[LTP] [PATCH] unshare03: using soft limit of NOFILE

[Petr Vorel]

> Hi!
> > I think it's safer to set NOFILE increasing from soft limit than from
> > hard limit.

> > Hard limit may lead to dup2 ENOMEM error which bring the result to
> > TBROK on little memory machine. (e.g. 2GB memory in my situation, hard
> > limit in /proc/sys/fs/nr_open come out to be 1073741816)

> > Signed-off-by: lufei <lufei@uniontech.com>
> > ---
> >  testcases/kernel/syscalls/unshare/unshare03.c | 14 ++++++--------
> >  1 file changed, 6 insertions(+), 8 deletions(-)

> > diff --git a/testcases/kernel/syscalls/unshare/unshare03.c b/testcases/kernel/syscalls/unshare/unshare03.c
> > index 7c5e71c4e..bb568264c 100644
> > --- a/testcases/kernel/syscalls/unshare/unshare03.c
> > +++ b/testcases/kernel/syscalls/unshare/unshare03.c
> > @@ -24,7 +24,7 @@

> >  static void run(void)
> >  {
> > -	int nr_open;
> > +	int rlim_max;
> >  	int nr_limit;
> >  	struct rlimit rlimit;
> >  	struct tst_clone_args args = {
> > @@ -32,14 +32,12 @@ static void run(void)
> >  		.exit_signal = SIGCHLD,
> >  	};

> > -	SAFE_FILE_SCANF(FS_NR_OPEN, "%d", &nr_open);
> > -	tst_res(TDEBUG, "Maximum number of file descriptors: %d", nr_open);
> > +	SAFE_GETRLIMIT(RLIMIT_NOFILE, &rlimit);
> > +	rlim_max = rlimit.rlim_max;

> > -	nr_limit = nr_open + NR_OPEN_LIMIT;
> > +	nr_limit = rlim_max + NR_OPEN_LIMIT;
> >  	SAFE_FILE_PRINTF(FS_NR_OPEN, "%d", nr_limit);

> > -	SAFE_GETRLIMIT(RLIMIT_NOFILE, &rlimit);
> > -
> >  	rlimit.rlim_cur = nr_limit;
> >  	rlimit.rlim_max = nr_limit;

> > @@ -47,10 +45,10 @@ static void run(void)
> >  	tst_res(TDEBUG, "Set new maximum number of file descriptors to : %d",
> >  		nr_limit);

> > -	SAFE_DUP2(2, nr_open + NR_OPEN_DUP);
> > +	SAFE_DUP2(2, rlim_max + NR_OPEN_DUP);

> >  	if (!SAFE_CLONE(&args)) {
> > -		SAFE_FILE_PRINTF(FS_NR_OPEN, "%d", nr_open);
> > +		SAFE_FILE_PRINTF(FS_NR_OPEN, "%d", rlim_max);
> >  		TST_EXP_FAIL(unshare(CLONE_FILES), EMFILE);
> >  		exit(0);
> >  	}

> Why do we bother with reading the /rpoc/sys/fs/nr_open file? All that we
> need to to do is to dup() a file descriptor and tnen set the nr_open
> limit to fd - 2. And if we do so we can drop the rlimit that increases
> the limit so that it's greater than nr_open as well.

IMHO file descriptor will be 3, fd - 2 == 1. And trying to set 1 to
/rpoc/sys/fs/nr_open leads to EINVAL. You probably mean something different.

The test is based on unshare_test.c [1] from Al Viro (611fbeb44a777 [2]). Both
tests IMHO use nr_open + 1024 nr_open + 1024 and then call dup2() on nr_open + 64
to avoid later EINVAL when writing /rpoc/sys/fs/nr_open.

Kind regards,
Petr

[1] https://web.git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/testing/selftests/core/unshare_test.c
[2] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=611fbeb44a777e5ab54ab3127ec85f72147911d8