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          *   READ ->data_tail                   READ ->data_head
  65          *   smp_mb()   (A)                     smp_rmb()       (C)
  66          *   WRITE $data                        READ $data
  67          *   smp_wmb()  (B)                     smp_mb()        (D)
  68          *   STORE ->data_head                  WRITE ->data_tail
  69          *
  70          * Where A pairs with D, and B pairs with C.
  71          *
  72          * I don't think A needs to be a full barrier because we won't in fact
  73          * write data until we see the store from userspace. So we simply don't
  74          * issue the data WRITE until we observe it. Be conservative for now.
  75          *
  76          * OTOH, D needs to be a full barrier since it separates the data READ
  77          * from the tail WRITE.
  78          *
  79          * For B a WMB is sufficient since it separates two WRITEs, and for C
  80          * an RMB is sufficient since it separates two READs.
  81          *
  82          * See perf_output_begin().
  83          */
  84         smp_wmb();
  85         rb->user_page->data_head = head;
  86
  87         /*
  88          * Now check if we missed an update -- rely on previous implied
  89          * compiler barriers to force a re-read.
  90          */
  91         if (unlikely(head != local_read(&rb->head))) {
  92                 local_inc(&rb->nest);
  93                 goto again;
  94         }
  95
  96         if (handle->wakeup != local_read(&rb->wakeup))
  97                 perf_output_wakeup(handle);
  98
  99 out:
 100         preempt_enable();
 101 }
 102
 103 int perf_output_begin(struct perf_output_handle *handle,
 104                       struct perf_event *event, unsigned int size)
 105 {
 106         struct ring_buffer *rb;
 107         unsigned long tail, offset, head;
 108         int have_lost, page_shift;
 109         struct {
 110                 struct perf_event_header header;
 111                 u64                      id;
 112                 u64                      lost;
 113         } lost_event;
 114
 115         rcu_read_lock();
 116         /*
 117          * For inherited events we send all the output towards the parent.
 118          */
 119         if (event->parent)
 120                 event = event->parent;
 121
 122         rb = rcu_dereference(event->rb);
 123         if (unlikely(!rb))
 124                 goto out;
 125
 126         if (unlikely(!rb->nr_pages))
 127                 goto out;
 128
 129         handle->rb    = rb;
 130         handle->event = event;
 131
 132         have_lost = local_read(&rb->lost);
 133         if (unlikely(have_lost)) {
 134                 size += sizeof(lost_event);
 135                 if (event->attr.sample_id_all)
 136                         size += event->id_header_size;
 137         }
 138
 139         perf_output_get_handle(handle);
 140
 141         do {
 142                 tail = ACCESS_ONCE(rb->user_page->data_tail);
 143                 offset = head = local_read(&rb->head);
 144                 if (!rb->overwrite &&
 145                     unlikely(CIRC_SPACE(head, tail, perf_data_size(rb)) < size))
 146                         goto fail;
 147                 head += size;
 148         } while (local_cmpxchg(&rb->head, offset, head) != offset);
 149
 150         /*
 151          * Separate the userpage->tail read from the data stores below.
 152          * Matches the MB userspace SHOULD issue after reading the data
 153          * and before storing the new tail position.
 154          *
 155          * See perf_output_put_handle().
 156          */
 157         smp_mb();
 158
 159         if (unlikely(head - local_read(&rb->wakeup) > rb->watermark))
 160                 local_add(rb->watermark, &rb->wakeup);
 161
 162         page_shift = PAGE_SHIFT + page_order(rb);
 163
 164         handle->page = (offset >> page_shift) & (rb->nr_pages - 1);
 165         offset &= (1UL << page_shift) - 1;
 166         handle->addr = rb->data_pages[handle->page] + offset;
 167         handle->size = (1UL << page_shift) - offset;
 168
 169         if (unlikely(have_lost)) {
 170                 struct perf_sample_data sample_data;
 171
 172                 lost_event.header.size = sizeof(lost_event);
 173                 lost_event.header.type = PERF_RECORD_LOST;
 174                 lost_event.header.misc = 0;
 175                 lost_event.id          = event->id;
 176                 lost_event.lost        = local_xchg(&rb->lost, 0);
 177
 178                 perf_event_header__init_id(&lost_event.header,
 179                                            &sample_data, event);
 180                 perf_output_put(handle, lost_event);
 181                 perf_event__output_id_sample(event, handle, &sample_data);
 182         }
 183
 184         return 0;
 185
 186 fail:
 187         local_inc(&rb->lost);
 188         perf_output_put_handle(handle);
 189 out:
 190         rcu_read_unlock();
 191
 192         return -ENOSPC;
 193 }
 194
 195 unsigned int perf_output_copy(struct perf_output_handle *handle,
 196                       const void *buf, unsigned int len)
 197 {
 198         return __output_copy(handle, buf, len);
 199 }
 200
 201 unsigned int perf_output_skip(struct perf_output_handle *handle,
 202                               unsigned int len)
 203 {
 204         return __output_skip(handle, NULL, len);
