Commit | Line | Data |
---|---|---|
a072738e | 1 | /* |
0d2eb44f | 2 | * Netburst Performance Events (P4, old Xeon) |
a072738e CG |
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 | ||
d814f301 | 14 | #define P4_CNTR_LIMIT 3 |
a072738e CG |
15 | /* |
16 | * array indices: 0,1 - HT threads, used with HT enabled cpu | |
17 | */ | |
d814f301 CG |
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 */ | |
c9cf4a01 CG |
21 | unsigned int escr_emask; /* valid ESCR EventMask bits */ |
22 | unsigned int shared; /* event is shared across threads */ | |
1ff3d7d7 | 23 | char cntr[2][P4_CNTR_LIMIT]; /* counter index (offset), -1 on abscence */ |
a072738e CG |
24 | }; |
25 | ||
39ef13a4 | 26 | struct p4_pebs_bind { |
d814f301 CG |
27 | unsigned int metric_pebs; |
28 | unsigned int metric_vert; | |
a072738e CG |
29 | }; |
30 | ||
39ef13a4 CG |
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, \ | |
d814f301 CG |
36 | } |
37 | ||
39ef13a4 CG |
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), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
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, | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
85 | .escr_emask = |
86 | P4_ESCR_EMASK_BIT(P4_EVENT_BPU_FETCH_REQUEST, TCMISS), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
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), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
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), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
109 | .escr_emask = |
110 | P4_ESCR_EMASK_BIT(P4_EVENT_MEMORY_COMPLETE, LSC) | | |
111 | P4_ESCR_EMASK_BIT(P4_EVENT_MEMORY_COMPLETE, SSC), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
117 | .escr_emask = |
118 | P4_ESCR_EMASK_BIT(P4_EVENT_LOAD_PORT_REPLAY, SPLIT_LD), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
124 | .escr_emask = |
125 | P4_ESCR_EMASK_BIT(P4_EVENT_STORE_PORT_REPLAY, SPLIT_ST), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
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), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
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, | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
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), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
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), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
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), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
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, | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
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), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
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), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
250 | .escr_emask = |
251 | P4_ESCR_EMASK_BIT(P4_EVENT_SSE_INPUT_ASSIST, ALL), | |
252 | .shared = 1, | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
258 | .escr_emask = |
259 | P4_ESCR_EMASK_BIT(P4_EVENT_PACKED_SP_UOP, ALL), | |
260 | .shared = 1, | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
266 | .escr_emask = |
267 | P4_ESCR_EMASK_BIT(P4_EVENT_PACKED_DP_UOP, ALL), | |
268 | .shared = 1, | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
274 | .escr_emask = |
275 | P4_ESCR_EMASK_BIT(P4_EVENT_SCALAR_SP_UOP, ALL), | |
276 | .shared = 1, | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
282 | .escr_emask = |
283 | P4_ESCR_EMASK_BIT(P4_EVENT_SCALAR_DP_UOP, ALL), | |
284 | .shared = 1, | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
290 | .escr_emask = |
291 | P4_ESCR_EMASK_BIT(P4_EVENT_64BIT_MMX_UOP, ALL), | |
292 | .shared = 1, | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
298 | .escr_emask = |
299 | P4_ESCR_EMASK_BIT(P4_EVENT_128BIT_MMX_UOP, ALL), | |
300 | .shared = 1, | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
306 | .escr_emask = |
307 | P4_ESCR_EMASK_BIT(P4_EVENT_X87_FP_UOP, ALL), | |
308 | .shared = 1, | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
314 | .escr_emask = |
315 | P4_ESCR_EMASK_BIT(P4_EVENT_TC_MISC, FLUSH), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
321 | .escr_emask = |
322 | P4_ESCR_EMASK_BIT(P4_EVENT_GLOBAL_POWER_EVENTS, RUNNING), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
328 | .escr_emask = |
329 | P4_ESCR_EMASK_BIT(P4_EVENT_TC_MS_XFER, CISC), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
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), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
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), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
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), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
364 | .escr_emask = |
365 | P4_ESCR_EMASK_BIT(P4_EVENT_RESOURCE_STALL, SBFULL), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
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, | |
d814f301 CG |
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 }, | |
c9cf4a01 | 380 | .escr_emask = 0, |
d814f301 CG |
