On Wed, Apr 06, 2022 at 09:51:32AM +0200, Christian König wrote:
> This should be possible now since we don't have the distinction
> between exclusive and shared fences any more.
>
> The only possible pitfall is that a dma_fence would be reused during the
> RCU grace period, but even that could be handled with a single extra check.
At worst this means we wait a bit longer than a perfect snapshot. For
anything where this would have resulted in dependency loops you need to
take the ww_mutex anyway.
>
> Signed-off-by: Christian König <christian.koe...@amd.com>
> ---
> drivers/dma-buf/dma-resv.c | 33 ++++++++++++---------------------
> drivers/dma-buf/st-dma-resv.c | 2 +-
> include/linux/dma-resv.h | 12 ------------
> 3 files changed, 13 insertions(+), 34 deletions(-)
>
> diff --git a/drivers/dma-buf/dma-resv.c b/drivers/dma-buf/dma-resv.c
> index 5b64aa554c36..0cce6e4ec946 100644
> --- a/drivers/dma-buf/dma-resv.c
> +++ b/drivers/dma-buf/dma-resv.c
> @@ -133,7 +133,6 @@ static void dma_resv_list_free(struct dma_resv_list *list)
> void dma_resv_init(struct dma_resv *obj)
> {
> ww_mutex_init(&obj->lock, &reservation_ww_class);
> - seqcount_ww_mutex_init(&obj->seq, &obj->lock);
This removes the last user, and I don't think we'll add one. Please also
add a patch to remove the seqcount_ww_mutex stuff and cc locking/rt folks
on this.
Reviewed-by: Daniel Vetter <daniel.vet...@ffwll.ch>
>
> RCU_INIT_POINTER(obj->fences, NULL);
> }
> @@ -292,28 +291,24 @@ void dma_resv_add_fence(struct dma_resv *obj, struct
> dma_fence *fence,
> fobj = dma_resv_fences_list(obj);
> count = fobj->num_fences;
>
> - write_seqcount_begin(&obj->seq);
> -
> for (i = 0; i < count; ++i) {
> enum dma_resv_usage old_usage;
>
> dma_resv_list_entry(fobj, i, obj, &old, &old_usage);
> if ((old->context == fence->context && old_usage >= usage) ||
> - dma_fence_is_signaled(old))
> - goto replace;
> + dma_fence_is_signaled(old)) {
> + dma_resv_list_set(fobj, i, fence, usage);
> + dma_fence_put(old);
> + return;
> + }
> }
>
> BUG_ON(fobj->num_fences >= fobj->max_fences);
> - old = NULL;
> count++;
>
> -replace:
> dma_resv_list_set(fobj, i, fence, usage);
> /* pointer update must be visible before we extend the num_fences */
> smp_store_mb(fobj->num_fences, count);
> -
> - write_seqcount_end(&obj->seq);
> - dma_fence_put(old);
> }
> EXPORT_SYMBOL(dma_resv_add_fence);
>
> @@ -341,7 +336,6 @@ void dma_resv_replace_fences(struct dma_resv *obj,
> uint64_t context,
> dma_resv_assert_held(obj);
>
> list = dma_resv_fences_list(obj);
> - write_seqcount_begin(&obj->seq);
> for (i = 0; list && i < list->num_fences; ++i) {
> struct dma_fence *old;
>
> @@ -352,14 +346,12 @@ void dma_resv_replace_fences(struct dma_resv *obj,
> uint64_t context,
> dma_resv_list_set(list, i, replacement, usage);
> dma_fence_put(old);
> }
> - write_seqcount_end(&obj->seq);
> }
> EXPORT_SYMBOL(dma_resv_replace_fences);
>
> /* Restart the unlocked iteration by initializing the cursor object. */
> static void dma_resv_iter_restart_unlocked(struct dma_resv_iter *cursor)
> {
> - cursor->seq = read_seqcount_begin(&cursor->obj->seq);
