Give a brief overview for how async discard is implemented. Signed-off-by: Dennis Zhou <dennis@xxxxxxxxxx> Reviewed-by: Josef Bacik <josef@xxxxxxxxxxxxxx> --- fs/btrfs/discard.c | 34 ++++++++++++++++++++++++++++++++++ 1 file changed, 34 insertions(+) diff --git a/fs/btrfs/discard.c b/fs/btrfs/discard.c index f95e437d7629..2ff284a8a760 100644 --- a/fs/btrfs/discard.c +++ b/fs/btrfs/discard.c @@ -1,5 +1,39 @@ /* * Copyright (C) 2019 Facebook. All rights reserved. + * + * This contains the logic to handle async discard. + * + * Async discard manages trimming of free space outside of transaction commit. + * Discarding is done by managing the block_groups on a LRU list based on free + * space recency. Two passes are used to first prioritize discarding extents + * and then allow for trimming in the bitmap the best opportunity to coalesce. + * The block_groups are maintained on multiple lists to allow for multiple + * passes with different discard filter requirements. A delayed work item is + * used to manage discarding with timeout determined by a max of the delay + * incurred by the iops rate limit, byte rate limit, and the timeout of max + * delay of BTRFS_DISCARD_MAX_DELAY. + * + * The first list is special to manage discarding of fully free block groups. + * This is necessary because we issue a final trim for a full free block group + * after forgetting it. When a block group becomes unused, instead of directly + * being added to the unused_bgs list, we add it to this first list. Then + * from there, if it becomes fully discarded, we place it onto the unused_bgs + * list. + * + * The in-memory free space cache serves as the backing state for discard. + * Consequently this means there is no persistence. We opt to load all the + * block groups in as not discarded, so the mount case degenerates to the + * crashing case. + * + * As the free space cache uses bitmaps, there exists a tradeoff between + * ease/efficiency for find_free_extent() and the accuracy of discard state. + * Here we opt to let untrimmed regions merge with everything while only letting + * trimmed regions merge with other trimmed regions. This can cause + * overtrimming, but the coalescing benefit seems to be worth it. Additionally, + * bitmap state is tracked as a whole. If we're able to fully trim a bitmap, + * the trimmed flag is set on the bitmap. Otherwise, if an allocation comes in, + * this resets the state and we will retry trimming the whole bitmap. This is a + * tradeoff between discard state accuracy and the cost of accounting. */ #include <linux/jiffies.h> -- 2.17.1
