		x86_pmu.pebs_no_isolation = 1;
3072 
3073 	if (x86_pmu.ds_pebs) {
3074 		char pebs_type = x86_pmu.intel_cap.pebs_trap ?  '+' : '-';
3075 		char *pebs_qual = "";
3076 		int format = x86_pmu.intel_cap.pebs_format;
3077 
3078 		if (format < 4)
3079 			x86_pmu.intel_cap.pebs_baseline = 0;
3080 
3081 		x86_pmu.pebs_enable = intel_pmu_pebs_enable;
3082 		x86_pmu.pebs_disable = intel_pmu_pebs_disable;
3083 		x86_pmu.pebs_enable_all = intel_pmu_pebs_enable_all;
3084 		x86_pmu.pebs_disable_all = intel_pmu_pebs_disable_all;
3085 
3086 		switch (format) {
3087 		case 0:
3088 			pr_cont("PEBS fmt0%c, ", pebs_type);
3089 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
3090 			/*
3091 			 * Using >PAGE_SIZE buffers makes the WRMSR to
3092 			 * PERF_GLOBAL_CTRL in intel_pmu_enable_all()
3093 			 * mysteriously hang on Core2.
3094 			 *
3095 			 * As a workaround, we don't do this.
3096 			 */
3097 			x86_pmu.pebs_buffer_size = PAGE_SIZE;
3098 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
3099 			break;
3100 
3101 		case 1:
3102 			pr_cont("PEBS fmt1%c, ", pebs_type);
3103 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
3104 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
3105 			break;
3106 
3107 		case 2:
3108 			pr_cont("PEBS fmt2%c, ", pebs_type);
3109 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_hsw);
3110 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
3111 			break;
3112 
3113 		case 3:
3114 			pr_cont("PEBS fmt3%c, ", pebs_type);
3115 			x86_pmu.pebs_record_size =
3116 						sizeof(struct pebs_record_skl);
3117 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
3118 			x86_pmu.large_pebs_flags |= PERF_SAMPLE_TIME;
3119 			break;
3120 
3121 		case 6:
3122 			if (x86_pmu.intel_cap.pebs_baseline)
3123 				x86_pmu.large_pebs_flags |= PERF_SAMPLE_READ;
3124 			fallthrough;
3125 		case 5:
3126 			x86_pmu.pebs_ept = 1;
3127 			fallthrough;
3128 		case 4:
3129 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_icl;
3130 			x86_pmu.pebs_record_size = sizeof(struct pebs_basic);
3131 			if (x86_pmu.intel_cap.pebs_baseline) {
3132 				x86_pmu.large_pebs_flags |=
3133 					PERF_SAMPLE_BRANCH_STACK |
3134 					PERF_SAMPLE_TIME;
3135 				x86_pmu.flags |= PMU_FL_PEBS_ALL;
3136 				x86_pmu.pebs_capable = ~0ULL;
3137 				pebs_qual = "-baseline";
3138 				x86_get_pmu(smp_processor_id())->capabilities |= PERF_PMU_CAP_EXTENDED_REGS;
3139 			} else {
3140 				/* Only basic record supported */
3141 				x86_pmu.large_pebs_flags &=
3142 					~(PERF_SAMPLE_ADDR |
3143 					  PERF_SAMPLE_TIME |
3144 					  PERF_SAMPLE_DATA_SRC |
3145 					  PERF_SAMPLE_TRANSACTION |
3146 					  PERF_SAMPLE_REGS_USER |
3147 					  PERF_SAMPLE_REGS_INTR);
3148 			}
3149 			pr_cont("PEBS fmt%d%c%s, ", format, pebs_type, pebs_qual);
3150 
3151 			/*
3152 			 * The PEBS-via-PT is not supported on hybrid platforms,
3153 			 * because not all CPUs of a hybrid machine support it.
3154 			 * The global x86_pmu.intel_cap, which only contains the
3155 			 * common capabilities, is used to check the availability
3156 			 * of the feature. The per-PMU pebs_output_pt_available
3157 			 * in a hybrid machine should be ignored.
3158 			 */
3159 			if (x86_pmu.intel_cap.pebs_output_pt_available) {
3160 				pr_cont("PEBS-via-PT, ");
3161 				x86_get_pmu(smp_processor_id())->capabilities |= PERF_PMU_CAP_AUX_OUTPUT;
3162 			}
3163 
3164 			break;
3165 
3166 		default:
3167 			pr_cont("no PEBS fmt%d%c, ", format, pebs_type);
3168 			x86_pmu.ds_pebs = 0;
3169 		}
3170 	}
3171 }
3172 
3173 void __init intel_pebs_init(void)
3174 {
3175 	if (x86_pmu.intel_cap.pebs_format == 0xf)
3176 		intel_arch_pebs_init();
3177 	else
3178 		intel_ds_pebs_init();
3179 }
3180 
3181 void perf_restore_debug_store(void)
3182 {
3183 	struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
3184 
3185 	if (!x86_pmu.bts && !x86_pmu.ds_pebs)
3186 		return;
3187 
3188 	wrmsrq(MSR_IA32_DS_AREA, (unsigned long)ds);
3189 }
3190