Loading...
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
130struct 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
145struct 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
169struct 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
187static int arm_ccn_node_to_xp(int node)
188{
189 return node / CCN_NUM_XP_PORTS;
190}
191
192static 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
212static 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
218static ssize_t arm_ccn_pmu_format_show(struct device *dev,
219 struct device_attribute *attr, char *buf)
220{
221 struct dev_ext_attribute *ea = container_of(attr,
222 struct dev_ext_attribute, attr);
223
224 return sysfs_emit(buf, "%s\n", (char *)ea->var);
225}
226
227#define CCN_FORMAT_ATTR(_name, _config) \
228 struct dev_ext_attribute arm_ccn_pmu_format_attr_##_name = \
229 { __ATTR(_name, S_IRUGO, arm_ccn_pmu_format_show, \
230 NULL), _config }
231
232static CCN_FORMAT_ATTR(node, "config:0-7");
233static CCN_FORMAT_ATTR(xp, "config:0-7");
234static CCN_FORMAT_ATTR(type, "config:8-15");
235static CCN_FORMAT_ATTR(event, "config:16-23");
236static CCN_FORMAT_ATTR(port, "config:24-25");
237static CCN_FORMAT_ATTR(bus, "config:24-25");
238static CCN_FORMAT_ATTR(vc, "config:26-28");
239static CCN_FORMAT_ATTR(dir, "config:29-29");
240static CCN_FORMAT_ATTR(mask, "config:30-33");
241static CCN_FORMAT_ATTR(cmp_l, "config1:0-62");
242static CCN_FORMAT_ATTR(cmp_h, "config2:0-59");
243
244static struct attribute *arm_ccn_pmu_format_attrs[] = {
245 &arm_ccn_pmu_format_attr_node.attr.attr,
246 &arm_ccn_pmu_format_attr_xp.attr.attr,
247 &arm_ccn_pmu_format_attr_type.attr.attr,
248 &arm_ccn_pmu_format_attr_event.attr.attr,
249 &arm_ccn_pmu_format_attr_port.attr.attr,
250 &arm_ccn_pmu_format_attr_bus.attr.attr,
251 &arm_ccn_pmu_format_attr_vc.attr.attr,
252 &arm_ccn_pmu_format_attr_dir.attr.attr,
253 &arm_ccn_pmu_format_attr_mask.attr.attr,
254 &arm_ccn_pmu_format_attr_cmp_l.attr.attr,
255 &arm_ccn_pmu_format_attr_cmp_h.attr.attr,
256 NULL
257};
258
259static const struct attribute_group arm_ccn_pmu_format_attr_group = {
260 .name = "format",
261 .attrs = arm_ccn_pmu_format_attrs,
262};
263
264
265struct arm_ccn_pmu_event {
266 struct device_attribute attr;
267 u32 type;
268 u32 event;
269 int num_ports;
270 int num_vcs;
271 const char *def;
272 int mask;
273};
274
275#define CCN_EVENT_ATTR(_name) \
276 __ATTR(_name, S_IRUGO, arm_ccn_pmu_event_show, NULL)
277
278/*
279 * Events defined in TRM for MN, HN-I and SBSX are actually watchpoints set on
280 * their ports in XP they are connected to. For the sake of usability they are
281 * explicitly defined here (and translated into a relevant watchpoint in
282 * arm_ccn_pmu_event_init()) so the user can easily request them without deep
283 * knowledge of the flit format.
284 */
285
286#define CCN_EVENT_MN(_name, _def, _mask) { .attr = CCN_EVENT_ATTR(mn_##_name), \
287 .type = CCN_TYPE_MN, .event = CCN_EVENT_WATCHPOINT, \
288 .num_ports = CCN_NUM_XP_PORTS, .num_vcs = CCN_NUM_VCS, \
289 .def = _def, .mask = _mask, }
290
291#define CCN_EVENT_HNI(_name, _def, _mask) { \
292 .attr = CCN_EVENT_ATTR(hni_##_name), .type = CCN_TYPE_HNI, \
293 .event = CCN_EVENT_WATCHPOINT, .num_ports = CCN_NUM_XP_PORTS, \
294 .num_vcs = CCN_NUM_VCS, .def = _def, .mask = _mask, }
295
296#define CCN_EVENT_SBSX(_name, _def, _mask) { \
297 .attr = CCN_EVENT_ATTR(sbsx_##_name), .type = CCN_TYPE_SBSX, \
298 .event = CCN_EVENT_WATCHPOINT, .num_ports = CCN_NUM_XP_PORTS, \
299 .num_vcs = CCN_NUM_VCS, .def = _def, .mask = _mask, }
300
301#define CCN_EVENT_HNF(_name, _event) { .attr = CCN_EVENT_ATTR(hnf_##_name), \
302 .type = CCN_TYPE_HNF, .event = _event, }
303
304#define CCN_EVENT_XP(_name, _event) { .attr = CCN_EVENT_ATTR(xp_##_name), \
305 .type = CCN_TYPE_XP, .event = _event, \
306 .num_ports = CCN_NUM_XP_PORTS, .num_vcs = CCN_NUM_VCS, }
307
308/*
309 * RN-I & RN-D (RN-D = RN-I + DVM) nodes have different type ID depending
310 * on configuration. One of them is picked to represent the whole group,
311 * as they all share the same event types.
312 */
313#define CCN_EVENT_RNI(_name, _event) { .attr = CCN_EVENT_ATTR(rni_##_name), \
314 .type = CCN_TYPE_RNI_3P, .event = _event, }
315
316#define CCN_EVENT_SBAS(_name, _event) { .attr = CCN_EVENT_ATTR(sbas_##_name), \
317 .type = CCN_TYPE_SBAS, .event = _event, }
318
319#define CCN_EVENT_CYCLES(_name) { .attr = CCN_EVENT_ATTR(_name), \
320 .type = CCN_TYPE_CYCLES }
321
322
323static ssize_t arm_ccn_pmu_event_show(struct device *dev,
324 struct device_attribute *attr, char *buf)
325{
326 struct arm_ccn *ccn = pmu_to_arm_ccn(dev_get_drvdata(dev));
327 struct arm_ccn_pmu_event *event = container_of(attr,
328 struct arm_ccn_pmu_event, attr);
329 int res;
330
331 res = sysfs_emit(buf, "type=0x%x", event->type);
332 if (event->event)
333 res += sysfs_emit_at(buf, res, ",event=0x%x", event->event);
334 if (event->def)
335 res += sysfs_emit_at(buf, res, ",%s", event->def);
336 if (event->mask)
337 res += sysfs_emit_at(buf, res, ",mask=0x%x", event->mask);
338
339 /* Arguments required by an event */
340 switch (event->type) {
341 case CCN_TYPE_CYCLES:
342 break;
343 case CCN_TYPE_XP:
344 res += sysfs_emit_at(buf, res, ",xp=?,vc=?");
345 if (event->event == CCN_EVENT_WATCHPOINT)
346 res += sysfs_emit_at(buf, res,
347 ",port=?,dir=?,cmp_l=?,cmp_h=?,mask=?");
348 else
349 res += sysfs_emit_at(buf, res, ",bus=?");
350
351 break;
352 case CCN_TYPE_MN:
353 res += sysfs_emit_at(buf, res, ",node=%d", ccn->mn_id);
354 break;
355 default:
356 res += sysfs_emit_at(buf, res, ",node=?");
357 break;
358 }
359
360 res += sysfs_emit_at(buf, res, "\n");
361
362 return res;
363}
364
365static umode_t arm_ccn_pmu_events_is_visible(struct kobject *kobj,
366 struct attribute *attr, int index)
367{
368 struct device *dev = kobj_to_dev(kobj);
369 struct arm_ccn *ccn = pmu_to_arm_ccn(dev_get_drvdata(dev));
370 struct device_attribute *dev_attr = container_of(attr,
371 struct device_attribute, attr);
372 struct arm_ccn_pmu_event *event = container_of(dev_attr,
373 struct arm_ccn_pmu_event, attr);
374
375 if (event->type == CCN_TYPE_SBAS && !ccn->sbas_present)
376 return 0;
377 if (event->type == CCN_TYPE_SBSX && !ccn->sbsx_present)
378 return 0;
379
380 return attr->mode;
381}
382
383static struct arm_ccn_pmu_event arm_ccn_pmu_events[] = {
384 CCN_EVENT_MN(eobarrier, "dir=1,vc=0,cmp_h=0x1c00", CCN_IDX_MASK_OPCODE),
385 CCN_EVENT_MN(ecbarrier, "dir=1,vc=0,cmp_h=0x1e00", CCN_IDX_MASK_OPCODE),
386 CCN_EVENT_MN(dvmop, "dir=1,vc=0,cmp_h=0x2800", CCN_IDX_MASK_OPCODE),
387 CCN_EVENT_HNI(txdatflits, "dir=1,vc=3", CCN_IDX_MASK_ANY),
388 CCN_EVENT_HNI(rxdatflits, "dir=0,vc=3", CCN_IDX_MASK_ANY),
389 CCN_EVENT_HNI(txreqflits, "dir=1,vc=0", CCN_IDX_MASK_ANY),
390 CCN_EVENT_HNI(rxreqflits, "dir=0,vc=0", CCN_IDX_MASK_ANY),
391 CCN_EVENT_HNI(rxreqflits_order, "dir=0,vc=0,cmp_h=0x8000",
392 CCN_IDX_MASK_ORDER),
393 CCN_EVENT_SBSX(txdatflits, "dir=1,vc=3", CCN_IDX_MASK_ANY),
394 CCN_EVENT_SBSX(rxdatflits, "dir=0,vc=3", CCN_IDX_MASK_ANY),
395 CCN_EVENT_SBSX(txreqflits, "dir=1,vc=0", CCN_IDX_MASK_ANY),
396 CCN_EVENT_SBSX(rxreqflits, "dir=0,vc=0", CCN_IDX_MASK_ANY),
397 CCN_EVENT_SBSX(rxreqflits_order, "dir=0,vc=0,cmp_h=0x8000",
398 CCN_IDX_MASK_ORDER),
399 CCN_EVENT_HNF(cache_miss, 0x1),
400 CCN_EVENT_HNF(l3_sf_cache_access, 0x02),
401 CCN_EVENT_HNF(cache_fill, 0x3),
402 CCN_EVENT_HNF(pocq_retry, 0x4),
403 CCN_EVENT_HNF(pocq_reqs_recvd, 0x5),
404 CCN_EVENT_HNF(sf_hit, 0x6),
405 CCN_EVENT_HNF(sf_evictions, 0x7),
406 CCN_EVENT_HNF(snoops_sent, 0x8),
407 CCN_EVENT_HNF(snoops_broadcast, 0x9),
408 CCN_EVENT_HNF(l3_eviction, 0xa),
409 CCN_EVENT_HNF(l3_fill_invalid_way, 0xb),
410 CCN_EVENT_HNF(mc_retries, 0xc),
411 CCN_EVENT_HNF(mc_reqs, 0xd),
412 CCN_EVENT_HNF(qos_hh_retry, 0xe),
413 CCN_EVENT_RNI(rdata_beats_p0, 0x1),
414 CCN_EVENT_RNI(rdata_beats_p1, 0x2),
415 CCN_EVENT_RNI(rdata_beats_p2, 0x3),
416 CCN_EVENT_RNI(rxdat_flits, 0x4),
417 CCN_EVENT_RNI(txdat_flits, 0x5),
418 CCN_EVENT_RNI(txreq_flits, 0x6),
419 CCN_EVENT_RNI(txreq_flits_retried, 0x7),
420 CCN_EVENT_RNI(rrt_full, 0x8),
421 CCN_EVENT_RNI(wrt_full, 0x9),
422 CCN_EVENT_RNI(txreq_flits_replayed, 0xa),
423 CCN_EVENT_XP(upload_starvation, 0x1),
424 CCN_EVENT_XP(download_starvation, 0x2),
425 CCN_EVENT_XP(respin, 0x3),
426 CCN_EVENT_XP(valid_flit, 0x4),
427 CCN_EVENT_XP(watchpoint, CCN_EVENT_WATCHPOINT),
428 CCN_EVENT_SBAS(rdata_beats_p0, 0x1),
429 CCN_EVENT_SBAS(rxdat_flits, 0x4),
430 CCN_EVENT_SBAS(txdat_flits, 0x5),
431 CCN_EVENT_SBAS(txreq_flits, 0x6),
432 CCN_EVENT_SBAS(txreq_flits_retried, 0x7),
433 CCN_EVENT_SBAS(rrt_full, 0x8),
434 CCN_EVENT_SBAS(wrt_full, 0x9),
435 CCN_EVENT_SBAS(txreq_flits_replayed, 0xa),
436 CCN_EVENT_CYCLES(cycles),
437};
438
439/* Populated in arm_ccn_init() */
440static struct attribute
441 *arm_ccn_pmu_events_attrs[ARRAY_SIZE(arm_ccn_pmu_events) + 1];
442
443static const struct attribute_group arm_ccn_pmu_events_attr_group = {
444 .name = "events",
445 .is_visible = arm_ccn_pmu_events_is_visible,
446 .attrs = arm_ccn_pmu_events_attrs,
447};
448
449
450static u64 *arm_ccn_pmu_get_cmp_mask(struct arm_ccn *ccn, const char *name)
451{
452 unsigned long i;
453
454 if (WARN_ON(!name || !name[0] || !isxdigit(name[0]) || !name[1]))
455 return NULL;
456 i = isdigit(name[0]) ? name[0] - '0' : 0xa + tolower(name[0]) - 'a';
457
458 switch (name[1]) {
459 case 'l':
460 return &ccn->dt.cmp_mask[i].l;
461 case 'h':
462 return &ccn->dt.cmp_mask[i].h;
463 default:
464 return NULL;
465 }
466}
467
468static ssize_t arm_ccn_pmu_cmp_mask_show(struct device *dev,
469 struct device_attribute *attr, char *buf)
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
474 return mask ? sysfs_emit(buf, "0x%016llx\n", *mask) : -EINVAL;
475}
476
477static ssize_t arm_ccn_pmu_cmp_mask_store(struct device *dev,
478 struct device_attribute *attr, const char *buf, size_t count)
479{
480 struct arm_ccn *ccn = pmu_to_arm_ccn(dev_get_drvdata(dev));
481 u64 *mask = arm_ccn_pmu_get_cmp_mask(ccn, attr->attr.name);
482 int err = -EINVAL;
483
484 if (mask)
485 err = kstrtoull(buf, 0, mask);
486
487 return err ? err : count;
488}
489
490#define CCN_CMP_MASK_ATTR(_name) \
491 struct device_attribute arm_ccn_pmu_cmp_mask_attr_##_name = \
492 __ATTR(_name, S_IRUGO | S_IWUSR, \
493 arm_ccn_pmu_cmp_mask_show, arm_ccn_pmu_cmp_mask_store)
494
495#define CCN_CMP_MASK_ATTR_RO(_name) \
496 struct device_attribute arm_ccn_pmu_cmp_mask_attr_##_name = \
497 __ATTR(_name, S_IRUGO, arm_ccn_pmu_cmp_mask_show, NULL)
498
499static CCN_CMP_MASK_ATTR(0l);
500static CCN_CMP_MASK_ATTR(0h);
501static CCN_CMP_MASK_ATTR(1l);
502static CCN_CMP_MASK_ATTR(1h);
503static CCN_CMP_MASK_ATTR(2l);
504static CCN_CMP_MASK_ATTR(2h);
505static CCN_CMP_MASK_ATTR(3l);
506static CCN_CMP_MASK_ATTR(3h);
507static CCN_CMP_MASK_ATTR(4l);
508static CCN_CMP_MASK_ATTR(4h);
509static CCN_CMP_MASK_ATTR(5l);
510static CCN_CMP_MASK_ATTR(5h);
511static CCN_CMP_MASK_ATTR(6l);
512static CCN_CMP_MASK_ATTR(6h);
513static CCN_CMP_MASK_ATTR(7l);
514static CCN_CMP_MASK_ATTR(7h);
515static CCN_CMP_MASK_ATTR_RO(8l);
516static CCN_CMP_MASK_ATTR_RO(8h);
517static CCN_CMP_MASK_ATTR_RO(9l);
518static CCN_CMP_MASK_ATTR_RO(9h);
519static CCN_CMP_MASK_ATTR_RO(al);
520static CCN_CMP_MASK_ATTR_RO(ah);
521static CCN_CMP_MASK_ATTR_RO(bl);
522static CCN_CMP_MASK_ATTR_RO(bh);
523
524static struct attribute *arm_ccn_pmu_cmp_mask_attrs[] = {
525 &arm_ccn_pmu_cmp_mask_attr_0l.attr, &arm_ccn_pmu_cmp_mask_attr_0h.attr,
526 &arm_ccn_pmu_cmp_mask_attr_1l.attr, &arm_ccn_pmu_cmp_mask_attr_1h.attr,
527 &arm_ccn_pmu_cmp_mask_attr_2l.attr, &arm_ccn_pmu_cmp_mask_attr_2h.attr,
528 &arm_ccn_pmu_cmp_mask_attr_3l.attr, &arm_ccn_pmu_cmp_mask_attr_3h.attr,
529 &arm_ccn_pmu_cmp_mask_attr_4l.attr, &arm_ccn_pmu_cmp_mask_attr_4h.attr,
530 &arm_ccn_pmu_cmp_mask_attr_5l.attr, &arm_ccn_pmu_cmp_mask_attr_5h.attr,
531 &arm_ccn_pmu_cmp_mask_attr_6l.attr, &arm_ccn_pmu_cmp_mask_attr_6h.attr,
532 &arm_ccn_pmu_cmp_mask_attr_7l.attr, &arm_ccn_pmu_cmp_mask_attr_7h.attr,
533 &arm_ccn_pmu_cmp_mask_attr_8l.attr, &arm_ccn_pmu_cmp_mask_attr_8h.attr,
534 &arm_ccn_pmu_cmp_mask_attr_9l.attr, &arm_ccn_pmu_cmp_mask_attr_9h.attr,
535 &arm_ccn_pmu_cmp_mask_attr_al.attr, &arm_ccn_pmu_cmp_mask_attr_ah.attr,
536 &arm_ccn_pmu_cmp_mask_attr_bl.attr, &arm_ccn_pmu_cmp_mask_attr_bh.attr,
537 NULL
538};
539
540static const struct attribute_group arm_ccn_pmu_cmp_mask_attr_group = {
541 .name = "cmp_mask",
542 .attrs = arm_ccn_pmu_cmp_mask_attrs,
543};
544
545static ssize_t arm_ccn_pmu_cpumask_show(struct device *dev,
546 struct device_attribute *attr, char *buf)
547{
548 struct arm_ccn *ccn = pmu_to_arm_ccn(dev_get_drvdata(dev));
549
550 return cpumap_print_to_pagebuf(true, buf, cpumask_of(ccn->dt.cpu));
551}
552
553static struct device_attribute arm_ccn_pmu_cpumask_attr =
554 __ATTR(cpumask, S_IRUGO, arm_ccn_pmu_cpumask_show, NULL);
555
556static struct attribute *arm_ccn_pmu_cpumask_attrs[] = {
557 &arm_ccn_pmu_cpumask_attr.attr,
558 NULL,
559};
560
561static const struct attribute_group arm_ccn_pmu_cpumask_attr_group = {
562 .attrs = arm_ccn_pmu_cpumask_attrs,
563};
564
565/*
566 * Default poll period is 10ms, which is way over the top anyway,
567 * as in the worst case scenario (an event every cycle), with 1GHz
568 * clocked bus, the smallest, 32 bit counter will overflow in
569 * more than 4s.
