pipeio: prevent race between SIGCHLD and open()

This test occasionally hangs on some machines. The hang has been observed mostly on single CPU ones.

pipeio code is using signal(2), setting by default SA_RESTART flag, which is also the case for SIGCHLD.

If last child manages to exit while parent is still at open(), parent gets SIGCHLD and open() is restarted. At this point test hangs.

Here's strace output from parent point of view:

=== snip ===

brk(0)                                  = 0x11bb000
brk(0x11dd000)                          = 0x11dd000
getpid()                                = 18826
stat("tpipe.18826", 0x7fff89e1d410)     = -1 ENOENT (No such file or

directory)
mknod("tpipe.18826", S_IFIFO|0777) = 0
rt_sigaction(SIGCHLD, {0x400a54, [CHLD], SA_RESTORER|SA_RESTART,
0x354aa32a20}, {SIG_DFL, [], 0}, 8) = 0
clone(child_stack=0, flags=CLONE_CHILD_CLEARTID|
CLONE_CHILD_SETTID|SIGCHLD, child_tidptr=0x7f53b9ddd9d0) = 18827
fstat(1, {st_mode=S_IFCHR|0620, st_rdev=makedev(136, 3), ...}) = 0 mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0)
= 0x7f53b9de5000
open("tpipe.18826", O_RDONLY
) = ? ERESTARTSYS (To be restarted)
--- SIGCHLD (Child exited) @ 0 (0) ---
wait4(-1, [{WIFEXITED(s) && WEXITSTATUS(s) == 0}], 0, NULL) = 18827
rt_sigreturn(0xffffffffffffffff) = 2
open("tpipe.18826", O_RDONLY
=== /snip ===

This patch is introducing semaphore, which prevents children from exiting until parent completes open(). It also adds timed wait, so parent waits for children to exit before it deletes pipe and semaphore.

---
 testcases/kernel/ipc/pipeio/pipeio.c |   45 
+++++++++++++++++++++++++++++++---
 1 files changed, 41 insertions(+), 4 deletions(-)

index 22ab3a7..dd52316 100644
@@ -158,6 +158,8 @@ char *av[];
 	  struct semid_ds *buf;
 	  unsigned short int *array;
 	} u;
+	unsigned int uwait_iter = 1000;
+	unsigned int uwait_total = 5000000;
 
 	u.val = 0;
 	format = HEX;

@@ -443,12 +445,16 @@ char *av[];
 
 	writebuf[size-1] = 'A';	/* to detect partial read/write problem */
 
-	if ((sem_id = semget(IPC_PRIVATE, 1, IPC_CREAT|S_IRWXU)) == -1) {
+	if ((sem_id = semget(IPC_PRIVATE, 2, IPC_CREAT|S_IRWXU)) == -1) {
 		tst_brkm(TBROK|TERRNO, NULL, "Couldn't allocate semaphore");
 	}
 
 	if (semctl(sem_id, 0, SETVAL, u) == -1)
-		tst_brkm(TBROK|TERRNO, NULL, "Couldn't initialize semaphore value");
+		tst_brkm(TBROK|TERRNO, NULL, "Couldn't initialize semaphore 0 value");
+
+	/* semaphore to hold off children from exiting until open() completes */
+	if (semctl(sem_id, 1, SETVAL, u) == -1)
+		tst_brkm(TBROK|TERRNO, NULL, "Couldn't initialize semaphore 1 value");
 
 	if (background) {
 	    if ((n=fork()) == -1) {

@@ -539,7 +545,7 @@ printf("child after fork pid = %d\n", getpid());
 		};
 
 		if (semop(sem_id, &sem_op, 1) == -1)
-			tst_brkm(TBROK|TERRNO, NULL, "Couldn't raise the semaphore");
+			tst_brkm(TBROK|TERRNO, NULL, "Couldn't raise the semaphore 0");
 
 		pid_word = (int *)&writebuf[0];
 		count_word = (int *)&writebuf[NBPW];

@@ -586,6 +592,15 @@ printf("child after fork pid = %d\n", getpid());
 			}
 			fflush(stderr);
 		}
+
+		/* child waits until parent completes open() */
+		sem_op = (struct sembuf) {
+			.sem_num = 1,
+			.sem_op = -1,
+			.sem_flg = 0
+		};
+		if (semop(sem_id, &sem_op, 1) == -1)
+			tst_brkm(TBROK|TERRNO, NULL, "Couldn't lower the semaphore 1");
 	}
 	if (c > 0) {	/***** if parent *****/
 

@@ -602,6 +617,15 @@ printf("child after fork pid = %d\n", getpid());
 			close(write_fd);
 		}
 
