xref: /linux/arch/x86/events/intel/ds.c (revision c847a208f43bfeb56943f2ca6fe2baf1db9dee7a)
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 omr_encoding {
38 	struct {
39 		u8 omr_source : 4;
40 		u8 omr_remote : 1;
41 		u8 omr_hitm : 1;
42 		u8 omr_snoop : 1;
43 		u8 omr_promoted : 1;
44 	};
45 	u8 omr_full;
46 };
47 
48 union intel_x86_pebs_dse {
49 	u64 val;
50 	struct {
51 		unsigned int ld_dse:4;
52 		unsigned int ld_stlb_miss:1;
53 		unsigned int ld_locked:1;
54 		unsigned int ld_data_blk:1;
55 		unsigned int ld_addr_blk:1;
56 		unsigned int ld_reserved:24;
57 	};
58 	struct {
59 		unsigned int st_l1d_hit:1;
60 		unsigned int st_reserved1:3;
61 		unsigned int st_stlb_miss:1;
62 		unsigned int st_locked:1;
63 		unsigned int st_reserved2:26;
64 	};
65 	struct {
66 		unsigned int st_lat_dse:4;
67 		unsigned int st_lat_stlb_miss:1;
68 		unsigned int st_lat_locked:1;
69 		unsigned int ld_reserved3:26;
70 	};
71 	struct {
72 		unsigned int mtl_dse:5;
73 		unsigned int mtl_locked:1;
74 		unsigned int mtl_stlb_miss:1;
75 		unsigned int mtl_fwd_blk:1;
76 		unsigned int ld_reserved4:24;
77 	};
78 	struct {
79 		unsigned int lnc_dse:8;
80 		unsigned int ld_reserved5:2;
81 		unsigned int lnc_stlb_miss:1;
82 		unsigned int lnc_locked:1;
83 		unsigned int lnc_data_blk:1;
84 		unsigned int lnc_addr_blk:1;
85 		unsigned int ld_reserved6:18;
86 	};
87 	struct {
88 		unsigned int pnc_dse: 8;
89 		unsigned int pnc_l2_miss:1;
90 		unsigned int pnc_stlb_clean_hit:1;
91 		unsigned int pnc_stlb_any_hit:1;
92 		unsigned int pnc_stlb_miss:1;
93 		unsigned int pnc_locked:1;
94 		unsigned int pnc_data_blk:1;
95 		unsigned int pnc_addr_blk:1;
96 		unsigned int pnc_fb_full:1;
97 		unsigned int ld_reserved8:16;
98 	};
99 	struct {
100 		unsigned int arw_dse:8;
101 		unsigned int arw_l2_miss:1;
102 		unsigned int arw_xq_promotion:1;
103 		unsigned int arw_reissue:1;
104 		unsigned int arw_stlb_miss:1;
105 		unsigned int arw_locked:1;
106 		unsigned int arw_data_blk:1;
107 		unsigned int arw_addr_blk:1;
108 		unsigned int arw_fb_full:1;
109 		unsigned int ld_reserved9:16;
110 	};
111 };
112 
113 
114 /*
115  * Map PEBS Load Latency Data Source encodings to generic
116  * memory data source information
117  */
118 #define P(a, b) PERF_MEM_S(a, b)
119 #define OP_LH (P(OP, LOAD) | P(LVL, HIT))
120 #define LEVEL(x) P(LVLNUM, x)
121 #define REM P(REMOTE, REMOTE)
122 #define SNOOP_NONE_MISS (P(SNOOP, NONE) | P(SNOOP, MISS))
123 
124 /* Version for Sandy Bridge and later */
125 static u64 pebs_data_source[PERF_PEBS_DATA_SOURCE_MAX] = {
126 	P(OP, LOAD) | P(LVL, MISS) | LEVEL(L3) | P(SNOOP, NA),/* 0x00:ukn L3 */
127 	OP_LH | P(LVL, L1)  | LEVEL(L1) | P(SNOOP, NONE),  /* 0x01: L1 local */
128 	OP_LH | P(LVL, LFB) | LEVEL(LFB) | P(SNOOP, NONE), /* 0x02: LFB hit */
129 	OP_LH | P(LVL, L2)  | LEVEL(L2) | P(SNOOP, NONE),  /* 0x03: L2 hit */
130 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, NONE),  /* 0x04: L3 hit */
131 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, MISS),  /* 0x05: L3 hit, snoop miss */
132 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HIT),   /* 0x06: L3 hit, snoop hit */
133 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HITM),  /* 0x07: L3 hit, snoop hitm */
134 	OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HIT),  /* 0x08: L3 miss snoop hit */
135 	OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HITM), /* 0x09: L3 miss snoop hitm*/
136 	OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | P(SNOOP, HIT),       /* 0x0a: L3 miss, shared */
137 	OP_LH | P(LVL, REM_RAM1) | REM | LEVEL(L3) | P(SNOOP, HIT),  /* 0x0b: L3 miss, shared */
138 	OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | SNOOP_NONE_MISS,     /* 0x0c: L3 miss, excl */
139 	OP_LH | P(LVL, REM_RAM1) | LEVEL(RAM) | REM | SNOOP_NONE_MISS, /* 0x0d: L3 miss, excl */
140 	OP_LH | P(LVL, IO)  | LEVEL(NA) | P(SNOOP, NONE), /* 0x0e: I/O */
141 	OP_LH | P(LVL, UNC) | LEVEL(NA) | P(SNOOP, NONE), /* 0x0f: uncached */
142 };
143 
144 /* Patch up minor differences in the bits */
145 void __init intel_pmu_pebs_data_source_nhm(void)
146 {
147 	pebs_data_source[0x05] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HIT);
148 	pebs_data_source[0x06] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
149 	pebs_data_source[0x07] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
150 }
151 
152 static void __init __intel_pmu_pebs_data_source_skl(bool pmem, u64 *data_source)
153 {
154 	u64 pmem_or_l4 = pmem ? LEVEL(PMEM) : LEVEL(L4);
155 
156 	data_source[0x08] = OP_LH | pmem_or_l4 | P(SNOOP, HIT);
157 	data_source[0x09] = OP_LH | pmem_or_l4 | REM | P(SNOOP, HIT);
158 	data_source[0x0b] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, NONE);
159 	data_source[0x0c] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOPX, FWD);
160 	data_source[0x0d] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOP, HITM);
161 }
162 
163 void __init intel_pmu_pebs_data_source_skl(bool pmem)
164 {
165 	__intel_pmu_pebs_data_source_skl(pmem, pebs_data_source);
166 }
167 
168 static void __init __intel_pmu_pebs_data_source_grt(u64 *data_source)
169 {
170 	data_source[0x05] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HIT);
171 	data_source[0x06] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
172 	data_source[0x08] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOPX, FWD);
173 }
174 
175 void __init intel_pmu_pebs_data_source_grt(void)
176 {
177 	__intel_pmu_pebs_data_source_grt(pebs_data_source);
178 }
179 
180 void __init intel_pmu_pebs_data_source_adl(void)
181 {
182 	u64 *data_source;
183 
184 	data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX].pebs_data_source;
185 	memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
186 	__intel_pmu_pebs_data_source_skl(false, data_source);
187 
188 	data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX].pebs_data_source;
189 	memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
190 	__intel_pmu_pebs_data_source_grt(data_source);
191 }
192 
193 static void __init __intel_pmu_pebs_data_source_cmt(u64 *data_source)
194 {
195 	data_source[0x07] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOPX, FWD);
196 	data_source[0x08] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
197 	data_source[0x0a] = OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | P(SNOOP, NONE);
198 	data_source[0x0b] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, NONE);
199 	data_source[0x0c] = OP_LH | LEVEL(RAM) | REM | P(SNOOPX, FWD);
200 	data_source[0x0d] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, HITM);
201 }
202 
203 void __init intel_pmu_pebs_data_source_mtl(void)
204 {
205 	u64 *data_source;
206 
207 	data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX].pebs_data_source;
208 	memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
209 	__intel_pmu_pebs_data_source_skl(false, data_source);
210 
211 	data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX].pebs_data_source;
212 	memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
213 	__intel_pmu_pebs_data_source_cmt(data_source);
214 }
215 
216 void __init intel_pmu_pebs_data_source_arl_h(void)
217 {
218 	u64 *data_source;
219 
220 	intel_pmu_pebs_data_source_lnl();
221 
222 	data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_TINY_IDX].pebs_data_source;
223 	memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
224 	__intel_pmu_pebs_data_source_cmt(data_source);
225 }
226 
227 void __init intel_pmu_pebs_data_source_cmt(void)
228 {
229 	__intel_pmu_pebs_data_source_cmt(pebs_data_source);
230 }
231 
232 /* Version for Lion Cove and later */
233 static u64 lnc_pebs_data_source[PERF_PEBS_DATA_SOURCE_MAX] = {
234 	P(OP, LOAD) | P(LVL, MISS) | LEVEL(L3) | P(SNOOP, NA),	/* 0x00: ukn L3 */
235 	OP_LH | P(LVL, L1)  | LEVEL(L1) | P(SNOOP, NONE),	/* 0x01: L1 hit */
236 	OP_LH | P(LVL, L1)  | LEVEL(L1) | P(SNOOP, NONE),	/* 0x02: L1 hit */
237 	OP_LH | P(LVL, LFB) | LEVEL(LFB) | P(SNOOP, NONE),	/* 0x03: LFB/L1 Miss Handling Buffer hit */
238 	0,							/* 0x04: Reserved */
239 	OP_LH | P(LVL, L2)  | LEVEL(L2) | P(SNOOP, NONE),	/* 0x05: L2 Hit */
240 	OP_LH | LEVEL(L2_MHB) | P(SNOOP, NONE),			/* 0x06: L2 Miss Handling Buffer Hit */
241 	0,							/* 0x07: Reserved */
242 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, NONE),	/* 0x08: L3 Hit */
243 	0,							/* 0x09: Reserved */
244 	0,							/* 0x0a: Reserved */
245 	0,							/* 0x0b: Reserved */
246 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOPX, FWD),	/* 0x0c: L3 Hit Snoop Fwd */
247 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HITM),	/* 0x0d: L3 Hit Snoop HitM */
248 	0,							/* 0x0e: Reserved */
249 	P(OP, LOAD) | P(LVL, MISS) | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HITM),	/* 0x0f: L3 Miss Snoop HitM */
250 	OP_LH | LEVEL(MSC) | P(SNOOP, NONE),			/* 0x10: Memory-side Cache Hit */
251 	OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | P(SNOOP, NONE), /* 0x11: Local Memory Hit */
252 };
253 
254 void __init intel_pmu_pebs_data_source_lnl(void)
255 {
256 	u64 *data_source;
257 
258 	data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX].pebs_data_source;
259 	memcpy(data_source, lnc_pebs_data_source, sizeof(lnc_pebs_data_source));
260 
261 	data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX].pebs_data_source;
262 	memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
263 	__intel_pmu_pebs_data_source_cmt(data_source);
264 }
265 
266 /* Version for Panthercove and later */
267 
268 /* L2 hit */
269 #define PNC_PEBS_DATA_SOURCE_MAX	16
270 static u64 pnc_pebs_l2_hit_data_source[PNC_PEBS_DATA_SOURCE_MAX] = {
271 	P(OP, LOAD) | P(LVL, NA) | LEVEL(NA) | P(SNOOP, NA),	/* 0x00: non-cache access */
272 	OP_LH               | LEVEL(L0) | P(SNOOP, NONE),	/* 0x01: L0 hit */
273 	OP_LH | P(LVL, L1)  | LEVEL(L1) | P(SNOOP, NONE),	/* 0x02: L1 hit */
274 	OP_LH | P(LVL, LFB) | LEVEL(LFB) | P(SNOOP, NONE),	/* 0x03: L1 Miss Handling Buffer hit */
275 	OP_LH | P(LVL, L2)  | LEVEL(L2) | P(SNOOP, NONE),	/* 0x04: L2 Hit Clean */
276 	0,							/* 0x05: Reserved */
277 	0,							/* 0x06: Reserved */
278 	OP_LH | P(LVL, L2)  | LEVEL(L2) | P(SNOOP, HIT),	/* 0x07: L2 Hit Snoop HIT */
279 	OP_LH | P(LVL, L2)  | LEVEL(L2) | P(SNOOP, HITM),	/* 0x08: L2 Hit Snoop Hit Modified */
280 	OP_LH | P(LVL, L2)  | LEVEL(L2) | P(SNOOP, MISS),	/* 0x09: Prefetch Promotion */
281 	OP_LH | P(LVL, L2)  | LEVEL(L2) | P(SNOOP, MISS),	/* 0x0a: Cross Core Prefetch Promotion */
282 	0,							/* 0x0b: Reserved */
283 	0,							/* 0x0c: Reserved */
284 	0,							/* 0x0d: Reserved */
285 	0,							/* 0x0e: Reserved */
286 	OP_LH | P(LVL, UNC) | LEVEL(NA) | P(SNOOP, NONE),	/* 0x0f: uncached */
287 };
288 
289 /* Version for Arctic Wolf and later */
290 
291 /* L2 hit */
292 #define ARW_PEBS_DATA_SOURCE_MAX	16
293 static u64 arw_pebs_l2_hit_data_source[ARW_PEBS_DATA_SOURCE_MAX] = {
294 	P(OP, LOAD) | P(LVL, NA) | LEVEL(NA) | P(SNOOP, NA),	/* 0x00: non-cache access */
295 	OP_LH | P(LVL, L1)  | LEVEL(L1) | P(SNOOP, NONE),	/* 0x01: L1 hit */
296 	OP_LH | P(LVL, LFB) | LEVEL(LFB) | P(SNOOP, NONE),	/* 0x02: WCB Hit */
297 	OP_LH | P(LVL, L2)  | LEVEL(L2) | P(SNOOP, NONE),	/* 0x03: L2 Hit Clean */
298 	OP_LH | P(LVL, L2)  | LEVEL(L2) | P(SNOOP, HIT),	/* 0x04: L2 Hit Snoop HIT */
299 	OP_LH | P(LVL, L2)  | LEVEL(L2) | P(SNOOP, HITM),	/* 0x05: L2 Hit Snoop Hit Modified */
300 	OP_LH | P(LVL, UNC) | LEVEL(NA) | P(SNOOP, NONE),	/* 0x06: uncached */
301 	0,							/* 0x07: Reserved */
302 	0,							/* 0x08: Reserved */
303 	0,							/* 0x09: Reserved */
304 	0,							/* 0x0a: Reserved */
305 	0,							/* 0x0b: Reserved */
306 	0,							/* 0x0c: Reserved */
307 	0,							/* 0x0d: Reserved */
308 	0,							/* 0x0e: Reserved */
309 	0,							/* 0x0f: Reserved */
310 };
311 
312 /* L2 miss */
313 #define OMR_DATA_SOURCE_MAX		16
314 static u64 omr_data_source[OMR_DATA_SOURCE_MAX] = {
315 	P(OP, LOAD) | P(LVL, NA) | LEVEL(NA) | P(SNOOP, NA),	/* 0x00: invalid */
316 	0,							/* 0x01: Reserved */
317 	OP_LH | P(LVL, L3) | LEVEL(L3) | P(REGION, L_SHARE),	/* 0x02: local CA shared cache */
318 	OP_LH | P(LVL, L3) | LEVEL(L3) | P(REGION, L_NON_SHARE),/* 0x03: local CA non-shared cache */
319 	OP_LH | P(LVL, L3) | LEVEL(L3) | P(REGION, O_IO),	/* 0x04: other CA IO agent */
320 	OP_LH | P(LVL, L3) | LEVEL(L3) | P(REGION, O_SHARE),	/* 0x05: other CA shared cache */
321 	OP_LH | P(LVL, L3) | LEVEL(L3) | P(REGION, O_NON_SHARE),/* 0x06: other CA non-shared cache */
322 	OP_LH | LEVEL(RAM) | P(REGION, MMIO),			/* 0x07: MMIO */
323 	OP_LH | LEVEL(RAM) | P(REGION, MEM0),			/* 0x08: Memory region 0 */
324 	OP_LH | LEVEL(RAM) | P(REGION, MEM1),			/* 0x09: Memory region 1 */
325 	OP_LH | LEVEL(RAM) | P(REGION, MEM2),			/* 0x0a: Memory region 2 */
326 	OP_LH | LEVEL(RAM) | P(REGION, MEM3),			/* 0x0b: Memory region 3 */
327 	OP_LH | LEVEL(RAM) | P(REGION, MEM4),			/* 0x0c: Memory region 4 */
328 	OP_LH | LEVEL(RAM) | P(REGION, MEM5),			/* 0x0d: Memory region 5 */
329 	OP_LH | LEVEL(RAM) | P(REGION, MEM6),			/* 0x0e: Memory region 6 */
330 	OP_LH | LEVEL(RAM) | P(REGION, MEM7),			/* 0x0f: Memory region 7 */
331 };
332 
333 static u64 parse_omr_data_source(u8 dse)
334 {
335 	union omr_encoding omr;
336 	u64 val = 0;
337 
338 	omr.omr_full = dse;
339 	val = omr_data_source[omr.omr_source];
340 	if (omr.omr_source > 0x1 && omr.omr_source < 0x7)
