On 8/31/20 10:56 AM, Darrick J. Wong wrote:
> On Mon, Aug 31, 2020 at 10:48:59AM -0500, Eric Sandeen wrote:
>> On 8/30/20 8:45 PM, Qian Cai wrote:
>>> It is trivial to trigger a WARN_ON_ONCE(1) in iomap_dio_actor() by
>>> unprivileged users which would taint the kernel, or worse - panic if
>>> panic_on_warn or panic_on_taint is set. Hence, just convert it to
>>> pr_warn_ratelimited() to let users know their workloads are racing.
>>> Thanks Dave Chinner for the initial analysis of the racing reproducers.
>>>
>>> Signed-off-by: Qian Cai <c...@lca.pw>
>>> ---
>>>
>>> v2: Record the path, pid and command as well.
>>>
>>> fs/iomap/direct-io.c | 17 ++++++++++++++++-
>>> 1 file changed, 16 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c
>>> index c1aafb2ab990..66a4502ef675 100644
>>> --- a/fs/iomap/direct-io.c
>>> +++ b/fs/iomap/direct-io.c
>>> @@ -374,6 +374,7 @@ iomap_dio_actor(struct inode *inode, loff_t pos, loff_t
>>> length,
>>> void *data, struct iomap *iomap, struct iomap *srcmap)
>>> {
>>> struct iomap_dio *dio = data;
>>> + char pathname[128], *path;
>>>
>>> switch (iomap->type) {
>>> case IOMAP_HOLE:
>>> @@ -389,7 +390,21 @@ iomap_dio_actor(struct inode *inode, loff_t pos,
>>> loff_t length,
>>> case IOMAP_INLINE:
>>> return iomap_dio_inline_actor(inode, pos, length, dio, iomap);
>>> default:
>>> - WARN_ON_ONCE(1);
>>
>> It seems like we should explicitly catch IOMAP_DELALLOC for this case, and
>> leave the
>> default: as a WARN_ON that is not user-triggerable? i.e.
>>
>> case IOMAP_DELALLOC:
>> <all the fancy warnings>
>> return -EIO;
>> default:
>> WARN_ON_ONCE(1);
>> return -EIO;
>>
>>> + /*
>>> + * DIO is not serialised against mmap() access at all, and so
>>> + * if the page_mkwrite occurs between the writeback and the
>>> + * iomap_apply() call in the DIO path, then it will see the
>>> + * DELALLOC block that the page-mkwrite allocated.
>>> + */
>>> + path = file_path(dio->iocb->ki_filp, pathname,
>>> + sizeof(pathname));
>>> + if (IS_ERR(path))
>>> + path = "(unknown)";
>>> +
>>> + pr_warn_ratelimited("page_mkwrite() is racing with DIO read
>>> (iomap->type = %u).\n"
>>> + "File: %s PID: %d Comm: %.20s\n",
>>> + iomap->type, path, current->pid,
>>> + current->comm);
>>
>> This is very specific ...
>>
>> Do we know that mmap/page_mkwrite is (and will always be) the only way to
>> reach this
>> point?
>>
>> It seems to me that this message won't be very useful for the admin;
>> "pg_mkwrite" may
>> mean something to us, but doubtful for the general public. And "type = 1"
>> won't mean
>> much to the reader, either.
>>
>> Maybe something like:
>>
>> "DIO encountered delayed allocation block, racing buffered+direct? File: %s
>> Comm: %.20s\n"
>>
>> It just seems that a user-facing warning should be something the admin has a
>> chance of
>> acting on without needing to file a bug for analysis by the developers.
>>
>> (though TBH "delayed allocation" probably doesn't mean much to the admin,
>> either)
>
> /me suggests
>
> "Direct I/O collision with buffered write! File: %s..."?
Sure, that sounds good to me. Terser is better.
> I concede that we ought to leave the nastier WARN for the default
> case since there are no other IOMAP_ types and so any other code is
> a sign of a serious screwup.
*nod* thanks.
-Eric
