Loading...
1/*
2 * ARMv7 Cortex-A8 and Cortex-A9 Performance Events handling code.
3 *
4 * ARMv7 support: Jean Pihet <jpihet@mvista.com>
5 * 2010 (c) MontaVista Software, LLC.
6 *
7 * Copied from ARMv6 code, with the low level code inspired
8 * by the ARMv7 Oprofile code.
9 *
10 * Cortex-A8 has up to 4 configurable performance counters and
11 * a single cycle counter.
12 * Cortex-A9 has up to 31 configurable performance counters and
13 * a single cycle counter.
14 *
15 * All counters can be enabled/disabled and IRQ masked separately. The cycle
16 * counter and all 4 performance counters together can be reset separately.
17 */
18
19#ifdef CONFIG_CPU_V7
20/*
21 * Common ARMv7 event types
22 *
23 * Note: An implementation may not be able to count all of these events
24 * but the encodings are considered to be `reserved' in the case that
25 * they are not available.
26 */
27enum armv7_perf_types {
28 ARMV7_PERFCTR_PMNC_SW_INCR = 0x00,
29 ARMV7_PERFCTR_IFETCH_MISS = 0x01,
30 ARMV7_PERFCTR_ITLB_MISS = 0x02,
31 ARMV7_PERFCTR_DCACHE_REFILL = 0x03, /* L1 */
32 ARMV7_PERFCTR_DCACHE_ACCESS = 0x04, /* L1 */
33 ARMV7_PERFCTR_DTLB_REFILL = 0x05,
34 ARMV7_PERFCTR_DREAD = 0x06,
35 ARMV7_PERFCTR_DWRITE = 0x07,
36 ARMV7_PERFCTR_INSTR_EXECUTED = 0x08,
37 ARMV7_PERFCTR_EXC_TAKEN = 0x09,
38 ARMV7_PERFCTR_EXC_EXECUTED = 0x0A,
39 ARMV7_PERFCTR_CID_WRITE = 0x0B,
40 /* ARMV7_PERFCTR_PC_WRITE is equivalent to HW_BRANCH_INSTRUCTIONS.
41 * It counts:
42 * - all branch instructions,
43 * - instructions that explicitly write the PC,
44 * - exception generating instructions.
45 */
46 ARMV7_PERFCTR_PC_WRITE = 0x0C,
47 ARMV7_PERFCTR_PC_IMM_BRANCH = 0x0D,
48 ARMV7_PERFCTR_PC_PROC_RETURN = 0x0E,
49 ARMV7_PERFCTR_UNALIGNED_ACCESS = 0x0F,
50
51 /* These events are defined by the PMUv2 supplement (ARM DDI 0457A). */
52 ARMV7_PERFCTR_PC_BRANCH_MIS_PRED = 0x10,
53 ARMV7_PERFCTR_CLOCK_CYCLES = 0x11,
54 ARMV7_PERFCTR_PC_BRANCH_PRED = 0x12,
55 ARMV7_PERFCTR_MEM_ACCESS = 0x13,
56 ARMV7_PERFCTR_L1_ICACHE_ACCESS = 0x14,
57 ARMV7_PERFCTR_L1_DCACHE_WB = 0x15,
58 ARMV7_PERFCTR_L2_DCACHE_ACCESS = 0x16,
59 ARMV7_PERFCTR_L2_DCACHE_REFILL = 0x17,
60 ARMV7_PERFCTR_L2_DCACHE_WB = 0x18,
61 ARMV7_PERFCTR_BUS_ACCESS = 0x19,
62 ARMV7_PERFCTR_MEMORY_ERROR = 0x1A,
63 ARMV7_PERFCTR_INSTR_SPEC = 0x1B,
64 ARMV7_PERFCTR_TTBR_WRITE = 0x1C,
65 ARMV7_PERFCTR_BUS_CYCLES = 0x1D,
66
67 ARMV7_PERFCTR_CPU_CYCLES = 0xFF
68};
69
70/* ARMv7 Cortex-A8 specific event types */
71enum armv7_a8_perf_types {
72 ARMV7_PERFCTR_WRITE_BUFFER_FULL = 0x40,
73 ARMV7_PERFCTR_L2_STORE_MERGED = 0x41,
74 ARMV7_PERFCTR_L2_STORE_BUFF = 0x42,
75 ARMV7_PERFCTR_L2_ACCESS = 0x43,
76 ARMV7_PERFCTR_L2_CACH_MISS = 0x44,
77 ARMV7_PERFCTR_AXI_READ_CYCLES = 0x45,
78 ARMV7_PERFCTR_AXI_WRITE_CYCLES = 0x46,
79 ARMV7_PERFCTR_MEMORY_REPLAY = 0x47,
80 ARMV7_PERFCTR_UNALIGNED_ACCESS_REPLAY = 0x48,
81 ARMV7_PERFCTR_L1_DATA_MISS = 0x49,
82 ARMV7_PERFCTR_L1_INST_MISS = 0x4A,
83 ARMV7_PERFCTR_L1_DATA_COLORING = 0x4B,
84 ARMV7_PERFCTR_L1_NEON_DATA = 0x4C,
85 ARMV7_PERFCTR_L1_NEON_CACH_DATA = 0x4D,
86 ARMV7_PERFCTR_L2_NEON = 0x4E,
87 ARMV7_PERFCTR_L2_NEON_HIT = 0x4F,
88 ARMV7_PERFCTR_L1_INST = 0x50,
89 ARMV7_PERFCTR_PC_RETURN_MIS_PRED = 0x51,
90 ARMV7_PERFCTR_PC_BRANCH_FAILED = 0x52,
91 ARMV7_PERFCTR_PC_BRANCH_TAKEN = 0x53,
92 ARMV7_PERFCTR_PC_BRANCH_EXECUTED = 0x54,
93 ARMV7_PERFCTR_OP_EXECUTED = 0x55,
94 ARMV7_PERFCTR_CYCLES_INST_STALL = 0x56,
95 ARMV7_PERFCTR_CYCLES_INST = 0x57,
96 ARMV7_PERFCTR_CYCLES_NEON_DATA_STALL = 0x58,
97 ARMV7_PERFCTR_CYCLES_NEON_INST_STALL = 0x59,
98 ARMV7_PERFCTR_NEON_CYCLES = 0x5A,
99
100 ARMV7_PERFCTR_PMU0_EVENTS = 0x70,
101 ARMV7_PERFCTR_PMU1_EVENTS = 0x71,
102 ARMV7_PERFCTR_PMU_EVENTS = 0x72,
103};
104
105/* ARMv7 Cortex-A9 specific event types */
106enum armv7_a9_perf_types {
107 ARMV7_PERFCTR_JAVA_HW_BYTECODE_EXEC = 0x40,
108 ARMV7_PERFCTR_JAVA_SW_BYTECODE_EXEC = 0x41,
109 ARMV7_PERFCTR_JAZELLE_BRANCH_EXEC = 0x42,
110
111 ARMV7_PERFCTR_COHERENT_LINE_MISS = 0x50,
112 ARMV7_PERFCTR_COHERENT_LINE_HIT = 0x51,
113
114 ARMV7_PERFCTR_ICACHE_DEP_STALL_CYCLES = 0x60,
115 ARMV7_PERFCTR_DCACHE_DEP_STALL_CYCLES = 0x61,
116 ARMV7_PERFCTR_TLB_MISS_DEP_STALL_CYCLES = 0x62,
117 ARMV7_PERFCTR_STREX_EXECUTED_PASSED = 0x63,
118 ARMV7_PERFCTR_STREX_EXECUTED_FAILED = 0x64,
119 ARMV7_PERFCTR_DATA_EVICTION = 0x65,
120 ARMV7_PERFCTR_ISSUE_STAGE_NO_INST = 0x66,
121 ARMV7_PERFCTR_ISSUE_STAGE_EMPTY = 0x67,
122 ARMV7_PERFCTR_INST_OUT_OF_RENAME_STAGE = 0x68,
123
124 ARMV7_PERFCTR_PREDICTABLE_FUNCT_RETURNS = 0x6E,
125
126 ARMV7_PERFCTR_MAIN_UNIT_EXECUTED_INST = 0x70,
127 ARMV7_PERFCTR_SECOND_UNIT_EXECUTED_INST = 0x71,
128 ARMV7_PERFCTR_LD_ST_UNIT_EXECUTED_INST = 0x72,
129 ARMV7_PERFCTR_FP_EXECUTED_INST = 0x73,
130 ARMV7_PERFCTR_NEON_EXECUTED_INST = 0x74,
131
132 ARMV7_PERFCTR_PLD_FULL_DEP_STALL_CYCLES = 0x80,
133 ARMV7_PERFCTR_DATA_WR_DEP_STALL_CYCLES = 0x81,
134 ARMV7_PERFCTR_ITLB_MISS_DEP_STALL_CYCLES = 0x82,
135 ARMV7_PERFCTR_DTLB_MISS_DEP_STALL_CYCLES = 0x83,
136 ARMV7_PERFCTR_MICRO_ITLB_MISS_DEP_STALL_CYCLES = 0x84,
137 ARMV7_PERFCTR_MICRO_DTLB_MISS_DEP_STALL_CYCLES = 0x85,
138 ARMV7_PERFCTR_DMB_DEP_STALL_CYCLES = 0x86,
139
140 ARMV7_PERFCTR_INTGR_CLK_ENABLED_CYCLES = 0x8A,
141 ARMV7_PERFCTR_DATA_ENGINE_CLK_EN_CYCLES = 0x8B,
142
143 ARMV7_PERFCTR_ISB_INST = 0x90,
144 ARMV7_PERFCTR_DSB_INST = 0x91,
145 ARMV7_PERFCTR_DMB_INST = 0x92,
146 ARMV7_PERFCTR_EXT_INTERRUPTS = 0x93,
147
148 ARMV7_PERFCTR_PLE_CACHE_LINE_RQST_COMPLETED = 0xA0,
149 ARMV7_PERFCTR_PLE_CACHE_LINE_RQST_SKIPPED = 0xA1,
150 ARMV7_PERFCTR_PLE_FIFO_FLUSH = 0xA2,
151 ARMV7_PERFCTR_PLE_RQST_COMPLETED = 0xA3,
152 ARMV7_PERFCTR_PLE_FIFO_OVERFLOW = 0xA4,
153 ARMV7_PERFCTR_PLE_RQST_PROG = 0xA5
154};
155
156/* ARMv7 Cortex-A5 specific event types */
157enum armv7_a5_perf_types {
158 ARMV7_PERFCTR_IRQ_TAKEN = 0x86,
159 ARMV7_PERFCTR_FIQ_TAKEN = 0x87,
160
161 ARMV7_PERFCTR_EXT_MEM_RQST = 0xc0,
162 ARMV7_PERFCTR_NC_EXT_MEM_RQST = 0xc1,
163 ARMV7_PERFCTR_PREFETCH_LINEFILL = 0xc2,
164 ARMV7_PERFCTR_PREFETCH_LINEFILL_DROP = 0xc3,
165 ARMV7_PERFCTR_ENTER_READ_ALLOC = 0xc4,
166 ARMV7_PERFCTR_READ_ALLOC = 0xc5,
167
168 ARMV7_PERFCTR_STALL_SB_FULL = 0xc9,
169};
170
171/* ARMv7 Cortex-A15 specific event types */
172enum armv7_a15_perf_types {
173 ARMV7_PERFCTR_L1_DCACHE_READ_ACCESS = 0x40,
174 ARMV7_PERFCTR_L1_DCACHE_WRITE_ACCESS = 0x41,
175 ARMV7_PERFCTR_L1_DCACHE_READ_REFILL = 0x42,
176 ARMV7_PERFCTR_L1_DCACHE_WRITE_REFILL = 0x43,
177
178 ARMV7_PERFCTR_L1_DTLB_READ_REFILL = 0x4C,
179 ARMV7_PERFCTR_L1_DTLB_WRITE_REFILL = 0x4D,
180
181 ARMV7_PERFCTR_L2_DCACHE_READ_ACCESS = 0x50,
182 ARMV7_PERFCTR_L2_DCACHE_WRITE_ACCESS = 0x51,
183 ARMV7_PERFCTR_L2_DCACHE_READ_REFILL = 0x52,
184 ARMV7_PERFCTR_L2_DCACHE_WRITE_REFILL = 0x53,
185
186 ARMV7_PERFCTR_SPEC_PC_WRITE = 0x76,
187};
188
189/*
190 * Cortex-A8 HW events mapping
191 *
192 * The hardware events that we support. We do support cache operations but
193 * we have harvard caches and no way to combine instruction and data
194 * accesses/misses in hardware.
195 */
196static const unsigned armv7_a8_perf_map[PERF_COUNT_HW_MAX] = {
197 [PERF_COUNT_HW_CPU_CYCLES] = ARMV7_PERFCTR_CPU_CYCLES,
198 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV7_PERFCTR_INSTR_EXECUTED,
199 [PERF_COUNT_HW_CACHE_REFERENCES] = HW_OP_UNSUPPORTED,
200 [PERF_COUNT_HW_CACHE_MISSES] = HW_OP_UNSUPPORTED,
201 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV7_PERFCTR_PC_WRITE,
202 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
203 [PERF_COUNT_HW_BUS_CYCLES] = ARMV7_PERFCTR_CLOCK_CYCLES,
204};
205
206static const unsigned armv7_a8_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
207 [PERF_COUNT_HW_CACHE_OP_MAX]
208 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
209 [C(L1D)] = {
210 /*
211 * The performance counters don't differentiate between read
212 * and write accesses/misses so this isn't strictly correct,
213 * but it's the best we can do. Writes and reads get
214 * combined.
215 */
216 [C(OP_READ)] = {
217 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_DCACHE_ACCESS,
218 [C(RESULT_MISS)] = ARMV7_PERFCTR_DCACHE_REFILL,
219 },
220 [C(OP_WRITE)] = {
221 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_DCACHE_ACCESS,
222 [C(RESULT_MISS)] = ARMV7_PERFCTR_DCACHE_REFILL,
223 },
224 [C(OP_PREFETCH)] = {
225 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
226 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
227 },
228 },
229 [C(L1I)] = {
230 [C(OP_READ)] = {
231 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_INST,
232 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_INST_MISS,
233 },
234 [C(OP_WRITE)] = {
235 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_INST,
236 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_INST_MISS,
237 },
238 [C(OP_PREFETCH)] = {
239 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
240 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
241 },
242 },
243 [C(LL)] = {
244 [C(OP_READ)] = {
245 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L2_ACCESS,
246 [C(RESULT_MISS)] = ARMV7_PERFCTR_L2_CACH_MISS,
247 },
248 [C(OP_WRITE)] = {
249 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L2_ACCESS,
250 [C(RESULT_MISS)] = ARMV7_PERFCTR_L2_CACH_MISS,
251 },
252 [C(OP_PREFETCH)] = {
253 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
254 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
255 },
256 },
257 [C(DTLB)] = {
258 [C(OP_READ)] = {
259 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
260 [C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
261 },
262 [C(OP_WRITE)] = {
263 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
264 [C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
265 },
266 [C(OP_PREFETCH)] = {
267 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
268 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
269 },
270 },
271 [C(ITLB)] = {
272 [C(OP_READ)] = {
273 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
274 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_MISS,
275 },
276 [C(OP_WRITE)] = {
277 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
278 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_MISS,
279 },
280 [C(OP_PREFETCH)] = {
281 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
282 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
283 },
284 },
285 [C(BPU)] = {
286 [C(OP_READ)] = {
287 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_WRITE,
288 [C(RESULT_MISS)]
289 = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
290 },
291 [C(OP_WRITE)] = {
292 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_WRITE,
293 [C(RESULT_MISS)]
294 = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
295 },
296 [C(OP_PREFETCH)] = {
297 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
298 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
299 },
300 },
301 [C(NODE)] = {
302 [C(OP_READ)] = {
303 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
304 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
305 },
306 [C(OP_WRITE)] = {
307 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
308 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
309 },
310 [C(OP_PREFETCH)] = {
311 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
312 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
313 },
314 },
315};
316
317/*
318 * Cortex-A9 HW events mapping
319 */
320static const unsigned armv7_a9_perf_map[PERF_COUNT_HW_MAX] = {
321 [PERF_COUNT_HW_CPU_CYCLES] = ARMV7_PERFCTR_CPU_CYCLES,
322 [PERF_COUNT_HW_INSTRUCTIONS] =
323 ARMV7_PERFCTR_INST_OUT_OF_RENAME_STAGE,
324 [PERF_COUNT_HW_CACHE_REFERENCES] = ARMV7_PERFCTR_DCACHE_ACCESS,
325 [PERF_COUNT_HW_CACHE_MISSES] = ARMV7_PERFCTR_DCACHE_REFILL,
326 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV7_PERFCTR_PC_WRITE,
327 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
328 [PERF_COUNT_HW_BUS_CYCLES] = ARMV7_PERFCTR_CLOCK_CYCLES,
329};
330
331static const unsigned armv7_a9_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
332 [PERF_COUNT_HW_CACHE_OP_MAX]
333 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
334 [C(L1D)] = {
335 /*
336 * The performance counters don't differentiate between read
337 * and write accesses/misses so this isn't strictly correct,
338 * but it's the best we can do. Writes and reads get
339 * combined.
