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>
6 #include <linux/arm-cci.h>
8 #include <linux/interrupt.h>
9 #include <linux/module.h>
10 #include <linux/of_address.h>
11 #include <linux/of_device.h>
12 #include <linux/of_irq.h>
13 #include <linux/of_platform.h>
14 #include <linux/perf_event.h>
15 #include <linux/platform_device.h>
16 #include <linux/slab.h>
17 #include <linux/spinlock.h>
19 #define DRIVER_NAME "ARM-CCI PMU"
21 #define CCI_PMCR 0x0100
22 #define CCI_PID2 0x0fe8
24 #define CCI_PMCR_CEN 0x00000001
25 #define CCI_PMCR_NCNT_MASK 0x0000f800
26 #define CCI_PMCR_NCNT_SHIFT 11
28 #define CCI_PID2_REV_MASK 0xf0
29 #define CCI_PID2_REV_SHIFT 4
31 #define CCI_PMU_EVT_SEL 0x000
32 #define CCI_PMU_CNTR 0x004
33 #define CCI_PMU_CNTR_CTRL 0x008
34 #define CCI_PMU_OVRFLW 0x00c
36 #define CCI_PMU_OVRFLW_FLAG 1
38 #define CCI_PMU_CNTR_SIZE(model) ((model)->cntr_size)
39 #define CCI_PMU_CNTR_BASE(model, idx) ((idx) * CCI_PMU_CNTR_SIZE(model))
40 #define CCI_PMU_CNTR_MASK ((1ULL << 32) -1)
41 #define CCI_PMU_CNTR_LAST(cci_pmu) (cci_pmu->num_cntrs - 1)
43 #define CCI_PMU_MAX_HW_CNTRS(model) \
44 ((model)->num_hw_cntrs + (model)->fixed_hw_cntrs)
46 /* Types of interfaces that can generate events */
50 #ifdef CONFIG_ARM_CCI5xx_PMU
56 #define NUM_HW_CNTRS_CII_4XX 4
57 #define NUM_HW_CNTRS_CII_5XX 8
58 #define NUM_HW_CNTRS_MAX NUM_HW_CNTRS_CII_5XX
60 #define FIXED_HW_CNTRS_CII_4XX 1
61 #define FIXED_HW_CNTRS_CII_5XX 0
62 #define FIXED_HW_CNTRS_MAX FIXED_HW_CNTRS_CII_4XX
64 #define HW_CNTRS_MAX (NUM_HW_CNTRS_MAX + FIXED_HW_CNTRS_MAX)
71 struct cci_pmu_hw_events {
72 struct perf_event **events;
73 unsigned long *used_mask;
74 raw_spinlock_t pmu_lock;
79 * struct cci_pmu_model:
80 * @fixed_hw_cntrs - Number of fixed event counters
81 * @num_hw_cntrs - Maximum number of programmable event counters
82 * @cntr_size - Size of an event counter mapping
84 struct cci_pmu_model {
89 struct attribute **format_attrs;
90 struct attribute **event_attrs;
91 struct event_range event_ranges[CCI_IF_MAX];
92 int (*validate_hw_event)(struct cci_pmu *, unsigned long);
93 int (*get_event_idx)(struct cci_pmu *, struct cci_pmu_hw_events *, unsigned long);
94 void (*write_counters)(struct cci_pmu *, unsigned long *);
97 static struct cci_pmu_model cci_pmu_models[];
101 void __iomem *ctrl_base;
106 unsigned long active_irqs;
107 const struct cci_pmu_model *model;
108 struct cci_pmu_hw_events hw_events;
109 struct platform_device *plat_device;
111 atomic_t active_events;
112 struct mutex reserve_mutex;
115 #define to_cci_pmu(c) (container_of(c, struct cci_pmu, pmu))
117 static struct cci_pmu *g_cci_pmu;
120 #ifdef CONFIG_ARM_CCI400_PMU
124 #ifdef CONFIG_ARM_CCI5xx_PMU
131 static void pmu_write_counters(struct cci_pmu *cci_pmu,
132 unsigned long *mask);
133 static ssize_t __maybe_unused cci_pmu_format_show(struct device *dev,
134 struct device_attribute *attr, char *buf);
135 static ssize_t __maybe_unused cci_pmu_event_show(struct device *dev,
136 struct device_attribute *attr, char *buf);
138 #define CCI_EXT_ATTR_ENTRY(_name, _func, _config) \
139 &((struct dev_ext_attribute[]) { \
140 { __ATTR(_name, S_IRUGO, _func, NULL), (void *)_config } \
143 #define CCI_FORMAT_EXT_ATTR_ENTRY(_name, _config) \
144 CCI_EXT_ATTR_ENTRY(_name, cci_pmu_format_show, (char *)_config)
145 #define CCI_EVENT_EXT_ATTR_ENTRY(_name, _config) \
146 CCI_EXT_ATTR_ENTRY(_name, cci_pmu_event_show, (unsigned long)_config)
148 /* CCI400 PMU Specific definitions */
150 #ifdef CONFIG_ARM_CCI400_PMU
153 #define CCI400_PORT_S0 0
154 #define CCI400_PORT_S1 1
155 #define CCI400_PORT_S2 2
156 #define CCI400_PORT_S3 3
157 #define CCI400_PORT_S4 4
158 #define CCI400_PORT_M0 5
159 #define CCI400_PORT_M1 6
160 #define CCI400_PORT_M2 7
162 #define CCI400_R1_PX 5
165 * Instead of an event id to monitor CCI cycles, a dedicated counter is
166 * provided. Use 0xff to represent CCI cycles and hope that no future revisions
167 * make use of this event in hardware.
169 enum cci400_perf_events {
170 CCI400_PMU_CYCLES = 0xff
173 #define CCI400_PMU_CYCLE_CNTR_IDX 0
174 #define CCI400_PMU_CNTR0_IDX 1
177 * CCI PMU event id is an 8-bit value made of two parts - bits 7:5 for one of 8
178 * ports and bits 4:0 are event codes. There are different event codes
179 * associated with each port type.
181 * Additionally, the range of events associated with the port types changed
182 * between Rev0 and Rev1.
184 * The constants below define the range of valid codes for each port type for
185 * the different revisions and are used to validate the event to be monitored.
