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.
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.
11 * Copyright (C) 2014 ARM Limited
14 #include <linux/ctype.h>
15 #include <linux/hrtimer.h>
16 #include <linux/idr.h>
17 #include <linux/interrupt.h>
19 #include <linux/module.h>
20 #include <linux/perf_event.h>
21 #include <linux/platform_device.h>
22 #include <linux/slab.h>
24 #define CCN_NUM_XP_PORTS 2
26 #define CCN_NUM_REGIONS 256
27 #define CCN_REGION_SIZE 0x10000
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
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)
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
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
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
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
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
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 */
123 #define CCN_EVENT_WATCHPOINT 0xfe /* Pseudoevent */
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
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)
136 struct arm_ccn_component
{
140 DECLARE_BITMAP(pmu_events_mask
, CCN_NUM_PMU_EVENTS
);
143 DECLARE_BITMAP(dt_cmp_mask
, CCN_NUM_XP_WATCHPOINTS
);
148 #define pmu_to_arm_ccn(_pmu) container_of(container_of(_pmu, \
149 struct arm_ccn_dt, pmu), struct arm_ccn, dt)
155 spinlock_t config_lock
;
157 DECLARE_BITMAP(pmu_counters_mask
, CCN_NUM_PMU_EVENT_COUNTERS
+ 1);
159 struct arm_ccn_component
*source
;
160 struct perf_event
*event
;
161 } pmu_counters
[CCN_NUM_PMU_EVENT_COUNTERS
+ 1];
165 } cmp_mask
[CCN_NUM_PMU_EVENT_COUNTERS
+ CCN_NUM_PREDEFINED_MASKS
];
167 struct hrtimer hrtimer
;
170 struct notifier_block cpu_nb
;
180 unsigned sbas_present
:1;
181 unsigned sbsx_present
:1;
184 struct arm_ccn_component
*node
;
187 struct arm_ccn_component
*xp
;
189 struct arm_ccn_dt dt
;
194 static int arm_ccn_node_to_xp(int node
)
196 return node
/ CCN_NUM_XP_PORTS
;
199 static int arm_ccn_node_to_xp_port(int node
)
201 return node
% CCN_NUM_XP_PORTS
;
206 * Bit shifts and masks in these defines must be kept in sync with
207 * arm_ccn_pmu_config_set() and CCN_FORMAT_ATTRs below!
209 #define CCN_CONFIG_NODE(_config) (((_config) >> 0) & 0xff)
210 #define CCN_CONFIG_XP(_config) (((_config) >> 0) & 0xff)
211 #define CCN_CONFIG_TYPE(_config) (((_config) >> 8) & 0xff)
212 #define CCN_CONFIG_EVENT(_config) (((_config) >> 16) & 0xff)
213 #define CCN_CONFIG_PORT(_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)
218 static void arm_ccn_pmu_config_set(u64
*config
, u32 node_xp
, u32 type
, u32 port
)
220 *config
&= ~((0xff << 0) | (0xff << 8) | (0x3 << 24));
221 *config
|= (node_xp
<< 0) | (type
<< 8) | (port
<< 24);
224 static ssize_t
arm_ccn_pmu_format_show(struct device
*dev
,
225 struct device_attribute
*attr
, char *buf
)
227 struct dev_ext_attribute
*ea
= container_of(attr
,
228 struct dev_ext_attribute
, attr
);
230 return snprintf(buf
, PAGE_SIZE
, "%s\n", (char *)ea
->var
);
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, \
238 static CCN_FORMAT_ATTR(node
, "config:0-7");
239 static CCN_FORMAT_ATTR(xp
, "config:0-7");
240 static CCN_FORMAT_ATTR(type
, "config:8-15");
241 static CCN_FORMAT_ATTR(event
, "config:16-23");
242 static CCN_FORMAT_ATTR(port
, "config:24-25");
243 static CCN_FORMAT_ATTR(vc
, "config:26-28");
244 static CCN_FORMAT_ATTR(dir
, "config:29-29");
245 static CCN_FORMAT_ATTR(mask
, "config:30-33");
246 static CCN_FORMAT_ATTR(cmp_l
, "config1:0-62");
247 static CCN_FORMAT_ATTR(cmp_h
, "config2:0-59");
249 static struct attribute
*arm_ccn_pmu_format_attrs
[] = {
250 &arm_ccn_pmu_format_attr_node
.attr
.attr
,
251 &arm_ccn_pmu_format_attr_xp
.attr
.attr
,
252 &arm_ccn_pmu_format_attr_type
.attr
.attr
,
253 &arm_ccn_pmu_format_attr_event
.attr
.attr
,
254 &arm_ccn_pmu_format_attr_port
.attr
.attr
,
255 &arm_ccn_pmu_format_attr_vc
.attr
.attr
,
256 &arm_ccn_pmu_format_attr_dir
.attr
.attr
,
257 &arm_ccn_pmu_format_attr_mask
.attr
.attr
,
258 &arm_ccn_pmu_format_attr_cmp_l
.attr
.attr
,
259 &arm_ccn_pmu_format_attr_cmp_h
.attr
.attr
,
263 static struct attribute_group arm_ccn_pmu_format_attr_group
= {
265 .attrs
= arm_ccn_pmu_format_attrs
,
269 struct arm_ccn_pmu_event
{
270 struct device_attribute attr
;
279 #define CCN_EVENT_ATTR(_name) \
280 __ATTR(_name, S_IRUGO, arm_ccn_pmu_event_show, NULL)
283 * Events defined in TRM for MN, HN-I and SBSX are actually watchpoints set on
284 * their ports in XP they are connected to. For the sake of usability they are
285 * explicitly defined here (and translated into a relevant watchpoint in
286 * arm_ccn_pmu_event_init()) so the user can easily request them without deep
287 * knowledge of the flit format.