570 */
571static unsigned int arm_ccn_pmu_poll_period_us = 10000;
572module_param_named(pmu_poll_period_us, arm_ccn_pmu_poll_period_us, uint,
573 S_IRUGO | S_IWUSR);
574
575static ktime_t arm_ccn_pmu_timer_period(void)
576{
577 return ns_to_ktime((u64)arm_ccn_pmu_poll_period_us * 1000);
578}
579
580
581static const struct attribute_group *arm_ccn_pmu_attr_groups[] = {
582 &arm_ccn_pmu_events_attr_group,
583 &arm_ccn_pmu_format_attr_group,
584 &arm_ccn_pmu_cmp_mask_attr_group,
585 &arm_ccn_pmu_cpumask_attr_group,
586 NULL
587};
588
589
590static int arm_ccn_pmu_alloc_bit(unsigned long *bitmap, unsigned long size)
591{
592 int bit;
593
594 do {
595 bit = find_first_zero_bit(bitmap, size);
596 if (bit >= size)
597 return -EAGAIN;
598 } while (test_and_set_bit(bit, bitmap));
599
600 return bit;
601}
602
603/* All RN-I and RN-D nodes have identical PMUs */
604static int arm_ccn_pmu_type_eq(u32 a, u32 b)
605{
606 if (a == b)
607 return 1;
608
609 switch (a) {
610 case CCN_TYPE_RNI_1P:
611 case CCN_TYPE_RNI_2P:
612 case CCN_TYPE_RNI_3P:
613 case CCN_TYPE_RND_1P:
614 case CCN_TYPE_RND_2P:
615 case CCN_TYPE_RND_3P:
616 switch (b) {
617 case CCN_TYPE_RNI_1P:
618 case CCN_TYPE_RNI_2P:
619 case CCN_TYPE_RNI_3P:
620 case CCN_TYPE_RND_1P:
621 case CCN_TYPE_RND_2P:
622 case CCN_TYPE_RND_3P:
623 return 1;
624 }
625 break;
626 }
627
628 return 0;
629}
630
631static int arm_ccn_pmu_event_alloc(struct perf_event *event)
632{
633 struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
634 struct hw_perf_event *hw = &event->hw;
635 u32 node_xp, type, event_id;
636 struct arm_ccn_component *source;
637 int bit;
638
639 node_xp = CCN_CONFIG_NODE(event->attr.config);
640 type = CCN_CONFIG_TYPE(event->attr.config);
641 event_id = CCN_CONFIG_EVENT(event->attr.config);
642
643 /* Allocate the cycle counter */
644 if (type == CCN_TYPE_CYCLES) {
645 if (test_and_set_bit(CCN_IDX_PMU_CYCLE_COUNTER,
646 ccn->dt.pmu_counters_mask))
647 return -EAGAIN;
648
649 hw->idx = CCN_IDX_PMU_CYCLE_COUNTER;
650 ccn->dt.pmu_counters[CCN_IDX_PMU_CYCLE_COUNTER].event = event;
651
652 return 0;
653 }
654
655 /* Allocate an event counter */
656 hw->idx = arm_ccn_pmu_alloc_bit(ccn->dt.pmu_counters_mask,
657 CCN_NUM_PMU_EVENT_COUNTERS);
658 if (hw->idx < 0) {
659 dev_dbg(ccn->dev, "No more counters available!\n");
660 return -EAGAIN;
661 }
662
663 if (type == CCN_TYPE_XP)
664 source = &ccn->xp[node_xp];
665 else
666 source = &ccn->node[node_xp];
667 ccn->dt.pmu_counters[hw->idx].source = source;
668
669 /* Allocate an event source or a watchpoint */
670 if (type == CCN_TYPE_XP && event_id == CCN_EVENT_WATCHPOINT)
671 bit = arm_ccn_pmu_alloc_bit(source->xp.dt_cmp_mask,
672 CCN_NUM_XP_WATCHPOINTS);
673 else
674 bit = arm_ccn_pmu_alloc_bit(source->pmu_events_mask,
675 CCN_NUM_PMU_EVENTS);
676 if (bit < 0) {
677 dev_dbg(ccn->dev, "No more event sources/watchpoints on node/XP %d!\n",
678 node_xp);
679 clear_bit(hw->idx, ccn->dt.pmu_counters_mask);
680 return -EAGAIN;
681 }
682 hw->config_base = bit;
683
684 ccn->dt.pmu_counters[hw->idx].event = event;
685
686 return 0;
687}
688
689static void arm_ccn_pmu_event_release(struct perf_event *event)
690{
691 struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
692 struct hw_perf_event *hw = &event->hw;
693
694 if (hw->idx == CCN_IDX_PMU_CYCLE_COUNTER) {
695 clear_bit(CCN_IDX_PMU_CYCLE_COUNTER, ccn->dt.pmu_counters_mask);
696 } else {
697 struct arm_ccn_component *source =
698 ccn->dt.pmu_counters[hw->idx].source;
699
700 if (CCN_CONFIG_TYPE(event->attr.config) == CCN_TYPE_XP &&
701 CCN_CONFIG_EVENT(event->attr.config) ==
702 CCN_EVENT_WATCHPOINT)
703 clear_bit(hw->config_base, source->xp.dt_cmp_mask);
704 else
705 clear_bit(hw->config_base, source->pmu_events_mask);
706 clear_bit(hw->idx, ccn->dt.pmu_counters_mask);
707 }
708
709 ccn->dt.pmu_counters[hw->idx].source = NULL;
710 ccn->dt.pmu_counters[hw->idx].event = NULL;
711}
712
713static int arm_ccn_pmu_event_init(struct perf_event *event)
714{
715 struct arm_ccn *ccn;
716 struct hw_perf_event *hw = &event->hw;
717 u32 node_xp, type, event_id;
718 int valid;
719 int i;
720 struct perf_event *sibling;
721
722 if (event->attr.type != event->pmu->type)
723 return -ENOENT;
724
725 ccn = pmu_to_arm_ccn(event->pmu);
726
727 if (hw->sample_period) {
728 dev_dbg(ccn->dev, "Sampling not supported!\n");
729 return -EOPNOTSUPP;
730 }
731
732 if (has_branch_stack(event)) {
733 dev_dbg(ccn->dev, "Can't exclude execution levels!\n");
734 return -EINVAL;
735 }
736
737 if (event->cpu < 0) {
738 dev_dbg(ccn->dev, "Can't provide per-task data!\n");
739 return -EOPNOTSUPP;
740 }
741 /*
742 * Many perf core operations (eg. events rotation) operate on a
743 * single CPU context. This is obvious for CPU PMUs, where one
744 * expects the same sets of events being observed on all CPUs,
745 * but can lead to issues for off-core PMUs, like CCN, where each
746 * event could be theoretically assigned to a different CPU. To
747 * mitigate this, we enforce CPU assignment to one, selected
748 * processor (the one described in the "cpumask" attribute).
749 */
750 event->cpu = ccn->dt.cpu;
751
752 node_xp = CCN_CONFIG_NODE(event->attr.config);
753 type = CCN_CONFIG_TYPE(event->attr.config);
754 event_id = CCN_CONFIG_EVENT(event->attr.config);
755
756 /* Validate node/xp vs topology */
757 switch (type) {
758 case CCN_TYPE_MN:
759 if (node_xp != ccn->mn_id) {
760 dev_dbg(ccn->dev, "Invalid MN ID %d!\n", node_xp);
761 return -EINVAL;
762 }
763 break;
764 case CCN_TYPE_XP:
765 if (node_xp >= ccn->num_xps) {
766 dev_dbg(ccn->dev, "Invalid XP ID %d!\n", node_xp);
767 return -EINVAL;
768 }
769 break;
770 case CCN_TYPE_CYCLES:
771 break;
772 default:
773 if (node_xp >= ccn->num_nodes) {
774 dev_dbg(ccn->dev, "Invalid node ID %d!\n", node_xp);
775 return -EINVAL;
776 }
777 if (!arm_ccn_pmu_type_eq(type, ccn->node[node_xp].type)) {
778 dev_dbg(ccn->dev, "Invalid type 0x%x for node %d!\n",
779 type, node_xp);
780 return -EINVAL;
781 }
782 break;
783 }
784
785 /* Validate event ID vs available for the type */
786 for (i = 0, valid = 0; i < ARRAY_SIZE(arm_ccn_pmu_events) && !valid;
787 i++) {
788 struct arm_ccn_pmu_event *e = &arm_ccn_pmu_events[i];
789 u32 port = CCN_CONFIG_PORT(event->attr.config);
790 u32 vc = CCN_CONFIG_VC(event->attr.config);
791
792 if (!arm_ccn_pmu_type_eq(type, e->type))
793 continue;
794 if (event_id != e->event)
795 continue;
796 if (e->num_ports && port >= e->num_ports) {
797 dev_dbg(ccn->dev, "Invalid port %d for node/XP %d!\n",
798 port, node_xp);
799 return -EINVAL;
800 }
801 if (e->num_vcs && vc >= e->num_vcs) {
802 dev_dbg(ccn->dev, "Invalid vc %d for node/XP %d!\n",
803 vc, node_xp);
804 return -EINVAL;
805 }
806 valid = 1;
807 }
808 if (!valid) {
809 dev_dbg(ccn->dev, "Invalid event 0x%x for node/XP %d!\n",
810 event_id, node_xp);
811 return -EINVAL;
812 }
813
814 /* Watchpoint-based event for a node is actually set on XP */
815 if (event_id == CCN_EVENT_WATCHPOINT && type != CCN_TYPE_XP) {
816 u32 port;
817
818 type = CCN_TYPE_XP;
819 port = arm_ccn_node_to_xp_port(node_xp);
820 node_xp = arm_ccn_node_to_xp(node_xp);
821
822 arm_ccn_pmu_config_set(&event->attr.config,
823 node_xp, type, port);
824 }
825
826 /*
827 * We must NOT create groups containing mixed PMUs, although software
828 * events are acceptable (for example to create a CCN group
829 * periodically read when a hrtimer aka cpu-clock leader triggers).
