From e8d7aa68c8fe68525d04a940ca59c1b5aae16b7e Mon Sep 17 00:00:00 2001 From: Alex Deucher Date: Tue, 20 Apr 2021 15:26:46 -0400 Subject: drm/amdgpu: Add graphics cache rinse packet for sdma 5.0 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Add emit mem sync callback for sdma_v5_0 In amdgpu sync object test, three threads created jobs to send GFX IB and SDMA IB in sequence. After the first GFX thread joined, sometimes the third thread will reuse the same physical page to store the SDMA IB. There will be a risk that SDMA will read GFX IB in the previous physical page. So it's better to flush the cache before commit sdma IB. Acked-by: Christian König Reviewed-by: Hawking Zhang Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c | 28 ++++++++++++++++++++++++++++ 1 file changed, 28 insertions(+) (limited to 'drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c') diff --git a/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c b/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c index 920fc6d4a127..d294ef6a625a 100644 --- a/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c +++ b/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c @@ -437,6 +437,33 @@ static void sdma_v5_0_ring_emit_ib(struct amdgpu_ring *ring, amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr)); } +/** + * sdma_v5_0_ring_emit_mem_sync - flush the IB by graphics cache rinse + * + * @ring: amdgpu ring pointer + * @job: job to retrieve vmid from + * @ib: IB object to schedule + * + * flush the IB by graphics cache rinse. + */ +static void sdma_v5_0_ring_emit_mem_sync(struct amdgpu_ring *ring) +{ + uint32_t gcr_cntl = + SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB | SDMA_GCR_GLM_INV | + SDMA_GCR_GL1_INV | SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV | + SDMA_GCR_GLI_INV(1); + + /* flush entire cache L0/L1/L2, this can be optimized by performance requirement */ + amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ)); + amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0)); + amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) | + SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0)); + amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) | + SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16)); + amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) | + SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0)); +} + /** * sdma_v5_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring * @@ -1643,6 +1670,7 @@ static const struct amdgpu_ring_funcs sdma_v5_0_ring_funcs = { 10 + 10 + 10, /* sdma_v5_0_ring_emit_fence x3 for user fence, vm fence */ .emit_ib_size = 5 + 7 + 6, /* sdma_v5_0_ring_emit_ib */ .emit_ib = sdma_v5_0_ring_emit_ib, + .emit_mem_sync = sdma_v5_0_ring_emit_mem_sync, .emit_fence = sdma_v5_0_ring_emit_fence, .emit_pipeline_sync = sdma_v5_0_ring_emit_pipeline_sync, .emit_vm_flush = sdma_v5_0_ring_emit_vm_flush, -- cgit v1.2.3-59-g8ed1b From 36f77e12a21eea07901bed8c6f0a170c7c7bf1b1 Mon Sep 17 00:00:00 2001 From: Alex Deucher Date: Tue, 4 May 2021 09:58:26 -0400 Subject: drm/amdgpu: drop the GCR packet from the emit_ib frame for sdma5.0 It's not needed here and has been added to the proper place in the previous patch. This aligns with what we do for sdma 5.2. Reviewed-by: Hawking Zhang Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c | 12 ------------ 1 file changed, 12 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c') diff --git a/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c b/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c index d294ef6a625a..04c68a79eca4 100644 --- a/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c +++ b/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c @@ -405,18 +405,6 @@ 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); - /* 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 -- cgit v1.2.3-59-g8ed1b