[PATCH 18/21] drm/amdgpu: add framework for HW specific priority settings v3
Christian König
deathsimple at vodafone.de
Fri Mar 3 13:46:09 UTC 2017
Am 02.03.2017 um 22:44 schrieb Andres Rodriguez:
> Add an initial framework for changing the HW priorities of rings. The
> framework allows requesting priority changes for the lifetime of an
> amdgpu_job. After the job completes the priority will decay to the next
> lowest priority for which a request is still valid.
>
> A new ring function set_priority() can now be populated to take care of
> the HW specific programming sequence for priority changes.
>
> v2: set priority before emitting IB, and take a ref on amdgpu_job
> v3: use AMD_SCHED_PRIORITY_* instead of AMDGPU_CTX_PRIORITY_*
> v4: plug amdgpu_ring_restore_priority_cb into amdgpu_job_free_cb
>
> Signed-off-by: Andres Rodriguez <andresx7 at gmail.com>
> ---
> drivers/gpu/drm/amd/amdgpu/amdgpu.h | 3 +-
> drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 2 +-
> drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 14 +++++-
> drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c | 73 ++++++++++++++++++++++++++++++++
> drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 12 ++++++
> 5 files changed, 100 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> index bf15373..e7e92fd 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> @@ -636,21 +636,21 @@ struct amdgpu_flip_work {
> struct amdgpu_ib {
> struct amdgpu_sa_bo *sa_bo;
> uint32_t length_dw;
> uint64_t gpu_addr;
> uint32_t *ptr;
> uint32_t flags;
> };
>
> extern const struct amd_sched_backend_ops amdgpu_sched_ops;
>
> -int amdgpu_job_alloc(struct amdgpu_device *adev, unsigned num_ibs,
> +int amdgpu_job_alloc(struct amdgpu_device *adev, unsigned num_ibs, int priority,
> struct amdgpu_job **job, struct amdgpu_vm *vm);
> int amdgpu_job_alloc_with_ib(struct amdgpu_device *adev, unsigned size,
> struct amdgpu_job **job);
>
> void amdgpu_job_free_resources(struct amdgpu_job *job);
> void amdgpu_job_free(struct amdgpu_job *job);
> int amdgpu_job_submit(struct amdgpu_job *job, struct amdgpu_ring *ring,
> struct amd_sched_entity *entity, void *owner,
> struct dma_fence **f);
>
> @@ -981,20 +981,21 @@ struct amdgpu_cs_parser {
> #define AMDGPU_VM_DOMAIN (1 << 3) /* bit set means in virtual memory context */
>
> struct amdgpu_job {
> struct amd_sched_job base;
> struct amdgpu_device *adev;
> struct amdgpu_vm *vm;
> struct amdgpu_ring *ring;
> struct amdgpu_sync sync;
> struct amdgpu_ib *ibs;
> struct dma_fence *fence; /* the hw fence */
> + int priority;
Actually that is unnecessary (and thinking more about it the priority in
the context is unnecessary as well).
(job->base.sched->sched_rq - job->base.s_entity->rq) should give you the
scheduler priority of the job as well.
We can add an inline helper in the scheduler for that.
Additional to that please use enum amd_sched_priority instead of int for
the priority all over the patch.
