From cadf97b196a1e5b2db2606d53f77714e3e9cf4bb Mon Sep 17 00:00:00 2001 From: Chunming Zhou Date: Fri, 15 Jan 2016 11:25:00 +0800 Subject: drm/amdgpu: clean up non-scheduler code path (v2) Non-scheduler code is longer supported. v2: agd: rebased on upstream Signed-off-by: Chunming Zhou Reviewed-by: Ken Wang Reviewed-by: Monk Liu Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 39 +++++++++++++++---------------- 1 file changed, 19 insertions(+), 20 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index 3671f9f220bd..cac03e743b58 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -472,6 +472,7 @@ int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring, int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring) { int i, r; + long timeout; ring->fence_drv.cpu_addr = NULL; ring->fence_drv.gpu_addr = 0; @@ -486,26 +487,24 @@ int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring) init_waitqueue_head(&ring->fence_drv.fence_queue); - if (amdgpu_enable_scheduler) { - long timeout = msecs_to_jiffies(amdgpu_lockup_timeout); - if (timeout == 0) { - /* - * FIXME: - * Delayed workqueue cannot use it directly, - * so the scheduler will not use delayed workqueue if - * MAX_SCHEDULE_TIMEOUT is set. - * Currently keep it simple and silly. - */ - timeout = MAX_SCHEDULE_TIMEOUT; - } - r = amd_sched_init(&ring->sched, &amdgpu_sched_ops, - amdgpu_sched_hw_submission, - timeout, ring->name); - if (r) { - DRM_ERROR("Failed to create scheduler on ring %s.\n", - ring->name); - return r; - } + timeout = msecs_to_jiffies(amdgpu_lockup_timeout); + if (timeout == 0) { + /* + * FIXME: + * Delayed workqueue cannot use it directly, + * so the scheduler will not use delayed workqueue if + * MAX_SCHEDULE_TIMEOUT is set. + * Currently keep it simple and silly. + */ + timeout = MAX_SCHEDULE_TIMEOUT; + } + r = amd_sched_init(&ring->sched, &amdgpu_sched_ops, + amdgpu_sched_hw_submission, + timeout, ring->name); + if (r) { + DRM_ERROR("Failed to create scheduler on ring %s.\n", + ring->name); + return r; } return 0; -- cgit v1.2.3 From 5907a0d8af71d17811be49f2c056b3a89660e188 Mon Sep 17 00:00:00 2001 From: Christian König Date: Mon, 18 Jan 2016 15:16:53 +0100 Subject: drm/amdgpu: cleanup sync_seq handling MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Not used any more without semaphores Signed-off-by: Christian König Reviewed-by: Chunming Zhou Acked-by: Alex Deucher Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 7 +-- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 94 ++++--------------------------- drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c | 2 +- 3 files changed, 13 insertions(+), 90 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h index 3e040cb20ebe..cd93104b4fbb 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h @@ -388,7 +388,7 @@ struct amdgpu_fence_driver { uint64_t gpu_addr; volatile uint32_t *cpu_addr; /* sync_seq is protected by ring emission lock */ - uint64_t sync_seq[AMDGPU_MAX_RINGS]; + uint64_t sync_seq; atomic64_t last_seq; bool initialized; struct amdgpu_irq_src *irq_src; @@ -441,11 +441,6 @@ int amdgpu_fence_wait_next(struct amdgpu_ring *ring); int amdgpu_fence_wait_empty(struct amdgpu_ring *ring); unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring); -bool amdgpu_fence_need_sync(struct amdgpu_fence *fence, - struct amdgpu_ring *ring); -void amdgpu_fence_note_sync(struct amdgpu_fence *fence, - struct amdgpu_ring *ring); - /* * TTM. */ diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index cac03e743b58..988a32d578a0 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -107,7 +107,7 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, void *owner, if ((*fence) == NULL) { return -ENOMEM; } - (*fence)->seq = ++ring->fence_drv.sync_seq[ring->idx]; + (*fence)->seq = ++ring->fence_drv.sync_seq; (*fence)->ring = ring; (*fence)->owner = owner; fence_init(&(*fence)->base, &amdgpu_fence_ops, @@ -171,7 +171,7 @@ static bool amdgpu_fence_activity(struct amdgpu_ring *ring) */ last_seq = atomic64_read(&ring->fence_drv.last_seq); do { - last_emitted = ring->fence_drv.sync_seq[ring->idx]; + last_emitted = ring->fence_drv.sync_seq; seq = amdgpu_fence_read(ring); seq |= last_seq & 0xffffffff00000000LL; if (seq < last_seq) { @@ -274,7 +274,7 @@ static int amdgpu_fence_ring_wait_seq(struct amdgpu_ring *ring, uint64_t seq) bool signaled = false; BUG_ON(!ring); - if (seq > ring->fence_drv.sync_seq[ring->idx]) + if (seq > ring->fence_drv.sync_seq) return -EINVAL; if (atomic64_read(&ring->fence_drv.last_seq) >= seq) @@ -304,7 +304,7 @@ int amdgpu_fence_wait_next(struct amdgpu_ring *ring) { uint64_t seq = atomic64_read(&ring->fence_drv.last_seq) + 1ULL; - if (seq >= ring->fence_drv.sync_seq[ring->idx]) + if (seq >= ring->fence_drv.sync_seq) return -ENOENT; return amdgpu_fence_ring_wait_seq(ring, seq); @@ -322,7 +322,7 @@ int amdgpu_fence_wait_next(struct amdgpu_ring *ring) */ int amdgpu_fence_wait_empty(struct amdgpu_ring *ring) { - uint64_t seq = ring->fence_drv.sync_seq[ring->idx]; + uint64_t seq = ring->fence_drv.sync_seq; if (!seq) return 0; @@ -347,7 +347,7 @@ unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring) * but it's ok to report slightly wrong fence count here. */ amdgpu_fence_process(ring); - emitted = ring->fence_drv.sync_seq[ring->idx] + emitted = ring->fence_drv.sync_seq - atomic64_read(&ring->fence_drv.last_seq); /* to avoid 32bits warp around */ if (emitted > 0x10000000) @@ -356,68 +356,6 @@ unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring) return (unsigned)emitted; } -/** - * amdgpu_fence_need_sync - do we need a semaphore - * - * @fence: amdgpu fence object - * @dst_ring: which ring to check against - * - * Check if the fence needs to be synced against another ring - * (all asics). If so, we need to emit a semaphore. - * Returns true if we need to sync with another ring, false if - * not. - */ -bool amdgpu_fence_need_sync(struct amdgpu_fence *fence, - struct amdgpu_ring *dst_ring) -{ - struct amdgpu_fence_driver *fdrv; - - if (!fence) - return false; - - if (fence->ring == dst_ring) - return false; - - /* we are protected by the ring mutex */ - fdrv = &dst_ring->fence_drv; - if (fence->seq <= fdrv->sync_seq[fence->ring->idx]) - return false; - - return true; -} - -/** - * amdgpu_fence_note_sync - record the sync point - * - * @fence: amdgpu fence object - * @dst_ring: which ring to check against - * - * Note the sequence number at which point the fence will - * be synced with the requested ring (all asics). - */ -void amdgpu_fence_note_sync(struct amdgpu_fence *fence, - struct amdgpu_ring *dst_ring) -{ - struct amdgpu_fence_driver *dst, *src; - unsigned i; - - if (!fence) - return; - - if (fence->ring == dst_ring) - return; - - /* we are protected by the ring mutex */ - src = &fence->ring->fence_drv; - dst = &dst_ring->fence_drv; - for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { - if (i == dst_ring->idx) - continue; - - dst->sync_seq[i] = max(dst->sync_seq[i], src->sync_seq[i]); - } -} - /** * amdgpu_fence_driver_start_ring - make the fence driver * ready for use on the requested ring. @@ -471,14 +409,12 @@ int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring, */ int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring) { - int i, r; long timeout; + int r; ring->fence_drv.cpu_addr = NULL; ring->fence_drv.gpu_addr = 0; - for (i = 0; i < AMDGPU_MAX_RINGS; ++i) - ring->fence_drv.sync_seq[i] = 0; - + ring->fence_drv.sync_seq = 0; atomic64_set(&ring->fence_drv.last_seq, 0); ring->fence_drv.initialized = false; @@ -650,7 +586,7 @@ void amdgpu_fence_driver_force_completion(struct amdgpu_device *adev) if (!ring || !ring->fence_drv.initialized) continue; - amdgpu_fence_write(ring, ring->fence_drv.sync_seq[i]); + amdgpu_fence_write(ring, ring->fence_drv.sync_seq); } } @@ -780,7 +716,7 @@ static int amdgpu_debugfs_fence_info(struct seq_file *m, void *data) struct drm_info_node *node = (struct drm_info_node *)m->private; struct drm_device *dev = node->minor->dev; struct amdgpu_device *adev = dev->dev_private; - int i, j; + int i; for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { struct amdgpu_ring *ring = adev->rings[i]; @@ -793,15 +729,7 @@ static int amdgpu_debugfs_fence_info(struct seq_file *m, void *data) seq_printf(m, "Last signaled fence 0x%016llx\n", (unsigned long long)atomic64_read(&ring->fence_drv.last_seq)); seq_printf(m, "Last emitted 0x%016llx\n", - ring->fence_drv.sync_seq[i]); - - for (j = 0; j < AMDGPU_MAX_RINGS; ++j) { - struct amdgpu_ring *other = adev->rings[j]; - if (i != j && other && other->fence_drv.initialized && - ring->fence_drv.sync_seq[j]) - seq_printf(m, "Last sync to ring %d 0x%016llx\n", - j, ring->fence_drv.sync_seq[j]); - } + ring->fence_drv.sync_seq); } return 0; } diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c index dea4ba9dfa6d..248271050d0a 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c @@ -4766,7 +4766,7 @@ 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); - uint32_t seq = ring->fence_drv.sync_seq[ring->idx]; + uint32_t seq = ring->fence_drv.sync_seq; uint64_t addr = ring->fence_drv.gpu_addr; amdgpu_ring_write(ring, PACKET3(PACKET3_WAIT_REG_MEM, 5)); -- cgit v1.2.3 From 18db89b4025a9fb71824d0ffd84594c9f6f41cec Mon Sep 17 00:00:00 2001 From: Alex Deucher Date: Thu, 14 Jan 2016 10:25:22 -0500 Subject: drm/amdgpu: add a debugfs property to trigger a GPU reset MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Ported from similar code in radeon. Reviewed-by: Junwei Zhang Reviewed-by: Christian König Reviewed-by: Ken Wang Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 20 +++++++++++++++++++- 1 file changed, 19 insertions(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index 988a32d578a0..08963fc83168 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -734,15 +734,33 @@ static int amdgpu_debugfs_fence_info(struct seq_file *m, void *data) return 0; } +/** + * amdgpu_debugfs_gpu_reset - manually trigger a gpu reset + * + * Manually trigger a gpu reset at the next fence wait. + */ +static int amdgpu_debugfs_gpu_reset(struct seq_file *m, void *data) +{ + struct drm_info_node *node = (struct drm_info_node *) m->private; + struct drm_device *dev = node->minor->dev; + struct amdgpu_device *adev = dev->dev_private; + + seq_printf(m, "gpu reset\n"); + amdgpu_gpu_reset(adev); + + return 0; +} + static struct drm_info_list