830 */
831 if (event->group_leader->pmu != event->pmu &&
832 !is_software_event(event->group_leader))
833 return -EINVAL;
834
835 for_each_sibling_event(sibling, event->group_leader) {
836 if (sibling->pmu != event->pmu &&
837 !is_software_event(sibling))
838 return -EINVAL;
839 }
840
841 return 0;
842}
843
844static u64 arm_ccn_pmu_read_counter(struct arm_ccn *ccn, int idx)
845{
846 u64 res;
847
848 if (idx == CCN_IDX_PMU_CYCLE_COUNTER) {
849#ifdef readq
850 res = readq(ccn->dt.base + CCN_DT_PMCCNTR);
851#else
852 /* 40 bit counter, can do snapshot and read in two parts */
853 writel(0x1, ccn->dt.base + CCN_DT_PMSR_REQ);
854 while (!(readl(ccn->dt.base + CCN_DT_PMSR) & 0x1))
855 ;
856 writel(0x1, ccn->dt.base + CCN_DT_PMSR_CLR);
857 res = readl(ccn->dt.base + CCN_DT_PMCCNTRSR + 4) & 0xff;
858 res <<= 32;
859 res |= readl(ccn->dt.base + CCN_DT_PMCCNTRSR);
860#endif
861 } else {
862 res = readl(ccn->dt.base + CCN_DT_PMEVCNT(idx));
863 }
864
865 return res;
866}
867
868static void arm_ccn_pmu_event_update(struct perf_event *event)
869{
870 struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
871 struct hw_perf_event *hw = &event->hw;
872 u64 prev_count, new_count, mask;
873
874 do {
875 prev_count = local64_read(&hw->prev_count);
876 new_count = arm_ccn_pmu_read_counter(ccn, hw->idx);
877 } while (local64_xchg(&hw->prev_count, new_count) != prev_count);
878
879 mask = (1LLU << (hw->idx == CCN_IDX_PMU_CYCLE_COUNTER ? 40 : 32)) - 1;
880
881 local64_add((new_count - prev_count) & mask, &event->count);
882}
883
884static void arm_ccn_pmu_xp_dt_config(struct perf_event *event, int enable)
885{
886 struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
887 struct hw_perf_event *hw = &event->hw;
888 struct arm_ccn_component *xp;
889 u32 val, dt_cfg;
890
891 /* Nothing to do for cycle counter */
892 if (hw->idx == CCN_IDX_PMU_CYCLE_COUNTER)
893 return;
894
895 if (CCN_CONFIG_TYPE(event->attr.config) == CCN_TYPE_XP)
896 xp = &ccn->xp[CCN_CONFIG_XP(event->attr.config)];
897 else
898 xp = &ccn->xp[arm_ccn_node_to_xp(
899 CCN_CONFIG_NODE(event->attr.config))];
900
901 if (enable)
902 dt_cfg = hw->event_base;
903 else
904 dt_cfg = CCN_XP_DT_CONFIG__DT_CFG__PASS_THROUGH;
905
906 spin_lock(&ccn->dt.config_lock);
907
908 val = readl(xp->base + CCN_XP_DT_CONFIG);
909 val &= ~(CCN_XP_DT_CONFIG__DT_CFG__MASK <<
910 CCN_XP_DT_CONFIG__DT_CFG__SHIFT(hw->idx));
911 val |= dt_cfg << CCN_XP_DT_CONFIG__DT_CFG__SHIFT(hw->idx);
912 writel(val, xp->base + CCN_XP_DT_CONFIG);
913
914 spin_unlock(&ccn->dt.config_lock);
915}
916
917static void arm_ccn_pmu_event_start(struct perf_event *event, int flags)
918{
919 struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
920 struct hw_perf_event *hw = &event->hw;
921
922 local64_set(&event->hw.prev_count,
923 arm_ccn_pmu_read_counter(ccn, hw->idx));
924 hw->state = 0;
925
926 /* Set the DT bus input, engaging the counter */
927 arm_ccn_pmu_xp_dt_config(event, 1);
928}
929
930static void arm_ccn_pmu_event_stop(struct perf_event *event, int flags)
931{
932 struct hw_perf_event *hw = &event->hw;
933
934 /* Disable counting, setting the DT bus to pass-through mode */
935 arm_ccn_pmu_xp_dt_config(event, 0);
936
937 if (flags & PERF_EF_UPDATE)
938 arm_ccn_pmu_event_update(event);
939
940 hw->state |= PERF_HES_STOPPED;
941}
942
943static void arm_ccn_pmu_xp_watchpoint_config(struct perf_event *event)
944{
945 struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
946 struct hw_perf_event *hw = &event->hw;
947 struct arm_ccn_component *source =
948 ccn->dt.pmu_counters[hw->idx].source;
949 unsigned long wp = hw->config_base;
950 u32 val;
951 u64 cmp_l = event->attr.config1;
952 u64 cmp_h = event->attr.config2;
953 u64 mask_l = ccn->dt.cmp_mask[CCN_CONFIG_MASK(event->attr.config)].l;
954 u64 mask_h = ccn->dt.cmp_mask[CCN_CONFIG_MASK(event->attr.config)].h;
955
956 hw->event_base = CCN_XP_DT_CONFIG__DT_CFG__WATCHPOINT(wp);
957
958 /* Direction (RX/TX), device (port) & virtual channel */
959 val = readl(source->base + CCN_XP_DT_INTERFACE_SEL);
960 val &= ~(CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__MASK <<
961 CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__SHIFT(wp));
962 val |= CCN_CONFIG_DIR(event->attr.config) <<
963 CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__SHIFT(wp);
964 val &= ~(CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__MASK <<
965 CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__SHIFT(wp));
966 val |= CCN_CONFIG_PORT(event->attr.config) <<
967 CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__SHIFT(wp);
968 val &= ~(CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__MASK <<
969 CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__SHIFT(wp));
970 val |= CCN_CONFIG_VC(event->attr.config) <<
971 CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__SHIFT(wp);
972 writel(val, source->base + CCN_XP_DT_INTERFACE_SEL);
973
974 /* Comparison values */
975 writel(cmp_l & 0xffffffff, source->base + CCN_XP_DT_CMP_VAL_L(wp));
976 writel((cmp_l >> 32) & 0x7fffffff,
977 source->base + CCN_XP_DT_CMP_VAL_L(wp) + 4);
978 writel(cmp_h & 0xffffffff, source->base + CCN_XP_DT_CMP_VAL_H(wp));
979 writel((cmp_h >> 32) & 0x0fffffff,
980 source->base + CCN_XP_DT_CMP_VAL_H(wp) + 4);
981
982 /* Mask */
983 writel(mask_l & 0xffffffff, source->base + CCN_XP_DT_CMP_MASK_L(wp));
984 writel((mask_l >> 32) & 0x7fffffff,
985 source->base + CCN_XP_DT_CMP_MASK_L(wp) + 4);
986 writel(mask_h & 0xffffffff, source->base + CCN_XP_DT_CMP_MASK_H(wp));
987 writel((mask_h >> 32) & 0x0fffffff,
988 source->base + CCN_XP_DT_CMP_MASK_H(wp) + 4);
989}
990
991static void arm_ccn_pmu_xp_event_config(struct perf_event *event)
992{
993 struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
994 struct hw_perf_event *hw = &event->hw;
995 struct arm_ccn_component *source =
996 ccn->dt.pmu_counters[hw->idx].source;
997 u32 val, id;
998
999 hw->event_base = CCN_XP_DT_CONFIG__DT_CFG__XP_PMU_EVENT(hw->config_base);
1000
1001 id = (CCN_CONFIG_VC(event->attr.config) << 4) |
1002 (CCN_CONFIG_BUS(event->attr.config) << 3) |
1003 (CCN_CONFIG_EVENT(event->attr.config) << 0);
1004
1005 val = readl(source->base + CCN_XP_PMU_EVENT_SEL);
1006 val &= ~(CCN_XP_PMU_EVENT_SEL__ID__MASK <<
1007 CCN_XP_PMU_EVENT_SEL__ID__SHIFT(hw->config_base));
1008 val |= id << CCN_XP_PMU_EVENT_SEL__ID__SHIFT(hw->config_base);
1009 writel(val, source->base + CCN_XP_PMU_EVENT_SEL);
1010}
1011
1012static void arm_ccn_pmu_node_event_config(struct perf_event *event)
1013{
1014 struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
1015 struct hw_perf_event *hw = &event->hw;
1016 struct arm_ccn_component *source =
1017 ccn->dt.pmu_counters[hw->idx].source;
1018 u32 type = CCN_CONFIG_TYPE(event->attr.config);
1019 u32 val, port;
1020
1021 port = arm_ccn_node_to_xp_port(CCN_CONFIG_NODE(event->attr.config));
1022 hw->event_base = CCN_XP_DT_CONFIG__DT_CFG__DEVICE_PMU_EVENT(port,
1023 hw->config_base);
1024
1025 /* These *_event_sel regs should be identical, but let's make sure... */
1026 BUILD_BUG_ON(CCN_HNF_PMU_EVENT_SEL != CCN_SBAS_PMU_EVENT_SEL);
1027 BUILD_BUG_ON(CCN_SBAS_PMU_EVENT_SEL != CCN_RNI_PMU_EVENT_SEL);
1028 BUILD_BUG_ON(CCN_HNF_PMU_EVENT_SEL__ID__SHIFT(1) !=
1029 CCN_SBAS_PMU_EVENT_SEL__ID__SHIFT(1));
1030 BUILD_BUG_ON(CCN_SBAS_PMU_EVENT_SEL__ID__SHIFT(1) !=
1031 CCN_RNI_PMU_EVENT_SEL__ID__SHIFT(1));
1032 BUILD_BUG_ON(CCN_HNF_PMU_EVENT_SEL__ID__MASK !=
1033 CCN_SBAS_PMU_EVENT_SEL__ID__MASK);
1034 BUILD_BUG_ON(CCN_SBAS_PMU_EVENT_SEL__ID__MASK !=
1035 CCN_RNI_PMU_EVENT_SEL__ID__MASK);
1036 if (WARN_ON(type != CCN_TYPE_HNF && type != CCN_TYPE_SBAS &&
1037 !arm_ccn_pmu_type_eq(type, CCN_TYPE_RNI_3P)))
1038 return;
1039
1040 /* Set the event id for the pre-allocated counter */
1041 val = readl(source->base + CCN_HNF_PMU_EVENT_SEL);
1042 val &= ~(CCN_HNF_PMU_EVENT_SEL__ID__MASK <<
1043 CCN_HNF_PMU_EVENT_SEL__ID__SHIFT(hw->config_base));
1044 val |= CCN_CONFIG_EVENT(event->attr.config) <<
1045 CCN_HNF_PMU_EVENT_SEL__ID__SHIFT(hw->config_base);
1046 writel(val, source->base + CCN_HNF_PMU_EVENT_SEL);
1047}
1048
1049static void arm_ccn_pmu_event_config(struct perf_event *event)
1050{
1051 struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
1052 struct hw_perf_event *hw = &event->hw;
1053 u32 xp, offset, val;
1054
1055 /* Cycle counter requires no setup */
1056 if (hw->idx == CCN_IDX_PMU_CYCLE_COUNTER)
1057 return;
1058
1059 if (CCN_CONFIG_TYPE(event->attr.config) == CCN_TYPE_XP)
1060 xp = CCN_CONFIG_XP(event->attr.config);
1061 else
1062 xp = arm_ccn_node_to_xp(CCN_CONFIG_NODE(event->attr.config));
1063
1064 spin_lock(&ccn->dt.config_lock);
1065
1066 /* Set the DT bus "distance" register */
1067 offset = (hw->idx / 4) * 4;
1068 val = readl(ccn->dt.base + CCN_DT_ACTIVE_DSM + offset);
1069 val &= ~(CCN_DT_ACTIVE_DSM__DSM_ID__MASK <<
1070 CCN_DT_ACTIVE_DSM__DSM_ID__SHIFT(hw->idx % 4));
1071 val |= xp << CCN_DT_ACTIVE_DSM__DSM_ID__SHIFT(hw->idx % 4);
1072 writel(val, ccn->dt.base + CCN_DT_ACTIVE_DSM + offset);
1073
1074 if (CCN_CONFIG_TYPE(event->attr.config) == CCN_TYPE_XP) {
1075 if (CCN_CONFIG_EVENT(event->attr.config) ==
1076 CCN_EVENT_WATCHPOINT)
1077 arm_ccn_pmu_xp_watchpoint_config(event);
1078 else
1079 arm_ccn_pmu_xp_event_config(event);
1080 } else {
1081 arm_ccn_pmu_node_event_config(event);
1082 }
1083
1084 spin_unlock(&ccn->dt.config_lock);
1085}
1086
1087static int arm_ccn_pmu_active_counters(struct arm_ccn *ccn)
1088{
1089 return bitmap_weight(ccn->dt.pmu_counters_mask,
1090 CCN_NUM_PMU_EVENT_COUNTERS + 1);
1091}
1092
1093static int arm_ccn_pmu_event_add(struct perf_event *event, int flags)
1094{
1095 int err;
1096 struct hw_perf_event *hw = &event->hw;
1097 struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
1098
1099 err = arm_ccn_pmu_event_alloc(event);
1100 if (err)
1101 return err;
1102
1103 /*
1104 * Pin the timer, so that the overflows are handled by the chosen
1105 * event->cpu (this is the same one as presented in "cpumask"
1106 * attribute).
1107 */
1108 if (!ccn->irq && arm_ccn_pmu_active_counters(ccn) == 1)
1109 hrtimer_start(&ccn->dt.hrtimer, arm_ccn_pmu_timer_period(),
1110 HRTIMER_MODE_REL_PINNED);
1111
1112 arm_ccn_pmu_event_config(event);
1113
1114 hw->state = PERF_HES_STOPPED;
1115
1116 if (flags & PERF_EF_START)
1117 arm_ccn_pmu_event_start(event, PERF_EF_UPDATE);
1118
1119 return 0;
1120}
1121
1122static void arm_ccn_pmu_event_del(struct perf_event *event, int flags)
1123{
1124 struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
1125
1126 arm_ccn_pmu_event_stop(event, PERF_EF_UPDATE);
1127
1128 arm_ccn_pmu_event_release(event);
1129
1130 if (!ccn->irq && arm_ccn_pmu_active_counters(ccn) == 0)
1131 hrtimer_cancel(&ccn->dt.hrtimer);
1132}
1133
1134static void arm_ccn_pmu_event_read(struct perf_event *event)
1135{
1136 arm_ccn_pmu_event_update(event);
1137}
1138
1139static void arm_ccn_pmu_enable(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
1148static void arm_ccn_pmu_disable(struct pmu *pmu)
1149{
1150 struct arm_ccn *ccn = pmu_to_arm_ccn(pmu);
1151
1152 u32 val = readl(ccn->dt.base + CCN_DT_PMCR);
1153 val &= ~CCN_DT_PMCR__PMU_EN;
1154 writel(val, ccn->dt.base + CCN_DT_PMCR);
1155}
1156
1157static irqreturn_t arm_ccn_pmu_overflow_handler(struct arm_ccn_dt *dt)
1158{
1159 u32 pmovsr = readl(dt->base + CCN_DT_PMOVSR);
1160 int idx;
1161
1162 if (!pmovsr)
1163 return IRQ_NONE;
1164
1165 writel(pmovsr, dt->base + CCN_DT_PMOVSR_CLR);
1166
1167 BUILD_BUG_ON(CCN_IDX_PMU_CYCLE_COUNTER != CCN_NUM_PMU_EVENT_COUNTERS);
1168
1169 for (idx = 0; idx < CCN_NUM_PMU_EVENT_COUNTERS + 1; idx++) {
1170 struct perf_event *event = dt->pmu_counters[idx].event;
1171 int overflowed = pmovsr & BIT(idx);
1172
1173 WARN_ON_ONCE(overflowed && !event &&
1174 idx != CCN_IDX_PMU_CYCLE_COUNTER);
1175
1176 if (!event || !overflowed)
1177 continue;
1178
1179 arm_ccn_pmu_event_update(event);
1180 }
1181
1182 return IRQ_HANDLED;
1183}
1184
1185static enum hrtimer_restart arm_ccn_pmu_timer_handler(struct hrtimer *hrtimer)
1186{
1187 struct arm_ccn_dt *dt = container_of(hrtimer, struct arm_ccn_dt,
1188 hrtimer);
1189 unsigned long flags;
1190
1191 local_irq_save(flags);
1192 arm_ccn_pmu_overflow_handler(dt);
1193 local_irq_restore(flags);
1194
1195 hrtimer_forward_now(hrtimer, arm_ccn_pmu_timer_period());
1196 return HRTIMER_RESTART;
1197}
1198
1199
1200static int arm_ccn_pmu_offline_cpu(unsigned int cpu, struct hlist_node *node)
1201{
1202 struct arm_ccn_dt *dt = hlist_entry_safe(node, struct arm_ccn_dt, node);
1203 struct arm_ccn *ccn = container_of(dt, struct arm_ccn, dt);
1204 unsigned int target;
1205
1206 if (cpu != dt->cpu)
1207 return 0;
1208 target = cpumask_any_but(cpu_online_mask, cpu);
1209 if (target >= nr_cpu_ids)
1210 return 0;
1211 perf_pmu_migrate_context(&dt->pmu, cpu, target);
1212 dt->cpu = target;
1213 if (ccn->irq)
1214 WARN_ON(irq_set_affinity(ccn->irq, cpumask_of(dt->cpu)));
1215 return 0;
1216}
1217
1218static DEFINE_IDA(arm_ccn_pmu_ida);
1219
1220static int arm_ccn_pmu_init(struct arm_ccn *ccn)
1221{
1222 int i;
1223 char *name;
1224 int err;
1225
1226 /* Initialize DT subsystem */
1227 ccn->dt.base = ccn->base + CCN_REGION_SIZE;
1228 spin_lock_init(&ccn->dt.config_lock);
1229 writel(CCN_DT_PMOVSR_CLR__MASK, ccn->dt.base + CCN_DT_PMOVSR_CLR);
1230 writel(CCN_DT_CTL__DT_EN, ccn->dt.base + CCN_DT_CTL);
1231 writel(CCN_DT_PMCR__OVFL_INTR_EN | CCN_DT_PMCR__PMU_EN,
1232 ccn->dt.base + CCN_DT_PMCR);
1233 writel(0x1, ccn->dt.base + CCN_DT_PMSR_CLR);
1234 for (i = 0; i < ccn->num_xps; i++) {
1235 writel(0, ccn->xp[i].base + CCN_XP_DT_CONFIG);
1236 writel((CCN_XP_DT_CONTROL__WP_ARM_SEL__ALWAYS <<
1237 CCN_XP_DT_CONTROL__WP_ARM_SEL__SHIFT(0)) |
1238 (CCN_XP_DT_CONTROL__WP_ARM_SEL__ALWAYS <<
1239 CCN_XP_DT_CONTROL__WP_ARM_SEL__SHIFT(1)) |
1240 CCN_XP_DT_CONTROL__DT_ENABLE,
1241 ccn->xp[i].base + CCN_XP_DT_CONTROL);
1242 }
1243 ccn->dt.cmp_mask[CCN_IDX_MASK_ANY].l = ~0;
1244 ccn->dt.cmp_mask[CCN_IDX_MASK_ANY].h = ~0;
1245 ccn->dt.cmp_mask[CCN_IDX_MASK_EXACT].l = 0;
1246 ccn->dt.cmp_mask[CCN_IDX_MASK_EXACT].h = 0;
1247 ccn->dt.cmp_mask[CCN_IDX_MASK_ORDER].l = ~0;
1248 ccn->dt.cmp_mask[CCN_IDX_MASK_ORDER].h = ~(0x1 << 15);
1249 ccn->dt.cmp_mask[CCN_IDX_MASK_OPCODE].l = ~0;
1250 ccn->dt.cmp_mask[CCN_IDX_MASK_OPCODE].h = ~(0x1f << 9);
1251
1252 /* Get a convenient /sys/event_source/devices/ name */
1253 ccn->dt.id = ida_alloc(&arm_ccn_pmu_ida, GFP_KERNEL);
1254 if (ccn->dt.id == 0) {
1255 name = "ccn";
1256 } else {
1257 name = devm_kasprintf(ccn->dev, GFP_KERNEL, "ccn_%d",
1258 ccn->dt.id);
1259 if (!name) {
1260 err = -ENOMEM;
1261 goto error_choose_name;
1262 }
1263 }
1264
1265 /* Perf driver registration */
1266 ccn->dt.pmu = (struct pmu) {
1267 .module = THIS_MODULE,
1268 .attr_groups = arm_ccn_pmu_attr_groups,
1269 .task_ctx_nr = perf_invalid_context,
1270 .event_init = arm_ccn_pmu_event_init,
1271 .add = arm_ccn_pmu_event_add,
1272 .del = arm_ccn_pmu_event_del,
1273 .start = arm_ccn_pmu_event_start,
1274 .stop = arm_ccn_pmu_event_stop,
1275 .read = arm_ccn_pmu_event_read,
1276 .pmu_enable = arm_ccn_pmu_enable,
1277 .pmu_disable = arm_ccn_pmu_disable,
1278 .capabilities = PERF_PMU_CAP_NO_EXCLUDE,
1279 };
1280
1281 /* No overflow interrupt? Have to use a timer instead. */
1282 if (!ccn->irq) {
