diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c')
| -rw-r--r-- | drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 1754 |
1 files changed, 1625 insertions, 129 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c index 35ed46b9249c..8b118c53f351 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c @@ -24,10 +24,17 @@ */ #include <linux/firmware.h> +#include <linux/pm_runtime.h> + #include "amdgpu.h" #include "amdgpu_gfx.h" #include "amdgpu_rlc.h" #include "amdgpu_ras.h" +#include "amdgpu_reset.h" +#include "amdgpu_xcp.h" +#include "amdgpu_xgmi.h" +#include "amdgpu_mes.h" +#include "nvd.h" /* delay 0.1 second to enable gfx off feature */ #define GFX_OFF_DELAY_ENABLE msecs_to_jiffies(100) @@ -63,35 +70,26 @@ void amdgpu_queue_mask_bit_to_mec_queue(struct amdgpu_device *adev, int bit, } bool amdgpu_gfx_is_mec_queue_enabled(struct amdgpu_device *adev, - int mec, int pipe, int queue) + int xcc_id, int mec, int pipe, int queue) { return test_bit(amdgpu_gfx_mec_queue_to_bit(adev, mec, pipe, queue), - adev->gfx.mec.queue_bitmap); + adev->gfx.mec_bitmap[xcc_id].queue_bitmap); } -int amdgpu_gfx_me_queue_to_bit(struct amdgpu_device *adev, - int me, int pipe, int queue) +static int amdgpu_gfx_me_queue_to_bit(struct amdgpu_device *adev, + int me, int pipe, int queue) { + int num_queue_per_pipe = 1; /* we only enable 1 KGQ per pipe */ int bit = 0; bit += me * adev->gfx.me.num_pipe_per_me - * adev->gfx.me.num_queue_per_pipe; - bit += pipe * adev->gfx.me.num_queue_per_pipe; + * num_queue_per_pipe; + bit += pipe * num_queue_per_pipe; bit += queue; return bit; } -void amdgpu_gfx_bit_to_me_queue(struct amdgpu_device *adev, int bit, - int *me, int *pipe, int *queue) -{ - *queue = bit % adev->gfx.me.num_queue_per_pipe; - *pipe = (bit / adev->gfx.me.num_queue_per_pipe) - % adev->gfx.me.num_pipe_per_me; - *me = (bit / adev->gfx.me.num_queue_per_pipe) - / adev->gfx.me.num_pipe_per_me; -} - bool amdgpu_gfx_is_me_queue_enabled(struct amdgpu_device *adev, int me, int pipe, int queue) { @@ -109,9 +107,9 @@ bool amdgpu_gfx_is_me_queue_enabled(struct amdgpu_device *adev, * The bitmask of CUs to be disabled in the shader array determined by se and * sh is stored in mask[se * max_sh + sh]. */ -void amdgpu_gfx_parse_disable_cu(unsigned *mask, unsigned max_se, unsigned max_sh) +void amdgpu_gfx_parse_disable_cu(unsigned int *mask, unsigned int max_se, unsigned int max_sh) { - unsigned se, sh, cu; + unsigned int se, sh, cu; const char *p; memset(mask, 0, sizeof(*mask) * max_se * max_sh); @@ -123,6 +121,7 @@ void amdgpu_gfx_parse_disable_cu(unsigned *mask, unsigned max_se, unsigned max_s for (;;) { char *next; int ret = sscanf(p, "%u.%u.%u", &se, &sh, &cu); + if (ret < 3) { DRM_ERROR("amdgpu: could not parse disable_cu\n"); return; @@ -151,12 +150,12 @@ static bool amdgpu_gfx_is_graphics_multipipe_capable(struct amdgpu_device *adev) static bool amdgpu_gfx_is_compute_multipipe_capable(struct amdgpu_device *adev) { if (amdgpu_compute_multipipe != -1) { - DRM_INFO("amdgpu: forcing compute pipe policy %d\n", + dev_info(adev->dev, "amdgpu: forcing compute pipe policy %d\n", amdgpu_compute_multipipe); return amdgpu_compute_multipipe == 1; } - if (adev->ip_versions[GC_HWIP][0] > IP_VERSION(9, 0, 0)) + if (amdgpu_ip_version(adev, GC_HWIP, 0) > IP_VERSION(9, 0, 0)) return true; /* FIXME: spreading the queues across pipes causes perf regressions @@ -204,37 +203,46 @@ bool amdgpu_gfx_is_high_priority_compute_queue(struct amdgpu_device *adev, void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev) { - int i, queue, pipe; + int i, j, queue, pipe; bool multipipe_policy = amdgpu_gfx_is_compute_multipipe_capable(adev); int max_queues_per_mec = min(adev->gfx.mec.num_pipe_per_mec * adev->gfx.mec.num_queue_per_pipe, adev->gfx.num_compute_rings); + int num_xcc = adev->gfx.xcc_mask ? NUM_XCC(adev->gfx.xcc_mask) : 1; if (multipipe_policy) { - /* policy: make queues evenly cross all pipes on MEC1 only */ - for (i = 0; i < max_queues_per_mec; i++) { - pipe = i % adev->gfx.mec.num_pipe_per_mec; - queue = (i / adev->gfx.mec.num_pipe_per_mec) % - adev->gfx.mec.num_queue_per_pipe; - - set_bit(pipe * adev->gfx.mec.num_queue_per_pipe + queue, - adev->gfx.mec.queue_bitmap); + /* policy: make queues evenly cross all pipes on MEC1 only + * for multiple xcc, just use the original policy for simplicity */ + for (j = 0; j < num_xcc; j++) { + for (i = 0; i < max_queues_per_mec; i++) { + pipe = i % adev->gfx.mec.num_pipe_per_mec; + queue = (i / adev->gfx.mec.num_pipe_per_mec) % + adev->gfx.mec.num_queue_per_pipe; + + set_bit(pipe * adev->gfx.mec.num_queue_per_pipe + queue, + adev->gfx.mec_bitmap[j].queue_bitmap); + } } } else { /* policy: amdgpu owns all queues in the given pipe */ - for (i = 0; i < max_queues_per_mec; ++i) - set_bit(i, adev->gfx.mec.queue_bitmap); + for (j = 0; j < num_xcc; j++) { + for (i = 0; i < max_queues_per_mec; ++i) + set_bit(i, adev->gfx.mec_bitmap[j].queue_bitmap); + } } - dev_dbg(adev->dev, "mec queue bitmap weight=%d\n", bitmap_weight(adev->gfx.mec.queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES)); + for (j = 0; j < num_xcc; j++) { + dev_dbg(adev->dev, "mec queue bitmap weight=%d\n", + bitmap_weight(adev->gfx.mec_bitmap[j].queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES)); + } } void amdgpu_gfx_graphics_queue_acquire(struct amdgpu_device *adev) { int i, queue, pipe; bool multipipe_policy = amdgpu_gfx_is_graphics_multipipe_capable(adev); - int max_queues_per_me = adev->gfx.me.num_pipe_per_me * - adev->gfx.me.num_queue_per_pipe; + int num_queue_per_pipe = 1; /* we only enable 1 KGQ per pipe */ + int max_queues_per_me = adev->gfx.me.num_pipe_per_me * num_queue_per_pipe; if (multipipe_policy) { /* policy: amdgpu owns the first queue per pipe at this stage @@ -242,9 +250,9 @@ void amdgpu_gfx_graphics_queue_acquire(struct amdgpu_device *adev) for (i = 0; i < max_queues_per_me; i++) { pipe = i % adev->gfx.me.num_pipe_per_me; queue = (i / adev->gfx.me.num_pipe_per_me) % - adev->gfx.me.num_queue_per_pipe; + num_queue_per_pipe; - set_bit(pipe * adev->gfx.me.num_queue_per_pipe + queue, + set_bit(pipe * num_queue_per_pipe + queue, adev->gfx.me.queue_bitmap); } } else { @@ -253,12 +261,13 @@ void amdgpu_gfx_graphics_queue_acquire(struct amdgpu_device *adev) } /* update the number of active graphics rings */ - adev->gfx.num_gfx_rings = - bitmap_weight(adev->gfx.me.queue_bitmap, AMDGPU_MAX_GFX_QUEUES); + if (adev->gfx.num_gfx_rings) + adev->gfx.num_gfx_rings = + bitmap_weight(adev->gfx.me.queue_bitmap, AMDGPU_MAX_GFX_QUEUES); } static int amdgpu_gfx_kiq_acquire(struct amdgpu_device *adev, - struct amdgpu_ring *ring) + struct amdgpu_ring *ring, int xcc_id) { int queue_bit; int mec, pipe, queue; @@ -268,7 +277,7 @@ static int amdgpu_gfx_kiq_acquire(struct amdgpu_device *adev, * adev->gfx.mec.num_queue_per_pipe; while (--queue_bit >= 0) { - if (test_bit(queue_bit, adev->gfx.mec.queue_bitmap)) + if (test_bit(queue_bit, adev->gfx.mec_bitmap[xcc_id].queue_bitmap)) continue; amdgpu_queue_mask_bit_to_mec_queue(adev, queue_bit, &mec, &pipe, &queue); @@ -292,11 +301,11 @@ static int amdgpu_gfx_kiq_acquire(struct amdgpu_device *adev, return -EINVAL; } -int amdgpu_gfx_kiq_init_ring(struct amdgpu_device *adev, - struct amdgpu_ring *ring, - struct amdgpu_irq_src *irq) +int amdgpu_gfx_kiq_init_ring(struct amdgpu_device *adev, int xcc_id) { - struct amdgpu_kiq *kiq = &adev->gfx.kiq; + struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id]; + struct amdgpu_irq_src *irq = &kiq->irq; + struct amdgpu_ring *ring = &kiq->ring; int r = 0; spin_lock_init(&kiq->ring_lock); @@ -304,15 +313,22 @@ int amdgpu_gfx_kiq_init_ring(struct amdgpu_device *adev, ring->adev = NULL; ring->ring_obj = NULL; ring->use_doorbell = true; - ring->doorbell_index = adev->doorbell_index.kiq; - - r = amdgpu_gfx_kiq_acquire(adev, ring); + ring->xcc_id = xcc_id; + ring->vm_hub = AMDGPU_GFXHUB(xcc_id); + ring->doorbell_index = + (adev->doorbell_index.kiq + + xcc_id * adev->doorbell_index.xcc_doorbell_range) + << 1; + + r = amdgpu_gfx_kiq_acquire(adev, ring, xcc_id); if (r) return r; ring->eop_gpu_addr = kiq->eop_gpu_addr; ring->no_scheduler = true; - sprintf(ring->name, "kiq_%d.