341 		val |= omr.omr_remote ? P(LVL, REM_CCE1) : 0;
342 	else if (omr.omr_source > 0x7)
343 		val |= omr.omr_remote ? P(LVL, REM_RAM1) : P(LVL, LOC_RAM);
344 
345 	if (omr.omr_remote)
346 		val |= REM;
347 
348 	val |= omr.omr_hitm ? P(SNOOP, HITM) : P(SNOOP, HIT);
349 
350 	if (omr.omr_source == 0x2) {
351 		u8 snoop = omr.omr_snoop | omr.omr_promoted;
352 
353 		if (snoop == 0x0)
354 			val |= P(SNOOP, NA);
355 		else if (snoop == 0x1)
356 			val |= P(SNOOP, MISS);
357 		else if (snoop == 0x2)
358 			val |= P(SNOOP, HIT);
359 		else if (snoop == 0x3)
360 			val |= P(SNOOP, NONE);
361 	} else if (omr.omr_source > 0x2 && omr.omr_source < 0x7) {
362 		val |= omr.omr_snoop ? P(SNOOPX, FWD) : 0;
363 	}
364 
365 	return val;
366 }
367 
368 static u64 precise_store_data(u64 status)
369 {
370 	union intel_x86_pebs_dse dse;
371 	u64 val = P(OP, STORE) | P(SNOOP, NA) | P(LVL, L1) | P(TLB, L2);
372 
373 	dse.val = status;
374 
375 	/*
376 	 * bit 4: TLB access
377 	 * 1 = stored missed 2nd level TLB
378 	 *
379 	 * so it either hit the walker or the OS
380 	 * otherwise hit 2nd level TLB
381 	 */
382 	if (dse.st_stlb_miss)
383 		val |= P(TLB, MISS);
384 	else
385 		val |= P(TLB, HIT);
386 
387 	/*
388 	 * bit 0: hit L1 data cache
389 	 * if not set, then all we know is that
390 	 * it missed L1D
391 	 */
392 	if (dse.st_l1d_hit)
393 		val |= P(LVL, HIT);
394 	else
395 		val |= P(LVL, MISS);
396 
397 	/*
398 	 * bit 5: Locked prefix
399 	 */
400 	if (dse.st_locked)
401 		val |= P(LOCK, LOCKED);
402 
403 	return val;
404 }
405 
406 static u64 precise_datala_hsw(struct perf_event *event, u64 status)
407 {
408 	union perf_mem_data_src dse;
409 
410 	dse.val = PERF_MEM_NA;
411 
412 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
413 		dse.mem_op = PERF_MEM_OP_STORE;
414 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_LD_HSW)
415 		dse.mem_op = PERF_MEM_OP_LOAD;
416 
417 	/*
418 	 * L1 info only valid for following events:
419 	 *
420 	 * MEM_UOPS_RETIRED.STLB_MISS_STORES
421 	 * MEM_UOPS_RETIRED.LOCK_STORES
422 	 * MEM_UOPS_RETIRED.SPLIT_STORES
423 	 * MEM_UOPS_RETIRED.ALL_STORES
424 	 */
425 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW) {
426 		if (status & 1)
427 			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_HIT;
428 		else
429 			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_MISS;
430 	}
431 	return dse.val;
432 }
433 
434 static inline void pebs_set_tlb_lock(u64 *val, bool tlb, bool lock)
435 {
436 	/*
437 	 * TLB access
438 	 * 0 = did not miss 2nd level TLB
439 	 * 1 = missed 2nd level TLB
440 	 */
441 	if (tlb)
442 		*val |= P(TLB, MISS) | P(TLB, L2);
443 	else
444 		*val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
445 
446 	/* locked prefix */
447 	if (lock)
448 		*val |= P(LOCK, LOCKED);
449 }
450 
451 /* Retrieve the latency data for e-core of ADL */
452 static u64 __grt_latency_data(struct perf_event *event, u64 status,
453 			       u8 dse, bool tlb, bool lock, bool blk)
454 {
455 	u64 val;
456 
457 	WARN_ON_ONCE(is_hybrid() &&
458 		     hybrid_pmu(event->pmu)->pmu_type == hybrid_big);
459 
460 	dse &= PERF_PEBS_DATA_SOURCE_GRT_MASK;
461 	val = hybrid_var(event->pmu, pebs_data_source)[dse];
462 
463 	pebs_set_tlb_lock(&val, tlb, lock);
464 
465 	if (blk)
466 		val |= P(BLK, DATA);
467 	else
468 		val |= P(BLK, NA);
469 
470 	return val;
471 }
472 
473 u64 grt_latency_data(struct perf_event *event, u64 status)
474 {
475 	union intel_x86_pebs_dse dse;
476 
477 	dse.val = status;
478 
479 	return __grt_latency_data(event, status, dse.ld_dse,
480 				  dse.ld_locked, dse.ld_stlb_miss,
481 				  dse.ld_data_blk);
482 }
483 
484 /* Retrieve the latency data for e-core of MTL */
485 u64 cmt_latency_data(struct perf_event *event, u64 status)
486 {
487 	union intel_x86_pebs_dse dse;
488 
489 	dse.val = status;
490 
491 	return __grt_latency_data(event, status, dse.mtl_dse,
492 				  dse.mtl_stlb_miss, dse.mtl_locked,
493 				  dse.mtl_fwd_blk);
494 }
495 
496 static u64 arw_latency_data(struct perf_event *event, u64 status)
497 {
498 	union intel_x86_pebs_dse dse;
499 	union perf_mem_data_src src;
500 	u64 val;
501 
502 	dse.val = status;
503 
504 	if (!dse.arw_l2_miss)
505 		val = arw_pebs_l2_hit_data_source[dse.arw_dse & 0xf];
506 	else
507 		val = parse_omr_data_source(dse.arw_dse);
508 
509 	if (!val)
510 		val = P(OP, LOAD) | LEVEL(NA) | P(SNOOP, NA);
511 
512 	if (dse.arw_stlb_miss)
513 		val |= P(TLB, MISS) | P(TLB, L2);
514 	else
515 		val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
516 
517 	if (dse.arw_locked)
518 		val |= P(LOCK, LOCKED);
519 
520 	if (dse.arw_data_blk)
521 		val |= P(BLK, DATA);
522 	if (dse.arw_addr_blk)
523 		val |= P(BLK, ADDR);
524 	if (!dse.arw_data_blk && !dse.arw_addr_blk)
525 		val |= P(BLK, NA);
526 
527 	src.val = val;
528 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
529 		src.mem_op = P(OP, STORE);
530 
531 	return src.val;
532 }
533 
534 static u64 lnc_latency_data(struct perf_event *event, u64 status)
535 {
536 	union intel_x86_pebs_dse dse;
537 	union perf_mem_data_src src;
538 	u64 val;
539 
540 	dse.val = status;
541 
542 	/* LNC core latency data */
543 	val = hybrid_var(event->pmu, pebs_data_source)[status & PERF_PEBS_DATA_SOURCE_MASK];
544 	if (!val)
545 		val = P(OP, LOAD) | LEVEL(NA) | P(SNOOP, NA);
546 
547 	if (dse.lnc_stlb_miss)
548 		val |= P(TLB, MISS) | P(TLB, L2);
549 	else
550 		val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
551 
552 	if (dse.lnc_locked)
553 		val |= P(LOCK, LOCKED);
554 
555 	if (dse.lnc_data_blk)
556 		val |= P(BLK, DATA);
557 	if (dse.lnc_addr_blk)
558 		val |= P(BLK, ADDR);
559 	if (!dse.lnc_data_blk && !dse.lnc_addr_blk)
560 		val |= P(BLK, NA);
561 
562 	src.val = val;
563 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
564 		src.mem_op = P(OP, STORE);
565 
566 	return src.val;
567 }
568 
569 u64 lnl_latency_data(struct perf_event *event, u64 status)
570 {
571 	struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu);
572 
573 	if (pmu->pmu_type == hybrid_small)
574 		return cmt_latency_data(event, status);
575 
576 	return lnc_latency_data(event, status);
577 }
578 
579 u64 arl_h_latency_data(struct perf_event *event, u64 status)
580 {
581 	struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu);
582 
583 	if (pmu->pmu_type == hybrid_tiny)
584 		return cmt_latency_data(event, status);
585 
586 	return lnl_latency_data(event, status);
587 }
588 
589 u64 pnc_latency_data(struct perf_event *event, u64 status)
590 {
591 	union intel_x86_pebs_dse dse;
592 	union perf_mem_data_src src;
593 	u64 val;
594 
595 	dse.val = status;
596 
597 	if (!dse.pnc_l2_miss)
598 		val = pnc_pebs_l2_hit_data_source[dse.pnc_dse & 0xf];
599 	else
600 		val = parse_omr_data_source(dse.pnc_dse);
601 
602 	if (!val)
603 		val = P(OP, LOAD) | LEVEL(NA) | P(SNOOP, NA);
604 
605 	if (dse.pnc_stlb_miss)
606 		val |= P(TLB, MISS) | P(TLB, L2);
607 	else
608 		val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
609 
610 	if (dse.pnc_locked)
611 		val |= P(LOCK, LOCKED);
612 
613 	if (dse.pnc_data_blk)
614 		val |= P(BLK, DATA);
615 	if (dse.pnc_addr_blk)
616 		val |= P(BLK, ADDR);
617 	if (!dse.pnc_data_blk && !dse.pnc_addr_blk)
618 		val |= P(BLK, NA);
619 
620 	src.val = val;
621 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
622 		src.mem_op = P(OP, STORE);
623 
624 	return src.val;
625 }
626 
627 u64 nvl_latency_data(struct perf_event *event, u64 status)
628 {
629 	struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu);
630 
631 	if (pmu->pmu_type == hybrid_small)
632 		return arw_latency_data(event, status);
633 
634 	return pnc_latency_data(event, status);
635 }
636 
637 static u64 load_latency_data(struct perf_event *event, u64 status)
638 {
639 	union intel_x86_pebs_dse dse;
640 	u64 val;
641 
642 	dse.val = status;
643 
644 	/*
645 	 * use the mapping table for bit 0-3
646 	 */
647 	val = hybrid_var(event->pmu, pebs_data_source)[dse.ld_dse];
648 
649 	/*
650 	 * Nehalem models do not support TLB, Lock infos
651 	 */
652 	if (x86_pmu.pebs_no_tlb) {
653 		val |= P(TLB, NA) | P(LOCK, NA);
654 		return val;
655 	}
656 
657 	pebs_set_tlb_lock(&val, dse.ld_stlb_miss, dse.ld_locked);
658 
659 	/*
660 	 * Ice Lake and earlier models do not support block infos.
661 	 */
662 	if (!x86_pmu.pebs_block) {
663 		val |= P(BLK, NA);
664 		return val;
665 	}
666 	/*
667 	 * bit 6: load was blocked since its data could not be forwarded
668 	 *        from a preceding store
669 	 */
670 	if (dse.ld_data_blk)
671 		val |= P(BLK, DATA);
672 
673 	/*
674 	 * bit 7: load was blocked due to potential address conflict with
675 	 *        a preceding store
676 	 */
677 	if (dse.ld_addr_blk)
678 		val |= P(BLK, ADDR);
679 
680 	if (!dse.ld_data_blk && !dse.ld_addr_blk)
681 		val |= P(BLK, NA);
682 
683 	return val;
684 }
685 
686 static u64 store_latency_data(struct perf_event *event, u64 status)
687 {
688 	union intel_x86_pebs_dse dse;
689 	union perf_mem_data_src src;
690 	u64 val;
691 
692 	dse.val = status;
693 
694 	/*
695 	 * use the mapping table for bit 0-3
696 	 */
697 	val = hybrid_var(event->pmu, pebs_data_source)[dse.st_lat_dse];
698 
699 	pebs_set_tlb_lock(&val, dse.st_lat_stlb_miss, dse.st_lat_locked);
700 
701 	val |= P(BLK, NA);
702 
703 	/*
704 	 * the pebs_data_source table is only for loads
705 	 * so override the mem_op to say STORE instead
706 	 */
707 	src.val = val;
708 	src.mem_op = P(OP,STORE);
709 
710 	return src.val;
711 }
712 
713 struct pebs_record_core {
714 	u64 flags, ip;
715 	u64 ax, bx, cx, dx;
716 	u64 si, di, bp, sp;
717 	u64 r8,  r9,  r10, r11;
718 	u64 r12, r13, r14, r15;
719 };
720 
721 struct pebs_record_nhm {
722 	u64 flags, ip;
723 	u64 ax, bx, cx, dx;
724 	u64 si, di, bp, sp;
725 	u64 r8,  r9,  r10, r11;
726 	u64 r12, r13, r14, r15;
727 	u64 status, dla, dse, lat;
728 };
729 
730 /*
731  * Same as pebs_record_nhm, with two additional fields.
732  */
733 struct pebs_record_hsw {
734 	u64 flags, ip;
735 	u64 ax, bx, cx, dx;
736 	u64 si, di, bp, sp;
737 	u64 r8,  r9,  r10, r11;
738 	u64 r12, r13, r14, r15;
739 	u64 status, dla, dse, lat;
740 	u64 real_ip, tsx_tuning;
741 };
742 
743 union hsw_tsx_tuning {
744 	struct {
745 		u32 cycles_last_block     : 32,
746 		    hle_abort		  : 1,
747 		    rtm_abort		  : 1,
748 		    instruction_abort     : 1,
749 		    non_instruction_abort : 1,
750 		    retry		  : 1,
751 		    data_conflict	  : 1,
752 		    capacity_writes	  : 1,
753 		    capacity_reads	  : 1;
754 	};
755 	u64	    value;
756 };
757 
758 #define PEBS_HSW_TSX_FLAGS	0xff00000000ULL
759 
760 /* Same as HSW, plus TSC */
761 
762 struct pebs_record_skl {
763 	u64 flags, ip;
764 	u64 ax, bx, cx, dx;
765 	u64 si, di, bp, sp;
766 	u64 r8,  r9,  r10, r11;
767 	u64 r12, r13, r14, r15;
768 	u64 status, dla, dse, lat;
769 	u64 real_ip, tsx_tuning;
770 	u64 tsc;
771 };
772 
773 void init_debug_store_on_cpu(int cpu)
774 {
775 	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
776 
777 	if (!ds)
778 		return;
779 
780 	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
781 		     (u32)((u64)(unsigned long)ds),
782 		     (u32)((u64)(unsigned long)ds >> 32));
783 }
784 
785 void fini_debug_store_on_cpu(int cpu)
786 {
787 	if (!per_cpu(cpu_hw_events, cpu).ds)
788 		return;
789 
790 	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
791 }
792 
793 static DEFINE_PER_CPU(void *, insn_buffer);
794 
795 static void ds_update_cea(void *cea, void *addr, size_t size, pgprot_t prot)
796 {
797 	unsigned long start = (unsigned long)cea;
798 	phys_addr_t pa;
799 	size_t msz = 0;
800 
801 	pa = virt_to_phys(addr);
802 
803 	preempt_disable();
804 	for (; msz < size; msz += PAGE_SIZE, pa += PAGE_SIZE, cea += PAGE_SIZE)
805 		cea_set_pte(cea, pa, prot);
806 
807 	/*
808 	 * This is a cross-CPU update of the cpu_entry_area, we must shoot down
809 	 * all TLB entries for it.
810 	 */
811 	flush_tlb_kernel_range(start, start + size);
812 	preempt_enable();
813 }
814 
815 static void ds_clear_cea(void *cea, size_t size)
816 {
817 	unsigned long start = (unsigned long)cea;
818 	size_t msz = 0;
819 
820 	preempt_disable();
821 	for (; msz < size; msz += PAGE_SIZE, cea += PAGE_SIZE)
822 		cea_set_pte(cea, 0, PAGE_NONE);
823 
824 	flush_tlb_kernel_range(start, start + size);
825 	preempt_enable();
826 }
827 
828 static void *dsalloc_pages(size_t size, gfp_t flags, int cpu)
829 {
830 	unsigned int order = get_order(size);
831 	int node = cpu_to_node(cpu);
832 	struct page *page;
833 
834 	page = __alloc_pages_node(node, flags | __GFP_ZERO, order);
835 	return page ? page_address(page) : NULL;
836 }
837 
838 static void dsfree_pages(const void *buffer, size_t size)
839 {
840 	if (buffer)
841 		free_pages((unsigned long)buffer, get_order(size));
842 }
843 
844 static int alloc_pebs_buffer(int cpu)
845 {
846 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
847 	struct debug_store *ds = hwev->ds;
848 	size_t bsiz = x86_pmu.pebs_buffer_size;
849 	int max, node = cpu_to_node(cpu);
850 	void *buffer, *insn_buff, *cea;
851 
852 	if (!intel_pmu_has_pebs())
853 		return 0;
854 
855 	buffer = dsalloc_pages(bsiz, GFP_KERNEL, cpu);
856 	if (unlikely(!buffer))
857 		return -ENOMEM;
858 
859 	if (x86_pmu.arch_pebs) {
860 		hwev->pebs_vaddr = buffer;
861 		return 0;
862 	}
863 
864 	/*
865 	 * HSW+ already provides us the eventing ip; no need to allocate this
866 	 * buffer then.
