1 // SPDX-License-Identifier: GPL-2.0
2 // Copyright (C) 2016-2020 Arm Limited
3 // CMN-600 Coherent Mesh Network PMU driver
5 #include <linux/acpi.h>
6 #include <linux/bitfield.h>
7 #include <linux/bitops.h>
8 #include <linux/interrupt.h>
10 #include <linux/kernel.h>
11 #include <linux/list.h>
12 #include <linux/module.h>
14 #include <linux/perf_event.h>
15 #include <linux/platform_device.h>
16 #include <linux/slab.h>
17 #include <linux/sort.h>
19 /* Common register stuff */
20 #define CMN_NODE_INFO 0x0000
21 #define CMN_NI_NODE_TYPE GENMASK_ULL(15, 0)
22 #define CMN_NI_NODE_ID GENMASK_ULL(31, 16)
23 #define CMN_NI_LOGICAL_ID GENMASK_ULL(47, 32)
25 #define CMN_NODEID_DEVID(reg) ((reg) & 3)
26 #define CMN_NODEID_PID(reg) (((reg) >> 2) & 1)
27 #define CMN_NODEID_X(reg, bits) ((reg) >> (3 + (bits)))
28 #define CMN_NODEID_Y(reg, bits) (((reg) >> 3) & ((1U << (bits)) - 1))
30 #define CMN_CHILD_INFO 0x0080
31 #define CMN_CI_CHILD_COUNT GENMASK_ULL(15, 0)
32 #define CMN_CI_CHILD_PTR_OFFSET GENMASK_ULL(31, 16)
34 #define CMN_CHILD_NODE_ADDR GENMASK(27, 0)
35 #define CMN_CHILD_NODE_EXTERNAL BIT(31)
37 #define CMN_ADDR_NODE_PTR GENMASK(27, 14)
39 #define CMN_NODE_PTR_DEVID(ptr) (((ptr) >> 2) & 3)
40 #define CMN_NODE_PTR_PID(ptr) ((ptr) & 1)
41 #define CMN_NODE_PTR_X(ptr, bits) ((ptr) >> (6 + (bits)))
42 #define CMN_NODE_PTR_Y(ptr, bits) (((ptr) >> 6) & ((1U << (bits)) - 1))
44 #define CMN_MAX_XPS (8 * 8)
46 /* The CFG node has one other useful purpose */
47 #define CMN_CFGM_PERIPH_ID_2 0x0010
48 #define CMN_CFGM_PID2_REVISION GENMASK(7, 4)
50 /* PMU registers occupy the 3rd 4KB page of each node's 16KB space */
51 #define CMN_PMU_OFFSET 0x2000
53 /* For most nodes, this is all there is */
54 #define CMN_PMU_EVENT_SEL 0x000
55 #define CMN_PMU_EVENTn_ID_SHIFT(n) ((n) * 8)
57 /* DTMs live in the PMU space of XP registers */
58 #define CMN_DTM_WPn(n) (0x1A0 + (n) * 0x18)
59 #define CMN_DTM_WPn_CONFIG(n) (CMN_DTM_WPn(n) + 0x00)
60 #define CMN_DTM_WPn_CONFIG_WP_COMBINE BIT(6)
61 #define CMN_DTM_WPn_CONFIG_WP_EXCLUSIVE BIT(5)
62 #define CMN_DTM_WPn_CONFIG_WP_GRP BIT(4)
63 #define CMN_DTM_WPn_CONFIG_WP_CHN_SEL GENMASK_ULL(3, 1)
64 #define CMN_DTM_WPn_CONFIG_WP_DEV_SEL BIT(0)
65 #define CMN_DTM_WPn_VAL(n) (CMN_DTM_WPn(n) + 0x08)
66 #define CMN_DTM_WPn_MASK(n) (CMN_DTM_WPn(n) + 0x10)
68 #define CMN_DTM_PMU_CONFIG 0x210
69 #define CMN__PMEVCNT0_INPUT_SEL GENMASK_ULL(37, 32)
70 #define CMN__PMEVCNT0_INPUT_SEL_WP 0x00
71 #define CMN__PMEVCNT0_INPUT_SEL_XP 0x04
72 #define CMN__PMEVCNT0_INPUT_SEL_DEV 0x10
73 #define CMN__PMEVCNT0_GLOBAL_NUM GENMASK_ULL(18, 16)
74 #define CMN__PMEVCNTn_GLOBAL_NUM_SHIFT(n) ((n) * 4)
75 #define CMN__PMEVCNT_PAIRED(n) BIT(4 + (n))
76 #define CMN__PMEVCNT23_COMBINED BIT(2)
77 #define CMN__PMEVCNT01_COMBINED BIT(1)
78 #define CMN_DTM_PMU_CONFIG_PMU_EN BIT(0)
80 #define CMN_DTM_PMEVCNT 0x220
82 #define CMN_DTM_PMEVCNTSR 0x240
84 #define CMN_DTM_NUM_COUNTERS 4
86 /* The DTC node is where the magic happens */
87 #define CMN_DT_DTC_CTL 0x0a00
88 #define CMN_DT_DTC_CTL_DT_EN BIT(0)
90 /* DTC counters are paired in 64-bit registers on a 16-byte stride. Yuck */
91 #define _CMN_DT_CNT_REG(n) ((((n) / 2) * 4 + (n) % 2) * 4)
92 #define CMN_DT_PMEVCNT(n) (CMN_PMU_OFFSET + _CMN_DT_CNT_REG(n))
93 #define CMN_DT_PMCCNTR (CMN_PMU_OFFSET + 0x40)
95 #define CMN_DT_PMEVCNTSR(n) (CMN_PMU_OFFSET + 0x50 + _CMN_DT_CNT_REG(n))
96 #define CMN_DT_PMCCNTRSR (CMN_PMU_OFFSET + 0x90)
98 #define CMN_DT_PMCR (CMN_PMU_OFFSET + 0x100)
99 #define CMN_DT_PMCR_PMU_EN BIT(0)
100 #define CMN_DT_PMCR_CNTR_RST BIT(5)
101 #define CMN_DT_PMCR_OVFL_INTR_EN BIT(6)
103 #define CMN_DT_PMOVSR (CMN_PMU_OFFSET + 0x118)
104 #define CMN_DT_PMOVSR_CLR (CMN_PMU_OFFSET + 0x120)
106 #define CMN_DT_PMSSR (CMN_PMU_OFFSET + 0x128)
107 #define CMN_DT_PMSSR_SS_STATUS(n) BIT(n)
109 #define CMN_DT_PMSRR (CMN_PMU_OFFSET + 0x130)
110 #define CMN_DT_PMSRR_SS_REQ BIT(0)
112 #define CMN_DT_NUM_COUNTERS 8
113 #define CMN_MAX_DTCS 4
116 * Even in the worst case a DTC counter can't wrap in fewer than 2^42 cycles,
117 * so throwing away one bit to make overflow handling easy is no big deal.
