src/gallium/drivers/panfrost/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 <stdio.h>
  27 #include <fcntl.h>
  28 #include <xf86drm.h>
  29 #include <pthread.h>
  30 #include "drm-uapi/panfrost_drm.h"
  31
  32 #include "pan_bo.h"
  33 #include "pan_screen.h"
  34 #include "pan_util.h"
  35 #include "pandecode/decode.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 /* Helper to calculate the bucket index of a BO */
  58
  59 static unsigned
  60 pan_bucket_index(unsigned size)
  61 {
  62         /* Round down to POT to compute a bucket index */
  63
  64         unsigned bucket_index = util_logbase2(size);
  65
  66         /* Clamp the bucket index; all huge allocations will be
  67          * sorted into the largest bucket */
  68
  69         bucket_index = MIN2(bucket_index, MAX_BO_CACHE_BUCKET);
  70
  71         /* The minimum bucket size must equal the minimum allocation
  72          * size; the maximum we clamped */
  73
  74         assert(bucket_index >= MIN_BO_CACHE_BUCKET);
  75         assert(bucket_index <= MAX_BO_CACHE_BUCKET);
  76
  77         /* Reindex from 0 */
  78         return (bucket_index - MIN_BO_CACHE_BUCKET);
  79 }
  80
  81 static struct list_head *
  82 pan_bucket(struct panfrost_screen *screen, unsigned size)
  83 {
  84         return &screen->bo_cache[pan_bucket_index(size)];
  85 }
  86
  87 /* Tries to fetch a BO of sufficient size with the appropriate flags from the
  88  * BO cache. If it succeeds, it returns that BO and removes the BO from the
  89  * cache. If it fails, it returns NULL signaling the caller to allocate a new
  90  * BO. */
  91
  92 static struct panfrost_bo *
  93 panfrost_bo_cache_fetch(
  94                 struct panfrost_screen *screen,
  95                 size_t size, uint32_t flags)
  96 {
  97         pthread_mutex_lock(&screen->bo_cache_lock);
  98         struct list_head *bucket = pan_bucket(screen, size);
  99         struct panfrost_bo *bo = NULL;
 100
 101         /* Iterate the bucket looking for something suitable */
 102         list_for_each_entry_safe(struct panfrost_bo, entry, bucket, link) {
 103                 if (entry->size >= size &&
 104                     entry->flags == flags) {
 105                         int ret;
 106                         struct drm_panfrost_madvise madv;
 107
 108                         /* This one works, splice it out of the cache */
 109                         list_del(&entry->link);
 110
 111                         madv.handle = entry->gem_handle;
 112                         madv.madv = PANFROST_MADV_WILLNEED;
 113                         madv.retained = 0;
 114
 115                         ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_MADVISE, &madv);
 116                         if (!ret && !madv.retained) {
 117                                 panfrost_bo_release(entry, false);
 118                                 continue;
 119                         }
 120                         /* Let's go! */
 121                         bo = entry;
 122                         break;
 123                 }
 124         }
 125         pthread_mutex_unlock(&screen->bo_cache_lock);
 126
 127         return bo;
 128 }
 129
 130 /* Tries to add a BO to the cache. Returns if it was
 131  * successful */
 132
 133 static bool
 134 panfrost_bo_cache_put(struct panfrost_bo *bo)
 135 {
 136         struct panfrost_screen *screen = bo->screen;
 137
 138         pthread_mutex_lock(&screen->bo_cache_lock);
 139         struct list_head *bucket = pan_bucket(screen, bo->size);
 140         struct drm_panfrost_madvise madv;
 141
 142         madv.handle = bo->gem_handle;
 143         madv.madv = PANFROST_MADV_DONTNEED;
 144         madv.retained = 0;
 145
 146         drmIoctl(screen->fd, DRM_IOCTL_PANFROST_MADVISE, &madv);
 147
 148         /* Add us to the bucket */
 149         list_addtail(&bo->link, bucket);
 150         pthread_mutex_unlock(&screen->bo_cache_lock);
