crypto: caam - fix S/G table passing page boundary

According to CAAM RM:
-crypto engine reads 4 S/G entries (64 bytes) at a time,
even if the S/G table has fewer entries
-it's the responsibility of the user / programmer to make sure
this HW behaviour has no side effect

The drivers do not take care of this currently, leading to IOMMU faults
when the S/G table ends close to a page boundary - since only one page
is DMA mapped, while CAAM's DMA engine accesses two pages.

Fix this by rounding up the number of allocated S/G table entries
to a multiple of 4.
Note that in case of two *contiguous* S/G tables, only the last table
might needs extra entries.

Signed-off-by: Horia Geantă <horia.geanta@nxp.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

1 files changed, 36 insertions, 4 deletions

diff --git a/drivers/crypto/caam/caamalg_qi.c b/drivers/crypto/caam/caamalg_qi.c
index 116cbc81fa8d..4ccaa4b7936c 100644
--- a/drivers/crypto/caam/caamalg_qi.c
+++ b/drivers/crypto/caam/caamalg_qi.c
@@ -4,7 +4,7 @@
  * Based on caamalg.c
  *
  * Copyright 2013-2016 Freescale Semiconductor, Inc.
- * Copyright 2016-2018 NXP
+ * Copyright 2016-2019 NXP
  */
 
 #include "compat.h"
@@ -1019,9 +1019,24 @@ static struct aead_edesc *aead_edesc_alloc(struct aead_request *req,
 	/*
 	 * Create S/G table: req->assoclen, [IV,] req->src [, req->dst].
 	 * Input is not contiguous.
+	 * HW reads 4 S/G entries at a time; make sure the reads don't go beyond
+	 * the end of the table by allocating more S/G entries. Logic:
+	 * if (src != dst && output S/G)
+	 *      pad output S/G, if needed
+	 * else if (src == dst && S/G)
+	 *      overlapping S/Gs; pad one of them
+	 * else if (input S/G) ...
+	 *      pad input S/G, if needed
 	 */
-	qm_sg_ents = 1 + !!ivsize + mapped_src_nents +
-		     (mapped_dst_nents > 1 ? mapped_dst_nents : 0);
+	qm_sg_ents = 1 + !!ivsize + mapped_src_nents;
+	if (mapped_dst_nents > 1)
+		qm_sg_ents += pad_sg_nents(mapped_dst_nents);
+	else if ((req->src == req->dst) && (mapped_src_nents > 1))
+		qm_sg_ents = max(pad_sg_nents(qm_sg_ents),
+				 1 + !!ivsize + pad_sg_nents(mapped_src_nents));
+	else
+		qm_sg_ents = pad_sg_nents(qm_sg_ents);
+
 	sg_table = &edesc->sgt[0];
 	qm_sg_bytes = qm_sg_ents * sizeof(*sg_table);
 	if (unlikely(offsetof(struct aead_edesc, sgt) + qm_sg_bytes + ivsize >
@@ -1276,7 +1291,24 @@ static struct skcipher_edesc *skcipher_edesc_alloc(struct skcipher_request *req,
 	qm_sg_ents = 1 + mapped_src_nents;
 	dst_sg_idx = qm_sg_ents;
 
-	qm_sg_ents += mapped_dst_nents > 1 ? mapped_dst_nents : 0;
+	/*
+	 * HW reads 4 S/G entries at a time; make sure the reads don't go beyond
+	 * the end of the table by allocating more S/G entries. Logic:
+	 * if (src != dst && output S/G)
+	 *      pad output S/G, if needed
+	 * else if (src == dst && S/G)
+	 *      overlapping S/Gs; pad one of them
+	 * else if (input S/G) ...
+	 *      pad input S/G, if needed
+	 */
+	if (mapped_dst_nents > 1)
+		qm_sg_ents += pad_sg_nents(mapped_dst_nents);
+	else if ((req->src == req->dst) && (mapped_src_nents > 1))
+		qm_sg_ents = max(pad_sg_nents(qm_sg_ents),
+				 1 + pad_sg_nents(mapped_src_nents));
+	else
+		qm_sg_ents = pad_sg_nents(qm_sg_ents);
+
 	qm_sg_bytes = qm_sg_ents * sizeof(struct qm_sg_entry);
 	if (unlikely(offsetof(struct skcipher_edesc, sgt) + qm_sg_bytes +
 		     ivsize > CAAM_QI_MEMCACHE_SIZE)) {