119 #define CMN_COUNTER_INIT 0x80000000
120 /* Similarly for the 40-bit cycle counter */
121 #define CMN_CC_INIT 0x8000000000ULL
124 /* Event attributes */
125 #define CMN_CONFIG_TYPE GENMASK(15, 0)
126 #define CMN_CONFIG_EVENTID GENMASK(23, 16)
127 #define CMN_CONFIG_OCCUPID GENMASK(27, 24)
128 #define CMN_CONFIG_BYNODEID BIT(31)
129 #define CMN_CONFIG_NODEID GENMASK(47, 32)
131 #define CMN_EVENT_TYPE(event) FIELD_GET(CMN_CONFIG_TYPE, (event)->attr.config)
132 #define CMN_EVENT_EVENTID(event) FIELD_GET(CMN_CONFIG_EVENTID, (event)->attr.config)
133 #define CMN_EVENT_OCCUPID(event) FIELD_GET(CMN_CONFIG_OCCUPID, (event)->attr.config)
134 #define CMN_EVENT_BYNODEID(event) FIELD_GET(CMN_CONFIG_BYNODEID, (event)->attr.config)
135 #define CMN_EVENT_NODEID(event) FIELD_GET(CMN_CONFIG_NODEID, (event)->attr.config)
137 #define CMN_CONFIG_WP_COMBINE GENMASK(27, 24)
138 #define CMN_CONFIG_WP_DEV_SEL BIT(48)
139 #define CMN_CONFIG_WP_CHN_SEL GENMASK(50, 49)
140 #define CMN_CONFIG_WP_GRP BIT(52)
141 #define CMN_CONFIG_WP_EXCLUSIVE BIT(53)
142 #define CMN_CONFIG1_WP_VAL GENMASK(63, 0)
143 #define CMN_CONFIG2_WP_MASK GENMASK(63, 0)
145 #define CMN_EVENT_WP_COMBINE(event) FIELD_GET(CMN_CONFIG_WP_COMBINE, (event)->attr.config)
146 #define CMN_EVENT_WP_DEV_SEL(event) FIELD_GET(CMN_CONFIG_WP_DEV_SEL, (event)->attr.config)
147 #define CMN_EVENT_WP_CHN_SEL(event) FIELD_GET(CMN_CONFIG_WP_CHN_SEL, (event)->attr.config)
148 #define CMN_EVENT_WP_GRP(event) FIELD_GET(CMN_CONFIG_WP_GRP, (event)->attr.config)
149 #define CMN_EVENT_WP_EXCLUSIVE(event) FIELD_GET(CMN_CONFIG_WP_EXCLUSIVE, (event)->attr.config)
150 #define CMN_EVENT_WP_VAL(event) FIELD_GET(CMN_CONFIG1_WP_VAL, (event)->attr.config1)
151 #define CMN_EVENT_WP_MASK(event) FIELD_GET(CMN_CONFIG2_WP_MASK, (event)->attr.config2)
153 /* Made-up event IDs for watchpoint direction */
155 #define CMN_WP_DOWN 2
158 /* r0px probably don't exist in silicon, thankfully */
179 CMN_TYPE_RNSAM = 0xf,
180 CMN_TYPE_CXRA = 0x100,
181 CMN_TYPE_CXHA = 0x101,
182 CMN_TYPE_CXLA = 0x102,
183 /* Not a real node type */
187 struct arm_cmn_node {
188 void __iomem *pmu_base;
190 enum cmn_node_type type;
197 unsigned int occupid_val;
198 unsigned int occupid_count;
206 __le32 pmu_config_high;
224 struct perf_event *counters[CMN_DT_NUM_COUNTERS];
225 struct perf_event *cycles;
228 #define CMN_STATE_DISABLED BIT(0)
229 #define CMN_STATE_TXN BIT(1)
235 enum cmn_revision rev;
240 struct arm_cmn_node *xps;
241 struct arm_cmn_node *dns;
243 struct arm_cmn_dtc *dtc;
244 unsigned int num_dtcs;
247 struct hlist_node cpuhp_node;
253 #define to_cmn(p) container_of(p, struct arm_cmn, pmu)
255 static int arm_cmn_hp_state;
257 struct arm_cmn_hw_event {
258 struct arm_cmn_node *dn;
260 unsigned int dtc_idx;
265 #define for_each_hw_dn(hw, dn, i) \
266 for (i = 0, dn = hw->dn; i < hw->num_dns; i++, dn++)
268 static struct arm_cmn_hw_event *to_cmn_hw(struct perf_event *event)
270 BUILD_BUG_ON(sizeof(struct arm_cmn_hw_event) > offsetof(struct hw_perf_event, target));
271 return (struct arm_cmn_hw_event *)&event->hw;
274 static void arm_cmn_set_index(u64 x[], unsigned int pos, unsigned int val)
276 x[pos / 32] |= (u64)val << ((pos % 32) * 2);
279 static unsigned int arm_cmn_get_index(u64 x[], unsigned int pos)
281 return (x[pos / 32] >> ((pos % 32) * 2)) & 3;
284 struct arm_cmn_event_attr {
285 struct device_attribute attr;
286 enum cmn_node_type type;
291 struct arm_cmn_format_attr {
292 struct device_attribute attr;
297 static int arm_cmn_xyidbits(const struct arm_cmn *cmn)
299 return cmn->mesh_x > 4 || cmn->mesh_y > 4 ? 3 : 2;
302 static void arm_cmn_init_node_to_xp(const struct arm_cmn *cmn,
303 struct arm_cmn_node *dn)
305 int bits = arm_cmn_xyidbits(cmn);
306 int x = CMN_NODEID_X(dn->id, bits);
307 int y = CMN_NODEID_Y(dn->id, bits);
308 int xp_idx = cmn->mesh_x * y + x;
310 dn->to_xp = (cmn->xps + xp_idx) - dn;
313 static struct arm_cmn_node *arm_cmn_node_to_xp(struct arm_cmn_node *dn)
315 return dn->type == CMN_TYPE_XP ? dn : dn + dn->to_xp;
318 static struct arm_cmn_node *arm_cmn_node(const struct arm_cmn *cmn,
319 enum cmn_node_type type)
323 for (i = 0; i < cmn->num_dns; i++)
324 if (cmn->dns[i].type == type)
329 #define CMN_EVENT_ATTR(_name, _type, _eventid, _occupid) \
330 (&((struct arm_cmn_event_attr[]) {{ \
331 .attr = __ATTR(_name, 0444, arm_cmn_event_show, NULL), \
333 .eventid = _eventid, \
334 .occupid = _occupid, \
337 static bool arm_cmn_is_occup_event(enum cmn_node_type type, unsigned int id)
339 return (type == CMN_TYPE_DVM && id == 0x05) ||
340 (type == CMN_TYPE_HNF && id == 0x0f);
343 static ssize_t arm_cmn_event_show(struct device *dev,
344 struct device_attribute *attr, char *buf)
346 struct arm_cmn_event_attr *eattr;
348 eattr = container_of(attr, typeof(*eattr), attr);
350 if (eattr->type == CMN_TYPE_DTC)
351 return sysfs_emit(buf, "type=0x%x\n", eattr->type);
353 if (eattr->type == CMN_TYPE_WP)
354 return sysfs_emit(buf,
355 "type=0x%x,eventid=0x%x,wp_dev_sel=?,wp_chn_sel=?,wp_grp=?,wp_val=?,wp_mask=?\n",
356 eattr->type, eattr->eventid);
358 if (arm_cmn_is_occup_event(eattr->type, eattr->eventid))
359 return sysfs_emit(buf, "type=0x%x,eventid=0x%x,occupid=0x%x\n",
360 eattr->type, eattr->eventid, eattr->occupid);
362 return sysfs_emit(buf, "type=0x%x,eventid=0x%x\n", eattr->type,
366 static umode_t arm_cmn_event_attr_is_visible(struct kobject *kobj,
367 struct attribute *attr,
370 struct device *dev = kobj_to_dev(kobj);
371 struct arm_cmn *cmn = to_cmn(dev_get_drvdata(dev));
372 struct arm_cmn_event_attr *eattr;
373 enum cmn_node_type type;
375 eattr = container_of(attr, typeof(*eattr), attr.attr);
378 /* Watchpoints aren't nodes */
379 if (type == CMN_TYPE_WP)
382 /* Revision-specific differences */
383 if (cmn->rev < CMN600_R1P2) {
384 if (type == CMN_TYPE_HNF && eattr->eventid == 0x1b)
388 if (!arm_cmn_node(cmn, type))
394 #define _CMN_EVENT_DVM(_name, _event, _occup) \
395 CMN_EVENT_ATTR(dn_##_name, CMN_TYPE_DVM, _event, _occup)
396 #define CMN_EVENT_DTC(_name) \
397 CMN_EVENT_ATTR(dtc_##_name, CMN_TYPE_DTC, 0, 0)
398 #define _CMN_EVENT_HNF(_name, _event, _occup) \
399 CMN_EVENT_ATTR(hnf_##_name, CMN_TYPE_HNF, _event, _occup)
400 #define CMN_EVENT_HNI(_name, _event) \
401 CMN_EVENT_ATTR(hni_##_name, CMN_TYPE_HNI, _event, 0)
402 #define __CMN_EVENT_XP(_name, _event) \
403 CMN_EVENT_ATTR(mxp_##_name, CMN_TYPE_XP, _event, 0)
404 #define CMN_EVENT_SBSX(_name, _event) \
405 CMN_EVENT_ATTR(sbsx_##_name, CMN_TYPE_SBSX, _event, 0)
406 #define CMN_EVENT_RNID(_name, _event) \
407 CMN_EVENT_ATTR(rnid_##_name, CMN_TYPE_RNI, _event, 0)
409 #define CMN_EVENT_DVM(_name, _event) \
410 _CMN_EVENT_DVM(_name, _event, 0)
411 #define CMN_EVENT_HNF(_name, _event) \
412 _CMN_EVENT_HNF(_name, _event, 0)
413 #define _CMN_EVENT_XP(_name, _event) \
414 __CMN_EVENT_XP(e_##_name, (_event) | (0 << 2)), \
415 __CMN_EVENT_XP(w_##_name, (_event) | (1 << 2)), \
416 __CMN_EVENT_XP(n_##_name, (_event) | (2 << 2)), \
417 __CMN_EVENT_XP(s_##_name, (_event) | (3 << 2)), \
418 __CMN_EVENT_XP(p0_##_name, (_event) | (4 << 2)), \
419 __CMN_EVENT_XP(p1_##_name, (_event) | (5 << 2))
421 /* Good thing there are only 3 fundamental XP events... */
422 #define CMN_EVENT_XP(_name, _event) \
423 _CMN_EVENT_XP(req_##_name, (_event) | (0 << 5)), \
424 _CMN_EVENT_XP(rsp_##_name, (_event) | (1 << 5)), \
425 _CMN_EVENT_XP(snp_##_name, (_event) | (2 << 5)), \
426 _CMN_EVENT_XP(dat_##_name, (_event) | (3 << 5))
429 static struct attribute *arm_cmn_event_attrs[] = {
430 CMN_EVENT_DTC(cycles),
433 * DVM node events conflict with HN-I events in the equivalent PMU
434 * slot, but our lazy short-cut of using the DTM counter index for
435 * the PMU index as well happens to avoid that by construction.
