src/panfrost/lib/pan_bo.c

   1 /*
   2  * Copyright 2019 Collabora, Ltd.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  *
  23  * Authors (Collabora):
  24  *   Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
  25  */
  26 #include <errno.h>
  27 #include <stdio.h>
  28 #include <fcntl.h>
  29 #include <xf86drm.h>
  30 #include <pthread.h>
  31 #include "drm-uapi/panfrost_drm.h"
  32
  33 #include "pan_bo.h"
  34 #include "pan_util.h"
  35 #include "wrap.h"
  36
  37 #include "os/os_mman.h"
  38
  39 #include "util/u_inlines.h"
  40 #include "util/u_math.h"
  41
  42 /* This file implements a userspace BO cache. Allocating and freeing
  43  * GPU-visible buffers is very expensive, and even the extra kernel roundtrips
  44  * adds more work than we would like at this point. So caching BOs in userspace
  45  * solves both of these problems and does not require kernel updates.
  46  *
  47  * Cached BOs are sorted into a bucket based on rounding their size down to the
  48  * nearest power-of-two. Each bucket contains a linked list of free panfrost_bo
  49  * objects. Putting a BO into the cache is accomplished by adding it to the
  50  * corresponding bucket. Getting a BO from the cache consists of finding the
  51  * appropriate bucket and sorting. A cache eviction is a kernel-level free of a
  52  * BO and removing it from the bucket. We special case evicting all BOs from
  53  * the cache, since that's what helpful in practice and avoids extra logic
  54  * around the linked list.
  55  */
  56
  57 static struct panfrost_bo *
  58 panfrost_bo_alloc(struct panfrost_device *dev, size_t size,
  59                   uint32_t flags)
  60 {
  61         struct drm_panfrost_create_bo create_bo = { .size = size };
  62         struct panfrost_bo *bo;
  63         int ret;
  64
  65         if (dev->kernel_version->version_major > 1 ||
  66             dev->kernel_version->version_minor >= 1) {
  67                 if (flags & PAN_BO_GROWABLE)
  68                         create_bo.flags |= PANFROST_BO_HEAP;
  69                 if (!(flags & PAN_BO_EXECUTE))
  70                         create_bo.flags |= PANFROST_BO_NOEXEC;
  71         }
  72
  73         ret = drmIoctl(dev->fd, DRM_IOCTL_PANFROST_CREATE_BO, &create_bo);
  74         if (ret) {
  75                 fprintf(stderr, "DRM_IOCTL_PANFROST_CREATE_BO failed: %m\n");
  76                 return NULL;
  77         }
  78
  79         bo = pan_lookup_bo(dev, create_bo.handle);
  80         assert(!memcmp(bo, &((struct panfrost_bo){}), sizeof(*bo)));
  81
  82         bo->size = create_bo.size;
  83         bo->gpu = create_bo.offset;
  84         bo->gem_handle = create_bo.handle;
  85         bo->flags = flags;
  86         bo->dev = dev;
  87         return bo;
  88 }
  89
  90 static void
  91 panfrost_bo_free(struct panfrost_bo *bo)
  92 {
  93         struct drm_gem_close gem_close = { .handle = bo->gem_handle };
  94         int ret;
  95
  96         ret = drmIoctl(bo->dev->fd, DRM_IOCTL_GEM_CLOSE, &gem_close);
  97         if (ret) {
  98                 fprintf(stderr, "DRM_IOCTL_GEM_CLOSE failed: %m\n");
  99                 assert(0);
 100         }
 101
 102         /* BO will be freed with the sparse array, but zero to indicate free */
 103         memset(bo, 0, sizeof(*bo));
 104 }
 105
 106 /* Returns true if the BO is ready, false otherwise.
 107  * access_type is encoding the type of access one wants to ensure is done.
 108  * Waiting is always done for writers, but if wait_readers is set then readers
 109  * are also waited for.
 110  */
 111 bool
 112 panfrost_bo_wait(struct panfrost_bo *bo, int64_t timeout_ns, bool wait_readers)
 113 {
 114         struct drm_panfrost_wait_bo req = {
 115                 .handle = bo->gem_handle,
 116                 .timeout_ns = timeout_ns,
 117         };
 118         int ret;
 119
 120         /* If the BO has been exported or imported we can't rely on the cached
 121          * state, we need to call the WAIT_BO ioctl.
 122          */
 123         if (!(bo->flags & PAN_BO_SHARED)) {
 124                 /* If ->gpu_access is 0, the BO is idle, no need to wait. */
 125                 if (!bo->gpu_access)
 126                         return true;
 127
 128                 /* If the caller only wants to wait for writers and no
 129                  * writes are pending, we don't have to wait.
 130                  */
 131                 if (!wait_readers && !(bo->gpu_access & PAN_BO_ACCESS_WRITE))
 132                         return true;
 133         }
 134
 135         /* The ioctl returns >= 0 value when the BO we are waiting for is ready
 136          * -1 otherwise.
 137          */
 138         ret = drmIoctl(bo->dev->fd, DRM_IOCTL_PANFROST_WAIT_BO, &req);
 139         if (ret != -1) {
 140                 /* Set gpu_access to 0 so that the next call to bo_wait()
 141                  * doesn't have to call the WAIT_BO ioctl.
 142                  */
 143                 bo->gpu_access = 0;
 144                 return true;
 145         }
 146
 147         /* If errno is not ETIMEDOUT or EBUSY that means the handle we passed
 148          * is invalid, which shouldn't happen here.
 149          */
 150         assert(errno == ETIMEDOUT || errno == EBUSY);
 151         return false;
 152 }
 153
 154 /* Helper to calculate the bucket index of a BO */
 155
 156 static unsigned
 157 pan_bucket_index(unsigned size)
 158 {
 159         /* Round down to POT to compute a bucket index */
 160
 161         unsigned bucket_index = util_logbase2(size);
 162
 163         /* Clamp the bucket index; all huge allocations will be
 164          * sorted into the largest bucket */
 165
 166         bucket_index = MIN2(bucket_index, MAX_BO_CACHE_BUCKET);
 167
 168         /* The minimum bucket size must equal the minimum allocation
 169          * size; the maximum we clamped */
 170
 171         assert(bucket_index >= MIN_BO_CACHE_BUCKET);
 172         assert(bucket_index <= MAX_BO_CACHE_BUCKET);
 173
 174         /* Reindex from 0 */
 175         return (bucket_index - MIN_BO_CACHE_BUCKET);
 176 }
 177
 178 static struct list_head *
 179 pan_bucket(struct panfrost_device *dev, unsigned size)
 180 {
 181         return &dev->bo_cache.buckets[pan_bucket_index(size)];
 182 }
 183
 184 /* Tries to fetch a BO of sufficient size with the appropriate flags from the
 185  * BO cache. If it succeeds, it returns that BO and removes the BO from the
 186  * cache. If it fails, it returns NULL signaling the caller to allocate a new
 187  * BO. */
 188
 189 static struct panfrost_bo *
 190 panfrost_bo_cache_fetch(struct panfrost_device *dev,
 191                         size_t size, uint32_t flags, bool dontwait)
 192 {
 193         pthread_mutex_lock(&dev->bo_cache.lock);
 194         struct list_head *bucket = pan_bucket(dev, size);
 195         struct panfrost_bo *bo = NULL;
 196
 197         /* Iterate the bucket looking for something suitable */
 198         list_for_each_entry_safe(struct panfrost_bo, entry, bucket,
 199                                  bucket_link) {
 200                 if (entry->size < size || entry->flags != flags)
 201                         continue;
 202
 203                 if (!panfrost_bo_wait(entry, dontwait ? 0 : INT64_MAX,
 204                                       PAN_BO_ACCESS_RW))
 205                         continue;
 206
 207                 struct drm_panfrost_madvise madv = {
 208                         .handle = entry->gem_handle,
 209                         .madv = PANFROST_MADV_WILLNEED,
 210                 };
 211                 int ret;
 212
 213                 /* This one works, splice it out of the cache */
 214                 list_del(&entry->bucket_link);
 215                 list_del(&entry->lru_link);
 216
 217                 ret = drmIoctl(dev->fd, DRM_IOCTL_PANFROST_MADVISE, &madv);
 218                 if (!ret && !madv.retained) {
 219                         panfrost_bo_free(entry);
 220                         continue;
 221                 }
 222                 /* Let's go! */
 223                 bo = entry;
 224                 break;
 225         }
 226         pthread_mutex_unlock(&dev->bo_cache.lock);
 227
 228         return bo;
 229 }
 230
 231 static void
 232 panfrost_bo_cache_evict_stale_bos(struct panfrost_device *dev)
 233 {
 234         struct timespec time;
 235
 236         clock_gettime(CLOCK_MONOTONIC, &time);
 237         list_for_each_entry_safe(struct panfrost_bo, entry,
 238                                  &dev->bo_cache.lru, lru_link) {
 239                 /* We want all entries that have been used more than 1 sec
 240                  * ago to be dropped, others can be kept.