1283 dev_info(ccn->dev, "No access to interrupts, using timer.\n");
1284 hrtimer_init(&ccn->dt.hrtimer, CLOCK_MONOTONIC,
1285 HRTIMER_MODE_REL);
1286 ccn->dt.hrtimer.function = arm_ccn_pmu_timer_handler;
1287 }
1288
1289 /* Pick one CPU which we will use to collect data from CCN... */
1290 ccn->dt.cpu = raw_smp_processor_id();
1291
1292 /* Also make sure that the overflow interrupt is handled by this CPU */
1293 if (ccn->irq) {
1294 err = irq_set_affinity(ccn->irq, cpumask_of(ccn->dt.cpu));
1295 if (err) {
1296 dev_err(ccn->dev, "Failed to set interrupt affinity!\n");
1297 goto error_set_affinity;
1298 }
1299 }
1300
1301 cpuhp_state_add_instance_nocalls(CPUHP_AP_PERF_ARM_CCN_ONLINE,
1302 &ccn->dt.node);
1303
1304 err = perf_pmu_register(&ccn->dt.pmu, name, -1);
1305 if (err)
1306 goto error_pmu_register;
1307
1308 return 0;
1309
1310error_pmu_register:
1311 cpuhp_state_remove_instance_nocalls(CPUHP_AP_PERF_ARM_CCN_ONLINE,
1312 &ccn->dt.node);
1313error_set_affinity:
1314error_choose_name:
1315 ida_free(&arm_ccn_pmu_ida, ccn->dt.id);
1316 for (i = 0; i < ccn->num_xps; i++)
1317 writel(0, ccn->xp[i].base + CCN_XP_DT_CONTROL);
1318 writel(0, ccn->dt.base + CCN_DT_PMCR);
1319 return err;
1320}
1321
1322static void arm_ccn_pmu_cleanup(struct arm_ccn *ccn)
1323{
1324 int i;
1325
1326 cpuhp_state_remove_instance_nocalls(CPUHP_AP_PERF_ARM_CCN_ONLINE,
1327 &ccn->dt.node);
1328 for (i = 0; i < ccn->num_xps; i++)
1329 writel(0, ccn->xp[i].base + CCN_XP_DT_CONTROL);
1330 writel(0, ccn->dt.base + CCN_DT_PMCR);
1331 perf_pmu_unregister(&ccn->dt.pmu);
1332 ida_free(&arm_ccn_pmu_ida, ccn->dt.id);
1333}
1334
1335static int arm_ccn_for_each_valid_region(struct arm_ccn *ccn,
1336 int (*callback)(struct arm_ccn *ccn, int region,
1337 void __iomem *base, u32 type, u32 id))
1338{
1339 int region;
1340
1341 for (region = 0; region < CCN_NUM_REGIONS; region++) {
1342 u32 val, type, id;
1343 void __iomem *base;
1344 int err;
1345
1346 val = readl(ccn->base + CCN_MN_OLY_COMP_LIST_63_0 +
1347 4 * (region / 32));
1348 if (!(val & (1 << (region % 32))))
1349 continue;
1350
1351 base = ccn->base + region * CCN_REGION_SIZE;
1352 val = readl(base + CCN_ALL_OLY_ID);
1353 type = (val >> CCN_ALL_OLY_ID__OLY_ID__SHIFT) &
1354 CCN_ALL_OLY_ID__OLY_ID__MASK;
1355 id = (val >> CCN_ALL_OLY_ID__NODE_ID__SHIFT) &
1356 CCN_ALL_OLY_ID__NODE_ID__MASK;
1357
1358 err = callback(ccn, region, base, type, id);
1359 if (err)
1360 return err;
1361 }
1362
1363 return 0;
1364}
1365
1366static int arm_ccn_get_nodes_num(struct arm_ccn *ccn, int region,
1367 void __iomem *base, u32 type, u32 id)
1368{
1369
1370 if (type == CCN_TYPE_XP && id >= ccn->num_xps)
1371 ccn->num_xps = id + 1;
1372 else if (id >= ccn->num_nodes)
1373 ccn->num_nodes = id + 1;
1374
1375 return 0;
1376}
1377
1378static int arm_ccn_init_nodes(struct arm_ccn *ccn, int region,
1379 void __iomem *base, u32 type, u32 id)
1380{
1381 struct arm_ccn_component *component;
1382
1383 dev_dbg(ccn->dev, "Region %d: id=%u, type=0x%02x\n", region, id, type);
1384
1385 switch (type) {
1386 case CCN_TYPE_MN:
1387 ccn->mn_id = id;
1388 return 0;
1389 case CCN_TYPE_DT:
1390 return 0;
1391 case CCN_TYPE_XP:
1392 component = &ccn->xp[id];
1393 break;
1394 case CCN_TYPE_SBSX:
1395 ccn->sbsx_present = 1;
1396 component = &ccn->node[id];
1397 break;
1398 case CCN_TYPE_SBAS:
1399 ccn->sbas_present = 1;
1400 fallthrough;
1401 default:
1402 component = &ccn->node[id];
1403 break;
1404 }
1405
1406 component->base = base;
1407 component->type = type;
1408
1409 return 0;
1410}
1411
1412
1413static irqreturn_t arm_ccn_error_handler(struct arm_ccn *ccn,
1414 const u32 *err_sig_val)
1415{
1416 /* This should be really handled by firmware... */
1417 dev_err(ccn->dev, "Error reported in %08x%08x%08x%08x%08x%08x.\n",
1418 err_sig_val[5], err_sig_val[4], err_sig_val[3],
1419 err_sig_val[2], err_sig_val[1], err_sig_val[0]);
1420 dev_err(ccn->dev, "Disabling interrupt generation for all errors.\n");
1421 writel(CCN_MN_ERRINT_STATUS__ALL_ERRORS__DISABLE,
1422 ccn->base + CCN_MN_ERRINT_STATUS);
1423
1424 return IRQ_HANDLED;
1425}
1426
1427
1428static irqreturn_t arm_ccn_irq_handler(int irq, void *dev_id)
1429{
1430 irqreturn_t res = IRQ_NONE;
1431 struct arm_ccn *ccn = dev_id;
1432 u32 err_sig_val[6];
1433 u32 err_or;
1434 int i;
1435
1436 /* PMU overflow is a special case */
1437 err_or = err_sig_val[0] = readl(ccn->base + CCN_MN_ERR_SIG_VAL_63_0);
1438 if (err_or & CCN_MN_ERR_SIG_VAL_63_0__DT) {
1439 err_or &= ~CCN_MN_ERR_SIG_VAL_63_0__DT;
1440 res = arm_ccn_pmu_overflow_handler(&ccn->dt);
1441 }
1442
1443 /* Have to read all err_sig_vals to clear them */
1444 for (i = 1; i < ARRAY_SIZE(err_sig_val); i++) {
1445 err_sig_val[i] = readl(ccn->base +
1446 CCN_MN_ERR_SIG_VAL_63_0 + i * 4);
1447 err_or |= err_sig_val[i];
1448 }
1449 if (err_or)
1450 res |= arm_ccn_error_handler(ccn, err_sig_val);
1451
1452 if (res != IRQ_NONE)
1453 writel(CCN_MN_ERRINT_STATUS__INTREQ__DESSERT,
1454 ccn->base + CCN_MN_ERRINT_STATUS);
1455
1456 return res;
1457}
1458
1459
1460static int arm_ccn_probe(struct platform_device *pdev)
1461{
1462 struct arm_ccn *ccn;
1463 int irq;
1464 int err;
1465
1466 ccn = devm_kzalloc(&pdev->dev, sizeof(*ccn), GFP_KERNEL);
1467 if (!ccn)
1468 return -ENOMEM;
1469 ccn->dev = &pdev->dev;
1470 platform_set_drvdata(pdev, ccn);
1471
1472 ccn->base = devm_platform_ioremap_resource(pdev, 0);
1473 if (IS_ERR(ccn->base))
1474 return PTR_ERR(ccn->base);
1475
1476 irq = platform_get_irq(pdev, 0);
1477 if (irq < 0)
1478 return irq;
1479
1480 /* Check if we can use the interrupt */
1481 writel(CCN_MN_ERRINT_STATUS__PMU_EVENTS__DISABLE,
1482 ccn->base + CCN_MN_ERRINT_STATUS);
1483 if (readl(ccn->base + CCN_MN_ERRINT_STATUS) &
1484 CCN_MN_ERRINT_STATUS__PMU_EVENTS__DISABLED) {
1485 /* Can set 'disable' bits, so can acknowledge interrupts */
1486 writel(CCN_MN_ERRINT_STATUS__PMU_EVENTS__ENABLE,
1487 ccn->base + CCN_MN_ERRINT_STATUS);
1488 err = devm_request_irq(ccn->dev, irq, arm_ccn_irq_handler,
1489 IRQF_NOBALANCING | IRQF_NO_THREAD,
1490 dev_name(ccn->dev), ccn);
1491 if (err)
1492 return err;
1493
1494 ccn->irq = irq;
1495 }
1496
1497
1498 /* Build topology */
1499
1500 err = arm_ccn_for_each_valid_region(ccn, arm_ccn_get_nodes_num);
1501 if (err)
1502 return err;
1503
1504 ccn->node = devm_kcalloc(ccn->dev, ccn->num_nodes, sizeof(*ccn->node),
1505 GFP_KERNEL);
1506 ccn->xp = devm_kcalloc(ccn->dev, ccn->num_xps, sizeof(*ccn->node),
1507 GFP_KERNEL);
1508 if (!ccn->node || !ccn->xp)
1509 return -ENOMEM;
1510
1511 err = arm_ccn_for_each_valid_region(ccn, arm_ccn_init_nodes);
1512 if (err)
1513 return err;
1514
1515 return arm_ccn_pmu_init(ccn);
1516}
1517
1518static int arm_ccn_remove(struct platform_device *pdev)
1519{
1520 struct arm_ccn *ccn = platform_get_drvdata(pdev);
1521
1522 arm_ccn_pmu_cleanup(ccn);
1523
1524 return 0;
1525}
1526
1527static const struct of_device_id arm_ccn_match[] = {
1528 { .compatible = "arm,ccn-502", },
1529 { .compatible = "arm,ccn-504", },
1530 { .compatible = "arm,ccn-512", },
1531 {},
1532};
1533MODULE_DEVICE_TABLE(of, arm_ccn_match);
1534
1535static struct platform_driver arm_ccn_driver = {
1536 .driver = {
1537 .name = "arm-ccn",
1538 .of_match_table = arm_ccn_match,
1539 .suppress_bind_attrs = true,
1540 },
1541 .probe = arm_ccn_probe,
1542 .remove = arm_ccn_remove,
1543};
1544
1545static int __init arm_ccn_init(void)
1546{
1547 int i, ret;
1548
1549 ret = cpuhp_setup_state_multi(CPUHP_AP_PERF_ARM_CCN_ONLINE,
1550 "perf/arm/ccn:online", NULL,
1551 arm_ccn_pmu_offline_cpu);
1552 if (ret)
1553 return ret;
1554
1555 for (i = 0; i < ARRAY_SIZE(arm_ccn_pmu_events); i++)
1556 arm_ccn_pmu_events_attrs[i] = &arm_ccn_pmu_events[i].attr.attr;
1557
1558 ret = platform_driver_register(&arm_ccn_driver);
1559 if (ret)
1560 cpuhp_remove_multi_state(CPUHP_AP_PERF_ARM_CCN_ONLINE);
1561 return ret;
1562}
1563
1564static void __exit arm_ccn_exit(void)
1565{
1566 platform_driver_unregister(&arm_ccn_driver);
1567 cpuhp_remove_multi_state(CPUHP_AP_PERF_ARM_CCN_ONLINE);
1568}
1569
1570module_init(arm_ccn_init);
1571module_exit(arm_ccn_exit);
1572
1573MODULE_AUTHOR("Pawel Moll <pawel.moll@arm.com>");
1574MODULE_LICENSE("GPL v2");
1/*
2 * This program is free software; you can redistribute it and/or modify
3 * it under the terms of the GNU General Public License version 2 as
4 * published by the Free Software Foundation.
5 *
6 * This program is distributed in the hope that it will be useful,
7 * but WITHOUT ANY WARRANTY; without even the implied warranty of
8 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
9 * GNU General Public License for more details.
10 *
11 * Copyright (C) 2014 ARM Limited
12 */
13
14#include <linux/ctype.h>
15#include <linux/hrtimer.h>
16#include <linux/idr.h>
17#include <linux/interrupt.h>
18#include <linux/io.h>
19#include <linux/module.h>
20#include <linux/perf_event.h>
21#include <linux/platform_device.h>
22#include <linux/slab.h>
23
24#define CCN_NUM_XP_PORTS 2
25#define CCN_NUM_VCS 4
26#define CCN_NUM_REGIONS 256
27#define CCN_REGION_SIZE 0x10000
28
29#define CCN_ALL_OLY_ID 0xff00
30#define CCN_ALL_OLY_ID__OLY_ID__SHIFT 0
31#define CCN_ALL_OLY_ID__OLY_ID__MASK 0x1f
32#define CCN_ALL_OLY_ID__NODE_ID__SHIFT 8
33#define CCN_ALL_OLY_ID__NODE_ID__MASK 0x3f
34
35#define CCN_MN_ERRINT_STATUS 0x0008
36#define CCN_MN_ERRINT_STATUS__INTREQ__DESSERT 0x11
37#define CCN_MN_ERRINT_STATUS__ALL_ERRORS__ENABLE 0x02
38#define CCN_MN_ERRINT_STATUS__ALL_ERRORS__DISABLED 0x20
39#define CCN_MN_ERRINT_STATUS__ALL_ERRORS__DISABLE 0x22
40#define CCN_MN_ERRINT_STATUS__CORRECTED_ERRORS_ENABLE 0x04
41#define CCN_MN_ERRINT_STATUS__CORRECTED_ERRORS_DISABLED 0x40
42#define CCN_MN_ERRINT_STATUS__CORRECTED_ERRORS_DISABLE 0x44
43#define CCN_MN_ERRINT_STATUS__PMU_EVENTS__ENABLE 0x08
44#define CCN_MN_ERRINT_STATUS__PMU_EVENTS__DISABLED 0x80
45#define CCN_MN_ERRINT_STATUS__PMU_EVENTS__DISABLE 0x88
46#define CCN_MN_OLY_COMP_LIST_63_0 0x01e0
47#define CCN_MN_ERR_SIG_VAL_63_0 0x0300
48#define CCN_MN_ERR_SIG_VAL_63_0__DT (1 << 1)
49
50#define CCN_DT_ACTIVE_DSM 0x0000
51#define CCN_DT_ACTIVE_DSM__DSM_ID__SHIFT(n) ((n) * 8)
52#define CCN_DT_ACTIVE_DSM__DSM_ID__MASK 0xff
53#define CCN_DT_CTL 0x0028
54#define CCN_DT_CTL__DT_EN (1 << 0)
55#define CCN_DT_PMEVCNT(n) (0x0100 + (n) * 0x8)
56#define CCN_DT_PMCCNTR 0x0140
57#define CCN_DT_PMCCNTRSR 0x0190
58#define CCN_DT_PMOVSR 0x0198
59#define CCN_DT_PMOVSR_CLR 0x01a0
60#define CCN_DT_PMOVSR_CLR__MASK 0x1f
61#define CCN_DT_PMCR 0x01a8
62#define CCN_DT_PMCR__OVFL_INTR_EN (1 << 6)
63#define CCN_DT_PMCR__PMU_EN (1 << 0)
64#define CCN_DT_PMSR 0x01b0
65#define CCN_DT_PMSR_REQ 0x01b8
66#define CCN_DT_PMSR_CLR 0x01c0
67
68#define CCN_HNF_PMU_EVENT_SEL 0x0600
69#define CCN_HNF_PMU_EVENT_SEL__ID__SHIFT(n) ((n) * 4)
70#define CCN_HNF_PMU_EVENT_SEL__ID__MASK 0xf
71
72#define CCN_XP_DT_CONFIG 0x0300
73#define CCN_XP_DT_CONFIG__DT_CFG__SHIFT(n) ((n) * 4)
74#define CCN_XP_DT_CONFIG__DT_CFG__MASK 0xf
75#define CCN_XP_DT_CONFIG__DT_CFG__PASS_THROUGH 0x0
76#define CCN_XP_DT_CONFIG__DT_CFG__WATCHPOINT_0_OR_1 0x1
77#define CCN_XP_DT_CONFIG__DT_CFG__WATCHPOINT(n) (0x2 + (n))
78#define CCN_XP_DT_CONFIG__DT_CFG__XP_PMU_EVENT(n) (0x4 + (n))
79#define CCN_XP_DT_CONFIG__DT_CFG__DEVICE_PMU_EVENT(d, n) (0x8 + (d) * 4 + (n))
80#define CCN_XP_DT_INTERFACE_SEL 0x0308
81#define CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__SHIFT(n) (0 + (n) * 8)
82#define CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__MASK 0x1
83#define CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__SHIFT(n) (1 + (n) * 8)
84#define CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__MASK 0x1
85#define CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__SHIFT(n) (2 + (n) * 8)
86#define CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__MASK 0x3
87#define CCN_XP_DT_CMP_VAL_L(n) (0x0310 + (n) * 0x40)
88#define CCN_XP_DT_CMP_VAL_H(n) (0x0318 + (n) * 0x40)
89#define CCN_XP_DT_CMP_MASK_L(n) (0x0320 + (n) * 0x40)
90#define CCN_XP_DT_CMP_MASK_H(n) (0x0328 + (n) * 0x40)
91#define CCN_XP_DT_CONTROL 0x0370
92#define CCN_XP_DT_CONTROL__DT_ENABLE (1 << 0)
93#define CCN_XP_DT_CONTROL__WP_ARM_SEL__SHIFT(n) (12 + (n) * 4)
94#define CCN_XP_DT_CONTROL__WP_ARM_SEL__MASK 0xf
95#define CCN_XP_DT_CONTROL__WP_ARM_SEL__ALWAYS 0xf
96#define CCN_XP_PMU_EVENT_SEL 0x0600
97#define CCN_XP_PMU_EVENT_SEL__ID__SHIFT(n) ((n) * 7)
98#define CCN_XP_PMU_EVENT_SEL__ID__MASK 0x3f
99
100#define CCN_SBAS_PMU_EVENT_SEL 0x0600
101#define CCN_SBAS_PMU_EVENT_SEL__ID__SHIFT(n) ((n) * 4)
102#define CCN_SBAS_PMU_EVENT_SEL__ID__MASK 0xf
103
104#define CCN_RNI_PMU_EVENT_SEL 0x0600
105#define CCN_RNI_PMU_EVENT_SEL__ID__SHIFT(n) ((n) * 4)
106#define CCN_RNI_PMU_EVENT_SEL__ID__MASK 0xf
107
108#define CCN_TYPE_MN 0x01
109#define CCN_TYPE_DT 0x02
110#define CCN_TYPE_HNF 0x04
111#define CCN_TYPE_HNI 0x05
112#define CCN_TYPE_XP 0x08
113#define CCN_TYPE_SBSX 0x0c
114#define CCN_TYPE_SBAS 0x10
115#define CCN_TYPE_RNI_1P 0x14
116#define CCN_TYPE_RNI_2P 0x15
117#define CCN_TYPE_RNI_3P 0x16
118#define CCN_TYPE_RND_1P 0x18 /* RN-D = RN-I + DVM */
119#define CCN_TYPE_RND_2P 0x19
120#define CCN_TYPE_RND_3P 0x1a
121#define CCN_TYPE_CYCLES 0xff /* Pseudotype */
122
123#define CCN_EVENT_WATCHPOINT 0xfe /* Pseudoevent */
124
125#define CCN_NUM_PMU_EVENTS 4
126#define CCN_NUM_XP_WATCHPOINTS 2 /* See DT.dbg_id.num_watchpoints */
127#define CCN_NUM_PMU_EVENT_COUNTERS 8 /* See DT.dbg_id.num_pmucntr */
128#define CCN_IDX_PMU_CYCLE_COUNTER CCN_NUM_PMU_EVENT_COUNTERS
129
130#define CCN_NUM_PREDEFINED_MASKS 4
131#define CCN_IDX_MASK_ANY (CCN_NUM_PMU_EVENT_COUNTERS + 0)
132#define CCN_IDX_MASK_EXACT (CCN_NUM_PMU_EVENT_COUNTERS + 1)
133#define CCN_IDX_MASK_ORDER (CCN_NUM_PMU_EVENT_COUNTERS + 2)
134#define CCN_IDX_MASK_OPCODE (CCN_NUM_PMU_EVENT_COUNTERS + 3)
135
136struct arm_ccn_component {
137 void __iomem *base;
138 u32 type;
139
140 DECLARE_BITMAP(pmu_events_mask, CCN_NUM_PMU_EVENTS);
141 union {
142 struct {
143 DECLARE_BITMAP(dt_cmp_mask, CCN_NUM_XP_WATCHPOINTS);
144 } xp;
145 };
146};
147
148#define pmu_to_arm_ccn(_pmu) container_of(container_of(_pmu, \
149 struct arm_ccn_dt, pmu), struct arm_ccn, dt)
150
151struct arm_ccn_dt {
152 int id;
153 void __iomem *base;
154
155 spinlock_t config_lock;
156
157 DECLARE_BITMAP(pmu_counters_mask, CCN_NUM_PMU_EVENT_COUNTERS + 1);
158 struct {
159 struct arm_ccn_component *source;
160 struct perf_event *event;
161 } pmu_counters[CCN_NUM_PMU_EVENT_COUNTERS + 1];
162
163 struct {
164 u64 l, h;
165 } cmp_mask[CCN_NUM_PMU_EVENT_COUNTERS + CCN_NUM_PREDEFINED_MASKS];
166
167 struct hrtimer hrtimer;
168
169 cpumask_t cpu;
170 struct hlist_node node;
171
172 struct pmu pmu;
173};
174
175struct arm_ccn {
176 struct device *dev;
177 void __iomem *base;
178 unsigned int irq;
179
180 unsigned sbas_present:1;
181 unsigned sbsx_present:1;
182
183 int num_nodes;
184 struct arm_ccn_component *node;
185
186 int num_xps;
187 struct arm_ccn_component *xp;
188
189 struct arm_ccn_dt dt;
190 int mn_id;
191};
192
193static int arm_ccn_node_to_xp(int node)
194{
195 return node / CCN_NUM_XP_PORTS;
196}
197
198static int arm_ccn_node_to_xp_port(int node)
199{
200 return node % CCN_NUM_XP_PORTS;
201}
202
203
204/*
205 * Bit shifts and masks in these defines must be kept in sync with
206 * arm_ccn_pmu_config_set() and CCN_FORMAT_ATTRs below!
