/**************************************************************************
*
* Copyright © 2007 Red Hat Inc.
- * Copyright © 2007 Intel Corporation
+ * Copyright © 2007-2012 Intel Corporation
* Copyright 2006 Tungsten Graphics, Inc., Bismarck, ND., USA
* All Rights Reserved.
*
#include <stdbool.h>
#include "errno.h"
+#ifndef ETIME
+#define ETIME ETIMEDOUT
+#endif
+#include "libdrm.h"
#include "libdrm_lists.h"
#include "intel_bufmgr.h"
#include "intel_bufmgr_priv.h"
#include "intel_chipset.h"
+#include "intel_aub.h"
#include "string.h"
#include "i915_drm.h"
+#ifdef HAVE_VALGRIND
+#include <valgrind.h>
+#include <memcheck.h>
+#define VG(x) x
+#else
+#define VG(x)
+#endif
+
+#define VG_CLEAR(s) VG(memset(&s, 0, sizeof(s)))
+
#define DBG(...) do { \
if (bufmgr_gem->bufmgr.debug) \
fprintf(stderr, __VA_ARGS__); \
typedef struct _drm_intel_bufmgr_gem {
drm_intel_bufmgr bufmgr;
+ atomic_t refcount;
+
int fd;
int max_relocs;
int num_buckets;
time_t time;
+ drmMMListHead managers;
+
drmMMListHead named;
+ drmMMListHead vma_cache;
+ int vma_count, vma_open, vma_max;
uint64_t gtt_size;
int available_fences;
unsigned int has_bsd : 1;
unsigned int has_blt : 1;
unsigned int has_relaxed_fencing : 1;
+ unsigned int has_llc : 1;
+ unsigned int has_wait_timeout : 1;
unsigned int bo_reuse : 1;
+ unsigned int no_exec : 1;
+ unsigned int has_vebox : 1;
bool fenced_relocs;
+
+ char *aub_filename;
+ FILE *aub_file;
+ uint32_t aub_offset;
} drm_intel_bufmgr_gem;
#define DRM_INTEL_RELOC_FENCE (1<<0)
/**
* Kenel-assigned global name for this object
+ *
+ * List contains both flink named and prime fd'd objects
*/
unsigned int global_name;
drmMMListHead name_list;
void *mem_virtual;
/** GTT virtual address for the buffer, saved across map/unmap cycles */
void *gtt_virtual;
+ /**
+ * Virtual address of the buffer allocated by user, used for userptr
+ * objects only.
+ */
+ void *user_virtual;
+ int map_count;
+ drmMMListHead vma_list;
/** BO cache list */
drmMMListHead head;
bool reusable;
/**
+ * Boolean of whether the GPU is definitely not accessing the buffer.
+ *
+ * This is only valid when reusable, since non-reusable
+ * buffers are those that have been shared wth other
+ * processes, so we don't know their state.
+ */
+ bool idle;
+
+ /**
+ * Boolean of whether this buffer was allocated with userptr
+ */
+ bool is_userptr;
+
+ /**
* Size in bytes of this buffer and its relocation descendents.
*
* Used to avoid costly tree walking in
* relocations.
*/
int reloc_tree_fences;
+
+ /** Flags that we may need to do the SW_FINSIH ioctl on unmap. */
+ bool mapped_cpu_write;
+
+ uint32_t aub_offset;
+
+ drm_intel_aub_annotation *aub_annotations;
+ unsigned aub_annotation_count;
};
static unsigned int
return ALIGN(pitch, 64);
if (*tiling_mode == I915_TILING_X
- || (IS_915(bufmgr_gem) && *tiling_mode == I915_TILING_Y))
+ || (IS_915(bufmgr_gem->pci_device)
+ && *tiling_mode == I915_TILING_Y))
tile_width = 512;
else
tile_width = 128;
(unsigned long long)bo_gem->relocs[j].offset,
target_gem->gem_handle,
target_gem->name,
- target_bo->offset,
+ target_bo->offset64,
bo_gem->relocs[j].delta);
}
}
struct drm_i915_gem_busy busy;
int ret;
- memset(&busy, 0, sizeof(busy));
+ if (bo_gem->reusable && bo_gem->idle)
+ return false;
+
+ VG_CLEAR(busy);
busy.handle = bo_gem->gem_handle;
ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_BUSY, &busy);
-
+ if (ret == 0) {
+ bo_gem->idle = !busy.busy;
+ return busy.busy;
+ } else {
+ return false;
+ }
return (ret == 0 && busy.busy);
}
{
struct drm_i915_gem_madvise madv;
+ VG_CLEAR(madv);
madv.handle = bo_gem->gem_handle;
madv.madv = state;
madv.retained = 1;
return NULL;
bo_gem->bo.size = bo_size;
- memset(&create, 0, sizeof(create));
+
+ VG_CLEAR(create);
create.size = bo_size;
ret = drmIoctl(bufmgr_gem->fd,
}
DRMINITLISTHEAD(&bo_gem->name_list);
+ DRMINITLISTHEAD(&bo_gem->vma_list);
}
bo_gem->name = name;
bo_gem->used_as_reloc_target = false;
bo_gem->has_error = false;
bo_gem->reusable = true;
+ bo_gem->aub_annotations = NULL;
+ bo_gem->aub_annotation_count = 0;
drm_intel_bo_gem_set_in_aperture_size(bufmgr_gem, bo_gem);
aligned_y = y;
height_alignment = 2;
- if (IS_GEN2(bufmgr_gem) && tiling != I915_TILING_NONE)
+ if ((bufmgr_gem->gen == 2) && tiling != I915_TILING_NONE)
height_alignment = 16;
else if (tiling == I915_TILING_X
- || (IS_915(bufmgr_gem) && tiling == I915_TILING_Y))
+ || (IS_915(bufmgr_gem->pci_device)
+ && tiling == I915_TILING_Y))
height_alignment = 8;
else if (tiling == I915_TILING_Y)
height_alignment = 32;
tiling, stride);
}
+static drm_intel_bo *
+drm_intel_gem_bo_alloc_userptr(drm_intel_bufmgr *bufmgr,
+ const char *name,
+ void *addr,
+ uint32_t tiling_mode,
+ uint32_t stride,
+ unsigned long size,
+ unsigned long flags)
+{
+ drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bufmgr;
+ drm_intel_bo_gem *bo_gem;
+ int ret;
+ struct drm_i915_gem_userptr userptr;
+
+ /* Tiling with userptr surfaces is not supported
+ * on all hardware so refuse it for time being.
+ */
+ if (tiling_mode != I915_TILING_NONE)
+ return NULL;
+
+ bo_gem = calloc(1, sizeof(*bo_gem));
+ if (!bo_gem)
+ return NULL;
+
+ bo_gem->bo.size = size;
+
+ VG_CLEAR(userptr);
+ userptr.user_ptr = (__u64)((unsigned long)addr);
+ userptr.user_size = size;
+ userptr.flags = flags;
+
+ ret = drmIoctl(bufmgr_gem->fd,
+ DRM_IOCTL_I915_GEM_USERPTR,
+ &userptr);
+ if (ret != 0) {
+ DBG("bo_create_userptr: "
+ "ioctl failed with user ptr %p size 0x%lx, "
+ "user flags 0x%lx\n", addr, size, flags);
+ free(bo_gem);
+ return NULL;
+ }
+
+ bo_gem->gem_handle = userptr.handle;
+ bo_gem->bo.handle = bo_gem->gem_handle;
+ bo_gem->bo.bufmgr = bufmgr;
+ bo_gem->is_userptr = true;
+ bo_gem->bo.virtual = addr;
+ /* Save the address provided by user */
+ bo_gem->user_virtual = addr;
+ bo_gem->tiling_mode = I915_TILING_NONE;
+ bo_gem->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
+ bo_gem->stride = 0;
+
+ DRMINITLISTHEAD(&bo_gem->name_list);
+ DRMINITLISTHEAD(&bo_gem->vma_list);
+
+ bo_gem->name = name;
+ atomic_set(&bo_gem->refcount, 1);
+ bo_gem->validate_index = -1;
+ bo_gem->reloc_tree_fences = 0;
+ bo_gem->used_as_reloc_target = false;
+ bo_gem->has_error = false;
+ bo_gem->reusable = false;
+
+ drm_intel_bo_gem_set_in_aperture_size(bufmgr_gem, bo_gem);
+
+ DBG("bo_create_userptr: "
+ "ptr %p buf %d (%s) size %ldb, stride 0x%x, tile mode %d\n",
+ addr, bo_gem->gem_handle, bo_gem->name,
+ size, stride, tiling_mode);
+
+ return &bo_gem->bo;
+}
+
/**
* Returns a drm_intel_bo wrapping the given buffer object handle.
*
* This can be used when one application needs to pass a buffer object
* to another.
*/
-drm_intel_bo *
+drm_public drm_intel_bo *
drm_intel_bo_gem_create_from_name(drm_intel_bufmgr *bufmgr,
const char *name,
unsigned int handle)
* alternating names for the front/back buffer a linear search
* provides a sufficiently fast match.
