/**
* cik_sdma_ring_test_ring - simple async dma engine test
*
* @ring: amdgpu_ring structure holding ring information
*
* Test the DMA engine by writing using it to write an
* value to memory. (CIK).
* Returns 0 for success, error for failure.
*/
static int cik_sdma_ring_test_ring(struct amdgpu_ring *ring)
{
struct amdgpu_device *adev = ring->adev;
unsigned i;
unsigned index;
int r;
u32 tmp;
u64 gpu_addr;
r = amdgpu_wb_get(adev, &index);
if (r) {
dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r);
return r;
}
gpu_addr = adev->wb.gpu_addr + (index * 4);
tmp = 0xCAFEDEAD;
adev->wb.wb[index] = cpu_to_le32(tmp);
r = amdgpu_ring_lock(ring, 5);
if (r) {
DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r);
amdgpu_wb_free(adev, index);
return r;
}
amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_WRITE, SDMA_WRITE_SUB_OPCODE_LINEAR, 0));
amdgpu_ring_write(ring, lower_32_bits(gpu_addr));
amdgpu_ring_write(ring, upper_32_bits(gpu_addr));
amdgpu_ring_write(ring, 1); /* number of DWs to follow */
amdgpu_ring_write(ring, 0xDEADBEEF);
amdgpu_ring_unlock_commit(ring);
for (i = 0; i < adev->usec_timeout; i++) {
tmp = le32_to_cpu(adev->wb.wb[index]);
if (tmp == 0xDEADBEEF)
break;
DRM_UDELAY(1);
}
if (i < adev->usec_timeout) {
DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
} else {
DRM_ERROR("amdgpu: ring %d test failed (0x%08X)\n",
ring->idx, tmp);
r = -EINVAL;
}
amdgpu_wb_free(adev, index);
return r;
}
static int amdgpu_test_create_and_emit_fence(struct amdgpu_device *adev,
struct amdgpu_ring *ring,
struct fence **fence)
{
uint32_t handle = ring->idx ^ 0xdeafbeef;
int r;
if (ring == &adev->uvd.ring) {
r = amdgpu_uvd_get_create_msg(ring, handle, NULL);
if (r) {
DRM_ERROR("Failed to get dummy create msg\n");
return r;
}
r = amdgpu_uvd_get_destroy_msg(ring, handle, fence);
if (r) {
DRM_ERROR("Failed to get dummy destroy msg\n");
return r;
}
} else if (ring == &adev->vce.ring[0] ||
ring == &adev->vce.ring[1]) {
r = amdgpu_vce_get_create_msg(ring, handle, NULL);
if (r) {
DRM_ERROR("Failed to get dummy create msg\n");
return r;
}
r = amdgpu_vce_get_destroy_msg(ring, handle, fence);
if (r) {
DRM_ERROR("Failed to get dummy destroy msg\n");
return r;
}
} else {
struct amdgpu_fence *a_fence = NULL;
r = amdgpu_ring_lock(ring, 64);
if (r) {
DRM_ERROR("Failed to lock ring A %d\n", ring->idx);
return r;
}
amdgpu_fence_emit(ring, AMDGPU_FENCE_OWNER_UNDEFINED, &a_fence);
amdgpu_ring_unlock_commit(ring);
*fence = &a_fence->base;
}
return 0;
}
void amdgpu_test_ring_sync(struct amdgpu_device *adev,
struct amdgpu_ring *ringA,
struct amdgpu_ring *ringB)
{
struct amdgpu_fence *fence1 = NULL, *fence2 = NULL;
struct amdgpu_semaphore *semaphore = NULL;
int r;
r = amdgpu_semaphore_create(adev, &semaphore);
if (r) {
DRM_ERROR("Failed to create semaphore\n");
goto out_cleanup;
}
r = amdgpu_ring_lock(ringA, 64);
if (r) {
DRM_ERROR("Failed to lock ring A %d\n
/**
* amdgpu_ib_schedule - schedule an IB (Indirect Buffer) on the ring
*
* @adev: amdgpu_device pointer
* @num_ibs: number of IBs to schedule
* @ibs: IB objects to schedule
* @owner: owner for creating the fences
*
* Schedule an IB on the associated ring (all asics).
* Returns 0 on success, error on failure.
*
* On SI, there are two parallel engines fed from the primary ring,
* the CE (Constant Engine) and the DE (Drawing Engine). Since
* resource descriptors have moved to memory, the CE allows you to
* prime the caches while the DE is updating register state so that
* the resource descriptors will be already in cache when the draw is
* processed. To accomplish this, the userspace driver submits two
* IBs, one for the CE and one for the DE. If there is a CE IB (called
* a CONST_IB), it will be put on the ring prior to the DE IB. Prior
* to SI there was just a DE IB.