207 */
208#define CCN_CONFIG_NODE(_config) (((_config) >> 0) & 0xff)
209#define CCN_CONFIG_XP(_config) (((_config) >> 0) & 0xff)
210#define CCN_CONFIG_TYPE(_config) (((_config) >> 8) & 0xff)
211#define CCN_CONFIG_EVENT(_config) (((_config) >> 16) & 0xff)
212#define CCN_CONFIG_PORT(_config) (((_config) >> 24) & 0x3)
213#define CCN_CONFIG_BUS(_config) (((_config) >> 24) & 0x3)
214#define CCN_CONFIG_VC(_config) (((_config) >> 26) & 0x7)
215#define CCN_CONFIG_DIR(_config) (((_config) >> 29) & 0x1)
216#define CCN_CONFIG_MASK(_config) (((_config) >> 30) & 0xf)
217
218static void arm_ccn_pmu_config_set(u64 *config, u32 node_xp, u32 type, u32 port)
219{
220 *config &= ~((0xff << 0) | (0xff << 8) | (0x3 << 24));
221 *config |= (node_xp << 0) | (type << 8) | (port << 24);
222}
223
224static ssize_t arm_ccn_pmu_format_show(struct device *dev,
225 struct device_attribute *attr, char *buf)
226{
227 struct dev_ext_attribute *ea = container_of(attr,
228 struct dev_ext_attribute, attr);
229
230 return snprintf(buf, PAGE_SIZE, "%s\n", (char *)ea->var);
231}
232
233#define CCN_FORMAT_ATTR(_name, _config) \
234 struct dev_ext_attribute arm_ccn_pmu_format_attr_##_name = \
235 { __ATTR(_name, S_IRUGO, arm_ccn_pmu_format_show, \
236 NULL), _config }
237
238static CCN_FORMAT_ATTR(node, "config:0-7");
239static CCN_FORMAT_ATTR(xp, "config:0-7");
240static CCN_FORMAT_ATTR(type, "config:8-15");
241static CCN_FORMAT_ATTR(event, "config:16-23");
242static CCN_FORMAT_ATTR(port, "config:24-25");
243static CCN_FORMAT_ATTR(bus, "config:24-25");
244static CCN_FORMAT_ATTR(vc, "config:26-28");
245static CCN_FORMAT_ATTR(dir, "config:29-29");
246static CCN_FORMAT_ATTR(mask, "config:30-33");
247static CCN_FORMAT_ATTR(cmp_l, "config1:0-62");
248static CCN_FORMAT_ATTR(cmp_h, "config2:0-59");
249
250static struct attribute *arm_ccn_pmu_format_attrs[] = {
251 &arm_ccn_pmu_format_attr_node.attr.attr,
252 &arm_ccn_pmu_format_attr_xp.attr.attr,
253 &arm_ccn_pmu_format_attr_type.attr.attr,
254 &arm_ccn_pmu_format_attr_event.attr.attr,
255 &arm_ccn_pmu_format_attr_port.attr.attr,
256 &arm_ccn_pmu_format_attr_bus.attr.attr,
257 &arm_ccn_pmu_format_attr_vc.attr.attr,
258 &arm_ccn_pmu_format_attr_dir.attr.attr,
259 &arm_ccn_pmu_format_attr_mask.attr.attr,
260 &arm_ccn_pmu_format_attr_cmp_l.attr.attr,
261 &arm_ccn_pmu_format_attr_cmp_h.attr.attr,
262 NULL
263};
264
265static const struct attribute_group arm_ccn_pmu_format_attr_group = {
266 .name = "format",
267 .attrs = arm_ccn_pmu_format_attrs,
268};
269
270
271struct arm_ccn_pmu_event {
272 struct device_attribute attr;
273 u32 type;
274 u32 event;
275 int num_ports;
276 int num_vcs;
277 const char *def;
278 int mask;
279};
280
281#define CCN_EVENT_ATTR(_name) \
282 __ATTR(_name, S_IRUGO, arm_ccn_pmu_event_show, NULL)
283
284/*
285 * Events defined in TRM for MN, HN-I and SBSX are actually watchpoints set on
286 * their ports in XP they are connected to. For the sake of usability they are
287 * explicitly defined here (and translated into a relevant watchpoint in
288 * arm_ccn_pmu_event_init()) so the user can easily request them without deep
289 * knowledge of the flit format.
290 */
291
292#define CCN_EVENT_MN(_name, _def, _mask) { .attr = CCN_EVENT_ATTR(mn_##_name), \
293 .type = CCN_TYPE_MN, .event = CCN_EVENT_WATCHPOINT, \
294 .num_ports = CCN_NUM_XP_PORTS, .num_vcs = CCN_NUM_VCS, \
295 .def = _def, .mask = _mask, }
296
297#define CCN_EVENT_HNI(_name, _def, _mask) { \
298 .attr = CCN_EVENT_ATTR(hni_##_name), .type = CCN_TYPE_HNI, \
299 .event = CCN_EVENT_WATCHPOINT, .num_ports = CCN_NUM_XP_PORTS, \
300 .num_vcs = CCN_NUM_VCS, .def = _def, .mask = _mask, }
301
302#define CCN_EVENT_SBSX(_name, _def, _mask) { \
303 .attr = CCN_EVENT_ATTR(sbsx_##_name), .type = CCN_TYPE_SBSX, \
304 .event = CCN_EVENT_WATCHPOINT, .num_ports = CCN_NUM_XP_PORTS, \
305 .num_vcs = CCN_NUM_VCS, .def = _def, .mask = _mask, }
306
307#define CCN_EVENT_HNF(_name, _event) { .attr = CCN_EVENT_ATTR(hnf_##_name), \
308 .type = CCN_TYPE_HNF, .event = _event, }
309
310#define CCN_EVENT_XP(_name, _event) { .attr = CCN_EVENT_ATTR(xp_##_name), \
311 .type = CCN_TYPE_XP, .event = _event, \
312 .num_ports = CCN_NUM_XP_PORTS, .num_vcs = CCN_NUM_VCS, }
313
314/*
315 * RN-I & RN-D (RN-D = RN-I + DVM) nodes have different type ID depending
316 * on configuration. One of them is picked to represent the whole group,
317 * as they all share the same event types.
318 */
319#define CCN_EVENT_RNI(_name, _event) { .attr = CCN_EVENT_ATTR(rni_##_name), \
320 .type = CCN_TYPE_RNI_3P, .event = _event, }
321
322#define CCN_EVENT_SBAS(_name, _event) { .attr = CCN_EVENT_ATTR(sbas_##_name), \
323 .type = CCN_TYPE_SBAS, .event = _event, }
324
325#define CCN_EVENT_CYCLES(_name) { .attr = CCN_EVENT_ATTR(_name), \
326 .type = CCN_TYPE_CYCLES }
327
328
329static ssize_t arm_ccn_pmu_event_show(struct device *dev,
330 struct device_attribute *attr, char *buf)
331{
332 struct arm_ccn *ccn = pmu_to_arm_ccn(dev_get_drvdata(dev));
333 struct arm_ccn_pmu_event *event = container_of(attr,
334 struct arm_ccn_pmu_event, attr);
335 ssize_t res;
336
337 res = snprintf(buf, PAGE_SIZE, "type=0x%x", event->type);
338 if (event->event)
339 res += snprintf(buf + res, PAGE_SIZE - res, ",event=0x%x",
340 event->event);
341 if (event->def)
342 res += snprintf(buf + res, PAGE_SIZE - res, ",%s",
343 event->def);
344 if (event->mask)
345 res += snprintf(buf + res, PAGE_SIZE - res, ",mask=0x%x",
346 event->mask);
347
348 /* Arguments required by an event */
349 switch (event->type) {
350 case CCN_TYPE_CYCLES:
351 break;
352 case CCN_TYPE_XP:
353 res += snprintf(buf + res, PAGE_SIZE - res,
354 ",xp=?,vc=?");
355 if (event->event == CCN_EVENT_WATCHPOINT)
356 res += snprintf(buf + res, PAGE_SIZE - res,
357 ",port=?,dir=?,cmp_l=?,cmp_h=?,mask=?");
358 else
359 res += snprintf(buf + res, PAGE_SIZE - res,
360 ",bus=?");
361
362 break;
363 case CCN_TYPE_MN:
364 res += snprintf(buf + res, PAGE_SIZE - res, ",node=%d", ccn->mn_id);
365 break;
366 default:
367 res += snprintf(buf + res, PAGE_SIZE - res, ",node=?");
368 break;
369 }
370
371 res += snprintf(buf + res, PAGE_SIZE - res, "\n");
372
373 return res;
374}
375
376static umode_t arm_ccn_pmu_events_is_visible(struct kobject *kobj,
377 struct attribute *attr, int index)
378{
379 struct device *dev = kobj_to_dev(kobj);
380 struct arm_ccn *ccn = pmu_to_arm_ccn(dev_get_drvdata(dev));
381 struct device_attribute *dev_attr = container_of(attr,
382 struct device_attribute, attr);
383 struct arm_ccn_pmu_event *event = container_of(dev_attr,
384 struct arm_ccn_pmu_event, attr);
385
386 if (event->type == CCN_TYPE_SBAS && !ccn->sbas_present)
387 return 0;
388 if (event->type == CCN_TYPE_SBSX && !ccn->sbsx_present)
389 return 0;
390
391 return attr->mode;
392}
393
394static struct arm_ccn_pmu_event arm_ccn_pmu_events[] = {
395 CCN_EVENT_MN(eobarrier, "dir=1,vc=0,cmp_h=0x1c00", CCN_IDX_MASK_OPCODE),
396 CCN_EVENT_MN(ecbarrier, "dir=1,vc=0,cmp_h=0x1e00", CCN_IDX_MASK_OPCODE),
397 CCN_EVENT_MN(dvmop, "dir=1,vc=0,cmp_h=0x2800", CCN_IDX_MASK_OPCODE),
398 CCN_EVENT_HNI(txdatflits, "dir=1,vc=3", CCN_IDX_MASK_ANY),
399 CCN_EVENT_HNI(rxdatflits, "dir=0,vc=3", CCN_IDX_MASK_ANY),
400 CCN_EVENT_HNI(txreqflits, "dir=1,vc=0", CCN_IDX_MASK_ANY),
401 CCN_EVENT_HNI(rxreqflits, "dir=0,vc=0", CCN_IDX_MASK_ANY),
402 CCN_EVENT_HNI(rxreqflits_order, "dir=0,vc=0,cmp_h=0x8000",
403 CCN_IDX_MASK_ORDER),
404 CCN_EVENT_SBSX(txdatflits, "dir=1,vc=3", CCN_IDX_MASK_ANY),
405 CCN_EVENT_SBSX(rxdatflits, "dir=0,vc=3", CCN_IDX_MASK_ANY),
406 CCN_EVENT_SBSX(txreqflits, "dir=1,vc=0", CCN_IDX_MASK_ANY),
407 CCN_EVENT_SBSX(rxreqflits, "dir=0,vc=0", CCN_IDX_MASK_ANY),
408 CCN_EVENT_SBSX(rxreqflits_order, "dir=0,vc=0,cmp_h=0x8000",
409 CCN_IDX_MASK_ORDER),
410 CCN_EVENT_HNF(cache_miss, 0x1),
411 CCN_EVENT_HNF(l3_sf_cache_access, 0x02),
412 CCN_EVENT_HNF(cache_fill, 0x3),
413 CCN_EVENT_HNF(pocq_retry, 0x4),
414 CCN_EVENT_HNF(pocq_reqs_recvd, 0x5),
415 CCN_EVENT_HNF(sf_hit, 0x6),
416 CCN_EVENT_HNF(sf_evictions, 0x7),
417 CCN_EVENT_HNF(snoops_sent, 0x8),
418 CCN_EVENT_HNF(snoops_broadcast, 0x9),
419 CCN_EVENT_HNF(l3_eviction, 0xa),
420 CCN_EVENT_HNF(l3_fill_invalid_way, 0xb),
421 CCN_EVENT_HNF(mc_retries, 0xc),
422 CCN_EVENT_HNF(mc_reqs, 0xd),
423 CCN_EVENT_HNF(qos_hh_retry, 0xe),
424 CCN_EVENT_RNI(rdata_beats_p0, 0x1),
425 CCN_EVENT_RNI(rdata_beats_p1, 0x2),
426 CCN_EVENT_RNI(rdata_beats_p2, 0x3),
427 CCN_EVENT_RNI(rxdat_flits, 0x4),
428 CCN_EVENT_RNI(txdat_flits, 0x5),
429 CCN_EVENT_RNI(txreq_flits, 0x6),
430 CCN_EVENT_RNI(txreq_flits_retried, 0x7),
431 CCN_EVENT_RNI(rrt_full, 0x8),
432 CCN_EVENT_RNI(wrt_full, 0x9),
433 CCN_EVENT_RNI(txreq_flits_replayed, 0xa),
434 CCN_EVENT_XP(upload_starvation, 0x1),
435 CCN_EVENT_XP(download_starvation, 0x2),
436 CCN_EVENT_XP(respin, 0x3),
437 CCN_EVENT_XP(valid_flit, 0x4),
438 CCN_EVENT_XP(watchpoint, CCN_EVENT_WATCHPOINT),
439 CCN_EVENT_SBAS(rdata_beats_p0, 0x1),
440 CCN_EVENT_SBAS(rxdat_flits, 0x4),
441 CCN_EVENT_SBAS(txdat_flits, 0x5),
442 CCN_EVENT_SBAS(txreq_flits, 0x6),
443 CCN_EVENT_SBAS(txreq_flits_retried, 0x7),
444 CCN_EVENT_SBAS(rrt_full, 0x8),
445 CCN_EVENT_SBAS(wrt_full, 0x9),
446 CCN_EVENT_SBAS(txreq_flits_replayed, 0xa),
447 CCN_EVENT_CYCLES(cycles),
448};
449
450/* Populated in arm_ccn_init() */
451static struct attribute
452 *arm_ccn_pmu_events_attrs[ARRAY_SIZE(arm_ccn_pmu_events) + 1];
453
454static const struct attribute_group arm_ccn_pmu_events_attr_group = {
455 .name = "events",
456 .is_visible = arm_ccn_pmu_events_is_visible,
457 .attrs = arm_ccn_pmu_events_attrs,
458};
459
460
461static u64 *arm_ccn_pmu_get_cmp_mask(struct arm_ccn *ccn, const char *name)
462{
463 unsigned long i;
464
465 if (WARN_ON(!name || !name[0] || !isxdigit(name[0]) || !name[1]))
466 return NULL;
467 i = isdigit(name[0]) ? name[0] - '0' : 0xa + tolower(name[0]) - 'a';
468
469 switch (name[1]) {
470 case 'l':
471 return &ccn->dt.cmp_mask[i].l;
472 case 'h':
473 return &ccn->dt.cmp_mask[i].h;
474 default:
475 return NULL;
476 }
477}
478
479static ssize_t arm_ccn_pmu_cmp_mask_show(struct device *dev,
480 struct device_attribute *attr, char *buf)
481{
482 struct arm_ccn *ccn = pmu_to_arm_ccn(dev_get_drvdata(dev));
483 u64 *mask = arm_ccn_pmu_get_cmp_mask(ccn, attr->attr.name);
484
485 return mask ? snprintf(buf, PAGE_SIZE, "0x%016llx\n", *mask) : -EINVAL;
486}
487
488static ssize_t arm_ccn_pmu_cmp_mask_store(struct device *dev,
489 struct device_attribute *attr, const char *buf, size_t count)
490{
491 struct arm_ccn *ccn = pmu_to_arm_ccn(dev_get_drvdata(dev));
492 u64 *mask = arm_ccn_pmu_get_cmp_mask(ccn, attr->attr.name);
493 int err = -EINVAL;
494
495 if (mask)
496 err = kstrtoull(buf, 0, mask);
497
498 return err ? err : count;
499}
500
501#define CCN_CMP_MASK_ATTR(_name) \