amdgpu_debugfs_fence_list[] = { {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL}, + {"amdgpu_gpu_reset", &amdgpu_debugfs_gpu_reset, 0, NULL} }; #endif int amdgpu_debugfs_fence_init(struct amdgpu_device *adev) { #if defined(CONFIG_DEBUG_FS) - return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list, 1); + return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list, 2); #else return 0; #endif -- cgit v1.2.3 From a27de35caab59bacf5c47713856739f86ec06e43 Mon Sep 17 00:00:00 2001 From: Christian König Date: Thu, 21 Jan 2016 11:28:53 +0100 Subject: drm/amdgpu: remove the ring lock v2 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit It's not needed any more because all access goes through the scheduler now. v2: Update commit message. Signed-off-by: Christian König Reviewed-by: Chunming Zhou Acked-by: Alex Deucher Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 5 -- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 1 - drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 6 --- drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c | 10 ++-- drivers/gpu/drm/amd/amdgpu/amdgpu_pm.c | 19 +++---- drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c | 80 +++--------------------------- drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c | 4 +- drivers/gpu/drm/amd/amdgpu/cik_sdma.c | 4 +- drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c | 8 +-- drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c | 8 +-- drivers/gpu/drm/amd/amdgpu/sdma_v2_4.c | 4 +- drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c | 4 +- drivers/gpu/drm/amd/amdgpu/uvd_v4_2.c | 8 +-- drivers/gpu/drm/amd/amdgpu/uvd_v5_0.c | 8 +-- drivers/gpu/drm/amd/amdgpu/uvd_v6_0.c | 8 +-- 15 files changed, 45 insertions(+), 132 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h index 3d4c2abfcfe8..081953e1ec6c 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h @@ -814,7 +814,6 @@ struct amdgpu_ring { struct amd_gpu_scheduler sched; spinlock_t fence_lock; - struct mutex *ring_lock; struct amdgpu_bo *ring_obj; volatile uint32_t *ring; unsigned rptr_offs; @@ -1190,12 +1189,9 @@ int amdgpu_ib_ring_tests(struct amdgpu_device *adev); /* Ring access between begin & end cannot sleep */ 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); -void amdgpu_ring_unlock_undo(struct amdgpu_ring *ring); unsigned amdgpu_ring_backup(struct amdgpu_ring *ring, uint32_t **data); int amdgpu_ring_restore(struct amdgpu_ring *ring, @@ -2009,7 +2005,6 @@ struct amdgpu_device { /* rings */ unsigned fence_context; - struct mutex ring_lock; unsigned num_rings; struct amdgpu_ring *rings[AMDGPU_MAX_RINGS]; bool ib_pool_ready; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c index bbe8023bf58f..cf8a3b37a111 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c @@ -1455,7 +1455,6 @@ int amdgpu_device_init(struct amdgpu_device *adev, /* mutex initialization are all done here so we * can recall function without having locking issues */ - mutex_init(&adev->ring_lock); mutex_init(&adev->vm_manager.lock); atomic_set(&adev->irq.ih.lock, 0); mutex_init(&adev->gem.mutex); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index 08963fc83168..72105020086c 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -487,7 +487,6 @@ void amdgpu_fence_driver_fini(struct amdgpu_device *adev) if (atomic_dec_and_test(&amdgpu_fence_slab_ref)) kmem_cache_destroy(amdgpu_fence_slab); - mutex_lock(&adev->ring_lock); for (i = 0; i < AMDGPU_MAX_RINGS; i++) { struct amdgpu_ring *ring = adev->rings[i]; @@ -505,7 +504,6 @@ void amdgpu_fence_driver_fini(struct amdgpu_device *adev) del_timer_sync(&ring->fence_drv.fallback_timer); ring->fence_drv.initialized = false; } - mutex_unlock(&adev->ring_lock); } /** @@ -520,7 +518,6 @@ void amdgpu_fence_driver_suspend(struct amdgpu_device *adev) { int i, r; - mutex_lock(&adev->ring_lock); for (i = 0; i < AMDGPU_MAX_RINGS; i++) { struct amdgpu_ring *ring = adev->rings[i]; if (!ring || !ring->fence_drv.initialized) @@ -537,7 +534,6 @@ void amdgpu_fence_driver_suspend(struct amdgpu_device *adev) amdgpu_irq_put(adev, ring->fence_drv.irq_src, ring->fence_drv.irq_type); } - mutex_unlock(&adev->ring_lock); } /** @@ -556,7 +552,6 @@ void amdgpu_fence_driver_resume(struct amdgpu_device *adev) { int i; - mutex_lock(&adev->ring_lock); for (i = 0; i < AMDGPU_MAX_RINGS; i++) { struct amdgpu_ring *ring = adev->rings[i]; if (!ring || !ring->fence_drv.initialized) @@ -566,7 +561,6 @@ void amdgpu_fence_driver_resume(struct amdgpu_device *adev) amdgpu_irq_get(adev, ring->fence_drv.irq_src, ring->fence_drv.irq_type); } - mutex_unlock(&adev->ring_lock); } /** diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c index 56ae9a58dbc5..40c9779993c8 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c @@ -147,7 +147,7 @@ int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs, return -EINVAL; } - r = amdgpu_ring_lock(ring, (256 + AMDGPU_NUM_SYNCS * 8) * num_ibs); + r = amdgpu_ring_alloc(ring, (256 + AMDGPU_NUM_SYNCS * 8) * num_ibs); if (r) { dev_err(adev->dev, "scheduling IB failed (%d).\n", r); return r; @@ -155,7 +155,7 @@ int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs, r = amdgpu_sync_wait(&ibs->sync); if (r) { - amdgpu_ring_unlock_undo(ring); + amdgpu_ring_undo(ring); dev_err(adev->dev, "failed to sync wait (%d)\n", r); return r; } @@ -180,7 +180,7 @@ int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs, if (ib->ring != ring || ib->ctx != ctx || ib->vm != vm) { ring->current_ctx = old_ctx; - amdgpu_ring_unlock_undo(ring); + amdgpu_ring_undo(ring); return -EINVAL; } amdgpu_ring_emit_ib(ring, ib); @@ -191,7 +191,7 @@ int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs, if (r) { dev_err(adev->dev, "failed to emit fence (%d)\n", r); ring->current_ctx = old_ctx; - amdgpu_ring_unlock_undo(ring); + amdgpu_ring_undo(ring); return r; } @@ -203,7 +203,7 @@ int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs, AMDGPU_FENCE_FLAG_64BIT); } - amdgpu_ring_unlock_commit(ring); + amdgpu_ring_commit(ring); return 0; } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_pm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_pm.c index 7d8d84eaea4a..a0da563c8c82 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_pm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_pm.c @@ -623,14 +623,12 @@ force: amdgpu_dpm_print_power_state(adev, adev->pm.dpm.requested_ps); } - mutex_lock(&adev->ring_lock); - /* update whether vce is active */ ps->vce_active = adev->pm.dpm.vce_active; ret = amdgpu_dpm_pre_set_power_state(adev); if (ret) - goto done; + return; /* update display watermarks based on new power state */ amdgpu_display_bandwidth_update(adev); @@ -667,9 +665,6 @@ force: amdgpu_dpm_force_performance_level(adev, adev->pm.dpm.forced_level); } } - -done: - mutex_unlock(&adev->ring_lock); } void amdgpu_dpm_enable_uvd(struct amdgpu_device *adev, bool enable) @@ -802,13 +797,11 @@ void amdgpu_pm_compute_clocks(struct amdgpu_device *adev) int i = 0; amdgpu_display_bandwidth_update(adev); - mutex_lock(&adev->ring_lock); - for (i = 0; i < AMDGPU_MAX_RINGS; i++) { - struct amdgpu_ring *ring = adev->rings[i]; - if (ring && ring->ready) - amdgpu_fence_wait_empty(ring); - } - mutex_unlock(&adev->ring_lock); + for (i = 0; i < AMDGPU_MAX_RINGS; i++) { + struct amdgpu_ring *ring = adev->rings[i]; + if (ring && ring->ready) + amdgpu_fence_wait_empty(ring); + } amdgpu_dpm_dispatch_task(adev, AMD_PP_EVENT_DISPLAY_CONFIG_CHANGE, NULL, NULL); } else { diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c index 66c6bbd27309..81d06d772dde 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c @@ -105,30 +105,6 @@ int amdgpu_ring_alloc(struct amdgpu_ring *ring, unsigned ndw) return 0; } -/** - * amdgpu_ring_lock - lock the ring and allocate space on it - * - * @adev: amdgpu_device pointer - * @ring: amdgpu_ring structure holding ring information - * @ndw: number of dwords to allocate in the ring buffer - * - * Lock the ring and allocate @ndw dwords in the ring buffer - * (all asics). - * Returns 0 on success, error on failure. - */ -int amdgpu_ring_lock(struct amdgpu_ring *ring, unsigned ndw) -{ - int r; - - mutex_lock(ring->ring_lock); - r = amdgpu_ring_alloc(ring, ndw); - if (r) { - mutex_unlock(ring->ring_lock); - return r; - } - return 0; -} - /** amdgpu_ring_insert_nop - insert NOP packets * * @ring: amdgpu_ring structure holding ring information @@ -167,20 +143,6 @@ void amdgpu_ring_commit(struct amdgpu_ring *ring) amdgpu_ring_set_wptr(ring); } -/** - * amdgpu_ring_unlock_commit - tell the GPU to execute the new - * commands on the ring buffer and unlock it - * - * @ring: amdgpu_ring structure holding ring information - * - * Call amdgpu_ring_commit() then unlock the ring (all asics). - */ -void amdgpu_ring_unlock_commit(struct amdgpu_ring *ring) -{ - amdgpu_ring_commit(ring); - mutex_unlock(ring->ring_lock); -} - /** * amdgpu_ring_undo - reset the wptr * @@ -193,19 +155,6 @@ void amdgpu_ring_undo(struct amdgpu_ring *ring) ring->wptr = ring->wptr_old; } -/** - * amdgpu_ring_unlock_undo - reset the wptr and unlock the ring - * - * @ring: amdgpu_ring structure holding ring information - * - * Call amdgpu_ring_undo() then unlock the ring (all asics). - */ -void amdgpu_ring_unlock_undo(struct amdgpu_ring *ring) -{ - amdgpu_ring_undo(ring); - mutex_unlock(ring->ring_lock); -} - /** * amdgpu_ring_backup - Back up the content of a ring * @@ -218,43 +167,32 @@ unsigned amdgpu_ring_backup(struct amdgpu_ring *ring, { unsigned size, ptr, i; - /* just in case lock the ring */ - mutex_lock(ring->ring_lock); *data = NULL; - if (ring->ring_obj == NULL) { - mutex_unlock(ring->ring_lock); + if (ring->ring_obj == NULL) return 0; - } /* it doesn't make sense to save anything if all fences are signaled */ - if (!amdgpu_fence_count_emitted(ring)) { - mutex_unlock(ring->ring_lock); + if (!amdgpu_fence_count_emitted(ring)) return 0; - } ptr = le32_to_cpu(*ring->next_rptr_cpu_addr); size = ring->wptr + (ring->ring_size / 4); size -= ptr; size &= ring->ptr_mask; - if (size == 0) { - mutex_unlock(ring->ring_lock); + if (size == 0) return 0; - } /* and then save the content of the ring */ *data = kmalloc_array(size, sizeof(uint32_t), GFP_KERNEL); - if (!*data) { - mutex_unlock(ring->ring_lock); + if (!