Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 1 | /* |
| 2 | * Performance events ring-buffer code: |
| 3 | * |
| 4 | * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de> |
| 5 | * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar |
Peter Zijlstra | 90eec10 | 2015-11-16 11:08:45 +0100 | [diff] [blame] | 6 | * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra |
Al Viro | d36b691 | 2011-12-29 17:09:01 -0500 | [diff] [blame] | 7 | * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com> |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 8 | * |
| 9 | * For licensing details see kernel-base/COPYING |
| 10 | */ |
| 11 | |
| 12 | #include <linux/perf_event.h> |
| 13 | #include <linux/vmalloc.h> |
| 14 | #include <linux/slab.h> |
Peter Zijlstra | 26c86da | 2013-10-31 10:19:59 +0100 | [diff] [blame] | 15 | #include <linux/circ_buf.h> |
Jiri Olsa | 7c60fc0 | 2015-01-28 18:54:38 +0100 | [diff] [blame] | 16 | #include <linux/poll.h> |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 17 | |
| 18 | #include "internal.h" |
| 19 | |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 20 | static void perf_output_wakeup(struct perf_output_handle *handle) |
| 21 | { |
Linus Torvalds | a9a0884 | 2018-02-11 14:34:03 -0800 | [diff] [blame^] | 22 | atomic_set(&handle->rb->poll, EPOLLIN); |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 23 | |
Peter Zijlstra | a8b0ca1 | 2011-06-27 14:41:57 +0200 | [diff] [blame] | 24 | handle->event->pending_wakeup = 1; |
| 25 | irq_work_queue(&handle->event->pending); |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 26 | } |
| 27 | |
| 28 | /* |
| 29 | * We need to ensure a later event_id doesn't publish a head when a former |
| 30 | * event isn't done writing. However since we need to deal with NMIs we |
| 31 | * cannot fully serialize things. |
| 32 | * |
| 33 | * We only publish the head (and generate a wakeup) when the outer-most |
| 34 | * event completes. |
| 35 | */ |
| 36 | static void perf_output_get_handle(struct perf_output_handle *handle) |
| 37 | { |
| 38 | struct ring_buffer *rb = handle->rb; |
| 39 | |
| 40 | preempt_disable(); |
| 41 | local_inc(&rb->nest); |
| 42 | handle->wakeup = local_read(&rb->wakeup); |
| 43 | } |
| 44 | |
| 45 | static void perf_output_put_handle(struct perf_output_handle *handle) |
| 46 | { |
| 47 | struct ring_buffer *rb = handle->rb; |
| 48 | unsigned long head; |
| 49 | |
| 50 | again: |
| 51 | head = local_read(&rb->head); |
| 52 | |
| 53 | /* |
| 54 | * IRQ/NMI can happen here, which means we can miss a head update. |
| 55 | */ |
| 56 | |
| 57 | if (!local_dec_and_test(&rb->nest)) |
| 58 | goto out; |
| 59 | |
| 60 | /* |
Peter Zijlstra | bf378d3 | 2013-10-28 13:55:29 +0100 | [diff] [blame] | 61 | * Since the mmap() consumer (userspace) can run on a different CPU: |
| 62 | * |
| 63 | * kernel user |
| 64 | * |
Peter Zijlstra | c7f2e3c | 2013-11-25 11:49:10 +0100 | [diff] [blame] | 65 | * if (LOAD ->data_tail) { LOAD ->data_head |
| 66 | * (A) smp_rmb() (C) |
| 67 | * STORE $data LOAD $data |
| 68 | * smp_wmb() (B) smp_mb() (D) |
| 69 | * STORE ->data_head STORE ->data_tail |
| 70 | * } |
Peter Zijlstra | bf378d3 | 2013-10-28 13:55:29 +0100 | [diff] [blame] | 71 | * |
| 72 | * Where A pairs with D, and B pairs with C. |
| 73 | * |
Peter Zijlstra | c7f2e3c | 2013-11-25 11:49:10 +0100 | [diff] [blame] | 74 | * In our case (A) is a control dependency that separates the load of |
| 75 | * the ->data_tail and the stores of $data. In case ->data_tail |
| 76 | * indicates there is no room in the buffer to store $data we do not. |
Peter Zijlstra | bf378d3 | 2013-10-28 13:55:29 +0100 | [diff] [blame] | 77 | * |
Peter Zijlstra | c7f2e3c | 2013-11-25 11:49:10 +0100 | [diff] [blame] | 78 | * D needs to be a full barrier since it separates the data READ |
Peter Zijlstra | bf378d3 | 2013-10-28 13:55:29 +0100 | [diff] [blame] | 79 | * from the tail WRITE. |
| 80 | * |
| 81 | * For B a WMB is sufficient since it separates two WRITEs, and for C |
| 82 | * an RMB is sufficient since it separates two READs. |
| 83 | * |
| 84 | * See perf_output_begin(). |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 85 | */ |
Peter Zijlstra | c7f2e3c | 2013-11-25 11:49:10 +0100 | [diff] [blame] | 86 | smp_wmb(); /* B, matches C */ |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 87 | rb->user_page->data_head = head; |
| 88 | |
| 89 | /* |
Peter Zijlstra | 394570b | 2013-10-31 17:41:23 +0100 | [diff] [blame] | 90 | * Now check if we missed an update -- rely on previous implied |
| 91 | * compiler barriers to force a re-read. |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 92 | */ |
| 93 | if (unlikely(head != local_read(&rb->head))) { |
| 94 | local_inc(&rb->nest); |
| 95 | goto again; |
| 96 | } |
| 97 | |
| 98 | if (handle->wakeup != local_read(&rb->wakeup)) |
| 99 | perf_output_wakeup(handle); |
| 100 | |
| 101 | out: |
| 102 | preempt_enable(); |
| 103 | } |
| 104 | |
Wang Nan | d1b26c7 | 2016-03-28 06:41:31 +0000 | [diff] [blame] | 105 | static bool __always_inline |
| 106 | ring_buffer_has_space(unsigned long head, unsigned long tail, |
| 107 | unsigned long data_size, unsigned int size, |