*/
+ pthread_mutex_lock(&bufmgr_gem->lock);
for (list = bufmgr_gem->named.next;
list != &bufmgr_gem->named;
list = list->next) {
bo_gem = DRMLISTENTRY(drm_intel_bo_gem, list, name_list);
if (bo_gem->global_name == handle) {
drm_intel_gem_bo_reference(&bo_gem->bo);
+ pthread_mutex_unlock(&bufmgr_gem->lock);
return &bo_gem->bo;
}
}
- bo_gem = calloc(1, sizeof(*bo_gem));
- if (!bo_gem)
- return NULL;
-
- memset(&open_arg, 0, sizeof(open_arg));
+ VG_CLEAR(open_arg);
open_arg.name = handle;
ret = drmIoctl(bufmgr_gem->fd,
DRM_IOCTL_GEM_OPEN,
if (ret != 0) {
DBG("Couldn't reference %s handle 0x%08x: %s\n",
name, handle, strerror(errno));
- free(bo_gem);
+ pthread_mutex_unlock(&bufmgr_gem->lock);
+ return NULL;
+ }
+ /* Now see if someone has used a prime handle to get this
+ * object from the kernel before by looking through the list
+ * again for a matching gem_handle
+ */
+ for (list = bufmgr_gem->named.next;
+ list != &bufmgr_gem->named;
+ list = list->next) {
+ bo_gem = DRMLISTENTRY(drm_intel_bo_gem, list, name_list);
+ if (bo_gem->gem_handle == open_arg.handle) {
+ drm_intel_gem_bo_reference(&bo_gem->bo);
+ pthread_mutex_unlock(&bufmgr_gem->lock);
+ return &bo_gem->bo;
+ }
+ }
+
+ bo_gem = calloc(1, sizeof(*bo_gem));
+ if (!bo_gem) {
+ pthread_mutex_unlock(&bufmgr_gem->lock);
return NULL;
}
+
bo_gem->bo.size = open_arg.size;
bo_gem->bo.offset = 0;
+ bo_gem->bo.offset64 = 0;
bo_gem->bo.virtual = NULL;
bo_gem->bo.bufmgr = bufmgr;
bo_gem->name = name;
bo_gem->global_name = handle;
bo_gem->reusable = false;
- memset(&get_tiling, 0, sizeof(get_tiling));
+ VG_CLEAR(get_tiling);
get_tiling.handle = bo_gem->gem_handle;
ret = drmIoctl(bufmgr_gem->fd,
DRM_IOCTL_I915_GEM_GET_TILING,
&get_tiling);
if (ret != 0) {
drm_intel_gem_bo_unreference(&bo_gem->bo);
+ pthread_mutex_unlock(&bufmgr_gem->lock);
return NULL;
}
bo_gem->tiling_mode = get_tiling.tiling_mode;
/* XXX stride is unknown */
drm_intel_bo_gem_set_in_aperture_size(bufmgr_gem, bo_gem);
+ DRMINITLISTHEAD(&bo_gem->vma_list);
DRMLISTADDTAIL(&bo_gem->name_list, &bufmgr_gem->named);
+ pthread_mutex_unlock(&bufmgr_gem->lock);
DBG("bo_create_from_handle: %d (%s)\n", handle, bo_gem->name);
return &bo_gem->bo;
struct drm_gem_close close;
int ret;
- if (bo_gem->mem_virtual)
+ DRMLISTDEL(&bo_gem->vma_list);
+ if (bo_gem->mem_virtual) {
+ VG(VALGRIND_FREELIKE_BLOCK(bo_gem->mem_virtual, 0));
munmap(bo_gem->mem_virtual, bo_gem->bo.size);
- if (bo_gem->gtt_virtual)
+ bufmgr_gem->vma_count--;
+ }
+ if (bo_gem->gtt_virtual) {
munmap(bo_gem->gtt_virtual, bo_gem->bo.size);
+ bufmgr_gem->vma_count--;
+ }
/* Close this object */
- memset(&close, 0, sizeof(close));
+ VG_CLEAR(close);
close.handle = bo_gem->gem_handle;
ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_GEM_CLOSE, &close);
if (ret != 0) {
DBG("DRM_IOCTL_GEM_CLOSE %d failed (%s): %s\n",
bo_gem->gem_handle, bo_gem->name, strerror(errno));
}
+ free(bo_gem->aub_annotations);
free(bo);
}
+static void
+drm_intel_gem_bo_mark_mmaps_incoherent(drm_intel_bo *bo)
+{
+#if HAVE_VALGRIND
+ drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
+
+ if (bo_gem->mem_virtual)
+ VALGRIND_MAKE_MEM_NOACCESS(bo_gem->mem_virtual, bo->size);
+
+ if (bo_gem->gtt_virtual)
+ VALGRIND_MAKE_MEM_NOACCESS(bo_gem->gtt_virtual, bo->size);
+#endif
+}
+
/** Frees all cached buffers significantly older than @time. */
static void
drm_intel_gem_cleanup_bo_cache(drm_intel_bufmgr_gem *bufmgr_gem, time_t time)
bufmgr_gem->time = time;
}
+static void drm_intel_gem_bo_purge_vma_cache(drm_intel_bufmgr_gem *bufmgr_gem)
+{
+ int limit;
+
+ DBG("%s: cached=%d, open=%d, limit=%d\n", __FUNCTION__,
+ bufmgr_gem->vma_count, bufmgr_gem->vma_open, bufmgr_gem->vma_max);
+
+ if (bufmgr_gem->vma_max < 0)
+ return;
+
+ /* We may need to evict a few entries in order to create new mmaps */
+ limit = bufmgr_gem->vma_max - 2*bufmgr_gem->vma_open;
+ if (limit < 0)
+ limit = 0;
+
+ while (bufmgr_gem->vma_count > limit) {
+ drm_intel_bo_gem *bo_gem;
+
+ bo_gem = DRMLISTENTRY(drm_intel_bo_gem,
+ bufmgr_gem->vma_cache.next,
+ vma_list);
+ assert(bo_gem->map_count == 0);
+ DRMLISTDELINIT(&bo_gem->vma_list);
+
+ if (bo_gem->mem_virtual) {
+ munmap(bo_gem->mem_virtual, bo_gem->bo.size);
+ bo_gem->mem_virtual = NULL;
+ bufmgr_gem->vma_count--;
+ }
+ if (bo_gem->gtt_virtual) {
+ munmap(bo_gem->gtt_virtual, bo_gem->bo.size);
+ bo_gem->gtt_virtual = NULL;
+ bufmgr_gem->vma_count--;
+ }
+ }
+}
+
+static void drm_intel_gem_bo_close_vma(drm_intel_bufmgr_gem *bufmgr_gem,
+ drm_intel_bo_gem *bo_gem)
+{
+ bufmgr_gem->vma_open--;
+ DRMLISTADDTAIL(&bo_gem->vma_list, &bufmgr_gem->vma_cache);
+ if (bo_gem->mem_virtual)
+ bufmgr_gem->vma_count++;
+ if (bo_gem->gtt_virtual)
+ bufmgr_gem->vma_count++;
+ drm_intel_gem_bo_purge_vma_cache(bufmgr_gem);
+}
+
+static void drm_intel_gem_bo_open_vma(drm_intel_bufmgr_gem *bufmgr_gem,
+ drm_intel_bo_gem *bo_gem)
+{
+ bufmgr_gem->vma_open++;
+ DRMLISTDEL(&bo_gem->vma_list);
+ if (bo_gem->mem_virtual)
+ bufmgr_gem->vma_count--;
+ if (bo_gem->gtt_virtual)
+ bufmgr_gem->vma_count--;
+ drm_intel_gem_bo_purge_vma_cache(bufmgr_gem);
+}
+
static void
drm_intel_gem_bo_unreference_final(drm_intel_bo *bo, time_t time)
{
bo_gem->relocs = NULL;
}
+ /* Clear any left-over mappings */
+ if (bo_gem->map_count) {
+ DBG("bo freed with non-zero map-count %d\n", bo_gem->map_count);
+ bo_gem->map_count = 0;
+ drm_intel_gem_bo_close_vma(bufmgr_gem, bo_gem);
+ drm_intel_gem_bo_mark_mmaps_incoherent(bo);
+ }
+
DRMLISTDEL(&bo_gem->name_list);
bucket = drm_intel_gem_bo_bucket_for_size(bufmgr_gem, bo->size);
drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
assert(atomic_read(&bo_gem->refcount) > 0);
- if (atomic_dec_and_test(&bo_gem->refcount)) {
+
+ if (atomic_add_unless(&bo_gem->refcount, -1, 1)) {
drm_intel_bufmgr_gem *bufmgr_gem =
(drm_intel_bufmgr_gem *) bo->bufmgr;
struct timespec time;
clock_gettime(CLOCK_MONOTONIC, &time);
pthread_mutex_lock(&bufmgr_gem->lock);
- drm_intel_gem_bo_unreference_final(bo, time.tv_sec);
- drm_intel_gem_cleanup_bo_cache(bufmgr_gem, time.tv_sec);
+
+ if (atomic_dec_and_test(&bo_gem->refcount)) {
+ drm_intel_gem_bo_unreference_final(bo, time.tv_sec);
+ drm_intel_gem_cleanup_bo_cache(bufmgr_gem, time.tv_sec);
+ }
+
pthread_mutex_unlock(&bufmgr_gem->lock);
}
}
struct drm_i915_gem_set_domain set_domain;
int ret;
+ if (bo_gem->is_userptr) {
+ /* Return the same user ptr */
+ bo->virtual = bo_gem->user_virtual;
+ return 0;
+ }
+
pthread_mutex_lock(&bufmgr_gem->lock);
- /* Allow recursive mapping. Mesa may recursively map buffers with
- * nested display loops.
