This is the next round of my reserve ticket refinements. Most of the changes are just fixing issues brought up by review. The updated diffstat is as follows
fs/btrfs/block-group.c | 5 +- fs/btrfs/block-rsv.c | 10 +-- fs/btrfs/delalloc-space.c | 4 -- fs/btrfs/delayed-ref.c | 2 +- fs/btrfs/extent-tree.c | 13 +--- fs/btrfs/space-info.c | 171 +++++++++++++++++++--------------------------- fs/btrfs/space-info.h | 30 +++++--- 7 files changed, 98 insertions(+), 137 deletions(-) v2->v3: - added 9/9 to rename btrfs_space_info_add_old_bytes as per discussions with Nikolay. - added a few comments. - made the logic clearer in the may_commit_transaction patch. - a few lockdep_assert_held()'s. - added the reviewed-by's. v1->v2: - added "btrfs: fix may_commit_transaction to deal with no partial filling" - fixed "btrfs: refactor the ticket wakeup code" to return true if we find a smaller ticket than our first ticket in the list. - Original email - While cleaning up some things around the global reserve and can_overcommit I started getting ENOSPC's with plenty of space to make reservations. The root cause of the problem has to do with how we satisfy ticket reservations. Previously we would add any space we were returning to the space info to the first ticket we found. The reason we did this was because new reservations just check the counters to see if they can continue, so we didn't want them to get reservations when we had waiters already queued up. So instead of returning the bytes to the space info, I'd add it to the ticket. Then if we failed to satisfy that ticket reservation we'd take any space we found and add it to the next guy in case it satisfied the next ticket reservation. This works generally well in practice, but there are several xfstests that run ENOSPC tests against very small file systems. These tests uncovered a corner case when it comes to overcommitting. If we overcommit the space, and then are no longer allowed to overcommit, we won't actually give any returned space to the tickets, because that would be really bad. Instead we return that space to the space_info and carry on. What was biting us in these test cases was the fact that we had very small metadata area, 8mib, and unlink asks for about 2mib of space. If we had overcommitted 8.1mib, we'd give back almost 2mib of space to the space_info, which could have instead been used for the reservation. This would result in an early ENOSPC. Since we are only doing this partial filling dance to avoid racing with new reservations we just fix that race by checking if we have pending reservations on the list, closing that race. Then we are free to use the normal checks to see if a ticket can be woken up. This simplifies the code a bunch, we no longer have to keep track of how much space the tickets were given and return those bytes, and I could consolidate the wakeup code into one function instead of two. The diffstat is as follows, this all passes xfstests, and sets us up nicely for the upcoming changesets. Thanks, Josef
