1 // SPDX-License-Identifier: GPL-2.0
2 // CCI Cache Coherent Interconnect PMU driver
3 // Copyright (C) 2013-2018 Arm Ltd.
4 // Author: Punit Agrawal <punit.agrawal@arm.com>, Suzuki Poulose <suzuki.poulose@arm.com>
5
6 #include <linux/arm-cci.h>
7 #include <linux/io.h>
8 #include <linux/interrupt.h>
9 #include <linux/module.h>
10 #include <linux/of.h>
11 #include <linux/perf_event.h>
12 #include <linux/platform_device.h>
13 #include <linux/slab.h>
14 #include <linux/spinlock.h>
15
16 #define DRIVER_NAME "ARM-CCI PMU"
17
18 #define CCI_PMCR 0x0100
19 #define CCI_PID2 0x0fe8
20
21 #define CCI_PMCR_CEN 0x00000001
22 #define CCI_PMCR_NCNT_MASK 0x0000f800
23 #define CCI_PMCR_NCNT_SHIFT 11
24
25 #define CCI_PID2_REV_MASK 0xf0
26 #define CCI_PID2_REV_SHIFT 4
27
28 #define CCI_PMU_EVT_SEL 0x000
29 #define CCI_PMU_CNTR 0x004
30 #define CCI_PMU_CNTR_CTRL 0x008
31 #define CCI_PMU_OVRFLW 0x00c
32
33 #define CCI_PMU_OVRFLW_FLAG 1
34
35 #define CCI_PMU_CNTR_SIZE(model) ((model)->cntr_size)
36 #define CCI_PMU_CNTR_BASE(model, idx) ((idx) * CCI_PMU_CNTR_SIZE(model))
37 #define CCI_PMU_CNTR_MASK ((1ULL << 32) - 1)
38 #define CCI_PMU_CNTR_LAST(cci_pmu) (cci_pmu->num_cntrs - 1)
39
40 #define CCI_PMU_MAX_HW_CNTRS(model) \
41 ((model)->num_hw_cntrs + (model)->fixed_hw_cntrs)
42
43 /* Types of interfaces that can generate events */
44 enum {
45 CCI_IF_SLAVE,
46 CCI_IF_MASTER,
47 #ifdef CONFIG_ARM_CCI5xx_PMU
48 CCI_IF_GLOBAL,
49 #endif
50 CCI_IF_MAX,
51 };
52
53 #define NUM_HW_CNTRS_CII_4XX 4
54 #define NUM_HW_CNTRS_CII_5XX 8
55 #define NUM_HW_CNTRS_MAX NUM_HW_CNTRS_CII_5XX
56
57 #define FIXED_HW_CNTRS_CII_4XX 1
58 #define FIXED_HW_CNTRS_CII_5XX 0
59 #define FIXED_HW_CNTRS_MAX FIXED_HW_CNTRS_CII_4XX
60
61 #define HW_CNTRS_MAX (NUM_HW_CNTRS_MAX + FIXED_HW_CNTRS_MAX)
62
63 struct event_range {
64 u32 min;
65 u32 max;
66 };
67
68 struct cci_pmu_hw_events {
69 struct perf_event **events;
70 unsigned long *used_mask;
71 raw_spinlock_t pmu_lock;
72 };
73
74 struct cci_pmu;
75 /*
76 * struct cci_pmu_model:
77 * @fixed_hw_cntrs - Number of fixed event counters
78 * @num_hw_cntrs - Maximum number of programmable event counters
79 * @cntr_size - Size of an event counter mapping
80 */
81 struct cci_pmu_model {
82 char *name;
83 u32 fixed_hw_cntrs;
84 u32 num_hw_cntrs;
85 u32 cntr_size;
86 struct attribute **format_attrs;
87 struct attribute **event_attrs;
88 struct event_range event_ranges[CCI_IF_MAX];
89 int (*validate_hw_event)(struct cci_pmu *, unsigned long);
90 int (*get_event_idx)(struct cci_pmu *, struct cci_pmu_hw_events *, unsigned long);
91 void (*write_counters)(struct cci_pmu *, unsigned long *);
92 };
93
94 static struct cci_pmu_model cci_pmu_models[];
95
96 struct cci_pmu {
97 void __iomem *base;
98 void __iomem *ctrl_base;
99 struct pmu pmu;
100 int cpu;
101 int nr_irqs;
102 int *irqs;
103 unsigned long active_irqs;
104 const struct cci_pmu_model *model;
105 struct cci_pmu_hw_events hw_events;
106 struct platform_device *plat_device;
107 int num_cntrs;
108 atomic_t active_events;
109 struct mutex reserve_mutex;
110 };
111
112 #define to_cci_pmu(c) (container_of(c, struct cci_pmu, pmu))
113
114 static struct cci_pmu *g_cci_pmu;
115
116 enum cci_models {
117 #ifdef CONFIG_ARM_CCI400_PMU
118 CCI400_R0,
119 CCI400_R1,
120 #endif
121 #ifdef CONFIG_ARM_CCI5xx_PMU
122 CCI500_R0,
123 CCI550_R0,
124 #endif
125 CCI_MODEL_MAX
126 };
127
128 static void pmu_write_counters(struct cci_pmu *cci_pmu,
129 unsigned long *mask);
130 static ssize_t __maybe_unused cci_pmu_event_show(struct device *dev,
131 struct device_attribute *attr, char *buf);
132
133 #define CCI_EXT_ATTR_ENTRY(_name, _func, _config) \
134 &((struct dev_ext_attribute[]) { \
135 { __ATTR(_name, S_IRUGO, _func, NULL), (void *)_config } \
136 })[0].attr.attr
137
138 #define CCI_FORMAT_EXT_ATTR_ENTRY(_name, _config) \
139 CCI_EXT_ATTR_ENTRY(_name, device_show_string, _config)
140 #define CCI_EVENT_EXT_ATTR_ENTRY(_name, _config) \
141 CCI_EXT_ATTR_ENTRY(_name, cci_pmu_event_show, (unsigned long)_config)
142
143 /* CCI400 PMU Specific definitions */
144
145 #ifdef CONFIG_ARM_CCI400_PMU
146
147 /* Port ids */
148 #define CCI400_PORT_S0 0
149 #define CCI400_PORT_S1 1
150 #define CCI400_PORT_S2 2
151 #define CCI400_PORT_S3 3
152 #define CCI400_PORT_S4 4
153 #define CCI400_PORT_M0 5
154 #define CCI400_PORT_M1 6
155 #define CCI400_PORT_M2 7
156
157 #define CCI400_R1_PX 5
158
159 /*
160 * Instead of an event id to monitor CCI cycles, a dedicated counter is
161 * provided. Use 0xff to represent CCI cycles and hope that no future revisions
162 * make use of this event in hardware.
163 */
164 enum cci400_perf_events {
165 CCI400_PMU_CYCLES = 0xff
166 };
167
168 #define CCI400_PMU_CYCLE_CNTR_IDX 0
169 #define CCI400_PMU_CNTR0_IDX 1
170
171 /*
172 * CCI PMU event id is an 8-bit value made of two parts - bits 7:5 for one of 8
173 * ports and bits 4:0 are event codes. There are different event codes
174 * associated with each port type.
175 *
176 * Additionally, the range of events associated with the port types changed
177 * between Rev0 and Rev1.
178 *
179 * The constants below define the range of valid codes for each port type for
180 * the different revisions and are used to validate the event to be monitored.
