drm/amdgpu: add coordinated MEC pipe reset for GFX compute queues
Subsystems (2)
PATCH v3 1/8 drm/amdgpu: add coordinated MEC pipe reset for GFX compute queues
Introduce a shared mutex and common helpers to serialize MEC pipe reset sequences between KGD (DRM scheduler) and KFD (AMDKFD) paths. This prevents races where one path could stop/start schedulers or reprogram hardware while the other is in the middle of a pipe reset, potentially leading to queue map/unmap corruption or HQD state mismatches.
The change adds:
- mec_pipe_reset_mutex to struct amdgpu_gfx, initialized during device init.
- amdgpu_gfx_mec_pipe_reset_prepare(): stops DRM schedulers and KFD scheduling for all compute rings on a given (xcc_id, me, pipe) tuple, backing up unprocessed commands except for an optional guilty queue that is already handled via the KGD ring reset path.
- amdgpu_gfx_mec_pipe_restart_schedulers(): restarts all schedulers and KFD scheduling for the affected pipe.
- amdgpu_gfx_mec_pipe_reset_recover_queues(): re-initializes and remaps each KCQ on the pipe, optionally using a timed-out fence for the guilty queue and collateral fences for others, then completes the ring reset helper sequence.
- amdgpu_gfx_mec_pipe_reset_run(): the core orchestration routine that takes the mutex, invokes prepare, performs the HW pipe reset via either a KFD or KGD callback, restarts schedulers on error, and recovers queues.
The implementation correctly handles single and multi-XCC configurations by offsetting into the compute_ring array per partition. The special queue value AMDGPU_MEC_PIPE_RESET_NO_QUEUE allows KFD-initiated resets where no single DRM KCQ is identified as the timeout victim.
--- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 1 + drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 196 +++++++++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h | 35 ++++ 3 files changed, 232 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -0
@@ -3742,6 +3742,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,amdgpu_sync_create (&adev->isolation[i].active);amdgpu_sync_create (&adev->isolation[i].prev); } +mutex_init (&adev->gfx.mec_pipe_reset_mutex);mutex_init (&adev->gfx.userq_sch_mutex);mutex_init (&adev->gfx.workload_profile_mutex);mutex_init (&adev->vcn.workload_profile_mutex);
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +196 -0
208 lines not on this page · raw message
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h +35 -0
@@ -527,6 +527,9 @@ struct amdgpu_gfx {const void *cleaner_shader_ptr;bool enable_cleaner_shader;struct amdgpu_isolation_work enforce_isolation[MAX_XCP ]; +/* Serialize MEC pipe reset prep/HW/recover between KGD and KFD */ +struct mutex mec_pipe_reset_mutex; +/* Mutex for synchronizing KFD scheduler operations */ struct mutex userq_sch_mutex;u64 userq_sch_req_count[MAX_XCP ];
@@ -603,6 +606,38 @@ int amdgpu_gfx_mec_queue_to_bit(struct amdgpu_device *adev, int mec, int pipe, int queue);void amdgpu_queue_mask_bit_to_mec_queue (struct amdgpu_device *adev,int bit,int *mec,int *pipe,int *queue); + +/* +* Pass @queue == AMDGPU_MEC_PIPE_RESET_NO_QUEUE when no DRM KCQ is the timeout +* victim (e.g. KFD-driven pipe reset); all queues on the pipe are backed up in +* prepare and recover uses collateral fences only. +*/ +#define AMDGPU_MEC_PIPE_RESET_NO_QUEUE U32_MAX + +typedef int (*amdgpu_gfx_kcq_init_queue_t )(struct amdgpu_ring *ring,bool clear); +typedef int (*amdgpu_gfx_kcq_mec_pipe_reset_t )(struct amdgpu_device *adev, +u32 me,u32 pipe,u32 queue); +typedef int (*amdgpu_gfx_kfd_mec_pipe_reset_t )(struct amdgpu_device *adev, +u32 xcc_id,u32 me,u32 pipe); + +int amdgpu_gfx_mec_pipe_reset_run (struct amdgpu_device *adev, +u32 xcc_id,u32 me,u32 pipe,u32 queue, +struct amdgpu_fence *timedout_fence, +amdgpu_gfx_kcq_mec_pipe_reset_t kcq_pipe_reset, +amdgpu_gfx_kfd_mec_pipe_reset_t kfd_pipe_reset, +amdgpu_gfx_kcq_init_queue_t kcq_init); + +void amdgpu_gfx_mec_pipe_reset_prepare (struct amdgpu_device *adev, +u32 xcc_id,u32 me,u32 pipe, +u32 guilty_queue); +void amdgpu_gfx_mec_pipe_restart_schedulers (struct amdgpu_device *adev, +u32 me,u32 pipe,u32 xcc_id); +int amdgpu_gfx_mec_pipe_reset_recover_queues ( +struct amdgpu_device *adev, +u32 xcc_id,u32 me,u32 pipe, +u32 guilty_queue, +struct amdgpu_fence *timedout_fence, +amdgpu_gfx_kcq_init_queue_t kcq_init);bool amdgpu_gfx_is_mec_queue_enabled (struct amdgpu_device *adev,int xcc_id,int mec,int pipe,int queue); bool amdgpu_gfx_is_high_priority_compute_queue(struct amdgpu_device *adev,
Signature · 1 line
2.49.0
PATCH v3 2/8 drm/amdgpu/gfx11: Refactor compute pipe reset and add HQD cleanup
Refactor gfx_v11_0_reset_compute_pipe() to accept explicit me, pipe, and queue parameters instead of deriving them from the ring structure. This enables the function to be used in generic pipe reset flows.
