2 * CCI cache coherent interconnect driver
4 * Copyright (C) 2013 ARM Ltd.
5 * Author: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 as
9 * published by the Free Software Foundation.
11 * This program is distributed "as is" WITHOUT ANY WARRANTY of any
12 * kind, whether express or implied; without even the implied warranty
13 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
17 #include <linux/arm-cci.h>
19 #include <linux/interrupt.h>
20 #include <linux/module.h>
21 #include <linux/of_address.h>
22 #include <linux/of_irq.h>
23 #include <linux/of_platform.h>
24 #include <linux/perf_event.h>
25 #include <linux/platform_device.h>
26 #include <linux/slab.h>
27 #include <linux/spinlock.h>
29 #include <asm/cacheflush.h>
30 #include <asm/smp_plat.h>
32 static void __iomem *cci_ctrl_base;
33 static unsigned long cci_ctrl_phys;
35 #ifdef CONFIG_ARM_CCI400_PORT_CTRL
38 unsigned int nb_ace_lite;
41 static const struct cci_nb_ports cci400_ports = {
46 #define CCI400_PORTS_DATA (&cci400_ports)
48 #define CCI400_PORTS_DATA (NULL)
51 static const struct of_device_id arm_cci_matches[] = {
52 #ifdef CONFIG_ARM_CCI400_COMMON
53 {.compatible = "arm,cci-400", .data = CCI400_PORTS_DATA },
55 #ifdef CONFIG_ARM_CCI5xx_PMU
56 { .compatible = "arm,cci-500", },
57 { .compatible = "arm,cci-550", },
62 #ifdef CONFIG_ARM_CCI_PMU
64 #define DRIVER_NAME "ARM-CCI"
65 #define DRIVER_NAME_PMU DRIVER_NAME " PMU"
67 #define CCI_PMCR 0x0100
68 #define CCI_PID2 0x0fe8
70 #define CCI_PMCR_CEN 0x00000001
71 #define CCI_PMCR_NCNT_MASK 0x0000f800
72 #define CCI_PMCR_NCNT_SHIFT 11
74 #define CCI_PID2_REV_MASK 0xf0
75 #define CCI_PID2_REV_SHIFT 4
77 #define CCI_PMU_EVT_SEL 0x000
78 #define CCI_PMU_CNTR 0x004
79 #define CCI_PMU_CNTR_CTRL 0x008
80 #define CCI_PMU_OVRFLW 0x00c
82 #define CCI_PMU_OVRFLW_FLAG 1
84 #define CCI_PMU_CNTR_SIZE(model) ((model)->cntr_size)
85 #define CCI_PMU_CNTR_BASE(model, idx) ((idx) * CCI_PMU_CNTR_SIZE(model))
86 #define CCI_PMU_CNTR_MASK ((1ULL << 32) -1)
87 #define CCI_PMU_CNTR_LAST(cci_pmu) (cci_pmu->num_cntrs - 1)
89 #define CCI_PMU_MAX_HW_CNTRS(model) \
90 ((model)->num_hw_cntrs + (model)->fixed_hw_cntrs)
92 /* Types of interfaces that can generate events */
96 #ifdef CONFIG_ARM_CCI5xx_PMU
107 struct cci_pmu_hw_events {
108 struct perf_event **events;
109 unsigned long *used_mask;
110 raw_spinlock_t pmu_lock;
115 * struct cci_pmu_model:
116 * @fixed_hw_cntrs - Number of fixed event counters
117 * @num_hw_cntrs - Maximum number of programmable event counters
118 * @cntr_size - Size of an event counter mapping
120 struct cci_pmu_model {
125 struct attribute **format_attrs;
126 struct attribute **event_attrs;
127 struct event_range event_ranges[CCI_IF_MAX];
128 int (*validate_hw_event)(struct cci_pmu *, unsigned long);
129 int (*get_event_idx)(struct cci_pmu *, struct cci_pmu_hw_events *, unsigned long);
130 void (*write_counters)(struct cci_pmu *, unsigned long *);
133 static struct cci_pmu_model cci_pmu_models[];
140 unsigned long active_irqs;
141 const struct cci_pmu_model *model;
142 struct cci_pmu_hw_events hw_events;
143 struct platform_device *plat_device;
145 atomic_t active_events;
146 struct mutex reserve_mutex;
147 struct hlist_node node;
151 #define to_cci_pmu(c) (container_of(c, struct cci_pmu, pmu))
154 #ifdef CONFIG_ARM_CCI400_PMU
158 #ifdef CONFIG_ARM_CCI5xx_PMU
165 static void pmu_write_counters(struct cci_pmu *cci_pmu,
166 unsigned long *mask);
167 static ssize_t cci_pmu_format_show(struct device *dev,
168 struct device_attribute *attr, char *buf);
169 static ssize_t cci_pmu_event_show(struct device *dev,
170 struct device_attribute *attr, char *buf);
172 #define CCI_EXT_ATTR_ENTRY(_name, _func, _config) \
173 &((struct dev_ext_attribute[]) { \
174 { __ATTR(_name, S_IRUGO, _func, NULL), (void *)_config } \
177 #define CCI_FORMAT_EXT_ATTR_ENTRY(_name, _config) \
178 CCI_EXT_ATTR_ENTRY(_name, cci_pmu_format_show, (char *)_config)
179 #define CCI_EVENT_EXT_ATTR_ENTRY(_name, _config) \
180 CCI_EXT_ATTR_ENTRY(_name, cci_pmu_event_show, (unsigned long)_config)
182 /* CCI400 PMU Specific definitions */
184 #ifdef CONFIG_ARM_CCI400_PMU
187 #define CCI400_PORT_S0 0
188 #define CCI400_PORT_S1 1
189 #define CCI400_PORT_S2 2
190 #define CCI400_PORT_S3 3
191 #define CCI400_PORT_S4 4
192 #define CCI400_PORT_M0 5
193 #define CCI400_PORT_M1 6
194 #define CCI400_PORT_M2 7
196 #define CCI400_R1_PX 5
199 * Instead of an event id to monitor CCI cycles, a dedicated counter is
200 * provided. Use 0xff to represent CCI cycles and hope that no future revisions
201 * make use of this event in hardware.
203 enum cci400_perf_events {
204 CCI400_PMU_CYCLES = 0xff
207 #define CCI400_PMU_CYCLE_CNTR_IDX 0
208 #define CCI400_PMU_CNTR0_IDX 1
211 * CCI PMU event id is an 8-bit value made of two parts - bits 7:5 for one of 8
212 * ports and bits 4:0 are event codes. There are different event codes
213 * associated with each port type.
215 * Additionally, the range of events associated with the port types changed
216 * between Rev0 and Rev1.
218 * The constants below define the range of valid codes for each port type for
219 * the different revisions and are used to validate the event to be monitored.
222 #define CCI400_PMU_EVENT_MASK 0xffUL
223 #define CCI400_PMU_EVENT_SOURCE_SHIFT 5
224 #define CCI400_PMU_EVENT_SOURCE_MASK 0x7
225 #define CCI400_PMU_EVENT_CODE_SHIFT 0
226 #define CCI400_PMU_EVENT_CODE_MASK 0x1f
227 #define CCI400_PMU_EVENT_SOURCE(event) \
228 ((event >> CCI400_PMU_EVENT_SOURCE_SHIFT) & \
229 CCI400_PMU_EVENT_SOURCE_MASK)
230 #define CCI400_PMU_EVENT_CODE(event) \
231 ((event >> CCI400_PMU_EVENT_CODE_SHIFT) & CCI400_PMU_EVENT_CODE_MASK)
233 #define CCI400_R0_SLAVE_PORT_MIN_EV 0x00
234 #define CCI400_R0_SLAVE_PORT_MAX_EV 0x13
235 #define CCI400_R0_MASTER_PORT_MIN_EV 0x14
236 #define CCI400_R0_MASTER_PORT_MAX_EV 0x1a
238 #define CCI400_R1_SLAVE_PORT_MIN_EV 0x00
239 #define CCI400_R1_SLAVE_PORT_MAX_EV 0x14
240 #define CCI400_R1_MASTER_PORT_MIN_EV 0x00
241 #define CCI400_R1_MASTER_PORT_MAX_EV 0x11
243 #define CCI400_CYCLE_EVENT_EXT_ATTR_ENTRY(_name, _config) \
244 CCI_EXT_ATTR_ENTRY(_name, cci400_pmu_cycle_event_show, \
245 (unsigned long)_config)
247 static ssize_t cci400_pmu_cycle_event_show(struct device *dev,
248 struct device_attribute *attr, char *buf);
250 static struct attribute *cci400_pmu_format_attrs[] = {
251 CCI_FORMAT_EXT_ATTR_ENTRY(event, "config:0-4"),
252 CCI_FORMAT_EXT_ATTR_ENTRY(source, "config:5-7"),
256 static struct attribute *cci400_r0_pmu_event_attrs[] = {
258 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_any, 0x0),
259 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_device, 0x01),
260 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_normal_or_nonshareable, 0x2),
261 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_inner_or_outershareable, 0x3),
262 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_cache_maintenance, 0x4),
263 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_mem_barrier, 0x5),
264 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_sync_barrier, 0x6),
265 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg, 0x7),
266 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg_sync, 0x8),
267 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_stall_tt_full, 0x9),
268 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_last_hs_snoop, 0xA),
269 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_stall_rvalids_h_rready_l, 0xB),
270 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_any, 0xC),
271 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_device, 0xD),
272 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_normal_or_nonshareable, 0xE),
273 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_inner_or_outershare_wback_wclean, 0xF),
274 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_write_unique, 0x10),
275 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_write_line_unique, 0x11),
276 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_evict, 0x12),
277 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_stall_tt_full, 0x13),
279 CCI_EVENT_EXT_ATTR_ENTRY(mi_retry_speculative_fetch, 0x14),
280 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_addr_hazard, 0x15),
281 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_id_hazard, 0x16),
282 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_tt_full, 0x17),
283 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_barrier_hazard, 0x18),
284 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_barrier_hazard, 0x19),
285 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_tt_full, 0x1A),
286 /* Special event for cycles counter */
287 CCI400_CYCLE_EVENT_EXT_ATTR_ENTRY(cycles, 0xff),
291 static struct attribute *cci400_r1_pmu_event_attrs[] = {
293 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_any, 0x0),
294 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_device, 0x01),
295 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_normal_or_nonshareable, 0x2),
296 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_inner_or_outershareable, 0x3),
297 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_cache_maintenance, 0x4),
298 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_mem_barrier, 0x5),
299 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_sync_barrier, 0x6),
300 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg, 0x7),
301 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg_sync, 0x8),
302 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_stall_tt_full, 0x9),
303 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_last_hs_snoop, 0xA),
304 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_stall_rvalids_h_rready_l, 0xB),
305 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_any, 0xC),
306 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_device, 0xD),
307 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_normal_or_nonshareable, 0xE),
308 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_inner_or_outershare_wback_wclean, 0xF),
309 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_write_unique, 0x10),
310 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_write_line_unique, 0x11),
311 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_evict, 0x12),
312 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_stall_tt_full, 0x13),
313 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_stall_slave_id_hazard, 0x14),
315 CCI_EVENT_EXT_ATTR_ENTRY(mi_retry_speculative_fetch, 0x0),
316 CCI_EVENT_EXT_ATTR_ENTRY(mi_stall_cycle_addr_hazard, 0x1),
317 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_master_id_hazard, 0x2),
318 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_hi_prio_rtq_full, 0x3),
319 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_barrier_hazard, 0x4),
320 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_barrier_hazard, 0x5),
321 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_wtq_full, 0x6),
322 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_low_prio_rtq_full, 0x7),
323 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_mid_prio_rtq_full, 0x8),
324 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_qvn_vn0, 0x9),
325 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_qvn_vn1, 0xA),
326 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_qvn_vn2, 0xB),
327 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_qvn_vn3, 0xC),
328 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_qvn_vn0, 0xD),
329 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_qvn_vn1, 0xE),
330 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_qvn_vn2, 0xF),
331 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_qvn_vn3, 0x10),
332 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_unique_or_line_unique_addr_hazard, 0x11),
333 /* Special event for cycles counter */
334 CCI400_CYCLE_EVENT_EXT_ATTR_ENTRY(cycles, 0xff),
338 static ssize_t cci400_pmu_cycle_event_show(struct device *dev,
339 struct device_attribute *attr, char *buf)
341 struct dev_ext_attribute *eattr = container_of(attr,
342 struct dev_ext_attribute, attr);
343 return snprintf(buf, PAGE_SIZE, "config=0x%lx\n", (unsigned long)eattr->var);
346 static int cci400_get_event_idx(struct cci_pmu *cci_pmu,
347 struct cci_pmu_hw_events *hw,
348 unsigned long cci_event)
352 /* cycles event idx is fixed */
353 if (cci_event == CCI400_PMU_CYCLES) {
354 if (test_and_set_bit(CCI400_PMU_CYCLE_CNTR_IDX, hw->used_mask))
357 return CCI400_PMU_CYCLE_CNTR_IDX;
360 for (idx = CCI400_PMU_CNTR0_IDX; idx <= CCI_PMU_CNTR_LAST(cci_pmu); ++idx)
361 if (!test_and_set_bit(idx, hw->used_mask))
364 /* No counters available */
368 static int cci400_validate_hw_event(struct cci_pmu *cci_pmu, unsigned long hw_event)
370 u8 ev_source = CCI400_PMU_EVENT_SOURCE(hw_event);
371 u8 ev_code = CCI400_PMU_EVENT_CODE(hw_event);
374 if (hw_event & ~CCI400_PMU_EVENT_MASK)
377 if (hw_event == CCI400_PMU_CYCLES)
386 /* Slave Interface */
387 if_type = CCI_IF_SLAVE;
392 /* Master Interface */
393 if_type = CCI_IF_MASTER;
399 if (ev_code >= cci_pmu->model->event_ranges[if_type].min &&
400 ev_code <= cci_pmu->model->event_ranges[if_type].max)
406 static int probe_cci400_revision(void)
409 rev = readl_relaxed(cci_ctrl_base + CCI_PID2) & CCI_PID2_REV_MASK;
410 rev >>= CCI_PID2_REV_SHIFT;
412 if (rev < CCI400_R1_PX)
418 static const struct cci_pmu_model *probe_cci_model(struct platform_device *pdev)
420 if (platform_has_secure_cci_access())
421 return &cci_pmu_models[probe_cci400_revision()];
424 #else /* !CONFIG_ARM_CCI400_PMU */
425 static inline struct cci_pmu_model *probe_cci_model(struct platform_device *pdev)
429 #endif /* CONFIG_ARM_CCI400_PMU */
431 #ifdef CONFIG_ARM_CCI5xx_PMU
434 * CCI5xx PMU event id is an 9-bit value made of two parts.
435 * bits [8:5] - Source for the event
436 * bits [4:0] - Event code (specific to type of interface)
442 #define CCI5xx_PORT_S0 0x0
443 #define CCI5xx_PORT_S1 0x1
444 #define CCI5xx_PORT_S2 0x2
445 #define CCI5xx_PORT_S3 0x3
446 #define CCI5xx_PORT_S4 0x4
447 #define CCI5xx_PORT_S5 0x5
448 #define CCI5xx_PORT_S6 0x6
450 #define CCI5xx_PORT_M0 0x8
451 #define CCI5xx_PORT_M1 0x9
452 #define CCI5xx_PORT_M2 0xa
453 #define CCI5xx_PORT_M3 0xb
454 #define CCI5xx_PORT_M4 0xc
455 #define CCI5xx_PORT_M5 0xd
456 #define CCI5xx_PORT_M6 0xe
458 #define CCI5xx_PORT_GLOBAL 0xf
460 #define CCI5xx_PMU_EVENT_MASK 0x1ffUL
461 #define CCI5xx_PMU_EVENT_SOURCE_SHIFT 0x5
462 #define CCI5xx_PMU_EVENT_SOURCE_MASK 0xf
463 #define CCI5xx_PMU_EVENT_CODE_SHIFT 0x0
464 #define CCI5xx_PMU_EVENT_CODE_MASK 0x1f
466 #define CCI5xx_PMU_EVENT_SOURCE(event) \
467 ((event >> CCI5xx_PMU_EVENT_SOURCE_SHIFT) & CCI5xx_PMU_EVENT_SOURCE_MASK)
468 #define CCI5xx_PMU_EVENT_CODE(event) \
469 ((event >> CCI5xx_PMU_EVENT_CODE_SHIFT) & CCI5xx_PMU_EVENT_CODE_MASK)
471 #define CCI5xx_SLAVE_PORT_MIN_EV 0x00
472 #define CCI5xx_SLAVE_PORT_MAX_EV 0x1f
473 #define CCI5xx_MASTER_PORT_MIN_EV 0x00
474 #define CCI5xx_MASTER_PORT_MAX_EV 0x06
475 #define CCI5xx_GLOBAL_PORT_MIN_EV 0x00
