Re: [PATCH v3 2/2] iommu/arm-smmu-v3: Adopt arm_smmu_event in handlers

[Pranjal Shrivastava <praan@google.com>]

On Fri, Oct 11, 2024 at 01:59:36PM +0100, Robin Murphy wrote:
> On 2024-10-11 12:06 pm, Pranjal Shrivastava wrote:
> > On Fri, Oct 11, 2024 at 10:45:52AM +0000, Pranjal Shrivastava wrote:
> > > On Fri, Oct 11, 2024 at 11:21:48AM +0100, Robin Murphy wrote:
> > > > On 2024-10-11 8:55 am, Pranjal Shrivastava wrote:
> > > > > On Thu, Oct 03, 2024 at 03:53:15PM -0700, Nicolin Chen wrote:
> > > > > > On Thu, Oct 03, 2024 at 09:34:35PM +0000, Pranjal Shrivastava wrote:
> > > > > > > > The master and master_name are not that necessary to be in the
> > > > > > > > arm_smmu_event. I would keep both of them as local variables in
> > > > > > > > arm_smmu_handle_evt and pass master_name in to dump().
> > > > > > > 
> > > > > > > Hmm, right or maybe ONLY have the master_name in arm_smmu_event?
> > > > > > 
> > > > > > Just move both out of arm_smmu_event. Keep the master_name as a
> > > > > > local variable like the "master". You'd just need to pass it in
> > > > > > to the dump(). They both are in that same handle_evt().
> > > > > > 
> > > > > 
> > > > > Ack.
> > > > > 
> > > > > > > > > Also, shall we rename it to `arm_smmu_read_evt_info` ?
> > > > > > > > 
> > > > > > > > I'd probably use arm_smmu_event_get_from_raw()? Trying to high-
> > > > > > > > light struct arm_smmu_event v.s. u64 evt[EVTQ_ENT_DWORDS]. Yet,
> > > > > > > > no strong feeling about that.
> > > > > > > > 
> > > > > > > 
> > > > > > > Ack. How about `read_arm_smmu_event` :)
> > > > > > > Unless, we wanna follow the arm_smmu_<func> convention?
> > > > > > 
> > > > > > I think it'd be nicer if we do.
> > > > > > 
> > > > > 
> > > > > Ack. `arm_smmu_get_evt_from_raw` it is!
> > > > 
> > > > FWIW that sounds needlessly overcomplicated to me - the "raw" event array
> > > > should be a member of arm_smmu_event itself, since it seems silly to have
> > > > them separate with one pointing to the other when they have the exact same
> > > > scope and lifetime anyway. There still shouldn't need to be more than a
> > > > single logical step to process an evtq record into a finished
> > > > arm_smmu_event, just that that processing is now going to go a bit further
> > > > than simply le64_to_cpu().
> > > 
> > > Hmm, are you suggesting something like queue_remove_raw(q, event->raw)?
> > > OR do you mean we should have our own queue_parse_evt(q, &event) that
> > > parses out the raw event into an `arm_smmu_event` record eliminating the
> > > need for the `arm_smmu_read_evt_info` altogether?
> > > 
> > > Because we'd still need to store the raw event anyway since we're still
> > > logging the raw event.
> > > 
> > > I think we can go with the former, i.e. queue_remove_raw(q, event->raw)
> > > and then use ``arm_smmu_read_evt_info` to populate other `event` fields
> > 
> > Based on your comment, I was thinking like the attached sample code.
> > LMK if you were referring to something like that?
> 
> Ah, apologies, it had slipped my mind that queue_read() is already buried
> two layers deep in an abstraction shared with the priq... so yeah, rather
> than start trying to turn that inside-out, I guess the neatest approach is
> to effectively flip this series instead - i.e. start with the minor shift
> to:
> 
> 	struct arm_smmu_event evt;
> 
> 	while (!queue_remove_raw(q, evt.raw)) {
> 		ret = arm_smmu_handle_evt(smmu, &evt);
> 		...
> 		dev_info(smmu->dev, "event 0x%02x received:\n", evt.id);
> 		//etc.
> 
> and leave arm_smmu_handle_evt() decoding the event as both it and
> arm_smmu_handle_ppr() already do, just now using the new structure instead
> of locals. Then we can come in to factor out and enhance the unhandled event
> report, and anything else specific to that reporting beyond the actual SMMU
> event record, like device names and ratelimiting, can remain private to
> arm_smmu_dump_event().

Ack. That makes sense. I'll re-organize the series accordingly.

> 
> AFAICS the locking concern seems avoidable as well - we don't actually need
> the whole arm_smmu_master in either case, just the device, so then end of
> arm_smmu_handle_evt() could easily be a little simpler:
> 
> 	...
> 	mutex_lock(&smmu->streams_mutex);
> 	master = arm_smmu_find_master(smmu, sid);
> 	if (master)
> 		event->dev = get_device(master->dev);
> 	mutex_unlock(&smmu->streams_mutex);
> 
> 	if (!event->dev)
> 		return -EINVAL;

Ahh, I see! My worry was a using `dev` after free but it looks like
`get_device(dev)` increments the reference count, preventing `dev` from
being freed prematurely.

Looking at the core code in drivers/base/core.c to confirm,
I see that the `device_release` invokes `dev->release` which frees the
`dev` struct and it seems that the `device_release` is called only
kobject's refcount reaches 0. Additionally, I see that `device_release`
is registered as the ->release callback in the device_ktype kobject.

So, looks like we should be good? LMK if I missed something above?

> 
> 	return iommu_report_device_fault(event->dev, &fault_evt);
> 
> then arm_smmu_dump_event() is safe to use a construct like:
> 
> 	event->dev ? dev_name(event->dev) : "(unassigned)"
> 
> and finally a cleanup in the evtq loop, which by that point probably
> condenses down to something like:
> 
> 	while (!queue_remove_raw(q, evt.raw)) {
> 		if (arm_smmu_handle_evt(smmu, &evt))
> 			arm_smmu_dump_event(&evt);
> 		put_device(evt.dev);

Ack. Although, I'm curious to know the concern while calling
`arm_smmu_dump_event` within `arm_smmu_handle_evt`, is it because we
might hold the lock for a longer duration due to the super-long prints?

> 		cond_resched();
> 	}
> 
> Thanks,
> Robin.

Thanks,
Pranjal