1 // SPDX-License-Identifier: GPL-2.0-only
4 * Copyright (C) 2014 ARM Limited
7 #include <linux/ctype.h>
8 #include <linux/hrtimer.h>
10 #include <linux/interrupt.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>
18 #define CCN_NUM_XP_PORTS 2
20 #define CCN_NUM_REGIONS 256
21 #define CCN_REGION_SIZE 0x10000
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
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)
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
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
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
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
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
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 */
117 #define CCN_EVENT_WATCHPOINT 0xfe /* Pseudoevent */
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
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)
130 struct arm_ccn_component
{
134 DECLARE_BITMAP(pmu_events_mask
, CCN_NUM_PMU_EVENTS
);
137 DECLARE_BITMAP(dt_cmp_mask
, CCN_NUM_XP_WATCHPOINTS
);
142 #define pmu_to_arm_ccn(_pmu) container_of(container_of(_pmu, \
143 struct arm_ccn_dt, pmu), struct arm_ccn, dt)
149 spinlock_t config_lock
;
151 DECLARE_BITMAP(pmu_counters_mask
, CCN_NUM_PMU_EVENT_COUNTERS
+ 1);
153 struct arm_ccn_component
*source
;
154 struct perf_event
*event
;
155 } pmu_counters
[CCN_NUM_PMU_EVENT_COUNTERS
+ 1];
159 } cmp_mask
[CCN_NUM_PMU_EVENT_COUNTERS
+ CCN_NUM_PREDEFINED_MASKS
];
161 struct hrtimer hrtimer
;
164 struct hlist_node node
;
174 unsigned sbas_present
:1;
175 unsigned sbsx_present
:1;
178 struct arm_ccn_component
*node
;
181 struct arm_ccn_component
*xp
;
183 struct arm_ccn_dt dt
;
187 static int arm_ccn_node_to_xp(int node
)
189 return node
/ CCN_NUM_XP_PORTS
;
192 static int arm_ccn_node_to_xp_port(int node
)
194 return node
% CCN_NUM_XP_PORTS
;
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!
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)
212 static void arm_ccn_pmu_config_set(u64
*config
, u32 node_xp
, u32 type
, u32 port
)
214 *config
&= ~((0xff << 0) | (0xff << 8) | (0x3 << 24));
215 *config
|= (node_xp
<< 0) | (type
<< 8) | (port
<< 24);
218 static ssize_t
arm_ccn_pmu_format_show(struct device
*dev
,
219 struct device_attribute
*attr
, char *buf
)
221 struct dev_ext_attribute
*ea
= container_of(attr
,
222 struct dev_ext_attribute
, attr
);
224 return snprintf(buf
, PAGE_SIZE
, "%s\n", (char *)ea
->var
);
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, \
232 static CCN_FORMAT_ATTR(node
, "config:0-7");
233 static CCN_FORMAT_ATTR(xp
, "config:0-7");
234 static CCN_FORMAT_ATTR(type
, "config:8-15");
235 static CCN_FORMAT_ATTR(event
, "config:16-23");
236 static CCN_FORMAT_ATTR(port
, "config:24-25");
237 static CCN_FORMAT_ATTR(bus
, "config:24-25");
238 static CCN_FORMAT_ATTR(vc
, "config:26-28");
239 static CCN_FORMAT_ATTR(dir
, "config:29-29");
240 static CCN_FORMAT_ATTR(mask
, "config:30-33");
241 static CCN_FORMAT_ATTR(cmp_l
, "config1:0-62");
242 static CCN_FORMAT_ATTR(cmp_h
, "config2:0-59");
244 static 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
,
259 static const struct attribute_group arm_ccn_pmu_format_attr_group
= {
261 .attrs
= arm_ccn_pmu_format_attrs
,
265 struct arm_ccn_pmu_event
{
266 struct device_attribute attr
;
275 #define CCN_EVENT_ATTR(_name) \
276 __ATTR(_name, S_IRUGO, arm_ccn_pmu_event_show, NULL)
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.
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, }
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, }
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, }
301 #define CCN_EVENT_HNF(_name, _event) { .attr = CCN_EVENT_ATTR(hnf_##_name), \
302 .type = CCN_TYPE_HNF, .event = _event, }
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, }
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.
313 #define CCN_EVENT_RNI(_name, _event) { .attr = CCN_EVENT_ATTR(rni_##_name), \
314 .type = CCN_TYPE_RNI_3P, .event = _event, }
316 #define CCN_EVENT_SBAS(_name, _event) { .attr = CCN_EVENT_ATTR(sbas_##_name), \
317 .type = CCN_TYPE_SBAS, .event = _event, }
319 #define CCN_EVENT_CYCLES(_name) { .attr = CCN_EVENT_ATTR(_name), \
320 .type = CCN_TYPE_CYCLES }
323 static ssize_t
arm_ccn_pmu_event_show(struct device
*dev
,
324 struct device_attribute
*attr
, char *buf
)
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
);
331 res
= scnprintf(buf
, PAGE_SIZE
, "type=0x%x", event
->type
);
333 res
+= scnprintf(buf
+ res
, PAGE_SIZE
- res
, ",event=0x%x",
336 res
+= scnprintf(buf
+ res
, PAGE_SIZE
- res
, ",%s",
339 res
+= scnprintf(buf
+ res
, PAGE_SIZE
- res
, ",mask=0x%x",
342 /* Arguments required by an event */
343 switch (event
->type
) {
344 case CCN_TYPE_CYCLES
:
347 res
+= scnprintf(buf
+ res
, PAGE_SIZE
- res
,
349 if (event
->event
== CCN_EVENT_WATCHPOINT
)
350 res
+= scnprintf(buf
+ res
, PAGE_SIZE
- res
,
351 ",port=?,dir=?,cmp_l=?,cmp_h=?,mask=?");
353 res
+= scnprintf(buf
+ res
, PAGE_SIZE
- res
,
