From 65f8682b9aaae20c2cdee993e6fe52374ad513c9 Mon Sep 17 00:00:00 2001 From: Danijel Slivka Date: Tue, 4 Oct 2022 15:39:44 +0200 Subject: drm/amdgpu: set vm_update_mode=0 as default for Sienna Cichlid in SRIOV case For asic with VF MMIO access protection avoid using CPU for VM table updates. CPU pagetable updates have issues with HDP flush as VF MMIO access protection blocks write to mmBIF_BX_DEV0_EPF0_VF0_HDP_MEM_COHERENCY_FLUSH_CNTL register during sriov runtime. v3: introduce virtualization capability flag AMDGPU_VF_MMIO_ACCESS_PROTECT which indicates that VF MMIO write access is not allowed in sriov runtime Signed-off-by: Danijel Slivka Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c index 83b0c5d86e48..2291aa14d888 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c @@ -2338,7 +2338,11 @@ void amdgpu_vm_manager_init(struct amdgpu_device *adev) */ #ifdef CONFIG_X86_64 if (amdgpu_vm_update_mode == -1) { - if (amdgpu_gmc_vram_full_visible(&adev->gmc)) + /* For asic with VF MMIO access protection + * avoid using CPU for VM table updates + */ + if (amdgpu_gmc_vram_full_visible(&adev->gmc) && + !amdgpu_sriov_vf_mmio_access_protection(adev)) adev->vm_manager.vm_update_mode = AMDGPU_VM_USE_CPU_FOR_COMPUTE; else -- cgit From 0bc71adc8b1b65a369d4b8ef098e74ed4aab62d9 Mon Sep 17 00:00:00 2001 From: Philip Yang Date: Wed, 2 Nov 2022 16:55:31 -0400 Subject: drm/amdgpu: Drop eviction lock when allocating PT BO MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Re-take the eviction lock immediately again after the allocation is completed, to fix circular locking warning with drm_buddy allocator. Move amdgpu_vm_eviction_lock/unlock/trylock to amdgpu_vm.h as they are called from multiple files. Signed-off-by: Philip Yang Reviewed-by: Christian König Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 26 -------------------------- drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h | 26 ++++++++++++++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c | 2 ++ 3 files changed, 28 insertions(+), 26 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c index 2291aa14d888..003aa9e47085 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c @@ -143,32 +143,6 @@ int amdgpu_vm_set_pasid(struct amdgpu_device *adev, struct amdgpu_vm *vm, return 0; } -/* - * vm eviction_lock can be taken in MMU notifiers. Make sure no reclaim-FS - * happens while holding this lock anywhere to prevent deadlocks when - * an MMU notifier runs in reclaim-FS context. - */ -static inline void amdgpu_vm_eviction_lock(struct amdgpu_vm *vm) -{ - mutex_lock(&vm->eviction_lock); - vm->saved_flags = memalloc_noreclaim_save(); -} - -static inline int amdgpu_vm_eviction_trylock(struct amdgpu_vm *vm) -{ - if (mutex_trylock(&vm->eviction_lock)) { - vm->saved_flags = memalloc_noreclaim_save(); - return 1; - } - return 0; -} - -static inline void amdgpu_vm_eviction_unlock(struct amdgpu_vm *vm) -{ - memalloc_noreclaim_restore(vm->saved_flags); - mutex_unlock(&vm->eviction_lock); -} - /** * amdgpu_vm_bo_evicted - vm_bo is evicted * diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h index 1d31771b4230..6546e786bf00 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h @@ -510,4 +510,30 @@ static inline uint64_t amdgpu_vm_tlb_seq(struct amdgpu_vm *vm) return atomic64_read(&vm->tlb_seq); } +/* + * vm eviction_lock can be taken in MMU notifiers. Make sure no reclaim-FS + * happens while holding this lock anywhere to prevent deadlocks when + * an MMU notifier runs in reclaim-FS context. + */ +static inline void amdgpu_vm_eviction_lock(struct amdgpu_vm *vm) +{ + mutex_lock(&vm->eviction_lock); + vm->saved_flags = memalloc_noreclaim_save(); +} + +static inline bool amdgpu_vm_eviction_trylock(struct amdgpu_vm *vm) +{ + if (mutex_trylock(&vm->eviction_lock)) { + vm->saved_flags = memalloc_noreclaim_save(); + return true; + } + return false; +} + +static inline void amdgpu_vm_eviction_unlock(struct amdgpu_vm *vm) +{ + memalloc_noreclaim_restore(vm->saved_flags); + mutex_unlock(&vm->eviction_lock); +} + #endif diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c index 358b91243e37..b5f3bba851db 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c @@ -597,7 +597,9 @@ static int amdgpu_vm_pt_alloc(struct amdgpu_device *adev, if (entry->bo) return 0; + amdgpu_vm_eviction_unlock(vm); r = amdgpu_vm_pt_create(adev, vm, cursor->level, immediate, &pt); + amdgpu_vm_eviction_lock(vm); if (r) return r; -- cgit From 3f4c175d62d89819121cbbd5a0a30f4b80862025 Mon Sep 17 00:00:00 2001 From: "Jiadong.Zhu" Date: Wed, 7 Sep 2022 10:24:37 +0800 Subject: drm/amdgpu: MCBP based on DRM scheduler (v9) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Trigger Mid-Command Buffer Preemption according to the priority of the software rings and the hw fence signalling condition. The muxer saves the locations of the indirect buffer frames from the software ring together with the fence sequence number in its fifo queue, and pops out those records when the fences are signalled. The locations are used to resubmit packages in preemption scenarios by coping the chunks from the software ring. v2: Update