On 10/01, Iago Toral wrote:
> On Thu, 2021-09-30 at 17:19 +0100, Melissa Wen wrote:
> > Using the generic extension from the previous patch, a specific
> > multisync
> > extension enables more than one in/out binary syncobj per job
> > submission.
> > Arrays of syncobjs are set in struct drm_v3d_multisync, that also
> > cares
> > of determining the stage for sync (wait deps) according to the job
> > queue.
> >
> > v2:
> > - subclass the generic extension struct (Daniel)
> > - simplify adding dependency conditions to make understandable (Iago)
> >
> > v3:
> > - fix conditions to consider single or multiples in/out_syncs (Iago)
> > - remove irrelevant comment (Iago)
> >
> > Signed-off-by: Melissa Wen <mwen@xxxxxxxxxx>
> > ---
> > drivers/gpu/drm/v3d/v3d_drv.c | 6 +-
> > drivers/gpu/drm/v3d/v3d_drv.h | 24 +++--
> > drivers/gpu/drm/v3d/v3d_gem.c | 185 ++++++++++++++++++++++++++++++
> > ----
> > include/uapi/drm/v3d_drm.h | 49 ++++++++-
> > 4 files changed, 232 insertions(+), 32 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/v3d/v3d_drv.c
> > b/drivers/gpu/drm/v3d/v3d_drv.c
> > index 3d6b9bcce2f7..bd46396a1ae0 100644
> > --- a/drivers/gpu/drm/v3d/v3d_drv.c
> > +++ b/drivers/gpu/drm/v3d/v3d_drv.c
> > @@ -96,6 +96,9 @@ static int v3d_get_param_ioctl(struct drm_device
>
> (...)
>
> >
> > @@ -516,9 +536,11 @@ v3d_attach_fences_and_unlock_reservation(struct
> > drm_file *file_priv,
> > struct v3d_job *job,
> > struct ww_acquire_ctx
> > *acquire_ctx,
> > u32 out_sync,
> > + struct v3d_submit_ext *se,
> > struct dma_fence *done_fence)
> > {
> > struct drm_syncobj *sync_out;
> > + bool has_multisync = se && (se->flags &
>
> We always pass the 'se' pointer from a local variable allocated in the
> stack of the caller so it is never going to be NULL.
>
> I am happy to keep the NULL check if you want to protect against future
> changes just in case, but if we do that then...
>
> > DRM_V3D_EXT_ID_MULTI_SYNC);
> > int i;
> >
> > for (i = 0; i < job->bo_count; i++) {
>
> (...)
>
> > +static void
> > +v3d_put_multisync_post_deps(struct v3d_submit_ext *se)
> > +{
> > + unsigned int i;
> > +
> > + if (!se->out_sync_count)
>
> ...we should also check for NULL here for consistency.
yes, consistency makes sense here.
>
> Also, I think there is another problem in the code: we always call
> v3d_job_cleanup for failed jobs, but only call v3d_job_put for
> successful jobs. However, reading the docs for drm_sched_job_init:
>
> "Drivers must make sure drm_sched_job_cleanup() if this function
> returns successfully, even when @job is aborted before
> drm_sched_job_arm() is called."
>
> So my understanding is that we should call v3d_job_cleanup instead of
> v3d_job_put for successful jobs or we would be leaking sched resources
> on every successful job, no?
When job_init is successful, v3d_job_cleanup is called by scheduler when
job is completed. drm_sched_job_cleanup describes how it works after
drm_sched_job_arm:
" After that point of no return @job is committed to be executed by the
scheduler, and this function should be called from the
&drm_sched_backend_ops.free_job callback."
On v3d_sched.c, .free_job points to v3d_sched_job_free(), which in turn
calls v3d_job_cleanup() (and then drm_sched_job_cleanup). So, it looks
ok.
Also, we can say that the very last v3d_job_put() is in charge to
decrement refcount initialized (set 1) in v3d_job_init(); while the
v3d_job_cleanup() from v3d_sched_job_free() callback decrements refcount
that was incremented when v3d_job_push() pushed the job to the
scheduler.
So, refcount pairs seem consistent, I mean, get and put. And about
drm_sched_job_cleanup, it is explicitly called when job_init fails or
after that by the scheduler.