*/
int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs,
struct amdgpu_ib *ibs, void *owner)
{
struct amdgpu_ib *ib = &ibs[0];
struct amdgpu_ring *ring;
struct amdgpu_ctx *ctx, *old_ctx;
struct amdgpu_vm *vm;
unsigned i;
int r = 0;
if (num_ibs == 0)
return -EINVAL;
ring = ibs->ring;
ctx = ibs->ctx;
vm = ibs->vm;
if (!ring->ready) {
dev_err(adev->dev, "couldn't schedule ib\n");
return -EINVAL;
}
r = amdgpu_sync_wait(&ibs->sync);
if (r) {
dev_err(adev->dev, "IB sync failed (%d).\n", r);
return r;
}
r = amdgpu_ring_lock(ring, (256 + AMDGPU_NUM_SYNCS * 8) * num_ibs);
if (r) {
dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
return r;
}
if (vm) {
/* grab a vm id if necessary */
r = amdgpu_vm_grab_id(ibs->vm, ibs->ring, &ibs->sync);
if (r) {
amdgpu_ring_unlock_undo(ring);
return r;
}
}
r = amdgpu_sync_rings(&ibs->sync, ring);
if (r) {
amdgpu_ring_unlock_undo(ring);
dev_err(adev->dev, "failed to sync rings (%d)\n", r);
return r;
}
if (vm) {
/* do context switch */
amdgpu_vm_flush(ring, vm, ib->sync.last_vm_update);
if (ring->funcs->emit_gds_switch)
amdgpu_ring_emit_gds_switch(ring, ib->vm->ids[ring->idx].id,
ib->gds_base, ib->gds_size,
ib->gws_base, ib->gws_size,
ib->oa_base, ib->oa_size);
if (ring->funcs->emit_hdp_flush)
amdgpu_ring_emit_hdp_flush(ring);
}
old_ctx = ring->current_ctx;
for (i = 0; i < num_ibs; ++i) {
ib = &ibs[i];
if (ib->ring != ring || ib->ctx != ctx || ib->vm != vm) {
ring->current_ctx = old_ctx;
amdgpu_ring_unlock_undo(ring);
return -EINVAL;
}
amdgpu_ring_emit_ib(ring, ib);
ring->current_ctx = ctx;
}
r = amdgpu_fence_emit(ring, owner, &ib->fence);
if (r) {
dev_err(adev->dev, "failed to emit fence (%d)\n", r);
ring->current_ctx = old_ctx;
amdgpu_ring_unlock_undo(ring);
return r;
}
if (!amdgpu_enable_scheduler && ib->ctx)
ib->sequence = amdgpu_ctx_add_fence(ib->ctx, ring,
&ib->fence->base);
/* wrap the last IB with fence */
if (ib->user) {
uint64_t addr = amdgpu_bo_gpu_offset(ib->user->bo);
addr += ib->user->offset;
amdgpu_ring_emit_fence(ring, addr, ib->sequence,
AMDGPU_FENCE_FLAG_64BIT);
}
if (ib->vm)
amdgpu_vm_fence(adev, ib->vm, &ib->fence->base);
amdgpu_ring_unlock_commit(ring);
return 0;
}
static void amdgpu_test_ring_sync2(struct amdgpu_device *adev,
struct amdgpu_ring *ringA,
struct amdgpu_ring *ringB,
struct amdgpu_ring *ringC)
{
struct fence *fenceA = NULL, *fenceB = NULL;
struct amdgpu_semaphore *semaphore = NULL;
bool sigA, sigB;
int i, r;
r = amdgpu_semaphore_create(adev, &semaphore);
if (r) {
DRM_ERROR("Failed to create semaphore\n");
goto out_cleanup;
}
r = amdgpu_ring_lock(ringA, 64);
if (r) {
DRM_ERROR("Failed to lock ring A %d\n", ringA->idx);
goto out_cleanup;
}
amdgpu_semaphore_emit_wait(ringA, semaphore);
amdgpu_ring_unlock_commit(ringA);
r = amdgpu_test_create_and_emit_fence(adev, ringA, &fenceA);
if (r)
goto out_cleanup;
r = amdgpu_ring_lock(ringB, 64);
if (r) {
DRM_ERROR("Failed to lock ring B %d\n", ringB->idx);
goto out_cleanup;
}
amdgpu_semaphore_emit_wait(ringB, semaphore);
amdgpu_ring_unlock_commit(ringB);
r = amdgpu_test_create_and_emit_fence(adev, ringB, &fenceB);
if (r)
goto out_cleanup;
mdelay(1000);
if (fence_is_signaled(fenceA)) {
DRM_ERROR("Fence A signaled without waiting for semaphore.\n");
goto out_cleanup;
}
if (fence_is_signaled(fenceB)) {
