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 the (compiler)
  89          * barrier in atomic_dec_and_test() to re-read rb->head.
  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;
 109         struct perf_sample_data sample_data;
 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                 lost_event.header.size = sizeof(lost_event);
 136                 perf_event_header__init_id(&lost_event.header, &sample_data,
 137                                            event);
 138                 size += lost_event.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                 head += size;
 150         } while (local_cmpxchg(&rb->head, offset, head) != offset);
 151
 152         /*
 153          * Separate the userpage->tail read from the data stores below.
 154          * Matches the MB userspace SHOULD issue after reading the data
 155          * and before storing the new tail position.
 156          *
 157          * See perf_output_put_handle().
 158          */
 159         smp_mb();
 160
 161         if (unlikely(head - local_read(&rb->wakeup) > rb->watermark))
 162                 local_add(rb->watermark, &rb->wakeup);
 163
 164         handle->page = offset >> (PAGE_SHIFT + page_order(rb));
 165         handle->page &= rb->nr_pages - 1;
 166         handle->size = offset & ((PAGE_SIZE << page_order(rb)) - 1);
 167         handle->addr = rb->data_pages[handle->page];
 168         handle->addr += handle->size;
 169         handle->size = (PAGE_SIZE << page_order(rb)) - handle->size;
 170
 171         if (unlikely(have_lost)) {
 172                 lost_event.header.type = PERF_RECORD_LOST;
 173                 lost_event.header.misc = 0;
 174                 lost_event.id          = event->id;
 175                 lost_event.lost        = local_xchg(&rb->lost, 0);
 176
 177                 perf_output_put(handle, lost_event);
 178                 perf_event__output_id_sample(event, handle, &sample_data);
 179         }
 180
 181         return 0;
 182
 183 fail:
 184         local_inc(&rb->lost);
 185         perf_output_put_handle(handle);
 186 out:
 187         rcu_read_unlock();
 188
 189         return -ENOSPC;
 190 }
 191
 192 unsigned int perf_output_copy(struct perf_output_handle *handle,
 193                       const void *buf, unsigned int len)
 194 {
 195         return __output_copy(handle, buf, len);
 196 }
 197
 198 unsigned int perf_output_skip(struct perf_output_handle *handle,
 199                               unsigned int len)
 200 {
 201         return __output_skip(handle, NULL, len);
 202 }
 203
 204 void perf_output_end(struct perf_output_handle *handle)
 205 {
 206         perf_output_put_handle(handle);
 207         rcu_read_unlock();
 208 }
 209
 210 static void
 211 ring_buffer_init(struct ring_buffer *rb, long watermark, int flags)
 212 {
 213         long max_size = perf_data_size(rb);
 214
 215         if (watermark)
 216                 rb->watermark = min(max_size, watermark);
 217
 218         if (!rb->watermark)
 219                 rb->watermark = max_size / 2;
 220
 221         if (flags & RING_BUFFER_WRITABLE)
 222                 rb->overwrite = 0;
 223         else
 224                 rb->overwrite = 1;
 225
 226         atomic_set(&rb->refcount, 1);
 227
 228         INIT_LIST_HEAD(&rb->event_list);
 229         spin_lock_init(&rb->event_lock);
 230 }
 231
 232 #ifndef CONFIG_PERF_USE_VMALLOC
 233
 234 /*
 235  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
 236  */
 237
 238 struct page *
 239 perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
 240 {
 241         if (pgoff > rb->nr_pages)
 242                 return NULL;
 243
 244         if (pgoff == 0)
 245                 return virt_to_page(rb->user_page);
 246
 247         return virt_to_page(rb->data_pages[pgoff - 1]);
 248 }
 249
 250 static void *perf_mmap_alloc_page(int cpu)
 251 {
 252         struct page *page;
 253         int node;
 254
 255         node = (cpu == -1) ? cpu : cpu_to_node(cpu);
