static void radeon_drm_cs_destroy(struct radeon_winsys_cs *rcs)
{
struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
radeon_drm_cs_sync_flush(rcs);
pipe_semaphore_destroy(&cs->flush_completed);
radeon_cs_context_cleanup(&cs->csc1);
radeon_cs_context_cleanup(&cs->csc2);
p_atomic_dec(&cs->ws->num_cs);
radeon_destroy_cs_context(&cs->csc1);
radeon_destroy_cs_context(&cs->csc2);
FREE(cs);
}
static void radeon_drm_cs_destroy(struct radeon_cmdbuf *rcs)
{
struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
radeon_drm_cs_sync_flush(rcs);
util_queue_fence_destroy(&cs->flush_completed);
radeon_cs_context_cleanup(&cs->csc1);
radeon_cs_context_cleanup(&cs->csc2);
p_atomic_dec(&cs->ws->num_cs);
radeon_destroy_cs_context(&cs->csc1);
radeon_destroy_cs_context(&cs->csc2);
radeon_fence_reference(&cs->next_fence, NULL);
FREE(cs);
}
void radeon_drm_cs_emit_ioctl_oneshot(struct radeon_drm_cs *cs, struct radeon_cs_context *csc)
{
unsigned i;
if (drmCommandWriteRead(csc->fd, DRM_RADEON_CS,
&csc->cs, sizeof(struct drm_radeon_cs))) {
if (debug_get_bool_option("RADEON_DUMP_CS", FALSE)) {
unsigned i;
fprintf(stderr, "radeon: The kernel rejected CS, dumping...\n");
for (i = 0; i < csc->chunks[0].length_dw; i++) {
fprintf(stderr, "0x%08X\n", csc->buf[i]);
}
} else {
fprintf(stderr, "radeon: The kernel rejected CS, "
"see dmesg for more information.\n");
}
}
if (cs->trace_buf) {
radeon_dump_cs_on_lockup(cs, csc);
}
for (i = 0; i < csc->crelocs; i++)
p_atomic_dec(&csc->relocs_bo[i]->num_active_ioctls);
radeon_cs_context_cleanup(csc);
}
static boolean radeon_drm_cs_validate(struct radeon_winsys_cs *rcs)
{
struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
boolean status =
cs->csc->used_gart < cs->ws->info.gart_size * 0.8 &&
cs->csc->used_vram < cs->ws->info.vram_size * 0.8;
if (status) {
cs->csc->validated_crelocs = cs->csc->crelocs;
} else {
/* Remove lately-added relocations. The validation failed with them
* and the CS is about to be flushed because of that. Keep only
* the already-validated relocations. */
unsigned i;
for (i = cs->csc->validated_crelocs; i < cs->csc->crelocs; i++) {
p_atomic_dec(&cs->csc->relocs_bo[i]->num_cs_references);
radeon_bo_reference(&cs->csc->relocs_bo[i], NULL);
}
cs->csc->crelocs = cs->csc->validated_crelocs;
/* Flush if there are any relocs. Clean up otherwise. */
if (cs->csc->crelocs) {
cs->flush_cs(cs->flush_data, RADEON_FLUSH_ASYNC);
} else {
radeon_cs_context_cleanup(cs->csc);
assert(cs->base.cdw == 0);
if (cs->base.cdw != 0) {
fprintf(stderr, "radeon: Unexpected error in %s.\n", __func__);
}
}
}
return status;
}
void radeon_drm_cs_emit_ioctl_oneshot(void *job, int thread_index)
{
struct radeon_cs_context *csc = ((struct radeon_drm_cs*)job)->cst;
unsigned i;
int r;
r = drmCommandWriteRead(csc->fd, DRM_RADEON_CS,
&csc->cs, sizeof(struct drm_radeon_cs));
if (r) {
if (r == -ENOMEM)