290 #define CCN_EVENT_MN(_name, _def, _mask) { .attr = CCN_EVENT_ATTR(mn_##_name), \
291 .type = CCN_TYPE_MN, .event = CCN_EVENT_WATCHPOINT, \
292 .num_ports = CCN_NUM_XP_PORTS, .num_vcs = CCN_NUM_VCS, \
293 .def = _def, .mask = _mask, }
295 #define CCN_EVENT_HNI(_name, _def, _mask) { \
296 .attr = CCN_EVENT_ATTR(hni_##_name), .type = CCN_TYPE_HNI, \
297 .event = CCN_EVENT_WATCHPOINT, .num_ports = CCN_NUM_XP_PORTS, \
298 .num_vcs = CCN_NUM_VCS, .def = _def, .mask = _mask, }
300 #define CCN_EVENT_SBSX(_name, _def, _mask) { \
301 .attr = CCN_EVENT_ATTR(sbsx_##_name), .type = CCN_TYPE_SBSX, \
302 .event = CCN_EVENT_WATCHPOINT, .num_ports = CCN_NUM_XP_PORTS, \
303 .num_vcs = CCN_NUM_VCS, .def = _def, .mask = _mask, }
305 #define CCN_EVENT_HNF(_name, _event) { .attr = CCN_EVENT_ATTR(hnf_##_name), \
306 .type = CCN_TYPE_HNF, .event = _event, }
308 #define CCN_EVENT_XP(_name, _event) { .attr = CCN_EVENT_ATTR(xp_##_name), \
309 .type = CCN_TYPE_XP, .event = _event, \
310 .num_ports = CCN_NUM_XP_PORTS, .num_vcs = CCN_NUM_VCS, }
313 * RN-I & RN-D (RN-D = RN-I + DVM) nodes have different type ID depending
314 * on configuration. One of them is picked to represent the whole group,
315 * as they all share the same event types.
317 #define CCN_EVENT_RNI(_name, _event) { .attr = CCN_EVENT_ATTR(rni_##_name), \
318 .type = CCN_TYPE_RNI_3P, .event = _event, }
320 #define CCN_EVENT_SBAS(_name, _event) { .attr = CCN_EVENT_ATTR(sbas_##_name), \
321 .type = CCN_TYPE_SBAS, .event = _event, }
323 #define CCN_EVENT_CYCLES(_name) { .attr = CCN_EVENT_ATTR(_name), \
324 .type = CCN_TYPE_CYCLES }
327 static ssize_t
arm_ccn_pmu_event_show(struct device
*dev
,
328 struct device_attribute
*attr
, char *buf
)
330 struct arm_ccn
*ccn
= pmu_to_arm_ccn(dev_get_drvdata(dev
));
331 struct arm_ccn_pmu_event
*event
= container_of(attr
,
332 struct arm_ccn_pmu_event
, attr
);
335 res
= snprintf(buf
, PAGE_SIZE
, "type=0x%x", event
->type
);
337 res
+= snprintf(buf
+ res
, PAGE_SIZE
- res
, ",event=0x%x",
340 res
+= snprintf(buf
+ res
, PAGE_SIZE
- res
, ",%s",
343 res
+= snprintf(buf
+ res
, PAGE_SIZE
- res
, ",mask=0x%x",
346 /* Arguments required by an event */
347 switch (event
->type
) {
348 case CCN_TYPE_CYCLES
:
351 res
+= snprintf(buf
+ res
, PAGE_SIZE
- res
,
352 ",xp=?,port=?,vc=?,dir=?");
353 if (event
->event
== CCN_EVENT_WATCHPOINT
)
354 res
+= snprintf(buf
+ res
, PAGE_SIZE
- res
,
355 ",cmp_l=?,cmp_h=?,mask=?");
358 res
+= snprintf(buf
+ res
, PAGE_SIZE
- res
, ",node=%d", ccn
->mn_id
);
361 res
+= snprintf(buf
+ res
, PAGE_SIZE
- res
, ",node=?");
365 res
+= snprintf(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 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 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
, &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 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_warn(ccn
->dev
, "Sampling not supported!\n");
737 if (has_branch_stack(event
) || event
->attr
.exclude_user
||
738 event
->attr
.exclude_kernel
|| event
->attr
.exclude_hv
||
739 event
->attr
.exclude_idle
) {
740 dev_warn(ccn
->dev
, "Can't exclude execution levels!\n");
744 if (event
->cpu
< 0) {
745 dev_warn(ccn
->dev
, "Can't provide per-task data!\n");
749 * Many perf core operations (eg. events rotation) operate on a
750 * single CPU context. This is obvious for CPU PMUs, where one
751 * expects the same sets of events being observed on all CPUs,
752 * but can lead to issues for off-core PMUs, like CCN, where each
753 * event could be theoretically assigned to a different CPU. To
754 * mitigate this, we enforce CPU assignment to one, selected
755 * processor (the one described in the "cpumask" attribute).
