Commit | Line | Data |
---|---|---|
e05b9b9e ME |
1 | /* |
2 | * Performance counter support for POWER8 processors. | |
3 | * | |
4 | * Copyright 2009 Paul Mackerras, IBM Corporation. | |
5 | * Copyright 2013 Michael Ellerman, IBM Corporation. | |
6 | * | |
7 | * This program is free software; you can redistribute it and/or | |
8 | * modify it under the terms of the GNU General Public License | |
9 | * as published by the Free Software Foundation; either version | |
10 | * 2 of the License, or (at your option) any later version. | |
11 | */ | |
12 | ||
c2e37a26 ME |
13 | #define pr_fmt(fmt) "power8-pmu: " fmt |
14 | ||
e05b9b9e ME |
15 | #include <linux/kernel.h> |
16 | #include <linux/perf_event.h> | |
17 | #include <asm/firmware.h> | |
18 | ||
19 | ||
20 | /* | |
21 | * Some power8 event codes. | |
22 | */ | |
23 | #define PM_CYC 0x0001e | |
24 | #define PM_GCT_NOSLOT_CYC 0x100f8 | |
25 | #define PM_CMPLU_STALL 0x4000a | |
26 | #define PM_INST_CMPL 0x00002 | |
27 | #define PM_BRU_FIN 0x10068 | |
28 | #define PM_BR_MPRED_CMPL 0x400f6 | |
29 | ||
2fdd313f ME |
30 | /* All L1 D cache load references counted at finish, gated by reject */ |
31 | #define PM_LD_REF_L1 0x100ee | |
32 | /* Load Missed L1 */ | |
33 | #define PM_LD_MISS_L1 0x3e054 | |
34 | /* Store Missed L1 */ | |
35 | #define PM_ST_MISS_L1 0x300f0 | |
36 | /* L1 cache data prefetches */ | |
37 | #define PM_L1_PREF 0x0d8b8 | |
38 | /* Instruction fetches from L1 */ | |
39 | #define PM_INST_FROM_L1 0x04080 | |
40 | /* Demand iCache Miss */ | |
41 | #define PM_L1_ICACHE_MISS 0x200fd | |
42 | /* Instruction Demand sectors wriittent into IL1 */ | |
43 | #define PM_L1_DEMAND_WRITE 0x0408c | |
44 | /* Instruction prefetch written into IL1 */ | |
45 | #define PM_IC_PREF_WRITE 0x0408e | |
46 | /* The data cache was reloaded from local core's L3 due to a demand load */ | |
47 | #define PM_DATA_FROM_L3 0x4c042 | |
48 | /* Demand LD - L3 Miss (not L2 hit and not L3 hit) */ | |
49 | #define PM_DATA_FROM_L3MISS 0x300fe | |
50 | /* All successful D-side store dispatches for this thread */ | |
51 | #define PM_L2_ST 0x17080 | |
52 | /* All successful D-side store dispatches for this thread that were L2 Miss */ | |
53 | #define PM_L2_ST_MISS 0x17082 | |
54 | /* Total HW L3 prefetches(Load+store) */ | |
55 | #define PM_L3_PREF_ALL 0x4e052 | |
56 | /* Data PTEG reload */ | |
57 | #define PM_DTLB_MISS 0x300fc | |
58 | /* ITLB Reloaded */ | |
59 | #define PM_ITLB_MISS 0x400fc | |
60 | ||
e05b9b9e ME |
61 | |
62 | /* | |
63 | * Raw event encoding for POWER8: | |
64 | * | |
65 | * 60 56 52 48 44 40 36 32 | |
66 | * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | | |
4df48999 ME |
67 | * | [ thresh_cmp ] [ thresh_ctl ] |
68 | * | | | |
69 | * *- EBB (Linux) thresh start/stop OR FAB match -* | |
e05b9b9e ME |
70 | * |
71 | * 28 24 20 16 12 8 4 0 | |
72 | * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | | |
73 | * [ ] [ sample ] [cache] [ pmc ] [unit ] c m [ pmcxsel ] | |
74 | * | | | | | | |
75 | * | | | | *- mark | |
76 | * | | *- L1/L2/L3 cache_sel | | |
77 | * | | | | |
78 | * | *- sampling mode for marked events *- combine | |
79 | * | | |
80 | * *- thresh_sel | |
81 | * | |
82 | * Below uses IBM bit numbering. | |
83 | * | |
84 | * MMCR1[x:y] = unit (PMCxUNIT) | |
85 | * MMCR1[x] = combine (PMCxCOMB) | |
86 | * | |
87 | * if pmc == 3 and unit == 0 and pmcxsel[0:6] == 0b0101011 | |
88 | * # PM_MRK_FAB_RSP_MATCH | |
89 | * MMCR1[20:27] = thresh_ctl (FAB_CRESP_MATCH / FAB_TYPE_MATCH) | |
90 | * else if pmc == 4 and unit == 0xf and pmcxsel[0:6] == 0b0101001 | |
91 | * # PM_MRK_FAB_RSP_MATCH_CYC | |
92 | * MMCR1[20:27] = thresh_ctl (FAB_CRESP_MATCH / FAB_TYPE_MATCH) | |
