From: Robin Murphy <robin.murphy@arm.com>
To: Jason Gunthorpe <jgg@nvidia.com>
Cc: Alex Williamson <alex.williamson@redhat.com>,
Cornelia Huck <cohuck@redhat.com>,
iommu@lists.linux.dev, Joerg Roedel <joro@8bytes.org>,
kvm@vger.kernel.org, Will Deacon <will@kernel.org>,
Qian Cai <cai@lca.pw>, Joerg Roedel <jroedel@suse.de>,
Marek Szyprowski <m.szyprowski@samsung.com>,
Matthew Rosato <mjrosato@linux.ibm.com>
Subject: Re: [PATCH 4/4] iommu: Fix ordering of iommu_release_device()
Date: Fri, 9 Sep 2022 10:05:58 +0100 [thread overview]
Message-ID: <38bac59a-808d-5e91-227a-a3a06633c091@arm.com> (raw)
In-Reply-To: <YxpiBEbGHECGGq5Q@nvidia.com>
On 2022-09-08 22:43, Jason Gunthorpe wrote:
> On Thu, Sep 08, 2022 at 10:27:06PM +0100, Robin Murphy wrote:
>
>> Oh, because s390 is using iommu_get_domain_for_dev() in its release_device
>> callback, which needs to dereference the group to work, and the current
>> domain may also be a non-default one which we can't prevent from
>> disappearing racily, that was why :(
>
> Hum, the issue there is the use of device->iommu_group - but that just
> means I didn't split properly. How about this incremental:
That did cross my mind, but it's a bit grim. In the light of the
morning, I'm not sure s390 actually *needs* the group anyway - AFAICS if
iommu_group_remove_device() has been processed first, that will have
synchronised against any concurrent attach/detach, so zdev->s390_domain
can be assumed to be up to date and used directly without the round trip
through iommu_get_domain_for_dev(). That then only leaves the issue that
that domain may still become invalid at any point after the group mutex
has been dropped.
> diff --git a/drivers/iommu/iommu.c b/drivers/iommu/iommu.c
> index c451bf715182ac..99ef799f3fe6b5 100644
> --- a/drivers/iommu/iommu.c
> +++ b/drivers/iommu/iommu.c
> @@ -351,6 +351,7 @@ void iommu_release_device(struct device *dev)
> * them until the have been detached. release_device() is expected to
> * detach all domains connected to the dev.
> */
> + dev->iommu_group = NULL;
> kobject_put(group->devices_kobj);
>
> module_put(ops->owner);
> @@ -980,7 +981,6 @@ static void __iommu_group_remove_device(struct device *dev)
>
> kfree(device->name);
> kfree(device);
> - dev->iommu_group = NULL;
> }
>
> /**
> @@ -995,6 +995,7 @@ void iommu_group_remove_device(struct device *dev)
> struct iommu_group *group = dev->iommu_group;
>
> __iommu_group_remove_device(dev);
> + dev->iommu_group = NULL;
> kobject_put(group->devices_kobj);
> }
> EXPORT_SYMBOL_GPL(iommu_group_remove_device);
>
> To me it makes sense that the driver should be able to continue to
> query the iommu_group during release anyhow..
I'm not so sure, release shouldn't be depending on a group since there
may never have been one anyway. Perhaps the answer is an extra
pre-release step to balance probe_finalize?
> And to your other question, the reason I split the function is because
> I couldn't really say WTF iommu_group_remove_device() was supposed to
> do. The __ version make ssense as part of the remove_device, due to
> the sequencing with ops->release()
>
> But the other one doesn't have that. So I want to put in a:
>
> WARN_ON(group->blocking_domain || group->default_domain);
>
> Because calling it after those domains are allocated looks broken to
> me.
I might be misunderstanding, but that sounds backwards - if a real
device is being hotplugged out, we absolutely expect that to happen
*after* its default domain has been set up. The external callers are
using fake groups where default domains aren't relevant, and I have no
idea what PAMU is doing but it's been doing it for long enough that it
most likely isn't a problem. Thus wherever that check would be it would
seem either wrong or unnecessary.
Thanks,
Robin.
next prev parent reply other threads:[~2022-09-09 9:06 UTC|newest]
Thread overview: 24+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-09-08 18:44 [PATCH 0/4] Fix splats releated to using the iommu_group after destroying devices Jason Gunthorpe
2022-09-08 18:44 ` [PATCH 1/4] vfio: Simplify vfio_create_group() Jason Gunthorpe
2022-09-20 19:45 ` Matthew Rosato
2022-09-08 18:44 ` [PATCH 2/4] vfio: Move the sanity check of the group to vfio_create_group() Jason Gunthorpe
2022-09-22 19:10 ` Alex Williamson
2022-09-22 19:36 ` Jason Gunthorpe
2022-09-22 21:23 ` Alex Williamson
2022-09-22 23:12 ` Jason Gunthorpe
2022-09-08 18:45 ` [PATCH 3/4] vfio: Follow a strict lifetime for struct iommu_group * Jason Gunthorpe
2022-09-20 19:32 ` Matthew Rosato
2022-09-08 18:45 ` [PATCH 4/4] iommu: Fix ordering of iommu_release_device() Jason Gunthorpe
2022-09-08 21:05 ` Robin Murphy
2022-09-08 21:27 ` Robin Murphy
2022-09-08 21:43 ` Jason Gunthorpe
2022-09-09 9:05 ` Robin Murphy [this message]
2022-09-09 13:25 ` Jason Gunthorpe
2022-09-09 17:57 ` Robin Murphy
2022-09-09 18:30 ` Jason Gunthorpe
2022-09-09 19:55 ` Robin Murphy
2022-09-09 23:45 ` Jason Gunthorpe
2022-09-12 11:13 ` Robin Murphy
2022-09-22 16:56 ` Jason Gunthorpe
2022-09-09 12:49 ` [PATCH 0/4] Fix splats releated to using the iommu_group after destroying devices Matthew Rosato
2022-09-09 16:24 ` Jason Gunthorpe
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=38bac59a-808d-5e91-227a-a3a06633c091@arm.com \
--to=robin.murphy@arm.com \
--cc=alex.williamson@redhat.com \
--cc=cai@lca.pw \
--cc=cohuck@redhat.com \
--cc=iommu@lists.linux.dev \
--cc=jgg@nvidia.com \
--cc=joro@8bytes.org \
--cc=jroedel@suse.de \
--cc=kvm@vger.kernel.org \
--cc=m.szyprowski@samsung.com \
--cc=mjrosato@linux.ibm.com \
--cc=will@kernel.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).