#include "amdgpu_cs.h"
+#include "util/hash_table.h"
#include "util/os_time.h"
#include "util/u_hash_table.h"
-#include "state_tracker/drm_driver.h"
-#include <amdgpu_drm.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)
-#endif
-
#ifndef AMDGPU_VA_RANGE_HIGH
#define AMDGPU_VA_RANGE_HIGH 0x2
#endif
uint32_t begin, end;
};
-static void amdgpu_bo_unmap(struct pb_buffer *buf);
-
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");
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);
- util_hash_table_remove(ws->bo_export_table, bo->bo);
+ _mesa_hash_table_remove_key(ws->bo_export_table, bo->bo);
simple_mtx_unlock(&ws->bo_export_table_lock);
if (bo->initial_domain & RADEON_DOMAIN_VRAM_GTT) {
static void amdgpu_clean_up_buffer_managers(struct amdgpu_winsys *ws)
{
- for (unsigned i = 0; i < NUM_SLAB_ALLOCATORS; i++)
+ 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);
}
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;
amdgpu_bo_handle buf_handle;
uint64_t va = 0;
struct amdgpu_winsys_bo *bo;
- amdgpu_va_handle va_handle;
+ amdgpu_va_handle va_handle = NULL;
int r;
/* VRAM or GTT must be specified, but not both at the same time. */
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.has_local_buffers)
- 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;
}
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)
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);
return amdgpu_bo_can_reclaim(&bo->base);
}
-static struct pb_slabs *get_slabs(struct amdgpu_winsys *ws, uint64_t size)
+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 = &ws->bo_slabs[i];
+ struct pb_slabs *slabs = &bo_slabs[i];
if (size <= 1 << (slabs->min_order + slabs->num_orders - 1))
return slabs;
assert(!bo->bo);
- pb_slab_free(get_slabs(bo->ws, bo->base.size), &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);
if (!slab)
return NULL;
+ 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++) {
- struct pb_slabs *slabs = &ws->bo_slabs[i];
- unsigned max_entry_size = 1 << (slabs->min_order + slabs->num_orders - 1);
+ 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. */
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);
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);
-
- md->u.gfx9.dcc_offset_256B = AMDGPU_TILING_GET(tiling_flags, DCC_OFFSET_256B);
- md->u.gfx9.dcc_pitch_max = AMDGPU_TILING_GET(tiling_flags, DCC_PITCH_MAX);
- md->u.gfx9.dcc_independent_64B = AMDGPU_TILING_GET(tiling_flags, DCC_INDEPENDENT_64B);
- } 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);
-
- tiling_flags |= AMDGPU_TILING_SET(DCC_OFFSET_256B, md->u.gfx9.dcc_offset_256B);
- tiling_flags |= AMDGPU_TILING_SET(DCC_PITCH_MAX, md->u.gfx9.dcc_pitch_max);
- tiling_flags |= AMDGPU_TILING_SET(DCC_INDEPENDENT_64B, md->u.gfx9.dcc_independent_64B);
- } 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));
/* Sparse buffers must have NO_CPU_ACCESS set. */
assert(!(flags & RADEON_FLAG_SPARSE) || flags & RADEON_FLAG_NO_CPU_ACCESS);
- struct pb_slabs *last_slab = &ws->bo_slabs[NUM_SLAB_ALLOCATORS - 1];
+ 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. */
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 << ws->bo_slabs[0].min_order, util_next_power_of_two(size))) {
+ 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;
- struct pb_slabs *slabs = get_slabs(ws, size);
+ struct pb_slabs *slabs = get_slabs(ws, size, flags);
entry = pb_slab_alloc(slabs, size, heap);
if (!entry) {
/* Clean up buffer managers and try again. */
bool use_reusable_pool = flags & RADEON_FLAG_NO_INTERPROCESS_SHARING;
if (use_reusable_pool) {
- 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);
/* Get a buffer from the cache. */
enum radeon_bo_domain domain,
enum radeon_bo_flag flags)
{
- return amdgpu_bo_create(amdgpu_winsys(ws), size, alignment, domain,
+ 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,
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;
switch (whandle->type) {
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);
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;
amdgpu_add_buffer_to_global_list(bo);
- util_hash_table_set(ws->bo_export_table, bo->bo, bo);
+ _mesa_hash_table_insert(ws->bo_export_table, bo->bo, bo);
simple_mtx_unlock(&ws->bo_export_table_lock);
return &bo->base;
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. */
type = amdgpu_bo_handle_type_gem_flink_name;
break;
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;
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);
- util_hash_table_set(ws->bo_export_table, bo->bo, bo);
+ _mesa_hash_table_insert(ws->bo_export_table, bo->bo, bo);
simple_mtx_unlock(&ws->bo_export_table_lock);
bo->is_shared = true;
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;
}