340 */
341 [C(OP_READ)] = {
342 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_DCACHE_ACCESS,
343 [C(RESULT_MISS)] = ARMV7_PERFCTR_DCACHE_REFILL,
344 },
345 [C(OP_WRITE)] = {
346 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_DCACHE_ACCESS,
347 [C(RESULT_MISS)] = ARMV7_PERFCTR_DCACHE_REFILL,
348 },
349 [C(OP_PREFETCH)] = {
350 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
351 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
352 },
353 },
354 [C(L1I)] = {
355 [C(OP_READ)] = {
356 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
357 [C(RESULT_MISS)] = ARMV7_PERFCTR_IFETCH_MISS,
358 },
359 [C(OP_WRITE)] = {
360 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
361 [C(RESULT_MISS)] = ARMV7_PERFCTR_IFETCH_MISS,
362 },
363 [C(OP_PREFETCH)] = {
364 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
365 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
366 },
367 },
368 [C(LL)] = {
369 [C(OP_READ)] = {
370 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
371 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
372 },
373 [C(OP_WRITE)] = {
374 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
375 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
376 },
377 [C(OP_PREFETCH)] = {
378 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
379 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
380 },
381 },
382 [C(DTLB)] = {
383 [C(OP_READ)] = {
384 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
385 [C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
386 },
387 [C(OP_WRITE)] = {
388 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
389 [C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
390 },
391 [C(OP_PREFETCH)] = {
392 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
393 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
394 },
395 },
396 [C(ITLB)] = {
397 [C(OP_READ)] = {
398 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
399 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_MISS,
400 },
401 [C(OP_WRITE)] = {
402 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
403 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_MISS,
404 },
405 [C(OP_PREFETCH)] = {
406 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
407 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
408 },
409 },
410 [C(BPU)] = {
411 [C(OP_READ)] = {
412 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_WRITE,
413 [C(RESULT_MISS)]
414 = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
415 },
416 [C(OP_WRITE)] = {
417 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_WRITE,
418 [C(RESULT_MISS)]
419 = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
420 },
421 [C(OP_PREFETCH)] = {
422 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
423 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
424 },
425 },
426 [C(NODE)] = {
427 [C(OP_READ)] = {
428 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
429 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
430 },
431 [C(OP_WRITE)] = {
432 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
433 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
434 },
435 [C(OP_PREFETCH)] = {
436 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
437 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
438 },
439 },
440};
441
442/*
443 * Cortex-A5 HW events mapping
444 */
445static const unsigned armv7_a5_perf_map[PERF_COUNT_HW_MAX] = {
446 [PERF_COUNT_HW_CPU_CYCLES] = ARMV7_PERFCTR_CPU_CYCLES,
447 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV7_PERFCTR_INSTR_EXECUTED,
448 [PERF_COUNT_HW_CACHE_REFERENCES] = HW_OP_UNSUPPORTED,
449 [PERF_COUNT_HW_CACHE_MISSES] = HW_OP_UNSUPPORTED,
450 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV7_PERFCTR_PC_WRITE,
451 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
452 [PERF_COUNT_HW_BUS_CYCLES] = HW_OP_UNSUPPORTED,
453};
454
455static const unsigned armv7_a5_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
456 [PERF_COUNT_HW_CACHE_OP_MAX]
457 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
458 [C(L1D)] = {
459 [C(OP_READ)] = {
460 [C(RESULT_ACCESS)]
461 = ARMV7_PERFCTR_DCACHE_ACCESS,
462 [C(RESULT_MISS)]
463 = ARMV7_PERFCTR_DCACHE_REFILL,
464 },
465 [C(OP_WRITE)] = {
466 [C(RESULT_ACCESS)]
467 = ARMV7_PERFCTR_DCACHE_ACCESS,
468 [C(RESULT_MISS)]
469 = ARMV7_PERFCTR_DCACHE_REFILL,
470 },
471 [C(OP_PREFETCH)] = {
472 [C(RESULT_ACCESS)]
473 = ARMV7_PERFCTR_PREFETCH_LINEFILL,
474 [C(RESULT_MISS)]
475 = ARMV7_PERFCTR_PREFETCH_LINEFILL_DROP,
476 },
477 },
478 [C(L1I)] = {
479 [C(OP_READ)] = {
480 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_ICACHE_ACCESS,
481 [C(RESULT_MISS)] = ARMV7_PERFCTR_IFETCH_MISS,
482 },
483 [C(OP_WRITE)] = {
484 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_ICACHE_ACCESS,
485 [C(RESULT_MISS)] = ARMV7_PERFCTR_IFETCH_MISS,
486 },
487 /*
488 * The prefetch counters don't differentiate between the I
489 * side and the D side.
490 */
491 [C(OP_PREFETCH)] = {
492 [C(RESULT_ACCESS)]
493 = ARMV7_PERFCTR_PREFETCH_LINEFILL,
494 [C(RESULT_MISS)]
495 = ARMV7_PERFCTR_PREFETCH_LINEFILL_DROP,
496 },
497 },
498 [C(LL)] = {
499 [C(OP_READ)] = {
500 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
501 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
502 },
503 [C(OP_WRITE)] = {
504 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
505 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
506 },
507 [C(OP_PREFETCH)] = {
508 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
509 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
510 },
511 },
512 [C(DTLB)] = {
513 [C(OP_READ)] = {
514 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
515 [C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
516 },
517 [C(OP_WRITE)] = {
518 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
519 [C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
520 },
521 [C(OP_PREFETCH)] = {
522 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
523 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
524 },
525 },
526 [C(ITLB)] = {
527 [C(OP_READ)] = {
528 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
529 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_MISS,
530 },
531 [C(OP_WRITE)] = {
532 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
533 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_MISS,
534 },
535 [C(OP_PREFETCH)] = {
536 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
537 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
538 },
539 },
540 [C(BPU)] = {
541 [C(OP_READ)] = {
542 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED,
543 [C(RESULT_MISS)]
544 = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
545 },
546 [C(OP_WRITE)] = {
547 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED,
548 [C(RESULT_MISS)]
549 = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
550 },
551 [C(OP_PREFETCH)] = {
552 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
553 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
554 },
555 },
556};
557
558/*
559 * Cortex-A15 HW events mapping
560 */
561static const unsigned armv7_a15_perf_map[PERF_COUNT_HW_MAX] = {
562 [PERF_COUNT_HW_CPU_CYCLES] = ARMV7_PERFCTR_CPU_CYCLES,
563 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV7_PERFCTR_INSTR_EXECUTED,
564 [PERF_COUNT_HW_CACHE_REFERENCES] = HW_OP_UNSUPPORTED,
565 [PERF_COUNT_HW_CACHE_MISSES] = HW_OP_UNSUPPORTED,
566 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV7_PERFCTR_SPEC_PC_WRITE,
567 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
568 [PERF_COUNT_HW_BUS_CYCLES] = ARMV7_PERFCTR_BUS_CYCLES,
569};
570
571static const unsigned armv7_a15_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
572 [PERF_COUNT_HW_CACHE_OP_MAX]
573 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
574 [C(L1D)] = {
575 [C(OP_READ)] = {
576 [C(RESULT_ACCESS)]
577 = ARMV7_PERFCTR_L1_DCACHE_READ_ACCESS,
578 [C(RESULT_MISS)]
579 = ARMV7_PERFCTR_L1_DCACHE_READ_REFILL,
580 },
581 [C(OP_WRITE)] = {
582 [C(RESULT_ACCESS)]
583 = ARMV7_PERFCTR_L1_DCACHE_WRITE_ACCESS,
584 [C(RESULT_MISS)]
585 = ARMV7_PERFCTR_L1_DCACHE_WRITE_REFILL,
586 },
587 [C(OP_PREFETCH)] = {
588 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
589 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
590 },
591 },
592 [C(L1I)] = {
593 /*
594 * Not all performance counters differentiate between read
595 * and write accesses/misses so we're not always strictly
596 * correct, but it's the best we can do. Writes and reads get
597 * combined in these cases.
598 */
599 [C(OP_READ)] = {
600 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_ICACHE_ACCESS,
601 [C(RESULT_MISS)] = ARMV7_PERFCTR_IFETCH_MISS,
602 },
603 [C(OP_WRITE)] = {
604 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_ICACHE_ACCESS,
605 [C(RESULT_MISS)] = ARMV7_PERFCTR_IFETCH_MISS,
606 },
607 [C(OP_PREFETCH)] = {
608 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
609 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
610 },
611 },
612 [C(LL)] = {
613 [C(OP_READ)] = {
614 [C(RESULT_ACCESS)]
615 = ARMV7_PERFCTR_L2_DCACHE_READ_ACCESS,
616 [C(RESULT_MISS)]
617 = ARMV7_PERFCTR_L2_DCACHE_READ_REFILL,
618 },
619 [C(OP_WRITE)] = {
620 [C(RESULT_ACCESS)]
621 = ARMV7_PERFCTR_L2_DCACHE_WRITE_ACCESS,
622 [C(RESULT_MISS)]
623 = ARMV7_PERFCTR_L2_DCACHE_WRITE_REFILL,
624 },
625 [C(OP_PREFETCH)] = {
626 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
627 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
628 },
629 },
630 [C(DTLB)] = {
631 [C(OP_READ)] = {
632 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
633 [C(RESULT_MISS)]
634 = ARMV7_PERFCTR_L1_DTLB_READ_REFILL,
635 },
636 [C(OP_WRITE)] = {
637 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
638 [C(RESULT_MISS)]
639 = ARMV7_PERFCTR_L1_DTLB_WRITE_REFILL,
640 },
641 [C(OP_PREFETCH)] = {
642 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
643 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
644 },
645 },
646 [C(ITLB)] = {
647 [C(OP_READ)] = {
648 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
649 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_MISS,
650 },
651 [C(OP_WRITE)] = {
652 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
653 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_MISS,
654 },
655 [C(OP_PREFETCH)] = {
656 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
657 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
658 },
659 },
660 [C(BPU)] = {
661 [C(OP_READ)] = {
662 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED,
663 [C(RESULT_MISS)]
664 = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
665 },
666 [C(OP_WRITE)] = {
667 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED,
668 [C(RESULT_MISS)]
669 = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
670 },
671 [C(OP_PREFETCH)] = {
672 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
673 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
674 },
675 },
676};
677
678/*
679 * Perf Events counters
680 */
681enum armv7_counters {
682 ARMV7_CYCLE_COUNTER = 1, /* Cycle counter */
683 ARMV7_COUNTER0 = 2, /* First event counter */
684};
685
686/*
687 * The cycle counter is ARMV7_CYCLE_COUNTER.
688 * The first event counter is ARMV7_COUNTER0.
689 * The last event counter is (ARMV7_COUNTER0 + armpmu->num_events - 1).
690 */
691#define ARMV7_COUNTER_LAST (ARMV7_COUNTER0 + armpmu->num_events - 1)
692
693/*
694 * ARMv7 low level PMNC access
695 */
696
697/*
698 * Per-CPU PMNC: config reg
699 */
700#define ARMV7_PMNC_E (1 << 0) /* Enable all counters */
701#define ARMV7_PMNC_P (1 << 1) /* Reset all counters */
702#define ARMV7_PMNC_C (1 << 2) /* Cycle counter reset */
703#define ARMV7_PMNC_D (1 << 3) /* CCNT counts every 64th cpu cycle */
704#define ARMV7_PMNC_X (1 << 4) /* Export to ETM */
705#define ARMV7_PMNC_DP (1 << 5) /* Disable CCNT if non-invasive debug*/
706#define ARMV7_PMNC_N_SHIFT 11 /* Number of counters supported */
707#define ARMV7_PMNC_N_MASK 0x1f
708#define ARMV7_PMNC_MASK 0x3f /* Mask for writable bits */
709
710/*
711 * Available counters
712 */
713#define ARMV7_CNT0 0 /* First event counter */
714#define ARMV7_CCNT 31 /* Cycle counter */
715
716/* Perf Event to low level counters mapping */
717#define ARMV7_EVENT_CNT_TO_CNTx (ARMV7_COUNTER0 - ARMV7_CNT0)
718
719/*
720 * CNTENS: counters enable reg
721 */
722#define ARMV7_CNTENS_P(idx) (1 << (idx - ARMV7_EVENT_CNT_TO_CNTx))
723#define ARMV7_CNTENS_C (1 << ARMV7_CCNT)
724
725/*
726 * CNTENC: counters disable reg
727 */
728#define ARMV7_CNTENC_P(idx) (1 << (idx - ARMV7_EVENT_CNT_TO_CNTx))
729#define ARMV7_CNTENC_C (1 << ARMV7_CCNT)
730
731/*
732 * INTENS: counters overflow interrupt enable reg
733 */
734#define ARMV7_INTENS_P(idx) (1 << (idx - ARMV7_EVENT_CNT_TO_CNTx))
735#define ARMV7_INTENS_C (1 << ARMV7_CCNT)
736
737/*
738 * INTENC: counters overflow interrupt disable reg
739 */
740#define ARMV7_INTENC_P(idx) (1 << (idx - ARMV7_EVENT_CNT_TO_CNTx))
741#define ARMV7_INTENC_C (1 << ARMV7_CCNT)
742
743/*
744 * EVTSEL: Event selection reg
745 */
746#define ARMV7_EVTSEL_MASK 0xff /* Mask for writable bits */
747
748/*
749 * SELECT: Counter selection reg
750 */
751#define ARMV7_SELECT_MASK 0x1f /* Mask for writable bits */
752
753/*
754 * FLAG: counters overflow flag status reg
755 */
756#define ARMV7_FLAG_P(idx) (1 << (idx - ARMV7_EVENT_CNT_TO_CNTx))
757#define ARMV7_FLAG_C (1 << ARMV7_CCNT)
758#define ARMV7_FLAG_MASK 0xffffffff /* Mask for writable bits */
759#define ARMV7_OVERFLOWED_MASK ARMV7_FLAG_MASK
760
761static inline unsigned long armv7_pmnc_read(void)
762{
763 u32 val;
764 asm volatile("mrc p15, 0, %0, c9, c12, 0" : "=r"(val));
765 return val;
766}
767
768static inline void armv7_pmnc_write(unsigned long val)
769{
770 val &= ARMV7_PMNC_MASK;
771 isb();
772 asm volatile("mcr p15, 0, %0, c9, c12, 0" : : "r"(val));
773}
774
775static inline int armv7_pmnc_has_overflowed(unsigned long pmnc)
776{
777 return pmnc & ARMV7_OVERFLOWED_MASK;
778}
779
780static inline int armv7_pmnc_counter_has_overflowed(unsigned long pmnc,
781 enum armv7_counters counter)
782{
783 int ret = 0;
784
785 if (counter == ARMV7_CYCLE_COUNTER)
786 ret = pmnc & ARMV7_FLAG_C;
787 else if ((counter >= ARMV7_COUNTER0) && (counter <= ARMV7_COUNTER_LAST))
