xref: /openbmc/linux/drivers/perf/arm-cmn.c (revision db4919ec)
1 // SPDX-License-Identifier: GPL-2.0
2 // Copyright (C) 2016-2020 Arm Limited
3 // CMN-600 Coherent Mesh Network PMU driver
4 
5 #include <linux/acpi.h>
6 #include <linux/bitfield.h>
7 #include <linux/bitops.h>
8 #include <linux/interrupt.h>
9 #include <linux/io.h>
10 #include <linux/kernel.h>
11 #include <linux/list.h>
12 #include <linux/module.h>
13 #include <linux/of.h>
14 #include <linux/perf_event.h>
15 #include <linux/platform_device.h>
16 #include <linux/slab.h>
17 #include <linux/sort.h>
18 
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)
24 
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))
29 
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)
33 
34 #define CMN_CHILD_NODE_ADDR		GENMASK(27,0)
35 #define CMN_CHILD_NODE_EXTERNAL		BIT(31)
36 
37 #define CMN_ADDR_NODE_PTR		GENMASK(27, 14)
38 
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))
43 
44 #define CMN_MAX_XPS			(8 * 8)
45 
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)
49 
50 /* PMU registers occupy the 3rd 4KB page of each node's 16KB space */
51 #define CMN_PMU_OFFSET			0x2000
52 
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)
56 
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)
67 
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)
79 
80 #define CMN_DTM_PMEVCNT			0x220
81 
82 #define CMN_DTM_PMEVCNTSR		0x240
83 
84 #define CMN_DTM_NUM_COUNTERS		4
85 
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)
89 
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)
94 
95 #define CMN_DT_PMEVCNTSR(n)		(CMN_PMU_OFFSET + 0x50 + _CMN_DT_CNT_REG(n))
96 #define CMN_DT_PMCCNTRSR		(CMN_PMU_OFFSET + 0x90)
97 
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)
102 
103 #define CMN_DT_PMOVSR			(CMN_PMU_OFFSET + 0x118)
104 #define CMN_DT_PMOVSR_CLR		(CMN_PMU_OFFSET + 0x120)
105 
106 #define CMN_DT_PMSSR			(CMN_PMU_OFFSET + 0x128)
107 #define CMN_DT_PMSSR_SS_STATUS(n)	BIT(n)
108 
109 #define CMN_DT_PMSRR			(CMN_PMU_OFFSET + 0x130)
110 #define CMN_DT_PMSRR_SS_REQ		BIT(0)
111 
112 #define CMN_DT_NUM_COUNTERS		8
113 #define CMN_MAX_DTCS			4
114 
115 /*
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.
118  */
119 #define CMN_COUNTER_INIT		0x80000000
120 /* Similarly for the 40-bit cycle counter */
121 #define CMN_CC_INIT			0x8000000000ULL
122 
123 
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)
130 
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)
136 
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)
144 
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)
152 
153 /* Made-up event IDs for watchpoint direction */
154 #define CMN_WP_UP			0
155 #define CMN_WP_DOWN			2
156 
157 
158 /* r0px probably don't exist in silicon, thankfully */
159 enum cmn_revision {
160 	CMN600_R1P0,
161 	CMN600_R1P1,
162 	CMN600_R1P2,
163 	CMN600_R1P3,
164 	CMN600_R2P0,
165 	CMN600_R3P0,
166 };
167 
168 enum cmn_node_type {
169 	CMN_TYPE_INVALID,
170 	CMN_TYPE_DVM,
171 	CMN_TYPE_CFG,
172 	CMN_TYPE_DTC,
173 	CMN_TYPE_HNI,
174 	CMN_TYPE_HNF,
175 	CMN_TYPE_XP,
176 	CMN_TYPE_SBSX,
177 	CMN_TYPE_RNI = 0xa,
178 	CMN_TYPE_RND = 0xd,
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 */
184 	CMN_TYPE_WP = 0x7770
185 };
186 
187 struct arm_cmn_node {
188 	void __iomem *pmu_base;
189 	u16 id, logid;
190 	enum cmn_node_type type;
191 
192 	union {
193 		/* Device node */
194 		struct {
195 			int to_xp;
196 			/* DN/HN-F/CXHA */
197 			unsigned int occupid_val;
198 			unsigned int occupid_count;
199 		};
200 		/* XP */
201 		struct {
202 			int dtc;
203 			u32 pmu_config_low;
204 			union {
205 				u8 input_sel[4];
206 				__le32 pmu_config_high;
207 			};
208 			s8 wp_event[4];
209 		};
210 	};
211 
212 	union {
213 		u8 event[4];
214 		__le32 event_sel;
215 	};
216 };
217 
218 struct arm_cmn_dtc {
219 	void __iomem *base;
220 	int irq;
221 	int irq_friend;
222 	bool cc_active;
223 
224 	struct perf_event *counters[CMN_DT_NUM_COUNTERS];
225 	struct perf_event *cycles;
226 };
227 
228 #define CMN_STATE_DISABLED	BIT(0)
229 #define CMN_STATE_TXN		BIT(1)
230 
231 struct arm_cmn {
232 	struct device *dev;
233 	void __iomem *base;
234 
235 	enum cmn_revision rev;
236 	u8 mesh_x;
237 	u8 mesh_y;
238 	u16 num_xps;
239 	u16 num_dns;
240 	struct arm_cmn_node *xps;
241 	struct arm_cmn_node *dns;
242 
243 	struct arm_cmn_dtc *dtc;
244 	unsigned int num_dtcs;
245 
246 	int cpu;
247 	struct hlist_node cpuhp_node;
248 
249 	unsigned int state;
250 	struct pmu pmu;
251 };
252 
253 #define to_cmn(p)	container_of(p, struct arm_cmn, pmu)
254 
255 static int arm_cmn_hp_state;
256 
257 struct arm_cmn_hw_event {
258 	struct arm_cmn_node *dn;
259 	u64 dtm_idx[2];
260 	unsigned int dtc_idx;
