Bug 1796120

+++ This bug was initially created as a clone of Bug #1771976 +++

--- Additional comment from Pino Toscano on 2020-01-28 15:08:41 UTC ---

(In reply to Lubomír Sedlář from comment #11)
> We have seen the same problem again. The directory it complained about was 
> /var/tmp/.guestfs-101448/appliance.d/root and right after the failure it did
> not exist.

Under a properly cached appliance, you will see three files, something like (with different sizes of course):
$ du -hcs /var/tmp/.guestfs-101448/appliance.d/
1.6M    tmp/.guestfs-1000/appliance.d/initrd
9.9M    tmp/.guestfs-1000/appliance.d/kernel
383M    tmp/.guestfs-1000/appliance.d/root
394M    total

Since 'root' is a big file (at least 300/350M) and that the directory is under /var/tmp, my theory is that systemd-tmpfiles is cleaning that to save space.
In particular, /usr/lib/tmpfiles.d/tmp.conf sets 30d (at least on Fedora 31) as age for /var/tmp, so in case libguestfs was not run for more than 30 days then I think that files might be cleaned.

When this happens, can you please check the timestamp of the appliance.d and all the files inside it?
- `stat /var/tmp/.guestfs-101448/appliance.d/`
- `stat /var/tmp/.guestfs-101448/appliance.d/*`

---

The issue I see is that supermin (that creates the appliance in appliance.d) does not check for the existance of all the files it creates, but only whether the directory exists. Most probably it should be smarter and check which files exists -- sadly ATM it is not easy to do so, as the files created change depending on the appliance type (tgz vs chroot) and depending on the files themselves.

--- Additional comment from Richard W.M. Jones on 2020-01-28 15:34:48 UTC ---

> supermin: if-newer: output does not need rebuilding

This message is consistent with what Pino says.

https://github.com/libguestfs/supermin/blob/62d5c774d6c8fcac11e28fcba99754b5478e5088/src/supermin.ml#L232

Note on line 239 it only checks the date of the output directory.

--- Additional comment from Lubomír Sedlář on 2020-01-29 13:35:07 UTC ---

In this case the appliance.d/root file is 4.0G.

Looking around the machine, I see there is /var/tmp/.guestfs-$UID/ for multiple users where the only content is an empty appliance.d directory.

I can replicate the error now:

/var/tmp/.guestfs-$UID is missing => guestmount works
/var/tmp/.guestfs-$UID contains all files => guestmount works
/var/tmp/.guestfs-$UID/appliance.d is empty directory and nothing else exists in the directory => error

/usr/lib/tmpfiles.d/tmp.conf defines that files should be deleted after 30 days. 
    v /var/tmp 1777 root root 30d

Is there something I can change about the guestmount command to avoid this? The only workaround I can probably do is to unconditionally remove /var/tmp/.guestfs-$UID before running it.

--- Additional comment from Richard W.M. Jones on 2020-01-29 16:17:46 UTC ---

I think this is an actual bug in supermin, as described in Pino above.
I wonder why we've not seen it before though - maybe the defaults in
tmpfiles changed recently?