Skip to contentSkip to conversation

9p: fix guest-triggered Treaddir/ACPI eject UAF

PATCH Superseded Christian Schoenebeck 5 participants 21 messages Jul 9 – Jul 20

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

cover 9p: fix guest-triggered Treaddir/ACPI eject UAF

A guest can trigger unplugging 9pfs server's virtio-pci device via ACPI eject. As a consequence the device is unrealized, server's internal state is freed while pending coroutines would still have access to them, causing a potential heap-use-after-free.

Overview Patches:

- Patch 1: this is the core fix, that drains all PDUs (i.e. coroutines that handle individual pending requests in parallel) before freeing server state.

- Patch 2: hardens security by disabling guest triggered ACPI rejects in general for 9pfs's virtio transport.

- Patch 3: fixes a similar identified issue with the Xen transport, even though not triggered via ACPI, it is also prone to UAF, plus a resource leak.

Independent of this series, it should be considered to globally change the default value of devices' hotpluggable property from default enabled, to default disabled. Because as this issue showed, it should be a conscious decision of developers to explicitly opt-in when knowing that their device DOES support hot-unplugging AND that it actually makes sense for the device category. As changing the default value affects a massive amount of devices in QEMU, I am still investigating this and this is of course not covered yet by this series.

Christian Schoenebeck (3):
hw/9pfs/virtio: drain in-flight PDUs before virtio-9p unrealize hw/9pfs/virtio: disable hotpluggable property of virtio-9p device hw/9pfs/xen: drain in-flight PDUs before xen-9p disconnect

 hw/9pfs/virtio-9p-device.c | 2 ++
 hw/9pfs/xen-9p-backend.c   | 4 ++++
 2 files changed, 6 insertions(+)
Signature · 1 line
2.47.3
#1 lore raw

PATCH 1/3 hw/9pfs/virtio: drain in-flight PDUs before virtio-9p unrealize

A guest can trigger a heap-use-after-free in the virtio transport unrealize path by submitting a Treaddir request and immediately ejecting the device via ACPI PCI hotplug. The unrealize path frees struct LocalData while a worker thread still holds a reference on it, causing a UAF in local_open_nofollow().

Fix this by draining all in-flight 9p PDUs by calling v9fs_reset() before final server cleanup. This ensures all coroutines completed, all FIDs are closed, and no worker thread still holds references on 9p server state when it is freed.

Fixes 6cecf09373 ("virtio-9p-device: add minimal unrealize handler")

Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/3937

---
 hw/9pfs/virtio-9p-device.c | 1 +
 1 file changed, 1 insertion(+)
hw/9pfs/virtio-9p-device.c +1 -0
index 50dc93091d..1ec48fc9e0 100644
@@ -243,6 +243,7 @@ static void virtio_9p_device_unrealize(DeviceState *dev)
     V9fsVirtioState *v = VIRTIO_9P(dev);
     V9fsState *s = &v->state;
 
+    v9fs_reset(s);
     virtio_delete_queue(v->vq);
     virtio_cleanup(vdev);
     v9fs_device_unrealize_common(s);
Signature · 1 line
2.47.3
#2 lore raw

Re: [PATCH 1/3] hw/9pfs/virtio: drain in-flight PDUs before virtio-9p unrealize

Quoted 11 lines · On Thursday, 9 July 2026 15:50:36 CEST Christian Schoenebeck wrote:

A guest can trigger a heap-use-after-free in the virtio transport unrealize path by submitting a Treaddir request and immediately ejecting the device via ACPI PCI hotplug. The unrealize path frees struct LocalData while a worker thread still holds a reference on it, causing a UAF in local_open_nofollow().

Fix this by draining all in-flight 9p PDUs by calling v9fs_reset() before final server cleanup. This ensures all coroutines completed, all FIDs are closed, and no worker thread still holds references on 9p server state when it is freed.

Quoted 6 lines
Fixes 6cecf09373 ("virtio-9p-device: add minimal unrealize handler")

Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/3937

---

Jia, if desired, I can make you the official author of this patch. If that's what you want, then please reply-to-all to this message with the same email address that you used for your report already, and acknowledge that you would be fine that I add a:

tag on behalf of you, for the reason described here:

https://www.qemu.org/docs/master/devel/submitting-a-patch.html

Quoted 15 lines
 hw/9pfs/virtio-9p-device.c | 1 +
 1 file changed, 1 insertion(+)
hw/9pfs/virtio-9p-device.c +1 -0
index 50dc93091d..1ec48fc9e0 100644
@@ -243,6 +243,7 @@ static void virtio_9p_device_unrealize(DeviceState *dev)
 V9fsVirtioState *v = VIRTIO_9P(dev);
     V9fsState *s = &v->state;
 
+    v9fs_reset(s);
     virtio_delete_queue(v->vq);
     virtio_cleanup(vdev);
     v9fs_device_unrealize_common(s);
#3 lore raw

Re: [PATCH 1/3] hw/9pfs/virtio: drain in-flight PDUs before virtio-9p unrealize

Hi Christian,

Sure, thanks for handling this.

Please make me the author of patch 1/3, and feel free to add:

on my behalf.

Thanks,
Jia

Quoted 52 lines · Christian Schoenebeck <[email protected]> 于2026年7月9日周四 22:20写道:
Quoted 11 lines · On Thursday, 9 July 2026 15:50:36 CEST Christian Schoenebeck wrote:

A guest can trigger a heap-use-after-free in the virtio transport unrealize path by submitting a Treaddir request and immediately ejecting the device via ACPI PCI hotplug. The unrealize path frees struct LocalData while a worker thread still holds a reference on it, causing a UAF in local_open_nofollow().

Fix this by draining all in-flight 9p PDUs by calling v9fs_reset() before final server cleanup. This ensures all coroutines completed, all FIDs are closed, and no worker thread still holds references on 9p server state when it is freed.

Quoted 6 lines
Fixes 6cecf09373 ("virtio-9p-device: add minimal unrealize handler")

Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/3937

---

Jia, if desired, I can make you the official author of this patch. If that's
what you want, then please reply-to-all to this message with the same email
address that you used for your report already, and acknowledge that you would
be fine that I add a:

tag on behalf of you, for the reason described here:

https://www.qemu.org/docs/master/devel/submitting-a-patch.html

hw/9pfs/virtio-9p-device.c static void virtio_9p_device_unrealize(DeviceState

Quoted 8 lines
 hw/9pfs/virtio-9p-device.c | 1 +
 1 file changed, 1 insertion(+)
hw/9pfs/virtio-9p-device.c +0 -0
index 50dc93091d..1ec48fc9e0 100644
@@ -243,6 +243,7 @@ static void virtio_9p_device_unrealize(DeviceState

*dev)

Quoted 7 lines
hw/9pfs/virtio-9p-device.c +1 -0
 V9fsVirtioState *v = VIRTIO_9P(dev);
     V9fsState *s = &v->state;
 
+    v9fs_reset(s);
     virtio_delete_queue(v->vq);
     virtio_cleanup(vdev);
     v9fs_device_unrealize_common(s);
#4 lore raw

PATCH 2/3 hw/9pfs/virtio: disable hotpluggable property of virtio-9p device

In reply to Christian Schoenebeck

Harden security by disabling hotpluggable property, to prevent issues like fixed in the previous commit.