261 	u8 dtcs_used;
262 	u8 num_dns;
263 };
264 
265 #define for_each_hw_dn(hw, dn, i) \
266 	for (i = 0, dn = hw->dn; i < hw->num_dns; i++, dn++)
267 
268 static struct arm_cmn_hw_event *to_cmn_hw(struct perf_event *event)
269 {
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;
272 }
273 
274 static void arm_cmn_set_index(u64 x[], unsigned int pos, unsigned int val)
275 {
276 	x[pos / 32] |= (u64)val << ((pos % 32) * 2);
277 }
278 
279 static unsigned int arm_cmn_get_index(u64 x[], unsigned int pos)
280 {
281 	return (x[pos / 32] >> ((pos % 32) * 2)) & 3;
282 }
283 
284 struct arm_cmn_event_attr {
285 	struct device_attribute attr;
286 	enum cmn_node_type type;
287 	u8 eventid;
288 	u8 occupid;
289 };
290 
291 struct arm_cmn_format_attr {
292 	struct device_attribute attr;
293 	u64 field;
294 	int config;
295 };
296 
297 static int arm_cmn_xyidbits(const struct arm_cmn *cmn)
298 {
299 	return cmn->mesh_x > 4 || cmn->mesh_y > 4 ? 3 : 2;
300 }
301 
302 static void arm_cmn_init_node_to_xp(const struct arm_cmn *cmn,
303 				    struct arm_cmn_node *dn)
304 {
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;
309 
310 	dn->to_xp = (cmn->xps + xp_idx) - dn;
311 }
312 
313 static struct arm_cmn_node *arm_cmn_node_to_xp(struct arm_cmn_node *dn)
314 {
315 	return dn->type == CMN_TYPE_XP ? dn : dn + dn->to_xp;
316 }
317 
318 static struct arm_cmn_node *arm_cmn_node(const struct arm_cmn *cmn,
319 					 enum cmn_node_type type)
320 {
321 	int i;
322 
323 	for (i = 0; i < cmn->num_dns; i++)
324 		if (cmn->dns[i].type == type)
325 			return &cmn->dns[i];
326 	return NULL;
327 }
328 
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),	\
332 		.type = _type,						\
333 		.eventid = _eventid,					\
334 		.occupid = _occupid,					\
335 	}})[0].attr.attr)
336 
337 static bool arm_cmn_is_occup_event(enum cmn_node_type type, unsigned int id)
338 {
339 	return (type == CMN_TYPE_DVM && id == 0x05) ||
340 	       (type == CMN_TYPE_HNF && id == 0x0f);
341 }
342 
343 static ssize_t arm_cmn_event_show(struct device *dev,
344 				  struct device_attribute *attr, char *buf)
345 {
346 	struct arm_cmn_event_attr *eattr;
347 
348 	eattr = container_of(attr, typeof(*eattr), attr);
349 
350 	if (eattr->type == CMN_TYPE_DTC)
351 		return snprintf(buf, PAGE_SIZE, "type=0x%x\n", eattr->type);
352 
353 	if (eattr->type == CMN_TYPE_WP)
354 		return snprintf(buf, PAGE_SIZE,
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);
357 
358 	if (arm_cmn_is_occup_event(eattr->type, eattr->eventid))
359 		return snprintf(buf, PAGE_SIZE, "type=0x%x,eventid=0x%x,occupid=0x%x\n",
360 				eattr->type, eattr->eventid, eattr->occupid);
361 
362 	return snprintf(buf, PAGE_SIZE, "type=0x%x,eventid=0x%x\n",
363 			eattr->type, eattr->eventid);
364 }
365 
366 static umode_t arm_cmn_event_attr_is_visible(struct kobject *kobj,
367 					     struct attribute *attr,
368 					     int unused)
369 {
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;
374 
375 	eattr = container_of(attr, typeof(*eattr), attr.attr);
376 	type = eattr->type;
377 
378 	/* Watchpoints aren't nodes */
379 	if (type == CMN_TYPE_WP)
380 		type = CMN_TYPE_XP;
381 
382 	/* Revision-specific differences */
383 	if (cmn->rev < CMN600_R1P2) {
384 		if (type == CMN_TYPE_HNF && eattr->eventid == 0x1b)
385 			return 0;
386 	}
387 
388 	if (!arm_cmn_node(cmn, type))
389 		return 0;
390 
391 	return attr->mode;
392 }
393 
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)
408 
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))
420 
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))
427 
428 
429 static struct attribute *arm_cmn_event_attrs[] = {
430 	CMN_EVENT_DTC(cycles),
431 
432 	/*
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.
436 	 */
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),
444 
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),
480 
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),
500 
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),
507 
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),
525 
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),
546 
547 	NULL
548 };
549 
550 static const struct attribute_group arm_cmn_event_attrs_group = {
551 	.name = "events",
552 	.attrs = arm_cmn_event_attrs,
553 	.is_visible = arm_cmn_event_attr_is_visible,
554 };
555 
556 static ssize_t arm_cmn_format_show(struct device *dev,
557 				   struct device_attribute *attr, char *buf)
558 {
559 	struct arm_cmn_format_attr *fmt = container_of(attr, typeof(*fmt), attr);
560 	int lo = __ffs(fmt->field), hi = __fls(fmt->field);
561 
562 	if (lo == hi)
563 		return snprintf(buf, PAGE_SIZE, "config:%d\n", lo);
564 
565 	if (!fmt->config)