93 | * else | |
94 | * MMCRA[48:55] = thresh_ctl (THRESH START/END) | |
95 | * | |
96 | * if thresh_sel: | |
97 | * MMCRA[45:47] = thresh_sel | |
98 | * | |
99 | * if thresh_cmp: | |
100 | * MMCRA[22:24] = thresh_cmp[0:2] | |
101 | * MMCRA[25:31] = thresh_cmp[3:9] | |
102 | * | |
103 | * if unit == 6 or unit == 7 | |
104 | * MMCRC[53:55] = cache_sel[1:3] (L2EVENT_SEL) | |
105 | * else if unit == 8 or unit == 9: | |
106 | * if cache_sel[0] == 0: # L3 bank | |
107 | * MMCRC[47:49] = cache_sel[1:3] (L3EVENT_SEL0) | |
108 | * else if cache_sel[0] == 1: | |
109 | * MMCRC[50:51] = cache_sel[2:3] (L3EVENT_SEL1) | |
110 | * else if cache_sel[1]: # L1 event | |
111 | * MMCR1[16] = cache_sel[2] | |
112 | * MMCR1[17] = cache_sel[3] | |
113 | * | |
114 | * if mark: | |
115 | * MMCRA[63] = 1 (SAMPLE_ENABLE) | |
116 | * MMCRA[57:59] = sample[0:2] (RAND_SAMP_ELIG) | |
117 | * MMCRA[61:62] = sample[3:4] (RAND_SAMP_MODE) | |
118 | * | |
119 | */ | |
120 | ||
4df48999 | 121 | #define EVENT_EBB_MASK 1ull |
fb568d76 | 122 | #define EVENT_EBB_SHIFT PERF_EVENT_CONFIG_EBB_SHIFT |
e05b9b9e ME |
123 | #define EVENT_THR_CMP_SHIFT 40 /* Threshold CMP value */ |
124 | #define EVENT_THR_CMP_MASK 0x3ff | |
125 | #define EVENT_THR_CTL_SHIFT 32 /* Threshold control value (start/stop) */ | |
126 | #define EVENT_THR_CTL_MASK 0xffull | |
127 | #define EVENT_THR_SEL_SHIFT 29 /* Threshold select value */ | |
128 | #define EVENT_THR_SEL_MASK 0x7 | |
129 | #define EVENT_THRESH_SHIFT 29 /* All threshold bits */ | |
130 | #define EVENT_THRESH_MASK 0x1fffffull | |
131 | #define EVENT_SAMPLE_SHIFT 24 /* Sampling mode & eligibility */ | |
132 | #define EVENT_SAMPLE_MASK 0x1f | |
133 | #define EVENT_CACHE_SEL_SHIFT 20 /* L2/L3 cache select */ | |
134 | #define EVENT_CACHE_SEL_MASK 0xf | |
135 | #define EVENT_IS_L1 (4 << EVENT_CACHE_SEL_SHIFT) | |
136 | #define EVENT_PMC_SHIFT 16 /* PMC number (1-based) */ | |
137 | #define EVENT_PMC_MASK 0xf | |
138 | #define EVENT_UNIT_SHIFT 12 /* Unit */ | |
139 | #define EVENT_UNIT_MASK 0xf | |
140 | #define EVENT_COMBINE_SHIFT 11 /* Combine bit */ | |
141 | #define EVENT_COMBINE_MASK 0x1 | |
142 | #define EVENT_MARKED_SHIFT 8 /* Marked bit */ | |
143 | #define EVENT_MARKED_MASK 0x1 | |
144 | #define EVENT_IS_MARKED (EVENT_MARKED_MASK << EVENT_MARKED_SHIFT) | |
145 | #define EVENT_PSEL_MASK 0xff /* PMCxSEL value */ | |
146 | ||
d8bec4c9 ME |
147 | #define EVENT_VALID_MASK \ |
148 | ((EVENT_THRESH_MASK << EVENT_THRESH_SHIFT) | \ | |
149 | (EVENT_SAMPLE_MASK << EVENT_SAMPLE_SHIFT) | \ | |
150 | (EVENT_CACHE_SEL_MASK << EVENT_CACHE_SEL_SHIFT) | \ | |
151 | (EVENT_PMC_MASK << EVENT_PMC_SHIFT) | \ | |
152 | (EVENT_UNIT_MASK << EVENT_UNIT_SHIFT) | \ | |
153 | (EVENT_COMBINE_MASK << EVENT_COMBINE_SHIFT) | \ | |
154 | (EVENT_MARKED_MASK << EVENT_MARKED_SHIFT) | \ | |
fb568d76 | 155 | (EVENT_EBB_MASK << EVENT_EBB_SHIFT) | \ |
d8bec4c9 ME |
156 | EVENT_PSEL_MASK) |
157 | ||
b1113557 AK |
158 | /* MMCRA IFM bits - POWER8 */ |
159 | #define POWER8_MMCRA_IFM1 0x0000000040000000UL | |
160 | #define POWER8_MMCRA_IFM2 0x0000000080000000UL | |
161 | #define POWER8_MMCRA_IFM3 0x00000000C0000000UL | |
162 | ||
163 | #define ONLY_PLM \ | |
164 | (PERF_SAMPLE_BRANCH_USER |\ | |
165 | PERF_SAMPLE_BRANCH_KERNEL |\ | |
166 | PERF_SAMPLE_BRANCH_HV) | |
167 | ||
e05b9b9e ME |
168 | /* |
169 | * Layout of constraint bits: | |
170 | * | |
171 | * 60 56 52 48 44 40 36 32 | |
172 | * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | | |
173 | * [ fab_match ] [ thresh_cmp ] [ thresh_ctl ] [ ] | |
174 | * | | |
175 | * thresh_sel -* | |
176 | * | |
177 | * 28 24 20 16 12 8 4 0 | |
178 | * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | | |
4df48999 ME |
179 | * | [ ] [ sample ] [ ] [6] [5] [4] [3] [2] [1] |
180 | * EBB -* | | | |
181 | * | | Count of events for each PMC. | |
182 | * L1 I/D qualifier -* | p1, p2, p3, p4, p5, p6. | |
e05b9b9e ME |
183 | * nc - number of counters -* |
184 | * | |
185 | * The PMC fields P1..P6, and NC, are adder fields. As we accumulate constraints | |
186 | * we want the low bit of each field to be added to any existing value. | |
187 | * | |
188 | * Everything else is a value field. | |
189 | */ | |
190 | ||
191 | #define CNST_FAB_MATCH_VAL(v) (((v) & EVENT_THR_CTL_MASK) << 56) | |
192 | #define CNST_FAB_MATCH_MASK CNST_FAB_MATCH_VAL(EVENT_THR_CTL_MASK) | |
193 | ||
194 | /* We just throw all the threshold bits into the constraint */ | |
195 | #define CNST_THRESH_VAL(v) (((v) & EVENT_THRESH_MASK) << 32) | |
196 | #define CNST_THRESH_MASK CNST_THRESH_VAL(EVENT_THRESH_MASK) | |
197 | ||
4df48999 ME |
198 | #define CNST_EBB_VAL(v) (((v) & EVENT_EBB_MASK) << 24) |
199 | #define CNST_EBB_MASK CNST_EBB_VAL(EVENT_EBB_MASK) | |
200 | ||
e05b9b9e ME |
201 | #define CNST_L1_QUAL_VAL(v) (((v) & 3) << 22) |
202 | #define CNST_L1_QUAL_MASK CNST_L1_QUAL_VAL(3) | |
203 | ||
204 | #define CNST_SAMPLE_VAL(v) (((v) & EVENT_SAMPLE_MASK) << 16) | |
205 | #define CNST_SAMPLE_MASK CNST_SAMPLE_VAL(EVENT_SAMPLE_MASK) | |
206 | ||
207 | /* | |
208 | * For NC we are counting up to 4 events. This requires three bits, and we need | |
209 | * the fifth event to overflow and set the 4th bit. To achieve that we bias the | |
210 | * fields by 3 in test_adder. | |
211 | */ | |
212 | #define CNST_NC_SHIFT 12 | |
213 | #define CNST_NC_VAL (1 << CNST_NC_SHIFT) | |
214 | #define CNST_NC_MASK (8 << CNST_NC_SHIFT) | |
215 | #define POWER8_TEST_ADDER (3 << CNST_NC_SHIFT) | |
216 | ||
217 | /* | |
218 | * For the per-PMC fields we have two bits. The low bit is added, so if two | |
219 | * events ask for the same PMC the sum will overflow, setting the high bit, | |
220 | * indicating an error. So our mask sets the high bit. | |
221 | */ | |
222 | #define CNST_PMC_SHIFT(pmc) ((pmc - 1) * 2) | |
223 | #define CNST_PMC_VAL(pmc) (1 << CNST_PMC_SHIFT(pmc)) | |
224 | #define CNST_PMC_MASK(pmc) (2 << CNST_PMC_SHIFT(pmc)) | |
225 | ||
226 | /* Our add_fields is defined as: */ | |
227 | #define POWER8_ADD_FIELDS \ | |
228 | CNST_PMC_VAL(1) | CNST_PMC_VAL(2) | CNST_PMC_VAL(3) | \ | |
229 | CNST_PMC_VAL(4) | CNST_PMC_VAL(5) | CNST_PMC_VAL(6) | CNST_NC_VAL | |
230 | ||
231 | ||
232 | /* Bits in MMCR1 for POWER8 */ | |
233 | #define MMCR1_UNIT_SHIFT(pmc) (60 - (4 * ((pmc) - 1))) | |
234 | #define MMCR1_COMBINE_SHIFT(pmc) (35 - ((pmc) - 1)) | |
235 | #define MMCR1_PMCSEL_SHIFT(pmc) (24 - (((pmc) - 1)) * 8) | |
a53b27b3 | 236 | #define MMCR1_FAB_SHIFT 36 |
e05b9b9e ME |
237 | #define MMCR1_DC_QUAL_SHIFT 47 |
238 | #define MMCR1_IC_QUAL_SHIFT 46 | |
239 | ||
240 | /* Bits in MMCRA for POWER8 */ | |
241 | #define MMCRA_SAMP_MODE_SHIFT 1 | |
242 | #define MMCRA_SAMP_ELIG_SHIFT 4 | |
243 | #define MMCRA_THR_CTL_SHIFT 8 | |
244 | #define MMCRA_THR_SEL_SHIFT 16 | |
245 | #define MMCRA_THR_CMP_SHIFT 32 | |
246 | #define MMCRA_SDAR_MODE_TLB (1ull << 42) | |
247 | ||
248 | ||
249 | static inline bool event_is_fab_match(u64 event) | |
250 | { | |
251 | /* Only check pmc, unit and pmcxsel, ignore the edge bit (0) */ | |
252 | event &= 0xff0fe; | |
253 | ||
254 | /* PM_MRK_FAB_RSP_MATCH & PM_MRK_FAB_RSP_MATCH_CYC */ | |
255 | return (event == 0x30056 || event == 0x4f052); | |
256 | } | |
257 | ||
258 | static int power8_get_constraint(u64 event, unsigned long *maskp, unsigned long *valp) | |
259 | { | |
4df48999 | 260 | unsigned int unit, pmc, cache, ebb; |
e05b9b9e ME |