%d.%d", ring->me, ring->pipe, ring->queue); + snprintf(ring->name, sizeof(ring->name), "kiq_%hhu.%hhu.%hhu.%hhu", + (unsigned char)xcc_id, (unsigned char)ring->me, + (unsigned char)ring->pipe, (unsigned char)ring->queue); r = amdgpu_ring_init(adev, ring, 1024, irq, AMDGPU_CP_KIQ_IRQ_DRIVER0, AMDGPU_RING_PRIO_DEFAULT, NULL); if (r) @@ -326,19 +342,19 @@ void amdgpu_gfx_kiq_free_ring(struct amdgpu_ring *ring) amdgpu_ring_fini(ring); } -void amdgpu_gfx_kiq_fini(struct amdgpu_device *adev) +void amdgpu_gfx_kiq_fini(struct amdgpu_device *adev, int xcc_id) { - struct amdgpu_kiq *kiq = &adev->gfx.kiq; + struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id]; amdgpu_bo_free_kernel(&kiq->eop_obj, &kiq->eop_gpu_addr, NULL); } int amdgpu_gfx_kiq_init(struct amdgpu_device *adev, - unsigned hpd_size) + unsigned int hpd_size, int xcc_id) { int r; u32 *hpd; - struct amdgpu_kiq *kiq = &adev->gfx.kiq; + struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id]; r = amdgpu_bo_create_kernel(adev, hpd_size, PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT, &kiq->eop_obj, @@ -361,13 +377,20 @@ int amdgpu_gfx_kiq_init(struct amdgpu_device *adev, /* create MQD for each compute/gfx queue */ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev, - unsigned mqd_size) + unsigned int mqd_size, int xcc_id) { - struct amdgpu_ring *ring = NULL; - int r, i; + int r, i, j; + struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id]; + struct amdgpu_ring *ring = &kiq->ring; + u32 domain = AMDGPU_GEM_DOMAIN_GTT; + +#if !defined(CONFIG_ARM) && !defined(CONFIG_ARM64) + /* Only enable on gfx10 and 11 for now to avoid changing behavior on older chips */ + if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(10, 0, 0)) + domain |= AMDGPU_GEM_DOMAIN_VRAM; +#endif /* create MQD for KIQ */ - ring = &adev->gfx.kiq.ring; if (!adev->enable_mes_kiq && !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 @@ -386,9 +409,12 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev, } /* prepare MQD backup */ - adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS] = kmalloc(mqd_size, GFP_KERNEL); - if (!adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS]) - dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name); + kiq->mqd_backup = kzalloc(mqd_size, GFP_KERNEL); + if (!kiq->mqd_backup) { + dev_warn(adev->dev, + "no memory to create MQD backup for ring %s\n", ring->name); + return -ENOMEM; + } } if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) { @@ -397,47 +423,55 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev, ring = &adev->gfx.gfx_ring[i]; if (!ring->mqd_obj) { r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE, - AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj, + domain, &ring->mqd_obj, &ring->mqd_gpu_addr, &ring->mqd_ptr); if (r) { dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r); return r; } + ring->mqd_size = mqd_size; /* prepare MQD backup */ - adev->gfx.me.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL); - if (!adev->gfx.me.mqd_backup[i]) + adev->gfx.me.mqd_backup[i] = kzalloc(mqd_size, GFP_KERNEL); + if (!adev->gfx.me.mqd_backup[i]) { dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name); + return -ENOMEM; + } } } } /* create MQD for each KCQ */ for (i = 0; i < adev->gfx.num_compute_rings; i++) { - ring = &adev->gfx.compute_ring[i]; + j = i + xcc_id * adev->gfx.num_compute_rings; + ring = &adev->gfx.compute_ring[j]; if (!ring->mqd_obj) { r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE, - AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj, + domain, &ring->mqd_obj, &ring->mqd_gpu_addr, &ring->mqd_ptr); if (r) { dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r); return r; } + ring->mqd_size = mqd_size; /* prepare MQD backup */ - adev->gfx.mec.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL); - if (!adev->gfx.mec.mqd_backup[i]) + adev->gfx.mec.mqd_backup[j] = kzalloc(mqd_size, GFP_KERNEL); + if (!adev->gfx.mec.mqd_backup[j]) { dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name); + return -ENOMEM; + } } } return 0; } -void amdgpu_gfx_mqd_sw_fini(struct amdgpu_device *adev) +void amdgpu_gfx_mqd_sw_fini(struct amdgpu_device *adev, int xcc_id) { struct amdgpu_ring *ring = NULL; - int i; + int i, j; + struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id]; if (adev->asic_type >= CHIP_NAVI10 && amdgpu_async_gfx_ring) { for (i = 0; i < adev->gfx.num_gfx_rings; i++) { @@ -450,43 +484,121 @@ void amdgpu_gfx_mqd_sw_fini(struct amdgpu_device *adev) } for (i = 0; i < adev->gfx.num_compute_rings; i++) { - ring = &adev->gfx.compute_ring[i]; - kfree(adev->gfx.mec.mqd_backup[i]); + j = i + xcc_id * adev->gfx.num_compute_rings; + ring = &adev->gfx.compute_ring[j]; + kfree(adev->gfx.mec.mqd_backup[j]); amdgpu_bo_free_kernel(&ring->mqd_obj, &ring->mqd_gpu_addr, &ring->mqd_ptr); } - ring = &adev->gfx.kiq.ring; - kfree(adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS]); + ring = &kiq->ring; + kfree(kiq->mqd_backup); amdgpu_bo_free_kernel(&ring->mqd_obj, &ring->mqd_gpu_addr, &ring->mqd_ptr); } -int amdgpu_gfx_disable_kcq(struct amdgpu_device *adev) +int amdgpu_gfx_disable_kcq(struct amdgpu_device *adev, int xcc_id) { - struct amdgpu_kiq *kiq = &adev->gfx.kiq; + struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id]; struct amdgpu_ring *kiq_ring = &kiq->ring; int i, r = 0; + int j; + + if (adev->enable_mes) { + for (i = 0; i < adev->gfx.num_compute_rings; i++) { + j = i + xcc_id * adev->gfx.num_compute_rings; + amdgpu_mes_unmap_legacy_queue(adev, + &adev->gfx.compute_ring[j], + RESET_QUEUES, 0, 0); + } + return 0; + } if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues) return -EINVAL; - spin_lock(&adev->gfx.kiq.ring_lock); + if (!kiq_ring->sched.ready || amdgpu_in_reset(adev)) + return 0; + + spin_lock(&kiq->ring_lock); if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size * adev->gfx.num_compute_rings)) { - spin_unlock(&adev->gfx.kiq.ring_lock); + spin_unlock(&kiq->ring_lock); return -ENOMEM; } - for (i = 0; i < adev->gfx.num_compute_rings; i++) - kiq->pmf->kiq_unmap_queues(kiq_ring, &adev->gfx.compute_ring[i], + for (i = 0; i < adev->gfx.num_compute_rings; i++) { + j = i + xcc_id * adev->gfx.num_compute_rings; + kiq->pmf->kiq_unmap_queues(kiq_ring, + &adev->gfx.compute_ring[j], RESET_QUEUES, 0, 0); + } + /* Submit unmap queue packet */ + amdgpu_ring_commit(kiq_ring); + /* + * Ring test will do a basic scratch register change check. Just run + * this to ensure that unmap queues that is submitted before got + * processed successfully before returning. + */ + r = amdgpu_ring_test_helper(kiq_ring); + + spin_unlock(&kiq->ring_lock); + + return r; +} + +int amdgpu_gfx_disable_kgq(struct amdgpu_device *adev, int xcc_id) +{ + struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id]; + struct amdgpu_ring *kiq_ring = &kiq->ring; + int i, r = 0; + int j; + + if (adev->enable_mes) { + if (amdgpu_gfx_is_master_xcc(adev, xcc_id)) { + for (i = 0; i < adev->gfx.num_gfx_rings; i++) { + j = i + xcc_id * adev->gfx.num_gfx_rings; + amdgpu_mes_unmap_legacy_queue(adev, + &adev->gfx.gfx_ring[j], + PREEMPT_QUEUES, 0, 0); + } + } + return 0; + } + + if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues) + return -EINVAL; + + if (!adev->gfx.kiq[0].ring.sched.ready || amdgpu_in_reset(adev)) + return 0; + + if (amdgpu_gfx_is_master_xcc(adev, xcc_id)) { + spin_lock(&kiq->ring_lock); + if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size * + adev->gfx.num_gfx_rings)) { + spin_unlock(&kiq->ring_lock); + return -ENOMEM; + } + + for (i = 0; i < adev->gfx.num_gfx_rings; i++) { + j = i + xcc_id * adev->gfx.num_gfx_rings; + kiq->pmf->kiq_unmap_queues(kiq_ring, + &adev->gfx.gfx_ring[j], + PREEMPT_QUEUES, 0, 0); + } + /* Submit unmap queue packet */ + amdgpu_ring_commit(kiq_ring); - if (adev->gfx.kiq.ring.sched.ready && !adev->job_hang) + /* + * Ring test will do a basic scratch register change check. + * Just run this to ensure that unmap queues that is submitted + * before got processed successfully before returning. + */ r = amdgpu_ring_test_helper(kiq_ring); - spin_unlock(&adev->gfx.kiq.ring_lock); + spin_unlock(&kiq->ring_lock); + } return r; } @@ -504,70 +616,167 @@ int amdgpu_queue_mask_bit_to_set_resource_bit(struct amdgpu_device *adev, return set_resource_bit; } -int amdgpu_gfx_enable_kcq(struct amdgpu_device *adev) +static int amdgpu_gfx_mes_enable_kcq(struct amdgpu_device *adev, int xcc_id) { - struct amdgpu_kiq *kiq = &adev->gfx.kiq; - struct amdgpu_ring *kiq_ring = &adev->gfx.kiq.ring; + struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id]; + struct amdgpu_ring *kiq_ring = &kiq->ring; + uint64_t queue_mask = ~0ULL; + int r, i, j; + + amdgpu_device_flush_hdp(adev, NULL); + + if (!adev->enable_uni_mes) { + spin_lock(&kiq->ring_lock); + r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->set_resources_size); + if (r) { + dev_err(adev->dev, "Failed to lock KIQ (%d).