xref: /linux/drivers/perf/arm-ccn.c (revision 7f71507851fc7764b36a3221839607d3a45c2025)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *
4  * Copyright (C) 2014 ARM Limited
5  */
6 
7 #include <linux/ctype.h>
8 #include <linux/hrtimer.h>
9 #include <linux/idr.h>
10 #include <linux/interrupt.h>
11 #include <linux/io.h>
12 #include <linux/module.h>
13 #include <linux/mod_devicetable.h>
14 #include <linux/perf_event.h>
15 #include <linux/platform_device.h>
16 #include <linux/slab.h>
17 
18 #define CCN_NUM_XP_PORTS 2
19 #define CCN_NUM_VCS 4
20 #define CCN_NUM_REGIONS	256
21 #define CCN_REGION_SIZE	0x10000
22 
23 #define CCN_ALL_OLY_ID			0xff00
24 #define CCN_ALL_OLY_ID__OLY_ID__SHIFT			0
25 #define CCN_ALL_OLY_ID__OLY_ID__MASK			0x1f
26 #define CCN_ALL_OLY_ID__NODE_ID__SHIFT			8
27 #define CCN_ALL_OLY_ID__NODE_ID__MASK			0x3f
28 
29 #define CCN_MN_ERRINT_STATUS		0x0008
30 #define CCN_MN_ERRINT_STATUS__INTREQ__DESSERT		0x11
31 #define CCN_MN_ERRINT_STATUS__ALL_ERRORS__ENABLE	0x02
32 #define CCN_MN_ERRINT_STATUS__ALL_ERRORS__DISABLED	0x20
33 #define CCN_MN_ERRINT_STATUS__ALL_ERRORS__DISABLE	0x22
34 #define CCN_MN_ERRINT_STATUS__CORRECTED_ERRORS_ENABLE	0x04
35 #define CCN_MN_ERRINT_STATUS__CORRECTED_ERRORS_DISABLED	0x40
36 #define CCN_MN_ERRINT_STATUS__CORRECTED_ERRORS_DISABLE	0x44
37 #define CCN_MN_ERRINT_STATUS__PMU_EVENTS__ENABLE	0x08
38 #define CCN_MN_ERRINT_STATUS__PMU_EVENTS__DISABLED	0x80
39 #define CCN_MN_ERRINT_STATUS__PMU_EVENTS__DISABLE	0x88
40 #define CCN_MN_OLY_COMP_LIST_63_0	0x01e0
41 #define CCN_MN_ERR_SIG_VAL_63_0		0x0300
42 #define CCN_MN_ERR_SIG_VAL_63_0__DT			(1 << 1)
43 
44 #define CCN_DT_ACTIVE_DSM		0x0000
45 #define CCN_DT_ACTIVE_DSM__DSM_ID__SHIFT(n)		((n) * 8)
46 #define CCN_DT_ACTIVE_DSM__DSM_ID__MASK			0xff
47 #define CCN_DT_CTL			0x0028
48 #define CCN_DT_CTL__DT_EN				(1 << 0)
49 #define CCN_DT_PMEVCNT(n)		(0x0100 + (n) * 0x8)
50 #define CCN_DT_PMCCNTR			0x0140
51 #define CCN_DT_PMCCNTRSR		0x0190
52 #define CCN_DT_PMOVSR			0x0198
53 #define CCN_DT_PMOVSR_CLR		0x01a0
54 #define CCN_DT_PMOVSR_CLR__MASK				0x1f
55 #define CCN_DT_PMCR			0x01a8
56 #define CCN_DT_PMCR__OVFL_INTR_EN			(1 << 6)
57 #define CCN_DT_PMCR__PMU_EN				(1 << 0)
58 #define CCN_DT_PMSR			0x01b0
59 #define CCN_DT_PMSR_REQ			0x01b8
60 #define CCN_DT_PMSR_CLR			0x01c0
61 
62 #define CCN_HNF_PMU_EVENT_SEL		0x0600
63 #define CCN_HNF_PMU_EVENT_SEL__ID__SHIFT(n)		((n) * 4)
64 #define CCN_HNF_PMU_EVENT_SEL__ID__MASK			0xf
65 
66 #define CCN_XP_DT_CONFIG		0x0300
67 #define CCN_XP_DT_CONFIG__DT_CFG__SHIFT(n)		((n) * 4)
68 #define CCN_XP_DT_CONFIG__DT_CFG__MASK			0xf
69 #define CCN_XP_DT_CONFIG__DT_CFG__PASS_THROUGH		0x0
70 #define CCN_XP_DT_CONFIG__DT_CFG__WATCHPOINT_0_OR_1	0x1
71 #define CCN_XP_DT_CONFIG__DT_CFG__WATCHPOINT(n)		(0x2 + (n))
72 #define CCN_XP_DT_CONFIG__DT_CFG__XP_PMU_EVENT(n)	(0x4 + (n))
73 #define CCN_XP_DT_CONFIG__DT_CFG__DEVICE_PMU_EVENT(d, n) (0x8 + (d) * 4 + (n))
74 #define CCN_XP_DT_INTERFACE_SEL		0x0308
75 #define CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__SHIFT(n)	(0 + (n) * 8)
76 #define CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__MASK	0x1
77 #define CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__SHIFT(n)	(1 + (n) * 8)
78 #define CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__MASK	0x1
79 #define CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__SHIFT(n)	(2 + (n) * 8)
80 #define CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__MASK	0x3
81 #define CCN_XP_DT_CMP_VAL_L(n)		(0x0310 + (n) * 0x40)
82 #define CCN_XP_DT_CMP_VAL_H(n)		(0x0318 + (n) * 0x40)
83 #define CCN_XP_DT_CMP_MASK_L(n)		(0x0320 + (n) * 0x40)
84 #define CCN_XP_DT_CMP_MASK_H(n)		(0x0328 + (n) * 0x40)
85 #define CCN_XP_DT_CONTROL		0x0370
86 #define CCN_XP_DT_CONTROL__DT_ENABLE			(1 << 0)
87 #define CCN_XP_DT_CONTROL__WP_ARM_SEL__SHIFT(n)		(12 + (n) * 4)
88 #define CCN_XP_DT_CONTROL__WP_ARM_SEL__MASK		0xf
89 #define CCN_XP_DT_CONTROL__WP_ARM_SEL__ALWAYS		0xf
90 #define CCN_XP_PMU_EVENT_SEL		0x0600
91 #define CCN_XP_PMU_EVENT_SEL__ID__SHIFT(n)		((n) * 7)
92 #define CCN_XP_PMU_EVENT_SEL__ID__MASK			0x3f
93 
94 #define CCN_SBAS_PMU_EVENT_SEL		0x0600
95 #define CCN_SBAS_PMU_EVENT_SEL__ID__SHIFT(n)		((n) * 4)
96 #define CCN_SBAS_PMU_EVENT_SEL__ID__MASK		0xf
97 
98 #define CCN_RNI_PMU_EVENT_SEL		0x0600
99 #define CCN_RNI_PMU_EVENT_SEL__ID__SHIFT(n)		((n) * 4)
100 #define CCN_RNI_PMU_EVENT_SEL__ID__MASK			0xf
101 
102 #define CCN_TYPE_MN	0x01
103 #define CCN_TYPE_DT	0x02
104 #define CCN_TYPE_HNF	0x04
105 #define CCN_TYPE_HNI	0x05
106 #define CCN_TYPE_XP	0x08
107 #define CCN_TYPE_SBSX	0x0c
108 #define CCN_TYPE_SBAS	0x10
109 #define CCN_TYPE_RNI_1P	0x14
110 #define CCN_TYPE_RNI_2P	0x15
111 #define CCN_TYPE_RNI_3P	0x16
112 #define CCN_TYPE_RND_1P	0x18 /* RN-D = RN-I + DVM */
113 #define CCN_TYPE_RND_2P	0x19
114 #define CCN_TYPE_RND_3P	0x1a
115 #define CCN_TYPE_CYCLES	0xff /* Pseudotype */
116 
117 #define CCN_EVENT_WATCHPOINT 0xfe /* Pseudoevent */
118 
119 #define CCN_NUM_PMU_EVENTS		4
120 #define CCN_NUM_XP_WATCHPOINTS		2 /* See DT.dbg_id.num_watchpoints */
121 #define CCN_NUM_PMU_EVENT_COUNTERS	8 /* See DT.dbg_id.num_pmucntr */
122 #define CCN_IDX_PMU_CYCLE_COUNTER	CCN_NUM_PMU_EVENT_COUNTERS
123 
124 #define CCN_NUM_PREDEFINED_MASKS	4
125 #define CCN_IDX_MASK_ANY		(CCN_NUM_PMU_EVENT_COUNTERS + 0)
126 #define CCN_IDX_MASK_EXACT		(CCN_NUM_PMU_EVENT_COUNTERS + 1)
127 #define CCN_IDX_MASK_ORDER		(CCN_NUM_PMU_EVENT_COUNTERS + 2)
128 #define CCN_IDX_MASK_OPCODE		(CCN_NUM_PMU_EVENT_COUNTERS + 3)
129 
130 struct arm_ccn_component {
131 	void __iomem *base;
132 	u32 type;
133 
134 	DECLARE_BITMAP(pmu_events_mask, CCN_NUM_PMU_EVENTS);
135 	union {
136 		struct {
137 			DECLARE_BITMAP(dt_cmp_mask, CCN_NUM_XP_WATCHPOINTS);
138 		} xp;
139 	};
140 };
141 
142 #define pmu_to_arm_ccn(_pmu) container_of(container_of(_pmu, \
143 	struct arm_ccn_dt, pmu), struct arm_ccn, dt)
144 
145 struct arm_ccn_dt {
146 	int id;
147 	void __iomem *base;
148 
149 	spinlock_t config_lock;
150 
151 	DECLARE_BITMAP(pmu_counters_mask, CCN_NUM_PMU_EVENT_COUNTERS + 1);
152 	struct {
153 		struct arm_ccn_component *source;
154 		struct perf_event *event;
155 	} pmu_counters[CCN_NUM_PMU_EVENT_COUNTERS + 1];
156 
157 	struct {
158 	       u64 l, h;
159 	} cmp_mask[CCN_NUM_PMU_EVENT_COUNTERS + CCN_NUM_PREDEFINED_MASKS];
160 
161 	struct hrtimer hrtimer;
162 
163 	unsigned int cpu;
164 	struct hlist_node node;
165 
166 	struct pmu pmu;
167 };
168 
169 struct arm_ccn {
170 	struct device *dev;
171 	void __iomem *base;
172 	unsigned int irq;
173 
174 	unsigned sbas_present:1;
175 	unsigned sbsx_present:1;
176 
177 	int num_nodes;
178 	struct arm_ccn_component *node;
179 
180 	int num_xps;
181 	struct arm_ccn_component *xp;
182 
183 	struct arm_ccn_dt dt;
184 	int mn_id;
185 };
186 
187 static int arm_ccn_node_to_xp(int node)
188 {
189 	return node / CCN_NUM_XP_PORTS;
190 }
191 
192 static int arm_ccn_node_to_xp_port(int node)
193 {
194 	return node % CCN_NUM_XP_PORTS;
195 }
196 
197 
198 /*
199  * Bit shifts and masks in these defines must be kept in sync with
200  * arm_ccn_pmu_config_set() and CCN_FORMAT_ATTRs below!
