From 6a9eb08a872ac0388aad2c901888888964f14559 Mon Sep 17 00:00:00 2001 From: Eric Anholt Date: Tue, 3 Jun 2008 09:27:37 -0700 Subject: Import bufmgr code to libdrm. Not yet hooked up to the build. --- libdrm/intel/intel_bufmgr_fake.c | 1177 ++++++++++++++++++++++++++++++++++++++ 1 file changed, 1177 insertions(+) create mode 100644 libdrm/intel/intel_bufmgr_fake.c (limited to 'libdrm/intel/intel_bufmgr_fake.c') diff --git a/libdrm/intel/intel_bufmgr_fake.c b/libdrm/intel/intel_bufmgr_fake.c new file mode 100644 index 00000000..2aed3d85 --- /dev/null +++ b/libdrm/intel/intel_bufmgr_fake.c @@ -0,0 +1,1177 @@ +/************************************************************************** + * + * Copyright 2006 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * 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, sub license, 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 NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS 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. + * + **************************************************************************/ + +/* Originally a fake version of the buffer manager so that we can + * prototype the changes in a driver fairly quickly, has been fleshed + * out to a fully functional interim solution. + * + * Basically wraps the old style memory management in the new + * programming interface, but is more expressive and avoids many of + * the bugs in the old texture manager. + */ +#include "mtypes.h" +#include "dri_bufmgr.h" +#include "intel_bufmgr_fake.h" +#include "drm.h" +#include "i915_drm.h" + +#include "simple_list.h" +#include "mm.h" +#include "imports.h" + +#define DBG(...) do { \ + if (bufmgr_fake->bufmgr.debug) \ + _mesa_printf(__VA_ARGS__); \ +} while (0) + +/* Internal flags: + */ +#define BM_NO_BACKING_STORE 0x00000001 +#define BM_NO_FENCE_SUBDATA 0x00000002 +#define BM_PINNED 0x00000004 + +/* Wrapper around mm.c's mem_block, which understands that you must + * wait for fences to expire before memory can be freed. This is + * specific to our use of memcpy for uploads - an upload that was + * processed through the command queue wouldn't need to care about + * fences. + */ +#define MAX_RELOCS 4096 + +struct fake_buffer_reloc +{ + /** Buffer object that the relocation points at. */ + dri_bo *target_buf; + /** Offset of the relocation entry within reloc_buf. */ + GLuint offset; + /** Cached value of the offset when we last performed this relocation. */ + GLuint last_target_offset; + /** Value added to target_buf's offset to get the relocation entry. */ + GLuint delta; + /** Cache domains the target buffer is read into. */ + uint32_t read_domains; + /** Cache domain the target buffer will have dirty cachelines in. */ + uint32_t write_domain; +}; + +struct block { + struct block *next, *prev; + struct mem_block *mem; /* BM_MEM_AGP */ + + /** + * Marks that the block is currently in the aperture and has yet to be + * fenced. + */ + unsigned on_hardware:1; + /** + * Marks that the block is currently fenced (being used by rendering) and + * can't be freed until @fence is passed. + */ + unsigned fenced:1; + + /** Fence cookie for the block. */ + unsigned fence; /* Split to read_fence, write_fence */ + + dri_bo *bo; + void *virtual; +}; + +typedef struct _bufmgr_fake { + dri_bufmgr bufmgr; + + unsigned long low_offset; + unsigned long size; + void *virtual; + + struct mem_block *heap; + struct block lru; /* only allocated, non-fence-pending blocks here */ + + unsigned buf_nr; /* for generating ids */ + + struct block on_hardware; /* after bmValidateBuffers */ + struct block fenced; /* after bmFenceBuffers (mi_flush, emit irq, write dword) */ + /* then to bufmgr->lru or free() */ + + unsigned int last_fence; + + unsigned fail:1; + unsigned need_fence:1; + GLboolean thrashing; + + /** + * Driver callback to emit a fence, returning the cookie. + * + * Currently, this also requires that a write flush be emitted before + * emitting the fence, but this should change. + */ + unsigned int (*fence_emit)(void *private); + /** Driver callback to wait for a fence cookie to have passed. */ + int (*fence_wait)(void *private, unsigned int fence_cookie); + /** Driver-supplied argument to driver callbacks */ + void *driver_priv; + + GLboolean debug; + + GLboolean performed_rendering; + + /* keep track of the current total size of objects we have relocs for */ + unsigned long current_total_size; +} dri_bufmgr_fake; + +typedef struct _dri_bo_fake { + dri_bo bo; + + unsigned id; /* debug only */ + const char *name; + + unsigned dirty:1; + unsigned size_accounted:1; /*this buffers size has been accounted against the aperture */ + unsigned card_dirty:1; /* has the card written to this buffer - we make need to copy it back */ + unsigned int refcount; + /* Flags may consist of any of the DRM_BO flags, plus + * DRM_BO_NO_BACKING_STORE and BM_NO_FENCE_SUBDATA, which are the first two + * driver private flags. + */ + uint64_t flags; + /** Cache domains the target buffer is read into. */ + uint32_t read_domains; + /** Cache domain the target buffer will have dirty cachelines in. */ + uint32_t write_domain; + + unsigned int alignment; + GLboolean is_static, validated; + unsigned int map_count; + + /** relocation list */ + struct fake_buffer_reloc *relocs; + GLuint nr_relocs; + + struct block *block; + void *backing_store; + void (*invalidate_cb)(dri_bo *bo, void *ptr); + void *invalidate_ptr; +} dri_bo_fake; + +static int clear_fenced(dri_bufmgr_fake *bufmgr_fake, + unsigned int fence_cookie); + +static int dri_fake_check_aperture_space(dri_bo *bo); + +#define MAXFENCE 0x7fffffff + +static GLboolean FENCE_LTE( unsigned a, unsigned b ) +{ + if (a == b) + return GL_TRUE; + + if (a < b && b - a < (1<<24)) + return GL_TRUE; + + if (a > b && MAXFENCE - a + b < (1<<24)) + return GL_TRUE; + + return GL_FALSE; +} + +static unsigned int +_fence_emit_internal(dri_bufmgr_fake *bufmgr_fake) +{ + bufmgr_fake->last_fence = bufmgr_fake->fence_emit(bufmgr_fake->driver_priv); + return bufmgr_fake->last_fence; +} + +static void +_fence_wait_internal(dri_bufmgr_fake *bufmgr_fake, unsigned int cookie) +{ + int ret; + + ret = bufmgr_fake->fence_wait(bufmgr_fake->driver_priv, cookie); + if (ret != 0) { + _mesa_printf("%s:%d: Error %d waiting for fence.\n", + __FILE__, __LINE__); + abort(); + } + clear_fenced(bufmgr_fake, cookie); +} + +static GLboolean +_fence_test(dri_bufmgr_fake *bufmgr_fake, unsigned fence) +{ + /* Slight problem with wrap-around: + */ + return fence == 0 || FENCE_LTE(fence, bufmgr_fake->last_fence); +} + +/** + * Allocate a memory manager block for the buffer. + */ +static GLboolean +alloc_block(dri_bo *bo) +{ + dri_bo_fake *bo_fake = (dri_bo_fake *)bo; + dri_bufmgr_fake *bufmgr_fake= (dri_bufmgr_fake *)bo->bufmgr; + struct block *block = (struct block *)calloc(sizeof *block, 1); + unsigned int align_log2 = _mesa_ffs(bo_fake->alignment) - 1; + GLuint sz; + + if (!block) + return GL_FALSE; + + sz = (bo->size + bo_fake->alignment - 1) & ~(bo_fake->alignment - 1); + + block->mem = mmAllocMem(bufmgr_fake->heap, sz, align_log2, 0); + if (!block->mem) { + free(block); + return GL_FALSE; + } + + make_empty_list(block); + + /* Insert at head or at tail??? + */ + insert_at_tail(&bufmgr_fake->lru, block); + + block->virtual = bufmgr_fake->virtual + + block->mem->ofs - bufmgr_fake->low_offset; + block->bo = bo; + + bo_fake->block = block; + + return GL_TRUE; +} + +/* Release the card storage associated with buf: + */ +static void free_block(dri_bufmgr_fake *bufmgr_fake, struct block *block) +{ + dri_bo_fake *bo_fake; + DBG("free block %p %08x %d %d\n", block, block->mem->ofs, block->on_hardware, block->fenced); + + if (!block) + return; + + bo_fake = (dri_bo_fake *)block->bo; + if (!