#include "amdgpu_cs.h"
+#include "util/hash_table.h"
#include "util/os_time.h"
-#include "state_tracker/drm_driver.h"
-#include <amdgpu_drm.h>
+#include "util/u_hash_table.h"
+#include "frontend/drm_driver.h"
+#include "drm-uapi/amdgpu_drm.h"
#include <xf86drm.h>
#include <stdio.h>
#include <inttypes.h>
-#ifndef AMDGPU_GEM_CREATE_VM_ALWAYS_VALID
-#define AMDGPU_GEM_CREATE_VM_ALWAYS_VALID (1 << 6)
+#ifndef AMDGPU_VA_RANGE_HIGH
+#define AMDGPU_VA_RANGE_HIGH 0x2
#endif
/* Set to 1 for verbose output showing committed sparse buffer ranges. */
uint32_t begin, end;
};
-static struct pb_buffer *
-amdgpu_bo_create(struct radeon_winsys *rws,
- uint64_t size,
- unsigned alignment,
- enum radeon_bo_domain domain,
- enum radeon_bo_flag flags);
-
static bool amdgpu_bo_wait(struct pb_buffer *_buf, uint64_t timeout,
enum radeon_bo_usage usage)
{
return ((struct amdgpu_winsys_bo*)buf)->initial_domain;
}
+static enum radeon_bo_flag amdgpu_bo_get_flags(
+ struct pb_buffer *buf)
+{
+ return ((struct amdgpu_winsys_bo*)buf)->flags;
+}
+
static void amdgpu_bo_remove_fences(struct amdgpu_winsys_bo *bo)
{
for (unsigned i = 0; i < bo->num_fences; ++i)
void amdgpu_bo_destroy(struct pb_buffer *_buf)
{
struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(_buf);
+ struct amdgpu_screen_winsys *sws_iter;
+ struct amdgpu_winsys *ws = bo->ws;
assert(bo->bo && "must not be called for slab entries");
- if (bo->ws->debug_all_bos) {
- simple_mtx_lock(&bo->ws->global_bo_list_lock);
- LIST_DEL(&bo->u.real.global_list_item);
- bo->ws->num_buffers--;
- simple_mtx_unlock(&bo->ws->global_bo_list_lock);
+ if (!bo->is_user_ptr && bo->cpu_ptr) {
+ bo->cpu_ptr = NULL;
+ amdgpu_bo_unmap(&bo->base);
+ }
+ assert(bo->is_user_ptr || bo->u.real.map_count == 0);
+
+ if (ws->debug_all_bos) {
+ simple_mtx_lock(&ws->global_bo_list_lock);
+ list_del(&bo->u.real.global_list_item);
+ ws->num_buffers--;
+ simple_mtx_unlock(&ws->global_bo_list_lock);
+ }
+
+ /* Close all KMS handles retrieved for other DRM file descriptions */
+ simple_mtx_lock(&ws->sws_list_lock);
+ for (sws_iter = ws->sws_list; sws_iter; sws_iter = sws_iter->next) {
+ struct hash_entry *entry;
+
+ if (!sws_iter->kms_handles)
+ continue;
+
+ entry = _mesa_hash_table_search(sws_iter->kms_handles, bo);
+ if (entry) {
+ struct drm_gem_close args = { .handle = (uintptr_t)entry->data };
+
+ drmIoctl(sws_iter->fd, DRM_IOCTL_GEM_CLOSE, &args);
+ _mesa_hash_table_remove(sws_iter->kms_handles, entry);
+ }
}
+ simple_mtx_unlock(&ws->sws_list_lock);
+
+ simple_mtx_lock(&ws->bo_export_table_lock);
+ _mesa_hash_table_remove_key(ws->bo_export_table, bo->bo);
+ simple_mtx_unlock(&ws->bo_export_table_lock);
- amdgpu_bo_va_op(bo->bo, 0, bo->base.size, bo->va, 0, AMDGPU_VA_OP_UNMAP);
- amdgpu_va_range_free(bo->u.real.va_handle);
+ if (bo->initial_domain & RADEON_DOMAIN_VRAM_GTT) {
+ amdgpu_bo_va_op(bo->bo, 0, bo->base.size, bo->va, 0, AMDGPU_VA_OP_UNMAP);
+ amdgpu_va_range_free(bo->u.real.va_handle);
+ }
amdgpu_bo_free(bo->bo);
amdgpu_bo_remove_fences(bo);
if (bo->initial_domain & RADEON_DOMAIN_VRAM)
- bo->ws->allocated_vram -= align64(bo->base.size, bo->ws->info.gart_page_size);
+ ws->allocated_vram -= align64(bo->base.size, ws->info.gart_page_size);
else if (bo->initial_domain & RADEON_DOMAIN_GTT)
- bo->ws->allocated_gtt -= align64(bo->base.size, bo->ws->info.gart_page_size);
-
- if (bo->u.real.map_count >= 1) {
- if (bo->initial_domain & RADEON_DOMAIN_VRAM)
- bo->ws->mapped_vram -= bo->base.size;
- else if (bo->initial_domain & RADEON_DOMAIN_GTT)
- bo->ws->mapped_gtt -= bo->base.size;
- bo->ws->num_mapped_buffers--;
- }
+ ws->allocated_gtt -= align64(bo->base.size, ws->info.gart_page_size);
+ simple_mtx_destroy(&bo->lock);
FREE(bo);
}
amdgpu_bo_destroy(_buf);
}
-static void *amdgpu_bo_map(struct pb_buffer *buf,
- struct radeon_winsys_cs *rcs,
- enum pipe_transfer_usage usage)
+static void amdgpu_clean_up_buffer_managers(struct amdgpu_winsys *ws)
+{
+ for (unsigned i = 0; i < NUM_SLAB_ALLOCATORS; i++) {
+ pb_slabs_reclaim(&ws->bo_slabs[i]);
+ if (ws->secure)
+ pb_slabs_reclaim(&ws->bo_slabs_encrypted[i]);
+ }
+
+ pb_cache_release_all_buffers(&ws->bo_cache);
+}
+
+static bool amdgpu_bo_do_map(struct amdgpu_winsys_bo *bo, void **cpu)
+{
+ assert(!bo->sparse && bo->bo && !bo->is_user_ptr);
+ int r = amdgpu_bo_cpu_map(bo->bo, cpu);
+ if (r) {
+ /* Clean up buffer managers and try again. */
+ amdgpu_clean_up_buffer_managers(bo->ws);
+ r = amdgpu_bo_cpu_map(bo->bo, cpu);
+ if (r)
+ return false;
+ }
+
