[dm-devel] [dm-level] master - libmultipath: fix use after free when iscsi logs in

[Martin Wilck]

Hello Lixiaokeng,

Thanks for the report and the analysis.

On Mon, 2020-07-13 at 10:22 +0800, lixiaokeng wrote:
> When two iscsi ips log in and out alternately and  the following
> scripts run
> at the same time,
> 
> #!/bin/bash
> interval=5
> while true
> do
>         iscsiadm -m node -p 9.41.147.171 &> /dev/null
>         iscsiadm -m node -p 9.41.148.172 &> /dev/null
>         iscsiadm -m session &> /dev/null
>         rescan-scsi-bus.sh &> /dev/null
>         multipath -v2 &> /dev/null
>         multipath -ll &> /dev/null
>         sleep $interval
> done
> 
> multipathd will have a segfault after about 30 mins.
> 
> The reason is that mpp->hwe is accessed after hwe is already freed.
> In
> extract_hwe_from_path func, mpp->hwe is set to pp->hwe, so they
> points to the
> same hwe. For some reasons, pp->mpp will be set to NULL in
> orphan_path func.
> Then, pp and hwe will be freed with (pp->mpp == NULL) in free_path
> func
> called by ev_remove_path func. However, mpp->hwe is not set to NULL
> while hwe
> is already freed. So, when iscsi device logs in and new path is added
> to mpp,
> mpp->hwe will be accessed in select_pgfailback func. Finally, use-
> after-free
> problem occurs.
> 
> The procedure details given as follows,
> 1.wait_dmevents thread
> wait_dmevents
> 	->dmevent_loop
> 		->update_multipath
> 			->__setup_multipath
> 				->update_multipath_strings
> 					 -> sync_paths
> 						->orphan_path
> 2.uevqloop thread  (iscsi log out, remove path)
> uevqloop
> ->uevent_dispatch
> 	->service_uevq
> 		->uev_remove_path
> 			->ev_remove_path  //pp->mpp is NULL
> 				->free_path(pp)
> 			//pp->hew are freed but mpp->hwe is not set to
> NULL
> 3.ev_remove_path  (iscsi log in, add path)
> uevqloop
> ->uevent_dispatch
> 	->service_uevq
> 		->ev_add_path
> 			->select_pgfailback //mpp->hwe is accessed
> 
> Here, we will set mpp->hwe to NULL before setting pp->map to NULL in
> orphan_path
> func.
> 
> Signed-off-by: Tianxiong Lu <lutianxiong at huawei.com>
> Signed-off-by: lixiaokeng <lixiaokeng at huawei.com>
> Signed-off-by: Zhiqiang Liu <liuzhiqiang26 at huawei.com>
> ---
>  libmultipath/structs_vec.c | 2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/libmultipath/structs_vec.c b/libmultipath/structs_vec.c
> index 3dbbaa0..9bbe5d1 100644
> --- a/libmultipath/structs_vec.c
> +++ b/libmultipath/structs_vec.c
> @@ -93,6 +93,8 @@ int adopt_paths(vector pathvec, struct multipath
> *mpp)
>  void orphan_path(struct path *pp, const char *reason)
>  {
>  	condlog(3, "%s: orphan path, %s", pp->dev, reason);
> +	if (pp->mpp && pp->mpp->hwe == pp->hwe)
> +		pp->mpp->hwe = NULL;
>  	pp->mpp = NULL;
>  	pp->dmstate = PSTATE_UNDEF;
>  	pp->uid_attribute = NULL;

This isn't wrong, but I think sync_paths() has to be fixed as well.
I'll send a small series soon, based on your patch.

Martin