Extended attribute readahead

When gfs2 allocates an inode and its extended attribute block next to each other at inode create time, the inode's directory entry indicates that in de_rahead. In that case, we can readahead the extended attribute block when we read in the inode.

Signed-off-by Andreas Gruenbacher <[email protected]>

---
 fs/gfs2/dir.c     | 15 +++++++++++----
 fs/gfs2/incore.h  |  1 +
 fs/gfs2/meta_io.c | 27 +++++++++++++++++++++++++--
 fs/gfs2/meta_io.h |  2 +-
 fs/gfs2/quota.c   |  2 +-
 fs/gfs2/rgrp.c    |  2 +-
 fs/gfs2/super.c   |  1 +
 fs/gfs2/xattr.c   | 10 +++++-----
 8 files changed, 46 insertions(+), 14 deletions(-)

----- Original Message -----

Quoted 6 lines

When gfs2 allocates an inode and its extended attribute block next to each other at inode create time, the inode's directory entry indicates that in de_rahead. In that case, we can readahead the extended attribute block when we read in the inode.

Signed-off-by Andreas Gruenbacher <[email protected]>

Quoted 11 lines

---
  fs/gfs2/dir.c     | 15 +++++++++++----
  fs/gfs2/incore.h  |  1 +
  fs/gfs2/meta_io.c | 27 +++++++++++++++++++++++++--
  fs/gfs2/meta_io.h |  2 +-
  fs/gfs2/quota.c   |  2 +-
  fs/gfs2/rgrp.c    |  2 +-
  fs/gfs2/super.c   |  1 +
  fs/gfs2/xattr.c   | 10 +++++-----
  8 files changed, 46 insertions(+), 14 deletions(-)

Hi Andreas,

Most of this looks good. However, two comments:

1. I don't like adding a new u16 to the gfs2_inode. I've been working to reduce the size of gfs2's inodes lately, so I'd rather see this implemented as a new GIF_RAHEAD (or similar) flag in gfs2_inode's i_flags.
2. It seems to me like we should take advantage of function gfs2_meta_ra() which already submits one block and a variable number of additional blocks for read-ahead, then waits for the first block IO to complete.

Regards,

Bob Peterson
Red Hat File Systems

Quoted 6 lines

Most of this looks good. However, two comments:

1. I don't like adding a new u16 to the gfs2_inode. I've been working to reduce the size of gfs2's inodes lately, so I'd rather see this implemented as a new GIF_RAHEAD (or similar) flag in gfs2_inode's
i_flags.

Quoted 3 lines

2. It seems to me like we should take advantage of function gfs2_meta_ra() which already submits one block and a variable number of additional blocks for read-ahead, then waits for the first block IO to complete.

The meta_ra thing is a bit of a hack, and best avoided for this use. We want to only send out a single I/O here rather than let the I/O stack do the merging after the fact,

Instead of submitting separate bio for the inode and its extended attributes, submit a single bio for both when possible. The entire request becomes a normal read, not a readahead.

To keep track of the buffer heads that make up the bio, we allocate temporary buffer head arrays: in the endio handler, it would also be possible to compute the buffer head block numbers from the bio and to grab the buffer heads with gfs2_getbuf, but the code would become even messier.

---
  fs/gfs2/meta_io.c | 94 ++++++++++++++++++++++++++++++++++++++++++++-----------
  1 file changed, 76 insertions(+), 18 deletions(-)

fs/gfs2/meta_io.c +32 -7

index 0f24828..e650127 100644
@@ -187,19 +187,63 @@ struct buffer_head *gfs2_meta_new(struct gfs2_glock *gl, u64 blkno)
  	return bh;
  }
 