867 	 */
868 	if (x86_pmu.intel_cap.pebs_format < 2) {
869 		insn_buff = kzalloc_node(PEBS_FIXUP_SIZE, GFP_KERNEL, node);
870 		if (!insn_buff) {
871 			dsfree_pages(buffer, bsiz);
872 			return -ENOMEM;
873 		}
874 		per_cpu(insn_buffer, cpu) = insn_buff;
875 	}
876 	hwev->pebs_vaddr = buffer;
877 	/* Update the cpu entry area mapping */
878 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
879 	ds->pebs_buffer_base = (unsigned long) cea;
880 	ds_update_cea(cea, buffer, bsiz, PAGE_KERNEL);
881 	ds->pebs_index = ds->pebs_buffer_base;
882 	max = x86_pmu.pebs_record_size * (bsiz / x86_pmu.pebs_record_size);
883 	ds->pebs_absolute_maximum = ds->pebs_buffer_base + max;
884 	return 0;
885 }
886 
887 static void release_pebs_buffer(int cpu)
888 {
889 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
890 	void *cea;
891 
892 	if (!intel_pmu_has_pebs())
893 		return;
894 
895 	if (x86_pmu.ds_pebs) {
896 		kfree(per_cpu(insn_buffer, cpu));
897 		per_cpu(insn_buffer, cpu) = NULL;
898 
899 		/* Clear the fixmap */
900 		cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
901 		ds_clear_cea(cea, x86_pmu.pebs_buffer_size);
902 	}
903 
904 	dsfree_pages(hwev->pebs_vaddr, x86_pmu.pebs_buffer_size);
905 	hwev->pebs_vaddr = NULL;
906 }
907 
908 static int alloc_bts_buffer(int cpu)
909 {
910 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
911 	struct debug_store *ds = hwev->ds;
912 	void *buffer, *cea;
913 	int max;
914 
915 	if (!x86_pmu.bts)
916 		return 0;
917 
918 	buffer = dsalloc_pages(BTS_BUFFER_SIZE, GFP_KERNEL | __GFP_NOWARN, cpu);
919 	if (unlikely(!buffer)) {
920 		WARN_ONCE(1, "%s: BTS buffer allocation failure\n", __func__);
921 		return -ENOMEM;
922 	}
923 	hwev->ds_bts_vaddr = buffer;
924 	/* Update the fixmap */
925 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
926 	ds->bts_buffer_base = (unsigned long) cea;
927 	ds_update_cea(cea, buffer, BTS_BUFFER_SIZE, PAGE_KERNEL);
928 	ds->bts_index = ds->bts_buffer_base;
929 	max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
930 	ds->bts_absolute_maximum = ds->bts_buffer_base +
931 					max * BTS_RECORD_SIZE;
932 	ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
933 					(max / 16) * BTS_RECORD_SIZE;
934 	return 0;
935 }
936 
937 static void release_bts_buffer(int cpu)
938 {
939 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
940 	void *cea;
941 
942 	if (!x86_pmu.bts)
943 		return;
944 
945 	/* Clear the fixmap */
946 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
947 	ds_clear_cea(cea, BTS_BUFFER_SIZE);
948 	dsfree_pages(hwev->ds_bts_vaddr, BTS_BUFFER_SIZE);
949 	hwev->ds_bts_vaddr = NULL;
950 }
951 
952 static int alloc_ds_buffer(int cpu)
953 {
954 	struct debug_store *ds = &get_cpu_entry_area(cpu)->cpu_debug_store;
955 
956 	memset(ds, 0, sizeof(*ds));
957 	per_cpu(cpu_hw_events, cpu).ds = ds;
958 	return 0;
959 }
960 
961 static void release_ds_buffer(int cpu)
962 {
963 	per_cpu(cpu_hw_events, cpu).ds = NULL;
964 }
965 
966 void release_ds_buffers(void)
967 {
968 	int cpu;
969 
970 	if (!x86_pmu.bts && !x86_pmu.ds_pebs)
971 		return;
972 
973 	for_each_possible_cpu(cpu)
974 		release_ds_buffer(cpu);
975 
976 	for_each_possible_cpu(cpu) {
977 		/*
978 		 * Again, ignore errors from offline CPUs, they will no longer
979 		 * observe cpu_hw_events.ds and not program the DS_AREA when
980 		 * they come up.
981 		 */
982 		fini_debug_store_on_cpu(cpu);
983 	}
984 
985 	for_each_possible_cpu(cpu) {
986 		if (x86_pmu.ds_pebs)
987 			release_pebs_buffer(cpu);
988 		release_bts_buffer(cpu);
989 	}
990 }
991 
992 void reserve_ds_buffers(void)
993 {
994 	int bts_err = 0, pebs_err = 0;
995 	int cpu;
996 
997 	x86_pmu.bts_active = 0;
998 
999 	if (x86_pmu.ds_pebs)
1000 		x86_pmu.pebs_active = 0;
1001 
1002 	if (!x86_pmu.bts && !x86_pmu.ds_pebs)
1003 		return;
1004 
1005 	if (!x86_pmu.bts)
1006 		bts_err = 1;
1007 
1008 	if (!x86_pmu.ds_pebs)
1009 		pebs_err = 1;
1010 
1011 	for_each_possible_cpu(cpu) {
1012 		if (alloc_ds_buffer(cpu)) {
1013 			bts_err = 1;
1014 			pebs_err = 1;
1015 		}
1016 
1017 		if (!bts_err && alloc_bts_buffer(cpu))
1018 			bts_err = 1;
1019 
1020 		if (x86_pmu.ds_pebs && !pebs_err &&
1021 		    alloc_pebs_buffer(cpu))
1022 			pebs_err = 1;
1023 
1024 		if (bts_err && pebs_err)
1025 			break;
1026 	}
1027 
1028 	if (bts_err) {
1029 		for_each_possible_cpu(cpu)
1030 			release_bts_buffer(cpu);
1031 	}
1032 
1033 	if (x86_pmu.ds_pebs && pebs_err) {
1034 		for_each_possible_cpu(cpu)
1035 			release_pebs_buffer(cpu);
1036 	}
1037 
1038 	if (bts_err && pebs_err) {
1039 		for_each_possible_cpu(cpu)
1040 			release_ds_buffer(cpu);
1041 	} else {
1042 		if (x86_pmu.bts && !bts_err)
1043 			x86_pmu.bts_active = 1;
1044 
1045 		if (x86_pmu.ds_pebs && !pebs_err)
1046 			x86_pmu.pebs_active = 1;
1047 
1048 		for_each_possible_cpu(cpu) {
1049 			/*
1050 			 * Ignores wrmsr_on_cpu() errors for offline CPUs they
1051 			 * will get this call through intel_pmu_cpu_starting().
1052 			 */
1053 			init_debug_store_on_cpu(cpu);
1054 		}
1055 	}
1056 }
1057 
1058 inline int alloc_arch_pebs_buf_on_cpu(int cpu)
1059 {
1060 	if (!x86_pmu.arch_pebs)
1061 		return 0;
1062 
1063 	return alloc_pebs_buffer(cpu);
1064 }
1065 
1066 inline void release_arch_pebs_buf_on_cpu(int cpu)
1067 {
1068 	if (!x86_pmu.arch_pebs)
1069 		return;
1070 
1071 	release_pebs_buffer(cpu);
1072 }
1073 
1074 void init_arch_pebs_on_cpu(int cpu)
1075 {
1076 	struct cpu_hw_events *cpuc = per_cpu_ptr(&cpu_hw_events, cpu);
1077 	u64 arch_pebs_base;
1078 
1079 	if (!x86_pmu.arch_pebs)
1080 		return;
1081 
1082 	if (!cpuc->pebs_vaddr) {
1083 		WARN(1, "Fail to allocate PEBS buffer on CPU %d\n", cpu);
1084 		x86_pmu.pebs_active = 0;
1085 		return;
1086 	}
1087 
1088 	/*
1089 	 * 4KB-aligned pointer of the output buffer
1090 	 * (__alloc_pages_node() return page aligned address)
1091 	 * Buffer Size = 4KB * 2^SIZE
1092 	 * contiguous physical buffer (__alloc_pages_node() with order)
1093 	 */
1094 	arch_pebs_base = virt_to_phys(cpuc->pebs_vaddr) | PEBS_BUFFER_SHIFT;
1095 	wrmsr_on_cpu(cpu, MSR_IA32_PEBS_BASE, (u32)arch_pebs_base,
1096 		     (u32)(arch_pebs_base >> 32));
1097 	x86_pmu.pebs_active = 1;
1098 }
1099 
1100 inline void fini_arch_pebs_on_cpu(int cpu)
1101 {
1102 	if (!x86_pmu.arch_pebs)
1103 		return;
1104 
1105 	wrmsr_on_cpu(cpu, MSR_IA32_PEBS_BASE, 0, 0);
1106 }
1107 
1108 /*
1109  * BTS
1110  */
1111 
1112 struct event_constraint bts_constraint =
1113 	EVENT_CONSTRAINT(0, 1ULL << INTEL_PMC_IDX_FIXED_BTS, 0);
1114 
1115 void intel_pmu_enable_bts(u64 config)
1116 {
1117 	unsigned long debugctlmsr;
1118 
1119 	debugctlmsr = get_debugctlmsr();
1120 
1121 	debugctlmsr |= DEBUGCTLMSR_TR;
1122 	debugctlmsr |= DEBUGCTLMSR_BTS;
1123 	if (config & ARCH_PERFMON_EVENTSEL_INT)
1124 		debugctlmsr |= DEBUGCTLMSR_BTINT;
1125 
1126 	if (!(config & ARCH_PERFMON_EVENTSEL_OS))
1127 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS;
1128 
1129 	if (!(config & ARCH_PERFMON_EVENTSEL_USR))
1130 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR;
1131 
1132 	update_debugctlmsr(debugctlmsr);
1133 }
1134 
1135 void intel_pmu_disable_bts(void)
1136 {
1137 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1138 	unsigned long debugctlmsr;
1139 
1140 	if (!cpuc->ds)
1141 		return;
1142 
1143 	debugctlmsr = get_debugctlmsr();
1144 
1145 	debugctlmsr &=
1146 		~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT |
1147 		  DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR);
1148 
1149 	update_debugctlmsr(debugctlmsr);
1150 }
1151 
1152 int intel_pmu_drain_bts_buffer(void)
1153 {
1154 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1155 	struct debug_store *ds = cpuc->ds;
1156 	struct bts_record {
1157 		u64	from;
1158 		u64	to;
1159 		u64	flags;
1160 	};
1161 	struct perf_event *event = cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
1162 	struct bts_record *at, *base, *top;
1163 	struct perf_output_handle handle;
1164 	struct perf_event_header header;
1165 	struct perf_sample_data data;
1166 	unsigned long skip = 0;
1167 	struct pt_regs regs;
1168 
1169 	if (!event)
1170 		return 0;
1171 
1172 	if (!x86_pmu.bts_active)
1173 		return 0;
1174 
1175 	base = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
1176 	top  = (struct bts_record *)(unsigned long)ds->bts_index;
1177 
1178 	if (top <= base)
1179 		return 0;
1180 
1181 	memset(&regs, 0, sizeof(regs));
1182 
1183 	ds->bts_index = ds->bts_buffer_base;
1184 
1185 	perf_sample_data_init(&data, 0, event->hw.last_period);
1186 
1187 	/*
1188 	 * BTS leaks kernel addresses in branches across the cpl boundary,
1189 	 * such as traps or system calls, so unless the user is asking for
1190 	 * kernel tracing (and right now it's not possible), we'd need to
1191 	 * filter them out. But first we need to count how many of those we
1192 	 * have in the current batch. This is an extra O(n) pass, however,
1193 	 * it's much faster than the other one especially considering that
1194 	 * n <= 2560 (BTS_BUFFER_SIZE / BTS_RECORD_SIZE * 15/16; see the
1195 	 * alloc_bts_buffer()).
1196 	 */
1197 	for (at = base; at < top; at++) {
1198 		/*
1199 		 * Note that right now *this* BTS code only works if
1200 		 * attr::exclude_kernel is set, but let's keep this extra
1201 		 * check here in case that changes.
1202 		 */
1203 		if (event->attr.exclude_kernel &&
1204 		    (kernel_ip(at->from) || kernel_ip(at->to)))
1205 			skip++;
1206 	}
1207 
1208 	/*
1209 	 * Prepare a generic sample, i.e. fill in the invariant fields.
1210 	 * We will overwrite the from and to address before we output
1211 	 * the sample.
1212 	 */
1213 	rcu_read_lock();
1214 	perf_prepare_sample(&data, event, &regs);
1215 	perf_prepare_header(&header, &data, event, &regs);
1216 
1217 	if (perf_output_begin(&handle, &data, event,
1218 			      header.size * (top - base - skip)))
1219 		goto unlock;
1220 
1221 	for (at = base; at < top; at++) {
1222 		/* Filter out any records that contain kernel addresses. */
1223 		if (event->attr.exclude_kernel &&
1224 		    (kernel_ip(at->from) || kernel_ip(at->to)))
1225 			continue;
1226 
1227 		data.ip		= at->from;
1228 		data.addr	= at->to;
1229 
1230 		perf_output_sample(&handle, &header, &data, event);
1231 	}
1232 
1233 	perf_output_end(&handle);
1234 
1235 	/* There's new data available. */
1236 	event->hw.interrupts++;
1237 	event->pending_kill = POLL_IN;
1238 unlock:
1239 	rcu_read_unlock();
1240 	return 1;
1241 }
1242 
1243 void intel_pmu_drain_pebs_buffer(void)
1244 {
1245 	struct perf_sample_data data;
1246 
1247 	static_call(x86_pmu_drain_pebs)(NULL, &data);
1248 }
1249 
1250 /*
1251  * PEBS
1252  */
1253 struct event_constraint intel_core2_pebs_event_constraints[] = {
1254 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
1255 	INTEL_FLAGS_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
1256 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
1257 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
1258 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
1259 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
1260 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01),
1261 	EVENT_CONSTRAINT_END
1262 };
1263 
1264 struct event_constraint intel_atom_pebs_event_constraints[] = {
1265 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
1266 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
1267 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
1268 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
1269 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01),
1270 	/* Allow all events as PEBS with no flags */
1271 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
1272 	EVENT_CONSTRAINT_END
1273 };
1274 
1275 struct event_constraint intel_slm_pebs_event_constraints[] = {
1276 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
1277 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x1),
1278 	/* Allow all events as PEBS with no flags */
1279 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
1280 	EVENT_CONSTRAINT_END
1281 };
1282 
1283 struct event_constraint intel_glm_pebs_event_constraints[] = {
1284 	/* Allow all events as PEBS with no flags */
1285 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
1286 	EVENT_CONSTRAINT_END
1287 };
1288 
1289 struct event_constraint intel_grt_pebs_event_constraints[] = {
1290 	/* Allow all events as PEBS with no flags */
1291 	INTEL_HYBRID_LAT_CONSTRAINT(0x5d0, 0x3),
1292 	INTEL_HYBRID_LAT_CONSTRAINT(0x6d0, 0xf),
1293 	EVENT_CONSTRAINT_END
1294 };
1295 
1296 struct event_constraint intel_arw_pebs_event_constraints[] = {
1297 	/* Allow all events as PEBS with no flags */
1298 	INTEL_HYBRID_LAT_CONSTRAINT(0x5d0, 0xff),
1299 	INTEL_HYBRID_LAT_CONSTRAINT(0x6d0, 0xff),
1300 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01d4, 0x1),
1301 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x02d4, 0x2),
1302 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x04d4, 0x4),
1303 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x08d4, 0x8),
1304 	EVENT_CONSTRAINT_END
1305 };
1306 
1307 struct event_constraint intel_nehalem_pebs_event_constraints[] = {
1308 	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
1309 	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
1310 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
1311 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INST_RETIRED.ANY */
1312 	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
1313 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
1314 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
1315 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
1316 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
1317 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
1318 	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
1319 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
1320 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
1321 	EVENT_CONSTRAINT_END
1322 };
1323 
1324 struct event_constraint intel_westmere_pebs_event_constraints[] = {
1325 	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
1326 	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
1327 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
1328 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INSTR_RETIRED.* */
1329 	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
1330 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
1331 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc5, 0xf),    /* BR_MISP_RETIRED.* */
1332 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
1333 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
1334 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
1335 	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
1336 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
1337 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
1338 	EVENT_CONSTRAINT_END
1339 };
1340 
1341 struct event_constraint intel_snb_pebs_event_constraints[] = {
1342 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
1343 	INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
1344 	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
1345 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
1346 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
1347         INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
1348         INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
1349         INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
1350         INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
1351 	/* Allow all events as PEBS with no flags */
1352 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
1353 	EVENT_CONSTRAINT_END
1354 };
1355 
1356 struct event_constraint intel_ivb_pebs_event_constraints[] = {
1357         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
1358         INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
1359 	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
1360 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
1361 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
1362 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
1363 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
1364 	INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
1365 	INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
1366 	INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
1367 	INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
1368 	/* Allow all events as PEBS with no flags */
1369 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
1370         EVENT_CONSTRAINT_END
1371 };
1372 
1373 struct event_constraint intel_hsw_pebs_event_constraints[] = {
1374 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
1375 	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
1376 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
1377 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
1378 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
1379 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
1380 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
1381 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
1382 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
1383 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
1384 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
1385 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
1386 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
1387 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
1388 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
1389 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
1390 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
1391 	/* Allow all events as PEBS with no flags */
1392 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
1393 	EVENT_CONSTRAINT_END
1394 };
1395 
1396 struct event_constraint intel_bdw_pebs_event_constraints[] = {
1397 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
1398 	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
1399 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
1400 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
1401 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
1402 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
1403 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
1404 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
1405 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
1406 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
1407 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
1408 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
1409 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
1410 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
1411 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
1412 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
1413 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
1414 	/* Allow all events as PEBS with no flags */
1415 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
1416 	EVENT_CONSTRAINT_END
1417 };
1418 
1419 
1420 struct event_constraint intel_skl_pebs_event_constraints[] = {
1421 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1c0, 0x2),	/* INST_RETIRED.PREC_DIST */
1422 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
1423 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
1424 	/* INST_RETIRED.TOTAL_CYCLES_PS (inv=1, cmask=16) (cycles:p). */
1425 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
1426 	INTEL_PLD_CONSTRAINT(0x1cd, 0xf),		      /* MEM_TRANS_RETIRED.* */
1427 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */
1428 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */
1429 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */
1430 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x22d0, 0xf), /* MEM_INST_RETIRED.LOCK_STORES */
1431 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */
1432 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */
1433 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */
1434 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */
1435 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_RETIRED.* */
1436 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_L3_HIT_RETIRED.* */
1437 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_L3_MISS_RETIRED.* */
1438 	/* Allow all events as PEBS with no flags */
1439 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
1440 	EVENT_CONSTRAINT_END
1441 };
1442 
1443 struct event_constraint intel_icl_pebs_event_constraints[] = {
1444 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x100000000ULL),	/* old INST_RETIRED.PREC_DIST */
1445 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x0100, 0x100000000ULL),	/* INST_RETIRED.PREC_DIST */
1446 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x0400, 0x800000000ULL),	/* SLOTS */
1447 
1448 	INTEL_PLD_CONSTRAINT(0x1cd, 0xff),			/* MEM_TRANS_RETIRED.LOAD_LATENCY */
1449 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf),	/* MEM_INST_RETIRED.STLB_MISS_LOADS */
1450 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf),	/* MEM_INST_RETIRED.STLB_MISS_STORES */
1451 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf),	/* MEM_INST_RETIRED.LOCK_LOADS */
1452 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf),	/* MEM_INST_RETIRED.SPLIT_LOADS */
1453 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf),	/* MEM_INST_RETIRED.SPLIT_STORES */
1454 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf),	/* MEM_INST_RETIRED.ALL_LOADS */
1455 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf),	/* MEM_INST_RETIRED.ALL_STORES */
1456 
1457 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(0xd1, 0xd4, 0xf), /* MEM_LOAD_*_RETIRED.* */
1458 
1459 	INTEL_FLAGS_EVENT_CONSTRAINT(0xd0, 0xf),		/* MEM_INST_RETIRED.* */
1460 
1461 	/*
1462 	 * Everything else is handled by PMU_FL_PEBS_ALL, because we
1463 	 * need the full constraints from the main table.