+ if (p_atomic_inc_return(&bo->u.real.map_count) == 1) {
+ if (bo->initial_domain & RADEON_DOMAIN_VRAM)
+ bo->ws->mapped_vram += bo->base.size;
+ else if (bo->initial_domain & RADEON_DOMAIN_GTT)
+ bo->ws->mapped_gtt += bo->base.size;
+ bo->ws->num_mapped_buffers++;
+ }
+
+ return true;
+}
+
+void *amdgpu_bo_map(struct pb_buffer *buf,
+ struct radeon_cmdbuf *rcs,
+ enum pipe_transfer_usage usage)
{
struct amdgpu_winsys_bo *bo = (struct amdgpu_winsys_bo*)buf;
struct amdgpu_winsys_bo *real;
struct amdgpu_cs *cs = (struct amdgpu_cs*)rcs;
- int r;
- void *cpu = NULL;
- uint64_t offset = 0;
assert(!bo->sparse);
* Only check whether the buffer is being used for write. */
if (cs && amdgpu_bo_is_referenced_by_cs_with_usage(cs, bo,
RADEON_USAGE_WRITE)) {
- cs->flush_cs(cs->flush_data, RADEON_FLUSH_ASYNC, NULL);
+ cs->flush_cs(cs->flush_data,
+ RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
return NULL;
}
}
} else {
if (cs && amdgpu_bo_is_referenced_by_cs(cs, bo)) {
- cs->flush_cs(cs->flush_data, RADEON_FLUSH_ASYNC, NULL);
+ cs->flush_cs(cs->flush_data,
+ RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
return NULL;
}
if (cs) {
if (amdgpu_bo_is_referenced_by_cs_with_usage(cs, bo,
RADEON_USAGE_WRITE)) {
- cs->flush_cs(cs->flush_data, 0, NULL);
+ cs->flush_cs(cs->flush_data,
+ RADEON_FLUSH_START_NEXT_GFX_IB_NOW, NULL);
} else {
/* Try to avoid busy-waiting in amdgpu_bo_wait. */
if (p_atomic_read(&bo->num_active_ioctls))
/* Mapping for write. */
if (cs) {
if (amdgpu_bo_is_referenced_by_cs(cs, bo)) {
- cs->flush_cs(cs->flush_data, 0, NULL);
+ cs->flush_cs(cs->flush_data,
+ RADEON_FLUSH_START_NEXT_GFX_IB_NOW, NULL);
} else {
/* Try to avoid busy-waiting in amdgpu_bo_wait. */
if (p_atomic_read(&bo->num_active_ioctls))
}
}
- /* If the buffer is created from user memory, return the user pointer. */
- if (bo->user_ptr)
- return bo->user_ptr;
+ /* Buffer synchronization has been checked, now actually map the buffer. */
+ void *cpu = NULL;
+ uint64_t offset = 0;
if (bo->bo) {
real = bo;
offset = bo->va - real->va;
}
- r = amdgpu_bo_cpu_map(real->bo, &cpu);
- if (r) {
- /* Clear the cache and try again. */
- pb_cache_release_all_buffers(&real->ws->bo_cache);
- r = amdgpu_bo_cpu_map(real->bo, &cpu);
- if (r)
- return NULL;
+ if (usage & RADEON_TRANSFER_TEMPORARY) {
+ if (real->is_user_ptr) {
+ cpu = real->cpu_ptr;
+ } else {
+ if (!amdgpu_bo_do_map(real, &cpu))
+ return NULL;
+ }
+ } else {
+ cpu = p_atomic_read(&real->cpu_ptr);
+ if (!cpu) {
+ simple_mtx_lock(&real->lock);
+ /* Must re-check due to the possibility of a race. Re-check need not
+ * be atomic thanks to the lock. */
+ cpu = real->cpu_ptr;
+ if (!cpu) {
+ if (!amdgpu_bo_do_map(real, &cpu)) {
+ simple_mtx_unlock(&real->lock);
+ return NULL;
+ }
+ p_atomic_set(&real->cpu_ptr, cpu);
+ }
+ simple_mtx_unlock(&real->lock);
+ }
}
- if (p_atomic_inc_return(&real->u.real.map_count) == 1) {
- if (real->initial_domain & RADEON_DOMAIN_VRAM)
- real->ws->mapped_vram += real->base.size;
- else if (real->initial_domain & RADEON_DOMAIN_GTT)
- real->ws->mapped_gtt += real->base.size;
- real->ws->num_mapped_buffers++;
- }
return (uint8_t*)cpu + offset;
}
-static void amdgpu_bo_unmap(struct pb_buffer *buf)
+void amdgpu_bo_unmap(struct pb_buffer *buf)
{
struct amdgpu_winsys_bo *bo = (struct amdgpu_winsys_bo*)buf;
struct amdgpu_winsys_bo *real;
assert(!bo->sparse);
- if (bo->user_ptr)
+ if (bo->is_user_ptr)
return;
real = bo->bo ? bo : bo->u.slab.real;
-
+ assert(real->u.real.map_count != 0 && "too many unmaps");
if (p_atomic_dec_zero(&real->u.real.map_count)) {
+ assert(!real->cpu_ptr &&
+ "too many unmaps or forgot RADEON_TRANSFER_TEMPORARY flag");
+
if (real->initial_domain & RADEON_DOMAIN_VRAM)
real->ws->mapped_vram -= real->base.size;
else if (real->initial_domain & RADEON_DOMAIN_GTT)
if (ws->debug_all_bos) {
simple_mtx_lock(&ws->global_bo_list_lock);
- LIST_ADDTAIL(&bo->u.real.global_list_item, &ws->global_bo_list);
+ list_addtail(&bo->u.real.global_list_item, &ws->global_bo_list);
ws->num_buffers++;
simple_mtx_unlock(&ws->global_bo_list_lock);
}
}
+static uint64_t amdgpu_get_optimal_vm_alignment(struct amdgpu_winsys *ws,
+ uint64_t size, unsigned alignment)
+{
+ uint64_t vm_alignment = alignment;
+
+ /* Increase the VM alignment for faster address translation. */
+ if (size >= ws->info.pte_fragment_size)
+ vm_alignment = MAX2(vm_alignment, ws->info.pte_fragment_size);
+
+ /* Gfx9: Increase the VM alignment to the most significant bit set
+ * in the size for faster address translation.
