Loading...
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_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>
18
19#define DRIVER_NAME "ARM-CCI PMU"
20
21#define CCI_PMCR 0x0100
22#define CCI_PID2 0x0fe8
23
24#define CCI_PMCR_CEN 0x00000001
25#define CCI_PMCR_NCNT_MASK 0x0000f800
26#define CCI_PMCR_NCNT_SHIFT 11
27
28#define CCI_PID2_REV_MASK 0xf0
29#define CCI_PID2_REV_SHIFT 4
30
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
35
36#define CCI_PMU_OVRFLW_FLAG 1
37
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)
42
43#define CCI_PMU_MAX_HW_CNTRS(model) \
44 ((model)->num_hw_cntrs + (model)->fixed_hw_cntrs)
45
46/* Types of interfaces that can generate events */
47enum {
48 CCI_IF_SLAVE,
49 CCI_IF_MASTER,
50#ifdef CONFIG_ARM_CCI5xx_PMU
51 CCI_IF_GLOBAL,
52#endif
53 CCI_IF_MAX,
54};
55
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
59
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
63
64#define HW_CNTRS_MAX (NUM_HW_CNTRS_MAX + FIXED_HW_CNTRS_MAX)
65
66struct event_range {
67 u32 min;
68 u32 max;
69};
70
71struct cci_pmu_hw_events {
72 struct perf_event **events;
73 unsigned long *used_mask;
74 raw_spinlock_t pmu_lock;
75};
76
77struct cci_pmu;
78/*
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
83 */
84struct cci_pmu_model {
85 char *name;
86 u32 fixed_hw_cntrs;
87 u32 num_hw_cntrs;
88 u32 cntr_size;
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 *);
95};
96
97static struct cci_pmu_model cci_pmu_models[];
98
99struct cci_pmu {
100 void __iomem *base;
101 void __iomem *ctrl_base;
102 struct pmu pmu;
103 int cpu;
104 int nr_irqs;
105 int *irqs;
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;
110 int num_cntrs;
111 atomic_t active_events;
112 struct mutex reserve_mutex;
113};
114
115#define to_cci_pmu(c) (container_of(c, struct cci_pmu, pmu))
116
117static struct cci_pmu *g_cci_pmu;
118
119enum cci_models {
120#ifdef CONFIG_ARM_CCI400_PMU
121 CCI400_R0,
122 CCI400_R1,
123#endif
124#ifdef CONFIG_ARM_CCI5xx_PMU
125 CCI500_R0,
126 CCI550_R0,
127#endif
128 CCI_MODEL_MAX
129};
130
131static void pmu_write_counters(struct cci_pmu *cci_pmu,
132 unsigned long *mask);
133static ssize_t __maybe_unused cci_pmu_format_show(struct device *dev,
134 struct device_attribute *attr, char *buf);
135static ssize_t __maybe_unused cci_pmu_event_show(struct device *dev,
136 struct device_attribute *attr, char *buf);
137
138#define CCI_EXT_ATTR_ENTRY(_name, _func, _config) \
139 &((struct dev_ext_attribute[]) { \
140 { __ATTR(_name, S_IRUGO, _func, NULL), (void *)_config } \
141 })[0].attr.attr
142
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)
147
148/* CCI400 PMU Specific definitions */
149
150#ifdef CONFIG_ARM_CCI400_PMU
151
152/* Port ids */
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
161
162#define CCI400_R1_PX 5
163
164/*
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.
168 */
169enum cci400_perf_events {
170 CCI400_PMU_CYCLES = 0xff
171};
172
173#define CCI400_PMU_CYCLE_CNTR_IDX 0
174#define CCI400_PMU_CNTR0_IDX 1
175
176/*
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.
180 *
181 * Additionally, the range of events associated with the port types changed
182 * between Rev0 and Rev1.
183 *
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.
186 */
187
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)
198
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
203
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
208
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)
212
213static ssize_t cci400_pmu_cycle_event_show(struct device *dev,
214 struct device_attribute *attr, char *buf);
215
216static 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"),
219 NULL
220};
221
222static struct attribute *cci400_r0_pmu_event_attrs[] = {
223 /* Slave events */
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),
244 /* Master events */
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),
254 NULL
255};
256
257static struct attribute *cci400_r1_pmu_event_attrs[] = {
258 /* Slave events */
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),
280 /* Master events */
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),
301 NULL
302};
303
304static ssize_t cci400_pmu_cycle_event_show(struct device *dev,
305 struct device_attribute *attr, char *buf)
306{
307 struct dev_ext_attribute *eattr = container_of(attr,
308 struct dev_ext_attribute, attr);
309 return sysfs_emit(buf, "config=0x%lx\n", (unsigned long)eattr->var);
310}
311
312static int cci400_get_event_idx(struct cci_pmu *cci_pmu,
313 struct cci_pmu_hw_events *hw,
314 unsigned long cci_event)
315{
316 int idx;
317
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))
321 return -EAGAIN;
322
323 return CCI400_PMU_CYCLE_CNTR_IDX;
324 }
325
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))
328 return idx;
329
330 /* No counters available */
331 return -EAGAIN;
332}
333
334static int cci400_validate_hw_event(struct cci_pmu *cci_pmu, unsigned long hw_event)
335{
336 u8 ev_source = CCI400_PMU_EVENT_SOURCE(hw_event);
337 u8 ev_code = CCI400_PMU_EVENT_CODE(hw_event);
338 int if_type;
339
340 if (hw_event & ~CCI400_PMU_EVENT_MASK)
341 return -ENOENT;
342
343 if (hw_event == CCI400_PMU_CYCLES)
344 return hw_event;
345
346 switch (ev_source) {
347 case CCI400_PORT_S0:
348 case CCI400_PORT_S1:
349 case CCI400_PORT_S2:
350 case CCI400_PORT_S3:
351 case CCI400_PORT_S4:
352 /* Slave Interface */
353 if_type = CCI_IF_SLAVE;
354 break;
355 case CCI400_PORT_M0:
356 case CCI400_PORT_M1:
357 case CCI400_PORT_M2:
358 /* Master Interface */
359 if_type = CCI_IF_MASTER;
360 break;
361 default:
362 return -ENOENT;
363 }
364
365 if (ev_code >= cci_pmu->model->event_ranges[if_type].min &&
366 ev_code <= cci_pmu->model->event_ranges[if_type].max)
367 return hw_event;
368
369 return -ENOENT;
370}
371
372static int probe_cci400_revision(struct cci_pmu *cci_pmu)
373{
374 int rev;
375 rev = readl_relaxed(cci_pmu->ctrl_base + CCI_PID2) & CCI_PID2_REV_MASK;
376 rev >>= CCI_PID2_REV_SHIFT;
377
378 if (rev < CCI400_R1_PX)
379 return CCI400_R0;
380 else
381 return CCI400_R1;
382}
383
384static const struct cci_pmu_model *probe_cci_model(struct cci_pmu *cci_pmu)
385{
386 if (platform_has_secure_cci_access())
387 return &cci_pmu_models[probe_cci400_revision(cci_pmu)];
388 return NULL;
389}
390#else /* !CONFIG_ARM_CCI400_PMU */
391static inline struct cci_pmu_model *probe_cci_model(struct cci_pmu *cci_pmu)
392{
393 return NULL;
394}
395#endif /* CONFIG_ARM_CCI400_PMU */
396
397#ifdef CONFIG_ARM_CCI5xx_PMU
398
399/*
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)
403 *
404 *
405 */
406
407/* Port ids */
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
415
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
423
424#define CCI5xx_PORT_GLOBAL 0xf
425
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
431
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)
436
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
443
444
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)
448
449static ssize_t cci5xx_pmu_global_event_show(struct device *dev,
450 struct device_attribute *attr, char *buf);
451
452static 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"),
455 NULL,
456};
457
458static struct attribute *cci5xx_pmu_event_attrs[] = {
459 /* Slave events */
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),
492
493 /* Master events */
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),
501
502 /* Global events */
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),
519 NULL
520};
521
522static ssize_t cci5xx_pmu_global_event_show(struct device *dev,
523 struct device_attribute *attr, char *buf)
524{
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 sysfs_emit(buf, "event=0x%lx,source=0x%x\n",
529 (unsigned long)eattr->var, CCI5xx_PORT_GLOBAL);
530}
531
532/*
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
539 * 0x7,0xe - Reserved
540 */
541static int cci500_validate_hw_event(struct cci_pmu *cci_pmu,
542 unsigned long hw_event)
543{
544 u32 ev_source = CCI5xx_PMU_EVENT_SOURCE(hw_event);
545 u32 ev_code = CCI5xx_PMU_EVENT_CODE(hw_event);
546 int if_type;
547
548 if (hw_event & ~CCI5xx_PMU_EVENT_MASK)
549 return -ENOENT;
550
551 switch (ev_source) {
552 case CCI5xx_PORT_S0:
553 case CCI5xx_PORT_S1:
554 case CCI5xx_PORT_S2:
555 case CCI5xx_PORT_S3:
556 case CCI5xx_PORT_S4:
557 case CCI5xx_PORT_S5:
558 case CCI5xx_PORT_S6:
559 if_type = CCI_IF_SLAVE;
560 break;
561 case CCI5xx_PORT_M0:
562 case CCI5xx_PORT_M1:
563 case CCI5xx_PORT_M2:
564 case CCI5xx_PORT_M3:
565 case CCI5xx_PORT_M4:
566 case CCI5xx_PORT_M5:
567 if_type = CCI_IF_MASTER;
568 break;
569 case CCI5xx_PORT_GLOBAL:
570 if_type = CCI_IF_GLOBAL;
571 break;
572 default:
573 return -ENOENT;
574 }
575
576 if (ev_code >= cci_pmu->model->event_ranges[if_type].min &&
577 ev_code <= cci_pmu->model->event_ranges[if_type].max)
578 return hw_event;
579
580 return -ENOENT;
581}
582
583/*
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
590 * 0x7 - Reserved
591 */
592static int cci550_validate_hw_event(struct cci_pmu *cci_pmu,
593 unsigned long hw_event)
594{
595 u32 ev_source = CCI5xx_PMU_EVENT_SOURCE(hw_event);
596 u32 ev_code = CCI5xx_PMU_EVENT_CODE(hw_event);
597 int if_type;
598
599 if (hw_event & ~CCI5xx_PMU_EVENT_MASK)
600 return -ENOENT;
601
602 switch (ev_source) {
603 case CCI5xx_PORT_S0:
604 case CCI5xx_PORT_S1:
605 case CCI5xx_PORT_S2:
606 case CCI5xx_PORT_S3:
607 case CCI5xx_PORT_S4:
608 case CCI5xx_PORT_S5:
609 case CCI5xx_PORT_S6:
610 if_type = CCI_IF_SLAVE;
611 break;
612 case CCI5xx_PORT_M0:
613 case CCI5xx_PORT_M1:
614 case CCI5xx_PORT_M2:
615 case CCI5xx_PORT_M3:
616 case CCI5xx_PORT_M4:
617 case CCI5xx_PORT_M5:
618 case CCI5xx_PORT_M6:
619 if_type = CCI_IF_MASTER;
620 break;
621 case CCI5xx_PORT_GLOBAL:
622 if_type = CCI_IF_GLOBAL;
623 break;
624 default:
625 return -ENOENT;
626 }
627
628 if (ev_code >= cci_pmu->model->event_ranges[if_type].min &&
629 ev_code <= cci_pmu->model->event_ranges[if_type].max)
630 return hw_event;
631
632 return -ENOENT;
633}
634
635#endif /* CONFIG_ARM_CCI5xx_PMU */
636
637/*
638 * Program the CCI PMU counters which have PERF_HES_ARCH set
639 * with the event period and mark them ready before we enable
640 * PMU.
641 */
642static void cci_pmu_sync_counters(struct cci_pmu *cci_pmu)
643{
644 int i;
645 struct cci_pmu_hw_events *cci_hw = &cci_pmu->hw_events;
646 DECLARE_BITMAP(mask, HW_CNTRS_MAX);
647
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];
651
652 if (WARN_ON(!event))
653 continue;
654
655 /* Leave the events which are not counting */
656 if (event->hw.state & PERF_HES_STOPPED)
657 continue;
658 if (event->hw.state & PERF_HES_ARCH) {
659 set_bit(i, mask);
660 event->hw.state &= ~PERF_HES_ARCH;
661 }
662 }
663
664 pmu_write_counters(cci_pmu, mask);
665}
666
667/* Should be called with cci_pmu->hw_events->pmu_lock held */
668static void __cci_pmu_enable_nosync(struct cci_pmu *cci_pmu)
669{
670 u32 val;
671
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);
675}
676
677/* Should be called with cci_pmu->hw_events->pmu_lock held */
678static void __cci_pmu_enable_sync(struct cci_pmu *cci_pmu)
679{
680 cci_pmu_sync_counters(cci_pmu);
681 __cci_pmu_enable_nosync(cci_pmu);
682}
683
684/* Should be called with cci_pmu->hw_events->pmu_lock held */
685static void __cci_pmu_disable(struct cci_pmu *cci_pmu)
686{
687 u32 val;
688
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);
692}
693
694static ssize_t cci_pmu_format_show(struct device *dev,
695 struct device_attribute *attr, char *buf)
696{
697 struct dev_ext_attribute *eattr = container_of(attr,
698 struct dev_ext_attribute, attr);
699 return sysfs_emit(buf, "%s\n", (char *)eattr->var);
700}
701
702static ssize_t cci_pmu_event_show(struct device *dev,
703 struct device_attribute *attr, char *buf)
704{
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 sysfs_emit(buf, "source=?,event=0x%lx\n",
709 (unsigned long)eattr->var);
710}
711
712static int pmu_is_valid_counter(struct cci_pmu *cci_pmu, int idx)
713{
714 return 0 <= idx && idx <= CCI_PMU_CNTR_LAST(cci_pmu);
715}
716
717static u32 pmu_read_register(struct cci_pmu *cci_pmu, int idx, unsigned int offset)
718{
719 return readl_relaxed(cci_pmu->base +
720 CCI_PMU_CNTR_BASE(cci_pmu->model, idx) + offset);
721}
722
723static void pmu_write_register(struct cci_pmu *cci_pmu, u32 value,
724 int idx, unsigned int offset)
725{
726 writel_relaxed(value, cci_pmu->base +
727 CCI_PMU_CNTR_BASE(cci_pmu->model, idx) + offset);
728}
729
730static void pmu_disable_counter(struct cci_pmu *cci_pmu, int idx)
731{
732 pmu_write_register(cci_pmu, 0, idx, CCI_PMU_CNTR_CTRL);
733}
734
735static void pmu_enable_counter(struct cci_pmu *cci_pmu, int idx)
736{
737 pmu_write_register(cci_pmu, 1, idx, CCI_PMU_CNTR_CTRL);
738}
739
740static bool __maybe_unused
741pmu_counter_is_enabled(struct cci_pmu *cci_pmu, int idx)
742{
743 return (pmu_read_register(cci_pmu, idx, CCI_PMU_CNTR_CTRL) & 0x1) != 0;
744}
745
746static void pmu_set_event(struct cci_pmu *cci_pmu, int idx, unsigned long event)
747{
748 pmu_write_register(cci_pmu, event, idx, CCI_PMU_EVT_SEL);
749}
750
751/*
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).