188 #define CCI400_PMU_EVENT_MASK 0xffUL
189 #define CCI400_PMU_EVENT_SOURCE_SHIFT 5
190 #define CCI400_PMU_EVENT_SOURCE_MASK 0x7
191 #define CCI400_PMU_EVENT_CODE_SHIFT 0
192 #define CCI400_PMU_EVENT_CODE_MASK 0x1f
193 #define CCI400_PMU_EVENT_SOURCE(event) \
194 ((event >> CCI400_PMU_EVENT_SOURCE_SHIFT) & \
195 CCI400_PMU_EVENT_SOURCE_MASK)
196 #define CCI400_PMU_EVENT_CODE(event) \
197 ((event >> CCI400_PMU_EVENT_CODE_SHIFT) & CCI400_PMU_EVENT_CODE_MASK)
199 #define CCI400_R0_SLAVE_PORT_MIN_EV 0x00
200 #define CCI400_R0_SLAVE_PORT_MAX_EV 0x13
201 #define CCI400_R0_MASTER_PORT_MIN_EV 0x14
202 #define CCI400_R0_MASTER_PORT_MAX_EV 0x1a
204 #define CCI400_R1_SLAVE_PORT_MIN_EV 0x00
205 #define CCI400_R1_SLAVE_PORT_MAX_EV 0x14
206 #define CCI400_R1_MASTER_PORT_MIN_EV 0x00
207 #define CCI400_R1_MASTER_PORT_MAX_EV 0x11
209 #define CCI400_CYCLE_EVENT_EXT_ATTR_ENTRY(_name, _config) \
210 CCI_EXT_ATTR_ENTRY(_name, cci400_pmu_cycle_event_show, \
211 (unsigned long)_config)
213 static ssize_t cci400_pmu_cycle_event_show(struct device *dev,
214 struct device_attribute *attr, char *buf);
216 static struct attribute *cci400_pmu_format_attrs[] = {
217 CCI_FORMAT_EXT_ATTR_ENTRY(event, "config:0-4"),
218 CCI_FORMAT_EXT_ATTR_ENTRY(source, "config:5-7"),
222 static struct attribute *cci400_r0_pmu_event_attrs[] = {
224 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_any, 0x0),
225 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_device, 0x01),
226 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_normal_or_nonshareable, 0x2),
227 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_inner_or_outershareable, 0x3),
228 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_cache_maintenance, 0x4),
229 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_mem_barrier, 0x5),
230 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_sync_barrier, 0x6),
231 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg, 0x7),
232 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg_sync, 0x8),
233 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_stall_tt_full, 0x9),
234 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_last_hs_snoop, 0xA),
235 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_stall_rvalids_h_rready_l, 0xB),
236 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_any, 0xC),
237 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_device, 0xD),
238 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_normal_or_nonshareable, 0xE),
239 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_inner_or_outershare_wback_wclean, 0xF),
240 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_write_unique, 0x10),
241 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_write_line_unique, 0x11),
242 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_evict, 0x12),
243 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_stall_tt_full, 0x13),
245 CCI_EVENT_EXT_ATTR_ENTRY(mi_retry_speculative_fetch, 0x14),
246 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_addr_hazard, 0x15),
247 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_id_hazard, 0x16),
248 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_tt_full, 0x17),
249 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_barrier_hazard, 0x18),
250 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_barrier_hazard, 0x19),
251 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_tt_full, 0x1A),
252 /* Special event for cycles counter */
253 CCI400_CYCLE_EVENT_EXT_ATTR_ENTRY(cycles, 0xff),
257 static struct attribute *cci400_r1_pmu_event_attrs[] = {
259 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_any, 0x0),
260 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_device, 0x01),
261 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_normal_or_nonshareable, 0x2),
262 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_inner_or_outershareable, 0x3),
263 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_cache_maintenance, 0x4),
264 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_mem_barrier, 0x5),
265 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_sync_barrier, 0x6),
266 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg, 0x7),
267 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg_sync, 0x8),
268 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_stall_tt_full, 0x9),
269 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_last_hs_snoop, 0xA),
270 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_stall_rvalids_h_rready_l, 0xB),
271 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_any, 0xC),
272 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_device, 0xD),
273 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_normal_or_nonshareable, 0xE),
274 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_inner_or_outershare_wback_wclean, 0xF),
275 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_write_unique, 0x10),
276 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_write_line_unique, 0x11),
277 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_evict, 0x12),
278 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_stall_tt_full, 0x13),
279 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_stall_slave_id_hazard, 0x14),
281 CCI_EVENT_EXT_ATTR_ENTRY(mi_retry_speculative_fetch, 0x0),
282 CCI_EVENT_EXT_ATTR_ENTRY(mi_stall_cycle_addr_hazard, 0x1),
283 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_master_id_hazard, 0x2),
284 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_hi_prio_rtq_full, 0x3),
285 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_barrier_hazard, 0x4),
286 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_barrier_hazard, 0x5),
287 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_wtq_full, 0x6),
288 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_low_prio_rtq_full, 0x7),
289 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_mid_prio_rtq_full, 0x8),
290 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_qvn_vn0, 0x9),
291 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_qvn_vn1, 0xA),
292 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_qvn_vn2, 0xB),
293 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_qvn_vn3, 0xC),
294 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_qvn_vn0, 0xD),
295 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_qvn_vn1, 0xE),
296 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_qvn_vn2, 0xF),
297 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_qvn_vn3, 0x10),
298 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_unique_or_line_unique_addr_hazard, 0x11),
299 /* Special event for cycles counter */
300 CCI400_CYCLE_EVENT_EXT_ATTR_ENTRY(cycles, 0xff),
304 static ssize_t cci400_pmu_cycle_event_show(struct device *dev,
305 struct device_attribute *attr, char *buf)
307 struct dev_ext_attribute *eattr = container_of(attr,
308 struct dev_ext_attribute, attr);
309 return snprintf(buf, PAGE_SIZE, "config=0x%lx\n", (unsigned long)eattr->var);
312 static int cci400_get_event_idx(struct cci_pmu *cci_pmu,
313 struct cci_pmu_hw_events *hw,
314 unsigned long cci_event)
318 /* cycles event idx is fixed */
319 if (cci_event == CCI400_PMU_CYCLES) {
320 if (test_and_set_bit(CCI400_PMU_CYCLE_CNTR_IDX, hw->used_mask))
323 return CCI400_PMU_CYCLE_CNTR_IDX;
326 for (idx = CCI400_PMU_CNTR0_IDX; idx <= CCI_PMU_CNTR_LAST(cci_pmu); ++idx)
327 if (!test_and_set_bit(idx, hw->used_mask))
330 /* No counters available */
334 static int cci400_validate_hw_event(struct cci_pmu *cci_pmu, unsigned long hw_event)
336 u8 ev_source = CCI400_PMU_EVENT_SOURCE(hw_event);
337 u8 ev_code = CCI400_PMU_EVENT_CODE(hw_event);
340 if (hw_event & ~CCI400_PMU_EVENT_MASK)
343 if (hw_event == CCI400_PMU_CYCLES)
352 /* Slave Interface */
353 if_type = CCI_IF_SLAVE;
358 /* Master Interface */
359 if_type = CCI_IF_MASTER;
365 if (ev_code >= cci_pmu->model->event_ranges[if_type].min &&
366 ev_code <= cci_pmu->model->event_ranges[if_type].max)
372 static int probe_cci400_revision(struct cci_pmu *cci_pmu)
375 rev = readl_relaxed(cci_pmu->ctrl_base + CCI_PID2) & CCI_PID2_REV_MASK;
376 rev >>= CCI_PID2_REV_SHIFT;
378 if (rev < CCI400_R1_PX)
384 static const struct cci_pmu_model *probe_cci_model(struct cci_pmu *cci_pmu)
386 if (platform_has_secure_cci_access())
387 return &cci_pmu_models[probe_cci400_revision(cci_pmu)];
390 #else /* !CONFIG_ARM_CCI400_PMU */
391 static inline struct cci_pmu_model *probe_cci_model(struct cci_pmu *cci_pmu)
395 #endif /* CONFIG_ARM_CCI400_PMU */
397 #ifdef CONFIG_ARM_CCI5xx_PMU
400 * CCI5xx PMU event id is an 9-bit value made of two parts.