201  */
202 #define CCN_CONFIG_NODE(_config)	(((_config) >> 0) & 0xff)
203 #define CCN_CONFIG_XP(_config)		(((_config) >> 0) & 0xff)
204 #define CCN_CONFIG_TYPE(_config)	(((_config) >> 8) & 0xff)
205 #define CCN_CONFIG_EVENT(_config)	(((_config) >> 16) & 0xff)
206 #define CCN_CONFIG_PORT(_config)	(((_config) >> 24) & 0x3)
207 #define CCN_CONFIG_BUS(_config)		(((_config) >> 24) & 0x3)
208 #define CCN_CONFIG_VC(_config)		(((_config) >> 26) & 0x7)
209 #define CCN_CONFIG_DIR(_config)		(((_config) >> 29) & 0x1)
210 #define CCN_CONFIG_MASK(_config)	(((_config) >> 30) & 0xf)
211 
212 static void arm_ccn_pmu_config_set(u64 *config, u32 node_xp, u32 type, u32 port)
213 {
214 	*config &= ~((0xff << 0) | (0xff << 8) | (0x3 << 24));
215 	*config |= (node_xp << 0) | (type << 8) | (port << 24);
216 }
217 
218 #define CCN_FORMAT_ATTR(_name, _config) \
219 	struct dev_ext_attribute arm_ccn_pmu_format_attr_##_name = \
220 			{ __ATTR(_name, S_IRUGO, device_show_string, \
221 			NULL), _config }
222 
223 static CCN_FORMAT_ATTR(node, "config:0-7");
224 static CCN_FORMAT_ATTR(xp, "config:0-7");
225 static CCN_FORMAT_ATTR(type, "config:8-15");
226 static CCN_FORMAT_ATTR(event, "config:16-23");
227 static CCN_FORMAT_ATTR(port, "config:24-25");
228 static CCN_FORMAT_ATTR(bus, "config:24-25");
229 static CCN_FORMAT_ATTR(vc, "config:26-28");
230 static CCN_FORMAT_ATTR(dir, "config:29-29");
231 static CCN_FORMAT_ATTR(mask, "config:30-33");
232 static CCN_FORMAT_ATTR(cmp_l, "config1:0-62");
233 static CCN_FORMAT_ATTR(cmp_h, "config2:0-59");
234 
235 static struct attribute *arm_ccn_pmu_format_attrs[] = {
236 	&arm_ccn_pmu_format_attr_node.attr.attr,
237 	&arm_ccn_pmu_format_attr_xp.attr.attr,
238 	&arm_ccn_pmu_format_attr_type.attr.attr,
239 	&arm_ccn_pmu_format_attr_event.attr.attr,
240 	&arm_ccn_pmu_format_attr_port.attr.attr,
241 	&arm_ccn_pmu_format_attr_bus.attr.attr,
242 	&arm_ccn_pmu_format_attr_vc.attr.attr,
243 	&arm_ccn_pmu_format_attr_dir.attr.attr,
244 	&arm_ccn_pmu_format_attr_mask.attr.attr,
245 	&arm_ccn_pmu_format_attr_cmp_l.attr.attr,
246 	&arm_ccn_pmu_format_attr_cmp_h.attr.attr,
247 	NULL
248 };
249 
250 static const struct attribute_group arm_ccn_pmu_format_attr_group = {
251 	.name = "format",
252 	.attrs = arm_ccn_pmu_format_attrs,
253 };
254 
255 
256 struct arm_ccn_pmu_event {
257 	struct device_attribute attr;
258 	u32 type;
259 	u32 event;
260 	int num_ports;
261 	int num_vcs;
262 	const char *def;
263 	int mask;
264 };
265 
266 #define CCN_EVENT_ATTR(_name) \
267 	__ATTR(_name, S_IRUGO, arm_ccn_pmu_event_show, NULL)
268 
269 /*
270  * Events defined in TRM for MN, HN-I and SBSX are actually watchpoints set on
271  * their ports in XP they are connected to. For the sake of usability they are
272  * explicitly defined here (and translated into a relevant watchpoint in
273  * arm_ccn_pmu_event_init()) so the user can easily request them without deep
274  * knowledge of the flit format.
275  */
276 
277 #define CCN_EVENT_MN(_name, _def, _mask) { .attr = CCN_EVENT_ATTR(mn_##_name), \
278 		.type = CCN_TYPE_MN, .event = CCN_EVENT_WATCHPOINT, \
279 		.num_ports = CCN_NUM_XP_PORTS, .num_vcs = CCN_NUM_VCS, \
280 		.def = _def, .mask = _mask, }
281 
282 #define CCN_EVENT_HNI(_name, _def, _mask) { \
283 		.attr = CCN_EVENT_ATTR(hni_##_name), .type = CCN_TYPE_HNI, \
284 		.event = CCN_EVENT_WATCHPOINT, .num_ports = CCN_NUM_XP_PORTS, \
285 		.num_vcs = CCN_NUM_VCS, .def = _def, .mask = _mask, }
286 
287 #define CCN_EVENT_SBSX(_name, _def, _mask) { \
288 		.attr = CCN_EVENT_ATTR(sbsx_##_name), .type = CCN_TYPE_SBSX, \
289 		.event = CCN_EVENT_WATCHPOINT, .num_ports = CCN_NUM_XP_PORTS, \
290 		.num_vcs = CCN_NUM_VCS, .def = _def, .mask = _mask, }
291 
292 #define CCN_EVENT_HNF(_name, _event) { .attr = CCN_EVENT_ATTR(hnf_##_name), \
293 		.type = CCN_TYPE_HNF, .event = _event, }
294 
295 #define CCN_EVENT_XP(_name, _event) { .attr = CCN_EVENT_ATTR(xp_##_name), \
296 		.type = CCN_TYPE_XP, .event = _event, \
297 		.num_ports = CCN_NUM_XP_PORTS, .num_vcs = CCN_NUM_VCS, }
298 
299 /*
300  * RN-I & RN-D (RN-D = RN-I + DVM) nodes have different type ID depending
301  * on configuration. One of them is picked to represent the whole group,
302  * as they all share the same event types.
