Comment 0 for bug 1796542

TLDR: commit 72ecad22d9f198aafee64218512e02ffa7818671 (in v4.10) introduced silent data corruption for O_DIRECT uses, it's fixed in 17d51b10d7773e4618bcac64648f30f12d4078fb (in v4.18)

A silent data corruption was introduced in v4.10-rc1 with commit 72ecad22d9f198aafee64218512e02ffa7818671 and was fixed in v4.18-rc7 with commit 17d51b10d7773e4618bcac64648f30f12d4078fb. It affects users of O_DIRECT, in our case a KVM virtual machine with drives which use qemu's "cache=none" option.

This is the commit which fixes the issue:
---------------------
commit 17d51b10d7773e4618bcac64648f30f12d4078fb
Author: Martin Wilck <email address hidden>
Date: Wed Jul 25 23:15:09 2018 +0200

    block: bio_iov_iter_get_pages: pin more pages for multi-segment IOs

    bio_iov_iter_get_pages() currently only adds pages for the next non-zero
    segment from the iov_iter to the bio. That's suboptimal for callers,
    which typically try to pin as many pages as fit into the bio. This patch
    converts the current bio_iov_iter_get_pages() into a static helper, and
    introduces a new helper that allocates as many pages as

     1) fit into the bio,
     2) are present in the iov_iter,
     3) and can be pinned by MM.

    Error is returned only if zero pages could be pinned. Because of 3), a
    zero return value doesn't necessarily mean all pages have been pinned.
    Callers that have to pin every page in the iov_iter must still call this
    function in a loop (this is currently the case).

    This change matters most for __blkdev_direct_IO_simple(), which calls
    bio_iov_iter_get_pages() only once. If it obtains less pages than
    requested, it returns a "short write" or "short read", and
    __generic_file_write_iter() falls back to buffered writes, which may
    lead to data corruption.

    Fixes: 72ecad22d9f1 ("block: support a full bio worth of IO for simplified bdev direct-io")
    Reviewed-by: Christoph Hellwig <email address hidden>
    Signed-off-by: Martin Wilck <email address hidden>
    Signed-off-by: Jens Axboe <email address hidden>
-------------------------

Since there were a lot of components involved in the initial report to us (xfs, guest kernel, guest virtio drivers, qemu, host kernel, storage system), we had to isolate it. This is the commit which fixes the data corruption bug. We created a reliable reproduction and tested with the patch and without the patch. We also created a version of the kernel which prints when the data-corrupting path in the kernel is triggered.

> 1) The release of Ubuntu you are using, via 'lsb_release -rd' or System -> About Ubuntu

# lsb_release -rd
Description: Ubuntu 18.04.1 LTS
Release: 18.04

> 2) The version of the package you are using, via 'apt-cache policy pkgname' or by checking in Software Center

# apt-cache policy linux-image-4.15.0-36-generic
linux-image-4.15.0-36-generic:
 Installed: 4.15.0-36.39
 Candidate: 4.15.0-36.39
 Version table:
*** 4.15.0-36.39 500
    500 http://archive.ubuntu.com/ubuntu bionic-security/main amd64 Packages
    500 http://archive.ubuntu.com/ubuntu bionic-updates/main amd64 Packages
    100 /var/lib/dpkg/status

> 3) What you expected to happen

We ran a fio random write workload over 8x 512MB files over XFS in guest OS, over qemu/kvm, over kernel 4.15.0-36.39-generic.
qemu-system was configured with cache=none, which means Direct IO. This is a very common configuration.
qemu-system was with aio=threads -- the default.

We were expecting no data corruption.

> 4) What happened instead

The guest filesystem was corrupted.