+ */
+ if (ws->info.chip_class >= GFX9) {
+ unsigned msb = util_last_bit64(size); /* 0 = no bit is set */
+ uint64_t msb_alignment = msb ? 1ull << (msb - 1) : 0;
+
+ vm_alignment = MAX2(vm_alignment, msb_alignment);
+ }
+ return vm_alignment;
+}
+
static struct amdgpu_winsys_bo *amdgpu_create_bo(struct amdgpu_winsys *ws,
uint64_t size,
unsigned alignment,
- unsigned usage,
enum radeon_bo_domain initial_domain,
unsigned flags,
- unsigned pb_cache_bucket)
+ int heap)
{
struct amdgpu_bo_alloc_request request = {0};
amdgpu_bo_handle buf_handle;
uint64_t va = 0;
struct amdgpu_winsys_bo *bo;
- amdgpu_va_handle va_handle;
- unsigned va_gap_size;
+ amdgpu_va_handle va_handle = NULL;
int r;
- assert(initial_domain & RADEON_DOMAIN_VRAM_GTT);
+ /* VRAM or GTT must be specified, but not both at the same time. */
+ assert(util_bitcount(initial_domain & (RADEON_DOMAIN_VRAM_GTT |
+ RADEON_DOMAIN_GDS |
+ RADEON_DOMAIN_OA)) == 1);
+
bo = CALLOC_STRUCT(amdgpu_winsys_bo);
if (!bo) {
return NULL;
}
- pb_cache_init_entry(&ws->bo_cache, &bo->u.real.cache_entry, &bo->base,
- pb_cache_bucket);
+ if (heap >= 0) {
+ pb_cache_init_entry(&ws->bo_cache, &bo->u.real.cache_entry, &bo->base,
+ heap);
+ }
request.alloc_size = size;
request.phys_alignment = alignment;
- if (initial_domain & RADEON_DOMAIN_VRAM)
+ if (initial_domain & RADEON_DOMAIN_VRAM) {
request.preferred_heap |= AMDGPU_GEM_DOMAIN_VRAM;
+
+ /* Since VRAM and GTT have almost the same performance on APUs, we could
+ * just set GTT. However, in order to decrease GTT(RAM) usage, which is
+ * shared with the OS, allow VRAM placements too. The idea is not to use
+ * VRAM usefully, but to use it so that it's not unused and wasted.
+ */
+ if (!ws->info.has_dedicated_vram)
+ request.preferred_heap |= AMDGPU_GEM_DOMAIN_GTT;
+ }
+
if (initial_domain & RADEON_DOMAIN_GTT)
request.preferred_heap |= AMDGPU_GEM_DOMAIN_GTT;
+ if (initial_domain & RADEON_DOMAIN_GDS)
+ request.preferred_heap |= AMDGPU_GEM_DOMAIN_GDS;
+ if (initial_domain & RADEON_DOMAIN_OA)
+ request.preferred_heap |= AMDGPU_GEM_DOMAIN_OA;
if (flags & RADEON_FLAG_NO_CPU_ACCESS)
request.flags |= AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
if (flags & RADEON_FLAG_GTT_WC)
request.flags |= AMDGPU_GEM_CREATE_CPU_GTT_USWC;
- if (flags & RADEON_FLAG_NO_INTERPROCESS_SHARING &&
- ws->info.drm_minor >= 20)
- request.flags |= AMDGPU_GEM_CREATE_VM_ALWAYS_VALID;
+ if (ws->zero_all_vram_allocs &&
+ (request.preferred_heap & AMDGPU_GEM_DOMAIN_VRAM))
+ request.flags |= AMDGPU_GEM_CREATE_VRAM_CLEARED;
+ if ((flags & RADEON_FLAG_ENCRYPTED) && ws->secure)
+ request.flags |= AMDGPU_GEM_CREATE_ENCRYPTED;
r = amdgpu_bo_alloc(ws->dev, &request, &buf_handle);
if (r) {
fprintf(stderr, "amdgpu: size : %"PRIu64" bytes\n", size);
fprintf(stderr, "amdgpu: alignment : %u bytes\n", alignment);
fprintf(stderr, "amdgpu: domains : %u\n", initial_domain);
+ fprintf(stderr, "amdgpu: flags : %" PRIx64 "\n", request.flags);
goto error_bo_alloc;
}
- va_gap_size = ws->check_vm ? MAX2(4 * alignment, 64 * 1024) : 0;
- if (size > ws->info.pte_fragment_size)
- alignment = MAX2(alignment, ws->info.pte_fragment_size);
- r = amdgpu_va_range_alloc(ws->dev, amdgpu_gpu_va_range_general,
- size + va_gap_size, alignment, 0, &va, &va_handle, 0);
- if (r)
- goto error_va_alloc;
+ if (initial_domain & RADEON_DOMAIN_VRAM_GTT) {
+ unsigned va_gap_size = ws->check_vm ? MAX2(4 * alignment, 64 * 1024) : 0;
- r = amdgpu_bo_va_op(buf_handle, 0, size, va, 0, AMDGPU_VA_OP_MAP);
- if (r)
- goto error_va_map;
+ r = amdgpu_va_range_alloc(ws->dev, amdgpu_gpu_va_range_general,
+ size + va_gap_size,
+ amdgpu_get_optimal_vm_alignment(ws, size, alignment),
+ 0, &va, &va_handle,
+ (flags & RADEON_FLAG_32BIT ? AMDGPU_VA_RANGE_32_BIT : 0) |
+ AMDGPU_VA_RANGE_HIGH);
+ if (r)
+ goto error_va_alloc;
+
+ unsigned vm_flags = AMDGPU_VM_PAGE_READABLE |
+ AMDGPU_VM_PAGE_EXECUTABLE;
+
+ if (!(flags & RADEON_FLAG_READ_ONLY))
+ vm_flags |= AMDGPU_VM_PAGE_WRITEABLE;
+ if (flags & RADEON_FLAG_UNCACHED)
+ vm_flags |= AMDGPU_VM_MTYPE_UC;
+
+ r = amdgpu_bo_va_op_raw(ws->dev, buf_handle, 0, size, va, vm_flags,
+ AMDGPU_VA_OP_MAP);
+ if (r)
+ goto error_va_map;
+ }
+
+ simple_mtx_init(&bo->lock, mtx_plain);
pipe_reference_init(&bo->base.reference, 1);
bo->base.alignment = alignment;
- bo->base.usage = usage;
+ bo->base.usage = 0;
bo->base.size = size;
bo->base.vtbl = &amdgpu_winsys_bo_vtbl;
bo->ws = ws;
bo->va = va;
bo->u.real.va_handle = va_handle;
bo->initial_domain = initial_domain;
+ bo->flags = flags;
bo->unique_id = __sync_fetch_and_add(&ws->next_bo_unique_id, 1);
- bo->is_local = !!(request.flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID);
if (initial_domain & RADEON_DOMAIN_VRAM)
ws->allocated_vram += align64(size, ws->info.gart_page_size);
else if (initial_domain & RADEON_DOMAIN_GTT)
ws->allocated_gtt += align64(size, ws->info.gart_page_size);
+ amdgpu_bo_export(bo->bo, amdgpu_bo_handle_type_kms, &bo->u.real.kms_handle);
+
amdgpu_add_buffer_to_global_list(bo);
return bo;
return amdgpu_bo_can_reclaim(&bo->base);
}
+static struct pb_slabs *get_slabs(struct amdgpu_winsys *ws, uint64_t size,
+ enum radeon_bo_flag flags)
+{
+ struct pb_slabs *bo_slabs = ((flags & RADEON_FLAG_ENCRYPTED) && ws->secure) ?