261 | unsigned long mask, value; |
262 | ||
263 | mask = value = 0; | |
264 | ||
d8bec4c9 ME |
265 | if (event & ~EVENT_VALID_MASK) |
266 | return -1; | |
267 | ||
4df48999 ME |
268 | pmc = (event >> EVENT_PMC_SHIFT) & EVENT_PMC_MASK; |
269 | unit = (event >> EVENT_UNIT_SHIFT) & EVENT_UNIT_MASK; | |
270 | cache = (event >> EVENT_CACHE_SEL_SHIFT) & EVENT_CACHE_SEL_MASK; | |
fb568d76 | 271 | ebb = (event >> EVENT_EBB_SHIFT) & EVENT_EBB_MASK; |
4df48999 | 272 | |
e05b9b9e | 273 | if (pmc) { |
7cbba630 ME |
274 | u64 base_event; |
275 | ||
e05b9b9e ME |
276 | if (pmc > 6) |
277 | return -1; | |
278 | ||
7cbba630 ME |
279 | /* Ignore Linux defined bits when checking event below */ |
280 | base_event = event & ~(EVENT_EBB_MASK << EVENT_EBB_SHIFT); | |
e05b9b9e | 281 | |
7cbba630 | 282 | if (pmc >= 5 && base_event != 0x500fa && base_event != 0x600f4) |
e05b9b9e | 283 | return -1; |
7cbba630 ME |
284 | |
285 | mask |= CNST_PMC_MASK(pmc); | |
286 | value |= CNST_PMC_VAL(pmc); | |
e05b9b9e ME |
287 | } |
288 | ||
289 | if (pmc <= 4) { | |
290 | /* | |
291 | * Add to number of counters in use. Note this includes events with | |
292 | * a PMC of 0 - they still need a PMC, it's just assigned later. | |
293 | * Don't count events on PMC 5 & 6, there is only one valid event | |
294 | * on each of those counters, and they are handled above. | |
295 | */ | |
296 | mask |= CNST_NC_MASK; | |
297 | value |= CNST_NC_VAL; | |
298 | } | |
299 | ||
300 | if (unit >= 6 && unit <= 9) { | |
301 | /* | |
302 | * L2/L3 events contain a cache selector field, which is | |
303 | * supposed to be programmed into MMCRC. However MMCRC is only | |
304 | * HV writable, and there is no API for guest kernels to modify | |
305 | * it. The solution is for the hypervisor to initialise the | |
306 | * field to zeroes, and for us to only ever allow events that | |
307 | * have a cache selector of zero. | |
308 | */ | |
309 | if (cache) | |
310 | return -1; | |
311 | ||
312 | } else if (event & EVENT_IS_L1) { | |
313 | mask |= CNST_L1_QUAL_MASK; | |
314 | value |= CNST_L1_QUAL_VAL(cache); | |
315 | } | |
316 | ||
317 | if (event & EVENT_IS_MARKED) { | |
318 | mask |= CNST_SAMPLE_MASK; | |
319 | value |= CNST_SAMPLE_VAL(event >> EVENT_SAMPLE_SHIFT); | |
320 | } | |
321 | ||
322 | /* | |
323 | * Special case for PM_MRK_FAB_RSP_MATCH and PM_MRK_FAB_RSP_MATCH_CYC, | |
324 | * the threshold control bits are used for the match value. | |
325 | */ | |
326 | if (event_is_fab_match(event)) { | |
327 | mask |= CNST_FAB_MATCH_MASK; | |
328 | value |= CNST_FAB_MATCH_VAL(event >> EVENT_THR_CTL_SHIFT); | |
329 | } else { | |
330 | /* | |
331 | * Check the mantissa upper two bits are not zero, unless the | |
332 | * exponent is also zero. See the THRESH_CMP_MANTISSA doc. | |
333 | */ | |
334 | unsigned int cmp, exp; | |
335 | ||
336 | cmp = (event >> EVENT_THR_CMP_SHIFT) & EVENT_THR_CMP_MASK; | |
337 | exp = cmp >> 7; | |
338 | ||
339 | if (exp && (cmp & 0x60) == 0) | |
340 | return -1; | |
341 | ||
342 | mask |= CNST_THRESH_MASK; | |
343 | value |= CNST_THRESH_VAL(event >> EVENT_THRESH_SHIFT); | |
344 | } | |
345 | ||
4df48999 ME |
346 | if (!pmc && ebb) |
347 | /* EBB events must specify the PMC */ | |
348 | return -1; | |
349 | ||
350 | /* | |
351 | * All events must agree on EBB, either all request it or none. | |
352 | * EBB events are pinned & exclusive, so this should never actually | |
353 | * hit, but we leave it as a fallback in case. | |
354 | */ | |
355 | mask |= CNST_EBB_VAL(ebb); | |
356 | value |= CNST_EBB_MASK; | |
357 | ||
e05b9b9e ME |
358 | *maskp = mask; |
359 | *valp = value; | |
360 | ||
361 | return 0; | |
362 | } | |
363 | ||