> uint32_t preamble_status;
> uint32_t num_ibs;
> void *owner;
> uint64_t fence_ctx; /* the fence_context this job uses */
> bool vm_needs_flush;
> unsigned vm_id;
> uint64_t vm_pd_addr;
> uint32_t gds_base, gds_size;
> uint32_t gws_base, gws_size;
> uint32_t oa_base, oa_size;
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
> index 605d40e..19ce202 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
> @@ -179,21 +179,21 @@ int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
>
> case AMDGPU_CHUNK_ID_DEPENDENCIES:
> break;
>
> default:
> ret = -EINVAL;
> goto free_partial_kdata;
> }
> }
>
> - ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
> + ret = amdgpu_job_alloc(p->adev, num_ibs, p->ctx->priority, &p->job, vm);
> if (ret)
> goto free_all_kdata;
>
> if (p->uf_entry.robj)
> p->job->uf_addr = uf_offset;
> kfree(chunk_array);
> return 0;
>
> free_all_kdata:
> i = p->nchunks - 1;
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
> index 86a1242..28f8a50 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
> @@ -32,50 +32,51 @@ static void amdgpu_job_timedout(struct amd_sched_job *s_job)
> {
> struct amdgpu_job *job = container_of(s_job, struct amdgpu_job, base);
>
> DRM_ERROR("ring %s timeout, last signaled seq=%u, last emitted seq=%u\n",
> job->base.sched->name,
> atomic_read(&job->ring->fence_drv.last_seq),
> job->ring->fence_drv.sync_seq);
> amdgpu_gpu_reset(job->adev);
> }
>
> -int amdgpu_job_alloc(struct amdgpu_device *adev, unsigned num_ibs,
> +int amdgpu_job_alloc(struct amdgpu_device *adev, unsigned num_ibs, int priority,
> struct amdgpu_job **job, struct amdgpu_vm *vm)
> {
> size_t size = sizeof(struct amdgpu_job);
>
> if (num_ibs == 0)
> return -EINVAL;
>
> size += sizeof(struct amdgpu_ib) * num_ibs;
>
> *job = kzalloc(size, GFP_KERNEL);
> if (!*job)
> return -ENOMEM;
>
> (*job)->adev = adev;
> (*job)->vm = vm;
> + (*job)->priority = priority;
> (*job)->ibs = (void *)&(*job)[1];
> (*job)->num_ibs = num_ibs;
>
> amdgpu_sync_create(&(*job)->sync);
>
> return 0;
> }
>
> int amdgpu_job_alloc_with_ib(struct amdgpu_device *adev, unsigned size,
> struct amdgpu_job **job)
> {
> int r;
>
> - r = amdgpu_job_alloc(adev, 1, job, NULL);
> + r = amdgpu_job_alloc(adev, 1, AMD_SCHED_PRIORITY_NORMAL, job, NULL);
> if (r)
> return r;
>
> r = amdgpu_ib_get(adev, NULL, size, &(*job)->ibs[0]);
> if (r)
> kfree(*job);
>
> return r;
> }
>
> @@ -88,20 +89,24 @@ void amdgpu_job_free_resources(struct amdgpu_job *job)
> f = job->base.s_fence ? &job->base.s_fence->finished : job->fence;
>
> for (i = 0; i < job->num_ibs; ++i)
> amdgpu_ib_free(job->adev, &job->ibs[i], f);
> }
>
> static void amdgpu_job_free_cb(struct amd_sched_job *s_job)
> {
> struct amdgpu_job *job = container_of(s_job, struct amdgpu_job, base);
>
> + /* no need to restore if we are already at the lowest priority */
> + if (job->priority < AMD_SCHED_PRIORITY_NORMAL)
> + amdgpu_ring_restore_priority_cb(job);
Please put that check into amdgpu_ring_restore_priority_cb(). And by the
way remove the _cb postfix from amdgpu_ring_restore_priority_cb().
> +
> dma_fence_put(job->fence);
> amdgpu_sync_free(&job->sync);
> kfree(job);
> }
>
> void amdgpu_job_free(struct amdgpu_job *job)
> {
> amdgpu_job_free_resources(job);
>
> dma_fence_put(job->fence);
> @@ -162,28 +167,33 @@ static struct dma_fence *amdgpu_job_run(struct amd_sched_job *sched_job)
> int r;
>
> if (!sched_job) {
> DRM_ERROR("job is null\n");
> return NULL;
> }
> job = to_amdgpu_job(sched_job);
>
> BUG_ON(amdgpu_sync_peek_fence(&job->sync, NULL));
>
> + r = amdgpu_ring_elevate_priority(job->ring, job->priority, job);
> + if (r)
> + DRM_ERROR("Failed to set job priority (%d)\n", r);
> +
> trace_amdgpu_sched_run_job(job);
> r = amdgpu_ib_schedule(job->ring, job->num_ibs, job->ibs, job, &fence);
> if (r)
> DRM_ERROR("Error scheduling IBs (%d)\n", r);
>
> /* if gpu reset, hw fence will be replaced here */
> dma_fence_put(job->fence);
> job->fence = dma_fence_get(fence);
> +
> amdgpu_job_free_resources(job);
> return fence;
> }
>
> const struct amd_sched_backend_ops amdgpu_sched_ops = {
> .dependency = amdgpu_job_dependency,