437 CMN_EVENT_DVM(rxreq_dvmop, 0x01),
438 CMN_EVENT_DVM(rxreq_dvmsync, 0x02),
439 CMN_EVENT_DVM(rxreq_dvmop_vmid_filtered, 0x03),
440 CMN_EVENT_DVM(rxreq_retried, 0x04),
441 _CMN_EVENT_DVM(rxreq_trk_occupancy_all, 0x05, 0),
442 _CMN_EVENT_DVM(rxreq_trk_occupancy_dvmop, 0x05, 1),
443 _CMN_EVENT_DVM(rxreq_trk_occupancy_dvmsync, 0x05, 2),
445 CMN_EVENT_HNF(cache_miss, 0x01),
446 CMN_EVENT_HNF(slc_sf_cache_access, 0x02),
447 CMN_EVENT_HNF(cache_fill, 0x03),
448 CMN_EVENT_HNF(pocq_retry, 0x04),
449 CMN_EVENT_HNF(pocq_reqs_recvd, 0x05),
450 CMN_EVENT_HNF(sf_hit, 0x06),
451 CMN_EVENT_HNF(sf_evictions, 0x07),
452 CMN_EVENT_HNF(dir_snoops_sent, 0x08),
453 CMN_EVENT_HNF(brd_snoops_sent, 0x09),
454 CMN_EVENT_HNF(slc_eviction, 0x0a),
455 CMN_EVENT_HNF(slc_fill_invalid_way, 0x0b),
456 CMN_EVENT_HNF(mc_retries, 0x0c),
457 CMN_EVENT_HNF(mc_reqs, 0x0d),
458 CMN_EVENT_HNF(qos_hh_retry, 0x0e),
459 _CMN_EVENT_HNF(qos_pocq_occupancy_all, 0x0f, 0),
460 _CMN_EVENT_HNF(qos_pocq_occupancy_read, 0x0f, 1),
461 _CMN_EVENT_HNF(qos_pocq_occupancy_write, 0x0f, 2),
462 _CMN_EVENT_HNF(qos_pocq_occupancy_atomic, 0x0f, 3),
463 _CMN_EVENT_HNF(qos_pocq_occupancy_stash, 0x0f, 4),
464 CMN_EVENT_HNF(pocq_addrhaz, 0x10),
465 CMN_EVENT_HNF(pocq_atomic_addrhaz, 0x11),
466 CMN_EVENT_HNF(ld_st_swp_adq_full, 0x12),
467 CMN_EVENT_HNF(cmp_adq_full, 0x13),
468 CMN_EVENT_HNF(txdat_stall, 0x14),
469 CMN_EVENT_HNF(txrsp_stall, 0x15),
470 CMN_EVENT_HNF(seq_full, 0x16),
471 CMN_EVENT_HNF(seq_hit, 0x17),
472 CMN_EVENT_HNF(snp_sent, 0x18),
473 CMN_EVENT_HNF(sfbi_dir_snp_sent, 0x19),
474 CMN_EVENT_HNF(sfbi_brd_snp_sent, 0x1a),
475 CMN_EVENT_HNF(snp_sent_untrk, 0x1b),
476 CMN_EVENT_HNF(intv_dirty, 0x1c),
477 CMN_EVENT_HNF(stash_snp_sent, 0x1d),
478 CMN_EVENT_HNF(stash_data_pull, 0x1e),
479 CMN_EVENT_HNF(snp_fwded, 0x1f),
481 CMN_EVENT_HNI(rrt_rd_occ_cnt_ovfl, 0x20),
482 CMN_EVENT_HNI(rrt_wr_occ_cnt_ovfl, 0x21),
483 CMN_EVENT_HNI(rdt_rd_occ_cnt_ovfl, 0x22),
484 CMN_EVENT_HNI(rdt_wr_occ_cnt_ovfl, 0x23),
485 CMN_EVENT_HNI(wdb_occ_cnt_ovfl, 0x24),
486 CMN_EVENT_HNI(rrt_rd_alloc, 0x25),
487 CMN_EVENT_HNI(rrt_wr_alloc, 0x26),
488 CMN_EVENT_HNI(rdt_rd_alloc, 0x27),
489 CMN_EVENT_HNI(rdt_wr_alloc, 0x28),
490 CMN_EVENT_HNI(wdb_alloc, 0x29),
491 CMN_EVENT_HNI(txrsp_retryack, 0x2a),
492 CMN_EVENT_HNI(arvalid_no_arready, 0x2b),
493 CMN_EVENT_HNI(arready_no_arvalid, 0x2c),
494 CMN_EVENT_HNI(awvalid_no_awready, 0x2d),
495 CMN_EVENT_HNI(awready_no_awvalid, 0x2e),
496 CMN_EVENT_HNI(wvalid_no_wready, 0x2f),
497 CMN_EVENT_HNI(txdat_stall, 0x30),
498 CMN_EVENT_HNI(nonpcie_serialization, 0x31),
499 CMN_EVENT_HNI(pcie_serialization, 0x32),
501 CMN_EVENT_XP(txflit_valid, 0x01),
502 CMN_EVENT_XP(txflit_stall, 0x02),
503 CMN_EVENT_XP(partial_dat_flit, 0x03),
504 /* We treat watchpoints as a special made-up class of XP events */
505 CMN_EVENT_ATTR(watchpoint_up, CMN_TYPE_WP, 0, 0),
506 CMN_EVENT_ATTR(watchpoint_down, CMN_TYPE_WP, 2, 0),
508 CMN_EVENT_SBSX(rd_req, 0x01),
509 CMN_EVENT_SBSX(wr_req, 0x02),
510 CMN_EVENT_SBSX(cmo_req, 0x03),
511 CMN_EVENT_SBSX(txrsp_retryack, 0x04),
512 CMN_EVENT_SBSX(txdat_flitv, 0x05),
513 CMN_EVENT_SBSX(txrsp_flitv, 0x06),
514 CMN_EVENT_SBSX(rd_req_trkr_occ_cnt_ovfl, 0x11),
515 CMN_EVENT_SBSX(wr_req_trkr_occ_cnt_ovfl, 0x12),
516 CMN_EVENT_SBSX(cmo_req_trkr_occ_cnt_ovfl, 0x13),
517 CMN_EVENT_SBSX(wdb_occ_cnt_ovfl, 0x14),
518 CMN_EVENT_SBSX(rd_axi_trkr_occ_cnt_ovfl, 0x15),
519 CMN_EVENT_SBSX(cmo_axi_trkr_occ_cnt_ovfl, 0x16),
520 CMN_EVENT_SBSX(arvalid_no_arready, 0x21),
521 CMN_EVENT_SBSX(awvalid_no_awready, 0x22),
522 CMN_EVENT_SBSX(wvalid_no_wready, 0x23),
523 CMN_EVENT_SBSX(txdat_stall, 0x24),
524 CMN_EVENT_SBSX(txrsp_stall, 0x25),
526 CMN_EVENT_RNID(s0_rdata_beats, 0x01),
527 CMN_EVENT_RNID(s1_rdata_beats, 0x02),
528 CMN_EVENT_RNID(s2_rdata_beats, 0x03),
529 CMN_EVENT_RNID(rxdat_flits, 0x04),
530 CMN_EVENT_RNID(txdat_flits, 0x05),
531 CMN_EVENT_RNID(txreq_flits_total, 0x06),
532 CMN_EVENT_RNID(txreq_flits_retried, 0x07),
533 CMN_EVENT_RNID(rrt_occ_ovfl, 0x08),
534 CMN_EVENT_RNID(wrt_occ_ovfl, 0x09),
535 CMN_EVENT_RNID(txreq_flits_replayed, 0x0a),
536 CMN_EVENT_RNID(wrcancel_sent, 0x0b),
537 CMN_EVENT_RNID(s0_wdata_beats, 0x0c),
538 CMN_EVENT_RNID(s1_wdata_beats, 0x0d),
539 CMN_EVENT_RNID(s2_wdata_beats, 0x0e),
540 CMN_EVENT_RNID(rrt_alloc, 0x0f),
541 CMN_EVENT_RNID(wrt_alloc, 0x10),
542 CMN_EVENT_RNID(rdb_unord, 0x11),
543 CMN_EVENT_RNID(rdb_replay, 0x12),
544 CMN_EVENT_RNID(rdb_hybrid, 0x13),
545 CMN_EVENT_RNID(rdb_ord, 0x14),
550 static const struct attribute_group arm_cmn_event_attrs_group = {
552 .attrs = arm_cmn_event_attrs,
553 .is_visible = arm_cmn_event_attr_is_visible,
556 static ssize_t arm_cmn_format_show(struct device *dev,
557 struct device_attribute *attr, char *buf)
559 struct arm_cmn_format_attr *fmt = container_of(attr, typeof(*fmt), attr);
560 int lo = __ffs(fmt->field), hi = __fls(fmt->field);
563 return sysfs_emit(buf, "config:%d\n", lo);
566 return sysfs_emit(buf, "config:%d-%d\n", lo, hi);
568 return sysfs_emit(buf, "config%d:%d-%d\n", fmt->config, lo, hi);