760 *
761 * @mask should be initialised to empty by the caller.
762 */
763static void __maybe_unused
764pmu_save_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
765{
766 int i;
767
768 for (i = 0; i < cci_pmu->num_cntrs; i++) {
769 if (pmu_counter_is_enabled(cci_pmu, i)) {
770 set_bit(i, mask);
771 pmu_disable_counter(cci_pmu, i);
772 }
773 }
774}
775
776/*
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.
779 */
780static void __maybe_unused
781pmu_restore_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
782{
783 int i;
784
785 for_each_set_bit(i, mask, cci_pmu->num_cntrs)
786 pmu_enable_counter(cci_pmu, i);
787}
788
789/*
790 * Returns the number of programmable counters actually implemented
791 * by the cci
792 */
793static u32 pmu_get_max_counters(struct cci_pmu *cci_pmu)
794{
795 return (readl_relaxed(cci_pmu->ctrl_base + CCI_PMCR) &
796 CCI_PMCR_NCNT_MASK) >> CCI_PMCR_NCNT_SHIFT;
797}
798
799static int pmu_get_event_idx(struct cci_pmu_hw_events *hw, struct perf_event *event)
800{
801 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
802 unsigned long cci_event = event->hw.config_base;
803 int idx;
804
805 if (cci_pmu->model->get_event_idx)
806 return cci_pmu->model->get_event_idx(cci_pmu, hw, cci_event);
807
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))
811 return idx;
812
813 /* No counters available */
814 return -EAGAIN;
815}
816
817static int pmu_map_event(struct perf_event *event)
818{
819 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
820
821 if (event->attr.type < PERF_TYPE_MAX ||
822 !cci_pmu->model->validate_hw_event)
823 return -ENOENT;
824
825 return cci_pmu->model->validate_hw_event(cci_pmu, event->attr.config);
826}
827
828static int pmu_request_irq(struct cci_pmu *cci_pmu, irq_handler_t handler)
829{
830 int i;
831 struct platform_device *pmu_device = cci_pmu->plat_device;
832
833 if (unlikely(!pmu_device))
834 return -ENODEV;
835
836 if (cci_pmu->nr_irqs < 1) {
837 dev_err(&pmu_device->dev, "no irqs for CCI PMUs defined\n");
838 return -ENODEV;
839 }
840
841 /*
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.
845 *
846 * This should allow handling of non-unique interrupt for the counters.
847 */
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);
851 if (err) {
852 dev_err(&pmu_device->dev, "unable to request IRQ%d for ARM CCI PMU counters\n",
853 cci_pmu->irqs[i]);
854 return err;
855 }
856
857 set_bit(i, &cci_pmu->active_irqs);
858 }
859
860 return 0;
861}
862
863static void pmu_free_irq(struct cci_pmu *cci_pmu)
864{
865 int i;
866
867 for (i = 0; i < cci_pmu->nr_irqs; i++) {
868 if (!test_and_clear_bit(i, &cci_pmu->active_irqs))
869 continue;
870
871 free_irq(cci_pmu->irqs[i], cci_pmu);
872 }
873}
874
875static u32 pmu_read_counter(struct perf_event *event)
876{
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;
880 u32 value;
881
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);
884 return 0;
885 }
886 value = pmu_read_register(cci_pmu, idx, CCI_PMU_CNTR);
887
888 return value;
889}
890
891static void pmu_write_counter(struct cci_pmu *cci_pmu, u32 value, int idx)
892{
893 pmu_write_register(cci_pmu, value, idx, CCI_PMU_CNTR);
894}
895
896static void __pmu_write_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
897{
898 int i;
899 struct cci_pmu_hw_events *cci_hw = &cci_pmu->hw_events;
900
901 for_each_set_bit(i, mask, cci_pmu->num_cntrs) {
902 struct perf_event *event = cci_hw->events[i];
903
904 if (WARN_ON(!event))
905 continue;
906 pmu_write_counter(cci_pmu, local64_read(&event->hw.prev_count), i);
907 }
908}
909
910static void pmu_write_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
911{
912 if (cci_pmu->model->write_counters)
913 cci_pmu->model->write_counters(cci_pmu, mask);
914 else
915 __pmu_write_counters(cci_pmu, mask);
916}
917
918#ifdef CONFIG_ARM_CCI5xx_PMU
919
920/*
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.
925 *
926 * So we do the following :
927 *
928 * 1) Disable all the PMU counters, saving their current state
929 * 2) Enable the global PMU profiling, now that all counters are
930 * disabled.
931 *
932 * For each counter to be programmed, repeat steps 3-7:
933 *
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
940 *
941 * 8) Disable the global PMU.
942 * 9) Restore the status of the rest of the counters.
943 *
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.
946 */
947#define CCI5xx_INVALID_EVENT ((CCI5xx_PORT_M0 << CCI5xx_PMU_EVENT_SOURCE_SHIFT) | \
948 (CCI5xx_PMU_EVENT_CODE_MASK << CCI5xx_PMU_EVENT_CODE_SHIFT))
949static void cci5xx_pmu_write_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
950{
951 int i;
952 DECLARE_BITMAP(saved_mask, HW_CNTRS_MAX);
953
954 bitmap_zero(saved_mask, cci_pmu->num_cntrs);
955 pmu_save_counters(cci_pmu, saved_mask);
956
957 /*
958 * Now that all the counters are disabled, we can safely turn the PMU on,
959 * without syncing the status of the counters
960 */
961 __cci_pmu_enable_nosync(cci_pmu);
962
963 for_each_set_bit(i, mask, cci_pmu->num_cntrs) {
964 struct perf_event *event = cci_pmu->hw_events.events[i];
965
966 if (WARN_ON(!event))
967 continue;
968
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);
974 }
975
976 __cci_pmu_disable(cci_pmu);
977
978 pmu_restore_counters(cci_pmu, saved_mask);
979}
980
981#endif /* CONFIG_ARM_CCI5xx_PMU */
982
983static u64 pmu_event_update(struct perf_event *event)
984{
985 struct hw_perf_event *hwc = &event->hw;
986 u64 delta, prev_raw_count, new_raw_count;
987
988 do {
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);
993
994 delta = (new_raw_count - prev_raw_count) & CCI_PMU_CNTR_MASK;
995
996 local64_add(delta, &event->count);
997
998 return new_raw_count;
999}
1000
1001static void pmu_read(struct perf_event *event)
1002{
1003 pmu_event_update(event);
1004}
1005
1006static void pmu_event_set_period(struct perf_event *event)
1007{
1008 struct hw_perf_event *hwc = &event->hw;
1009 /*
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.
1014 */
1015 u64 val = 1ULL << 31;
1016 local64_set(&hwc->prev_count, val);
1017
1018 /*
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
1021 * enabled.
1022 * Mark this counter for sync.
1023 */
1024 hwc->state |= PERF_HES_ARCH;
1025}
1026
1027static irqreturn_t pmu_handle_irq(int irq_num, void *dev)
1028{
1029 struct cci_pmu *cci_pmu = dev;
1030 struct cci_pmu_hw_events *events = &cci_pmu->hw_events;
1031 int idx, handled = IRQ_NONE;
1032
1033 raw_spin_lock(&events->pmu_lock);
1034
1035 /* Disable the PMU while we walk through the counters */
1036 __cci_pmu_disable(cci_pmu);
1037 /*
1038 * Iterate over counters and update the corresponding perf events.
1039 * This should work regardless of whether we have per-counter overflow
1040 * interrupt or a combined overflow interrupt.
1041 */
1042 for (idx = 0; idx <= CCI_PMU_CNTR_LAST(cci_pmu); idx++) {
1043 struct perf_event *event = events->events[idx];
1044
1045 if (!event)
1046 continue;
1047
1048 /* Did this counter overflow? */
1049 if (!(pmu_read_register(cci_pmu, idx, CCI_PMU_OVRFLW) &
1050 CCI_PMU_OVRFLW_FLAG))
1051 continue;
1052
1053 pmu_write_register(cci_pmu, CCI_PMU_OVRFLW_FLAG, idx,
1054 CCI_PMU_OVRFLW);
1055
1056 pmu_event_update(event);
1057 pmu_event_set_period(event);
1058 handled = IRQ_HANDLED;
1059 }
1060
1061 /* Enable the PMU and sync possibly overflowed counters */
1062 __cci_pmu_enable_sync(cci_pmu);
1063 raw_spin_unlock(&events->pmu_lock);
1064
1065 return IRQ_RETVAL(handled);
1066}
1067
1068static int cci_pmu_get_hw(struct cci_pmu *cci_pmu)
1069{
1070 int ret = pmu_request_irq(cci_pmu, pmu_handle_irq);
1071 if (ret) {
1072 pmu_free_irq(cci_pmu);
1073 return ret;
1074 }
1075 return 0;
1076}
1077
1078static void cci_pmu_put_hw(struct cci_pmu *cci_pmu)
1079{
1080 pmu_free_irq(cci_pmu);
1081}
1082
1083static void hw_perf_event_destroy(struct perf_event *event)
1084{
1085 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1086 atomic_t *active_events = &cci_pmu->active_events;
1087 struct mutex *reserve_mutex = &cci_pmu->reserve_mutex;
1088
1089 if (atomic_dec_and_mutex_lock(active_events, reserve_mutex)) {
1090 cci_pmu_put_hw(cci_pmu);
1091 mutex_unlock(reserve_mutex);
1092 }
1093}
1094
1095static void cci_pmu_enable(struct pmu *pmu)
1096{
1097 struct cci_pmu *cci_pmu = to_cci_pmu(pmu);
1098 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1099 bool enabled = !bitmap_empty(hw_events->used_mask, cci_pmu->num_cntrs);
1100 unsigned long flags;
1101
1102 if (!enabled)
1103 return;
1104
1105 raw_spin_lock_irqsave(&hw_events->pmu_lock, flags);
1106 __cci_pmu_enable_sync(cci_pmu);
1107 raw_spin_unlock_irqrestore(&hw_events->pmu_lock, flags);
1108
1109}
1110
1111static void cci_pmu_disable(struct pmu *pmu)
1112{
1113 struct cci_pmu *cci_pmu = to_cci_pmu(pmu);
1114 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1115 unsigned long flags;
1116
1117 raw_spin_lock_irqsave(&hw_events->pmu_lock, flags);
1118 __cci_pmu_disable(cci_pmu);
1119 raw_spin_unlock_irqrestore(&hw_events->pmu_lock, flags);
1120}
1121
1122/*
1123 * Check if the idx represents a non-programmable counter.
1124 * All the fixed event counters are mapped before the programmable
1125 * counters.
1126 */
1127static bool pmu_fixed_hw_idx(struct cci_pmu *cci_pmu, int idx)
1128{
1129 return (idx >= 0) && (idx < cci_pmu->model->fixed_hw_cntrs);
1130}
1131
1132static void cci_pmu_start(struct perf_event *event, int pmu_flags)
1133{
1134 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1135 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1136 struct hw_perf_event *hwc = &event->hw;
1137 int idx = hwc->idx;
1138 unsigned long flags;
1139
1140 /*
1141 * To handle interrupt latency, we always reprogram the period
1142 * regardless of PERF_EF_RELOAD.
1143 */
1144 if (pmu_flags & PERF_EF_RELOAD)
1145 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
1146
1147 hwc->state = 0;
1148
1149 if (unlikely(!pmu_is_valid_counter(cci_pmu, idx))) {
1150 dev_err(&cci_pmu->plat_device->dev, "Invalid CCI PMU counter %d\n", idx);
1151 return;
1152 }
1153
1154 raw_spin_lock_irqsave(&hw_events->pmu_lock, flags);
1155
1156 /* Configure the counter unless you are counting a fixed event */
1157 if (!pmu_fixed_hw_idx(cci_pmu, idx))
1158 pmu_set_event(cci_pmu, idx, hwc->config_base);
1159
1160 pmu_event_set_period(event);
1161 pmu_enable_counter(cci_pmu, idx);
1162
1163 raw_spin_unlock_irqrestore(&hw_events->pmu_lock, flags);
1164}
1165
1166static void cci_pmu_stop(struct perf_event *event, int pmu_flags)
1167{
1168 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1169 struct hw_perf_event *hwc = &event->hw;
1170 int idx = hwc->idx;
1171
1172 if (hwc->state & PERF_HES_STOPPED)
1173 return;
1174
1175 if (unlikely(!pmu_is_valid_counter(cci_pmu, idx))) {
1176 dev_err(&cci_pmu->plat_device->dev, "Invalid CCI PMU counter %d\n", idx);
1177 return;
1178 }
1179
1180 /*
1181 * We always reprogram the counter, so ignore PERF_EF_UPDATE. See
1182 * cci_pmu_start()
1183 */
1184 pmu_disable_counter(cci_pmu, idx);
1185 pmu_event_update(event);
1186 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
1187}
1188
1189static int cci_pmu_add(struct perf_event *event, int flags)
1190{
1191 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1192 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1193 struct hw_perf_event *hwc = &event->hw;
1194 int idx;
1195
1196 /* If we don't have a space for the counter then finish early. */
1197 idx = pmu_get_event_idx(hw_events, event);
1198 if (idx < 0)
1199 return idx;
1200
1201 event->hw.idx = idx;
1202 hw_events->events[idx] = event;
1203
1204 hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
1205 if (flags & PERF_EF_START)
1206 cci_pmu_start(event, PERF_EF_RELOAD);
1207
1208 /* Propagate our changes to the userspace mapping. */
1209 perf_event_update_userpage(event);
1210
1211 return 0;
1212}
1213
1214static void cci_pmu_del(struct perf_event *event, int flags)
1215{
1216 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1217 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1218 struct hw_perf_event *hwc = &event->hw;
1219 int idx = hwc->idx;
1220
1221 cci_pmu_stop(event, PERF_EF_UPDATE);
1222 hw_events->events[idx] = NULL;
1223 clear_bit(idx, hw_events->used_mask);
1224
1225 perf_event_update_userpage(event);
1226}
1227
1228static int validate_event(struct pmu *cci_pmu,
1229 struct cci_pmu_hw_events *hw_events,
1230 struct perf_event *event)
1231{
1232 if (is_software_event(event))
1233 return 1;
1234
1235 /*
1236 * Reject groups spanning multiple HW PMUs (e.g. CPU + CCI). The
1237 * core perf code won't check that the pmu->ctx == leader->ctx
1238 * until after pmu->event_init(event).
