From: Xu Yilun <yilun.xu@linux.intel.com>
To: Jason Gunthorpe <jgg@nvidia.com>
Cc: "Tian, Kevin" <kevin.tian@intel.com>,
"will@kernel.org" <will@kernel.org>,
"aneesh.kumar@kernel.org" <aneesh.kumar@kernel.org>,
"iommu@lists.linux.dev" <iommu@lists.linux.dev>,
"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
"joro@8bytes.org" <joro@8bytes.org>,
"robin.murphy@arm.com" <robin.murphy@arm.com>,
"shuah@kernel.org" <shuah@kernel.org>,
"nicolinc@nvidia.com" <nicolinc@nvidia.com>,
"aik@amd.com" <aik@amd.com>,
"Williams, Dan J" <dan.j.williams@intel.com>,
"baolu.lu@linux.intel.com" <baolu.lu@linux.intel.com>,
"Xu, Yilun" <yilun.xu@intel.com>
Subject: Re: [PATCH v3 2/5] iommufd: Destroy vdevice on idevice destroy
Date: Mon, 7 Jul 2025 18:58:45 +0800 [thread overview]
Message-ID: <aGuoZUaNOgclC2uU@yilunxu-OptiPlex-7050> (raw)
In-Reply-To: <20250702124042.GD1051729@nvidia.com>
On Wed, Jul 02, 2025 at 09:40:42AM -0300, Jason Gunthorpe wrote:
> On Wed, Jul 02, 2025 at 09:13:50AM +0000, Tian, Kevin wrote:
> > > > Yes, you can't touch idev inside the destroy function at all, under
> > > > any version. idev is only valid if you have a refcount on vdev.
> > > >
> > > > But why are you touching this lock? Arrange things so abort doesn't
> > > > touch the idev??
> > >
> > > idev has a pointer idev->vdev to track the vdev's lifecycle.
> > > idev->igroup->lock protects the pointer. At the end of
> > > iommufd_vdevice_destroy() this pointer should be NULLed so that idev
> > > knows vdev is really destroyed.
>
> Well, that is destroy, not abort, but OK, there is an issue with
> destroy.
>
> > but comparing to that I'd prefer to the original wait approach...
>
> Okay, but lets try to keep the wait hidden inside the refcounting..
>
> The issue here is we don't hold a refcount on idev while working with
> idev. Let's fix that and then things should work properly?
>
> Maybe something like this:
>
> diff --git a/drivers/iommu/iommufd/device.c b/drivers/iommu/iommufd/device.c
> index 4e781aa9fc6329..9174fa7c972b80 100644
> --- a/drivers/iommu/iommufd/device.c
> +++ b/drivers/iommu/iommufd/device.c
> @@ -178,12 +178,20 @@ static void iommufd_device_remove_vdev(struct iommufd_device *idev)
> mutex_unlock(&idev->igroup->lock);
> }
>
> +void iommufd_device_pre_destroy(struct iommufd_object *obj)
> +{
> + struct iommufd_device *idev =
> + container_of(obj, struct iommufd_device, obj);
> +
> + /* Release the short term users on this */
> + iommufd_device_remove_vdev(idev);
> +}
> +
> void iommufd_device_destroy(struct iommufd_object *obj)
> {
> struct iommufd_device *idev =
> container_of(obj, struct iommufd_device, obj);
>
> - iommufd_device_remove_vdev(idev);
> iommu_device_release_dma_owner(idev->dev);
> iommufd_put_group(idev->igroup);
> if (!iommufd_selftest_is_mock_dev(idev->dev))
> diff --git a/drivers/iommu/iommufd/main.c b/drivers/iommu/iommufd/main.c
> index b2e8e106c16158..387c630fdabfbd 100644
> --- a/drivers/iommu/iommufd/main.c
> +++ b/drivers/iommu/iommufd/main.c
> @@ -151,6 +151,9 @@ static int iommufd_object_dec_wait_shortterm(struct iommufd_ctx *ictx,
> if (refcount_dec_and_test(&to_destroy->shortterm_users))
> return 0;
>
> + if (iommufd_object_ops[to_destroy->type].pre_destroy)
> + iommufd_object_ops[to_destroy->type].pre_destroy(to_destroy);
> +
> if (wait_event_timeout(ictx->destroy_wait,
> refcount_read(&to_destroy->shortterm_users) ==i 0,
I still see concerns here. The pre_destroy() and wait_event is before the
idev's users count decreasing to 0, which means a new user could get in.
The worse thing is the new user could hold the shortterm_users until
pre_destroy(), but there isn't a second pre_destroy()...
I think extending the life of shortterm_users at runtime makes things
complex. My idea is keep the original definition of shortterm_users at
runtime, only repurpose it as a wait completion after users refcount == 0.
