On May 08, 2008 22:18 -0400, Theodore Ts'o wrote: > On Thu, May 08, 2008 at 04:48:47PM -0600, Andreas Dilger wrote: > > I just noticed lazy_itable_init in the mke2fs.8.in man page. I think > > a warning needs to be added there that this is not currently safe to > > use, because the kernel does not yet do the background zeroing. There > > is nothing in the man page to indicate that this is unsafe... > > Yeah, I was hoping we would actually get this fixed before 1.41 was > released.... (i.e., implement the background zeroing). It would still be an issue for 2.6.24 and 2.6.25 kernels, so I think it at least deserves a warning until there is a specific kernel that can be referenced that has this functionality. > One of the > things I was thinking about was whether we could avoid needing to go > through the jbd layer when zeroing out an entire inode table block, > and then in the completion callback function when the block group was > completely initiaized, we could clear the ITABLE_UNINIT flag. > > It doesn't need to go through the journal, because if we crash without > having the flag set, its not a big deal; the inode table will just not > be marked initialized. The only thing which might require a little > care is if buffer head referencing part of the inode table which is > getting zero'ed out is in flight when an inode allocation happens, an > inode gets marked dirty, and fs/ext4/inode.c wants to write out an > inode table block that is in the middle of being zero'ed. Given that > we've bypassed the jbd layer for efficiency's sake, something bad > could happy unless we protect it with some kind of lock. > > Or we could just say that this initialization pass is relatively rare, > so do it the cheap cheasy way, even if the blocks end up going through > the journal. The upside is that it should be pretty quick and easy to > code it this way. It is only a once-per-filesystem-lifetime operation, and while it is a fair amount of IO, it could be done efficient because of sequential IO. I believe that the unwritten extent code added a function "ext4_zero_blocks" or similar that maps the ZERO_PAGE into a bio and submits that for IO to zero out the extent. This could be used for the inode tables also, avoiding the major problem of dirtying thousands of blocks in memory. The risk, as you write, is the locking, and the only locks I see on the inode table are the bh locks on the itable blocks in __ext4_get_inode_loc(). We can't hold the per-group hashed spinlock for this IO. We might consider adding an rw semaphore, and in the very common case there will only ever be readers on this lock. If there is a zeroing happening then there can be a "trylock" operation and the entire group skipped to avoid blocking the caller for a long time. Cheers, Andreas -- Andreas Dilger Sr. Staff Engineer, Lustre Group Sun Microsystems of Canada, Inc. -- To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html