diff options
Diffstat (limited to 'drivers/gpu/drm/amd')
64 files changed, 1480 insertions, 1216 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h index 2fc58e658986..6647fb26ef25 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h @@ -82,6 +82,7 @@ extern int amdgpu_vm_block_size; extern int amdgpu_enable_scheduler; extern int amdgpu_sched_jobs; extern int amdgpu_sched_hw_submission; +extern int amdgpu_enable_semaphores; #define AMDGPU_WAIT_IDLE_TIMEOUT_IN_MS 3000 #define AMDGPU_MAX_USEC_TIMEOUT 100000 /* 100 ms */ @@ -98,6 +99,9 @@ extern int amdgpu_sched_hw_submission; #define AMDGPU_MAX_COMPUTE_RINGS 8 #define AMDGPU_MAX_VCE_RINGS 2 +/* max number of IP instances */ +#define AMDGPU_MAX_SDMA_INSTANCES 2 + /* number of hw syncs before falling back on blocking */ #define AMDGPU_NUM_SYNCS 4 @@ -183,6 +187,7 @@ struct amdgpu_vm; struct amdgpu_ring; struct amdgpu_semaphore; struct amdgpu_cs_parser; +struct amdgpu_job; struct amdgpu_irq_src; struct amdgpu_fpriv; @@ -246,7 +251,7 @@ struct amdgpu_buffer_funcs { unsigned copy_num_dw; /* used for buffer migration */ - void (*emit_copy_buffer)(struct amdgpu_ring *ring, + void (*emit_copy_buffer)(struct amdgpu_ib *ib, /* src addr in bytes */ uint64_t src_offset, /* dst addr in bytes */ @@ -261,7 +266,7 @@ struct amdgpu_buffer_funcs { unsigned fill_num_dw; /* used for buffer clearing */ - void (*emit_fill_buffer)(struct amdgpu_ring *ring, + void (*emit_fill_buffer)(struct amdgpu_ib *ib, /* value to write to memory */ uint32_t src_data, /* dst addr in bytes */ @@ -339,6 +344,8 @@ struct amdgpu_ring_funcs { int (*test_ring)(struct amdgpu_ring *ring); int (*test_ib)(struct amdgpu_ring *ring); bool (*is_lockup)(struct amdgpu_ring *ring); + /* insert NOP packets */ + void (*insert_nop)(struct amdgpu_ring *ring, uint32_t count); }; /* @@ -426,7 +433,7 @@ int amdgpu_fence_driver_init(struct amdgpu_device *adev); void amdgpu_fence_driver_fini(struct amdgpu_device *adev); void amdgpu_fence_driver_force_completion(struct amdgpu_device *adev); -void amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring); +int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring); int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring, struct amdgpu_irq_src *irq_src, unsigned irq_type); @@ -440,8 +447,10 @@ int amdgpu_fence_wait_empty(struct amdgpu_ring *ring); unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring); signed long amdgpu_fence_wait_any(struct amdgpu_device *adev, - struct amdgpu_fence **fences, - bool intr, long t); + struct fence **array, + uint32_t count, + bool intr, + signed long t); struct amdgpu_fence *amdgpu_fence_ref(struct amdgpu_fence *fence); void amdgpu_fence_unref(struct amdgpu_fence **fence); @@ -514,7 +523,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t dst_offset, uint32_t byte_count, struct reservation_object *resv, - struct amdgpu_fence **fence); + struct fence **fence); int amdgpu_mmap(struct file *filp, struct vm_area_struct *vma); struct amdgpu_bo_list_entry { @@ -650,7 +659,7 @@ struct amdgpu_sa_bo { struct amdgpu_sa_manager *manager; unsigned soffset; unsigned eoffset; - struct amdgpu_fence *fence; + struct fence *fence; }; /* @@ -692,7 +701,7 @@ bool amdgpu_semaphore_emit_wait(struct amdgpu_ring *ring, struct amdgpu_semaphore *semaphore); void amdgpu_semaphore_free(struct amdgpu_device *adev, struct amdgpu_semaphore **semaphore, - struct amdgpu_fence *fence); + struct fence *fence); /* * Synchronization @@ -700,7 +709,8 @@ void amdgpu_semaphore_free(struct amdgpu_device *adev, struct amdgpu_sync { struct amdgpu_semaphore *semaphores[AMDGPU_NUM_SYNCS]; struct amdgpu_fence *sync_to[AMDGPU_MAX_RINGS]; - struct amdgpu_fence *last_vm_update; + DECLARE_HASHTABLE(fences, 4); + struct fence *last_vm_update; }; void amdgpu_sync_create(struct amdgpu_sync *sync); @@ -712,8 +722,10 @@ int amdgpu_sync_resv(struct amdgpu_device *adev, void *owner); int amdgpu_sync_rings(struct amdgpu_sync *sync, struct amdgpu_ring *ring); +struct fence *amdgpu_sync_get_fence(struct amdgpu_sync *sync); +int amdgpu_sync_wait(struct amdgpu_sync *sync); void amdgpu_sync_free(struct amdgpu_device *adev, struct amdgpu_sync *sync, - struct amdgpu_fence *fence); + struct fence *fence); /* * GART structures, functions & helpers @@ -871,7 +883,7 @@ int amdgpu_sched_ib_submit_kernel_helper(struct amdgpu_device *adev, struct amdgpu_ring *ring, struct amdgpu_ib *ibs, unsigned num_ibs, - int (*free_job)(struct amdgpu_cs_parser *), + int (*free_job)(struct amdgpu_job *), void *owner, struct fence **fence); @@ -879,7 +891,7 @@ struct amdgpu_ring { struct amdgpu_device *adev; const struct amdgpu_ring_funcs *funcs; struct amdgpu_fence_driver fence_drv; - struct amd_gpu_scheduler *scheduler; + struct amd_gpu_scheduler sched; spinlock_t fence_lock; struct mutex *ring_lock; @@ -957,7 +969,7 @@ struct amdgpu_vm_id { unsigned id; uint64_t pd_gpu_addr; /* last flushed PD/PT update */ - struct amdgpu_fence *flushed_updates; + struct fence *flushed_updates; /* last use of vmid */ struct amdgpu_fence *last_id_use; }; @@ -1042,7 +1054,7 @@ struct amdgpu_ctx *amdgpu_ctx_get(struct amdgpu_fpriv *fpriv, uint32_t id); int amdgpu_ctx_put(struct amdgpu_ctx *ctx); uint64_t amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx, struct amdgpu_ring *ring, - struct fence *fence, uint64_t queued_seq); + struct fence *fence); struct fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx, struct amdgpu_ring *ring, uint64_t seq); @@ -1078,8 +1090,6 @@ struct amdgpu_bo_list { }; struct amdgpu_bo_list * -amdgpu_bo_list_clone(struct amdgpu_bo_list *list); -struct amdgpu_bo_list * amdgpu_bo_list_get(struct amdgpu_fpriv *fpriv, int id); void amdgpu_bo_list_put(struct amdgpu_bo_list *list); void amdgpu_bo_list_free(struct amdgpu_bo_list *list); @@ -1192,8 +1202,6 @@ struct amdgpu_gfx { struct amdgpu_irq_src priv_inst_irq; /* gfx status */ uint32_t gfx_current_status; - /* sync signal for const engine */ - unsigned ce_sync_offs; /* ce ram size*/ unsigned ce_ram_size; }; @@ -1210,6 +1218,7 @@ int amdgpu_ib_ring_tests(struct amdgpu_device *adev); void amdgpu_ring_free_size(struct amdgpu_ring *ring); int amdgpu_ring_alloc(struct amdgpu_ring *ring, unsigned ndw); int amdgpu_ring_lock(struct amdgpu_ring *ring, unsigned ndw); +void amdgpu_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count); void amdgpu_ring_commit(struct amdgpu_ring *ring); void amdgpu_ring_unlock_commit(struct amdgpu_ring *ring); void amdgpu_ring_undo(struct amdgpu_ring *ring); @@ -1255,15 +1264,19 @@ struct amdgpu_cs_parser { /* user fence */ struct amdgpu_user_fence uf; +}; - struct amdgpu_ring *ring; - struct mutex job_lock; - struct work_struct job_work; - int (*prepare_job)(struct amdgpu_cs_parser *sched_job); - int (*run_job)(struct amdgpu_cs_parser *sched_job); - int (*free_job)(struct amdgpu_cs_parser *sched_job); - struct amd_sched_fence *s_fence; +struct amdgpu_job { + struct amd_sched_job base; + struct amdgpu_device *adev; + struct amdgpu_ib *ibs; + uint32_t num_ibs; + struct mutex job_lock; + struct amdgpu_user_fence uf; + int (*free_job)(struct amdgpu_job *job); }; +#define to_amdgpu_job(sched_job) \ + container_of((sched_job), struct amdgpu_job, base) static inline u32 amdgpu_get_ib_value(struct amdgpu_cs_parser *p, uint32_t ib_idx, int idx) { @@ -1659,7 +1672,6 @@ struct amdgpu_uvd { struct amdgpu_bo *vcpu_bo; void *cpu_addr; uint64_t gpu_addr; - void *saved_bo; atomic_t handles[AMDGPU_MAX_UVD_HANDLES]; struct drm_file *filp[AMDGPU_MAX_UVD_HANDLES]; struct delayed_work idle_work; @@ -1703,6 +1715,7 @@ struct amdgpu_sdma { uint32_t feature_version; struct amdgpu_ring ring; + bool burst_nop; }; /* @@ -2051,7 +2064,7 @@ struct amdgpu_device { struct amdgpu_gfx gfx; /* sdma */ - struct amdgpu_sdma sdma[2]; + struct amdgpu_sdma sdma[AMDGPU_MAX_SDMA_INSTANCES]; struct amdgpu_irq_src sdma_trap_irq; struct amdgpu_irq_src sdma_illegal_inst_irq; @@ -2190,6 +2203,21 @@ static inline void amdgpu_ring_write(struct amdgpu_ring *ring, uint32_t v) ring->ring_free_dw--; } +static inline struct amdgpu_sdma * amdgpu_get_sdma_instance(struct amdgpu_ring *ring) +{ + struct amdgpu_device *adev = ring->adev; + int i; + + for (i = 0; i < AMDGPU_MAX_SDMA_INSTANCES; i++) + if (&adev->sdma[i].ring == ring) + break; + + if (i < AMDGPU_MAX_SDMA_INSTANCES) + return &adev->sdma[i]; + else + return NULL; +} + /* * ASICs macro. */ @@ -2241,8 +2269,8 @@ static inline void amdgpu_ring_write(struct amdgpu_ring *ring, uint32_t v) #define amdgpu_display_add_connector(adev, ci, sd, ct, ib, coi, h, r) (adev)->mode_info.funcs->add_connector((adev), (ci), (sd), (ct), (ib), (coi), (h), (r)) #define amdgpu_display_stop_mc_access(adev, s) (adev)->mode_info.funcs->stop_mc_access((adev), (s)) #define amdgpu_display_resume_mc_access(adev, s) (adev)->mode_info.funcs->resume_mc_access((adev), (s)) -#define amdgpu_emit_copy_buffer(adev, r, s, d, b) (adev)->mman.buffer_funcs->emit_copy_buffer((r), (s), (d), (b)) -#define amdgpu_emit_fill_buffer(adev, r, s, d, b) (adev)->mman.buffer_funcs->emit_fill_buffer((r), (s), (d), (b)) +#define amdgpu_emit_copy_buffer(adev, ib, s, d, b) (adev)->mman.buffer_funcs->emit_copy_buffer((ib), (s), (d), (b)) +#define amdgpu_emit_fill_buffer(adev, ib, s, d, b) (adev)->mman.buffer_funcs->emit_fill_buffer((ib), (s), (d), (b)) #define amdgpu_dpm_get_temperature(adev) (adev)->pm.funcs->get_temperature((adev)) #define amdgpu_dpm_pre_set_power_state(adev) (adev)->pm.funcs->pre_set_power_state((adev)) #define amdgpu_dpm_set_power_state(adev) (adev)->pm.funcs->set_power_state((adev)) @@ -2343,7 +2371,7 @@ int amdgpu_vm_grab_id(struct amdgpu_vm *vm, struct amdgpu_ring *ring, struct amdgpu_sync *sync); void amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_vm *vm, - struct amdgpu_fence *updates); + struct fence *updates); void amdgpu_vm_fence(struct amdgpu_device *adev, struct amdgpu_vm *vm, struct amdgpu_fence *fence); @@ -2373,7 +2401,7 @@ int amdgpu_vm_bo_unmap(struct amdgpu_device *adev, uint64_t addr); void amdgpu_vm_bo_rmv(struct amdgpu_device *adev, struct amdgpu_bo_va *bo_va); - +int amdgpu_vm_free_job(struct amdgpu_job *job); /* * functions used by amdgpu_encoder.c */ diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c index 496ed2192eba..84d68d658f8a 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c @@ -183,7 +183,7 @@ int alloc_gtt_mem(struct kgd_dev *kgd, size_t size, return -ENOMEM; r = amdgpu_bo_create(rdev, size, PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_GTT, - AMDGPU_GEM_CREATE_CPU_GTT_USWC, NULL, &(*mem)->bo); + AMDGPU_GEM_CREATE_CPU_GTT_USWC, NULL, NULL, &(*mem)->bo); if (r) { dev_err(rdev->dev, "failed to allocate BO for amdkfd (%d)\n", r); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_benchmark.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_benchmark.c index 759482e4300d..cd639c362df3 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_benchmark.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_benchmark.c @@ -33,7 +33,7 @@ static int amdgpu_benchmark_do_move(struct amdgpu_device *adev, unsigned size, { unsigned long start_jiffies; unsigned long end_jiffies; - struct amdgpu_fence *fence = NULL; + struct fence *fence = NULL; int i, r; start_jiffies = jiffies; @@ -42,17 +42,17 @@ static int amdgpu_benchmark_do_move(struct amdgpu_device *adev, unsigned size, r = amdgpu_copy_buffer(ring, saddr, daddr, size, NULL, &fence); if (r) goto exit_do_move; - r = fence_wait(&fence->base, false); + r = fence_wait(fence, false); if (r) goto exit_do_move; - amdgpu_fence_unref(&fence); + fence_put(fence); } end_jiffies = jiffies; r = jiffies_to_msecs(end_jiffies - start_jiffies); exit_do_move: if (fence) - amdgpu_fence_unref(&fence); + fence_put(fence); return r; } @@ -79,7 +79,8 @@ static void amdgpu_benchmark_move(struct amdgpu_device *adev, unsigned size, int time; n = AMDGPU_BENCHMARK_ITERATIONS; - r = amdgpu_bo_create(adev, size, PAGE_SIZE, true, sdomain, 0, NULL, &sobj); + r = amdgpu_bo_create(adev, size, PAGE_SIZE, true, sdomain, 0, NULL, + NULL, &sobj); if (r) { goto out_cleanup; } @@ -91,7 +92,8 @@ static void amdgpu_benchmark_move(struct amdgpu_device *adev, unsigned size, if (r) { goto out_cleanup; } - r = amdgpu_bo_create(adev, size, PAGE_SIZE, true, ddomain, 0, NULL, &dobj); + r = amdgpu_bo_create(adev, size, PAGE_SIZE, true, ddomain, 0, NULL, + NULL, &dobj); if (r) { goto out_cleanup; } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_bo_list.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_bo_list.c index 7eed523bf28f..f82a2dd83874 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_bo_list.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_bo_list.c @@ -62,39 +62,6 @@ static int amdgpu_bo_list_create(struct amdgpu_fpriv *fpriv, return 0; } -struct amdgpu_bo_list * -amdgpu_bo_list_clone(struct amdgpu_bo_list *list) -{ - struct amdgpu_bo_list *result; - unsigned i; - - result = kmalloc(sizeof(struct amdgpu_bo_list), GFP_KERNEL); - if (!result) - return NULL; - - result->array = drm_calloc_large(list->num_entries, - sizeof(struct amdgpu_bo_list_entry)); - if (!result->array) { - kfree(result); - return NULL; - } - - mutex_init(&result->lock); - result->gds_obj = list->gds_obj; - result->gws_obj = list->gws_obj; - result->oa_obj = list->oa_obj; - result->has_userptr = list->has_userptr; - result->num_entries = list->num_entries; - - memcpy(result->array, list->array, list->num_entries * - sizeof(struct amdgpu_bo_list_entry)); - - for (i = 0; i < result->num_entries; ++i) - amdgpu_bo_ref(result->array[i].robj); - - return result; -} - static void amdgpu_bo_list_destroy(struct amdgpu_fpriv *fpriv, int id) { struct amdgpu_bo_list *list; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cgs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cgs.c index 6b1243f9f86d..1c3fc99c5465 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cgs.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cgs.c @@ -86,7 +86,7 @@ static int amdgpu_cgs_gmap_kmem(void *cgs_device, void *kmem, struct sg_table *sg = drm_prime_pages_to_sg(&kmem_page, npages); ret = amdgpu_bo_create(adev, size, PAGE_SIZE, false, - AMDGPU_GEM_DOMAIN_GTT, 0, sg, &bo); + AMDGPU_GEM_DOMAIN_GTT, 0, sg, NULL, &bo); if (ret) return ret; ret = amdgpu_bo_reserve(bo, false); @@ -197,7 +197,8 @@ static int amdgpu_cgs_alloc_gpu_mem(void *cgs_device, ret = amdgpu_bo_create_restricted(adev, size, PAGE_SIZE, true, domain, flags, - NULL, &placement, &obj); + NULL, &placement, NULL, + &obj); if (ret) { DRM_ERROR("(%d) bo create failed\n", ret); return ret; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_connectors.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_connectors.c index 27df17a0e620..89c3dd62ba21 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_connectors.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_connectors.c @@ -75,6 +75,11 @@ void amdgpu_connector_hotplug(struct drm_connector *connector) if (!amdgpu_display_hpd_sense(adev, amdgpu_connector->hpd.hpd)) { drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF); } else if (amdgpu_atombios_dp_needs_link_train(amdgpu_connector)) { + /* Don't try to start link training before we + * have the dpcd */ + if (!amdgpu_atombios_dp_get_dpcd(amdgpu_connector)) + return; + /* set it to OFF so that drm_helper_connector_dpms() * won't return immediately since the current state * is ON at this point. diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c index e4424b4db5d3..749420f1ea6f 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c @@ -126,19 +126,6 @@ int amdgpu_cs_get_ring(struct amdgpu_device *adev, u32 ip_type, return 0; } -static void amdgpu_job_work_func(struct work_struct *work) -{ - struct amdgpu_cs_parser *sched_job = - container_of(work, struct amdgpu_cs_parser, - job_work); - mutex_lock(&sched_job->job_lock); - if (sched_job->free_job) - sched_job->free_job(sched_job); - mutex_unlock(&sched_job->job_lock); - /* after processing job, free memory */ - fence_put(&sched_job->s_fence->base); - kfree(sched_job); -} struct amdgpu_cs_parser *amdgpu_cs_parser_create(struct amdgpu_device *adev, struct drm_file *filp, struct amdgpu_ctx *ctx, @@ -157,10 +144,6 @@ struct amdgpu_cs_parser *amdgpu_cs_parser_create(struct amdgpu_device *adev, parser->ctx = ctx; parser->ibs = ibs; parser->num_ibs = num_ibs; - if (amdgpu_enable_scheduler) { - mutex_init(&parser->job_lock); - INIT_WORK(&parser->job_work, amdgpu_job_work_func); - } for (i = 0; i < num_ibs; i++) ibs[i].ctx = ctx; @@ -171,56 +154,41 @@ int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data) { union drm_amdgpu_cs *cs = data; uint64_t *chunk_array_user; - uint64_t *chunk_array = NULL; + uint64_t *chunk_array; struct amdgpu_fpriv *fpriv = p->filp->driver_priv; - struct amdgpu_bo_list *bo_list = NULL; unsigned size, i; - int r = 0; + int ret; - if (!cs->in.num_chunks) - goto out; + if (cs->in.num_chunks == 0) + return 0; + + chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL); + if (!chunk_array) + return -ENOMEM; p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id); if (!p->ctx) { - r = -EINVAL; - goto out; - } - bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle); - if (!amdgpu_enable_scheduler) - p->bo_list = bo_list; - else { - if (bo_list && !bo_list->has_userptr) { - p->bo_list = amdgpu_bo_list_clone(bo_list); - amdgpu_bo_list_put(bo_list); - if (!p->bo_list) - return -ENOMEM; - } else if (bo_list && bo_list->has_userptr) - p->bo_list = bo_list; - else - p->bo_list = NULL; + ret = -EINVAL; + goto free_chunk; } + p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle); + /* get chunks */ INIT_LIST_HEAD(&p->validated); - chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL); - if (chunk_array == NULL) { - r = -ENOMEM; - goto out; - } - chunk_array_user = (uint64_t __user *)(cs->in.chunks); if (copy_from_user(chunk_array, chunk_array_user, sizeof(uint64_t)*cs->in.num_chunks)) { - r = -EFAULT; - goto out; + ret = -EFAULT; + goto put_bo_list; } p->nchunks = cs->in.num_chunks; p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk), GFP_KERNEL); - if (p->chunks == NULL) { - r = -ENOMEM; - goto out; + if (!p->chunks) { + ret = -ENOMEM; + goto put_bo_list; } for (i = 0; i < p->nchunks; i++) { @@ -231,8 +199,9 @@ int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data) chunk_ptr = (void __user *)chunk_array[i]; if (copy_from_user(&user_chunk, chunk_ptr, sizeof(struct drm_amdgpu_cs_chunk))) { - r = -EFAULT; - goto out; + ret = -EFAULT; + i--; + goto free_partial_kdata; } p->chunks[i].chunk_id = user_chunk.chunk_id; p->chunks[i].length_dw = user_chunk.length_dw; @@ -243,13 +212,14 @@ int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data) p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t)); if (p->chunks[i].kdata == NULL) { - r = -ENOMEM; - goto out; + ret = -ENOMEM; + i--; + goto free_partial_kdata; } size *= sizeof(uint32_t); if (copy_from_user(p->chunks[i].kdata, cdata, size)) { - r = -EFAULT; - goto out; + ret = -EFAULT; + goto free_partial_kdata; } switch (p->chunks[i].chunk_id) { @@ -269,15 +239,15 @@ int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data) gobj = drm_gem_object_lookup(p->adev->ddev, p->filp, handle); if (gobj == NULL) { - r = -EINVAL; - goto out; + ret = -EINVAL; + goto free_partial_kdata; } p->uf.bo = gem_to_amdgpu_bo(gobj); p->uf.offset = fence_data->offset; } else { - r = -EINVAL; - goto out; + ret = -EINVAL; + goto free_partial_kdata; } break; @@ -285,19 +255,35 @@ int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data) break; default: - r = -EINVAL; - goto out; + ret = -EINVAL; + goto free_partial_kdata; } } - p->ibs = kmalloc_array(p->num_ibs, sizeof(struct amdgpu_ib), GFP_KERNEL); - if (!p->ibs) - r = -ENOMEM; + p->ibs = kcalloc(p->num_ibs, sizeof(struct amdgpu_ib), GFP_KERNEL); + if (!p->ibs) { + ret = -ENOMEM; + goto free_all_kdata; + } -out: kfree(chunk_array); - return r; + return 0; + +free_all_kdata: + i = p->nchunks - 1; +free_partial_kdata: + for (; i >= 0; i--) + drm_free_large(p->chunks[i].kdata); + kfree(p->chunks); +put_bo_list: + if (p->bo_list) + amdgpu_bo_list_put(p->bo_list); + amdgpu_ctx_put(p->ctx); +free_chunk: + kfree(chunk_array); + + return ret; } /* Returns how many bytes TTM can move per IB. @@ -352,25 +338,17 @@ static u64 amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev) return max(bytes_moved_threshold, 1024*1024ull); } -int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p) +int amdgpu_cs_list_validate(struct amdgpu_device *adev, + struct amdgpu_vm *vm, + struct list_head *validated) { - struct amdgpu_fpriv *fpriv = p->filp->driver_priv; - struct amdgpu_vm *vm = &fpriv->vm; - struct amdgpu_device *adev = p->adev; struct amdgpu_bo_list_entry *lobj; - struct list_head duplicates; struct amdgpu_bo *bo; u64 bytes_moved = 0, initial_bytes_moved; u64 bytes_moved_threshold = amdgpu_cs_get_threshold_for_moves(adev); int r; - INIT_LIST_HEAD(&duplicates); - r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true, &duplicates); - if (unlikely(r != 0)) { - return r; - } - - list_for_each_entry(lobj, &p->validated, tv.head) { + list_for_each_entry(lobj, validated, tv.head) { bo = lobj->robj; if (!bo->pin_count) { u32 domain = lobj->prefered_domains; @@ -385,7 +363,7 @@ int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p) * into account. We don't want to disallow buffer moves * completely. */ - if (current_domain != AMDGPU_GEM_DOMAIN_CPU && + if ((lobj->allowed_domains & current_domain) != 0 && (domain & current_domain) == 0 && /* will be moved */ bytes_moved > bytes_moved_threshold) { /* don't move it */ @@ -404,7 +382,6 @@ int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p) domain = lobj->allowed_domains; goto retry; } - ttm_eu_backoff_reservation(&p->ticket, &p->validated); return r; } } @@ -417,6 +394,7 @@ static int amdgpu_cs_parser_relocs(struct amdgpu_cs_parser *p) { struct amdgpu_fpriv *fpriv = p->filp->driver_priv; struct amdgpu_cs_buckets buckets; + struct list_head duplicates; bool need_mmap_lock = false; int i, r; @@ -436,8 +414,22 @@ static int amdgpu_cs_parser_relocs(struct amdgpu_cs_parser *p) if (need_mmap_lock) down_read(¤t->mm->mmap_sem); - r = amdgpu_cs_list_validate(p); + INIT_LIST_HEAD(&duplicates); + r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true, &duplicates); + if (unlikely(r != 0)) + goto error_reserve; + + r = amdgpu_cs_list_validate(p->adev, &fpriv->vm, &p->validated); + if (r) + goto error_validate; + r = amdgpu_cs_list_validate(p->adev, &fpriv->vm, &duplicates); + +error_validate: + if (r) + ttm_eu_backoff_reservation(&p->ticket, &p->validated); + +error_reserve: if (need_mmap_lock) up_read(¤t->mm->mmap_sem); @@ -498,25 +490,24 @@ static void amdgpu_cs_parser_fini_late(struct amdgpu_cs_parser *parser) unsigned i; if (parser->ctx) amdgpu_ctx_put(parser->ctx); - if (parser->bo_list) { - if (amdgpu_enable_scheduler && !parser->bo_list->has_userptr) - amdgpu_bo_list_free(parser->bo_list); - else - amdgpu_bo_list_put(parser->bo_list); - } + if (parser->bo_list) + amdgpu_bo_list_put(parser->bo_list); + drm_free_large(parser->vm_bos); for (i = 0; i < parser->nchunks; i++) drm_free_large(parser->chunks[i].kdata); kfree(parser->chunks); - if (parser->ibs) - for (i = 0; i < parser->num_ibs; i++) - amdgpu_ib_free(parser->adev, &parser->ibs[i]); - kfree(parser->ibs); - if (parser->uf.bo) - drm_gem_object_unreference_unlocked(&parser->uf.bo->gem_base); - if (!amdgpu_enable_scheduler) - kfree(parser); + { + if (parser->ibs) + for (i = 0; i < parser->num_ibs; i++) + amdgpu_ib_free(parser->adev, &parser->ibs[i]); + kfree(parser->ibs); + if (parser->uf.bo) + drm_gem_object_unreference_unlocked(&parser->uf.bo->gem_base); + } + + kfree(parser); } /** @@ -533,12 +524,6 @@ static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bo amdgpu_cs_parser_fini_late(parser); } -static int amdgpu_cs_parser_free_job(struct amdgpu_cs_parser *sched_job) -{ - amdgpu_cs_parser_fini_late(sched_job); - return 0; -} - static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p, struct amdgpu_vm *vm) { @@ -810,68 +795,16 @@ static int amdgpu_cs_dependencies(struct amdgpu_device *adev, return 0; } -static int amdgpu_cs_parser_prepare_job(struct amdgpu_cs_parser *sched_job) +static int amdgpu_cs_free_job(struct amdgpu_job *job) { - int r, i; - struct amdgpu_cs_parser *parser = sched_job; - struct amdgpu_device *adev = sched_job->adev; - bool reserved_buffers = false; - - r = amdgpu_cs_parser_relocs(parser); - if (r) { - if (r != -ERESTARTSYS) { - if (r == -ENOMEM) - DRM_ERROR("Not enough memory for command submission!