I.e. idev increase its own shortterm_users, waiting for vdev to decrease
it.
diff --git a/drivers/iommu/iommufd/device.c b/drivers/iommu/iommufd/device.c
index 9876bf70980a..49b787bc0688 100644
--- a/drivers/iommu/iommufd/device.c
+++ b/drivers/iommu/iommufd/device.c
@@ -147,8 +147,15 @@ static void iommufd_device_remove_vdev(struct iommufd_device *idev)
goto out_unlock;
vdev = iommufd_get_vdevice(idev->ictx, idev->vdev->obj.id);
- if (IS_ERR(vdev))
+ if (IS_ERR(vdev)) {
+ /*
+ * vdev is removed from xarray by userspace, but is not
+ * destroyed. We should wait for vdev real destruction before
+ * idev destruction.
+ */
+ iommufd_object_destroy_wait_prepare(&idev->obj);
goto out_unlock;
+ }
/* Should never happen */
if (WARN_ON(vdev != idev->vdev)) {
diff --git a/drivers/iommu/iommufd/iommufd_private.h b/drivers/iommu/iommufd/iommufd_private.h
index 08f3c66366ea..6620006df8d1 100644
--- a/drivers/iommu/iommufd/iommufd_private.h
+++ b/drivers/iommu/iommufd/iommufd_private.h
@@ -180,6 +180,47 @@ static inline void iommufd_put_object(struct iommufd_ctx *ictx,
wake_up_interruptible_all(&ictx->destroy_wait);
}
+/*
+ * The iommufd_object_destroy_wait_xx() are used to synchronize the last minute
+ * destructions of interrelated objects. E.g. the vdev's destroy() should
+ * always be called before idev's destroy().
+ *
+ * iommufd_object_destroy_wait_prepare() should only be called by
+ * later-destroyer in its pre_destroy() ops, where the users & shortterm_users
+ * refcounts are all zeroed. I.e. no new user is possible. Use "users == 0 &&
+ * shortterm_users == 1" as the "destroy wait" state, as this combination
+ * is not used in runtime usecase and it also prevents new users.
+ *
+ * The later-destroyer should use their own locking to ensure the
+ * first-destroyer still exists when calling this function.
+ */
+static inline void
+iommufd_object_destroy_wait_prepare(struct iommufd_object *to_destroy)
+{
+ if (WARN_ON(refcount_read(&to_destroy->users) != 0 ||
+ refcount_read(&to_destroy->shortterm_users) != 0))
+ return;
+
+ refcount_set(&to_destroy->shortterm_users, 1);
+}
+
+/*
+ * iommufd_object_destroy_wait_complete() should be called by first-destroyer
+ * in its destroy() ops. The first-destroyer should not access the
+ * later-destroyer after calling this function.
+ */
+static inline void
+iommufd_object_destroy_wait_complete(struct iommufd_ctx *ictx,
+ struct iommufd_object *to_destroy)
+{
+ if (refcount_read(&to_destroy->users) != 0 ||
+ refcount_read(&to_destroy->shortterm_users) != 1)
+ return;
+
+ refcount_dec(&to_destroy->shortterm_users);
+ wake_up_interruptible_all(&ictx->destroy_wait);
+}
+
void iommufd_object_abort(struct iommufd_ctx *ictx, struct iommufd_object *obj);
void iommufd_object_abort_and_destroy(struct iommufd_ctx *ictx,
struct iommufd_object *obj);
diff --git a/drivers/iommu/iommufd/main.c b/drivers/iommu/iommufd/main.c
index fbccfbd8fb03..148218293f55 100644
--- a/drivers/iommu/iommufd/main.c
+++ b/drivers/iommu/iommufd/main.c
@@ -96,15 +96,37 @@ struct iommufd_object *iommufd_get_object(struct iommufd_ctx *ictx, u32 id,
return obj;
}
+static int iommufd_object_pre_destroy_wait(struct iommufd_ctx *ictx,
+ struct iommufd_object *to_destroy)
+{
+ if (!iommufd_object_ops[to_destroy->type].pre_destroy)
+ return 0;
+
+ /*
+ * pre_destroy() may put the object in destroy wait state (users == 0
+ * && shortterm_users == 1). Another thread completes the waiting by
+ * decreasing the shortterm_users back to 0. The helpers
+ * iommufd_object_destroy_wait_prepare() and
+ * iommufd_object_destroy_wait_complete() should be used for this
+ * purpose.
