Re: [PATCH RFC 1/3] kfence: Add a module parameter to adjust kfence objects

From: Marco Elver <elver@google.com>
To: "liupeng (DM)" <liupeng256@huawei.com>
Cc: glider@google.com, dvyukov@google.com, corbet@lwn.net,
	 sumit.semwal@linaro.org, christian.koenig@amd.com,
	akpm@linux-foundation.org,  kasan-dev@googlegroups.com,
	linux-doc@vger.kernel.org,  linux-kernel@vger.kernel.org,
	linaro-mm-sig@lists.linaro.org,  linux-mm@kvack.org
Subject: Re: [PATCH RFC 1/3] kfence: Add a module parameter to adjust kfence objects
Date: Mon, 24 Jan 2022 12:55:28 +0100	[thread overview]
Message-ID: <CANpmjNO8g_MB-5T9YxLKHOe=Mo8AWTmSFGh5jmr479s=j-v0Pg@mail.gmail.com> (raw)
In-Reply-To: <CANpmjNM_bp03RvWYr+PaOxx0DS3LryChweG90QXci3iBgzW4wQ@mail.gmail.com>

On Mon, 24 Jan 2022 at 12:45, Marco Elver <elver@google.com> wrote:
>
> [ FYI, your reply was not plain text, so LKML may have rejected it. I
> advise that you switch your email client for LKML emails to plain
> text. ]
>
> On Mon, 24 Jan 2022 at 12:24, liupeng (DM) <liupeng256@huawei.com> wrote:
> [...]
> > > I think the only reasonable way forward is if you add immediate patching
> > > support to the kernel as the "Note" suggests.
> >
> > May you give us more details about "immediate patching"?
> [...]
> > Thank you for your patient suggestions, it's actually helpful and inspired.
> > We have integrated your latest work "skipping already covered allocations",
> > and will do more experiments about KFENCE. Finally, we really hope you can
> > give us more introductions about "immediate patching".
>
> "Immediate patching" would, similar to "static branches" or
> "alternatives" be based on code hot patching.
>
> https://www.kernel.org/doc/html/latest/staging/static-keys.html
>
> "Patching immediates" would essentially patch the immediate operands
> of certain (limited) instructions. I think designing this properly to
> work across various architectures (like static_keys/jump_label) is
> very complex. So it may not be a viable near-term option.
>
> What Dmitry suggests using a constant virtual address carveout is more
> realistic. But this means having to discuss with arch maintainers
> which virtual address ranges can be reserved. The nice thing about
> just relying on memblock and nothing else is that it is very portable
> and simple. You can have a look at how KASAN deals with organizing its
> shadow memory if you are interested.

Hmm, there may be more issues lurking here:

https://lore.kernel.org/all/20200929140226.GB53442@C02TD0UTHF1T.local/
https://lore.kernel.org/all/20200929142411.GC53442@C02TD0UTHF1T.local/

... and I'm guessing if we assign a fixed virtual address range it'll
live outside the linear mapping, which is likely to break certain
requirements of kmalloc()'d allocations in certain situations (a
problem we had with v1 of KFENCE on arm64).

So I don't even know if that's feasible. :-/

Thanks,
-- Marco