xref: /openbmc/linux/drivers/perf/arm-cci.c (revision a17922de)
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 */
47 enum {
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 struct event_range {
57 	u32 min;
58 	u32 max;
59 };
60 
61 struct cci_pmu_hw_events {
62 	struct perf_event **events;
63 	unsigned long *used_mask;
64 	raw_spinlock_t pmu_lock;
65 };
66 
67 struct 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  */
74 struct 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 
87 static struct cci_pmu_model cci_pmu_models[];
88 
89 struct 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 
107 static struct cci_pmu *g_cci_pmu;
108 
109 enum 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 
121 static void pmu_write_counters(struct cci_pmu *cci_pmu,
122 				 unsigned long *mask);
123 static ssize_t __maybe_unused cci_pmu_format_show(struct device *dev,
124 			struct device_attribute *attr, char *buf);
125 static ssize_t __maybe_unused 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  */
159 enum 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 
203 static ssize_t cci400_pmu_cycle_event_show(struct device *dev,
204 			struct device_attribute *attr, char *buf);
205 
206 static 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 
212 static 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 
247 static 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 
294 static 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 
302 static 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 
324 static 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 
362 static 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 
374 static 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 */
381 static 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 
439 static ssize_t cci5xx_pmu_global_event_show(struct device *dev,
440 				struct device_attribute *attr, char *buf);
441 
442 static 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 
448 static 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 
512 static 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  */
531 static 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  */
582 static 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  */
632 static 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 */
659 static 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 */
669 static 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 */
676 static 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 
685 static 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 
693 static 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 
703 static 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 
708 static 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 
714 static 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 
721 static 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 
726 static 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 
731 static bool __maybe_unused
732 pmu_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 
737 static 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  */
754 static void __maybe_unused
755 pmu_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  */
771 static void __maybe_unused
772 pmu_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  */
784 static 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 
790 static 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 
808 static 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 
819 static 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 
854 static 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 
866 static 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 
882 static 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 
887 static 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 
901 static 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))
940 static 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 
974 static 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 
992 static void pmu_read(struct perf_event *event)
993 {
994 	pmu_event_update(event);
995 }
996 
997 static 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 
1018 static 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 
1060 static 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 
1070 static void cci_pmu_put_hw(struct cci_pmu *cci_pmu)
1071 {
1072 	pmu_free_irq(cci_pmu);
1073 }
1074 
1075 static 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 
1087 static 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 
1103 static 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  */
1119 static 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 
1124 static 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 
1158 static 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 
1181 static 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 
1188 	/* If we don't have a space for the counter then finish early. */
1189 	idx = pmu_get_event_idx(hw_events, event);
1190 	if (idx < 0)
1191 		return idx;
1192 
1193 	event->hw.idx = idx;
1194 	hw_events->events[idx] = event;
1195 
1196 	hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
1197 	if (flags & PERF_EF_START)
1198 		cci_pmu_start(event, PERF_EF_RELOAD);
1199 
1200 	/* Propagate our changes to the userspace mapping. */
1201 	perf_event_update_userpage(event);
1202 
1203 	return 0;
1204 }
1205 
1206 static void cci_pmu_del(struct perf_event *event, int flags)
1207 {
1208 	struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1209 	struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1210 	struct hw_perf_event *hwc = &event->hw;
1211 	int idx = hwc->idx;
1212 
1213 	cci_pmu_stop(event, PERF_EF_UPDATE);
1214 	hw_events->events[idx] = NULL;
1215 	clear_bit(idx, hw_events->used_mask);
1216 
1217 	perf_event_update_userpage(event);
1218 }
1219 
1220 static int validate_event(struct pmu *cci_pmu,
1221 			  struct cci_pmu_hw_events *hw_events,
1222 			  struct perf_event *event)
1223 {
1224 	if (is_software_event(event))
1225 		return 1;
1226 
1227 	/*
1228 	 * Reject groups spanning multiple HW PMUs (e.g. CPU + CCI). The
1229 	 * core perf code won't check that the pmu->ctx == leader->ctx
1230 	 * until after pmu->event_init(event).
