#include <string.h>
#include <unistd.h>
#include <assert.h>
-#include <pthread.h>
#include <sys/ioctl.h>
#include <sys/stat.h>
#include <sys/types.h>
#include <stdbool.h>
#include "errno.h"
-#ifndef ETIME
-#define ETIME ETIMEDOUT
-#endif
#include "common/gen_clflush.h"
-#include "common/gen_debug.h"
-#include "common/gen_device_info.h"
+#include "dev/gen_debug.h"
+#include "common/gen_gem.h"
+#include "dev/gen_device_info.h"
#include "libdrm_macros.h"
#include "main/macros.h"
#include "util/macros.h"
#include "util/hash_table.h"
#include "util/list.h"
+#include "util/u_dynarray.h"
+#include "util/vma.h"
#include "brw_bufmgr.h"
#include "brw_context.h"
#include "string.h"
-#include "i915_drm.h"
+#include "drm-uapi/i915_drm.h"
#ifdef HAVE_VALGRIND
#include <valgrind.h>
#define VG_DEFINED(ptr, size) VG(VALGRIND_MAKE_MEM_DEFINED(ptr, size))
#define VG_NOACCESS(ptr, size) VG(VALGRIND_MAKE_MEM_NOACCESS(ptr, size))
-#define memclear(s) memset(&s, 0, sizeof(s))
+#define PAGE_SIZE 4096
#define FILE_DEBUG_FLAG DEBUG_BUFMGR
return c == unless;
}
+/**
+ * i965 fixed-size bucketing VMA allocator.
+ *
+ * The BO cache maintains "cache buckets" for buffers of various sizes.
+ * All buffers in a given bucket are identically sized - when allocating,
+ * we always round up to the bucket size. This means that virtually all
+ * allocations are fixed-size; only buffers which are too large to fit in
+ * a bucket can be variably-sized.
+ *
+ * We create an allocator for each bucket. Each contains a free-list, where
+ * each node contains a <starting address, 64-bit bitmap> pair. Each bit
+ * represents a bucket-sized block of memory. (At the first level, each
+ * bit corresponds to a page. For the second bucket, bits correspond to
+ * two pages, and so on.) 1 means a block is free, and 0 means it's in-use.
+ * The lowest bit in the bitmap is for the first block.
+ *
+ * This makes allocations cheap - any bit of any node will do. We can pick
+ * the head of the list and use ffs() to find a free block. If there are
+ * none, we allocate 64 blocks from a larger allocator - either a bigger
+ * bucketing allocator, or a fallback top-level allocator for large objects.
+ */
+struct vma_bucket_node {
+ uint64_t start_address;
+ uint64_t bitmap;
+};
+
struct bo_cache_bucket {
+ /** List of cached BOs. */
struct list_head head;
+
+ /** Size of this bucket, in bytes. */
uint64_t size;
+
+ /** List of vma_bucket_nodes. */
+ struct util_dynarray vma_list[BRW_MEMZONE_COUNT];
};
struct brw_bufmgr {
+ uint32_t refcount;
+
+ struct list_head link;
+
int fd;
- pthread_mutex_t lock;
+ mtx_t lock;
/** Array of lists of cached gem objects of power-of-two sizes */
struct bo_cache_bucket cache_bucket[14 * 4];
struct hash_table *name_table;
struct hash_table *handle_table;
+ struct util_vma_heap vma_allocator[BRW_MEMZONE_COUNT];
+
bool has_llc:1;
bool has_mmap_wc:1;
+ bool has_mmap_offset:1;
bool bo_reuse:1;
+
+ uint64_t initial_kflags;
+};
+
+static mtx_t global_bufmgr_list_mutex = _MTX_INITIALIZER_NP;
+static struct list_head global_bufmgr_list = {
+ .next = &global_bufmgr_list,
+ .prev = &global_bufmgr_list,
};
static int bo_set_tiling_internal(struct brw_bo *bo, uint32_t tiling_mode,
static void bo_free(struct brw_bo *bo);
-static uint32_t
-key_hash_uint(const void *key)
-{
- return _mesa_hash_data(key, 4);
-}
-
-static bool
-key_uint_equal(const void *a, const void *b)
-{
- return *((unsigned *) a) == *((unsigned *) b);
-}
+static uint64_t vma_alloc(struct brw_bufmgr *bufmgr,
+ enum brw_memory_zone memzone,
+ uint64_t size, uint64_t alignment);
static struct brw_bo *
hash_find_bo(struct hash_table *ht, unsigned int key)
return size;
/* 965+ just need multiples of page size for tiling */
- return ALIGN(size, 4096);
+ return ALIGN(size, PAGE_SIZE);
}
/*
return ALIGN(pitch, tile_width);
}
+/**
+ * This function finds the correct bucket fit for the input size.
+ * The function works with O(1) complexity when the requested size
+ * was queried instead of iterating the size through all the buckets.
+ */
static struct bo_cache_bucket *
bucket_for_size(struct brw_bufmgr *bufmgr, uint64_t size)
{
- int i;
+ /* Calculating the pages and rounding up to the page size. */
+ const unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
+
+ /* Row Bucket sizes clz((x-1) | 3) Row Column
+ * in pages stride size
+ * 0: 1 2 3 4 -> 30 30 30 30 4 1
+ * 1: 5 6 7 8 -> 29 29 29 29 4 1
+ * 2: 10 12 14 16 -> 28 28 28 28 8 2
+ * 3: 20 24 28 32 -> 27 27 27 27 16 4
+ */
+ const unsigned row = 30 - __builtin_clz((pages - 1) | 3);
+ const unsigned row_max_pages = 4 << row;
- for (i = 0; i < bufmgr->num_buckets; i++) {
- struct bo_cache_bucket *bucket = &bufmgr->cache_bucket[i];
- if (bucket->size >= size) {
- return bucket;
+ /* The '& ~2' is the special case for row 1. In row 1, max pages /
+ * 2 is 2, but the previous row maximum is zero (because there is
+ * no previous row). All row maximum sizes are power of 2, so that
+ * is the only case where that bit will be set.
