]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - arch/arc/kernel/perf_event.c
ARC: perf: Add some comments/debug stuff
[karo-tx-linux.git] / arch / arc / kernel / perf_event.c
1 /*
2  * Linux performance counter support for ARC700 series
3  *
4  * Copyright (C) 2013 Synopsys, Inc. (www.synopsys.com)
5  *
6  * This code is inspired by the perf support of various other architectures.
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License version 2 as
10  * published by the Free Software Foundation.
11  *
12  */
13 #include <linux/errno.h>
14 #include <linux/module.h>
15 #include <linux/of.h>
16 #include <linux/perf_event.h>
17 #include <linux/platform_device.h>
18 #include <asm/arcregs.h>
19
20 struct arc_pmu {
21         struct pmu      pmu;
22         int             counter_size;   /* in bits */
23         int             n_counters;
24         unsigned long   used_mask[BITS_TO_LONGS(ARC_PMU_MAX_HWEVENTS)];
25         int             ev_hw_idx[PERF_COUNT_ARC_HW_MAX];
26 };
27
28 static struct arc_pmu *arc_pmu;
29
30 /* read counter #idx; note that counter# != event# on ARC! */
31 static uint64_t arc_pmu_read_counter(int idx)
32 {
33         uint32_t tmp;
34         uint64_t result;
35
36         /*
37          * ARC supports making 'snapshots' of the counters, so we don't
38          * need to care about counters wrapping to 0 underneath our feet
39          */
40         write_aux_reg(ARC_REG_PCT_INDEX, idx);
41         tmp = read_aux_reg(ARC_REG_PCT_CONTROL);
42         write_aux_reg(ARC_REG_PCT_CONTROL, tmp | ARC_REG_PCT_CONTROL_SN);
43         result = (uint64_t) (read_aux_reg(ARC_REG_PCT_SNAPH)) << 32;
44         result |= read_aux_reg(ARC_REG_PCT_SNAPL);
45
46         return result;
47 }
48
49 static void arc_perf_event_update(struct perf_event *event,
50                                   struct hw_perf_event *hwc, int idx)
51 {
52         uint64_t prev_raw_count, new_raw_count;
53         int64_t delta;
54
55         do {
56                 prev_raw_count = local64_read(&hwc->prev_count);
57                 new_raw_count = arc_pmu_read_counter(idx);
58         } while (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
59                                  new_raw_count) != prev_raw_count);
60
61         delta = (new_raw_count - prev_raw_count) &
62                 ((1ULL << arc_pmu->counter_size) - 1ULL);
63
64         local64_add(delta, &event->count);
65         local64_sub(delta, &hwc->period_left);
66 }
67
68 static void arc_pmu_read(struct perf_event *event)
69 {
70         arc_perf_event_update(event, &event->hw, event->hw.idx);
71 }
72
73 static int arc_pmu_cache_event(u64 config)
74 {
75         unsigned int cache_type, cache_op, cache_result;
76         int ret;
77
78         cache_type      = (config >>  0) & 0xff;
79         cache_op        = (config >>  8) & 0xff;
80         cache_result    = (config >> 16) & 0xff;
81         if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
82                 return -EINVAL;
83         if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
84                 return -EINVAL;
85         if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
86                 return -EINVAL;
87
88         ret = arc_pmu_cache_map[cache_type][cache_op][cache_result];
89
90         if (ret == CACHE_OP_UNSUPPORTED)
91                 return -ENOENT;
92
93         pr_debug("init cache event: type/op/result %d/%d/%d with h/w %d \'%s\'\n",
94                  cache_type, cache_op, cache_result, ret,
95                  arc_pmu_ev_hw_map[ret]);
96
97         return ret;
98 }
99
100 /* initializes hw_perf_event structure if event is supported */
101 static int arc_pmu_event_init(struct perf_event *event)
102 {
103         struct hw_perf_event *hwc = &event->hw;
104         int ret;
105
106         switch (event->attr.type) {
107         case PERF_TYPE_HARDWARE:
108                 if (event->attr.config >= PERF_COUNT_HW_MAX)
109                         return -ENOENT;
110                 if (arc_pmu->ev_hw_idx[event->attr.config] < 0)
111                         return -ENOENT;
112                 hwc->config = arc_pmu->ev_hw_idx[event->attr.config];
113                 pr_debug("init event %d with h/w %d \'%s\'\n",
114                          (int) event->attr.config, (int) hwc->config,
115                          arc_pmu_ev_hw_map[event->attr.config]);
116                 return 0;
117         case PERF_TYPE_HW_CACHE:
118                 ret = arc_pmu_cache_event(event->attr.config);
119                 if (ret < 0)
120                         return ret;
121                 hwc->config = arc_pmu->ev_hw_idx[ret];
122                 return 0;
123         default:
124                 return -ENOENT;
125         }
126 }
127
128 /* starts all counters */
129 static void arc_pmu_enable(struct pmu *pmu)
130 {
131         uint32_t tmp;
132         tmp = read_aux_reg(ARC_REG_PCT_CONTROL);
133         write_aux_reg(ARC_REG_PCT_CONTROL, (tmp & 0xffff0000) | 0x1);
134 }
135
136 /* stops all counters */
137 static void arc_pmu_disable(struct pmu *pmu)
138 {
139         uint32_t tmp;
140         tmp = read_aux_reg(ARC_REG_PCT_CONTROL);
141         write_aux_reg(ARC_REG_PCT_CONTROL, (tmp & 0xffff0000) | 0x0);
142 }
143
144 /*
145  * Assigns hardware counter to hardware condition.