Thanks,
Melissa
>
> Iago
>
> > + return;
> > +
> > + for (i = 0; i < se->out_sync_count; i++)
> > + drm_syncobj_put(se->out_syncs[i].syncobj);
> > + kvfree(se->out_syncs);
> > +}
> > +
> > +static int
> > +v3d_get_multisync_post_deps(struct drm_file *file_priv,
> > + struct v3d_submit_ext *se,
> > + u32 count, u64 handles)
> > +{
> > + struct drm_v3d_sem __user *post_deps;
> > + int i, ret;
> > +
> > + if (!count)
> > + return 0;
> > +
> > + se->out_syncs = (struct v3d_submit_outsync *)
> > + kvmalloc_array(count,
> > + sizeof(struct
> > v3d_submit_outsync),
> > + GFP_KERNEL);
> > + if (!se->out_syncs)
> > + return -ENOMEM;
> > +
> > + post_deps = u64_to_user_ptr(handles);
> > +
> > + for (i = 0; i < count; i++) {
> > + struct drm_v3d_sem out;
> > +
> > + ret = copy_from_user(&out, post_deps++, sizeof(out));
> > + if (ret) {
> > + DRM_DEBUG("Failed to copy post dep handles\n");
> > + goto fail;
> > + }
> > +
> > + se->out_syncs[i].syncobj = drm_syncobj_find(file_priv,
> > + out.handle)
> > ;
> > + if (!se->out_syncs[i].syncobj) {
> > + ret = -EINVAL;
> > + goto fail;
> > + }
> > }
> > + se->out_sync_count = count;
> > +
> > + return 0;
> > +
> > +fail:
> > + for (i--; i >= 0; i--)
> > + drm_syncobj_put(se->out_syncs[i].syncobj);
> > + kvfree(se->out_syncs);
> > +
> > + return ret;
> > +}
> > +
> > +/* Get data for multiple binary semaphores synchronization. Parse
> > syncobj
> > + * to be signaled when job completes (out_sync).
> > + */
> > +static int
> > +v3d_get_multisync_submit_deps(struct drm_file *file_priv,
> > + struct drm_v3d_extension __user *ext,
> > + void *data)
> > +{
> > + struct drm_v3d_multi_sync multisync;
> > + struct v3d_submit_ext *se = data;
> > + int ret;
> > +
> > + ret = copy_from_user(&multisync, ext, sizeof(multisync));
> > + if (ret)
> > + return ret;
> > +
> > + if (multisync.pad)
> > + return -EINVAL;
> > +
> > + ret = v3d_get_multisync_post_deps(file_priv, data,
> > multisync.out_sync_count,
> > + multisync.out_syncs);
> > + if (ret)
> > + return ret;
> > +
> > + se->in_sync_count = multisync.in_sync_count;
> > + se->in_syncs = multisync.in_syncs;
> > + se->flags |= DRM_V3D_EXT_ID_MULTI_SYNC;
> > + se->wait_stage = multisync.wait_stage;
> > +
> > + return 0;
> > }
> >
> > /* Whenever userspace sets ioctl extensions, v3d_get_extensions
> > parses data
> > * according to the extension id (name).
> > */
> > static int
> > -v3d_get_extensions(struct drm_file *file_priv, u64 ext_handles)
> > +v3d_get_extensions(struct drm_file *file_priv,
> > + u64 ext_handles,
> > + void *data)
> > {
> > struct drm_v3d_extension __user *user_ext;
> > + int ret;
> >
> > user_ext = u64_to_user_ptr(ext_handles);
> > while (user_ext) {
> > @@ -555,7 +687,11 @@ v3d_get_extensions(struct drm_file *file_priv,
> > u64 ext_handles)
> > }
> >
> > switch (ext.id) {
> > - case 0:
> > + case DRM_V3D_EXT_ID_MULTI_SYNC:
> > + ret = v3d_get_multisync_submit_deps(file_priv,
> > user_ext, data);
> > + if (ret)
> > + return ret;
> > + break;
> > default:
> > DRM_DEBUG_DRIVER("Unknown extension id: %d\n",
> > ext.id);
> > return -EINVAL;
> > @@ -586,6 +722,7 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void
> > *data,
> > struct v3d_dev *v3d = to_v3d_dev(dev);
> > struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
> > struct drm_v3d_submit_cl *args = data;
> > + struct v3d_submit_ext se = {0};
> > struct v3d_bin_job *bin = NULL;
> > struct v3d_render_job *render = NULL;
> > struct v3d_job *clean_job = NULL;
> > @@ -606,7 +743,7 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void
> > *data,
> > }
> >