1464 	 */
1465 
1466 	EVENT_CONSTRAINT_END
1467 };
1468 
1469 struct event_constraint intel_glc_pebs_event_constraints[] = {
1470 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x100, 0x100000000ULL),	/* INST_RETIRED.PREC_DIST */
1471 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x0400, 0x800000000ULL),
1472 
1473 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xfe),
1474 	INTEL_PLD_CONSTRAINT(0x1cd, 0xfe),
1475 	INTEL_PSD_CONSTRAINT(0x2cd, 0x1),
1476 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf),	/* MEM_INST_RETIRED.STLB_MISS_LOADS */
1477 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf),	/* MEM_INST_RETIRED.STLB_MISS_STORES */
1478 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf),	/* MEM_INST_RETIRED.LOCK_LOADS */
1479 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf),	/* MEM_INST_RETIRED.SPLIT_LOADS */
1480 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf),	/* MEM_INST_RETIRED.SPLIT_STORES */
1481 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf),	/* MEM_INST_RETIRED.ALL_LOADS */
1482 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf),	/* MEM_INST_RETIRED.ALL_STORES */
1483 
1484 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(0xd1, 0xd4, 0xf),
1485 
1486 	INTEL_FLAGS_EVENT_CONSTRAINT(0xd0, 0xf),
1487 
1488 	/*
1489 	 * Everything else is handled by PMU_FL_PEBS_ALL, because we
1490 	 * need the full constraints from the main table.
1491 	 */
1492 
1493 	EVENT_CONSTRAINT_END
1494 };
1495 
1496 struct event_constraint intel_lnc_pebs_event_constraints[] = {
1497 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x100, 0x100000000ULL),	/* INST_RETIRED.PREC_DIST */
1498 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x0400, 0x800000000ULL),
1499 
1500 	INTEL_HYBRID_LDLAT_CONSTRAINT(0x1cd, 0x3fc),
1501 	INTEL_HYBRID_STLAT_CONSTRAINT(0x2cd, 0x3),
1502 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf),	/* MEM_INST_RETIRED.STLB_MISS_LOADS */
1503 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf),	/* MEM_INST_RETIRED.STLB_MISS_STORES */
1504 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf),	/* MEM_INST_RETIRED.LOCK_LOADS */
1505 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf),	/* MEM_INST_RETIRED.SPLIT_LOADS */
1506 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf),	/* MEM_INST_RETIRED.SPLIT_STORES */
1507 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf),	/* MEM_INST_RETIRED.ALL_LOADS */
1508 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf),	/* MEM_INST_RETIRED.ALL_STORES */
1509 
1510 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(0xd1, 0xd4, 0xf),
1511 
1512 	INTEL_FLAGS_EVENT_CONSTRAINT(0xd0, 0xf),
1513 
1514 	/*
1515 	 * Everything else is handled by PMU_FL_PEBS_ALL, because we
1516 	 * need the full constraints from the main table.
1517 	 */
1518 
1519 	EVENT_CONSTRAINT_END
1520 };
1521 
1522 struct event_constraint intel_pnc_pebs_event_constraints[] = {
1523 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x100, 0x100000000ULL),	/* INST_RETIRED.PREC_DIST */
1524 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x0400, 0x800000000ULL),
1525 
1526 	INTEL_HYBRID_LDLAT_CONSTRAINT(0x1cd, 0xfc),
1527 	INTEL_HYBRID_STLAT_CONSTRAINT(0x2cd, 0x3),
1528 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf),	/* MEM_INST_RETIRED.STLB_MISS_LOADS */
1529 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf),	/* MEM_INST_RETIRED.STLB_MISS_STORES */
1530 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf),	/* MEM_INST_RETIRED.LOCK_LOADS */
1531 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf),	/* MEM_INST_RETIRED.SPLIT_LOADS */
1532 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf),	/* MEM_INST_RETIRED.SPLIT_STORES */
1533 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf),	/* MEM_INST_RETIRED.ALL_LOADS */
1534 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf),	/* MEM_INST_RETIRED.ALL_STORES */
1535 
1536 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(0xd1, 0xd4, 0xf),
1537 
1538 	INTEL_FLAGS_EVENT_CONSTRAINT(0xd0, 0xf),
1539 	INTEL_FLAGS_EVENT_CONSTRAINT(0xd6, 0xf),
1540 
1541 	/*
1542 	 * Everything else is handled by PMU_FL_PEBS_ALL, because we
1543 	 * need the full constraints from the main table.
1544 	 */
1545 
1546 	EVENT_CONSTRAINT_END
1547 };
1548 
1549 struct event_constraint *intel_pebs_constraints(struct perf_event *event)
1550 {
1551 	struct event_constraint *pebs_constraints = hybrid(event->pmu, pebs_constraints);
1552 	struct event_constraint *c;
1553 
1554 	if (!event->attr.precise_ip)
1555 		return NULL;
1556 
1557 	if (pebs_constraints) {
1558 		for_each_event_constraint(c, pebs_constraints) {
1559 			if (constraint_match(c, event->hw.config)) {
1560 				event->hw.flags |= c->flags;
1561 				return c;
1562 			}
1563 		}
1564 	}
1565 
1566 	/*
1567 	 * Extended PEBS support
1568 	 * Makes the PEBS code search the normal constraints.
1569 	 */
1570 	if (x86_pmu.flags & PMU_FL_PEBS_ALL)
1571 		return NULL;
1572 
1573 	return &emptyconstraint;
1574 }
1575 
1576 /*
1577  * We need the sched_task callback even for per-cpu events when we use
1578  * the large interrupt threshold, such that we can provide PID and TID
1579  * to PEBS samples.
1580  */
1581 static inline bool pebs_needs_sched_cb(struct cpu_hw_events *cpuc)
1582 {
1583 	if (cpuc->n_pebs == cpuc->n_pebs_via_pt)
1584 		return false;
1585 
1586 	return cpuc->n_pebs && (cpuc->n_pebs == cpuc->n_large_pebs);
1587 }
1588 
1589 void intel_pmu_pebs_sched_task(struct perf_event_pmu_context *pmu_ctx, bool sched_in)
1590 {
1591 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1592 
1593 	if (!sched_in && pebs_needs_sched_cb(cpuc))
1594 		intel_pmu_drain_pebs_buffer();
1595 }
1596 
1597 static inline void pebs_update_threshold(struct cpu_hw_events *cpuc)
1598 {
1599 	struct debug_store *ds = cpuc->ds;
1600 	int max_pebs_events = intel_pmu_max_num_pebs(cpuc->pmu);
1601 	u64 threshold;
1602 	int reserved;
1603 
1604 	if (cpuc->n_pebs_via_pt)
1605 		return;
1606 
1607 	if (x86_pmu.flags & PMU_FL_PEBS_ALL)
1608 		reserved = max_pebs_events + x86_pmu_max_num_counters_fixed(cpuc->pmu);
1609 	else
1610 		reserved = max_pebs_events;
1611 
1612 	if (cpuc->n_pebs == cpuc->n_large_pebs) {
1613 		threshold = ds->pebs_absolute_maximum -
1614 			reserved * cpuc->pebs_record_size;
1615 	} else {
1616 		threshold = ds->pebs_buffer_base + cpuc->pebs_record_size;
1617 	}
1618 
1619 	ds->pebs_interrupt_threshold = threshold;
1620 }
1621 
1622 #define PEBS_DATACFG_CNTRS(x)						\
1623 	((x >> PEBS_DATACFG_CNTR_SHIFT) & PEBS_DATACFG_CNTR_MASK)
1624 
1625 #define PEBS_DATACFG_CNTR_BIT(x)					\
1626 	(((1ULL << x) & PEBS_DATACFG_CNTR_MASK) << PEBS_DATACFG_CNTR_SHIFT)
1627 
1628 #define PEBS_DATACFG_FIX(x)						\
1629 	((x >> PEBS_DATACFG_FIX_SHIFT) & PEBS_DATACFG_FIX_MASK)
1630 
1631 #define PEBS_DATACFG_FIX_BIT(x)						\
1632 	(((1ULL << (x)) & PEBS_DATACFG_FIX_MASK)			\
1633 	 << PEBS_DATACFG_FIX_SHIFT)
1634 
1635 static void adaptive_pebs_record_size_update(void)
1636 {
1637 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1638 	u64 pebs_data_cfg = cpuc->pebs_data_cfg;
1639 	int sz = sizeof(struct pebs_basic);
1640 
1641 	if (pebs_data_cfg & PEBS_DATACFG_MEMINFO)
1642 		sz += sizeof(struct pebs_meminfo);
1643 	if (pebs_data_cfg & PEBS_DATACFG_GP)
1644 		sz += sizeof(struct pebs_gprs);
1645 	if (pebs_data_cfg & PEBS_DATACFG_XMMS)
1646 		sz += sizeof(struct pebs_xmm);
1647 	if (pebs_data_cfg & PEBS_DATACFG_LBRS)
1648 		sz += x86_pmu.lbr_nr * sizeof(struct lbr_entry);
1649 	if (pebs_data_cfg & (PEBS_DATACFG_METRICS | PEBS_DATACFG_CNTR)) {
1650 		sz += sizeof(struct pebs_cntr_header);
1651 
1652 		/* Metrics base and Metrics Data */
1653 		if (pebs_data_cfg & PEBS_DATACFG_METRICS)
1654 			sz += 2 * sizeof(u64);
1655 
1656 		if (pebs_data_cfg & PEBS_DATACFG_CNTR) {
1657 			sz += (hweight64(PEBS_DATACFG_CNTRS(pebs_data_cfg)) +
1658 			       hweight64(PEBS_DATACFG_FIX(pebs_data_cfg))) *
1659 			      sizeof(u64);
1660 		}
1661 	}
1662 
1663 	cpuc->pebs_record_size = sz;
1664 }
1665 
1666 static void __intel_pmu_pebs_update_cfg(struct perf_event *event,
1667 					int idx, u64 *pebs_data_cfg)
1668 {
1669 	if (is_metric_event(event)) {
1670 		*pebs_data_cfg |= PEBS_DATACFG_METRICS;
1671 		return;
1672 	}
1673 
1674 	*pebs_data_cfg |= PEBS_DATACFG_CNTR;
1675 
1676 	if (idx >= INTEL_PMC_IDX_FIXED)
1677 		*pebs_data_cfg |= PEBS_DATACFG_FIX_BIT(idx - INTEL_PMC_IDX_FIXED);
1678 	else
1679 		*pebs_data_cfg |= PEBS_DATACFG_CNTR_BIT(idx);
1680 }
1681 
1682 
1683 void intel_pmu_pebs_late_setup(struct cpu_hw_events *cpuc)
1684 {
1685 	struct perf_event *event;
1686 	u64 pebs_data_cfg = 0;
1687 	int i;
1688 
1689 	for (i = 0; i < cpuc->n_events; i++) {
1690 		event = cpuc->event_list[i];
1691 		if (!is_pebs_counter_event_group(event))
1692 			continue;
1693 		__intel_pmu_pebs_update_cfg(event, cpuc->assign[i], &pebs_data_cfg);
1694 	}
1695 
1696 	if (pebs_data_cfg & ~cpuc->pebs_data_cfg)
1697 		cpuc->pebs_data_cfg |= pebs_data_cfg | PEBS_UPDATE_DS_SW;
1698 }
1699 
1700 #define PERF_PEBS_MEMINFO_TYPE	(PERF_SAMPLE_ADDR | PERF_SAMPLE_DATA_SRC |   \
1701 				PERF_SAMPLE_PHYS_ADDR |			     \
1702 				PERF_SAMPLE_WEIGHT_TYPE |		     \
1703 				PERF_SAMPLE_TRANSACTION |		     \
1704 				PERF_SAMPLE_DATA_PAGE_SIZE)
1705 
1706 static u64 pebs_update_adaptive_cfg(struct perf_event *event)
1707 {
1708 	struct perf_event_attr *attr = &event->attr;
1709 	u64 sample_type = attr->sample_type;
1710 	u64 pebs_data_cfg = 0;
1711 	bool gprs, tsx_weight;
1712 
1713 	if (!(sample_type & ~(PERF_SAMPLE_IP|PERF_SAMPLE_TIME)) &&
1714 	    attr->precise_ip > 1)
1715 		return pebs_data_cfg;
1716 
1717 	if (sample_type & PERF_PEBS_MEMINFO_TYPE)
1718 		pebs_data_cfg |= PEBS_DATACFG_MEMINFO;
1719 
1720 	/*
1721 	 * We need GPRs when:
1722 	 * + user requested them
1723 	 * + precise_ip < 2 for the non event IP
1724 	 * + For RTM TSX weight we need GPRs for the abort code.
1725 	 */
1726 	gprs = ((sample_type & PERF_SAMPLE_REGS_INTR) &&
1727 		(attr->sample_regs_intr & PEBS_GP_REGS)) ||
1728 	       ((sample_type & PERF_SAMPLE_REGS_USER) &&
1729 		(attr->sample_regs_user & PEBS_GP_REGS));
1730 
1731 	tsx_weight = (sample_type & PERF_SAMPLE_WEIGHT_TYPE) &&
1732 		     ((attr->config & INTEL_ARCH_EVENT_MASK) ==
1733 		      x86_pmu.rtm_abort_event);
1734 
1735 	if (gprs || (attr->precise_ip < 2) || tsx_weight)
1736 		pebs_data_cfg |= PEBS_DATACFG_GP;
1737 
1738 	if ((sample_type & PERF_SAMPLE_REGS_INTR) &&
1739 	    (attr->sample_regs_intr & PERF_REG_EXTENDED_MASK))
1740 		pebs_data_cfg |= PEBS_DATACFG_XMMS;
1741 
1742 	if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
1743 		/*
1744 		 * For now always log all LBRs. Could configure this
1745 		 * later.
