Have you tried the same file 2x and verified the corruption is in the same places and looks the same? I have not as of yet seen write corruption (except when a vendors disk was resetting and it was lying about having written the data prior to the crash, these were ssds, if your disk write cache is on and you have a disk reset this can also happen), but have not seen "lost writes" otherwise, but would expect the 2 read corruption I have seen to also be able to cause write issues. So for that look for scsi notifications for disk resets that should not happen. I have had a "bad" controller cause read corruptions, those corruptions would move around, replacing the controller resolved it, so there may be lack of error checking "inside" some paths in the card. Lucky I had a number of these controllers and had cold spares for them. The give away here was 2 separate buses with almost identical load with 6 separate disks each and all12 disks on 2 buses had between 47-52 scsi errors, which points to the only component shared (the controller). The backplane and cables are unlikely in general cause this, there is too much error checking between the controller and the disk from what I know. I have had pre-pcie bus (PCI-X bus, 2 slots shared, both set to 133 cause random read corruptions, lowering speed to 100 fixed it), this one was duplicated on multiple identical pieces of hw with all different parts on the duplication machine. I have also seen lost writes (from software) because someone did a seek without doing a flush which in some versions of the libs loses the unfilled block when the seek happens (this is noted in the man page, and I saw it 20years ago, it is still noted in the man page, so no idea if it was ever fixed). So has more than one application been noted to see the corruption? So one question, have you seen the corruption in a path that would rely on one controller, or all corruptions you have seen involving more than one controller? Isolate and test each controller if you can, or if you can afford to replace it and see if it continues. On Thu, May 7, 2020 at 12:33 PM Michal Soltys <msoltyspl@xxxxxxxxx> wrote: > > Note: this is just general question - if anyone experienced something similar or could suggest how to pinpoint / verify the actual cause. > > Thanks to btrfs's checksumming we discovered somewhat (even if quite rare) nasty silent corruption going on on one of our hosts. Or perhaps "corruption" is not the correct word - the files simply have precise 4kb (1 page) of incorrect data. The incorrect pieces of data look on their own fine - as something that was previously in the place, or written from wrong source. > > The hardware is (can provide more detailed info of course): > > - Supermicro X9DR7-LN4F > - onboard LSI SAS2308 controller (2 sff-8087 connectors, 1 connected to backplane) > - 96 gb ram (ecc) > - 24 disk backplane > > - 1 array connected directly to lsi controller (4 disks, mdraid5, internal bitmap, 512kb chunk) > - 1 array on the backplane (4 disks, mdraid5, journaled) > - journal for the above array is: mdraid1, 2 ssd disks (micron 5300 pro disks) > - 1 btrfs raid1 boot array on motherboard's sata ports (older but still fine intel ssds from DC 3500 series) > > Raid 5 arrays are in lvm volume group, and the logical volumes are used by VMs. Some of the volumes are linear, some are using thin-pools (with metadata on the aforementioned intel ssds, in mirrored config). LVM > uses large extent sizes (120m) and the chunk-size of thin-pools is set to 1.5m to match underlying raid stripe. Everything is cleanly aligned as well. > > With a doze of testing we managed to roughly rule out the following elements as being the cause: > > - qemu/kvm (issue occured directly on host) > - backplane (issue occured on disks directly connected via LSI's 2nd connector) > - cable (as a above, two different cables) > - memory (unlikely - ECC for once, thoroughly tested, no errors ever reported via edac-util or memtest) > - mdadm journaling (issue occured on plain mdraid configuration as well) > - disks themselves (issue occured on two separate mdadm arrays) > - filesystem (issue occured on both btrfs and ext4 (checksumed manually) ) > > We did not manage to rule out (though somewhat _highly_ unlikely): > > - lvm thin (issue always - so far - occured on lvm thin pools) > - mdraid (issue always - so far - on mdraid managed arrays) > - kernel (tested with - in this case - debian's 5.2 and 5.4 kernels, happened with both - so it would imply rather already longstanding bug somewhere) > > And finally - so far - the issue never occured: > > - directly on a disk > - directly on mdraid > - on linear lvm volume on top of mdraid > > As far as the issue goes it's: > > - always a 4kb chunk that is incorrect - in a ~1 tb file it can be from a few to few dozens of such chunks > - we also found (or rather btrfs scrub did) a few small damaged files as well > - the chunks look like a correct piece of different or previous data > > The 4kb is well, weird ? Doesn't really matter any chunk/stripes sizes anywhere across the stack (lvm - 120m extents, 1.5m chunks on thin pools; mdraid - default 512kb chunks). It does nicely fit a page though ... > > Anyway, if anyone has any ideas or suggestions what could be happening (perhaps with this particular motherboard or vendor) or how to pinpoint the cause - I'll be grateful for any.