502 struct device_attribute arm_ccn_pmu_cmp_mask_attr_##_name = \
503 __ATTR(_name, S_IRUGO | S_IWUSR, \
504 arm_ccn_pmu_cmp_mask_show, arm_ccn_pmu_cmp_mask_store)
505
506#define CCN_CMP_MASK_ATTR_RO(_name) \
507 struct device_attribute arm_ccn_pmu_cmp_mask_attr_##_name = \
508 __ATTR(_name, S_IRUGO, arm_ccn_pmu_cmp_mask_show, NULL)
509
510static CCN_CMP_MASK_ATTR(0l);
511static CCN_CMP_MASK_ATTR(0h);
512static CCN_CMP_MASK_ATTR(1l);
513static CCN_CMP_MASK_ATTR(1h);
514static CCN_CMP_MASK_ATTR(2l);
515static CCN_CMP_MASK_ATTR(2h);
516static CCN_CMP_MASK_ATTR(3l);
517static CCN_CMP_MASK_ATTR(3h);
518static CCN_CMP_MASK_ATTR(4l);
519static CCN_CMP_MASK_ATTR(4h);
520static CCN_CMP_MASK_ATTR(5l);
521static CCN_CMP_MASK_ATTR(5h);
522static CCN_CMP_MASK_ATTR(6l);
523static CCN_CMP_MASK_ATTR(6h);
524static CCN_CMP_MASK_ATTR(7l);
525static CCN_CMP_MASK_ATTR(7h);
526static CCN_CMP_MASK_ATTR_RO(8l);
527static CCN_CMP_MASK_ATTR_RO(8h);
528static CCN_CMP_MASK_ATTR_RO(9l);
529static CCN_CMP_MASK_ATTR_RO(9h);
530static CCN_CMP_MASK_ATTR_RO(al);
531static CCN_CMP_MASK_ATTR_RO(ah);
532static CCN_CMP_MASK_ATTR_RO(bl);
533static CCN_CMP_MASK_ATTR_RO(bh);
534
535static struct attribute *arm_ccn_pmu_cmp_mask_attrs[] = {
536 &arm_ccn_pmu_cmp_mask_attr_0l.attr, &arm_ccn_pmu_cmp_mask_attr_0h.attr,
537 &arm_ccn_pmu_cmp_mask_attr_1l.attr, &arm_ccn_pmu_cmp_mask_attr_1h.attr,
538 &arm_ccn_pmu_cmp_mask_attr_2l.attr, &arm_ccn_pmu_cmp_mask_attr_2h.attr,
539 &arm_ccn_pmu_cmp_mask_attr_3l.attr, &arm_ccn_pmu_cmp_mask_attr_3h.attr,
540 &arm_ccn_pmu_cmp_mask_attr_4l.attr, &arm_ccn_pmu_cmp_mask_attr_4h.attr,
541 &arm_ccn_pmu_cmp_mask_attr_5l.attr, &arm_ccn_pmu_cmp_mask_attr_5h.attr,
542 &arm_ccn_pmu_cmp_mask_attr_6l.attr, &arm_ccn_pmu_cmp_mask_attr_6h.attr,
543 &arm_ccn_pmu_cmp_mask_attr_7l.attr, &arm_ccn_pmu_cmp_mask_attr_7h.attr,
544 &arm_ccn_pmu_cmp_mask_attr_8l.attr, &arm_ccn_pmu_cmp_mask_attr_8h.attr,
545 &arm_ccn_pmu_cmp_mask_attr_9l.attr, &arm_ccn_pmu_cmp_mask_attr_9h.attr,
546 &arm_ccn_pmu_cmp_mask_attr_al.attr, &arm_ccn_pmu_cmp_mask_attr_ah.attr,
547 &arm_ccn_pmu_cmp_mask_attr_bl.attr, &arm_ccn_pmu_cmp_mask_attr_bh.attr,
548 NULL
549};
550
551static const struct attribute_group arm_ccn_pmu_cmp_mask_attr_group = {
552 .name = "cmp_mask",
553 .attrs = arm_ccn_pmu_cmp_mask_attrs,
554};
555
556static ssize_t arm_ccn_pmu_cpumask_show(struct device *dev,
557 struct device_attribute *attr, char *buf)
558{
559 struct arm_ccn *ccn = pmu_to_arm_ccn(dev_get_drvdata(dev));
560
561 return cpumap_print_to_pagebuf(true, buf, &ccn->dt.cpu);
562}
563
564static struct device_attribute arm_ccn_pmu_cpumask_attr =
565 __ATTR(cpumask, S_IRUGO, arm_ccn_pmu_cpumask_show, NULL);
566
567static struct attribute *arm_ccn_pmu_cpumask_attrs[] = {
568 &arm_ccn_pmu_cpumask_attr.attr,
569 NULL,
570};
571
572static const struct attribute_group arm_ccn_pmu_cpumask_attr_group = {
573 .attrs = arm_ccn_pmu_cpumask_attrs,
574};
575
576/*
577 * Default poll period is 10ms, which is way over the top anyway,
578 * as in the worst case scenario (an event every cycle), with 1GHz
579 * clocked bus, the smallest, 32 bit counter will overflow in
580 * more than 4s.
581 */
582static unsigned int arm_ccn_pmu_poll_period_us = 10000;
583module_param_named(pmu_poll_period_us, arm_ccn_pmu_poll_period_us, uint,
584 S_IRUGO | S_IWUSR);
585
586static ktime_t arm_ccn_pmu_timer_period(void)
587{
588 return ns_to_ktime((u64)arm_ccn_pmu_poll_period_us * 1000);
589}
590
591
592static const struct attribute_group *arm_ccn_pmu_attr_groups[] = {
593 &arm_ccn_pmu_events_attr_group,
594 &arm_ccn_pmu_format_attr_group,
595 &arm_ccn_pmu_cmp_mask_attr_group,
596 &arm_ccn_pmu_cpumask_attr_group,
597 NULL
598};
599
600
601static int arm_ccn_pmu_alloc_bit(unsigned long *bitmap, unsigned long size)
602{
603 int bit;
604
605 do {
606 bit = find_first_zero_bit(bitmap, size);
607 if (bit >= size)
608 return -EAGAIN;
609 } while (test_and_set_bit(bit, bitmap));
610
611 return bit;
612}
613
614/* All RN-I and RN-D nodes have identical PMUs */
615static int arm_ccn_pmu_type_eq(u32 a, u32 b)
616{
617 if (a == b)
618 return 1;
619
620 switch (a) {
621 case CCN_TYPE_RNI_1P:
622 case CCN_TYPE_RNI_2P:
623 case CCN_TYPE_RNI_3P:
624 case CCN_TYPE_RND_1P:
625 case CCN_TYPE_RND_2P:
626 case CCN_TYPE_RND_3P:
627 switch (b) {
628 case CCN_TYPE_RNI_1P:
629 case CCN_TYPE_RNI_2P:
630 case CCN_TYPE_RNI_3P:
631 case CCN_TYPE_RND_1P:
632 case CCN_TYPE_RND_2P:
633 case CCN_TYPE_RND_3P:
634 return 1;
635 }
636 break;
637 }
638
639 return 0;
640}
641
642static int arm_ccn_pmu_event_alloc(struct perf_event *event)
643{
644 struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
645 struct hw_perf_event *hw = &event->hw;
646 u32 node_xp, type, event_id;
647 struct arm_ccn_component *source;
648 int bit;
649
650 node_xp = CCN_CONFIG_NODE(event->attr.config);
651 type = CCN_CONFIG_TYPE(event->attr.config);
652 event_id = CCN_CONFIG_EVENT(event->attr.config);
653
654 /* Allocate the cycle counter */
655 if (type == CCN_TYPE_CYCLES) {
656 if (test_and_set_bit(CCN_IDX_PMU_CYCLE_COUNTER,
657 ccn->dt.pmu_counters_mask))
658 return -EAGAIN;
659
660 hw->idx = CCN_IDX_PMU_CYCLE_COUNTER;
661 ccn->dt.pmu_counters[CCN_IDX_PMU_CYCLE_COUNTER].event = event;
662
663 return 0;
664 }
665
666 /* Allocate an event counter */
667 hw->idx = arm_ccn_pmu_alloc_bit(ccn->dt.pmu_counters_mask,
668 CCN_NUM_PMU_EVENT_COUNTERS);
669 if (hw->idx < 0) {
670 dev_dbg(ccn->dev, "No more counters available!\n");
671 return -EAGAIN;
672 }
673
674 if (type == CCN_TYPE_XP)
675 source = &ccn->xp[node_xp];
676 else
677 source = &ccn->node[node_xp];
678 ccn->dt.pmu_counters[hw->idx].source = source;
679
680 /* Allocate an event source or a watchpoint */
681 if (type == CCN_TYPE_XP && event_id == CCN_EVENT_WATCHPOINT)
682 bit = arm_ccn_pmu_alloc_bit(source->xp.dt_cmp_mask,
683 CCN_NUM_XP_WATCHPOINTS);
684 else
685 bit = arm_ccn_pmu_alloc_bit(source->pmu_events_mask,
686 CCN_NUM_PMU_EVENTS);
687 if (bit < 0) {
688 dev_dbg(ccn->dev, "No more event sources/watchpoints on node/XP %d!\n",
689 node_xp);
690 clear_bit(hw->idx, ccn->dt.pmu_counters_mask);
691 return -EAGAIN;
692 }
693 hw->config_base = bit;
694
695 ccn->dt.pmu_counters[hw->idx].event = event;
696
697 return 0;
698}
699
700static void arm_ccn_pmu_event_release(struct perf_event *event)
701{
702 struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
703 struct hw_perf_event *hw = &event->hw;
704
705 if (hw->idx == CCN_IDX_PMU_CYCLE_COUNTER) {
706 clear_bit(CCN_IDX_PMU_CYCLE_COUNTER, ccn->dt.pmu_counters_mask);
707 } else {
708 struct arm_ccn_component *source =
709 ccn->dt.pmu_counters[hw->idx].source;
710
711 if (CCN_CONFIG_TYPE(event->attr.config) == CCN_TYPE_XP &&
712 CCN_CONFIG_EVENT(event->attr.config) ==
713 CCN_EVENT_WATCHPOINT)
714 clear_bit(hw->config_base, source->xp.dt_cmp_mask);
715 else
716 clear_bit(hw->config_base, source->pmu_events_mask);
717 clear_bit(hw->idx, ccn->dt.pmu_counters_mask);
718 }
719
720 ccn->dt.pmu_counters[hw->idx].source = NULL;
721 ccn->dt.pmu_counters[hw->idx].event = NULL;
722}
723
724static int arm_ccn_pmu_event_init(struct perf_event *event)
725{
726 struct arm_ccn *ccn;
727 struct hw_perf_event *hw = &event->hw;
728 u32 node_xp, type, event_id;
729 int valid;
730 int i;
731 struct perf_event *sibling;
732
733 if (event->attr.type != event->pmu->type)
734 return -ENOENT;
735
736 ccn = pmu_to_arm_ccn(event->pmu);
737
738 if (hw->sample_period) {
739 dev_warn(ccn->dev, "Sampling not supported!\n");
740 return -EOPNOTSUPP;
741 }
742
743 if (has_branch_stack(event) || event->attr.exclude_user ||
744 event->attr.exclude_kernel || event->attr.exclude_hv ||
745 event->attr.exclude_idle || event->attr.exclude_host ||
746 event->attr.exclude_guest) {
747 dev_warn(ccn->dev, "Can't exclude execution levels!\n");
748 return -EINVAL;
749 }
750
751 if (event->cpu < 0) {
752 dev_warn(ccn->dev, "Can't provide per-task data!\n");
753 return -EOPNOTSUPP;
754 }
755 /*
756 * Many perf core operations (eg. events rotation) operate on a
757 * single CPU context. This is obvious for CPU PMUs, where one
758 * expects the same sets of events being observed on all CPUs,
759 * but can lead to issues for off-core PMUs, like CCN, where each
760 * event could be theoretically assigned to a different CPU. To
761 * mitigate this, we enforce CPU assignment to one, selected
762 * processor (the one described in the "cpumask" attribute).
763 */
764 event->cpu = cpumask_first(&ccn->dt.cpu);
765
766 node_xp = CCN_CONFIG_NODE(event->attr.config);
767 type = CCN_CONFIG_TYPE(event->attr.config);
768 event_id = CCN_CONFIG_EVENT(event->attr.config);
769
770 /* Validate node/xp vs topology */
771 switch (type) {
772 case CCN_TYPE_MN:
773 if (node_xp != ccn->mn_id) {
774 dev_warn(ccn->dev, "Invalid MN ID %d!\n", node_xp);
775 return -EINVAL;
776 }
777 break;
778 case CCN_TYPE_XP:
779 if (node_xp >= ccn->num_xps) {
780 dev_warn(ccn->dev, "Invalid XP ID %d!\n", node_xp);
781 return -EINVAL;
782 }
783 break;
784 case CCN_TYPE_CYCLES:
785 break;
786 default:
787 if (node_xp >= ccn->num_nodes) {
788 dev_warn(ccn->dev, "Invalid node ID %d!\n", node_xp);
789 return -EINVAL;
790 }
791 if (!arm_ccn_pmu_type_eq(type, ccn->node[node_xp].type)) {
792 dev_warn(ccn->dev, "Invalid type 0x%x for node %d!\n",
793 type, node_xp);
794 return -EINVAL;
795 }
796 break;
797 }
798
799 /* Validate event ID vs available for the type */
800 for (i = 0, valid = 0; i < ARRAY_SIZE(arm_ccn_pmu_events) && !valid;
801 i++) {
802 struct arm_ccn_pmu_event *e = &arm_ccn_pmu_events[i];
803 u32 port = CCN_CONFIG_PORT(event->attr.config);
804 u32 vc = CCN_CONFIG_VC(event->attr.config);
805
806 if (!arm_ccn_pmu_type_eq(type, e->type))
807 continue;
808 if (event_id != e->event)
809 continue;
810 if (e->num_ports && port >= e->num_ports) {
811 dev_warn(ccn->dev, "Invalid port %d for node/XP %d!\n",
812 port, node_xp);
813 return -EINVAL;
814 }
815 if (e->num_vcs && vc >= e->num_vcs) {
816 dev_warn(ccn->dev, "Invalid vc %d for node/XP %d!\n",
817 vc, node_xp);
818 return -EINVAL;
819 }
820 valid = 1;
821 }
822 if (!valid) {
823 dev_warn(ccn->dev, "Invalid event 0x%x for node/XP %d!\n",
824 event_id, node_xp);
825 return -EINVAL;
826 }
827
828 /* Watchpoint-based event for a node is actually set on XP */
829 if (event_id == CCN_EVENT_WATCHPOINT && type != CCN_TYPE_XP) {
830 u32 port;
831
832 type = CCN_TYPE_XP;
833 port = arm_ccn_node_to_xp_port(node_xp);
834 node_xp = arm_ccn_node_to_xp(node_xp);
835
836 arm_ccn_pmu_config_set(&event->attr.config,
837 node_xp, type, port);
838 }
839
840 /*
841 * We must NOT create groups containing mixed PMUs, although software
842 * events are acceptable (for example to create a CCN group
843 * periodically read when a hrtimer aka cpu-clock leader triggers).
