On 05/21/2010 02:17 PM, Eric Blake wrote:
> On 05/21/2010 11:03 AM, Cole Robinson wrote:
>> If a directory pool contains pipes or sockets, a pool start can fail or hang:
>>
>>
https://bugzilla.redhat.com/show_bug.cgi?id=589577
>>
>> We already try to avoid these special files, but only attempt after
>> opening the path, which is where the problems lie. Unify volume opening
>> into a single function which runs stat() before any open() call. Directory
>
> This sentence of the commit log is out-of-date...
>
>> pools can then proceed along, ignoring the invalid files.
>>
>> v2: Use ATTRIBUTE_NONNULL. Drop stat check, just open with
>> O_NONBLOCK|O_NOCTTY.
>
> ...given this sentence.
>
>> +int
>> +virStorageBackendVolOpen(const char *path)
>> +{
>> + int fd;
>> +
>> + if ((fd = open(path, O_RDONLY|O_NONBLOCK|O_NOCTTY)) < 0) {
>> + virReportSystemError(errno,
>> + _("cannot open volume '%s'"),
>> + path);
>> + return -1;
>> + }
>
> I'm thinking that we still want to use fstat() (after the open()) and
> reject directories, sockets, and pipes (or more accurately, accept only
> regular files, block devices, and possibly character devices).
>
> In other words, you got rid of the stat/open race (good), but also got
> rid of the strictness provided by validating [f]stat (not so good).
>
Sorry, I should have pointed out that there is already code that does
this, see virStorageBackendUpdateVolTargetInfoFD, which is always called
after VolOpen. The existing code is all confusingly arranged though, so
maybe I should take a harder look at organizing.
That check isn't sufficient. We need to explicitly *allow* directories
to be reported, since a pool may point at a directory containing trees
for container based virt root filesystems. We should drop block+char
devices for directory based pools too, since those aren't relevant,
being dealt with by other storage pool backends.
Daniel
--
|: Red Hat, Engineering, London -o-
:|
|: GnuPG: 7D3B9505 -o- F3C9 553F A1DA 4AC2 5648 23C1 B3DF F742 7D3B 9505 :|