> > if (args->flags & DRM_V3D_SUBMIT_EXTENSION) {
> > - ret = v3d_get_extensions(file_priv, args->extensions);
> > + ret = v3d_get_extensions(file_priv, args->extensions,
> > &se);
> > if (ret) {
> > DRM_DEBUG("Failed to get extensions.\n");
> > return ret;
> > @@ -614,7 +751,7 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void
> > *data,
> > }
> >
> > ret = v3d_job_init(v3d, file_priv, (void *)&render,
> > sizeof(*render),
> > - v3d_render_job_free, args->in_sync_rcl,
> > V3D_RENDER);
> > + v3d_render_job_free, args->in_sync_rcl, &se,
> > V3D_RENDER);
> > if (ret)
> > goto fail;
> >
> > @@ -624,7 +761,7 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void
> > *data,
> >
> > if (args->bcl_start != args->bcl_end) {
> > ret = v3d_job_init(v3d, file_priv, (void *)&bin,
> > sizeof(*bin),
> > - v3d_job_free, args->in_sync_bcl,
> > V3D_BIN);
> > + v3d_job_free, args->in_sync_bcl,
> > &se, V3D_BIN);
> > if (ret)
> > goto fail;
> >
> > @@ -638,7 +775,7 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void
> > *data,
> >
> > if (args->flags & DRM_V3D_SUBMIT_CL_FLUSH_CACHE) {
> > ret = v3d_job_init(v3d, file_priv, (void *)&clean_job,
> > sizeof(*clean_job),
> > - v3d_job_free, 0, V3D_CACHE_CLEAN);
> > + v3d_job_free, 0, 0,
> > V3D_CACHE_CLEAN);
> > if (ret)
> > goto fail;
> >
> > @@ -698,6 +835,7 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void
> > *data,
> > last_job,
> > &acquire_ctx,
> > args->out_sync,
> > + &se,
> > last_job->done_fence);
> >
> > if (bin)
> > @@ -716,6 +854,7 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void
> > *data,
> > v3d_job_cleanup((void *)bin);
> > v3d_job_cleanup((void *)render);
> > v3d_job_cleanup(clean_job);
> > + v3d_put_multisync_post_deps(&se);
> >
> > return ret;
> > }
> > @@ -735,6 +874,7 @@ v3d_submit_tfu_ioctl(struct drm_device *dev, void
> > *data,
> > {
> > struct v3d_dev *v3d = to_v3d_dev(dev);
> > struct drm_v3d_submit_tfu *args = data;
> > + struct v3d_submit_ext se = {0};
> > struct v3d_tfu_job *job = NULL;
> > struct ww_acquire_ctx acquire_ctx;
> > int ret = 0;
> > @@ -747,7 +887,7 @@ v3d_submit_tfu_ioctl(struct drm_device *dev, void
> > *data,
> > }
> >
> > if (args->flags & DRM_V3D_SUBMIT_EXTENSION) {
> > - ret = v3d_get_extensions(file_priv, args->extensions);
> > + ret = v3d_get_extensions(file_priv, args->extensions,
> > &se);
> > if (ret) {
> > DRM_DEBUG("Failed to get extensions.\n");
> > return ret;
> > @@ -755,7 +895,7 @@ v3d_submit_tfu_ioctl(struct drm_device *dev, void
> > *data,
> > }
> >
> > ret = v3d_job_init(v3d, file_priv, (void *)&job, sizeof(*job),
> > - v3d_job_free, args->in_sync, V3D_TFU);
> > + v3d_job_free, args->in_sync, &se, V3D_TFU);
> > if (ret)
> > goto fail;
> >
> > @@ -803,6 +943,7 @@ v3d_submit_tfu_ioctl(struct drm_device *dev, void
> > *data,
> > v3d_attach_fences_and_unlock_reservation(file_priv,
> > &job->base,
> > &acquire_ctx,
> > args->out_sync,
> > + &se,
> > job->base.done_fence);
> >
> > v3d_job_put(&job->base);
> > @@ -811,6 +952,7 @@ v3d_submit_tfu_ioctl(struct drm_device *dev, void
> > *data,
> >
> > fail:
> > v3d_job_cleanup((void *)job);
> > + v3d_put_multisync_post_deps(&se);
> >
> > return ret;
> > }
> > @@ -831,6 +973,7 @@ v3d_submit_csd_ioctl(struct drm_device *dev, void
> > *data,
> > struct v3d_dev *v3d = to_v3d_dev(dev);
> > struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
> > struct drm_v3d_submit_csd *args = data;
> > + struct v3d_submit_ext se = {0};
> > struct v3d_csd_job *job = NULL;
> > struct v3d_job *clean_job = NULL;
> > struct ww_acquire_ctx acquire_ctx;
> > @@ -852,7 +995,7 @@ v3d_submit_csd_ioctl(struct drm_device *dev, void