401 * bits [8:5] - Source for the event
402 * bits [4:0] - Event code (specific to type of interface)
408 #define CCI5xx_PORT_S0 0x0
409 #define CCI5xx_PORT_S1 0x1
410 #define CCI5xx_PORT_S2 0x2
411 #define CCI5xx_PORT_S3 0x3
412 #define CCI5xx_PORT_S4 0x4
413 #define CCI5xx_PORT_S5 0x5
414 #define CCI5xx_PORT_S6 0x6
416 #define CCI5xx_PORT_M0 0x8
417 #define CCI5xx_PORT_M1 0x9
418 #define CCI5xx_PORT_M2 0xa
419 #define CCI5xx_PORT_M3 0xb
420 #define CCI5xx_PORT_M4 0xc
421 #define CCI5xx_PORT_M5 0xd
422 #define CCI5xx_PORT_M6 0xe
424 #define CCI5xx_PORT_GLOBAL 0xf
426 #define CCI5xx_PMU_EVENT_MASK 0x1ffUL
427 #define CCI5xx_PMU_EVENT_SOURCE_SHIFT 0x5
428 #define CCI5xx_PMU_EVENT_SOURCE_MASK 0xf
429 #define CCI5xx_PMU_EVENT_CODE_SHIFT 0x0
430 #define CCI5xx_PMU_EVENT_CODE_MASK 0x1f
432 #define CCI5xx_PMU_EVENT_SOURCE(event) \
433 ((event >> CCI5xx_PMU_EVENT_SOURCE_SHIFT) & CCI5xx_PMU_EVENT_SOURCE_MASK)
434 #define CCI5xx_PMU_EVENT_CODE(event) \
435 ((event >> CCI5xx_PMU_EVENT_CODE_SHIFT) & CCI5xx_PMU_EVENT_CODE_MASK)
437 #define CCI5xx_SLAVE_PORT_MIN_EV 0x00
438 #define CCI5xx_SLAVE_PORT_MAX_EV 0x1f
439 #define CCI5xx_MASTER_PORT_MIN_EV 0x00
440 #define CCI5xx_MASTER_PORT_MAX_EV 0x06
441 #define CCI5xx_GLOBAL_PORT_MIN_EV 0x00
442 #define CCI5xx_GLOBAL_PORT_MAX_EV 0x0f
445 #define CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(_name, _config) \
446 CCI_EXT_ATTR_ENTRY(_name, cci5xx_pmu_global_event_show, \
447 (unsigned long) _config)
449 static ssize_t cci5xx_pmu_global_event_show(struct device *dev,
450 struct device_attribute *attr, char *buf);
452 static struct attribute *cci5xx_pmu_format_attrs[] = {
453 CCI_FORMAT_EXT_ATTR_ENTRY(event, "config:0-4"),
454 CCI_FORMAT_EXT_ATTR_ENTRY(source, "config:5-8"),
458 static struct attribute *cci5xx_pmu_event_attrs[] = {
460 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_arvalid, 0x0),
461 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_dev, 0x1),
462 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_nonshareable, 0x2),
463 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_shareable_non_alloc, 0x3),
464 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_shareable_alloc, 0x4),
465 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_invalidate, 0x5),
466 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_cache_maint, 0x6),
467 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg, 0x7),
468 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_rval, 0x8),
469 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_rlast_snoop, 0x9),
470 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_awalid, 0xA),
471 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_dev, 0xB),
472 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_non_shareable, 0xC),
473 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_share_wb, 0xD),
474 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_share_wlu, 0xE),
475 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_share_wunique, 0xF),
476 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_evict, 0x10),
477 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_wrevict, 0x11),
478 CCI_EVENT_EXT_ATTR_ENTRY(si_w_data_beat, 0x12),
479 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_acvalid, 0x13),
480 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_read, 0x14),
481 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_clean, 0x15),
482 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_data_transfer_low, 0x16),
483 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_stall_arvalid, 0x17),
484 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_stall, 0x18),
485 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_stall, 0x19),
486 CCI_EVENT_EXT_ATTR_ENTRY(si_w_data_stall, 0x1A),
487 CCI_EVENT_EXT_ATTR_ENTRY(si_w_resp_stall, 0x1B),
488 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_stall, 0x1C),
489 CCI_EVENT_EXT_ATTR_ENTRY(si_s_data_stall, 0x1D),
490 CCI_EVENT_EXT_ATTR_ENTRY(si_rq_stall_ot_limit, 0x1E),
491 CCI_EVENT_EXT_ATTR_ENTRY(si_r_stall_arbit, 0x1F),
494 CCI_EVENT_EXT_ATTR_ENTRY(mi_r_data_beat_any, 0x0),
495 CCI_EVENT_EXT_ATTR_ENTRY(mi_w_data_beat_any, 0x1),
496 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall, 0x2),
497 CCI_EVENT_EXT_ATTR_ENTRY(mi_r_data_stall, 0x3),
498 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall, 0x4),
499 CCI_EVENT_EXT_ATTR_ENTRY(mi_w_data_stall, 0x5),
500 CCI_EVENT_EXT_ATTR_ENTRY(mi_w_resp_stall, 0x6),
503 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_filter_bank_0_1, 0x0),
504 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_filter_bank_2_3, 0x1),
505 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_filter_bank_4_5, 0x2),
506 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_filter_bank_6_7, 0x3),
507 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_miss_filter_bank_0_1, 0x4),
508 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_miss_filter_bank_2_3, 0x5),
509 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_miss_filter_bank_4_5, 0x6),
510 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_miss_filter_bank_6_7, 0x7),
511 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_back_invalidation, 0x8),
512 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_stall_alloc_busy, 0x9),
513 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_stall_tt_full, 0xA),
514 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_wrq, 0xB),
515 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_cd_hs, 0xC),
516 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_rq_stall_addr_hazard, 0xD),
517 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_rq_stall_tt_full, 0xE),
518 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_rq_tzmp1_prot, 0xF),
522 static ssize_t cci5xx_pmu_global_event_show(struct device *dev,
523 struct device_attribute *attr, char *buf)
525 struct dev_ext_attribute *eattr = container_of(attr,
526 struct dev_ext_attribute, attr);
527 /* Global events have single fixed source code */
528 return snprintf(buf, PAGE_SIZE, "event=0x%lx,source=0x%x\n",
529 (unsigned long)eattr->var, CCI5xx_PORT_GLOBAL);
533 * CCI500 provides 8 independent event counters that can count
534 * any of the events available.
535 * CCI500 PMU event source ids
536 * 0x0-0x6 - Slave interfaces
537 * 0x8-0xD - Master interfaces
538 * 0xf - Global Events
541 static int cci500_validate_hw_event(struct cci_pmu *cci_pmu,
542 unsigned long hw_event)
544 u32 ev_source = CCI5xx_PMU_EVENT_SOURCE(hw_event);
545 u32 ev_code = CCI5xx_PMU_EVENT_CODE(hw_event);
548 if (hw_event & ~CCI5xx_PMU_EVENT_MASK)
559 if_type = CCI_IF_SLAVE;
567 if_type = CCI_IF_MASTER;
569 case CCI5xx_PORT_GLOBAL:
570 if_type = CCI_IF_GLOBAL;
576 if (ev_code >= cci_pmu->model->event_ranges[if_type].min &&
577 ev_code <= cci_pmu->model->event_ranges[if_type].max)
584 * CCI550 provides 8 independent event counters that can count
585 * any of the events available.
