?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
x86, numa: Reduce minimum fake node size to 32M
[linux/fpc-iii.git] / arch / x86 / kernel / cpu / perf_event_p4.c
blob81400b93e69483e18bedf1c541753e2f915910cd
1 /*
2 * Netburst Perfomance Events (P4, old Xeon)
3 *
4 * Copyright (C) 2010 Parallels, Inc., Cyrill Gorcunov <gorcunov@openvz.org>
5 * Copyright (C) 2010 Intel Corporation, Lin Ming <ming.m.lin@intel.com>
6 *
7 * For licencing details see kernel-base/COPYING
8 */
9
10 #ifdef CONFIG_CPU_SUP_INTEL
11
12 #include <asm/perf_event_p4.h>
13
14 #define P4_CNTR_LIMIT 3
15 /*
16 * array indices: 0,1 - HT threads, used with HT enabled cpu
17 */
18 struct p4_event_bind {
19 unsigned int opcode; /* Event code and ESCR selector */
20 unsigned int escr_msr[2]; /* ESCR MSR for this event */
21 unsigned int escr_emask; /* valid ESCR EventMask bits */
22 unsigned int shared; /* event is shared across threads */
23 char cntr[2][P4_CNTR_LIMIT]; /* counter index (offset), -1 on abscence */
24 };
25
26 struct p4_pebs_bind {
27 unsigned int metric_pebs;
28 unsigned int metric_vert;
29 };
30
31 /* it sets P4_PEBS_ENABLE_UOP_TAG as well */
32 #define P4_GEN_PEBS_BIND(name, pebs, vert) \
33 [P4_PEBS_METRIC__##name] = { \
34 .metric_pebs = pebs | P4_PEBS_ENABLE_UOP_TAG, \
35 .metric_vert = vert, \
36 }
37
38 /*
39 * note we have P4_PEBS_ENABLE_UOP_TAG always set here
40 *
41 * it's needed for mapping P4_PEBS_CONFIG_METRIC_MASK bits of
42 * event configuration to find out which values are to be
43 * written into MSR_IA32_PEBS_ENABLE and MSR_P4_PEBS_MATRIX_VERT
44 * resgisters
45 */
46 static struct p4_pebs_bind p4_pebs_bind_map[] = {
47 P4_GEN_PEBS_BIND(1stl_cache_load_miss_retired, 0x0000001, 0x0000001),
48 P4_GEN_PEBS_BIND(2ndl_cache_load_miss_retired, 0x0000002, 0x0000001),
49 P4_GEN_PEBS_BIND(dtlb_load_miss_retired, 0x0000004, 0x0000001),
50 P4_GEN_PEBS_BIND(dtlb_store_miss_retired, 0x0000004, 0x0000002),
51 P4_GEN_PEBS_BIND(dtlb_all_miss_retired, 0x0000004, 0x0000003),
52 P4_GEN_PEBS_BIND(tagged_mispred_branch, 0x0018000, 0x0000010),
53 P4_GEN_PEBS_BIND(mob_load_replay_retired, 0x0000200, 0x0000001),
54 P4_GEN_PEBS_BIND(split_load_retired, 0x0000400, 0x0000001),
55 P4_GEN_PEBS_BIND(split_store_retired, 0x0000400, 0x0000002),
56 };
57
58 /*
59 * Note that we don't use CCCR1 here, there is an
60 * exception for P4_BSQ_ALLOCATION but we just have
61 * no workaround
62 *
63 * consider this binding as resources which particular
64 * event may borrow, it doesn't contain EventMask,
65 * Tags and friends -- they are left to a caller
66 */
67 static struct p4_event_bind p4_event_bind_map[] = {
68 [P4_EVENT_TC_DELIVER_MODE] = {
69 .opcode = P4_OPCODE(P4_EVENT_TC_DELIVER_MODE),
70 .escr_msr = { MSR_P4_TC_ESCR0, MSR_P4_TC_ESCR1 },
71 .escr_emask =
72 P4_ESCR_EMASK_BIT(P4_EVENT_TC_DELIVER_MODE, DD) |
73 P4_ESCR_EMASK_BIT(P4_EVENT_TC_DELIVER_MODE, DB) |
74 P4_ESCR_EMASK_BIT(P4_EVENT_TC_DELIVER_MODE, DI) |
75 P4_ESCR_EMASK_BIT(P4_EVENT_TC_DELIVER_MODE, BD) |
76 P4_ESCR_EMASK_BIT(P4_EVENT_TC_DELIVER_MODE, BB) |
77 P4_ESCR_EMASK_BIT(P4_EVENT_TC_DELIVER_MODE, BI) |
78 P4_ESCR_EMASK_BIT(P4_EVENT_TC_DELIVER_MODE, ID),
79 .shared = 1,
80 .cntr = { {4, 5, -1}, {6, 7, -1} },
81 },
82 [P4_EVENT_BPU_FETCH_REQUEST] = {
83 .opcode = P4_OPCODE(P4_EVENT_BPU_FETCH_REQUEST),
84 .escr_msr = { MSR_P4_BPU_ESCR0, MSR_P4_BPU_ESCR1 },
85 .escr_emask =
86 P4_ESCR_EMASK_BIT(P4_EVENT_BPU_FETCH_REQUEST, TCMISS),
87 .cntr = { {0, -1, -1}, {2, -1, -1} },
88 },
89 [P4_EVENT_ITLB_REFERENCE] = {
90 .opcode = P4_OPCODE(P4_EVENT_ITLB_REFERENCE),
91 .escr_msr = { MSR_P4_ITLB_ESCR0, MSR_P4_ITLB_ESCR1 },
92 .escr_emask =
93 P4_ESCR_EMASK_BIT(P4_EVENT_ITLB_REFERENCE, HIT) |
94 P4_ESCR_EMASK_BIT(P4_EVENT_ITLB_REFERENCE, MISS) |
95 P4_ESCR_EMASK_BIT(P4_EVENT_ITLB_REFERENCE, HIT_UK),
96 .cntr = { {0, -1, -1}, {2, -1, -1} },
97 },
98 [P4_EVENT_MEMORY_CANCEL] = {
99 .opcode = P4_OPCODE(P4_EVENT_MEMORY_CANCEL),
100 .escr_msr = { MSR_P4_DAC_ESCR0, MSR_P4_DAC_ESCR1 },
101 .escr_emask =
102 P4_ESCR_EMASK_BIT(P4_EVENT_MEMORY_CANCEL, ST_RB_FULL) |
103 P4_ESCR_EMASK_BIT(P4_EVENT_MEMORY_CANCEL, 64K_CONF),
104 .cntr = { {8, 9, -1}, {10, 11, -1} },
105 },
106 [P4_EVENT_MEMORY_COMPLETE] = {
107 .opcode = P4_OPCODE(P4_EVENT_MEMORY_COMPLETE),
108 .escr_msr = { MSR_P4_SAAT_ESCR0 , MSR_P4_SAAT_ESCR1 },
109 .escr_emask =
110 P4_ESCR_EMASK_BIT(P4_EVENT_MEMORY_COMPLETE, LSC) |
111 P4_ESCR_EMASK_BIT(P4_EVENT_MEMORY_COMPLETE, SSC),
112 .cntr = { {8, 9, -1}, {10, 11, -1} },
113 },
114 [P4_EVENT_LOAD_PORT_REPLAY] = {
115 .opcode = P4_OPCODE(P4_EVENT_LOAD_PORT_REPLAY),
116 .escr_msr = { MSR_P4_SAAT_ESCR0, MSR_P4_SAAT_ESCR1 },
117 .escr_emask =
118 P4_ESCR_EMASK_BIT(P4_EVENT_LOAD_PORT_REPLAY, SPLIT_LD),
119 .cntr = { {8, 9, -1}, {10, 11, -1} },
120 },
121 [P4_EVENT_STORE_PORT_REPLAY] = {
122 .opcode = P4_OPCODE(P4_EVENT_STORE_PORT_REPLAY),
123 .escr_msr = { MSR_P4_SAAT_ESCR0 , MSR_P4_SAAT_ESCR1 },
124 .escr_emask =
125 P4_ESCR_EMASK_BIT(P4_EVENT_STORE_PORT_REPLAY, SPLIT_ST),
126 .cntr = { {8, 9, -1}, {10, 11, -1} },
127 },
128 [P4_EVENT_MOB_LOAD_REPLAY] = {
129 .opcode = P4_OPCODE(P4_EVENT_MOB_LOAD_REPLAY),