\n"); - else - DRM_ERROR("Failed to process the buffer list %d!\n", r); - } - } - - if (!r) { - reserved_buffers = true; - r = amdgpu_cs_ib_fill(adev, parser); - } - if (!r) { - r = amdgpu_cs_dependencies(adev, parser); - if (r) - DRM_ERROR("Failed in the dependencies handling %d!\n", r); - } - if (r) { - amdgpu_cs_parser_fini(parser, r, reserved_buffers); - return r; - } - - for (i = 0; i < parser->num_ibs; i++) - trace_amdgpu_cs(parser, i); - - r = amdgpu_cs_ib_vm_chunk(adev, parser); - return r; -} - -static struct amdgpu_ring *amdgpu_cs_parser_get_ring( - struct amdgpu_device *adev, - struct amdgpu_cs_parser *parser) -{ - int i, r; - - struct amdgpu_cs_chunk *chunk; - struct drm_amdgpu_cs_chunk_ib *chunk_ib; - struct amdgpu_ring *ring; - for (i = 0; i < parser->nchunks; i++) { - chunk = &parser->chunks[i]; - chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata; - - if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB) - continue; - - r = amdgpu_cs_get_ring(adev, chunk_ib->ip_type, - chunk_ib->ip_instance, chunk_ib->ring, - &ring); - if (r) - return NULL; - break; - } - return ring; + int i; + if (job->ibs) + for (i = 0; i < job->num_ibs; i++) + amdgpu_ib_free(job->adev, &job->ibs[i]); + kfree(job->ibs); + if (job->uf.bo) + drm_gem_object_unreference_unlocked(&job->uf.bo->gem_base); + return 0; } int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) @@ -879,7 +812,8 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) struct amdgpu_device *adev = dev->dev_private; union drm_amdgpu_cs *cs = data; struct amdgpu_cs_parser *parser; - int r; + bool reserved_buffers = false; + int i, r; down_read(&adev->exclusive_lock); if (!adev->accel_working) { @@ -893,50 +827,85 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) r = amdgpu_cs_parser_init(parser, data); if (r) { DRM_ERROR("Failed to initialize parser !\n"); - amdgpu_cs_parser_fini(parser, r, false); + kfree(parser); up_read(&adev->exclusive_lock); r = amdgpu_cs_handle_lockup(adev, r); return r; } - if (amdgpu_enable_scheduler && parser->num_ibs) { - struct amdgpu_ring * ring = - amdgpu_cs_parser_get_ring(adev, parser); - r = amdgpu_cs_parser_prepare_job(parser); + r = amdgpu_cs_parser_relocs(parser); + if (r == -ENOMEM) + DRM_ERROR("Not enough memory for command submission!\n"); + else if (r && r != -ERESTARTSYS) + DRM_ERROR("Failed to process the buffer list %d!\n", r); + else if (!r) { + reserved_buffers = true; + r = amdgpu_cs_ib_fill(adev, parser); + } + + if (!r) { + r = amdgpu_cs_dependencies(adev, parser); if (r) - goto out; - parser->ring = ring; - parser->free_job = amdgpu_cs_parser_free_job; - mutex_lock(&parser->job_lock); - r = amd_sched_push_job(ring->scheduler, - &parser->ctx->rings[ring->idx].entity, - parser, - &parser->s_fence); + DRM_ERROR("Failed in the dependencies handling %d!\n", r); + } + + if (r) + goto out; + + for (i = 0; i < parser->num_ibs; i++) + trace_amdgpu_cs(parser, i); + + r = amdgpu_cs_ib_vm_chunk(adev, parser); + if (r) + goto out; + + if (amdgpu_enable_scheduler && parser->num_ibs) { + struct amdgpu_job *job; + struct amdgpu_ring * ring = parser->ibs->ring; + job = kzalloc(sizeof(struct amdgpu_job), GFP_KERNEL); + if (!job) + return -ENOMEM; + job->base.sched = &ring->sched; + job->base.s_entity = &parser->ctx->rings[ring->idx].entity; + job->adev = parser->adev; + job->ibs = parser->ibs; + job->num_ibs = parser->num_ibs; + job->base.owner = parser->filp; + mutex_init(&job->job_lock); + if (job->ibs[job->num_ibs - 1].user) { + memcpy(&job->uf, &parser->uf, + sizeof(struct amdgpu_user_fence)); + job->ibs[job->num_ibs - 1].user = &job->uf; + } + + job->free_job = amdgpu_cs_free_job; + mutex_lock(&job->job_lock); + r = amd_sched_entity_push_job(&job->base); if (r) { - mutex_unlock(&parser->job_lock); + mutex_unlock(&job->job_lock); + amdgpu_cs_free_job(job); + kfree(job); goto out; } - parser->ibs[parser->num_ibs - 1].sequence = + cs->out.handle = amdgpu_ctx_add_fence(parser->ctx, ring, - &parser->s_fence->base, - parser->s_fence->v_seq); - cs->out.handle = parser->s_fence->v_seq; + &job->base.s_fence->base); + parser->ibs[parser->num_ibs - 1].sequence = cs->out.handle; + list_sort(NULL, &parser->validated, cmp_size_smaller_first); ttm_eu_fence_buffer_objects(&parser->ticket, &parser->validated, - &parser->s_fence->base); + &job->base.s_fence->base); - mutex_unlock(&parser->job_lock); + mutex_unlock(&job->job_lock); + amdgpu_cs_parser_fini_late(parser); up_read(&adev->exclusive_lock); return 0; } - r = amdgpu_cs_parser_prepare_job(parser); - if (r) - goto out; cs->out.handle = parser->ibs[parser->num_ibs - 1].sequence; out: - amdgpu_cs_parser_fini(parser, r, true); + amdgpu_cs_parser_fini(parser, r, reserved_buffers); up_read(&adev->exclusive_lock); r = amdgpu_cs_handle_lockup(adev, r); return r; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c index 08bc7722ddb8..e0b80ccdfe8a 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c @@ -43,10 +43,10 @@ int amdgpu_ctx_init(struct amdgpu_device *adev, bool kernel, for (i = 0; i < adev->num_rings; i++) { struct amd_sched_rq *rq; if (kernel) - rq = &adev->rings[i]->scheduler->kernel_rq; + rq = &adev->rings[i]->sched.kernel_rq; else - rq = &adev->rings[i]->scheduler->sched_rq; - r = amd_sched_entity_init(adev->rings[i]->scheduler, + rq = &adev->rings[i]->sched.sched_rq; + r = amd_sched_entity_init(&adev->rings[i]->sched, &ctx->rings[i].entity, rq, amdgpu_sched_jobs); if (r) @@ -55,7 +55,7 @@ int amdgpu_ctx_init(struct amdgpu_device *adev, bool kernel, if (i < adev->num_rings) { for (j = 0; j < i; j++) - amd_sched_entity_fini(adev->rings[j]->scheduler, + amd_sched_entity_fini(&adev->rings[j]->sched, &ctx->rings[j].entity); kfree(ctx); return r; @@ -75,7 +75,7 @@ void amdgpu_ctx_fini(struct amdgpu_ctx *ctx) if (amdgpu_enable_scheduler) { for (i = 0; i < adev->num_rings; i++) - amd_sched_entity_fini(adev->rings[i]->scheduler, + amd_sched_entity_fini(&adev->rings[i]->sched, &ctx->rings[i].entity); } } @@ -229,17 +229,13 @@ int amdgpu_ctx_put(struct amdgpu_ctx *ctx) } uint64_t amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx, struct amdgpu_ring *ring, - struct fence *fence, uint64_t queued_seq) + struct fence *fence) { struct amdgpu_ctx_ring *cring = & ctx->rings[ring->idx]; - uint64_t seq = 0; + uint64_t seq = cring->sequence; unsigned idx = 0; struct fence *other = NULL; - if (amdgpu_enable_scheduler) - seq = queued_seq; - else - seq = cring->sequence; idx = seq % AMDGPU_CTX_MAX_CS_PENDING; other = cring->fences[idx]; if (other) { @@ -253,8 +249,7 @@ uint64_t amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx, struct amdgpu_ring *ring, spin_lock(&ctx->ring_lock); cring->fences[idx] = fence; - if (!amdgpu_enable_scheduler) - cring->sequence++; + cring->sequence++; spin_unlock(&ctx->ring_lock); fence_put(other); @@ -267,21 +262,16 @@ struct fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx, { struct amdgpu_ctx_ring *cring = & ctx->rings[ring->idx]; struct fence *fence; - uint64_t queued_seq; spin_lock(&ctx->ring_lock); - if (amdgpu_enable_scheduler) - queued_seq = amd_sched_next_queued_seq(&cring->entity); - else - queued_seq = cring->sequence; - if (seq >= queued_seq) { + if (seq >= cring->sequence) { spin_unlock(&ctx->ring_lock); return ERR_PTR(-EINVAL); } - if (seq + AMDGPU_CTX_MAX_CS_PENDING < queued_seq) { + if (seq + AMDGPU_CTX_MAX_CS_PENDING < cring->sequence) { spin_unlock(&ctx->ring_lock); return NULL; } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c index 42d1a22c1199..6068d8207d10 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c @@ -244,8 +244,9 @@ static int amdgpu_vram_scratch_init(struct amdgpu_device *adev) if (adev->vram_scratch.robj == NULL) { r = amdgpu_bo_create(adev, AMDGPU_GPU_PAGE_SIZE, - PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_VRAM, 0, - NULL, &adev->vram_scratch.robj); + PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_VRAM, + AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED, + NULL, NULL, &adev->vram_scratch.robj); if (r) { return r; } @@ -448,7 +449,8 @@ static int amdgpu_wb_init(struct amdgpu_device *adev) if (adev->wb.wb_obj == NULL) { r = amdgpu_bo_create(adev, AMDGPU_MAX_WB * 4, PAGE_SIZE, true, - AMDGPU_GEM_DOMAIN_GTT, 0, NULL, &adev->wb.wb_obj); + AMDGPU_GEM_DOMAIN_GTT, 0, NULL, NULL, + &adev->wb.wb_obj); if (r) { dev_warn(adev->dev, "(%d) create WB bo failed\n", r); return r; @@ -1649,9 +1651,11 @@ int amdgpu_suspend_kms(struct drm_device *dev, bool suspend, bool fbcon) drm_kms_helper_poll_disable(dev); /* turn off display hw */ + drm_modeset_lock_all(dev); list_for_each_entry(connector, &dev->mode_config.connector_list, head) { drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF); } + drm_modeset_unlock_all(dev); /* unpin the front buffers */ list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) { @@ -1746,9 +1750,11 @@ int amdgpu_resume_kms(struct drm_device *dev, bool resume, bool fbcon) if (fbcon) { drm_helper_resume_force_mode(dev); /* turn on display hw */ + drm_modeset_lock_all(dev); list_for_each_entry(connector, &dev->mode_config.connector_list, head) { drm_helper_connector_dpms(connector, DRM_MODE_DPMS_ON); } + drm_modeset_unlock_all(dev); } drm_kms_helper_poll_enable(dev); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_display.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_display.c index e3d70772b531..9b34a3410c32 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_display.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_display.c @@ -745,7 +745,8 @@ bool amdgpu_crtc_scaling_mode_fixup(struct drm_crtc *crtc, * */ int amdgpu_get_crtc_scanoutpos(struct drm_device *dev, int crtc, unsigned int flags, - int *vpos, int *hpos, ktime_t *stime, ktime_t *etime) + int *vpos, int *hpos, ktime_t *stime, ktime_t *etime, + const struct drm_display_mode *mode) { u32 vbl = 0, position = 0; int vbl_start, vbl_end, vtotal, ret = 0; @@ -781,7 +782,7 @@ int amdgpu_get_crtc_scanoutpos(struct drm_device *dev, int crtc, unsigned int fl } else { /* No: Fake something reasonable which gives at least ok results. */ - vbl_start = adev->mode_info.crtcs[crtc]->base.hwmode.crtc_vdisplay; + vbl_start = mode->crtc_vdisplay; vbl_end = 0; } @@ -797,7 +798,7 @@ int amdgpu_get_crtc_scanoutpos(struct drm_device *dev, int crtc, unsigned int fl /* Inside "upper part" of vblank area? Apply corrective offset if so: */ if (in_vbl && (*vpos >= vbl_start)) { - vtotal = adev->mode_info.crtcs[crtc]->base.hwmode.crtc_vtotal; + vtotal = mode->crtc_vtotal; *vpos = *vpos - vtotal; } @@ -819,8 +820,8 @@ int amdgpu_get_crtc_scanoutpos(struct drm_device *dev, int crtc, unsigned int fl * We only do this if DRM_CALLED_FROM_VBLIRQ. */ if ((flags & DRM_CALLED_FROM_VBLIRQ) && !in_vbl) { - vbl_start = adev->mode_info.crtcs[crtc]->base.hwmode.crtc_vdisplay; - vtotal = adev->mode_info.crtcs[crtc]->base.hwmode.crtc_vtotal; + vbl_start = mode->crtc_vdisplay; + vtotal = mode->crtc_vtotal; if (vbl_start - *vpos < vtotal / 100) { *vpos -= vtotal; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c index e6fa27805207..adb48353f2e1 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c @@ -49,9 +49,10 @@ /* * KMS wrapper. * - 3.0.0 - initial driver + * - 3.1.0 - allow reading more status registers (GRBM, SRBM, SDMA, CP) */ #define KMS_DRIVER_MAJOR 3 -#define KMS_DRIVER_MINOR 0 +#define KMS_DRIVER_MINOR 1 #define KMS_DRIVER_PATCHLEVEL 0 int amdgpu_vram_limit = 0; @@ -78,6 +79,7 @@ int amdgpu_exp_hw_support = 0; int amdgpu_enable_scheduler = 0; int amdgpu_sched_jobs = 16; int amdgpu_sched_hw_submission = 2; +int amdgpu_enable_semaphores = 1; MODULE_PARM_DESC(vramlimit, "Restrict VRAM for testing, in megabytes"); module_param_named(vramlimit, amdgpu_vram_limit, int, 0600); @@ -151,6 +153,9 @@ module_param_named(sched_jobs, amdgpu_sched_jobs, int, 0444); MODULE_PARM_DESC(sched_hw_submission, "the max number of HW submissions (default 2)"); module_param_named(sched_hw_submission, amdgpu_sched_hw_submission, int, 0444); +MODULE_PARM_DESC(enable_semaphores, "Enable semaphores (1 = enable (default), 0 = disable)"); +module_param_named(enable_semaphores, amdgpu_enable_semaphores, int, 0644); + static struct pci_device_id pciidlist[] = { #ifdef CONFIG_DRM_AMDGPU_CIK /* Kaveri */ diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fb.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fb.c index 81b821247dde..8a122b1b7786 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fb.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fb.c @@ -126,8 +126,8 @@ static int amdgpufb_create_pinned_object(struct amdgpu_fbdev *rfbdev, aligned_size = ALIGN(size, PAGE_SIZE); ret = amdgpu_gem_object_create(adev, aligned_size, 0, AMDGPU_GEM_DOMAIN_VRAM, - 0, true, - &gobj); + AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED, + true, &gobj); if (ret) { printk(KERN_ERR "failed to allocate framebuffer (%d)\n", aligned_size); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index 98500f1756f7..b3fc26c59787 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -609,9 +609,9 @@ int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring, * Init the fence driver for the requested ring (all asics). * Helper function for amdgpu_fence_driver_init(). */ -void amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring) +int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring) { - int i; + int i, r; ring->fence_drv.cpu_addr = NULL; ring->fence_drv.gpu_addr = 0; @@ -625,15 +625,19 @@ void amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring) amdgpu_fence_check_lockup); ring->fence_drv.ring = ring; + init_waitqueue_head(&ring->fence_drv.fence_queue); + if (amdgpu_enable_scheduler) { - ring->scheduler = amd_sched_create((void *)ring->adev, - &amdgpu_sched_ops, - ring->idx, 5, 0, - amdgpu_sched_hw_submission); - if (!ring->scheduler) - DRM_ERROR("Failed to create scheduler on ring %d.\n", - ring->idx); + r = amd_sched_init(&ring->sched, &amdgpu_sched_ops, + amdgpu_sched_hw_submission, ring->name); + if (r) { + DRM_ERROR("Failed to create scheduler on ring %s.\n", + ring->name); + return r; + } } + + return 0; } /** @@ -681,8 +685,7 @@ void amdgpu_fence_driver_fini(struct amdgpu_device *adev) wake_up_all(&ring->fence_drv.fence_queue); amdgpu_irq_put(adev, ring->fence_drv.irq_src, ring->fence_drv.irq_type); - if (ring->scheduler) - amd_sched_destroy(ring->scheduler); + amd_sched_fini(&ring->sched); ring->fence_drv.initialized = false; } mutex_unlock(&adev->ring_lock); @@ -836,16 +839,15 @@ static inline bool amdgpu_test_signaled(struct amdgpu_fence *fence) return test_bit(FENCE_FLAG_SIGNALED_BIT, &fence->base.flags); } -static inline bool amdgpu_test_signaled_any(struct amdgpu_fence **fences) +static bool amdgpu_test_signaled_any(struct fence **fences, uint32_t count) { int idx; - struct amdgpu_fence *fence; + struct fence *fence; - idx = 0; - for (idx = 0; idx < AMDGPU_MAX_RINGS; ++idx) { + for (idx = 0; idx < count; ++idx) { fence = fences[idx]; if (fence) { - if (test_bit(FENCE_FLAG_SIGNALED_BIT, &fence->base.flags)) + if (test_bit(FENCE_FLAG_SIGNALED_BIT, &fence->flags)) return true; } } @@ -867,33 +869,48 @@ static void amdgpu_fence_wait_cb(struct fence *fence, struct fence_cb *cb) static signed long amdgpu_fence_default_wait(struct fence *f, bool intr, signed long t) { - struct amdgpu_fence *array[AMDGPU_MAX_RINGS]; struct amdgpu_fence *fence = to_amdgpu_fence(f); struct amdgpu_device *adev = fence->ring->adev; - memset(&array[0], 0, sizeof(array)); - array[0] = fence; - - return amdgpu_fence_wait_any(adev, array, intr, t); + return amdgpu_fence_wait_any(adev, &f, 1, intr, t); } -/* wait until any fence in array signaled */ +/** + * Wait the fence array with timeout + * + * @adev: amdgpu device + * @array: the fence array with amdgpu fence pointer + * @count: the number of the fence array + * @intr: when sleep, set the current task interruptable or not + * @t: timeout to wait + * + * It will return when any fence is signaled or timeout. + */ signed long amdgpu_fence_wait_any(struct amdgpu_device *adev, - struct amdgpu_fence **array, bool intr, signed long t) + struct fence **array, uint32_t count, + bool intr, signed long t) { - long idx = 0; - struct amdgpu_wait_cb cb[AMDGPU_MAX_RINGS]; - struct amdgpu_fence *fence; + struct amdgpu_wait_cb *cb; + struct fence *fence; + unsigned idx; BUG_ON(!array); - for (idx = 0; idx < AMDGPU_MAX_RINGS; ++idx) { + cb = kcalloc(count, sizeof(struct amdgpu_wait_cb), GFP_KERNEL); + if (cb == NULL) { + t = -ENOMEM; + goto err_free_cb; + } + + for (idx = 0; idx < count; ++idx) { fence = array[idx]; if (fence) { cb[idx].task = current; - if (fence_add_callback(&fence->base, - &cb[idx].base, amdgpu_fence_wait_cb)) - return t; /* return if fence is already signaled */ + if (fence_add_callback(fence, + &cb[idx].base, amdgpu_fence_wait_cb)) { + /* The fence is already signaled */ + goto fence_rm_cb; + } } } @@ -907,7 +924,7 @@ signed long amdgpu_fence_wait_any(struct amdgpu_device *adev, * amdgpu_test_signaled_any must be called after * set_current_state to prevent a race with wake_up_process */ - if (amdgpu_test_signaled_any(array)) + if (amdgpu_test_signaled_any(array, count)) break; if (adev->needs_reset) { @@ -923,13 +940,16 @@ signed long amdgpu_fence_wait_any(struct amdgpu_device *adev, __set_current_state(TASK_RUNNING); - idx = 0; - for (idx = 0; idx < AMDGPU_MAX_RINGS; ++idx) { +fence_rm_cb: + for (idx = 0; idx < count; ++idx) { fence = array[idx]; - if (fence) - fence_remove_callback(&fence->base, &cb[idx].base); + if (fence && cb[idx].base.func) + fence_remove_callback(fence, &cb[idx].base); } +err_free_cb: + kfree(cb); + return t; } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c index e02db0b2e839..7312d729d300 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c @@ -125,8 +125,9 @@ int amdgpu_gart_table_vram_alloc(struct amdgpu_device *adev) if (adev->gart.robj == NULL) { r = amdgpu_bo_create(adev, adev->gart.table_size, - PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_VRAM, 0, - NULL, &adev->gart.robj); + PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_VRAM, + AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED, + NULL, NULL, &adev->gart.robj); if (r) { return r; } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c index 4afc507820c0..7297ca3a0ba7 