1 files changed, 144 insertions, 19 deletions

diff --git a/module/zfs/zio_checksum.c b/module/zfs/zio_checksum.c
index b05e787dc..59871c50e 100644
--- a/module/zfs/zio_checksum.c
+++ b/module/zfs/zio_checksum.c
@@ -21,10 +21,12 @@
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  * Copyright (c) 2013 by Delphix. All rights reserved.
+ * Copyright 2013 Saso Kiselkov. All rights reserved.
  */
 
 #include <sys/zfs_context.h>
 #include <sys/spa.h>
+#include <sys/spa_impl.h>
 #include <sys/zio.h>
 #include <sys/zio_checksum.h>
 #include <sys/zil.h>
@@ -58,28 +60,96 @@
  * checksum function of the appropriate strength.  When reading a block,
  * we compare the expected checksum against the actual checksum, which we
  * compute via the checksum function specified by BP_GET_CHECKSUM(bp).
+ *
+ * SALTED CHECKSUMS
+ *
+ * To enable the use of less secure hash algorithms with dedup, we
+ * introduce the notion of salted checksums (MACs, really).  A salted
+ * checksum is fed both a random 256-bit value (the salt) and the data
+ * to be checksummed.  This salt is kept secret (stored on the pool, but
+ * never shown to the user).  Thus even if an attacker knew of collision
+ * weaknesses in the hash algorithm, they won't be able to mount a known
+ * plaintext attack on the DDT, since the actual hash value cannot be
+ * known ahead of time.  How the salt is used is algorithm-specific
+ * (some might simply prefix it to the data block, others might need to
+ * utilize a full-blown HMAC).  On disk the salt is stored in a ZAP
+ * object in the MOS (DMU_POOL_CHECKSUM_SALT).
+ *
+ * CONTEXT TEMPLATES
+ *
+ * Some hashing algorithms need to perform a substantial amount of
+ * initialization work (e.g. salted checksums above may need to pre-hash
+ * the salt) before being able to process data.  Performing this
+ * redundant work for each block would be wasteful, so we instead allow
+ * a checksum algorithm to do the work once (the first time it's used)
+ * and then keep this pre-initialized context as a template inside the
+ * spa_t (spa_cksum_tmpls).  If the zio_checksum_info_t contains
+ * non-NULL ci_tmpl_init and ci_tmpl_free callbacks, they are used to
+ * construct and destruct the pre-initialized checksum context.  The
+ * pre-initialized context is then reused during each checksum
+ * invocation and passed to the checksum function.
  */
 
 /*ARGSUSED*/
 static void
-zio_checksum_off(const void *buf, uint64_t size, zio_cksum_t *zcp)
+zio_checksum_off(const void *buf, uint64_t size,
+    const void *ctx_template, zio_cksum_t *zcp)
 {
 	ZIO_SET_CHECKSUM(zcp, 0, 0, 0, 0);
 }
 