1746 		 */
1747 		pebs_data_cfg |= PEBS_DATACFG_LBRS |
1748 			((x86_pmu.lbr_nr-1) << PEBS_DATACFG_LBR_SHIFT);
1749 	}
1750 
1751 	return pebs_data_cfg;
1752 }
1753 
1754 static void
1755 pebs_update_state(bool needed_cb, struct cpu_hw_events *cpuc,
1756 		  struct perf_event *event, bool add)
1757 {
1758 	struct pmu *pmu = event->pmu;
1759 
1760 	/*
1761 	 * Make sure we get updated with the first PEBS event.
1762 	 * During removal, ->pebs_data_cfg is still valid for
1763 	 * the last PEBS event. Don't clear it.
1764 	 */
1765 	if ((cpuc->n_pebs == 1) && add)
1766 		cpuc->pebs_data_cfg = PEBS_UPDATE_DS_SW;
1767 
1768 	if (needed_cb != pebs_needs_sched_cb(cpuc)) {
1769 		if (!needed_cb)
1770 			perf_sched_cb_inc(pmu);
1771 		else
1772 			perf_sched_cb_dec(pmu);
1773 
1774 		cpuc->pebs_data_cfg |= PEBS_UPDATE_DS_SW;
1775 	}
1776 
1777 	/*
1778 	 * The PEBS record doesn't shrink on pmu::del(). Doing so would require
1779 	 * iterating all remaining PEBS events to reconstruct the config.
1780 	 */
1781 	if (x86_pmu.intel_cap.pebs_baseline && add) {
1782 		u64 pebs_data_cfg;
1783 
1784 		pebs_data_cfg = pebs_update_adaptive_cfg(event);
1785 		/*
1786 		 * Be sure to update the thresholds when we change the record.
1787 		 */
1788 		if (pebs_data_cfg & ~cpuc->pebs_data_cfg)
1789 			cpuc->pebs_data_cfg |= pebs_data_cfg | PEBS_UPDATE_DS_SW;
1790 	}
1791 }
1792 
1793 u64 intel_get_arch_pebs_data_config(struct perf_event *event)
1794 {
1795 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1796 	u64 pebs_data_cfg = 0;
1797 	u64 cntr_mask;
1798 
1799 	if (WARN_ON(event->hw.idx < 0 || event->hw.idx >= X86_PMC_IDX_MAX))
1800 		return 0;
1801 
1802 	pebs_data_cfg |= pebs_update_adaptive_cfg(event);
1803 
1804 	cntr_mask = (PEBS_DATACFG_CNTR_MASK << PEBS_DATACFG_CNTR_SHIFT) |
1805 		    (PEBS_DATACFG_FIX_MASK << PEBS_DATACFG_FIX_SHIFT) |
1806 		    PEBS_DATACFG_CNTR | PEBS_DATACFG_METRICS;
1807 	pebs_data_cfg |= cpuc->pebs_data_cfg & cntr_mask;
1808 
1809 	return pebs_data_cfg;
1810 }
1811 
1812 void intel_pmu_pebs_add(struct perf_event *event)
1813 {
1814 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1815 	struct hw_perf_event *hwc = &event->hw;
1816 	bool needed_cb = pebs_needs_sched_cb(cpuc);
1817 
1818 	cpuc->n_pebs++;
1819 	if (hwc->flags & PERF_X86_EVENT_LARGE_PEBS)
1820 		cpuc->n_large_pebs++;
1821 	if (hwc->flags & PERF_X86_EVENT_PEBS_VIA_PT)
1822 		cpuc->n_pebs_via_pt++;
1823 
1824 	pebs_update_state(needed_cb, cpuc, event, true);
1825 }
1826 
1827 static void intel_pmu_pebs_via_pt_disable(struct perf_event *event)
1828 {
1829 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1830 
1831 	if (!is_pebs_pt(event))
1832 		return;
1833 
1834 	if (!(cpuc->pebs_enabled & ~PEBS_VIA_PT_MASK))
1835 		cpuc->pebs_enabled &= ~PEBS_VIA_PT_MASK;
1836 }
1837 
1838 static void intel_pmu_pebs_via_pt_enable(struct perf_event *event)
1839 {
1840 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1841 	struct hw_perf_event *hwc = &event->hw;
1842 	struct debug_store *ds = cpuc->ds;
1843 	u64 value = ds->pebs_event_reset[hwc->idx];
1844 	u32 base = MSR_RELOAD_PMC0;
1845 	unsigned int idx = hwc->idx;
1846 
1847 	if (!is_pebs_pt(event))
1848 		return;
1849 
1850 	if (!(event->hw.flags & PERF_X86_EVENT_LARGE_PEBS))
1851 		cpuc->pebs_enabled |= PEBS_PMI_AFTER_EACH_RECORD;
1852 
1853 	cpuc->pebs_enabled |= PEBS_OUTPUT_PT;
1854 
1855 	if (hwc->idx >= INTEL_PMC_IDX_FIXED) {
1856 		base = MSR_RELOAD_FIXED_CTR0;
1857 		idx = hwc->idx - INTEL_PMC_IDX_FIXED;
1858 		if (x86_pmu.intel_cap.pebs_format < 5)
1859 			value = ds->pebs_event_reset[MAX_PEBS_EVENTS_FMT4 + idx];
1860 		else
1861 			value = ds->pebs_event_reset[MAX_PEBS_EVENTS + idx];
1862 	}
1863 	wrmsrq(base + idx, value);
1864 }
1865 
1866 static inline void intel_pmu_drain_large_pebs(struct cpu_hw_events *cpuc)
1867 {
1868 	if (cpuc->n_pebs == cpuc->n_large_pebs &&
1869 	    cpuc->n_pebs != cpuc->n_pebs_via_pt)
1870 		intel_pmu_drain_pebs_buffer();
1871 }
1872 
1873 static void __intel_pmu_pebs_enable(struct perf_event *event)
1874 {
1875 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1876 	struct hw_perf_event *hwc = &event->hw;
1877 
1878 	hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
1879 	cpuc->pebs_enabled |= 1ULL << hwc->idx;
1880 }
1881 
1882 void intel_pmu_pebs_enable(struct perf_event *event)
1883 {
1884 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1885 	u64 pebs_data_cfg = cpuc->pebs_data_cfg & ~PEBS_UPDATE_DS_SW;
1886 	struct hw_perf_event *hwc = &event->hw;
1887 	struct debug_store *ds = cpuc->ds;
1888 	unsigned int idx = hwc->idx;
1889 
1890 	__intel_pmu_pebs_enable(event);
1891 
1892 	if ((event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT) && (x86_pmu.version < 5))
1893 		cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32);
1894 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
1895 		cpuc->pebs_enabled |= 1ULL << 63;
1896 
1897 	if (x86_pmu.intel_cap.pebs_baseline) {
1898 		hwc->config |= ICL_EVENTSEL_ADAPTIVE;
1899 		if (pebs_data_cfg != cpuc->active_pebs_data_cfg) {
1900 			/*
1901 			 * drain_pebs() assumes uniform record size;
1902 			 * hence we need to drain when changing said
1903 			 * size.
1904 			 */
1905 			intel_pmu_drain_pebs_buffer();
1906 			adaptive_pebs_record_size_update();
1907 			wrmsrq(MSR_PEBS_DATA_CFG, pebs_data_cfg);
1908 			cpuc->active_pebs_data_cfg = pebs_data_cfg;
1909 		}
1910 	}
1911 	if (cpuc->pebs_data_cfg & PEBS_UPDATE_DS_SW) {
1912 		cpuc->pebs_data_cfg = pebs_data_cfg;
1913 		pebs_update_threshold(cpuc);
1914 	}
1915 
1916 	if (idx >= INTEL_PMC_IDX_FIXED) {
1917 		if (x86_pmu.intel_cap.pebs_format < 5)
1918 			idx = MAX_PEBS_EVENTS_FMT4 + (idx - INTEL_PMC_IDX_FIXED);
1919 		else
1920 			idx = MAX_PEBS_EVENTS + (idx - INTEL_PMC_IDX_FIXED);
1921 	}
1922 
1923 	/*
1924 	 * Use auto-reload if possible to save a MSR write in the PMI.
1925 	 * This must be done in pmu::start(), because PERF_EVENT_IOC_PERIOD.
1926 	 */
1927 	if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
1928 		ds->pebs_event_reset[idx] =
1929 			(u64)(-hwc->sample_period) & x86_pmu.cntval_mask;
1930 	} else {
1931 		ds->pebs_event_reset[idx] = 0;
1932 	}
1933 
1934 	intel_pmu_pebs_via_pt_enable(event);
1935 }
1936 
1937 void intel_pmu_pebs_del(struct perf_event *event)
1938 {
1939 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1940 	struct hw_perf_event *hwc = &event->hw;
1941 	bool needed_cb = pebs_needs_sched_cb(cpuc);
1942 
1943 	cpuc->n_pebs--;
1944 	if (hwc->flags & PERF_X86_EVENT_LARGE_PEBS)
1945 		cpuc->n_large_pebs--;
1946 	if (hwc->flags & PERF_X86_EVENT_PEBS_VIA_PT)
1947 		cpuc->n_pebs_via_pt--;
1948 
1949 	pebs_update_state(needed_cb, cpuc, event, false);
1950 }
1951 
1952 static void __intel_pmu_pebs_disable(struct perf_event *event)
1953 {
1954 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1955 	struct hw_perf_event *hwc = &event->hw;
1956 
1957 	intel_pmu_drain_large_pebs(cpuc);
1958 	cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
1959 	hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
1960 }
1961 
1962 void intel_pmu_pebs_disable(struct perf_event *event)
1963 {
1964 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1965 	struct hw_perf_event *hwc = &event->hw;
1966 
1967 	__intel_pmu_pebs_disable(event);
1968 
1969 	if ((event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT) &&
1970 	    (x86_pmu.version < 5))
1971 		cpuc->pebs_enabled &= ~(1ULL << (hwc->idx + 32));
1972 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
1973 		cpuc->pebs_enabled &= ~(1ULL << 63);
1974 
1975 	intel_pmu_pebs_via_pt_disable(event);
1976 
1977 	if (cpuc->enabled)
1978 		wrmsrq(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
1979 }
1980 
1981 void intel_pmu_pebs_enable_all(void)
1982 {
1983 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1984 
1985 	if (cpuc->pebs_enabled)
1986 		wrmsrq(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
1987 }
1988 
1989 void intel_pmu_pebs_disable_all(void)
1990 {
1991 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1992 
1993 	if (cpuc->pebs_enabled)
1994 		__intel_pmu_pebs_disable_all();
1995 }
1996 
1997 static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
1998 {
1999 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2000 	unsigned long from = cpuc->lbr_entries[0].from;
2001 	unsigned long old_to, to = cpuc->lbr_entries[0].to;
2002 	unsigned long ip = regs->ip;
2003 	int is_64bit = 0;
2004 	void *kaddr;
2005 	int size;
2006 
2007 	/*
2008 	 * We don't need to fixup if the PEBS assist is fault like
2009 	 */
2010 	if (!x86_pmu.intel_cap.pebs_trap)
2011 		return 1;
2012 
2013 	/*
2014 	 * No LBR entry, no basic block, no rewinding
2015 	 */
2016 	if (!cpuc->lbr_stack.nr || !from || !to)
2017 		return 0;
2018 
2019 	/*
2020 	 * Basic blocks should never cross user/kernel boundaries
2021 	 */
2022 	if (kernel_ip(ip) != kernel_ip(to))
2023 		return 0;
2024 
2025 	/*
2026 	 * unsigned math, either ip is before the start (impossible) or
2027 	 * the basic block is larger than 1 page (sanity)
2028 	 */
2029 	if ((ip - to) > PEBS_FIXUP_SIZE)
2030 		return 0;
2031 
2032 	/*
2033 	 * We sampled a branch insn, rewind using the LBR stack
2034 	 */
2035 	if (ip == to) {
2036 		set_linear_ip(regs, from);
2037 		return 1;
2038 	}
2039 
2040 	size = ip - to;
2041 	if (!kernel_ip(ip)) {
2042 		int bytes;
2043 		u8 *buf = this_cpu_read(insn_buffer);
2044 
2045 		/* 'size' must fit our buffer, see above */
2046 		bytes = copy_from_user_nmi(buf, (void __user *)to, size);
2047 		if (bytes != 0)
2048 			return 0;
2049 
2050 		kaddr = buf;
2051 	} else {
2052 		kaddr = (void *)to;
2053 	}
2054 
2055 	do {
2056 		struct insn insn;
2057 
2058 		old_to = to;
2059 
2060 #ifdef CONFIG_X86_64
2061 		is_64bit = kernel_ip(to) || any_64bit_mode(regs);
2062 #endif
2063 		insn_init(&insn, kaddr, size, is_64bit);
2064 
2065 		/*
2066 		 * Make sure there was not a problem decoding the instruction.
2067 		 * This is doubly important because we have an infinite loop if
2068 		 * insn.length=0.
2069 		 */
2070 		if (insn_get_length(&insn))
2071 			break;
2072 
2073 		to += insn.length;
2074 		kaddr += insn.length;
2075 		size -= insn.length;
2076 	} while (to < ip);
2077 
2078 	if (to == ip) {
2079 		set_linear_ip(regs, old_to);
2080 		return 1;
2081 	}
2082 
2083 	/*
2084 	 * Even though we decoded the basic block, the instruction stream
2085 	 * never matched the given IP, either the TO or the IP got corrupted.
2086 	 */
2087 	return 0;
2088 }
2089 
2090 static inline u64 intel_get_tsx_weight(u64 tsx_tuning)
2091 {
2092 	if (tsx_tuning) {
2093 		union hsw_tsx_tuning tsx = { .value = tsx_tuning };
2094 		return tsx.cycles_last_block;
2095 	}
2096 	return 0;
2097 }
2098 
2099 static inline u64 intel_get_tsx_transaction(u64 tsx_tuning, u64 ax)
2100 {
2101 	u64 txn = (tsx_tuning & PEBS_HSW_TSX_FLAGS) >> 32;
2102 
2103 	/* For RTM XABORTs also log the abort code from AX */
2104 	if ((txn & PERF_TXN_TRANSACTION) && (ax & 1))
2105 		txn |= ((ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT;
2106 	return txn;
2107 }
2108 
2109 static inline u64 get_pebs_status(void *n)
2110 {
2111 	if (x86_pmu.intel_cap.pebs_format < 4)
2112 		return ((struct pebs_record_nhm *)n)->status;
2113 	return ((struct pebs_basic *)n)->applicable_counters;
2114 }
2115 
2116 #define PERF_X86_EVENT_PEBS_HSW_PREC \
2117 		(PERF_X86_EVENT_PEBS_ST_HSW | \
2118 		 PERF_X86_EVENT_PEBS_LD_HSW | \
2119 		 PERF_X86_EVENT_PEBS_NA_HSW)
2120 
2121 static u64 get_data_src(struct perf_event *event, u64 aux)
2122 {
2123 	u64 val = PERF_MEM_NA;
2124 	int fl = event->hw.flags;
2125 	bool fst = fl & (PERF_X86_EVENT_PEBS_ST | PERF_X86_EVENT_PEBS_HSW_PREC);
2126 
2127 	if (fl & PERF_X86_EVENT_PEBS_LDLAT)
2128 		val = load_latency_data(event, aux);
2129 	else if (fl & PERF_X86_EVENT_PEBS_STLAT)
2130 		val = store_latency_data(event, aux);
2131 	else if (fl & PERF_X86_EVENT_PEBS_LAT_HYBRID)
2132 		val = x86_pmu.pebs_latency_data(event, aux);
2133 	else if (fst && (fl & PERF_X86_EVENT_PEBS_HSW_PREC))
2134 		val = precise_datala_hsw(event, aux);
2135 	else if (fst)
2136 		val = precise_store_data(aux);
2137 	return val;
2138 }
2139 
2140 static void setup_pebs_time(struct perf_event *event,
2141 			    struct perf_sample_data *data,
2142 			    u64 tsc)
2143 {
2144 	/* Converting to a user-defined clock is not supported yet. */
2145 	if (event->attr.use_clockid != 0)
2146 		return;
2147 
2148 	/*
2149 	 * Doesn't support the conversion when the TSC is unstable.
2150 	 * The TSC unstable case is a corner case and very unlikely to
2151 	 * happen. If it happens, the TSC in a PEBS record will be
2152 	 * dropped and fall back to perf_event_clock().
2153 	 */
2154 	if (!using_native_sched_clock() || !sched_clock_stable())
2155 		return;
2156 
2157 	data->time = native_sched_clock_from_tsc(tsc) + __sched_clock_offset;
2158 	data->sample_flags |= PERF_SAMPLE_TIME;
2159 }
2160 
2161 #define PERF_SAMPLE_ADDR_TYPE	(PERF_SAMPLE_ADDR |		\
2162 				 PERF_SAMPLE_PHYS_ADDR |	\
2163 				 PERF_SAMPLE_DATA_PAGE_SIZE)
2164 
2165 static void setup_pebs_fixed_sample_data(struct perf_event *event,
2166 				   struct pt_regs *iregs, void *__pebs,
2167 				   struct perf_sample_data *data,
2168 				   struct pt_regs *regs)
2169 {
2170 	/*
2171 	 * We cast to the biggest pebs_record but are careful not to
2172 	 * unconditionally access the 'extra' entries.