586 * CCI550 PMU event source ids
587 * 0x0-0x6 - Slave interfaces
588 * 0x8-0xe - Master interfaces
589 * 0xf - Global Events
592 static int cci550_validate_hw_event(struct cci_pmu *cci_pmu,
593 unsigned long hw_event)
595 u32 ev_source = CCI5xx_PMU_EVENT_SOURCE(hw_event);
596 u32 ev_code = CCI5xx_PMU_EVENT_CODE(hw_event);
599 if (hw_event & ~CCI5xx_PMU_EVENT_MASK)
610 if_type = CCI_IF_SLAVE;
619 if_type = CCI_IF_MASTER;
621 case CCI5xx_PORT_GLOBAL:
622 if_type = CCI_IF_GLOBAL;
628 if (ev_code >= cci_pmu->model->event_ranges[if_type].min &&
629 ev_code <= cci_pmu->model->event_ranges[if_type].max)
635 #endif /* CONFIG_ARM_CCI5xx_PMU */
638 * Program the CCI PMU counters which have PERF_HES_ARCH set
639 * with the event period and mark them ready before we enable
642 static void cci_pmu_sync_counters(struct cci_pmu *cci_pmu)
645 struct cci_pmu_hw_events *cci_hw = &cci_pmu->hw_events;
646 DECLARE_BITMAP(mask, HW_CNTRS_MAX);
648 bitmap_zero(mask, cci_pmu->num_cntrs);
649 for_each_set_bit(i, cci_pmu->hw_events.used_mask, cci_pmu->num_cntrs) {
650 struct perf_event *event = cci_hw->events[i];
655 /* Leave the events which are not counting */
656 if (event->hw.state & PERF_HES_STOPPED)
658 if (event->hw.state & PERF_HES_ARCH) {
660 event->hw.state &= ~PERF_HES_ARCH;
664 pmu_write_counters(cci_pmu, mask);
667 /* Should be called with cci_pmu->hw_events->pmu_lock held */
668 static void __cci_pmu_enable_nosync(struct cci_pmu *cci_pmu)
672 /* Enable all the PMU counters. */
673 val = readl_relaxed(cci_pmu->ctrl_base + CCI_PMCR) | CCI_PMCR_CEN;
674 writel(val, cci_pmu->ctrl_base + CCI_PMCR);
677 /* Should be called with cci_pmu->hw_events->pmu_lock held */
678 static void __cci_pmu_enable_sync(struct cci_pmu *cci_pmu)
680 cci_pmu_sync_counters(cci_pmu);
681 __cci_pmu_enable_nosync(cci_pmu);
684 /* Should be called with cci_pmu->hw_events->pmu_lock held */
685 static void __cci_pmu_disable(struct cci_pmu *cci_pmu)
689 /* Disable all the PMU counters. */
690 val = readl_relaxed(cci_pmu->ctrl_base + CCI_PMCR) & ~CCI_PMCR_CEN;
691 writel(val, cci_pmu->ctrl_base + CCI_PMCR);
694 static ssize_t cci_pmu_format_show(struct device *dev,
695 struct device_attribute *attr, char *buf)
697 struct dev_ext_attribute *eattr = container_of(attr,
698 struct dev_ext_attribute, attr);
699 return snprintf(buf, PAGE_SIZE, "%s\n", (char *)eattr->var);
702 static ssize_t cci_pmu_event_show(struct device *dev,
703 struct device_attribute *attr, char *buf)
705 struct dev_ext_attribute *eattr = container_of(attr,
706 struct dev_ext_attribute, attr);
707 /* source parameter is mandatory for normal PMU events */
708 return snprintf(buf, PAGE_SIZE, "source=?,event=0x%lx\n",
709 (unsigned long)eattr->var);
712 static int pmu_is_valid_counter(struct cci_pmu *cci_pmu, int idx)
714 return 0 <= idx && idx <= CCI_PMU_CNTR_LAST(cci_pmu);
717 static u32 pmu_read_register(struct cci_pmu *cci_pmu, int idx, unsigned int offset)
719 return readl_relaxed(cci_pmu->base +
720 CCI_PMU_CNTR_BASE(cci_pmu->model, idx) + offset);
723 static void pmu_write_register(struct cci_pmu *cci_pmu, u32 value,
724 int idx, unsigned int offset)
726 writel_relaxed(value, cci_pmu->base +
727 CCI_PMU_CNTR_BASE(cci_pmu->model, idx) + offset);
730 static void pmu_disable_counter(struct cci_pmu *cci_pmu, int idx)
732 pmu_write_register(cci_pmu, 0, idx, CCI_PMU_CNTR_CTRL);
735 static void pmu_enable_counter(struct cci_pmu *cci_pmu, int idx)
737 pmu_write_register(cci_pmu, 1, idx, CCI_PMU_CNTR_CTRL);
740 static bool __maybe_unused
741 pmu_counter_is_enabled(struct cci_pmu *cci_pmu, int idx)
743 return (pmu_read_register(cci_pmu, idx, CCI_PMU_CNTR_CTRL) & 0x1) != 0;
746 static void pmu_set_event(struct cci_pmu *cci_pmu, int idx, unsigned long event)
748 pmu_write_register(cci_pmu, event, idx, CCI_PMU_EVT_SEL);
752 * For all counters on the CCI-PMU, disable any 'enabled' counters,
753 * saving the changed counters in the mask, so that we can restore
754 * it later using pmu_restore_counters. The mask is private to the
755 * caller. We cannot rely on the used_mask maintained by the CCI_PMU
756 * as it only tells us if the counter is assigned to perf_event or not.
757 * The state of the perf_event cannot be locked by the PMU layer, hence
758 * we check the individual counter status (which can be locked by
759 * cci_pm->hw_events->pmu_lock).
761 * @mask should be initialised to empty by the caller.
763 static void __maybe_unused
764 pmu_save_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
768 for (i = 0; i < cci_pmu->num_cntrs; i++) {
769 if (pmu_counter_is_enabled(cci_pmu, i)) {
771 pmu_disable_counter(cci_pmu, i);
777 * Restore the status of the counters. Reversal of the pmu_save_counters().
778 * For each counter set in the mask, enable the counter back.