+ */
+ const unsigned prev_row_max_pages = (row_max_pages / 2) & ~2;
+ int col_size_log2 = row - 1;
+ col_size_log2 += (col_size_log2 < 0);
+
+ const unsigned col = (pages - prev_row_max_pages +
+ ((1 << col_size_log2) - 1)) >> col_size_log2;
+
+ /* Calculating the index based on the row and column. */
+ const unsigned index = (row * 4) + (col - 1);
+
+ return (index < bufmgr->num_buckets) ?
+ &bufmgr->cache_bucket[index] : NULL;
+}
+
+static enum brw_memory_zone
+memzone_for_address(uint64_t address)
+{
+ const uint64_t _4GB = 1ull << 32;
+
+ if (address >= _4GB)
+ return BRW_MEMZONE_OTHER;
+
+ return BRW_MEMZONE_LOW_4G;
+}
+
+static uint64_t
+bucket_vma_alloc(struct brw_bufmgr *bufmgr,
+ struct bo_cache_bucket *bucket,
+ enum brw_memory_zone memzone)
+{
+ struct util_dynarray *vma_list = &bucket->vma_list[memzone];
+ struct vma_bucket_node *node;
+
+ if (vma_list->size == 0) {
+ /* This bucket allocator is out of space - allocate a new block of
+ * memory for 64 blocks from a larger allocator (either a larger
+ * bucket or util_vma).
+ *
+ * We align the address to the node size (64 blocks) so that
+ * bucket_vma_free can easily compute the starting address of this
+ * block by rounding any address we return down to the node size.
+ *
+ * Set the first bit used, and return the start address.
+ */
+ uint64_t node_size = 64ull * bucket->size;
+ node = util_dynarray_grow(vma_list, struct vma_bucket_node, 1);
+
+ if (unlikely(!node))
+ return 0ull;
+
+ uint64_t addr = vma_alloc(bufmgr, memzone, node_size, node_size);
+ node->start_address = gen_48b_address(addr);
+ node->bitmap = ~1ull;
+ return node->start_address;
+ }
+
+ /* Pick any bit from any node - they're all the right size and free. */
+ node = util_dynarray_top_ptr(vma_list, struct vma_bucket_node);
+ int bit = ffsll(node->bitmap) - 1;
+ assert(bit >= 0 && bit <= 63);
+
+ /* Reserve the memory by clearing the bit. */
+ assert((node->bitmap & (1ull << bit)) != 0ull);
+ node->bitmap &= ~(1ull << bit);
+
+ uint64_t addr = node->start_address + bit * bucket->size;
+
+ /* If this node is now completely full, remove it from the free list. */
+ if (node->bitmap == 0ull) {
+ (void) util_dynarray_pop(vma_list, struct vma_bucket_node);
+ }
+
+ return addr;
+}
+
+static void
+bucket_vma_free(struct bo_cache_bucket *bucket, uint64_t address)
+{
+ enum brw_memory_zone memzone = memzone_for_address(address);
+ struct util_dynarray *vma_list = &bucket->vma_list[memzone];
+ const uint64_t node_bytes = 64ull * bucket->size;
+ struct vma_bucket_node *node = NULL;
+
+ /* bucket_vma_alloc allocates 64 blocks at a time, and aligns it to
+ * that 64 block size. So, we can round down to get the starting address.
+ */
+ uint64_t start = (address / node_bytes) * node_bytes;
+
+ /* Dividing the offset from start by bucket size gives us the bit index. */
+ int bit = (address - start) / bucket->size;
+
+ assert(start + bit * bucket->size == address);
+
+ util_dynarray_foreach(vma_list, struct vma_bucket_node, cur) {
+ if (cur->start_address == start) {
+ node = cur;
+ break;
}
}
+ if (!node) {
+ /* No node - the whole group of 64 blocks must have been in-use. */
+ node = util_dynarray_grow(vma_list, struct vma_bucket_node, 1);
+
+ if (unlikely(!node))
+ return; /* bogus, leaks some GPU VMA, but nothing we can do... */
+
+ node->start_address = start;
+ node->bitmap = 0ull;
+ }
+
+ /* Set the bit to return the memory. */
+ assert((node->bitmap & (1ull << bit)) == 0ull);
+ node->bitmap |= 1ull << bit;
+
+ /* The block might be entirely free now, and if so, we could return it
+ * to the larger allocator. But we may as well hang on to it, in case
+ * we get more allocations at this block size.
+ */
+}
+
+static struct bo_cache_bucket *
+get_bucket_allocator(struct brw_bufmgr *bufmgr, uint64_t size)
+{
+ /* Skip using the bucket allocator for very large sizes, as it allocates
+ * 64 of them and this can balloon rather quickly.
+ */
+ if (size > 1024 * PAGE_SIZE)
+ return NULL;
+
+ struct bo_cache_bucket *bucket = bucket_for_size(bufmgr, size);
+
+ if (bucket && bucket->size == size)
+ return bucket;
+
return NULL;
}
+/**
+ * Allocate a section of virtual memory for a buffer, assigning an address.
+ *
+ * This uses either the bucket allocator for the given size, or the large
+ * object allocator (util_vma).
+ */
+static uint64_t
+vma_alloc(struct brw_bufmgr *bufmgr,
+ enum brw_memory_zone memzone,
+ uint64_t size,
+ uint64_t alignment)
+{
+ /* Without softpin support, we let the kernel assign addresses. */
+ assert(brw_using_softpin(bufmgr));
+
+ alignment = ALIGN(alignment, PAGE_SIZE);
+
+ struct bo_cache_bucket *bucket = get_bucket_allocator(bufmgr, size);
+ uint64_t addr;
+
+ if (bucket) {
+ addr = bucket_vma_alloc(bufmgr, bucket, memzone);
+ } else {
+ addr = util_vma_heap_alloc(&bufmgr->vma_allocator[memzone], size,
+ alignment);
+ }
+
+ assert((addr >> 48ull) == 0);
+ assert((addr % alignment) == 0);
+
+ return gen_canonical_address(addr);
+}
+
+/**
+ * Free a virtual memory area, allowing the address to be reused.