Virtio-9p is a pass-through file-sharing device that provides a guest mount of a host filesystem tree. Unlike block or network devices, guest-triggered hotplug of the 9p device has no practical use case: the guest cannot recover from the device removal, and there is no protocol-level device-loss scenario as with block or ethernet devices.

Prevent guest-triggered hotplug by marking the device as non-hotpluggable. Device removal via QMP is not affected by this, as the QMP removal path is not reading the device's hotpluggable field at all.

---
 hw/9pfs/virtio-9p-device.c | 1 +
 1 file changed, 1 insertion(+)
hw/9pfs/virtio-9p-device.c +1 -0
index 1ec48fc9e0..ba43fd22d2 100644
@@ -273,6 +273,7 @@ static void virtio_9p_class_init(ObjectClass *klass, const void *data)
 
     device_class_set_props(dc, virtio_9p_properties);
     dc->vmsd = &vmstate_virtio_9p;
+    dc->hotpluggable = false;
     set_bit(DEVICE_CATEGORY_STORAGE, dc->categories);
     vdc->realize = virtio_9p_device_realize;
     vdc->unrealize = virtio_9p_device_unrealize;
Signature · 1 line
2.47.3
#5 lore raw

Re: [PATCH 2/3] hw/9pfs/virtio: disable hotpluggable property of virtio-9p device

Quoted 9 lines · On Thu, 9 Jul 2026 15:50:36 +0200 Christian Schoenebeck…

Harden security by disabling hotpluggable property, to prevent issues like fixed in the previous commit.

Virtio-9p is a pass-through file-sharing device that provides a guest mount of a host filesystem tree. Unlike block or network devices, guest-triggered hotplug of the 9p device has no practical use case: the guest cannot recover from the device removal, and there is no protocol-level device-loss scenario as with block or ethernet devices.

hmh, I'm no maintainer of 9pfs, but to me it looks like any other storage device.
One should be able to unmount/stop using it and unplug
(it doesn't really matter if unplug is triggered by guest or host side).

Quoted 22 lines

Prevent guest-triggered hotplug by marking the device as non-hotpluggable. Device removal via QMP is not affected by this, as the QMP removal path is not reading the device's hotpluggable field at all.

---
 hw/9pfs/virtio-9p-device.c | 1 +
 1 file changed, 1 insertion(+)
hw/9pfs/virtio-9p-device.c +1 -0
index 1ec48fc9e0..ba43fd22d2 100644
@@ -273,6 +273,7 @@ static void virtio_9p_class_init(ObjectClass *klass, const void *data)
 
     device_class_set_props(dc, virtio_9p_properties);
     dc->vmsd = &vmstate_virtio_9p;
+    dc->hotpluggable = false;
     set_bit(DEVICE_CATEGORY_STORAGE, dc->categories);
     vdc->realize = virtio_9p_device_realize;
     vdc->unrealize = virtio_9p_device_unrealize;
#6 lore raw

Re: [PATCH 2/3] hw/9pfs/virtio: disable hotpluggable property of virtio-9p device

Quoted 13 lines · On Friday, 10 July 2026 09:37:37 CEST Igor Mammedov wrote:

On Thu, 9 Jul 2026 15:50:36 +0200

Quoted 9 lines · Christian Schoenebeck

Harden security by disabling hotpluggable property, to prevent issues like fixed in the previous commit.

Virtio-9p is a pass-through file-sharing device that provides a guest mount of a host filesystem tree. Unlike block or network devices, guest-triggered hotplug of the 9p device has no practical use case: the guest cannot recover from the device removal, and there is no protocol-level device-loss scenario as with block or ethernet devices.

Quoted 4 lines

hmh, I'm no maintainer of 9pfs, but to me it looks like any other storage device.
One should be able to unmount/stop using it and unplug
(it doesn't really matter if unplug is triggered by guest or host side).

Guest could still unmount and stop using the 9pfs device. But why should guest be able to unrealize the 9pfs device at any time? What should be the purpose for this particular device?

/Christian

#7 lore raw

Re: [PATCH 2/3] hw/9pfs/virtio: disable hotpluggable property of virtio-9p device

Quoted 14 lines · On Fri, 10 Jul 2026 10:06:29 +0200 Christian Schoenebeck…
Quoted 13 lines · On Friday, 10 July 2026 09:37:37 CEST Igor Mammedov wrote:

On Thu, 9 Jul 2026 15:50:36 +0200

Quoted 9 lines · Christian Schoenebeck

Harden security by disabling hotpluggable property, to prevent issues like fixed in the previous commit.

Virtio-9p is a pass-through file-sharing device that provides a guest mount of a host filesystem tree. Unlike block or network devices, guest-triggered hotplug of the 9p device has no practical use case: the guest cannot recover from the device removal, and there is no protocol-level device-loss scenario as with block or ethernet devices.

Quoted 8 lines
Quoted 4 lines

hmh, I'm no maintainer of 9pfs, but to me it looks like any other storage device.
One should be able to unmount/stop using it and unplug
(it doesn't really matter if unplug is triggered by guest or host side).

Guest could still unmount and stop using the 9pfs device. But why should guest be able to unrealize the 9pfs device at any time? What should be the purpose for this particular device?

unrealize is part of unplug process, at the end of unplug no device should be left attached. You might not need/use it but others might be using it actively.

---

In QEMU impl. it's cooperative process (we don't do surprise removal of PCI devices), where host triggers plug/unplug process and guest side completes it, by:
1. freeing resources on its side 1st (hopefully)
2. issuing eject/poweroff (either via ACPI or native PCIE-HP/SHPC interface)

If guest fails to do #1 it is able to abort process, or it might continue and do #2 with a risk of loosing data but it's guest's business/policy.

If you need to disable hotplug for a specific PCI device instance, use 'hotplug' CLI option on pcie bridge. If you use other than PCI variant of virtio device, then unplug as you describe it isn't reachable, and if it should be disabled in code, it's up to actual frontend device that exposes underlying virtio one.

Quoted 3 lines

/Christian

#8 lore raw

Re: [PATCH 2/3] hw/9pfs/virtio: disable hotpluggable property of virtio-9p device

Quoted 22 lines · On Friday, 10 July 2026 12:23:45 CEST Igor Mammedov wrote:

On Fri, 10 Jul 2026 10:06:29 +0200

Quoted 19 lines · Christian Schoenebeck
Quoted 17 lines · On Friday, 10 July 2026 09:37:37 CEST Igor Mammedov wrote:

On Thu, 9 Jul 2026 15:50:36 +0200

Quoted 9 lines · Christian Schoenebeck

Harden security by disabling hotpluggable property, to prevent issues like fixed in the previous commit.

