After a semaphore is created the sem_otime member of semid_ds returned
by an IPC_STAT will be initialized to 0. After the first semop call it
is set to non-zero and AFAIK should never be reset to 0. However this
is precisely what I'm seeing. Attached is a program [attachment removed
and message resent due to list moderation - will send semc.c on demand
directly] that performs a variety of semaphore operations.
This command opens a semaphore with IPC_CREAT, IPC_EXCL, mode 777,
1 sem in the array, and initializes it's value to 10:
$ ./semc /tmp/foo -o 1 1 777 1 10
CMD_OPEN: semid: 0 semnum: 0 path: /tmp/foo
flags: oflags: O_CREAT|O_EXCL mode: 0777 numsems: 1 value: 10
393216 = cmd_open("/tmp/foo", 1, O_CREAT|O_EXCL, 0777, 10)
We can query the value with:
$ ./semc 393216 0 -g
CMD_GETVAL: semid: 393216 semnum: 0 path:
flags: oflags: mode: 0000 numsems: 0 value: 0
10 = cmd_getvalue(393216, 0, GETVAL)
10 is right. So far so good. But if we try to open the semaphore *without
O_EXCL* there's a problem:
$ ./semc /tmp/foo -o 1 0 777 1 10
CMD_OPEN: semid: 0 semnum: 0 path: /tmp/foo
flags: oflags: O_CREAT mode: 0777 numsems: 1 value: 10
149: semid_get: Operation timed out
Command failed
The problem is that it is necessary to check sem_otime to make sure it is
!= 0 to avoid a race condition where a caller A creates a sem with semget
but a second caller B swoops in, opens the sem and performs and operation
(e.g. post, wait, getval, etc) before caller A has the opportunity to
initialize the sem value with semop. The code to do this in the attached
semc.c follows:
130 if ((semid = semget(key, nsems, 0)) != -1) {
131 struct semid_ds buf;
132
133/* This inner try-loop ensures that the semaphore is initialized
before
134 * we return even if the semaphore has been created with semget but
not
135 * yet initialized with semctl. See Stevens' UNPv2 p274.
136 */
137 arg.buf = &buf;
138 for (max = MAX_TRIES; max; max--) {
139 if (semctl(semid, 0, IPC_STAT, arg) == -1) {
140 ERR(errno, "semctl");
141 return -1;
142 }
143 if (buf.sem_otime != 0) {
144 return semid;
145 }
146 sleep(1);
147 }
148
149 ERR(errno = ETIMEDOUT, "semid_get");
150 return -1;
This hole in the API is described very well in Stevens' Unix Network
Programming Volume 2 page 274.
So the problem with Darwin is that creating a sem, exiting, and then
opening the sem a second time appears to reset the sem_otime value.
This code works as expected (sem_otime does not get reset to 0) on Linux
2.4, HP-UX, OSF1, and BSD. It's only Darwin that behaves differently.
This is a major problem because there caller opening an existing
semaphore will just hang waiting for the other non-existant program to
initialize it.
Pls advise,
Mike
--
IRC - where men are men, women are men, and the boys are FBI agents.
_______________________________________________
Do not post admin requests to the list. They will be ignored.
Darwin-dev mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
http://lists.apple.com/mailman/options/darwin-dev/email@hidden
This email sent to email@hidden
