On Tue, Nov 28, 2023 at 12:11:33PM -0800, Darrick J. Wong wrote: > And now that I've dug further through my notes, I've realized that > there's a better reason for this unexplained _get_buf -> _read_buf > transition and the setting of XBF_DONE in _delwri_queue_here. > > This patch introduces the behavior that we flush the delwri list to disk > every 256k. Where "the delwri list" is the one used for writing stage btrees I think. > Flushing the buffers releases them, which means that > reclaim could free the buffer before xfs_btree_bload_node needs it again > to build the next level up. Oh, indeed. > If that's the case, then _get_buf will get > us a !DONE buffer with zeroes instead of reading the (freshly written) > buffer back in from disk. We'll then end up formatting garbage keys > into the node block, which is bad. Yeah. > /* > * Read the lower-level block in case the buffer for it has > * been reclaimed. LRU refs will be set on the block, which is > * desirable if the new btree commits. > */ > ret = xfs_btree_read_buf_block(cur, child_ptr, 0, &child_block, > &child_bp); > > The behavior of setting XBF_DONE in xfs_buf_delwri_queue_here is an > optimization if _delwri_submit releases the buffer and it is /not/ > reclaimed. In that case, xfs_btree_read_buf_block will find the buffer > without the DONE flag set and reread the contents from disk, which is > unnecessary. Yeah. I still find it weird to set it in the delwri_submit_here helper, but maybe that's a discussion for the other thread.