Virtio-9p is a pass-through file-sharing device that provides a guest mount of a host filesystem tree. Unlike block or network devices, guest-triggered hotplug of the 9p device has no practical use case: the guest cannot recover from the device removal, and there is no protocol-level device-loss scenario as with block or ethernet devices.

hmh, I'm no maintainer of 9pfs, but to me it looks like any other storage device.
One should be able to unmount/stop using it and unplug
(it doesn't really matter if unplug is triggered by guest or host side).

Quoted 7 lines
Quoted 3 lines

Guest could still unmount and stop using the 9pfs device. But why should guest be able to unrealize the 9pfs device at any time? What should be the purpose for this particular device?

unrealize is part of unplug process, at the end of unplug no device should be left attached. You might not need/use it but others might be using it actively.

I got that, but my question was: why would they want to do that with an 9pfs device? Can you make up any real-life scenario why a guest might want to remove a 9p device, given there is no way for them to bring the device back?

Looking at the fixed issue (patch 1), my impression was that original 9p server developers were unaware that guest can actually trigger a device unrealize via ACPI eject. Most probably because hotplugging is enabled by default for all devices in QEMU

Independent of this patch here, I'm therefore currently investigating whether enabling hotplugging by default in QEMU actually make sense. To me, this should be an opt-in, not the other way around. Because device developers should make sure that ejecting a device a) makes sense for the device type in the first place and most importantly b) that ejecting the device works (without data loss and without negative security impact, as it was the case here).

Quoted 8 lines

---

In QEMU impl. it's cooperative process (we don't do surprise removal of PCI devices), where host triggers plug/unplug process and guest side completes it, by: 1. freeing resources on its side 1st (hopefully)
2. issuing eject/poweroff (either via ACPI or native PCIE-HP/SHPC
interface)

Quoted 8 lines

If guest fails to do #1 it is able to abort process, or it might continue and do #2 with a risk of loosing data but it's guest's business/policy.

If you need to disable hotplug for a specific PCI device instance, use 'hotplug' CLI option on pcie bridge. If you use other than PCI variant of virtio device, then unplug as you describe it isn't reachable, and if it should be disabled in code, it's up to actual frontend device that exposes underlying virtio one.

What does that have to do with this issue here? The normal 9pfs live-cycle is to create the 9p device at QEMU startup, and unrealizing the 9p device at QEMU shutdown.

If guest does an ACPI eject, then the only way to bring the 9p device back is to restart QEMU with the required -virtfs / -fsdev -device CLI options.

The Linux 9p client (guest) does never trigger an ACPI reject on unmount. It just marks the virtio channel as unused. If client would eject it, it would not be possible for guest to re-mount it later on.

/Christian

#9 lore raw

Re: [PATCH 2/3] hw/9pfs/virtio: disable hotpluggable property of virtio-9p device

Quoted 27 lines · On Fri, 10 Jul 2026 12:49:06 +0200 Christian Schoenebeck…
Quoted 26 lines · On Friday, 10 July 2026 12:23:45 CEST Igor Mammedov wrote:

On Fri, 10 Jul 2026 10:06:29 +0200

Quoted 22 lines · Christian Schoenebeck
Quoted 13 lines · On Friday, 10 July 2026 09:37:37 CEST Igor Mammedov wrote:

On Thu, 9 Jul 2026 15:50:36 +0200

Quoted 9 lines · Christian Schoenebeck

Harden security by disabling hotpluggable property, to prevent issues like fixed in the previous commit.

Virtio-9p is a pass-through file-sharing device that provides a guest mount of a host filesystem tree. Unlike block or network devices, guest-triggered hotplug of the 9p device has no practical use case: the guest cannot recover from the device removal, and there is no protocol-level device-loss scenario as with block or ethernet devices.

Quoted 4 lines

hmh, I'm no maintainer of 9pfs, but to me it looks like any other storage device.
One should be able to unmount/stop using it and unplug
(it doesn't really matter if unplug is triggered by guest or host side).

Guest could still unmount and stop using the 9pfs device. But why should guest be able to unrealize the 9pfs device at any time? What should be the purpose for this particular device?

Quoted 7 lines
Quoted 3 lines

unrealize is part of unplug process, at the end of unplug no device should be left attached. You might not need/use it but others might be using it actively.

I got that, but my question was: why would they want to do that with an 9pfs device? Can you make up any real-life scenario why a guest might want to remove a 9p device, given there is no way for them to bring the device back?

note: I'm looking from pov of hotpluggable PCI device and generic hotplug infra, only.

it's not guest users directly, it's how hotplug flow works for various guest OSes:

1. host plugs device in (-device or device_add)
2. guest OS get's notified one way or another and does what ever guest side init needed (incl. mounting share in 9pfs case)

opposite flow:
1. host does device_del (basically notify guest to remove device)
2. guest OS frees resources and tells qemu to delete device
3. qemu process remove event (which incl. unrealize as part of destroying device)

Of cause guest if free to eject device without signal from host, (I could imagine: get some file from share once and then guest releases no longer need resource).

Quoted 3 lines

Looking at the fixed issue (patch 1), my impression was that original 9p server developers were unaware that guest can actually trigger a device unrealize via ACPI eject.

I'd say it's a bug, and you are trying to fix it in patch #1

Quoted 2 lines

Most probably because hotplugging is enabled by
default for all devices in QEMU

it is on by default for PCI devices but also heavily depends on used configuration (where/what is plugged).

Quoted 7 lines

Independent of this patch here, I'm therefore currently investigating whether enabling hotplugging by default in QEMU actually make sense. To me, this should be an opt-in, not the other way around. Because device developers should make sure that ejecting a device a) makes sense for the device type in the first place and most importantly b) that ejecting the device works (without data loss and without negative security impact, as it was the case here).

well, hotplug is on by default for PCI devices. It is unlikely we would flip default now.

It might make sense to disable hotplug in mmio frontend if such exists, but then it's pointless as there is no way to trigger unplug for it (both host or guest side)

Quoted 18 lines
Quoted 9 lines

---

In QEMU impl. it's cooperative process (we don't do surprise removal of PCI devices), where host triggers plug/unplug process and guest side completes it, by: 1. freeing resources on its side 1st (hopefully)
2. issuing eject/poweroff (either via ACPI or native PCIE-HP/SHPC
interface)

Quoted 8 lines

If guest fails to do #1 it is able to abort process, or it might continue and do #2 with a risk of loosing data but it's guest's business/policy.