(bo_fake->flags & BM_NO_BACKING_STORE) && (bo_fake->card_dirty == 1)) { + memcpy(bo_fake->backing_store, block->virtual, block->bo->size); + bo_fake->card_dirty = 1; + bo_fake->dirty = 1; + } + + if (block->on_hardware) { + block->bo = NULL; + } + else if (block->fenced) { + block->bo = NULL; + } + else { + DBG(" - free immediately\n"); + remove_from_list(block); + + mmFreeMem(block->mem); + free(block); + } +} + +static void +alloc_backing_store(dri_bo *bo) +{ + dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr; + dri_bo_fake *bo_fake = (dri_bo_fake *)bo; + assert(!bo_fake->backing_store); + assert(!(bo_fake->flags & (BM_PINNED|BM_NO_BACKING_STORE))); + + bo_fake->backing_store = ALIGN_MALLOC(bo->size, 64); + + DBG("alloc_backing - buf %d %p %d\n", bo_fake->id, bo_fake->backing_store, bo->size); + assert(bo_fake->backing_store); +} + +static void +free_backing_store(dri_bo *bo) +{ + dri_bo_fake *bo_fake = (dri_bo_fake *)bo; + + if (bo_fake->backing_store) { + assert(!(bo_fake->flags & (BM_PINNED|BM_NO_BACKING_STORE))); + ALIGN_FREE(bo_fake->backing_store); + bo_fake->backing_store = NULL; + } +} + +static void +set_dirty(dri_bo *bo) +{ + dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr; + dri_bo_fake *bo_fake = (dri_bo_fake *)bo; + + if (bo_fake->flags & BM_NO_BACKING_STORE && bo_fake->invalidate_cb != NULL) + bo_fake->invalidate_cb(bo, bo_fake->invalidate_ptr); + + assert(!(bo_fake->flags & BM_PINNED)); + + DBG("set_dirty - buf %d\n", bo_fake->id); + bo_fake->dirty = 1; +} + +static GLboolean +evict_lru(dri_bufmgr_fake *bufmgr_fake, GLuint max_fence) +{ + struct block *block, *tmp; + + DBG("%s\n", __FUNCTION__); + + foreach_s(block, tmp, &bufmgr_fake->lru) { + dri_bo_fake *bo_fake = (dri_bo_fake *)block->bo; + + if (bo_fake != NULL && (bo_fake->flags & BM_NO_FENCE_SUBDATA)) + continue; + + if (block->fence && max_fence && !FENCE_LTE(block->fence, max_fence)) + return 0; + + set_dirty(&bo_fake->bo); + bo_fake->block = NULL; + + free_block(bufmgr_fake, block); + return GL_TRUE; + } + + return GL_FALSE; +} + +#define foreach_s_rev(ptr, t, list) \ + for(ptr=(list)->prev,t=(ptr)->prev; list != ptr; ptr=t, t=(t)->prev) + +static GLboolean +evict_mru(dri_bufmgr_fake *bufmgr_fake) +{ + struct block *block, *tmp; + + DBG("%s\n", __FUNCTION__); + + foreach_s_rev(block, tmp, &bufmgr_fake->lru) { + dri_bo_fake *bo_fake = (dri_bo_fake *)block->bo; + + if (bo_fake && (bo_fake->flags & BM_NO_FENCE_SUBDATA)) + continue; + + set_dirty(&bo_fake->bo); + bo_fake->block = NULL; + + free_block(bufmgr_fake, block); + return GL_TRUE; + } + + return GL_FALSE; +} + +/** + * Removes all objects from the fenced list older than the given fence. + */ +static int clear_fenced(dri_bufmgr_fake *bufmgr_fake, + unsigned int fence_cookie) +{ + struct block *block, *tmp; + int ret = 0; + + foreach_s(block, tmp, &bufmgr_fake->fenced) { + assert(block->fenced); + + if (_fence_test(bufmgr_fake, block->fence)) { + + block->fenced = 0; + + if (!block->bo) { + DBG("delayed free: offset %x sz %x\n", + block->mem->ofs, block->mem->size); + remove_from_list(block); + mmFreeMem(block->mem); + free(block); + } + else { + DBG("return to lru: offset %x sz %x\n", + block->mem->ofs, block->mem->size); + move_to_tail(&bufmgr_fake->lru, block); + } + + ret = 1; + } + else { + /* Blocks are ordered by fence, so if one fails, all from + * here will fail also: + */ + DBG("fence not passed: offset %x sz %x %d %d \n", + block->mem->ofs, block->mem->size, block->fence, bufmgr_fake->last_fence); + break; + } + } + + DBG("%s: %d\n", __FUNCTION__, ret); + return ret; +} + +static void fence_blocks(dri_bufmgr_fake *bufmgr_fake, unsigned fence) +{ + struct block *block, *tmp; + + foreach_s (block, tmp, &bufmgr_fake->on_hardware) { + DBG("Fence block %p (sz 0x%x ofs %x buf %p) with fence %d\n", block, + block->mem->size, block->mem->ofs, block->bo, fence); + block->fence = fence; + + block->on_hardware = 0; + block->fenced = 1; + + /* Move to tail of pending list here + */ + move_to_tail(&bufmgr_fake->fenced, block); + } + + assert(is_empty_list(&bufmgr_fake->on_hardware)); +} + +static GLboolean evict_and_alloc_block(dri_bo *bo) +{ + dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr; + dri_bo_fake *bo_fake = (dri_bo_fake *)bo; + + assert(bo_fake->block == NULL); + + /* Search for already free memory: + */ + if (alloc_block(bo)) + return GL_TRUE; + + /* If we're not thrashing, allow lru eviction to dig deeper into + * recently used textures. We'll probably be thrashing soon: + */ + if (!bufmgr_fake->thrashing) { + while (evict_lru(bufmgr_fake, 0)) + if (alloc_block(bo)) + return GL_TRUE; + } + + /* Keep thrashing counter alive? + */ + if (bufmgr_fake->thrashing) + bufmgr_fake->thrashing = 20; + + /* Wait on any already pending fences - here we are waiting for any + * freed memory that has been submitted to hardware and fenced to + * become available: + */ + while (!is_empty_list(&bufmgr_fake->fenced)) { + GLuint fence = bufmgr_fake->fenced.next->fence; + _fence_wait_internal(bufmgr_fake, fence); + + if (alloc_block(bo)) + return GL_TRUE; + } + + if (!is_empty_list(&bufmgr_fake->on_hardware)) { + while (!is_empty_list(&bufmgr_fake->fenced)) { + GLuint fence = bufmgr_fake->fenced.next->fence; + _fence_wait_internal(bufmgr_fake, fence); + } + + if (!bufmgr_fake->thrashing) { + DBG("thrashing\n"); + } + bufmgr_fake->thrashing = 20; + + if (alloc_block(bo)) + return GL_TRUE; + } + + while (evict_mru(bufmgr_fake)) + if (alloc_block(bo)) + return GL_TRUE; + + DBG("%s 0x%x bytes failed\n", __FUNCTION__, bo->size); + + return GL_FALSE; +} + +/*********************************************************************** + * Public functions + */ + +/** + * Wait for hardware idle by emitting a fence and waiting for it. + */ +static void +dri_bufmgr_fake_wait_idle(dri_bufmgr_fake *bufmgr_fake) +{ + unsigned int cookie; + + cookie = bufmgr_fake->fence_emit(bufmgr_fake->driver_priv); + _fence_wait_internal(bufmgr_fake, cookie); +} + +/** + * Wait for rendering to a buffer to complete. + * + * It is assumed that the bathcbuffer which performed the rendering included + * the necessary flushing. + */ +static void +dri_fake_bo_wait_rendering(dri_bo *bo) +{ + dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr; + dri_bo_fake *bo_fake = (dri_bo_fake *)bo; + + if (bo_fake->block == NULL || !bo_fake->block->fenced) + return; + + _fence_wait_internal(bufmgr_fake, bo_fake->block->fence); +} + +/* Specifically ignore texture memory sharing. + * -- just evict everything + * -- and wait for idle + */ +void +dri_bufmgr_fake_contended_lock_take(dri_bufmgr *bufmgr) +{ + dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bufmgr; + struct block *block, *tmp; + + bufmgr_fake->need_fence = 1; + bufmgr_fake->fail = 0; + + /* Wait for hardware idle. We don't know where acceleration has been + * happening, so we'll need to wait anyway before letting anything get + * put on the card again. + */ + dri_bufmgr_fake_wait_idle(bufmgr_fake); + + /* Check that we hadn't released the lock without having fenced the last + * set of buffers. + */ + assert(is_empty_list(&bufmgr_fake->fenced)); + assert(is_empty_list(&bufmgr_fake->on_hardware)); + + foreach_s(block, tmp, &bufmgr_fake->lru) { + assert(_fence_test(bufmgr_fake, block->fence)); + set_dirty(block->bo); + } +} + +static