- */
+ if (bo_gem->map_count++ == 0)
+ drm_intel_gem_bo_open_vma(bufmgr_gem, bo_gem);
+
if (!bo_gem->mem_virtual) {
struct drm_i915_gem_mmap mmap_arg;
- DBG("bo_map: %d (%s)\n", bo_gem->gem_handle, bo_gem->name);
+ DBG("bo_map: %d (%s), map_count=%d\n",
+ bo_gem->gem_handle, bo_gem->name, bo_gem->map_count);
- memset(&mmap_arg, 0, sizeof(mmap_arg));
+ VG_CLEAR(mmap_arg);
mmap_arg.handle = bo_gem->gem_handle;
mmap_arg.offset = 0;
mmap_arg.size = bo->size;
DBG("%s:%d: Error mapping buffer %d (%s): %s .\n",
__FILE__, __LINE__, bo_gem->gem_handle,
bo_gem->name, strerror(errno));
+ if (--bo_gem->map_count == 0)
+ drm_intel_gem_bo_close_vma(bufmgr_gem, bo_gem);
pthread_mutex_unlock(&bufmgr_gem->lock);
return ret;
}
+ VG(VALGRIND_MALLOCLIKE_BLOCK(mmap_arg.addr_ptr, mmap_arg.size, 0, 1));
bo_gem->mem_virtual = (void *)(uintptr_t) mmap_arg.addr_ptr;
}
DBG("bo_map: %d (%s) -> %p\n", bo_gem->gem_handle, bo_gem->name,
bo_gem->mem_virtual);
bo->virtual = bo_gem->mem_virtual;
+ VG_CLEAR(set_domain);
set_domain.handle = bo_gem->gem_handle;
set_domain.read_domains = I915_GEM_DOMAIN_CPU;
if (write_enable)
strerror(errno));
}
+ if (write_enable)
+ bo_gem->mapped_cpu_write = true;
+
+ drm_intel_gem_bo_mark_mmaps_incoherent(bo);
+ VG(VALGRIND_MAKE_MEM_DEFINED(bo_gem->mem_virtual, bo->size));
pthread_mutex_unlock(&bufmgr_gem->lock);
return 0;
}
-int drm_intel_gem_bo_map_gtt(drm_intel_bo *bo)
+static int
+map_gtt(drm_intel_bo *bo)
{
drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
- struct drm_i915_gem_set_domain set_domain;
int ret;
- pthread_mutex_lock(&bufmgr_gem->lock);
+ if (bo_gem->is_userptr)
+ return -EINVAL;
+
+ if (bo_gem->map_count++ == 0)
+ drm_intel_gem_bo_open_vma(bufmgr_gem, bo_gem);
/* Get a mapping of the buffer if we haven't before. */
if (bo_gem->gtt_virtual == NULL) {
struct drm_i915_gem_mmap_gtt mmap_arg;
- DBG("bo_map_gtt: mmap %d (%s)\n", bo_gem->gem_handle,
- bo_gem->name);
+ DBG("bo_map_gtt: mmap %d (%s), map_count=%d\n",
+ bo_gem->gem_handle, bo_gem->name, bo_gem->map_count);
- memset(&mmap_arg, 0, sizeof(mmap_arg));
+ VG_CLEAR(mmap_arg);
mmap_arg.handle = bo_gem->gem_handle;
/* Get the fake offset back... */
__FILE__, __LINE__,
bo_gem->gem_handle, bo_gem->name,
strerror(errno));
- pthread_mutex_unlock(&bufmgr_gem->lock);
+ if (--bo_gem->map_count == 0)
+ drm_intel_gem_bo_close_vma(bufmgr_gem, bo_gem);
return ret;
}
__FILE__, __LINE__,
bo_gem->gem_handle, bo_gem->name,
strerror(errno));
- pthread_mutex_unlock(&bufmgr_gem->lock);
+ if (--bo_gem->map_count == 0)
+ drm_intel_gem_bo_close_vma(bufmgr_gem, bo_gem);
return ret;
}
}
DBG("bo_map_gtt: %d (%s) -> %p\n", bo_gem->gem_handle, bo_gem->name,
bo_gem->gtt_virtual);
- /* Now move it to the GTT domain so that the CPU caches are flushed */
+ return 0;
+}
+
+drm_public int
+drm_intel_gem_bo_map_gtt(drm_intel_bo *bo)
+{
+ drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
+ drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
+ struct drm_i915_gem_set_domain set_domain;
+ int ret;
+
+ pthread_mutex_lock(&bufmgr_gem->lock);
+
+ ret = map_gtt(bo);
+ if (ret) {
+ pthread_mutex_unlock(&bufmgr_gem->lock);
+ return ret;
+ }
+
+ /* Now move it to the GTT domain so that the GPU and CPU
+ * caches are flushed and the GPU isn't actively using the
+ * buffer.
+ *
+ * The pagefault handler does this domain change for us when
+ * it has unbound the BO from the GTT, but it's up to us to
+ * tell it when we're about to use things if we had done
+ * rendering and it still happens to be bound to the GTT.
+ */
+ VG_CLEAR(set_domain);
set_domain.handle = bo_gem->gem_handle;
set_domain.read_domains = I915_GEM_DOMAIN_GTT;
set_domain.write_domain = I915_GEM_DOMAIN_GTT;
strerror(errno));
}
+ drm_intel_gem_bo_mark_mmaps_incoherent(bo);
+ VG(VALGRIND_MAKE_MEM_DEFINED(bo_gem->gtt_virtual, bo->size));
pthread_mutex_unlock(&bufmgr_gem->lock);
return 0;
}
-int drm_intel_gem_bo_unmap_gtt(drm_intel_bo *bo)
+/**
+ * Performs a mapping of the buffer object like the normal GTT
+ * mapping, but avoids waiting for the GPU to be done reading from or
+ * rendering to the buffer.
+ *
+ * This is used in the implementation of GL_ARB_map_buffer_range: The
+ * user asks to create a buffer, then does a mapping, fills some
+ * space, runs a drawing command, then asks to map it again without
+ * synchronizing because it guarantees that it won't write over the
+ * data that the GPU is busy using (or, more specifically, that if it
+ * does write over the data, it acknowledges that rendering is
+ * undefined).
+ */
+
+drm_public int
+drm_intel_gem_bo_map_unsynchronized(drm_intel_bo *bo)
{
drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
- int ret = 0;
+#ifdef HAVE_VALGRIND
+ drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
+#endif
+ int ret;
- if (bo == NULL)
- return 0;
+ /* If the CPU cache isn't coherent with the GTT, then use a
+ * regular synchronized mapping. The problem is that we don't
+ * track where the buffer was last used on the CPU side in
+ * terms of drm_intel_bo_map vs drm_intel_gem_bo_map_gtt, so
+ * we would potentially corrupt the buffer even when the user
+ * does reasonable things.
+ */
+ if (!bufmgr_gem->has_llc)
+ return drm_intel_gem_bo_map_gtt(bo);
pthread_mutex_lock(&bufmgr_gem->lock);
- bo->virtual = NULL;
+
+ ret = map_gtt(bo);
+ if (ret == 0) {
+ drm_intel_gem_bo_mark_mmaps_incoherent(bo);
+ VG(VALGRIND_MAKE_MEM_DEFINED(bo_gem->gtt_virtual, bo->size));
+ }
+
pthread_mutex_unlock(&bufmgr_gem->lock);
return ret;
static int drm_intel_gem_bo_unmap(drm_intel_bo *bo)
{
- drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
+ drm_intel_bufmgr_gem *bufmgr_gem;
drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
- struct drm_i915_gem_sw_finish sw_finish;
- int ret;
+ int ret = 0;
if (bo == NULL)
return 0;
+ if (bo_gem->is_userptr)
+ return 0;
+
+ bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
+
pthread_mutex_lock(&bufmgr_gem->lock);
- /* Cause a flush to happen if the buffer's pinned for scanout, so the
- * results show up in a timely manner.
- */
- sw_finish.handle = bo_gem->gem_handle;
- ret = drmIoctl(bufmgr_gem->fd,
- DRM_IOCTL_I915_GEM_SW_FINISH,
- &sw_finish);
- ret = ret == -1 ? -errno : 0;
+ if (bo_gem->map_count <= 0) {
+ DBG("attempted to unmap an unmapped bo\n");
+ pthread_mutex_unlock(&bufmgr_gem->lock);
+ /* Preserve the old behaviour of just treating this as a
+ * no-op rather than reporting the error.
+ */
+ return 0;
+ }
+
+ if (bo_gem->mapped_cpu_write) {
+ struct drm_i915_gem_sw_finish sw_finish;
+
+ /* Cause a flush to happen if the buffer's pinned for
+ * scanout, so the results show up in a timely manner.
+ * Unlike GTT set domains, this only does work if the
+ * buffer should be scanout-related.
+ */
+ VG_CLEAR(sw_finish);
+ sw_finish.handle = bo_gem->gem_handle;
+ ret = drmIoctl(bufmgr_gem->fd,
+ DRM_IOCTL_I915_GEM_SW_FINISH,
+ &sw_finish);
+ ret = ret == -1 ? -errno : 0;
+
+ bo_gem->mapped_cpu_write = false;
+ }
- bo->virtual = NULL;
+ /* We need to unmap after every innovation as we cannot track
+ * an open vma for every bo as that will exhaasut the system
+ * limits and cause later failures.