+ ws->bo_slabs_encrypted : ws->bo_slabs;
+ /* Find the correct slab allocator for the given size. */
+ for (unsigned i = 0; i < NUM_SLAB_ALLOCATORS; i++) {
+ struct pb_slabs *slabs = &bo_slabs[i];
+
+ if (size <= 1 << (slabs->min_order + slabs->num_orders - 1))
+ return slabs;
+ }
+
+ assert(0);
+ return NULL;
+}
+
static void amdgpu_bo_slab_destroy(struct pb_buffer *_buf)
{
struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(_buf);
assert(!bo->bo);
- pb_slab_free(&bo->ws->bo_slabs, &bo->u.slab.entry);
+ if (bo->flags & RADEON_FLAG_ENCRYPTED)
+ pb_slab_free(get_slabs(bo->ws,
+ bo->base.size,
+ RADEON_FLAG_ENCRYPTED), &bo->u.slab.entry);
+ else
+ pb_slab_free(get_slabs(bo->ws,
+ bo->base.size,
+ 0), &bo->u.slab.entry);
}
static const struct pb_vtbl amdgpu_winsys_bo_slab_vtbl = {
/* other functions are never called */
};
-struct pb_slab *amdgpu_bo_slab_alloc(void *priv, unsigned heap,
- unsigned entry_size,
- unsigned group_index)
+static struct pb_slab *amdgpu_bo_slab_alloc(void *priv, unsigned heap,
+ unsigned entry_size,
+ unsigned group_index,
+ bool encrypted)
{
struct amdgpu_winsys *ws = priv;
struct amdgpu_slab *slab = CALLOC_STRUCT(amdgpu_slab);
enum radeon_bo_domain domains = radeon_domain_from_heap(heap);
enum radeon_bo_flag flags = radeon_flags_from_heap(heap);
uint32_t base_id;
+ unsigned slab_size = 0;
if (!slab)
return NULL;
- unsigned slab_size = 1 << AMDGPU_SLAB_BO_SIZE_LOG2;
- slab->buffer = amdgpu_winsys_bo(amdgpu_bo_create(&ws->base,
+ if (encrypted)
+ flags |= RADEON_FLAG_ENCRYPTED;
+
+ struct pb_slabs *slabs = (flags & RADEON_FLAG_ENCRYPTED && ws->secure) ?
+ ws->bo_slabs_encrypted : ws->bo_slabs;
+
+ /* Determine the slab buffer size. */
+ for (unsigned i = 0; i < NUM_SLAB_ALLOCATORS; i++) {
+ unsigned max_entry_size = 1 << (slabs[i].min_order + slabs[i].num_orders - 1);
+
+ if (entry_size <= max_entry_size) {
+ /* The slab size is twice the size of the largest possible entry. */
+ slab_size = max_entry_size * 2;
+
+ /* The largest slab should have the same size as the PTE fragment
+ * size to get faster address translation.
+ */
+ if (i == NUM_SLAB_ALLOCATORS - 1 &&
+ slab_size < ws->info.pte_fragment_size)
+ slab_size = ws->info.pte_fragment_size;
+ break;
+ }
+ }
+ assert(slab_size != 0);
+
+ slab->buffer = amdgpu_winsys_bo(amdgpu_bo_create(ws,
slab_size, slab_size,
domains, flags));
if (!slab->buffer)
goto fail;
- assert(slab->buffer->bo);
-
slab->base.num_entries = slab->buffer->base.size / entry_size;
slab->base.num_free = slab->base.num_entries;
slab->entries = CALLOC(slab->base.num_entries, sizeof(*slab->entries));
if (!slab->entries)
goto fail_buffer;
- LIST_INITHEAD(&slab->base.free);
+ list_inithead(&slab->base.free);
base_id = __sync_fetch_and_add(&ws->next_bo_unique_id, slab->base.num_entries);
for (unsigned i = 0; i < slab->base.num_entries; ++i) {
struct amdgpu_winsys_bo *bo = &slab->entries[i];
+ simple_mtx_init(&bo->lock, mtx_plain);
bo->base.alignment = entry_size;
bo->base.usage = slab->buffer->base.usage;
bo->base.size = entry_size;
bo->unique_id = base_id + i;
bo->u.slab.entry.slab = &slab->base;
bo->u.slab.entry.group_index = group_index;
- bo->u.slab.real = slab->buffer;
- LIST_ADDTAIL(&bo->u.slab.entry.head, &slab->base.free);
+ if (slab->buffer->bo) {
+ /* The slab is not suballocated. */
+ bo->u.slab.real = slab->buffer;
+ } else {
+ /* The slab is allocated out of a bigger slab. */
+ bo->u.slab.real = slab->buffer->u.slab.real;
+ assert(bo->u.slab.real->bo);
+ }
+
+ list_addtail(&bo->u.slab.entry.head, &slab->base.free);
}
return &slab->base;
return NULL;
}
+struct pb_slab *amdgpu_bo_slab_alloc_encrypted(void *priv, unsigned heap,
+ unsigned entry_size,
+ unsigned group_index)
+{
+ return amdgpu_bo_slab_alloc(priv, heap, entry_size, group_index, true);
+}
+
+struct pb_slab *amdgpu_bo_slab_alloc_normal(void *priv, unsigned heap,
+ unsigned entry_size,
+ unsigned group_index)
+{
+ return amdgpu_bo_slab_alloc(priv, heap, entry_size, group_index, false);
+}
+
void amdgpu_bo_slab_free(void *priv, struct pb_slab *pslab)
{
struct amdgpu_slab *slab = amdgpu_slab(pslab);
- for (unsigned i = 0; i < slab->base.num_entries; ++i)
+ for (unsigned i = 0; i < slab->base.num_entries; ++i) {
amdgpu_bo_remove_fences(&slab->entries[i]);
+ simple_mtx_destroy(&slab->entries[i].lock);
+ }
FREE(slab->entries);
amdgpu_winsys_bo_reference(&slab->buffer, NULL);
bo->base.size - (uint64_t)bo->u.sparse.num_backing_pages * RADEON_SPARSE_PAGE_SIZE);
