Re: [PATCH v2 0/2] arm64: Cut rebuild time when changing CONFIG_BLK_DEV_INITRD

[Rob Herring <robh@kernel.org>]

On Thu, Oct 25, 2018 at 12:30 PM Mike Rapoport <rppt@linux.ibm.com> wrote:
>
> On Thu, Oct 25, 2018 at 08:15:15AM -0500, Rob Herring wrote:
> > +Ard
> >
> > On Thu, Oct 25, 2018 at 4:38 AM Mike Rapoport <rppt@linux.ibm.com> wrote:
> > >
> > > On Wed, Oct 24, 2018 at 02:55:17PM -0500, Rob Herring wrote:
> > > > On Wed, Oct 24, 2018 at 2:33 PM Florian Fainelli <f.fainelli@gmail.com> wrote:
> > > > >
> > > > > Hi all,
> > > > >
> > > > > While investigating why ARM64 required a ton of objects to be rebuilt
> > > > > when toggling CONFIG_DEV_BLK_INITRD, it became clear that this was
> > > > > because we define __early_init_dt_declare_initrd() differently and we do
> > > > > that in arch/arm64/include/asm/memory.h which gets included by a fair
> > > > > amount of other header files, and translation units as well.
> > > >
> > > > I scratch my head sometimes as to why some config options rebuild so
> > > > much stuff. One down, ? to go. :)
> > > >
> > > > > Changing the value of CONFIG_DEV_BLK_INITRD is a common thing with build
> > > > > systems that generate two kernels: one with the initramfs and one
> > > > > without. buildroot is one of these build systems, OpenWrt is also
> > > > > another one that does this.
> > > > >
> > > > > This patch series proposes adding an empty initrd.h to satisfy the need
> > > > > for drivers/of/fdt.c to unconditionally include that file, and moves the
> > > > > custom __early_init_dt_declare_initrd() definition away from
> > > > > asm/memory.h
> > > > >
> > > > > This cuts the number of objects rebuilds from 1920 down to 26, so a
> > > > > factor 73 approximately.
> > > > >
> > > > > Apologies for the long CC list, please let me know how you would go
> > > > > about merging that and if another approach would be preferable, e.g:
> > > > > introducing a CONFIG_ARCH_INITRD_BELOW_START_OK Kconfig option or
> > > > > something like that.
> > > >
> > > > There may be a better way as of 4.20 because bootmem is now gone and
> > > > only memblock is used. This should unify what each arch needs to do
> > > > with initrd early. We need the physical address early for memblock
> > > > reserving. Then later on we need the virtual address to access the
> > > > initrd. Perhaps we should just change initrd_start and initrd_end to
> > > > physical addresses (or add 2 new variables would be less invasive and
> > > > allow for different translation than __va()). The sanity checks and
> > > > memblock reserve could also perhaps be moved to a common location.
> > > >
> > > > Alternatively, given arm64 is the only oddball, I'd be fine with an
> > > > "if (IS_ENABLED(CONFIG_ARM64))" condition in the default
> > > > __early_init_dt_declare_initrd as long as we have a path to removing
> > > > it like the above option.
> > >
> > > I think arm64 does not have to redefine __early_init_dt_declare_initrd().
> > > Something like this might be just all we need (completely untested,
> > > probably it won't even compile):
> > >
> > > diff --git a/arch/arm64/mm/init.c b/arch/arm64/mm/init.c
> > > index 9d9582c..e9ca238 100644
> > > --- a/arch/arm64/mm/init.c
> > > +++ b/arch/arm64/mm/init.c
> > > @@ -62,6 +62,9 @@ s64 memstart_addr __ro_after_init = -1;
> > >  phys_addr_t arm64_dma_phys_limit __ro_after_init;
> > >
> > >  #ifdef CONFIG_BLK_DEV_INITRD
> > > +
> > > +static phys_addr_t initrd_start_phys, initrd_end_phys;
> > > +
> > >  static int __init early_initrd(char *p)
> > >  {
> > >         unsigned long start, size;
> > > @@ -71,8 +74,8 @@ static int __init early_initrd(char *p)
> > >         if (*endp == ',') {
> > >                 size = memparse(endp + 1, NULL);
> > >
> > > -               initrd_start = start;
> > > -               initrd_end = start + size;
> > > +               initrd_start_phys = start;
> > > +               initrd_end_phys = end;
> > >         }
> > >         return 0;
> > >  }
> > > @@ -407,14 +410,27 @@ void __init arm64_memblock_init(void)
> > >                 memblock_add(__pa_symbol(_text), (u64)(_end - _text));
> > >         }
> > >
> > > -       if (IS_ENABLED(CONFIG_BLK_DEV_INITRD) && initrd_start) {
> > > +       if (IS_ENABLED(CONFIG_BLK_DEV_INITRD) &&
> > > +           (initrd_start || initrd_start_phys)) {
> > > +               /*
> > > +                * FIXME: ensure proper precendence between
> > > +                * early_initrd and DT when both are present
> >
> > Command line takes precedence, so just reverse the order.
> >
> > > +                */
> > > +               if (initrd_start) {
> > > +                       initrd_start_phys = __phys_to_virt(initrd_start);
> > > +                       initrd_end_phys = __phys_to_virt(initrd_end);

BTW, I think you meant virt_to_phys() here?

> >
> > AIUI, the original issue was doing the P2V translation was happening
> > too early and the VA could be wrong if the linear range is adjusted.
> > So I don't think this would work.
>
> Probably things have changed since then, but in the current code there is
>
>                 initrd_start = __phys_to_virt(initrd_start);
>
> and in between only the code related to CONFIG_RANDOMIZE_BASE, so I believe
> it's safe to use __phys_to_virt() here as well.

Here is fine yes, but I believe it was the the phys to virt in the DT
code before adjusting the linear range that was the problem.

Rob