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 sz = perf_data_size(rb);
  22         unsigned long mask = sz - 1;
  23
  24         /*
  25          * check if user-writable
  26          * overwrite : over-write its own tail
  27          * !overwrite: buffer possibly drops events.
  28          */
  29         if (rb->overwrite)
  30                 return true;
  31
  32         /*
  33          * verify that payload is not bigger than buffer
  34          * otherwise masking logic may fail to detect
  35          * the "not enough space" condition
  36          */
  37         if ((head - offset) > sz)
  38                 return false;
  39
  40         offset = (offset - tail) & mask;
  41         head   = (head   - tail) & mask;
  42
  43         if ((int)(head - offset) < 0)
  44                 return false;
  45
  46         return true;
  47 }
  48
  49 static void perf_output_wakeup(struct perf_output_handle *handle)
  50 {
  51         atomic_set(&handle->rb->poll, POLL_IN);
  52
  53         handle->event->pending_wakeup = 1;
  54         irq_work_queue(&handle->event->pending);
  55 }
  56
  57 /*
  58  * We need to ensure a later event_id doesn't publish a head when a former
  59  * event isn't done writing. However since we need to deal with NMIs we
  60  * cannot fully serialize things.
  61  *
  62  * We only publish the head (and generate a wakeup) when the outer-most
  63  * event completes.
  64  */
  65 static void perf_output_get_handle(struct perf_output_handle *handle)
  66 {
  67         struct ring_buffer *rb = handle->rb;
  68
  69         preempt_disable();
  70         local_inc(&rb->nest);
  71         handle->wakeup = local_read(&rb->wakeup);
  72 }
  73
  74 static void perf_output_put_handle(struct perf_output_handle *handle)
  75 {
  76         struct ring_buffer *rb = handle->rb;
  77         unsigned long head;
  78
  79 again:
  80         head = local_read(&rb->head);
  81
  82         /*
  83          * IRQ/NMI can happen here, which means we can miss a head update.
  84          */
  85
  86         if (!local_dec_and_test(&rb->nest))
  87                 goto out;
  88
  89         /*
  90          * Since the mmap() consumer (userspace) can run on a different CPU:
  91          *
  92          *   kernel                             user
  93          *
  94          *   READ ->data_tail                   READ ->data_head
  95          *   smp_mb()   (A)                     smp_rmb()       (C)
  96          *   WRITE $data                        READ $data
  97          *   smp_wmb()  (B)                     smp_mb()        (D)
  98          *   STORE ->data_head                  WRITE ->data_tail
  99          *
 100          * Where A pairs with D, and B pairs with C.
 101          *
 102          * I don't think A needs to be a full barrier because we won't in fact
 103          * write data until we see the store from userspace. So we simply don't
 104          * issue the data WRITE until we observe it. Be conservative for now.
 105          *
 106          * OTOH, D needs to be a full barrier since it separates the data READ
 107          * from the tail WRITE.
 108          *
 109          * For B a WMB is sufficient since it separates two WRITEs, and for C
 110          * an RMB is sufficient since it separates two READs.
 111          *
 112          * See perf_output_begin().
 113          */
 114         smp_wmb();
 115         rb->user_page->data_head = head;
 116
 117         /*
 118          * Now check if we missed an update, rely on the (compiler)
 119          * barrier in atomic_dec_and_test() to re-read rb->head.
 120          */
 121         if (unlikely(head != local_read(&rb->head))) {
 122                 local_inc(&rb->nest);
 123                 goto again;
 124         }
 125
 126         if (handle->wakeup != local_read(&rb->wakeup))
 127                 perf_output_wakeup(handle);
 128
 129 out:
 130         preempt_enable();
 131 }
 132
 133 int perf_output_begin(struct perf_output_handle *handle,
 134                       struct perf_event *event, unsigned int size)
 135 {
 136         struct ring_buffer *rb;
 137         unsigned long tail, offset, head;
 138         int have_lost;
 139         struct perf_sample_data sample_data;
 140         struct {
 141                 struct perf_event_header header;
 142                 u64                      id;
 143                 u64                      lost;
 144         } lost_event;
 145
 146         rcu_read_lock();
 147         /*
 148          * For inherited events we send all the output towards the parent.