130 .escr_msr = { MSR_P4_MOB_ESCR0, MSR_P4_MOB_ESCR1 },
131 .escr_emask =
132 P4_ESCR_EMASK_BIT(P4_EVENT_MOB_LOAD_REPLAY, NO_STA) |
133 P4_ESCR_EMASK_BIT(P4_EVENT_MOB_LOAD_REPLAY, NO_STD) |
134 P4_ESCR_EMASK_BIT(P4_EVENT_MOB_LOAD_REPLAY, PARTIAL_DATA) |
135 P4_ESCR_EMASK_BIT(P4_EVENT_MOB_LOAD_REPLAY, UNALGN_ADDR),
136 .cntr = { {0, -1, -1}, {2, -1, -1} },
137 },
138 [P4_EVENT_PAGE_WALK_TYPE] = {
139 .opcode = P4_OPCODE(P4_EVENT_PAGE_WALK_TYPE),
140 .escr_msr = { MSR_P4_PMH_ESCR0, MSR_P4_PMH_ESCR1 },
141 .escr_emask =
142 P4_ESCR_EMASK_BIT(P4_EVENT_PAGE_WALK_TYPE, DTMISS) |
143 P4_ESCR_EMASK_BIT(P4_EVENT_PAGE_WALK_TYPE, ITMISS),
144 .shared = 1,
145 .cntr = { {0, -1, -1}, {2, -1, -1} },
146 },
147 [P4_EVENT_BSQ_CACHE_REFERENCE] = {
148 .opcode = P4_OPCODE(P4_EVENT_BSQ_CACHE_REFERENCE),
149 .escr_msr = { MSR_P4_BSU_ESCR0, MSR_P4_BSU_ESCR1 },
150 .escr_emask =
151 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_2ndL_HITS) |
152 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_2ndL_HITE) |
153 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_2ndL_HITM) |
154 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_3rdL_HITS) |
155 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_3rdL_HITE) |
156 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_3rdL_HITM) |
157 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_2ndL_MISS) |
158 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_3rdL_MISS) |
159 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, WR_2ndL_MISS),
160 .cntr = { {0, -1, -1}, {2, -1, -1} },
161 },
162 [P4_EVENT_IOQ_ALLOCATION] = {
163 .opcode = P4_OPCODE(P4_EVENT_IOQ_ALLOCATION),
164 .escr_msr = { MSR_P4_FSB_ESCR0, MSR_P4_FSB_ESCR1 },
165 .escr_emask =
166 P4_ESCR_EMASK_BIT(P4_EVENT_IOQ_ALLOCATION, DEFAULT) |
167 P4_ESCR_EMASK_BIT(P4_EVENT_IOQ_ALLOCATION, ALL_READ) |
168 P4_ESCR_EMASK_BIT(P4_EVENT_IOQ_ALLOCATION, ALL_WRITE) |
169 P4_ESCR_EMASK_BIT(P4_EVENT_IOQ_ALLOCATION, MEM_UC) |
170 P4_ESCR_EMASK_BIT(P4_EVENT_IOQ_ALLOCATION, MEM_WC) |
171 P4_ESCR_EMASK_BIT(P4_EVENT_IOQ_ALLOCATION, MEM_WT) |
172 P4_ESCR_EMASK_BIT(P4_EVENT_IOQ_ALLOCATION, MEM_WP) |
173 P4_ESCR_EMASK_BIT(P4_EVENT_IOQ_ALLOCATION, MEM_WB) |
174 P4_ESCR_EMASK_BIT(P4_EVENT_IOQ_ALLOCATION, OWN) |
175 P4_ESCR_EMASK_BIT(P4_EVENT_IOQ_ALLOCATION, OTHER) |
176 P4_ESCR_EMASK_BIT(P4_EVENT_IOQ_ALLOCATION, PREFETCH),
177 .cntr = { {0, -1, -1}, {2, -1, -1} },
178 },
179 [P4_EVENT_IOQ_ACTIVE_ENTRIES] = { /* shared ESCR */
180 .opcode = P4_OPCODE(P4_EVENT_IOQ_ACTIVE_ENTRIES),
181 .escr_msr = { MSR_P4_FSB_ESCR1, MSR_P4_FSB_ESCR1 },
182 .escr_emask =
183 P4_ESCR_EMASK_BIT(P4_EVENT_IOQ_ACTIVE_ENTRIES, DEFAULT) |
184 P4_ESCR_EMASK_BIT(P4_EVENT_IOQ_ACTIVE_ENTRIES, ALL_READ) |
185 P4_ESCR_EMASK_BIT(P4_EVENT_IOQ_ACTIVE_ENTRIES, ALL_WRITE) |
186 P4_ESCR_EMASK_BIT(P4_EVENT_IOQ_ACTIVE_ENTRIES, MEM_UC) |
187 P4_ESCR_EMASK_BIT(P4_EVENT_IOQ_ACTIVE_ENTRIES, MEM_WC) |
188 P4_ESCR_EMASK_BIT(P4_EVENT_IOQ_ACTIVE_ENTRIES, MEM_WT) |
189 P4_ESCR_EMASK_BIT(P4_EVENT_IOQ_ACTIVE_ENTRIES, MEM_WP) |
190 P4_ESCR_EMASK_BIT(P4_EVENT_IOQ_ACTIVE_ENTRIES, MEM_WB) |
191 P4_ESCR_EMASK_BIT(P4_EVENT_IOQ_ACTIVE_ENTRIES, OWN) |
192 P4_ESCR_EMASK_BIT(P4_EVENT_IOQ_ACTIVE_ENTRIES, OTHER) |
193 P4_ESCR_EMASK_BIT(P4_EVENT_IOQ_ACTIVE_ENTRIES, PREFETCH),
194 .cntr = { {2, -1, -1}, {3, -1, -1} },
195 },
196 [P4_EVENT_FSB_DATA_ACTIVITY] = {
197 .opcode = P4_OPCODE(P4_EVENT_FSB_DATA_ACTIVITY),
198 .escr_msr = { MSR_P4_FSB_ESCR0, MSR_P4_FSB_ESCR1 },
199 .escr_emask =
200 P4_ESCR_EMASK_BIT(P4_EVENT_FSB_DATA_ACTIVITY, DRDY_DRV) |
201 P4_ESCR_EMASK_BIT(P4_EVENT_FSB_DATA_ACTIVITY, DRDY_OWN) |
202 P4_ESCR_EMASK_BIT(P4_EVENT_FSB_DATA_ACTIVITY, DRDY_OTHER) |
203 P4_ESCR_EMASK_BIT(P4_EVENT_FSB_DATA_ACTIVITY, DBSY_DRV) |
204 P4_ESCR_EMASK_BIT(P4_EVENT_FSB_DATA_ACTIVITY, DBSY_OWN) |
205 P4_ESCR_EMASK_BIT(P4_EVENT_FSB_DATA_ACTIVITY, DBSY_OTHER),
206 .shared = 1,
207 .cntr = { {0, -1, -1}, {2, -1, -1} },
208 },
209 [P4_EVENT_BSQ_ALLOCATION] = { /* shared ESCR, broken CCCR1 */
210 .opcode = P4_OPCODE(P4_EVENT_BSQ_ALLOCATION),
211 .escr_msr = { MSR_P4_BSU_ESCR0, MSR_P4_BSU_ESCR0 },
212 .escr_emask =
213 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ALLOCATION, REQ_TYPE0) |
214 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ALLOCATION, REQ_TYPE1) |
215 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ALLOCATION, REQ_LEN0) |
216 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ALLOCATION, REQ_LEN1) |
217 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ALLOCATION, REQ_IO_TYPE) |
218 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ALLOCATION, REQ_LOCK_TYPE) |
219 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ALLOCATION, REQ_CACHE_TYPE) |
220 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ALLOCATION, REQ_SPLIT_TYPE) |
221 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ALLOCATION, REQ_DEM_TYPE) |
222 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ALLOCATION, REQ_ORD_TYPE) |
223 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ALLOCATION, MEM_TYPE0) |
224 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ALLOCATION, MEM_TYPE1) |
