Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Subject: Re: [PATCH] irqchip/gic-v4.1: Properly lock VPEs when doing a
 directLPI invalidation
To:     Marc Zyngier <maz@kernel.org>
CC:     <linux-kernel@vger.kernel.org>,
        Thomas Gleixner <tglx@linutronix.de>,
        Kunkun Jiang <jiangkunkun@huawei.com>,
        <wanghaibin.wang@huawei.com>
References: <20230617073242.3199746-1-maz@kernel.org>
 <b665e102-eba9-cd4f-5b4e-ccbd77444233@huawei.com>
 <86wmzgx1va.wl-maz@kernel.org>
From:   Zenghui Yu <yuzenghui@huawei.com>
Message-ID: <ee9123ca-95fc-4482-cee1-965cfe6b89b4@huawei.com>
Date:   Tue, 4 Jul 2023 23:42:32 +0800
User-Agent: Mozilla/5.0 (Windows NT 10.0; WOW64; rv:68.0) Gecko/20100101
 Thunderbird/68.9.0
MIME-Version: 1.0
In-Reply-To: <86wmzgx1va.wl-maz@kernel.org>
Content-Type: text/plain; charset="utf-8"; format=flowed
Content-Language: en-US
Content-Transfer-Encoding: 7bit
Precedence: bulk

Hi Marc,

On 2023/7/4 2:54, Marc Zyngier wrote:
> On Thu, 29 Jun 2023 15:52:24 +0100,
> Zenghui Yu <yuzenghui@huawei.com> wrote:
>>
>> Hi Marc,
>>
>> Nit: I think the Subject header can be changed to 'irqchip/gic-v4' as
>> the bug it fixes only affects GICv4 HW. v4.1 is unaffected.
> 
> I'm not so sure.
> 
> v4.0 didn't allow direct invalidation of VPE doorbells (we had to use
> the fake device hack), except for the HiSi special that implemented
> DirectLPI despite the presence of multiple ITSs. It was a violation of
> the architecture, but it really saved the day by making invalidations
> cheap enough.

[ I should've mentioned that I had reproduced the bug and tested your
patch on my 920, which is, yeah, a HiSi implementation of GICv4.0 with
DirectLPI supported. But ]

> 
> Only with v4.1 did we get architectural support for doorbell
> invalidation via a register instead of a command for a fake device.
> 
> So as far as the architecture is concerned, this should only affect
> v4.1. As a side effect, it also affect HiSi's v4.0 implementations.

... iiuc the bug we're fixing is that we perform a register based
invalidation for doorbells (via its_vpe_[un]mask_irq/its_vpe_send_inv),
acquire and release the per-RD lock with a *race* against a concurrent
its_map_vm(), which would modify the vpe->col_idx behind our backs and
affect the lock we're looking for.

its_vpe_[un]mask_irq() are callbacks for the v4.0 irqchip, i.e.,
its_vpe_irq_chip.

With v4.1, we switch to use its_vpe_4_1_irq_chip and invalidate
doorbells by sending the new INVDB command (and shouldn't be affected by
this bug).

Thanks,
Zenghui