size = MAX2(size, RADEON_SPARSE_PAGE_SIZE);
- buf = amdgpu_bo_create(&bo->ws->base, size, RADEON_SPARSE_PAGE_SIZE,
+ buf = amdgpu_bo_create(bo->ws, size, RADEON_SPARSE_PAGE_SIZE,
bo->initial_domain,
bo->u.sparse.flags | RADEON_FLAG_NO_SUBALLOC);
if (!buf) {
fprintf(stderr, "amdgpu: clearing PRT VA region on destroy failed (%d)\n", r);
}
- while (!list_empty(&bo->u.sparse.backing)) {
+ while (!list_is_empty(&bo->u.sparse.backing)) {
struct amdgpu_sparse_backing *dummy = NULL;
sparse_free_backing_buffer(bo,
container_of(bo->u.sparse.backing.next,
}
amdgpu_va_range_free(bo->u.sparse.va_handle);
- simple_mtx_destroy(&bo->u.sparse.commit_lock);
FREE(bo->u.sparse.commitments);
+ simple_mtx_destroy(&bo->lock);
FREE(bo);
}
if (!bo)
return NULL;
+ simple_mtx_init(&bo->lock, mtx_plain);
pipe_reference_init(&bo->base.reference, 1);
bo->base.alignment = RADEON_SPARSE_PAGE_SIZE;
bo->base.size = size;
if (!bo->u.sparse.commitments)
goto error_alloc_commitments;
- simple_mtx_init(&bo->u.sparse.commit_lock, mtx_plain);
- LIST_INITHEAD(&bo->u.sparse.backing);
+ list_inithead(&bo->u.sparse.backing);
/* For simplicity, we always map a multiple of the page size. */
map_size = align64(size, RADEON_SPARSE_PAGE_SIZE);
va_gap_size = ws->check_vm ? 4 * RADEON_SPARSE_PAGE_SIZE : 0;
r = amdgpu_va_range_alloc(ws->dev, amdgpu_gpu_va_range_general,
map_size + va_gap_size, RADEON_SPARSE_PAGE_SIZE,
- 0, &bo->va, &bo->u.sparse.va_handle, 0);
+ 0, &bo->va, &bo->u.sparse.va_handle,
+ AMDGPU_VA_RANGE_HIGH);
if (r)
goto error_va_alloc;
error_va_map:
amdgpu_va_range_free(bo->u.sparse.va_handle);
error_va_alloc:
- simple_mtx_destroy(&bo->u.sparse.commit_lock);
FREE(bo->u.sparse.commitments);
error_alloc_commitments:
+ simple_mtx_destroy(&bo->lock);
FREE(bo);
return NULL;
}
va_page = offset / RADEON_SPARSE_PAGE_SIZE;
end_va_page = va_page + DIV_ROUND_UP(size, RADEON_SPARSE_PAGE_SIZE);
- simple_mtx_lock(&bo->u.sparse.commit_lock);
+ simple_mtx_lock(&bo->lock);
#if DEBUG_SPARSE_COMMITS
sparse_dump(bo, __func__);
}
out:
- simple_mtx_unlock(&bo->u.sparse.commit_lock);
+ simple_mtx_unlock(&bo->lock);
return ok;
}
-static unsigned eg_tile_split(unsigned tile_split)
-{
- switch (tile_split) {
- case 0: tile_split = 64; break;
- case 1: tile_split = 128; break;
- case 2: tile_split = 256; break;
- case 3: tile_split = 512; break;
- default:
- case 4: tile_split = 1024; break;
- case 5: tile_split = 2048; break;
- case 6: tile_split = 4096; break;
- }
- return tile_split;
-}
-
-static unsigned eg_tile_split_rev(unsigned eg_tile_split)
-{
- switch (eg_tile_split) {
- case 64: return 0;
- case 128: return 1;
- case 256: return 2;
- case 512: return 3;
- default:
- case 1024: return 4;
- case 2048: return 5;
- case 4096: return 6;
- }
-}
-
static void amdgpu_buffer_get_metadata(struct pb_buffer *_buf,
- struct radeon_bo_metadata *md)
+ struct radeon_bo_metadata *md,
+ struct radeon_surf *surf)
{
struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(_buf);
struct amdgpu_bo_info info = {0};
- uint64_t tiling_flags;
int r;
assert(bo->bo && "must not be called for slab entries");
if (r)
return;
- tiling_flags = info.metadata.tiling_info;
-
- if (bo->ws->info.chip_class >= GFX9) {
- md->u.gfx9.swizzle_mode = AMDGPU_TILING_GET(tiling_flags, SWIZZLE_MODE);
- } else {
- md->u.legacy.microtile = RADEON_LAYOUT_LINEAR;
- md->u.legacy.macrotile = RADEON_LAYOUT_LINEAR;
-
- if (AMDGPU_TILING_GET(tiling_flags, ARRAY_MODE) == 4) /* 2D_TILED_THIN1 */
- md->u.legacy.macrotile = RADEON_LAYOUT_TILED;
- else if (AMDGPU_TILING_GET(tiling_flags, ARRAY_MODE) == 2) /* 1D_TILED_THIN1 */
- md->u.legacy.microtile = RADEON_LAYOUT_TILED;
-
- md->u.legacy.pipe_config = AMDGPU_TILING_GET(tiling_flags, PIPE_CONFIG);
- md->u.legacy.bankw = 1 << AMDGPU_TILING_GET(tiling_flags, BANK_WIDTH);
- md->u.legacy.bankh = 1 << AMDGPU_TILING_GET(tiling_flags, BANK_HEIGHT);
- md->u.legacy.tile_split = eg_tile_split(AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT));
- md->u.legacy.mtilea = 1 << AMDGPU_TILING_GET(tiling_flags, MACRO_TILE_ASPECT);
- md->u.legacy.num_banks = 2 << AMDGPU_TILING_GET(tiling_flags, NUM_BANKS);
- md->u.legacy.scanout = AMDGPU_TILING_GET(tiling_flags, MICRO_TILE_MODE) == 0; /* DISPLAY */
- }
+ ac_surface_set_bo_metadata(&bo->ws->info, surf, info.metadata.tiling_info,
+ &md->mode);
md->size_metadata = info.metadata.size_metadata;
memcpy(md->metadata, info.metadata.umd_metadata, sizeof(md->metadata));
}
static void amdgpu_buffer_set_metadata(struct pb_buffer *_buf,