If you need to disable hotplug for a specific PCI device instance, use 'hotplug' CLI option on pcie bridge. If you use other than PCI variant of virtio device, then unplug as you describe it isn't reachable, and if it should be disabled in code, it's up to actual frontend device that exposes underlying virtio one.

Quoted 6 lines

What does that have to do with this issue here? The normal 9pfs live-cycle is to create the 9p device at QEMU startup, and unrealizing the 9p device at QEMU shutdown.

If guest does an ACPI eject, then the only way to bring the 9p device back is to restart QEMU with the required -virtfs / -fsdev -device CLI options.

one can use device_add to put it back into running machine.

given mention of ACPI eject, 9pfs flavor actually used is PCI variant => all PCI device features apply incl. hot(un)plug.
If hotplug is not desirable for particular device instance, one can disable it on CLI.

that said,
if 9pfs maintainer deem that it shouldn't be hotpluggable and it's safe to remove capability, I'm fine with that too, just fix commit message section wrt QMP removal and perhaps move the flag into front-end impl of the device.

BTW: always describe a way to reproduce problem (cover letter and/or patch) so that reviewer could see whole flow. Otherwise one has to guess what/how it gets broken based on context.

Quoted 3 lines

The Linux 9p client (guest) does never trigger an ACPI reject on unmount. It just marks the virtio channel as unused. If client would eject it, it would not be possible for guest to re-mount it later on.

no argument here

Quoted 5 lines

/Christian

#10 lore raw

Re: [PATCH 2/3] hw/9pfs/virtio: disable hotpluggable property of virtio-9p device

Quoted 4 lines · On Friday, 10 July 2026 14:51:49 CEST Igor Mammedov wrote:

On Fri, 10 Jul 2026 12:49:06 +0200

Quoted 1 line · Christian Schoenebeck

On Friday, 10 July 2026 12:23:45 CEST Igor Mammedov wrote:

[...]

Quoted 2 lines

note: I'm looking from pov of hotpluggable PCI device and generic hotplug infra, only.

That's okay, but so far I don't see the relevance for this particular 9p device.

Quoted 6 lines

it's not guest users directly, it's how hotplug flow works for various guest OSes:

1. host plugs device in (-device or device_add)
2. guest OS get's notified one way or another and does what ever guest side init needed (incl. mounting share in 9pfs case)

That's not affected by this patch, right?

Quoted 5 lines

opposite flow:
1. host does device_del (basically notify guest to remove device)
2. guest OS frees resources and tells qemu to delete device
3. qemu process remove event (which incl. unrealize as part of destroying device)

And that's not affected by this patch either, right?

Quoted 1 line

Of cause guest if free to eject device without signal from host,

OK, here is the point where we deviate: you are apparently seeing this from a purely theoretical PoV.

I am facing reality: for several years I'm the only person taking care about this piece of code at all (on a side channel, for free, next to my actual work). And for several months I get AI generated security reports thrown at me, where I have to a) filter legit ones, and b) fix those legit security issues.

For that reason, I am tightening security wherever I can, to prevent further flood.

So the question here is: are you concerned about a real-life issue being introduced by disabling hotplugging for 9pfs specifically?

Quoted 2 lines

(I could imagine: get some file from share once and then guest releases no longer need resource).

It's a pass-through file system. Ejecting the device does not really free a noteworthy amount of resources.

Quoted 11 lines
Quoted 3 lines

Looking at the fixed issue (patch 1), my impression was that original 9p server developers were unaware that guest can actually trigger a device unrealize via ACPI eject.

I'd say it's a bug, and you are trying to fix it in patch #1

Quoted 2 lines

Most probably because hotplugging is enabled by
default for all devices in QEMU

it is on by default for PCI devices but also heavily depends on used configuration (where/what is plugged).

Exactly! And I was trying to explain, that for this particular use case (9pfs), I don't see any real-live use-case for allowing a guest to eject the 9p device at runtime.

So why allowing it? Just for fun?

Quoted 10 lines
Quoted 7 lines

Independent of this patch here, I'm therefore currently investigating whether enabling hotplugging by default in QEMU actually make sense. To me, this should be an opt-in, not the other way around. Because device developers should make sure that ejecting a device a) makes sense for the device type in the first place and most importantly b) that ejecting the device works (without data loss and without negative security impact, as it was the case here).

well, hotplug is on by default for PCI devices. It is unlikely we would flip default now.

I am investigating to flip the default value for good reasons: I made a quick scan on the huge list of devices in QEMU and found quite a number that are affected by a similar UAF issue on guest ACPI eject. What do those devices have in common? Hotplug is enabled by default, not explicitly by the respective device.

Quoted 2 lines

that said, if 9pfs maintainer deem that it shouldn't be hotpluggable and it's safe to remove capability,

That's me, unfortunately. Singular.

Quoted 3 lines

I'm fine with that too, just fix commit message
section wrt QMP removal and perhaps move the flag into front-end impl of the device.

Yeah, I just noticed that as well, that this should rather go to the PCI bridge device. This patch applied it to the inner device, which is wrong.

Quoted 3 lines

BTW: always describe a way to reproduce problem (cover letter and/or patch) so that reviewer could see whole flow. Otherwise one has to guess what/how it gets broken based on context.

The original report (private) is linked by patch 1:
https://gitlab.com/qemu-project/qemu/-/work_items/3937

Anyway, I'll think some more days about this, whether I should fix or drop the patch.

If you have more feedback, always appreciated of course.

Thanks!

/Christian

#11 lore raw

Re: [PATCH 2/3] hw/9pfs/virtio: disable hotpluggable property of virtio-9p device

Quoted 61 lines · Christian Schoenebeck
Quoted 4 lines · On Friday, 10 July 2026 14:51:49 CEST Igor Mammedov wrote:

On Fri, 10 Jul 2026 12:49:06 +0200

Quoted 1 line · Christian Schoenebeck

On Friday, 10 July 2026 12:23:45 CEST Igor Mammedov wrote:

[...]

Quoted 2 lines

note: I'm looking from pov of hotpluggable PCI device and generic hotplug infra, only.

That's okay, but so far I don't see the relevance for this particular 9p device.

Quoted 6 lines

it's not guest users directly, it's how hotplug flow works for various guest OSes:

1. host plugs device in (-device or device_add)
2. guest OS get's notified one way or another and does what ever guest side init needed (incl. mounting share in 9pfs case)

That's not affected by this patch, right?

Quoted 5 lines

opposite flow:
1. host does device_del (basically notify guest to remove device)
2. guest OS frees resources and tells qemu to delete device
3. qemu process remove event (which incl. unrealize as part of destroying device)

And that's not affected by this patch either, right?

Quoted 1 line

Of cause guest if free to eject device without signal from host,

OK, here is the point where we deviate: you are apparently seeing this from a purely theoretical PoV.

I am facing reality: for several years I'm the only person taking care about this piece of code at all (on a side channel, for free, next to my actual work). And for several months I get AI generated security reports thrown at me, where I have to a) filter legit ones, and b) fix those legit security issues.

