dgaudet 97/11/04 00:33:13
Modified: src/test time-sem.c
Log:
latest greatest time-sem.c
Revision Changes Path
1.3 +64 -15 apachen/src/test/time-sem.c
Index: time-sem.c
===================================================================
RCS file: /export/home/cvs/apachen/src/test/time-sem.c,v
retrieving revision 1.2
retrieving revision 1.3
diff -u -r1.2 -r1.3
--- time-sem.c 1997/11/04 07:57:43 1.2
+++ time-sem.c 1997/11/04 08:33:13 1.3
@@ -1,23 +1,38 @@
/*
-Date: Sat, 1 Nov 1997 16:53:52 -0800 (PST)
-From: Dean Gaudet <[EMAIL PROTECTED]>
+time-sem.c has the basics of the semaphores we use in http_main.c. It's
+intended for timing differences between various methods on an
+architecture. In practice we've found many things affect which semaphore
+to be used:
+
+ - NFS filesystems absolutely suck for fcntl() and flock()
+
+ - uslock absolutely sucks on single-processor IRIX boxes, but
+ absolutely rocks on multi-processor boxes. The converse
+ is true for fcntl. sysvsem seems a moderate balance.
+
+ - Under Solaris you can't have too many processes use SEM_UNDO, there
+ might be a tuneable somewhere that increases the limit from 29.
+ We're not sure what the tunable is, so there's a define
+ NO_SEM_UNDO which can be used to simulate us trapping/blocking
+ signals to be able to properly release the semaphore on a clean
+ child death.
+
+argv[1] is the #children, argv[2] is the #iterations per child
+
+You should run each over many different #children inputs, and choose
+#iter such that the program runs for at least a second or so... or even
+longer depending on your patience.
-This time-sem.c includes the necessary signal manipulations to allow us to
-continue to use pthreads. It appears to still be and order of magnitude
-faster than fcntl locking. I'm a little too busy to make a patch though,
-so if someone could take the pthread code in time-sem.c here and compare
-it against the server ... and generate a patch, that'd be great. You also
-have to properly release the mutex on any of the three unblocked signals --
-which you should do in the handlers we already have for those signals,
-so that we don't have to also instate a handler during the critical
-section ...
-
-Dean
+compile with:
gcc -o time-FCNTL -Wall -O time-sem.c -DUSE_FCNTL_SERIALIZED_ACCEPT
gcc -o time-FLOCK -Wall -O time-sem.c -DUSE_FLOCK_SERIALIZED_ACCEPT
-gcc -o time-SEM -Wall -O time-sem.c -DUSE_SYSVSEM_SERIALIZED_ACCEPT
+gcc -o time-SYSVSEM -Wall -O time-sem.c -DUSE_SYSVSEM_SERIALIZED_ACCEPT
+gcc -o time-SYSVSEM2 -Wall -O time-sem.c -DUSE_SYSVSEM_SERIALIZED_ACCEPT
-DNO_SEM_UNDO
gcc -o time-PTHREAD -Wall -O time-sem.c -DUSE_PTHREAD_SERIALIZED_ACCEPT
-lpthread
+gcc -o time-USLOCK -Wall -O time-sem.c -DUSE_USLOCK_SERIALIZED_ACCEPT
+
+not all versions work on all systems.
*/
#include <errno.h>
@@ -146,6 +161,10 @@
#include <sys/sem.h>
static int sem_id = -1;
+#ifdef NO_SEM_UNDO
+static sigset_t accept_block_mask;
+static sigset_t accept_previous_mask;
+#endif
void accept_mutex_init(void)
{
@@ -167,15 +186,29 @@
perror ("semctl");
exit(1);
}
+#ifdef NO_SEM_UNDO
+ sigfillset(&accept_block_mask);
+ sigdelset(&accept_block_mask, SIGHUP);
+ sigdelset(&accept_block_mask, SIGTERM);
+ sigdelset(&accept_block_mask, SIGUSR1);
+#endif
}
void accept_mutex_on()
{
struct sembuf op;
+#ifdef NO_SEM_UNDO
+ if (sigprocmask(SIG_BLOCK, &accept_block_mask, &accept_previous_mask)) {
+ perror("sigprocmask(SIG_BLOCK)");
+ exit (1);
+ }
+ op.sem_flg = 0;
+#else
+ op.sem_flg = SEM_UNDO;
+#endif
op.sem_num = 0;
op.sem_op = -1;
- op.sem_flg = SEM_UNDO;
if (semop(sem_id, &op, 1) < 0) {
perror ("accept_mutex_on");
exit (1);
@@ -188,15 +221,30 @@
op.sem_num = 0;
op.sem_op = 1;
+#ifdef NO_SEM_UNDO
+ op.sem_flg = 0;
+#else
op.sem_flg = SEM_UNDO;
+#endif
if (semop(sem_id, &op, 1) < 0) {
perror ("accept_mutex_off");
exit (1);
}
+#ifdef NO_SEM_UNDO
+ if (sigprocmask(SIG_SETMASK, &accept_previous_mask, NULL)) {
+ perror("sigprocmask(SIG_SETMASK)");
+ exit (1);
+ }
+#endif
}
#elif defined (USE_PTHREAD_SERIALIZED_ACCEPT)
+/* note: pthread mutexes aren't released on child death, hence the
+ * signal goop ... in a real implementation we'd do special things
+ * during hup, term, usr1.
+ */
+
#include <pthread.h>
static pthread_mutex_t *mutex;
@@ -403,5 +451,6 @@
}
last.tv_usec = ms;
printf ("%8lu.%06lu\n", last.tv_sec, last.tv_usec);
+ exit(0);
}