Introduce gfx_v11_0_clear_hqds_on_mec_pipe() to properly clear CP_HQD_ACTIVE and CP_HQD_DEQUEUE_REQUEST for all queues on a given MEC pipe while the pipe reset is asserted, ensuring the HQDs are torn down correctly before deasserting reset.
Switch the KCQ reset path to use the common MEC pipe reset helper amdgpu_gfx_mec_pipe_reset_run(), which coordinates the reset sequence including KFD suspend/resume to avoid conflicts with user mode queues.
--- drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c | 177 +++++++++++++++---------- 1 file changed, 109 insertions(+), 68 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c +109 -68
@@ -6906,11 +6906,39 @@ static int gfx_v11_0_reset_kgq(struct amdgpu_ring *ring,return amdgpu_ring_reset_helper_end (ring, timedout_fence); } -static int gfx_v11_0_reset_compute_pipe (struct amdgpu_ring *ring) +/* +* With MEC pipe reset asserted, clear CP_HQD_ACTIVE / CP_HQD_DEQUEUE_REQUEST for +* every queue on (me, pipe). HQDs must be torn down while pipe reset stays +* asserted; only then clear the pipe reset bit. +* Caller must hold adev->srbm_mutex. +*/ +static void gfx_v11_0_clear_hqds_on_mec_pipe (struct amdgpu_device *adev,u32 me, +u32 pipe) { +unsigned int q; +int j; -struct amdgpu_device *adev = ring->adev; -uint32_t reset_pipe =0 , clean_pipe =0 ; +for (q =0 ; q < adev->gfx.mec.num_queue_per_pipe; q++) { +soc21_grbm_select (adev, me, pipe, q,0 ); +/* Start from a clean HQD dequeue state before forcing HQD inactive. */ +WREG32_SOC15 (GC ,0 , regCP_HQD_ACTIVE,0 ); +if (RREG32_SOC15 (GC ,0 , regCP_HQD_ACTIVE) &1 ) { +WREG32_SOC15 (GC ,0 , regCP_HQD_DEQUEUE_REQUEST,1 ); +for (j =0 ; j < adev->usec_timeout; j++) { +if (!(RREG32_SOC15 (GC ,0 , regCP_HQD_ACTIVE) &1 )) +break ; +udelay (1 ); + } + } + +WREG32_SOC15 (GC ,0 , regCP_HQD_DEQUEUE_REQUEST,0 ); + } +} + +static int gfx_v11_0_reset_compute_pipe (struct amdgpu_device *adev, +u32 me,u32 pipe,u32 queue) +{ +uint32_t reset_val, clean_val;int r;if (!gfx_v11_pipe_reset_support (adev))
@@ -6918,109 +6946,115 @@ static int gfx_v11_0_reset_compute_pipe(struct amdgpu_ring *ring)gfx_v11_0_set_safe_mode (adev,0 );mutex_lock (&adev->srbm_mutex); -soc21_grbm_select (adev, ring->me, ring->pipe, ring->queue,0 ); - - reset_pipe =RREG32_SOC15 (GC ,0 , regCP_MEC_RS64_CNTL); - clean_pipe = reset_pipe; +soc21_grbm_select (adev, me, pipe, queue,0 );if (adev->gfx.rs64_enable) { + reset_val =RREG32_SOC15 (GC ,0 , regCP_MEC_RS64_CNTL); + clean_val = reset_val; -switch (ring->pipe) { +switch (pipe) {case 0 : - reset_pipe =REG_SET_FIELD (reset_pipe,CP_MEC_RS64_CNTL , -MEC_PIPE0_RESET ,1 ); - clean_pipe =REG_SET_FIELD (clean_pipe,CP_MEC_RS64_CNTL , -MEC_PIPE0_RESET ,0 ); + reset_val =REG_SET_FIELD (reset_val,CP_MEC_RS64_CNTL , +MEC_PIPE0_RESET ,1 ); + clean_val =REG_SET_FIELD (clean_val,CP_MEC_RS64_CNTL , +MEC_PIPE0_RESET ,0 );break ;case 1 : - reset_pipe =REG_SET_FIELD (reset_pipe,CP_MEC_RS64_CNTL , -MEC_PIPE1_RESET ,1 ); - clean_pipe =REG_SET_FIELD (clean_pipe,CP_MEC_RS64_CNTL , -MEC_PIPE1_RESET ,0 ); + reset_val =REG_SET_FIELD (reset_val,CP_MEC_RS64_CNTL , +MEC_PIPE1_RESET ,1 ); + clean_val =REG_SET_FIELD (clean_val,CP_MEC_RS64_CNTL , +MEC_PIPE1_RESET ,0 );break ;case 2 : - reset_pipe =REG_SET_FIELD (reset_pipe,CP_MEC_RS64_CNTL , -MEC_PIPE2_RESET ,1 ); - clean_pipe =REG_SET_FIELD (clean_pipe,CP_MEC_RS64_CNTL , -MEC_PIPE2_RESET ,0 ); + reset_val =REG_SET_FIELD (reset_val,CP_MEC_RS64_CNTL , +MEC_PIPE2_RESET ,1 ); + clean_val =REG_SET_FIELD (clean_val,CP_MEC_RS64_CNTL , +MEC_PIPE2_RESET ,0 );break ;case 3 : - reset_pipe =REG_SET_FIELD (reset_pipe,CP_MEC_RS64_CNTL , -MEC_PIPE3_RESET ,1 ); - clean_pipe =REG_SET_FIELD (clean_pipe,CP_MEC_RS64_CNTL , -MEC_PIPE3_RESET ,0 ); + reset_val =REG_SET_FIELD (reset_val,CP_MEC_RS64_CNTL , +MEC_PIPE3_RESET ,1 ); + clean_val =REG_SET_FIELD (clean_val,CP_MEC_RS64_CNTL , +MEC_PIPE3_RESET ,0 );break ;default :break ; } -WREG32_SOC15 (GC ,0 , regCP_MEC_RS64_CNTL, reset_pipe); -WREG32_SOC15 (GC ,0 , regCP_MEC_RS64_CNTL, clean_pipe); +WREG32_SOC15 (GC ,0 , regCP_MEC_RS64_CNTL, reset_val); +gfx_v11_0_clear_hqds_on_mec_pipe (adev, me, pipe); +soc21_grbm_select (adev, me, pipe, queue,0 ); +WREG32_SOC15 (GC ,0 , regCP_MEC_RS64_CNTL, clean_val); r = (RREG32_SOC15 (GC ,0 , regCP_MEC_RS64_INSTR_PNTR) <<2 ) -RS64_FW_UC_START_ADDR_LO ; }else { -if (ring->me ==1 ) { -switch (ring->pipe) { + reset_val =RREG32_SOC15 (GC ,0 , regCP_MEC_CNTL); + clean_val = reset_val; + +if (me ==1 ) { +switch (pipe) {case 0 : - reset_pipe =REG_SET_FIELD (reset_pipe,CP_MEC_CNTL , -MEC_ME1_PIPE0_RESET ,1 ); - clean_pipe =REG_SET_FIELD (clean_pipe,CP_MEC_CNTL , -MEC_ME1_PIPE0_RESET ,0 ); + reset_val =REG_SET_FIELD (reset_val,CP_MEC_CNTL , +MEC_ME1_PIPE0_RESET ,1 ); + clean_val =REG_SET_FIELD (clean_val,CP_MEC_CNTL , +MEC_ME1_PIPE0_RESET ,0 );break ;case 1 : - reset_pipe =REG_SET_FIELD (reset_pipe,CP_MEC_CNTL , -MEC_ME1_PIPE1_RESET ,1 ); - clean_pipe =REG_SET_FIELD (clean_pipe,CP_MEC_CNTL , -MEC_ME1_PIPE1_RESET ,0 ); + reset_val =REG_SET_FIELD (reset_val,CP_MEC_CNTL , +MEC_ME1_PIPE1_RESET ,1 ); + clean_val =REG_SET_FIELD (clean_val,CP_MEC_CNTL , +MEC_ME1_PIPE1_RESET ,0 );break ;case 2 : - reset_pipe =REG_SET_FIELD (reset_pipe,CP_MEC_CNTL , -MEC_ME1_PIPE2_RESET ,1 ); - clean_pipe =REG_SET_FIELD (clean_pipe,CP_MEC_CNTL , -MEC_ME1_PIPE2_RESET ,0 ); + reset_val =REG_SET_FIELD (reset_val,CP_MEC_CNTL , +MEC_ME1_PIPE2_RESET ,1 ); + clean_val =REG_SET_FIELD (clean_val,CP_MEC_CNTL , +MEC_ME1_PIPE2_RESET ,0 );break ;case 3 : - reset_pipe =REG_SET_FIELD (reset_pipe,CP_MEC_CNTL , -MEC_ME1_PIPE3_RESET ,1 ); - clean_pipe =REG_SET_FIELD (clean_pipe,CP_MEC_CNTL , -MEC_ME1_PIPE3_RESET ,0 ); + reset_val =REG_SET_FIELD (reset_val,CP_MEC_CNTL , +MEC_ME1_PIPE3_RESET ,1 ); + clean_val =REG_SET_FIELD (clean_val,CP_MEC_CNTL , +MEC_ME1_PIPE3_RESET ,0 );break ;default :break ; }/* mec1 fw pc: CP_MEC1_INSTR_PNTR */ }else { -switch (ring->pipe) { +switch (pipe) {case 0 : - reset_pipe =REG_SET_FIELD (reset_pipe,CP_MEC_CNTL , -MEC_ME2_PIPE0_RESET ,1 ); - clean_pipe =REG_SET_FIELD (clean_pipe,CP_MEC_CNTL , -MEC_ME2_PIPE0_RESET ,0 ); + reset_val =REG_SET_FIELD (reset_val,CP_MEC_CNTL , +MEC_ME2_PIPE0_RESET ,1 ); + clean_val =REG_SET_FIELD (clean_val,CP_MEC_CNTL , +MEC_ME2_PIPE0_RESET ,0 );break ;case 1 : - reset_pipe =REG_SET_FIELD (reset_pipe,CP_MEC_CNTL , -MEC_ME2_PIPE1_RESET ,1 ); - clean_pipe =REG_SET_FIELD (clean_pipe,CP_MEC_CNTL , -MEC_ME2_PIPE1_RESET ,0 ); + reset_val =REG_SET_FIELD (reset_val,CP_MEC_CNTL , +MEC_ME2_PIPE1_RESET ,1 ); + clean_val =REG_SET_FIELD (clean_val,CP_MEC_CNTL , +MEC_ME2_PIPE1_RESET ,0 );break ;case 2 : - reset_pipe =REG_SET_FIELD (reset_pipe,CP_MEC_CNTL , -MEC_ME2_PIPE2_RESET ,1 ); - clean_pipe =REG_SET_FIELD (clean_pipe,CP_MEC_CNTL , -MEC_ME2_PIPE2_RESET ,0 ); + reset_val =REG_SET_FIELD (reset_val,CP_MEC_CNTL , +MEC_ME2_PIPE2_RESET ,1 ); + clean_val =REG_SET_FIELD (clean_val,CP_MEC_CNTL , +MEC_ME2_PIPE2_RESET ,0 );break ;case 3 : - reset_pipe =REG_SET_FIELD (reset_pipe,CP_MEC_CNTL , -MEC_ME2_PIPE3_RESET ,1 ); - clean_pipe =REG_SET_FIELD (clean_pipe,CP_MEC_CNTL , -MEC_ME2_PIPE3_RESET ,0 ); + reset_val =REG_SET_FIELD (reset_val,CP_MEC_CNTL , +MEC_ME2_PIPE3_RESET ,1 ); + clean_val =REG_SET_FIELD (clean_val,CP_MEC_CNTL , +MEC_ME2_PIPE3_RESET ,0 );break ;default :break ; }/* mec2 fw pc: CP:CP_MEC2_INSTR_PNTR */ } -WREG32_SOC15 (GC ,0 , regCP_MEC_CNTL, reset_pipe); -WREG32_SOC15 (GC ,0 , regCP_MEC_CNTL, clean_pipe); +WREG32_SOC15 (GC ,0 , regCP_MEC_CNTL, reset_val); +gfx_v11_0_clear_hqds_on_mec_pipe (adev, me, pipe); +soc21_grbm_select (adev, me, pipe, queue,0 ); +WREG32_SOC15 (GC ,0 , regCP_MEC_CNTL, clean_val); r =RREG32 (SOC15_REG_OFFSET (GC ,0 , regCP_MEC1_INSTR_PNTR)); }
@@ -7028,8 +7062,8 @@ static int gfx_v11_0_reset_compute_pipe(struct amdgpu_ring *ring)mutex_unlock (&adev->srbm_mutex);gfx_v11_0_unset_safe_mode (adev,0 ); -dev_info (adev->dev,"The ring %s pipe resets to MEC FW start PC: %s\n" , ring->name, - r ==0 ?"successfully" :"failed" ); +dev_dbg (adev->dev,"MEC pipe me%u pipe%u queue%u resets to MEC FW start PC: %s\n" , + me, pipe, queue, r ==0 ?"successfully" :"failed" );/*FIXME:Sometimes driver can't cache the MEC firmware start PC correctly, so the pipe * reset status relies on the compute ring test result. */
@@ -7048,9 +7082,16 @@ static int gfx_v11_0_reset_kcq(struct amdgpu_ring *ring, r =amdgpu_mes_reset_legacy_queue (ring->adev, ring, vmid, true,0 );if (r) {dev_warn (adev->dev,"fail(%d) to reset kcq and try pipe reset\n" , r); - r =gfx_v11_0_reset_compute_pipe (ring); -if (r) -return r; + +amdgpu_amdkfd_suspend (adev, true); + r =amdgpu_gfx_mec_pipe_reset_run (adev, + ring->xcc_id, ring->me, ring->pipe, + ring->queue, timedout_fence, + gfx_v11_0_reset_compute_pipe, +NULL , + gfx_v11_0_kcq_init_queue); +amdgpu_amdkfd_resume (adev, true); +return r; } r =gfx_v11_0_kcq_init_queue (ring, true);
Signature · 1 line
2.49.0
PATCH v3 3/8 drm/amdgpu/gfx11: Fall back to pipe reset if per-queue reset ring test fails
In reply to Jesse ZhangAfter a per-queue reset via MES, verify that the queue is functional by performing a ring test. If the test fails, fall back to a full pipe reset to ensure proper recovery.
This adds a fallback path similar to the one already present when the initial per-queue reset attempt fails, improving the robustness of KCQ reset handling on GFX11 hardware.
--- drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c | 11 +++++++++++ 1 file changed, 11 insertions(+)
drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c +11 -0
@@ -7075,14 +7075,18 @@ static int gfx_v11_0_reset_kcq(struct amdgpu_ring *ring, struct amdgpu_fence *timedout_fence) {struct amdgpu_device *adev = ring->adev; +int reset_mode =AMDGPU_RESET_TYPE_PER_QUEUE ;int r =0 ;amdgpu_ring_reset_helper_begin (ring, timedout_fence); r =amdgpu_mes_reset_legacy_queue (ring->adev, ring, vmid, true,0 ); + +pipe_reset:if (r) {dev_warn (adev->dev,"fail(%d) to reset kcq and try pipe reset\n" , r); + reset_mode =AMDGPU_RESET_TYPE_PER_PIPE ;amdgpu_amdkfd_suspend (adev, true); r = amdgpu_gfx_mec_pipe_reset_run(adev, ring->xcc_id, ring->me, ring->pipe,
@@ -7105,6 +7109,13 @@ static int gfx_v11_0_reset_kcq(struct amdgpu_ring *ring,return r; } +if (reset_mode ==AMDGPU_RESET_TYPE_PER_QUEUE ) { +if (amdgpu_ring_reset_helper_end (ring, timedout_fence)) + goto pipe_reset; +else +return 0 ; + } +return amdgpu_ring_reset_helper_end (ring, timedout_fence); }
Signature · 1 line
2.49.0
PATCH v3 4/8 drm/amdgpu/gfx11: enable per-pipe reset support for compute queues
In reply to Jesse ZhangThis allows the driver to fall back to pipe-level reset when per-queue reset fails, improving recovery success for hung compute or graphics rings.