 151
 152         return true;
 153 }
 154
 155 /* Evicts all BOs from the cache. Called during context
 156  * destroy or during low-memory situations (to free up
 157  * memory that may be unused by us just sitting in our
 158  * cache, but still reserved from the perspective of the
 159  * OS) */
 160
 161 void
 162 panfrost_bo_cache_evict_all(
 163                 struct panfrost_screen *screen)
 164 {
 165         pthread_mutex_lock(&screen->bo_cache_lock);
 166         for (unsigned i = 0; i < ARRAY_SIZE(screen->bo_cache); ++i) {
 167                 struct list_head *bucket = &screen->bo_cache[i];
 168
 169                 list_for_each_entry_safe(struct panfrost_bo, entry, bucket, link) {
 170                         list_del(&entry->link);
 171                         panfrost_bo_release(entry, false);
 172                 }
 173         }
 174         pthread_mutex_unlock(&screen->bo_cache_lock);
 175 }
 176
 177 void
 178 panfrost_bo_mmap(struct panfrost_bo *bo)
 179 {
 180         struct drm_panfrost_mmap_bo mmap_bo = { .handle = bo->gem_handle };
 181         int ret;
 182
 183         if (bo->cpu)
 184                 return;
 185
 186         ret = drmIoctl(bo->screen->fd, DRM_IOCTL_PANFROST_MMAP_BO, &mmap_bo);
 187         if (ret) {
 188                 fprintf(stderr, "DRM_IOCTL_PANFROST_MMAP_BO failed: %m\n");
 189                 assert(0);
 190         }
 191
 192         bo->cpu = os_mmap(NULL, bo->size, PROT_READ | PROT_WRITE, MAP_SHARED,
 193                           bo->screen->fd, mmap_bo.offset);
 194         if (bo->cpu == MAP_FAILED) {
 195                 fprintf(stderr, "mmap failed: %p %m\n", bo->cpu);
 196                 assert(0);
 197         }
 198
 199         /* Record the mmap if we're tracing */
 200         if (pan_debug & PAN_DBG_TRACE)
 201                 pandecode_inject_mmap(bo->gpu, bo->cpu, bo->size, NULL);
 202 }
 203
 204 static void
 205 panfrost_bo_munmap(struct panfrost_bo *bo)
 206 {
 207         if (!bo->cpu)
 208                 return;
 209
 210         if (os_munmap((void *) (uintptr_t)bo->cpu, bo->size)) {
 211                 perror("munmap");
 212                 abort();
 213         }
 214
 215         bo->cpu = NULL;
 216 }
 217
 218 struct panfrost_bo *
 219 panfrost_bo_create(struct panfrost_screen *screen, size_t size,
 220                    uint32_t flags)
 221 {
 222         struct panfrost_bo *bo;
 223
 224         /* Kernel will fail (confusingly) with EPERM otherwise */
 225         assert(size > 0);
 226
 227         /* To maximize BO cache usage, don't allocate tiny BOs */
 228         size = MAX2(size, 4096);
 229
 230         /* GROWABLE BOs cannot be mmapped */
 231         if (flags & PAN_BO_GROWABLE)
 232                 assert(flags & PAN_BO_INVISIBLE);
 233
 234         unsigned translated_flags = 0;
 235
 236         if (screen->kernel_version->version_major > 1 ||
 237             screen->kernel_version->version_minor >= 1) {
 238                 if (flags & PAN_BO_GROWABLE)
 239                         translated_flags |= PANFROST_BO_HEAP;
 240                 if (!(flags & PAN_BO_EXECUTE))
 241                         translated_flags |= PANFROST_BO_NOEXEC;
 242         }
 243
 244         struct drm_panfrost_create_bo create_bo = {
 245                 .size = size,
 246                 .flags = translated_flags,
 247         };
 248
 249         /* Before creating a BO, we first want to check the cache */
 250
 251         bo = panfrost_bo_cache_fetch(screen, size, flags);
 252
 253         if (bo == NULL) {
 254                 /* Otherwise, the cache misses and we need to allocate a BO fresh from
 255                  * the kernel */
 256
 257                 int ret;
 258
 259                 ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_CREATE_BO, &create_bo);
 260                 if (ret) {
 261                         fprintf(stderr, "DRM_IOCTL_PANFROST_CREATE_BO failed: %m\n");