571 #define _CMN_FORMAT_ATTR(_name, _cfg, _fld) \
572 (&((struct arm_cmn_format_attr[]) {{ \
573 .attr = __ATTR(_name, 0444, arm_cmn_format_show, NULL), \
577 #define CMN_FORMAT_ATTR(_name, _fld) _CMN_FORMAT_ATTR(_name, 0, _fld)
579 static struct attribute *arm_cmn_format_attrs[] = {
580 CMN_FORMAT_ATTR(type, CMN_CONFIG_TYPE),
581 CMN_FORMAT_ATTR(eventid, CMN_CONFIG_EVENTID),
582 CMN_FORMAT_ATTR(occupid, CMN_CONFIG_OCCUPID),
583 CMN_FORMAT_ATTR(bynodeid, CMN_CONFIG_BYNODEID),
584 CMN_FORMAT_ATTR(nodeid, CMN_CONFIG_NODEID),
586 CMN_FORMAT_ATTR(wp_dev_sel, CMN_CONFIG_WP_DEV_SEL),
587 CMN_FORMAT_ATTR(wp_chn_sel, CMN_CONFIG_WP_CHN_SEL),
588 CMN_FORMAT_ATTR(wp_grp, CMN_CONFIG_WP_GRP),
589 CMN_FORMAT_ATTR(wp_exclusive, CMN_CONFIG_WP_EXCLUSIVE),
590 CMN_FORMAT_ATTR(wp_combine, CMN_CONFIG_WP_COMBINE),
592 _CMN_FORMAT_ATTR(wp_val, 1, CMN_CONFIG1_WP_VAL),
593 _CMN_FORMAT_ATTR(wp_mask, 2, CMN_CONFIG2_WP_MASK),
598 static const struct attribute_group arm_cmn_format_attrs_group = {
600 .attrs = arm_cmn_format_attrs,
603 static ssize_t arm_cmn_cpumask_show(struct device *dev,
604 struct device_attribute *attr, char *buf)
606 struct arm_cmn *cmn = to_cmn(dev_get_drvdata(dev));
608 return cpumap_print_to_pagebuf(true, buf, cpumask_of(cmn->cpu));
611 static struct device_attribute arm_cmn_cpumask_attr =
612 __ATTR(cpumask, 0444, arm_cmn_cpumask_show, NULL);
614 static struct attribute *arm_cmn_cpumask_attrs[] = {
615 &arm_cmn_cpumask_attr.attr,
619 static const struct attribute_group arm_cmn_cpumask_attr_group = {
620 .attrs = arm_cmn_cpumask_attrs,
623 static const struct attribute_group *arm_cmn_attr_groups[] = {
624 &arm_cmn_event_attrs_group,
625 &arm_cmn_format_attrs_group,
626 &arm_cmn_cpumask_attr_group,
630 static int arm_cmn_wp_idx(struct perf_event *event)
632 return CMN_EVENT_EVENTID(event) + CMN_EVENT_WP_GRP(event);
635 static u32 arm_cmn_wp_config(struct perf_event *event)
638 u32 dev = CMN_EVENT_WP_DEV_SEL(event);
639 u32 chn = CMN_EVENT_WP_CHN_SEL(event);
640 u32 grp = CMN_EVENT_WP_GRP(event);
641 u32 exc = CMN_EVENT_WP_EXCLUSIVE(event);
642 u32 combine = CMN_EVENT_WP_COMBINE(event);
644 config = FIELD_PREP(CMN_DTM_WPn_CONFIG_WP_DEV_SEL, dev) |
645 FIELD_PREP(CMN_DTM_WPn_CONFIG_WP_CHN_SEL, chn) |
646 FIELD_PREP(CMN_DTM_WPn_CONFIG_WP_GRP, grp) |
647 FIELD_PREP(CMN_DTM_WPn_CONFIG_WP_EXCLUSIVE, exc);
649 config |= CMN_DTM_WPn_CONFIG_WP_COMBINE;
654 static void arm_cmn_set_state(struct arm_cmn *cmn, u32 state)
657 writel_relaxed(0, cmn->dtc[0].base + CMN_DT_PMCR);
661 static void arm_cmn_clear_state(struct arm_cmn *cmn, u32 state)
663 cmn->state &= ~state;
665 writel_relaxed(CMN_DT_PMCR_PMU_EN | CMN_DT_PMCR_OVFL_INTR_EN,
666 cmn->dtc[0].base + CMN_DT_PMCR);
669 static void arm_cmn_pmu_enable(struct pmu *pmu)
671 arm_cmn_clear_state(to_cmn(pmu), CMN_STATE_DISABLED);
674 static void arm_cmn_pmu_disable(struct pmu *pmu)
676 arm_cmn_set_state(to_cmn(pmu), CMN_STATE_DISABLED);
679 static u64 arm_cmn_read_dtm(struct arm_cmn *cmn, struct arm_cmn_hw_event *hw,
682 struct arm_cmn_node *dn;
683 unsigned int i, offset;
686 offset = snapshot ? CMN_DTM_PMEVCNTSR : CMN_DTM_PMEVCNT;
687 for_each_hw_dn(hw, dn, i) {
688 struct arm_cmn_node *xp = arm_cmn_node_to_xp(dn);
689 int dtm_idx = arm_cmn_get_index(hw->dtm_idx, i);
690 u64 reg = readq_relaxed(xp->pmu_base + offset);
691 u16 dtm_count = reg >> (dtm_idx * 16);
698 static u64 arm_cmn_read_cc(struct arm_cmn_dtc *dtc)
700 u64 val = readq_relaxed(dtc->base + CMN_DT_PMCCNTR);
702 writeq_relaxed(CMN_CC_INIT, dtc->base + CMN_DT_PMCCNTR);
703 return (val - CMN_CC_INIT) & ((CMN_CC_INIT << 1) - 1);
706 static u32 arm_cmn_read_counter(struct arm_cmn_dtc *dtc, int idx)
708 u32 val, pmevcnt = CMN_DT_PMEVCNT(idx);
710 val = readl_relaxed(dtc->base + pmevcnt);
711 writel_relaxed(CMN_COUNTER_INIT, dtc->base + pmevcnt);
712 return val - CMN_COUNTER_INIT;
715 static void arm_cmn_init_counter(struct perf_event *event)
717 struct arm_cmn *cmn = to_cmn(event->pmu);
718 struct arm_cmn_hw_event *hw = to_cmn_hw(event);
719 unsigned int i, pmevcnt = CMN_DT_PMEVCNT(hw->dtc_idx);
722 for (i = 0; hw->dtcs_used & (1U << i); i++) {
723 writel_relaxed(CMN_COUNTER_INIT, cmn->dtc[i].base + pmevcnt);
724 cmn->dtc[i].counters[hw->dtc_idx] = event;
727 count = arm_cmn_read_dtm(cmn, hw, false);
728 local64_set(&event->hw.prev_count, count);
731 static void arm_cmn_event_read(struct perf_event *event)
733 struct arm_cmn *cmn = to_cmn(event->pmu);
734 struct arm_cmn_hw_event *hw = to_cmn_hw(event);
735 u64 delta, new, prev;
739 if (hw->dtc_idx == CMN_DT_NUM_COUNTERS) {
740 i = __ffs(hw->dtcs_used);
741 delta = arm_cmn_read_cc(cmn->dtc + i);
742 local64_add(delta, &event->count);
745 new = arm_cmn_read_dtm(cmn, hw, false);
746 prev = local64_xchg(&event->hw.prev_count, new);
750 local_irq_save(flags);
751 for (i = 0; hw->dtcs_used & (1U << i); i++) {
752 new = arm_cmn_read_counter(cmn->dtc + i, hw->dtc_idx);
755 local_irq_restore(flags);
756 local64_add(delta, &event->count);