V2: replace both gfx_v11_compute_pipe_reset_support() and gfx_v11_pipe_reset_support() with amdgpu_ring_is_reset_type_supported (Alex)
--- drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c | 25 +++++++++++-------------- 1 file changed, 11 insertions(+), 14 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c +11 -14
@@ -1851,6 +1851,8 @@ static int gfx_v11_0_sw_init(struct amdgpu_ip_block *ip_block) !adev->debug_disable_gpu_ring_reset) { adev->gfx.compute_supported_reset |=AMDGPU_RESET_TYPE_PER_QUEUE ; adev->gfx.gfx_supported_reset |=AMDGPU_RESET_TYPE_PER_QUEUE ; +if (adev->gfx.mec_fw_version >=2670 ) + adev->gfx.compute_supported_reset |=AMDGPU_RESET_TYPE_PER_PIPE ; }break ;default :
@@ -1858,6 +1860,7 @@ static int gfx_v11_0_sw_init(struct amdgpu_ip_block *ip_block) !adev->debug_disable_gpu_ring_reset) { adev->gfx.compute_supported_reset |=AMDGPU_RESET_TYPE_PER_QUEUE ; adev->gfx.gfx_supported_reset |=AMDGPU_RESET_TYPE_PER_QUEUE ; + adev->gfx.compute_supported_reset |=AMDGPU_RESET_TYPE_PER_PIPE ; }break ; }
@@ -6807,13 +6810,6 @@ static void gfx_v11_0_emit_mem_sync(struct amdgpu_ring *ring)amdgpu_ring_write (ring, gcr_cntl);/* GCR_CNTL */ } -static bool gfx_v11_pipe_reset_support (struct amdgpu_device *adev) -{ -/* Disable the pipe reset until the CPFW fully support it.*/ -dev_warn_once (adev->dev,"The CPFW hasn't support pipe reset yet.\n" ); -return false; -} -static int gfx_v11_reset_gfx_pipe (struct amdgpu_ring *ring) {
@@ -6821,9 +6817,6 @@ static int gfx_v11_reset_gfx_pipe(struct amdgpu_ring *ring)uint32_t reset_pipe =0 , clean_pipe =0 ;int r; -if (!gfx_v11_pipe_reset_support (adev)) -return -EOPNOTSUPP ; -gfx_v11_0_set_safe_mode (adev,0 );mutex_lock (&adev->srbm_mutex);soc21_grbm_select (adev, ring->me, ring->pipe, ring->queue,0 );
@@ -6884,6 +6877,10 @@ static int gfx_v11_0_reset_kgq(struct amdgpu_ring *ring,if (r) {dev_warn (adev->dev,"reset via MES failed and try pipe reset %d\n" , r); +if (!amdgpu_ring_is_reset_type_supported (ring, +AMDGPU_RESET_TYPE_PER_PIPE )) +return r; + r =gfx_v11_reset_gfx_pipe (ring);if (r)return r;
@@ -6941,9 +6938,6 @@ static int gfx_v11_0_reset_compute_pipe(struct amdgpu_device *adev,uint32_t reset_val, clean_val;int r; -if (!gfx_v11_pipe_reset_support (adev)) -return -EOPNOTSUPP ; -gfx_v11_0_set_safe_mode (adev,0 );mutex_lock (&adev->srbm_mutex);soc21_grbm_select (adev, me, pipe, queue,0 );
@@ -7085,8 +7079,11 @@ static int gfx_v11_0_reset_kcq(struct amdgpu_ring *ring, pipe_reset:if (r) {dev_warn (adev->dev,"fail(%d) to reset kcq and try pipe reset\n" , r); - reset_mode =AMDGPU_RESET_TYPE_PER_PIPE ; +if (!amdgpu_ring_is_reset_type_supported (ring, +AMDGPU_RESET_TYPE_PER_PIPE )) +return r; +amdgpu_amdkfd_suspend (adev, true); r = amdgpu_gfx_mec_pipe_reset_run(adev, ring->xcc_id, ring->me, ring->pipe,
Signature · 1 line
2.49.0
PATCH v3 5/8 drm/amdgpu/gfx12: Refactor compute pipe reset and add HQD cleanup
In reply to Jesse ZhangRefactor gfx_v12_0_reset_compute_pipe() to accept explicit me, pipe, and queue parameters instead of deriving them from the ring structure. This enables the function to be used in generic pipe reset flows.
Introduce gfx_v12_0_clear_hqds_on_mec_pipe() to properly clear CP_HQD_ACTIVE and CP_HQD_DEQUEUE_REQUEST for all queues on a given MEC pipe while the pipe reset is asserted, ensuring the HQDs are torn down correctly before deasserting reset.
Switch the KCQ reset path to use the common MEC pipe reset helper amdgpu_gfx_mec_pipe_reset_run(), which coordinates the reset sequence including KFD suspend/resume to avoid conflicts with user mode queues.