225 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ALLOCATION, MEM_TYPE2),
226 .cntr = { {0, -1, -1}, {1, -1, -1} },
227 },
228 [P4_EVENT_BSQ_ACTIVE_ENTRIES] = { /* shared ESCR */
229 .opcode = P4_OPCODE(P4_EVENT_BSQ_ACTIVE_ENTRIES),
230 .escr_msr = { MSR_P4_BSU_ESCR1 , MSR_P4_BSU_ESCR1 },
231 .escr_emask =
232 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ACTIVE_ENTRIES, REQ_TYPE0) |
233 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ACTIVE_ENTRIES, REQ_TYPE1) |
234 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ACTIVE_ENTRIES, REQ_LEN0) |
235 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ACTIVE_ENTRIES, REQ_LEN1) |
236 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ACTIVE_ENTRIES, REQ_IO_TYPE) |
237 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ACTIVE_ENTRIES, REQ_LOCK_TYPE) |
238 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ACTIVE_ENTRIES, REQ_CACHE_TYPE) |
239 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ACTIVE_ENTRIES, REQ_SPLIT_TYPE) |
240 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ACTIVE_ENTRIES, REQ_DEM_TYPE) |
241 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ACTIVE_ENTRIES, REQ_ORD_TYPE) |
242 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ACTIVE_ENTRIES, MEM_TYPE0) |
243 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ACTIVE_ENTRIES, MEM_TYPE1) |
244 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_ACTIVE_ENTRIES, MEM_TYPE2),
245 .cntr = { {2, -1, -1}, {3, -1, -1} },
246 },
247 [P4_EVENT_SSE_INPUT_ASSIST] = {
248 .opcode = P4_OPCODE(P4_EVENT_SSE_INPUT_ASSIST),
249 .escr_msr = { MSR_P4_FIRM_ESCR0, MSR_P4_FIRM_ESCR1 },
250 .escr_emask =
251 P4_ESCR_EMASK_BIT(P4_EVENT_SSE_INPUT_ASSIST, ALL),
252 .shared = 1,
253 .cntr = { {8, 9, -1}, {10, 11, -1} },
254 },
255 [P4_EVENT_PACKED_SP_UOP] = {
256 .opcode = P4_OPCODE(P4_EVENT_PACKED_SP_UOP),
257 .escr_msr = { MSR_P4_FIRM_ESCR0, MSR_P4_FIRM_ESCR1 },
258 .escr_emask =
259 P4_ESCR_EMASK_BIT(P4_EVENT_PACKED_SP_UOP, ALL),
260 .shared = 1,
261 .cntr = { {8, 9, -1}, {10, 11, -1} },
262 },
263 [P4_EVENT_PACKED_DP_UOP] = {
264 .opcode = P4_OPCODE(P4_EVENT_PACKED_DP_UOP),
265 .escr_msr = { MSR_P4_FIRM_ESCR0, MSR_P4_FIRM_ESCR1 },
266 .escr_emask =
267 P4_ESCR_EMASK_BIT(P4_EVENT_PACKED_DP_UOP, ALL),
268 .shared = 1,
269 .cntr = { {8, 9, -1}, {10, 11, -1} },
270 },
271 [P4_EVENT_SCALAR_SP_UOP] = {
272 .opcode = P4_OPCODE(P4_EVENT_SCALAR_SP_UOP),
273 .escr_msr = { MSR_P4_FIRM_ESCR0, MSR_P4_FIRM_ESCR1 },
274 .escr_emask =
275 P4_ESCR_EMASK_BIT(P4_EVENT_SCALAR_SP_UOP, ALL),
276 .shared = 1,
277 .cntr = { {8, 9, -1}, {10, 11, -1} },
278 },
279 [P4_EVENT_SCALAR_DP_UOP] = {
280 .opcode = P4_OPCODE(P4_EVENT_SCALAR_DP_UOP),
281 .escr_msr = { MSR_P4_FIRM_ESCR0, MSR_P4_FIRM_ESCR1 },
282 .escr_emask =
283 P4_ESCR_EMASK_BIT(P4_EVENT_SCALAR_DP_UOP, ALL),
284 .shared = 1,
285 .cntr = { {8, 9, -1}, {10, 11, -1} },
286 },
287 [P4_EVENT_64BIT_MMX_UOP] = {
288 .opcode = P4_OPCODE(P4_EVENT_64BIT_MMX_UOP),
289 .escr_msr = { MSR_P4_FIRM_ESCR0, MSR_P4_FIRM_ESCR1 },
290 .escr_emask =
291 P4_ESCR_EMASK_BIT(P4_EVENT_64BIT_MMX_UOP, ALL),
292 .shared = 1,
293 .cntr = { {8, 9, -1}, {10, 11, -1} },
294 },
295 [P4_EVENT_128BIT_MMX_UOP] = {
296 .opcode = P4_OPCODE(P4_EVENT_128BIT_MMX_UOP),
297 .escr_msr = { MSR_P4_FIRM_ESCR0, MSR_P4_FIRM_ESCR1 },
298 .escr_emask =
299 P4_ESCR_EMASK_BIT(P4_EVENT_128BIT_MMX_UOP, ALL),
300 .shared = 1,
301 .cntr = { {8, 9, -1}, {10, 11, -1} },
302 },
303 [P4_EVENT_X87_FP_UOP] = {
304 .opcode = P4_OPCODE(P4_EVENT_X87_FP_UOP),
305 .escr_msr = { MSR_P4_FIRM_ESCR0, MSR_P4_FIRM_ESCR1 },
306 .escr_emask =
307 P4_ESCR_EMASK_BIT(P4_EVENT_X87_FP_UOP, ALL),
308 .shared = 1,
309 .cntr = { {8, 9, -1}, {10, 11, -1} },
310 },
311 [P4_EVENT_TC_MISC] = {
312 .opcode = P4_OPCODE(P4_EVENT_TC_MISC),
313 .escr_msr = { MSR_P4_TC_ESCR0, MSR_P4_TC_ESCR1 },
314 .escr_emask =
315 P4_ESCR_EMASK_BIT(P4_EVENT_TC_MISC, FLUSH),
316 .cntr = { {4, 5, -1}, {6, 7, -1} },
317 },
318 [P4_EVENT_GLOBAL_POWER_EVENTS] = {
319 .opcode = P4_OPCODE(P4_EVENT_GLOBAL_POWER_EVENTS),
320 .escr_msr = { MSR_P4_FSB_ESCR0, MSR_P4_FSB_ESCR1 },
321 .escr_emask =
322 P4_ESCR_EMASK_BIT(P4_EVENT_GLOBAL_POWER_EVENTS, RUNNING),
323 .cntr = { {0, -1, -1}, {2, -1, -1} },
324 },
325 [P4_EVENT_TC_MS_XFER] = {
326 .opcode = P4_OPCODE(P4_EVENT_TC_MS_XFER),
327 .escr_msr = { MSR_P4_MS_ESCR0, MSR_P4_MS_ESCR1 },
328 .escr_emask =
329 P4_ESCR_EMASK_BIT(P4_EVENT_TC_MS_XFER, CISC),
330 .cntr = { {4, 5, -1}, {6, 7, -1} },
331 },
332 [P4_EVENT_UOP_QUEUE_WRITES] = {
333 .opcode = P4_OPCODE(P4_EVENT_UOP_QUEUE_WRITES),
334 .escr_msr = { MSR_P4_MS_ESCR0, MSR_P4_MS_ESCR1 },
335 .escr_emask =
336 P4_ESCR_EMASK_BIT(P4_EVENT_UOP_QUEUE_WRITES, FROM_TC_BUILD) |
337 P4_ESCR_EMASK_BIT(P4_EVENT_UOP_QUEUE_WRITES, FROM_TC_DELIVER) |
338 P4_ESCR_EMASK_BIT(P4_EVENT_UOP_QUEUE_WRITES, FROM_ROM),
339 .cntr = { {4, 5, -1}, {6, 7, -1} },
340 },
341 [P4_EVENT_RETIRED_MISPRED_BRANCH_TYPE] = {
342 .opcode = P4_OPCODE(P4_EVENT_RETIRED_MISPRED_BRANCH_TYPE),
343 .escr_msr = { MSR_P4_TBPU_ESCR0 , MSR_P4_TBPU_ESCR0 },
344 .escr_emask =
