diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h')
| -rw-r--r-- | drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h | 252 |
1 files changed, 238 insertions, 14 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h index 7334982ea702..25590b301f25 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h @@ -24,6 +24,11 @@ #ifndef __AMDGPU_MES_H__ #define __AMDGPU_MES_H__ +#include "amdgpu_irq.h" +#include "kgd_kfd_interface.h" +#include "amdgpu_gfx.h" +#include <linux/sched/mm.h> + #define AMDGPU_MES_MAX_COMPUTE_PIPES 8 #define AMDGPU_MES_MAX_GFX_PIPES 2 #define AMDGPU_MES_MAX_SDMA_PIPES 2 @@ -37,11 +42,29 @@ enum amdgpu_mes_priority_level { AMDGPU_MES_PRIORITY_NUM_LEVELS }; +#define AMDGPU_MES_PROC_CTX_SIZE 0x1000 /* one page area */ +#define AMDGPU_MES_GANG_CTX_SIZE 0x1000 /* one page area */ + struct amdgpu_mes_funcs; +enum admgpu_mes_pipe { + AMDGPU_MES_SCHED_PIPE = 0, + AMDGPU_MES_KIQ_PIPE, + AMDGPU_MAX_MES_PIPES = 2, +}; + struct amdgpu_mes { struct amdgpu_device *adev; + struct mutex mutex_hidden; + + struct idr pasid_idr; + struct idr gang_id_idr; + struct idr queue_id_idr; + struct ida doorbell_ida; + + spinlock_t queue_id_lock; + uint32_t total_max_queue; uint32_t doorbell_id_offset; uint32_t max_doorbell_slices; @@ -51,27 +74,28 @@ struct amdgpu_mes { struct amdgpu_ring ring; - const struct firmware *fw; + const struct firmware *fw[AMDGPU_MAX_MES_PIPES]; /* mes ucode */ - struct amdgpu_bo *ucode_fw_obj; - uint64_t ucode_fw_gpu_addr; - uint32_t *ucode_fw_ptr; - uint32_t ucode_fw_version; - uint64_t uc_start_addr; + struct amdgpu_bo *ucode_fw_obj[AMDGPU_MAX_MES_PIPES]; + uint64_t ucode_fw_gpu_addr[AMDGPU_MAX_MES_PIPES]; + uint32_t *ucode_fw_ptr[AMDGPU_MAX_MES_PIPES]; + uint32_t ucode_fw_version[AMDGPU_MAX_MES_PIPES]; + uint64_t uc_start_addr[AMDGPU_MAX_MES_PIPES]; /* mes ucode data */ - struct amdgpu_bo *data_fw_obj; - uint64_t data_fw_gpu_addr; - uint32_t *data_fw_ptr; - uint32_t data_fw_version; - uint64_t data_start_addr; + struct amdgpu_bo *data_fw_obj[AMDGPU_MAX_MES_PIPES]; + uint64_t data_fw_gpu_addr[AMDGPU_MAX_MES_PIPES]; + uint32_t *data_fw_ptr[AMDGPU_MAX_MES_PIPES]; + uint32_t data_fw_version[AMDGPU_MAX_MES_PIPES]; + uint64_t data_start_addr[AMDGPU_MAX_MES_PIPES]; /* eop gpu obj */ - struct amdgpu_bo *eop_gpu_obj; - uint64_t eop_gpu_addr; + struct amdgpu_bo *eop_gpu_obj[AMDGPU_MAX_MES_PIPES]; + uint64_t eop_gpu_addr[AMDGPU_MAX_MES_PIPES]; - void *mqd_backup; + void *mqd_backup[AMDGPU_MAX_MES_PIPES]; + struct amdgpu_irq_src irq[AMDGPU_MAX_MES_PIPES]; uint32_t vmid_mask_gfxhub; uint32_t vmid_mask_mmhub; @@ -85,11 +109,81 @@ 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; }; +struct amdgpu_mes_process { + int pasid; + struct amdgpu_vm *vm; + uint64_t pd_gpu_addr; + struct amdgpu_bo *proc_ctx_bo; + uint64_t proc_ctx_gpu_addr; + void *proc_ctx_cpu_ptr; + uint64_t process_quantum; + struct list_head gang_list; + uint32_t doorbell_index; + unsigned long *doorbell_bitmap; + struct mutex doorbell_lock; +}; + +struct amdgpu_mes_gang { + int gang_id; + int priority; + int inprocess_gang_priority; + int