See below [HM]. On Wed, Apr 02, 2008 at 04:23:41PM -0400, Mikulas Patocka wrote: > Hi > > Unfortunatelly, the bug with desychnronizing raid1 that someone pointed out > on Monday, is real. The bug happens when you modify the page while its > being written to raid1 device --- old version can be written to one mirror > leg, the new versions to the other mirror leg. Raid1 code does not notice > this, marks the region clean after the writes finish, and the volume stays > desynchronized. > > The possibilities, how data can be modified while they are being written. > > 1. an application does O_DIRECT IO and modifies the memory underway. > > --- this is a problem of the application and we don't have to care about > it. > > 2. an application maps file for writing. pdflush or kswapd daemon writes > the page on background while the application is modifying it. > > 3. an application writes to a page with write() syscall. This syscall can > race with pdflush or kswapd as well. > > 4. a filesystem modifies the buffer while its being written by pdflush or > kswapd daemons. > > > The pdflush and kswapd daemons run in background and do periodic writes of > the modified data. pdflush is triggered regularly and writes data in > specified interval (about 30 seconds), so that in case of crash, the image > on disk is not too old. kswapd is triggered when the free memory goes low > --- it writes file pages and filesystem buffers too. > > In cases 2,3,4 the data may be modified while they are being written, but > the kernel writes them later again. The sequence is something like: > clear dirty bit > submit IO > --- if the data are modified while the IO is in progress, the dirty bit is > turned on again and the data will be written later and possible data > corruption is corrected. --- so as long as the system does not crash, there > can't be desynchronized mirror. > > But if the system crashes before the data are written second time, the > blocks may stay desynchronized. > > An example of data corruption on ext2: > > We have a dirty bitmap buffer > Pdflush clears the dirty flag and starts writing the buffer > The write is submitted to dm-raid1, it makes two requests and submits them > to two mirror devices > > This operation races with another thread allocating a block on ext2 and > doing: [HM] And taking out a copy unlocked in the RAID driver ain't help application data integrity, because it could still change the data while the RAID driver is copying, hence leading to coherency on the RAID set but holding incorrect application data. One can argue that this is ok in case of a crash, because the application failed to flush any page changes and hence has to be capable to recover from this. We always will end up with consistent mirrors (either on multiplicated successful writes to all legs or after resynchronization of the mirror) at the cost of internal caching of pages. > ext2_new_blocks > calling read_block_bitmap > calling sb_getblk > calling bh_uptodate_or_lock --- sees that the buffer is uptodate (even if > it's under write), so it returns. > calling ext2_try_to_allocate_with_rsv > calling ext2_try_to_allocate > calling ext2_set_bit_atomic --- this modifies the bitmap > *** now suppose that 2nd mirror device already finished > its write and don't get updated bit, while 1st mirror > device writes the updated bit to disk. > calling mark_buffer_dirty --- this schedules new update of the buffer > (after several seconds) > > Both writes finished, dm-raid1 driver turns off the dirty bit for the > region. > > Before pdflush writes the buffer second time, we get a > ***CRASH*** > > After new boot, dm-raid1 doesn't update the region, because the region's > bit is off. fsck scans the device. It reads the bitmap from the first > device, sees that the bit is correctly set and doesn't write the bitmap. > > Some times later, the administrator removes the 1st disk, the kernel starts > reading from 2nd mirror. Ext2 allocates another file, it reads the bitmap > from the 2nd device, sees the bit is off and allocates another block there. > Now there is data corruption => two files pointing to the same block. > > > Ideas how to fix it: > > 1. lock the buffers and unmap the pages while they are being written. > --- upstream developers would likely reject it. No other driver than > dm-raid1 has problems with this and they wouldn't damp performance because > of one driver. [HM] md RAID456 and dm RAID45 don't have the raid1 problem, because they utilize stripe caches, hence tacking page copies. Application pages can change nonetheless vs. stripe cache pages. > > 2. never turn the region dirty bit off until the filesystem is unmounted. > --- simplest fix. If the computer crashes after a long time, it > resynchronizes the whole device. md-raid resynchronizes the whole device > after a crash too. [HM] We wouldn't resync the whole device, just dirty regions. Of course the whole device would be the worst case with a huge write data set. For obvious reasons this is not what we want performamce-wise... > > 3. turn off the bit if the block wasn't written in one pdflush period > --- requires an interaction with pdflush, rather complex, I wouldn't > recommend it. > > 4. make more region states. > --- If the region is in RH_DIRTY state and all writes drain, the state is > changed to RH_MAYBE_DIRTY. (we don't know if the region is synchronized or > not). The disk dirty flag is kept. > --- periodically (once in few minutes, so that it doesn't affect > performance much), the change all regions in RH_MAYBE_DIRTY state to > RH_CLEAN_CANDIDATE, then issue sync() on all filesystems. If, after the > sync(), the region is still in RH_CLEAN_CANDIDATE (i.e. it hasn't been > written during the sync()), it is moved to RH_CLEAN state and the on-disk > bit for the region is turned off. [HM] This is essentially one technical approach for my comment on 2. above. RH_MAYBE_DIRTY sounds superfluous at first glance, because when all writes to a region drained, we can set RH_CLEAN_CANDIDATE, run the sync() and check if that state persists in order to trigger the dirty log update. Heinz > > If one of the above scenarios 2,3,4 happened (modifying a buffer while it's > under the disk write), the the sync() would have written the buffer again > and kicked the region out of RH_CLEAN_CANDIDATE state. If the sync() didn't > touch the buffer than we are sure that both on-disk copies are > synchronized. > > > Do you have any other ideas on this? > > Mikulas > > -- > dm-devel mailing list > dm-devel@xxxxxxxxxx > https://www.redhat.com/mailman/listinfo/dm-devel -- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel
