On Tue, Mar 8, 2011 at 10:28 PM, Andreas Dilger <adilger@xxxxxxxxx> wrote: > [I've removed LSF-PC from the CC list, since I believe the attendee/topic selection is long past] I think topics for discussion are always welcome (there is no sealed agenda for LSF) > > One reason that we didn't use this patch is that it is essentially the opposite of "ordered" mode when it comes to data reliability. It requires that the journal be updated with the blocks to be modified _before_ the data blocks are written. > > This means that stale data may be exposed to userspace in case of a crash, unless a two-phase commit is done where the declare blocks are committed to one transaction, and only after that the data blocks are written. I wouldn't go as far as calling it 2 separate transactions, but certainly, a three-phase commit is possible: 1. write declared and metadata blocks descriptors --- barrier/flush disk cache --- 2. write data blocks to disk and metadata blocks to journal --- barrier/flush disk cache --- 3. write commit record --- barrier/flush disk cache --- The journal recovery will guide RAID to re-sync the blocks in the "current" (uncommitted) transaction. > > The other option is a WAFL mode where writes are done into RAID stripe-aligned free space so that parity loss is not harmful to existing data. This is not unreasonable with mballoc for new blocks, but not necessarily overwrites. It may tie in with snapshots nicely, since it has the same COW requirements. Now this is too crazy even in my standards ;-) > > Cheers, Andreas > > On 2011-03-08, at 8:35, Amir Goldstein <amir73il@xxxxxxxxx> wrote: > >> Hi All, >> >> I have picked up the old ext3/jbd patches a while ago and >> was trying to figure out how difficult would it be to port them >> to ext4/jbd2. >> >> The gain from these patches to anyone using software RAID >> should be clear, see: >> http://lwn.net/Articles/363490/ >> >> What is not clear to me at this point is what are the >> performance implications, if any, of withholding "declared" data >> writeback until journal commit. >> >> I also did not get to estimating the effort involved in the porting >> to ext4/jbd2. The absence of ext3/jbd "dirty data" buffers list is >> going to be one of the issues to deal with. >> >> Does anyone out there have interest in pushing this forward? >> >> Amir. >> -- >> To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in >> the body of a message to majordomo@xxxxxxxxxxxxxxx >> More majordomo info at http://vger.kernel.org/majordomo-info.html > -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