364 | static int power8_compute_mmcr(u64 event[], int n_ev, | |
365 | unsigned int hwc[], unsigned long mmcr[]) | |
366 | { | |
367 | unsigned long mmcra, mmcr1, unit, combine, psel, cache, val; | |
368 | unsigned int pmc, pmc_inuse; | |
369 | int i; | |
370 | ||
371 | pmc_inuse = 0; | |
372 | ||
373 | /* First pass to count resource use */ | |
374 | for (i = 0; i < n_ev; ++i) { | |
375 | pmc = (event[i] >> EVENT_PMC_SHIFT) & EVENT_PMC_MASK; | |
376 | if (pmc) | |
377 | pmc_inuse |= 1 << pmc; | |
378 | } | |
379 | ||
380 | /* In continous sampling mode, update SDAR on TLB miss */ | |
381 | mmcra = MMCRA_SDAR_MODE_TLB; | |
382 | mmcr1 = 0; | |
383 | ||
384 | /* Second pass: assign PMCs, set all MMCR1 fields */ | |
385 | for (i = 0; i < n_ev; ++i) { | |
386 | pmc = (event[i] >> EVENT_PMC_SHIFT) & EVENT_PMC_MASK; | |
387 | unit = (event[i] >> EVENT_UNIT_SHIFT) & EVENT_UNIT_MASK; | |
388 | combine = (event[i] >> EVENT_COMBINE_SHIFT) & EVENT_COMBINE_MASK; | |
389 | psel = event[i] & EVENT_PSEL_MASK; | |
390 | ||
391 | if (!pmc) { | |
392 | for (pmc = 1; pmc <= 4; ++pmc) { | |
393 | if (!(pmc_inuse & (1 << pmc))) | |
394 | break; | |
395 | } | |
396 | ||
397 | pmc_inuse |= 1 << pmc; | |
398 | } | |
399 | ||
400 | if (pmc <= 4) { | |
401 | mmcr1 |= unit << MMCR1_UNIT_SHIFT(pmc); | |
402 | mmcr1 |= combine << MMCR1_COMBINE_SHIFT(pmc); | |
403 | mmcr1 |= psel << MMCR1_PMCSEL_SHIFT(pmc); | |
404 | } | |
405 | ||
406 | if (event[i] & EVENT_IS_L1) { | |
407 | cache = event[i] >> EVENT_CACHE_SEL_SHIFT; | |
408 | mmcr1 |= (cache & 1) << MMCR1_IC_QUAL_SHIFT; | |
409 | cache >>= 1; | |
410 | mmcr1 |= (cache & 1) << MMCR1_DC_QUAL_SHIFT; | |
411 | } | |
412 | ||
413 | if (event[i] & EVENT_IS_MARKED) { | |
414 | mmcra |= MMCRA_SAMPLE_ENABLE; | |
415 | ||
416 | val = (event[i] >> EVENT_SAMPLE_SHIFT) & EVENT_SAMPLE_MASK; | |
417 | if (val) { | |
418 | mmcra |= (val & 3) << MMCRA_SAMP_MODE_SHIFT; | |
419 | mmcra |= (val >> 2) << MMCRA_SAMP_ELIG_SHIFT; | |
420 | } | |
421 | } | |
422 | ||
423 | /* | |
424 | * PM_MRK_FAB_RSP_MATCH and PM_MRK_FAB_RSP_MATCH_CYC, | |
425 | * the threshold bits are used for the match value. | |
426 | */ | |
427 | if (event_is_fab_match(event[i])) { | |
a53b27b3 ME |
428 | mmcr1 |= ((event[i] >> EVENT_THR_CTL_SHIFT) & |
429 | EVENT_THR_CTL_MASK) << MMCR1_FAB_SHIFT; | |
e05b9b9e ME |
430 | } else { |
431 | val = (event[i] >> EVENT_THR_CTL_SHIFT) & EVENT_THR_CTL_MASK; | |
432 | mmcra |= val << MMCRA_THR_CTL_SHIFT; | |
433 | val = (event[i] >> EVENT_THR_SEL_SHIFT) & EVENT_THR_SEL_MASK; | |
434 | mmcra |= val << MMCRA_THR_SEL_SHIFT; | |
435 | val = (event[i] >> EVENT_THR_CMP_SHIFT) & EVENT_THR_CMP_MASK; | |
436 | mmcra |= val << MMCRA_THR_CMP_SHIFT; | |
437 | } | |
438 | ||
439 | hwc[i] = pmc - 1; | |
440 | } | |
441 | ||
442 | /* Return MMCRx values */ | |
443 | mmcr[0] = 0; | |
444 | ||
445 | /* pmc_inuse is 1-based */ | |
446 | if (pmc_inuse & 2) | |
447 | mmcr[0] = MMCR0_PMC1CE; | |
448 | ||
449 | if (pmc_inuse & 0x7c) | |
450 | mmcr[0] |= MMCR0_PMCjCE; | |
451 | ||
7a7a41f9 ME |
452 | /* If we're not using PMC 5 or 6, freeze them */ |
453 | if (!(pmc_inuse & 0x60)) | |
454 | mmcr[0] |= MMCR0_FC56; | |
455 | ||
e05b9b9e ME |
456 | mmcr[1] = mmcr1; |
457 | mmcr[2] = mmcra; | |
458 | ||
459 | return 0; | |
460 | } | |
461 | ||
462 | #define MAX_ALT 2 | |
463 | ||
464 | /* Table of alternatives, sorted by column 0 */ | |
465 | static const unsigned int event_alternatives[][MAX_ALT] = { | |
466 | { 0x10134, 0x301e2 }, /* PM_MRK_ST_CMPL */ | |
467 | { 0x10138, 0x40138 }, /* PM_BR_MRK_2PATH */ | |
468 | { 0x18082, 0x3e05e }, /* PM_L3_CO_MEPF */ | |
469 | { 0x1d14e, 0x401e8 }, /* PM_MRK_DATA_FROM_L2MISS */ | |