> .run_job = amdgpu_job_run,
> .timedout_job = amdgpu_job_timedout,
> .free_job = amdgpu_job_free_cb
> };
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c
> index 31c6274..fa6e72f2 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c
> @@ -193,20 +193,91 @@ void amdgpu_ring_commit(struct amdgpu_ring *ring)
> */
> void amdgpu_ring_undo(struct amdgpu_ring *ring)
> {
> ring->wptr = ring->wptr_old;
>
> if (ring->funcs->end_use)
> ring->funcs->end_use(ring);
> }
>
> /**
> + * amdgpu_ring_restore_priority_cb - restore a ring's priority
> + *
> + * @job: job which completed execution
> + *
> + * If necessary, restore a ring's priority after a job finishes executing
> + */
> +void amdgpu_ring_restore_priority_cb(struct amdgpu_job *job)
> +{
> + int i;
> + struct amdgpu_ring *ring = job->ring;
> +
> + spin_lock(&ring->priority_lock);
> +
> + /* remove ourselves from the list if necessary */
> + if (job == ring->last_job[job->priority])
> + ring->last_job[job->priority] = NULL;
> +
> + /* something higher prio is executing, no need to decay */
> + if (ring->priority < job->priority)
> + goto out_unlock;
> +
> + /* decay priority to the next level with a job available */
> + for (i = job->priority; i < AMD_SCHED_MAX_PRIORITY; i++) {
> + if (i == AMD_SCHED_PRIORITY_NORMAL || ring->last_job[i]) {
> + ring->priority = i;
> + if (ring->funcs->set_priority)
> + ring->funcs->set_priority(ring, i);
> +
> + break;
> + }
> + }
> +
> +out_unlock:
> + spin_unlock(&ring->priority_lock);
> +}
If you use an atomic to count the number of jobs for each priority
running you would only need the spinlock for updating the priority.
Additional to that I would abort early if ring->funcs->set_priority
isn't available (e.g. before doing anything else).
Regards,
Christian.
> +
> +/**
> + * amdgpu_ring_elevate_priority - change the ring's priority
> + *
> + * @ring: amdgpu_ring structure holding the information
> + * @priority: target priority
> + * @job: priority should remain elevated for the duration of this job
> + *
> + * Use HW specific mechanism's to elevate the ring's priority while @job
> + * is executing. Once @job finishes executing, the ring will reset back
> + * to normal priority.
> + * Returns 0 on success, error otherwise
> + */
> +int amdgpu_ring_elevate_priority(struct amdgpu_ring *ring, int priority,
> + struct amdgpu_job *job)
> +{
> + if (priority < 0 || priority >= AMD_SCHED_MAX_PRIORITY)
> + return -EINVAL;
> +
> + spin_lock(&ring->priority_lock);
> + ring->last_job[priority] = job;
> +
> + /* lower number means higher priority */
> + if (priority >= ring->priority)
> + goto out_unlock;
> +
> + ring->priority = priority;
> + if (ring->funcs->set_priority)
> + ring->funcs->set_priority(ring, priority);
> +
> +out_unlock:
> + spin_unlock(&ring->priority_lock);
> + return 0;
> +}
> +
> +/**
> * amdgpu_ring_init - init driver ring struct.
> *
> * @adev: amdgpu_device pointer
> * @ring: amdgpu_ring structure holding ring information
> * @max_ndw: maximum number of dw for ring alloc
> * @nop: nop packet for this ring
> *
> * Initialize the driver information for the selected ring (all asics).
> * Returns 0 on success, error on failure.
> */
> @@ -274,20 +345,22 @@ int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
> &ring->gpu_addr,
> (void **)&ring->ring);
> if (r) {
> dev_err(adev->dev, "(%d) ring create failed\n", r);
> return r;
> }
> memset((void *)ring->ring, 0, ring->ring_size);
> }
> ring->ptr_mask = (ring->ring_size / 4) - 1;
> ring->max_dw = max_dw;
> + ring->priority = AMD_SCHED_PRIORITY_NORMAL;
> + spin_lock_init(&ring->priority_lock);
> INIT_LIST_HEAD(&ring->lru_list);
> amdgpu_ring_lru_touch(adev, ring);
>
> if (amdgpu_debugfs_ring_init(adev, ring)) {
> DRM_ERROR("Failed to register debugfs file for rings !\n");
> }
> return 0;
> }
>
> /**
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
> index b51bdcd..ee6be0c 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
> @@ -17,20 +17,21 @@
> * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
> * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> * OTHER DEALINGS IN THE SOFTWARE.