303  */
304 #define CCN_EVENT_RNI(_name, _event) { .attr = CCN_EVENT_ATTR(rni_##_name), \
305 		.type = CCN_TYPE_RNI_3P, .event = _event, }
306 
307 #define CCN_EVENT_SBAS(_name, _event) { .attr = CCN_EVENT_ATTR(sbas_##_name), \
308 		.type = CCN_TYPE_SBAS, .event = _event, }
309 
310 #define CCN_EVENT_CYCLES(_name) { .attr = CCN_EVENT_ATTR(_name), \
311 		.type = CCN_TYPE_CYCLES }
312 
313 
314 static ssize_t arm_ccn_pmu_event_show(struct device *dev,
315 		struct device_attribute *attr, char *buf)
316 {
317 	struct arm_ccn *ccn = pmu_to_arm_ccn(dev_get_drvdata(dev));
318 	struct arm_ccn_pmu_event *event = container_of(attr,
319 			struct arm_ccn_pmu_event, attr);
320 	int res;
321 
322 	res = sysfs_emit(buf, "type=0x%x", event->type);
323 	if (event->event)
324 		res += sysfs_emit_at(buf, res, ",event=0x%x", event->event);
325 	if (event->def)
326 		res += sysfs_emit_at(buf, res, ",%s", event->def);
327 	if (event->mask)
328 		res += sysfs_emit_at(buf, res, ",mask=0x%x", event->mask);
329 
330 	/* Arguments required by an event */
331 	switch (event->type) {
332 	case CCN_TYPE_CYCLES:
333 		break;
334 	case CCN_TYPE_XP:
335 		res += sysfs_emit_at(buf, res, ",xp=?,vc=?");
336 		if (event->event == CCN_EVENT_WATCHPOINT)
337 			res += sysfs_emit_at(buf, res,
338 					",port=?,dir=?,cmp_l=?,cmp_h=?,mask=?");
339 		else
340 			res += sysfs_emit_at(buf, res, ",bus=?");
341 
342 		break;
343 	case CCN_TYPE_MN:
344 		res += sysfs_emit_at(buf, res, ",node=%d", ccn->mn_id);
345 		break;
346 	default:
347 		res += sysfs_emit_at(buf, res, ",node=?");
348 		break;
349 	}
350 
351 	res += sysfs_emit_at(buf, res, "\n");
352 
353 	return res;
354 }
355 
356 static umode_t arm_ccn_pmu_events_is_visible(struct kobject *kobj,
357 				     struct attribute *attr, int index)
358 {
359 	struct device *dev = kobj_to_dev(kobj);
360 	struct arm_ccn *ccn = pmu_to_arm_ccn(dev_get_drvdata(dev));
361 	struct device_attribute *dev_attr = container_of(attr,
362 			struct device_attribute, attr);
363 	struct arm_ccn_pmu_event *event = container_of(dev_attr,
364 			struct arm_ccn_pmu_event, attr);
365 
366 	if (event->type == CCN_TYPE_SBAS && !ccn->sbas_present)
367 		return 0;
368 	if (event->type == CCN_TYPE_SBSX && !ccn->sbsx_present)
369 		return 0;
370 
371 	return attr->mode;
372 }
373 
374 static struct arm_ccn_pmu_event arm_ccn_pmu_events[] = {
375 	CCN_EVENT_MN(eobarrier, "dir=1,vc=0,cmp_h=0x1c00", CCN_IDX_MASK_OPCODE),
376 	CCN_EVENT_MN(ecbarrier, "dir=1,vc=0,cmp_h=0x1e00", CCN_IDX_MASK_OPCODE),
377 	CCN_EVENT_MN(dvmop, "dir=1,vc=0,cmp_h=0x2800", CCN_IDX_MASK_OPCODE),
378 	CCN_EVENT_HNI(txdatflits, "dir=1,vc=3", CCN_IDX_MASK_ANY),
379 	CCN_EVENT_HNI(rxdatflits, "dir=0,vc=3", CCN_IDX_MASK_ANY),
380 	CCN_EVENT_HNI(txreqflits, "dir=1,vc=0", CCN_IDX_MASK_ANY),
381 	CCN_EVENT_HNI(rxreqflits, "dir=0,vc=0", CCN_IDX_MASK_ANY),
382 	CCN_EVENT_HNI(rxreqflits_order, "dir=0,vc=0,cmp_h=0x8000",
383 			CCN_IDX_MASK_ORDER),
384 	CCN_EVENT_SBSX(txdatflits, "dir=1,vc=3", CCN_IDX_MASK_ANY),
385 	CCN_EVENT_SBSX(rxdatflits, "dir=0,vc=3", CCN_IDX_MASK_ANY),
386 	CCN_EVENT_SBSX(txreqflits, "dir=1,vc=0", CCN_IDX_MASK_ANY),
387 	CCN_EVENT_SBSX(rxreqflits, "dir=0,vc=0", CCN_IDX_MASK_ANY),
388 	CCN_EVENT_SBSX(rxreqflits_order, "dir=0,vc=0,cmp_h=0x8000",
389 			CCN_IDX_MASK_ORDER),
390 	CCN_EVENT_HNF(cache_miss, 0x1),
391 	CCN_EVENT_HNF(l3_sf_cache_access, 0x02),
392 	CCN_EVENT_HNF(cache_fill, 0x3),
393 	CCN_EVENT_HNF(pocq_retry, 0x4),
394 	CCN_EVENT_HNF(pocq_reqs_recvd, 0x5),
395 	CCN_EVENT_HNF(sf_hit, 0x6),
396 	CCN_EVENT_HNF(sf_evictions, 0x7),
397 	CCN_EVENT_HNF(snoops_sent, 0x8),
398 	CCN_EVENT_HNF(snoops_broadcast, 0x9),
399 	CCN_EVENT_HNF(l3_eviction, 0xa),
400 	CCN_EVENT_HNF(l3_fill_invalid_way, 0xb),
401 	CCN_EVENT_HNF(mc_retries, 0xc),
402 	CCN_EVENT_HNF(mc_reqs, 0xd),
403 	CCN_EVENT_HNF(qos_hh_retry, 0xe),
404 	CCN_EVENT_RNI(rdata_beats_p0, 0x1),
405 	CCN_EVENT_RNI(rdata_beats_p1, 0x2),
406 	CCN_EVENT_RNI(rdata_beats_p2, 0x3),
407 	CCN_EVENT_RNI(rxdat_flits, 0x4),
408 	CCN_EVENT_RNI(txdat_flits, 0x5),
409 	CCN_EVENT_RNI(txreq_flits, 0x6),
410 	CCN_EVENT_RNI(txreq_flits_retried, 0x7),
411 	CCN_EVENT_RNI(rrt_full, 0x8),
412 	CCN_EVENT_RNI(wrt_full, 0x9),
413 	CCN_EVENT_RNI(txreq_flits_replayed, 0xa),
414 	CCN_EVENT_XP(upload_starvation, 0x1),
415 	CCN_EVENT_XP(download_starvation, 0x2),
416 	CCN_EVENT_XP(respin, 0x3),
417 	CCN_EVENT_XP(valid_flit, 0x4),
418 	CCN_EVENT_XP(watchpoint, CCN_EVENT_WATCHPOINT),
419 	CCN_EVENT_SBAS(rdata_beats_p0, 0x1),
420 	CCN_EVENT_SBAS(rxdat_flits, 0x4),
421 	CCN_EVENT_SBAS(txdat_flits, 0x5),
422 	CCN_EVENT_SBAS(txreq_flits, 0x6),
423 	CCN_EVENT_SBAS(txreq_flits_retried, 0x7),
424 	CCN_EVENT_SBAS(rrt_full, 0x8),
425 	CCN_EVENT_SBAS(wrt_full, 0x9),
426 	CCN_EVENT_SBAS(txreq_flits_replayed, 0xa),
427 	CCN_EVENT_CYCLES(cycles),
428 };
429 
430 /* Populated in arm_ccn_init() */
431 static struct attribute
432 		*arm_ccn_pmu_events_attrs[ARRAY_SIZE(arm_ccn_pmu_events) + 1];
433 
434 static const struct attribute_group arm_ccn_pmu_events_attr_group = {
435 	.name = "events",
436 	.is_visible = arm_ccn_pmu_events_is_visible,
437 	.attrs = arm_ccn_pmu_events_attrs,
438 };
439 
440 
441 static u64 *arm_ccn_pmu_get_cmp_mask(struct arm_ccn *ccn, const char *name)
442 {
443 	unsigned long i;
444 
445 	if (WARN_ON(!name || !name[0] || !isxdigit(name[0]) || !name[1]))
446 		return NULL;
447 	i = isdigit(name[0]) ? name[0] - '0' : 0xa + tolower(name[0]) - 'a';
448 
449 	switch (name[1]) {
450 	case 'l':
451 		return &ccn->dt.cmp_mask[i].l;
452 	case 'h':
453 		return &ccn->dt.cmp_mask[i].h;
454 	default:
455 		return NULL;
456 	}
457 }
458 
459 static ssize_t arm_ccn_pmu_cmp_mask_show(struct device *dev,
460 		struct device_attribute *attr, char *buf)
461 {
462 	struct arm_ccn *ccn = pmu_to_arm_ccn(dev_get_drvdata(dev));
463 	u64 *mask = arm_ccn_pmu_get_cmp_mask(ccn, attr->attr.name);
464 
465 	return mask ? sysfs_emit(buf, "0x%016llx\n", *mask) : -EINVAL;
466 }
467 
468 static ssize_t arm_ccn_pmu_cmp_mask_store(struct device *dev,
469 		struct device_attribute *attr, const char *buf, size_t count)
470 {
471 	struct arm_ccn *ccn = pmu_to_arm_ccn(dev_get_drvdata(dev));
472 	u64 *mask = arm_ccn_pmu_get_cmp_mask(ccn, attr->attr.name);
473 	int err = -EINVAL;
474 
475 	if (mask)
476 		err = kstrtoull(buf, 0, mask);
477 
478 	return err ? err : count;
479 }
480 
481 #define CCN_CMP_MASK_ATTR(_name) \
482 	struct device_attribute arm_ccn_pmu_cmp_mask_attr_##_name = \
483 			__ATTR(_name, S_IRUGO | S_IWUSR, \
484 			arm_ccn_pmu_cmp_mask_show, arm_ccn_pmu_cmp_mask_store)
485 
486 #define CCN_CMP_MASK_ATTR_RO(_name) \
487 	struct device_attribute arm_ccn_pmu_cmp_mask_attr_##_name = \
488 			__ATTR(_name, S_IRUGO, arm_ccn_pmu_cmp_mask_show, NULL)
489 
490 static CCN_CMP_MASK_ATTR(0l);
491 static CCN_CMP_MASK_ATTR(0h);
492 static CCN_CMP_MASK_ATTR(1l);
493 static CCN_CMP_MASK_ATTR(1h);
494 static CCN_CMP_MASK_ATTR(2l);
495 static CCN_CMP_MASK_ATTR(2h);
496 static CCN_CMP_MASK_ATTR(3l);
497 static CCN_CMP_MASK_ATTR(3h);
498 static CCN_CMP_MASK_ATTR(4l);
499 static CCN_CMP_MASK_ATTR(4h);
500 static CCN_CMP_MASK_ATTR(5l);
501 static CCN_CMP_MASK_ATTR(5h);
502 static CCN_CMP_MASK_ATTR(6l);
503 static CCN_CMP_MASK_ATTR(6h);
504 static CCN_CMP_MASK_ATTR(7l);
505 static CCN_CMP_MASK_ATTR(7h);
506 static CCN_CMP_MASK_ATTR_RO(8l);
507 static CCN_CMP_MASK_ATTR_RO(8h);
508 static CCN_CMP_MASK_ATTR_RO(9l);
509 static CCN_CMP_MASK_ATTR_RO(9h);
510 static CCN_CMP_MASK_ATTR_RO(al);