fprintf(stderr, "radeon: Not enough memory for command submission.\n");
else if (debug_get_bool_option("RADEON_DUMP_CS", false)) {
unsigned i;
fprintf(stderr, "radeon: The kernel rejected CS, dumping...\n");
for (i = 0; i < csc->chunks[0].length_dw; i++) {
fprintf(stderr, "0x%08X\n", csc->buf[i]);
}
} else {
fprintf(stderr, "radeon: The kernel rejected CS, "
"see dmesg for more information (%i).\n", r);
}
}
for (i = 0; i < csc->num_relocs; i++)
p_atomic_dec(&csc->relocs_bo[i].bo->num_active_ioctls);
for (i = 0; i < csc->num_slab_buffers; i++)
p_atomic_dec(&csc->slab_buffers[i].bo->num_active_ioctls);
radeon_cs_context_cleanup(csc);
}
static void radeon_destroy_cs_context(struct radeon_cs_context *csc)
{
radeon_cs_context_cleanup(csc);
FREE(csc->slab_buffers);
FREE(csc->relocs_bo);
FREE(csc->relocs);
}
static void radeon_drm_cs_flush(struct radeon_winsys_cs *rcs, unsigned flags)
{
struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
struct radeon_cs_context *tmp;
if (rcs->cdw > RADEON_MAX_CMDBUF_DWORDS) {
fprintf(stderr, "radeon: command stream overflowed\n");
}
radeon_drm_cs_sync_flush(cs);
/* Flip command streams. */
tmp = cs->csc;
cs->csc = cs->cst;
cs->cst = tmp;
/* If the CS is not empty or overflowed, emit it in a separate thread. */
if (cs->base.cdw && cs->base.cdw <= RADEON_MAX_CMDBUF_DWORDS) {
unsigned i, crelocs = cs->cst->crelocs;
cs->cst->chunks[0].length_dw = cs->base.cdw;
for (i = 0; i < crelocs; i++) {
/* Update the number of active asynchronous CS ioctls for the buffer. */
p_atomic_inc(&cs->cst->relocs_bo[i]->num_active_ioctls);
}
cs->cst->flags[0] = 0;
cs->cst->flags[1] = RADEON_CS_RING_GFX;
cs->cst->cs.num_chunks = 2;
if (flags & RADEON_FLUSH_KEEP_TILING_FLAGS) {
cs->cst->flags[0] |= RADEON_CS_KEEP_TILING_FLAGS;
cs->cst->cs.num_chunks = 3;
}
if (cs->ws->info.r600_virtual_address) {
cs->cst->flags[0] |= RADEON_CS_USE_VM;
cs->cst->cs.num_chunks = 3;
}
if (flags & RADEON_FLUSH_COMPUTE) {
cs->cst->flags[1] = RADEON_CS_RING_COMPUTE;
cs->cst->cs.num_chunks = 3;
}
if (cs->thread &&
(flags & RADEON_FLUSH_ASYNC)) {
cs->flush_started = 1;
pipe_semaphore_signal(&cs->flush_queued);
} else {
radeon_drm_cs_emit_ioctl_oneshot(cs->cst);
}
} else {
radeon_cs_context_cleanup(cs->cst);
}
/* Prepare a new CS. */
cs->base.buf = cs->csc->buf;
cs->base.cdw = 0;
}
static void radeon_drm_cs_destroy(struct radeon_winsys_cs *rcs)
{
struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
radeon_drm_cs_sync_flush(cs);
if (cs->thread) {
cs->kill_thread = 1;
pipe_semaphore_signal(&cs->flush_queued);
pipe_semaphore_wait(&cs->flush_completed);
pipe_thread_wait(cs->thread);
}
pipe_semaphore_destroy(&cs->flush_queued);
pipe_semaphore_destroy(&cs->flush_completed);
radeon_cs_context_cleanup(&cs->csc1);
radeon_cs_context_cleanup(&cs->csc2);
p_atomic_dec(&cs->ws->num_cs);
radeon_destroy_cs_context(&cs->csc1);
