syscalls/shmctl05.c: Fix ENOSPC error

Message ID 1533108893-13078-1-git-send-email-yangx.jy@cn.fujitsu.com
State Accepted
Headers show
Series
  • syscalls/shmctl05.c: Fix ENOSPC error
Related show

Commit Message

Xiao Yang Aug. 1, 2018, 7:34 a.m.
Running shmctl05 got the following error on some distros(e.g. RHEL7.5):
------------------------------------------------------
tst_safe_sysv_ipc.c:111: BROK: shmctl05.c:51: shmget(61455, 4096, 3c0) failed: ENOSPC
shmctl05.c:104: WARN: pthread_join(..., (nil)) failed: EDEADLK
tst_safe_sysv_ipc.c:111: BROK: shmctl05.c:81: shmget(61455, 4096, 3c0) failed: ENOSPC
-------------------------------------------------------

From shmctl(2) manpage, shmctl(IPC_RMID) just marks a shm segment to
be destroyed, and the segment will only actually be destroyed after
the last process detaches it (i.e., when the shm_nattch member of the
associated structure shmid_ds is zero).  So it is possible for the
number of created shm segments to exceed the system-wide maximum number
(e.g. shmmni is 4096) if only shmctl(IPC_RMID) has been called.

From shmdt(2) manpage, a successful shmdt(2) call will decrement
the shm_nattch by one.  So we should call shmdt(2) to decrement
the shm_nattch to zero before calling shmctl(IPC_RMID).

Add shmdt(2) to ensure that all shm segments are actually destroyed.

Signed-off-by: Xiao Yang <yangx.jy@cn.fujitsu.com>
---
 testcases/kernel/syscalls/ipc/shmctl/shmctl05.c | 2 ++
 1 file changed, 2 insertions(+)

Comments

Eric Biggers Aug. 24, 2018, 4:27 a.m. | #1
Hi Xiao,

On Wed, Aug 01, 2018 at 03:34:53PM +0800, Xiao Yang wrote:
> Running shmctl05 got the following error on some distros(e.g. RHEL7.5):
> ------------------------------------------------------
> tst_safe_sysv_ipc.c:111: BROK: shmctl05.c:51: shmget(61455, 4096, 3c0) failed: ENOSPC
> shmctl05.c:104: WARN: pthread_join(..., (nil)) failed: EDEADLK
> tst_safe_sysv_ipc.c:111: BROK: shmctl05.c:81: shmget(61455, 4096, 3c0) failed: ENOSPC
> -------------------------------------------------------
> 
> From shmctl(2) manpage, shmctl(IPC_RMID) just marks a shm segment to
> be destroyed, and the segment will only actually be destroyed after
> the last process detaches it (i.e., when the shm_nattch member of the
> associated structure shmid_ds is zero).  So it is possible for the
> number of created shm segments to exceed the system-wide maximum number
> (e.g. shmmni is 4096) if only shmctl(IPC_RMID) has been called.
> 
> From shmdt(2) manpage, a successful shmdt(2) call will decrement
> the shm_nattch by one.  So we should call shmdt(2) to decrement
> the shm_nattch to zero before calling shmctl(IPC_RMID).
> 
> Add shmdt(2) to ensure that all shm segments are actually destroyed.
> 
> Signed-off-by: Xiao Yang <yangx.jy@cn.fujitsu.com>
> ---
>  testcases/kernel/syscalls/ipc/shmctl/shmctl05.c | 2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/testcases/kernel/syscalls/ipc/shmctl/shmctl05.c b/testcases/kernel/syscalls/ipc/shmctl/shmctl05.c
> index a960cc9..0eb0776 100644
> --- a/testcases/kernel/syscalls/ipc/shmctl/shmctl05.c
> +++ b/testcases/kernel/syscalls/ipc/shmctl/shmctl05.c
> @@ -92,6 +92,8 @@ static void do_test(void)
>  				"Unexpected remap_file_pages() error");
>  		}
>  		tst_fzsync_wait_a(&fzsync_pair);
> +		/* ensure that a shm segment will actually be destroyed */
> +		SAFE_SHMDT(addr);
>  	}
>  
>  	tst_res(TPASS, "didn't crash");
> -- 

I think you missed part of the explanation for why this test (apparently) fails
on old kernels.  On recent kernels, remap_file_pages() *is* unmapping the shm
segment, so the test passes.  Perhaps the behavior of remap_file_pages() changed
in v4.0 when its implementation was replaced with an emulation.