> > *data,
> > }
> >
> > if (args->flags & DRM_V3D_SUBMIT_EXTENSION) {
> > - ret = v3d_get_extensions(file_priv, args->extensions);
> > + ret = v3d_get_extensions(file_priv, args->extensions,
> > &se);
> > if (ret) {
> > DRM_DEBUG("Failed to get extensions.\n");
> > return ret;
> > @@ -860,12 +1003,12 @@ v3d_submit_csd_ioctl(struct drm_device *dev,
> > void *data,
> > }
> >
> > ret = v3d_job_init(v3d, file_priv, (void *)&job, sizeof(*job),
> > - v3d_job_free, args->in_sync, V3D_CSD);
> > + v3d_job_free, args->in_sync, &se, V3D_CSD);
> > if (ret)
> > goto fail;
> >
> > ret = v3d_job_init(v3d, file_priv, (void *)&clean_job,
> > sizeof(*clean_job),
> > - v3d_job_free, 0, V3D_CACHE_CLEAN);
> > + v3d_job_free, 0, 0, V3D_CACHE_CLEAN);
> > if (ret)
> > goto fail;
> >
> > @@ -904,6 +1047,7 @@ v3d_submit_csd_ioctl(struct drm_device *dev,
> > void *data,
> > clean_job,
> > &acquire_ctx,
> > args->out_sync,
> > + &se,
> > clean_job-
> > >done_fence);
> >
> > v3d_job_put(&job->base);
> > @@ -918,6 +1062,7 @@ v3d_submit_csd_ioctl(struct drm_device *dev,
> > void *data,
> > fail:
> > v3d_job_cleanup((void *)job);
> > v3d_job_cleanup(clean_job);
> > + v3d_put_multisync_post_deps(&se);
> >
> > return ret;
> > }
> > diff --git a/include/uapi/drm/v3d_drm.h b/include/uapi/drm/v3d_drm.h
> > index 55b443ca6c0b..3dfc0af8756a 100644
> > --- a/include/uapi/drm/v3d_drm.h
> > +++ b/include/uapi/drm/v3d_drm.h
> > @@ -73,6 +73,53 @@ struct drm_v3d_extension {
> > __u32 flags; /* mbz */
> > };
> >
> > +/* struct drm_v3d_sem - wait/signal semaphore
> > + *
> > + * If binary semaphore, it only takes syncobj handle and ignores
> > flags and
> > + * point fields. Point is defined for timeline syncobj feature.
> > + */
> > +struct drm_v3d_sem {
> > + __u32 handle; /* syncobj */
> > + /* rsv below, for future uses */
> > + __u32 flags;
> > + __u64 point; /* for timeline sem support */
> > + __u64 mbz[2]; /* must be zero, rsv */
> > +};
> > +
> > +/* Enum for each of the V3D queues. */
> > +enum v3d_queue {
> > + V3D_BIN,
> > + V3D_RENDER,
> > + V3D_TFU,
> > + V3D_CSD,
> > + V3D_CACHE_CLEAN,
> > +};
> > +
> > +/**
> > + * struct drm_v3d_multi_sync - ioctl extension to add support
> > multiples
> > + * syncobjs for commands submission.
> > + *
> > + * When an extension of DRM_V3D_EXT_ID_MULTI_SYNC id is defined, it
> > points to
> > + * this extension to define wait and signal dependencies, instead of
> > single
> > + * in/out sync entries on submitting commands. The field flags is
> > used to
> > + * determine the stage to set wait dependencies.
> > + */
> > +struct drm_v3d_multi_sync {
> > + struct drm_v3d_extension base;
> > + /* Array of wait and signal semaphores */
> > + __u64 in_syncs;
> > + __u64 out_syncs;
> > +
> > + /* Number of entries */
> > + __u32 in_sync_count;
> > + __u32 out_sync_count;
> > +
> > + /* set the stage (v3d_queue) to sync */
> > + __u32 wait_stage;
> > +
> > + __u32 pad; /* mbz */
> > +};
> > +
> > /**
> > * struct drm_v3d_submit_cl - ioctl argument for submitting commands
> > to the 3D
> > * engine.
> > @@ -228,6 +275,7 @@ enum drm_v3d_param {
> > DRM_V3D_PARAM_SUPPORTS_CSD,
> > DRM_V3D_PARAM_SUPPORTS_CACHE_FLUSH,
> > DRM_V3D_PARAM_SUPPORTS_PERFMON,
> > + DRM_V3D_PARAM_SUPPORTS_MULTISYNC_EXT,
> > };
> >
> > struct drm_v3d_get_param {
> > @@ -271,7 +319,6 @@ struct drm_v3d_submit_tfu {
> >
> > /* Pointer to an array of ioctl extensions*/
> > __u64 extensions;
> > -
> > };
> >
> > /* Submits a compute shader for dispatch. This job will block on
> > any
>
Attachment:
signature.asc
Description: PGP signature