\n", r); + spin_unlock(&kiq->ring_lock); + return r; + } + + kiq->pmf->kiq_set_resources(kiq_ring, queue_mask); + r = amdgpu_ring_test_helper(kiq_ring); + spin_unlock(&kiq->ring_lock); + if (r) + dev_err(adev->dev, "KIQ failed to set resources\n"); + } + + for (i = 0; i < adev->gfx.num_compute_rings; i++) { + j = i + xcc_id * adev->gfx.num_compute_rings; + r = amdgpu_mes_map_legacy_queue(adev, + &adev->gfx.compute_ring[j]); + if (r) { + dev_err(adev->dev, "failed to map compute queue\n"); + return r; + } + } + + return 0; +} + +int amdgpu_gfx_enable_kcq(struct amdgpu_device *adev, int xcc_id) +{ + struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id]; + struct amdgpu_ring *kiq_ring = &kiq->ring; uint64_t queue_mask = 0; - int r, i; + int r, i, j; + + if (adev->mes.enable_legacy_queue_map) + return amdgpu_gfx_mes_enable_kcq(adev, xcc_id); if (!kiq->pmf || !kiq->pmf->kiq_map_queues || !kiq->pmf->kiq_set_resources) return -EINVAL; for (i = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; ++i) { - if (!test_bit(i, adev->gfx.mec.queue_bitmap)) + if (!test_bit(i, adev->gfx.mec_bitmap[xcc_id].queue_bitmap)) continue; /* This situation may be hit in the future if a new HW * generation exposes more than 64 queues. If so, the * definition of queue_mask needs updating */ if (WARN_ON(i > (sizeof(queue_mask)*8))) { - DRM_ERROR("Invalid KCQ enabled: %d\n", i); + dev_err(adev->dev, "Invalid KCQ enabled: %d\n", i); break; } queue_mask |= (1ull << amdgpu_queue_mask_bit_to_set_resource_bit(adev, i)); } - DRM_INFO("kiq ring mec %d pipe %d q %d\n", kiq_ring->me, kiq_ring->pipe, - kiq_ring->queue); - spin_lock(&adev->gfx.kiq.ring_lock); + amdgpu_device_flush_hdp(adev, NULL); + + dev_info(adev->dev, "kiq ring mec %d pipe %d q %d\n", kiq_ring->me, + kiq_ring->pipe, kiq_ring->queue); + + spin_lock(&kiq->ring_lock); r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size * adev->gfx.num_compute_rings + kiq->pmf->set_resources_size); if (r) { - DRM_ERROR("Failed to lock KIQ (%d).\n", r); - spin_unlock(&adev->gfx.kiq.ring_lock); + dev_err(adev->dev, "Failed to lock KIQ (%d).\n", r); + spin_unlock(&kiq->ring_lock); return r; } - if (adev->enable_mes) - queue_mask = ~0ULL; - kiq->pmf->kiq_set_resources(kiq_ring, queue_mask); - for (i = 0; i < adev->gfx.num_compute_rings; i++) - kiq->pmf->kiq_map_queues(kiq_ring, &adev->gfx.compute_ring[i]); - + for (i = 0; i < adev->gfx.num_compute_rings; i++) { + j = i + xcc_id * adev->gfx.num_compute_rings; + kiq->pmf->kiq_map_queues(kiq_ring, + &adev->gfx.compute_ring[j]); + } + /* Submit map queue packet */ + amdgpu_ring_commit(kiq_ring); + /* + * Ring test will do a basic scratch register change check. Just run + * this to ensure that map queues that is submitted before got + * processed successfully before returning. + */ r = amdgpu_ring_test_helper(kiq_ring); - spin_unlock(&adev->gfx.kiq.ring_lock); + spin_unlock(&kiq->ring_lock); if (r) - DRM_ERROR("KCQ enable failed\n"); + dev_err(adev->dev, "KCQ enable failed\n"); return r; } -/* amdgpu_gfx_off_ctrl - Handle gfx off feature enable/disable - * - * @adev: amdgpu_device pointer - * @bool enable true: enable gfx off feature, false: disable gfx off feature - * - * 1. gfx off feature will be enabled by gfx ip after gfx cg gp enabled. - * 2. other client can send request to disable gfx off feature, the request should be honored. - * 3. other client can cancel their request of disable gfx off feature - * 4. other client should not send request to enable gfx off feature before disable gfx off feature. - */ +int amdgpu_gfx_enable_kgq(struct amdgpu_device *adev, int xcc_id) +{ + struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id]; + struct amdgpu_ring *kiq_ring = &kiq->ring; + int r, i, j; -void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable) + if (!kiq->pmf || !kiq->pmf->kiq_map_queues) + return -EINVAL; + + amdgpu_device_flush_hdp(adev, NULL); + + if (adev->mes.enable_legacy_queue_map) { + for (i = 0; i < adev->gfx.num_gfx_rings; i++) { + j = i + xcc_id * adev->gfx.num_gfx_rings; + r = amdgpu_mes_map_legacy_queue(adev, + &adev->gfx.gfx_ring[j]); + if (r) { + dev_err(adev->dev, "failed to map gfx queue\n"); + return r; + } + } + + return 0; + } + + spin_lock(&kiq->ring_lock); + /* No need to map kcq on the slave */ + if (amdgpu_gfx_is_master_xcc(adev, xcc_id)) { + r = amdgpu_ring_alloc(kiq_ring, kiq->pmf->map_queues_size * + adev->gfx.num_gfx_rings); + if (r) { + dev_err(adev->dev, "Failed to lock KIQ (%d).\n", r); + spin_unlock(&kiq->ring_lock); + return r; + } + + for (i = 0; i < adev->gfx.num_gfx_rings; i++) { + j = i + xcc_id * adev->gfx.num_gfx_rings; + kiq->pmf->kiq_map_queues(kiq_ring, + &adev->gfx.gfx_ring[j]); + } + } + /* Submit map queue packet */ + amdgpu_ring_commit(kiq_ring); + /* + * Ring test will do a basic scratch register change check. Just run + * this to ensure that map queues that is submitted before got + * processed successfully before returning. + */ + r = amdgpu_ring_test_helper(kiq_ring); + spin_unlock(&kiq->ring_lock); + if (r) + dev_err(adev->dev, "KGQ enable failed\n"); + + return r; +} + +static void amdgpu_gfx_do_off_ctrl(struct amdgpu_device *adev, bool enable, + bool no_delay) { unsigned long delay = GFX_OFF_DELAY_ENABLE; @@ -589,9 +798,9 @@ void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable) if (adev->gfx.gfx_off_req_count == 0 && !adev->gfx.gfx_off_state) { /* If going to s2idle, no need to wait */ - if (adev->in_s0ix) { + if (no_delay) { if (!amdgpu_dpm_set_powergating_by_smu(adev, - AMD_IP_BLOCK_TYPE_GFX, true)) + AMD_IP_BLOCK_TYPE_GFX, true, 0)) adev->gfx.gfx_off_state = true; } else { schedule_delayed_work(&adev->gfx.gfx_off_delay_work, @@ -603,7 +812,7 @@ void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable) cancel_delayed_work_sync(&adev->gfx.gfx_off_delay_work); if (adev->gfx.gfx_off_state && - !amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false)) { + !amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false, 0)) { adev->gfx.gfx_off_state = false; if (adev->gfx.funcs->init_spm_golden) { @@ -621,6 +830,43 @@ unlock: mutex_unlock(&adev->gfx.gfx_off_mutex); } +/* amdgpu_gfx_off_ctrl - Handle gfx off feature enable/disable + * + * @adev: amdgpu_device pointer + * @bool enable true: enable gfx off feature, false: disable gfx off feature + * + * 1. gfx off feature will be enabled by gfx ip after gfx cg pg enabled. + * 2. other client can send request to disable gfx off feature, the request should be honored. + * 3. other client can cancel their request of disable gfx off feature + * 4. other client should not send request to enable gfx off feature before disable gfx off feature. + * + * gfx off allow will be delayed by GFX_OFF_DELAY_ENABLE ms. + */ +void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable) +{ + /* If going to s2idle, no need to wait */ + bool no_delay = adev->in_s0ix ? true : false; + + amdgpu_gfx_do_off_ctrl(adev, enable, no_delay); +} + +/* amdgpu_gfx_off_ctrl_immediate - Handle gfx off feature enable/disable + * + * @adev: amdgpu_device pointer + * @bool enable true: enable gfx off feature, false: disable gfx off feature + * + * 1. gfx off feature will be enabled by gfx ip after gfx cg pg enabled. + * 2. other client can send request to disable