 241                  * Note the <= 2 check and not <= 1. It's here to account for
 242                  * the fact that we're only testing ->tv_sec, not ->tv_nsec.
 243                  * That means we might keep entries that are between 1 and 2
 244                  * seconds old, but we don't really care, as long as unused BOs
 245                  * are dropped at some point.
 246                  */
 247                 if (time.tv_sec - entry->last_used <= 2)
 248                         break;
 249
 250                 list_del(&entry->bucket_link);
 251                 list_del(&entry->lru_link);
 252                 panfrost_bo_free(entry);
 253         }
 254 }
 255
 256 /* Tries to add a BO to the cache. Returns if it was
 257  * successful */
 258
 259 static bool
 260 panfrost_bo_cache_put(struct panfrost_bo *bo)
 261 {
 262         struct panfrost_device *dev = bo->dev;
 263
 264         if (bo->flags & PAN_BO_SHARED)
 265                 return false;
 266
 267         pthread_mutex_lock(&dev->bo_cache.lock);
 268         struct list_head *bucket = pan_bucket(dev, MAX2(bo->size, 4096));
 269         struct drm_panfrost_madvise madv;
 270         struct timespec time;
 271
 272         madv.handle = bo->gem_handle;
 273         madv.madv = PANFROST_MADV_DONTNEED;
 274         madv.retained = 0;
 275
 276         drmIoctl(dev->fd, DRM_IOCTL_PANFROST_MADVISE, &madv);
 277
 278         /* Add us to the bucket */
 279         list_addtail(&bo->bucket_link, bucket);
 280
 281         /* Add us to the LRU list and update the last_used field. */
 282         list_addtail(&bo->lru_link, &dev->bo_cache.lru);
 283         clock_gettime(CLOCK_MONOTONIC, &time);
 284         bo->last_used = time.tv_sec;
 285
 286         /* Let's do some cleanup in the BO cache while we hold the
 287          * lock.