511 static CCN_CMP_MASK_ATTR_RO(ah);
512 static CCN_CMP_MASK_ATTR_RO(bl);
513 static CCN_CMP_MASK_ATTR_RO(bh);
514 
515 static struct attribute *arm_ccn_pmu_cmp_mask_attrs[] = {
516 	&arm_ccn_pmu_cmp_mask_attr_0l.attr, &arm_ccn_pmu_cmp_mask_attr_0h.attr,
517 	&arm_ccn_pmu_cmp_mask_attr_1l.attr, &arm_ccn_pmu_cmp_mask_attr_1h.attr,
518 	&arm_ccn_pmu_cmp_mask_attr_2l.attr, &arm_ccn_pmu_cmp_mask_attr_2h.attr,
519 	&arm_ccn_pmu_cmp_mask_attr_3l.attr, &arm_ccn_pmu_cmp_mask_attr_3h.attr,
520 	&arm_ccn_pmu_cmp_mask_attr_4l.attr, &arm_ccn_pmu_cmp_mask_attr_4h.attr,
521 	&arm_ccn_pmu_cmp_mask_attr_5l.attr, &arm_ccn_pmu_cmp_mask_attr_5h.attr,
522 	&arm_ccn_pmu_cmp_mask_attr_6l.attr, &arm_ccn_pmu_cmp_mask_attr_6h.attr,
523 	&arm_ccn_pmu_cmp_mask_attr_7l.attr, &arm_ccn_pmu_cmp_mask_attr_7h.attr,
524 	&arm_ccn_pmu_cmp_mask_attr_8l.attr, &arm_ccn_pmu_cmp_mask_attr_8h.attr,
525 	&arm_ccn_pmu_cmp_mask_attr_9l.attr, &arm_ccn_pmu_cmp_mask_attr_9h.attr,
526 	&arm_ccn_pmu_cmp_mask_attr_al.attr, &arm_ccn_pmu_cmp_mask_attr_ah.attr,
527 	&arm_ccn_pmu_cmp_mask_attr_bl.attr, &arm_ccn_pmu_cmp_mask_attr_bh.attr,
528 	NULL
529 };
530 
531 static const struct attribute_group arm_ccn_pmu_cmp_mask_attr_group = {
532 	.name = "cmp_mask",
533 	.attrs = arm_ccn_pmu_cmp_mask_attrs,
534 };
535 
536 static ssize_t arm_ccn_pmu_cpumask_show(struct device *dev,
537 				     struct device_attribute *attr, char *buf)
538 {
539 	struct arm_ccn *ccn = pmu_to_arm_ccn(dev_get_drvdata(dev));
540 
541 	return cpumap_print_to_pagebuf(true, buf, cpumask_of(ccn->dt.cpu));
542 }
543 
544 static struct device_attribute arm_ccn_pmu_cpumask_attr =
545 		__ATTR(cpumask, S_IRUGO, arm_ccn_pmu_cpumask_show, NULL);
546 
547 static struct attribute *arm_ccn_pmu_cpumask_attrs[] = {
548 	&arm_ccn_pmu_cpumask_attr.attr,
549 	NULL,
550 };
551 
552 static const struct attribute_group arm_ccn_pmu_cpumask_attr_group = {
553 	.attrs = arm_ccn_pmu_cpumask_attrs,
554 };
555 
556 /*
557  * Default poll period is 10ms, which is way over the top anyway,
558  * as in the worst case scenario (an event every cycle), with 1GHz
559  * clocked bus, the smallest, 32 bit counter will overflow in
560  * more than 4s.
561  */
562 static unsigned int arm_ccn_pmu_poll_period_us = 10000;
563 module_param_named(pmu_poll_period_us, arm_ccn_pmu_poll_period_us, uint,
564 		S_IRUGO | S_IWUSR);
565 
566 static ktime_t arm_ccn_pmu_timer_period(void)
567 {
568 	return ns_to_ktime((u64)arm_ccn_pmu_poll_period_us * 1000);
569 }
570 
571 
572 static const struct attribute_group *arm_ccn_pmu_attr_groups[] = {
573 	&arm_ccn_pmu_events_attr_group,
574 	&arm_ccn_pmu_format_attr_group,
575 	&arm_ccn_pmu_cmp_mask_attr_group,
576 	&arm_ccn_pmu_cpumask_attr_group,
577 	NULL
578 };
579 
580 
581 static int arm_ccn_pmu_alloc_bit(unsigned long *bitmap, unsigned long size)
582 {
583 	int bit;
584 
585 	do {
586 		bit = find_first_zero_bit(bitmap, size);
587 		if (bit >= size)
588 			return -EAGAIN;
589 	} while (test_and_set_bit(bit, bitmap));
590 
591 	return bit;
592 }
593 
594 /* All RN-I and RN-D nodes have identical PMUs */
595 static int arm_ccn_pmu_type_eq(u32 a, u32 b)
596 {
597 	if (a == b)
598 		return 1;
599 
600 	switch (a) {
601 	case CCN_TYPE_RNI_1P:
602 	case CCN_TYPE_RNI_2P:
603 	case CCN_TYPE_RNI_3P:
604 	case CCN_TYPE_RND_1P:
605 	case CCN_TYPE_RND_2P:
606 	case CCN_TYPE_RND_3P:
607 		switch (b) {
608 		case CCN_TYPE_RNI_1P:
609 		case CCN_TYPE_RNI_2P:
610 		case CCN_TYPE_RNI_3P:
611 		case CCN_TYPE_RND_1P:
612 		case CCN_TYPE_RND_2P:
613 		case CCN_TYPE_RND_3P:
614 			return 1;
615 		}
616 		break;
617 	}
618 
619 	return 0;
620 }
621 
622 static int arm_ccn_pmu_event_alloc(struct perf_event *event)
623 {
624 	struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
625 	struct hw_perf_event *hw = &event->hw;
626 	u32 node_xp, type, event_id;
627 	struct arm_ccn_component *source;
628 	int bit;
629 
630 	node_xp = CCN_CONFIG_NODE(event->attr.config);
631 	type = CCN_CONFIG_TYPE(event->attr.config);
632 	event_id = CCN_CONFIG_EVENT(event->attr.config);
633 
634 	/* Allocate the cycle counter */
635 	if (type == CCN_TYPE_CYCLES) {
636 		if (test_and_set_bit(CCN_IDX_PMU_CYCLE_COUNTER,
637 				ccn->dt.pmu_counters_mask))
638 			return -EAGAIN;
639 
640 		hw->idx = CCN_IDX_PMU_CYCLE_COUNTER;
641 		ccn->dt.pmu_counters[CCN_IDX_PMU_CYCLE_COUNTER].event = event;
642 
643 		return 0;
644 	}
645 
646 	/* Allocate an event counter */
647 	hw->idx = arm_ccn_pmu_alloc_bit(ccn->dt.pmu_counters_mask,
648 			CCN_NUM_PMU_EVENT_COUNTERS);
649 	if (hw->idx < 0) {
650 		dev_dbg(ccn->dev, "No more counters available!\n");
651 		return -EAGAIN;
652 	}
653 
654 	if (type == CCN_TYPE_XP)
655 		source = &ccn->xp[node_xp];
656 	else
657 		source = &ccn->node[node_xp];
658 	ccn->dt.pmu_counters[hw->idx].source = source;
659 
660 	/* Allocate an event source or a watchpoint */
661 	if (type == CCN_TYPE_XP && event_id == CCN_EVENT_WATCHPOINT)
662 		bit = arm_ccn_pmu_alloc_bit(source->xp.dt_cmp_mask,
663 				CCN_NUM_XP_WATCHPOINTS);
664 	else
665 		bit = arm_ccn_pmu_alloc_bit(source->pmu_events_mask,
666 				CCN_NUM_PMU_EVENTS);
667 	if (bit < 0) {
668 		dev_dbg(ccn->dev, "No more event sources/watchpoints on node/XP %d!\n",
669 				node_xp);
670 		clear_bit(hw->idx, ccn->dt.pmu_counters_mask);
671 		return -EAGAIN;
672 	}
673 	hw->config_base = bit;
674 
675 	ccn->dt.pmu_counters[hw->idx].event = event;
676 
677 	return 0;
678 }
679 
680 static void arm_ccn_pmu_event_release(struct perf_event *event)
681 {
682 	struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
683 	struct hw_perf_event *hw = &event->hw;
684 
685 	if (hw->idx == CCN_IDX_PMU_CYCLE_COUNTER) {
686 		clear_bit(CCN_IDX_PMU_CYCLE_COUNTER, ccn->dt.pmu_counters_mask);
687 	} else {
688 		struct arm_ccn_component *source =
689 				ccn->dt.pmu_counters[hw->idx].source;
690 
691 		if (CCN_CONFIG_TYPE(event->attr.config) == CCN_TYPE_XP &&
692 				CCN_CONFIG_EVENT(event->attr.config) ==
693 				CCN_EVENT_WATCHPOINT)
694 			clear_bit(hw->config_base, source->xp.dt_cmp_mask);
695 		else
696 			clear_bit(hw->config_base, source->pmu_events_mask);
697 		clear_bit(hw->idx, ccn->dt.pmu_counters_mask);
698 	}
699 
700 	ccn->dt.pmu_counters[hw->idx].source = NULL;
701 	ccn->dt.pmu_counters[hw->idx].event = NULL;
702 }
703 
704 static int arm_ccn_pmu_event_init(struct perf_event *event)
705 {
706 	struct arm_ccn *ccn;
707 	struct hw_perf_event *hw = &event->hw;
708 	u32 node_xp, type, event_id;
709 	int valid;
710 	int i;
711 	struct perf_event *sibling;
712 
713 	if (event->attr.type != event->pmu->type)
714 		return -ENOENT;
715 
716 	ccn = pmu_to_arm_ccn(event->pmu);
717 
718 	if (hw->sample_period) {
719 		dev_dbg(ccn->dev, "Sampling not supported!\n");
720 		return -EOPNOTSUPP;
721 	}
722 
723 	if (has_branch_stack(event)) {
724 		dev_dbg(ccn->dev, "Can't exclude execution levels!\n");
725 		return -EINVAL;
726 	}
727 
728 	if (event->cpu < 0) {
729 		dev_dbg(ccn->dev, "Can't provide per-task data!\n");
730 		return -EOPNOTSUPP;
731 	}
732 	/*
733 	 * Many perf core operations (eg. events rotation) operate on a
734 	 * single CPU context. This is obvious for CPU PMUs, where one
735 	 * expects the same sets of events being observed on all CPUs,
736 	 * but can lead to issues for off-core PMUs, like CCN, where each
737 	 * event could be theoretically assigned to a different CPU. To
738 	 * mitigate this, we enforce CPU assignment to one, selected
739 	 * processor (the one described in the "cpumask" attribute).
