/* * Copyright 2019 Collabora, Ltd. * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the "Software"), * to deal in the Software without restriction, including without limitation * the rights to use, copy, modify, merge, publish, distribute, sublicense, * and/or sell copies of the Software, and to permit persons to whom the * Software is furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice (including the next * paragraph) shall be included in all copies or substantial portions of the * Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. * * Authors (Collabora): * Alyssa Rosenzweig */ #include #include #include #include #include #include "drm-uapi/panfrost_drm.h" #include "pan_bo.h" #include "pan_screen.h" #include "pan_util.h" #include "pandecode/decode.h" #include "os/os_mman.h" #include "util/u_inlines.h" #include "util/u_math.h" /* This file implements a userspace BO cache. Allocating and freeing * GPU-visible buffers is very expensive, and even the extra kernel roundtrips * adds more work than we would like at this point. So caching BOs in userspace * solves both of these problems and does not require kernel updates. * * Cached BOs are sorted into a bucket based on rounding their size down to the * nearest power-of-two. Each bucket contains a linked list of free panfrost_bo * objects. Putting a BO into the cache is accomplished by adding it to the * corresponding bucket. Getting a BO from the cache consists of finding the * appropriate bucket and sorting. A cache eviction is a kernel-level free of a * BO and removing it from the bucket. We special case evicting all BOs from * the cache, since that's what helpful in practice and avoids extra logic * around the linked list. */ static struct panfrost_bo * panfrost_bo_alloc(struct panfrost_screen *screen, size_t size, uint32_t flags) { struct drm_panfrost_create_bo create_bo = { .size = size }; struct panfrost_bo *bo; int ret; if (screen->kernel_version->version_major > 1 || screen->kernel_version->version_minor >= 1) { if (flags & PAN_BO_GROWABLE) create_bo.flags |= PANFROST_BO_HEAP; if (!(flags & PAN_BO_EXECUTE)) create_bo.flags |= PANFROST_BO_NOEXEC; } ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_CREATE_BO, &create_bo); if (ret) { DBG("DRM_IOCTL_PANFROST_CREATE_BO failed: %m\n"); return NULL; } bo = rzalloc(screen, struct panfrost_bo); assert(bo); bo->size = create_bo.size; bo->gpu = create_bo.offset; bo->gem_handle = create_bo.handle; bo->flags = flags; bo->screen = screen; return bo; } static void panfrost_bo_free(struct panfrost_bo *bo) { struct drm_gem_close gem_close = { .handle = bo->gem_handle }; int ret; ret = drmIoctl(bo->screen->fd, DRM_IOCTL_GEM_CLOSE, &gem_close); if (ret) { fprintf(stderr, "DRM_IOCTL_GEM_CLOSE failed: %m\n"); assert(0); } ralloc_free(bo); } /* Returns true if the BO is ready, false otherwise. * access_type is encoding the type of access one wants to ensure is done. * Say you want to make sure all writers are done writing, you should pass * PAN_BO_ACCESS_WRITE. * If you want to wait for all users, you should pass PAN_BO_ACCESS_RW. * PAN_BO_ACCESS_READ would work too as waiting for readers implies * waiting for writers as well, but we want to make things explicit and waiting * only for readers is impossible. */ bool panfrost_bo_wait(struct panfrost_bo *bo, int64_t timeout_ns, uint32_t access_type) { struct drm_panfrost_wait_bo req = { .handle = bo->gem_handle, .timeout_ns = timeout_ns, }; int ret; assert(access_type == PAN_BO_ACCESS_WRITE || access_type == PAN_BO_ACCESS_RW); /* If the BO has been exported or imported we can't rely on the cached * state, we need to call the WAIT_BO ioctl. */ if (!