bug#54557: “fakechroot” execution engine doesn’t work for Bash

bug#54557: “fakechroot” execution engine doesn’t work for Bash

[Ludovic Courtès]

Hello!

The “fakechroot” engine fails when applied to ‘bash’:

--8<---------------cut here---------------start------------->8---
$ guix pack -RR bash -S /bin=bin 
/gnu/store/mnbmklbvd1pk7yby0k8h26msk6z11m1m-bash-tarball-pack.tar.gz
$ (cd /tmp/pack; tar xf /gnu/store/mnbmklbvd1pk7yby0k8h26msk6z11m1m-bash-tarball-pack.tar.gz)
$ unshare -mrf sh -c 'mount -t tmpfs -o ro none /gnu/store; GUIX_EXECUTION_ENGINE=fakechroot /tmp/pack/bin/bash --version'
/tmp/pack/gnu/store/d99ykvj3axzzidygsmdmzxah4lvxd6hw-bash-5.1.8/bin//bash: out of memory to store relocation results for /tmp/pack/bin/bash
--8<---------------cut here---------------end--------------->8---

The message comes from ld.so.  My guess is that the problem comes from
Bash’s terrible ‘malloc’ replacement:

--8<---------------cut here---------------start------------->8---
$ objdump -T /gnu/store/d99ykvj3axzzidygsmdmzxah4lvxd6hw-bash-5.1.8/bin/bash | grep malloc
00000000004ae6f0 g    DF .text	000000000000003b  Base        malloc_usable_size
00000000004ae850 g    DF .text	0000000000000009  Base        malloc
0000000000484e70 g    DF .text	000000000000005b  Base        xmalloc
00000000004ee020 g    DO .bss	0000000000000004  Base        malloc_trace_at_exit
00000000004e8c24 g    DO .data	0000000000000004  Base        malloc_mmap_threshold
0000000000484f70 g    DF .text	00000000000000dd  Base        sh_xmalloc
00000000004f7a04 g    DO .bss	0000000000000004  Base        malloc_trace
00000000004f7a08 g    DO .bss	0000000000000004  Base        malloc_flags
00000000004ae730 g    DF .text	0000000000000005  Base        sh_malloc
00000000004f7a00 g    DO .bss	0000000000000004  Base        malloc_register
00000000004ae660 g    DF .text	000000000000000c  Base        _malloc_unblock_signals
00000000004ae630 g    DF .text	000000000000002e  Base        _malloc_block_signals
--8<---------------cut here---------------end--------------->8---

[Time passes…]  I confirmed this hypothesis by running:

  guix pack -RR -S /bin=bin -m manifest.scm

on this manifest:

--8<---------------cut here---------------start------------->8---
(use-modules (guix) (guix utils)
             (guix profiles)
             (gnu packages bash))

(define bash-sans-malloc
  (package/inherit bash
    (name "bash-sans-malloc")
    (arguments
     (substitute-keyword-arguments (package-arguments bash)
       ((#:configure-flags flags ''())
        `(cons "--without-bash-malloc" ,flags))))))

(packages->manifest (list bash-sans-malloc))
--8<---------------cut here---------------end--------------->8---

Works just fine:

--8<---------------cut here---------------start------------->8---
$ unshare -mrf sh -c 'mount -t tmpfs -o ro none /gnu/store; GUIX_EXECUTION_ENGINE=fakechroot /tmp/pack/bin/bash --version'
GNU bash, version 5.1.8(1)-release (x86_64-unknown-linux-gnu)
Copyright (C) 2020 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later <http://gnu.org/licenses/gpl.html>

This is free software; you are free to change and redistribute it.
There is NO WARRANTY, to the extent permitted by law.
--8<---------------cut here---------------end--------------->8---

So in the end, it’s a bug report that’s kinda closed already.

Perhaps we should build Bash ‘--without-bash-malloc’ unconditionally in
‘core-updates’?  The ‘INSTALL’ file reads:

--8<---------------cut here---------------start------------->8---
'--with-bash-malloc'
     Use the Bash version of 'malloc' in the directory 'lib/malloc'.
     This is not the same 'malloc' that appears in GNU libc, but an
     older version originally derived from the 4.2 BSD 'malloc'.  This
     'malloc' is very fast, but wastes some space on each allocation.
     This option is enabled by default.  The 'NOTES' file contains a
     list of systems for which this should be turned off, and
     'configure' disables this option automatically for a number of
     systems.
--8<---------------cut here---------------end--------------->8---

There might be other options if we want to keep it, such as changing the
ELF visibility of those symbols, but I wonder if it’s worth the effort.

Thoughts?

