Just updating the numbers for Focal / Original in comment #9 (the issue did happen more times overnight.)
[ 10.018315] rc.local[744]: 61 FAILED [ 10.018882] rc.local[744]: 1186 PASSED -- 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: In Progress 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: In Progress 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
