2007-05-06 08:51:46

by Ollie Wild

[permalink] [raw]
Subject: [patch] removes MAX_ARG_PAGES

A while back, I sent out a preliminary patch
(http://thread.gmane.org/gmane.linux.ports.hppa/752) to remove the
MAX_ARG_PAGES limit on command line sizes. Since then, Peter Zijlstra
and I have fixed a number of bugs and addressed the various
outstanding issues.

The attached patch incorporates the following changes:

- Fixes a BUG_ON() assertion failure discovered by Ingo Molnar.
- Adds CONFIG_STACK_GROWSUP (parisc) support.
- Adds auditing support.
- Reverts to the old behavior on architectures with no MMU.
- Fixes broken execution of 64-bit binaries from 32-bit binaries.
- Adds elf_fdpic support.
- Fixes cache coherency bugs.

We've tested the following architectures: i386, x86_64, um/i386,
parisc, and frv. These are representative of the various scenarios
which this patch addresses, but other architecture teams should try it
out to make sure there aren't any unexpected gotchas.

Ollie


Attachments:
(No filename) (903.00 B)
no_MAX_ARG_PAGES.patch (45.96 kB)
Download all attachments

2007-05-07 17:47:04

by Tony Luck

[permalink] [raw]
Subject: RE: [patch] removes MAX_ARG_PAGES

> We've tested the following architectures: i386, x86_64, um/i386,
> parisc, and frv. These are representative of the various scenarios
> which this patch addresses, but other architecture teams should try it
> out to make sure there aren't any unexpected gotchas.

Doesn't build on ia64: complaints from arch/ia64/ia32/binfmt_elf.c
(which #includes ../../../fs/binfmt_elf.c) ...

arch/ia64/ia32/binfmt_elf32.c: In function `ia32_setup_arg_pages':
arch/ia64/ia32/binfmt_elf32.c:206: error: `MAX_ARG_PAGES' undeclared (first use in this function)
arch/ia64/ia32/binfmt_elf32.c:206: error: (Each undeclared identifier is reported only once
arch/ia64/ia32/binfmt_elf32.c:206: error: for each function it appears in.)
arch/ia64/ia32/binfmt_elf32.c:240: error: structure has no member named `page'
arch/ia64/ia32/binfmt_elf32.c:242: error: structure has no member named `page'
arch/ia64/ia32/binfmt_elf32.c:243: warning: implicit declaration of function `install_arg_page'
make[1]: *** [arch/ia64/ia32/binfmt_elf32.o] Error 1

Turning off CONFIG_IA32-SUPPORT, the kernel built, but oops'd during boot.
My serial connection to my test machine is currently broken, so I didn't
get a capture of the stack trace, sorry.

-Tony

2007-05-07 19:00:39

by William Lee Irwin III

[permalink] [raw]
Subject: Re: [patch] removes MAX_ARG_PAGES

At some point in the past, Ollie Wild wrote:
>> We've tested the following architectures: i386, x86_64, um/i386,
>> parisc, and frv. These are representative of the various scenarios
>> which this patch addresses, but other architecture teams should try it
>> out to make sure there aren't any unexpected gotchas.

On Mon, May 07, 2007 at 10:46:49AM -0700, Luck, Tony wrote:
> Doesn't build on ia64: complaints from arch/ia64/ia32/binfmt_elf.c
> (which #includes ../../../fs/binfmt_elf.c) ...
[...]
> Turning off CONFIG_IA32-SUPPORT, the kernel built, but oops'd during boot.
> My serial connection to my test machine is currently broken, so I didn't
> get a capture of the stack trace, sorry.

It needs to sweep 32-bit emulation code more generally.