Ludo’.

bug#54557: “fakechroot” execution engine doesn’t work for Bash

[Philippe SWARTVAGHER]

Le 24/03/2022 à 22:02, Ludovic Courtès a écrit :
> Hello!
>
> The “fakechroot” engine fails when applied to ‘bash’:
>
> --8<---------------cut here---------------start------------->8---
> $ guix pack -RR bash -S /bin=bin
> /gnu/store/mnbmklbvd1pk7yby0k8h26msk6z11m1m-bash-tarball-pack.tar.gz
> $ (cd /tmp/pack; tar xf /gnu/store/mnbmklbvd1pk7yby0k8h26msk6z11m1m-bash-tarball-pack.tar.gz)
> $ unshare -mrf sh -c 'mount -t tmpfs -o ro none /gnu/store; GUIX_EXECUTION_ENGINE=fakechroot /tmp/pack/bin/bash --version'
> /tmp/pack/gnu/store/d99ykvj3axzzidygsmdmzxah4lvxd6hw-bash-5.1.8/bin//bash: out of memory to store relocation results for /tmp/pack/bin/bash
> --8<---------------cut here---------------end--------------->8---
>
> The message comes from ld.so.  My guess is that the problem comes from
> Bash’s terrible ‘malloc’ replacement:
>
> --8<---------------cut here---------------start------------->8---
> $ objdump -T /gnu/store/d99ykvj3axzzidygsmdmzxah4lvxd6hw-bash-5.1.8/bin/bash | grep malloc
> 00000000004ae6f0 g    DF .text	000000000000003b  Base        malloc_usable_size
> 00000000004ae850 g    DF .text	0000000000000009  Base        malloc
> 0000000000484e70 g    DF .text	000000000000005b  Base        xmalloc
> 00000000004ee020 g    DO .bss	0000000000000004  Base        malloc_trace_at_exit
> 00000000004e8c24 g    DO .data	0000000000000004  Base        malloc_mmap_threshold
> 0000000000484f70 g    DF .text	00000000000000dd  Base        sh_xmalloc
> 00000000004f7a04 g    DO .bss	0000000000000004  Base        malloc_trace
> 00000000004f7a08 g    DO .bss	0000000000000004  Base        malloc_flags
> 00000000004ae730 g    DF .text	0000000000000005  Base        sh_malloc
> 00000000004f7a00 g    DO .bss	0000000000000004  Base        malloc_register
> 00000000004ae660 g    DF .text	000000000000000c  Base        _malloc_unblock_signals
> 00000000004ae630 g    DF .text	000000000000002e  Base        _malloc_block_signals
> --8<---------------cut here---------------end--------------->8---
>
> [Time passes…]  I confirmed this hypothesis by running:
>
>    guix pack -RR -S /bin=bin -m manifest.scm
>
> on this manifest:
>
> --8<---------------cut here---------------start------------->8---
> (use-modules (guix) (guix utils)
>               (guix profiles)
>               (gnu packages bash))
>
> (define bash-sans-malloc
>    (package/inherit bash
>      (name "bash-sans-malloc")
>      (arguments
>       (substitute-keyword-arguments (package-arguments bash)
>         ((#:configure-flags flags ''())
>          `(cons "--without-bash-malloc" ,flags))))))
>
> (packages->manifest (list bash-sans-malloc))
> --8<---------------cut here---------------end--------------->8---
>
> Works just fine:
>
> --8<---------------cut here---------------start------------->8---
> $ unshare -mrf sh -c 'mount -t tmpfs -o ro none /gnu/store; GUIX_EXECUTION_ENGINE=fakechroot /tmp/pack/bin/bash --version'
> GNU bash, version 5.1.8(1)-release (x86_64-unknown-linux-gnu)
> Copyright (C) 2020 Free Software Foundation, Inc.
> License GPLv3+: GNU GPL version 3 or later <http://gnu.org/licenses/gpl.html>
>
> This is free software; you are free to change and redistribute it.
> There is NO WARRANTY, to the extent permitted by law.
> --8<---------------cut here---------------end--------------->8---
>
> So in the end, it’s a bug report that’s kinda closed already.
>
> Perhaps we should build Bash ‘--without-bash-malloc’ unconditionally in
> ‘core-updates’?  The ‘INSTALL’ file reads:
>
> --8<---------------cut here---------------start------------->8---
> '--with-bash-malloc'
>       Use the Bash version of 'malloc' in the directory 'lib/malloc'.
>       This is not the same 'malloc' that appears in GNU libc, but an
>       older version originally derived from the 4.2 BSD 'malloc'.  This
>       'malloc' is very fast, but wastes some space on each allocation.
>       This option is enabled by default.  The 'NOTES' file contains a
>       list of systems for which this should be turned off, and
>       'configure' disables this option automatically for a number of
>       systems.
> --8<---------------cut here---------------end--------------->8---
>
> There might be other options if we want to keep it, such as changing the
> ELF visibility of those symbols, but I wonder if it’s worth the effort.
>
> Thoughts?
>
> Ludo’.