1231 	 */
1232 	if (event->pmu != cci_pmu)
1233 		return 0;
1234 
1235 	if (event->state < PERF_EVENT_STATE_OFF)
1236 		return 1;
1237 
1238 	if (event->state == PERF_EVENT_STATE_OFF && !event->attr.enable_on_exec)
1239 		return 1;
1240 
1241 	return pmu_get_event_idx(hw_events, event) >= 0;
1242 }
1243 
1244 static int validate_group(struct perf_event *event)
1245 {
1246 	struct perf_event *sibling, *leader = event->group_leader;
1247 	struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1248 	unsigned long mask[BITS_TO_LONGS(cci_pmu->num_cntrs)];
1249 	struct cci_pmu_hw_events fake_pmu = {
1250 		/*
1251 		 * Initialise the fake PMU. We only need to populate the
1252 		 * used_mask for the purposes of validation.
1253 		 */
1254 		.used_mask = mask,
1255 	};
1256 	memset(mask, 0, BITS_TO_LONGS(cci_pmu->num_cntrs) * sizeof(unsigned long));
1257 
1258 	if (!validate_event(event->pmu, &fake_pmu, leader))
1259 		return -EINVAL;
1260 
1261 	for_each_sibling_event(sibling, leader) {
1262 		if (!validate_event(event->pmu, &fake_pmu, sibling))
1263 			return -EINVAL;
1264 	}
1265 
1266 	if (!validate_event(event->pmu, &fake_pmu, event))
1267 		return -EINVAL;
1268 
1269 	return 0;
1270 }
1271 
1272 static int __hw_perf_event_init(struct perf_event *event)
1273 {
1274 	struct hw_perf_event *hwc = &event->hw;
1275 	int mapping;
1276 
1277 	mapping = pmu_map_event(event);
1278 
1279 	if (mapping < 0) {
1280 		pr_debug("event %x:%llx not supported\n", event->attr.type,
1281 			 event->attr.config);
1282 		return mapping;
1283 	}
1284 
1285 	/*
1286 	 * We don't assign an index until we actually place the event onto
1287 	 * hardware. Use -1 to signify that we haven't decided where to put it
1288 	 * yet.
1289 	 */
1290 	hwc->idx		= -1;
1291 	hwc->config_base	= 0;
1292 	hwc->config		= 0;
1293 	hwc->event_base		= 0;
1294 
1295 	/*
1296 	 * Store the event encoding into the config_base field.
1297 	 */
1298 	hwc->config_base	    |= (unsigned long)mapping;
1299 
1300 	if (event->group_leader != event) {
1301 		if (validate_group(event) != 0)
1302 			return -EINVAL;
1303 	}
1304 
1305 	return 0;
1306 }
1307 
1308 static int cci_pmu_event_init(struct perf_event *event)
1309 {
1310 	struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1311 	atomic_t *active_events = &cci_pmu->active_events;
1312 	int err = 0;
1313 
1314 	if (event->attr.type != event->pmu->type)
1315 		return -ENOENT;
1316 
1317 	/* Shared by all CPUs, no meaningful state to sample */
1318 	if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
1319 		return -EOPNOTSUPP;
1320 
1321 	/* We have no filtering of any kind */
1322 	if (event->attr.exclude_user	||
1323 	    event->attr.exclude_kernel	||
1324 	    event->attr.exclude_hv	||
1325 	    event->attr.exclude_idle	||
1326 	    event->attr.exclude_host	||
1327 	    event->attr.exclude_guest)
1328 		return -EINVAL;
1329 
1330 	/*
1331 	 * Following the example set by other "uncore" PMUs, we accept any CPU
1332 	 * and rewrite its affinity dynamically rather than having perf core
1333 	 * handle cpu == -1 and pid == -1 for this case.
1334 	 *
1335 	 * The perf core will pin online CPUs for the duration of this call and
1336 	 * the event being installed into its context, so the PMU's CPU can't
1337 	 * change under our feet.
