Hello Chris, Am 19.03.21 um 02:11 schrieb Chris Johns:
On 3/3/21 7:41 pm, Christian MAUDERER wrote:Hello Chris,Am 03.03.21 um 02:17 schrieb Chris Johns:On 2/3/21 7:26 pm, Christian MAUDERER wrote:Hello Chris, Am 02.03.21 um 01:03 schrieb Chris Johns:On 1/3/21 7:24 pm, Christian MAUDERER wrote: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 */Yes.Sounds like it _can_ be the same as the maximum file number but doesn't have to.I think we need to have them be the same value.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_iopsAh OK. I knew it had been assigned somewhere and yes it looks like it is local to that file.So question is: Where and how is maxfiles defined?I have provided a value in the rtemsbsd init file as part of the set og globals we need to maintained.Somehow I missed that. Where can I find it?Again sorry, I was in a rush and I was not clear. I have add this in my new changes for NFSv4 support and this is what I sorted out.Ah, OK. In that case I'll keep the rtems_libio_number_iops at the moment if it is OK for you.Yes that is fine.
Thanks.
+ 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.Yes kqueue is nice but we need to support existing code and this is a primary role libbsd needs to perform. The lack is signals in libbsd is another source of issue.I didn't want to say that we should not support select. I only wanted to say that FreeBSD most likely just avoids that problem in kernel space so that they don't have to support a select with 1037243 files. In user space they limit a select to 1024 files.Yes they do. My suggestion of us using 1024 is FreeBSD applications are fine with this then they should be portable to RTEMS and libbsd and the need for us to patch code is removed.1024 for a select is quite much for an embedded system. It means that we will need 128 bytes of stack for each select call in the future. It's a quite big change.The select call in LibBSD double this internally.
Doesn't make it better.
Another possibility would be 256 (32 bytes) like some other *BSD do. Maybe we should start a new discussion thread about the right FD_SETSIZE so that more developers get involved?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#endYes.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.Yes. There are other checks and arrays in the kernel's select that handle the size of 1024 or smaller. We should align ourselves with FreeBSD. A select call uses a lot of stack.I have no problem if we align more with FreeBSD. But I'm not sure how much effort that would be. And we have to be careful not to increase the memory usage for small targets too much.We need to find a path between that sort of space optimization and the functionality libbsd brings. The select call is costly in terms of stack size as the kern_select call has: /* * The magic 2048 here is chosen to be just enough for FD_SETSIZE * infds with the new FD_SETSIZE of 1024, and more than enough for * FD_SETSIZE infds, outfds and exceptfds with the old FD_SETSIZE * of 256. */ fd_mask s_selbits[howmany(2048, NFDBITS)]; I do not know if this means we need to be more concerned about the size of less.Hm. They could have used FD_SETSIZE instead of the magic value with a comment...It might pay to check the surrounding code a litte more before raising something like that on the FreeBSds hackers list. There are subtle optimisations happening to make this call perform.
I didn't plan to rise it on the FreeBSD mailing list. If I plan to fight against existing magic numbers I'll start in RTEMS before I fight them in FreeBSD ;-)
I think a big problem for RTEMS is that we have shared file numbers for all processes. This makes 1024 a lot less compared to starting from 0 for every process.We are a single process so I do not follow. We are no different to other processes. We have lived with a small file set size for years until libio was changed to cycle over the fd numbers
We started to include libbsd with some applications that are more used to desktop or server environments like wpa supplicant or racoon. These applications originally didn't expect to run in a single process environment but have an own process for each of them.
I'm just not sure whether increasing FD_SETSIZE is the best solution for that problem. It only moves it so that it does happen in fewer applications at the cost that all applications that use select and can work with the current limit will need more stack.
That means: I wouldn't object to such a patch but I don't want to be the one creating and suggesting it.
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 am looking 6-freebsd-12 and master. Thses changes are a major rework and not suitable for 5.I assumed that.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.Sorry I missed this for 5. I will need to think about this on 5. I do not think it is a problem but I am not sure about it on 6 and beyond. I would prefer we consider a better fix for select that the limited 64 descriptors we currently have.Maybe we should take a multi step approach: Use the workaround like I suggested for 5 and (at the moment) for master. A similar workaround is already used in quite some other locations in libbsd. So it's not a new method. Start a ticket and a discussion for a better solution.Sounds a good approach. I am fine with the changes because it is what we have to do and this was not meant to halt the progress, rather I wanted to have the discussion and raise the concerns.OK. Thanks.
If you don't add further objections, I'll create a ticket for the patch and apply it in the next days. I'll create an extra ticket and mail thread for applying it to 5 afterwards.
Best regards Christian
Like I said: Maybe we should create a ticket and start an extra discussion thread to get more ideas?I'm not sure yet how that could look like. Some impressions from other systems: - FreeBSD, OpenBSD, NetBSD and DragonFly BSD are using the same method we use (or most likely the other way round) only with different FD_SETSIZEs: https://cgit.freebsd.org/src/tree/sys/sys/select.h https://github.com/NetBSD/src/blob/trunk/sys/sys/fd_set.h https://github.com/openbsd/src/blob/master/sys/sys/select.h https://gitweb.dragonflybsd.org/dragonfly.git/blob/HEAD:/sys/sys/_fd_set.h - Linux Kernels nolibc.h used in Tests does the same again (GPL or MIT): https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/include/nolibc/nolibc.h#n2522 - Apple is the same again. I'm not sure about the Apple license therefore no link here so I don't spoil anyone beneath myself. There are others that could be worth a look like Haiku, Musl-Libc, and so on. But my general impression is that there is no much better solution at least in the bigger systems except for avoiding select or setting a bigger FD_SETSIZE. The most common ones seem to be 256 or 1024. But that only moves the problem a bit further away.Yes select does not offer a great range of solutions. As Sebastian has stated a number of times kqueue does and this should be encouraged. And kqueue by passes the signal issue that seems to acompany select, eg timer signals. I see the central quesiton being the role libbsd plays and I feel it is to aid porting code to RTEMS and easily as possible. New code written for RTEMS should avoid select.Agreed: Select is evil ;-) Just kidding: It's clear that we have to handle application that use select too due to portability reasons. I'm just not sure what a good solution would be. It feels like every solution is a bad one: If we increase FD_SETSIZE to typical values on desktop systems (256 or 1024), we will need a lot of stack and only cover a part of the applications. If we change every location in the libraries it increases maintenance effort.We also keep in mind if an application uses select _and_ signals it will be broken on LibBSD. PTP2 was like that. 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
