kernel/events/ring_buffer.c

   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
   6  *  Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
   7  *  Copyright  ©  2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
   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>
  15 #include <linux/circ_buf.h>
  16
  17 #include "internal.h"
  18
  19 static void perf_output_wakeup(struct perf_output_handle *handle)
  20 {
  21         atomic_set(&handle->rb->poll, POLL_IN);
  22
  23         handle->event->pending_wakeup = 1;
  24         irq_work_queue(&handle->event->pending);
  25 }
  26
  27 /*
  28  * We need to ensure a later event_id doesn't publish a head when a former
  29  * event isn't done writing. However since we need to deal with NMIs we
  30  * cannot fully serialize things.
  31  *
  32  * We only publish the head (and generate a wakeup) when the outer-most
  33  * event completes.
  34  */
  35 static void perf_output_get_handle(struct perf_output_handle *handle)
  36 {
  37         struct ring_buffer *rb = handle->rb;
  38
  39         preempt_disable();
  40         local_inc(&rb->nest);
  41         handle->wakeup = local_read(&rb->wakeup);
  42 }
  43
  44 static void perf_output_put_handle(struct perf_output_handle *handle)
  45 {
  46         struct ring_buffer *rb = handle->rb;
  47         unsigned long head;
  48
  49 again:
  50         head = local_read(&rb->head);
  51
  52         /*
  53          * IRQ/NMI can happen here, which means we can miss a head update.
  54          */
  55
  56         if (!local_dec_and_test(&rb->nest))
  57                 goto out;
  58
  59         /*
  60          * Since the mmap() consumer (userspace) can run on a different CPU:
  61          *
  62          *   kernel                             user
  63          *
  64          *   if (LOAD ->data_tail) {            LOAD ->data_head
  65          *                      (A)             smp_rmb()       (C)
  66          *      STORE $data                     LOAD $data
  67          *      smp_wmb()       (B)             smp_mb()        (D)
  68          *      STORE ->data_head               STORE ->data_tail
  69          *   }
  70          *
  71          * Where A pairs with D, and B pairs with C.
  72          *
  73          * In our case (A) is a control dependency that separates the load of
  74          * the ->data_tail and the stores of $data. In case ->data_tail
  75          * indicates there is no room in the buffer to store $data we do not.
  76          *
  77          * D needs to be a full barrier since it separates the data READ
  78          * from the tail WRITE.
  79          *
  80          * For B a WMB is sufficient since it separates two WRITEs, and for C
  81          * an RMB is sufficient since it separates two READs.
  82          *
  83          * See perf_output_begin().
  84          */
  85         smp_wmb(); /* B, matches C */
  86         rb->user_page->data_head = head;
  87
  88         /*
  89          * Now check if we missed an update -- rely on previous implied
  90          * compiler barriers to force a re-read.
  91          */
  92         if (unlikely(head != local_read(&rb->head))) {
  93                 local_inc(&rb->nest);
  94                 goto again;
  95         }
  96
  97         if (handle->wakeup != local_read(&rb->wakeup))
  98                 perf_output_wakeup(handle);
  99
 100 out:
 101         preempt_enable();
 102 }
 103
 104 int perf_output_begin(struct perf_output_handle *handle,
 105                       struct perf_event *event, unsigned int size)
 106 {
 107         struct ring_buffer *rb;
 108         unsigned long tail, offset, head;
 109         int have_lost, page_shift;
 110         struct {
 111                 struct perf_event_header header;
 112                 u64                      id;
 113                 u64                      lost;
 114         } lost_event;
 115
 116         rcu_read_lock();
 117         /*
 118          * For inherited events we send all the output towards the parent.
 119          */
 120         if (event->parent)
 121                 event = event->parent;
 122
 123         rb = rcu_dereference(event->rb);
 124         if (unlikely(!rb))
 125                 goto out;
 126
 127         if (unlikely(!rb->nr_pages))
 128                 goto out;
 129
 130         handle->rb    = rb;
 131         handle->event = event;
 132
 133         have_lost = local_read(&rb->lost);
 134         if (unlikely(have_lost)) {
 135                 size += sizeof(lost_event);
 136                 if (event->attr.sample_id_all)
 137                         size += event->id_header_size;
 138         }
 139
 140         perf_output_get_handle(handle);
 141
 142         do {
 143                 tail = ACCESS_ONCE(rb->user_page->data_tail);
 144                 offset = head = local_read(&rb->head);
 145                 if (!rb->overwrite &&
 146                     unlikely(CIRC_SPACE(head, tail, perf_data_size(rb)) < size))
 147                         goto fail;
 148
 149                 /*
 150                  * The above forms a control dependency barrier separating the
 151                  * @tail load above from the data stores below. Since the @tail
 152                  * load is required to compute the branch to fail below.
