Re: [PATCH 08/15] efi/x86: Move command-line initrd loading to efi_main

From: Arvind Sankar <nivedita@alum.mit.edu>
To: Dan Williams <dan.j.williams@intel.com>
Cc: Arvind Sankar <nivedita@alum.mit.edu>,
	"linux-efi@vger.kernel.org" <linux-efi@vger.kernel.org>,
	"mingo@kernel.org" <mingo@kernel.org>,
	"tglx@linutronix.de" <tglx@linutronix.de>,
	"ardb@kernel.org" <ardb@kernel.org>,
	"linux@roeck-us.net" <linux@roeck-us.net>,
	"joe@perches.com" <joe@perches.com>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	"arnd@arndb.de" <arnd@arndb.de>
Subject: Re: [PATCH 08/15] efi/x86: Move command-line initrd loading to efi_main
Date: Wed, 27 May 2020 19:02:09 -0400	[thread overview]
Message-ID: <20200527230209.GA3575079@rani.riverdale.lan> (raw)
In-Reply-To: <CAPcyv4jOCY=kxeZVWsS0Xc36jmPr7DSR_sFrsMeoiEs+iEfbEA@mail.gmail.com>

On Wed, May 27, 2020 at 03:56:45PM -0700, Dan Williams wrote:
> On Wed, May 27, 2020 at 3:47 PM Arvind Sankar <nivedita@alum.mit.edu> wrote:
> >
> > On Wed, May 27, 2020 at 10:30:18PM +0000, Williams, Dan J wrote:
> > > On Fri, 2020-05-08 at 20:01 +0200, Ard Biesheuvel wrote:
> > > > From: Arvind Sankar <nivedita@alum.mit.edu>
> > > >
> > > > Consolidate the initrd loading in efi_main.
> > > >
> > > > The command line options now need to be parsed only once.
> > > >
> > > > Signed-off-by: Arvind Sankar <nivedita@alum.mit.edu>
> > > > Link:
> > > > https://lore.kernel.org/r/20200430182843.2510180-9-nivedita@alum.mit.edu
> > > > Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
> > >
> > > Hi,
> > >
> > > This patch patch in tip/master as:
> > >
> > > 987053a30016 efi/x86: Move command-line initrd loading to efi_main
> > >
> > > ...regresses my nfs root configuration. It hangs trying to mount the
> > > nfs root filesystem "root=/dev/nfs ip=dhcp".
> > >
> > > It does not revert cleanly.
> > >
> > >
> >
> > Does this fix it?
> >
> > diff --git a/drivers/firmware/efi/libstub/x86-stub.c b/drivers/firmware/efi/libstub/x86-stub.c
> > index defeb6035109..f53362efef84 100644
> > --- a/drivers/firmware/efi/libstub/x86-stub.c
> > +++ b/drivers/firmware/efi/libstub/x86-stub.c
> > @@ -771,10 +771,12 @@ unsigned long efi_main(efi_handle_t handle,
> >                         efi_err("Failed to load initrd!\n");
> >                         goto fail;
> >                 }
> > -               efi_set_u64_split(addr, &hdr->ramdisk_image,
> > -                                 &boot_params->ext_ramdisk_image);
> > -               efi_set_u64_split(size, &hdr->ramdisk_size,
> > -                                 &boot_params->ext_ramdisk_size);
> > +               if (size > 0) {
> > +                       efi_set_u64_split(addr, &hdr->ramdisk_image,
> > +                                         &boot_params->ext_ramdisk_image);
> > +                       efi_set_u64_split(size, &hdr->ramdisk_size,
> > +                                         &boot_params->ext_ramdisk_size);
> > +               }
> 
> I'll give it a shot, but my guess would have been something related to
> the fact that this patch moves the initrd loading relative to when the
> command line is being parsed. In this case it's a dracut initrd built
> by:
> 
>     dracut -m "nfs network base"
> 
> ...with a kernel built with:
> 
> CONFIG_IP_PNP_DHCP=y
> 
> ...and a built-in network interface. The behavior seems to be that the
> kernel gets an IP address just fine, but there's no initrd userspace
> to mount nfs and the kernel eventually gives up looking for root.

It's an oversight in this patch: I set addr/size to 0 in the case where
the EFI stub is not supposed to handle the initrd loading (because a
bootloader ran before it and was responsible for handling the loading),
but then those 0's get written into the bootparams structure anyway,
blowing away whatever the bootloader had loaded.