345 P4_ESCR_EMASK_BIT(P4_EVENT_RETIRED_MISPRED_BRANCH_TYPE, CONDITIONAL) |
346 P4_ESCR_EMASK_BIT(P4_EVENT_RETIRED_MISPRED_BRANCH_TYPE, CALL) |
347 P4_ESCR_EMASK_BIT(P4_EVENT_RETIRED_MISPRED_BRANCH_TYPE, RETURN) |
348 P4_ESCR_EMASK_BIT(P4_EVENT_RETIRED_MISPRED_BRANCH_TYPE, INDIRECT),
349 .cntr = { {4, 5, -1}, {6, 7, -1} },
350 },
351 [P4_EVENT_RETIRED_BRANCH_TYPE] = {
352 .opcode = P4_OPCODE(P4_EVENT_RETIRED_BRANCH_TYPE),
353 .escr_msr = { MSR_P4_TBPU_ESCR0 , MSR_P4_TBPU_ESCR1 },
354 .escr_emask =
355 P4_ESCR_EMASK_BIT(P4_EVENT_RETIRED_BRANCH_TYPE, CONDITIONAL) |
356 P4_ESCR_EMASK_BIT(P4_EVENT_RETIRED_BRANCH_TYPE, CALL) |
357 P4_ESCR_EMASK_BIT(P4_EVENT_RETIRED_BRANCH_TYPE, RETURN) |
358 P4_ESCR_EMASK_BIT(P4_EVENT_RETIRED_BRANCH_TYPE, INDIRECT),
359 .cntr = { {4, 5, -1}, {6, 7, -1} },
360 },
361 [P4_EVENT_RESOURCE_STALL] = {
362 .opcode = P4_OPCODE(P4_EVENT_RESOURCE_STALL),
363 .escr_msr = { MSR_P4_ALF_ESCR0, MSR_P4_ALF_ESCR1 },
364 .escr_emask =
365 P4_ESCR_EMASK_BIT(P4_EVENT_RESOURCE_STALL, SBFULL),
366 .cntr = { {12, 13, 16}, {14, 15, 17} },
367 },
368 [P4_EVENT_WC_BUFFER] = {
369 .opcode = P4_OPCODE(P4_EVENT_WC_BUFFER),
370 .escr_msr = { MSR_P4_DAC_ESCR0, MSR_P4_DAC_ESCR1 },
371 .escr_emask =
372 P4_ESCR_EMASK_BIT(P4_EVENT_WC_BUFFER, WCB_EVICTS) |
373 P4_ESCR_EMASK_BIT(P4_EVENT_WC_BUFFER, WCB_FULL_EVICTS),
374 .shared = 1,
375 .cntr = { {8, 9, -1}, {10, 11, -1} },
376 },
377 [P4_EVENT_B2B_CYCLES] = {
378 .opcode = P4_OPCODE(P4_EVENT_B2B_CYCLES),
379 .escr_msr = { MSR_P4_FSB_ESCR0, MSR_P4_FSB_ESCR1 },
380 .escr_emask = 0,
381 .cntr = { {0, -1, -1}, {2, -1, -1} },
382 },
383 [P4_EVENT_BNR] = {
384 .opcode = P4_OPCODE(P4_EVENT_BNR),
385 .escr_msr = { MSR_P4_FSB_ESCR0, MSR_P4_FSB_ESCR1 },
386 .escr_emask = 0,
387 .cntr = { {0, -1, -1}, {2, -1, -1} },
388 },
389 [P4_EVENT_SNOOP] = {
390 .opcode = P4_OPCODE(P4_EVENT_SNOOP),
391 .escr_msr = { MSR_P4_FSB_ESCR0, MSR_P4_FSB_ESCR1 },
392 .escr_emask = 0,
393 .cntr = { {0, -1, -1}, {2, -1, -1} },
394 },
395 [P4_EVENT_RESPONSE] = {
396 .opcode = P4_OPCODE(P4_EVENT_RESPONSE),
397 .escr_msr = { MSR_P4_FSB_ESCR0, MSR_P4_FSB_ESCR1 },
398 .escr_emask = 0,
399 .cntr = { {0, -1, -1}, {2, -1, -1} },
400 },
401 [P4_EVENT_FRONT_END_EVENT] = {
402 .opcode = P4_OPCODE(P4_EVENT_FRONT_END_EVENT),
403 .escr_msr = { MSR_P4_CRU_ESCR2, MSR_P4_CRU_ESCR3 },
404 .escr_emask =
405 P4_ESCR_EMASK_BIT(P4_EVENT_FRONT_END_EVENT, NBOGUS) |
406 P4_ESCR_EMASK_BIT(P4_EVENT_FRONT_END_EVENT, BOGUS),
407 .cntr = { {12, 13, 16}, {14, 15, 17} },
408 },
409 [P4_EVENT_EXECUTION_EVENT] = {
410 .opcode = P4_OPCODE(P4_EVENT_EXECUTION_EVENT),
411 .escr_msr = { MSR_P4_CRU_ESCR2, MSR_P4_CRU_ESCR3 },
412 .escr_emask =
413 P4_ESCR_EMASK_BIT(P4_EVENT_EXECUTION_EVENT, NBOGUS0) |
414 P4_ESCR_EMASK_BIT(P4_EVENT_EXECUTION_EVENT, NBOGUS1) |
415 P4_ESCR_EMASK_BIT(P4_EVENT_EXECUTION_EVENT, NBOGUS2) |
416 P4_ESCR_EMASK_BIT(P4_EVENT_EXECUTION_EVENT, NBOGUS3) |
417 P4_ESCR_EMASK_BIT(P4_EVENT_EXECUTION_EVENT, BOGUS0) |
418 P4_ESCR_EMASK_BIT(P4_EVENT_EXECUTION_EVENT, BOGUS1) |
419 P4_ESCR_EMASK_BIT(P4_EVENT_EXECUTION_EVENT, BOGUS2) |
420 P4_ESCR_EMASK_BIT(P4_EVENT_EXECUTION_EVENT, BOGUS3),
421 .cntr = { {12, 13, 16}, {14, 15, 17} },
422 },
423 [P4_EVENT_REPLAY_EVENT] = {
424 .opcode = P4_OPCODE(P4_EVENT_REPLAY_EVENT),
425 .escr_msr = { MSR_P4_CRU_ESCR2, MSR_P4_CRU_ESCR3 },
426 .escr_emask =
427 P4_ESCR_EMASK_BIT(P4_EVENT_REPLAY_EVENT, NBOGUS) |
428 P4_ESCR_EMASK_BIT(P4_EVENT_REPLAY_EVENT, BOGUS),
429 .cntr = { {12, 13, 16}, {14, 15, 17} },
430 },
431 [P4_EVENT_INSTR_RETIRED] = {
432 .opcode = P4_OPCODE(P4_EVENT_INSTR_RETIRED),
433 .escr_msr = { MSR_P4_CRU_ESCR0, MSR_P4_CRU_ESCR1 },
434 .escr_emask =
435 P4_ESCR_EMASK_BIT(P4_EVENT_INSTR_RETIRED, NBOGUSNTAG) |
436 P4_ESCR_EMASK_BIT(P4_EVENT_INSTR_RETIRED, NBOGUSTAG) |
437 P4_ESCR_EMASK_BIT(P4_EVENT_INSTR_RETIRED, BOGUSNTAG) |
438 P4_ESCR_EMASK_BIT(P4_EVENT_INSTR_RETIRED, BOGUSTAG),
439 .cntr = { {12, 13, 16}, {14, 15, 17} },
440 },
441 [P4_EVENT_UOPS_RETIRED] = {
442 .opcode = P4_OPCODE(P4_EVENT_UOPS_RETIRED),
443 .escr_msr = { MSR_P4_CRU_ESCR0, MSR_P4_CRU_ESCR1 },
444 .escr_emask =
445 P4_ESCR_EMASK_BIT(P4_EVENT_UOPS_RETIRED, NBOGUS) |
446 P4_ESCR_EMASK_BIT(P4_EVENT_UOPS_RETIRED, BOGUS),
447 .cntr = { {12, 13, 16}, {14, 15, 17} },
448 },
449 [P4_EVENT_UOP_TYPE] = {
450 .opcode = P4_OPCODE(P4_EVENT_UOP_TYPE),
451 .escr_msr = { MSR_P4_RAT_ESCR0, MSR_P4_RAT_ESCR1 },
452 .escr_emask =
453 P4_ESCR_EMASK_BIT(P4_EVENT_UOP_TYPE, TAGLOADS) |
454 P4_ESCR_EMASK_BIT(P4_EVENT_UOP_TYPE, TAGSTORES),
455 .cntr = { {12, 13, 16}, {14, 15, 17} },
456 },
457 [P4_EVENT_BRANCH_RETIRED] = {
458 .opcode = P4_OPCODE(P4_EVENT_BRANCH_RETIRED),
459 .escr_msr = { MSR_P4_CRU_ESCR2, MSR_P4_CRU_ESCR3 },
460 .escr_emask =
461 P4_ESCR_EMASK_BIT(P4_EVENT_BRANCH_RETIRED, MMNP) |
462 P4_ESCR_EMASK_BIT(P4_EVENT_BRANCH_RETIRED, MMNM) |
463 P4_ESCR_EMASK_BIT(P4_EVENT_BRANCH_RETIRED, MMTP) |
464 P4_ESCR_EMASK_BIT(P4_EVENT_BRANCH_RETIRED, MMTM),
465 .cntr = { {12, 13, 16}, {14, 15, 17} },
466 },
467 [P4_EVENT_MISPRED_BRANCH_RETIRED] = {
468 .opcode = P4_OPCODE(P4_EVENT_MISPRED_BRANCH_RETIRED),
469 .escr_msr = { MSR_P4_CRU_ESCR0, MSR_P4_CRU_ESCR1 },
