You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@httpd.apache.org by Igor Kovalenko <Ig...@motorola.com> on 2001/11/15 09:47:15 UTC
[PATCH] support for QNX6 - take 2
Here it goes again. Since last version was not quite welcomed, I'll defend
it in advance. First, it fixes the undefined 'NOFD' constant. Of course, I
could also make it totally QNX-specific, but I'd hate to pollute code with
#ifdef SOME_OS unless necessary and I don't think any of you here would like
that either. After all, that is the reason to have OS-specific sections in
one header which would turn on or off appropriate 'features'.
One of the things objected in last version was addition of MAP_ANON clause
in PTHREAD_SERIALIZED_ACCEPT code. As I said, I could agree to remove that,
but ... the PTHREAD_SERIALIZED_ACCEPT is OS-specific feature to begin with,
so it would only matter if it is enabled. Looking at 1.3.22 code, no
platform actually defines USE_PTHREAD_SERIALIZED_ACCEPT by default and only
Solaris, HP-UX (commented out) and AIX appear to even mention it. And the
implementation appears to be Solaris-specific, in fact. That makes me think
I am not hurting anyone by adding MAP_ANON version, I would actually guess
it might make some platforms be able to use that code. And it still works on
Solaris (2.8 has MAP_ANON and it works, 2.5 has no MAP_ANON, so it still
works old way) and saves unnecessary open() and close() calls.
I don't want to waste any more of anyone's time discussing this, the issue
is not worth it. If that part of patch is still unaceptable, I am fine
without it.
Another non-specific change is addition of USE_POSIX_SCOREBOARD condition to
the 'optimize timeouts'. That is because 1) it is logical, since it is same
thing as USE_MMAP_SCOREBOARD & USE_SHMGET_SCOREBOARD and 2) I wrote the
USE_POSIX_SCOREBOARD code.
regards,
- igor
diff -ru apache_1.3.22/src/Configure apache_1.3.22/src/Configure
--- apache_1.3.22/src/Configure Mon Oct 8 14:59:36 2001
+++ apache_1.3.22/src/Configure Thu Oct 18 00:21:27 2001
@@ -554,6 +554,11 @@
CFLAGS="$CFLAGS -DQNX -mf -3"
LIBS="$LIBS -N128k -lsocket -lunix"
DEF_WANTHSREGEX=yes
+ ;;
+ *-*-qnx6)
+ OS='NTO'
+ CFLAGS="$CFLAGS -DNTO"
+ LIBS="$LIBS -lm -lsocket"
;;
*-isc4*)
OS='ISC 4'
@@ -1076,6 +1081,14 @@
LDFLAGS_SHLIB='-b -a archive'
LDFLAGS_MOD_SHLIB=$LDFLAGS_SHLIB
;;
+ *-*-qnx6*)
+ LD_SHLIB="gcc"
+ CFLAGS_SHLIB="-fPIC"
+ LDFLAGS_SHLIB="-shared"
+ LDFLAGS_MOD_SHLIB=$LDFLAGS_SHLIB
+ LDFLAGS_SHLIB_EXPORT="-Wl,-E"
+ SHLIB_SUFFIX_DEPTH=0
+ ;;
*-linux1)
CFLAGS_SHLIB="-fpic"
LDFLAGS_SHLIB="-Bshareable"
diff -ru apache_1.3.22/src/helpers/GuessOS apache_1.3.22/src/helpers/GuessOS
--- apache_1.3.22/src/helpers/GuessOS Mon Oct 8 09:49:33 2001
+++ apache_1.3.22/src/helpers/GuessOS Thu Oct 18 00:30:28 2001
@@ -181,12 +181,22 @@
echo "${MACHINE}-tandem-oss"; exit 0;
;;
- QNX:*)
- if [ "$VERSION" -gt 422 ]; then
- echo "${MACHINE}-qssl-qnx32"
- else
- echo "${MACHINE}-qssl-qnx"
+ *nto:* | QNX:*)
+ if test "${MACHINE}" = "x86pc" || test "${MACHINE}" = "PCI" ; then
+ MACHINE="pc"
fi
+ case $RELEASE in
+ [0-9]*)
+ echo "${MACHINE}-qssl-qnx6"
+ ;;
+ [A-Z]*)
+ if [ "$VERSION" -gt 422 ]; then
+ echo "${MACHINE}-qssl-qnx32"
+ else
+ echo "${MACHINE}-qssl-qnx"
+ fi
+ ;;
+ esac
exit 0
;;
diff -ru apache_1.3.22/src/include/ap_config.h
apache_1.3.22/src/include/ap_config.h
--- apache_1.3.22/src/include/ap_config.h Mon Oct 8 09:49:33 2001
+++ apache_1.3.22/src/include/ap_config.h Sun Oct 21 00:15:14 2001
@@ -745,6 +745,36 @@
#define NET_SIZE_T socklen_t
#endif
+#elif defined(NTO)
+
+#define HAVE_MMAP 1
+#define HAVE_GMTOFF 1
+#define HAVE_SYSLOG 1
+
+#define USE_MMAP_SCOREBOARD
+#define USE_MMAP_FILES 1
+#define USE_PTHREAD_SERIALIZED_ACCEPT
+
+#undef MAP_TMPFILE
+#undef NO_TIMES
+
+/* needs prototypes for execXX */
+#define NEED_PROCESS_H
+
+/*
+ Careful here.