-static void gfs2_meta_readahead(struct gfs2_glock *gl, u64 blkno)
+struct gfs2_meta_read {
+	int num;
+	struct buffer_head *bhs[0];
+};
+
+static void gfs2_meta_read_endio(struct bio *bio) {
+	struct gfs2_meta_read *r = bio->bi_private;
+	int i;
+
+	for (i = 0; i < r->num; i++) {
+		struct buffer_head *bh = r->bhs[i];
+
+		if (unlikely(bio_flagged(bio, BIO_QUIET)))
+			set_bit(BH_Quiet, &bh->b_state);
+
+		bh->b_end_io(bh, !bio->bi_error);
+	}
+	bio_put(bio);
+	kfree(r);
+}
+
+/*
+ * (See submit_bh_wbc.)
+ */
+static void gfs2_submit_bhs(int rw, struct buffer_head *bhs[], int num)
  {
-	struct buffer_head *bh;
+	struct gfs2_meta_read *r;
+	struct buffer_head *bh = bhs[0];
+	struct bio *bio;
+	int i;
 
-	bh = gfs2_getbuf(gl, blkno, 1);
-	lock_buffer(bh);
-	if (buffer_uptodate(bh)) {
-		unlock_buffer(bh);
-		brelse(bh);
+	if (!num)
+		return;
+

fs/gfs2/meta_io.c +38 -9

+	r->num = num;
+	for (i = 0; i < num; i++)
+		r->bhs[i] = bhs[i];
+
+	bio = bio_alloc(GFP_NOIO, num);
+	bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
+	bio->bi_bdev = bh->b_bdev;
+	for (i = 0; i < num; i++) {
+		bh = bhs[i];
+		bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
+	}
+	bio->bi_end_io = gfs2_meta_read_endio;
+	bio->bi_private = r;
+
+	submit_bio(rw, bio);
  }
 
  /**

@@ -216,7 +260,8 @@ int gfs2_meta_read(struct gfs2_glock *gl, u64 blkno, int flags,
  		   int rahead, struct buffer_head **bhp)
  {
  	struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
-	struct buffer_head *bh;
+	struct buffer_head *bh, *bhs[2];
+	int num = 0;
 
  	if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags))) {
  		*bhp = NULL;

@@ -228,18 +273,31 @@ int gfs2_meta_read(struct gfs2_glock *gl, u64 blkno, int flags,
  	lock_buffer(bh);
  	if (buffer_uptodate(bh)) {
  		unlock_buffer(bh);
-		if (rahead)
-			gfs2_meta_readahead(gl, blkno + 1);
-		return 0;
+		flags &= ~DIO_WAIT;
+	} else {
+		bh->b_end_io = end_buffer_read_sync;
+		get_bh(bh);
+		bhs[num++] = bh;
  	}
-	bh->b_end_io = end_buffer_read_sync;
-	get_bh(bh);
-	submit_bh(READ_SYNC | REQ_META | REQ_PRIO, bh);
-	if (rahead)
-		gfs2_meta_readahead(gl, blkno + 1);
+
+	if (rahead) {
+		bh = gfs2_getbuf(gl, blkno + 1, CREATE);
+
+		lock_buffer(bh);
+		if (buffer_uptodate(bh)) {
+			unlock_buffer(bh);
+			brelse(bh);
+		} else {
+			bh->b_end_io = end_buffer_read_sync;
+			bhs[num++] = bh;
+		}
+	}
+
+	gfs2_submit_bhs(READ_SYNC | REQ_META | REQ_PRIO, bhs, num);
  	if (!(flags & DIO_WAIT))
  		return 0;
 
+	bh = *bhp;
  	wait_on_buffer(bh);
  	if (unlikely(!buffer_uptodate(bh))) {
  		struct gfs2_trans *tr = current->journal_info;

Hi,

Quoted 71 lines · Andreas Gruenbacher

Instead of submitting separate bio for the inode and its extended attributes, submit a single bio for both when possible. The entire request becomes a normal read, not a readahead.

To keep track of the buffer heads that make up the bio, we allocate temporary buffer head arrays: in the endio handler, it would also be possible to compute the buffer head block numbers from the bio and to grab the buffer heads with gfs2_getbuf, but the code would become even messier.