 zio_checksum_info_t zio_checksum_table[ZIO_CHECKSUM_FUNCTIONS] = {
-	{{NULL,			NULL},			0, 0, 0, "inherit"},
-	{{NULL,			NULL},			0, 0, 0, "on"},
-	{{zio_checksum_off,	zio_checksum_off},	0, 0, 0, "off"},
-	{{zio_checksum_SHA256,	zio_checksum_SHA256},	1, 1, 0, "label"},
-	{{zio_checksum_SHA256,	zio_checksum_SHA256},	1, 1, 0, "gang_header"},
-	{{fletcher_2_native,	fletcher_2_byteswap},	0, 1, 0, "zilog"},
-	{{fletcher_2_native,	fletcher_2_byteswap},	0, 0, 0, "fletcher2"},
-	{{fletcher_4_native,	fletcher_4_byteswap},	1, 0, 0, "fletcher4"},
-	{{zio_checksum_SHA256,	zio_checksum_SHA256},	1, 0, 1, "sha256"},
-	{{fletcher_4_native,	fletcher_4_byteswap},	0, 1, 0, "zilog2"},
+	{{NULL, NULL}, NULL, NULL, 0, "inherit"},
+	{{NULL, NULL}, NULL, NULL, 0, "on"},
+	{{zio_checksum_off,		zio_checksum_off},
+	    NULL, NULL, 0, "off"},
+	{{zio_checksum_SHA256,		zio_checksum_SHA256},
+	    NULL, NULL, ZCHECKSUM_FLAG_METADATA | ZCHECKSUM_FLAG_EMBEDDED,
+	    "label"},
+	{{zio_checksum_SHA256,		zio_checksum_SHA256},
+	    NULL, NULL, ZCHECKSUM_FLAG_METADATA | ZCHECKSUM_FLAG_EMBEDDED,
+	    "gang_header"},
+	{{fletcher_2_native,		fletcher_2_byteswap},
+	    NULL, NULL, ZCHECKSUM_FLAG_EMBEDDED, "zilog"},
+	{{fletcher_2_native,		fletcher_2_byteswap},
+	    NULL, NULL, 0, "fletcher2"},
+	{{fletcher_4_native,		fletcher_4_byteswap},
+	    NULL, NULL, ZCHECKSUM_FLAG_METADATA, "fletcher4"},
+	{{zio_checksum_SHA256,		zio_checksum_SHA256},
+	    NULL, NULL, ZCHECKSUM_FLAG_METADATA | ZCHECKSUM_FLAG_DEDUP |
+	    ZCHECKSUM_FLAG_NOPWRITE, "sha256"},
+	{{fletcher_4_native,		fletcher_4_byteswap},
+	    NULL, NULL, ZCHECKSUM_FLAG_EMBEDDED, "zilog2"},
+	{{zio_checksum_off,		zio_checksum_off},
+	    NULL, NULL, 0, "noparity"},
+	{{zio_checksum_SHA512_native,	zio_checksum_SHA512_byteswap},
+	    NULL, NULL, ZCHECKSUM_FLAG_METADATA | ZCHECKSUM_FLAG_DEDUP |
+	    ZCHECKSUM_FLAG_NOPWRITE, "sha512"},
+	{{zio_checksum_skein_native,	zio_checksum_skein_byteswap},
+	    zio_checksum_skein_tmpl_init, zio_checksum_skein_tmpl_free,
+	    ZCHECKSUM_FLAG_METADATA | ZCHECKSUM_FLAG_DEDUP |
+	    ZCHECKSUM_FLAG_SALTED | ZCHECKSUM_FLAG_NOPWRITE, "skein"},
+	{{zio_checksum_edonr_native,	zio_checksum_edonr_byteswap},
+	    zio_checksum_edonr_tmpl_init, zio_checksum_edonr_tmpl_free,
+	    ZCHECKSUM_FLAG_METADATA | ZCHECKSUM_FLAG_SALTED |
+	    ZCHECKSUM_FLAG_NOPWRITE, "edonr"},
 };
 
+spa_feature_t
+zio_checksum_to_feature(enum zio_checksum cksum)
+{
+	switch (cksum) {
+	case ZIO_CHECKSUM_SHA512:
+		return (SPA_FEATURE_SHA512);
+	case ZIO_CHECKSUM_SKEIN:
+		return (SPA_FEATURE_SKEIN);
+	case ZIO_CHECKSUM_EDONR:
+		return (SPA_FEATURE_EDONR);
+	default:
+		return (SPA_FEATURE_NONE);
+	}
+}
+
 enum zio_checksum
 zio_checksum_select(enum zio_checksum child, enum zio_checksum parent)
 {
@@ -113,7 +183,8 @@ zio_checksum_dedup_select(spa_t *spa, enum zio_checksum child,
 	if (child == (ZIO_CHECKSUM_ON | ZIO_CHECKSUM_VERIFY))
 		return (spa_dedup_checksum(spa) | ZIO_CHECKSUM_VERIFY);
 
-	ASSERT(zio_checksum_table[child & ZIO_CHECKSUM_MASK].ci_dedup ||
+	ASSERT((zio_checksum_table[child & ZIO_CHECKSUM_MASK].ci_flags &
+	    ZCHECKSUM_FLAG_DEDUP) ||
 	    (child & ZIO_CHECKSUM_VERIFY) || child == ZIO_CHECKSUM_OFF);
 
 	return (child);
@@ -146,6 +217,30 @@ zio_checksum_label_verifier(zio_cksum_t *zcp, uint64_t offset)
 }
 
 /*
+ * Calls the template init function of a checksum which supports context
+ * templates and installs the template into the spa_t.
+ */
+static void
+zio_checksum_template_init(enum zio_checksum checksum, spa_t *spa)
+{
+	zio_checksum_info_t *ci = &zio_checksum_table[checksum];
+
+	if (ci->ci_tmpl_init == NULL)
+		return;
+	if (spa->spa_cksum_tmpls[checksum] != NULL)
+		return;
+
+	VERIFY(ci->ci_tmpl_free != NULL);
+	mutex_enter(&spa->spa_cksum_tmpls_lock);
+	if (spa->spa_cksum_tmpls[checksum] == NULL) {
+		spa->spa_cksum_tmpls[checksum] =
+		    ci->ci_tmpl_init(&spa->spa_cksum_salt);
+		VERIFY(spa->spa_cksum_tmpls[checksum] != NULL);
+	}
+	mutex_exit(&spa->spa_cksum_tmpls_lock);
+}
+
+/*
  * Generate the checksum.
  */
 void
@@ -156,11 +251,14 @@ zio_checksum_compute(zio_t *zio, enum zio_checksum checksum,
 	uint64_t offset = zio->io_offset;
 	zio_checksum_info_t *ci = &zio_checksum_table[checksum];
 	zio_cksum_t cksum;
+	spa_t *spa = zio->io_spa;
 