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c @@ -69,7 +69,8 @@ int amdgpu_gem_object_create(struct amdgpu_device *adev, unsigned long size, } } retry: - r = amdgpu_bo_create(adev, size, alignment, kernel, initial_domain, flags, NULL, &robj); + r = amdgpu_bo_create(adev, size, alignment, kernel, initial_domain, + flags, NULL, NULL, &robj); if (r) { if (r != -ERESTARTSYS) { if (initial_domain == AMDGPU_GEM_DOMAIN_VRAM) { @@ -426,6 +427,10 @@ int amdgpu_gem_metadata_ioctl(struct drm_device *dev, void *data, &args->data.data_size_bytes, &args->data.flags); } else if (args->op == AMDGPU_GEM_METADATA_OP_SET_METADATA) { + if (args->data.data_size_bytes > sizeof(args->data.data)) { + r = -EINVAL; + goto unreserve; + } r = amdgpu_bo_set_tiling_flags(robj, args->data.tiling_info); if (!r) r = amdgpu_bo_set_metadata(robj, args->data.data, @@ -433,6 +438,7 @@ int amdgpu_gem_metadata_ioctl(struct drm_device *dev, void *data, args->data.flags); } +unreserve: amdgpu_bo_unreserve(robj); out: drm_gem_object_unreference_unlocked(gobj); @@ -454,11 +460,12 @@ static void amdgpu_gem_va_update_vm(struct amdgpu_device *adev, struct ttm_validate_buffer tv, *entry; struct amdgpu_bo_list_entry *vm_bos; struct ww_acquire_ctx ticket; - struct list_head list; + struct list_head list, duplicates; unsigned domain; int r; INIT_LIST_HEAD(&list); + INIT_LIST_HEAD(&duplicates); tv.bo = &bo_va->bo->tbo; tv.shared = true; @@ -468,7 +475,8 @@ static void amdgpu_gem_va_update_vm(struct amdgpu_device *adev, if (!vm_bos) return; - r = ttm_eu_reserve_buffers(&ticket, &list, true, NULL); + /* Provide duplicates to avoid -EALREADY */ + r = ttm_eu_reserve_buffers(&ticket, &list, true, &duplicates); if (r) goto error_free; @@ -615,6 +623,7 @@ int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data, info.alignment = robj->tbo.mem.page_alignment << PAGE_SHIFT; info.domains = robj->initial_domain; info.domain_flags = robj->flags; + amdgpu_bo_unreserve(robj); if (copy_to_user(out, &info, sizeof(info))) r = -EFAULT; break; @@ -622,17 +631,19 @@ int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data, case AMDGPU_GEM_OP_SET_PLACEMENT: if (amdgpu_ttm_tt_has_userptr(robj->tbo.ttm)) { r = -EPERM; + amdgpu_bo_unreserve(robj); break; } robj->initial_domain = args->value & (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT | AMDGPU_GEM_DOMAIN_CPU); + amdgpu_bo_unreserve(robj); break; default: + amdgpu_bo_unreserve(robj); r = -EINVAL; } - amdgpu_bo_unreserve(robj); out: drm_gem_object_unreference_unlocked(gobj); return r; @@ -648,12 +659,13 @@ int amdgpu_mode_dumb_create(struct drm_file *file_priv, int r; args->pitch = amdgpu_align_pitch(adev, args->width, args->bpp, 0) * ((args->bpp + 1) / 8); - args->size = args->pitch * args->height; + args->size = (u64)args->pitch * args->height; args->size = ALIGN(args->size, PAGE_SIZE); r = amdgpu_gem_object_create(adev, args->size, 0, AMDGPU_GEM_DOMAIN_VRAM, - 0, ttm_bo_type_device, + AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED, + ttm_bo_type_device, &gobj); if (r) return -ENOMEM; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c index 5104e64e9ad8..c439735ee670 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c @@ -73,29 +73,12 @@ int amdgpu_ib_get(struct amdgpu_ring *ring, struct amdgpu_vm *vm, if (!vm) ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo); - else - ib->gpu_addr = 0; - - } else { - ib->sa_bo = NULL; - ib->ptr = NULL; - ib->gpu_addr = 0; } amdgpu_sync_create(&ib->sync); ib->ring = ring; - ib->fence = NULL; - ib->user = NULL; ib->vm = vm; - ib->ctx = NULL; - ib->gds_base = 0; - ib->gds_size = 0; - ib->gws_base = 0; - ib->gws_size = 0; - ib->oa_base = 0; - ib->oa_size = 0; - ib->flags = 0; return 0; } @@ -110,8 +93,8 @@ int amdgpu_ib_get(struct amdgpu_ring *ring, struct amdgpu_vm *vm, */ void amdgpu_ib_free(struct amdgpu_device *adev, struct amdgpu_ib *ib) { - amdgpu_sync_free(adev, &ib->sync, ib->fence); - amdgpu_sa_bo_free(adev, &ib->sa_bo, ib->fence); + amdgpu_sync_free(adev, &ib->sync, &ib->fence->base); + amdgpu_sa_bo_free(adev, &ib->sa_bo, &ib->fence->base); amdgpu_fence_unref(&ib->fence); } @@ -143,7 +126,6 @@ int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs, struct amdgpu_ring *ring; struct amdgpu_ctx *ctx, *old_ctx; struct amdgpu_vm *vm; - uint64_t sequence; unsigned i; int r = 0; @@ -158,7 +140,11 @@ int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs, dev_err(adev->dev, "couldn't schedule ib\n"); return -EINVAL; } - + r = amdgpu_sync_wait(&ibs->sync); + if (r) { + dev_err(adev->dev, "IB sync failed (%d).\n", r); + return r; + } r = amdgpu_ring_lock(ring, (256 + AMDGPU_NUM_SYNCS * 8) * num_ibs); if (r) { dev_err(adev->dev, "scheduling IB failed (%d).\n", r); @@ -216,12 +202,9 @@ int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs, return r; } - sequence = amdgpu_enable_scheduler ? ib->sequence : 0; - if (!amdgpu_enable_scheduler && ib->ctx) ib->sequence = amdgpu_ctx_add_fence(ib->ctx, ring, - &ib->fence->base, - sequence); + &ib->fence->base); /* wrap the last IB with fence */ if (ib->user) { diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ih.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ih.c index 90044b254404..534fc04e80fd 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ih.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ih.c @@ -43,7 +43,7 @@ static int amdgpu_ih_ring_alloc(struct amdgpu_device *adev) r = amdgpu_bo_create(adev, adev->irq.ih.ring_size, PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_GTT, 0, - NULL, &adev->irq.ih.ring_obj); + NULL, NULL, &adev->irq.ih.ring_obj); if (r) { DRM_ERROR("amdgpu: failed to create ih ring buffer (%d).\n", r); return r; @@ -98,18 +98,12 @@ int amdgpu_ih_ring_init(struct amdgpu_device *adev, unsigned ring_size, /* add 8 bytes for the rptr/wptr shadows and * add them to the end of the ring allocation. */ - adev->irq.ih.ring = kzalloc(adev->irq.ih.ring_size + 8, GFP_KERNEL); + adev->irq.ih.ring = pci_alloc_consistent(adev->pdev, + adev->irq.ih.ring_size + 8, + &adev->irq.ih.rb_dma_addr); if (adev->irq.ih.ring == NULL) return -ENOMEM; - adev->irq.ih.rb_dma_addr = pci_map_single(adev->pdev, - (void *)adev->irq.ih.ring, - adev->irq.ih.ring_size, - PCI_DMA_BIDIRECTIONAL); - if (pci_dma_mapping_error(adev->pdev, adev->irq.ih.rb_dma_addr)) { - dev_err(&adev->pdev->dev, "Failed to DMA MAP the IH RB page\n"); - kfree((void *)adev->irq.ih.ring); - return -ENOMEM; - } + memset((void *)adev->irq.ih.ring, 0, adev->irq.ih.ring_size + 8); adev->irq.ih.wptr_offs = (adev->irq.ih.ring_size / 4) + 0; adev->irq.ih.rptr_offs = (adev->irq.ih.ring_size / 4) + 1; } @@ -149,9 +143,9 @@ void amdgpu_ih_ring_fini(struct amdgpu_device *adev) /* add 8 bytes for the rptr/wptr shadows and * add them to the end of the ring allocation. */ - pci_unmap_single(adev->pdev, adev->irq.ih.rb_dma_addr, - adev->irq.ih.ring_size + 8, PCI_DMA_BIDIRECTIONAL); - kfree((void *)adev->irq.ih.ring); + pci_free_consistent(adev->pdev, adev->irq.ih.ring_size + 8, + (void *)adev->irq.ih.ring, + adev->irq.ih.rb_dma_addr); adev->irq.ih.ring = NULL; } } else { diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_irq.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_irq.c index 0aba8e9bc8a0..7c42ff670080 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_irq.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_irq.c @@ -140,7 +140,7 @@ void amdgpu_irq_preinstall(struct drm_device *dev) */ int amdgpu_irq_postinstall(struct drm_device *dev) { - dev->max_vblank_count = 0x001fffff; + dev->max_vblank_count = 0x00ffffff; return 0; } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c index 87da6b1848fd..275f1c3dbba0 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c @@ -390,7 +390,7 @@ static int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file min((size_t)size, sizeof(vram_gtt))) ? -EFAULT : 0; } case AMDGPU_INFO_READ_MMR_REG: { - unsigned n, alloc_size = info->read_mmr_reg.count * 4; + unsigned n, alloc_size; uint32_t *regs; unsigned se_num = (info->read_mmr_reg.instance >> AMDGPU_INFO_MMR_SE_INDEX_SHIFT) & @@ -406,9 +406,10 @@ static int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file if (sh_num == AMDGPU_INFO_MMR_SH_INDEX_MASK) sh_num = 0xffffffff; - regs = kmalloc(alloc_size, GFP_KERNEL); + regs = kmalloc_array(info->read_mmr_reg.count, sizeof(*regs), GFP_KERNEL); if (!regs) return -ENOMEM; + alloc_size = info->read_mmr_reg.count * sizeof(*regs); for (i = 0; i < info->read_mmr_reg.count; i++) if (amdgpu_asic_read_register(adev, se_num, sh_num, @@ -560,6 +561,8 @@ void amdgpu_driver_postclose_kms(struct drm_device *dev, if (!fpriv) return; + amdgpu_ctx_mgr_fini(&fpriv->ctx_mgr); + amdgpu_vm_fini(adev, &fpriv->vm); idr_for_each_entry(&fpriv->bo_list_handles, list, handle) @@ -568,8 +571,6 @@ void amdgpu_driver_postclose_kms(struct drm_device *dev, idr_destroy(&fpriv->bo_list_handles); mutex_destroy(&fpriv->bo_list_lock); - amdgpu_ctx_mgr_fini(&fpriv->ctx_mgr); - kfree(fpriv); file_priv->driver_priv = NULL; } @@ -681,7 +682,7 @@ int amdgpu_get_vblank_timestamp_kms(struct drm_device *dev, int crtc, /* Helper routine in DRM core does all the work: */ return drm_calc_vbltimestamp_from_scanoutpos(dev, crtc, max_error, vblank_time, flags, - drmcrtc, &drmcrtc->hwmode); + &drmcrtc->hwmode); } const struct drm_ioctl_desc amdgpu_ioctls_kms[] = { diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mode.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_mode.h index 64efe5b52e65..2b03425f9740 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_mode.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mode.h @@ -543,7 +543,8 @@ void amdgpu_encoder_set_active_device(struct drm_encoder *encoder); int amdgpu_get_crtc_scanoutpos(struct drm_device *dev, int crtc, unsigned int flags, int *vpos, int *hpos, ktime_t *stime, - ktime_t *etime); + ktime_t *etime, + const struct drm_display_mode *mode); int amdgpu_framebuffer_init(struct drm_device *dev, struct amdgpu_framebuffer *rfb, diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c index 57adcad2f7ba..1a7708f365f3 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c @@ -127,7 +127,7 @@ static void amdgpu_ttm_placement_init(struct amdgpu_device *adev, placements[c].fpfn = adev->mc.visible_vram_size >> PAGE_SHIFT; placements[c++].flags = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED | - TTM_PL_FLAG_VRAM; + TTM_PL_FLAG_VRAM | TTM_PL_FLAG_TOPDOWN; } placements[c].fpfn = 0; placements[c++].flags = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED | @@ -215,6 +215,7 @@ int amdgpu_bo_create_restricted(struct amdgpu_device *adev, bool kernel, u32 domain, u64 flags, struct sg_table *sg, struct ttm_placement *placement, + struct reservation_object *resv, struct amdgpu_bo **bo_ptr) { struct amdgpu_bo *bo; @@ -261,7 +262,7 @@ int amdgpu_bo_create_restricted(struct amdgpu_device *adev, /* Kernel allocation are uninterruptible */ r = ttm_bo_init(&adev->mman.bdev, &bo->tbo, size, type, &bo->placement, page_align, !kernel, NULL, - acc_size, sg, NULL, &amdgpu_ttm_bo_destroy); + acc_size, sg, resv, &amdgpu_ttm_bo_destroy); if (unlikely(r != 0)) { return r; } @@ -275,7 +276,9 @@ int amdgpu_bo_create_restricted(struct amdgpu_device *adev, int amdgpu_bo_create(struct amdgpu_device *adev, unsigned long size, int byte_align, bool kernel, u32 domain, u64 flags, - struct sg_table *sg, struct amdgpu_bo **bo_ptr) + struct sg_table *sg, + struct reservation_object *resv, + struct amdgpu_bo **bo_ptr) { struct ttm_placement placement = {0}; struct ttm_place placements[AMDGPU_GEM_DOMAIN_MAX + 1]; @@ -286,11 +289,9 @@ int amdgpu_bo_create(struct amdgpu_device *adev, amdgpu_ttm_placement_init(adev, &placement, placements, domain, flags); - return amdgpu_bo_create_restricted(adev, size, byte_align, - kernel, domain, flags, - sg, - &placement, - bo_ptr); + return amdgpu_bo_create_restricted(adev, size, byte_align, kernel, + domain, flags, sg, &placement, + resv, bo_ptr); } int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr) @@ -535,12 +536,10 @@ int amdgpu_bo_set_metadata (struct amdgpu_bo *bo, void *metadata, if (metadata == NULL) return -EINVAL; - buffer = kzalloc(metadata_size, GFP_KERNEL); + buffer = kmemdup(metadata, metadata_size, GFP_KERNEL); if (buffer == NULL) return -ENOMEM; - memcpy(buffer, metadata, metadata_size); - kfree(bo->metadata); bo->metadata_flags = flags; bo->metadata = buffer; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h index 238465a9ac55..3c2ff4567798 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h @@ -129,12 +129,14 @@ int amdgpu_bo_create(struct amdgpu_device *adev, unsigned long size, int byte_align, bool kernel, u32 domain, u64 flags, struct sg_table *sg, + struct reservation_object *resv, struct amdgpu_bo **bo_ptr); int amdgpu_bo_create_restricted(struct amdgpu_device *adev, unsigned long size, int byte_align, bool kernel, u32 domain, u64 flags, struct sg_table *sg, struct ttm_placement *placement, + struct reservation_object *resv, struct amdgpu_bo **bo_ptr); int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr); void amdgpu_bo_kunmap(struct amdgpu_bo *bo); @@ -193,7 +195,7 @@ int amdgpu_sa_bo_new(struct amdgpu_device *adev, unsigned size, unsigned align); void amdgpu_sa_bo_free(struct amdgpu_device *adev, struct amdgpu_sa_bo **sa_bo, - struct amdgpu_fence *fence); + struct fence *fence); #if defined(CONFIG_DEBUG_FS) void amdgpu_sa_bo_dump_debug_info(struct amdgpu_sa_manager *sa_manager, struct seq_file *m); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_prime.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_prime.c index d9652fe32d6a..59f735a933a9 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_prime.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_prime.c @@ -61,12 +61,15 @@ struct drm_gem_object *amdgpu_gem_prime_import_sg_table(struct drm_device *dev, struct dma_buf_attachment *attach, struct sg_table *sg) { + struct reservation_object *resv = attach->dmabuf->resv; struct amdgpu_device *adev = dev->dev_private; struct amdgpu_bo *bo; int ret; + ww_mutex_lock(&resv->lock, NULL); ret = amdgpu_bo_create(adev, attach->dmabuf->size, PAGE_SIZE, false, - AMDGPU_GEM_DOMAIN_GTT, 0, sg, &bo); + AMDGPU_GEM_DOMAIN_GTT, 0, sg, resv, &bo); + ww_mutex_unlock(&resv->lock); if (ret) return ERR_PTR(ret); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c index 7d442c51063e..30dce235ddeb 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c @@ -131,6 +131,21 @@ int amdgpu_ring_lock(struct amdgpu_ring *ring, unsigned ndw) return 0; } +/** amdgpu_ring_insert_nop - insert NOP packets + * + * @ring: amdgpu_ring structure holding ring information + * @count: the number of NOP packets to insert + * + * This is the generic insert_nop function for rings except SDMA + */ +void amdgpu_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count) +{ + int i; + + for (i = 0; i < count; i++) + amdgpu_ring_write(ring, ring->nop); +} + /** * amdgpu_ring_commit - tell the GPU to execute the new * commands on the ring buffer @@ -143,10 +158,13 @@ int amdgpu_ring_lock(struct amdgpu_ring *ring, unsigned ndw) */ void amdgpu_ring_commit(struct amdgpu_ring *ring) { + uint32_t count; + /* We pad to match fetch size */ - while (ring->wptr & ring->align_mask) { - amdgpu_ring_write(ring, ring->nop); - } + count = ring->align_mask + 1 - (ring->wptr & ring->align_mask); + count %= ring->align_mask + 1; + ring->funcs->insert_nop(ring, count); + mb(); amdgpu_ring_set_wptr(ring); } @@ -339,11 +357,11 @@ int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring, ring->adev = adev; ring->idx = adev->num_rings++; adev->rings[ring->idx] = ring; - amdgpu_fence_driver_init_ring(ring); + r = amdgpu_fence_driver_init_ring(ring); + if (r) + return r; } - init_waitqueue_head(&ring->fence_drv.fence_queue); - r = amdgpu_wb_get(adev, &ring->rptr_offs); if (r) { dev_err(adev->dev, "(%d) ring rptr_offs wb alloc failed\n", r); @@ -389,7 +407,7 @@ int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring, if (ring->ring_obj == NULL) { r = amdgpu_bo_create(adev, ring->ring_size, PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_GTT, 0, - NULL, &ring->ring_obj); + NULL, NULL, &ring->ring_obj); if (r) { dev_err(adev->dev, "(%d) ring create failed\n", r); return r; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_sa.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_sa.c index d6398cf45f24..e90712443fe9 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_sa.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_sa.c @@ -64,8 +64,8 @@ int amdgpu_sa_bo_manager_init(struct amdgpu_device *adev, INIT_LIST_HEAD(&sa_manager->flist[i]); } - r = amdgpu_bo_create(adev, size, align, true, - domain, 0, NULL, &sa_manager->bo); + r = amdgpu_bo_create(adev, size, align, true, domain, + 0, NULL, NULL, &sa_manager->bo); if (r) { dev_err(adev->dev, "(%d) failed to allocate bo for manager\n", r); return r; @@ -139,6 +139,25 @@ int amdgpu_sa_bo_manager_suspend(struct amdgpu_device *adev, return r; } +static uint32_t amdgpu_sa_get_ring_from_fence(struct fence *f) +{ + struct amdgpu_fence *a_fence; + struct amd_sched_fence *s_fence; + + s_fence = to_amd_sched_fence(f); + if (s_fence) { + struct amdgpu_ring *ring; + + ring = container_of(s_fence->sched, struct amdgpu_ring, sched); + return ring->idx; + } + + a_fence = to_amdgpu_fence(f); + if (a_fence) + return a_fence->ring->idx; + return 0; +} + static void amdgpu_sa_bo_remove_locked(struct amdgpu_sa_bo *sa_bo) { struct amdgpu_sa_manager *sa_manager = sa_bo->manager; @@ -147,7 +166,7 @@ static void amdgpu_sa_bo_remove_locked(struct amdgpu_sa_bo *sa_bo) } list_del_init(&sa_bo->olist); list_del_init(&sa_bo->flist); - amdgpu_fence_unref(&sa_bo->fence); + fence_put(sa_bo->fence); kfree(sa_bo); } @@ -161,7 +180,7 @@ static void amdgpu_sa_bo_try_free(struct amdgpu_sa_manager *sa_manager) sa_bo = list_entry(sa_manager->hole->next, struct amdgpu_sa_bo, olist); list_for_each_entry_safe_from(sa_bo, tmp, &sa_manager->olist, olist) { if (sa_bo->fence == NULL || - !fence_is_signaled(&sa_bo->fence->base)) { + !fence_is_signaled(sa_bo->fence)) { return; } amdgpu_sa_bo_remove_locked(sa_bo); @@ -246,7 +265,7 @@ static bool amdgpu_sa_event(struct amdgpu_sa_manager *sa_manager, } static bool amdgpu_sa_bo_next_hole(struct amdgpu_sa_manager *sa_manager, - struct amdgpu_fence **fences, + struct fence **fences, unsigned *tries) { struct amdgpu_sa_bo *best_bo = NULL; @@ -275,7 +294,7 @@ static bool amdgpu_sa_bo_next_hole(struct amdgpu_sa_manager *sa_manager, sa_bo = list_first_entry(&sa_manager->flist[i], struct amdgpu_sa_bo, flist); - if (!fence_is_signaled(&sa_bo->fence->base)) { + if (!fence_is_signaled(sa_bo->fence)) { fences[i] = sa_bo->fence; continue; } @@ -299,7 +318,8 @@ static bool amdgpu_sa_bo_next_hole(struct amdgpu_sa_manager *sa_manager, } if (best_bo) { - ++tries[best_bo->fence->ring->idx]; + uint32_t idx = amdgpu_sa_get_ring_from_fence(best_bo->fence); + ++tries[idx]; sa_manager->hole = best_bo->olist.prev; /* we knew that this one is signaled, @@ -315,7 +335,7 @@ int amdgpu_sa_bo_new(struct amdgpu_device *adev, struct amdgpu_sa_bo **sa_bo, unsigned size, unsigned align) { - struct amdgpu_fence *fences[AMDGPU_MAX_RINGS]; + struct fence *fences[AMDGPU_MAX_RINGS]; unsigned tries[AMDGPU_MAX_RINGS]; int i, r; signed long t; @@ -352,7 +372,8 @@ int amdgpu_sa_bo_new(struct amdgpu_device *adev, } while (amdgpu_sa_bo_next_hole(sa_manager, fences, tries)); spin_unlock(&sa_manager->wq.lock); - t = amdgpu_fence_wait_any(adev, fences, false, MAX_SCHEDULE_TIMEOUT); + t = amdgpu_fence_wait_any(adev, fences, AMDGPU_MAX_RINGS, + false, MAX_SCHEDULE_TIMEOUT); r = (t > 0) ? 0 : t; spin_lock(&sa_manager->wq.lock); /* if we have nothing to wait for block */ @@ -372,7 +393,7 @@ int amdgpu_sa_bo_new(struct amdgpu_device *adev, } void amdgpu_sa_bo_free(struct amdgpu_device *adev, struct amdgpu_sa_bo **sa_bo, - struct amdgpu_fence *fence) + struct fence *fence) { struct amdgpu_sa_manager *sa_manager; @@ -382,10 +403,11 @@ void amdgpu_sa_bo_free(struct amdgpu_device *adev, struct amdgpu_sa_bo **sa_bo, sa_manager = (*sa_bo)->manager; spin_lock(&sa_manager->wq.lock); - if (fence && !fence_is_signaled(&fence->base)) { - (*sa_bo)->fence = amdgpu_fence_ref(fence); - list_add_tail(&(*sa_bo)->flist, - &sa_manager->flist[fence->ring->idx]); + if (fence && !fence_is_signaled(fence)) { + uint32_t idx; + (*sa_bo)->fence = fence_get(fence); + idx = amdgpu_sa_get_ring_from_fence(fence); + list_add_tail(&(*sa_bo)->flist, &sa_manager->flist[idx]); } else { amdgpu_sa_bo_remove_locked(*sa_bo); } @@ -395,6 +417,26 @@ void amdgpu_sa_bo_free(struct amdgpu_device *adev, struct amdgpu_sa_bo **sa_bo, } #if defined(CONFIG_DEBUG_FS) + +static void amdgpu_sa_bo_dump_fence(struct fence *fence, struct seq_file *m) +{ + struct amdgpu_fence *a_fence = to_amdgpu_fence(fence); + struct amd_sched_fence *s_fence = to_amd_sched_fence(fence); + + if (a_fence) + seq_printf(m, " protected by 0x%016llx on ring %d", + a_fence->seq, a_fence->ring->idx); + + if (s_fence) { + struct amdgpu_ring *ring; + + + ring = container_of(s_fence->sched, struct amdgpu_ring, sched); + seq_printf(m, " protected by 0x%016x on ring %d", + s_fence->base.seqno, ring->idx); + } +} + void amdgpu_sa_bo_dump_debug_info(struct amdgpu_sa_manager *sa_manager, struct seq_file *m) { @@ -411,10 +453,8 @@ void amdgpu_sa_bo_dump_debug_info(struct amdgpu_sa_manager *sa_manager, } seq_printf(m, "[0x%010llx 0x%010llx] size %8lld", soffset, eoffset, eoffset - soffset); - if (i->fence) { - seq_printf(m, " protected by 0x%016llx on ring %d", - i->fence->seq, i->fence->ring->idx); - } + if (i->fence) + amdgpu_sa_bo_dump_fence(i->fence, m); seq_printf(m, "\n"); } spin_unlock(&sa_manager->wq.lock); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_sched.