 288          */
 289         panfrost_bo_cache_evict_stale_bos(dev);
 290         pthread_mutex_unlock(&dev->bo_cache.lock);
 291
 292         return true;
 293 }
 294
 295 /* Evicts all BOs from the cache. Called during context
 296  * destroy or during low-memory situations (to free up
 297  * memory that may be unused by us just sitting in our
 298  * cache, but still reserved from the perspective of the
 299  * OS) */
 300
 301 void
 302 panfrost_bo_cache_evict_all(
 303                 struct panfrost_device *dev)
 304 {
 305         pthread_mutex_lock(&dev->bo_cache.lock);
 306         for (unsigned i = 0; i < ARRAY_SIZE(dev->bo_cache.buckets); ++i) {
 307                 struct list_head *bucket = &dev->bo_cache.buckets[i];
 308
 309                 list_for_each_entry_safe(struct panfrost_bo, entry, bucket,
 310                                          bucket_link) {
 311                         list_del(&entry->bucket_link);
 312                         list_del(&entry->lru_link);
 313                         panfrost_bo_free(entry);
 314                 }
 315         }
 316         pthread_mutex_unlock(&dev->bo_cache.lock);
 317 }
 318
 319 void
 320 panfrost_bo_mmap(struct panfrost_bo *bo)
 321 {
 322         struct drm_panfrost_mmap_bo mmap_bo = { .handle = bo->gem_handle };
 323         int ret;
 324
 325         if (bo->cpu)
 326                 return;
 327
 328         ret = drmIoctl(bo->dev->fd, DRM_IOCTL_PANFROST_MMAP_BO, &mmap_bo);
 329         if (ret) {
 330                 fprintf(stderr, "DRM_IOCTL_PANFROST_MMAP_BO failed: %m\n");
 331                 assert(0);
 332         }
 333
 334         bo->cpu = os_mmap(NULL, bo->size, PROT_READ | PROT_WRITE, MAP_SHARED,
 335                           bo->dev->fd, mmap_bo.offset);
 336         if (bo->cpu == MAP_FAILED) {
 337                 fprintf(stderr, "mmap failed: %p %m\n", bo->cpu);
 338                 assert(0);
 339         }
 340 }
 341
 342 static void
 343 panfrost_bo_munmap(struct panfrost_bo *bo)
 344 {
 345         if (!bo->cpu)
 346                 return;
 347
 348         if (os_munmap((void *) (uintptr_t)bo->cpu, bo->size)) {
 349                 perror("munmap");
 350                 abort();
 351         }
 352
 353         bo->cpu = NULL;
 354 }
 355
 356 struct panfrost_bo *
 357 panfrost_bo_create(struct panfrost_device *dev, size_t size,
 358                    uint32_t flags)
 359 {
 360         struct panfrost_bo *bo;
 361
 362         /* Kernel will fail (confusingly) with EPERM otherwise */
 363         assert(size > 0);
 364
 365         /* To maximize BO cache usage, don't allocate tiny BOs */
 366         size = MAX2(size, 4096);
 367
 368         /* GROWABLE BOs cannot be mmapped */
 369         if (flags & PAN_BO_GROWABLE)
 370                 assert(flags & PAN_BO_INVISIBLE);
 371
 372         /* Before creating a BO, we first want to check the cache but without
 373          * waiting for BO readiness (BOs in the cache can still be referenced
 374          * by jobs that are not finished yet).
 375          * If the cached allocation fails we fall back on fresh BO allocation,
 376          * and if that fails too, we try one more time to allocate from the
 377          * cache, but this time we accept to wait.
 378          */
 379         bo = panfrost_bo_cache_fetch(dev, size, flags, true);
 380         if (!bo)
 381                 bo = panfrost_bo_alloc(dev, size, flags);
 382         if (!bo)
 383                 bo = panfrost_bo_cache_fetch(dev, size, flags, false);
 384
 385         if (!bo)
 386                 fprintf(stderr, "BO creation failed\n");
 387
 388         assert(bo);
 389
 390         /* Only mmap now if we know we need to. For CPU-invisible buffers, we
 391          * never map since we don't care about their contents; they're purely
 392          * for GPU-internal use. But we do trace them anyway. */
 393
 394         if (!(flags & (PAN_BO_INVISIBLE | PAN_BO_DELAY_MMAP)))
 395                 panfrost_bo_mmap(bo);
 396
 397         p_atomic_set(&bo->refcnt, 1);
 398
 399         if (dev->debug & (PAN_DBG_TRACE | PAN_DBG_SYNC)) {
 400                 if (flags & PAN_BO_INVISIBLE)
 401                         pandecode_inject_mmap(bo->gpu, NULL, bo->size, NULL);
 402                 else if (!(flags & PAN_BO_DELAY_MMAP))
 403                         pandecode_inject_mmap(bo->gpu, bo->cpu, bo->size, NULL);
 404         }
 405
 406         return bo;
 407 }
 408
 409 void
 410 panfrost_bo_reference(struct panfrost_bo *bo)
 411 {
 412         if (bo) {
 413                 ASSERTED int count = p_atomic_inc_return(&bo->refcnt);
 414                 assert(count != 1);
 415         }
 416 }
 417
 418 void
 419 panfrost_bo_unreference(struct panfrost_bo *bo)
 420 {
 421         if (!bo)
 422                 return;
 423
 424         /* Don't return to cache if there are still references */
 425         if (p_atomic_dec_return(&bo->refcnt))
 426                 return;
 427
 428         struct panfrost_device *dev = bo->dev;
 429
 430         pthread_mutex_lock(&dev->bo_map_lock);
 431
 432         /* Someone might have imported this BO while we were waiting for the
 433          * lock, let's make sure it's still not referenced before freeing it.