+		/* raise semaphore so children can exit */
+		sem_op = (struct sembuf) {
+			.sem_num = 1,
+			.sem_op = num_wrters,
+			.sem_flg = 0
+		};
+		if (semop(sem_id, &sem_op, 1) == -1)
+			tst_brkm(TBROK|TERRNO, NULL, "Couldn't raise the semaphore 1");
+
 		sem_op = (struct sembuf) {
 			.sem_num = 0,
 			.sem_op = -num_wrters,

@@ -612,7 +636,7 @@ printf("child after fork pid = %d\n", getpid());
 			if (errno == EINTR) {
 				continue;
 			}
-			tst_brkm(TBROK|TERRNO, NULL, "Couldn't wait on semaphore");
+			tst_brkm(TBROK|TERRNO, NULL, "Couldn't wait on semaphore 0");
 		}
 
 		for (i=num_wrters*num_writes; i > 0 || loop; --i) {

@@ -694,6 +718,19 @@ output:
 				tst_resm(TPASS, "1 PASS %d pipe reads complete, read size = %d, %s %s",
 				          count+1,size,pipe_type,blk_type);
 
+		/* wait for all children to finish, timeout after uwait_total
+		   semtimedop might not be available everywhere */
+		for (i=0; i<uwait_total; i+=uwait_iter) {
+			if (semctl(sem_id, 1, GETVAL) == 0) {
+				break;
+			}
+			usleep(uwait_iter);
+		}
+
+		if (i > uwait_total) {
+			tst_resm(TWARN, "Timed out waiting for child processes to exit");
+		}
+
 		semctl(sem_id, 0, IPC_RMID);
 
 		if (!unpipe)

------------------------------------------------------------------------------
10 Tips for Better Server Consolidation
Server virtualization is being driven by many needs. But none more important than the need to reduce IT complexity while improving strategic productivity. Learn More!
http://www.accelacomm.com/jaw/sdnl/114/51507609/

_______________________________________________
Ltp-list mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/ltp-list

Committed -- thanks!
-Garrett

------------------------------------------------------------------------------
10 Tips for Better Server Consolidation
Server virtualization is being driven by many needs. But none more important than the need to reduce IT complexity while improving strategic productivity. Learn More!
http://www.accelacomm.com/jaw/sdnl/114/51507609/
_______________________________________________
Ltp-list mailing list [email protected]
https://lists.sourceforge.net/lists/listinfo/ltp-list

Garrett,

I'm not sure what happened, but I see only first patch from series committed with description belonging to second one.

Regards,
Jan

----- Original Message -----

------------------------------------------------------------------------------
10 Tips for Better Server Consolidation
Server virtualization is being driven by many needs. But none more important than the need to reduce IT complexity while improving strategic productivity. Learn More!
http://www.accelacomm.com/jaw/sdnl/114/51507609/
_______________________________________________
Ltp-list mailing list [email protected]
https://lists.sourceforge.net/lists/listinfo/ltp-list

I got confused by the multiple patches. Is this thread the one I
need to look at to rectify the other issues identified? Thanks,
-Garrett

------------------------------------------------------------------------------
10 Tips for Better Server Consolidation
Server virtualization is being driven by many needs. But none more important than the need to reduce IT complexity while improving strategic productivity. Learn More!
http://www.accelacomm.com/jaw/sdnl/114/51507609/
_______________________________________________
Ltp-list mailing list [email protected]
https://lists.sourceforge.net/lists/listinfo/ltp-list

----- Original Message -----

Yes, this is the one.

I posted single patch yesterday:
[LTP] [PATCH] pipeio: prevent race between SIGCHLD and open()

You requested to use TERRNO instead of strerror(errno), so while I was at it, I made this change for all existing code in pipeio.c:
[LTP] [PATCH v2 1/2] pipeio: convert strerror(errno) to TERRNO

and then put my original patch (with minor changes) on top of that:
[LTP] [PATCH v2 2/2] pipeio: prevent race between SIGCHLD and open()

Regards,
Jan

------------------------------------------------------------------------------
10 Tips for Better Server Consolidation
Server virtualization is being driven by many needs. But none more important than the need to reduce IT complexity while improving strategic productivity. Learn More!
http://www.accelacomm.com/jaw/sdnl/114/51507609/
_______________________________________________
Ltp-list mailing list [email protected]
https://lists.sourceforge.net/lists/listinfo/ltp-list

Hi!

Should be fixed in latest commit, thanks!

Cyril Hrubis
[email protected]

------------------------------------------------------------------------------
Write once. Port to many.
Get the SDK and tools to simplify cross-platform app development. Create 
new or port existing apps to sell to consumers worldwide. Explore the 
Intel AppUpSM program developer opportunity. appdeveloper.intel.com/join
http://p.sf.net/sfu/intel-appdev
_______________________________________________
Ltp-list mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/ltp-list