On 09/07/2026 13:19, Boris Brezillon wrote:
On Thu, 9 Jul 2026 11:56:48 +0100
Tvrtko Ursulin <[email protected]> wrote:
On 09/07/2026 07:45, Boris Brezillon wrote:
On Wed, 8 Jul 2026 17:47:36 +0100
Tvrtko Ursulin <[email protected]> wrote:
On 06/07/2026 15:18, Boris Brezillon wrote:
On Mon, 6 Jul 2026 13:03:33 +0100
Tvrtko Ursulin <[email protected]> wrote:
On 02/07/2026 16:31, Boris Brezillon wrote:
On Thu, 2 Jul 2026 15:37:39 +0100
Tvrtko Ursulin <[email protected]> wrote:
Currently an unordered workqueue is used for the DRM scheduler which means
its concurrency is externally managed, and given there is one scheduler
instance per userspace queue, that means workqueue management logic is
within its rights to spawn many kernel threads to submit their respective
jobs.
Problem there is that all run job callbacks are serialized on the device
global mutex,
I think we should address that instead, and either shorten the scope of
the locked section, or make it so we don't make it a contention point
for concurrent job submission from different contexts (with a rwsem
instead of a lock, for instance).
making the potential thread storm just causing lock
contention.
If we add a separate ordered workqueue for the DRM scheduler integration
we can avoid this problem, since the ordered property directly expresses
the nature of the submission backend implementation.
Yep, except that's not how it was meant to work. The goal was to allow
contexts to submit their jobs concurrently to the FW. The only reason we
take the lock is to:
1. make sure the context is still allowed to take jobs
2. kick the group scheduler if the context is not resident
For #1, I believe we can come up with either a lockless solution, or a
solution where the lock protecting the state belongs to the group
instead of being externally protected by the device-wide scheduler lock.
For #2, the rwsem approach, and narrowing down the locked section to
just this part of the code should do the trick.
Out of curiosity how much CPU side parallelism you think is required to
keep these GPUs fed? Both today (with the greater lock contention) and
in the future (with the reduced contention) I guess would be interesting
data points.
The maximum is known: it's the amount of FW CSG slot we have available.
I think the theoretical limit is 16, but IIRC, we never had more than 8
exposed by the FW.
Yeah but is it _really_ required to have 8 CPU threads feed these slots?
8 is indeed the number of SW slots, but there are multiple HW queues
under the hood (and multiple cores to dispatch jobs to), making it so
multiple GPU context can effectively be scheduled in parallel. This
number is lower than the number of FW slots though (I need to check if
it's exposed through some RO regs).
Got it, thanks!
So it is desirable to keep the FW slots filled with up to N jobs of each
queue type, where N is the HW parallelism of that queue type.
Jobs are one level below that. Each FW slot is attached a group
(that's Arm's naming), which is basically backing a VkQueue. The group
contains 1 to 8 queues (1 for fragment processing, 1 for geometry
processing, 1 for compute, 1 for transfers, ...) which can run
concurrently, and each of these queues has a ring buffer to push jobs
to. The ring buffer size is customizable and determines the maximum
number of jobs you can have queued before you have to wait for a ringbuf
slot to become available again. I know I use the term "slot" a lot and
it gets confusing, but FW slot (a way to make a group/VkQueue resident)
must not be confused with a ringbuf slot (a slot where you store you job
targeting a specific queue under this group). You can fill up your
ringbuf slots while the group is not resident (when that happens, we
kick the second-stage scheduler to make sure the group is going to
become resident at some point). And equally, you can have a group that's
resident, but has all its queue ringbufs empty (in that case, it should
sooner or later be evicted and replaced by another group that wants to
execute things).
Thank you for this explanation - I am marking it with my "reference
material" tag.
So, back to your initial statement, it is desirable to keep the queues
under a resident group (one that's attached to a FW slot) filled with
jobs to maximize utilization. And ideally, we want the group with the
highest priority to be refilled faster than the other resident groups.
At least assuming that the MCU is not significantly slower than the main
CPU in dequeing the FW slots into HW queues.
So, FW is scheduling queues under the resident groups, but the ring
buffers are consumed by a HW component (CSHWIF in Arm's naming), which I
expect to run faster than the MCU itself, even though, the MCU being
dedicated to just that scheduling task with very little to no OS
overhead, I'd expect the turnaround time to be quite fast compared to
the APU, even if the MCU is clocked at a way lower freq.