1338 	 */
1339 	if (event->cpu < 0)
1340 		return -EINVAL;
1341 	event->cpu = cci_pmu->cpu;
1342 
1343 	event->destroy = hw_perf_event_destroy;
1344 	if (!atomic_inc_not_zero(active_events)) {
1345 		mutex_lock(&cci_pmu->reserve_mutex);
1346 		if (atomic_read(active_events) == 0)
1347 			err = cci_pmu_get_hw(cci_pmu);
1348 		if (!err)
1349 			atomic_inc(active_events);
1350 		mutex_unlock(&cci_pmu->reserve_mutex);
1351 	}
1352 	if (err)
1353 		return err;
1354 
1355 	err = __hw_perf_event_init(event);
1356 	if (err)
1357 		hw_perf_event_destroy(event);
1358 
1359 	return err;
1360 }
1361 
1362 static ssize_t pmu_cpumask_attr_show(struct device *dev,
1363 				     struct device_attribute *attr, char *buf)
1364 {
1365 	struct pmu *pmu = dev_get_drvdata(dev);
1366 	struct cci_pmu *cci_pmu = to_cci_pmu(pmu);
1367 
1368 	return cpumap_print_to_pagebuf(true, buf, cpumask_of(cci_pmu->cpu));
1369 }
1370 
1371 static struct device_attribute pmu_cpumask_attr =
1372 	__ATTR(cpumask, S_IRUGO, pmu_cpumask_attr_show, NULL);
1373 
1374 static struct attribute *pmu_attrs[] = {
1375 	&pmu_cpumask_attr.attr,
1376 	NULL,
1377 };
1378 
1379 static struct attribute_group pmu_attr_group = {
1380 	.attrs = pmu_attrs,
1381 };
1382 
1383 static struct attribute_group pmu_format_attr_group = {
1384 	.name = "format",
1385 	.attrs = NULL,		/* Filled in cci_pmu_init_attrs */
1386 };
1387 
1388 static struct attribute_group pmu_event_attr_group = {
1389 	.name = "events",
1390 	.attrs = NULL,		/* Filled in cci_pmu_init_attrs */
1391 };
1392 
1393 static const struct attribute_group *pmu_attr_groups[] = {
1394 	&pmu_attr_group,
1395 	&pmu_format_attr_group,
1396 	&pmu_event_attr_group,
1397 	NULL
1398 };
1399 
1400 static int cci_pmu_init(struct cci_pmu *cci_pmu, struct platform_device *pdev)
1401 {
1402 	const struct cci_pmu_model *model = cci_pmu->model;
1403 	char *name = model->name;
1404 	u32 num_cntrs;
1405 
1406 	pmu_event_attr_group.attrs = model->event_attrs;
1407 	pmu_format_attr_group.attrs = model->format_attrs;
1408 
1409 	cci_pmu->pmu = (struct pmu) {
1410 		.module		= THIS_MODULE,
1411 		.name		= cci_pmu->model->name,
1412 		.task_ctx_nr	= perf_invalid_context,
1413 		.pmu_enable	= cci_pmu_enable,
1414 		.pmu_disable	= cci_pmu_disable,
1415 		.event_init	= cci_pmu_event_init,
1416 		.add		= cci_pmu_add,
1417 		.del		= cci_pmu_del,
1418 		.start		= cci_pmu_start,
1419 		.stop		= cci_pmu_stop,
1420 		.read		= pmu_read,
1421 		.attr_groups	= pmu_attr_groups,
1422 	};
1423 
1424 	cci_pmu->plat_device = pdev;
1425 	num_cntrs = pmu_get_max_counters(cci_pmu);
1426 	if (num_cntrs > cci_pmu->model->num_hw_cntrs) {
1427 		dev_warn(&pdev->dev,
1428 			"PMU implements more counters(%d) than supported by"
1429 			" the model(%d), truncated.",
1430 			num_cntrs, cci_pmu->model->num_hw_cntrs);
