From 373008bfc9cdb0f050258947fa5a095f0657e1bc Mon Sep 17 00:00:00 2001 From: Dusica Milinkovic Date: Wed, 10 Aug 2022 09:43:15 +0200 Subject: drm/amdgpu: Increase tlb flush timeout for sriov [Why] During multi-vf executing benchmark (Luxmark) observed kiq error timeout. It happenes because all of VFs do the tlb invalidation at the same time. Although each VF has the invalidate register set, from hardware side the invalidate requests are queue to execute. [How] In case of 12 VF increase timeout on 12*100ms Signed-off-by: Dusica Milinkovic Acked-by: Shaoyun Liu Acked-by: Alex Deucher Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu.h') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h index e146810c700b..d597e2656c47 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h @@ -317,7 +317,7 @@ enum amdgpu_kiq_irq { AMDGPU_CP_KIQ_IRQ_DRIVER0 = 0, AMDGPU_CP_KIQ_IRQ_LAST }; - +#define SRIOV_USEC_TIMEOUT 1200000 /* wait 12 * 100ms for SRIOV */ #define MAX_KIQ_REG_WAIT 5000 /* in usecs, 5ms */ #define MAX_KIQ_REG_BAILOUT_INTERVAL 5 /* in msecs, 5ms */ #define MAX_KIQ_REG_TRY 1000 -- cgit v1.2.3-59-g8ed1b From 5bd8d53f6fa53eab5433698d1362dae2aa53c1cc Mon Sep 17 00:00:00 2001 From: Victor Zhao Date: Tue, 14 Jun 2022 10:48:11 +0800 Subject: drm/amdgpu: add debugfs amdgpu_reset_level Introduce amdgpu_reset_level debugfs in order to help debug and test specific type of reset. Also helps blocking unwanted type of resets. By default, mode2 reset will not be enabled v2: make this debugfs in adev and use debugfs_create_u32 Signed-off-by: Victor Zhao Acked-by: Andrey Grodzovsky Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 5 +++++ drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c | 2 ++ drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c | 8 ++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c | 3 +++ 4 files changed, 18 insertions(+) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu.h') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h index d597e2656c47..66a26f2c51c1 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h @@ -274,6 +274,9 @@ extern int amdgpu_vcnfw_log; #define AMDGPU_RESET_VCE (1 << 13) #define AMDGPU_RESET_VCE1 (1 << 14) +#define AMDGPU_RESET_LEVEL_SOFT_RECOVERY (1 << 0) +#define AMDGPU_RESET_LEVEL_MODE2 (1 << 1) + /* max cursor sizes (in pixels) */ #define CIK_CURSOR_WIDTH 128 #define CIK_CURSOR_HEIGHT 128 @@ -1060,6 +1063,8 @@ struct amdgpu_device { uint32_t scpm_status; struct work_struct reset_work; + + uint32_t amdgpu_reset_level_mask; }; static inline struct amdgpu_device *drm_to_adev(struct drm_device *ddev) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c index cb00c7d6f50b..4e54569fa223 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c @@ -1786,6 +1786,8 @@ int amdgpu_debugfs_init(struct amdgpu_device *adev) return PTR_ERR(ent); } + debugfs_create_u32("amdgpu_reset_level", 0600, root, &adev->amdgpu_reset_level_mask); + /* Register debugfs entries for amdgpu_ttm */ amdgpu_ttm_debugfs_init(adev); amdgpu_debugfs_pm_init(adev); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c index 831fb222139c..9da5ead50c90 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c @@ -37,6 +37,8 @@ int amdgpu_reset_init(struct amdgpu_device *adev) { int ret = 0; + adev->amdgpu_reset_level_mask = 0x1; + switch (adev->ip_versions[MP1_HWIP][0]) { case IP_VERSION(13, 0, 2): ret = aldebaran_reset_init(adev); @@ -74,6 +76,9 @@ int amdgpu_reset_prepare_hwcontext(struct amdgpu_device *adev, { struct amdgpu_reset_handler *reset_handler = NULL; + if (!