]> git.proxmox.com Git - mirror_ubuntu-kernels.git/commit
net/smc: fix sk_refcnt underflow on linkdown and fallback
authorDust Li <dust.li@linux.alibaba.com>
Wed, 10 Nov 2021 07:02:34 +0000 (15:02 +0800)
committerDavid S. Miller <davem@davemloft.net>
Wed, 10 Nov 2021 14:43:59 +0000 (14:43 +0000)
commite5d5aadcf3cd59949316df49c27cb21788d7efe4
treedc705348ed5335458e23bdcb3a96146c8d9bb374
parentc7ebe23cee350fb187ee00ff445b01e11de0bfe9
net/smc: fix sk_refcnt underflow on linkdown and fallback

We got the following WARNING when running ab/nginx
test with RDMA link flapping (up-down-up).
The reason is when smc_sock fallback and at linkdown
happens simultaneously, we may got the following situation:

__smc_lgr_terminate()
 --> smc_conn_kill()
    --> smc_close_active_abort()
           smc_sock->sk_state = SMC_CLOSED
           sock_put(smc_sock)

smc_sock was set to SMC_CLOSED and sock_put() been called
when terminate the link group. But later application call
close() on the socket, then we got:

__smc_release():
    if (smc_sock->fallback)
        smc_sock->sk_state = SMC_CLOSED
        sock_put(smc_sock)

Again we set the smc_sock to CLOSED through it's already
in CLOSED state, and double put the refcnt, so the following
warning happens:

refcount_t: underflow; use-after-free.
WARNING: CPU: 5 PID: 860 at lib/refcount.c:28 refcount_warn_saturate+0x8d/0xf0
Modules linked in:
CPU: 5 PID: 860 Comm: nginx Not tainted 5.10.46+ #403
Hardware name: Alibaba Cloud Alibaba Cloud ECS, BIOS 8c24b4c 04/01/2014
RIP: 0010:refcount_warn_saturate+0x8d/0xf0
Code: 05 5c 1e b5 01 01 e8 52 25 bc ff 0f 0b c3 80 3d 4f 1e b5 01 00 75 ad 48

RSP: 0018:ffffc90000527e50 EFLAGS: 00010286
RAX: 0000000000000026 RBX: ffff8881300df2c0 RCX: 0000000000000027
RDX: 0000000000000000 RSI: ffff88813bd58040 RDI: ffff88813bd58048
RBP: 0000000000000000 R08: 0000000000000003 R09: 0000000000000001
R10: ffff8881300df2c0 R11: ffffc90000527c78 R12: ffff8881300df340
R13: ffff8881300df930 R14: ffff88810b3dad80 R15: ffff8881300df4f8
FS:  00007f739de8fb80(0000) GS:ffff88813bd40000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 000000000a01b008 CR3: 0000000111b64003 CR4: 00000000003706e0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
Call Trace:
 smc_release+0x353/0x3f0
 __sock_release+0x3d/0xb0
 sock_close+0x11/0x20
 __fput+0x93/0x230
 task_work_run+0x65/0xa0
 exit_to_user_mode_prepare+0xf9/0x100
 syscall_exit_to_user_mode+0x27/0x190
 entry_SYSCALL_64_after_hwframe+0x44/0xa9

This patch adds check in __smc_release() to make
sure we won't do an extra sock_put() and set the
socket to CLOSED when its already in CLOSED state.

Fixes: 51f1de79ad8e (net/smc: replace sock_put worker by socket refcounting)
Signed-off-by: Dust Li <dust.li@linux.alibaba.com>
Reviewed-by: Tony Lu <tonylu@linux.alibaba.com>
Signed-off-by: Dust Li <dust.li@linux.alibaba.com>
Acked-by: Karsten Graul <kgraul@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/smc/af_smc.c