Re: 32bit kernel is broken for Linux-5.1-rc1 due to GCC cmodel=medlow

From: Anup Patel <anup@brainfault.org>
To: Mike Rapoport <rppt@linux.ibm.com>
Cc: Damien Le Moal <Damien.LeMoal@wdc.com>,
	Anup Patel <anup.patel@wdc.com>,
	Palmer Dabbelt <palmer@sifive.com>,
	Christoph Hellwig <hch@infradead.org>,
	Atish Patra <atish.patra@wdc.com>,
	linux-riscv@lists.infradead.org
Subject: Re: 32bit kernel is broken for Linux-5.1-rc1 due to GCC cmodel=medlow
Date: Fri, 22 Mar 2019 17:59:07 +0530	[thread overview]
Message-ID: <CAAhSdy3UoX9gS=9YN6m0=qA2g-m=bw-Y-N-5CAFQZcnKA04C5g@mail.gmail.com> (raw)
In-Reply-To: <20190322104209.GB24367@rapoport-lnx>

On Fri, Mar 22, 2019 at 4:12 PM Mike Rapoport <rppt@linux.ibm.com> wrote:
>
> Hi,
>
> On Fri, Mar 22, 2019 at 11:46:24AM +0530, Anup Patel wrote:
> > Hi Palmer,
> >
> > The 32bit kernel booting is broken for Linux-5.1-rc1 due to GCC cmodel=medlow
> > affecting setup_vm() movement from kernel/setup.c to mm/init.c.
> >
> > There is no issue with 64bit kernel booting.
> >
> > The "[PATCH v2 2/5] RISC-V: Make setup_vm() independent of GCC code model"
> > fixes this issue.
> >
> > https://www.mail-archive.com/linux-kernel@vger.kernel.org/msg1959102.html
> >
> > If possible please include above patch as Linux-5.1-rc1 fix.
>
> I'm not convinced that it's the best solution. Multiple __load_pa() and
> __load_va() conversions make code unreadable.

We have many __load_pa() and __load_va() in setup_vm() here because
this patch only get's things working for 32bit RISC-V kernel.

If you look at PATCH3 of my series then you will see that use of __load_pa()
and __load_va() is simplified a lot and is easily readable. This was possible
by rewriting the initial page table setup code.

> Is there any reason swapper_pg_dir cannot be setup after 'relocate'? It'll
> save a lot of churn for the current fix and for the addition of 4K mappings
>
> I've drafted a patch that separates trampoline_pd_dir and swapper_pg_dir
> setup, if it works, the __load_pa() and __load_va() conversions can be
> applied only to the trampoline initialization. I don't have riscv hardware,
> so it's compile tested only.

I had similar views as yours but I found that trampoling_pg_dir is totally
redundant and since it is part of __init section it will break runtime hart
hotplug in future. All mapping in trampoling_pg_dir are already covered
by swapper_pg_dir. The PATCH4 in my v2 series removes trampoline_pg_dir.

We just need to map vmlinux_start to vmlinux_end and DTB in setup_vm()
that's all. Complete memory is only mapped in setup_vm_final().

Regards,
Anup

_______________________________________________
linux-riscv mailing list
linux-riscv@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-riscv