comment style. v3: Fix conflict caused by previous modifications. v4: Remove unnecessary prints. v5: Fix corner cases for resubmission cases. v6: Refactor functions for resubmission, calling fence_process in irq handler. v7: Solve conflict for removing amdgpu_sw_ring.c. v8: Add time threshold to judge if preemption request is needed. v9: Correct comment spelling. Set fence emit timestamp before rsu assignment. Cc: Christian Koenig Cc: Luben Tuikov Cc: Andrey Grodzovsky Cc: Michel Dänzer Signed-off-by: Jiadong.Zhu Acked-by: Luben Tuikov Acked-by: Huang Rui Acked-by: Christian König Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 54 +++++ drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c | 2 + drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c | 12 + drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 8 + drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c | 351 ++++++++++++++++++++++++--- drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.h | 29 ++- drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 2 + drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c | 7 +- 8 files changed, 422 insertions(+), 43 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index d0d99ed607dd..00444203220d 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -55,6 +55,7 @@ struct amdgpu_fence { /* RB, DMA, etc. */ struct amdgpu_ring *ring; + ktime_t start_timestamp; }; static struct kmem_cache *amdgpu_fence_slab; @@ -199,6 +200,8 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amd } } + to_amdgpu_fence(fence)->start_timestamp = ktime_get(); + /* This function can't be called concurrently anyway, otherwise * emitting the fence would mess up the hardware ring buffer. */ @@ -406,6 +409,57 @@ unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring) return lower_32_bits(emitted); } +/** + * amdgpu_fence_last_unsignaled_time_us - the time fence emitted until now + * @ring: ring the fence is associated with + * + * Find the earliest fence unsignaled until now, calculate the time delta + * between the time fence emitted and now. + */ +u64 amdgpu_fence_last_unsignaled_time_us(struct amdgpu_ring *ring) +{ + struct amdgpu_fence_driver *drv = &ring->fence_drv; + struct dma_fence *fence; + uint32_t last_seq, sync_seq; + + last_seq = atomic_read(&ring->fence_drv.last_seq); + sync_seq = READ_ONCE(ring->fence_drv.sync_seq); + if (last_seq == sync_seq) + return 0; + + ++last_seq; + last_seq &= drv->num_fences_mask; + fence = drv->fences[last_seq]; + if (!fence) + return 0; + + return ktime_us_delta(ktime_get(), + to_amdgpu_fence(fence)->start_timestamp); +} + +/** + * amdgpu_fence_update_start_timestamp - update the timestamp of the fence + * @ring: ring the fence is associated with + * @seq: the fence seq number to update. + * @timestamp: the start timestamp to update. + * + * The function called at the time the fence and related ib is about to + * resubmit to gpu in MCBP scenario. Thus we do not consider race condition + * with amdgpu_fence_process to modify the same fence. + */ +void amdgpu_fence_update_start_timestamp(struct amdgpu_ring *ring, uint32_t seq, ktime_t timestamp) +{ + struct amdgpu_fence_driver *drv = &ring->fence_drv; + struct dma_fence *fence; + + seq &= drv->num_fences_mask; + fence = drv->fences[seq]; + if (!fence) + return; + + to_amdgpu_fence(fence)->start_timestamp = timestamp; +} + /** * amdgpu_fence_driver_start_ring - make the fence driver * ready for use on the requested ring. diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c index 774c77bb8f4e..bcccc348dbe2 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c @@ -211,6 +211,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs, } } + amdgpu_ring_ib_begin(ring); if (job && ring->funcs->init_cond_exec) patch_offset = amdgpu_ring_init_cond_exec(ring); @@ -285,6 +286,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs, ring->hw_prio == AMDGPU_GFX_PIPE_PRIO_HIGH) ring->funcs->emit_wave_limit(ring, false); + amdgpu_ring_ib_end(ring); amdgpu_ring_commit(ring); return 0; } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c index d3558c34d406..dc474b809604 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c @@ -569,3 +569,15 @@ int amdgpu_ring_init_mqd(struct amdgpu_ring *ring) return mqd_mgr->init_mqd(adev, ring->mqd_ptr, &prop); } + +void amdgpu_ring_ib_begin(struct amdgpu_ring *ring) +{ + if (ring->is_sw_ring) + amdgpu_sw_ring_ib_begin(ring); +} + +void amdgpu_ring_ib_end(struct amdgpu_ring *ring) +{ + if (ring->is_sw_ring) + amdgpu_sw_ring_ib_end(ring); +} diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h index 073e767c057f..f752c7ae7f60 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h @@ -145,8 +145,13 @@ signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring, uint32_t wait_seq, signed long timeout); unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring); + void amdgpu_fence_driver_isr_toggle(struct amdgpu_device *adev, bool stop); +u64 amdgpu_fence_last_unsignaled_time_us(struct amdgpu_ring *ring); +void amdgpu_fence_update_start_timestamp(struct amdgpu_ring *ring, uint32_t seq, + ktime_t