1239 */
1240 if (event->pmu != cci_pmu)
1241 return 0;
1242
1243 if (event->state < PERF_EVENT_STATE_OFF)
1244 return 1;
1245
1246 if (event->state == PERF_EVENT_STATE_OFF && !event->attr.enable_on_exec)
1247 return 1;
1248
1249 return pmu_get_event_idx(hw_events, event) >= 0;
1250}
1251
1252static int validate_group(struct perf_event *event)
1253{
1254 struct perf_event *sibling, *leader = event->group_leader;
1255 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1256 unsigned long mask[BITS_TO_LONGS(HW_CNTRS_MAX)];
1257 struct cci_pmu_hw_events fake_pmu = {
1258 /*
1259 * Initialise the fake PMU. We only need to populate the
1260 * used_mask for the purposes of validation.
1261 */
1262 .used_mask = mask,
1263 };
1264 bitmap_zero(mask, cci_pmu->num_cntrs);
1265
1266 if (!validate_event(event->pmu, &fake_pmu, leader))
1267 return -EINVAL;
1268
1269 for_each_sibling_event(sibling, leader) {
1270 if (!validate_event(event->pmu, &fake_pmu, sibling))
1271 return -EINVAL;
1272 }
1273
1274 if (!validate_event(event->pmu, &fake_pmu, event))
1275 return -EINVAL;
1276
1277 return 0;
1278}
1279
1280static int __hw_perf_event_init(struct perf_event *event)
1281{
1282 struct hw_perf_event *hwc = &event->hw;
1283 int mapping;
1284
1285 mapping = pmu_map_event(event);
1286
1287 if (mapping < 0) {
1288 pr_debug("event %x:%llx not supported\n", event->attr.type,
1289 event->attr.config);
1290 return mapping;
1291 }
1292
1293 /*
1294 * We don't assign an index until we actually place the event onto
1295 * hardware. Use -1 to signify that we haven't decided where to put it
1296 * yet.
1297 */
1298 hwc->idx = -1;
1299 hwc->config_base = 0;
1300 hwc->config = 0;
1301 hwc->event_base = 0;
1302
1303 /*
1304 * Store the event encoding into the config_base field.
1305 */
1306 hwc->config_base |= (unsigned long)mapping;
1307
1308 if (event->group_leader != event) {
1309 if (validate_group(event) != 0)
1310 return -EINVAL;
1311 }
1312
1313 return 0;
1314}
1315
1316static int cci_pmu_event_init(struct perf_event *event)
1317{
1318 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1319 atomic_t *active_events = &cci_pmu->active_events;
1320 int err = 0;
1321
1322 if (event->attr.type != event->pmu->type)
1323 return -ENOENT;
1324
1325 /* Shared by all CPUs, no meaningful state to sample */
1326 if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
1327 return -EOPNOTSUPP;
1328
1329 /*
1330 * Following the example set by other "uncore" PMUs, we accept any CPU
1331 * and rewrite its affinity dynamically rather than having perf core
1332 * handle cpu == -1 and pid == -1 for this case.
1333 *
1334 * The perf core will pin online CPUs for the duration of this call and
1335 * the event being installed into its context, so the PMU's CPU can't
1336 * change under our feet.
1337 */
1338 if (event->cpu < 0)
1339 return -EINVAL;
1340 event->cpu = cci_pmu->cpu;
1341
1342 event->destroy = hw_perf_event_destroy;
1343 if (!atomic_inc_not_zero(active_events)) {
1344 mutex_lock(&cci_pmu->reserve_mutex);
1345 if (atomic_read(active_events) == 0)
1346 err = cci_pmu_get_hw(cci_pmu);
1347 if (!err)
1348 atomic_inc(active_events);
1349 mutex_unlock(&cci_pmu->reserve_mutex);
1350 }
1351 if (err)
1352 return err;
1353
1354 err = __hw_perf_event_init(event);
1355 if (err)
1356 hw_perf_event_destroy(event);
1357
1358 return err;
1359}
1360
1361static ssize_t pmu_cpumask_attr_show(struct device *dev,
1362 struct device_attribute *attr, char *buf)
1363{
1364 struct pmu *pmu = dev_get_drvdata(dev);
1365 struct cci_pmu *cci_pmu = to_cci_pmu(pmu);
1366
1367 return cpumap_print_to_pagebuf(true, buf, cpumask_of(cci_pmu->cpu));
1368}
1369
1370static struct device_attribute pmu_cpumask_attr =
1371 __ATTR(cpumask, S_IRUGO, pmu_cpumask_attr_show, NULL);
1372
1373static struct attribute *pmu_attrs[] = {
1374 &pmu_cpumask_attr.attr,
1375 NULL,
1376};
1377
1378static const struct attribute_group pmu_attr_group = {
1379 .attrs = pmu_attrs,
1380};
1381
1382static struct attribute_group pmu_format_attr_group = {
1383 .name = "format",
1384 .attrs = NULL, /* Filled in cci_pmu_init_attrs */
1385};
1386
1387static struct attribute_group pmu_event_attr_group = {
1388 .name = "events",
1389 .attrs = NULL, /* Filled in cci_pmu_init_attrs */
1390};
1391
1392static const struct attribute_group *pmu_attr_groups[] = {
1393 &pmu_attr_group,
1394 &pmu_format_attr_group,
1395 &pmu_event_attr_group,
1396 NULL
1397};
1398
1399static int cci_pmu_init(struct cci_pmu *cci_pmu, struct platform_device *pdev)
1400{
1401 const struct cci_pmu_model *model = cci_pmu->model;
1402 char *name = model->name;
1403 u32 num_cntrs;
1404
1405 if (WARN_ON(model->num_hw_cntrs > NUM_HW_CNTRS_MAX))
1406 return -EINVAL;
1407 if (WARN_ON(model->fixed_hw_cntrs > FIXED_HW_CNTRS_MAX))
1408 return -EINVAL;
1409
1410 pmu_event_attr_group.attrs = model->event_attrs;
1411 pmu_format_attr_group.attrs = model->format_attrs;
1412
1413 cci_pmu->pmu = (struct pmu) {
1414 .module = THIS_MODULE,
1415 .name = cci_pmu->model->name,
1416 .task_ctx_nr = perf_invalid_context,
1417 .pmu_enable = cci_pmu_enable,
1418 .pmu_disable = cci_pmu_disable,
1419 .event_init = cci_pmu_event_init,
1420 .add = cci_pmu_add,
1421 .del = cci_pmu_del,
1422 .start = cci_pmu_start,
1423 .stop = cci_pmu_stop,
1424 .read = pmu_read,
1425 .attr_groups = pmu_attr_groups,
1426 .capabilities = PERF_PMU_CAP_NO_EXCLUDE,
1427 };
1428
1429 cci_pmu->plat_device = pdev;
1430 num_cntrs = pmu_get_max_counters(cci_pmu);
1431 if (num_cntrs > cci_pmu->model->num_hw_cntrs) {
1432 dev_warn(&pdev->dev,
1433 "PMU implements more counters(%d) than supported by"
1434 " the model(%d), truncated.",
1435 num_cntrs, cci_pmu->model->num_hw_cntrs);
1436 num_cntrs = cci_pmu->model->num_hw_cntrs;
1437 }
1438 cci_pmu->num_cntrs = num_cntrs + cci_pmu->model->fixed_hw_cntrs;
1439
1440 return perf_pmu_register(&cci_pmu->pmu, name, -1);
1441}
1442
1443static int cci_pmu_offline_cpu(unsigned int cpu)
1444{
1445 int target;
1446
1447 if (!g_cci_pmu || cpu != g_cci_pmu->cpu)
1448 return 0;
1449
1450 target = cpumask_any_but(cpu_online_mask, cpu);
1451 if (target >= nr_cpu_ids)
1452 return 0;
1453
1454 perf_pmu_migrate_context(&g_cci_pmu->pmu, cpu, target);
1455 g_cci_pmu->cpu = target;
1456 return 0;
1457}
1458
1459static __maybe_unused struct cci_pmu_model cci_pmu_models[] = {
1460#ifdef CONFIG_ARM_CCI400_PMU
1461 [CCI400_R0] = {
1462 .name = "CCI_400",
1463 .fixed_hw_cntrs = FIXED_HW_CNTRS_CII_4XX, /* Cycle counter */
1464 .num_hw_cntrs = NUM_HW_CNTRS_CII_4XX,
1465 .cntr_size = SZ_4K,
1466 .format_attrs = cci400_pmu_format_attrs,
1467 .event_attrs = cci400_r0_pmu_event_attrs,
1468 .event_ranges = {
1469 [CCI_IF_SLAVE] = {
1470 CCI400_R0_SLAVE_PORT_MIN_EV,
1471 CCI400_R0_SLAVE_PORT_MAX_EV,
1472 },
1473 [CCI_IF_MASTER] = {
1474 CCI400_R0_MASTER_PORT_MIN_EV,
1475 CCI400_R0_MASTER_PORT_MAX_EV,
1476 },
1477 },
1478 .validate_hw_event = cci400_validate_hw_event,
1479 .get_event_idx = cci400_get_event_idx,
1480 },
1481 [CCI400_R1] = {
1482 .name = "CCI_400_r1",
1483 .fixed_hw_cntrs = FIXED_HW_CNTRS_CII_4XX, /* Cycle counter */
1484 .num_hw_cntrs = NUM_HW_CNTRS_CII_4XX,
1485 .cntr_size = SZ_4K,
1486 .format_attrs = cci400_pmu_format_attrs,
1487 .event_attrs = cci400_r1_pmu_event_attrs,
1488 .event_ranges = {
1489 [CCI_IF_SLAVE] = {
1490 CCI400_R1_SLAVE_PORT_MIN_EV,
1491 CCI400_R1_SLAVE_PORT_MAX_EV,
1492 },
1493 [CCI_IF_MASTER] = {
1494 CCI400_R1_MASTER_PORT_MIN_EV,
1495 CCI400_R1_MASTER_PORT_MAX_EV,
1496 },
1497 },
1498 .validate_hw_event = cci400_validate_hw_event,
1499 .get_event_idx = cci400_get_event_idx,
1500 },
1501#endif
1502#ifdef CONFIG_ARM_CCI5xx_PMU
1503 [CCI500_R0] = {
1504 .name = "CCI_500",
1505 .fixed_hw_cntrs = FIXED_HW_CNTRS_CII_5XX,
1506 .num_hw_cntrs = NUM_HW_CNTRS_CII_5XX,
1507 .cntr_size = SZ_64K,
1508 .format_attrs = cci5xx_pmu_format_attrs,
1509 .event_attrs = cci5xx_pmu_event_attrs,
1510 .event_ranges = {
1511 [CCI_IF_SLAVE] = {
1512 CCI5xx_SLAVE_PORT_MIN_EV,
1513 CCI5xx_SLAVE_PORT_MAX_EV,
1514 },
1515 [CCI_IF_MASTER] = {
1516 CCI5xx_MASTER_PORT_MIN_EV,
1517 CCI5xx_MASTER_PORT_MAX_EV,
1518 },
1519 [CCI_IF_GLOBAL] = {
1520 CCI5xx_GLOBAL_PORT_MIN_EV,
1521 CCI5xx_GLOBAL_PORT_MAX_EV,
1522 },
1523 },
1524 .validate_hw_event = cci500_validate_hw_event,
1525 .write_counters = cci5xx_pmu_write_counters,
1526 },
1527 [CCI550_R0] = {
1528 .name = "CCI_550",
1529 .fixed_hw_cntrs = FIXED_HW_CNTRS_CII_5XX,
1530 .num_hw_cntrs = NUM_HW_CNTRS_CII_5XX,
1531 .cntr_size = SZ_64K,
1532 .format_attrs = cci5xx_pmu_format_attrs,
1533 .event_attrs = cci5xx_pmu_event_attrs,
1534 .event_ranges = {
1535 [CCI_IF_SLAVE] = {
1536 CCI5xx_SLAVE_PORT_MIN_EV,
1537 CCI5xx_SLAVE_PORT_MAX_EV,
1538 },
1539 [CCI_IF_MASTER] = {
1540 CCI5xx_MASTER_PORT_MIN_EV,
1541 CCI5xx_MASTER_PORT_MAX_EV,
1542 },
1543 [CCI_IF_GLOBAL] = {
1544 CCI5xx_GLOBAL_PORT_MIN_EV,
1545 CCI5xx_GLOBAL_PORT_MAX_EV,
1546 },
1547 },
1548 .validate_hw_event = cci550_validate_hw_event,
1549 .write_counters = cci5xx_pmu_write_counters,
1550 },
1551#endif
1552};
1553
1554static const struct of_device_id arm_cci_pmu_matches[] = {
1555#ifdef CONFIG_ARM_CCI400_PMU
1556 {
1557 .compatible = "arm,cci-400-pmu",
1558 .data = NULL,
1559 },
1560 {
1561 .compatible = "arm,cci-400-pmu,r0",
1562 .data = &cci_pmu_models[CCI400_R0],
1563 },
1564 {
1565 .compatible = "arm,cci-400-pmu,r1",
1566 .data = &cci_pmu_models[CCI400_R1],
1567 },
1568#endif
1569#ifdef CONFIG_ARM_CCI5xx_PMU
1570 {
1571 .compatible = "arm,cci-500-pmu,r0",
1572 .data = &cci_pmu_models[CCI500_R0],
1573 },
1574 {
1575 .compatible = "arm,cci-550-pmu,r0",
1576 .data = &cci_pmu_models[CCI550_R0],
1577 },
1578#endif
1579 {},
1580};
1581MODULE_DEVICE_TABLE(of, arm_cci_pmu_matches);
1582
1583static bool is_duplicate_irq(int irq, int *irqs, int nr_irqs)
1584{
1585 int i;
1586
1587 for (i = 0; i < nr_irqs; i++)
1588 if (irq == irqs[i])
1589 return true;
1590
1591 return false;
1592}
1593
1594static struct cci_pmu *cci_pmu_alloc(struct device *dev)
1595{
1596 struct cci_pmu *cci_pmu;
1597 const struct cci_pmu_model *model;
1598
1599 /*
1600 * All allocations are devm_* hence we don't have to free
1601 * them explicitly on an error, as it would end up in driver
1602 * detach.