181 */
182
183 #define CCI400_PMU_EVENT_MASK 0xffUL
184 #define CCI400_PMU_EVENT_SOURCE_SHIFT 5
185 #define CCI400_PMU_EVENT_SOURCE_MASK 0x7
186 #define CCI400_PMU_EVENT_CODE_SHIFT 0
187 #define CCI400_PMU_EVENT_CODE_MASK 0x1f
188 #define CCI400_PMU_EVENT_SOURCE(event) \
189 ((event >> CCI400_PMU_EVENT_SOURCE_SHIFT) & \
190 CCI400_PMU_EVENT_SOURCE_MASK)
191 #define CCI400_PMU_EVENT_CODE(event) \
192 ((event >> CCI400_PMU_EVENT_CODE_SHIFT) & CCI400_PMU_EVENT_CODE_MASK)
193
194 #define CCI400_R0_SLAVE_PORT_MIN_EV 0x00
195 #define CCI400_R0_SLAVE_PORT_MAX_EV 0x13
196 #define CCI400_R0_MASTER_PORT_MIN_EV 0x14
197 #define CCI400_R0_MASTER_PORT_MAX_EV 0x1a
198
199 #define CCI400_R1_SLAVE_PORT_MIN_EV 0x00
200 #define CCI400_R1_SLAVE_PORT_MAX_EV 0x14
201 #define CCI400_R1_MASTER_PORT_MIN_EV 0x00
202 #define CCI400_R1_MASTER_PORT_MAX_EV 0x11
203
204 #define CCI400_CYCLE_EVENT_EXT_ATTR_ENTRY(_name, _config) \
205 CCI_EXT_ATTR_ENTRY(_name, cci400_pmu_cycle_event_show, \
206 (unsigned long)_config)
207
208 static ssize_t cci400_pmu_cycle_event_show(struct device *dev,
209 struct device_attribute *attr, char *buf);
210
211 static struct attribute *cci400_pmu_format_attrs[] = {
212 CCI_FORMAT_EXT_ATTR_ENTRY(event, "config:0-4"),
213 CCI_FORMAT_EXT_ATTR_ENTRY(source, "config:5-7"),
214 NULL
215 };
216
217 static struct attribute *cci400_r0_pmu_event_attrs[] = {
218 /* Slave events */
219 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_any, 0x0),
220 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_device, 0x01),
221 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_normal_or_nonshareable, 0x2),
222 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_inner_or_outershareable, 0x3),
223 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_cache_maintenance, 0x4),
224 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_mem_barrier, 0x5),
225 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_sync_barrier, 0x6),
226 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg, 0x7),
227 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg_sync, 0x8),
228 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_stall_tt_full, 0x9),
229 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_last_hs_snoop, 0xA),
230 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_stall_rvalids_h_rready_l, 0xB),
231 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_any, 0xC),
232 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_device, 0xD),
233 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_normal_or_nonshareable, 0xE),
234 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_inner_or_outershare_wback_wclean, 0xF),
235 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_write_unique, 0x10),
236 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_write_line_unique, 0x11),
237 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_evict, 0x12),
238 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_stall_tt_full, 0x13),
239 /* Master events */
240 CCI_EVENT_EXT_ATTR_ENTRY(mi_retry_speculative_fetch, 0x14),
241 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_addr_hazard, 0x15),
242 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_id_hazard, 0x16),
243 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_tt_full, 0x17),
244 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_barrier_hazard, 0x18),
245 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_barrier_hazard, 0x19),
246 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_tt_full, 0x1A),
247 /* Special event for cycles counter */
248 CCI400_CYCLE_EVENT_EXT_ATTR_ENTRY(cycles, 0xff),
249 NULL
250 };
251
252 static struct attribute *cci400_r1_pmu_event_attrs[] = {
253 /* Slave events */
254 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_any, 0x0),
255 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_device, 0x01),
256 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_normal_or_nonshareable, 0x2),
257 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_inner_or_outershareable, 0x3),
258 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_cache_maintenance, 0x4),
259 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_mem_barrier, 0x5),
260 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_sync_barrier, 0x6),
261 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg, 0x7),
262 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg_sync, 0x8),
263 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_stall_tt_full, 0x9),
264 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_last_hs_snoop, 0xA),
265 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_stall_rvalids_h_rready_l, 0xB),
266 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_any, 0xC),
267 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_device, 0xD),
268 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_normal_or_nonshareable, 0xE),
269 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_inner_or_outershare_wback_wclean, 0xF),
270 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_write_unique, 0x10),
271 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_write_line_unique, 0x11),
272 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_evict, 0x12),
273 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_stall_tt_full, 0x13),
274 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_stall_slave_id_hazard, 0x14),
275 /* Master events */
276 CCI_EVENT_EXT_ATTR_ENTRY(mi_retry_speculative_fetch, 0x0),
277 CCI_EVENT_EXT_ATTR_ENTRY(mi_stall_cycle_addr_hazard, 0x1),
278 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_master_id_hazard, 0x2),
279 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_hi_prio_rtq_full, 0x3),
280 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_barrier_hazard, 0x4),
281 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_barrier_hazard, 0x5),
282 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_wtq_full, 0x6),
283 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_low_prio_rtq_full, 0x7),
284 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_mid_prio_rtq_full, 0x8),
285 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_qvn_vn0, 0x9),
286 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_qvn_vn1, 0xA),
287 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_qvn_vn2, 0xB),
288 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_qvn_vn3, 0xC),
289 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_qvn_vn0, 0xD),
290 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_qvn_vn1, 0xE),
291 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_qvn_vn2, 0xF),
292 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_qvn_vn3, 0x10),
293 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_unique_or_line_unique_addr_hazard, 0x11),
294 /* Special event for cycles counter */
295 CCI400_CYCLE_EVENT_EXT_ATTR_ENTRY(cycles, 0xff),
296 NULL
297 };
298
cci400_pmu_cycle_event_show(struct device * dev,struct device_attribute * attr,char * buf)299 static ssize_t cci400_pmu_cycle_event_show(struct device *dev,
300 struct device_attribute *attr, char *buf)
301 {
302 struct dev_ext_attribute *eattr = container_of(attr,
303 struct dev_ext_attribute, attr);
304 return sysfs_emit(buf, "config=0x%lx\n", (unsigned long)eattr->var);
305 }
306
cci400_get_event_idx(struct cci_pmu * cci_pmu,struct cci_pmu_hw_events * hw,unsigned long cci_event)307 static int cci400_get_event_idx(struct cci_pmu *cci_pmu,
308 struct cci_pmu_hw_events *hw,
309 unsigned long cci_event)
310 {
311 int idx;
312
313 /* cycles event idx is fixed */
314 if (cci_event == CCI400_PMU_CYCLES) {
315 if (test_and_set_bit(CCI400_PMU_CYCLE_CNTR_IDX, hw->used_mask))
316 return -EAGAIN;
317
318 return CCI400_PMU_CYCLE_CNTR_IDX;
319 }
320
321 for (idx = CCI400_PMU_CNTR0_IDX; idx <= CCI_PMU_CNTR_LAST(cci_pmu); ++idx)
322 if (!test_and_set_bit(idx, hw->used_mask))
323 return idx;
324
325 /* No counters available */
326 return -EAGAIN;
327 }
328
cci400_validate_hw_event(struct cci_pmu * cci_pmu,unsigned long hw_event)329 static int cci400_validate_hw_event(struct cci_pmu *cci_pmu, unsigned long hw_event)
330 {
331 u8 ev_source = CCI400_PMU_EVENT_SOURCE(hw_event);
332 u8 ev_code = CCI400_PMU_EVENT_CODE(hw_event);
333 int if_type;
334
335 if (hw_event & ~CCI400_PMU_EVENT_MASK)
336 return -ENOENT;
337
338 if (hw_event == CCI400_PMU_CYCLES)
339 return hw_event;
340
341 switch (ev_source) {
342 case CCI400_PORT_S0:
343 case CCI400_PORT_S1:
344 case CCI400_PORT_S2:
345 case CCI400_PORT_S3:
346 case CCI400_PORT_S4:
347 /* Slave Interface */
348 if_type = CCI_IF_SLAVE;
349 break;
350 case CCI400_PORT_M0:
351 case CCI400_PORT_M1:
352 case CCI400_PORT_M2:
353 /* Master Interface */
354 if_type = CCI_IF_MASTER;
355 break;
356 default:
357 return -ENOENT;
358 }
359
360 if (ev_code >= cci_pmu->model->event_ranges[if_type].min &&
361 ev_code <= cci_pmu->model->event_ranges[if_type].max)
362 return hw_event;
363
364 return -ENOENT;
365 }
366
probe_cci400_revision(struct cci_pmu * cci_pmu)367 static int probe_cci400_revision(struct cci_pmu *cci_pmu)
368 {
369 int rev;
370 rev = readl_relaxed(cci_pmu->ctrl_base + CCI_PID2) & CCI_PID2_REV_MASK;
371 rev >>= CCI_PID2_REV_SHIFT;
372
373 if (rev < CCI400_R1_PX)
374 return CCI400_R0;
375 else
376 return CCI400_R1;
377 }
378
probe_cci_model(struct cci_pmu * cci_pmu)379 static const struct cci_pmu_model *probe_cci_model(struct cci_pmu *cci_pmu)
380 {
381 if (platform_has_secure_cci_access())
382 return &cci_pmu_models[probe_cci400_revision(cci_pmu)];
383 return NULL;
384 }
385 #else /* !CONFIG_ARM_CCI400_PMU */
probe_cci_model(struct cci_pmu * cci_pmu)386 static inline struct cci_pmu_model *probe_cci_model(struct cci_pmu *cci_pmu)
387 {
388 return NULL;
389 }
390 #endif /* CONFIG_ARM_CCI400_PMU */
391
392 #ifdef CONFIG_ARM_CCI5xx_PMU
393
394 /*
395 * CCI5xx PMU event id is an 9-bit value made of two parts.