740 	 */
741 	event->cpu = ccn->dt.cpu;
742 
743 	node_xp = CCN_CONFIG_NODE(event->attr.config);
744 	type = CCN_CONFIG_TYPE(event->attr.config);
745 	event_id = CCN_CONFIG_EVENT(event->attr.config);
746 
747 	/* Validate node/xp vs topology */
748 	switch (type) {
749 	case CCN_TYPE_MN:
750 		if (node_xp != ccn->mn_id) {
751 			dev_dbg(ccn->dev, "Invalid MN ID %d!\n", node_xp);
752 			return -EINVAL;
753 		}
754 		break;
755 	case CCN_TYPE_XP:
756 		if (node_xp >= ccn->num_xps) {
757 			dev_dbg(ccn->dev, "Invalid XP ID %d!\n", node_xp);
758 			return -EINVAL;
759 		}
760 		break;
761 	case CCN_TYPE_CYCLES:
762 		break;
763 	default:
764 		if (node_xp >= ccn->num_nodes) {
765 			dev_dbg(ccn->dev, "Invalid node ID %d!\n", node_xp);
766 			return -EINVAL;
767 		}
768 		if (!arm_ccn_pmu_type_eq(type, ccn->node[node_xp].type)) {
769 			dev_dbg(ccn->dev, "Invalid type 0x%x for node %d!\n",
770 					type, node_xp);
771 			return -EINVAL;
772 		}
773 		break;
774 	}
775 
776 	/* Validate event ID vs available for the type */
777 	for (i = 0, valid = 0; i < ARRAY_SIZE(arm_ccn_pmu_events) && !valid;
778 			i++) {
779 		struct arm_ccn_pmu_event *e = &arm_ccn_pmu_events[i];
780 		u32 port = CCN_CONFIG_PORT(event->attr.config);
781 		u32 vc = CCN_CONFIG_VC(event->attr.config);
782 
783 		if (!arm_ccn_pmu_type_eq(type, e->type))
784 			continue;
785 		if (event_id != e->event)
786 			continue;
787 		if (e->num_ports && port >= e->num_ports) {
788 			dev_dbg(ccn->dev, "Invalid port %d for node/XP %d!\n",
789 					port, node_xp);
790 			return -EINVAL;
791 		}
792 		if (e->num_vcs && vc >= e->num_vcs) {
793 			dev_dbg(ccn->dev, "Invalid vc %d for node/XP %d!\n",
794 					vc, node_xp);
795 			return -EINVAL;
796 		}
797 		valid = 1;
798 	}
799 	if (!valid) {
800 		dev_dbg(ccn->dev, "Invalid event 0x%x for node/XP %d!\n",
801 				event_id, node_xp);
802 		return -EINVAL;
803 	}
804 
805 	/* Watchpoint-based event for a node is actually set on XP */
806 	if (event_id == CCN_EVENT_WATCHPOINT && type != CCN_TYPE_XP) {
807 		u32 port;
808 
809 		type = CCN_TYPE_XP;
810 		port = arm_ccn_node_to_xp_port(node_xp);
811 		node_xp = arm_ccn_node_to_xp(node_xp);
812 
813 		arm_ccn_pmu_config_set(&event->attr.config,
814 				node_xp, type, port);
815 	}
816 
817 	/*
818 	 * We must NOT create groups containing mixed PMUs, although software
819 	 * events are acceptable (for example to create a CCN group
820 	 * periodically read when a hrtimer aka cpu-clock leader triggers).
821 	 */
822 	if (event->group_leader->pmu != event->pmu &&
823 			!is_software_event(event->group_leader))
824 		return -EINVAL;
825 
826 	for_each_sibling_event(sibling, event->group_leader) {
827 		if (sibling->pmu != event->pmu &&
828 				!is_software_event(sibling))
829 			return -EINVAL;
830 	}
831 
832 	return 0;
833 }
834 
835 static u64 arm_ccn_pmu_read_counter(struct arm_ccn *ccn, int idx)
836 {
837 	u64 res;
838 
839 	if (idx == CCN_IDX_PMU_CYCLE_COUNTER) {
840 #ifdef readq
841 		res = readq(ccn->dt.base + CCN_DT_PMCCNTR);
842 #else
843 		/* 40 bit counter, can do snapshot and read in two parts */
844 		writel(0x1, ccn->dt.base + CCN_DT_PMSR_REQ);
845 		while (!(readl(ccn->dt.base + CCN_DT_PMSR) & 0x1))
846 			;
847 		writel(0x1, ccn->dt.base + CCN_DT_PMSR_CLR);
848 		res = readl(ccn->dt.base + CCN_DT_PMCCNTRSR + 4) & 0xff;
849 		res <<= 32;
850 		res |= readl(ccn->dt.base + CCN_DT_PMCCNTRSR);
851 #endif
852 	} else {
853 		res = readl(ccn->dt.base + CCN_DT_PMEVCNT(idx));
854 	}
855 
856 	return res;
857 }
858 
859 static void arm_ccn_pmu_event_update(struct perf_event *event)
860 {
861 	struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
862 	struct hw_perf_event *hw = &event->hw;
863 	u64 prev_count, new_count, mask;
864 
865 	do {
866 		prev_count = local64_read(&hw->prev_count);
867 		new_count = arm_ccn_pmu_read_counter(ccn, hw->idx);
868 	} while (local64_xchg(&hw->prev_count, new_count) != prev_count);
869 
870 	mask = (1LLU << (hw->idx == CCN_IDX_PMU_CYCLE_COUNTER ? 40 : 32)) - 1;
871 
872 	local64_add((new_count - prev_count) & mask, &event->count);
873 }
874 
875 static void arm_ccn_pmu_xp_dt_config(struct perf_event *event, int enable)
876 {
877 	struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
878 	struct hw_perf_event *hw = &event->hw;
879 	struct arm_ccn_component *xp;
880 	u32 val, dt_cfg;
881 
882 	/* Nothing to do for cycle counter */
883 	if (hw->idx == CCN_IDX_PMU_CYCLE_COUNTER)
884 		return;
885 
886 	if (CCN_CONFIG_TYPE(event->attr.config) == CCN_TYPE_XP)
887 		xp = &ccn->xp[CCN_CONFIG_XP(event->attr.config)];
888 	else
889 		xp = &ccn->xp[arm_ccn_node_to_xp(
890 				CCN_CONFIG_NODE(event->attr.config))];
891 
892 	if (enable)
893 		dt_cfg = hw->event_base;
894 	else
895 		dt_cfg = CCN_XP_DT_CONFIG__DT_CFG__PASS_THROUGH;
896 
897 	spin_lock(&ccn->dt.config_lock);
898 
899 	val = readl(xp->base + CCN_XP_DT_CONFIG);
900 	val &= ~(CCN_XP_DT_CONFIG__DT_CFG__MASK <<
901 			CCN_XP_DT_CONFIG__DT_CFG__SHIFT(hw->idx));
902 	val |= dt_cfg << CCN_XP_DT_CONFIG__DT_CFG__SHIFT(hw->idx);
903 	writel(val, xp->base + CCN_XP_DT_CONFIG);
904 
905 	spin_unlock(&ccn->dt.config_lock);
906 }
907 
908 static void arm_ccn_pmu_event_start(struct perf_event *event, int flags)
909 {
910 	struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
911 	struct hw_perf_event *hw = &event->hw;
912 
913 	local64_set(&event->hw.prev_count,
914 			arm_ccn_pmu_read_counter(ccn, hw->idx));
915 	hw->state = 0;
916 
917 	/* Set the DT bus input, engaging the counter */
918 	arm_ccn_pmu_xp_dt_config(event, 1);
919 }
920 
921 static void arm_ccn_pmu_event_stop(struct perf_event *event, int flags)
922 {
923 	struct hw_perf_event *hw = &event->hw;
924 
925 	/* Disable counting, setting the DT bus to pass-through mode */
926 	arm_ccn_pmu_xp_dt_config(event, 0);
927 
928 	if (flags & PERF_EF_UPDATE)
929 		arm_ccn_pmu_event_update(event);
930 
931 	hw->state |= PERF_HES_STOPPED;
932 }
933 
934 static void arm_ccn_pmu_xp_watchpoint_config(struct perf_event *event)
935 {
936 	struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
937 	struct hw_perf_event *hw = &event->hw;
938 	struct arm_ccn_component *source =
939 			ccn->dt.pmu_counters[hw->idx].source;
940 	unsigned long wp = hw->config_base;
941 	u32 val;
942 	u64 cmp_l = event->attr.config1;
943 	u64 cmp_h = event->attr.config2;
944 	u64 mask_l = ccn->dt.cmp_mask[CCN_CONFIG_MASK(event->attr.config)].l;