Makes sense.
What would that number be in practice? I am going back to what Tejun
mentioned, that if we can provide some bound number of how many RT
workers we may need, then he may be able to provide the RT facility.
As I say below, I think we can start with just a single-worker
wq per priority, and take it from there if we see the resident queues
are not filled fast enough compared to the non-resident ones. What's
really important for us is for jobs targeting high-prio queues to hit
their ringbuf before those targeting lower prio queues.
GPU will still take one at a time and preemption is not that fast, no?
Preemption on the FW side is pretty simple: each slot gets a unique
prio, and lower prio slots only get HW queues and GPU resources if
higher prio ones are idle and accept to give up their resources for a
bit. We then have a 10ms tick in panthor to rotate the FW slot
priorities. So yes, preemption is not very granular, but that's not
really the problem I'm worried about. What I'm worried about is having
just one thread for everything, with the first-queued/first-served
model that the kthread_worker infrastructure provides. If we're talking
about one thread per-priority level, that's already better, and then I
agree that the contention on GPU contexts with the same priority is less
of an issue, especially since the run_job work has to run before being
rescheduled, which gives you this natural FIFO behavior, thus leaving
other contexts a chance to queue their run_job in the meantime.
But this WQ_UNBOUND -> WQ_SINGLE_THREAD transition, where the wq is
shared among the entire device is not that. It's actually serializing
work submission for all GPU contexts regardless of their priority.
TLDR; I'd be happy if we start with just one kthread per-prio + the
narrowing of the locked section in the run_job() implementation, so
that context submission actually happens concurrently, and low prio
context don't starve high-prio/RT ones in the submission path.
I don't think this is actually priority starvation but plain FIFO
starvation.
The FIFO is per group-queue, not global to the device though (we have
one sched_entity and one scheduler per group-queue, not one for the
entire system). The starvation, if any, would come from the
serialization in the run_job() implementation, as we try to acquire the
device-wide panthor_scheduler::lock. And that's not by design, but
rather a flaw in the original implementation.
It can happen even today with WQ_UNBOUND, which does nothing
about breaking the FIFO order based on priorities, just that some
parallelism alleviates it.
Well, it's indeed not providing a guarantee that things will be
pushed in the proper prio order, but the fact threads can be
spawned makes it less likely for low-prio queues to block high-prio
ones, at least.
But in principle I am fine with going with N workers. It's just a matter
of what is N derived from and how big it is. It could be even be passed
to alloc_workqueue in the today's code base but I accept there is not
much value to that since I don't think there are SoC's with a Mali GPU
and server level number of CPU cores.
I think we can start with one ordered-wq per prio level. So basically
what you intended to do in this patch, but instead of having a single
ordered-wq, we have four of them (one per prio). It doesn't address the
fact run_job() on non-resident groups might hit their ringbuf before
resident ones, but that's probably good enough as a first step, and as
you pointed out, the current locking forces this serialization with no
more guarantee regarding who's going to be served first anyway, so it
can't be worse than it already is.
Lets for a moment assume the kthread_worker idea will not fly due to xe,
or any other reason really. Also since the option for RT workqueues is
unexpectedly on the table. In that case, and assuming RT workqueues will
happen, could a feasible plan for panthor be to create three workqueues:
1. One unbound with max_active = 2 for low and medium group priority.
2. Another unbound + WQ_HIGHPRI, also with max_active = 2 for high.
3. And one unbound + WQ_REALTIME, again max_active = 2 for realtime.
For high and realtime max_active either 2 or 1 on dual core, if there
are such SoCs. Two threads ensure same priority clients are able to keep
the GPU fed.
One issue is that you mentioned you would like dynamic priority changes
and with this it may be tricky. But ignoring that for the moment, and
the discussion on how to handle other panthor workers which take part in
the execution flow post submit, this should pretty much address the
submit latency from userspace to ->run_job(). What do you think, is it
worth entertaining this alternative?
Regards,
Tvrtko
Note that this is probably not enough to make a difference, we also
need panthor-specific work items to go to the per-prio wqs based on
where the event comes from (we have per-FW-slot interrupts), and the
priority of the group attached to the FW slot. Or we just consider any
event from the FW as high-prio regardless of the slot its pointing to,
dunno.