757 event
->cpu
= cpumask_first(&ccn
->dt
.cpu
);
759 node_xp
= CCN_CONFIG_NODE(event
->attr
.config
);
760 type
= CCN_CONFIG_TYPE(event
->attr
.config
);
761 event_id
= CCN_CONFIG_EVENT(event
->attr
.config
);
763 /* Validate node/xp vs topology */
766 if (node_xp
!= ccn
->mn_id
) {
767 dev_warn(ccn
->dev
, "Invalid MN ID %d!\n", node_xp
);
772 if (node_xp
>= ccn
->num_xps
) {
773 dev_warn(ccn
->dev
, "Invalid XP ID %d!\n", node_xp
);
777 case CCN_TYPE_CYCLES
:
780 if (node_xp
>= ccn
->num_nodes
) {
781 dev_warn(ccn
->dev
, "Invalid node ID %d!\n", node_xp
);
784 if (!arm_ccn_pmu_type_eq(type
, ccn
->node
[node_xp
].type
)) {
785 dev_warn(ccn
->dev
, "Invalid type 0x%x for node %d!\n",
792 /* Validate event ID vs available for the type */
793 for (i
= 0, valid
= 0; i
< ARRAY_SIZE(arm_ccn_pmu_events
) && !valid
;
795 struct arm_ccn_pmu_event
*e
= &arm_ccn_pmu_events
[i
];
796 u32 port
= CCN_CONFIG_PORT(event
->attr
.config
);
797 u32 vc
= CCN_CONFIG_VC(event
->attr
.config
);
799 if (!arm_ccn_pmu_type_eq(type
, e
->type
))
801 if (event_id
!= e
->event
)
803 if (e
->num_ports
&& port
>= e
->num_ports
) {
804 dev_warn(ccn
->dev
, "Invalid port %d for node/XP %d!\n",
808 if (e
->num_vcs
&& vc
>= e
->num_vcs
) {
809 dev_warn(ccn
->dev
, "Invalid vc %d for node/XP %d!\n",
816 dev_warn(ccn
->dev
, "Invalid event 0x%x for node/XP %d!\n",
821 /* Watchpoint-based event for a node is actually set on XP */
822 if (event_id
== CCN_EVENT_WATCHPOINT
&& type
!= CCN_TYPE_XP
) {
826 port
= arm_ccn_node_to_xp_port(node_xp
);
827 node_xp
= arm_ccn_node_to_xp(node_xp
);
829 arm_ccn_pmu_config_set(&event
->attr
.config
,
830 node_xp
, type
, port
);
834 * We must NOT create groups containing mixed PMUs, although software
835 * events are acceptable (for example to create a CCN group
836 * periodically read when a hrtimer aka cpu-clock leader triggers).
838 if (event
->group_leader
->pmu
!= event
->pmu
&&
839 !is_software_event(event
->group_leader
))
842 list_for_each_entry(sibling
, &event
->group_leader
->sibling_list
,
844 if (sibling
->pmu
!= event
->pmu
&&
845 !is_software_event(sibling
))
851 static u64
arm_ccn_pmu_read_counter(struct arm_ccn
*ccn
, int idx
)
855 if (idx
== CCN_IDX_PMU_CYCLE_COUNTER
) {
857 res
= readq(ccn
->dt
.base
+ CCN_DT_PMCCNTR
);
859 /* 40 bit counter, can do snapshot and read in two parts */
860 writel(0x1, ccn
->dt
.base
+ CCN_DT_PMSR_REQ
);
861 while (!(readl(ccn
->dt
.base
+ CCN_DT_PMSR
) & 0x1))
863 writel(0x1, ccn
->dt
.base
+ CCN_DT_PMSR_CLR
);
864 res
= readl(ccn
->dt
.base
+ CCN_DT_PMCCNTRSR
+ 4) & 0xff;
866 res
|= readl(ccn
->dt
.base
+ CCN_DT_PMCCNTRSR
);
869 res
= readl(ccn
->dt
.base
+ CCN_DT_PMEVCNT(idx
));
875 static void arm_ccn_pmu_event_update(struct perf_event
*event
)
877 struct arm_ccn
*ccn
= pmu_to_arm_ccn(event
->pmu
);
878 struct hw_perf_event
*hw
= &event
->hw
;
879 u64 prev_count
, new_count
, mask
;
882 prev_count
= local64_read(&hw
->prev_count
);
883 new_count
= arm_ccn_pmu_read_counter(ccn
, hw
->idx
);
884 } while (local64_xchg(&hw
->prev_count
, new_count
) != prev_count
);
886 mask
= (1LLU << (hw
->idx
== CCN_IDX_PMU_CYCLE_COUNTER
? 40 : 32)) - 1;
888 local64_add((new_count
- prev_count
) & mask
, &event
->count
);
891 static void arm_ccn_pmu_xp_dt_config(struct perf_event
*event
, int enable
)
893 struct arm_ccn
*ccn
= pmu_to_arm_ccn(event
->pmu
);
894 struct hw_perf_event
*hw
= &event
->hw
;
895 struct arm_ccn_component
*xp
;
898 /* Nothing to do for cycle counter */
899 if (hw
->idx
== CCN_IDX_PMU_CYCLE_COUNTER
)