(adev->amdgpu_reset_level_mask & AMDGPU_RESET_LEVEL_MODE2)) + return -ENOSYS; + if (test_bit(AMDGPU_SKIP_MODE2_RESET, &reset_context->flags)) return -ENOSYS; @@ -93,6 +98,9 @@ int amdgpu_reset_perform_reset(struct amdgpu_device *adev, int ret; struct amdgpu_reset_handler *reset_handler = NULL; + if (!(adev->amdgpu_reset_level_mask & AMDGPU_RESET_LEVEL_MODE2)) + return -ENOSYS; + if (test_bit(AMDGPU_SKIP_MODE2_RESET, &reset_context->flags)) return -ENOSYS; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c index d3558c34d406..3e316b013fd9 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c @@ -405,6 +405,9 @@ bool amdgpu_ring_soft_recovery(struct amdgpu_ring *ring, unsigned int vmid, { ktime_t deadline = ktime_add_us(ktime_get(), 10000); + if (!(ring->adev->amdgpu_reset_level_mask & AMDGPU_RESET_LEVEL_SOFT_RECOVERY)) + return false; + if (amdgpu_sriov_vf(ring->adev) || !ring->funcs->soft_recovery || !fence) return false; -- cgit v1.2.3-59-g8ed1b From 194eb174cbe4fe2b3376ac30acca2dc8c8beca00 Mon Sep 17 00:00:00 2001 From: Victor Zhao Date: Fri, 24 Jun 2022 12:00:06 +0800 Subject: drm/amdgpu: reduce reset time In multi container use case, reset time is important, so skip ring tests and cp halt wait during ip suspending for reset as they are going to fail and cost more time on reset v2: add a hang flag to indicate the reset comes from a job timeout, skip ring test and cp halt wait in this case v3: move hang flag to adev Signed-off-by: Victor Zhao Acked-by: Andrey Grodzovsky Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 1 + drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 2 ++ drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c | 9 +++++++-- 4 files changed, 11 insertions(+), 3 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu.h') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h index 66a26f2c51c1..79bb6fd83094 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h @@ -1065,6 +1065,7 @@ struct amdgpu_device { struct work_struct reset_work; uint32_t amdgpu_reset_level_mask; + bool job_hang; }; static inline struct amdgpu_device *drm_to_adev(struct drm_device *ddev) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c index 222d3d7ea076..454a78ba60d4 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c @@ -477,7 +477,7 @@ int amdgpu_gfx_disable_kcq(struct amdgpu_device *adev) kiq->pmf->kiq_unmap_queues(kiq_ring, &adev->gfx.compute_ring[i], RESET_QUEUES, 0, 0); - if (adev->gfx.kiq.ring.sched.ready) + if (adev->gfx.kiq.ring.sched.ready && !adev->job_hang) r = amdgpu_ring_test_helper(kiq_ring); spin_unlock(&adev->gfx.kiq.ring_lock); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c index 36f49da15800..8f51adf3b329 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c @@ -49,6 +49,7 @@ static enum drm_gpu_sched_stat amdgpu_job_timedout(struct drm_sched_job *s_job) } memset(&ti, 0, sizeof(struct amdgpu_task_info)); + adev->job_hang = true; if (amdgpu_gpu_recovery && amdgpu_ring_soft_recovery(ring, job->vmid, s_job->s_fence->parent)) { @@ -83,6 +84,7 @@ static enum drm_gpu_sched_stat amdgpu_job_timedout(struct drm_sched_job *s_job) } exit: + adev->job_hang = false; drm_dev_exit(idx); return DRM_GPU_SCHED_STAT_NOMINAL; } diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c index a2a4dc1844c0..1a915edccb92 100644 --- a/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c @@ -5971,6 +5971,9 @@ static int gfx_v10_0_cp_gfx_enable(struct