2173 	 */
2174 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2175 	struct pebs_record_skl *pebs = __pebs;
2176 	u64 sample_type;
2177 	int fll;
2178 
2179 	if (pebs == NULL)
2180 		return;
2181 
2182 	sample_type = event->attr.sample_type;
2183 	fll = event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT;
2184 
2185 	perf_sample_data_init(data, 0, event->hw.last_period);
2186 
2187 	/*
2188 	 * Use latency for weight (only avail with PEBS-LL)
2189 	 */
2190 	if (fll && (sample_type & PERF_SAMPLE_WEIGHT_TYPE)) {
2191 		data->weight.full = pebs->lat;
2192 		data->sample_flags |= PERF_SAMPLE_WEIGHT_TYPE;
2193 	}
2194 
2195 	/*
2196 	 * data.data_src encodes the data source
2197 	 */
2198 	if (sample_type & PERF_SAMPLE_DATA_SRC) {
2199 		data->data_src.val = get_data_src(event, pebs->dse);
2200 		data->sample_flags |= PERF_SAMPLE_DATA_SRC;
2201 	}
2202 
2203 	/*
2204 	 * We must however always use iregs for the unwinder to stay sane; the
2205 	 * record BP,SP,IP can point into thin air when the record is from a
2206 	 * previous PMI context or an (I)RET happened between the record and
2207 	 * PMI.
2208 	 */
2209 	perf_sample_save_callchain(data, event, iregs);
2210 
2211 	/*
2212 	 * We use the interrupt regs as a base because the PEBS record does not
2213 	 * contain a full regs set, specifically it seems to lack segment
2214 	 * descriptors, which get used by things like user_mode().
2215 	 *
2216 	 * In the simple case fix up only the IP for PERF_SAMPLE_IP.
2217 	 */
2218 	*regs = *iregs;
2219 
2220 	/*
2221 	 * Initialize regs_>flags from PEBS,
2222 	 * Clear exact bit (which uses x86 EFLAGS Reserved bit 3),
2223 	 * i.e., do not rely on it being zero:
2224 	 */
2225 	regs->flags = pebs->flags & ~PERF_EFLAGS_EXACT;
2226 
2227 	if (sample_type & PERF_SAMPLE_REGS_INTR) {
2228 		regs->ax = pebs->ax;
2229 		regs->bx = pebs->bx;
2230 		regs->cx = pebs->cx;
2231 		regs->dx = pebs->dx;
2232 		regs->si = pebs->si;
2233 		regs->di = pebs->di;
2234 
2235 		regs->bp = pebs->bp;
2236 		regs->sp = pebs->sp;
2237 
2238 #ifndef CONFIG_X86_32
2239 		regs->r8 = pebs->r8;
2240 		regs->r9 = pebs->r9;
2241 		regs->r10 = pebs->r10;
2242 		regs->r11 = pebs->r11;
2243 		regs->r12 = pebs->r12;
2244 		regs->r13 = pebs->r13;
2245 		regs->r14 = pebs->r14;
2246 		regs->r15 = pebs->r15;
2247 #endif
2248 	}
2249 
2250 	if (event->attr.precise_ip > 1) {
2251 		/*
2252 		 * Haswell and later processors have an 'eventing IP'
2253 		 * (real IP) which fixes the off-by-1 skid in hardware.
2254 		 * Use it when precise_ip >= 2 :
2255 		 */
2256 		if (x86_pmu.intel_cap.pebs_format >= 2) {
2257 			set_linear_ip(regs, pebs->real_ip);
2258 			regs->flags |= PERF_EFLAGS_EXACT;
2259 		} else {
2260 			/* Otherwise, use PEBS off-by-1 IP: */
2261 			set_linear_ip(regs, pebs->ip);
2262 
2263 			/*
2264 			 * With precise_ip >= 2, try to fix up the off-by-1 IP
2265 			 * using the LBR. If successful, the fixup function
2266 			 * corrects regs->ip and calls set_linear_ip() on regs:
2267 			 */
2268 			if (intel_pmu_pebs_fixup_ip(regs))
2269 				regs->flags |= PERF_EFLAGS_EXACT;
2270 		}
2271 	} else {
2272 		/*
2273 		 * When precise_ip == 1, return the PEBS off-by-1 IP,
2274 		 * no fixup attempted:
2275 		 */
2276 		set_linear_ip(regs, pebs->ip);
2277 	}
2278 
2279 
2280 	if ((sample_type & PERF_SAMPLE_ADDR_TYPE) &&
2281 	    x86_pmu.intel_cap.pebs_format >= 1) {
2282 		data->addr = pebs->dla;
2283 		data->sample_flags |= PERF_SAMPLE_ADDR;
2284 	}
2285 
2286 	if (x86_pmu.intel_cap.pebs_format >= 2) {
2287 		/* Only set the TSX weight when no memory weight. */
2288 		if ((sample_type & PERF_SAMPLE_WEIGHT_TYPE) && !fll) {
2289 			data->weight.full = intel_get_tsx_weight(pebs->tsx_tuning);
2290 			data->sample_flags |= PERF_SAMPLE_WEIGHT_TYPE;
2291 		}
2292 		if (sample_type & PERF_SAMPLE_TRANSACTION) {
2293 			data->txn = intel_get_tsx_transaction(pebs->tsx_tuning,
2294 							      pebs->ax);
2295 			data->sample_flags |= PERF_SAMPLE_TRANSACTION;
2296 		}
2297 	}
2298 
2299 	/*
2300 	 * v3 supplies an accurate time stamp, so we use that
2301 	 * for the time stamp.
2302 	 *
2303 	 * We can only do this for the default trace clock.
2304 	 */
2305 	if (x86_pmu.intel_cap.pebs_format >= 3)
2306 		setup_pebs_time(event, data, pebs->tsc);
2307 
2308 	perf_sample_save_brstack(data, event, &cpuc->lbr_stack, NULL);
2309 }
2310 
2311 static void adaptive_pebs_save_regs(struct pt_regs *regs,
2312 				    struct pebs_gprs *gprs)
2313 {
2314 	regs->ax = gprs->ax;
2315 	regs->bx = gprs->bx;
2316 	regs->cx = gprs->cx;
2317 	regs->dx = gprs->dx;
2318 	regs->si = gprs->si;
2319 	regs->di = gprs->di;
2320 	regs->bp = gprs->bp;
2321 	regs->sp = gprs->sp;
2322 #ifndef CONFIG_X86_32
2323 	regs->r8 = gprs->r8;
2324 	regs->r9 = gprs->r9;
2325 	regs->r10 = gprs->r10;
2326 	regs->r11 = gprs->r11;
2327 	regs->r12 = gprs->r12;
2328 	regs->r13 = gprs->r13;
2329 	regs->r14 = gprs->r14;
2330 	regs->r15 = gprs->r15;
2331 #endif
2332 }
2333 
2334 static void intel_perf_event_update_pmc(struct perf_event *event, u64 pmc)
2335 {
2336 	int shift = 64 - x86_pmu.cntval_bits;
2337 	struct hw_perf_event *hwc;
2338 	u64 delta, prev_pmc;
2339 
2340 	/*
2341 	 * A recorded counter may not have an assigned event in the
2342 	 * following cases. The value should be dropped.
2343 	 * - An event is deleted. There is still an active PEBS event.
2344 	 *   The PEBS record doesn't shrink on pmu::del().
2345 	 *   If the counter of the deleted event once occurred in a PEBS
2346 	 *   record, PEBS still records the counter until the counter is
2347 	 *   reassigned.
2348 	 * - An event is stopped for some reason, e.g., throttled.
2349 	 *   During this period, another event is added and takes the
2350 	 *   counter of the stopped event. The stopped event is assigned
2351 	 *   to another new and uninitialized counter, since the
2352 	 *   x86_pmu_start(RELOAD) is not invoked for a stopped event.
2353 	 *   The PEBS__DATA_CFG is updated regardless of the event state.
2354 	 *   The uninitialized counter can be recorded in a PEBS record.
2355 	 *   But the cpuc->events[uninitialized_counter] is always NULL,
2356 	 *   because the event is stopped. The uninitialized value is
2357 	 *   safely dropped.
2358 	 */
2359 	if (!event)
2360 		return;
2361 
2362 	hwc = &event->hw;
2363 	prev_pmc = local64_read(&hwc->prev_count);
2364 
2365 	/* Only update the count when the PMU is disabled */
2366 	WARN_ON(this_cpu_read(cpu_hw_events.enabled));
2367 	local64_set(&hwc->prev_count, pmc);
2368 
2369 	delta = (pmc << shift) - (prev_pmc << shift);
2370 	delta >>= shift;
2371 
2372 	local64_add(delta, &event->count);
2373 	local64_sub(delta, &hwc->period_left);
2374 }
2375 
2376 static inline void __setup_pebs_counter_group(struct cpu_hw_events *cpuc,
2377 					      struct perf_event *event,
2378 					      struct pebs_cntr_header *cntr,
2379 					      void *next_record)
2380 {
2381 	int bit;
2382 
2383 	for_each_set_bit(bit, (unsigned long *)&cntr->cntr, INTEL_PMC_MAX_GENERIC) {
2384 		intel_perf_event_update_pmc(cpuc->events[bit], *(u64 *)next_record);
2385 		next_record += sizeof(u64);
2386 	}
2387 
2388 	for_each_set_bit(bit, (unsigned long *)&cntr->fixed, INTEL_PMC_MAX_FIXED) {
2389 		/* The slots event will be handled with perf_metric later */
2390 		if ((cntr->metrics == INTEL_CNTR_METRICS) &&
2391 		    (bit + INTEL_PMC_IDX_FIXED == INTEL_PMC_IDX_FIXED_SLOTS)) {
2392 			next_record += sizeof(u64);
2393 			continue;
2394 		}
2395 		intel_perf_event_update_pmc(cpuc->events[bit + INTEL_PMC_IDX_FIXED],
2396 					    *(u64 *)next_record);
2397 		next_record += sizeof(u64);
2398 	}
2399 
2400 	/* HW will reload the value right after the overflow. */
2401 	if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
2402 		local64_set(&event->hw.prev_count, (u64)-event->hw.sample_period);
2403 
2404 	if (cntr->metrics == INTEL_CNTR_METRICS) {
2405 		static_call(intel_pmu_update_topdown_event)
2406 			   (cpuc->events[INTEL_PMC_IDX_FIXED_SLOTS],
2407 			    (u64 *)next_record);
2408 		next_record += 2 * sizeof(u64);
2409 	}
2410 }
2411 
2412 #define PEBS_LATENCY_MASK			0xffff
2413 
2414 static inline void __setup_perf_sample_data(struct perf_event *event,
2415 					    struct pt_regs *iregs,
2416 					    struct perf_sample_data *data)
2417 {
2418 	perf_sample_data_init(data, 0, event->hw.last_period);
2419 
2420 	/*
2421 	 * We must however always use iregs for the unwinder to stay sane; the
2422 	 * record BP,SP,IP can point into thin air when the record is from a
2423 	 * previous PMI context or an (I)RET happened between the record and
2424 	 * PMI.
2425 	 */
2426 	perf_sample_save_callchain(data, event, iregs);
2427 }
2428 
2429 static inline void __setup_pebs_basic_group(struct perf_event *event,
2430 					    struct pt_regs *regs,
2431 					    struct perf_sample_data *data,
2432 					    u64 sample_type, u64 ip,
2433 					    u64 tsc, u16 retire)
2434 {
2435 	/* The ip in basic is EventingIP */
2436 	set_linear_ip(regs, ip);
2437 	regs->flags = PERF_EFLAGS_EXACT;
2438 	setup_pebs_time(event, data, tsc);
2439 
2440 	if (sample_type & PERF_SAMPLE_WEIGHT_STRUCT)
2441 		data->weight.var3_w = retire;
2442 }
2443 
2444 static inline void __setup_pebs_gpr_group(struct perf_event *event,
2445 					  struct pt_regs *regs,
2446 					  struct pebs_gprs *gprs,
2447 					  u64 sample_type)
2448 {
2449 	if (event->attr.precise_ip < 2) {
2450 		set_linear_ip(regs, gprs->ip);
2451 		regs->flags &= ~PERF_EFLAGS_EXACT;
2452 	}
2453 
2454 	if (sample_type & (PERF_SAMPLE_REGS_INTR | PERF_SAMPLE_REGS_USER))
2455 		adaptive_pebs_save_regs(regs, gprs);
2456 }
2457 
2458 static inline void __setup_pebs_meminfo_group(struct perf_event *event,
2459 					      struct perf_sample_data *data,
2460 					      u64 sample_type, u64 latency,
2461 					      u16 instr_latency, u64 address,
2462 					      u64 aux, u64 tsx_tuning, u64 ax)
2463 {
2464 	if (sample_type & PERF_SAMPLE_WEIGHT_TYPE) {
2465 		u64 tsx_latency = intel_get_tsx_weight(tsx_tuning);
2466 
2467 		data->weight.var2_w = instr_latency;
2468 
2469 		/*
2470 		 * Although meminfo::latency is defined as a u64,
2471 		 * only the lower 32 bits include the valid data
2472 		 * in practice on Ice Lake and earlier platforms.
2473 		 */
2474 		if (sample_type & PERF_SAMPLE_WEIGHT)
2475 			data->weight.full = latency ?: tsx_latency;
2476 		else
2477 			data->weight.var1_dw = (u32)latency ?: tsx_latency;
2478 
2479 		data->sample_flags |= PERF_SAMPLE_WEIGHT_TYPE;
2480 	}
2481 
2482 	if (sample_type & PERF_SAMPLE_DATA_SRC) {
2483 		data->data_src.val = get_data_src(event, aux);
2484 		data->sample_flags |= PERF_SAMPLE_DATA_SRC;
2485 	}
2486 
2487 	if (sample_type & PERF_SAMPLE_ADDR_TYPE) {
2488 		data->addr = address;
2489 		data->sample_flags |= PERF_SAMPLE_ADDR;
2490 	}
2491 
2492 	if (sample_type & PERF_SAMPLE_TRANSACTION) {
2493 		data->txn = intel_get_tsx_transaction(tsx_tuning, ax);
2494 		data->sample_flags |= PERF_SAMPLE_TRANSACTION;
2495 	}
2496 }
2497 
2498 /*
2499  * With adaptive PEBS the layout depends on what fields are configured.
2500  */
2501 static void setup_pebs_adaptive_sample_data(struct perf_event *event,
2502 					    struct pt_regs *iregs, void *__pebs,
2503 					    struct perf_sample_data *data,
2504 					    struct pt_regs *regs)
2505 {
2506 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2507 	u64 sample_type = event->attr.sample_type;
2508 	struct pebs_basic *basic = __pebs;
2509 	void *next_record = basic + 1;
2510 	struct pebs_meminfo *meminfo = NULL;
2511 	struct pebs_gprs *gprs = NULL;
2512 	struct x86_perf_regs *perf_regs;
2513 	u64 format_group;
2514 	u16 retire;
2515 
2516 	if (basic == NULL)
2517 		return;
2518 
2519 	perf_regs = container_of(regs, struct x86_perf_regs, regs);
2520 	perf_regs->xmm_regs = NULL;
2521 
2522 	format_group = basic->format_group;
2523 
2524 	__setup_perf_sample_data(event, iregs, data);
2525 
2526 	*regs = *iregs;
2527 
2528 	/* basic group */
2529 	retire = x86_pmu.flags & PMU_FL_RETIRE_LATENCY ?
2530 			basic->retire_latency : 0;
2531 	__setup_pebs_basic_group(event, regs, data, sample_type,
2532 				 basic->ip, basic->tsc, retire);
2533 
2534 	/*
2535 	 * The record for MEMINFO is in front of GP
2536 	 * But PERF_SAMPLE_TRANSACTION needs gprs->ax.
2537 	 * Save the pointer here but process later.
2538 	 */
2539 	if (format_group & PEBS_DATACFG_MEMINFO) {
2540 		meminfo = next_record;
2541 		next_record = meminfo + 1;
2542 	}
2543 
2544 	if (format_group & PEBS_DATACFG_GP) {
2545 		gprs = next_record;
2546 		next_record = gprs + 1;
2547 
2548 		__setup_pebs_gpr_group(event, regs, gprs, sample_type);
2549 	}
2550 
2551 	if (format_group & PEBS_DATACFG_MEMINFO) {
2552 		u64 latency = x86_pmu.flags & PMU_FL_INSTR_LATENCY ?
2553 				meminfo->cache_latency : meminfo->mem_latency;
2554 		u64 instr_latency = x86_pmu.flags & PMU_FL_INSTR_LATENCY ?
