Re: [PATCH] pstore: ramoops: support pmsg size larger than kmalloc limitation

From: "Guilherme G. Piccoli" <gpiccoli@igalia.com>
To: Kees Cook <keescook@chromium.org>
Cc: Yuxiao Zhang <yuxiaozhang@google.com>,
	Greg KH <gregkh@linuxfoundation.org>,
	Tony Luck <tony.luck@intel.com>,
	linux-hardening@vger.kernel.org, linux-kernel@vger.kernel.org,
	wak@google.com
Subject: Re: [PATCH] pstore: ramoops: support pmsg size larger than kmalloc limitation
Date: Thu, 29 Jun 2023 16:22:40 -0300	[thread overview]
Message-ID: <80bf6cef-7824-124f-fe69-d820ae3416f3@igalia.com> (raw)
In-Reply-To: <202306281615.360B12EFC@keescook>

On 28/06/2023 20:24, Kees Cook wrote:
> On Wed, Jun 28, 2023 at 03:12:10PM -0300, Guilherme G. Piccoli wrote:
>> Also - Kees certainly knows that way better, but - are we 100% sure that
>> the region for pstore can be non-contiguous? For some reason, I always
>> thought this was a requirement - grepping the code, I found this
>> (wrong?) comment:
>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/fs/pstore/zone.c#n3
> 
> The contiguous requirements are entirely for the RAM backend's storage,
> so these intermediate memory regions can use non-contiguous physical
> backing memory (i.e. vmalloc).
> 
> Even the special case of crash-dumping should be fine for the large
> buffer used to hold the crash before doing compression.
> 
> There are effectively 4 types of allocations in pstore:
> 
> 1- a physical -> virtual mapping for the RAM backend
> 2- the allocations (if any) for non-RAM backends to hold a copy of pstore
>    records when extracted from the backend storage (e.g NVRAM, EFI vars,
>    etc).
> 3- incoming allocations that are used temporarily to hand data to the
>    backend (e.g. the write_compat used in this patch)
> 4- the allocation used for holding the pstorefs data contents (which I
>    need to double-check, but is entirely defined by the backends)
> 
> Changes aren't needed for (1), it's fine on its own. This patch changes
> allocations for (2) and (3) for pmsg and the RAM backend, if I'm reading
> it correctly. I think (4) is included as part of (2), but I need to
> check. As long as all paths use kvfree() for the record buffers,
> everything should Just Work for RAM. Switching the other backends to
> also use kvalloc() would allow for greater flexibility, though.
> 
> Anyway, it's on my list to review and test. I'm still working through
> other things related to the merge window opening, so I may be a bit slow
> for the next week. :)
> 
> -Kees
> 

Thanks a bunch for the clarification Kees, much appreciated!
Now I understand it way better =)