 262                         assert(0);
 263                 }
 264
 265                 /* We have a BO allocated from the kernel; fill in the userspace
 266                  * version */
 267
 268                 bo = rzalloc(screen, struct panfrost_bo);
 269                 bo->size = create_bo.size;
 270                 bo->gpu = create_bo.offset;
 271                 bo->gem_handle = create_bo.handle;
 272                 bo->flags = flags;
 273         }
 274
 275         /* Only mmap now if we know we need to. For CPU-invisible buffers, we
 276          * never map since we don't care about their contents; they're purely
 277          * for GPU-internal use. But we do trace them anyway. */
 278
 279         if (!(flags & (PAN_BO_INVISIBLE | PAN_BO_DELAY_MMAP)))
 280                 panfrost_bo_mmap(bo);
 281         else if (flags & PAN_BO_INVISIBLE) {
 282                 if (pan_debug & PAN_DBG_TRACE)
 283                         pandecode_inject_mmap(bo->gpu, NULL, bo->size, NULL);
 284         }
 285
 286         pipe_reference_init(&bo->reference, 1);
 287         return bo;
 288 }
 289
 290 void
 291 panfrost_bo_release(struct panfrost_bo *bo, bool cacheable)
 292 {
 293         if (!bo)
 294                 return;
 295
 296         struct drm_gem_close gem_close = { .handle = bo->gem_handle };
 297         int ret;
 298
 299         /* Rather than freeing the BO now, we'll cache the BO for later
 300          * allocations if we're allowed to */
 301
 302         panfrost_bo_munmap(bo);
 303
 304         if (cacheable) {
 305                 bool cached = panfrost_bo_cache_put(bo);
 306
 307                 if (cached)
 308                         return;
 309         }
 310
 311         /* Otherwise, if the BO wasn't cached, we'll legitimately free the BO */
 312
 313         ret = drmIoctl(bo->screen->fd, DRM_IOCTL_GEM_CLOSE, &gem_close);
 314         if (ret) {
 315                 fprintf(stderr, "DRM_IOCTL_GEM_CLOSE failed: %m\n");
 316                 assert(0);
 317         }
 318
 319         ralloc_free(bo);
 320 }
 321
 322 void
 323 panfrost_bo_reference(struct panfrost_bo *bo)
 324 {
 325         if (bo)
 326                 pipe_reference(NULL, &bo->reference);
 327 }
 328
 329 void
 330 panfrost_bo_unreference(struct panfrost_bo *bo)
 331 {
 332         if (!bo)
 333                 return;
 334
 335         /* When the reference count goes to zero, we need to cleanup */
 336
 337         if (pipe_reference(&bo->reference, NULL))
 338                 panfrost_bo_release(bo, true);
 339 }
 340
 341 struct panfrost_bo *
 342 panfrost_bo_import(struct panfrost_screen *screen, int fd)
 343 {
 344         struct panfrost_bo *bo = rzalloc(screen, struct panfrost_bo);
 345         struct drm_panfrost_get_bo_offset get_bo_offset = {0,};
 346         ASSERTED int ret;
 347         unsigned gem_handle;
 348
 349         ret = drmPrimeFDToHandle(screen->fd, fd, &gem_handle);
 350         assert(!ret);
 351
 352         get_bo_offset.handle = gem_handle;
 353         ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_GET_BO_OFFSET, &get_bo_offset);
 354         assert(!ret);
 355
 356         bo->screen = screen;
 357         bo->gem_handle = gem_handle;
 358         bo->gpu = (mali_ptr) get_bo_offset.offset;
 359         bo->size = lseek(fd, 0, SEEK_END);
 360         assert(bo->size > 0);
 361         pipe_reference_init(&bo->reference, 1);
 362
 363         // TODO map and unmap on demand?
 364         panfrost_bo_mmap(bo);
 365         return bo;
 366 }
 367
 368 int
 369 panfrost_bo_export(struct panfrost_bo *bo)
 370 {
 371         struct drm_prime_handle args = {
 372                 .handle = bo->gem_handle,
 373                 .flags = DRM_CLOEXEC,
 374         };
 375
 376         int ret = drmIoctl(bo->screen->fd, DRM_IOCTL_PRIME_HANDLE_TO_FD, &args);
 377         if (ret == -1)
 378                 return -1;
 379
 380         return args.fd;
 381 }
 382