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;
From:   Michael Ellerman <mpe@ellerman.id.au>
To:     Catalin Marinas <catalin.marinas@arm.com>
Cc:     Qian Cai <cai@lca.pw>, paulus@ozlabs.org, benh@kernel.crashing.org,
        kvm-ppc@vger.kernel.org, linuxppc-dev@lists.ozlabs.org,
        linux-kernel@vger.kernel.org
Subject: Re: [PATCH] powerpc/kvm: silence kmemleak false positives
In-Reply-To: <20200511112829.GB19176@gaia>
References: <20200509015538.3183-1-cai@lca.pw> <87y2pybu38.fsf@mpe.ellerman.id.au> <20200511112829.GB19176@gaia>
Date:   Wed, 13 May 2020 14:00:44 +1000
Message-ID: <87k11gbi1f.fsf@mpe.ellerman.id.au>
MIME-Version: 1.0
Content-Type: text/plain
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

Catalin Marinas <catalin.marinas@arm.com> writes:
> On Mon, May 11, 2020 at 09:15:55PM +1000, Michael Ellerman wrote:
>> Qian Cai <cai@lca.pw> writes:
>> > kvmppc_pmd_alloc() and kvmppc_pte_alloc() allocate some memory but then
>> > pud_populate() and pmd_populate() will use __pa() to reference the newly
>> > allocated memory. The same is in xive_native_provision_pages().
>> >
>> > Since kmemleak is unable to track the physical memory resulting in false
>> > positives, silence those by using kmemleak_ignore().
>> 
>> There is kmemleak_alloc_phys(), which according to the docs can be used
>> for tracking a phys address.
>
> This won't help. While kmemleak_alloc_phys() allows passing a physical
> address, it doesn't track physical address references to this object. It
> still expects VA pointing to it, otherwise the object would be reported
> as a leak.

OK, thanks for clarifying that.

> We currently only call this from the memblock code with a min_count of
> 0, meaning it will not be reported as a leak if no references are found.
>
> We don't have this issue with page tables on other architectures since
> most of them use whole page allocations which aren't tracked by
> kmemleak. These powerpc functions use kmem_cache_alloc() which would be
> tracked automatically by kmemleak. While we could add a phys alias to
> kmemleak (another search tree), I think the easiest is as per Qian's
> patch, just ignore those objects.

Agreed.

cheers