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 }, | |
c9cf4a01 | 386 | .escr_emask = 0, |
d814f301 CG |
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 }, | |
c9cf4a01 | 392 | .escr_emask = 0, |
d814f301 CG |
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 }, | |
c9cf4a01 | 398 | .escr_emask = 0, |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
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), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
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), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
426 | .escr_emask = |
427 | P4_ESCR_EMASK_BIT(P4_EVENT_REPLAY_EVENT, NBOGUS) | | |
428 | P4_ESCR_EMASK_BIT(P4_EVENT_REPLAY_EVENT, BOGUS), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
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), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
444 | .escr_emask = |
445 | P4_ESCR_EMASK_BIT(P4_EVENT_UOPS_RETIRED, NBOGUS) | | |
446 | P4_ESCR_EMASK_BIT(P4_EVENT_UOPS_RETIRED, BOGUS), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
452 | .escr_emask = |
453 | P4_ESCR_EMASK_BIT(P4_EVENT_UOP_TYPE, TAGLOADS) | | |
454 | P4_ESCR_EMASK_BIT(P4_EVENT_UOP_TYPE, TAGSTORES), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
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), | |
d814f301 CG |
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 }, | |
c9cf4a01 | 470 | .escr_emask = |
1d321881 | 471 | P4_ESCR_EMASK_BIT(P4_EVENT_MISPRED_BRANCH_RETIRED, NBOGUS), |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
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), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
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), | |
d814f301 CG |
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 }, | |
c9cf4a01 CG |
497 | .escr_emask = |
498 | P4_ESCR_EMASK_BIT(P4_EVENT_INSTR_COMPLETED, NBOGUS) | | |
499 | P4_ESCR_EMASK_BIT(P4_EVENT_INSTR_COMPLETED, BOGUS), | |
d814f301 CG |
500 | .cntr = { {12, 13, 16}, {14, 15, 17} }, |
501 | }, | |
502 | }; | |
a072738e | 503 | |
39ef13a4 | 504 | #define P4_GEN_CACHE_EVENT(event, bit, metric) \ |
d814f301 CG |
505 | p4_config_pack_escr(P4_ESCR_EVENT(event) | \ |
506 | P4_ESCR_EMASK_BIT(event, bit)) | \ | |
39ef13a4 | 507 | p4_config_pack_cccr(metric | \ |
d814f301 | 508 | P4_CCCR_ESEL(P4_OPCODE_ESEL(P4_OPCODE(event)))) |
cb7d6b50 | 509 | |
caaa8be3 | 510 | static __initconst const u64 p4_hw_cache_event_ids |
cb7d6b50 LM |
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, | |
d814f301 | 518 | [ C(RESULT_MISS) ] = P4_GEN_CACHE_EVENT(P4_EVENT_REPLAY_EVENT, NBOGUS, |
39ef13a4 | 519 | P4_PEBS_METRIC__1stl_cache_load_miss_retired), |
cb7d6b50 LM |
520 | }, |
521 | }, | |
522 | [ C(LL ) ] = { | |
523 | [ C(OP_READ) ] = { | |
524 | [ C(RESULT_ACCESS) ] = 0x0, | |
d814f301 | 525 | [ C(RESULT_MISS) ] = P4_GEN_CACHE_EVENT(P4_EVENT_REPLAY_EVENT, NBOGUS, |
39ef13a4 | 526 | P4_PEBS_METRIC__2ndl_cache_load_miss_retired), |
cb7d6b50 | 527 | }, |
d814f301 | 528 | }, |
cb7d6b50 LM |
529 | [ C(DTLB) ] = { |
530 | [ C(OP_READ) ] = { | |
531 | [ C(RESULT_ACCESS) ] = 0x0, | |
d814f301 | 532 | [ C(RESULT_MISS) ] = P4_GEN_CACHE_EVENT(P4_EVENT_REPLAY_EVENT, NBOGUS, |
39ef13a4 | 533 | P4_PEBS_METRIC__dtlb_load_miss_retired), |
cb7d6b50 LM |
534 | }, |
535 | [ C(OP_WRITE) ] = { | |
536 | [ C(RESULT_ACCESS) ] = 0x0, | |
d814f301 | 537 | [ C(RESULT_MISS) ] = P4_GEN_CACHE_EVENT(P4_EVENT_REPLAY_EVENT, NBOGUS, |
39ef13a4 | 538 | P4_PEBS_METRIC__dtlb_store_miss_retired), |
cb7d6b50 LM |
539 | }, |
540 | }, | |
541 | [ C(ITLB) ] = { | |
542 | [ C(OP_READ) ] = { | |
d814f301 | 543 | [ C(RESULT_ACCESS) ] = P4_GEN_CACHE_EVENT(P4_EVENT_ITLB_REFERENCE, HIT, |
39ef13a4 | 544 | P4_PEBS_METRIC__none), |
d814f301 | 545 | [ C(RESULT_MISS) ] = P4_GEN_CACHE_EVENT(P4_EVENT_ITLB_REFERENCE, MISS, |
39ef13a4 | 546 | P4_PEBS_METRIC__none), |
cb7d6b50 LM |
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 | }, | |
89d6c0b5 PZ |
557 | [ C(NODE) ] = { |
558 | [ C(OP_READ) ] = { | |
559 | [ C(RESULT_ACCESS) ] = -1, | |
560 | [ C(RESULT_MISS) ] = -1, | |
561 | }, | |
562 | [ C(OP_WRITE) ] = { | |
563 | [ C(RESULT_ACCESS) ] = -1, | |
564 | [ C(RESULT_MISS) ] = -1, | |
565 | }, | |
566 | [ C(OP_PREFETCH) ] = { | |
567 | [ C(RESULT_ACCESS) ] = -1, | |
568 | [ C(RESULT_MISS) ] = -1, | |
569 | }, | |
570 | }, | |
cb7d6b50 LM |
571 | }; |
572 | ||
f9129870 CG |
573 | /* |
574 | * Because of Netburst being quite restricted in now | |
575 | * many same events can run simultaneously, we use | |
576 | * event aliases, ie different events which have the | |
577 | * same functionallity but use non-intersected resources | |
578 | * (ESCR/CCCR/couter registers). This allow us to run | |
579 | * two or more semi-same events together. It is done | |
580 | * transparently to a user space. | |
581 | * | |
582 | * Never set any cusom internal bits such as P4_CONFIG_HT, | |
583 | * P4_CONFIG_ALIASABLE or bits for P4_PEBS_METRIC, they are | |
584 | * either up-to-dated automatically either not appliable | |
585 | * at all. | |
586 | * | |
587 | * And be really carefull choosing aliases! | |
588 | */ | |
589 | struct p4_event_alias { | |