- struct radeon_bo_metadata *md)
+ struct radeon_bo_metadata *md,
+ struct radeon_surf *surf)
{
struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(_buf);
struct amdgpu_bo_metadata metadata = {0};
- uint64_t tiling_flags = 0;
assert(bo->bo && "must not be called for slab entries");
- if (bo->ws->info.chip_class >= GFX9) {
- tiling_flags |= AMDGPU_TILING_SET(SWIZZLE_MODE, md->u.gfx9.swizzle_mode);
- } else {
- if (md->u.legacy.macrotile == RADEON_LAYOUT_TILED)
- tiling_flags |= AMDGPU_TILING_SET(ARRAY_MODE, 4); /* 2D_TILED_THIN1 */
- else if (md->u.legacy.microtile == RADEON_LAYOUT_TILED)
- tiling_flags |= AMDGPU_TILING_SET(ARRAY_MODE, 2); /* 1D_TILED_THIN1 */
- else
- tiling_flags |= AMDGPU_TILING_SET(ARRAY_MODE, 1); /* LINEAR_ALIGNED */
-
- tiling_flags |= AMDGPU_TILING_SET(PIPE_CONFIG, md->u.legacy.pipe_config);
- tiling_flags |= AMDGPU_TILING_SET(BANK_WIDTH, util_logbase2(md->u.legacy.bankw));
- tiling_flags |= AMDGPU_TILING_SET(BANK_HEIGHT, util_logbase2(md->u.legacy.bankh));
- if (md->u.legacy.tile_split)
- tiling_flags |= AMDGPU_TILING_SET(TILE_SPLIT, eg_tile_split_rev(md->u.legacy.tile_split));
- tiling_flags |= AMDGPU_TILING_SET(MACRO_TILE_ASPECT, util_logbase2(md->u.legacy.mtilea));
- tiling_flags |= AMDGPU_TILING_SET(NUM_BANKS, util_logbase2(md->u.legacy.num_banks)-1);
-
- if (md->u.legacy.scanout)
- tiling_flags |= AMDGPU_TILING_SET(MICRO_TILE_MODE, 0); /* DISPLAY_MICRO_TILING */
- else
- tiling_flags |= AMDGPU_TILING_SET(MICRO_TILE_MODE, 1); /* THIN_MICRO_TILING */
- }
+ ac_surface_get_bo_metadata(&bo->ws->info, surf, &metadata.tiling_info);
- metadata.tiling_info = tiling_flags;
metadata.size_metadata = md->size_metadata;
memcpy(metadata.umd_metadata, md->metadata, sizeof(md->metadata));
amdgpu_bo_set_metadata(bo->bo, &metadata);
}
-static struct pb_buffer *
-amdgpu_bo_create(struct radeon_winsys *rws,
+struct pb_buffer *
+amdgpu_bo_create(struct amdgpu_winsys *ws,
uint64_t size,
unsigned alignment,
enum radeon_bo_domain domain,
enum radeon_bo_flag flags)
{
- struct amdgpu_winsys *ws = amdgpu_winsys(rws);
struct amdgpu_winsys_bo *bo;
- unsigned usage = 0, pb_cache_bucket = 0;
+ int heap = -1;
+
+ if (domain & (RADEON_DOMAIN_GDS | RADEON_DOMAIN_OA))
+ flags |= RADEON_FLAG_NO_CPU_ACCESS | RADEON_FLAG_NO_SUBALLOC;
/* VRAM implies WC. This is not optional. */
assert(!(domain & RADEON_DOMAIN_VRAM) || flags & RADEON_FLAG_GTT_WC);
- /* NO_CPU_ACCESS is valid with VRAM only. */
- assert(domain == RADEON_DOMAIN_VRAM || !(flags & RADEON_FLAG_NO_CPU_ACCESS));
+ /* NO_CPU_ACCESS is not valid with GTT. */
+ assert(!(domain & RADEON_DOMAIN_GTT) || !(flags & RADEON_FLAG_NO_CPU_ACCESS));
+
+ /* Sparse buffers must have NO_CPU_ACCESS set. */
+ assert(!(flags & RADEON_FLAG_SPARSE) || flags & RADEON_FLAG_NO_CPU_ACCESS);
+
+ struct pb_slabs *slabs = (flags & RADEON_FLAG_ENCRYPTED && ws->secure) ?
+ ws->bo_slabs_encrypted : ws->bo_slabs;
+ struct pb_slabs *last_slab = &slabs[NUM_SLAB_ALLOCATORS - 1];
+ unsigned max_slab_entry_size = 1 << (last_slab->min_order + last_slab->num_orders - 1);
/* Sub-allocate small buffers from slabs. */
if (!(flags & (RADEON_FLAG_NO_SUBALLOC | RADEON_FLAG_SPARSE)) &&
- size <= (1 << AMDGPU_SLAB_MAX_SIZE_LOG2) &&
- alignment <= MAX2(1 << AMDGPU_SLAB_MIN_SIZE_LOG2, util_next_power_of_two(size))) {
+ size <= max_slab_entry_size &&
+ /* The alignment must be at most the size of the smallest slab entry or
+ * the next power of two. */
+ alignment <= MAX2(1 << slabs[0].min_order, util_next_power_of_two(size))) {
struct pb_slab_entry *entry;
int heap = radeon_get_heap_index(domain, flags);
if (heap < 0 || heap >= RADEON_MAX_SLAB_HEAPS)
goto no_slab;
- entry = pb_slab_alloc(&ws->bo_slabs, size, heap);
+ struct pb_slabs *slabs = get_slabs(ws, size, flags);
+ entry = pb_slab_alloc(slabs, size, heap);
if (!entry) {
- /* Clear the cache and try again. */
- pb_cache_release_all_buffers(&ws->bo_cache);
+ /* Clean up buffer managers and try again. */
+ amdgpu_clean_up_buffer_managers(ws);
- entry = pb_slab_alloc(&ws->bo_slabs, size, heap);
+ entry = pb_slab_alloc(slabs, size, heap);
}
if (!entry)
return NULL;
if (flags & RADEON_FLAG_SPARSE) {
assert(RADEON_SPARSE_PAGE_SIZE % alignment == 0);
- flags |= RADEON_FLAG_NO_CPU_ACCESS;
-
return amdgpu_bo_sparse_create(ws, size, domain, flags);
}
* BOs. Aligning this here helps the cached bufmgr. Especially small BOs,
* like constant/uniform buffers, can benefit from better and more reuse.