844 */
845 if (event->group_leader->pmu != event->pmu &&
846 !is_software_event(event->group_leader))
847 return -EINVAL;
848
849 for_each_sibling_event(sibling, event->group_leader) {
850 if (sibling->pmu != event->pmu &&
851 !is_software_event(sibling))
852 return -EINVAL;
853 }
854
855 return 0;
856}
857
858static u64 arm_ccn_pmu_read_counter(struct arm_ccn *ccn, int idx)
859{
860 u64 res;
861
862 if (idx == CCN_IDX_PMU_CYCLE_COUNTER) {
863#ifdef readq
864 res = readq(ccn->dt.base + CCN_DT_PMCCNTR);
865#else
866 /* 40 bit counter, can do snapshot and read in two parts */
867 writel(0x1, ccn->dt.base + CCN_DT_PMSR_REQ);
868 while (!(readl(ccn->dt.base + CCN_DT_PMSR) & 0x1))
869 ;
870 writel(0x1, ccn->dt.base + CCN_DT_PMSR_CLR);
871 res = readl(ccn->dt.base + CCN_DT_PMCCNTRSR + 4) & 0xff;
872 res <<= 32;
873 res |= readl(ccn->dt.base + CCN_DT_PMCCNTRSR);
874#endif
875 } else {
876 res = readl(ccn->dt.base + CCN_DT_PMEVCNT(idx));
877 }
878
879 return res;
880}
881
882static void arm_ccn_pmu_event_update(struct perf_event *event)
883{
884 struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
885 struct hw_perf_event *hw = &event->hw;
886 u64 prev_count, new_count, mask;
887
888 do {
889 prev_count = local64_read(&hw->prev_count);
890 new_count = arm_ccn_pmu_read_counter(ccn, hw->idx);
891 } while (local64_xchg(&hw->prev_count, new_count) != prev_count);
892
893 mask = (1LLU << (hw->idx == CCN_IDX_PMU_CYCLE_COUNTER ? 40 : 32)) - 1;
894
895 local64_add((new_count - prev_count) & mask, &event->count);
896}
897
898static void arm_ccn_pmu_xp_dt_config(struct perf_event *event, int enable)
899{
900 struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
901 struct hw_perf_event *hw = &event->hw;
902 struct arm_ccn_component *xp;
903 u32 val, dt_cfg;
904
905 /* Nothing to do for cycle counter */
906 if (hw->idx == CCN_IDX_PMU_CYCLE_COUNTER)
907 return;
908
909 if (CCN_CONFIG_TYPE(event->attr.config) == CCN_TYPE_XP)
910 xp = &ccn->xp[CCN_CONFIG_XP(event->attr.config)];
911 else
912 xp = &ccn->xp[arm_ccn_node_to_xp(
913 CCN_CONFIG_NODE(event->attr.config))];
914
915 if (enable)
916 dt_cfg = hw->event_base;
917 else
918 dt_cfg = CCN_XP_DT_CONFIG__DT_CFG__PASS_THROUGH;
919
920 spin_lock(&ccn->dt.config_lock);
921
922 val = readl(xp->base + CCN_XP_DT_CONFIG);
923 val &= ~(CCN_XP_DT_CONFIG__DT_CFG__MASK <<
924 CCN_XP_DT_CONFIG__DT_CFG__SHIFT(hw->idx));
925 val |= dt_cfg << CCN_XP_DT_CONFIG__DT_CFG__SHIFT(hw->idx);
926 writel(val, xp->base + CCN_XP_DT_CONFIG);
927
928 spin_unlock(&ccn->dt.config_lock);
929}
930
931static void arm_ccn_pmu_event_start(struct perf_event *event, int flags)
932{
933 struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
934 struct hw_perf_event *hw = &event->hw;
935
936 local64_set(&event->hw.prev_count,
937 arm_ccn_pmu_read_counter(ccn, hw->idx));
938 hw->state = 0;
939
940 /* Set the DT bus input, engaging the counter */
941 arm_ccn_pmu_xp_dt_config(event, 1);
942}
943
944static void arm_ccn_pmu_event_stop(struct perf_event *event, int flags)
945{
946 struct hw_perf_event *hw = &event->hw;
947
948 /* Disable counting, setting the DT bus to pass-through mode */
949 arm_ccn_pmu_xp_dt_config(event, 0);
950
951 if (flags & PERF_EF_UPDATE)
952 arm_ccn_pmu_event_update(event);
953
954 hw->state |= PERF_HES_STOPPED;
955}
956
957static void arm_ccn_pmu_xp_watchpoint_config(struct perf_event *event)
958{
959 struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
960 struct hw_perf_event *hw = &event->hw;
961 struct arm_ccn_component *source =
962 ccn->dt.pmu_counters[hw->idx].source;
963 unsigned long wp = hw->config_base;
964 u32 val;
965 u64 cmp_l = event->attr.config1;
966 u64 cmp_h = event->attr.config2;
967 u64 mask_l = ccn->dt.cmp_mask[CCN_CONFIG_MASK(event->attr.config)].l;
968 u64 mask_h = ccn->dt.cmp_mask[CCN_CONFIG_MASK(event->attr.config)].h;
969
970 hw->event_base = CCN_XP_DT_CONFIG__DT_CFG__WATCHPOINT(wp);
971
972 /* Direction (RX/TX), device (port) & virtual channel */
973 val = readl(source->base + CCN_XP_DT_INTERFACE_SEL);
974 val &= ~(CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__MASK <<
975 CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__SHIFT(wp));
976 val |= CCN_CONFIG_DIR(event->attr.config) <<
977 CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__SHIFT(wp);
978 val &= ~(CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__MASK <<
979 CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__SHIFT(wp));
980 val |= CCN_CONFIG_PORT(event->attr.config) <<
981 CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__SHIFT(wp);
982 val &= ~(CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__MASK <<
983 CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__SHIFT(wp));
984 val |= CCN_CONFIG_VC(event->attr.config) <<
985 CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__SHIFT(wp);
986 writel(val, source->base + CCN_XP_DT_INTERFACE_SEL);
987
988 /* Comparison values */
989 writel(cmp_l & 0xffffffff, source->base + CCN_XP_DT_CMP_VAL_L(wp));
990 writel((cmp_l >> 32) & 0x7fffffff,
991 source->base + CCN_XP_DT_CMP_VAL_L(wp) + 4);
992 writel(cmp_h & 0xffffffff, source->base + CCN_XP_DT_CMP_VAL_H(wp));
993 writel((cmp_h >> 32) & 0x0fffffff,
994 source->base + CCN_XP_DT_CMP_VAL_H(wp) + 4);
995
996 /* Mask */
997 writel(mask_l & 0xffffffff, source->base + CCN_XP_DT_CMP_MASK_L(wp));
998 writel((mask_l >> 32) & 0x7fffffff,
999 source->base + CCN_XP_DT_CMP_MASK_L(wp) + 4);
1000 writel(mask_h & 0xffffffff, source->base + CCN_XP_DT_CMP_MASK_H(wp));
1001 writel((mask_h >> 32) & 0x0fffffff,
1002 source->base + CCN_XP_DT_CMP_MASK_H(wp) + 4);
1003}
1004
1005static void arm_ccn_pmu_xp_event_config(struct perf_event *event)
1006{
1007 struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
1008 struct hw_perf_event *hw = &event->hw;
1009 struct arm_ccn_component *source =
1010 ccn->dt.pmu_counters[hw->idx].source;
1011 u32 val, id;
1012
1013 hw->event_base = CCN_XP_DT_CONFIG__DT_CFG__XP_PMU_EVENT(hw->config_base);
1014
1015 id = (CCN_CONFIG_VC(event->attr.config) << 4) |
1016 (CCN_CONFIG_BUS(event->attr.config) << 3) |
1017 (CCN_CONFIG_EVENT(event->attr.config) << 0);
1018
1019 val = readl(source->base + CCN_XP_PMU_EVENT_SEL);
1020 val &= ~(CCN_XP_PMU_EVENT_SEL__ID__MASK <<
1021 CCN_XP_PMU_EVENT_SEL__ID__SHIFT(hw->config_base));
1022 val |= id << CCN_XP_PMU_EVENT_SEL__ID__SHIFT(hw->config_base);
1023 writel(val, source->base + CCN_XP_PMU_EVENT_SEL);
1024}
1025
1026static void arm_ccn_pmu_node_event_config(struct perf_event *event)
1027{
1028 struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
1029 struct hw_perf_event *hw = &event->hw;
1030 struct arm_ccn_component *source =
1031 ccn->dt.pmu_counters[hw->idx].source;
1032 u32 type = CCN_CONFIG_TYPE(event->attr.config);
1033 u32 val, port;
1034
1035 port = arm_ccn_node_to_xp_port(CCN_CONFIG_NODE(event->attr.config));
1036 hw->event_base = CCN_XP_DT_CONFIG__DT_CFG__DEVICE_PMU_EVENT(port,
1037 hw->config_base);
1038
1039 /* These *_event_sel regs should be identical, but let's make sure... */
1040 BUILD_BUG_ON(CCN_HNF_PMU_EVENT_SEL != CCN_SBAS_PMU_EVENT_SEL);
1041 BUILD_BUG_ON(CCN_SBAS_PMU_EVENT_SEL != CCN_RNI_PMU_EVENT_SEL);
1042 BUILD_BUG_ON(CCN_HNF_PMU_EVENT_SEL__ID__SHIFT(1) !=
1043 CCN_SBAS_PMU_EVENT_SEL__ID__SHIFT(1));
1044 BUILD_BUG_ON(CCN_SBAS_PMU_EVENT_SEL__ID__SHIFT(1) !=
1045 CCN_RNI_PMU_EVENT_SEL__ID__SHIFT(1));
1046 BUILD_BUG_ON(CCN_HNF_PMU_EVENT_SEL__ID__MASK !=
1047 CCN_SBAS_PMU_EVENT_SEL__ID__MASK);
1048 BUILD_BUG_ON(CCN_SBAS_PMU_EVENT_SEL__ID__MASK !=
1049 CCN_RNI_PMU_EVENT_SEL__ID__MASK);
1050 if (WARN_ON(type != CCN_TYPE_HNF && type != CCN_TYPE_SBAS &&
1051 !arm_ccn_pmu_type_eq(type, CCN_TYPE_RNI_3P)))
1052 return;
1053
1054 /* Set the event id for the pre-allocated counter */
1055 val = readl(source->base + CCN_HNF_PMU_EVENT_SEL);
1056 val &= ~(CCN_HNF_PMU_EVENT_SEL__ID__MASK <<
1057 CCN_HNF_PMU_EVENT_SEL__ID__SHIFT(hw->config_base));
1058 val |= CCN_CONFIG_EVENT(event->attr.config) <<
1059 CCN_HNF_PMU_EVENT_SEL__ID__SHIFT(hw->config_base);
1060 writel(val, source->base + CCN_HNF_PMU_EVENT_SEL);
1061}
1062
1063static void arm_ccn_pmu_event_config(struct perf_event *event)
1064{
1065 struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
1066 struct hw_perf_event *hw = &event->hw;
1067 u32 xp, offset, val;
1068
1069 /* Cycle counter requires no setup */
1070 if (hw->idx == CCN_IDX_PMU_CYCLE_COUNTER)
1071 return;
1072
1073 if (CCN_CONFIG_TYPE(event->attr.config) == CCN_TYPE_XP)
1074 xp = CCN_CONFIG_XP(event->attr.config);
1075 else
1076 xp = arm_ccn_node_to_xp(CCN_CONFIG_NODE(event->attr.config));
1077
1078 spin_lock(&ccn->dt.config_lock);
1079
1080 /* Set the DT bus "distance" register */
1081 offset = (hw->idx / 4) * 4;
1082 val = readl(ccn->dt.base + CCN_DT_ACTIVE_DSM + offset);
1083 val &= ~(CCN_DT_ACTIVE_DSM__DSM_ID__MASK <<
1084 CCN_DT_ACTIVE_DSM__DSM_ID__SHIFT(hw->idx % 4));
1085 val |= xp << CCN_DT_ACTIVE_DSM__DSM_ID__SHIFT(hw->idx % 4);
1086 writel(val, ccn->dt.base + CCN_DT_ACTIVE_DSM + offset);
1087
1088 if (CCN_CONFIG_TYPE(event->attr.config) == CCN_TYPE_XP) {
1089 if (CCN_CONFIG_EVENT(event->attr.config) ==
1090 CCN_EVENT_WATCHPOINT)
1091 arm_ccn_pmu_xp_watchpoint_config(event);
1092 else
1093 arm_ccn_pmu_xp_event_config(event);
1094 } else {
1095 arm_ccn_pmu_node_event_config(event);
1096 }
1097
1098 spin_unlock(&ccn->dt.config_lock);
1099}
1100
1101static int arm_ccn_pmu_active_counters(struct arm_ccn *ccn)
1102{
1103 return bitmap_weight(ccn->dt.pmu_counters_mask,
1104 CCN_NUM_PMU_EVENT_COUNTERS + 1);
1105}
1106
1107static int arm_ccn_pmu_event_add(struct perf_event *event, int flags)
1108{
1109 int err;
1110 struct hw_perf_event *hw = &event->hw;
1111 struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
1112
1113 err = arm_ccn_pmu_event_alloc(event);
1114 if (err)
1115 return err;
1116
1117 /*
1118 * Pin the timer, so that the overflows are handled by the chosen
1119 * event->cpu (this is the same one as presented in "cpumask"
1120 * attribute).
