lists.openwall.net   lists  /  announce  owl-users  owl-dev  john-users  john-dev  passwdqc-users  yescrypt  popa3d-users  /  oss-security  kernel-hardening  musl  sabotage  tlsify  passwords  /  crypt-dev  xvendor  /  Bugtraq  Full-Disclosure  linux-kernel  linux-netdev  linux-ext4  PHC 
Open Source and information security mailing list archives
 
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Date:   Fri, 19 Jun 2020 18:23:14 +0000
From:   Alex Zhuravlev <azhuravlev@...mcloud.com>
To:     "Theodore Y. Ts'o" <tytso@....edu>
CC:     Alex Zhuravlev <azhuravlev@...mcloud.com>,
        Благодаренко Артём 
        <artem.blagodarenko@...il.com>,
        "linux-ext4@...r.kernel.org" <linux-ext4@...r.kernel.org>
Subject: Re: [PATCH 1/2] ext4:  mballoc - prefetching for bitmaps

Please find it attached, rebased against latest tree.

Thanks, Alex


> On 19 Jun 2020, at 20:16, Theodore Y. Ts'o <tytso@....edu> wrote:
> 
> I'm not seeing any newer version of this patch since the one on this
> thread, posted on May 15th.
> 
> Am I missing something?   What's the latest version of this patch that you have?
> 
>     	     		  	     	    - Ted


>From 023a44b30b1c77873401ee310705933482a89b02 Mon Sep 17 00:00:00 2001
From: Alex Zhuravlev <bzzz@...mcloud.com>
Date: Tue, 21 Apr 2020 10:54:07 +0300
Subject: [PATCH] [PATCH 1/2] ext4:  mballoc - prefetching for bitmaps

This should significantly improve bitmap loading, especially for flex
groups as it tries to load all bitmaps within a flex.group instead of
one by one synchronously.

Prefetching is done in 8 * flex_bg groups, so it should be 8 read-ahead
reads for a single allocating thread. At the end of allocation the
thread waits for read-ahead completion and initializes buddy information
so that read-aheads are not lost in case of memory pressure.

At cr=0 the number of prefetching IOs is limited per allocation context
to prevent a situation when mballoc loads thousands of bitmaps looking
for a perfect group and ignoring groups with good chunks.

Together with the patch "ext4: limit scanning of uninitialized groups"
the mount time (which includes few tiny allocations) of a 1PB filesystem
is reduced significantly:

               0% full    50%-full unpatched    patched
  mount time       33s                9279s       563s

Signed-off-by: Alex Zhuravlev <bzzz@...mcloud.com>
Reviewed-by: Andreas Dilger <adilger@...mcloud.com>
---
 fs/ext4/balloc.c  |  12 ++++-
 fs/ext4/ext4.h    |   8 +++-
 fs/ext4/mballoc.c | 113 +++++++++++++++++++++++++++++++++++++++++++++-
 fs/ext4/mballoc.h |   2 +
 fs/ext4/sysfs.c   |   4 ++
 5 files changed, 135 insertions(+), 4 deletions(-)

diff --git a/fs/ext4/balloc.c b/fs/ext4/balloc.c
index 1ba46d87cdf1..c52011af4812 100644
--- a/fs/ext4/balloc.c
+++ b/fs/ext4/balloc.c
@@ -413,7 +413,8 @@ static int ext4_validate_block_bitmap(struct super_block *sb,
  * Return buffer_head on success or an ERR_PTR 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,
+				 bool ignore_locked)
 {
 	struct ext4_group_desc *desc;
 	struct ext4_sb_info *sbi = EXT4_SB(sb);
@@ -444,6 +445,13 @@ 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 */
+		put_bh(bh);
+		return NULL;
+	}
+
 	lock_buffer(bh);
 	if (bitmap_uptodate(bh)) {
 		unlock_buffer(bh);
@@ -534,7 +542,7 @@ 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, false);
 	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 42f5060f3cdf..7451662e092a 100644
--- a/fs/ext4/ext4.h
+++ b/fs/ext4/ext4.h
@@ -1505,6 +1505,8 @@ struct ext4_sb_info {
 	/* where last allocation was done - for stream allocation */
 	unsigned long s_mb_last_group;
 	unsigned long s_mb_last_start;
+	unsigned int s_mb_prefetch;
+	unsigned int s_mb_prefetch_limit;
 