*/
- size = align64(size, ws->info.gart_page_size);
- alignment = align(alignment, ws->info.gart_page_size);
+ if (domain & RADEON_DOMAIN_VRAM_GTT) {
+ size = align64(size, ws->info.gart_page_size);
+ alignment = align(alignment, ws->info.gart_page_size);
+ }
bool use_reusable_pool = flags & RADEON_FLAG_NO_INTERPROCESS_SHARING;
if (use_reusable_pool) {
- int heap = radeon_get_heap_index(domain, flags);
+ heap = radeon_get_heap_index(domain, flags & ~RADEON_FLAG_ENCRYPTED);
assert(heap >= 0 && heap < RADEON_MAX_CACHED_HEAPS);
- usage = 1 << heap; /* Only set one usage bit for each heap. */
-
- pb_cache_bucket = radeon_get_pb_cache_bucket_index(heap);
- assert(pb_cache_bucket < ARRAY_SIZE(ws->bo_cache.buckets));
/* Get a buffer from the cache. */
bo = (struct amdgpu_winsys_bo*)
- pb_cache_reclaim_buffer(&ws->bo_cache, size, alignment, usage,
- pb_cache_bucket);
+ pb_cache_reclaim_buffer(&ws->bo_cache, size, alignment, 0, heap);
if (bo)
return &bo->base;
}
/* Create a new one. */
- bo = amdgpu_create_bo(ws, size, alignment, usage, domain, flags,
- pb_cache_bucket);
+ bo = amdgpu_create_bo(ws, size, alignment, domain, flags, heap);
if (!bo) {
- /* Clear the cache and try again. */
- pb_slabs_reclaim(&ws->bo_slabs);
- pb_cache_release_all_buffers(&ws->bo_cache);
- bo = amdgpu_create_bo(ws, size, alignment, usage, domain, flags,
- pb_cache_bucket);
+ /* Clean up buffer managers and try again. */
+ amdgpu_clean_up_buffer_managers(ws);
+
+ bo = amdgpu_create_bo(ws, size, alignment, domain, flags, heap);
if (!bo)
return NULL;
}
return &bo->base;
}
+static struct pb_buffer *
+amdgpu_buffer_create(struct radeon_winsys *ws,
+ uint64_t size,
+ unsigned alignment,
+ enum radeon_bo_domain domain,
+ enum radeon_bo_flag flags)
+{
+ struct pb_buffer * res = amdgpu_bo_create(amdgpu_winsys(ws), size, alignment, domain,
+ flags);
+ return res;
+}
+
static struct pb_buffer *amdgpu_bo_from_handle(struct radeon_winsys *rws,
struct winsys_handle *whandle,
- unsigned *stride,
- unsigned *offset)
+ unsigned vm_alignment)
{
struct amdgpu_winsys *ws = amdgpu_winsys(rws);
- struct amdgpu_winsys_bo *bo;
+ struct amdgpu_winsys_bo *bo = NULL;
enum amdgpu_bo_handle_type type;
struct amdgpu_bo_import_result result = {0};
uint64_t va;
- amdgpu_va_handle va_handle;
+ amdgpu_va_handle va_handle = NULL;
struct amdgpu_bo_info info = {0};
enum radeon_bo_domain initial = 0;
+ enum radeon_bo_flag flags = 0;
int r;
- /* Initialize the structure. */
- bo = CALLOC_STRUCT(amdgpu_winsys_bo);
- if (!bo) {
- return NULL;
- }
-
switch (whandle->type) {
- case DRM_API_HANDLE_TYPE_SHARED:
+ case WINSYS_HANDLE_TYPE_SHARED:
type = amdgpu_bo_handle_type_gem_flink_name;
break;
- case DRM_API_HANDLE_TYPE_FD:
+ case WINSYS_HANDLE_TYPE_FD:
type = amdgpu_bo_handle_type_dma_buf_fd;
break;
default:
r = amdgpu_bo_import(ws->dev, type, whandle->handle, &result);
if (r)
- goto error;
+ return NULL;
+
+ simple_mtx_lock(&ws->bo_export_table_lock);
+ bo = util_hash_table_get(ws->bo_export_table, result.buf_handle);
+
+ /* If the amdgpu_winsys_bo instance already exists, bump the reference
+ * counter and return it.
+ */
+ if (bo) {
+ p_atomic_inc(&bo->base.reference.count);
+ simple_mtx_unlock(&ws->bo_export_table_lock);
+
+ /* Release the buffer handle, because we don't need it anymore.
+ * This function is returning an existing buffer, which has its own
+ * handle.
+ */
+ amdgpu_bo_free(result.buf_handle);
+ return &bo->base;
+ }
/* Get initial domains. */
r = amdgpu_bo_query_info(result.buf_handle, &info);
if (r)
- goto error_query;
+ goto error;
r = amdgpu_va_range_alloc(ws->dev, amdgpu_gpu_va_range_general,
- result.alloc_size, 1 << 20, 0, &va, &va_handle, 0);
+ result.alloc_size,
+ amdgpu_get_optimal_vm_alignment(ws, result.alloc_size,
+ vm_alignment),
+ 0, &va, &va_handle, AMDGPU_VA_RANGE_HIGH);
if (r)
- goto error_query;
+ goto error;
+
+ bo = CALLOC_STRUCT(amdgpu_winsys_bo);
+ if (!bo)
+ goto error;
r = amdgpu_bo_va_op(result.buf_handle, 0, result.alloc_size, va, 0, AMDGPU_VA_OP_MAP);
if (r)
- goto error_va_map;
+ goto error;
if (info.preferred_heap & AMDGPU_GEM_DOMAIN_VRAM)
initial |= RADEON_DOMAIN_VRAM;
if (info.preferred_heap & AMDGPU_GEM_DOMAIN_GTT)
initial |= RADEON_DOMAIN_GTT;
+ if (info.alloc_flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
+ flags |= RADEON_FLAG_NO_CPU_ACCESS;
+ if (info.alloc_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
+ flags |= RADEON_FLAG_GTT_WC;
+ if (info.alloc_flags & AMDGPU_GEM_CREATE_ENCRYPTED)
+ flags |= RADEON_FLAG_ENCRYPTED;
-
+ /* Initialize the structure. */
+ simple_mtx_init(&bo->lock, mtx_plain);
pipe_reference_init(&bo->base.reference, 1);
bo->base.alignment = info.phys_alignment;
bo->bo = result.buf_handle;
bo->va = va;
bo->u.real.va_handle = va_handle;
bo->initial_domain = initial;
+ bo->flags = flags;
bo->unique_id = __sync_fetch_and_add(&ws->next_bo_unique_id, 1);
bo->is_shared = true;
- if (stride)
- *stride = whandle->stride;
- if (offset)
- *offset = whandle->offset;
-
if (bo->initial_domain & RADEON_DOMAIN_VRAM)
ws->allocated_vram += align64(bo->base.size, ws->info.gart_page_size);
else if (bo->initial_domain & RADEON_DOMAIN_GTT)
ws->allocated_gtt += align64(bo->base.size, ws->info.gart_page_size);
+ amdgpu_bo_export(bo->bo, amdgpu_bo_handle_type_kms, &bo->u.real.kms_handle);