dri_bo * +dri_fake_bo_alloc(dri_bufmgr *bufmgr, const char *name, + unsigned long size, unsigned int alignment, + uint64_t location_mask) +{ + dri_bufmgr_fake *bufmgr_fake; + dri_bo_fake *bo_fake; + + bufmgr_fake = (dri_bufmgr_fake *)bufmgr; + + assert(size != 0); + + bo_fake = calloc(1, sizeof(*bo_fake)); + if (!bo_fake) + return NULL; + + bo_fake->bo.size = size; + bo_fake->bo.offset = -1; + bo_fake->bo.virtual = NULL; + bo_fake->bo.bufmgr = bufmgr; + bo_fake->refcount = 1; + + /* Alignment must be a power of two */ + assert((alignment & (alignment - 1)) == 0); + if (alignment == 0) + alignment = 1; + bo_fake->alignment = alignment; + bo_fake->id = ++bufmgr_fake->buf_nr; + bo_fake->name = name; + bo_fake->flags = 0; + bo_fake->is_static = GL_FALSE; + + DBG("drm_bo_alloc: (buf %d: %s, %d kb)\n", bo_fake->id, bo_fake->name, + bo_fake->bo.size / 1024); + + return &bo_fake->bo; +} + +static dri_bo * +dri_fake_bo_alloc_static(dri_bufmgr *bufmgr, const char *name, + unsigned long offset, unsigned long size, + void *virtual, uint64_t location_mask) +{ + dri_bufmgr_fake *bufmgr_fake; + dri_bo_fake *bo_fake; + + bufmgr_fake = (dri_bufmgr_fake *)bufmgr; + + assert(size != 0); + + bo_fake = calloc(1, sizeof(*bo_fake)); + if (!bo_fake) + return NULL; + + bo_fake->bo.size = size; + bo_fake->bo.offset = offset; + bo_fake->bo.virtual = virtual; + bo_fake->bo.bufmgr = bufmgr; + bo_fake->refcount = 1; + bo_fake->id = ++bufmgr_fake->buf_nr; + bo_fake->name = name; + bo_fake->flags = BM_PINNED | DRM_BO_FLAG_NO_MOVE; + bo_fake->is_static = GL_TRUE; + + DBG("drm_bo_alloc_static: (buf %d: %s, %d kb)\n", bo_fake->id, bo_fake->name, + bo_fake->bo.size / 1024); + + return &bo_fake->bo; +} + +static void +dri_fake_bo_reference(dri_bo *bo) +{ + dri_bo_fake *bo_fake = (dri_bo_fake *)bo; + + bo_fake->refcount++; +} + +static void +dri_fake_bo_unreference(dri_bo *bo) +{ + dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr; + dri_bo_fake *bo_fake = (dri_bo_fake *)bo; + int i; + + if (!bo) + return; + + if (--bo_fake->refcount == 0) { + assert(bo_fake->map_count == 0); + /* No remaining references, so free it */ + if (bo_fake->block) + free_block(bufmgr_fake, bo_fake->block); + free_backing_store(bo); + + for (i = 0; i < bo_fake->nr_relocs; i++) + dri_bo_unreference(bo_fake->relocs[i].target_buf); + + DBG("drm_bo_unreference: free buf %d %s\n", bo_fake->id, bo_fake->name); + + free(bo_fake->relocs); + free(bo); + + return; + } +} + +/** + * Set the buffer as not requiring backing store, and instead get the callback + * invoked whenever it would be set dirty. + */ +void dri_bo_fake_disable_backing_store(dri_bo *bo, + void (*invalidate_cb)(dri_bo *bo, + void *ptr), + void *ptr) +{ + dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr; + dri_bo_fake *bo_fake = (dri_bo_fake *)bo; + + if (bo_fake->backing_store) + free_backing_store(bo); + + bo_fake->flags |= BM_NO_BACKING_STORE; + + DBG("disable_backing_store set buf %d dirty\n", bo_fake->id); + bo_fake->dirty = 1; + bo_fake->invalidate_cb = invalidate_cb; + bo_fake->invalidate_ptr = ptr; + + /* Note that it is invalid right from the start. Also note + * invalidate_cb is called with the bufmgr locked, so cannot + * itself make bufmgr calls. + */ + if (invalidate_cb != NULL) + invalidate_cb(bo, ptr); +} + +/** + * Map a buffer into bo->virtual, allocating either card memory space (If + * BM_NO_BACKING_STORE or BM_PINNED) or backing store, as necessary. + */ +static int +dri_fake_bo_map(dri_bo *bo, GLboolean write_enable) +{ + dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr; + dri_bo_fake *bo_fake = (dri_bo_fake *)bo; + + /* Static buffers are always mapped. */ + if (bo_fake->is_static) + return 0; + + /* Allow recursive mapping. Mesa may recursively map buffers with + * nested display loops, and it is used internally in bufmgr_fake + * for relocation. + */ + if (bo_fake->map_count++ != 0) + return 0; + + { + DBG("drm_bo_map: (buf %d: %s, %d kb)\n", bo_fake->id, bo_fake->name, + bo_fake->bo.size / 1024); + + if (bo->virtual != NULL) { + _mesa_printf("%s: already mapped\n", __FUNCTION__); + abort(); + } + else if (bo_fake->flags & (BM_NO_BACKING_STORE|BM_PINNED)) { + + if (!bo_fake->block && !evict_and_alloc_block(bo)) { + DBG("%s: alloc failed\n", __FUNCTION__); + bufmgr_fake->fail = 1; + return 1; + } + else { + assert(bo_fake->block); + bo_fake->dirty = 0; + + if (!(bo_fake->flags & BM_NO_FENCE_SUBDATA) && + bo_fake->block->fenced) { + dri_fake_bo_wait_rendering(bo); + } + + bo->virtual = bo_fake->block->virtual; + } + } + else { + if (write_enable) + set_dirty(bo); + + if (bo_fake->backing_store == 0) + alloc_backing_store(bo); + + bo->virtual = bo_fake->backing_store; + } + } + + return 0; +} + +static int +dri_fake_bo_unmap(dri_bo *bo) +{ + dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr; + dri_bo_fake *bo_fake = (dri_bo_fake *)bo; + + /* Static buffers are always mapped. */ + if (bo_fake->is_static) + return 0; + + assert(bo_fake->map_count != 0); + if (--bo_fake->map_count != 0) + return 0; + + DBG("drm_bo_unmap: (buf %d: %s, %d kb)\n", bo_fake->id, bo_fake->name, + bo_fake->bo.size / 1024); + + bo->virtual = NULL; + + return 0; +} + +static void +dri_fake_kick_all(dri_bufmgr_fake *bufmgr_fake) +{ + struct block *block, *tmp; + + bufmgr_fake->performed_rendering = GL_FALSE; + /* okay for ever BO that is on the HW kick it off. + seriously not afraid of the POLICE right now */ + foreach_s(block, tmp, &bufmgr_fake->on_hardware) { + dri_bo_fake *bo_fake = (dri_bo_fake *)block->bo; + + block->on_hardware = 0; + free_block(bufmgr_fake, block); + bo_fake->block = NULL; + bo_fake->validated = GL_FALSE; + if (!(bo_fake->flags & BM_NO_BACKING_STORE)) + bo_fake->dirty = 1; + } +} + +static int +dri_fake_bo_validate(dri_bo *bo) +{ + dri_bufmgr_fake *bufmgr_fake; + dri_bo_fake *bo_fake = (dri_bo_fake *)bo; + + /* XXX: Sanity-check whether we've already validated this one under + * different flags. See drmAddValidateItem(). + */ + bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr; + + DBG("drm_bo_validate: (buf %d: %s, %d kb)\n", bo_fake->id, bo_fake->name, + bo_fake->bo.size / 1024); + + /* Sanity check: Buffers should be unmapped before being validated. + * This is not so much of a problem for bufmgr_fake, but TTM refuses, + * and the problem is harder to debug there. + */ + assert(bo_fake->map_count == 0); + + if (bo_fake->is_static) { + /* Add it to the needs-fence list */ + bufmgr_fake->need_fence = 1; + return 0; + } + + /* reset size accounted */ + bo_fake->size_accounted = 0; + + /* Allocate the card memory */ + if (!bo_fake->block && !evict_and_alloc_block(bo)) { + bufmgr_fake->fail = 1; + DBG("Failed to validate buf %d:%s\n", bo_fake->id, bo_fake->name); + return -1; + } + + assert(bo_fake->block); + assert(bo_fake->block->bo == &bo_fake->bo); + + bo->offset = bo_fake->block->mem->ofs; + + /* Upload the buffer contents if necessary */ + if (bo_fake->dirty) { + DBG("Upload dirty buf %d:%s, sz %d offset 0x%x\n", bo_fake->id, + bo_fake->name, bo->size, bo_fake->block->mem->ofs); + + assert(!