+ */
+static void
+vma_free(struct brw_bufmgr *bufmgr,
+ uint64_t address,
+ uint64_t size)
+{
+ assert(brw_using_softpin(bufmgr));
+
+ /* Un-canonicalize the address. */
+ address = gen_48b_address(address);
+
+ if (address == 0ull)
+ return;
+
+ struct bo_cache_bucket *bucket = get_bucket_allocator(bufmgr, size);
+
+ if (bucket) {
+ bucket_vma_free(bucket, address);
+ } else {
+ enum brw_memory_zone memzone = memzone_for_address(address);
+ util_vma_heap_free(&bufmgr->vma_allocator[memzone], address, size);
+ }
+}
+
int
brw_bo_busy(struct brw_bo *bo)
{
struct brw_bufmgr *bufmgr = bo->bufmgr;
- struct drm_i915_gem_busy busy;
- int ret;
+ struct drm_i915_gem_busy busy = { .handle = bo->gem_handle };
- memclear(busy);
- busy.handle = bo->gem_handle;
-
- ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_BUSY, &busy);
+ int ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_BUSY, &busy);
if (ret == 0) {
bo->idle = !busy.busy;
return busy.busy;
int
brw_bo_madvise(struct brw_bo *bo, int state)
{
- struct drm_i915_gem_madvise madv;
+ struct drm_i915_gem_madvise madv = {
+ .handle = bo->gem_handle,
+ .madv = state,
+ .retained = 1,
+ };
- memclear(madv);
- madv.handle = bo->gem_handle;
- madv.madv = state;
- madv.retained = 1;
drmIoctl(bo->bufmgr->fd, DRM_IOCTL_I915_GEM_MADVISE, &madv);
return madv.retained;
bo_alloc_internal(struct brw_bufmgr *bufmgr,
const char *name,
uint64_t size,
+ enum brw_memory_zone memzone,
unsigned flags,
uint32_t tiling_mode,
- uint32_t stride, uint64_t alignment)
+ uint32_t stride)
{
struct brw_bo *bo;
- unsigned int page_size = getpagesize();
int ret;
struct bo_cache_bucket *bucket;
bool alloc_from_cache;
* allocation up.
*/
if (bucket == NULL) {
- bo_size = size;
- if (bo_size < page_size)
- bo_size = page_size;
+ unsigned int page_size = getpagesize();
+ bo_size = size == 0 ? page_size : ALIGN(size, page_size);
} else {
bo_size = bucket->size;
}
+ assert(bo_size);
- pthread_mutex_lock(&bufmgr->lock);
+ mtx_lock(&bufmgr->lock);
/* Get a buffer out of the cache if available */
retry:
alloc_from_cache = false;
- if (bucket != NULL && !list_empty(&bucket->head)) {
+ if (bucket != NULL && !list_is_empty(&bucket->head)) {
if (busy && !zeroed) {
/* Allocate new render-target BOs from the tail (MRU)
* of the list, as it will likely be hot in the GPU
bo = LIST_ENTRY(struct brw_bo, bucket->head.prev, head);
list_del(&bo->head);
alloc_from_cache = true;
- bo->align = alignment;
} else {
- assert(alignment == 0);
/* For non-render-target BOs (where we're probably
* going to map it first thing in order to fill it
* with data), check if the last BO in the cache is
}
}
- if (!alloc_from_cache) {
- struct drm_i915_gem_create create;
-
+ if (alloc_from_cache) {
+ /* If the cache BO isn't in the right memory zone, free the old
+ * memory and assign it a new address.
+ */
+ if ((bo->kflags & EXEC_OBJECT_PINNED) &&
+ memzone != memzone_for_address(bo->gtt_offset)) {
+ vma_free(bufmgr, bo->gtt_offset, bo->size);
+ bo->gtt_offset = 0ull;
+ }
+ } else {
bo = calloc(1, sizeof(*bo));
if (!bo)
goto err;
bo->size = bo_size;
bo->idle = true;
- memclear(create);
- create.size = bo_size;
+ struct drm_i915_gem_create create = { .size = bo_size };
/* All new BOs we get from the kernel are zeroed, so we don't need to
* worry about that here.
}
bo->gem_handle = create.handle;
- _mesa_hash_table_insert(bufmgr->handle_table, &bo->gem_handle, bo);
bo->bufmgr = bufmgr;
- bo->align = alignment;
bo->tiling_mode = I915_TILING_NONE;
bo->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
bo->reusable = true;
bo->cache_coherent = bufmgr->has_llc;
bo->index = -1;
+ bo->kflags = bufmgr->initial_kflags;
- pthread_mutex_unlock(&bufmgr->lock);
+ if ((bo->kflags & EXEC_OBJECT_PINNED) && bo->gtt_offset == 0ull) {
+ bo->gtt_offset = vma_alloc(bufmgr, memzone, bo->size, 1);
+
+ if (bo->gtt_offset == 0ull)
+ goto err_free;
+ }
+
+ mtx_unlock(&bufmgr->lock);
DBG("bo_create: buf %d (%s) %llub\n", bo->gem_handle, bo->name,
(unsigned long long) size);
err_free:
bo_free(bo);
err:
- pthread_mutex_unlock(&bufmgr->lock);
+ mtx_unlock(&bufmgr->lock);
return NULL;
}
struct brw_bo *
brw_bo_alloc(struct brw_bufmgr *bufmgr,
- const char *name, uint64_t size, uint64_t alignment)
+ const char *name, uint64_t size,
+ enum brw_memory_zone memzone)
{
- return bo_alloc_internal(bufmgr, name, size, 0, I915_TILING_NONE, 0, 0);
+ return bo_alloc_internal(bufmgr, name, size, memzone,
+ 0, I915_TILING_NONE, 0);
}
struct brw_bo *
brw_bo_alloc_tiled(struct brw_bufmgr *bufmgr, const char *name,
- uint64_t size, uint32_t tiling_mode, uint32_t pitch,
+ uint64_t size, enum brw_memory_zone memzone,
+ uint32_t tiling_mode, uint32_t pitch,
unsigned flags)
{
- return bo_alloc_internal(bufmgr, name, size, flags, tiling_mode, pitch, 0);
+ return bo_alloc_internal(bufmgr, name, size, memzone,
+ flags, tiling_mode, pitch);
}
struct brw_bo *
brw_bo_alloc_tiled_2d(struct brw_bufmgr *bufmgr, const char *name,
- int x, int y, int cpp, uint32_t tiling,
- uint32_t *pitch, unsigned flags)
+ int x, int y, int cpp, enum brw_memory_zone memzone,
+ uint32_t tiling, uint32_t *pitch, unsigned flags)
{
uint64_t size;
uint32_t stride;
if (tiling == I915_TILING_NONE)
stride = 0;
- return bo_alloc_internal(bufmgr, name, size, flags, tiling, stride, 0);
+ return bo_alloc_internal(bufmgr, name, size, memzone,
+ flags, tiling, stride);
}
/**
const char *name, unsigned int handle)
{
struct brw_bo *bo;
- int ret;
- struct drm_gem_open open_arg;
- struct drm_i915_gem_get_tiling get_tiling;
/* At the moment most applications only have a few named bo.
