Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
From:   "liwei (GF)" <liwei391@huawei.com>
Subject: Re: [RFC PATCH] irqchip/gic-v3: Make gic_handle_irq() notrace
To:     Julien Thierry <julien.thierry@arm.com>,
        Zenghui Yu <yuzenghui@huawei.com>,
        Marc Zyngier <marc.zyngier@arm.com>, <rostedt@goodmis.org>
CC:     <mark.rutland@arm.com>, <suzuki.poulose@arm.com>,
        <catalin.marinas@arm.com>, <will.deacon@arm.com>,
        <linux@armlinux.org.uk>, <linux-kernel@vger.kernel.org>,
        <mingo@redhat.com>, <james.morse@arm.com>,
        <wanghaibin.wang@huawei.com>,
        <linux-arm-kernel@lists.infradead.org>, <guohanjun@huawei.com>,
        "huawei.libin@huawei.com >> libin 00196512" <huawei.libin@huawei.com>
References: <1553865808-30604-1-git-send-email-yuzenghui@huawei.com>
 <237b00d0-8a8d-ede6-91b1-c3e23e61925c@arm.com>
 <bf434b0c-2b17-b92d-3171-a57203486c84@huawei.com>
 <aad77b33-7eb0-57a6-2d60-1cb717c1e1c7@arm.com>
Message-ID: <73ec5893-ae0e-64a4-a914-18a09059a5ce@huawei.com>
Date:   Wed, 3 Apr 2019 23:23:29 +0800
User-Agent: Mozilla/5.0 (Windows NT 6.1; Win64; x64; rv:60.0) Gecko/20100101
 Thunderbird/60.4.0
MIME-Version: 1.0
In-Reply-To: <aad77b33-7eb0-57a6-2d60-1cb717c1e1c7@arm.com>
Content-Type: text/plain; charset="utf-8"
Content-Transfer-Encoding: 8bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

Hi Julien,

On 2019/4/2 22:00, Julien Thierry wrote:
I meet this issue by coincidence before too.

> I finally found out what happens.
> 
> When using interrupt priority masking, at the begining of
> gic_handle_irq(), we are in this awkward state where we still have the I
> bit set and PMR unmasked (this is because we cannot ack the interrupt
> that was signaled if it has lower priority than the current priority mask).
> 
> To try and keep things simple, we decided that local_irq_*() would only
> deal with PMR (when using priority masking). With one additional case
> being that, if PSR.I is set when saving flags, we'd represent it in the
> form of a value of PMR (i.e. GIC_PRIO_IRQOFF), so that irqs_disabled()
> and such would still accurately state that no interrupt could happen in
> those sections. The assumption was that in the few sections were we are
> having the PSR.I set, we wouldn't care about having interrupts disabled
> with PSR.I or PMR. And now that assumption appears to be wrong:
I am confused of the logic of the save and restore here, why can't we do the
save & restore exactly?

> trace_graph_entry(), called at the begining of gic_handle_irq() when
> enabling the tracer, does use local_irq_save/restore(). The save
> operation returns flags GIC_PRIO_IRQOFF (because PSR.I is set when we
> enter gic_handle_irq() ). The restore operation will then proceed to
> mask PMR, once we get back to gic_handle_irq() we can't acknowledge the
> interrupt we just received...
> 
> The function tracer does not appear to save/restore flags on function
> entry (I saw save/restore operations in the stack tracer but for some
> reason couldn't get them to trigger the issue).
> 
> To confirm this, I checked with the following diff (which is not a fix,
> it is better to mark gic_handle_irq() as notrace if I don't find
> something more suitable).
> 
I am convinced a notrace marking here will walkaround this issue，but i am
afraid there is still some trouble like this else.
I send my modification to solve this issue just now, it is tested ok so far,
but i am not particularly sure, could you please have a look?

Thanks,
Wei