1 From patchwork Thu Apr 26 23:28:34 2018
2 Content-Type: text/plain; charset="utf-8"
4 Content-Transfer-Encoding: 7bit
5 Subject: [v2] MIPS: c-r4k: fix data corruption related to cache coherence.
6 X-Patchwork-Submitter: NeilBrown <neil@brown.name>
8 Message-Id: <87vacdlf8d.fsf@notabene.neil.brown.name>
9 To: James Hogan <jhogan@kernel.org>
10 Cc: Ralf Baechle <ralf@linux-mips.org>,
11 Paul Burton <paul.burton@mips.com>, linux-mips@linux-mips.org,
12 linux-kernel@vger.kernel.org
13 Date: Fri, 27 Apr 2018 09:28:34 +1000
14 From: NeilBrown <neil@brown.name>
15 List-Id: linux-mips <linux-mips.eddie.linux-mips.org>
17 When DMA will be performed to a MIPS32 1004K CPS, the
18 L1-cache for the range needs to be flushed and invalidated
20 The code currently takes one of two approaches.
21 1/ If the range is less than the size of the dcache, then
22 HIT type requests flush/invalidate cache lines for the
23 particular addresses. HIT-type requests a globalised
24 by the CPS so this is safe on SMP.
26 2/ If the range is larger than the size of dcache, then
27 INDEX type requests flush/invalidate the whole cache.
28 INDEX type requests affect the local cache only. CPS
29 does not propagate them in any way. So this invalidation
30 is not safe on SMP CPS systems.
32 Data corruption due to '2' can quite easily be demonstrated by
33 repeatedly "echo 3 > /proc/sys/vm/drop_caches" and then sha1sum
34 a file that is several times the size of available memory.
35 Dropping caches means that large contiguous extents (large than
36 dcache) are more likely.
38 This was not a problem before Linux-4.8 because option 2 was
39 never used if CONFIG_MIPS_CPS was defined. The commit
40 which removed that apparently didn't appreciate the full
41 consequence of the change.
43 We could, in theory, globalize the INDEX based flush by sending an IPI
44 to other cores. These cache invalidation routines can be called with
45 interrupts disabled and synchronous IPI require interrupts to be
46 enabled. Asynchronous IPI may not trigger writeback soon enough.
47 So we cannot use IPI in practice.
49 We can already test is IPI would be needed for an INDEX operation
50 with r4k_op_needs_ipi(R4K_INDEX). If this is True then we mustn't try
51 the INDEX approach as we cannot use IPI. If this is False (e.g. when
52 there is only one core and hence one L1 cache) then it is safe to
53 use the INDEX approach without IPI.
55 This patch avoids options 2 if r4k_op_needs_ipi(R4K_INDEX), and so
56 eliminates the corruption.
58 Fixes: c00ab4896ed5 ("MIPS: Remove cpu_has_safe_index_cacheops")
59 Cc: stable@vger.kernel.org # v4.8+
60 Signed-off-by: NeilBrown <neil@brown.name>
62 arch/mips/mm/c-r4k.c | 9 ++++++---
63 1 file changed, 6 insertions(+), 3 deletions(-)
65 --- a/arch/mips/mm/c-r4k.c
66 +++ b/arch/mips/mm/c-r4k.c
67 @@ -851,9 +851,12 @@ static void r4k_dma_cache_wback_inv(unsi
69 * Either no secondary cache or the available caches don't have the
70 * subset property so we have to flush the primary caches
73 + * If we would need IPI to perform an INDEX-type operation, then
74 + * we have to use the HIT-type alternative as IPI cannot be used
75 + * here due to interrupts possibly being disabled.
77 - if (size >= dcache_size) {
78 + if (!r4k_op_needs_ipi(R4K_INDEX) && size >= dcache_size) {
81 R4600_HIT_CACHEOP_WAR_IMPL;
82 @@ -890,7 +893,7 @@ static void r4k_dma_cache_inv(unsigned l
86 - if (size >= dcache_size) {
87 + if (!r4k_op_needs_ipi(R4K_INDEX) && size >= dcache_size) {
90 R4600_HIT_CACHEOP_WAR_IMPL;