Re: [PATCH v2] MIPS: c-r4k: fix data corruption related to cache coherence.

[NeilBrown]

Hi James,
 this hasn't appear in linux-next yet, or in any branch
 of
   git://git.kernel.org/pub/scm/linux/kernel/git/jhogan/mips.git

 Should I expect it to?

Thanks,
NeilBrown

On Fri, Apr 27 2018, NeilBrown wrote:

> When DMA will be performed to a MIPS32 1004K CPS, the
> L1-cache for the range needs to be flushed and invalidated
> first.
> The code currently takes one of two approaches.
> 1/ If the range is less than the size of the dcache, then
>    HIT type requests flush/invalidate cache lines for the
>    particular addresses.  HIT-type requests a globalised
>    by the CPS so this is safe on SMP.
>
> 2/ If the range is larger than the size of dcache, then
>    INDEX type requests flush/invalidate the whole cache.
>    INDEX type requests affect the local cache only. CPS
>    does not propagate them in any way.  So this invalidation
>    is not safe on SMP CPS systems.
>
> Data corruption due to '2' can quite easily be demonstrated by
> repeatedly "echo 3 > /proc/sys/vm/drop_caches" and then sha1sum
> a file that is several times the size of available memory.
> Dropping caches means that large contiguous extents (large than
> dcache) are more likely.
>
> This was not a problem before Linux-4.8 because option 2 was
> never used if CONFIG_MIPS_CPS was defined.  The commit
> which removed that apparently didn't appreciate the full
> consequence of the change.
>
> We could, in theory, globalize the INDEX based flush by sending an IPI
> to other cores.  These cache invalidation routines can be called with
> interrupts disabled and synchronous IPI require interrupts to be
> enabled.  Asynchronous IPI may not trigger writeback soon enough.
> So we cannot use IPI in practice.
>
> We can already test is IPI would be needed for an INDEX operation
> with r4k_op_needs_ipi(R4K_INDEX).  If this is True then we mustn't try
> the INDEX approach as we cannot use IPI.  If this is False (e.g. when
> there is only one core and hence one L1 cache) then it is safe to
> use the INDEX approach without IPI.
>
> This patch avoids options 2 if r4k_op_needs_ipi(R4K_INDEX), and so
> eliminates the corruption.
>
> Fixes: c00ab4896ed5 ("MIPS: Remove cpu_has_safe_index_cacheops")
> Cc: stable@xxxxxxxxxxxxxxx # v4.8+
> Signed-off-by: NeilBrown <neil@xxxxxxxxxx>
> ---
>  arch/mips/mm/c-r4k.c | 9 ++++++---
>  1 file changed, 6 insertions(+), 3 deletions(-)
>
> diff --git a/arch/mips/mm/c-r4k.c b/arch/mips/mm/c-r4k.c
> index 6f534b209971..e12dfa48b478 100644
> --- a/arch/mips/mm/c-r4k.c
> +++ b/arch/mips/mm/c-r4k.c
> @@ -851,9 +851,12 @@ static void r4k_dma_cache_wback_inv(unsigned long addr, unsigned long size)
>  	/*
>  	 * Either no secondary cache or the available caches don't have the
>  	 * subset property so we have to flush the primary caches
> -	 * explicitly
> +	 * explicitly.
> +	 * If we would need IPI to perform an INDEX-type operation, then
> +	 * we have to use the HIT-type alternative as IPI cannot be used
> +	 * here due to interrupts possibly being disabled.
>  	 */
> -	if (size >= dcache_size) {
> +	if (!r4k_op_needs_ipi(R4K_INDEX) && size >= dcache_size) {
>  		r4k_blast_dcache();
>  	} else {
>  		R4600_HIT_CACHEOP_WAR_IMPL;
> @@ -890,7 +893,7 @@ static void r4k_dma_cache_inv(unsigned long addr, unsigned long size)
>  		return;
>  	}
>  
> -	if (size >= dcache_size) {
> +	if (!r4k_op_needs_ipi(R4K_INDEX) && size >= dcache_size) {
>  		r4k_blast_dcache();
>  	} else {
>  		R4600_HIT_CACHEOP_WAR_IMPL;
> -- 
> 2.14.0.rc0.dirty

Attachment: signature.asc
Description: PGP signature