On Wed, Mar 18, 2009 at 02:12:15PM +0100, Jan Kara wrote: > > Wouldn't this cause a deadlock ? We want to commit a transaction because > > we don't have enough journal space (via journal_start) and now that would cause block > > allocation which would do another journal_start() > Yes, that's exactly what I think. We cannot start a transaction while > committing another transaction. Also you must put the block allocation > into the transaction you're going to commit because of data consistency > guarantees. > So if you want to do "alloc on commit" you have to reserve enough > credits to the running transaction at the "block reservation" time and > then use them for allocation at commit time. But this gets complex > because the number of needed credits is hard to estimate (we don't know how > many bitmaps / group descriptors we're going to modify). I'm not yet > sure how to solve this problem... Yeah, agreed, this is going to get tricky. What we would have to do is estimate a worst case, and include that in the running tally, and then subtract it off when we start allocating the data blocks. But the problem then is what happens to new file system operations? If we stall them, it will be a major performance hit. We can't let them start a new transaction, because we can't have to open transactions at the same time. If we let them continue to run against the current transaction, then #1, we could run out of space (although the we give ourselves 25% of the journal as "slop" space which is extremely generous), and #2, there is a race where the new file system operations that do delayed allocation won't get allocated on the commit. So this is not going to be an easy problem to solve, not without massively complicating the jbd2 layer... - Ted -- 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
