Commit 64d6f56a ("panfrost: Allocate syncobjs in panfrost_flush")
aimed at optimizing the fencing logic but it looks it also broke the
fence-based synchronization in subtle ways.

Indeed, now that the fence only waits on a single syncobj, we're not
guaranteed that all jobs queued in panfrost_flush_all_batches() will
be done when the fence is signaled, because jobs at the top level
(those stored in the batches hashmap) have not inter-dependencies.

Commit 9e397956 ("panfrost: signal syncobj if nothing is going to
be flushed") made this even more apparent by signaling the fence right
away if nothing was left to be drawn in the current context, thus
ignoring any of the batches left to flushed in the ->batches map.

If we want to keep relying the existing kernel APIs there's clearly no
ideal solution here. We can either go back to the original fencing
mechanism where each fence contained an array of syncobjs to be tested
or serialize jobs that have no explicit dependencies so we know the last
submitted job will also be the last one to return. The orginal approach
has proven to add quite a significant overhead (caused by the amount of
ioctls and the time spent in kernel space to gather dma fences attached
to those syncobjs and test them). So let's go for the simple solution
where we have a single syncobj bound to the context which we update to
point to the last job out_sync every time we submit a top-level job.

This approach implies reworking the way we create fences since we
need to capture the syncobj state at the time the fence is created.
Unfortunately, there's not SYNCOBJ_CLONE ioctl, which forces us to
export/create/import a fence so we have a new object that's not
subject to changes done to the context syncobj.

If we want to further optimize the logic, we should probably explore
some of those options:

1/ Adding array based SYNCOBJ ioctls (SYNCOBJ_{CREATE,DESTROY,CLONE}_ARRAY)
   so we can mitigate the cost of ioctls when we need to manipulate
   arrays of syncobjs
2/ Support synchronization jobs. That is, jobs that have a NULL job chain
   but an array of sync_in and a sync_out to allow creating
   synchronization points
3/ Add syncobj aggregators so we only have to wait on one syncobj from
   userspace. The syncobj aggregator would wait for all sub syncobjs to
   be signaled before signaling the top-level one.

Fixes: 64d6f56a

 ("panfrost: Allocate syncobjs in panfrost_flush")
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <!7831>