Re: [PATCH rtems-libbsd 1/2] racoon/session: Honor file descriptor maximum

[Christian MAUDERER]

Hello Chris,

thanks for the review.
Am 26.02.21 um 19:04 schrieb Chris Johns:
On 26/2/21 2:01 am, Christian Mauderer wrote:
Dynamically allocate a big enough file descriptor set for select(). A
better solution would be to use kqueue() instead of select().
---
  .../racoon/rtems-bsd-racoon-session-data.h    |  6 +--
  ipsec-tools/src/racoon/session.c              | 40 +++++++++++++++++++
  2 files changed, 43 insertions(+), 3 deletions(-)

diff --git a/ipsec-tools/src/racoon/rtems-bsd-racoon-session-data.h 
b/ipsec-tools/src/racoon/rtems-bsd-racoon-session-data.h
index b869a1518..196107a35 100644
--- a/ipsec-tools/src/racoon/rtems-bsd-racoon-session-data.h
+++ b/ipsec-tools/src/racoon/rtems-bsd-racoon-session-data.h
@@ -2,11 +2,11 @@
  #include <rtems/linkersets.h>
  #include "rtems-bsd-racoon-data.h"
  /* session.c */
-RTEMS_LINKER_RWSET_CONTENT(bsd_prog_racoon, static fd_set active_mask);
-RTEMS_LINKER_RWSET_CONTENT(bsd_prog_racoon, static fd_set preset_mask);
+RTEMS_LINKER_RWSET_CONTENT(bsd_prog_racoon, static _types_fd_set 
*allocated_active_mask);
+RTEMS_LINKER_RWSET_CONTENT(bsd_prog_racoon, static _types_fd_set 
*allocated_preset_mask);
  RTEMS_LINKER_RWSET_CONTENT(bsd_prog_racoon, static int nfds);
  RTEMS_LINKER_RWSET_CONTENT(bsd_prog_racoon, static int signals[]);
  RTEMS_LINKER_RWSET_CONTENT(bsd_prog_racoon, static sig_atomic_t volatile 
volatile sigreq[]);
-RTEMS_LINKER_RWSET_CONTENT(bsd_prog_racoon, static struct fd_monitor 
fd_monitors[]);
+RTEMS_LINKER_RWSET_CONTENT(bsd_prog_racoon, static struct fd_monitor 
*allocated_fd_monitors);
  RTEMS_LINKER_RWSET_CONTENT(bsd_prog_racoon, static struct fd_monitor_list 
fd_monitor_tree[]);
  RTEMS_LINKER_RWSET_CONTENT(bsd_prog_racoon, static struct sched scflushsa);
diff --git a/ipsec-tools/src/racoon/session.c b/ipsec-tools/src/racoon/session.c
index 65124c15e..90120c761 100644
--- a/ipsec-tools/src/racoon/session.c
+++ b/ipsec-tools/src/racoon/session.c
@@ -65,6 +65,10 @@
  #include <sys/stat.h>
  #include <paths.h>
  #include <err.h>
+#ifdef __rtems__
+#include <sys/param.h>
+#include <rtems/libio_.h>
+#endif /* __rtems__ */
  
  #include <netinet/in.h>
  #include <resolv.h>
@@ -123,8 +127,16 @@ static void check_sigreq __P((void));
  static void check_flushsa __P((void));
  static int close_sockets __P((void));
  
+#ifndef __rtems__
  static fd_set preset_mask, active_mask;
  static struct fd_monitor fd_monitors[FD_SETSIZE];
+#else /* __rtems__ */
+static fd_set *allocated_preset_mask, *allocated_active_mask;
+static struct fd_monitor *allocated_fd_monitors;
+#define preset_mask (*allocated_preset_mask)
+#define active_mask (*allocated_active_mask)
+#define fd_monitors (allocated_fd_monitors)
+#endif /* __rtems__ */
  static TAILQ_HEAD(fd_monitor_list, fd_monitor) 
fd_monitor_tree[NUM_PRIORITIES];
  static int nfds = 0;
  
