[moving to tech@] On Tuesday, 20 September 2016 08:03:32 CEST Stefan Fritsch wrote: > On Tue, 20 Sep 2016, Darren Tucker wrote: > > On Tue, Sep 20, 2016 at 1:43 AM, Jan Stary <h...@stare.cz> wrote: > > > This is current/i386 on an ALIX.1E (demsg below). > > > I have an USB disk connected for /backup. > > > > > > Upon every reboot, the filesystem on that disk is dirty: > > > WARNING: R/W mount of /backup denied. Filesystem is not clean - run > > > fsck > > > > I saw something similar on an APU where the root disk was on > > (USB-attached) sdcard: > > http://marc.info/?l=openbsd-misc&m=144237305322074&w=2 > > > > It seems to be a race. There used to be a 4sec pause in the kernel > > that was removed: > > > > """ > > Remove 4 second delay on reboot/shutdown that was added 8 years > > ago to "workaround MP timeout/splhigh/scsi race at reboot time". > > """ > > I think before we re-add some arbitrary delays, we should check if we are > actually sending an explicit cache flush command to the disk controllers. > I have some code somewhere that does this for umount and mount -ur. I will > look for it and send it to tech@, but probably not today.
I found a few cases, where we should send a cache flush but don't. Unfortunately, none of these cases explain the problem seen by Jan and Darren. The cases I have found are: * When a R/W mount is updated to R/O. I will send patches for this in a separate mail. * When a R/W mount is unmounted but there is still another partition from the same disk mounted. * When sync(2) is called. Though I am not 100% sure if we really want to do a cache flush for every sync. Thoughts? For the usb disk issue, some more debugging is necessary. Cheers, Stefan
