Arch Pacman broken (fails reading list of installed packages)

[fpqc]

Build 16237

Installing packages with pacman in an Arch userspace (manual setup instructions here ) now fails to see that installed packages are installed, so installing a new package will install all dependencies up to the very top of the dependency tree. I think this also happened in the previous build (16236) (actually no longer sure), and I'm not sure if it happened due to a change in the code of Pacman or a change in WSL (I don't have a Win10 computer on hand running CU, just insider builds).

The workaround is to delete the local database (/var/lib/pacman/local) before every manual installation of a package, but this completely breaks updating with -Su.

[sunilmut]

@fpqc - Are you able to downgrade the pacman version to see if that changes anything? That will reduce one of the variable here.

[DDoSolitary]

@sunilmut I also have this problem too after upgrading build 16237. I didn't upgrade the pacman.

[fpqc]

@sunilmut Grabbed 5.0.1-5 from my remote package cache downgrading now.

Edit: Still fails in the same way. I'm using this bootstrap tarball

https://mirrors.kernel.org/archlinux/iso/latest/archlinux-bootstrap-2017.07.01-x86_64.tar.gz

Here's the one from May, which definitely worked at the time on earlier builds of Windows:

https://mirrors.kernel.org/archlinux/iso/2017.05.01/archlinux-bootstrap-2017.05.01-x86_64.tar.gz

It fails when you chroot in to see the existing pacman database, meaning that it must be a WSL regression and not a pacman or glibc regression.

[fpqc]

@sunilmut If you untar and chroot into the bootstrap environment, pacman -Q will show no output.

Previously (on earlier builds like 16215), pacman -Q while chrooted would output this list for 05.01 bootstrap tarball and this list for 07.01 bootstrap tarball.

The output on WSL 16237 is nonexistent.

Since the updated versions of strace are still not working due to #555 , I'm guessing you can get the traces directly out of the driver to see what syscalls are getting hit and perhaps what's breaking.

[sunilmut]

Looking into this now, starting with a repro.

[sunilmut]

Looks like I have a repro. Will start digging.

[sunilmut]

Found the culprit.

syscall trace of pacman -Q on build 16215, where things are working:

GETDENTS - 3, {000000000154DD40}, 32768 = 3107
GETDENTS - 3, {000000000154DD40}, 32768 = 0
LSEEK - 3, 0, 0 = 0
GETDENTS - 3, {000000000154DD40}, 32768 = 3107
GETDENTS - 3, {000000000154DD40}, 32768 = 0

syscall trace of pacman -Q on build 16237, where things are broken:

GETDENTS - 3, {0000000001E13D40}, 32768 = 3424
GETDENTS - 3, {0000000001E13D40}, 32768 = 0
LSEEK - 3, 0, 0 = 0
GETDENTS - 3, {0000000001E13D40}, 32768 = 0

Note how after the LSEEK call, the cursor is not rewinding. We did some changes to the getdents syscall recently and regressed this. I tested some privates and the issue is resolved. I will get this in and try to get it fast tracked. Thanks @fpqc for reporting the issue. Also adding @SvenGroot as FYI.

[sunilmut]

The fix has been checked in the dev branch and we will try to get it up as soon as possible.

Additionally, are there any other known blockers in getting strace to work on Arch, other than #555 ?

[therealkenc]

Additionally, are there any other known blockers in getting strace to work on Arch

Nah PTRACE_O_TRACEEXIT will make strace 4.15 happy AFAIK. The breakage was this change here per #1578.

[sunilmut]

@SvenGroot recently added support for PTRACE_O_TRACEEXIT, which is making it's way to the Insider build.

[sunilmut]

#555 is "fixinbound" :)

[mateuszlewko]

In which build we can expect this fix?

[benhillis]

Most recent Insider build has the fix.

[meltdown03]

Do we have to uninstall / reinstall to see the chages? Running pacman -Q still shows nothing for me on latest build.

[benhillis]

No you do not have to reinstall. I read over this thread again and it looks like there is one additional fix that will be coming soon.

[fpqc]

@meltdown03 If you used my workaround it is highly recommended to reinstall because the database will be broken and inaccurate.

[meltdown03]

I've been using the workaround everytime I install/update anything. There's no way to rebuild the database somehow?

[fpqc]

Nope. This was sort of implicit when I warned that -Su is going to break. You could fix it I guess by doing pacman -S base base-devel --force and then manually force-reinstalling each of the packages you had previously installed.

[fpqc]

@benhillis You should just be able to mark this fixed (not even fixedininsider because it was a fast-ring only regression). I tested in the latest build and all is well.

[benhillis]

This is fixed in 16257.