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 {
 103         struct ring_buffer *rb;
 104         unsigned long tail, offset, head;
 105         int have_lost;
 106         struct perf_sample_data sample_data;
 107         struct {
 108                 struct perf_event_header header;
 109                 u64                      id;
 110                 u64                      lost;
 111         } lost_event;
 112
 113         rcu_read_lock();
 114         /*
 115          * For inherited events we send all the output towards the parent.
 116          */
 117         if (event->parent)
 118                 event = event->parent;
 119
 120         rb = rcu_dereference(event->rb);
 121         if (!rb)
 122                 goto out;
 123
 124         handle->rb      = rb;
 125         handle->event   = event;
 126
 127         if (!rb->nr_pages)
 128                 goto out;
 129
 130         have_lost = local_read(&rb->lost);
 131         if (have_lost) {
 132                 lost_event.header.size = sizeof(lost_event);
 133                 perf_event_header__init_id(&lost_event.header, &sample_data,
 134                                            event);
 135                 size += lost_event.header.size;
 136         }
 137
 138         perf_output_get_handle(handle);
 139
 140         do {
 141                 /*
 142                  * Userspace could choose to issue a mb() before updating the
 143                  * tail pointer. So that all reads will be completed before the
 144                  * write is issued.
 145                  */
 146                 tail = ACCESS_ONCE(rb->user_page->data_tail);
 147                 smp_rmb();
 148                 offset = head = local_read(&rb->head);
 149                 head += size;
 150                 if (unlikely(!perf_output_space(rb, tail, offset, head)))
 151                         goto fail;
 152         } while (local_cmpxchg(&rb->head, offset, head) != offset);
 153
 154         if (head - local_read(&rb->wakeup) > rb->watermark)
 155                 local_add(rb->watermark, &rb->wakeup);
 156
 157         handle->page = offset >> (PAGE_SHIFT + page_order(rb));
 158         handle->page &= rb->nr_pages - 1;
 159         handle->size = offset & ((PAGE_SIZE << page_order(rb)) - 1);
 160         handle->addr = rb->data_pages[handle->page];
 161         handle->addr += handle->size;
 162         handle->size = (PAGE_SIZE << page_order(rb)) - handle->size;
 163
 164         if (have_lost) {
 165                 lost_event.header.type = PERF_RECORD_LOST;
 166                 lost_event.header.misc = 0;
 167                 lost_event.id          = event->id;
 168                 lost_event.lost        = local_xchg(&rb->lost, 0);
 169
 170                 perf_output_put(handle, lost_event);
 171                 perf_event__output_id_sample(event, handle, &sample_data);
 172         }
 173
 174         return 0;
 175
 176 fail:
 177         local_inc(&rb->lost);
 178         perf_output_put_handle(handle);
 179 out:
 180         rcu_read_unlock();
 181
 182         return -ENOSPC;
 183 }
 184
 185 unsigned int perf_output_copy(struct perf_output_handle *handle,
 186                       const void *buf, unsigned int len)
 187 {
 188         return __output_copy(handle, buf, len);
 189 }
 190
 191 unsigned int perf_output_skip(struct perf_output_handle *handle,
 192                               unsigned int len)
 193 {
 194         return __output_skip(handle, NULL, len);
 195 }
 196
 197 void perf_output_end(struct perf_output_handle *handle)
 198 {
 199         perf_output_put_handle(handle);
 200         rcu_read_unlock();
 201 }
 202
 203 static void
 204 ring_buffer_init(struct ring_buffer *rb, long watermark, int flags)
 205 {
 206         long max_size = perf_data_size(rb);
 207
 208         if (watermark)
 209                 rb->watermark = min(max_size, watermark);
 210
 211         if (!rb->watermark)
 212                 rb->watermark = max_size / 2;
 213
 214         if (flags & RING_BUFFER_WRITABLE)
 215                 rb->writable = 1;
 216
 217         atomic_set(&rb->refcount, 1);
 218
 219         INIT_LIST_HEAD(&rb->event_list);
 220         spin_lock_init(&rb->event_lock);
 221 }
 222
 223 #ifndef CONFIG_PERF_USE_VMALLOC
 224
 225 /*
 226  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