*data) return 0; - } for (i = 0; i < size; ++i) { (*data)[i] = ring->ring[ptr++]; ptr &= ring->ptr_mask; } - mutex_unlock(ring->ring_lock); return size; } @@ -276,7 +214,7 @@ int amdgpu_ring_restore(struct amdgpu_ring *ring, return 0; /* restore the saved ring content */ - r = amdgpu_ring_lock(ring, size); + r = amdgpu_ring_alloc(ring, size); if (r) return r; @@ -284,7 +222,7 @@ int amdgpu_ring_restore(struct amdgpu_ring *ring, amdgpu_ring_write(ring, data[i]); } - amdgpu_ring_unlock_commit(ring); + amdgpu_ring_commit(ring); kfree(data); return 0; } @@ -352,7 +290,6 @@ int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring, return r; } - ring->ring_lock = &adev->ring_lock; /* Align ring size */ rb_bufsz = order_base_2(ring_size / 8); ring_size = (1 << (rb_bufsz + 1)) * 4; @@ -410,15 +347,10 @@ void amdgpu_ring_fini(struct amdgpu_ring *ring) int r; struct amdgpu_bo *ring_obj; - if (ring->ring_lock == NULL) - return; - - mutex_lock(ring->ring_lock); ring_obj = ring->ring_obj; ring->ready = false; ring->ring = NULL; ring->ring_obj = NULL; - mutex_unlock(ring->ring_lock); amdgpu_wb_free(ring->adev, ring->fence_offs); amdgpu_wb_free(ring->adev, ring->rptr_offs); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c index 16fbde9c5f56..c90517f61210 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c @@ -788,14 +788,14 @@ int amdgpu_vce_ring_test_ring(struct amdgpu_ring *ring) unsigned i; int r; - r = amdgpu_ring_lock(ring, 16); + r = amdgpu_ring_alloc(ring, 16); if (r) { DRM_ERROR("amdgpu: vce failed to lock ring %d (%d).\n", ring->idx, r); return r; } amdgpu_ring_write(ring, VCE_CMD_END); - amdgpu_ring_unlock_commit(ring); + amdgpu_ring_commit(ring); for (i = 0; i < adev->usec_timeout; i++) { if (amdgpu_ring_get_rptr(ring) != rptr) diff --git a/drivers/gpu/drm/amd/amdgpu/cik_sdma.c b/drivers/gpu/drm/amd/amdgpu/cik_sdma.c index 538d6a26e529..e8a48ae8d360 100644 --- a/drivers/gpu/drm/amd/amdgpu/cik_sdma.c +++ b/drivers/gpu/drm/amd/amdgpu/cik_sdma.c @@ -560,7 +560,7 @@ static int cik_sdma_ring_test_ring(struct amdgpu_ring *ring) tmp = 0xCAFEDEAD; adev->wb.wb[index] = cpu_to_le32(tmp); - r = amdgpu_ring_lock(ring, 5); + r = amdgpu_ring_alloc(ring, 5); if (r) { DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r); amdgpu_wb_free(adev, index); @@ -571,7 +571,7 @@ static int cik_sdma_ring_test_ring(struct amdgpu_ring *ring) amdgpu_ring_write(ring, upper_32_bits(gpu_addr)); amdgpu_ring_write(ring, 1); /* number of DWs to follow */ amdgpu_ring_write(ring, 0xDEADBEEF); - amdgpu_ring_unlock_commit(ring); + amdgpu_ring_commit(ring); for (i = 0; i < adev->usec_timeout; i++) { tmp = le32_to_cpu(adev->wb.wb[index]); diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c index b20eb7969e05..1d796eac421b 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c @@ -2379,7 +2379,7 @@ static int gfx_v7_0_ring_test_ring(struct amdgpu_ring *ring) return r; } WREG32(scratch, 0xCAFEDEAD); - r = amdgpu_ring_lock(ring, 3); + r = amdgpu_ring_alloc(ring, 3); if (r) { DRM_ERROR("amdgpu: cp failed to lock ring %d (%d).\n", ring->idx, r); amdgpu_gfx_scratch_free(adev, scratch); @@ -2388,7 +2388,7 @@ static int gfx_v7_0_ring_test_ring(struct amdgpu_ring *ring) amdgpu_ring_write(ring, PACKET3(PACKET3_SET_UCONFIG_REG, 1)); amdgpu_ring_write(ring, (scratch - PACKET3_SET_UCONFIG_REG_START)); amdgpu_ring_write(ring, 0xDEADBEEF); - amdgpu_ring_unlock_commit(ring); + amdgpu_ring_commit(ring); for (i = 0; i < adev->usec_timeout; i++) { tmp = RREG32(scratch); @@ -2812,7 +2812,7 @@ static int gfx_v7_0_cp_gfx_start(struct amdgpu_device *adev) gfx_v7_0_cp_gfx_enable(adev, true); - r = amdgpu_ring_lock(ring, gfx_v7_0_get_csb_size(adev) + 8); + r = amdgpu_ring_alloc(ring, gfx_v7_0_get_csb_size(adev) + 8); if (r) { DRM_ERROR("amdgpu: cp failed to lock ring (%d).\n", r); return r; @@ -2881,7 +2881,7 @@ static int gfx_v7_0_cp_gfx_start(struct amdgpu_device *adev) amdgpu_ring_write(ring, 0x0000000e); /* VGT_VERTEX_REUSE_BLOCK_CNTL */ amdgpu_ring_write(ring, 0x00000010); /* VGT_OUT_DEALLOC_CNTL */ - amdgpu_ring_unlock_commit(ring); + amdgpu_ring_commit(ring); return 0; } diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c index 248271050d0a..0d9ef524cf71 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c @@ -652,7 +652,7 @@ static int gfx_v8_0_ring_test_ring(struct amdgpu_ring *ring) return r; } WREG32(scratch, 0xCAFEDEAD); - r = amdgpu_ring_lock(ring, 3); + r = amdgpu_ring_alloc(ring, 3); if (r) { DRM_ERROR("amdgpu: cp failed to lock ring %d (%d).\n", ring->idx, r); @@ -662,7 +662,7 @@ static int gfx_v8_0_ring_test_ring(struct amdgpu_ring *ring) amdgpu_ring_write(ring, PACKET3(PACKET3_SET_UCONFIG_REG, 1)); amdgpu_ring_write(ring, (scratch - PACKET3_SET_UCONFIG_REG_START)); amdgpu_ring_write(ring, 0xDEADBEEF); - amdgpu_ring_unlock_commit(ring); + amdgpu_ring_commit(ring); for (i = 0; i < adev->usec_timeout; i++) { tmp = RREG32(scratch); @@ -3062,7 +3062,7 @@ static int gfx_v8_0_cp_gfx_start(struct amdgpu_device *adev) gfx_v8_0_cp_gfx_enable(adev, true); - r = amdgpu_ring_lock(ring, gfx_v8_0_get_csb_size(adev) + 4); + r = amdgpu_ring_alloc(ring, gfx_v8_0_get_csb_size(adev) + 4); if (r) { DRM_ERROR("amdgpu: cp failed to lock ring (%d).\n", r); return r; @@ -3126,7 +3126,7 @@ static int gfx_v8_0_cp_gfx_start(struct amdgpu_device *adev) amdgpu_ring_write(ring, 0x8000); amdgpu_ring_write(ring, 0x8000); - amdgpu_ring_unlock_commit(ring); + amdgpu_ring_commit(ring); return 0; } diff --git a/drivers/gpu/drm/amd/amdgpu/sdma_v2_4.c b/drivers/gpu/drm/amd/amdgpu/sdma_v2_4.c index 622f5dad6f2a..9fae4bf1a6c6 100644 --- a/drivers/gpu/drm/amd/amdgpu/sdma_v2_4.c +++ b/drivers/gpu/drm/amd/amdgpu/sdma_v2_4.c @@ -611,7 +611,7 @@ static int sdma_v2_4_ring_test_ring(struct amdgpu_ring *ring) tmp = 0xCAFEDEAD; adev->wb.wb[index] = cpu_to_le32(tmp); - r = amdgpu_ring_lock(ring, 5); + r = amdgpu_ring_alloc(ring, 5); if (r) { DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r); amdgpu_wb_free(adev, index); @@ -624,7 +624,7 @@ static int sdma_v2_4_ring_test_ring(struct amdgpu_ring *ring) amdgpu_ring_write(ring, upper_32_bits(gpu_addr)); amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1)); amdgpu_ring_write(ring, 0xDEADBEEF); - amdgpu_ring_unlock_commit(ring); + amdgpu_ring_commit(ring); for (i = 0; i < adev->usec_timeout; i++) { tmp = le32_to_cpu(adev->wb.wb[index]); diff --git a/drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c b/drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c index d6170e69eb34..b2fbf96dad7a 100644 --- a/drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c +++ b/drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c @@ -762,7 +762,7 @@ static int sdma_v3_0_ring_test_ring(struct amdgpu_ring *ring) tmp = 0xCAFEDEAD; adev->wb.wb[index] = cpu_to_le32(tmp); - r = amdgpu_ring_lock(ring, 5); + r = amdgpu_ring_alloc(ring, 5); if (r) { DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r); amdgpu_wb_free(adev, index); @@ -775,7 +775,7 @@ static int sdma_v3_0_ring_test_ring(struct amdgpu_ring *ring) amdgpu_ring_write(ring, upper_32_bits(gpu_addr)); amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(1)); amdgpu_ring_write(ring, 0xDEADBEEF); - amdgpu_ring_unlock_commit(ring); + amdgpu_ring_commit(ring); for (i = 0; i < adev->usec_timeout; i++) { tmp = le32_to_cpu(adev->wb.wb[index]); diff --git a/drivers/gpu/drm/amd/amdgpu/uvd_v4_2.c b/drivers/gpu/drm/amd/amdgpu/uvd_v4_2.c index c4753726a79f..e7a141c75467 100644 --- a/drivers/gpu/drm/amd/amdgpu/uvd_v4_2.c +++ b/drivers/gpu/drm/amd/amdgpu/uvd_v4_2.c @@ -164,7 +164,7 @@ static int uvd_v4_2_hw_init(void *handle) goto done; } - r = amdgpu_ring_lock(ring, 10); + r = amdgpu_ring_alloc(ring, 10); if (r) { DRM_ERROR("amdgpu: ring failed to lock UVD ring (%d).\n", r); goto done; @@ -189,7 +189,7 @@ static int uvd_v4_2_hw_init(void *handle) amdgpu_ring_write(ring, PACKET0(mmUVD_SEMA_CNTL, 0)); amdgpu_ring_write(ring, 3); - amdgpu_ring_unlock_commit(ring); + amdgpu_ring_commit(ring); done: /* lower clocks again */ @@ -453,7 +453,7 @@ static int uvd_v4_2_ring_test_ring(struct amdgpu_ring *ring) int r; WREG32(mmUVD_CONTEXT_ID, 0xCAFEDEAD); - r = amdgpu_ring_lock(ring, 3); + r = amdgpu_ring_alloc(ring, 3); if (r) { DRM_ERROR("amdgpu: cp failed to lock ring %d (%d).\n", ring->idx, r); @@ -461,7 +461,7 @@ static int uvd_v4_2_ring_test_ring(struct amdgpu_ring *ring) } amdgpu_ring_write(ring, PACKET0(mmUVD_CONTEXT_ID, 0)); amdgpu_ring_write(ring, 0xDEADBEEF); - amdgpu_ring_unlock_commit(ring); + amdgpu_ring_commit(ring); for (i = 0; i < adev->usec_timeout; i++) { tmp = RREG32(mmUVD_CONTEXT_ID); if (tmp == 0xDEADBEEF) diff --git a/drivers/gpu/drm/amd/amdgpu/uvd_v5_0.c b/drivers/gpu/drm/amd/amdgpu/uvd_v5_0.c index 774033ab2b86..3775f7756cf2 100644 --- a/drivers/gpu/drm/amd/amdgpu/uvd_v5_0.c +++ b/drivers/gpu/drm/amd/amdgpu/uvd_v5_0.c @@ -160,7 +160,7 @@ static int uvd_v5_0_hw_init(void *handle) goto done; } - r = amdgpu_ring_lock(ring, 10); + r = amdgpu_ring_alloc(ring, 10); if (r) { DRM_ERROR("amdgpu: ring failed to lock UVD ring (%d).\n", r); goto done; @@ -185,7 +185,7 @@ static int uvd_v5_0_hw_init(void *handle) amdgpu_ring_write(ring, PACKET0(mmUVD_SEMA_CNTL, 0)); amdgpu_ring_write(ring, 3); - amdgpu_ring_unlock_commit(ring); + amdgpu_ring_commit(ring); done: /* lower clocks again */ @@ -497,7 +497,7 @@ static int uvd_v5_0_ring_test_ring(struct amdgpu_ring *ring) int r; WREG32(mmUVD_CONTEXT_ID, 0xCAFEDEAD); - r = amdgpu_ring_lock(ring, 3); + r = amdgpu_ring_alloc(ring, 3); if (r) { DRM_ERROR("amdgpu: cp failed to lock ring %d (%d).\n", ring->idx, r); @@ -505,7 +505,7 @@ static int uvd_v5_0_ring_test_ring(struct amdgpu_ring *ring) } amdgpu_ring_write(ring, PACKET0(mmUVD_CONTEXT_ID, 0)); amdgpu_ring_write(ring, 0xDEADBEEF); - amdgpu_ring_unlock_commit(ring); + amdgpu_ring_commit(ring); for (i = 0; i < adev->usec_timeout; i++) { tmp = RREG32(mmUVD_CONTEXT_ID); if (tmp == 0xDEADBEEF) diff --git a/drivers/gpu/drm/amd/amdgpu/uvd_v6_0.c b/drivers/gpu/drm/amd/amdgpu/uvd_v6_0.c index eb0bbbab5b35..0b2fccad8e9a 100644 --- a/drivers/gpu/drm/amd/amdgpu/uvd_v6_0.c +++ b/drivers/gpu/drm/amd/amdgpu/uvd_v6_0.c @@ -157,7 +157,7 @@ static int uvd_v6_0_hw_init(void *handle) goto done; } - r = amdgpu_ring_lock(ring, 10); + r = amdgpu_ring_alloc(ring, 10); if (r) { DRM_ERROR("amdgpu: ring failed to lock UVD ring (%d).