Re: [PATCH] 9p: fix crash when transaction killed

From: Schspa Shi
Date: Tue Nov 29 2022 - 22:04:45 EST



asmadeus@xxxxxxxxxxxxx writes:

> Schspa Shi wrote on Wed, Nov 30, 2022 at 12:22:51AM +0800:
>> The transport layer of fs does not fully support the cancel request.
>> When the request is in the REQ_STATUS_SENT state, p9_fd_cancelled
>> will forcibly delete the request, and at this time p9_[read/write]_work
>> may continue to use the request. Therefore, it causes UAF .
>>
>> There is the logs from syzbot.
>>
>> Corrupted memory at 0xffff88807eade00b [ 0xff 0x07 0x00 0x00 0x00 0x00
>> 0x00 0x00 . . . . . . . . ] (in kfence-#110):
>> p9_fcall_fini net/9p/client.c:248 [inline]
>> p9_req_put net/9p/client.c:396 [inline]
>> p9_req_put+0x208/0x250 net/9p/client.c:390
>> p9_client_walk+0x247/0x540 net/9p/client.c:1165
>> clone_fid fs/9p/fid.h:21 [inline]
>> v9fs_fid_xattr_set+0xe4/0x2b0 fs/9p/xattr.c:118
>> v9fs_xattr_set fs/9p/xattr.c:100 [inline]
>> v9fs_xattr_handler_set+0x6f/0x120 fs/9p/xattr.c:159
>> __vfs_setxattr+0x119/0x180 fs/xattr.c:182
>> __vfs_setxattr_noperm+0x129/0x5f0 fs/xattr.c:216
>> __vfs_setxattr_locked+0x1d3/0x260 fs/xattr.c:277
>> vfs_setxattr+0x143/0x340 fs/xattr.c:309
>> setxattr+0x146/0x160 fs/xattr.c:617
>> path_setxattr+0x197/0x1c0 fs/xattr.c:636
>> __do_sys_setxattr fs/xattr.c:652 [inline]
>> __se_sys_setxattr fs/xattr.c:648 [inline]
>> __ia32_sys_setxattr+0xc0/0x160 fs/xattr.c:648
>> do_syscall_32_irqs_on arch/x86/entry/common.c:112 [inline]
>> __do_fast_syscall_32+0x65/0xf0 arch/x86/entry/common.c:178
>> do_fast_syscall_32+0x33/0x70 arch/x86/entry/common.c:203
>> entry_SYSENTER_compat_after_hwframe+0x70/0x82
>>
>> Below is a similar scenario, the scenario in the syzbot log looks more
>> complicated than this one, but the root cause seems to be the same.
>>
>> T21124 p9_write_work p9 read_work
>> ======================== first trans =================================
>> p9_client_walk
>> p9_client_rpc
>> p9_client_prepare_req
>> /* req->refcount == 2 */
>> c->trans_mod->request(c, req);
>> p9_fd_request
>> req move to unsent_req_list
>> req->status = REQ_STATUS_SENT;
>> req move to req_list
>> << send to server >>
>> wait_event_killable
>> << get kill signal >>
>> if (c->trans_mod->cancel(c, req))
>> p9_client_flush(c, req);
>> /* send flush request */
>> req = p9_client_rpc(c, P9_TFLUSH, "w", oldtag);
>> if (c->trans_mod->cancelled)
>> c->trans_mod->cancelled(c, oldreq);
>> /* old req was deleted from req_list */
>> /* req->refcount == 1 */
>> p9_req_put
>> /* req->refcount == 0 */
>> << preempted >>
>> << get response, UAF here >>
>> m->rreq = p9_tag_lookup(m->client, m->rc.tag);
>> /* req->refcount == 1 */
>> << do response >>
>> p9_client_cb(m->client, m->rreq, REQ_STATUS_RCVD);
>> /* req->refcount == 0 */
>> p9_fcall_fini
>> /* request have been freed */
>> p9_fcall_fini
>> /* double free */
>> p9_req_put(m->client, m->rreq);
>> /* req->refcount == 1 */
>>
>> To fix it, we can wait the request with status REQ_STATUS_SENT returned.
>
> Christian replied on this (we cannot wait) but I agree with him -- the

Yes, this is where I worry about too, this wait maybe cause a deadlock.

> scenario you describe is proteced by p9_tag_lookup checking for refcount
> with refcount_inc_not_zero (p9_req_try_get).

Thanks for pointing out the zero value check here, the scene in the
commit message does not hold.

>
> The normal scenarii for flush are as follow:
> - cancel before request is sent: no flush, just free
> - flush is ignored and reply comes first: we get reply from original
> request then reply from flush
> - flush is handled and reply never comes: we only get reply from flush
>
> Protocol-wise, we can safely reuse the tag after the flush reply got
> received; and as far as I can follow the code we only ever free the tag
> (last p9_call_fini) after flush has returned so the entry should be
> protected.
>
> If we receive a response on the given tag between cancelled and the main
> thread going out the request has been marked as FLSHD and should be
> ignored. . . here is one p9_req_put in p9_read_work() in this case but
> it corresponds to the ref obtained by p9_tag_lookup() so it should be
> valid.
>
>
> I'm happy to believe we have a race somewhere (even if no sane server
> would produce it), but right now I don't see it looking at the code.. :/

And I think there is a race too. because the syzbot report about 9p fs
memory corruption multi times.

As for the problem, the p9_tag_lookup only takes the rcu_read_lock when
accessing the IDR, why it doesn't take the p9_client->lock? Maybe the
root cause is that a lock is missing here.

--
BRs
Schspa Shi