146  * Note that there is no separate start/stop mechanism;
147  * stopping is achieved by assigning the 'never' condition
148  */
149 static void arc_pmu_start(struct perf_event *event, int flags)
150 {
151         struct hw_perf_event *hwc = &event->hw;
152         int idx = hwc->idx;
153
154         if (WARN_ON_ONCE(idx == -1))
155                 return;
156
157         if (flags & PERF_EF_RELOAD)
158                 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
159
160         event->hw.state = 0;
161
162         /* enable ARC pmu here */
163         write_aux_reg(ARC_REG_PCT_INDEX, idx);
164         write_aux_reg(ARC_REG_PCT_CONFIG, hwc->config);
165 }
166
167 static void arc_pmu_stop(struct perf_event *event, int flags)
168 {
169         struct hw_perf_event *hwc = &event->hw;
170         int idx = hwc->idx;
171
172         if (!(event->hw.state & PERF_HES_STOPPED)) {
173                 /* stop ARC pmu here */
174                 write_aux_reg(ARC_REG_PCT_INDEX, idx);
175
176                 /* condition code #0 is always "never" */
177                 write_aux_reg(ARC_REG_PCT_CONFIG, 0);
178
179                 event->hw.state |= PERF_HES_STOPPED;
180         }
181
182         if ((flags & PERF_EF_UPDATE) &&
183             !(event->hw.state & PERF_HES_UPTODATE)) {
184                 arc_perf_event_update(event, &event->hw, idx);
185                 event->hw.state |= PERF_HES_UPTODATE;
186         }
187 }
188
189 static void arc_pmu_del(struct perf_event *event, int flags)
190 {
191         arc_pmu_stop(event, PERF_EF_UPDATE);
192         __clear_bit(event->hw.idx, arc_pmu->used_mask);
193
194         perf_event_update_userpage(event);
195 }
196
197 /* allocate hardware counter and optionally start counting */
198 static int arc_pmu_add(struct perf_event *event, int flags)
199 {
200         struct hw_perf_event *hwc = &event->hw;
201         int idx = hwc->idx;
202
203         if (__test_and_set_bit(idx, arc_pmu->used_mask)) {
204                 idx = find_first_zero_bit(arc_pmu->used_mask,
205                                           arc_pmu->n_counters);
206                 if (idx == arc_pmu->n_counters)
207                         return -EAGAIN;
208
209                 __set_bit(idx, arc_pmu->used_mask);
210                 hwc->idx = idx;
211         }
212
213         write_aux_reg(ARC_REG_PCT_INDEX, idx);
214         write_aux_reg(ARC_REG_PCT_CONFIG, 0);
215         write_aux_reg(ARC_REG_PCT_COUNTL, 0);
216         write_aux_reg(ARC_REG_PCT_COUNTH, 0);
217         local64_set(&hwc->prev_count, 0);
218
219         hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
220         if (flags & PERF_EF_START)
221                 arc_pmu_start(event, PERF_EF_RELOAD);
222
223         perf_event_update_userpage(event);
224
225         return 0;
226 }
227
228 static int arc_pmu_device_probe(struct platform_device *pdev)
229 {
230         struct arc_pmu *arc_pmu;
231         struct arc_reg_pct_build pct_bcr;
232         struct arc_reg_cc_build cc_bcr;
233         int i, j, ret;
234
235         union cc_name {
236                 struct {
237                         uint32_t word0, word1;
238                         char sentinel;
239                 } indiv;
240                 char str[9];
241         } cc_name;
242
243
244         READ_BCR(ARC_REG_PCT_BUILD, pct_bcr);