timestamp); + /* * Rings. */ @@ -313,6 +318,9 @@ struct amdgpu_ring { #define amdgpu_ring_preempt_ib(r) (r)->funcs->preempt_ib(r) int amdgpu_ring_alloc(struct amdgpu_ring *ring, unsigned ndw); +void amdgpu_ring_ib_begin(struct amdgpu_ring *ring); +void amdgpu_ring_ib_end(struct amdgpu_ring *ring); + void amdgpu_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count); void amdgpu_ring_generic_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib); void amdgpu_ring_commit(struct amdgpu_ring *ring); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c index a52c4e7e8c39..3348337379da 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.c @@ -28,6 +28,7 @@ #include "amdgpu.h" #define AMDGPU_MUX_RESUBMIT_JIFFIES_TIMEOUT (HZ / 2) +#define AMDGPU_MAX_LAST_UNSIGNALED_THRESHOLD_US 10000 static const struct ring_info { unsigned int hw_pio; @@ -37,23 +38,145 @@ static const struct ring_info { { AMDGPU_RING_PRIO_2, "gfx_high"}, }; +static struct kmem_cache *amdgpu_mux_chunk_slab; + +static inline struct amdgpu_mux_entry *amdgpu_ring_mux_sw_entry(struct amdgpu_ring_mux *mux, + struct amdgpu_ring *ring) +{ + return ring->entry_index < mux->ring_entry_size ? + &mux->ring_entry[ring->entry_index] : NULL; +} + +/* copy packages on sw ring range[begin, end) */ +static void amdgpu_ring_mux_copy_pkt_from_sw_ring(struct amdgpu_ring_mux *mux, + struct amdgpu_ring *ring, + u64 s_start, u64 s_end) +{ + u64 start, end; + struct amdgpu_ring *real_ring = mux->real_ring; + + start = s_start & ring->buf_mask; + end = s_end & ring->buf_mask; + + if (start == end) { + DRM_ERROR("no more data copied from sw ring\n"); + return; + } + if (start > end) { + amdgpu_ring_alloc(real_ring, (ring->ring_size >> 2) + end - start); + amdgpu_ring_write_multiple(real_ring, (void *)&ring->ring[start], + (ring->ring_size >> 2) - start); + amdgpu_ring_write_multiple(real_ring, (void *)&ring->ring[0], end); + } else { + amdgpu_ring_alloc(real_ring, end - start); + amdgpu_ring_write_multiple(real_ring, (void *)&ring->ring[start], end - start); + } +} + +static void amdgpu_mux_resubmit_chunks(struct amdgpu_ring_mux *mux) +{ + struct amdgpu_mux_entry *e = NULL; + struct amdgpu_mux_chunk *chunk; + uint32_t seq, last_seq; + int i; + + /*find low priority entries:*/ + if (!mux->s_resubmit) + return; + + for (i = 0; i < mux->num_ring_entries; i++) { + if (mux->ring_entry[i].ring->hw_prio <= AMDGPU_RING_PRIO_DEFAULT) { + e = &mux->ring_entry[i]; + break; + } + } + + if (!e) { + DRM_ERROR("%s no low priority ring found\n", __func__); + return; + } + + last_seq = atomic_read(&e->ring->fence_drv.last_seq); + seq = mux->seqno_to_resubmit; + if (last_seq < seq) { + /*resubmit all the fences between (last_seq, seq]*/ + list_for_each_entry(chunk, &e->list, entry) { + if (chunk->sync_seq > last_seq && chunk->sync_seq <= seq) { + amdgpu_fence_update_start_timestamp(e->ring, + chunk->sync_seq, + ktime_get()); + amdgpu_ring_mux_copy_pkt_from_sw_ring(mux, e->ring, + chunk->start, + chunk->end); + mux->wptr_resubmit = chunk->end; + amdgpu_ring_commit(mux->real_ring); + } + } + } + + del_timer(&mux->resubmit_timer); + mux->s_resubmit = false; +} + +static void amdgpu_ring_mux_schedule_resubmit(struct amdgpu_ring_mux *mux) +{ + mod_timer(&mux->resubmit_timer, jiffies + AMDGPU_MUX_RESUBMIT_JIFFIES_TIMEOUT); +} + +static void amdgpu_mux_resubmit_fallback(struct timer_list *t) +{ + struct amdgpu_ring_mux *mux = from_timer(mux, t, resubmit_timer); + + if (!spin_trylock(&mux->lock)) { + amdgpu_ring_mux_schedule_resubmit(mux); + DRM_ERROR("reschedule resubmit\n"); + return; + } + amdgpu_mux_resubmit_chunks(mux); + spin_unlock(&mux->lock); +} + int amdgpu_ring_mux_init(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring, unsigned int entry_size) { mux->real_ring = ring; mux->num_ring_entries = 0; + mux->ring_entry = kcalloc(entry_size, sizeof(struct amdgpu_mux_entry), GFP_KERNEL); if (!mux->ring_entry) return -ENOMEM; mux->ring_entry_size = entry_size; + mux->s_resubmit = false; + + amdgpu_mux_chunk_slab = kmem_cache_create("amdgpu_mux_chunk", + sizeof(struct amdgpu_mux_chunk), 0, + SLAB_HWCACHE_ALIGN, NULL); + if (!amdgpu_mux_chunk_slab) { + DRM_ERROR("create amdgpu_mux_chunk cache failed\n"); + return -ENOMEM; + } + spin_lock_init(&mux->lock); + timer_setup(&mux->resubmit_timer, amdgpu_mux_resubmit_fallback, 0); return 0; } void amdgpu_ring_mux_fini(struct amdgpu_ring_mux *mux) { + struct amdgpu_mux_entry *e; + struct amdgpu_mux_chunk *chunk, *chunk2; + int i; + + for (i = 0; i < mux->num_ring_entries; i++) { + e = &mux->ring_entry[i]; + list_for_each_entry_safe(chunk, chunk2, &e->list, entry) { + list_del(&chunk->entry); + kmem_cache_free(amdgpu_mux_chunk_slab, chunk); + } + } + kmem_cache_destroy(amdgpu_mux_chunk_slab); kfree(mux->ring_entry); mux->ring_entry = NULL; mux->num_ring_entries = 0; @@ -73,62 +196,48 @@ int amdgpu_ring_mux_add_sw_ring(struct amdgpu_ring_mux *mux, struct amdgpu_ring ring->entry_index = mux->num_ring_entries; e->ring = ring; + INIT_LIST_HEAD(&e->list); mux->num_ring_entries += 1; return 0; } -static inline struct amdgpu_mux_entry *amdgpu_ring_mux_sw_entry(struct amdgpu_ring_mux *mux, - struct amdgpu_ring *ring) -{ - return ring->entry_index < mux->ring_entry_size ? - &mux->ring_entry[ring->entry_index] : NULL; -} - -/* copy packages on sw ring range[begin, end) */ -static void amdgpu_ring_mux_copy_pkt_from_sw_ring(struct amdgpu_ring_mux *mux, - struct amdgpu_ring *ring, - u64 s_start, u64 s_end) -{ - u64 start, end; - struct amdgpu_ring *real_ring = mux->real_ring; - - start = s_start & ring->buf_mask; - end = s_end & ring->buf_mask; - - if (start == end) { - DRM_ERROR("no more data copied from sw ring\n"); - return; - } - if (start > end) { - amdgpu_ring_alloc(real_ring, (ring->ring_size >> 2) + end - start); - amdgpu_ring_write_multiple(real_ring, (void *)&ring->ring[start], - (ring->ring_size >> 2) - start); - amdgpu_ring_write_multiple(real_ring, (void *)&ring->ring[0], end); - } else { - amdgpu_ring_alloc(real_ring, end - start); - amdgpu_ring_write_multiple(real_ring, (void *)&ring->ring[start], end - start); - } -} - void amdgpu_ring_mux_set_wptr(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring, u64 wptr) { struct amdgpu_mux_entry *e; + spin_lock(&mux->lock); + + if (ring->hw_prio <= AMDGPU_RING_PRIO_DEFAULT) + amdgpu_mux_resubmit_chunks(mux); + e = amdgpu_ring_mux_sw_entry(mux, ring); if (!e) { DRM_ERROR("cannot find entry for sw ring\n"); + spin_unlock(&mux->lock); + return; + } + + /* We could skip this set wptr as preemption in process. */ + if (ring->hw_prio <= AMDGPU_RING_PRIO_DEFAULT && mux->pending_trailing_fence_signaled) { + spin_unlock(&mux->lock); return; } - spin_lock(&mux->lock); e->sw_cptr = e->sw_wptr; + /* Update cptr if the package already copied in resubmit functions */ + if (ring->hw_prio <= AMDGPU_RING_PRIO_DEFAULT && e->sw_cptr < mux->wptr_resubmit) + e->sw_cptr = mux->wptr_resubmit; e->sw_wptr = wptr; e->start_ptr_in_hw_ring = mux->real_ring->wptr; - amdgpu_ring_mux_copy_pkt_from_sw_ring(mux, ring, e->sw_cptr, wptr); - e->end_ptr_in_hw_ring = mux->real_ring->wptr; - amdgpu_ring_commit(mux->real_ring); - + /* Skip copying for the packages already resubmitted.*/ + if (ring->hw_prio > AMDGPU_RING_PRIO_DEFAULT || mux->wptr_resubmit < wptr) { + amdgpu_ring_mux_copy_pkt_from_sw_ring(mux, ring, e->sw_cptr, wptr); + e->end_ptr_in_hw_ring = mux->real_ring->wptr; + amdgpu_ring_commit(mux->real_ring); + } else { + e->end_ptr_in_hw_ring = mux->real_ring->wptr; + } spin_unlock(&mux->lock); } @@ -239,3 +348,169 @@ unsigned int amdgpu_sw_ring_priority(int idx) return idx < ARRAY_SIZE(sw_ring_info) ? sw_ring_info[idx].hw_pio : AMDGPU_RING_PRIO_DEFAULT; } + +/*Scan on low prio rings to have unsignaled fence and high ring has no fence.*/ +int amdgpu_mcbp_scan(struct amdgpu_ring_mux *mux) +{ + struct amdgpu_ring *ring; + int i, need_preempt; + + need_preempt = 0; + for (i = 0; i < mux->num_ring_entries; i++) { + ring = mux->ring_entry[i].ring; + if (ring->hw_prio > AMDGPU_RING_PRIO_DEFAULT && + amdgpu_fence_count_emitted(ring) > 0) + return 0; + if (ring->hw_prio <= AMDGPU_RING_PRIO_DEFAULT && + amdgpu_fence_last_unsignaled_time_us(ring) > + AMDGPU_MAX_LAST_UNSIGNALED_THRESHOLD_US) + need_preempt = 1; + } + return need_preempt && !mux->s_resubmit; +} + +/* Trigger Mid-Command Buffer Preemption (MCBP) and find if we need to resubmit. */ +int amdgpu_mcbp_trigger_preempt(struct amdgpu_ring_mux *mux) +{ + int r; + + spin_lock(&mux->lock); + mux->pending_trailing_fence_signaled = true; + r = amdgpu_ring_preempt_ib(mux->real_ring); + spin_unlock(&mux->lock); + return r; +} + +void amdgpu_sw_ring_ib_begin(struct amdgpu_ring *ring) +{ + struct amdgpu_device *adev = ring->adev; + struct amdgpu_ring_mux *mux = &adev->gfx.muxer; + + WARN_ON(!ring->is_sw_ring); + if (ring->hw_prio > AMDGPU_RING_PRIO_DEFAULT) { + if (amdgpu_mcbp_scan(mux) > 0) + amdgpu_mcbp_trigger_preempt(mux); + return; + } + + amdgpu_ring_mux_start_ib(mux, ring); +} + +void amdgpu_sw_ring_ib_end(struct amdgpu_ring *ring) +{ + struct amdgpu_device *adev = ring->adev; + struct amdgpu_ring_mux *mux = &adev->gfx.muxer; + + WARN_ON(!ring->is_sw_ring); + if (ring->hw_prio > AMDGPU_RING_PRIO_DEFAULT) + return; + amdgpu_ring_mux_end_ib(mux, ring); +} + +void amdgpu_ring_mux_start_ib(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring) +{ + struct amdgpu_mux_entry *e; + struct amdgpu_mux_chunk *chunk; + + spin_lock(&mux->lock); + amdgpu_mux_resubmit_chunks(mux); + spin_unlock(&mux->lock); + + e = amdgpu_ring_mux_sw_entry(mux, ring); + if (!e) { + DRM_ERROR("cannot find entry!\n"); + return; + } + + chunk = kmem_cache_alloc(amdgpu_mux_chunk_slab, GFP_KERNEL); + if (!chunk) { + DRM_ERROR("alloc amdgpu_mux_chunk_slab failed\n"); + return; + } + + chunk->start = ring->wptr; + list_add_tail(&chunk->entry, &e->list); +} + +static void scan_and_remove_signaled_chunk(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring) +{ + uint32_t last_seq, size = 0; + struct amdgpu_mux_entry *e; + struct amdgpu_mux_chunk *chunk, *tmp; + + e = amdgpu_ring_mux_sw_entry(mux, ring); + if (!e) { + DRM_ERROR("cannot find entry!