/* * Copyright (C) 2012-2018 Rob Clark * * 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: * Rob Clark */ #include #include #include "util/hash_table.h" #include "util/set.h" #include "util/slab.h" #include "drm/freedreno_ringbuffer.h" #include "msm_priv.h" /* The legacy implementation of submit/ringbuffer, which still does the * traditional reloc and cmd tracking */ #define INIT_SIZE 0x1000 static pthread_mutex_t idx_lock = PTHREAD_MUTEX_INITIALIZER; struct msm_submit { struct fd_submit base; DECLARE_ARRAY(struct drm_msm_gem_submit_bo, submit_bos); DECLARE_ARRAY(struct fd_bo *, bos); unsigned seqno; /* maps fd_bo to idx in bos table: */ struct hash_table *bo_table; struct slab_mempool ring_pool; /* hash-set of associated rings: */ struct set *ring_set; struct fd_ringbuffer *primary; /* Allow for sub-allocation of stateobj ring buffers (ie. sharing * the same underlying bo).. * * We also rely on previous stateobj having been fully constructed * so we can reclaim extra space at it's end. */ struct fd_ringbuffer *suballoc_ring; }; FD_DEFINE_CAST(fd_submit, msm_submit); /* for FD_RINGBUFFER_GROWABLE rb's, tracks the 'finalized' cmdstream buffers * and sizes. Ie. a finalized buffer can have no more commands appended to * it. */ struct msm_cmd { struct fd_bo *ring_bo; unsigned size; DECLARE_ARRAY(struct drm_msm_gem_submit_reloc, relocs); }; static struct msm_cmd * cmd_new(struct fd_bo *ring_bo) { struct msm_cmd *cmd = malloc(sizeof(*cmd)); cmd->ring_bo = fd_bo_ref(ring_bo); cmd->size = 0; cmd->nr_relocs = cmd->max_relocs = 0; cmd->relocs = NULL; return cmd; } static void cmd_free(struct msm_cmd *cmd) { fd_bo_del(cmd->ring_bo); free(cmd); } /* for _FD_RINGBUFFER_OBJECT rb's we need to track the bo's and flags to * later copy into the submit when the stateobj rb is later referenced by * a regular rb: */ struct msm_reloc_bo { struct fd_bo *bo; unsigned flags; }; struct msm_ringbuffer { struct fd_ringbuffer base; /* for FD_RINGBUFFER_STREAMING rb's which are sub-allocated */ unsigned offset; union { /* for _FD_RINGBUFFER_OBJECT case: */ struct { struct fd_pipe *pipe; DECLARE_ARRAY(struct msm_reloc_bo, reloc_bos); struct set *ring_set; }; /* for other cases: */ struct { struct fd_submit *submit; DECLARE_ARRAY(struct msm_cmd *, cmds); }; } u; struct msm_cmd *cmd; /* current cmd */ struct fd_bo *ring_bo; }; FD_DEFINE_CAST(fd_ringbuffer, msm_ringbuffer); static void finalize_current_cmd(struct fd_ringbuffer *ring); static struct fd_ringbuffer * msm_ringbuffer_init( struct msm_ringbuffer *msm_ring, uint32_t size, enum fd_ringbuffer_flags flags); /* add (if needed) bo to submit and return index: */ static uint32_t append_bo(struct msm_submit *submit, struct fd_bo *bo, uint32_t flags) { struct msm_bo *msm_bo = to_msm_bo(bo); uint32_t idx; pthread_mutex_lock(&idx_lock); if (likely(msm_bo->current_submit_seqno == submit->seqno)) { idx = msm_bo->idx; } else { uint32_t hash = _mesa_hash_pointer(bo); struct hash_entry *entry; entry = _mesa_hash_table_search_pre_hashed(submit->bo_table, hash, bo); if (entry) { /* found */ idx = (uint32_t)(uintptr_t)entry->data; } else { idx = APPEND(submit, submit_bos); idx = APPEND(submit, bos); submit->submit_bos[idx].flags = 0; submit->submit_bos[idx].handle = bo->handle; submit->submit_bos[idx].presumed = 0; submit->bos[idx] = fd_bo_ref(bo); _mesa_hash_table_insert_pre_hashed(submit->bo_table, hash, bo, (void *)(uintptr_t)idx); } msm_bo->current_submit_seqno = submit->seqno; msm_bo->idx = idx; } pthread_mutex_unlock(&idx_lock); if (flags & FD_RELOC_READ) submit->submit_bos[idx].flags |= MSM_SUBMIT_BO_READ; if (flags & FD_RELOC_WRITE) submit->submit_bos[idx].flags |= MSM_SUBMIT_BO_WRITE; return idx; } static void append_ring(struct set *set, struct fd_ringbuffer *ring) { uint32_t hash = _mesa_hash_pointer(ring); if (!