902 if (CCN_CONFIG_TYPE(event
->attr
.config
) == CCN_TYPE_XP
)
903 xp
= &ccn
->xp
[CCN_CONFIG_XP(event
->attr
.config
)];
905 xp
= &ccn
->xp
[arm_ccn_node_to_xp(
906 CCN_CONFIG_NODE(event
->attr
.config
))];
909 dt_cfg
= hw
->event_base
;
911 dt_cfg
= CCN_XP_DT_CONFIG__DT_CFG__PASS_THROUGH
;
913 spin_lock(&ccn
->dt
.config_lock
);
915 val
= readl(xp
->base
+ CCN_XP_DT_CONFIG
);
916 val
&= ~(CCN_XP_DT_CONFIG__DT_CFG__MASK
<<
917 CCN_XP_DT_CONFIG__DT_CFG__SHIFT(hw
->idx
));
918 val
|= dt_cfg
<< CCN_XP_DT_CONFIG__DT_CFG__SHIFT(hw
->idx
);
919 writel(val
, xp
->base
+ CCN_XP_DT_CONFIG
);
921 spin_unlock(&ccn
->dt
.config_lock
);
924 static void arm_ccn_pmu_event_start(struct perf_event
*event
, int flags
)
926 struct arm_ccn
*ccn
= pmu_to_arm_ccn(event
->pmu
);
927 struct hw_perf_event
*hw
= &event
->hw
;
929 local64_set(&event
->hw
.prev_count
,
930 arm_ccn_pmu_read_counter(ccn
, hw
->idx
));
934 * Pin the timer, so that the overflows are handled by the chosen
935 * event->cpu (this is the same one as presented in "cpumask"
939 hrtimer_start(&ccn
->dt
.hrtimer
, arm_ccn_pmu_timer_period(),
940 HRTIMER_MODE_REL_PINNED
);
942 /* Set the DT bus input, engaging the counter */
943 arm_ccn_pmu_xp_dt_config(event
, 1);
946 static void arm_ccn_pmu_event_stop(struct perf_event
*event
, int flags
)
948 struct arm_ccn
*ccn
= pmu_to_arm_ccn(event
->pmu
);
949 struct hw_perf_event
*hw
= &event
->hw
;
952 /* Disable counting, setting the DT bus to pass-through mode */
953 arm_ccn_pmu_xp_dt_config(event
, 0);
956 hrtimer_cancel(&ccn
->dt
.hrtimer
);
958 /* Let the DT bus drain */
959 timeout
= arm_ccn_pmu_read_counter(ccn
, CCN_IDX_PMU_CYCLE_COUNTER
) +
961 while (arm_ccn_pmu_read_counter(ccn
, CCN_IDX_PMU_CYCLE_COUNTER
) <
965 if (flags
& PERF_EF_UPDATE
)
966 arm_ccn_pmu_event_update(event
);
968 hw
->state
|= PERF_HES_STOPPED
;
971 static void arm_ccn_pmu_xp_watchpoint_config(struct perf_event
*event
)
973 struct arm_ccn
*ccn
= pmu_to_arm_ccn(event
->pmu
);
974 struct hw_perf_event
*hw
= &event
->hw
;
975 struct arm_ccn_component
*source
=
976 ccn
->dt
.pmu_counters
[hw
->idx
].source
;
977 unsigned long wp
= hw
->config_base
;
979 u64 cmp_l
= event
->attr
.config1
;
980 u64 cmp_h
= event
->attr
.config2
;
981 u64 mask_l
= ccn
->dt
.cmp_mask
[CCN_CONFIG_MASK(event
->attr
.config
)].l
;
982 u64 mask_h
= ccn
->dt
.cmp_mask
[CCN_CONFIG_MASK(event
->attr
.config
)].h
;
984 hw
->event_base
= CCN_XP_DT_CONFIG__DT_CFG__WATCHPOINT(wp
);
986 /* Direction (RX/TX), device (port) & virtual channel */
987 val
= readl(source
->base
+ CCN_XP_DT_INTERFACE_SEL
);
988 val
&= ~(CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__MASK
<<
989 CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__SHIFT(wp
));
990 val
|= CCN_CONFIG_DIR(event
->attr
.config
) <<
991 CCN_XP_DT_INTERFACE_SEL__DT_IO_SEL__SHIFT(wp
);
992 val
&= ~(CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__MASK
<<
993 CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__SHIFT(wp
));
994 val
|= CCN_CONFIG_PORT(event
->attr
.config
) <<
995 CCN_XP_DT_INTERFACE_SEL__DT_DEV_SEL__SHIFT(wp
);
996 val
&= ~(CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__MASK
<<
997 CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__SHIFT(wp
));
998 val
|= CCN_CONFIG_VC(event
->attr
.config
) <<
999 CCN_XP_DT_INTERFACE_SEL__DT_VC_SEL__SHIFT(wp
);
1000 writel(val
, source
->base
+ CCN_XP_DT_INTERFACE_SEL
);
1002 /* Comparison values */
1003 writel(cmp_l
& 0xffffffff, source
->base
+ CCN_XP_DT_CMP_VAL_L(wp
));
1004 writel((cmp_l
>> 32) & 0x7fffffff,