(bo->flags & (PAN_BO_IMPORTED | PAN_BO_EXPORTED))) { /* If ->gpu_access is 0, the BO is idle, no need to wait. */ if (!bo->gpu_access) return true; /* If the caller only wants to wait for writers and no * writes are pending, we don't have to wait. */ if (access_type == PAN_BO_ACCESS_WRITE && !(bo->gpu_access & PAN_BO_ACCESS_WRITE)) return true; } /* The ioctl returns >= 0 value when the BO we are waiting for is ready * -1 otherwise. */ ret = drmIoctl(bo->screen->fd, DRM_IOCTL_PANFROST_WAIT_BO, &req); if (ret != -1) { /* Set gpu_access to 0 so that the next call to bo_wait() * doesn't have to call the WAIT_BO ioctl. */ bo->gpu_access = 0; return true; } /* If errno is not ETIMEDOUT or EBUSY that means the handle we passed * is invalid, which shouldn't happen here. */ assert(errno == ETIMEDOUT || errno == EBUSY); return false; } /* Helper to calculate the bucket index of a BO */ static unsigned pan_bucket_index(unsigned size) { /* Round down to POT to compute a bucket index */ unsigned bucket_index = util_logbase2(size); /* Clamp the bucket index; all huge allocations will be * sorted into the largest bucket */ bucket_index = MIN2(bucket_index, MAX_BO_CACHE_BUCKET); /* The minimum bucket size must equal the minimum allocation * size; the maximum we clamped */ assert(bucket_index >= MIN_BO_CACHE_BUCKET); assert(bucket_index <= MAX_BO_CACHE_BUCKET); /* Reindex from 0 */ return (bucket_index - MIN_BO_CACHE_BUCKET); } static struct list_head * pan_bucket(struct panfrost_screen *screen, unsigned size) { return &screen->bo_cache.buckets[pan_bucket_index(size)]; } /* Tries to fetch a BO of sufficient size with the appropriate flags from the * BO cache. If it succeeds, it returns that BO and removes the BO from the * cache. If it fails, it returns NULL signaling the caller to allocate a new * BO. */ static struct panfrost_bo * panfrost_bo_cache_fetch(struct panfrost_screen *screen, size_t size, uint32_t flags, bool dontwait) { pthread_mutex_lock(&screen->bo_cache.lock); struct list_head *bucket = pan_bucket(screen, size); struct panfrost_bo *bo = NULL; /* Iterate the bucket looking for something suitable */ list_for_each_entry_safe(struct panfrost_bo, entry, bucket, bucket_link) { if (entry->size < size || entry->flags != flags) continue; if (!panfrost_bo_wait(entry, dontwait ? 0 : INT64_MAX, PAN_BO_ACCESS_RW)) continue; struct drm_panfrost_madvise madv = { .handle = entry->gem_handle, .madv = PANFROST_MADV_WILLNEED, }; int ret; /* This one works, splice it out of the cache */ list_del(&entry->bucket_link); list_del(&entry->lru_link); ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_MADVISE, &madv); if (!ret && !madv.retained) { panfrost_bo_free(entry); continue; } /* Let's go! */ bo = entry; break; } pthread_mutex_unlock(&screen->bo_cache.lock); return bo; } static void panfrost_bo_cache_evict_stale_bos(struct panfrost_screen *screen) { struct timespec time; clock_gettime(CLOCK_MONOTONIC, &time); list_for_each_entry_safe(struct panfrost_bo, entry, &screen->bo_cache.lru, lru_link) { /* We want all entries that have been used more than 1 sec * ago to be dropped, others can be kept. * Note the <= 2 check and not <= 1. It's here to account for * the fact that we're only testing ->tv_sec, not ->tv_nsec. * That means we might keep entries that are between 1 and 2 * seconds old, but we don't really care, as long as unused BOs * are dropped at some point. */ if (time.tv_sec - entry->last_used <= 2) break; list_del(&entry->bucket_link); list_del(&entry->lru_link); panfrost_bo_free(entry); } } /* Tries to add a BO to the cache. Returns if it was * successful */ static bool panfrost_bo_cache_put(struct panfrost_bo *bo) { struct panfrost_screen *screen = bo->screen; if (bo->flags & PAN_BO_DONT_REUSE) return false; pthread_mutex_lock(&screen->bo_cache.lock); struct list_head *bucket = pan_bucket(screen, bo->size); struct drm_panfrost_madvise madv; struct timespec time; madv.handle = bo->gem_handle; madv.madv = PANFROST_MADV_DONTNEED; madv.retained = 0; drmIoctl(screen->fd, DRM_IOCTL_PANFROST_MADVISE, &madv); /* Add us to the bucket */ list_addtail(&bo->bucket_link, bucket); /* Add us to the LRU list and update the last_used field. */ list_addtail(&bo->lru_link, &screen->bo_cache.lru); clock_gettime(CLOCK_MONOTONIC, &time); bo->last_used = time.tv_sec; /* Let's do some cleanup in the BO cache while we hold the * lock. */ panfrost_bo_cache_evict_stale_bos(screen); pthread_mutex_unlock(&screen->bo_cache.lock); return true; } /* Evicts all BOs from the cache. Called during context * destroy or during low-memory situations (to free up * memory that may be unused by us just sitting in our * cache, but still reserved from the perspective of the * OS) */ void panfrost_bo_cache_evict_all( struct panfrost_screen *screen) { pthread_mutex_lock(&screen->bo_cache.lock); for (unsigned i = 0; i < ARRAY_SIZE(screen->bo_cache.buckets); ++i) { struct list_head *bucket = &screen->bo_cache.buckets[i]; list_for_each_entry_safe(struct panfrost_bo, entry, bucket, bucket_link) { list_del(&entry->bucket_link); list_del(&entry->lru_link); panfrost_bo_free(entry); } } pthread_mutex_unlock(&screen->bo_cache.lock); } void panfrost_bo_mmap(struct panfrost_bo *bo) { struct drm_panfrost_mmap_bo mmap_bo = { .handle = bo->gem_handle }; int ret; if (bo->cpu) return; ret = drmIoctl(bo->screen->fd, DRM_IOCTL_PANFROST_MMAP_BO, &mmap_bo); if (ret) { fprintf(stderr, "DRM_IOCTL_PANFROST_MMAP_BO failed: %m\n"); assert(0); } bo->cpu = os_mmap(NULL, bo->size, PROT_READ | PROT_WRITE, MAP_SHARED, bo->screen->fd, mmap_bo.offset); if (bo->cpu == MAP_FAILED) { fprintf(stderr, "mmap failed: %p %m\n", bo->cpu); assert(0); } /* Record the mmap if we're tracing */ if (pan_debug & (PAN_DBG_TRACE | PAN_DBG_SYNC)) pandecode_inject_mmap(bo->gpu, bo->cpu, bo->size, NULL); } static void panfrost_bo_munmap(struct panfrost_bo *bo) { if (!bo->cpu) return; if (os_munmap((void *) (uintptr_t)bo->cpu, bo->size)) { perror("munmap"); abort(); } bo->cpu = NULL; } struct panfrost_bo * panfrost_bo_create(struct panfrost_screen *screen, size_t size, uint32_t flags) { struct panfrost_bo *bo; /* Kernel will fail (confusingly) with EPERM otherwise */ assert(size > 0); /* To maximize BO cache usage, don't allocate tiny BOs */ size = MAX2(size, 4096); /* GROWABLE BOs cannot be mmapped */ if (flags & PAN_BO_GROWABLE) assert(flags & PAN_BO_INVISIBLE); /* Before creating a BO, we first want to check the cache but without * waiting for BO readiness (BOs in the cache can still be referenced * by jobs that are not finished yet). * If the cached allocation fails we fall back on fresh BO allocation, * and if that fails too, we try one more time to allocate from the * cache, but this time we accept to wait. */ bo = panfrost_bo_cache_fetch(screen, size, flags, true); if (!bo) bo = panfrost_bo_alloc(screen, size, flags); if (!bo) bo = panfrost_bo_cache_fetch(screen, size, flags, false); if (!bo) fprintf(stderr, "BO creation failed\n"); assert(bo); /* Only mmap now if we know we need to. For CPU-invisible buffers, we * never map since we don't care about their contents; they're purely * for GPU-internal use. But we do trace them anyway. */ if (!