566 		return snprintf(buf, PAGE_SIZE, "config:%d-%d\n", lo, hi);
567 
568 	return snprintf(buf, PAGE_SIZE, "config%d:%d-%d\n", fmt->config, lo, hi);
569 }
570 
571 #define _CMN_FORMAT_ATTR(_name, _cfg, _fld)				\
572 	(&((struct arm_cmn_format_attr[]) {{				\
573 		.attr = __ATTR(_name, 0444, arm_cmn_format_show, NULL),	\
574 		.config = _cfg,						\
575 		.field = _fld,						\
576 	}})[0].attr.attr)
577 #define CMN_FORMAT_ATTR(_name, _fld)	_CMN_FORMAT_ATTR(_name, 0, _fld)
578 
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),
585 
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),
591 
592 	_CMN_FORMAT_ATTR(wp_val, 1, CMN_CONFIG1_WP_VAL),
593 	_CMN_FORMAT_ATTR(wp_mask, 2, CMN_CONFIG2_WP_MASK),
594 
595 	NULL
596 };
597 
598 static const struct attribute_group arm_cmn_format_attrs_group = {
599 	.name = "format",
600 	.attrs = arm_cmn_format_attrs,
601 };
602 
603 static ssize_t arm_cmn_cpumask_show(struct device *dev,
604 				    struct device_attribute *attr, char *buf)
605 {
606 	struct arm_cmn *cmn = to_cmn(dev_get_drvdata(dev));
607 
608 	return cpumap_print_to_pagebuf(true, buf, cpumask_of(cmn->cpu));
609 }
610 
611 static struct device_attribute arm_cmn_cpumask_attr =
612 		__ATTR(cpumask, 0444, arm_cmn_cpumask_show, NULL);
613 
614 static struct attribute *arm_cmn_cpumask_attrs[] = {
615 	&arm_cmn_cpumask_attr.attr,
616 	NULL,
617 };
618 
619 static struct attribute_group arm_cmn_cpumask_attr_group = {
620 	.attrs = arm_cmn_cpumask_attrs,
621 };
622 
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,
627 	NULL
628 };
629 
630 static int arm_cmn_wp_idx(struct perf_event *event)
631 {
632 	return CMN_EVENT_EVENTID(event) + CMN_EVENT_WP_GRP(event);
633 }
634 
635 static u32 arm_cmn_wp_config(struct perf_event *event)
636 {
637 	u32 config;
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);
643 
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);
648 	if (combine && !grp)
649 		config |= CMN_DTM_WPn_CONFIG_WP_COMBINE;
650 
651 	return config;
652 }
653 
654 static void arm_cmn_set_state(struct arm_cmn *cmn, u32 state)
655 {
656 	if (!cmn->state)
657 		writel_relaxed(0, cmn->dtc[0].base + CMN_DT_PMCR);
658 	cmn->state |= state;
659 }
660 
661 static void arm_cmn_clear_state(struct arm_cmn *cmn, u32 state)
662 {
663 	cmn->state &= ~state;
664 	if (!cmn->state)
665 		writel_relaxed(CMN_DT_PMCR_PMU_EN | CMN_DT_PMCR_OVFL_INTR_EN,
666 			       cmn->dtc[0].base + CMN_DT_PMCR);
667 }
668 
669 static void arm_cmn_pmu_enable(struct pmu *pmu)
670 {
671 	arm_cmn_clear_state(to_cmn(pmu), CMN_STATE_DISABLED);
672 }
673 
674 static void arm_cmn_pmu_disable(struct pmu *pmu)
675 {
676 	arm_cmn_set_state(to_cmn(pmu), CMN_STATE_DISABLED);
677 }
678 
679 static u64 arm_cmn_read_dtm(struct arm_cmn *cmn, struct arm_cmn_hw_event *hw,
680 			    bool snapshot)
681 {
682 	struct arm_cmn_node *dn;
683 	unsigned int i, offset;
684 	u64 count = 0;
685 
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);
692 
693 		count += dtm_count;
694 	}
695 	return count;
696 }
697 
698 static u64 arm_cmn_read_cc(struct arm_cmn_dtc *dtc)
699 {
700 	u64 val = readq_relaxed(dtc->base + CMN_DT_PMCCNTR);
701 
702 	writeq_relaxed(CMN_CC_INIT, dtc->base + CMN_DT_PMCCNTR);
703 	return (val - CMN_CC_INIT) & ((CMN_CC_INIT << 1) - 1);
704 }
705 
706 static u32 arm_cmn_read_counter(struct arm_cmn_dtc *dtc, int idx)
707 {
708 	u32 val, pmevcnt = CMN_DT_PMEVCNT(idx);
709 
710 	val = readl_relaxed(dtc->base + pmevcnt);
711 	writel_relaxed(CMN_COUNTER_INIT, dtc->base + pmevcnt);
712 	return val - CMN_COUNTER_INIT;
713 }
714 
715 static void arm_cmn_init_counter(struct perf_event *event)
716 {
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);
720 	u64 count;
721 
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;
725 	}
726 
727 	count = arm_cmn_read_dtm(cmn, hw, false);
728 	local64_set(&event->hw.prev_count, count);
729 }
730 
731 static void arm_cmn_event_read(struct perf_event *event)
732 {
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;
736 	unsigned long flags;
737 	unsigned int i;
738 
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);
743 		return;
744 	}
745 	new = arm_cmn_read_dtm(cmn, hw, false);
746 	prev = local64_xchg(&event->hw.prev_count, new);
747 
748 	delta = new - prev;
749 
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);
753 		delta += new << 16;
754 	}
755 	local_irq_restore(flags);
756 	local64_add(delta, &event->count);
757 }
758 
759 static void arm_cmn_event_start(struct perf_event *event, int flags)
760 {
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);
765 	int i;
766 
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);
775 
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));
779 		}