amdgpu_device *adev, bool enable) WREG32_SOC15(GC, 0, mmCP_ME_CNTL, tmp); } + if (adev->job_hang && !enable) + return 0; + for (i = 0; i < adev->usec_timeout; i++) { if (RREG32_SOC15(GC, 0, mmCP_STAT) == 0) break; @@ -7569,8 +7572,10 @@ static int gfx_v10_0_kiq_disable_kgq(struct amdgpu_device *adev) for (i = 0; i < adev->gfx.num_gfx_rings; i++) kiq->pmf->kiq_unmap_queues(kiq_ring, &adev->gfx.gfx_ring[i], PREEMPT_QUEUES, 0, 0); - - return amdgpu_ring_test_helper(kiq_ring); + if (!adev->job_hang) + return amdgpu_ring_test_helper(kiq_ring); + else + return 0; } #endif -- cgit v1.2.3-59-g8ed1b From 68ce8b242242651eb3cb4ff29b79c44d02f752c9 Mon Sep 17 00:00:00 2001 From: Christian König Date: Wed, 2 Mar 2022 16:26:53 +0100 Subject: drm/amdgpu: add gang submit backend v2 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Allows submitting jobs as gang which needs to run on multiple engines at the same time. Basic idea is that we have a global gang submit fence representing when the gang leader is finally pushed to run on the hardware last. Jobs submitted as gang are never re-submitted in case of a GPU reset since this won't work and will just deadlock the hardware immediately again. v2: fix logic inversion, improve documentation, fix rcu Signed-off-by: Christian König Reviewed-by: Alex Deucher Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 3 +++ drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 35 ++++++++++++++++++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 28 ++++++++++++++++++++++-- drivers/gpu/drm/amd/amdgpu/amdgpu_job.h | 3 +++ 4 files changed, 67 insertions(+), 2 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu.h') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h index 79bb6fd83094..ae9371b172e3 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h @@ -885,6 +885,7 @@ struct amdgpu_device { u64 fence_context; unsigned num_rings; struct amdgpu_ring *rings[AMDGPU_MAX_RINGS]; + struct dma_fence __rcu *gang_submit; bool ib_pool_ready; struct amdgpu_sa_manager ib_pools[AMDGPU_IB_POOL_MAX]; struct amdgpu_sched gpu_sched[AMDGPU_HW_IP_NUM][AMDGPU_RING_PRIO_MAX]; @@ -1294,6 +1295,8 @@ u32 amdgpu_device_pcie_port_rreg(struct amdgpu_device *adev, u32 reg); void amdgpu_device_pcie_port_wreg(struct amdgpu_device *adev, u32 reg, u32 v); +struct dma_fence *amdgpu_device_switch_gang(struct amdgpu_device *adev, + struct dma_fence *gang); /* atpx handler */ #if defined(CONFIG_VGA_SWITCHEROO) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c index 869c843c1d58..c04ea7f1e819 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c @@ -3511,6 +3511,7 @@ int amdgpu_device_init(struct amdgpu_device *adev, adev->gmc.gart_size = 512 * 1024 * 1024; adev->accel_working = false; adev->num_rings = 0; + RCU_INIT_POINTER(adev->gang_submit, dma_fence_get_stub()); adev->mman.buffer_funcs = NULL; adev->mman.buffer_funcs_ring = NULL; adev->vm_manager.vm_pte_funcs = NULL; @@ -3992,6 +3993,7 @@ void amdgpu_device_fini_sw(struct amdgpu_device *adev) release_firmware(adev->firmware.gpu_info_fw); adev->firmware.gpu_info_fw = NULL; adev->accel_working = false; + dma_fence_put(rcu_dereference_protected(adev->gang_submit, true)); amdgpu_reset_fini(adev); @@ -5960,3 +5962,36 @@ void amdgpu_device_pcie_port_wreg(struct amdgpu_device *adev, (void)RREG32(data); spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); } + +/** + * amdgpu_device_switch_gang - switch to a new gang + * @adev: amdgpu_device pointer + * @gang: the gang