396 * bits [8:5] - Source for the event
397 * bits [4:0] - Event code (specific to type of interface)
398 *
399 *
400 */
401
402 /* Port ids */
403 #define CCI5xx_PORT_S0 0x0
404 #define CCI5xx_PORT_S1 0x1
405 #define CCI5xx_PORT_S2 0x2
406 #define CCI5xx_PORT_S3 0x3
407 #define CCI5xx_PORT_S4 0x4
408 #define CCI5xx_PORT_S5 0x5
409 #define CCI5xx_PORT_S6 0x6
410
411 #define CCI5xx_PORT_M0 0x8
412 #define CCI5xx_PORT_M1 0x9
413 #define CCI5xx_PORT_M2 0xa
414 #define CCI5xx_PORT_M3 0xb
415 #define CCI5xx_PORT_M4 0xc
416 #define CCI5xx_PORT_M5 0xd
417 #define CCI5xx_PORT_M6 0xe
418
419 #define CCI5xx_PORT_GLOBAL 0xf
420
421 #define CCI5xx_PMU_EVENT_MASK 0x1ffUL
422 #define CCI5xx_PMU_EVENT_SOURCE_SHIFT 0x5
423 #define CCI5xx_PMU_EVENT_SOURCE_MASK 0xf
424 #define CCI5xx_PMU_EVENT_CODE_SHIFT 0x0
425 #define CCI5xx_PMU_EVENT_CODE_MASK 0x1f
426
427 #define CCI5xx_PMU_EVENT_SOURCE(event) \
428 ((event >> CCI5xx_PMU_EVENT_SOURCE_SHIFT) & CCI5xx_PMU_EVENT_SOURCE_MASK)
429 #define CCI5xx_PMU_EVENT_CODE(event) \
430 ((event >> CCI5xx_PMU_EVENT_CODE_SHIFT) & CCI5xx_PMU_EVENT_CODE_MASK)
431
432 #define CCI5xx_SLAVE_PORT_MIN_EV 0x00
433 #define CCI5xx_SLAVE_PORT_MAX_EV 0x1f
434 #define CCI5xx_MASTER_PORT_MIN_EV 0x00
435 #define CCI5xx_MASTER_PORT_MAX_EV 0x06
436 #define CCI5xx_GLOBAL_PORT_MIN_EV 0x00
437 #define CCI5xx_GLOBAL_PORT_MAX_EV 0x0f
438
439
440 #define CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(_name, _config) \
441 CCI_EXT_ATTR_ENTRY(_name, cci5xx_pmu_global_event_show, \
442 (unsigned long) _config)
443
444 static ssize_t cci5xx_pmu_global_event_show(struct device *dev,
445 struct device_attribute *attr, char *buf);
446
447 static struct attribute *cci5xx_pmu_format_attrs[] = {
448 CCI_FORMAT_EXT_ATTR_ENTRY(event, "config:0-4"),
449 CCI_FORMAT_EXT_ATTR_ENTRY(source, "config:5-8"),
450 NULL,
451 };
452
453 static struct attribute *cci5xx_pmu_event_attrs[] = {
454 /* Slave events */
455 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_arvalid, 0x0),
456 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_dev, 0x1),
457 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_nonshareable, 0x2),
458 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_shareable_non_alloc, 0x3),
459 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_shareable_alloc, 0x4),
460 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_invalidate, 0x5),
461 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_cache_maint, 0x6),
462 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg, 0x7),
463 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_rval, 0x8),
464 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_rlast_snoop, 0x9),
465 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_awalid, 0xA),
466 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_dev, 0xB),
467 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_non_shareable, 0xC),
468 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_share_wb, 0xD),
469 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_share_wlu, 0xE),
470 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_share_wunique, 0xF),
471 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_evict, 0x10),
472 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_wrevict, 0x11),
473 CCI_EVENT_EXT_ATTR_ENTRY(si_w_data_beat, 0x12),
474 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_acvalid, 0x13),
475 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_read, 0x14),
476 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_clean, 0x15),
477 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_data_transfer_low, 0x16),
478 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_stall_arvalid, 0x17),
479 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_stall, 0x18),
480 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_stall, 0x19),
481 CCI_EVENT_EXT_ATTR_ENTRY(si_w_data_stall, 0x1A),
482 CCI_EVENT_EXT_ATTR_ENTRY(si_w_resp_stall, 0x1B),
483 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_stall, 0x1C),
484 CCI_EVENT_EXT_ATTR_ENTRY(si_s_data_stall, 0x1D),
485 CCI_EVENT_EXT_ATTR_ENTRY(si_rq_stall_ot_limit, 0x1E),
486 CCI_EVENT_EXT_ATTR_ENTRY(si_r_stall_arbit, 0x1F),
487
488 /* Master events */
489 CCI_EVENT_EXT_ATTR_ENTRY(mi_r_data_beat_any, 0x0),
490 CCI_EVENT_EXT_ATTR_ENTRY(mi_w_data_beat_any, 0x1),
491 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall, 0x2),
492 CCI_EVENT_EXT_ATTR_ENTRY(mi_r_data_stall, 0x3),
493 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall, 0x4),
494 CCI_EVENT_EXT_ATTR_ENTRY(mi_w_data_stall, 0x5),
495 CCI_EVENT_EXT_ATTR_ENTRY(mi_w_resp_stall, 0x6),
496
497 /* Global events */
498 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_filter_bank_0_1, 0x0),
499 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_filter_bank_2_3, 0x1),
500 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_filter_bank_4_5, 0x2),
501 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_filter_bank_6_7, 0x3),
502 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_miss_filter_bank_0_1, 0x4),
503 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_miss_filter_bank_2_3, 0x5),
504 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_miss_filter_bank_4_5, 0x6),
505 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_miss_filter_bank_6_7, 0x7),
506 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_back_invalidation, 0x8),
507 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_stall_alloc_busy, 0x9),
508 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_stall_tt_full, 0xA),
509 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_wrq, 0xB),
510 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_cd_hs, 0xC),
511 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_rq_stall_addr_hazard, 0xD),
512 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_rq_stall_tt_full, 0xE),
513 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_rq_tzmp1_prot, 0xF),
514 NULL
515 };
516
cci5xx_pmu_global_event_show(struct device * dev,struct device_attribute * attr,char * buf)517 static ssize_t cci5xx_pmu_global_event_show(struct device *dev,
518 struct device_attribute *attr, char *buf)
519 {
520 struct dev_ext_attribute *eattr = container_of(attr,
521 struct dev_ext_attribute, attr);
522 /* Global events have single fixed source code */
523 return sysfs_emit(buf, "event=0x%lx,source=0x%x\n",
524 (unsigned long)eattr->var, CCI5xx_PORT_GLOBAL);
525 }
526
527 /*
528 * CCI500 provides 8 independent event counters that can count
529 * any of the events available.
530 * CCI500 PMU event source ids
531 * 0x0-0x6 - Slave interfaces
532 * 0x8-0xD - Master interfaces
533 * 0xf - Global Events
534 * 0x7,0xe - Reserved
535 */
cci500_validate_hw_event(struct cci_pmu * cci_pmu,unsigned long hw_event)536 static int cci500_validate_hw_event(struct cci_pmu *cci_pmu,
537 unsigned long hw_event)
538 {
539 u32 ev_source = CCI5xx_PMU_EVENT_SOURCE(hw_event);
540 u32 ev_code = CCI5xx_PMU_EVENT_CODE(hw_event);
541 int if_type;
542
543 if (hw_event & ~CCI5xx_PMU_EVENT_MASK)
544 return -ENOENT;
545
546 switch (ev_source) {
547 case CCI5xx_PORT_S0:
548 case CCI5xx_PORT_S1:
549 case CCI5xx_PORT_S2:
550 case CCI5xx_PORT_S3:
551 case CCI5xx_PORT_S4:
552 case CCI5xx_PORT_S5:
553 case CCI5xx_PORT_S6:
554 if_type = CCI_IF_SLAVE;
555 break;
556 case CCI5xx_PORT_M0:
557 case CCI5xx_PORT_M1:
558 case CCI5xx_PORT_M2:
559 case CCI5xx_PORT_M3:
560 case CCI5xx_PORT_M4:
561 case CCI5xx_PORT_M5:
562 if_type = CCI_IF_MASTER;
563 break;
564 case CCI5xx_PORT_GLOBAL:
565 if_type = CCI_IF_GLOBAL;
566 break;
567 default:
568 return -ENOENT;
569 }
570
571 if (ev_code >= cci_pmu->model->event_ranges[if_type].min &&
572 ev_code <= cci_pmu->model->event_ranges[if_type].max)
573 return hw_event;
574
575 return -ENOENT;
576 }
577
578 /*
579 * CCI550 provides 8 independent event counters that can count
580 * any of the events available.
581 * CCI550 PMU event source ids
582 * 0x0-0x6 - Slave interfaces
583 * 0x8-0xe - Master interfaces
584 * 0xf - Global Events
585 * 0x7 - Reserved
586 */
cci550_validate_hw_event(struct cci_pmu * cci_pmu,unsigned long hw_event)587 static int cci550_validate_hw_event(struct cci_pmu *cci_pmu,
588 unsigned long hw_event)
589 {
590 u32 ev_source = CCI5xx_PMU_EVENT_SOURCE(hw_event);
591 u32 ev_code = CCI5xx_PMU_EVENT_CODE(hw_event);
592 int if_type;
593
594 if (hw_event & ~CCI5xx_PMU_EVENT_MASK)
595 return -ENOENT;
596
597 switch (ev_source) {
598 case CCI5xx_PORT_S0:
599 case CCI5xx_PORT_S1:
600 case CCI5xx_PORT_S2:
601 case CCI5xx_PORT_S3:
602 case CCI5xx_PORT_S4:
603 case CCI5xx_PORT_S5:
604 case CCI5xx_PORT_S6:
605 if_type = CCI_IF_SLAVE;
606 break;
607 case CCI5xx_PORT_M0:
608 case CCI5xx_PORT_M1:
609 case CCI5xx_PORT_M2:
610 case CCI5xx_PORT_M3:
611 case CCI5xx_PORT_M4:
612 case CCI5xx_PORT_M5:
613 case CCI5xx_PORT_M6:
614 if_type = CCI_IF_MASTER;
615 break;
616 case CCI5xx_PORT_GLOBAL:
617 if_type = CCI_IF_GLOBAL;
618 break;
619 default:
620 return -ENOENT;
621 }
622
623 if (ev_code >= cci_pmu->model->event_ranges[if_type].min &&
624 ev_code <= cci_pmu->model->event_ranges[if_type].max)
625 return hw_event;
626
627 return -ENOENT;
628 }
629
630 #endif /* CONFIG_ARM_CCI5xx_PMU */
631
632 /*
633 * Program the CCI PMU counters which have PERF_HES_ARCH set
634 * with the event period and mark them ready before we enable
635 * PMU.