For that reason, I am tightening security wherever I can, to prevent further flood.

So the question here is: are you concerned about a real-life issue being introduced by disabling hotplugging for 9pfs specifically?

Quoted 2 lines

(I could imagine: get some file from share once and then guest releases no longer need resource).

It's a pass-through file system. Ejecting the device does not really free a noteworthy amount of resources.

Quoted 9 lines
Quoted 3 lines

Looking at the fixed issue (patch 1), my impression was that original 9p server developers were unaware that guest can actually trigger a device unrealize via ACPI eject.

I'd say it's a bug, and you are trying to fix it in patch #1

Quoted 2 lines

Most probably because hotplugging is enabled by
default for all devices in QEMU

Quoted 8 lines
Quoted 2 lines

it is on by default for PCI devices but also heavily depends on used configuration (where/what is plugged).

Exactly! And I was trying to explain, that for this particular use case (9pfs), I don't see any real-live use-case for allowing a guest to eject the 9p device at runtime.

So why allowing it? Just for fun?

I don't think it would be usual for a guest to unilaterally "eject" a 9p device. A host initiated PCI hot-unplug request from QEMU would ultimately trigger the guest eject code path though. IMHO host initiated hotplug+unplug is a valid use for any device. For 9p the use is to provide a time limited data share to some workload in the guest. There are many other ways that can be done too, USB MTP, or virtiofs, or in guest with SFTP or WebDAV, etc.

With regards,
Daniel

Signature · 3 lines
|: https://berrange.com       ~~        https://hachyderm.io/@berrange :|
|: https://libvirt.org          ~~          https://entangle-photo.org :|
|: https://pixelfed.art/berrange   ~~    https://fstop138.berrange.com :|
#12 lore raw

Re: [PATCH 2/3] hw/9pfs/virtio: disable hotpluggable property of virtio-9p device

Quoted 22 lines · On Fri, 10 Jul 2026 15:40:04 +0100 Daniel P. Berrangé <[email protected]>…
Quoted 21 lines · Christian Schoenebeck
Quoted 4 lines · On Friday, 10 July 2026 14:51:49 CEST Igor Mammedov wrote:

On Fri, 10 Jul 2026 12:49:06 +0200

Quoted 1 line · Christian Schoenebeck

On Friday, 10 July 2026 12:23:45 CEST Igor Mammedov wrote:

[...]

Quoted 2 lines

note: I'm looking from pov of hotpluggable PCI device and generic hotplug infra, only.

That's okay, but so far I don't see the relevance for this particular 9p device.

Quoted 6 lines

it's not guest users directly, it's how hotplug flow works for various guest OSes:

1. host plugs device in (-device or device_add)
2. guest OS get's notified one way or another and does what ever guest side init needed (incl. mounting share in 9pfs case)

That's not affected by this patch, right?

Quoted 7 lines
Quoted 7 lines
Quoted 5 lines

opposite flow:
1. host does device_del (basically notify guest to remove device)
2. guest OS frees resources and tells qemu to delete device
3. qemu process remove event (which incl. unrealize as part of destroying device)

And that's not affected by this patch either, right?

the patch would break unplug flow by effectively removing 'eject' knob from guest side, which is part of unplug flow.

Quoted 12 lines
Quoted 12 lines
Quoted 2 lines

Of cause guest if free to eject device without signal from host,

OK, here is the point where we deviate: you are apparently seeing this from a purely theoretical PoV.

I am facing reality: for several years I'm the only person taking care about this piece of code at all (on a side channel, for free, next to my actual work). And for several months I get AI generated security reports thrown at me, where I have to a) filter legit ones, and b) fix those legit security issues.

Quoted 5 lines
Quoted 5 lines

For that reason, I am tightening security wherever I can, to prevent further flood.

So the question here is: are you concerned about a real-life issue being introduced by disabling hotplugging for 9pfs specifically?

my concern is that there might be users that use unplug despite present bug(s), and the patch would regress their usecase.

If unplug 'works', then we shouldn't disable it.
If it's never worked to begin with, it should be safe to disable it.

Quoted 35 lines
Quoted 15 lines
Quoted 2 lines

(I could imagine: get some file from share once and then guest releases no longer need resource).

It's a pass-through file system. Ejecting the device does not really free a noteworthy amount of resources.

Quoted 9 lines
Quoted 3 lines

Looking at the fixed issue (patch 1), my impression was that original 9p server developers were unaware that guest can actually trigger a device unrealize via ACPI eject.

I'd say it's a bug, and you are trying to fix it in patch #1

Quoted 2 lines

Most probably because hotplugging is enabled by
default for all devices in QEMU

Quoted 8 lines
Quoted 2 lines

it is on by default for PCI devices but also heavily depends on used configuration (where/what is plugged).

Exactly! And I was trying to explain, that for this particular use case (9pfs), I don't see any real-live use-case for allowing a guest to eject the 9p device at runtime.

So why allowing it? Just for fun?

I don't think it would be usual for a guest to unilaterally "eject" a 9p device. A host initiated PCI hot-unplug request from QEMU would ultimately trigger the guest eject code path though. IMHO host initiated hotplug+unplug is a valid use for any device. For 9p the use is to provide a time limited data share to some workload in the guest. There are many other ways that can be done too, USB MTP, or virtiofs, or in guest with SFTP or WebDAV, etc.

With regards,
Daniel

#13 lore raw

Re: [PATCH 2/3] hw/9pfs/virtio: disable hotpluggable property of virtio-9p device

Quoted 36 lines · On Monday, 13 July 2026 09:49:35 CEST Igor Mammedov wrote:

On Fri, 10 Jul 2026 15:40:04 +0100

Quoted 33 lines · Daniel P. Berrangé
Quoted 32 lines · Christian Schoenebeck
Quoted 4 lines · On Friday, 10 July 2026 14:51:49 CEST Igor Mammedov wrote:

On Fri, 10 Jul 2026 12:49:06 +0200

Quoted 1 line · Christian Schoenebeck

On Friday, 10 July 2026 12:23:45 CEST Igor Mammedov wrote:

[...]

Quoted 3 lines

note: I'm looking from pov of hotpluggable PCI device and generic hotplug
infra, only.

That's okay, but so far I don't see the relevance for this particular 9p device.

Quoted 8 lines

it's not guest users directly, it's how hotplug flow works for various guest OSes:

1. host plugs device in (-device or device_add)
2. guest OS get's notified one way or another and does what ever guest side

init needed (incl. mounting share in 9pfs case)

That's not affected by this patch, right?

Quoted 6 lines

opposite flow:
1. host does device_del (basically notify guest to remove device)
2. guest OS frees resources and tells qemu to delete device
3. qemu process remove event (which incl. unrealize as part of destroying
device)

Quoted 4 lines
Quoted 1 line
Quoted 1 line

And that's not affected by this patch either, right?

the patch would break unplug flow by effectively removing 'eject' knob from guest side, which is part of unplug flow.