radeon_destroy_cs_context(&cs->csc2);
FREE(cs);
}
static void radeon_drm_cs_flush(struct radeon_winsys_cs *rcs, unsigned flags)
{
struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
struct radeon_cs_context *tmp;
radeon_drm_cs_sync_flush(cs);
/* Flip command streams. */
tmp = cs->csc;
cs->csc = cs->cst;
cs->cst = tmp;
/* If the CS is not empty, emit it in a separate thread. */
if (cs->base.cdw) {
unsigned i, crelocs = cs->cst->crelocs;
cs->cst->chunks[0].length_dw = cs->base.cdw;
for (i = 0; i < crelocs; i++) {
/* Update the number of active asynchronous CS ioctls for the buffer. */
p_atomic_inc(&cs->cst->relocs_bo[i]->num_active_ioctls);
}
if (flags & RADEON_FLUSH_KEEP_TILING_FLAGS) {
cs->cst->cs.num_chunks = 3;
cs->cst->flags = RADEON_CS_KEEP_TILING_FLAGS;
} else {
cs->cst->cs.num_chunks = 2;
}
if (cs->thread &&
(flags & RADEON_FLUSH_ASYNC)) {
cs->flush_started = 1;
pipe_semaphore_signal(&cs->flush_queued);
} else {
radeon_drm_cs_emit_ioctl_oneshot(cs->cst);
}
} else {
radeon_cs_context_cleanup(cs->cst);
}
/* Prepare a new CS. */
cs->base.buf = cs->csc->buf;
cs->base.cdw = 0;
}
static void radeon_drm_cs_flush(struct radeon_winsys_cs *rcs, unsigned flags, uint32_t cs_trace_id)
{
struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
struct radeon_cs_context *tmp;
switch (cs->base.ring_type) {
case RING_DMA:
/* pad DMA ring to 8 DWs */
if (cs->ws->info.chip_class <= SI) {
while (rcs->cdw & 7)
OUT_CS(&cs->base, 0xf0000000); /* NOP packet */
} else {
while (rcs->cdw & 7)
OUT_CS(&cs->base, 0x00000000); /* NOP packet */
}
break;
case RING_GFX:
/* pad DMA ring to 8 DWs to meet CP fetch alignment requirements
* r6xx, requires at least 4 dw alignment to avoid a hw bug.
*/
if (flags & RADEON_FLUSH_COMPUTE) {
if (cs->ws->info.chip_class <= SI) {
while (rcs->cdw & 7)
OUT_CS(&cs->base, 0x80000000); /* type2 nop packet */
} else {
while (rcs->cdw & 7)
OUT_CS(&cs->base, 0xffff1000); /* type3 nop packet */
}
} else {
while (rcs->cdw & 7)
OUT_CS(&cs->base, 0x80000000); /* type2 nop packet */
}
break;
}
if (rcs->cdw > RADEON_MAX_CMDBUF_DWORDS) {
fprintf(stderr, "radeon: command stream overflowed\n");
}
radeon_drm_cs_sync_flush(rcs);
/* Flip command streams. */
tmp = cs->csc;
cs->csc = cs->cst;
cs->cst = tmp;
cs->cst->cs_trace_id = cs_trace_id;
/* If the CS is not empty or overflowed, emit it in a separate thread. */
if (cs->base.cdw && cs->base.cdw <= RADEON_MAX_CMDBUF_DWORDS && !debug_get_option_noop()) {
unsigned i, crelocs = cs->cst->crelocs;
cs->cst->chunks[0].length_dw = cs->base.cdw;
for (i = 0; i < crelocs; i++) {
/* Update the number of active asynchronous CS ioctls for the buffer. */
p_atomic_inc(&cs->cst->relocs_bo[i]->num_active_ioctls);
}
switch (cs->base.ring_type) {
case RING_DMA:
cs->cst->flags[0] = 0;
cs->cst->flags[1] = RADEON_CS_RING_DMA;
cs->cst->cs.num_chunks = 3;
if (cs->ws->info.r600_virtual_address) {
