Re: [RFC] improve malloc for large filesystems

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 




> On 20 Nov 2019, at 21:13, Theodore Y. Ts'o <tytso@xxxxxxx> wrote:
> 
> Hi Alex,
> 
> A couple of comments.  First, please separate this patch so that these
> two separate pieces of functionality can be reviewed and tested
> separately:
> 

And this is the second one

Thanks, Alex

>From d2ff76d76320ade5f53002aa522b6eccfa058d47 Mon Sep 17 00:00:00 2001
From: Alex Zhuravlev <bzzz@xxxxxxxxxxxxx>
Date: Thu, 21 Nov 2019 10:00:07 +0300
Subject: [PATCH 2/2] ext4: prefetch bitmaps during block allocation

when the cache is cold reading bitmaps one by one can slowdown
the process significantly, especially on legacy rotating drives.
---
 fs/ext4/balloc.c  | 12 ++++++++++--
 fs/ext4/ext4.h    |  4 +++-
 fs/ext4/mballoc.c | 50 ++++++++++++++++++++++++++++++++++++++++++++++-
 fs/ext4/mballoc.h |  1 +
 fs/ext4/sysfs.c   |  2 ++
 5 files changed, 65 insertions(+), 4 deletions(-)

diff --git a/fs/ext4/balloc.c b/fs/ext4/balloc.c
index 0b202e00d93f..76547601384b 100644
--- a/fs/ext4/balloc.c
+++ b/fs/ext4/balloc.c
@@ -404,7 +404,8 @@ static int ext4_validate_block_bitmap(struct super_block *sb,
  * Return buffer_head on success or NULL in case of failure.
  */
 struct buffer_head *
-ext4_read_block_bitmap_nowait(struct super_block *sb, ext4_group_t block_group)
+ext4_read_block_bitmap_nowait(struct super_block *sb, ext4_group_t block_group,
+				 int ignore_locked)
 {
 	struct ext4_group_desc *desc;
 	struct ext4_sb_info *sbi = EXT4_SB(sb);
@@ -435,6 +436,13 @@ ext4_read_block_bitmap_nowait(struct super_block *sb, ext4_group_t block_group)
 	if (bitmap_uptodate(bh))
 		goto verify;
 
+	if (ignore_locked && buffer_locked(bh)) {
+		/* buffer under IO already, do not wait
+		 * if called for prefetching */
+		err = 0;
+		goto out;
+	}
+
 	lock_buffer(bh);
 	if (bitmap_uptodate(bh)) {
 		unlock_buffer(bh);
@@ -524,7 +532,7 @@ ext4_read_block_bitmap(struct super_block *sb, ext4_group_t block_group)
 	struct buffer_head *bh;
 	int err;
 
-	bh = ext4_read_block_bitmap_nowait(sb, block_group);
+	bh = ext4_read_block_bitmap_nowait(sb, block_group, 1);
 	if (IS_ERR(bh))
 		return bh;
 	err = ext4_wait_block_bitmap(sb, block_group, bh);
diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
index d4e47fdad87c..2320d7e2f8d6 100644
--- a/fs/ext4/ext4.h
+++ b/fs/ext4/ext4.h
@@ -1482,6 +1482,7 @@ struct ext4_sb_info {
 	unsigned long s_mb_last_start;
 	unsigned int s_mb_toscan0;
 	unsigned int s_mb_toscan1;
+	unsigned int s_mb_prefetch;
 
 	/* stats for buddy allocator */
 	atomic_t s_bal_reqs;	/* number of reqs with len > 1 */
@@ -2335,7 +2336,8 @@ extern struct ext4_group_desc * ext4_get_group_desc(struct super_block * sb,
 extern int ext4_should_retry_alloc(struct super_block *sb, int *retries);
 
 extern struct buffer_head *ext4_read_block_bitmap_nowait(struct super_block *sb,
-						ext4_group_t block_group);
+						ext4_group_t block_group,
+						int ignore_locked);
 extern int ext4_wait_block_bitmap(struct super_block *sb,
 				  ext4_group_t block_group,
 				  struct buffer_head *bh);
diff --git a/fs/ext4/mballoc.c b/fs/ext4/mballoc.c
index cebd7d8df0b8..eac4ee225527 100644
--- a/fs/ext4/mballoc.c
+++ b/fs/ext4/mballoc.c
@@ -861,7 +861,7 @@ static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
 			bh[i] = NULL;
 			continue;
 		}
-		bh[i] = ext4_read_block_bitmap_nowait(sb, group);
+		bh[i] = ext4_read_block_bitmap_nowait(sb, group, 0);
 		if (IS_ERR(bh[i])) {
 			err = PTR_ERR(bh[i]);
 			bh[i] = NULL;
@@ -2095,6 +2095,48 @@ static int ext4_mb_good_group(struct ext4_allocation_context *ac,
 	return 0;
 }
 