476 #define CCI5xx_GLOBAL_PORT_MAX_EV 0x0f
479 #define CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(_name, _config) \
480 CCI_EXT_ATTR_ENTRY(_name, cci5xx_pmu_global_event_show, \
481 (unsigned long) _config)
483 static ssize_t cci5xx_pmu_global_event_show(struct device *dev,
484 struct device_attribute *attr, char *buf);
486 static struct attribute *cci5xx_pmu_format_attrs[] = {
487 CCI_FORMAT_EXT_ATTR_ENTRY(event, "config:0-4"),
488 CCI_FORMAT_EXT_ATTR_ENTRY(source, "config:5-8"),
492 static struct attribute *cci5xx_pmu_event_attrs[] = {
494 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_arvalid, 0x0),
495 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_dev, 0x1),
496 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_nonshareable, 0x2),
497 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_shareable_non_alloc, 0x3),
498 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_shareable_alloc, 0x4),
499 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_invalidate, 0x5),
500 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_cache_maint, 0x6),
501 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg, 0x7),
502 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_rval, 0x8),
503 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_rlast_snoop, 0x9),
504 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_awalid, 0xA),
505 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_dev, 0xB),
506 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_non_shareable, 0xC),
507 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_share_wb, 0xD),
508 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_share_wlu, 0xE),
509 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_share_wunique, 0xF),
510 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_evict, 0x10),
511 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_wrevict, 0x11),
512 CCI_EVENT_EXT_ATTR_ENTRY(si_w_data_beat, 0x12),
513 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_acvalid, 0x13),
514 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_read, 0x14),
515 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_clean, 0x15),
516 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_data_transfer_low, 0x16),
517 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_stall_arvalid, 0x17),
518 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_stall, 0x18),
519 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_stall, 0x19),
520 CCI_EVENT_EXT_ATTR_ENTRY(si_w_data_stall, 0x1A),
521 CCI_EVENT_EXT_ATTR_ENTRY(si_w_resp_stall, 0x1B),
522 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_stall, 0x1C),
523 CCI_EVENT_EXT_ATTR_ENTRY(si_s_data_stall, 0x1D),
524 CCI_EVENT_EXT_ATTR_ENTRY(si_rq_stall_ot_limit, 0x1E),
525 CCI_EVENT_EXT_ATTR_ENTRY(si_r_stall_arbit, 0x1F),
528 CCI_EVENT_EXT_ATTR_ENTRY(mi_r_data_beat_any, 0x0),
529 CCI_EVENT_EXT_ATTR_ENTRY(mi_w_data_beat_any, 0x1),
530 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall, 0x2),
531 CCI_EVENT_EXT_ATTR_ENTRY(mi_r_data_stall, 0x3),
532 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall, 0x4),
533 CCI_EVENT_EXT_ATTR_ENTRY(mi_w_data_stall, 0x5),
534 CCI_EVENT_EXT_ATTR_ENTRY(mi_w_resp_stall, 0x6),
537 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_filter_bank_0_1, 0x0),
538 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_filter_bank_2_3, 0x1),
539 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_filter_bank_4_5, 0x2),
540 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_filter_bank_6_7, 0x3),
541 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_miss_filter_bank_0_1, 0x4),
542 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_miss_filter_bank_2_3, 0x5),
543 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_miss_filter_bank_4_5, 0x6),
544 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_miss_filter_bank_6_7, 0x7),
545 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_back_invalidation, 0x8),
546 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_stall_alloc_busy, 0x9),
547 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_stall_tt_full, 0xA),
548 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_wrq, 0xB),
549 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_cd_hs, 0xC),
550 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_rq_stall_addr_hazard, 0xD),
551 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_rq_stall_tt_full, 0xE),
552 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_rq_tzmp1_prot, 0xF),
556 static ssize_t cci5xx_pmu_global_event_show(struct device *dev,
557 struct device_attribute *attr, char *buf)
559 struct dev_ext_attribute *eattr = container_of(attr,
560 struct dev_ext_attribute, attr);
561 /* Global events have single fixed source code */
562 return snprintf(buf, PAGE_SIZE, "event=0x%lx,source=0x%x\n",
563 (unsigned long)eattr->var, CCI5xx_PORT_GLOBAL);
567 * CCI500 provides 8 independent event counters that can count
568 * any of the events available.
569 * CCI500 PMU event source ids
570 * 0x0-0x6 - Slave interfaces
571 * 0x8-0xD - Master interfaces
572 * 0xf - Global Events
575 static int cci500_validate_hw_event(struct cci_pmu *cci_pmu,
576 unsigned long hw_event)
578 u32 ev_source = CCI5xx_PMU_EVENT_SOURCE(hw_event);
579 u32 ev_code = CCI5xx_PMU_EVENT_CODE(hw_event);
582 if (hw_event & ~CCI5xx_PMU_EVENT_MASK)
593 if_type = CCI_IF_SLAVE;
601 if_type = CCI_IF_MASTER;
603 case CCI5xx_PORT_GLOBAL:
604 if_type = CCI_IF_GLOBAL;
610 if (ev_code >= cci_pmu->model->event_ranges[if_type].min &&
611 ev_code <= cci_pmu->model->event_ranges[if_type].max)
618 * CCI550 provides 8 independent event counters that can count
619 * any of the events available.
620 * CCI550 PMU event source ids
621 * 0x0-0x6 - Slave interfaces
622 * 0x8-0xe - Master interfaces
623 * 0xf - Global Events
626 static int cci550_validate_hw_event(struct cci_pmu *cci_pmu,
627 unsigned long hw_event)
629 u32 ev_source = CCI5xx_PMU_EVENT_SOURCE(hw_event);
630 u32 ev_code = CCI5xx_PMU_EVENT_CODE(hw_event);
633 if (hw_event & ~CCI5xx_PMU_EVENT_MASK)
644 if_type = CCI_IF_SLAVE;
653 if_type = CCI_IF_MASTER;
655 case CCI5xx_PORT_GLOBAL:
656 if_type = CCI_IF_GLOBAL;
662 if (ev_code >= cci_pmu->model->event_ranges[if_type].min &&
663 ev_code <= cci_pmu->model->event_ranges[if_type].max)
669 #endif /* CONFIG_ARM_CCI5xx_PMU */
672 * Program the CCI PMU counters which have PERF_HES_ARCH set
673 * with the event period and mark them ready before we enable
676 static void cci_pmu_sync_counters(struct cci_pmu *cci_pmu)
679 struct cci_pmu_hw_events *cci_hw = &cci_pmu->hw_events;
681 DECLARE_BITMAP(mask, cci_pmu->num_cntrs);
683 bitmap_zero(mask, cci_pmu->num_cntrs);
684 for_each_set_bit(i, cci_pmu->hw_events.used_mask, cci_pmu->num_cntrs) {
685 struct perf_event *event = cci_hw->events[i];
690 /* Leave the events which are not counting */
691 if (event->hw.state & PERF_HES_STOPPED)
693 if (event->hw.state & PERF_HES_ARCH) {
695 event->hw.state &= ~PERF_HES_ARCH;
699 pmu_write_counters(cci_pmu, mask);
702 /* Should be called with cci_pmu->hw_events->pmu_lock held */
703 static void __cci_pmu_enable_nosync(struct cci_pmu *cci_pmu)
707 /* Enable all the PMU counters. */
708 val = readl_relaxed(cci_ctrl_base + CCI_PMCR) | CCI_PMCR_CEN;
709 writel(val, cci_ctrl_base + CCI_PMCR);
712 /* Should be called with cci_pmu->hw_events->pmu_lock held */
713 static void __cci_pmu_enable_sync(struct cci_pmu *cci_pmu)
715 cci_pmu_sync_counters(cci_pmu);
716 __cci_pmu_enable_nosync(cci_pmu);
719 /* Should be called with cci_pmu->hw_events->pmu_lock held */
720 static void __cci_pmu_disable(void)
724 /* Disable all the PMU counters. */
725 val = readl_relaxed(cci_ctrl_base + CCI_PMCR) & ~CCI_PMCR_CEN;
726 writel(val, cci_ctrl_base + CCI_PMCR);
729 static ssize_t cci_pmu_format_show(struct device *dev,
730 struct device_attribute *attr, char *buf)
732 struct dev_ext_attribute *eattr = container_of(attr,
733 struct dev_ext_attribute, attr);
734 return snprintf(buf, PAGE_SIZE, "%s\n", (char *)eattr->var);
737 static ssize_t cci_pmu_event_show(struct device *dev,
738 struct device_attribute *attr, char *buf)
740 struct dev_ext_attribute *eattr = container_of(attr,
741 struct dev_ext_attribute, attr);
742 /* source parameter is mandatory for normal PMU events */
743 return snprintf(buf, PAGE_SIZE, "source=?,event=0x%lx\n",
744 (unsigned long)eattr->var);
747 static int pmu_is_valid_counter(struct cci_pmu *cci_pmu, int idx)
749 return 0 <= idx && idx <= CCI_PMU_CNTR_LAST(cci_pmu);
752 static u32 pmu_read_register(struct cci_pmu *cci_pmu, int idx, unsigned int offset)
754 return readl_relaxed(cci_pmu->base +
755 CCI_PMU_CNTR_BASE(cci_pmu->model, idx) + offset);
758 static void pmu_write_register(struct cci_pmu *cci_pmu, u32 value,
759 int idx, unsigned int offset)
761 writel_relaxed(value, cci_pmu->base +
762 CCI_PMU_CNTR_BASE(cci_pmu->model, idx) + offset);
765 static void pmu_disable_counter(struct cci_pmu *cci_pmu, int idx)
767 pmu_write_register(cci_pmu, 0, idx, CCI_PMU_CNTR_CTRL);
770 static void pmu_enable_counter(struct cci_pmu *cci_pmu, int idx)
772 pmu_write_register(cci_pmu, 1, idx, CCI_PMU_CNTR_CTRL);
775 static bool __maybe_unused
776 pmu_counter_is_enabled(struct cci_pmu *cci_pmu, int idx)
778 return (pmu_read_register(cci_pmu, idx, CCI_PMU_CNTR_CTRL) & 0x1) != 0;
781 static void pmu_set_event(struct cci_pmu *cci_pmu, int idx, unsigned long event)
783 pmu_write_register(cci_pmu, event, idx, CCI_PMU_EVT_SEL);
787 * For all counters on the CCI-PMU, disable any 'enabled' counters,
788 * saving the changed counters in the mask, so that we can restore
789 * it later using pmu_restore_counters. The mask is private to the
790 * caller. We cannot rely on the used_mask maintained by the CCI_PMU
791 * as it only tells us if the counter is assigned to perf_event or not.