470 .escr_emask =
471 P4_ESCR_EMASK_BIT(P4_EVENT_MISPRED_BRANCH_RETIRED, NBOGUS),
472 .cntr = { {12, 13, 16}, {14, 15, 17} },
473 },
474 [P4_EVENT_X87_ASSIST] = {
475 .opcode = P4_OPCODE(P4_EVENT_X87_ASSIST),
476 .escr_msr = { MSR_P4_CRU_ESCR2, MSR_P4_CRU_ESCR3 },
477 .escr_emask =
478 P4_ESCR_EMASK_BIT(P4_EVENT_X87_ASSIST, FPSU) |
479 P4_ESCR_EMASK_BIT(P4_EVENT_X87_ASSIST, FPSO) |
480 P4_ESCR_EMASK_BIT(P4_EVENT_X87_ASSIST, POAO) |
481 P4_ESCR_EMASK_BIT(P4_EVENT_X87_ASSIST, POAU) |
482 P4_ESCR_EMASK_BIT(P4_EVENT_X87_ASSIST, PREA),
483 .cntr = { {12, 13, 16}, {14, 15, 17} },
484 },
485 [P4_EVENT_MACHINE_CLEAR] = {
486 .opcode = P4_OPCODE(P4_EVENT_MACHINE_CLEAR),
487 .escr_msr = { MSR_P4_CRU_ESCR2, MSR_P4_CRU_ESCR3 },
488 .escr_emask =
489 P4_ESCR_EMASK_BIT(P4_EVENT_MACHINE_CLEAR, CLEAR) |
490 P4_ESCR_EMASK_BIT(P4_EVENT_MACHINE_CLEAR, MOCLEAR) |
491 P4_ESCR_EMASK_BIT(P4_EVENT_MACHINE_CLEAR, SMCLEAR),
492 .cntr = { {12, 13, 16}, {14, 15, 17} },
493 },
494 [P4_EVENT_INSTR_COMPLETED] = {
495 .opcode = P4_OPCODE(P4_EVENT_INSTR_COMPLETED),
496 .escr_msr = { MSR_P4_CRU_ESCR0, MSR_P4_CRU_ESCR1 },
497 .escr_emask =
498 P4_ESCR_EMASK_BIT(P4_EVENT_INSTR_COMPLETED, NBOGUS) |
499 P4_ESCR_EMASK_BIT(P4_EVENT_INSTR_COMPLETED, BOGUS),
500 .cntr = { {12, 13, 16}, {14, 15, 17} },
501 },
502 };
503
504 #define P4_GEN_CACHE_EVENT(event, bit, metric) \
505 p4_config_pack_escr(P4_ESCR_EVENT(event) | \
506 P4_ESCR_EMASK_BIT(event, bit)) | \
507 p4_config_pack_cccr(metric | \
508 P4_CCCR_ESEL(P4_OPCODE_ESEL(P4_OPCODE(event))))
509
510 static __initconst const u64 p4_hw_cache_event_ids
511 [PERF_COUNT_HW_CACHE_MAX]
512 [PERF_COUNT_HW_CACHE_OP_MAX]
513 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
514 {
515 [ C(L1D ) ] = {
516 [ C(OP_READ) ] = {
517 [ C(RESULT_ACCESS) ] = 0x0,
518 [ C(RESULT_MISS) ] = P4_GEN_CACHE_EVENT(P4_EVENT_REPLAY_EVENT, NBOGUS,
519 P4_PEBS_METRIC__1stl_cache_load_miss_retired),
520 },
521 },
522 [ C(LL ) ] = {
523 [ C(OP_READ) ] = {
524 [ C(RESULT_ACCESS) ] = 0x0,
525 [ C(RESULT_MISS) ] = P4_GEN_CACHE_EVENT(P4_EVENT_REPLAY_EVENT, NBOGUS,
526 P4_PEBS_METRIC__2ndl_cache_load_miss_retired),
527 },
528 },
529 [ C(DTLB) ] = {
530 [ C(OP_READ) ] = {
531 [ C(RESULT_ACCESS) ] = 0x0,
532 [ C(RESULT_MISS) ] = P4_GEN_CACHE_EVENT(P4_EVENT_REPLAY_EVENT, NBOGUS,
533 P4_PEBS_METRIC__dtlb_load_miss_retired),
534 },
535 [ C(OP_WRITE) ] = {
536 [ C(RESULT_ACCESS) ] = 0x0,
537 [ C(RESULT_MISS) ] = P4_GEN_CACHE_EVENT(P4_EVENT_REPLAY_EVENT, NBOGUS,
538 P4_PEBS_METRIC__dtlb_store_miss_retired),
539 },
540 },
541 [ C(ITLB) ] = {
542 [ C(OP_READ) ] = {
543 [ C(RESULT_ACCESS) ] = P4_GEN_CACHE_EVENT(P4_EVENT_ITLB_REFERENCE, HIT,
544 P4_PEBS_METRIC__none),
545 [ C(RESULT_MISS) ] = P4_GEN_CACHE_EVENT(P4_EVENT_ITLB_REFERENCE, MISS,
546 P4_PEBS_METRIC__none),
547 },
548 [ C(OP_WRITE) ] = {
549 [ C(RESULT_ACCESS) ] = -1,
550 [ C(RESULT_MISS) ] = -1,
551 },
552 [ C(OP_PREFETCH) ] = {
553 [ C(RESULT_ACCESS) ] = -1,
554 [ C(RESULT_MISS) ] = -1,
555 },
556 },
557 };
558
559 static u64 p4_general_events[PERF_COUNT_HW_MAX] = {
560 /* non-halted CPU clocks */
561 [PERF_COUNT_HW_CPU_CYCLES] =
562 p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_GLOBAL_POWER_EVENTS) |
563 P4_ESCR_EMASK_BIT(P4_EVENT_GLOBAL_POWER_EVENTS, RUNNING)),
564
565 /*
566 * retired instructions
567 * in a sake of simplicity we don't use the FSB tagging
568 */
569 [PERF_COUNT_HW_INSTRUCTIONS] =
570 p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_INSTR_RETIRED) |
571 P4_ESCR_EMASK_BIT(P4_EVENT_INSTR_RETIRED, NBOGUSNTAG) |
572 P4_ESCR_EMASK_BIT(P4_EVENT_INSTR_RETIRED, BOGUSNTAG)),
573
574 /* cache hits */
575 [PERF_COUNT_HW_CACHE_REFERENCES] =
576 p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_BSQ_CACHE_REFERENCE) |
577 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_2ndL_HITS) |
578 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_2ndL_HITE) |
579 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_2ndL_HITM) |
580 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_3rdL_HITS) |
581 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_3rdL_HITE) |
582 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_3rdL_HITM)),
583
584 /* cache misses */
585 [PERF_COUNT_HW_CACHE_MISSES] =
586 p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_BSQ_CACHE_REFERENCE) |
587 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_2ndL_MISS) |
588 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_3rdL_MISS) |
589 P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, WR_2ndL_MISS)),
590
591 /* branch instructions retired */
592 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] =
593 p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_RETIRED_BRANCH_TYPE) |
594 P4_ESCR_EMASK_BIT(P4_EVENT_RETIRED_BRANCH_TYPE, CONDITIONAL) |
595 P4_ESCR_EMASK_BIT(P4_EVENT_RETIRED_BRANCH_TYPE, CALL) |
596 P4_ESCR_EMASK_BIT(P4_EVENT_RETIRED_BRANCH_TYPE, RETURN) |
597 P4_ESCR_EMASK_BIT(P4_EVENT_RETIRED_BRANCH_TYPE, INDIRECT)),
598
599 /* mispredicted branches retired */
600 [PERF_COUNT_HW_BRANCH_MISSES] =
601 p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_MISPRED_BRANCH_RETIRED) |