These are two different things: device_del would not be affected by this.

1. Ejecting the device from host side e.g. via QMP would still work.

vs.

2. Ejecting from guest side OTOH would be disabled.

It is also different from regular block devices where you have convenient ways to eject a block device on guest OSes. For a virtio-9p device it is not that easy.

Quoted 9 lines
Quoted 9 lines
Quoted 9 lines

OK, here is the point where we deviate: you are apparently seeing this from a purely theoretical PoV.

I am facing reality: for several years I'm the only person taking care about this piece of code at all (on a side channel, for free, next to my actual work). And for several months I get AI generated security reports thrown at me, where I have to a) filter legit ones, and b) fix those legit security issues.

Quoted 8 lines
Quoted 5 lines
Quoted 5 lines

For that reason, I am tightening security wherever I can, to prevent further flood.

So the question here is: are you concerned about a real-life issue being introduced by disabling hotplugging for 9pfs specifically?

my concern is that there might be users that use unplug despite present bug(s), and the patch would regress their usecase.

I already got that with your first message, and directly asked which use-case that would be exactly.

Just saying there might be one theoretical unknown person on this planet who might be using it, and nobody even being capable to imagine what for, does not justify keeping an exotic feature (again: context 9pfs) that already had negative impact on security for anybody else, especially for such a low man power codebase like 9pfs.

This is similar to a discussion in the past about retaining support for older macOS versions in QEMU:

https://lore.kernel.org/qemu-devel/CAFEAcA9VnQCG3r28BOLr_qXLRM2V68r3oK4ZfY9+bVz2j1oSyA@mail.gmail.com/

Anyway, I decided to drop this patch and retaining guest eject for now.

However if there is any other kind of bug, especially of security nature, then I will definitely disable guest eject unless proven there are justified real- world users or somebody willing to step-in and keeping care of.

/Christian

#14 lore raw

Re: [PATCH 2/3] hw/9pfs/virtio: disable hotpluggable property of virtio-9p device

Quoted 46 lines · Christian Schoenebeck
Quoted 36 lines · On Monday, 13 July 2026 09:49:35 CEST Igor Mammedov wrote:

On Fri, 10 Jul 2026 15:40:04 +0100

Quoted 33 lines · Daniel P. Berrangé
Quoted 32 lines · Christian Schoenebeck
Quoted 4 lines · On Friday, 10 July 2026 14:51:49 CEST Igor Mammedov wrote:

On Fri, 10 Jul 2026 12:49:06 +0200

Quoted 1 line · Christian Schoenebeck

On Friday, 10 July 2026 12:23:45 CEST Igor Mammedov wrote:

[...]

Quoted 3 lines

note: I'm looking from pov of hotpluggable PCI device and generic hotplug
infra, only.

That's okay, but so far I don't see the relevance for this particular 9p device.

Quoted 8 lines

it's not guest users directly, it's how hotplug flow works for various guest OSes:

1. host plugs device in (-device or device_add)
2. guest OS get's notified one way or another and does what ever guest side

init needed (incl. mounting share in 9pfs case)

That's not affected by this patch, right?

Quoted 6 lines

opposite flow:
1. host does device_del (basically notify guest to remove device)
2. guest OS frees resources and tells qemu to delete device
3. qemu process remove event (which incl. unrealize as part of destroying
device)

Quoted 4 lines
Quoted 1 line
Quoted 1 line

And that's not affected by this patch either, right?

the patch would break unplug flow by effectively removing 'eject' knob from guest side, which is part of unplug flow.

These are two different things: device_del would not be affected by this.

1. Ejecting the device from host side e.g. via QMP would still work.

Quoted 7 lines

vs.

2. Ejecting from guest side OTOH would be disabled.

It is also different from regular block devices where you have convenient ways to eject a block device on guest OSes. For a virtio-9p device it is not that easy.

Igor can correct me if I'm wrong, but IIUC there's no distinction to those 2 scenarios from the guest POV.

A "device_del" in QMP does not actually delete a device from the QEMU POV. Rather it triggers an ACPI notification to the geust OS, that the operator wants to unplug the device. The guest OS then honours (or not) that by initiating a guest side "eject". IOW disabling guest initiated eject would break device_del too AFAIK.

With regards,
Daniel

Signature · 3 lines
|: https://berrange.com       ~~        https://hachyderm.io/@berrange :|
|: https://libvirt.org          ~~          https://entangle-photo.org :|
|: https://pixelfed.art/berrange   ~~    https://fstop138.berrange.com :|
#15 lore raw

Re: [PATCH 2/3] hw/9pfs/virtio: disable hotpluggable property of virtio-9p device

Quoted 58 lines · On Fri, 17 Jul 2026 15:16:14 +0100 Daniel P. Berrangé <[email protected]>…
Quoted 46 lines · Christian Schoenebeck
Quoted 36 lines · On Monday, 13 July 2026 09:49:35 CEST Igor Mammedov wrote:

On Fri, 10 Jul 2026 15:40:04 +0100

Quoted 33 lines · Daniel P. Berrangé
Quoted 32 lines · Christian Schoenebeck
Quoted 4 lines · On Friday, 10 July 2026 14:51:49 CEST Igor Mammedov wrote:

On Fri, 10 Jul 2026 12:49:06 +0200

Quoted 1 line · Christian Schoenebeck

On Friday, 10 July 2026 12:23:45 CEST Igor Mammedov wrote:

[...]

Quoted 3 lines

note: I'm looking from pov of hotpluggable PCI device and generic hotplug
infra, only.

That's okay, but so far I don't see the relevance for this particular 9p device.

Quoted 8 lines

it's not guest users directly, it's how hotplug flow works for various guest OSes:

1. host plugs device in (-device or device_add)
2. guest OS get's notified one way or another and does what ever guest side

init needed (incl. mounting share in 9pfs case)

That's not affected by this patch, right?

Quoted 6 lines

opposite flow:
1. host does device_del (basically notify guest to remove device)
2. guest OS frees resources and tells qemu to delete device
3. qemu process remove event (which incl. unrealize as part of destroying
device)

Quoted 4 lines
Quoted 1 line
Quoted 1 line

And that's not affected by this patch either, right?

the patch would break unplug flow by effectively removing 'eject' knob from guest side, which is part of unplug flow.

These are two different things: device_del would not be affected by this.

1. Ejecting the device from host side e.g. via QMP would still work.

Quoted 7 lines

vs.

2. Ejecting from guest side OTOH would be disabled.

It is also different from regular block devices where you have convenient ways to eject a block device on guest OSes. For a virtio-9p device it is not that easy.

Igor can correct me if I'm wrong, but IIUC there's no distinction to those 2 scenarios from the guest POV.