(bo_fake->flags & + (BM_NO_BACKING_STORE|BM_PINNED))); + + /* Actually, should be able to just wait for a fence on the memory, + * which we would be tracking when we free it. Waiting for idle is + * a sufficiently large hammer for now. + */ + dri_bufmgr_fake_wait_idle(bufmgr_fake); + + /* we may never have mapped this BO so it might not have any backing + * store if this happens it should be rare, but 0 the card memory + * in any case */ + if (bo_fake->backing_store) + memcpy(bo_fake->block->virtual, bo_fake->backing_store, bo->size); + else + memset(bo_fake->block->virtual, 0, bo->size); + + bo_fake->dirty = 0; + } + + bo_fake->block->fenced = 0; + bo_fake->block->on_hardware = 1; + move_to_tail(&bufmgr_fake->on_hardware, bo_fake->block); + + bo_fake->validated = GL_TRUE; + bufmgr_fake->need_fence = 1; + + return 0; +} + +static void +dri_fake_fence_validated(dri_bufmgr *bufmgr) +{ + dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bufmgr; + unsigned int cookie; + + cookie = _fence_emit_internal(bufmgr_fake); + fence_blocks(bufmgr_fake, cookie); + + DBG("drm_fence_validated: 0x%08x cookie\n", cookie); +} + +static void +dri_fake_destroy(dri_bufmgr *bufmgr) +{ + dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bufmgr; + + mmDestroy(bufmgr_fake->heap); + free(bufmgr); +} + +static int +dri_fake_emit_reloc(dri_bo *reloc_buf, + uint32_t read_domains, uint32_t write_domain, + uint32_t delta, uint32_t offset, dri_bo *target_buf) +{ + dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)reloc_buf->bufmgr; + struct fake_buffer_reloc *r; + dri_bo_fake *reloc_fake = (dri_bo_fake *)reloc_buf; + dri_bo_fake *target_fake = (dri_bo_fake *)target_buf; + int i; + + assert(reloc_buf); + assert(target_buf); + + assert(target_fake->is_static || target_fake->size_accounted); + + if (reloc_fake->relocs == NULL) { + reloc_fake->relocs = malloc(sizeof(struct fake_buffer_reloc) * + MAX_RELOCS); + } + + r = &reloc_fake->relocs[reloc_fake->nr_relocs++]; + + assert(reloc_fake->nr_relocs <= MAX_RELOCS); + + dri_bo_reference(target_buf); + + r->target_buf = target_buf; + r->offset = offset; + r->last_target_offset = target_buf->offset; + r->delta = delta; + r->read_domains = read_domains; + r->write_domain = write_domain; + + if (bufmgr_fake->debug) { + /* Check that a conflicting relocation hasn't already been emitted. */ + for (i = 0; i < reloc_fake->nr_relocs - 1; i++) { + struct fake_buffer_reloc *r2 = &reloc_fake->relocs[i]; + + assert(r->offset != r2->offset); + } + } + + return 0; +} + +/** + * Incorporates the validation flags associated with each relocation into + * the combined validation flags for the buffer on this batchbuffer submission. + */ +static void +dri_fake_calculate_domains(dri_bo *bo) +{ + dri_bo_fake *bo_fake = (dri_bo_fake *)bo; + int i; + + for (i = 0; i < bo_fake->nr_relocs; i++) { + struct fake_buffer_reloc *r = &bo_fake->relocs[i]; + dri_bo_fake *target_fake = (dri_bo_fake *)r->target_buf; + + /* Do the same for the tree of buffers we depend on */ + dri_fake_calculate_domains(r->target_buf); + + target_fake->read_domains |= r->read_domains; + if (target_fake->write_domain != 0) + target_fake->write_domain = r->write_domain; + } +} + + +static int +dri_fake_reloc_and_validate_buffer(dri_bo *bo) +{ + dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr; + dri_bo_fake *bo_fake = (dri_bo_fake *)bo; + int i, ret; + + assert(bo_fake->map_count == 0); + + for (i = 0; i < bo_fake->nr_relocs; i++) { + struct fake_buffer_reloc *r = &bo_fake->relocs[i]; + dri_bo_fake *target_fake = (dri_bo_fake *)r->target_buf; + uint32_t reloc_data; + + /* Validate the target buffer if that hasn't been done. */ + if (!target_fake->validated) { + ret = dri_fake_reloc_and_validate_buffer(r->target_buf); + if (ret != 0) { + if (bo->virtual != NULL) + dri_bo_unmap(bo); + return ret; + } + } + + /* Calculate the value of the relocation entry. */ + if (r->target_buf->offset != r->last_target_offset) { + reloc_data = r->target_buf->offset + r->delta; + + if (bo->virtual == NULL) + dri_bo_map(bo, GL_TRUE); + + *(uint32_t *)(bo->virtual + r->offset) = reloc_data; + + r->last_target_offset = r->target_buf->offset; + } + } + + if (bo->virtual != NULL) + dri_bo_unmap(bo); + + if (bo_fake->write_domain != 0) { + if (!(bo_fake->flags & (BM_NO_BACKING_STORE|BM_PINNED))) { + if (bo_fake->backing_store == 0) + alloc_backing_store(bo); + + bo_fake->card_dirty = 1; + } + bufmgr_fake->performed_rendering = GL_TRUE; + } + + return dri_fake_bo_validate(bo); +} + +static void * +dri_fake_process_relocs(dri_bo *batch_buf) +{ + dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)batch_buf->bufmgr; + dri_bo_fake *batch_fake = (dri_bo_fake *)batch_buf; + int ret; + int retry_count = 0; + + bufmgr_fake->performed_rendering = GL_FALSE; + + dri_fake_calculate_domains(batch_buf); + + batch_fake->read_domains = DRM_GEM_DOMAIN_I915_COMMAND; + + /* we've ran out of RAM so blow the whole lot away and retry */ + restart: + ret = dri_fake_reloc_and_validate_buffer(batch_buf); + if (bufmgr_fake->fail == 1) { + if (retry_count == 0) { + retry_count++; + dri_fake_kick_all(bufmgr_fake); + bufmgr_fake->fail = 0; + goto restart; + } else /* dump out the memory here */ + mmDumpMemInfo(bufmgr_fake->heap); + } + + assert(ret == 0); + + bufmgr_fake->current_total_size = 0; + return NULL; +} + +static void +dri_bo_fake_post_submit(dri_bo *bo) +{ + dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr; + dri_bo_fake *bo_fake = (dri_bo_fake *)bo; + int i; + + for (i = 0; i < bo_fake->nr_relocs; i++) { + struct fake_buffer_reloc *r = &bo_fake->relocs[i]; + dri_bo_fake *target_fake = (dri_bo_fake *)r->target_buf; + + if (target_fake->validated) + dri_bo_fake_post_submit(r->target_buf); + + DBG("%s@0x%08x + 0x%08x -> %s@0x%08x + 0x%08x\n", + bo_fake->name, (uint32_t)bo->offset, r->offset, + target_fake->name, (uint32_t)r->target_buf->offset, r->delta); + } + + assert(bo_fake->map_count == 0); + bo_fake->validated = GL_FALSE; + bo_fake->read_domains = 0; + bo_fake->write_domain = 0; +} + + +static void +dri_fake_post_submit(dri_bo *batch_buf) +{ + dri_fake_fence_validated(batch_buf->bufmgr); + + dri_bo_fake_post_submit(batch_buf); +} + +static int +dri_fake_check_aperture_space(dri_bo *bo) +{ + dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr; + dri_bo_fake *bo_fake = (dri_bo_fake *)bo; + GLuint sz; + + sz = (bo->size + bo_fake->alignment - 1) & ~(bo_fake->alignment - 1); + + if (bo_fake->size_accounted || bo_fake->is_static) + return 0; + + if (bufmgr_fake->current_total_size + sz > bufmgr_fake->size) { + DBG("check_space: %s bo %d %d overflowed bufmgr size %d\n", bo_fake->name, bo_fake->id, sz, bufmgr_fake->size); + return -1; + } + + bufmgr_fake->current_total_size += sz; + bo_fake->size_accounted = 1; + DBG("drm_check_space: buf %d, %s %d %d\n", bo_fake->id, bo_fake->name, bo->size, bufmgr_fake->current_total_size); + return 0; +} + +dri_bufmgr * +dri_bufmgr_fake_init(unsigned long low_offset, void *low_virtual, + unsigned long size, + unsigned int (*fence_emit)(void *private), + int (*fence_wait)(void *private, unsigned int cookie), + void *driver_priv) +{ + dri_bufmgr_fake *bufmgr_fake; + + bufmgr_fake = calloc(1, sizeof(*bufmgr_fake)); + + /* Initialize allocator */ + make_empty_list(&bufmgr_fake->fenced); + make_empty_list(&bufmgr_fake->on_hardware); + make_empty_list(&bufmgr_fake->lru); + + bufmgr_fake->low_offset = low_offset; + bufmgr_fake->virtual = low_virtual; + bufmgr_fake->size = size; + bufmgr_fake->heap = mmInit(low_offset, size); + + /* Hook in methods */ + bufmgr_fake->bufmgr.bo_alloc = dri_fake_bo_alloc; + bufmgr_fake->bufmgr.bo_alloc_static = dri_fake_bo_alloc_static; + bufmgr_fake->bufmgr.bo_reference = dri_fake_bo_reference; + bufmgr_fake->bufmgr.bo_unreference = dri_fake_bo_unreference; + bufmgr_fake->bufmgr.bo_map = dri_fake_bo_map; + bufmgr_fake->bufmgr.bo_unmap = dri_fake_bo_unmap; + bufmgr_fake->bufmgr.bo_wait_rendering = dri_fake_bo_wait_rendering; + bufmgr_fake->bufmgr.destroy = dri_fake_destroy; + bufmgr_fake->bufmgr.emit_reloc = dri_fake_emit_reloc; + bufmgr_fake->bufmgr.process_relocs = dri_fake_process_relocs; + bufmgr_fake->bufmgr.post_submit = dri_fake_post_submit; + bufmgr_fake->bufmgr.check_aperture_space = dri_fake_check_aperture_space; + bufmgr_fake->bufmgr.debug = GL_FALSE; + + bufmgr_fake->fence_emit = fence_emit; + bufmgr_fake->fence_wait = fence_wait; + bufmgr_fake->driver_priv = driver_priv; + + return &bufmgr_fake->bufmgr; +} + -- cgit v1.2.3 From c4857429c716f35e1fa054d1990cae28055d96d7 Mon Sep 17 00:00:00 2001 From: Eric Anholt Date: Tue, 3 Jun 2008 10:20:49 -0700 Subject: Fix and hook up bufmgr code to the build. --- libdrm/intel/intel_bufmgr_fake.c | 234 ++++++++++++++++++++------------------- 1 file changed, 121 insertions(+), 113 deletions(-) (limited to 'libdrm/intel/intel_bufmgr_fake.c') diff --git a/libdrm/intel/intel_bufmgr_fake.c b/libdrm/intel/intel_bufmgr_fake.c index 2aed3d85..1bddbeab 100644 --- a/libdrm/intel/intel_bufmgr_fake.c +++ b/libdrm/intel/intel_bufmgr_fake.c @@ -33,19 +33,19 @@ * programming interface, but is more expressive and avoids many of * the bugs in the old texture manager. */ -#include "mtypes.h" + +#include +#include +#include #include "dri_bufmgr.h" -#include "intel_bufmgr_fake.h" +#include "intel_bufmgr.h" #include "drm.h" #include "i915_drm.h" - -#include "simple_list.h" #include "mm.h" -#include "imports.h" #define DBG(...) do { \ if (bufmgr_fake->bufmgr.debug) \ - _mesa_printf(__VA_ARGS__); \ + drmMsg(__VA_ARGS__); \ } while (0) /* Internal flags: @@ -67,11 +67,11 @@ struct fake_buffer_reloc /** Buffer object that the relocation points at. */ dri_bo *target_buf; /** Offset of the relocation entry within reloc_buf. */ - GLuint offset; + uint32_t offset; /** Cached value of the offset when we last performed this relocation. */ - GLuint last_target_offset; + uint32_t last_target_offset; /** Value added to target_buf's offset to get the relocation entry. */ - GLuint delta; + uint32_t delta; /** Cache domains the target buffer is read into. */ uint32_t read_domains; /** Cache domain the target buffer will have dirty cachelines in. */ @@ -102,25 +102,36 @@ struct block { typedef struct _bufmgr_fake { dri_bufmgr bufmgr; + struct intel_bufmgr intel_bufmgr; unsigned long low_offset; unsigned long size; void *virtual; struct mem_block *heap; - struct block lru; /* only allocated, non-fence-pending blocks here */ unsigned buf_nr; /* for generating ids */ - struct block on_hardware; /* after bmValidateBuffers */ - struct block fenced; /* after bmFenceBuffers (mi_flush, emit irq, write dword) */ + /** + * List of blocks which are currently in the GART but haven't been + * fenced yet. + */ + struct block on_hardware; + /** + * List of blocks which are in the GART and have an active fence on them. + */ + struct block fenced; + /** + * List of blocks which have an expired fence and are ready to be evicted. + */ + struct block lru; /* then to bufmgr->lru or free() */ unsigned int last_fence; unsigned fail:1; unsigned need_fence:1; - GLboolean thrashing; + int thrashing; /** * Driver callback to emit a fence, returning the cookie. @@ -134,9 +145,9 @@ typedef struct _bufmgr_fake { /** Driver-supplied argument to driver callbacks */ void *driver_priv; - GLboolean debug; + int debug; - GLboolean performed_rendering; + int performed_rendering; /* keep track of the current total size of objects we have relocs for */ unsigned long current_total_size; @@ -163,12 +174,12 @@ typedef struct _dri_bo_fake { uint32_t write_domain; unsigned int alignment; - GLboolean is_static, validated; + int is_static, validated; unsigned int map_count; /** relocation list */ struct fake_buffer_reloc *relocs; - GLuint nr_relocs; + int nr_relocs; struct block *block; void *backing_store; @@ -183,18 +194,18 @@ static int dri_fake_check_aperture_space(dri_bo *bo); #define MAXFENCE 0x7fffffff -static GLboolean FENCE_LTE( unsigned a, unsigned b ) +static int FENCE_LTE( unsigned a, unsigned b ) { if (a == b) - return GL_TRUE; + return 1; if (a < b && b - a < (1<<24)) - return GL_TRUE; + return 1; if (a > b && MAXFENCE - a + b < (1<<24)) - return GL_TRUE; + return 1; - return GL_FALSE; + return 0; } static unsigned int @@ -211,14 +222,13 @@ _fence_wait_internal(dri_bufmgr_fake *bufmgr_fake, unsigned int cookie) ret = bufmgr_fake->fence_wait(bufmgr_fake->driver_priv, cookie); if (ret != 0) { - _mesa_printf("%s:%d: Error %d waiting for fence.\n", - __FILE__, __LINE__); + drmMsg("%s:%d: Error %d waiting for fence.\n", __FILE__, __LINE__); abort(); } clear_fenced(bufmgr_fake, cookie); } -static GLboolean +static int _fence_test(dri_bufmgr_fake *bufmgr_fake, unsigned fence) { /* Slight problem with wrap-around: @@ -229,39 +239,39 @@ _fence_test(dri_bufmgr_fake *bufmgr_fake, unsigned fence) /** * Allocate a memory manager block for the buffer. */ -static GLboolean +static int alloc_block(dri_bo *bo) { dri_bo_fake *bo_fake = (dri_bo_fake *)bo; dri_bufmgr_fake *bufmgr_fake= (dri_bufmgr_fake *)bo->bufmgr; struct block *block = (struct block *)calloc(sizeof *block, 1); - unsigned int align_log2 = _mesa_ffs(bo_fake->alignment) - 1; - GLuint sz; + unsigned int align_log2 = ffs(bo_fake->alignment) - 1; + unsigned int sz; if (!block) - return GL_FALSE; + return 1; sz = (bo->size + bo_fake->alignment - 1) & ~(bo_fake->alignment - 1); - block->mem = mmAllocMem(bufmgr_fake->heap, sz, align_log2, 0); + block->mem = drmmmAllocMem(bufmgr_fake->heap, sz, align_log2, 0); if (!block->mem) { free(block); - return GL_FALSE; + return 0; } - make_empty_list(block); + DRMINITLISTHEAD(block); /* Insert at head or at tail??? */ - insert_at_tail(&bufmgr_fake->lru, block); + DRMLISTADDTAIL(block, &bufmgr_fake->lru); - block->virtual = bufmgr_fake->virtual + + block->virtual = (uint8_t *)bufmgr_fake->virtual + block->mem->ofs - bufmgr_fake->low_offset; block->bo = bo; bo_fake->block = block; - return GL_TRUE; + return 1; } /* Release the card storage associated with buf: @@ -289,9 +299,9 @@ static void free_block(dri_bufmgr_fake *bufmgr_fake, struct block *block) } else { DBG(" - free immediately\n"); - remove_from_list(block); + DRMLISTDEL(block); - mmFreeMem(block->mem); + drmmmFreeMem(block->mem); free(block); } } @@ -304,7 +314,7 @@ alloc_backing_store(dri_bo *bo) assert(!bo_fake->backing_store); assert(!(bo_fake->flags & (BM_PINNED|BM_NO_BACKING_STORE))); - bo_fake->backing_store = ALIGN_MALLOC(bo->size, 64); + bo_fake->backing_store = malloc(bo->size); DBG("alloc_backing - buf %d %p %d\n", bo_fake->id, bo_fake->backing_store, bo->size); assert(bo_fake->backing_store); @@ -317,7 +327,7 @@ free_backing_store(dri_bo *bo) if (bo_fake->backing_store) { assert(!