 149          */
 150         if (event->parent)
 151                 event = event->parent;
 152
 153         rb = rcu_dereference(event->rb);
 154         if (!rb)
 155                 goto out;
 156
 157         handle->rb      = rb;
 158         handle->event   = event;
 159
 160         if (!rb->nr_pages)
 161                 goto out;
 162
 163         have_lost = local_read(&rb->lost);
 164         if (have_lost) {
 165                 lost_event.header.size = sizeof(lost_event);
 166                 perf_event_header__init_id(&lost_event.header, &sample_data,
 167                                            event);
 168                 size += lost_event.header.size;
 169         }
 170
 171         perf_output_get_handle(handle);
 172
 173         do {
 174                 /*
 175                  * Userspace could choose to issue a mb() before updating the
 176                  * tail pointer. So that all reads will be completed before the
 177                  * write is issued.
 178                  *
 179                  * See perf_output_put_handle().
 180                  */
 181                 tail = ACCESS_ONCE(rb->user_page->data_tail);
 182                 smp_mb();
 183                 offset = head = local_read(&rb->head);
 184                 head += size;
 185                 if (unlikely(!perf_output_space(rb, tail, offset, head)))
 186                         goto fail;
 187         } while (local_cmpxchg(&rb->head, offset, head) != offset);
 188
 189         if (head - local_read(&rb->wakeup) > rb->watermark)
 190                 local_add(rb->watermark, &rb->wakeup);
 191
 192         handle->page = offset >> (PAGE_SHIFT + page_order(rb));
 193         handle->page &= rb->nr_pages - 1;
 194         handle->size = offset & ((PAGE_SIZE << page_order(rb)) - 1);
 195         handle->addr = rb->data_pages[handle->page];
 196         handle->addr += handle->size;
 197         handle->size = (PAGE_SIZE << page_order(rb)) - handle->size;
 198
 199         if (have_lost) {
 200                 lost_event.header.type = PERF_RECORD_LOST;
 201                 lost_event.header.misc = 0;
 202                 lost_event.id          = event->id;
 203                 lost_event.lost        = local_xchg(&rb->lost, 0);
 204
 205                 perf_output_put(handle, lost_event);
 206                 perf_event__output_id_sample(event, handle, &sample_data);
 207         }
 208
 209         return 0;
 210
 211 fail:
 212         local_inc(&rb->lost);
 213         perf_output_put_handle(handle);
 214 out:
 215         rcu_read_unlock();
 216
 217         return -ENOSPC;
 218 }
 219
 220 unsigned int perf_output_copy(struct perf_output_handle *handle,
 221                       const void *buf, unsigned int len)
 222 {
 223         return __output_copy(handle, buf, len);
 224 }
 225
 226 unsigned int perf_output_skip(struct perf_output_handle *handle,
 227                               unsigned int len)
 228 {
 229         return __output_skip(handle, NULL, len);
 230 }
 231
 232 void perf_output_end(struct perf_output_handle *handle)
 233 {
 234         perf_output_put_handle(handle);
 235         rcu_read_unlock();
 236 }
 237
 238 static void
 239 ring_buffer_init(struct ring_buffer *rb, long watermark, int flags)
 240 {
 241         long max_size = perf_data_size(rb);
 242
 243         if (watermark)
 244                 rb->watermark = min(max_size, watermark);
 245
 246         if (!rb->watermark)
 247                 rb->watermark = max_size / 2;
 248
 249         if (flags & RING_BUFFER_WRITABLE)
 250                 rb->overwrite = 0;
 251         else
 252                 rb->overwrite = 1;
 253
 254         atomic_set(&rb->refcount, 1);
 255
 256         INIT_LIST_HEAD(&rb->event_list);
 257         spin_lock_init(&rb->event_lock);
 258 }
 259
 260 #ifndef CONFIG_PERF_USE_VMALLOC
 261
 262 /*
 263  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
 264  */
 265
 266 struct page *
 267 perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
 268 {
 269         if (pgoff > rb->nr_pages)
 270                 return NULL;
 271
 272         if (pgoff == 0)
 273                 return virt_to_page(rb->user_page);
 274