358 res
+= scnprintf(buf
+ res
, PAGE_SIZE
- res
, ",node=%d", ccn
->mn_id
);
361 res
+= scnprintf(buf
+ res
, PAGE_SIZE
- res
, ",node=?");
365 res
+= scnprintf(buf
+ res
, PAGE_SIZE
- res
, "\n");
370 static umode_t
arm_ccn_pmu_events_is_visible(struct kobject
*kobj
,
371 struct attribute
*attr
, int index
)
373 struct device
*dev
= kobj_to_dev(kobj
);
374 struct arm_ccn
*ccn
= pmu_to_arm_ccn(dev_get_drvdata(dev
));
375 struct device_attribute
*dev_attr
= container_of(attr
,
376 struct device_attribute
, attr
);
377 struct arm_ccn_pmu_event
*event
= container_of(dev_attr
,
378 struct arm_ccn_pmu_event
, attr
);
380 if (event
->type
== CCN_TYPE_SBAS
&& !ccn
->sbas_present
)
382 if (event
->type
== CCN_TYPE_SBSX
&& !ccn
->sbsx_present
)
388 static struct arm_ccn_pmu_event arm_ccn_pmu_events
[] = {
389 CCN_EVENT_MN(eobarrier
, "dir=1,vc=0,cmp_h=0x1c00", CCN_IDX_MASK_OPCODE
),
390 CCN_EVENT_MN(ecbarrier
, "dir=1,vc=0,cmp_h=0x1e00", CCN_IDX_MASK_OPCODE
),
391 CCN_EVENT_MN(dvmop
, "dir=1,vc=0,cmp_h=0x2800", CCN_IDX_MASK_OPCODE
),
392 CCN_EVENT_HNI(txdatflits
, "dir=1,vc=3", CCN_IDX_MASK_ANY
),
393 CCN_EVENT_HNI(rxdatflits
, "dir=0,vc=3", CCN_IDX_MASK_ANY
),
394 CCN_EVENT_HNI(txreqflits
, "dir=1,vc=0", CCN_IDX_MASK_ANY
),
395 CCN_EVENT_HNI(rxreqflits
, "dir=0,vc=0", CCN_IDX_MASK_ANY
),
396 CCN_EVENT_HNI(rxreqflits_order
, "dir=0,vc=0,cmp_h=0x8000",
398 CCN_EVENT_SBSX(txdatflits
, "dir=1,vc=3", CCN_IDX_MASK_ANY
),
399 CCN_EVENT_SBSX(rxdatflits
, "dir=0,vc=3", CCN_IDX_MASK_ANY
),
400 CCN_EVENT_SBSX(txreqflits
, "dir=1,vc=0", CCN_IDX_MASK_ANY
),
401 CCN_EVENT_SBSX(rxreqflits
, "dir=0,vc=0", CCN_IDX_MASK_ANY
),
402 CCN_EVENT_SBSX(rxreqflits_order
, "dir=0,vc=0,cmp_h=0x8000",
404 CCN_EVENT_HNF(cache_miss
, 0x1),
405 CCN_EVENT_HNF(l3_sf_cache_access
, 0x02),
406 CCN_EVENT_HNF(cache_fill
, 0x3),
407 CCN_EVENT_HNF(pocq_retry
, 0x4),
408 CCN_EVENT_HNF(pocq_reqs_recvd
, 0x5),
409 CCN_EVENT_HNF(sf_hit
, 0x6),
410 CCN_EVENT_HNF(sf_evictions
, 0x7),
411 CCN_EVENT_HNF(snoops_sent
, 0x8),
412 CCN_EVENT_HNF(snoops_broadcast
, 0x9),
413 CCN_EVENT_HNF(l3_eviction
, 0xa),
414 CCN_EVENT_HNF(l3_fill_invalid_way
, 0xb),
415 CCN_EVENT_HNF(mc_retries
, 0xc),
416 CCN_EVENT_HNF(mc_reqs
, 0xd),
417 CCN_EVENT_HNF(qos_hh_retry
, 0xe),
418 CCN_EVENT_RNI(rdata_beats_p0
, 0x1),
419 CCN_EVENT_RNI(rdata_beats_p1
, 0x2),
420 CCN_EVENT_RNI(rdata_beats_p2
, 0x3),
421 CCN_EVENT_RNI(rxdat_flits
, 0x4),
422 CCN_EVENT_RNI(txdat_flits
, 0x5),
423 CCN_EVENT_RNI(txreq_flits
, 0x6),
424 CCN_EVENT_RNI(txreq_flits_retried
, 0x7),
425 CCN_EVENT_RNI(rrt_full
, 0x8),
426 CCN_EVENT_RNI(wrt_full
, 0x9),
427 CCN_EVENT_RNI(txreq_flits_replayed
, 0xa),
428 CCN_EVENT_XP(upload_starvation
, 0x1),
429 CCN_EVENT_XP(download_starvation
, 0x2),
430 CCN_EVENT_XP(respin
, 0x3),
431 CCN_EVENT_XP(valid_flit
, 0x4),
432 CCN_EVENT_XP(watchpoint
, CCN_EVENT_WATCHPOINT
),
433 CCN_EVENT_SBAS(rdata_beats_p0
, 0x1),
434 CCN_EVENT_SBAS(rxdat_flits
, 0x4),
435 CCN_EVENT_SBAS(txdat_flits
, 0x5),
436 CCN_EVENT_SBAS(txreq_flits
, 0x6),
437 CCN_EVENT_SBAS(txreq_flits_retried
, 0x7),
438 CCN_EVENT_SBAS(rrt_full
, 0x8),
439 CCN_EVENT_SBAS(wrt_full
, 0x9),
440 CCN_EVENT_SBAS(txreq_flits_replayed
, 0xa),
441 CCN_EVENT_CYCLES(cycles
),
444 /* Populated in arm_ccn_init() */
445 static struct attribute
446 *arm_ccn_pmu_events_attrs
[ARRAY_SIZE(arm_ccn_pmu_events
) + 1];
448 static const struct attribute_group arm_ccn_pmu_events_attr_group
= {
450 .is_visible
= arm_ccn_pmu_events_is_visible
,
451 .attrs
= arm_ccn_pmu_events_attrs
,
455 static u64
*arm_ccn_pmu_get_cmp_mask(struct arm_ccn
*ccn
, const char *name
)
459 if (WARN_ON(!name
|| !name
[0] || !isxdigit(name
[0]) || !name
[1]))
461 i
= isdigit(name
[0]) ? name
[0] - '0' : 0xa + tolower(name
[0]) - 'a';
465 return &ccn
->dt
.cmp_mask
[i
].l
;
467 return &ccn
->dt
.cmp_mask
[i
].h
;
473 static ssize_t
arm_ccn_pmu_cmp_mask_show(struct device
*dev
,
474 struct device_attribute
*attr
, char *buf
)
476 struct arm_ccn
*ccn
= pmu_to_arm_ccn(dev_get_drvdata(dev
));
477 u64
*mask
= arm_ccn_pmu_get_cmp_mask(ccn
, attr
->attr
.name
);
479 return mask
? snprintf(buf
, PAGE_SIZE
, "0x%016llx\n", *mask
) : -EINVAL
;
482 static ssize_t
arm_ccn_pmu_cmp_mask_store(struct device
*dev
,
483 struct device_attribute
*attr
, const char *buf
, size_t count
)
485 struct arm_ccn
*ccn
= pmu_to_arm_ccn(dev_get_drvdata(dev
));
486 u64
*mask
= arm_ccn_pmu_get_cmp_mask(ccn
, attr
->attr
.name
);
490 err
= kstrtoull(buf
, 0, mask
);
492 return err
? err
: count
;
495 #define CCN_CMP_MASK_ATTR(_name) \
496 struct device_attribute arm_ccn_pmu_cmp_mask_attr_##_name = \
497 __ATTR(_name, S_IRUGO | S_IWUSR, \
498 arm_ccn_pmu_cmp_mask_show, arm_ccn_pmu_cmp_mask_store)
500 #define CCN_CMP_MASK_ATTR_RO(_name) \
501 struct device_attribute arm_ccn_pmu_cmp_mask_attr_##_name = \
502 __ATTR(_name, S_IRUGO, arm_ccn_pmu_cmp_mask_show, NULL)
504 static CCN_CMP_MASK_ATTR(0l);
505 static CCN_CMP_MASK_ATTR(0h
);
506 static CCN_CMP_MASK_ATTR(1l);
507 static CCN_CMP_MASK_ATTR(1h
);