1005 source
->base
+ CCN_XP_DT_CMP_VAL_L(wp
) + 4);
1006 writel(cmp_h
& 0xffffffff, source
->base
+ CCN_XP_DT_CMP_VAL_H(wp
));
1007 writel((cmp_h
>> 32) & 0x0fffffff,
1008 source
->base
+ CCN_XP_DT_CMP_VAL_H(wp
) + 4);
1011 writel(mask_l
& 0xffffffff, source
->base
+ CCN_XP_DT_CMP_MASK_L(wp
));
1012 writel((mask_l
>> 32) & 0x7fffffff,
1013 source
->base
+ CCN_XP_DT_CMP_MASK_L(wp
) + 4);
1014 writel(mask_h
& 0xffffffff, source
->base
+ CCN_XP_DT_CMP_MASK_H(wp
));
1015 writel((mask_h
>> 32) & 0x0fffffff,
1016 source
->base
+ CCN_XP_DT_CMP_MASK_H(wp
) + 4);
1019 static void arm_ccn_pmu_xp_event_config(struct perf_event
*event
)
1021 struct arm_ccn
*ccn
= pmu_to_arm_ccn(event
->pmu
);
1022 struct hw_perf_event
*hw
= &event
->hw
;
1023 struct arm_ccn_component
*source
=
1024 ccn
->dt
.pmu_counters
[hw
->idx
].source
;
1027 hw
->event_base
= CCN_XP_DT_CONFIG__DT_CFG__XP_PMU_EVENT(hw
->config_base
);
1029 id
= (CCN_CONFIG_VC(event
->attr
.config
) << 4) |
1030 (CCN_CONFIG_PORT(event
->attr
.config
) << 3) |
1031 (CCN_CONFIG_EVENT(event
->attr
.config
) << 0);
1033 val
= readl(source
->base
+ CCN_XP_PMU_EVENT_SEL
);
1034 val
&= ~(CCN_XP_PMU_EVENT_SEL__ID__MASK
<<
1035 CCN_XP_PMU_EVENT_SEL__ID__SHIFT(hw
->config_base
));
1036 val
|= id
<< CCN_XP_PMU_EVENT_SEL__ID__SHIFT(hw
->config_base
);
1037 writel(val
, source
->base
+ CCN_XP_PMU_EVENT_SEL
);
1040 static void arm_ccn_pmu_node_event_config(struct perf_event
*event
)
1042 struct arm_ccn
*ccn
= pmu_to_arm_ccn(event
->pmu
);
1043 struct hw_perf_event
*hw
= &event
->hw
;
1044 struct arm_ccn_component
*source
=
1045 ccn
->dt
.pmu_counters
[hw
->idx
].source
;
1046 u32 type
= CCN_CONFIG_TYPE(event
->attr
.config
);
1049 port
= arm_ccn_node_to_xp_port(CCN_CONFIG_NODE(event
->attr
.config
));
1050 hw
->event_base
= CCN_XP_DT_CONFIG__DT_CFG__DEVICE_PMU_EVENT(port
,
1053 /* These *_event_sel regs should be identical, but let's make sure... */
1054 BUILD_BUG_ON(CCN_HNF_PMU_EVENT_SEL
!= CCN_SBAS_PMU_EVENT_SEL
);
1055 BUILD_BUG_ON(CCN_SBAS_PMU_EVENT_SEL
!= CCN_RNI_PMU_EVENT_SEL
);
1056 BUILD_BUG_ON(CCN_HNF_PMU_EVENT_SEL__ID__SHIFT(1) !=
1057 CCN_SBAS_PMU_EVENT_SEL__ID__SHIFT(1));
1058 BUILD_BUG_ON(CCN_SBAS_PMU_EVENT_SEL__ID__SHIFT(1) !=
1059 CCN_RNI_PMU_EVENT_SEL__ID__SHIFT(1));
1060 BUILD_BUG_ON(CCN_HNF_PMU_EVENT_SEL__ID__MASK
!=
1061 CCN_SBAS_PMU_EVENT_SEL__ID__MASK
);
1062 BUILD_BUG_ON(CCN_SBAS_PMU_EVENT_SEL__ID__MASK
!=
1063 CCN_RNI_PMU_EVENT_SEL__ID__MASK
);
1064 if (WARN_ON(type
!= CCN_TYPE_HNF
&& type
!= CCN_TYPE_SBAS
&&
1065 !arm_ccn_pmu_type_eq(type
, CCN_TYPE_RNI_3P
)))
1068 /* Set the event id for the pre-allocated counter */
1069 val
= readl(source
->base
+ CCN_HNF_PMU_EVENT_SEL
);
1070 val
&= ~(CCN_HNF_PMU_EVENT_SEL__ID__MASK
<<
1071 CCN_HNF_PMU_EVENT_SEL__ID__SHIFT(hw
->config_base
));
1072 val
|= CCN_CONFIG_EVENT(event
->attr
.config
) <<
1073 CCN_HNF_PMU_EVENT_SEL__ID__SHIFT(hw
->config_base
);
1074 writel(val
, source
->base
+ CCN_HNF_PMU_EVENT_SEL
);
1077 static void arm_ccn_pmu_event_config(struct perf_event
*event
)
1079 struct arm_ccn
*ccn
= pmu_to_arm_ccn(event
->pmu
);
1080 struct hw_perf_event
*hw
= &event
->hw
;
1081 u32 xp
, offset
, val
;
1083 /* Cycle counter requires no setup */
1084 if (hw
->idx
== CCN_IDX_PMU_CYCLE_COUNTER
)
1087 if (CCN_CONFIG_TYPE(event
->attr
.config
) == CCN_TYPE_XP
)
1088 xp
= CCN_CONFIG_XP(event
->attr
.config
);
1090 xp
= arm_ccn_node_to_xp(CCN_CONFIG_NODE(event
->attr
.config
));
1092 spin_lock(&ccn
->dt
.config_lock
);
1094 /* Set the DT bus "distance" register */
1095 offset
= (hw
->idx