+ */
+ if (--bo_gem->map_count == 0) {
+ drm_intel_gem_bo_close_vma(bufmgr_gem, bo_gem);
+ drm_intel_gem_bo_mark_mmaps_incoherent(bo);
+ bo->virtual = NULL;
+ }
pthread_mutex_unlock(&bufmgr_gem->lock);
return ret;
}
+drm_public int
+drm_intel_gem_bo_unmap_gtt(drm_intel_bo *bo)
+{
+ return drm_intel_gem_bo_unmap(bo);
+}
+
static int
drm_intel_gem_bo_subdata(drm_intel_bo *bo, unsigned long offset,
unsigned long size, const void *data)
struct drm_i915_gem_pwrite pwrite;
int ret;
- memset(&pwrite, 0, sizeof(pwrite));
+ if (bo_gem->is_userptr)
+ return -EINVAL;
+
+ VG_CLEAR(pwrite);
pwrite.handle = bo_gem->gem_handle;
pwrite.offset = offset;
pwrite.size = size;
struct drm_i915_get_pipe_from_crtc_id get_pipe_from_crtc_id;
int ret;
+ VG_CLEAR(get_pipe_from_crtc_id);
get_pipe_from_crtc_id.crtc_id = crtc_id;
ret = drmIoctl(bufmgr_gem->fd,
DRM_IOCTL_I915_GET_PIPE_FROM_CRTC_ID,
struct drm_i915_gem_pread pread;
int ret;
- memset(&pread, 0, sizeof(pread));
+ if (bo_gem->is_userptr)
+ return -EINVAL;
+
+ VG_CLEAR(pread);
pread.handle = bo_gem->gem_handle;
pread.offset = offset;
pread.size = size;
}
/**
+ * Waits on a BO for the given amount of time.
+ *
+ * @bo: buffer object to wait for
+ * @timeout_ns: amount of time to wait in nanoseconds.
+ * If value is less than 0, an infinite wait will occur.
+ *
+ * Returns 0 if the wait was successful ie. the last batch referencing the
+ * object has completed within the allotted time. Otherwise some negative return
+ * value describes the error. Of particular interest is -ETIME when the wait has
+ * failed to yield the desired result.
+ *
+ * Similar to drm_intel_gem_bo_wait_rendering except a timeout parameter allows
+ * the operation to give up after a certain amount of time. Another subtle
+ * difference is the internal locking semantics are different (this variant does
+ * not hold the lock for the duration of the wait). This makes the wait subject
+ * to a larger userspace race window.
+ *
+ * The implementation shall wait until the object is no longer actively
+ * referenced within a batch buffer at the time of the call. The wait will
+ * not guarantee that the buffer is re-issued via another thread, or an flinked
+ * handle. Userspace must make sure this race does not occur if such precision
+ * is important.
+ */
+drm_public int
+drm_intel_gem_bo_wait(drm_intel_bo *bo, int64_t timeout_ns)
+{
+ drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
+ drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
+ struct drm_i915_gem_wait wait;
+ int ret;
+
+ if (!bufmgr_gem->has_wait_timeout) {
+ DBG("%s:%d: Timed wait is not supported. Falling back to "
+ "infinite wait\n", __FILE__, __LINE__);
+ if (timeout_ns) {
+ drm_intel_gem_bo_wait_rendering(bo);
+ return 0;
+ } else {
+ return drm_intel_gem_bo_busy(bo) ? -ETIME : 0;
+ }
+ }
+
+ wait.bo_handle = bo_gem->gem_handle;
+ wait.timeout_ns = timeout_ns;
+ wait.flags = 0;
+ ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_WAIT, &wait);
+ if (ret == -1)
+ return -errno;
+
+ return ret;
+}
+
+/**
* Sets the object to the GTT read and possibly write domain, used by the X
* 2D driver in the absence of kernel support to do drm_intel_gem_bo_map_gtt().
*
* In combination with drm_intel_gem_bo_pin() and manual fence management, we
* can do tiled pixmaps this way.
*/
-void
+drm_public void
drm_intel_gem_bo_start_gtt_access(drm_intel_bo *bo, int write_enable)
{
drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
struct drm_i915_gem_set_domain set_domain;
int ret;
+ VG_CLEAR(set_domain);
set_domain.handle = bo_gem->gem_handle;
set_domain.read_domains = I915_GEM_DOMAIN_GTT;
set_domain.write_domain = write_enable ? I915_GEM_DOMAIN_GTT : 0;
free(bufmgr_gem->exec2_objects);
free(bufmgr_gem->exec_objects);
free(bufmgr_gem->exec_bos);
+ free(bufmgr_gem->aub_filename);
pthread_mutex_destroy(&bufmgr_gem->lock);
target_bo_gem->gem_handle;
bo_gem->relocs[bo_gem->reloc_count].read_domains = read_domains;
bo_gem->relocs[bo_gem->reloc_count].write_domain = write_domain;
- bo_gem->relocs[bo_gem->reloc_count].presumed_offset = target_bo->offset;
+ bo_gem->relocs[bo_gem->reloc_count].presumed_offset = target_bo->offset64;
bo_gem->reloc_target_info[bo_gem->reloc_count].bo = target_bo;
if (target_bo != bo)
read_domains, write_domain, true);
}
+drm_public int
+drm_intel_gem_bo_get_reloc_count(drm_intel_bo *bo)
+{
+ drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
+
+ return bo_gem->reloc_count;
+}
+
+/**
+ * Removes existing relocation entries in the BO after "start".
+ *
+ * This allows a user to avoid a two-step process for state setup with
+ * counting up all the buffer objects and doing a
+ * drm_intel_bufmgr_check_aperture_space() before emitting any of the
+ * relocations for the state setup. Instead, save the state of the
+ * batchbuffer including drm_intel_gem_get_reloc_count(), emit all the
+ * state, and then check if it still fits in the aperture.
+ *
+ * Any further drm_intel_bufmgr_check_aperture_space() queries
+ * involving this buffer in the tree are undefined after this call.
+ */
+drm_public void
+drm_intel_gem_bo_clear_relocs(drm_intel_bo *bo, int start)
+{
+ drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
+ drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
+ int i;
+ struct timespec time;
+
+ clock_gettime(CLOCK_MONOTONIC, &time);
+
+ assert(bo_gem->reloc_count >= start);
+
+ /* Unreference the cleared target buffers */
+ pthread_mutex_lock(&bufmgr_gem->lock);
+
+ for (i = start; i < bo_gem->reloc_count; i++) {
+ drm_intel_bo_gem *target_bo_gem = (drm_intel_bo_gem *) bo_gem->reloc_target_info[i].bo;
+ if (&target_bo_gem->bo != bo) {
+ bo_gem->reloc_tree_fences -= target_bo_gem->reloc_tree_fences;
+ drm_intel_gem_bo_unreference_locked_timed(&target_bo_gem->bo,
+ time.tv_sec);
+ }
+ }
+ bo_gem->reloc_count = start;
+
+ pthread_mutex_unlock(&bufmgr_gem->lock);
+
+}
+
/**
* Walk the tree of relocations rooted at BO and accumulate the list of
* validations to be performed and update the relocation buffers with
if (target_bo == bo)
continue;
+ drm_intel_gem_bo_mark_mmaps_incoherent(bo);
+
/* Continue walking the tree depth-first. */
drm_intel_gem_bo_process_reloc(target_bo);
if (target_bo == bo)
continue;
+ drm_intel_gem_bo_mark_mmaps_incoherent(bo);
+
/* Continue walking the tree depth-first. */
drm_intel_gem_bo_process_reloc2(target_bo);
drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
/* Update the buffer offset */
- if (bufmgr_gem->exec_objects[i].offset != bo->offset) {
+ if (bufmgr_gem->exec_objects[i].offset != bo->offset64) {
DBG("BO %d (%s) migrated: 0x%08lx -> 0x%08llx\n",
- bo_gem->gem_handle, bo_gem->name, bo->offset,
+ bo_gem->gem_handle, bo_gem->name, bo->offset64,
(unsigned long long)bufmgr_gem->exec_objects[i].
offset);
+ bo->offset64 = bufmgr_gem->exec_objects[i].offset;
bo->offset = bufmgr_gem->exec_objects[i].offset;
}
}
drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *)bo;
/* Update the buffer offset */
- if (bufmgr_gem->exec2_objects[i].offset != bo->offset) {
+ if (bufmgr_gem->exec2_objects[i].offset != bo->offset64) {
DBG("BO %d (%s) migrated: 0x%08lx -> 0x%08llx\n",
- bo_gem->gem_handle, bo_gem->name, bo->offset,
+ bo_gem->gem_handle, bo_gem->name, bo->offset64,
(unsigned long long)bufmgr_gem->exec2_objects[i].offset);
+ bo->offset64 = bufmgr_gem->exec2_objects[i].offset;
bo->offset = bufmgr_gem->exec2_objects[i].offset;
}
}
}
+static void
+aub_out(drm_intel_bufmgr_gem *bufmgr_gem, uint32_t data)
+{
+ fwrite(&data, 1, 4, bufmgr_gem->aub_file);
+}
+
+static void
+aub_out_data(drm_intel_bufmgr_gem *bufmgr_gem, void *data, size_t size)
+{
+ fwrite(data, 1, size, bufmgr_gem->aub_file);
+}
+
+static void
+aub_write_bo_data(drm_intel_bo *bo, uint32_t offset, uint32_t size)
+{
+ drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
+ drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
+ uint32_t *data;
+ unsigned int i;
+
+ data = malloc(bo->size);
+ drm_intel_bo_get_subdata(bo, offset, size, data);
+
+ /* Easy mode: write out bo with no relocations */
+ if (!bo_gem->reloc_count) {
+ aub_out_data(bufmgr_gem, data, size);
+ free(data);
+ return;
+ }
+
+ /* Otherwise, handle the relocations while writing. */
+ for (i = 0; i < size / 4; i++) {
+ int r;
+ for (r = 0; r < bo_gem->reloc_count; r++) {
+ struct drm_i915_gem_relocation_entry *reloc;
+ drm_intel_reloc_target *info;
+
+ reloc = &bo_gem->relocs[r];
+ info = &bo_gem->reloc_target_info[r];
+
+ if (reloc->offset == offset + i * 4) {
+ drm_intel_bo_gem *target_gem;
+ uint32_t val;
+
+ target_gem = (drm_intel_bo_gem *)info->bo;
+
+ val = reloc->delta;
+ val += target_gem->aub_offset;
+
+ aub_out(bufmgr_gem, val);
+ data[i] = val;
+ break;
+ }
+ }
+ if (r == bo_gem->reloc_count) {
+ /* no relocation, just the data */
+ aub_out(bufmgr_gem, data[i]);
+ }
+ }
+
+ free(data);
+}
+
+static void
+aub_bo_get_address(drm_intel_bo *bo)
+{
+ drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
+ drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
+
+ /* Give the object a graphics address in the AUB file. We
+ * don't just use the GEM object address because we do AUB
+ * dumping before execution -- we want to successfully log
+ * when the hardware might hang, and we might even want to aub
+ * capture for a driver trying to execute on a different
+ * generation of hardware by disabling the actual kernel exec
+ * call.
