[BK SUMMARY] fix NGROUPS hard limit (resend)

[BK SUMMARY] fix NGROUPS hard limit (resend)

[Timothy Hockin]

Linus,

This patchset removes the hard NGROUPS limit.  It has been in use in a similar
form (but with a sysctl-set limit) on our systems for some time.

I have a separate patch to convert XFS to the generic qsort(), which I will
bounce to SGI if/when this gets pulled.

There is a small change needed for glibc, and I will send that patch to the
glibc people if/when this gets pulled.

Lastly, this does not fixup all the architectures.  I have other patchsets for
that, which need to be reviewed by arch maintainers.

Tim


Please do a

	bk pull http://suncobalt.bkbits.net/ngroups-2.5

This will update the following files:

 include/linux/nfsiod.h         |   52 -----------
 fs/nfsd/auth.c                 |   11 +-
 fs/proc/array.c                |    2 
 include/asm-i386/param.h       |    4 
 include/linux/init_task.h      |    1 
 include/linux/kernel.h         |    5 +
 include/linux/limits.h         |    3 
 include/linux/sched.h          |    3 
 include/linux/sunrpc/svcauth.h |    3 
 kernel/exit.c                  |    6 +
 kernel/fork.c                  |    4 
 kernel/sys.c                   |   88 +++++++++++++++-----
 kernel/uid16.c                 |   63 ++++++++++----
 lib/Makefile                   |    5 -
 lib/bsearch.c                  |   49 +++++++++++
 lib/qsort.c                    |  180 +++++++++++++++++++++++++++++++++++++++++
 net/sunrpc/svcauth_unix.c      |    4 
 17 files changed, 380 insertions(+), 103 deletions(-)

through these ChangeSets (diffs in separate email):

<thockin@freakshow.cobalt.com> (02/10/21 1.812)
   no one references nfsiod.h anymore - nix it.

<thockin@freakshow.cobalt.com> (02/10/21 1.811)
   fix usage of NGROUPS in nfsd and svcauth

<thockin@freakshow.cobalt.com> (02/10/21 1.810)
   Remove the limit of 32 groups.  We now have a per-task, dynamic array of
   groups, which is kept sorted and refcounted.
   
   This ChangeSet incorporates all the core functionality. but does not fixup
   all the incorrect usages of groups.  That is in a seperate ChangeSet.

<thockin@freakshow.cobalt.com> (02/10/21 1.809)
   Add generic qsort() and bsearch(): qsort() from BSD, bsearch() from glibc

Re: [BK SUMMARY] fix NGROUPS hard limit (resend)

[Alan Cox]

On Tue, 2002-10-29 at 19:32, Timothy Hockin wrote:
> Linus,
> 
> This patchset removes the hard NGROUPS limit.  It has been in use in a similar
> form (but with a sysctl-set limit) on our systems for some time.
> 
> I have a separate patch to convert XFS to the generic qsort(), which I will
> bounce to SGI if/when this gets pulled.

What is the worst case stack usage of your qsort ?

Re: [BK SUMMARY] fix NGROUPS hard limit (resend)

[Tim Hockin]

Alan Cox wrote:
> On Tue, 2002-10-29 at 19:32, Timothy Hockin wrote:
> 
>>Linus,
>>
>>This patchset removes the hard NGROUPS limit.  It has been in use in a similar
>>form (but with a sysctl-set limit) on our systems for some time.
>>
>>I have a separate patch to convert XFS to the generic qsort(), which I will
>>bounce to SGI if/when this gets pulled.
> 
> 
> What is the worst case stack usage of your qsort ?

Unfortunately, I haven't done an analysis of this algorithm, but quick 
empirical tests for random, reversed, and sorted data show stack usage 
to be about 50% less than glibc's qsort() for large data sets.  We were 
using the qsort() as exists in XFS, but when discussing with Cristoph, 
he asked that we use this qsort() implementation instead.  It seems to 
perform markedly better for large sets, too.




-- 
Tim Hockin
Systems Software Engineer
Sun Microsystems, Linux Kernel Engineering
thockin@sun.com

[BK SUMMARY] fix NGROUPS hard limit (resend)

[Timothy Hockin]

Linus,

This patchset removes the hard NGROUPS limit.  It has been in use in a similar
form (but with a sysctl-set limit) on our systems for some time.

The last changeset is of questionable value - nfsiod is not included by or
referenced by anything.  Perhaps it should just be removed?

I have a separate patch to convert XFS to the generic qsort(), which I will
bounce to SGI if/when this gets pulled.

There is a small change needed for glibc, and I will send that patch to the
glibc people if/when this gets pulled.

Lastly, this does not fixup all the architectures.  I have other patchsets for
that, which need to be reviewed by arch maintainers.

Tim


Please do a

	bk pull http://suncobalt.bkbits.net/ngroups-2.5

This will update the following files:

 include/linux/nfsiod.h         |   52 -----------
 fs/nfsd/auth.c                 |   11 +-
 fs/proc/array.c                |    2 
 include/asm-i386/param.h       |    4 
 include/linux/init_task.h      |    1 
 include/linux/kernel.h         |    5 +
 include/linux/limits.h         |    3 
 include/linux/sched.h          |    3 
 include/linux/sunrpc/svcauth.h |    3 
 kernel/exit.c                  |    6 +
 kernel/fork.c                  |    4 
 kernel/sys.c                   |   88 +++++++++++++++-----
 kernel/uid16.c                 |   63 ++++++++++----
 lib/Makefile                   |    5 -
 lib/bsearch.c                  |   49 +++++++++++
 lib/qsort.c                    |  180 +++++++++++++++++++++++++++++++++++++++++
 net/sunrpc/svcauth_unix.c      |    4 
 17 files changed, 380 insertions(+), 103 deletions(-)

through these ChangeSets (diffs in separate email):

<thockin@freakshow.cobalt.com> (02/10/21 1.812)
   no one references nfsiod.h anymore - nix it.

<thockin@freakshow.cobalt.com> (02/10/21 1.811)
   fix usage of NGROUPS in nfsd and svcauth

<thockin@freakshow.cobalt.com> (02/10/21 1.810)
   Remove the limit of 32 groups.  We now have a per-task, dynamic array of
   groups, which is kept sorted and refcounted.
   
   This ChangeSet incorporates all the core functionality. but does not fixup
   all the incorrect usages of groups.  That is in a seperate ChangeSet.

<thockin@freakshow.cobalt.com> (02/10/21 1.809)
   Add generic qsort() and bsearch(): qsort() from BSD, bsearch() from glibc