945 	u64 mask_h = ccn->dt.cmp_mask[CCN_CONFIG_MASK(event->attr.config)].h;
946 
947 	hw->event_base = CCN_XP_DT_CONFIG__DT_CFG__WATCHPOINT(wp);
948 
949 	/* Direction (RX/TX), device (port) & virtual channel */
950 	val = readl(source->base + CCN_XP_DT_INTERFACE_SEL);
951 	val &= ~(CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__MASK <<
952 			CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__SHIFT(wp));
953 	val |= CCN_CONFIG_DIR(event->attr.config) <<
954 			CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__SHIFT(wp);
955 	val &= ~(CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__MASK <<
956 			CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__SHIFT(wp));
957 	val |= CCN_CONFIG_PORT(event->attr.config) <<
958 			CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__SHIFT(wp);
959 	val &= ~(CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__MASK <<
960 			CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__SHIFT(wp));
961 	val |= CCN_CONFIG_VC(event->attr.config) <<
962 			CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__SHIFT(wp);
963 	writel(val, source->base + CCN_XP_DT_INTERFACE_SEL);
964 
965 	/* Comparison values */
966 	writel(cmp_l & 0xffffffff, source->base + CCN_XP_DT_CMP_VAL_L(wp));
967 	writel((cmp_l >> 32) & 0x7fffffff,
968 			source->base + CCN_XP_DT_CMP_VAL_L(wp) + 4);
969 	writel(cmp_h & 0xffffffff, source->base + CCN_XP_DT_CMP_VAL_H(wp));
970 	writel((cmp_h >> 32) & 0x0fffffff,
971 			source->base + CCN_XP_DT_CMP_VAL_H(wp) + 4);
972 
973 	/* Mask */
974 	writel(mask_l & 0xffffffff, source->base + CCN_XP_DT_CMP_MASK_L(wp));
975 	writel((mask_l >> 32) & 0x7fffffff,
976 			source->base + CCN_XP_DT_CMP_MASK_L(wp) + 4);
977 	writel(mask_h & 0xffffffff, source->base + CCN_XP_DT_CMP_MASK_H(wp));
978 	writel((mask_h >> 32) & 0x0fffffff,
979 			source->base + CCN_XP_DT_CMP_MASK_H(wp) + 4);
980 }
981 
982 static void arm_ccn_pmu_xp_event_config(struct perf_event *event)
983 {
984 	struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
985 	struct hw_perf_event *hw = &event->hw;
986 	struct arm_ccn_component *source =
987 			ccn->dt.pmu_counters[hw->idx].source;
988 	u32 val, id;
989 
990 	hw->event_base = CCN_XP_DT_CONFIG__DT_CFG__XP_PMU_EVENT(hw->config_base);
991 
992 	id = (CCN_CONFIG_VC(event->attr.config) << 4) |
993 			(CCN_CONFIG_BUS(event->attr.config) << 3) |
994 			(CCN_CONFIG_EVENT(event->attr.config) << 0);
995 
996 	val = readl(source->base + CCN_XP_PMU_EVENT_SEL);
997 	val &= ~(CCN_XP_PMU_EVENT_SEL__ID__MASK <<
998 			CCN_XP_PMU_EVENT_SEL__ID__SHIFT(hw->config_base));
999 	val |= id << CCN_XP_PMU_EVENT_SEL__ID__SHIFT(hw->config_base);
1000 	writel(val, source->base + CCN_XP_PMU_EVENT_SEL);
1001 }
1002 
1003 static void arm_ccn_pmu_node_event_config(struct perf_event *event)
1004 {
1005 	struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
1006 	struct hw_perf_event *hw = &event->hw;
1007 	struct arm_ccn_component *source =
1008 			ccn->dt.pmu_counters[hw->idx].source;
1009 	u32 type = CCN_CONFIG_TYPE(event->attr.config);
1010 	u32 val, port;
1011 
1012 	port = arm_ccn_node_to_xp_port(CCN_CONFIG_NODE(event->attr.config));
1013 	hw->event_base = CCN_XP_DT_CONFIG__DT_CFG__DEVICE_PMU_EVENT(port,
1014 			hw->config_base);
1015 
1016 	/* These *_event_sel regs should be identical, but let's make sure... */
1017 	BUILD_BUG_ON(CCN_HNF_PMU_EVENT_SEL != CCN_SBAS_PMU_EVENT_SEL);
1018 	BUILD_BUG_ON(CCN_SBAS_PMU_EVENT_SEL != CCN_RNI_PMU_EVENT_SEL);
1019 	BUILD_BUG_ON(CCN_HNF_PMU_EVENT_SEL__ID__SHIFT(1) !=
1020 			CCN_SBAS_PMU_EVENT_SEL__ID__SHIFT(1));
1021 	BUILD_BUG_ON(CCN_SBAS_PMU_EVENT_SEL__ID__SHIFT(1) !=
1022 			CCN_RNI_PMU_EVENT_SEL__ID__SHIFT(1));
1023 	BUILD_BUG_ON(CCN_HNF_PMU_EVENT_SEL__ID__MASK !=
1024 			CCN_SBAS_PMU_EVENT_SEL__ID__MASK);
1025 	BUILD_BUG_ON(CCN_SBAS_PMU_EVENT_SEL__ID__MASK !=
1026 			CCN_RNI_PMU_EVENT_SEL__ID__MASK);
1027 	if (WARN_ON(type != CCN_TYPE_HNF && type != CCN_TYPE_SBAS &&
1028 			!arm_ccn_pmu_type_eq(type, CCN_TYPE_RNI_3P)))
1029 		return;
1030 
1031 	/* Set the event id for the pre-allocated counter */
1032 	val = readl(source->base + CCN_HNF_PMU_EVENT_SEL);
1033 	val &= ~(CCN_HNF_PMU_EVENT_SEL__ID__MASK <<
1034 		CCN_HNF_PMU_EVENT_SEL__ID__SHIFT(hw->config_base));
1035 	val |= CCN_CONFIG_EVENT(event->attr.config) <<
1036 		CCN_HNF_PMU_EVENT_SEL__ID__SHIFT(hw->config_base);
1037 	writel(val, source->base + CCN_HNF_PMU_EVENT_SEL);
1038 }
1039 
1040 static void arm_ccn_pmu_event_config(struct perf_event *event)
1041 {
1042 	struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
1043 	struct hw_perf_event *hw = &event->hw;
1044 	u32 xp, offset, val;
1045 
1046 	/* Cycle counter requires no setup */
1047 	if (hw->idx == CCN_IDX_PMU_CYCLE_COUNTER)
1048 		return;
1049 
1050 	if (CCN_CONFIG_TYPE(event->attr.config) == CCN_TYPE_XP)
1051 		xp = CCN_CONFIG_XP(event->attr.config);
1052 	else
1053 		xp = arm_ccn_node_to_xp(CCN_CONFIG_NODE(event->attr.config));
1054 
1055 	spin_lock(&ccn->dt.config_lock);
1056 
1057 	/* Set the DT bus "distance" register */
1058 	offset = (hw->idx / 4) * 4;
1059 	val = readl(ccn->dt.base + CCN_DT_ACTIVE_DSM + offset);
1060 	val &= ~(CCN_DT_ACTIVE_DSM__DSM_ID__MASK <<
1061 			CCN_DT_ACTIVE_DSM__DSM_ID__SHIFT(hw->idx % 4));
1062 	val |= xp << CCN_DT_ACTIVE_DSM__DSM_ID__SHIFT(hw->idx % 4);
1063 	writel(val, ccn->dt.base + CCN_DT_ACTIVE_DSM + offset);
1064 
1065 	if (CCN_CONFIG_TYPE(event->attr.config) == CCN_TYPE_XP) {
1066 		if (CCN_CONFIG_EVENT(event->attr.config) ==
1067 				CCN_EVENT_WATCHPOINT)
1068 			arm_ccn_pmu_xp_watchpoint_config(event);
1069 		else
1070 			arm_ccn_pmu_xp_event_config(event);
1071 	} else {
1072 		arm_ccn_pmu_node_event_config(event);
1073 	}
1074 
1075 	spin_unlock(&ccn->dt.config_lock);
1076 }
1077 
1078 static int arm_ccn_pmu_active_counters(struct arm_ccn *ccn)
1079 {
1080 	return bitmap_weight(ccn->dt.pmu_counters_mask,
1081 			     CCN_NUM_PMU_EVENT_COUNTERS + 1);
1082 }
1083 
1084 static int arm_ccn_pmu_event_add(struct perf_event *event, int flags)
1085 {
1086 	int err;
1087 	struct hw_perf_event *hw = &event->hw;
1088 	struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
1089 
1090 	err = arm_ccn_pmu_event_alloc(event);
1091 	if (err)
1092 		return err;
1093 
1094 	/*
1095 	 * Pin the timer, so that the overflows are handled by the chosen
1096 	 * event->cpu (this is the same one as presented in "cpumask"
1097 	 * attribute).