\n", r); goto done; @@ -182,7 +182,7 @@ static int uvd_v6_0_hw_init(void *handle) amdgpu_ring_write(ring, PACKET0(mmUVD_SEMA_CNTL, 0)); amdgpu_ring_write(ring, 3); - amdgpu_ring_unlock_commit(ring); + amdgpu_ring_commit(ring); done: if (!r) @@ -736,7 +736,7 @@ static int uvd_v6_0_ring_test_ring(struct amdgpu_ring *ring) int r; WREG32(mmUVD_CONTEXT_ID, 0xCAFEDEAD); - r = amdgpu_ring_lock(ring, 3); + r = amdgpu_ring_alloc(ring, 3); if (r) { DRM_ERROR("amdgpu: cp failed to lock ring %d (%d).\n", ring->idx, r); @@ -744,7 +744,7 @@ static int uvd_v6_0_ring_test_ring(struct amdgpu_ring *ring) } amdgpu_ring_write(ring, PACKET0(mmUVD_CONTEXT_ID, 0)); amdgpu_ring_write(ring, 0xDEADBEEF); - amdgpu_ring_unlock_commit(ring); + amdgpu_ring_commit(ring); for (i = 0; i < adev->usec_timeout; i++) { tmp = RREG32(mmUVD_CONTEXT_ID); if (tmp == 0xDEADBEEF) -- cgit v1.2.3 From 9b3896686fab2be860500ab44770071f15873aa1 Mon Sep 17 00:00:00 2001 From: Christian König Date: Thu, 11 Feb 2016 14:42:33 +0100 Subject: drm/amdgpu: remove fence reset detection leftovers MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit wait_event() never returns before the fence was signaled. Signed-off-by: Christian König Acked-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 14 ++++---------- 1 file changed, 4 insertions(+), 10 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index 72105020086c..97db196dc6f8 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -260,19 +260,16 @@ static bool amdgpu_fence_seq_signaled(struct amdgpu_ring *ring, u64 seq) } /* - * amdgpu_ring_wait_seq_timeout - wait for seq of the specific ring to signal + * amdgpu_ring_wait_seq - wait for seq of the specific ring to signal * @ring: ring to wait on for the seq number * @seq: seq number wait for * * return value: * 0: seq signaled, and gpu not hang - * -EDEADL: GPU hang detected * -EINVAL: some paramter is not valid */ static int amdgpu_fence_ring_wait_seq(struct amdgpu_ring *ring, uint64_t seq) { - bool signaled = false; - BUG_ON(!ring); if (seq > ring->fence_drv.sync_seq) return -EINVAL; @@ -281,13 +278,10 @@ static int amdgpu_fence_ring_wait_seq(struct amdgpu_ring *ring, uint64_t seq) return 0; amdgpu_fence_schedule_fallback(ring); - wait_event(ring->fence_drv.fence_queue, ( - (signaled = amdgpu_fence_seq_signaled(ring, seq)))); + wait_event(ring->fence_drv.fence_queue, + amdgpu_fence_seq_signaled(ring, seq)); - if (signaled) - return 0; - else - return -EDEADLK; + return 0; } /** -- cgit v1.2.3 From 336d1f5efe93db3d997a6d105760dd613d7ecdce Mon Sep 17 00:00:00 2001 From: Christian König Date: Tue, 16 Feb 2016 10:57:10 +0100 Subject: drm/amdgpu: remove HW fence owner MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Not used any more since we now always use the sheduler. Signed-off-by: Christian König Reviewed-by: Alex Deucher Reviewed-by: Chunming Zhou --- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 9 ++------- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 6 +----- drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c | 6 ++---- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c | 8 +------- drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c | 3 +-- drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c | 7 ++----- drivers/gpu/drm/amd/amdgpu/cik_sdma.c | 3 +-- drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c | 3 +-- drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c | 6 ++---- drivers/gpu/drm/amd/amdgpu/sdma_v2_4.c | 3 +-- drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c | 3 +-- 12 files changed, 16 insertions(+), 43 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h index 11a9f5899995..fab6ddb26b5b 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h @@ -409,9 +409,6 @@ struct amdgpu_fence { struct amdgpu_ring *ring; uint64_t seq; - /* filp or special value for fence creator */ - void *owner; - wait_queue_t fence_wake; }; @@ -432,8 +429,7 @@ int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring, unsigned irq_type); void amdgpu_fence_driver_suspend(struct amdgpu_device *adev); void amdgpu_fence_driver_resume(struct amdgpu_device *adev); -int amdgpu_fence_emit(struct amdgpu_ring *ring, void *owner, - struct amdgpu_fence **fence); +int amdgpu_fence_emit(struct amdgpu_ring *ring, struct amdgpu_fence **fence); void amdgpu_fence_process(struct amdgpu_ring *ring); int amdgpu_fence_wait_next(struct amdgpu_ring *ring); int amdgpu_fence_wait_empty(struct amdgpu_ring *ring); @@ -1177,8 +1173,7 @@ int amdgpu_ib_get(struct amdgpu_device *adev, struct amdgpu_vm *vm, unsigned size, struct amdgpu_ib *ib); void amdgpu_ib_free(struct amdgpu_device *adev, struct amdgpu_ib *ib); int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs, - struct amdgpu_ib *ib, void *owner, - struct fence *last_vm_update, + struct amdgpu_ib *ib, struct fence *last_vm_update, struct fence **f); int amdgpu_ib_pool_init(struct amdgpu_device *adev); void amdgpu_ib_pool_fini(struct amdgpu_device *adev); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index 97db196dc6f8..d94b13ac290f 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -91,25 +91,21 @@ static u32 amdgpu_fence_read(struct amdgpu_ring *ring) * amdgpu_fence_emit - emit a fence on the requested ring * * @ring: ring the fence is associated with - * @owner: creator of the fence * @fence: amdgpu fence object * * Emits a fence command on the requested ring (all asics). * Returns 0 on success, -ENOMEM on failure. */ -int amdgpu_fence_emit(struct amdgpu_ring *ring, void *owner, - struct amdgpu_fence **fence) +int amdgpu_fence_emit(struct amdgpu_ring *ring, struct amdgpu_fence **fence) { struct amdgpu_device *adev = ring->adev; - /* we are protected by the ring emission mutex */ *fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL); if ((*fence) == NULL) { return -ENOMEM; } (*fence)->seq = ++ring->fence_drv.sync_seq; (*fence)->ring = ring; - (*fence)->owner = owner; fence_init(&(*fence)->base, &amdgpu_fence_ops, &ring->fence_drv.fence_queue.lock, adev->fence_context + ring->idx, diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c index 4e978e7aa1b8..9550247b030d 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c @@ -101,7 +101,6 @@ void amdgpu_ib_free(struct amdgpu_device *adev, struct amdgpu_ib *ib) * @adev: amdgpu_device pointer * @num_ibs: number of IBs to schedule * @ibs: IB objects to schedule - * @owner: owner for creating the fences * @f: fence created during this submission * * Schedule an IB on the associated ring (all asics). @@ -118,8 +117,7 @@ void amdgpu_ib_free(struct amdgpu_device *adev, struct amdgpu_ib *ib) * to SI there was just a DE IB. */ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs, - struct amdgpu_ib *ibs, void *owner, - struct fence *last_vm_update, + struct amdgpu_ib *ibs, struct fence *last_vm_update, struct fence **f) { struct amdgpu_device *adev = ring->adev; @@ -183,7 +181,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs, amdgpu_ring_emit_hdp_invalidate(ring); } - r = amdgpu_fence_emit(ring, owner, &ib->fence); + r = amdgpu_fence_emit(ring, &ib->fence); if (r) { dev_err(adev->dev, "failed to emit fence (%d)\n", r); ring->current_ctx = old_ctx; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c index 90e52f7e17a0..692b45560d0a 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c @@ -148,7 +148,7 @@ static struct fence *amdgpu_job_run(struct amd_sched_job *sched_job) } trace_amdgpu_sched_run_job(job); - r = amdgpu_ib_schedule(job->ring, job->num_ibs, job->ibs, job->owner, + r = amdgpu_ib_schedule(job->ring, job->num_ibs, job->ibs, job->sync.last_vm_update, &fence); if (r) { DRM_ERROR("Error scheduling IBs (%d)\n", r); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c index 87690cc57206..e3673422aac8 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c @@ -60,12 +60,8 @@ void amdgpu_sync_create(struct amdgpu_sync *sync) */ 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; @@ -85,13 +81,11 @@ static bool amdgpu_sync_same_dev(struct amdgpu_device *adev, struct fence *f) */ 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; } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c index 1de82bf4fc79..e5f0a5e29551 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c @@ -886,8 +886,7 @@ static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo, ib->length_dw = 16; if (direct) { - r = amdgpu_ib_schedule(ring, 1, ib, - AMDGPU_FENCE_OWNER_UNDEFINED, NULL, &f); + r = amdgpu_ib_schedule(ring, 1, ib, NULL, &f); if (r) goto err_free; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c index 39c3aa60381a..6d191fb40b38 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c @@ -425,8 +425,7 @@ int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle, for (i = ib->length_dw; i < ib_size_dw; ++i) ib->ptr[i] = 0x0; - r = amdgpu_ib_schedule(ring, 1, ib, AMDGPU_FENCE_OWNER_UNDEFINED, - NULL, &f); + r = amdgpu_ib_schedule(ring, 1, ib, NULL, &f); if (r) goto err; @@ -487,9 +486,7 @@ int amdgpu_vce_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle, ib->ptr[i] = 0x0; if (direct) { - r = amdgpu_ib_schedule(ring, 1, ib, - AMDGPU_FENCE_OWNER_UNDEFINED, - NULL, &f); + r = amdgpu_ib_schedule(ring, 1, ib, NULL, &f); if (r) goto err; diff --git a/drivers/gpu/drm/amd/amdgpu/cik_sdma.c b/drivers/gpu/drm/amd/amdgpu/cik_sdma.c index c4101ebd2b3f..b5b4220a6141 100644 --- a/drivers/gpu/drm/amd/amdgpu/cik_sdma.c +++ b/drivers/gpu/drm/amd/amdgpu/cik_sdma.c @@ -643,8 +643,7 @@ static int cik_sdma_ring_test_ib(struct amdgpu_ring *ring) ib.ptr[3] = 1; ib.ptr[4] = 0xDEADBEEF; ib.length_dw = 5; - r = amdgpu_ib_schedule(ring, 1, &ib, AMDGPU_FENCE_OWNER_UNDEFINED, - NULL, &f); + r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f); if (r) goto err1; diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c index 1b713009c5d2..46c2436d74bd 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c @@ -2136,8 +2136,7 @@ static int gfx_v7_0_ring_test_ib(struct amdgpu_ring *ring) ib.ptr[2] = 0xDEADBEEF; ib.length_dw = 3; - r = amdgpu_ib_schedule(ring, 1, &ib, AMDGPU_FENCE_OWNER_UNDEFINED, - NULL, &f); + r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f); if (r) goto err2; diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c index aeb35f504999..e0b64de9b5af 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c @@ -706,8 +706,7 @@ static int gfx_v8_0_ring_test_ib(struct amdgpu_ring *ring) ib.ptr[2] = 0xDEADBEEF; ib.length_dw = 3; - r = amdgpu_ib_schedule(ring, 1, &ib, AMDGPU_FENCE_OWNER_UNDEFINED, - NULL, &f); + r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f); if (r) goto err2; @@ -1262,8 +1261,7 @@ static int gfx_v8_0_do_edc_gpr_workarounds(struct amdgpu_device *adev) ib.ptr[ib.length_dw++] = EVENT_TYPE(7) | EVENT_INDEX(4); /* shedule the ib on the ring */ - r = amdgpu_ib_schedule(ring, 1, &ib, AMDGPU_FENCE_OWNER_UNDEFINED, - NULL, &f); + r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f); if (r) { DRM_ERROR("amdgpu: ib submit failed (%d).