On 2019/08/06 5:31, Jens Axboe wrote:
> On 8/5/19 11:31 AM, Jens Axboe wrote:
>> On 8/5/19 11:15 AM, Darrick J. Wong wrote:
>>> Hi Damien,
>>>
>>> I noticed a regression in xfs/747 (an unreleased xfstest for the
>>> xfs_scrub media scanning feature) on 5.3-rc3. I'll condense that down
>>> to a simpler reproducer:
>>>
>>> # dmsetup table
>>> error-test: 0 209 linear 8:48 0
>>> error-test: 209 1 error
>>> error-test: 210 6446894 linear 8:48 210
>>>
>>> Basically we have a ~3G /dev/sdd and we set up device mapper to fail IO
>>> for sector 209 and to pass the io to the scsi device everywhere else.
>>>
>>> On 5.3-rc3, performing a directio pread of this range with a < 1M buffer
>>> (in other words, a request for fewer than MAX_BIO_PAGES bytes) yields
>>> EIO like you'd expect:
>>>
>>> # strace -e pread64 xfs_io -d -c 'pread -b 1024k 0k 1120k' /dev/mapper/error-test
>>> pread64(3, 0x7f880e1c7000, 1048576, 0) = -1 EIO (Input/output error)
>>> pread: Input/output error
>>> +++ exited with 0 +++
>>>
>>> But doing it with a larger buffer succeeds(!):
>>>
>>> # strace -e pread64 xfs_io -d -c 'pread -b 2048k 0k 1120k' /dev/mapper/error-test
>>> pread64(3, "XFSB\0\0\20\0\0\0\0\0\0\fL\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 1146880, 0) = 1146880
>>> read 1146880/1146880 bytes at offset 0
>>> 1 MiB, 1 ops; 0.0009 sec (1.124 GiB/sec and 1052.6316 ops/sec)
>>> +++ exited with 0 +++
>>>
>>> (Note that the part of the buffer corresponding to the dm-error area is
>>> uninitialized)
>>>
>>> On 5.3-rc2, both commands would fail with EIO like you'd expect. The
>>> only change between rc2 and rc3 is commit 0eb6ddfb865c ("block: Fix
>>> __blkdev_direct_IO() for bio fragments").
>>>
>>> AFAICT we end up in __blkdev_direct_IO with a 1120K buffer, which gets
>>> split into two bios: one for the first BIO_MAX_PAGES worth of data (1MB)
>>> and a second one for the 96k after that.
>>>
>>> I think the problem is that every time we submit a bio, we increase ret
>>> by the size of that bio, but at the time we do that we have no idea if
>>> the bio is going to succeed or not. At the end of the function we do:
>>>
>>> if (!ret)
>>> ret = blk_status_to_errno(dio->bio.bi_status);
>>>
>>> Which means that we only pick up the IO error if we haven't already set
>>> ret. I suppose that was useful for being able to return a short read,
>>> but now that we always increment ret by the size of the bio, we act like
>>> the whole buffer was read. I tried a -rc2 kernel and found that 40% of
>>> the time I'd get an EIO and the rest of the time I got a short read.
>>>
>>> Not sure where to go from here, but something's not right...
>>
>> I'll take a look.
>
> How about this? The old code did:
>
> if (!ret)
> ret = blk_status_to_errno(dio->bio.bi_status);
> if (likely(!ret))
> ret = dio->size;
>
> where 'ret' was just tracking the error. With 'ret' now being the
> positive IO size, we should overwrite it if ret is >= 0, not just if
> it's zero.
>
> Also kill a use-after-free.
