1 files changed, 39 insertions, 21 deletions

diff --git a/fs/btrfs/file-item.c b/fs/btrfs/file-item.c
index c09fbc257634..a42e6d54e7cd 100644
--- a/fs/btrfs/file-item.c
+++ b/fs/btrfs/file-item.c
@@ -397,6 +397,36 @@ int btrfs_lookup_bio_sums(struct btrfs_bio *bbio)
 		path->skip_locking = 1;
 	}
 
+	/*
+	 * If we are searching for a csum of an extent from a past
+	 * transaction, we can search in the commit root and reduce
+	 * lock contention on the csum tree extent buffers.
+	 *
+	 * This is important because that lock is an rwsem which gets
+	 * pretty heavy write load under memory pressure and sustained
+	 * csum overwrites, unlike the commit_root_sem. (Memory pressure
+	 * makes us writeback the nodes multiple times per transaction,
+	 * which makes us cow them each time, taking the write lock.)
+	 *
+	 * Due to how rwsem is implemented, there is a possible
+	 * priority inversion where the readers holding the lock don't
+	 * get scheduled (say they're in a cgroup stuck in heavy reclaim)
+	 * which then blocks writers, including transaction commit. By
+	 * using a semaphore with fewer writers (only a commit switching
+	 * the roots), we make this issue less likely.
+	 *
+	 * Note that we don't rely on btrfs_search_slot to lock the
+	 * commit root csum. We call search_slot multiple times, which would
+	 * create a potential race where a commit comes in between searches
+	 * while we are not holding the commit_root_sem, and we get csums
+	 * from across transactions.
+	 */
+	if (bbio->csum_search_commit_root) {
+		path->search_commit_root = 1;
+		path->skip_locking = 1;
+		down_read(&fs_info->commit_root_sem);
+	}
+
 	while (bio_offset < orig_len) {
 		int count;
 		u64 cur_disk_bytenr = orig_disk_bytenr + bio_offset;
@@ -442,6 +472,8 @@ int btrfs_lookup_bio_sums(struct btrfs_bio *bbio)
 		bio_offset += count * sectorsize;
 	}
 
+	if (bbio->csum_search_commit_root)
+		up_read(&fs_info->commit_root_sem);
 	return ret;
 }
 
@@ -743,12 +775,10 @@ int btrfs_csum_one_bio(struct btrfs_bio *bbio)
 	SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
 	struct bio *bio = &bbio->bio;
 	struct btrfs_ordered_sum *sums;
-	char *data;
-	struct bvec_iter iter;
-	struct bio_vec bvec;
+	struct bvec_iter iter = bio->bi_iter;
+	phys_addr_t paddr;
+	const u32 blocksize = fs_info->sectorsize;
 	int index;
-	unsigned int blockcount;
-	int i;
 	unsigned nofs_flag;
 
 	nofs_flag = memalloc_nofs_save();
@@ -767,21 +797,9 @@ int btrfs_csum_one_bio(struct btrfs_bio *bbio)
 
 	shash->tfm = fs_info->csum_shash;
 
-	bio_for_each_segment(bvec, bio, iter) {
-		blockcount = BTRFS_BYTES_TO_BLKS(fs_info,
-						 bvec.bv_len + fs_info->sectorsize
-						 - 1);
-
-		for (i = 0; i < blockcount; i++) {
-			data = bvec_kmap_local(&bvec);
-			crypto_shash_digest(shash,
-					    data + (i * fs_info->sectorsize),
-					    fs_info->sectorsize,
-					    sums->sums + index);
-			kunmap_local(data);
-			index += fs_info->csum_size;
-		}
-
+	btrfs_bio_for_each_block(paddr, bio, &iter, blocksize) {
+		btrfs_calculate_block_csum(fs_info, paddr, sums->sums + index);
+		index += fs_info->csum_size;
 	}
 
 	bbio->sums = sums;
@@ -993,7 +1011,7 @@ int btrfs_del_csums(struct btrfs_trans_handle *trans,
 			 * item changed size or key
 			 */
 			ret = btrfs_split_item(trans, root, path, &key, offset);
-			if (ret && ret != -EAGAIN) {
+			if (unlikely(ret && ret != -EAGAIN)) {
 				btrfs_abort_transaction(trans, ret);
 				break;
 			}