1121 */
1122 if (!ccn->irq && arm_ccn_pmu_active_counters(ccn) == 1)
1123 hrtimer_start(&ccn->dt.hrtimer, arm_ccn_pmu_timer_period(),
1124 HRTIMER_MODE_REL_PINNED);
1125
1126 arm_ccn_pmu_event_config(event);
1127
1128 hw->state = PERF_HES_STOPPED;
1129
1130 if (flags & PERF_EF_START)
1131 arm_ccn_pmu_event_start(event, PERF_EF_UPDATE);
1132
1133 return 0;
1134}
1135
1136static void arm_ccn_pmu_event_del(struct perf_event *event, int flags)
1137{
1138 struct arm_ccn *ccn = pmu_to_arm_ccn(event->pmu);
1139
1140 arm_ccn_pmu_event_stop(event, PERF_EF_UPDATE);
1141
1142 arm_ccn_pmu_event_release(event);
1143
1144 if (!ccn->irq && arm_ccn_pmu_active_counters(ccn) == 0)
1145 hrtimer_cancel(&ccn->dt.hrtimer);
1146}
1147
1148static void arm_ccn_pmu_event_read(struct perf_event *event)
1149{
1150 arm_ccn_pmu_event_update(event);
1151}
1152
1153static void arm_ccn_pmu_enable(struct pmu *pmu)
1154{
1155 struct arm_ccn *ccn = pmu_to_arm_ccn(pmu);
1156
1157 u32 val = readl(ccn->dt.base + CCN_DT_PMCR);
1158 val |= CCN_DT_PMCR__PMU_EN;
1159 writel(val, ccn->dt.base + CCN_DT_PMCR);
1160}
1161
1162static void arm_ccn_pmu_disable(struct pmu *pmu)
1163{
1164 struct arm_ccn *ccn = pmu_to_arm_ccn(pmu);
1165
1166 u32 val = readl(ccn->dt.base + CCN_DT_PMCR);
1167 val &= ~CCN_DT_PMCR__PMU_EN;
1168 writel(val, ccn->dt.base + CCN_DT_PMCR);
1169}
1170
1171static irqreturn_t arm_ccn_pmu_overflow_handler(struct arm_ccn_dt *dt)
1172{
1173 u32 pmovsr = readl(dt->base + CCN_DT_PMOVSR);
1174 int idx;
1175
1176 if (!pmovsr)
1177 return IRQ_NONE;
1178
1179 writel(pmovsr, dt->base + CCN_DT_PMOVSR_CLR);
1180
1181 BUILD_BUG_ON(CCN_IDX_PMU_CYCLE_COUNTER != CCN_NUM_PMU_EVENT_COUNTERS);
1182
1183 for (idx = 0; idx < CCN_NUM_PMU_EVENT_COUNTERS + 1; idx++) {
1184 struct perf_event *event = dt->pmu_counters[idx].event;
1185 int overflowed = pmovsr & BIT(idx);
1186
1187 WARN_ON_ONCE(overflowed && !event &&
1188 idx != CCN_IDX_PMU_CYCLE_COUNTER);
1189
1190 if (!event || !overflowed)
1191 continue;
1192
1193 arm_ccn_pmu_event_update(event);
1194 }
1195
1196 return IRQ_HANDLED;
1197}
1198
1199static enum hrtimer_restart arm_ccn_pmu_timer_handler(struct hrtimer *hrtimer)
1200{
1201 struct arm_ccn_dt *dt = container_of(hrtimer, struct arm_ccn_dt,
1202 hrtimer);
1203 unsigned long flags;
1204
1205 local_irq_save(flags);
1206 arm_ccn_pmu_overflow_handler(dt);
1207 local_irq_restore(flags);
1208
1209 hrtimer_forward_now(hrtimer, arm_ccn_pmu_timer_period());
1210 return HRTIMER_RESTART;
1211}
1212
1213
1214static int arm_ccn_pmu_offline_cpu(unsigned int cpu, struct hlist_node *node)
1215{
1216 struct arm_ccn_dt *dt = hlist_entry_safe(node, struct arm_ccn_dt, node);
1217 struct arm_ccn *ccn = container_of(dt, struct arm_ccn, dt);
1218 unsigned int target;
1219
1220 if (!cpumask_test_and_clear_cpu(cpu, &dt->cpu))
1221 return 0;
1222 target = cpumask_any_but(cpu_online_mask, cpu);
1223 if (target >= nr_cpu_ids)
1224 return 0;
1225 perf_pmu_migrate_context(&dt->pmu, cpu, target);
1226 cpumask_set_cpu(target, &dt->cpu);
1227 if (ccn->irq)
1228 WARN_ON(irq_set_affinity_hint(ccn->irq, &dt->cpu) != 0);
1229 return 0;
1230}
1231
1232static DEFINE_IDA(arm_ccn_pmu_ida);
1233
1234static int arm_ccn_pmu_init(struct arm_ccn *ccn)
1235{
1236 int i;
1237 char *name;
1238 int err;
1239
1240 /* Initialize DT subsystem */
1241 ccn->dt.base = ccn->base + CCN_REGION_SIZE;
1242 spin_lock_init(&ccn->dt.config_lock);
1243 writel(CCN_DT_PMOVSR_CLR__MASK, ccn->dt.base + CCN_DT_PMOVSR_CLR);
1244 writel(CCN_DT_CTL__DT_EN, ccn->dt.base + CCN_DT_CTL);
1245 writel(CCN_DT_PMCR__OVFL_INTR_EN | CCN_DT_PMCR__PMU_EN,
1246 ccn->dt.base + CCN_DT_PMCR);
1247 writel(0x1, ccn->dt.base + CCN_DT_PMSR_CLR);
1248 for (i = 0; i < ccn->num_xps; i++) {
1249 writel(0, ccn->xp[i].base + CCN_XP_DT_CONFIG);
1250 writel((CCN_XP_DT_CONTROL__WP_ARM_SEL__ALWAYS <<
1251 CCN_XP_DT_CONTROL__WP_ARM_SEL__SHIFT(0)) |
1252 (CCN_XP_DT_CONTROL__WP_ARM_SEL__ALWAYS <<
1253 CCN_XP_DT_CONTROL__WP_ARM_SEL__SHIFT(1)) |
1254 CCN_XP_DT_CONTROL__DT_ENABLE,
1255 ccn->xp[i].base + CCN_XP_DT_CONTROL);
1256 }
1257 ccn->dt.cmp_mask[CCN_IDX_MASK_ANY].l = ~0;
1258 ccn->dt.cmp_mask[CCN_IDX_MASK_ANY].h = ~0;
1259 ccn->dt.cmp_mask[CCN_IDX_MASK_EXACT].l = 0;
1260 ccn->dt.cmp_mask[CCN_IDX_MASK_EXACT].h = 0;
1261 ccn->dt.cmp_mask[CCN_IDX_MASK_ORDER].l = ~0;
1262 ccn->dt.cmp_mask[CCN_IDX_MASK_ORDER].h = ~(0x1 << 15);
1263 ccn->dt.cmp_mask[CCN_IDX_MASK_OPCODE].l = ~0;
1264 ccn->dt.cmp_mask[CCN_IDX_MASK_OPCODE].h = ~(0x1f << 9);
1265
1266 /* Get a convenient /sys/event_source/devices/ name */
1267 ccn->dt.id = ida_simple_get(&arm_ccn_pmu_ida, 0, 0, GFP_KERNEL);
1268 if (ccn->dt.id == 0) {
1269 name = "ccn";
1270 } else {
1271 name = devm_kasprintf(ccn->dev, GFP_KERNEL, "ccn_%d",
1272 ccn->dt.id);
1273 if (!name) {
1274 err = -ENOMEM;
1275 goto error_choose_name;
1276 }
1277 }
1278
1279 /* Perf driver registration */
1280 ccn->dt.pmu = (struct pmu) {
1281 .module = THIS_MODULE,
1282 .attr_groups = arm_ccn_pmu_attr_groups,
1283 .task_ctx_nr = perf_invalid_context,
1284 .event_init = arm_ccn_pmu_event_init,
1285 .add = arm_ccn_pmu_event_add,
1286 .del = arm_ccn_pmu_event_del,
1287 .start = arm_ccn_pmu_event_start,
1288 .stop = arm_ccn_pmu_event_stop,
1289 .read = arm_ccn_pmu_event_read,
1290 .pmu_enable = arm_ccn_pmu_enable,
1291 .pmu_disable = arm_ccn_pmu_disable,
1292 };
1293
1294 /* No overflow interrupt? Have to use a timer instead. */
1295 if (!ccn->irq) {
1296 dev_info(ccn->dev, "No access to interrupts, using timer.\n");
1297 hrtimer_init(&ccn->dt.hrtimer, CLOCK_MONOTONIC,
1298 HRTIMER_MODE_REL);
1299 ccn->dt.hrtimer.function = arm_ccn_pmu_timer_handler;
1300 }
1301
1302 /* Pick one CPU which we will use to collect data from CCN... */
1303 cpumask_set_cpu(get_cpu(), &ccn->dt.cpu);
1304
1305 /* Also make sure that the overflow interrupt is handled by this CPU */
1306 if (ccn->irq) {
1307 err = irq_set_affinity_hint(ccn->irq, &ccn->dt.cpu);
1308 if (err) {
1309 dev_err(ccn->dev, "Failed to set interrupt affinity!\n");
1310 goto error_set_affinity;
1311 }
1312 }
1313
1314 err = perf_pmu_register(&ccn->dt.pmu, name, -1);
1315 if (err)
1316 goto error_pmu_register;
1317
1318 cpuhp_state_add_instance_nocalls(CPUHP_AP_PERF_ARM_CCN_ONLINE,
1319 &ccn->dt.node);
1320 put_cpu();
1321 return 0;
1322
1323error_pmu_register:
1324error_set_affinity:
1325 put_cpu();
1326error_choose_name:
1327 ida_simple_remove(&arm_ccn_pmu_ida, ccn->dt.id);
1328 for (i = 0; i < ccn->num_xps; i++)
1329 writel(0, ccn->xp[i].base + CCN_XP_DT_CONTROL);
1330 writel(0, ccn->dt.base + CCN_DT_PMCR);
1331 return err;
1332}
1333
1334static void arm_ccn_pmu_cleanup(struct arm_ccn *ccn)
1335{
1336 int i;
1337
1338 cpuhp_state_remove_instance_nocalls(CPUHP_AP_PERF_ARM_CCN_ONLINE,
1339 &ccn->dt.node);
1340 if (ccn->irq)
1341 irq_set_affinity_hint(ccn->irq, NULL);
1342 for (i = 0; i < ccn->num_xps; i++)
1343 writel(0, ccn->xp[i].base + CCN_XP_DT_CONTROL);
1344 writel(0, ccn->dt.base + CCN_DT_PMCR);
1345 perf_pmu_unregister(&ccn->dt.pmu);
1346 ida_simple_remove(&arm_ccn_pmu_ida, ccn->dt.id);
1347}
1348
1349static int arm_ccn_for_each_valid_region(struct arm_ccn *ccn,
1350 int (*callback)(struct arm_ccn *ccn, int region,
1351 void __iomem *base, u32 type, u32 id))
1352{
1353 int region;
1354
1355 for (region = 0; region < CCN_NUM_REGIONS; region++) {
1356 u32 val, type, id;
1357 void __iomem *base;
1358 int err;
1359
1360 val = readl(ccn->base + CCN_MN_OLY_COMP_LIST_63_0 +
1361 4 * (region / 32));
1362 if (!(val & (1 << (region % 32))))
1363 continue;
1364
1365 base = ccn->base + region * CCN_REGION_SIZE;
1366 val = readl(base + CCN_ALL_OLY_ID);
1367 type = (val >> CCN_ALL_OLY_ID__OLY_ID__SHIFT) &
1368 CCN_ALL_OLY_ID__OLY_ID__MASK;
1369 id = (val >> CCN_ALL_OLY_ID__NODE_ID__SHIFT) &
1370 CCN_ALL_OLY_ID__NODE_ID__MASK;
1371
1372 err = callback(ccn, region, base, type, id);
1373 if (err)
1374 return err;
1375 }
1376
1377 return 0;
1378}
1379
1380static int arm_ccn_get_nodes_num(struct arm_ccn *ccn, int region,
1381 void __iomem *base, u32 type, u32 id)
1382{
1383
1384 if (type == CCN_TYPE_XP && id >= ccn->num_xps)
1385 ccn->num_xps = id + 1;
1386 else if (id >= ccn->num_nodes)
1387 ccn->num_nodes = id + 1;
1388
1389 return 0;
1390}
1391
1392static int arm_ccn_init_nodes(struct arm_ccn *ccn, int region,
1393 void __iomem *base, u32 type, u32 id)
1394{
1395 struct arm_ccn_component *component;
1396
1397 dev_dbg(ccn->dev, "Region %d: id=%u, type=0x%02x\n", region, id, type);
1398
1399 switch (type) {
1400 case CCN_TYPE_MN:
1401 ccn->mn_id = id;
1402 return 0;
1403 case CCN_TYPE_DT:
1404 return 0;
1405 case CCN_TYPE_XP:
1406 component = &ccn->xp[id];
1407 break;
1408 case CCN_TYPE_SBSX:
1409 ccn->sbsx_present = 1;
1410 component = &ccn->node[id];
1411 break;
1412 case CCN_TYPE_SBAS:
1413 ccn->sbas_present = 1;
1414 /* Fall-through */
1415 default:
1416 component = &ccn->node[id];
1417 break;
1418 }
1419
1420 component->base = base;
1421 component->type = type;
1422
1423 return 0;
1424}
1425
1426
1427static irqreturn_t arm_ccn_error_handler(struct arm_ccn *ccn,
1428 const u32 *err_sig_val)
1429{
1430 /* This should be really handled by firmware... */
1431 dev_err(ccn->dev, "Error reported in %08x%08x%08x%08x%08x%08x.\n",
1432 err_sig_val[5], err_sig_val[4], err_sig_val[3],
1433 err_sig_val[2], err_sig_val[1], err_sig_val[0]);
1434 dev_err(ccn->dev, "Disabling interrupt generation for all errors.\n");
1435 writel(CCN_MN_ERRINT_STATUS__ALL_ERRORS__DISABLE,
1436 ccn->base + CCN_MN_ERRINT_STATUS);
1437
1438 return IRQ_HANDLED;
1439}
1440
1441
1442static irqreturn_t arm_ccn_irq_handler(int irq, void *dev_id)
1443{
1444 irqreturn_t res = IRQ_NONE;
1445 struct arm_ccn *ccn = dev_id;
1446 u32 err_sig_val[6];
1447 u32 err_or;
1448 int i;
1449
1450 /* PMU overflow is a special case */
1451 err_or = err_sig_val[0] = readl(ccn->base + CCN_MN_ERR_SIG_VAL_63_0);
1452 if (err_or & CCN_MN_ERR_SIG_VAL_63_0__DT) {
1453 err_or &= ~CCN_MN_ERR_SIG_VAL_63_0__DT;
1454 res = arm_ccn_pmu_overflow_handler(&ccn->dt);
1455 }
1456
1457 /* Have to read all err_sig_vals to clear them */
1458 for (i = 1; i < ARRAY_SIZE(err_sig_val); i++) {
1459 err_sig_val[i] = readl(ccn->base +
1460 CCN_MN_ERR_SIG_VAL_63_0 + i * 4);
1461 err_or |= err_sig_val[i];
1462 }
1463 if (err_or)
1464 res |= arm_ccn_error_handler(ccn, err_sig_val);
1465
1466 if (res != IRQ_NONE)
1467 writel(CCN_MN_ERRINT_STATUS__INTREQ__DESSERT,
1468 ccn->base + CCN_MN_ERRINT_STATUS);
1469
1470 return res;
1471}
1472
1473
1474static int arm_ccn_probe(struct platform_device *pdev)
1475{
1476 struct arm_ccn *ccn;
1477 struct resource *res;
1478 unsigned int irq;
1479 int err;
1480
1481 ccn = devm_kzalloc(&pdev->dev, sizeof(*ccn), GFP_KERNEL);
1482 if (!ccn)
1483 return -ENOMEM;
1484 ccn->dev = &pdev->dev;
1485 platform_set_drvdata(pdev, ccn);
1486
1487 res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
1488 if (!res)
1489 return -EINVAL;
1490
1491 if (!devm_request_mem_region(ccn->dev, res->start,
1492 resource_size(res), pdev->name))
1493 return -EBUSY;
1494
1495 ccn->base = devm_ioremap(ccn->dev, res->start,
1496 resource_size(res));
1497 if (!ccn->base)
1498 return -EFAULT;
1499
1500 res = platform_get_resource(pdev, IORESOURCE_IRQ, 0);
1501 if (!res)
1502 return -EINVAL;
1503 irq = res->start;
1504
1505 /* Check if we can use the interrupt */
1506 writel(CCN_MN_ERRINT_STATUS__PMU_EVENTS__DISABLE,
1507 ccn->base + CCN_MN_ERRINT_STATUS);
1508 if (readl(ccn->base + CCN_MN_ERRINT_STATUS) &
1509 CCN_MN_ERRINT_STATUS__PMU_EVENTS__DISABLED) {
1510 /* Can set 'disable' bits, so can acknowledge interrupts */
1511 writel(CCN_MN_ERRINT_STATUS__PMU_EVENTS__ENABLE,
1512 ccn->base + CCN_MN_ERRINT_STATUS);
1513 err = devm_request_irq(ccn->dev, irq, arm_ccn_irq_handler,
1514 IRQF_NOBALANCING | IRQF_NO_THREAD,
1515 dev_name(ccn->dev), ccn);
1516 if (err)
1517 return err;
1518
1519 ccn->irq = irq;
1520 }
1521
1522
1523 /* Build topology */
1524
1525 err = arm_ccn_for_each_valid_region(ccn, arm_ccn_get_nodes_num);
1526 if (err)
1527 return err;
1528
1529 ccn->node = devm_kcalloc(ccn->dev, ccn->num_nodes, sizeof(*ccn->node),
1530 GFP_KERNEL);
1531 ccn->xp = devm_kcalloc(ccn->dev, ccn->num_xps, sizeof(*ccn->node),
1532 GFP_KERNEL);
1533 if (!ccn->node || !ccn->xp)
1534 return -ENOMEM;
1535
1536 err = arm_ccn_for_each_valid_region(ccn, arm_ccn_init_nodes);
1537 if (err)
1538 return err;
1539
1540 return arm_ccn_pmu_init(ccn);
1541}
1542
1543static int arm_ccn_remove(struct platform_device *pdev)
1544{
1545 struct arm_ccn *ccn = platform_get_drvdata(pdev);
1546
1547 arm_ccn_pmu_cleanup(ccn);
1548
1549 return 0;
1550}
1551
1552static const struct of_device_id arm_ccn_match[] = {
1553 { .compatible = "arm,ccn-502", },
1554 { .compatible = "arm,ccn-504", },
1555 {},
1556};
1557MODULE_DEVICE_TABLE(of, arm_ccn_match);
1558
1559static struct platform_driver arm_ccn_driver = {
1560 .driver = {
1561 .name = "arm-ccn",
1562 .of_match_table = arm_ccn_match,
1563 },
1564 .probe = arm_ccn_probe,
1565 .remove = arm_ccn_remove,
1566};
1567
1568static int __init arm_ccn_init(void)
1569{
1570 int i, ret;
1571
1572 ret = cpuhp_setup_state_multi(CPUHP_AP_PERF_ARM_CCN_ONLINE,
1573 "perf/arm/ccn:online", NULL,
1574 arm_ccn_pmu_offline_cpu);
1575 if (ret)
1576 return ret;
1577
1578 for (i = 0; i < ARRAY_SIZE(arm_ccn_pmu_events); i++)
1579 arm_ccn_pmu_events_attrs[i] = &arm_ccn_pmu_events[i].attr.attr;
1580
1581 ret = platform_driver_register(&arm_ccn_driver);
1582 if (ret)
1583 cpuhp_remove_multi_state(CPUHP_AP_PERF_ARM_CCN_ONLINE);
1584 return ret;
1585}
1586
1587static void __exit arm_ccn_exit(void)
1588{
1589 platform_driver_unregister(&arm_ccn_driver);
1590 cpuhp_remove_multi_state(CPUHP_AP_PERF_ARM_CCN_ONLINE);
1591}
1592
1593module_init(arm_ccn_init);
1594module_exit(arm_ccn_exit);
1595
1596MODULE_AUTHOR("Pawel Moll <pawel.moll@arm.com>");
1597MODULE_LICENSE("GPL");