+ Things have changed a lot between 6.0 and 6.1
+ We'll assume 6.1+ is default
+*/
+/* loader does not support #! before 6.1 */
+/* #define NEED_HASHBANG_EMUL */
+/* kernel does not implement getrlimit/setrlimit before 6.1 */
+/* #undef HAVE_SYS_RESOURCE_H */
+/* tiny stack does not implement SO_LINGER */
+/* #undef USE_SO_LINGER */
+/* lingering close is broken with tiny TCP stack */
+/* #define NO_LINGCLOSE */
+
#elif defined(QNX)
#ifndef crypt
char *crypt(const char *pw, const char *salt);
@@ -1156,7 +1186,7 @@
#undef NSIG
#endif
#include <errno.h>
-#if !defined(QNX) && !defined(CONVEXOS11) && !defined(NEXT) &&
!defined(TPF) && !defined(NETWARE) && !defined(MPE)
+#if !defined(QNX) && !defined(NTO) && !defined(CONVEXOS11) &&
!defined(NEXT) && !defined(TPF) && !defined(NETWARE) && !defined(MPE)
#include <memory.h>
#endif
diff -ru apache_1.3.22/src/include/httpd.h apache_1.3.22/src/include/httpd.h
--- apache_1.3.22/src/include/httpd.h Mon Oct 8 21:56:05 2001
+++ apache_1.3.22/src/include/httpd.h Thu Oct 18 01:06:10 2001
@@ -1189,7 +1189,7 @@
* also not using a scoreboard file
*/
#if !defined (MULTITHREAD) && \
- (defined (USE_MMAP_SCOREBOARD) || defined (USE_SHMGET_SCOREBOARD))
+ (defined (USE_MMAP_SCOREBOARD) || defined (USE_SHMGET_SCOREBOARD) ||
defined(USE_POSIX_SCOREBOARD))
#define OPTIMIZE_TIMEOUTS
#endif
diff -ru apache_1.3.22/src/main/http_core.c
apache_1.3.22/src/main/http_core.c
--- apache_1.3.22/src/main/http_core.c Tue Sep 4 12:15:15 2001
+++ apache_1.3.22/src/main/http_core.c Sun Nov 4 05:27:38 2001
@@ -3666,8 +3666,18 @@
&& (!r->header_only || (d->content_md5 & 1))) {
/* we need to protect ourselves in case we die while we've got the
* file mmapped */
+#ifdef NTO
+ /*
+ MAP_NOSYNCFILE is required for mmap() on regular files,
+ MAP_ELF is just a workaround for a bug in 6.1 (fails on MAP_PRIVATE)
+ It appears to give some 10% performance boost nevertheless
+ */
+ mm = mmap(NULL, r->finfo.st_size, PROT_READ,
+ MAP_PRIVATE|MAP_NOSYNCFILE|MAP_ELF, fileno(f), 0);
+#else
mm = mmap(NULL, r->finfo.st_size, PROT_READ, MAP_PRIVATE,
fileno(f), 0);
+#endif
if (mm == (caddr_t)-1) {
ap_log_rerror(APLOG_MARK, APLOG_CRIT, r,
"default_handler: mmap failed: %s", r->filename);
diff -ru apache_1.3.22/src/main/http_main.c
apache_1.3.22/src/main/http_main.c
--- apache_1.3.22/src/main/http_main.c Fri Oct 5 20:21:11 2001
+++ apache_1.3.22/src/main/http_main.c Wed Nov 14 01:35:24 2001
@@ -611,6 +611,8 @@
* on Solaris. Note that pthread mutexes are *NOT* released when a task
* dies ... the task has to free it itself. So we block signals and
* try to be nice about releasing the mutex.
