Re: [PATCH] generic/233,270: unlimit the max locked memory size for io_uring

[Eryu Guan <guan@eryu.me>]

On Thu, Jan 28, 2021 at 11:02:09AM +0800, Zorro Lang wrote:
> On Wed, Jan 27, 2021 at 03:43:40PM -0800, Darrick J. Wong wrote:
> > On Wed, Jan 27, 2021 at 03:21:43PM +0800, Zorro Lang wrote:
> > > The ltp/fsstress always fails on io_uring_queue_init() by returnning
> > > ENOMEM. Due to io_uring accounts memory it needs under the rlimit
> > > memlocked option, which can be quite low on some setups, especially
> > > on 64K pagesize machine. root isn't under this restriction, but
> > > regular users are. So only g/233 and g/270 which use $qa_user to run
> > > fsstress are failed.
> > > 
> > > To avoid this failure, set max locked memory to unlimited before doing
> > > fsstress, then restore it after test done.
> > > 
> > > Signed-off-by: Zorro Lang <zlang@redhat.com>
> > > ---
> > >  tests/generic/233 | 8 ++++++++
> > >  tests/generic/270 | 8 ++++++++
> > >  2 files changed, 16 insertions(+)
> > > 
> > > diff --git a/tests/generic/233 b/tests/generic/233
> > > index 7eda5774..342690c2 100755
> > > --- a/tests/generic/233
> > > +++ b/tests/generic/233
> > > @@ -43,6 +43,13 @@ _fsstress()
> > >  -f rename=10 -f fsync=2 -f write=15 -f dwrite=15 \
> > >  -n $count -d $out -p 7`
> > >  
> > > +	# io_uring accounts memory it needs under the rlimit memlocked option,
> > > +	# which can be quite low on some setups (especially 64K pagesize). root
> > > +	# isn't under this restriction, but regular users are. To avoid the
> > > +	# io_uring_queue_init fail on ENOMEM, set max locked memory to unlimited
> > > +	# temporarily.
> > > +	lmem=`ulimit -l`
> > > +	ulimit -l unlimited
> > 
> > Should this apply to all the other tests that run fsstress?
> > 
> > $ grep -l FSSTRESS_PROG tests/ | wc -l
> > 94
> 
> The root user isn't under this restriction, only g/233 and g/270 run fsstress
> with non-root user ($qa_user).
> Hmm... is that possible to run xfstests with a non-root user? Or you'd like to

No, fstests is required to be run by root user, there's a check in
'check' script :)

> write a common helper _fsstress() (if run fsx with non-root user, I doubt it's
> needed too)...

Perhaps we should raise the limit in fsstress and fsx while setting up
io_uring_queue_init()? So we don't have to add this ulimit settings to
every test that runs fsstress/fsx with regular user.

Thanks,
Eryu