gfx off feature, the request should be honored. + * 3. other client can cancel their request of disable gfx off feature + * 4. other client should not send request to enable gfx off feature before disable gfx off feature. + * + * gfx off allow will be issued immediately. + */ +void amdgpu_gfx_off_ctrl_immediate(struct amdgpu_device *adev, bool enable) +{ + amdgpu_gfx_do_off_ctrl(adev, enable, true); +} + int amdgpu_set_gfx_off_residency(struct amdgpu_device *adev, bool value) { int r = 0; @@ -679,16 +925,24 @@ int amdgpu_gfx_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *r int r; if (amdgpu_ras_is_supported(adev, ras_block->block)) { - if (!amdgpu_persistent_edc_harvesting_supported(adev)) - amdgpu_ras_reset_error_status(adev, AMDGPU_RAS_BLOCK__GFX); + if (!amdgpu_persistent_edc_harvesting_supported(adev)) { + r = amdgpu_ras_reset_error_status(adev, AMDGPU_RAS_BLOCK__GFX); + if (r) + return r; + } r = amdgpu_ras_block_late_init(adev, ras_block); if (r) return r; - r = amdgpu_irq_get(adev, &adev->gfx.cp_ecc_error_irq, 0); - if (r) - goto late_fini; + if (amdgpu_sriov_vf(adev)) + return r; + + if (adev->gfx.cp_ecc_error_irq.funcs) { + r = amdgpu_irq_get(adev, &adev->gfx.cp_ecc_error_irq, 0); + if (r) + goto late_fini; + } } else { amdgpu_ras_feature_enable_on_boot(adev, ras_block, 0); } @@ -725,7 +979,7 @@ int amdgpu_gfx_ras_sw_init(struct amdgpu_device *adev) /* If not define special ras_late_init function, use gfx default ras_late_init */ if (!ras->ras_block.ras_late_init) - ras->ras_block.ras_late_init = amdgpu_ras_block_late_init; + ras->ras_block.ras_late_init = amdgpu_gfx_ras_late_init; /* If not defined special ras_cb function, use default ras_cb */ if (!ras->ras_block.ras_cb) @@ -777,23 +1031,42 @@ int amdgpu_gfx_cp_ecc_error_irq(struct amdgpu_device *adev, ih_data.head = *ras_if; - DRM_ERROR("CP ECC ERROR IRQ\n"); + dev_err(adev->dev, "CP ECC ERROR IRQ\n"); amdgpu_ras_interrupt_dispatch(adev, &ih_data); return 0; } -uint32_t amdgpu_kiq_rreg(struct amdgpu_device *adev, uint32_t reg) +void amdgpu_gfx_ras_error_func(struct amdgpu_device *adev, + void *ras_error_status, + void (*func)(struct amdgpu_device *adev, void *ras_error_status, + int xcc_id)) +{ + int i; + int num_xcc = adev->gfx.xcc_mask ? NUM_XCC(adev->gfx.xcc_mask) : 1; + uint32_t xcc_mask = GENMASK(num_xcc - 1, 0); + struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status; + + if (err_data) { + err_data->ue_count = 0; + err_data->ce_count = 0; + } + + for_each_inst(i, xcc_mask) + func(adev, ras_error_status, i); +} + +uint32_t amdgpu_kiq_rreg(struct amdgpu_device *adev, uint32_t reg, uint32_t xcc_id) { signed long r, cnt = 0; unsigned long flags; uint32_t seq, reg_val_offs = 0, value = 0; - struct amdgpu_kiq *kiq = &adev->gfx.kiq; + struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id]; struct amdgpu_ring *ring = &kiq->ring; if (amdgpu_device_skip_hw_access(adev)) return 0; - if (adev->mes.ring.sched.ready) + if (adev->mes.ring[0].sched.ready) return amdgpu_mes_rreg(adev, reg); BUG_ON(!ring->funcs->emit_rreg); @@ -803,7 +1076,10 @@ uint32_t amdgpu_kiq_rreg(struct amdgpu_device *adev, uint32_t reg) pr_err("critical bug! too many kiq readers\n"); goto failed_unlock; } - amdgpu_ring_alloc(ring, 32); + r = amdgpu_ring_alloc(ring, 32); + if (r) + goto failed_unlock; + amdgpu_ring_emit_rreg(ring, reg, reg_val_offs); r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT); if (r) @@ -827,6 +1103,9 @@ uint32_t amdgpu_kiq_rreg(struct amdgpu_device *adev, uint32_t reg) might_sleep(); while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) { + if (amdgpu_in_reset(adev)) + goto failed_kiq_read; + msleep(MAX_KIQ_REG_BAILOUT_INTERVAL); r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT); } @@ -850,12 +1129,12 @@ failed_kiq_read: return ~0; } -void amdgpu_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v) +void amdgpu_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v, uint32_t xcc_id) { signed long r, cnt = 0; unsigned long flags; uint32_t seq; - struct amdgpu_kiq *kiq = &adev->gfx.kiq; + struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id]; struct amdgpu_ring *ring = &kiq->ring; BUG_ON(!ring->funcs->emit_wreg); @@ -863,13 +1142,16 @@ void amdgpu_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v) if (amdgpu_device_skip_hw_access(adev)) return; - if (adev->mes.ring.sched.ready) { + if (adev->mes.ring[0].sched.ready) { amdgpu_mes_wreg(adev, reg, v); return; } spin_lock_irqsave(&kiq->ring_lock, flags); - amdgpu_ring_alloc(ring, 32); + r = amdgpu_ring_alloc(ring, 32); + if (r) + goto failed_unlock; + amdgpu_ring_emit_wreg(ring, reg, v); r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT); if (r) @@ -893,6 +1175,8 @@ void amdgpu_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v) might_sleep(); while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) { + if (amdgpu_in_reset(adev)) + goto failed_kiq_write; msleep(MAX_KIQ_REG_BAILOUT_INTERVAL); r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT); @@ -905,11 +1189,81 @@ void amdgpu_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v) failed_undo: amdgpu_ring_undo(ring); +failed_unlock: spin_unlock_irqrestore(&kiq->ring_lock, flags); failed_kiq_write: dev_err(adev->dev, "failed to write reg:%x\n", reg); } +int amdgpu_kiq_hdp_flush(struct amdgpu_device *adev) +{ + signed long r, cnt = 0; + unsigned long flags; + uint32_t seq; + struct amdgpu_kiq *kiq = &adev->gfx.kiq[0]; + struct amdgpu_ring *ring = &kiq->ring; + + if (amdgpu_device_skip_hw_access(adev)) + return 0; + + if (adev->enable_mes_kiq && adev->mes.ring[0].sched.ready) + return amdgpu_mes_hdp_flush(adev); + + if (!ring->funcs->emit_hdp_flush) { + return -EOPNOTSUPP; + } + + spin_lock_irqsave(&kiq->ring_lock, flags); + r = amdgpu_ring_alloc(ring, 32); + if (r) + goto failed_unlock; + + amdgpu_ring_emit_hdp_flush(ring); + r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT); + if (r) + goto failed_undo; + + amdgpu_ring_commit(ring); + spin_unlock_irqrestore(&kiq->ring_lock, flags); + + r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT); + + /* don't wait anymore for gpu reset case because this way may + * block gpu_recover() routine forever, e.g. this virt_kiq_rreg + * is triggered in TTM and ttm_bo_lock_delayed_workqueue() will + * never return if we keep waiting in virt_kiq_rreg, which cause + * gpu_recover() hang there. + * + * also don't wait anymore for IRQ context + * */ + if (r < 1 && (amdgpu_in_reset(adev) || in_interrupt())) + goto failed_kiq_hdp_flush; + + might_sleep(); + while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) { + if (amdgpu_in_reset(adev)) + goto failed_kiq_hdp_flush; + + msleep(MAX_KIQ_REG_BAILOUT_INTERVAL); + r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT); + } + + if (cnt > MAX_KIQ_REG_TRY) { + dev_err(adev->dev, "failed to flush HDP via KIQ timeout\n"); + return -ETIMEDOUT; + } + + return 0; + +failed_undo: + amdgpu_ring_undo(ring); +failed_unlock: + spin_unlock_irqrestore(&kiq->ring_lock, flags); +failed_kiq_hdp_flush: + dev_err(adev->dev, "failed to flush HDP via KIQ\n"); + return r < 0 ? r : -EIO; +} + int amdgpu_gfx_get_num_kcq(struct amdgpu_device *adev) { if (amdgpu_num_kcq == -1) { @@ -1049,7 +1403,8 @@ void amdgpu_gfx_cp_init_microcode(struct amdgpu_device *adev, fw_size = le32_to_cpu(cp_hdr_v2_0->data_size_bytes); break; default: - break; + dev_err(adev->dev, "Invalid ucode id %u\n", ucode_id); + return; } if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) { @@ -1059,3 +1414,1144 @@ void amdgpu_gfx_cp_init_microcode(struct amdgpu_device *adev, adev->firmware.fw_size += ALIGN(fw_size, PAGE_SIZE); } } + +bool amdgpu_gfx_is_master_xcc(struct amdgpu_device *adev, int xcc_id) +{ + return !(xcc_id % (adev->gfx.num_xcc_per_xcp ? + adev->gfx.num_xcc_per_xcp : 1)); +} + +static ssize_t amdgpu_gfx_get_current_compute_partition(struct device *dev, + struct device_attribute *addr, + char *buf) +{ + struct drm_device *ddev = dev_get_drvdata(dev); + struct amdgpu_device *adev = drm_to_adev(ddev); + int mode; + + /* Only minimal precaution taken to reject requests while in reset.