---
  fs/gfs2/meta_io.c | 94
++++++++++++++++++++++++++++++++++++++++++++-----------
  1 file changed, 76 insertions(+), 18 deletions(-)

fs/gfs2/meta_io.c +36 -7

index 0f24828..e650127 100644
@@ -187,19 +187,63 @@ struct buffer_head *gfs2_meta_new(struct gfs2_glock
 *gl, u64 blkno)
        return bh;
  }
  -static void gfs2_meta_readahead(struct gfs2_glock *gl, u64 blkno)
+struct gfs2_meta_read {
+       int num;
+       struct buffer_head *bhs[0];
+};
+
+static void gfs2_meta_read_endio(struct bio *bio) {
+       struct gfs2_meta_read *r = bio->bi_private;
+       int i;
+
+       for (i = 0; i < r->num; i++) {
+               struct buffer_head *bh = r->bhs[i];
+
+               if (unlikely(bio_flagged(bio, BIO_QUIET)))
+                       set_bit(BH_Quiet, &bh->b_state);
+
+               bh->b_end_io(bh, !bio->bi_error);
+       }
+       bio_put(bio);
+       kfree(r);
+}
+
+/*
+ * (See submit_bh_wbc.)
+ */
+static void gfs2_submit_bhs(int rw, struct buffer_head *bhs[], int num)
  {
-       struct buffer_head *bh;
+       struct gfs2_meta_read *r;
+       struct buffer_head *bh = bhs[0];
+       struct bio *bio;
+       int i;
  -     bh = gfs2_getbuf(gl, blkno, 1);
-       lock_buffer(bh);
-       if (buffer_uptodate(bh)) {
-               unlock_buffer(bh);
-               brelse(bh);
+       if (!num)
+               return;
+
+       if (num == 1) {
+               bh->b_end_io = end_buffer_read_sync;
+               submit_bh(rw, bh);
                return;
        }
-       bh->b_end_io = end_buffer_read_sync;
-       submit_bh(READA | REQ_META | REQ_PRIO, bh);
+

fs/gfs2/meta_io.c struct buffer_head *gfs2_meta_new(struct gfs2_glock *gl, u64 blkno) · line 231

Quoted 2 lines

fs/gfs2/meta_io.c +2 -0

+       r = kmalloc(sizeof(*r) + num * sizeof(r->bhs[0]),
+                   GFP_NOIO | __GFP_NOFAIL);

Can we avoid doing this I wonder? I would like to avoid adding GFP_NOFAIL allocations, and I think it should be possible to figure out where the buffers are using the bio itself (i.e. iterating over it, similar to gfs2_end_log_write() in lops.c) so that we don't need to allocate this additional memory.

Here is an updated version of this patch, please review.

Thanks,
Andreas

--

Instead of submitting a READ_SYNC bio for the inode and a READA bio for the inode's extended attributes through submit_bh, submit a single READ bio for both strough submit_bio when possible. This can be more efficient on some kinds of block devices.

Signed-off-by Andreas Gruenbacher <[email protected]>

---
  fs/gfs2/meta_io.c | 81 ++++++++++++++++++++++++++++++++++++++++++-------------
  1 file changed, 63 insertions(+), 18 deletions(-)

fs/gfs2/meta_io.c +63 -18

index 0f24828..e137d96 100644
@@ -187,19 +187,50 @@ struct buffer_head *gfs2_meta_new(struct gfs2_glock *gl, u64 blkno)
  	return bh;
  }
 
-static void gfs2_meta_readahead(struct gfs2_glock *gl, u64 blkno)
+static void gfs2_meta_read_endio(struct bio *bio)
  {
-	struct buffer_head *bh;
+	struct bio_vec *bvec;
+	int i;
+
+	bio_for_each_segment_all(bvec, bio, i) {
+		struct page *page = bvec->bv_page;
+		struct buffer_head *bh = page_buffers(page);
+		unsigned int len = bvec->bv_len;
+
+		while (bh_offset(bh) < bvec->bv_offset)
+			bh = bh->b_this_page;
+		do {
+			struct buffer_head *next = bh->b_this_page;
+			len -= bh->b_size;
+			bh->b_end_io(bh, !bio->bi_error);
+			bh = next;
+		} while (bh && len);
+	}
+	bio_put(bio);
+}
 