 205 }
 206
 207 void perf_output_end(struct perf_output_handle *handle)
 208 {
 209         perf_output_put_handle(handle);
 210         rcu_read_unlock();
 211 }
 212
 213 static void
 214 ring_buffer_init(struct ring_buffer *rb, long watermark, int flags)
 215 {
 216         long max_size = perf_data_size(rb);
 217
 218         if (watermark)
 219                 rb->watermark = min(max_size, watermark);
 220
 221         if (!rb->watermark)
 222                 rb->watermark = max_size / 2;
 223
 224         if (flags & RING_BUFFER_WRITABLE)
 225                 rb->overwrite = 0;
 226         else
 227                 rb->overwrite = 1;
 228
 229         atomic_set(&rb->refcount, 1);
 230
 231         INIT_LIST_HEAD(&rb->event_list);
 232         spin_lock_init(&rb->event_lock);
 233 }
 234
 235 #ifndef CONFIG_PERF_USE_VMALLOC
 236
 237 /*
 238  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
 239  */
 240
 241 struct page *
 242 perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
 243 {
 244         if (pgoff > rb->nr_pages)
 245                 return NULL;
 246
 247         if (pgoff == 0)
 248                 return virt_to_page(rb->user_page);
 249
 250         return virt_to_page(rb->data_pages[pgoff - 1]);
 251 }
 252
 253 static void *perf_mmap_alloc_page(int cpu)
 254 {
 255         struct page *page;
 256         int node;
 257
 258         node = (cpu == -1) ? cpu : cpu_to_node(cpu);
 259         page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
 260         if (!page)
 261                 return NULL;
 262
 263         return page_address(page);
 264 }
 265
 266 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
 267 {
 268         struct ring_buffer *rb;
 269         unsigned long size;
 270         int i;
 271
 272         size = sizeof(struct ring_buffer);
 273         size += nr_pages * sizeof(void *);
 274
 275         rb = kzalloc(size, GFP_KERNEL);
 276         if (!rb)
 277                 goto fail;
 278
 279         rb->user_page = perf_mmap_alloc_page(cpu);
 280         if (!rb->user_page)
 281                 goto fail_user_page;
 282
 283         for (i = 0; i < nr_pages; i++) {
 284                 rb->data_pages[i] = perf_mmap_alloc_page(cpu);
 285                 if (!rb->data_pages[i])
 286                         goto fail_data_pages;
 287         }
 288
 289         rb->nr_pages = nr_pages;
 290
 291         ring_buffer_init(rb, watermark, flags);
 292
 293         return rb;
 294
 295 fail_data_pages:
 296         for (i--; i >= 0; i--)
 297                 free_page((unsigned long)rb->data_pages[i]);
 298
 299         free_page((unsigned long)rb->user_page);
 300
 301 fail_user_page:
 302         kfree(rb);
 303
 304 fail:
 305         return NULL;
 306 }
 307
 308 static void perf_mmap_free_page(unsigned long addr)
 309 {
 310         struct page *page = virt_to_page((void *)addr);
 311
 312         page->mapping = NULL;
 313         __free_page(page);
 314 }
 315
 316 void rb_free(struct ring_buffer *rb)
 317 {
 318         int i;
 319
 320         perf_mmap_free_page((unsigned long)rb->user_page);
 321         for (i = 0; i < rb->nr_pages; i++)
 322                 perf_mmap_free_page((unsigned long)rb->data_pages[i]);
 323         kfree(rb);
 324 }
 325
 326 #else
 327 static int data_page_nr(struct ring_buffer *rb)
 328 {
 329         return rb->nr_pages << page_order(rb);
 330 }
 331
 332 struct page *
 333 perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
 334 {
 335         /* The '>' counts in the user page. */
 336         if (pgoff > data_page_nr(rb))
 337                 return NULL;
 338
 339         return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE);
 340 }
 341
 342 static void perf_mmap_unmark_page(void *addr)
 343 {
 344         struct page *page = vmalloc_to_page(addr);
 345
 346         page->mapping = NULL;
 347 }
 348
 349 static void rb_free_work(struct work_struct *work)
 350 {
 351         struct ring_buffer *rb;
 352         void *base;
 353         int i, nr;
 354
 355         rb = container_of(work, struct ring_buffer, work);
 356         nr = data_page_nr(rb);
 357
 358         base = rb->user_page;
 359         /* The '<=' counts in the user page. */
 360         for (i = 0; i <= nr; i++)
 361                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
 362
 363         vfree(base);
 364         kfree(rb);
 365 }
 366
 367 void rb_free(struct ring_buffer *rb)
 368 {
 369         schedule_work(&rb->work);
 370 }
 371
 372 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
 373 {
 374         struct ring_buffer *rb;
 375         unsigned long size;
 376         void *all_buf;
 377
 378         size = sizeof(struct ring_buffer);
 379         size += sizeof(void *);
 380
 381         rb = kzalloc(size, GFP_KERNEL);
 382         if (!rb)
 383                 goto fail;
 384
 385         INIT_WORK(&rb->work, rb_free_work);
 386
 387         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
 388         if (!all_buf)
 389                 goto fail_all_buf;
 390
 391         rb->user_page = all_buf;
 392         rb->data_pages[0] = all_buf + PAGE_SIZE;
 393         rb->page_order = ilog2(nr_pages);
 394         rb->nr_pages = !!nr_pages;
 395
 396         ring_buffer_init(rb, watermark, flags);
 397
 398         return rb;
 399
 400 fail_all_buf:
 401         kfree(rb);
 402
 403 fail:
 404         return NULL;
 405 }
 406
 407 #endif