\n"); + return; + } + + last_seq = atomic_read(&ring->fence_drv.last_seq); + + list_for_each_entry_safe(chunk, tmp, &e->list, entry) { + if (chunk->sync_seq <= last_seq) { + list_del(&chunk->entry); + kmem_cache_free(amdgpu_mux_chunk_slab, chunk); + } else { + size++; + } + } +} + +void amdgpu_ring_mux_end_ib(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring) +{ + struct amdgpu_mux_entry *e; + struct amdgpu_mux_chunk *chunk; + + e = amdgpu_ring_mux_sw_entry(mux, ring); + if (!e) { + DRM_ERROR("cannot find entry!\n"); + return; + } + + chunk = list_last_entry(&e->list, struct amdgpu_mux_chunk, entry); + if (!chunk) { + DRM_ERROR("cannot find chunk!\n"); + return; + } + + chunk->end = ring->wptr; + chunk->sync_seq = READ_ONCE(ring->fence_drv.sync_seq); + + scan_and_remove_signaled_chunk(mux, ring); +} + +bool amdgpu_mcbp_handle_trailing_fence_irq(struct amdgpu_ring_mux *mux) +{ + struct amdgpu_mux_entry *e; + struct amdgpu_ring *ring = NULL; + int i; + + if (!mux->pending_trailing_fence_signaled) + return false; + + if (mux->real_ring->trail_seq != le32_to_cpu(*mux->real_ring->trail_fence_cpu_addr)) + return false; + + for (i = 0; i < mux->num_ring_entries; i++) { + e = &mux->ring_entry[i]; + if (e->ring->hw_prio <= AMDGPU_RING_PRIO_DEFAULT) { + ring = e->ring; + break; + } + } + + if (!ring) { + DRM_ERROR("cannot find low priority ring\n"); + return false; + } + + amdgpu_fence_process(ring); + if (amdgpu_fence_count_emitted(ring) > 0) { + mux->s_resubmit = true; + mux->seqno_to_resubmit = ring->fence_drv.sync_seq; + amdgpu_ring_mux_schedule_resubmit(mux); + } + + mux->pending_trailing_fence_signaled = false; + return true; +} diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.h index f67970dc3dbd..4be45fc14954 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring_mux.h @@ -29,6 +29,7 @@ #include "amdgpu_ring.h" struct amdgpu_ring; + /** * struct amdgpu_mux_entry - the entry recording software rings copying information. * @ring: the pointer to the software ring. @@ -37,6 +38,7 @@ struct amdgpu_ring; * @sw_cptr: the position of the copy pointer in the sw ring. * @sw_rptr: the read pointer in software ring. * @sw_wptr: the write pointer in software ring. + * @list: list head for amdgpu_mux_chunk */ struct amdgpu_mux_entry { struct amdgpu_ring *ring; @@ -45,6 +47,7 @@ struct amdgpu_mux_entry { u64 sw_cptr; u64 sw_rptr; u64 sw_wptr; + struct list_head list; }; struct amdgpu_ring_mux { @@ -55,6 +58,26 @@ struct amdgpu_ring_mux { unsigned int ring_entry_size; /*the lock for copy data from different software rings*/ spinlock_t lock; + bool s_resubmit; + uint32_t seqno_to_resubmit; + u64 wptr_resubmit; + struct timer_list resubmit_timer; + + bool pending_trailing_fence_signaled; +}; + +/** + * struct amdgpu_mux_chunk - save the location of indirect buffer's package on softare rings. + * @entry: the list entry. + * @sync_seq: the fence seqno related with the saved IB. + * @start:- start location on the software ring. + * @end:- end location on the software ring. + */ +struct amdgpu_mux_chunk { + struct list_head entry; + uint32_t sync_seq; + u64 start; + u64 end; }; int amdgpu_ring_mux_init(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring, @@ -64,15 +87,17 @@ int amdgpu_ring_mux_add_sw_ring(struct amdgpu_ring_mux *mux, struct amdgpu_ring void amdgpu_ring_mux_set_wptr(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring, u64 wptr); u64 amdgpu_ring_mux_get_wptr(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring); u64 amdgpu_ring_mux_get_rptr(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring); +void amdgpu_ring_mux_start_ib(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring); +void amdgpu_ring_mux_end_ib(struct amdgpu_ring_mux *mux, struct amdgpu_ring *ring); +bool amdgpu_mcbp_handle_trailing_fence_irq(struct amdgpu_ring_mux *mux); u64 amdgpu_sw_ring_get_rptr_gfx(struct amdgpu_ring *ring); u64 amdgpu_sw_ring_get_wptr_gfx(struct amdgpu_ring *ring); void amdgpu_sw_ring_set_wptr_gfx(struct amdgpu_ring *ring); - void amdgpu_sw_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count); void amdgpu_sw_ring_ib_begin(struct amdgpu_ring *ring); void amdgpu_sw_ring_ib_end(struct amdgpu_ring *ring); - const char *amdgpu_sw_ring_name(int idx); unsigned int amdgpu_sw_ring_priority(int idx); + #endif diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c index 003aa9e47085..37b461a19117 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c @@ -541,6 +541,7 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, if (!vm_flush_needed && !gds_switch_needed && !need_pipe_sync) return 0; + amdgpu_ring_ib_begin(ring); if (ring->funcs->init_cond_exec) patch_offset = amdgpu_ring_init_cond_exec(ring); @@ -601,6 +602,7 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, amdgpu_ring_emit_switch_buffer(ring); amdgpu_ring_emit_switch_buffer(ring); } + amdgpu_ring_ib_end(ring); return 0; } diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c index afcc6e397c72..f202b45c413c 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c @@ -5456,7 +5456,7 @@ static int gfx_v9_0_ring_preempt_ib(struct amdgpu_ring *ring) ring->trail_seq += 1; amdgpu_ring_alloc(ring, 