508 static CCN_CMP_MASK_ATTR(2l);
509 static CCN_CMP_MASK_ATTR(2h
);
510 static CCN_CMP_MASK_ATTR(3l);
511 static CCN_CMP_MASK_ATTR(3h
);
512 static CCN_CMP_MASK_ATTR(4l);
513 static CCN_CMP_MASK_ATTR(4h
);
514 static CCN_CMP_MASK_ATTR(5l);
515 static CCN_CMP_MASK_ATTR(5h
);
516 static CCN_CMP_MASK_ATTR(6l);
517 static CCN_CMP_MASK_ATTR(6h
);
518 static CCN_CMP_MASK_ATTR(7l);
519 static CCN_CMP_MASK_ATTR(7h
);
520 static CCN_CMP_MASK_ATTR_RO(8l);
521 static CCN_CMP_MASK_ATTR_RO(8h
);
522 static CCN_CMP_MASK_ATTR_RO(9l);
523 static CCN_CMP_MASK_ATTR_RO(9h
);
524 static CCN_CMP_MASK_ATTR_RO(al
);
525 static CCN_CMP_MASK_ATTR_RO(ah
);
526 static CCN_CMP_MASK_ATTR_RO(bl
);
527 static CCN_CMP_MASK_ATTR_RO(bh
);
529 static struct attribute
*arm_ccn_pmu_cmp_mask_attrs
[] = {
530 &arm_ccn_pmu_cmp_mask_attr_0l
.attr
, &arm_ccn_pmu_cmp_mask_attr_0h
.attr
,
531 &arm_ccn_pmu_cmp_mask_attr_1l
.attr
, &arm_ccn_pmu_cmp_mask_attr_1h
.attr
,
532 &arm_ccn_pmu_cmp_mask_attr_2l
.attr
, &arm_ccn_pmu_cmp_mask_attr_2h
.attr
,
533 &arm_ccn_pmu_cmp_mask_attr_3l
.attr
, &arm_ccn_pmu_cmp_mask_attr_3h
.attr
,
534 &arm_ccn_pmu_cmp_mask_attr_4l
.attr
, &arm_ccn_pmu_cmp_mask_attr_4h
.attr
,
535 &arm_ccn_pmu_cmp_mask_attr_5l
.attr
, &arm_ccn_pmu_cmp_mask_attr_5h
.attr
,
536 &arm_ccn_pmu_cmp_mask_attr_6l
.attr
, &arm_ccn_pmu_cmp_mask_attr_6h
.attr
,
537 &arm_ccn_pmu_cmp_mask_attr_7l
.attr
, &arm_ccn_pmu_cmp_mask_attr_7h
.attr
,
538 &arm_ccn_pmu_cmp_mask_attr_8l
.attr
, &arm_ccn_pmu_cmp_mask_attr_8h
.attr
,
539 &arm_ccn_pmu_cmp_mask_attr_9l
.attr
, &arm_ccn_pmu_cmp_mask_attr_9h
.attr
,
540 &arm_ccn_pmu_cmp_mask_attr_al
.attr
, &arm_ccn_pmu_cmp_mask_attr_ah
.attr
,
541 &arm_ccn_pmu_cmp_mask_attr_bl
.attr
, &arm_ccn_pmu_cmp_mask_attr_bh
.attr
,
545 static const struct attribute_group arm_ccn_pmu_cmp_mask_attr_group
= {
547 .attrs
= arm_ccn_pmu_cmp_mask_attrs
,
550 static ssize_t
arm_ccn_pmu_cpumask_show(struct device
*dev
,
551 struct device_attribute
*attr
, char *buf
)
553 struct arm_ccn
*ccn
= pmu_to_arm_ccn(dev_get_drvdata(dev
));
555 return cpumap_print_to_pagebuf(true, buf
, cpumask_of(ccn
->dt
.cpu
));
558 static struct device_attribute arm_ccn_pmu_cpumask_attr
=
559 __ATTR(cpumask
, S_IRUGO
, arm_ccn_pmu_cpumask_show
, NULL
);
561 static struct attribute
*arm_ccn_pmu_cpumask_attrs
[] = {
562 &arm_ccn_pmu_cpumask_attr
.attr
,
566 static const struct attribute_group arm_ccn_pmu_cpumask_attr_group
= {
567 .attrs
= arm_ccn_pmu_cpumask_attrs
,
571 * Default poll period is 10ms, which is way over the top anyway,
572 * as in the worst case scenario (an event every cycle), with 1GHz
573 * clocked bus, the smallest, 32 bit counter will overflow in
576 static unsigned int arm_ccn_pmu_poll_period_us
= 10000;
577 module_param_named(pmu_poll_period_us
, arm_ccn_pmu_poll_period_us
, uint
,
580 static ktime_t
arm_ccn_pmu_timer_period(void)
582 return ns_to_ktime((u64
)arm_ccn_pmu_poll_period_us
* 1000);
586 static const struct attribute_group
*arm_ccn_pmu_attr_groups
[] = {
587 &arm_ccn_pmu_events_attr_group
,
588 &arm_ccn_pmu_format_attr_group
,
589 &arm_ccn_pmu_cmp_mask_attr_group
,
590 &arm_ccn_pmu_cpumask_attr_group
,
595 static int arm_ccn_pmu_alloc_bit(unsigned long *bitmap
, unsigned long size
)
600 bit
= find_first_zero_bit(bitmap
, size
);
603 } while (test_and_set_bit(bit
, bitmap
));
608 /* All RN-I and RN-D nodes have identical PMUs */
609 static int arm_ccn_pmu_type_eq(u32 a
, u32 b
)
615 case CCN_TYPE_RNI_1P
:
616 case CCN_TYPE_RNI_2P
:
617 case CCN_TYPE_RNI_3P
:
618 case CCN_TYPE_RND_1P
:
619 case CCN_TYPE_RND_2P
:
620 case CCN_TYPE_RND_3P
:
622 case CCN_TYPE_RNI_1P
:
623 case CCN_TYPE_RNI_2P
:
624 case CCN_TYPE_RNI_3P
:
625 case CCN_TYPE_RND_1P
:
626 case CCN_TYPE_RND_2P
:
627 case CCN_TYPE_RND_3P
:
636 static int arm_ccn_pmu_event_alloc(struct perf_event
*event
)
638 struct arm_ccn
*ccn
= pmu_to_arm_ccn(event
->pmu
);
639 struct hw_perf_event
*hw
= &event
->hw
;
640 u32 node_xp
, type
, event_id
;
641 struct arm_ccn_component
*source
;
644 node_xp
= CCN_CONFIG_NODE(event
->attr
.config
);
645 type
= CCN_CONFIG_TYPE(event
->attr
.config
);
646 event_id
= CCN_CONFIG_EVENT(event
->attr
.config
);
648 /* Allocate the cycle counter */
649 if (type
== CCN_TYPE_CYCLES
) {
650 if (test_and_set_bit(CCN_IDX_PMU_CYCLE_COUNTER
,
651 ccn
->dt
.pmu_counters_mask
))
654 hw
->idx
= CCN_IDX_PMU_CYCLE_COUNTER
;
655 ccn
->dt
.pmu_counters
[CCN_IDX_PMU_CYCLE_COUNTER
].event
= event
;
660 /* Allocate an event counter */
661 hw
->idx
= arm_ccn_pmu_alloc_bit(ccn
->dt
.pmu_counters_mask
,
662 CCN_NUM_PMU_EVENT_COUNTERS
);
664 dev_dbg(ccn
->dev
, "No more counters available!\n");
668 if (type
== CCN_TYPE_XP
)
669 source
= &ccn
->xp
[node_xp
];
671 source
= &ccn
->node
[node_xp
];
672 ccn
->dt
.pmu_counters
[hw
->idx
].source
= source
;
674 /* Allocate an event source or a watchpoint */
675 if (type
== CCN_TYPE_XP
&& event_id
== CCN_EVENT_WATCHPOINT
)
676 bit
= arm_ccn_pmu_alloc_bit(source
->xp
.dt_cmp_mask
,
677 CCN_NUM_XP_WATCHPOINTS
);
679 bit
