kernel_cmdline quoting equivalence

[jeckersb]

Side quest spawned from investigating ostreedev/ostree#3544 ...

Quoted kargs end up getting transformed between grub and the kernel, here's an example:

# Original kargs does not have a "foo" value
[root@ibm-p8-kvm-03-guest-02 ~]# rpm-ostree kargs
root=UUID=4f68bce3-e8cd-4db1-96e7-fbcaf984b709 rw ostree=/ostree/boot.1/default/944761db707ac4bce2b4e1362b79cfea486bcd764ac56131a30780b4927ee26b/0
# Append a quoted arg
[root@ibm-p8-kvm-03-guest-02 ~]# rpm-ostree kargs --append='foo="quoted value"'
Staging deployment... done
Changes queued for next boot. Run "systemctl reboot" to start a reboot
# Sanity check that the quoted arg was added, and in the format we expect
[root@ibm-p8-kvm-03-guest-02 ~]# rpm-ostree kargs
root=UUID=4f68bce3-e8cd-4db1-96e7-fbcaf984b709 rw ostree=/ostree/boot.1/default/944761db707ac4bce2b4e1362b79cfea486bcd764ac56131a30780b4927ee26b/0 foo="quoted value"
[root@ibm-p8-kvm-03-guest-02 ~]# systemctl reboot
[...]
# After reboot, boot loader entry looks correct
[root@ibm-p8-kvm-03-guest-02 ~]# grep options /boot/loader/entries/ostree-2.conf
options root=UUID=4f68bce3-e8cd-4db1-96e7-fbcaf984b709 rw ostree=/ostree/boot.0/default/944761db707ac4bce2b4e1362b79cfea486bcd764ac56131a30780b4927ee26b/0 foo="quoted value"
# However, the kernel sees the value as "externally" quoted in early boot
[root@ibm-p8-kvm-03-guest-02 ~]# journalctl -b --grep 'Command line:' --no-pager
Nov 06 18:55:24 localhost kernel: Command line: BOOT_IMAGE=(hd0,gpt3)/boot/ostree/default-944761db707ac4bce2b4e1362b79cfea486bcd764ac56131a30780b4927ee26b/vmlinuz-6.12.0-150.el10.x86_64 root=UUID=4f68bce3-e8cd-4db1-96e7-fbcaf984b709 rw ostree=/ostree/boot.0/default/944761db707ac4bce2b4e1362b79cfea486bcd764ac56131a30780b4927ee26b/0 "foo=quoted value"
# And the same format persists via /proc/cmdline
[root@ibm-p8-kvm-03-guest-02 ~]# cat /proc/cmdline
BOOT_IMAGE=(hd0,gpt3)/boot/ostree/default-944761db707ac4bce2b4e1362b79cfea486bcd764ac56131a30780b4927ee26b/vmlinuz-6.12.0-150.el10.x86_64 root=UUID=4f68bce3-e8cd-4db1-96e7-fbcaf984b709 rw ostree=/ostree/boot.0/default/944761db707ac4bce2b4e1362b79cfea486bcd764ac56131a30780b4927ee26b/0 "foo=quoted value"

This is a bit of a landmine. I noticed it while trying to write an integration tests for the above linked issue and it failed because it didn't match how I expected it to. At the very least I need to update the kernel_cmdline crate so these are equivalent:

let outside_quoted = param("\"foo=quoted value\"");
let value_quoted = param("foo=\"quoted value\"");
assert_eq!(outside_quoted, value_quoted);

As currently implemented, this does not parse correctly and these are considered two different values.

[cgwalters]

And the Linux kernel parses them as equivalent? If so I guess it argues we do need to treat them equivalently too.

As far as where they get transformed...I guess we might need to dig into whether that's in the grub or the kernel parser

[jeckersb]

And the Linux kernel parses them as equivalent? If so I guess it argues we do need to treat them equivalently too.

Yeah it's equivalent.

As far as where they get transformed...I guess we might need to dig into whether that's in the grub or the kernel parser

I'm 99% sure it's grub that's transforming them, because as far as I can tell (with claude at my side) the value that gets logged as Command line: is directly memcpy'd from the pointer that the bootloader hands off to the kernel (at least for x86 since that's all arch-dependent). The kernel hasn't done any processing on the kargs yet.

Edit: would be an interesting exercise to see how other bootloaders behave with the same input. A quest for another day...

[jeckersb]

I spent a bit more time looking into this with claude assistance, and in a nutshell:

  The flow is:
  1. User types: linux /vmlinuz foo="quoted value"
  2. GRUB parser tokenizes it and strips quotes: argv[1] = foo=quoted value
  3. grub_create_loader_cmdline() sees a space, wraps whole thing: "foo=quoted value"
  4. Result passed to kernel: "foo=quoted value"

Or with links to relevant parts:

1. The lexer defines DQSTR as the type for double-quoted strings
2. grub_script_arglist_to_argv converts the lexed input into argv-style input
   2.1 DQSTR is expanded to arg->str which is the (unquoted) body of the double-quoted string.
3. grub_create_loader_cmdline adds quotes to the start and end if there's any spaces in the arg.