 153                  *
 154                  * A, matches D; the full memory barrier userspace SHOULD issue
 155                  * after reading the data and before storing the new tail
 156                  * position.
 157                  *
 158                  * See perf_output_put_handle().
 159                  */
 160
 161                 head += size;
 162         } while (local_cmpxchg(&rb->head, offset, head) != offset);
 163
 164         /*
 165          * We rely on the implied barrier() by local_cmpxchg() to ensure
 166          * none of the data stores below can be lifted up by the compiler.
 167          */
 168
 169         if (unlikely(head - local_read(&rb->wakeup) > rb->watermark))
 170                 local_add(rb->watermark, &rb->wakeup);
 171
 172         page_shift = PAGE_SHIFT + page_order(rb);
 173
 174         handle->page = (offset >> page_shift) & (rb->nr_pages - 1);
 175         offset &= (1UL << page_shift) - 1;
 176         handle->addr = rb->data_pages[handle->page] + offset;
 177         handle->size = (1UL << page_shift) - offset;
 178
 179         if (unlikely(have_lost)) {
 180                 struct perf_sample_data sample_data;
 181
 182                 lost_event.header.size = sizeof(lost_event);
 183                 lost_event.header.type = PERF_RECORD_LOST;
 184                 lost_event.header.misc = 0;
 185                 lost_event.id          = event->id;
 186                 lost_event.lost        = local_xchg(&rb->lost, 0);
 187
 188                 perf_event_header__init_id(&lost_event.header,
 189                                            &sample_data, event);
 190                 perf_output_put(handle, lost_event);
 191                 perf_event__output_id_sample(event, handle, &sample_data);
 192         }
 193
 194         return 0;
 195
 196 fail:
 197         local_inc(&rb->lost);
 198         perf_output_put_handle(handle);
 199 out:
 200         rcu_read_unlock();
 201
 202         return -ENOSPC;
 203 }
 204
 205 unsigned int perf_output_copy(struct perf_output_handle *handle,
 206                       const void *buf, unsigned int len)
 207 {
 208         return __output_copy(handle, buf, len);
 209 }
 210
 211 unsigned int perf_output_skip(struct perf_output_handle *handle,
 212                               unsigned int len)
 213 {
 214         return __output_skip(handle, NULL, len);
 215 }
 216
 217 void perf_output_end(struct perf_output_handle *handle)
 218 {
 219         perf_output_put_handle(handle);
 220         rcu_read_unlock();
 221 }
 222
 223 static void
 224 ring_buffer_init(struct ring_buffer *rb, long watermark, int flags)
 225 {
 226         long max_size = perf_data_size(rb);
 227
 228         if (watermark)
 229                 rb->watermark = min(max_size, watermark);
 230
 231         if (!rb->watermark)
 232                 rb->watermark = max_size / 2;
 233
 234         if (flags & RING_BUFFER_WRITABLE)
 235                 rb->overwrite = 0;
 236         else
 237                 rb->overwrite = 1;
 238
 239         atomic_set(&rb->refcount, 1);
 240
 241         INIT_LIST_HEAD(&rb->event_list);
 242         spin_lock_init(&rb->event_lock);
 243 }
 244
 245 #ifndef CONFIG_PERF_USE_VMALLOC
 246
 247 /*
 248  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
 249  */
 250
 251 struct page *
 252 perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
 253 {
 254         if (pgoff > rb->nr_pages)
 255                 return NULL;
 256
 257         if (pgoff == 0)
 258                 return virt_to_page(rb->user_page);
 259
 260         return virt_to_page(rb->data_pages[pgoff - 1]);
 261 }
 262
 263 static void *perf_mmap_alloc_page(int cpu)
 264 {
 265         struct page *page;
 266         int node;
 267
 268         node = (cpu == -1) ? cpu : cpu_to_node(cpu);
 269         page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
 270         if (!page)