>
> diff --git a/fs/block_dev.c b/fs/block_dev.c
> index a6f7c892cb4a..67c8e87c9481 100644
> --- a/fs/block_dev.c
> +++ b/fs/block_dev.c
> @@ -386,6 +386,7 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter, int nr_pages)
>
> ret = 0;
> for (;;) {
> + ssize_t this_size;
> int err;
>
> bio_set_dev(bio, bdev);
> @@ -433,13 +434,14 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter, int nr_pages)
> polled = true;
> }
>
> + this_size = bio->bi_iter.bi_size;
> qc = submit_bio(bio);
> if (qc == BLK_QC_T_EAGAIN) {
> if (!ret)
> ret = -EAGAIN;
> goto error;
> }
> - ret = dio->size;
> + ret += this_size;
>
> if (polled)
> WRITE_ONCE(iocb->ki_cookie, qc);
> @@ -460,13 +462,14 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter, int nr_pages)
> atomic_inc(&dio->ref);
> }
>
> + this_size = bio->bi_iter.bi_size;
> qc = submit_bio(bio);
> if (qc == BLK_QC_T_EAGAIN) {
> if (!ret)
> ret = -EAGAIN;
> goto error;
> }
> - ret = dio->size;
> + ret += this_size;
>
> bio = bio_alloc(gfp, nr_pages);
> if (!bio) {
> @@ -494,7 +497,7 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter, int nr_pages)
> __set_current_state(TASK_RUNNING);
>
> out:
> - if (!ret)
> + if (ret >= 0)
> ret = blk_status_to_errno(dio->bio.bi_status);
>
> bio_put(&dio->bio);
>
Jens,
I would set "this_size" when dio->size is being incremented though, to avoid
repeating it.
if (nowait)
bio->bi_opf |= (REQ_NOWAIT | REQ_NOWAIT_INLINE);
+ this_size = bio->bi_iter.bi_size;
- dio->size += bio->bi_iter.bi_size;
+ dio->size += this_size;
pos += bio->bi_iter.bi_size;
In any case, looking again at this code, it looks like there is a problem with
dio->size being incremented early, even for fragments that get BLK_QC_T_EAGAIN,
because dio->size is being used in blkdev_bio_end_io(). So an incorrect size can
be reported to user space in that case on completion (e.g. large asynchronous
no-wait dio that cannot be issued in one go).
So maybe something like this ? (completely untested)
diff --git a/fs/block_dev.c b/fs/block_dev.c
index 75cc7f424b3a..77714e03c21e 100644
--- a/fs/block_dev.c
+++ b/fs/block_dev.c
@@ -349,7 +349,7 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter
*iter, int nr_pages)
loff_t pos = iocb->ki_pos;
blk_qc_t qc = BLK_QC_T_NONE;
gfp_t gfp;
- ssize_t ret;
+ ssize_t ret = 0;
if ((pos | iov_iter_alignment(iter)) &
(bdev_logical_block_size(bdev) - 1))
@@ -386,6 +386,7 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter
*iter, int nr_pages)
ret = 0;
for (;;) {
+ size_t this_size;
int err;
bio_set_dev(bio, bdev);
@@ -421,7 +422,7 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter
*iter, int nr_pages)
if (nowait)
bio->bi_opf |= (REQ_NOWAIT | REQ_NOWAIT_INLINE);
- dio->size += bio->bi_iter.bi_size;
+ this_size = bio->bi_iter.bi_size;
pos += bio->bi_iter.bi_size;
nr_pages = iov_iter_npages(iter, BIO_MAX_PAGES);
@@ -435,11 +436,11 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter
*iter, int nr_pages)
qc = submit_bio(bio);
if (qc == BLK_QC_T_EAGAIN) {
- if (!ret)
+ if (!dio->size)
ret = -EAGAIN;
goto error;
}
- ret = dio->size;
+ dio->size += this_size;
if (polled)
WRITE_ONCE(iocb->ki_cookie, qc);
@@ -462,15 +463,15 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter
*iter, int nr_pages)
qc = submit_bio(bio);
if (qc == BLK_QC_T_EAGAIN) {
- if (!ret)
+ if (!dio->size)
ret = -EAGAIN;
goto error;
}
- ret = dio->size;
+ dio->size += this_size;
bio = bio_alloc(gfp, nr_pages);
if (!bio) {
- if (!ret)
+ if (!dio->size)
ret = -EAGAIN;
goto error;
}
@@ -496,10 +497,15 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter
*iter, int nr_pages)
out:
if (!ret)
ret = blk_status_to_errno(dio->bio.bi_status);
+ if (likely(!ret))
+ ret = dio->size;
bio_put(&dio->bio);
return ret;
--
Damien Le Moal
Western Digital Research