*/ + if (amdgpu_in_reset(adev)) + return -EPERM; + + mode = amdgpu_xcp_query_partition_mode(adev->xcp_mgr, + AMDGPU_XCP_FL_NONE); + + return sysfs_emit(buf, "%s\n", amdgpu_gfx_compute_mode_desc(mode)); +} + +static ssize_t amdgpu_gfx_set_compute_partition(struct device *dev, + struct device_attribute *addr, + const char *buf, size_t count) +{ + struct drm_device *ddev = dev_get_drvdata(dev); + struct amdgpu_device *adev = drm_to_adev(ddev); + enum amdgpu_gfx_partition mode; + int ret = 0, num_xcc; + + num_xcc = NUM_XCC(adev->gfx.xcc_mask); + if (num_xcc % 2 != 0) + return -EINVAL; + + if (!strncasecmp("SPX", buf, strlen("SPX"))) { + mode = AMDGPU_SPX_PARTITION_MODE; + } else if (!strncasecmp("DPX", buf, strlen("DPX"))) { + /* + * DPX mode needs AIDs to be in multiple of 2. + * Each AID connects 2 XCCs. + */ + if (num_xcc%4) + return -EINVAL; + mode = AMDGPU_DPX_PARTITION_MODE; + } else if (!strncasecmp("TPX", buf, strlen("TPX"))) { + if (num_xcc != 6) + return -EINVAL; + mode = AMDGPU_TPX_PARTITION_MODE; + } else if (!strncasecmp("QPX", buf, strlen("QPX"))) { + if (num_xcc != 8) + return -EINVAL; + mode = AMDGPU_QPX_PARTITION_MODE; + } else if (!strncasecmp("CPX", buf, strlen("CPX"))) { + mode = AMDGPU_CPX_PARTITION_MODE; + } else { + return -EINVAL; + } + + /* Don't allow a switch while under reset */ + if (!down_read_trylock(&adev->reset_domain->sem)) + return -EPERM; + + ret = amdgpu_xcp_switch_partition_mode(adev->xcp_mgr, mode); + + up_read(&adev->reset_domain->sem); + + if (ret) + return ret; + + return count; +} + +static const char *xcp_desc[] = { + [AMDGPU_SPX_PARTITION_MODE] = "SPX", + [AMDGPU_DPX_PARTITION_MODE] = "DPX", + [AMDGPU_TPX_PARTITION_MODE] = "TPX", + [AMDGPU_QPX_PARTITION_MODE] = "QPX", + [AMDGPU_CPX_PARTITION_MODE] = "CPX", +}; + +static ssize_t amdgpu_gfx_get_available_compute_partition(struct device *dev, + struct device_attribute *addr, + char *buf) +{ + struct drm_device *ddev = dev_get_drvdata(dev); + struct amdgpu_device *adev = drm_to_adev(ddev); + struct amdgpu_xcp_mgr *xcp_mgr = adev->xcp_mgr; + int size = 0, mode; + char *sep = ""; + + if (!xcp_mgr || !xcp_mgr->avail_xcp_modes) + return sysfs_emit(buf, "Not supported\n"); + + for_each_inst(mode, xcp_mgr->avail_xcp_modes) { + size += sysfs_emit_at(buf, size, "%s%s", sep, xcp_desc[mode]); + sep = ", "; + } + + size += sysfs_emit_at(buf, size, "\n"); + + return size; +} + +static int amdgpu_gfx_run_cleaner_shader_job(struct amdgpu_ring *ring) +{ + struct amdgpu_device *adev = ring->adev; + struct drm_gpu_scheduler *sched = &ring->sched; + struct drm_sched_entity entity; + static atomic_t counter; + struct dma_fence *f; + struct amdgpu_job *job; + struct amdgpu_ib *ib; + void *owner; + int i, r; + + /* Initialize the scheduler entity */ + r = drm_sched_entity_init(&entity, DRM_SCHED_PRIORITY_NORMAL, + &sched, 1, NULL); + if (r) { + dev_err(adev->dev, "Failed setting up GFX kernel entity.\n"); + goto err; + } + + /* + * Use some unique dummy value as the owner to make sure we execute + * the cleaner shader on each submission. The value just need to change + * for each submission and is otherwise meaningless. + */ + owner = (void *)(unsigned long)atomic_inc_return(&counter); + + r = amdgpu_job_alloc_with_ib(ring->adev, &entity, owner, + 64, 0, &job, + AMDGPU_KERNEL_JOB_ID_CLEANER_SHADER); + if (r) + goto err; + + job->enforce_isolation = true; + /* always run the cleaner shader */ + job->run_cleaner_shader = true; + + ib = &job->ibs[0]; + for (i = 0; i <= ring->funcs->align_mask; ++i) + ib->ptr[i] = ring->funcs->nop; + ib->length_dw = ring->funcs->align_mask + 1; + + f = amdgpu_job_submit(job); + + r = dma_fence_wait(f, false); + if (r) + goto err; + + dma_fence_put(f); + + /* Clean up the scheduler entity */ + drm_sched_entity_destroy(&entity); + return 0; + +err: + return r; +} + +static int amdgpu_gfx_run_cleaner_shader(struct amdgpu_device *adev, int xcp_id) +{ + int num_xcc = NUM_XCC(adev->gfx.xcc_mask); + struct amdgpu_ring *ring; + int num_xcc_to_clear; + int i, r, xcc_id; + + if (adev->gfx.num_xcc_per_xcp) + num_xcc_to_clear = adev->gfx.num_xcc_per_xcp; + else + num_xcc_to_clear = 1; + + for (xcc_id = 0; xcc_id < num_xcc; xcc_id++) { + for (i = 0; i < adev->gfx.num_compute_rings; i++) { + ring = &adev->gfx.compute_ring[i + xcc_id * adev->gfx.num_compute_rings]; + if ((ring->xcp_id == xcp_id) && ring->sched.ready) { + r = amdgpu_gfx_run_cleaner_shader_job(ring); + if (r) + return r; + num_xcc_to_clear--; + break; + } + } + } + + if (num_xcc_to_clear) + return -ENOENT; + + return 0; +} + +/** + * amdgpu_gfx_set_run_cleaner_shader - Execute the AMDGPU GFX Cleaner Shader + * @dev: The device structure + * @attr: The device attribute structure + * @buf: The buffer containing the input data + * @count: The size of the input data + * + * Provides the sysfs interface to manually run a cleaner shader, which is + * used to clear the GPU state between different tasks. Writing a value to the + * 'run_cleaner_shader' sysfs file triggers the cleaner shader execution. + * The value written corresponds to the partition index on multi-partition + * devices. On single-partition devices, the value should be '0'. + * + * The cleaner shader clears the Local Data Store (LDS) and General Purpose + * Registers (GPRs) to ensure data isolation between GPU workloads. + * + * Return: The number of bytes written to the sysfs file. + */ +static ssize_t amdgpu_gfx_set_run_cleaner_shader(struct device *dev, + struct device_attribute *attr, + const char *buf, + size_t count) +{ + struct drm_device *ddev = dev_get_drvdata(dev); + struct amdgpu_device *adev = drm_to_adev(ddev); + int ret; + long value; + + if (amdgpu_in_reset(adev)) + return -EPERM; + if (adev->in_suspend && !adev->in_runpm) + return -EPERM; + + if (adev->gfx.disable_kq) + return -EPERM; + + ret = kstrtol(buf, 0, &value); + + if (ret) + return -EINVAL; + + if (value < 0) + return -EINVAL; + + if (adev->xcp_mgr) { + if (value >= adev->xcp_mgr->num_xcps) + return -EINVAL; + } else { + if (value > 1) + return -EINVAL; + } + + ret = pm_runtime_get_sync(ddev->dev); + if (ret < 0) { + pm_runtime_put_autosuspend(ddev->dev); + return ret; + } + + ret = amdgpu_gfx_run_cleaner_shader(adev, value); + + pm_runtime_put_autosuspend(ddev->dev); + + if (ret) + return ret; + + return count; +} + +/** + * amdgpu_gfx_get_enforce_isolation - Query AMDGPU GFX Enforce Isolation Settings + * @dev: The device structure + * @attr: The device attribute structure + * @buf: The buffer to store the output data + * + * Provides the sysfs read interface to get the current settings of the 'enforce_isolation' + * feature for each GPU partition. Reading from the 'enforce_isolation' + * sysfs file returns the isolation settings for all partitions, where '0' + * indicates disabled, '1' indicates enabled, and '2' indicates enabled in legacy mode, + * and '3' indicates enabled without cleaner shader. + * + * Return: The number of bytes read from the sysfs file. + */ +static ssize_t amdgpu_gfx_get_enforce_isolation(struct device *dev, + struct device_attribute *attr, + char *buf) +{ + struct drm_device *ddev = dev_get_drvdata(dev); + struct amdgpu_device *adev = drm_to_adev(ddev); + int i; + ssize_t size = 0; + + if (adev->xcp_mgr) { + for (i = 0; i < adev->xcp_mgr->num_xcps; i++) { + size += sysfs_emit_at(buf, size, "%u", adev->enforce_isolation[i]); + if (i < (adev->xcp_mgr->num_xcps - 1)) + size += sysfs_emit_at(buf, size, " "); + } + buf[size++] = '\n'; + } else { + size = sysfs_emit_at(buf, 0, "%u\n", adev->enforce_isolation[0]); + } + + return size; +} + +/** + * amdgpu_gfx_set_enforce_isolation - Control AMDGPU GFX Enforce Isolation + * @dev: The device structure + * @attr: The device attribute structure + * @buf: The buffer containing the input data + * @count: The size of the input data + * + * This function allows control over the 'enforce_isolation' feature, which + * serializes access to the graphics engine. Writing '0' to disable, '1' to + * enable isolation with cleaner shader, '2' to enable legacy isolation without + * cleaner shader, or '3' to enable process isolation without submitting