Received-SPF: pass (google.com: domain of linux-kernel+bounces-33128-linux.lists.archive=gmail.com@vger.kernel.org designates 139.178.88.99 as permitted sender) client-ip=139.178.88.99;
Date: Mon, 22 Jan 2024 14:02:29 +0000
Message-ID: <86r0i98o0a.wl-maz@kernel.org>
From: Marc Zyngier <maz@kernel.org>
To: Tangnianyao <tangnianyao@huawei.com>
Cc: <tglx@linutronix.de>,
	<linux-kernel@vger.kernel.org>,
	<linux-arm-kernel@lists.infradead.org>,
	<guoyang2@huawei.com>,
	<wangwudi@hisilicon.com>
Subject: Re: [PATCH] irqchip/gic-v4.1:Check whether indirect table is supported in allocate_vpe_l1_table
In-Reply-To: <5de3da53-9c0d-2a2d-876b-2181e540fa2f@huawei.com>
References: <20240122160607.1078960-1-tangnianyao@huawei.com>
	<86sf2p91zt.wl-maz@kernel.org>
	<5de3da53-9c0d-2a2d-876b-2181e540fa2f@huawei.com>
User-Agent: Wanderlust/2.15.9 (Almost Unreal) SEMI-EPG/1.14.7 (Harue)
 FLIM-LB/1.14.9 (=?UTF-8?B?R29qxY0=?=) APEL-LB/10.8 EasyPG/1.0.0 Emacs/29.1
 (aarch64-unknown-linux-gnu) MULE/6.0 (HANACHIRUSATO)
Precedence: bulk
MIME-Version: 1.0 (generated by SEMI-EPG 1.14.7 - "Harue")
Content-Type: text/plain; charset=US-ASCII

On Mon, 22 Jan 2024 13:13:09 +0000,
Tangnianyao <tangnianyao@huawei.com> wrote:
>
> On 1/22/2024 17:00, Marc Zyngier wrote:
> > [Fixing the LKML address, which has bits of Stephan's address embedded
> > in it...]
> >
> > On Mon, 22 Jan 2024 16:06:07 +0000,
> > Nianyao Tang <tangnianyao@huawei.com> wrote:
> >> In allocate_vpe_l1_table, when we fail to inherit VPE table from other
> >> redistributors or ITSs, and we allocate a new vpe table for current common 
> >> affinity field without checking whether indirect table is supported.
> >> Let's fix it.
> > Is there an actual implementation that doesn't support the indirect
> > property for the VPE table? I know this is allowed for consistency
> > with the original revision of the architecture, but I never expected
> > an actual GICv4.1 implementation to be *that* bad.
> >
> > If that's the case, I'm a bit puzzled/worried.
> 
> I met this problem in a developing implementation and find it's allowed by GIC spec.
> In such environment,  in a common affinity field with only redistributors and without
> any ITS in it, forcing its_vpe_id_alloc to allocate a large vpeid(like 65000), and there
> comes an error message "VPE IRQ allocation failure". It originally comes from
> allocate_vpe_l2_table, reading GICR_VPROPBASER with GICR_VPROPBASER_4_1_SIZE=1
> and GICR_VPROPBASER_4_1_INDIRECT=0.

Really, you should get your HW engineers to fix their GIC
implementation.  I'm OK with working around this issue for
completeness, but shipping such an implementation would be a mistake.

[...]

> I have another question here. The max number of pages  for GITS_BASER
> and GICR_VPROPBASER is different here, while GITS_BASER.Size is
> bit[7:0] with max 256, and GICR_4_1_VPROPBASER.Size is bit[6:0] with max 128.
> Kernel usually probe ITS basers first and then probe GICR_4_1_VPROPBASER in
> a common affinity group. Maybe we need to check this in "inherit_vpe_l1_table_from_its" ?

This is because GITS_BASER[] is generic (also works for devices and
collections), while GICR_VPROPBASER is tailored to the VPE table which
is usually smaller.

I would expect that GICD_TYPER2.VID reports something that cannot
result in something going wrong (in this case, the L1 allocation
cannot be more than 128 pages).

Overall, the kernel isn't a validation suite for the HW, and we expect
it to have some level of sanity. So if none of this is in shipping HW
but only in some model with crazy parameters, I don't think we should
go out of our way to support it.

Thanks,

	M.

-- 
Without deviation from the norm, progress is not possible.