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_sched.c index a86e38158afa..2e946b2cad88 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_sched.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_sched.c @@ -27,119 +27,87 @@ #include <drm/drmP.h> #include "amdgpu.h" -static int amdgpu_sched_prepare_job(struct amd_gpu_scheduler *sched, - struct amd_sched_entity *entity, - struct amd_sched_job *job) +static struct fence *amdgpu_sched_dependency(struct amd_sched_job *sched_job) { - int r = 0; - struct amdgpu_cs_parser *sched_job; - if (!job || !job->data) { - DRM_ERROR("job is null\n"); - return -EINVAL; - } - - sched_job = (struct amdgpu_cs_parser *)job->data; - if (sched_job->prepare_job) { - r = sched_job->prepare_job(sched_job); - if (r) { - DRM_ERROR("Prepare job error\n"); - schedule_work(&sched_job->job_work); - } - } - return r; + struct amdgpu_job *job = to_amdgpu_job(sched_job); + return amdgpu_sync_get_fence(&job->ibs->sync); } -static struct fence *amdgpu_sched_run_job(struct amd_gpu_scheduler *sched, - struct amd_sched_entity *entity, - struct amd_sched_job *job) +static struct fence *amdgpu_sched_run_job(struct amd_sched_job *sched_job) { - int r = 0; - struct amdgpu_cs_parser *sched_job; - struct amdgpu_fence *fence; + struct amdgpu_fence *fence = NULL; + struct amdgpu_job *job; + int r; - if (!job || !job->data) { + if (!sched_job) { DRM_ERROR("job is null\n"); return NULL; } - sched_job = (struct amdgpu_cs_parser *)job->data; - mutex_lock(&sched_job->job_lock); - r = amdgpu_ib_schedule(sched_job->adev, - sched_job->num_ibs, - sched_job->ibs, - sched_job->filp); - if (r) + job = to_amdgpu_job(sched_job); + mutex_lock(&job->job_lock); + r = amdgpu_ib_schedule(job->adev, + job->num_ibs, + job->ibs, + job->base.owner); + if (r) { + DRM_ERROR("Error scheduling IBs (%d)\n", r); goto err; - fence = amdgpu_fence_ref(sched_job->ibs[sched_job->num_ibs - 1].fence); - - if (sched_job->run_job) { - r = sched_job->run_job(sched_job); - if (r) - goto err; } - mutex_unlock(&sched_job->job_lock); - return &fence->base; + fence = amdgpu_fence_ref(job->ibs[job->num_ibs - 1].fence); err: - DRM_ERROR("Run job error\n"); - mutex_unlock(&sched_job->job_lock); - schedule_work(&sched_job->job_work); - return NULL; -} - -static void amdgpu_sched_process_job(struct amd_gpu_scheduler *sched, - struct amd_sched_job *job) -{ - struct amdgpu_cs_parser *sched_job; + if (job->free_job) + job->free_job(job); - if (!job || !job->data) { - DRM_ERROR("job is null\n"); - return; - } - sched_job = (struct amdgpu_cs_parser *)job->data; - schedule_work(&sched_job->job_work); + mutex_unlock(&job->job_lock); + fence_put(&job->base.s_fence->base); + kfree(job); + return fence ? &fence->base : NULL; } struct amd_sched_backend_ops amdgpu_sched_ops = { - .prepare_job = amdgpu_sched_prepare_job, + .dependency = amdgpu_sched_dependency, .run_job = amdgpu_sched_run_job, - .process_job = amdgpu_sched_process_job }; int amdgpu_sched_ib_submit_kernel_helper(struct amdgpu_device *adev, struct amdgpu_ring *ring, struct amdgpu_ib *ibs, unsigned num_ibs, - int (*free_job)(struct amdgpu_cs_parser *), + int (*free_job)(struct amdgpu_job *), void *owner, struct fence **f) { int r = 0; if (amdgpu_enable_scheduler) { - struct amdgpu_cs_parser *sched_job = - amdgpu_cs_parser_create(adev, owner, &adev->kernel_ctx, - ibs, num_ibs); - if(!sched_job) { + struct amdgpu_job *job = + kzalloc(sizeof(struct amdgpu_job), GFP_KERNEL); + if (!job) return -ENOMEM; - } - sched_job->free_job = free_job; - mutex_lock(&sched_job->job_lock); - r = amd_sched_push_job(ring->scheduler, - &adev->kernel_ctx.rings[ring->idx].entity, - sched_job, &sched_job->s_fence); + job->base.sched = &ring->sched; + job->base.s_entity = &adev->kernel_ctx.rings[ring->idx].entity; + job->adev = adev; + job->ibs = ibs; + job->num_ibs = num_ibs; + job->base.owner = owner; + mutex_init(&job->job_lock); + job->free_job = free_job; + mutex_lock(&job->job_lock); + r = amd_sched_entity_push_job(&job->base); if (r) { - mutex_unlock(&sched_job->job_lock); - kfree(sched_job); + mutex_unlock(&job->job_lock); + kfree(job); return r; } - ibs[num_ibs - 1].sequence = sched_job->s_fence->v_seq; - *f = fence_get(&sched_job->s_fence->base); - mutex_unlock(&sched_job->job_lock); + *f = fence_get(&job->base.s_fence->base); + mutex_unlock(&job->job_lock); } else { r = amdgpu_ib_schedule(adev, num_ibs, ibs, owner); if (r) return r; *f = fence_get(&ibs[num_ibs - 1].fence->base); } + return 0; } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_semaphore.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_semaphore.c index d6d41a42ab65..ff3ca52ec6fe 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_semaphore.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_semaphore.c @@ -87,7 +87,7 @@ bool amdgpu_semaphore_emit_wait(struct amdgpu_ring *ring, void amdgpu_semaphore_free(struct amdgpu_device *adev, struct amdgpu_semaphore **semaphore, - struct amdgpu_fence *fence) + struct fence *fence) { if (semaphore == NULL || *semaphore == NULL) { return; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c index 7cb711fc1ee2..4921de15b451 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c @@ -32,6 +32,11 @@ #include "amdgpu.h" #include "amdgpu_trace.h" +struct amdgpu_sync_entry { + struct hlist_node node; + struct fence *fence; +}; + /** * amdgpu_sync_create - zero init sync object * @@ -49,9 +54,39 @@ void amdgpu_sync_create(struct amdgpu_sync *sync) for (i = 0; i < AMDGPU_MAX_RINGS; ++i) sync->sync_to[i] = NULL; + hash_init(sync->fences); sync->last_vm_update = NULL; } +static bool amdgpu_sync_same_dev(struct amdgpu_device *adev, struct fence *f) +{ + struct amdgpu_fence *a_fence = to_amdgpu_fence(f); + struct amd_sched_fence *s_fence = to_amd_sched_fence(f); + + if (a_fence) + return a_fence->ring->adev == adev; + + if (s_fence) { + struct amdgpu_ring *ring; + + ring = container_of(s_fence->sched, struct amdgpu_ring, sched); + return ring->adev == adev; + } + + return false; +} + +static bool amdgpu_sync_test_owner(struct fence *f, void *owner) +{ + struct amdgpu_fence *a_fence = to_amdgpu_fence(f); + struct amd_sched_fence *s_fence = to_amd_sched_fence(f); + if (s_fence) + return s_fence->owner == owner; + if (a_fence) + return a_fence->owner == owner; + return false; +} + /** * amdgpu_sync_fence - remember to sync to this fence * @@ -62,31 +97,69 @@ void amdgpu_sync_create(struct amdgpu_sync *sync) int amdgpu_sync_fence(struct amdgpu_device *adev, struct amdgpu_sync *sync, struct fence *f) { + struct amdgpu_sync_entry *e; struct amdgpu_fence *fence; struct amdgpu_fence *other; + struct fence *tmp, *later; if (!f) return 0; + if (amdgpu_sync_same_dev(adev, f) && + amdgpu_sync_test_owner(f, AMDGPU_FENCE_OWNER_VM)) { + if (sync->last_vm_update) { + tmp = sync->last_vm_update; + BUG_ON(f->context != tmp->context); + later = (f->seqno - tmp->seqno <= INT_MAX) ? f : tmp; + sync->last_vm_update = fence_get(later); + fence_put(tmp); + } else + sync->last_vm_update = fence_get(f); + } + fence = to_amdgpu_fence(f); - if (!fence || fence->ring->adev != adev) - return fence_wait(f, true); + if (!fence || fence->ring->adev != adev) { + hash_for_each_possible(sync->fences, e, node, f->context) { + struct fence *new; + if (unlikely(e->fence->context != f->context)) + continue; + new = fence_get(fence_later(e->fence, f)); + if (new) { + fence_put(e->fence); + e->fence = new; + } + return 0; + } + + e = kmalloc(sizeof(struct amdgpu_sync_entry), GFP_KERNEL); + if (!e) + return -ENOMEM; + + hash_add(sync->fences, &e->node, f->context); + e->fence = fence_get(f); + return 0; + } other = sync->sync_to[fence->ring->idx]; sync->sync_to[fence->ring->idx] = amdgpu_fence_ref( amdgpu_fence_later(fence, other)); amdgpu_fence_unref(&other); - if (fence->owner == AMDGPU_FENCE_OWNER_VM) { - other = sync->last_vm_update; - sync->last_vm_update = amdgpu_fence_ref( - amdgpu_fence_later(fence, other)); - amdgpu_fence_unref(&other); - } - return 0; } +static void *amdgpu_sync_get_owner(struct fence *f) +{ + struct amdgpu_fence *a_fence = to_amdgpu_fence(f); + struct amd_sched_fence *s_fence = to_amd_sched_fence(f); + + if (s_fence) + return s_fence->owner; + else if (a_fence) + return a_fence->owner; + return AMDGPU_FENCE_OWNER_UNDEFINED; +} + /** * amdgpu_sync_resv - use the semaphores to sync to a reservation object * @@ -103,7 +176,7 @@ int amdgpu_sync_resv(struct amdgpu_device *adev, { struct reservation_object_list *flist; struct fence *f; - struct amdgpu_fence *fence; + void *fence_owner; unsigned i; int r = 0; @@ -121,22 +194,22 @@ int amdgpu_sync_resv(struct amdgpu_device *adev, for (i = 0; i < flist->shared_count; ++i) { f = rcu_dereference_protected(flist->shared[i], reservation_object_held(resv)); - fence = f ? to_amdgpu_fence(f) : NULL; - if (fence && fence->ring->adev == adev) { + if (amdgpu_sync_same_dev(adev, f)) { /* VM updates are only interesting * for other VM updates and moves. */ + fence_owner = amdgpu_sync_get_owner(f); if ((owner != AMDGPU_FENCE_OWNER_MOVE) && - (fence->owner != AMDGPU_FENCE_OWNER_MOVE) && + (fence_owner != AMDGPU_FENCE_OWNER_MOVE) && ((owner == AMDGPU_FENCE_OWNER_VM) != - (fence->owner == AMDGPU_FENCE_OWNER_VM))) + (fence_owner == AMDGPU_FENCE_OWNER_VM))) continue; /* Ignore fence from the same owner as * long as it isn't undefined. */ if (owner != AMDGPU_FENCE_OWNER_UNDEFINED && - fence->owner == owner) + fence_owner == owner) continue; } @@ -147,6 +220,60 @@ int amdgpu_sync_resv(struct amdgpu_device *adev, return r; } +struct fence *amdgpu_sync_get_fence(struct amdgpu_sync *sync) +{ + struct amdgpu_sync_entry *e; + struct hlist_node *tmp; + struct fence *f; + int i; + + hash_for_each_safe(sync->fences, i, tmp, e, node) { + + f = e->fence; + + hash_del(&e->node); + kfree(e); + + if (!fence_is_signaled(f)) + return f; + + fence_put(f); + } + return NULL; +} + +int amdgpu_sync_wait(struct amdgpu_sync *sync) +{ + struct amdgpu_sync_entry *e; + struct hlist_node *tmp; + int i, r; + + hash_for_each_safe(sync->fences, i, tmp, e, node) { + r = fence_wait(e->fence, false); + if (r) + return r; + + hash_del(&e->node); + fence_put(e->fence); + kfree(e); + } + + if (amdgpu_enable_semaphores) + return 0; + + for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { + struct amdgpu_fence *fence = sync->sync_to[i]; + if (!fence) + continue; + + r = fence_wait(&fence->base, false); + if (r) + return r; + } + + return 0; +} + /** * amdgpu_sync_rings - sync ring to all registered fences * @@ -178,7 +305,8 @@ int amdgpu_sync_rings(struct amdgpu_sync *sync, return -EINVAL; } - if (amdgpu_enable_scheduler || (count >= AMDGPU_NUM_SYNCS)) { + if (amdgpu_enable_scheduler || !amdgpu_enable_semaphores || + (count >= AMDGPU_NUM_SYNCS)) { /* not enough room, wait manually */ r = fence_wait(&fence->base, false); if (r) @@ -234,15 +362,23 @@ int amdgpu_sync_rings(struct amdgpu_sync *sync, */ void amdgpu_sync_free(struct amdgpu_device *adev, struct amdgpu_sync *sync, - struct amdgpu_fence *fence) + struct fence *fence) { + struct amdgpu_sync_entry *e; + struct hlist_node *tmp; unsigned i; + hash_for_each_safe(sync->fences, i, tmp, e, node) { + hash_del(&e->node); + fence_put(e->fence); + kfree(e); + } + for (i = 0; i < AMDGPU_NUM_SYNCS; ++i) amdgpu_semaphore_free(adev, &sync->semaphores[i], fence); for (i = 0; i < AMDGPU_MAX_RINGS; ++i) amdgpu_fence_unref(&sync->sync_to[i]); - amdgpu_fence_unref(&sync->last_vm_update); + fence_put(sync->last_vm_update); } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_test.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_test.c index 962dd5552137..4865615e9c06 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_test.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_test.c @@ -59,8 +59,9 @@ static void amdgpu_do_test_moves(struct amdgpu_device *adev) goto out_cleanup; } - r = amdgpu_bo_create(adev, size, PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_VRAM, 0, - NULL, &vram_obj); + r = amdgpu_bo_create(adev, size, PAGE_SIZE, true, + AMDGPU_GEM_DOMAIN_VRAM, 0, + NULL, NULL, &vram_obj); if (r) { DRM_ERROR("Failed to create VRAM object\n"); goto out_cleanup; @@ -77,10 +78,11 @@ static void amdgpu_do_test_moves(struct amdgpu_device *adev) void *gtt_map, *vram_map; void **gtt_start, **gtt_end; void **vram_start, **vram_end; - struct amdgpu_fence *fence = NULL; + struct fence *fence = NULL; r = amdgpu_bo_create(adev, size, PAGE_SIZE, true, - AMDGPU_GEM_DOMAIN_GTT, 0, NULL, gtt_obj + i); + AMDGPU_GEM_DOMAIN_GTT, 0, NULL, + NULL, gtt_obj + i); if (r) { DRM_ERROR("Failed to create GTT object %d\n", i); goto out_lclean; @@ -116,13 +118,13 @@ static void amdgpu_do_test_moves(struct amdgpu_device *adev) goto out_lclean_unpin; } - r = fence_wait(&fence->base, false); + r = fence_wait(fence, false); if (r) { DRM_ERROR("Failed to wait for GTT->VRAM fence %d\n", i); goto out_lclean_unpin; } - amdgpu_fence_unref(&fence); + fence_put(fence); r = amdgpu_bo_kmap(vram_obj, &vram_map); if (r) { @@ -161,13 +163,13 @@ static void amdgpu_do_test_moves(struct amdgpu_device *adev) goto out_lclean_unpin; } - r = fence_wait(&fence->base, false); + r = fence_wait(fence, false); if (r) { DRM_ERROR("Failed to wait for VRAM->GTT fence %d\n", i); goto out_lclean_unpin; } - amdgpu_fence_unref(&fence); + fence_put(fence); r = amdgpu_bo_kmap(gtt_obj[i], >t_map); if (r) { @@ -214,7 +216,7 @@ out_lclean: amdgpu_bo_unref(>t_obj[i]); } if (fence) - amdgpu_fence_unref(&fence); + fence_put(fence); break; } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c index dd3415d2e45d..364cbe975332 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c @@ -228,7 +228,7 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo, struct amdgpu_device *adev; struct amdgpu_ring *ring; uint64_t old_start, new_start; - struct amdgpu_fence *fence; + struct fence *fence; int r; adev = amdgpu_get_adev(bo->bdev); @@ -269,9 +269,9 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo, new_mem->num_pages * PAGE_SIZE, /* bytes */ bo->resv, &fence); /* FIXME: handle copy error */ - r = ttm_bo_move_accel_cleanup(bo, &fence->base, + r = ttm_bo_move_accel_cleanup(bo, fence, evict, no_wait_gpu, new_mem); - amdgpu_fence_unref(&fence); + fence_put(fence); return r; } @@ -859,8 +859,9 @@ int amdgpu_ttm_init(struct amdgpu_device *adev) amdgpu_ttm_set_active_vram_size(adev, adev->mc.visible_vram_size); r = amdgpu_bo_create(adev, 256 * 1024, PAGE_SIZE, true, - AMDGPU_GEM_DOMAIN_VRAM, 0, - NULL, &adev->stollen_vga_memory); + AMDGPU_GEM_DOMAIN_VRAM, + AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED, + NULL, NULL, &adev->stollen_vga_memory); if (r) { return r; } @@ -987,46 +988,48 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t dst_offset, uint32_t byte_count, struct reservation_object *resv, - struct amdgpu_fence **fence) + struct fence **fence) { struct amdgpu_device *adev = ring->adev; - struct amdgpu_sync sync; uint32_t max_bytes; unsigned num_loops, num_dw; + struct amdgpu_ib *ib; unsigned i; int r; - /* sync other rings */ - amdgpu_sync_create(&sync); - if (resv) { - r = amdgpu_sync_resv(adev, &sync, resv, false); - if (r) { - DRM_ERROR("sync failed (%d).\n", r); - amdgpu_sync_free(adev, &sync, NULL); - return r; - } - } - max_bytes = adev->mman.buffer_funcs->copy_max_bytes; num_loops = DIV_ROUND_UP(byte_count, max_bytes); num_dw = num_loops * adev->mman.buffer_funcs->copy_num_dw; - /* for fence and sync */ - num_dw += 64 + AMDGPU_NUM_SYNCS * 8; + /* for IB padding */ + while (num_dw & 0x7) + num_dw++; + + ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL); + if (!ib) + return -ENOMEM; - r = amdgpu_ring_lock(ring, num_dw); + r = amdgpu_ib_get(ring, NULL, num_dw * 4, ib); if (r) { - DRM_ERROR("ring lock failed (%d).\n", r); - amdgpu_sync_free(adev, &sync, NULL); + kfree(ib); return r; } - amdgpu_sync_rings(&sync, ring); + ib->length_dw = 0; + + if (resv) { + r = amdgpu_sync_resv(adev, &ib->sync, resv, + AMDGPU_FENCE_OWNER_UNDEFINED); + if (r) { + DRM_ERROR("sync failed (%d).\n", r); + goto error_free; + } + } for (i = 0; i < num_loops; i++) { uint32_t cur_size_in_bytes = min(byte_count, max_bytes); - amdgpu_emit_copy_buffer(adev, ring, src_offset, dst_offset, + amdgpu_emit_copy_buffer(adev, ib, src_offset, dst_offset, cur_size_in_bytes); src_offset += cur_size_in_bytes; @@ -1034,17 +1037,24 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring, byte_count -= cur_size_in_bytes; } - r = amdgpu_fence_emit(ring, AMDGPU_FENCE_OWNER_MOVE, fence); - if (r) { - amdgpu_ring_unlock_undo(ring); - amdgpu_sync_free(adev, &sync, NULL); - return r; - } - - amdgpu_ring_unlock_commit(ring); - amdgpu_sync_free(adev, &sync, *fence); + amdgpu_vm_pad_ib(adev, ib); + WARN_ON(ib->length_dw > num_dw); + r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1, + &amdgpu_vm_free_job, + AMDGPU_FENCE_OWNER_MOVE, + fence); + if (r) + goto error_free; + if (!amdgpu_enable_scheduler) { + amdgpu_ib_free(adev, ib); + kfree(ib); + } return 0; +error_free: + amdgpu_ib_free(adev, ib); + kfree(ib); + return r; } #if defined(CONFIG_DEBUG_FS) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ucode.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ucode.c index 482e66797ae6..5cc95f1a7dab 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ucode.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ucode.c @@ -247,7 +247,7 @@ int amdgpu_ucode_init_bo(struct amdgpu_device *adev) const struct common_firmware_header *header = NULL; err = amdgpu_bo_create(adev, adev->firmware.fw_size, PAGE_SIZE, true, - AMDGPU_GEM_DOMAIN_GTT, 0, NULL, bo); + AMDGPU_GEM_DOMAIN_GTT, 0, NULL, NULL, bo); if (err) { dev_err(adev->dev, "(%d) Firmware buffer allocate failed\n", err); err = -ENOMEM; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c index 68369cf1e318..d0312364d950 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c @@ -154,7 +154,9 @@ int amdgpu_uvd_sw_init(struct amdgpu_device *adev) bo_size = AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8) + AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE; r = amdgpu_bo_create(adev, bo_size, PAGE_SIZE, true, - AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, &adev->uvd.vcpu_bo); + AMDGPU_GEM_DOMAIN_VRAM, + AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED, + NULL, NULL, &adev->uvd.vcpu_bo); if (r) { dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r); return r; @@ -221,31 +223,32 @@ int amdgpu_uvd_sw_fini(struct amdgpu_device *adev) int amdgpu_uvd_suspend(struct amdgpu_device *adev) { - unsigned size; - void *ptr; - const struct common_firmware_header *hdr; - int i; + struct amdgpu_ring *ring = &adev->uvd.ring; + int i, r; if (adev->uvd.vcpu_bo == NULL) return 0; - for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) - if (atomic_read(&adev->uvd.handles[i])) - break; - - if (i == AMDGPU_MAX_UVD_HANDLES) - return 0; + for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) { + uint32_t handle = atomic_read(&adev->uvd.handles[i]); + if (handle != 0) { + struct fence *fence; - hdr = (const struct common_firmware_header *)adev->uvd.fw->data; + amdgpu_uvd_note_usage(adev); - size = amdgpu_bo_size(adev->uvd.vcpu_bo); - size -= le32_to_cpu(hdr->ucode_size_bytes); + r = amdgpu_uvd_get_destroy_msg(ring, handle, &fence); + if (r) { + DRM_ERROR("Error destroying UVD (%d)!\n", r); + continue; + } - ptr = adev->uvd.cpu_addr; - ptr += le32_to_cpu(hdr->ucode_size_bytes); + fence_wait(fence, false); + fence_put(fence); - adev->uvd.saved_bo = kmalloc(size, GFP_KERNEL); - memcpy(adev->uvd.saved_bo, ptr, size); + adev->uvd.filp[i] = NULL; + atomic_set(&adev->uvd.handles[i], 0); + } + } return 0; } @@ -270,12 +273,7 @@ int amdgpu_uvd_resume(struct amdgpu_device *adev) ptr = adev->uvd.cpu_addr; ptr += le32_to_cpu(hdr->ucode_size_bytes); - if (adev->uvd.saved_bo != NULL) { - memcpy(ptr, adev->uvd.saved_bo, size); - kfree(adev->uvd.saved_bo); - adev->uvd.saved_bo = NULL; - } else - memset(ptr, 0, size); + memset(ptr, 0, size); return 0; } @@ -545,46 +543,60 @@ static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx, return -EINVAL; } - if (msg_type == 1) { + switch (msg_type) { + case 0: + /* it's a create msg, calc image size (width * height) */ + amdgpu_bo_kunmap(bo); + + /* try to alloc a new handle */ + for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) { + if (atomic_read(&adev->uvd.handles[i]) == handle) { + DRM_ERROR("Handle 0x%x already in use!\n", handle); + return -EINVAL; + } + + if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) { + adev->uvd.filp[i] = ctx->parser->filp; + return 0; + } + } + + DRM_ERROR("No more free UVD handles!\n"); + return -EINVAL; + + case 1: /* it's a decode msg, calc buffer sizes */ r = amdgpu_uvd_cs_msg_decode(msg, ctx->buf_sizes); amdgpu_bo_kunmap(bo); if (r) return r; - } else if (msg_type == 2) { + /* validate the handle */ + for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) { + if (atomic_read(&adev->uvd.handles[i]) == handle) { + if (adev->uvd.filp[i] != ctx->parser->filp) { + DRM_ERROR("UVD handle collision detected!\n"); + return -EINVAL; + } + return 0; + } + } + + DRM_ERROR("Invalid UVD handle 0x%x!