On Fri, Oct 11, 2024 at 09:59:03AM +0200, Christoph Hellwig wrote: > On Thu, Oct 10, 2024 at 12:01:47PM -0700, Darrick J. Wong wrote: > > What if instead this took the form of a new defer_ops type? The > > xfs_prepare_sb_update function would allocate a tracking object where > > we'd pin the sb buffer and record which fields get changed, as well as > > the new values. xfs_commit_sb_update then xfs_defer_add()s it to the > > transaction and commits it. (The ->create_intent function would return > > NULL so that no log item is created.) > > > > The ->finish_item function would then bhold the sb buffer, update the > > ondisk super like how xfs_commit_sb_update does in this patch, set > > XFS_SB_TRANS_SYNC, and return -EAGAIN. The defer ops would commit and > > flush that transaction and call ->finish_item again, at which point it > > would recompute the incore/cached geometry as necessary, bwrite the sb > > buffer, and release it. > > > > The downside is that it's more complexity, but the upside is that the > > geometry changes are contained in one place instead of being scattered > > around, and the incore changes only happen if the synchronous > > transaction actually gets written to disk. IOWs, the end result is the > > same as what you propose here, but structured differently. > > That sounds overkill at first, but if we want to move all sb updates > to that model more strutured infrastructure might be very useful. <nod> We could just take this as-is and refactor it into the defer item code once we're done making all the other sb geometry growfsrt updates. I'd rather do that than rebase *two* entire patchsets just to get the same results. --D