1431 		num_cntrs = cci_pmu->model->num_hw_cntrs;
1432 	}
1433 	cci_pmu->num_cntrs = num_cntrs + cci_pmu->model->fixed_hw_cntrs;
1434 
1435 	return perf_pmu_register(&cci_pmu->pmu, name, -1);
1436 }
1437 
1438 static int cci_pmu_offline_cpu(unsigned int cpu)
1439 {
1440 	int target;
1441 
1442 	if (!g_cci_pmu || cpu != g_cci_pmu->cpu)
1443 		return 0;
1444 
1445 	target = cpumask_any_but(cpu_online_mask, cpu);
1446 	if (target >= nr_cpu_ids)
1447 		return 0;
1448 
1449 	perf_pmu_migrate_context(&g_cci_pmu->pmu, cpu, target);
1450 	g_cci_pmu->cpu = target;
1451 	return 0;
1452 }
1453 
1454 static __maybe_unused struct cci_pmu_model cci_pmu_models[] = {
1455 #ifdef CONFIG_ARM_CCI400_PMU
1456 	[CCI400_R0] = {
1457 		.name = "CCI_400",
1458 		.fixed_hw_cntrs = 1,	/* Cycle counter */
1459 		.num_hw_cntrs = 4,
1460 		.cntr_size = SZ_4K,
1461 		.format_attrs = cci400_pmu_format_attrs,
1462 		.event_attrs = cci400_r0_pmu_event_attrs,
1463 		.event_ranges = {
1464 			[CCI_IF_SLAVE] = {
1465 				CCI400_R0_SLAVE_PORT_MIN_EV,
1466 				CCI400_R0_SLAVE_PORT_MAX_EV,
1467 			},
1468 			[CCI_IF_MASTER] = {
1469 				CCI400_R0_MASTER_PORT_MIN_EV,
1470 				CCI400_R0_MASTER_PORT_MAX_EV,
1471 			},
1472 		},
1473 		.validate_hw_event = cci400_validate_hw_event,
1474 		.get_event_idx = cci400_get_event_idx,
1475 	},
1476 	[CCI400_R1] = {
1477 		.name = "CCI_400_r1",
1478 		.fixed_hw_cntrs = 1,	/* Cycle counter */
1479 		.num_hw_cntrs = 4,
1480 		.cntr_size = SZ_4K,
1481 		.format_attrs = cci400_pmu_format_attrs,
1482 		.event_attrs = cci400_r1_pmu_event_attrs,
1483 		.event_ranges = {
1484 			[CCI_IF_SLAVE] = {
1485 				CCI400_R1_SLAVE_PORT_MIN_EV,
1486 				CCI400_R1_SLAVE_PORT_MAX_EV,
1487 			},
1488 			[CCI_IF_MASTER] = {
1489 				CCI400_R1_MASTER_PORT_MIN_EV,
1490 				CCI400_R1_MASTER_PORT_MAX_EV,
1491 			},
1492 		},
1493 		.validate_hw_event = cci400_validate_hw_event,
1494 		.get_event_idx = cci400_get_event_idx,
1495 	},
1496 #endif
1497 #ifdef CONFIG_ARM_CCI5xx_PMU
1498 	[CCI500_R0] = {
1499 		.name = "CCI_500",
1500 		.fixed_hw_cntrs = 0,
1501 		.num_hw_cntrs = 8,
1502 		.cntr_size = SZ_64K,
1503 		.format_attrs = cci5xx_pmu_format_attrs,
1504 		.event_attrs = cci5xx_pmu_event_attrs,
1505 		.event_ranges = {
1506 			[CCI_IF_SLAVE] = {
1507 				CCI5xx_SLAVE_PORT_MIN_EV,
1508 				CCI5xx_SLAVE_PORT_MAX_EV,
1509 			},
1510 			[CCI_IF_MASTER] = {
1511 				CCI5xx_MASTER_PORT_MIN_EV,
1512 				CCI5xx_MASTER_PORT_MAX_EV,
1513 			},
1514 			[CCI_IF_GLOBAL] = {
1515 				CCI5xx_GLOBAL_PORT_MIN_EV,
1516 				CCI5xx_GLOBAL_PORT_MAX_EV,
1517 			},
1518 		},
1519 		.validate_hw_event = cci500_validate_hw_event,
1520 		.write_counters	= cci5xx_pmu_write_counters,
1521 	},
1522 	[CCI550_R0] = {
1523 		.name = "CCI_550",
1524 		.fixed_hw_cntrs = 0,
1525 		.num_hw_cntrs = 8,