792 * The state of the perf_event cannot be locked by the PMU layer, hence
793 * we check the individual counter status (which can be locked by
794 * cci_pm->hw_events->pmu_lock).
796 * @mask should be initialised to empty by the caller.
798 static void __maybe_unused
799 pmu_save_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
803 for (i = 0; i < cci_pmu->num_cntrs; i++) {
804 if (pmu_counter_is_enabled(cci_pmu, i)) {
806 pmu_disable_counter(cci_pmu, i);
812 * Restore the status of the counters. Reversal of the pmu_save_counters().
813 * For each counter set in the mask, enable the counter back.
815 static void __maybe_unused
816 pmu_restore_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
820 for_each_set_bit(i, mask, cci_pmu->num_cntrs)
821 pmu_enable_counter(cci_pmu, i);
825 * Returns the number of programmable counters actually implemented
828 static u32 pmu_get_max_counters(void)
830 return (readl_relaxed(cci_ctrl_base + CCI_PMCR) &
831 CCI_PMCR_NCNT_MASK) >> CCI_PMCR_NCNT_SHIFT;
834 static int pmu_get_event_idx(struct cci_pmu_hw_events *hw, struct perf_event *event)
836 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
837 unsigned long cci_event = event->hw.config_base;
840 if (cci_pmu->model->get_event_idx)
841 return cci_pmu->model->get_event_idx(cci_pmu, hw, cci_event);
843 /* Generic code to find an unused idx from the mask */
844 for(idx = 0; idx <= CCI_PMU_CNTR_LAST(cci_pmu); idx++)
845 if (!test_and_set_bit(idx, hw->used_mask))
848 /* No counters available */
852 static int pmu_map_event(struct perf_event *event)
854 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
856 if (event->attr.type < PERF_TYPE_MAX ||
857 !cci_pmu->model->validate_hw_event)
860 return cci_pmu->model->validate_hw_event(cci_pmu, event->attr.config);
863 static int pmu_request_irq(struct cci_pmu *cci_pmu, irq_handler_t handler)
866 struct platform_device *pmu_device = cci_pmu->plat_device;
868 if (unlikely(!pmu_device))
871 if (cci_pmu->nr_irqs < 1) {
872 dev_err(&pmu_device->dev, "no irqs for CCI PMUs defined\n");
877 * Register all available CCI PMU interrupts. In the interrupt handler
878 * we iterate over the counters checking for interrupt source (the
879 * overflowing counter) and clear it.
881 * This should allow handling of non-unique interrupt for the counters.
883 for (i = 0; i < cci_pmu->nr_irqs; i++) {
884 int err = request_irq(cci_pmu->irqs[i], handler, IRQF_SHARED,
885 "arm-cci-pmu", cci_pmu);
887 dev_err(&pmu_device->dev, "unable to request IRQ%d for ARM CCI PMU counters\n",
892 set_bit(i, &cci_pmu->active_irqs);
898 static void pmu_free_irq(struct cci_pmu *cci_pmu)
902 for (i = 0; i < cci_pmu->nr_irqs; i++) {
903 if (!test_and_clear_bit(i, &cci_pmu->active_irqs))
906 free_irq(cci_pmu->irqs[i], cci_pmu);
910 static u32 pmu_read_counter(struct perf_event *event)
912 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
913 struct hw_perf_event *hw_counter = &event->hw;
914 int idx = hw_counter->idx;
917 if (unlikely(!pmu_is_valid_counter(cci_pmu, idx))) {
918 dev_err(&cci_pmu->plat_device->dev, "Invalid CCI PMU counter %d\n", idx);
921 value = pmu_read_register(cci_pmu, idx, CCI_PMU_CNTR);
926 static void pmu_write_counter(struct cci_pmu *cci_pmu, u32 value, int idx)
928 pmu_write_register(cci_pmu, value, idx, CCI_PMU_CNTR);
931 static void __pmu_write_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
934 struct cci_pmu_hw_events *cci_hw = &cci_pmu->hw_events;
936 for_each_set_bit(i, mask, cci_pmu->num_cntrs) {
937 struct perf_event *event = cci_hw->events[i];
941 pmu_write_counter(cci_pmu, local64_read(&event->hw.prev_count), i);
945 static void pmu_write_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
947 if (cci_pmu->model->write_counters)
948 cci_pmu->model->write_counters(cci_pmu, mask);
950 __pmu_write_counters(cci_pmu, mask);
953 #ifdef CONFIG_ARM_CCI5xx_PMU
956 * CCI-500/CCI-550 has advanced power saving policies, which could gate the
957 * clocks to the PMU counters, which makes the writes to them ineffective.
958 * The only way to write to those counters is when the global counters
959 * are enabled and the particular counter is enabled.
961 * So we do the following :
963 * 1) Disable all the PMU counters, saving their current state
964 * 2) Enable the global PMU profiling, now that all counters are
967 * For each counter to be programmed, repeat steps 3-7:
969 * 3) Write an invalid event code to the event control register for the
970 counter, so that the counters are not modified.
971 * 4) Enable the counter control for the counter.
972 * 5) Set the counter value
973 * 6) Disable the counter
974 * 7) Restore the event in the target counter
976 * 8) Disable the global PMU.
977 * 9) Restore the status of the rest of the counters.
979 * We choose an event which for CCI-5xx is guaranteed not to count.
980 * We use the highest possible event code (0x1f) for the master interface 0.
982 #define CCI5xx_INVALID_EVENT ((CCI5xx_PORT_M0 << CCI5xx_PMU_EVENT_SOURCE_SHIFT) | \
983 (CCI5xx_PMU_EVENT_CODE_MASK << CCI5xx_PMU_EVENT_CODE_SHIFT))
984 static void cci5xx_pmu_write_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
987 DECLARE_BITMAP(saved_mask, cci_pmu->num_cntrs);
989 bitmap_zero(saved_mask, cci_pmu->num_cntrs);
990 pmu_save_counters(cci_pmu, saved_mask);
993 * Now that all the counters are disabled, we can safely turn the PMU on,
994 * without syncing the status of the counters
996 __cci_pmu_enable_nosync(cci_pmu);
998 for_each_set_bit(i, mask, cci_pmu->num_cntrs) {
999 struct perf_event *event = cci_pmu->hw_events.events[i];
1001 if (WARN_ON(!event))
1004 pmu_set_event(cci_pmu, i, CCI5xx_INVALID_EVENT);
1005 pmu_enable_counter(cci_pmu, i);
1006 pmu_write_counter(cci_pmu, local64_read(&event->hw.prev_count), i);
1007 pmu_disable_counter(cci_pmu, i);
1008 pmu_set_event(cci_pmu, i, event->hw.config_base);
1011 __cci_pmu_disable();
1013 pmu_restore_counters(cci_pmu, saved_mask);
1016 #endif /* CONFIG_ARM_CCI5xx_PMU */
1018 static u64 pmu_event_update(struct perf_event *event)
1020 struct hw_perf_event *hwc = &event->hw;
1021 u64 delta, prev_raw_count, new_raw_count;
1024 prev_raw_count = local64_read(&hwc->prev_count);
1025 new_raw_count = pmu_read_counter(event);
1026 } while (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
1027 new_raw_count) != prev_raw_count);
1029 delta = (new_raw_count - prev_raw_count) & CCI_PMU_CNTR_MASK;
1031 local64_add(delta, &event->count);
1033 return new_raw_count;
1036 static void pmu_read(struct perf_event *event)
1038 pmu_event_update(event);
1041 static void pmu_event_set_period(struct perf_event *event)
1043 struct hw_perf_event *hwc = &event->hw;
1045 * The CCI PMU counters have a period of 2^32. To account for the
1046 * possiblity of extreme interrupt latency we program for a period of
1047 * half that. Hopefully we can handle the interrupt before another 2^31
1048 * events occur and the counter overtakes its previous value.
1050 u64 val = 1ULL << 31;
1051 local64_set(&hwc->prev_count, val);
1054 * CCI PMU uses PERF_HES_ARCH to keep track of the counters, whose
1055 * values needs to be sync-ed with the s/w state before the PMU is
1057 * Mark this counter for sync.
1059 hwc->state |= PERF_HES_ARCH;
1062 static irqreturn_t pmu_handle_irq(int irq_num, void *dev)
1064 unsigned long flags;
1065 struct cci_pmu *cci_pmu = dev;
1066 struct cci_pmu_hw_events *events = &cci_pmu->hw_events;
1067 int idx, handled = IRQ_NONE;
1069 raw_spin_lock_irqsave(&events->pmu_lock, flags);
1071 /* Disable the PMU while we walk through the counters */
1072 __cci_pmu_disable();
1074 * Iterate over counters and update the corresponding perf events.
1075 * This should work regardless of whether we have per-counter overflow
1076 * interrupt or a combined overflow interrupt.