2555 				meminfo->instr_latency : 0;
2556 		u64 ax = gprs ? gprs->ax : 0;
2557 
2558 		__setup_pebs_meminfo_group(event, data, sample_type, latency,
2559 					   instr_latency, meminfo->address,
2560 					   meminfo->aux, meminfo->tsx_tuning,
2561 					   ax);
2562 	}
2563 
2564 	if (format_group & PEBS_DATACFG_XMMS) {
2565 		struct pebs_xmm *xmm = next_record;
2566 
2567 		next_record = xmm + 1;
2568 		perf_regs->xmm_regs = xmm->xmm;
2569 	}
2570 
2571 	if (format_group & PEBS_DATACFG_LBRS) {
2572 		struct lbr_entry *lbr = next_record;
2573 		int num_lbr = ((format_group >> PEBS_DATACFG_LBR_SHIFT)
2574 					& 0xff) + 1;
2575 		next_record = next_record + num_lbr * sizeof(struct lbr_entry);
2576 
2577 		if (has_branch_stack(event)) {
2578 			intel_pmu_store_pebs_lbrs(lbr);
2579 			intel_pmu_lbr_save_brstack(data, cpuc, event);
2580 		}
2581 	}
2582 
2583 	if (format_group & (PEBS_DATACFG_CNTR | PEBS_DATACFG_METRICS)) {
2584 		struct pebs_cntr_header *cntr = next_record;
2585 		unsigned int nr;
2586 
2587 		next_record += sizeof(struct pebs_cntr_header);
2588 		/*
2589 		 * The PEBS_DATA_CFG is a global register, which is the
2590 		 * superset configuration for all PEBS events.
2591 		 * For the PEBS record of non-sample-read group, ignore
2592 		 * the counter snapshot fields.
2593 		 */
2594 		if (is_pebs_counter_event_group(event)) {
2595 			__setup_pebs_counter_group(cpuc, event, cntr, next_record);
2596 			data->sample_flags |= PERF_SAMPLE_READ;
2597 		}
2598 
2599 		nr = hweight32(cntr->cntr) + hweight32(cntr->fixed);
2600 		if (cntr->metrics == INTEL_CNTR_METRICS)
2601 			nr += 2;
2602 		next_record += nr * sizeof(u64);
2603 	}
2604 
2605 	WARN_ONCE(next_record != __pebs + basic->format_size,
2606 			"PEBS record size %u, expected %llu, config %llx\n",
2607 			basic->format_size,
2608 			(u64)(next_record - __pebs),
2609 			format_group);
2610 }
2611 
2612 static inline bool arch_pebs_record_continued(struct arch_pebs_header *header)
2613 {
2614 	/* Continue bit or null PEBS record indicates fragment follows. */
2615 	return header->cont || !(header->format & GENMASK_ULL(63, 16));
2616 }
2617 
2618 static void setup_arch_pebs_sample_data(struct perf_event *event,
2619 					struct pt_regs *iregs,
2620 					void *__pebs,
2621 					struct perf_sample_data *data,
2622 					struct pt_regs *regs)
2623 {
2624 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2625 	u64 sample_type = event->attr.sample_type;
2626 	struct arch_pebs_header *header = NULL;
2627 	struct arch_pebs_aux *meminfo = NULL;
2628 	struct arch_pebs_gprs *gprs = NULL;
2629 	struct x86_perf_regs *perf_regs;
2630 	void *next_record;
2631 	void *at = __pebs;
2632 
2633 	if (at == NULL)
2634 		return;
2635 
2636 	perf_regs = container_of(regs, struct x86_perf_regs, regs);
2637 	perf_regs->xmm_regs = NULL;
2638 
2639 	__setup_perf_sample_data(event, iregs, data);
2640 
2641 	*regs = *iregs;
2642 
2643 again:
2644 	header = at;
2645 	next_record = at + sizeof(struct arch_pebs_header);
2646 	if (header->basic) {
2647 		struct arch_pebs_basic *basic = next_record;
2648 		u16 retire = 0;
2649 
2650 		next_record = basic + 1;
2651 
2652 		if (sample_type & PERF_SAMPLE_WEIGHT_STRUCT)
2653 			retire = basic->valid ? basic->retire : 0;
2654 		__setup_pebs_basic_group(event, regs, data, sample_type,
2655 				 basic->ip, basic->tsc, retire);
2656 	}
2657 
2658 	/*
2659 	 * The record for MEMINFO is in front of GP
2660 	 * But PERF_SAMPLE_TRANSACTION needs gprs->ax.
2661 	 * Save the pointer here but process later.
2662 	 */
2663 	if (header->aux) {
2664 		meminfo = next_record;
2665 		next_record = meminfo + 1;
2666 	}
2667 
2668 	if (header->gpr) {
2669 		gprs = next_record;
2670 		next_record = gprs + 1;
2671 
2672 		__setup_pebs_gpr_group(event, regs,
2673 				       (struct pebs_gprs *)gprs,
2674 				       sample_type);
2675 	}
2676 
2677 	if (header->aux) {
2678 		u64 ax = gprs ? gprs->ax : 0;
2679 
2680 		__setup_pebs_meminfo_group(event, data, sample_type,
2681 					   meminfo->cache_latency,
2682 					   meminfo->instr_latency,
2683 					   meminfo->address, meminfo->aux,
2684 					   meminfo->tsx_tuning, ax);
2685 	}
2686 
2687 	if (header->xmm) {
2688 		struct pebs_xmm *xmm;
2689 
2690 		next_record += sizeof(struct arch_pebs_xer_header);
2691 
2692 		xmm = next_record;
2693 		perf_regs->xmm_regs = xmm->xmm;
2694 		next_record = xmm + 1;
2695 	}
2696 
2697 	if (header->lbr) {
2698 		struct arch_pebs_lbr_header *lbr_header = next_record;
2699 		struct lbr_entry *lbr;
2700 		int num_lbr;
2701 
2702 		next_record = lbr_header + 1;
2703 		lbr = next_record;
2704 
2705 		num_lbr = header->lbr == ARCH_PEBS_LBR_NUM_VAR ?
2706 				lbr_header->depth :
2707 				header->lbr * ARCH_PEBS_BASE_LBR_ENTRIES;
2708 		next_record += num_lbr * sizeof(struct lbr_entry);
2709 
2710 		if (has_branch_stack(event)) {
2711 			intel_pmu_store_pebs_lbrs(lbr);
2712 			intel_pmu_lbr_save_brstack(data, cpuc, event);
2713 		}
2714 	}
2715 
2716 	if (header->cntr) {
2717 		struct arch_pebs_cntr_header *cntr = next_record;
2718 		unsigned int nr;
2719 
2720 		next_record += sizeof(struct arch_pebs_cntr_header);
2721 
2722 		if (is_pebs_counter_event_group(event)) {
2723 			__setup_pebs_counter_group(cpuc, event,
2724 				(struct pebs_cntr_header *)cntr, next_record);
2725 			data->sample_flags |= PERF_SAMPLE_READ;
2726 		}
2727 
2728 		nr = hweight32(cntr->cntr) + hweight32(cntr->fixed);
2729 		if (cntr->metrics == INTEL_CNTR_METRICS)
2730 			nr += 2;
2731 		next_record += nr * sizeof(u64);
2732 	}
2733 
2734 	/* Parse followed fragments if there are. */
2735 	if (arch_pebs_record_continued(header)) {
2736 		at = at + header->size;
2737 		goto again;
2738 	}
2739 }
2740 
2741 static inline void *
2742 get_next_pebs_record_by_bit(void *base, void *top, int bit)
2743 {
2744 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2745 	void *at;
2746 	u64 pebs_status;
2747 
2748 	/*
2749 	 * fmt0 does not have a status bitfield (does not use
2750 	 * perf_record_nhm format)
2751 	 */
2752 	if (x86_pmu.intel_cap.pebs_format < 1)
2753 		return base;
2754 
2755 	if (base == NULL)
2756 		return NULL;
2757 
2758 	for (at = base; at < top; at += cpuc->pebs_record_size) {
2759 		unsigned long status = get_pebs_status(at);
2760 
2761 		if (test_bit(bit, (unsigned long *)&status)) {
2762 			/* PEBS v3 has accurate status bits */
2763 			if (x86_pmu.intel_cap.pebs_format >= 3)
2764 				return at;
2765 
2766 			if (status == (1 << bit))
2767 				return at;
2768 
2769 			/* clear non-PEBS bit and re-check */
2770 			pebs_status = status & cpuc->pebs_enabled;
2771 			pebs_status &= PEBS_COUNTER_MASK;
2772 			if (pebs_status == (1 << bit))
2773 				return at;
2774 		}
2775 	}
2776 	return NULL;
2777 }
2778 
2779 /*
2780  * Special variant of intel_pmu_save_and_restart() for auto-reload.
2781  */
2782 static int
2783 intel_pmu_save_and_restart_reload(struct perf_event *event, int count)
2784 {
2785 	struct hw_perf_event *hwc = &event->hw;
2786 	int shift = 64 - x86_pmu.cntval_bits;
2787 	u64 period = hwc->sample_period;
2788 	u64 prev_raw_count, new_raw_count;
2789 	s64 new, old;
2790 
2791 	WARN_ON(!period);
2792 
2793 	/*
2794 	 * drain_pebs() only happens when the PMU is disabled.
2795 	 */
2796 	WARN_ON(this_cpu_read(cpu_hw_events.enabled));
2797 
2798 	prev_raw_count = local64_read(&hwc->prev_count);
2799 	new_raw_count = rdpmc(hwc->event_base_rdpmc);
2800 	local64_set(&hwc->prev_count, new_raw_count);
2801 
2802 	/*
2803 	 * Since the counter increments a negative counter value and
2804 	 * overflows on the sign switch, giving the interval:
2805 	 *
2806 	 *   [-period, 0]
2807 	 *
2808 	 * the difference between two consecutive reads is:
2809 	 *
2810 	 *   A) value2 - value1;
2811 	 *      when no overflows have happened in between,
2812 	 *
2813 	 *   B) (0 - value1) + (value2 - (-period));
2814 	 *      when one overflow happened in between,
2815 	 *
2816 	 *   C) (0 - value1) + (n - 1) * (period) + (value2 - (-period));
2817 	 *      when @n overflows happened in between.
2818 	 *
2819 	 * Here A) is the obvious difference, B) is the extension to the
2820 	 * discrete interval, where the first term is to the top of the
2821 	 * interval and the second term is from the bottom of the next
2822 	 * interval and C) the extension to multiple intervals, where the
2823 	 * middle term is the whole intervals covered.
2824 	 *
2825 	 * An equivalent of C, by reduction, is:
2826 	 *
2827 	 *   value2 - value1 + n * period
2828 	 */
2829 	new = ((s64)(new_raw_count << shift) >> shift);
2830 	old = ((s64)(prev_raw_count << shift) >> shift);
2831 	local64_add(new - old + count * period, &event->count);
2832 
2833 	local64_set(&hwc->period_left, -new);
2834 
2835 	perf_event_update_userpage(event);
2836 
2837 	return 0;
2838 }
2839 
2840 typedef void (*setup_fn)(struct perf_event *, struct pt_regs *, void *,
2841 			 struct perf_sample_data *, struct pt_regs *);
2842 
2843 static struct pt_regs dummy_iregs;
2844 
2845 static __always_inline void
2846 __intel_pmu_pebs_event(struct perf_event *event,
2847 		       struct pt_regs *iregs,
2848 		       struct pt_regs *regs,
2849 		       struct perf_sample_data *data,
2850 		       void *at,
2851 		       setup_fn setup_sample)
2852 {
2853 	setup_sample(event, iregs, at, data, regs);
2854 	perf_event_output(event, data, regs);
2855 }
2856 
2857 static __always_inline void
2858 __intel_pmu_pebs_last_event(struct perf_event *event,
2859 			    struct pt_regs *iregs,
2860 			    struct pt_regs *regs,
2861 			    struct perf_sample_data *data,
2862 			    void *at,
2863 			    int count,
2864 			    setup_fn setup_sample)
2865 {
2866 	struct hw_perf_event *hwc = &event->hw;
2867 
2868 	setup_sample(event, iregs, at, data, regs);
2869 	if (iregs == &dummy_iregs) {
2870 		/*
2871 		 * The PEBS records may be drained in the non-overflow context,
2872 		 * e.g., large PEBS + context switch. Perf should treat the
2873 		 * last record the same as other PEBS records, and doesn't
2874 		 * invoke the generic overflow handler.
2875 		 */
2876 		perf_event_output(event, data, regs);
2877 	} else {
2878 		/*
2879 		 * All but the last records are processed.
2880 		 * The last one is left to be able to call the overflow handler.
2881 		 */
2882 		perf_event_overflow(event, data, regs);
2883 	}
2884 
2885 	if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
2886 		if ((is_pebs_counter_event_group(event))) {
2887 			/*
2888 			 * The value of each sample has been updated when setup
2889 			 * the corresponding sample data.
2890 			 */
2891 			perf_event_update_userpage(event);
2892 		} else {
2893 			/*
2894 			 * Now, auto-reload is only enabled in fixed period mode.
2895 			 * The reload value is always hwc->sample_period.
2896 			 * May need to change it, if auto-reload is enabled in
2897 			 * freq mode later.
2898 			 */
2899 			intel_pmu_save_and_restart_reload(event, count);
2900 		}
2901 	} else {
2902 		/*
2903 		 * For a non-precise event, it's possible the
2904 		 * counters-snapshotting records a positive value for the
2905 		 * overflowed event. Then the HW auto-reload mechanism
2906 		 * reset the counter to 0 immediately, because the
2907 		 * pebs_event_reset is cleared if the PERF_X86_EVENT_AUTO_RELOAD
2908 		 * is not set. The counter backwards may be observed in a
2909 		 * PMI handler.
2910 		 *
2911 		 * Since the event value has been updated when processing the
2912 		 * counters-snapshotting record, only needs to set the new
2913 		 * period for the counter.
2914 		 */
2915 		if (is_pebs_counter_event_group(event))
2916 			static_call(x86_pmu_set_period)(event);
2917 		else
2918 			intel_pmu_save_and_restart(event);
2919 	}
2920 }
2921 
2922 static __always_inline void
2923 __intel_pmu_pebs_events(struct perf_event *event,
2924 			struct pt_regs *iregs,
2925 			struct perf_sample_data *data,
2926 			void *base, void *top,
2927 			int bit, int count,
2928 			setup_fn setup_sample)
2929 {
2930 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2931 	struct x86_perf_regs perf_regs;
2932 	struct pt_regs *regs = &perf_regs.regs;
2933 	void *at = get_next_pebs_record_by_bit(base, top, bit);
2934 	int cnt = count;
2935 
2936 	if (!iregs)
2937 		iregs = &dummy_iregs;
2938 
2939 	while (cnt > 1) {
2940 		__intel_pmu_pebs_event(event, iregs, regs, data, at, setup_sample);
2941 		at += cpuc->pebs_record_size;
2942 		at = get_next_pebs_record_by_bit(at, top, bit);
2943 		cnt--;
2944 	}
2945 
2946 	__intel_pmu_pebs_last_event(event, iregs, regs, data, at, count, setup_sample);
2947 }
2948 
2949 static void intel_pmu_drain_pebs_core(struct pt_regs *iregs, struct perf_sample_data *data)
2950 {
2951 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2952 	struct debug_store *ds = cpuc->ds;
2953 	struct perf_event *event = cpuc->events[0]; /* PMC0 only */
2954 	struct pebs_record_core *at, *top;
2955 	int n;
2956 
2957 	if (!x86_pmu.pebs_active)
2958 		return;
2959 
2960 	at  = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
2961 	top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
2962 
2963 	/*
2964 	 * Whatever else happens, drain the thing
2965 	 */
2966 	ds->pebs_index = ds->pebs_buffer_base;
2967 
2968 	if (!test_bit(0, cpuc->active_mask))
2969 		return;
2970 
2971 	WARN_ON_ONCE(!event);
2972 
2973 	if (!event->attr.precise_ip)
2974 		return;
2975 
2976 	n = top - at;
2977 	if (n <= 0) {
2978 		if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
2979 			intel_pmu_save_and_restart_reload(event, 0);
2980 		return;
2981 	}
2982 
2983 	__intel_pmu_pebs_events(event, iregs, data, at, top, 0, n,
2984 				setup_pebs_fixed_sample_data);
2985 }
2986 
2987 static void intel_pmu_pebs_event_update_no_drain(struct cpu_hw_events *cpuc, u64 mask)
2988 {
2989 	u64 pebs_enabled = cpuc->pebs_enabled & mask;
2990 	struct perf_event *event;
2991 	int bit;
2992 
2993 	/*
2994 	 * The drain_pebs() could be called twice in a short period
2995 	 * for auto-reload event in pmu::read(). There are no
2996 	 * overflows have happened in between.
2997 	 * It needs to call intel_pmu_save_and_restart_reload() to
2998 	 * update the event->count for this case.