/ 4) * 4;
1096 val
= readl(ccn
->dt
.base
+ CCN_DT_ACTIVE_DSM
+ offset
);
1097 val
&= ~(CCN_DT_ACTIVE_DSM__DSM_ID__MASK
<<
1098 CCN_DT_ACTIVE_DSM__DSM_ID__SHIFT(hw
->idx
% 4));
1099 val
|= xp
<< CCN_DT_ACTIVE_DSM__DSM_ID__SHIFT(hw
->idx
% 4);
1100 writel(val
, ccn
->dt
.base
+ CCN_DT_ACTIVE_DSM
+ offset
);
1102 if (CCN_CONFIG_TYPE(event
->attr
.config
) == CCN_TYPE_XP
) {
1103 if (CCN_CONFIG_EVENT(event
->attr
.config
) ==
1104 CCN_EVENT_WATCHPOINT
)
1105 arm_ccn_pmu_xp_watchpoint_config(event
);
1107 arm_ccn_pmu_xp_event_config(event
);
1109 arm_ccn_pmu_node_event_config(event
);
1112 spin_unlock(&ccn
->dt
.config_lock
);
1115 static int arm_ccn_pmu_event_add(struct perf_event
*event
, int flags
)
1118 struct hw_perf_event
*hw
= &event
->hw
;
1120 err
= arm_ccn_pmu_event_alloc(event
);
1124 arm_ccn_pmu_event_config(event
);
1126 hw
->state
= PERF_HES_STOPPED
;
1128 if (flags
& PERF_EF_START
)
1129 arm_ccn_pmu_event_start(event
, PERF_EF_UPDATE
);
1134 static void arm_ccn_pmu_event_del(struct perf_event
*event
, int flags
)
1136 arm_ccn_pmu_event_stop(event
, PERF_EF_UPDATE
);
1138 arm_ccn_pmu_event_release(event
);
1141 static void arm_ccn_pmu_event_read(struct perf_event
*event
)
1143 arm_ccn_pmu_event_update(event
);
1146 static irqreturn_t
arm_ccn_pmu_overflow_handler(struct arm_ccn_dt
*dt
)
1148 u32 pmovsr
= readl(dt
->base
+ CCN_DT_PMOVSR
);
1154 writel(pmovsr
, dt
->base
+ CCN_DT_PMOVSR_CLR
);
1156 BUILD_BUG_ON(CCN_IDX_PMU_CYCLE_COUNTER
!= CCN_NUM_PMU_EVENT_COUNTERS
);
1158 for (idx
= 0; idx
< CCN_NUM_PMU_EVENT_COUNTERS
+ 1; idx
++) {
1159 struct perf_event
*event
= dt
->pmu_counters
[idx
].event
;
1160 int overflowed
= pmovsr
& BIT(idx
);
1162 WARN_ON_ONCE(overflowed
&& !event
&&
1163 idx
!= CCN_IDX_PMU_CYCLE_COUNTER
);
1165 if (!event
|| !overflowed
)
1168 arm_ccn_pmu_event_update(event
);
1174 static enum hrtimer_restart
arm_ccn_pmu_timer_handler(struct hrtimer
*hrtimer
)
1176 struct arm_ccn_dt
*dt
= container_of(hrtimer
, struct arm_ccn_dt
,
1178 unsigned long flags
;
1180 local_irq_save(flags
);
1181 arm_ccn_pmu_overflow_handler(dt
);
1182 local_irq_restore(flags
);
1184 hrtimer_forward_now(hrtimer
, arm_ccn_pmu_timer_period());
1185 return HRTIMER_RESTART
;
1189 static int arm_ccn_pmu_cpu_notifier(struct notifier_block
*nb
,
1190 unsigned long action
, void *hcpu
)
1192 struct arm_ccn_dt
*dt
= container_of(nb
, struct arm_ccn_dt
, cpu_nb
);
1193 struct arm_ccn
*ccn
= container_of(dt
, struct arm_ccn
, dt
);
1194 unsigned int cpu
= (long)hcpu
; /* for (long) see kernel/cpu.c */
1195 unsigned int target
;
1197 switch (action
& ~CPU_TASKS_FROZEN
) {
1198 case CPU_DOWN_PREPARE
:
1199 if (!cpumask_test_and_clear_cpu(cpu
, &dt
->cpu
))
1201 target
= cpumask_any_but(cpu_online_mask
, cpu
);
1202 if (target
>= nr_cpu_ids
)
1204 perf_pmu_migrate_context(&dt
->pmu
, cpu
, target
);
1205 cpumask_set_cpu(target
, &dt
->cpu
);
1207 WARN_ON(irq_set_affinity(ccn
->irq
, &dt
->cpu
) != 0);
1216 static DEFINE_IDA(arm_ccn_pmu_ida
);
1218 static int arm_ccn_pmu_init(struct arm_ccn
*ccn
)
1224 /* Initialize DT subsystem */
1225 ccn
->dt
.base
= ccn
->base
+ CCN_REGION_SIZE
;
1226 spin_lock_init(&ccn
->dt
.config_lock
);
1227 writel(CCN_DT_PMOVSR_CLR__MASK
, ccn
->dt
.base
+ CCN_DT_PMOVSR_CLR
);
1228 writel(CCN_DT_CTL__DT_EN
, ccn
->dt
.base
+ CCN_DT_CTL
);
1229 writel(CCN_DT_PMCR__OVFL_INTR_EN
| CCN_DT_PMCR__PMU_EN
,
1230 ccn
->dt
.base
+ CCN_DT_PMCR
);
1231 writel(0x1, ccn
->dt
.base
+ CCN_DT_PMSR_CLR
);
1232 for (i
= 0; i
< ccn
->num_xps
; i
++) {
1233 writel(0, ccn
->xp
[i
].base