* For instance, in a DRI client only the render buffers passed
* alternating names for the front/back buffer a linear search
* provides a sufficiently fast match.
*/
- pthread_mutex_lock(&bufmgr->lock);
+ mtx_lock(&bufmgr->lock);
bo = hash_find_bo(bufmgr->name_table, handle);
if (bo) {
brw_bo_reference(bo);
goto out;
}
- memclear(open_arg);
- open_arg.name = handle;
- ret = drmIoctl(bufmgr->fd, DRM_IOCTL_GEM_OPEN, &open_arg);
+ struct drm_gem_open open_arg = { .name = handle };
+ int ret = drmIoctl(bufmgr->fd, DRM_IOCTL_GEM_OPEN, &open_arg);
if (ret != 0) {
DBG("Couldn't reference %s handle 0x%08x: %s\n",
name, handle, strerror(errno));
bo->global_name = handle;
bo->reusable = false;
bo->external = true;
+ bo->kflags = bufmgr->initial_kflags;
+
+ if (bo->kflags & EXEC_OBJECT_PINNED)
+ bo->gtt_offset = vma_alloc(bufmgr, BRW_MEMZONE_OTHER, bo->size, 1);
_mesa_hash_table_insert(bufmgr->handle_table, &bo->gem_handle, bo);
_mesa_hash_table_insert(bufmgr->name_table, &bo->global_name, bo);
- memclear(get_tiling);
- get_tiling.handle = bo->gem_handle;
+ struct drm_i915_gem_get_tiling get_tiling = { .handle = bo->gem_handle };
ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_GET_TILING, &get_tiling);
if (ret != 0)
goto err_unref;
DBG("bo_create_from_handle: %d (%s)\n", handle, bo->name);
out:
- pthread_mutex_unlock(&bufmgr->lock);
+ mtx_unlock(&bufmgr->lock);
return bo;
err_unref:
bo_free(bo);
- pthread_mutex_unlock(&bufmgr->lock);
+ mtx_unlock(&bufmgr->lock);
return NULL;
}
bo_free(struct brw_bo *bo)
{
struct brw_bufmgr *bufmgr = bo->bufmgr;
- struct drm_gem_close close;
- struct hash_entry *entry;
- int ret;
if (bo->map_cpu) {
VG_NOACCESS(bo->map_cpu, bo->size);
drm_munmap(bo->map_gtt, bo->size);
}
- if (bo->global_name) {
- entry = _mesa_hash_table_search(bufmgr->name_table, &bo->global_name);
- _mesa_hash_table_remove(bufmgr->name_table, entry);
+ if (bo->external) {
+ struct hash_entry *entry;
+
+ if (bo->global_name) {
+ entry = _mesa_hash_table_search(bufmgr->name_table, &bo->global_name);
+ _mesa_hash_table_remove(bufmgr->name_table, entry);
+ }
+
+ entry = _mesa_hash_table_search(bufmgr->handle_table, &bo->gem_handle);
+ _mesa_hash_table_remove(bufmgr->handle_table, entry);
}
- entry = _mesa_hash_table_search(bufmgr->handle_table, &bo->gem_handle);
- _mesa_hash_table_remove(bufmgr->handle_table, entry);
/* Close this object */
- memclear(close);
- close.handle = bo->gem_handle;
- ret = drmIoctl(bufmgr->fd, DRM_IOCTL_GEM_CLOSE, &close);
+ struct drm_gem_close close = { .handle = bo->gem_handle };
+ int ret = drmIoctl(bufmgr->fd, DRM_IOCTL_GEM_CLOSE, &close);
if (ret != 0) {
DBG("DRM_IOCTL_GEM_CLOSE %d failed (%s): %s\n",
bo->gem_handle, bo->name, strerror(errno));
}
+
+ if (bo->kflags & EXEC_OBJECT_PINNED)
+ vma_free(bo->bufmgr, bo->gtt_offset, bo->size);
+
free(bo);
}
bo->free_time = time;
bo->name = NULL;
- bo->kflags = 0;
list_addtail(&bo->head, &bucket->head);
} else {
clock_gettime(CLOCK_MONOTONIC, &time);
- pthread_mutex_lock(&bufmgr->lock);
+ mtx_lock(&bufmgr->lock);
if (p_atomic_dec_zero(&bo->refcount)) {
bo_unreference_final(bo, time.tv_sec);
cleanup_bo_cache(bufmgr, time.tv_sec);
}
- pthread_mutex_unlock(&bufmgr->lock);
+ mtx_unlock(&bufmgr->lock);
}
}
struct brw_bo *bo,
const char *action)
{
- double elapsed = unlikely(brw && brw->perf_debug) ? -get_time() : 0.0;
+ bool busy = brw && brw->perf_debug && !bo->idle;
+ double elapsed = unlikely(busy) ? -get_time() : 0.0;
brw_bo_wait_rendering(bo);
- if (unlikely(brw && brw->perf_debug)) {
+ if (unlikely(busy)) {
elapsed += get_time();
if (elapsed > 1e-5) /* 0.01ms */
perf_debug("%s a busy \"%s\" BO stalled and took %.03f ms.\n",
}
static void *
-brw_bo_map_cpu(struct brw_context *brw, struct brw_bo *bo, unsigned flags)
+brw_bo_gem_mmap_legacy(struct brw_context *brw, struct brw_bo *bo, bool wc)
+{
+ struct brw_bufmgr *bufmgr = bo->bufmgr;
+
+ struct drm_i915_gem_mmap mmap_arg = {
+ .handle = bo->gem_handle,
+ .size = bo->size,
+ .flags = wc ? I915_MMAP_WC : 0,
+ };
+
+ int ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_MMAP, &mmap_arg);
+ if (ret != 0) {
+ DBG("%s:%d: Error mapping buffer %d (%s): %s .\n",
+ __FILE__, __LINE__, bo->gem_handle, bo->name, strerror(errno));
+ return NULL;
+ }
+ void *map = (void *) (uintptr_t) mmap_arg.addr_ptr;
+
+ return map;
+}
+
+static void *
+brw_bo_gem_mmap_offset(struct brw_context *brw, struct brw_bo *bo, bool wc)
+{
+ struct brw_bufmgr *bufmgr = bo->bufmgr;
+
+ struct drm_i915_gem_mmap_offset mmap_arg = {