470 | { 0x1e054, 0x4000a }, /* PM_CMPLU_STALL */ | |
471 | { 0x20036, 0x40036 }, /* PM_BR_2PATH */ | |
472 | { 0x200f2, 0x300f2 }, /* PM_INST_DISP */ | |
473 | { 0x200f4, 0x600f4 }, /* PM_RUN_CYC */ | |
474 | { 0x2013c, 0x3012e }, /* PM_MRK_FILT_MATCH */ | |
475 | { 0x3e054, 0x400f0 }, /* PM_LD_MISS_L1 */ | |
476 | { 0x400fa, 0x500fa }, /* PM_RUN_INST_CMPL */ | |
477 | }; | |
478 | ||
479 | /* | |
480 | * Scan the alternatives table for a match and return the | |
481 | * index into the alternatives table if found, else -1. | |
482 | */ | |
483 | static int find_alternative(u64 event) | |
484 | { | |
485 | int i, j; | |
486 | ||
487 | for (i = 0; i < ARRAY_SIZE(event_alternatives); ++i) { | |
488 | if (event < event_alternatives[i][0]) | |
489 | break; | |
490 | ||
491 | for (j = 0; j < MAX_ALT && event_alternatives[i][j]; ++j) | |
492 | if (event == event_alternatives[i][j]) | |
493 | return i; | |
494 | } | |
495 | ||
496 | return -1; | |
497 | } | |
498 | ||
499 | static int power8_get_alternatives(u64 event, unsigned int flags, u64 alt[]) | |
500 | { | |
501 | int i, j, num_alt = 0; | |
502 | u64 alt_event; | |
503 | ||
504 | alt[num_alt++] = event; | |
505 | ||
506 | i = find_alternative(event); | |
507 | if (i >= 0) { | |
508 | /* Filter out the original event, it's already in alt[0] */ | |
509 | for (j = 0; j < MAX_ALT; ++j) { | |
510 | alt_event = event_alternatives[i][j]; | |
511 | if (alt_event && alt_event != event) | |
512 | alt[num_alt++] = alt_event; | |
513 | } | |
514 | } | |
515 | ||
516 | if (flags & PPMU_ONLY_COUNT_RUN) { | |
517 | /* | |
518 | * We're only counting in RUN state, so PM_CYC is equivalent to | |
519 | * PM_RUN_CYC and PM_INST_CMPL === PM_RUN_INST_CMPL. | |
520 | */ | |
521 | j = num_alt; | |
522 | for (i = 0; i < num_alt; ++i) { | |
523 | switch (alt[i]) { | |
524 | case 0x1e: /* PM_CYC */ | |
525 | alt[j++] = 0x600f4; /* PM_RUN_CYC */ | |
526 | break; | |
527 | case 0x600f4: /* PM_RUN_CYC */ | |
528 | alt[j++] = 0x1e; | |
529 | break; | |
530 | case 0x2: /* PM_PPC_CMPL */ | |
531 | alt[j++] = 0x500fa; /* PM_RUN_INST_CMPL */ | |
532 | break; | |
533 | case 0x500fa: /* PM_RUN_INST_CMPL */ | |
534 | alt[j++] = 0x2; /* PM_PPC_CMPL */ | |
535 | break; | |
536 | } | |
537 | } | |
538 | num_alt = j; | |
539 | } | |
540 | ||
541 | return num_alt; | |
542 | } | |
543 | ||
544 | static void power8_disable_pmc(unsigned int pmc, unsigned long mmcr[]) | |
545 | { | |
546 | if (pmc <= 3) | |
547 | mmcr[1] &= ~(0xffUL << MMCR1_PMCSEL_SHIFT(pmc + 1)); | |
548 | } | |
549 | ||
550 | PMU_FORMAT_ATTR(event, "config:0-49"); | |
551 | PMU_FORMAT_ATTR(pmcxsel, "config:0-7"); | |
552 | PMU_FORMAT_ATTR(mark, "config:8"); | |
553 | PMU_FORMAT_ATTR(combine, "config:11"); | |
554 | PMU_FORMAT_ATTR(unit, "config:12-15"); | |
555 | PMU_FORMAT_ATTR(pmc, "config:16-19"); | |
556 | PMU_FORMAT_ATTR(cache_sel, "config:20-23"); | |
557 | PMU_FORMAT_ATTR(sample_mode, "config:24-28"); | |
558 | PMU_FORMAT_ATTR(thresh_sel, "config:29-31"); | |
559 | PMU_FORMAT_ATTR(thresh_stop, "config:32-35"); | |
560 | PMU_FORMAT_ATTR(thresh_start, "config:36-39"); | |
561 | PMU_FORMAT_ATTR(thresh_cmp, "config:40-49"); | |
562 | ||
563 | static struct attribute *power8_pmu_format_attr[] = { | |
564 | &format_attr_event.attr, | |
565 | &format_attr_pmcxsel.attr, | |
566 | &format_attr_mark.attr, | |
567 | &format_attr_combine.attr, | |
568 | &format_attr_unit.attr, | |
569 | &format_attr_pmc.attr, | |
570 | &format_attr_cache_sel.attr, | |
571 | &format_attr_sample_mode.attr, | |
572 | &format_attr_thresh_sel.attr, | |
573 | &format_attr_thresh_stop.attr, | |
574 | &format_attr_thresh_start.attr, | |
575 | &format_attr_thresh_cmp.attr, | |
576 | NULL, | |
