Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Tue, 4 Jul 2023 08:01:06 +0200
From:   Andrew Jones <ajones@ventanamicro.com>
To:     John Hubbard <jhubbard@nvidia.com>
Cc:     Andrew Morton <akpm@linux-foundation.org>,
        Albert Ou <aou@eecs.berkeley.edu>,
        Alexandre Ghiti <alexghiti@rivosinc.com>,
        Hugh Dickins <hughd@google.com>,
        Palmer Dabbelt <palmer@dabbelt.com>,
        Paul Walmsley <paul.walmsley@sifive.com>,
        Qinglin Pan <panqinglin2020@iscas.ac.cn>,
        linux-riscv@lists.infradead.org, linux-mm@kvack.org,
        LKML <linux-kernel@vger.kernel.org>,
        James Houghton <jthoughton@google.com>,
        Ryan Roberts <ryan.roberts@arm.com>
Subject: Re: [PATCH] mm: riscv: fix an unsafe pte read in huge_pte_alloc()
Message-ID: <20230704-f273e5ba6c440dff03d07101@orel>
References: <20230703190044.311730-1-jhubbard@nvidia.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20230703190044.311730-1-jhubbard@nvidia.com>
Precedence: bulk

On Mon, Jul 03, 2023 at 12:00:44PM -0700, John Hubbard wrote:
> The WARN_ON_ONCE() statement in riscv's huge_pte_alloc() is susceptible
> to false positives, because the pte is read twice at the C language
> level, locklessly, within the same conditional statement. Depending on
> compiler behavior, this can lead to generated machine code that actually
> reads the pte just once, or twice. Reading twice will expose the code to
> changing pte values and cause incorrect behavior.
> 
> In [1], similar code actually caused a kernel crash on 64-bit x86, when
> using clang to build the kernel, but only after the conversion from *pte
> reads, to ptep_get(pte). The latter uses READ_ONCE(), which forced a
> double read of *pte.
> 
> Rather than waiting for the upcoming ptep_get() conversion, just convert
> this part of the code now, but in a way that avoids the above problem:
> take a single snapshot of the pte before using it in the WARN
> conditional.
> 
> As expected, this preparatory step does not actually change the
> generated code ("make mm/hugetlbpage.s"), on riscv64, when using a gcc
> 12.2 cross compiler.
> 
> [1] https://lore.kernel.org/20230630013203.1955064-1-jhubbard@nvidia.com
> 
> Suggested-by: James Houghton <jthoughton@google.com>
> Cc: Ryan Roberts <ryan.roberts@arm.com>
> Signed-off-by: John Hubbard <jhubbard@nvidia.com>
> ---
>  arch/riscv/mm/hugetlbpage.c | 6 +++++-
>  1 file changed, 5 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/riscv/mm/hugetlbpage.c b/arch/riscv/mm/hugetlbpage.c
> index 542883b3b49b..96225a8533ad 100644
> --- a/arch/riscv/mm/hugetlbpage.c
> +++ b/arch/riscv/mm/hugetlbpage.c
> @@ -73,7 +73,11 @@ pte_t *huge_pte_alloc(struct mm_struct *mm,
>  	}
>  
>  out:
> -	WARN_ON_ONCE(pte && pte_present(*pte) && !pte_huge(*pte));
> +	if (pte) {
> +		pte_t pteval = ptep_get_lockless(pte);

I think ptep_get_lockless() on riscv (even riscv32) will always just be
ptep_get(), since pte_t is unsigned long, which can be read atomically.

> +
> +		WARN_ON_ONCE(pte_present(pteval) && !pte_huge(pteval));

Ensuring we only read the pte once is good though.

Reviewed-by: Andrew Jones <ajones@ventanamicro.com>

Thanks,
drew


> +	}
>  	return pte;
>  }
>  
> 
> base-commit: 0a8d6c9c7128a93689fba384cdd7f72b0ce19abd
> -- 
> 2.41.0
>