1603 */
1604 cci_pmu = devm_kzalloc(dev, sizeof(*cci_pmu), GFP_KERNEL);
1605 if (!cci_pmu)
1606 return ERR_PTR(-ENOMEM);
1607
1608 cci_pmu->ctrl_base = *(void __iomem **)dev->platform_data;
1609
1610 model = of_device_get_match_data(dev);
1611 if (!model) {
1612 dev_warn(dev,
1613 "DEPRECATED compatible property, requires secure access to CCI registers");
1614 model = probe_cci_model(cci_pmu);
1615 }
1616 if (!model) {
1617 dev_warn(dev, "CCI PMU version not supported\n");
1618 return ERR_PTR(-ENODEV);
1619 }
1620
1621 cci_pmu->model = model;
1622 cci_pmu->irqs = devm_kcalloc(dev, CCI_PMU_MAX_HW_CNTRS(model),
1623 sizeof(*cci_pmu->irqs), GFP_KERNEL);
1624 if (!cci_pmu->irqs)
1625 return ERR_PTR(-ENOMEM);
1626 cci_pmu->hw_events.events = devm_kcalloc(dev,
1627 CCI_PMU_MAX_HW_CNTRS(model),
1628 sizeof(*cci_pmu->hw_events.events),
1629 GFP_KERNEL);
1630 if (!cci_pmu->hw_events.events)
1631 return ERR_PTR(-ENOMEM);
1632 cci_pmu->hw_events.used_mask = devm_bitmap_zalloc(dev,
1633 CCI_PMU_MAX_HW_CNTRS(model),
1634 GFP_KERNEL);
1635 if (!cci_pmu->hw_events.used_mask)
1636 return ERR_PTR(-ENOMEM);
1637
1638 return cci_pmu;
1639}
1640
1641static int cci_pmu_probe(struct platform_device *pdev)
1642{
1643 struct cci_pmu *cci_pmu;
1644 int i, ret, irq;
1645
1646 cci_pmu = cci_pmu_alloc(&pdev->dev);
1647 if (IS_ERR(cci_pmu))
1648 return PTR_ERR(cci_pmu);
1649
1650 cci_pmu->base = devm_platform_ioremap_resource(pdev, 0);
1651 if (IS_ERR(cci_pmu->base))
1652 return -ENOMEM;
1653
1654 /*
1655 * CCI PMU has one overflow interrupt per counter; but some may be tied
1656 * together to a common interrupt.
1657 */
1658 cci_pmu->nr_irqs = 0;
1659 for (i = 0; i < CCI_PMU_MAX_HW_CNTRS(cci_pmu->model); i++) {
1660 irq = platform_get_irq(pdev, i);
1661 if (irq < 0)
1662 break;
1663
1664 if (is_duplicate_irq(irq, cci_pmu->irqs, cci_pmu->nr_irqs))
1665 continue;
1666
1667 cci_pmu->irqs[cci_pmu->nr_irqs++] = irq;
1668 }
1669
1670 /*
1671 * Ensure that the device tree has as many interrupts as the number
1672 * of counters.
1673 */
1674 if (i < CCI_PMU_MAX_HW_CNTRS(cci_pmu->model)) {
1675 dev_warn(&pdev->dev, "In-correct number of interrupts: %d, should be %d\n",
1676 i, CCI_PMU_MAX_HW_CNTRS(cci_pmu->model));
1677 return -EINVAL;
1678 }
1679
1680 raw_spin_lock_init(&cci_pmu->hw_events.pmu_lock);
1681 mutex_init(&cci_pmu->reserve_mutex);
1682 atomic_set(&cci_pmu->active_events, 0);
1683
1684 cci_pmu->cpu = raw_smp_processor_id();
1685 g_cci_pmu = cci_pmu;
1686 cpuhp_setup_state_nocalls(CPUHP_AP_PERF_ARM_CCI_ONLINE,
1687 "perf/arm/cci:online", NULL,
1688 cci_pmu_offline_cpu);
1689
1690 ret = cci_pmu_init(cci_pmu, pdev);
1691 if (ret)
1692 goto error_pmu_init;
1693
1694 pr_info("ARM %s PMU driver probed", cci_pmu->model->name);
1695 return 0;
1696
1697error_pmu_init:
1698 cpuhp_remove_state(CPUHP_AP_PERF_ARM_CCI_ONLINE);
1699 g_cci_pmu = NULL;
1700 return ret;
1701}
1702
1703static int cci_pmu_remove(struct platform_device *pdev)
1704{
1705 if (!g_cci_pmu)
1706 return 0;
1707
1708 cpuhp_remove_state(CPUHP_AP_PERF_ARM_CCI_ONLINE);
1709 perf_pmu_unregister(&g_cci_pmu->pmu);
1710 g_cci_pmu = NULL;
1711
1712 return 0;
1713}
1714
1715static struct platform_driver cci_pmu_driver = {
1716 .driver = {
1717 .name = DRIVER_NAME,
1718 .of_match_table = arm_cci_pmu_matches,
1719 .suppress_bind_attrs = true,
1720 },
1721 .probe = cci_pmu_probe,
1722 .remove = cci_pmu_remove,
1723};
1724
1725module_platform_driver(cci_pmu_driver);
1726MODULE_LICENSE("GPL v2");
1727MODULE_DESCRIPTION("ARM CCI PMU support");
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_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>
18
19#define DRIVER_NAME "ARM-CCI PMU"
20
21#define CCI_PMCR 0x0100
22#define CCI_PID2 0x0fe8
23
24#define CCI_PMCR_CEN 0x00000001
25#define CCI_PMCR_NCNT_MASK 0x0000f800
26#define CCI_PMCR_NCNT_SHIFT 11
27
28#define CCI_PID2_REV_MASK 0xf0
29#define CCI_PID2_REV_SHIFT 4
30
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
35
36#define CCI_PMU_OVRFLW_FLAG 1
37
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)
42
43#define CCI_PMU_MAX_HW_CNTRS(model) \
44 ((model)->num_hw_cntrs + (model)->fixed_hw_cntrs)
45
46/* Types of interfaces that can generate events */
47enum {
48 CCI_IF_SLAVE,
49 CCI_IF_MASTER,
50#ifdef CONFIG_ARM_CCI5xx_PMU
51 CCI_IF_GLOBAL,
52#endif
53 CCI_IF_MAX,
54};
55
56struct event_range {
57 u32 min;
58 u32 max;
59};
60
61struct cci_pmu_hw_events {
62 struct perf_event **events;
63 unsigned long *used_mask;
64 raw_spinlock_t pmu_lock;
65};
66
67struct cci_pmu;
68/*
69 * struct cci_pmu_model:
70 * @fixed_hw_cntrs - Number of fixed event counters
71 * @num_hw_cntrs - Maximum number of programmable event counters
72 * @cntr_size - Size of an event counter mapping
73 */
74struct cci_pmu_model {
75 char *name;
76 u32 fixed_hw_cntrs;
77 u32 num_hw_cntrs;
78 u32 cntr_size;
79 struct attribute **format_attrs;
80 struct attribute **event_attrs;
81 struct event_range event_ranges[CCI_IF_MAX];
82 int (*validate_hw_event)(struct cci_pmu *, unsigned long);
83 int (*get_event_idx)(struct cci_pmu *, struct cci_pmu_hw_events *, unsigned long);
84 void (*write_counters)(struct cci_pmu *, unsigned long *);
85};
86
87static struct cci_pmu_model cci_pmu_models[];
88
89struct cci_pmu {
90 void __iomem *base;
91 void __iomem *ctrl_base;
92 struct pmu pmu;
93 int cpu;
94 int nr_irqs;
95 int *irqs;
96 unsigned long active_irqs;
97 const struct cci_pmu_model *model;
98 struct cci_pmu_hw_events hw_events;
99 struct platform_device *plat_device;
100 int num_cntrs;
101 atomic_t active_events;
102 struct mutex reserve_mutex;
103};
104
105#define to_cci_pmu(c) (container_of(c, struct cci_pmu, pmu))
106
107static struct cci_pmu *g_cci_pmu;
108
109enum cci_models {
110#ifdef CONFIG_ARM_CCI400_PMU
111 CCI400_R0,
112 CCI400_R1,
113#endif
114#ifdef CONFIG_ARM_CCI5xx_PMU
115 CCI500_R0,
116 CCI550_R0,
117#endif
118 CCI_MODEL_MAX
119};
120
121static void pmu_write_counters(struct cci_pmu *cci_pmu,
122 unsigned long *mask);
123static ssize_t cci_pmu_format_show(struct device *dev,
124 struct device_attribute *attr, char *buf);
125static ssize_t cci_pmu_event_show(struct device *dev,
126 struct device_attribute *attr, char *buf);
127
128#define CCI_EXT_ATTR_ENTRY(_name, _func, _config) \
129 &((struct dev_ext_attribute[]) { \
130 { __ATTR(_name, S_IRUGO, _func, NULL), (void *)_config } \
131 })[0].attr.attr
132
133#define CCI_FORMAT_EXT_ATTR_ENTRY(_name, _config) \
134 CCI_EXT_ATTR_ENTRY(_name, cci_pmu_format_show, (char *)_config)
135#define CCI_EVENT_EXT_ATTR_ENTRY(_name, _config) \
136 CCI_EXT_ATTR_ENTRY(_name, cci_pmu_event_show, (unsigned long)_config)
137
138/* CCI400 PMU Specific definitions */
139
140#ifdef CONFIG_ARM_CCI400_PMU
141
142/* Port ids */
143#define CCI400_PORT_S0 0
144#define CCI400_PORT_S1 1
145#define CCI400_PORT_S2 2
146#define CCI400_PORT_S3 3
147#define CCI400_PORT_S4 4
148#define CCI400_PORT_M0 5
149#define CCI400_PORT_M1 6
150#define CCI400_PORT_M2 7
151
152#define CCI400_R1_PX 5
153
154/*
155 * Instead of an event id to monitor CCI cycles, a dedicated counter is
156 * provided. Use 0xff to represent CCI cycles and hope that no future revisions
157 * make use of this event in hardware.
158 */
159enum cci400_perf_events {
160 CCI400_PMU_CYCLES = 0xff
161};
162
163#define CCI400_PMU_CYCLE_CNTR_IDX 0
164#define CCI400_PMU_CNTR0_IDX 1
165
166/*
167 * CCI PMU event id is an 8-bit value made of two parts - bits 7:5 for one of 8
168 * ports and bits 4:0 are event codes. There are different event codes
169 * associated with each port type.
170 *
171 * Additionally, the range of events associated with the port types changed
172 * between Rev0 and Rev1.
173 *
174 * The constants below define the range of valid codes for each port type for
175 * the different revisions and are used to validate the event to be monitored.