1526 		.cntr_size = SZ_64K,
1527 		.format_attrs = cci5xx_pmu_format_attrs,
1528 		.event_attrs = cci5xx_pmu_event_attrs,
1529 		.event_ranges = {
1530 			[CCI_IF_SLAVE] = {
1531 				CCI5xx_SLAVE_PORT_MIN_EV,
1532 				CCI5xx_SLAVE_PORT_MAX_EV,
1533 			},
1534 			[CCI_IF_MASTER] = {
1535 				CCI5xx_MASTER_PORT_MIN_EV,
1536 				CCI5xx_MASTER_PORT_MAX_EV,
1537 			},
1538 			[CCI_IF_GLOBAL] = {
1539 				CCI5xx_GLOBAL_PORT_MIN_EV,
1540 				CCI5xx_GLOBAL_PORT_MAX_EV,
1541 			},
1542 		},
1543 		.validate_hw_event = cci550_validate_hw_event,
1544 		.write_counters	= cci5xx_pmu_write_counters,
1545 	},
1546 #endif
1547 };
1548 
1549 static const struct of_device_id arm_cci_pmu_matches[] = {
1550 #ifdef CONFIG_ARM_CCI400_PMU
1551 	{
1552 		.compatible = "arm,cci-400-pmu",
1553 		.data	= NULL,
1554 	},
1555 	{
1556 		.compatible = "arm,cci-400-pmu,r0",
1557 		.data	= &cci_pmu_models[CCI400_R0],
1558 	},
1559 	{
1560 		.compatible = "arm,cci-400-pmu,r1",
1561 		.data	= &cci_pmu_models[CCI400_R1],
1562 	},
1563 #endif
1564 #ifdef CONFIG_ARM_CCI5xx_PMU
1565 	{
1566 		.compatible = "arm,cci-500-pmu,r0",
1567 		.data = &cci_pmu_models[CCI500_R0],
1568 	},
1569 	{
1570 		.compatible = "arm,cci-550-pmu,r0",
1571 		.data = &cci_pmu_models[CCI550_R0],
1572 	},
1573 #endif
1574 	{},
1575 };
1576 MODULE_DEVICE_TABLE(of, arm_cci_pmu_matches);
1577 
1578 static bool is_duplicate_irq(int irq, int *irqs, int nr_irqs)
1579 {
1580 	int i;
1581 
1582 	for (i = 0; i < nr_irqs; i++)
1583 		if (irq == irqs[i])
1584 			return true;
1585 
1586 	return false;
1587 }
1588 
1589 static struct cci_pmu *cci_pmu_alloc(struct device *dev)
1590 {
1591 	struct cci_pmu *cci_pmu;
1592 	const struct cci_pmu_model *model;
1593 
1594 	/*
1595 	 * All allocations are devm_* hence we don't have to free
1596 	 * them explicitly on an error, as it would end up in driver
1597 	 * detach.
1598 	 */
1599 	cci_pmu = devm_kzalloc(dev, sizeof(*cci_pmu), GFP_KERNEL);
1600 	if (!cci_pmu)
1601 		return ERR_PTR(-ENOMEM);
1602 
1603 	cci_pmu->ctrl_base = *(void __iomem **)dev->platform_data;
1604 
1605 	model = of_device_get_match_data(dev);
1606 	if (!model) {
1607 		dev_warn(dev,
1608 			 "DEPRECATED compatible property, requires secure access to CCI registers");
1609 		model = probe_cci_model(cci_pmu);
1610 	}
1611 	if (!model) {
1612 		dev_warn(dev, "CCI PMU version not supported\n");
1613 		return ERR_PTR(-ENODEV);
1614 	}
1615 
1616 	cci_pmu->model = model;
1617 	cci_pmu->irqs = devm_kcalloc(dev, CCI_PMU_MAX_HW_CNTRS(model),
1618 					sizeof(*cci_pmu->irqs), GFP_KERNEL);
1619 	if (!cci_pmu->irqs)
1620 		return ERR_PTR(-ENOMEM);
1621 	cci_pmu->hw_events.events = devm_kcalloc(dev,
1622 					     CCI_PMU_MAX_HW_CNTRS(model),
1623 					     sizeof(*cci_pmu->hw_events.events),
1624 					     GFP_KERNEL);
1625 	if (!cci_pmu->hw_events.events)