Quoted 6 lines

A "device_del" in QMP does not actually delete a device from the QEMU POV. Rather it triggers an ACPI notification to the geust OS, that the operator wants to unplug the device. The guest OS then honours (or not) that by initiating a guest side "eject". IOW disabling guest initiated eject would break device_del too AFAIK.

That's how it works.
Sorry if I wasn't able to convey the message in my previous replies.

Quoted 3 lines

With regards,
Daniel

#16 lore raw

Re: [PATCH 2/3] hw/9pfs/virtio: disable hotpluggable property of virtio-9p device

Quoted 5 lines · On Monday, 20 July 2026 13:23:37 CEST Igor Mammedov wrote:

On Fri, 17 Jul 2026 15:16:14 +0100

Quoted 2 lines · Daniel P. Berrangé
Quoted 1 line · Christian Schoenebeck

On Monday, 13 July 2026 09:49:35 CEST Igor Mammedov wrote:

[...]

Quoted 26 lines
Quoted 19 lines
Quoted 8 lines
Quoted 2 lines

the patch would break unplug flow by effectively removing 'eject' knob from guest side, which is part of unplug flow.

These are two different things: device_del would not be affected by this.

1. Ejecting the device from host side e.g. via QMP would still work.

Quoted 7 lines

vs.

2. Ejecting from guest side OTOH would be disabled.

It is also different from regular block devices where you have convenient ways to eject a block device on guest OSes. For a virtio-9p device it is not that easy.

Igor can correct me if I'm wrong, but IIUC there's no distinction to those 2 scenarios from the guest POV.

Quoted 6 lines

A "device_del" in QMP does not actually delete a device from the QEMU POV. Rather it triggers an ACPI notification to the geust OS, that the operator wants to unplug the device. The guest OS then honours (or not) that by initiating a guest side "eject". IOW disabling guest initiated eject would break device_del too AFAIK.

Quoted 2 lines

That's how it works.
Sorry if I wasn't able to convey the message in my previous replies.

OK, I just realized I was wrong. The QMP device_del path also checks for dc->hotpluggable early via qdev_hotplug_unplug_allowed_common() and would already abort there. And I can also see the flow that it just triggers ACPI and guest eventually completing device eject.

So yes, it would disable both guest *and* host side eject.

I already dropped this patch from v2 anyway.

Thanks!

/Christian

#17 lore raw

PATCH 3/3 hw/9pfs/xen: drain in-flight PDUs before xen-9p disconnect

In reply to Christian Schoenebeck

The xen-9p disconnect path has two issues:

1. It frees the Xen9pfsRing structures while in-flight PDUs may still reference them via pdu->tag to index rings[]. This causes a UAF in xen_9pfs_push_and_notify() when worker threads resume after completing filesystem operations.

2. It never calls v9fs_device_unrealize_common(), which means server state (struct LocalData, mountfd, FIDs) is never cleaned up on disconnect, causing a resource leak on every guest-initiated disconnect.

Fix both by draining in-flight PDUs via v9fs_reset() before tearing down rings, and calling v9fs_device_unrealize_common() to clean up server state.

Fixes b37eeb0201 ("xen/9pfs: introduce Xen 9pfs backend")
---
 hw/9pfs/xen-9p-backend.c | 4 ++++
 1 file changed, 4 insertions(+)
hw/9pfs/xen-9p-backend.c +4 -0
index 24c90d97ec..3b7a68779a 100644
@@ -368,10 +368,14 @@ static void xen_9pfs_evtchn_event(void *opaque)
 static void xen_9pfs_disconnect(struct XenLegacyDevice *xendev)
 {
     Xen9pfsDev *xen_9pdev = container_of(xendev, Xen9pfsDev, xendev);
+    V9fsState *s = &xen_9pdev->state;
     int i;
 
     trace_xen_9pfs_disconnect(xendev->name);
 
+    v9fs_reset(s);
+    v9fs_device_unrealize_common(s);
+
     for (i = 0; i < xen_9pdev->num_rings; i++) {
         if (xen_9pdev->rings[i].evtchndev != NULL) {
             qemu_set_fd_handler(qemu_xen_evtchn_fd(xen_9pdev->rings[i].evtchndev),
Signature · 1 line
2.47.3
#18 lore raw

Re: [PATCH 3/3] hw/9pfs/xen: drain in-flight PDUs before xen-9p disconnect

Hi Stefano,

do you might have a chance to look at this Xen patch?

/Christian

Quoted 42 lines · On Thursday, 9 July 2026 15:50:36 CEST Christian Schoenebeck wrote:

The xen-9p disconnect path has two issues:

1. It frees the Xen9pfsRing structures while in-flight PDUs may still reference them via pdu->tag to index rings[]. This causes a UAF in xen_9pfs_push_and_notify() when worker threads resume after completing filesystem operations.

2. It never calls v9fs_device_unrealize_common(), which means server state (struct LocalData, mountfd, FIDs) is never cleaned up on disconnect, causing a resource leak on every guest-initiated disconnect.

Fix both by draining in-flight PDUs via v9fs_reset() before tearing down rings, and calling v9fs_device_unrealize_common() to clean up server state.

Fixes b37eeb0201 ("xen/9pfs: introduce Xen 9pfs backend")
---
 hw/9pfs/xen-9p-backend.c | 4 ++++
 1 file changed, 4 insertions(+)
hw/9pfs/xen-9p-backend.c +4 -0
index 24c90d97ec..3b7a68779a 100644
@@ -368,10 +368,14 @@ static void xen_9pfs_evtchn_event(void *opaque)
 static void xen_9pfs_disconnect(struct XenLegacyDevice *xendev)
 {
     Xen9pfsDev *xen_9pdev = container_of(xendev, Xen9pfsDev, xendev);
+    V9fsState *s = &xen_9pdev->state;
     int i;
 
     trace_xen_9pfs_disconnect(xendev->name);
 
+    v9fs_reset(s);
+    v9fs_device_unrealize_common(s);
+
     for (i = 0; i < xen_9pdev->num_rings; i++) {
         if (xen_9pdev->rings[i].evtchndev != NULL) {
 

qemu_set_fd_handler(qemu_xen_evtchn_fd(xen_9pdev->rings[i].evtchndev),

#19 lore raw

Re: [PATCH 3/3] hw/9pfs/xen: drain in-flight PDUs before xen-9p disconnect

Quoted 3 lines · Christian Schoenebeck

Hi Stefano,

do you might have a chance to look at this Xen patch?

Sorry for the delay

Quoted 36 lines

/Christian

Quoted 33 lines · On Thursday, 9 July 2026 15:50:36 CEST Christian Schoenebeck wrote:

The xen-9p disconnect path has two issues:

1. It frees the Xen9pfsRing structures while in-flight PDUs may still reference them via pdu->tag to index rings[]. This causes a UAF in xen_9pfs_push_and_notify() when worker threads resume after completing filesystem operations.