+ */
+ bo_gem->aub_offset = bufmgr_gem->aub_offset;
+ bufmgr_gem->aub_offset += bo->size;
+ /* XXX: Handle aperture overflow. */
+ assert(bufmgr_gem->aub_offset < 256 * 1024 * 1024);
+}
+
+static void
+aub_write_trace_block(drm_intel_bo *bo, uint32_t type, uint32_t subtype,
+ uint32_t offset, uint32_t size)
+{
+ drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
+ drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
+
+ aub_out(bufmgr_gem,
+ CMD_AUB_TRACE_HEADER_BLOCK |
+ ((bufmgr_gem->gen >= 8 ? 6 : 5) - 2));
+ aub_out(bufmgr_gem,
+ AUB_TRACE_MEMTYPE_GTT | type | AUB_TRACE_OP_DATA_WRITE);
+ aub_out(bufmgr_gem, subtype);
+ aub_out(bufmgr_gem, bo_gem->aub_offset + offset);
+ aub_out(bufmgr_gem, size);
+ if (bufmgr_gem->gen >= 8)
+ aub_out(bufmgr_gem, 0);
+ aub_write_bo_data(bo, offset, size);
+}
+
+/**
+ * Break up large objects into multiple writes. Otherwise a 128kb VBO
+ * would overflow the 16 bits of size field in the packet header and
+ * everything goes badly after that.
+ */
+static void
+aub_write_large_trace_block(drm_intel_bo *bo, uint32_t type, uint32_t subtype,
+ uint32_t offset, uint32_t size)
+{
+ uint32_t block_size;
+ uint32_t sub_offset;
+
+ for (sub_offset = 0; sub_offset < size; sub_offset += block_size) {
+ block_size = size - sub_offset;
+
+ if (block_size > 8 * 4096)
+ block_size = 8 * 4096;
+
+ aub_write_trace_block(bo, type, subtype, offset + sub_offset,
+ block_size);
+ }
+}
+
+static void
+aub_write_bo(drm_intel_bo *bo)
+{
+ drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
+ uint32_t offset = 0;
+ unsigned i;
+
+ aub_bo_get_address(bo);
+
+ /* Write out each annotated section separately. */
+ for (i = 0; i < bo_gem->aub_annotation_count; ++i) {
+ drm_intel_aub_annotation *annotation =
+ &bo_gem->aub_annotations[i];
+ uint32_t ending_offset = annotation->ending_offset;
+ if (ending_offset > bo->size)
+ ending_offset = bo->size;
+ if (ending_offset > offset) {
+ aub_write_large_trace_block(bo, annotation->type,
+ annotation->subtype,
+ offset,
+ ending_offset - offset);
+ offset = ending_offset;
+ }
+ }
+
+ /* Write out any remaining unannotated data */
+ if (offset < bo->size) {
+ aub_write_large_trace_block(bo, AUB_TRACE_TYPE_NOTYPE, 0,
+ offset, bo->size - offset);
+ }
+}
+
+/*
+ * Make a ringbuffer on fly and dump it
+ */
+static void
+aub_build_dump_ringbuffer(drm_intel_bufmgr_gem *bufmgr_gem,
+ uint32_t batch_buffer, int ring_flag)
+{
+ uint32_t ringbuffer[4096];
+ int ring = AUB_TRACE_TYPE_RING_PRB0; /* The default ring */
+ int ring_count = 0;
+
+ if (ring_flag == I915_EXEC_BSD)
+ ring = AUB_TRACE_TYPE_RING_PRB1;
+ else if (ring_flag == I915_EXEC_BLT)
+ ring = AUB_TRACE_TYPE_RING_PRB2;
+
+ /* Make a ring buffer to execute our batchbuffer. */
+ memset(ringbuffer, 0, sizeof(ringbuffer));
+ if (bufmgr_gem->gen >= 8) {
+ ringbuffer[ring_count++] = AUB_MI_BATCH_BUFFER_START | (3 - 2);
+ ringbuffer[ring_count++] = batch_buffer;
+ ringbuffer[ring_count++] = 0;
+ } else {
+ ringbuffer[ring_count++] = AUB_MI_BATCH_BUFFER_START;
+ ringbuffer[ring_count++] = batch_buffer;
+ }
+
+ /* Write out the ring. This appears to trigger execution of
+ * the ring in the simulator.
+ */
+ aub_out(bufmgr_gem,
+ CMD_AUB_TRACE_HEADER_BLOCK |
+ ((bufmgr_gem->gen >= 8 ? 6 : 5) - 2));
+ aub_out(bufmgr_gem,
+ AUB_TRACE_MEMTYPE_GTT | ring | AUB_TRACE_OP_COMMAND_WRITE);
+ aub_out(bufmgr_gem, 0); /* general/surface subtype */
+ aub_out(bufmgr_gem, bufmgr_gem->aub_offset);
+ aub_out(bufmgr_gem, ring_count * 4);
+ if (bufmgr_gem->gen >= 8)
+ aub_out(bufmgr_gem, 0);
+
+ /* FIXME: Need some flush operations here? */
+ aub_out_data(bufmgr_gem, ringbuffer, ring_count * 4);
+
+ /* Update offset pointer */
+ bufmgr_gem->aub_offset += 4096;
+}
+
+drm_public void
+drm_intel_gem_bo_aub_dump_bmp(drm_intel_bo *bo,
+ int x1, int y1, int width, int height,
+ enum aub_dump_bmp_format format,
+ int pitch, int offset)
+{
+ drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
+ drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *)bo;
+ uint32_t cpp;
+
+ switch (format) {
+ case AUB_DUMP_BMP_FORMAT_8BIT:
+ cpp = 1;
+ break;
+ case AUB_DUMP_BMP_FORMAT_ARGB_4444:
+ cpp = 2;
+ break;
+ case AUB_DUMP_BMP_FORMAT_ARGB_0888:
+ case AUB_DUMP_BMP_FORMAT_ARGB_8888:
+ cpp = 4;
+ break;
+ default:
+ printf("Unknown AUB dump format %d\n", format);
+ return;
+ }
+
+ if (!bufmgr_gem->aub_file)
+ return;
+
+ aub_out(bufmgr_gem, CMD_AUB_DUMP_BMP | 4);
+ aub_out(bufmgr_gem, (y1 << 16) | x1);
+ aub_out(bufmgr_gem,
+ (format << 24) |
+ (cpp << 19) |
+ pitch / 4);
+ aub_out(bufmgr_gem, (height << 16) | width);
+ aub_out(bufmgr_gem, bo_gem->aub_offset + offset);
+ aub_out(bufmgr_gem,
+ ((bo_gem->tiling_mode != I915_TILING_NONE) ? (1 << 2) : 0) |
+ ((bo_gem->tiling_mode == I915_TILING_Y) ? (1 << 3) : 0));
+}
+
+static void
+aub_exec(drm_intel_bo *bo, int ring_flag, int used)
+{
+ drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
+ drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
+ int i;
+ bool batch_buffer_needs_annotations;
+
+ if (!bufmgr_gem->aub_file)
+ return;
+
+ /* If batch buffer is not annotated, annotate it the best we
+ * can.
+ */
+ batch_buffer_needs_annotations = bo_gem->aub_annotation_count == 0;
+ if (batch_buffer_needs_annotations) {
+ drm_intel_aub_annotation annotations[2] = {
+ { AUB_TRACE_TYPE_BATCH, 0, used },
+ { AUB_TRACE_TYPE_NOTYPE, 0, bo->size }
+ };
+ drm_intel_bufmgr_gem_set_aub_annotations(bo, annotations, 2);
+ }
+
+ /* Write out all buffers to AUB memory */
+ for (i = 0; i < bufmgr_gem->exec_count; i++) {
+ aub_write_bo(bufmgr_gem->exec_bos[i]);
+ }
+
+ /* Remove any annotations we added */
+ if (batch_buffer_needs_annotations)
+ drm_intel_bufmgr_gem_set_aub_annotations(bo, NULL, 0);
+
+ /* Dump ring buffer */
+ aub_build_dump_ringbuffer(bufmgr_gem, bo_gem->aub_offset, ring_flag);
+
+ fflush(bufmgr_gem->aub_file);
+
+ /*
+ * One frame has been dumped. So reset the aub_offset for the next frame.
+ *
+ * FIXME: Can we do this?