759 static void arm_cmn_event_start(struct perf_event *event, int flags)
761 struct arm_cmn *cmn = to_cmn(event->pmu);
762 struct arm_cmn_hw_event *hw = to_cmn_hw(event);
763 struct arm_cmn_node *dn;
764 enum cmn_node_type type = CMN_EVENT_TYPE(event);
767 if (type == CMN_TYPE_DTC) {
768 i = __ffs(hw->dtcs_used);
769 writeq_relaxed(CMN_CC_INIT, cmn->dtc[i].base + CMN_DT_PMCCNTR);
770 cmn->dtc[i].cc_active = true;
771 } else if (type == CMN_TYPE_WP) {
772 int wp_idx = arm_cmn_wp_idx(event);
773 u64 val = CMN_EVENT_WP_VAL(event);
774 u64 mask = CMN_EVENT_WP_MASK(event);
776 for_each_hw_dn(hw, dn, i) {
777 writeq_relaxed(val, dn->pmu_base + CMN_DTM_WPn_VAL(wp_idx));
778 writeq_relaxed(mask, dn->pmu_base + CMN_DTM_WPn_MASK(wp_idx));
780 } else for_each_hw_dn(hw, dn, i) {
781 int dtm_idx = arm_cmn_get_index(hw->dtm_idx, i);
783 dn->event[dtm_idx] = CMN_EVENT_EVENTID(event);
784 writel_relaxed(le32_to_cpu(dn->event_sel), dn->pmu_base + CMN_PMU_EVENT_SEL);
788 static void arm_cmn_event_stop(struct perf_event *event, int flags)
790 struct arm_cmn *cmn = to_cmn(event->pmu);
791 struct arm_cmn_hw_event *hw = to_cmn_hw(event);
792 struct arm_cmn_node *dn;
793 enum cmn_node_type type = CMN_EVENT_TYPE(event);
796 if (type == CMN_TYPE_DTC) {
797 i = __ffs(hw->dtcs_used);
798 cmn->dtc[i].cc_active = false;
799 } else if (type == CMN_TYPE_WP) {
800 int wp_idx = arm_cmn_wp_idx(event);
802 for_each_hw_dn(hw, dn, i) {
803 writeq_relaxed(0, dn->pmu_base + CMN_DTM_WPn_MASK(wp_idx));
804 writeq_relaxed(~0ULL, dn->pmu_base + CMN_DTM_WPn_VAL(wp_idx));
806 } else for_each_hw_dn(hw, dn, i) {
807 int dtm_idx = arm_cmn_get_index(hw->dtm_idx, i);
809 dn->event[dtm_idx] = 0;
810 writel_relaxed(le32_to_cpu(dn->event_sel), dn->pmu_base + CMN_PMU_EVENT_SEL);
813 arm_cmn_event_read(event);
817 u8 dtm_count[CMN_MAX_XPS];
818 u8 occupid[CMN_MAX_XPS];
819 u8 wp[CMN_MAX_XPS][4];
824 static void arm_cmn_val_add_event(struct arm_cmn_val *val, struct perf_event *event)
826 struct arm_cmn_hw_event *hw = to_cmn_hw(event);
827 struct arm_cmn_node *dn;
828 enum cmn_node_type type;
832 if (is_software_event(event))
835 type = CMN_EVENT_TYPE(event);
836 if (type == CMN_TYPE_DTC) {
842 if (arm_cmn_is_occup_event(type, CMN_EVENT_EVENTID(event)))
843 occupid = CMN_EVENT_OCCUPID(event) + 1;
847 for_each_hw_dn(hw, dn, i) {
848 int wp_idx, xp = arm_cmn_node_to_xp(dn)->logid;
850 val->dtm_count[xp]++;
851 val->occupid[xp] = occupid;
853 if (type != CMN_TYPE_WP)
856 wp_idx = arm_cmn_wp_idx(event);
857 val->wp[xp][wp_idx] = CMN_EVENT_WP_COMBINE(event) + 1;
861 static int arm_cmn_validate_group(struct perf_event *event)
863 struct arm_cmn_hw_event *hw = to_cmn_hw(event);
864 struct arm_cmn_node *dn;
865 struct perf_event *sibling, *leader = event->group_leader;
866 enum cmn_node_type type;
867 struct arm_cmn_val val;
874 if (event->pmu != leader->pmu && !is_software_event(leader))
877 memset(&val, 0, sizeof(val));
879 arm_cmn_val_add_event(&val, leader);
880 for_each_sibling_event(sibling, leader)
881 arm_cmn_val_add_event(&val, sibling);
883 type = CMN_EVENT_TYPE(event);
884 if (type == CMN_TYPE_DTC)
885 return val.cycles ? -EINVAL : 0;
887 if (val.dtc_count == CMN_DT_NUM_COUNTERS)
890 if (arm_cmn_is_occup_event(type, CMN_EVENT_EVENTID(event)))
891 occupid = CMN_EVENT_OCCUPID(event) + 1;
895 for_each_hw_dn(hw, dn, i) {
896 int wp_idx, wp_cmb, xp = arm_cmn_node_to_xp(dn)->logid;
898 if (val.dtm_count[xp] == CMN_DTM_NUM_COUNTERS)
901 if (occupid && val.occupid[xp] && occupid != val.occupid[xp])
904 if (type != CMN_TYPE_WP)
907 wp_idx = arm_cmn_wp_idx(event);
908 if (val.wp[xp][wp_idx])
911 wp_cmb = val.wp[xp][wp_idx ^ 1];
912 if (wp_cmb && wp_cmb != CMN_EVENT_WP_COMBINE(event) + 1)
919 static int arm_cmn_event_init(struct perf_event *event)
921 struct arm_cmn *cmn = to_cmn(event->pmu);
922 struct arm_cmn_hw_event *hw = to_cmn_hw(event);
923 enum cmn_node_type type;
928 if (event->attr.type != event->pmu->type)
931 if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
934 event->cpu = cmn->cpu;
938 type = CMN_EVENT_TYPE(event);
939 /* DTC events (i.e. cycles) already have everything they need */
940 if (type == CMN_TYPE_DTC)
943 /* For watchpoints we need the actual XP node here */
944 if (type == CMN_TYPE_WP) {
946 /* ...and we need a "real" direction */
947 eventid = CMN_EVENT_EVENTID(event);
948 if (eventid != CMN_WP_UP && eventid != CMN_WP_DOWN)
952 bynodeid = CMN_EVENT_BYNODEID(event);
953 nodeid = CMN_EVENT_NODEID(event);
955 hw->dn = arm_cmn_node(cmn, type);
956 for (i = hw->dn - cmn->dns; i < cmn->num_dns && cmn->dns[i].type == type; i++) {
959 } else if (cmn->dns[i].id != nodeid) {
968 int bits = arm_cmn_xyidbits(cmn);
970 dev_dbg(cmn->dev, "invalid node 0x%x (%d,%d,%d,%d) type 0x%x\n",
971 nodeid, CMN_NODEID_X(nodeid, bits), CMN_NODEID_Y(nodeid, bits),
972 CMN_NODEID_PID(nodeid), CMN_NODEID_DEVID(nodeid), type);
976 * By assuming events count in all DTC domains, we cunningly avoid
977 * needing to know anything about how XPs are assigned to domains.