(bo_fake->flags & (BM_PINNED|BM_NO_BACKING_STORE))); - ALIGN_FREE(bo_fake->backing_store); + free(bo_fake->backing_store); bo_fake->backing_store = NULL; } } @@ -337,14 +347,14 @@ set_dirty(dri_bo *bo) bo_fake->dirty = 1; } -static GLboolean -evict_lru(dri_bufmgr_fake *bufmgr_fake, GLuint max_fence) +static int +evict_lru(dri_bufmgr_fake *bufmgr_fake, unsigned int max_fence) { struct block *block, *tmp; DBG("%s\n", __FUNCTION__); - foreach_s(block, tmp, &bufmgr_fake->lru) { + DRMLISTFOREACHSAFE(block, tmp, &bufmgr_fake->lru) { dri_bo_fake *bo_fake = (dri_bo_fake *)block->bo; if (bo_fake != NULL && (bo_fake->flags & BM_NO_FENCE_SUBDATA)) @@ -357,23 +367,20 @@ evict_lru(dri_bufmgr_fake *bufmgr_fake, GLuint max_fence) bo_fake->block = NULL; free_block(bufmgr_fake, block); - return GL_TRUE; + return 1; } - return GL_FALSE; + return 0; } -#define foreach_s_rev(ptr, t, list) \ - for(ptr=(list)->prev,t=(ptr)->prev; list != ptr; ptr=t, t=(t)->prev) - -static GLboolean +static int evict_mru(dri_bufmgr_fake *bufmgr_fake) { struct block *block, *tmp; DBG("%s\n", __FUNCTION__); - foreach_s_rev(block, tmp, &bufmgr_fake->lru) { + DRMLISTFOREACHSAFEREVERSE(block, tmp, &bufmgr_fake->lru) { dri_bo_fake *bo_fake = (dri_bo_fake *)block->bo; if (bo_fake && (bo_fake->flags & BM_NO_FENCE_SUBDATA)) @@ -383,10 +390,10 @@ evict_mru(dri_bufmgr_fake *bufmgr_fake) bo_fake->block = NULL; free_block(bufmgr_fake, block); - return GL_TRUE; + return 1; } - return GL_FALSE; + return 0; } /** @@ -398,7 +405,7 @@ static int clear_fenced(dri_bufmgr_fake *bufmgr_fake, struct block *block, *tmp; int ret = 0; - foreach_s(block, tmp, &bufmgr_fake->fenced) { + DRMLISTFOREACHSAFE(block, tmp, &bufmgr_fake->fenced) { assert(block->fenced); if (_fence_test(bufmgr_fake, block->fence)) { @@ -408,14 +415,15 @@ static int clear_fenced(dri_bufmgr_fake *bufmgr_fake, if (!block->bo) { DBG("delayed free: offset %x sz %x\n", block->mem->ofs, block->mem->size); - remove_from_list(block); - mmFreeMem(block->mem); + DRMLISTDEL(block); + drmmmFreeMem(block->mem); free(block); } else { DBG("return to lru: offset %x sz %x\n", block->mem->ofs, block->mem->size); - move_to_tail(&bufmgr_fake->lru, block); + DRMLISTDEL(block); + DRMLISTADDTAIL(block, &bufmgr_fake->lru); } ret = 1; @@ -438,7 +446,7 @@ static void fence_blocks(dri_bufmgr_fake *bufmgr_fake, unsigned fence) { struct block *block, *tmp; - foreach_s (block, tmp, &bufmgr_fake->on_hardware) { + DRMLISTFOREACHSAFE(block, tmp, &bufmgr_fake->on_hardware) { DBG("Fence block %p (sz 0x%x ofs %x buf %p) with fence %d\n", block, block->mem->size, block->mem->ofs, block->bo, fence); block->fence = fence; @@ -448,13 +456,14 @@ static void fence_blocks(dri_bufmgr_fake *bufmgr_fake, unsigned fence) /* Move to tail of pending list here */ - move_to_tail(&bufmgr_fake->fenced, block); + DRMLISTDEL(block); + DRMLISTADDTAIL(block, &bufmgr_fake->fenced); } - assert(is_empty_list(&bufmgr_fake->on_hardware)); + assert(DRMLISTEMPTY(&bufmgr_fake->on_hardware)); } -static GLboolean evict_and_alloc_block(dri_bo *bo) +static int evict_and_alloc_block(dri_bo *bo) { dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr; dri_bo_fake *bo_fake = (dri_bo_fake *)bo; @@ -464,7 +473,7 @@ static GLboolean evict_and_alloc_block(dri_bo *bo) /* Search for already free memory: */ if (alloc_block(bo)) - return GL_TRUE; + return 1; /* If we're not thrashing, allow lru eviction to dig deeper into * recently used textures. We'll probably be thrashing soon: @@ -472,7 +481,7 @@ static GLboolean evict_and_alloc_block(dri_bo *bo) if (!bufmgr_fake->thrashing) { while (evict_lru(bufmgr_fake, 0)) if (alloc_block(bo)) - return GL_TRUE; + return 1; } /* Keep thrashing counter alive? @@ -484,17 +493,17 @@ static GLboolean evict_and_alloc_block(dri_bo *bo) * freed memory that has been submitted to hardware and fenced to * become available: */ - while (!is_empty_list(&bufmgr_fake->fenced)) { - GLuint fence = bufmgr_fake->fenced.next->fence; + while (!DRMLISTEMPTY(&bufmgr_fake->fenced)) { + uint32_t fence = bufmgr_fake->fenced.next->fence; _fence_wait_internal(bufmgr_fake, fence); if (alloc_block(bo)) - return GL_TRUE; + return 1; } - if (!is_empty_list(&bufmgr_fake->on_hardware)) { - while (!is_empty_list(&bufmgr_fake->fenced)) { - GLuint fence = bufmgr_fake->fenced.next->fence; + if (!DRMLISTEMPTY(&bufmgr_fake->on_hardware)) { + while (!DRMLISTEMPTY(&bufmgr_fake->fenced)) { + uint32_t fence = bufmgr_fake->fenced.next->fence; _fence_wait_internal(bufmgr_fake, fence); } @@ -504,16 +513,16 @@ static GLboolean evict_and_alloc_block(dri_bo *bo) bufmgr_fake->thrashing = 20; if (alloc_block(bo)) - return GL_TRUE; + return 1; } while (evict_mru(bufmgr_fake)) if (alloc_block(bo)) - return GL_TRUE; + return 1; DBG("%s 0x%x bytes failed\n", __FUNCTION__, bo->size); - return GL_FALSE; + return 0; } /*********************************************************************** @@ -555,7 +564,7 @@ dri_fake_bo_wait_rendering(dri_bo *bo) * -- and wait for idle */ void -dri_bufmgr_fake_contended_lock_take(dri_bufmgr *bufmgr) +intel_bufmgr_fake_contended_lock_take(dri_bufmgr *bufmgr) { dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bufmgr; struct block *block, *tmp; @@ -572,10 +581,10 @@ dri_bufmgr_fake_contended_lock_take(dri_bufmgr *bufmgr) /* Check that we hadn't released the lock without having fenced the last * set of buffers. */ - assert(is_empty_list(&bufmgr_fake->fenced)); - assert(is_empty_list(&bufmgr_fake->on_hardware)); + assert(DRMLISTEMPTY(&bufmgr_fake->fenced)); + assert(DRMLISTEMPTY(&bufmgr_fake->on_hardware)); - foreach_s(block, tmp, &bufmgr_fake->lru) { + DRMLISTFOREACHSAFE(block, tmp, &bufmgr_fake->lru) { assert(_fence_test(bufmgr_fake, block->fence)); set_dirty(block->bo); } @@ -583,8 +592,7 @@ dri_bufmgr_fake_contended_lock_take(dri_bufmgr *bufmgr) static dri_bo * dri_fake_bo_alloc(dri_bufmgr *bufmgr, const char *name, - unsigned long size, unsigned int alignment, - uint64_t location_mask) + unsigned long size, unsigned int alignment) { dri_bufmgr_fake *bufmgr_fake; dri_bo_fake *bo_fake; @@ -611,7 +619,7 @@ dri_fake_bo_alloc(dri_bufmgr *bufmgr, const char *name, bo_fake->id = ++bufmgr_fake->buf_nr; bo_fake->name = name; bo_fake->flags = 0; - bo_fake->is_static = GL_FALSE; + bo_fake->is_static = 0; DBG("drm_bo_alloc: (buf %d: %s, %d kb)\n", bo_fake->id, bo_fake->name, bo_fake->bo.size / 1024); @@ -619,10 +627,10 @@ dri_fake_bo_alloc(dri_bufmgr *bufmgr, const char *name, return &bo_fake->bo; } -static dri_bo * -dri_fake_bo_alloc_static(dri_bufmgr *bufmgr, const char *name, - unsigned long offset, unsigned long size, - void *virtual, uint64_t location_mask) +dri_bo * +intel_bo_fake_alloc_static(dri_bufmgr *bufmgr, const char *name, + unsigned long offset, unsigned long size, + void *virtual) { dri_bufmgr_fake *bufmgr_fake; dri_bo_fake *bo_fake; @@ -643,7 +651,7 @@ dri_fake_bo_alloc_static(dri_bufmgr *bufmgr, const char *name, bo_fake->id = ++bufmgr_fake->buf_nr; bo_fake->name = name; bo_fake->flags = BM_PINNED | DRM_BO_FLAG_NO_MOVE; - bo_fake->is_static = GL_TRUE; + bo_fake->is_static = 1; DBG("drm_bo_alloc_static: (buf %d: %s, %d kb)\n", bo_fake->id, bo_fake->name, bo_fake->bo.size / 1024); @@ -692,10 +700,10 @@ dri_fake_bo_unreference(dri_bo *bo) * Set the buffer as not requiring backing store, and instead get the callback * invoked whenever it would be set dirty. */ -void dri_bo_fake_disable_backing_store(dri_bo *bo, - void (*invalidate_cb)(dri_bo *bo, - void *ptr), - void *ptr) +void intel_bo_fake_disable_backing_store(dri_bo *bo, + void (*invalidate_cb)(dri_bo *bo, + void *ptr), + void *ptr) { dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr; dri_bo_fake *bo_fake = (dri_bo_fake *)bo; @@ -723,7 +731,7 @@ void dri_bo_fake_disable_backing_store(dri_bo *bo, * BM_NO_BACKING_STORE or BM_PINNED) or backing store, as necessary. */ static int -dri_fake_bo_map(dri_bo *bo, GLboolean write_enable) +dri_fake_bo_map(dri_bo *bo, int write_enable) { dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr; dri_bo_fake *bo_fake = (dri_bo_fake *)bo; @@ -744,7 +752,7 @@ dri_fake_bo_map(dri_bo *bo, GLboolean write_enable) bo_fake->bo.size / 1024); if (bo->virtual != NULL) { - _mesa_printf("%s: already mapped\n", __FUNCTION__); + drmMsg("%s: already mapped\n", __FUNCTION__); abort(); } else if (bo_fake->flags & (BM_NO_BACKING_STORE|BM_PINNED)) { @@ -807,16 +815,16 @@ dri_fake_kick_all(dri_bufmgr_fake *bufmgr_fake) { struct block *block, *tmp; - bufmgr_fake->performed_rendering = GL_FALSE; + bufmgr_fake->performed_rendering = 0; /* okay for ever BO that is on the HW kick it off. seriously not afraid of the POLICE right now */ - foreach_s(block, tmp, &bufmgr_fake->on_hardware) { + DRMLISTFOREACHSAFE(block, tmp, &bufmgr_fake->on_hardware) { dri_bo_fake *bo_fake = (dri_bo_fake *)block->bo; block->on_hardware = 0; free_block(bufmgr_fake, block); bo_fake->block = NULL; - bo_fake->validated = GL_FALSE; + bo_fake->validated = 0; if (!