+ */
+ bufmgr_gem->aub_offset = 0x10000;
+}
+
static int
drm_intel_gem_bo_exec(drm_intel_bo *bo, int used,
drm_clip_rect_t * cliprects, int num_cliprects, int DR4)
*/
drm_intel_add_validate_buffer(bo);
+ VG_CLEAR(execbuf);
execbuf.buffers_ptr = (uintptr_t) bufmgr_gem->exec_objects;
execbuf.buffer_count = bufmgr_gem->exec_count;
execbuf.batch_start_offset = 0;
drm_intel_bo *bo = bufmgr_gem->exec_bos[i];
drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
+ bo_gem->idle = false;
+
/* Disconnect the buffer from the validate list */
bo_gem->validate_index = -1;
bufmgr_gem->exec_bos[i] = NULL;
}
static int
-drm_intel_gem_bo_mrb_exec2(drm_intel_bo *bo, int used,
- drm_clip_rect_t *cliprects, int num_cliprects, int DR4,
- unsigned int flags)
+do_exec2(drm_intel_bo *bo, int used, drm_intel_context *ctx,
+ drm_clip_rect_t *cliprects, int num_cliprects, int DR4,
+ unsigned int flags)
{
drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bo->bufmgr;
struct drm_i915_gem_execbuffer2 execbuf;
- int ret, i;
+ int ret = 0;
+ int i;
switch (flags & 0x7) {
default:
if (!bufmgr_gem->has_bsd)
return -EINVAL;
break;
+ case I915_EXEC_VEBOX:
+ if (!bufmgr_gem->has_vebox)
+ return -EINVAL;
+ break;
case I915_EXEC_RENDER:
case I915_EXEC_DEFAULT:
break;
*/
drm_intel_add_validate_buffer2(bo, 0);
+ VG_CLEAR(execbuf);
execbuf.buffers_ptr = (uintptr_t)bufmgr_gem->exec2_objects;
execbuf.buffer_count = bufmgr_gem->exec_count;
execbuf.batch_start_offset = 0;
execbuf.DR1 = 0;
execbuf.DR4 = DR4;
execbuf.flags = flags;
- execbuf.rsvd1 = 0;
+ if (ctx == NULL)
+ i915_execbuffer2_set_context_id(execbuf, 0);
+ else
+ i915_execbuffer2_set_context_id(execbuf, ctx->ctx_id);
execbuf.rsvd2 = 0;
+ aub_exec(bo, flags, used);
+
+ if (bufmgr_gem->no_exec)
+ goto skip_execution;
+
ret = drmIoctl(bufmgr_gem->fd,
DRM_IOCTL_I915_GEM_EXECBUFFER2,
&execbuf);
}
drm_intel_update_buffer_offsets2(bufmgr_gem);
+skip_execution:
if (bufmgr_gem->bufmgr.debug)
drm_intel_gem_dump_validation_list(bufmgr_gem);
drm_intel_bo *bo = bufmgr_gem->exec_bos[i];
drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *)bo;
+ bo_gem->idle = false;
+
/* Disconnect the buffer from the validate list */
bo_gem->validate_index = -1;
bufmgr_gem->exec_bos[i] = NULL;
drm_clip_rect_t *cliprects, int num_cliprects,
int DR4)
{
- return drm_intel_gem_bo_mrb_exec2(bo, used,
- cliprects, num_cliprects, DR4,
- I915_EXEC_RENDER);
+ return do_exec2(bo, used, NULL, cliprects, num_cliprects, DR4,
+ I915_EXEC_RENDER);
+}
+
+static int
+drm_intel_gem_bo_mrb_exec2(drm_intel_bo *bo, int used,
+ drm_clip_rect_t *cliprects, int num_cliprects, int DR4,
+ unsigned int flags)
+{
+ return do_exec2(bo, used, NULL, cliprects, num_cliprects, DR4,
+ flags);
+}
+
+drm_public int
+drm_intel_gem_bo_context_exec(drm_intel_bo *bo, drm_intel_context *ctx,
+ int used, unsigned int flags)
+{
+ return do_exec2(bo, used, ctx, NULL, 0, 0, flags);
}
static int
struct drm_i915_gem_pin pin;
int ret;
- memset(&pin, 0, sizeof(pin));
+ VG_CLEAR(pin);
pin.handle = bo_gem->gem_handle;
pin.alignment = alignment;
if (ret != 0)
return -errno;
+ bo->offset64 = pin.offset;
bo->offset = pin.offset;
return 0;
}
struct drm_i915_gem_unpin unpin;
int ret;
- memset(&unpin, 0, sizeof(unpin));
+ VG_CLEAR(unpin);
unpin.handle = bo_gem->gem_handle;
ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_UNPIN, &unpin);
drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
int ret;
+ /* Tiling with userptr surfaces is not supported
+ * on all hardware so refuse it for time being.
+ */
+ if (bo_gem->is_userptr)
+ return -EINVAL;
+
/* Linear buffers have no stride. By ensuring that we only ever use
* stride 0 with linear buffers, we simplify our code.
*/
return 0;
}
+drm_public drm_intel_bo *
+drm_intel_bo_gem_create_from_prime(drm_intel_bufmgr *bufmgr, int prime_fd, int size)
+{
+ drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bufmgr;
+ int ret;
+ uint32_t handle;
+ drm_intel_bo_gem *bo_gem;
+ struct drm_i915_gem_get_tiling get_tiling;
+ drmMMListHead *list;
+
+ ret = drmPrimeFDToHandle(bufmgr_gem->fd, prime_fd, &handle);
+
+ /*
+ * See if the kernel has already returned this buffer to us. Just as
+ * for named buffers, we must not create two bo's pointing at the same
+ * kernel object
+ */
+ pthread_mutex_lock(&bufmgr_gem->lock);
+ for (list = bufmgr_gem->named.next;
+ list != &bufmgr_gem->named;
+ list = list->next) {
+ bo_gem = DRMLISTENTRY(drm_intel_bo_gem, list, name_list);
+ if (bo_gem->gem_handle == handle) {
+ drm_intel_gem_bo_reference(&bo_gem->bo);
+ pthread_mutex_unlock(&bufmgr_gem->lock);
+ return &bo_gem->bo;
+ }
+ }
+
+ if (ret) {
+ fprintf(stderr,"ret is %d %d\n", ret, errno);
+ pthread_mutex_unlock(&bufmgr_gem->lock);
+ return NULL;
+ }
+
+ bo_gem = calloc(1, sizeof(*bo_gem));
+ if (!bo_gem) {
+ pthread_mutex_unlock(&bufmgr_gem->lock);
+ return NULL;
+ }
+ /* Determine size of bo. The fd-to-handle ioctl really should
+ * return the size, but it doesn't. If we have kernel 3.12 or
+ * later, we can lseek on the prime fd to get the size. Older
+ * kernels will just fail, in which case we fall back to the
+ * provided (estimated or guess size). */
+ ret = lseek(prime_fd, 0, SEEK_END);
+ if (ret != -1)
+ bo_gem->bo.size = ret;
+ else
+ bo_gem->bo.size = size;
+
+ bo_gem->bo.handle = handle;
+ bo_gem->bo.bufmgr = bufmgr;
+
+ bo_gem->gem_handle = handle;
+
+ atomic_set(&bo_gem->refcount, 1);
+
+ bo_gem->name = "prime";
+ bo_gem->validate_index = -1;
+ bo_gem->reloc_tree_fences = 0;
+ bo_gem->used_as_reloc_target = false;
+ bo_gem->has_error = false;
+ bo_gem->reusable = false;
+
+ DRMINITLISTHEAD(&bo_gem->vma_list);
+ DRMLISTADDTAIL(&bo_gem->name_list, &bufmgr_gem->named);
+ pthread_mutex_unlock(&bufmgr_gem->lock);
+
+ VG_CLEAR(get_tiling);
+ get_tiling.handle = bo_gem->gem_handle;
+ ret = drmIoctl(bufmgr_gem->fd,
+ DRM_IOCTL_I915_GEM_GET_TILING,
+ &get_tiling);
+ if (ret != 0) {
+ drm_intel_gem_bo_unreference(&bo_gem->bo);
+ return NULL;
+ }
+ bo_gem->tiling_mode = get_tiling.tiling_mode;
+ bo_gem->swizzle_mode = get_tiling.swizzle_mode;
+ /* XXX stride is unknown */
+ drm_intel_bo_gem_set_in_aperture_size(bufmgr_gem, bo_gem);
+
+ return &bo_gem->bo;
+}
+
+drm_public int
+drm_intel_bo_gem_export_to_prime(drm_intel_bo *bo, int *prime_fd)
+{
+ drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
+ drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
+
+ pthread_mutex_lock(&bufmgr_gem->lock);
+ if (DRMLISTEMPTY(&bo_gem->name_list))
+ DRMLISTADDTAIL(&bo_gem->name_list, &bufmgr_gem->named);
+ pthread_mutex_unlock(&bufmgr_gem->lock);
+
+ if (drmPrimeHandleToFD(bufmgr_gem->fd, bo_gem->gem_handle,
+ DRM_CLOEXEC, prime_fd) != 0)
+ return -errno;
+
+ bo_gem->reusable = false;
+
+ return 0;
+}
+
static int
drm_intel_gem_bo_flink(drm_intel_bo *bo, uint32_t * name)
{
drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bo->bufmgr;
drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
- struct drm_gem_flink flink;
int ret;
if (!bo_gem->global_name) {
- memset(&flink, 0, sizeof(flink));
+ struct drm_gem_flink flink;
+
+ VG_CLEAR(flink);
flink.handle = bo_gem->gem_handle;
+ pthread_mutex_lock(&bufmgr_gem->lock);
+
ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_GEM_FLINK, &flink);
- if (ret != 0)
+ if (ret != 0) {
+ pthread_mutex_unlock(&bufmgr_gem->lock);
return -errno;
+ }
+
bo_gem->global_name = flink.name;
bo_gem->reusable = false;
- DRMLISTADDTAIL(&bo_gem->name_list, &bufmgr_gem->named);
+ if (DRMLISTEMPTY(&bo_gem->name_list))
+ DRMLISTADDTAIL(&bo_gem->name_list, &bufmgr_gem->named);
+ pthread_mutex_unlock(&bufmgr_gem->lock);
}
*name = bo_gem->global_name;
* size is only bounded by how many buffers of that size we've managed to have
* in flight at once.