590 | u64 orig; | |
591 | u64 alter; | |
592 | } p4_event_aliases[] = { | |
593 | { | |
594 | /* | |
595 | * Non-halted cycles can be substituted with | |
596 | * non-sleeping cycles (see Intel SDM Vol3b for | |
597 | * details). | |
598 | */ | |
599 | .orig = | |
600 | p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_GLOBAL_POWER_EVENTS) | | |
601 | P4_ESCR_EMASK_BIT(P4_EVENT_GLOBAL_POWER_EVENTS, RUNNING)), | |
602 | .alter = | |
603 | p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_EXECUTION_EVENT) | | |
604 | P4_ESCR_EMASK_BIT(P4_EVENT_EXECUTION_EVENT, NBOGUS0)| | |
605 | P4_ESCR_EMASK_BIT(P4_EVENT_EXECUTION_EVENT, NBOGUS1)| | |
606 | P4_ESCR_EMASK_BIT(P4_EVENT_EXECUTION_EVENT, NBOGUS2)| | |
607 | P4_ESCR_EMASK_BIT(P4_EVENT_EXECUTION_EVENT, NBOGUS3)| | |
608 | P4_ESCR_EMASK_BIT(P4_EVENT_EXECUTION_EVENT, BOGUS0) | | |
609 | P4_ESCR_EMASK_BIT(P4_EVENT_EXECUTION_EVENT, BOGUS1) | | |
610 | P4_ESCR_EMASK_BIT(P4_EVENT_EXECUTION_EVENT, BOGUS2) | | |
611 | P4_ESCR_EMASK_BIT(P4_EVENT_EXECUTION_EVENT, BOGUS3))| | |
612 | p4_config_pack_cccr(P4_CCCR_THRESHOLD(15) | P4_CCCR_COMPLEMENT | | |
613 | P4_CCCR_COMPARE), | |
614 | }, | |
615 | }; | |
616 | ||
617 | static u64 p4_get_alias_event(u64 config) | |
618 | { | |
619 | u64 config_match; | |
620 | int i; | |
621 | ||
622 | /* | |
623 | * Probably we're lucky and don't have to do | |
624 | * matching over all config bits. | |
625 | */ | |
626 | if (!(config & P4_CONFIG_ALIASABLE)) | |
627 | return 0; | |
628 | ||
629 | config_match = config & P4_CONFIG_EVENT_ALIAS_MASK; | |
630 | ||
631 | /* | |
632 | * If an event was previously swapped to the alter config | |
633 | * we should swap it back otherwise contnention on registers | |
634 | * will return back. | |
635 | */ | |
636 | for (i = 0; i < ARRAY_SIZE(p4_event_aliases); i++) { | |
637 | if (config_match == p4_event_aliases[i].orig) { | |
638 | config_match = p4_event_aliases[i].alter; | |
639 | break; | |
640 | } else if (config_match == p4_event_aliases[i].alter) { | |
641 | config_match = p4_event_aliases[i].orig; | |
642 | break; | |
643 | } | |
644 | } | |
645 | ||
646 | if (i >= ARRAY_SIZE(p4_event_aliases)) | |
647 | return 0; | |
648 | ||
649 | return config_match | | |
650 | (config & P4_CONFIG_EVENT_ALIAS_IMMUTABLE_BITS); | |
651 | } | |
652 | ||
d814f301 CG |
653 | static u64 p4_general_events[PERF_COUNT_HW_MAX] = { |
654 | /* non-halted CPU clocks */ | |
655 | [PERF_COUNT_HW_CPU_CYCLES] = | |
656 | p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_GLOBAL_POWER_EVENTS) | | |
f9129870 CG |
657 | P4_ESCR_EMASK_BIT(P4_EVENT_GLOBAL_POWER_EVENTS, RUNNING)) | |
658 | P4_CONFIG_ALIASABLE, | |
d814f301 CG |
659 | |
660 | /* | |
661 | * retired instructions | |
662 | * in a sake of simplicity we don't use the FSB tagging | |
663 | */ | |
664 | [PERF_COUNT_HW_INSTRUCTIONS] = | |
665 | p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_INSTR_RETIRED) | | |
666 | P4_ESCR_EMASK_BIT(P4_EVENT_INSTR_RETIRED, NBOGUSNTAG) | | |
667 | P4_ESCR_EMASK_BIT(P4_EVENT_INSTR_RETIRED, BOGUSNTAG)), | |
668 | ||
669 | /* cache hits */ | |
670 | [PERF_COUNT_HW_CACHE_REFERENCES] = | |
671 | p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_BSQ_CACHE_REFERENCE) | | |
672 | P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_2ndL_HITS) | | |
673 | P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_2ndL_HITE) | | |
674 | P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_2ndL_HITM) | | |
675 | P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_3rdL_HITS) | | |
676 | P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_3rdL_HITE) | | |
677 | P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_3rdL_HITM)), | |
678 | ||
679 | /* cache misses */ | |
680 | [PERF_COUNT_HW_CACHE_MISSES] = | |
681 | p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_BSQ_CACHE_REFERENCE) | | |
682 | P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_2ndL_MISS) | | |
683 | P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, RD_3rdL_MISS) | | |
684 | P4_ESCR_EMASK_BIT(P4_EVENT_BSQ_CACHE_REFERENCE, WR_2ndL_MISS)), | |
685 | ||
686 | /* branch instructions retired */ | |
687 | [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = | |
688 | p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_RETIRED_BRANCH_TYPE) | | |
689 | P4_ESCR_EMASK_BIT(P4_EVENT_RETIRED_BRANCH_TYPE, CONDITIONAL) | | |
690 | P4_ESCR_EMASK_BIT(P4_EVENT_RETIRED_BRANCH_TYPE, CALL) | | |
691 | P4_ESCR_EMASK_BIT(P4_EVENT_RETIRED_BRANCH_TYPE, RETURN) | | |
692 | P4_ESCR_EMASK_BIT(P4_EVENT_RETIRED_BRANCH_TYPE, INDIRECT)), | |
693 | ||
694 | /* mispredicted branches retired */ | |
695 | [PERF_COUNT_HW_BRANCH_MISSES] = | |
696 | p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_MISPRED_BRANCH_RETIRED) | | |
697 | P4_ESCR_EMASK_BIT(P4_EVENT_MISPRED_BRANCH_RETIRED, NBOGUS)), | |
698 | ||
699 | /* bus ready clocks (cpu is driving #DRDY_DRV\#DRDY_OWN): */ | |
700 | [PERF_COUNT_HW_BUS_CYCLES] = | |
701 | p4_config_pack_escr(P4_ESCR_EVENT(P4_EVENT_FSB_DATA_ACTIVITY) | | |
702 | P4_ESCR_EMASK_BIT(P4_EVENT_FSB_DATA_ACTIVITY, DRDY_DRV) | | |
703 | P4_ESCR_EMASK_BIT(P4_EVENT_FSB_DATA_ACTIVITY, DRDY_OWN)) | | |