780 	} else for_each_hw_dn(hw, dn, i) {
781 		int dtm_idx = arm_cmn_get_index(hw->dtm_idx, i);
782 
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);
785 	}
786 }
787 
788 static void arm_cmn_event_stop(struct perf_event *event, int flags)
789 {
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);
794 	int i;
795 
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);
801 
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));
805 		}
806 	} else for_each_hw_dn(hw, dn, i) {
807 		int dtm_idx = arm_cmn_get_index(hw->dtm_idx, i);
808 
809 		dn->event[dtm_idx] = 0;
810 		writel_relaxed(le32_to_cpu(dn->event_sel), dn->pmu_base + CMN_PMU_EVENT_SEL);
811 	}
812 
813 	arm_cmn_event_read(event);
814 }
815 
816 struct arm_cmn_val {
817 	u8 dtm_count[CMN_MAX_XPS];
818 	u8 occupid[CMN_MAX_XPS];
819 	u8 wp[CMN_MAX_XPS][4];
820 	int dtc_count;
821 	bool cycles;
822 };
823 
824 static void arm_cmn_val_add_event(struct arm_cmn_val *val, struct perf_event *event)
825 {
826 	struct arm_cmn_hw_event *hw = to_cmn_hw(event);
827 	struct arm_cmn_node *dn;
828 	enum cmn_node_type type;
829 	int i;
830 	u8 occupid;
831 
832 	if (is_software_event(event))
833 		return;
834 
835 	type = CMN_EVENT_TYPE(event);
836 	if (type == CMN_TYPE_DTC) {
837 		val->cycles = true;
838 		return;
839 	}
840 
841 	val->dtc_count++;
842 	if (arm_cmn_is_occup_event(type, CMN_EVENT_EVENTID(event)))
843 		occupid = CMN_EVENT_OCCUPID(event) + 1;
844 	else
845 		occupid = 0;
846 
847 	for_each_hw_dn(hw, dn, i) {
848 		int wp_idx, xp = arm_cmn_node_to_xp(dn)->logid;
849 
850 		val->dtm_count[xp]++;
851 		val->occupid[xp] = occupid;
852 
853 		if (type != CMN_TYPE_WP)
854 			continue;
855 
856 		wp_idx = arm_cmn_wp_idx(event);
857 		val->wp[xp][wp_idx] = CMN_EVENT_WP_COMBINE(event) + 1;
858 	}
859 }
860 
861 static int arm_cmn_validate_group(struct perf_event *event)
862 {
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;
868 	int i;
869 	u8 occupid;
870 
871 	if (leader == event)
872 		return 0;
873 
874 	if (event->pmu != leader->pmu && !is_software_event(leader))
875 		return -EINVAL;
876 
877 	memset(&val, 0, sizeof(val));
878 
879 	arm_cmn_val_add_event(&val, leader);
880 	for_each_sibling_event(sibling, leader)
881 		arm_cmn_val_add_event(&val, sibling);
882 
883 	type = CMN_EVENT_TYPE(event);
884 	if (type == CMN_TYPE_DTC)
885 		return val.cycles ? -EINVAL : 0;
886 
887 	if (val.dtc_count == CMN_DT_NUM_COUNTERS)
888 		return -EINVAL;
889 
890 	if (arm_cmn_is_occup_event(type, CMN_EVENT_EVENTID(event)))
891 		occupid = CMN_EVENT_OCCUPID(event) + 1;
892 	else
893 		occupid = 0;
894 
895 	for_each_hw_dn(hw, dn, i) {
896 		int wp_idx, wp_cmb, xp = arm_cmn_node_to_xp(dn)->logid;
897 
898 		if (val.dtm_count[xp] == CMN_DTM_NUM_COUNTERS)
899 			return -EINVAL;
900 
901 		if (occupid && val.occupid[xp] && occupid != val.occupid[xp])
902 			return -EINVAL;
903 
904 		if (type != CMN_TYPE_WP)
905 			continue;
906 
907 		wp_idx = arm_cmn_wp_idx(event);
908 		if (val.wp[xp][wp_idx])
909 			return -EINVAL;
910 
911 		wp_cmb = val.wp[xp][wp_idx ^ 1];
912 		if (wp_cmb && wp_cmb != CMN_EVENT_WP_COMBINE(event) + 1)
913 			return -EINVAL;
914 	}
915 
916 	return 0;
917 }
918 
919 static int arm_cmn_event_init(struct perf_event *event)
920 {
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;
924 	unsigned int i;
925 	bool bynodeid;
926 	u16 nodeid, eventid;
927 
928 	if (event->attr.type != event->pmu->type)
929 		return -ENOENT;
930 
931 	if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
932 		return -EINVAL;
933 
934 	event->cpu = cmn->cpu;
935 	if (event->cpu < 0)
936 		return -EINVAL;
937 
938 	type = CMN_EVENT_TYPE(event);
939 	/* DTC events (i.e. cycles) already have everything they need */
940 	if (type == CMN_TYPE_DTC)
941 		return 0;
942 
943 	/* For watchpoints we need the actual XP node here */
944 	if (type == CMN_TYPE_WP) {
945 		type = CMN_TYPE_XP;
946 		/* ...and we need a "real" direction */
947 		eventid = CMN_EVENT_EVENTID(event);
948 		if (eventid != CMN_WP_UP && eventid != CMN_WP_DOWN)
949 			return -EINVAL;
950 	}
951 
952 	bynodeid = CMN_EVENT_BYNODEID(event);
953 	nodeid = CMN_EVENT_NODEID(event);
954 
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++) {
957 		if (!bynodeid) {
958 			hw->num_dns++;
959 		} else if (cmn->dns[i].id != nodeid) {
960 			hw->dn++;
961 		} else {
962 			hw->num_dns = 1;
963 			break;
964 		}
965 	}
966 
967 	if (!hw->num_dns) {
968 		int bits = arm_cmn_xyidbits(cmn);
969 
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);
973 		return -EINVAL;
974 	}
975 	/*
976 	 * By assuming events count in all DTC domains, we cunningly avoid
977 	 * needing to know anything about how XPs are assigned to domains.