*/
-void
+drm_public void
drm_intel_bufmgr_gem_enable_reuse(drm_intel_bufmgr *bufmgr)
{
drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *) bufmgr;
* allocation. If this option is not enabled, all relocs will have fence
* register allocated.
*/
-void
+drm_public void
drm_intel_bufmgr_gem_enable_fenced_relocs(drm_intel_bufmgr *bufmgr)
{
drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr;
}
}
+drm_public void
+drm_intel_bufmgr_gem_set_vma_cache_size(drm_intel_bufmgr *bufmgr, int limit)
+{
+ drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr;
+
+ bufmgr_gem->vma_max = limit;
+
+ drm_intel_gem_bo_purge_vma_cache(bufmgr_gem);
+}
+
+/**
+ * Get the PCI ID for the device. This can be overridden by setting the
+ * INTEL_DEVID_OVERRIDE environment variable to the desired ID.
+ */
+static int
+get_pci_device_id(drm_intel_bufmgr_gem *bufmgr_gem)
+{
+ char *devid_override;
+ int devid;
+ int ret;
+ drm_i915_getparam_t gp;
+
+ if (geteuid() == getuid()) {
+ devid_override = getenv("INTEL_DEVID_OVERRIDE");
+ if (devid_override) {
+ bufmgr_gem->no_exec = true;
+ return strtod(devid_override, NULL);
+ }
+ }
+
+ VG_CLEAR(devid);
+ VG_CLEAR(gp);
+ gp.param = I915_PARAM_CHIPSET_ID;
+ gp.value = &devid;
+ ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp);
+ if (ret) {
+ fprintf(stderr, "get chip id failed: %d [%d]\n", ret, errno);
+ fprintf(stderr, "param: %d, val: %d\n", gp.param, *gp.value);
+ }
+ return devid;
+}
+
+drm_public int
+drm_intel_bufmgr_gem_get_devid(drm_intel_bufmgr *bufmgr)
+{
+ drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr;
+
+ return bufmgr_gem->pci_device;
+}
+
+/**
+ * Sets the AUB filename.
+ *
+ * This function has to be called before drm_intel_bufmgr_gem_set_aub_dump()
+ * for it to have any effect.
+ */
+drm_public void
+drm_intel_bufmgr_gem_set_aub_filename(drm_intel_bufmgr *bufmgr,
+ const char *filename)
+{
+ drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr;
+
+ free(bufmgr_gem->aub_filename);
+ if (filename)
+ bufmgr_gem->aub_filename = strdup(filename);
+}
+
+/**
+ * Sets up AUB dumping.
+ *
+ * This is a trace file format that can be used with the simulator.
+ * Packets are emitted in a format somewhat like GPU command packets.
+ * You can set up a GTT and upload your objects into the referenced
+ * space, then send off batchbuffers and get BMPs out the other end.
+ */
+drm_public void
+drm_intel_bufmgr_gem_set_aub_dump(drm_intel_bufmgr *bufmgr, int enable)
+{
+ drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr;
+ int entry = 0x200003;
+ int i;
+ int gtt_size = 0x10000;
+ const char *filename;
+
+ if (!enable) {
+ if (bufmgr_gem->aub_file) {
+ fclose(bufmgr_gem->aub_file);
+ bufmgr_gem->aub_file = NULL;
+ }
+ return;
+ }
+
+ if (geteuid() != getuid())
+ return;
+
+ if (bufmgr_gem->aub_filename)
+ filename = bufmgr_gem->aub_filename;
+ else
+ filename = "intel.aub";
+ bufmgr_gem->aub_file = fopen(filename, "w+");
+ if (!bufmgr_gem->aub_file)
+ return;
+
+ /* Start allocating objects from just after the GTT. */
+ bufmgr_gem->aub_offset = gtt_size;
+
+ /* Start with a (required) version packet. */
+ aub_out(bufmgr_gem, CMD_AUB_HEADER | (13 - 2));
+ aub_out(bufmgr_gem,
+ (4 << AUB_HEADER_MAJOR_SHIFT) |
+ (0 << AUB_HEADER_MINOR_SHIFT));
+ for (i = 0; i < 8; i++) {
+ aub_out(bufmgr_gem, 0); /* app name */
+ }
+ aub_out(bufmgr_gem, 0); /* timestamp */
+ aub_out(bufmgr_gem, 0); /* timestamp */
+ aub_out(bufmgr_gem, 0); /* comment len */
+
+ /* Set up the GTT. The max we can handle is 256M */
+ aub_out(bufmgr_gem, CMD_AUB_TRACE_HEADER_BLOCK | ((bufmgr_gem->gen >= 8 ? 6 : 5) - 2));
+ aub_out(bufmgr_gem, AUB_TRACE_MEMTYPE_NONLOCAL | 0 | AUB_TRACE_OP_DATA_WRITE);
+ aub_out(bufmgr_gem, 0); /* subtype */
+ aub_out(bufmgr_gem, 0); /* offset */
+ aub_out(bufmgr_gem, gtt_size); /* size */
+ if (bufmgr_gem->gen >= 8)
+ aub_out(bufmgr_gem, 0);
+ for (i = 0x000; i < gtt_size; i += 4, entry += 0x1000) {
+ aub_out(bufmgr_gem, entry);
+ }
+}
+
+drm_public drm_intel_context *
+drm_intel_gem_context_create(drm_intel_bufmgr *bufmgr)
+{
+ drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr;
+ struct drm_i915_gem_context_create create;
+ drm_intel_context *context = NULL;
+ int ret;
+
+ context = calloc(1, sizeof(*context));
+ if (!context)
+ return NULL;
+
+ VG_CLEAR(create);
+ ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_CONTEXT_CREATE, &create);
+ if (ret != 0) {
+ DBG("DRM_IOCTL_I915_GEM_CONTEXT_CREATE failed: %s\n",
+ strerror(errno));
+ free(context);
+ return NULL;
+ }
+
+ context->ctx_id = create.ctx_id;
+ context->bufmgr = bufmgr;
+
+ return context;
+}
+
+drm_public void
+drm_intel_gem_context_destroy(drm_intel_context *ctx)
+{
+ drm_intel_bufmgr_gem *bufmgr_gem;
+ struct drm_i915_gem_context_destroy destroy;
+ int ret;
+
+ if (ctx == NULL)
+ return;
+
+ VG_CLEAR(destroy);
+
+ bufmgr_gem = (drm_intel_bufmgr_gem *)ctx->bufmgr;
+ destroy.ctx_id = ctx->ctx_id;
+ ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_CONTEXT_DESTROY,
+ &destroy);
+ if (ret != 0)
+ fprintf(stderr, "DRM_IOCTL_I915_GEM_CONTEXT_DESTROY failed: %s\n",
+ strerror(errno));
+
+ free(ctx);
+}
+
+drm_public int
+drm_intel_get_reset_stats(drm_intel_context *ctx,
+ uint32_t *reset_count,
+ uint32_t *active,
+ uint32_t *pending)
+{
+ drm_intel_bufmgr_gem *bufmgr_gem;
+ struct drm_i915_reset_stats stats;
+ int ret;
+
+ if (ctx == NULL)
+ return -EINVAL;
+
+ memset(&stats, 0, sizeof(stats));
+
+ bufmgr_gem = (drm_intel_bufmgr_gem *)ctx->bufmgr;
+ stats.ctx_id = ctx->ctx_id;
+ ret = drmIoctl(bufmgr_gem->fd,
+ DRM_IOCTL_I915_GET_RESET_STATS,
+ &stats);
+ if (ret == 0) {
+ if (reset_count != NULL)
+ *reset_count = stats.reset_count;
+
+ if (active != NULL)
+ *active = stats.batch_active;
+
+ if (pending != NULL)
+ *pending = stats.batch_pending;
+ }
+
+ return ret;
+}
+
+drm_public int
+drm_intel_reg_read(drm_intel_bufmgr *bufmgr,
+ uint32_t offset,
+ uint64_t *result)
+{
+ drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr;
+ struct drm_i915_reg_read reg_read;
+ int ret;
+
+ VG_CLEAR(reg_read);
+ reg_read.offset = offset;
+
+ ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_REG_READ, ®_read);
+
+ *result = reg_read.val;
+ return ret;
+}
+
+
+/**
+ * Annotate the given bo for use in aub dumping.
+ *
+ * \param annotations is an array of drm_intel_aub_annotation objects
+ * describing the type of data in various sections of the bo. Each
+ * element of the array specifies the type and subtype of a section of
+ * the bo, and the past-the-end offset of that section. The elements
+ * of \c annotations must be sorted so that ending_offset is
+ * increasing.
+ *
+ * \param count is the number of elements in the \c annotations array.
+ * If \c count is zero, then \c annotations will not be dereferenced.
+ *
+ * Annotations are copied into a private data structure, so caller may
+ * re-use the memory pointed to by \c annotations after the call
+ * returns.