DRM_ERROR("Fence B signaled without waiting for semaphore.\n");
goto out_cleanup;
}
r = amdgpu_ring_lock(ringC, 64);
if (r) {
DRM_ERROR("Failed to lock ring B %p\n", ringC);
goto out_cleanup;
}
amdgpu_semaphore_emit_signal(ringC, semaphore);
amdgpu_ring_unlock_commit(ringC);
for (i = 0; i < 30; ++i) {
mdelay(100);
sigA = fence_is_signaled(fenceA);
sigB = fence_is_signaled(fenceB);
if (sigA || sigB)
break;
}
if (!sigA && !sigB) {
DRM_ERROR("Neither fence A nor B has been signaled\n");
goto out_cleanup;
} else if (sigA && sigB) {
DRM_ERROR("Both fence A and B has been signaled\n");
goto out_cleanup;
}
DRM_INFO("Fence %c was first signaled\n", sigA ? 'A' : 'B');
r = amdgpu_ring_lock(ringC, 64);
if (r) {
DRM_ERROR("Failed to lock ring B %p\n", ringC);
goto out_cleanup;
}
amdgpu_semaphore_emit_signal(ringC, semaphore);
amdgpu_ring_unlock_commit(ringC);
mdelay(1000);
r = fence_wait(fenceA, false);
if (r) {
DRM_ERROR("Failed to wait for sync fence A\n");
goto out_cleanup;
}
r = fence_wait(fenceB, false);
if (r) {
DRM_ERROR("Failed to wait for sync fence B\n");
goto out_cleanup;
}
out_cleanup:
amdgpu_semaphore_free(adev, &semaphore, NULL);
if (fenceA)
fence_put(fenceA);
if (fenceB)
fence_put(fenceB);
if (r)
printk(KERN_WARNING "Error while testing ring sync (%d).\n", r);
}
void amdgpu_test_ring_sync(struct amdgpu_device *adev,
struct amdgpu_ring *ringA,
struct amdgpu_ring *ringB)
{
struct fence *fence1 = NULL, *fence2 = NULL;
struct amdgpu_semaphore *semaphore = NULL;
int r;
r = amdgpu_semaphore_create(adev, &semaphore);
if (r) {
DRM_ERROR("Failed to create semaphore\n");
goto out_cleanup;
}
r = amdgpu_ring_lock(ringA, 64);
if (r) {
DRM_ERROR("Failed to lock ring A %d\n", ringA->idx);
goto out_cleanup;
}
amdgpu_semaphore_emit_wait(ringA, semaphore);
amdgpu_ring_unlock_commit(ringA);
r = amdgpu_test_create_and_emit_fence(adev, ringA, &fence1);
if (r)
goto out_cleanup;
r = amdgpu_ring_lock(ringA, 64);
if (r) {
DRM_ERROR("Failed to lock ring A %d\n", ringA->idx);
goto out_cleanup;
}
amdgpu_semaphore_emit_wait(ringA, semaphore);
amdgpu_ring_unlock_commit(ringA);
r = amdgpu_test_create_and_emit_fence(adev, ringA, &fence2);
if (r)
goto out_cleanup;
mdelay(1000);
if (fence_is_signaled(fence1)) {
DRM_ERROR("Fence 1 signaled without waiting for semaphore.\n");
goto out_cleanup;
}
r = amdgpu_ring_lock(ringB, 64);
if (r) {
DRM_ERROR("Failed to lock ring B %p\n", ringB);
goto out_cleanup;
}
amdgpu_semaphore_emit_signal(ringB, semaphore);
amdgpu_ring_unlock_commit(ringB);
r = fence_wait(fence1, false);
if (r) {
DRM_ERROR("Failed to wait for sync fence 1\n");
goto out_cleanup;
}
mdelay(1000);
if (fence_is_signaled(fence2)) {
DRM_ERROR("Fence 2 signaled without waiting for semaphore.\n");
goto out_cleanup;
}
r = amdgpu_ring_lock(ringB, 64);
if (r) {
DRM_ERROR("Failed to lock ring B %p\n", ringB);
goto out_cleanup;
}
amdgpu_semaphore_emit_signal(ringB, semaphore);
amdgpu_ring_unlock_commit(ringB);
r = fence_wait(fence2, false);
if (r) {
DRM_ERROR("Failed to wait for sync fence 1\n");
goto out_cleanup;
}
out_cleanup:
amdgpu_semaphore_free(adev, &semaphore, NULL);
if (fence1)
fence_put(fence1);
if (fence2)
fence_put(fence2);
if (r)
printk(KERN_WARNING "Error while testing ring sync (%d).\n", r);
}