704 | p4_config_pack_cccr(P4_CCCR_EDGE | P4_CCCR_COMPARE), | |
a072738e CG |
705 | }; |
706 | ||
d814f301 CG |
707 | static struct p4_event_bind *p4_config_get_bind(u64 config) |
708 | { | |
709 | unsigned int evnt = p4_config_unpack_event(config); | |
710 | struct p4_event_bind *bind = NULL; | |
711 | ||
712 | if (evnt < ARRAY_SIZE(p4_event_bind_map)) | |
713 | bind = &p4_event_bind_map[evnt]; | |
714 | ||
715 | return bind; | |
716 | } | |
717 | ||
a072738e CG |
718 | static u64 p4_pmu_event_map(int hw_event) |
719 | { | |
d814f301 CG |
720 | struct p4_event_bind *bind; |
721 | unsigned int esel; | |
a072738e CG |
722 | u64 config; |
723 | ||
d814f301 CG |
724 | config = p4_general_events[hw_event]; |
725 | bind = p4_config_get_bind(config); | |
726 | esel = P4_OPCODE_ESEL(bind->opcode); | |
727 | config |= p4_config_pack_cccr(P4_CCCR_ESEL(esel)); | |
a072738e | 728 | |
a072738e CG |
729 | return config; |
730 | } | |
731 | ||
c9cf4a01 CG |
732 | /* check cpu model specifics */ |
733 | static bool p4_event_match_cpu_model(unsigned int event_idx) | |
734 | { | |
735 | /* INSTR_COMPLETED event only exist for model 3, 4, 6 (Prescott) */ | |
736 | if (event_idx == P4_EVENT_INSTR_COMPLETED) { | |
737 | if (boot_cpu_data.x86_model != 3 && | |
738 | boot_cpu_data.x86_model != 4 && | |
739 | boot_cpu_data.x86_model != 6) | |
740 | return false; | |
741 | } | |
742 | ||
743 | /* | |
744 | * For info | |
745 | * - IQ_ESCR0, IQ_ESCR1 only for models 1 and 2 | |
746 | */ | |
747 | ||
748 | return true; | |
749 | } | |
750 | ||
39ef13a4 CG |
751 | static int p4_validate_raw_event(struct perf_event *event) |
752 | { | |
c9cf4a01 | 753 | unsigned int v, emask; |
39ef13a4 | 754 | |
c9cf4a01 | 755 | /* User data may have out-of-bound event index */ |
39ef13a4 | 756 | v = p4_config_unpack_event(event->attr.config); |
c9cf4a01 CG |
757 | if (v >= ARRAY_SIZE(p4_event_bind_map)) |
758 | return -EINVAL; | |
759 | ||
760 | /* It may be unsupported: */ | |
761 | if (!p4_event_match_cpu_model(v)) | |
39ef13a4 | 762 | return -EINVAL; |
c9cf4a01 CG |
763 | |
764 | /* | |
765 | * NOTE: P4_CCCR_THREAD_ANY has not the same meaning as | |
766 | * in Architectural Performance Monitoring, it means not | |
767 | * on _which_ logical cpu to count but rather _when_, ie it | |
768 | * depends on logical cpu state -- count event if one cpu active, | |
769 | * none, both or any, so we just allow user to pass any value | |
770 | * desired. | |
771 | * | |
772 | * In turn we always set Tx_OS/Tx_USR bits bound to logical | |
773 | * cpu without their propagation to another cpu | |
774 | */ | |
775 | ||
776 | /* | |
0d2eb44f | 777 | * if an event is shared across the logical threads |
c9cf4a01 CG |
778 | * the user needs special permissions to be able to use it |
779 | */ | |
d038b12c | 780 | if (p4_ht_active() && p4_event_bind_map[v].shared) { |
c9cf4a01 CG |
781 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
782 | return -EACCES; | |
39ef13a4 CG |
783 | } |
784 | ||
c9cf4a01 CG |
785 | /* ESCR EventMask bits may be invalid */ |
786 | emask = p4_config_unpack_escr(event->attr.config) & P4_ESCR_EVENTMASK_MASK; | |
787 | if (emask & ~p4_event_bind_map[v].escr_emask) | |
788 | return -EINVAL; | |
789 | ||
39ef13a4 | 790 | /* |
c9cf4a01 | 791 | * it may have some invalid PEBS bits |
39ef13a4 | 792 | */ |
c9cf4a01 | 793 | if (p4_config_pebs_has(event->attr.config, P4_PEBS_CONFIG_ENABLE)) |
39ef13a4 | 794 | return -EINVAL; |
c9cf4a01 | 795 | |
39ef13a4 | 796 | v = p4_config_unpack_metric(event->attr.config); |
c9cf4a01 | 797 | if (v >= ARRAY_SIZE(p4_pebs_bind_map)) |
39ef13a4 | 798 | return -EINVAL; |
39ef13a4 CG |
799 | |
800 | return 0; | |
801 | } | |
802 | ||
b4cdc5c2 | 803 | static int p4_hw_config(struct perf_event *event) |
a072738e | 804 | { |
137351e0 CG |
805 | int cpu = get_cpu(); |
806 | int rc = 0; | |
d814f301 | 807 | u32 escr, cccr; |
a072738e CG |
808 | |
809 | /* | |
810 | * the reason we use cpu that early is that: if we get scheduled | |
811 | * first time on the same cpu -- we will not need swap thread | |
812 | * specific flags in config (and will save some cpu cycles) | |
813 | */ | |
814 | ||
d814f301 | 815 | cccr = p4_default_cccr_conf(cpu); |
b4cdc5c2 PZ |
816 | escr = p4_default_escr_conf(cpu, event->attr.exclude_kernel, |
817 | event->attr.exclude_user); | |
818 | event->hw.config = p4_config_pack_escr(escr) | | |
819 | p4_config_pack_cccr(cccr); | |
a072738e | 820 | |
cb7d6b50 | 821 | if (p4_ht_active() && p4_ht_thread(cpu)) |
b4cdc5c2 PZ |
822 | event->hw.config = p4_set_ht_bit(event->hw.config); |
823 | ||
de902d96 | 824 | if (event->attr.type == PERF_TYPE_RAW) { |
d038b12c SE |
825 | struct p4_event_bind *bind; |
826 | unsigned int esel; | |
c9cf4a01 CG |
827 | /* |
828 | * Clear bits we reserve to be managed by kernel itself | |
829 | * and never allowed from a user space | |
830 | */ | |
831 | event->attr.config &= P4_CONFIG_MASK; | |
832 | ||
39ef13a4 CG |
833 | rc = p4_validate_raw_event(event); |
834 | if (rc) | |
c7993165 | 835 | goto out; |
c7993165 | 836 | |
de902d96 | 837 | /* |
39ef13a4 CG |