+ CCN_XP_DT_CONFIG
);
1234 writel((CCN_XP_DT_CONTROL__WP_ARM_SEL__ALWAYS
<<
1235 CCN_XP_DT_CONTROL__WP_ARM_SEL__SHIFT(0)) |
1236 (CCN_XP_DT_CONTROL__WP_ARM_SEL__ALWAYS
<<
1237 CCN_XP_DT_CONTROL__WP_ARM_SEL__SHIFT(1)) |
1238 CCN_XP_DT_CONTROL__DT_ENABLE
,
1239 ccn
->xp
[i
].base
+ CCN_XP_DT_CONTROL
);
1241 ccn
->dt
.cmp_mask
[CCN_IDX_MASK_ANY
].l
= ~0;
1242 ccn
->dt
.cmp_mask
[CCN_IDX_MASK_ANY
].h
= ~0;
1243 ccn
->dt
.cmp_mask
[CCN_IDX_MASK_EXACT
].l
= 0;
1244 ccn
->dt
.cmp_mask
[CCN_IDX_MASK_EXACT
].h
= 0;
1245 ccn
->dt
.cmp_mask
[CCN_IDX_MASK_ORDER
].l
= ~0;
1246 ccn
->dt
.cmp_mask
[CCN_IDX_MASK_ORDER
].h
= ~(0x1 << 15);
1247 ccn
->dt
.cmp_mask
[CCN_IDX_MASK_OPCODE
].l
= ~0;
1248 ccn
->dt
.cmp_mask
[CCN_IDX_MASK_OPCODE
].h
= ~(0x1f << 9);
1250 /* Get a convenient /sys/event_source/devices/ name */
1251 ccn
->dt
.id
= ida_simple_get(&arm_ccn_pmu_ida
, 0, 0, GFP_KERNEL
);
1252 if (ccn
->dt
.id
== 0) {
1255 int len
= snprintf(NULL
, 0, "ccn_%d", ccn
->dt
.id
);
1257 name
= devm_kzalloc(ccn
->dev
, len
+ 1, GFP_KERNEL
);
1258 snprintf(name
, len
+ 1, "ccn_%d", ccn
->dt
.id
);
1261 /* Perf driver registration */
1262 ccn
->dt
.pmu
= (struct pmu
) {
1263 .module
= THIS_MODULE
,
1264 .attr_groups
= arm_ccn_pmu_attr_groups
,
1265 .task_ctx_nr
= perf_invalid_context
,
1266 .event_init
= arm_ccn_pmu_event_init
,
1267 .add
= arm_ccn_pmu_event_add
,
1268 .del
= arm_ccn_pmu_event_del
,
1269 .start
= arm_ccn_pmu_event_start
,
1270 .stop
= arm_ccn_pmu_event_stop
,
1271 .read
= arm_ccn_pmu_event_read
,
1274 /* No overflow interrupt? Have to use a timer instead. */
1276 dev_info(ccn
->dev
, "No access to interrupts, using timer.\n");
1277 hrtimer_init(&ccn
->dt
.hrtimer
, CLOCK_MONOTONIC
,
1279 ccn
->dt
.hrtimer
.function
= arm_ccn_pmu_timer_handler
;
1282 /* Pick one CPU which we will use to collect data from CCN... */
1283 cpumask_set_cpu(smp_processor_id(), &ccn
->dt
.cpu
);
1286 * ... and change the selection when it goes offline. Priority is
1287 * picked to have a chance to migrate events before perf is notified.
1289 ccn
->dt
.cpu_nb
.notifier_call
= arm_ccn_pmu_cpu_notifier
;
1290 ccn
->dt
.cpu_nb
.priority
= CPU_PRI_PERF
+ 1,
1291 err
= register_cpu_notifier(&ccn
->dt
.cpu_nb
);
1293 goto error_cpu_notifier
;
1295 /* Also make sure that the overflow interrupt is handled by this CPU */
1297 err
= irq_set_affinity(ccn
->irq
, &ccn
->dt
.cpu
);
1299 dev_err(ccn
->dev
, "Failed to set interrupt affinity!\n");
1300 goto error_set_affinity
;
1304 err
= perf_pmu_register(&ccn
->dt
.pmu
, name
, -1);
1306 goto error_pmu_register
;
1312 unregister_cpu_notifier(&ccn
->dt
.cpu_nb
);
1314 ida_simple_remove(&arm_ccn_pmu_ida
, ccn
->dt
.id
);
1315 for (i
= 0; i
< ccn
->num_xps
; i
++)
1316 writel(0, ccn
->xp
[i
].base
+ CCN_XP_DT_CONTROL
);
1317 writel(0, ccn
->dt
.base
+ CCN_DT_PMCR
);
1321 static void arm_ccn_pmu_cleanup(struct arm_ccn
*ccn
)
1325 irq_set_affinity(ccn
->irq
, cpu_possible_mask
);
1326 unregister_cpu_notifier(&ccn
->dt
.cpu_nb
);
1327 for (i
= 0; i
< ccn
->num_xps
; i
++)
1328 writel(0, ccn
->xp
[i
].base
+ CCN_XP_DT_CONTROL
);
1329 writel(0, ccn
->dt
.base
+ CCN_DT_PMCR
);
1330 perf_pmu_unregister(&ccn
->dt
.pmu
);
1331 ida_simple_remove(&arm_ccn_pmu_ida
, ccn
->dt
.id
);
1335 static 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
))
1341 for (region
= 0; region
< CCN_NUM_REGIONS
; region
++) {
1346 val
= readl(ccn
->base
+ CCN_MN_OLY_COMP_LIST_63_0
+
1348 if (!(val
& (1 << (region
% 32))))
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
;
1358 err
= callback(ccn