979 hw->dtcs_used = (1U << cmn->num_dtcs) - 1;
981 return arm_cmn_validate_group(event);
984 static void arm_cmn_event_clear(struct arm_cmn *cmn, struct perf_event *event,
987 struct arm_cmn_hw_event *hw = to_cmn_hw(event);
988 enum cmn_node_type type = CMN_EVENT_TYPE(event);
991 struct arm_cmn_node *xp = arm_cmn_node_to_xp(hw->dn + i);
992 unsigned int dtm_idx = arm_cmn_get_index(hw->dtm_idx, i);
994 if (type == CMN_TYPE_WP)
995 hw->dn[i].wp_event[arm_cmn_wp_idx(event)] = -1;
997 if (arm_cmn_is_occup_event(type, CMN_EVENT_EVENTID(event)))
998 hw->dn[i].occupid_count--;
1000 xp->pmu_config_low &= ~CMN__PMEVCNT_PAIRED(dtm_idx);
1001 writel_relaxed(xp->pmu_config_low, xp->pmu_base + CMN_DTM_PMU_CONFIG);
1003 memset(hw->dtm_idx, 0, sizeof(hw->dtm_idx));
1005 for (i = 0; hw->dtcs_used & (1U << i); i++)
1006 cmn->dtc[i].counters[hw->dtc_idx] = NULL;
1009 static int arm_cmn_event_add(struct perf_event *event, int flags)
1011 struct arm_cmn *cmn = to_cmn(event->pmu);
1012 struct arm_cmn_hw_event *hw = to_cmn_hw(event);
1013 struct arm_cmn_dtc *dtc = &cmn->dtc[0];
1014 struct arm_cmn_node *dn;
1015 enum cmn_node_type type = CMN_EVENT_TYPE(event);
1016 unsigned int i, dtc_idx, input_sel;
1018 if (type == CMN_TYPE_DTC) {
1020 while (cmn->dtc[i].cycles)
1021 if (++i == cmn->num_dtcs)
1024 cmn->dtc[i].cycles = event;
1025 hw->dtc_idx = CMN_DT_NUM_COUNTERS;
1026 hw->dtcs_used = 1U << i;
1028 if (flags & PERF_EF_START)
1029 arm_cmn_event_start(event, 0);
1033 /* Grab a free global counter first... */
1035 while (dtc->counters[dtc_idx])
1036 if (++dtc_idx == CMN_DT_NUM_COUNTERS)
1039 hw->dtc_idx = dtc_idx;
1041 /* ...then the local counters to feed it. */
1042 for_each_hw_dn(hw, dn, i) {
1043 struct arm_cmn_node *xp = arm_cmn_node_to_xp(dn);
1044 unsigned int dtm_idx, shift;
1048 while (xp->pmu_config_low & CMN__PMEVCNT_PAIRED(dtm_idx))
1049 if (++dtm_idx == CMN_DTM_NUM_COUNTERS)
1052 if (type == CMN_TYPE_XP) {
1053 input_sel = CMN__PMEVCNT0_INPUT_SEL_XP + dtm_idx;
1054 } else if (type == CMN_TYPE_WP) {
1055 int tmp, wp_idx = arm_cmn_wp_idx(event);
1056 u32 cfg = arm_cmn_wp_config(event);
1058 if (dn->wp_event[wp_idx] >= 0)
1061 tmp = dn->wp_event[wp_idx ^ 1];
1062 if (tmp >= 0 && CMN_EVENT_WP_COMBINE(event) !=
1063 CMN_EVENT_WP_COMBINE(dtc->counters[tmp]))
1066 input_sel = CMN__PMEVCNT0_INPUT_SEL_WP + wp_idx;
1067 dn->wp_event[wp_idx] = dtc_idx;
1068 writel_relaxed(cfg, dn->pmu_base + CMN_DTM_WPn_CONFIG(wp_idx));
1070 unsigned int port = CMN_NODEID_PID(dn->id);
1071 unsigned int dev = CMN_NODEID_DEVID(dn->id);
1073 input_sel = CMN__PMEVCNT0_INPUT_SEL_DEV + dtm_idx +
1074 (port << 4) + (dev << 2);
1076 if (arm_cmn_is_occup_event(type, CMN_EVENT_EVENTID(event))) {
1077 int occupid = CMN_EVENT_OCCUPID(event);
1079 if (dn->occupid_count == 0) {
1080 dn->occupid_val = occupid;
1081 writel_relaxed(occupid,
1082 dn->pmu_base + CMN_PMU_EVENT_SEL + 4);
1083 } else if (dn->occupid_val != occupid) {
1086 dn->occupid_count++;
1090 arm_cmn_set_index(hw->dtm_idx, i, dtm_idx);
1092 xp->input_sel[dtm_idx] = input_sel;
1093 shift = CMN__PMEVCNTn_GLOBAL_NUM_SHIFT(dtm_idx);
1094 xp->pmu_config_low &= ~(CMN__PMEVCNT0_GLOBAL_NUM << shift);
1095 xp->pmu_config_low |= FIELD_PREP(CMN__PMEVCNT0_GLOBAL_NUM, dtc_idx) << shift;
1096 xp->pmu_config_low |= CMN__PMEVCNT_PAIRED(dtm_idx);
1097 reg = (u64)le32_to_cpu(xp->pmu_config_high) << 32 | xp->pmu_config_low;
1098 writeq_relaxed(reg, xp->pmu_base + CMN_DTM_PMU_CONFIG);
1102 arm_cmn_init_counter(event);
1104 if (flags & PERF_EF_START)
1105 arm_cmn_event_start(event, 0);
1110 arm_cmn_event_clear(cmn, event, i);
1114 static void arm_cmn_event_del(struct perf_event *event, int flags)
1116 struct arm_cmn *cmn = to_cmn(event->pmu);
1117 struct arm_cmn_hw_event *hw = to_cmn_hw(event);
1118 enum cmn_node_type type = CMN_EVENT_TYPE(event);
1120 arm_cmn_event_stop(event, PERF_EF_UPDATE);
1122 if (type == CMN_TYPE_DTC)
1123 cmn->dtc[__ffs(hw->dtcs_used)].cycles = NULL;
1125 arm_cmn_event_clear(cmn, event, hw->num_dns);
1129 * We stop the PMU for both add and read, to avoid skew across DTM counters.