_mesa_set_search_pre_hashed(set, hash, ring)) { fd_ringbuffer_ref(ring); _mesa_set_add_pre_hashed(set, hash, ring); } } static void msm_submit_suballoc_ring_bo(struct fd_submit *submit, struct msm_ringbuffer *msm_ring, uint32_t size) { struct msm_submit *msm_submit = to_msm_submit(submit); unsigned suballoc_offset = 0; struct fd_bo *suballoc_bo = NULL; if (msm_submit->suballoc_ring) { struct msm_ringbuffer *suballoc_ring = to_msm_ringbuffer(msm_submit->suballoc_ring); suballoc_bo = suballoc_ring->ring_bo; suballoc_offset = fd_ringbuffer_size(msm_submit->suballoc_ring) + suballoc_ring->offset; suballoc_offset = align(suballoc_offset, 0x10); if ((size + suballoc_offset) > suballoc_bo->size) { suballoc_bo = NULL; } } if (!suballoc_bo) { // TODO possibly larger size for streaming bo? msm_ring->ring_bo = fd_bo_new_ring( submit->pipe->dev, 0x8000, 0); msm_ring->offset = 0; } else { msm_ring->ring_bo = fd_bo_ref(suballoc_bo); msm_ring->offset = suballoc_offset; } struct fd_ringbuffer *old_suballoc_ring = msm_submit->suballoc_ring; msm_submit->suballoc_ring = fd_ringbuffer_ref(&msm_ring->base); if (old_suballoc_ring) fd_ringbuffer_del(old_suballoc_ring); } static struct fd_ringbuffer * msm_submit_new_ringbuffer(struct fd_submit *submit, uint32_t size, enum fd_ringbuffer_flags flags) { struct msm_submit *msm_submit = to_msm_submit(submit); struct msm_ringbuffer *msm_ring; msm_ring = slab_alloc_st(&msm_submit->ring_pool); msm_ring->u.submit = submit; /* NOTE: needs to be before _suballoc_ring_bo() since it could * increment the refcnt of the current ring */ msm_ring->base.refcnt = 1; if (flags & FD_RINGBUFFER_STREAMING) { msm_submit_suballoc_ring_bo(submit, msm_ring, size); } else { if (flags & FD_RINGBUFFER_GROWABLE) size = INIT_SIZE; msm_ring->offset = 0; msm_ring->ring_bo = fd_bo_new_ring(submit->pipe->dev, size, 0); } if (!msm_ringbuffer_init(msm_ring, size, flags)) return NULL; if (flags & FD_RINGBUFFER_PRIMARY) { debug_assert(!msm_submit->primary); msm_submit->primary = fd_ringbuffer_ref(&msm_ring->base); } return &msm_ring->base; } static struct drm_msm_gem_submit_reloc * handle_stateobj_relocs(struct msm_submit *submit, struct msm_ringbuffer *ring) { struct msm_cmd *cmd = ring->cmd; struct drm_msm_gem_submit_reloc *relocs; relocs = malloc(cmd->nr_relocs * sizeof(*relocs)); for (unsigned i = 0; i < cmd->nr_relocs; i++) { unsigned idx = cmd->relocs[i].reloc_idx; struct fd_bo *bo = ring->u.reloc_bos[idx].bo; unsigned flags = 0; if (ring->u.reloc_bos[idx].flags & MSM_SUBMIT_BO_READ) flags |= FD_RELOC_READ; if (ring->u.reloc_bos[idx].flags & MSM_SUBMIT_BO_WRITE) flags |= FD_RELOC_WRITE; relocs[i] = cmd->relocs[i]; relocs[i].reloc_idx = append_bo(submit, bo, flags); } return relocs; } static int msm_submit_flush(struct fd_submit *submit, int in_fence_fd, int *out_fence_fd, uint32_t *out_fence) { struct msm_submit *msm_submit = to_msm_submit(submit); struct msm_pipe *msm_pipe = to_msm_pipe(submit->pipe); struct drm_msm_gem_submit req = { .flags = msm_pipe->pipe, .queueid = msm_pipe->queue_id, }; int ret; debug_assert(msm_submit->primary); finalize_current_cmd(msm_submit->primary); append_ring(msm_submit->ring_set, msm_submit->primary); struct set_entry *entry; unsigned nr_cmds = 0; unsigned nr_objs = 0; set_foreach(msm_submit->ring_set, entry) { struct fd_ringbuffer *ring = (void *)entry->key; if (ring->flags & _FD_RINGBUFFER_OBJECT) { nr_cmds += 1; nr_objs += 1; } else { if (ring != msm_submit->primary) finalize_current_cmd(ring); nr_cmds += to_msm_ringbuffer(ring)->u.nr_cmds; } } void *obj_relocs[nr_objs]; struct drm_msm_gem_submit_cmd cmds[nr_cmds]; unsigned i = 0, o = 0; set_foreach(msm_submit->ring_set, entry) { struct fd_ringbuffer *ring = (void *)entry->key; struct msm_ringbuffer *msm_ring = to_msm_ringbuffer(ring); debug_assert(i < nr_cmds); // TODO handle relocs: if (ring->flags & _FD_RINGBUFFER_OBJECT) { debug_assert(o < nr_objs); void *relocs = handle_stateobj_relocs(msm_submit, msm_ring); obj_relocs[o++] = relocs; cmds[i].type = MSM_SUBMIT_CMD_IB_TARGET_BUF; cmds[i].submit_idx = append_bo(msm_submit, msm_ring->ring_bo, FD_RELOC_READ); cmds[i].submit_offset = msm_ring->offset; cmds[i].size = offset_bytes(ring->cur, ring->start); cmds[i].pad = 0; cmds[i].nr_relocs = msm_ring->cmd->nr_relocs; cmds[i].relocs = VOID2U64(relocs); i++; } else { for (unsigned j = 0; j < msm_ring->u.nr_cmds; j++) { if (ring->flags & FD_RINGBUFFER_PRIMARY) { cmds[i].type = MSM_SUBMIT_CMD_BUF; } else { cmds[i].type = MSM_SUBMIT_CMD_IB_TARGET_BUF; } cmds[i].submit_idx = append_bo(msm_submit, msm_ring->u.cmds[j]->ring_bo, FD_RELOC_READ); cmds[i].submit_offset = msm_ring->offset; cmds[i].size = msm_ring->u.cmds[j]->size; cmds[i].pad = 0; cmds[i].nr_relocs = msm_ring->u.cmds[j]->nr_relocs; cmds[i].relocs = VOID2U64(msm_ring->u.cmds[j]->relocs); i++; } } } if (in_fence_fd != -1) { req.flags |= MSM_SUBMIT_FENCE_FD_IN | MSM_SUBMIT_NO_IMPLICIT; req.fence_fd = in_fence_fd; } if (out_fence_fd) { req.flags |= MSM_SUBMIT_FENCE_FD_OUT; } /* needs to be after get_cmd() as that could create bos/cmds table: */ req.bos = VOID2U64(msm_submit->submit_bos), req.nr_bos = msm_submit->nr_submit_bos; req.cmds = VOID2U64(cmds), req.nr_cmds = nr_cmds; DEBUG_MSG("nr_cmds=%u, nr_bos=%u", req.nr_cmds, req.nr_bos); ret = drmCommandWriteRead(submit->pipe->dev->fd, DRM_MSM_GEM_SUBMIT, &req, sizeof(req)); if (ret) { ERROR_MSG("submit failed: %d (%s)", ret, strerror(errno)); msm_dump_submit(&req); } else if (!ret) { if (out_fence) *out_fence = req.fence; if (out_fence_fd) *out_fence_fd = req.fence_fd; } for (unsigned o = 0; o < nr_objs; o++) free(obj_relocs[o]); return ret; } static void unref_rings(struct set_entry *entry) { struct fd_ringbuffer *ring = (void *)entry->key; fd_ringbuffer_del(ring); } static void msm_submit_destroy(struct fd_submit *submit) { struct msm_submit *msm_submit = to_msm_submit(submit); if (msm_submit->primary) fd_ringbuffer_del(msm_submit->primary); if (msm_submit->suballoc_ring) fd_ringbuffer_del(msm_submit->suballoc_ring); _mesa_hash_table_destroy(msm_submit->bo_table, NULL); _mesa_set_destroy(msm_submit->ring_set, unref_rings); // TODO it would be nice to have a way to debug_assert() if all // rb's haven't been free'd back to the slab, because that is // an indication that we are leaking bo's slab_destroy(&msm_submit->ring_pool); for (unsigned i = 0; i < msm_submit->nr_bos; i++) fd_bo_del(msm_submit->bos[i]); free(msm_submit->submit_bos); free(msm_submit->bos); free(msm_submit); } static const struct fd_submit_funcs submit_funcs = { .new_ringbuffer = msm_submit_new_ringbuffer, .flush = msm_submit_flush, .destroy = msm_submit_destroy, }; struct fd_submit * msm_submit_new(struct fd_pipe *pipe) { struct msm_submit *msm_submit = calloc(1, sizeof(*msm_submit)); struct fd_submit *submit; static unsigned submit_cnt = 0; msm_submit->seqno = ++submit_cnt; msm_submit->bo_table = _mesa_hash_table_create(NULL, _mesa_hash_pointer, _mesa_key_pointer_equal); msm_submit->ring_set = _mesa_set_create(NULL, _mesa_hash_pointer, _mesa_key_pointer_equal); // TODO tune size: slab_create(&msm_submit->ring_pool, sizeof(struct msm_ringbuffer), 16); submit = &msm_submit->base; submit->pipe = pipe; submit->funcs = &submit_funcs; return submit; } static void finalize_current_cmd(struct fd_ringbuffer *ring) { struct msm_ringbuffer *msm_ring = to_msm_ringbuffer(ring); debug_assert(!(ring->flags & _FD_RINGBUFFER_OBJECT)); if (!msm_ring->cmd) return; debug_assert(msm_ring->cmd->ring_bo == msm_ring->ring_bo); unsigned idx = APPEND(&msm_ring->u, cmds); msm_ring->u.cmds[idx] = msm_ring->cmd; msm_ring->cmd = NULL; msm_ring->u.cmds[idx]->size = offset_bytes(ring->cur, ring->start); } static void msm_ringbuffer_grow(struct fd_ringbuffer *ring, uint32_t size) { struct msm_ringbuffer *msm_ring = to_msm_ringbuffer(ring); struct fd_pipe *pipe = msm_ring->u.submit->pipe; debug_assert(ring->flags & FD_RINGBUFFER_GROWABLE); finalize_current_cmd(ring); fd_bo_del(msm_ring->ring_bo); msm_ring->ring_bo = fd_bo_new_ring(pipe->dev, size, 0); msm_ring->cmd = cmd_new(msm_ring->ring_bo); ring->start = fd_bo_map(msm_ring->ring_bo); ring->end = &(ring->start[size/4]); ring->cur = ring->start; ring->size = size; } static void msm_ringbuffer_emit_reloc(struct fd_ringbuffer *ring, const struct fd_reloc *reloc) { struct msm_ringbuffer *msm_ring = to_msm_ringbuffer(ring); struct fd_pipe *pipe; unsigned reloc_idx; if (ring->flags & _FD_RINGBUFFER_OBJECT) { unsigned idx = APPEND(&msm_ring->u, reloc_bos); msm_ring->u.reloc_bos[idx].bo = fd_bo_ref(reloc->bo); msm_ring->u.reloc_bos[idx].flags = reloc->flags; /* this gets fixed up at submit->flush() time, since this state- * object rb can be used with many different submits */ reloc_idx = idx; pipe = msm_ring->u.pipe; } else { struct msm_submit *msm_submit = to_msm_submit(msm_ring->u.submit); reloc_idx = append_bo(msm_submit, reloc->bo, reloc->flags); pipe = msm_ring->u.submit->pipe; } struct drm_msm_gem_submit_reloc *r; unsigned idx = APPEND(msm_ring->cmd, relocs); r = &msm_ring->cmd->relocs[idx]; r->reloc_idx = reloc_idx; r->reloc_offset = reloc->offset; r->or = reloc->or; r->shift = reloc->shift; r->submit_offset = offset_bytes(ring->cur, ring->start) + msm_ring->offset; ring->cur++; if (pipe->gpu_id >= 500) { idx = APPEND(msm_ring->cmd, relocs); r = &msm_ring->cmd->relocs[idx]; r->reloc_idx = reloc_idx; r->reloc_offset = reloc->offset; r->or = reloc->orhi; r->shift = reloc->shift - 32; r->submit_offset = offset_bytes(ring->cur, ring->start) + msm_ring->offset; ring->cur++; } } static void append_stateobj_rings(struct msm_submit *submit, struct fd_ringbuffer *target) { struct msm_ringbuffer *msm_target = to_msm_ringbuffer(target); debug_assert(target->flags & _FD_RINGBUFFER_OBJECT); struct set_entry *entry; set_foreach(msm_target->u.ring_set, entry) { struct