This series contains fixes for locking bugs in the recently introduced
pds-vfio-pci driver. There was an initial bug reported by Dan Carpenter
at:
https://lore.kernel.org/kvm/[email protected]/
However, more locking bugs were found when looking into the previously
mentioned issue. So, all fixes are included in this series.
v2:
https://lore.kernel.org/kvm/[email protected]/
- Trim the OOPs in the patch commit messages
- Rework patch 3/3 to only unlock the spinlock once
- Destroy the state_mutex in the driver specific vfio_device_ops.release
callback
Brett Creeley (3):
pds/vfio: Fix spinlock bad magic BUG
pds/vfio: Fix mutex lock->magic != lock warning
pds/vfio: Fix possible sleep while in atomic context
drivers/vfio/pci/pds/vfio_dev.c | 27 ++++++++++++++++++++-------
1 file changed, 20 insertions(+), 7 deletions(-)
--
2.17.1
The following BUG was found when running on a kernel with
CONFIG_DEBUG_MUTEXES=y set:
DEBUG_LOCKS_WARN_ON(lock->magic != lock)
RIP: 0010:mutex_trylock+0x10d/0x120
Call Trace:
<TASK>
? __warn+0x85/0x140
? mutex_trylock+0x10d/0x120
? report_bug+0xfc/0x1e0
? handle_bug+0x3f/0x70
? exc_invalid_op+0x17/0x70
? asm_exc_invalid_op+0x1a/0x20
? mutex_trylock+0x10d/0x120
? mutex_trylock+0x10d/0x120
pds_vfio_reset+0x3a/0x60 [pds_vfio_pci]
pci_reset_function+0x4b/0x70
reset_store+0x5b/0xa0
kernfs_fop_write_iter+0x137/0x1d0
vfs_write+0x2de/0x410
ksys_write+0x5d/0xd0
do_syscall_64+0x3b/0x90
entry_SYSCALL_64_after_hwframe+0x6e/0xd8
As shown, lock->magic != lock. This is because
mutex_init(&pds_vfio->state_mutex) is called in the VFIO open path. So,
if a reset is initiated before the VFIO device is opened the mutex will
have never been initialized. Fix this by calling
mutex_init(&pds_vfio->state_mutex) in the VFIO init path.
Also, don't destroy the mutex on close because the device may
be re-opened, which would cause mutex to be uninitialized. Fix this by
implementing a driver specific vfio_device_ops.release callback that
destroys the mutex before calling vfio_pci_core_release_dev().
Signed-off-by: Brett Creeley <[email protected]>
Reviewed-by: Shannon Nelson <[email protected]>
---
drivers/vfio/pci/pds/vfio_dev.c | 15 ++++++++++++---
1 file changed, 12 insertions(+), 3 deletions(-)
diff --git a/drivers/vfio/pci/pds/vfio_dev.c b/drivers/vfio/pci/pds/vfio_dev.c
index c351f588fa13..306b1c25f016 100644
--- a/drivers/vfio/pci/pds/vfio_dev.c
+++ b/drivers/vfio/pci/pds/vfio_dev.c
@@ -155,6 +155,7 @@ static int pds_vfio_init_device(struct vfio_device *vdev)
pds_vfio->vf_id = vf_id;
+ mutex_init(&pds_vfio->state_mutex);
spin_lock_init(&pds_vfio->reset_lock);
vdev->migration_flags = VFIO_MIGRATION_STOP_COPY | VFIO_MIGRATION_P2P;
@@ -170,6 +171,16 @@ static int pds_vfio_init_device(struct vfio_device *vdev)
return 0;
}
+static void pds_vfio_release_device(struct vfio_device *vdev)
+{
+ struct pds_vfio_pci_device *pds_vfio =
+ container_of(vdev, struct pds_vfio_pci_device,
+ vfio_coredev.vdev);
+
+ mutex_destroy(&pds_vfio->state_mutex);
+ vfio_pci_core_release_dev(vdev);
+}
+
static int pds_vfio_open_device(struct vfio_device *vdev)
{
struct pds_vfio_pci_device *pds_vfio =
@@ -181,7 +192,6 @@ static int pds_vfio_open_device(struct vfio_device *vdev)
if (err)
return err;
- mutex_init(&pds_vfio->state_mutex);
pds_vfio->state = VFIO_DEVICE_STATE_RUNNING;
pds_vfio->deferred_reset_state = VFIO_DEVICE_STATE_RUNNING;
@@ -201,14 +211,13 @@ static void pds_vfio_close_device(struct vfio_device *vdev)
pds_vfio_put_save_file(pds_vfio);
pds_vfio_dirty_disable(pds_vfio, true);
mutex_unlock(&pds_vfio->state_mutex);
- mutex_destroy(&pds_vfio->state_mutex);
vfio_pci_core_close_device(vdev);
}
static const struct vfio_device_ops pds_vfio_ops = {
.name = "pds-vfio",
.init = pds_vfio_init_device,