 271                 return NULL;
 272
 273         return page_address(page);
 274 }
 275
 276 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
 277 {
 278         struct ring_buffer *rb;
 279         unsigned long size;
 280         int i;
 281
 282         size = sizeof(struct ring_buffer);
 283         size += nr_pages * sizeof(void *);
 284
 285         rb = kzalloc(size, GFP_KERNEL);
 286         if (!rb)
 287                 goto fail;
 288
 289         rb->user_page = perf_mmap_alloc_page(cpu);
 290         if (!rb->user_page)
 291                 goto fail_user_page;
 292
 293         for (i = 0; i < nr_pages; i++) {
 294                 rb->data_pages[i] = perf_mmap_alloc_page(cpu);
 295                 if (!rb->data_pages[i])
 296                         goto fail_data_pages;
 297         }
 298
 299         rb->nr_pages = nr_pages;
 300
 301         ring_buffer_init(rb, watermark, flags);
 302
 303         return rb;
 304
 305 fail_data_pages:
 306         for (i--; i >= 0; i--)
 307                 free_page((unsigned long)rb->data_pages[i]);
 308
 309         free_page((unsigned long)rb->user_page);
 310
 311 fail_user_page:
 312         kfree(rb);
 313
 314 fail:
 315         return NULL;
 316 }
 317
 318 static void perf_mmap_free_page(unsigned long addr)
 319 {
 320         struct page *page = virt_to_page((void *)addr);
 321
 322         page->mapping = NULL;
 323         __free_page(page);
 324 }
 325
 326 void rb_free(struct ring_buffer *rb)
 327 {
 328         int i;
 329
 330         perf_mmap_free_page((unsigned long)rb->user_page);
 331         for (i = 0; i < rb->nr_pages; i++)
 332                 perf_mmap_free_page((unsigned long)rb->data_pages[i]);
 333         kfree(rb);
 334 }
 335
 336 #else
 337 static int data_page_nr(struct ring_buffer *rb)
 338 {
 339         return rb->nr_pages << page_order(rb);
 340 }
 341
 342 struct page *
 343 perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
 344 {
 345         /* The '>' counts in the user page. */
 346         if (pgoff > data_page_nr(rb))
 347                 return NULL;
 348
 349         return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE);
 350 }
 351
 352 static void perf_mmap_unmark_page(void *addr)
 353 {
 354         struct page *page = vmalloc_to_page(addr);
 355
 356         page->mapping = NULL;
 357 }
 358
 359 static void rb_free_work(struct work_struct *work)
 360 {
 361         struct ring_buffer *rb;
 362         void *base;
 363         int i, nr;
 364
 365         rb = container_of(work, struct ring_buffer, work);
 366         nr = data_page_nr(rb);
 367
 368         base = rb->user_page;
 369         /* The '<=' counts in the user page. */
 370         for (i = 0; i <= nr; i++)
 371                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
 372
 373         vfree(base);
 374         kfree(rb);
 375 }
 376
 377 void rb_free(struct ring_buffer *rb)
 378 {
 379         schedule_work(&rb->work);
 380 }
 381
 382 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
 383 {
 384         struct ring_buffer *rb;
 385         unsigned long size;
 386         void *all_buf;
 387
 388         size = sizeof(struct ring_buffer);
 389         size += sizeof(void *);
 390
 391         rb = kzalloc(size, GFP_KERNEL);
 392         if (!rb)
 393                 goto fail;
 394
 395         INIT_WORK(&rb->work, rb_free_work);
 396
 397         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
 398         if (!all_buf)
 399                 goto fail_all_buf;
 400
 401         rb->user_page = all_buf;
 402         rb->data_pages[0] = all_buf + PAGE_SIZE;
 403         rb->page_order = ilog2(nr_pages);
 404         rb->nr_pages = !!nr_pages;
 405
 406         ring_buffer_init(rb, watermark, flags);
 407
 408         return rb;
 409
 410 fail_all_buf:
 411         kfree(rb);
 412
 413 fail:
 414         return NULL;
 415 }
 416
 417 #endif