978 	 */
979 	hw->dtcs_used = (1U << cmn->num_dtcs) - 1;
980 
981 	return arm_cmn_validate_group(event);
982 }
983 
984 static void arm_cmn_event_clear(struct arm_cmn *cmn, struct perf_event *event,
985 				int i)
986 {
987 	struct arm_cmn_hw_event *hw = to_cmn_hw(event);
988 	enum cmn_node_type type = CMN_EVENT_TYPE(event);
989 
990 	while (i--) {
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);
993 
994 		if (type == CMN_TYPE_WP)
995 			hw->dn[i].wp_event[arm_cmn_wp_idx(event)] = -1;
996 
997 		if (arm_cmn_is_occup_event(type, CMN_EVENT_EVENTID(event)))
998 			hw->dn[i].occupid_count--;
999 
1000 		xp->pmu_config_low &= ~CMN__PMEVCNT_PAIRED(dtm_idx);
1001 		writel_relaxed(xp->pmu_config_low, xp->pmu_base + CMN_DTM_PMU_CONFIG);
1002 	}
1003 	memset(hw->dtm_idx, 0, sizeof(hw->dtm_idx));
1004 
1005 	for (i = 0; hw->dtcs_used & (1U << i); i++)
1006 		cmn->dtc[i].counters[hw->dtc_idx] = NULL;
1007 }
1008 
1009 static int arm_cmn_event_add(struct perf_event *event, int flags)
1010 {
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;
1017 
1018 	if (type == CMN_TYPE_DTC) {
1019 		i = 0;
1020 		while (cmn->dtc[i].cycles)
1021 			if (++i == cmn->num_dtcs)
1022 				return -ENOSPC;
1023 
1024 		cmn->dtc[i].cycles = event;
1025 		hw->dtc_idx = CMN_DT_NUM_COUNTERS;
1026 		hw->dtcs_used = 1U << i;
1027 
1028 		if (flags & PERF_EF_START)
1029 			arm_cmn_event_start(event, 0);
1030 		return 0;
1031 	}
1032 
1033 	/* Grab a free global counter first... */
1034 	dtc_idx = 0;
1035 	while (dtc->counters[dtc_idx])
1036 		if (++dtc_idx == CMN_DT_NUM_COUNTERS)
1037 			return -ENOSPC;
1038 
1039 	hw->dtc_idx = dtc_idx;
1040 
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;
1045 		u64 reg;
1046 
1047 		dtm_idx = 0;
1048 		while (xp->pmu_config_low & CMN__PMEVCNT_PAIRED(dtm_idx))
1049 			if (++dtm_idx == CMN_DTM_NUM_COUNTERS)
1050 				goto free_dtms;
1051 
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);
1057 
1058 			if (dn->wp_event[wp_idx] >= 0)
1059 				goto free_dtms;
1060 
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]))
1064 				goto free_dtms;
1065 
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));
1069 		} else {
1070 			unsigned int port = CMN_NODEID_PID(dn->id);
1071 			unsigned int dev = CMN_NODEID_DEVID(dn->id);
1072 
1073 			input_sel = CMN__PMEVCNT0_INPUT_SEL_DEV + dtm_idx +
1074 				    (port << 4) + (dev << 2);
1075 
1076 			if (arm_cmn_is_occup_event(type, CMN_EVENT_EVENTID(event))) {
1077 				int occupid = CMN_EVENT_OCCUPID(event);
1078 
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) {
1084 					goto free_dtms;
1085 				}
1086 				dn->occupid_count++;
1087 			}
1088 		}
1089 
1090 		arm_cmn_set_index(hw->dtm_idx, i, dtm_idx);
1091 
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);
1099 	}
1100 
1101 	/* Go go go! */
1102 	arm_cmn_init_counter(event);
1103 
1104 	if (flags & PERF_EF_START)
1105 		arm_cmn_event_start(event, 0);
1106 
1107 	return 0;
1108 
1109 free_dtms:
1110 	arm_cmn_event_clear(cmn, event, i);
1111 	return -ENOSPC;
1112 }
1113 
1114 static void arm_cmn_event_del(struct perf_event *event, int flags)
1115 {
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);
1119 
1120 	arm_cmn_event_stop(event, PERF_EF_UPDATE);
1121 
1122 	if (type == CMN_TYPE_DTC)
1123 		cmn->dtc[__ffs(hw->dtcs_used)].cycles = NULL;
1124 	else
1125 		arm_cmn_event_clear(cmn, event, hw->num_dns);
1126 }
1127 
1128 /*
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 :(
1133  */
1134 static void arm_cmn_start_txn(struct pmu *pmu, unsigned int flags)
1135 {
1136 	arm_cmn_set_state(to_cmn(pmu), CMN_STATE_TXN);
1137 }
1138 
1139 static void arm_cmn_end_txn(struct pmu *pmu)
1140 {
1141 	arm_cmn_clear_state(to_cmn(pmu), CMN_STATE_TXN);
1142 }
1143 
1144 static int arm_cmn_commit_txn(struct pmu *pmu)
1145 {
1146 	arm_cmn_end_txn(pmu);
1147 	return 0;
1148 }
1149 
1150 static int arm_cmn_pmu_offline_cpu(unsigned int cpu, struct hlist_node *node)
1151 {
1152 	struct arm_cmn *cmn;
1153 	unsigned int target;
1154 
1155 	cmn = hlist_entry_safe(node, struct arm_cmn, cpuhp_node);
1156 	if (cpu != cmn->cpu)
1157 		return 0;
1158 
1159 	target = cpumask_any_but(cpu_online_mask, cpu);
1160 	if (target >= nr_cpu_ids)
1161 		return 0;
1162 
1163 	perf_pmu_migrate_context(&cmn->pmu, cpu, target);
1164 	cmn->cpu = target;
1165 	return 0;
1166 }
1167 
1168 static irqreturn_t arm_cmn_handle_irq(int irq, void *dev_id)
1169 {
1170 	struct arm_cmn_dtc *dtc = dev_id;
1171 	irqreturn_t ret = IRQ_NONE;
1172 
1173 	for (;;) {
1174 		u32 status = readl_relaxed(dtc->base + CMN_DT_PMOVSR);
1175 		u64 delta;
1176 		int i;
1177 
1178 		for (i = 0; i < CMN_DTM_NUM_COUNTERS; i++) {
1179 			if (status & (1U << i)) {
1180 				ret = IRQ_HANDLED;
1181 				if (WARN_ON(!dtc->counters[i]))
1182 					continue;
1183 				delta = (u64)arm_cmn_read_counter(dtc, i) << 16;
1184 				local64_add(delta, &dtc->counters[i]->count);
1185 			}
1186 		}
1187 
1188 		if (status & (1U << CMN_DT_NUM_COUNTERS)) {
1189 			ret = IRQ_HANDLED;
1190 			if (dtc->cc_active && !WARN_ON(!dtc->cycles)) {
1191 				delta = arm_cmn_read_cc(dtc);
1192 				local64_add(delta, &dtc->cycles->count);
1193 			}
1194 		}
1195 
1196 		writel_relaxed(status, dtc->base + CMN_DT_PMOVSR_CLR);
1197 
1198 		if (!dtc->irq_friend)
1199 			return ret;
1200 		dtc += dtc->irq_friend;
1201 	}
1202 }
1203 
1204 /* We can reasonably accommodate DTCs of the same CMN sharing IRQs */
1205 static int arm_cmn_init_irqs(struct arm_cmn *cmn)
1206 {
1207 	int i, j, irq, err;
1208 
1209 	for (i = 0; i < cmn->num_dtcs; i++) {
1210 		irq = cmn->dtc[i].irq;
1211 		for (j = i; j--; ) {