= arm_ccn_pmu_alloc_bit(source
->pmu_events_mask
,
682 dev_dbg(ccn
->dev
, "No more event sources/watchpoints on node/XP %d!\n",
684 clear_bit(hw
->idx
, ccn
->dt
.pmu_counters_mask
);
687 hw
->config_base
= bit
;
689 ccn
->dt
.pmu_counters
[hw
->idx
].event
= event
;
694 static void arm_ccn_pmu_event_release(struct perf_event
*event
)
696 struct arm_ccn
*ccn
= pmu_to_arm_ccn(event
->pmu
);
697 struct hw_perf_event
*hw
= &event
->hw
;
699 if (hw
->idx
== CCN_IDX_PMU_CYCLE_COUNTER
) {
700 clear_bit(CCN_IDX_PMU_CYCLE_COUNTER
, ccn
->dt
.pmu_counters_mask
);
702 struct arm_ccn_component
*source
=
703 ccn
->dt
.pmu_counters
[hw
->idx
].source
;
705 if (CCN_CONFIG_TYPE(event
->attr
.config
) == CCN_TYPE_XP
&&
706 CCN_CONFIG_EVENT(event
->attr
.config
) ==
707 CCN_EVENT_WATCHPOINT
)
708 clear_bit(hw
->config_base
, source
->xp
.dt_cmp_mask
);
710 clear_bit(hw
->config_base
, source
->pmu_events_mask
);
711 clear_bit(hw
->idx
, ccn
->dt
.pmu_counters_mask
);
714 ccn
->dt
.pmu_counters
[hw
->idx
].source
= NULL
;
715 ccn
->dt
.pmu_counters
[hw
->idx
].event
= NULL
;
718 static int arm_ccn_pmu_event_init(struct perf_event
*event
)
721 struct hw_perf_event
*hw
= &event
->hw
;
722 u32 node_xp
, type
, event_id
;
725 struct perf_event
*sibling
;
727 if (event
->attr
.type
!= event
->pmu
->type
)
730 ccn
= pmu_to_arm_ccn(event
->pmu
);
732 if (hw
->sample_period
) {
733 dev_dbg(ccn
->dev
, "Sampling not supported!\n");
737 if (has_branch_stack(event
)) {
738 dev_dbg(ccn
->dev
, "Can't exclude execution levels!\n");
742 if (event
->cpu
< 0) {
743 dev_dbg(ccn
->dev
, "Can't provide per-task data!\n");
747 * Many perf core operations (eg. events rotation) operate on a
748 * single CPU context. This is obvious for CPU PMUs, where one
749 * expects the same sets of events being observed on all CPUs,
750 * but can lead to issues for off-core PMUs, like CCN, where each
751 * event could be theoretically assigned to a different CPU. To
752 * mitigate this, we enforce CPU assignment to one, selected
753 * processor (the one described in the "cpumask" attribute).
755 event
->cpu
= ccn
->dt
.cpu
;
757 node_xp
= CCN_CONFIG_NODE(event
->attr
.config
);
758 type
= CCN_CONFIG_TYPE(event
->attr
.config
);
759 event_id
= CCN_CONFIG_EVENT(event
->attr
.config
);
761 /* Validate node/xp vs topology */
764 if (node_xp
!= ccn
->mn_id
) {
765 dev_dbg(ccn
->dev
, "Invalid MN ID %d!\n", node_xp
);
770 if (node_xp
>= ccn
->num_xps
) {
771 dev_dbg(ccn
->dev
, "Invalid XP ID %d!\n", node_xp
);
775 case CCN_TYPE_CYCLES
:
778 if (node_xp
>= ccn
->num_nodes
) {
779 dev_dbg(ccn
->dev
, "Invalid node ID %d!\n", node_xp
);
782 if (!arm_ccn_pmu_type_eq(type
, ccn
->node
[node_xp
].type
)) {
783 dev_dbg(ccn
->dev
, "Invalid type 0x%x for node %d!\n",
790 /* Validate event ID vs available for the type */
791 for (i
= 0, valid
= 0; i
< ARRAY_SIZE(arm_ccn_pmu_events
) && !valid
;
793 struct arm_ccn_pmu_event
*e
= &arm_ccn_pmu_events
[i
];
794 u32 port
= CCN_CONFIG_PORT(event
->attr
.config
);
795 u32 vc
= CCN_CONFIG_VC(event
->attr
.config
);
797 if (!arm_ccn_pmu_type_eq(type
, e
->type
))
799 if (event_id
!= e
->event
)
801 if (e
->num_ports
&& port
>= e
->num_ports
) {
802 dev_dbg(ccn
->dev
, "Invalid port %d for node/XP %d!\n",
806 if (e
->num_vcs
&& vc
>= e
->num_vcs
) {
807 dev_dbg(ccn
->dev
, "Invalid vc %d for node/XP %d!\n",
814 dev_dbg(ccn
->dev
, "Invalid event 0x%x for node/XP %d!\n",
819 /* Watchpoint-based event for a node is actually set on XP */
820 if (event_id
== CCN_EVENT_WATCHPOINT
&& type
!= CCN_TYPE_XP
) {
824 port
= arm_ccn_node_to_xp_port(node_xp
);
825 node_xp
= arm_ccn_node_to_xp(node_xp
);
827 arm_ccn_pmu_config_set(&event
->attr
.config
,
828 node_xp
, type
, port
);
832 * We must NOT create groups containing mixed PMUs, although software
833 * events are acceptable (for example to create a CCN group
834 * periodically read when a hrtimer aka cpu-clock leader triggers).
836 if (event
->group_leader
->pmu
!= event
->pmu
&&
837 !is_software_event(event
->group_leader
))
840 for_each_sibling_event(sibling
, event
->group_leader
) {
841 if (sibling
->pmu
!= event
->pmu
&&
842 !is_software_event(sibling
))
849 static u64
arm_ccn_pmu_read_counter(struct arm_ccn
*ccn
, int idx
)
853 if (idx
== CCN_IDX_PMU_CYCLE_COUNTER
) {
855 res
= readq(ccn
->dt
.base
+ CCN_DT_PMCCNTR
);
857 /* 40 bit counter, can do snapshot and read in two parts */
858 writel(0x1, ccn
->dt
.base
+ CCN_DT_PMSR_REQ
);
859 while (!(readl(ccn
->dt
.base
+ CCN_DT_PMSR
) & 0x1))
861 writel(0x1, ccn
->dt
.base
+ CCN_DT_PMSR_CLR
);
862 res
= readl(ccn
->dt
.base
+ CCN_DT_PMCCNTRSR
+ 4) & 0xff;
864 res
|= readl(ccn
->dt
.base
+ CCN_DT_PMCCNTRSR
);
867 res
= readl(ccn
->dt
.base
+ CCN_DT_PMEVCNT(idx
));
873 static void arm_ccn_pmu_event_update(struct perf_event
*event
)
875 struct arm_ccn
*ccn
= pmu_to_arm_ccn(event
->pmu
);
876 struct hw_perf_event
*hw
= &event
->hw
;
877 u64 prev_count
, new_count
, mask
;
880 prev_count
= local64_read(&hw
->prev_count
);
881 new_count
= arm_ccn_pmu_read_counter(ccn