| 108 | bool backward) |
| 109 | { |
| 110 | if (!backward) |
| 111 | return CIRC_SPACE(head, tail, data_size) >= size; |
| 112 | else |
| 113 | return CIRC_SPACE(tail, head, data_size) >= size; |
| 114 | } |
| 115 | |
| 116 | static int __always_inline |
| 117 | __perf_output_begin(struct perf_output_handle *handle, |
| 118 | struct perf_event *event, unsigned int size, |
| 119 | bool backward) |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 120 | { |
| 121 | struct ring_buffer *rb; |
| 122 | unsigned long tail, offset, head; |
Peter Zijlstra | 524feca | 2013-10-31 17:36:25 +0100 | [diff] [blame] | 123 | int have_lost, page_shift; |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 124 | struct { |
| 125 | struct perf_event_header header; |
| 126 | u64 id; |
| 127 | u64 lost; |
| 128 | } lost_event; |
| 129 | |
| 130 | rcu_read_lock(); |
| 131 | /* |
| 132 | * For inherited events we send all the output towards the parent. |
| 133 | */ |
| 134 | if (event->parent) |
| 135 | event = event->parent; |
| 136 | |
| 137 | rb = rcu_dereference(event->rb); |
Peter Zijlstra | c72b42a | 2013-10-31 17:20:25 +0100 | [diff] [blame] | 138 | if (unlikely(!rb)) |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 139 | goto out; |
| 140 | |
Wang Nan | 86e7972 | 2016-03-28 06:41:29 +0000 | [diff] [blame] | 141 | if (unlikely(rb->paused)) { |
| 142 | if (rb->nr_pages) |
| 143 | local_inc(&rb->lost); |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 144 | goto out; |
Wang Nan | 86e7972 | 2016-03-28 06:41:29 +0000 | [diff] [blame] | 145 | } |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 146 | |
Peter Zijlstra | c72b42a | 2013-10-31 17:20:25 +0100 | [diff] [blame] | 147 | handle->rb = rb; |
| 148 | handle->event = event; |
| 149 | |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 150 | have_lost = local_read(&rb->lost); |
Peter Zijlstra | c72b42a | 2013-10-31 17:20:25 +0100 | [diff] [blame] | 151 | if (unlikely(have_lost)) { |
Peter Zijlstra | d20a973 | 2013-10-31 17:29:29 +0100 | [diff] [blame] | 152 | size += sizeof(lost_event); |
| 153 | if (event->attr.sample_id_all) |
| 154 | size += event->id_header_size; |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 155 | } |
| 156 | |
| 157 | perf_output_get_handle(handle); |
| 158 | |
| 159 | do { |
Linus Torvalds | 105ff3c | 2015-11-03 17:22:17 -0800 | [diff] [blame] | 160 | tail = READ_ONCE(rb->user_page->data_tail); |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 161 | offset = head = local_read(&rb->head); |
Wang Nan | d1b26c7 | 2016-03-28 06:41:31 +0000 | [diff] [blame] | 162 | if (!rb->overwrite) { |
| 163 | if (unlikely(!ring_buffer_has_space(head, tail, |
| 164 | perf_data_size(rb), |
| 165 | size, backward))) |
| 166 | goto fail; |
| 167 | } |
Peter Zijlstra | c7f2e3c | 2013-11-25 11:49:10 +0100 | [diff] [blame] | 168 | |
| 169 | /* |
| 170 | * The above forms a control dependency barrier separating the |
| 171 | * @tail load above from the data stores below. Since the @tail |
| 172 | * load is required to compute the branch to fail below. |
| 173 | * |
| 174 | * A, matches D; the full memory barrier userspace SHOULD issue |
| 175 | * after reading the data and before storing the new tail |
| 176 | * position. |
| 177 | * |
| 178 | * See perf_output_put_handle(). |
| 179 | */ |
| 180 | |
Wang Nan | d1b26c7 | 2016-03-28 06:41:31 +0000 | [diff] [blame] | 181 | if (!backward) |
| 182 | head += size; |
| 183 | else |
| 184 | head -= size; |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 185 | } while (local_cmpxchg(&rb->head, offset, head) != offset); |
| 186 | |
Wang Nan | d1b26c7 | 2016-03-28 06:41:31 +0000 | [diff] [blame] | 187 | if (backward) { |
| 188 | offset = head; |
| 189 | head = (u64)(-head); |
| 190 | } |
| 191 | |
Peter Zijlstra | 85f59ed | 2013-10-31 17:25:38 +0100 | [diff] [blame] | 192 | /* |
Peter Zijlstra | c7f2e3c | 2013-11-25 11:49:10 +0100 | [diff] [blame] | 193 | * We rely on the implied barrier() by local_cmpxchg() to ensure |
| 194 | * none of the data stores below can be lifted up by the compiler. |
Peter Zijlstra | 85f59ed | 2013-10-31 17:25:38 +0100 | [diff] [blame] | 195 | */ |
Peter Zijlstra | 85f59ed | 2013-10-31 17:25:38 +0100 | [diff] [blame] | 196 | |
Peter Zijlstra | c72b42a | 2013-10-31 17:20:25 +0100 | [diff] [blame] | 197 | if (unlikely(head - local_read(&rb->wakeup) > rb->watermark)) |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 198 | local_add(rb->watermark, &rb->wakeup); |
| 199 | |
Peter Zijlstra | 524feca | 2013-10-31 17:36:25 +0100 | [diff] [blame] | 200 | page_shift = PAGE_SHIFT + page_order(rb); |
| 201 | |
| 202 | handle->page = (offset >> page_shift) & (rb->nr_pages - 1); |
| 203 | offset &= (1UL << page_shift) - 1; |
| 204 | handle->addr = rb->data_pages[handle->page] + offset; |
| 205 | handle->size = (1UL << page_shift) - offset; |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 206 | |
Peter Zijlstra | c72b42a | 2013-10-31 17:20:25 +0100 | [diff] [blame] | 207 | if (unlikely(have_lost)) { |
Peter Zijlstra | d20a973 | 2013-10-31 17:29:29 +0100 | [diff] [blame] | 208 | struct perf_sample_data sample_data; |
| 209 | |