577 | }; | |
578 | ||
579 | struct attribute_group power8_pmu_format_group = { | |
580 | .name = "format", | |
581 | .attrs = power8_pmu_format_attr, | |
582 | }; | |
583 | ||
584 | static const struct attribute_group *power8_pmu_attr_groups[] = { | |
585 | &power8_pmu_format_group, | |
586 | NULL, | |
587 | }; | |
588 | ||
589 | static int power8_generic_events[] = { | |
590 | [PERF_COUNT_HW_CPU_CYCLES] = PM_CYC, | |
591 | [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = PM_GCT_NOSLOT_CYC, | |
592 | [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = PM_CMPLU_STALL, | |
593 | [PERF_COUNT_HW_INSTRUCTIONS] = PM_INST_CMPL, | |
594 | [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = PM_BRU_FIN, | |
595 | [PERF_COUNT_HW_BRANCH_MISSES] = PM_BR_MPRED_CMPL, | |
2fdd313f ME |
596 | [PERF_COUNT_HW_CACHE_REFERENCES] = PM_LD_REF_L1, |
597 | [PERF_COUNT_HW_CACHE_MISSES] = PM_LD_MISS_L1, | |
e05b9b9e ME |
598 | }; |
599 | ||
b1113557 AK |
600 | static u64 power8_bhrb_filter_map(u64 branch_sample_type) |
601 | { | |
602 | u64 pmu_bhrb_filter = 0; | |
b1113557 | 603 | |
7689bdca | 604 | /* BHRB and regular PMU events share the same privilege state |
b1113557 | 605 | * filter configuration. BHRB is always recorded along with a |
7689bdca AK |
606 | * regular PMU event. As the privilege state filter is handled |
607 | * in the basic PMC configuration of the accompanying regular | |
608 | * PMU event, we ignore any separate BHRB specific request. | |
b1113557 | 609 | */ |
b1113557 AK |
610 | |
611 | /* No branch filter requested */ | |
612 | if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY) | |
613 | return pmu_bhrb_filter; | |
614 | ||
615 | /* Invalid branch filter options - HW does not support */ | |
616 | if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY_RETURN) | |
617 | return -1; | |
618 | ||
619 | if (branch_sample_type & PERF_SAMPLE_BRANCH_IND_CALL) | |
620 | return -1; | |
621 | ||
622 | if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY_CALL) { | |
623 | pmu_bhrb_filter |= POWER8_MMCRA_IFM1; | |
624 | return pmu_bhrb_filter; | |
625 | } | |
626 | ||
627 | /* Every thing else is unsupported */ | |
628 | return -1; | |
629 | } | |
630 | ||
631 | static void power8_config_bhrb(u64 pmu_bhrb_filter) | |
632 | { | |
633 | /* Enable BHRB filter in PMU */ | |
634 | mtspr(SPRN_MMCRA, (mfspr(SPRN_MMCRA) | pmu_bhrb_filter)); | |
635 | } | |
636 | ||
2fdd313f ME |
637 | #define C(x) PERF_COUNT_HW_CACHE_##x |
638 | ||
639 | /* | |
640 | * Table of generalized cache-related events. | |
641 | * 0 means not supported, -1 means nonsensical, other values | |
642 | * are event codes. | |
643 | */ | |
644 | static int power8_cache_events[C(MAX)][C(OP_MAX)][C(RESULT_MAX)] = { | |
645 | [ C(L1D) ] = { | |
646 | [ C(OP_READ) ] = { | |
647 | [ C(RESULT_ACCESS) ] = PM_LD_REF_L1, | |
648 | [ C(RESULT_MISS) ] = PM_LD_MISS_L1, | |
649 | }, | |
650 | [ C(OP_WRITE) ] = { | |
651 | [ C(RESULT_ACCESS) ] = 0, | |
652 | [ C(RESULT_MISS) ] = PM_ST_MISS_L1, | |
653 | }, | |
654 | [ C(OP_PREFETCH) ] = { | |
655 | [ C(RESULT_ACCESS) ] = PM_L1_PREF, | |
656 | [ C(RESULT_MISS) ] = 0, | |
657 | }, | |
658 | }, | |
659 | [ C(L1I) ] = { | |
660 | [ C(OP_READ) ] = { | |
661 | [ C(RESULT_ACCESS) ] = PM_INST_FROM_L1, | |
662 | [ C(RESULT_MISS) ] = PM_L1_ICACHE_MISS, | |
663 | }, | |
664 | [ C(OP_WRITE) ] = { | |
665 | [ C(RESULT_ACCESS) ] = PM_L1_DEMAND_WRITE, | |
666 | [ C(RESULT_MISS) ] = -1, | |
667 | }, | |
668 | [ C(OP_PREFETCH) ] = { | |
669 | [ C(RESULT_ACCESS) ] = PM_IC_PREF_WRITE, | |
670 | [ C(RESULT_MISS) ] = 0, | |
671 | }, | |
672 | }, | |
673 | [ C(LL) ] = { | |
674 | [ C(OP_READ) ] = { | |