13); gfx_v9_0_ring_emit_fence(ring, ring->trail_fence_gpu_addr, - ring->trail_seq, AMDGPU_FENCE_FLAG_EXEC); + ring->trail_seq, AMDGPU_FENCE_FLAG_EXEC | AMDGPU_FENCE_FLAG_INT); /*reset the CP_VMID_PREEMPT after trailing fence*/ amdgpu_ring_emit_wreg(ring, SOC15_REG_OFFSET(GC, 0, mmCP_VMID_PREEMPT), @@ -5882,8 +5882,9 @@ static int gfx_v9_0_eop_irq(struct amdgpu_device *adev, switch (me_id) { case 0: - /* Fence signals are handled on the software rings*/ - if (adev->gfx.num_gfx_rings) { + if (adev->gfx.num_gfx_rings && + !amdgpu_mcbp_handle_trailing_fence_irq(&adev->gfx.muxer)) { + /* Fence signals are handled on the software rings*/ for (i = 0; i < GFX9_NUM_SW_GFX_RINGS; i++) amdgpu_fence_process(&adev->gfx.sw_gfx_ring[i]); } -- cgit From 4670ac706ff9b3d0adb766ef9e93cc36d9dda474 Mon Sep 17 00:00:00 2001 From: Alex Deucher Date: Thu, 1 Dec 2022 16:41:53 -0500 Subject: drm/amdgpu: expand on GPUVM documentation MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Expand the GPUVM documentation to better describe the hardware functionality and use cases it serves. v2: Fixed a couple of spelling mistakes. Reviewed-by: Christian König Signed-off-by: Alex Deucher Link: https://lore.kernel.org/r/20221201214153.8453-2-alexander.deucher@amd.com Signed-off-by: Luben Tuikov Reviewed-by: Luben Tuikov --- drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 41 +++++++++++++++++++++++++--------- 1 file changed, 31 insertions(+), 10 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c index 37b461a19117..c05cff979004 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c @@ -45,22 +45,43 @@ /** * DOC: GPUVM * - * GPUVM is similar to the legacy gart on older asics, however - * rather than there being a single global gart table - * for the entire GPU, there are multiple VM page tables active - * at any given time. The VM page tables can contain a mix - * vram pages and system memory pages and system memory pages + * GPUVM is the MMU functionality provided on the GPU. + * GPUVM is similar to the legacy GART on older asics, however + * rather than there being a single global GART table + * for the entire GPU, there can be multiple GPUVM page tables active + * at any given time. The GPUVM page tables can contain a mix + * VRAM pages and system pages (both memory and MMIO) and system pages * can be mapped as snooped (cached system pages) or unsnooped * (uncached system pages). - * Each VM has an ID associated with it and there is a page table - * associated with each VMID. When executing a command buffer, - * the kernel tells the ring what VMID to use for that command + * + * Each active GPUVM has an ID associated with it and there is a page table + * linked with each VMID. When executing a command buffer, + * the kernel tells the engine what VMID to use for that command * buffer. VMIDs are allocated dynamically as commands are submitted. * The userspace drivers maintain their own address space and the kernel * sets up their pages tables accordingly when they submit their * command buffers and a VMID is assigned. - * Cayman/Trinity support up to 8 active VMs at any given time; - * SI supports 16. + * The hardware supports up to 16 active GPUVMs at any given time. + * + * Each GPUVM is represented by a 1-2 or 1-5 level page table, depending + * on the ASIC family. GPUVM supports RWX attributes on each page as well + * as other features such as encryption and caching attributes. + * + * VMID 0 is special. It is the GPUVM used for the kernel driver. In + * addition to an aperture managed by a page table, VMID 0 also has + * several other apertures. There is an aperture for direct access to VRAM + * and there is a legacy AGP aperture which just forwards accesses directly + * to the matching system physical addresses (or IOVAs when an IOMMU is + * present). These apertures provide direct access to these memories without + * incurring the overhead of a page table. VMID 0 is used by the kernel + * driver for tasks like memory management. + * + * GPU clients (i.e., engines on the GPU) use GPUVM VMIDs to access memory. + * For user applications, each application can have their own unique GPUVM + * address space. The application manages the address space and the kernel + * driver manages the GPUVM page tables for each process. If an GPU client + * accesses an invalid page, it will generate a GPU page fault, similar to + * accessing an invalid page on a CPU. */ #define START(node) ((node)->start) -- cgit From 56b0989e2939811c11ed9c449ff84cf85878ffe3 Mon Sep 17 00:00:00 2001 From: Christian König Date: Fri, 25 Nov 2022 16:04:25 +0100 Subject: drm/amdgpu: fix GDS/GWS/OA switch handling MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Bas pointed out that this isn't working as expected and could cause crashes. Fix the handling by storing the marker that a switch is needed inside the job instead. Reported-by: Bas Nieuwenhuizen Signed-off-by: Christian König Reviewed-by: Alex Deucher Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c | 42 ++++++++++++++++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_job.h | 1 + drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 54 ++++++++++++--------------------- 3 files changed, 