+ *
+ * --- This works on QNX6 as well, even faster perhaps ---
*/
#include <pthread.h>
@@ -645,8 +647,20 @@
static void accept_mutex_init_pthread(pool *p)
{
pthread_mutexattr_t mattr;
- int fd;
+#ifdef MAP_ANON
+#ifndef NOFD
+#define NOFD -1
+#endif
+ accept_mutex = (pthread_mutex_t *) mmap((caddr_t) 0,
sizeof(*accept_mutex),
+ PROT_READ | PROT_WRITE, MAP_ANON|MAP_SHARED, NOFD, 0);
+ if (accept_mutex == (void *) (caddr_t) - 1) {
+ perror("mmap");
+ exit(APEXIT_INIT);
+ }
+#else
+ int fd;
+
fd = open("/dev/zero", O_RDWR);
if (fd == -1) {
perror("open(/dev/zero)");
@@ -659,6 +673,7 @@
exit(APEXIT_INIT);
}
close(fd);
+#endif
if ((errno = pthread_mutexattr_init(&mattr))) {
perror("pthread_mutexattr_init");
exit(APEXIT_INIT);
@@ -2042,30 +2057,13 @@
/*
* POSIX 1003.4 style
*
- * Note 1:
- * As of version 4.23A, shared memory in QNX must reside under /dev/shmem,
- * where no subdirectories allowed.
+ * This code is only used by QNX4 probably, and from version 4.25
+ * QNX4 can actually use USE_MMAP_SCOREBOARD. So this is only still
+ * useful if you're stuck with one of earlier versions. Otherwise
+ * just define USE_MMAP_SCOREBOARD in QNX section of Configure.
*
- * POSIX shm_open() and shm_unlink() will take care about this issue,
- * but to avoid confusion, I suggest to redefine scoreboard file name
- * in httpd.conf to cut "logs/" from it. With default setup actual name
- * will be "/dev/shmem/logs.apache_status".
- *
- * If something went wrong and Apache did not unlinked this object upon
- * exit, you can remove it manually, using "rm -f" command.
- *
- * Note 2:
- * <sys/mman.h> in QNX defines MAP_ANON, but current implementation
- * does NOT support BSD style anonymous mapping. So, the order of
- * conditional compilation is important:
- * this #ifdef section must be ABOVE the next one (BSD style).
- *
- * I tested this stuff and it works fine for me, but if it provides
- * trouble for you, just comment out USE_MMAP_SCOREBOARD in QNX section
- * of ap_config.h
- *
- * June 5, 1997,
- * Igor N. Kovalenko -- infoh@mail.wplus.net
+ * Oct 18, 2001,
+ * Igor Kovalenko -- ikovalenko@sourceforge.net
*/
static void cleanup_shared_mem(void *d)
diff -ru apache_1.3.22/src/modules/standard/mod_mime_magic.c
apache_1.3.22/src/modules/standard/mod_mime_magic.c
--- apache_1.3.22/src/modules/standard/mod_mime_magic.c Wed Oct 3 23:55:01
2001
+++ apache_1.3.22/src/modules/standard/mod_mime_magic.c Thu Oct 18 02:11:56
2001
@@ -255,7 +255,7 @@
static int match(request_rec *, unsigned char *, int);
static int mget(request_rec *, union VALUETYPE *, unsigned char *,
struct magic *, int);
-static int mcheck(request_rec *, union VALUETYPE *, struct magic *);
+static int ap_mcheck(request_rec *, union VALUETYPE *, struct magic *);
static void mprint(request_rec *, union VALUETYPE *, struct magic *);
static int uncompress(request_rec *, int,
@@ -1600,7 +1600,7 @@
/* check if main entry matches */
if (!mget(r, &p, s, m, nbytes) ||
- !mcheck(r, &p, m)) {
+ !ap_mcheck(r, &p, m)) {
struct magic *m_cont;
/*
@@ -1671,7 +1671,7 @@
cont_level = m->cont_level;
}
if (mget(r, &p, s, m, nbytes) &&
- mcheck(r, &p, m)) {
+ ap_mcheck(r, &p, m)) {
/*
* This continuation matched. Print its message, with a
* blank before it if the previous item printed and this
@@ -1845,7 +1845,7 @@
return 1;
}
-static int mcheck(request_rec *r, union VALUETYPE *p, struct magic *m)
+static int ap_mcheck(request_rec *r, union VALUETYPE *p, struct magic *m)
{
register unsigned long l = m->value.l;
register unsigned long v;
@@ -1898,7 +1898,7 @@
default:
/* bogosity, pretend that it just wasn't a match */
ap_log_rerror(APLOG_MARK, APLOG_NOERRNO | APLOG_ERR, r,
- MODNAME ": invalid type %d in mcheck().", m->type);
+ MODNAME ": invalid type %d in ap_mcheck().", m->type);
return 0;
}
@@ -1983,7 +1983,7 @@
/* bogosity, pretend it didn't match */
matched = 0;
ap_log_rerror(APLOG_MARK, APLOG_NOERRNO | APLOG_ERR, r,
- MODNAME ": mcheck: can't happen: invalid relation %d.",
+ MODNAME ": ap_mcheck: can't happen: invalid relation %d.",
m->reln);
break;
}
diff -ru apache_1.3.22/src/support/htdigest.c
apache_1.3.22/src/support/htdigest.c
--- apache_1.3.22/src/support/htdigest.c Wed Jul 25 09:08:04 2001
+++ apache_1.3.22/src/support/htdigest.c Thu Oct 18 01:46:57 2001
@@ -73,7 +73,7 @@
#endif
#include "ap.h"
#include "ap_md5.h"
-#if defined(MPE) || defined(QNX) || defined(WIN32) || defined(__TANDEM) ||
defined(OS390) || defined(BEOS) || defined(BONE)
+#if defined(MPE) || defined(QNX) || defined(NTO) || defined(WIN32) ||
defined(__TANDEM) || defined(OS390) || defined(BEOS) || defined(BONE)
#include <signal.h>
#else
#include <sys/signal.h>
diff -ru apache_1.3.22/src/support/suexec.c
apache_1.3.22/src/support/suexec.c
--- apache_1.3.22/src/support/suexec.c Thu Mar 8 12:11:43 2001
+++ apache_1.3.22/src/support/suexec.c Thu Oct 18 01:54:14 2001
@@ -95,11 +95,11 @@
/*
***********************************************************************
- * There is no initgroups() in QNX, so I believe this is safe :-)
- * Use cc -osuexec -3 -O -mf -DQNX suexec.c to compile.