| 210 | lost_event.header.size = sizeof(lost_event); |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 211 | lost_event.header.type = PERF_RECORD_LOST; |
| 212 | lost_event.header.misc = 0; |
| 213 | lost_event.id = event->id; |
| 214 | lost_event.lost = local_xchg(&rb->lost, 0); |
| 215 | |
Peter Zijlstra | d20a973 | 2013-10-31 17:29:29 +0100 | [diff] [blame] | 216 | perf_event_header__init_id(&lost_event.header, |
| 217 | &sample_data, event); |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 218 | perf_output_put(handle, lost_event); |
| 219 | perf_event__output_id_sample(event, handle, &sample_data); |
| 220 | } |
| 221 | |
| 222 | return 0; |
| 223 | |
| 224 | fail: |
| 225 | local_inc(&rb->lost); |
| 226 | perf_output_put_handle(handle); |
| 227 | out: |
| 228 | rcu_read_unlock(); |
| 229 | |
| 230 | return -ENOSPC; |
| 231 | } |
| 232 | |
Wang Nan | 9ecda41 | 2016-04-05 14:11:18 +0000 | [diff] [blame] | 233 | int perf_output_begin_forward(struct perf_output_handle *handle, |
| 234 | struct perf_event *event, unsigned int size) |
| 235 | { |
| 236 | return __perf_output_begin(handle, event, size, false); |
| 237 | } |
| 238 | |
| 239 | int perf_output_begin_backward(struct perf_output_handle *handle, |
| 240 | struct perf_event *event, unsigned int size) |
| 241 | { |
| 242 | return __perf_output_begin(handle, event, size, true); |
| 243 | } |
| 244 | |
Wang Nan | d1b26c7 | 2016-03-28 06:41:31 +0000 | [diff] [blame] | 245 | int perf_output_begin(struct perf_output_handle *handle, |
| 246 | struct perf_event *event, unsigned int size) |
| 247 | { |
Wang Nan | 9ecda41 | 2016-04-05 14:11:18 +0000 | [diff] [blame] | 248 | |
| 249 | return __perf_output_begin(handle, event, size, |
| 250 | unlikely(is_write_backward(event))); |
Wang Nan | d1b26c7 | 2016-03-28 06:41:31 +0000 | [diff] [blame] | 251 | } |
| 252 | |
Frederic Weisbecker | 91d7753 | 2012-08-07 15:20:38 +0200 | [diff] [blame] | 253 | unsigned int perf_output_copy(struct perf_output_handle *handle, |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 254 | const void *buf, unsigned int len) |
| 255 | { |
Frederic Weisbecker | 91d7753 | 2012-08-07 15:20:38 +0200 | [diff] [blame] | 256 | return __output_copy(handle, buf, len); |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 257 | } |
| 258 | |
Jiri Olsa | 5685e0f | 2012-08-07 15:20:39 +0200 | [diff] [blame] | 259 | unsigned int perf_output_skip(struct perf_output_handle *handle, |
| 260 | unsigned int len) |
| 261 | { |
| 262 | return __output_skip(handle, NULL, len); |
| 263 | } |
| 264 | |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 265 | void perf_output_end(struct perf_output_handle *handle) |
| 266 | { |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 267 | perf_output_put_handle(handle); |
| 268 | rcu_read_unlock(); |
| 269 | } |
| 270 | |
| 271 | static void |
| 272 | ring_buffer_init(struct ring_buffer *rb, long watermark, int flags) |
| 273 | { |
| 274 | long max_size = perf_data_size(rb); |
| 275 | |
| 276 | if (watermark) |
| 277 | rb->watermark = min(max_size, watermark); |
| 278 | |
| 279 | if (!rb->watermark) |
| 280 | rb->watermark = max_size / 2; |
| 281 | |
| 282 | if (flags & RING_BUFFER_WRITABLE) |
Stephane Eranian | dd9c086 | 2013-03-18 14:33:28 +0100 | [diff] [blame] | 283 | rb->overwrite = 0; |
| 284 | else |
| 285 | rb->overwrite = 1; |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 286 | |
| 287 | atomic_set(&rb->refcount, 1); |
Peter Zijlstra | 10c6db1 | 2011-11-26 02:47:31 +0100 | [diff] [blame] | 288 | |
| 289 | INIT_LIST_HEAD(&rb->event_list); |
| 290 | spin_lock_init(&rb->event_lock); |
Wang Nan | 86e7972 | 2016-03-28 06:41:29 +0000 | [diff] [blame] | 291 | |
| 292 | /* |
| 293 | * perf_output_begin() only checks rb->paused, therefore |
| 294 | * rb->paused must be true if we have no pages for output. |
| 295 | */ |
| 296 | if (!rb->nr_pages) |
| 297 | rb->paused = 1; |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 298 | } |
| 299 | |
Will Deacon | f4c0b0a | 2017-02-20 15:33:50 +0200 | [diff] [blame] | 300 | void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags) |
| 301 | { |
| 302 | /* |
| 303 | * OVERWRITE is determined by perf_aux_output_end() and can't |
| 304 | * be passed in directly. |
| 305 | */ |
| 306 | if (WARN_ON_ONCE(flags & PERF_AUX_FLAG_OVERWRITE)) |
| 307 | return; |
| 308 | |
| 309 | handle->aux_flags |= flags; |
| 310 | } |
| 311 | EXPORT_SYMBOL_GPL(perf_aux_output_flag); |
| 312 | |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 313 | /* |
| 314 | * This is called before hardware starts writing to the AUX area to |
| 315 | * obtain an output handle and make sure there's room in the buffer. |
| 316 | * When the capture completes, call perf_aux_output_end() to commit |
| 317 | * the recorded data to the buffer. |
| 318 | * |
| 319 | * The ordering is similar to that of perf_output_{begin,end}, with |
| 320 | * the exception of (B), which should be taken care of by the pmu |
| 321 | * driver, since ordering rules will differ depending on hardware. |
Alexander Shishkin | af5bb4e | 2016-03-04 15:42:47 +0200 | [diff] [blame] | 322 | * |
