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:   Tue, 5 May 2020 15:15:29 +0100
From:   Daniel Thompson <daniel.thompson@linaro.org>
To:     Will Deacon <will@kernel.org>
Cc:     Catalin Marinas <catalin.marinas@arm.com>,
        Douglas Anderson <dianders@chromium.org>,
        Jason Wessel <jason.wessel@windriver.com>,
        linux-arm-kernel@lists.infradead.org, linux-kernel@vger.kernel.org,
        patches@linaro.org
Subject: Re: [PATCH v2] arm64: cacheflush: Fix KGDB trap detection
Message-ID: <20200505141529.2qb35fbqudtac3ys@holly.lan>
References: <20200504170518.2959478-1-daniel.thompson@linaro.org>
 <20200504204803.GA5657@willie-the-truck>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20200504204803.GA5657@willie-the-truck>
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Mon, May 04, 2020 at 09:48:04PM +0100, Will Deacon wrote:
> On Mon, May 04, 2020 at 06:05:18PM +0100, Daniel Thompson wrote:
> > flush_icache_range() contains a bodge to avoid issuing IPIs when the kgdb
> > trap handler is running because issuing IPIs is unsafe (and not needed)
> > in this execution context. However the current test, based on
> > kgdb_connected is flawed: it both over-matches and under-matches.
> > 
> > The over match occurs because kgdb_connected is set when gdb attaches
> > to the stub and remains set during normal running. This is relatively
> > harmelss because in almost all cases irq_disabled() will be false.
> > 
> > The under match is more serious. When kdb is used instead of kgdb to access
> > the debugger then kgdb_connected is not set in all the places that the
> > debug core updates sw breakpoints (and hence flushes the icache). This
> > can lead to deadlock.
> > 
> > Fix by replacing the ad-hoc check with the proper kgdb macro. This also
> > allows us to drop the #ifdef wrapper.
> > 
> > Fixes: 3b8c9f1cdfc5 ("arm64: IPI each CPU after invalidating the I-cache for kernel mappings")
> > Signed-off-by: Daniel Thompson <daniel.thompson@linaro.org>
> > Reviewed-by: Douglas Anderson <dianders@chromium.org>
> > ---
> > 
> > Notes:
> >     v2: Improve the commit message based based on feedback from Doug
> >         Anderson
> > 
> >  arch/arm64/include/asm/cacheflush.h | 6 +++---
> >  1 file changed, 3 insertions(+), 3 deletions(-)
> > 
> > diff --git a/arch/arm64/include/asm/cacheflush.h b/arch/arm64/include/asm/cacheflush.h
> > index e6cca3d4acf7..ce50c1f1f1ea 100644
> > --- a/arch/arm64/include/asm/cacheflush.h
> > +++ b/arch/arm64/include/asm/cacheflush.h
> > @@ -79,7 +79,7 @@ static inline void flush_icache_range(unsigned long start, unsigned long end)
> >  	 * IPI all online CPUs so that they undergo a context synchronization
> >  	 * event and are forced to refetch the new instructions.
> >  	 */
> > -#ifdef CONFIG_KGDB
> > +
> >  	/*
> >  	 * KGDB performs cache maintenance with interrupts disabled, so we
> >  	 * will deadlock trying to IPI the secondary CPUs. In theory, we can
> > @@ -89,9 +89,9 @@ static inline void flush_icache_range(unsigned long start, unsigned long end)
> >  	 * the patching operation, so we don't need extra IPIs here anyway.
> >  	 * In which case, add a KGDB-specific bodge and return early.
> >  	 */
> > -	if (kgdb_connected && irqs_disabled())
> > +	if (in_dbg_master())
> 
> Does this imply that irqs are disabled?

Yes.

Assuming CONFIG_KGDB is enabled then in_dbg_master() expands to:

    (raw_smp_processor_id() == atomic_read(&kgdb_active))

kgdb_active is written to from exactly four locations in the kernel and
all are within a single function, albeit a very big function with control
flow the that could politely be called "quirky". I try not to think about
what it might be impolitely called.

kgdb_active is only ever set to a value other than -1 when we are
executing the kgdb exception handler and interrupts have been
explicitly disabled.


Daniel.