cs->cst->flags[0] |= RADEON_CS_USE_VM;
}
break;
case RING_UVD:
cs->cst->flags[0] = 0;
cs->cst->flags[1] = RADEON_CS_RING_UVD;
cs->cst->cs.num_chunks = 3;
break;
default:
case RING_GFX:
cs->cst->flags[0] = 0;
cs->cst->flags[1] = RADEON_CS_RING_GFX;
cs->cst->cs.num_chunks = 2;
if (flags & RADEON_FLUSH_KEEP_TILING_FLAGS) {
cs->cst->flags[0] |= RADEON_CS_KEEP_TILING_FLAGS;
cs->cst->cs.num_chunks = 3;
}
if (cs->ws->info.r600_virtual_address) {
cs->cst->flags[0] |= RADEON_CS_USE_VM;
cs->cst->cs.num_chunks = 3;
}
if (flags & RADEON_FLUSH_END_OF_FRAME) {
cs->cst->flags[0] |= RADEON_CS_END_OF_FRAME;
cs->cst->cs.num_chunks = 3;
}
if (flags & RADEON_FLUSH_COMPUTE) {
cs->cst->flags[1] = RADEON_CS_RING_COMPUTE;
cs->cst->cs.num_chunks = 3;
}
break;
}
if (cs->ws->thread && (flags & RADEON_FLUSH_ASYNC)) {
cs->flush_started = 1;
radeon_drm_ws_queue_cs(cs->ws, cs);
} else {
pipe_mutex_lock(cs->ws->cs_stack_lock);
if (cs->ws->thread) {
while (p_atomic_read(&cs->ws->ncs)) {
pipe_condvar_wait(cs->ws->cs_queue_empty, cs->ws->cs_stack_lock);
}
}
pipe_mutex_unlock(cs->ws->cs_stack_lock);
radeon_drm_cs_emit_ioctl_oneshot(cs, cs->cst);
}
} else {
radeon_cs_context_cleanup(cs->cst);
}
/* Prepare a new CS. */
cs->base.buf = cs->csc->buf;
cs->base.cdw = 0;
}
static int radeon_drm_cs_flush(struct radeon_cmdbuf *rcs,
unsigned flags,
struct pipe_fence_handle **pfence)
{
struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
struct radeon_cs_context *tmp;
switch (cs->ring_type) {
case RING_DMA:
/* pad DMA ring to 8 DWs */
if (cs->ws->info.chip_class <= GFX6) {
while (rcs->current.cdw & 7)
radeon_emit(&cs->base, 0xf0000000); /* NOP packet */
} else {
while (rcs->current.cdw & 7)
radeon_emit(&cs->base, 0x00000000); /* NOP packet */
}
break;
case RING_GFX:
/* pad GFX ring to 8 DWs to meet CP fetch alignment requirements
* r6xx, requires at least 4 dw alignment to avoid a hw bug.
*/
if (cs->ws->info.gfx_ib_pad_with_type2) {
while (rcs->current.cdw & 7)
radeon_emit(&cs->base, 0x80000000); /* type2 nop packet */
} else {
while (rcs->current.cdw & 7)
radeon_emit(&cs->base, 0xffff1000); /* type3 nop packet */
}
break;
case RING_UVD:
while (rcs->current.cdw & 15)
radeon_emit(&cs->base, 0x80000000); /* type2 nop packet */
break;
default:
break;
}
if (rcs->current.cdw > rcs->current.max_dw) {
fprintf(stderr, "radeon: command stream overflowed\n");
}
if (pfence || cs->csc->num_slab_buffers) {
struct pipe_fence_handle *fence;
if (cs->next_fence) {
fence = cs->next_fence;
cs->next_fence = NULL;
} else {
fence = radeon_cs_create_fence(rcs);
}
if (fence) {
if (pfence)
radeon_fence_reference(pfence, fence);
mtx_lock(&cs->ws->bo_fence_lock);
for (unsigned i = 0; i < cs->csc->num_slab_buffers; ++i) {
struct radeon_bo *bo = cs->csc->slab_buffers[i].bo;
p_atomic_inc(&bo->num_active_ioctls);
radeon_bo_slab_fence(bo, (struct radeon_bo *)fence);
}