, hw
->idx
);
882 } while (local64_xchg(&hw
->prev_count
, new_count
) != prev_count
);
884 mask
= (1LLU << (hw
->idx
== CCN_IDX_PMU_CYCLE_COUNTER
? 40 : 32)) - 1;
886 local64_add((new_count
- prev_count
) & mask
, &event
->count
);
889 static void arm_ccn_pmu_xp_dt_config(struct perf_event
*event
, int enable
)
891 struct arm_ccn
*ccn
= pmu_to_arm_ccn(event
->pmu
);
892 struct hw_perf_event
*hw
= &event
->hw
;
893 struct arm_ccn_component
*xp
;
896 /* Nothing to do for cycle counter */
897 if (hw
->idx
== CCN_IDX_PMU_CYCLE_COUNTER
)
900 if (CCN_CONFIG_TYPE(event
->attr
.config
) == CCN_TYPE_XP
)
901 xp
= &ccn
->xp
[CCN_CONFIG_XP(event
->attr
.config
)];
903 xp
= &ccn
->xp
[arm_ccn_node_to_xp(
904 CCN_CONFIG_NODE(event
->attr
.config
))];
907 dt_cfg
= hw
->event_base
;
909 dt_cfg
= CCN_XP_DT_CONFIG__DT_CFG__PASS_THROUGH
;
911 spin_lock(&ccn
->dt
.config_lock
);
913 val
= readl(xp
->base
+ CCN_XP_DT_CONFIG
);
914 val
&= ~(CCN_XP_DT_CONFIG__DT_CFG__MASK
<<
915 CCN_XP_DT_CONFIG__DT_CFG__SHIFT(hw
->idx
));
916 val
|= dt_cfg
<< CCN_XP_DT_CONFIG__DT_CFG__SHIFT(hw
->idx
);
917 writel(val
, xp
->base
+ CCN_XP_DT_CONFIG
);
919 spin_unlock(&ccn
->dt
.config_lock
);
922 static void arm_ccn_pmu_event_start(struct perf_event
*event
, int flags
)
924 struct arm_ccn
*ccn
= pmu_to_arm_ccn(event
->pmu
);
925 struct hw_perf_event
*hw
= &event
->hw
;
927 local64_set(&event
->hw
.prev_count
,
928 arm_ccn_pmu_read_counter(ccn
, hw
->idx
));
931 /* Set the DT bus input, engaging the counter */
932 arm_ccn_pmu_xp_dt_config(event
, 1);
935 static void arm_ccn_pmu_event_stop(struct perf_event
*event
, int flags
)
937 struct hw_perf_event
*hw
= &event
->hw
;
939 /* Disable counting, setting the DT bus to pass-through mode */
940 arm_ccn_pmu_xp_dt_config(event
, 0);
942 if (flags
& PERF_EF_UPDATE
)
943 arm_ccn_pmu_event_update(event
);
945 hw
->state
|= PERF_HES_STOPPED
;
948 static void arm_ccn_pmu_xp_watchpoint_config(struct perf_event
*event
)
950 struct arm_ccn
*ccn
= pmu_to_arm_ccn(event
->pmu
);
951 struct hw_perf_event
*hw
= &event
->hw
;
952 struct arm_ccn_component
*source
=
953 ccn
->dt
.pmu_counters
[hw
->idx
].source
;
954 unsigned long wp
= hw
->config_base
;
956 u64 cmp_l
= event
->attr
.config1
;
957 u64 cmp_h
= event
->attr
.config2
;
958 u64 mask_l
= ccn
->dt
.cmp_mask
[CCN_CONFIG_MASK(event
->attr
.config
)].l
;
959 u64 mask_h
= ccn
->dt
.cmp_mask
[CCN_CONFIG_MASK(event
->attr
.config
)].h
;
961 hw
->event_base
= CCN_XP_DT_CONFIG__DT_CFG__WATCHPOINT(wp
);
963 /* Direction (RX/TX), device (port) & virtual channel */
964 val
= readl(source
->base
+ CCN_XP_DT_INTERFACE_SEL
);
965 val
&= ~(CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__MASK
<<
966 CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__SHIFT(wp
));
967 val
|= CCN_CONFIG_DIR(event
->attr
.config
) <<
968 CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__SHIFT(wp
);
969 val
&= ~(CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__MASK
<<
970 CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__SHIFT(wp
));
971 val
|= CCN_CONFIG_PORT(event
->attr
.config
) <<
972 CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__SHIFT(wp
);
973 val
&= ~(CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__MASK
<<
974 CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__SHIFT(wp
));
975 val
|= CCN_CONFIG_VC(event
->attr
.config
) <<
976 CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__SHIFT(wp
);
977 writel(val
, source
->base
+ CCN_XP_DT_INTERFACE_SEL
);
979 /* Comparison values */
980 writel(cmp_l
& 0xffffffff, source
->base
+ CCN_XP_DT_CMP_VAL_L(wp
));
981 writel((cmp_l
>> 32) & 0x7fffffff,
982 source
->base
+ CCN_XP_DT_CMP_VAL_L(wp
) + 4);
983 writel(cmp_h
& 0xffffffff, source
->base
+ CCN_XP_DT_CMP_VAL_H(wp
));
984 writel((cmp_h
>> 32) & 0x0fffffff,
985 source
->base
+ CCN_XP_DT_CMP_VAL_H(wp
) + 4);
988 writel(mask_l
& 0xffffffff, source
->base
+ CCN_XP_DT_CMP_MASK_L(wp
));
989 writel((mask_l
>> 32) & 0x7fffffff,
990 source
->base
+ CCN_XP_DT_CMP_MASK_L(wp
) + 4);
991 writel(mask_h
& 0xffffffff, source
->base
+ CCN_XP_DT_CMP_MASK_H(wp
));
992 writel((mask_h
>> 32) & 0x0fffffff,
993 source
->base
+ CCN_XP_DT_CMP_MASK_H(wp
) + 4);
996 static void arm_ccn_pmu_xp_event_config(struct perf_event
*event
)
998 struct arm_ccn
*ccn
= pmu_to_arm_ccn(event
->pmu
);
999 struct hw_perf_event
*hw
= &event
->hw
;
1000 struct arm_ccn_component
*source
=
1001 ccn
->dt
.pmu_counters
[hw
->idx
].source
;
1004 hw
->event_base
= CCN_XP_DT_CONFIG__DT_CFG__XP_PMU_EVENT(hw
->config_base
);
1006 id
= (CCN_CONFIG_VC(event
->attr
.config
) << 4) |
1007 (CCN_CONFIG_BUS(event
->attr
.config
) << 3) |
1008 (CCN_CONFIG_EVENT(event
->attr
.config
) << 0);
1010 val
= readl(source
->base
+ CCN_XP_PMU_EVENT_SEL
);
1011 val
&= ~(CCN_XP_PMU_EVENT_SEL__ID__MASK
<<
1012 CCN_XP_PMU_EVENT_SEL__ID__SHIFT(hw
->config_base
));
1013 val
|= id
<< CCN_XP_PMU_EVENT_SEL__ID__SHIFT(hw
->config_base
);
1014 writel(val
, source
->base
+ CCN_XP_PMU_EVENT_SEL
);