788 ret = pmnc & ARMV7_FLAG_P(counter);
789 else
790 pr_err("CPU%u checking wrong counter %d overflow status\n",
791 smp_processor_id(), counter);
792
793 return ret;
794}
795
796static inline int armv7_pmnc_select_counter(unsigned int idx)
797{
798 u32 val;
799
800 if ((idx < ARMV7_COUNTER0) || (idx > ARMV7_COUNTER_LAST)) {
801 pr_err("CPU%u selecting wrong PMNC counter"
802 " %d\n", smp_processor_id(), idx);
803 return -1;
804 }
805
806 val = (idx - ARMV7_EVENT_CNT_TO_CNTx) & ARMV7_SELECT_MASK;
807 asm volatile("mcr p15, 0, %0, c9, c12, 5" : : "r" (val));
808 isb();
809
810 return idx;
811}
812
813static inline u32 armv7pmu_read_counter(int idx)
814{
815 unsigned long value = 0;
816
817 if (idx == ARMV7_CYCLE_COUNTER)
818 asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r" (value));
819 else if ((idx >= ARMV7_COUNTER0) && (idx <= ARMV7_COUNTER_LAST)) {
820 if (armv7_pmnc_select_counter(idx) == idx)
821 asm volatile("mrc p15, 0, %0, c9, c13, 2"
822 : "=r" (value));
823 } else
824 pr_err("CPU%u reading wrong counter %d\n",
825 smp_processor_id(), idx);
826
827 return value;
828}
829
830static inline void armv7pmu_write_counter(int idx, u32 value)
831{
832 if (idx == ARMV7_CYCLE_COUNTER)
833 asm volatile("mcr p15, 0, %0, c9, c13, 0" : : "r" (value));
834 else if ((idx >= ARMV7_COUNTER0) && (idx <= ARMV7_COUNTER_LAST)) {
835 if (armv7_pmnc_select_counter(idx) == idx)
836 asm volatile("mcr p15, 0, %0, c9, c13, 2"
837 : : "r" (value));
838 } else
839 pr_err("CPU%u writing wrong counter %d\n",
840 smp_processor_id(), idx);
841}
842
843static inline void armv7_pmnc_write_evtsel(unsigned int idx, u32 val)
844{
845 if (armv7_pmnc_select_counter(idx) == idx) {
846 val &= ARMV7_EVTSEL_MASK;
847 asm volatile("mcr p15, 0, %0, c9, c13, 1" : : "r" (val));
848 }
849}
850
851static inline u32 armv7_pmnc_enable_counter(unsigned int idx)
852{
853 u32 val;
854
855 if ((idx != ARMV7_CYCLE_COUNTER) &&
856 ((idx < ARMV7_COUNTER0) || (idx > ARMV7_COUNTER_LAST))) {
857 pr_err("CPU%u enabling wrong PMNC counter"
858 " %d\n", smp_processor_id(), idx);
859 return -1;
860 }
861
862 if (idx == ARMV7_CYCLE_COUNTER)
863 val = ARMV7_CNTENS_C;
864 else
865 val = ARMV7_CNTENS_P(idx);
866
867 asm volatile("mcr p15, 0, %0, c9, c12, 1" : : "r" (val));
868
869 return idx;
870}
871
872static inline u32 armv7_pmnc_disable_counter(unsigned int idx)
873{
874 u32 val;
875
876
877 if ((idx != ARMV7_CYCLE_COUNTER) &&
878 ((idx < ARMV7_COUNTER0) || (idx > ARMV7_COUNTER_LAST))) {
879 pr_err("CPU%u disabling wrong PMNC counter"
880 " %d\n", smp_processor_id(), idx);
881 return -1;
882 }
883
884 if (idx == ARMV7_CYCLE_COUNTER)
885 val = ARMV7_CNTENC_C;
886 else
887 val = ARMV7_CNTENC_P(idx);
888
889 asm volatile("mcr p15, 0, %0, c9, c12, 2" : : "r" (val));
890
891 return idx;
892}
893
894static inline u32 armv7_pmnc_enable_intens(unsigned int idx)
895{
896 u32 val;
897
898 if ((idx != ARMV7_CYCLE_COUNTER) &&
899 ((idx < ARMV7_COUNTER0) || (idx > ARMV7_COUNTER_LAST))) {
900 pr_err("CPU%u enabling wrong PMNC counter"
901 " interrupt enable %d\n", smp_processor_id(), idx);
902 return -1;
903 }
904
905 if (idx == ARMV7_CYCLE_COUNTER)
906 val = ARMV7_INTENS_C;
907 else
908 val = ARMV7_INTENS_P(idx);
909
910 asm volatile("mcr p15, 0, %0, c9, c14, 1" : : "r" (val));
911
912 return idx;
913}
914
915static inline u32 armv7_pmnc_disable_intens(unsigned int idx)
916{
917 u32 val;
918
919 if ((idx != ARMV7_CYCLE_COUNTER) &&
920 ((idx < ARMV7_COUNTER0) || (idx > ARMV7_COUNTER_LAST))) {
921 pr_err("CPU%u disabling wrong PMNC counter"
922 " interrupt enable %d\n", smp_processor_id(), idx);
923 return -1;
924 }
925
926 if (idx == ARMV7_CYCLE_COUNTER)
927 val = ARMV7_INTENC_C;
928 else
929 val = ARMV7_INTENC_P(idx);
930
931 asm volatile("mcr p15, 0, %0, c9, c14, 2" : : "r" (val));
932
933 return idx;
934}
935
936static inline u32 armv7_pmnc_getreset_flags(void)
937{
938 u32 val;
939
940 /* Read */
941 asm volatile("mrc p15, 0, %0, c9, c12, 3" : "=r" (val));
942
943 /* Write to clear flags */
944 val &= ARMV7_FLAG_MASK;
945 asm volatile("mcr p15, 0, %0, c9, c12, 3" : : "r" (val));
946
947 return val;
948}
949
950#ifdef DEBUG
951static void armv7_pmnc_dump_regs(void)
952{
953 u32 val;
954 unsigned int cnt;
955
956 printk(KERN_INFO "PMNC registers dump:\n");
957
958 asm volatile("mrc p15, 0, %0, c9, c12, 0" : "=r" (val));
959 printk(KERN_INFO "PMNC =0x%08x\n", val);
960
961 asm volatile("mrc p15, 0, %0, c9, c12, 1" : "=r" (val));
962 printk(KERN_INFO "CNTENS=0x%08x\n", val);
963
964 asm volatile("mrc p15, 0, %0, c9, c14, 1" : "=r" (val));
965 printk(KERN_INFO "INTENS=0x%08x\n", val);
966
967 asm volatile("mrc p15, 0, %0, c9, c12, 3" : "=r" (val));
968 printk(KERN_INFO "FLAGS =0x%08x\n", val);
969
970 asm volatile("mrc p15, 0, %0, c9, c12, 5" : "=r" (val));
971 printk(KERN_INFO "SELECT=0x%08x\n", val);
972
973 asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r" (val));
974 printk(KERN_INFO "CCNT =0x%08x\n", val);
975
976 for (cnt = ARMV7_COUNTER0; cnt < ARMV7_COUNTER_LAST; cnt++) {
977 armv7_pmnc_select_counter(cnt);
978 asm volatile("mrc p15, 0, %0, c9, c13, 2" : "=r" (val));
979 printk(KERN_INFO "CNT[%d] count =0x%08x\n",
980 cnt-ARMV7_EVENT_CNT_TO_CNTx, val);
981 asm volatile("mrc p15, 0, %0, c9, c13, 1" : "=r" (val));
982 printk(KERN_INFO "CNT[%d] evtsel=0x%08x\n",
983 cnt-ARMV7_EVENT_CNT_TO_CNTx, val);
984 }
985}
986#endif
987
988static void armv7pmu_enable_event(struct hw_perf_event *hwc, int idx)
989{
990 unsigned long flags;
991
992 /*
993 * Enable counter and interrupt, and set the counter to count
994 * the event that we're interested in.
995 */
996 raw_spin_lock_irqsave(&pmu_lock, flags);
997
998 /*
999 * Disable counter
1000 */
1001 armv7_pmnc_disable_counter(idx);
1002
1003 /*
1004 * Set event (if destined for PMNx counters)
1005 * We don't need to set the event if it's a cycle count
1006 */
1007 if (idx != ARMV7_CYCLE_COUNTER)
1008 armv7_pmnc_write_evtsel(idx, hwc->config_base);
1009
1010 /*
1011 * Enable interrupt for this counter
1012 */
1013 armv7_pmnc_enable_intens(idx);
1014
1015 /*
1016 * Enable counter
1017 */
1018 armv7_pmnc_enable_counter(idx);
1019
1020 raw_spin_unlock_irqrestore(&pmu_lock, flags);
1021}
1022
1023static void armv7pmu_disable_event(struct hw_perf_event *hwc, int idx)
1024{
1025 unsigned long flags;
1026
1027 /*
1028 * Disable counter and interrupt
1029 */
1030 raw_spin_lock_irqsave(&pmu_lock, flags);
1031
1032 /*
1033 * Disable counter
1034 */
1035 armv7_pmnc_disable_counter(idx);
1036
1037 /*
1038 * Disable interrupt for this counter
1039 */
1040 armv7_pmnc_disable_intens(idx);
1041
1042 raw_spin_unlock_irqrestore(&pmu_lock, flags);
1043}
1044
1045static irqreturn_t armv7pmu_handle_irq(int irq_num, void *dev)
1046{
1047 unsigned long pmnc;
1048 struct perf_sample_data data;
1049 struct cpu_hw_events *cpuc;
1050 struct pt_regs *regs;
1051 int idx;
1052
1053 /*
1054 * Get and reset the IRQ flags
1055 */
1056 pmnc = armv7_pmnc_getreset_flags();
1057
1058 /*
1059 * Did an overflow occur?
1060 */
1061 if (!armv7_pmnc_has_overflowed(pmnc))
1062 return IRQ_NONE;
1063
1064 /*
1065 * Handle the counter(s) overflow(s)
1066 */
1067 regs = get_irq_regs();
1068
1069 perf_sample_data_init(&data, 0);
1070
1071 cpuc = &__get_cpu_var(cpu_hw_events);
1072 for (idx = 0; idx <= armpmu->num_events; ++idx) {
1073 struct perf_event *event = cpuc->events[idx];
1074 struct hw_perf_event *hwc;
1075
1076 if (!test_bit(idx, cpuc->active_mask))
1077 continue;
1078
1079 /*
1080 * We have a single interrupt for all counters. Check that
1081 * each counter has overflowed before we process it.
1082 */
1083 if (!armv7_pmnc_counter_has_overflowed(pmnc, idx))
1084 continue;
1085
1086 hwc = &event->hw;
1087 armpmu_event_update(event, hwc, idx, 1);
1088 data.period = event->hw.last_period;
1089 if (!armpmu_event_set_period(event, hwc, idx))
1090 continue;
1091
1092 if (perf_event_overflow(event, &data, regs))
1093 armpmu->disable(hwc, idx);
1094 }
1095
1096 /*
1097 * Handle the pending perf events.
1098 *
1099 * Note: this call *must* be run with interrupts disabled. For
1100 * platforms that can have the PMU interrupts raised as an NMI, this
1101 * will not work.
1102 */
1103 irq_work_run();
1104
1105 return IRQ_HANDLED;
1106}
1107
1108static void armv7pmu_start(void)
1109{
1110 unsigned long flags;
1111
1112 raw_spin_lock_irqsave(&pmu_lock, flags);
1113 /* Enable all counters */
1114 armv7_pmnc_write(armv7_pmnc_read() | ARMV7_PMNC_E);
1115 raw_spin_unlock_irqrestore(&pmu_lock, flags);
1116}
1117
1118static void armv7pmu_stop(void)
1119{
1120 unsigned long flags;
1121
1122 raw_spin_lock_irqsave(&pmu_lock, flags);
1123 /* Disable all counters */
1124 armv7_pmnc_write(armv7_pmnc_read() & ~ARMV7_PMNC_E);
1125 raw_spin_unlock_irqrestore(&pmu_lock, flags);
1126}
1127
1128static int armv7pmu_get_event_idx(struct cpu_hw_events *cpuc,
1129 struct hw_perf_event *event)
1130{
1131 int idx;
1132
1133 /* Always place a cycle counter into the cycle counter. */
1134 if (event->config_base == ARMV7_PERFCTR_CPU_CYCLES) {
1135 if (test_and_set_bit(ARMV7_CYCLE_COUNTER, cpuc->used_mask))
1136 return -EAGAIN;
1137
1138 return ARMV7_CYCLE_COUNTER;
1139 } else {
1140 /*
1141 * For anything other than a cycle counter, try and use
1142 * the events counters
1143 */
1144 for (idx = ARMV7_COUNTER0; idx <= armpmu->num_events; ++idx) {
1145 if (!test_and_set_bit(idx, cpuc->used_mask))
1146 return idx;
1147 }
1148
1149 /* The counters are all in use. */
1150 return -EAGAIN;
1151 }
1152}
1153
1154static void armv7pmu_reset(void *info)
1155{
1156 u32 idx, nb_cnt = armpmu->num_events;
1157
1158 /* The counter and interrupt enable registers are unknown at reset. */
1159 for (idx = 1; idx < nb_cnt; ++idx)
1160 armv7pmu_disable_event(NULL, idx);
1161
1162 /* Initialize & Reset PMNC: C and P bits */
1163 armv7_pmnc_write(ARMV7_PMNC_P | ARMV7_PMNC_C);
1164}
1165
1166static struct arm_pmu armv7pmu = {
1167 .handle_irq = armv7pmu_handle_irq,
1168 .enable = armv7pmu_enable_event,
1169 .disable = armv7pmu_disable_event,
1170 .read_counter = armv7pmu_read_counter,
1171 .write_counter = armv7pmu_write_counter,
1172 .get_event_idx = armv7pmu_get_event_idx,
1173 .start = armv7pmu_start,
1174 .stop = armv7pmu_stop,
1175 .reset = armv7pmu_reset,
1176 .raw_event_mask = 0xFF,
1177 .max_period = (1LLU << 32) - 1,
1178};
1179
1180static u32 __init armv7_read_num_pmnc_events(void)
1181{
1182 u32 nb_cnt;
1183
1184 /* Read the nb of CNTx counters supported from PMNC */
1185 nb_cnt = (armv7_pmnc_read() >> ARMV7_PMNC_N_SHIFT) & ARMV7_PMNC_N_MASK;
1186
1187 /* Add the CPU cycles counter and return */
1188 return nb_cnt + 1;
1189}
1190
1191static const struct arm_pmu *__init armv7_a8_pmu_init(void)
1192{
1193 armv7pmu.id = ARM_PERF_PMU_ID_CA8;
1194 armv7pmu.name = "ARMv7 Cortex-A8";
1195 armv7pmu.cache_map = &armv7_a8_perf_cache_map;
1196 armv7pmu.event_map = &armv7_a8_perf_map;
1197 armv7pmu.num_events = armv7_read_num_pmnc_events();
1198 return &armv7pmu;
1199}
1200
1201static const struct arm_pmu *__init armv7_a9_pmu_init(void)
1202{
1203 armv7pmu.id = ARM_PERF_PMU_ID_CA9;
1204 armv7pmu.name = "ARMv7 Cortex-A9";
1205 armv7pmu.cache_map = &armv7_a9_perf_cache_map;
1206 armv7pmu.event_map = &armv7_a9_perf_map;
1207 armv7pmu.num_events = armv7_read_num_pmnc_events();
1208 return &armv7pmu;
1209}
1210
1211static const struct arm_pmu *__init armv7_a5_pmu_init(void)
1212{
1213 armv7pmu.id = ARM_PERF_PMU_ID_CA5;
1214 armv7pmu.name = "ARMv7 Cortex-A5";
1215 armv7pmu.cache_map = &armv7_a5_perf_cache_map;
1216 armv7pmu.event_map = &armv7_a5_perf_map;
1217 armv7pmu.num_events = armv7_read_num_pmnc_events();
1218 return &armv7pmu;
1219}
1220
1221static const struct arm_pmu *__init armv7_a15_pmu_init(void)
1222{
1223 armv7pmu.id = ARM_PERF_PMU_ID_CA15;
1224 armv7pmu.name = "ARMv7 Cortex-A15";
1225 armv7pmu.cache_map = &armv7_a15_perf_cache_map;
1226 armv7pmu.event_map = &armv7_a15_perf_map;
1227 armv7pmu.num_events = armv7_read_num_pmnc_events();
1228 return &armv7pmu;
1229}
1230#else
1231static const struct arm_pmu *__init armv7_a8_pmu_init(void)
1232{
1233 return NULL;
1234}
1235
1236static const struct arm_pmu *__init armv7_a9_pmu_init(void)
1237{
1238 return NULL;
1239}
1240
1241static const struct arm_pmu *__init armv7_a5_pmu_init(void)
1242{
1243 return NULL;
1244}
1245
1246static const struct arm_pmu *__init armv7_a15_pmu_init(void)
1247{
1248 return NULL;
1249}
1250#endif /* CONFIG_CPU_V7 */
1/*
2 * ARMv7 Cortex-A8 and Cortex-A9 Performance Events handling code.
3 *
4 * ARMv7 support: Jean Pihet <jpihet@mvista.com>
5 * 2010 (c) MontaVista Software, LLC.
6 *
7 * Copied from ARMv6 code, with the low level code inspired
8 * by the ARMv7 Oprofile code.
9 *
10 * Cortex-A8 has up to 4 configurable performance counters and
11 * a single cycle counter.
12 * Cortex-A9 has up to 31 configurable performance counters and
13 * a single cycle counter.
14 *
15 * All counters can be enabled/disabled and IRQ masked separately. The cycle
16 * counter and all 4 performance counters together can be reset separately.
17 */
18
19#ifdef CONFIG_CPU_V7
20
21#include <asm/cp15.h>
22#include <asm/cputype.h>
23#include <asm/irq_regs.h>
24#include <asm/vfp.h>
25#include "../vfp/vfpinstr.h"
26
27#include <linux/of.h>
28#include <linux/perf/arm_pmu.h>
29#include <linux/platform_device.h>
30
31/*
32 * Common ARMv7 event types
33 *
34 * Note: An implementation may not be able to count all of these events
35 * but the encodings are considered to be `reserved' in the case that
36 * they are not available.
37 */
38#define ARMV7_PERFCTR_PMNC_SW_INCR 0x00
39#define ARMV7_PERFCTR_L1_ICACHE_REFILL 0x01
40#define ARMV7_PERFCTR_ITLB_REFILL 0x02
41#define ARMV7_PERFCTR_L1_DCACHE_REFILL 0x03
42#define ARMV7_PERFCTR_L1_DCACHE_ACCESS 0x04
43#define ARMV7_PERFCTR_DTLB_REFILL 0x05
44#define ARMV7_PERFCTR_MEM_READ 0x06
45#define ARMV7_PERFCTR_MEM_WRITE 0x07
46#define ARMV7_PERFCTR_INSTR_EXECUTED 0x08
47#define ARMV7_PERFCTR_EXC_TAKEN 0x09
48#define ARMV7_PERFCTR_EXC_EXECUTED 0x0A
49#define ARMV7_PERFCTR_CID_WRITE 0x0B
50
51/*
52 * ARMV7_PERFCTR_PC_WRITE is equivalent to HW_BRANCH_INSTRUCTIONS.
53 * It counts:
54 * - all (taken) branch instructions,
55 * - instructions that explicitly write the PC,
56 * - exception generating instructions.
