Hi, At Thu, 06 May 2010 03:23:27 +0900 (JST), Ryusuke Konishi wrote: > > On Thu, 06 May 2010 00:10:53 +0900, Jiro SEKIBA wrote: > > Hi, > > > > OK, it looks like a flag should be introduced for older flash device > > not to write super block whenever checkpoint is created, but maintain > > it less frequently or even not writing at all until unmount. > > > > In any cases, at least super block disk format is not modified. > > > > At Tue, 04 May 2010 23:01:35 +0900 (JST), > > Ryusuke Konishi wrote: > > > > > > On Tue, 4 May 2010 14:02:22 +0200, Reinoud Zandijk wrote: > > > > On Mon, May 03, 2010 at 03:35:03PM -0400, Jay Carlson wrote: > > > > > On May 3, 2010, at 11:54 AM, Jiro SEKIBA wrote: > > > > > Low-end consumer flash (like USB thumb drives or MMC) often uses a simple > > > > > zone-based FTL with bad block replacement. I am told really cheap ones > > > > > allocate zones with 1000 flash blocks with 24 held as replacements for > > > > > failing blocks and wear leveling. > > > > > > > > > > If I understand nilfs, its superblock is both fixed location and "hot": it > > > > > is written fairly frequently (on every fsync?) On cheap flash, it will wear > > > > > out the flash block it lives on in 25 write lifetimes or less. > > > > > > > > I'd opt for NOT writing out the super block but on unmount or when the > > > > roll-forward chain is disturbed by the garbage collector.i > > > > > > > > Another option is to never update it; it should at most take a few secs to > > > > locate the latest segsum by just scanning trough the segement summaries; > > > > especially now the segment summaries have the checkpoint number incorporated. > > > > > > > > If on mount all first segment summaries are read (say 4000 to 8000 sectors) > > > > its clear wich is the newest and then follow that chain until you reach the > > > > end... and you can mount. I agree its not optimal but i dont see a reason as > > > > to why it shouldn't work :) > > > > > > I was just thinking the same thing. > > > > > > The reason nilfs frequently updates super blocks is for maintaining a > > > pointer to recent logs. And, the new checkpoint number field allows > > > it to find them by scanning through summary headers of each segments. > > > > > > This may be expensive for hard drives, but may be acceptable for flash > > > devices. I think it's worth adding a new mount option (or a flag) for > > > this. > > > > This leads that boot loader is required to take care of searching > > valid super root by the scanning all the segments, unlike just > > roll forwarding the log. > > > > From boot loader (grub2) point of view, any access to the filesystem > > requires "mount" operation, which means leading two files (initrd and kernel) > > requires at least two mount operations. That means in case of unclean unmount, > > it requires to scan whole disks twice. > > > > So I opt to write super block less frequently, to maintain super root > > pointer so that roll forwarding likely finds correct super root. > > > > Or maybe specifying frequency as mount option, say like write back > > per 100 check points. > > > > thanks, > > > > regards, > > As an alternative, I'm thinking to add a new state flag which > indicates segments are allocated physically continuously from > the super root to which super blocks point. > > > The aim of this flag is allowing nilfs to find out the latest segment > with bisection search. > > Envisioned changes are as follows: > > * add the new flag (for example, NILFS_INORDER_FS) for sbp->s_state. > > * set the flag if a new mount options is specified (for example > "-o bisect-root"). > > * Do not update super block when the filesystem is unmounted, and keep > the state: s_state.NILFS_INORDER_FS = 1, s_state.NILFS_VALID_FS = 0. > > * Stop periodic update of super blocks if the flag is set. > > * If s_state.NILFS_VALID_FS = 0 && s_state.NILFS_INORDER_FS = 1 when > the filesystem is mounted, then do bisect search to find out the > latest segment. Here are the question, to do bisect search, you need to know the end of the segments to divide physically continuous blocks. Which is the end of the block for bisect? Is this going to be the physical partition end? > * If a new segment is allocated discontinuously and a new super root > is created, then write out super blocks to catch up the position. > > * If GC breaks the series of physically continous segments, then > update super blocks to catch up the latest super root position. > > * Add a new option to cleanerd to pass the "-o bisect-root" mode > and let it select the rotational GC algorithm (current default). > > * If the "-o bisect-root" option is not specified, then use a > conventional algorithm and drop the NILFS_INORDER_FS flag. I prefer option name like "-o async_sb" or "-o no_sync_sb" or so. Because users would rather curious how to udpate super block than how to find the latest log. > This doesn't break forward compatibility, because the "next segment > chain" is still maintained and older implementations will ignore the > NILFS_INORDER_FS flag. Older implementations and the current grub2 > module can find the latest super root though they incur penalty for > mount time. > > Another drawback of this approach is that it depends on garbage > collection algorithm, but at least at present, this seems not to > matter. > > How does that sound? It would be much better than scanning whole segments from boot loader point of view. > Thanks, > Ryusuke Konishi > -- > To unsubscribe from this list: send the line "unsubscribe linux-nilfs" in > the body of a message to majordomo@xxxxxxxxxxxxxxx > More majordomo info at http://vger.kernel.org/majordomo-info.html > > > -- Jiro SEKIBA <jir@xxxxxxxxx> -- To unsubscribe from this list: send the line "unsubscribe linux-nilfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