 	/* stats for buddy allocator */
 	atomic_t s_bal_reqs;	/* number of reqs with len > 1 */
@@ -2446,7 +2448,8 @@ 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,
+						bool ignore_locked);
 extern int ext4_wait_block_bitmap(struct super_block *sb,
 				  ext4_group_t block_group,
 				  struct buffer_head *bh);
@@ -3145,6 +3148,7 @@ struct ext4_group_info {
 	(1 << EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT)
 #define EXT4_GROUP_INFO_IBITMAP_CORRUPT		\
 	(1 << EXT4_GROUP_INFO_IBITMAP_CORRUPT_BIT)
+#define EXT4_GROUP_INFO_BBITMAP_READ_BIT	4
 
 #define EXT4_MB_GRP_NEED_INIT(grp)	\
 	(test_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &((grp)->bb_state)))
@@ -3159,6 +3163,8 @@ struct ext4_group_info {
 	(set_bit(EXT4_GROUP_INFO_WAS_TRIMMED_BIT, &((grp)->bb_state)))
 #define EXT4_MB_GRP_CLEAR_TRIMMED(grp)	\
 	(clear_bit(EXT4_GROUP_INFO_WAS_TRIMMED_BIT, &((grp)->bb_state)))
+#define EXT4_MB_GRP_TEST_AND_SET_READ(grp)	\
+	(test_and_set_bit(EXT4_GROUP_INFO_BBITMAP_READ_BIT, &((grp)->bb_state)))
 
 #define EXT4_MAX_CONTENTION		8
 #define EXT4_CONTENTION_THRESHOLD	2
diff --git a/fs/ext4/mballoc.c b/fs/ext4/mballoc.c
index 9d69f0a1372d..88af980e945e 100644
--- a/fs/ext4/mballoc.c
+++ b/fs/ext4/mballoc.c
@@ -922,7 +922,7 @@ 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, false);
 		if (IS_ERR(bh[i])) {
 			err = PTR_ERR(bh[i]);
 			bh[i] = NULL;
@@ -2244,6 +2244,91 @@ static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac,
 	return ret;
 }
 
+/*
+ * 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)
+{
+	struct super_block *sb = ac->ac_sb;
+	ext4_group_t ngroups = ext4_get_groups_count(sb);
+	struct ext4_sb_info *sbi = EXT4_SB(sb);
+	struct ext4_group_info *grp;
+	ext4_group_t nr, group = start;
+	struct buffer_head *bh;
+
+	/* limit prefetching at cr=0/1, otherwise mballoc can
+	 * spend a lot of time loading imperfect groups */
+	if (ac->ac_criteria < 2 && ac->ac_prefetch_ios >= sbi->s_mb_prefetch_limit)
+		return;
+
+	/* batch prefetching to get few READs in flight */
+	if (ac->ac_prefetch != group)
+		return;
+
+	nr = sbi->s_mb_prefetch;
+	if (ext4_has_feature_flex_bg(sb)) {
+		/* align to flex_bg to get more bitmaps with a single IO */
+		nr = (group / sbi->s_mb_prefetch) * sbi->s_mb_prefetch;
+		nr = nr + sbi->s_mb_prefetch - group;
+	}
+	while (nr-- > 0) {
+		grp = ext4_get_group_info(sb, group);
+
+		/* prevent expensive getblk() on groups w/ IO in progress */
+		if (EXT4_MB_GRP_TEST_AND_SET_READ(grp))
+			goto next;
+
+		/* ignore empty groups - those will be skipped
+		 * during the scanning as well */
+		if (grp->bb_free > 0 && EXT4_MB_GRP_NEED_INIT(grp)) {
+			bh = ext4_read_block_bitmap_nowait(sb, group, true);
+			if (bh && !IS_ERR(bh)) {
+				if (!buffer_uptodate(bh))
+					ac->ac_prefetch_ios++;
+				brelse(bh);
+			}
+		}
+next:
+		if (++group >= ngroups)
+			group = 0;
+	}
+	ac->ac_prefetch = group;
+}
+
+static void
+ext4_mb_prefetch_fini(struct ext4_allocation_context *ac)
+{
+	struct ext4_group_info *grp;
+	ext4_group_t group;
+	int nr, rc;
+
+	/* initialize last window of prefetched groups */
+	nr = ac->ac_prefetch_ios;
+	if (nr > EXT4_SB(ac->ac_sb)->s_mb_prefetch)
+		nr = EXT4_SB(ac->ac_sb)->s_mb_prefetch;
+	group = ac->ac_prefetch;
+	if (!group)
+		group = ext4_get_groups_count(ac->ac_sb);
+	group--;
+	while (nr-- > 0) {
+		grp = ext4_get_group_info(ac->ac_sb, group);
+		if (grp->bb_free > 0 && EXT4_MB_GRP_NEED_INIT(grp)) {
+			rc = ext4_mb_init_group(ac->ac_sb, group, GFP_NOFS);
+			if (rc)
+				break;
+		}
+		if (!group)
+			group = ext4_get_groups_count(ac->ac_sb);
+		group--;
+	}
+}
+
 static noinline_for_stack int
 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
 {
@@ -2317,6 +2402,7 @@ ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
 		 * from the goal value specified
 		 */
 		group = ac->ac_g_ex.fe_group;
+		ac->ac_prefetch = group;
 