602 P4_ESCR_EMASK_BIT(P4_EVENT_MISPRED_BRANCH_RETIRED, NBOGUS)),
603
604 /* bus ready clocks (cpu is driving #DRDY_DRV\#DRDY_OWN): */
605 [PERF_COUNT_HW_BUS_CYCLES] =
606 p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_FSB_DATA_ACTIVITY) |
607 P4_ESCR_EMASK_BIT(P4_EVENT_FSB_DATA_ACTIVITY, DRDY_DRV) |
608 P4_ESCR_EMASK_BIT(P4_EVENT_FSB_DATA_ACTIVITY, DRDY_OWN)) |
609 p4_config_pack_cccr(P4_CCCR_EDGE | P4_CCCR_COMPARE),
610 };
611
612 static struct p4_event_bind *p4_config_get_bind(u64 config)
613 {
614 unsigned int evnt = p4_config_unpack_event(config);
615 struct p4_event_bind *bind = NULL;
616
617 if (evnt < ARRAY_SIZE(p4_event_bind_map))
618 bind = &p4_event_bind_map[evnt];
619
620 return bind;
621 }
622
623 static u64 p4_pmu_event_map(int hw_event)
624 {
625 struct p4_event_bind *bind;
626 unsigned int esel;
627 u64 config;
628
629 config = p4_general_events[hw_event];
630 bind = p4_config_get_bind(config);
631 esel = P4_OPCODE_ESEL(bind->opcode);
632 config |= p4_config_pack_cccr(P4_CCCR_ESEL(esel));
633
634 return config;
635 }
636
637 /* check cpu model specifics */
638 static bool p4_event_match_cpu_model(unsigned int event_idx)
639 {
640 /* INSTR_COMPLETED event only exist for model 3, 4, 6 (Prescott) */
641 if (event_idx == P4_EVENT_INSTR_COMPLETED) {
642 if (boot_cpu_data.x86_model != 3 &&
643 boot_cpu_data.x86_model != 4 &&
644 boot_cpu_data.x86_model != 6)
645 return false;
646 }
647
648 /*
649 * For info
650 * - IQ_ESCR0, IQ_ESCR1 only for models 1 and 2
651 */
652
653 return true;
654 }
655
656 static int p4_validate_raw_event(struct perf_event *event)
657 {
658 unsigned int v, emask;
659
660 /* User data may have out-of-bound event index */
661 v = p4_config_unpack_event(event->attr.config);
662 if (v >= ARRAY_SIZE(p4_event_bind_map))
663 return -EINVAL;
664
665 /* It may be unsupported: */
666 if (!p4_event_match_cpu_model(v))
667 return -EINVAL;
668
669 /*
670 * NOTE: P4_CCCR_THREAD_ANY has not the same meaning as
671 * in Architectural Performance Monitoring, it means not
672 * on _which_ logical cpu to count but rather _when_, ie it
673 * depends on logical cpu state -- count event if one cpu active,
674 * none, both or any, so we just allow user to pass any value
675 * desired.
676 *
677 * In turn we always set Tx_OS/Tx_USR bits bound to logical
678 * cpu without their propagation to another cpu
679 */
680
681 /*
682 * if an event is shared accross the logical threads
683 * the user needs special permissions to be able to use it
684 */
685 if (p4_event_bind_map[v].shared) {
686 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
687 return -EACCES;
688 }
689
690 /* ESCR EventMask bits may be invalid */
691 emask = p4_config_unpack_escr(event->attr.config) & P4_ESCR_EVENTMASK_MASK;
692 if (emask & ~p4_event_bind_map[v].escr_emask)
693 return -EINVAL;
694
695 /*
696 * it may have some invalid PEBS bits
697 */
698 if (p4_config_pebs_has(event->attr.config, P4_PEBS_CONFIG_ENABLE))
699 return -EINVAL;
700
701 v = p4_config_unpack_metric(event->attr.config);
702 if (v >= ARRAY_SIZE(p4_pebs_bind_map))
703 return -EINVAL;
704
705 return 0;
706 }
707
708 static int p4_hw_config(struct perf_event *event)
709 {
710 int cpu = get_cpu();
711 int rc = 0;
712 u32 escr, cccr;
713
714 /*
715 * the reason we use cpu that early is that: if we get scheduled
716 * first time on the same cpu -- we will not need swap thread
717 * specific flags in config (and will save some cpu cycles)
718 */
719
720 cccr = p4_default_cccr_conf(cpu);
721 escr = p4_default_escr_conf(cpu, event->attr.exclude_kernel,
722 event->attr.exclude_user);
723 event->hw.config = p4_config_pack_escr(escr) |
724 p4_config_pack_cccr(cccr);
725
726 if (p4_ht_active() && p4_ht_thread(cpu))
727 event->hw.config = p4_set_ht_bit(event->hw.config);
728
729 if (event->attr.type == PERF_TYPE_RAW) {
730
731 /*
732 * Clear bits we reserve to be managed by kernel itself
733 * and never allowed from a user space
734 */
735 event->attr.config &= P4_CONFIG_MASK;
736
737 rc = p4_validate_raw_event(event);
738 if (rc)
739 goto out;
740
741 /*
742 * Note that for RAW events we allow user to use P4_CCCR_RESERVED
743 * bits since we keep additional info here (for cache events and etc)
744 */
745 event->hw.config |= event->attr.config;
746 }
747
748 rc = x86_setup_perfctr(event);
749 out:
750 put_cpu();
751 return rc;
752 }
753
754 static inline int p4_pmu_clear_cccr_ovf(struct hw_perf_event *hwc)
755 {
756 int overflow = 0;
757 u32 low, high;
758
759 rdmsr(hwc->config_base + hwc->idx, low, high);
760
761 /* we need to check high bit for unflagged overflows */
762 if ((low & P4_CCCR_OVF) || !(high & (1 << 31))) {
763 overflow = 1;
764 (void)checking_wrmsrl(hwc->config_base + hwc->idx,
765 ((u64)low) & ~P4_CCCR_OVF);
766 }
767
768 return overflow;
769 }
770
771 static void p4_pmu_disable_pebs(void)
772 {
773 /*
774 * FIXME
775 *
776 * It's still allowed that two threads setup same cache
777 * events so we can't simply clear metrics until we knew
778 * noone is depending on us, so we need kind of counter
779 * for "ReplayEvent" users.