- .release = vfio_pci_core_release_dev,
+ .release = pds_vfio_release_device,
.open_device = pds_vfio_open_device,
.close_device = pds_vfio_close_device,
.ioctl = vfio_pci_core_ioctl,
--
2.17.1
The driver could possibly sleep while in atomic context resulting
in the following call trace while CONFIG_DEBUG_ATOMIC_SLEEP=y is
set:
BUG: sleeping function called from invalid context at kernel/locking/mutex.c:283
in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 2817, name: bash
preempt_count: 1, expected: 0
RCU nest depth: 0, expected: 0
Call Trace:
<TASK>
dump_stack_lvl+0x36/0x50
__might_resched+0x123/0x170
mutex_lock+0x1e/0x50
pds_vfio_put_lm_file+0x1e/0xa0 [pds_vfio_pci]
pds_vfio_put_save_file+0x19/0x30 [pds_vfio_pci]
pds_vfio_state_mutex_unlock+0x2e/0x80 [pds_vfio_pci]
pci_reset_function+0x4b/0x70
reset_store+0x5b/0xa0
kernfs_fop_write_iter+0x137/0x1d0
vfs_write+0x2de/0x410
ksys_write+0x5d/0xd0
do_syscall_64+0x3b/0x90
entry_SYSCALL_64_after_hwframe+0x6e/0xd8
This can happen if pds_vfio_put_restore_file() and/or
pds_vfio_put_save_file() grab the mutex_lock(&lm_file->lock)
while the spin_lock(&pds_vfio->reset_lock) is held, which can
happen during while calling pds_vfio_state_mutex_unlock().
Fix this by using a local state variable to represent the current
vfio device state and update all the next state values. Then
release the spin_unlock(&pds_vfio->reset_lock) before checking if
the current vfio device state is VFIO_DEVICE_STATE_ERROR. Then
calling pds_vfio_put_restore_file() and pds_vfio_put_save_file() are
no longer being done while holding the reset_lock.
The only possible concerns are other threads that may call
pds_vfio_put_restore_file(), pds_vfio_put_save_file(), and/or
pds_vfio_dirty_disable(). However, those paths are already protected
by the state mutex_lock(), which is held in this context.
Reported-by: Dan Carpenter <[email protected]>
Closes: https://lore.kernel.org/kvm/[email protected]/
Signed-off-by: Brett Creeley <[email protected]>
Reviewed-by: Shannon Nelson <[email protected]>
---
drivers/vfio/pci/pds/vfio_dev.c | 10 ++++++----
1 file changed, 6 insertions(+), 4 deletions(-)
diff --git a/drivers/vfio/pci/pds/vfio_dev.c b/drivers/vfio/pci/pds/vfio_dev.c
index 306b1c25f016..cf7f639ba0ec 100644
--- a/drivers/vfio/pci/pds/vfio_dev.c
+++ b/drivers/vfio/pci/pds/vfio_dev.c
@@ -31,15 +31,17 @@ void pds_vfio_state_mutex_unlock(struct pds_vfio_pci_device *pds_vfio)
again:
spin_lock(&pds_vfio->reset_lock);
if (pds_vfio->deferred_reset) {
+ enum vfio_device_mig_state current_state = pds_vfio->state;
+
pds_vfio->deferred_reset = false;
- if (pds_vfio->state == VFIO_DEVICE_STATE_ERROR) {
+ pds_vfio->state = pds_vfio->deferred_reset_state;
+ pds_vfio->deferred_reset_state = VFIO_DEVICE_STATE_RUNNING;
+ spin_unlock(&pds_vfio->reset_lock);
+ if (current_state == VFIO_DEVICE_STATE_ERROR) {
pds_vfio_put_restore_file(pds_vfio);
pds_vfio_put_save_file(pds_vfio);
pds_vfio_dirty_disable(pds_vfio, false);
}
- pds_vfio->state = pds_vfio->deferred_reset_state;
- pds_vfio->deferred_reset_state = VFIO_DEVICE_STATE_RUNNING;
- spin_unlock(&pds_vfio->reset_lock);
goto again;
}
mutex_unlock(&pds_vfio->state_mutex);
--
2.17.1
> -----Original Message-----
> From: Brett Creeley [mailto:[email protected]]
> Sent: 12 October 2023 00:01
> To: [email protected]; [email protected]; Shameerali Kolothum Thodi
> <[email protected]>; [email protected];
> [email protected]; [email protected]
> Cc: [email protected]; [email protected];
> [email protected]; [email protected]
> Subject: [PATCH v2 vfio 2/3] pds/vfio: Fix mutex lock->magic != lock warning
>
> The following BUG was found when running on a kernel with
> CONFIG_DEBUG_MUTEXES=y set:
>
> DEBUG_LOCKS_WARN_ON(lock->magic != lock)