(bo_fake->flags & BM_NO_BACKING_STORE)) bo_fake->dirty = 1; } @@ -890,9 +898,10 @@ dri_fake_bo_validate(dri_bo *bo) bo_fake->block->fenced = 0; bo_fake->block->on_hardware = 1; - move_to_tail(&bufmgr_fake->on_hardware, bo_fake->block); + DRMLISTDEL(bo_fake->block); + DRMLISTADDTAIL(bo_fake->block, &bufmgr_fake->on_hardware); - bo_fake->validated = GL_TRUE; + bo_fake->validated = 1; bufmgr_fake->need_fence = 1; return 0; @@ -915,7 +924,7 @@ dri_fake_destroy(dri_bufmgr *bufmgr) { dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bufmgr; - mmDestroy(bufmgr_fake->heap); + drmmmDestroy(bufmgr_fake->heap); free(bufmgr); } @@ -1018,9 +1027,9 @@ dri_fake_reloc_and_validate_buffer(dri_bo *bo) reloc_data = r->target_buf->offset + r->delta; if (bo->virtual == NULL) - dri_bo_map(bo, GL_TRUE); + dri_bo_map(bo, 1); - *(uint32_t *)(bo->virtual + r->offset) = reloc_data; + *(uint32_t *)((uint8_t *)bo->virtual + r->offset) = reloc_data; r->last_target_offset = r->target_buf->offset; } @@ -1036,7 +1045,7 @@ dri_fake_reloc_and_validate_buffer(dri_bo *bo) bo_fake->card_dirty = 1; } - bufmgr_fake->performed_rendering = GL_TRUE; + bufmgr_fake->performed_rendering = 1; } return dri_fake_bo_validate(bo); @@ -1050,7 +1059,7 @@ dri_fake_process_relocs(dri_bo *batch_buf) int ret; int retry_count = 0; - bufmgr_fake->performed_rendering = GL_FALSE; + bufmgr_fake->performed_rendering = 0; dri_fake_calculate_domains(batch_buf); @@ -1066,7 +1075,7 @@ dri_fake_process_relocs(dri_bo *batch_buf) bufmgr_fake->fail = 0; goto restart; } else /* dump out the memory here */ - mmDumpMemInfo(bufmgr_fake->heap); + drmmmDumpMemInfo(bufmgr_fake->heap); } assert(ret == 0); @@ -1095,7 +1104,7 @@ dri_bo_fake_post_submit(dri_bo *bo) } assert(bo_fake->map_count == 0); - bo_fake->validated = GL_FALSE; + bo_fake->validated = 0; bo_fake->read_domains = 0; bo_fake->write_domain = 0; } @@ -1114,7 +1123,7 @@ dri_fake_check_aperture_space(dri_bo *bo) { dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr; dri_bo_fake *bo_fake = (dri_bo_fake *)bo; - GLuint sz; + unsigned int sz; sz = (bo->size + bo_fake->alignment - 1) & ~(bo_fake->alignment - 1); @@ -1133,40 +1142,39 @@ dri_fake_check_aperture_space(dri_bo *bo) } dri_bufmgr * -dri_bufmgr_fake_init(unsigned long low_offset, void *low_virtual, - unsigned long size, - unsigned int (*fence_emit)(void *private), - int (*fence_wait)(void *private, unsigned int cookie), - void *driver_priv) +intel_bufmgr_fake_init(unsigned long low_offset, void *low_virtual, + unsigned long size, + unsigned int (*fence_emit)(void *private), + int (*fence_wait)(void *private, unsigned int cookie), + void *driver_priv) { dri_bufmgr_fake *bufmgr_fake; bufmgr_fake = calloc(1, sizeof(*bufmgr_fake)); /* Initialize allocator */ - make_empty_list(&bufmgr_fake->fenced); - make_empty_list(&bufmgr_fake->on_hardware); - make_empty_list(&bufmgr_fake->lru); + DRMINITLISTHEAD(&bufmgr_fake->fenced); + DRMINITLISTHEAD(&bufmgr_fake->on_hardware); + DRMINITLISTHEAD(&bufmgr_fake->lru); bufmgr_fake->low_offset = low_offset; bufmgr_fake->virtual = low_virtual; bufmgr_fake->size = size; - bufmgr_fake->heap = mmInit(low_offset, size); + bufmgr_fake->heap = drmmmInit(low_offset, size); /* Hook in methods */ bufmgr_fake->bufmgr.bo_alloc = dri_fake_bo_alloc; - bufmgr_fake->bufmgr.bo_alloc_static = dri_fake_bo_alloc_static; bufmgr_fake->bufmgr.bo_reference = dri_fake_bo_reference; bufmgr_fake->bufmgr.bo_unreference = dri_fake_bo_unreference; bufmgr_fake->bufmgr.bo_map = dri_fake_bo_map; bufmgr_fake->bufmgr.bo_unmap = dri_fake_bo_unmap; bufmgr_fake->bufmgr.bo_wait_rendering = dri_fake_bo_wait_rendering; bufmgr_fake->bufmgr.destroy = dri_fake_destroy; - bufmgr_fake->bufmgr.emit_reloc = dri_fake_emit_reloc; bufmgr_fake->bufmgr.process_relocs = dri_fake_process_relocs; bufmgr_fake->bufmgr.post_submit = dri_fake_post_submit; bufmgr_fake->bufmgr.check_aperture_space = dri_fake_check_aperture_space; - bufmgr_fake->bufmgr.debug = GL_FALSE; + bufmgr_fake->bufmgr.debug = 0; + bufmgr_fake->intel_bufmgr.emit_reloc = dri_fake_emit_reloc; bufmgr_fake->fence_emit = fence_emit; bufmgr_fake->fence_wait = fence_wait; -- cgit v1.2.3 From d198e9b091e919a90895e551e48fb30d36a849ef Mon Sep 17 00:00:00 2001 From: Eric Anholt Date: Thu, 5 Jun 2008 08:44:46 -0700 Subject: Add a function to bufmgr_fake to evict all buffers in the GTT. This will be used by the X Server for VT switch. --- libdrm/intel/intel_bufmgr_fake.c | 35 ++++++++++++++++++++++++++++++++++- 1 file changed, 34 insertions(+), 1 deletion(-) (limited to 'libdrm/intel/intel_bufmgr_fake.c') diff --git a/libdrm/intel/intel_bufmgr_fake.c b/libdrm/intel/intel_bufmgr_fake.c index 1bddbeab..3f5a22d3 100644 --- a/libdrm/intel/intel_bufmgr_fake.c +++ b/libdrm/intel/intel_bufmgr_fake.c @@ -125,7 +125,6 @@ typedef struct _bufmgr_fake { * List of blocks which have an expired fence and are ready to be evicted. */ struct block lru; - /* then to bufmgr->lru or free() */ unsigned int last_fence; @@ -1141,6 +1140,40 @@ dri_fake_check_aperture_space(dri_bo *bo) return 0; } +/** + * Evicts all buffers, waiting for fences to pass and copying contents out + * as necessary. + * + * Used by the X Server on LeaveVT, when the card memory is no longer our + * own. + */ +void +intel_bufmgr_fake_evict_all(dri_bufmgr *bufmgr) +{ + dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bufmgr; + struct block *block, *tmp; + + bufmgr_fake->need_fence = 1; + bufmgr_fake->fail = 0; + + /* Wait for hardware idle. We don't know where acceleration has been + * happening, so we'll need to wait anyway before letting anything get + * put on the card again. + */ + dri_bufmgr_fake_wait_idle(bufmgr_fake); + + /* Check that we hadn't released the lock without having fenced the last + * set of buffers. + */ + assert(DRMLISTEMPTY(&bufmgr_fake->fenced)); + assert(DRMLISTEMPTY(&bufmgr_fake->on_hardware)); + + DRMLISTFOREACHSAFE(block, tmp, &bufmgr_fake->lru) { + /* Releases the memory, and memcpys dirty contents out if necessary. */ + free_block(bufmgr_fake, block); + } +} + dri_bufmgr * intel_bufmgr_fake_init(unsigned long low_offset, void *low_virtual, unsigned long size, -- cgit v1.2.3 From 2655005762b8915d5f44d1d1ee7e6c2eb34841d7 Mon Sep 17 00:00:00 2001 From: Eric Anholt Date: Wed, 11 Jun 2008 14:42:40 -0700 Subject: [gem] Move potentially device-specific ioctls to the intel driver. This is the create (may want location flags), pread/pwrite/mmap (performance tuning hints), and set_domain (will 32 bits be enough for everyone?) ioctls. Left in the generic set are just flink/open/close. The 2D driver must be updated for this change, and API but not ABI is broken for 3D. The driver version is bumped to mark this. --- libdrm/intel/intel_bufmgr_fake.