linux-kernel - Re: [syzbot] KASAN: use-after-free Read in addr

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Message-ID: <CACT4Y+aq+m8DkB+8Zvh0G5Ehmtd2kgapDr9eJEP0Mq5WLCOWMg@mail.gmail.com>
Date:   Mon, 20 Sep 2021 10:13:10 +0200
From:   Dmitry Vyukov <dvyukov@...gle.com>
To:     Jason Gunthorpe <jgg@...pe.ca>
Cc:     syzbot <syzbot+dc3dfba010d7671e05f5@...kaller.appspotmail.com>,
        dledford@...hat.com, leon@...nel.org, linux-kernel@...r.kernel.org,
        linux-rdma@...r.kernel.org, syzkaller-bugs@...glegroups.com,
        Aleksandr Nogikh <nogikh@...gle.com>
Subject: Re: [syzbot] KASAN: use-after-free Read in addr_handler (4)

On Thu, 16 Sept 2021 at 18:28, Jason Gunthorpe <jgg@...pe.ca> wrote:
>
> On Thu, Sep 16, 2021 at 04:45:27PM +0200, Dmitry Vyukov wrote:
>
> > Answering your question re what was running concurrently with what.
> > Each of the syscalls in these programs can run up to 2 times and
> > ultimately any of these calls can race with any. Potentially syzkaller
> > can predict values kernel will return (e.g. id's) before kernel
> > actually returned them. I guess this does not restrict search area for
> > the bug a lot...
>
> I have a reasonable theory now..
>
> Based on the ops you provided this FSM sequence is possible
>
> RDMA_USER_CM_CMD_RESOLVE_IP
>   RDMA_CM_IDLE -> RDMA_CM_ADDR_QUERY
>   does rdma_resolve_ip(addr_handler)
>
>                           addr_handler
>                             RDMA_CM_ADDR_QUERY -> RDMA_CM_ADDR_BOUND
>                             [.. handler still running ..]
>
> RDMA_USER_CM_CMD_RESOLVE_IP
>   RDMA_CM_ADDR_BOUND -> RDMA_CM_ADDR_QUERY
>   does rdma_resolve_ip(addr_handler)
>
> RDMA_DESTROY_ID
>   rdma_addr_cancel()
>
> Which, if it happens fast enough, could trigger a situation where the
> '&id_priv->id.route.addr.dev_addr' "handle" is in the req_list twice
> beacause the addr_handler work queue hasn't yet got to the point of
> deleting it from the req_list before the the 2nd one is added.
>
> The issue is rdma_addr_cancel() has to be called rdma_resolve_ip() can
> be called again.
>
> Skipping it will cause 'req_list' to have two items in the internal
> linked list with the same key and it will not cancel the newest one
> with the active timer. This would cause the use after free syndrome
> like this trace is showing.
>
> I can make a patch, but have no way to know if it is any good :\

Good detective work!

But if you have a theory of what happens, it's usually easy to write a
reproducer that aims at triggering this exact scenario. Isn't it the
case here? I would assume you need it to add as a test anyway.