| 323 | * Call this from pmu::start(); see the comment in perf_aux_output_end() |
| 324 | * about its use in pmu callbacks. Both can also be called from the PMI |
| 325 | * handler if needed. |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 326 | */ |
| 327 | void *perf_aux_output_begin(struct perf_output_handle *handle, |
| 328 | struct perf_event *event) |
| 329 | { |
| 330 | struct perf_event *output_event = event; |
| 331 | unsigned long aux_head, aux_tail; |
| 332 | struct ring_buffer *rb; |
| 333 | |
| 334 | if (output_event->parent) |
| 335 | output_event = output_event->parent; |
| 336 | |
| 337 | /* |
| 338 | * Since this will typically be open across pmu::add/pmu::del, we |
| 339 | * grab ring_buffer's refcount instead of holding rcu read lock |
| 340 | * to make sure it doesn't disappear under us. |
| 341 | */ |
| 342 | rb = ring_buffer_get(output_event); |
| 343 | if (!rb) |
| 344 | return NULL; |
| 345 | |
Alexander Shishkin | b79ccad | 2016-09-06 16:23:50 +0300 | [diff] [blame] | 346 | if (!rb_has_aux(rb)) |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 347 | goto err; |
| 348 | |
| 349 | /* |
Alexander Shishkin | b79ccad | 2016-09-06 16:23:50 +0300 | [diff] [blame] | 350 | * If aux_mmap_count is zero, the aux buffer is in perf_mmap_close(), |
| 351 | * about to get freed, so we leave immediately. |
| 352 | * |
| 353 | * Checking rb::aux_mmap_count and rb::refcount has to be done in |
| 354 | * the same order, see perf_mmap_close. Otherwise we end up freeing |
| 355 | * aux pages in this path, which is a bug, because in_atomic(). |
Alexander Shishkin | dcb10a9 | 2016-03-04 15:42:45 +0200 | [diff] [blame] | 356 | */ |
| 357 | if (!atomic_read(&rb->aux_mmap_count)) |
Alexander Shishkin | b79ccad | 2016-09-06 16:23:50 +0300 | [diff] [blame] | 358 | goto err; |
| 359 | |
| 360 | if (!atomic_inc_not_zero(&rb->aux_refcount)) |
| 361 | goto err; |
Alexander Shishkin | dcb10a9 | 2016-03-04 15:42:45 +0200 | [diff] [blame] | 362 | |
| 363 | /* |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 364 | * Nesting is not supported for AUX area, make sure nested |
| 365 | * writers are caught early |
| 366 | */ |
| 367 | if (WARN_ON_ONCE(local_xchg(&rb->aux_nest, 1))) |
| 368 | goto err_put; |
| 369 | |
Will Deacon | 2ab346c | 2017-08-16 17:18:16 +0100 | [diff] [blame] | 370 | aux_head = rb->aux_head; |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 371 | |
| 372 | handle->rb = rb; |
| 373 | handle->event = event; |
| 374 | handle->head = aux_head; |
Alexander Shishkin | 2023a0d | 2015-01-14 14:18:17 +0200 | [diff] [blame] | 375 | handle->size = 0; |
Will Deacon | f4c0b0a | 2017-02-20 15:33:50 +0200 | [diff] [blame] | 376 | handle->aux_flags = 0; |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 377 | |
| 378 | /* |
Alexander Shishkin | 2023a0d | 2015-01-14 14:18:17 +0200 | [diff] [blame] | 379 | * In overwrite mode, AUX data stores do not depend on aux_tail, |
| 380 | * therefore (A) control dependency barrier does not exist. The |
| 381 | * (B) <-> (C) ordering is still observed by the pmu driver. |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 382 | */ |
Alexander Shishkin | 2023a0d | 2015-01-14 14:18:17 +0200 | [diff] [blame] | 383 | if (!rb->aux_overwrite) { |
Mark Rutland | 6aa7de0 | 2017-10-23 14:07:29 -0700 | [diff] [blame] | 384 | aux_tail = READ_ONCE(rb->user_page->aux_tail); |
Will Deacon | 2ab346c | 2017-08-16 17:18:16 +0100 | [diff] [blame] | 385 | handle->wakeup = rb->aux_wakeup + rb->aux_watermark; |
Alexander Shishkin | 2023a0d | 2015-01-14 14:18:17 +0200 | [diff] [blame] | 386 | if (aux_head - aux_tail < perf_aux_size(rb)) |
| 387 | handle->size = CIRC_SPACE(aux_head, aux_tail, perf_aux_size(rb)); |
| 388 | |
| 389 | /* |
| 390 | * handle->size computation depends on aux_tail load; this forms a |
| 391 | * control dependency barrier separating aux_tail load from aux data |
| 392 | * store that will be enabled on successful return |
| 393 | */ |
| 394 | if (!handle->size) { /* A, matches D */ |
| 395 | event->pending_disable = 1; |
| 396 | perf_output_wakeup(handle); |
| 397 | local_set(&rb->aux_nest, 0); |
| 398 | goto err_put; |
| 399 | } |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 400 | } |
| 401 | |
| 402 | return handle->rb->aux_priv; |
| 403 | |
| 404 | err_put: |
Alexander Shishkin | af5bb4e | 2016-03-04 15:42:47 +0200 | [diff] [blame] | 405 | /* can't be last */ |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 406 | rb_free_aux(rb); |
| 407 | |
| 408 | err: |
Alexander Shishkin | 95ff4ca | 2015-12-02 18:41:11 +0200 | [diff] [blame] | 409 | ring_buffer_put(rb); |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 410 | handle->event = NULL; |
| 411 | |
| 412 | return NULL; |
| 413 | } |
Will Deacon | bc1d202 | 2016-08-16 16:53:15 +0100 | [diff] [blame] | 414 | EXPORT_SYMBOL_GPL(perf_aux_output_begin); |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 415 | |
Alexander Shishkin | 441430e | 2017-09-06 19:08:11 +0300 | [diff] [blame] | 416 | static bool __always_inline rb_need_aux_wakeup(struct ring_buffer *rb) |
| 417 | { |
| 418 | if (rb->aux_overwrite) |
| 419 | return false; |
| 420 | |
| 421 | if (rb->aux_head - rb->aux_wakeup >= rb->aux_watermark) { |