+ * There is no initgroups() in QNX4 (but there is in QNX6)
+ * On QNX4 compile with cc -osuexec -3 -O -mf -DQNX suexec.c
*
- * May 17, 1997.
- * Igor N. Kovalenko -- infoh@mail.wplus.net
+ * Oct 18, 2001.
+ * Igor Kovalenko -- ikovalenko@sourceforge.net
***********************************************************************
*/
Re: [PATCH] support for QNX6 - take 2
Posted by Igor Kovalenko <Ig...@motorola.com>.
I've checked AIX and HPUX. AIX has /dev/zero, so it might have chance of
using PTHREAD_SERIALIZED_ACCEPT as is, but HP-UX 10.x does not. They however
both work using MAP_ANON, except HP-UX calls it MAP_ANONYMOUS.
The ap_config.h actually defines MAP_ANON if MAP_ANONYMOUS is defined ...
but only if USE_MMAP_SCOREBOARD is defined. It also includes sys/mman.h only
in that case, all other cases when it is needed (USE_POSIX_SCOREBOARD,
USE_MMAP_FILES and mod_static) include it directly in source. Which is sad.
And on HPUX scoreboard is USE_SHMGET_SCOREBOARD, so MAP_whatever would be
undefined at all.
So the code for PTHREAD_SERIALIZED_ACCEPT actually should work at least on
Solaris, HP-UX 10.x and 11.x, AIX 4.3 and QNX6.x. Since it is probably
faster than anything else, I guess it would worth trouble streamlining the
includes/defines a bit. I could do patch for that, perhaps separately from
QNX6 patch but it surely will 'affect other platforms' ;)
Comments?
- igor
----- Original Message -----
From: "Igor Kovalenko" <Ig...@email.mot.com>
To: <de...@httpd.apache.org>
Sent: Thursday, November 15, 2001 2:47 AM
Subject: [PATCH] support for QNX6 - take 2
> Here it goes again. Since last version was not quite welcomed, I'll defend
> it in advance. First, it fixes the undefined 'NOFD' constant. Of course, I
> could also make it totally QNX-specific, but I'd hate to pollute code with
> #ifdef SOME_OS unless necessary and I don't think any of you here would
like
> that either. After all, that is the reason to have OS-specific sections in
> one header which would turn on or off appropriate 'features'.
>
> One of the things objected in last version was addition of MAP_ANON clause
> in PTHREAD_SERIALIZED_ACCEPT code. As I said, I could agree to remove
that,
> but ... the PTHREAD_SERIALIZED_ACCEPT is OS-specific feature to begin
with,
> so it would only matter if it is enabled. Looking at 1.3.22 code, no
> platform actually defines USE_PTHREAD_SERIALIZED_ACCEPT by default and
only
> Solaris, HP-UX (commented out) and AIX appear to even mention it. And the
> implementation appears to be Solaris-specific, in fact. That makes me
think
> I am not hurting anyone by adding MAP_ANON version, I would actually guess
> it might make some platforms be able to use that code. And it still works
on
> Solaris (2.8 has MAP_ANON and it works, 2.5 has no MAP_ANON, so it still
> works old way) and saves unnecessary open() and close() calls.
>
> I don't want to waste any more of anyone's time discussing this, the issue
> is not worth it. If that part of patch is still unaceptable, I am fine
> without it.
>
> Another non-specific change is addition of USE_POSIX_SCOREBOARD condition
to
> the 'optimize timeouts'. That is because 1) it is logical, since it is
same
> thing as USE_MMAP_SCOREBOARD & USE_SHMGET_SCOREBOARD and 2) I wrote the
> USE_POSIX_SCOREBOARD code.