-- wli

2007-05-09 20:49:08

by Andrew Morton

[permalink] [raw]
Subject: Re: [patch] removes MAX_ARG_PAGES

On Sun, 6 May 2007 01:51:34 -0700
"Ollie Wild" <[email protected]> wrote:

> A while back, I sent out a preliminary patch
> (http://thread.gmane.org/gmane.linux.ports.hppa/752) to remove the
> MAX_ARG_PAGES limit on command line sizes. Since then, Peter Zijlstra
> and I have fixed a number of bugs and addressed the various
> outstanding issues.
>
> The attached patch incorporates the following changes:
>
> - Fixes a BUG_ON() assertion failure discovered by Ingo Molnar.
> - Adds CONFIG_STACK_GROWSUP (parisc) support.
> - Adds auditing support.
> - Reverts to the old behavior on architectures with no MMU.
> - Fixes broken execution of 64-bit binaries from 32-bit binaries.
> - Adds elf_fdpic support.
> - Fixes cache coherency bugs.
>
> We've tested the following architectures: i386, x86_64, um/i386,
> parisc, and frv. These are representative of the various scenarios
> which this patch addresses, but other architecture teams should try it
> out to make sure there aren't any unexpected gotchas.

I'll duck this for now, given the couple of problems which people have reported.

But please keep going ;) We sorely need this.

2007-05-10 01:05:34

by Rob Landley

[permalink] [raw]
Subject: Re: [patch] removes MAX_ARG_PAGES

On Wednesday 09 May 2007 4:48 pm, Andrew Morton wrote:
> On Sun, 6 May 2007 01:51:34 -0700
> "Ollie Wild" <[email protected]> wrote:
>
> > A while back, I sent out a preliminary patch
> > (http://thread.gmane.org/gmane.linux.ports.hppa/752) to remove the
> > MAX_ARG_PAGES limit on command line sizes. Since then, Peter Zijlstra
> > and I have fixed a number of bugs and addressed the various
> > outstanding issues.
> >
> > The attached patch incorporates the following changes:
> >
> > - Fixes a BUG_ON() assertion failure discovered by Ingo Molnar.
> > - Adds CONFIG_STACK_GROWSUP (parisc) support.
> > - Adds auditing support.
> > - Reverts to the old behavior on architectures with no MMU.
> > - Fixes broken execution of 64-bit binaries from 32-bit binaries.
> > - Adds elf_fdpic support.
> > - Fixes cache coherency bugs.
> >
> > We've tested the following architectures: i386, x86_64, um/i386,
> > parisc, and frv. These are representative of the various scenarios
> > which this patch addresses, but other architecture teams should try it
> > out to make sure there aren't any unexpected gotchas.
>
> I'll duck this for now, given the couple of problems which people have
> reported.

Just FYI, a really really quick and dirty way of testing this sort of thing on
more architectures and you're likely to physically have?

1) Install QEMU.

2) Grab http://landley.net/code/firmware (releases in the downloads directory,
or tarball of most recent repository snapshot is
wget "http://landley.net/hg/firmware?ca=tip;type=gz").

3) Edit "download.sh" to point at the URL of your tarball instead of whatever
kernel.org version it's using. (Or add your patch to sources/patches if it
applies to the version it's already using. Note that if you set SHA1= blank
in download.sh it'll skip the checksum test, so you don't have to recalculate
the sha1sum if you don't want to.)

4) Run ./build.sh for the architecture you're interested in. (I suggest
armv4l, i686, mipsel, and x86_64. Both sparc and ppc are currently broken
for different reasons; I'm working on it.) Wait a longish time for it to
finish compiling. :)

5) "cd build; ./run-armv4l.sh" and your shell prompt should now be in qemu
running a kernel for the appropriate architecture. (You even have a native
version of gcc you can build stuff with, although you may have
to "ln -s /tools/lib /lib" to run the results, for reasons Linux From Scratch
developers will recognize. :)

This won't help you test real hardware (at least hardware qemu doesn't
emulate), but for stuff like filesystems or executable file formats, it's
handy. :)

Email me if something doesn't work...

Rob

2007-05-10 04:07:16

by Ollie Wild

[permalink] [raw]
Subject: Re: [patch] removes MAX_ARG_PAGES

On 5/9/07, Rob Landley <[email protected]> wrote:
> Just FYI, a really really quick and dirty way of testing this sort of thing on
> more architectures and you're likely to physically have?