-	bh = gfs2_getbuf(gl, blkno, 1);
-	lock_buffer(bh);
-	if (buffer_uptodate(bh)) {
-		unlock_buffer(bh);
-		brelse(bh);
+/*
+ * Submit several consecutive buffer head I/O requests as a single bio I/O
+ * request.  (See submit_bh_wbc.)
+ */
+static void gfs2_submit_bhs(int rw, struct buffer_head *bhs[], int num)
+{
+	struct buffer_head *bh = bhs[0];
+	struct bio *bio;
+	int i;
+
+	if (!num)
  		return;
+
+	bio = bio_alloc(GFP_NOIO, num);
+	bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
+	bio->bi_bdev = bh->b_bdev;
+	for (i = 0; i < num; i++) {
+		bh = bhs[i];
+		bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
  	}
-	bh->b_end_io = end_buffer_read_sync;
-	submit_bh(READA | REQ_META | REQ_PRIO, bh);
+	bio->bi_end_io = gfs2_meta_read_endio;
+	submit_bio(rw, bio);
  }
 
  /**

@@ -216,7 +247,8 @@ int gfs2_meta_read(struct gfs2_glock *gl, u64 blkno, int flags,
  		   int rahead, struct buffer_head **bhp)
  {
  	struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
-	struct buffer_head *bh;
+	struct buffer_head *bh, *bhs[2];
+	int num = 0;
 
  	if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags))) {
  		*bhp = NULL;

@@ -228,18 +260,31 @@ int gfs2_meta_read(struct gfs2_glock *gl, u64 blkno, int flags,
  	lock_buffer(bh);
  	if (buffer_uptodate(bh)) {
  		unlock_buffer(bh);
-		if (rahead)
-			gfs2_meta_readahead(gl, blkno + 1);
-		return 0;
+		flags &= ~DIO_WAIT;
+	} else {
+		bh->b_end_io = end_buffer_read_sync;
+		get_bh(bh);
+		bhs[num++] = bh;
  	}
-	bh->b_end_io = end_buffer_read_sync;
-	get_bh(bh);
-	submit_bh(READ_SYNC | REQ_META | REQ_PRIO, bh);
-	if (rahead)
-		gfs2_meta_readahead(gl, blkno + 1);
+
+	if (rahead) {
+		bh = gfs2_getbuf(gl, blkno + 1, CREATE);
+
+		lock_buffer(bh);
+		if (buffer_uptodate(bh)) {
+			unlock_buffer(bh);
+			brelse(bh);
+		} else {
+			bh->b_end_io = end_buffer_read_sync;
+			bhs[num++] = bh;
+		}
+	}
+
+	gfs2_submit_bhs(READ_SYNC | REQ_META | REQ_PRIO, bhs, num);
  	if (!(flags & DIO_WAIT))
  		return 0;
 
+	bh = *bhp;
  	wait_on_buffer(bh);
  	if (unlikely(!buffer_uptodate(bh))) {
  		struct gfs2_trans *tr = current->journal_info;

That looks good to me, and I assume that it should be faster too?

Quoted 1 line · Steven Whitehouse

That looks good to me, and I assume that it should be faster too?

I did some tests with a directory tree with 3439 directories and 51556 files in it.

In that tree, 47313 or 86% of the 54995 files and directories ended up with readahead enabled, while 7682 didn't.

On a slow SATA magnetic disc, I compared the time required for running "find", then "find | xargs stat", then "find | xargs getfattr". I booted with selinux=0 for these measurements, so "find" and "find | xargs stat" didn't use the xattrs.

The results are attached as a spreadsheet. They show that xattr readahead can save quite some time (here, about -75%). The optimized version of xattr readahead is about 3% faster than the simple version in this test, so I'll leave it up to you whether that's worth it for you.

It makes no difference whether the scheduler is "noop" or "deadline" here; the "noop" scheduler still seems to merge adjacent I/O requests (which makes sense).

Andreas

Here is an updated version of this patch, please review.

Thanks,
Andreas

--

Instead of submitting a READ_SYNC bio for the inode and a READA bio for the inode's extended attributes through submit_bh, submit a single READ bio for both strough submit_bio when possible. This can be more efficient on some kinds of block devices.