+ *
+ * Annotations are stored for the lifetime of the bo; to reset to the
+ * default state (no annotations), call this function with a \c count
+ * of zero.
+ */
+drm_public void
+drm_intel_bufmgr_gem_set_aub_annotations(drm_intel_bo *bo,
+ drm_intel_aub_annotation *annotations,
+ unsigned count)
+{
+ drm_intel_bo_gem *bo_gem = (drm_intel_bo_gem *) bo;
+ unsigned size = sizeof(*annotations) * count;
+ drm_intel_aub_annotation *new_annotations =
+ count > 0 ? realloc(bo_gem->aub_annotations, size) : NULL;
+ if (new_annotations == NULL) {
+ free(bo_gem->aub_annotations);
+ bo_gem->aub_annotations = NULL;
+ bo_gem->aub_annotation_count = 0;
+ return;
+ }
+ memcpy(new_annotations, annotations, size);
+ bo_gem->aub_annotations = new_annotations;
+ bo_gem->aub_annotation_count = count;
+}
+
+static pthread_mutex_t bufmgr_list_mutex = PTHREAD_MUTEX_INITIALIZER;
+static drmMMListHead bufmgr_list = { &bufmgr_list, &bufmgr_list };
+
+static drm_intel_bufmgr_gem *
+drm_intel_bufmgr_gem_find(int fd)
+{
+ drm_intel_bufmgr_gem *bufmgr_gem;
+
+ DRMLISTFOREACHENTRY(bufmgr_gem, &bufmgr_list, managers) {
+ if (bufmgr_gem->fd == fd) {
+ atomic_inc(&bufmgr_gem->refcount);
+ return bufmgr_gem;
+ }
+ }
+
+ return NULL;
+}
+
+static void
+drm_intel_bufmgr_gem_unref(drm_intel_bufmgr *bufmgr)
+{
+ drm_intel_bufmgr_gem *bufmgr_gem = (drm_intel_bufmgr_gem *)bufmgr;
+
+ if (atomic_add_unless(&bufmgr_gem->refcount, -1, 1)) {
+ pthread_mutex_lock(&bufmgr_list_mutex);
+
+ if (atomic_dec_and_test(&bufmgr_gem->refcount)) {
+ DRMLISTDEL(&bufmgr_gem->managers);
+ drm_intel_bufmgr_gem_destroy(bufmgr);
+ }
+
+ pthread_mutex_unlock(&bufmgr_list_mutex);
+ }
+}
+
+static bool
+has_userptr(drm_intel_bufmgr_gem *bufmgr_gem)
+{
+ int ret;
+ void *ptr;
+ long pgsz;
+ struct drm_i915_gem_userptr userptr;
+ struct drm_gem_close close_bo;
+
+ pgsz = sysconf(_SC_PAGESIZE);
+ assert(pgsz > 0);
+
+ ret = posix_memalign(&ptr, pgsz, pgsz);
+ if (ret) {
+ DBG("Failed to get a page (%ld) for userptr detection!\n",
+ pgsz);
+ return false;
+ }
+
+ memset(&userptr, 0, sizeof(userptr));
+ userptr.user_ptr = (__u64)(unsigned long)ptr;
+ userptr.user_size = pgsz;
+
+retry:
+ ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_USERPTR, &userptr);
+ if (ret) {
+ if (errno == ENODEV && userptr.flags == 0) {
+ userptr.flags = I915_USERPTR_UNSYNCHRONIZED;
+ goto retry;
+ }
+ free(ptr);
+ return false;
+ }
+
+ close_bo.handle = userptr.handle;
+ ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_GEM_CLOSE, &close_bo);
+ free(ptr);
+ if (ret) {
+ fprintf(stderr, "Failed to release test userptr object! (%d) "
+ "i915 kernel driver may not be sane!\n", errno);
+ return false;
+ }
+
+ return true;
+}
+
/**
* 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.
*/
-drm_intel_bufmgr *
+drm_public drm_intel_bufmgr *
drm_intel_bufmgr_gem_init(int fd, int batch_size)
{
drm_intel_bufmgr_gem *bufmgr_gem;
int ret, tmp;
bool exec2 = false;
+ pthread_mutex_lock(&bufmgr_list_mutex);
+
+ bufmgr_gem = drm_intel_bufmgr_gem_find(fd);
+ if (bufmgr_gem)
+ goto exit;
+
bufmgr_gem = calloc(1, sizeof(*bufmgr_gem));
if (bufmgr_gem == NULL)
- return NULL;
+ goto exit;
bufmgr_gem->fd = fd;
+ atomic_set(&bufmgr_gem->refcount, 1);
if (pthread_mutex_init(&bufmgr_gem->lock, NULL) != 0) {
free(bufmgr_gem);
- return NULL;
+ bufmgr_gem = NULL;
+ goto exit;
}
ret = drmIoctl(bufmgr_gem->fd,
(int)bufmgr_gem->gtt_size / 1024);
}
- gp.param = I915_PARAM_CHIPSET_ID;
- gp.value = &bufmgr_gem->pci_device;
- ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp);
- if (ret) {
- fprintf(stderr, "get chip id failed: %d [%d]\n", ret, errno);
- fprintf(stderr, "param: %d, val: %d\n", gp.param, *gp.value);
- }
+ bufmgr_gem->pci_device = get_pci_device_id(bufmgr_gem);
- if (IS_GEN2(bufmgr_gem))
+ if (IS_GEN2(bufmgr_gem->pci_device))
bufmgr_gem->gen = 2;
- else if (IS_GEN3(bufmgr_gem))
+ else if (IS_GEN3(bufmgr_gem->pci_device))
bufmgr_gem->gen = 3;
- else if (IS_GEN4(bufmgr_gem))
+ else if (IS_GEN4(bufmgr_gem->pci_device))
bufmgr_gem->gen = 4;
- else
+ else if (IS_GEN5(bufmgr_gem->pci_device))
+ bufmgr_gem->gen = 5;
+ else if (IS_GEN6(bufmgr_gem->pci_device))
bufmgr_gem->gen = 6;
+ else if (IS_GEN7(bufmgr_gem->pci_device))
+ bufmgr_gem->gen = 7;
+ else if (IS_GEN8(bufmgr_gem->pci_device))
+ bufmgr_gem->gen = 8;
+ else {
+ free(bufmgr_gem);
+ bufmgr_gem = NULL;
+ goto exit;
+ }
+ if (IS_GEN3(bufmgr_gem->pci_device) &&
+ bufmgr_gem->gtt_size > 256*1024*1024) {
+ /* The unmappable part of gtt on gen 3 (i.e. above 256MB) can't
+ * be used for tiled blits. To simplify the accounting, just
+ * substract the unmappable part (fixed to 256MB on all known
+ * gen3 devices) if the kernel advertises it. */
+ bufmgr_gem->gtt_size -= 256*1024*1024;
+ }
+
+ VG_CLEAR(gp);
gp.value = &tmp;
gp.param = I915_PARAM_HAS_EXECBUF2;
ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp);
bufmgr_gem->has_relaxed_fencing = ret == 0;
+ if (has_userptr(bufmgr_gem))
+ bufmgr_gem->bufmgr.bo_alloc_userptr =
+ drm_intel_gem_bo_alloc_userptr;
+
+ gp.param = I915_PARAM_HAS_WAIT_TIMEOUT;
+ ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp);
+ bufmgr_gem->has_wait_timeout = ret == 0;
+
+ gp.param = I915_PARAM_HAS_LLC;
+ ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp);
+ if (ret != 0) {
+ /* Kernel does not supports HAS_LLC query, fallback to GPU
+ * generation detection and assume that we have LLC on GEN6/7
+ */
+ bufmgr_gem->has_llc = (IS_GEN6(bufmgr_gem->pci_device) |
+ IS_GEN7(bufmgr_gem->pci_device));
+ } else
+ bufmgr_gem->has_llc = *gp.value;
+
+ gp.param = I915_PARAM_HAS_VEBOX;
+ ret = drmIoctl(bufmgr_gem->fd, DRM_IOCTL_I915_GETPARAM, &gp);
+ bufmgr_gem->has_vebox = (ret == 0) & (*gp.value > 0);
+
if (bufmgr_gem->gen < 4) {
gp.param = I915_PARAM_NUM_FENCES_AVAIL;
gp.value = &bufmgr_gem->available_fences;
bufmgr_gem->bufmgr.bo_exec = drm_intel_gem_bo_exec;
bufmgr_gem->bufmgr.bo_busy = drm_intel_gem_bo_busy;
bufmgr_gem->bufmgr.bo_madvise = drm_intel_gem_bo_madvise;
- bufmgr_gem->bufmgr.destroy = drm_intel_bufmgr_gem_destroy;
+ bufmgr_gem->bufmgr.destroy = drm_intel_bufmgr_gem_unref;
bufmgr_gem->bufmgr.debug = 0;
bufmgr_gem->bufmgr.check_aperture_space =
drm_intel_gem_check_aperture_space;
DRMINITLISTHEAD(&bufmgr_gem->named);
init_cache_buckets(bufmgr_gem);
- return &bufmgr_gem->bufmgr;
+ DRMINITLISTHEAD(&bufmgr_gem->vma_cache);
+ bufmgr_gem->vma_max = -1; /* unlimited by default */
+
+ DRMLISTADD(&bufmgr_gem->managers, &bufmgr_list);
+
+exit:
+ pthread_mutex_unlock(&bufmgr_list_mutex);
+
+ return bufmgr_gem != NULL ? &bufmgr_gem->bufmgr : NULL;
}
projects / platform / upstream / libdrm.git / blobdiff