Re: NFS: handle source server reboot

From: Olga Kornievskaia
Date: Fri Nov 15 2019 - 14:52:01 EST


On Fri, Nov 15, 2019 at 5:06 AM Colin Ian King <colin.king@xxxxxxxxxxxxx> wrote:
>
> Hi,
>
> Static analysis with Coverity has detected a memory leak in the
> following commit:
>
> commit 0e65a32c8a569db363048e17a708b1a0913adbef
> Author: Olga Kornievskaia <kolga@xxxxxxxxxx>
> Date: Fri Jun 14 14:38:40 2019 -0400
>
> NFS: handle source server reboot
>
> In function __nfs4_copy_file_range() in fs/nfs/nfs4file.c, analysis is
> as follows:
>
>
> 155retry:
> 5. Condition !nfs42_files_from_same_server(file_in, file_out), taking
> false branch.
> 9. Condition !nfs42_files_from_same_server(file_in, file_out), taking
> false branch.
>
> 156 if (!nfs42_files_from_same_server(file_in, file_out)) {
> 157 /* for inter copy, if copy size if smaller than 12 RPC
> 158 * payloads, fallback to traditional copy. There are
> 159 * 14 RPCs during an NFSv4.x mount between source/dest
> 160 * servers.
> 161 */
> 162 if (sync ||
> 163 count <= 14 *
> NFS_SERVER(file_inode(file_in))->rsize)
> 164 return -EOPNOTSUPP;
> 165 cn_resp = kzalloc(sizeof(struct nfs42_copy_notify_res),
> 166 GFP_NOFS);
> 167 if (unlikely(cn_resp == NULL))
> 168 return -ENOMEM;
> 169
> 170 ret = nfs42_proc_copy_notify(file_in, file_out, cn_resp);
> 171 if (ret) {
> 172 ret = -EOPNOTSUPP;
> 173 goto out;
> 174 }
> 175 nss = &cn_resp->cnr_src;
> 176 cnrs = &cn_resp->cnr_stateid;
> 177 }
> 178 ret = nfs42_proc_copy(file_in, pos_in, file_out, pos_out, count,
> 179 nss, cnrs, sync);
> 180out:
> 6. freed_arg: kfree frees cn_resp.
>
> CID 91571 (#1 of 1): Double free (USE_AFTER_FREE)10. double_free:
> Calling kfree frees pointer cn_resp which has already been freed.
>
> 181 kfree(cn_resp);
>
> 7. Condition ret == -11, taking true branch.
>
> 182 if (ret == -EAGAIN)
> 8. Jumping to label retry.
>
> 183 goto retry;
> 184 return ret;
> 185}
> 186
>
> On the 2nd iteration of the retry loop, cn_resp is being free'd twice if
> the call to nfs42_files_from_same_server() returns zero since cn_resp is
> not kalloc'd in the 2nd iteration. A naive fix would be to set cn_resp
> to NULL after the kfree on line 181, but I'm not sure if there is a
> better way to resolve this.
>

If a definition of double free include freeing a null pointer twice,
then I agree this is a valid catch. Cases are when servers are the
same kfree(cn_resp) is called with a null argument and if retried will
be called with a null argument again (since kfree doesn't care about
null it shouldn't be a problem). If servers are not the same, then
memory is allocated then freed then allocated again. When code is
re-executed (on retry), it's not possible for condition to change from
servers being same or different.

I'll send a patch that conditions the kfree() to only when it was
allocated. Hopefully, it won't trip the analysis that way.