Re: [RFC PATCH -next V2 5/7] arm64: add get_user to machine check safe

From: Mark Rutland <mark.rutland@arm.com>
To: Tong Tiangen <tongtiangen@huawei.com>
Cc: Andrew Morton <akpm@linux-foundation.org>,
	Thomas Gleixner <tglx@linutronix.de>,
	Ingo Molnar <mingo@redhat.com>, Borislav Petkov <bp@alien8.de>,
	Dave Hansen <dave.hansen@linux.intel.com>,
	Catalin Marinas <catalin.marinas@arm.com>,
	Will Deacon <will@kernel.org>,
	Alexander Viro <viro@zeniv.linux.org.uk>,
	x86@kernel.org, "H. Peter Anvin" <hpa@zytor.com>,
	linux-arm-kernel@lists.infradead.org,
	linux-kernel@vger.kernel.org, linux-mm@kvack.org
Subject: Re: [RFC PATCH -next V2 5/7] arm64: add get_user to machine check safe
Date: Fri, 8 Apr 2022 16:22:22 +0100	[thread overview]
Message-ID: <YlBTLn1cf7+gJmiM@lakrids> (raw)
In-Reply-To: <ac80045b-24f0-a163-ea96-be0d01c48d29@huawei.com>

On Thu, Apr 07, 2022 at 10:38:04PM +0800, Tong Tiangen wrote:
> 在 2022/4/6 19:22, Mark Rutland 写道:
> > On Wed, Apr 06, 2022 at 09:13:09AM +0000, Tong Tiangen wrote:
> > > Add scenarios get_user to machine check safe. The processing of
> > > EX_TYPE_UACCESS_ERR_ZERO and EX_TYPE_UACCESS_ERR_ZERO_UCE_RECOVERY is same
> > > and both return -EFAULT.
> > 
> > Which uaccess cases do we expect to *not* be recoverable?
> > 
> > Naively I would assume that if we're going to treat a memory error on a uaccess
> > as fatal to userspace we should be able to do that for *any* uacesses.
> > 
> > The commit message should explain why we need the distinction between a
> > recoverable uaccess and a non-recoverable uaccess.
> > 
> > Thanks,
> > Mark.
> 
> Currently, any memory error consumed in kernel mode will lead to panic
> (do_sea()).
> 
> My idea is that not all memory errors consumed in kernel mode are fatal,
> such as copy_ from_ user/get_ user is a memory error consumed when
> reading user data in the process context. In this case, we can not let the
> kernel panic, just kill the process without affecting the operation
> of the system.

I understood this part.

> However, not all uaccess can be recovered without affecting the normal
> operation of the system. The key is not whether it is uaccess, but whether
> there are key data affecting the normal operation of the system in the read
> page.

Ok. Can you give an example of such a case where the a uaccess that hits
a memory error must be fatal?

I think you might be trying to say that for copy_{to,from}_user() we can
make that judgement, but those are combined user+kernel access
primitives, and the *uaccess* part should never be reading from a page
with "key data affecting the normal operation of the system", since
that's userspace memory.

Is there any *userspace access* (e.g. where we use LDTR/STTR today)
where we must treat a memory error as fatal to the system?

Thanks,
Mark.