Hi Doug, > -----Original Message----- > From: Doug Ledford <[email protected]> > Sent: Thursday, August 23, 2018 11:56 AM > To: Parav Pandit <[email protected]>; Jason Gunthorpe <[email protected]>; > Eric Biggers <[email protected]> > Cc: [email protected]; [email protected]; > Leon Romanovsky <[email protected]>; [email protected]; > Mark Bloch <[email protected]>; Moni Shoua <[email protected]>; > [email protected]; syzbot > <[email protected]> > Subject: Re: [RDMA bug] KASAN: use-after-free Read in __list_del_entry_valid > (4) > > On Thu, 2018-08-23 at 16:39 +0000, Parav Pandit wrote: > > > -----Original Message----- > > > From: Jason Gunthorpe <[email protected]> > > > Sent: Thursday, August 23, 2018 9:55 AM > > > To: Eric Biggers <[email protected]> > > > Cc: Doug Ledford <[email protected]>; [email protected]; > > > [email protected]; Leon Romanovsky > > > <[email protected]>; [email protected]; Mark Bloch > > > <[email protected]>; Moni Shoua <[email protected]>; Parav > Pandit > > > <[email protected]>; [email protected]; syzbot > > > <[email protected]> > > > Subject: Re: [RDMA bug] KASAN: use-after-free Read in > > > __list_del_entry_valid > > > (4) > > > > > > On Wed, Aug 22, 2018 at 11:16:31PM -0700, Eric Biggers wrote: > > > > Hello RDMA / InfiniBand maintainers, > > > > > > > > This is an RDMA bug and it still occurs on Linus' tree as of today > > > > (commit 815f0ddb346c1960). > > > > > > > > I've also simplified the reproducer for it; see below after the original > report. > > > > Apparently it involves a race between RDMA_USER_CM_CMD_RESOLVE_IP > > > > > > and > > > > RDMA_USER_CM_CMD_LISTEN. > > > > > > That is an amazing reproducer! > > > > > > I have a feeling this is the same cause as all the other syzkaller bugs > > > in this > code: > > > lack of any sane locking at all :\ > > > > > > We've talked about chucking a big lock around this whole thing, but > > > nobody has done it yet.. It isn't so simple. > > > > > > > I had some code in which reduces three locks (handler_lock, qp_mutex, > id_lock) to single mutex to protect the cm_id and protects every exported > symbol of rdmacm which works on cm_id. > > But not ready enough to post it as patch yet. Lot of tests required before > > I get > there and some refactor too before that. > > Does it finally address the fact that the rdmacm code was written so that it > was > always synchronous but RoCE src gid (I think that's what it was, I'm typing > this > from long ago memory) lookup broke that assumption? > I am not sure. To me it is unlikely, because rdma_resolve_route() for InfiniBand is not synchronous either which needs to query the SA. But qp_mutex existed long before that which doesn't provide any performance improvements. ( by splitting as 3rd lock instead of id_lock and handler_lock) and so on.
> -- > Doug Ledford <[email protected]> > GPG KeyID: B826A3330E572FDD > Key fingerprint = AE6B 1BDA 122B 23B4 265B 1274 B826 A333 0E57 2FDD
