From 2b10cb58d7a3fd621ec9b2ba765a092e562ef998 Mon Sep 17 00:00:00 2001 From: David Rosca Date: Mon, 18 Aug 2025 09:06:58 +0200 Subject: drm/amdgpu/vcn4: Fix IB parsing with multiple engine info packages There can be multiple engine info packages in one IB and the first one may be common engine, not decode/encode. We need to parse the entire IB instead of stopping after finding first engine info. Signed-off-by: David Rosca Reviewed-by: Leo Liu Signed-off-by: Alex Deucher (cherry picked from commit dc8f9f0f45166a6b37864e7a031c726981d6e5fc) Cc: stable@vger.kernel.org --- drivers/gpu/drm/amd/amdgpu/vcn_v4_0.c | 52 ++++++++++++++--------------------- 1 file changed, 21 insertions(+), 31 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/vcn_v4_0.c') diff --git a/drivers/gpu/drm/amd/amdgpu/vcn_v4_0.c b/drivers/gpu/drm/amd/amdgpu/vcn_v4_0.c index 1924e075b66f..311fb595bd69 100644 --- a/drivers/gpu/drm/amd/amdgpu/vcn_v4_0.c +++ b/drivers/gpu/drm/amd/amdgpu/vcn_v4_0.c @@ -1907,22 +1907,16 @@ out: #define RADEON_VCN_ENGINE_TYPE_ENCODE (0x00000002) #define RADEON_VCN_ENGINE_TYPE_DECODE (0x00000003) - #define RADEON_VCN_ENGINE_INFO (0x30000001) -#define RADEON_VCN_ENGINE_INFO_MAX_OFFSET 16 - #define RENCODE_ENCODE_STANDARD_AV1 2 #define RENCODE_IB_PARAM_SESSION_INIT 0x00000003 -#define RENCODE_IB_PARAM_SESSION_INIT_MAX_OFFSET 64 -/* return the offset in ib if id is found, -1 otherwise - * to speed up the searching we only search upto max_offset - */ -static int vcn_v4_0_enc_find_ib_param(struct amdgpu_ib *ib, uint32_t id, int max_offset) +/* return the offset in ib if id is found, -1 otherwise */ +static int vcn_v4_0_enc_find_ib_param(struct amdgpu_ib *ib, uint32_t id, int start) { int i; - for (i = 0; i < ib->length_dw && i < max_offset && ib->ptr[i] >= 8; i += ib->ptr[i]/4) { + for (i = start; i < ib->length_dw && ib->ptr[i] >= 8; i += ib->ptr[i] / 4) { if (ib->ptr[i + 1] == id) return i; } @@ -1937,33 +1931,29 @@ static int vcn_v4_0_ring_patch_cs_in_place(struct amdgpu_cs_parser *p, struct amdgpu_vcn_decode_buffer *decode_buffer; uint64_t addr; uint32_t val; - int idx; + int idx = 0, sidx; /* The first instance can decode anything */ if (!ring->me) return 0; - /* RADEON_VCN_ENGINE_INFO is at the top of ib block */ - idx = vcn_v4_0_enc_find_ib_param(ib, RADEON_VCN_ENGINE_INFO, - RADEON_VCN_ENGINE_INFO_MAX_OFFSET); - if (idx < 0) /* engine info is missing */ - return 0; - - val = amdgpu_ib_get_value(ib, idx + 2); /* RADEON_VCN_ENGINE_TYPE */ - if (val == RADEON_VCN_ENGINE_TYPE_DECODE) { - decode_buffer = (struct amdgpu_vcn_decode_buffer *)&ib->ptr[idx + 6]; - - if (!(decode_buffer->valid_buf_flag & 0x1)) - return 0; - - addr = ((u64)decode_buffer->msg_buffer_address_hi) << 32 | - decode_buffer->msg_buffer_address_lo; - return vcn_v4_0_dec_msg(p, job, addr); - } else if (val == RADEON_VCN_ENGINE_TYPE_ENCODE) { - idx = vcn_v4_0_enc_find_ib_param(ib, RENCODE_IB_PARAM_SESSION_INIT, - RENCODE_IB_PARAM_SESSION_INIT_MAX_OFFSET); - if (idx >= 0 && ib->ptr[idx + 2] == RENCODE_ENCODE_STANDARD_AV1) - return vcn_v4_0_limit_sched(p, job); + while ((idx = vcn_v4_0_enc_find_ib_param(ib, RADEON_VCN_ENGINE_INFO, idx)) >= 0) { + val = amdgpu_ib_get_value(ib, idx + 2); /* RADEON_VCN_ENGINE_TYPE */ + if (val == RADEON_VCN_ENGINE_TYPE_DECODE) { + decode_buffer = (struct amdgpu_vcn_decode_buffer *)&ib->ptr[idx + 6]; + + if (!