fd_ringbuffer *ring = (void *)entry->key; append_ring(submit->ring_set, ring); if (ring->flags & _FD_RINGBUFFER_OBJECT) { append_stateobj_rings(submit, ring); } } } static uint32_t msm_ringbuffer_emit_reloc_ring(struct fd_ringbuffer *ring, struct fd_ringbuffer *target, uint32_t cmd_idx) { struct msm_ringbuffer *msm_target = to_msm_ringbuffer(target); struct msm_ringbuffer *msm_ring = to_msm_ringbuffer(ring); struct fd_bo *bo; uint32_t size; if ((target->flags & FD_RINGBUFFER_GROWABLE) && (cmd_idx < msm_target->u.nr_cmds)) { bo = msm_target->u.cmds[cmd_idx]->ring_bo; size = msm_target->u.cmds[cmd_idx]->size; } else { bo = msm_target->ring_bo; size = offset_bytes(target->cur, target->start); } msm_ringbuffer_emit_reloc(ring, &(struct fd_reloc){ .bo = bo, .flags = FD_RELOC_READ, .offset = msm_target->offset, }); if ((target->flags & _FD_RINGBUFFER_OBJECT) && !(ring->flags & _FD_RINGBUFFER_OBJECT)) { struct msm_submit *msm_submit = to_msm_submit(msm_ring->u.submit); append_stateobj_rings(msm_submit, target); } if (ring->flags & _FD_RINGBUFFER_OBJECT) { append_ring(msm_ring->u.ring_set, target); } else { struct msm_submit *msm_submit = to_msm_submit(msm_ring->u.submit); append_ring(msm_submit->ring_set, target); } return size; } static uint32_t msm_ringbuffer_cmd_count(struct fd_ringbuffer *ring) { if (ring->flags & FD_RINGBUFFER_GROWABLE) return to_msm_ringbuffer(ring)->u.nr_cmds + 1; return 1; } static void msm_ringbuffer_destroy(struct fd_ringbuffer *ring) { struct msm_ringbuffer *msm_ring = to_msm_ringbuffer(ring); fd_bo_del(msm_ring->ring_bo); if (msm_ring->cmd) cmd_free(msm_ring->cmd); if (ring->flags & _FD_RINGBUFFER_OBJECT) { for (unsigned i = 0; i < msm_ring->u.nr_reloc_bos; i++) { fd_bo_del(msm_ring->u.reloc_bos[i].bo); } _mesa_set_destroy(msm_ring->u.ring_set, unref_rings); free(msm_ring); } else { struct fd_submit *submit = msm_ring->u.submit; for (unsigned i = 0; i < msm_ring->u.nr_cmds; i++) { cmd_free(msm_ring->u.cmds[i]); } slab_free_st(&to_msm_submit(submit)->ring_pool, msm_ring); } } static const struct fd_ringbuffer_funcs ring_funcs = { .grow = msm_ringbuffer_grow, .emit_reloc = msm_ringbuffer_emit_reloc, .emit_reloc_ring = msm_ringbuffer_emit_reloc_ring, .cmd_count = msm_ringbuffer_cmd_count, .destroy = msm_ringbuffer_destroy, }; static inline struct fd_ringbuffer * msm_ringbuffer_init(struct msm_ringbuffer *msm_ring, uint32_t size, enum fd_ringbuffer_flags flags) { struct fd_ringbuffer *ring = &msm_ring->base; debug_assert(msm_ring->ring_bo); uint8_t *base = fd_bo_map(msm_ring->ring_bo); ring->start = (void *)(base + msm_ring->offset); ring->end = &(ring->start[size/4]); ring->cur = ring->start; ring->size = size; ring->flags = flags; ring->funcs = &ring_funcs; msm_ring->u.cmds = NULL; msm_ring->u.nr_cmds = msm_ring->u.max_cmds = 0; msm_ring->cmd = cmd_new(msm_ring->ring_bo); return ring; } struct fd_ringbuffer * msm_ringbuffer_new_object(struct fd_pipe *pipe, uint32_t size) { struct msm_ringbuffer *msm_ring = malloc(sizeof(*msm_ring)); msm_ring->u.pipe = pipe; msm_ring->offset = 0; msm_ring->ring_bo = fd_bo_new_ring(pipe->dev, size, 0); msm_ring->base.refcnt = 1; msm_ring->u.reloc_bos = NULL; msm_ring->u.nr_reloc_bos = msm_ring->u.max_reloc_bos = 0; msm_ring->u.ring_set = _mesa_set_create(NULL, _mesa_hash_pointer, _mesa_key_pointer_equal); return msm_ringbuffer_init(msm_ring, size, _FD_RINGBUFFER_OBJECT); }