636 */
cci_pmu_sync_counters(struct cci_pmu * cci_pmu)637 static void cci_pmu_sync_counters(struct cci_pmu *cci_pmu)
638 {
639 int i;
640 struct cci_pmu_hw_events *cci_hw = &cci_pmu->hw_events;
641 DECLARE_BITMAP(mask, HW_CNTRS_MAX);
642
643 bitmap_zero(mask, HW_CNTRS_MAX);
644 for_each_set_bit(i, cci_pmu->hw_events.used_mask, cci_pmu->num_cntrs) {
645 struct perf_event *event = cci_hw->events[i];
646
647 if (WARN_ON(!event))
648 continue;
649
650 /* Leave the events which are not counting */
651 if (event->hw.state & PERF_HES_STOPPED)
652 continue;
653 if (event->hw.state & PERF_HES_ARCH) {
654 __set_bit(i, mask);
655 event->hw.state &= ~PERF_HES_ARCH;
656 }
657 }
658
659 pmu_write_counters(cci_pmu, mask);
660 }
661
662 /* Should be called with cci_pmu->hw_events->pmu_lock held */
__cci_pmu_enable_nosync(struct cci_pmu * cci_pmu)663 static void __cci_pmu_enable_nosync(struct cci_pmu *cci_pmu)
664 {
665 u32 val;
666
667 /* Enable all the PMU counters. */
668 val = readl_relaxed(cci_pmu->ctrl_base + CCI_PMCR) | CCI_PMCR_CEN;
669 writel(val, cci_pmu->ctrl_base + CCI_PMCR);
670 }
671
672 /* Should be called with cci_pmu->hw_events->pmu_lock held */
__cci_pmu_enable_sync(struct cci_pmu * cci_pmu)673 static void __cci_pmu_enable_sync(struct cci_pmu *cci_pmu)
674 {
675 cci_pmu_sync_counters(cci_pmu);
676 __cci_pmu_enable_nosync(cci_pmu);
677 }
678
679 /* Should be called with cci_pmu->hw_events->pmu_lock held */
__cci_pmu_disable(struct cci_pmu * cci_pmu)680 static void __cci_pmu_disable(struct cci_pmu *cci_pmu)
681 {
682 u32 val;
683
684 /* Disable all the PMU counters. */
685 val = readl_relaxed(cci_pmu->ctrl_base + CCI_PMCR) & ~CCI_PMCR_CEN;
686 writel(val, cci_pmu->ctrl_base + CCI_PMCR);
687 }
688
cci_pmu_event_show(struct device * dev,struct device_attribute * attr,char * buf)689 static ssize_t cci_pmu_event_show(struct device *dev,
690 struct device_attribute *attr, char *buf)
691 {
692 struct dev_ext_attribute *eattr = container_of(attr,
693 struct dev_ext_attribute, attr);
694 /* source parameter is mandatory for normal PMU events */
695 return sysfs_emit(buf, "source=?,event=0x%lx\n",
696 (unsigned long)eattr->var);
697 }
698
pmu_is_valid_counter(struct cci_pmu * cci_pmu,int idx)699 static int pmu_is_valid_counter(struct cci_pmu *cci_pmu, int idx)
700 {
701 return 0 <= idx && idx <= CCI_PMU_CNTR_LAST(cci_pmu);
702 }
703
pmu_read_register(struct cci_pmu * cci_pmu,int idx,unsigned int offset)704 static u32 pmu_read_register(struct cci_pmu *cci_pmu, int idx, unsigned int offset)
705 {
706 return readl_relaxed(cci_pmu->base +
707 CCI_PMU_CNTR_BASE(cci_pmu->model, idx) + offset);
708 }
709
pmu_write_register(struct cci_pmu * cci_pmu,u32 value,int idx,unsigned int offset)710 static void pmu_write_register(struct cci_pmu *cci_pmu, u32 value,
711 int idx, unsigned int offset)
712 {
713 writel_relaxed(value, cci_pmu->base +
714 CCI_PMU_CNTR_BASE(cci_pmu->model, idx) + offset);
715 }
716
pmu_disable_counter(struct cci_pmu * cci_pmu,int idx)717 static void pmu_disable_counter(struct cci_pmu *cci_pmu, int idx)
718 {
719 pmu_write_register(cci_pmu, 0, idx, CCI_PMU_CNTR_CTRL);
720 }
721
pmu_enable_counter(struct cci_pmu * cci_pmu,int idx)722 static void pmu_enable_counter(struct cci_pmu *cci_pmu, int idx)
723 {
724 pmu_write_register(cci_pmu, 1, idx, CCI_PMU_CNTR_CTRL);
725 }
726
727 static bool __maybe_unused
pmu_counter_is_enabled(struct cci_pmu * cci_pmu,int idx)728 pmu_counter_is_enabled(struct cci_pmu *cci_pmu, int idx)
729 {
730 return (pmu_read_register(cci_pmu, idx, CCI_PMU_CNTR_CTRL) & 0x1) != 0;
731 }
732
pmu_set_event(struct cci_pmu * cci_pmu,int idx,unsigned long event)733 static void pmu_set_event(struct cci_pmu *cci_pmu, int idx, unsigned long event)
734 {
735 pmu_write_register(cci_pmu, event, idx, CCI_PMU_EVT_SEL);
736 }
737
738 /*
739 * For all counters on the CCI-PMU, disable any 'enabled' counters,
740 * saving the changed counters in the mask, so that we can restore
741 * it later using pmu_restore_counters. The mask is private to the
742 * caller. We cannot rely on the used_mask maintained by the CCI_PMU
743 * as it only tells us if the counter is assigned to perf_event or not.
744 * The state of the perf_event cannot be locked by the PMU layer, hence
745 * we check the individual counter status (which can be locked by
746 * cci_pm->hw_events->pmu_lock).
747 *
748 * @mask should be initialised to empty by the caller.
749 */
750 static void __maybe_unused
pmu_save_counters(struct cci_pmu * cci_pmu,unsigned long * mask)751 pmu_save_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
752 {
753 int i;
754
755 for (i = 0; i < cci_pmu->num_cntrs; i++) {
756 if (pmu_counter_is_enabled(cci_pmu, i)) {
757 set_bit(i, mask);
758 pmu_disable_counter(cci_pmu, i);
759 }
760 }
761 }
762
763 /*
764 * Restore the status of the counters. Reversal of the pmu_save_counters().
765 * For each counter set in the mask, enable the counter back.
766 */
767 static void __maybe_unused
pmu_restore_counters(struct cci_pmu * cci_pmu,unsigned long * mask)768 pmu_restore_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
769 {
770 int i;
771
772 for_each_set_bit(i, mask, cci_pmu->num_cntrs)
773 pmu_enable_counter(cci_pmu, i);
774 }
775
776 /*
777 * Returns the number of programmable counters actually implemented
778 * by the cci
779 */
pmu_get_max_counters(struct cci_pmu * cci_pmu)780 static u32 pmu_get_max_counters(struct cci_pmu *cci_pmu)
781 {
782 return (readl_relaxed(cci_pmu->ctrl_base + CCI_PMCR) &
783 CCI_PMCR_NCNT_MASK) >> CCI_PMCR_NCNT_SHIFT;
784 }
785
pmu_get_event_idx(struct cci_pmu_hw_events * hw,struct perf_event * event)786 static int pmu_get_event_idx(struct cci_pmu_hw_events *hw, struct perf_event *event)
787 {
788 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
789 unsigned long cci_event = event->hw.config_base;
790 int idx;
791
792 if (cci_pmu->model->get_event_idx)
793 return cci_pmu->model->get_event_idx(cci_pmu, hw, cci_event);
794
795 /* Generic code to find an unused idx from the mask */
796 for (idx = 0; idx <= CCI_PMU_CNTR_LAST(cci_pmu); idx++)
797 if (!test_and_set_bit(idx, hw->used_mask))
798 return idx;
799
800 /* No counters available */
801 return -EAGAIN;
802 }
803
pmu_map_event(struct perf_event * event)804 static int pmu_map_event(struct perf_event *event)
805 {
806 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
807
808 if (event->attr.type < PERF_TYPE_MAX ||
809 !cci_pmu->model->validate_hw_event)
810 return -ENOENT;
811
812 return cci_pmu->model->validate_hw_event(cci_pmu, event->attr.config);
813 }
814
pmu_request_irq(struct cci_pmu * cci_pmu,irq_handler_t handler)815 static int pmu_request_irq(struct cci_pmu *cci_pmu, irq_handler_t handler)
816 {
817 int i;
818 struct platform_device *pmu_device = cci_pmu->plat_device;
819
820 if (unlikely(!pmu_device))
821 return -ENODEV;
822
823 if (cci_pmu->nr_irqs < 1) {
824 dev_err(&pmu_device->dev, "no irqs for CCI PMUs defined\n");
825 return -ENODEV;
826 }
827
828 /*
829 * Register all available CCI PMU interrupts. In the interrupt handler
830 * we iterate over the counters checking for interrupt source (the
831 * overflowing counter) and clear it.
