[Ocfs2-devel] [PATCH 1/1] Ocfs2: Optimize punching-hole codes v3.
Tristan Ye
tristan.ye at oracle.com
Mon Mar 29 22:49:46 PDT 2010
Changes from v2 to v3:
1. Fix a bug when crossing extent blocks.
2. Apply joel's comments to wrap loose codes as appropriate helpers.
V3 patch has survived with fill_verify_holes testcase in ocfs2-test,
it also passed my manual sanity check and stress tests with enormous
extent records, besides, more corner testcases succeeded than v2.
Currently punching hole on a file with 3+ extent tree depth was
really a performance disaster, it even caused several hours to
go, though we may not hit this in real life with such a huge extent
number.
One simple way to improve the performance is quite straightforward,
by learning the logic of truncating codes, means we'd punch hole from
hole_end to hole_start, which reduce the overhead of btree operation
in a significant way, such as tree rotation and moving.
Following is the testing result when punching hole from 0 to file end
in bytes, on a 1G file, 1G file consists of 256k extent records, each record
cover 4k data(just one cluster, clustersize is 4k):
===========================================================================
* Former punching-hole mechanism:
===========================================================================
I waited 1 hour for its completion, unfortunately it's still ongoing.
===========================================================================
* Patched punching-hode mechanism:
===========================================================================
real 0m2.518s
user 0m0.000s
sys 0m2.445s
That means we've gained up to 1000 times improvement on performance in this
case, whee! It's fairly cool. and it looks like that performance gain will
be raising when extent records grow.
The patch was based on my former 2 patches, which were about truncating
codes optimization and fixup to handle CoW on punching hole.
Signed-off-by: Tristan Ye <tristan.ye at oracle.com>
---
fs/ocfs2/file.c | 192 +++++++++++++++++++++++++++++++++++++++++++++++--------
1 files changed, 164 insertions(+), 28 deletions(-)
diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c
index db2e0c9..5ccfc7e 100644
--- a/fs/ocfs2/file.c
+++ b/fs/ocfs2/file.c
@@ -1423,18 +1423,124 @@ out:
return ret;
}
+/*
+ * Hepler to find the rightmost record which contains 'pos' cpos,
+ * skip the holes if any, also adjust the 'pos' accordingly.
+ */
+static void ocfs2_find_rec_with_pos(struct ocfs2_extent_list *el,
+ struct ocfs2_extent_rec **rec_got,
+ u32 *pos)
+{
+ int i;
+ u32 range;
+ struct ocfs2_extent_rec *rec = NULL;
+
+ for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
+
+ rec = &el->l_recs[i];
+ range = le32_to_cpu(rec->e_cpos) +
+ ocfs2_rec_clusters(el, rec);
+
+ if (le32_to_cpu(rec->e_cpos) < *pos) {
+ /*
+ * Skip a hole.
+ */
+ if (range < *pos)
+ *pos = range;
+
+ break;
+ }
+
+ /*
+ * Simply jump to previous record if the pos is
+ * the start of a record.
+ */
+ if (le32_to_cpu(rec->e_cpos) == *pos) {
+ i--;
+ if (i < 0) {
+ *rec_got = NULL;
+ return;
+ }
+
+ rec = &el->l_recs[i];
+ range = le32_to_cpu(rec->e_cpos) +
+ ocfs2_rec_clusters(el, rec);
+ /*
+ * Skip a hole.
+ */
+ if (range < *pos)
+ *pos = range;
+
+ break;
+ }
+ }
+
+ *rec_got = &el->l_recs[i];
+}
+
+/*
+ * Helper to calculate the punching pos and length in one run, we handle the
+ * following three cases in order:
+ *
+ * - remove the entire record
+ * - remove a partial record
+ * - no record needs to be removed (hole-punching completed)
+*/
+static void ocfs2_calc_trunc_pos(struct inode *inode,
+ struct ocfs2_extent_list *el,
+ struct ocfs2_extent_rec *rec,
+ u32 trunc_start, u32 *trunc_cpos,
+ u32 *trunc_len, u32 *trunc_end,
+ u64 *blkno, int *done)
+{
+ int ret = 0;
+ u32 coff, range;
+
+ range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
+
+ if (le32_to_cpu(rec->e_cpos) >= trunc_start) {
+ *trunc_cpos = le32_to_cpu(rec->e_cpos);
+ *trunc_len = *trunc_end - le32_to_cpu(rec->e_cpos);
+ *blkno = le64_to_cpu(rec->e_blkno);
+ *trunc_end = le32_to_cpu(rec->e_cpos);
+ } else if (range > trunc_start) {
+ *trunc_cpos = trunc_start;
+ *trunc_len = range - trunc_start;
+ coff = trunc_start - le32_to_cpu(rec->e_cpos);
+ *blkno = le64_to_cpu(rec->e_blkno) +
+ ocfs2_clusters_to_blocks(inode->i_sb, coff);
+ *trunc_end = trunc_start;
+ } else {
+ /*
+ * It may have two following possibilities:
+ *
+ * - last record has been removed
+ * - trunc_start was within a hole
+ *
+ * both two cases mean the completion of hole punching.
