diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h')
| -rw-r--r-- | drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h | 85 |
1 files changed, 84 insertions, 1 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h index 548015bb6ee7..25590b301f25 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h @@ -56,7 +56,7 @@ enum admgpu_mes_pipe { struct amdgpu_mes { struct amdgpu_device *adev; - struct mutex mutex; + struct mutex mutex_hidden; struct idr pasid_idr; struct idr gang_id_idr; @@ -109,9 +109,11 @@ struct amdgpu_mes { uint32_t query_status_fence_offs; uint64_t query_status_fence_gpu_addr; uint64_t *query_status_fence_ptr; + uint32_t saved_flags; /* initialize kiq pipe */ int (*kiq_hw_init)(struct amdgpu_device *adev); + int (*kiq_hw_fini)(struct amdgpu_device *adev); /* ip specific functions */ const struct amdgpu_mes_funcs *funcs; @@ -198,6 +200,10 @@ struct mes_add_queue_input { uint64_t wptr_addr; uint32_t queue_type; uint32_t paging; + uint32_t gws_base; + uint32_t gws_size; + uint64_t tba_addr; + uint64_t tma_addr; }; struct mes_remove_queue_input { @@ -205,6 +211,16 @@ struct mes_remove_queue_input { uint64_t gang_context_addr; }; +struct mes_unmap_legacy_queue_input { + enum amdgpu_unmap_queues_action action; + uint32_t queue_type; + uint32_t doorbell_offset; + uint32_t pipe_id; + uint32_t queue_id; + uint64_t trail_fence_addr; + uint64_t trail_fence_data; +}; + struct mes_suspend_gang_input { bool suspend_all_gangs; uint64_t gang_context_addr; @@ -224,6 +240,9 @@ struct amdgpu_mes_funcs { int (*remove_hw_queue)(struct amdgpu_mes *mes, struct mes_remove_queue_input *input); + int (*unmap_legacy_queue)(struct amdgpu_mes *mes, + struct mes_unmap_legacy_queue_input *input); + int (*suspend_gang)(struct amdgpu_mes *mes, struct mes_suspend_gang_input *input); @@ -232,6 +251,7 @@ struct amdgpu_mes_funcs { }; #define amdgpu_mes_kiq_hw_init(adev) (adev)->mes.kiq_hw_init((adev)) +#define amdgpu_mes_kiq_hw_fini(adev) (adev)->mes.kiq_hw_fini((adev)) int amdgpu_mes_ctx_get_offs(struct amdgpu_ring *ring, unsigned int id_offs); @@ -255,6 +275,11 @@ int amdgpu_mes_add_hw_queue(struct amdgpu_device *adev, int gang_id, int *queue_id); int amdgpu_mes_remove_hw_queue(struct amdgpu_device *adev, int queue_id); +int amdgpu_mes_unmap_legacy_queue(struct amdgpu_device *adev, + struct amdgpu_ring *ring, + enum amdgpu_unmap_queues_action action, + u64 gpu_addr, u64 seq); + int amdgpu_mes_add_ring(struct amdgpu_device *adev, int gang_id, int queue_type, int idx, struct amdgpu_mes_ctx_data *ctx_data, @@ -280,4 +305,62 @@ unsigned int amdgpu_mes_get_doorbell_dw_offset_in_bar( uint32_t doorbell_index, unsigned int doorbell_id); int amdgpu_mes_doorbell_process_slice(struct amdgpu_device *adev); + +/* + * MES lock can be taken in MMU notifiers. + * + * A bit more detail about why to set no-FS reclaim with MES lock: + * + * The purpose of the MMU notifier is to stop GPU access to memory so + * that the Linux VM subsystem can move pages around safely. This is + * done by preempting user mode queues for the affected process. When + * MES is used, MES lock needs to be taken to preempt the queues. + * + * The MMU notifier callback entry point in the driver is + * amdgpu_mn_invalidate_range_start_hsa. The relevant call chain from + * there is: + * amdgpu_amdkfd_evict_userptr -> kgd2kfd_quiesce_mm -> + * kfd_process_evict_queues -> pdd->dev->dqm->ops.evict_process_queues + * + * The last part of the chain is a function pointer where we take the + * MES lock. + * + * The problem with taking locks in the MMU notifier is, that MMU + * notifiers can be called in reclaim-FS context. That's where the + * kernel frees up pages to make room for new page allocations under + * memory pressure. While we are running in reclaim-FS context, we must + * not trigger another memory reclaim operation because that would + * recursively reenter the reclaim code and cause a deadlock. The + * memalloc_nofs_save/restore calls guarantee that. + * + * In addition we also need to avoid lock dependencies on other locks taken + * under the MES lock, for example reservation locks. Here is a possible + * scenario of a deadlock: + * Thread A: takes and holds reservation lock | triggers reclaim-FS | + * MMU notifier | blocks trying to take MES lock + * Thread B: takes and holds MES lock | blocks trying to take reservation lock + * + * In this scenario Thread B gets involved in a deadlock even without + * triggering a reclaim-FS operation itself. + * To fix this and break the lock dependency chain you'd need to either: + * 1. protect reservation locks with memalloc_nofs_save/restore, or + * 2. avoid taking reservation locks under the MES lock. + * + * Reservation locks are taken all over the kernel in different subsystems, we + * have no control over them and their lock dependencies.So the only workable + * solution is to avoid taking other locks under the MES lock. + * As a result, make sure no reclaim-FS happens while holding this lock anywhere + * to prevent deadlocks when an MMU notifier runs in reclaim-FS context. + */ +static inline void amdgpu_mes_lock(struct amdgpu_mes *mes) +{ + mutex_lock(&mes->mutex_hidden); + mes->saved_flags = memalloc_noreclaim_save(); +} + +static inline void amdgpu_mes_unlock(struct amdgpu_mes *mes) +{ + memalloc_noreclaim_restore(mes->saved_flags); + mutex_unlock(&mes->mutex_hidden); +} #endif /* __AMDGPU_MES_H__ */ |