to switch to + * + * Try to switch to a new gang. + * Returns: NULL if we switched to the new gang or a reference to the current + * gang leader. + */ +struct dma_fence *amdgpu_device_switch_gang(struct amdgpu_device *adev, + struct dma_fence *gang) +{ + struct dma_fence *old = NULL; + + do { + dma_fence_put(old); + rcu_read_lock(); + old = dma_fence_get_rcu_safe(&adev->gang_submit); + rcu_read_unlock(); + + if (old == gang) + break; + + if (!dma_fence_is_signaled(old)) + return old; + + } while (cmpxchg((struct dma_fence __force **)&adev->gang_submit, + old, gang) != old); + + dma_fence_put(old); + return NULL; +} diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c index 877b3c22a8a8..cfbe19cfe9af 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c @@ -173,11 +173,29 @@ static void amdgpu_job_free_cb(struct drm_sched_job *s_job) dma_fence_put(&job->hw_fence); } +void amdgpu_job_set_gang_leader(struct amdgpu_job *job, + struct amdgpu_job *leader) +{ + struct dma_fence *fence = &leader->base.s_fence->scheduled; + + WARN_ON(job->gang_submit); + + /* + * Don't add a reference when we are the gang leader to avoid circle + * dependency. + */ + if (job != leader) + dma_fence_get(fence); + job->gang_submit = fence; +} + void amdgpu_job_free(struct amdgpu_job *job) { amdgpu_job_free_resources(job); amdgpu_sync_free(&job->sync); amdgpu_sync_free(&job->sched_sync); + if (job->gang_submit != &job->base.s_fence->scheduled) + dma_fence_put(job->gang_submit); if (!job->hw_fence.ops) kfree(job); @@ -247,12 +265,16 @@ static struct dma_fence *amdgpu_job_dependency(struct drm_sched_job *sched_job, fence = amdgpu_sync_get_fence(&job->sync); } + if (!fence && job->gang_submit) + fence = amdgpu_device_switch_gang(ring->adev, job->gang_submit); + return fence; } static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) { struct amdgpu_ring *ring = to_amdgpu_ring(sched_job->sched); + struct amdgpu_device *adev = ring->adev; struct dma_fence *fence = NULL, *finished; struct amdgpu_job *job; int r = 0; @@ -264,8 +286,10 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job) trace_amdgpu_sched_run_job(job); - if (job->vram_lost_counter != atomic_read(&ring->adev->vram_lost_counter)) - dma_fence_set_error(finished, -ECANCELED);/* skip IB as well if VRAM lost */ + /* Skip job if VRAM is lost and never resubmit gangs */ + if (job->vram_lost_counter != atomic_read(&adev->vram_lost_counter) || + (job->job_run_counter && job->gang_submit)) + dma_fence_set_error(finished, -ECANCELED); if (finished->error < 0) { DRM_INFO("Skip scheduling IBs!\n"); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h index 866d35bbe073..ab7b150e5d50 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h @@ -50,6 +50,7 @@ struct amdgpu_job { struct amdgpu_sync sync; struct amdgpu_sync sched_sync; struct dma_fence hw_fence; + struct dma_fence *gang_submit; uint32_t preamble_status; uint32_t preemption_status; bool vm_needs_flush; @@ -84,6 +85,8 @@ int amdgpu_job_alloc_with_ib(struct amdgpu_device *adev, unsigned size, void amdgpu_job_set_resources(struct amdgpu_job *job, struct amdgpu_bo *gds, struct amdgpu_bo *gws, struct amdgpu_bo *oa); void amdgpu_job_free_resources(struct amdgpu_job *job); +void amdgpu_job_set_gang_leader(struct amdgpu_job *job, + struct amdgpu_job *leader); void amdgpu_job_free(struct amdgpu_job *job); int amdgpu_job_submit(struct amdgpu_job *job, struct drm_sched_entity *entity, void *owner, struct dma_fence **f); -- cgit v1.2.3-59-g8ed1b