\n", handle); + return -ENOENT; + + case 2: /* it's a destroy msg, free the handle */ for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) atomic_cmpxchg(&adev->uvd.handles[i], handle, 0); amdgpu_bo_kunmap(bo); return 0; - } else { - /* it's a create msg */ - amdgpu_bo_kunmap(bo); - - if (msg_type != 0) { - DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type); - return -EINVAL; - } - - /* it's a create msg, no special handling needed */ - } - - /* create or decode, validate the handle */ - for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) { - if (atomic_read(&adev->uvd.handles[i]) == handle) - return 0; - } - /* handle not found try to alloc a new one */ - for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) { - if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) { - adev->uvd.filp[i] = ctx->parser->filp; - return 0; - } + default: + DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type); + return -EINVAL; } - - DRM_ERROR("No more free UVD handles!\n"); + BUG(); return -EINVAL; } @@ -807,10 +819,10 @@ int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx) } static int amdgpu_uvd_free_job( - struct amdgpu_cs_parser *sched_job) + struct amdgpu_job *job) { - amdgpu_ib_free(sched_job->adev, sched_job->ibs); - kfree(sched_job->ibs); + amdgpu_ib_free(job->adev, job->ibs); + kfree(job->ibs); return 0; } @@ -905,7 +917,9 @@ int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle, int r, i; r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true, - AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, &bo); + AMDGPU_GEM_DOMAIN_VRAM, + AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED, + NULL, NULL, &bo); if (r) return r; @@ -952,7 +966,9 @@ int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle, int r, i; r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true, - AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, &bo); + AMDGPU_GEM_DOMAIN_VRAM, + AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED, + NULL, NULL, &bo); if (r) return r; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c index 33ee6ae28f37..74f2038ac747 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c @@ -141,7 +141,9 @@ int amdgpu_vce_sw_init(struct amdgpu_device *adev, unsigned long size) /* allocate firmware, stack and heap BO */ r = amdgpu_bo_create(adev, size, PAGE_SIZE, true, - AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, &adev->vce.vcpu_bo); + AMDGPU_GEM_DOMAIN_VRAM, + AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED, + NULL, NULL, &adev->vce.vcpu_bo); if (r) { dev_err(adev->dev, "(%d) failed to allocate VCE bo\n", r); return r; @@ -340,10 +342,10 @@ void amdgpu_vce_free_handles(struct amdgpu_device *adev, struct drm_file *filp) } static int amdgpu_vce_free_job( - struct amdgpu_cs_parser *sched_job) + struct amdgpu_job *job) { - amdgpu_ib_free(sched_job->adev, sched_job->ibs); - kfree(sched_job->ibs); + amdgpu_ib_free(job->adev, job->ibs); + kfree(job->ibs); return 0; } @@ -836,6 +838,10 @@ int amdgpu_vce_ring_test_ib(struct amdgpu_ring *ring) struct fence *fence = NULL; int r; + /* skip vce ring1 ib test for now, since it's not reliable */ + if (ring == &ring->adev->vce.ring[1]) + return 0; + r = amdgpu_vce_get_create_msg(ring, 1, NULL); if (r) { DRM_ERROR("amdgpu: failed to get create msg (%d).\n", r); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c index a78a206e176e..1e14531353e0 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c @@ -200,19 +200,29 @@ int amdgpu_vm_grab_id(struct amdgpu_vm *vm, struct amdgpu_ring *ring, */ void amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_vm *vm, - struct amdgpu_fence *updates) + struct fence *updates) { uint64_t pd_addr = amdgpu_bo_gpu_offset(vm->page_directory); struct amdgpu_vm_id *vm_id = &vm->ids[ring->idx]; - struct amdgpu_fence *flushed_updates = vm_id->flushed_updates; + struct fence *flushed_updates = vm_id->flushed_updates; + bool is_earlier = false; + + if (flushed_updates && updates) { + BUG_ON(flushed_updates->context != updates->context); + is_earlier = (updates->seqno - flushed_updates->seqno <= + INT_MAX) ? true : false; + } if (pd_addr != vm_id->pd_gpu_addr || !flushed_updates || - (updates && amdgpu_fence_is_earlier(flushed_updates, updates))) { + is_earlier) { trace_amdgpu_vm_flush(pd_addr, ring->idx, vm_id->id); - vm_id->flushed_updates = amdgpu_fence_ref( - amdgpu_fence_later(flushed_updates, updates)); - amdgpu_fence_unref(&flushed_updates); + if (is_earlier) { + vm_id->flushed_updates = fence_get(updates); + fence_put(flushed_updates); + } + if (!flushed_updates) + vm_id->flushed_updates = fence_get(updates); vm_id->pd_gpu_addr = pd_addr; amdgpu_ring_emit_vm_flush(ring, vm_id->id, vm_id->pd_gpu_addr); } @@ -306,13 +316,12 @@ static void amdgpu_vm_update_pages(struct amdgpu_device *adev, } } -static int amdgpu_vm_free_job( - struct amdgpu_cs_parser *sched_job) +int amdgpu_vm_free_job(struct amdgpu_job *job) { int i; - for (i = 0; i < sched_job->num_ibs; i++) - amdgpu_ib_free(sched_job->adev, &sched_job->ibs[i]); - kfree(sched_job->ibs); + for (i = 0; i < job->num_ibs; i++) + amdgpu_ib_free(job->adev, &job->ibs[i]); + kfree(job->ibs); return 0; } @@ -618,9 +627,14 @@ static int amdgpu_vm_update_ptes(struct amdgpu_device *adev, { uint64_t mask = AMDGPU_VM_PTE_COUNT - 1; uint64_t last_pte = ~0, last_dst = ~0; + void *owner = AMDGPU_FENCE_OWNER_VM; unsigned count = 0; uint64_t addr; + /* sync to everything on unmapping */ + if (!(flags & AMDGPU_PTE_VALID)) + owner = AMDGPU_FENCE_OWNER_UNDEFINED; + /* walk over the address space and update the page tables */ for (addr = start; addr < end; ) { uint64_t pt_idx = addr >> amdgpu_vm_block_size; @@ -629,8 +643,7 @@ static int amdgpu_vm_update_ptes(struct amdgpu_device *adev, uint64_t pte; int r; - amdgpu_sync_resv(adev, &ib->sync, pt->tbo.resv, - AMDGPU_FENCE_OWNER_VM); + amdgpu_sync_resv(adev, &ib->sync, pt->tbo.resv, owner); r = reservation_object_reserve_shared(pt->tbo.resv); if (r) return r; @@ -673,31 +686,6 @@ static int amdgpu_vm_update_ptes(struct amdgpu_device *adev, } /** - * amdgpu_vm_fence_pts - fence page tables after an update - * - * @vm: requested vm - * @start: start of GPU address range - * @end: end of GPU address range - * @fence: fence to use - * - * Fence the page tables in the range @start - @end (cayman+). - * - * Global and local mutex must be locked! - */ -static void amdgpu_vm_fence_pts(struct amdgpu_vm *vm, - uint64_t start, uint64_t end, - struct fence *fence) -{ - unsigned i; - - start >>= amdgpu_vm_block_size; - end >>= amdgpu_vm_block_size; - - for (i = start; i <= end; ++i) - amdgpu_bo_fence(vm->page_tables[i].bo, fence, true); -} - -/** * amdgpu_vm_bo_update_mapping - update a mapping in the vm page table * * @adev: amdgpu_device pointer @@ -781,17 +769,6 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev, ib->length_dw = 0; - if (!(flags & AMDGPU_PTE_VALID)) { - unsigned i; - - for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { - struct amdgpu_fence *f = vm->ids[i].last_id_use; - r = amdgpu_sync_fence(adev, &ib->sync, &f->base); - if (r) - return r; - } - } - r = amdgpu_vm_update_ptes(adev, vm, ib, mapping->it.start, mapping->it.last + 1, addr + mapping->offset, flags, gtt_flags); @@ -811,8 +788,7 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev, if (r) goto error_free; - amdgpu_vm_fence_pts(vm, mapping->it.start, - mapping->it.last + 1, f); + amdgpu_bo_fence(vm->page_directory, f, true); if (fence) { fence_put(*fence); *fence = fence_get(f); @@ -853,7 +829,7 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev, int r; if (mem) { - addr = mem->start << PAGE_SHIFT; + addr = (u64)mem->start << PAGE_SHIFT; if (mem->mem_type != TTM_PL_TT) addr += adev->vm_manager.vram_base_offset; } else { @@ -1087,6 +1063,7 @@ int amdgpu_vm_bo_map(struct amdgpu_device *adev, /* walk over the address space and allocate the page tables */ for (pt_idx = saddr; pt_idx <= eaddr; ++pt_idx) { + struct reservation_object *resv = vm->page_directory->tbo.resv; struct amdgpu_bo *pt; if (vm->page_tables[pt_idx].bo) @@ -1095,9 +1072,13 @@ int amdgpu_vm_bo_map(struct amdgpu_device *adev, /* drop mutex to allocate and clear page table */ mutex_unlock(&vm->mutex); + ww_mutex_lock(&resv->lock, NULL); r = amdgpu_bo_create(adev, AMDGPU_VM_PTE_COUNT * 8, AMDGPU_GPU_PAGE_SIZE, true, - AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, &pt); + AMDGPU_GEM_DOMAIN_VRAM, + AMDGPU_GEM_CREATE_NO_CPU_ACCESS, + NULL, resv, &pt); + ww_mutex_unlock(&resv->lock); if (r) goto error_free; @@ -1297,8 +1278,9 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm) vm->page_directory_fence = NULL; r = amdgpu_bo_create(adev, pd_size, align, true, - AMDGPU_GEM_DOMAIN_VRAM, 0, - NULL, &vm->page_directory); + AMDGPU_GEM_DOMAIN_VRAM, + AMDGPU_GEM_CREATE_NO_CPU_ACCESS, + NULL, NULL, &vm->page_directory); if (r) return r; @@ -1347,7 +1329,7 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm) fence_put(vm->page_directory_fence); for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { - amdgpu_fence_unref(&vm->ids[i].flushed_updates); + fence_put(vm->ids[i].flushed_updates); amdgpu_fence_unref(&vm->ids[i].last_id_use); } diff --git a/drivers/gpu/drm/amd/amdgpu/atombios_dp.c b/drivers/gpu/drm/amd/amdgpu/atombios_dp.c index 9ba0a7d5bc8e..92b6acadfc52 100644 --- a/drivers/gpu/drm/amd/amdgpu/atombios_dp.c +++ b/drivers/gpu/drm/amd/amdgpu/atombios_dp.c @@ -139,7 +139,8 @@ amdgpu_atombios_dp_aux_transfer(struct drm_dp_aux *aux, struct drm_dp_aux_msg *m tx_buf[0] = msg->address & 0xff; tx_buf[1] = msg->address >> 8; - tx_buf[2] = msg->request << 4; + tx_buf[2] = (msg->request << 4) | + ((msg->address >> 16) & 0xf); tx_buf[3] = msg->size ? (msg->size - 1) : 0; switch (msg->request & ~DP_AUX_I2C_MOT) { diff --git a/drivers/gpu/drm/amd/amdgpu/cik_sdma.c b/drivers/gpu/drm/amd/amdgpu/cik_sdma.c index 2b4242b39b0a..9ea9de457da3 100644 --- a/drivers/gpu/drm/amd/amdgpu/cik_sdma.c +++ b/drivers/gpu/drm/amd/amdgpu/cik_sdma.c @@ -188,6 +188,19 @@ static void cik_sdma_ring_set_wptr(struct amdgpu_ring *ring) WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me], (ring->wptr << 2) & 0x3fffc); } +static void cik_sdma_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count) +{ + struct amdgpu_sdma *sdma = amdgpu_get_sdma_instance(ring); + int i; + + for (i = 0; i < count; i++) + if (sdma && sdma->burst_nop && (i == 0)) + amdgpu_ring_write(ring, ring->nop | + SDMA_NOP_COUNT(count - 1)); + else + amdgpu_ring_write(ring, ring->nop); +} + /** * cik_sdma_ring_emit_ib - Schedule an IB on the DMA engine * @@ -213,8 +226,8 @@ static void cik_sdma_ring_emit_ib(struct amdgpu_ring *ring, amdgpu_ring_write(ring, next_rptr); /* IB packet must end on a 8 DW boundary */ - while ((ring->wptr & 7) != 4) - amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0)); + cik_sdma_ring_insert_nop(ring, (12 - (ring->wptr & 7)) % 8); + amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_INDIRECT_BUFFER, 0, extra_bits)); amdgpu_ring_write(ring, ib->gpu_addr & 0xffffffe0); /* base must be 32 byte aligned */ amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xffffffff); @@ -501,6 +514,8 @@ static int cik_sdma_load_microcode(struct amdgpu_device *adev) fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4; adev->sdma[i].fw_version = le32_to_cpu(hdr->header.ucode_version); adev->sdma[i].feature_version = le32_to_cpu(hdr->ucode_feature_version); + if (adev->sdma[i].feature_version >= 20) + adev->sdma[i].burst_nop = true; fw_data = (const __le32 *) (adev->sdma[i].fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes)); WREG32(mmSDMA0_UCODE_ADDR + sdma_offsets[i], 0); @@ -630,6 +645,7 @@ static int cik_sdma_ring_test_ib(struct amdgpu_ring *ring) gpu_addr = adev->wb.gpu_addr + (index * 4); tmp = 0xCAFEDEAD; adev->wb.wb[index] = cpu_to_le32(tmp); + memset(&ib, 0, sizeof(ib)); r = amdgpu_ib_get(ring, NULL, 256, &ib); if (r) { DRM_ERROR("amdgpu: failed to get ib (%d).\n", r); @@ -814,8 +830,19 @@ static void cik_sdma_vm_set_pte_pde(struct amdgpu_ib *ib, */ static void cik_sdma_vm_pad_ib(struct amdgpu_ib *ib) { - while (ib->length_dw & 0x7) - ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0); + struct amdgpu_sdma *sdma = amdgpu_get_sdma_instance(ib->ring); + u32 pad_count; + int i; + + pad_count = (8 - (ib->length_dw & 0x7)) % 8; + for (i = 0; i < pad_count; i++) + if (sdma && sdma->burst_nop && (i == 0)) + ib->ptr[ib->length_dw++] = + SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0) | + SDMA_NOP_COUNT(pad_count - 1); + else + ib->ptr[ib->length_dw++] = + SDMA_PACKET(SDMA_OPCODE_NOP, 0, 0); } /** @@ -1302,6 +1329,7 @@ static const struct amdgpu_ring_funcs cik_sdma_ring_funcs = { .test_ring = cik_sdma_ring_test_ring, .test_ib = cik_sdma_ring_test_ib, .is_lockup = cik_sdma_ring_is_lockup, + .insert_nop = cik_sdma_ring_insert_nop, }; static void cik_sdma_set_ring_funcs(struct amdgpu_device *adev) @@ -1338,18 +1366,18 @@ static void cik_sdma_set_irq_funcs(struct amdgpu_device *adev) * Used by the amdgpu ttm implementation to move pages if * registered as the asic copy callback. */ -static void cik_sdma_emit_copy_buffer(struct amdgpu_ring *ring, +static void cik_sdma_emit_copy_buffer(struct amdgpu_ib *ib, uint64_t src_offset, uint64_t dst_offset, uint32_t byte_count) { - amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_COPY, SDMA_COPY_SUB_OPCODE_LINEAR, 0)); - amdgpu_ring_write(ring, byte_count); - amdgpu_ring_write(ring, 0); /* src/dst endian swap */ - amdgpu_ring_write(ring, lower_32_bits(src_offset)); - amdgpu_ring_write(ring, upper_32_bits(src_offset)); - amdgpu_ring_write(ring, lower_32_bits(dst_offset)); - amdgpu_ring_write(ring, upper_32_bits(dst_offset)); + ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_COPY, SDMA_COPY_SUB_OPCODE_LINEAR, 0); + ib->ptr[ib->length_dw++] = byte_count; + ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */ + ib->ptr[ib->length_dw++] = lower_32_bits(src_offset); + ib->ptr[ib->length_dw++] = upper_32_bits(src_offset); + ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset); + ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset); } /** @@ -1362,16 +1390,16 @@ static void cik_sdma_emit_copy_buffer(struct amdgpu_ring *ring, * * Fill GPU buffers using the DMA engine (CIK). */ -static void cik_sdma_emit_fill_buffer(struct amdgpu_ring *ring, +static void cik_sdma_emit_fill_buffer(struct amdgpu_ib *ib, uint32_t src_data, uint64_t dst_offset, uint32_t byte_count) { - amdgpu_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_CONSTANT_FILL, 0, 0)); - amdgpu_ring_write(ring, lower_32_bits(dst_offset)); - amdgpu_ring_write(ring, upper_32_bits(dst_offset)); - amdgpu_ring_write(ring, src_data); - amdgpu_ring_write(ring, byte_count); + ib->ptr[ib->length_dw++] = SDMA_PACKET(SDMA_OPCODE_CONSTANT_FILL, 0, 0); + ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset); + ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset); + ib->ptr[ib->length_dw++] = src_data; + ib->ptr[ib->length_dw++] = byte_count; } static const struct amdgpu_buffer_funcs cik_sdma_buffer_funcs = { diff --git a/drivers/gpu/drm/amd/amdgpu/cikd.h b/drivers/gpu/drm/amd/amdgpu/cikd.h index a3e3dfaa01a4..7f6d457f250a 100644 --- a/drivers/gpu/drm/amd/amdgpu/cikd.h +++ b/drivers/gpu/drm/amd/amdgpu/cikd.h @@ -487,6 +487,7 @@ (((op) & 0xFF) << 0)) /* sDMA opcodes */ #define SDMA_OPCODE_NOP 0 +# define SDMA_NOP_COUNT(x) (((x) & 0x3FFF) << 16) #define SDMA_OPCODE_COPY 1 # define SDMA_COPY_SUB_OPCODE_LINEAR 0 # define SDMA_COPY_SUB_OPCODE_TILED 1 diff --git a/drivers/gpu/drm/amd/amdgpu/cz_dpm.c b/drivers/gpu/drm/amd/amdgpu/cz_dpm.c index ace870afc7d4..44fa96ad4709 100644 --- a/drivers/gpu/drm/amd/amdgpu/cz_dpm.c +++ b/drivers/gpu/drm/amd/amdgpu/cz_dpm.c @@ -1596,9 +1596,9 @@ static int cz_dpm_update_low_memory_pstate(struct amdgpu_device *adev) if (pi->sys_info.nb_dpm_enable) { if (ps->force_high) - cz_dpm_nbdpm_lm_pstate_enable(adev, true); - else cz_dpm_nbdpm_lm_pstate_enable(adev, false); + else + cz_dpm_nbdpm_lm_pstate_enable(adev, true); } return ret; diff --git a/drivers/gpu/drm/amd/amdgpu/cz_smc.c b/drivers/gpu/drm/amd/amdgpu/cz_smc.c index a72ffc7d6c26..e33180d3314a 100644 --- a/drivers/gpu/drm/amd/amdgpu/cz_smc.c +++ b/drivers/gpu/drm/amd/amdgpu/cz_smc.c @@ -814,7 +814,8 @@ int cz_smu_init(struct amdgpu_device *adev) * 3. map kernel virtual address */ ret = amdgpu_bo_create(adev, priv->toc_buffer.data_size, PAGE_SIZE, - true, AMDGPU_GEM_DOMAIN_GTT, 0, NULL, toc_buf); + true, AMDGPU_GEM_DOMAIN_GTT, 0, NULL, NULL, + toc_buf); if (ret) { dev_err(adev->dev, "(%d) SMC TOC buffer allocation failed\n", ret); @@ -822,7 +823,8 @@ int cz_smu_init(struct amdgpu_device *adev) } ret = amdgpu_bo_create(adev, priv->smu_buffer.data_size, PAGE_SIZE, - true, AMDGPU_GEM_DOMAIN_GTT, 0, NULL, smu_buf); + true, AMDGPU_GEM_DOMAIN_GTT, 0, NULL, NULL, + smu_buf); if (ret) { dev_err(adev->dev, "(%d) SMC Internal buffer allocation failed\n", ret); diff --git a/drivers/gpu/drm/amd/amdgpu/dce_v10_0.c b/drivers/gpu/drm/amd/amdgpu/dce_v10_0.c index 4b255ac3043c..e4d101b1252a 100644 --- a/drivers/gpu/drm/amd/amdgpu/dce_v10_0.c +++ b/drivers/gpu/drm/amd/amdgpu/dce_v10_0.c @@ -1353,7 +1353,7 @@ static void dce_v10_0_program_watermarks(struct amdgpu_device *adev, tmp = REG_SET_FIELD(wm_mask, DPG_WATERMARK_MASK_CONTROL, URGENCY_WATERMARK_MASK, 2); WREG32(mmDPG_WATERMARK_MASK_CONTROL + amdgpu_crtc->crtc_offset, tmp); tmp = RREG32(mmDPG_PIPE_URGENCY_CONTROL + amdgpu_crtc->crtc_offset); - tmp = REG_SET_FIELD(tmp, DPG_PIPE_URGENCY_CONTROL, URGENCY_LOW_WATERMARK, latency_watermark_a); + tmp = REG_SET_FIELD(tmp, DPG_PIPE_URGENCY_CONTROL, URGENCY_LOW_WATERMARK, latency_watermark_b); tmp = REG_SET_FIELD(tmp, DPG_PIPE_URGENCY_CONTROL, URGENCY_HIGH_WATERMARK, line_time); WREG32(mmDPG_PIPE_URGENCY_CONTROL + amdgpu_crtc->crtc_offset, tmp); /* restore original selection */ diff --git a/drivers/gpu/drm/amd/amdgpu/dce_v11_0.c b/drivers/gpu/drm/amd/amdgpu/dce_v11_0.c index 70eee807421f..6411e8244671 100644 --- a/drivers/gpu/drm/amd/amdgpu/dce_v11_0.c +++ b/drivers/gpu/drm/amd/amdgpu/dce_v11_0.c @@ -1329,7 +1329,7 @@ static void dce_v11_0_program_watermarks(struct amdgpu_device *adev, tmp = REG_SET_FIELD(wm_mask, DPG_WATERMARK_MASK_CONTROL, URGENCY_WATERMARK_MASK, 2); WREG32(mmDPG_WATERMARK_MASK_CONTROL + amdgpu_crtc->crtc_offset, tmp); tmp = RREG32(mmDPG_PIPE_URGENCY_CONTROL + amdgpu_crtc->crtc_offset); - tmp = REG_SET_FIELD(tmp, DPG_PIPE_URGENCY_CONTROL, URGENCY_LOW_WATERMARK, latency_watermark_a); + tmp = REG_SET_FIELD(tmp, DPG_PIPE_URGENCY_CONTROL, URGENCY_LOW_WATERMARK, latency_watermark_b); tmp = REG_SET_FIELD(tmp, DPG_PIPE_URGENCY_CONTROL, URGENCY_HIGH_WATERMARK, line_time); WREG32(mmDPG_PIPE_URGENCY_CONTROL + amdgpu_crtc->crtc_offset, tmp); /* restore original selection */ diff --git a/drivers/gpu/drm/amd/amdgpu/fiji_smc.c b/drivers/gpu/drm/amd/amdgpu/fiji_smc.c index 493c8c9c7faa..bda1249eb871 100644 --- a/drivers/gpu/drm/amd/amdgpu/fiji_smc.c +++ b/drivers/gpu/drm/amd/amdgpu/fiji_smc.c @@ -762,7 +762,9 @@ int fiji_smu_init(struct amdgpu_device *adev) /* Allocate FW image data structure and header buffer */ ret = amdgpu_bo_create(adev, image_size, PAGE_SIZE, - true, AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, toc_buf); + true, AMDGPU_GEM_DOMAIN_VRAM, + AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED, + NULL, NULL, toc_buf); if (ret) { DRM_ERROR("Failed to allocate memory for TOC buffer\n"); return -ENOMEM; @@ -770,7 +772,9 @@ int fiji_smu_init(struct amdgpu_device *adev) /* Allocate buffer for SMU internal buffer */ ret = amdgpu_bo_create(adev, smu_internal_buffer_size, PAGE_SIZE, - true, AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, smu_buf); + true, AMDGPU_GEM_DOMAIN_VRAM, + AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED, + NULL, NULL, smu_buf); if (ret) { DRM_ERROR("Failed to allocate memory for SMU internal buffer\n"); return -ENOMEM; diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c index 9b0cab413677..e992bf2ff66c 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c @@ -2660,6 +2660,7 @@ static int gfx_v7_0_ring_test_ib(struct amdgpu_ring *ring) return r; } WREG32(scratch, 0xCAFEDEAD); + memset(&ib, 0, sizeof(ib)); r = amdgpu_ib_get(ring, NULL, 256, &ib); if (r) { DRM_ERROR("amdgpu: failed to get ib (%d).