the + * cleaner shader to the 'enforce_isolation' sysfs file sets the isolation mode + * for each partition. The input should specify the setting for all + * partitions. + * + * Return: The number of bytes written to the sysfs file. + */ +static ssize_t amdgpu_gfx_set_enforce_isolation(struct device *dev, + struct device_attribute *attr, + const char *buf, size_t count) +{ + struct drm_device *ddev = dev_get_drvdata(dev); + struct amdgpu_device *adev = drm_to_adev(ddev); + long partition_values[MAX_XCP] = {0}; + int ret, i, num_partitions; + const char *input_buf = buf; + + for (i = 0; i < (adev->xcp_mgr ? adev->xcp_mgr->num_xcps : 1); i++) { + ret = sscanf(input_buf, "%ld", &partition_values[i]); + if (ret <= 0) + break; + + /* Move the pointer to the next value in the string */ + input_buf = strchr(input_buf, ' '); + if (input_buf) { + input_buf++; + } else { + i++; + break; + } + } + num_partitions = i; + + if (adev->xcp_mgr && num_partitions != adev->xcp_mgr->num_xcps) + return -EINVAL; + + if (!adev->xcp_mgr && num_partitions != 1) + return -EINVAL; + + for (i = 0; i < num_partitions; i++) { + if (partition_values[i] != 0 && + partition_values[i] != 1 && + partition_values[i] != 2 && + partition_values[i] != 3) + return -EINVAL; + } + + mutex_lock(&adev->enforce_isolation_mutex); + for (i = 0; i < num_partitions; i++) { + switch (partition_values[i]) { + case 0: + default: + adev->enforce_isolation[i] = AMDGPU_ENFORCE_ISOLATION_DISABLE; + break; + case 1: + adev->enforce_isolation[i] = + AMDGPU_ENFORCE_ISOLATION_ENABLE; + break; + case 2: + adev->enforce_isolation[i] = + AMDGPU_ENFORCE_ISOLATION_ENABLE_LEGACY; + break; + case 3: + adev->enforce_isolation[i] = + AMDGPU_ENFORCE_ISOLATION_NO_CLEANER_SHADER; + break; + } + } + mutex_unlock(&adev->enforce_isolation_mutex); + + amdgpu_mes_update_enforce_isolation(adev); + + return count; +} + +static ssize_t amdgpu_gfx_get_gfx_reset_mask(struct device *dev, + struct device_attribute *attr, + char *buf) +{ + struct drm_device *ddev = dev_get_drvdata(dev); + struct amdgpu_device *adev = drm_to_adev(ddev); + + if (!adev) + return -ENODEV; + + return amdgpu_show_reset_mask(buf, adev->gfx.gfx_supported_reset); +} + +static ssize_t amdgpu_gfx_get_compute_reset_mask(struct device *dev, + struct device_attribute *attr, + char *buf) +{ + struct drm_device *ddev = dev_get_drvdata(dev); + struct amdgpu_device *adev = drm_to_adev(ddev); + + if (!adev) + return -ENODEV; + + return amdgpu_show_reset_mask(buf, adev->gfx.compute_supported_reset); +} + +static DEVICE_ATTR(run_cleaner_shader, 0200, + NULL, amdgpu_gfx_set_run_cleaner_shader); + +static DEVICE_ATTR(enforce_isolation, 0644, + amdgpu_gfx_get_enforce_isolation, + amdgpu_gfx_set_enforce_isolation); + +static DEVICE_ATTR(current_compute_partition, 0644, + amdgpu_gfx_get_current_compute_partition, + amdgpu_gfx_set_compute_partition); + +static DEVICE_ATTR(available_compute_partition, 0444, + amdgpu_gfx_get_available_compute_partition, NULL); +static DEVICE_ATTR(gfx_reset_mask, 0444, + amdgpu_gfx_get_gfx_reset_mask, NULL); + +static DEVICE_ATTR(compute_reset_mask, 0444, + amdgpu_gfx_get_compute_reset_mask, NULL); + +static int amdgpu_gfx_sysfs_xcp_init(struct amdgpu_device *adev) +{ + struct amdgpu_xcp_mgr *xcp_mgr = adev->xcp_mgr; + bool xcp_switch_supported; + int r; + + if (!xcp_mgr) + return 0; + + xcp_switch_supported = + (xcp_mgr->funcs && xcp_mgr->funcs->switch_partition_mode); + + if (!xcp_switch_supported) + dev_attr_current_compute_partition.attr.mode &= + ~(S_IWUSR | S_IWGRP | S_IWOTH); + + r = device_create_file(adev->dev, &dev_attr_current_compute_partition); + if (r) + return r; + + if (xcp_switch_supported) + r = device_create_file(adev->dev, + &dev_attr_available_compute_partition); + + return r; +} + +static void amdgpu_gfx_sysfs_xcp_fini(struct amdgpu_device *adev) +{ + struct amdgpu_xcp_mgr *xcp_mgr = adev->xcp_mgr; + bool xcp_switch_supported; + + if (!xcp_mgr) + return; + + xcp_switch_supported = + (xcp_mgr->funcs && xcp_mgr->funcs->switch_partition_mode); + device_remove_file(adev->dev, &dev_attr_current_compute_partition); + + if (xcp_switch_supported) + device_remove_file(adev->dev, + &dev_attr_available_compute_partition); +} + +static int amdgpu_gfx_sysfs_isolation_shader_init(struct amdgpu_device *adev) +{ + int r; + + r = device_create_file(adev->dev, &dev_attr_enforce_isolation); + if (r) + return r; + if (adev->gfx.enable_cleaner_shader) + r = device_create_file(adev->dev, &dev_attr_run_cleaner_shader); + + return r; +} + +static void amdgpu_gfx_sysfs_isolation_shader_fini(struct amdgpu_device *adev) +{ + device_remove_file(adev->dev, &dev_attr_enforce_isolation); + if (adev->gfx.enable_cleaner_shader) + device_remove_file(adev->dev, &dev_attr_run_cleaner_shader); +} + +static int amdgpu_gfx_sysfs_reset_mask_init(struct amdgpu_device *adev) +{ + int r = 0; + + if (!amdgpu_gpu_recovery) + return r; + + if (adev->gfx.num_gfx_rings) { + r = device_create_file(adev->dev, &dev_attr_gfx_reset_mask); + if (r) + return r; + } + + if (adev->gfx.num_compute_rings) { + r = device_create_file(adev->dev, &dev_attr_compute_reset_mask); + if (r) + return r; + } + + return r; +} + +static void amdgpu_gfx_sysfs_reset_mask_fini(struct amdgpu_device *adev) +{ + if (!amdgpu_gpu_recovery) + return; + + if (adev->gfx.num_gfx_rings) + device_remove_file(adev->dev, &dev_attr_gfx_reset_mask); + + if (adev->gfx.num_compute_rings) + device_remove_file(adev->dev, &dev_attr_compute_reset_mask); +} + +int amdgpu_gfx_sysfs_init(struct amdgpu_device *adev) +{ + int r; + + r = amdgpu_gfx_sysfs_xcp_init(adev); + if (r) { + dev_err(adev->dev, "failed to create xcp sysfs files"); + return r; + } + + r = amdgpu_gfx_sysfs_isolation_shader_init(adev); + if (r) + dev_err(adev->dev, "failed to create isolation sysfs files"); + + r = amdgpu_gfx_sysfs_reset_mask_init(adev); + if (r) + dev_err(adev->dev, "failed to create reset mask sysfs files"); + + return r; +} + +void amdgpu_gfx_sysfs_fini(struct amdgpu_device *adev) +{ + if (adev->dev->kobj.sd) { + amdgpu_gfx_sysfs_xcp_fini(adev); + amdgpu_gfx_sysfs_isolation_shader_fini(adev); + amdgpu_gfx_sysfs_reset_mask_fini(adev); + } +} + +int amdgpu_gfx_cleaner_shader_sw_init(struct amdgpu_device *adev, + unsigned int cleaner_shader_size) +{ + if (!adev->gfx.enable_cleaner_shader) + return -EOPNOTSUPP; + + return amdgpu_bo_create_kernel(adev, cleaner_shader_size, PAGE_SIZE, + AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT, + &adev->gfx.cleaner_shader_obj, + &adev->gfx.cleaner_shader_gpu_addr, + (void **)&adev->gfx.cleaner_shader_cpu_ptr); +} + +void amdgpu_gfx_cleaner_shader_sw_fini(struct amdgpu_device *adev) +{ + if (!adev->gfx.enable_cleaner_shader) + return; + + amdgpu_bo_free_kernel(&adev->gfx.cleaner_shader_obj, + &adev->gfx.cleaner_shader_gpu_addr, + (void **)&adev->gfx.cleaner_shader_cpu_ptr); +} + +void amdgpu_gfx_cleaner_shader_init(struct amdgpu_device *adev, + unsigned int cleaner_shader_size, + const void *cleaner_shader_ptr) +{ + if (!adev->gfx.enable_cleaner_shader) + return; + + if (adev->gfx.cleaner_shader_cpu_ptr && cleaner_shader_ptr) + memcpy_toio(adev->gfx.cleaner_shader_cpu_ptr, cleaner_shader_ptr, + cleaner_shader_size); +} + +/** + * amdgpu_gfx_kfd_sch_ctrl - Control the KFD scheduler from the KGD (Graphics Driver) + * @adev: amdgpu_device pointer + * @idx: Index of the scheduler to control + * @enable: Whether to enable or disable the KFD scheduler + * + * This function is used to control the KFD (Kernel Fusion Driver) scheduler + * from the KGD. It is part of the cleaner shader feature. This function plays + * a key role in enforcing process isolation on the GPU. + * + * The function uses a reference count mechanism (kfd_sch_req_count) to keep + * track of the number of requests to enable the KFD scheduler. When a request + * to enable the KFD scheduler is made, the reference count is decremented. + * When the reference count reaches zero, a delayed work is scheduled to + * enforce isolation after a delay of GFX_SLICE_PERIOD. + * + * When a request to disable the KFD scheduler is made, the function first + * checks