675 | [ C(RESULT_ACCESS) ] = PM_DATA_FROM_L3, | |
676 | [ C(RESULT_MISS) ] = PM_DATA_FROM_L3MISS, | |
677 | }, | |
678 | [ C(OP_WRITE) ] = { | |
679 | [ C(RESULT_ACCESS) ] = PM_L2_ST, | |
680 | [ C(RESULT_MISS) ] = PM_L2_ST_MISS, | |
681 | }, | |
682 | [ C(OP_PREFETCH) ] = { | |
683 | [ C(RESULT_ACCESS) ] = PM_L3_PREF_ALL, | |
684 | [ C(RESULT_MISS) ] = 0, | |
685 | }, | |
686 | }, | |
687 | [ C(DTLB) ] = { | |
688 | [ C(OP_READ) ] = { | |
689 | [ C(RESULT_ACCESS) ] = 0, | |
690 | [ C(RESULT_MISS) ] = PM_DTLB_MISS, | |
691 | }, | |
692 | [ C(OP_WRITE) ] = { | |
693 | [ C(RESULT_ACCESS) ] = -1, | |
694 | [ C(RESULT_MISS) ] = -1, | |
695 | }, | |
696 | [ C(OP_PREFETCH) ] = { | |
697 | [ C(RESULT_ACCESS) ] = -1, | |
698 | [ C(RESULT_MISS) ] = -1, | |
699 | }, | |
700 | }, | |
701 | [ C(ITLB) ] = { | |
702 | [ C(OP_READ) ] = { | |
703 | [ C(RESULT_ACCESS) ] = 0, | |
704 | [ C(RESULT_MISS) ] = PM_ITLB_MISS, | |
705 | }, | |
706 | [ C(OP_WRITE) ] = { | |
707 | [ C(RESULT_ACCESS) ] = -1, | |
708 | [ C(RESULT_MISS) ] = -1, | |
709 | }, | |
710 | [ C(OP_PREFETCH) ] = { | |
711 | [ C(RESULT_ACCESS) ] = -1, | |
712 | [ C(RESULT_MISS) ] = -1, | |
713 | }, | |
714 | }, | |
715 | [ C(BPU) ] = { | |
716 | [ C(OP_READ) ] = { | |
717 | [ C(RESULT_ACCESS) ] = PM_BRU_FIN, | |
718 | [ C(RESULT_MISS) ] = PM_BR_MPRED_CMPL, | |
719 | }, | |
720 | [ C(OP_WRITE) ] = { | |
721 | [ C(RESULT_ACCESS) ] = -1, | |
722 | [ C(RESULT_MISS) ] = -1, | |
723 | }, | |
724 | [ C(OP_PREFETCH) ] = { | |
725 | [ C(RESULT_ACCESS) ] = -1, | |
726 | [ C(RESULT_MISS) ] = -1, | |
727 | }, | |
728 | }, | |
729 | [ C(NODE) ] = { | |
730 | [ C(OP_READ) ] = { | |
731 | [ C(RESULT_ACCESS) ] = -1, | |
732 | [ C(RESULT_MISS) ] = -1, | |
733 | }, | |
734 | [ C(OP_WRITE) ] = { | |
735 | [ C(RESULT_ACCESS) ] = -1, | |
736 | [ C(RESULT_MISS) ] = -1, | |
737 | }, | |
738 | [ C(OP_PREFETCH) ] = { | |
739 | [ C(RESULT_ACCESS) ] = -1, | |
740 | [ C(RESULT_MISS) ] = -1, | |
741 | }, | |
742 | }, | |
743 | }; | |
744 | ||
745 | #undef C | |
746 | ||
e05b9b9e ME |
747 | static struct power_pmu power8_pmu = { |
748 | .name = "POWER8", | |
749 | .n_counter = 6, | |
750 | .max_alternatives = MAX_ALT + 1, | |
751 | .add_fields = POWER8_ADD_FIELDS, | |
752 | .test_adder = POWER8_TEST_ADDER, | |
753 | .compute_mmcr = power8_compute_mmcr, | |
b1113557 AK |
754 | .config_bhrb = power8_config_bhrb, |
755 | .bhrb_filter_map = power8_bhrb_filter_map, | |
e05b9b9e ME |
756 | .get_constraint = power8_get_constraint, |
757 | .get_alternatives = power8_get_alternatives, | |
758 | .disable_pmc = power8_disable_pmc, | |
4df48999 | 759 | .flags = PPMU_HAS_SSLOT | PPMU_HAS_SIER | PPMU_BHRB | PPMU_EBB, |
e05b9b9e ME |
760 | .n_generic = ARRAY_SIZE(power8_generic_events), |
761 | .generic_events = power8_generic_events, | |
2fdd313f | 762 | .cache_events = &power8_cache_events, |
e05b9b9e | 763 | .attr_groups = power8_pmu_attr_groups, |
b1113557 | 764 | .bhrb_nr = 32, |
e05b9b9e ME |
765 | }; |
766 | ||
767 | static int __init init_power8_pmu(void) | |
768 | { | |
5d7ead00 ME |
769 | int rc; |
770 | ||
e05b9b9e ME |
771 | if (!cur_cpu_spec->oprofile_cpu_type || |
772 | strcmp(cur_cpu_spec->oprofile_cpu_type, "ppc64/power8")) | |
773 | return -ENODEV; | |
774 | ||
5d7ead00 ME |
775 | rc = register_power_pmu(&power8_pmu); |
776 | if (rc) | |
777 | return rc; | |
778 | ||
779 | /* Tell userspace that EBB is supported */ | |
780 | cur_cpu_spec->cpu_user_features2 |= PPC_FEATURE2_EBB; | |
781 | ||
c2e37a26 ME |
782 | if (cpu_has_feature(CPU_FTR_PMAO_BUG)) |
783 | pr_info("PMAO restore workaround active.\n"); | |
784 | ||
5d7ead00 | 785 | return 0; |
e05b9b9e ME |
786 | } |
787 | early_initcall(init_power8_pmu); |