\n", r); @@ -3205,7 +3206,7 @@ static int gfx_v7_0_mec_init(struct amdgpu_device *adev) r = amdgpu_bo_create(adev, adev->gfx.mec.num_mec *adev->gfx.mec.num_pipe * MEC_HPD_SIZE * 2, PAGE_SIZE, true, - AMDGPU_GEM_DOMAIN_GTT, 0, NULL, + AMDGPU_GEM_DOMAIN_GTT, 0, NULL, NULL, &adev->gfx.mec.hpd_eop_obj); if (r) { dev_warn(adev->dev, "(%d) create HDP EOP bo failed\n", r); @@ -3372,7 +3373,7 @@ static int gfx_v7_0_cp_compute_resume(struct amdgpu_device *adev) r = amdgpu_bo_create(adev, sizeof(struct bonaire_mqd), PAGE_SIZE, true, - AMDGPU_GEM_DOMAIN_GTT, 0, NULL, + AMDGPU_GEM_DOMAIN_GTT, 0, NULL, NULL, &ring->mqd_obj); if (r) { dev_warn(adev->dev, "(%d) create MQD bo failed\n", r); @@ -3609,41 +3610,6 @@ static int gfx_v7_0_cp_resume(struct amdgpu_device *adev) return 0; } -static void gfx_v7_0_ce_sync_me(struct amdgpu_ring *ring) -{ - struct amdgpu_device *adev = ring->adev; - u64 gpu_addr = adev->wb.gpu_addr + adev->gfx.ce_sync_offs * 4; - - /* instruct DE to set a magic number */ - amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3)); - amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) | - WRITE_DATA_DST_SEL(5))); - amdgpu_ring_write(ring, gpu_addr & 0xfffffffc); - amdgpu_ring_write(ring, upper_32_bits(gpu_addr) & 0xffffffff); - amdgpu_ring_write(ring, 1); - - /* let CE wait till condition satisfied */ - amdgpu_ring_write(ring, PACKET3(PACKET3_WAIT_REG_MEM, 5)); - amdgpu_ring_write(ring, (WAIT_REG_MEM_OPERATION(0) | /* wait */ - WAIT_REG_MEM_MEM_SPACE(1) | /* memory */ - WAIT_REG_MEM_FUNCTION(3) | /* == */ - WAIT_REG_MEM_ENGINE(2))); /* ce */ - amdgpu_ring_write(ring, gpu_addr & 0xfffffffc); - amdgpu_ring_write(ring, upper_32_bits(gpu_addr) & 0xffffffff); - amdgpu_ring_write(ring, 1); - amdgpu_ring_write(ring, 0xffffffff); - amdgpu_ring_write(ring, 4); /* poll interval */ - - /* instruct CE to reset wb of ce_sync to zero */ - amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3)); - amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(2) | - WRITE_DATA_DST_SEL(5) | - WR_CONFIRM)); - amdgpu_ring_write(ring, gpu_addr & 0xfffffffc); - amdgpu_ring_write(ring, upper_32_bits(gpu_addr) & 0xffffffff); - amdgpu_ring_write(ring, 0); -} - /* * vm * VMID 0 is the physical GPU addresses as used by the kernel. @@ -3662,6 +3628,13 @@ static void gfx_v7_0_ring_emit_vm_flush(struct amdgpu_ring *ring, unsigned vm_id, uint64_t pd_addr) { int usepfp = (ring->type == AMDGPU_RING_TYPE_GFX); + if (usepfp) { + /* synce CE with ME to prevent CE fetch CEIB before context switch done */ + amdgpu_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0)); + amdgpu_ring_write(ring, 0); + amdgpu_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0)); + amdgpu_ring_write(ring, 0); + } amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3)); amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(usepfp) | @@ -3702,7 +3675,10 @@ static void gfx_v7_0_ring_emit_vm_flush(struct amdgpu_ring *ring, amdgpu_ring_write(ring, 0x0); /* synce CE with ME to prevent CE fetch CEIB before context switch done */ - gfx_v7_0_ce_sync_me(ring); + amdgpu_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0)); + amdgpu_ring_write(ring, 0); + amdgpu_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0)); + amdgpu_ring_write(ring, 0); } } @@ -3785,7 +3761,10 @@ static int gfx_v7_0_rlc_init(struct amdgpu_device *adev) /* save restore block */ if (adev->gfx.rlc.save_restore_obj == NULL) { r = amdgpu_bo_create(adev, dws * 4, PAGE_SIZE, true, - AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, &adev->gfx.rlc.save_restore_obj); + AMDGPU_GEM_DOMAIN_VRAM, + AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED, + NULL, NULL, + &adev->gfx.rlc.save_restore_obj); if (r) { dev_warn(adev->dev, "(%d) create RLC sr bo failed\n", r); return r; @@ -3826,7 +3805,10 @@ static int gfx_v7_0_rlc_init(struct amdgpu_device *adev) if (adev->gfx.rlc.clear_state_obj == NULL) { r = amdgpu_bo_create(adev, dws * 4, PAGE_SIZE, true, - AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, &adev->gfx.rlc.clear_state_obj); + AMDGPU_GEM_DOMAIN_VRAM, + AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED, + NULL, NULL, + &adev->gfx.rlc.clear_state_obj); if (r) { dev_warn(adev->dev, "(%d) create RLC c bo failed\n", r); gfx_v7_0_rlc_fini(adev); @@ -3863,7 +3845,10 @@ static int gfx_v7_0_rlc_init(struct amdgpu_device *adev) if (adev->gfx.rlc.cp_table_size) { if (adev->gfx.rlc.cp_table_obj == NULL) { r = amdgpu_bo_create(adev, adev->gfx.rlc.cp_table_size, PAGE_SIZE, true, - AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, &adev->gfx.rlc.cp_table_obj); + AMDGPU_GEM_DOMAIN_VRAM, + AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED, + NULL, NULL, + &adev->gfx.rlc.cp_table_obj); if (r) { dev_warn(adev->dev, "(%d) create RLC cp table bo failed\n", r); gfx_v7_0_rlc_fini(adev); @@ -4795,12 +4780,6 @@ static int gfx_v7_0_sw_init(void *handle) return r; } - r = amdgpu_wb_get(adev, &adev->gfx.ce_sync_offs); - if (r) { - DRM_ERROR("(%d) gfx.ce_sync_offs wb alloc failed\n", r); - return r; - } - for (i = 0; i < adev->gfx.num_gfx_rings; i++) { ring = &adev->gfx.gfx_ring[i]; ring->ring_obj = NULL; @@ -4844,21 +4823,21 @@ static int gfx_v7_0_sw_init(void *handle) r = amdgpu_bo_create(adev, adev->gds.mem.gfx_partition_size, PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_GDS, 0, - NULL, &adev->gds.gds_gfx_bo); + NULL, NULL, &adev->gds.gds_gfx_bo); if (r) return r; r = amdgpu_bo_create(adev, adev->gds.gws.gfx_partition_size, PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_GWS, 0, - NULL, &adev->gds.gws_gfx_bo); + NULL, NULL, &adev->gds.gws_gfx_bo); if (r) return r; r = amdgpu_bo_create(adev, adev->gds.oa.gfx_partition_size, PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_OA, 0, - NULL, &adev->gds.oa_gfx_bo); + NULL, NULL, &adev->gds.oa_gfx_bo); if (r) return r; @@ -4879,8 +4858,6 @@ static int gfx_v7_0_sw_fini(void *handle) for (i = 0; i < adev->gfx.num_compute_rings; i++) amdgpu_ring_fini(&adev->gfx.compute_ring[i]); - amdgpu_wb_free(adev, adev->gfx.ce_sync_offs); - gfx_v7_0_cp_compute_fini(adev); gfx_v7_0_rlc_fini(adev); gfx_v7_0_mec_fini(adev); @@ -5597,6 +5574,7 @@ static const struct amdgpu_ring_funcs gfx_v7_0_ring_funcs_gfx = { .test_ring = gfx_v7_0_ring_test_ring, .test_ib = gfx_v7_0_ring_test_ib, .is_lockup = gfx_v7_0_ring_is_lockup, + .insert_nop = amdgpu_ring_insert_nop, }; static const struct amdgpu_ring_funcs gfx_v7_0_ring_funcs_compute = { @@ -5613,6 +5591,7 @@ static const struct amdgpu_ring_funcs gfx_v7_0_ring_funcs_compute = { .test_ring = gfx_v7_0_ring_test_ring, .test_ib = gfx_v7_0_ring_test_ib, .is_lockup = gfx_v7_0_ring_is_lockup, + .insert_nop = amdgpu_ring_insert_nop, }; static void gfx_v7_0_set_ring_funcs(struct amdgpu_device *adev) diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c index 4b68e6306f40..cb4f68f53f24 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c @@ -622,6 +622,7 @@ static int gfx_v8_0_ring_test_ib(struct amdgpu_ring *ring) return r; } WREG32(scratch, 0xCAFEDEAD); + memset(&ib, 0, sizeof(ib)); r = amdgpu_ib_get(ring, NULL, 256, &ib); if (r) { DRM_ERROR("amdgpu: failed to get ib (%d).\n", r); @@ -867,7 +868,7 @@ static int gfx_v8_0_mec_init(struct amdgpu_device *adev) r = amdgpu_bo_create(adev, adev->gfx.mec.num_mec *adev->gfx.mec.num_pipe * MEC_HPD_SIZE * 2, PAGE_SIZE, true, - AMDGPU_GEM_DOMAIN_GTT, 0, NULL, + AMDGPU_GEM_DOMAIN_GTT, 0, NULL, NULL, &adev->gfx.mec.hpd_eop_obj); if (r) { dev_warn(adev->dev, "(%d) create HDP EOP bo failed\n", r); @@ -939,12 +940,6 @@ static int gfx_v8_0_sw_init(void *handle) return r; } - r = amdgpu_wb_get(adev, &adev->gfx.ce_sync_offs); - if (r) { - DRM_ERROR("(%d) gfx.ce_sync_offs wb alloc failed\n", r); - return r; - } - /* set up the gfx ring */ for (i = 0; i < adev->gfx.num_gfx_rings; i++) { ring = &adev->gfx.gfx_ring[i]; @@ -994,21 +989,21 @@ static int gfx_v8_0_sw_init(void *handle) /* reserve GDS, GWS and OA resource for gfx */ r = amdgpu_bo_create(adev, adev->gds.mem.gfx_partition_size, PAGE_SIZE, true, - AMDGPU_GEM_DOMAIN_GDS, 0, + AMDGPU_GEM_DOMAIN_GDS, 0, NULL, NULL, &adev->gds.gds_gfx_bo); if (r) return r; r = amdgpu_bo_create(adev, adev->gds.gws.gfx_partition_size, PAGE_SIZE, true, - AMDGPU_GEM_DOMAIN_GWS, 0, + AMDGPU_GEM_DOMAIN_GWS, 0, NULL, NULL, &adev->gds.gws_gfx_bo); if (r) return r; r = amdgpu_bo_create(adev, adev->gds.oa.gfx_partition_size, PAGE_SIZE, true, - AMDGPU_GEM_DOMAIN_OA, 0, + AMDGPU_GEM_DOMAIN_OA, 0, NULL, NULL, &adev->gds.oa_gfx_bo); if (r) return r; @@ -1032,8 +1027,6 @@ static int gfx_v8_0_sw_fini(void *handle) for (i = 0; i < adev->gfx.num_compute_rings; i++) amdgpu_ring_fini(&adev->gfx.compute_ring[i]); - amdgpu_wb_free(adev, adev->gfx.ce_sync_offs); - gfx_v8_0_mec_fini(adev); return 0; @@ -2004,7 +1997,7 @@ static void gfx_v8_0_setup_rb(struct amdgpu_device *adev, } /** - * gmc_v8_0_init_compute_vmid - gart enable + * gfx_v8_0_init_compute_vmid - gart enable * * @rdev: amdgpu_device pointer * @@ -2014,7 +2007,7 @@ static void gfx_v8_0_setup_rb(struct amdgpu_device *adev, #define DEFAULT_SH_MEM_BASES (0x6000) #define FIRST_COMPUTE_VMID (8) #define LAST_COMPUTE_VMID (16) -static void gmc_v8_0_init_compute_vmid(struct amdgpu_device *adev) +static void gfx_v8_0_init_compute_vmid(struct amdgpu_device *adev) { int i; uint32_t sh_mem_config; @@ -2281,7 +2274,7 @@ static void gfx_v8_0_gpu_init(struct amdgpu_device *adev) vi_srbm_select(adev, 0, 0, 0, 0); mutex_unlock(&adev->srbm_mutex); - gmc_v8_0_init_compute_vmid(adev); + gfx_v8_0_init_compute_vmid(adev); mutex_lock(&adev->grbm_idx_mutex); /* @@ -3105,7 +3098,7 @@ static int gfx_v8_0_cp_compute_resume(struct amdgpu_device *adev) sizeof(struct vi_mqd), PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_GTT, 0, NULL, - &ring->mqd_obj); + NULL, &ring->mqd_obj); if (r) { dev_warn(adev->dev, "(%d) create MQD bo failed\n", r); return r; @@ -3239,7 +3232,8 @@ static int gfx_v8_0_cp_compute_resume(struct amdgpu_device *adev) /* enable the doorbell if requested */ if (use_doorbell) { - if (adev->asic_type == CHIP_CARRIZO) { + if ((adev->asic_type == CHIP_CARRIZO) || + (adev->asic_type == CHIP_FIJI)) { WREG32(mmCP_MEC_DOORBELL_RANGE_LOWER, AMDGPU_DOORBELL_KIQ << 2); WREG32(mmCP_MEC_DOORBELL_RANGE_UPPER, @@ -3963,6 +3957,7 @@ static void gfx_v8_0_ring_emit_fence_gfx(struct amdgpu_ring *ring, u64 addr, DATA_SEL(write64bit ? 2 : 1) | INT_SEL(int_sel ? 2 : 0)); amdgpu_ring_write(ring, lower_32_bits(seq)); amdgpu_ring_write(ring, upper_32_bits(seq)); + } /** @@ -4003,49 +3998,34 @@ static bool gfx_v8_0_ring_emit_semaphore(struct amdgpu_ring *ring, return true; } -static void gfx_v8_0_ce_sync_me(struct amdgpu_ring *ring) +static void gfx_v8_0_ring_emit_vm_flush(struct amdgpu_ring *ring, + unsigned vm_id, uint64_t pd_addr) { - struct amdgpu_device *adev = ring->adev; - u64 gpu_addr = adev->wb.gpu_addr + adev->gfx.ce_sync_offs * 4; - - /* instruct DE to set a magic number */ - amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3)); - amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) | - WRITE_DATA_DST_SEL(5))); - amdgpu_ring_write(ring, gpu_addr & 0xfffffffc); - amdgpu_ring_write(ring, upper_32_bits(gpu_addr) & 0xffffffff); - amdgpu_ring_write(ring, 1); + int usepfp = (ring->type == AMDGPU_RING_TYPE_GFX); + uint32_t seq = ring->fence_drv.sync_seq[ring->idx]; + uint64_t addr = ring->fence_drv.gpu_addr; - /* let CE wait till condition satisfied */ amdgpu_ring_write(ring, PACKET3(PACKET3_WAIT_REG_MEM, 5)); - amdgpu_ring_write(ring, (WAIT_REG_MEM_OPERATION(0) | /* wait */ - WAIT_REG_MEM_MEM_SPACE(1) | /* memory */ - WAIT_REG_MEM_FUNCTION(3) | /* == */ - WAIT_REG_MEM_ENGINE(2))); /* ce */ - amdgpu_ring_write(ring, gpu_addr & 0xfffffffc); - amdgpu_ring_write(ring, upper_32_bits(gpu_addr) & 0xffffffff); - amdgpu_ring_write(ring, 1); + amdgpu_ring_write(ring, (WAIT_REG_MEM_MEM_SPACE(1) | /* memory */ + WAIT_REG_MEM_FUNCTION(3))); /* equal */ + amdgpu_ring_write(ring, addr & 0xfffffffc); + amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff); + amdgpu_ring_write(ring, seq); amdgpu_ring_write(ring, 0xffffffff); amdgpu_ring_write(ring, 4); /* poll interval */ - /* instruct CE to reset wb of ce_sync to zero */ - amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3)); - amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(2) | - WRITE_DATA_DST_SEL(5) | - WR_CONFIRM)); - amdgpu_ring_write(ring, gpu_addr & 0xfffffffc); - amdgpu_ring_write(ring, upper_32_bits(gpu_addr) & 0xffffffff); - amdgpu_ring_write(ring, 0); -} - -static void gfx_v8_0_ring_emit_vm_flush(struct amdgpu_ring *ring, - unsigned vm_id, uint64_t pd_addr) -{ - int usepfp = (ring->type == AMDGPU_RING_TYPE_GFX); + if (usepfp) { + /* synce CE with ME to prevent CE fetch CEIB before context switch done */ + amdgpu_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0)); + amdgpu_ring_write(ring, 0); + amdgpu_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0)); + amdgpu_ring_write(ring, 0); + } amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3)); amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(usepfp) | - WRITE_DATA_DST_SEL(0))); + WRITE_DATA_DST_SEL(0)) | + WR_CONFIRM); if (vm_id < 8) { amdgpu_ring_write(ring, (mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vm_id)); @@ -4081,9 +4061,10 @@ static void gfx_v8_0_ring_emit_vm_flush(struct amdgpu_ring *ring, /* sync PFP to ME, otherwise we might get invalid PFP reads */ amdgpu_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0)); amdgpu_ring_write(ring, 0x0); - - /* synce CE with ME to prevent CE fetch CEIB before context switch done */ - gfx_v8_0_ce_sync_me(ring); + amdgpu_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0)); + amdgpu_ring_write(ring, 0); + amdgpu_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0)); + amdgpu_ring_write(ring, 0); } } @@ -4377,6 +4358,7 @@ static const struct amdgpu_ring_funcs gfx_v8_0_ring_funcs_gfx = { .test_ring = gfx_v8_0_ring_test_ring, .test_ib = gfx_v8_0_ring_test_ib, .is_lockup = gfx_v8_0_ring_is_lockup, + .insert_nop = amdgpu_ring_insert_nop, }; static const struct amdgpu_ring_funcs gfx_v8_0_ring_funcs_compute = { @@ -4393,6 +4375,7 @@ static const struct amdgpu_ring_funcs gfx_v8_0_ring_funcs_compute = { .test_ring = gfx_v8_0_ring_test_ring, .test_ib = gfx_v8_0_ring_test_ib, .is_lockup = gfx_v8_0_ring_is_lockup, + .insert_nop = amdgpu_ring_insert_nop, }; static void gfx_v8_0_set_ring_funcs(struct amdgpu_device *adev) diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c index 10218828face..774528ab8704 100644 --- a/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c @@ -523,17 +523,11 @@ static int gmc_v7_0_gart_enable(struct amdgpu_device *adev) tmp = RREG32(mmVM_CONTEXT1_CNTL); tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, ENABLE_CONTEXT, 1); tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_DEPTH, 1); - tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT, 1); tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); - tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT, 1); tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); - tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT, 1); tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, 1); - tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, VALID_PROTECTION_FAULT_ENABLE_INTERRUPT, 1); tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, VALID_PROTECTION_FAULT_ENABLE_DEFAULT, 1); - tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, READ_PROTECTION_FAULT_ENABLE_INTERRUPT, 1); tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, READ_PROTECTION_FAULT_ENABLE_DEFAULT, 1); - tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT, 1); tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_BLOCK_SIZE, amdgpu_vm_block_size - 9); @@ -852,6 +846,13 @@ static int gmc_v7_0_early_init(void *handle) return 0; } +static int gmc_v7_0_late_init(void *handle) +{ + struct amdgpu_device *adev = (struct amdgpu_device *)handle; + + return amdgpu_irq_get(adev, &adev->mc.vm_fault, 0); +} + static int gmc_v7_0_sw_init(void *handle) { int r; @@ -976,6 +977,7 @@ static int gmc_v7_0_hw_fini(void *handle) { struct amdgpu_device *adev = (struct amdgpu_device *)handle; + amdgpu_irq_put(adev, &adev->mc.vm_fault, 0); gmc_v7_0_gart_disable(adev); return 0; @@ -1301,7 +1303,7 @@ static int gmc_v7_0_set_powergating_state(void *handle, const struct amd_ip_funcs gmc_v7_0_ip_funcs = { .early_init = gmc_v7_0_early_init, - .late_init = NULL, + .late_init = gmc_v7_0_late_init, .sw_init = gmc_v7_0_sw_init, .sw_fini = gmc_v7_0_sw_fini, .hw_init = gmc_v7_0_hw_init, diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c index 78109b750d29..9a07742620d0 100644 --- a/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c @@ -653,19 +653,12 @@ static int gmc_v8_0_gart_enable(struct amdgpu_device *adev) tmp = RREG32(mmVM_CONTEXT1_CNTL); tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, ENABLE_CONTEXT, 1); tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_DEPTH, 1); - tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT, 1); tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); - tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT, 1); tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); - tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT, 1); tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, 1); - tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, VALID_PROTECTION_FAULT_ENABLE_INTERRUPT, 1); tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, VALID_PROTECTION_FAULT_ENABLE_DEFAULT, 1); - tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, READ_PROTECTION_FAULT_ENABLE_INTERRUPT, 1); tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, READ_PROTECTION_FAULT_ENABLE_DEFAULT, 1); - tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT, 1); tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); - tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT, 1); tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, EXECUTE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_BLOCK_SIZE, amdgpu_vm_block_size - 9); @@ -852,6 +845,13 @@ static int gmc_v8_0_early_init(void *handle) return 0; } +static int gmc_v8_0_late_init(void *handle) +{ + struct amdgpu_device *adev = (struct amdgpu_device *)handle; + + return amdgpu_irq_get(adev, &adev->mc.vm_fault, 0); +} + static int gmc_v8_0_sw_init(void *handle) { int r; @@ -978,6 +978,7 @@ static int gmc_v8_0_hw_fini(void *handle) { struct amdgpu_device *adev = (struct amdgpu_device *)handle; + amdgpu_irq_put(adev, &adev->mc.vm_fault, 0); gmc_v8_0_gart_disable(adev); return 0; @@ -1288,7 +1289,7 @@ static int gmc_v8_0_set_powergating_state(void *handle, const struct amd_ip_funcs gmc_v8_0_ip_funcs = { .early_init = gmc_v8_0_early_init, - .late_init = NULL, + .late_init = gmc_v8_0_late_init, .sw_init = gmc_v8_0_sw_init, .sw_fini = gmc_v8_0_sw_fini, .hw_init = gmc_v8_0_hw_init, diff --git a/drivers/gpu/drm/amd/amdgpu/iceland_sdma_pkt_open.h b/drivers/gpu/drm/amd/amdgpu/iceland_sdma_pkt_open.h index c723602c7b0c..ee6a041cb288 100644 --- a/drivers/gpu/drm/amd/amdgpu/iceland_sdma_pkt_open.h +++ b/drivers/gpu/drm/amd/amdgpu/iceland_sdma_pkt_open.h @@ -2163,5 +2163,10 @@ #define SDMA_PKT_NOP_HEADER_sub_op_shift 8 #define SDMA_PKT_NOP_HEADER_SUB_OP(x) (((x) & SDMA_PKT_NOP_HEADER_sub_op_mask) << SDMA_PKT_NOP_HEADER_sub_op_shift) +/*define for count field*/ +#define SDMA_PKT_NOP_HEADER_count_offset 0 +#define SDMA_PKT_NOP_HEADER_count_mask 0x00003FFF +#define SDMA_PKT_NOP_HEADER_count_shift 16 +#define SDMA_PKT_NOP_HEADER_COUNT(x) (((x) & SDMA_PKT_NOP_HEADER_count_mask) << SDMA_PKT_NOP_HEADER_count_shift) #endif /* __ICELAND_SDMA_PKT_OPEN_H_ */ diff --git a/drivers/gpu/drm/amd/amdgpu/iceland_smc.c b/drivers/gpu/drm/amd/amdgpu/iceland_smc.c index c6f1e2f12b5f..966d4b2ed9da 100644 --- a/drivers/gpu/drm/amd/amdgpu/iceland_smc.c +++ b/drivers/gpu/drm/amd/amdgpu/iceland_smc.c @@ -623,7 +623,9 @@ int iceland_smu_init(struct amdgpu_device *adev) /* Allocate FW image data structure and header buffer */ ret = amdgpu_bo_create(adev, image_size, PAGE_SIZE, - true, AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, toc_buf); + true, AMDGPU_GEM_DOMAIN_VRAM, + AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED, + NULL, NULL, toc_buf); if (ret) { DRM_ERROR("Failed to allocate memory for TOC buffer\n"); return -ENOMEM; diff --git a/drivers/gpu/drm/amd/amdgpu/sdma_v2_4.c b/drivers/gpu/drm/amd/amdgpu/sdma_v2_4.c index 9de8104eddeb..14e87234171a 100644 --- a/drivers/gpu/drm/amd/amdgpu/sdma_v2_4.c +++ b/drivers/gpu/drm/amd/amdgpu/sdma_v2_4.c @@ -146,6 +146,8 @@ static int sdma_v2_4_init_microcode(struct amdgpu_device *adev) hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma[i].fw->data; adev->sdma[i].fw_version = le32_to_cpu(hdr->header.ucode_version); adev->sdma[i].feature_version = le32_to_cpu(hdr->ucode_feature_version); + if (adev->sdma[i].feature_version >= 20) + adev->sdma[i].burst_nop = true; if (adev->firmware.smu_load) { info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i]; @@ -218,6 +220,19 @@ static void sdma_v2_4_ring_set_wptr(struct amdgpu_ring *ring) WREG32(mmSDMA0_GFX_RB_WPTR + sdma_offsets[me], ring->wptr << 2); } +static void sdma_v2_4_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count) +{ + struct amdgpu_sdma *sdma = amdgpu_get_sdma_instance(ring); + int i; + + for (i = 0; i < count; i++) + if (sdma && sdma->burst_nop && (i == 0)) + amdgpu_ring_write(ring, ring->nop | + SDMA_PKT_NOP_HEADER_COUNT(count - 1)); + else + amdgpu_ring_write(ring, ring->nop); +} + /** * sdma_v2_4_ring_emit_ib - Schedule an IB on the DMA engine * @@ -245,8 +260,8 @@ static void sdma_v2_4_ring_emit_ib(struct amdgpu_ring *ring, amdgpu_ring_write(ring, next_rptr); /* IB packet must end on a 8 DW boundary */ - while ((ring->wptr & 7) != 2) - amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_NOP)); + sdma_v2_4_ring_insert_nop(ring, (10 - (ring->wptr & 7)) % 8); + amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) | SDMA_PKT_INDIRECT_HEADER_VMID(vmid)); /* base must be 32 byte aligned */ @@ -689,6 +704,7 @@ static int sdma_v2_4_ring_test_ib(struct amdgpu_ring *ring) gpu_addr = adev->wb.gpu_addr + (index * 4); tmp = 0xCAFEDEAD; adev->wb.wb[index] = cpu_to_le32(tmp); + memset(&ib, 0, sizeof(ib)); r = amdgpu_ib_get(ring, NULL, 256, &ib); if (r) { DRM_ERROR("amdgpu: failed to get ib (%d).