+ .handle = bo->gem_handle,
+ .flags = wc ? I915_MMAP_OFFSET_WC : I915_MMAP_OFFSET_WB,
+ };
+
+ /* Get the fake offset back */
+ int ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_MMAP_OFFSET, &mmap_arg);
+ if (ret != 0) {
+ DBG("%s:%d: Error preparing buffer %d (%s): %s .\n",
+ __FILE__, __LINE__, bo->gem_handle, bo->name, strerror(errno));
+ return NULL;
+ }
+
+ /* And map it */
+ void *map = drm_mmap(0, bo->size, PROT_READ | PROT_WRITE, MAP_SHARED,
+ bufmgr->fd, mmap_arg.offset);
+ if (map == MAP_FAILED) {
+ DBG("%s:%d: Error mapping buffer %d (%s): %s .\n",
+ __FILE__, __LINE__, bo->gem_handle, bo->name, strerror(errno));
+ return NULL;
+ }
+
+ return map;
+}
+
+static void *
+brw_bo_gem_mmap(struct brw_context *brw, struct brw_bo *bo, bool wc)
{
struct brw_bufmgr *bufmgr = bo->bufmgr;
+ if (bufmgr->has_mmap_offset)
+ return brw_bo_gem_mmap_offset(brw, bo, wc);
+ else
+ return brw_bo_gem_mmap_legacy(brw, bo, wc);
+}
+
+static void *
+brw_bo_map_cpu(struct brw_context *brw, struct brw_bo *bo, unsigned flags)
+{
/* We disallow CPU maps for writing to non-coherent buffers, as the
* CPU map can become invalidated when a batch is flushed out, which
* can happen at unpredictable times. You should use WC maps instead.
assert(bo->cache_coherent || !(flags & MAP_WRITE));
if (!bo->map_cpu) {
- struct drm_i915_gem_mmap mmap_arg;
- void *map;
-
DBG("brw_bo_map_cpu: %d (%s)\n", bo->gem_handle, bo->name);
- memclear(mmap_arg);
- mmap_arg.handle = bo->gem_handle;
- mmap_arg.size = bo->size;
- int ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_MMAP, &mmap_arg);
- if (ret != 0) {
- ret = -errno;
- DBG("%s:%d: Error mapping buffer %d (%s): %s .\n",
- __FILE__, __LINE__, bo->gem_handle, bo->name, strerror(errno));
- return NULL;
- }
- map = (void *) (uintptr_t) mmap_arg.addr_ptr;
+ void *map = brw_bo_gem_mmap(brw, bo, false);
VG_DEFINED(map, bo->size);
if (p_atomic_cmpxchg(&bo->map_cpu, NULL, map)) {
return NULL;
if (!bo->map_wc) {
- struct drm_i915_gem_mmap mmap_arg;
- void *map;
-
DBG("brw_bo_map_wc: %d (%s)\n", bo->gem_handle, bo->name);
-
- memclear(mmap_arg);
- mmap_arg.handle = bo->gem_handle;
- mmap_arg.size = bo->size;
- mmap_arg.flags = I915_MMAP_WC;
- int ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_MMAP, &mmap_arg);
- if (ret != 0) {
- ret = -errno;
- DBG("%s:%d: Error mapping buffer %d (%s): %s .\n",
- __FILE__, __LINE__, bo->gem_handle, bo->name, strerror(errno));
- return NULL;
- }
-
- map = (void *) (uintptr_t) mmap_arg.addr_ptr;
+ void *map = brw_bo_gem_mmap(brw, bo, true);
VG_DEFINED(map, bo->size);
if (p_atomic_cmpxchg(&bo->map_wc, NULL, map)) {
/* Get a mapping of the buffer if we haven't before. */
if (bo->map_gtt == NULL) {
- struct drm_i915_gem_mmap_gtt mmap_arg;
- void *map;
-
DBG("bo_map_gtt: mmap %d (%s)\n", bo->gem_handle, bo->name);
- memclear(mmap_arg);
- mmap_arg.handle = bo->gem_handle;
+ struct drm_i915_gem_mmap_gtt mmap_arg = { .handle = bo->gem_handle };
/* Get the fake offset back... */
int ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_MMAP_GTT, &mmap_arg);
}
/* and mmap it. */
- map = drm_mmap(0, bo->size, PROT_READ | PROT_WRITE,
- MAP_SHARED, bufmgr->fd, mmap_arg.offset);
+ void *map = drm_mmap(0, bo->size, PROT_READ | PROT_WRITE,
+ MAP_SHARED, bufmgr->fd, mmap_arg.offset);
if (map == MAP_FAILED) {
DBG("%s:%d: Error mapping buffer %d (%s): %s .\n",
__FILE__, __LINE__, bo->gem_handle, bo->name, strerror(errno));
uint64_t size, const void *data)
{
struct brw_bufmgr *bufmgr = bo->bufmgr;
- struct drm_i915_gem_pwrite pwrite;
- int ret;
- memclear(pwrite);
- pwrite.handle = bo->gem_handle;
- pwrite.offset = offset;
- pwrite.size = size;
- pwrite.data_ptr = (uint64_t) (uintptr_t) data;
- ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_PWRITE, &pwrite);
+ struct drm_i915_gem_pwrite pwrite = {
+ .handle = bo->gem_handle,
+ .offset = offset,
+ .size = size,
+ .data_ptr = (uint64_t) (uintptr_t) data,
+ };
+
+ int ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_PWRITE, &pwrite);
if (ret != 0) {
ret = -errno;
DBG("%s:%d: Error writing data to buffer %d: "
brw_bo_wait(struct brw_bo *bo, int64_t timeout_ns)
{
struct brw_bufmgr *bufmgr = bo->bufmgr;
- struct drm_i915_gem_wait wait;
- int ret;
/* If we know it's idle, don't bother with the kernel round trip */
if (bo->idle && !bo->external)
return 0;
- memclear(wait);
- wait.bo_handle = bo->gem_handle;
- wait.timeout_ns = timeout_ns;
- ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_WAIT, &wait);
- if (ret == -1)
