[RFC PATCH 0/2] mkfs: large fs IO scalability improvements

[RFC PATCH 0/2] mkfs: large fs IO scalability improvements

[Dave Chinner]

Hi folks,

These are the patches that fell out of looking at why making an 8EB
filesystem was taking so long on one of my faster VMs. Sure, it's
millions of IOs, but there was 3x as much IO as there should have
been, and a large amount of it was unexpected, inefficient RMW
cycles cycles during free list initialisation.

Details in the patches, I don't think there's anything controversial
in them. Let me know what you think.

Cheers,

Dave.

[PATCH 1/2] mkfs: stop zeroing old superblocks excessively

[Dave Chinner]

From: Dave Chinner <dchinner@redhat.com>

When making a new filesystem, don't zero superblocks way beyond the
end of the new filesystem. If the old filesystem was an EB scale
filesytsem, then this zeroing requires millions of IOs to complete.
We don't want to do this if the new filesystem on the device is only
going to be 100TB. Sure, zeroing old superblocks a good distance
beyond the new size is a good idea, as is zeroing the ones in the
middle and end, but the other 7,999,000 superblocks? Not so much.

Make a sane cut-off decision - zero out to 10x the size of the new
filesystem, then zero the middle AGs in the old filesystem, then
zero the last ones.

The initial zeroing out to 10x the new fs size means that this code
will only ever trigger in rare corner cases outside a testing
environment - there are very few production workloads where a huge
block device is reused immediately and permanently for a tiny much
smaller filesystem. Those that do this (e.g. on thing provisioned
devices) discard the in use blocks anyway and so the zeroing won't
actually do anything useful.

Time to mkfs a 1TB filsystem on a big device after it held another
larger filesystem:

previous FS size	10PB	100PB	 1EB
old mkfs time		1.95s	8.9s	81.3s
patched			0.95s	1.2s	 1.2s


Signed-off-by: Dave Chinner <dchinner@redhat.com>
---
 mkfs/xfs_mkfs.c | 64 +++++++++++++++++++++++++++++++++++++++++++++----
 1 file changed, 59 insertions(+), 5 deletions(-)

diff --git a/mkfs/xfs_mkfs.c b/mkfs/xfs_mkfs.c
index 2e53c1e83b6a..c153592c705e 100644
--- a/mkfs/xfs_mkfs.c
+++ b/mkfs/xfs_mkfs.c
@@ -1155,14 +1155,15 @@ validate_ag_geometry(
 
 static void
 zero_old_xfs_structures(
-	libxfs_init_t		*xi,
-	xfs_sb_t		*new_sb)
+	struct libxfs_xinit	*xi,
+	struct xfs_sb		*new_sb)
 {
-	void 			*buf;
-	xfs_sb_t 		sb;
+	void			*buf;
+	struct xfs_sb		sb;
 	uint32_t		bsize;
 	int			i;
 	xfs_off_t		off;
+	xfs_off_t		end;
 
 	/*
 	 * We open regular files with O_TRUNC|O_CREAT. Nothing to do here...
@@ -1220,15 +1221,68 @@ zero_old_xfs_structures(
 