global_priority_level; + struct list_head list; + struct amdgpu_mes_process *process; + struct amdgpu_bo *gang_ctx_bo; + uint64_t gang_ctx_gpu_addr; + void *gang_ctx_cpu_ptr; + uint64_t gang_quantum; + struct list_head queue_list; +}; + +struct amdgpu_mes_queue { + struct list_head list; + struct amdgpu_mes_gang *gang; + int queue_id; + uint64_t doorbell_off; + struct amdgpu_bo *mqd_obj; + void *mqd_cpu_ptr; + uint64_t mqd_gpu_addr; + uint64_t wptr_gpu_addr; + int queue_type; + int paging; + struct amdgpu_ring *ring; +}; + +struct amdgpu_mes_queue_properties { + int queue_type; + uint64_t hqd_base_gpu_addr; + uint64_t rptr_gpu_addr; + uint64_t wptr_gpu_addr; + uint32_t queue_size; + uint64_t eop_gpu_addr; + uint32_t hqd_pipe_priority; + uint32_t hqd_queue_priority; + bool paging; + struct amdgpu_ring *ring; + /* out */ + uint64_t doorbell_off; +}; + +struct amdgpu_mes_gang_properties { + uint32_t priority; + uint32_t gang_quantum; + uint32_t inprocess_gang_priority; + uint32_t priority_level; + int global_priority_level; +}; + struct mes_add_queue_input { uint32_t process_id; uint64_t page_table_base_addr; @@ -106,6 +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 { @@ -113,6 +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; @@ -132,6 +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); @@ -139,4 +250,117 @@ struct amdgpu_mes_funcs { struct mes_resume_gang_input *input); }; +#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); + +int amdgpu_mes_init(struct amdgpu_device *adev); +void amdgpu_mes_fini(struct amdgpu_device *adev); + +int amdgpu_mes_create_process(struct amdgpu_device *adev, int pasid, + struct amdgpu_vm *vm); +void amdgpu_mes_destroy_process(struct amdgpu_device *adev, int pasid); + +int amdgpu_mes_add_gang(struct amdgpu_device *adev, int pasid, + struct amdgpu_mes_gang_properties *gprops, + int *gang_id); +int amdgpu_mes_remove_gang(struct amdgpu_device *adev, int gang_id); + +int amdgpu_mes_suspend(struct amdgpu_device *adev); +int amdgpu_mes_resume(struct amdgpu_device *adev); + +int amdgpu_mes_add_hw_queue(struct amdgpu_device *adev, int gang_id, + struct amdgpu_mes_queue_properties *qprops, + 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, + struct amdgpu_ring **out); +void amdgpu_mes_remove_ring(struct amdgpu_device *adev, + struct amdgpu_ring *ring); + +int amdgpu_mes_ctx_alloc_meta_data(struct amdgpu_device *adev, + struct amdgpu_mes_ctx_data *ctx_data); +void amdgpu_mes_ctx_free_meta_data(struct amdgpu_mes_ctx_data *ctx_data); +int amdgpu_mes_ctx_map_meta_data(struct amdgpu_device *adev, + struct amdgpu_vm *vm, + struct amdgpu_mes_ctx_data *ctx_data); + +int amdgpu_mes_self_test(struct amdgpu_device *adev); + +int amdgpu_mes_alloc_process_doorbells(struct amdgpu_device *adev, + unsigned int *doorbell_index); +void amdgpu_mes_free_process_doorbells(struct amdgpu_device *adev, + unsigned int doorbell_index); +unsigned int amdgpu_mes_get_doorbell_dw_offset_in_bar( + struct amdgpu_device *adev, + 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__ */ |