(decode_buffer->valid_buf_flag & 0x1)) + return 0; + + addr = ((u64)decode_buffer->msg_buffer_address_hi) << 32 | + decode_buffer->msg_buffer_address_lo; + return vcn_v4_0_dec_msg(p, job, addr); + } else if (val == RADEON_VCN_ENGINE_TYPE_ENCODE) { + sidx = vcn_v4_0_enc_find_ib_param(ib, RENCODE_IB_PARAM_SESSION_INIT, idx); + if (sidx >= 0 && ib->ptr[sidx + 2] == RENCODE_ENCODE_STANDARD_AV1) + return vcn_v4_0_limit_sched(p, job); + } + idx += ib->ptr[idx] / 4; } return 0; } -- cgit v1.2.3 From 3318f2d20ce48849855df5e190813826d0bc3653 Mon Sep 17 00:00:00 2001 From: David Rosca Date: Mon, 18 Aug 2025 09:18:37 +0200 Subject: drm/amdgpu/vcn: Allow limiting ctx to instance 0 for AV1 at any time There is no reason to require this to happen on first submitted IB only. We need to wait for the queue to be idle, but it can be done at any time (including when there are multiple video sessions active). Signed-off-by: David Rosca Reviewed-by: Leo Liu Signed-off-by: Alex Deucher (cherry picked from commit 8908fdce0634a623404e9923ed2f536101a39db5) Cc: stable@vger.kernel.org --- drivers/gpu/drm/amd/amdgpu/vcn_v3_0.c | 12 ++++++++---- drivers/gpu/drm/amd/amdgpu/vcn_v4_0.c | 12 ++++++++---- 2 files changed, 16 insertions(+), 8 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/vcn_v4_0.c') diff --git a/drivers/gpu/drm/amd/amdgpu/vcn_v3_0.c b/drivers/gpu/drm/amd/amdgpu/vcn_v3_0.c index 4b8f4407047f..2811226b0ea5 100644 --- a/drivers/gpu/drm/amd/amdgpu/vcn_v3_0.c +++ b/drivers/gpu/drm/amd/amdgpu/vcn_v3_0.c @@ -1888,15 +1888,19 @@ static int vcn_v3_0_limit_sched(struct amdgpu_cs_parser *p, struct amdgpu_job *job) { struct drm_gpu_scheduler **scheds; - - /* The create msg must be in the first IB submitted */ - if (atomic_read(&job->base.entity->fence_seq)) - return -EINVAL; + struct dma_fence *fence; /* if VCN0 is harvested, we can't support AV1 */ if (p->adev->vcn.harvest_config & AMDGPU_VCN_HARVEST_VCN0) return -EINVAL; + /* wait for all jobs to finish before switching to instance 0 */ + fence = amdgpu_ctx_get_fence(p->ctx, job->base.entity, ~0ull); + if (fence) { + dma_fence_wait(fence, false); + dma_fence_put(fence); + } + scheds = p->adev->gpu_sched[AMDGPU_HW_IP_VCN_DEC] [AMDGPU_RING_PRIO_DEFAULT].sched; drm_sched_entity_modify_sched(job->base.entity, scheds, 1); diff --git a/drivers/gpu/drm/amd/amdgpu/vcn_v4_0.c b/drivers/gpu/drm/amd/amdgpu/vcn_v4_0.c index 311fb595bd69..706f3b2f484f 100644 --- a/drivers/gpu/drm/amd/amdgpu/vcn_v4_0.c +++ b/drivers/gpu/drm/amd/amdgpu/vcn_v4_0.c @@ -1808,15 +1808,19 @@ static int vcn_v4_0_limit_sched(struct amdgpu_cs_parser *p, struct amdgpu_job *job) { struct drm_gpu_scheduler **scheds; - - /* The create msg must be in the first IB submitted */ - if (atomic_read(&job->base.entity->fence_seq)) - return -EINVAL; + struct dma_fence *fence; /* if VCN0 is harvested, we can't support AV1 */ if (p->adev->vcn.harvest_config & AMDGPU_VCN_HARVEST_VCN0) return -EINVAL; + /* wait for all jobs to finish before switching to instance 0 */ + fence = amdgpu_ctx_get_fence(p->ctx, job->base.entity, ~0ull); + if (fence) { + dma_fence_wait(fence, false); + dma_fence_put(fence); + } + scheds = p->adev->gpu_sched[AMDGPU_HW_IP_VCN_ENC] [AMDGPU_RING_PRIO_0].sched; drm_sched_entity_modify_sched(job->base.entity, scheds, 1); -- cgit v1.2.3