 	/*
 	 * block size and basic geometry seems alright, zero the secondaries.
+	 *
+	 * Don't be insane when it comes to overwriting really large filesystems
+	 * as it could take millions of IOs to zero every secondary
+	 * superblock. If we are remaking a huge filesystem, then do the
+	 * zeroing, but if we are replacing it with a small one (typically done
+	 * in test environments, limit the zeroing to:
+	 *
+	 *	- around the range of the new filesystem
+	 *	- the middle of the old filesystem
+	 *	- the end of the old filesystem
+	 *
+	 * Killing the middle and end of the old filesystem will prevent repair
+	 * from finding it with it's fast secondary sb scan algorithm. The slow
+	 * scan algorithm will then confirm the small filesystem geometry by
+	 * brute force scans.
 	 */
 	memset(buf, 0, new_sb->sb_sectsize);
+
+	/* this carefully avoids integer overflows */
+	end = sb.sb_dblocks;
+	if (sb.sb_agcount > 10000 &&
+	    new_sb->sb_dblocks < end / 10)
+		end = new_sb->sb_dblocks * 10;
 	off = 0;
-	for (i = 1; i < sb.sb_agcount; i++)  {
+	for (i = 1; i < sb.sb_agcount && off < end; i++)  {
+		off += sb.sb_agblocks;
+		if (pwrite(xi->dfd, buf, new_sb->sb_sectsize,
+					off << sb.sb_blocklog) == -1)
+			break;
+	}
+
+	if (end == sb.sb_dblocks)
+		return;
+
+	/*
+	 * Trash the middle 1000 AGs of the old fs, which we know has at least
+	 * 10000 AGs at this point. Cast to make sure we are doing 64bit
+	 * multiplies, otherwise off gets truncated to 32 bit. I hate C.
+	 */
+	i = (sb.sb_agcount / 2) - 500;
+	off = (xfs_off_t)sb.sb_agblocks * i;
+	off = (xfs_off_t)sb.sb_agblocks * ((sb.sb_agcount / 2) - 500);
+	end = off + 1000 * sb.sb_agblocks;
+	while (off < end) {
+		if (pwrite(xi->dfd, buf, new_sb->sb_sectsize,
+					off << sb.sb_blocklog) == -1)
+			break;
 		off += sb.sb_agblocks;
+	}
+
+	/*
+	 * Trash the last 1000 AGs of the old fs
+	 */
+	off = (xfs_off_t)sb.sb_agblocks * (sb.sb_agcount - 1000);
+	end = sb.sb_dblocks;
+	while (off < end) {
 		if (pwrite(xi->dfd, buf, new_sb->sb_sectsize,
 					off << sb.sb_blocklog) == -1)
 			break;
+		off += sb.sb_agblocks;
 	}
+
 done:
 	free(buf);
 }
-- 
2.17.0

[PATCH 2/2] mkfs: rework AG header initialisation ordering

[Dave Chinner]

From: Dave Chinner <dchinner@redhat.com>

When observing the behaviour of an 8EB mkfs execution, I noticed
that a phase where there are a massive number of read/modify/write
cycles occurring. I didn't wait for it to complete - it was obvious
that it was after all the AG headers had been written. That left the
AGFL initialisation as the likely cause.

When all the AG headers don't fit in the libxfs buffer cache, the
AGFL init requires re-reading the AGF, the AGFL, the free space tree
root blocks and the rmap tree root block. They all then get
modified and written back out. 10 IOs per AG. When you have 8
million AGs, that's a lot of extra IO.

Change the initialisation algorithm to initialise the AGFL
immediately after initialising the rest of the headers and
calculating the minimum AGFL size for that AG. This means the
modifications will all hit the buffer cache and this will remove the
IO penalty.

The "worst_freelist" size calculation doesn't change from AG to AG -
it's based on the physical configuration of the AG, and all AGs have
the same configuration. hence we only need to calculate this once,
not for every AG. That allows us to initialise the AGFL immediately
after the rest of the AG has been initialised rather than in a
separate pass.

TIme to make a filesystem from scratch, using a zeroed device so the
force overwrite algorithms are not triggered and -K to avoid
discards:

FS size		10PB	100PB	 1EB
current mkfs	26.9s	214.8s	2484s
patched		11.3s	 70.3s	 709s

In both cases, the IO profile looks identical for the initial AG
header writeout loop. The difference is that the old code then
does the RMW loop to init the AGFL, and that runs at about half the
speed. Hence runtime of the new code is reduce by around 65-70%
simply by avoiding all that IO.