--- drivers/gpu/drm/amd/amdgpu/gfx_v12_0.c | 133 ++++++++++++++++--------- 1 file changed, 85 insertions(+), 48 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v12_0.c +85 -48
@@ -5355,10 +5355,38 @@ static int gfx_v12_0_reset_kgq(struct amdgpu_ring *ring,return amdgpu_ring_reset_helper_end (ring, timedout_fence); } -static int gfx_v12_0_reset_compute_pipe (struct amdgpu_ring *ring) +/* +* With MEC pipe reset asserted, clear CP_HQD_ACTIVE / CP_HQD_DEQUEUE_REQUEST for +* every queue on (me, pipe). HQDs must be torn down while pipe reset stays +* asserted; only then clear the pipe reset bit. +* Caller must hold adev->srbm_mutex. +*/ +static void gfx_v12_0_clear_hqds_on_mec_pipe (struct amdgpu_device *adev,u32 me, +u32 pipe) { -struct amdgpu_device *adev = ring->adev; -uint32_t reset_pipe =0 , clean_pipe =0 ; +unsigned int q; +int j; + +for (q =0 ; q < adev->gfx.mec.num_queue_per_pipe; q++) { +soc24_grbm_select (adev, me, pipe, q,0 ); +/* Start from a clean HQD dequeue state before forcing HQD inactive. */ +WREG32_SOC15 (GC ,0 , regCP_HQD_ACTIVE,0 ); +if (RREG32_SOC15 (GC ,0 , regCP_HQD_ACTIVE) &1 ) { +WREG32_SOC15 (GC ,0 , regCP_HQD_DEQUEUE_REQUEST,1 ); +for (j =0 ; j < adev->usec_timeout; j++) { +if (!(RREG32_SOC15 (GC ,0 , regCP_HQD_ACTIVE) &1 )) +break ; +udelay (1 ); + } + } +WREG32_SOC15 (GC ,0 , regCP_HQD_DEQUEUE_REQUEST,0 ); + } +} + +static int gfx_v12_0_reset_compute_pipe (struct amdgpu_device *adev, +u32 me,u32 pipe,u32 queue) +{ +uint32_t reset_val, clean_val;int r =0 ;if (!gfx_v12_pipe_reset_support (adev))
@@ -5366,75 +5394,78 @@ static int gfx_v12_0_reset_compute_pipe(struct amdgpu_ring *ring)gfx_v12_0_set_safe_mode (adev,0 );mutex_lock (&adev->srbm_mutex); -soc24_grbm_select (adev, ring->me, ring->pipe, ring->queue,0 ); - - reset_pipe =RREG32_SOC15 (GC ,0 , regCP_MEC_RS64_CNTL); - clean_pipe = reset_pipe; - +soc24_grbm_select (adev, me, pipe, queue,0 );if (adev->gfx.rs64_enable) { -switch (ring->pipe) { + reset_val =RREG32_SOC15 (GC ,0 , regCP_MEC_RS64_CNTL); + clean_val = reset_val; + +switch (pipe) {case 0 : - reset_pipe =REG_SET_FIELD (reset_pipe,CP_MEC_RS64_CNTL , -MEC_PIPE0_RESET ,1 ); - clean_pipe =REG_SET_FIELD (clean_pipe,CP_MEC_RS64_CNTL , -MEC_PIPE0_RESET ,0 ); + reset_val =REG_SET_FIELD (reset_val,CP_MEC_RS64_CNTL , +MEC_PIPE0_RESET ,1 ); + clean_val =REG_SET_FIELD (clean_val,CP_MEC_RS64_CNTL , +MEC_PIPE0_RESET ,0 );break ;case 1 : - reset_pipe =REG_SET_FIELD (reset_pipe,CP_MEC_RS64_CNTL , -MEC_PIPE1_RESET ,1 ); - clean_pipe =REG_SET_FIELD (clean_pipe,CP_MEC_RS64_CNTL , -MEC_PIPE1_RESET ,0 ); + reset_val =REG_SET_FIELD (reset_val,CP_MEC_RS64_CNTL , +MEC_PIPE1_RESET ,1 ); + clean_val =REG_SET_FIELD (clean_val,CP_MEC_RS64_CNTL , +MEC_PIPE1_RESET ,0 );break ;case 2 : - reset_pipe =REG_SET_FIELD (reset_pipe,CP_MEC_RS64_CNTL , -MEC_PIPE2_RESET ,1 ); - clean_pipe =REG_SET_FIELD (clean_pipe,CP_MEC_RS64_CNTL , -MEC_PIPE2_RESET ,0 ); + reset_val =REG_SET_FIELD (reset_val,CP_MEC_RS64_CNTL , +MEC_PIPE2_RESET ,1 ); + clean_val =REG_SET_FIELD (clean_val,CP_MEC_RS64_CNTL , +MEC_PIPE2_RESET ,0 );break ;case 3 : - reset_pipe =REG_SET_FIELD (reset_pipe,CP_MEC_RS64_CNTL , -MEC_PIPE3_RESET ,1 ); - clean_pipe =REG_SET_FIELD (clean_pipe,CP_MEC_RS64_CNTL , -MEC_PIPE3_RESET ,0 ); + reset_val =REG_SET_FIELD (reset_val,CP_MEC_RS64_CNTL , +MEC_PIPE3_RESET ,1 ); + clean_val =REG_SET_FIELD (clean_val,CP_MEC_RS64_CNTL , +MEC_PIPE3_RESET ,0 );break ;default :break ; } -WREG32_SOC15 (GC ,0 , regCP_MEC_RS64_CNTL, reset_pipe); -WREG32_SOC15 (GC ,0 , regCP_MEC_RS64_CNTL, clean_pipe); +WREG32_SOC15 (GC ,0 , regCP_MEC_RS64_CNTL, reset_val); +gfx_v12_0_clear_hqds_on_mec_pipe (adev, me, pipe); +soc24_grbm_select (adev, me, pipe, queue,0 ); +WREG32_SOC15 (GC ,0 , regCP_MEC_RS64_CNTL, clean_val); r = (RREG32_SOC15 (GC ,0 , regCP_MEC_RS64_INSTR_PNTR) <<2 ) -RS64_FW_UC_START_ADDR_LO ; }else { -switch (ring->pipe) { + reset_val =RREG32_SOC15 (GC ,0 , regCP_MEC_CNTL); + clean_val = reset_val; + +switch (pipe) {case 0 : - reset_pipe =REG_SET_FIELD (reset_pipe,CP_MEC_CNTL , -MEC_ME1_PIPE0_RESET ,1 ); - clean_pipe =REG_SET_FIELD (clean_pipe,CP_MEC_CNTL , -MEC_ME1_PIPE0_RESET ,0 ); + reset_val =REG_SET_FIELD (reset_val,CP_MEC_CNTL , +MEC_ME1_PIPE0_RESET ,1 ); + clean_val =REG_SET_FIELD (clean_val,CP_MEC_CNTL , +MEC_ME1_PIPE0_RESET ,0 );break ;case 1 : - reset_pipe =REG_SET_FIELD (reset_pipe,CP_MEC_CNTL , -MEC_ME1_PIPE1_RESET ,1 ); - clean_pipe =REG_SET_FIELD (clean_pipe,CP_MEC_CNTL , -MEC_ME1_PIPE1_RESET ,0 ); + reset_val =REG_SET_FIELD (reset_val,CP_MEC_CNTL , +MEC_ME1_PIPE1_RESET ,1 ); + clean_val =REG_SET_FIELD (clean_val,CP_MEC_CNTL , +MEC_ME1_PIPE1_RESET ,0 );break ;default : -break ; +break ; } -WREG32_SOC15 (GC ,0 , regCP_MEC_CNTL, reset_pipe); -WREG32_SOC15 (GC ,0 , regCP_MEC_CNTL, clean_pipe); -/* Doesn't find the F32 MEC instruction pointer register, and suppose -* the driver won't run into the F32 mode. -*/ + +WREG32_SOC15 (GC ,0 , regCP_MEC_CNTL, reset_val); +gfx_v12_0_clear_hqds_on_mec_pipe (adev, me, pipe); +soc24_grbm_select (adev, me, pipe, queue,0 ); +WREG32_SOC15 (GC ,0 , regCP_MEC_CNTL, clean_val); }soc24_grbm_select (adev,0 ,0 ,0 ,0 );mutex_unlock (&adev->srbm_mutex);gfx_v12_0_unset_safe_mode (adev,0 ); -dev_info (adev->dev,"The ring %s pipe resets: %s\n" , ring->name, - r ==0 ?"successfully" :"failed" ); -/* Need the ring test to verify the pipe reset result.*/ +dev_dbg (adev->dev,"MEC pipe me%u pipe%u queue%u resets to MEC FW start PC: %s\n" , + me, pipe, queue, r ==0 ?"successfully" :"failed" );return 0 ; }
@@ -5450,9 +5481,15 @@ static int gfx_v12_0_reset_kcq(struct amdgpu_ring *ring, r =amdgpu_mes_reset_legacy_queue (ring->adev, ring, vmid, true,0 );if (r) {dev_warn (adev->dev,"fail(%d) to reset kcq and try pipe reset\n" , r); - r =gfx_v12_0_reset_compute_pipe (ring); -if (r) -return r; +amdgpu_amdkfd_suspend (adev, true); + r =amdgpu_gfx_mec_pipe_reset_run (adev, + ring->xcc_id, ring->me, ring->pipe, + ring->queue, timedout_fence, + gfx_v12_0_reset_compute_pipe, +NULL , + gfx_v12_0_kcq_init_queue); +amdgpu_amdkfd_resume (adev, true); +return r; } r =gfx_v12_0_kcq_init_queue (ring, true);
Signature · 1 line
2.49.0
PATCH v3 6/8 drm/amdgpu/gfx12: Fall back to pipe reset if per-queue reset ring test fails
In reply to Jesse ZhangAfter a per-queue reset via MES, verify that the queue is functional by performing a ring test. If the test fails, fall back to a full pipe reset to ensure proper recovery.
This adds a fallback path similar to the one already present when the initial per-queue reset attempt fails, improving the robustness of KCQ reset handling on GFX11 hardware.
--- drivers/gpu/drm/amd/amdgpu/gfx_v12_0.c | 11 +++++++++++ 1 file changed, 11 insertions(+)
drivers/gpu/drm/amd/amdgpu/gfx_v12_0.c +11 -0
@@ -5474,13 +5474,17 @@ static int gfx_v12_0_reset_kcq(struct amdgpu_ring *ring, struct amdgpu_fence *timedout_fence) {struct amdgpu_device *adev = ring->adev; +int reset_mode =AMDGPU_RESET_TYPE_PER_QUEUE ;int r;amdgpu_ring_reset_helper_begin (ring, timedout_fence); r =amdgpu_mes_reset_legacy_queue (ring->adev, ring, vmid, true,0 ); + +pipe_reset:if (r) {dev_warn (adev->dev,"fail(%d) to reset kcq and try pipe reset\n" , r); + reset_mode =AMDGPU_RESET_TYPE_PER_PIPE ;amdgpu_amdkfd_suspend (adev, true); r = amdgpu_gfx_mec_pipe_reset_run(adev, ring->xcc_id, ring->me, ring->pipe,
@@ -5503,6 +5507,13 @@ static int gfx_v12_0_reset_kcq(struct amdgpu_ring *ring,return r; } +if (reset_mode ==AMDGPU_RESET_TYPE_PER_QUEUE ) { +if (amdgpu_ring_reset_helper_end (ring, timedout_fence)) + goto pipe_reset; +else +return 0 ; + } +return amdgpu_ring_reset_helper_end (ring, timedout_fence); }
Signature · 1 line
2.49.0
PATCH v3 7/8 drm/amdgpu/gfx12: enable per-pipe reset support for compute queues
In reply to Jesse ZhangThis allows the driver to fall back to pipe-level reset when per-queue reset fails, improving recovery success for hung compute or graphics rings.
