Re: bootconfig length parse error in kernel

[Masami Hiramatsu]

On Thu, 12 Nov 2020 14:49:16 +0800
Chen Yu <yu.chen.surf@xxxxxxxxx> wrote:

> On Thu, Nov 12, 2020 at 1:50 PM Masami Hiramatsu <mhiramat@xxxxxxxxxx> wrote:
> >
> > Hi Chen,
> >
> > On Thu, 12 Nov 2020 12:34:36 +0800
> > Chen Yu <yu.chen.surf@xxxxxxxxx> wrote:
> >
> > > Hi Masami,
> > >
> > > On Wed, Nov 11, 2020 at 5:37 PM Masami Hiramatsu <mhiramat@xxxxxxxxxx> wrote:
> > > >
> > > > Hi Chen,
> > > >
> > > > On Tue, 10 Nov 2020 23:39:53 +0800
> > > > Chen Yu <yu.chen.surf@xxxxxxxxx> wrote:
> > > >
> > > > > Hi Masami,
> > > > > Thanks for writing bootconfig and it is useful for boot up trace event
> > > > > debugging.
> > > >
> > > > Thanks for testing!
> > > >
> > > > > However it was found that on 5.10-rc2 the bootconfig does not work and it shows
> > > > > "'bootconfig' found on command line, but no bootconfig found"
> > > > > And the reason for this is the kernel found the magic number to be incorrect.
> > > > > I've added some hack in kernel to dump the first 12 bytes, it shows:
> > > > > "OTCONFIG". So printed more content ahead we can find
> > > > > "#BOOTCONFIG" ahead. So it looks that there is some alignment during
> > > > > initrd load, and get_boot_config_from_initrd() might also deal with it. That is
> > > > > to say:
> > > > > data = (char *)initrd_end - BOOTCONFIG_MAGIC_LEN;
> > > > > might do some alignment?
> > > >
> > > > Hrm, interesting. So initrd_end might be aligned. Could you print out the
> > > > actuall address of initrd_end?
> > > I've done some investigation, it looks like this issue is not related
> > > to alignment, but related to
> > > the bootloader that has provided an inaccurate ramdisk size via
> > > boot_params.hdr.ramdisk_size.
> >
> > Yeah, it seems to happen. bootloader can pass wrong (bigger) size
> > to kernel. BTW, what bootloader would you use?
> >
> It is
> $ grub-install --version
> grub-install (GRUB) 2.04-1ubuntu26.2
> > > The actual size of initrd is:
> > > ls /boot/initrd.img-5.10.0-rc3-e1000e-hw+ -l
> > > -rw-r--r-- 1 root root 48689230 11月 12 00:08
> > > /boot/initrd.img-5.10.0-rc3-e1000e-hw+
> > > while the ramdisk size provided by bootloader via
> > > boot_params.hdr.ramdisk_size is
> > > 48689232, which is 2 bytes bigger than the actual size, and this is
> > > why the initrd_end
> > > is bigger than expected and causing the missmatch of magic number.
> >
> > OK. It seems that the bootloader might cut it up to 16 bytes
> > aligned. (But I think that's wrong behavior, there is no reason
> > to do it)
> Agree.
> >
> > > Since there is no guarantee that bootloader provides the accurate
> > > ramdisk size, an compromised
> > > proposal might be that to search for the magic number a little ahead.
> >
> > If the bootloader does such wrong behavior, there is no guarantee
> > that the size is "a little" bigger. IOW, it can be aligned to the
> > page size (4KB-)
> >
> Right. How about inserting the bootconfig at initrd_start if
> initrd_end could not be trusted?
> > > For example, the
> > > following patch works for me:
> > > diff --git a/init/main.c b/init/main.c
> > > index 130376ec10ba..60fb125d44f4 100644
> > > --- a/init/main.c
> > > +++ b/init/main.c
> > > @@ -273,7 +273,10 @@ static void * __init
> > > get_boot_config_from_initrd(u32 *_size, u32 *_csum)
> > >         if (!initrd_end)
> > >                 return NULL;
> > >
> > > -       data = (char *)initrd_end - BOOTCONFIG_MAGIC_LEN;
> > > +       data = memchr((char *)initrd_end - 2 * BOOTCONFIG_MAGIC_LEN,
> > > +                      '#', BOOTCONFIG_MAGIC_LEN);
> > > +       if (!data)
> > > +               return NULL;
> >
> > So this also does not guarantee that we can find "#" in BOOTCONFIG_MAGIC_LEN.
> > We need to find actual code in the bootloader, what it does.
> >
> Indeed.
> > >         if (memcmp(data, BOOTCONFIG_MAGIC, BOOTCONFIG_MAGIC_LEN))
> > >                 return NULL;
> > >
> > >
> > > > And could you tell me which platform are you tested?
> > > >
> > > It is HP ZHAN 99 Mobile Workstation G1 with i5-8300H, Ubuntu 20.04.
> >
> > Hmm, this means x86 Grub2 does this change. Let me check it.
> >

I found the 4 byte alignment code in the grub 
(grub_initrd_init()grub-core/loader/linux.c), but that seems to happen
only when load a file with newc:/PACKAGE/FILE format. (And this is not
documented.) At a glance, u-boot may not do that (but of course user
can pass different size directly by command), EDK2 doesn't too.
So, we should check at least 3 byte back for grub. 

BTW, just out of curious, what is your "initrd" command line in grub.conf? :)

Thank you,

-- 
Masami Hiramatsu <mhiramat@xxxxxxxxxx>