Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Sat, 09 May 2020 18:43:30 +1000
From:   Nicholas Piggin <npiggin@gmail.com>
Subject: Re: ioremap() called early from pnv_pci_init_ioda_phb()
To:     Qian Cai <cai@lca.pw>, Oliver O'Halloran <oohall@gmail.com>
Cc:     LKML <linux-kernel@vger.kernel.org>,
        linuxppc-dev <linuxppc-dev@lists.ozlabs.org>
References: <B183CDAA-DA88-4760-9C1B-F73A8F7840E7@lca.pw>
        <CAOSf1CFNp6+k_y_87r7p2e8cKfX0rK-9wBxeR+K0e0y8R0_TNg@mail.gmail.com>
In-Reply-To: <CAOSf1CFNp6+k_y_87r7p2e8cKfX0rK-9wBxeR+K0e0y8R0_TNg@mail.gmail.com>
MIME-Version: 1.0
Message-Id: <1589013519.0fzm2px5cz.astroid@bobo.none>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: quoted-printable
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

Excerpts from Oliver O'Halloran's message of May 9, 2020 6:11 pm:
> On Sat, May 9, 2020 at 12:41 AM Qian Cai <cai@lca.pw> wrote:
>>
>>  Booting POWER9 PowerNV has this message,
>>
>> "ioremap() called early from pnv_pci_init_ioda_phb+0x420/0xdfc. Use earl=
y_ioremap() instead=E2=80=9D
>>
>> but use the patch below will result in leaks because it will never call =
early_iounmap() anywhere. However, it looks me it was by design that phb->r=
egs mapping would be there forever where it would be used in pnv_ioda_get_i=
nval_reg(), so is just that check_early_ioremap_leak() initcall too strong?
>=20
> The warning there is junk. The PHBs are setup at boot and never torn
> down so we're not "leaking" the mapping. It's supposed to be there for
> the lifetime of the kernel.
>=20
> That said, we could probably move the PCI setup to a point later in
> boot where the normal ioremap can be be used. We would have to check
> for initcalls which depend on the PHBs being setup and delay those too
> though.

I think it helps to unify code a bit more and take special cases out of=20
ioremap to have all these early calls use early_ioremap.

We actually do want to move these later if possible too, on radix they=20
use memblock for page tables, and on hash they don't even set up proper=20
kernel page tables but just bolt PTEs into the hash table.

Thanks,
Nick