1017 static void arm_ccn_pmu_node_event_config(struct perf_event
*event
)
1019 struct arm_ccn
*ccn
= pmu_to_arm_ccn(event
->pmu
);
1020 struct hw_perf_event
*hw
= &event
->hw
;
1021 struct arm_ccn_component
*source
=
1022 ccn
->dt
.pmu_counters
[hw
->idx
].source
;
1023 u32 type
= CCN_CONFIG_TYPE(event
->attr
.config
);
1026 port
= arm_ccn_node_to_xp_port(CCN_CONFIG_NODE(event
->attr
.config
));
1027 hw
->event_base
= CCN_XP_DT_CONFIG__DT_CFG__DEVICE_PMU_EVENT(port
,
1030 /* These *_event_sel regs should be identical, but let's make sure... */
1031 BUILD_BUG_ON(CCN_HNF_PMU_EVENT_SEL
!= CCN_SBAS_PMU_EVENT_SEL
);
1032 BUILD_BUG_ON(CCN_SBAS_PMU_EVENT_SEL
!= CCN_RNI_PMU_EVENT_SEL
);
1033 BUILD_BUG_ON(CCN_HNF_PMU_EVENT_SEL__ID__SHIFT(1) !=
1034 CCN_SBAS_PMU_EVENT_SEL__ID__SHIFT(1));
1035 BUILD_BUG_ON(CCN_SBAS_PMU_EVENT_SEL__ID__SHIFT(1) !=
1036 CCN_RNI_PMU_EVENT_SEL__ID__SHIFT(1));
1037 BUILD_BUG_ON(CCN_HNF_PMU_EVENT_SEL__ID__MASK
!=
1038 CCN_SBAS_PMU_EVENT_SEL__ID__MASK
);
1039 BUILD_BUG_ON(CCN_SBAS_PMU_EVENT_SEL__ID__MASK
!=
1040 CCN_RNI_PMU_EVENT_SEL__ID__MASK
);
1041 if (WARN_ON(type
!= CCN_TYPE_HNF
&& type
!= CCN_TYPE_SBAS
&&
1042 !arm_ccn_pmu_type_eq(type
, CCN_TYPE_RNI_3P
)))
1045 /* Set the event id for the pre-allocated counter */
1046 val
= readl(source
->base
+ CCN_HNF_PMU_EVENT_SEL
);
1047 val
&= ~(CCN_HNF_PMU_EVENT_SEL__ID__MASK
<<
1048 CCN_HNF_PMU_EVENT_SEL__ID__SHIFT(hw
->config_base
));
1049 val
|= CCN_CONFIG_EVENT(event
->attr
.config
) <<
1050 CCN_HNF_PMU_EVENT_SEL__ID__SHIFT(hw
->config_base
);
1051 writel(val
, source
->base
+ CCN_HNF_PMU_EVENT_SEL
);
1054 static void arm_ccn_pmu_event_config(struct perf_event
*event
)
1056 struct arm_ccn
*ccn
= pmu_to_arm_ccn(event
->pmu
);
1057 struct hw_perf_event
*hw
= &event
->hw
;
1058 u32 xp
, offset
, val
;
1060 /* Cycle counter requires no setup */
1061 if (hw
->idx
== CCN_IDX_PMU_CYCLE_COUNTER
)
1064 if (CCN_CONFIG_TYPE(event
->attr
.config
) == CCN_TYPE_XP
)
1065 xp
= CCN_CONFIG_XP(event
->attr
.config
);
1067 xp
= arm_ccn_node_to_xp(CCN_CONFIG_NODE(event
->attr
.config
));
1069 spin_lock(&ccn
->dt
.config_lock
);
1071 /* Set the DT bus "distance" register */
1072 offset
= (hw
->idx
/ 4) * 4;
1073 val
= readl(ccn
->dt
.base
+ CCN_DT_ACTIVE_DSM
+ offset
);
1074 val
&= ~(CCN_DT_ACTIVE_DSM__DSM_ID__MASK
<<
1075 CCN_DT_ACTIVE_DSM__DSM_ID__SHIFT(hw
->idx
% 4));
1076 val
|= xp
<< CCN_DT_ACTIVE_DSM__DSM_ID__SHIFT(hw
->idx
% 4);
1077 writel(val
, ccn
->dt
.base
+ CCN_DT_ACTIVE_DSM
+ offset
);
1079 if (CCN_CONFIG_TYPE(event
->attr
.config
) == CCN_TYPE_XP
) {
1080 if (CCN_CONFIG_EVENT(event
->attr
.config
) ==
1081 CCN_EVENT_WATCHPOINT
)
1082 arm_ccn_pmu_xp_watchpoint_config(event
);
1084 arm_ccn_pmu_xp_event_config(event
);
1086 arm_ccn_pmu_node_event_config(event
);
1089 spin_unlock(&ccn
->dt
.config_lock
);
1092 static int arm_ccn_pmu_active_counters(struct arm_ccn
*ccn
)
1094 return bitmap_weight(ccn
->dt
.pmu_counters_mask
,
1095 CCN_NUM_PMU_EVENT_COUNTERS
+ 1);
1098 static int arm_ccn_pmu_event_add(struct perf_event
*event
, int flags
)
1101 struct hw_perf_event
*hw
= &event
->hw
;
1102 struct arm_ccn
*ccn
= pmu_to_arm_ccn(event
->pmu
);
1104 err
= arm_ccn_pmu_event_alloc(event
);
1109 * Pin the timer, so that the overflows are handled by the chosen
1110 * event->cpu (this is the same one as presented in "cpumask"
1113 if (!ccn
->irq
&& arm_ccn_pmu_active_counters(ccn
) == 1)
1114 hrtimer_start(&ccn
->dt
.hrtimer
, arm_ccn_pmu_timer_period(),
1115 HRTIMER_MODE_REL_PINNED
);
1117 arm_ccn_pmu_event_config(event
);
1119 hw
->state
= PERF_HES_STOPPED
;
1121 if (flags
& PERF_EF_START
)
1122 arm_ccn_pmu_event_start(event
, PERF_EF_UPDATE
);
1127 static void arm_ccn_pmu_event_del(struct perf_event
*event
, int flags
)
1129 struct arm_ccn
*ccn
= pmu_to_arm_ccn(event
->pmu
);
1131 arm_ccn_pmu_event_stop(event
, PERF_EF_UPDATE
);
1133 arm_ccn_pmu_event_release(event
);
1135 if (!ccn
->irq
&& arm_ccn_pmu_active_counters(ccn
) == 0)
1136 hrtimer_cancel(&ccn
->dt
.hrtimer
);
1139 static void arm_ccn_pmu_event_read(struct perf_event
*event
)
1141 arm_ccn_pmu_event_update(event
);
1144 static void arm_ccn_pmu_enable(struct pmu
*pmu
)
1146 struct arm_ccn
*ccn
= pmu_to_arm_ccn(pmu
);
1148 u32 val
= readl(ccn
->dt
.base
+ CCN_DT_PMCR
);
1149 val
|= CCN_DT_PMCR__PMU_EN
;
1150 writel(val
, ccn
->dt
.base
+ CCN_DT_PMCR
);
1153 static void arm_ccn_pmu_disable(struct pmu
*pmu
)
1155 struct arm_ccn
*ccn
= pmu_to_arm_ccn(pmu
);
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
);
1162 static irqreturn_t
arm_ccn_pmu_overflow_handler(struct arm_ccn_dt
*dt
)
1164 u32 pmovsr
= readl(dt
->base
+ CCN_DT_PMOVSR
);
1170 writel(pmovsr
, dt
->base
+ CCN_DT_PMOVSR_CLR
);
1172 BUILD_BUG_ON(CCN_IDX_PMU_CYCLE_COUNTER
!= CCN_NUM_PMU_EVENT_COUNTERS
);
1174 for (idx
= 0; idx
< CCN_NUM_PMU_EVENT_COUNTERS
+ 1; idx
++) {
1175 struct perf_event
*event
= dt
->pmu_counters
[idx
].event
;
1176 int overflowed
= pmovsr
& BIT(idx
);
1178 WARN_ON_ONCE(overflowed
&& !event
&&
1179 idx
!= CCN_IDX_PMU_CYCLE_COUNTER
