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:   Wed, 16 Jun 2021 11:36:39 -0700
From:   Rustam Kovhaev <rkovhaev@gmail.com>
To:     Catalin Marinas <catalin.marinas@arm.com>
Cc:     Andrew Morton <akpm@linux-foundation.org>, linux-mm@kvack.org,
        linux-kernel@vger.kernel.org, dvyukov@google.com,
        gregkh@linuxfoundation.org
Subject: Re: kmemleak memory scanning
Message-ID: <YMpEtyUP0vuWo9HV@nuc10>
References: <YMe8ktUsdtwFKHuF@nuc10>
 <20210615101515.GC26027@arm.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20210615101515.GC26027@arm.com>
Precedence: bulk

hi Catalin,

On Tue, Jun 15, 2021 at 11:15:15AM +0100, Catalin Marinas wrote:
> Hi Rustam,
> 
> On Mon, Jun 14, 2021 at 01:31:14PM -0700, Rustam Kovhaev wrote:
> > a) kmemleak scans struct page (kmemleak.c:1462), but it does not scan
> > the actual contents (page_address(page)) of the page.
> > if we allocate an object with kmalloc(), then allocate page with
> > alloc_page(), and if we put kmalloc pointer somewhere inside that page, 
> > kmemleak will report kmalloc pointer as a false positive.
> > should we improve kmemleak and make it scan page contents?
> > or will this bring too many false negatives?
> 
> This is indeed on purpose otherwise (1) we'd get a lot of false
> negatives and (2) the scanning would take significantly longer. There
> are a lot more pages allocated for user processes than used in the
> kernel, we don't need to scan them all.
> 
> We do have a few places where we explicitly call kmemleak_alloc():
> neigh_hash_alloc(), alloc_page_ext(), blk_mq_alloc_rqs(),
> early_amd_iommu_init().

makes sense, tyvm!

> > b) when kmemleak object gets created (kmemleak.c:598) it gets checksum
> > of 0, by the time user requests kmemleak "scan" via debugfs the pointer
> > will be most likely changed to some value by the kernel and during
> > first scan kmemleak won't report the object as orphan even if it did not
> > find any reference to it, because it will execute update_checksum() and
> > after that will proceed to updating object->count (kmemleak.c:1502).
> > and so the user will have to initiate a second "scan" via debugfs and
> > only then kmemleak will produce the report.
> > should we document this?
> 
> That's a mitigation against false positives. Lot's of objects that get
> allocated just prior to a memory scan have a tendency to be reported as
> leaks before they get referenced via e.g. a list (and the in-object
> list_head structure updated). So you'd need to get the checksum
> identical in two consecutive scans to report it as a leak.
> 
> We should probably document this.

thanks, i'll send a documentation patch for this