Re: [PATCH v2 6/8] fsstress: implement the clonerange/deduperange ioctls

["Darrick J. Wong" <darrick.wong@oracle.com>]

On Thu, Feb 22, 2018 at 06:17:31PM +0000, Luis Henriques wrote:
> On Thu, Feb 22, 2018 at 09:27:41AM -0800, Darrick J. Wong wrote:
> > On Thu, Feb 22, 2018 at 04:06:14PM +0000, Luis Henriques wrote:
> > > On Thu, Dec 14, 2017 at 06:07:31PM -0800, Darrick J. Wong wrote:
> > > 
> > > <snip>
> > > 
> > > > +void
> > > > +clonerange_f(
> > > > +	int			opno,
> > > > +	long			r)
> > > > +{
> > > 
> > > <snip>
> > > 
> > > > +	/* Calculate offsets */
> > > > +	len = (random() % FILELEN_MAX) + 1;
> > > > +	len &= ~(stat1.st_blksize - 1);
> > > > +	if (len == 0)
> > > > +		len = stat1.st_blksize;
> > > > +	if (len > stat1.st_size)
> > > > +		len = stat1.st_size;
> > > > +
> > > > +	lr = ((__int64_t)random() << 32) + random();
> > > > +	if (stat1.st_size == len)
> > > > +		off1 = 0;
> > > > +	else
> > > > +		off1 = (off64_t)(lr % MIN(stat1.st_size - len, MAXFSIZE));
> > > > +	off1 %= maxfsize;
> > > > +	off1 &= ~(stat1.st_blksize - 1);
> > > > +
> > > > +	/*
> > > > +	 * If srcfile == destfile, randomly generate destination ranges
> > > > +	 * until we find one that doesn't overlap the source range.
> > > > +	 */
> > > > +	do {
> > > > +		lr = ((__int64_t)random() << 32) + random();
> > > > +		off2 = (off64_t)(lr % MIN(stat2.st_size + (1024 * 1024), MAXFSIZE));
> > > > +		off2 %= maxfsize;
> > > > +		off2 &= ~(stat2.st_blksize - 1);
> > > > +	} while (stat1.st_ino == stat2.st_ino && llabs(off2 - off1) < len);
> > > 
> > > I started seeing hangs in generic/013 on cephfs.  After spending some
> > > time looking, I found that this loops forever.  And the reason seems to
> > > be that stat1.st_blksize is too big for this filesystem (4M) -- when
> > > doing:
> > 
> > "Too big for this filesystem"?
> > 
> > Uh... maybe you'd better start by giving me more stat buffer info --
> > what's st_size?
> > 
> > > 	off1 &= ~(stat1.st_blksize - 1);
> > 
> > These bits round the start offset down to block granularity, since clone
> > range implementations generally require that the ranges align to block
> > boundaries.
> > 
> > (Though AFAICT ceph doesn't support clone range anyway...)
> > 
> > So reading between the lines, is the problem here that ceph advertises a
> > blocksize of 4M and fsstress calls clonerange_f with files that are
> > smaller than 4M in size, so the only possible offsets with a 4M
> > blocksize are zero and that's why we end up looping forever?
> 
> Brilliantly described!  That is *exactly* what I'm seeing and failed to
> describe.  I guess I could use FSSTRESS_AVOID to work around this issue,
> but there are probably better options.

Better to patch fsstress.c against this bug. :)

Does the following patch help?

--D

diff --git a/ltp/fsstress.c b/ltp/fsstress.c
index 935f5de..e107099 100644
--- a/ltp/fsstress.c
+++ b/ltp/fsstress.c
@@ -2222,6 +2222,7 @@ clonerange_f(
 	off64_t			lr;
 	off64_t			off1;
 	off64_t			off2;
+	off64_t			max_off2;
 	size_t			len;
 	int			v1;
 	int			v2;
@@ -2305,9 +2306,10 @@ clonerange_f(
 	 * If srcfile == destfile, randomly generate destination ranges
 	 * until we find one that doesn't overlap the source range.
 	 */
+	max_off2 = MIN(stat2.st_size + (1024ULL * stat2.st_blksize), MAXFSIZE);
 	do {
 		lr = ((int64_t)random() << 32) + random();
-		off2 = (off64_t)(lr % MIN(stat2.st_size + (1024 * 1024), MAXFSIZE));
+		off2 = (off64_t)(lr % max_off2);
 		off2 %= maxfsize;
 		off2 &= ~(stat2.st_blksize - 1);
 	} while (stat1.st_ino == stat2.st_ino && llabs(off2 - off1) < len);