@@ -134,7 +146,11 @@ static struct sched scflushsa = SCHED_INITIALIZER();
  void
  monitor_fd(int fd, int (*callback)(void *, int), void *ctx, int priority)
  {
+#ifndef __rtems__
        if (fd < 0 || fd >= FD_SETSIZE) {
+#else /* __rtems__ */
+       if (fd < 0 || fd >= rtems_libio_number_iops) {
+#endif /* __rtems__ */
                plog(LLV_ERROR, LOCATION, NULL, "fd_set overrun");
                exit(1);
        }
@@ -158,7 +174,11 @@ monitor_fd(int fd, int (*callback)(void *, int), void 
*ctx, int priority)
  void
  unmonitor_fd(int fd)
  {
+#ifndef __rtems__
        if (fd < 0 || fd >= FD_SETSIZE) {
+#else /* __rtems__ */
+       if (fd < 0 || fd >= rtems_libio_number_iops) {
+#endif /* __rtems__ */
                plog(LLV_ERROR, LOCATION, NULL, "fd_set overrun");
                exit(1);
        }
@@ -186,7 +206,22 @@ session(void)
        struct fd_monitor *fdm;
  
  	nfds = 0;
+#ifndef __rtems__
        FD_ZERO(&preset_mask);
+#else /* __rtems__ */
+       allocated_preset_mask = calloc(sizeof(fd_set),
+           howmany(rtems_libio_number_iops, sizeof(fd_set) * 8));

Does `maxfiles` work here?


To be honest: I'm not sure.

According to the comment in file.h:

extern int maxfiles; /* kernel limit on number of open files */

Sounds like it _can_ be the same as the maximum file number but doesn't 
have to. I didn't find where we implement it. It's declared as an extern 
int maxfiles but I didn't find any definition. I found it only in libbsd 
in freebsd/sys/kern/uipc_socket.c where it is defined like follows:

#define maxfiles rtems_libio_number_iops

So question is: Where and how is maxfiles defined?

+       if (allocated_preset_mask == NULL)
+               errx(1, "failed to allocate preset_mask");
+       allocated_active_mask = calloc(sizeof(fd_set),
+           howmany(rtems_libio_number_iops, sizeof(fd_set) * 8));
+       if (allocated_active_mask == NULL)
+               errx(1, "failed to allocate active_mask");
+       allocated_fd_monitors = calloc(
+           rtems_libio_number_iops, sizeof(struct fd_monitor));
+       if (allocated_fd_monitors == NULL)
+               errx(1, "failed to allocate fd_monitors");

At the core of this issue is the rotating fd allocation that we have in libio. A
report from a FreeBSD machine I have is:

$ sysctl -a | grep maxfiles
kern.maxfiles: 1037243
kern.maxfilesperproc: 933516

I doubt a select process in FreeBSD needs a select array with that many bits. I
have added similar code else where but I am wondering if this is really what we
want to do. A side effect for us is the stack usage is not bounded and that is a
problem. I think our newlib based max setting is too small.

I think we have to distinguish between FreeBSD kernel space and user 
space. If I have seen it correctly, FreeBSD uses kqueues or maybe 
sometimes poll instead of select in kernel space most of the time. That 
avoids the problem with big file numbers.

Default for FreeBSD seems to be a FD_SETSIZE of 1024. See the Notes 
section in the select man page:

https://www.freebsd.org/cgi/man.cgi?query=select&sektion=2&manpath=freebsd-release-ports#end

I think select is used in user space applications most of the time. I 
would guess that FreeBSD has some file number mapping between kernel and 
user space that would allow every application to open 1024 files.


FYI I have a major set of changes to libbsd that enables FreeBSD descriptor
support and moves the libio support to specific interfaces.

OK. When do you plan to add that and to which branches? I would like to 
add this patch set to 5 sooner or later because it fixes a potential bug 
in racoon for that branch too.

The changes I have
made are in tarballs in nfsv4 in my ftp area. The support includes the mapping
of libio descriptors to FB ones. The benefit of these changes is VFS and FreeBSD
file system support.

FreeBSD file system support sounds like a great feature. It would allow 
access to a whole bunch of stable and well tested file system 
implementations.

Best regards

Christian


Chris


--
--------------------------------------------
embedded brains GmbH
Herr Christian MAUDERER
Dornierstr. 4
82178 Puchheim
Germany
email: christian.maude...@embedded-brains.de
phone: +49-89-18 94 741 - 18
fax:   +49-89-18 94 741 - 08

Registergericht: Amtsgericht München
Registernummer: HRB 157899
Vertretungsberechtigte Geschäftsführer: Peter Rasmussen, Thomas Dörfler
Unsere Datenschutzerklärung finden Sie hier:
https://embedded-brains.de/datenschutzerklaerung/
_______________________________________________
devel mailing list
devel@rtems.org
http://lists.rtems.org/mailman/listinfo/devel