Dragos Tatulea
2023-Jul-27 16:02 UTC
[PATCH] vdpa/mlx5: Fix crash on shutdown for when no ndev exists
On Wed, 2023-07-26 at 15:26 -0400, Michael S. Tsirkin wrote:> On Wed, Jul 26, 2023 at 10:07:38PM +0300, Dragos Tatulea wrote: > > The ndev was accessed on shutdown without a check if it actually exists. > > This triggered the crash pasted below. This patch simply adds a check > > before using ndev. > > > > ?BUG: kernel NULL pointer dereference, address: 0000000000000300 > > ?#PF: supervisor read access in kernel mode > > ?#PF: error_code(0x0000) - not-present page > > ?PGD 0 P4D 0 > > ?Oops: 0000 [#1] SMP > > ?CPU: 0 PID: 1 Comm: systemd-shutdow Not tainted 6.5.0- > > rc2_for_upstream_min_debug_2023_07_17_15_05 #1 > > ?Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.13.0-0- > > gf21b5a4aeb02-prebuilt.qemu.org 04/01/2014 > > ?RIP: 0010:mlx5v_shutdown+0xe/0x50 [mlx5_vdpa] > > ?RSP: 0018:ffff8881003bfdc0 EFLAGS: 00010286 > > ?RAX: ffff888103befba0 RBX: ffff888109d28008 RCX: 0000000000000017 > > ?RDX: 0000000000000001 RSI: 0000000000000212 RDI: ffff888109d28000 > > ?RBP: 0000000000000000 R08: 0000000d3a3a3882 R09: 0000000000000001 > > ?R10: 0000000000000000 R11: 0000000000000000 R12: ffff888109d28000 > > ?R13: ffff888109d28080 R14: 00000000fee1dead R15: 0000000000000000 > > ?FS:? 00007f4969e0be40(0000) GS:ffff88852c800000(0000) > > knlGS:0000000000000000 > > ?CS:? 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > > ?CR2: 0000000000000300 CR3: 00000001051cd006 CR4: 0000000000370eb0 > > ?DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 > > ?DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 > > ?Call Trace: > > ? <TASK> > > ? ? __die+0x20/0x60 > > ? ? page_fault_oops+0x14c/0x3c0 > > ? ? exc_page_fault+0x75/0x140 > > ? ? asm_exc_page_fault+0x22/0x30 > > ? ? mlx5v_shutdown+0xe/0x50 [mlx5_vdpa] > > ? device_shutdown+0x13e/0x1e0 > > ? kernel_restart+0x36/0x90 > > ? __do_sys_reboot+0x141/0x210 > > ? ? vfs_writev+0xcd/0x140 > > ? ? handle_mm_fault+0x161/0x260 > > ? ? do_writev+0x6b/0x110 > > ? do_syscall_64+0x3d/0x90 > > ? entry_SYSCALL_64_after_hwframe+0x46/0xb0 > > ?RIP: 0033:0x7f496990fb56 > > ?RSP: 002b:00007fffc7bdde88 EFLAGS: 00000206 ORIG_RAX: 00000000000000a9 > > ?RAX: ffffffffffffffda RBX: 0000000000000000 RCX: 00007f496990fb56 > > ?RDX: 0000000001234567 RSI: 0000000028121969 RDI: fffffffffee1dead > > ?RBP: 00007fffc7bde1d0 R08: 0000000000000000 R09: 0000000000000000 > > ?R10: 0000000000000000 R11: 0000000000000206 R12: 0000000000000000 > > ?R13: 00007fffc7bddf10 R14: 0000000000000000 R15: 00007fffc7bde2b8 > > ? </TASK> > > ?CR2: 0000000000000300 > > ?