780 *
781 * What is more complex -- RAW events, if user (for some
782 * reason) will pass some cache event metric with improper
783 * event opcode -- it's fine from hardware point of view
784 * but completely nonsence from "meaning" of such action.
785 *
786 * So at moment let leave metrics turned on forever -- it's
787 * ok for now but need to be revisited!
788 *
789 * (void)checking_wrmsrl(MSR_IA32_PEBS_ENABLE, (u64)0);
790 * (void)checking_wrmsrl(MSR_P4_PEBS_MATRIX_VERT, (u64)0);
791 */
792 }
793
794 static inline void p4_pmu_disable_event(struct perf_event *event)
795 {
796 struct hw_perf_event *hwc = &event->hw;
797
798 /*
799 * If event gets disabled while counter is in overflowed
800 * state we need to clear P4_CCCR_OVF, otherwise interrupt get
801 * asserted again and again
802 */
803 (void)checking_wrmsrl(hwc->config_base + hwc->idx,
804 (u64)(p4_config_unpack_cccr(hwc->config)) &
805 ~P4_CCCR_ENABLE & ~P4_CCCR_OVF & ~P4_CCCR_RESERVED);
806 }
807
808 static void p4_pmu_disable_all(void)
809 {
810 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
811 int idx;
812
813 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
814 struct perf_event *event = cpuc->events[idx];
815 if (!test_bit(idx, cpuc->active_mask))
816 continue;
817 p4_pmu_disable_event(event);
818 }
819
820 p4_pmu_disable_pebs();
821 }
822
823 /* configuration must be valid */
824 static void p4_pmu_enable_pebs(u64 config)
825 {
826 struct p4_pebs_bind *bind;
827 unsigned int idx;
828
829 BUILD_BUG_ON(P4_PEBS_METRIC__max > P4_PEBS_CONFIG_METRIC_MASK);
830
831 idx = p4_config_unpack_metric(config);
832 if (idx == P4_PEBS_METRIC__none)
833 return;
834
835 bind = &p4_pebs_bind_map[idx];
836
837 (void)checking_wrmsrl(MSR_IA32_PEBS_ENABLE, (u64)bind->metric_pebs);
838 (void)checking_wrmsrl(MSR_P4_PEBS_MATRIX_VERT, (u64)bind->metric_vert);
839 }
840
841 static void p4_pmu_enable_event(struct perf_event *event)
842 {
843 struct hw_perf_event *hwc = &event->hw;
844 int thread = p4_ht_config_thread(hwc->config);
845 u64 escr_conf = p4_config_unpack_escr(p4_clear_ht_bit(hwc->config));
846 unsigned int idx = p4_config_unpack_event(hwc->config);
847 struct p4_event_bind *bind;
848 u64 escr_addr, cccr;
849
850 bind = &p4_event_bind_map[idx];
851 escr_addr = (u64)bind->escr_msr[thread];
852
853 /*
854 * - we dont support cascaded counters yet
855 * - and counter 1 is broken (erratum)
856 */
857 WARN_ON_ONCE(p4_is_event_cascaded(hwc->config));
858 WARN_ON_ONCE(hwc->idx == 1);
859
860 /* we need a real Event value */
861 escr_conf &= ~P4_ESCR_EVENT_MASK;
862 escr_conf |= P4_ESCR_EVENT(P4_OPCODE_EVNT(bind->opcode));
863
864 cccr = p4_config_unpack_cccr(hwc->config);
865
866 /*
867 * it could be Cache event so we need to write metrics
868 * into additional MSRs
869 */
870 p4_pmu_enable_pebs(hwc->config);
871
872 (void)checking_wrmsrl(escr_addr, escr_conf);
873 (void)checking_wrmsrl(hwc->config_base + hwc->idx,
874 (cccr & ~P4_CCCR_RESERVED) | P4_CCCR_ENABLE);
875 }
876
877 static void p4_pmu_enable_all(int added)
878 {
879 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
880 int idx;
881
882 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
883 struct perf_event *event = cpuc->events[idx];
884 if (!test_bit(idx, cpuc->active_mask))
885 continue;
886 p4_pmu_enable_event(event);
887 }
888 }
889
890 static int p4_pmu_handle_irq(struct pt_regs *regs)
891 {
892 struct perf_sample_data data;
893 struct cpu_hw_events *cpuc;
894 struct perf_event *event;
895 struct hw_perf_event *hwc;
896 int idx, handled = 0;
897 u64 val;
898
899 data.addr = 0;
900 data.raw = NULL;
901
902 cpuc = &__get_cpu_var(cpu_hw_events);
903
904 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
905 int overflow;
906
907 if (!test_bit(idx, cpuc->active_mask)) {
908 /* catch in-flight IRQs */
909 if (__test_and_clear_bit(idx, cpuc->running))
910 handled++;
911 continue;
912 }
913
914 event = cpuc->events[idx];
915 hwc = &event->hw;
916
917 WARN_ON_ONCE(hwc->idx != idx);
918
919 /* it might be unflagged overflow */
920 overflow = p4_pmu_clear_cccr_ovf(hwc);
921
922 val = x86_perf_event_update(event);
923 if (!overflow && (val & (1ULL << (x86_pmu.cntval_bits - 1))))
924 continue;
925
926 handled += overflow;
927
928 /* event overflow for sure */
929 data.period = event->hw.last_period;
930
931 if (!x86_perf_event_set_period(event))
932 continue;
933 if (perf_event_overflow(event, 1, &data, regs))
934 p4_pmu_disable_event(event);
935 }
936
937 if (handled) {
938 /* p4 quirk: unmask it again */
939 apic_write(APIC_LVTPC, apic_read(APIC_LVTPC) & ~APIC_LVT_MASKED);
940 inc_irq_stat(apic_perf_irqs);
941 }
942
943 return handled;
944 }
945
946 /*
947 * swap thread specific fields according to a thread
948 * we are going to run on
949 */
950 static void p4_pmu_swap_config_ts(struct hw_perf_event *hwc, int cpu)
951 {
952 u32 escr, cccr;
953
954 /*
955 * we either lucky and continue on same cpu or no HT support
956 */
957 if (!p4_should_swap_ts(hwc->config, cpu))
958 return;
959
960 /*
961 * the event is migrated from an another logical
962 * cpu, so we need to swap thread specific flags
963 */
964
965 escr = p4_config_unpack_escr(hwc->config);
966 cccr = p4_config_unpack_cccr(hwc->config);
967
968 if (p4_ht_thread(cpu)) {
969 cccr &= ~P4_CCCR_OVF_PMI_T0;
970 cccr |= P4_CCCR_OVF_PMI_T1;
971 if (escr & P4_ESCR_T0_OS) {
972 escr &= ~P4_ESCR_T0_OS;
973 escr |= P4_ESCR_T1_OS;
974 }
975 if (escr & P4_ESCR_T0_USR) {
976 escr &= ~P4_ESCR_T0_USR;
977 escr |= P4_ESCR_T1_USR;
978 }
979 hwc->config = p4_config_pack_escr(escr);
980 hwc->config |= p4_config_pack_cccr(cccr);
981 hwc->config |= P4_CONFIG_HT;
982 } else {
983 cccr &= ~P4_CCCR_OVF_PMI_T1;
984 cccr |= P4_CCCR_OVF_PMI_T0;
985 if (escr & P4_ESCR_T1_OS) {
986 escr &= ~P4_ESCR_T1_OS;
987 escr |= P4_ESCR_T0_OS;
988 }
989 if (escr & P4_ESCR_T1_USR) {
990 escr &= ~P4_ESCR_T1_USR;
991 escr |= P4_ESCR_T0_USR;
992 }
993 hwc->config = p4_config_pack_escr(escr);
994 hwc->config |= p4_config_pack_cccr(cccr);
995 hwc->config &= ~P4_CONFIG_HT;
996 }
997 }
998
999 /*
1000 * ESCR address hashing is tricky, ESCRs are not sequential
1001 * in memory but all starts from MSR_P4_BSU_ESCR0 (0x03a0) and
1002 * the metric between any ESCRs is laid in range [0xa0,0xe1]
1003 *
1004 * so we make ~70% filled hashtable
1005 */
1006
1007 #define P4_ESCR_MSR_BASE 0x000003a0
1008 #define P4_ESCR_MSR_MAX 0x000003e1
1009 #define P4_ESCR_MSR_TABLE_SIZE (P4_ESCR_MSR_MAX - P4_ESCR_MSR_BASE + 1)
1010 #define P4_ESCR_MSR_IDX(msr) (msr - P4_ESCR_MSR_BASE)
1011 #define P4_ESCR_MSR_TABLE_ENTRY(msr) [P4_ESCR_MSR_IDX(msr)] = msr
1012
1013 static const unsigned int p4_escr_table[P4_ESCR_MSR_TABLE_SIZE] = {
1014 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_ALF_ESCR0),
1015 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_ALF_ESCR1),
1016 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_BPU_ESCR0),