1130 * In theory we could use snapshots to read without stopping, but then it
1131 * becomes a lot trickier to deal with overlow and racing against interrupts,
1132 * plus it seems they don't work properly on some hardware anyway :(
1134 static void arm_cmn_start_txn(struct pmu *pmu, unsigned int flags)
1136 arm_cmn_set_state(to_cmn(pmu), CMN_STATE_TXN);
1139 static void arm_cmn_end_txn(struct pmu *pmu)
1141 arm_cmn_clear_state(to_cmn(pmu), CMN_STATE_TXN);
1144 static int arm_cmn_commit_txn(struct pmu *pmu)
1146 arm_cmn_end_txn(pmu);
1150 static int arm_cmn_pmu_offline_cpu(unsigned int cpu, struct hlist_node *node)
1152 struct arm_cmn *cmn;
1153 unsigned int i, target;
1155 cmn = hlist_entry_safe(node, struct arm_cmn, cpuhp_node);
1156 if (cpu != cmn->cpu)
1159 target = cpumask_any_but(cpu_online_mask, cpu);
1160 if (target >= nr_cpu_ids)
1163 perf_pmu_migrate_context(&cmn->pmu, cpu, target);
1164 for (i = 0; i < cmn->num_dtcs; i++)
1165 irq_set_affinity(cmn->dtc[i].irq, cpumask_of(target));
1170 static irqreturn_t arm_cmn_handle_irq(int irq, void *dev_id)
1172 struct arm_cmn_dtc *dtc = dev_id;
1173 irqreturn_t ret = IRQ_NONE;
1176 u32 status = readl_relaxed(dtc->base + CMN_DT_PMOVSR);
1180 for (i = 0; i < CMN_DTM_NUM_COUNTERS; i++) {
1181 if (status & (1U << i)) {
1183 if (WARN_ON(!dtc->counters[i]))
1185 delta = (u64)arm_cmn_read_counter(dtc, i) << 16;
1186 local64_add(delta, &dtc->counters[i]->count);
1190 if (status & (1U << CMN_DT_NUM_COUNTERS)) {
1192 if (dtc->cc_active && !WARN_ON(!dtc->cycles)) {
1193 delta = arm_cmn_read_cc(dtc);
1194 local64_add(delta, &dtc->cycles->count);
1198 writel_relaxed(status, dtc->base + CMN_DT_PMOVSR_CLR);
1200 if (!dtc->irq_friend)
1202 dtc += dtc->irq_friend;
1206 /* We can reasonably accommodate DTCs of the same CMN sharing IRQs */
1207 static int arm_cmn_init_irqs(struct arm_cmn *cmn)
1211 for (i = 0; i < cmn->num_dtcs; i++) {
1212 irq = cmn->dtc[i].irq;
1213 for (j = i; j--; ) {
1214 if (cmn->dtc[j].irq == irq) {
1215 cmn->dtc[j].irq_friend = i - j;
1219 err = devm_request_irq(cmn->dev, irq, arm_cmn_handle_irq,
1220 IRQF_NOBALANCING | IRQF_NO_THREAD,
1221 dev_name(cmn->dev), &cmn->dtc[i]);
1225 err = irq_set_affinity(irq, cpumask_of(cmn->cpu));
1229 ; /* isn't C great? */
1234 static void arm_cmn_init_dtm(struct arm_cmn_node *xp)
1238 for (i = 0; i < 4; i++) {
1239 xp->wp_event[i] = -1;
1240 writeq_relaxed(0, xp->pmu_base + CMN_DTM_WPn_MASK(i));
1241 writeq_relaxed(~0ULL, xp->pmu_base + CMN_DTM_WPn_VAL(i));
1243 xp->pmu_config_low = CMN_DTM_PMU_CONFIG_PMU_EN;
1247 static int arm_cmn_init_dtc(struct arm_cmn *cmn, struct arm_cmn_node *dn, int idx)
1249 struct arm_cmn_dtc *dtc = cmn->dtc + idx;
1250 struct arm_cmn_node *xp;
1252 dtc->base = dn->pmu_base - CMN_PMU_OFFSET;
1253 dtc->irq = platform_get_irq(to_platform_device(cmn->dev), idx);
1257 writel_relaxed(0, dtc->base + CMN_DT_PMCR);
1258 writel_relaxed(0x1ff, dtc->base + CMN_DT_PMOVSR_CLR);
1259 writel_relaxed(CMN_DT_PMCR_OVFL_INTR_EN, dtc->base + CMN_DT_PMCR);
1261 /* We do at least know that a DTC's XP must be in that DTC's domain */
1262 xp = arm_cmn_node_to_xp(dn);
1268 static int arm_cmn_node_cmp(const void *a, const void *b)
1270 const struct arm_cmn_node *dna = a, *dnb = b;
1273 cmp = dna->type - dnb->type;
1275 cmp = dna->logid - dnb->logid;
1279 static int arm_cmn_init_dtcs(struct arm_cmn *cmn)
1281 struct arm_cmn_node *dn;
1284 cmn->dtc = devm_kcalloc(cmn->dev, cmn->num_dtcs, sizeof(cmn->dtc[0]), GFP_KERNEL);
1288 sort(cmn->dns, cmn->num_dns, sizeof(cmn->dns[0]), arm_cmn_node_cmp, NULL);
1290 cmn->xps = arm_cmn_node(cmn, CMN_TYPE_XP);
1292 for (dn = cmn->dns; dn < cmn->dns + cmn->num_dns; dn++) {
1293 if (dn->type != CMN_TYPE_XP)
1294 arm_cmn_init_node_to_xp(cmn, dn);
1295 else if (cmn->num_dtcs == 1)
1298 if (dn->type == CMN_TYPE_DTC)
1299 arm_cmn_init_dtc(cmn, dn, dtc_idx++);
1301 /* To the PMU, RN-Ds don't add anything over RN-Is, so smoosh them together */
1302 if (dn->type == CMN_TYPE_RND)
1303 dn->type = CMN_TYPE_RNI;
1306 writel_relaxed(CMN_DT_DTC_CTL_DT_EN, cmn->dtc[0].base + CMN_DT_DTC_CTL);
1311 static void arm_cmn_init_node_info(struct arm_cmn *cmn, u32 offset, struct arm_cmn_node *node)
1314 u64 reg = readq_relaxed(cmn->base + offset + CMN_NODE_INFO);
1316 node->type = FIELD_GET(CMN_NI_NODE_TYPE, reg);
1317 node->id = FIELD_GET(CMN_NI_NODE_ID, reg);
1318 node->logid = FIELD_GET(CMN_NI_LOGICAL_ID, reg);
1320 node->pmu_base = cmn->base + offset + CMN_PMU_OFFSET;
1322 if (node->type == CMN_TYPE_CFG)
1324 else if (node->type == CMN_TYPE_XP)
1329 dev_dbg(cmn->dev, "node%*c%#06hx%*ctype:%-#6x id:%-4hd off:%#x\n",
1330 (level * 2) + 1, ' ', node->id, 5 - (level * 2), ' ',
1331 node->type, node->logid, offset);
1334 static int arm_cmn_discover(struct arm_cmn *cmn, unsigned int rgn_offset)
1336 void __iomem *cfg_region;
1337 struct arm_cmn_node cfg, *dn;
1338 u16 child_count, child_poff;
1339 u32 xp_offset[CMN_MAX_XPS];
1343 cfg_region = cmn->base + rgn_offset;
1344 reg = readl_relaxed(cfg_region + CMN_CFGM_PERIPH_ID_2);
1345 cmn->rev = FIELD_GET(CMN_CFGM_PID2_REVISION, reg);
1346 dev_dbg(cmn->dev, "periph_id_2 revision: %d\n", cmn->rev);
1348 arm_cmn_init_node_info(cmn, rgn_offset, &cfg);
1349 if (cfg.type != CMN_TYPE_CFG)
1352 reg = readq_relaxed(cfg_region + CMN_CHILD_INFO);
1353 child_count = FIELD_GET(CMN_CI_CHILD_COUNT, reg);
1354 child_poff = FIELD_GET(CMN_CI_CHILD_PTR_OFFSET, reg);
1356 cmn->num_xps = child_count;
1357 cmn->num_dns = cmn->num_xps;
1359 /* Pass 1: visit the XPs, enumerate their children */
1360 for (i = 0; i < cmn->num_xps; i++) {
1361 reg = readq_relaxed(cfg_region + child_poff + i * 8);
1362 xp_offset[i] = reg & CMN_CHILD_NODE_ADDR;
1364 reg = readq_relaxed(cmn->base + xp_offset[i] + CMN_CHILD_INFO);
1365 cmn->num_dns += FIELD_GET(CMN_CI_CHILD_COUNT, reg);
1368 /* Cheeky +1 to help terminate pointer-based iteration */
1369 cmn->dns = devm_kcalloc(cmn->dev, cmn->num_dns + 1,
1370 sizeof(*cmn->dns), GFP_KERNEL);
1374 /* Pass 2: now we can actually populate the nodes */
1376 for (i = 0; i < cmn->num_xps; i++) {
1377 void __iomem *xp_region = cmn->base + xp_offset[i];
1378 struct arm_cmn_node *xp = dn++;
1380 arm_cmn_init_node_info(cmn, xp_offset[i], xp);
1381 arm_cmn_init_dtm(xp);
1383 * Thanks to the order in which XP logical IDs seem to be
1384 * assigned, we can handily infer the mesh X dimension by
1385 * looking out for the XP at (0,1) without needing to know
1386 * the exact node ID format, which we can later derive.