>
> regards,
> - igor
>
> diff -ru apache_1.3.22/src/Configure apache_1.3.22/src/Configure
>
> --- apache_1.3.22/src/Configure Mon Oct 8 14:59:36 2001
>
> +++ apache_1.3.22/src/Configure Thu Oct 18 00:21:27 2001
>
> @@ -554,6 +554,11 @@
>
> CFLAGS="$CFLAGS -DQNX -mf -3"
>
> LIBS="$LIBS -N128k -lsocket -lunix"
>
> DEF_WANTHSREGEX=yes
>
> + ;;
>
> + *-*-qnx6)
>
> + OS='NTO'
>
> + CFLAGS="$CFLAGS -DNTO"
>
> + LIBS="$LIBS -lm -lsocket"
>
> ;;
>
> *-isc4*)
>
> OS='ISC 4'
>
> @@ -1076,6 +1081,14 @@
>
> LDFLAGS_SHLIB='-b -a archive'
>
> LDFLAGS_MOD_SHLIB=$LDFLAGS_SHLIB
>
> ;;
>
> + *-*-qnx6*)
>
> + LD_SHLIB="gcc"
>
> + CFLAGS_SHLIB="-fPIC"
>
> + LDFLAGS_SHLIB="-shared"
>
> + LDFLAGS_MOD_SHLIB=$LDFLAGS_SHLIB
>
> + LDFLAGS_SHLIB_EXPORT="-Wl,-E"
>
> + SHLIB_SUFFIX_DEPTH=0
>
> + ;;
>
> *-linux1)
>
> CFLAGS_SHLIB="-fpic"
>
> LDFLAGS_SHLIB="-Bshareable"
>
> diff -ru apache_1.3.22/src/helpers/GuessOS
apache_1.3.22/src/helpers/GuessOS
>
> --- apache_1.3.22/src/helpers/GuessOS Mon Oct 8 09:49:33 2001
>
> +++ apache_1.3.22/src/helpers/GuessOS Thu Oct 18 00:30:28 2001
>
> @@ -181,12 +181,22 @@
>
> echo "${MACHINE}-tandem-oss"; exit 0;
>
> ;;
>
>
> - QNX:*)
>
> - if [ "$VERSION" -gt 422 ]; then
>
> - echo "${MACHINE}-qssl-qnx32"
>
> - else
>
> - echo "${MACHINE}-qssl-qnx"
>
> + *nto:* | QNX:*)
>
> + if test "${MACHINE}" = "x86pc" || test "${MACHINE}" = "PCI" ; then
>
> + MACHINE="pc"
>
> fi
>
> + case $RELEASE in
>
> + [0-9]*)
>
> + echo "${MACHINE}-qssl-qnx6"
>
> + ;;
>
> + [A-Z]*)
>
> + if [ "$VERSION" -gt 422 ]; then
>
> + echo "${MACHINE}-qssl-qnx32"
>
> + else
>
> + echo "${MACHINE}-qssl-qnx"
>
> + fi
>
> + ;;
>
> + esac
>
> exit 0
>
> ;;
>
>
> diff -ru apache_1.3.22/src/include/ap_config.h
> apache_1.3.22/src/include/ap_config.h
>
> --- apache_1.3.22/src/include/ap_config.h Mon Oct 8 09:49:33 2001
>
> +++ apache_1.3.22/src/include/ap_config.h Sun Oct 21 00:15:14 2001
>
> @@ -745,6 +745,36 @@
>
> #define NET_SIZE_T socklen_t
>
> #endif
>
>
> +#elif defined(NTO)
>
> +
>
> +#define HAVE_MMAP 1
>
> +#define HAVE_GMTOFF 1
>
> +#define HAVE_SYSLOG 1
>
> +
>
> +#define USE_MMAP_SCOREBOARD
>
> +#define USE_MMAP_FILES 1
>
> +#define USE_PTHREAD_SERIALIZED_ACCEPT
>
> +
>
> +#undef MAP_TMPFILE
>
> +#undef NO_TIMES
>
> +
>
> +/* needs prototypes for execXX */
>
> +#define NEED_PROCESS_H
>
> +
>
> +/*
>
> + Careful here.