54 insertions(+), 43 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c index 2a9a2593dc18..01878145a586 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c @@ -165,6 +165,26 @@ bool amdgpu_vmid_had_gpu_reset(struct amdgpu_device *adev, atomic_read(&adev->gpu_reset_counter); } +/* Check if we need to switch to another set of resources */ +static bool amdgpu_vmid_gds_switch_needed(struct amdgpu_vmid *id, + struct amdgpu_job *job) +{ + return id->gds_base != job->gds_base || + id->gds_size != job->gds_size || + id->gws_base != job->gws_base || + id->gws_size != job->gws_size || + id->oa_base != job->oa_base || + id->oa_size != job->oa_size; +} + +/* Check if the id is compatible with the job */ +static bool amdgpu_vmid_compatible(struct amdgpu_vmid *id, + struct amdgpu_job *job) +{ + return id->pd_gpu_addr == job->vm_pd_addr && + !amdgpu_vmid_gds_switch_needed(id, job); +} + /** * amdgpu_vmid_grab_idle - grab idle VMID * @@ -265,7 +285,7 @@ static int amdgpu_vmid_grab_reserved(struct amdgpu_vm *vm, *id = vm->reserved_vmid[vmhub]; if ((*id)->owner != vm->immediate.fence_context || - (*id)->pd_gpu_addr != job->vm_pd_addr || + !amdgpu_vmid_compatible(*id, job) || (*id)->flushed_updates < updates || !(*id)->last_flush || ((*id)->last_flush->context != fence_context && @@ -294,7 +314,6 @@ static int amdgpu_vmid_grab_reserved(struct amdgpu_vm *vm, if (r) return r; - (*id)->flushed_updates = updates; job->vm_needs_flush = needs_flush; return 0; } @@ -333,7 +352,7 @@ static int amdgpu_vmid_grab_used(struct amdgpu_vm *vm, if ((*id)->owner != vm->immediate.fence_context) continue; - if ((*id)->pd_gpu_addr != job->vm_pd_addr) + if (!amdgpu_vmid_compatible(*id, job)) continue; if (!(*id)->last_flush || @@ -355,7 +374,6 @@ static int amdgpu_vmid_grab_used(struct amdgpu_vm *vm, if (r) return r; - (*id)->flushed_updates = updates; job->vm_needs_flush |= needs_flush; return 0; } @@ -408,22 +426,30 @@ int amdgpu_vmid_grab(struct amdgpu_vm *vm, struct amdgpu_ring *ring, if (r) goto error; - id->flushed_updates = amdgpu_vm_tlb_seq(vm); job->vm_needs_flush = true; } list_move_tail(&id->list, &id_mgr->ids_lru); } - id->pd_gpu_addr = job->vm_pd_addr; - id->owner = vm->immediate.fence_context; - + job->gds_switch_needed = amdgpu_vmid_gds_switch_needed(id, job); if (job->vm_needs_flush) { + id->flushed_updates = amdgpu_vm_tlb_seq(vm); dma_fence_put(id->last_flush); id->last_flush = NULL; } job->vmid = id - id_mgr->ids; job->pasid = vm->pasid; + + id->gds_base = job->gds_base; + id->gds_size = job->gds_size; + id->gws_base = job->gws_base; + id->gws_size = job->gws_size; + id->oa_base = job->oa_base; + id->oa_size = job->oa_size; + id->pd_gpu_addr = job->vm_pd_addr; + id->owner = vm->immediate.fence_context; + trace_amdgpu_vm_grab_id(vm, ring, job); error: diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h index a372802ea4e0..02e85b040baf 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h @@ -53,6 +53,7 @@ struct amdgpu_job { uint32_t preamble_status; uint32_t preemption_status; bool vm_needs_flush; + bool gds_switch_needed; uint64_t vm_pd_addr; unsigned vmid; unsigned pasid; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c index c05cff979004..245c66ea10e7 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c @@ -484,25 +484,20 @@ bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring, struct amdgpu_device *adev = ring->adev; unsigned vmhub = ring->funcs->vmhub; struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub]; - struct amdgpu_vmid *id; - bool gds_switch_needed; - bool vm_flush_needed = job->vm_needs_flush || ring->has_compute_vm_bug; if (job->vmid == 0) return false; - id = &id_mgr->ids[job->vmid]; - gds_switch_needed = ring->funcs->emit_gds_switch && ( - id->gds_base != job->gds_base || - id->gds_size != job->gds_size || - id->gws_base != job->gws_base || - id->gws_size != job->gws_size || - id->oa_base != job->oa_base || - id->oa_size != job->oa_size); - - if (amdgpu_vmid_had_gpu_reset(adev, id)) + + if (job->vm_needs_flush || ring->has_compute_vm_bug) + return true; + + if (ring->funcs->emit_gds_switch && job->gds_switch_needed) return true; - return vm_flush_needed || gds_switch_needed; + if (amdgpu_vmid_had_gpu_reset(adev, &id_mgr->ids[job->vmid])) + return true; + + return false; } /** @@ -524,13 +519,8 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, unsigned vmhub = ring->funcs->vmhub; struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub]; struct amdgpu_vmid *id = &id_mgr->ids[job->vmid]; - bool gds_switch_needed = ring->funcs->emit_gds_switch && ( - id->gds_base != job->gds_base || - id->gds_size != job->gds_size || - id->gws_base != job->gws_base || - id->gws_size != job->gws_size || - id->oa_base != job->oa_base || - id->oa_size != job->oa_size); + bool gds_switch_needed = ring->funcs->emit_gds_switch && + job->gds_switch_needed; bool vm_flush_needed = job->vm_needs_flush; struct dma_fence *fence = NULL; bool pasid_mapping_needed = false; @@ -577,6 +567,14 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, if (pasid_mapping_needed) amdgpu_gmc_emit_pasid_mapping(ring, job->vmid, job->pasid); + if (!ring->is_mes_queue && ring->funcs->emit_gds_switch && + gds_switch_needed) { + amdgpu_ring_emit_gds_switch(ring, job->vmid, job->gds_base, + job->gds_size, job->gws_base, + job->gws_size, job->oa_base, + job->oa_size); + } + if (vm_flush_needed || pasid_mapping_needed) { r = amdgpu_fence_emit(ring, &fence, NULL, 0); if (r) @@ -601,20 +599,6 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, } dma_fence_put(fence); - if (!ring->is_mes_queue && ring->funcs->emit_gds_switch && - gds_switch_needed) { - id->gds_base = job->gds_base; - id->gds_size = job->gds_size; - id->gws_base = job->gws_base; - id->gws_size = job->gws_size; - id->oa_base = job->oa_base; - id->oa_size = job->oa_size; - amdgpu_ring_emit_gds_switch(ring, job->vmid, job->gds_base, - job->gds_size, job->gws_base, - job->gws_size, job->oa_base, - job->oa_size); - } - if (ring->funcs->patch_cond_exec) amdgpu_ring_patch_cond_exec(ring, patch_offset); -- cgit From 5f3c40e9e2460c42f5bf6c51b1e393d7159241c3 Mon Sep 17 00:00:00 2001 From: Christian König Date: Fri, 25 Nov 2022 16:42:45 +0100 Subject: drm/amdgpu: cleanup SPM support a bit MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit This should probably not access job->vm and also emit the SPM switch under the conditional execute. Signed-off-by: Christian König Reviewed-by: Alex Deucher Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c | 1 + drivers/gpu/drm/amd/amdgpu/amdgpu_job.h | 1 + drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 9 +++++---- 3 files changed, 7 insertions(+), 4 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c index 01878145a586..6949dfec75d5 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c @@ -315,6 +315,7 @@ static int amdgpu_vmid_grab_reserved(struct amdgpu_vm *vm, return r; job->vm_needs_flush = needs_flush; + job->spm_update_needed = true; return 0; } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h index 02e85b040baf..52f2e313ea17 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h @@ -54,6 +54,7 @@ struct amdgpu_job { uint32_t preemption_status; bool vm_needs_flush; bool gds_switch_needed; + bool spm_update_needed; uint64_t vm_pd_addr; unsigned vmid; unsigned pasid; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c index 245c66ea10e7..a05cce3f3170 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c @@ -519,22 +519,20 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, unsigned vmhub = ring->funcs->vmhub; struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub]; struct amdgpu_vmid *id = &id_mgr->ids[job->vmid]; + bool spm_update_needed = job->spm_update_needed; bool gds_switch_needed = ring->funcs->emit_gds_switch && job->gds_switch_needed; bool vm_flush_needed = job->vm_needs_flush; struct dma_fence *fence = NULL; bool pasid_mapping_needed = false; unsigned patch_offset = 0; - bool update_spm_vmid_needed = (job->vm && (job->vm->reserved_vmid[vmhub] != NULL)); int r; - if (update_spm_vmid_needed && adev->gfx.rlc.funcs->update_spm_vmid) - adev->gfx.rlc.funcs->update_spm_vmid(adev, job->vmid); - if (amdgpu_vmid_had_gpu_reset(adev, id)) { gds_switch_needed = true; vm_flush_needed = true; pasid_mapping_needed = true; + spm_update_needed = true; } mutex_lock(&id_mgr->lock); @@ -567,6 +565,9 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, if (pasid_mapping_needed) amdgpu_gmc_emit_pasid_mapping(ring, job->vmid, job->pasid); + if (spm_update_needed && adev->gfx.rlc.funcs->update_spm_vmid) + adev->gfx.rlc.funcs->update_spm_vmid(adev, job->vmid); + if (!ring->is_mes_queue && ring->funcs->emit_gds_switch && gds_switch_needed) { amdgpu_ring_emit_gds_switch(ring, job->vmid, job->gds_base, -- cgit From 053499f7b45dc56758240615569b349fe9e2fc8d Mon Sep 17 00:00:00 2001 From: Christian König Date: Fri, 25 Nov 2022 16:45:09 +0100 Subject: drm/amdgpu: stop waiting for the VM during unreserve MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit This is completely pointless since the VMID always stays allocated until the VM is idle. Signed-off-by: Christian König Reviewed-by: Alex Deucher Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 16 ---------------- 1 file changed, 16 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c index a05cce3f3170..dc379dc22c77 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c @@ -2368,7 +2368,6 @@ int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) union drm_amdgpu_vm *args = data; struct amdgpu_device *adev = drm_to_adev(dev); struct amdgpu_fpriv *fpriv = filp->driver_priv; - long timeout = msecs_to_jiffies(2000); int r; switch (args->in.op) { @@ -2380,21 +2379,6 @@ int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) return r; break; case AMDGPU_VM_OP_UNRESERVE_VMID: - if (amdgpu_sriov_runtime(adev)) - timeout = 8 * timeout; - - /* Wait vm idle to make sure the vmid set in SPM_VMID is - * not referenced anymore. - */ - r = amdgpu_bo_reserve(fpriv->vm.root.bo, true); - if (r) - return r; - - r = amdgpu_vm_wait_idle(&fpriv->vm, timeout); - if (r < 0) - return r; - - amdgpu_bo_unreserve(fpriv->vm.root.bo); amdgpu_vmid_free_reserved(adev, &fpriv->vm, AMDGPU_GFXHUB_0); break; default: -- cgit