2999 	 */
3000 	for_each_set_bit(bit, (unsigned long *)&pebs_enabled, X86_PMC_IDX_MAX) {
3001 		event = cpuc->events[bit];
3002 		if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
3003 			intel_pmu_save_and_restart_reload(event, 0);
3004 	}
3005 }
3006 
3007 static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs, struct perf_sample_data *data)
3008 {
3009 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
3010 	struct debug_store *ds = cpuc->ds;
3011 	struct perf_event *event;
3012 	void *base, *at, *top;
3013 	short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
3014 	short error[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
3015 	int max_pebs_events = intel_pmu_max_num_pebs(NULL);
3016 	int bit, i, size;
3017 	u64 mask;
3018 
3019 	if (!x86_pmu.pebs_active)
3020 		return;
3021 
3022 	base = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
3023 	top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
3024 
3025 	ds->pebs_index = ds->pebs_buffer_base;
3026 
3027 	mask = x86_pmu.pebs_events_mask;
3028 	size = max_pebs_events;
3029 	if (x86_pmu.flags & PMU_FL_PEBS_ALL) {
3030 		mask |= x86_pmu.fixed_cntr_mask64 << INTEL_PMC_IDX_FIXED;
3031 		size = INTEL_PMC_IDX_FIXED + x86_pmu_max_num_counters_fixed(NULL);
3032 	}
3033 
3034 	if (unlikely(base >= top)) {
3035 		intel_pmu_pebs_event_update_no_drain(cpuc, mask);
3036 		return;
3037 	}
3038 
3039 	for (at = base; at < top; at += x86_pmu.pebs_record_size) {
3040 		struct pebs_record_nhm *p = at;
3041 		u64 pebs_status;
3042 
3043 		pebs_status = p->status & cpuc->pebs_enabled;
3044 		pebs_status &= mask;
3045 
3046 		/* PEBS v3 has more accurate status bits */
3047 		if (x86_pmu.intel_cap.pebs_format >= 3) {
3048 			for_each_set_bit(bit, (unsigned long *)&pebs_status, size)
3049 				counts[bit]++;
3050 
3051 			continue;
3052 		}
3053 
3054 		/*
3055 		 * On some CPUs the PEBS status can be zero when PEBS is
3056 		 * racing with clearing of GLOBAL_STATUS.
3057 		 *
3058 		 * Normally we would drop that record, but in the
3059 		 * case when there is only a single active PEBS event
3060 		 * we can assume it's for that event.
3061 		 */
3062 		if (!pebs_status && cpuc->pebs_enabled &&
3063 			!(cpuc->pebs_enabled & (cpuc->pebs_enabled-1)))
3064 			pebs_status = p->status = cpuc->pebs_enabled;
3065 
3066 		bit = find_first_bit((unsigned long *)&pebs_status,
3067 				     max_pebs_events);
3068 
3069 		if (!(x86_pmu.pebs_events_mask & (1 << bit)))
3070 			continue;
3071 
3072 		/*
3073 		 * The PEBS hardware does not deal well with the situation
3074 		 * when events happen near to each other and multiple bits
3075 		 * are set. But it should happen rarely.
3076 		 *
3077 		 * If these events include one PEBS and multiple non-PEBS
3078 		 * events, it doesn't impact PEBS record. The record will
3079 		 * be handled normally. (slow path)
3080 		 *
3081 		 * If these events include two or more PEBS events, the
3082 		 * records for the events can be collapsed into a single
3083 		 * one, and it's not possible to reconstruct all events
3084 		 * that caused the PEBS record. It's called collision.
3085 		 * If collision happened, the record will be dropped.
3086 		 */
3087 		if (pebs_status != (1ULL << bit)) {
3088 			for_each_set_bit(i, (unsigned long *)&pebs_status, size)
3089 				error[i]++;
3090 			continue;
3091 		}
3092 
3093 		counts[bit]++;
3094 	}
3095 
3096 	for_each_set_bit(bit, (unsigned long *)&mask, size) {
3097 		if ((counts[bit] == 0) && (error[bit] == 0))
3098 			continue;
3099 
3100 		event = cpuc->events[bit];
3101 		if (WARN_ON_ONCE(!event))
3102 			continue;
3103 
3104 		if (WARN_ON_ONCE(!event->attr.precise_ip))
3105 			continue;
3106 
3107 		/* log dropped samples number */
3108 		if (error[bit]) {
3109 			perf_log_lost_samples(event, error[bit]);
3110 
3111 			if (iregs)
3112 				perf_event_account_interrupt(event);
3113 		}
3114 
3115 		if (counts[bit]) {
3116 			__intel_pmu_pebs_events(event, iregs, data, base,
3117 						top, bit, counts[bit],
3118 						setup_pebs_fixed_sample_data);
3119 		}
3120 	}
3121 }
3122 
3123 static __always_inline void
3124 __intel_pmu_handle_pebs_record(struct pt_regs *iregs,
3125 			       struct pt_regs *regs,
3126 			       struct perf_sample_data *data,
3127 			       void *at, u64 pebs_status,
3128 			       short *counts, void **last,
3129 			       setup_fn setup_sample)
3130 {
3131 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
3132 	struct perf_event *event;
3133 	int bit;
3134 
3135 	for_each_set_bit(bit, (unsigned long *)&pebs_status, X86_PMC_IDX_MAX) {
3136 		event = cpuc->events[bit];
3137 
3138 		if (WARN_ON_ONCE(!event) ||
3139 		    WARN_ON_ONCE(!event->attr.precise_ip))
3140 			continue;
3141 
3142 		if (counts[bit]++) {
3143 			__intel_pmu_pebs_event(event, iregs, regs, data,
3144 					       last[bit], setup_sample);
3145 		}
3146 
3147 		last[bit] = at;
3148 	}
3149 }
3150 
3151 static __always_inline void
3152 __intel_pmu_handle_last_pebs_record(struct pt_regs *iregs,
3153 				    struct pt_regs *regs,
3154 				    struct perf_sample_data *data,
3155 				    u64 mask, short *counts, void **last,
3156 				    setup_fn setup_sample)
3157 {
3158 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
3159 	struct perf_event *event;
3160 	int bit;
3161 
3162 	for_each_set_bit(bit, (unsigned long *)&mask, X86_PMC_IDX_MAX) {
3163 		if (!counts[bit])
3164 			continue;
3165 
3166 		event = cpuc->events[bit];
3167 
3168 		__intel_pmu_pebs_last_event(event, iregs, regs, data, last[bit],
3169 					    counts[bit], setup_sample);
3170 	}
3171 
3172 }
3173 
3174 static void intel_pmu_drain_pebs_icl(struct pt_regs *iregs, struct perf_sample_data *data)
3175 {
3176 	short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
3177 	void *last[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS];
3178 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
3179 	struct debug_store *ds = cpuc->ds;
3180 	struct x86_perf_regs perf_regs;
3181 	struct pt_regs *regs = &perf_regs.regs;
3182 	struct pebs_basic *basic;
3183 	void *base, *at, *top;
3184 	u64 mask;
3185 
3186 	if (!x86_pmu.pebs_active)
3187 		return;
3188 
3189 	base = (struct pebs_basic *)(unsigned long)ds->pebs_buffer_base;
3190 	top = (struct pebs_basic *)(unsigned long)ds->pebs_index;
3191 
3192 	ds->pebs_index = ds->pebs_buffer_base;
3193 
3194 	mask = hybrid(cpuc->pmu, pebs_events_mask) |
3195 	       (hybrid(cpuc->pmu, fixed_cntr_mask64) << INTEL_PMC_IDX_FIXED);
3196 	mask &= cpuc->pebs_enabled;
3197 
3198 	if (unlikely(base >= top)) {
3199 		intel_pmu_pebs_event_update_no_drain(cpuc, mask);
3200 		return;
3201 	}
3202 
3203 	if (!iregs)
3204 		iregs = &dummy_iregs;
3205 
3206 	/* Process all but the last event for each counter. */
3207 	for (at = base; at < top; at += basic->format_size) {
3208 		u64 pebs_status;
3209 
3210 		basic = at;
3211 		if (basic->format_size != cpuc->pebs_record_size)
3212 			continue;
3213 
3214 		pebs_status = mask & basic->applicable_counters;
3215 		__intel_pmu_handle_pebs_record(iregs, regs, data, at,
3216 					       pebs_status, counts, last,
3217 					       setup_pebs_adaptive_sample_data);
3218 	}
3219 
3220 	__intel_pmu_handle_last_pebs_record(iregs, regs, data, mask, counts, last,
3221 					    setup_pebs_adaptive_sample_data);
3222 }
3223 
3224 static void intel_pmu_drain_arch_pebs(struct pt_regs *iregs,
3225 				      struct perf_sample_data *data)
3226 {
3227 	short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
3228 	void *last[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS];
3229 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
3230 	union arch_pebs_index index;
3231 	struct x86_perf_regs perf_regs;
3232 	struct pt_regs *regs = &perf_regs.regs;
3233 	void *base, *at, *top;
3234 	u64 mask;
3235 
3236 	rdmsrq(MSR_IA32_PEBS_INDEX, index.whole);
3237 
3238 	if (unlikely(!index.wr)) {
3239 		intel_pmu_pebs_event_update_no_drain(cpuc, X86_PMC_IDX_MAX);
3240 		return;
3241 	}
3242 
3243 	base = cpuc->pebs_vaddr;
3244 	top = cpuc->pebs_vaddr + (index.wr << ARCH_PEBS_INDEX_WR_SHIFT);
3245 
3246 	index.wr = 0;
3247 	index.full = 0;
3248 	index.en = 1;
3249 	if (cpuc->n_pebs == cpuc->n_large_pebs)
3250 		index.thresh = ARCH_PEBS_THRESH_MULTI;
3251 	else
3252 		index.thresh = ARCH_PEBS_THRESH_SINGLE;
3253 	wrmsrq(MSR_IA32_PEBS_INDEX, index.whole);
3254 
3255 	mask = hybrid(cpuc->pmu, arch_pebs_cap).counters & cpuc->pebs_enabled;
3256 
3257 	if (!iregs)
3258 		iregs = &dummy_iregs;
3259 
3260 	/* Process all but the last event for each counter. */
3261 	for (at = base; at < top;) {
3262 		struct arch_pebs_header *header;
3263 		struct arch_pebs_basic *basic;
3264 		u64 pebs_status;
3265 
3266 		header = at;
3267 
3268 		if (WARN_ON_ONCE(!header->size))
3269 			break;
3270 
3271 		/* 1st fragment or single record must have basic group */
3272 		if (!header->basic) {
3273 			at += header->size;
3274 			continue;
3275 		}
3276 
3277 		basic = at + sizeof(struct arch_pebs_header);
3278 		pebs_status = mask & basic->applicable_counters;
3279 		__intel_pmu_handle_pebs_record(iregs, regs, data, at,
3280 					       pebs_status, counts, last,
3281 					       setup_arch_pebs_sample_data);
3282 
3283 		/* Skip non-last fragments */
3284 		while (arch_pebs_record_continued(header)) {
3285 			if (!header->size)
3286 				break;
3287 			at += header->size;
3288 			header = at;
3289 		}
3290 
3291 		/* Skip last fragment or the single record */
3292 		at += header->size;
3293 	}
3294 
3295 	__intel_pmu_handle_last_pebs_record(iregs, regs, data, mask,
3296 					    counts, last,
3297 					    setup_arch_pebs_sample_data);
3298 }
3299 
3300 static void __init intel_arch_pebs_init(void)
3301 {
3302 	/*
3303 	 * Current hybrid platforms always both support arch-PEBS or not
3304 	 * on all kinds of cores. So directly set x86_pmu.arch_pebs flag
3305 	 * if boot cpu supports arch-PEBS.
3306 	 */
3307 	x86_pmu.arch_pebs = 1;
3308 	x86_pmu.pebs_buffer_size = PEBS_BUFFER_SIZE;
3309 	x86_pmu.drain_pebs = intel_pmu_drain_arch_pebs;
3310 	x86_pmu.pebs_capable = ~0ULL;
3311 	x86_pmu.flags |= PMU_FL_PEBS_ALL;
3312 
3313 	x86_pmu.pebs_enable = __intel_pmu_pebs_enable;
3314 	x86_pmu.pebs_disable = __intel_pmu_pebs_disable;
3315 }
3316 
3317 /*
3318  * PEBS probe and setup
3319  */
3320 
3321 static void __init intel_ds_pebs_init(void)
3322 {
3323 	/*
3324 	 * No support for 32bit formats
3325 	 */
3326 	if (!boot_cpu_has(X86_FEATURE_DTES64))
3327 		return;
3328 
3329 	x86_pmu.ds_pebs = boot_cpu_has(X86_FEATURE_PEBS);
3330 	x86_pmu.pebs_buffer_size = PEBS_BUFFER_SIZE;
3331 	if (x86_pmu.version <= 4)
3332 		x86_pmu.pebs_no_isolation = 1;
3333 
3334 	if (x86_pmu.ds_pebs) {
3335 		char pebs_type = x86_pmu.intel_cap.pebs_trap ?  '+' : '-';
3336 		char *pebs_qual = "";
3337 		int format = x86_pmu.intel_cap.pebs_format;
3338 
3339 		if (format < 4)
3340 			x86_pmu.intel_cap.pebs_baseline = 0;
3341 
3342 		x86_pmu.pebs_enable = intel_pmu_pebs_enable;
3343 		x86_pmu.pebs_disable = intel_pmu_pebs_disable;
3344 		x86_pmu.pebs_enable_all = intel_pmu_pebs_enable_all;
3345 		x86_pmu.pebs_disable_all = intel_pmu_pebs_disable_all;
3346 
3347 		switch (format) {
3348 		case 0:
3349 			pr_cont("PEBS fmt0%c, ", pebs_type);
3350 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
3351 			/*
3352 			 * Using >PAGE_SIZE buffers makes the WRMSR to
3353 			 * PERF_GLOBAL_CTRL in intel_pmu_enable_all()
3354 			 * mysteriously hang on Core2.
3355 			 *
3356 			 * As a workaround, we don't do this.
3357 			 */
3358 			x86_pmu.pebs_buffer_size = PAGE_SIZE;
3359 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
3360 			break;
3361 
3362 		case 1:
3363 			pr_cont("PEBS fmt1%c, ", pebs_type);
3364 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
3365 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
3366 			break;
3367 
3368 		case 2:
3369 			pr_cont("PEBS fmt2%c, ", pebs_type);
3370 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_hsw);
3371 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
3372 			break;
3373 
3374 		case 3:
3375 			pr_cont("PEBS fmt3%c, ", pebs_type);
3376 			x86_pmu.pebs_record_size =
3377 						sizeof(struct pebs_record_skl);
3378 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
3379 			x86_pmu.large_pebs_flags |= PERF_SAMPLE_TIME;
3380 			break;
3381 
3382 		case 6:
3383 			if (x86_pmu.intel_cap.pebs_baseline)
3384 				x86_pmu.large_pebs_flags |= PERF_SAMPLE_READ;
3385 			fallthrough;
3386 		case 5:
3387 			x86_pmu.pebs_ept = 1;
3388 			fallthrough;
3389 		case 4:
3390 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_icl;
3391 			x86_pmu.pebs_record_size = sizeof(struct pebs_basic);
3392 			if (x86_pmu.intel_cap.pebs_baseline) {
3393 				x86_pmu.large_pebs_flags |=
3394 					PERF_SAMPLE_BRANCH_STACK |
3395 					PERF_SAMPLE_TIME;
3396 				x86_pmu.flags |= PMU_FL_PEBS_ALL;
3397 				x86_pmu.pebs_capable = ~0ULL;
3398 				pebs_qual = "-baseline";
3399 				x86_get_pmu(smp_processor_id())->capabilities |= PERF_PMU_CAP_EXTENDED_REGS;
3400 			} else {
3401 				/* Only basic record supported */
3402 				x86_pmu.large_pebs_flags &=
3403 					~(PERF_SAMPLE_ADDR |
3404 					  PERF_SAMPLE_TIME |
3405 					  PERF_SAMPLE_DATA_SRC |
3406 					  PERF_SAMPLE_TRANSACTION |
3407 					  PERF_SAMPLE_REGS_USER |
3408 					  PERF_SAMPLE_REGS_INTR);
3409 			}
3410 			pr_cont("PEBS fmt%d%c%s, ", format, pebs_type, pebs_qual);
3411 
3412 			/*
3413 			 * The PEBS-via-PT is not supported on hybrid platforms,
3414 			 * because not all CPUs of a hybrid machine support it.
3415 			 * The global x86_pmu.intel_cap, which only contains the
3416 			 * common capabilities, is used to check the availability
3417 			 * of the feature. The per-PMU pebs_output_pt_available
3418 			 * in a hybrid machine should be ignored.
3419 			 */
3420 			if (x86_pmu.intel_cap.pebs_output_pt_available) {
3421 				pr_cont("PEBS-via-PT, ");
3422 				x86_get_pmu(smp_processor_id())->capabilities |= PERF_PMU_CAP_AUX_OUTPUT;
3423 			}
3424 
3425 			break;
3426 
3427 		default:
3428 			pr_cont("no PEBS fmt%d%c, ", format, pebs_type);
3429 			x86_pmu.ds_pebs = 0;
3430 		}
3431 	}
3432 }
3433 
3434 void __init intel_pebs_init(void)
3435 {
3436 	if (x86_pmu.intel_cap.pebs_format == 0xf)
3437 		intel_arch_pebs_init();
3438 	else
3439 		intel_ds_pebs_init();
3440 }
3441 
3442 void perf_restore_debug_store(void)
3443 {
3444 	struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
3445 
3446 	if (!x86_pmu.bts && !x86_pmu.ds_pebs)
3447 		return;
3448 
3449 	wrmsrq(MSR_IA32_DS_AREA, (unsigned long)ds);
3450 }
3451