From mboxrd@z Thu Jan 1 00:00:00 1970 From: David Brown Subject: Re: Recovering from a Bad Resilver / Rebuild Date: Tue, 27 Sep 2011 11:27:43 +0200 Message-ID: References: <09f356fa46129bd08dd45752c0f736de.squirrel@www.maxstr.com> <20110926145248.6ffc5f02@notabene.brown> <20110926130351.63adc330@natsu> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: Sender: linux-raid-owner@vger.kernel.org To: linux-raid@vger.kernel.org List-Id: linux-raid.ids On 27/09/2011 01:46, Kenn wrote: >>> On Mon, 26 Sep 2011 14:52:48 +1000 >>> NeilBrown wrote: >>> >>> On Sun, 25 Sep 2011 21:23:31 -0700 "Kenn" wrote: >>> >>> So that brings up another point -- I've been reading through your blog, >>> and I acknowledge your thoughts on not having much benefit to checksums on >>> every block (http://neil.brown.name/blog/20110227114201), but sometimes >>> people like to having that extra lock on their door even though it takes >>> more effort to go in and out of their home. In my five-drive array, if >>> the last five words were the checksums of the blocks on every drive, the >>> checksums off each drive could vote on trusting the blocks of every other >>> drive during the rebuild process, and prevent an idiot (me) from killing >>> his data. It would force wasteful sectors on the drive, perhaps harm >>> performance by squeezing 2+n bytes out of each sector, but if someone >>> wants to protect their data as much as possible, it would be a welcome >>> option where performance is not a priority. >>> >>> Also, the checksums do provide some protection: first, against against >>> partial media failure, which is a major flaw in raid 456 design according >>> to http://www.miracleas.com/BAARF/RAID5_versus_RAID10.txt , and checksum >>> voting could protect against the Atomicity/write-in-place flaw outlined in >>> http://en.wikipedia.org/wiki/RAID#Problems_with_RAID . >>> >>> What do you think? >>> >>> Kenn >> On Sun, 26 Sep 2011 19:56:50 -0700 "David Brown" > wrote: >> >> /raid/ protects against partial media flaws. If one disk in a raid5 >> stripe has a bad sector, that sector will be ignored and the missing >> data will be re-created from the other disks using the raid recovery >> algorithm. If you want to have such protection even when doing a resync >> (as many people do), then use raid6 - it has two parity blocks. >> >> As Neil points out in his blog, it is impossible to fully recover from a >> failure part way through a write - checksum voting or majority voting >> /may/ give you the right answer, but it may not. If you need protection >> against that, you have to have filesystem level control (data logging >> and journalling as well as metafile journalling), or perhaps use raid >> systems with battery backed write caches. > > From what I understand of basic RAID theory, the "If one disk in a raid5 > stripe has a bad sector," is the part that's based on too much faith in > the hardware. RAID trusts the hardware to send it errors when there are > read failures, and it's helpless when the drive reads garbage without an > error and returns it as a good read. During a rebuild this will destroy a > good array. This is the argument against RAID in the articles I listed, > and why checksums in the blocks would be helpful as they get around this > blind spot. And they give early warning on reads that something is dying. > Having each block's checksums in all the other blocks in the stripe lets > md detect a previously failed atomic write and give another early warning. > > I think for people coming from the "can't be too safe" mindset, these > checksums would be welcome, and basically, anyone who signs up for RAID5/6 > already is choosing safety over performance. > I think you have to be very clear on the difference between /unrecoverable/ read errors and /undetected/ read errors. Unrecoverable read errors means the disk controller has seen more bit errors on the disk surface than it is able to correct. These are not a problem for raid, because the disk controller returns an error message - the raid system then re-creates the missing data from the rest of the stripe. This is one of the main reasons for using raid in the first place. It /is/ a problem if such an URE occurs while you are already resyncing a missing disk - and is therefore a major motivation behind raid6 (and also Neil's "hotsync" plans). /Undetected/ read errors are when the disk controller reads errors from the disk surface, and the incorrect data passes the disk's RS and CRC checksums. The chances of this happening are absurdly small unless there are faults in the drive electronics or firmware (in which case all bets are off anyway). Higher level checksums are one way to detect such errors, as is regular data scrubbing. It is correct that there is always a chance of incorrect data getting through somewhere with undetected read errors. I don't have any figures on me, but I suspect that before these are a realistic worry then you have bigger concerns about memory bit errors going undetected despite ECC ram, and undetected network errors despite Ethernet and IP checksums.