838 | * Note that for RAW events we allow user to use P4_CCCR_RESERVED |
839 | * bits since we keep additional info here (for cache events and etc) | |
de902d96 | 840 | */ |
c9cf4a01 | 841 | event->hw.config |= event->attr.config; |
d038b12c SE |
842 | bind = p4_config_get_bind(event->attr.config); |
843 | if (!bind) { | |
844 | rc = -EINVAL; | |
845 | goto out; | |
846 | } | |
847 | esel = P4_OPCODE_ESEL(bind->opcode); | |
848 | event->hw.config |= p4_config_pack_cccr(P4_CCCR_ESEL(esel)); | |
de902d96 | 849 | } |
cb7d6b50 | 850 | |
137351e0 | 851 | rc = x86_setup_perfctr(event); |
c7993165 | 852 | out: |
137351e0 | 853 | put_cpu(); |
137351e0 | 854 | return rc; |
a072738e CG |
855 | } |
856 | ||
0db1a7bc | 857 | static inline int p4_pmu_clear_cccr_ovf(struct hw_perf_event *hwc) |
a072738e | 858 | { |
047a3772 | 859 | u64 v; |
a072738e | 860 | |
047a3772 | 861 | /* an official way for overflow indication */ |
73d6e522 | 862 | rdmsrl(hwc->config_base, v); |
047a3772 | 863 | if (v & P4_CCCR_OVF) { |
73d6e522 | 864 | wrmsrl(hwc->config_base, v & ~P4_CCCR_OVF); |
047a3772 | 865 | return 1; |
a072738e | 866 | } |
0db1a7bc | 867 | |
7d44ec19 CG |
868 | /* |
869 | * In some circumstances the overflow might issue an NMI but did | |
870 | * not set P4_CCCR_OVF bit. Because a counter holds a negative value | |
871 | * we simply check for high bit being set, if it's cleared it means | |
872 | * the counter has reached zero value and continued counting before | |
873 | * real NMI signal was received: | |
874 | */ | |
242214f9 | 875 | rdmsrl(hwc->event_base, v); |
7d44ec19 | 876 | if (!(v & ARCH_P4_UNFLAGGED_BIT)) |
047a3772 CG |
877 | return 1; |
878 | ||
879 | return 0; | |
a072738e CG |
880 | } |
881 | ||
39ef13a4 CG |
882 | static void p4_pmu_disable_pebs(void) |
883 | { | |
884 | /* | |
885 | * FIXME | |
886 | * | |
887 | * It's still allowed that two threads setup same cache | |
888 | * events so we can't simply clear metrics until we knew | |
0d2eb44f | 889 | * no one is depending on us, so we need kind of counter |
39ef13a4 CG |
890 | * for "ReplayEvent" users. |
891 | * | |
892 | * What is more complex -- RAW events, if user (for some | |
893 | * reason) will pass some cache event metric with improper | |
894 | * event opcode -- it's fine from hardware point of view | |
0d2eb44f | 895 | * but completely nonsense from "meaning" of such action. |
39ef13a4 CG |
896 | * |
897 | * So at moment let leave metrics turned on forever -- it's | |
898 | * ok for now but need to be revisited! | |
899 | * | |
900 | * (void)checking_wrmsrl(MSR_IA32_PEBS_ENABLE, (u64)0); | |
901 | * (void)checking_wrmsrl(MSR_P4_PEBS_MATRIX_VERT, (u64)0); | |
902 | */ | |
903 | } | |
904 | ||
a072738e CG |
905 | static inline void p4_pmu_disable_event(struct perf_event *event) |
906 | { | |
907 | struct hw_perf_event *hwc = &event->hw; | |
908 | ||
909 | /* | |
910 | * If event gets disabled while counter is in overflowed | |
911 | * state we need to clear P4_CCCR_OVF, otherwise interrupt get | |
912 | * asserted again and again | |
913 | */ | |
73d6e522 | 914 | (void)checking_wrmsrl(hwc->config_base, |
a072738e | 915 | (u64)(p4_config_unpack_cccr(hwc->config)) & |
d814f301 | 916 | ~P4_CCCR_ENABLE & ~P4_CCCR_OVF & ~P4_CCCR_RESERVED); |
a072738e CG |
917 | } |
918 | ||
919 | static void p4_pmu_disable_all(void) | |
920 | { | |
921 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); | |
922 | int idx; | |
923 | ||
948b1bb8 | 924 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { |
a072738e CG |
925 | struct perf_event *event = cpuc->events[idx]; |
926 | if (!test_bit(idx, cpuc->active_mask)) | |
927 | continue; | |
928 | p4_pmu_disable_event(event); | |
929 | } | |
39ef13a4 CG |
930 | |
931 | p4_pmu_disable_pebs(); | |
932 | } | |
933 | ||
934 | /* configuration must be valid */ | |
935 | static void p4_pmu_enable_pebs(u64 config) | |
936 | { | |
937 | struct p4_pebs_bind *bind; | |
938 | unsigned int idx; | |
939 | ||
940 | BUILD_BUG_ON(P4_PEBS_METRIC__max > P4_PEBS_CONFIG_METRIC_MASK); | |
941 | ||
942 | idx = p4_config_unpack_metric(config); | |
943 | if (idx == P4_PEBS_METRIC__none) | |
944 | return; | |
945 | ||
946 | bind = &p4_pebs_bind_map[idx]; | |
947 | ||
948 | (void)checking_wrmsrl(MSR_IA32_PEBS_ENABLE, (u64)bind->metric_pebs); | |
949 | (void)checking_wrmsrl(MSR_P4_PEBS_MATRIX_VERT, (u64)bind->metric_vert); | |
a072738e CG |
950 | } |
951 | ||
952 | static void p4_pmu_enable_event(struct perf_event *event) | |
953 | { | |
954 | struct hw_perf_event *hwc = &event->hw; | |
955 | int thread = p4_ht_config_thread(hwc->config); | |
956 | u64 escr_conf = p4_config_unpack_escr(p4_clear_ht_bit(hwc->config)); | |
d814f301 | 957 | unsigned int idx = p4_config_unpack_event(hwc->config); |
d814f301 | 958 | struct p4_event_bind *bind; |
d814f301 | 959 | u64 escr_addr, cccr; |
a072738e | 960 | |
d814f301 CG |
961 | bind = &p4_event_bind_map[idx]; |
962 | escr_addr = (u64)bind->escr_msr[thread]; | |
a072738e CG |
963 | |
964 | /* | |
965 | * - we dont support cascaded counters yet | |
966 | * - and counter 1 is broken (erratum) | |
967 | */ | |
968 | WARN_ON_ONCE(p4_is_event_cascaded(hwc->config)); | |