1098 	 */
1099 	if (!ccn->irq && arm_ccn_pmu_active_counters(ccn) == 1)
1100 		hrtimer_start(&ccn->dt.hrtimer, arm_ccn_pmu_timer_period(),
1101 			      HRTIMER_MODE_REL_PINNED);
1102 
1103 	arm_ccn_pmu_event_config(event);
1104 
1105 	hw->state = PERF_HES_STOPPED;
1106 
1107 	if (flags & PERF_EF_START)
1108 		arm_ccn_pmu_event_start(event, PERF_EF_UPDATE);
1109 
1110 	return 0;
1111 }
1112 
1113 static void arm_ccn_pmu_event_del(struct perf_event *event, int flags)
1114 {
1115 	struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
1116 
1117 	arm_ccn_pmu_event_stop(event, PERF_EF_UPDATE);
1118 
1119 	arm_ccn_pmu_event_release(event);
1120 
1121 	if (!ccn->irq && arm_ccn_pmu_active_counters(ccn) == 0)
1122 		hrtimer_cancel(&ccn->dt.hrtimer);
1123 }
1124 
1125 static void arm_ccn_pmu_event_read(struct perf_event *event)
1126 {
1127 	arm_ccn_pmu_event_update(event);
1128 }
1129 
1130 static void arm_ccn_pmu_enable(struct pmu *pmu)
1131 {
1132 	struct arm_ccn *ccn = pmu_to_arm_ccn(pmu);
1133 
1134 	u32 val = readl(ccn->dt.base + CCN_DT_PMCR);
1135 	val |= CCN_DT_PMCR__PMU_EN;
1136 	writel(val, ccn->dt.base + CCN_DT_PMCR);
1137 }
1138 
1139 static void arm_ccn_pmu_disable(struct pmu *pmu)
1140 {
1141 	struct arm_ccn *ccn = pmu_to_arm_ccn(pmu);
1142 
1143 	u32 val = readl(ccn->dt.base + CCN_DT_PMCR);
1144 	val &= ~CCN_DT_PMCR__PMU_EN;
1145 	writel(val, ccn->dt.base + CCN_DT_PMCR);
1146 }
1147 
1148 static irqreturn_t arm_ccn_pmu_overflow_handler(struct arm_ccn_dt *dt)
1149 {
1150 	u32 pmovsr = readl(dt->base + CCN_DT_PMOVSR);
1151 	int idx;
1152 
1153 	if (!pmovsr)
1154 		return IRQ_NONE;
1155 
1156 	writel(pmovsr, dt->base + CCN_DT_PMOVSR_CLR);
1157 
1158 	BUILD_BUG_ON(CCN_IDX_PMU_CYCLE_COUNTER != CCN_NUM_PMU_EVENT_COUNTERS);
1159 
1160 	for (idx = 0; idx < CCN_NUM_PMU_EVENT_COUNTERS + 1; idx++) {
1161 		struct perf_event *event = dt->pmu_counters[idx].event;
1162 		int overflowed = pmovsr & BIT(idx);
1163 
1164 		WARN_ON_ONCE(overflowed && !event &&
1165 				idx != CCN_IDX_PMU_CYCLE_COUNTER);
1166 
1167 		if (!event || !overflowed)
1168 			continue;
1169 
1170 		arm_ccn_pmu_event_update(event);
1171 	}
1172 
1173 	return IRQ_HANDLED;
1174 }
1175 
1176 static enum hrtimer_restart arm_ccn_pmu_timer_handler(struct hrtimer *hrtimer)
1177 {
1178 	struct arm_ccn_dt *dt = container_of(hrtimer, struct arm_ccn_dt,
1179 			hrtimer);
1180 	unsigned long flags;
1181 
1182 	local_irq_save(flags);
1183 	arm_ccn_pmu_overflow_handler(dt);
1184 	local_irq_restore(flags);
1185 
1186 	hrtimer_forward_now(hrtimer, arm_ccn_pmu_timer_period());
1187 	return HRTIMER_RESTART;
1188 }
1189 
1190 
1191 static int arm_ccn_pmu_offline_cpu(unsigned int cpu, struct hlist_node *node)
1192 {
1193 	struct arm_ccn_dt *dt = hlist_entry_safe(node, struct arm_ccn_dt, node);
1194 	struct arm_ccn *ccn = container_of(dt, struct arm_ccn, dt);
1195 	unsigned int target;
1196 
1197 	if (cpu != dt->cpu)
1198 		return 0;
1199 	target = cpumask_any_but(cpu_online_mask, cpu);
1200 	if (target >= nr_cpu_ids)
1201 		return 0;
1202 	perf_pmu_migrate_context(&dt->pmu, cpu, target);
1203 	dt->cpu = target;
1204 	if (ccn->irq)
1205 		WARN_ON(irq_set_affinity(ccn->irq, cpumask_of(dt->cpu)));
1206 	return 0;
1207 }
1208 
1209 static DEFINE_IDA(arm_ccn_pmu_ida);
1210 
1211 static int arm_ccn_pmu_init(struct arm_ccn *ccn)
1212 {
1213 	int i;
1214 	char *name;
1215 	int err;
1216 
1217 	/* Initialize DT subsystem */
1218 	ccn->dt.base = ccn->base + CCN_REGION_SIZE;
1219 	spin_lock_init(&ccn->dt.config_lock);
1220 	writel(CCN_DT_PMOVSR_CLR__MASK, ccn->dt.base + CCN_DT_PMOVSR_CLR);
1221 	writel(CCN_DT_CTL__DT_EN, ccn->dt.base + CCN_DT_CTL);
1222 	writel(CCN_DT_PMCR__OVFL_INTR_EN | CCN_DT_PMCR__PMU_EN,
1223 			ccn->dt.base + CCN_DT_PMCR);
1224 	writel(0x1, ccn->dt.base + CCN_DT_PMSR_CLR);
1225 	for (i = 0; i < ccn->num_xps; i++) {
1226 		writel(0, ccn->xp[i].base + CCN_XP_DT_CONFIG);
1227 		writel((CCN_XP_DT_CONTROL__WP_ARM_SEL__ALWAYS <<
1228 				CCN_XP_DT_CONTROL__WP_ARM_SEL__SHIFT(0)) |
1229 				(CCN_XP_DT_CONTROL__WP_ARM_SEL__ALWAYS <<
1230 				CCN_XP_DT_CONTROL__WP_ARM_SEL__SHIFT(1)) |
1231 				CCN_XP_DT_CONTROL__DT_ENABLE,
1232 				ccn->xp[i].base + CCN_XP_DT_CONTROL);
1233 	}
1234 	ccn->dt.cmp_mask[CCN_IDX_MASK_ANY].l = ~0;
1235 	ccn->dt.cmp_mask[CCN_IDX_MASK_ANY].h = ~0;
1236 	ccn->dt.cmp_mask[CCN_IDX_MASK_EXACT].l = 0;
1237 	ccn->dt.cmp_mask[CCN_IDX_MASK_EXACT].h = 0;
1238 	ccn->dt.cmp_mask[CCN_IDX_MASK_ORDER].l = ~0;
1239 	ccn->dt.cmp_mask[CCN_IDX_MASK_ORDER].h = ~(0x1 << 15);
1240 	ccn->dt.cmp_mask[CCN_IDX_MASK_OPCODE].l = ~0;
1241 	ccn->dt.cmp_mask[CCN_IDX_MASK_OPCODE].h = ~(0x1f << 9);
1242 
1243 	/* Get a convenient /sys/event_source/devices/ name */
1244 	ccn->dt.id = ida_alloc(&arm_ccn_pmu_ida, GFP_KERNEL);
1245 	if (ccn->dt.id == 0) {
1246 		name = "ccn";
1247 	} else {
1248 		name = devm_kasprintf(ccn->dev, GFP_KERNEL, "ccn_%d",
1249 				      ccn->dt.id);
1250 		if (!name) {
1251 			err = -ENOMEM;
1252 			goto error_choose_name;
1253 		}
1254 	}
1255 
1256 	/* Perf driver registration */
1257 	ccn->dt.pmu = (struct pmu) {
1258 		.module = THIS_MODULE,
1259 		.parent = ccn->dev,
1260 		.attr_groups = arm_ccn_pmu_attr_groups,
1261 		.task_ctx_nr = perf_invalid_context,
1262 		.event_init = arm_ccn_pmu_event_init,
1263 		.add = arm_ccn_pmu_event_add,
1264 		.del = arm_ccn_pmu_event_del,
1265 		.start = arm_ccn_pmu_event_start,
1266 		.stop = arm_ccn_pmu_event_stop,
1267 		.read = arm_ccn_pmu_event_read,
1268 		.pmu_enable = arm_ccn_pmu_enable,
1269 		.pmu_disable = arm_ccn_pmu_disable,
1270 		.capabilities = PERF_PMU_CAP_NO_EXCLUDE,
1271 	};
1272 
1273 	/* No overflow interrupt? Have to use a timer instead. */
1274 	if (!ccn->irq) {
1275 		dev_info(ccn->dev, "No access to interrupts, using timer.\n");
1276 		hrtimer_init(&ccn->dt.hrtimer, CLOCK_MONOTONIC,
1277 				HRTIMER_MODE_REL);
1278 		ccn->dt.hrtimer.function = arm_ccn_pmu_timer_handler;
1279 	}
1280 
1281 	/* Pick one CPU which we will use to collect data from CCN... */
1282 	ccn->dt.cpu = raw_smp_processor_id();
1283 
1284 	/* Also make sure that the overflow interrupt is handled by this CPU */
1285 	if (ccn->irq) {
1286 		err = irq_set_affinity(ccn->irq, cpumask_of(ccn->dt.cpu));
1287 		if (err) {
1288 			dev_err(ccn->dev, "Failed to set interrupt affinity!\n");
1289 			goto error_set_affinity;
1290 		}
1291 	}
1292 
1293 	cpuhp_state_add_instance_nocalls(CPUHP_AP_PERF_ARM_CCN_ONLINE,
1294 					 &ccn->dt.node);
1295 
1296 	err = perf_pmu_register(&ccn->dt.pmu, name, -1);
1297 	if (err)
1298 		goto error_pmu_register;
1299 
1300 	return 0;
1301 
1302 error_pmu_register:
1303 	cpuhp_state_remove_instance_nocalls(CPUHP_AP_PERF_ARM_CCN_ONLINE,
1304 					    &ccn->dt.node);
1305 error_set_affinity:
1306 error_choose_name:
1307 	ida_free(&arm_ccn_pmu_ida, ccn->dt.id);
1308 	for (i = 0; i < ccn->num_xps; i++)
1309 		writel(0, ccn->xp[i].base + CCN_XP_DT_CONTROL);
1310 	writel(0, ccn->dt.base + CCN_DT_PMCR);
1311 	return err;
1312 }
1313 
1314 static void arm_ccn_pmu_cleanup(struct arm_ccn *ccn)
1315 {
1316 	int i;
1317 
1318 	cpuhp_state_remove_instance_nocalls(CPUHP_AP_PERF_ARM_CCN_ONLINE,
1319 					    &ccn->dt.node);
1320 	for (i = 0; i < ccn->num_xps; i++)