mtx_unlock(&cs->ws->bo_fence_lock);
radeon_fence_reference(&fence, NULL);
}
} else {
radeon_fence_reference(&cs->next_fence, NULL);
}
radeon_drm_cs_sync_flush(rcs);
/* Swap command streams. */
tmp = cs->csc;
cs->csc = cs->cst;
cs->cst = tmp;
/* If the CS is not empty or overflowed, emit it in a separate thread. */
if (cs->base.current.cdw && cs->base.current.cdw <= cs->base.current.max_dw && !debug_get_option_noop()) {
unsigned i, num_relocs;
num_relocs = cs->cst->num_relocs;
cs->cst->chunks[0].length_dw = cs->base.current.cdw;
for (i = 0; i < num_relocs; i++) {
/* Update the number of active asynchronous CS ioctls for the buffer. */
p_atomic_inc(&cs->cst->relocs_bo[i].bo->num_active_ioctls);
}
switch (cs->ring_type) {
case RING_DMA:
cs->cst->flags[0] = 0;
cs->cst->flags[1] = RADEON_CS_RING_DMA;
cs->cst->cs.num_chunks = 3;
if (cs->ws->info.r600_has_virtual_memory) {
cs->cst->flags[0] |= RADEON_CS_USE_VM;
}
break;
case RING_UVD:
cs->cst->flags[0] = 0;
cs->cst->flags[1] = RADEON_CS_RING_UVD;
cs->cst->cs.num_chunks = 3;
break;
case RING_VCE:
cs->cst->flags[0] = 0;
cs->cst->flags[1] = RADEON_CS_RING_VCE;
cs->cst->cs.num_chunks = 3;
break;
default:
case RING_GFX:
case RING_COMPUTE:
cs->cst->flags[0] = RADEON_CS_KEEP_TILING_FLAGS;
cs->cst->flags[1] = RADEON_CS_RING_GFX;
cs->cst->cs.num_chunks = 3;
if (cs->ws->info.r600_has_virtual_memory) {
cs->cst->flags[0] |= RADEON_CS_USE_VM;
cs->cst->cs.num_chunks = 3;
}
if (flags & PIPE_FLUSH_END_OF_FRAME) {
cs->cst->flags[0] |= RADEON_CS_END_OF_FRAME;
cs->cst->cs.num_chunks = 3;
}
if (cs->ring_type == RING_COMPUTE) {
cs->cst->flags[1] = RADEON_CS_RING_COMPUTE;
cs->cst->cs.num_chunks = 3;
}
break;
}
if (util_queue_is_initialized(&cs->ws->cs_queue)) {
util_queue_add_job(&cs->ws->cs_queue, cs, &cs->flush_completed,
radeon_drm_cs_emit_ioctl_oneshot, NULL);
if (!(flags & PIPE_FLUSH_ASYNC))
radeon_drm_cs_sync_flush(rcs);
} else {
radeon_drm_cs_emit_ioctl_oneshot(cs, 0);
}
} else {
radeon_cs_context_cleanup(cs->cst);
}
/* Prepare a new CS. */
cs->base.current.buf = cs->csc->buf;
cs->base.current.cdw = 0;
cs->base.used_vram = 0;
cs->base.used_gart = 0;
if (cs->ring_type == RING_GFX)
cs->ws->num_gfx_IBs++;
else if (cs->ring_type == RING_DMA)
cs->ws->num_sdma_IBs++;
return 0;
}
static void radeon_drm_cs_flush(struct radeon_winsys_cs *rcs,
unsigned flags,
struct pipe_fence_handle **fence,
uint32_t cs_trace_id)
{
struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
struct radeon_cs_context *tmp;
switch (cs->base.ring_type) {
case RING_DMA:
/* pad DMA ring to 8 DWs */
if (cs->ws->info.chip_class <= SI) {
while (rcs->cdw & 7)
OUT_CS(&cs->base, 0xf0000000); /* NOP packet */
} else {
while (rcs->cdw & 7)
OUT_CS(&cs->base, 0x00000000); /* NOP packet */
}
break;
case RING_GFX:
/* pad DMA ring to 8 DWs to meet CP fetch alignment requirements
* r6xx, requires at least 4 dw alignment to avoid a hw bug.