969 | WARN_ON_ONCE(hwc->idx == 1); | |
970 | ||
d814f301 CG |
971 | /* we need a real Event value */ |
972 | escr_conf &= ~P4_ESCR_EVENT_MASK; | |
973 | escr_conf |= P4_ESCR_EVENT(P4_OPCODE_EVNT(bind->opcode)); | |
974 | ||
975 | cccr = p4_config_unpack_cccr(hwc->config); | |
976 | ||
977 | /* | |
39ef13a4 CG |
978 | * it could be Cache event so we need to write metrics |
979 | * into additional MSRs | |
d814f301 | 980 | */ |
39ef13a4 | 981 | p4_pmu_enable_pebs(hwc->config); |
d814f301 CG |
982 | |
983 | (void)checking_wrmsrl(escr_addr, escr_conf); | |
73d6e522 | 984 | (void)checking_wrmsrl(hwc->config_base, |
d814f301 | 985 | (cccr & ~P4_CCCR_RESERVED) | P4_CCCR_ENABLE); |
a072738e CG |
986 | } |
987 | ||
11164cd4 | 988 | static void p4_pmu_enable_all(int added) |
a072738e CG |
989 | { |
990 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); | |
991 | int idx; | |
992 | ||
948b1bb8 | 993 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { |
a072738e CG |
994 | struct perf_event *event = cpuc->events[idx]; |
995 | if (!test_bit(idx, cpuc->active_mask)) | |
996 | continue; | |
997 | p4_pmu_enable_event(event); | |
998 | } | |
999 | } | |
1000 | ||
1001 | static int p4_pmu_handle_irq(struct pt_regs *regs) | |
1002 | { | |
1003 | struct perf_sample_data data; | |
1004 | struct cpu_hw_events *cpuc; | |
1005 | struct perf_event *event; | |
1006 | struct hw_perf_event *hwc; | |
1007 | int idx, handled = 0; | |
1008 | u64 val; | |
1009 | ||
103b3934 | 1010 | perf_sample_data_init(&data, 0); |
a072738e CG |
1011 | |
1012 | cpuc = &__get_cpu_var(cpu_hw_events); | |
1013 | ||
948b1bb8 | 1014 | for (idx = 0; idx < x86_pmu.num_counters; idx++) { |
1c250d70 | 1015 | int overflow; |
a072738e | 1016 | |
03e22198 CG |
1017 | if (!test_bit(idx, cpuc->active_mask)) { |
1018 | /* catch in-flight IRQs */ | |
1019 | if (__test_and_clear_bit(idx, cpuc->running)) | |
1020 | handled++; | |
a072738e | 1021 | continue; |
03e22198 | 1022 | } |
a072738e CG |
1023 | |
1024 | event = cpuc->events[idx]; | |
1025 | hwc = &event->hw; | |
1026 | ||
1027 | WARN_ON_ONCE(hwc->idx != idx); | |
1028 | ||
0db1a7bc | 1029 | /* it might be unflagged overflow */ |
1c250d70 | 1030 | overflow = p4_pmu_clear_cccr_ovf(hwc); |
a072738e CG |
1031 | |
1032 | val = x86_perf_event_update(event); | |
1c250d70 | 1033 | if (!overflow && (val & (1ULL << (x86_pmu.cntval_bits - 1)))) |
a072738e CG |
1034 | continue; |
1035 | ||
1c250d70 CG |
1036 | handled += overflow; |
1037 | ||
0db1a7bc CG |
1038 | /* event overflow for sure */ |
1039 | data.period = event->hw.last_period; | |
a072738e CG |
1040 | |
1041 | if (!x86_perf_event_set_period(event)) | |
1042 | continue; | |
a8b0ca17 | 1043 | if (perf_event_overflow(event, &data, regs)) |
1ea5a6af | 1044 | x86_pmu_stop(event, 0); |
a072738e CG |
1045 | } |
1046 | ||
2bce5dac | 1047 | if (handled) |
a072738e | 1048 | inc_irq_stat(apic_perf_irqs); |
2bce5dac DZ |
1049 | |
1050 | /* | |
1051 | * When dealing with the unmasking of the LVTPC on P4 perf hw, it has | |
1052 | * been observed that the OVF bit flag has to be cleared first _before_ | |
1053 | * the LVTPC can be unmasked. | |
1054 | * | |
1055 | * The reason is the NMI line will continue to be asserted while the OVF | |
1056 | * bit is set. This causes a second NMI to generate if the LVTPC is | |
1057 | * unmasked before the OVF bit is cleared, leading to unknown NMI | |
1058 | * messages. | |
1059 | */ | |
1060 | apic_write(APIC_LVTPC, APIC_DM_NMI); | |
a072738e | 1061 | |
de725dec | 1062 | return handled; |
a072738e CG |
1063 | } |
1064 | ||
1065 | /* | |
1066 | * swap thread specific fields according to a thread | |
1067 | * we are going to run on | |
1068 | */ | |
1069 | static void p4_pmu_swap_config_ts(struct hw_perf_event *hwc, int cpu) | |
1070 | { | |
1071 | u32 escr, cccr; | |
1072 | ||
1073 | /* | |
1074 | * we either lucky and continue on same cpu or no HT support | |
1075 | */ | |
1076 | if (!p4_should_swap_ts(hwc->config, cpu)) | |
1077 | return; | |
1078 | ||
1079 | /* | |
1080 | * the event is migrated from an another logical | |
1081 | * cpu, so we need to swap thread specific flags | |
1082 | */ | |
1083 | ||
1084 | escr = p4_config_unpack_escr(hwc->config); | |
1085 | cccr = p4_config_unpack_cccr(hwc->config); | |
1086 | ||
1087 | if (p4_ht_thread(cpu)) { | |
1088 | cccr &= ~P4_CCCR_OVF_PMI_T0; | |
1089 | cccr |= P4_CCCR_OVF_PMI_T1; | |
d814f301 CG |
1090 | if (escr & P4_ESCR_T0_OS) { |
1091 | escr &= ~P4_ESCR_T0_OS; | |
1092 | escr |= P4_ESCR_T1_OS; | |
a072738e | 1093 | } |
d814f301 CG |
1094 | if (escr & P4_ESCR_T0_USR) { |
1095 | escr &= ~P4_ESCR_T0_USR; | |
1096 | escr |= P4_ESCR_T1_USR; | |
a072738e CG |
1097 | } |
1098 | hwc->config = p4_config_pack_escr(escr); | |
1099 | hwc->config |= p4_config_pack_cccr(cccr); | |
1100 | hwc->config |= P4_CONFIG_HT; | |
1101 | } else { | |
1102 | cccr &= ~P4_CCCR_OVF_PMI_T1; | |
1103 | cccr |= P4_CCCR_OVF_PMI_T0; | |
d814f301 CG |
1104 | if (escr & P4_ESCR_T1_OS) { |
1105 | escr &= ~P4_ESCR_T1_OS; | |
1106 | escr |= P4_ESCR_T0_OS; | |
a072738e | 1107 | } |
d814f301 CG |
1108 | if (escr & P4_ESCR_T1_USR) { |
1109 | escr &= ~P4_ESCR_T1_USR; | |