\n", r); @@ -878,8 +894,19 @@ static void sdma_v2_4_vm_set_pte_pde(struct amdgpu_ib *ib, */ static void sdma_v2_4_vm_pad_ib(struct amdgpu_ib *ib) { - while (ib->length_dw & 0x7) - ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP); + struct amdgpu_sdma *sdma = amdgpu_get_sdma_instance(ib->ring); + u32 pad_count; + int i; + + pad_count = (8 - (ib->length_dw & 0x7)) % 8; + for (i = 0; i < pad_count; i++) + if (sdma && sdma->burst_nop && (i == 0)) + ib->ptr[ib->length_dw++] = + SDMA_PKT_HEADER_OP(SDMA_OP_NOP) | + SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1); + else + ib->ptr[ib->length_dw++] = + SDMA_PKT_HEADER_OP(SDMA_OP_NOP); } /** @@ -1313,6 +1340,7 @@ static const struct amdgpu_ring_funcs sdma_v2_4_ring_funcs = { .test_ring = sdma_v2_4_ring_test_ring, .test_ib = sdma_v2_4_ring_test_ib, .is_lockup = sdma_v2_4_ring_is_lockup, + .insert_nop = sdma_v2_4_ring_insert_nop, }; static void sdma_v2_4_set_ring_funcs(struct amdgpu_device *adev) @@ -1349,19 +1377,19 @@ static void sdma_v2_4_set_irq_funcs(struct amdgpu_device *adev) * Used by the amdgpu ttm implementation to move pages if * registered as the asic copy callback. */ -static void sdma_v2_4_emit_copy_buffer(struct amdgpu_ring *ring, +static void sdma_v2_4_emit_copy_buffer(struct amdgpu_ib *ib, uint64_t src_offset, uint64_t dst_offset, uint32_t byte_count) { - amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_COPY) | - SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR)); - amdgpu_ring_write(ring, byte_count); - amdgpu_ring_write(ring, 0); /* src/dst endian swap */ - amdgpu_ring_write(ring, lower_32_bits(src_offset)); - amdgpu_ring_write(ring, upper_32_bits(src_offset)); - amdgpu_ring_write(ring, lower_32_bits(dst_offset)); - amdgpu_ring_write(ring, upper_32_bits(dst_offset)); + ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) | + SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR); + ib->ptr[ib->length_dw++] = byte_count; + ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */ + ib->ptr[ib->length_dw++] = lower_32_bits(src_offset); + ib->ptr[ib->length_dw++] = upper_32_bits(src_offset); + ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset); + ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset); } /** @@ -1374,16 +1402,16 @@ static void sdma_v2_4_emit_copy_buffer(struct amdgpu_ring *ring, * * Fill GPU buffers using the DMA engine (VI). */ -static void sdma_v2_4_emit_fill_buffer(struct amdgpu_ring *ring, +static void sdma_v2_4_emit_fill_buffer(struct amdgpu_ib *ib, uint32_t src_data, uint64_t dst_offset, uint32_t byte_count) { - amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL)); - amdgpu_ring_write(ring, lower_32_bits(dst_offset)); - amdgpu_ring_write(ring, upper_32_bits(dst_offset)); - amdgpu_ring_write(ring, src_data); - amdgpu_ring_write(ring, byte_count); + ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL); + ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset); + ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset); + ib->ptr[ib->length_dw++] = src_data; + ib->ptr[ib->length_dw++] = byte_count; } static const struct amdgpu_buffer_funcs sdma_v2_4_buffer_funcs = { diff --git a/drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c b/drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c index 029f3455f9f9..9bfe92df15f7 100644 --- a/drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c +++ b/drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c @@ -218,6 +218,8 @@ static int sdma_v3_0_init_microcode(struct amdgpu_device *adev) hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma[i].fw->data; adev->sdma[i].fw_version = le32_to_cpu(hdr->header.ucode_version); adev->sdma[i].feature_version = le32_to_cpu(hdr->ucode_feature_version); + if (adev->sdma[i].feature_version >= 20) + adev->sdma[i].burst_nop = true; if (adev->firmware.smu_load) { info = &adev->firmware.ucode[AMDGPU_UCODE_ID_SDMA0 + i]; @@ -304,6 +306,19 @@ static void sdma_v3_0_ring_set_wptr(struct amdgpu_ring *ring) } } +static void sdma_v3_0_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count) +{ + struct amdgpu_sdma *sdma = amdgpu_get_sdma_instance(ring); + int i; + + for (i = 0; i < count; i++) + if (sdma && sdma->burst_nop && (i == 0)) + amdgpu_ring_write(ring, ring->nop | + SDMA_PKT_NOP_HEADER_COUNT(count - 1)); + else + amdgpu_ring_write(ring, ring->nop); +} + /** * sdma_v3_0_ring_emit_ib - Schedule an IB on the DMA engine * @@ -330,8 +345,7 @@ static void sdma_v3_0_ring_emit_ib(struct amdgpu_ring *ring, amdgpu_ring_write(ring, next_rptr); /* IB packet must end on a 8 DW boundary */ - while ((ring->wptr & 7) != 2) - amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_NOP)); + sdma_v3_0_ring_insert_nop(ring, (10 - (ring->wptr & 7)) % 8); amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) | SDMA_PKT_INDIRECT_HEADER_VMID(vmid)); @@ -810,6 +824,7 @@ static int sdma_v3_0_ring_test_ib(struct amdgpu_ring *ring) gpu_addr = adev->wb.gpu_addr + (index * 4); tmp = 0xCAFEDEAD; adev->wb.wb[index] = cpu_to_le32(tmp); + memset(&ib, 0, sizeof(ib)); r = amdgpu_ib_get(ring, NULL, 256, &ib); if (r) { DRM_ERROR("amdgpu: failed to get ib (%d).\n", r); @@ -998,8 +1013,19 @@ static void sdma_v3_0_vm_set_pte_pde(struct amdgpu_ib *ib, */ static void sdma_v3_0_vm_pad_ib(struct amdgpu_ib *ib) { - while (ib->length_dw & 0x7) - ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP); + struct amdgpu_sdma *sdma = amdgpu_get_sdma_instance(ib->ring); + u32 pad_count; + int i; + + pad_count = (8 - (ib->length_dw & 0x7)) % 8; + for (i = 0; i < pad_count; i++) + if (sdma && sdma->burst_nop && (i == 0)) + ib->ptr[ib->length_dw++] = + SDMA_PKT_HEADER_OP(SDMA_OP_NOP) | + SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1); + else + ib->ptr[ib->length_dw++] = + SDMA_PKT_HEADER_OP(SDMA_OP_NOP); } /** @@ -1437,6 +1463,7 @@ static const struct amdgpu_ring_funcs sdma_v3_0_ring_funcs = { .test_ring = sdma_v3_0_ring_test_ring, .test_ib = sdma_v3_0_ring_test_ib, .is_lockup = sdma_v3_0_ring_is_lockup, + .insert_nop = sdma_v3_0_ring_insert_nop, }; static void sdma_v3_0_set_ring_funcs(struct amdgpu_device *adev) @@ -1473,19 +1500,19 @@ static void sdma_v3_0_set_irq_funcs(struct amdgpu_device *adev) * Used by the amdgpu ttm implementation to move pages if * registered as the asic copy callback. */ -static void sdma_v3_0_emit_copy_buffer(struct amdgpu_ring *ring, +static void sdma_v3_0_emit_copy_buffer(struct amdgpu_ib *ib, uint64_t src_offset, uint64_t dst_offset, uint32_t byte_count) { - amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_COPY) | - SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR)); - amdgpu_ring_write(ring, byte_count); - amdgpu_ring_write(ring, 0); /* src/dst endian swap */ - amdgpu_ring_write(ring, lower_32_bits(src_offset)); - amdgpu_ring_write(ring, upper_32_bits(src_offset)); - amdgpu_ring_write(ring, lower_32_bits(dst_offset)); - amdgpu_ring_write(ring, upper_32_bits(dst_offset)); + ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) | + SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR); + ib->ptr[ib->length_dw++] = byte_count; + ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */ + ib->ptr[ib->length_dw++] = lower_32_bits(src_offset); + ib->ptr[ib->length_dw++] = upper_32_bits(src_offset); + ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset); + ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset); } /** @@ -1498,16 +1525,16 @@ static void sdma_v3_0_emit_copy_buffer(struct amdgpu_ring *ring, * * Fill GPU buffers using the DMA engine (VI). */ -static void sdma_v3_0_emit_fill_buffer(struct amdgpu_ring *ring, +static void sdma_v3_0_emit_fill_buffer(struct amdgpu_ib *ib, uint32_t src_data, uint64_t dst_offset, uint32_t byte_count) { - amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL)); - amdgpu_ring_write(ring, lower_32_bits(dst_offset)); - amdgpu_ring_write(ring, upper_32_bits(dst_offset)); - amdgpu_ring_write(ring, src_data); - amdgpu_ring_write(ring, byte_count); + ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL); + ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset); + ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset); + ib->ptr[ib->length_dw++] = src_data; + ib->ptr[ib->length_dw++] = byte_count; } static const struct amdgpu_buffer_funcs sdma_v3_0_buffer_funcs = { diff --git a/drivers/gpu/drm/amd/amdgpu/tonga_sdma_pkt_open.h b/drivers/gpu/drm/amd/amdgpu/tonga_sdma_pkt_open.h index 099b7b56113c..e5ebd084288d 100644 --- a/drivers/gpu/drm/amd/amdgpu/tonga_sdma_pkt_open.h +++ b/drivers/gpu/drm/amd/amdgpu/tonga_sdma_pkt_open.h @@ -2236,5 +2236,10 @@ #define SDMA_PKT_NOP_HEADER_sub_op_shift 8 #define SDMA_PKT_NOP_HEADER_SUB_OP(x) (((x) & SDMA_PKT_NOP_HEADER_sub_op_mask) << SDMA_PKT_NOP_HEADER_sub_op_shift) +/*define for count field*/ +#define SDMA_PKT_NOP_HEADER_count_offset 0 +#define SDMA_PKT_NOP_HEADER_count_mask 0x00003FFF +#define SDMA_PKT_NOP_HEADER_count_shift 16 +#define SDMA_PKT_NOP_HEADER_COUNT(x) (((x) & SDMA_PKT_NOP_HEADER_count_mask) << SDMA_PKT_NOP_HEADER_count_shift) #endif /* __TONGA_SDMA_PKT_OPEN_H_ */ diff --git a/drivers/gpu/drm/amd/amdgpu/tonga_smc.c b/drivers/gpu/drm/amd/amdgpu/tonga_smc.c index 5fc53a40c7ac..5421309c1862 100644 --- a/drivers/gpu/drm/amd/amdgpu/tonga_smc.c +++ b/drivers/gpu/drm/amd/amdgpu/tonga_smc.c @@ -761,7 +761,9 @@ int tonga_smu_init(struct amdgpu_device *adev) /* Allocate FW image data structure and header buffer */ ret = amdgpu_bo_create(adev, image_size, PAGE_SIZE, - true, AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, toc_buf); + true, AMDGPU_GEM_DOMAIN_VRAM, + AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED, + NULL, NULL, toc_buf); if (ret) { DRM_ERROR("Failed to allocate memory for TOC buffer\n"); return -ENOMEM; @@ -769,7 +771,9 @@ int tonga_smu_init(struct amdgpu_device *adev) /* Allocate buffer for SMU internal buffer */ ret = amdgpu_bo_create(adev, smu_internal_buffer_size, PAGE_SIZE, - true, AMDGPU_GEM_DOMAIN_VRAM, 0, NULL, smu_buf); + true, AMDGPU_GEM_DOMAIN_VRAM, + AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED, + NULL, NULL, smu_buf); if (ret) { DRM_ERROR("Failed to allocate memory for SMU internal buffer\n"); return -ENOMEM; diff --git a/drivers/gpu/drm/amd/amdgpu/uvd_v4_2.c b/drivers/gpu/drm/amd/amdgpu/uvd_v4_2.c index 9ac383bc6c1f..ed50dd725788 100644 --- a/drivers/gpu/drm/amd/amdgpu/uvd_v4_2.c +++ b/drivers/gpu/drm/amd/amdgpu/uvd_v4_2.c @@ -224,11 +224,11 @@ static int uvd_v4_2_suspend(void *handle) int r; struct amdgpu_device *adev = (struct amdgpu_device *)handle; - r = uvd_v4_2_hw_fini(adev); + r = amdgpu_uvd_suspend(adev); if (r) return r; - r = amdgpu_uvd_suspend(adev); + r = uvd_v4_2_hw_fini(adev); if (r) return r; @@ -886,6 +886,7 @@ static const struct amdgpu_ring_funcs uvd_v4_2_ring_funcs = { .test_ring = uvd_v4_2_ring_test_ring, .test_ib = uvd_v4_2_ring_test_ib, .is_lockup = amdgpu_ring_test_lockup, + .insert_nop = amdgpu_ring_insert_nop, }; static void uvd_v4_2_set_ring_funcs(struct amdgpu_device *adev) diff --git a/drivers/gpu/drm/amd/amdgpu/uvd_v5_0.c b/drivers/gpu/drm/amd/amdgpu/uvd_v5_0.c index de4b3f57902d..9ad8b9906c0b 100644 --- a/drivers/gpu/drm/amd/amdgpu/uvd_v5_0.c +++ b/drivers/gpu/drm/amd/amdgpu/uvd_v5_0.c @@ -220,11 +220,11 @@ static int uvd_v5_0_suspend(void *handle) int r; struct amdgpu_device *adev = (struct amdgpu_device *)handle; - r = uvd_v5_0_hw_fini(adev); + r = amdgpu_uvd_suspend(adev); if (r) return r; - r = amdgpu_uvd_suspend(adev); + r = uvd_v5_0_hw_fini(adev); if (r) return r; @@ -825,6 +825,7 @@ static const struct amdgpu_ring_funcs uvd_v5_0_ring_funcs = { .test_ring = uvd_v5_0_ring_test_ring, .test_ib = uvd_v5_0_ring_test_ib, .is_lockup = amdgpu_ring_test_lockup, + .insert_nop = amdgpu_ring_insert_nop, }; static void uvd_v5_0_set_ring_funcs(struct amdgpu_device *adev) diff --git a/drivers/gpu/drm/amd/amdgpu/uvd_v6_0.c b/drivers/gpu/drm/amd/amdgpu/uvd_v6_0.c index 66c975870e97..7e9934fa4193 100644 --- a/drivers/gpu/drm/amd/amdgpu/uvd_v6_0.c +++ b/drivers/gpu/drm/amd/amdgpu/uvd_v6_0.c @@ -214,14 +214,16 @@ static int uvd_v6_0_suspend(void *handle) int r; struct amdgpu_device *adev = (struct amdgpu_device *)handle; + /* Skip this for APU for now */ + if (!(adev->flags & AMD_IS_APU)) { + r = amdgpu_uvd_suspend(adev); + if (r) + return r; + } r = uvd_v6_0_hw_fini(adev); if (r) return r; - r = amdgpu_uvd_suspend(adev); - if (r) - return r; - return r; } @@ -230,10 +232,12 @@ static int uvd_v6_0_resume(void *handle) int r; struct amdgpu_device *adev = (struct amdgpu_device *)handle; - r = amdgpu_uvd_resume(adev); - if (r) - return r; - + /* Skip this for APU for now */ + if (!(adev->flags & AMD_IS_APU)) { + r = amdgpu_uvd_resume(adev); + if (r) + return r; + } r = uvd_v6_0_hw_init(adev); if (r) return r; @@ -805,6 +809,7 @@ static const struct amdgpu_ring_funcs uvd_v6_0_ring_funcs = { .test_ring = uvd_v6_0_ring_test_ring, .test_ib = uvd_v6_0_ring_test_ib, .is_lockup = amdgpu_ring_test_lockup, + .insert_nop = amdgpu_ring_insert_nop, }; static void uvd_v6_0_set_ring_funcs(struct amdgpu_device *adev) diff --git a/drivers/gpu/drm/amd/amdgpu/vce_v2_0.c b/drivers/gpu/drm/amd/amdgpu/vce_v2_0.c index 303d961d57bd..cd16df543f64 100644 --- a/drivers/gpu/drm/amd/amdgpu/vce_v2_0.c +++ b/drivers/gpu/drm/amd/amdgpu/vce_v2_0.c @@ -643,6 +643,7 @@ static const struct amdgpu_ring_funcs vce_v2_0_ring_funcs = { .test_ring = amdgpu_vce_ring_test_ring, .test_ib = amdgpu_vce_ring_test_ib, .is_lockup = amdgpu_ring_test_lockup, + .insert_nop = amdgpu_ring_insert_nop, }; static void vce_v2_0_set_ring_funcs(struct amdgpu_device *adev) diff --git a/drivers/gpu/drm/amd/amdgpu/vce_v3_0.c b/drivers/gpu/drm/amd/amdgpu/vce_v3_0.c index 4349658081ff..f0656dfb53f3 100644 --- a/drivers/gpu/drm/amd/amdgpu/vce_v3_0.c +++ b/drivers/gpu/drm/amd/amdgpu/vce_v3_0.c @@ -32,8 +32,8 @@ #include "vid.h" #include "vce/vce_3_0_d.h" #include "vce/vce_3_0_sh_mask.h" -#include "oss/oss_2_0_d.h" -#include "oss/oss_2_0_sh_mask.h" +#include "oss/oss_3_0_d.h" +#include "oss/oss_3_0_sh_mask.h" #include "gca/gfx_8_0_d.h" #include "smu/smu_7_1_2_d.h" #include "smu/smu_7_1_2_sh_mask.h" @@ -426,17 +426,41 @@ static void vce_v3_0_mc_resume(struct amdgpu_device *adev, int idx) static bool vce_v3_0_is_idle(void *handle) { struct amdgpu_device *adev = (struct amdgpu_device *)handle; + u32 mask = 0; + int idx; - return !(RREG32(mmSRBM_STATUS2) & SRBM_STATUS2__VCE_BUSY_MASK); + for (idx = 0; idx < 2; ++idx) { + if (adev->vce.harvest_config & (1 << idx)) + continue; + + if (idx == 0) + mask |= SRBM_STATUS2__VCE0_BUSY_MASK; + else + mask |= SRBM_STATUS2__VCE1_BUSY_MASK; + } + + return !(RREG32(mmSRBM_STATUS2) & mask); } static int vce_v3_0_wait_for_idle(void *handle) { unsigned i; struct amdgpu_device *adev = (struct amdgpu_device *)handle; + u32 mask = 0; + int idx; + + for (idx = 0; idx < 2; ++idx) { + if (adev->vce.harvest_config & (1 << idx)) + continue; + + if (idx == 0) + mask |= SRBM_STATUS2__VCE0_BUSY_MASK; + else + mask |= SRBM_STATUS2__VCE1_BUSY_MASK; + } for (i = 0; i < adev->usec_timeout; i++) { - if (!(RREG32(mmSRBM_STATUS2) & SRBM_STATUS2__VCE_BUSY_MASK)) + if (!(RREG32(mmSRBM_STATUS2) & mask)) return 0; } return -ETIMEDOUT; @@ -445,9 +469,21 @@ static int vce_v3_0_wait_for_idle(void *handle) static int vce_v3_0_soft_reset(void *handle) { struct amdgpu_device *adev = (struct amdgpu_device *)handle; + u32 mask = 0; + int idx; + + for (idx = 0; idx < 2; ++idx) { + if (adev->vce.harvest_config & (1 << idx)) + continue; - WREG32_P(mmSRBM_SOFT_RESET, SRBM_SOFT_RESET__SOFT_RESET_VCE_MASK, - ~SRBM_SOFT_RESET__SOFT_RESET_VCE_MASK); + if (idx == 0) + mask |= SRBM_SOFT_RESET__SOFT_RESET_VCE0_MASK; + else + mask |= SRBM_SOFT_RESET__SOFT_RESET_VCE1_MASK; + } + WREG32_P(mmSRBM_SOFT_RESET, mask, + ~(SRBM_SOFT_RESET__SOFT_RESET_VCE0_MASK | + SRBM_SOFT_RESET__SOFT_RESET_VCE1_MASK)); mdelay(5); return vce_v3_0_start(adev); @@ -608,6 +644,7 @@ static const struct amdgpu_ring_funcs vce_v3_0_ring_funcs = { .test_ring = amdgpu_vce_ring_test_ring, .test_ib = amdgpu_vce_ring_test_ib, .is_lockup = amdgpu_ring_test_lockup, + .insert_nop = amdgpu_ring_insert_nop, }; static void vce_v3_0_set_ring_funcs(struct amdgpu_device *adev) diff --git a/drivers/gpu/drm/amd/amdgpu/vi.c b/drivers/gpu/drm/amd/amdgpu/vi.c index 552d9e75ad1b..b55ceb14fdcd 100644 --- a/drivers/gpu/drm/amd/amdgpu/vi.c +++ b/drivers/gpu/drm/amd/amdgpu/vi.c @@ -1400,7 +1400,8 @@ static int vi_common_early_init(void *handle) case CHIP_CARRIZO: adev->has_uvd = true; adev->cg_flags = 0; - adev->pg_flags = AMDGPU_PG_SUPPORT_UVD | AMDGPU_PG_SUPPORT_VCE; + /* Disable UVD pg */ + adev->pg_flags = /* AMDGPU_PG_SUPPORT_UVD | */AMDGPU_PG_SUPPORT_VCE; adev->external_rev_id = adev->rev_id + 0x1; if (amdgpu_smc_load_fw && smc_enabled) adev->firmware.smu_load = true; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index c991973019d0..c6a1b4cc6458 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -31,7 +31,7 @@ #include <uapi/linux/kfd_ioctl.h> #include <linux/time.h> #include <linux/mm.h> -#include <uapi/asm-generic/mman-common.h> +#include <linux/mman.h> #include <asm/processor.h> #include "kfd_priv.h" #include "kfd_device_queue_manager.h" diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c b/drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c index 35b987574633..2b655103ba79 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c @@ -33,7 +33,7 @@ #include <linux/time.h> #include "kfd_priv.h" #include <linux/mm.h> -#include <uapi/asm-generic/mman-common.h> +#include <linux/mman.h> #include <asm/processor.h> /* diff --git a/drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h b/drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h new file mode 100644 index 000000000000..144f50acc971 --- /dev/null +++ b/drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h @@ -0,0 +1,41 @@ +#if !defined(_GPU_SCHED_TRACE_H) || defined(TRACE_HEADER_MULTI_READ) +#define _GPU_SCHED_TRACE_H_ + +#include <linux/stringify.h> +#include <linux/types.h> +#include <linux/tracepoint.h> + +#include <drm/drmP.h> + +#undef TRACE_SYSTEM +#define TRACE_SYSTEM gpu_sched +#define TRACE_INCLUDE_FILE gpu_sched_trace + +TRACE_EVENT(amd_sched_job, + TP_PROTO(struct amd_sched_job *sched_job), + TP_ARGS(sched_job), + TP_STRUCT__entry( + __field(struct amd_sched_entity *, entity) + __field(const char *, name) + __field(u32, job_count) + __field(int, hw_job_count) + ), + + TP_fast_assign( + __entry->entity = sched_job->s_entity; + __entry->name = sched_job->sched->name; + __entry->job_count = kfifo_len( + &sched_job->s_entity->job_queue) / sizeof(sched_job); + __entry->hw_job_count = atomic_read( + &sched_job->sched->hw_rq_count); + ), + TP_printk("entity=%p, ring=%s, job count:%u, hw job count:%d", + __entry->entity, __entry->name, __entry->job_count, + __entry->hw_job_count) +); +#endif + +/* This part must be outside protection */ +#undef TRACE_INCLUDE_PATH +#define TRACE_INCLUDE_PATH . +#include <trace/define_trace.h> diff --git a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c index 265d3e2f63cc..3697eeeecf82 100644 --- a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c +++ b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c @@ -27,132 +27,79 @@ #include <drm/drmP.h> #include "gpu_scheduler.h" +#define CREATE_TRACE_POINTS +#include "gpu_sched_trace.h" + +static struct amd_sched_job * +amd_sched_entity_pop_job(struct amd_sched_entity *entity); +static void amd_sched_wakeup(struct amd_gpu_scheduler *sched); + /* Initialize a given run queue struct */ static void amd_sched_rq_init(struct amd_sched_rq *rq) { + spin_lock_init(&rq->lock); INIT_LIST_HEAD(&rq->entities); - mutex_init(&rq->lock); rq->current_entity = NULL; } static void amd_sched_rq_add_entity(struct amd_sched_rq *rq, struct amd_sched_entity *entity) { - mutex_lock(&rq->lock); + spin_lock(&rq->lock); list_add_tail(&entity->list, &rq->entities); - mutex_unlock(&rq->lock); + spin_unlock(&rq->lock); } static void amd_sched_rq_remove_entity(struct amd_sched_rq *rq, struct amd_sched_entity *entity) { - mutex_lock(&rq->lock); + spin_lock(&rq->lock); list_del_init(&entity->list); if (rq->current_entity == entity) rq->current_entity = NULL; - mutex_unlock(&rq->lock); + spin_unlock(&rq->lock); } /** - * Select next entity from a specified run queue with round robin policy. - * It could return the same entity as current one if current is the only - * available one in the queue. Return NULL if nothing available. + * Select next job from a specified run queue with round robin policy. + * Return NULL if nothing available. */ -static struct amd_sched_entity * -amd_sched_rq_select_entity(struct amd_sched_rq *rq) +static struct amd_sched_job * +amd_sched_rq_select_job(struct amd_sched_rq *rq) { - struct amd_sched_entity *entity = rq->current_entity; + struct amd_sched_entity *entity; + struct amd_sched_job *sched_job; + spin_lock(&rq->lock); + + entity = rq->current_entity; if (entity) { list_for_each_entry_continue(entity, &rq->entities, list) { - if (!kfifo_is_empty(&entity->job_queue)) { + sched_job = amd_sched_entity_pop_job(entity); + if (sched_job) { rq->current_entity = entity; - return rq->current_entity; + spin_unlock(&rq->lock); + return sched_job; } } } list_for_each_entry(entity, &rq->entities, list) { - if (!kfifo_is_empty(&entity->job_queue)) { + sched_job = amd_sched_entity_pop_job(entity); + if (sched_job) { rq->current_entity = entity; - return rq->current_entity; + spin_unlock(&rq->lock); + return sched_job; } if (entity == rq->current_entity) break; } - return NULL; -} - -/** - * Note: This function should only been called inside scheduler main - * function for thread safety, there is no other protection here. - * return ture if scheduler has something ready to run. - * - * For active_hw_rq, there is only one producer(scheduler thread) and - * one consumer(ISR). It should be safe to use this function in scheduler - * main thread to decide whether to continue emit more IBs. -*/ -static bool is_scheduler_ready(struct amd_gpu_scheduler *sched) -{ - unsigned long flags; - bool full; - - spin_lock_irqsave(&sched->queue_lock, flags); - full = atomic64_read(&sched->hw_rq_count) < - sched->hw_submission_limit ? true : false; - spin_unlock_irqrestore(&sched->queue_lock, flags); - - return full; -} - -/** - * Select next entity from the kernel run queue, if not available, - * return null. -*/ -static struct amd_sched_entity * -kernel_rq_select_context(struct amd_gpu_scheduler *sched) -{ - struct amd_sched_entity *sched_entity; - struct amd_sched_rq *rq = &sched->kernel_rq; - - mutex_lock(&rq->lock); - sched_entity = amd_sched_rq_select_entity(rq); - mutex_unlock(&rq->lock); - return sched_entity; -} - -/** - * Select next entity containing real IB submissions -*/ -static struct amd_sched_entity * -select_context(struct amd_gpu_scheduler *sched) -{ - struct amd_sched_entity *wake_entity = NULL; - struct amd_sched_entity *tmp; - struct amd_sched_rq *rq; + spin_unlock(&rq->lock); - if (!is_scheduler_ready(sched)) - return NULL; - - /* Kernel run queue has higher priority than normal run queue*/ - tmp = kernel_rq_select_context(sched); - if (tmp != NULL) - goto exit; - - rq = &sched->sched_rq; - mutex_lock(&rq->lock); - tmp = amd_sched_rq_select_entity(rq); - mutex_unlock(&rq->lock); -exit: - if (sched->current_entity && (sched->current_entity != tmp)) - wake_entity = sched->current_entity; - sched->current_entity = tmp; - if (wake_entity && wake_entity->need_wakeup) - wake_up(&wake_entity->wait_queue); - return tmp; + return NULL; } /** @@ -171,34 +118,27 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched, struct amd_sched_rq *rq, uint32_t jobs) { - uint64_t seq_ring = 0; - char name[20]; + int r; if (!