 256         page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
 257         if (!page)
 258                 return NULL;
 259
 260         return page_address(page);
 261 }
 262
 263 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
 264 {
 265         struct ring_buffer *rb;
 266         unsigned long size;
 267         int i;
 268
 269         size = sizeof(struct ring_buffer);
 270         size += nr_pages * sizeof(void *);
 271
 272         rb = kzalloc(size, GFP_KERNEL);
 273         if (!rb)
 274                 goto fail;
 275
 276         rb->user_page = perf_mmap_alloc_page(cpu);
 277         if (!rb->user_page)
 278                 goto fail_user_page;
 279
 280         for (i = 0; i < nr_pages; i++) {
 281                 rb->data_pages[i] = perf_mmap_alloc_page(cpu);
 282                 if (!rb->data_pages[i])
 283                         goto fail_data_pages;
 284         }
 285
 286         rb->nr_pages = nr_pages;
 287
 288         ring_buffer_init(rb, watermark, flags);
 289
 290         return rb;
 291
 292 fail_data_pages:
 293         for (i--; i >= 0; i--)
 294                 free_page((unsigned long)rb->data_pages[i]);
 295
 296         free_page((unsigned long)rb->user_page);
 297
 298 fail_user_page:
 299         kfree(rb);
 300
 301 fail:
 302         return NULL;
 303 }
 304
 305 static void perf_mmap_free_page(unsigned long addr)
 306 {
 307         struct page *page = virt_to_page((void *)addr);
 308
 309         page->mapping = NULL;
 310         __free_page(page);
 311 }
 312
 313 void rb_free(struct ring_buffer *rb)
 314 {
 315         int i;
 316
 317         perf_mmap_free_page((unsigned long)rb->user_page);
 318         for (i = 0; i < rb->nr_pages; i++)
 319                 perf_mmap_free_page((unsigned long)rb->data_pages[i]);
 320         kfree(rb);
 321 }
 322
 323 #else
 324 static int data_page_nr(struct ring_buffer *rb)
 325 {
 326         return rb->nr_pages << page_order(rb);
 327 }
 328
 329 struct page *
 330 perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
 331 {
 332         /* The '>' counts in the user page. */
 333         if (pgoff > data_page_nr(rb))
 334                 return NULL;
 335
 336         return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE);
 337 }
 338
 339 static void perf_mmap_unmark_page(void *addr)
 340 {
 341         struct page *page = vmalloc_to_page(addr);
 342
 343         page->mapping = NULL;
 344 }
 345
 346 static void rb_free_work(struct work_struct *work)
 347 {
 348         struct ring_buffer *rb;
 349         void *base;
 350         int i, nr;
 351
 352         rb = container_of(work, struct ring_buffer, work);
 353         nr = data_page_nr(rb);
 354
 355         base = rb->user_page;
 356         /* The '<=' counts in the user page. */
 357         for (i = 0; i <= nr; i++)
 358                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
 359
 360         vfree(base);
 361         kfree(rb);
 362 }
 363
 364 void rb_free(struct ring_buffer *rb)
 365 {
 366         schedule_work(&rb->work);
 367 }
 368
 369 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
 370 {
 371         struct ring_buffer *rb;
 372         unsigned long size;
 373         void *all_buf;
 374
 375         size = sizeof(struct ring_buffer);
 376         size += sizeof(void *);
 377
 378         rb = kzalloc(size, GFP_KERNEL);
 379         if (!rb)
 380                 goto fail;
 381
 382         INIT_WORK(&rb->work, rb_free_work);
 383
 384         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
 385         if (!all_buf)
 386                 goto fail_all_buf;
 387
 388         rb->user_page = all_buf;
 389         rb->data_pages[0] = all_buf + PAGE_SIZE;
 390         rb->page_order = ilog2(nr_pages);
 391         rb->nr_pages = !!nr_pages;
 392
 393         ring_buffer_init(rb, watermark, flags);
 394
 395         return rb;
 396
 397 fail_all_buf:
 398         kfree(rb);
 399
 400 fail:
 401         return NULL;
 402 }
 403
 404 #endif