Sergey Bugaev, le jeu. 15 juin 2023 11:52:24 +0300, a ecrit: > On Thu, Jun 15, 2023 at 2:14 AM Samuel Thibault <samuel.thiba...@gnu.org> > wrote: > > > So the installed system is not to blame, it must be that the installer > > > is corrupting the FS somehow. > > > > Possibly some missing disk cache flush. I believe we have done fixes > > there in the past, but possibly some things are still missing... > > Hmm, I might be looking in the wrong places for it, but where does > disk flushing happen at all?
I thought we had it somewhere, but I don't find it either (no 0xE7 or 0xEA in linux/src/include/linux/hdreg.h). It could be called from host_reboot() at least. > What I can see from looking through the source is: sync(1) or whatever > -> fsys_syncfs -> diskfs_sync_everything -> pager_sync -> > file_pager_write_page -> pending_blocks_write -> store_write -> > device_write -> (*dev->emul_ops->write) -> (*bd->ds->fops->write) -> > block_write The device interface itself would need to get an additionnal RPC to trigger such a flush. Then libstore support it, libpager support it, etc. That being said, in principle a machine reboot shouldn't require such flush? When a DMA write is reported as done, the device should be supposed actually write data even if another OS instance comes up? > > Or perhaps. > > (truncated message?) No, I was thinking about something but eventually dropped it but missed dropping that part. Samuel