780 static void __maybe_unused
781 pmu_restore_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
785 for_each_set_bit(i, mask, cci_pmu->num_cntrs)
786 pmu_enable_counter(cci_pmu, i);
790 * Returns the number of programmable counters actually implemented
793 static u32 pmu_get_max_counters(struct cci_pmu *cci_pmu)
795 return (readl_relaxed(cci_pmu->ctrl_base + CCI_PMCR) &
796 CCI_PMCR_NCNT_MASK) >> CCI_PMCR_NCNT_SHIFT;
799 static int pmu_get_event_idx(struct cci_pmu_hw_events *hw, struct perf_event *event)
801 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
802 unsigned long cci_event = event->hw.config_base;
805 if (cci_pmu->model->get_event_idx)
806 return cci_pmu->model->get_event_idx(cci_pmu, hw, cci_event);
808 /* Generic code to find an unused idx from the mask */
809 for(idx = 0; idx <= CCI_PMU_CNTR_LAST(cci_pmu); idx++)
810 if (!test_and_set_bit(idx, hw->used_mask))
813 /* No counters available */
817 static int pmu_map_event(struct perf_event *event)
819 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
821 if (event->attr.type < PERF_TYPE_MAX ||
822 !cci_pmu->model->validate_hw_event)
825 return cci_pmu->model->validate_hw_event(cci_pmu, event->attr.config);
828 static int pmu_request_irq(struct cci_pmu *cci_pmu, irq_handler_t handler)
831 struct platform_device *pmu_device = cci_pmu->plat_device;
833 if (unlikely(!pmu_device))
836 if (cci_pmu->nr_irqs < 1) {
837 dev_err(&pmu_device->dev, "no irqs for CCI PMUs defined\n");
842 * Register all available CCI PMU interrupts. In the interrupt handler
843 * we iterate over the counters checking for interrupt source (the
844 * overflowing counter) and clear it.
846 * This should allow handling of non-unique interrupt for the counters.
848 for (i = 0; i < cci_pmu->nr_irqs; i++) {
849 int err = request_irq(cci_pmu->irqs[i], handler, IRQF_SHARED,
850 "arm-cci-pmu", cci_pmu);
852 dev_err(&pmu_device->dev, "unable to request IRQ%d for ARM CCI PMU counters\n",
857 set_bit(i, &cci_pmu->active_irqs);
863 static void pmu_free_irq(struct cci_pmu *cci_pmu)
867 for (i = 0; i < cci_pmu->nr_irqs; i++) {
868 if (!test_and_clear_bit(i, &cci_pmu->active_irqs))
871 free_irq(cci_pmu->irqs[i], cci_pmu);
875 static u32 pmu_read_counter(struct perf_event *event)
877 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
878 struct hw_perf_event *hw_counter = &event->hw;
879 int idx = hw_counter->idx;
882 if (unlikely(!pmu_is_valid_counter(cci_pmu, idx))) {
883 dev_err(&cci_pmu->plat_device->dev, "Invalid CCI PMU counter %d\n", idx);
886 value = pmu_read_register(cci_pmu, idx, CCI_PMU_CNTR);
891 static void pmu_write_counter(struct cci_pmu *cci_pmu, u32 value, int idx)
893 pmu_write_register(cci_pmu, value, idx, CCI_PMU_CNTR);
896 static void __pmu_write_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
899 struct cci_pmu_hw_events *cci_hw = &cci_pmu->hw_events;
901 for_each_set_bit(i, mask, cci_pmu->num_cntrs) {
902 struct perf_event *event = cci_hw->events[i];
906 pmu_write_counter(cci_pmu, local64_read(&event->hw.prev_count), i);
910 static void pmu_write_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
912 if (cci_pmu->model->write_counters)
913 cci_pmu->model->write_counters(cci_pmu, mask);
915 __pmu_write_counters(cci_pmu, mask);
918 #ifdef CONFIG_ARM_CCI5xx_PMU
921 * CCI-500/CCI-550 has advanced power saving policies, which could gate the
922 * clocks to the PMU counters, which makes the writes to them ineffective.
923 * The only way to write to those counters is when the global counters
924 * are enabled and the particular counter is enabled.
926 * So we do the following :
928 * 1) Disable all the PMU counters, saving their current state
929 * 2) Enable the global PMU profiling, now that all counters are
932 * For each counter to be programmed, repeat steps 3-7:
934 * 3) Write an invalid event code to the event control register for the
935 counter, so that the counters are not modified.
936 * 4) Enable the counter control for the counter.
937 * 5) Set the counter value
938 * 6) Disable the counter
939 * 7) Restore the event in the target counter
941 * 8) Disable the global PMU.
942 * 9) Restore the status of the rest of the counters.
944 * We choose an event which for CCI-5xx is guaranteed not to count.
945 * We use the highest possible event code (0x1f) for the master interface 0.
947 #define CCI5xx_INVALID_EVENT ((CCI5xx_PORT_M0 << CCI5xx_PMU_EVENT_SOURCE_SHIFT) | \
948 (CCI5xx_PMU_EVENT_CODE_MASK << CCI5xx_PMU_EVENT_CODE_SHIFT))
949 static void cci5xx_pmu_write_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
952 DECLARE_BITMAP(saved_mask, HW_CNTRS_MAX);
954 bitmap_zero(saved_mask, cci_pmu->num_cntrs);
955 pmu_save_counters(cci_pmu, saved_mask);
958 * Now that all the counters are disabled, we can safely turn the PMU on,
959 * without syncing the status of the counters
961 __cci_pmu_enable_nosync(cci_pmu);
963 for_each_set_bit(i, mask, cci_pmu->num_cntrs) {
964 struct perf_event *event = cci_pmu->hw_events.events[i];
969 pmu_set_event(cci_pmu, i, CCI5xx_INVALID_EVENT);
970 pmu_enable_counter(cci_pmu, i);
971 pmu_write_counter(cci_pmu, local64_read(&event->hw.prev_count), i);
972 pmu_disable_counter(cci_pmu, i);
973 pmu_set_event(cci_pmu, i, event->hw.config_base);
976 __cci_pmu_disable(cci_pmu);
978 pmu_restore_counters(cci_pmu, saved_mask);
981 #endif /* CONFIG_ARM_CCI5xx_PMU */
983 static u64 pmu_event_update(struct perf_event *event)
985 struct hw_perf_event *hwc = &event->hw;
986 u64 delta, prev_raw_count, new_raw_count;
989 prev_raw_count = local64_read(&hwc->prev_count);
990 new_raw_count = pmu_read_counter(event);
991 } while (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
992 new_raw_count) != prev_raw_count);
994 delta = (new_raw_count - prev_raw_count) & CCI_PMU_CNTR_MASK;
996 local64_add(delta, &event->count);
998 return new_raw_count;
1001 static void pmu_read(struct perf_event *event)
1003 pmu_event_update(event);
1006 static void pmu_event_set_period(struct perf_event *event)
1008 struct hw_perf_event *hwc = &event->hw;
1010 * The CCI PMU counters have a period of 2^32. To account for the
1011 * possiblity of extreme interrupt latency we program for a period of
1012 * half that. Hopefully we can handle the interrupt before another 2^31
1013 * events occur and the counter overtakes its previous value.
1015 u64 val = 1ULL << 31;
1016 local64_set(&hwc->prev_count, val);
1019 * CCI PMU uses PERF_HES_ARCH to keep track of the counters, whose
1020 * values needs to be sync-ed with the s/w state before the PMU is
1022 * Mark this counter for sync.
1024 hwc->state |= PERF_HES_ARCH;
1027 static irqreturn_t pmu_handle_irq(int irq_num, void *dev)
1029 unsigned long flags;
1030 struct cci_pmu *cci_pmu = dev;
1031 struct cci_pmu_hw_events *events = &cci_pmu->hw_events;
1032 int idx, handled = IRQ_NONE;
1034 raw_spin_lock_irqsave(&events->pmu_lock, flags);
1036 /* Disable the PMU while we walk through the counters */
1037 __cci_pmu_disable(cci_pmu);
1039 * Iterate over counters and update the corresponding perf events.