\n", r); goto fail; diff --git a/drivers/gpu/drm/amd/amdgpu/sdma_v2_4.c b/drivers/gpu/drm/amd/amdgpu/sdma_v2_4.c index 134d5d0bb77f..747ef558033a 100644 --- a/drivers/gpu/drm/amd/amdgpu/sdma_v2_4.c +++ b/drivers/gpu/drm/amd/amdgpu/sdma_v2_4.c @@ -701,8 +701,7 @@ static int sdma_v2_4_ring_test_ib(struct amdgpu_ring *ring) ib.ptr[7] = SDMA_PKT_HEADER_OP(SDMA_OP_NOP); ib.length_dw = 8; - r = amdgpu_ib_schedule(ring, 1, &ib, AMDGPU_FENCE_OWNER_UNDEFINED, - NULL, &f); + r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f); if (r) goto err1; diff --git a/drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c b/drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c index ca12d6faa8c2..11a544fbf197 100644 --- a/drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c +++ b/drivers/gpu/drm/amd/amdgpu/sdma_v3_0.c @@ -853,8 +853,7 @@ static int sdma_v3_0_ring_test_ib(struct amdgpu_ring *ring) ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP); ib.length_dw = 8; - r = amdgpu_ib_schedule(ring, 1, &ib, AMDGPU_FENCE_OWNER_UNDEFINED, - NULL, &f); + r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f); if (r) goto err1; -- cgit v1.2.3 From 364beb2cc45247e980a097e53d0932e143873333 Mon Sep 17 00:00:00 2001 From: Christian König Date: Tue, 16 Feb 2016 17:39:39 +0100 Subject: drm/amdgpu: return the common fence from amdgpu_fence_emit MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Try to avoid using the hardware specific fences even more. Signed-off-by: Christian König Reviewed-by: Alex Deucher Reviewed-by: Chunming Zhou --- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 4 ++-- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 27 ++++++++++++++------------- drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c | 7 +++---- 3 files changed, 19 insertions(+), 19 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h index 3e4ec56919c7..128eba604f97 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h @@ -429,7 +429,7 @@ int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring, unsigned irq_type); void amdgpu_fence_driver_suspend(struct amdgpu_device *adev); void amdgpu_fence_driver_resume(struct amdgpu_device *adev); -int amdgpu_fence_emit(struct amdgpu_ring *ring, struct amdgpu_fence **fence); +int amdgpu_fence_emit(struct amdgpu_ring *ring, struct fence **fence); void amdgpu_fence_process(struct amdgpu_ring *ring); int amdgpu_fence_wait_next(struct amdgpu_ring *ring); int amdgpu_fence_wait_empty(struct amdgpu_ring *ring); @@ -766,7 +766,7 @@ struct amdgpu_ib { uint32_t length_dw; uint64_t gpu_addr; uint32_t *ptr; - struct amdgpu_fence *fence; + struct fence *fence; struct amdgpu_user_fence *user; struct amdgpu_vm *vm; unsigned vm_id; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index d94b13ac290f..83599f2a0387 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -91,28 +91,29 @@ static u32 amdgpu_fence_read(struct amdgpu_ring *ring) * amdgpu_fence_emit - emit a fence on the requested ring * * @ring: ring the fence is associated with - * @fence: amdgpu fence object + * @f: resulting fence object * * Emits a fence command on the requested ring (all asics). * Returns 0 on success, -ENOMEM on failure. */ -int amdgpu_fence_emit(struct amdgpu_ring *ring, struct amdgpu_fence **fence) +int amdgpu_fence_emit(struct amdgpu_ring *ring, struct fence **f) { struct amdgpu_device *adev = ring->adev; + struct amdgpu_fence *fence; - *fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL); - if ((*fence) == NULL) { + fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL); + if (fence == NULL) return -ENOMEM; - } - (*fence)->seq = ++ring->fence_drv.sync_seq; - (*fence)->ring = ring; - fence_init(&(*fence)->base, &amdgpu_fence_ops, - &ring->fence_drv.fence_queue.lock, - adev->fence_context + ring->idx, - (*fence)->seq); + + fence->seq = ++ring->fence_drv.sync_seq; + fence->ring = ring; + fence_init(&fence->base, &amdgpu_fence_ops, + &ring->fence_drv.fence_queue.lock, + adev->fence_context + ring->idx, + fence->seq); amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr, - (*fence)->seq, - AMDGPU_FENCE_FLAG_INT); + fence->seq, AMDGPU_FENCE_FLAG_INT); + *f = &fence->base; return 0; } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c index 9550247b030d..979c445f8096 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c @@ -90,9 +90,8 @@ int amdgpu_ib_get(struct amdgpu_device *adev, struct amdgpu_vm *vm, */ void amdgpu_ib_free(struct amdgpu_device *adev, struct amdgpu_ib *ib) { - amdgpu_sa_bo_free(adev, &ib->sa_bo, &ib->fence->base); - if (ib->fence) - fence_put(&ib->fence->base); + amdgpu_sa_bo_free(adev, &ib->sa_bo, ib->fence); + fence_put(ib->fence); } /** @@ -198,7 +197,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs, } if (f) - *f = fence_get(&ib->fence->base); + *f = fence_get(ib->fence); amdgpu_ring_commit(ring); return 0; -- cgit v1.2.3 From 77163f074aa68b06f6de2f2fc271f4a828d56da2 Mon Sep 17 00:00:00 2001 From: Christian König Date: Fri, 11 Mar 2016 15:15:02 +0100 Subject: drm/amdgpu: remove amdgpu_fence_wait_next MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Not used any more. Signed-off-by: Christian König Acked-by: Alex Deucher Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 1 - drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 20 -------------------- 2 files changed, 21 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h index 1059c04fd9df..d4411f60d1ce 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h @@ -432,7 +432,6 @@ void amdgpu_fence_driver_suspend(struct amdgpu_device *adev); void amdgpu_fence_driver_resume(struct amdgpu_device *adev); int amdgpu_fence_emit(struct amdgpu_ring *ring, struct fence **fence); void amdgpu_fence_process(struct amdgpu_ring *ring); -int amdgpu_fence_wait_next(struct amdgpu_ring *ring); int amdgpu_fence_wait_empty(struct amdgpu_ring *ring); unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index 83599f2a0387..fabb01e8c8d5 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -281,26 +281,6 @@ static int amdgpu_fence_ring_wait_seq(struct amdgpu_ring *ring, uint64_t seq) return 0; } -/** - * amdgpu_fence_wait_next - wait for the next fence to signal - * - * @adev: amdgpu device pointer - * @ring: ring index the fence is associated with - * - * Wait for the next fence on the requested ring to signal (all asics). - * Returns 0 if the next fence has passed, error for all other cases. - * Caller must hold ring lock. - */ -int amdgpu_fence_wait_next(struct amdgpu_ring *ring) -{ - uint64_t seq = atomic64_read(&ring->fence_drv.last_seq) + 1ULL; - - if (seq >= ring->fence_drv.sync_seq) - return -ENOENT; - - return amdgpu_fence_ring_wait_seq(ring, seq); -} - /** * amdgpu_fence_wait_empty - wait for all fences to signal * -- cgit v1.2.3 From 22e5a2f46a26adc6822c54af946b384e14930417 Mon Sep 17 00:00:00 2001 From: Christian König Date: Fri, 11 Mar 2016 15:12:53 +0100 Subject: drm/amdgpu: move fence structure into amdgpu_fence.c MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit No need to have that in the header file any more. Signed-off-by: Christian König Reviewed-by: Alex Deucher Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 25 ------------------------- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 26 +++++++++++++++++++++++++- 2 files changed, 25 insertions(+), 26 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h index d4411f60d1ce..7dee9af33196 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h @@ -180,7 +180,6 @@ extern int amdgpu_powerplay; #define CIK_CURSOR_HEIGHT 128 struct amdgpu_device; -struct amdgpu_fence; struct amdgpu_ib; struct amdgpu_vm; struct amdgpu_ring; @@ -403,16 +402,6 @@ struct amdgpu_fence_driver { #define AMDGPU_FENCE_FLAG_64BIT (1 << 0) #define AMDGPU_FENCE_FLAG_INT (1 << 1) -struct amdgpu_fence { - struct fence base; - - /* RB, DMA, etc. */ - struct amdgpu_ring *ring; - uint64_t seq; - - wait_queue_t fence_wake; -}; - struct amdgpu_user_fence { /* write-back bo */ struct amdgpu_bo *bo; @@ -2108,20 +2097,6 @@ void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v); u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index); void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v); -/* - * Cast helper - */ -extern const struct fence_ops amdgpu_fence_ops; -static inline struct amdgpu_fence *to_amdgpu_fence(struct fence *f) -{ - struct amdgpu_fence *__f = container_of(f, struct amdgpu_fence, base); - - if (__f->base.ops == &amdgpu_fence_ops) - return __f; - - return NULL; -} - /* * Registers read & write functions. */ diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index fabb01e8c8d5..3db18f42c5c9 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -47,9 +47,33 @@ * that the the relevant GPU caches have been flushed. */ +struct amdgpu_fence { + struct fence base; + + /* RB, DMA, etc. */ + struct amdgpu_ring *ring; + uint64_t seq; + + wait_queue_t fence_wake; +}; + static struct kmem_cache *amdgpu_fence_slab; static atomic_t amdgpu_fence_slab_ref = ATOMIC_INIT(0); +/* + * Cast helper + */ +static const struct fence_ops amdgpu_fence_ops; +static inline struct amdgpu_fence *to_amdgpu_fence(struct fence *f) +{ + struct amdgpu_fence *__f = container_of(f, struct amdgpu_fence, base); + + if (__f->base.ops == &amdgpu_fence_ops) + return __f; + + return NULL; +} + /** * amdgpu_fence_write - write a fence value * @@ -663,7 +687,7 @@ static void amdgpu_fence_release(struct fence *f) kmem_cache_free(amdgpu_fence_slab, fence); } -const struct fence_ops amdgpu_fence_ops = { +static const struct fence_ops amdgpu_fence_ops = { .get_driver_name = amdgpu_fence_get_driver_name, .get_timeline_name = amdgpu_fence_get_timeline_name, .enable_signaling = amdgpu_fence_enable_signaling, -- cgit v1.2.3 From d9713ef6b99e43b93bf8e73fc210d2925322eacc Mon Sep 17 00:00:00 2001 From: Christian König Date: Fri, 11 Mar 2016 17:49:58 +0100 Subject: drm/amdgpu: cleanup amdgpu_fence_activity MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The comment about the loop counter was never valid, even