On 2019/2/12 下午2:22, Steve Leung wrote: > > > ----- Original Message ----- >> From: "Qu Wenruo" >> To: "STEVE LEUNG" , linux-btrfs@vger.kernel.org >> Sent: Sunday, February 10, 2019 6:52:23 AM >> Subject: Re: corruption with multi-device btrfs + single bcache, won't mount > >> ----- Original Message ----- >> From: "Qu Wenruo" >> On 2019/2/10 下午2:56, STEVE LEUNG wrote: >>> Hi all, >>> >>> I decided to try something a bit crazy, and try multi-device raid1 btrfs on >>> top of dm-crypt and bcache. That is: >>> >>> btrfs -> dm-crypt -> bcache -> physical disks >>> >>> I have a single cache device in front of 4 disks. Maybe this wasn't >>> that good of an idea, because the filesystem went read-only a few >>> days after setting it up, and now it won't mount. I'd been running >>> btrfs on top of 4 dm-crypt-ed disks for some time without any >>> problems, and only added bcache (taking one device out at a time, >>> converting it over, adding it back) recently. >>> >>> This was on Arch Linux x86-64, kernel 4.20.1. >>> >>> dmesg from a mount attempt (using -o usebackuproot,nospace_cache,clear_cache): >>> >>> [ 267.355024] BTRFS info (device dm-5): trying to use backup root at mount time >>> [ 267.355027] BTRFS info (device dm-5): force clearing of disk cache >>> [ 267.355030] BTRFS info (device dm-5): disabling disk space caching >>> [ 267.355032] BTRFS info (device dm-5): has skinny extents >>> [ 271.446808] BTRFS error (device dm-5): parent transid verify failed on >>> 13069706166272 wanted 4196588 found 4196585 >>> [ 271.447485] BTRFS error (device dm-5): parent transid verify failed on >>> 13069706166272 wanted 4196588 found 4196585 >> >> When this happens, there is no good way to completely recover (btrfs >> check pass after the recovery) the fs. >> >> We should enhance btrfs-progs to handle it, but it will take some time. >> >>> [ 271.447491] BTRFS error (device dm-5): failed to read block groups: -5 >>> [ 271.455868] BTRFS error (device dm-5): open_ctree failed >>> >>> btrfs check: >>> >>> parent transid verify failed on 13069706166272 wanted 4196588 found 4196585 >>> parent transid verify failed on 13069706166272 wanted 4196588 found 4196585 >>> parent transid verify failed on 13069706166272 wanted 4196588 found 4196585 >>> parent transid verify failed on 13069706166272 wanted 4196588 found 4196585 >>> Ignoring transid failure >>> ERROR: child eb corrupted: parent bytenr=13069708722176 item=7 parent level=2 >>> child level=0 >>> ERROR: cannot open file system >>> >>> Any simple fix for the filesystem? It'd be nice to recover the data >>> that's hopefully still intact. I have some backups that I can dust >>> off if it really comes down to it, but it's more convenient to >>> recover the data in-place. >> >> However there is a patch to address this kinda "common" corruption scenario. >> >> https://lwn.net/Articles/777265/ >> >> In that patchset, there is a new rescue=bg_skip mount option (needs to >> be used with ro), which should allow you to access whatever you still >> have from the fs. >> >> From other reporters, such corruption is mainly related to extent tree, >> thus data damage should be pretty small. > > Ok I think I spoke too soon. Some files are recoverable, but many > cannot be read. Userspace gets back an I/O error, and the kernel log > reports similar parent transid verify failed errors, with what seem > to be similar generation numbers to what I saw in my original mount > error. > > i.e. wants 4196588, found something that's off by usually 2 or 3. > Occasionally there's one that's off by about 1300. That's more or less expected for such transid corruption. The fs is already screwed up. The lowest generation you found during all these error message could be when the first corruption happens. (And it may date back to very old days) > > There are multiple snapshots on this filesystem (going back a few > days), and the same file in each snapshot seems to be equally > affected, even if the file hasn't changed in many months.> > Metadata seems to be intact - I can stat every file in one of the > snapshots and I don't get any errors back. > > Any other ideas? It kind of seems like "btrfs restore" would be > suitable here, but it sounds like it would need to be taught about > rescue=bg_skip first. Since v4.16.1, btrfs-restore should be OK to ignore extent tree completely, thus you can try btrfs-restore. Btrfs-restore may be a little better, since it can ignore csum errors completely. Thanks, Qu > > Thanks for all the help. Even a partial recovery is a lot better > than what I was facing before. > > Steve >