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'libdrm/intel/intel_bufmgr_fake.c') diff --git a/libdrm/intel/intel_bufmgr_fake.c b/libdrm/intel/intel_bufmgr_fake.c index 3f5a22d3..06e85b2b 100644 --- a/libdrm/intel/intel_bufmgr_fake.c +++ b/libdrm/intel/intel_bufmgr_fake.c @@ -1062,7 +1062,7 @@ dri_fake_process_relocs(dri_bo *batch_buf) dri_fake_calculate_domains(batch_buf); - batch_fake->read_domains = DRM_GEM_DOMAIN_I915_COMMAND; + batch_fake->read_domains = I915_GEM_DOMAIN_COMMAND; /* we've ran out of RAM so blow the whole lot away and retry */ restart: -- cgit v1.2.3 From 57b4c4c32d3b474939775042bb8a17423def446d Mon Sep 17 00:00:00 2001 From: Eric Anholt Date: Wed, 11 Jun 2008 15:58:33 -0700 Subject: Move the renaming of mm.c symbols to #defines in the header. This reduces the diff from Mesa and reduces the illegibility of what I did. --- libdrm/intel/intel_bufmgr_fake.c | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) (limited to 'libdrm/intel/intel_bufmgr_fake.c') diff --git a/libdrm/intel/intel_bufmgr_fake.c b/libdrm/intel/intel_bufmgr_fake.c index 06e85b2b..e988eb58 100644 --- a/libdrm/intel/intel_bufmgr_fake.c +++ b/libdrm/intel/intel_bufmgr_fake.c @@ -252,7 +252,7 @@ alloc_block(dri_bo *bo) sz = (bo->size + bo_fake->alignment - 1) & ~(bo_fake->alignment - 1); - block->mem = drmmmAllocMem(bufmgr_fake->heap, sz, align_log2, 0); + block->mem = mmAllocMem(bufmgr_fake->heap, sz, align_log2, 0); if (!block->mem) { free(block); return 0; @@ -300,7 +300,7 @@ static void free_block(dri_bufmgr_fake *bufmgr_fake, struct block *block) DBG(" - free immediately\n"); DRMLISTDEL(block); - drmmmFreeMem(block->mem); + mmFreeMem(block->mem); free(block); } } @@ -415,7 +415,7 @@ static int clear_fenced(dri_bufmgr_fake *bufmgr_fake, DBG("delayed free: offset %x sz %x\n", block->mem->ofs, block->mem->size); DRMLISTDEL(block); - drmmmFreeMem(block->mem); + mmFreeMem(block->mem); free(block); } else { @@ -923,7 +923,7 @@ dri_fake_destroy(dri_bufmgr *bufmgr) { dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bufmgr; - drmmmDestroy(bufmgr_fake->heap); + mmDestroy(bufmgr_fake->heap); free(bufmgr); } @@ -1074,7 +1074,7 @@ dri_fake_process_relocs(dri_bo *batch_buf) bufmgr_fake->fail = 0; goto restart; } else /* dump out the memory here */ - drmmmDumpMemInfo(bufmgr_fake->heap); + mmDumpMemInfo(bufmgr_fake->heap); } assert(ret == 0); @@ -1193,7 +1193,7 @@ intel_bufmgr_fake_init(unsigned long low_offset, void *low_virtual, bufmgr_fake->low_offset = low_offset; bufmgr_fake->virtual = low_virtual; bufmgr_fake->size = size; - bufmgr_fake->heap = drmmmInit(low_offset, size); + bufmgr_fake->heap = mmInit(low_offset, size); /* Hook in methods */ bufmgr_fake->bufmgr.bo_alloc = dri_fake_bo_alloc; -- cgit v1.2.3 From 46e9274e8538e5b0517f611dca99dde611f4e95d Mon Sep 17 00:00:00 2001 From: Eric Anholt Date: Fri, 8 Aug 2008 13:13:46 -0700 Subject: Replace the check_aperture API with one we can make thread-safe. While the bufmgr isn't thread-safe at the moment, we need it to be for shared objects between contexts. --- libdrm/intel/intel_bufmgr_fake.c | 63 ++++++++++++++++++++++++---------------- 1 file changed, 38 insertions(+), 25 deletions(-) (limited to 'libdrm/intel/intel_bufmgr_fake.c') diff --git a/libdrm/intel/intel_bufmgr_fake.c b/libdrm/intel/intel_bufmgr_fake.c index e988eb58..63dd9bef 100644 --- a/libdrm/intel/intel_bufmgr_fake.c +++ b/libdrm/intel/intel_bufmgr_fake.c @@ -43,6 +43,8 @@ #include "i915_drm.h" #include "mm.h" +#define ALIGN(value, alignment) ((value + alignment - 1) & ~(alignment - 1)) + #define DBG(...) do { \ if (bufmgr_fake->bufmgr.debug) \ drmMsg(__VA_ARGS__); \ @@ -147,9 +149,6 @@ typedef struct _bufmgr_fake { int debug; int performed_rendering; - - /* keep track of the current total size of objects we have relocs for */ - unsigned long current_total_size; } dri_bufmgr_fake; typedef struct _dri_bo_fake { @@ -159,8 +158,8 @@ typedef struct _dri_bo_fake { const char *name; unsigned dirty:1; - unsigned size_accounted:1; /*this buffers size has been accounted against the aperture */ - unsigned card_dirty:1; /* has the card written to this buffer - we make need to copy it back */ + /** has the card written to this buffer - we make need to copy it back */ + unsigned card_dirty:1; unsigned int refcount; /* Flags may consist of any of the DRM_BO flags, plus * DRM_BO_NO_BACKING_STORE and BM_NO_FENCE_SUBDATA, which are the first two @@ -179,6 +178,12 @@ typedef struct _dri_bo_fake { /** relocation list */ struct fake_buffer_reloc *relocs; int nr_relocs; + /** + * Total size of the target_bos of this buffer. + * + * Used for estimation in check_aperture. + */ + unsigned int child_size; struct block *block; void *backing_store; @@ -189,8 +194,6 @@ typedef struct _dri_bo_fake { static int clear_fenced(dri_bufmgr_fake *bufmgr_fake, unsigned int fence_cookie); -static int dri_fake_check_aperture_space(dri_bo *bo); - #define MAXFENCE 0x7fffffff static int FENCE_LTE( unsigned a, unsigned b ) @@ -855,9 +858,6 @@ dri_fake_bo_validate(dri_bo *bo) return 0; } - /* reset size accounted */ - bo_fake->size_accounted = 0; - /* Allocate the card memory */ if (!bo_fake->block && !evict_and_alloc_block(bo)) { bufmgr_fake->fail = 1; @@ -941,8 +941,6 @@ dri_fake_emit_reloc(dri_bo *reloc_buf, assert(reloc_buf); assert(target_buf); - assert(target_fake->is_static || target_fake->size_accounted); - if (reloc_fake->relocs == NULL) { reloc_fake->relocs = malloc(sizeof(struct fake_buffer_reloc) * MAX_RELOCS); @@ -954,6 +952,9 @@ dri_fake_emit_reloc(dri_bo *reloc_buf, dri_bo_reference(target_buf); + if (!target_fake->is_static) + reloc_fake->child_size += ALIGN(target_buf->size, target_fake->alignment); + r->target_buf = target_buf; r->offset = offset; r->last_target_offset = target_buf->offset; @@ -1079,7 +1080,6 @@ dri_fake_process_relocs(dri_bo *batch_buf) assert(ret == 0); - bufmgr_fake->current_total_size = 0; return NULL; } @@ -1117,26 +1117,39 @@ dri_fake_post_submit(dri_bo *batch_buf) dri_bo_fake_post_submit(batch_buf); } +/** + * Return an error if the list of BOs will exceed the aperture size. + * + * This is a rough guess and likely to fail, as during the validate sequence we + * may place a buffer in an inopportune spot early on and then fail to fit + * a set smaller than the aperture. + */ static int -dri_fake_check_aperture_space(dri_bo *bo) +dri_fake_check_aperture_space(dri_bo **bo_array, int count) { - dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr; - dri_bo_fake *bo_fake = (dri_bo_fake *)bo; - unsigned int sz; + dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo_array[0]->bufmgr; + unsigned int sz = 0; + int i; - sz = (bo->size + bo_fake->alignment - 1) & ~(bo_fake->alignment - 1); + for (i = 0; i < count; i++) { + dri_bo_fake *bo_fake = (dri_bo_fake *)bo_array[i]; - if (bo_fake->size_accounted || bo_fake->is_static) - return 0; + if (bo_fake == NULL) + continue; + + if (!bo_fake->is_static) + sz += ALIGN(bo_array[i]->size, bo_fake->alignment); + sz += bo_fake->child_size; + } - if (bufmgr_fake->current_total_size + sz > bufmgr_fake->size) { - DBG("check_space: %s bo %d %d overflowed bufmgr size %d\n", bo_fake->name, bo_fake->id, sz, bufmgr_fake->size); + if (sz > bufmgr_fake->size) { + DBG("check_space: overflowed bufmgr size, %dkb vs %dkb\n", + sz / 1024, bufmgr_fake->size / 1024); return -1; } - bufmgr_fake->current_total_size += sz; - bo_fake->size_accounted = 1; - DBG("drm_check_space: buf %d, %s %d %d\n", bo_fake->id, bo_fake->name, bo->size, bufmgr_fake->current_total_size); + DBG("drm_check_space: sz %dkb vs bufgr %dkb\n", sz / 1024 , + bufmgr_fake->size / 1024); return 0; } -- cgit v1.2.3