 	ASSERT((uint_t)checksum < ZIO_CHECKSUM_FUNCTIONS);
 	ASSERT(ci->ci_func[0] != NULL);
 
-	if (ci->ci_eck) {
+	zio_checksum_template_init(checksum, spa);
+
+	if (ci->ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
 		zio_eck_t *eck;
 
 		if (checksum == ZIO_CHECKSUM_ZILOG2) {
@@ -179,10 +277,12 @@ zio_checksum_compute(zio_t *zio, enum zio_checksum checksum,
 		else
 			bp->blk_cksum = eck->zec_cksum;
 		eck->zec_magic = ZEC_MAGIC;
-		ci->ci_func[0](data, size, &cksum);
+		ci->ci_func[0](data, size, spa->spa_cksum_tmpls[checksum],
+		    &cksum);
 		eck->zec_cksum = cksum;
 	} else {
-		ci->ci_func[0](data, size, &bp->blk_cksum);
+		ci->ci_func[0](data, size, spa->spa_cksum_tmpls[checksum],
+		    &bp->blk_cksum);
 	}
 }
 
@@ -191,13 +291,15 @@ zio_checksum_error_impl(spa_t *spa, blkptr_t *bp, enum zio_checksum checksum,
     void *data, uint64_t size, uint64_t offset, zio_bad_cksum_t *info)
 {
 	zio_checksum_info_t *ci = &zio_checksum_table[checksum];
-	zio_cksum_t actual_cksum, expected_cksum;
 	int byteswap;
+	zio_cksum_t actual_cksum, expected_cksum;
 
 	if (checksum >= ZIO_CHECKSUM_FUNCTIONS || ci->ci_func[0] == NULL)
 		return (SET_ERROR(EINVAL));
 
-	if (ci->ci_eck) {
+	zio_checksum_template_init(checksum, spa);
+
+	if (ci->ci_flags & ZCHECKSUM_FLAG_EMBEDDED) {
 		zio_eck_t *eck;
 		zio_cksum_t verifier;
 
@@ -235,7 +337,8 @@ zio_checksum_error_impl(spa_t *spa, blkptr_t *bp, enum zio_checksum checksum,
 
 		expected_cksum = eck->zec_cksum;
 		eck->zec_cksum = verifier;
-		ci->ci_func[byteswap](data, size, &actual_cksum);
+		ci->ci_func[byteswap](data, size,
+		    spa->spa_cksum_tmpls[checksum], &actual_cksum);
 		eck->zec_cksum = expected_cksum;
 
 		if (byteswap) {
@@ -245,7 +348,8 @@ zio_checksum_error_impl(spa_t *spa, blkptr_t *bp, enum zio_checksum checksum,
 	} else {
 		byteswap = BP_SHOULD_BYTESWAP(bp);
 		expected_cksum = bp->blk_cksum;
-		ci->ci_func[byteswap](data, size, &actual_cksum);
+		ci->ci_func[byteswap](data, size,
+		    spa->spa_cksum_tmpls[checksum], &actual_cksum);
 	}
 
 	if (info != NULL) {
@@ -286,3 +390,24 @@ zio_checksum_error(zio_t *zio, zio_bad_cksum_t *info)
 	}
 	return (error);
 }
+
+/*
+ * Called by a spa_t that's about to be deallocated. This steps through
+ * all of the checksum context templates and deallocates any that were
+ * initialized using the algorithm-specific template init function.
+ */
+void
+zio_checksum_templates_free(spa_t *spa)
+{
+	enum zio_checksum checksum;
+	for (checksum = 0; checksum < ZIO_CHECKSUM_FUNCTIONS;
+	    checksum++) {
+		if (spa->spa_cksum_tmpls[checksum] != NULL) {
+			zio_checksum_info_t *ci = &zio_checksum_table[checksum];
+
+			VERIFY(ci->ci_tmpl_free != NULL);
+			ci->ci_tmpl_free(spa->spa_cksum_tmpls[checksum]);
+			spa->spa_cksum_tmpls[checksum] = NULL;
+		}
+	}
+}