[PATCH] Drivers: hv: vmbus: Fix kernel crash upon unbinding a device from uio_hv_generic driver

The following crash happens when a never-used device is unbound from
uio_hv_generic driver:

 kernel BUG at mm/slub.c:321!
 invalid opcode: 0000 [#1] SMP PTI
 CPU: 0 PID: 4001 Comm: bash Kdump: loaded Tainted: G               X --------- ---  5.14.0-0.rc2.23.el9.x86_64 #1
 Hardware name: Microsoft Corporation Virtual Machine/Virtual Machine, BIOS 090008  12/07/2018
 RIP: 0010:__slab_free+0x1d5/0x3d0
...
 Call Trace:
  ? pick_next_task_fair+0x18e/0x3b0
  ? __cond_resched+0x16/0x40
  ? vunmap_pmd_range.isra.0+0x154/0x1c0
  ? __vunmap+0x22d/0x290
  ? hv_ringbuffer_cleanup+0x36/0x40 [hv_vmbus]
  kfree+0x331/0x380
  ? hv_uio_remove+0x43/0x60 [uio_hv_generic]
  hv_ringbuffer_cleanup+0x36/0x40 [hv_vmbus]
  vmbus_free_ring+0x21/0x60 [hv_vmbus]
  hv_uio_remove+0x4f/0x60 [uio_hv_generic]
  vmbus_remove+0x23/0x30 [hv_vmbus]
  __device_release_driver+0x17a/0x230
  device_driver_detach+0x3c/0xa0
  unbind_store+0x113/0x130
...

The problem appears to be that we free 'ring_info->pkt_buffer' twice:
first, when the device is unbound from in-kernel driver (netvsc in this
case) and second from hv_uio_remove(). Normally, ring buffer is supposed
to be re-initialized from hv_uio_open() but this happens when UIO device
is being opened and this is not guaranteed to happen.

Generally, it is OK to call hv_ringbuffer_cleanup() twice for the same
channel (which is being handed over between in-kernel drivers and UIO) even
if we didn't call hv_ringbuffer_init() in between. We, however, need to
avoid kfree() call for an already freed pointer.

Fixes: adae1e931acd ("Drivers: hv: vmbus: Copy packets sent by Hyper-V out of the ring buffer")
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
---
 drivers/hv/ring_buffer.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/drivers/hv/ring_buffer.c b/drivers/hv/ring_buffer.c
index 2aee356840a2..314015d9e912 100644
--- a/drivers/hv/ring_buffer.c
+++ b/drivers/hv/ring_buffer.c

@@ -245,6 +245,7 @@ void hv_ringbuffer_cleanup(struct hv_ring_buffer_info *ring_info)
 	mutex_unlock(&ring_info->ring_buffer_mutex);
 
 	kfree(ring_info->pkt_buffer);
+	ring_info->pkt_buffer = NULL;
 	ring_info->pkt_buffer_size = 0;
 }
 

-- 
2.31.1

On Tue, Aug 31, 2021 at 04:39:16PM +0200, Vitaly Kuznetsov wrote:

This certainly won't hurt.

I would however like to wait till Andrea and Michael go over the
reasoning of this patch before doing anything.

Wei.

 	ring_info->pkt_buffer_size = 0;
 }
 
-- 
2.31.1

Wei Liu [off-list ref] writes:

On Tue, Aug 31, 2021 at 04:39:16PM +0200, Vitaly Kuznetsov wrote:

quoted

The following crash happens when a never-used device is unbound from
uio_hv_generic driver:

 kernel BUG at mm/slub.c:321!
 invalid opcode: 0000 [#1] SMP PTI
 CPU: 0 PID: 4001 Comm: bash Kdump: loaded Tainted: G               X --------- ---  5.14.0-0.rc2.23.el9.x86_64 #1
 Hardware name: Microsoft Corporation Virtual Machine/Virtual Machine, BIOS 090008  12/07/2018
 RIP: 0010:__slab_free+0x1d5/0x3d0
...
 Call Trace:
  ? pick_next_task_fair+0x18e/0x3b0
  ? __cond_resched+0x16/0x40
  ? vunmap_pmd_range.isra.0+0x154/0x1c0
  ? __vunmap+0x22d/0x290
  ? hv_ringbuffer_cleanup+0x36/0x40 [hv_vmbus]
  kfree+0x331/0x380
  ? hv_uio_remove+0x43/0x60 [uio_hv_generic]
  hv_ringbuffer_cleanup+0x36/0x40 [hv_vmbus]
  vmbus_free_ring+0x21/0x60 [hv_vmbus]
  hv_uio_remove+0x4f/0x60 [uio_hv_generic]
  vmbus_remove+0x23/0x30 [hv_vmbus]
  __device_release_driver+0x17a/0x230
  device_driver_detach+0x3c/0xa0
  unbind_store+0x113/0x130
...