1040 * This should work regardless of whether we have per-counter overflow
1041 * interrupt or a combined overflow interrupt.
1043 for (idx = 0; idx <= CCI_PMU_CNTR_LAST(cci_pmu); idx++) {
1044 struct perf_event *event = events->events[idx];
1049 /* Did this counter overflow? */
1050 if (!(pmu_read_register(cci_pmu, idx, CCI_PMU_OVRFLW) &
1051 CCI_PMU_OVRFLW_FLAG))
1054 pmu_write_register(cci_pmu, CCI_PMU_OVRFLW_FLAG, idx,
1057 pmu_event_update(event);
1058 pmu_event_set_period(event);
1059 handled = IRQ_HANDLED;
1062 /* Enable the PMU and sync possibly overflowed counters */
1063 __cci_pmu_enable_sync(cci_pmu);
1064 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
1066 return IRQ_RETVAL(handled);
1069 static int cci_pmu_get_hw(struct cci_pmu *cci_pmu)
1071 int ret = pmu_request_irq(cci_pmu, pmu_handle_irq);
1073 pmu_free_irq(cci_pmu);
1079 static void cci_pmu_put_hw(struct cci_pmu *cci_pmu)
1081 pmu_free_irq(cci_pmu);
1084 static void hw_perf_event_destroy(struct perf_event *event)
1086 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1087 atomic_t *active_events = &cci_pmu->active_events;
1088 struct mutex *reserve_mutex = &cci_pmu->reserve_mutex;
1090 if (atomic_dec_and_mutex_lock(active_events, reserve_mutex)) {
1091 cci_pmu_put_hw(cci_pmu);
1092 mutex_unlock(reserve_mutex);
1096 static void cci_pmu_enable(struct pmu *pmu)
1098 struct cci_pmu *cci_pmu = to_cci_pmu(pmu);
1099 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1100 int enabled = bitmap_weight(hw_events->used_mask, cci_pmu->num_cntrs);
1101 unsigned long flags;
1106 raw_spin_lock_irqsave(&hw_events->pmu_lock, flags);
1107 __cci_pmu_enable_sync(cci_pmu);
1108 raw_spin_unlock_irqrestore(&hw_events->pmu_lock, flags);
1112 static void cci_pmu_disable(struct pmu *pmu)
1114 struct cci_pmu *cci_pmu = to_cci_pmu(pmu);
1115 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1116 unsigned long flags;
1118 raw_spin_lock_irqsave(&hw_events->pmu_lock, flags);
1119 __cci_pmu_disable(cci_pmu);
1120 raw_spin_unlock_irqrestore(&hw_events->pmu_lock, flags);
1124 * Check if the idx represents a non-programmable counter.
1125 * All the fixed event counters are mapped before the programmable
1128 static bool pmu_fixed_hw_idx(struct cci_pmu *cci_pmu, int idx)
1130 return (idx >= 0) && (idx < cci_pmu->model->fixed_hw_cntrs);
1133 static void cci_pmu_start(struct perf_event *event, int pmu_flags)
1135 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1136 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1137 struct hw_perf_event *hwc = &event->hw;
1139 unsigned long flags;
1142 * To handle interrupt latency, we always reprogram the period
1143 * regardlesss of PERF_EF_RELOAD.
1145 if (pmu_flags & PERF_EF_RELOAD)
1146 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
1150 if (unlikely(!pmu_is_valid_counter(cci_pmu, idx))) {
1151 dev_err(&cci_pmu->plat_device->dev, "Invalid CCI PMU counter %d\n", idx);
1155 raw_spin_lock_irqsave(&hw_events->pmu_lock, flags);
1157 /* Configure the counter unless you are counting a fixed event */
1158 if (!pmu_fixed_hw_idx(cci_pmu, idx))
1159 pmu_set_event(cci_pmu, idx, hwc->config_base);
1161 pmu_event_set_period(event);
1162 pmu_enable_counter(cci_pmu, idx);
1164 raw_spin_unlock_irqrestore(&hw_events->pmu_lock, flags);
1167 static void cci_pmu_stop(struct perf_event *event, int pmu_flags)
1169 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1170 struct hw_perf_event *hwc = &event->hw;
1173 if (hwc->state & PERF_HES_STOPPED)
1176 if (unlikely(!pmu_is_valid_counter(cci_pmu, idx))) {
1177 dev_err(&cci_pmu->plat_device->dev, "Invalid CCI PMU counter %d\n", idx);
1182 * We always reprogram the counter, so ignore PERF_EF_UPDATE. See
1185 pmu_disable_counter(cci_pmu, idx);
1186 pmu_event_update(event);
1187 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
1190 static int cci_pmu_add(struct perf_event *event, int flags)
1192 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1193 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1194 struct hw_perf_event *hwc = &event->hw;
1197 /* If we don't have a space for the counter then finish early. */
1198 idx = pmu_get_event_idx(hw_events, event);
1202 event->hw.idx = idx;
1203 hw_events->events[idx] = event;
1205 hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
1206 if (flags & PERF_EF_START)
1207 cci_pmu_start(event, PERF_EF_RELOAD);
1209 /* Propagate our changes to the userspace mapping. */
1210 perf_event_update_userpage(event);
1215 static void cci_pmu_del(struct perf_event *event, int flags)
1217 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1218 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1219 struct hw_perf_event *hwc = &event->hw;
1222 cci_pmu_stop(event, PERF_EF_UPDATE);
1223 hw_events->events[idx] = NULL;
1224 clear_bit(idx, hw_events->used_mask);
1226 perf_event_update_userpage(event);
1229 static int validate_event(struct pmu *cci_pmu,
1230 struct cci_pmu_hw_events *hw_events,
1231 struct perf_event *event)
1233 if (is_software_event(event))
1237 * Reject groups spanning multiple HW PMUs (e.g. CPU + CCI). The
1238 * core perf code won't check that the pmu->ctx == leader->ctx
1239 * until after pmu->event_init(event).
1241 if (event->pmu != cci_pmu)
1244 if (event->state < PERF_EVENT_STATE_OFF)
1247 if (event->state == PERF_EVENT_STATE_OFF && !event->attr.enable_on_exec)
1250 return pmu_get_event_idx(hw_events, event) >= 0;
1253 static int validate_group(struct perf_event *event)
1255 struct perf_event *sibling, *leader = event->group_leader;
1256 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1257 unsigned long mask[BITS_TO_LONGS(HW_CNTRS_MAX)];
1258 struct cci_pmu_hw_events fake_pmu = {
1260 * Initialise the fake PMU. We only need to populate the
1261 * used_mask for the purposes of validation.