);
1181 if (!event
|| !overflowed
)
1184 arm_ccn_pmu_event_update(event
);
1190 static enum hrtimer_restart
arm_ccn_pmu_timer_handler(struct hrtimer
*hrtimer
)
1192 struct arm_ccn_dt
*dt
= container_of(hrtimer
, struct arm_ccn_dt
,
1194 unsigned long flags
;
1196 local_irq_save(flags
);
1197 arm_ccn_pmu_overflow_handler(dt
);
1198 local_irq_restore(flags
);
1200 hrtimer_forward_now(hrtimer
, arm_ccn_pmu_timer_period());
1201 return HRTIMER_RESTART
;
1205 static int arm_ccn_pmu_offline_cpu(unsigned int cpu
, struct hlist_node
*node
)
1207 struct arm_ccn_dt
*dt
= hlist_entry_safe(node
, struct arm_ccn_dt
, node
);
1208 struct arm_ccn
*ccn
= container_of(dt
, struct arm_ccn
, dt
);
1209 unsigned int target
;
1213 target
= cpumask_any_but(cpu_online_mask
, cpu
);
1214 if (target
>= nr_cpu_ids
)
1216 perf_pmu_migrate_context(&dt
->pmu
, cpu
, target
);
1219 WARN_ON(irq_set_affinity_hint(ccn
->irq
, cpumask_of(dt
->cpu
)));
1223 static DEFINE_IDA(arm_ccn_pmu_ida
);
1225 static int arm_ccn_pmu_init(struct arm_ccn
*ccn
)
1231 /* Initialize DT subsystem */
1232 ccn
->dt
.base
= ccn
->base
+ CCN_REGION_SIZE
;
1233 spin_lock_init(&ccn
->dt
.config_lock
);
1234 writel(CCN_DT_PMOVSR_CLR__MASK
, ccn
->dt
.base
+ CCN_DT_PMOVSR_CLR
);
1235 writel(CCN_DT_CTL__DT_EN
, ccn
->dt
.base
+ CCN_DT_CTL
);
1236 writel(CCN_DT_PMCR__OVFL_INTR_EN
| CCN_DT_PMCR__PMU_EN
,
1237 ccn
->dt
.base
+ CCN_DT_PMCR
);
1238 writel(0x1, ccn
->dt
.base
+ CCN_DT_PMSR_CLR
);
1239 for (i
= 0; i
< ccn
->num_xps
; i
++) {
1240 writel(0, ccn
->xp
[i
].base
+ CCN_XP_DT_CONFIG
);
1241 writel((CCN_XP_DT_CONTROL__WP_ARM_SEL__ALWAYS
<<
1242 CCN_XP_DT_CONTROL__WP_ARM_SEL__SHIFT(0)) |
1243 (CCN_XP_DT_CONTROL__WP_ARM_SEL__ALWAYS
<<
1244 CCN_XP_DT_CONTROL__WP_ARM_SEL__SHIFT(1)) |
1245 CCN_XP_DT_CONTROL__DT_ENABLE
,
1246 ccn
->xp
[i
].base
+ CCN_XP_DT_CONTROL
);
1248 ccn
->dt
.cmp_mask
[CCN_IDX_MASK_ANY
].l
= ~0;
1249 ccn
->dt
.cmp_mask
[CCN_IDX_MASK_ANY
].h
= ~0;
1250 ccn
->dt
.cmp_mask
[CCN_IDX_MASK_EXACT
].l
= 0;
1251 ccn
->dt
.cmp_mask
[CCN_IDX_MASK_EXACT
].h
= 0;
1252 ccn
->dt
.cmp_mask
[CCN_IDX_MASK_ORDER
].l
= ~0;
1253 ccn
->dt
.cmp_mask
[CCN_IDX_MASK_ORDER
].h
= ~(0x1 << 15);
1254 ccn
->dt
.cmp_mask
[CCN_IDX_MASK_OPCODE
].l
= ~0;
1255 ccn
->dt
.cmp_mask
[CCN_IDX_MASK_OPCODE
].h
= ~(0x1f << 9);
1257 /* Get a convenient /sys/event_source/devices/ name */
1258 ccn
->dt
.id
= ida_simple_get(&arm_ccn_pmu_ida
, 0, 0, GFP_KERNEL
);
1259 if (ccn
->dt
.id
== 0) {
1262 name
= devm_kasprintf(ccn
->dev
, GFP_KERNEL
, "ccn_%d",
1266 goto error_choose_name
;
1270 /* Perf driver registration */
1271 ccn
->dt
.pmu
= (struct pmu
) {
1272 .module
= THIS_MODULE
,
1273 .attr_groups
= arm_ccn_pmu_attr_groups
,
1274 .task_ctx_nr
= perf_invalid_context
,
1275 .event_init
= arm_ccn_pmu_event_init
,
1276 .add
= arm_ccn_pmu_event_add
,
1277 .del
= arm_ccn_pmu_event_del
,
1278 .start
= arm_ccn_pmu_event_start
,
1279 .stop
= arm_ccn_pmu_event_stop
,
1280 .read
= arm_ccn_pmu_event_read
,
1281 .pmu_enable
= arm_ccn_pmu_enable
,
1282 .pmu_disable
= arm_ccn_pmu_disable
,
1283 .capabilities
= PERF_PMU_CAP_NO_EXCLUDE
,
1286 /* No overflow interrupt? Have to use a timer instead. */
1288 dev_info(ccn
->dev
, "No access to interrupts, using timer.\n");
1289 hrtimer_init(&ccn
->dt
.hrtimer
, CLOCK_MONOTONIC
,
1291 ccn
->dt
.hrtimer
.function
= arm_ccn_pmu_timer_handler
;
1294 /* Pick one CPU which we will use to collect data from CCN... */
1295 ccn
->dt
.cpu
= raw_smp_processor_id();
1297 /* Also make sure that the overflow interrupt is handled by this CPU */
1299 err
= irq_set_affinity_hint(ccn
->irq
, cpumask_of(ccn
->dt
.cpu
));
1301 dev_err(ccn
->dev
, "Failed to set interrupt affinity!\n");
1302 goto error_set_affinity
;
1306 cpuhp_state_add_instance_nocalls(CPUHP_AP_PERF_ARM_CCN_ONLINE
,
1309 err
= perf_pmu_register(&ccn
->dt
.pmu
, name
, -1);
1311 goto error_pmu_register
;
1316 cpuhp_state_remove_instance_nocalls(CPUHP_AP_PERF_ARM_CCN_ONLINE
,
1320 ida_simple_remove(&arm_ccn_pmu_ida
, ccn
->dt
.id
);
1321 for (i
= 0; i
< ccn
->num_xps
; i
++)
1322 writel(0, ccn
->xp
[i
].base
+ CCN_XP_DT_CONTROL
);
1323 writel(0, ccn
->dt
.base
+ CCN_DT_PMCR
);
1327 static void arm_ccn_pmu_cleanup(struct arm_ccn
*ccn
)
1331 cpuhp_state_remove_instance_nocalls(CPUHP_AP_PERF_ARM_CCN_ONLINE
,
1334 irq_set_affinity_hint(ccn
->irq
, NULL
);
1335 for (i
= 0; i
< ccn
->num_xps
; i
++)
1336 writel(0, ccn
->xp
[i
].base
+ CCN_XP_DT_CONTROL
);
1337 writel(0, ccn
->dt
.base
+ CCN_DT_PMCR
);
1338 perf_pmu_unregister(&ccn
->dt
.pmu
);
1339 ida_simple_remove(&arm_ccn_pmu_ida
, ccn
->dt
.id
);
1342 static int arm_ccn_for_each_valid_region(struct arm_ccn
*ccn
,
1343 int (*callback
)(struct arm_ccn
*ccn
, int region
,
1344 void __iomem
*base
, u32 type
, u32 id
))
1348 for (region
= 0; region
< CCN_NUM_REGIONS
; region
++) {
1353 val
= readl(ccn
->base
+ CCN_MN_OLY_COMP_LIST_63_0
+
1355 if (!(val
& (1 << (region
% 32))))
1358 base
= ccn
->base
+ region
* CCN_REGION_SIZE
;
1359 val
= readl(base
+ CCN_ALL_OLY_ID
