Re: [PATCH liveupdates] apply-live-updates.sh: handle updates for /run

[Will Woods <wwoods-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>]

On Thu, 2012-09-06 at 12:17 +0200, Harald Hoyer wrote:
> Am 30.08.2012 00:58, schrieb Will Woods:
> > /run will get mounted at $NEWROOT/run after switch_root, but it's not
> > there yet. bind-mount it in place so updates for /run actually land in
> > /run.
> > 
> 
> Huh? Why do you want updates in /run ?????

Short answer: because /tmp gets covered up by an empty tmpfs at (real)
system startup, and we need to put them somewhere!


Longer explanation: 

Historically, anaconda's runtime environment was read-only[1]. This
makes sense, since it's usually squashfs (which doesn't do writes), and
it's frequently running off a shiny, read-only plastic disc.

But how do you fix bugs (or test bugfixes) without regenerating the
image and burning a new disk? With updates images!

To make updates work, anaconda sets its PYTHONPATH set to something
like:
  /tmp/updates:$PYTHONPATH
and sets GLADEPATH, PIXMAPPATH, LD_LIBRARY_PATH, etc. similarly.

This way, updates images fetched with 'updates=http://...' get unpacked
into /tmp/updates, and those files override the existing contents of the
runtime image. Hooray! Bugfixes for read-only media!

These days we *do* have the magic device-mapper overlay hack, so we can
make a squashfs image read-write[2] and just overwrite files. But we
still use the old update method for a couple reasons:

1) It already works,
2) it works with read-only filesystems (like squashfs), 
3) overwriting existing files can be tricky[3].

====== A SIDE DISCUSSION ABOUT UPDATES AND READONLY ROOT/USR ======
I admit this is all kind of a hack.

It'd be much better if there was a generally supported systemwide place
to put 'updates' - that is, things that override the existing system
libraries/binaries/data/config in /, /etc, and /usr.

Maybe a tmpfs mount at /usr/local would work, but that doesn't cover
systemd configuration or ld.so. LD_LIBRARY_PATH could fix the latter,
but it'd be better if we didn't have to change half a dozen $XXX_PATH
variables to make this work.

I seem to remember there was some work on readonly-root systems with the
Stateless Linux stuff, but I'm not sure how/if that solves the problem
here..
===================================================================

Anyway, to make the existing method keep working, we need to unpack the
updates image to a tmpfs that stays around after switch-root. And that's
what /run is for. So that's why we want to put the updates in /run.

Does that answer the question?

-w

[1] Mostly because squashfs is read-only, but even if we weren't using
squashfs the image might be on a CDROM or read-only NFS mount or
something.

[2] This actually requires that we put the runtime image into an ext4
image, and put *that* into squashfs - squashfs doesn't support writes at
all, at the filesystem layer, and the device-mapper stuff redirects
writes at the *block* layer..

[3] for example, some updates images contained /lib/libXXX.so. But then
we overwrote the /lib -> /usr/lib symlink with the new '/lib' directory,
thus breaking all the other libraries..