| 422 | rb->aux_wakeup = rounddown(rb->aux_head, rb->aux_watermark); |
| 423 | return true; |
| 424 | } |
| 425 | |
| 426 | return false; |
| 427 | } |
| 428 | |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 429 | /* |
| 430 | * Commit the data written by hardware into the ring buffer by adjusting |
| 431 | * aux_head and posting a PERF_RECORD_AUX into the perf buffer. It is the |
| 432 | * pmu driver's responsibility to observe ordering rules of the hardware, |
| 433 | * so that all the data is externally visible before this is called. |
Alexander Shishkin | af5bb4e | 2016-03-04 15:42:47 +0200 | [diff] [blame] | 434 | * |
| 435 | * Note: this has to be called from pmu::stop() callback, as the assumption |
| 436 | * of the AUX buffer management code is that after pmu::stop(), the AUX |
| 437 | * transaction must be stopped and therefore drop the AUX reference count. |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 438 | */ |
Will Deacon | f4c0b0a | 2017-02-20 15:33:50 +0200 | [diff] [blame] | 439 | void perf_aux_output_end(struct perf_output_handle *handle, unsigned long size) |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 440 | { |
Alexander Shishkin | ae0c2d9 | 2017-02-20 15:33:51 +0200 | [diff] [blame] | 441 | bool wakeup = !!(handle->aux_flags & PERF_AUX_FLAG_TRUNCATED); |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 442 | struct ring_buffer *rb = handle->rb; |
Alexander Shishkin | 2023a0d | 2015-01-14 14:18:17 +0200 | [diff] [blame] | 443 | unsigned long aux_head; |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 444 | |
Alexander Shishkin | 2023a0d | 2015-01-14 14:18:17 +0200 | [diff] [blame] | 445 | /* in overwrite mode, driver provides aux_head via handle */ |
| 446 | if (rb->aux_overwrite) { |
Will Deacon | f4c0b0a | 2017-02-20 15:33:50 +0200 | [diff] [blame] | 447 | handle->aux_flags |= PERF_AUX_FLAG_OVERWRITE; |
Alexander Shishkin | 2023a0d | 2015-01-14 14:18:17 +0200 | [diff] [blame] | 448 | |
| 449 | aux_head = handle->head; |
Will Deacon | 2ab346c | 2017-08-16 17:18:16 +0100 | [diff] [blame] | 450 | rb->aux_head = aux_head; |
Alexander Shishkin | 2023a0d | 2015-01-14 14:18:17 +0200 | [diff] [blame] | 451 | } else { |
Will Deacon | f4c0b0a | 2017-02-20 15:33:50 +0200 | [diff] [blame] | 452 | handle->aux_flags &= ~PERF_AUX_FLAG_OVERWRITE; |
| 453 | |
Will Deacon | 2ab346c | 2017-08-16 17:18:16 +0100 | [diff] [blame] | 454 | aux_head = rb->aux_head; |
| 455 | rb->aux_head += size; |
Alexander Shishkin | 2023a0d | 2015-01-14 14:18:17 +0200 | [diff] [blame] | 456 | } |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 457 | |
Will Deacon | f4c0b0a | 2017-02-20 15:33:50 +0200 | [diff] [blame] | 458 | if (size || handle->aux_flags) { |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 459 | /* |
| 460 | * Only send RECORD_AUX if we have something useful to communicate |
| 461 | */ |
| 462 | |
Will Deacon | f4c0b0a | 2017-02-20 15:33:50 +0200 | [diff] [blame] | 463 | perf_event_aux_event(handle->event, aux_head, size, |
| 464 | handle->aux_flags); |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 465 | } |
| 466 | |
Will Deacon | 2ab346c | 2017-08-16 17:18:16 +0100 | [diff] [blame] | 467 | rb->user_page->aux_head = rb->aux_head; |
Alexander Shishkin | 441430e | 2017-09-06 19:08:11 +0300 | [diff] [blame] | 468 | if (rb_need_aux_wakeup(rb)) |
Alexander Shishkin | 3f56e68 | 2016-05-10 16:18:33 +0300 | [diff] [blame] | 469 | wakeup = true; |
Alexander Shishkin | 3f56e68 | 2016-05-10 16:18:33 +0300 | [diff] [blame] | 470 | |
| 471 | if (wakeup) { |
Will Deacon | f4c0b0a | 2017-02-20 15:33:50 +0200 | [diff] [blame] | 472 | if (handle->aux_flags & PERF_AUX_FLAG_TRUNCATED) |
Alexander Shishkin | 3f56e68 | 2016-05-10 16:18:33 +0300 | [diff] [blame] | 473 | handle->event->pending_disable = 1; |
| 474 | perf_output_wakeup(handle); |
| 475 | } |
| 476 | |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 477 | handle->event = NULL; |
| 478 | |
| 479 | local_set(&rb->aux_nest, 0); |
Alexander Shishkin | af5bb4e | 2016-03-04 15:42:47 +0200 | [diff] [blame] | 480 | /* can't be last */ |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 481 | rb_free_aux(rb); |
Alexander Shishkin | 95ff4ca | 2015-12-02 18:41:11 +0200 | [diff] [blame] | 482 | ring_buffer_put(rb); |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 483 | } |
Will Deacon | bc1d202 | 2016-08-16 16:53:15 +0100 | [diff] [blame] | 484 | EXPORT_SYMBOL_GPL(perf_aux_output_end); |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 485 | |
| 486 | /* |
| 487 | * Skip over a given number of bytes in the AUX buffer, due to, for example, |
| 488 | * hardware's alignment constraints. |
| 489 | */ |
| 490 | int perf_aux_output_skip(struct perf_output_handle *handle, unsigned long size) |
| 491 | { |
| 492 | struct ring_buffer *rb = handle->rb; |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 493 | |
| 494 | if (size > handle->size) |
| 495 | return -ENOSPC; |
| 496 | |
Will Deacon | 2ab346c | 2017-08-16 17:18:16 +0100 | [diff] [blame] | 497 | rb->aux_head += size; |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 498 | |
