Commit | Line | Data |
---|---|---|
ac44e669 PM |
1 | /* |
2 | * Performance event support framework for SuperH hardware counters. | |
3 | * | |
4 | * Copyright (C) 2009 Paul Mundt | |
5 | * | |
6 | * Heavily based on the x86 and PowerPC implementations. | |
7 | * | |
8 | * x86: | |
9 | * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de> | |
10 | * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar | |
11 | * Copyright (C) 2009 Jaswinder Singh Rajput | |
12 | * Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter | |
13 | * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> | |
14 | * Copyright (C) 2009 Intel Corporation, <markus.t.metzger@intel.com> | |
15 | * | |
16 | * ppc: | |
17 | * Copyright 2008-2009 Paul Mackerras, IBM Corporation. | |
18 | * | |
19 | * This file is subject to the terms and conditions of the GNU General Public | |
20 | * License. See the file "COPYING" in the main directory of this archive | |
21 | * for more details. | |
22 | */ | |
23 | #include <linux/kernel.h> | |
24 | #include <linux/init.h> | |
25 | #include <linux/io.h> | |
26 | #include <linux/irq.h> | |
27 | #include <linux/perf_event.h> | |
28 | #include <asm/processor.h> | |
29 | ||
30 | struct cpu_hw_events { | |
31 | struct perf_event *events[MAX_HWEVENTS]; | |
32 | unsigned long used_mask[BITS_TO_LONGS(MAX_HWEVENTS)]; | |
33 | unsigned long active_mask[BITS_TO_LONGS(MAX_HWEVENTS)]; | |
34 | }; | |
35 | ||
36 | DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events); | |
37 | ||
38 | static struct sh_pmu *sh_pmu __read_mostly; | |
39 | ||
40 | /* Number of perf_events counting hardware events */ | |
41 | static atomic_t num_events; | |
42 | /* Used to avoid races in calling reserve/release_pmc_hardware */ | |
43 | static DEFINE_MUTEX(pmc_reserve_mutex); | |
44 | ||
45 | /* | |
46 | * Stub these out for now, do something more profound later. | |
47 | */ | |
48 | int reserve_pmc_hardware(void) | |
49 | { | |
50 | return 0; | |
51 | } | |
52 | ||
53 | void release_pmc_hardware(void) | |
54 | { | |
55 | } | |
56 | ||
57 | static inline int sh_pmu_initialized(void) | |
58 | { | |
59 | return !!sh_pmu; | |
60 | } | |
61 | ||
62 | /* | |
63 | * Release the PMU if this is the last perf_event. | |
64 | */ | |
65 | static void hw_perf_event_destroy(struct perf_event *event) | |
66 | { | |
67 | if (!atomic_add_unless(&num_events, -1, 1)) { | |
68 | mutex_lock(&pmc_reserve_mutex); | |
69 | if (atomic_dec_return(&num_events) == 0) | |
70 | release_pmc_hardware(); | |
71 | mutex_unlock(&pmc_reserve_mutex); | |
72 | } | |
73 | } | |
74 | ||
75 | static int hw_perf_cache_event(int config, int *evp) | |
76 | { | |
77 | unsigned long type, op, result; | |
78 | int ev; | |
79 | ||
80 | if (!sh_pmu->cache_events) | |
81 | return -EINVAL; | |
82 | ||
83 | /* unpack config */ | |
84 | type = config & 0xff; | |
85 | op = (config >> 8) & 0xff; | |
86 | result = (config >> 16) & 0xff; | |
87 | ||
88 | if (type >= PERF_COUNT_HW_CACHE_MAX || | |
89 | op >= PERF_COUNT_HW_CACHE_OP_MAX || | |
90 | result >= PERF_COUNT_HW_CACHE_RESULT_MAX) | |
91 | return -EINVAL; | |
92 | ||
93 | ev = (*sh_pmu->cache_events)[type][op][result]; | |
94 | if (ev == 0) | |
95 | return -EOPNOTSUPP; | |
96 | if (ev == -1) | |
97 | return -EINVAL; | |
98 | *evp = ev; | |
99 | return 0; | |
100 | } | |
101 | ||
102 | static int __hw_perf_event_init(struct perf_event *event) | |
103 | { | |
104 | struct perf_event_attr *attr = &event->attr; | |