1265 memset(mask, 0, BITS_TO_LONGS(cci_pmu->num_cntrs) * sizeof(unsigned long));
1267 if (!validate_event(event->pmu, &fake_pmu, leader))
1270 for_each_sibling_event(sibling, leader) {
1271 if (!validate_event(event->pmu, &fake_pmu, sibling))
1275 if (!validate_event(event->pmu, &fake_pmu, event))
1281 static int __hw_perf_event_init(struct perf_event *event)
1283 struct hw_perf_event *hwc = &event->hw;
1286 mapping = pmu_map_event(event);
1289 pr_debug("event %x:%llx not supported\n", event->attr.type,
1290 event->attr.config);
1295 * We don't assign an index until we actually place the event onto
1296 * hardware. Use -1 to signify that we haven't decided where to put it
1300 hwc->config_base = 0;
1302 hwc->event_base = 0;
1305 * Store the event encoding into the config_base field.
1307 hwc->config_base |= (unsigned long)mapping;
1309 if (event->group_leader != event) {
1310 if (validate_group(event) != 0)
1317 static int cci_pmu_event_init(struct perf_event *event)
1319 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1320 atomic_t *active_events = &cci_pmu->active_events;
1323 if (event->attr.type != event->pmu->type)
1326 /* Shared by all CPUs, no meaningful state to sample */
1327 if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
1331 * Following the example set by other "uncore" PMUs, we accept any CPU
1332 * and rewrite its affinity dynamically rather than having perf core
1333 * handle cpu == -1 and pid == -1 for this case.
1335 * The perf core will pin online CPUs for the duration of this call and
1336 * the event being installed into its context, so the PMU's CPU can't
1337 * change under our feet.
1341 event->cpu = cci_pmu->cpu;
1343 event->destroy = hw_perf_event_destroy;
1344 if (!atomic_inc_not_zero(active_events)) {
1345 mutex_lock(&cci_pmu->reserve_mutex);
1346 if (atomic_read(active_events) == 0)
1347 err = cci_pmu_get_hw(cci_pmu);
1349 atomic_inc(active_events);
1350 mutex_unlock(&cci_pmu->reserve_mutex);
1355 err = __hw_perf_event_init(event);
1357 hw_perf_event_destroy(event);
1362 static ssize_t pmu_cpumask_attr_show(struct device *dev,
1363 struct device_attribute *attr, char *buf)
1365 struct pmu *pmu = dev_get_drvdata(dev);
1366 struct cci_pmu *cci_pmu = to_cci_pmu(pmu);
1368 return cpumap_print_to_pagebuf(true, buf, cpumask_of(cci_pmu->cpu));
1371 static struct device_attribute pmu_cpumask_attr =
1372 __ATTR(cpumask, S_IRUGO, pmu_cpumask_attr_show, NULL);
1374 static struct attribute *pmu_attrs[] = {
1375 &pmu_cpumask_attr.attr,
1379 static struct attribute_group pmu_attr_group = {
1383 static struct attribute_group pmu_format_attr_group = {
1385 .attrs = NULL, /* Filled in cci_pmu_init_attrs */
1388 static struct attribute_group pmu_event_attr_group = {
1390 .attrs = NULL, /* Filled in cci_pmu_init_attrs */
1393 static const struct attribute_group *pmu_attr_groups[] = {
1395 &pmu_format_attr_group,
1396 &pmu_event_attr_group,
1400 static int cci_pmu_init(struct cci_pmu *cci_pmu, struct platform_device *pdev)
1402 const struct cci_pmu_model *model = cci_pmu->model;
1403 char *name = model->name;
1406 if (WARN_ON(model->num_hw_cntrs > NUM_HW_CNTRS_MAX))
1408 if (WARN_ON(model->fixed_hw_cntrs > FIXED_HW_CNTRS_MAX))
1411 pmu_event_attr_group.attrs = model->event_attrs;
1412 pmu_format_attr_group.attrs = model->format_attrs;
1414 cci_pmu->pmu = (struct pmu) {
1415 .module = THIS_MODULE,
1416 .name = cci_pmu->model->name,
1417 .task_ctx_nr = perf_invalid_context,
1418 .pmu_enable = cci_pmu_enable,
1419 .pmu_disable = cci_pmu_disable,
1420 .event_init = cci_pmu_event_init,
1423 .start = cci_pmu_start,
1424 .stop = cci_pmu_stop,
1426 .attr_groups = pmu_attr_groups,
1427 .capabilities = PERF_PMU_CAP_NO_EXCLUDE,
1430 cci_pmu->plat_device = pdev;
1431 num_cntrs = pmu_get_max_counters(cci_pmu);
1432 if (num_cntrs > cci_pmu->model->num_hw_cntrs) {
1433 dev_warn(&pdev->dev,
1434 "PMU implements more counters(%d) than supported by"
1435 " the model(%d), truncated.",
1436 num_cntrs, cci_pmu->model->num_hw_cntrs);
1437 num_cntrs = cci_pmu->model->num_hw_cntrs;
1439 cci_pmu->num_cntrs = num_cntrs + cci_pmu->model->fixed_hw_cntrs;
1441 return perf_pmu_register(&cci_pmu->pmu, name, -1);
1444 static int cci_pmu_offline_cpu(unsigned int cpu)
1448 if (!g_cci_pmu || cpu != g_cci_pmu->cpu)
1451 target = cpumask_any_but(cpu_online_mask, cpu);
1452 if (target >= nr_cpu_ids)
1455 perf_pmu_migrate_context(&g_cci_pmu->pmu, cpu, target);
1456 g_cci_pmu->cpu = target;
1460 static __maybe_unused struct cci_pmu_model cci_pmu_models[] = {
1461 #ifdef CONFIG_ARM_CCI400_PMU
1464 .fixed_hw_cntrs = FIXED_HW_CNTRS_CII_4XX, /* Cycle counter */
1465 .num_hw_cntrs = NUM_HW_CNTRS_CII_4XX,
1467 .format_attrs = cci400_pmu_format_attrs,
1468 .event_attrs = cci400_r0_pmu_event_attrs,
1471 CCI400_R0_SLAVE_PORT_MIN_EV,
1472 CCI400_R0_SLAVE_PORT_MAX_EV,
1475 CCI400_R0_MASTER_PORT_MIN_EV,
1476 CCI400_R0_MASTER_PORT_MAX_EV,
1479 .validate_hw_event = cci400_validate_hw_event,
1480 .get_event_idx = cci400_get_event_idx,
1483 .name = "CCI_400_r1",
1484 .fixed_hw_cntrs = FIXED_HW_CNTRS_CII_4XX, /* Cycle counter */
1485 .num_hw_cntrs = NUM_HW_CNTRS_CII_4XX,