+ */
+ ret = 1;
+ }
+
+ *done = ret;
+}
+
static int ocfs2_remove_inode_range(struct inode *inode,
struct buffer_head *di_bh, u64 byte_start,
u64 byte_len)
{
- int ret = 0, flags = 0;
- u32 trunc_start, trunc_len, cpos, phys_cpos, alloc_size;
+ int ret = 0, flags = 0, done = 0;
+ u32 trunc_start, trunc_len, trunc_end, trunc_cpos, phys_cpos;
+ u32 cluster_within_list;
struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
struct ocfs2_cached_dealloc_ctxt dealloc;
struct address_space *mapping = inode->i_mapping;
struct ocfs2_extent_tree et;
+ struct ocfs2_path *path = NULL;
+ struct ocfs2_extent_list *el = NULL;
+ struct ocfs2_extent_rec *rec = NULL;
struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
- u64 refcount_loc = le64_to_cpu(di->i_refcount_loc);
+ u64 blkno, refcount_loc = le64_to_cpu(di->i_refcount_loc);
ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
ocfs2_init_dealloc_ctxt(&dealloc);
@@ -1482,16 +1588,13 @@ static int ocfs2_remove_inode_range(struct inode *inode,
}
trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
- trunc_len = (byte_start + byte_len) >> osb->s_clustersize_bits;
- if (trunc_len >= trunc_start)
- trunc_len -= trunc_start;
- else
- trunc_len = 0;
+ trunc_end = (byte_start + byte_len) >> osb->s_clustersize_bits;
+ cluster_within_list = trunc_end;
- mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u\n",
+ mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, cend: %u\n",
(unsigned long long)OCFS2_I(inode)->ip_blkno,
(unsigned long long)byte_start,
- (unsigned long long)byte_len, trunc_start, trunc_len);
+ (unsigned long long)byte_len, trunc_start, trunc_end);
ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
if (ret) {
@@ -1499,32 +1602,65 @@ static int ocfs2_remove_inode_range(struct inode *inode,
goto out;
}
- cpos = trunc_start;
- while (trunc_len) {
- ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
- &alloc_size, &flags);
+ path = ocfs2_new_path_from_et(&et);
+ if (!path) {
+ ret = -ENOMEM;
+ mlog_errno(ret);
+ goto out;
+ }
+
+ while (trunc_end > 0) {
+ /*
+ * Unlike truncate codes, here we want to find a path which
+ * contains (trunc_end - 1) cpos, and then trunc_end will be
+ * decreased after each removal of a record range.
+ *
+ * Why not using trunc_end to search the path?
+ * The reason is simple, think about the situation of crossing
+ * the extent block, we need to find the adjacent block by
+ * decreasing one cluster, otherwise, it will run into a loop.
+ */
+ ret = ocfs2_find_path(INODE_CACHE(inode), path,
+ cluster_within_list);
if (ret) {
mlog_errno(ret);
goto out;
}
- if (alloc_size > trunc_len)
- alloc_size = trunc_len;
+ el = path_leaf_el(path);
- /* Only do work for non-holes */
- if (phys_cpos != 0) {
- ret = ocfs2_remove_btree_range(inode, &et, cpos,
- phys_cpos, alloc_size,
- &dealloc, refcount_loc,
- flags);
- if (ret) {
- mlog_errno(ret);
- goto out;
- }
+ ocfs2_find_rec_with_pos(el, &rec, &trunc_end);
+ /*
+ * Go to next extent block directly when last rec was truncated.
+ */
+ if (!rec) {
+ cluster_within_list = trunc_end - 1;
+ ocfs2_reinit_path(path, 1);
+ break;
}
- cpos += alloc_size;
- trunc_len -= alloc_size;
+ ocfs2_calc_trunc_pos(inode, el, rec, trunc_start, &trunc_cpos,
+ &trunc_len, &trunc_end, &blkno, &done);
+ if (done)
+ break;
+
+ flags = rec->e_flags;
+ phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
+
+ ret = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
+ phys_cpos, trunc_len, &dealloc,
+ refcount_loc, flags);
+ if (ret < 0) {
+ mlog_errno(ret);
+ goto out;
+ }
+
+ if (trunc_end > 0)
+ cluster_within_list = trunc_end - 1;
+ else
+ break;
+
+ ocfs2_reinit_path(path, 1);
}
ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
--
1.5.5
More information about the Ocfs2-devel
mailing list