Signed-off-by Andreas Gruenbacher <[email protected]>

---
 fs/gfs2/meta_io.c | 81
 ++++++++++++++++++++++++++++++++++++++++++-------------
 1 file changed, 63 insertions(+), 18 deletions(-)

fs/gfs2/meta_io.c +40 -9

index 0f24828..e137d96 100644
@@ -187,19 +187,50 @@ struct buffer_head *gfs2_meta_new(struct gfs2_glock
 *gl, u64 blkno)
 	return bh;
 }
 
-static void gfs2_meta_readahead(struct gfs2_glock *gl, u64 blkno)
+static void gfs2_meta_read_endio(struct bio *bio)
 {
-	struct buffer_head *bh;
+	struct bio_vec *bvec;
+	int i;
+
+	bio_for_each_segment_all(bvec, bio, i) {
+		struct page *page = bvec->bv_page;
+		struct buffer_head *bh = page_buffers(page);
+		unsigned int len = bvec->bv_len;
+
+		while (bh_offset(bh) < bvec->bv_offset)
+			bh = bh->b_this_page;
+		do {
+			struct buffer_head *next = bh->b_this_page;
+			len -= bh->b_size;
+			bh->b_end_io(bh, !bio->bi_error);
+			bh = next;
+		} while (bh && len);
+	}
+	bio_put(bio);
+}
 
-	bh = gfs2_getbuf(gl, blkno, 1);
-	lock_buffer(bh);
-	if (buffer_uptodate(bh)) {
-		unlock_buffer(bh);
-		brelse(bh);
+/*
+ * Submit several consecutive buffer head I/O requests as a single bio I/O
+ * request.  (See submit_bh_wbc.)
+ */
+static void gfs2_submit_bhs(int rw, struct buffer_head *bhs[], int num)
+{
+	struct buffer_head *bh = bhs[0];
+	struct bio *bio;
+	int i;
+
+	if (!num)
 		return;
+
+	bio = bio_alloc(GFP_NOIO, num);
+	bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
+	bio->bi_bdev = bh->b_bdev;
+	for (i = 0; i < num; i++) {
+		bh = bhs[i];
+		bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
 	}
-	bh->b_end_io = end_buffer_read_sync;
-	submit_bh(READA | REQ_META | REQ_PRIO, bh);
+	bio->bi_end_io = gfs2_meta_read_endio;
+	submit_bio(rw, bio);
 }
 

/**

fs/gfs2/meta_io.c +2 -1

@@ -216,7 +247,8 @@ int gfs2_meta_read(struct gfs2_glock *gl, u64 blkno, int
 flags,
 		   int rahead, struct buffer_head **bhp)
 {
 	struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
-	struct buffer_head *bh;
+	struct buffer_head *bh, *bhs[2];
+	int num = 0;
 
 	if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags))) {

fs/gfs2/meta_io.c +21 -8

 		*bhp = NULL;

@@ -228,18 +260,31 @@ int gfs2_meta_read(struct gfs2_glock *gl, u64 blkno,
 int flags,
 	lock_buffer(bh);
 	if (buffer_uptodate(bh)) {
 		unlock_buffer(bh);
-		if (rahead)
-			gfs2_meta_readahead(gl, blkno + 1);
-		return 0;
+		flags &= ~DIO_WAIT;
+	} else {
+		bh->b_end_io = end_buffer_read_sync;
+		get_bh(bh);
+		bhs[num++] = bh;
 	}
-	bh->b_end_io = end_buffer_read_sync;
-	get_bh(bh);
-	submit_bh(READ_SYNC | REQ_META | REQ_PRIO, bh);
-	if (rahead)
-		gfs2_meta_readahead(gl, blkno + 1);
+
+	if (rahead) {
+		bh = gfs2_getbuf(gl, blkno + 1, CREATE);
+
+		lock_buffer(bh);
+		if (buffer_uptodate(bh)) {
+			unlock_buffer(bh);
+			brelse(bh);
+		} else {
+			bh->b_end_io = end_buffer_read_sync;
+			bhs[num++] = bh;
+		}
+	}
+
+	gfs2_submit_bhs(READ_SYNC | REQ_META | REQ_PRIO, bhs, num);
 	if (!(flags & DIO_WAIT))
 		return 0;
 
+	bh = *bhp;
 	wait_on_buffer(bh);
 	if (unlikely(!buffer_uptodate(bh))) {

struct gfs2_trans *tr = current->journal_info;
--
2.5.0