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
  16 #include "internal.h"
  17
  18 static bool perf_output_space(struct ring_buffer *rb, unsigned long tail,
  19                               unsigned long offset, unsigned long head)
  20 {
  21         unsigned long mask;
  22
  23         if (!rb->writable)
  24                 return true;
  25
  26         mask = perf_data_size(rb) - 1;
  27
  28         offset = (offset - tail) & mask;
  29         head   = (head   - tail) & mask;
  30
  31         if ((int)(head - offset) < 0)
  32                 return false;
  33
  34         return true;
  35 }
  36
  37 static void perf_output_wakeup(struct perf_output_handle *handle)
  38 {
  39         atomic_set(&handle->rb->poll, POLL_IN);
  40
  41         handle->event->pending_wakeup = 1;
  42         irq_work_queue(&handle->event->pending);
  43 }
  44
  45 /*
  46  * We need to ensure a later event_id doesn't publish a head when a former
  47  * event isn't done writing. However since we need to deal with NMIs we
  48  * cannot fully serialize things.
  49  *
  50  * We only publish the head (and generate a wakeup) when the outer-most
  51  * event completes.
  52  */
  53 static void perf_output_get_handle(struct perf_output_handle *handle)
  54 {
  55         struct ring_buffer *rb = handle->rb;
  56
  57         preempt_disable();
  58         local_inc(&rb->nest);
  59         handle->wakeup = local_read(&rb->wakeup);
  60 }
  61
  62 static void perf_output_put_handle(struct perf_output_handle *handle)
  63 {
  64         struct ring_buffer *rb = handle->rb;
  65         unsigned long head;
  66
  67 again:
  68         head = local_read(&rb->head);
  69
  70         /*
  71          * IRQ/NMI can happen here, which means we can miss a head update.
  72          */
  73
  74         if (!local_dec_and_test(&rb->nest))
  75                 goto out;
  76
  77         /*
  78          * Publish the known good head. Rely on the full barrier implied
  79          * by atomic_dec_and_test() order the rb->head read and this
  80          * write.
  81          */
  82         rb->user_page->data_head = head;
  83
  84         /*
  85          * Now check if we missed an update, rely on the (compiler)
  86          * barrier in atomic_dec_and_test() to re-read rb->head.
  87          */
  88         if (unlikely(head != local_read(&rb->head))) {
  89                 local_inc(&rb->nest);
  90                 goto again;
  91         }
  92
  93         if (handle->wakeup != local_read(&rb->wakeup))
  94                 perf_output_wakeup(handle);
  95
  96 out:
  97         preempt_enable();
  98 }
  99
 100 int perf_output_begin(struct perf_output_handle *handle,
 101                       struct perf_event *event, unsigned int size,
 102                       int sample)
 103 {
 104         struct ring_buffer *rb;
 105         unsigned long tail, offset, head;
 106         int have_lost;
 107         struct perf_sample_data sample_data;
 108         struct {
 109                 struct perf_event_header header;
 110                 u64                      id;
 111                 u64                      lost;
 112         } lost_event;
 113
 114         rcu_read_lock();
 115         /*
 116          * For inherited events we send all the output towards the parent.
 117          */
 118         if (event->parent)
 119                 event = event->parent;
 120
 121         rb = rcu_dereference(event->rb);
 122         if (!rb)
 123                 goto out;
 124
 125         handle->rb      = rb;
 126         handle->event   = event;
 127         handle->sample  = sample;
 128
 129         if (!rb->nr_pages)
 130                 goto out;
 131
 132         have_lost = local_read(&rb->lost);
 133         if (have_lost) {
 134                 lost_event.header.size = sizeof(lost_event);
 135                 perf_event_header__init_id(&lost_event.header, &sample_data,
 136                                            event);
 137                 size += lost_event.header.size;
 138         }
 139
 140         perf_output_get_handle(handle);
 141
 142         do {
 143                 /*
 144                  * Userspace could choose to issue a mb() before updating the
 145                  * tail pointer. So that all reads will be completed before the
 146                  * write is issued.
 147                  */
 148                 tail = ACCESS_ONCE(rb->user_page->data_tail);
 149                 smp_rmb();
 150                 offset = head = local_read(&rb->head);
 151                 head += size;
 152                 if (unlikely(!perf_output_space(rb, tail, offset, head)))
 153                         goto fail;
 154         } while (local_cmpxchg(&rb->head, offset, head) != offset);
 155
 156         if (head - local_read(&rb->wakeup) > rb->watermark)
 157                 local_add(rb->watermark, &rb->wakeup);
 158
 159         handle->page = offset >> (PAGE_SHIFT + page_order(rb));
 160         handle->page &= rb->nr_pages - 1;
 161         handle->size = offset & ((PAGE_SIZE << page_order(rb)) - 1);
 162         handle->addr = rb->data_pages[handle->page];
 163         handle->addr += handle->size;
 164         handle->size = (PAGE_SIZE << page_order(rb)) - handle->size;
 165
 166         if (have_lost) {
 167                 lost_event.header.type = PERF_RECORD_LOST;
 168                 lost_event.header.misc = 0;
 169                 lost_event.id          = event->id;
 170                 lost_event.lost        = local_xchg(&rb->lost, 0);
 171
 172                 perf_output_put(handle, lost_event);
 173                 perf_event__output_id_sample(event, handle, &sample_data);
 174         }
 175
 176         return 0;
 177
 178 fail:
 179         local_inc(&rb->lost);
 180         perf_output_put_handle(handle);
 181 out:
 182         rcu_read_unlock();
 183
