[edk2-devel] [PATCH 1/1] OvmfPkg/VirtioFsDxe: fix SimpleFileOpen
Laszlo Ersek
lersek at redhat.com
Wed Oct 18 11:20:22 UTC 2023
On 10/18/23 12:33, Gerd Hoffmann wrote:
> VirtiofsDxe throws an error in case the caller tries to open a file or
> directory using an handle with is not a directory, claiming that opening
> something relative to a file does not make sense.
>
> The claim is correct, but the code throws errors for both relative and
> absolute paths. Add a check to fix that.
>
> Signed-off-by: Gerd Hoffmann <kraxel at redhat.com>
> ---
> OvmfPkg/VirtioFsDxe/SimpleFsOpen.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/OvmfPkg/VirtioFsDxe/SimpleFsOpen.c b/OvmfPkg/VirtioFsDxe/SimpleFsOpen.c
> index a13d4f6a1e2d..1729ea2f5cf2 100644
> --- a/OvmfPkg/VirtioFsDxe/SimpleFsOpen.c
> +++ b/OvmfPkg/VirtioFsDxe/SimpleFsOpen.c
> @@ -397,7 +397,7 @@ VirtioFsSimpleFileOpen (
> // it cannot be implemented consistently with how a file is referred to
> // relative to a directory).
> //
> - if (!VirtioFsFile->IsDirectory) {
> + if (!VirtioFsFile->IsDirectory && FileName[0] != '\\') {
> DEBUG ((
> DEBUG_ERROR,
> ("%a: Label=\"%s\" CanonicalPathname=\"%a\" FileName=\"%s\": "
It's nice to see this topic pop up on edk2-devel; apparently you started
testing shim on top of virtio-fs. :)
I have had the following patch in my local repo, on a separate branch,
since April this year:
> commit cb4a6d1664ea6cabd14d2af0e5d9abb114973870
> Author: Laszlo Ersek <lersek at redhat.com>
> Date: Sat Apr 8 22:50:50 2023 +0200
>
> OvmfPkg/VirtioFsDxe: tolerate opening an abs. pathname rel. to a reg. file
>
> Referring to a file relative to a regular file makes no sense (or at least
> it cannot be implemented consistently with how a file is referred to
> relative to a directory). VirtioFsSimpleFileOpen() has enforced this
> strictly since the beginning, and a few months ago I reported USWG Mantis
> ticket #2367 [1] too, for clearing up the related confusion in the UEFI
> spec.
>
> Unfortunately, the shim boot loader contains such a bug [2] [3]. I don't
> believe the shim bug is ever going to be fixed. We can however relax the
> check in VirtioFsSimpleFileOpen() a bit: if the pathname that's being
> opened relative to a regular file is absolute, then the base file is going
> to be ignored anyway, so we can let the caller's bug slide. This happens
> to make shim work.
>
> Why this matters: UEFI-bootable Linux installer ISOs tend to come with
> shim and grub in the embedded (ElTorito) FAT image (ESP). Sometimes you
> want to build upstream shim/grub binaries, but boot the same ISO
> otherwise. The fastest way for overriding the ESP for this purpose is to
> copy its original contents to a virtio filesystem, then overwrite the shim
> and grub binaries from the host side. Note that this is different from
> direct-booting a kernel (via fw_cfg); the point is to check whether the
> just-built shim and grub are able to boot the rest of the ISO.
>
> [1] https://mantis.uefi.org/mantis/view.php?id=2367
> [2] https://bugzilla.redhat.com/show_bug.cgi?id=1966973
> [3] https://github.com/rhboot/shim/issues/382
>
> Cc: Ard Biesheuvel <ardb+tianocore at kernel.org>
> Cc: Gerd Hoffmann <kraxel at redhat.com>
> Cc: Jiewen Yao <jiewen.yao at intel.com>
> Cc: Jordan Justen <jordan.l.justen at intel.com>
> Signed-off-by: Laszlo Ersek <lersek at redhat.com>
>
> Notes:
> context:-U4
>
> diff --git a/OvmfPkg/VirtioFsDxe/SimpleFsOpen.c b/OvmfPkg/VirtioFsDxe/SimpleFsOpen.c
> index d479f76f5bc3..ec0521ac3703 100644
> --- a/OvmfPkg/VirtioFsDxe/SimpleFsOpen.c
> +++ b/OvmfPkg/VirtioFsDxe/SimpleFsOpen.c
> @@ -394,22 +394,33 @@ VirtioFsSimpleFileOpen (
>
> //
> // Referring to a file relative to a regular file makes no sense (or at least
> // it cannot be implemented consistently with how a file is referred to
> - // relative to a directory).