 275         return virt_to_page(rb->data_pages[pgoff - 1]);
 276 }
 277
 278 static void *perf_mmap_alloc_page(int cpu)
 279 {
 280         struct page *page;
 281         int node;
 282
 283         node = (cpu == -1) ? cpu : cpu_to_node(cpu);
 284         page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
 285         if (!page)
 286                 return NULL;
 287
 288         return page_address(page);
 289 }
 290
 291 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
 292 {
 293         struct ring_buffer *rb;
 294         unsigned long size;
 295         int i;
 296
 297         size = sizeof(struct ring_buffer);
 298         size += nr_pages * sizeof(void *);
 299
 300         rb = kzalloc(size, GFP_KERNEL);
 301         if (!rb)
 302                 goto fail;
 303
 304         rb->user_page = perf_mmap_alloc_page(cpu);
 305         if (!rb->user_page)
 306                 goto fail_user_page;
 307
 308         for (i = 0; i < nr_pages; i++) {
 309                 rb->data_pages[i] = perf_mmap_alloc_page(cpu);
 310                 if (!rb->data_pages[i])
 311                         goto fail_data_pages;
 312         }
 313
 314         rb->nr_pages = nr_pages;
 315
 316         ring_buffer_init(rb, watermark, flags);
 317
 318         return rb;
 319
 320 fail_data_pages:
 321         for (i--; i >= 0; i--)
 322                 free_page((unsigned long)rb->data_pages[i]);
 323
 324         free_page((unsigned long)rb->user_page);
 325
 326 fail_user_page:
 327         kfree(rb);
 328
 329 fail:
 330         return NULL;
 331 }
 332
 333 static void perf_mmap_free_page(unsigned long addr)
 334 {
 335         struct page *page = virt_to_page((void *)addr);
 336
 337         page->mapping = NULL;
 338         __free_page(page);
 339 }
 340
 341 void rb_free(struct ring_buffer *rb)
 342 {
 343         int i;
 344
 345         perf_mmap_free_page((unsigned long)rb->user_page);
 346         for (i = 0; i < rb->nr_pages; i++)
 347                 perf_mmap_free_page((unsigned long)rb->data_pages[i]);
 348         kfree(rb);
 349 }
 350
 351 #else
 352 static int data_page_nr(struct ring_buffer *rb)
 353 {
 354         return rb->nr_pages << page_order(rb);
 355 }
 356
 357 struct page *
 358 perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
 359 {
 360         /* The '>' counts in the user page. */
 361         if (pgoff > data_page_nr(rb))
 362                 return NULL;
 363
 364         return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE);
 365 }
 366
 367 static void perf_mmap_unmark_page(void *addr)
 368 {
 369         struct page *page = vmalloc_to_page(addr);
 370
 371         page->mapping = NULL;
 372 }
 373
 374 static void rb_free_work(struct work_struct *work)
 375 {
 376         struct ring_buffer *rb;
 377         void *base;
 378         int i, nr;
 379
 380         rb = container_of(work, struct ring_buffer, work);
 381         nr = data_page_nr(rb);
 382
 383         base = rb->user_page;
 384         /* The '<=' counts in the user page. */
 385         for (i = 0; i <= nr; i++)
 386                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
 387
 388         vfree(base);
 389         kfree(rb);
 390 }
 391
 392 void rb_free(struct ring_buffer *rb)
 393 {
 394         schedule_work(&rb->work);
 395 }
 396
 397 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
 398 {
 399         struct ring_buffer *rb;
 400         unsigned long size;
 401         void *all_buf;
 402
 403         size = sizeof(struct ring_buffer);
 404         size += sizeof(void *);
 405
 406         rb = kzalloc(size, GFP_KERNEL);
 407         if (!rb)
 408                 goto fail;
 409
 410         INIT_WORK(&rb->work, rb_free_work);
 411
 412         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
 413         if (!all_buf)
 414                 goto fail_all_buf;
 415
 416         rb->user_page = all_buf;
 417         rb->data_pages[0] = all_buf + PAGE_SIZE;
 418         rb->page_order = ilog2(nr_pages);
 419         rb->nr_pages = !!nr_pages;
 420
 421         ring_buffer_init(rb, watermark, flags);
 422
 423         return rb;
 424
 425 fail_all_buf:
 426         kfree(rb);
 427
 428 fail:
 429         return NULL;
 430 }
 431
 432 #endif