176 */
177
178#define CCI400_PMU_EVENT_MASK 0xffUL
179#define CCI400_PMU_EVENT_SOURCE_SHIFT 5
180#define CCI400_PMU_EVENT_SOURCE_MASK 0x7
181#define CCI400_PMU_EVENT_CODE_SHIFT 0
182#define CCI400_PMU_EVENT_CODE_MASK 0x1f
183#define CCI400_PMU_EVENT_SOURCE(event) \
184 ((event >> CCI400_PMU_EVENT_SOURCE_SHIFT) & \
185 CCI400_PMU_EVENT_SOURCE_MASK)
186#define CCI400_PMU_EVENT_CODE(event) \
187 ((event >> CCI400_PMU_EVENT_CODE_SHIFT) & CCI400_PMU_EVENT_CODE_MASK)
188
189#define CCI400_R0_SLAVE_PORT_MIN_EV 0x00
190#define CCI400_R0_SLAVE_PORT_MAX_EV 0x13
191#define CCI400_R0_MASTER_PORT_MIN_EV 0x14
192#define CCI400_R0_MASTER_PORT_MAX_EV 0x1a
193
194#define CCI400_R1_SLAVE_PORT_MIN_EV 0x00
195#define CCI400_R1_SLAVE_PORT_MAX_EV 0x14
196#define CCI400_R1_MASTER_PORT_MIN_EV 0x00
197#define CCI400_R1_MASTER_PORT_MAX_EV 0x11
198
199#define CCI400_CYCLE_EVENT_EXT_ATTR_ENTRY(_name, _config) \
200 CCI_EXT_ATTR_ENTRY(_name, cci400_pmu_cycle_event_show, \
201 (unsigned long)_config)
202
203static ssize_t cci400_pmu_cycle_event_show(struct device *dev,
204 struct device_attribute *attr, char *buf);
205
206static struct attribute *cci400_pmu_format_attrs[] = {
207 CCI_FORMAT_EXT_ATTR_ENTRY(event, "config:0-4"),
208 CCI_FORMAT_EXT_ATTR_ENTRY(source, "config:5-7"),
209 NULL
210};
211
212static struct attribute *cci400_r0_pmu_event_attrs[] = {
213 /* Slave events */
214 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_any, 0x0),
215 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_device, 0x01),
216 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_normal_or_nonshareable, 0x2),
217 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_inner_or_outershareable, 0x3),
218 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_cache_maintenance, 0x4),
219 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_mem_barrier, 0x5),
220 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_sync_barrier, 0x6),
221 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg, 0x7),
222 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg_sync, 0x8),
223 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_stall_tt_full, 0x9),
224 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_last_hs_snoop, 0xA),
225 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_stall_rvalids_h_rready_l, 0xB),
226 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_any, 0xC),
227 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_device, 0xD),
228 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_normal_or_nonshareable, 0xE),
229 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_inner_or_outershare_wback_wclean, 0xF),
230 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_write_unique, 0x10),
231 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_write_line_unique, 0x11),
232 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_evict, 0x12),
233 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_stall_tt_full, 0x13),
234 /* Master events */
235 CCI_EVENT_EXT_ATTR_ENTRY(mi_retry_speculative_fetch, 0x14),
236 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_addr_hazard, 0x15),
237 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_id_hazard, 0x16),
238 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_tt_full, 0x17),
239 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_barrier_hazard, 0x18),
240 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_barrier_hazard, 0x19),
241 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_tt_full, 0x1A),
242 /* Special event for cycles counter */
243 CCI400_CYCLE_EVENT_EXT_ATTR_ENTRY(cycles, 0xff),
244 NULL
245};
246
247static struct attribute *cci400_r1_pmu_event_attrs[] = {
248 /* Slave events */
249 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_any, 0x0),
250 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_device, 0x01),
251 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_normal_or_nonshareable, 0x2),
252 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_inner_or_outershareable, 0x3),
253 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_cache_maintenance, 0x4),
254 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_mem_barrier, 0x5),
255 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_sync_barrier, 0x6),
256 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg, 0x7),
257 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg_sync, 0x8),
258 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_stall_tt_full, 0x9),
259 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_last_hs_snoop, 0xA),
260 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_stall_rvalids_h_rready_l, 0xB),
261 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_any, 0xC),
262 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_device, 0xD),
263 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_normal_or_nonshareable, 0xE),
264 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_inner_or_outershare_wback_wclean, 0xF),
265 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_write_unique, 0x10),
266 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_write_line_unique, 0x11),
267 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_evict, 0x12),
268 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_stall_tt_full, 0x13),
269 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_stall_slave_id_hazard, 0x14),
270 /* Master events */
271 CCI_EVENT_EXT_ATTR_ENTRY(mi_retry_speculative_fetch, 0x0),
272 CCI_EVENT_EXT_ATTR_ENTRY(mi_stall_cycle_addr_hazard, 0x1),
273 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_master_id_hazard, 0x2),
274 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_hi_prio_rtq_full, 0x3),
275 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_barrier_hazard, 0x4),
276 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_barrier_hazard, 0x5),
277 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_wtq_full, 0x6),
278 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_low_prio_rtq_full, 0x7),
279 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_mid_prio_rtq_full, 0x8),
280 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_qvn_vn0, 0x9),
281 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_qvn_vn1, 0xA),
282 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_qvn_vn2, 0xB),
283 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_qvn_vn3, 0xC),
284 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_qvn_vn0, 0xD),
285 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_qvn_vn1, 0xE),
286 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_qvn_vn2, 0xF),
287 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_qvn_vn3, 0x10),
288 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_unique_or_line_unique_addr_hazard, 0x11),
289 /* Special event for cycles counter */
290 CCI400_CYCLE_EVENT_EXT_ATTR_ENTRY(cycles, 0xff),
291 NULL
292};
293
294static ssize_t cci400_pmu_cycle_event_show(struct device *dev,
295 struct device_attribute *attr, char *buf)
296{
297 struct dev_ext_attribute *eattr = container_of(attr,
298 struct dev_ext_attribute, attr);
299 return snprintf(buf, PAGE_SIZE, "config=0x%lx\n", (unsigned long)eattr->var);
300}
301
302static int cci400_get_event_idx(struct cci_pmu *cci_pmu,
303 struct cci_pmu_hw_events *hw,
304 unsigned long cci_event)
305{
306 int idx;
307
308 /* cycles event idx is fixed */
309 if (cci_event == CCI400_PMU_CYCLES) {
310 if (test_and_set_bit(CCI400_PMU_CYCLE_CNTR_IDX, hw->used_mask))
311 return -EAGAIN;
312
313 return CCI400_PMU_CYCLE_CNTR_IDX;
314 }
315
316 for (idx = CCI400_PMU_CNTR0_IDX; idx <= CCI_PMU_CNTR_LAST(cci_pmu); ++idx)
317 if (!test_and_set_bit(idx, hw->used_mask))
318 return idx;
319
320 /* No counters available */
321 return -EAGAIN;
322}
323
324static int cci400_validate_hw_event(struct cci_pmu *cci_pmu, unsigned long hw_event)
325{
326 u8 ev_source = CCI400_PMU_EVENT_SOURCE(hw_event);
327 u8 ev_code = CCI400_PMU_EVENT_CODE(hw_event);
328 int if_type;
329
330 if (hw_event & ~CCI400_PMU_EVENT_MASK)
331 return -ENOENT;
332
333 if (hw_event == CCI400_PMU_CYCLES)
334 return hw_event;
335
336 switch (ev_source) {
337 case CCI400_PORT_S0:
338 case CCI400_PORT_S1:
339 case CCI400_PORT_S2:
340 case CCI400_PORT_S3:
341 case CCI400_PORT_S4:
342 /* Slave Interface */
343 if_type = CCI_IF_SLAVE;
344 break;
345 case CCI400_PORT_M0:
346 case CCI400_PORT_M1:
347 case CCI400_PORT_M2:
348 /* Master Interface */
349 if_type = CCI_IF_MASTER;
350 break;
351 default:
352 return -ENOENT;
353 }
354
355 if (ev_code >= cci_pmu->model->event_ranges[if_type].min &&
356 ev_code <= cci_pmu->model->event_ranges[if_type].max)
357 return hw_event;
358
359 return -ENOENT;
360}
361
362static int probe_cci400_revision(struct cci_pmu *cci_pmu)
363{
364 int rev;
365 rev = readl_relaxed(cci_pmu->ctrl_base + CCI_PID2) & CCI_PID2_REV_MASK;
366 rev >>= CCI_PID2_REV_SHIFT;
367
368 if (rev < CCI400_R1_PX)
369 return CCI400_R0;
370 else
371 return CCI400_R1;
372}
373
374static const struct cci_pmu_model *probe_cci_model(struct cci_pmu *cci_pmu)
375{
376 if (platform_has_secure_cci_access())
377 return &cci_pmu_models[probe_cci400_revision(cci_pmu)];
378 return NULL;
379}
380#else /* !CONFIG_ARM_CCI400_PMU */
381static inline struct cci_pmu_model *probe_cci_model(struct cci_pmu *cci_pmu)
382{
383 return NULL;
384}
385#endif /* CONFIG_ARM_CCI400_PMU */
386
387#ifdef CONFIG_ARM_CCI5xx_PMU
388
389/*
390 * CCI5xx PMU event id is an 9-bit value made of two parts.
391 * bits [8:5] - Source for the event
392 * bits [4:0] - Event code (specific to type of interface)
393 *
394 *
395 */
396
397/* Port ids */
398#define CCI5xx_PORT_S0 0x0
399#define CCI5xx_PORT_S1 0x1
400#define CCI5xx_PORT_S2 0x2
401#define CCI5xx_PORT_S3 0x3
402#define CCI5xx_PORT_S4 0x4
403#define CCI5xx_PORT_S5 0x5
404#define CCI5xx_PORT_S6 0x6
405
406#define CCI5xx_PORT_M0 0x8
407#define CCI5xx_PORT_M1 0x9
408#define CCI5xx_PORT_M2 0xa
409#define CCI5xx_PORT_M3 0xb
410#define CCI5xx_PORT_M4 0xc
411#define CCI5xx_PORT_M5 0xd
412#define CCI5xx_PORT_M6 0xe
413
414#define CCI5xx_PORT_GLOBAL 0xf
415
416#define CCI5xx_PMU_EVENT_MASK 0x1ffUL
417#define CCI5xx_PMU_EVENT_SOURCE_SHIFT 0x5
418#define CCI5xx_PMU_EVENT_SOURCE_MASK 0xf
419#define CCI5xx_PMU_EVENT_CODE_SHIFT 0x0
420#define CCI5xx_PMU_EVENT_CODE_MASK 0x1f
421
422#define CCI5xx_PMU_EVENT_SOURCE(event) \
423 ((event >> CCI5xx_PMU_EVENT_SOURCE_SHIFT) & CCI5xx_PMU_EVENT_SOURCE_MASK)
424#define CCI5xx_PMU_EVENT_CODE(event) \
425 ((event >> CCI5xx_PMU_EVENT_CODE_SHIFT) & CCI5xx_PMU_EVENT_CODE_MASK)
426
427#define CCI5xx_SLAVE_PORT_MIN_EV 0x00
428#define CCI5xx_SLAVE_PORT_MAX_EV 0x1f
429#define CCI5xx_MASTER_PORT_MIN_EV 0x00
430#define CCI5xx_MASTER_PORT_MAX_EV 0x06
431#define CCI5xx_GLOBAL_PORT_MIN_EV 0x00
432#define CCI5xx_GLOBAL_PORT_MAX_EV 0x0f
433
434
435#define CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(_name, _config) \
436 CCI_EXT_ATTR_ENTRY(_name, cci5xx_pmu_global_event_show, \
437 (unsigned long) _config)
438
439static ssize_t cci5xx_pmu_global_event_show(struct device *dev,
440 struct device_attribute *attr, char *buf);
441
442static struct attribute *cci5xx_pmu_format_attrs[] = {
443 CCI_FORMAT_EXT_ATTR_ENTRY(event, "config:0-4"),
444 CCI_FORMAT_EXT_ATTR_ENTRY(source, "config:5-8"),
445 NULL,
446};
447
448static struct attribute *cci5xx_pmu_event_attrs[] = {
449 /* Slave events */
450 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_arvalid, 0x0),
451 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_dev, 0x1),
452 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_nonshareable, 0x2),
453 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_shareable_non_alloc, 0x3),
454 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_shareable_alloc, 0x4),
455 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_invalidate, 0x5),
456 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_cache_maint, 0x6),
457 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg, 0x7),
458 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_rval, 0x8),
459 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_rlast_snoop, 0x9),
460 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_awalid, 0xA),
461 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_dev, 0xB),
462 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_non_shareable, 0xC),
463 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_share_wb, 0xD),
464 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_share_wlu, 0xE),
465 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_share_wunique, 0xF),
466 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_evict, 0x10),
467 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_wrevict, 0x11),
468 CCI_EVENT_EXT_ATTR_ENTRY(si_w_data_beat, 0x12),
469 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_acvalid, 0x13),
470 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_read, 0x14),
471 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_clean, 0x15),
472 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_data_transfer_low, 0x16),
473 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_stall_arvalid, 0x17),
474 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_stall, 0x18),
475 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_stall, 0x19),
476 CCI_EVENT_EXT_ATTR_ENTRY(si_w_data_stall, 0x1A),
477 CCI_EVENT_EXT_ATTR_ENTRY(si_w_resp_stall, 0x1B),
478 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_stall, 0x1C),
479 CCI_EVENT_EXT_ATTR_ENTRY(si_s_data_stall, 0x1D),
480 CCI_EVENT_EXT_ATTR_ENTRY(si_rq_stall_ot_limit, 0x1E),
481 CCI_EVENT_EXT_ATTR_ENTRY(si_r_stall_arbit, 0x1F),
482
483 /* Master events */
484 CCI_EVENT_EXT_ATTR_ENTRY(mi_r_data_beat_any, 0x0),
485 CCI_EVENT_EXT_ATTR_ENTRY(mi_w_data_beat_any, 0x1),
486 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall, 0x2),
487 CCI_EVENT_EXT_ATTR_ENTRY(mi_r_data_stall, 0x3),
488 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall, 0x4),
489 CCI_EVENT_EXT_ATTR_ENTRY(mi_w_data_stall, 0x5),
490 CCI_EVENT_EXT_ATTR_ENTRY(mi_w_resp_stall, 0x6),
491
492 /* Global events */
493 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_filter_bank_0_1, 0x0),
494 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_filter_bank_2_3, 0x1),
495 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_filter_bank_4_5, 0x2),
496 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_filter_bank_6_7, 0x3),
497 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_miss_filter_bank_0_1, 0x4),
498 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_miss_filter_bank_2_3, 0x5),
499 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_miss_filter_bank_4_5, 0x6),
500 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_miss_filter_bank_6_7, 0x7),
501 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_back_invalidation, 0x8),
502 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_stall_alloc_busy, 0x9),
503 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_stall_tt_full, 0xA),
504 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_wrq, 0xB),
505 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_cd_hs, 0xC),
506 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_rq_stall_addr_hazard, 0xD),
507 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_rq_stall_tt_full, 0xE),
508 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_rq_tzmp1_prot, 0xF),
509 NULL
510};
511
512static ssize_t cci5xx_pmu_global_event_show(struct device *dev,
513 struct device_attribute *attr, char *buf)
514{
515 struct dev_ext_attribute *eattr = container_of(attr,
516 struct dev_ext_attribute, attr);
517 /* Global events have single fixed source code */
518 return snprintf(buf, PAGE_SIZE, "event=0x%lx,source=0x%x\n",
519 (unsigned long)eattr->var, CCI5xx_PORT_GLOBAL);
520}
521
522/*
523 * CCI500 provides 8 independent event counters that can count
524 * any of the events available.
525 * CCI500 PMU event source ids
526 * 0x0-0x6 - Slave interfaces
527 * 0x8-0xD - Master interfaces
528 * 0xf - Global Events
529 * 0x7,0xe - Reserved
530 */
531static int cci500_validate_hw_event(struct cci_pmu *cci_pmu,
532 unsigned long hw_event)
533{
534 u32 ev_source = CCI5xx_PMU_EVENT_SOURCE(hw_event);
535 u32 ev_code = CCI5xx_PMU_EVENT_CODE(hw_event);
536 int if_type;
537
538 if (hw_event & ~CCI5xx_PMU_EVENT_MASK)
539 return -ENOENT;
540
541 switch (ev_source) {
542 case CCI5xx_PORT_S0:
543 case CCI5xx_PORT_S1:
544 case CCI5xx_PORT_S2:
545 case CCI5xx_PORT_S3:
546 case CCI5xx_PORT_S4:
547 case CCI5xx_PORT_S5:
548 case CCI5xx_PORT_S6:
549 if_type = CCI_IF_SLAVE;
550 break;
551 case CCI5xx_PORT_M0:
552 case CCI5xx_PORT_M1:
553 case CCI5xx_PORT_M2:
554 case CCI5xx_PORT_M3:
555 case CCI5xx_PORT_M4:
556 case CCI5xx_PORT_M5:
557 if_type = CCI_IF_MASTER;
558 break;
559 case CCI5xx_PORT_GLOBAL:
560 if_type = CCI_IF_GLOBAL;
561 break;
562 default:
563 return -ENOENT;
564 }
565
566 if (ev_code >= cci_pmu->model->event_ranges[if_type].min &&
567 ev_code <= cci_pmu->model->event_ranges[if_type].max)
568 return hw_event;
569
570 return -ENOENT;
571}
572
573/*
574 * CCI550 provides 8 independent event counters that can count
575 * any of the events available.