>
> + Things have changed a lot between 6.0 and 6.1
>
> + We'll assume 6.1+ is default
>
> +*/
>
> +/* loader does not support #! before 6.1 */
>
> +/* #define NEED_HASHBANG_EMUL */
>
> +/* kernel does not implement getrlimit/setrlimit before 6.1 */
>
> +/* #undef HAVE_SYS_RESOURCE_H */
>
> +/* tiny stack does not implement SO_LINGER */
>
> +/* #undef USE_SO_LINGER */
>
> +/* lingering close is broken with tiny TCP stack */
>
> +/* #define NO_LINGCLOSE */
>
> +
>
> #elif defined(QNX)
>
> #ifndef crypt
>
> char *crypt(const char *pw, const char *salt);
>
> @@ -1156,7 +1186,7 @@
>
> #undef NSIG
>
> #endif
>
> #include <errno.h>
>
> -#if !defined(QNX) && !defined(CONVEXOS11) && !defined(NEXT) &&
> !defined(TPF) && !defined(NETWARE) && !defined(MPE)
>
> +#if !defined(QNX) && !defined(NTO) && !defined(CONVEXOS11) &&
> !defined(NEXT) && !defined(TPF) && !defined(NETWARE) && !defined(MPE)
>
> #include <memory.h>
>
> #endif
>
>
> diff -ru apache_1.3.22/src/include/httpd.h
apache_1.3.22/src/include/httpd.h
>
> --- apache_1.3.22/src/include/httpd.h Mon Oct 8 21:56:05 2001
>
> +++ apache_1.3.22/src/include/httpd.h Thu Oct 18 01:06:10 2001
>
> @@ -1189,7 +1189,7 @@
>
> * also not using a scoreboard file
>
> */
>
> #if !defined (MULTITHREAD) && \
>
> - (defined (USE_MMAP_SCOREBOARD) || defined (USE_SHMGET_SCOREBOARD))
>
> + (defined (USE_MMAP_SCOREBOARD) || defined (USE_SHMGET_SCOREBOARD) ||
> defined(USE_POSIX_SCOREBOARD))
>
> #define OPTIMIZE_TIMEOUTS
>
> #endif
>
>
> diff -ru apache_1.3.22/src/main/http_core.c
> apache_1.3.22/src/main/http_core.c
>
> --- apache_1.3.22/src/main/http_core.c Tue Sep 4 12:15:15 2001
>
> +++ apache_1.3.22/src/main/http_core.c Sun Nov 4 05:27:38 2001
>
> @@ -3666,8 +3666,18 @@
>
> && (!r->header_only || (d->content_md5 & 1))) {
>
> /* we need to protect ourselves in case we die while we've got the
>
> * file mmapped */
>
> +#ifdef NTO
>
> + /*
>
> + MAP_NOSYNCFILE is required for mmap() on regular files,
>
> + MAP_ELF is just a workaround for a bug in 6.1 (fails on MAP_PRIVATE)
>
> + It appears to give some 10% performance boost nevertheless
>
> + */
>
> + mm = mmap(NULL, r->finfo.st_size, PROT_READ,
>
> + MAP_PRIVATE|MAP_NOSYNCFILE|MAP_ELF, fileno(f), 0);
>
> +#else
>
> mm = mmap(NULL, r->finfo.st_size, PROT_READ, MAP_PRIVATE,
>
> fileno(f), 0);
>
> +#endif
>
> if (mm == (caddr_t)-1) {
>
> ap_log_rerror(APLOG_MARK, APLOG_CRIT, r,
>
> "default_handler: mmap failed: %s", r->filename);
>
> diff -ru apache_1.3.22/src/main/http_main.c
> apache_1.3.22/src/main/http_main.c
>
> --- apache_1.3.22/src/main/http_main.c Fri Oct 5 20:21:11 2001
>
> +++ apache_1.3.22/src/main/http_main.c Wed Nov 14 01:35:24 2001
>
> @@ -611,6 +611,8 @@
>
> * on Solaris. Note that pthread mutexes are *NOT* released when a task
>
> * dies ... the task has to free it itself. So we block signals and
>
> * try to be nice about releasing the mutex.
>
> + *
>
> + * --- This works on QNX6 as well, even faster perhaps ---
>
> */
>
>
> #include <pthread.h>
>
> @@ -645,8 +647,20 @@
>
> static void accept_mutex_init_pthread(pool *p)
>
> {
>
> pthread_mutexattr_t mattr;
>
> - int fd;
>
>
> +#ifdef MAP_ANON
>
> +#ifndef NOFD
>
> +#define NOFD -1
>
> +#endif
>
> + accept_mutex = (pthread_mutex_t *) mmap((caddr_t) 0,
> sizeof(*accept_mutex),
>
> + PROT_READ | PROT_WRITE, MAP_ANON|MAP_SHARED, NOFD, 0);
>
> + if (accept_mutex == (void *) (caddr_t) - 1) {
>
> + perror("mmap");
>
> + exit(APEXIT_INIT);
>
> + }
>
> +#else
>
> + int fd;
>
> +
>
> fd = open("/dev/zero", O_RDWR);
>
> if (fd == -1) {
>
> perror("open(/dev/zero)");
>
> @@ -659,6 +673,7 @@
>
> exit(APEXIT_INIT);
>
> }
>
> close(fd);
>
> +#endif
>
> if ((errno = pthread_mutexattr_init(&mattr))) {
>
> perror("pthread_mutexattr_init");
>
> exit(APEXIT_INIT);
>
> @@ -2042,30 +2057,13 @@
>
> /*
>
> * POSIX 1003.4 style
>
> *
>
> - * Note 1:
>
> - * As of version 4.23A, shared memory in QNX must reside under
/dev/shmem,
>
> - * where no subdirectories allowed.