1017 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_BPU_ESCR1),
1018 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_BSU_ESCR0),
1019 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_BSU_ESCR1),
1020 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_CRU_ESCR0),
1021 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_CRU_ESCR1),
1022 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_CRU_ESCR2),
1023 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_CRU_ESCR3),
1024 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_CRU_ESCR4),
1025 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_CRU_ESCR5),
1026 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_DAC_ESCR0),
1027 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_DAC_ESCR1),
1028 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_FIRM_ESCR0),
1029 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_FIRM_ESCR1),
1030 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_FLAME_ESCR0),
1031 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_FLAME_ESCR1),
1032 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_FSB_ESCR0),
1033 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_FSB_ESCR1),
1034 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_IQ_ESCR0),
1035 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_IQ_ESCR1),
1036 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_IS_ESCR0),
1037 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_IS_ESCR1),
1038 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_ITLB_ESCR0),
1039 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_ITLB_ESCR1),
1040 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_IX_ESCR0),
1041 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_IX_ESCR1),
1042 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_MOB_ESCR0),
1043 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_MOB_ESCR1),
1044 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_MS_ESCR0),
1045 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_MS_ESCR1),
1046 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_PMH_ESCR0),
1047 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_PMH_ESCR1),
1048 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_RAT_ESCR0),
1049 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_RAT_ESCR1),
1050 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_SAAT_ESCR0),
1051 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_SAAT_ESCR1),
1052 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_SSU_ESCR0),
1053 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_SSU_ESCR1),
1054 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_TBPU_ESCR0),
1055 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_TBPU_ESCR1),
1056 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_TC_ESCR0),
1057 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_TC_ESCR1),
1058 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_U2L_ESCR0),
1059 P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_U2L_ESCR1),
1060 };
1061
1062 static int p4_get_escr_idx(unsigned int addr)
1063 {
1064 unsigned int idx = P4_ESCR_MSR_IDX(addr);
1065
1066 if (unlikely(idx >= P4_ESCR_MSR_TABLE_SIZE ||
1067 !p4_escr_table[idx] ||
1068 p4_escr_table[idx] != addr)) {
1069 WARN_ONCE(1, "P4 PMU: Wrong address passed: %x\n", addr);
1070 return -1;
1071 }
1072
1073 return idx;
1074 }
1075
1076 static int p4_next_cntr(int thread, unsigned long *used_mask,
1077 struct p4_event_bind *bind)
1078 {
1079 int i, j;
1080
1081 for (i = 0; i < P4_CNTR_LIMIT; i++) {
1082 j = bind->cntr[thread][i];
1083 if (j != -1 && !test_bit(j, used_mask))
1084 return j;
1085 }
1086
1087 return -1;
1088 }
1089
1090 static int p4_pmu_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
1091 {
1092 unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
1093 unsigned long escr_mask[BITS_TO_LONGS(P4_ESCR_MSR_TABLE_SIZE)];
1094 int cpu = smp_processor_id();
1095 struct hw_perf_event *hwc;
1096 struct p4_event_bind *bind;
1097 unsigned int i, thread, num;
1098 int cntr_idx, escr_idx;
1099
1100 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
1101 bitmap_zero(escr_mask, P4_ESCR_MSR_TABLE_SIZE);
1102
1103 for (i = 0, num = n; i < n; i++, num--) {
1104
1105 hwc = &cpuc->event_list[i]->hw;
1106 thread = p4_ht_thread(cpu);
1107 bind = p4_config_get_bind(hwc->config);
1108 escr_idx = p4_get_escr_idx(bind->escr_msr[thread]);
1109 if (unlikely(escr_idx == -1))
1110 goto done;
1111
1112 if (hwc->idx != -1 && !p4_should_swap_ts(hwc->config, cpu)) {
1113 cntr_idx = hwc->idx;
1114 if (assign)
1115 assign[i] = hwc->idx;
1116 goto reserve;
1117 }
1118
1119 cntr_idx = p4_next_cntr(thread, used_mask, bind);
1120 if (cntr_idx == -1 || test_bit(escr_idx, escr_mask))
1121 goto done;
1122
1123 p4_pmu_swap_config_ts(hwc, cpu);
1124 if (assign)
1125 assign[i] = cntr_idx;
1126 reserve:
1127 set_bit(cntr_idx, used_mask);
1128 set_bit(escr_idx, escr_mask);
1129 }
1130
1131 done:
1132 return num ? -ENOSPC : 0;
1133 }
1134
1135 static __initconst const struct x86_pmu p4_pmu = {
1136 .name = "Netburst P4/Xeon",
1137 .handle_irq = p4_pmu_handle_irq,
1138 .disable_all = p4_pmu_disable_all,
1139 .enable_all = p4_pmu_enable_all,
1140 .enable = p4_pmu_enable_event,
1141 .disable = p4_pmu_disable_event,
1142 .eventsel = MSR_P4_BPU_CCCR0,
1143 .perfctr = MSR_P4_BPU_PERFCTR0,
1144 .event_map = p4_pmu_event_map,
1145 .max_events = ARRAY_SIZE(p4_general_events),
1146 .get_event_constraints = x86_get_event_constraints,
1147 /*
1148 * IF HT disabled we may need to use all
1149 * ARCH_P4_MAX_CCCR counters simulaneously
1150 * though leave it restricted at moment assuming
1151 * HT is on
1152 */
1153 .num_counters = ARCH_P4_MAX_CCCR,
1154 .apic = 1,
1155 .cntval_bits = 40,
1156 .cntval_mask = (1ULL << 40) - 1,
1157 .max_period = (1ULL << 39) - 1,
1158 .hw_config = p4_hw_config,
1159 .schedule_events = p4_pmu_schedule_events,
1160 /*
1161 * This handles erratum N15 in intel doc 249199-029,
1162 * the counter may not be updated correctly on write
1163 * so we need a second write operation to do the trick
1164 * (the official workaround didn't work)
1165 *
1166 * the former idea is taken from OProfile code
1167 */
1168 .perfctr_second_write = 1,
1169 };
1170
1171 static __init int p4_pmu_init(void)
1172 {
1173 unsigned int low, high;
1174
1175 /* If we get stripped -- indexig fails */
1176 BUILD_BUG_ON(ARCH_P4_MAX_CCCR > X86_PMC_MAX_GENERIC);
1177
1178 rdmsr(MSR_IA32_MISC_ENABLE, low, high);
1179 if (!(low & (1 << 7))) {
1180 pr_cont("unsupported Netburst CPU model %d ",
1181 boot_cpu_data.x86_model);
1182 return -ENODEV;
1183 }
1184
1185 memcpy(hw_cache_event_ids, p4_hw_cache_event_ids,
1186 sizeof(hw_cache_event_ids));
1187
1188 pr_cont("Netburst events, ");
1189
1190 x86_pmu = p4_pmu;
1191
1192 return 0;
1193 }
1194
1195 #endif /* CONFIG_CPU_SUP_INTEL */
Linux with added FPC-III support