> cursor->index = 0;
> cursor->num_fences = 0;
> cursor->fences = dma_resv_fences_list(cursor->obj);
> @@ -388,8 +380,10 @@ static void dma_resv_iter_walk_unlocked(struct
> dma_resv_iter *cursor)
> cursor->obj, &cursor->fence,
> &cursor->fence_usage);
> cursor->fence = dma_fence_get_rcu(cursor->fence);
> - if (!cursor->fence)
> - break;
> + if (!cursor->fence) {
> + dma_resv_iter_restart_unlocked(cursor);
> + continue;
> + }
>
> if (!dma_fence_is_signaled(cursor->fence) &&
> cursor->usage >= cursor->fence_usage)
> @@ -415,7 +409,7 @@ struct dma_fence *dma_resv_iter_first_unlocked(struct
> dma_resv_iter *cursor)
> do {
> dma_resv_iter_restart_unlocked(cursor);
> dma_resv_iter_walk_unlocked(cursor);
> - } while (read_seqcount_retry(&cursor->obj->seq, cursor->seq));
> + } while (dma_resv_fences_list(cursor->obj) != cursor->fences);
> rcu_read_unlock();
>
> return cursor->fence;
> @@ -438,13 +432,13 @@ struct dma_fence *dma_resv_iter_next_unlocked(struct
> dma_resv_iter *cursor)
>
> rcu_read_lock();
> cursor->is_restarted = false;
> - restart = read_seqcount_retry(&cursor->obj->seq, cursor->seq);
> + restart = dma_resv_fences_list(cursor->obj) != cursor->fences;
> do {
> if (restart)
> dma_resv_iter_restart_unlocked(cursor);
> dma_resv_iter_walk_unlocked(cursor);
> restart = true;
> - } while (read_seqcount_retry(&cursor->obj->seq, cursor->seq));
> + } while (dma_resv_fences_list(cursor->obj) != cursor->fences);
> rcu_read_unlock();
>
> return cursor->fence;
> @@ -540,10 +534,7 @@ int dma_resv_copy_fences(struct dma_resv *dst, struct
> dma_resv *src)
> }
> dma_resv_iter_end(&cursor);
>
> - write_seqcount_begin(&dst->seq);
> list = rcu_replace_pointer(dst->fences, list, dma_resv_held(dst));
> - write_seqcount_end(&dst->seq);
> -
> dma_resv_list_free(list);
> return 0;
> }
> diff --git a/drivers/dma-buf/st-dma-resv.c b/drivers/dma-buf/st-dma-resv.c
> index 8ace9e84c845..813779e3c9be 100644
> --- a/drivers/dma-buf/st-dma-resv.c
> +++ b/drivers/dma-buf/st-dma-resv.c
> @@ -217,7 +217,7 @@ static int test_for_each_unlocked(void *arg)
> if (r == -ENOENT) {
> r = -EINVAL;
> /* That should trigger an restart */
> - cursor.seq--;
> + cursor.fences = (void*)~0;
> } else if (r == -EINVAL) {
> r = 0;
> }
> diff --git a/include/linux/dma-resv.h b/include/linux/dma-resv.h
> index 1db759eacc98..c8ccbc94d5d2 100644
> --- a/include/linux/dma-resv.h
> +++ b/include/linux/dma-resv.h
> @@ -155,15 +155,6 @@ struct dma_resv {
> */
> struct ww_mutex lock;
>
> - /**
> - * @seq:
> - *
> - * Sequence count for managing RCU read-side synchronization, allows
> - * read-only access to @fences while ensuring we take a consistent
> - * snapshot.
> - */
> - seqcount_ww_mutex_t seq;
> -
> /**
> * @fences:
> *
> @@ -202,9 +193,6 @@ struct dma_resv_iter {
> /** @fence_usage: the usage of the current fence */
> enum dma_resv_usage fence_usage;
>
> - /** @seq: sequence number to check for modifications */
> - unsigned int seq;
> -
> /** @index: index into the shared fences */
> unsigned int index;
>
> --
> 2.25.1
>
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