1212 			if (cmn->dtc[j].irq == irq) {
1213 				cmn->dtc[j].irq_friend = j - i;
1214 				goto next;
1215 			}
1216 		}
1217 		err = devm_request_irq(cmn->dev, irq, arm_cmn_handle_irq,
1218 				       IRQF_NOBALANCING | IRQF_NO_THREAD,
1219 				       dev_name(cmn->dev), &cmn->dtc[i]);
1220 		if (err)
1221 			return err;
1222 
1223 		err = irq_set_affinity_hint(irq, cpumask_of(cmn->cpu));
1224 		if (err)
1225 			return err;
1226 	next:
1227 		; /* isn't C great? */
1228 	}
1229 	return 0;
1230 }
1231 
1232 static void arm_cmn_init_dtm(struct arm_cmn_node *xp)
1233 {
1234 	int i;
1235 
1236 	for (i = 0; i < 4; i++) {
1237 		xp->wp_event[i] = -1;
1238 		writeq_relaxed(0, xp->pmu_base + CMN_DTM_WPn_MASK(i));
1239 		writeq_relaxed(~0ULL, xp->pmu_base + CMN_DTM_WPn_VAL(i));
1240 	}
1241 	xp->pmu_config_low = CMN_DTM_PMU_CONFIG_PMU_EN;
1242 	xp->dtc = -1;
1243 }
1244 
1245 static int arm_cmn_init_dtc(struct arm_cmn *cmn, struct arm_cmn_node *dn, int idx)
1246 {
1247 	struct arm_cmn_dtc *dtc = cmn->dtc + idx;
1248 	struct arm_cmn_node *xp;
1249 
1250 	dtc->base = dn->pmu_base - CMN_PMU_OFFSET;
1251 	dtc->irq = platform_get_irq(to_platform_device(cmn->dev), idx);
1252 	if (dtc->irq < 0)
1253 		return dtc->irq;
1254 
1255 	writel_relaxed(0, dtc->base + CMN_DT_PMCR);
1256 	writel_relaxed(0x1ff, dtc->base + CMN_DT_PMOVSR_CLR);
1257 	writel_relaxed(CMN_DT_PMCR_OVFL_INTR_EN, dtc->base + CMN_DT_PMCR);
1258 
1259 	/* We do at least know that a DTC's XP must be in that DTC's domain */
1260 	xp = arm_cmn_node_to_xp(dn);
1261 	xp->dtc = idx;
1262 
1263 	return 0;
1264 }
1265 
1266 static int arm_cmn_node_cmp(const void *a, const void *b)
1267 {
1268 	const struct arm_cmn_node *dna = a, *dnb = b;
1269 	int cmp;
1270 
1271 	cmp = dna->type - dnb->type;
1272 	if (!cmp)
1273 		cmp = dna->logid - dnb->logid;
1274 	return cmp;
1275 }
1276 
1277 static int arm_cmn_init_dtcs(struct arm_cmn *cmn)
1278 {
1279 	struct arm_cmn_node *dn;
1280 	int dtc_idx = 0;
1281 
1282 	cmn->dtc = devm_kcalloc(cmn->dev, cmn->num_dtcs, sizeof(cmn->dtc[0]), GFP_KERNEL);
1283 	if (!cmn->dtc)
1284 		return -ENOMEM;
1285 
1286 	sort(cmn->dns, cmn->num_dns, sizeof(cmn->dns[0]), arm_cmn_node_cmp, NULL);
1287 
1288 	cmn->xps = arm_cmn_node(cmn, CMN_TYPE_XP);
1289 
1290 	for (dn = cmn->dns; dn < cmn->dns + cmn->num_dns; dn++) {
1291 		if (dn->type != CMN_TYPE_XP)
1292 			arm_cmn_init_node_to_xp(cmn, dn);
1293 		else if (cmn->num_dtcs == 1)
1294 			dn->dtc = 0;
1295 
1296 		if (dn->type == CMN_TYPE_DTC)
1297 			arm_cmn_init_dtc(cmn, dn, dtc_idx++);
1298 
1299 		/* To the PMU, RN-Ds don't add anything over RN-Is, so smoosh them together */
1300 		if (dn->type == CMN_TYPE_RND)
1301 			dn->type = CMN_TYPE_RNI;
1302 	}
1303 
1304 	writel_relaxed(CMN_DT_DTC_CTL_DT_EN, cmn->dtc[0].base + CMN_DT_DTC_CTL);
1305 
1306 	return 0;
1307 }
1308 
1309 static void arm_cmn_init_node_info(struct arm_cmn *cmn, u32 offset, struct arm_cmn_node *node)
1310 {
1311 	int level;
1312 	u64 reg = readq_relaxed(cmn->base + offset + CMN_NODE_INFO);
1313 
1314 	node->type = FIELD_GET(CMN_NI_NODE_TYPE, reg);
1315 	node->id = FIELD_GET(CMN_NI_NODE_ID, reg);
1316 	node->logid = FIELD_GET(CMN_NI_LOGICAL_ID, reg);
1317 
1318 	node->pmu_base = cmn->base + offset + CMN_PMU_OFFSET;
1319 
1320 	if (node->type == CMN_TYPE_CFG)
1321 		level = 0;
1322 	else if (node->type == CMN_TYPE_XP)
1323 		level = 1;
1324 	else
1325 		level = 2;
1326 
1327 	dev_dbg(cmn->dev, "node%*c%#06hx%*ctype:%-#6x id:%-4hd off:%#x\n",
1328 			(level * 2) + 1, ' ', node->id, 5 - (level * 2), ' ',
1329 			node->type, node->logid, offset);
1330 }
1331 
1332 static int arm_cmn_discover(struct arm_cmn *cmn, unsigned int rgn_offset)
1333 {
1334 	void __iomem *cfg_region;
1335 	struct arm_cmn_node cfg, *dn;
1336 	u16 child_count, child_poff;
1337 	u32 xp_offset[CMN_MAX_XPS];
1338 	u64 reg;
1339 	int i, j;
1340 
1341 	cfg_region = cmn->base + rgn_offset;
1342 	reg = readl_relaxed(cfg_region + CMN_CFGM_PERIPH_ID_2);
1343 	cmn->rev = FIELD_GET(CMN_CFGM_PID2_REVISION, reg);
1344 	dev_dbg(cmn->dev, "periph_id_2 revision: %d\n", cmn->rev);
1345 
1346 	arm_cmn_init_node_info(cmn, rgn_offset, &cfg);
1347 	if (cfg.type != CMN_TYPE_CFG)
1348 		return -ENODEV;
1349 
1350 	reg = readq_relaxed(cfg_region + CMN_CHILD_INFO);
1351 	child_count = FIELD_GET(CMN_CI_CHILD_COUNT, reg);
1352 	child_poff = FIELD_GET(CMN_CI_CHILD_PTR_OFFSET, reg);
1353 
1354 	cmn->num_xps = child_count;
1355 	cmn->num_dns = cmn->num_xps;
1356 
1357 	/* Pass 1: visit the XPs, enumerate their children */
1358 	for (i = 0; i < cmn->num_xps; i++) {
1359 		reg = readq_relaxed(cfg_region + child_poff + i * 8);
1360 		xp_offset[i] = reg & CMN_CHILD_NODE_ADDR;
1361 
1362 		reg = readq_relaxed(cmn->base + xp_offset[i] + CMN_CHILD_INFO);
1363 		cmn->num_dns += FIELD_GET(CMN_CI_CHILD_COUNT, reg);
1364 	}
1365 
1366 	/* Cheeky +1 to help terminate pointer-based iteration */
1367 	cmn->dns = devm_kcalloc(cmn->dev, cmn->num_dns + 1,
1368 				sizeof(*cmn->dns), GFP_KERNEL);
1369 	if (!cmn->dns)
1370 		return -ENOMEM;
1371 
1372 	/* Pass 2: now we can actually populate the nodes */
1373 	dn = cmn->dns;
1374 	for (i = 0; i < cmn->num_xps; i++) {
1375 		void __iomem *xp_region = cmn->base + xp_offset[i];
1376 		struct arm_cmn_node *xp = dn++;
1377 
1378 		arm_cmn_init_node_info(cmn, xp_offset[i], xp);
1379 		arm_cmn_init_dtm(xp);
1380 		/*
1381 		 * Thanks to the order in which XP logical IDs seem to be
1382 		 * assigned, we can handily infer the mesh X dimension by
1383 		 * looking out for the XP at (0,1) without needing to know
1384 		 * the exact node ID format, which we can later derive.
