Re: [PATCH v2] 9p/fd: set req refcount to zero to avoid uninitialized usage

From: asmadeus
Date: Wed Nov 30 2022 - 08:31:42 EST


Schspa Shi wrote on Wed, Nov 30, 2022 at 09:08:31PM +0800:
> When the transport layer of fs cancels the request, it is deleted from the
> client side. But the server can send a response with the freed tag.
>
> When the new request allocated, we add it to idr, and use the id form idr
> as tag, which will have the same tag with high probability. Then initialize
> the refcount after adding it to idr.

ultimately this bug has nothing to do with tag reuse -- we don't
actually need flush at all to trigger it.

- thread1 starts new request; idr initialized with tag X
- thread2 receives something for tag X, increments refcount before
refcount init
- thread1 resets refcount to two incorrectly

This could happen on any new message where the server voluntarily sends
a reply with tag X before the request has been sent; in a cyclic model
as suggested in the other thread it would be easy to guess just last+1
for an hypothetical attacker.

This scenario with flush is just how syzbot happened to trigger it, but
I think it's just superfluous to this commit message.

A few more nitpicks on wording below; happy to adjust things myself as I
apply patches but might as well comment first...

> If the p9_read_work got a response before the refcount initiated. It will
> use a uninitialized req, which will result in a bad request data struct.
>
> There is the logs from syzbot.

English: Here is ...

> 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 this patch seems can fix it.

English: seems to fix it?
(thanks for checking!)

>
> T21124 p9_read_work
> ======================== second trans =================================
> p9_client_walk
> p9_client_rpc
> p9_client_prepare_req
> p9_tag_alloc
> req = kmem_cache_alloc(p9_req_cache, GFP_NOFS);
> tag = idr_alloc
> << preempted >>
> req->tc.tag = tag;
> /* req->[refcount/tag] == uninitilzed */

typo: uninitialized

> m->rreq = p9_tag_lookup(m->client, m->rc.tag);

it's not obvious for someone reading this not familiar with 9p that
lookup will increment refcount

>
> refcount_set(&req->refcount, 2);
> << do response/error >>
> p9_req_put(m->client, m->rreq);
> /* req->refcount == 1 */
>
> /* req->refcount == 1 */
> << got a bad refcount >>

it's not obvious from this going back to thread 1 with a refcount of 1
would be a bad refcount, either.
One possible scenario would be:

/* increments uninitalized refcount */
req = p9_tag_lookup(tag)
refcount_set(req->refcount, 2)
/* cb drops one ref */
p9_client_cb(req)
/* reader thread drops its ref:
request is incorrectly freed */
p9_req_put(req)
/* use after free and ref underflow */
p9_req_put(req)



> To fix it, we can initize the refcount to zero before add to idr.
>
> Reported-by: syzbot+8f1060e2aaf8ca55220b@xxxxxxxxxxxxxxxxxxxxxxxxx
>

There should be no empty line between the tags; tags are part of the
"trailer" and some tools handle it as such (like git interpret-trailers),
which would ignore that Reported-by as it is not part of the last block
of text.

> Signed-off-by: Schspa Shi <schspa@xxxxxxxxx>
> ---
> net/9p/client.c | 4 ++++
> 1 file changed, 4 insertions(+)
>
> diff --git a/net/9p/client.c b/net/9p/client.c
> index aaa37b07e30a..a72cb597a8ab 100644
> --- a/net/9p/client.c
> +++ b/net/9p/client.c
> @@ -297,6 +297,10 @@ p9_tag_alloc(struct p9_client *c, int8_t type, uint t_size, uint r_size,
> p9pdu_reset(&req->rc);
> req->t_err = 0;
> req->status = REQ_STATUS_ALLOC;
> + /* p9_tag_lookup relies on this refcount to be zero to avoid
> + * getting a freed request.

A freed request would have 0 by definition, if it isn't zero then this
is a newly allocated uninit request, so this comment is incorrect.

How about:
/* refcount needs to be set to 0 before inserting into the idr
* so p9_tag_lookup does not accept a request that is not fully
* initialized. refcount_set to 2 below will mark request live.
*/

> + */
> + refcount_set(&req->refcount, 0);
> init_waitqueue_head(&req->wq);
> INIT_LIST_HEAD(&req->req_list);
>