aboutsummaryrefslogtreecommitdiffstats
path: root/target/linux/generic/pending-4.14/103-MIPS-c-r4k-fix-data-corruption-related-to-cache-coherence.patch
blob: 69d926bd5bf078d2cdbdf1b32e4a9289b167ff8c (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
From patchwork Thu Apr 26 23:28:34 2018
Content-Type: text/plain; charset="utf-8"
MIME-Version: 1.0
Content-Transfer-Encoding: 7bit
Subject: [v2] MIPS: c-r4k: fix data corruption related to cache coherence.
X-Patchwork-Submitter: NeilBrown <neil@brown.name>
X-Patchwork-Id: 19259
Message-Id: <87vacdlf8d.fsf@notabene.neil.brown.name>
To: James Hogan <jhogan@kernel.org>
Cc: Ralf Baechle <ralf@linux-mips.org>,
 Paul Burton <paul.burton@mips.com>, linux-mips@linux-mips.org,
 linux-kernel@vger.kernel.org
Date: Fri, 27 Apr 2018 09:28:34 +1000
From: NeilBrown <neil@brown.name>
List-Id: linux-mips <linux-mips.eddie.linux-mips.org>

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@vger.kernel.org # v4.8+
Signed-off-by: NeilBrown <neil@brown.name>
---
 arch/mips/mm/c-r4k.c | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

--- 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(unsi
 	/*
 	 * 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 l
 		return;
 	}
 
-	if (size >= dcache_size) {
+	if (!r4k_op_needs_ipi(R4K_INDEX) && size >= dcache_size) {
 		r4k_blast_dcache();
 	} else {
 		R4600_HIT_CACHEOP_WAR_IMPL;