 227  */
 228
 229 struct page *
 230 perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
 231 {
 232         if (pgoff > rb->nr_pages)
 233                 return NULL;
 234
 235         if (pgoff == 0)
 236                 return virt_to_page(rb->user_page);
 237
 238         return virt_to_page(rb->data_pages[pgoff - 1]);
 239 }
 240
 241 static void *perf_mmap_alloc_page(int cpu)
 242 {
 243         struct page *page;
 244         int node;
 245
 246         node = (cpu == -1) ? cpu : cpu_to_node(cpu);
 247         page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
 248         if (!page)
 249                 return NULL;
 250
 251         return page_address(page);
 252 }
 253
 254 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
 255 {
 256         struct ring_buffer *rb;
 257         unsigned long size;
 258         int i;
 259
 260         size = sizeof(struct ring_buffer);
 261         size += nr_pages * sizeof(void *);
 262
 263         rb = kzalloc(size, GFP_KERNEL);
 264         if (!rb)
 265                 goto fail;
 266
 267         rb->user_page = perf_mmap_alloc_page(cpu);
 268         if (!rb->user_page)
 269                 goto fail_user_page;
 270
 271         for (i = 0; i < nr_pages; i++) {
 272                 rb->data_pages[i] = perf_mmap_alloc_page(cpu);
 273                 if (!rb->data_pages[i])
 274                         goto fail_data_pages;
 275         }
 276
 277         rb->nr_pages = nr_pages;
 278
 279         ring_buffer_init(rb, watermark, flags);
 280
 281         return rb;
 282
 283 fail_data_pages:
 284         for (i--; i >= 0; i--)
 285                 free_page((unsigned long)rb->data_pages[i]);
 286
 287         free_page((unsigned long)rb->user_page);
 288
 289 fail_user_page:
 290         kfree(rb);
 291
 292 fail:
 293         return NULL;
 294 }
 295
 296 static void perf_mmap_free_page(unsigned long addr)
 297 {
 298         struct page *page = virt_to_page((void *)addr);
 299
 300         page->mapping = NULL;
 301         __free_page(page);
 302 }
 303
 304 void rb_free(struct ring_buffer *rb)
 305 {
 306         int i;
 307
 308         perf_mmap_free_page((unsigned long)rb->user_page);
 309         for (i = 0; i < rb->nr_pages; i++)
 310                 perf_mmap_free_page((unsigned long)rb->data_pages[i]);
 311         kfree(rb);
 312 }
 313
 314 #else
 315
 316 struct page *
 317 perf_mmap_to_page(struct ring_buffer *rb, unsigned long pgoff)
 318 {
 319         if (pgoff > (1UL << page_order(rb)))
 320                 return NULL;
 321
 322         return vmalloc_to_page((void *)rb->user_page + pgoff * PAGE_SIZE);
 323 }
 324
 325 static void perf_mmap_unmark_page(void *addr)
 326 {
 327         struct page *page = vmalloc_to_page(addr);
 328
 329         page->mapping = NULL;
 330 }
 331
 332 static void rb_free_work(struct work_struct *work)
 333 {
 334         struct ring_buffer *rb;
 335         void *base;
 336         int i, nr;
 337
 338         rb = container_of(work, struct ring_buffer, work);
 339         nr = 1 << page_order(rb);
 340
 341         base = rb->user_page;
 342         for (i = 0; i < nr + 1; i++)
 343                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
 344
 345         vfree(base);
 346         kfree(rb);
 347 }
 348
 349 void rb_free(struct ring_buffer *rb)
 350 {
 351         schedule_work(&rb->work);
 352 }
 353
 354 struct ring_buffer *rb_alloc(int nr_pages, long watermark, int cpu, int flags)
 355 {
 356         struct ring_buffer *rb;
 357         unsigned long size;
 358         void *all_buf;
 359
 360         size = sizeof(struct ring_buffer);
 361         size += sizeof(void *);
 362
 363         rb = kzalloc(size, GFP_KERNEL);
 364         if (!rb)
 365                 goto fail;
 366
 367         INIT_WORK(&rb->work, rb_free_work);
 368
 369         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
 370         if (!all_buf)
 371                 goto fail_all_buf;
 372
 373         rb->user_page = all_buf;
 374         rb->data_pages[0] = all_buf + PAGE_SIZE;
 375         rb->page_order = ilog2(nr_pages);
 376         rb->nr_pages = 1;
 377
 378         ring_buffer_init(rb, watermark, flags);
 379
 380         return rb;
 381
 382 fail_all_buf:
 383         kfree(rb);
 384
 385 fail:
 386         return NULL;
 387 }
 388
 389 #endif