Setting tag verification-done-bionic.
Test results with stress-ng: 2420 PASSED, 0 FAILED.
Test results w/ (x)fstests: no regressions in data=ordered and
data=journal (below).
---
Bionic, data=ordered: no regression.
$ ./fstests-check-logs.sh prop-ordered/xfstests.log*
Kernel version: 4.15.0-160-generic #168-Ubuntu
- Number of Runs: 10
- Always fail (failed every time)
10 ext4/045
10 generic/044
10 generic/045
10 generic/046
10 generic/082
10 generic/095
10 generic/219
10 generic/230
10 generic/231
10 generic/232
10 generic/233
10 generic/235
10 generic/270
10 generic/299
10 generic/300
10 generic/381
10 generic/382
10 generic/491
10 generic/553
10 generic/554
10 generic/555
10 generic/565
10 generic/566
10 generic/569
10 generic/587
10 generic/598
10 generic/600
10 generic/623
10 generic/629
10 shared/298
- Flaky (failed, but not every time)
2 generic/475
7 generic/388
Bionic, data=journal: no regression.
$ ./fstests-check-logs.sh prop-journal/xfstests.log*
Kernel version: 4.15.0-160-generic #168-Ubuntu
- Number of Runs: 10
- Always fail (failed every time)
10 ext4/045
10 generic/082
10 generic/219
10 generic/223
10 generic/230
10 generic/231
10 generic/232
10 generic/233
10 generic/235
10 generic/270
10 generic/381
10 generic/382
10 generic/441
10 generic/491
10 generic/553
10 generic/554
10 generic/555
10 generic/565
10 generic/566
10 generic/569
10 generic/598
10 generic/600
10 generic/623
10 generic/629
10 shared/298
- Flaky (failed, but not every time)
1 generic/547
5 generic/371 (* improved from Always fail)
7 generic/347 (* improved from Always fail)
7 generic/388
** Tags removed: verification-needed-bionic
** Tags added: verification-done-bionic
--
You received this bug notification because you are a member of Kernel
Packages, which is subscribed to linux in Ubuntu.
https://bugs.launchpad.net/bugs/1847340
Title:
ext4 journal recovery fails w/ data=journal + mmap
Status in linux package in Ubuntu:
Incomplete
Status in linux source package in Bionic:
Fix Committed
Status in linux source package in Disco:
Won't Fix
Status in linux source package in Eoan:
Won't Fix
Status in linux source package in Focal:
Fix Committed
Status in linux source package in Hirsute:
Fix Released
Bug description:
[Impact]
With mmap()ed files on ext4's data journaling it's possible to change
a mapped page's buffers contents during their jbd2 transaction commit
(as currently nothing prevents/blocks the write access at that time.)
This might happen between the buffers checksum calculation and actual
write to journal, so the (old) checksum is invalid for the (new) data.
If the system crashes after that, but before such journal entry makes
it to the filesystem, the journal replay on the next mount just fails,
and the filesystem now requires fsck. (apparently curtin might set up
/etc/fstab with passno=0, requiring manual intervention.)
[39751.096455] EXT4-fs: Warning: mounting with data=journal disables
delayed allocation and O_DIRECT support!
[39751.114435] JBD2: Invalid checksum recovering block 87305 in log
[39751.146133] JBD2: Invalid checksum recovering block 88039 in log
[39751.195950] JBD2: Invalid checksum recovering block 49633 in log
[39751.265158] JBD2: recovery failed
[39751.265163] EXT4-fs (vdc): error loading journal
[Fix]
The fix is to write-protect the pages during journal transaction commit,
so that writes to mapped pages hit a page fault, then ext4's page_mkwrite
hook can block until the commit finishes and the buffers can be modified.
In order to do that, add jbd2 journal callbacks that the filesystems can
customize, called before/after the critical region in transaction commit,
then have ext4 in data journaling mode to write-protect the pages whose
buffers are being committed (and handle cases that need pages redirtied.)
The changes are restricted to the data journaling mode and page_mkwrite
hook, and other modes/paths use the same code/behavior in the callbacks.
[Test Case]
Set up an ext4 filesystem in data journaling mode, and run stress-ng's
mmap file test on it, then crash the system after a bit; check whether
the filesystem can mount again or not (i.e., with jbd2 checksum errors.)
# mkfs.ext4 $DEV
# mount -o data=journal $DEV $DIR
# cd $DIR
# stress-ng --mmap $((4*$(nproc))) --mmap-file &
# sleep 60
# echo c >/proc/sysrq-trigger
...
# mount -o data=journal $DEV $DIR # PASS/FAIL.
# dmesg | tail
[Regression Potential]
Regressions would likely manifest in ext4 data journaling mode (which
is not the default mode, 'ordered') with memory mapped access, as the
other modes/paths are largely unaffected by the changes/same behavior.
This has been tested with (x)fstests, that showed no regressions on
data=ordered and data=journal on both Bionic and Focal (with kernel
versions 4.15.0-156-generic and 5.4.0-84-generic) w/in 10 runs each.
And the stress-ng test-case as well. (Numbers/details in the LP bug.)
[Other info]
The patchset is applied on 5.10, so Hirsute (5.11) is already fixed;
only Focal and Bionic need it.
There are little changes in the patches between Focal and Bionic
(mostly minor backport adjustments, mainly due to no vm_fault_t)
but unfortunately that needs separate versions for most patches.
...
[Original Bug Description]
[Impact]
In the event of a loss of power, ext4 filesystems mounted w/
data=journal,journal_checksum are subject to a corruption issue that requires a
fsck to recover. This is exacerbated by installations by curtin that set
passno=0 in /etc/fstab, preventing fsck from running automatically and thus
requiring a manual recovery. And *that* is further exacerbated because
initramfs-tools is smart enough to not include fsck.ext4 when passno=0 is
detected in /etc/fstab, requiring the user to boot from recovery media.
[Test Case]
Forcibly power cycle a system running 'stress-ng --dir 0'. I've created a
package to automate the reproduction:
https://git.launchpad.net/~dannf/+git/dgx2-ext4-csum-repro?h=master
[Fix]
[Regression Risk]
To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1847340/+subscriptions
--
Mailing list: https://launchpad.net/~kernel-packages
Post to : kernel-packages@lists.launchpad.net
Unsubscribe : https://launchpad.net/~kernel-packages
More help : https://help.launchpad.net/ListHelp