1321 		writel(0, ccn->xp[i].base + CCN_XP_DT_CONTROL);
1322 	writel(0, ccn->dt.base + CCN_DT_PMCR);
1323 	perf_pmu_unregister(&ccn->dt.pmu);
1324 	ida_free(&arm_ccn_pmu_ida, ccn->dt.id);
1325 }
1326 
1327 static int arm_ccn_for_each_valid_region(struct arm_ccn *ccn,
1328 		int (*callback)(struct arm_ccn *ccn, int region,
1329 		void __iomem *base, u32 type, u32 id))
1330 {
1331 	int region;
1332 
1333 	for (region = 0; region < CCN_NUM_REGIONS; region++) {
1334 		u32 val, type, id;
1335 		void __iomem *base;
1336 		int err;
1337 
1338 		val = readl(ccn->base + CCN_MN_OLY_COMP_LIST_63_0 +
1339 				4 * (region / 32));
1340 		if (!(val & (1 << (region % 32))))
1341 			continue;
1342 
1343 		base = ccn->base + region * CCN_REGION_SIZE;
1344 		val = readl(base + CCN_ALL_OLY_ID);
1345 		type = (val >> CCN_ALL_OLY_ID__OLY_ID__SHIFT) &
1346 				CCN_ALL_OLY_ID__OLY_ID__MASK;
1347 		id = (val >> CCN_ALL_OLY_ID__NODE_ID__SHIFT) &
1348 				CCN_ALL_OLY_ID__NODE_ID__MASK;
1349 
1350 		err = callback(ccn, region, base, type, id);
1351 		if (err)
1352 			return err;
1353 	}
1354 
1355 	return 0;
1356 }
1357 
1358 static int arm_ccn_get_nodes_num(struct arm_ccn *ccn, int region,
1359 		void __iomem *base, u32 type, u32 id)
1360 {
1361 
1362 	if (type == CCN_TYPE_XP && id >= ccn->num_xps)
1363 		ccn->num_xps = id + 1;
1364 	else if (id >= ccn->num_nodes)
1365 		ccn->num_nodes = id + 1;
1366 
1367 	return 0;
1368 }
1369 
1370 static int arm_ccn_init_nodes(struct arm_ccn *ccn, int region,
1371 		void __iomem *base, u32 type, u32 id)
1372 {
1373 	struct arm_ccn_component *component;
1374 
1375 	dev_dbg(ccn->dev, "Region %d: id=%u, type=0x%02x\n", region, id, type);
1376 
1377 	switch (type) {
1378 	case CCN_TYPE_MN:
1379 		ccn->mn_id = id;
1380 		return 0;
1381 	case CCN_TYPE_DT:
1382 		return 0;
1383 	case CCN_TYPE_XP:
1384 		component = &ccn->xp[id];
1385 		break;
1386 	case CCN_TYPE_SBSX:
1387 		ccn->sbsx_present = 1;
1388 		component = &ccn->node[id];
1389 		break;
1390 	case CCN_TYPE_SBAS:
1391 		ccn->sbas_present = 1;
1392 		fallthrough;
1393 	default:
1394 		component = &ccn->node[id];
1395 		break;
1396 	}
1397 
1398 	component->base = base;
1399 	component->type = type;
1400 
1401 	return 0;
1402 }
1403 
1404 
1405 static irqreturn_t arm_ccn_error_handler(struct arm_ccn *ccn,
1406 		const u32 *err_sig_val)
1407 {
1408 	/* This should be really handled by firmware... */
1409 	dev_err(ccn->dev, "Error reported in %08x%08x%08x%08x%08x%08x.\n",
1410 			err_sig_val[5], err_sig_val[4], err_sig_val[3],
1411 			err_sig_val[2], err_sig_val[1], err_sig_val[0]);
1412 	dev_err(ccn->dev, "Disabling interrupt generation for all errors.\n");
1413 	writel(CCN_MN_ERRINT_STATUS__ALL_ERRORS__DISABLE,
1414 			ccn->base + CCN_MN_ERRINT_STATUS);
1415 
1416 	return IRQ_HANDLED;
1417 }
1418 
1419 
1420 static irqreturn_t arm_ccn_irq_handler(int irq, void *dev_id)
1421 {
1422 	irqreturn_t res = IRQ_NONE;
1423 	struct arm_ccn *ccn = dev_id;
1424 	u32 err_sig_val[6];
1425 	u32 err_or;
1426 	int i;
1427 
1428 	/* PMU overflow is a special case */
1429 	err_or = err_sig_val[0] = readl(ccn->base + CCN_MN_ERR_SIG_VAL_63_0);
1430 	if (err_or & CCN_MN_ERR_SIG_VAL_63_0__DT) {
1431 		err_or &= ~CCN_MN_ERR_SIG_VAL_63_0__DT;
1432 		res = arm_ccn_pmu_overflow_handler(&ccn->dt);
1433 	}
1434 
1435 	/* Have to read all err_sig_vals to clear them */
1436 	for (i = 1; i < ARRAY_SIZE(err_sig_val); i++) {
1437 		err_sig_val[i] = readl(ccn->base +
1438 				CCN_MN_ERR_SIG_VAL_63_0 + i * 4);
1439 		err_or |= err_sig_val[i];
1440 	}
1441 	if (err_or)
1442 		res |= arm_ccn_error_handler(ccn, err_sig_val);
1443 
1444 	if (res != IRQ_NONE)
1445 		writel(CCN_MN_ERRINT_STATUS__INTREQ__DESSERT,
1446 				ccn->base + CCN_MN_ERRINT_STATUS);
1447 
1448 	return res;
1449 }
1450 
1451 
1452 static int arm_ccn_probe(struct platform_device *pdev)
1453 {
1454 	struct arm_ccn *ccn;
1455 	int irq;
1456 	int err;
1457 
1458 	ccn = devm_kzalloc(&pdev->dev, sizeof(*ccn), GFP_KERNEL);
1459 	if (!ccn)
1460 		return -ENOMEM;
1461 	ccn->dev = &pdev->dev;
1462 	platform_set_drvdata(pdev, ccn);
1463 
1464 	ccn->base = devm_platform_ioremap_resource(pdev, 0);
1465 	if (IS_ERR(ccn->base))
1466 		return PTR_ERR(ccn->base);
1467 
1468 	irq = platform_get_irq(pdev, 0);
1469 	if (irq < 0)
1470 		return irq;
1471 
1472 	/* Check if we can use the interrupt */
1473 	writel(CCN_MN_ERRINT_STATUS__PMU_EVENTS__DISABLE,
1474 			ccn->base + CCN_MN_ERRINT_STATUS);
1475 	if (readl(ccn->base + CCN_MN_ERRINT_STATUS) &
1476 			CCN_MN_ERRINT_STATUS__PMU_EVENTS__DISABLED) {
1477 		/* Can set 'disable' bits, so can acknowledge interrupts */
1478 		writel(CCN_MN_ERRINT_STATUS__PMU_EVENTS__ENABLE,
1479 				ccn->base + CCN_MN_ERRINT_STATUS);
1480 		err = devm_request_irq(ccn->dev, irq, arm_ccn_irq_handler,
1481 				       IRQF_NOBALANCING | IRQF_NO_THREAD,
1482 				       dev_name(ccn->dev), ccn);
1483 		if (err)
1484 			return err;
1485 
1486 		ccn->irq = irq;
1487 	}
1488 
1489 
1490 	/* Build topology */
1491 
1492 	err = arm_ccn_for_each_valid_region(ccn, arm_ccn_get_nodes_num);
1493 	if (err)
1494 		return err;
1495 
1496 	ccn->node = devm_kcalloc(ccn->dev, ccn->num_nodes, sizeof(*ccn->node),
1497 				 GFP_KERNEL);
1498 	ccn->xp = devm_kcalloc(ccn->dev, ccn->num_xps, sizeof(*ccn->node),
1499 			       GFP_KERNEL);
1500 	if (!ccn->node || !ccn->xp)
1501 		return -ENOMEM;
1502 
1503 	err = arm_ccn_for_each_valid_region(ccn, arm_ccn_init_nodes);
1504 	if (err)
1505 		return err;
1506 
1507 	return arm_ccn_pmu_init(ccn);
1508 }
1509 
1510 static void arm_ccn_remove(struct platform_device *pdev)
1511 {
1512 	struct arm_ccn *ccn = platform_get_drvdata(pdev);
1513 
1514 	arm_ccn_pmu_cleanup(ccn);
1515 }
1516 
1517 static const struct of_device_id arm_ccn_match[] = {
1518 	{ .compatible = "arm,ccn-502", },
1519 	{ .compatible = "arm,ccn-504", },
1520 	{ .compatible = "arm,ccn-512", },
1521 	{},
1522 };
1523 MODULE_DEVICE_TABLE(of, arm_ccn_match);
1524 
1525 static struct platform_driver arm_ccn_driver = {
1526 	.driver = {
1527 		.name = "arm-ccn",
1528 		.of_match_table = arm_ccn_match,
1529 		.suppress_bind_attrs = true,
1530 	},
1531 	.probe = arm_ccn_probe,
1532 	.remove = arm_ccn_remove,
1533 };
1534 
1535 static int __init arm_ccn_init(void)
1536 {
1537 	int i, ret;
1538 
1539 	ret = cpuhp_setup_state_multi(CPUHP_AP_PERF_ARM_CCN_ONLINE,
1540 				      "perf/arm/ccn:online", NULL,
1541 				      arm_ccn_pmu_offline_cpu);
1542 	if (ret)
1543 		return ret;
1544 
1545 	for (i = 0; i < ARRAY_SIZE(arm_ccn_pmu_events); i++)
1546 		arm_ccn_pmu_events_attrs[i] = &arm_ccn_pmu_events[i].attr.attr;
1547 
1548 	ret = platform_driver_register(&arm_ccn_driver);
1549 	if (ret)
1550 		cpuhp_remove_multi_state(CPUHP_AP_PERF_ARM_CCN_ONLINE);
1551 	return ret;
1552 }
1553 
1554 static void __exit arm_ccn_exit(void)
1555 {
1556 	platform_driver_unregister(&arm_ccn_driver);
1557 	cpuhp_remove_multi_state(CPUHP_AP_PERF_ARM_CCN_ONLINE);
1558 }
1559 
1560 module_init(arm_ccn_init);
1561 module_exit(arm_ccn_exit);
1562 
1563 MODULE_AUTHOR("Pawel Moll <pawel.moll@arm.com>");
1564 MODULE_DESCRIPTION("ARM CCN (Cache Coherent Network) Performance Monitor Driver");
1565 MODULE_LICENSE("GPL v2");
1566