guile magic fail ;-)

[dimaqq]

pacmanned autoconf, autogen, automake, libtool; pacman pulled in guile-1.8.8-2-arm and gave out this warning:

/sbin/ldconfig: /usr/lib/libguile.so is not an ELF file - it has the wrong magic bytes at the start.

looked at the file:
# hexdump -C /usr/lib/libguile.so.17.4.0 | head
00000000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
*
000a2a20 00 00 00 00 00 00 00 00 00 20 93 e5 18 00 83 e5 |......... ......|
000a2a30 01 20 82 e2 14 20 83 e5 00 b0 e0 e3 58 ff ff ea |. ... ......X...|
000a2a40 04 30 94 e5 00 20 a0 e3 18 20 83 e5 00 b0 e0 e3 |.0... ... ......|
000a2a50 53 ff ff ea 20 20 9f e5 01 30 d0 e5 02 20 95 e7 |S... ...0... ..|
000a2a60 34 10 a0 e3 00 20 92 e5 91 23 23 e0 1c 30 93 e5 |4.... ...##..0..|
000a2a70 33 ff 2f e1 b1 ff ff ea 1c 61 02 00 14 1a 00 00 |3./......a......|
000a2a80 41 0f 50 e3 70 40 2d e9 00 40 a0 e1 2b 00 00 0a |A.P.p@-..@..+...|
000a2a90 0c 00 50 e3 2c 00 00 0a ff 30 00 e2 0c 00 53 e3 |..P.,....0....S.|

Those zeros at the start look rather suspicious.

[dimaqq]

found out that /usr/share/file/magic.mgc also has a bunch of zeroes in the start of file.
file package was upgraded via pacman -Syu today too, to 5.11-1

Unpacked file-5.11-1-arm.pkg.tar.xz (md5: 7371420657d1567edd92c98b0b451833) manually on amd64 and verified that magic.mgc is broken inside the arm package. Is there any way to verify that file on the repo server?

makes me wonder if many today's upgrades are broken.

I compared amd64 to arm, file-5.11-1, magic.mgc, same file size, arm appears to have first 666100 bytes overwritten with zeros. Amount of zeros is same as in libguile.so

[WarheadsSE]

You might want to provide this thread to upstream as well, if you haven't already!

[dimaqq]

Not sure how to do that, perhaps here?
Packager : Kevin Mihelich <kevin@archlinuxarm.org>

I'm pretty sure only arm builds are affected. Amd64 package of the same version, file-5.11-1, is ok

[WarheadsSE]

I must have misunderstood "on amd64 and verified that magic.mgc is broken inside the package"

[dimaqq]

Indeed it was a bit confusing, expanded a bit.

[kmihelich]

From looking into this quickly this morning, file and guile are definitely broken on v5.. but it doesn't seem to have propagated into other packages. I pulled a random sampling of other lib* and program packages, and they're reporting correctly under a fixed magic.

I'll get rebuilt packages uploaded later.

[dimaqq]

Thanks!

I checked the file system, there are only very few files where first 1024 bytes are zeros, assumed broken, here's the list:

/usr/lib/libgmp.a is owned by gmp 5.0.4-1
/usr/lib/libguile.so.17.4.0 is owned by guile 1.8.8-2
/usr/share/hwdata/pci.ids is owned by pciutils 3.1.9-1
/usr/share/file/magic.mgc is owned by file 5.11-1

The oldest of these was built on Feb 16th.

Of course I don't have too many packages installed, can't say anything about other packages.

I volunteer to install and run same quick check on all if needed, doesn't cost me anything I run self-hosted arm in qemu chroot on my desktop.

Any idea how this came to be? Perhaps it's time to fsck the build machine?

[kmihelich]

Not sure how it happened, though it's very much reproducible. Worth noting, the issue doesn't happen when packages are rebuilt in a v5 chroot running on a Pandaboard, so there might be something deep that got hit by something.

[kmihelich]

Actually, finding bad stuff back to the middle of February makes sense. One of the main v5 builders, a GoFlex Net of mine, died with RAM failure around Feb 12th. Its hard drive was transplanted to another device to keep things building, but there may be something wrong with it somewhere. I'm doing some further testing against a known-good v5 builder to see if I can isolate the issue, hopefully this turns out to be relatively painless to fix.