Does this properly emulate caching? On parisc, cache coherency was
the main issue we ran into. I suspect this might be the case with
other architectures as well.

Ollie

2007-05-10 09:21:11

by Rob Landley

[permalink] [raw]
Subject: Re: [patch] removes MAX_ARG_PAGES

On Thursday 10 May 2007 12:06 am, Ollie Wild wrote:
> On 5/9/07, Rob Landley <[email protected]> wrote:
> > Just FYI, a really really quick and dirty way of testing this sort of
thing on
> > more architectures and you're likely to physically have?
>
> Does this properly emulate caching? On parisc, cache coherency was
> the main issue we ran into. I suspect this might be the case with
> other architectures as well.

This is really a QEMU question. I've been focused on making cross-compilers
and using those to create kernels and a minimal native build environment I
could use to natively compile packages with. (The way I designed the thing
you could substitute real hardware for the qemu step, assuming you had it.
Or another emulator like armulator for a specific platform.)

I don't believe QEMU emulates parisc yet, although it adds new platforms all
the time. (It just grew an alpha emulation last month.) It's under very
active development.

Rob

2007-05-22 12:24:20

by Peter Zijlstra

[permalink] [raw]
Subject: RE: [patch] removes MAX_ARG_PAGES

On Mon, 2007-05-07 at 10:46 -0700, Luck, Tony wrote:
> > We've tested the following architectures: i386, x86_64, um/i386,
> > parisc, and frv. These are representative of the various scenarios
> > which this patch addresses, but other architecture teams should try it
> > out to make sure there aren't any unexpected gotchas.
>
> Doesn't build on ia64: complaints from arch/ia64/ia32/binfmt_elf.c
> (which #includes ../../../fs/binfmt_elf.c) ...
>
> arch/ia64/ia32/binfmt_elf32.c: In function `ia32_setup_arg_pages':
> arch/ia64/ia32/binfmt_elf32.c:206: error: `MAX_ARG_PAGES' undeclared (first use in this function)
> arch/ia64/ia32/binfmt_elf32.c:206: error: (Each undeclared identifier is reported only once
> arch/ia64/ia32/binfmt_elf32.c:206: error: for each function it appears in.)
> arch/ia64/ia32/binfmt_elf32.c:240: error: structure has no member named `page'
> arch/ia64/ia32/binfmt_elf32.c:242: error: structure has no member named `page'
> arch/ia64/ia32/binfmt_elf32.c:243: warning: implicit declaration of function `install_arg_page'
> make[1]: *** [arch/ia64/ia32/binfmt_elf32.o] Error 1
>
> Turning off CONFIG_IA32-SUPPORT, the kernel built, but oops'd during boot.
> My serial connection to my test machine is currently broken, so I didn't
> get a capture of the stack trace, sorry.


Ok, I found the problem. IA64 places constraints on virtual address
space. We initially place the stack at TASK_SIZE, and once the binfmt
tells us where it should have gone, we move it down to the new location.

However IA64 has v-space carved up in regions, and the top of the user
accessible address space is reserved for hugetlbfs.

So we should be using STACK_TOP, which provides the highest stack
address, however, some arches have conditions in the STACK_TOP macros
such that the result is not what is expected until the binfmt
personality is set.

In order to solve this, I added a STACK_TOP_MAX macro for each arch and
use that. This made IA64 boot properly.

The second patch makes the compat stuff compile, untested though, as I
have no idea how that works on ia64.



Attachments:
stack_top_max.patch (11.56 kB)
setup_arg_pages.patch (4.65 kB)
Download all attachments

2007-05-24 15:20:46

by Peter Zijlstra

[permalink] [raw]
Subject: RE: [patch] removes MAX_ARG_PAGES

On Tue, 2007-05-22 at 16:49 -0700, Luck, Tony wrote:
> > In order to solve this, I added a STACK_TOP_MAX macro for each arch and
> > use that. This made IA64 boot properly.
>
> Yes. Patches apply to 2.6.21 cleanly, build cleanly and boot for me too.
> I can even run:
>
> $ ls -lrt `find /usr/share -type f -print`
>
> on the new kernel (thats 48020 files for a total of 2.3MB of args).
>
> > The second patch makes the compat stuff compile, untested though, as I
> > have no idea how that works on ia64.
>
> Not so good on this part. Running an ia32 binary on this patched kernel
> just hung my system :-(