Will Deacon | 2ab346c | 2017-08-16 17:18:16 +0100 | [diff] [blame] | 499 | rb->user_page->aux_head = rb->aux_head; |
Alexander Shishkin | 441430e | 2017-09-06 19:08:11 +0300 | [diff] [blame] | 500 | if (rb_need_aux_wakeup(rb)) { |
Alexander Shishkin | 1a59413 | 2015-01-14 14:18:18 +0200 | [diff] [blame] | 501 | perf_output_wakeup(handle); |
Will Deacon | 2ab346c | 2017-08-16 17:18:16 +0100 | [diff] [blame] | 502 | handle->wakeup = rb->aux_wakeup + rb->aux_watermark; |
Alexander Shishkin | 1a59413 | 2015-01-14 14:18:18 +0200 | [diff] [blame] | 503 | } |
| 504 | |
Will Deacon | 2ab346c | 2017-08-16 17:18:16 +0100 | [diff] [blame] | 505 | handle->head = rb->aux_head; |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 506 | handle->size -= size; |
| 507 | |
| 508 | return 0; |
| 509 | } |
Will Deacon | bc1d202 | 2016-08-16 16:53:15 +0100 | [diff] [blame] | 510 | EXPORT_SYMBOL_GPL(perf_aux_output_skip); |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 511 | |
| 512 | void *perf_get_aux(struct perf_output_handle *handle) |
| 513 | { |
| 514 | /* this is only valid between perf_aux_output_begin and *_end */ |
| 515 | if (!handle->event) |
| 516 | return NULL; |
| 517 | |
| 518 | return handle->rb->aux_priv; |
| 519 | } |
Will Deacon | bc1d202 | 2016-08-16 16:53:15 +0100 | [diff] [blame] | 520 | EXPORT_SYMBOL_GPL(perf_get_aux); |
Alexander Shishkin | fdc2670 | 2015-01-14 14:18:16 +0200 | [diff] [blame] | 521 | |
Alexander Shishkin | 0a4e38e | 2015-01-14 14:18:12 +0200 | [diff] [blame] | 522 | #define PERF_AUX_GFP (GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_NORETRY) |
| 523 | |
| 524 | static struct page *rb_alloc_aux_page(int node, int order) |
| 525 | { |
| 526 | struct page *page; |
| 527 | |
| 528 | if (order > MAX_ORDER) |
| 529 | order = MAX_ORDER; |
| 530 | |
| 531 | do { |
| 532 | page = alloc_pages_node(node, PERF_AUX_GFP, order); |
| 533 | } while (!page && order--); |
| 534 | |
| 535 | if (page && order) { |
| 536 | /* |
Alexander Shishkin | c2ad6b5 | 2015-07-28 09:00:04 +0300 | [diff] [blame] | 537 | * Communicate the allocation size to the driver: |
| 538 | * if we managed to secure a high-order allocation, |
| 539 | * set its first page's private to this order; |
| 540 | * !PagePrivate(page) means it's just a normal page. |
Alexander Shishkin | 0a4e38e | 2015-01-14 14:18:12 +0200 | [diff] [blame] | 541 | */ |
| 542 | split_page(page, order); |
| 543 | SetPagePrivate(page); |
| 544 | set_page_private(page, order); |
| 545 | } |
| 546 | |
| 547 | return page; |
| 548 | } |
| 549 | |
| 550 | static void rb_free_aux_page(struct ring_buffer *rb, int idx) |
| 551 | { |
| 552 | struct page *page = virt_to_page(rb->aux_pages[idx]); |
| 553 | |
| 554 | ClearPagePrivate(page); |
| 555 | page->mapping = NULL; |
| 556 | __free_page(page); |
| 557 | } |
| 558 | |
Alexander Shishkin | 45c815f | 2016-01-19 17:14:29 +0200 | [diff] [blame] | 559 | static void __rb_free_aux(struct ring_buffer *rb) |
| 560 | { |
| 561 | int pg; |
| 562 | |
Alexander Shishkin | 95ff4ca | 2015-12-02 18:41:11 +0200 | [diff] [blame] | 563 | /* |
| 564 | * Should never happen, the last reference should be dropped from |
| 565 | * perf_mmap_close() path, which first stops aux transactions (which |
| 566 | * in turn are the atomic holders of aux_refcount) and then does the |
| 567 | * last rb_free_aux(). |
| 568 | */ |
| 569 | WARN_ON_ONCE(in_atomic()); |
| 570 | |
Alexander Shishkin | 45c815f | 2016-01-19 17:14:29 +0200 | [diff] [blame] | 571 | if (rb->aux_priv) { |
| 572 | rb->free_aux(rb->aux_priv); |
| 573 | rb->free_aux = NULL; |
| 574 | rb->aux_priv = NULL; |
| 575 | } |
| 576 | |
| 577 | if (rb->aux_nr_pages) { |
| 578 | for (pg = 0; pg < rb->aux_nr_pages; pg++) |
| 579 | rb_free_aux_page(rb, pg); |
| 580 | |
| 581 | kfree(rb->aux_pages); |
| 582 | rb->aux_nr_pages = 0; |
| 583 | } |
| 584 | } |
| 585 | |
Peter Zijlstra | 45bfb2e | 2015-01-14 14:18:11 +0200 | [diff] [blame] | 586 | int rb_alloc_aux(struct ring_buffer *rb, struct perf_event *event, |
Alexander Shishkin | 1a59413 | 2015-01-14 14:18:18 +0200 | [diff] [blame] | 587 | pgoff_t pgoff, int nr_pages, long watermark, int flags) |
Peter Zijlstra | 45bfb2e | 2015-01-14 14:18:11 +0200 | [diff] [blame] | 588 | { |
| 589 | bool overwrite = !(flags & RING_BUFFER_WRITABLE); |
| 590 | int node = (event->cpu == -1) ? -1 : cpu_to_node(event->cpu); |
Alexander Shishkin | 0a4e38e | 2015-01-14 14:18:12 +0200 | [diff] [blame] | 591 | int ret = -ENOMEM, max_order = 0; |
Peter Zijlstra | 45bfb2e | 2015-01-14 14:18:11 +0200 | [diff] [blame] | 592 | |
| 593 | if (!has_aux(event)) |
Hendrik Brueckner | 8a1898d | 2017-06-20 12:26:39 +0200 | [diff] [blame] | 594 | return -EOPNOTSUPP; |
Peter Zijlstra | 45bfb2e | 2015-01-14 14:18:11 +0200 | [diff] [blame] | 595 | |
Alexander Shishkin | 6a27923 | 2015-01-14 14:18:13 +0200 | [diff] [blame] | 596 | if (event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) { |
Alexander Shishkin | 0a4e38e | 2015-01-14 14:18:12 +0200 | [diff] [blame] | 597 | /* |
| 598 | * We need to start with the max_order that fits in nr_pages, |
| 599 | * not the other way around, hence ilog2() and not get_order. |
| 600 | */ |