57 */
58#define ARMV7_PERFCTR_PC_WRITE 0x0C
59#define ARMV7_PERFCTR_PC_IMM_BRANCH 0x0D
60#define ARMV7_PERFCTR_PC_PROC_RETURN 0x0E
61#define ARMV7_PERFCTR_MEM_UNALIGNED_ACCESS 0x0F
62#define ARMV7_PERFCTR_PC_BRANCH_MIS_PRED 0x10
63#define ARMV7_PERFCTR_CLOCK_CYCLES 0x11
64#define ARMV7_PERFCTR_PC_BRANCH_PRED 0x12
65
66/* These events are defined by the PMUv2 supplement (ARM DDI 0457A). */
67#define ARMV7_PERFCTR_MEM_ACCESS 0x13
68#define ARMV7_PERFCTR_L1_ICACHE_ACCESS 0x14
69#define ARMV7_PERFCTR_L1_DCACHE_WB 0x15
70#define ARMV7_PERFCTR_L2_CACHE_ACCESS 0x16
71#define ARMV7_PERFCTR_L2_CACHE_REFILL 0x17
72#define ARMV7_PERFCTR_L2_CACHE_WB 0x18
73#define ARMV7_PERFCTR_BUS_ACCESS 0x19
74#define ARMV7_PERFCTR_MEM_ERROR 0x1A
75#define ARMV7_PERFCTR_INSTR_SPEC 0x1B
76#define ARMV7_PERFCTR_TTBR_WRITE 0x1C
77#define ARMV7_PERFCTR_BUS_CYCLES 0x1D
78
79#define ARMV7_PERFCTR_CPU_CYCLES 0xFF
80
81/* ARMv7 Cortex-A8 specific event types */
82#define ARMV7_A8_PERFCTR_L2_CACHE_ACCESS 0x43
83#define ARMV7_A8_PERFCTR_L2_CACHE_REFILL 0x44
84#define ARMV7_A8_PERFCTR_L1_ICACHE_ACCESS 0x50
85#define ARMV7_A8_PERFCTR_STALL_ISIDE 0x56
86
87/* ARMv7 Cortex-A9 specific event types */
88#define ARMV7_A9_PERFCTR_INSTR_CORE_RENAME 0x68
89#define ARMV7_A9_PERFCTR_STALL_ICACHE 0x60
90#define ARMV7_A9_PERFCTR_STALL_DISPATCH 0x66
91
92/* ARMv7 Cortex-A5 specific event types */
93#define ARMV7_A5_PERFCTR_PREFETCH_LINEFILL 0xc2
94#define ARMV7_A5_PERFCTR_PREFETCH_LINEFILL_DROP 0xc3
95
96/* ARMv7 Cortex-A15 specific event types */
97#define ARMV7_A15_PERFCTR_L1_DCACHE_ACCESS_READ 0x40
98#define ARMV7_A15_PERFCTR_L1_DCACHE_ACCESS_WRITE 0x41
99#define ARMV7_A15_PERFCTR_L1_DCACHE_REFILL_READ 0x42
100#define ARMV7_A15_PERFCTR_L1_DCACHE_REFILL_WRITE 0x43
101
102#define ARMV7_A15_PERFCTR_DTLB_REFILL_L1_READ 0x4C
103#define ARMV7_A15_PERFCTR_DTLB_REFILL_L1_WRITE 0x4D
104
105#define ARMV7_A15_PERFCTR_L2_CACHE_ACCESS_READ 0x50
106#define ARMV7_A15_PERFCTR_L2_CACHE_ACCESS_WRITE 0x51
107#define ARMV7_A15_PERFCTR_L2_CACHE_REFILL_READ 0x52
108#define ARMV7_A15_PERFCTR_L2_CACHE_REFILL_WRITE 0x53
109
110#define ARMV7_A15_PERFCTR_PC_WRITE_SPEC 0x76
111
112/* ARMv7 Cortex-A12 specific event types */
113#define ARMV7_A12_PERFCTR_L1_DCACHE_ACCESS_READ 0x40
114#define ARMV7_A12_PERFCTR_L1_DCACHE_ACCESS_WRITE 0x41
115
116#define ARMV7_A12_PERFCTR_L2_CACHE_ACCESS_READ 0x50
117#define ARMV7_A12_PERFCTR_L2_CACHE_ACCESS_WRITE 0x51
118
119#define ARMV7_A12_PERFCTR_PC_WRITE_SPEC 0x76
120
121#define ARMV7_A12_PERFCTR_PF_TLB_REFILL 0xe7
122
123/* ARMv7 Krait specific event types */
124#define KRAIT_PMRESR0_GROUP0 0xcc
125#define KRAIT_PMRESR1_GROUP0 0xd0
126#define KRAIT_PMRESR2_GROUP0 0xd4
127#define KRAIT_VPMRESR0_GROUP0 0xd8
128
129#define KRAIT_PERFCTR_L1_ICACHE_ACCESS 0x10011
130#define KRAIT_PERFCTR_L1_ICACHE_MISS 0x10010
131
132#define KRAIT_PERFCTR_L1_ITLB_ACCESS 0x12222
133#define KRAIT_PERFCTR_L1_DTLB_ACCESS 0x12210
134
135/* ARMv7 Scorpion specific event types */
136#define SCORPION_LPM0_GROUP0 0x4c
137#define SCORPION_LPM1_GROUP0 0x50
138#define SCORPION_LPM2_GROUP0 0x54
139#define SCORPION_L2LPM_GROUP0 0x58
140#define SCORPION_VLPM_GROUP0 0x5c
141
142#define SCORPION_ICACHE_ACCESS 0x10053
143#define SCORPION_ICACHE_MISS 0x10052
144
145#define SCORPION_DTLB_ACCESS 0x12013
146#define SCORPION_DTLB_MISS 0x12012
147
148#define SCORPION_ITLB_MISS 0x12021
149
150/*
151 * Cortex-A8 HW events mapping
152 *
153 * The hardware events that we support. We do support cache operations but
154 * we have harvard caches and no way to combine instruction and data
155 * accesses/misses in hardware.
156 */
157static const unsigned armv7_a8_perf_map[PERF_COUNT_HW_MAX] = {
158 PERF_MAP_ALL_UNSUPPORTED,
159 [PERF_COUNT_HW_CPU_CYCLES] = ARMV7_PERFCTR_CPU_CYCLES,
160 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV7_PERFCTR_INSTR_EXECUTED,
161 [PERF_COUNT_HW_CACHE_REFERENCES] = ARMV7_PERFCTR_L1_DCACHE_ACCESS,
162 [PERF_COUNT_HW_CACHE_MISSES] = ARMV7_PERFCTR_L1_DCACHE_REFILL,
163 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV7_PERFCTR_PC_WRITE,
164 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
165 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = ARMV7_A8_PERFCTR_STALL_ISIDE,
166};
167
168static const unsigned armv7_a8_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
169 [PERF_COUNT_HW_CACHE_OP_MAX]
170 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
171 PERF_CACHE_MAP_ALL_UNSUPPORTED,
172
173 /*
174 * The performance counters don't differentiate between read and write
175 * accesses/misses so this isn't strictly correct, but it's the best we
176 * can do. Writes and reads get combined.
177 */
178 [C(L1D)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_DCACHE_ACCESS,
179 [C(L1D)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_L1_DCACHE_REFILL,
180 [C(L1D)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_DCACHE_ACCESS,
181 [C(L1D)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_L1_DCACHE_REFILL,
182
183 [C(L1I)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_A8_PERFCTR_L1_ICACHE_ACCESS,
184 [C(L1I)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_L1_ICACHE_REFILL,
185
186 [C(LL)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_A8_PERFCTR_L2_CACHE_ACCESS,
187 [C(LL)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_A8_PERFCTR_L2_CACHE_REFILL,
188 [C(LL)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV7_A8_PERFCTR_L2_CACHE_ACCESS,
189 [C(LL)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_A8_PERFCTR_L2_CACHE_REFILL,
190
191 [C(DTLB)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
192 [C(DTLB)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
193
194 [C(ITLB)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_REFILL,
195 [C(ITLB)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_REFILL,
196
197 [C(BPU)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED,
198 [C(BPU)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
199 [C(BPU)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED,
200 [C(BPU)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
201};
202
203/*
204 * Cortex-A9 HW events mapping
205 */
206static const unsigned armv7_a9_perf_map[PERF_COUNT_HW_MAX] = {
207 PERF_MAP_ALL_UNSUPPORTED,
208 [PERF_COUNT_HW_CPU_CYCLES] = ARMV7_PERFCTR_CPU_CYCLES,
209 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV7_A9_PERFCTR_INSTR_CORE_RENAME,
210 [PERF_COUNT_HW_CACHE_REFERENCES] = ARMV7_PERFCTR_L1_DCACHE_ACCESS,
211 [PERF_COUNT_HW_CACHE_MISSES] = ARMV7_PERFCTR_L1_DCACHE_REFILL,
212 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV7_PERFCTR_PC_WRITE,
213 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
214 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = ARMV7_A9_PERFCTR_STALL_ICACHE,
215 [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = ARMV7_A9_PERFCTR_STALL_DISPATCH,
216};
217
218static const unsigned armv7_a9_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
219 [PERF_COUNT_HW_CACHE_OP_MAX]
220 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
221 PERF_CACHE_MAP_ALL_UNSUPPORTED,
222
223 /*
224 * The performance counters don't differentiate between read and write
225 * accesses/misses so this isn't strictly correct, but it's the best we
226 * can do. Writes and reads get combined.
227 */
228 [C(L1D)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_DCACHE_ACCESS,
229 [C(L1D)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_L1_DCACHE_REFILL,
230 [C(L1D)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_DCACHE_ACCESS,
231 [C(L1D)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_L1_DCACHE_REFILL,
232
233 [C(L1I)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_L1_ICACHE_REFILL,
234
235 [C(DTLB)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
236 [C(DTLB)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
237
238 [C(ITLB)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_REFILL,
239 [C(ITLB)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_REFILL,
240
241 [C(BPU)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED,
242 [C(BPU)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
243 [C(BPU)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED,
244 [C(BPU)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
245};
246
247/*
248 * Cortex-A5 HW events mapping
249 */
250static const unsigned armv7_a5_perf_map[PERF_COUNT_HW_MAX] = {
251 PERF_MAP_ALL_UNSUPPORTED,
252 [PERF_COUNT_HW_CPU_CYCLES] = ARMV7_PERFCTR_CPU_CYCLES,
253 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV7_PERFCTR_INSTR_EXECUTED,
254 [PERF_COUNT_HW_CACHE_REFERENCES] = ARMV7_PERFCTR_L1_DCACHE_ACCESS,
255 [PERF_COUNT_HW_CACHE_MISSES] = ARMV7_PERFCTR_L1_DCACHE_REFILL,
256 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV7_PERFCTR_PC_WRITE,
257 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
258};
259
260static const unsigned armv7_a5_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
261 [PERF_COUNT_HW_CACHE_OP_MAX]
262 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
263 PERF_CACHE_MAP_ALL_UNSUPPORTED,
264
265 [C(L1D)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_DCACHE_ACCESS,
266 [C(L1D)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_L1_DCACHE_REFILL,
267 [C(L1D)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_DCACHE_ACCESS,
268 [C(L1D)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_L1_DCACHE_REFILL,
269 [C(L1D)][C(OP_PREFETCH)][C(RESULT_ACCESS)] = ARMV7_A5_PERFCTR_PREFETCH_LINEFILL,
270 [C(L1D)][C(OP_PREFETCH)][C(RESULT_MISS)] = ARMV7_A5_PERFCTR_PREFETCH_LINEFILL_DROP,
271
272 [C(L1I)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_ICACHE_ACCESS,
273 [C(L1I)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_L1_ICACHE_REFILL,
274 /*
275 * The prefetch counters don't differentiate between the I side and the
276 * D side.
277 */
278 [C(L1I)][C(OP_PREFETCH)][C(RESULT_ACCESS)] = ARMV7_A5_PERFCTR_PREFETCH_LINEFILL,
279 [C(L1I)][C(OP_PREFETCH)][C(RESULT_MISS)] = ARMV7_A5_PERFCTR_PREFETCH_LINEFILL_DROP,
280
281 [C(DTLB)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
282 [C(DTLB)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
283
284 [C(ITLB)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_REFILL,
285 [C(ITLB)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_REFILL,
286
287 [C(BPU)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED,
288 [C(BPU)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
289 [C(BPU)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED,
290 [C(BPU)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
291};
292
293/*
294 * Cortex-A15 HW events mapping
295 */
296static const unsigned armv7_a15_perf_map[PERF_COUNT_HW_MAX] = {
297 PERF_MAP_ALL_UNSUPPORTED,
298 [PERF_COUNT_HW_CPU_CYCLES] = ARMV7_PERFCTR_CPU_CYCLES,
299 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV7_PERFCTR_INSTR_EXECUTED,
300 [PERF_COUNT_HW_CACHE_REFERENCES] = ARMV7_PERFCTR_L1_DCACHE_ACCESS,
301 [PERF_COUNT_HW_CACHE_MISSES] = ARMV7_PERFCTR_L1_DCACHE_REFILL,
302 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV7_A15_PERFCTR_PC_WRITE_SPEC,
303 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
304 [PERF_COUNT_HW_BUS_CYCLES] = ARMV7_PERFCTR_BUS_CYCLES,
305};
306
307static const unsigned armv7_a15_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
308 [PERF_COUNT_HW_CACHE_OP_MAX]
309 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
310 PERF_CACHE_MAP_ALL_UNSUPPORTED,
311
312 [C(L1D)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_A15_PERFCTR_L1_DCACHE_ACCESS_READ,
313 [C(L1D)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_A15_PERFCTR_L1_DCACHE_REFILL_READ,
314 [C(L1D)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV7_A15_PERFCTR_L1_DCACHE_ACCESS_WRITE,
315 [C(L1D)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_A15_PERFCTR_L1_DCACHE_REFILL_WRITE,
316
317 /*
318 * Not all performance counters differentiate between read and write
319 * accesses/misses so we're not always strictly correct, but it's the
320 * best we can do. Writes and reads get combined in these cases.
321 */
322 [C(L1I)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_ICACHE_ACCESS,
323 [C(L1I)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_L1_ICACHE_REFILL,
324
325 [C(LL)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_A15_PERFCTR_L2_CACHE_ACCESS_READ,
326 [C(LL)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_A15_PERFCTR_L2_CACHE_REFILL_READ,
327 [C(LL)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV7_A15_PERFCTR_L2_CACHE_ACCESS_WRITE,
328 [C(LL)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_A15_PERFCTR_L2_CACHE_REFILL_WRITE,
329
330 [C(DTLB)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_A15_PERFCTR_DTLB_REFILL_L1_READ,
331 [C(DTLB)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_A15_PERFCTR_DTLB_REFILL_L1_WRITE,
332
333 [C(ITLB)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_REFILL,
334 [C(ITLB)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_REFILL,
335
336 [C(BPU)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED,
337 [C(BPU)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
338 [C(BPU)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED,
339 [C(BPU)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
340};
341
342/*
343 * Cortex-A7 HW events mapping
344 */
345static const unsigned armv7_a7_perf_map[PERF_COUNT_HW_MAX] = {
346 PERF_MAP_ALL_UNSUPPORTED,
347 [PERF_COUNT_HW_CPU_CYCLES] = ARMV7_PERFCTR_CPU_CYCLES,
348 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV7_PERFCTR_INSTR_EXECUTED,
349 [PERF_COUNT_HW_CACHE_REFERENCES] = ARMV7_PERFCTR_L1_DCACHE_ACCESS,
350 [PERF_COUNT_HW_CACHE_MISSES] = ARMV7_PERFCTR_L1_DCACHE_REFILL,
351 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV7_PERFCTR_PC_WRITE,
352 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
353 [PERF_COUNT_HW_BUS_CYCLES] = ARMV7_PERFCTR_BUS_CYCLES,
354};
355
356static const unsigned armv7_a7_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
357 [PERF_COUNT_HW_CACHE_OP_MAX]
358 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
359 PERF_CACHE_MAP_ALL_UNSUPPORTED,
360
361 /*
362 * The performance counters don't differentiate between read and write
363 * accesses/misses so this isn't strictly correct, but it's the best we
364 * can do. Writes and reads get combined.
365 */
366 [C(L1D)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_DCACHE_ACCESS,
367 [C(L1D)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_L1_DCACHE_REFILL,
368 [C(L1D)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_DCACHE_ACCESS,
369 [C(L1D)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_L1_DCACHE_REFILL,
370
371 [C(L1I)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_ICACHE_ACCESS,
372 [C(L1I)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_L1_ICACHE_REFILL,
373
374 [C(LL)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_L2_CACHE_ACCESS,
375 [C(LL)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_L2_CACHE_REFILL,
376 [C(LL)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_L2_CACHE_ACCESS,
377 [C(LL)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_L2_CACHE_REFILL,
378
379 [C(DTLB)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
380 [C(DTLB)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
381
382 [C(ITLB)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_REFILL,
383 [C(ITLB)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_REFILL,
384
385 [C(BPU)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED,
386 [C(BPU)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
387 [C(BPU)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED,
388 [C(BPU)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
389};
390
391/*
392 * Cortex-A12 HW events mapping
393 */
394static const unsigned armv7_a12_perf_map[PERF_COUNT_HW_MAX] = {
395 PERF_MAP_ALL_UNSUPPORTED,
396 [PERF_COUNT_HW_CPU_CYCLES] = ARMV7_PERFCTR_CPU_CYCLES,
397 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV7_PERFCTR_INSTR_EXECUTED,
398 [PERF_COUNT_HW_CACHE_REFERENCES] = ARMV7_PERFCTR_L1_DCACHE_ACCESS,
399 [PERF_COUNT_HW_CACHE_MISSES] = ARMV7_PERFCTR_L1_DCACHE_REFILL,
400 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV7_A12_PERFCTR_PC_WRITE_SPEC,
401 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
402 [PERF_COUNT_HW_BUS_CYCLES] = ARMV7_PERFCTR_BUS_CYCLES,
403};
404
405static const unsigned armv7_a12_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
406 [PERF_COUNT_HW_CACHE_OP_MAX]
407 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
408 PERF_CACHE_MAP_ALL_UNSUPPORTED,
409
410 [C(L1D)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_A12_PERFCTR_L1_DCACHE_ACCESS_READ,
411 [C(L1D)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_L1_DCACHE_REFILL,
412 [C(L1D)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV7_A12_PERFCTR_L1_DCACHE_ACCESS_WRITE,
413 [C(L1D)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_L1_DCACHE_REFILL,
414
415 /*
416 * Not all performance counters differentiate between read and write
417 * accesses/misses so we're not always strictly correct, but it's the
418 * best we can do. Writes and reads get combined in these cases.