> RIP: 0010:mutex_trylock+0x10d/0x120
> Call Trace:
> <TASK>
> ? __warn+0x85/0x140
> ? mutex_trylock+0x10d/0x120
> ? report_bug+0xfc/0x1e0
> ? handle_bug+0x3f/0x70
> ? exc_invalid_op+0x17/0x70
> ? asm_exc_invalid_op+0x1a/0x20
> ? mutex_trylock+0x10d/0x120
> ? mutex_trylock+0x10d/0x120
> pds_vfio_reset+0x3a/0x60 [pds_vfio_pci]
> pci_reset_function+0x4b/0x70
> reset_store+0x5b/0xa0
> kernfs_fop_write_iter+0x137/0x1d0
> vfs_write+0x2de/0x410
> ksys_write+0x5d/0xd0
> do_syscall_64+0x3b/0x90
> entry_SYSCALL_64_after_hwframe+0x6e/0xd8
>
> As shown, lock->magic != lock. This is because
> mutex_init(&pds_vfio->state_mutex) is called in the VFIO open path. So,
> if a reset is initiated before the VFIO device is opened the mutex will
> have never been initialized. Fix this by calling
> mutex_init(&pds_vfio->state_mutex) in the VFIO init path.
>
> Also, don't destroy the mutex on close because the device may
> be re-opened, which would cause mutex to be uninitialized. Fix this by
> implementing a driver specific vfio_device_ops.release callback that
> destroys the mutex before calling vfio_pci_core_release_dev().
>
> Signed-off-by: Brett Creeley <[email protected]>
> Reviewed-by: Shannon Nelson <[email protected]>
Reviewed-by: Shameer Kolothum <[email protected]>
Looks like mutex destruction logic needs to be added to HiSilicon driver as well.
From a quick look mlx5 also doesn't destroy the state_mutex.
Thanks,
Shameer
> ---
> drivers/vfio/pci/pds/vfio_dev.c | 15 ++++++++++++---
> 1 file changed, 12 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/vfio/pci/pds/vfio_dev.c b/drivers/vfio/pci/pds/vfio_dev.c
> index c351f588fa13..306b1c25f016 100644
> --- a/drivers/vfio/pci/pds/vfio_dev.c
> +++ b/drivers/vfio/pci/pds/vfio_dev.c
> @@ -155,6 +155,7 @@ static int pds_vfio_init_device(struct vfio_device
> *vdev)
>
> pds_vfio->vf_id = vf_id;
>
> + mutex_init(&pds_vfio->state_mutex);
> spin_lock_init(&pds_vfio->reset_lock);
>
> vdev->migration_flags = VFIO_MIGRATION_STOP_COPY |
> VFIO_MIGRATION_P2P;
> @@ -170,6 +171,16 @@ static int pds_vfio_init_device(struct vfio_device
> *vdev)
> return 0;
> }
>
> +static void pds_vfio_release_device(struct vfio_device *vdev)
> +{
> + struct pds_vfio_pci_device *pds_vfio =
> + container_of(vdev, struct pds_vfio_pci_device,
> + vfio_coredev.vdev);
> +
> + mutex_destroy(&pds_vfio->state_mutex);
> + vfio_pci_core_release_dev(vdev);
> +}
> +
> static int pds_vfio_open_device(struct vfio_device *vdev)
> {
> struct pds_vfio_pci_device *pds_vfio =
> @@ -181,7 +192,6 @@ static int pds_vfio_open_device(struct vfio_device
> *vdev)
> if (err)
> return err;
>
> - mutex_init(&pds_vfio->state_mutex);
> pds_vfio->state = VFIO_DEVICE_STATE_RUNNING;
> pds_vfio->deferred_reset_state = VFIO_DEVICE_STATE_RUNNING;
>
> @@ -201,14 +211,13 @@ static void pds_vfio_close_device(struct
> vfio_device *vdev)
> pds_vfio_put_save_file(pds_vfio);
> pds_vfio_dirty_disable(pds_vfio, true);
> mutex_unlock(&pds_vfio->state_mutex);
> - mutex_destroy(&pds_vfio->state_mutex);
> vfio_pci_core_close_device(vdev);
> }
>
> static const struct vfio_device_ops pds_vfio_ops = {
> .name = "pds-vfio",
> .init = pds_vfio_init_device,
> - .release = vfio_pci_core_release_dev,
> + .release = pds_vfio_release_device,
> .open_device = pds_vfio_open_device,
> .close_device = pds_vfio_close_device,
> .ioctl = vfio_pci_core_ioctl,
> --
> 2.17.1
On Thu, Oct 12, 2023 at 09:00:09AM +0000, Shameerali Kolothum Thodi wrote:
> > Also, don't destroy the mutex on close because the device may
> > be re-opened, which would cause mutex to be uninitialized. Fix this by
> > implementing a driver specific vfio_device_ops.release callback that
> > destroys the mutex before calling vfio_pci_core_release_dev().
