Re: [PATCH] efivarfs: Iterate variables with increasing name buffer sizes

[Ard Biesheuvel]

On Tue, 23 Jan 2024 at 14:55, Tim Schumacher <timschumi@xxxxxx> wrote:
>
> On 23.01.24 12:24, Ard Biesheuvel wrote:
> > On Tue, 23 Jan 2024 at 00:15, Tim Schumacher <timschumi@xxxxxx> wrote:
> >>
> >> This sidesteps a quirk in a few old (2011-ish) UEFI implementations,
> >> where a call to `GetNextVariableName` with a buffer size larger than 512
> >> bytes will always return `EFI_INVALID_PARAMETER`.
> >
> > I wonder if we might just reduce this to 512 and be done with it.
> > Presumably, Windows boots fine in UEFI mode on these machines, which
> > suggests that it passes a value <= 512 too, and I don't recall ever
> > encountering systems using extremely long variable names (i.e., longer
> > than 512 byte)
>
> I'd rather avoid introducing deviations from the specifications on the
> kernel side as well.

Which specification would this deviate from?

> Someone or something might legitimately set a large
> variable name, so we'd have to have resize logic anyways (to resize from
> 512 to 512+). Also, as mentioned on the patch, I'm entirely unsure what
> the size ends up being used for, so I'd rather err on the side of
> caution (most importantly in regards to the buffer size).
>
> Windows _does_ boot fine (and is able to read all the variables), so
> they presumably start off with 512 or smaller. FreeBSD definitely starts
> from 512, but they also implement additional resize logic.
>
> In regards to complexity of the proposed solution, how about we approach
> this from the other side? Start off with advertising 1024 bytes of
> buffer storage, and cut that value in half (without actually resizing
> the buffer) as long as we get `EFI_INVALID_PARAMETER` while on the first
> run.
>
> If we ever get `EFI_BUFFER_TOO_SMALL`, we know that something is wrong
> with the UEFI implementation (because that either means that something
> claims to be larger than 1024 bytes, or that our assumptions about the
> quirk don't hold up) and can bail out and log as appropriate. That would
> limit the complexity to the machines that need it, completely omit the
> need for resize logic, and would still be specification compliant.

Yes, I would prefer to keep this as simple as possible.