Re: [syzbot] [rdma?] possible deadlock in rds_trans_get_preferred

[Eduard Zateev]

Hi Hillf,

I tested your asynchronous CMA ID destruction patch from the #syz test
reply on the current upstream checkout at 21a7f44fd076 (Linux 7.3.0).

The kernel was built with KASAN, lockdep, RDMA_RXE and RDS enabled. I ran
the syzbot C reproducer in QEMU with four vCPUs. On the unpatched kernel it
reproduced the reported circular locking dependency between rds_trans_sem
and the RDMA CMA handler_mutex. With your patch applied, the same
reproducer completed successfully, established both active and passive RDS
connections, and produced no possible circular locking, possible recursive
locking, WARNING, Oops, or BUG output.

The build completed successfully and checkpatch.pl --strict reported no
errors or warnings. The complete test logs and the applied diff are
available on request.

Tested-by: Eduard Zateev <hackerowskiy@xxxxxxxxx>