1487 .format_attrs = cci400_pmu_format_attrs,
1488 .event_attrs = cci400_r1_pmu_event_attrs,
1491 CCI400_R1_SLAVE_PORT_MIN_EV,
1492 CCI400_R1_SLAVE_PORT_MAX_EV,
1495 CCI400_R1_MASTER_PORT_MIN_EV,
1496 CCI400_R1_MASTER_PORT_MAX_EV,
1499 .validate_hw_event = cci400_validate_hw_event,
1500 .get_event_idx = cci400_get_event_idx,
1503 #ifdef CONFIG_ARM_CCI5xx_PMU
1506 .fixed_hw_cntrs = FIXED_HW_CNTRS_CII_5XX,
1507 .num_hw_cntrs = NUM_HW_CNTRS_CII_5XX,
1508 .cntr_size = SZ_64K,
1509 .format_attrs = cci5xx_pmu_format_attrs,
1510 .event_attrs = cci5xx_pmu_event_attrs,
1513 CCI5xx_SLAVE_PORT_MIN_EV,
1514 CCI5xx_SLAVE_PORT_MAX_EV,
1517 CCI5xx_MASTER_PORT_MIN_EV,
1518 CCI5xx_MASTER_PORT_MAX_EV,
1521 CCI5xx_GLOBAL_PORT_MIN_EV,
1522 CCI5xx_GLOBAL_PORT_MAX_EV,
1525 .validate_hw_event = cci500_validate_hw_event,
1526 .write_counters = cci5xx_pmu_write_counters,
1530 .fixed_hw_cntrs = FIXED_HW_CNTRS_CII_5XX,
1531 .num_hw_cntrs = NUM_HW_CNTRS_CII_5XX,
1532 .cntr_size = SZ_64K,
1533 .format_attrs = cci5xx_pmu_format_attrs,
1534 .event_attrs = cci5xx_pmu_event_attrs,
1537 CCI5xx_SLAVE_PORT_MIN_EV,
1538 CCI5xx_SLAVE_PORT_MAX_EV,
1541 CCI5xx_MASTER_PORT_MIN_EV,
1542 CCI5xx_MASTER_PORT_MAX_EV,
1545 CCI5xx_GLOBAL_PORT_MIN_EV,
1546 CCI5xx_GLOBAL_PORT_MAX_EV,
1549 .validate_hw_event = cci550_validate_hw_event,
1550 .write_counters = cci5xx_pmu_write_counters,
1555 static const struct of_device_id arm_cci_pmu_matches[] = {
1556 #ifdef CONFIG_ARM_CCI400_PMU
1558 .compatible = "arm,cci-400-pmu",
1562 .compatible = "arm,cci-400-pmu,r0",
1563 .data = &cci_pmu_models[CCI400_R0],
1566 .compatible = "arm,cci-400-pmu,r1",
1567 .data = &cci_pmu_models[CCI400_R1],
1570 #ifdef CONFIG_ARM_CCI5xx_PMU
1572 .compatible = "arm,cci-500-pmu,r0",
1573 .data = &cci_pmu_models[CCI500_R0],
1576 .compatible = "arm,cci-550-pmu,r0",
1577 .data = &cci_pmu_models[CCI550_R0],
1582 MODULE_DEVICE_TABLE(of, arm_cci_pmu_matches);
1584 static bool is_duplicate_irq(int irq, int *irqs, int nr_irqs)
1588 for (i = 0; i < nr_irqs; i++)
1595 static struct cci_pmu *cci_pmu_alloc(struct device *dev)
1597 struct cci_pmu *cci_pmu;
1598 const struct cci_pmu_model *model;
1601 * All allocations are devm_* hence we don't have to free
1602 * them explicitly on an error, as it would end up in driver
1605 cci_pmu = devm_kzalloc(dev, sizeof(*cci_pmu), GFP_KERNEL);
1607 return ERR_PTR(-ENOMEM);
1609 cci_pmu->ctrl_base = *(void __iomem **)dev->platform_data;
1611 model = of_device_get_match_data(dev);
1614 "DEPRECATED compatible property, requires secure access to CCI registers");
1615 model = probe_cci_model(cci_pmu);
1618 dev_warn(dev, "CCI PMU version not supported\n");
1619 return ERR_PTR(-ENODEV);
1622 cci_pmu->model = model;
1623 cci_pmu->irqs = devm_kcalloc(dev, CCI_PMU_MAX_HW_CNTRS(model),
1624 sizeof(*cci_pmu->irqs), GFP_KERNEL);
1626 return ERR_PTR(-ENOMEM);
1627 cci_pmu->hw_events.events = devm_kcalloc(dev,
1628 CCI_PMU_MAX_HW_CNTRS(model),
1629 sizeof(*cci_pmu->hw_events.events),
1631 if (!cci_pmu->hw_events.events)
1632 return ERR_PTR(-ENOMEM);
1633 cci_pmu->hw_events.used_mask = devm_kcalloc(dev,
1634 BITS_TO_LONGS(CCI_PMU_MAX_HW_CNTRS(model)),
1635 sizeof(*cci_pmu->hw_events.used_mask),
1637 if (!cci_pmu->hw_events.used_mask)
1638 return ERR_PTR(-ENOMEM);
1643 static int cci_pmu_probe(struct platform_device *pdev)
1645 struct cci_pmu *cci_pmu;
1648 cci_pmu = cci_pmu_alloc(&pdev->dev);
1649 if (IS_ERR(cci_pmu))
1650 return PTR_ERR(cci_pmu);
1652 cci_pmu->base = devm_platform_ioremap_resource(pdev, 0);
1653 if (IS_ERR(cci_pmu->base))
1657 * CCI PMU has one overflow interrupt per counter; but some may be tied
1658 * together to a common interrupt.
1660 cci_pmu->nr_irqs = 0;
1661 for (i = 0; i < CCI_PMU_MAX_HW_CNTRS(cci_pmu->model); i++) {
1662 irq = platform_get_irq(pdev, i);
1666 if (is_duplicate_irq(irq, cci_pmu->irqs, cci_pmu->nr_irqs))
1669 cci_pmu->irqs[cci_pmu->nr_irqs++] = irq;
1673 * Ensure that the device tree has as many interrupts as the number
1676 if (i < CCI_PMU_MAX_HW_CNTRS(cci_pmu->model)) {
1677 dev_warn(&pdev->dev, "In-correct number of interrupts: %d, should be %d\n",
1678 i, CCI_PMU_MAX_HW_CNTRS(cci_pmu->model));
1682 raw_spin_lock_init(&cci_pmu->hw_events.pmu_lock);
1683 mutex_init(&cci_pmu->reserve_mutex);
1684 atomic_set(&cci_pmu->active_events, 0);
1686 cci_pmu->cpu = raw_smp_processor_id();
1687 g_cci_pmu = cci_pmu;
1688 cpuhp_setup_state_nocalls(CPUHP_AP_PERF_ARM_CCI_ONLINE,
1689 "perf/arm/cci:online", NULL,
1690 cci_pmu_offline_cpu);
1692 ret = cci_pmu_init(cci_pmu, pdev);
1694 goto error_pmu_init;
1696 pr_info("ARM %s PMU driver probed", cci_pmu->model->name);
1700 cpuhp_remove_state(CPUHP_AP_PERF_ARM_CCI_ONLINE);
1705 static int cci_pmu_remove(struct platform_device *pdev)
1710 cpuhp_remove_state(CPUHP_AP_PERF_ARM_CCI_ONLINE);
1711 perf_pmu_unregister(&g_cci_pmu->pmu);
1717 static struct platform_driver cci_pmu_driver = {
1719 .name = DRIVER_NAME,
1720 .of_match_table = arm_cci_pmu_matches,
1721 .suppress_bind_attrs = true,
1723 .probe = cci_pmu_probe,
1724 .remove = cci_pmu_remove,
1727 module_platform_driver(cci_pmu_driver);
1728 MODULE_LICENSE("GPL v2");
1729 MODULE_DESCRIPTION("ARM CCI PMU support");