+ struct drm_i915_gem_wait wait = {
+ .bo_handle = bo->gem_handle,
+ .timeout_ns = timeout_ns,
+ };
+ int ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_WAIT, &wait);
+ if (ret != 0)
return -errno;
bo->idle = true;
}
void
-brw_bufmgr_destroy(struct brw_bufmgr *bufmgr)
+brw_bufmgr_unref(struct brw_bufmgr *bufmgr)
{
- pthread_mutex_destroy(&bufmgr->lock);
+ mtx_lock(&global_bufmgr_list_mutex);
+ if (p_atomic_dec_zero(&bufmgr->refcount)) {
+ list_del(&bufmgr->link);
+ } else {
+ bufmgr = NULL;
+ }
+ mtx_unlock(&global_bufmgr_list_mutex);
+
+ if (!bufmgr)
+ return;
+
+ mtx_destroy(&bufmgr->lock);
/* Free any cached buffer objects we were going to reuse */
for (int i = 0; i < bufmgr->num_buckets; i++) {
bo_free(bo);
}
+
+ if (brw_using_softpin(bufmgr)) {
+ for (int z = 0; z < BRW_MEMZONE_COUNT; z++) {
+ util_dynarray_fini(&bucket->vma_list[z]);
+ }
+ }
}
_mesa_hash_table_destroy(bufmgr->name_table, NULL);
_mesa_hash_table_destroy(bufmgr->handle_table, NULL);
+ if (brw_using_softpin(bufmgr)) {
+ for (int z = 0; z < BRW_MEMZONE_COUNT; z++) {
+ util_vma_heap_finish(&bufmgr->vma_allocator[z]);
+ }
+ }
+
+ close(bufmgr->fd);
+ bufmgr->fd = -1;
+
free(bufmgr);
}
return 0;
}
-struct brw_bo *
-brw_bo_gem_create_from_prime(struct brw_bufmgr *bufmgr, int prime_fd)
+static struct brw_bo *
+brw_bo_gem_create_from_prime_internal(struct brw_bufmgr *bufmgr, int prime_fd,
+ int tiling_mode, uint32_t stride)
{
- int ret;
uint32_t handle;
struct brw_bo *bo;
- struct drm_i915_gem_get_tiling get_tiling;
- pthread_mutex_lock(&bufmgr->lock);
- ret = drmPrimeFDToHandle(bufmgr->fd, prime_fd, &handle);
+ mtx_lock(&bufmgr->lock);
+ int ret = drmPrimeFDToHandle(bufmgr->fd, prime_fd, &handle);
if (ret) {
DBG("create_from_prime: failed to obtain handle from fd: %s\n",
strerror(errno));
- pthread_mutex_unlock(&bufmgr->lock);
+ mtx_unlock(&bufmgr->lock);
return NULL;
}
bo->name = "prime";
bo->reusable = false;
bo->external = true;
+ bo->kflags = bufmgr->initial_kflags;
- memclear(get_tiling);
- get_tiling.handle = bo->gem_handle;
- if (drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_GET_TILING, &get_tiling))
- goto err;
+ if (bo->kflags & EXEC_OBJECT_PINNED) {
+ assert(bo->size > 0);
+ bo->gtt_offset = vma_alloc(bufmgr, BRW_MEMZONE_OTHER, bo->size, 1);
+ }
- bo->tiling_mode = get_tiling.tiling_mode;
- bo->swizzle_mode = get_tiling.swizzle_mode;
- /* XXX stride is unknown */
+ if (tiling_mode < 0) {
+ struct drm_i915_gem_get_tiling get_tiling = { .handle = bo->gem_handle };
+ if (drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_GET_TILING, &get_tiling))
+ goto err;
+
+ bo->tiling_mode = get_tiling.tiling_mode;
+ bo->swizzle_mode = get_tiling.swizzle_mode;
+ /* XXX stride is unknown */
+ } else {
+ bo_set_tiling_internal(bo, tiling_mode, stride);
+ }
out:
- pthread_mutex_unlock(&bufmgr->lock);
+ mtx_unlock(&bufmgr->lock);
return bo;
err:
bo_free(bo);
- pthread_mutex_unlock(&bufmgr->lock);
+ mtx_unlock(&bufmgr->lock);
return NULL;
}
+struct brw_bo *
+brw_bo_gem_create_from_prime(struct brw_bufmgr *bufmgr, int prime_fd)
+{
+ return brw_bo_gem_create_from_prime_internal(bufmgr, prime_fd, -1, 0);
+}
+
+struct brw_bo *
+brw_bo_gem_create_from_prime_tiled(struct brw_bufmgr *bufmgr, int prime_fd,
+ uint32_t tiling_mode, uint32_t stride)
+{
+ assert(tiling_mode == I915_TILING_NONE ||
+ tiling_mode == I915_TILING_X ||
+ tiling_mode == I915_TILING_Y);
+
+ return brw_bo_gem_create_from_prime_internal(bufmgr, prime_fd,
+ tiling_mode, stride);
+}
+
+static void
+brw_bo_make_external(struct brw_bo *bo)
+{
+ struct brw_bufmgr *bufmgr = bo->bufmgr;
+
+ if (!bo->external) {
+ mtx_lock(&bufmgr->lock);
+ if (!bo->external) {
+ _mesa_hash_table_insert(bufmgr->handle_table, &bo->gem_handle, bo);
+ bo->external = true;
+ }
+ mtx_unlock(&bufmgr->lock);
+ }
+}
+
int
brw_bo_gem_export_to_prime(struct brw_bo *bo, int *prime_fd)
{
struct brw_bufmgr *bufmgr = bo->bufmgr;
+ brw_bo_make_external(bo);
+
if (drmPrimeHandleToFD(bufmgr->fd, bo->gem_handle,
DRM_CLOEXEC, prime_fd) != 0)
return -errno;
bo->reusable = false;
- bo->external = true;
return 0;
}
+uint32_t
+brw_bo_export_gem_handle(struct brw_bo *bo)
+{
+ brw_bo_make_external(bo);
+
+ return bo->gem_handle;
+}
+
int
brw_bo_flink(struct brw_bo *bo, uint32_t *name)
{
struct brw_bufmgr *bufmgr = bo->bufmgr;
if (!bo->global_name) {
- struct drm_gem_flink flink;
+ struct drm_gem_flink flink = { .handle = bo->gem_handle };
- memclear(flink);
- flink.handle = bo->gem_handle;
if (drmIoctl(bufmgr->fd, DRM_IOCTL_GEM_FLINK, &flink))