419 */
420 [C(L1I)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_ICACHE_ACCESS,
421 [C(L1I)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_L1_ICACHE_REFILL,
422
423 [C(LL)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_A12_PERFCTR_L2_CACHE_ACCESS_READ,
424 [C(LL)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_L2_CACHE_REFILL,
425 [C(LL)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV7_A12_PERFCTR_L2_CACHE_ACCESS_WRITE,
426 [C(LL)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_L2_CACHE_REFILL,
427
428 [C(DTLB)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
429 [C(DTLB)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
430 [C(DTLB)][C(OP_PREFETCH)][C(RESULT_MISS)] = ARMV7_A12_PERFCTR_PF_TLB_REFILL,
431
432 [C(ITLB)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_REFILL,
433 [C(ITLB)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_REFILL,
434
435 [C(BPU)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED,
436 [C(BPU)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
437 [C(BPU)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED,
438 [C(BPU)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
439};
440
441/*
442 * Krait HW events mapping
443 */
444static const unsigned krait_perf_map[PERF_COUNT_HW_MAX] = {
445 PERF_MAP_ALL_UNSUPPORTED,
446 [PERF_COUNT_HW_CPU_CYCLES] = ARMV7_PERFCTR_CPU_CYCLES,
447 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV7_PERFCTR_INSTR_EXECUTED,
448 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV7_PERFCTR_PC_WRITE,
449 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
450 [PERF_COUNT_HW_BUS_CYCLES] = ARMV7_PERFCTR_CLOCK_CYCLES,
451};
452
453static const unsigned krait_perf_map_no_branch[PERF_COUNT_HW_MAX] = {
454 PERF_MAP_ALL_UNSUPPORTED,
455 [PERF_COUNT_HW_CPU_CYCLES] = ARMV7_PERFCTR_CPU_CYCLES,
456 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV7_PERFCTR_INSTR_EXECUTED,
457 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
458 [PERF_COUNT_HW_BUS_CYCLES] = ARMV7_PERFCTR_CLOCK_CYCLES,
459};
460
461static const unsigned krait_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
462 [PERF_COUNT_HW_CACHE_OP_MAX]
463 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
464 PERF_CACHE_MAP_ALL_UNSUPPORTED,
465
466 /*
467 * The performance counters don't differentiate between read and write
468 * accesses/misses so this isn't strictly correct, but it's the best we
469 * can do. Writes and reads get combined.
470 */
471 [C(L1D)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_DCACHE_ACCESS,
472 [C(L1D)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_L1_DCACHE_REFILL,
473 [C(L1D)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_DCACHE_ACCESS,
474 [C(L1D)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_L1_DCACHE_REFILL,
475
476 [C(L1I)][C(OP_READ)][C(RESULT_ACCESS)] = KRAIT_PERFCTR_L1_ICACHE_ACCESS,
477 [C(L1I)][C(OP_READ)][C(RESULT_MISS)] = KRAIT_PERFCTR_L1_ICACHE_MISS,
478
479 [C(DTLB)][C(OP_READ)][C(RESULT_ACCESS)] = KRAIT_PERFCTR_L1_DTLB_ACCESS,
480 [C(DTLB)][C(OP_WRITE)][C(RESULT_ACCESS)] = KRAIT_PERFCTR_L1_DTLB_ACCESS,
481
482 [C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = KRAIT_PERFCTR_L1_ITLB_ACCESS,
483 [C(ITLB)][C(OP_WRITE)][C(RESULT_ACCESS)] = KRAIT_PERFCTR_L1_ITLB_ACCESS,
484
485 [C(BPU)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED,
486 [C(BPU)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
487 [C(BPU)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED,
488 [C(BPU)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
489};
490
491/*
492 * Scorpion HW events mapping
493 */
494static const unsigned scorpion_perf_map[PERF_COUNT_HW_MAX] = {
495 PERF_MAP_ALL_UNSUPPORTED,
496 [PERF_COUNT_HW_CPU_CYCLES] = ARMV7_PERFCTR_CPU_CYCLES,
497 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV7_PERFCTR_INSTR_EXECUTED,
498 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV7_PERFCTR_PC_WRITE,
499 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
500 [PERF_COUNT_HW_BUS_CYCLES] = ARMV7_PERFCTR_CLOCK_CYCLES,
501};
502
503static const unsigned scorpion_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
504 [PERF_COUNT_HW_CACHE_OP_MAX]
505 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
506 PERF_CACHE_MAP_ALL_UNSUPPORTED,
507 /*
508 * The performance counters don't differentiate between read and write
509 * accesses/misses so this isn't strictly correct, but it's the best we
510 * can do. Writes and reads get combined.
511 */
512 [C(L1D)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_DCACHE_ACCESS,
513 [C(L1D)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_L1_DCACHE_REFILL,
514 [C(L1D)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_DCACHE_ACCESS,
515 [C(L1D)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_L1_DCACHE_REFILL,
516 [C(L1I)][C(OP_READ)][C(RESULT_ACCESS)] = SCORPION_ICACHE_ACCESS,
517 [C(L1I)][C(OP_READ)][C(RESULT_MISS)] = SCORPION_ICACHE_MISS,
518 /*
519 * Only ITLB misses and DTLB refills are supported. If users want the
520 * DTLB refills misses a raw counter must be used.
521 */
522 [C(DTLB)][C(OP_READ)][C(RESULT_ACCESS)] = SCORPION_DTLB_ACCESS,
523 [C(DTLB)][C(OP_READ)][C(RESULT_MISS)] = SCORPION_DTLB_MISS,
524 [C(DTLB)][C(OP_WRITE)][C(RESULT_ACCESS)] = SCORPION_DTLB_ACCESS,
525 [C(DTLB)][C(OP_WRITE)][C(RESULT_MISS)] = SCORPION_DTLB_MISS,
526 [C(ITLB)][C(OP_READ)][C(RESULT_MISS)] = SCORPION_ITLB_MISS,
527 [C(ITLB)][C(OP_WRITE)][C(RESULT_MISS)] = SCORPION_ITLB_MISS,
528 [C(BPU)][C(OP_READ)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED,
529 [C(BPU)][C(OP_READ)][C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
530 [C(BPU)][C(OP_WRITE)][C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED,
531 [C(BPU)][C(OP_WRITE)][C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
532};
533
534PMU_FORMAT_ATTR(event, "config:0-7");
535
536static struct attribute *armv7_pmu_format_attrs[] = {
537 &format_attr_event.attr,
538 NULL,
539};
540
541static struct attribute_group armv7_pmu_format_attr_group = {
542 .name = "format",
543 .attrs = armv7_pmu_format_attrs,
544};
545
546#define ARMV7_EVENT_ATTR_RESOLVE(m) #m
547#define ARMV7_EVENT_ATTR(name, config) \
548 PMU_EVENT_ATTR_STRING(name, armv7_event_attr_##name, \
549 "event=" ARMV7_EVENT_ATTR_RESOLVE(config))
550
551ARMV7_EVENT_ATTR(sw_incr, ARMV7_PERFCTR_PMNC_SW_INCR);
552ARMV7_EVENT_ATTR(l1i_cache_refill, ARMV7_PERFCTR_L1_ICACHE_REFILL);
553ARMV7_EVENT_ATTR(l1i_tlb_refill, ARMV7_PERFCTR_ITLB_REFILL);
554ARMV7_EVENT_ATTR(l1d_cache_refill, ARMV7_PERFCTR_L1_DCACHE_REFILL);
555ARMV7_EVENT_ATTR(l1d_cache, ARMV7_PERFCTR_L1_DCACHE_ACCESS);
556ARMV7_EVENT_ATTR(l1d_tlb_refill, ARMV7_PERFCTR_DTLB_REFILL);
557ARMV7_EVENT_ATTR(ld_retired, ARMV7_PERFCTR_MEM_READ);
558ARMV7_EVENT_ATTR(st_retired, ARMV7_PERFCTR_MEM_WRITE);
559ARMV7_EVENT_ATTR(inst_retired, ARMV7_PERFCTR_INSTR_EXECUTED);
560ARMV7_EVENT_ATTR(exc_taken, ARMV7_PERFCTR_EXC_TAKEN);
561ARMV7_EVENT_ATTR(exc_return, ARMV7_PERFCTR_EXC_EXECUTED);
562ARMV7_EVENT_ATTR(cid_write_retired, ARMV7_PERFCTR_CID_WRITE);
563ARMV7_EVENT_ATTR(pc_write_retired, ARMV7_PERFCTR_PC_WRITE);
564ARMV7_EVENT_ATTR(br_immed_retired, ARMV7_PERFCTR_PC_IMM_BRANCH);
565ARMV7_EVENT_ATTR(br_return_retired, ARMV7_PERFCTR_PC_PROC_RETURN);
566ARMV7_EVENT_ATTR(unaligned_ldst_retired, ARMV7_PERFCTR_MEM_UNALIGNED_ACCESS);
567ARMV7_EVENT_ATTR(br_mis_pred, ARMV7_PERFCTR_PC_BRANCH_MIS_PRED);
568ARMV7_EVENT_ATTR(cpu_cycles, ARMV7_PERFCTR_CLOCK_CYCLES);
569ARMV7_EVENT_ATTR(br_pred, ARMV7_PERFCTR_PC_BRANCH_PRED);
570
571static struct attribute *armv7_pmuv1_event_attrs[] = {
572 &armv7_event_attr_sw_incr.attr.attr,
573 &armv7_event_attr_l1i_cache_refill.attr.attr,
574 &armv7_event_attr_l1i_tlb_refill.attr.attr,
575 &armv7_event_attr_l1d_cache_refill.attr.attr,
576 &armv7_event_attr_l1d_cache.attr.attr,
577 &armv7_event_attr_l1d_tlb_refill.attr.attr,
578 &armv7_event_attr_ld_retired.attr.attr,
579 &armv7_event_attr_st_retired.attr.attr,
580 &armv7_event_attr_inst_retired.attr.attr,
581 &armv7_event_attr_exc_taken.attr.attr,
582 &armv7_event_attr_exc_return.attr.attr,
583 &armv7_event_attr_cid_write_retired.attr.attr,
584 &armv7_event_attr_pc_write_retired.attr.attr,
585 &armv7_event_attr_br_immed_retired.attr.attr,
586 &armv7_event_attr_br_return_retired.attr.attr,
587 &armv7_event_attr_unaligned_ldst_retired.attr.attr,
588 &armv7_event_attr_br_mis_pred.attr.attr,
589 &armv7_event_attr_cpu_cycles.attr.attr,
590 &armv7_event_attr_br_pred.attr.attr,
591 NULL,
592};
593
594static struct attribute_group armv7_pmuv1_events_attr_group = {
595 .name = "events",
596 .attrs = armv7_pmuv1_event_attrs,
597};
598
599static const struct attribute_group *armv7_pmuv1_attr_groups[] = {
600 &armv7_pmuv1_events_attr_group,
601 &armv7_pmu_format_attr_group,
602 NULL,
603};
604
605ARMV7_EVENT_ATTR(mem_access, ARMV7_PERFCTR_MEM_ACCESS);
606ARMV7_EVENT_ATTR(l1i_cache, ARMV7_PERFCTR_L1_ICACHE_ACCESS);
607ARMV7_EVENT_ATTR(l1d_cache_wb, ARMV7_PERFCTR_L1_DCACHE_WB);
608ARMV7_EVENT_ATTR(l2d_cache, ARMV7_PERFCTR_L2_CACHE_ACCESS);
609ARMV7_EVENT_ATTR(l2d_cache_refill, ARMV7_PERFCTR_L2_CACHE_REFILL);
610ARMV7_EVENT_ATTR(l2d_cache_wb, ARMV7_PERFCTR_L2_CACHE_WB);
611ARMV7_EVENT_ATTR(bus_access, ARMV7_PERFCTR_BUS_ACCESS);
612ARMV7_EVENT_ATTR(memory_error, ARMV7_PERFCTR_MEM_ERROR);
613ARMV7_EVENT_ATTR(inst_spec, ARMV7_PERFCTR_INSTR_SPEC);
614ARMV7_EVENT_ATTR(ttbr_write_retired, ARMV7_PERFCTR_TTBR_WRITE);
615ARMV7_EVENT_ATTR(bus_cycles, ARMV7_PERFCTR_BUS_CYCLES);
616
617static struct attribute *armv7_pmuv2_event_attrs[] = {
618 &armv7_event_attr_sw_incr.attr.attr,
619 &armv7_event_attr_l1i_cache_refill.attr.attr,
620 &armv7_event_attr_l1i_tlb_refill.attr.attr,
621 &armv7_event_attr_l1d_cache_refill.attr.attr,
622 &armv7_event_attr_l1d_cache.attr.attr,
623 &armv7_event_attr_l1d_tlb_refill.attr.attr,
624 &armv7_event_attr_ld_retired.attr.attr,
625 &armv7_event_attr_st_retired.attr.attr,
626 &armv7_event_attr_inst_retired.attr.attr,
627 &armv7_event_attr_exc_taken.attr.attr,
628 &armv7_event_attr_exc_return.attr.attr,
629 &armv7_event_attr_cid_write_retired.attr.attr,
630 &armv7_event_attr_pc_write_retired.attr.attr,
631 &armv7_event_attr_br_immed_retired.attr.attr,
632 &armv7_event_attr_br_return_retired.attr.attr,
633 &armv7_event_attr_unaligned_ldst_retired.attr.attr,
634 &armv7_event_attr_br_mis_pred.attr.attr,
635 &armv7_event_attr_cpu_cycles.attr.attr,
636 &armv7_event_attr_br_pred.attr.attr,
637 &armv7_event_attr_mem_access.attr.attr,
638 &armv7_event_attr_l1i_cache.attr.attr,
639 &armv7_event_attr_l1d_cache_wb.attr.attr,
640 &armv7_event_attr_l2d_cache.attr.attr,
641 &armv7_event_attr_l2d_cache_refill.attr.attr,
642 &armv7_event_attr_l2d_cache_wb.attr.attr,
643 &armv7_event_attr_bus_access.attr.attr,
644 &armv7_event_attr_memory_error.attr.attr,
645 &armv7_event_attr_inst_spec.attr.attr,
646 &armv7_event_attr_ttbr_write_retired.attr.attr,
647 &armv7_event_attr_bus_cycles.attr.attr,
648 NULL,
649};
650
651static struct attribute_group armv7_pmuv2_events_attr_group = {
652 .name = "events",
653 .attrs = armv7_pmuv2_event_attrs,
654};
655
656static const struct attribute_group *armv7_pmuv2_attr_groups[] = {
657 &armv7_pmuv2_events_attr_group,
658 &armv7_pmu_format_attr_group,
659 NULL,
660};
661
662/*
663 * Perf Events' indices
664 */
665#define ARMV7_IDX_CYCLE_COUNTER 0
666#define ARMV7_IDX_COUNTER0 1
667#define ARMV7_IDX_COUNTER_LAST(cpu_pmu) \
668 (ARMV7_IDX_CYCLE_COUNTER + cpu_pmu->num_events - 1)
669
670#define ARMV7_MAX_COUNTERS 32
671#define ARMV7_COUNTER_MASK (ARMV7_MAX_COUNTERS - 1)
672
673/*
674 * ARMv7 low level PMNC access
675 */
676
677/*
678 * Perf Event to low level counters mapping
679 */
680#define ARMV7_IDX_TO_COUNTER(x) \
681 (((x) - ARMV7_IDX_COUNTER0) & ARMV7_COUNTER_MASK)
682
683/*
684 * Per-CPU PMNC: config reg