(flags & (PAN_BO_INVISIBLE | PAN_BO_DELAY_MMAP))) panfrost_bo_mmap(bo); else if (flags & PAN_BO_INVISIBLE) { if (pan_debug & (PAN_DBG_TRACE | PAN_DBG_SYNC)) pandecode_inject_mmap(bo->gpu, NULL, bo->size, NULL); } pipe_reference_init(&bo->reference, 1); pthread_mutex_lock(&screen->active_bos_lock); _mesa_set_add(bo->screen->active_bos, bo); pthread_mutex_unlock(&screen->active_bos_lock); return bo; } void panfrost_bo_reference(struct panfrost_bo *bo) { if (bo) pipe_reference(NULL, &bo->reference); } void panfrost_bo_unreference(struct panfrost_bo *bo) { if (!bo) return; if (!pipe_reference(&bo->reference, NULL)) return; struct panfrost_screen *screen = bo->screen; pthread_mutex_lock(&screen->active_bos_lock); /* Someone might have imported this BO while we were waiting for the * lock, let's make sure it's still not referenced before freeing it. */ if (!pipe_is_referenced(&bo->reference)) { _mesa_set_remove_key(bo->screen->active_bos, bo); /* When the reference count goes to zero, we need to cleanup */ panfrost_bo_munmap(bo); /* Rather than freeing the BO now, we'll cache the BO for later * allocations if we're allowed to. */ if (!panfrost_bo_cache_put(bo)) panfrost_bo_free(bo); } pthread_mutex_unlock(&screen->active_bos_lock); } struct panfrost_bo * panfrost_bo_import(struct panfrost_screen *screen, int fd) { struct panfrost_bo *bo, *newbo = rzalloc(screen, struct panfrost_bo); struct drm_panfrost_get_bo_offset get_bo_offset = {0,}; struct set_entry *entry; ASSERTED int ret; unsigned gem_handle; newbo->screen = screen; ret = drmPrimeFDToHandle(screen->fd, fd, &gem_handle); assert(!ret); newbo->gem_handle = gem_handle; pthread_mutex_lock(&screen->active_bos_lock); entry = _mesa_set_search_or_add(screen->active_bos, newbo); assert(entry); bo = (struct panfrost_bo *)entry->key; if (newbo == bo) { get_bo_offset.handle = gem_handle; ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_GET_BO_OFFSET, &get_bo_offset); assert(!ret); newbo->gpu = (mali_ptr) get_bo_offset.offset; newbo->size = lseek(fd, 0, SEEK_END); newbo->flags |= PAN_BO_DONT_REUSE | PAN_BO_IMPORTED; assert(newbo->size > 0); pipe_reference_init(&newbo->reference, 1); // TODO map and unmap on demand? panfrost_bo_mmap(newbo); } else { ralloc_free(newbo); /* !pipe_is_referenced(&bo->reference) can happen if the BO * was being released but panfrost_bo_import() acquired the * lock before panfrost_bo_unreference(). In that case, refcnt * is 0 and we can't use panfrost_bo_reference() directly, we * have to re-initialize it with pipe_reference_init(). * Note that panfrost_bo_unreference() checks * pipe_is_referenced() value just after acquiring the lock to * make sure the object is not freed if panfrost_bo_import() * acquired it in the meantime. */ if (!pipe_is_referenced(&bo->reference)) pipe_reference_init(&newbo->reference, 1); else panfrost_bo_reference(bo); assert(bo->cpu); } pthread_mutex_unlock(&screen->active_bos_lock); return bo; } int panfrost_bo_export(struct panfrost_bo *bo) { struct drm_prime_handle args = { .handle = bo->gem_handle, .flags = DRM_CLOEXEC, }; int ret = drmIoctl(bo->screen->fd, DRM_IOCTL_PRIME_HANDLE_TO_FD, &args); if (ret == -1) return -1; bo->flags |= PAN_BO_DONT_REUSE | PAN_BO_EXPORTED; return args.fd; }