1626 		return ERR_PTR(-ENOMEM);
1627 	cci_pmu->hw_events.used_mask = devm_kcalloc(dev,
1628 						BITS_TO_LONGS(CCI_PMU_MAX_HW_CNTRS(model)),
1629 						sizeof(*cci_pmu->hw_events.used_mask),
1630 						GFP_KERNEL);
1631 	if (!cci_pmu->hw_events.used_mask)
1632 		return ERR_PTR(-ENOMEM);
1633 
1634 	return cci_pmu;
1635 }
1636 
1637 static int cci_pmu_probe(struct platform_device *pdev)
1638 {
1639 	struct resource *res;
1640 	struct cci_pmu *cci_pmu;
1641 	int i, ret, irq;
1642 
1643 	cci_pmu = cci_pmu_alloc(&pdev->dev);
1644 	if (IS_ERR(cci_pmu))
1645 		return PTR_ERR(cci_pmu);
1646 
1647 	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
1648 	cci_pmu->base = devm_ioremap_resource(&pdev->dev, res);
1649 	if (IS_ERR(cci_pmu->base))
1650 		return -ENOMEM;
1651 
1652 	/*
1653 	 * CCI PMU has one overflow interrupt per counter; but some may be tied
1654 	 * together to a common interrupt.
1655 	 */
1656 	cci_pmu->nr_irqs = 0;
1657 	for (i = 0; i < CCI_PMU_MAX_HW_CNTRS(cci_pmu->model); i++) {
1658 		irq = platform_get_irq(pdev, i);
1659 		if (irq < 0)
1660 			break;
1661 
1662 		if (is_duplicate_irq(irq, cci_pmu->irqs, cci_pmu->nr_irqs))
1663 			continue;
1664 
1665 		cci_pmu->irqs[cci_pmu->nr_irqs++] = irq;
1666 	}
1667 
1668 	/*
1669 	 * Ensure that the device tree has as many interrupts as the number
1670 	 * of counters.
1671 	 */
1672 	if (i < CCI_PMU_MAX_HW_CNTRS(cci_pmu->model)) {
1673 		dev_warn(&pdev->dev, "In-correct number of interrupts: %d, should be %d\n",
1674 			i, CCI_PMU_MAX_HW_CNTRS(cci_pmu->model));
1675 		return -EINVAL;
1676 	}
1677 
1678 	raw_spin_lock_init(&cci_pmu->hw_events.pmu_lock);
1679 	mutex_init(&cci_pmu->reserve_mutex);
1680 	atomic_set(&cci_pmu->active_events, 0);
1681 	cci_pmu->cpu = get_cpu();
1682 
1683 	ret = cci_pmu_init(cci_pmu, pdev);
1684 	if (ret) {
1685 		put_cpu();
1686 		return ret;
1687 	}
1688 
1689 	cpuhp_setup_state_nocalls(CPUHP_AP_PERF_ARM_CCI_ONLINE,
1690 				  "perf/arm/cci:online", NULL,
1691 				  cci_pmu_offline_cpu);
1692 	put_cpu();
1693 	g_cci_pmu = cci_pmu;
1694 	pr_info("ARM %s PMU driver probed", cci_pmu->model->name);
1695 	return 0;
1696 }
1697 
1698 static int cci_pmu_remove(struct platform_device *pdev)
1699 {
1700 	if (!g_cci_pmu)
1701 		return 0;
1702 
1703 	cpuhp_remove_state(CPUHP_AP_PERF_ARM_CCI_ONLINE);
1704 	perf_pmu_unregister(&g_cci_pmu->pmu);
1705 	g_cci_pmu = NULL;
1706 
1707 	return 0;
1708 }
1709 
1710 static struct platform_driver cci_pmu_driver = {
1711 	.driver = {
1712 		   .name = DRIVER_NAME,
1713 		   .of_match_table = arm_cci_pmu_matches,
1714 		  },
1715 	.probe = cci_pmu_probe,
1716 	.remove = cci_pmu_remove,
1717 };
1718 
1719 module_platform_driver(cci_pmu_driver);
1720 MODULE_LICENSE("GPL v2");
1721 MODULE_DESCRIPTION("ARM CCI PMU support");
1722