105 | struct hw_perf_event *hwc = &event->hw; | |
8820002c | 106 | int config = -1; |
ac44e669 PM |
107 | int err; |
108 | ||
109 | if (!sh_pmu_initialized()) | |
110 | return -ENODEV; | |
111 | ||
112 | /* | |
113 | * All of the on-chip counters are "limited", in that they have | |
114 | * no interrupts, and are therefore unable to do sampling without | |
115 | * further work and timer assistance. | |
116 | */ | |
117 | if (hwc->sample_period) | |
118 | return -EINVAL; | |
119 | ||
120 | /* | |
121 | * See if we need to reserve the counter. | |
122 | * | |
123 | * If no events are currently in use, then we have to take a | |
124 | * mutex to ensure that we don't race with another task doing | |
125 | * reserve_pmc_hardware or release_pmc_hardware. | |
126 | */ | |
127 | err = 0; | |
128 | if (!atomic_inc_not_zero(&num_events)) { | |
129 | mutex_lock(&pmc_reserve_mutex); | |
130 | if (atomic_read(&num_events) == 0 && | |
131 | reserve_pmc_hardware()) | |
132 | err = -EBUSY; | |
133 | else | |
134 | atomic_inc(&num_events); | |
135 | mutex_unlock(&pmc_reserve_mutex); | |
136 | } | |
137 | ||
138 | if (err) | |
139 | return err; | |
140 | ||
141 | event->destroy = hw_perf_event_destroy; | |
142 | ||
143 | switch (attr->type) { | |
144 | case PERF_TYPE_RAW: | |
145 | config = attr->config & sh_pmu->raw_event_mask; | |
146 | break; | |
147 | case PERF_TYPE_HW_CACHE: | |
148 | err = hw_perf_cache_event(attr->config, &config); | |
149 | if (err) | |
150 | return err; | |
151 | break; | |
152 | case PERF_TYPE_HARDWARE: | |
153 | if (attr->config >= sh_pmu->max_events) | |
154 | return -EINVAL; | |
155 | ||
156 | config = sh_pmu->event_map(attr->config); | |
157 | break; | |
ac44e669 PM |
158 | } |
159 | ||
160 | if (config == -1) | |
161 | return -EINVAL; | |
162 | ||
163 | hwc->config |= config; | |
164 | ||
165 | return 0; | |
166 | } | |
167 | ||
168 | static void sh_perf_event_update(struct perf_event *event, | |
169 | struct hw_perf_event *hwc, int idx) | |
170 | { | |
171 | u64 prev_raw_count, new_raw_count; | |
172 | s64 delta; | |
173 | int shift = 0; | |
174 | ||
175 | /* | |
176 | * Depending on the counter configuration, they may or may not | |
177 | * be chained, in which case the previous counter value can be | |
178 | * updated underneath us if the lower-half overflows. | |
179 | * | |
180 | * Our tactic to handle this is to first atomically read and | |
181 | * exchange a new raw count - then add that new-prev delta | |
182 | * count to the generic counter atomically. | |
183 | * | |
184 | * As there is no interrupt associated with the overflow events, | |
185 | * this is the simplest approach for maintaining consistency. | |
186 | */ | |
187 | again: | |
188 | prev_raw_count = atomic64_read(&hwc->prev_count); | |
189 | new_raw_count = sh_pmu->read(idx); | |
190 | ||
191 | if (atomic64_cmpxchg(&hwc->prev_count, prev_raw_count, | |
192 | new_raw_count) != prev_raw_count) | |
193 | goto again; | |
194 | ||
195 | /* | |
196 | * Now we have the new raw value and have updated the prev | |
197 | * timestamp already. We can now calculate the elapsed delta | |
198 | * (counter-)time and add that to the generic counter. | |
199 | * | |
200 | * Careful, not all hw sign-extends above the physical width | |
201 | * of the count. | |
202 | */ | |
203 | delta = (new_raw_count << shift) - (prev_raw_count << shift); | |
204 | delta >>= shift; | |
205 | ||
206 | atomic64_add(delta, &event->count); | |
207 | } | |