685 */
686#define ARMV7_PMNC_E (1 << 0) /* Enable all counters */
687#define ARMV7_PMNC_P (1 << 1) /* Reset all counters */
688#define ARMV7_PMNC_C (1 << 2) /* Cycle counter reset */
689#define ARMV7_PMNC_D (1 << 3) /* CCNT counts every 64th cpu cycle */
690#define ARMV7_PMNC_X (1 << 4) /* Export to ETM */
691#define ARMV7_PMNC_DP (1 << 5) /* Disable CCNT if non-invasive debug*/
692#define ARMV7_PMNC_N_SHIFT 11 /* Number of counters supported */
693#define ARMV7_PMNC_N_MASK 0x1f
694#define ARMV7_PMNC_MASK 0x3f /* Mask for writable bits */
695
696/*
697 * FLAG: counters overflow flag status reg
698 */
699#define ARMV7_FLAG_MASK 0xffffffff /* Mask for writable bits */
700#define ARMV7_OVERFLOWED_MASK ARMV7_FLAG_MASK
701
702/*
703 * PMXEVTYPER: Event selection reg
704 */
705#define ARMV7_EVTYPE_MASK 0xc80000ff /* Mask for writable bits */
706#define ARMV7_EVTYPE_EVENT 0xff /* Mask for EVENT bits */
707
708/*
709 * Event filters for PMUv2
710 */
711#define ARMV7_EXCLUDE_PL1 (1 << 31)
712#define ARMV7_EXCLUDE_USER (1 << 30)
713#define ARMV7_INCLUDE_HYP (1 << 27)
714
715/*
716 * Secure debug enable reg
717 */
718#define ARMV7_SDER_SUNIDEN BIT(1) /* Permit non-invasive debug */
719
720static inline u32 armv7_pmnc_read(void)
721{
722 u32 val;
723 asm volatile("mrc p15, 0, %0, c9, c12, 0" : "=r"(val));
724 return val;
725}
726
727static inline void armv7_pmnc_write(u32 val)
728{
729 val &= ARMV7_PMNC_MASK;
730 isb();
731 asm volatile("mcr p15, 0, %0, c9, c12, 0" : : "r"(val));
732}
733
734static inline int armv7_pmnc_has_overflowed(u32 pmnc)
735{
736 return pmnc & ARMV7_OVERFLOWED_MASK;
737}
738
739static inline int armv7_pmnc_counter_valid(struct arm_pmu *cpu_pmu, int idx)
740{
741 return idx >= ARMV7_IDX_CYCLE_COUNTER &&
742 idx <= ARMV7_IDX_COUNTER_LAST(cpu_pmu);
743}
744
745static inline int armv7_pmnc_counter_has_overflowed(u32 pmnc, int idx)
746{
747 return pmnc & BIT(ARMV7_IDX_TO_COUNTER(idx));
748}
749
750static inline void armv7_pmnc_select_counter(int idx)
751{
752 u32 counter = ARMV7_IDX_TO_COUNTER(idx);
753 asm volatile("mcr p15, 0, %0, c9, c12, 5" : : "r" (counter));
754 isb();
755}
756
757static inline u32 armv7pmu_read_counter(struct perf_event *event)
758{
759 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
760 struct hw_perf_event *hwc = &event->hw;
761 int idx = hwc->idx;
762 u32 value = 0;
763
764 if (!armv7_pmnc_counter_valid(cpu_pmu, idx)) {
765 pr_err("CPU%u reading wrong counter %d\n",
766 smp_processor_id(), idx);
767 } else if (idx == ARMV7_IDX_CYCLE_COUNTER) {
768 asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r" (value));
769 } else {
770 armv7_pmnc_select_counter(idx);
771 asm volatile("mrc p15, 0, %0, c9, c13, 2" : "=r" (value));
772 }
773
774 return value;
775}
776
777static inline void armv7pmu_write_counter(struct perf_event *event, u32 value)
778{
779 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
780 struct hw_perf_event *hwc = &event->hw;
781 int idx = hwc->idx;
782
783 if (!armv7_pmnc_counter_valid(cpu_pmu, idx)) {
784 pr_err("CPU%u writing wrong counter %d\n",
785 smp_processor_id(), idx);
786 } else if (idx == ARMV7_IDX_CYCLE_COUNTER) {
787 asm volatile("mcr p15, 0, %0, c9, c13, 0" : : "r" (value));
788 } else {
789 armv7_pmnc_select_counter(idx);
790 asm volatile("mcr p15, 0, %0, c9, c13, 2" : : "r" (value));
791 }
792}
793
794static inline void armv7_pmnc_write_evtsel(int idx, u32 val)
795{
796 armv7_pmnc_select_counter(idx);
797 val &= ARMV7_EVTYPE_MASK;
798 asm volatile("mcr p15, 0, %0, c9, c13, 1" : : "r" (val));
799}
800
801static inline void armv7_pmnc_enable_counter(int idx)
802{
803 u32 counter = ARMV7_IDX_TO_COUNTER(idx);
804 asm volatile("mcr p15, 0, %0, c9, c12, 1" : : "r" (BIT(counter)));
805}
806
807static inline void armv7_pmnc_disable_counter(int idx)
808{
809 u32 counter = ARMV7_IDX_TO_COUNTER(idx);
810 asm volatile("mcr p15, 0, %0, c9, c12, 2" : : "r" (BIT(counter)));
811}
812
813static inline void armv7_pmnc_enable_intens(int idx)
814{
815 u32 counter = ARMV7_IDX_TO_COUNTER(idx);
816 asm volatile("mcr p15, 0, %0, c9, c14, 1" : : "r" (BIT(counter)));
817}
818
819static inline void armv7_pmnc_disable_intens(int idx)
820{
821 u32 counter = ARMV7_IDX_TO_COUNTER(idx);
822 asm volatile("mcr p15, 0, %0, c9, c14, 2" : : "r" (BIT(counter)));
823 isb();
824 /* Clear the overflow flag in case an interrupt is pending. */
825 asm volatile("mcr p15, 0, %0, c9, c12, 3" : : "r" (BIT(counter)));
826 isb();
827}
828
829static inline u32 armv7_pmnc_getreset_flags(void)
830{
831 u32 val;
832
833 /* Read */
834 asm volatile("mrc p15, 0, %0, c9, c12, 3" : "=r" (val));
835
836 /* Write to clear flags */
837 val &= ARMV7_FLAG_MASK;
838 asm volatile("mcr p15, 0, %0, c9, c12, 3" : : "r" (val));
839
840 return val;
841}
842
843#ifdef DEBUG
844static void armv7_pmnc_dump_regs(struct arm_pmu *cpu_pmu)
845{
846 u32 val;
847 unsigned int cnt;
848
849 pr_info("PMNC registers dump:\n");
850
851 asm volatile("mrc p15, 0, %0, c9, c12, 0" : "=r" (val));
852 pr_info("PMNC =0x%08x\n", val);
853
854 asm volatile("mrc p15, 0, %0, c9, c12, 1" : "=r" (val));
855 pr_info("CNTENS=0x%08x\n", val);
856
857 asm volatile("mrc p15, 0, %0, c9, c14, 1" : "=r" (val));
858 pr_info("INTENS=0x%08x\n", val);
859
860 asm volatile("mrc p15, 0, %0, c9, c12, 3" : "=r" (val));
861 pr_info("FLAGS =0x%08x\n", val);
862
863 asm volatile("mrc p15, 0, %0, c9, c12, 5" : "=r" (val));
864 pr_info("SELECT=0x%08x\n", val);
865
866 asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r" (val));
867 pr_info("CCNT =0x%08x\n", val);
868
869 for (cnt = ARMV7_IDX_COUNTER0;
870 cnt <= ARMV7_IDX_COUNTER_LAST(cpu_pmu); cnt++) {
871 armv7_pmnc_select_counter(cnt);
872 asm volatile("mrc p15, 0, %0, c9, c13, 2" : "=r" (val));
873 pr_info("CNT[%d] count =0x%08x\n",
874 ARMV7_IDX_TO_COUNTER(cnt), val);
875 asm volatile("mrc p15, 0, %0, c9, c13, 1" : "=r" (val));
876 pr_info("CNT[%d] evtsel=0x%08x\n",
877 ARMV7_IDX_TO_COUNTER(cnt), val);
878 }
879}
880#endif
881
882static void armv7pmu_enable_event(struct perf_event *event)
883{
884 unsigned long flags;
885 struct hw_perf_event *hwc = &event->hw;
886 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
887 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
888 int idx = hwc->idx;
889
890 if (!armv7_pmnc_counter_valid(cpu_pmu, idx)) {
891 pr_err("CPU%u enabling wrong PMNC counter IRQ enable %d\n",
892 smp_processor_id(), idx);
893 return;
894 }
895
896 /*
897 * Enable counter and interrupt, and set the counter to count
898 * the event that we're interested in.
899 */
900 raw_spin_lock_irqsave(&events->pmu_lock, flags);
901
902 /*
903 * Disable counter
904 */
905 armv7_pmnc_disable_counter(idx);
906
907 /*
908 * Set event (if destined for PMNx counters)
909 * We only need to set the event for the cycle counter if we
910 * have the ability to perform event filtering.
911 */
912 if (cpu_pmu->set_event_filter || idx != ARMV7_IDX_CYCLE_COUNTER)
913 armv7_pmnc_write_evtsel(idx, hwc->config_base);
914
915 /*
916 * Enable interrupt for this counter
917 */
918 armv7_pmnc_enable_intens(idx);
919
920 /*
921 * Enable counter
922 */
923 armv7_pmnc_enable_counter(idx);
924
925 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
926}
927
928static void armv7pmu_disable_event(struct perf_event *event)
929{
930 unsigned long flags;
931 struct hw_perf_event *hwc = &event->hw;
932 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
933 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
934 int idx = hwc->idx;
935
936 if (!armv7_pmnc_counter_valid(cpu_pmu, idx)) {
937 pr_err("CPU%u disabling wrong PMNC counter IRQ enable %d\n",
938 smp_processor_id(), idx);
939 return;
940 }
941
942 /*
943 * Disable counter and interrupt
944 */
945 raw_spin_lock_irqsave(&events->pmu_lock, flags);
946
947 /*
948 * Disable counter
949 */
950 armv7_pmnc_disable_counter(idx);
951
952 /*
953 * Disable interrupt for this counter
954 */
955 armv7_pmnc_disable_intens(idx);
956
957 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
958}
959
960static irqreturn_t armv7pmu_handle_irq(int irq_num, void *dev)
961{
962 u32 pmnc;
963 struct perf_sample_data data;
964 struct arm_pmu *cpu_pmu = (struct arm_pmu *)dev;
965 struct pmu_hw_events *cpuc = this_cpu_ptr(cpu_pmu->hw_events);
966 struct pt_regs *regs;
967 int idx;
968
969 /*
970 * Get and reset the IRQ flags
971 */
972 pmnc = armv7_pmnc_getreset_flags();
973
974 /*
975 * Did an overflow occur?
976 */
977 if (!armv7_pmnc_has_overflowed(pmnc))
978 return IRQ_NONE;
979
980 /*
981 * Handle the counter(s) overflow(s)
982 */
983 regs = get_irq_regs();
984
985 for (idx = 0; idx < cpu_pmu->num_events; ++idx) {
986 struct perf_event *event = cpuc->events[idx];
987 struct hw_perf_event *hwc;
988
989 /* Ignore if we don't have an event. */
990 if (!event)
991 continue;
992
993 /*
994 * We have a single interrupt for all counters. Check that
995 * each counter has overflowed before we process it.
996 */
997 if (!armv7_pmnc_counter_has_overflowed(pmnc, idx))
998 continue;
999
1000 hwc = &event->hw;
1001 armpmu_event_update(event);
1002 perf_sample_data_init(&data, 0, hwc->last_period);
1003 if (!armpmu_event_set_period(event))
1004 continue;
1005
1006 if (perf_event_overflow(event, &data, regs))
1007 cpu_pmu->disable(event);
1008 }
1009
1010 /*
1011 * Handle the pending perf events.
1012 *
1013 * Note: this call *must* be run with interrupts disabled. For
1014 * platforms that can have the PMU interrupts raised as an NMI, this
1015 * will not work.
1016 */
1017 irq_work_run();
1018
1019 return IRQ_HANDLED;
1020}
1021
1022static void armv7pmu_start(struct arm_pmu *cpu_pmu)
1023{
1024 unsigned long flags;
1025 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
1026
1027 raw_spin_lock_irqsave(&events->pmu_lock, flags);
1028 /* Enable all counters */
1029 armv7_pmnc_write(armv7_pmnc_read() | ARMV7_PMNC_E);
1030 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
1031}
1032
1033static void armv7pmu_stop(struct arm_pmu *cpu_pmu)
1034{
1035 unsigned long flags;
1036 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
1037
1038 raw_spin_lock_irqsave(&events->pmu_lock, flags);
1039 /* Disable all counters */
1040 armv7_pmnc_write(armv7_pmnc_read() & ~ARMV7_PMNC_E);
1041 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
1042}
1043
1044static int armv7pmu_get_event_idx(struct pmu_hw_events *cpuc,
1045 struct perf_event *event)
1046{
1047 int idx;
1048 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
1049 struct hw_perf_event *hwc = &event->hw;
1050 unsigned long evtype = hwc->config_base & ARMV7_EVTYPE_EVENT;
1051
1052 /* Always place a cycle counter into the cycle counter. */
1053 if (evtype == ARMV7_PERFCTR_CPU_CYCLES) {
1054 if (test_and_set_bit(ARMV7_IDX_CYCLE_COUNTER, cpuc->used_mask))
1055 return -EAGAIN;
1056
1057 return ARMV7_IDX_CYCLE_COUNTER;
1058 }
1059
1060 /*
1061 * For anything other than a cycle counter, try and use
1062 * the events counters
1063 */
1064 for (idx = ARMV7_IDX_COUNTER0; idx < cpu_pmu->num_events; ++idx) {
1065 if (!test_and_set_bit(idx, cpuc->used_mask))
1066 return idx;
1067 }
1068
1069 /* The counters are all in use. */
1070 return -EAGAIN;
1071}
1072
1073/*
1074 * Add an event filter to a given event. This will only work for PMUv2 PMUs.
1075 */
1076static int armv7pmu_set_event_filter(struct hw_perf_event *event,
1077 struct perf_event_attr *attr)
1078{
1079 unsigned long config_base = 0;
1080
1081 if (attr->exclude_idle)
1082 return -EPERM;
1083 if (attr->exclude_user)
1084 config_base |= ARMV7_EXCLUDE_USER;
1085 if (attr->exclude_kernel)
1086 config_base |= ARMV7_EXCLUDE_PL1;
1087 if (!attr->exclude_hv)
1088 config_base |= ARMV7_INCLUDE_HYP;
1089
1090 /*
1091 * Install the filter into config_base as this is used to
1092 * construct the event type.