> *
> * Authors: Christian König
> */
> #ifndef __AMDGPU_RING_H__
> #define __AMDGPU_RING_H__
>
> +#include <drm/amdgpu_drm.h>
> #include "gpu_scheduler.h"
>
> /* max number of rings */
> #define AMDGPU_MAX_RINGS 16
> #define AMDGPU_MAX_GFX_RINGS 1
> #define AMDGPU_MAX_COMPUTE_RINGS 8
> #define AMDGPU_MAX_VCE_RINGS 3
>
> /* some special values for the owner field */
> #define AMDGPU_FENCE_OWNER_UNDEFINED ((void*)0ul)
> @@ -130,20 +131,22 @@ struct amdgpu_ring_funcs {
> void (*pad_ib)(struct amdgpu_ring *ring, struct amdgpu_ib *ib);
> unsigned (*init_cond_exec)(struct amdgpu_ring *ring);
> void (*patch_cond_exec)(struct amdgpu_ring *ring, unsigned offset);
> /* note usage for clock and power gating */
> void (*begin_use)(struct amdgpu_ring *ring);
> void (*end_use)(struct amdgpu_ring *ring);
> void (*emit_switch_buffer) (struct amdgpu_ring *ring);
> void (*emit_cntxcntl) (struct amdgpu_ring *ring, uint32_t flags);
> void (*emit_rreg)(struct amdgpu_ring *ring, uint32_t reg);
> void (*emit_wreg)(struct amdgpu_ring *ring, uint32_t reg, uint32_t val);
> + /* priority functions */
> + void (*set_priority) (struct amdgpu_ring *ring, int priority);
> };
>
> struct amdgpu_ring {
> struct amdgpu_device *adev;
> const struct amdgpu_ring_funcs *funcs;
> struct amdgpu_fence_driver fence_drv;
> struct amd_gpu_scheduler sched;
> struct list_head lru_list;
>
> struct amdgpu_bo *ring_obj;
> @@ -164,31 +167,40 @@ struct amdgpu_ring {
> struct amdgpu_bo *mqd_obj;
> u32 doorbell_index;
> bool use_doorbell;
> unsigned wptr_offs;
> unsigned fence_offs;
> uint64_t current_ctx;
> char name[16];
> unsigned cond_exe_offs;
> u64 cond_exe_gpu_addr;
> volatile u32 *cond_exe_cpu_addr;
> +
> + spinlock_t priority_lock;
> + /* protected by priority_lock */
> + struct amdgpu_job *last_job[AMD_SCHED_MAX_PRIORITY];
> + int priority;
> +
> #if defined(CONFIG_DEBUG_FS)
> struct dentry *ent;
> #endif
> };
>
> int amdgpu_ring_is_valid_index(struct amdgpu_device *adev,
> int hw_ip, int ring);
> int amdgpu_ring_alloc(struct amdgpu_ring *ring, unsigned ndw);
> void amdgpu_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count);
> void amdgpu_ring_generic_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib);
> void amdgpu_ring_commit(struct amdgpu_ring *ring);
> void amdgpu_ring_undo(struct amdgpu_ring *ring);
> +int amdgpu_ring_elevate_priority(struct amdgpu_ring *ring, int priority,
> + struct amdgpu_job *job);
> +void amdgpu_ring_restore_priority_cb(struct amdgpu_job *job);
> int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
> unsigned ring_size, struct amdgpu_irq_src *irq_src,
> unsigned irq_type);
> void amdgpu_ring_fini(struct amdgpu_ring *ring);
> int amdgpu_ring_lru_get(struct amdgpu_device *adev, int hw_ip,
> struct amdgpu_ring **ring);
> void amdgpu_ring_lru_touch(struct amdgpu_device *adev, struct amdgpu_ring *ring);
>
> #endif
More information about the amd-gfx
mailing list