diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c')
-rw-r--r-- | drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c | 31 |
1 files changed, 25 insertions, 6 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c b/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c index 8493bfbbc148..bd715012185c 100644 --- a/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c +++ b/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c @@ -382,8 +382,27 @@ static void sdma_v5_0_ring_emit_ib(struct amdgpu_ring *ring, unsigned vmid = AMDGPU_JOB_GET_VMID(job); uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid); - /* IB packet must end on a 8 DW boundary */ - sdma_v5_0_ring_insert_nop(ring, (10 - (lower_32_bits(ring->wptr) & 7)) % 8); + /* Invalidate L2, because if we don't do it, we might get stale cache + * lines from previous IBs. + */ + amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ)); + amdgpu_ring_write(ring, 0); + amdgpu_ring_write(ring, (SDMA_GCR_GL2_INV | + SDMA_GCR_GL2_WB | + SDMA_GCR_GLM_INV | + SDMA_GCR_GLM_WB) << 16); + amdgpu_ring_write(ring, 0xffffff80); + amdgpu_ring_write(ring, 0xffff); + + /* An IB packet must end on a 8 DW boundary--the next dword + * must be on a 8-dword boundary. Our IB packet below is 6 + * dwords long, thus add x number of NOPs, such that, in + * modular arithmetic, + * wptr + 6 + x = 8k, k >= 0, which in C is, + * (wptr + 6 + x) % 8 = 0. + * The expression below, is a solution of x. + */ + sdma_v5_0_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7); amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) | SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf)); @@ -1086,10 +1105,10 @@ static void sdma_v5_0_vm_set_pte_pde(struct amdgpu_ib *ib, } /** - * sdma_v5_0_ring_pad_ib - pad the IB to the required number of dw - * + * sdma_v5_0_ring_pad_ib - pad the IB * @ib: indirect buffer to fill with padding * + * Pad the IB with NOPs to a boundary multiple of 8. */ static void sdma_v5_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib) { @@ -1097,7 +1116,7 @@ static void sdma_v5_0_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib u32 pad_count; int i; - pad_count = (8 - (ib->length_dw & 0x7)) % 8; + pad_count = (-ib->length_dw) & 0x7; for (i = 0; i < pad_count; i++) if (sdma && sdma->burst_nop && (i == 0)) ib->ptr[ib->length_dw++] = @@ -1600,7 +1619,7 @@ static const struct amdgpu_ring_funcs sdma_v5_0_ring_funcs = { SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 + SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 * 2 + 10 + 10 + 10, /* sdma_v5_0_ring_emit_fence x3 for user fence, vm fence */ - .emit_ib_size = 7 + 6, /* sdma_v5_0_ring_emit_ib */ + .emit_ib_size = 5 + 7 + 6, /* sdma_v5_0_ring_emit_ib */ .emit_ib = sdma_v5_0_ring_emit_ib, .emit_fence = sdma_v5_0_ring_emit_fence, .emit_pipeline_sync = sdma_v5_0_ring_emit_pipeline_sync, |