The problem appears to be that we free 'ring_info->pkt_buffer' twice:
first, when the device is unbound from in-kernel driver (netvsc in this
case) and second from hv_uio_remove(). Normally, ring buffer is supposed
to be re-initialized from hv_uio_open() but this happens when UIO device
is being opened and this is not guaranteed to happen.

Generally, it is OK to call hv_ringbuffer_cleanup() twice for the same
channel (which is being handed over between in-kernel drivers and UIO) even
if we didn't call hv_ringbuffer_init() in between. We, however, need to
avoid kfree() call for an already freed pointer.

Fixes: adae1e931acd ("Drivers: hv: vmbus: Copy packets sent by Hyper-V out of the ring buffer")
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
---
 drivers/hv/ring_buffer.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/drivers/hv/ring_buffer.c b/drivers/hv/ring_buffer.c
index 2aee356840a2..314015d9e912 100644
--- a/drivers/hv/ring_buffer.c
+++ b/drivers/hv/ring_buffer.c

@@ -245,6 +245,7 @@ void hv_ringbuffer_cleanup(struct hv_ring_buffer_info *ring_info)
 	mutex_unlock(&ring_info->ring_buffer_mutex);
 
 	kfree(ring_info->pkt_buffer);
+	ring_info->pkt_buffer = NULL;

This certainly won't hurt.

I would however like to wait till Andrea and Michael go over the
reasoning of this patch before doing anything.

Thanks,

the counter-intuitive thing here is that the channel sometimes continues
to live after hv_ringbuffer_cleanup(): when we unbind a device from
in-kernel driver and give it to uio_hv_generic the channel is handed
over from one driver to another.

Wei.

quoted

 	ring_info->pkt_buffer_size = 0;
 }
 
-- 
2.31.1

-- 
Vitaly

On Tue, Aug 31, 2021 at 04:39:16PM +0200, Vitaly Kuznetsov wrote:

The following crash happens when a never-used device is unbound from
uio_hv_generic driver:

 kernel BUG at mm/slub.c:321!
 invalid opcode: 0000 [#1] SMP PTI
 CPU: 0 PID: 4001 Comm: bash Kdump: loaded Tainted: G               X --------- ---  5.14.0-0.rc2.23.el9.x86_64 #1
 Hardware name: Microsoft Corporation Virtual Machine/Virtual Machine, BIOS 090008  12/07/2018
 RIP: 0010:__slab_free+0x1d5/0x3d0
...
 Call Trace:
  ? pick_next_task_fair+0x18e/0x3b0
  ? __cond_resched+0x16/0x40
  ? vunmap_pmd_range.isra.0+0x154/0x1c0
  ? __vunmap+0x22d/0x290
  ? hv_ringbuffer_cleanup+0x36/0x40 [hv_vmbus]
  kfree+0x331/0x380
  ? hv_uio_remove+0x43/0x60 [uio_hv_generic]
  hv_ringbuffer_cleanup+0x36/0x40 [hv_vmbus]
  vmbus_free_ring+0x21/0x60 [hv_vmbus]
  hv_uio_remove+0x4f/0x60 [uio_hv_generic]
  vmbus_remove+0x23/0x30 [hv_vmbus]
  __device_release_driver+0x17a/0x230
  device_driver_detach+0x3c/0xa0
  unbind_store+0x113/0x130
...

The problem appears to be that we free 'ring_info->pkt_buffer' twice:
first, when the device is unbound from in-kernel driver (netvsc in this
case) and second from hv_uio_remove(). Normally, ring buffer is supposed
to be re-initialized from hv_uio_open() but this happens when UIO device
is being opened and this is not guaranteed to happen.

Generally, it is OK to call hv_ringbuffer_cleanup() twice for the same
channel (which is being handed over between in-kernel drivers and UIO) even
if we didn't call hv_ringbuffer_init() in between. We, however, need to
avoid kfree() call for an already freed pointer.

Fixes: adae1e931acd ("Drivers: hv: vmbus: Copy packets sent by Hyper-V out of the ring buffer")
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>

LGTM.

Reviewed-by: Andrea Parri <parri.andrea@gmail.com>

ae6935ed7d424f appears to have anticipated this problem on ->ring_buffer
and adopted the solution proposed here by Vitaly.

  Andrea

From: Vitaly Kuznetsov <vkuznets@redhat.com> Sent: Tuesday, August 31, 2021 7:39 AM

Reviewed-by: Michael Kelley <redacted>

On Tue, Aug 31, 2021 at 04:39:16PM +0200, Vitaly Kuznetsov wrote:

The following crash happens when a never-used device is unbound from
uio_hv_generic driver:

[...]

Fixes: adae1e931acd ("Drivers: hv: vmbus: Copy packets sent by Hyper-V out of the ring buffer")
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>

Applied to hyperv-fixes. Thanks.