if the reference count is zero. If it is, it cancels the delayed work + * for enforcing isolation and checks if the KFD scheduler is active. If the + * KFD scheduler is active, it sends a request to stop the KFD scheduler and + * sets the KFD scheduler state to inactive. Then, it increments the reference + * count. + * + * The function is synchronized using the kfd_sch_mutex to ensure that the KFD + * scheduler state and reference count are updated atomically. + * + * Note: If the reference count is already zero when a request to enable the + * KFD scheduler is made, it means there's an imbalance bug somewhere. The + * function triggers a warning in this case. + */ +static void amdgpu_gfx_kfd_sch_ctrl(struct amdgpu_device *adev, u32 idx, + bool enable) +{ + mutex_lock(&adev->gfx.userq_sch_mutex); + + if (enable) { + /* If the count is already 0, it means there's an imbalance bug somewhere. + * Note that the bug may be in a different caller than the one which triggers the + * WARN_ON_ONCE. + */ + if (WARN_ON_ONCE(adev->gfx.userq_sch_req_count[idx] == 0)) { + dev_err(adev->dev, "Attempted to enable KFD scheduler when reference count is already zero\n"); + goto unlock; + } + + adev->gfx.userq_sch_req_count[idx]--; + + if (adev->gfx.userq_sch_req_count[idx] == 0 && + adev->gfx.userq_sch_inactive[idx]) { + schedule_delayed_work(&adev->gfx.enforce_isolation[idx].work, + msecs_to_jiffies(adev->gfx.enforce_isolation_time[idx])); + } + } else { + if (adev->gfx.userq_sch_req_count[idx] == 0) { + cancel_delayed_work_sync(&adev->gfx.enforce_isolation[idx].work); + if (!adev->gfx.userq_sch_inactive[idx]) { + amdgpu_userq_stop_sched_for_enforce_isolation(adev, idx); + if (adev->kfd.init_complete) + amdgpu_amdkfd_stop_sched(adev, idx); + adev->gfx.userq_sch_inactive[idx] = true; + } + } + + adev->gfx.userq_sch_req_count[idx]++; + } + +unlock: + mutex_unlock(&adev->gfx.userq_sch_mutex); +} + +/** + * amdgpu_gfx_enforce_isolation_handler - work handler for enforcing shader isolation + * + * @work: work_struct. + * + * This function is the work handler for enforcing shader isolation on AMD GPUs. + * It counts the number of emitted fences for each GFX and compute ring. If there + * are any fences, it schedules the `enforce_isolation_work` to be run after a + * delay of `GFX_SLICE_PERIOD`. If there are no fences, it signals the Kernel Fusion + * Driver (KFD) to resume the runqueue. The function is synchronized using the + * `enforce_isolation_mutex`. + */ +void amdgpu_gfx_enforce_isolation_handler(struct work_struct *work) +{ + struct amdgpu_isolation_work *isolation_work = + container_of(work, struct amdgpu_isolation_work, work.work); + struct amdgpu_device *adev = isolation_work->adev; + u32 i, idx, fences = 0; + + if (isolation_work->xcp_id == AMDGPU_XCP_NO_PARTITION) + idx = 0; + else + idx = isolation_work->xcp_id; + + if (idx >= MAX_XCP) + return; + + mutex_lock(&adev->enforce_isolation_mutex); + for (i = 0; i < AMDGPU_MAX_GFX_RINGS; ++i) { + if (isolation_work->xcp_id == adev->gfx.gfx_ring[i].xcp_id) + fences += amdgpu_fence_count_emitted(&adev->gfx.gfx_ring[i]); + } + for (i = 0; i < (AMDGPU_MAX_COMPUTE_RINGS * AMDGPU_MAX_GC_INSTANCES); ++i) { + if (isolation_work->xcp_id == adev->gfx.compute_ring[i].xcp_id) + fences += amdgpu_fence_count_emitted(&adev->gfx.compute_ring[i]); + } + if (fences) { + /* we've already had our timeslice, so let's wrap this up */ + schedule_delayed_work(&adev->gfx.enforce_isolation[idx].work, + msecs_to_jiffies(1)); + } else { + /* Tell KFD to resume the runqueue */ + WARN_ON_ONCE(!adev->gfx.userq_sch_inactive[idx]); + WARN_ON_ONCE(adev->gfx.userq_sch_req_count[idx]); + + amdgpu_userq_start_sched_for_enforce_isolation(adev, idx); + if (adev->kfd.init_complete) + amdgpu_amdkfd_start_sched(adev, idx); + adev->gfx.userq_sch_inactive[idx] = false; + } + mutex_unlock(&adev->enforce_isolation_mutex); +} + +/** + * amdgpu_gfx_enforce_isolation_wait_for_kfd - Manage KFD wait period for process isolation + * @adev: amdgpu_device pointer + * @idx: Index of the GPU partition + * + * When kernel submissions come in, the jobs are given a time slice and once + * that time slice is up, if there are KFD user queues active, kernel + * submissions are blocked until KFD has had its time slice. Once the KFD time + * slice is up, KFD user queues are preempted and kernel submissions are + * unblocked and allowed to run again. + */ +static void +amdgpu_gfx_enforce_isolation_wait_for_kfd(struct amdgpu_device *adev, + u32 idx) +{ + unsigned long cjiffies; + bool wait = false; + + mutex_lock(&adev->enforce_isolation_mutex); + if (adev->enforce_isolation[idx] == AMDGPU_ENFORCE_ISOLATION_ENABLE) { + /* set the initial values if nothing is set */ + if (!adev->gfx.enforce_isolation_jiffies[idx]) { + adev->gfx.enforce_isolation_jiffies[idx] = jiffies; + adev->gfx.enforce_isolation_time[idx] = GFX_SLICE_PERIOD_MS; + } + /* Make sure KFD gets a chance to run */ + if (amdgpu_amdkfd_compute_active(adev, idx)) { + cjiffies = jiffies; + if (time_after(cjiffies, adev->gfx.enforce_isolation_jiffies[idx])) { + cjiffies -= adev->gfx.enforce_isolation_jiffies[idx]; + if ((jiffies_to_msecs(cjiffies) >= GFX_SLICE_PERIOD_MS)) { + /* if our time is up, let KGD work drain before scheduling more */ + wait = true; + /* reset the timer period */ + adev->gfx.enforce_isolation_time[idx] = GFX_SLICE_PERIOD_MS; + } else { + /* set the timer period to what's left in our time slice */ + adev->gfx.enforce_isolation_time[idx] = + GFX_SLICE_PERIOD_MS - jiffies_to_msecs(cjiffies); + } + } else { + /* if jiffies wrap around we will just wait a little longer */ + adev->gfx.enforce_isolation_jiffies[idx] = jiffies; + } + } else { + /* if there is no KFD work, then set the full slice period */ + adev->gfx.enforce_isolation_jiffies[idx] = jiffies; + adev->gfx.enforce_isolation_time[idx] = GFX_SLICE_PERIOD_MS; + } + } + mutex_unlock(&adev->enforce_isolation_mutex); + + if (wait) + msleep(GFX_SLICE_PERIOD_MS); +} + +/** + * amdgpu_gfx_enforce_isolation_ring_begin_use - Begin use of a ring with enforced isolation + * @ring: Pointer to the amdgpu_ring structure + * + * Ring begin_use helper implementation for gfx which serializes access to the + * gfx IP between kernel submission IOCTLs and KFD user queues when isolation + * enforcement is enabled. The kernel submission IOCTLs and KFD user queues + * each get a time slice when both are active. + */ +void amdgpu_gfx_enforce_isolation_ring_begin_use(struct amdgpu_ring *ring) +{ + struct amdgpu_device *adev = ring->adev; + u32 idx; + bool sched_work = false; + + if (!adev->gfx.enable_cleaner_shader) + return; + + if (ring->xcp_id == AMDGPU_XCP_NO_PARTITION) + idx = 0; + else + idx = ring->xcp_id; + + if (idx >= MAX_XCP) + return; + + /* Don't submit more work until KFD has had some time */ + amdgpu_gfx_enforce_isolation_wait_for_kfd(adev, idx); + + mutex_lock(&adev->enforce_isolation_mutex); + if (adev->enforce_isolation[idx] == AMDGPU_ENFORCE_ISOLATION_ENABLE) { + if (adev->kfd.init_complete) + sched_work = true; + } + mutex_unlock(&adev->enforce_isolation_mutex); + + if (sched_work) + amdgpu_gfx_kfd_sch_ctrl(adev, idx, false); +} + +/** + * amdgpu_gfx_enforce_isolation_ring_end_use - End use of a ring with enforced isolation + * @ring: Pointer to the amdgpu_ring structure + * + * Ring end_use helper implementation for gfx which serializes access to the + * gfx IP between kernel submission IOCTLs and KFD user queues when isolation + * enforcement is enabled. The kernel submission IOCTLs and KFD user queues + * each get a time slice when both are active. + */ +void amdgpu_gfx_enforce_isolation_ring_end_use(struct amdgpu_ring *ring) +{ + struct amdgpu_device *adev = ring->adev; + u32 idx; + bool sched_work = false; + + if (!adev->gfx.enable_cleaner_shader) + return; + + if (ring->xcp_id == AMDGPU_XCP_NO_PARTITION) + idx = 0; + else + idx = ring->xcp_id; + + if (idx >= MAX_XCP) + return; + + mutex_lock(&adev->enforce_isolation_mutex); + if (adev->enforce_isolation[idx] == AMDGPU_ENFORCE_ISOLATION_ENABLE) { + if (adev->kfd.init_complete) + sched_work = true; + } + mutex_unlock(&adev->enforce_isolation_mutex); + + if (sched_work) + amdgpu_gfx_kfd_sch_ctrl(adev, idx, true); +} + +void amdgpu_gfx_profile_idle_work_handler(struct work_struct *work) +{ + struct amdgpu_device *adev = + container_of(work, struct amdgpu_device, gfx.idle_work.work); + enum PP_SMC_POWER_PROFILE profile; + u32 i, fences = 0; + int r; + + if (adev->gfx.num_gfx_rings) + profile = PP_SMC_POWER_PROFILE_FULLSCREEN3D; + else + profile = PP_SMC_POWER_PROFILE_COMPUTE; + + for (i = 0; i < AMDGPU_MAX_GFX_RINGS; ++i) + fences += amdgpu_fence_count_emitted(&adev->gfx.gfx_ring[i]); + for (i = 0; i < (AMDGPU_MAX_COMPUTE_RINGS * AMDGPU_MAX_GC_INSTANCES); ++i) + fences += amdgpu_fence_count_emitted(&adev->gfx.compute_ring[i]); + if (!fences && !atomic_read(&adev->gfx.total_submission_cnt)) { + mutex_lock(&adev->gfx.workload_profile_mutex); + if (adev->gfx.workload_profile_active) { + r = amdgpu_dpm_switch_power_profile(adev, profile, false); + if (r) + dev_warn(adev->dev, "(%d) failed to disable %s power profile mode\n", r, + profile == PP_SMC_POWER_PROFILE_FULLSCREEN3D ? + "fullscreen 3D" : "compute"); + adev->gfx.workload_profile_active = false; + } + mutex_unlock(&adev->gfx.workload_profile_mutex); + } else { + schedule_delayed_work(&adev->gfx.idle_work, GFX_PROFILE_IDLE_TIMEOUT); + } +} + +void amdgpu_gfx_profile_ring_begin_use(struct amdgpu_ring *ring) +{ + struct amdgpu_device *adev = ring->adev; + enum PP_SMC_POWER_PROFILE profile; + int r; + + if (amdgpu_dpm_is_overdrive_enabled(adev)) + return; + + if (adev->gfx.num_gfx_rings) + profile = PP_SMC_POWER_PROFILE_FULLSCREEN3D; + else + profile = PP_SMC_POWER_PROFILE_COMPUTE; + + atomic_inc(&adev->gfx.total_submission_cnt); + + cancel_delayed_work_sync(&adev->gfx.idle_work); + + /* We can safely return early here because we've cancelled the + * the delayed work so there is no one else to set it to false + * and we don't care if someone else sets it to true. + */ + if (adev->gfx.workload_profile_active) + return; + + mutex_lock(&adev->gfx.workload_profile_mutex); + if (!adev->gfx.workload_profile_active) { + r = amdgpu_dpm_switch_power_profile(adev, profile, true); + if (r) + dev_warn(adev->dev, "(%d) failed to disable %s power profile mode\n", r, + profile == PP_SMC_POWER_PROFILE_FULLSCREEN3D ? + "fullscreen 3D" : "compute"); + adev->gfx.workload_profile_active = true; + } + mutex_unlock(&adev->gfx.workload_profile_mutex); +} + +void amdgpu_gfx_profile_ring_end_use(struct amdgpu_ring *ring) +{ + struct amdgpu_device *adev = ring->adev; + + if (amdgpu_dpm_is_overdrive_enabled(adev)) + return; + + atomic_dec(&ring->adev->gfx.total_submission_cnt); + + schedule_delayed_work(&ring->adev->gfx.idle_work, GFX_PROFILE_IDLE_TIMEOUT); +} + +/** + * amdgpu_gfx_csb_preamble_start - Set CSB preamble start + * + * @buffer: This is an output variable that gets the PACKET3 preamble setup. + * + * Return: + * return the latest index. + */ +u32 amdgpu_gfx_csb_preamble_start(u32 *buffer) +{ + u32 count = 0; + + buffer[count++] = cpu_to_le32(PACKET3(PACKET3_PREAMBLE_CNTL, 0)); + buffer[count++] = cpu_to_le32(PACKET3_PREAMBLE_BEGIN_CLEAR_STATE); + + buffer[count++] = cpu_to_le32(PACKET3(PACKET3_CONTEXT_CONTROL, 1)); + buffer[count++] = cpu_to_le32(0x80000000); + buffer[count++] = cpu_to_le32(0x80000000); + + return count; +} + +/** + * amdgpu_gfx_csb_data_parser - Parser CS data + * + * @adev: amdgpu_device pointer used to get the CS data and other gfx info. + * @buffer: This is an output variable that gets the PACKET3 preamble end. + * @count: Index to start set the preemble end. + * + * Return: + * return the latest index. + */ +u32 amdgpu_gfx_csb_data_parser(struct amdgpu_device *adev, u32 *buffer, u32 count) +{ + const struct cs_section_def *sect = NULL; + const struct cs_extent_def *ext = NULL; + u32 i; + + for (sect = adev->gfx.rlc.cs_data; sect->section != NULL; ++sect) { + for (ext = sect->section; ext->extent != NULL; ++ext) { + if (sect->id == SECT_CONTEXT) { + buffer[count++] = cpu_to_le32(PACKET3(PACKET3_SET_CONTEXT_REG, ext->reg_count)); + buffer[count++] = cpu_to_le32(ext->reg_index - PACKET3_SET_CONTEXT_REG_START); + + for (i = 0; i < ext->reg_count; i++) + buffer[count++] = cpu_to_le32(ext->extent[i]); + } + } + } + + return count; +} + +/** + * amdgpu_gfx_csb_preamble_end - Set CSB preamble end + * + * @buffer: This is an output variable that gets the PACKET3 preamble end. + * @count: Index to start set the preemble end. + */ +void amdgpu_gfx_csb_preamble_end(u32 *buffer, u32 count) +{ + buffer[count++] = cpu_to_le32(PACKET3(PACKET3_PREAMBLE_CNTL, 0)); + buffer[count++] = cpu_to_le32(PACKET3_PREAMBLE_END_CLEAR_STATE); + + buffer[count++] = cpu_to_le32(PACKET3(PACKET3_CLEAR_STATE, 0)); + buffer[count++] = cpu_to_le32(0); +} + +/* + * debugfs for to enable/disable gfx job submission to specific core. + */ +#if defined(CONFIG_DEBUG_FS) +static int amdgpu_debugfs_gfx_sched_mask_set(void *data, u64 val) +{ + struct amdgpu_device *adev = (struct amdgpu_device *)data; + u32 i; + u64 mask = 0; + struct amdgpu_ring *ring; + + if (!adev) + return -ENODEV; + + mask = (1ULL << adev->gfx.num_gfx_rings) - 1; + if ((val & mask) == 0) + return -EINVAL; + + for (i = 0; i < adev->gfx.num_gfx_rings; ++i) { + ring = &adev->gfx.gfx_ring[i]; + if (val & (1 << i)) + ring->sched.ready = true; + else + ring->sched.ready = false; + } + /* publish sched.ready flag update effective immediately across smp */ + smp_rmb(); + return 0; +} + +static int amdgpu_debugfs_gfx_sched_mask_get(void *data, u64 *val) +{ + struct amdgpu_device *adev = (struct amdgpu_device *)data; + u32 i; + u64 mask = 0; + struct amdgpu_ring *ring; + + if (!adev) + return -ENODEV; + for (i = 0; i < adev->gfx.num_gfx_rings; ++i) { + ring = &adev->gfx.gfx_ring[i]; + if (ring->sched.ready) + mask |= 1ULL << i; + } + + *val = mask; + return 0; +} + +DEFINE_DEBUGFS_ATTRIBUTE(amdgpu_debugfs_gfx_sched_mask_fops, + amdgpu_debugfs_gfx_sched_mask_get, + amdgpu_debugfs_gfx_sched_mask_set, "%llx\n"); + +#endif + +void amdgpu_debugfs_gfx_sched_mask_init(struct amdgpu_device *adev) +{ +#if defined(CONFIG_DEBUG_FS) + struct drm_minor *minor = adev_to_drm(adev)->primary; + struct dentry *root = minor->debugfs_root; + char name[32]; + + if (!(adev->gfx.num_gfx_rings > 1)) + return; + sprintf(name, "amdgpu_gfx_sched_mask"); + debugfs_create_file(name, 0600, root, adev, + &amdgpu_debugfs_gfx_sched_mask_fops); +#endif +} + +/* + * debugfs for to enable/disable compute job submission to specific core. + */ +#if defined(CONFIG_DEBUG_FS) +static int amdgpu_debugfs_compute_sched_mask_set(void *data, u64 val) +{ + struct amdgpu_device *adev = (struct amdgpu_device *)data; + u32 i; + u64 mask = 0; + struct amdgpu_ring *ring; + + if (!adev) + return -ENODEV; + + mask = (1ULL << adev->gfx.num_compute_rings) - 1; + if ((val & mask) == 0) + return -EINVAL; + + for (i = 0; i < adev->gfx.num_compute_rings; ++i) { + ring = &adev->gfx.compute_ring[i]; + if (val & (1 << i)) + ring->sched.ready = true; + else + ring->sched.ready = false; + } + + /* publish sched.ready flag update effective immediately across smp */ + smp_rmb(); + return 0; +} + +static int amdgpu_debugfs_compute_sched_mask_get(void *data, u64 *val) +{ + struct amdgpu_device *adev = (struct amdgpu_device *)data; + u32 i; + u64 mask = 0; + struct amdgpu_ring *ring; + + if (!adev) + return -ENODEV; + for (i = 0; i < adev->gfx.num_compute_rings; ++i) { + ring = &adev->gfx.compute_ring[i]; + if (ring->sched.ready) + mask |= 1ULL << i; + } + + *val = mask; + return 0; +} + +DEFINE_DEBUGFS_ATTRIBUTE(amdgpu_debugfs_compute_sched_mask_fops, + amdgpu_debugfs_compute_sched_mask_get, + amdgpu_debugfs_compute_sched_mask_set, "%llx\n"); + +#endif + +void amdgpu_debugfs_compute_sched_mask_init(struct amdgpu_device *adev) +{ +#if defined(CONFIG_DEBUG_FS) + struct drm_minor *minor = adev_to_drm(adev)->primary; + struct dentry *root = minor->debugfs_root; + char name[32]; + + if (!(adev->gfx.num_compute_rings > 1)) + return; + sprintf(name, "amdgpu_compute_sched_mask"); + debugfs_create_file(name, 0600, root, adev, + &amdgpu_debugfs_compute_sched_mask_fops); +#endif +} + |