2. It never calls v9fs_device_unrealize_common(), which means server state (struct LocalData, mountfd, FIDs) is never cleaned up on disconnect, causing a resource leak on every guest-initiated disconnect.

Fix both by draining in-flight PDUs via v9fs_reset() before tearing down rings, and calling v9fs_device_unrealize_common() to clean up server state.

Fixes b37eeb0201 ("xen/9pfs: introduce Xen 9pfs backend")
---
 hw/9pfs/xen-9p-backend.c | 4 ++++
 1 file changed, 4 insertions(+)
hw/9pfs/xen-9p-backend.c +1 -0
index 24c90d97ec..3b7a68779a 100644
@@ -368,10 +368,14 @@ static void xen_9pfs_evtchn_event(void *opaque)
 static void xen_9pfs_disconnect(struct XenLegacyDevice *xendev)
 {
     Xen9pfsDev *xen_9pdev = container_of(xendev, Xen9pfsDev, xendev);
+    V9fsState *s = &xen_9pdev->state;
     int i;
 
Quoted 4 lines

hw/9pfs/xen-9p-backend.c static void xen_9pfs_evtchn_event(void *opaque) · line 377

Quoted 4 lines
hw/9pfs/xen-9p-backend.c +2 -0
     trace_xen_9pfs_disconnect(xendev->name);
 
+    v9fs_reset(s);
+    v9fs_device_unrealize_common(s);

While the patch addresses a real issue, xen_9pfs_disconnect() is also called from the error paths of xen_9pfs_pdu_vmarshal/vunmarshal, which run inside the PDU's own coroutine. v9fs_reset() is not safe to be called there.

We could remove the direct call to xen_9pfs_disconnect() from xen_9pfs_pdu_vmarshal/vunmarshal because they already change the state to XenbusStateClosing, which should result in the same

Quoted 8 lines
Quoted 4 lines
hw/9pfs/xen-9p-backend.c +0 -0
     for (i = 0; i < xen_9pdev->num_rings; i++) {
         if (xen_9pdev->rings[i].evtchndev != NULL) {
 

qemu_set_fd_handler(qemu_xen_evtchn_fd(xen_9pdev->rings[i].evtchndev),

#20 lore raw

Re: [PATCH 3/3] hw/9pfs/xen: drain in-flight PDUs before xen-9p disconnect

Quoted 6 lines · On Saturday, 18 July 2026 02:46:36 CEST Stefano Stabellini wrote:
Quoted 3 lines · Christian Schoenebeck

Hi Stefano,

do you might have a chance to look at this Xen patch?

Sorry for the delay

Thanks for looking at this, much appreciated!

Quoted 43 lines
Quoted 43 lines · On Thursday, 9 July 2026 15:50:36 CEST Christian Schoenebeck wrote:
Quoted 42 lines

The xen-9p disconnect path has two issues:

1. It frees the Xen9pfsRing structures while in-flight PDUs may still

reference them via pdu->tag to index rings[]. This causes a UAF in xen_9pfs_push_and_notify() when worker threads resume after completing filesystem operations.

2. It never calls v9fs_device_unrealize_common(), which means server

state (struct LocalData, mountfd, FIDs) is never cleaned up on disconnect, causing a resource leak on every guest-initiated disconnect.

Fix both by draining in-flight PDUs via v9fs_reset() before tearing down rings, and calling v9fs_device_unrealize_common() to clean up server state.

Fixes b37eeb0201 ("xen/9pfs: introduce Xen 9pfs backend")

---

 hw/9pfs/xen-9p-backend.c | 4 ++++
 1 file changed, 4 insertions(+)
hw/9pfs/xen-9p-backend.c +1 -0
index 24c90d97ec..3b7a68779a 100644
@@ -368,10 +368,14 @@ static void xen_9pfs_evtchn_event(void *opaque)
 
 static void xen_9pfs_disconnect(struct XenLegacyDevice *xendev)
 {
 
     Xen9pfsDev *xen_9pdev = container_of(xendev, Xen9pfsDev, xendev);
 
+    V9fsState *s = &xen_9pdev->state;
 
     int i;
 
     trace_xen_9pfs_disconnect(xendev->name);
Quoted 7 lines
Quoted 2 lines

hw/9pfs/xen-9p-backend.c static void xen_9pfs_evtchn_event(void *opaque) · line 377

Quoted 2 lines
hw/9pfs/xen-9p-backend.c +2 -0
+    v9fs_reset(s);
+    v9fs_device_unrealize_common(s);

While the patch addresses a real issue, xen_9pfs_disconnect() is also called from the error paths of xen_9pfs_pdu_vmarshal/vunmarshal, which run inside the PDU's own coroutine. v9fs_reset() is not safe to be called there.

Well, xen_9pfs_pdu_vmarshal() / xen_9pfs_pdu_vunmarshal() are both called on main thread, and xen_be_disconnect() has a guard that prevents the disconnect handler being called twice (more about that later), however you are right that these two functions run within coroutines. Which then leads to the following problem in v9fs_reset():

    while (!QLIST_EMPTY(&s->active_list)) {
        aio_poll(qemu_get_aio_context(), true);
    }

This loop would then run forever, as the coroutine calling v9fs_reset() is on the active_list, hence this list would never turn empty and the loop condition never turn false.

Furthermore aio_poll() is marked as no_coroutine_fn:

include/qemu/aio.h:bool no_coroutine_fn aio_poll(AioContext *ctx, bool blocking);

So aio_poll() should actually not be called in a coroutine context at all.

Quoted 3 lines

We could remove the direct call to xen_9pfs_disconnect() from xen_9pfs_pdu_vmarshal/vunmarshal because they already change the state to XenbusStateClosing, which should result in the same

The problem with this approach is that xen_9pfs_pdu_vmarshal() / xen_9pfs_pdu_vunmarshal() would still call:

    xen_be_set_state(&xen_9pfs->xendev, XenbusStateClosing);

And the previously mentioned guard in xen_be_disconnect() for not calling the disconnect callback more than once, looks like this:

    if (xendev->be_state != XenbusStateClosing &&
        xendev->be_state != XenbusStateClosed  &&
        xendev->ops->disconnect) {
        xendev->ops->disconnect(xendev);
    }

Hence if this (v)unmarshal error path was triggered, the disconnect handler would then actually never be executed. And that's apparently the reason why these two functions do call xen_9pfs_disconnect() explicitly in their error path right now.

What I think might work though, was replacing the current direct xen_9pfs_disconnect() calls by deferred ones, something like:

    aio_bh_schedule_oneshot(... xen_9pfs_disconnect ...);

Because xen_9pfs_disconnect() would then run both on main thread and not run in a coroutine context. So it should turn it safe.

I'll prepare a v2 with the latter solution.

/Christian