| 601 | max_order = ilog2(nr_pages); |
| 602 | |
Alexander Shishkin | 6a27923 | 2015-01-14 14:18:13 +0200 | [diff] [blame] | 603 | /* |
| 604 | * PMU requests more than one contiguous chunks of memory |
| 605 | * for SW double buffering |
| 606 | */ |
| 607 | if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_SW_DOUBLEBUF) && |
| 608 | !overwrite) { |
| 609 | if (!max_order) |
| 610 | return -EINVAL; |
| 611 | |
| 612 | max_order--; |
| 613 | } |
| 614 | } |
| 615 | |
Peter Zijlstra | 45bfb2e | 2015-01-14 14:18:11 +0200 | [diff] [blame] | 616 | rb->aux_pages = kzalloc_node(nr_pages * sizeof(void *), GFP_KERNEL, node); |
| 617 | if (!rb->aux_pages) |
| 618 | return -ENOMEM; |
| 619 | |
| 620 | rb->free_aux = event->pmu->free_aux; |
Alexander Shishkin | 0a4e38e | 2015-01-14 14:18:12 +0200 | [diff] [blame] | 621 | for (rb->aux_nr_pages = 0; rb->aux_nr_pages < nr_pages;) { |
Peter Zijlstra | 45bfb2e | 2015-01-14 14:18:11 +0200 | [diff] [blame] | 622 | struct page *page; |
Alexander Shishkin | 0a4e38e | 2015-01-14 14:18:12 +0200 | [diff] [blame] | 623 | int last, order; |
Peter Zijlstra | 45bfb2e | 2015-01-14 14:18:11 +0200 | [diff] [blame] | 624 | |
Alexander Shishkin | 0a4e38e | 2015-01-14 14:18:12 +0200 | [diff] [blame] | 625 | order = min(max_order, ilog2(nr_pages - rb->aux_nr_pages)); |
| 626 | page = rb_alloc_aux_page(node, order); |
Peter Zijlstra | 45bfb2e | 2015-01-14 14:18:11 +0200 | [diff] [blame] | 627 | if (!page) |
| 628 | goto out; |
| 629 | |
Alexander Shishkin | 0a4e38e | 2015-01-14 14:18:12 +0200 | [diff] [blame] | 630 | for (last = rb->aux_nr_pages + (1 << page_private(page)); |
| 631 | last > rb->aux_nr_pages; rb->aux_nr_pages++) |
| 632 | rb->aux_pages[rb->aux_nr_pages] = page_address(page++); |
Peter Zijlstra | 45bfb2e | 2015-01-14 14:18:11 +0200 | [diff] [blame] | 633 | } |
| 634 | |
Alexander Shishkin | aa319bc | 2015-05-22 18:30:20 +0300 | [diff] [blame] | 635 | /* |
| 636 | * In overwrite mode, PMUs that don't support SG may not handle more |
| 637 | * than one contiguous allocation, since they rely on PMI to do double |
| 638 | * buffering. In this case, the entire buffer has to be one contiguous |
| 639 | * chunk. |
| 640 | */ |
| 641 | if ((event->pmu->capabilities & PERF_PMU_CAP_AUX_NO_SG) && |
| 642 | overwrite) { |
| 643 | struct page *page = virt_to_page(rb->aux_pages[0]); |
| 644 | |
| 645 | if (page_private(page) != max_order) |
| 646 | goto out; |
| 647 | } |
| 648 | |
Peter Zijlstra | 45bfb2e | 2015-01-14 14:18:11 +0200 | [diff] [blame] | 649 | rb->aux_priv = event->pmu->setup_aux(event->cpu, rb->aux_pages, nr_pages, |
| 650 | overwrite); |
| 651 | if (!rb->aux_priv) |
| 652 | goto out; |
| 653 | |
| 654 | ret = 0; |
| 655 | |
| 656 | /* |
| 657 | * aux_pages (and pmu driver's private data, aux_priv) will be |
| 658 | * referenced in both producer's and consumer's contexts, thus |
| 659 | * we keep a refcount here to make sure either of the two can |
| 660 | * reference them safely. |
| 661 | */ |
| 662 | atomic_set(&rb->aux_refcount, 1); |
| 663 | |
Alexander Shishkin | 2023a0d | 2015-01-14 14:18:17 +0200 | [diff] [blame] | 664 | rb->aux_overwrite = overwrite; |
Alexander Shishkin | 1a59413 | 2015-01-14 14:18:18 +0200 | [diff] [blame] | 665 | rb->aux_watermark = watermark; |
| 666 | |
| 667 | if (!rb->aux_watermark && !rb->aux_overwrite) |
| 668 | rb->aux_watermark = nr_pages << (PAGE_SHIFT - 1); |
Alexander Shishkin | 2023a0d | 2015-01-14 14:18:17 +0200 | [diff] [blame] | 669 | |
Peter Zijlstra | 45bfb2e | 2015-01-14 14:18:11 +0200 | [diff] [blame] | 670 | out: |
| 671 | if (!ret) |
| 672 | rb->aux_pgoff = pgoff; |
| 673 | else |
Alexander Shishkin | 45c815f | 2016-01-19 17:14:29 +0200 | [diff] [blame] | 674 | __rb_free_aux(rb); |
Peter Zijlstra | 45bfb2e | 2015-01-14 14:18:11 +0200 | [diff] [blame] | 675 | |
| 676 | return ret; |
| 677 | } |
| 678 | |
Peter Zijlstra | 45bfb2e | 2015-01-14 14:18:11 +0200 | [diff] [blame] | 679 | void rb_free_aux(struct ring_buffer *rb) |
| 680 | { |
| 681 | if (atomic_dec_and_test(&rb->aux_refcount)) |
| 682 | __rb_free_aux(rb); |
| 683 | } |
| 684 | |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 685 | #ifndef CONFIG_PERF_USE_VMALLOC |
| 686 | |
| 687 | /* |
| 688 | * Back perf_mmap() with regular GFP_KERNEL-0 pages. |
| 689 | */ |
| 690 | |
Peter Zijlstra | 45bfb2e | 2015-01-14 14:18:11 +0200 | [diff] [blame] | 691 | static struct page * |
| 692 | __perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff) |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 693 | { |
| 694 | if (pgoff > rb->nr_pages) |
| 695 | return NULL; |
| 696 | |
| 697 | if (pgoff == 0) |
| 698 | return virt_to_page(rb->user_page); |
| 699 | |
| 700 | return virt_to_page(rb->data_pages[pgoff - 1]); |
| 701 | } |
| 702 | |
| 703 | static void *perf_mmap_alloc_page(int cpu) |
| 704 | { |
| 705 | struct page *page; |
| 706 | int node; |
| 707 | |
| 708 | node = (cpu == -1) ? cpu : cpu_to_node(cpu); |
| 709 | page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0); |
| 710 | if (!page) |
| 711 | return NULL; |
| 712 | |
| 713 | return page_address(page); |
| 714 | } |
| 715 | |
| 716 | struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags) |