1093 */
1094 event->config_base = config_base;
1095
1096 return 0;
1097}
1098
1099static void armv7pmu_reset(void *info)
1100{
1101 struct arm_pmu *cpu_pmu = (struct arm_pmu *)info;
1102 u32 idx, nb_cnt = cpu_pmu->num_events, val;
1103
1104 if (cpu_pmu->secure_access) {
1105 asm volatile("mrc p15, 0, %0, c1, c1, 1" : "=r" (val));
1106 val |= ARMV7_SDER_SUNIDEN;
1107 asm volatile("mcr p15, 0, %0, c1, c1, 1" : : "r" (val));
1108 }
1109
1110 /* The counter and interrupt enable registers are unknown at reset. */
1111 for (idx = ARMV7_IDX_CYCLE_COUNTER; idx < nb_cnt; ++idx) {
1112 armv7_pmnc_disable_counter(idx);
1113 armv7_pmnc_disable_intens(idx);
1114 }
1115
1116 /* Initialize & Reset PMNC: C and P bits */
1117 armv7_pmnc_write(ARMV7_PMNC_P | ARMV7_PMNC_C);
1118}
1119
1120static int armv7_a8_map_event(struct perf_event *event)
1121{
1122 return armpmu_map_event(event, &armv7_a8_perf_map,
1123 &armv7_a8_perf_cache_map, 0xFF);
1124}
1125
1126static int armv7_a9_map_event(struct perf_event *event)
1127{
1128 return armpmu_map_event(event, &armv7_a9_perf_map,
1129 &armv7_a9_perf_cache_map, 0xFF);
1130}
1131
1132static int armv7_a5_map_event(struct perf_event *event)
1133{
1134 return armpmu_map_event(event, &armv7_a5_perf_map,
1135 &armv7_a5_perf_cache_map, 0xFF);
1136}
1137
1138static int armv7_a15_map_event(struct perf_event *event)
1139{
1140 return armpmu_map_event(event, &armv7_a15_perf_map,
1141 &armv7_a15_perf_cache_map, 0xFF);
1142}
1143
1144static int armv7_a7_map_event(struct perf_event *event)
1145{
1146 return armpmu_map_event(event, &armv7_a7_perf_map,
1147 &armv7_a7_perf_cache_map, 0xFF);
1148}
1149
1150static int armv7_a12_map_event(struct perf_event *event)
1151{
1152 return armpmu_map_event(event, &armv7_a12_perf_map,
1153 &armv7_a12_perf_cache_map, 0xFF);
1154}
1155
1156static int krait_map_event(struct perf_event *event)
1157{
1158 return armpmu_map_event(event, &krait_perf_map,
1159 &krait_perf_cache_map, 0xFFFFF);
1160}
1161
1162static int krait_map_event_no_branch(struct perf_event *event)
1163{
1164 return armpmu_map_event(event, &krait_perf_map_no_branch,
1165 &krait_perf_cache_map, 0xFFFFF);
1166}
1167
1168static int scorpion_map_event(struct perf_event *event)
1169{
1170 return armpmu_map_event(event, &scorpion_perf_map,
1171 &scorpion_perf_cache_map, 0xFFFFF);
1172}
1173
1174static void armv7pmu_init(struct arm_pmu *cpu_pmu)
1175{
1176 cpu_pmu->handle_irq = armv7pmu_handle_irq;
1177 cpu_pmu->enable = armv7pmu_enable_event;
1178 cpu_pmu->disable = armv7pmu_disable_event;
1179 cpu_pmu->read_counter = armv7pmu_read_counter;
1180 cpu_pmu->write_counter = armv7pmu_write_counter;
1181 cpu_pmu->get_event_idx = armv7pmu_get_event_idx;
1182 cpu_pmu->start = armv7pmu_start;
1183 cpu_pmu->stop = armv7pmu_stop;
1184 cpu_pmu->reset = armv7pmu_reset;
1185 cpu_pmu->max_period = (1LLU << 32) - 1;
1186};
1187
1188static void armv7_read_num_pmnc_events(void *info)
1189{
1190 int *nb_cnt = info;
1191
1192 /* Read the nb of CNTx counters supported from PMNC */
1193 *nb_cnt = (armv7_pmnc_read() >> ARMV7_PMNC_N_SHIFT) & ARMV7_PMNC_N_MASK;
1194
1195 /* Add the CPU cycles counter */
1196 *nb_cnt += 1;
1197}
1198
1199static int armv7_probe_num_events(struct arm_pmu *arm_pmu)
1200{
1201 return smp_call_function_any(&arm_pmu->supported_cpus,
1202 armv7_read_num_pmnc_events,
1203 &arm_pmu->num_events, 1);
1204}
1205
1206static int armv7_a8_pmu_init(struct arm_pmu *cpu_pmu)
1207{
1208 armv7pmu_init(cpu_pmu);
1209 cpu_pmu->name = "armv7_cortex_a8";
1210 cpu_pmu->map_event = armv7_a8_map_event;
1211 cpu_pmu->pmu.attr_groups = armv7_pmuv1_attr_groups;
1212 return armv7_probe_num_events(cpu_pmu);
1213}
1214
1215static int armv7_a9_pmu_init(struct arm_pmu *cpu_pmu)
1216{
1217 armv7pmu_init(cpu_pmu);
1218 cpu_pmu->name = "armv7_cortex_a9";
1219 cpu_pmu->map_event = armv7_a9_map_event;
1220 cpu_pmu->pmu.attr_groups = armv7_pmuv1_attr_groups;
1221 return armv7_probe_num_events(cpu_pmu);
1222}
1223
1224static int armv7_a5_pmu_init(struct arm_pmu *cpu_pmu)
1225{
1226 armv7pmu_init(cpu_pmu);
1227 cpu_pmu->name = "armv7_cortex_a5";
1228 cpu_pmu->map_event = armv7_a5_map_event;
1229 cpu_pmu->pmu.attr_groups = armv7_pmuv1_attr_groups;
1230 return armv7_probe_num_events(cpu_pmu);
1231}
1232
1233static int armv7_a15_pmu_init(struct arm_pmu *cpu_pmu)
1234{
1235 armv7pmu_init(cpu_pmu);
1236 cpu_pmu->name = "armv7_cortex_a15";
1237 cpu_pmu->map_event = armv7_a15_map_event;
1238 cpu_pmu->set_event_filter = armv7pmu_set_event_filter;
1239 cpu_pmu->pmu.attr_groups = armv7_pmuv2_attr_groups;
1240 return armv7_probe_num_events(cpu_pmu);
1241}
1242
1243static int armv7_a7_pmu_init(struct arm_pmu *cpu_pmu)
1244{
1245 armv7pmu_init(cpu_pmu);
1246 cpu_pmu->name = "armv7_cortex_a7";
1247 cpu_pmu->map_event = armv7_a7_map_event;
1248 cpu_pmu->set_event_filter = armv7pmu_set_event_filter;
1249 cpu_pmu->pmu.attr_groups = armv7_pmuv2_attr_groups;
1250 return armv7_probe_num_events(cpu_pmu);
1251}
1252
1253static int armv7_a12_pmu_init(struct arm_pmu *cpu_pmu)
1254{
1255 armv7pmu_init(cpu_pmu);
1256 cpu_pmu->name = "armv7_cortex_a12";
1257 cpu_pmu->map_event = armv7_a12_map_event;
1258 cpu_pmu->set_event_filter = armv7pmu_set_event_filter;
1259 cpu_pmu->pmu.attr_groups = armv7_pmuv2_attr_groups;
1260 return armv7_probe_num_events(cpu_pmu);
1261}
1262
1263static int armv7_a17_pmu_init(struct arm_pmu *cpu_pmu)
1264{
1265 int ret = armv7_a12_pmu_init(cpu_pmu);
1266 cpu_pmu->name = "armv7_cortex_a17";
1267 cpu_pmu->pmu.attr_groups = armv7_pmuv2_attr_groups;
1268 return ret;
1269}
1270
1271/*
1272 * Krait Performance Monitor Region Event Selection Register (PMRESRn)
1273 *
1274 * 31 30 24 16 8 0
1275 * +--------------------------------+
1276 * PMRESR0 | EN | CC | CC | CC | CC | N = 1, R = 0
1277 * +--------------------------------+
1278 * PMRESR1 | EN | CC | CC | CC | CC | N = 1, R = 1
1279 * +--------------------------------+
1280 * PMRESR2 | EN | CC | CC | CC | CC | N = 1, R = 2
1281 * +--------------------------------+
1282 * VPMRESR0 | EN | CC | CC | CC | CC | N = 2, R = ?
1283 * +--------------------------------+
1284 * EN | G=3 | G=2 | G=1 | G=0
1285 *
1286 * Event Encoding:
1287 *
1288 * hwc->config_base = 0xNRCCG
1289 *
1290 * N = prefix, 1 for Krait CPU (PMRESRn), 2 for Venum VFP (VPMRESR)
1291 * R = region register
1292 * CC = class of events the group G is choosing from
1293 * G = group or particular event
1294 *
1295 * Example: 0x12021 is a Krait CPU event in PMRESR2's group 1 with code 2
1296 *
1297 * A region (R) corresponds to a piece of the CPU (execution unit, instruction
1298 * unit, etc.) while the event code (CC) corresponds to a particular class of
1299 * events (interrupts for example). An event code is broken down into
1300 * groups (G) that can be mapped into the PMU (irq, fiqs, and irq+fiqs for
1301 * example).
1302 */
1303
1304#define KRAIT_EVENT (1 << 16)
1305#define VENUM_EVENT (2 << 16)
1306#define KRAIT_EVENT_MASK (KRAIT_EVENT | VENUM_EVENT)
1307#define PMRESRn_EN BIT(31)
1308
1309#define EVENT_REGION(event) (((event) >> 12) & 0xf) /* R */
1310#define EVENT_GROUP(event) ((event) & 0xf) /* G */
1311#define EVENT_CODE(event) (((event) >> 4) & 0xff) /* CC */
1312#define EVENT_VENUM(event) (!!(event & VENUM_EVENT)) /* N=2 */
1313#define EVENT_CPU(event) (!!(event & KRAIT_EVENT)) /* N=1 */
1314
1315static u32 krait_read_pmresrn(int n)
1316{
1317 u32 val;
1318
1319 switch (n) {
1320 case 0:
1321 asm volatile("mrc p15, 1, %0, c9, c15, 0" : "=r" (val));
1322 break;
1323 case 1:
1324 asm volatile("mrc p15, 1, %0, c9, c15, 1" : "=r" (val));
1325 break;
1326 case 2:
1327 asm volatile("mrc p15, 1, %0, c9, c15, 2" : "=r" (val));
1328 break;
1329 default:
1330 BUG(); /* Should be validated in krait_pmu_get_event_idx() */
1331 }
1332
1333 return val;
1334}
1335
1336static void krait_write_pmresrn(int n, u32 val)
1337{
1338 switch (n) {
1339 case 0:
1340 asm volatile("mcr p15, 1, %0, c9, c15, 0" : : "r" (val));
1341 break;
1342 case 1:
1343 asm volatile("mcr p15, 1, %0, c9, c15, 1" : : "r" (val));
1344 break;
1345 case 2:
1346 asm volatile("mcr p15, 1, %0, c9, c15, 2" : : "r" (val));
1347 break;
1348 default:
1349 BUG(); /* Should be validated in krait_pmu_get_event_idx() */
1350 }
1351}
1352
1353static u32 venum_read_pmresr(void)
1354{
1355 u32 val;
1356 asm volatile("mrc p10, 7, %0, c11, c0, 0" : "=r" (val));
1357 return val;
1358}
1359
1360static void venum_write_pmresr(u32 val)
1361{
1362 asm volatile("mcr p10, 7, %0, c11, c0, 0" : : "r" (val));
1363}
1364
1365static void venum_pre_pmresr(u32 *venum_orig_val, u32 *fp_orig_val)
1366{
1367 u32 venum_new_val;
1368 u32 fp_new_val;
1369
1370 BUG_ON(preemptible());
1371 /* CPACR Enable CP10 and CP11 access */
1372 *venum_orig_val = get_copro_access();
1373 venum_new_val = *venum_orig_val | CPACC_SVC(10) | CPACC_SVC(11);
1374 set_copro_access(venum_new_val);
1375
1376 /* Enable FPEXC */
1377 *fp_orig_val = fmrx(FPEXC);
1378 fp_new_val = *fp_orig_val | FPEXC_EN;
1379 fmxr(FPEXC, fp_new_val);
1380}
1381
1382static void venum_post_pmresr(u32 venum_orig_val, u32 fp_orig_val)
1383{
1384 BUG_ON(preemptible());
1385 /* Restore FPEXC */
1386 fmxr(FPEXC, fp_orig_val);
1387 isb();
1388 /* Restore CPACR */
1389 set_copro_access(venum_orig_val);
1390}
1391
1392static u32 krait_get_pmresrn_event(unsigned int region)
1393{
1394 static const u32 pmresrn_table[] = { KRAIT_PMRESR0_GROUP0,
1395 KRAIT_PMRESR1_GROUP0,
1396 KRAIT_PMRESR2_GROUP0 };
1397 return pmresrn_table[region];
1398}
1399
1400static void krait_evt_setup(int idx, u32 config_base)
1401{
1402 u32 val;
1403 u32 mask;
1404 u32 vval, fval;
1405 unsigned int region = EVENT_REGION(config_base);
1406 unsigned int group = EVENT_GROUP(config_base);
1407 unsigned int code = EVENT_CODE(config_base);
1408 unsigned int group_shift;
1409 bool venum_event = EVENT_VENUM(config_base);
1410
1411 group_shift = group * 8;
1412 mask = 0xff << group_shift;
1413
1414 /* Configure evtsel for the region and group */
1415 if (venum_event)
1416 val = KRAIT_VPMRESR0_GROUP0;
1417 else
1418 val = krait_get_pmresrn_event(region);
1419 val += group;
1420 /* Mix in mode-exclusion bits */
1421 val |= config_base & (ARMV7_EXCLUDE_USER | ARMV7_EXCLUDE_PL1);
1422 armv7_pmnc_write_evtsel(idx, val);
1423
1424 if (venum_event) {
1425 venum_pre_pmresr(&vval, &fval);
1426 val = venum_read_pmresr();
1427 val &= ~mask;
1428 val |= code << group_shift;
1429 val |= PMRESRn_EN;
1430 venum_write_pmresr(val);
1431 venum_post_pmresr(vval, fval);
1432 } else {
1433 val = krait_read_pmresrn(region);
1434 val &= ~mask;
1435 val |= code << group_shift;
1436 val |= PMRESRn_EN;
1437 krait_write_pmresrn(region, val);
1438 }
1439}
1440
1441static u32 clear_pmresrn_group(u32 val, int group)
1442{
1443 u32 mask;
1444 int group_shift;
1445
1446 group_shift = group * 8;
1447 mask = 0xff << group_shift;
1448 val &= ~mask;
1449
1450 /* Don't clear enable bit if entire region isn't disabled */
1451 if (val & ~PMRESRn_EN)
1452 return val |= PMRESRn_EN;
1453
1454 return 0;
1455}
1456
1457static void krait_clearpmu(u32 config_base)
1458{
1459 u32 val;
1460 u32 vval, fval;
1461 unsigned int region = EVENT_REGION(config_base);
1462 unsigned int group = EVENT_GROUP(config_base);
1463 bool venum_event = EVENT_VENUM(config_base);
1464
1465 if (venum_event) {
1466 venum_pre_pmresr(&vval, &fval);
1467 val = venum_read_pmresr();
1468 val = clear_pmresrn_group(val, group);
1469 venum_write_pmresr(val);
1470 venum_post_pmresr(vval, fval);
1471 } else {
1472 val = krait_read_pmresrn(region);
1473 val = clear_pmresrn_group(val, group);
1474 krait_write_pmresrn(region, val);
1475 }
1476}
1477
1478static void krait_pmu_disable_event(struct perf_event *event)
1479{
1480 unsigned long flags;
1481 struct hw_perf_event *hwc = &event->hw;
1482 int idx = hwc->idx;
1483 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
1484 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
1485
1486 /* Disable counter and interrupt */
1487 raw_spin_lock_irqsave(&events->pmu_lock, flags);
1488
1489 /* Disable counter */
1490 armv7_pmnc_disable_counter(idx);
1491
1492 /*
1493 * Clear pmresr code (if destined for PMNx counters)
1494 */
1495 if (hwc->config_base & KRAIT_EVENT_MASK)
1496 krait_clearpmu(hwc->config_base);
1497
1498 /* Disable interrupt for this counter */
1499 armv7_pmnc_disable_intens(idx);
1500
1501 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
1502}
1503
1504static void krait_pmu_enable_event(struct perf_event *event)
1505{
1506 unsigned long flags;
1507 struct hw_perf_event *hwc = &event->hw;
1508 int idx = hwc->idx;
1509 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
1510 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
1511
1512 /*
1513 * Enable counter and interrupt, and set the counter to count
1514 * the event that we're interested in.
1515 */
1516 raw_spin_lock_irqsave(&events->pmu_lock, flags);
1517
1518 /* Disable counter */
1519 armv7_pmnc_disable_counter(idx);
1520
1521 /*
1522 * Set event (if destined for PMNx counters)
1523 * We set the event for the cycle counter because we
1524 * have the ability to perform event filtering.
1525 */
1526 if (hwc->config_base & KRAIT_EVENT_MASK)
1527 krait_evt_setup(idx, hwc->config_base);
1528 else
1529 armv7_pmnc_write_evtsel(idx, hwc->config_base);
1530
1531 /* Enable interrupt for this counter */
1532 armv7_pmnc_enable_intens(idx);
1533
1534 /* Enable counter */
1535 armv7_pmnc_enable_counter(idx);
1536
1537 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
1538}
1539
1540static void krait_pmu_reset(void *info)
1541{
1542 u32 vval, fval;
1543 struct arm_pmu *cpu_pmu = info;
1544 u32 idx, nb_cnt = cpu_pmu->num_events;
1545
1546 armv7pmu_reset(info);
1547
1548 /* Clear all pmresrs */
1549 krait_write_pmresrn(0, 0);
1550 krait_write_pmresrn(1, 0);
1551 krait_write_pmresrn(2, 0);
1552
1553 venum_pre_pmresr(&vval, &fval);
1554 venum_write_pmresr(0);
1555 venum_post_pmresr(vval, fval);
1556
1557 /* Reset PMxEVNCTCR to sane default */
1558 for (idx = ARMV7_IDX_CYCLE_COUNTER; idx < nb_cnt; ++idx) {
1559 armv7_pmnc_select_counter(idx);
1560 asm volatile("mcr p15, 0, %0, c9, c15, 0" : : "r" (0));
1561 }
1562
1563}
1564
1565static int krait_event_to_bit(struct perf_event *event, unsigned int region,
1566 unsigned int group)
1567{
1568 int bit;
1569 struct hw_perf_event *hwc = &event->hw;
1570 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
1571
1572 if (hwc->config_base & VENUM_EVENT)
1573 bit = KRAIT_VPMRESR0_GROUP0;
1574 else
1575 bit = krait_get_pmresrn_event(region);
1576 bit -= krait_get_pmresrn_event(0);
1577 bit += group;
1578 /*
1579 * Lower bits are reserved for use by the counters (see
1580 * armv7pmu_get_event_idx() for more info)
1581 */
1582 bit += ARMV7_IDX_COUNTER_LAST(cpu_pmu) + 1;
1583
1584 return bit;
1585}
1586
1587/*
1588 * We check for column exclusion constraints here.
1589 * Two events cant use the same group within a pmresr register.