Signed-off-by: Dave Chinner <dchinner@redhat.com>
---
 mkfs/xfs_mkfs.c | 40 +++++++++++++++++++++++++---------------
 1 file changed, 25 insertions(+), 15 deletions(-)

diff --git a/mkfs/xfs_mkfs.c b/mkfs/xfs_mkfs.c
index c153592c705e..d70fbdb6b15a 100644
--- a/mkfs/xfs_mkfs.c
+++ b/mkfs/xfs_mkfs.c
@@ -3374,7 +3374,7 @@ initialise_ag_headers(
 	struct xfs_mount	*mp,
 	struct xfs_sb		*sbp,
 	xfs_agnumber_t		agno,
-	int			*worst_freelist)
+	int			*freelist_size)
 {
 	struct xfs_perag	*pag = libxfs_perag_get(mp, agno);
 	struct xfs_agfl		*agfl;
@@ -3453,8 +3453,22 @@ initialise_ag_headers(
 		agf->agf_longest = cpu_to_be32(agsize -
 			XFS_FSB_TO_AGBNO(mp, cfg->logstart) - cfg->logblocks);
 	}
-	if (libxfs_alloc_min_freelist(mp, pag) > *worst_freelist)
-		*worst_freelist = libxfs_alloc_min_freelist(mp, pag);
+
+	/*
+	 * The AGFL size is the same for all AGs because all AGs have the same
+	 * layout. If this AG sameness ever changes in the future, we'll need to
+	 * revisit how we initialise the AGFLs.
+	 */
+	if (*freelist_size == 0)
+		*freelist_size = libxfs_alloc_min_freelist(mp, pag);
+	else if (*freelist_size < libxfs_alloc_min_freelist(mp, pag)) {
+		fprintf(stderr,
+_("%s: Abort! Freelist size (%u) for AG %u not constant (%u)!\n"),
+			progname, libxfs_alloc_min_freelist(mp, pag),
+			agno, *freelist_size);
+		exit(1);
+	}
+
 	libxfs_writebuf(buf, LIBXFS_EXIT_ON_FAILURE);
 
 	/*
@@ -3724,14 +3738,14 @@ static void
 initialise_ag_freespace(
 	struct xfs_mount	*mp,
 	xfs_agnumber_t		agno,
-	int			worst_freelist)
+	int			freelist_size)
 {
 	struct xfs_alloc_arg	args;
 	struct xfs_trans	*tp;
 	struct xfs_trans_res tres = {0};
 	int			c;
 
-	c = libxfs_trans_alloc(mp, &tres, worst_freelist, 0, 0, &tp);
+	c = libxfs_trans_alloc(mp, &tres, freelist_size, 0, 0, &tp);
 	if (c)
 		res_failed(c);
 
@@ -3797,7 +3811,7 @@ main(
 	int			quiet = 0;
 	char			*protofile = NULL;
 	char			*protostring = NULL;
-	int			worst_freelist = 0;
+	int			freelist_size = 0;
 
 	struct libxfs_xinit	xi = {
 		.isdirect = LIBXFS_DIRECT,
@@ -4025,16 +4039,12 @@ main(
 	}
 
 	/*
-	 * Initialise all the static on disk metadata.
+	 * Initialise all the AG headers on disk.
 	 */
-	for (agno = 0; agno < cfg.agcount; agno++)
-		initialise_ag_headers(&cfg, mp, sbp, agno, &worst_freelist);
-
-	/*
-	 * Initialise the freespace freelists (i.e. AGFLs) in each AG.
-	 */
-	for (agno = 0; agno < cfg.agcount; agno++)
-		initialise_ag_freespace(mp, agno, worst_freelist);
+	for (agno = 0; agno < cfg.agcount; agno++) {
+		initialise_ag_headers(&cfg, mp, sbp, agno, &freelist_size);
+		initialise_ag_freespace(mp, agno, freelist_size);
+	}
 
 	/*
 	 * Allocate the root inode and anything else in the proto file.
-- 
2.17.0