| 717 | { |
| 718 | struct ring_buffer *rb; |
| 719 | unsigned long size; |
| 720 | int i; |
| 721 | |
| 722 | size = sizeof(struct ring_buffer); |
| 723 | size += nr_pages * sizeof(void *); |
| 724 | |
| 725 | rb = kzalloc(size, GFP_KERNEL); |
| 726 | if (!rb) |
| 727 | goto fail; |
| 728 | |
| 729 | rb->user_page = perf_mmap_alloc_page(cpu); |
| 730 | if (!rb->user_page) |
| 731 | goto fail_user_page; |
| 732 | |
| 733 | for (i = 0; i < nr_pages; i++) { |
| 734 | rb->data_pages[i] = perf_mmap_alloc_page(cpu); |
| 735 | if (!rb->data_pages[i]) |
| 736 | goto fail_data_pages; |
| 737 | } |
| 738 | |
| 739 | rb->nr_pages = nr_pages; |
| 740 | |
| 741 | ring_buffer_init(rb, watermark, flags); |
| 742 | |
| 743 | return rb; |
| 744 | |
| 745 | fail_data_pages: |
| 746 | for (i--; i >= 0; i--) |
| 747 | free_page((unsigned long)rb->data_pages[i]); |
| 748 | |
| 749 | free_page((unsigned long)rb->user_page); |
| 750 | |
| 751 | fail_user_page: |
| 752 | kfree(rb); |
| 753 | |
| 754 | fail: |
| 755 | return NULL; |
| 756 | } |
| 757 | |
| 758 | static void perf_mmap_free_page(unsigned long addr) |
| 759 | { |
| 760 | struct page *page = virt_to_page((void *)addr); |
| 761 | |
| 762 | page->mapping = NULL; |
| 763 | __free_page(page); |
| 764 | } |
| 765 | |
| 766 | void rb_free(struct ring_buffer *rb) |
| 767 | { |
| 768 | int i; |
| 769 | |
| 770 | perf_mmap_free_page((unsigned long)rb->user_page); |
| 771 | for (i = 0; i < rb->nr_pages; i++) |
| 772 | perf_mmap_free_page((unsigned long)rb->data_pages[i]); |
| 773 | kfree(rb); |
| 774 | } |
| 775 | |
| 776 | #else |
Jiri Olsa | 5919b30 | 2013-03-19 15:35:09 +0100 | [diff] [blame] | 777 | static int data_page_nr(struct ring_buffer *rb) |
| 778 | { |
| 779 | return rb->nr_pages << page_order(rb); |
| 780 | } |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 781 | |
Peter Zijlstra | 45bfb2e | 2015-01-14 14:18:11 +0200 | [diff] [blame] | 782 | static struct page * |
| 783 | __perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff) |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 784 | { |
Jiri Olsa | 5919b30 | 2013-03-19 15:35:09 +0100 | [diff] [blame] | 785 | /* The '>' counts in the user page. */ |
| 786 | if (pgoff > data_page_nr(rb)) |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 787 | return NULL; |
| 788 | |
| 789 | return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE); |
| 790 | } |
| 791 | |
| 792 | static void perf_mmap_unmark_page(void *addr) |
| 793 | { |
| 794 | struct page *page = vmalloc_to_page(addr); |
| 795 | |
| 796 | page->mapping = NULL; |
| 797 | } |
| 798 | |
| 799 | static void rb_free_work(struct work_struct *work) |
| 800 | { |
| 801 | struct ring_buffer *rb; |
| 802 | void *base; |
| 803 | int i, nr; |
| 804 | |
| 805 | rb = container_of(work, struct ring_buffer, work); |
Jiri Olsa | 5919b30 | 2013-03-19 15:35:09 +0100 | [diff] [blame] | 806 | nr = data_page_nr(rb); |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 807 | |
| 808 | base = rb->user_page; |
Jiri Olsa | 5919b30 | 2013-03-19 15:35:09 +0100 | [diff] [blame] | 809 | /* The '<=' counts in the user page. */ |
| 810 | for (i = 0; i <= nr; i++) |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 811 | perf_mmap_unmark_page(base + (i * PAGE_SIZE)); |
| 812 | |
| 813 | vfree(base); |
| 814 | kfree(rb); |
| 815 | } |
| 816 | |
| 817 | void rb_free(struct ring_buffer *rb) |
| 818 | { |
| 819 | schedule_work(&rb->work); |
| 820 | } |
| 821 | |
| 822 | struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags) |
| 823 | { |
| 824 | struct ring_buffer *rb; |
| 825 | unsigned long size; |
| 826 | void *all_buf; |
| 827 | |
| 828 | size = sizeof(struct ring_buffer); |
| 829 | size += sizeof(void *); |
| 830 | |
| 831 | rb = kzalloc(size, GFP_KERNEL); |
| 832 | if (!rb) |
| 833 | goto fail; |
| 834 | |
| 835 | INIT_WORK(&rb->work, rb_free_work); |
| 836 | |
| 837 | all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE); |
| 838 | if (!all_buf) |
| 839 | goto fail_all_buf; |
| 840 | |
| 841 | rb->user_page = all_buf; |
| 842 | rb->data_pages[0] = all_buf + PAGE_SIZE; |
Peter Zijlstra | 8184059 | 2016-01-29 15:17:51 +0100 | [diff] [blame] | 843 | if (nr_pages) { |
| 844 | rb->nr_pages = 1; |
| 845 | rb->page_order = ilog2(nr_pages); |
| 846 | } |
Frederic Weisbecker | 7636913 | 2011-05-19 19:55:04 +0200 | [diff] [blame] | 847 | |
| 848 | ring_buffer_init(rb, watermark, flags); |
| 849 | |
| 850 | return rb; |
| 851 | |
| 852 | fail_all_buf: |
| 853 | kfree(rb); |
| 854 | |
| 855 | fail: |
| 856 | return NULL; |
| 857 | } |
| 858 | |
| 859 | #endif |
Peter Zijlstra | 45bfb2e | 2015-01-14 14:18:11 +0200 | [diff] [blame] | 860 | |
| 861 | struct page * |
| 862 | perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff) |
| 863 | { |
| 864 | if (rb->aux_nr_pages) { |
| 865 | /* above AUX space */ |
| 866 | if (pgoff > rb->aux_pgoff + rb->aux_nr_pages) |
| 867 | return NULL; |
| 868 | |
| 869 | /* AUX space */ |
| 870 | if (pgoff >= rb->aux_pgoff) |
| 871 | return virt_to_page(rb->aux_pages[pgoff - rb->aux_pgoff]); |
| 872 | } |
| 873 | |
| 874 | return __perf_mmap_to_page(rb, pgoff); |
| 875 | } |