832 *
833 * This should allow handling of non-unique interrupt for the counters.
834 */
835 for (i = 0; i < cci_pmu->nr_irqs; i++) {
836 int err = request_irq(cci_pmu->irqs[i], handler, IRQF_SHARED,
837 "arm-cci-pmu", cci_pmu);
838 if (err) {
839 dev_err(&pmu_device->dev, "unable to request IRQ%d for ARM CCI PMU counters\n",
840 cci_pmu->irqs[i]);
841 return err;
842 }
843
844 set_bit(i, &cci_pmu->active_irqs);
845 }
846
847 return 0;
848 }
849
pmu_free_irq(struct cci_pmu * cci_pmu)850 static void pmu_free_irq(struct cci_pmu *cci_pmu)
851 {
852 int i;
853
854 for (i = 0; i < cci_pmu->nr_irqs; i++) {
855 if (!test_and_clear_bit(i, &cci_pmu->active_irqs))
856 continue;
857
858 free_irq(cci_pmu->irqs[i], cci_pmu);
859 }
860 }
861
pmu_read_counter(struct perf_event * event)862 static u32 pmu_read_counter(struct perf_event *event)
863 {
864 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
865 struct hw_perf_event *hw_counter = &event->hw;
866 int idx = hw_counter->idx;
867 u32 value;
868
869 if (unlikely(!pmu_is_valid_counter(cci_pmu, idx))) {
870 dev_err(&cci_pmu->plat_device->dev, "Invalid CCI PMU counter %d\n", idx);
871 return 0;
872 }
873 value = pmu_read_register(cci_pmu, idx, CCI_PMU_CNTR);
874
875 return value;
876 }
877
pmu_write_counter(struct cci_pmu * cci_pmu,u32 value,int idx)878 static void pmu_write_counter(struct cci_pmu *cci_pmu, u32 value, int idx)
879 {
880 pmu_write_register(cci_pmu, value, idx, CCI_PMU_CNTR);
881 }
882
__pmu_write_counters(struct cci_pmu * cci_pmu,unsigned long * mask)883 static void __pmu_write_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
884 {
885 int i;
886 struct cci_pmu_hw_events *cci_hw = &cci_pmu->hw_events;
887
888 for_each_set_bit(i, mask, cci_pmu->num_cntrs) {
889 struct perf_event *event = cci_hw->events[i];
890
891 if (WARN_ON(!event))
892 continue;
893 pmu_write_counter(cci_pmu, local64_read(&event->hw.prev_count), i);
894 }
895 }
896
pmu_write_counters(struct cci_pmu * cci_pmu,unsigned long * mask)897 static void pmu_write_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
898 {
899 if (cci_pmu->model->write_counters)
900 cci_pmu->model->write_counters(cci_pmu, mask);
901 else
902 __pmu_write_counters(cci_pmu, mask);
903 }
904
905 #ifdef CONFIG_ARM_CCI5xx_PMU
906
907 /*
908 * CCI-500/CCI-550 has advanced power saving policies, which could gate the
909 * clocks to the PMU counters, which makes the writes to them ineffective.
910 * The only way to write to those counters is when the global counters
911 * are enabled and the particular counter is enabled.
912 *
913 * So we do the following :
914 *
915 * 1) Disable all the PMU counters, saving their current state
916 * 2) Enable the global PMU profiling, now that all counters are
917 * disabled.
918 *
919 * For each counter to be programmed, repeat steps 3-7:
920 *
921 * 3) Write an invalid event code to the event control register for the
922 counter, so that the counters are not modified.
923 * 4) Enable the counter control for the counter.
924 * 5) Set the counter value
925 * 6) Disable the counter
926 * 7) Restore the event in the target counter
927 *
928 * 8) Disable the global PMU.
929 * 9) Restore the status of the rest of the counters.
930 *
931 * We choose an event which for CCI-5xx is guaranteed not to count.
932 * We use the highest possible event code (0x1f) for the master interface 0.
933 */
934 #define CCI5xx_INVALID_EVENT ((CCI5xx_PORT_M0 << CCI5xx_PMU_EVENT_SOURCE_SHIFT) | \
935 (CCI5xx_PMU_EVENT_CODE_MASK << CCI5xx_PMU_EVENT_CODE_SHIFT))
cci5xx_pmu_write_counters(struct cci_pmu * cci_pmu,unsigned long * mask)936 static void cci5xx_pmu_write_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
937 {
938 int i;
939 DECLARE_BITMAP(saved_mask, HW_CNTRS_MAX);
940
941 bitmap_zero(saved_mask, cci_pmu->num_cntrs);
942 pmu_save_counters(cci_pmu, saved_mask);
943
944 /*
945 * Now that all the counters are disabled, we can safely turn the PMU on,
946 * without syncing the status of the counters
947 */
948 __cci_pmu_enable_nosync(cci_pmu);
949
950 for_each_set_bit(i, mask, cci_pmu->num_cntrs) {
951 struct perf_event *event = cci_pmu->hw_events.events[i];
952
953 if (WARN_ON(!event))
954 continue;
955
956 pmu_set_event(cci_pmu, i, CCI5xx_INVALID_EVENT);
957 pmu_enable_counter(cci_pmu, i);
958 pmu_write_counter(cci_pmu, local64_read(&event->hw.prev_count), i);
959 pmu_disable_counter(cci_pmu, i);
960 pmu_set_event(cci_pmu, i, event->hw.config_base);
961 }
962
963 __cci_pmu_disable(cci_pmu);
964
965 pmu_restore_counters(cci_pmu, saved_mask);
966 }
967
968 #endif /* CONFIG_ARM_CCI5xx_PMU */
969
pmu_event_update(struct perf_event * event)970 static u64 pmu_event_update(struct perf_event *event)
971 {
972 struct hw_perf_event *hwc = &event->hw;
973 u64 delta, prev_raw_count, new_raw_count;
974
975 do {
976 prev_raw_count = local64_read(&hwc->prev_count);
977 new_raw_count = pmu_read_counter(event);
978 } while (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
979 new_raw_count) != prev_raw_count);
980
981 delta = (new_raw_count - prev_raw_count) & CCI_PMU_CNTR_MASK;
982
983 local64_add(delta, &event->count);
984
985 return new_raw_count;
986 }
987
pmu_read(struct perf_event * event)988 static void pmu_read(struct perf_event *event)
989 {
990 pmu_event_update(event);
991 }
992
pmu_event_set_period(struct perf_event * event)993 static void pmu_event_set_period(struct perf_event *event)
994 {
995 struct hw_perf_event *hwc = &event->hw;
996 /*
997 * The CCI PMU counters have a period of 2^32. To account for the
998 * possiblity of extreme interrupt latency we program for a period of
999 * half that. Hopefully we can handle the interrupt before another 2^31
1000 * events occur and the counter overtakes its previous value.
1001 */
1002 u64 val = 1ULL << 31;
1003 local64_set(&hwc->prev_count, val);
1004
1005 /*
1006 * CCI PMU uses PERF_HES_ARCH to keep track of the counters, whose
1007 * values needs to be sync-ed with the s/w state before the PMU is
1008 * enabled.
1009 * Mark this counter for sync.
1010 */
1011 hwc->state |= PERF_HES_ARCH;
1012 }
1013
pmu_handle_irq(int irq_num,void * dev)1014 static irqreturn_t pmu_handle_irq(int irq_num, void *dev)
1015 {
1016 struct cci_pmu *cci_pmu = dev;
1017 struct cci_pmu_hw_events *events = &cci_pmu->hw_events;
1018 int idx, handled = IRQ_NONE;
1019
1020 raw_spin_lock(&events->pmu_lock);
1021
1022 /* Disable the PMU while we walk through the counters */
1023 __cci_pmu_disable(cci_pmu);
1024 /*
1025 * Iterate over counters and update the corresponding perf events.
1026 * This should work regardless of whether we have per-counter overflow
1027 * interrupt or a combined overflow interrupt.