1078 for (idx = 0; idx <= CCI_PMU_CNTR_LAST(cci_pmu); idx++) {
1079 struct perf_event *event = events->events[idx];
1084 /* Did this counter overflow? */
1085 if (!(pmu_read_register(cci_pmu, idx, CCI_PMU_OVRFLW) &
1086 CCI_PMU_OVRFLW_FLAG))
1089 pmu_write_register(cci_pmu, CCI_PMU_OVRFLW_FLAG, idx,
1092 pmu_event_update(event);
1093 pmu_event_set_period(event);
1094 handled = IRQ_HANDLED;
1097 /* Enable the PMU and sync possibly overflowed counters */
1098 __cci_pmu_enable_sync(cci_pmu);
1099 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
1101 return IRQ_RETVAL(handled);
1104 static int cci_pmu_get_hw(struct cci_pmu *cci_pmu)
1106 int ret = pmu_request_irq(cci_pmu, pmu_handle_irq);
1108 pmu_free_irq(cci_pmu);
1114 static void cci_pmu_put_hw(struct cci_pmu *cci_pmu)
1116 pmu_free_irq(cci_pmu);
1119 static void hw_perf_event_destroy(struct perf_event *event)
1121 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1122 atomic_t *active_events = &cci_pmu->active_events;
1123 struct mutex *reserve_mutex = &cci_pmu->reserve_mutex;
1125 if (atomic_dec_and_mutex_lock(active_events, reserve_mutex)) {
1126 cci_pmu_put_hw(cci_pmu);
1127 mutex_unlock(reserve_mutex);
1131 static void cci_pmu_enable(struct pmu *pmu)
1133 struct cci_pmu *cci_pmu = to_cci_pmu(pmu);
1134 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1135 int enabled = bitmap_weight(hw_events->used_mask, cci_pmu->num_cntrs);
1136 unsigned long flags;
1141 raw_spin_lock_irqsave(&hw_events->pmu_lock, flags);
1142 __cci_pmu_enable_sync(cci_pmu);
1143 raw_spin_unlock_irqrestore(&hw_events->pmu_lock, flags);
1147 static void cci_pmu_disable(struct pmu *pmu)
1149 struct cci_pmu *cci_pmu = to_cci_pmu(pmu);
1150 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1151 unsigned long flags;
1153 raw_spin_lock_irqsave(&hw_events->pmu_lock, flags);
1154 __cci_pmu_disable();
1155 raw_spin_unlock_irqrestore(&hw_events->pmu_lock, flags);
1159 * Check if the idx represents a non-programmable counter.
1160 * All the fixed event counters are mapped before the programmable
1163 static bool pmu_fixed_hw_idx(struct cci_pmu *cci_pmu, int idx)
1165 return (idx >= 0) && (idx < cci_pmu->model->fixed_hw_cntrs);
1168 static void cci_pmu_start(struct perf_event *event, int pmu_flags)
1170 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1171 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1172 struct hw_perf_event *hwc = &event->hw;
1174 unsigned long flags;
1177 * To handle interrupt latency, we always reprogram the period
1178 * regardlesss of PERF_EF_RELOAD.
1180 if (pmu_flags & PERF_EF_RELOAD)
1181 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
1185 if (unlikely(!pmu_is_valid_counter(cci_pmu, idx))) {
1186 dev_err(&cci_pmu->plat_device->dev, "Invalid CCI PMU counter %d\n", idx);
1190 raw_spin_lock_irqsave(&hw_events->pmu_lock, flags);
1192 /* Configure the counter unless you are counting a fixed event */
1193 if (!pmu_fixed_hw_idx(cci_pmu, idx))
1194 pmu_set_event(cci_pmu, idx, hwc->config_base);
1196 pmu_event_set_period(event);
1197 pmu_enable_counter(cci_pmu, idx);
1199 raw_spin_unlock_irqrestore(&hw_events->pmu_lock, flags);
1202 static void cci_pmu_stop(struct perf_event *event, int pmu_flags)
1204 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1205 struct hw_perf_event *hwc = &event->hw;
1208 if (hwc->state & PERF_HES_STOPPED)
1211 if (unlikely(!pmu_is_valid_counter(cci_pmu, idx))) {
1212 dev_err(&cci_pmu->plat_device->dev, "Invalid CCI PMU counter %d\n", idx);
1217 * We always reprogram the counter, so ignore PERF_EF_UPDATE. See
1220 pmu_disable_counter(cci_pmu, idx);
1221 pmu_event_update(event);
1222 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
1225 static int cci_pmu_add(struct perf_event *event, int flags)
1227 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1228 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1229 struct hw_perf_event *hwc = &event->hw;
1233 perf_pmu_disable(event->pmu);
1235 /* If we don't have a space for the counter then finish early. */
1236 idx = pmu_get_event_idx(hw_events, event);
1242 event->hw.idx = idx;
1243 hw_events->events[idx] = event;
1245 hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
1246 if (flags & PERF_EF_START)
1247 cci_pmu_start(event, PERF_EF_RELOAD);
1249 /* Propagate our changes to the userspace mapping. */
1250 perf_event_update_userpage(event);
1253 perf_pmu_enable(event->pmu);
1257 static void cci_pmu_del(struct perf_event *event, int flags)
1259 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1260 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1261 struct hw_perf_event *hwc = &event->hw;
1264 cci_pmu_stop(event, PERF_EF_UPDATE);
1265 hw_events->events[idx] = NULL;
1266 clear_bit(idx, hw_events->used_mask);
1268 perf_event_update_userpage(event);
1272 validate_event(struct pmu *cci_pmu,
1273 struct cci_pmu_hw_events *hw_events,
1274 struct perf_event *event)
1276 if (is_software_event(event))
1280 * Reject groups spanning multiple HW PMUs (e.g. CPU + CCI). The
1281 * core perf code won't check that the pmu->ctx == leader->ctx
1282 * until after pmu->event_init(event).
1284 if (event->pmu != cci_pmu)
1287 if (event->state < PERF_EVENT_STATE_OFF)
1290 if (event->state == PERF_EVENT_STATE_OFF && !event->attr.enable_on_exec)
1293 return pmu_get_event_idx(hw_events, event) >= 0;
1297 validate_group(struct perf_event *event)
1299 struct perf_event *sibling, *leader = event->group_leader;
1300 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1301 unsigned long mask[BITS_TO_LONGS(cci_pmu->num_cntrs)];
1302 struct cci_pmu_hw_events fake_pmu = {
1304 * Initialise the fake PMU. We only need to populate the
1305 * used_mask for the purposes of validation.
1309 memset(mask, 0, BITS_TO_LONGS(cci_pmu->num_cntrs) * sizeof(unsigned long));
1311 if (!validate_event(event->pmu, &fake_pmu, leader))
1314 list_for_each_entry(sibling, &leader->sibling_list, group_entry) {
1315 if (!validate_event(event->pmu, &fake_pmu, sibling))
1319 if (!validate_event(event->pmu, &fake_pmu, event))
1326 __hw_perf_event_init(struct perf_event *event)
1328 struct hw_perf_event *hwc = &event->hw;
1331 mapping = pmu_map_event(event);
1334 pr_debug("event %x:%llx not supported\n", event->attr.type,
1335 event->attr.config);
1340 * We don't assign an index until we actually place the event onto
1341 * hardware. Use -1 to signify that we haven't decided where to put it
1345 hwc->config_base = 0;
1347 hwc->event_base = 0;
1350 * Store the event encoding into the config_base field.
1352 hwc->config_base |= (unsigned long)mapping;
1355 * Limit the sample_period to half of the counter width. That way, the
1356 * new counter value is far less likely to overtake the previous one
1357 * unless you have some serious IRQ latency issues.
1359 hwc->sample_period = CCI_PMU_CNTR_MASK >> 1;
1360 hwc->last_period = hwc->sample_period;
1361 local64_set(&hwc->period_left, hwc->sample_period);
1363 if (event->group_leader != event) {
1364 if (validate_group(event) != 0)
1371 static int cci_pmu_event_init(struct perf_event *event)
1373 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1374 atomic_t *active_events = &cci_pmu->active_events;
1378 if (event->attr.type != event->pmu->type)
1381 /* Shared by all CPUs, no meaningful state to sample */
1382 if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
1385 /* We have no filtering of any kind */
1386 if (event->attr.exclude_user ||
1387 event->attr.exclude_kernel ||
1388 event->attr.exclude_hv ||
1389 event->attr.exclude_idle ||
1390 event->attr.exclude_host ||
1391 event->attr.exclude_guest)
1395 * Following the example set by other "uncore" PMUs, we accept any CPU
1396 * and rewrite its affinity dynamically rather than having perf core
1397 * handle cpu == -1 and pid == -1 for this case.
1399 * The perf core will pin online CPUs for the duration of this call and
1400 * the event being installed into its context, so the PMU's CPU can't
1401 * change under our feet.