 434          */
 435         if (p_atomic_read(&bo->refcnt) == 0) {
 436                 /* When the reference count goes to zero, we need to cleanup */
 437                 panfrost_bo_munmap(bo);
 438
 439                 /* Rather than freeing the BO now, we'll cache the BO for later
 440                  * allocations if we're allowed to.
 441                  */
 442                 if (!panfrost_bo_cache_put(bo))
 443                         panfrost_bo_free(bo);
 444
 445         }
 446         pthread_mutex_unlock(&dev->bo_map_lock);
 447 }
 448
 449 struct panfrost_bo *
 450 panfrost_bo_import(struct panfrost_device *dev, int fd)
 451 {
 452         struct panfrost_bo *bo;
 453         struct drm_panfrost_get_bo_offset get_bo_offset = {0,};
 454         ASSERTED int ret;
 455         unsigned gem_handle;
 456
 457         ret = drmPrimeFDToHandle(dev->fd, fd, &gem_handle);
 458         assert(!ret);
 459
 460         pthread_mutex_lock(&dev->bo_map_lock);
 461         bo = pan_lookup_bo(dev, gem_handle);
 462
 463         if (!bo->dev) {
 464                 get_bo_offset.handle = gem_handle;
 465                 ret = drmIoctl(dev->fd, DRM_IOCTL_PANFROST_GET_BO_OFFSET, &get_bo_offset);
 466                 assert(!ret);
 467
 468                 bo->dev = dev;
 469                 bo->gpu = (mali_ptr) get_bo_offset.offset;
 470                 bo->size = lseek(fd, 0, SEEK_END);
 471                 bo->flags = PAN_BO_SHARED;
 472                 bo->gem_handle = gem_handle;
 473                 assert(bo->size > 0);
 474                 p_atomic_set(&bo->refcnt, 1);
 475                 // TODO map and unmap on demand?
 476                 panfrost_bo_mmap(bo);
 477         } else {
 478                 /* bo->refcnt == 0 can happen if the BO
 479                  * was being released but panfrost_bo_import() acquired the
 480                  * lock before panfrost_bo_unreference(). In that case, refcnt
 481                  * is 0 and we can't use panfrost_bo_reference() directly, we
 482                  * have to re-initialize the refcnt().
 483                  * Note that panfrost_bo_unreference() checks
 484                  * refcnt value just after acquiring the lock to
 485                  * make sure the object is not freed if panfrost_bo_import()
 486                  * acquired it in the meantime.
 487                  */
 488                 if (p_atomic_read(&bo->refcnt) == 0)
 489                         p_atomic_set(&bo->refcnt, 1);
 490                 else
 491                         panfrost_bo_reference(bo);
 492                 assert(bo->cpu);
 493         }
 494         pthread_mutex_unlock(&dev->bo_map_lock);
 495
 496         return bo;
 497 }
 498
 499 int
 500 panfrost_bo_export(struct panfrost_bo *bo)
 501 {
 502         struct drm_prime_handle args = {
 503                 .handle = bo->gem_handle,
 504                 .flags = DRM_CLOEXEC,
 505         };
 506
 507         int ret = drmIoctl(bo->dev->fd, DRM_IOCTL_PRIME_HANDLE_TO_FD, &args);
 508         if (ret == -1)
 509                 return -1;
 510
 511         bo->flags |= PAN_BO_SHARED;
 512         return args.fd;
 513 }
 514