Calling shmdt() is probably the right fix for the test, but you shouldn't call
the SAFE_* version since shmdt() will fail with an error on recent kernels,
which with the SAFE_* version would fail the test.

- Eric
Jan Stancek Aug. 28, 2018, 11:27 a.m. | #2
----- Original Message -----
> Hi Xiao,
> 
> On Wed, Aug 01, 2018 at 03:34:53PM +0800, Xiao Yang wrote:
> > Running shmctl05 got the following error on some distros(e.g. RHEL7.5):
> > ------------------------------------------------------
> > tst_safe_sysv_ipc.c:111: BROK: shmctl05.c:51: shmget(61455, 4096, 3c0)
> > failed: ENOSPC
> > shmctl05.c:104: WARN: pthread_join(..., (nil)) failed: EDEADLK
> > tst_safe_sysv_ipc.c:111: BROK: shmctl05.c:81: shmget(61455, 4096, 3c0)
> > failed: ENOSPC
> > -------------------------------------------------------
> > 
> > From shmctl(2) manpage, shmctl(IPC_RMID) just marks a shm segment to
> > be destroyed, and the segment will only actually be destroyed after
> > the last process detaches it (i.e., when the shm_nattch member of the
> > associated structure shmid_ds is zero).  So it is possible for the
> > number of created shm segments to exceed the system-wide maximum number
> > (e.g. shmmni is 4096) if only shmctl(IPC_RMID) has been called.
> > 
> > From shmdt(2) manpage, a successful shmdt(2) call will decrement
> > the shm_nattch by one.  So we should call shmdt(2) to decrement
> > the shm_nattch to zero before calling shmctl(IPC_RMID).
> > 
> > Add shmdt(2) to ensure that all shm segments are actually destroyed.
> > 
> > Signed-off-by: Xiao Yang <yangx.jy@cn.fujitsu.com>
> > ---
> >  testcases/kernel/syscalls/ipc/shmctl/shmctl05.c | 2 ++
> >  1 file changed, 2 insertions(+)
> > 
> > diff --git a/testcases/kernel/syscalls/ipc/shmctl/shmctl05.c
> > b/testcases/kernel/syscalls/ipc/shmctl/shmctl05.c
> > index a960cc9..0eb0776 100644
> > --- a/testcases/kernel/syscalls/ipc/shmctl/shmctl05.c
> > +++ b/testcases/kernel/syscalls/ipc/shmctl/shmctl05.c
> > @@ -92,6 +92,8 @@ static void do_test(void)
> >  				"Unexpected remap_file_pages() error");
> >  		}
> >  		tst_fzsync_wait_a(&fzsync_pair);
> > +		/* ensure that a shm segment will actually be destroyed */
> > +		SAFE_SHMDT(addr);
> >  	}
> >  
> >  	tst_res(TPASS, "didn't crash");
> > --
> 
> I think you missed part of the explanation for why this test (apparently)
> fails
> on old kernels.  On recent kernels, remap_file_pages() *is* unmapping the shm
> segment, so the test passes.  Perhaps the behavior of remap_file_pages()
> changed
> in v4.0 when its implementation was replaced with an emulation.
> 
> Calling shmdt() is probably the right fix for the test, but you shouldn't
> call
> the SAFE_* version since shmdt() will fail with an error on recent kernels,
> which with the SAFE_* version would fail the test.

Pushed with plain shmdt(), expanded comment and commit message.

Thanks,
Jan

Patch

diff --git a/testcases/kernel/syscalls/ipc/shmctl/shmctl05.c b/testcases/kernel/syscalls/ipc/shmctl/shmctl05.c
index a960cc9..0eb0776 100644
--- a/testcases/kernel/syscalls/ipc/shmctl/shmctl05.c
+++ b/testcases/kernel/syscalls/ipc/shmctl/shmctl05.c
@@ -92,6 +92,8 @@  static void do_test(void)
 				"Unexpected remap_file_pages() error");
 		}
 		tst_fzsync_wait_a(&fzsync_pair);
+		/* ensure that a shm segment will actually be destroyed */
+		SAFE_SHMDT(addr);
 	}
 
 	tst_res(TPASS, "didn't crash");