1385 		 */
1386 		if (xp->id == (1 << 3))
1387 			cmn->mesh_x = xp->logid;
1388 
1389 		reg = readq_relaxed(xp_region + CMN_CHILD_INFO);
1390 		child_count = FIELD_GET(CMN_CI_CHILD_COUNT, reg);
1391 		child_poff = FIELD_GET(CMN_CI_CHILD_PTR_OFFSET, reg);
1392 
1393 		for (j = 0; j < child_count; j++) {
1394 			reg = readq_relaxed(xp_region + child_poff + j * 8);
1395 			/*
1396 			 * Don't even try to touch anything external, since in general
1397 			 * we haven't a clue how to power up arbitrary CHI requesters.
1398 			 * As of CMN-600r1 these could only be RN-SAMs or CXLAs,
1399 			 * neither of which have any PMU events anyway.
1400 			 * (Actually, CXLAs do seem to have grown some events in r1p2,
1401 			 * but they don't go to regular XP DTMs, and they depend on
1402 			 * secure configuration which we can't easily deal with)
1403 			 */
1404 			if (reg & CMN_CHILD_NODE_EXTERNAL) {
1405 				dev_dbg(cmn->dev, "ignoring external node %llx\n", reg);
1406 				continue;
1407 			}
1408 
1409 			arm_cmn_init_node_info(cmn, reg & CMN_CHILD_NODE_ADDR, dn);
1410 
1411 			switch (dn->type) {
1412 			case CMN_TYPE_DTC:
1413 				cmn->num_dtcs++;
1414 				dn++;
1415 				break;
1416 			/* These guys have PMU events */
1417 			case CMN_TYPE_DVM:
1418 			case CMN_TYPE_HNI:
1419 			case CMN_TYPE_HNF:
1420 			case CMN_TYPE_SBSX:
1421 			case CMN_TYPE_RNI:
1422 			case CMN_TYPE_RND:
1423 			case CMN_TYPE_CXRA:
1424 			case CMN_TYPE_CXHA:
1425 				dn++;
1426 				break;
1427 			/* Nothing to see here */
1428 			case CMN_TYPE_RNSAM:
1429 			case CMN_TYPE_CXLA:
1430 				break;
1431 			/* Something has gone horribly wrong */
1432 			default:
1433 				dev_err(cmn->dev, "invalid device node type: 0x%x\n", dn->type);
1434 				return -ENODEV;
1435 			}
1436 		}
1437 	}
1438 
1439 	/* Correct for any nodes we skipped */
1440 	cmn->num_dns = dn - cmn->dns;
1441 
1442 	/*
1443 	 * If mesh_x wasn't set during discovery then we never saw
1444 	 * an XP at (0,1), thus we must have an Nx1 configuration.
1445 	 */
1446 	if (!cmn->mesh_x)
1447 		cmn->mesh_x = cmn->num_xps;
1448 	cmn->mesh_y = cmn->num_xps / cmn->mesh_x;
1449 
1450 	dev_dbg(cmn->dev, "mesh %dx%d, ID width %d\n",
1451 		cmn->mesh_x, cmn->mesh_y, arm_cmn_xyidbits(cmn));
1452 
1453 	return 0;
1454 }
1455 
1456 static int arm_cmn_acpi_probe(struct platform_device *pdev, struct arm_cmn *cmn)
1457 {
1458 	struct resource *cfg, *root;
1459 
1460 	cfg = platform_get_resource(pdev, IORESOURCE_MEM, 0);
1461 	if (!cfg)
1462 		return -EINVAL;
1463 
1464 	root = platform_get_resource(pdev, IORESOURCE_MEM, 1);
1465 	if (!root)
1466 		return -EINVAL;
1467 
1468 	if (!resource_contains(cfg, root))
1469 		swap(cfg, root);
1470 	/*
1471 	 * Note that devm_ioremap_resource() is dumb and won't let the platform
1472 	 * device claim cfg when the ACPI companion device has already claimed
1473 	 * root within it. But since they *are* already both claimed in the
1474 	 * appropriate name, we don't really need to do it again here anyway.