I just tried this on an Altix from the test lab, and ia32 bash just
started.

native 64 bit gave this:

$ ls `find / -xdev -type f` | wc
1 89847 4200288

but the 32 bit bash kept complaining about his arg list being too long.
This might be a 32->64 issue (/bin/ls is 64 bit after all - and I
couldn't find a 32 bit ls in the compat packages), but I really don't
know. Will look into this.

The difference I had with what I posted last is below; these changes
were just me cleaning up. (and fixing one bug on STACK_GROWS_UP
introduced by the previous patches - but AFAIK that is not relevant to
ia64)

Signed-off-by: Peter Zijlstra <[email protected]>
---
fs/exec.c | 15 ++++++---------
1 file changed, 6 insertions(+), 9 deletions(-)

Index: linux-2.6-2/fs/exec.c
===================================================================
--- linux-2.6-2.orig/fs/exec.c 2007-05-24 09:15:54.000000000 +0200
+++ linux-2.6-2/fs/exec.c 2007-05-24 09:41:49.000000000 +0200
@@ -532,26 +532,23 @@ int setup_arg_pages(struct linux_binprm
stack_base = current->signal->rlim[RLIMIT_STACK].rlim_max;
if (stack_base > (1 << 30))
stack_base = 1 << 30;
- stack_base = PAGE_ALIGN(stack_top - stack_base);

/* Make sure we didn't let the argument array grow too large. */
if (vma->vm_end - vma->vm_start > stack_base)
return -ENOMEM;

+ stack_base = PAGE_ALIGN(stack_top - stack_base);
+
stack_shift = stack_base - vma->vm_start;
mm->arg_start = bprm->p + stack_shift;
bprm->p = vma->vm_end + stack_shift;
#else
BUG_ON(stack_top & ~PAGE_MASK);

- stack_base = arch_align_stack(stack_top - mm->stack_vm*PAGE_SIZE);
- stack_base = PAGE_ALIGN(stack_base);
-
- /* Make sure we didn't let the argument array grow too large. */
- if (stack_base > stack_top)
- return -ENOMEM;
+ stack_top = arch_align_stack(stack_top);
+ stack_top = PAGE_ALIGN(stack_top);
+ stack_shift = stack_top - vma->vm_end;

- stack_shift = stack_base - (bprm->p & PAGE_MASK);
bprm->p += stack_shift;
mm->arg_start = bprm->p;
#endif
@@ -598,7 +595,7 @@ int setup_arg_pages(struct linux_binprm
return -EFAULT;
}
#else
- if (expand_stack(vma, stack_base -
+ if (expand_stack(vma, vma->vm_start -
EXTRA_STACK_VM_PAGES * PAGE_SIZE)) {
up_write(&mm->mmap_sem);
return -EFAULT;


2007-05-25 18:48:36

by Tony Luck

[permalink] [raw]
Subject: RE: [patch] removes MAX_ARG_PAGES

> I just tried this on an Altix from the test lab, and ia32 bash just
> started.

I don't have any native x86 binaries on my Madison-based testbox, so my
test case was to compile a simple program that counted total length of
argument strings on an x86 box, and copy it to my ia64 box. So that I
wouldn't have to copy over a bunch of libraries too, I compiled it
with -static. This is the test case that "hung" my system (re-running
it today from /dev/tty1 instead of from an xterm, I see that it actually
oopsed in rb_next()). I wasn't even running with a long arglist. Just
"*" for my home directory (19 files/directories = ~170 bytes).

-Tony

My test program. Compile on ia32 box with "cc -static -o args args.c"

---- begin args.c ----
main(int argc, char **argv)
{
int n;

printf("argc = %d\n", argc);

n = 0;
while (--argc)
n += strlen(*++argv);

printf("bytes = %d\n", n);
}