Hi all, Here's a second RFC of the mechanism to defer AGFL block frees. The purpose of this change is to minimize log reservation consumption (and thus reduce the likelihood of overruns) due to too many allocation operations being performed in a single transaction context. The major difference in this version is that rather than explicitly plumb an xfs_defer_ops structure to the allocator code from specific contexts where we want to defer AGFL block frees, we update struct xfs_trans to carry an optional dfops reference and use that to decide whether to defer such frees during AGFL fixups. This approach implies a few additional thoughts: - We eventually want to defer AGFL block frees by default (or at least from any code that already has a use for xfs_defer_ops). - With most transactions converted, we can facilitate further cleanups where we've already had to manually plumb xfs_defer_ops to places that already have access to the associated transaction. - Going further, we ultimately may be able to consider some further dfops interface cleanups, such as associating dfops to the tx via xfs_defer_init(), processing deferred ops at transaction commit time, rethinking how dfops memory is allocated, etc. If the broader ideas to defer AGFL block frees and to do so via refactoring the transaction/dfops relationship are acceptable, I'd like to take an incremental approach (starting with enabling deferred AGFL frees from the inode-related contexts modified in this series) to incorporate this change so we can enable the behavior in the contexts where it's known to be most effective without having to refactor the world all at once. With regard to testing, this series survives some basic xfstests testing so far. I've also run some fairly basic "performance" tests to try and assess how much of an effect this has on transaction regrants and whether or not it's worth bumping transaction log counts to cover regrants due to deferred AGFL frees. The final two patches are included for reference purposes and discuss this further. In summary, the testing I've done so far[1] suggests that bumping logcounts is probably overkill given the low frequency of AGFL block frees relative to all transaction commits. Thoughts, reviews, flames appreciated. Brian rfcv2: - Use transaction to carry deferred ops struct to allocation context. - Remove dfops param from dir interfaces where possible. - Defer AGFL block frees from more contexts. - Define runtime stat for transaction regrant and logcount patch (to be potentially removed). rfcv1: https://marc.info/?l=linux-xfs&m=151267309423883&w=2 [1] Some notes around regrant testing using xs_regrant_logspace. * Sustained file removal test (known to reproduce log res. overruns): - for-next - 6641554 - deferred AGFL frees - 6742268 - deferred AGFL w/ tr_inactive - 6715692 - deferred AGFL w/ inactive+truncate - 74138 Deferring AGFL block frees increases the regrant count by about ~1.5% in this test. Upping the tr_inactive logcount reduced this by 0.4%. The significant effect from the bump of the tr_truncate logcount shows that this workload is dominated by bmap extent frees, so I don't think this particular number is relevant to the patchset. * fsstress test against a 15GB fs (runs for ~5m and consumes ~9GB of available space, repeated a couple runs, all using the same seed value): - for-next - 35469, 35616 - deferred AGFL frees - 35446, 35445 - deferred AGFL w/ patch 9 (AGFL count == 1) - 35510, 35412 This short, but more mixed workload doesn't show much of a difference between the baseline and deferred AGFL free code (in fact the counts drop, which suggests this is all just noise). Brian Foster (9): xfs: create agfl block free helper function xfs: allow attach of dfops to transaction xfs: defer agfl block frees when dfops is available xfs: defer agfl block frees from deferred ops processing context xfs: defer agfl frees from inode inactivation xfs: defer frees from common inode allocation paths xfs: defer agfl frees from directory op transactions xfs: add a runtime stat for extra transaction log regrants xfs: add extra log count for transactions that defer agfl frees fs/xfs/libxfs/xfs_alloc.c | 82 ++++++++++++++++++++++++++++++++++++------ fs/xfs/libxfs/xfs_alloc.h | 2 ++ fs/xfs/libxfs/xfs_defer.c | 9 +++++ fs/xfs/libxfs/xfs_defer.h | 1 + fs/xfs/libxfs/xfs_dir2.c | 16 ++++----- fs/xfs/libxfs/xfs_dir2.h | 9 ++--- fs/xfs/libxfs/xfs_ialloc.c | 6 ++-- fs/xfs/libxfs/xfs_ialloc.h | 1 - fs/xfs/libxfs/xfs_trans_resv.h | 19 +++++----- fs/xfs/xfs_inode.c | 69 ++++++++++++++++++----------------- fs/xfs/xfs_inode.h | 3 +- fs/xfs/xfs_log.c | 1 + fs/xfs/xfs_stats.h | 3 +- fs/xfs/xfs_symlink.c | 3 +- fs/xfs/xfs_trace.h | 2 ++ fs/xfs/xfs_trans.c | 7 ++-- fs/xfs/xfs_trans.h | 1 + fs/xfs/xfs_trans_extfree.c | 70 ++++++++++++++++++++++++++++++++++++ 18 files changed, 227 insertions(+), 77 deletions(-) -- 2.13.6 -- To unsubscribe from this list: send the line "unsubscribe linux-xfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