> + // relative to a directory). See USWG Mantis ticket #2367.
> //
> if (!VirtioFsFile->IsDirectory) {
> + BOOLEAN BugCompat;
> +
> + //
> + // Tolerate this bug in the caller if FileName is absolute. If FileName is
> + // absolute, then VirtioFsAppendPath() below will disregard
> + // VirtioFsFile->CanonicalPathname.
> + //
> + BugCompat = (FileName[0] == L'\\');
> +
> DEBUG ((
> - DEBUG_ERROR,
> + BugCompat ? DEBUG_WARN : DEBUG_ERROR,
> ("%a: Label=\"%s\" CanonicalPathname=\"%a\" FileName=\"%s\": "
> "nonsensical request to open a file or directory relative to a regular "
> "file\n"),
> __FUNCTION__,
> VirtioFs->Label,
> VirtioFsFile->CanonicalPathname,
> FileName
> ));
> - return EFI_INVALID_PARAMETER;
> + if (!BugCompat) {
> + return EFI_INVALID_PARAMETER;
> + }
> }
>
> //
> // Allocate the new VIRTIO_FS_FILE object.
Note that I'm adamant that this is a shim (and UEFI spec) bug, and that
the current upstream code is right, *regardless* of whether the pathname
to open starts with a backslash or not. The spec bug is reference [1]
above, and the original incarnation of my shim bug report is reference
[2]. Reference [3] is just the original RHBZ [2] having been migrated /
copied to the upstream tracker.
In other words, the patch is expressly a bug-compat patch.
There are two reasons why I never posted the patch:
(1) The (non-)treatment I received from the shim maintainers in ticket
<https://github.com/rhboot/shim/issues/382> discouraged me from doing
anything with, or for, shim.
(2) With this modification in place, shim is happy, but grub isn't. When
I realized that, I looked relatively deeply into making grub work on top
of virtio-fs as well -- and my findings were horrendous.
I wrote up my findings in a private email to some colleagues; you were
among the recipients. The Message-Id of that email is
<244b4c0f-8c79-7cd6-193e-54046ecf323c at redhat.com>, and the date is
"4/19/23, 15:18". My main statement in that email was that grub2 was
*architecturally incompatible* with UEFI, and I added:
> What I mean by architecturally incompatible: grub2 is designed from a
> perspective where it thinks it is an operating system; in other words,
> that it *owns* the computer. The problem is that UEFI thinks the exact
> same thing of itself, and of course the two conflict.
I elaborated a great detail on that, providing various examples, in
particular in relation to how grub used EFI_SIMPLE_FILE_SYSTEM_PROTOCOL.
Those findings were what made me deem my shim bugcompat patch futile,
after all -- I didn't post the patch because grub was unsalvageable
anyway, so I didn't see the point.
If you have a use case where you rely on shim but *not* on Grub (UKIs?),
then I'm OK relaxing the strictness of VirtioFsDxe. In that case, I'd
prefer upstreaming my above patch, from April, rather than taking yours.
What do you think about that?
--*--
Here's a further (independent) caveat: if you are using VirtioFsDxe with
the rust language virtiofsd, then you might experience hangs in
VirtioFsInit. For fixing that, you need the following *qemu* patch set:
[PATCH v3 0/7] vhost-user: call VHOST_USER_SET_VRING_ENABLE synchronously
https://patchew.org/QEMU/20231002203221.17241-1-lersek@redhat.com/
(This patch set has been on qemu-devel for nearly 2 months now, counting
from v1; I'm going to ping MST again. It's been ready for merging for
weeks now!)
Thanks!
Laszlo
-=-=-=-=-=-=-=-=-=-=-=-
Groups.io Links: You receive all messages sent to this group.
View/Reply Online (#109729): https://edk2.groups.io/g/devel/message/109729
Mute This Topic: https://groups.io/mt/102036263/1813853
Group Owner: devel+owner at edk2.groups.io
Unsubscribe: https://edk2.groups.io/g/devel/leave/3943202/1813853/130120423/xyzzy [edk2-devel-archive at redhat.com]
-=-=-=-=-=-=-=-=-=-=-=-
More information about the edk2-devel-archive
mailing list