245         if (!pct_bcr.v) {
246                 pr_err("This core does not have performance counters!\n");
247                 return -ENODEV;
248         }
249         BUG_ON(pct_bcr.c > ARC_PMU_MAX_HWEVENTS);
250
251         READ_BCR(ARC_REG_CC_BUILD, cc_bcr);
252         if (!cc_bcr.v) {
253                 pr_err("Performance counters exist, but no countable conditions?\n");
254                 return -ENODEV;
255         }
256
257         arc_pmu = devm_kzalloc(&pdev->dev, sizeof(struct arc_pmu), GFP_KERNEL);
258         if (!arc_pmu)
259                 return -ENOMEM;
260
261         arc_pmu->n_counters = pct_bcr.c;
262         arc_pmu->counter_size = 32 + (pct_bcr.s << 4);
263
264         pr_info("ARC perf\t: %d counters (%d bits), %d countable conditions\n",
265                 arc_pmu->n_counters, arc_pmu->counter_size, cc_bcr.c);
266
267         cc_name.str[8] = 0;
268         for (i = 0; i < PERF_COUNT_ARC_HW_MAX; i++)
269                 arc_pmu->ev_hw_idx[i] = -1;
270
271         /* loop thru all available h/w condition indexes */
272         for (j = 0; j < cc_bcr.c; j++) {
273                 write_aux_reg(ARC_REG_CC_INDEX, j);
274                 cc_name.indiv.word0 = read_aux_reg(ARC_REG_CC_NAME0);
275                 cc_name.indiv.word1 = read_aux_reg(ARC_REG_CC_NAME1);
276
277                 /* See if it has been mapped to a perf event_id */
278                 for (i = 0; i < ARRAY_SIZE(arc_pmu_ev_hw_map); i++) {
279                         if (arc_pmu_ev_hw_map[i] &&
280                             !strcmp(arc_pmu_ev_hw_map[i], cc_name.str) &&
281                             strlen(arc_pmu_ev_hw_map[i])) {
282                                 pr_debug("mapping perf event %2d to h/w event \'%8s\' (idx %d)\n",
283                                          i, cc_name.str, j);
284                                 arc_pmu->ev_hw_idx[i] = j;
285                         }
286                 }
287         }
288
289         arc_pmu->pmu = (struct pmu) {
290                 .pmu_enable     = arc_pmu_enable,
291                 .pmu_disable    = arc_pmu_disable,
292                 .event_init     = arc_pmu_event_init,
293                 .add            = arc_pmu_add,
294                 .del            = arc_pmu_del,
295                 .start          = arc_pmu_start,
296                 .stop           = arc_pmu_stop,
297                 .read           = arc_pmu_read,
298         };
299
300         /* ARC 700 PMU does not support sampling events */
301         arc_pmu->pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
302
303         ret = perf_pmu_register(&arc_pmu->pmu, pdev->name, PERF_TYPE_RAW);
304
305         return ret;
306 }
307
308 #ifdef CONFIG_OF
309 static const struct of_device_id arc_pmu_match[] = {
310         { .compatible = "snps,arc700-pmu" },
311         {},
312 };
313 MODULE_DEVICE_TABLE(of, arc_pmu_match);
314 #endif
315
316 static struct platform_driver arc_pmu_driver = {
317         .driver = {
318                 .name           = "arc700-pmu",
319                 .of_match_table = of_match_ptr(arc_pmu_match),
320         },
321         .probe          = arc_pmu_device_probe,
322 };
323
324 module_platform_driver(arc_pmu_driver);
325
326 MODULE_LICENSE("GPL");
327 MODULE_AUTHOR("Mischa Jonker <mjonker@synopsys.com>");
328 MODULE_DESCRIPTION("ARC PMU driver");