+ */
+ iommufd_object_ops[to_destroy->type].pre_destroy(to_destroy);
+ if (wait_event_timeout(ictx->destroy_wait,
+ refcount_read(&to_destroy->shortterm_users) ==
+ 0,
+ msecs_to_jiffies(60000)))
+ return 0;
+
+ pr_crit("Time out waiting for iommufd object to become free\n");
+ return -EBUSY;
+}
+
static int iommufd_object_dec_wait_shortterm(struct iommufd_ctx *ictx,
struct iommufd_object *to_destroy)
{
if (refcount_dec_and_test(&to_destroy->shortterm_users))
return 0;
- if (iommufd_object_ops[to_destroy->type].pre_destroy)
- iommufd_object_ops[to_destroy->type].pre_destroy(to_destroy);
-
if (wait_event_timeout(ictx->destroy_wait,
refcount_read(&to_destroy->shortterm_users) ==
0,
@@ -184,6 +206,10 @@ int iommufd_object_remove(struct iommufd_ctx *ictx,
ret = iommufd_object_dec_wait_shortterm(ictx, obj);
if (WARN_ON(ret))
return ret;
+ } else {
+ ret = iommufd_object_pre_destroy_wait(ictx, obj);
+ if (WARN_ON(ret))
+ return ret;
}
iommufd_object_ops[obj->type].destroy(obj);
diff --git a/drivers/iommu/iommufd/viommu.c b/drivers/iommu/iommufd/viommu.c
index d49e8f0c8bb3..a7feefd5e472 100644
--- a/drivers/iommu/iommufd/viommu.c
+++ b/drivers/iommu/iommufd/viommu.c
@@ -103,11 +103,12 @@ void iommufd_vdevice_destroy(struct iommufd_object *obj)
{
struct iommufd_vdevice *vdev =
container_of(obj, struct iommufd_vdevice, obj);
+ struct iommufd_device *idev = vdev->idev;
- mutex_lock(&vdev->idev->igroup->lock);
+ mutex_lock(&idev->igroup->lock);
iommufd_vdevice_abort(obj);
- mutex_unlock(&vdev->idev->igroup->lock);
- iommufd_put_object(vdev->viommu->ictx, &vdev->idev->obj);
+ mutex_unlock(&idev->igroup->lock);
+ iommufd_object_destroy_wait_complete(idev->ictx, &idev->obj);
}
int iommufd_vdevice_alloc_ioctl(struct iommufd_ucmd *ucmd)
@@ -185,8 +186,7 @@ int iommufd_vdevice_alloc_ioctl(struct iommufd_ucmd *ucmd)
out_unlock_igroup:
mutex_unlock(&idev->igroup->lock);
out_put_idev:
- if (rc)
- iommufd_put_object(ucmd->ictx, &idev->obj);
+ iommufd_put_object(ucmd->ictx, &idev->obj);
out_put_viommu:
iommufd_put_object(ucmd->ictx, &viommu->obj);
return rc;
next prev parent reply other threads:[~2025-07-07 11:07 UTC|newest]
Thread overview: 32+ messages / expand[flat|nested] mbox.gz Atom feed top
2025-06-27 3:38 [PATCH v3 0/5] iommufd: Destroy vdevice on device unbind Xu Yilun
2025-06-27 3:38 ` [PATCH v3 1/5] iommufd: Add iommufd_object_tombstone_user() helper Xu Yilun
2025-06-30 3:08 ` Baolu Lu
2025-06-30 7:24 ` Xu Yilun
2025-06-30 5:52 ` Tian, Kevin
2025-06-30 6:41 ` Xu Yilun
2025-06-30 19:50 ` Nicolin Chen
2025-07-08 8:45 ` Xu Yilun
2025-06-27 3:38 ` [PATCH v3 2/5] iommufd: Destroy vdevice on idevice destroy Xu Yilun
2025-06-30 5:08 ` Baolu Lu
2025-07-08 8:34 ` Xu Yilun
2025-06-30 6:27 ` Tian, Kevin
2025-06-30 10:18 ` Xu Yilun
2025-06-30 14:50 ` Jason Gunthorpe
2025-07-01 9:19 ` Xu Yilun
2025-07-01 12:13 ` Jason Gunthorpe
2025-07-02 2:23 ` Xu Yilun
2025-07-02 9:13 ` Tian, Kevin
2025-07-02 12:40 ` Jason Gunthorpe
2025-07-03 4:32 ` Tian, Kevin
2025-07-03 11:21 ` Jason Gunthorpe
2025-07-07 10:58 ` Xu Yilun [this message]
2025-07-07 12:25 ` Jason Gunthorpe
2025-07-07 19:41 ` Xu Yilun
2025-06-30 19:34 ` Nicolin Chen
2025-07-08 8:12 ` Xu Yilun
2025-06-27 3:38 ` [PATCH v3 3/5] iommufd/vdevice: Remove struct device reference from struct vdevice Xu Yilun
2025-06-30 5:11 ` Baolu Lu
2025-07-04 15:06 ` Jason Gunthorpe
2025-06-27 3:38 ` [PATCH v3 4/5] iommufd/selftest: Explicitly skip tests for inapplicable variant Xu Yilun
2025-07-04 15:07 ` Jason Gunthorpe
2025-06-27 3:38 ` [PATCH v3 5/5] iommufd/selftest: Add coverage for vdevice tombstone Xu Yilun
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=aGuoZUaNOgclC2uU@yilunxu-OptiPlex-7050 \
--to=yilun.xu@linux.intel.com \
--cc=aik@amd.com \
--cc=aneesh.kumar@kernel.org \
--cc=baolu.lu@linux.intel.com \
--cc=dan.j.williams@intel.com \
--cc=iommu@lists.linux.dev \
--cc=jgg@nvidia.com \
--cc=joro@8bytes.org \
--cc=kevin.tian@intel.com \
--cc=linux-kernel@vger.kernel.org \
--cc=nicolinc@nvidia.com \
--cc=robin.murphy@arm.com \
--cc=shuah@kernel.org \
--cc=will@kernel.org \
--cc=yilun.xu@intel.com \
/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
all inboxes | Powered by JetHome®