>
> + * This code is only used by QNX4 probably, and from version 4.25
>
> + * QNX4 can actually use USE_MMAP_SCOREBOARD. So this is only still
>
> + * useful if you're stuck with one of earlier versions. Otherwise
>
> + * just define USE_MMAP_SCOREBOARD in QNX section of Configure.
>
> *
>
> - * POSIX shm_open() and shm_unlink() will take care about this issue,
>
> - * but to avoid confusion, I suggest to redefine scoreboard file name
>
> - * in httpd.conf to cut "logs/" from it. With default setup actual name
>
> - * will be "/dev/shmem/logs.apache_status".
>
> - *
>
> - * If something went wrong and Apache did not unlinked this object upon
>
> - * exit, you can remove it manually, using "rm -f" command.
>
> - *
>
> - * Note 2:
>
> - * <sys/mman.h> in QNX defines MAP_ANON, but current implementation
>
> - * does NOT support BSD style anonymous mapping. So, the order of
>
> - * conditional compilation is important:
>
> - * this #ifdef section must be ABOVE the next one (BSD style).
>
> - *
>
> - * I tested this stuff and it works fine for me, but if it provides
>
> - * trouble for you, just comment out USE_MMAP_SCOREBOARD in QNX section
>
> - * of ap_config.h
>
> - *
>
> - * June 5, 1997,
>
> - * Igor N. Kovalenko -- infoh@mail.wplus.net
>
> + * Oct 18, 2001,
>
> + * Igor Kovalenko -- ikovalenko@sourceforge.net
>
> */
>
>
> static void cleanup_shared_mem(void *d)
>
> diff -ru apache_1.3.22/src/modules/standard/mod_mime_magic.c
> apache_1.3.22/src/modules/standard/mod_mime_magic.c
>
> --- apache_1.3.22/src/modules/standard/mod_mime_magic.c Wed Oct 3 23:55:01
> 2001
>
> +++ apache_1.3.22/src/modules/standard/mod_mime_magic.c Thu Oct 18
02:11:56
> 2001
>
> @@ -255,7 +255,7 @@
>
> static int match(request_rec *, unsigned char *, int);
>
> static int mget(request_rec *, union VALUETYPE *, unsigned char *,
>
> struct magic *, int);
>
> -static int mcheck(request_rec *, union VALUETYPE *, struct magic *);
>
> +static int ap_mcheck(request_rec *, union VALUETYPE *, struct magic *);
>
> static void mprint(request_rec *, union VALUETYPE *, struct magic *);
>
>
> static int uncompress(request_rec *, int,
>
> @@ -1600,7 +1600,7 @@
>
>
> /* check if main entry matches */
>
> if (!mget(r, &p, s, m, nbytes) ||
>
> - !mcheck(r, &p, m)) {
>
> + !ap_mcheck(r, &p, m)) {
>
> struct magic *m_cont;
>
>
> /*
>
> @@ -1671,7 +1671,7 @@
>
> cont_level = m->cont_level;
>
> }
>
> if (mget(r, &p, s, m, nbytes) &&
>
> - mcheck(r, &p, m)) {
>
> + ap_mcheck(r, &p, m)) {
>
> /*
>
> * This continuation matched. Print its message, with a
>
> * blank before it if the previous item printed and this
>
> @@ -1845,7 +1845,7 @@
>
> return 1;
>
> }
>
>
> -static int mcheck(request_rec *r, union VALUETYPE *p, struct magic *m)
>
> +static int ap_mcheck(request_rec *r, union VALUETYPE *p, struct magic *m)
>
> {
>
> register unsigned long l = m->value.l;
>
> register unsigned long v;
>
> @@ -1898,7 +1898,7 @@
>
> default:
>
> /* bogosity, pretend that it just wasn't a match */
>
> ap_log_rerror(APLOG_MARK, APLOG_NOERRNO | APLOG_ERR, r,
>
> - MODNAME ": invalid type %d in mcheck().", m->type);
>
> + MODNAME ": invalid type %d in ap_mcheck().", m->type);
>
> return 0;
>
> }
>
>
> @@ -1983,7 +1983,7 @@
>
> /* bogosity, pretend it didn't match */
>
> matched = 0;
>
> ap_log_rerror(APLOG_MARK, APLOG_NOERRNO | APLOG_ERR, r,
>
> - MODNAME ": mcheck: can't happen: invalid relation %d.",
>
> + MODNAME ": ap_mcheck: can't happen: invalid relation %d.",
>
> m->reln);
>
> break;
>
> }
>
> diff -ru apache_1.3.22/src/support/htdigest.c
> apache_1.3.22/src/support/htdigest.c
>
> --- apache_1.3.22/src/support/htdigest.c Wed Jul 25 09:08:04 2001
>
> +++ apache_1.3.22/src/support/htdigest.c Thu Oct 18 01:46:57 2001
>
> @@ -73,7 +73,7 @@
>
> #endif
>
> #include "ap.h"
>
> #include "ap_md5.h"