when you have multiple threads this loop only runs as long as the sequence increases. Signed-off-by: Christian König Reviewed-by: Chunming Zhou --- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 35 +++---------------------------- 1 file changed, 3 insertions(+), 32 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index 3db18f42c5c9..35fbc8874514 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -166,30 +166,8 @@ static void amdgpu_fence_schedule_fallback(struct amdgpu_ring *ring) static bool amdgpu_fence_activity(struct amdgpu_ring *ring) { uint64_t seq, last_seq, last_emitted; - unsigned count_loop = 0; bool wake = false; - /* Note there is a scenario here for an infinite loop but it's - * very unlikely to happen. For it to happen, the current polling - * process need to be interrupted by another process and another - * process needs to update the last_seq btw the atomic read and - * xchg of the current process. - * - * More over for this to go in infinite loop there need to be - * continuously new fence signaled ie amdgpu_fence_read needs - * to return a different value each time for both the currently - * polling process and the other process that xchg the last_seq - * btw atomic read and xchg of the current process. And the - * value the other process set as last seq must be higher than - * the seq value we just read. Which means that current process - * need to be interrupted after amdgpu_fence_read and before - * atomic xchg. - * - * To be even more safe we count the number of time we loop and - * we bail after 10 loop just accepting the fact that we might - * have temporarly set the last_seq not to the true real last - * seq but to an older one. - */ last_seq = atomic64_read(&ring->fence_drv.last_seq); do { last_emitted = ring->fence_drv.sync_seq; @@ -200,23 +178,16 @@ static bool amdgpu_fence_activity(struct amdgpu_ring *ring) seq |= last_emitted & 0xffffffff00000000LL; } - if (seq <= last_seq || seq > last_emitted) { + if (seq <= last_seq || seq > last_emitted) break; - } + /* If we loop over we don't want to return without * checking if a fence is signaled as it means that the * seq we just read is different from the previous on. */ wake = true; last_seq = seq; - if ((count_loop++) > 10) { - /* We looped over too many time leave with the - * fact that we might have set an older fence - * seq then the current real last seq as signaled - * by the hw. - */ - break; - } + } while (atomic64_xchg(&ring->fence_drv.last_seq, seq) > seq); if (seq < last_emitted) -- cgit v1.2.3 From ca08e04d5642c429ec891fa17bf379be988dea6b Mon Sep 17 00:00:00 2001 From: Christian König Date: Fri, 11 Mar 2016 17:57:56 +0100 Subject: drm/amdgpu: merge amdgpu_fence_process and _activity MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit No need to keep the two separate any more. Signed-off-by: Christian König Reviewed-by: Chunming Zhou --- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 24 +++++------------------- 1 file changed, 5 insertions(+), 19 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index 35fbc8874514..d8bbe85c22a2 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -155,15 +155,15 @@ static void amdgpu_fence_schedule_fallback(struct amdgpu_ring *ring) } /** - * amdgpu_fence_activity - check for fence activity + * amdgpu_fence_process - check for fence activity * * @ring: pointer to struct amdgpu_ring * * Checks the current fence value and calculates the last - * signalled fence value. Returns true if activity occured - * on the ring, and the fence_queue should be waken up. + * signalled fence value. Wakes the fence queue if the + * sequence number has increased. */ -static bool amdgpu_fence_activity(struct amdgpu_ring *ring) +void amdgpu_fence_process(struct amdgpu_ring *ring) { uint64_t seq, last_seq, last_emitted; bool wake = false; @@ -193,21 +193,7 @@ static bool amdgpu_fence_activity(struct amdgpu_ring *ring) if (seq < last_emitted) amdgpu_fence_schedule_fallback(ring); - return wake; -} - -/** - * amdgpu_fence_process - process a fence - * - * @adev: amdgpu_device pointer - * @ring: ring index the fence is associated with - * - * Checks the current fence value and wakes the fence queue - * if the sequence number has increased (all asics). - */ -void amdgpu_fence_process(struct amdgpu_ring *ring) -{ - if (amdgpu_fence_activity(ring)) + if (wake) wake_up_all(&ring->fence_drv.fence_queue); } -- cgit v1.2.3 From b44135351a3a5cfc60f5b6729445311c2d5e141c Mon Sep 17 00:00:00 2001 From: Christian König Date: Tue, 15 Mar 2016 13:40:17 +0100 Subject: drm/amdgpu: RCU protected amdgpu_fence_release MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Fences must be freed RCU protected, otherwise the reservation_object_*_rcu() functions can run into problems. Signed-off-by: Christian König Reviewed-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 23 ++++++++++++++++++++++- 1 file changed, 22 insertions(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index d8bbe85c22a2..ec73ab1c177d 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -638,12 +638,33 @@ static bool amdgpu_fence_enable_signaling(struct fence *f) return true; } -static void amdgpu_fence_release(struct fence *f) +/** + * amdgpu_fence_free - free up the fence memory + * + * @rcu: RCU callback head + * + * Free up the fence memory after the RCU grace period. + */ +static void amdgpu_fence_free(struct rcu_head *rcu) { + struct fence *f = container_of(rcu, struct fence, rcu); struct amdgpu_fence *fence = to_amdgpu_fence(f); kmem_cache_free(amdgpu_fence_slab, fence); } +/** + * amdgpu_fence_release - callback that fence can be freed + * + * @fence: fence + * + * This function is called when the reference count becomes zero. + * It just RCU schedules freeing up the fence. + */ +static void amdgpu_fence_release(struct fence *f) +{ + call_rcu(&f->rcu, amdgpu_fence_free); +} + static const struct fence_ops amdgpu_fence_ops = { .get_driver_name = amdgpu_fence_get_driver_name, .get_timeline_name = amdgpu_fence_get_timeline_name, -- cgit v1.2.3 From e6151a08bbb3c85cd0b23813432690939e143131 Mon Sep 17 00:00:00 2001 From: Christian König Date: Tue, 15 Mar 2016 14:52:26 +0100 Subject: drm/amdgpu: add number of hardware submissions to amdgpu_fence_driver_init_ring MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Make this a parameter instead of using the global variable directly. Signed-off-by: Christian König Acked-by: Alex Deucher Reviewed-by: Chunming Zhou --- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 3 ++- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 10 ++++++++-- drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c | 3 ++- 3 files changed, 12 insertions(+), 4 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h index 7dee9af33196..26b137801aea 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h @@ -413,7 +413,8 @@ 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); -int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring); +int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring, + unsigned num_hw_submission); int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring, struct amdgpu_irq_src *irq_src, unsigned irq_type); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index ec73ab1c177d..4654113ecfae 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -355,15 +355,21 @@ int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring, * for the requested ring. * * @ring: ring to init the fence driver on + * @num_hw_submission: number of entries on the hardware queue * * Init the fence driver for the requested ring (all asics). * Helper function for amdgpu_fence_driver_init(). */ -int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring) +int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring, + unsigned num_hw_submission) { long timeout; int r; + /* Check that num_hw_submission is a power of two */ + if ((num_hw_submission & (num_hw_submission - 1)) != 0) + return -EINVAL; + ring->fence_drv.cpu_addr = NULL; ring->fence_drv.gpu_addr = 0; ring->fence_drv.sync_seq = 0; @@ -387,7 +393,7 @@ int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring) timeout = MAX_SCHEDULE_TIMEOUT; } r = amd_sched_init(&ring->sched, &amdgpu_sched_ops, - amdgpu_sched_hw_submission, + num_hw_submission, timeout, ring->name); if (r) { DRM_ERROR("Failed to create scheduler on ring %s.\n", diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c index cc3c7adb994c..972eed2ef787 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c @@ -236,7 +236,8 @@ 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; - r = amdgpu_fence_driver_init_ring(ring); + r = amdgpu_fence_driver_init_ring(ring, + amdgpu_sched_hw_submission); if (r) return r; } -- cgit v1.2.3 From c89377d10a11e5d8be11525f220dc624574c1aa5 Mon Sep 17 00:00:00 2001 From: Christian König Date: Sun, 13 Mar 2016 19:19:48 +0100 Subject: drm/amdgpu: keep all fences in an RCU protected array v2 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Just keep all HW fences in a RCU protected array as a first step to replace the wait queue. v2: update commit message, move fixes into separate patch. Signed-off-by: Christian König Acked-by: Alex Deucher Reviewed-by: Chunming Zhou --- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 2 ++ drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 32 ++++++++++++++++++++++++++++--- 2 files changed, 31 insertions(+), 3 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h index 26b137801aea..2fea597f8b6c 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h @@ -393,6 +393,8 @@ struct amdgpu_fence_driver { unsigned irq_type; struct timer_list fallback_timer; wait_queue_head_t fence_queue; + unsigned num_fences_mask; + struct fence **fences; }; /* some special values for the owner field */ diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index 4654113ecfae..44eac91163eb 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -124,6 +124,8 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct fence **f) { struct amdgpu_device *adev = ring->adev; struct amdgpu_fence *fence; + struct fence *old, **ptr; + unsigned idx; fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL); if (fence == NULL) @@ -137,7 +139,21 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct fence **f) fence->seq); amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr, fence->seq, AMDGPU_FENCE_FLAG_INT); + + idx = fence->seq & ring->fence_drv.num_fences_mask; + ptr = &ring->fence_drv.fences[idx]; + /* This function can't be called concurrently anyway, otherwise + * emitting the fence would mess up the hardware ring buffer. + */ + old = rcu_dereference_protected(*ptr, 1); + + rcu_assign_pointer(*ptr, fence_get(&fence->base)); + + BUG_ON(old && !fence_is_signaled(old)); + fence_put(old); + *f = &fence->base; + return 0; } @@ -380,6 +396,11 @@ int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring, (unsigned long)ring); init_waitqueue_head(&ring->fence_drv.fence_queue); + ring->fence_drv.num_fences_mask = num_hw_submission - 1; + ring->fence_drv.fences = kcalloc(num_hw_submission, sizeof(void *), + GFP_KERNEL); + if (!ring->fence_drv.fences) + return -ENOMEM; timeout = msecs_to_jiffies(amdgpu_lockup_timeout); if (timeout == 0) { @@ -441,10 +462,9 @@ int amdgpu_fence_driver_init(struct amdgpu_device *adev) */ void amdgpu_fence_driver_fini(struct amdgpu_device *adev) { - int i, r; + unsigned i, j; + int r; - if (atomic_dec_and_test(&amdgpu_fence_slab_ref)) - kmem_cache_destroy(amdgpu_fence_slab); for (i = 0; i < AMDGPU_MAX_RINGS; i++) { struct amdgpu_ring *ring = adev->rings[i]; @@ -460,8 +480,14 @@ void amdgpu_fence_driver_fini(struct amdgpu_device *adev) ring->fence_drv.irq_type); amd_sched_fini(&ring->sched); del_timer_sync(&ring->fence_drv.fallback_timer); + for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j) + fence_put(ring->fence_drv.fences[i]); + kfree(ring->fence_drv.fences); ring->fence_drv.initialized = false; } + + if (atomic_dec_and_test(&amdgpu_fence_slab_ref)) + kmem_cache_destroy(amdgpu_fence_slab); } /** -- cgit v1.2.3 From f09c2be4d4804963f18417abd5b51bc8a8330851 Mon Sep 17 00:00:00 2001 From: Christian König Date: Sun, 13 Mar 2016 19:37:01 +0100 Subject: drm/amdgpu: cleanup amdgpu_fence_wait_empty v2 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Just wait for last fence instead of waiting for the sequence manually. v2: don't use amdgpu_sched_jobs for the mask Signed-off-by: Christian König Acked-by: Alex Deucher Reviewed-by: Chunming Zhou --- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 69 +++++++------------------------ 1 file changed, 15 insertions(+), 54 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index 44eac91163eb..d5bdd9633c85 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -227,57 +227,6 @@ static void amdgpu_fence_fallback(unsigned long arg) amdgpu_fence_process(ring); } -/** - * amdgpu_fence_seq_signaled - check if a fence sequence number has signaled - * - * @ring: ring the fence is associated with - * @seq: sequence number - * - * Check if the last signaled fence sequnce number is >= the requested - * sequence number (all asics). - * Returns true if the fence has signaled (current fence value - * is >= requested value) or false if it has not (current fence - * value is < the requested value. Helper function for - * amdgpu_fence_signaled(). - */ -static bool amdgpu_fence_seq_signaled(struct amdgpu_ring *ring, u64 seq) -{ - if (atomic64_read(&ring->fence_drv.last_seq) >= seq) - return true; - - /* poll new last sequence at least once */ - amdgpu_fence_process(ring); - if (atomic64_read(&ring->fence_drv.last_seq) >= seq) - return true; - - return false; -} - -/* - * amdgpu_ring_wait_seq - wait for seq of the specific ring to signal - * @ring: ring to wait on for the seq number - * @seq: seq number wait for - * - * return value: - * 0: seq signaled, and gpu not hang - * -EINVAL: some paramter is not valid - */ -static int amdgpu_fence_ring_wait_seq(struct amdgpu_ring *ring, uint64_t seq) -{ - BUG_ON(!ring); - if (seq > ring->fence_drv.sync_seq) - return -EINVAL; - - if (atomic64_read(&ring->fence_drv.last_seq) >= seq) - return 0; - - amdgpu_fence_schedule_fallback(ring); - wait_event(ring->fence_drv.fence_queue, - amdgpu_fence_seq_signaled(ring, seq)); - - return 0; -} - /** * amdgpu_fence_wait_empty - wait for all fences to signal * @@ -286,16 +235,28 @@ static int amdgpu_fence_ring_wait_seq(struct amdgpu_ring *ring, uint64_t seq) * * Wait for all fences on the requested ring to signal (all asics). * Returns 0 if the fences have passed, error for all other cases. - * Caller must hold ring lock. */ int amdgpu_fence_wait_empty(struct amdgpu_ring *ring) { - uint64_t seq = ring->fence_drv.sync_seq; + uint64_t seq = ACCESS_ONCE(ring->fence_drv.sync_seq); + struct fence *fence, **ptr; + int r; if (!seq) return 0; - return amdgpu_fence_ring_wait_seq(ring, seq); + ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask]; + rcu_read_lock(); + fence = rcu_dereference(*ptr); + if (!fence || !fence_get_rcu(fence)) { + rcu_read_unlock(); + return 0; + } + rcu_read_unlock(); + + r = fence_wait(fence, false); + fence_put(fence); + return r; } /** -- cgit v1.2.3 From 4a7d74f1763edd96badba1fe1c99f07ddf0376a6 Mon Sep 17 00:00:00 2001 From: Christian König Date: Mon, 14 Mar 2016 14:29:46 +0100 Subject: drm/amdgpu: signal fences directly in amdgpu_fence_process MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Because of the scheduler we need to signal all fences immediately anyway, so try to avoid the waitqueue overhead. Signed-off-by: Christian König Acked-by: Alex Deucher Reviewed-by: Chunming Zhou --- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 97 ++++++++++--------------------- 2 files changed, 31 insertions(+), 68 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h index 2fea597f8b6c..c91156c7bda1 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h @@ -392,8 +392,8 @@ struct amdgpu_fence_driver { struct amdgpu_irq_src *irq_src; unsigned irq_type; struct timer_list fallback_timer; - wait_queue_head_t fence_queue; unsigned num_fences_mask; + spinlock_t lock; struct fence **fences; }; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index d5bdd9633c85..c5980c4133a6 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -53,8 +53,6 @@ struct amdgpu_fence { /* RB, DMA, etc. */ struct amdgpu_ring *ring; uint64_t seq; - - wait_queue_t fence_wake; }; static struct kmem_cache *amdgpu_fence_slab; @@ -124,7 +122,7 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct fence **f) { struct amdgpu_device *adev = ring->adev; struct amdgpu_fence *fence; - struct fence *old, **ptr; + struct fence **ptr; unsigned idx; fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL); @@ -134,7 +132,7 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct fence **f) fence->seq = ++ring->fence_drv.sync_seq; fence->ring = ring; fence_init(&fence->base, &amdgpu_fence_ops, - &ring->fence_drv.fence_queue.lock, + &ring->fence_drv.lock, adev->fence_context + ring->idx, fence->seq); amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr, @@ -145,13 +143,10 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct fence **f) /* This function can't be called concurrently anyway, otherwise * emitting the fence would mess up the hardware ring buffer. */ - old = rcu_dereference_protected(*ptr, 1); + BUG_ON(rcu_dereference_protected(*ptr, 1)); rcu_assign_pointer(*ptr, fence_get(&fence->base)); - BUG_ON(old && !fence_is_signaled(old)); - fence_put(old); - *f = &fence->base; return 0; @@ -181,11 +176,12 @@ static void amdgpu_fence_schedule_fallback(struct amdgpu_ring *ring) */ void amdgpu_fence_process(struct amdgpu_ring *ring) { + struct amdgpu_fence_driver *drv = &ring->fence_drv; uint64_t seq, last_seq, last_emitted; - bool wake = false; + int r; - last_seq = atomic64_read(&ring->fence_drv.last_seq); do { + last_seq = atomic64_read(&ring->fence_drv.last_seq); last_emitted = ring->fence_drv.sync_seq; seq = amdgpu_fence_read(ring); seq |= last_seq & 0xffffffff00000000LL; @@ -195,22 +191,32 @@ void amdgpu_fence_process(struct amdgpu_ring *ring) } if (seq <= last_seq || seq > last_emitted) - break; + return; - /* If we loop over we don't want to return without - * checking if a fence is signaled as it means that the - * seq we just read is different from the previous on. - */ - wake = true; - last_seq = seq; - - } while (atomic64_xchg(&ring->fence_drv.last_seq, seq) > seq); + } while (atomic64_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq); if (seq < last_emitted) amdgpu_fence_schedule_fallback(ring); - if (wake) - wake_up_all(&ring->fence_drv.fence_queue); + while (last_seq != seq) { + struct fence *fence, **ptr; + + ptr = &drv->fences[++last_seq & drv->num_fences_mask]; + + /* There is always exactly one thread signaling this fence slot */ + fence = rcu_dereference_protected(*ptr, 1); + rcu_assign_pointer(*ptr, NULL); + + BUG_ON(!fence); + + r = fence_signal(fence); + if (!r) + FENCE_TRACE(fence, "signaled from irq context\n"); + else + BUG(); + + fence_put(fence); + } } /** @@ -356,8 +362,8 @@ int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring, setup_timer(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback, (unsigned long)ring); - init_waitqueue_head(&ring->fence_drv.fence_queue); ring->fence_drv.num_fences_mask = num_hw_submission - 1; + spin_lock_init(&ring->fence_drv.lock); ring->fence_drv.fences = kcalloc(num_hw_submission, sizeof(void *), GFP_KERNEL); if (!ring->fence_drv.fences) @@ -436,7 +442,6 @@ void amdgpu_fence_driver_fini(struct amdgpu_device *adev) /* no need to trigger GPU reset as we are unloading */ amdgpu_fence_driver_force_completion(adev); } - wake_up_all(&ring->fence_drv.fence_queue); amdgpu_irq_put(adev, ring->fence_drv.irq_src, ring->fence_drv.irq_type); amd_sched_fini(&ring->sched); @@ -568,42 +573,6 @@ static bool amdgpu_fence_is_signaled(struct fence *f) return false; } -/** - * amdgpu_fence_check_signaled - callback from fence_queue - * - * this function is called with fence_queue lock held, which is also used - * for the fence locking itself, so unlocked variants are used for - * fence_signal, and remove_wait_queue. - */ -static int amdgpu_fence_check_signaled(wait_queue_t *wait, unsigned mode, int flags, void *key) -{ - struct amdgpu_fence *fence; - struct amdgpu_device *adev; - u64 seq; - int ret; - - fence = container_of(wait, struct amdgpu_fence, fence_wake); - adev = fence->ring->adev; - - /* - * We cannot use amdgpu_fence_process here because we're already - * in the waitqueue, in a call from wake_up_all. - */ - seq = atomic64_read(&fence->ring->fence_drv.last_seq); - if (seq >= fence->seq) { - ret = fence_signal_locked(&fence->base); - if (!ret) - FENCE_TRACE(&fence->base, "signaled from irq context\n"); - else - FENCE_TRACE(&fence->base, "was already signaled\n"); - - __remove_wait_queue(&fence->ring->fence_drv.fence_queue, &fence->fence_wake); - fence_put(&fence->base); - } else - FENCE_TRACE(&fence->base, "pending\n"); - return 0; -} - /** * amdgpu_fence_enable_signaling - enable signalling on fence * @fence: fence @@ -617,17 +586,11 @@ static bool amdgpu_fence_enable_signaling(struct fence *f) struct amdgpu_fence *fence = to_amdgpu_fence(f); struct amdgpu_ring *ring = fence->ring; - if (atomic64_read(&ring->fence_drv.last_seq) >= fence->seq) - return false; - - fence->fence_wake.flags = 0; - fence->fence_wake.private = NULL; - fence->fence_wake.func = amdgpu_fence_check_signaled; - __add_wait_queue(&ring->fence_drv.fence_queue, &fence->fence_wake); - fence_get(f); if (!timer_pending(&ring->fence_drv.fallback_timer)) amdgpu_fence_schedule_fallback(ring); + FENCE_TRACE(&fence->base, "armed on ring %i!