 		for (i = 0; i < ngroups; group++, i++) {
 			int ret = 0;
@@ -2328,6 +2414,8 @@ 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_nolock(ac, group, cr);
 			if (ret <= 0) {
@@ -2402,6 +2490,10 @@ ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
 	mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n",
 		 ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status,
 		 ac->ac_flags, cr, err);
+
+	/* use prefetched bitmaps to init buddy so that read info is not lost */
+	ext4_mb_prefetch_fini(ac);
+
 	return err;
 }
 
@@ -2648,6 +2740,25 @@ static int ext4_mb_init_backend(struct super_block *sb)
 			goto err_freebuddy;
 	}
 
+	if (ext4_has_feature_flex_bg(sb)) {
+		/* a single flex group is supposed to be read by a single IO */
+		sbi->s_mb_prefetch = 1 << sbi->s_es->s_log_groups_per_flex;
+		sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */
+	} else {
+		sbi->s_mb_prefetch = 32;
+	}
+	if (sbi->s_mb_prefetch > ext4_get_groups_count(sb))
+		sbi->s_mb_prefetch = ext4_get_groups_count(sb);
+	/* now many real IOs to prefetch within a single allocation at cr=0
+	 * given cr=0 is an CPU-related optimization we shouldn't try to
+	 * load too many groups, at some point we should start to use what
+	 * we've got in memory.
+	 * with an average random access time 5ms, it'd take a second to get
+	 * 200 groups (* N with flex_bg), so let's make this limit 4 */
+	sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4;
+	if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb))
+		sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb);
+
 	return 0;
 
 err_freebuddy:
diff --git a/fs/ext4/mballoc.h b/fs/ext4/mballoc.h
index 6b4d17c2935d..45103cdf08cb 100644
--- a/fs/ext4/mballoc.h
+++ b/fs/ext4/mballoc.h
@@ -171,6 +171,8 @@ 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;
+	int ac_prefetch_ios; /* number of initialied prefetch IO */
 };
 
 #define AC_STATUS_CONTINUE	1
diff --git a/fs/ext4/sysfs.c b/fs/ext4/sysfs.c
index 6c9fc9e21c13..31e0db726d21 100644
--- a/fs/ext4/sysfs.c
+++ b/fs/ext4/sysfs.c
@@ -240,6 +240,8 @@ EXT4_RO_ATTR_ES_STRING(last_error_func, s_last_error_func, 32);
 EXT4_ATTR(first_error_time, 0444, first_error_time);
 EXT4_ATTR(last_error_time, 0444, last_error_time);
 EXT4_ATTR(journal_task, 0444, journal_task);
+EXT4_RW_ATTR_SBI_UI(mb_prefetch, s_mb_prefetch);
+EXT4_RW_ATTR_SBI_UI(mb_prefetch_limit, s_mb_prefetch_limit);
 
 static unsigned int old_bump_val = 128;
 EXT4_ATTR_PTR(max_writeback_mb_bump, 0444, pointer_ui, &old_bump_val);
@@ -283,6 +285,8 @@ static struct attribute *ext4_attrs[] = {
 #ifdef CONFIG_EXT4_DEBUG
 	ATTR_LIST(simulate_fail),
 #endif
+	ATTR_LIST(mb_prefetch),
+	ATTR_LIST(mb_prefetch_limit),
 	NULL,
 };
 ATTRIBUTE_GROUPS(ext4);
-- 
2.21.3

Powered by blists - more mailing lists