return -errno;
- pthread_mutex_lock(&bufmgr->lock);
+ brw_bo_make_external(bo);
+ mtx_lock(&bufmgr->lock);
if (!bo->global_name) {
bo->global_name = flink.name;
- bo->reusable = false;
- bo->external = true;
-
_mesa_hash_table_insert(bufmgr->name_table, &bo->global_name, bo);
}
- pthread_mutex_unlock(&bufmgr->lock);
+ mtx_unlock(&bufmgr->lock);
+
+ bo->reusable = false;
}
*name = bo->global_name;
return 0;
}
-/**
- * Enables unlimited caching of buffer objects for reuse.
- *
- * This is potentially very memory expensive, as the cache at each bucket
- * size is only bounded by how many buffers of that size we've managed to have
- * in flight at once.
- */
-void
-brw_bufmgr_enable_reuse(struct brw_bufmgr *bufmgr)
-{
- bufmgr->bo_reuse = true;
-}
-
static void
add_bucket(struct brw_bufmgr *bufmgr, int size)
{
assert(i < ARRAY_SIZE(bufmgr->cache_bucket));
list_inithead(&bufmgr->cache_bucket[i].head);
+ if (brw_using_softpin(bufmgr)) {
+ for (int z = 0; z < BRW_MEMZONE_COUNT; z++)
+ util_dynarray_init(&bufmgr->cache_bucket[i].vma_list[z], NULL);
+ }
bufmgr->cache_bucket[i].size = size;
bufmgr->num_buckets++;
+
+ assert(bucket_for_size(bufmgr, size) == &bufmgr->cache_bucket[i]);
+ assert(bucket_for_size(bufmgr, size - 2048) == &bufmgr->cache_bucket[i]);
+ assert(bucket_for_size(bufmgr, size + 1) != &bufmgr->cache_bucket[i]);
}
static void
* width/height alignment and rounding of sizes to pages will
* get us useful cache hit rates anyway)
*/
- add_bucket(bufmgr, 4096);
- add_bucket(bufmgr, 4096 * 2);
- add_bucket(bufmgr, 4096 * 3);
+ add_bucket(bufmgr, PAGE_SIZE);
+ add_bucket(bufmgr, PAGE_SIZE * 2);
+ add_bucket(bufmgr, PAGE_SIZE * 3);
/* Initialize the linked lists for BO reuse cache. */
- for (size = 4 * 4096; size <= cache_max_size; size *= 2) {
+ for (size = 4 * PAGE_SIZE; size <= cache_max_size; size *= 2) {
add_bucket(bufmgr, size);
add_bucket(bufmgr, size + size * 1 / 4);
uint32_t
brw_create_hw_context(struct brw_bufmgr *bufmgr)
{
- struct drm_i915_gem_context_create create;
- int ret;
-
- memclear(create);
- ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_CONTEXT_CREATE, &create);
+ struct drm_i915_gem_context_create create = { };
+ int ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_CONTEXT_CREATE, &create);
if (ret != 0) {
DBG("DRM_IOCTL_I915_GEM_CONTEXT_CREATE failed: %s\n", strerror(errno));
return 0;
return create.ctx_id;
}
+int
+brw_hw_context_set_priority(struct brw_bufmgr *bufmgr,
+ uint32_t ctx_id,
+ int priority)
+{
+ struct drm_i915_gem_context_param p = {
+ .ctx_id = ctx_id,
+ .param = I915_CONTEXT_PARAM_PRIORITY,
+ .value = priority,
+ };
+ int err;
+
+ err = 0;
+ if (drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_CONTEXT_SETPARAM, &p))
+ err = -errno;
+
+ return err;
+}
+
void
brw_destroy_hw_context(struct brw_bufmgr *bufmgr, uint32_t ctx_id)
{
- struct drm_i915_gem_context_destroy d = {.ctx_id = ctx_id };
+ struct drm_i915_gem_context_destroy d = { .ctx_id = ctx_id };
if (ctx_id != 0 &&
drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_CONTEXT_DESTROY, &d) != 0) {
int
brw_reg_read(struct brw_bufmgr *bufmgr, uint32_t offset, uint64_t *result)
{
- struct drm_i915_reg_read reg_read;
- int ret;
-
- memclear(reg_read);
- reg_read.offset = offset;
-
- ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_REG_READ, ®_read);
+ struct drm_i915_reg_read reg_read = { .offset = offset };
+ int ret = drmIoctl(bufmgr->fd, DRM_IOCTL_I915_REG_READ, ®_read);
*result = reg_read.val;
return ret;
static int
gem_param(int fd, int name)
{
- drm_i915_getparam_t gp;
int v = -1; /* No param uses (yet) the sign bit, reserve it for errors */
- memset(&gp, 0, sizeof(gp));
- gp.param = name;
- gp.value = &v;
+ struct drm_i915_getparam gp = { .param = name, .value = &v };
if (drmIoctl(fd, DRM_IOCTL_I915_GETPARAM, &gp))
return -1;
return v;
}
+static int
+gem_context_getparam(int fd, uint32_t context, uint64_t param, uint64_t *value)
+{
+ struct drm_i915_gem_context_param gp = {
+ .ctx_id = context,
+ .param = param,
+ };
+
+ if (drmIoctl(fd, DRM_IOCTL_I915_GEM_CONTEXT_GETPARAM, &gp))
+ return -1;
+
+ *value = gp.value;
+
+ return 0;
+}
+
+bool
+brw_using_softpin(struct brw_bufmgr *bufmgr)
+{
+ return bufmgr->initial_kflags & EXEC_OBJECT_PINNED;
+}
+
+static struct brw_bufmgr *
+brw_bufmgr_ref(struct brw_bufmgr *bufmgr)
+{
+ p_atomic_inc(&bufmgr->refcount);
+ return bufmgr;
+}
+
/**
* Initializes the GEM buffer manager, which uses the kernel to allocate, map,
* and manage map buffer objections.