1110 | escr |= P4_ESCR_T0_USR; | |
a072738e CG |
1111 | } |
1112 | hwc->config = p4_config_pack_escr(escr); | |
1113 | hwc->config |= p4_config_pack_cccr(cccr); | |
1114 | hwc->config &= ~P4_CONFIG_HT; | |
1115 | } | |
1116 | } | |
1117 | ||
72001990 CG |
1118 | /* |
1119 | * ESCR address hashing is tricky, ESCRs are not sequential | |
623aab89 | 1120 | * in memory but all starts from MSR_P4_BSU_ESCR0 (0x03a0) and |
72001990 CG |
1121 | * the metric between any ESCRs is laid in range [0xa0,0xe1] |
1122 | * | |
1123 | * so we make ~70% filled hashtable | |
1124 | */ | |
1125 | ||
1126 | #define P4_ESCR_MSR_BASE 0x000003a0 | |
1127 | #define P4_ESCR_MSR_MAX 0x000003e1 | |
1128 | #define P4_ESCR_MSR_TABLE_SIZE (P4_ESCR_MSR_MAX - P4_ESCR_MSR_BASE + 1) | |
1129 | #define P4_ESCR_MSR_IDX(msr) (msr - P4_ESCR_MSR_BASE) | |
1130 | #define P4_ESCR_MSR_TABLE_ENTRY(msr) [P4_ESCR_MSR_IDX(msr)] = msr | |
1131 | ||
1132 | static const unsigned int p4_escr_table[P4_ESCR_MSR_TABLE_SIZE] = { | |
1133 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_ALF_ESCR0), | |
1134 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_ALF_ESCR1), | |
1135 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_BPU_ESCR0), | |
1136 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_BPU_ESCR1), | |
1137 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_BSU_ESCR0), | |
1138 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_BSU_ESCR1), | |
1139 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_CRU_ESCR0), | |
1140 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_CRU_ESCR1), | |
1141 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_CRU_ESCR2), | |
1142 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_CRU_ESCR3), | |
1143 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_CRU_ESCR4), | |
1144 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_CRU_ESCR5), | |
1145 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_DAC_ESCR0), | |
1146 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_DAC_ESCR1), | |
1147 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_FIRM_ESCR0), | |
1148 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_FIRM_ESCR1), | |
1149 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_FLAME_ESCR0), | |
1150 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_FLAME_ESCR1), | |
1151 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_FSB_ESCR0), | |
1152 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_FSB_ESCR1), | |
1153 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_IQ_ESCR0), | |
1154 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_IQ_ESCR1), | |
1155 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_IS_ESCR0), | |
1156 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_IS_ESCR1), | |
1157 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_ITLB_ESCR0), | |
1158 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_ITLB_ESCR1), | |
1159 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_IX_ESCR0), | |
1160 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_IX_ESCR1), | |
1161 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_MOB_ESCR0), | |
1162 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_MOB_ESCR1), | |
1163 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_MS_ESCR0), | |
1164 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_MS_ESCR1), | |
1165 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_PMH_ESCR0), | |
1166 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_PMH_ESCR1), | |
1167 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_RAT_ESCR0), | |
1168 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_RAT_ESCR1), | |
1169 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_SAAT_ESCR0), | |
1170 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_SAAT_ESCR1), | |
1171 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_SSU_ESCR0), | |
1172 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_SSU_ESCR1), | |
1173 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_TBPU_ESCR0), | |
1174 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_TBPU_ESCR1), | |
1175 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_TC_ESCR0), | |
1176 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_TC_ESCR1), | |
1177 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_U2L_ESCR0), | |
1178 | P4_ESCR_MSR_TABLE_ENTRY(MSR_P4_U2L_ESCR1), | |
a072738e CG |
1179 | }; |
1180 | ||
1181 | static int p4_get_escr_idx(unsigned int addr) | |
1182 | { | |
72001990 | 1183 | unsigned int idx = P4_ESCR_MSR_IDX(addr); |
a072738e | 1184 | |
623aab89 CG |
1185 | if (unlikely(idx >= P4_ESCR_MSR_TABLE_SIZE || |
1186 | !p4_escr_table[idx] || | |
1187 | p4_escr_table[idx] != addr)) { | |
72001990 CG |
1188 | WARN_ONCE(1, "P4 PMU: Wrong address passed: %x\n", addr); |
1189 | return -1; | |
a072738e CG |
1190 | } |
1191 | ||
72001990 | 1192 | return idx; |
a072738e CG |
1193 | } |
1194 | ||
d814f301 CG |
1195 | static int p4_next_cntr(int thread, unsigned long *used_mask, |
1196 | struct p4_event_bind *bind) | |
1197 | { | |
1ff3d7d7 | 1198 | int i, j; |
d814f301 CG |
1199 | |