(sched && entity && rq)) return -EINVAL; memset(entity, 0, sizeof(struct amd_sched_entity)); - seq_ring = ((uint64_t)sched->ring_id) << 60; - spin_lock_init(&entity->lock); - entity->belongto_rq = rq; - entity->scheduler = sched; - init_waitqueue_head(&entity->wait_queue); - init_waitqueue_head(&entity->wait_emit); - entity->fence_context = fence_context_alloc(1); - snprintf(name, sizeof(name), "c_entity[%llu]", entity->fence_context); - memcpy(entity->name, name, 20); - entity->need_wakeup = false; - if(kfifo_alloc(&entity->job_queue, - jobs * sizeof(void *), - GFP_KERNEL)) - return -EINVAL; + INIT_LIST_HEAD(&entity->list); + entity->rq = rq; + entity->sched = sched; spin_lock_init(&entity->queue_lock); - atomic64_set(&entity->last_queued_v_seq, seq_ring); - atomic64_set(&entity->last_signaled_v_seq, seq_ring); + r = kfifo_alloc(&entity->job_queue, jobs * sizeof(void *), GFP_KERNEL); + if (r) + return r; + + atomic_set(&entity->fence_seq, 0); + entity->fence_context = fence_context_alloc(1); /* Add the entity to the run queue */ amd_sched_rq_add_entity(rq, entity); + return 0; } @@ -210,23 +150,24 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched, * * return true if entity is initialized, false otherwise */ -static bool is_context_entity_initialized(struct amd_gpu_scheduler *sched, - struct amd_sched_entity *entity) +static bool amd_sched_entity_is_initialized(struct amd_gpu_scheduler *sched, + struct amd_sched_entity *entity) { - return entity->scheduler == sched && - entity->belongto_rq != NULL; + return entity->sched == sched && + entity->rq != NULL; } -static bool is_context_entity_idle(struct amd_gpu_scheduler *sched, - struct amd_sched_entity *entity) +/** + * Check if entity is idle + * + * @entity The pointer to a valid scheduler entity + * + * Return true if entity don't has any unscheduled jobs. + */ +static bool amd_sched_entity_is_idle(struct amd_sched_entity *entity) { - /** - * Idle means no pending IBs, and the entity is not - * currently being used. - */ - barrier(); - if ((sched->current_entity != entity) && - kfifo_is_empty(&entity->job_queue)) + rmb(); + if (kfifo_is_empty(&entity->job_queue)) return true; return false; @@ -238,225 +179,247 @@ static bool is_context_entity_idle(struct amd_gpu_scheduler *sched, * @sched Pointer to scheduler instance * @entity The pointer to a valid scheduler entity * - * return 0 if succeed. negative error code on failure + * Cleanup and free the allocated resources. */ -int amd_sched_entity_fini(struct amd_gpu_scheduler *sched, - struct amd_sched_entity *entity) +void amd_sched_entity_fini(struct amd_gpu_scheduler *sched, + struct amd_sched_entity *entity) { - int r = 0; - struct amd_sched_rq *rq = entity->belongto_rq; + struct amd_sched_rq *rq = entity->rq; + + if (!amd_sched_entity_is_initialized(sched, entity)) + return; - if (!is_context_entity_initialized(sched, entity)) - return 0; - entity->need_wakeup = true; /** * The client will not queue more IBs during this fini, consume existing * queued IBs */ - r = wait_event_timeout( - entity->wait_queue, - is_context_entity_idle(sched, entity), - msecs_to_jiffies(AMD_GPU_WAIT_IDLE_TIMEOUT_IN_MS) - ) ? 0 : -1; - - if (r) { - if (entity->is_pending) - DRM_INFO("Entity %p is in waiting state during fini,\ - all pending ibs will be canceled.\n", - entity); - } + wait_event(sched->job_scheduled, amd_sched_entity_is_idle(entity)); amd_sched_rq_remove_entity(rq, entity); kfifo_free(&entity->job_queue); - return r; +} + +static void amd_sched_entity_wakeup(struct fence *f, struct fence_cb *cb) +{ + struct amd_sched_entity *entity = + container_of(cb, struct amd_sched_entity, cb); + entity->dependency = NULL; + fence_put(f); + amd_sched_wakeup(entity->sched); +} + +static struct amd_sched_job * +amd_sched_entity_pop_job(struct amd_sched_entity *entity) +{ + struct amd_gpu_scheduler *sched = entity->sched; + struct amd_sched_job *sched_job; + + if (ACCESS_ONCE(entity->dependency)) + return NULL; + + if (!kfifo_out_peek(&entity->job_queue, &sched_job, sizeof(sched_job))) + return NULL; + + while ((entity->dependency = sched->ops->dependency(sched_job))) { + + if (fence_add_callback(entity->dependency, &entity->cb, + amd_sched_entity_wakeup)) + fence_put(entity->dependency); + else + return NULL; + } + + return sched_job; } /** - * Submit a normal job to the job queue + * Helper to submit a job to the job queue * - * @sched The pointer to the scheduler - * @c_entity The pointer to amd_sched_entity - * @job The pointer to job required to submit - * return 0 if succeed. -1 if failed. - * -2 indicate queue is full for this client, client should wait untill - * scheduler consum some queued command. - * -1 other fail. -*/ -int amd_sched_push_job(struct amd_gpu_scheduler *sched, - struct amd_sched_entity *c_entity, - void *data, - struct amd_sched_fence **fence) + * @sched_job The pointer to job required to submit + * + * Returns true if we could submit the job. + */ +static bool amd_sched_entity_in(struct amd_sched_job *sched_job) { - struct amd_sched_job *job; + struct amd_sched_entity *entity = sched_job->s_entity; + bool added, first = false; + + spin_lock(&entity->queue_lock); + added = kfifo_in(&entity->job_queue, &sched_job, + sizeof(sched_job)) == sizeof(sched_job); + + if (added && kfifo_len(&entity->job_queue) == sizeof(sched_job)) + first = true; + + spin_unlock(&entity->queue_lock); + + /* first job wakes up scheduler */ + if (first) + amd_sched_wakeup(sched_job->sched); + + return added; +} + +/** + * Submit a job to the job queue + * + * @sched_job The pointer to job required to submit + * + * Returns 0 for success, negative error code otherwise. + */ +int amd_sched_entity_push_job(struct amd_sched_job *sched_job) +{ + struct amd_sched_entity *entity = sched_job->s_entity; + struct amd_sched_fence *fence = amd_sched_fence_create( + entity, sched_job->owner); if (!fence) - return -EINVAL; - job = kzalloc(sizeof(struct amd_sched_job), GFP_KERNEL); - if (!job) return -ENOMEM; - job->sched = sched; - job->s_entity = c_entity; - job->data = data; - *fence = amd_sched_fence_create(c_entity); - if ((*fence) == NULL) { - kfree(job); - return -EINVAL; - } - fence_get(&(*fence)->base); - job->s_fence = *fence; - while (kfifo_in_spinlocked(&c_entity->job_queue, &job, sizeof(void *), - &c_entity->queue_lock) != sizeof(void *)) { - /** - * Current context used up all its IB slots - * wait here, or need to check whether GPU is hung - */ - schedule(); - } - /* first job wake up scheduler */ - if ((kfifo_len(&c_entity->job_queue) / sizeof(void *)) == 1) - wake_up_interruptible(&sched->wait_queue); + + fence_get(&fence->base); + sched_job->s_fence = fence; + + wait_event(entity->sched->job_scheduled, + amd_sched_entity_in(sched_job)); + trace_amd_sched_job(sched_job); return 0; } +/** + * Return ture if we can push more jobs to the hw. + */ +static bool amd_sched_ready(struct amd_gpu_scheduler *sched) +{ + return atomic_read(&sched->hw_rq_count) < + sched->hw_submission_limit; +} + +/** + * Wake up the scheduler when it is ready + */ +static void amd_sched_wakeup(struct amd_gpu_scheduler *sched) +{ + if (amd_sched_ready(sched)) + wake_up_interruptible(&sched->wake_up_worker); +} + +/** + * Select next to run +*/ +static struct amd_sched_job * +amd_sched_select_job(struct amd_gpu_scheduler *sched) +{ + struct amd_sched_job *sched_job; + + if (!amd_sched_ready(sched)) + return NULL; + + /* Kernel run queue has higher priority than normal run queue*/ + sched_job = amd_sched_rq_select_job(&sched->kernel_rq); + if (sched_job == NULL) + sched_job = amd_sched_rq_select_job(&sched->sched_rq); + + return sched_job; +} + static void amd_sched_process_job(struct fence *f, struct fence_cb *cb) { - struct amd_sched_job *sched_job = - container_of(cb, struct amd_sched_job, cb); - struct amd_gpu_scheduler *sched; - unsigned long flags; - - sched = sched_job->sched; - atomic64_set(&sched_job->s_entity->last_signaled_v_seq, - sched_job->s_fence->v_seq); - amd_sched_fence_signal(sched_job->s_fence); - spin_lock_irqsave(&sched->queue_lock, flags); - list_del(&sched_job->list); - atomic64_dec(&sched->hw_rq_count); - spin_unlock_irqrestore(&sched->queue_lock, flags); - - sched->ops->process_job(sched, sched_job); - fence_put(&sched_job->s_fence->base); - kfree(sched_job); - wake_up_interruptible(&sched->wait_queue); + struct amd_sched_fence *s_fence = + container_of(cb, struct amd_sched_fence, cb); + struct amd_gpu_scheduler *sched = s_fence->sched; + + atomic_dec(&sched->hw_rq_count); + amd_sched_fence_signal(s_fence); + fence_put(&s_fence->base); + wake_up_interruptible(&sched->wake_up_worker); } static int amd_sched_main(void *param) { - int r; - struct amd_sched_job *job; struct sched_param sparam = {.sched_priority = 1}; - struct amd_sched_entity *c_entity = NULL; struct amd_gpu_scheduler *sched = (struct amd_gpu_scheduler *)param; + int r, count; sched_setscheduler(current, SCHED_FIFO, &sparam); while (!kthread_should_stop()) { + struct amd_sched_entity *entity; + struct amd_sched_fence *s_fence; + struct amd_sched_job *sched_job; struct fence *fence; - wait_event_interruptible(sched->wait_queue, - is_scheduler_ready(sched) && - (c_entity = select_context(sched))); - r = kfifo_out(&c_entity->job_queue, &job, sizeof(void *)); - if (r != sizeof(void *)) + wait_event_interruptible(sched->wake_up_worker, + kthread_should_stop() || + (sched_job = amd_sched_select_job(sched))); + + if (!sched_job) continue; - r = sched->ops->prepare_job(sched, c_entity, job); - if (!r) { - unsigned long flags; - spin_lock_irqsave(&sched->queue_lock, flags); - list_add_tail(&job->list, &sched->active_hw_rq); - atomic64_inc(&sched->hw_rq_count); - spin_unlock_irqrestore(&sched->queue_lock, flags); - } - mutex_lock(&sched->sched_lock); - fence = sched->ops->run_job(sched, c_entity, job); + + entity = sched_job->s_entity; + s_fence = sched_job->s_fence; + atomic_inc(&sched->hw_rq_count); + fence = sched->ops->run_job(sched_job); if (fence) { - r = fence_add_callback(fence, &job->cb, + r = fence_add_callback(fence, &s_fence->cb, amd_sched_process_job); if (r == -ENOENT) - amd_sched_process_job(fence, &job->cb); + amd_sched_process_job(fence, &s_fence->cb); else if (r) DRM_ERROR("fence add callback failed (%d)\n", r); fence_put(fence); + } else { + DRM_ERROR("Failed to run job!\n"); + amd_sched_process_job(NULL, &s_fence->cb); } - mutex_unlock(&sched->sched_lock); + + count = kfifo_out(&entity->job_queue, &sched_job, + sizeof(sched_job)); + WARN_ON(count != sizeof(sched_job)); + wake_up(&sched->job_scheduled); } return 0; } /** - * Create a gpu scheduler + * Init a gpu scheduler instance * - * @device The device context for this scheduler - * @ops The backend operations for this scheduler. - * @id The scheduler is per ring, here is ring id. - * @granularity The minumum ms unit the scheduler will scheduled. - * @preemption Indicate whether this ring support preemption, 0 is no. + * @sched The pointer to the scheduler + * @ops The backend operations for this scheduler. + * @hw_submissions Number of hw submissions to do. + * @name Name used for debugging * - * return the pointer to scheduler for success, otherwise return NULL + * Return 0 on success, otherwise error code. */ -struct amd_gpu_scheduler *amd_sched_create(void *device, - struct amd_sched_backend_ops *ops, - unsigned ring, - unsigned granularity, - unsigned preemption, - unsigned hw_submission) +int amd_sched_init(struct amd_gpu_scheduler *sched, + struct amd_sched_backend_ops *ops, + unsigned hw_submission, const char *name) { - struct amd_gpu_scheduler *sched; - char name[20]; - - sched = kzalloc(sizeof(struct amd_gpu_scheduler), GFP_KERNEL); - if (!sched) - return NULL; - - sched->device = device; sched->ops = ops; - sched->granularity = granularity; - sched->ring_id = ring; - sched->preemption = preemption; sched->hw_submission_limit = hw_submission; - snprintf(name, sizeof(name), "gpu_sched[%d]", ring); - mutex_init(&sched->sched_lock); - spin_lock_init(&sched->queue_lock); + sched->name = name; amd_sched_rq_init(&sched->sched_rq); amd_sched_rq_init(&sched->kernel_rq); - init_waitqueue_head(&sched->wait_queue); - INIT_LIST_HEAD(&sched->active_hw_rq); - atomic64_set(&sched->hw_rq_count, 0); + init_waitqueue_head(&sched->wake_up_worker); + init_waitqueue_head(&sched->job_scheduled); + atomic_set(&sched->hw_rq_count, 0); + /* Each scheduler will run on a seperate kernel thread */ - sched->thread = kthread_create(amd_sched_main, sched, name); - if (sched->thread) { - wake_up_process(sched->thread); - return sched; + sched->thread = kthread_run(amd_sched_main, sched, sched->name); + if (IS_ERR(sched->thread)) { + DRM_ERROR("Failed to create scheduler for %s.\n", name); + return PTR_ERR(sched->thread); } - DRM_ERROR("Failed to create scheduler for id %d.\n", ring); - kfree(sched); - return NULL; + return 0; } /** * Destroy a gpu scheduler * * @sched The pointer to the scheduler - * - * return 0 if succeed. -1 if failed. */ -int amd_sched_destroy(struct amd_gpu_scheduler *sched) +void amd_sched_fini(struct amd_gpu_scheduler *sched) { kthread_stop(sched->thread); - kfree(sched); - return 0; -} - -/** - * Get next queued sequence number - * - * @entity The context entity - * - * return the next queued sequence number -*/ -uint64_t amd_sched_next_queued_seq(struct amd_sched_entity *c_entity) -{ - return atomic64_read(&c_entity->last_queued_v_seq) + 1; } diff --git a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h index ceb5918bfbeb..80b64dc22214 100644 --- a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h +++ b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h @@ -27,8 +27,6 @@ #include <linux/kfifo.h> #include <linux/fence.h> -#define AMD_GPU_WAIT_IDLE_TIMEOUT_IN_MS 3000 - struct amd_gpu_scheduler; struct amd_sched_rq; @@ -40,21 +38,17 @@ struct amd_sched_rq; */ struct amd_sched_entity { struct list_head list; - struct amd_sched_rq *belongto_rq; - spinlock_t lock; - /* the virtual_seq is unique per context per ring */ - atomic64_t last_queued_v_seq; - atomic64_t last_signaled_v_seq; - /* the job_queue maintains the jobs submitted by clients */ - struct kfifo job_queue; + struct amd_sched_rq *rq; + struct amd_gpu_scheduler *sched; + spinlock_t queue_lock; - struct amd_gpu_scheduler *scheduler; - wait_queue_head_t wait_queue; - wait_queue_head_t wait_emit; - bool is_pending; + struct kfifo job_queue; + + atomic_t fence_seq; uint64_t fence_context; - char name[20]; - bool need_wakeup; + + struct fence *dependency; + struct fence_cb cb; }; /** @@ -63,7 +57,7 @@ struct amd_sched_entity { * the next entity to emit commands from. */ struct amd_sched_rq { - struct mutex lock; + spinlock_t lock; struct list_head entities; struct amd_sched_entity *current_entity; }; @@ -71,18 +65,16 @@ struct amd_sched_rq { struct amd_sched_fence { struct fence base; struct fence_cb cb; - struct amd_sched_entity *entity; - uint64_t v_seq; + struct amd_gpu_scheduler *sched; spinlock_t lock; + void *owner; }; struct amd_sched_job { - struct list_head list; - struct fence_cb cb; struct amd_gpu_scheduler *sched; struct amd_sched_entity *s_entity; - void *data; struct amd_sched_fence *s_fence; + void *owner; }; extern const struct fence_ops amd_sched_fence_ops; @@ -101,61 +93,40 @@ static inline struct amd_sched_fence *to_amd_sched_fence(struct fence *f) * these functions should be implemented in driver side */ struct amd_sched_backend_ops { - int (*prepare_job)(struct amd_gpu_scheduler *sched, - struct amd_sched_entity *c_entity, - struct amd_sched_job *job); - struct fence *(*run_job)(struct amd_gpu_scheduler *sched, - struct amd_sched_entity *c_entity, - struct amd_sched_job *job); - void (*process_job)(struct amd_gpu_scheduler *sched, - struct amd_sched_job *job); + struct fence *(*dependency)(struct amd_sched_job *sched_job); + struct fence *(*run_job)(struct amd_sched_job *sched_job); }; /** * One scheduler is implemented for each hardware ring */ struct amd_gpu_scheduler { - void *device; - struct task_struct *thread; + struct amd_sched_backend_ops *ops; + uint32_t hw_submission_limit; + const char *name; struct amd_sched_rq sched_rq; struct amd_sched_rq kernel_rq; - struct list_head active_hw_rq; - atomic64_t hw_rq_count; - struct amd_sched_backend_ops *ops; - uint32_t ring_id; - uint32_t granularity; /* in ms unit */ - uint32_t preemption; - wait_queue_head_t wait_queue; - struct amd_sched_entity *current_entity; - struct mutex sched_lock; - spinlock_t queue_lock; - uint32_t hw_submission_limit; + wait_queue_head_t wake_up_worker; + wait_queue_head_t job_scheduled; + atomic_t hw_rq_count; + struct task_struct *thread; }; -struct amd_gpu_scheduler *amd_sched_create(void *device, - struct amd_sched_backend_ops *ops, - uint32_t ring, - uint32_t granularity, - uint32_t preemption, - uint32_t hw_submission); -int amd_sched_destroy(struct amd_gpu_scheduler *sched); - -int amd_sched_push_job(struct amd_gpu_scheduler *sched, - struct amd_sched_entity *c_entity, - void *data, - struct amd_sched_fence **fence); +int amd_sched_init(struct amd_gpu_scheduler *sched, + struct amd_sched_backend_ops *ops, + uint32_t hw_submission, const char *name); +void amd_sched_fini(struct amd_gpu_scheduler *sched); int amd_sched_entity_init(struct amd_gpu_scheduler *sched, struct amd_sched_entity *entity, struct amd_sched_rq *rq, uint32_t jobs); -int amd_sched_entity_fini(struct amd_gpu_scheduler *sched, - struct amd_sched_entity *entity); - -uint64_t amd_sched_next_queued_seq(struct amd_sched_entity *c_entity); +void amd_sched_entity_fini(struct amd_gpu_scheduler *sched, + struct amd_sched_entity *entity); +int amd_sched_entity_push_job(struct amd_sched_job *sched_job); struct amd_sched_fence *amd_sched_fence_create( - struct amd_sched_entity *s_entity); + struct amd_sched_entity *s_entity, void *owner); void amd_sched_fence_signal(struct amd_sched_fence *fence); diff --git a/drivers/gpu/drm/amd/scheduler/sched_fence.c b/drivers/gpu/drm/amd/scheduler/sched_fence.c index a4751598c0b4..d802638094f4 100644 --- a/drivers/gpu/drm/amd/scheduler/sched_fence.c +++ b/drivers/gpu/drm/amd/scheduler/sched_fence.c @@ -27,19 +27,22 @@ #include <drm/drmP.h> #include "gpu_scheduler.h" -struct amd_sched_fence *amd_sched_fence_create(struct amd_sched_entity *s_entity) +struct amd_sched_fence *amd_sched_fence_create(struct amd_sched_entity *s_entity, void *owner) { struct amd_sched_fence *fence = NULL; + unsigned seq; + fence = kzalloc(sizeof(struct amd_sched_fence), GFP_KERNEL); if (fence == NULL) return NULL; - fence->v_seq = atomic64_inc_return(&s_entity->last_queued_v_seq); - fence->entity = s_entity; + fence->owner = owner; + fence->sched = s_entity->sched; spin_lock_init(&fence->lock); - fence_init(&fence->base, &amd_sched_fence_ops, - &fence->lock, - s_entity->fence_context, - fence->v_seq); + + seq = atomic_inc_return(&s_entity->fence_seq); + fence_init(&fence->base, &amd_sched_fence_ops, &fence->lock, + s_entity->fence_context, seq); + return fence; } @@ -60,7 +63,7 @@ static const char *amd_sched_fence_get_driver_name(struct fence *fence) static const char *amd_sched_fence_get_timeline_name(struct fence *f) { struct amd_sched_fence *fence = to_amd_sched_fence(f); - return (const char *)fence->entity->name; + return (const char *)fence->sched->name; } static bool amd_sched_fence_enable_signaling(struct fence *f) |