Hello,

FTR, the --without-bash-malloc is used in the Debian bash package:

apt source bash

cd bash-5.1/debian

grep -Irn without-bash-malloc
changelog:145:  * Configure the normal build --without-bash-malloc as well.
changelog:1125:  * Configure the static build --without-bash-malloc.
changelog:1462:  * Disable the GNU/kFreeBSD kludge 
(--without-bash-malloc). Closes: #234137.
changelog:1546:  * Configure --without-bash-malloc on GNU/FreeBSD 
(closes: #194182).
changelog:1739:  * Configure --without-bash-malloc. At least on hppa, 
this fixes an error,
rules:79:    --without-bash-malloc


This option is also advised in Linux From Scratch: 
https://www.linuxfromscratch.org/lfs/view/stable/chapter08/bash.html

-- 
Philippe

bug#54557: “fakechroot” execution engine doesn’t work for Bash

[Maxim Cournoyer]

Hi Ludovic,

Ludovic Courtès <ludovic.courtes@inria.fr> writes:

[...]

> The message comes from ld.so.  My guess is that the problem comes from
> Bash’s terrible ‘malloc’ replacement:
>

[...]

> [Time passes…]  I confirmed this hypothesis by running:
>
>   guix pack -RR -S /bin=bin -m manifest.scm
>
> on this manifest:
>
> (use-modules (guix) (guix utils)
>              (guix profiles)
>              (gnu packages bash))
>
> (define bash-sans-malloc
>   (package/inherit bash
>     (name "bash-sans-malloc")
>     (arguments
>      (substitute-keyword-arguments (package-arguments bash)
>        ((#:configure-flags flags ''())
>         `(cons "--without-bash-malloc" ,flags))))))
>
> (packages->manifest (list bash-sans-malloc))
>
>
> Works just fine:
>
> $ unshare -mrf sh -c 'mount -t tmpfs -o ro none /gnu/store; GUIX_EXECUTION_ENGINE=fakechroot /tmp/pack/bin/bash --version'
> GNU bash, version 5.1.8(1)-release (x86_64-unknown-linux-gnu)
> Copyright (C) 2020 Free Software Foundation, Inc.
> License GPLv3+: GNU GPL version 3 or later <http://gnu.org/licenses/gpl.html>

[...]

> Perhaps we should build Bash ‘--without-bash-malloc’ unconditionally in
> ‘core-updates’?  The ‘INSTALL’ file reads:
>
> '--with-bash-malloc'
>      Use the Bash version of 'malloc' in the directory 'lib/malloc'.
>      This is not the same 'malloc' that appears in GNU libc, but an
>      older version originally derived from the 4.2 BSD 'malloc'.  This
>      'malloc' is very fast, but wastes some space on each allocation.
>      This option is enabled by default.  The 'NOTES' file contains a
>      list of systems for which this should be turned off, and
>      'configure' disables this option automatically for a number of
>      systems.
>
> There might be other options if we want to keep it, such as changing the
> ELF visibility of those symbols, but I wonder if it’s worth the effort.
>
> Thoughts?

I'd be OK with --without-bash-malloc; it seems we'll pay a bit in terms
of Bash performance in exchange for better memory usage.  It also brings
benefits such as solving this issue and may benefit from
advances/bugfixes to glibc's malloc in the future, if there are any.

Well done investigating!

Maxim

bug#54557: “fakechroot” execution engine doesn’t work for Bash

[Ludovic Courtès]

Hi,

Philippe SWARTVAGHER <philippe.swartvagher@inria.fr> skribis:

> FTR, the --without-bash-malloc is used in the Debian bash package:
>
> apt source bash
>
> cd bash-5.1/debian
>
> grep -Irn without-bash-malloc
> changelog:145:  * Configure the normal build --without-bash-malloc as well.
> changelog:1125:  * Configure the static build --without-bash-malloc.
> changelog:1462:  * Disable the GNU/kFreeBSD kludge
> (--without-bash-malloc). Closes: #234137.
> changelog:1546:  * Configure --without-bash-malloc on GNU/FreeBSD
> (closes: #194182).
> changelog:1739:  * Configure --without-bash-malloc. At least on hppa,
> this fixes an error,
> rules:79:    --without-bash-malloc
>
>
> This option is also advised in Linux From Scratch:
> https://www.linuxfromscratch.org/lfs/view/stable/chapter08/bash.html

Good to know, thanks for sharing.

I just realized that ‘bash-minimal’ in Guix already uses it, so that’s
another way to work around the ‘guix pack -RR’ issue at hand.

Ludo’.

bug#54557: “fakechroot” execution engine doesn’t work for Bash

[Ludovic Courtès]

Hi,

Maxim Cournoyer <maxim.cournoyer@gmail.com> skribis:

> I'd be OK with --without-bash-malloc; it seems we'll pay a bit in terms
> of Bash performance in exchange for better memory usage.  It also brings
> benefits such as solving this issue and may benefit from
> advances/bugfixes to glibc's malloc in the future, if there are any.

We might actually benefit from the improvements made in glibc’s malloc
over the years (it’s more actively developed than that of Bash), who
knows.

Anyway, pushed to ‘core-updates’ as
c6b5161e97ed1010d61331874b09c3231af3b1f9.

Thanks,
Ludo’.