Date:   Thu, 2 Dec 2021 13:03:41 -0400
From:   Jason Gunthorpe <jgg@ziepe.ca>
To:     Kees Cook <keescook@chromium.org>
Cc:     Bixuan Cui <cuibixuan@linux.alibaba.com>, linux-mm@kvack.org,
        linux-kernel@vger.kernel.org, torvalds@linux-foundation.org,
        leon@kernel.org, akpm@linux-foundation.org, w@1wt.eu
Subject: Re: [PATCH -next] mm: delete oversized WARN_ON() in kvmalloc() calls
Message-ID: <20211202170341.GO5112@ziepe.ca>
References: <1638410784-48646-1-git-send-email-cuibixuan@linux.alibaba.com>
 <202112011944.28EF2FC44@keescook>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <202112011944.28EF2FC44@keescook>
Precedence: bulk

On Wed, Dec 01, 2021 at 07:46:01PM -0800, Kees Cook wrote:

> If we're rejecting the value, then it's still a pathological size, so
> shouldn't the check be happening in the caller? I think the WARN is
> doing exactly what it was supposed to do: find the places where bad
> sizes can reach vmalloc.

I think it meshes very poorly with the overflow work:

  p = kzalloc(struct_size(p, regions, num_regions), GFP_KERNEL);

If num_regions is user controlled data why should the calling driver
hvae to somehow sanitize num_regions (without bugs!) instead of
relying on struct_size() and kzalloc() to contain all the sanitation?

What you are suggesting just pushes security sensitive coding into
drivers, which I think is the opposite of what we all want?

Jason