208 | ||
209 | static void sh_pmu_disable(struct perf_event *event) | |
210 | { | |
211 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); | |
212 | struct hw_perf_event *hwc = &event->hw; | |
213 | int idx = hwc->idx; | |
214 | ||
215 | clear_bit(idx, cpuc->active_mask); | |
216 | sh_pmu->disable(hwc, idx); | |
217 | ||
218 | barrier(); | |
219 | ||
220 | sh_perf_event_update(event, &event->hw, idx); | |
221 | ||
222 | cpuc->events[idx] = NULL; | |
223 | clear_bit(idx, cpuc->used_mask); | |
224 | ||
225 | perf_event_update_userpage(event); | |
226 | } | |
227 | ||
228 | static int sh_pmu_enable(struct perf_event *event) | |
229 | { | |
230 | struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events); | |
231 | struct hw_perf_event *hwc = &event->hw; | |
232 | int idx = hwc->idx; | |
233 | ||
234 | if (test_and_set_bit(idx, cpuc->used_mask)) { | |
235 | idx = find_first_zero_bit(cpuc->used_mask, sh_pmu->num_events); | |
236 | if (idx == sh_pmu->num_events) | |
237 | return -EAGAIN; | |
238 | ||
239 | set_bit(idx, cpuc->used_mask); | |
240 | hwc->idx = idx; | |
241 | } | |
242 | ||
243 | sh_pmu->disable(hwc, idx); | |
244 | ||
245 | cpuc->events[idx] = event; | |
246 | set_bit(idx, cpuc->active_mask); | |
247 | ||
248 | sh_pmu->enable(hwc, idx); | |
249 | ||
250 | perf_event_update_userpage(event); | |
251 | ||
252 | return 0; | |
253 | } | |
254 | ||
255 | static void sh_pmu_read(struct perf_event *event) | |
256 | { | |
257 | sh_perf_event_update(event, &event->hw, event->hw.idx); | |
258 | } | |
259 | ||
260 | static const struct pmu pmu = { | |
261 | .enable = sh_pmu_enable, | |
262 | .disable = sh_pmu_disable, | |
263 | .read = sh_pmu_read, | |
264 | }; | |
265 | ||
266 | const struct pmu *hw_perf_event_init(struct perf_event *event) | |
267 | { | |
268 | int err = __hw_perf_event_init(event); | |
269 | if (unlikely(err)) { | |
270 | if (event->destroy) | |
271 | event->destroy(event); | |
272 | return ERR_PTR(err); | |
273 | } | |
274 | ||
275 | return &pmu; | |
276 | } | |
277 | ||
3f6da390 | 278 | static void sh_pmu_setup(int cpu) |
ac44e669 PM |
279 | { |
280 | struct cpu_hw_events *cpuhw = &per_cpu(cpu_hw_events, cpu); | |
281 | ||
282 | memset(cpuhw, 0, sizeof(struct cpu_hw_events)); | |
283 | } | |
284 | ||
3f6da390 PZ |
285 | static int __cpuinit |
286 | sh_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu) | |
287 | { | |
288 | unsigned int cpu = (long)hcpu; | |
289 | ||
290 | switch (action & ~CPU_TASKS_FROZEN) { | |
291 | case CPU_UP_PREPARE: | |
292 | sh_pmu_setup(cpu); | |
293 | break; | |
294 | ||
295 | default: | |
296 | break; | |
297 | } | |
298 | ||
299 | return NOTIFY_OK; | |
300 | } | |
301 | ||
ac44e669 PM |
302 | void hw_perf_enable(void) |
303 | { | |
304 | if (!sh_pmu_initialized()) | |
305 | return; | |
306 | ||
307 | sh_pmu->enable_all(); | |
308 | } | |
309 | ||
310 | void hw_perf_disable(void) | |
311 | { | |
312 | if (!sh_pmu_initialized()) | |
313 | return; | |
314 | ||
315 | sh_pmu->disable_all(); | |
316 | } | |
317 | ||
90851c40 | 318 | int __cpuinit register_sh_pmu(struct sh_pmu *pmu) |
ac44e669 PM |
319 | { |
320 | if (sh_pmu) | |
321 | return -EBUSY; | |
322 | sh_pmu = pmu; | |
323 | ||
324 | pr_info("Performance Events: %s support registered\n", pmu->name); | |
325 | ||
1d317f90 | 326 | WARN_ON(pmu->num_events > MAX_HWEVENTS); |
ac44e669 | 327 | |
3f6da390 | 328 | perf_cpu_notifier(sh_pmu_notifier); |
ac44e669 PM |
329 | return 0; |
330 | } |