, region
, base
, type
, id
);
1366 static int arm_ccn_get_nodes_num(struct arm_ccn
*ccn
, int region
,
1367 void __iomem
*base
, u32 type
, u32 id
)
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;
1378 static int arm_ccn_init_nodes(struct arm_ccn
*ccn
, int region
,
1379 void __iomem
*base
, u32 type
, u32 id
)
1381 struct arm_ccn_component
*component
;
1383 dev_dbg(ccn
->dev
, "Region %d: id=%u, type=0x%02x\n", region
, id
, type
);
1392 component
= &ccn
->xp
[id
];
1395 ccn
->sbsx_present
= 1;
1396 component
= &ccn
->node
[id
];
1399 ccn
->sbas_present
= 1;
1402 component
= &ccn
->node
[id
];
1406 component
->base
= base
;
1407 component
->type
= type
;
1413 static irqreturn_t
arm_ccn_error_handler(struct arm_ccn
*ccn
,
1414 const u32
*err_sig_val
)
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
);
1428 static irqreturn_t
arm_ccn_irq_handler(int irq
, void *dev_id
)
1430 irqreturn_t res
= IRQ_NONE
;
1431 struct arm_ccn
*ccn
= dev_id
;
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
);
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
];
1450 res
|= arm_ccn_error_handler(ccn
, err_sig_val
);
1452 if (res
!= IRQ_NONE
)
1453 writel(CCN_MN_ERRINT_STATUS__INTREQ__DESSERT
,
1454 ccn
->base
+ CCN_MN_ERRINT_STATUS
);
1460 static int arm_ccn_probe(struct platform_device
*pdev
)
1462 struct arm_ccn
*ccn
;
1463 struct resource
*res
;
1467 ccn
= devm_kzalloc(&pdev
->dev
, sizeof(*ccn
), GFP_KERNEL
);
1470 ccn
->dev
= &pdev
->dev
;
1471 platform_set_drvdata(pdev
, ccn
);
1473 res
= platform_get_resource(pdev
, IORESOURCE_MEM
, 0);
1477 if (!devm_request_mem_region(ccn
->dev
, res
->start
,
1478 resource_size(res
), pdev
->name
))
1481 ccn
->base
= devm_ioremap(ccn
->dev
, res
->start
,
1482 resource_size(res
));
1486 res
= platform_get_resource(pdev
, IORESOURCE_IRQ
, 0);
1491 /* Check if we can use the interrupt */
1492 writel(CCN_MN_ERRINT_STATUS__PMU_EVENTS__DISABLE
,
1493 ccn
->base
+ CCN_MN_ERRINT_STATUS
);
1494 if (readl(ccn
->base
+ CCN_MN_ERRINT_STATUS
) &
1495 CCN_MN_ERRINT_STATUS__PMU_EVENTS__DISABLED
) {
1496 /* Can set 'disable' bits, so can acknowledge interrupts */
1497 writel(CCN_MN_ERRINT_STATUS__PMU_EVENTS__ENABLE
,
1498 ccn
->base
+ CCN_MN_ERRINT_STATUS
);
1499 err
= devm_request_irq(ccn
->dev
, irq
, arm_ccn_irq_handler
, 0,
1500 dev_name(ccn
->dev
), ccn
);
1508 /* Build topology */
1510 err
= arm_ccn_for_each_valid_region(ccn
, arm_ccn_get_nodes_num
);
1514 ccn
->node
= devm_kzalloc(ccn
->dev
, sizeof(*ccn
->node
) * ccn
->num_nodes
,
1516 ccn
->xp
= devm_kzalloc(ccn
->dev
, sizeof(*ccn
->node
) * ccn
->num_xps
,
1518 if (!ccn
->node
|| !ccn
->xp
)
1521 err
= arm_ccn_for_each_valid_region(ccn
, arm_ccn_init_nodes
);
1525 return arm_ccn_pmu_init(ccn
);
1528 static int arm_ccn_remove(struct platform_device
*pdev
)
1530 struct arm_ccn
*ccn
= platform_get_drvdata(pdev
);
1532 arm_ccn_pmu_cleanup(ccn
);
1537 static const struct of_device_id arm_ccn_match
[] = {
1538 { .compatible
= "arm,ccn-504", },
1542 static struct platform_driver arm_ccn_driver
= {
1545 .of_match_table
= arm_ccn_match
,
1547 .probe
= arm_ccn_probe
,
1548 .remove
= arm_ccn_remove
,
1551 static int __init
arm_ccn_init(void)
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
;
1558 return platform_driver_register(&arm_ccn_driver
);
1561 static void __exit
arm_ccn_exit(void)
1563 platform_driver_unregister(&arm_ccn_driver
);
1566 module_init(arm_ccn_init
);
1567 module_exit(arm_ccn_exit
);
1569 MODULE_AUTHOR("Pawel Moll <pawel.moll@arm.com>");
1570 MODULE_LICENSE("GPL");