[Cluster-devel] [PATCH] dlm: use sock_create_lite inside tcp_accept_from_sock

[Steven Whitehouse]

Hi,


On 07/08/17 07:31, Guoqing Jiang wrote:
> With commit 0ffdaf5b41cf ("net/sock: add WARN_ON(parent->sk)
> in sock_graft()"), a calltrace happened as follows:
>
> [  457.018340] WARNING: CPU: 0 PID: 15623 at ./include/net/sock.h:1703 inet_accept+0x135/0x140
> ...
> [  457.018381] RIP: 0010:inet_accept+0x135/0x140
> [  457.018381] RSP: 0018:ffffc90001727d18 EFLAGS: 00010286
> [  457.018383] RAX: 0000000000000001 RBX: ffff880012413000 RCX: 0000000000000001
> [  457.018384] RDX: 000000000000018a RSI: 00000000fffffe01 RDI: ffffffff8156fae8
> [  457.018384] RBP: ffffc90001727d38 R08: 0000000000000000 R09: 0000000000004305
> [  457.018385] R10: 0000000000000001 R11: 0000000000004304 R12: ffff880035ae7a00
> [  457.018386] R13: ffff88001282af10 R14: ffff880034e4e200 R15: 0000000000000000
> [  457.018387] FS:  0000000000000000(0000) GS:ffff88003fc00000(0000) knlGS:0000000000000000
> [  457.018388] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> [  457.018389] CR2: 00007fdec22f9000 CR3: 0000000002b5a000 CR4: 00000000000006f0
> [  457.018395] Call Trace:
> [  457.018402]  tcp_accept_from_sock.part.8+0x12d/0x449 [dlm]
> [  457.018405]  ? vprintk_emit+0x248/0x2d0
> [  457.018409]  tcp_accept_from_sock+0x3f/0x50 [dlm]
> [  457.018413]  process_recv_sockets+0x3b/0x50 [dlm]
> [  457.018415]  process_one_work+0x138/0x370
> [  457.018417]  worker_thread+0x4d/0x3b0
> [  457.018419]  kthread+0x109/0x140
> [  457.018421]  ? rescuer_thread+0x320/0x320
> [  457.018422]  ? kthread_park+0x60/0x60
> [  457.018424]  ret_from_fork+0x25/0x30
>
> Since newsocket created by sock_create_kern sets it's
> sock by the path:
>
> 	sock_create_kern -> __sock_creat
> 			 ->pf->create => inet_create
> 			 -> sock_init_data
>
> Then WARN_ON is triggered by "con->sock->ops->accept =>
> inet_accept -> sock_graft", it also means newsock->sk
> is leaked since sock_graft will replace it with a new
> sk.
>
> To resolve the issue, we need to use sock_create_lite
> instead of sock_create_kern, like commit 0933a578cd55
> ("rds: tcp: use sock_create_lite() to create the accept
> socket") did.
Good spotting!

Bob, is this the reason that you had so much trouble figuring out what 
was going on with the sk callbacks? You will need to review your patches 
for that I think, in case this makes a difference to them.

Acked-by: Steven Whitehouse <swhiteho at redhat.com>

Steve.

> Signed-off-by: Guoqing Jiang <gqjiang at suse.com>
> ---
>   fs/dlm/lowcomms.c | 2 +-
>   1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/fs/dlm/lowcomms.c b/fs/dlm/lowcomms.c
> index 9382db9..4813d0e 100644
> --- a/fs/dlm/lowcomms.c
> +++ b/fs/dlm/lowcomms.c
> @@ -729,7 +729,7 @@ static int tcp_accept_from_sock(struct connection *con)
>   	mutex_unlock(&connections_lock);
>   
>   	memset(&peeraddr, 0, sizeof(peeraddr));
> -	result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
> +	result = sock_create_lite(dlm_local_addr[0]->ss_family,
>   				  SOCK_STREAM, IPPROTO_TCP, &newsock);
>   	if (result < 0)
>   		return -ENOMEM;