+/*
+ * each allocation context (i.e. a thread doing allocation) has own
+ * sliding prefetch window of @s_mb_prefetch size which starts at the
+ * very first goal and moves ahead of scaning.
+ * a side effect is that subsequent allocations will likely find
+ * the bitmaps in cache or at least in-flight.
+ */
+static void
+ext4_mb_prefetch(struct ext4_allocation_context *ac,
+		    ext4_group_t start)
+{
+	ext4_group_t ngroups = ext4_get_groups_count(ac->ac_sb);
+	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
+	struct ext4_group_info *grp;
+	ext4_group_t group = start;
+	struct buffer_head *bh;
+	int nr;
+
+	/* batch prefetching to get few READs in flight */
+	if (group + (sbi->s_mb_prefetch >> 1) < ac->ac_prefetch)
+		return;
+
+	nr = sbi->s_mb_prefetch;
+	while (nr > 0) {
+		if (++group >= ngroups)
+			group = 0;
+		if (unlikely(group == start))
+			break;
+		grp = ext4_get_group_info(ac->ac_sb, group);
+		/* ignore empty groups - those will be skipped
+		 * during the scanning as well */
+		if (grp->bb_free == 0)
+			continue;
+		nr--;
+		if (!EXT4_MB_GRP_NEED_INIT(grp))
+			continue;
+		bh = ext4_read_block_bitmap_nowait(ac->ac_sb, group, 1);
+		brelse(bh);
+	}
+	ac->ac_prefetch = group;
+}
+
 static noinline_for_stack int
 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
 {
@@ -2160,6 +2202,9 @@ ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
 	 * cr == 0 try to get exact allocation,
 	 * cr == 3  try to get anything
 	 */
+
+	ac->ac_prefetch = ac->ac_g_ex.fe_group;
+
 repeat:
 	for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
 		ac->ac_criteria = cr;
@@ -2187,6 +2232,8 @@ ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
 			if (group >= ngroups)
 				group = 0;
 
+			ext4_mb_prefetch(ac, group);
+
 			/* This now checks without needing the buddy page */
 			ret = ext4_mb_good_group(ac, group, cr);
 			if (ret <= 0) {
@@ -2882,6 +2929,7 @@ void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid)
 	}
 	sbi->s_mb_toscan0 = 1024;
 	sbi->s_mb_toscan1 = 4096;
+	sbi->s_mb_prefetch = 32;
 
 	list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list)
 		ext4_free_data_in_buddy(sb, entry);
diff --git a/fs/ext4/mballoc.h b/fs/ext4/mballoc.h
index 88c98f17e3d9..9ba5c75e6490 100644
--- a/fs/ext4/mballoc.h
+++ b/fs/ext4/mballoc.h
@@ -175,6 +175,7 @@ struct ext4_allocation_context {
 	struct page *ac_buddy_page;
 	struct ext4_prealloc_space *ac_pa;
 	struct ext4_locality_group *ac_lg;
+	ext4_group_t ac_prefetch;
 };
 
 #define AC_STATUS_CONTINUE	1
diff --git a/fs/ext4/sysfs.c b/fs/ext4/sysfs.c
index c96ee20f5487..4476d828439b 100644
--- a/fs/ext4/sysfs.c
+++ b/fs/ext4/sysfs.c
@@ -200,6 +200,7 @@ EXT4_ATTR(last_error_time, 0444, last_error_time);
 EXT4_ATTR(journal_task, 0444, journal_task);
 EXT4_RW_ATTR_SBI_UI(mb_toscan0, s_mb_toscan0);
 EXT4_RW_ATTR_SBI_UI(mb_toscan1, s_mb_toscan1);
+EXT4_RW_ATTR_SBI_UI(mb_prefetch, s_mb_prefetch);
 
 static unsigned int old_bump_val = 128;
 EXT4_ATTR_PTR(max_writeback_mb_bump, 0444, pointer_ui, &old_bump_val);
@@ -232,6 +233,7 @@ static struct attribute *ext4_attrs[] = {
 	ATTR_LIST(journal_task),
 	ATTR_LIST(mb_toscan0),
 	ATTR_LIST(mb_toscan1),
+	ATTR_LIST(mb_prefetch),
 	NULL,
 };
 ATTRIBUTE_GROUPS(ext4);
-- 
2.20.1







[Index of Archives]     [Reiser Filesystem Development]     [Ceph FS]     [Kernel Newbies]     [Security]     [Netfilter]     [Bugtraq]     [Linux FS]     [Yosemite National Park]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Device Mapper]     [Linux Media]

  Powered by Linux