);
1360 type
= (val
>> CCN_ALL_OLY_ID__OLY_ID__SHIFT
) &
1361 CCN_ALL_OLY_ID__OLY_ID__MASK
;
1362 id
= (val
>> CCN_ALL_OLY_ID__NODE_ID__SHIFT
) &
1363 CCN_ALL_OLY_ID__NODE_ID__MASK
;
1365 err
= callback(ccn
, region
, base
, type
, id
);
1373 static int arm_ccn_get_nodes_num(struct arm_ccn
*ccn
, int region
,
1374 void __iomem
*base
, u32 type
, u32 id
)
1377 if (type
== CCN_TYPE_XP
&& id
>= ccn
->num_xps
)
1378 ccn
->num_xps
= id
+ 1;
1379 else if (id
>= ccn
->num_nodes
)
1380 ccn
->num_nodes
= id
+ 1;
1385 static int arm_ccn_init_nodes(struct arm_ccn
*ccn
, int region
,
1386 void __iomem
*base
, u32 type
, u32 id
)
1388 struct arm_ccn_component
*component
;
1390 dev_dbg(ccn
->dev
, "Region %d: id=%u, type=0x%02x\n", region
, id
, type
);
1399 component
= &ccn
->xp
[id
];
1402 ccn
->sbsx_present
= 1;
1403 component
= &ccn
->node
[id
];
1406 ccn
->sbas_present
= 1;
1409 component
= &ccn
->node
[id
];
1413 component
->base
= base
;
1414 component
->type
= type
;
1420 static irqreturn_t
arm_ccn_error_handler(struct arm_ccn
*ccn
,
1421 const u32
*err_sig_val
)
1423 /* This should be really handled by firmware... */
1424 dev_err(ccn
->dev
, "Error reported in %08x%08x%08x%08x%08x%08x.\n",
1425 err_sig_val
[5], err_sig_val
[4], err_sig_val
[3],
1426 err_sig_val
[2], err_sig_val
[1], err_sig_val
[0]);
1427 dev_err(ccn
->dev
, "Disabling interrupt generation for all errors.\n");
1428 writel(CCN_MN_ERRINT_STATUS__ALL_ERRORS__DISABLE
,
1429 ccn
->base
+ CCN_MN_ERRINT_STATUS
);
1435 static irqreturn_t
arm_ccn_irq_handler(int irq
, void *dev_id
)
1437 irqreturn_t res
= IRQ_NONE
;
1438 struct arm_ccn
*ccn
= dev_id
;
1443 /* PMU overflow is a special case */
1444 err_or
= err_sig_val
[0] = readl(ccn
->base
+ CCN_MN_ERR_SIG_VAL_63_0
);
1445 if (err_or
& CCN_MN_ERR_SIG_VAL_63_0__DT
) {
1446 err_or
&= ~CCN_MN_ERR_SIG_VAL_63_0__DT
;
1447 res
= arm_ccn_pmu_overflow_handler(&ccn
->dt
);
1450 /* Have to read all err_sig_vals to clear them */
1451 for (i
= 1; i
< ARRAY_SIZE(err_sig_val
); i
++) {
1452 err_sig_val
[i
] = readl(ccn
->base
+
1453 CCN_MN_ERR_SIG_VAL_63_0
+ i
* 4);
1454 err_or
|= err_sig_val
[i
];
1457 res
|= arm_ccn_error_handler(ccn
, err_sig_val
);
1459 if (res
!= IRQ_NONE
)
1460 writel(CCN_MN_ERRINT_STATUS__INTREQ__DESSERT
,
1461 ccn
->base
+ CCN_MN_ERRINT_STATUS
);
1467 static int arm_ccn_probe(struct platform_device
*pdev
)
1469 struct arm_ccn
*ccn
;
1470 struct resource
*res
;
1474 ccn
= devm_kzalloc(&pdev
->dev
, sizeof(*ccn
), GFP_KERNEL
);
1477 ccn
->dev
= &pdev
->dev
;
1478 platform_set_drvdata(pdev
, ccn
);
1480 ccn
->base
= devm_platform_ioremap_resource(pdev
, 0);
1481 if (IS_ERR(ccn
->base
))
1482 return PTR_ERR(ccn
->base
);
1484 res
= platform_get_resource(pdev
, IORESOURCE_IRQ
, 0);
1489 /* Check if we can use the interrupt */
1490 writel(CCN_MN_ERRINT_STATUS__PMU_EVENTS__DISABLE
,
1491 ccn
->base
+ CCN_MN_ERRINT_STATUS
);
1492 if (readl(ccn
->base
+ CCN_MN_ERRINT_STATUS
) &
1493 CCN_MN_ERRINT_STATUS__PMU_EVENTS__DISABLED
) {
1494 /* Can set 'disable' bits, so can acknowledge interrupts */
1495 writel(CCN_MN_ERRINT_STATUS__PMU_EVENTS__ENABLE
,
1496 ccn
->base
+ CCN_MN_ERRINT_STATUS
);
1497 err
= devm_request_irq(ccn
->dev
, irq
, arm_ccn_irq_handler
,
1498 IRQF_NOBALANCING
| IRQF_NO_THREAD
,
1499 dev_name(ccn
->dev
), ccn
);
1507 /* Build topology */
1509 err
= arm_ccn_for_each_valid_region(ccn
, arm_ccn_get_nodes_num
);
1513 ccn
->node
= devm_kcalloc(ccn
->dev
, ccn
->num_nodes
, sizeof(*ccn
->node
),
1515 ccn
->xp
= devm_kcalloc(ccn
->dev
, ccn
->num_xps
, sizeof(*ccn
->node
),
1517 if (!ccn
->node
|| !ccn
->xp
)
1520 err
= arm_ccn_for_each_valid_region(ccn
, arm_ccn_init_nodes
);
1524 return arm_ccn_pmu_init(ccn
);
1527 static int arm_ccn_remove(struct platform_device
*pdev
)
1529 struct arm_ccn
*ccn
= platform_get_drvdata(pdev
);
1531 arm_ccn_pmu_cleanup(ccn
);
1536 static const struct of_device_id arm_ccn_match
[] = {
1537 { .compatible
= "arm,ccn-502", },
1538 { .compatible
= "arm,ccn-504", },
1539 { .compatible
= "arm,ccn-512", },
1542 MODULE_DEVICE_TABLE(of
, arm_ccn_match
);
1544 static struct platform_driver arm_ccn_driver
= {
1547 .of_match_table
= arm_ccn_match
,
1548 .suppress_bind_attrs
= true,
1550 .probe
= arm_ccn_probe
,
1551 .remove
= arm_ccn_remove
,
1554 static int __init
arm_ccn_init(void)
1558 ret
= cpuhp_setup_state_multi(CPUHP_AP_PERF_ARM_CCN_ONLINE
,
1559 "perf/arm/ccn:online", NULL
,
1560 arm_ccn_pmu_offline_cpu
);
1564 for (i
= 0; i
< ARRAY_SIZE(arm_ccn_pmu_events
); i
++)
1565 arm_ccn_pmu_events_attrs
[i
] = &arm_ccn_pmu_events
[i
].attr
.attr
;
1567 ret
= platform_driver_register(&arm_ccn_driver
);
1569 cpuhp_remove_multi_state(CPUHP_AP_PERF_ARM_CCN_ONLINE
);
1573 static void __exit
arm_ccn_exit(void)
1575 platform_driver_unregister(&arm_ccn_driver
);
1576 cpuhp_remove_multi_state(CPUHP_AP_PERF_ARM_CCN_ONLINE
);
1579 module_init(arm_ccn_init
);
1580 module_exit(arm_ccn_exit
);
1582 MODULE_AUTHOR("Pawel Moll <pawel.moll@arm.com>");
1583 MODULE_LICENSE("GPL v2");