Re: [RFC PATCH] RDMA/srp: Fix use-after-free in srp_exit_cmd_priv

[Jason Gunthorpe]

On Fri, Jun 24, 2022 at 12:02:53PM +0800, Li Zhijian wrote:
> srp_exit_cmd_priv() will try to access srp_device by Scsi_Host like below:
> 
>  Scsi_Host                srp_target_port     srp_host         srp_device
> +------------------+ +-- +--------------+  +>+----------+   +->+---------+
> |                  | |   |              |  | |          |   |  |         |
> |                  | |   |  *srp_host   +--+ | *srp_dev +---+  | *dev    |
> +-+hostdata--------+-+   |              |    |          |      |         |
> | | srp_target_port|     |              |    |          |      |         |
> | |                |     |              |    |          |      |         |
> | |                |     |              |    |          |      |         |
> +-+----------------+---- +--------------+    +----------+      +---------+
> 
> But sometims Scsi_Host still keeps the reference to srp_host that is
> possible released already. This could be happend if i frequently abort
> (Ctrl-c) the blktests during it was running and then cause below error:
> 
> [  952.299153] Freed by task 17289:
> [  952.299156]  kasan_save_stack+0x1e/0x40
> [  952.299160]  kasan_set_track+0x21/0x30
> [  952.299164]  kasan_set_free_info+0x20/0x30
> [  952.299169]  __kasan_slab_free+0x108/0x170
> [  952.299173]  kfree+0x9a/0x320
> [  952.299177]  srp_remove_one+0x114/0x180 [ib_srp]
> [  952.299189]  remove_client_context+0x8f/0xd0 [ib_core]
> [  952.299269]  disable_device+0xee/0x1e0 [ib_core]
> [  952.299348]  __ib_unregister_device+0x59/0xf0 [ib_core]
> [  952.299429]  ib_unregister_device_and_put+0x3b/0x50 [ib_core]
> [  952.299509]  nldev_dellink+0x126/0x1b0 [ib_core]
> [  952.299592]  rdma_nl_rcv_msg+0x1cc/0x310 [ib_core]
> [  952.299673]  rdma_nl_rcv+0x172/0x200 [ib_core]
> [  952.299760]  netlink_unicast+0x36b/0x4a0
> [  952.299770]  netlink_sendmsg+0x3a9/0x6d0
> [  952.299774]  sock_sendmsg+0x91/0xa0
> [  952.299783]  __sys_sendto+0x16f/0x210
> [  952.299788]  __x64_sys_sendto+0x6f/0x80
> [  952.299792]  do_syscall_64+0x3b/0x90
> [  952.299795]  entry_SYSCALL_64_after_hwframe+0x46/0xb0

I don't even understand how get_device() prevents this call chain??

It looks to me like the problem is srp_remove_one() is not waiting for
or canceling some outstanding work.

Jason