* hawaii with old firmware needs type2 nop packet.
* accel_working2 with value 3 indicates the new firmware.
*/
if (cs->ws->info.chip_class <= SI ||
(cs->ws->info.family == CHIP_HAWAII &&
cs->ws->accel_working2 < 3)) {
while (rcs->cdw & 7)
OUT_CS(&cs->base, 0x80000000); /* type2 nop packet */
} else {
while (rcs->cdw & 7)
OUT_CS(&cs->base, 0xffff1000); /* type3 nop packet */
}
break;
case RING_UVD:
while (rcs->cdw & 15)
OUT_CS(&cs->base, 0x80000000); /* type2 nop packet */
break;
default:
break;
}
if (rcs->cdw > RADEON_MAX_CMDBUF_DWORDS) {
fprintf(stderr, "radeon: command stream overflowed\n");
}
if (fence) {
radeon_fence_reference(fence, NULL);
*fence = radeon_cs_create_fence(rcs);
}
radeon_drm_cs_sync_flush(rcs);
/* Swap command streams. */
tmp = cs->csc;
cs->csc = cs->cst;
cs->cst = tmp;
cs->cst->cs_trace_id = cs_trace_id;
/* If the CS is not empty or overflowed, emit it in a separate thread. */
if (cs->base.cdw && cs->base.cdw <= RADEON_MAX_CMDBUF_DWORDS && !debug_get_option_noop()) {
unsigned i, crelocs;
crelocs = cs->cst->crelocs;
cs->cst->chunks[0].length_dw = cs->base.cdw;
for (i = 0; i < crelocs; i++) {
/* Update the number of active asynchronous CS ioctls for the buffer. */
p_atomic_inc(&cs->cst->relocs_bo[i]->num_active_ioctls);
}
switch (cs->base.ring_type) {
case RING_DMA:
cs->cst->flags[0] = 0;
cs->cst->flags[1] = RADEON_CS_RING_DMA;
cs->cst->cs.num_chunks = 3;
if (cs->ws->info.r600_virtual_address) {
cs->cst->flags[0] |= RADEON_CS_USE_VM;
}
break;
case RING_UVD:
cs->cst->flags[0] = 0;
cs->cst->flags[1] = RADEON_CS_RING_UVD;
cs->cst->cs.num_chunks = 3;
break;
case RING_VCE:
cs->cst->flags[0] = 0;
cs->cst->flags[1] = RADEON_CS_RING_VCE;
cs->cst->cs.num_chunks = 3;
break;
default:
case RING_GFX:
cs->cst->flags[0] = 0;
cs->cst->flags[1] = RADEON_CS_RING_GFX;
cs->cst->cs.num_chunks = 2;
if (flags & RADEON_FLUSH_KEEP_TILING_FLAGS) {
cs->cst->flags[0] |= RADEON_CS_KEEP_TILING_FLAGS;
cs->cst->cs.num_chunks = 3;
}
if (cs->ws->info.r600_virtual_address) {
cs->cst->flags[0] |= RADEON_CS_USE_VM;
cs->cst->cs.num_chunks = 3;
}
if (flags & RADEON_FLUSH_END_OF_FRAME) {
cs->cst->flags[0] |= RADEON_CS_END_OF_FRAME;
cs->cst->cs.num_chunks = 3;
}
if (flags & RADEON_FLUSH_COMPUTE) {
cs->cst->flags[1] = RADEON_CS_RING_COMPUTE;
cs->cst->cs.num_chunks = 3;
}
break;
}
if (cs->ws->thread) {
pipe_semaphore_wait(&cs->flush_completed);
radeon_drm_ws_queue_cs(cs->ws, cs);
if (!(flags & RADEON_FLUSH_ASYNC))
radeon_drm_cs_sync_flush(rcs);
} else {
radeon_drm_cs_emit_ioctl_oneshot(cs, cs->cst);
}
} else {
radeon_cs_context_cleanup(cs->cst);
}
/* Prepare a new CS. */
cs->base.buf = cs->csc->buf;
cs->base.cdw = 0;
cs->ws->num_cs_flushes++;
}