Hi Ted, turns out that I've been wrong with my assumption that we can count with metadata reservations for the unwritten extent conversion. So the patch I've proposed: ext4: Reserve metadata if writing into uninitialized which I think you've already merged into the devel branch is not particularly useful and it should be reverted. Reverting this will not cause any problem to reappear and in fact it should get rid of some of the warnings about more metadata blocks being allocated than reserved. The problem is that even though we're able to reserve the metadata for the possible conversion, we can never know when to free the metadata reservation. Currently we're freeing the metadata reservation (which is usually overestimated because we never know for sure how much metadata will actually be needed) when the number of reserved data blocks drop to zero. However since we only need to reserve metadata (and not data) blocks when writing into unwritten extent, this approach does not work and we would free reserved metadata even if we might still need it - there is no way to tell whether we're going to need it or not in writeback path. However even after the revert we're still left with the problem with unwrittent extent conversion in dealloc path. I've attempted to solved it with similar mechanism which xfs has - block reserve pool. Blocks from this reserve pool should be removed from the global pool and no one would be able to allocate from it, unless: 1. we're in delalloc path and we need space for metadata allocation in unwritten extent conversion, This should solve the problem with not having enough metadata in ENOSPC condition in delayed allocation writeback path. 2. we're in punch hole path and we need space for metadata allocation when splitting extents This should solve the problem when we do not have _any_ space at all and we attempt to punch a hole into the file resulting in the need for new extent tree block. Punch hole should succeed even in ENOSPC conditions since we're actually freeing space. 3. we're writing into unwritten (preallocated) extent and we're in ENOSPC condition - not having any blocks to allocated metadata for unwritten extent conversion. This should solve the problem when we can return ENOSPC even when writing into preallocated space which is certainly unexpected (xfstest 274). I already have a patch which implements that, but it need some tweaks and some more testing, so I'll send it probably later this week, as well as more metadata reservation fixes (not related to unwritten extent conversion). Let me know what do you think. Thanks! -Lukas -- 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