576 * CCI550 PMU event source ids
577 * 0x0-0x6 - Slave interfaces
578 * 0x8-0xe - Master interfaces
579 * 0xf - Global Events
580 * 0x7 - Reserved
581 */
582static int cci550_validate_hw_event(struct cci_pmu *cci_pmu,
583 unsigned long hw_event)
584{
585 u32 ev_source = CCI5xx_PMU_EVENT_SOURCE(hw_event);
586 u32 ev_code = CCI5xx_PMU_EVENT_CODE(hw_event);
587 int if_type;
588
589 if (hw_event & ~CCI5xx_PMU_EVENT_MASK)
590 return -ENOENT;
591
592 switch (ev_source) {
593 case CCI5xx_PORT_S0:
594 case CCI5xx_PORT_S1:
595 case CCI5xx_PORT_S2:
596 case CCI5xx_PORT_S3:
597 case CCI5xx_PORT_S4:
598 case CCI5xx_PORT_S5:
599 case CCI5xx_PORT_S6:
600 if_type = CCI_IF_SLAVE;
601 break;
602 case CCI5xx_PORT_M0:
603 case CCI5xx_PORT_M1:
604 case CCI5xx_PORT_M2:
605 case CCI5xx_PORT_M3:
606 case CCI5xx_PORT_M4:
607 case CCI5xx_PORT_M5:
608 case CCI5xx_PORT_M6:
609 if_type = CCI_IF_MASTER;
610 break;
611 case CCI5xx_PORT_GLOBAL:
612 if_type = CCI_IF_GLOBAL;
613 break;
614 default:
615 return -ENOENT;
616 }
617
618 if (ev_code >= cci_pmu->model->event_ranges[if_type].min &&
619 ev_code <= cci_pmu->model->event_ranges[if_type].max)
620 return hw_event;
621
622 return -ENOENT;
623}
624
625#endif /* CONFIG_ARM_CCI5xx_PMU */
626
627/*
628 * Program the CCI PMU counters which have PERF_HES_ARCH set
629 * with the event period and mark them ready before we enable
630 * PMU.
631 */
632static void cci_pmu_sync_counters(struct cci_pmu *cci_pmu)
633{
634 int i;
635 struct cci_pmu_hw_events *cci_hw = &cci_pmu->hw_events;
636
637 DECLARE_BITMAP(mask, cci_pmu->num_cntrs);
638
639 bitmap_zero(mask, cci_pmu->num_cntrs);
640 for_each_set_bit(i, cci_pmu->hw_events.used_mask, cci_pmu->num_cntrs) {
641 struct perf_event *event = cci_hw->events[i];
642
643 if (WARN_ON(!event))
644 continue;
645
646 /* Leave the events which are not counting */
647 if (event->hw.state & PERF_HES_STOPPED)
648 continue;
649 if (event->hw.state & PERF_HES_ARCH) {
650 set_bit(i, mask);
651 event->hw.state &= ~PERF_HES_ARCH;
652 }
653 }
654
655 pmu_write_counters(cci_pmu, mask);
656}
657
658/* Should be called with cci_pmu->hw_events->pmu_lock held */
659static void __cci_pmu_enable_nosync(struct cci_pmu *cci_pmu)
660{
661 u32 val;
662
663 /* Enable all the PMU counters. */
664 val = readl_relaxed(cci_pmu->ctrl_base + CCI_PMCR) | CCI_PMCR_CEN;
665 writel(val, cci_pmu->ctrl_base + CCI_PMCR);
666}
667
668/* Should be called with cci_pmu->hw_events->pmu_lock held */
669static void __cci_pmu_enable_sync(struct cci_pmu *cci_pmu)
670{
671 cci_pmu_sync_counters(cci_pmu);
672 __cci_pmu_enable_nosync(cci_pmu);
673}
674
675/* Should be called with cci_pmu->hw_events->pmu_lock held */
676static void __cci_pmu_disable(struct cci_pmu *cci_pmu)
677{
678 u32 val;
679
680 /* Disable all the PMU counters. */
681 val = readl_relaxed(cci_pmu->ctrl_base + CCI_PMCR) & ~CCI_PMCR_CEN;
682 writel(val, cci_pmu->ctrl_base + CCI_PMCR);
683}
684
685static ssize_t cci_pmu_format_show(struct device *dev,
686 struct device_attribute *attr, char *buf)
687{
688 struct dev_ext_attribute *eattr = container_of(attr,
689 struct dev_ext_attribute, attr);
690 return snprintf(buf, PAGE_SIZE, "%s\n", (char *)eattr->var);
691}
692
693static ssize_t cci_pmu_event_show(struct device *dev,
694 struct device_attribute *attr, char *buf)
695{
696 struct dev_ext_attribute *eattr = container_of(attr,
697 struct dev_ext_attribute, attr);
698 /* source parameter is mandatory for normal PMU events */
699 return snprintf(buf, PAGE_SIZE, "source=?,event=0x%lx\n",
700 (unsigned long)eattr->var);
701}
702
703static int pmu_is_valid_counter(struct cci_pmu *cci_pmu, int idx)
704{
705 return 0 <= idx && idx <= CCI_PMU_CNTR_LAST(cci_pmu);
706}
707
708static u32 pmu_read_register(struct cci_pmu *cci_pmu, int idx, unsigned int offset)
709{
710 return readl_relaxed(cci_pmu->base +
711 CCI_PMU_CNTR_BASE(cci_pmu->model, idx) + offset);
712}
713
714static void pmu_write_register(struct cci_pmu *cci_pmu, u32 value,
715 int idx, unsigned int offset)
716{
717 writel_relaxed(value, cci_pmu->base +
718 CCI_PMU_CNTR_BASE(cci_pmu->model, idx) + offset);
719}
720
721static void pmu_disable_counter(struct cci_pmu *cci_pmu, int idx)
722{
723 pmu_write_register(cci_pmu, 0, idx, CCI_PMU_CNTR_CTRL);
724}
725
726static void pmu_enable_counter(struct cci_pmu *cci_pmu, int idx)
727{
728 pmu_write_register(cci_pmu, 1, idx, CCI_PMU_CNTR_CTRL);
729}
730
731static bool __maybe_unused
732pmu_counter_is_enabled(struct cci_pmu *cci_pmu, int idx)
733{
734 return (pmu_read_register(cci_pmu, idx, CCI_PMU_CNTR_CTRL) & 0x1) != 0;
735}
736
737static void pmu_set_event(struct cci_pmu *cci_pmu, int idx, unsigned long event)
738{
739 pmu_write_register(cci_pmu, event, idx, CCI_PMU_EVT_SEL);
740}
741
742/*
743 * For all counters on the CCI-PMU, disable any 'enabled' counters,
744 * saving the changed counters in the mask, so that we can restore
745 * it later using pmu_restore_counters. The mask is private to the
746 * caller. We cannot rely on the used_mask maintained by the CCI_PMU
747 * as it only tells us if the counter is assigned to perf_event or not.
748 * The state of the perf_event cannot be locked by the PMU layer, hence
749 * we check the individual counter status (which can be locked by
750 * cci_pm->hw_events->pmu_lock).
751 *
752 * @mask should be initialised to empty by the caller.
753 */
754static void __maybe_unused
755pmu_save_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
756{
757 int i;
758
759 for (i = 0; i < cci_pmu->num_cntrs; i++) {
760 if (pmu_counter_is_enabled(cci_pmu, i)) {
761 set_bit(i, mask);
762 pmu_disable_counter(cci_pmu, i);
763 }
764 }
765}
766
767/*
768 * Restore the status of the counters. Reversal of the pmu_save_counters().
769 * For each counter set in the mask, enable the counter back.
770 */
771static void __maybe_unused
772pmu_restore_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
773{
774 int i;
775
776 for_each_set_bit(i, mask, cci_pmu->num_cntrs)
777 pmu_enable_counter(cci_pmu, i);
778}
779
780/*
781 * Returns the number of programmable counters actually implemented
782 * by the cci
783 */
784static u32 pmu_get_max_counters(struct cci_pmu *cci_pmu)
785{
786 return (readl_relaxed(cci_pmu->ctrl_base + CCI_PMCR) &
787 CCI_PMCR_NCNT_MASK) >> CCI_PMCR_NCNT_SHIFT;
788}
789
790static int pmu_get_event_idx(struct cci_pmu_hw_events *hw, struct perf_event *event)
791{
792 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
793 unsigned long cci_event = event->hw.config_base;
794 int idx;
795
796 if (cci_pmu->model->get_event_idx)
797 return cci_pmu->model->get_event_idx(cci_pmu, hw, cci_event);
798
799 /* Generic code to find an unused idx from the mask */
800 for(idx = 0; idx <= CCI_PMU_CNTR_LAST(cci_pmu); idx++)
801 if (!test_and_set_bit(idx, hw->used_mask))
802 return idx;
803
804 /* No counters available */
805 return -EAGAIN;
806}
807
808static int pmu_map_event(struct perf_event *event)
809{
810 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
811
812 if (event->attr.type < PERF_TYPE_MAX ||
813 !cci_pmu->model->validate_hw_event)
814 return -ENOENT;
815
816 return cci_pmu->model->validate_hw_event(cci_pmu, event->attr.config);
817}
818
819static int pmu_request_irq(struct cci_pmu *cci_pmu, irq_handler_t handler)
820{
821 int i;
822 struct platform_device *pmu_device = cci_pmu->plat_device;
823
824 if (unlikely(!pmu_device))
825 return -ENODEV;
826
827 if (cci_pmu->nr_irqs < 1) {
828 dev_err(&pmu_device->dev, "no irqs for CCI PMUs defined\n");
829 return -ENODEV;
830 }
831
832 /*
833 * Register all available CCI PMU interrupts. In the interrupt handler
834 * we iterate over the counters checking for interrupt source (the
835 * overflowing counter) and clear it.
836 *
837 * This should allow handling of non-unique interrupt for the counters.
838 */
839 for (i = 0; i < cci_pmu->nr_irqs; i++) {
840 int err = request_irq(cci_pmu->irqs[i], handler, IRQF_SHARED,
841 "arm-cci-pmu", cci_pmu);
842 if (err) {
843 dev_err(&pmu_device->dev, "unable to request IRQ%d for ARM CCI PMU counters\n",
844 cci_pmu->irqs[i]);
845 return err;
846 }
847
848 set_bit(i, &cci_pmu->active_irqs);
849 }
850
851 return 0;
852}
853
854static void pmu_free_irq(struct cci_pmu *cci_pmu)
855{
856 int i;
857
858 for (i = 0; i < cci_pmu->nr_irqs; i++) {
859 if (!test_and_clear_bit(i, &cci_pmu->active_irqs))
860 continue;
861
862 free_irq(cci_pmu->irqs[i], cci_pmu);
863 }
864}
865
866static u32 pmu_read_counter(struct perf_event *event)
867{
868 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
869 struct hw_perf_event *hw_counter = &event->hw;
870 int idx = hw_counter->idx;
871 u32 value;
872
873 if (unlikely(!pmu_is_valid_counter(cci_pmu, idx))) {
874 dev_err(&cci_pmu->plat_device->dev, "Invalid CCI PMU counter %d\n", idx);
875 return 0;
876 }
877 value = pmu_read_register(cci_pmu, idx, CCI_PMU_CNTR);
878
879 return value;
880}
881
882static void pmu_write_counter(struct cci_pmu *cci_pmu, u32 value, int idx)
883{
884 pmu_write_register(cci_pmu, value, idx, CCI_PMU_CNTR);
885}
886
887static void __pmu_write_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
888{
889 int i;
890 struct cci_pmu_hw_events *cci_hw = &cci_pmu->hw_events;
891
892 for_each_set_bit(i, mask, cci_pmu->num_cntrs) {
893 struct perf_event *event = cci_hw->events[i];
894
895 if (WARN_ON(!event))
896 continue;
897 pmu_write_counter(cci_pmu, local64_read(&event->hw.prev_count), i);
898 }
899}
900
901static void pmu_write_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
902{
903 if (cci_pmu->model->write_counters)
904 cci_pmu->model->write_counters(cci_pmu, mask);
905 else
906 __pmu_write_counters(cci_pmu, mask);
907}
908
909#ifdef CONFIG_ARM_CCI5xx_PMU
910
911/*
912 * CCI-500/CCI-550 has advanced power saving policies, which could gate the
913 * clocks to the PMU counters, which makes the writes to them ineffective.
914 * The only way to write to those counters is when the global counters
915 * are enabled and the particular counter is enabled.
916 *
917 * So we do the following :
918 *
919 * 1) Disable all the PMU counters, saving their current state
920 * 2) Enable the global PMU profiling, now that all counters are
921 * disabled.
922 *
923 * For each counter to be programmed, repeat steps 3-7:
924 *
925 * 3) Write an invalid event code to the event control register for the
926 counter, so that the counters are not modified.
927 * 4) Enable the counter control for the counter.
928 * 5) Set the counter value
929 * 6) Disable the counter
930 * 7) Restore the event in the target counter
931 *
932 * 8) Disable the global PMU.
933 * 9) Restore the status of the rest of the counters.
934 *
935 * We choose an event which for CCI-5xx is guaranteed not to count.
936 * We use the highest possible event code (0x1f) for the master interface 0.
