Re: Semaphores and threads anomaly and bug?

Re: Semaphores and threads anomaly and bug?

[Manfred Spraul]

John wrote:

>I know this isn't defined anywhere but the seems to be an ambiguity and 
>discrepancy between versions of Unix and Linux over threads and semaphores.
>
>Do the "SEM_UNDO"s get applied when a thread terminates or when the 
>"whole thing" terminates?
>  
>
According to the Unix spec: per-process.
Older Linux kernels applied it per-thread. Newer kernels can handle it 
per-process, and AFAIK it's the default for NPTL.

>I think that in ipc/sem.c line 1062 the line should be made 
>conditional on "u->semadj[i]" being non-zero.
>  
>
Fixed in 2.6. But there is another bug in that block: undos can increase 
the semaphore value above SEMVMX.

>There is a potential problem here in that the code in ipc/sem.c doesn't 
>allow the adjustment to yield a negative value but what if it starts at 
>zero, thread A increments it, thread B decrements it back to zero (both 
>with SEM_UNDO) and thread A exits first? Thread A's undo won't work and 
>then thread B's undo will increment it again leaving it in an incorrect 
>state which is different from thread B exiting first.
>  
>
Correct. undo operations should never try to decrease the semaphore 
value - an attempt to decrease below 0 is either silently ignored, or 
the semaphore value is set to 0.

--
    Manfred

Re: Semaphores and threads anomaly and bug?

[John M Collins]

On Tuesday 04 November 2003 17:29, Linus Torvalds wrote:
> On Tue, 4 Nov 2003, John M Collins wrote:
> > I know this isn't defined anywhere but the seems to be an ambiguity and
> > discrepancy between versions of Unix and Linux over threads and
> > semaphores.
> >
> > Do the "SEM_UNDO"s get applied when a thread terminates or when the
> > "whole thing" terminates?
>
> It's entirely up to you. That's what CLONE_SYSVSEM is supposed to
> determine.
>
> However, CLONE_SYSVSEM is fairly recent, and I think you will need to use
> the new threading libraries to see it.

Thanks.

Could I possibly draw your attention to the other point I made about the 
updating of "sempid" which does seem to be a mistake to me?

-- 
John Collins Xi Software Ltd www.xisl.com

Re: Semaphores and threads anomaly and bug?

[Linus Torvalds]

On Tue, 4 Nov 2003, John M Collins wrote:
> 
> I know this isn't defined anywhere but the seems to be an ambiguity and 
> discrepancy between versions of Unix and Linux over threads and semaphores.
> 
> Do the "SEM_UNDO"s get applied when a thread terminates or when the 
> "whole thing" terminates?

It's entirely up to you. That's what CLONE_SYSVSEM is supposed to 
determine.

However, CLONE_SYSVSEM is fairly recent, and I think you will need to use 
the new threading libraries to see it.

		Linus

Semaphores and threads anomaly and bug?

[John M Collins]

(Please CC me in any reply as I'm not subscribed)

I know this isn't defined anywhere but the seems to be an ambiguity and 
discrepancy between versions of Unix and Linux over threads and semaphores.

Do the "SEM_UNDO"s get applied when a thread terminates or when the 
"whole thing" terminates?

I tried the following C++ program

#include <iostream>
#include <sys/ipc.h>
#include <sys/sem.h>
#include <pthread.h>

using    namespace  std;

int    semchan;

union semun {
      int val;                  /* value for SETVAL */
      struct semid_ds *buf;     /* buffer for IPC_STAT, IPC_SET */
      unsigned short *array;    /* array for GETALL, SETALL */
                                /* Linux specific part: */
      struct seminfo *__buf;    /* buffer for IPC_INFO */
};

int    getsemv()
{
    return  semctl(semchan, 0, GETVAL, 0);
}

void    *tfunc(void *arg)
{
    cout << "About to set sema4 current=" << getsemv() << endl;
    sembuf  sv;
    sv.sem_op = 77;
    sv.sem_flg = SEM_UNDO;
    sv.sem_num = 0;
    semop(semchan, &sv, 1);
    cout << "Done,sem now=" << getsemv() << endl;
    return  0;
}

int    main()
{
    semchan = semget(999, 1, 0666 | IPC_CREAT);
    semun  z;
    z.val = 1;
    semctl(semchan, 0, SETVAL, z);
    cout << "Created sema4 initial value=" << getsemv() << endl;
    cout << "Creating thread" << endl;
    pthread_t  th;
    pthread_create(&th, 0, tfunc, 0);
    pthread_join(th, 0);
    cout << "After thread value=" << getsemv() << endl;
    return  0;
}

Trying it on Linux (2.4.21 kernel) it says:

Created sema4 initial value=1
Creating thread
About to set sema4 current=1
Done,sem now=78
After thread value=1

Implying that the thread exit applies the SEM_UNDOs whereas trying on 
Solaris 2.9 and HP/UX 11 I get

Created sema4 initial value=1
Creating thread
About to set sema4 current=1
Done,sem now=78
After thread value=78

After which the value is 1, implying that the SEM_UNDOs get applied on 
process exit.

There is another anomaly which applies to SEM_UNDO in that when a 
process exits, every semaphore in the set has its "sempid" set to the 
process id of the exiting process, even ones that the process left 
alone. I think that in ipc/sem.c line 1062 the line should be made 
conditional on "u->semadj[i]" being non-zero.

I know all this isn't defined anywhere but I hit on this trying to write 
an application involving a server process using threads and a 
semaphore-protected shared memory segment accessed by client processes. 
The semaphore might be set to "locked" by a different thread from the 
one that "unlocks" it (I'm using "Mutexes" inside the server process). I 
know you probably don't want SEM_UNDO in a fully-debugged server 
process, only in the clients but it's all a question of getting the said 
server process into the glorious state of being "fully-debugged" or 
somewhere near there.

There is a potential problem here in that the code in ipc/sem.c doesn't 
allow the adjustment to yield a negative value but what if it starts at 
zero, thread A increments it, thread B decrements it back to zero (both 
with SEM_UNDO) and thread A exits first? Thread A's undo won't work and 
then thread B's undo will increment it again leaving it in an incorrect 
state which is different from thread B exiting first.

-- 
John Collins Xi Software Ltd www.xisl.com