Message-ID: <20180718053818.GF11258@linaro.org>
Date:   Wed, 18 Jul 2018 14:38:20 +0900
From:   AKASHI Takahiro <takahiro.akashi@...aro.org>
To:     Dave Young <dyoung@...hat.com>
Cc:     James Morse <james.morse@....com>, catalin.marinas@....com,
        will.deacon@....com, dhowells@...hat.com, vgoyal@...hat.com,
        herbert@...dor.apana.org.au, davem@...emloft.net, bhe@...hat.com,
        arnd@...db.de, ard.biesheuvel@...aro.org, bhsharma@...hat.com,
        kexec@...ts.infradead.org, linux-arm-kernel@...ts.infradead.org,
        linux-kernel@...r.kernel.org,
        "Eric W. Biederman" <ebiederm@...ssion.com>
Subject: Re: [PATCH v11 03/15] powerpc, kexec_file: factor out memblock-based
 arch_kexec_walk_mem()

Dave,

On Tue, Jul 17, 2018 at 03:49:23PM +0800, Dave Young wrote:
> Hi AKASHI,
> On 07/17/18 at 02:31pm, AKASHI Takahiro wrote:
> > Hi Dave,
> > 
> > On Mon, Jul 16, 2018 at 08:24:12PM +0800, Dave Young wrote:
> > > On 07/16/18 at 12:04pm, James Morse wrote:
> > > > Hi Dave,
> > > > 
> > > > On 14/07/18 02:52, Dave Young wrote:
> > > > > On 07/11/18 at 04:41pm, AKASHI Takahiro wrote:
> > > > >> Memblock list is another source for usable system memory layout.
> > > > >> So powerpc's arch_kexec_walk_mem() is moved to kexec_file.c so that
> > > > >> other memblock-based architectures, particularly arm64, can also utilise
> > > > >> it. A moved function is now renamed to kexec_walk_memblock() and merged
> > > > >> into the existing arch_kexec_walk_mem() for general use, either resource
> > > > >> list or memblock list.
> > > > >>
> > > > >> A consequent function will not work for kdump with memblock list, but
> > > > >> this will be fixed in the next patch.
> > > > 
> > > > >> diff --git a/kernel/kexec_file.c b/kernel/kexec_file.c
> > > > 
> > > > >> @@ -513,6 +563,10 @@ static int locate_mem_hole_callback(struct resource *res, void *arg)
> > > > >>  int __weak arch_kexec_walk_mem(struct kexec_buf *kbuf,
> > > > >>  			       int (*func)(struct resource *, void *))
> > > > >>  {
> > > > >> +	if (IS_ENABLED(CONFIG_HAVE_MEMBLOCK) &&
> > > > >> +			!IS_ENABLED(CONFIG_ARCH_DISCARD_MEMBLOCK))
> > > > >> +		return kexec_walk_memblock(kbuf, func);
> > > > > 
> > > > > AKASHI, I'm not sure if this works on all arches, for example I chekced
> > > > > the .config on my Nokia N900 kernel tree, there is HAVE_MEMBLOCK=y and
> > > > > no CONFIG_ARCH_DISCARD_MEMBLOCK, in 32bit arm code no arch_kexec_walk_mem()
> > > > By doesn't work you mean it's a change in behaviour?
> > > > I think this is fine because 32bit arm doesn't support KEXEC_FILE, (this file is
> > > > kexec_file specific right?).
> > > 
> > > Ah, replied on a train, I forgot this is only for kexec_file, sorry
> > > about that.  Please ignore the comment.
> > > 
> > > But since we have a weak function arch_kexec_walk_mem, adding another
> > > condition branch within this weak function looks not good.
> > > Something like below would be better:
> > 
> > I see your concern here, but
> > 
> > 
> > > int kexec_locate_mem_hole(struct kexec_buf *kbuf)
> > > {
> > >         int ret;
> > > 
> > > 	+ if use memblock
> > > 	+	ret = kexec_walk_memblock()
> > > 	+ else
> > >         	ret = arch_kexec_walk_mem(kbuf, locate_mem_hole_callback);
> > > 
> > >         return ret == 1 ? 0 : -EADDRNOTAVAIL;
> > > }
> > 
> > what if yet another architecture comes to kexec_file and wanna
> > take a third approach? How can it override those functions?
> > Depending on kernel configuration, it might re-define either
> > kexec_walk_memblock() or arch_kexec_walk_mem(). It sounds weird to me.
> 
> I also feel this weird, but it is slightly better because currently no
> user need another overriding requirement, and I feel it is not expected to have in
> the future for the memblock use.
> 
> Rethinking about this issue, we can just remove the weak function and
> just use general function.

Do you really want to remove "weak" attribute?

> Currently with your patch applied only s390 use arch_kexec_walk_mem like
> below:
> /*
>  * The kernel is loaded to a fixed location. Turn off kexec_locate_mem_hole
>  * and provide kbuf->mem by hand.
>  */
> int arch_kexec_walk_mem(struct kexec_buf *kbuf,
>                         int (*func)(struct resource *, void *))
> {
>         return 1;
> }
> 
> AFAIK, all other users initialize kbuf->mem as NULL, so we can check

As a matter of fact, nobody initializes kbuf->mem before calling
kexec_add_buffer (in turn, kexec_locate_mem_hole()).

> kbuf->mem in int kexec_locate_mem_hole:
> 
> if (kbuf->mem)
> 	return 0;
> 
> if use memblock
> 	kexec_walk_memblock
> else
> 	kexec_walk_mem

I think that your solution will work for existing architectures
with appropriate patches, but to take your approach, as I said above,
we will have to modify every call site on all kexec_file-capable architectures.

If this is what you expect, I will work on it, but I don't think
that it would be a better idea.

Thanks,
-Takahiro AKASHI

> > 
> > Thanks,
> > -Takahiro AKASHI
> > 
> > > 
> > > > 
> > > > It only affects architectures with MEMBLOCK and KEXEC_FILE: powerpc, s390 and
> > > > soon arm64. s390 keeps its behaviour because it provides arch_kexec_walk_mem(),
> > > > and powerpc's is copied in here as its generic 'memblock describes my memory'
> > > > stuff. The implementation would be the same on arm64, so we're doing this to
> > > > avoid duplicating otherwise generic arch code. I think 32bit arm should be able
> > > > to use this too if it gets KEXEC_FILE support. (32bit arms' KEXEC already
> > > > depends on MEMBLOCK).
> > > > 
> > > > 
> > > > Thanks,
> > > > 
> > > > James
> > > 
> > > Thanks
> > > Dave
> 
> Thanks
> Dave

Powered by blists - more mailing lists