1028 */
1029 for (idx = 0; idx <= CCI_PMU_CNTR_LAST(cci_pmu); idx++) {
1030 struct perf_event *event = events->events[idx];
1031
1032 if (!event)
1033 continue;
1034
1035 /* Did this counter overflow? */
1036 if (!(pmu_read_register(cci_pmu, idx, CCI_PMU_OVRFLW) &
1037 CCI_PMU_OVRFLW_FLAG))
1038 continue;
1039
1040 pmu_write_register(cci_pmu, CCI_PMU_OVRFLW_FLAG, idx,
1041 CCI_PMU_OVRFLW);
1042
1043 pmu_event_update(event);
1044 pmu_event_set_period(event);
1045 handled = IRQ_HANDLED;
1046 }
1047
1048 /* Enable the PMU and sync possibly overflowed counters */
1049 __cci_pmu_enable_sync(cci_pmu);
1050 raw_spin_unlock(&events->pmu_lock);
1051
1052 return IRQ_RETVAL(handled);
1053 }
1054
cci_pmu_get_hw(struct cci_pmu * cci_pmu)1055 static int cci_pmu_get_hw(struct cci_pmu *cci_pmu)
1056 {
1057 int ret = pmu_request_irq(cci_pmu, pmu_handle_irq);
1058 if (ret) {
1059 pmu_free_irq(cci_pmu);
1060 return ret;
1061 }
1062 return 0;
1063 }
1064
cci_pmu_put_hw(struct cci_pmu * cci_pmu)1065 static void cci_pmu_put_hw(struct cci_pmu *cci_pmu)
1066 {
1067 pmu_free_irq(cci_pmu);
1068 }
1069
hw_perf_event_destroy(struct perf_event * event)1070 static void hw_perf_event_destroy(struct perf_event *event)
1071 {
1072 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1073 atomic_t *active_events = &cci_pmu->active_events;
1074 struct mutex *reserve_mutex = &cci_pmu->reserve_mutex;
1075
1076 if (atomic_dec_and_mutex_lock(active_events, reserve_mutex)) {
1077 cci_pmu_put_hw(cci_pmu);
1078 mutex_unlock(reserve_mutex);
1079 }
1080 }
1081
cci_pmu_enable(struct pmu * pmu)1082 static void cci_pmu_enable(struct pmu *pmu)
1083 {
1084 struct cci_pmu *cci_pmu = to_cci_pmu(pmu);
1085 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1086 bool enabled = !bitmap_empty(hw_events->used_mask, cci_pmu->num_cntrs);
1087 unsigned long flags;
1088
1089 if (!enabled)
1090 return;
1091
1092 raw_spin_lock_irqsave(&hw_events->pmu_lock, flags);
1093 __cci_pmu_enable_sync(cci_pmu);
1094 raw_spin_unlock_irqrestore(&hw_events->pmu_lock, flags);
1095
1096 }
1097
cci_pmu_disable(struct pmu * pmu)1098 static void cci_pmu_disable(struct pmu *pmu)
1099 {
1100 struct cci_pmu *cci_pmu = to_cci_pmu(pmu);
1101 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1102 unsigned long flags;
1103
1104 raw_spin_lock_irqsave(&hw_events->pmu_lock, flags);
1105 __cci_pmu_disable(cci_pmu);
1106 raw_spin_unlock_irqrestore(&hw_events->pmu_lock, flags);
1107 }
1108
1109 /*
1110 * Check if the idx represents a non-programmable counter.
1111 * All the fixed event counters are mapped before the programmable
1112 * counters.
1113 */
pmu_fixed_hw_idx(struct cci_pmu * cci_pmu,int idx)1114 static bool pmu_fixed_hw_idx(struct cci_pmu *cci_pmu, int idx)
1115 {
1116 return (idx >= 0) && (idx < cci_pmu->model->fixed_hw_cntrs);
1117 }
1118
cci_pmu_start(struct perf_event * event,int pmu_flags)1119 static void cci_pmu_start(struct perf_event *event, int pmu_flags)
1120 {
1121 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1122 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1123 struct hw_perf_event *hwc = &event->hw;
1124 int idx = hwc->idx;
1125 unsigned long flags;
1126
1127 /*
1128 * To handle interrupt latency, we always reprogram the period
1129 * regardless of PERF_EF_RELOAD.
1130 */
1131 if (pmu_flags & PERF_EF_RELOAD)
1132 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
1133
1134 hwc->state = 0;
1135
1136 if (unlikely(!pmu_is_valid_counter(cci_pmu, idx))) {
1137 dev_err(&cci_pmu->plat_device->dev, "Invalid CCI PMU counter %d\n", idx);
1138 return;
1139 }
1140
1141 raw_spin_lock_irqsave(&hw_events->pmu_lock, flags);
1142
1143 /* Configure the counter unless you are counting a fixed event */
1144 if (!pmu_fixed_hw_idx(cci_pmu, idx))
1145 pmu_set_event(cci_pmu, idx, hwc->config_base);
1146
1147 pmu_event_set_period(event);
1148 pmu_enable_counter(cci_pmu, idx);
1149
1150 raw_spin_unlock_irqrestore(&hw_events->pmu_lock, flags);
1151 }
1152
cci_pmu_stop(struct perf_event * event,int pmu_flags)1153 static void cci_pmu_stop(struct perf_event *event, int pmu_flags)
1154 {
1155 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1156 struct hw_perf_event *hwc = &event->hw;
1157 int idx = hwc->idx;
1158
1159 if (hwc->state & PERF_HES_STOPPED)
1160 return;
1161
1162 if (unlikely(!pmu_is_valid_counter(cci_pmu, idx))) {
1163 dev_err(&cci_pmu->plat_device->dev, "Invalid CCI PMU counter %d\n", idx);
1164 return;
1165 }
1166
1167 /*
1168 * We always reprogram the counter, so ignore PERF_EF_UPDATE. See
1169 * cci_pmu_start()
1170 */
1171 pmu_disable_counter(cci_pmu, idx);
1172 pmu_event_update(event);
1173 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
1174 }
1175
cci_pmu_add(struct perf_event * event,int flags)1176 static int cci_pmu_add(struct perf_event *event, int flags)
1177 {
1178 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1179 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1180 struct hw_perf_event *hwc = &event->hw;
1181 int idx;
1182
1183 /* If we don't have a space for the counter then finish early. */
1184 idx = pmu_get_event_idx(hw_events, event);
1185 if (idx < 0)
1186 return idx;
1187
1188 event->hw.idx = idx;
1189 hw_events->events[idx] = event;
1190
1191 hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
1192 if (flags & PERF_EF_START)
1193 cci_pmu_start(event, PERF_EF_RELOAD);
1194
1195 /* Propagate our changes to the userspace mapping. */
1196 perf_event_update_userpage(event);
1197
1198 return 0;
1199 }
1200
cci_pmu_del(struct perf_event * event,int flags)1201 static void cci_pmu_del(struct perf_event *event, int flags)
1202 {
1203 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1204 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1205 struct hw_perf_event *hwc = &event->hw;
1206 int idx = hwc->idx;
1207
1208 cci_pmu_stop(event, PERF_EF_UPDATE);
1209 hw_events->events[idx] = NULL;
1210 clear_bit(idx, hw_events->used_mask);
1211
1212 perf_event_update_userpage(event);
1213 }
1214
validate_event(struct pmu * cci_pmu,struct cci_pmu_hw_events * hw_events,struct perf_event * event)1215 static int validate_event(struct pmu *cci_pmu,
1216 struct cci_pmu_hw_events *hw_events,
1217 struct perf_event *event)
1218 {
1219 if (is_software_event(event))
1220 return 1;
1221
1222 /*
1223 * Reject groups spanning multiple HW PMUs (e.g. CPU + CCI). The
1224 * core perf code won't check that the pmu->ctx == leader->ctx
1225 * until after pmu->event_init(event).
1226 */
1227 if (event->pmu != cci_pmu)
1228 return 0;
1229
1230 if (event->state < PERF_EVENT_STATE_OFF)
1231 return 1;
1232
1233 if (event->state == PERF_EVENT_STATE_OFF && !event->attr.enable_on_exec)
1234 return 1;
1235
1236 return pmu_get_event_idx(hw_events, event) >= 0;
1237 }
1238
validate_group(struct perf_event * event)1239 static int validate_group(struct perf_event *event)
1240 {
1241 struct perf_event *sibling, *leader = event->group_leader;
1242 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1243 unsigned long mask[BITS_TO_LONGS(HW_CNTRS_MAX)];
1244 struct cci_pmu_hw_events fake_pmu = {
1245 /*
1246 * Initialise the fake PMU. We only need to populate the
1247 * used_mask for the purposes of validation.
1248 */
1249 .used_mask = mask,
1250 };
1251 bitmap_zero(mask, cci_pmu->num_cntrs);
1252
1253 if (!validate_event(event->pmu, &fake_pmu, leader))
1254 return -EINVAL;
1255
1256 for_each_sibling_event(sibling, leader) {
1257 if (!validate_event(event->pmu, &fake_pmu, sibling))
1258 return -EINVAL;
1259 }
1260
1261 if (!validate_event(event->pmu, &fake_pmu, event))
1262 return -EINVAL;
1263
1264 return 0;
1265 }
1266
__hw_perf_event_init(struct perf_event * event)1267 static int __hw_perf_event_init(struct perf_event *event)
1268 {
1269 struct hw_perf_event *hwc = &event->hw;
1270 int mapping;
1271
1272 mapping = pmu_map_event(event);
1273
1274 if (mapping < 0) {
1275 pr_debug("event %x:%llx not supported\n", event->attr.type,
1276 event->attr.config);
1277 return mapping;
1278 }
1279
1280 /*
1281 * We don't assign an index until we actually place the event onto
1282 * hardware. Use -1 to signify that we haven't decided where to put it
1283 * yet.