1475 	 */
1476 	cmn->base = devm_ioremap(cmn->dev, cfg->start, resource_size(cfg));
1477 	if (!cmn->base)
1478 		return -ENOMEM;
1479 
1480 	return root->start - cfg->start;
1481 }
1482 
1483 static int arm_cmn_of_probe(struct platform_device *pdev, struct arm_cmn *cmn)
1484 {
1485 	struct device_node *np = pdev->dev.of_node;
1486 	u32 rootnode;
1487 	int ret;
1488 
1489 	cmn->base = devm_platform_ioremap_resource(pdev, 0);
1490 	if (IS_ERR(cmn->base))
1491 		return PTR_ERR(cmn->base);
1492 
1493 	ret = of_property_read_u32(np, "arm,root-node", &rootnode);
1494 	if (ret)
1495 		return ret;
1496 
1497 	return rootnode;
1498 }
1499 
1500 static int arm_cmn_probe(struct platform_device *pdev)
1501 {
1502 	struct arm_cmn *cmn;
1503 	const char *name;
1504 	static atomic_t id;
1505 	int err, rootnode, this_id;
1506 
1507 	cmn = devm_kzalloc(&pdev->dev, sizeof(*cmn), GFP_KERNEL);
1508 	if (!cmn)
1509 		return -ENOMEM;
1510 
1511 	cmn->dev = &pdev->dev;
1512 	platform_set_drvdata(pdev, cmn);
1513 
1514 	if (has_acpi_companion(cmn->dev))
1515 		rootnode = arm_cmn_acpi_probe(pdev, cmn);
1516 	else
1517 		rootnode = arm_cmn_of_probe(pdev, cmn);
1518 	if (rootnode < 0)
1519 		return rootnode;
1520 
1521 	err = arm_cmn_discover(cmn, rootnode);
1522 	if (err)
1523 		return err;
1524 
1525 	err = arm_cmn_init_dtcs(cmn);
1526 	if (err)
1527 		return err;
1528 
1529 	err = arm_cmn_init_irqs(cmn);
1530 	if (err)
1531 		return err;
1532 
1533 	cmn->cpu = raw_smp_processor_id();
1534 	cmn->pmu = (struct pmu) {
1535 		.module = THIS_MODULE,
1536 		.attr_groups = arm_cmn_attr_groups,
1537 		.capabilities = PERF_PMU_CAP_NO_EXCLUDE,
1538 		.task_ctx_nr = perf_invalid_context,
1539 		.pmu_enable = arm_cmn_pmu_enable,
1540 		.pmu_disable = arm_cmn_pmu_disable,
1541 		.event_init = arm_cmn_event_init,
1542 		.add = arm_cmn_event_add,
1543 		.del = arm_cmn_event_del,
1544 		.start = arm_cmn_event_start,
1545 		.stop = arm_cmn_event_stop,
1546 		.read = arm_cmn_event_read,
1547 		.start_txn = arm_cmn_start_txn,
1548 		.commit_txn = arm_cmn_commit_txn,
1549 		.cancel_txn = arm_cmn_end_txn,
1550 	};
1551 
1552 	this_id = atomic_fetch_inc(&id);
1553 	if (this_id == 0) {
1554 		name = "arm_cmn";
1555 	} else {
1556 		name = devm_kasprintf(cmn->dev, GFP_KERNEL, "arm_cmn_%d", this_id);
1557 		if (!name)
1558 			return -ENOMEM;
1559 	}
1560 
1561 	err = cpuhp_state_add_instance(arm_cmn_hp_state, &cmn->cpuhp_node);
1562 	if (err)
1563 		return err;
1564 
1565 	err = perf_pmu_register(&cmn->pmu, name, -1);
1566 	if (err)
1567 		cpuhp_state_remove_instance(arm_cmn_hp_state, &cmn->cpuhp_node);
1568 	return err;
1569 }
1570 
1571 static int arm_cmn_remove(struct platform_device *pdev)
1572 {
1573 	struct arm_cmn *cmn = platform_get_drvdata(pdev);
1574 	int i;
1575 
1576 	writel_relaxed(0, cmn->dtc[0].base + CMN_DT_DTC_CTL);
1577 
1578 	perf_pmu_unregister(&cmn->pmu);
1579 	cpuhp_state_remove_instance(arm_cmn_hp_state, &cmn->cpuhp_node);
1580 
1581 	for (i = 0; i < cmn->num_dtcs; i++)
1582 		irq_set_affinity_hint(cmn->dtc[i].irq, NULL);
1583 
1584 	return 0;
1585 }
1586 
1587 #ifdef CONFIG_OF
1588 static const struct of_device_id arm_cmn_of_match[] = {
1589 	{ .compatible = "arm,cmn-600", },
1590 	{}
1591 };
1592 MODULE_DEVICE_TABLE(of, arm_cmn_of_match);
1593 #endif
1594 
1595 #ifdef CONFIG_ACPI
1596 static const struct acpi_device_id arm_cmn_acpi_match[] = {
1597 	{ "ARMHC600", },
1598 	{}
1599 };
1600 MODULE_DEVICE_TABLE(acpi, arm_cmn_acpi_match);
1601 #endif
1602 
1603 static struct platform_driver arm_cmn_driver = {
1604 	.driver = {
1605 		.name = "arm-cmn",
1606 		.of_match_table = of_match_ptr(arm_cmn_of_match),
1607 		.acpi_match_table = ACPI_PTR(arm_cmn_acpi_match),
1608 	},
1609 	.probe = arm_cmn_probe,
1610 	.remove = arm_cmn_remove,
1611 };
1612 
1613 static int __init arm_cmn_init(void)
1614 {
1615 	int ret;
1616 
1617 	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
1618 				      "perf/arm/cmn:online", NULL,
1619 				      arm_cmn_pmu_offline_cpu);
1620 	if (ret < 0)
1621 		return ret;
1622 
1623 	arm_cmn_hp_state = ret;
1624 	ret = platform_driver_register(&arm_cmn_driver);
1625 	if (ret)
1626 		cpuhp_remove_multi_state(arm_cmn_hp_state);
1627 	return ret;
1628 }
1629 
1630 static void __exit arm_cmn_exit(void)
1631 {
1632 	platform_driver_unregister(&arm_cmn_driver);
1633 	cpuhp_remove_multi_state(arm_cmn_hp_state);
1634 }
1635 
1636 module_init(arm_cmn_init);
1637 module_exit(arm_cmn_exit);
1638 
1639 MODULE_AUTHOR("Robin Murphy <robin.murphy@arm.com>");
1640 MODULE_DESCRIPTION("Arm CMN-600 PMU driver");
1641 MODULE_LICENSE("GPL v2");
1642