937 */
938#define CCI5xx_INVALID_EVENT ((CCI5xx_PORT_M0 << CCI5xx_PMU_EVENT_SOURCE_SHIFT) | \
939 (CCI5xx_PMU_EVENT_CODE_MASK << CCI5xx_PMU_EVENT_CODE_SHIFT))
940static void cci5xx_pmu_write_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
941{
942 int i;
943 DECLARE_BITMAP(saved_mask, cci_pmu->num_cntrs);
944
945 bitmap_zero(saved_mask, cci_pmu->num_cntrs);
946 pmu_save_counters(cci_pmu, saved_mask);
947
948 /*
949 * Now that all the counters are disabled, we can safely turn the PMU on,
950 * without syncing the status of the counters
951 */
952 __cci_pmu_enable_nosync(cci_pmu);
953
954 for_each_set_bit(i, mask, cci_pmu->num_cntrs) {
955 struct perf_event *event = cci_pmu->hw_events.events[i];
956
957 if (WARN_ON(!event))
958 continue;
959
960 pmu_set_event(cci_pmu, i, CCI5xx_INVALID_EVENT);
961 pmu_enable_counter(cci_pmu, i);
962 pmu_write_counter(cci_pmu, local64_read(&event->hw.prev_count), i);
963 pmu_disable_counter(cci_pmu, i);
964 pmu_set_event(cci_pmu, i, event->hw.config_base);
965 }
966
967 __cci_pmu_disable(cci_pmu);
968
969 pmu_restore_counters(cci_pmu, saved_mask);
970}
971
972#endif /* CONFIG_ARM_CCI5xx_PMU */
973
974static u64 pmu_event_update(struct perf_event *event)
975{
976 struct hw_perf_event *hwc = &event->hw;
977 u64 delta, prev_raw_count, new_raw_count;
978
979 do {
980 prev_raw_count = local64_read(&hwc->prev_count);
981 new_raw_count = pmu_read_counter(event);
982 } while (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
983 new_raw_count) != prev_raw_count);
984
985 delta = (new_raw_count - prev_raw_count) & CCI_PMU_CNTR_MASK;
986
987 local64_add(delta, &event->count);
988
989 return new_raw_count;
990}
991
992static void pmu_read(struct perf_event *event)
993{
994 pmu_event_update(event);
995}
996
997static void pmu_event_set_period(struct perf_event *event)
998{
999 struct hw_perf_event *hwc = &event->hw;
1000 /*
1001 * The CCI PMU counters have a period of 2^32. To account for the
1002 * possiblity of extreme interrupt latency we program for a period of
1003 * half that. Hopefully we can handle the interrupt before another 2^31
1004 * events occur and the counter overtakes its previous value.
1005 */
1006 u64 val = 1ULL << 31;
1007 local64_set(&hwc->prev_count, val);
1008
1009 /*
1010 * CCI PMU uses PERF_HES_ARCH to keep track of the counters, whose
1011 * values needs to be sync-ed with the s/w state before the PMU is
1012 * enabled.
1013 * Mark this counter for sync.
1014 */
1015 hwc->state |= PERF_HES_ARCH;
1016}
1017
1018static irqreturn_t pmu_handle_irq(int irq_num, void *dev)
1019{
1020 unsigned long flags;
1021 struct cci_pmu *cci_pmu = dev;
1022 struct cci_pmu_hw_events *events = &cci_pmu->hw_events;
1023 int idx, handled = IRQ_NONE;
1024
1025 raw_spin_lock_irqsave(&events->pmu_lock, flags);
1026
1027 /* Disable the PMU while we walk through the counters */
1028 __cci_pmu_disable(cci_pmu);
1029 /*
1030 * Iterate over counters and update the corresponding perf events.
1031 * This should work regardless of whether we have per-counter overflow
1032 * interrupt or a combined overflow interrupt.
1033 */
1034 for (idx = 0; idx <= CCI_PMU_CNTR_LAST(cci_pmu); idx++) {
1035 struct perf_event *event = events->events[idx];
1036
1037 if (!event)
1038 continue;
1039
1040 /* Did this counter overflow? */
1041 if (!(pmu_read_register(cci_pmu, idx, CCI_PMU_OVRFLW) &
1042 CCI_PMU_OVRFLW_FLAG))
1043 continue;
1044
1045 pmu_write_register(cci_pmu, CCI_PMU_OVRFLW_FLAG, idx,
1046 CCI_PMU_OVRFLW);
1047
1048 pmu_event_update(event);
1049 pmu_event_set_period(event);
1050 handled = IRQ_HANDLED;
1051 }
1052
1053 /* Enable the PMU and sync possibly overflowed counters */
1054 __cci_pmu_enable_sync(cci_pmu);
1055 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
1056
1057 return IRQ_RETVAL(handled);
1058}
1059
1060static int cci_pmu_get_hw(struct cci_pmu *cci_pmu)
1061{
1062 int ret = pmu_request_irq(cci_pmu, pmu_handle_irq);
1063 if (ret) {
1064 pmu_free_irq(cci_pmu);
1065 return ret;
1066 }
1067 return 0;
1068}
1069
1070static void cci_pmu_put_hw(struct cci_pmu *cci_pmu)
1071{
1072 pmu_free_irq(cci_pmu);
1073}
1074
1075static void hw_perf_event_destroy(struct perf_event *event)
1076{
1077 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1078 atomic_t *active_events = &cci_pmu->active_events;
1079 struct mutex *reserve_mutex = &cci_pmu->reserve_mutex;
1080
1081 if (atomic_dec_and_mutex_lock(active_events, reserve_mutex)) {
1082 cci_pmu_put_hw(cci_pmu);
1083 mutex_unlock(reserve_mutex);
1084 }
1085}
1086
1087static void cci_pmu_enable(struct pmu *pmu)
1088{
1089 struct cci_pmu *cci_pmu = to_cci_pmu(pmu);
1090 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1091 int enabled = bitmap_weight(hw_events->used_mask, cci_pmu->num_cntrs);
1092 unsigned long flags;
1093
1094 if (!enabled)
1095 return;
1096
1097 raw_spin_lock_irqsave(&hw_events->pmu_lock, flags);
1098 __cci_pmu_enable_sync(cci_pmu);
1099 raw_spin_unlock_irqrestore(&hw_events->pmu_lock, flags);
1100
1101}
1102
1103static void cci_pmu_disable(struct pmu *pmu)
1104{
1105 struct cci_pmu *cci_pmu = to_cci_pmu(pmu);
1106 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1107 unsigned long flags;
1108
1109 raw_spin_lock_irqsave(&hw_events->pmu_lock, flags);
1110 __cci_pmu_disable(cci_pmu);
1111 raw_spin_unlock_irqrestore(&hw_events->pmu_lock, flags);
1112}
1113
1114/*
1115 * Check if the idx represents a non-programmable counter.
1116 * All the fixed event counters are mapped before the programmable
1117 * counters.
1118 */
1119static bool pmu_fixed_hw_idx(struct cci_pmu *cci_pmu, int idx)
1120{
1121 return (idx >= 0) && (idx < cci_pmu->model->fixed_hw_cntrs);
1122}
1123
1124static void cci_pmu_start(struct perf_event *event, int pmu_flags)
1125{
1126 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1127 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1128 struct hw_perf_event *hwc = &event->hw;
1129 int idx = hwc->idx;
1130 unsigned long flags;
1131
1132 /*
1133 * To handle interrupt latency, we always reprogram the period
1134 * regardlesss of PERF_EF_RELOAD.
1135 */
1136 if (pmu_flags & PERF_EF_RELOAD)
1137 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
1138
1139 hwc->state = 0;
1140
1141 if (unlikely(!pmu_is_valid_counter(cci_pmu, idx))) {
1142 dev_err(&cci_pmu->plat_device->dev, "Invalid CCI PMU counter %d\n", idx);
1143 return;
1144 }
1145
1146 raw_spin_lock_irqsave(&hw_events->pmu_lock, flags);
1147
1148 /* Configure the counter unless you are counting a fixed event */
1149 if (!pmu_fixed_hw_idx(cci_pmu, idx))
1150 pmu_set_event(cci_pmu, idx, hwc->config_base);
1151
1152 pmu_event_set_period(event);
1153 pmu_enable_counter(cci_pmu, idx);
1154
1155 raw_spin_unlock_irqrestore(&hw_events->pmu_lock, flags);
1156}
1157
1158static void cci_pmu_stop(struct perf_event *event, int pmu_flags)
1159{
1160 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1161 struct hw_perf_event *hwc = &event->hw;
1162 int idx = hwc->idx;
1163
1164 if (hwc->state & PERF_HES_STOPPED)
1165 return;
1166
1167 if (unlikely(!pmu_is_valid_counter(cci_pmu, idx))) {
1168 dev_err(&cci_pmu->plat_device->dev, "Invalid CCI PMU counter %d\n", idx);
1169 return;
1170 }
1171
1172 /*
1173 * We always reprogram the counter, so ignore PERF_EF_UPDATE. See
1174 * cci_pmu_start()
1175 */
1176 pmu_disable_counter(cci_pmu, idx);
1177 pmu_event_update(event);
1178 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
1179}
1180
1181static int cci_pmu_add(struct perf_event *event, int flags)
1182{
1183 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1184 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1185 struct hw_perf_event *hwc = &event->hw;
1186 int idx;
1187 int err = 0;
1188
1189 perf_pmu_disable(event->pmu);
1190
1191 /* If we don't have a space for the counter then finish early. */
1192 idx = pmu_get_event_idx(hw_events, event);
1193 if (idx < 0) {
1194 err = idx;
1195 goto out;
1196 }
1197
1198 event->hw.idx = idx;
1199 hw_events->events[idx] = event;
1200
1201 hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
1202 if (flags & PERF_EF_START)
1203 cci_pmu_start(event, PERF_EF_RELOAD);
1204
1205 /* Propagate our changes to the userspace mapping. */
1206 perf_event_update_userpage(event);
1207
1208out:
1209 perf_pmu_enable(event->pmu);
1210 return err;
1211}
1212
1213static void cci_pmu_del(struct perf_event *event, int flags)
1214{
1215 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1216 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1217 struct hw_perf_event *hwc = &event->hw;
1218 int idx = hwc->idx;
1219
1220 cci_pmu_stop(event, PERF_EF_UPDATE);
1221 hw_events->events[idx] = NULL;
1222 clear_bit(idx, hw_events->used_mask);
1223
1224 perf_event_update_userpage(event);
1225}
1226
1227static int validate_event(struct pmu *cci_pmu,
1228 struct cci_pmu_hw_events *hw_events,
1229 struct perf_event *event)
1230{
1231 if (is_software_event(event))
1232 return 1;
1233
1234 /*
1235 * Reject groups spanning multiple HW PMUs (e.g. CPU + CCI). The
1236 * core perf code won't check that the pmu->ctx == leader->ctx
1237 * until after pmu->event_init(event).
1238 */
1239 if (event->pmu != cci_pmu)
1240 return 0;
1241
1242 if (event->state < PERF_EVENT_STATE_OFF)
1243 return 1;
1244
1245 if (event->state == PERF_EVENT_STATE_OFF && !event->attr.enable_on_exec)
1246 return 1;
1247
1248 return pmu_get_event_idx(hw_events, event) >= 0;
1249}
1250
1251static int validate_group(struct perf_event *event)
1252{
1253 struct perf_event *sibling, *leader = event->group_leader;
1254 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1255 unsigned long mask[BITS_TO_LONGS(cci_pmu->num_cntrs)];
1256 struct cci_pmu_hw_events fake_pmu = {
1257 /*
1258 * Initialise the fake PMU. We only need to populate the
1259 * used_mask for the purposes of validation.
1260 */
1261 .used_mask = mask,
1262 };
1263 memset(mask, 0, BITS_TO_LONGS(cci_pmu->num_cntrs) * sizeof(unsigned long));
1264
1265 if (!validate_event(event->pmu, &fake_pmu, leader))
1266 return -EINVAL;
1267
1268 for_each_sibling_event(sibling, leader) {
1269 if (!validate_event(event->pmu, &fake_pmu, sibling))
1270 return -EINVAL;
1271 }
1272
1273 if (!validate_event(event->pmu, &fake_pmu, event))
1274 return -EINVAL;
1275
1276 return 0;
1277}
1278
1279static int __hw_perf_event_init(struct perf_event *event)
1280{
1281 struct hw_perf_event *hwc = &event->hw;
1282 int mapping;
1283
1284 mapping = pmu_map_event(event);
1285
1286 if (mapping < 0) {
1287 pr_debug("event %x:%llx not supported\n", event->attr.type,
1288 event->attr.config);
1289 return mapping;
1290 }
1291
1292 /*
1293 * We don't assign an index until we actually place the event onto
1294 * hardware. Use -1 to signify that we haven't decided where to put it
1295 * yet.
1296 */
1297 hwc->idx = -1;
1298 hwc->config_base = 0;
1299 hwc->config = 0;
1300 hwc->event_base = 0;
1301
1302 /*
1303 * Store the event encoding into the config_base field.
1304 */
1305 hwc->config_base |= (unsigned long)mapping;
1306
1307 /*
1308 * Limit the sample_period to half of the counter width. That way, the
1309 * new counter value is far less likely to overtake the previous one
1310 * unless you have some serious IRQ latency issues.
1311 */
1312 hwc->sample_period = CCI_PMU_CNTR_MASK >> 1;
1313 hwc->last_period = hwc->sample_period;
1314 local64_set(&hwc->period_left, hwc->sample_period);
1315
1316 if (event->group_leader != event) {
1317 if (validate_group(event) != 0)
1318 return -EINVAL;
1319 }
1320
1321 return 0;
1322}
1323
1324static int cci_pmu_event_init(struct perf_event *event)
1325{
1326 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1327 atomic_t *active_events = &cci_pmu->active_events;
1328 int err = 0;
1329
1330 if (event->attr.type != event->pmu->type)
1331 return -ENOENT;
1332
1333 /* Shared by all CPUs, no meaningful state to sample */
1334 if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
1335 return -EOPNOTSUPP;
1336
1337 /* We have no filtering of any kind */
1338 if (event->attr.exclude_user ||
1339 event->attr.exclude_kernel ||
1340 event->attr.exclude_hv ||
1341 event->attr.exclude_idle ||
1342 event->attr.exclude_host ||
1343 event->attr.exclude_guest)
1344 return -EINVAL;
1345
1346 /*
1347 * Following the example set by other "uncore" PMUs, we accept any CPU
1348 * and rewrite its affinity dynamically rather than having perf core
1349 * handle cpu == -1 and pid == -1 for this case.
1350 *
1351 * The perf core will pin online CPUs for the duration of this call and
1352 * the event being installed into its context, so the PMU's CPU can't
1353 * change under our feet.
