Currently, vk_sync_timeline_wait() may report idle while we still have
time points outstanding in vk_queue. This race between vk_sync_timeline
and vk_queue is problematic as vk_queue_submit_cleanup() can end-up
trying to unref a timeline point that was already destroyed in
vk_sync_timeline_finish(), leading to an use-after-free.
Address this race-condition by introducing a reference counter to the
timeline. Each timeline point takes a reference to the timeline when
it's checked out and drops the reference when it's freed. Now, the
timeline is only destroyed when all the references had been dropped.
To guarantee that all references will be dropped and the timeline will
be destroyed, vk_sync_timeline_finish() waits for all pending points
to be checked out.
This commit fixes several Mesa CI flakes in v3dv related to timeline
semaphores.
Cc: mesa-stable
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/12648
Signed-off-by: Maíra Canal <mcanal@igalia.com>
Reviewed-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35921>
Previously, we had a sort of double reference count with refcount being
used to track how many waits were holding a reference and pending being
used to track whether or not it was in the pending list. This meant
that the unref operation actually had to look at both refcount and
pending in order to determine if it could free the time point. This is
way too confusing. Instead, we should just always take a reference
while we're pending.
This also simplifies the over-all interface because there's no longer a
difference between free and release. `struct vk_sync_timeline_point` is
now just a reference-counted API.
Reviewed-by: Maíra Canal <mcanal@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35921>
The lifetime of `struct vk_sync_timeline` is tied to the lifetime of
`vk_fence` or `vk_semaphore` objects. To better manage this lifecycle,
create a wrapper struct `vk_sync_timeline_state` that contains the
timeline state and is dynamically allocated when the timeline is
initialized. This separation allows the timeline state to persist
independently of the sync object lifecycle.
Signed-off-by: Maíra Canal <mcanal@igalia.com>
Reviewed-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35921>
The code we had for handling this with c11 cnd_t was gross. Let's use a
primitive that actually works.
Reviewed-by: Derek Foreman <derek.foreman@collabora.com>
Reviewed-by: Tatsuyuki Ishi <ishitatsuyuki@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/29924>
This is mostly copy+paste (and a bit of re-typing) from ANV.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Acked-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13427>