+
amdgpu_add_buffer_to_global_list(bo);
- return &bo->base;
+ _mesa_hash_table_insert(ws->bo_export_table, bo->bo, bo);
+ simple_mtx_unlock(&ws->bo_export_table_lock);
-error_va_map:
- amdgpu_va_range_free(va_handle);
-
-error_query:
- amdgpu_bo_free(result.buf_handle);
+ return &bo->base;
error:
- FREE(bo);
+ simple_mtx_unlock(&ws->bo_export_table_lock);
+ if (bo)
+ FREE(bo);
+ if (va_handle)
+ amdgpu_va_range_free(va_handle);
+ amdgpu_bo_free(result.buf_handle);
return NULL;
}
-static bool amdgpu_bo_get_handle(struct pb_buffer *buffer,
- unsigned stride, unsigned offset,
- unsigned slice_size,
+static bool amdgpu_bo_get_handle(struct radeon_winsys *rws,
+ struct pb_buffer *buffer,
struct winsys_handle *whandle)
{
+ struct amdgpu_screen_winsys *sws = amdgpu_screen_winsys(rws);
struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(buffer);
+ struct amdgpu_winsys *ws = bo->ws;
enum amdgpu_bo_handle_type type;
+ struct hash_entry *entry;
int r;
/* Don't allow exports of slab entries and sparse buffers. */
bo->u.real.use_reusable_pool = false;
switch (whandle->type) {
- case DRM_API_HANDLE_TYPE_SHARED:
+ case WINSYS_HANDLE_TYPE_SHARED:
type = amdgpu_bo_handle_type_gem_flink_name;
break;
- case DRM_API_HANDLE_TYPE_FD:
+ case WINSYS_HANDLE_TYPE_KMS:
+ if (sws->fd == ws->fd) {
+ whandle->handle = bo->u.real.kms_handle;
+
+ if (bo->is_shared)
+ return true;
+
+ goto hash_table_set;
+ }
+
+ simple_mtx_lock(&ws->sws_list_lock);
+ entry = _mesa_hash_table_search(sws->kms_handles, bo);
+ simple_mtx_unlock(&ws->sws_list_lock);
+ if (entry) {
+ whandle->handle = (uintptr_t)entry->data;
+ return true;
+ }
+ /* Fall through */
+ case WINSYS_HANDLE_TYPE_FD:
type = amdgpu_bo_handle_type_dma_buf_fd;
break;
- case DRM_API_HANDLE_TYPE_KMS:
- type = amdgpu_bo_handle_type_kms;
- break;
default:
return false;
}
if (r)
return false;
- whandle->stride = stride;
- whandle->offset = offset;
- whandle->offset += slice_size * whandle->layer;
+ if (whandle->type == WINSYS_HANDLE_TYPE_KMS) {
+ int dma_fd = whandle->handle;
+
+ r = drmPrimeFDToHandle(sws->fd, dma_fd, &whandle->handle);
+ close(dma_fd);
+
+ if (r)
+ return false;
+
+ simple_mtx_lock(&ws->sws_list_lock);
+ _mesa_hash_table_insert_pre_hashed(sws->kms_handles,
+ bo->u.real.kms_handle, bo,
+ (void*)(uintptr_t)whandle->handle);
+ simple_mtx_unlock(&ws->sws_list_lock);
+ }
+
+ hash_table_set:
+ simple_mtx_lock(&ws->bo_export_table_lock);
+ _mesa_hash_table_insert(ws->bo_export_table, bo->bo, bo);
+ simple_mtx_unlock(&ws->bo_export_table_lock);
+
bo->is_shared = true;
return true;
}
struct amdgpu_winsys_bo *bo;
uint64_t va;
amdgpu_va_handle va_handle;
+ /* Avoid failure when the size is not page aligned */
+ uint64_t aligned_size = align64(size, ws->info.gart_page_size);
bo = CALLOC_STRUCT(amdgpu_winsys_bo);
if (!bo)
return NULL;
- if (amdgpu_create_bo_from_user_mem(ws->dev, pointer, size, &buf_handle))
+ if (amdgpu_create_bo_from_user_mem(ws->dev, pointer,
+ aligned_size, &buf_handle))
goto error;
if (amdgpu_va_range_alloc(ws->dev, amdgpu_gpu_va_range_general,
- size, 1 << 12, 0, &va, &va_handle, 0))
+ aligned_size,
+ amdgpu_get_optimal_vm_alignment(ws, aligned_size,
+ ws->info.gart_page_size),
+ 0, &va, &va_handle, AMDGPU_VA_RANGE_HIGH))
goto error_va_alloc;
- if (amdgpu_bo_va_op(buf_handle, 0, size, va, 0, AMDGPU_VA_OP_MAP))
+ if (amdgpu_bo_va_op(buf_handle, 0, aligned_size, va, 0, AMDGPU_VA_OP_MAP))
goto error_va_map;
/* Initialize it. */
+ bo->is_user_ptr = true;
pipe_reference_init(&bo->base.reference, 1);
+ simple_mtx_init(&bo->lock, mtx_plain);
bo->bo = buf_handle;
bo->base.alignment = 0;
bo->base.size = size;
bo->base.vtbl = &amdgpu_winsys_bo_vtbl;
bo->ws = ws;
- bo->user_ptr = pointer;
+ bo->cpu_ptr = pointer;
bo->va = va;
bo->u.real.va_handle = va_handle;
bo->initial_domain = RADEON_DOMAIN_GTT;
bo->unique_id = __sync_fetch_and_add(&ws->next_bo_unique_id, 1);
- ws->allocated_gtt += align64(bo->base.size, ws->info.gart_page_size);
+ ws->allocated_gtt += aligned_size;
amdgpu_add_buffer_to_global_list(bo);
+ amdgpu_bo_export(bo->bo, amdgpu_bo_handle_type_kms, &bo->u.real.kms_handle);
+
return (struct pb_buffer*)bo;
error_va_map:
static bool amdgpu_bo_is_user_ptr(struct pb_buffer *buf)
{
- return ((struct amdgpu_winsys_bo*)buf)->user_ptr != NULL;
+ return ((struct amdgpu_winsys_bo*)buf)->is_user_ptr;
}
static bool amdgpu_bo_is_suballocated(struct pb_buffer *buf)
return ((struct amdgpu_winsys_bo*)buf)->va;
}
-void amdgpu_bo_init_functions(struct amdgpu_winsys *ws)
+void amdgpu_bo_init_functions(struct amdgpu_screen_winsys *ws)
{
ws->base.buffer_set_metadata = amdgpu_buffer_set_metadata;
ws->base.buffer_get_metadata = amdgpu_buffer_get_metadata;
ws->base.buffer_map = amdgpu_bo_map;
ws->base.buffer_unmap = amdgpu_bo_unmap;
ws->base.buffer_wait = amdgpu_bo_wait;
- ws->base.buffer_create = amdgpu_bo_create;
+ ws->base.buffer_create = amdgpu_buffer_create;
ws->base.buffer_from_handle = amdgpu_bo_from_handle;
ws->base.buffer_from_ptr = amdgpu_bo_from_ptr;
ws->base.buffer_is_user_ptr = amdgpu_bo_is_user_ptr;
ws->base.buffer_commit = amdgpu_bo_sparse_commit;
ws->base.buffer_get_virtual_address = amdgpu_bo_get_va;
ws->base.buffer_get_initial_domain = amdgpu_bo_get_initial_domain;
+ ws->base.buffer_get_flags = amdgpu_bo_get_flags;
}