\n", ring->idx); + return true; } -- cgit v1.2.3 From 91cc6418a0ea14633bdc1d2ea91d174fc1e9187d Mon Sep 17 00:00:00 2001 From: Christian König Date: Mon, 14 Mar 2016 14:49:33 +0100 Subject: drm/amdgpu: drop the extra fence range check v2 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Amdgpu doesn't support using scratch registers for fences any more. So we won't see values like 0xdeadbeef as fence value any more. v2: reschedule timer even if no change detected Signed-off-by: Christian König Acked-by: Alex Deucher Reviewed-by: Chunming Zhou --- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 3 --- 1 file changed, 3 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index c5980c4133a6..fa4eabeee71d 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -190,9 +190,6 @@ void amdgpu_fence_process(struct amdgpu_ring *ring) seq |= last_emitted & 0xffffffff00000000LL; } - if (seq <= last_seq || seq > last_emitted) - return; - } while (atomic64_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq); if (seq < last_emitted) -- cgit v1.2.3 From 480d0bf07ecf27da0a5f12774d9642072c364fa9 Mon Sep 17 00:00:00 2001 From: Christian König Date: Mon, 14 Mar 2016 15:23:11 +0100 Subject: drm/amdgpu: remove amdgpu_fence_is_signaled MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit It's just overhead to check the fence value when we signal them directly anyway. Signed-off-by: Christian König Acked-by: Alex Deucher Reviewed-by: Chunming Zhou --- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 25 ------------------------- 1 file changed, 25 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index fa4eabeee71d..da9a155a622c 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -546,30 +546,6 @@ static const char *amdgpu_fence_get_timeline_name(struct fence *f) return (const char *)fence->ring->name; } -/** - * amdgpu_fence_is_signaled - test if fence is signaled - * - * @f: fence to test - * - * Test the fence sequence number if it is already signaled. If it isn't - * signaled start fence processing. Returns True if the fence is signaled. - */ -static bool amdgpu_fence_is_signaled(struct fence *f) -{ - struct amdgpu_fence *fence = to_amdgpu_fence(f); - struct amdgpu_ring *ring = fence->ring; - - if (atomic64_read(&ring->fence_drv.last_seq) >= fence->seq) - return true; - - amdgpu_fence_process(ring); - - if (atomic64_read(&ring->fence_drv.last_seq) >= fence->seq) - return true; - - return false; -} - /** * amdgpu_fence_enable_signaling - enable signalling on fence * @fence: fence @@ -622,7 +598,6 @@ static const struct fence_ops amdgpu_fence_ops = { .get_driver_name = amdgpu_fence_get_driver_name, .get_timeline_name = amdgpu_fence_get_timeline_name, .enable_signaling = amdgpu_fence_enable_signaling, - .signaled = amdgpu_fence_is_signaled, .wait = fence_default_wait, .release = amdgpu_fence_release, }; -- cgit v1.2.3 From 742c085fa86345ae9de259c7f15c652322da4e83 Mon Sep 17 00:00:00 2001 From: Christian König Date: Mon, 14 Mar 2016 15:46:06 +0100 Subject: drm/amdgpu: switch back to 32bit hw fences v2 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit We don't need to extend them to 64bits any more, so avoid the extra overhead. v2: update commit message. Signed-off-by: Christian König Acked-by: Alex Deucher Reviewed-by: Chunming Zhou --- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 4 +-- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 49 ++++++++++++------------------- 2 files changed, 21 insertions(+), 32 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h index c91156c7bda1..31fd7ee62e69 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h @@ -386,8 +386,8 @@ struct amdgpu_fence_driver { uint64_t gpu_addr; volatile uint32_t *cpu_addr; /* sync_seq is protected by ring emission lock */ - uint64_t sync_seq; - atomic64_t last_seq; + uint32_t sync_seq; + atomic_t last_seq; bool initialized; struct amdgpu_irq_src *irq_src; unsigned irq_type; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index da9a155a622c..4303b447efe8 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -52,7 +52,6 @@ struct amdgpu_fence { /* RB, DMA, etc. */ struct amdgpu_ring *ring; - uint64_t seq; }; static struct kmem_cache *amdgpu_fence_slab; @@ -104,7 +103,7 @@ static u32 amdgpu_fence_read(struct amdgpu_ring *ring) if (drv->cpu_addr) seq = le32_to_cpu(*drv->cpu_addr); else - seq = lower_32_bits(atomic64_read(&drv->last_seq)); + seq = atomic_read(&drv->last_seq); return seq; } @@ -123,23 +122,22 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct fence **f) struct amdgpu_device *adev = ring->adev; struct amdgpu_fence *fence; struct fence **ptr; - unsigned idx; + uint32_t seq; fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL); if (fence == NULL) return -ENOMEM; - fence->seq = ++ring->fence_drv.sync_seq; + seq = ++ring->fence_drv.sync_seq; fence->ring = ring; fence_init(&fence->base, &amdgpu_fence_ops, &ring->fence_drv.lock, adev->fence_context + ring->idx, - fence->seq); + seq); amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr, - fence->seq, AMDGPU_FENCE_FLAG_INT); + seq, AMDGPU_FENCE_FLAG_INT); - idx = fence->seq & ring->fence_drv.num_fences_mask; - ptr = &ring->fence_drv.fences[idx]; + ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask]; /* This function can't be called concurrently anyway, otherwise * emitting the fence would mess up the hardware ring buffer. */ @@ -177,22 +175,16 @@ static void amdgpu_fence_schedule_fallback(struct amdgpu_ring *ring) void amdgpu_fence_process(struct amdgpu_ring *ring) { struct amdgpu_fence_driver *drv = &ring->fence_drv; - uint64_t seq, last_seq, last_emitted; + uint32_t seq, last_seq; int r; do { - last_seq = atomic64_read(&ring->fence_drv.last_seq); - last_emitted = ring->fence_drv.sync_seq; + last_seq = atomic_read(&ring->fence_drv.last_seq); seq = amdgpu_fence_read(ring); - seq |= last_seq & 0xffffffff00000000LL; - if (seq < last_seq) { - seq &= 0xffffffff; - seq |= last_emitted & 0xffffffff00000000LL; - } - } while (atomic64_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq); + } while (atomic_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq); - if (seq < last_emitted) + if (seq != ring->fence_drv.sync_seq) amdgpu_fence_schedule_fallback(ring); while (last_seq != seq) { @@ -279,13 +271,10 @@ unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring) * but it's ok to report slightly wrong fence count here. */ amdgpu_fence_process(ring); - emitted = ring->fence_drv.sync_seq - - atomic64_read(&ring->fence_drv.last_seq); - /* to avoid 32bits warp around */ - if (emitted > 0x10000000) - emitted = 0x10000000; - - return (unsigned)emitted; + emitted = 0x100000000ull; + emitted -= atomic_read(&ring->fence_drv.last_seq); + emitted += ACCESS_ONCE(ring->fence_drv.sync_seq); + return lower_32_bits(emitted); } /** @@ -317,7 +306,7 @@ int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring, ring->fence_drv.cpu_addr = adev->uvd.cpu_addr + index; ring->fence_drv.gpu_addr = adev->uvd.gpu_addr + index; } - amdgpu_fence_write(ring, atomic64_read(&ring->fence_drv.last_seq)); + amdgpu_fence_write(ring, atomic_read(&ring->fence_drv.last_seq)); amdgpu_irq_get(adev, irq_src, irq_type); ring->fence_drv.irq_src = irq_src; @@ -353,7 +342,7 @@ int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring, ring->fence_drv.cpu_addr = NULL; ring->fence_drv.gpu_addr = 0; ring->fence_drv.sync_seq = 0; - atomic64_set(&ring->fence_drv.last_seq, 0); + atomic_set(&ring->fence_drv.last_seq, 0); ring->fence_drv.initialized = false; setup_timer(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback, @@ -621,9 +610,9 @@ static int amdgpu_debugfs_fence_info(struct seq_file *m, void *data) amdgpu_fence_process(ring); seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name); - seq_printf(m, "Last signaled fence 0x%016llx\n", - (unsigned long long)atomic64_read(&ring->fence_drv.last_seq)); - seq_printf(m, "Last emitted 0x%016llx\n", + seq_printf(m, "Last signaled fence 0x%08x\n", + atomic_read(&ring->fence_drv.last_seq)); + seq_printf(m, "Last emitted 0x%08x\n", ring->fence_drv.sync_seq); } return 0; -- cgit v1.2.3 From fc387a0b183150b5beb953d740441f19b49a1fb3 Mon Sep 17 00:00:00 2001 From: Chunming Zhou Date: Thu, 31 Mar 2016 11:07:14 +0800 Subject: drm/amdgpu: fence wait old rcu slot MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit since the rcu slot was initialized to be num_hw_submission, if command submission doesn't use scheduler, this limitation will be invalid like uvd test. Signed-off-by: Chunming Zhou Reviewed-by: Christian König Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index 4303b447efe8..d81f1f4883a6 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -121,7 +121,7 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct fence **f) { struct amdgpu_device *adev = ring->adev; struct amdgpu_fence *fence; - struct fence **ptr; + struct fence *old, **ptr; uint32_t seq; fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL); @@ -141,7 +141,11 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct fence **f) /* This function can't be called concurrently anyway, otherwise * emitting the fence would mess up the hardware ring buffer. */ - BUG_ON(rcu_dereference_protected(*ptr, 1)); + old = rcu_dereference_protected(*ptr, 1); + if (old && !fence_is_signaled(old)) { + DRM_INFO("rcu slot is busy\n"); + fence_wait(old, false); + } rcu_assign_pointer(*ptr, fence_get(&fence->base)); -- cgit v1.2.3