V2: replace both gfx_v11_compute_pipe_reset_support() and gfx_v11_pipe_reset_support() with amdgpu_ring_is_reset_type_supported (Alex)
--- drivers/gpu/drm/amd/amdgpu/gfx_v12_0.c | 23 ++++++++++------------- 1 file changed, 10 insertions(+), 13 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v12_0.c +10 -13
@@ -1557,6 +1557,8 @@ static int gfx_v12_0_sw_init(struct amdgpu_ip_block *ip_block) !adev->debug_disable_gpu_ring_reset) { adev->gfx.compute_supported_reset |=AMDGPU_RESET_TYPE_PER_QUEUE ; adev->gfx.gfx_supported_reset |=AMDGPU_RESET_TYPE_PER_QUEUE ; +if (adev->gfx.mec_fw_version >=3190 ) + adev->gfx.compute_supported_reset |=AMDGPU_RESET_TYPE_PER_PIPE ; }break ;default :
@@ -5257,22 +5259,12 @@ static void gfx_v12_ip_dump(struct amdgpu_ip_block *ip_block)amdgpu_gfx_off_ctrl (adev, true); } -static bool gfx_v12_pipe_reset_support (struct amdgpu_device *adev) -{ -/* Disable the pipe reset until the CPFW fully support it.*/ -dev_warn_once (adev->dev,"The CPFW hasn't support pipe reset yet.\n" ); -return false; -} -static int gfx_v12_reset_gfx_pipe (struct amdgpu_ring *ring) {struct amdgpu_device *adev = ring->adev;uint32_t reset_pipe =0 , clean_pipe =0 ;int r; -if (!gfx_v12_pipe_reset_support (adev)) -return -EOPNOTSUPP ; -gfx_v12_0_set_safe_mode (adev,0 );mutex_lock (&adev->srbm_mutex);soc24_grbm_select (adev, ring->me, ring->pipe, ring->queue,0 );
@@ -5333,6 +5325,10 @@ static int gfx_v12_0_reset_kgq(struct amdgpu_ring *ring, r =amdgpu_mes_reset_legacy_queue (ring->adev, ring, vmid, use_mmio,0 );if (r) {dev_warn (adev->dev,"reset via MES failed and try pipe reset %d\n" , r); +if (!amdgpu_ring_is_reset_type_supported (ring, +AMDGPU_RESET_TYPE_PER_PIPE )) +return r; + r =gfx_v12_reset_gfx_pipe (ring);if (r)return r;
@@ -5389,9 +5385,6 @@ static int gfx_v12_0_reset_compute_pipe(struct amdgpu_device *adev,uint32_t reset_val, clean_val;int r =0 ; -if (!gfx_v12_pipe_reset_support (adev)) -return -EOPNOTSUPP ; -gfx_v12_0_set_safe_mode (adev,0 );mutex_lock (&adev->srbm_mutex);soc24_grbm_select (adev, me, pipe, queue,0 );
@@ -5485,6 +5478,10 @@ static int gfx_v12_0_reset_kcq(struct amdgpu_ring *ring,if (r) {dev_warn (adev->dev,"fail(%d) to reset kcq and try pipe reset\n" , r); reset_mode =AMDGPU_RESET_TYPE_PER_PIPE ; +if (!amdgpu_ring_is_reset_type_supported (ring, +AMDGPU_RESET_TYPE_PER_PIPE )) +return r; +amdgpu_amdkfd_suspend (adev, true); r = amdgpu_gfx_mec_pipe_reset_run(adev, ring->xcc_id, ring->me, ring->pipe,
Signature · 1 line
2.49.0
PATCH v3 8/8 drm/amdgpu/gfx_v12_0: set gfx.rs64_enable from PFP header on GFX12
In reply to Jesse Zhanggfx_v12_0_init_microcode() always loads RS64 CP ucode but never set adev->gfx.rs64_enable, so it stayed false and code that branches on it (e.g. MEC pipe reset) used the legacy CP_MEC_CNTL path incorrectly.
Match GFX11: derive RS64 mode from the PFP firmware header (v2.0) via amdgpu_ucode_hdr_version(). Log at debug when RS64 is enabled.
--- drivers/gpu/drm/amd/amdgpu/gfx_v12_0.c | 7 +++++++ 1 file changed, 7 insertions(+)
drivers/gpu/drm/amd/amdgpu/gfx_v12_0.c +7 -0
@@ -602,6 +602,13 @@ static int gfx_v12_0_init_microcode(struct amdgpu_device *adev) "amdgpu/%s_pfp.bin", ucode_prefix);if (err) goto out; + + adev->gfx.rs64_enable =amdgpu_ucode_hdr_version ( + (union amdgpu_firmware_header *) + adev->gfx.pfp_fw->data,2 ,0 ); +if (adev->gfx.rs64_enable) +dev_dbg (adev->dev,"CP RS64 enable\n" ); +amdgpu_gfx_cp_init_microcode (adev,AMDGPU_UCODE_ID_CP_RS64_PFP );amdgpu_gfx_cp_init_microcode (adev,AMDGPU_UCODE_ID_CP_RS64_PFP_P0_STACK );
Signature · 1 line
2.49.0