---[ end trace 0000000000000000 ]--- > > > > Fixes: bc9a2b3e686e ("vdpa/mlx5: Support interrupt bypassing") > > Signed-off-by: Dragos Tatulea <dtatulea at nvidia.com> > > --- > > ?drivers/vdpa/mlx5/net/mlx5_vnet.c | 3 ++- > > ?1 file changed, 2 insertions(+), 1 deletion(-) > > > > diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c > > b/drivers/vdpa/mlx5/net/mlx5_vnet.c > > index 9138ef2fb2c8..e2e7ebd71798 100644 > > --- a/drivers/vdpa/mlx5/net/mlx5_vnet.c > > +++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c > > @@ -3556,7 +3556,8 @@ static void mlx5v_shutdown(struct auxiliary_device > > *auxdev) > > ????????mgtdev = auxiliary_get_drvdata(auxdev); > > ????????ndev = mgtdev->ndev; > > ? > > -???????free_irqs(ndev); > > +???????if (ndev) > > +???????????????free_irqs(ndev); > > ?} > > ? > > something I don't get: > irqs are allocated in mlx5_vdpa_dev_add > why are they not freed in mlx5_vdpa_dev_del? >That is a good point. I will try to find out. I also don't get why free_irq is called in the vdpa dev .free op instead of mlx5_vdpa_dev_del. Maybe I can change that in a different refactoring.> this is what's creating all this mess. > >Not quite: mlx5_vdpa_dev_del (which is a .dev_del of for struct vdpa_mgmtdev_ops) doesn't get called on shutdown. At least that's what I see. Or am I missing something?> > ?static const struct auxiliary_device_id mlx5v_id_table[] = { > > -- > > 2.41.0 >
Michael S. Tsirkin
2023-Jul-27 16:28 UTC
[PATCH] vdpa/mlx5: Fix crash on shutdown for when no ndev exists
On Thu, Jul 27, 2023 at 04:02:16PM +0000, Dragos Tatulea wrote:> On Wed, 2023-07-26 at 15:26 -0400, Michael S. Tsirkin wrote: > > On Wed, Jul 26, 2023 at 10:07:38PM +0300, Dragos Tatulea wrote: > > > The ndev was accessed on shutdown without a check if it actually exists. > > > This triggered the crash pasted below. This patch simply adds a check > > > before using ndev. > > > > > > ?BUG: kernel NULL pointer dereference, address: 0000000000000300 > > > ?#PF: supervisor read access in kernel mode > > > ?#PF: error_code(0x0000) - not-present page > > > ?PGD 0 P4D 0 > > > ?Oops: 0000 [#1] SMP > > > ?CPU: 0 PID: 1 Comm: systemd-shutdow Not tainted 6.5.0- > > > rc2_for_upstream_min_debug_2023_07_17_15_05 #1 > > > ?Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.13.0-0- > > > gf21b5a4aeb02-prebuilt.qemu.org 04/01/2014 > > > ?RIP: 0010:mlx5v_shutdown+0xe/0x50 [mlx5_vdpa] > > > ?RSP: 0018:ffff8881003bfdc0 EFLAGS: 00010286 > > > ?RAX: ffff888103befba0 RBX: ffff888109d28008 RCX: 0000000000000017 > > > ?RDX: 0000000000000001 RSI: 0000000000000212 RDI: ffff888109d28000 > > > ?RBP: 0000000000000000 R08: 0000000d3a3a3882 R09: 0000000000000001 > > > ?R10: 0000000000000000 R11: 0000000000000000 R12: ffff888109d28000 > > > ?R13: ffff888109d28080 R14: 00000000fee1dead R15: 0000000000000000 > > > ?FS:? 00007f4969e0be40(0000) GS:ffff88852c800000(0000) > > > knlGS:0000000000000000 > > > ?CS:? 