1284 */
1285 hwc->idx = -1;
1286 hwc->config_base = 0;
1287 hwc->config = 0;
1288 hwc->event_base = 0;
1289
1290 /*
1291 * Store the event encoding into the config_base field.
1292 */
1293 hwc->config_base |= (unsigned long)mapping;
1294
1295 if (event->group_leader != event) {
1296 if (validate_group(event) != 0)
1297 return -EINVAL;
1298 }
1299
1300 return 0;
1301 }
1302
cci_pmu_event_init(struct perf_event * event)1303 static int cci_pmu_event_init(struct perf_event *event)
1304 {
1305 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1306 atomic_t *active_events = &cci_pmu->active_events;
1307 int err = 0;
1308
1309 if (event->attr.type != event->pmu->type)
1310 return -ENOENT;
1311
1312 /* Shared by all CPUs, no meaningful state to sample */
1313 if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
1314 return -EOPNOTSUPP;
1315
1316 /*
1317 * Following the example set by other "uncore" PMUs, we accept any CPU
1318 * and rewrite its affinity dynamically rather than having perf core
1319 * handle cpu == -1 and pid == -1 for this case.
1320 *
1321 * The perf core will pin online CPUs for the duration of this call and
1322 * the event being installed into its context, so the PMU's CPU can't
1323 * change under our feet.
1324 */
1325 if (event->cpu < 0)
1326 return -EINVAL;
1327 event->cpu = cci_pmu->cpu;
1328
1329 event->destroy = hw_perf_event_destroy;
1330 if (!atomic_inc_not_zero(active_events)) {
1331 mutex_lock(&cci_pmu->reserve_mutex);
1332 if (atomic_read(active_events) == 0)
1333 err = cci_pmu_get_hw(cci_pmu);
1334 if (!err)
1335 atomic_inc(active_events);
1336 mutex_unlock(&cci_pmu->reserve_mutex);
1337 }
1338 if (err)
1339 return err;
1340
1341 err = __hw_perf_event_init(event);
1342 if (err)
1343 hw_perf_event_destroy(event);
1344
1345 return err;
1346 }
1347
pmu_cpumask_attr_show(struct device * dev,struct device_attribute * attr,char * buf)1348 static ssize_t pmu_cpumask_attr_show(struct device *dev,
1349 struct device_attribute *attr, char *buf)
1350 {
1351 struct pmu *pmu = dev_get_drvdata(dev);
1352 struct cci_pmu *cci_pmu = to_cci_pmu(pmu);
1353
1354 return cpumap_print_to_pagebuf(true, buf, cpumask_of(cci_pmu->cpu));
1355 }
1356
1357 static struct device_attribute pmu_cpumask_attr =
1358 __ATTR(cpumask, S_IRUGO, pmu_cpumask_attr_show, NULL);
1359
1360 static struct attribute *pmu_attrs[] = {
1361 &pmu_cpumask_attr.attr,
1362 NULL,
1363 };
1364
1365 static const struct attribute_group pmu_attr_group = {
1366 .attrs = pmu_attrs,
1367 };
1368
1369 static struct attribute_group pmu_format_attr_group = {
1370 .name = "format",
1371 .attrs = NULL, /* Filled in cci_pmu_init_attrs */
1372 };
1373
1374 static struct attribute_group pmu_event_attr_group = {
1375 .name = "events",
1376 .attrs = NULL, /* Filled in cci_pmu_init_attrs */
1377 };
1378
1379 static const struct attribute_group *pmu_attr_groups[] = {
1380 &pmu_attr_group,
1381 &pmu_format_attr_group,
1382 &pmu_event_attr_group,
1383 NULL
1384 };
1385
cci_pmu_init(struct cci_pmu * cci_pmu,struct platform_device * pdev)1386 static int cci_pmu_init(struct cci_pmu *cci_pmu, struct platform_device *pdev)
1387 {
1388 const struct cci_pmu_model *model = cci_pmu->model;
1389 char *name = model->name;
1390 u32 num_cntrs;
1391
1392 if (WARN_ON(model->num_hw_cntrs > NUM_HW_CNTRS_MAX))
1393 return -EINVAL;
1394 if (WARN_ON(model->fixed_hw_cntrs > FIXED_HW_CNTRS_MAX))
1395 return -EINVAL;
1396
1397 pmu_event_attr_group.attrs = model->event_attrs;
1398 pmu_format_attr_group.attrs = model->format_attrs;
1399
1400 cci_pmu->pmu = (struct pmu) {
1401 .module = THIS_MODULE,
1402 .parent = &pdev->dev,
1403 .name = cci_pmu->model->name,
1404 .task_ctx_nr = perf_invalid_context,
1405 .pmu_enable = cci_pmu_enable,
1406 .pmu_disable = cci_pmu_disable,
1407 .event_init = cci_pmu_event_init,
1408 .add = cci_pmu_add,
1409 .del = cci_pmu_del,
1410 .start = cci_pmu_start,
1411 .stop = cci_pmu_stop,
1412 .read = pmu_read,
1413 .attr_groups = pmu_attr_groups,
1414 .capabilities = PERF_PMU_CAP_NO_EXCLUDE,
1415 };
1416
1417 cci_pmu->plat_device = pdev;
1418 num_cntrs = pmu_get_max_counters(cci_pmu);
1419 if (num_cntrs > cci_pmu->model->num_hw_cntrs) {
1420 dev_warn(&pdev->dev,
1421 "PMU implements more counters(%d) than supported by"
1422 " the model(%d), truncated.",
1423 num_cntrs, cci_pmu->model->num_hw_cntrs);
1424 num_cntrs = cci_pmu->model->num_hw_cntrs;
1425 }
1426 cci_pmu->num_cntrs = num_cntrs + cci_pmu->model->fixed_hw_cntrs;
1427
1428 return perf_pmu_register(&cci_pmu->pmu, name, -1);
1429 }
1430
cci_pmu_offline_cpu(unsigned int cpu)1431 static int cci_pmu_offline_cpu(unsigned int cpu)
1432 {
1433 int target;
1434
1435 if (!g_cci_pmu || cpu != g_cci_pmu->cpu)
1436 return 0;
1437
1438 target = cpumask_any_but(cpu_online_mask, cpu);
1439 if (target >= nr_cpu_ids)
1440 return 0;
1441
1442 perf_pmu_migrate_context(&g_cci_pmu->pmu, cpu, target);
1443 g_cci_pmu->cpu = target;
1444 return 0;
1445 }
1446
1447 static __maybe_unused struct cci_pmu_model cci_pmu_models[] = {
1448 #ifdef CONFIG_ARM_CCI400_PMU
1449 [CCI400_R0] = {
1450 .name = "CCI_400",
1451 .fixed_hw_cntrs = FIXED_HW_CNTRS_CII_4XX, /* Cycle counter */
1452 .num_hw_cntrs = NUM_HW_CNTRS_CII_4XX,
1453 .cntr_size = SZ_4K,
1454 .format_attrs = cci400_pmu_format_attrs,
1455 .event_attrs = cci400_r0_pmu_event_attrs,
1456 .event_ranges = {
1457 [CCI_IF_SLAVE] = {
1458 CCI400_R0_SLAVE_PORT_MIN_EV,
1459 CCI400_R0_SLAVE_PORT_MAX_EV,
1460 },
1461 [CCI_IF_MASTER] = {
1462 CCI400_R0_MASTER_PORT_MIN_EV,
1463 CCI400_R0_MASTER_PORT_MAX_EV,
1464 },
1465 },
1466 .validate_hw_event = cci400_validate_hw_event,
1467 .get_event_idx = cci400_get_event_idx,
1468 },
1469 [CCI400_R1] = {
1470 .name = "CCI_400_r1",
1471 .fixed_hw_cntrs = FIXED_HW_CNTRS_CII_4XX, /* Cycle counter */
1472 .num_hw_cntrs = NUM_HW_CNTRS_CII_4XX,
1473 .cntr_size = SZ_4K,
1474 .format_attrs = cci400_pmu_format_attrs,
1475 .event_attrs = cci400_r1_pmu_event_attrs,
1476 .event_ranges = {
1477 [CCI_IF_SLAVE] = {
1478 CCI400_R1_SLAVE_PORT_MIN_EV,
1479 CCI400_R1_SLAVE_PORT_MAX_EV,
1480 },
1481 [CCI_IF_MASTER] = {
1482 CCI400_R1_MASTER_PORT_MIN_EV,
1483 CCI400_R1_MASTER_PORT_MAX_EV,
1484 },
1485 },
1486 .validate_hw_event = cci400_validate_hw_event,
1487 .get_event_idx = cci400_get_event_idx,
1488 },
1489 #endif
