On Mon, Oct 25, 2021 at 03:11:57PM +0200, Ard Biesheuvel wrote:
> On Mon, 25 Oct 2021 at 14:47, Chen Yu <yu.c.chen@xxxxxxxxx> wrote:
> >
> > On Mon, Oct 25, 2021 at 02:02:24PM +0200, Greg Kroah-Hartman wrote:
> > > On Mon, Oct 25, 2021 at 07:45:19PM +0800, Chen Yu wrote:
> > > > On Mon, Oct 25, 2021 at 08:44:00AM +0200, Greg Kroah-Hartman wrote:
> > > > > On Mon, Oct 25, 2021 at 02:25:04PM +0800, Chen Yu wrote:
> > > > > > Platform Firmware Runtime Update image starts with UEFI headers, and the
> > > > > > headers are defined in UEFI specification, but some of them have not been
> > > > > > defined in the kernel yet.
> > > > > >
> > > > > > For example, the header layout of a capsule file looks like this:
> > > > > >
> > > > > > EFI_CAPSULE_HEADER
> > > > > > EFI_FIRMWARE_MANAGEMENT_CAPSULE_HEADER
> > > > > > EFI_FIRMWARE_MANAGEMENT_CAPSULE_IMAGE_HEADER
> > > > > > EFI_FIRMWARE_IMAGE_AUTHENTICATION
> > > > > >
> > > > > > These structures would be used by the Platform Firmware Runtime Update
> > > > > > driver to parse the format of capsule file to verify if the corresponding
> > > > > > version number is valid. The EFI_CAPSULE_HEADER has been defined in the
> > > > > > kernel, however the rest are not, thus introduce corresponding UEFI
> > > > > > structures accordingly. Besides, EFI_FIRMWARE_MANAGEMENT_CAPSULE_HEADER
> > > > > > and EFI_FIRMWARE_MANAGEMENT_CAPSULE_IMAGE_HEADER need not be aligned and
> > > > > > so the corresponding data types should be packed.
> > > > > >
> > > > > > Signed-off-by: Chen Yu <yu.c.chen@xxxxxxxxx>
> > > > > > ---
> > > > > > v6: No change since v5.
> > > > > > v5: No change since v4.
> > > > > > v4: Revise the commit log to make it more clear. (Rafael J. Wysocki)
> > > > > > ---
> > > > > > include/linux/efi.h | 50 +++++++++++++++++++++++++++++++++++++++++++++
> > > > > > 1 file changed, 50 insertions(+)
> > > > > >
> > > > > > diff --git a/include/linux/efi.h b/include/linux/efi.h
> > > > > > index 6b5d36babfcc..19ff834e1388 100644
> > > > > > --- a/include/linux/efi.h
> > > > > > +++ b/include/linux/efi.h
> > > > > > @@ -148,6 +148,56 @@ typedef struct {
> > > > > > u32 imagesize;
> > > > > > } efi_capsule_header_t;
> > > > > >
> > > > > > +#pragma pack(1)
> > > > >
> > > > > Why is this pragma suddenly needed now in this file?
> > > > >
> > > > > If you really need this for a specific structure, use the "__packed"
> > > > > attribute please.
> > > > >
> > > > These two structures are required to be packed in the uefi spec, I'll change
> > > > them to "__packed".
> > >
> > > And they are the _only_ ones in this .h file that require this? I would
> > > think that they all require this.
> > >
> > I did a search in the uefi specification, and found 42 pack(1) structures,
> > while the other structures do not have pack(1) attribute.
> >
> > It seems to me that whether the structures are required to be strictly packed
> > depends on the use case. Here's my understanding and I might be wrong: In this
> > patch, according to the skeleton of capsule file described in
> > [Figure 23-6 Firmware Management and Firmware Image Management headers]
> > in the uefi spec [1], the two structures are located at the beginning of
> > the capsule file, and followed by real payload. If these structure are packed
> > then the the adjacent binary payload could start on byte boundary without
> > padding, which might save space for capsule file.
> >
>
> Packing only affects internal padding, and a struct's size is never
> padded to be a multiple of its alignment (which equals the largest
> alignment of all its members). This of course assumes that you don't
> abuse array indexing as a sizeof() operator.
>
> However, the __packed attribute does indicate to the compiler that the
> entire thing can appear misaligned in memory. So if one follows the
> other in the capsule header, the __packed attribute may be appropriate
> to ensure that the second one is not accessed using misaligned loads
> and stores.
>
> And then there is of course the ambiguity in alignment of uint64_t on
> x86, which could be either 4 or 8 bytes depending on the context (and
> UEFI targets all of them). So __packed may be used to disambiguate
> between those if a uint64_t field appears on a boundary whose offset %
> 8 == 4.
>
> So please use __packed rather than the pragma(), and apply it wherever
> it is applied in the spec.
>
Thanks for the explaination in detail! Ard. Will do in next version.
Thanks,
Chenyu