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > > > ?CR2: 0000000000000300 CR3: 00000001051cd006 CR4: 0000000000370eb0 > > > ?DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 > > > ?DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 > > > ?Call Trace: > > > ? <TASK> > > > ? ? __die+0x20/0x60 > > > ? ? page_fault_oops+0x14c/0x3c0 > > > ? ? exc_page_fault+0x75/0x140 > > > ? ? asm_exc_page_fault+0x22/0x30 > > > ? ? mlx5v_shutdown+0xe/0x50 [mlx5_vdpa] > > > ? device_shutdown+0x13e/0x1e0 > > > ? kernel_restart+0x36/0x90 > > > ? __do_sys_reboot+0x141/0x210 > > > ? ? vfs_writev+0xcd/0x140 > > > ? ? handle_mm_fault+0x161/0x260 > > > ? ? do_writev+0x6b/0x110 > > > ? do_syscall_64+0x3d/0x90 > > > ? entry_SYSCALL_64_after_hwframe+0x46/0xb0 > > > ?RIP: 0033:0x7f496990fb56 > > > ?RSP: 002b:00007fffc7bdde88 EFLAGS: 00000206 ORIG_RAX: 00000000000000a9 > > > ?RAX: ffffffffffffffda RBX: 0000000000000000 RCX: 00007f496990fb56 > > > ?RDX: 0000000001234567 RSI: 0000000028121969 RDI: fffffffffee1dead > > > ?RBP: 00007fffc7bde1d0 R08: 0000000000000000 R09: 0000000000000000 > > > ?R10: 0000000000000000 R11: 0000000000000206 R12: 0000000000000000 > > > ?R13: 00007fffc7bddf10 R14: 0000000000000000 R15: 00007fffc7bde2b8 > > > ? </TASK> > > > ?CR2: 0000000000000300 > > > ?---[ end trace 0000000000000000 ]--- > > > > > > Fixes: bc9a2b3e686e ("vdpa/mlx5: Support interrupt bypassing") > > > Signed-off-by: Dragos Tatulea <dtatulea at nvidia.com> > > > --- > > > ?drivers/vdpa/mlx5/net/mlx5_vnet.c | 3 ++- > > > ?1 file changed, 2 insertions(+), 1 deletion(-) > > > > > > diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c > > > b/drivers/vdpa/mlx5/net/mlx5_vnet.c > > > index 9138ef2fb2c8..e2e7ebd71798 100644 > > > --- a/drivers/vdpa/mlx5/net/mlx5_vnet.c > > > +++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c > > > @@ -3556,7 +3556,8 @@ static void mlx5v_shutdown(struct auxiliary_device > > > *auxdev) > > > ????????mgtdev = auxiliary_get_drvdata(auxdev); > > > ????????ndev = mgtdev->ndev; > > > ? > > > -???????free_irqs(ndev); > > > +???????if (ndev) > > > +???????????????free_irqs(ndev); > > > ?} > > > ? > > > > something I don't get: > > irqs are allocated in mlx5_vdpa_dev_add > > why are they not freed in mlx5_vdpa_dev_del? > > > That is a good point. I will try to find out. I also don't get why free_irq is > called in the vdpa dev .free op instead of mlx5_vdpa_dev_del. Maybe I can change > that in a different refactoring.as it is I have no idea whether e.g. ndev can change between these two call sites. that would make the check pointless.> > this is what's creating all this mess. > > > > > Not quite: mlx5_vdpa_dev_del (which is a .dev_del of for struct > vdpa_mgmtdev_ops) doesn't get called on shutdown. At least that's what I see. Or > am I missing something?and why do we care whether irqs are freed on shutdown?> > > ?static const struct auxiliary_device_id mlx5v_id_table[] = { > > > -- > > > 2.41.0 > > >
Reasonably Related Threads
- [PATCH] vdpa/mlx5: Fix crash on shutdown for when no ndev exists
- [PATCH] vdpa/mlx5: Fix crash on shutdown for when no ndev exists
- [PATCH] vdpa/mlx5: Fix crash on shutdown for when no ndev exists
- [PATCH] vdpa/mlx5: Fix crash on shutdown for when no ndev exists
- [PATCH] vdpa/mlx5: Fix crash on shutdown for when no ndev exists