1200 | for (i = 0; i < P4_CNTR_LIMIT; i++) { | |
1ff3d7d7 CG |
1201 | j = bind->cntr[thread][i]; |
1202 | if (j != -1 && !test_bit(j, used_mask)) | |
d814f301 CG |
1203 | return j; |
1204 | } | |
1205 | ||
1206 | return -1; | |
1207 | } | |
1208 | ||
a072738e CG |
1209 | static int p4_pmu_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign) |
1210 | { | |
1211 | unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)]; | |
72001990 | 1212 | unsigned long escr_mask[BITS_TO_LONGS(P4_ESCR_MSR_TABLE_SIZE)]; |
9d36dfcf | 1213 | int cpu = smp_processor_id(); |
d814f301 CG |
1214 | struct hw_perf_event *hwc; |
1215 | struct p4_event_bind *bind; | |
1216 | unsigned int i, thread, num; | |
1217 | int cntr_idx, escr_idx; | |
f9129870 CG |
1218 | u64 config_alias; |
1219 | int pass; | |
a072738e CG |
1220 | |
1221 | bitmap_zero(used_mask, X86_PMC_IDX_MAX); | |
72001990 | 1222 | bitmap_zero(escr_mask, P4_ESCR_MSR_TABLE_SIZE); |
a072738e | 1223 | |
a072738e | 1224 | for (i = 0, num = n; i < n; i++, num--) { |
d814f301 | 1225 | |
a072738e | 1226 | hwc = &cpuc->event_list[i]->hw; |
a072738e | 1227 | thread = p4_ht_thread(cpu); |
f9129870 CG |
1228 | pass = 0; |
1229 | ||
1230 | again: | |
1231 | /* | |
1232 | * Aliases are swappable so we may hit circular | |
1233 | * lock if both original config and alias need | |
1234 | * resources (MSR registers) which already busy. | |
1235 | */ | |
1236 | if (pass > 2) | |
1237 | goto done; | |
1238 | ||
d814f301 CG |
1239 | bind = p4_config_get_bind(hwc->config); |
1240 | escr_idx = p4_get_escr_idx(bind->escr_msr[thread]); | |
72001990 CG |
1241 | if (unlikely(escr_idx == -1)) |
1242 | goto done; | |
a072738e | 1243 | |
a072738e | 1244 | if (hwc->idx != -1 && !p4_should_swap_ts(hwc->config, cpu)) { |
d814f301 | 1245 | cntr_idx = hwc->idx; |
a072738e CG |
1246 | if (assign) |
1247 | assign[i] = hwc->idx; | |
a072738e CG |
1248 | goto reserve; |
1249 | } | |
1250 | ||
d814f301 | 1251 | cntr_idx = p4_next_cntr(thread, used_mask, bind); |
f9129870 CG |
1252 | if (cntr_idx == -1 || test_bit(escr_idx, escr_mask)) { |
1253 | /* | |
1254 | * Probably an event alias is still available. | |
1255 | */ | |
1256 | config_alias = p4_get_alias_event(hwc->config); | |
1257 | if (!config_alias) | |
1258 | goto done; | |
1259 | hwc->config = config_alias; | |
1260 | pass++; | |
1261 | goto again; | |
1262 | } | |
a072738e | 1263 | |
a072738e | 1264 | p4_pmu_swap_config_ts(hwc, cpu); |
d814f301 CG |
1265 | if (assign) |
1266 | assign[i] = cntr_idx; | |
a072738e | 1267 | reserve: |
d814f301 | 1268 | set_bit(cntr_idx, used_mask); |
a072738e CG |
1269 | set_bit(escr_idx, escr_mask); |
1270 | } | |
1271 | ||
1272 | done: | |
1273 | return num ? -ENOSPC : 0; | |
1274 | } | |
1275 | ||
caaa8be3 | 1276 | static __initconst const struct x86_pmu p4_pmu = { |
a072738e CG |
1277 | .name = "Netburst P4/Xeon", |
1278 | .handle_irq = p4_pmu_handle_irq, | |
1279 | .disable_all = p4_pmu_disable_all, | |
1280 | .enable_all = p4_pmu_enable_all, | |
1281 | .enable = p4_pmu_enable_event, | |
1282 | .disable = p4_pmu_disable_event, | |
1283 | .eventsel = MSR_P4_BPU_CCCR0, | |
1284 | .perfctr = MSR_P4_BPU_PERFCTR0, | |
1285 | .event_map = p4_pmu_event_map, | |
d814f301 | 1286 | .max_events = ARRAY_SIZE(p4_general_events), |
a072738e CG |
1287 | .get_event_constraints = x86_get_event_constraints, |
1288 | /* | |
1289 | * IF HT disabled we may need to use all | |
1290 | * ARCH_P4_MAX_CCCR counters simulaneously | |
1291 | * though leave it restricted at moment assuming | |
1292 | * HT is on | |
1293 | */ | |
948b1bb8 | 1294 | .num_counters = ARCH_P4_MAX_CCCR, |
a072738e | 1295 | .apic = 1, |
047a3772 CG |
1296 | .cntval_bits = ARCH_P4_CNTRVAL_BITS, |
1297 | .cntval_mask = ARCH_P4_CNTRVAL_MASK, | |
1298 | .max_period = (1ULL << (ARCH_P4_CNTRVAL_BITS - 1)) - 1, | |
a072738e CG |
1299 | .hw_config = p4_hw_config, |
1300 | .schedule_events = p4_pmu_schedule_events, | |
68aa00ac CG |
1301 | /* |
1302 | * This handles erratum N15 in intel doc 249199-029, | |
1303 | * the counter may not be updated correctly on write | |
1304 | * so we need a second write operation to do the trick | |
1305 | * (the official workaround didn't work) | |
1306 | * | |
1307 | * the former idea is taken from OProfile code | |
1308 | */ | |
1309 | .perfctr_second_write = 1, | |
a072738e CG |
1310 | }; |
1311 | ||
1312 | static __init int p4_pmu_init(void) | |
1313 | { | |
1314 | unsigned int low, high; | |
1315 | ||
fa7b6947 | 1316 | /* If we get stripped -- indexing fails */ |
a072738e CG |
1317 | BUILD_BUG_ON(ARCH_P4_MAX_CCCR > X86_PMC_MAX_GENERIC); |
1318 | ||
1319 | rdmsr(MSR_IA32_MISC_ENABLE, low, high); | |
1320 | if (!(low & (1 << 7))) { | |
1321 | pr_cont("unsupported Netburst CPU model %d ", | |
1322 | boot_cpu_data.x86_model); | |
1323 | return -ENODEV; | |
1324 | } | |
1325 | ||
cb7d6b50 | 1326 | memcpy(hw_cache_event_ids, p4_hw_cache_event_ids, |
d814f301 | 1327 | sizeof(hw_cache_event_ids)); |
cb7d6b50 | 1328 | |
a072738e CG |
1329 | pr_cont("Netburst events, "); |
1330 | ||
1331 | x86_pmu = p4_pmu; | |
1332 | ||
1333 | return 0; | |
1334 | } | |
1335 | ||
1336 | #endif /* CONFIG_CPU_SUP_INTEL */ |