 184         return -ENOSPC;
 185 }
 186
 187 void perf_output_copy(struct perf_output_handle *handle,
 188                       const void *buf, unsigned int len)
 189 {
 190         __output_copy(handle, buf, len);
 191 }
 192
 193 void perf_output_end(struct perf_output_handle *handle)
 194 {
 195         struct perf_event *event = handle->event;
 196         struct ring_buffer *rb = handle->rb;
 197
 198         if (handle->sample && !event->attr.watermark) {
 199                 int wakeup_events = event->attr.wakeup_events;
 200
 201                 if (wakeup_events) {
 202                         int events = local_inc_return(&rb->events);
 203                         if (events >= wakeup_events) {
 204                                 local_sub(wakeup_events, &rb->events);
 205                                 local_inc(&rb->wakeup);
 206                         }
 207                 }
 208         }
 209
 210         perf_output_put_handle(handle);
 211         rcu_read_unlock();
 212 }
 213
 214 static void
 215 ring_buffer_init(struct ring_buffer *rb, long watermark, int flags)
 216 {
 217         long max_size = perf_data_size(rb);
 218
 219         if (watermark)
 220                 rb->watermark = min(max_size, watermark);
 221
 222         if (!rb->watermark)
 223                 rb->watermark = max_size / 2;
 224
 225         if (flags & RING_BUFFER_WRITABLE)
 226                 rb->writable = 1;
 227
 228         atomic_set(&rb->refcount, 1);
 229 }
 230
 231 #ifndef CONFIG_PERF_USE_VMALLOC
 232
 233 /*
 234  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
 235  */
 236
 237 struct page *
 238 perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
 239 {
 240         if (pgoff > rb->nr_pages)
 241                 return NULL;
 242
 243         if (pgoff == 0)
 244                 return virt_to_page(rb->user_page);
 245
 246         return virt_to_page(rb->data_pages[pgoff - 1]);
 247 }
 248
 249 static void *perf_mmap_alloc_page(int cpu)
 250 {
 251         struct page *page;
 252         int node;
 253
 254         node = (cpu == -1) ? cpu : cpu_to_node(cpu);
 255         page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
 256         if (!page)
 257                 return NULL;
 258
 259         return page_address(page);
 260 }
 261
 262 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
 263 {
 264         struct ring_buffer *rb;
 265         unsigned long size;
 266         int i;
 267
 268         size = sizeof(struct ring_buffer);
 269         size += nr_pages * sizeof(void *);
 270
 271         rb = kzalloc(size, GFP_KERNEL);
 272         if (!rb)
 273                 goto fail;
 274
 275         rb->user_page = perf_mmap_alloc_page(cpu);
 276         if (!rb->user_page)
 277                 goto fail_user_page;
 278
 279         for (i = 0; i < nr_pages; i++) {
 280                 rb->data_pages[i] = perf_mmap_alloc_page(cpu);
 281                 if (!rb->data_pages[i])
 282                         goto fail_data_pages;
 283         }
 284
 285         rb->nr_pages = nr_pages;
 286
 287         ring_buffer_init(rb, watermark, flags);
 288
 289         return rb;
 290
 291 fail_data_pages:
 292         for (i--; i >= 0; i--)
 293                 free_page((unsigned long)rb->data_pages[i]);
 294
 295         free_page((unsigned long)rb->user_page);
 296
 297 fail_user_page:
 298         kfree(rb);
 299
 300 fail:
 301         return NULL;
 302 }
 303
 304 static void perf_mmap_free_page(unsigned long addr)
 305 {
 306         struct page *page = virt_to_page((void *)addr);
 307
 308         page->mapping = NULL;
 309         __free_page(page);
 310 }
 311
 312 void rb_free(struct ring_buffer *rb)
 313 {
 314         int i;
 315
 316         perf_mmap_free_page((unsigned long)rb->user_page);
 317         for (i = 0; i < rb->nr_pages; i++)
 318                 perf_mmap_free_page((unsigned long)rb->data_pages[i]);
 319         kfree(rb);
 320 }
 321
 322 #else
 323
 324 struct page *
 325 perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
 326 {
 327         if (pgoff > (1UL << page_order(rb)))
 328                 return NULL;
 329
 330         return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE);
 331 }
 332
 333 static void perf_mmap_unmark_page(void *addr)
 334 {
 335         struct page *page = vmalloc_to_page(addr);
 336
 337         page->mapping = NULL;
 338 }
 339
 340 static void rb_free_work(struct work_struct *work)
 341 {
 342         struct ring_buffer *rb;
 343         void *base;
 344         int i, nr;
 345
 346         rb = container_of(work, struct ring_buffer, work);
 347         nr = 1 << page_order(rb);
 348
 349         base = rb->user_page;
 350         for (i = 0; i < nr + 1; i++)
 351                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
 352
 353         vfree(base);
 354         kfree(rb);
 355 }
 356
 357 void rb_free(struct ring_buffer *rb)
 358 {
 359         schedule_work(&rb->work);
 360 }
 361
 362 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
 363 {
 364         struct ring_buffer *rb;
 365         unsigned long size;
 366         void *all_buf;
 367
 368         size = sizeof(struct ring_buffer);
 369         size += sizeof(void *);
 370
 371         rb = kzalloc(size, GFP_KERNEL);
 372         if (!rb)
 373                 goto fail;
 374
 375         INIT_WORK(&rb->work, rb_free_work);
 376
 377         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
 378         if (!all_buf)
 379                 goto fail_all_buf;
 380
 381         rb->user_page = all_buf;
 382         rb->data_pages[0] = all_buf + PAGE_SIZE;
 383         rb->page_order = ilog2(nr_pages);
 384         rb->nr_pages = 1;
 385
 386         ring_buffer_init(rb, watermark, flags);
 387
 388         return rb;
 389
 390 fail_all_buf:
 391         kfree(rb);
 392
 393 fail:
 394         return NULL;
 395 }
 396
 397 #endif