From beb841028480a13a664f868688102f3cee762a6b Mon Sep 17 00:00:00 2001 From: Monk Liu Date: Thu, 21 Sep 2017 15:10:06 +0800 Subject: drm/amdgpu/sriov:alloc KIQ MQD in VRAM(v2) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit this way after KIQ MQD released in drv unloading, CPC can still let KIQ access this MQD thus RLCV SAVE_VF will not fail v2: always use VRAM domain for KIQ MQD no matter BM or SRIOV Signed-off-by: Monk Liu Acked-by: Alex Deucher Acked-by: Christian König Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c index 4f6c68fc1dd9..4fcd98e65998 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c @@ -260,8 +260,13 @@ int amdgpu_gfx_compute_mqd_sw_init(struct amdgpu_device *adev, /* create MQD for KIQ */ ring = &adev->gfx.kiq.ring; if (!ring->mqd_obj) { + /* originaly the KIQ MQD is put in GTT domain, but for SRIOV VRAM domain is a must + * otherwise hypervisor trigger SAVE_VF fail after driver unloaded which mean MQD + * deallocated and gart_unbind, to strict diverage we decide to use VRAM domain for + * KIQ MQD no matter SRIOV or Bare-metal + */ r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE, - AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj, + AMDGPU_GEM_DOMAIN_VRAM, &ring->mqd_obj, &ring->mqd_gpu_addr, &ring->mqd_ptr); if (r) { dev_warn(adev->dev, "failed to create ring mqd ob (%d)", r); -- cgit v1.2.3 From 0f7607d484f57c31a3e0b5b4e75ff1366cc90b6b Mon Sep 17 00:00:00 2001 From: Andres Rodriguez Date: Tue, 26 Sep 2017 12:22:45 -0400 Subject: drm/amdgpu: use multipipe compute policy on non PL11 asics A performance regression for OpenCL tests on Polaris11 had this feature disabled for all asics. Instead, disable it selectively on the affected asics. Acked-by: Felix Kuehling Signed-off-by: Andres Rodriguez Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 14 ++++++++++++-- 1 file changed, 12 insertions(+), 2 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c index 4fcd98e65998..3930fcc3e344 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c @@ -109,9 +109,20 @@ void amdgpu_gfx_parse_disable_cu(unsigned *mask, unsigned max_se, unsigned max_s } } +static bool amdgpu_gfx_is_multipipe_capable(struct amdgpu_device *adev) +{ + /* FIXME: spreading the queues across pipes causes perf regressions + * on POLARIS11 compute workloads */ + if (adev->asic_type == CHIP_POLARIS11) + return false; + + return adev->gfx.mec.num_mec > 1; +} + void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev) { int i, queue, pipe, mec; + bool multipipe_policy = amdgpu_gfx_is_multipipe_capable(adev); /* policy for amdgpu compute queue ownership */ for (i = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; ++i) { @@ -125,8 +136,7 @@ void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev) if (mec >= adev->gfx.mec.num_mec) break; - /* FIXME: spreading the queues across pipes causes perf regressions */ - if (0) { + if (multipipe_policy) { /* policy: amdgpu owns the first two queues of the first MEC */ if (mec == 0 && queue < 2) set_bit(i, adev->gfx.mec.queue_bitmap); -- cgit v1.2.3 From 4a75aefe3feb99ff87c1ea594b4db377b98f50e8 Mon Sep 17 00:00:00 2001 From: Andres Rodriguez Date: Tue, 26 Sep 2017 12:22:46 -0400 Subject: drm/amdgpu: add option for force enable multipipe policy for compute Useful for testing the effects of multipipe compute without recompiling. Acked-by: Felix Kuehling Signed-off-by: Andres Rodriguez Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 1 + drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 4 ++++ drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 6 ++++++ 3 files changed, 11 insertions(+) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h index ebfc267467ee..005b9823e468 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h @@ -121,6 +121,7 @@ extern int amdgpu_cntl_sb_buf_per_se; extern int amdgpu_param_buf_per_se; extern int amdgpu_job_hang_limit; extern int amdgpu_lbpw; +extern int amdgpu_compute_multipipe; #ifdef CONFIG_DRM_AMDGPU_SI extern int amdgpu_si_support; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c index 4f98960e47f9..9816f50a7498 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c @@ -122,6 +122,7 @@ int amdgpu_cntl_sb_buf_per_se = 0; int amdgpu_param_buf_per_se = 0; int amdgpu_job_hang_limit = 0; int amdgpu_lbpw = -1; +int