1490 #ifdef CONFIG_ARM_CCI5xx_PMU
1491 [CCI500_R0] = {
1492 .name = "CCI_500",
1493 .fixed_hw_cntrs = FIXED_HW_CNTRS_CII_5XX,
1494 .num_hw_cntrs = NUM_HW_CNTRS_CII_5XX,
1495 .cntr_size = SZ_64K,
1496 .format_attrs = cci5xx_pmu_format_attrs,
1497 .event_attrs = cci5xx_pmu_event_attrs,
1498 .event_ranges = {
1499 [CCI_IF_SLAVE] = {
1500 CCI5xx_SLAVE_PORT_MIN_EV,
1501 CCI5xx_SLAVE_PORT_MAX_EV,
1502 },
1503 [CCI_IF_MASTER] = {
1504 CCI5xx_MASTER_PORT_MIN_EV,
1505 CCI5xx_MASTER_PORT_MAX_EV,
1506 },
1507 [CCI_IF_GLOBAL] = {
1508 CCI5xx_GLOBAL_PORT_MIN_EV,
1509 CCI5xx_GLOBAL_PORT_MAX_EV,
1510 },
1511 },
1512 .validate_hw_event = cci500_validate_hw_event,
1513 .write_counters = cci5xx_pmu_write_counters,
1514 },
1515 [CCI550_R0] = {
1516 .name = "CCI_550",
1517 .fixed_hw_cntrs = FIXED_HW_CNTRS_CII_5XX,
1518 .num_hw_cntrs = NUM_HW_CNTRS_CII_5XX,
1519 .cntr_size = SZ_64K,
1520 .format_attrs = cci5xx_pmu_format_attrs,
1521 .event_attrs = cci5xx_pmu_event_attrs,
1522 .event_ranges = {
1523 [CCI_IF_SLAVE] = {
1524 CCI5xx_SLAVE_PORT_MIN_EV,
1525 CCI5xx_SLAVE_PORT_MAX_EV,
1526 },
1527 [CCI_IF_MASTER] = {
1528 CCI5xx_MASTER_PORT_MIN_EV,
1529 CCI5xx_MASTER_PORT_MAX_EV,
1530 },
1531 [CCI_IF_GLOBAL] = {
1532 CCI5xx_GLOBAL_PORT_MIN_EV,
1533 CCI5xx_GLOBAL_PORT_MAX_EV,
1534 },
1535 },
1536 .validate_hw_event = cci550_validate_hw_event,
1537 .write_counters = cci5xx_pmu_write_counters,
1538 },
1539 #endif
1540 };
1541
1542 static const struct of_device_id arm_cci_pmu_matches[] = {
1543 #ifdef CONFIG_ARM_CCI400_PMU
1544 {
1545 .compatible = "arm,cci-400-pmu",
1546 .data = NULL,
1547 },
1548 {
1549 .compatible = "arm,cci-400-pmu,r0",
1550 .data = &cci_pmu_models[CCI400_R0],
1551 },
1552 {
1553 .compatible = "arm,cci-400-pmu,r1",
1554 .data = &cci_pmu_models[CCI400_R1],
1555 },
1556 #endif
1557 #ifdef CONFIG_ARM_CCI5xx_PMU
1558 {
1559 .compatible = "arm,cci-500-pmu,r0",
1560 .data = &cci_pmu_models[CCI500_R0],
1561 },
1562 {
1563 .compatible = "arm,cci-550-pmu,r0",
1564 .data = &cci_pmu_models[CCI550_R0],
1565 },
1566 #endif
1567 {},
1568 };
1569 MODULE_DEVICE_TABLE(of, arm_cci_pmu_matches);
1570
is_duplicate_irq(int irq,int * irqs,int nr_irqs)1571 static bool is_duplicate_irq(int irq, int *irqs, int nr_irqs)
1572 {
1573 int i;
1574
1575 for (i = 0; i < nr_irqs; i++)
1576 if (irq == irqs[i])
1577 return true;
1578
1579 return false;
1580 }
1581
cci_pmu_alloc(struct device * dev)1582 static struct cci_pmu *cci_pmu_alloc(struct device *dev)
1583 {
1584 struct cci_pmu *cci_pmu;
1585 const struct cci_pmu_model *model;
1586
1587 /*
1588 * All allocations are devm_* hence we don't have to free
1589 * them explicitly on an error, as it would end up in driver
1590 * detach.
1591 */
1592 cci_pmu = devm_kzalloc(dev, sizeof(*cci_pmu), GFP_KERNEL);
1593 if (!cci_pmu)
1594 return ERR_PTR(-ENOMEM);
1595
1596 cci_pmu->ctrl_base = *(void __iomem **)dev->platform_data;
1597
1598 model = of_device_get_match_data(dev);
1599 if (!model) {
1600 dev_warn(dev,
1601 "DEPRECATED compatible property, requires secure access to CCI registers");
1602 model = probe_cci_model(cci_pmu);
1603 }
1604 if (!model) {
1605 dev_warn(dev, "CCI PMU version not supported\n");
1606 return ERR_PTR(-ENODEV);
1607 }
1608
1609 cci_pmu->model = model;
1610 cci_pmu->irqs = devm_kcalloc(dev, CCI_PMU_MAX_HW_CNTRS(model),
1611 sizeof(*cci_pmu->irqs), GFP_KERNEL);
1612 if (!cci_pmu->irqs)
1613 return ERR_PTR(-ENOMEM);
1614 cci_pmu->hw_events.events = devm_kcalloc(dev,
1615 CCI_PMU_MAX_HW_CNTRS(model),
1616 sizeof(*cci_pmu->hw_events.events),
1617 GFP_KERNEL);
1618 if (!cci_pmu->hw_events.events)
1619 return ERR_PTR(-ENOMEM);
1620 cci_pmu->hw_events.used_mask = devm_bitmap_zalloc(dev,
1621 CCI_PMU_MAX_HW_CNTRS(model),
1622 GFP_KERNEL);
1623 if (!cci_pmu->hw_events.used_mask)
1624 return ERR_PTR(-ENOMEM);
1625
1626 return cci_pmu;
1627 }
1628
cci_pmu_probe(struct platform_device * pdev)1629 static int cci_pmu_probe(struct platform_device *pdev)
1630 {
1631 struct cci_pmu *cci_pmu;
1632 int i, ret, irq;
1633
1634 cci_pmu = cci_pmu_alloc(&pdev->dev);
1635 if (IS_ERR(cci_pmu))
1636 return PTR_ERR(cci_pmu);
1637
1638 cci_pmu->base = devm_platform_ioremap_resource(pdev, 0);
1639 if (IS_ERR(cci_pmu->base))
1640 return -ENOMEM;
1641
1642 /*
1643 * CCI PMU has one overflow interrupt per counter; but some may be tied
1644 * together to a common interrupt.
1645 */
1646 cci_pmu->nr_irqs = 0;
1647 for (i = 0; i < CCI_PMU_MAX_HW_CNTRS(cci_pmu->model); i++) {
1648 irq = platform_get_irq(pdev, i);
1649 if (irq < 0)
1650 break;
1651
1652 if (is_duplicate_irq(irq, cci_pmu->irqs, cci_pmu->nr_irqs))
1653 continue;
1654
1655 cci_pmu->irqs[cci_pmu->nr_irqs++] = irq;
1656 }
1657
1658 /*
1659 * Ensure that the device tree has as many interrupts as the number
1660 * of counters.
1661 */
1662 if (i < CCI_PMU_MAX_HW_CNTRS(cci_pmu->model)) {
1663 dev_warn(&pdev->dev, "In-correct number of interrupts: %d, should be %d\n",
1664 i, CCI_PMU_MAX_HW_CNTRS(cci_pmu->model));
1665 return -EINVAL;
1666 }
1667
1668 raw_spin_lock_init(&cci_pmu->hw_events.pmu_lock);
1669 mutex_init(&cci_pmu->reserve_mutex);
1670 atomic_set(&cci_pmu->active_events, 0);
1671
1672 cci_pmu->cpu = raw_smp_processor_id();
1673 g_cci_pmu = cci_pmu;
1674 cpuhp_setup_state_nocalls(CPUHP_AP_PERF_ARM_CCI_ONLINE,
1675 "perf/arm/cci:online", NULL,
1676 cci_pmu_offline_cpu);
1677
1678 ret = cci_pmu_init(cci_pmu, pdev);
1679 if (ret)
1680 goto error_pmu_init;
1681
1682 pr_info("ARM %s PMU driver probed", cci_pmu->model->name);
1683 return 0;
1684
1685 error_pmu_init:
1686 cpuhp_remove_state(CPUHP_AP_PERF_ARM_CCI_ONLINE);
1687 g_cci_pmu = NULL;
1688 return ret;
1689 }
1690
cci_pmu_remove(struct platform_device * pdev)1691 static void cci_pmu_remove(struct platform_device *pdev)
1692 {
1693 if (!g_cci_pmu)
1694 return;
1695
1696 cpuhp_remove_state(CPUHP_AP_PERF_ARM_CCI_ONLINE);
1697 perf_pmu_unregister(&g_cci_pmu->pmu);
1698 g_cci_pmu = NULL;
1699 }
1700
1701 static struct platform_driver cci_pmu_driver = {
1702 .driver = {
1703 .name = DRIVER_NAME,
1704 .of_match_table = arm_cci_pmu_matches,
1705 .suppress_bind_attrs = true,
1706 },
1707 .probe = cci_pmu_probe,
1708 .remove_new = cci_pmu_remove,
1709 };
1710
1711 module_platform_driver(cci_pmu_driver);
1712 MODULE_LICENSE("GPL v2");
1713 MODULE_DESCRIPTION("ARM CCI PMU support");
1714