*
* \param fd File descriptor of the opened DRM device.
*/
-struct brw_bufmgr *
-brw_bufmgr_init(struct gen_device_info *devinfo, int fd)
+static struct brw_bufmgr *
+brw_bufmgr_create(struct gen_device_info *devinfo, int fd, bool bo_reuse)
{
struct brw_bufmgr *bufmgr;
* Don't do this! Ensure that each library/bufmgr has its own device
* fd so that its namespace does not clash with another.
*/
- bufmgr->fd = fd;
+ bufmgr->fd = dup(fd);
+ if (bufmgr->fd < 0) {
+ free(bufmgr);
+ return NULL;
+ }
- if (pthread_mutex_init(&bufmgr->lock, NULL) != 0) {
+ p_atomic_set(&bufmgr->refcount, 1);
+
+ if (mtx_init(&bufmgr->lock, mtx_plain) != 0) {
+ close(bufmgr->fd);
free(bufmgr);
return NULL;
}
+ uint64_t gtt_size;
+ if (gem_context_getparam(fd, 0, I915_CONTEXT_PARAM_GTT_SIZE, >t_size))
+ gtt_size = 0;
+
bufmgr->has_llc = devinfo->has_llc;
bufmgr->has_mmap_wc = gem_param(fd, I915_PARAM_MMAP_VERSION) > 0;
+ bufmgr->bo_reuse = bo_reuse;
+ bufmgr->has_mmap_offset = gem_param(fd, I915_PARAM_MMAP_GTT_VERSION) >= 4;
+
+ const uint64_t _4GB = 4ull << 30;
+
+ /* The STATE_BASE_ADDRESS size field can only hold 1 page shy of 4GB */
+ const uint64_t _4GB_minus_1 = _4GB - PAGE_SIZE;
+
+ if (devinfo->gen >= 8 && gtt_size > _4GB) {
+ bufmgr->initial_kflags |= EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
+
+ /* Allocate VMA in userspace if we have softpin and full PPGTT. */
+ if (gem_param(fd, I915_PARAM_HAS_EXEC_SOFTPIN) > 0 &&
+ gem_param(fd, I915_PARAM_HAS_ALIASING_PPGTT) > 1) {
+ bufmgr->initial_kflags |= EXEC_OBJECT_PINNED;
+
+ util_vma_heap_init(&bufmgr->vma_allocator[BRW_MEMZONE_LOW_4G],
+ PAGE_SIZE, _4GB_minus_1);
+
+ /* Leave the last 4GB out of the high vma range, so that no state
+ * base address + size can overflow 48 bits.
+ */
+ util_vma_heap_init(&bufmgr->vma_allocator[BRW_MEMZONE_OTHER],
+ 1 * _4GB, gtt_size - 2 * _4GB);
+ } else if (devinfo->gen >= 10) {
+ /* Softpin landed in 4.5, but GVT used an aliasing PPGTT until
+ * kernel commit 6b3816d69628becb7ff35978aa0751798b4a940a in
+ * 4.14. Gen10+ GVT hasn't landed yet, so it's not actually a
+ * problem - but extending this requirement back to earlier gens
+ * might actually mean requiring 4.14.
+ */
+ fprintf(stderr, "i965 requires softpin (Kernel 4.5) on Gen10+.");
+ close(bufmgr->fd);
+ free(bufmgr);
+ return NULL;
+ }
+ }
init_cache_buckets(bufmgr);
bufmgr->name_table =
- _mesa_hash_table_create(NULL, key_hash_uint, key_uint_equal);
+ _mesa_hash_table_create(NULL, _mesa_hash_uint, _mesa_key_uint_equal);
bufmgr->handle_table =
- _mesa_hash_table_create(NULL, key_hash_uint, key_uint_equal);
+ _mesa_hash_table_create(NULL, _mesa_hash_uint, _mesa_key_uint_equal);
return bufmgr;
}
+
+struct brw_bufmgr *
+brw_bufmgr_get_for_fd(struct gen_device_info *devinfo, int fd, bool bo_reuse)
+{
+ struct stat st;
+
+ if (fstat(fd, &st))
+ return NULL;
+
+ struct brw_bufmgr *bufmgr = NULL;
+
+ mtx_lock(&global_bufmgr_list_mutex);
+ list_for_each_entry(struct brw_bufmgr, iter_bufmgr, &global_bufmgr_list, link) {
+ struct stat iter_st;
+ if (fstat(iter_bufmgr->fd, &iter_st))
+ continue;
+
+ if (st.st_rdev == iter_st.st_rdev) {
+ assert(iter_bufmgr->bo_reuse == bo_reuse);
+ bufmgr = brw_bufmgr_ref(iter_bufmgr);
+ goto unlock;
+ }
+ }
+
+ bufmgr = brw_bufmgr_create(devinfo, fd, bo_reuse);
+ list_addtail(&bufmgr->link, &global_bufmgr_list);
+
+ unlock:
+ mtx_unlock(&global_bufmgr_list_mutex);
+
+ return bufmgr;
+}
+
+int
+brw_bufmgr_get_fd(struct brw_bufmgr *bufmgr)
+{
+ return bufmgr->fd;
+}
projects / mesa.git / blobdiff