1590 */
1591static int krait_pmu_get_event_idx(struct pmu_hw_events *cpuc,
1592 struct perf_event *event)
1593{
1594 int idx;
1595 int bit = -1;
1596 struct hw_perf_event *hwc = &event->hw;
1597 unsigned int region = EVENT_REGION(hwc->config_base);
1598 unsigned int code = EVENT_CODE(hwc->config_base);
1599 unsigned int group = EVENT_GROUP(hwc->config_base);
1600 bool venum_event = EVENT_VENUM(hwc->config_base);
1601 bool krait_event = EVENT_CPU(hwc->config_base);
1602
1603 if (venum_event || krait_event) {
1604 /* Ignore invalid events */
1605 if (group > 3 || region > 2)
1606 return -EINVAL;
1607 if (venum_event && (code & 0xe0))
1608 return -EINVAL;
1609
1610 bit = krait_event_to_bit(event, region, group);
1611 if (test_and_set_bit(bit, cpuc->used_mask))
1612 return -EAGAIN;
1613 }
1614
1615 idx = armv7pmu_get_event_idx(cpuc, event);
1616 if (idx < 0 && bit >= 0)
1617 clear_bit(bit, cpuc->used_mask);
1618
1619 return idx;
1620}
1621
1622static void krait_pmu_clear_event_idx(struct pmu_hw_events *cpuc,
1623 struct perf_event *event)
1624{
1625 int bit;
1626 struct hw_perf_event *hwc = &event->hw;
1627 unsigned int region = EVENT_REGION(hwc->config_base);
1628 unsigned int group = EVENT_GROUP(hwc->config_base);
1629 bool venum_event = EVENT_VENUM(hwc->config_base);
1630 bool krait_event = EVENT_CPU(hwc->config_base);
1631
1632 if (venum_event || krait_event) {
1633 bit = krait_event_to_bit(event, region, group);
1634 clear_bit(bit, cpuc->used_mask);
1635 }
1636}
1637
1638static int krait_pmu_init(struct arm_pmu *cpu_pmu)
1639{
1640 armv7pmu_init(cpu_pmu);
1641 cpu_pmu->name = "armv7_krait";
1642 /* Some early versions of Krait don't support PC write events */
1643 if (of_property_read_bool(cpu_pmu->plat_device->dev.of_node,
1644 "qcom,no-pc-write"))
1645 cpu_pmu->map_event = krait_map_event_no_branch;
1646 else
1647 cpu_pmu->map_event = krait_map_event;
1648 cpu_pmu->set_event_filter = armv7pmu_set_event_filter;
1649 cpu_pmu->reset = krait_pmu_reset;
1650 cpu_pmu->enable = krait_pmu_enable_event;
1651 cpu_pmu->disable = krait_pmu_disable_event;
1652 cpu_pmu->get_event_idx = krait_pmu_get_event_idx;
1653 cpu_pmu->clear_event_idx = krait_pmu_clear_event_idx;
1654 return armv7_probe_num_events(cpu_pmu);
1655}
1656
1657/*
1658 * Scorpion Local Performance Monitor Register (LPMn)
1659 *
1660 * 31 30 24 16 8 0
1661 * +--------------------------------+
1662 * LPM0 | EN | CC | CC | CC | CC | N = 1, R = 0
1663 * +--------------------------------+
1664 * LPM1 | EN | CC | CC | CC | CC | N = 1, R = 1
1665 * +--------------------------------+
1666 * LPM2 | EN | CC | CC | CC | CC | N = 1, R = 2
1667 * +--------------------------------+
1668 * L2LPM | EN | CC | CC | CC | CC | N = 1, R = 3
1669 * +--------------------------------+
1670 * VLPM | EN | CC | CC | CC | CC | N = 2, R = ?
1671 * +--------------------------------+
1672 * EN | G=3 | G=2 | G=1 | G=0
1673 *
1674 *
1675 * Event Encoding:
1676 *
1677 * hwc->config_base = 0xNRCCG
1678 *
1679 * N = prefix, 1 for Scorpion CPU (LPMn/L2LPM), 2 for Venum VFP (VLPM)
1680 * R = region register
1681 * CC = class of events the group G is choosing from
1682 * G = group or particular event
1683 *
1684 * Example: 0x12021 is a Scorpion CPU event in LPM2's group 1 with code 2
1685 *
1686 * A region (R) corresponds to a piece of the CPU (execution unit, instruction
1687 * unit, etc.) while the event code (CC) corresponds to a particular class of
1688 * events (interrupts for example). An event code is broken down into
1689 * groups (G) that can be mapped into the PMU (irq, fiqs, and irq+fiqs for
1690 * example).
1691 */
1692
1693static u32 scorpion_read_pmresrn(int n)
1694{
1695 u32 val;
1696
1697 switch (n) {
1698 case 0:
1699 asm volatile("mrc p15, 0, %0, c15, c0, 0" : "=r" (val));
1700 break;
1701 case 1:
1702 asm volatile("mrc p15, 1, %0, c15, c0, 0" : "=r" (val));
1703 break;
1704 case 2:
1705 asm volatile("mrc p15, 2, %0, c15, c0, 0" : "=r" (val));
1706 break;
1707 case 3:
1708 asm volatile("mrc p15, 3, %0, c15, c2, 0" : "=r" (val));
1709 break;
1710 default:
1711 BUG(); /* Should be validated in scorpion_pmu_get_event_idx() */
1712 }
1713
1714 return val;
1715}
1716
1717static void scorpion_write_pmresrn(int n, u32 val)
1718{
1719 switch (n) {
1720 case 0:
1721 asm volatile("mcr p15, 0, %0, c15, c0, 0" : : "r" (val));
1722 break;
1723 case 1:
1724 asm volatile("mcr p15, 1, %0, c15, c0, 0" : : "r" (val));
1725 break;
1726 case 2:
1727 asm volatile("mcr p15, 2, %0, c15, c0, 0" : : "r" (val));
1728 break;
1729 case 3:
1730 asm volatile("mcr p15, 3, %0, c15, c2, 0" : : "r" (val));
1731 break;
1732 default:
1733 BUG(); /* Should be validated in scorpion_pmu_get_event_idx() */
1734 }
1735}
1736
1737static u32 scorpion_get_pmresrn_event(unsigned int region)
1738{
1739 static const u32 pmresrn_table[] = { SCORPION_LPM0_GROUP0,
1740 SCORPION_LPM1_GROUP0,
1741 SCORPION_LPM2_GROUP0,
1742 SCORPION_L2LPM_GROUP0 };
1743 return pmresrn_table[region];
1744}
1745
1746static void scorpion_evt_setup(int idx, u32 config_base)
1747{
1748 u32 val;
1749 u32 mask;
1750 u32 vval, fval;
1751 unsigned int region = EVENT_REGION(config_base);
1752 unsigned int group = EVENT_GROUP(config_base);
1753 unsigned int code = EVENT_CODE(config_base);
1754 unsigned int group_shift;
1755 bool venum_event = EVENT_VENUM(config_base);
1756
1757 group_shift = group * 8;
1758 mask = 0xff << group_shift;
1759
1760 /* Configure evtsel for the region and group */
1761 if (venum_event)
1762 val = SCORPION_VLPM_GROUP0;
1763 else
1764 val = scorpion_get_pmresrn_event(region);
1765 val += group;
1766 /* Mix in mode-exclusion bits */
1767 val |= config_base & (ARMV7_EXCLUDE_USER | ARMV7_EXCLUDE_PL1);
1768 armv7_pmnc_write_evtsel(idx, val);
1769
1770 asm volatile("mcr p15, 0, %0, c9, c15, 0" : : "r" (0));
1771
1772 if (venum_event) {
1773 venum_pre_pmresr(&vval, &fval);
1774 val = venum_read_pmresr();
1775 val &= ~mask;
1776 val |= code << group_shift;
1777 val |= PMRESRn_EN;
1778 venum_write_pmresr(val);
1779 venum_post_pmresr(vval, fval);
1780 } else {
1781 val = scorpion_read_pmresrn(region);
1782 val &= ~mask;
1783 val |= code << group_shift;
1784 val |= PMRESRn_EN;
1785 scorpion_write_pmresrn(region, val);
1786 }
1787}
1788
1789static void scorpion_clearpmu(u32 config_base)
1790{
1791 u32 val;
1792 u32 vval, fval;
1793 unsigned int region = EVENT_REGION(config_base);
1794 unsigned int group = EVENT_GROUP(config_base);
1795 bool venum_event = EVENT_VENUM(config_base);
1796
1797 if (venum_event) {
1798 venum_pre_pmresr(&vval, &fval);
1799 val = venum_read_pmresr();
1800 val = clear_pmresrn_group(val, group);
1801 venum_write_pmresr(val);
1802 venum_post_pmresr(vval, fval);
1803 } else {
1804 val = scorpion_read_pmresrn(region);
1805 val = clear_pmresrn_group(val, group);
1806 scorpion_write_pmresrn(region, val);
1807 }
1808}
1809
1810static void scorpion_pmu_disable_event(struct perf_event *event)
1811{
1812 unsigned long flags;
1813 struct hw_perf_event *hwc = &event->hw;
1814 int idx = hwc->idx;
1815 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
1816 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
1817
1818 /* Disable counter and interrupt */
1819 raw_spin_lock_irqsave(&events->pmu_lock, flags);
1820
1821 /* Disable counter */
1822 armv7_pmnc_disable_counter(idx);
1823
1824 /*
1825 * Clear pmresr code (if destined for PMNx counters)
1826 */
1827 if (hwc->config_base & KRAIT_EVENT_MASK)
1828 scorpion_clearpmu(hwc->config_base);
1829
1830 /* Disable interrupt for this counter */
1831 armv7_pmnc_disable_intens(idx);
1832
1833 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
1834}
1835
1836static void scorpion_pmu_enable_event(struct perf_event *event)
1837{
1838 unsigned long flags;
1839 struct hw_perf_event *hwc = &event->hw;
1840 int idx = hwc->idx;
1841 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
1842 struct pmu_hw_events *events = this_cpu_ptr(cpu_pmu->hw_events);
1843
1844 /*
1845 * Enable counter and interrupt, and set the counter to count
1846 * the event that we're interested in.
1847 */
1848 raw_spin_lock_irqsave(&events->pmu_lock, flags);
1849
1850 /* Disable counter */
1851 armv7_pmnc_disable_counter(idx);
1852
1853 /*
1854 * Set event (if destined for PMNx counters)
1855 * We don't set the event for the cycle counter because we
1856 * don't have the ability to perform event filtering.
1857 */
1858 if (hwc->config_base & KRAIT_EVENT_MASK)
1859 scorpion_evt_setup(idx, hwc->config_base);
1860 else if (idx != ARMV7_IDX_CYCLE_COUNTER)
1861 armv7_pmnc_write_evtsel(idx, hwc->config_base);
1862
1863 /* Enable interrupt for this counter */
1864 armv7_pmnc_enable_intens(idx);
1865
1866 /* Enable counter */
1867 armv7_pmnc_enable_counter(idx);
1868
1869 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
1870}
1871
1872static void scorpion_pmu_reset(void *info)
1873{
1874 u32 vval, fval;
1875 struct arm_pmu *cpu_pmu = info;
1876 u32 idx, nb_cnt = cpu_pmu->num_events;
1877
1878 armv7pmu_reset(info);
1879
1880 /* Clear all pmresrs */
1881 scorpion_write_pmresrn(0, 0);
1882 scorpion_write_pmresrn(1, 0);
1883 scorpion_write_pmresrn(2, 0);
1884 scorpion_write_pmresrn(3, 0);
1885
1886 venum_pre_pmresr(&vval, &fval);
1887 venum_write_pmresr(0);
1888 venum_post_pmresr(vval, fval);
1889
1890 /* Reset PMxEVNCTCR to sane default */
1891 for (idx = ARMV7_IDX_CYCLE_COUNTER; idx < nb_cnt; ++idx) {
1892 armv7_pmnc_select_counter(idx);
1893 asm volatile("mcr p15, 0, %0, c9, c15, 0" : : "r" (0));
1894 }
1895}
1896
1897static int scorpion_event_to_bit(struct perf_event *event, unsigned int region,
1898 unsigned int group)
1899{
1900 int bit;
1901 struct hw_perf_event *hwc = &event->hw;
1902 struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
1903
1904 if (hwc->config_base & VENUM_EVENT)
1905 bit = SCORPION_VLPM_GROUP0;
1906 else
1907 bit = scorpion_get_pmresrn_event(region);
1908 bit -= scorpion_get_pmresrn_event(0);
1909 bit += group;
1910 /*
1911 * Lower bits are reserved for use by the counters (see
1912 * armv7pmu_get_event_idx() for more info)
1913 */
1914 bit += ARMV7_IDX_COUNTER_LAST(cpu_pmu) + 1;
1915
1916 return bit;
1917}
1918
1919/*
1920 * We check for column exclusion constraints here.
1921 * Two events cant use the same group within a pmresr register.
1922 */
1923static int scorpion_pmu_get_event_idx(struct pmu_hw_events *cpuc,
1924 struct perf_event *event)
1925{
1926 int idx;
1927 int bit = -1;
1928 struct hw_perf_event *hwc = &event->hw;
1929 unsigned int region = EVENT_REGION(hwc->config_base);
1930 unsigned int group = EVENT_GROUP(hwc->config_base);
1931 bool venum_event = EVENT_VENUM(hwc->config_base);
1932 bool scorpion_event = EVENT_CPU(hwc->config_base);
1933
1934 if (venum_event || scorpion_event) {
1935 /* Ignore invalid events */
1936 if (group > 3 || region > 3)
1937 return -EINVAL;
1938
1939 bit = scorpion_event_to_bit(event, region, group);
1940 if (test_and_set_bit(bit, cpuc->used_mask))
1941 return -EAGAIN;
1942 }
1943
1944 idx = armv7pmu_get_event_idx(cpuc, event);
1945 if (idx < 0 && bit >= 0)
1946 clear_bit(bit, cpuc->used_mask);
1947
1948 return idx;
1949}
1950
1951static void scorpion_pmu_clear_event_idx(struct pmu_hw_events *cpuc,
1952 struct perf_event *event)
1953{
1954 int bit;
1955 struct hw_perf_event *hwc = &event->hw;
1956 unsigned int region = EVENT_REGION(hwc->config_base);
1957 unsigned int group = EVENT_GROUP(hwc->config_base);
1958 bool venum_event = EVENT_VENUM(hwc->config_base);
1959 bool scorpion_event = EVENT_CPU(hwc->config_base);
1960
1961 if (venum_event || scorpion_event) {
1962 bit = scorpion_event_to_bit(event, region, group);
1963 clear_bit(bit, cpuc->used_mask);
1964 }
1965}
1966
1967static int scorpion_pmu_init(struct arm_pmu *cpu_pmu)
1968{
1969 armv7pmu_init(cpu_pmu);
1970 cpu_pmu->name = "armv7_scorpion";
1971 cpu_pmu->map_event = scorpion_map_event;
1972 cpu_pmu->reset = scorpion_pmu_reset;
1973 cpu_pmu->enable = scorpion_pmu_enable_event;
1974 cpu_pmu->disable = scorpion_pmu_disable_event;
1975 cpu_pmu->get_event_idx = scorpion_pmu_get_event_idx;
1976 cpu_pmu->clear_event_idx = scorpion_pmu_clear_event_idx;
1977 return armv7_probe_num_events(cpu_pmu);
1978}
1979
1980static int scorpion_mp_pmu_init(struct arm_pmu *cpu_pmu)
1981{
1982 armv7pmu_init(cpu_pmu);
1983 cpu_pmu->name = "armv7_scorpion_mp";
1984 cpu_pmu->map_event = scorpion_map_event;
1985 cpu_pmu->reset = scorpion_pmu_reset;
1986 cpu_pmu->enable = scorpion_pmu_enable_event;
1987 cpu_pmu->disable = scorpion_pmu_disable_event;
1988 cpu_pmu->get_event_idx = scorpion_pmu_get_event_idx;
1989 cpu_pmu->clear_event_idx = scorpion_pmu_clear_event_idx;
1990 return armv7_probe_num_events(cpu_pmu);
1991}
1992
1993static const struct of_device_id armv7_pmu_of_device_ids[] = {
1994 {.compatible = "arm,cortex-a17-pmu", .data = armv7_a17_pmu_init},
1995 {.compatible = "arm,cortex-a15-pmu", .data = armv7_a15_pmu_init},
1996 {.compatible = "arm,cortex-a12-pmu", .data = armv7_a12_pmu_init},
1997 {.compatible = "arm,cortex-a9-pmu", .data = armv7_a9_pmu_init},
1998 {.compatible = "arm,cortex-a8-pmu", .data = armv7_a8_pmu_init},
1999 {.compatible = "arm,cortex-a7-pmu", .data = armv7_a7_pmu_init},
2000 {.compatible = "arm,cortex-a5-pmu", .data = armv7_a5_pmu_init},
2001 {.compatible = "qcom,krait-pmu", .data = krait_pmu_init},
2002 {.compatible = "qcom,scorpion-pmu", .data = scorpion_pmu_init},
2003 {.compatible = "qcom,scorpion-mp-pmu", .data = scorpion_mp_pmu_init},
2004 {},
2005};
2006
2007static const struct pmu_probe_info armv7_pmu_probe_table[] = {
2008 ARM_PMU_PROBE(ARM_CPU_PART_CORTEX_A8, armv7_a8_pmu_init),
2009 ARM_PMU_PROBE(ARM_CPU_PART_CORTEX_A9, armv7_a9_pmu_init),
2010 { /* sentinel value */ }
2011};
2012
2013
2014static int armv7_pmu_device_probe(struct platform_device *pdev)
2015{
2016 return arm_pmu_device_probe(pdev, armv7_pmu_of_device_ids,
2017 armv7_pmu_probe_table);
2018}
2019
2020static struct platform_driver armv7_pmu_driver = {
2021 .driver = {
2022 .name = "armv7-pmu",
2023 .of_match_table = armv7_pmu_of_device_ids,
2024 },
2025 .probe = armv7_pmu_device_probe,
2026};
2027
2028static int __init register_armv7_pmu_driver(void)
2029{
2030 return platform_driver_register(&armv7_pmu_driver);
2031}
2032device_initcall(register_armv7_pmu_driver);
2033#endif /* CONFIG_CPU_V7 */