1354 */
1355 if (event->cpu < 0)
1356 return -EINVAL;
1357 event->cpu = cci_pmu->cpu;
1358
1359 event->destroy = hw_perf_event_destroy;
1360 if (!atomic_inc_not_zero(active_events)) {
1361 mutex_lock(&cci_pmu->reserve_mutex);
1362 if (atomic_read(active_events) == 0)
1363 err = cci_pmu_get_hw(cci_pmu);
1364 if (!err)
1365 atomic_inc(active_events);
1366 mutex_unlock(&cci_pmu->reserve_mutex);
1367 }
1368 if (err)
1369 return err;
1370
1371 err = __hw_perf_event_init(event);
1372 if (err)
1373 hw_perf_event_destroy(event);
1374
1375 return err;
1376}
1377
1378static ssize_t pmu_cpumask_attr_show(struct device *dev,
1379 struct device_attribute *attr, char *buf)
1380{
1381 struct pmu *pmu = dev_get_drvdata(dev);
1382 struct cci_pmu *cci_pmu = to_cci_pmu(pmu);
1383
1384 return cpumap_print_to_pagebuf(true, buf, cpumask_of(cci_pmu->cpu));
1385}
1386
1387static struct device_attribute pmu_cpumask_attr =
1388 __ATTR(cpumask, S_IRUGO, pmu_cpumask_attr_show, NULL);
1389
1390static struct attribute *pmu_attrs[] = {
1391 &pmu_cpumask_attr.attr,
1392 NULL,
1393};
1394
1395static struct attribute_group pmu_attr_group = {
1396 .attrs = pmu_attrs,
1397};
1398
1399static struct attribute_group pmu_format_attr_group = {
1400 .name = "format",
1401 .attrs = NULL, /* Filled in cci_pmu_init_attrs */
1402};
1403
1404static struct attribute_group pmu_event_attr_group = {
1405 .name = "events",
1406 .attrs = NULL, /* Filled in cci_pmu_init_attrs */
1407};
1408
1409static const struct attribute_group *pmu_attr_groups[] = {
1410 &pmu_attr_group,
1411 &pmu_format_attr_group,
1412 &pmu_event_attr_group,
1413 NULL
1414};
1415
1416static int cci_pmu_init(struct cci_pmu *cci_pmu, struct platform_device *pdev)
1417{
1418 const struct cci_pmu_model *model = cci_pmu->model;
1419 char *name = model->name;
1420 u32 num_cntrs;
1421
1422 pmu_event_attr_group.attrs = model->event_attrs;
1423 pmu_format_attr_group.attrs = model->format_attrs;
1424
1425 cci_pmu->pmu = (struct pmu) {
1426 .name = cci_pmu->model->name,
1427 .task_ctx_nr = perf_invalid_context,
1428 .pmu_enable = cci_pmu_enable,
1429 .pmu_disable = cci_pmu_disable,
1430 .event_init = cci_pmu_event_init,
1431 .add = cci_pmu_add,
1432 .del = cci_pmu_del,
1433 .start = cci_pmu_start,
1434 .stop = cci_pmu_stop,
1435 .read = pmu_read,
1436 .attr_groups = pmu_attr_groups,
1437 };
1438
1439 cci_pmu->plat_device = pdev;
1440 num_cntrs = pmu_get_max_counters(cci_pmu);
1441 if (num_cntrs > cci_pmu->model->num_hw_cntrs) {
1442 dev_warn(&pdev->dev,
1443 "PMU implements more counters(%d) than supported by"
1444 " the model(%d), truncated.",
1445 num_cntrs, cci_pmu->model->num_hw_cntrs);
1446 num_cntrs = cci_pmu->model->num_hw_cntrs;
1447 }
1448 cci_pmu->num_cntrs = num_cntrs + cci_pmu->model->fixed_hw_cntrs;
1449
1450 return perf_pmu_register(&cci_pmu->pmu, name, -1);
1451}
1452
1453static int cci_pmu_offline_cpu(unsigned int cpu)
1454{
1455 int target;
1456
1457 if (!g_cci_pmu || cpu != g_cci_pmu->cpu)
1458 return 0;
1459
1460 target = cpumask_any_but(cpu_online_mask, cpu);
1461 if (target >= nr_cpu_ids)
1462 return 0;
1463
1464 perf_pmu_migrate_context(&g_cci_pmu->pmu, cpu, target);
1465 g_cci_pmu->cpu = target;
1466 return 0;
1467}
1468
1469static struct cci_pmu_model cci_pmu_models[] = {
1470#ifdef CONFIG_ARM_CCI400_PMU
1471 [CCI400_R0] = {
1472 .name = "CCI_400",
1473 .fixed_hw_cntrs = 1, /* Cycle counter */
1474 .num_hw_cntrs = 4,
1475 .cntr_size = SZ_4K,
1476 .format_attrs = cci400_pmu_format_attrs,
1477 .event_attrs = cci400_r0_pmu_event_attrs,
1478 .event_ranges = {
1479 [CCI_IF_SLAVE] = {
1480 CCI400_R0_SLAVE_PORT_MIN_EV,
1481 CCI400_R0_SLAVE_PORT_MAX_EV,
1482 },
1483 [CCI_IF_MASTER] = {
1484 CCI400_R0_MASTER_PORT_MIN_EV,
1485 CCI400_R0_MASTER_PORT_MAX_EV,
1486 },
1487 },
1488 .validate_hw_event = cci400_validate_hw_event,
1489 .get_event_idx = cci400_get_event_idx,
1490 },
1491 [CCI400_R1] = {
1492 .name = "CCI_400_r1",
1493 .fixed_hw_cntrs = 1, /* Cycle counter */
1494 .num_hw_cntrs = 4,
1495 .cntr_size = SZ_4K,
1496 .format_attrs = cci400_pmu_format_attrs,
1497 .event_attrs = cci400_r1_pmu_event_attrs,
1498 .event_ranges = {
1499 [CCI_IF_SLAVE] = {
1500 CCI400_R1_SLAVE_PORT_MIN_EV,
1501 CCI400_R1_SLAVE_PORT_MAX_EV,
1502 },
1503 [CCI_IF_MASTER] = {
1504 CCI400_R1_MASTER_PORT_MIN_EV,
1505 CCI400_R1_MASTER_PORT_MAX_EV,
1506 },
1507 },
1508 .validate_hw_event = cci400_validate_hw_event,
1509 .get_event_idx = cci400_get_event_idx,
1510 },
1511#endif
1512#ifdef CONFIG_ARM_CCI5xx_PMU
1513 [CCI500_R0] = {
1514 .name = "CCI_500",
1515 .fixed_hw_cntrs = 0,
1516 .num_hw_cntrs = 8,
1517 .cntr_size = SZ_64K,
1518 .format_attrs = cci5xx_pmu_format_attrs,
1519 .event_attrs = cci5xx_pmu_event_attrs,
1520 .event_ranges = {
1521 [CCI_IF_SLAVE] = {
1522 CCI5xx_SLAVE_PORT_MIN_EV,
1523 CCI5xx_SLAVE_PORT_MAX_EV,
1524 },
1525 [CCI_IF_MASTER] = {
1526 CCI5xx_MASTER_PORT_MIN_EV,
1527 CCI5xx_MASTER_PORT_MAX_EV,
1528 },
1529 [CCI_IF_GLOBAL] = {
1530 CCI5xx_GLOBAL_PORT_MIN_EV,
1531 CCI5xx_GLOBAL_PORT_MAX_EV,
1532 },
1533 },
1534 .validate_hw_event = cci500_validate_hw_event,
1535 .write_counters = cci5xx_pmu_write_counters,
1536 },
1537 [CCI550_R0] = {
1538 .name = "CCI_550",
1539 .fixed_hw_cntrs = 0,
1540 .num_hw_cntrs = 8,
1541 .cntr_size = SZ_64K,
1542 .format_attrs = cci5xx_pmu_format_attrs,
1543 .event_attrs = cci5xx_pmu_event_attrs,
1544 .event_ranges = {
1545 [CCI_IF_SLAVE] = {
1546 CCI5xx_SLAVE_PORT_MIN_EV,
1547 CCI5xx_SLAVE_PORT_MAX_EV,
1548 },
1549 [CCI_IF_MASTER] = {
1550 CCI5xx_MASTER_PORT_MIN_EV,
1551 CCI5xx_MASTER_PORT_MAX_EV,
1552 },
1553 [CCI_IF_GLOBAL] = {
1554 CCI5xx_GLOBAL_PORT_MIN_EV,
1555 CCI5xx_GLOBAL_PORT_MAX_EV,
1556 },
1557 },
1558 .validate_hw_event = cci550_validate_hw_event,
1559 .write_counters = cci5xx_pmu_write_counters,
1560 },
1561#endif
1562};
1563
1564static const struct of_device_id arm_cci_pmu_matches[] = {
1565#ifdef CONFIG_ARM_CCI400_PMU
1566 {
1567 .compatible = "arm,cci-400-pmu",
1568 .data = NULL,
1569 },
1570 {
1571 .compatible = "arm,cci-400-pmu,r0",
1572 .data = &cci_pmu_models[CCI400_R0],
1573 },
1574 {
1575 .compatible = "arm,cci-400-pmu,r1",
1576 .data = &cci_pmu_models[CCI400_R1],
1577 },
1578#endif
1579#ifdef CONFIG_ARM_CCI5xx_PMU
1580 {
1581 .compatible = "arm,cci-500-pmu,r0",
1582 .data = &cci_pmu_models[CCI500_R0],
1583 },
1584 {
1585 .compatible = "arm,cci-550-pmu,r0",
1586 .data = &cci_pmu_models[CCI550_R0],
1587 },
1588#endif
1589 {},
1590};
1591
1592static bool is_duplicate_irq(int irq, int *irqs, int nr_irqs)
1593{
1594 int i;
1595
1596 for (i = 0; i < nr_irqs; i++)
1597 if (irq == irqs[i])
1598 return true;
1599
1600 return false;
1601}
1602
1603static struct cci_pmu *cci_pmu_alloc(struct device *dev)
1604{
1605 struct cci_pmu *cci_pmu;
1606 const struct cci_pmu_model *model;
1607
1608 /*
1609 * All allocations are devm_* hence we don't have to free
1610 * them explicitly on an error, as it would end up in driver
1611 * detach.
1612 */
1613 cci_pmu = devm_kzalloc(dev, sizeof(*cci_pmu), GFP_KERNEL);
1614 if (!cci_pmu)
1615 return ERR_PTR(-ENOMEM);
1616
1617 cci_pmu->ctrl_base = *(void __iomem **)dev->platform_data;
1618
1619 model = of_device_get_match_data(dev);
1620 if (!model) {
1621 dev_warn(dev,
1622 "DEPRECATED compatible property, requires secure access to CCI registers");
1623 model = probe_cci_model(cci_pmu);
1624 }
1625 if (!model) {
1626 dev_warn(dev, "CCI PMU version not supported\n");
1627 return ERR_PTR(-ENODEV);
1628 }
1629
1630 cci_pmu->model = model;
1631 cci_pmu->irqs = devm_kcalloc(dev, CCI_PMU_MAX_HW_CNTRS(model),
1632 sizeof(*cci_pmu->irqs), GFP_KERNEL);
1633 if (!cci_pmu->irqs)
1634 return ERR_PTR(-ENOMEM);
1635 cci_pmu->hw_events.events = devm_kcalloc(dev,
1636 CCI_PMU_MAX_HW_CNTRS(model),
1637 sizeof(*cci_pmu->hw_events.events),
1638 GFP_KERNEL);
1639 if (!cci_pmu->hw_events.events)
1640 return ERR_PTR(-ENOMEM);
1641 cci_pmu->hw_events.used_mask = devm_kcalloc(dev,
1642 BITS_TO_LONGS(CCI_PMU_MAX_HW_CNTRS(model)),
1643 sizeof(*cci_pmu->hw_events.used_mask),
1644 GFP_KERNEL);
1645 if (!cci_pmu->hw_events.used_mask)
1646 return ERR_PTR(-ENOMEM);
1647
1648 return cci_pmu;
1649}
1650
1651static int cci_pmu_probe(struct platform_device *pdev)
1652{
1653 struct resource *res;
1654 struct cci_pmu *cci_pmu;
1655 int i, ret, irq;
1656
1657 cci_pmu = cci_pmu_alloc(&pdev->dev);
1658 if (IS_ERR(cci_pmu))
1659 return PTR_ERR(cci_pmu);
1660
1661 res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
1662 cci_pmu->base = devm_ioremap_resource(&pdev->dev, res);
1663 if (IS_ERR(cci_pmu->base))
1664 return -ENOMEM;
1665
1666 /*
1667 * CCI PMU has one overflow interrupt per counter; but some may be tied
1668 * together to a common interrupt.
1669 */
1670 cci_pmu->nr_irqs = 0;
1671 for (i = 0; i < CCI_PMU_MAX_HW_CNTRS(cci_pmu->model); i++) {
1672 irq = platform_get_irq(pdev, i);
1673 if (irq < 0)
1674 break;
1675
1676 if (is_duplicate_irq(irq, cci_pmu->irqs, cci_pmu->nr_irqs))
1677 continue;
1678
1679 cci_pmu->irqs[cci_pmu->nr_irqs++] = irq;
1680 }
1681
1682 /*
1683 * Ensure that the device tree has as many interrupts as the number
1684 * of counters.
1685 */
1686 if (i < CCI_PMU_MAX_HW_CNTRS(cci_pmu->model)) {
1687 dev_warn(&pdev->dev, "In-correct number of interrupts: %d, should be %d\n",
1688 i, CCI_PMU_MAX_HW_CNTRS(cci_pmu->model));
1689 return -EINVAL;
1690 }
1691
1692 raw_spin_lock_init(&cci_pmu->hw_events.pmu_lock);
1693 mutex_init(&cci_pmu->reserve_mutex);
1694 atomic_set(&cci_pmu->active_events, 0);
1695 cci_pmu->cpu = get_cpu();
1696
1697 ret = cci_pmu_init(cci_pmu, pdev);
1698 if (ret) {
1699 put_cpu();
1700 return ret;
1701 }
1702
1703 cpuhp_setup_state_nocalls(CPUHP_AP_PERF_ARM_CCI_ONLINE,
1704 "perf/arm/cci:online", NULL,
1705 cci_pmu_offline_cpu);
1706 put_cpu();
1707 g_cci_pmu = cci_pmu;
1708 pr_info("ARM %s PMU driver probed", cci_pmu->model->name);
1709 return 0;
1710}
1711
1712static struct platform_driver cci_pmu_driver = {
1713 .driver = {
1714 .name = DRIVER_NAME,
1715 .of_match_table = arm_cci_pmu_matches,
1716 },
1717 .probe = cci_pmu_probe,
1718};
1719
1720builtin_platform_driver(cci_pmu_driver);
1721MODULE_LICENSE("GPL");
1722MODULE_DESCRIPTION("ARM CCI PMU support");