>
> -#if defined(MPE) || defined(QNX) || defined(WIN32) || defined(__TANDEM)
||
> defined(OS390) || defined(BEOS) || defined(BONE)
>
> +#if defined(MPE) || defined(QNX) || defined(NTO) || defined(WIN32) ||
> defined(__TANDEM) || defined(OS390) || defined(BEOS) || defined(BONE)
>
> #include <signal.h>
>
> #else
>
> #include <sys/signal.h>
>
> diff -ru apache_1.3.22/src/support/suexec.c
> apache_1.3.22/src/support/suexec.c
>
> --- apache_1.3.22/src/support/suexec.c Thu Mar 8 12:11:43 2001
>
> +++ apache_1.3.22/src/support/suexec.c Thu Oct 18 01:54:14 2001
>
> @@ -95,11 +95,11 @@
>
>
> /*
>
> ***********************************************************************
>
> - * There is no initgroups() in QNX, so I believe this is safe :-)
>
> - * Use cc -osuexec -3 -O -mf -DQNX suexec.c to compile.
>
> + * There is no initgroups() in QNX4 (but there is in QNX6)
>
> + * On QNX4 compile with cc -osuexec -3 -O -mf -DQNX suexec.c
>
> *
>
> - * May 17, 1997.
>
> - * Igor N. Kovalenko -- infoh@mail.wplus.net
>
> + * Oct 18, 2001.
>
> + * Igor Kovalenko -- ikovalenko@sourceforge.net
>
> ***********************************************************************
>
> */
>
>
>
>
Re: [PATCH] support for QNX6 - take 2
Posted by Igor Kovalenko <Ig...@motorola.com>.
Jeff Trawick wrote:
>
> "Igor Kovalenko" <Ig...@motorola.com> writes:
>
> > One of the things objected in last version was addition of MAP_ANON clause
> > in PTHREAD_SERIALIZED_ACCEPT code. As I said, I could agree to remove that,
> > but ... the PTHREAD_SERIALIZED_ACCEPT is OS-specific feature to begin with,
> > so it would only matter if it is enabled. Looking at 1.3.22 code, no
> > platform actually defines USE_PTHREAD_SERIALIZED_ACCEPT by default and only
> > Solaris, HP-UX (commented out) and AIX appear to even mention it. And the
> > implementation appears to be Solaris-specific, in fact.
>
> FYI... AIX uses it by default, but by setting
> -DUSE_PTHREAD_SERIALIZED_ACCEPT in src/Configure.
>
Thanks, i see now. But as I said, MAP_ANON code works on AIX.
> The comments in the code about it being Solaris-specific aren't
> actually true. Some systems have that form of shared memory, some
> systems have that flavor of pthread_mutex, and where those sets of
> systems overlap that code ought to work :)
That's what I thought, so I am somewhat surprized nobody cares to try.
This probably is the fastest way to handle serialized accept.
> I wish I had a chance now to review your patch but alas... :(
So how do I get that patch in? Do I need to jump through hoops and keep
bothering here or what?
regards,
- igor
Re: [PATCH] support for QNX6 - take 2
Posted by Jeff Trawick <tr...@attglobal.net>.
"Igor Kovalenko" <Ig...@motorola.com> writes:
> One of the things objected in last version was addition of MAP_ANON clause
> in PTHREAD_SERIALIZED_ACCEPT code. As I said, I could agree to remove that,
> but ... the PTHREAD_SERIALIZED_ACCEPT is OS-specific feature to begin with,
> so it would only matter if it is enabled. Looking at 1.3.22 code, no
> platform actually defines USE_PTHREAD_SERIALIZED_ACCEPT by default and only
> Solaris, HP-UX (commented out) and AIX appear to even mention it. And the
> implementation appears to be Solaris-specific, in fact.
FYI... AIX uses it by default, but by setting
-DUSE_PTHREAD_SERIALIZED_ACCEPT in src/Configure.
The comments in the code about it being Solaris-specific aren't
actually true. Some systems have that form of shared memory, some
systems have that flavor of pthread_mutex, and where those sets of
systems overlap that code ought to work :)
I wish I had a chance now to review your patch but alas... :(
--
Jeff Trawick | trawick@attglobal.net | PGP public key at web site:
http://www.geocities.com/SiliconValley/Park/9289/
Born in Roswell... married an alien...