1403 cpu = cpumask_first(&cci_pmu->cpus);
1404 if (event->cpu < 0 || cpu < 0)
1408 event->destroy = hw_perf_event_destroy;
1409 if (!atomic_inc_not_zero(active_events)) {
1410 mutex_lock(&cci_pmu->reserve_mutex);
1411 if (atomic_read(active_events) == 0)
1412 err = cci_pmu_get_hw(cci_pmu);
1414 atomic_inc(active_events);
1415 mutex_unlock(&cci_pmu->reserve_mutex);
1420 err = __hw_perf_event_init(event);
1422 hw_perf_event_destroy(event);
1427 static ssize_t pmu_cpumask_attr_show(struct device *dev,
1428 struct device_attribute *attr, char *buf)
1430 struct pmu *pmu = dev_get_drvdata(dev);
1431 struct cci_pmu *cci_pmu = to_cci_pmu(pmu);
1433 int n = scnprintf(buf, PAGE_SIZE - 1, "%*pbl",
1434 cpumask_pr_args(&cci_pmu->cpus));
1440 static struct device_attribute pmu_cpumask_attr =
1441 __ATTR(cpumask, S_IRUGO, pmu_cpumask_attr_show, NULL);
1443 static struct attribute *pmu_attrs[] = {
1444 &pmu_cpumask_attr.attr,
1448 static struct attribute_group pmu_attr_group = {
1452 static struct attribute_group pmu_format_attr_group = {
1454 .attrs = NULL, /* Filled in cci_pmu_init_attrs */
1457 static struct attribute_group pmu_event_attr_group = {
1459 .attrs = NULL, /* Filled in cci_pmu_init_attrs */
1462 static const struct attribute_group *pmu_attr_groups[] = {
1464 &pmu_format_attr_group,
1465 &pmu_event_attr_group,
1469 static int cci_pmu_init(struct cci_pmu *cci_pmu, struct platform_device *pdev)
1471 const struct cci_pmu_model *model = cci_pmu->model;
1472 char *name = model->name;
1475 pmu_event_attr_group.attrs = model->event_attrs;
1476 pmu_format_attr_group.attrs = model->format_attrs;
1478 cci_pmu->pmu = (struct pmu) {
1479 .name = cci_pmu->model->name,
1480 .task_ctx_nr = perf_invalid_context,
1481 .pmu_enable = cci_pmu_enable,
1482 .pmu_disable = cci_pmu_disable,
1483 .event_init = cci_pmu_event_init,
1486 .start = cci_pmu_start,
1487 .stop = cci_pmu_stop,
1489 .attr_groups = pmu_attr_groups,
1492 cci_pmu->plat_device = pdev;
1493 num_cntrs = pmu_get_max_counters();
1494 if (num_cntrs > cci_pmu->model->num_hw_cntrs) {
1495 dev_warn(&pdev->dev,
1496 "PMU implements more counters(%d) than supported by"
1497 " the model(%d), truncated.",
1498 num_cntrs, cci_pmu->model->num_hw_cntrs);
1499 num_cntrs = cci_pmu->model->num_hw_cntrs;
1501 cci_pmu->num_cntrs = num_cntrs + cci_pmu->model->fixed_hw_cntrs;
1503 return perf_pmu_register(&cci_pmu->pmu, name, -1);
1506 static int cci_pmu_offline_cpu(unsigned int cpu, struct hlist_node *node)
1508 struct cci_pmu *cci_pmu = hlist_entry_safe(node, struct cci_pmu, node);
1509 unsigned int target;
1511 if (!cpumask_test_and_clear_cpu(cpu, &cci_pmu->cpus))
1513 target = cpumask_any_but(cpu_online_mask, cpu);
1514 if (target >= nr_cpu_ids)
1517 * TODO: migrate context once core races on event->ctx have
1520 cpumask_set_cpu(target, &cci_pmu->cpus);
1524 static struct cci_pmu_model cci_pmu_models[] = {
1525 #ifdef CONFIG_ARM_CCI400_PMU
1528 .fixed_hw_cntrs = 1, /* Cycle counter */
1531 .format_attrs = cci400_pmu_format_attrs,
1532 .event_attrs = cci400_r0_pmu_event_attrs,
1535 CCI400_R0_SLAVE_PORT_MIN_EV,
1536 CCI400_R0_SLAVE_PORT_MAX_EV,
1539 CCI400_R0_MASTER_PORT_MIN_EV,
1540 CCI400_R0_MASTER_PORT_MAX_EV,
1543 .validate_hw_event = cci400_validate_hw_event,
1544 .get_event_idx = cci400_get_event_idx,
1547 .name = "CCI_400_r1",
1548 .fixed_hw_cntrs = 1, /* Cycle counter */
1551 .format_attrs = cci400_pmu_format_attrs,
1552 .event_attrs = cci400_r1_pmu_event_attrs,
1555 CCI400_R1_SLAVE_PORT_MIN_EV,
1556 CCI400_R1_SLAVE_PORT_MAX_EV,
1559 CCI400_R1_MASTER_PORT_MIN_EV,
1560 CCI400_R1_MASTER_PORT_MAX_EV,
1563 .validate_hw_event = cci400_validate_hw_event,
1564 .get_event_idx = cci400_get_event_idx,
1567 #ifdef CONFIG_ARM_CCI5xx_PMU
1570 .fixed_hw_cntrs = 0,
1572 .cntr_size = SZ_64K,
1573 .format_attrs = cci5xx_pmu_format_attrs,
1574 .event_attrs = cci5xx_pmu_event_attrs,
1577 CCI5xx_SLAVE_PORT_MIN_EV,
1578 CCI5xx_SLAVE_PORT_MAX_EV,
1581 CCI5xx_MASTER_PORT_MIN_EV,
1582 CCI5xx_MASTER_PORT_MAX_EV,
1585 CCI5xx_GLOBAL_PORT_MIN_EV,
1586 CCI5xx_GLOBAL_PORT_MAX_EV,
1589 .validate_hw_event = cci500_validate_hw_event,
1590 .write_counters = cci5xx_pmu_write_counters,
1594 .fixed_hw_cntrs = 0,
1596 .cntr_size = SZ_64K,
1597 .format_attrs = cci5xx_pmu_format_attrs,
1598 .event_attrs = cci5xx_pmu_event_attrs,
1601 CCI5xx_SLAVE_PORT_MIN_EV,
1602 CCI5xx_SLAVE_PORT_MAX_EV,
1605 CCI5xx_MASTER_PORT_MIN_EV,
1606 CCI5xx_MASTER_PORT_MAX_EV,
1609 CCI5xx_GLOBAL_PORT_MIN_EV,
1610 CCI5xx_GLOBAL_PORT_MAX_EV,
1613 .validate_hw_event = cci550_validate_hw_event,
1614 .write_counters = cci5xx_pmu_write_counters,
1619 static const struct of_device_id arm_cci_pmu_matches[] = {
1620 #ifdef CONFIG_ARM_CCI400_PMU
1622 .compatible = "arm,cci-400-pmu",
1626 .compatible = "arm,cci-400-pmu,r0",
1627 .data = &cci_pmu_models[CCI400_R0],
1630 .compatible = "arm,cci-400-pmu,r1",
1631 .data = &cci_pmu_models[CCI400_R1],
1634 #ifdef CONFIG_ARM_CCI5xx_PMU
1636 .compatible = "arm,cci-500-pmu,r0",
1637 .data = &cci_pmu_models[CCI500_R0],
1640 .compatible = "arm,cci-550-pmu,r0",
1641 .data = &cci_pmu_models[CCI550_R0],
1647 static inline const struct cci_pmu_model *get_cci_model(struct platform_device *pdev)
1649 const struct of_device_id *match = of_match_node(arm_cci_pmu_matches,
1656 dev_warn(&pdev->dev, "DEPRECATED compatible property,"
1657 "requires secure access to CCI registers");
1658 return probe_cci_model(pdev);
1661 static bool is_duplicate_irq(int irq, int *irqs, int nr_irqs)
1665 for (i = 0; i < nr_irqs; i++)
1672 static struct cci_pmu *cci_pmu_alloc(struct platform_device *pdev)
1674 struct cci_pmu *cci_pmu;
1675 const struct cci_pmu_model *model;
1678 * All allocations are devm_* hence we don't have to free
1679 * them explicitly on an error, as it would end up in driver
1682 model = get_cci_model(pdev);
1684 dev_warn(&pdev->dev, "CCI PMU version not supported\n");
1685 return ERR_PTR(-ENODEV);
1688 cci_pmu = devm_kzalloc(&pdev->dev, sizeof(*cci_pmu), GFP_KERNEL);
1690 return ERR_PTR(-ENOMEM);
1692 cci_pmu->model = model;
1693 cci_pmu->irqs = devm_kcalloc(&pdev->dev, CCI_PMU_MAX_HW_CNTRS(model),
1694 sizeof(*cci_pmu->irqs), GFP_KERNEL);
1696 return ERR_PTR(-ENOMEM);
1697 cci_pmu->hw_events.events = devm_kcalloc(&pdev->dev,
1698 CCI_PMU_MAX_HW_CNTRS(model),
1699 sizeof(*cci_pmu->hw_events.events),
1701 if (!cci_pmu->hw_events.events)
1702 return ERR_PTR(-ENOMEM);
1703 cci_pmu->hw_events.used_mask = devm_kcalloc(&pdev->dev,
1704 BITS_TO_LONGS(CCI_PMU_MAX_HW_CNTRS(model)),
1705 sizeof(*cci_pmu->hw_events.used_mask),
1707 if (!cci_pmu->hw_events.used_mask)
1708 return ERR_PTR(-ENOMEM);
1714 static int cci_pmu_probe(struct platform_device *pdev)
1716 struct resource *res;
1717 struct cci_pmu *cci_pmu;
1720 cci_pmu = cci_pmu_alloc(pdev);
1721 if (IS_ERR(cci_pmu))
1722 return PTR_ERR(cci_pmu);
1724 res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
1725 cci_pmu->base = devm_ioremap_resource(&pdev->dev, res);
1726 if (IS_ERR(cci_pmu->base))
1730 * CCI PMU has one overflow interrupt per counter; but some may be tied
1731 * together to a common interrupt.
