[PATCHv5 0/8] arm64: zboot support

[Dave Young]

On Thu, 20 Jul 2023 at 16:59, Pingfan Liu <piliu at redhat.com> wrote:
>
> On Thu, Jul 20, 2023 at 3:27 PM Dave Young <dyoung at redhat.com> wrote:
> >
> > Hi Pingfan,
> >
> > On Thu, 20 Jul 2023 at 10:05, Pingfan Liu <piliu at redhat.com> wrote:
> > >
> > > Hi Dave,
> > >
> > > Thanks for your insight. Please see the comments inline below.
> > >
> > > On Wed, Jul 19, 2023 at 11:00 AM Dave Young <dyoung at redhat.com> wrote:
> > > >
> > > > Hi Pingfan, Simon,
> > > >
> > > > On 07/17/23 at 09:07pm, Pingfan Liu wrote:
> > > > > As more complicated capsule kernel format occurs like zboot, where the
> > > > > compressed kernel is stored as a payload. The straight forward
> > > > > decompression can not meet the demand.
> > > > >
> > > > > As the first step, on aarch64, reading in the kernel file in a probe
> > > > > method and decide how to unfold the content by the method itself.
> > > > >
> > > > > This series introduce a new image probe interface probe2(), which
> > > > > returns three factors: kernel buffer, kernel size and kernel fd through
> > > > > a struct parsed_info.
> > > > > -1. the parsed kernel_buf should be returned so that it can be used by
> > > > > the image load method later.
> > > > > -2. the final fd passed to sys_kexec_file_load, since aarch64 kernel can
> > > > > only work with Image format, the outer payload should be stripped and a
> > > > > temporary file of Image should be created.
> > > >
> > > > I took a look at the Image.gz file load code, the current code can be
> > > > simplified with passing a fd directly instead of creating temp files via
> > > > memfd_create with the already decompressed kernel_buf.
> > > >
> > > > The current file load is like below:
> > > >
> > > > do_kexec_file_load():
> > > >   1.slurp_decompress_file
> > > >     2. probe
> > > >       3. load
> > > >         4. kexec_file_load
> > > >
> > > > In step 1, the Image.gz has been decompressed to kernel_buf, so just
> > > > create a virtual memfd copy to it, then save the virtual fd for step 4
> > > > use.
> > > >
> > > > Otherwise in step 2 it is some sanity checking, step 3 is setting
> > > > something else eg. initrd_fd, cmdline. With the changes below Image and
> > > > Image.gz will share same code. I think you can add the zboot
> > > > detection/checking code in the Image probe, load functions, with a new
> > > > info->kernel_fd, you can decompress the zboot kernel_buf and save to
> > > > another virtual memfd, and set to the info->kernel_fd.  Then in step 4
> > > > the kexec_file_load can just use it.
> > > >
> > >
> > > This only results in a minor change in the interface, not like mine. I
> > > prefer this method.
> > >
> > > > The kernel_buf itself is only used for sanity checking of the formats,
> > > > kernel only needs a file fd, so I think it should be fine and easier
> > > > than the original ways.
> > > >
> > >
> > > Overall, this new method minimally affects the function interface, in
> > > addition to the code simplification.
> > > I will try to take it in the next version.
> > >
> > Ok, great.  Btw, it would be ideal if some of the pez_prepare work can
> > be done early before the slurp_decompress_file in
> > do_kexec_file_load(), then it could be just arch independent :)
> >
>
> I assume that you suggest the following big picture:
>     if trying slurp_decompress_file()
>     else trying pez_prepare()
>     for (i = 0; i < file_types; i++) {
>       if (file_type[i].probe(kernel_buf, kernel_size) >= 0)
>    }
>
> But there is some check before pez_prepare() in pez_arm64_probe(), I
> am not sure whether they can be skipped or not.

Ok, got it, if the zimage header of arm64 is arch specific , it should
still go to the arch part.

Thanks
Dave