amdgpu_compute_multipipe = -1; MODULE_PARM_DESC(vramlimit, "Restrict VRAM for testing, in megabytes"); module_param_named(vramlimit, amdgpu_vram_limit, int, 0600); @@ -265,6 +266,9 @@ module_param_named(job_hang_limit, amdgpu_job_hang_limit, int ,0444); MODULE_PARM_DESC(lbpw, "Load Balancing Per Watt (LBPW) support (1 = enable, 0 = disable, -1 = auto)"); module_param_named(lbpw, amdgpu_lbpw, int, 0444); +MODULE_PARM_DESC(compute_multipipe, "Force compute queues to be spread across pipes (1 = enable, 0 = disable, -1 = auto)"); +module_param_named(compute_multipipe, amdgpu_compute_multipipe, int, 0444); + #ifdef CONFIG_DRM_AMDGPU_SI #if defined(CONFIG_DRM_RADEON) || defined(CONFIG_DRM_RADEON_MODULE) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c index 3930fcc3e344..83435ccbad44 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c @@ -111,6 +111,12 @@ void amdgpu_gfx_parse_disable_cu(unsigned *mask, unsigned max_se, unsigned max_s static bool amdgpu_gfx_is_multipipe_capable(struct amdgpu_device *adev) { + if (amdgpu_compute_multipipe != -1) { + DRM_INFO("amdgpu: forcing compute pipe policy %d\n", + amdgpu_compute_multipipe); + return amdgpu_compute_multipipe == 1; + } + /* FIXME: spreading the queues across pipes causes perf regressions * on POLARIS11 compute workloads */ if (adev->asic_type == CHIP_POLARIS11) -- cgit v1.2.3 From 43ca8efa46d9b1c4defa1b27c4dd1ef3866aaad9 Mon Sep 17 00:00:00 2001 From: pding Date: Fri, 13 Oct 2017 15:38:35 +0800 Subject: drm/amdgpu: busywait KIQ register accessing (v4) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Register accessing is performed when IRQ is disabled. Never sleep in this function. Known issue: dead sleep in many use cases of index/data registers. v2: - wrap polling fence functions. - don't trigger IRQ for polling in case of wrongly fence signal. v3: - handle wrap round gracefully. - add comments for polling function v4: - don't return negative timeout confused with error code Signed-off-by: pding Reviewed-by: Christian König Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 8 ++--- drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 50 ++++++++++++++++++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 4 +++ drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c | 30 ++++++++---------- 6 files changed, 71 insertions(+), 25 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h index b8ba1f5ae5e7..cbcb6a153aba 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h @@ -879,7 +879,7 @@ struct amdgpu_mec { struct amdgpu_kiq { u64 eop_gpu_addr; struct amdgpu_bo *eop_obj; - struct mutex ring_mutex; + spinlock_t ring_lock; struct amdgpu_ring ring; struct amdgpu_irq_src irq; }; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c index 3a6ce6386ad0..0b9332e65a4c 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c @@ -109,10 +109,8 @@ uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg, { uint32_t ret; - if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev)) { - BUG_ON(in_interrupt()); + if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev)) return amdgpu_virt_kiq_rreg(adev, reg); - } if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX)) ret = readl(((void __iomem *)adev->rmmio) + (reg * 4)); @@ -137,10 +135,8 @@ void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v, adev->last_mm_index = v; } - if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev)) { - BUG_ON(in_interrupt()); + if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev)) return amdgpu_virt_kiq_wreg(adev, reg, v); - } if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX)) writel(v, ((void __iomem *)adev->rmmio) + (reg * 4)); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c index 7bdedd788f5a..fb9f88ef6059 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c @@ -168,6 +168,32 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f) return 0; } +/** + * amdgpu_fence_emit_polling - emit a fence on the requeste ring + * + * @ring: ring the fence is associated with + * @s: resulting sequence number + * + * Emits a fence command on the requested ring (all asics). + * Used For polling fence. + * Returns 0 on success, -ENOMEM on failure. + */ +int amdgpu_fence_emit_polling(struct amdgpu_ring *ring, uint32_t *s) +{ + uint32_t seq; + + if (!s) + return -EINVAL; + + seq = ++ring->fence_drv.sync_seq; + amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr, + seq, AMDGPU_FENCE_FLAG_INT); + + *s = seq; + + return 0; +} + /** * amdgpu_fence_schedule_fallback - schedule fallback check * @@ -281,6 +307,30 @@ int amdgpu_fence_wait_empty(struct amdgpu_ring *ring) return r; } +/** + * amdgpu_fence_wait_polling - busy wait for givn sequence number + * + * @ring: ring index the fence is associated with + * @wait_seq: sequence number to wait + * @timeout: the timeout for waiting in usecs + * + * Wait for all fences on the requested ring to signal (all asics). + * Returns left time if no timeout, 0 or minus if timeout. + */ +signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring, + uint32_t wait_seq, + signed long timeout) +{ + uint32_t seq; + + do { + seq = amdgpu_fence_read(ring); + udelay(5); + timeout -= 5; + } while ((int32_t)(wait_seq - seq) > 0 && timeout > 0); + + return timeout > 0 ? timeout : 0; +} /** * amdgpu_fence_count_emitted - get the count of emitted fences * diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c index 83435ccbad44..ef043361009f 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c @@ -201,7 +201,7 @@ int amdgpu_gfx_kiq_init_ring(struct amdgpu_device *adev, struct amdgpu_kiq *kiq = &adev->gfx.kiq; int r = 0; - mutex_init(&kiq->ring_mutex); + spin_lock_init(&kiq->ring_lock); r = amdgpu_wb_get(adev, &adev->virt.reg_val_offs); if (r) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h index 0d9ce141404c..b18c2b96691f 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h @@ -90,8 +90,12 @@ int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring, 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 dma_fence **fence); +int amdgpu_fence_emit_polling(struct amdgpu_ring *ring, uint32_t *s); void amdgpu_fence_process(struct amdgpu_ring *ring); int amdgpu_fence_wait_empty(struct amdgpu_ring *ring); +signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring, + uint32_t wait_seq, + signed long timeout); unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring); /* diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c index ed7be2eb24b0..e97f80f86005 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c @@ -22,7 +22,7 @@ */ #include "amdgpu.h" -#define MAX_KIQ_REG_WAIT 100000 +#define MAX_KIQ_REG_WAIT 100000000 /* in usecs */ int amdgpu_allocate_static_csa(struct amdgpu_device *adev) { @@ -114,27 +114,24 @@ void amdgpu_virt_init_setting(struct amdgpu_device *adev) uint32_t amdgpu_virt_kiq_rreg(struct amdgpu_device *adev, uint32_t reg) { signed long r; - uint32_t val; - struct dma_fence *f; + uint32_t val, seq; struct amdgpu_kiq *kiq = &adev->gfx.kiq; struct amdgpu_ring *ring = &kiq->ring; BUG_ON(!ring->funcs->emit_rreg); - mutex_lock(&kiq->ring_mutex); + spin_lock(&kiq->ring_lock); amdgpu_ring_alloc(ring, 32); amdgpu_ring_emit_rreg(ring, reg); - amdgpu_fence_emit(ring, &f); + amdgpu_fence_emit_polling(ring, &seq); amdgpu_ring_commit(ring); - mutex_unlock(&kiq->ring_mutex); + spin_unlock(&kiq->ring_lock); - r = dma_fence_wait_timeout(f, false, msecs_to_jiffies(MAX_KIQ_REG_WAIT)); - dma_fence_put(f); + r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT); if (r < 1) { - DRM_ERROR("wait for kiq fence error: %ld.\n", r); + DRM_ERROR("wait for kiq fence error: %ld\n", r); return ~0; } - val = adev->wb.wb[adev->virt.reg_val_offs]; return val; @@ -143,23 +140,22 @@ uint32_t amdgpu_virt_kiq_rreg(struct amdgpu_device *adev, uint32_t reg) void amdgpu_virt_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v) { signed long r; - struct dma_fence *f; + uint32_t seq; struct amdgpu_kiq *kiq = &adev->gfx.kiq; struct amdgpu_ring *ring = &kiq->ring; BUG_ON(!ring->funcs->emit_wreg); - mutex_lock(&kiq->ring_mutex); + spin_lock(&kiq->ring_lock); amdgpu_ring_alloc(ring, 32); amdgpu_ring_emit_wreg(ring, reg, v); - amdgpu_fence_emit(ring, &f); + amdgpu_fence_emit_polling(ring, &seq); amdgpu_ring_commit(ring); - mutex_unlock(&kiq->ring_mutex); + spin_unlock(&kiq->ring_lock); - r = dma_fence_wait_timeout(f, false, msecs_to_jiffies(MAX_KIQ_REG_WAIT)); + r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT); if (r < 1) - DRM_ERROR("wait for kiq fence error: %ld.\n", r); - dma_fence_put(f); + DRM_ERROR("wait for kiq fence error: %ld\n", r); } /** -- cgit v1.2.3