1733 cci_pmu->nr_irqs = 0;
1734 for (i = 0; i < CCI_PMU_MAX_HW_CNTRS(cci_pmu->model); i++) {
1735 irq = platform_get_irq(pdev, i);
1739 if (is_duplicate_irq(irq, cci_pmu->irqs, cci_pmu->nr_irqs))
1742 cci_pmu->irqs[cci_pmu->nr_irqs++] = irq;
1746 * Ensure that the device tree has as many interrupts as the number
1749 if (i < CCI_PMU_MAX_HW_CNTRS(cci_pmu->model)) {
1750 dev_warn(&pdev->dev, "In-correct number of interrupts: %d, should be %d\n",
1751 i, CCI_PMU_MAX_HW_CNTRS(cci_pmu->model));
1755 raw_spin_lock_init(&cci_pmu->hw_events.pmu_lock);
1756 mutex_init(&cci_pmu->reserve_mutex);
1757 atomic_set(&cci_pmu->active_events, 0);
1758 cpumask_set_cpu(get_cpu(), &cci_pmu->cpus);
1760 ret = cci_pmu_init(cci_pmu, pdev);
1766 cpuhp_state_add_instance_nocalls(CPUHP_AP_PERF_ARM_CCI_ONLINE,
1769 pr_info("ARM %s PMU driver probed", cci_pmu->model->name);
1773 static int cci_platform_probe(struct platform_device *pdev)
1778 return of_platform_populate(pdev->dev.of_node, NULL, NULL, &pdev->dev);
1781 static struct platform_driver cci_pmu_driver = {
1783 .name = DRIVER_NAME_PMU,
1784 .of_match_table = arm_cci_pmu_matches,
1786 .probe = cci_pmu_probe,
1789 static struct platform_driver cci_platform_driver = {
1791 .name = DRIVER_NAME,
1792 .of_match_table = arm_cci_matches,
1794 .probe = cci_platform_probe,
1797 static int __init cci_platform_init(void)
1801 ret = cpuhp_setup_state_multi(CPUHP_AP_PERF_ARM_CCI_ONLINE,
1802 "perf/arm/cci:online", NULL,
1803 cci_pmu_offline_cpu);
1807 ret = platform_driver_register(&cci_pmu_driver);
1811 return platform_driver_register(&cci_platform_driver);
1814 #else /* !CONFIG_ARM_CCI_PMU */
1816 static int __init cci_platform_init(void)
1821 #endif /* CONFIG_ARM_CCI_PMU */
1823 #ifdef CONFIG_ARM_CCI400_PORT_CTRL
1825 #define CCI_PORT_CTRL 0x0
1826 #define CCI_CTRL_STATUS 0xc
1828 #define CCI_ENABLE_SNOOP_REQ 0x1
1829 #define CCI_ENABLE_DVM_REQ 0x2
1830 #define CCI_ENABLE_REQ (CCI_ENABLE_SNOOP_REQ | CCI_ENABLE_DVM_REQ)
1832 enum cci_ace_port_type {
1833 ACE_INVALID_PORT = 0x0,
1838 struct cci_ace_port {
1841 enum cci_ace_port_type type;
1842 struct device_node *dn;
1845 static struct cci_ace_port *ports;
1846 static unsigned int nb_cci_ports;
1854 * Use the port MSB as valid flag, shift can be made dynamic
1855 * by computing number of bits required for port indexes.
1856 * Code disabling CCI cpu ports runs with D-cache invalidated
1857 * and SCTLR bit clear so data accesses must be kept to a minimum
1858 * to improve performance; for now shift is left static to
1859 * avoid one more data access while disabling the CCI port.
1861 #define PORT_VALID_SHIFT 31
1862 #define PORT_VALID (0x1 << PORT_VALID_SHIFT)
1864 static inline void init_cpu_port(struct cpu_port *port, u32 index, u64 mpidr)
1866 port->port = PORT_VALID | index;
1867 port->mpidr = mpidr;
1870 static inline bool cpu_port_is_valid(struct cpu_port *port)
1872 return !!(port->port & PORT_VALID);
1875 static inline bool cpu_port_match(struct cpu_port *port, u64 mpidr)
1877 return port->mpidr == (mpidr & MPIDR_HWID_BITMASK);
1880 static struct cpu_port cpu_port[NR_CPUS];
1883 * __cci_ace_get_port - Function to retrieve the port index connected to
1886 * @dn: device node of the device to look-up
1890 * - CCI port index if success
1891 * - -ENODEV if failure
1893 static int __cci_ace_get_port(struct device_node *dn, int type)
1897 struct device_node *cci_portn;
1899 cci_portn = of_parse_phandle(dn, "cci-control-port", 0);
1900 for (i = 0; i < nb_cci_ports; i++) {
1901 ace_match = ports[i].type == type;
1902 if (ace_match && cci_portn == ports[i].dn)
1908 int cci_ace_get_port(struct device_node *dn)
1910 return __cci_ace_get_port(dn, ACE_LITE_PORT);
1912 EXPORT_SYMBOL_GPL(cci_ace_get_port);
1914 static void cci_ace_init_ports(void)
1917 struct device_node *cpun;
1920 * Port index look-up speeds up the function disabling ports by CPU,
1921 * since the logical to port index mapping is done once and does
1922 * not change after system boot.
1923 * The stashed index array is initialized for all possible CPUs
1926 for_each_possible_cpu(cpu) {
1927 /* too early to use cpu->of_node */
1928 cpun = of_get_cpu_node(cpu, NULL);
1930 if (WARN(!cpun, "Missing cpu device node\n"))
1933 port = __cci_ace_get_port(cpun, ACE_PORT);
1937 init_cpu_port(&cpu_port[cpu], port, cpu_logical_map(cpu));
1940 for_each_possible_cpu(cpu) {
1941 WARN(!cpu_port_is_valid(&cpu_port[cpu]),
1942 "CPU %u does not have an associated CCI port\n",
1947 * Functions to enable/disable a CCI interconnect slave port
1949 * They are called by low-level power management code to disable slave
1950 * interfaces snoops and DVM broadcast.
1951 * Since they may execute with cache data allocation disabled and
1952 * after the caches have been cleaned and invalidated the functions provide
1953 * no explicit locking since they may run with D-cache disabled, so normal
1954 * cacheable kernel locks based on ldrex/strex may not work.
1955 * Locking has to be provided by BSP implementations to ensure proper
1960 * cci_port_control() - function to control a CCI port
1962 * @port: index of the port to setup
1963 * @enable: if true enables the port, if false disables it
1965 static void notrace cci_port_control(unsigned int port, bool enable)
1967 void __iomem *base = ports[port].base;
1969 writel_relaxed(enable ? CCI_ENABLE_REQ : 0, base + CCI_PORT_CTRL);
1971 * This function is called from power down procedures
1972 * and must not execute any instruction that might
1973 * cause the processor to be put in a quiescent state
1974 * (eg wfi). Hence, cpu_relax() can not be added to this
1975 * read loop to optimize power, since it might hide possibly
1976 * disruptive operations.
1978 while (readl_relaxed(cci_ctrl_base + CCI_CTRL_STATUS) & 0x1)
1983 * cci_disable_port_by_cpu() - function to disable a CCI port by CPU
1986 * @mpidr: mpidr of the CPU whose CCI port should be disabled
1988 * Disabling a CCI port for a CPU implies disabling the CCI port
1989 * controlling that CPU cluster. Code disabling CPU CCI ports
1990 * must make sure that the CPU running the code is the last active CPU
1991 * in the cluster ie all other CPUs are quiescent in a low power state.
1995 * -ENODEV on port look-up failure
1997 int notrace cci_disable_port_by_cpu(u64 mpidr)
2001 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
2002 is_valid = cpu_port_is_valid(&cpu_port[cpu]);
2003 if (is_valid && cpu_port_match(&cpu_port[cpu], mpidr)) {
2004 cci_port_control(cpu_port[cpu].port, false);
2010 EXPORT_SYMBOL_GPL(cci_disable_port_by_cpu);
2013 * cci_enable_port_for_self() - enable a CCI port for calling CPU
2015 * Enabling a CCI port for the calling CPU implies enabling the CCI
2016 * port controlling that CPU's cluster. Caller must make sure that the
2017 * CPU running the code is the first active CPU in the cluster and all
2018 * other CPUs are quiescent in a low power state or waiting for this CPU
2019 * to complete the CCI initialization.
2021 * Because this is called when the MMU is still off and with no stack,
2022 * the code must be position independent and ideally rely on callee
2023 * clobbered registers only. To achieve this we must code this function
2024 * entirely in assembler.
2026 * On success this returns with the proper CCI port enabled. In case of
2027 * any failure this never returns as the inability to enable the CCI is
2028 * fatal and there is no possible recovery at this stage.