> >
> > Signed-off-by: Brett Creeley <[email protected]>
> > Reviewed-by: Shannon Nelson <[email protected]>
>
> Reviewed-by: Shameer Kolothum <[email protected]>
>
> Looks like mutex destruction logic needs to be added to HiSilicon driver as well.
> From a quick look mlx5 also doesn't destroy the state_mutex.
FWIW, mutex_destroy is a debugging feature, it is nice if you can do it, but
not a bug if it is missing..
Jason
On Wed, 11 Oct 2023 16:01:12 -0700
Brett Creeley <[email protected]> wrote:
> This series contains fixes for locking bugs in the recently introduced
> pds-vfio-pci driver. There was an initial bug reported by Dan Carpenter
> at:
>
> https://lore.kernel.org/kvm/[email protected]/
>
> However, more locking bugs were found when looking into the previously
> mentioned issue. So, all fixes are included in this series.
>
> v2:
> https://lore.kernel.org/kvm/[email protected]/
> - Trim the OOPs in the patch commit messages
> - Rework patch 3/3 to only unlock the spinlock once
I thought we determined the spinlock, and thus the atomic context, was
an arbitrary choice. I would have figured we simply convert it to a
mutex. Why didn't we take that route? Thanks,
Alex
> - Destroy the state_mutex in the driver specific vfio_device_ops.release
> callback
>
> Brett Creeley (3):
> pds/vfio: Fix spinlock bad magic BUG
> pds/vfio: Fix mutex lock->magic != lock warning
> pds/vfio: Fix possible sleep while in atomic context
>
> drivers/vfio/pci/pds/vfio_dev.c | 27 ++++++++++++++++++++-------
> 1 file changed, 20 insertions(+), 7 deletions(-)
>
On 10/13/2023 2:09 PM, Alex Williamson wrote:
> Caution: This message originated from an External Source. Use proper caution when opening attachments, clicking links, or responding.
>
>
> On Wed, 11 Oct 2023 16:01:12 -0700
> Brett Creeley <[email protected]> wrote:
>
>> This series contains fixes for locking bugs in the recently introduced
>> pds-vfio-pci driver. There was an initial bug reported by Dan Carpenter
>> at:
>>
>> https://lore.kernel.org/kvm/[email protected]/
>>
>> However, more locking bugs were found when looking into the previously
>> mentioned issue. So, all fixes are included in this series.
>>
>> v2:
>> https://lore.kernel.org/kvm/[email protected]/
>> - Trim the OOPs in the patch commit messages
>> - Rework patch 3/3 to only unlock the spinlock once
>
> I thought we determined the spinlock, and thus the atomic context, was
> an arbitrary choice. I would have figured we simply convert it to a
> mutex. Why didn't we take that route? Thanks,
>
> Alex
Hmm. I guess it wasn't completely clear that was the expected solution
even after Jason's response. However, I can resubmit a v3 with that change.
Thanks,
Brett
>
>> - Destroy the state_mutex in the driver specific vfio_device_ops.release
>> callback
>>
>> Brett Creeley (3):
>> pds/vfio: Fix spinlock bad magic BUG
>> pds/vfio: Fix mutex lock->magic != lock warning
>> pds/vfio: Fix possible sleep while in atomic context
>>
>> drivers/vfio/pci/pds/vfio_dev.c | 27 ++++++++++++++++++++-------
>> 1 file changed, 20 insertions(+), 7 deletions(-)
>>
>