On 12/8/20 12:36 AM, David Howells wrote:
> Randy Dunlap <rdunlap@xxxxxxxxxxxxx> wrote:
>
>> Otherwise please look at the patch below.
>
> The patch won't help, since it's not going through sys_fsconfig() - worse, it
> introduces two new errors.
>
>> fc->source = param->string;
>> - param->string = NULL;
>
> This will cause the string now attached to fc->source to be freed by the
> caller. No, the original is doing the correct thing here. The point is to
> steal the string.
>
>> @@ -262,7 +262,9 @@ static int vfs_fsconfig_locked(struct fs
>>
>> - return vfs_parse_fs_param(fc, param);
>> + ret = vfs_parse_fs_param(fc, param);
>> + kfree(param->string);
>> + return ret;
>
> But your stack trace shows you aren't going through sys_fsconfig(), so this
> function isn't involved. Further, this introduces a double free, since
> sys_fsconfig() frees param.string after it drops uapi_mutex.
>
> Looking at the backtrace:
>
>> kmemdup_nul+0x2d/0x70 mm/util.c:151
>> vfs_parse_fs_string+0x6e/0xd0 fs/fs_context.c:155
>> generic_parse_monolithic+0xe0/0x130 fs/fs_context.c:201
>> do_new_mount fs/namespace.c:2871 [inline]
>> path_mount+0xbbb/0x1170 fs/namespace.c:3205
>> do_mount fs/namespace.c:3218 [inline]
>> __do_sys_mount fs/namespace.c:3426 [inline]
>> __se_sys_mount fs/namespace.c:3403 [inline]
>> __x64_sys_mount+0x18e/0x1d0 fs/namespace.c:3403
>
> A couple of possibilities spring to mind from that: maybe
> vfs_parse_fs_string() is not releasing the param.string - but that's not the
> problem since we stole the string and the free is definitely there at the
> bottom of the function:
>
> int vfs_parse_fs_string(struct fs_context *fc, const char *key,
> const char *value, size_t v_size)
> {
> ...
> kfree(param.string);
> return ret;
> }
>
> or fc->source is not being cleaned up in vfs_clean_context() - but that's
> there as well:
>
> void vfs_clean_context(struct fs_context *fc)
> {
> ...
> kfree(fc->source);
> fc->source = NULL;
>
> In either of these cases, I would expect this to have already become evident
> from other filesystem mounts as there would be a lot of leaking going on,
> particularly with the first.
>
> Now the backtrace only shows what the state was when the string was allocated;
> it doesn't show what happened to it after that, so another possibility is that
> the filesystem being mounted nicked what vfs_parse_fs_param() had rightfully
> stolen, transferring fc->source somewhere else and then failed to release it -
> most likely on mount failure (ie. it's an error handling bug in the
> filesystem).
>
> Do we know what filesystem it was?
Yes, it's call AFS (or kAFS).
Thanks for your comments & help.
--
~Randy