2030 asmlinkage void __naked cci_enable_port_for_self(void)
2034 " mrc p15, 0, r0, c0, c0, 5 @ get MPIDR value \n"
2035 " and r0, r0, #"__stringify(MPIDR_HWID_BITMASK)" \n"
2038 " add r1, r1, r2 @ &cpu_port \n"
2039 " add ip, r1, %[sizeof_cpu_port] \n"
2041 /* Loop over the cpu_port array looking for a matching MPIDR */
2042 "1: ldr r2, [r1, %[offsetof_cpu_port_mpidr_lsb]] \n"
2043 " cmp r2, r0 @ compare MPIDR \n"
2046 /* Found a match, now test port validity */
2047 " ldr r3, [r1, %[offsetof_cpu_port_port]] \n"
2048 " tst r3, #"__stringify(PORT_VALID)" \n"
2051 /* no match, loop with the next cpu_port entry */
2052 "2: add r1, r1, %[sizeof_struct_cpu_port] \n"
2053 " cmp r1, ip @ done? \n"
2056 /* CCI port not found -- cheaply try to stall this CPU */
2057 "cci_port_not_found: \n"
2060 " b cci_port_not_found \n"
2062 /* Use matched port index to look up the corresponding ports entry */
2063 "3: bic r3, r3, #"__stringify(PORT_VALID)" \n"
2065 " ldmia r0, {r1, r2} \n"
2066 " sub r1, r1, r0 @ virt - phys \n"
2067 " ldr r0, [r0, r2] @ *(&ports) \n"
2068 " mov r2, %[sizeof_struct_ace_port] \n"
2069 " mla r0, r2, r3, r0 @ &ports[index] \n"
2070 " sub r0, r0, r1 @ virt_to_phys() \n"
2072 /* Enable the CCI port */
2073 " ldr r0, [r0, %[offsetof_port_phys]] \n"
2074 " mov r3, %[cci_enable_req]\n"
2075 " str r3, [r0, #"__stringify(CCI_PORT_CTRL)"] \n"
2077 /* poll the status reg for completion */
2080 " ldr r0, [r0, r1] @ cci_ctrl_base \n"
2081 "4: ldr r1, [r0, #"__stringify(CCI_CTRL_STATUS)"] \n"
2082 " tst r1, %[cci_control_status_bits] \n"
2089 "5: .word cpu_port - . \n"
2091 " .word ports - 6b \n"
2092 "7: .word cci_ctrl_phys - . \n"
2094 [sizeof_cpu_port] "i" (sizeof(cpu_port)),
2095 [cci_enable_req] "i" cpu_to_le32(CCI_ENABLE_REQ),
2096 [cci_control_status_bits] "i" cpu_to_le32(1),
2098 [offsetof_cpu_port_mpidr_lsb] "i" (offsetof(struct cpu_port, mpidr)),
2100 [offsetof_cpu_port_mpidr_lsb] "i" (offsetof(struct cpu_port, mpidr)+4),
2102 [offsetof_cpu_port_port] "i" (offsetof(struct cpu_port, port)),
2103 [sizeof_struct_cpu_port] "i" (sizeof(struct cpu_port)),
2104 [sizeof_struct_ace_port] "i" (sizeof(struct cci_ace_port)),
2105 [offsetof_port_phys] "i" (offsetof(struct cci_ace_port, phys)) );
2111 * __cci_control_port_by_device() - function to control a CCI port by device
2114 * @dn: device node pointer of the device whose CCI port should be
2116 * @enable: if true enables the port, if false disables it
2120 * -ENODEV on port look-up failure
2122 int notrace __cci_control_port_by_device(struct device_node *dn, bool enable)
2129 port = __cci_ace_get_port(dn, ACE_LITE_PORT);
2130 if (WARN_ONCE(port < 0, "node %pOF ACE lite port look-up failure\n",
2133 cci_port_control(port, enable);
2136 EXPORT_SYMBOL_GPL(__cci_control_port_by_device);
2139 * __cci_control_port_by_index() - function to control a CCI port by port index
2141 * @port: port index previously retrieved with cci_ace_get_port()
2142 * @enable: if true enables the port, if false disables it
2146 * -ENODEV on port index out of range
2147 * -EPERM if operation carried out on an ACE PORT
2149 int notrace __cci_control_port_by_index(u32 port, bool enable)
2151 if (port >= nb_cci_ports || ports[port].type == ACE_INVALID_PORT)
2154 * CCI control for ports connected to CPUS is extremely fragile
2155 * and must be made to go through a specific and controlled
2156 * interface (ie cci_disable_port_by_cpu(); control by general purpose
2157 * indexing is therefore disabled for ACE ports.
2159 if (ports[port].type == ACE_PORT)
2162 cci_port_control(port, enable);
2165 EXPORT_SYMBOL_GPL(__cci_control_port_by_index);
2167 static const struct of_device_id arm_cci_ctrl_if_matches[] = {
2168 {.compatible = "arm,cci-400-ctrl-if", },
2172 static int cci_probe_ports(struct device_node *np)
2174 struct cci_nb_ports const *cci_config;
2175 int ret, i, nb_ace = 0, nb_ace_lite = 0;
2176 struct device_node *cp;
2177 struct resource res;
2178 const char *match_str;
2182 cci_config = of_match_node(arm_cci_matches, np)->data;
2186 nb_cci_ports = cci_config->nb_ace + cci_config->nb_ace_lite;
2188 ports = kcalloc(nb_cci_ports, sizeof(*ports), GFP_KERNEL);
2192 for_each_child_of_node(np, cp) {
2193 if (!of_match_node(arm_cci_ctrl_if_matches, cp))
2196 if (!of_device_is_available(cp))
2199 i = nb_ace + nb_ace_lite;
2201 if (i >= nb_cci_ports)
2204 if (of_property_read_string(cp, "interface-type",
2206 WARN(1, "node %pOF missing interface-type property\n",
2210 is_ace = strcmp(match_str, "ace") == 0;
2211 if (!is_ace && strcmp(match_str, "ace-lite")) {
2212 WARN(1, "node %pOF containing invalid interface-type property, skipping it\n",
2217 ret = of_address_to_resource(cp, 0, &res);
2219 ports[i].base = ioremap(res.start, resource_size(&res));
2220 ports[i].phys = res.start;
2222 if (ret || !ports[i].base) {
2223 WARN(1, "unable to ioremap CCI port %d\n", i);
2228 if (WARN_ON(nb_ace >= cci_config->nb_ace))
2230 ports[i].type = ACE_PORT;
2233 if (WARN_ON(nb_ace_lite >= cci_config->nb_ace_lite))
2235 ports[i].type = ACE_LITE_PORT;
2242 * If there is no CCI port that is under kernel control
2243 * return early and report probe status.
2245 if (!nb_ace && !nb_ace_lite)
2248 /* initialize a stashed array of ACE ports to speed-up look-up */
2249 cci_ace_init_ports();
2252 * Multi-cluster systems may need this data when non-coherent, during
2253 * cluster power-up/power-down. Make sure it reaches main memory.
2255 sync_cache_w(&cci_ctrl_base);
2256 sync_cache_w(&cci_ctrl_phys);
2257 sync_cache_w(&ports);
2258 sync_cache_w(&cpu_port);
2259 __sync_cache_range_w(ports, sizeof(*ports) * nb_cci_ports);
2260 pr_info("ARM CCI driver probed\n");
2264 #else /* !CONFIG_ARM_CCI400_PORT_CTRL */
2265 static inline int cci_probe_ports(struct device_node *np)
2269 #endif /* CONFIG_ARM_CCI400_PORT_CTRL */
2271 static int cci_probe(void)
2274 struct device_node *np;
2275 struct resource res;
2277 np = of_find_matching_node(NULL, arm_cci_matches);
2278 if(!np || !of_device_is_available(np))
2281 ret = of_address_to_resource(np, 0, &res);
2283 cci_ctrl_base = ioremap(res.start, resource_size(&res));
2284 cci_ctrl_phys = res.start;
2286 if (ret || !cci_ctrl_base) {
2287 WARN(1, "unable to ioremap CCI ctrl\n");
2291 return cci_probe_ports(np);
2294 static int cci_init_status = -EAGAIN;
2295 static DEFINE_MUTEX(cci_probing);
2297 static int cci_init(void)
2299 if (cci_init_status != -EAGAIN)
2300 return cci_init_status;
2302 mutex_lock(&cci_probing);
2303 if (cci_init_status == -EAGAIN)
2304 cci_init_status = cci_probe();
2305 mutex_unlock(&cci_probing);
2306 return cci_init_status;
2310 * To sort out early init calls ordering a helper function is provided to
2311 * check if the CCI driver has beed initialized. Function check if the driver
2312 * has been initialized, if not it calls the init function that probes
2313 * the driver and updates the return value.
2315 bool cci_probed(void)
2317 return cci_init() == 0;
2319 EXPORT_SYMBOL_GPL(cci_probed);
2321 early_initcall(cci_init);
2322 core_initcall(cci_platform_init);
2323 MODULE_LICENSE("GPL");
2324 MODULE_DESCRIPTION("ARM CCI support");