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