1388 if (xp->id == (1 << 3))
1389 cmn->mesh_x = xp->logid;
1391 reg = readq_relaxed(xp_region + CMN_CHILD_INFO);
1392 child_count = FIELD_GET(CMN_CI_CHILD_COUNT, reg);
1393 child_poff = FIELD_GET(CMN_CI_CHILD_PTR_OFFSET, reg);
1395 for (j = 0; j < child_count; j++) {
1396 reg = readq_relaxed(xp_region + child_poff + j * 8);
1398 * Don't even try to touch anything external, since in general
1399 * we haven't a clue how to power up arbitrary CHI requesters.
1400 * As of CMN-600r1 these could only be RN-SAMs or CXLAs,
1401 * neither of which have any PMU events anyway.
1402 * (Actually, CXLAs do seem to have grown some events in r1p2,
1403 * but they don't go to regular XP DTMs, and they depend on
1404 * secure configuration which we can't easily deal with)
1406 if (reg & CMN_CHILD_NODE_EXTERNAL) {
1407 dev_dbg(cmn->dev, "ignoring external node %llx\n", reg);
1411 arm_cmn_init_node_info(cmn, reg & CMN_CHILD_NODE_ADDR, dn);
1418 /* These guys have PMU events */
1429 /* Nothing to see here */
1430 case CMN_TYPE_RNSAM:
1433 /* Something has gone horribly wrong */
1435 dev_err(cmn->dev, "invalid device node type: 0x%x\n", dn->type);
1441 /* Correct for any nodes we skipped */
1442 cmn->num_dns = dn - cmn->dns;
1445 * If mesh_x wasn't set during discovery then we never saw
1446 * an XP at (0,1), thus we must have an Nx1 configuration.
1449 cmn->mesh_x = cmn->num_xps;
1450 cmn->mesh_y = cmn->num_xps / cmn->mesh_x;
1452 dev_dbg(cmn->dev, "mesh %dx%d, ID width %d\n",
1453 cmn->mesh_x, cmn->mesh_y, arm_cmn_xyidbits(cmn));
1458 static int arm_cmn_acpi_probe(struct platform_device *pdev, struct arm_cmn *cmn)
1460 struct resource *cfg, *root;
1462 cfg = platform_get_resource(pdev, IORESOURCE_MEM, 0);
1466 root = platform_get_resource(pdev, IORESOURCE_MEM, 1);
1470 if (!resource_contains(cfg, root))
1473 * Note that devm_ioremap_resource() is dumb and won't let the platform
1474 * device claim cfg when the ACPI companion device has already claimed
1475 * root within it. But since they *are* already both claimed in the
1476 * appropriate name, we don't really need to do it again here anyway.
1478 cmn->base = devm_ioremap(cmn->dev, cfg->start, resource_size(cfg));
1482 return root->start - cfg->start;
1485 static int arm_cmn_of_probe(struct platform_device *pdev, struct arm_cmn *cmn)
1487 struct device_node *np = pdev->dev.of_node;
1491 cmn->base = devm_platform_ioremap_resource(pdev, 0);
1492 if (IS_ERR(cmn->base))
1493 return PTR_ERR(cmn->base);
1495 ret = of_property_read_u32(np, "arm,root-node", &rootnode);
1502 static int arm_cmn_probe(struct platform_device *pdev)
1504 struct arm_cmn *cmn;
1509 cmn = devm_kzalloc(&pdev->dev, sizeof(*cmn), GFP_KERNEL);
1513 cmn->dev = &pdev->dev;
1514 platform_set_drvdata(pdev, cmn);
1516 if (has_acpi_companion(cmn->dev))
1517 rootnode = arm_cmn_acpi_probe(pdev, cmn);
1519 rootnode = arm_cmn_of_probe(pdev, cmn);
1523 err = arm_cmn_discover(cmn, rootnode);
1527 err = arm_cmn_init_dtcs(cmn);
1531 err = arm_cmn_init_irqs(cmn);
1535 cmn->cpu = raw_smp_processor_id();
1536 cmn->pmu = (struct pmu) {
1537 .module = THIS_MODULE,
1538 .attr_groups = arm_cmn_attr_groups,
1539 .capabilities = PERF_PMU_CAP_NO_EXCLUDE,
1540 .task_ctx_nr = perf_invalid_context,
1541 .pmu_enable = arm_cmn_pmu_enable,
1542 .pmu_disable = arm_cmn_pmu_disable,
1543 .event_init = arm_cmn_event_init,
1544 .add = arm_cmn_event_add,
1545 .del = arm_cmn_event_del,
1546 .start = arm_cmn_event_start,
1547 .stop = arm_cmn_event_stop,
1548 .read = arm_cmn_event_read,
1549 .start_txn = arm_cmn_start_txn,
1550 .commit_txn = arm_cmn_commit_txn,
1551 .cancel_txn = arm_cmn_end_txn,
1554 name = devm_kasprintf(cmn->dev, GFP_KERNEL, "arm_cmn_%d", atomic_fetch_inc(&id));
1558 err = cpuhp_state_add_instance(arm_cmn_hp_state, &cmn->cpuhp_node);
1562 err = perf_pmu_register(&cmn->pmu, name, -1);
1564 cpuhp_state_remove_instance(arm_cmn_hp_state, &cmn->cpuhp_node);
1568 static int arm_cmn_remove(struct platform_device *pdev)
1570 struct arm_cmn *cmn = platform_get_drvdata(pdev);
1572 writel_relaxed(0, cmn->dtc[0].base + CMN_DT_DTC_CTL);
1574 perf_pmu_unregister(&cmn->pmu);
1575 cpuhp_state_remove_instance(arm_cmn_hp_state, &cmn->cpuhp_node);
1580 static const struct of_device_id arm_cmn_of_match[] = {
1581 { .compatible = "arm,cmn-600", },
1584 MODULE_DEVICE_TABLE(of, arm_cmn_of_match);
1588 static const struct acpi_device_id arm_cmn_acpi_match[] = {
1592 MODULE_DEVICE_TABLE(acpi, arm_cmn_acpi_match);
1595 static struct platform_driver arm_cmn_driver = {
1598 .of_match_table = of_match_ptr(arm_cmn_of_match),
1599 .acpi_match_table = ACPI_PTR(arm_cmn_acpi_match),
1601 .probe = arm_cmn_probe,
1602 .remove = arm_cmn_remove,
1605 static int __init arm_cmn_init(void)
1609 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
1610 "perf/arm/cmn:online", NULL,
1611 arm_cmn_pmu_offline_cpu);
1615 arm_cmn_hp_state = ret;
1616 ret = platform_driver_register(&arm_cmn_driver);
1618 cpuhp_remove_multi_state(arm_cmn_hp_state);
1622 static void __exit arm_cmn_exit(void)
1624 platform_driver_unregister(&arm_cmn_driver);
1625 cpuhp_remove_multi_state(arm_cmn_hp_state);
1628 module_init(arm_cmn_init);
1629 module_exit(arm_cmn_exit);
1631 MODULE_AUTHOR("Robin Murphy <robin.murphy@arm.com>");
1632 MODULE_DESCRIPTION("Arm CMN-600 PMU driver");
1633 MODULE_LICENSE("GPL v2");