OpenZFS 1300 - filename normalization doesn't work for removes

Authored by: Kevin Crowe <kevin.crowe@nexenta.com>
Reviewed by: Yuri Pankov <yuri.pankov@nexenta.com>
Reviewed by: Pavel Zakharov <pavel.zakharov@delphix.com>
Reviewed by: Matt Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Ported-by: George Melikov <mail@gmelikov.ru>

OpenZFS-issue: https://www.illumos.org/issues/1300
OpenZFS-commit: https://github.com/openzfs/openzfs/commit/8f1750d
Closes #5725 

Porting notes:
- zap_micro.c: all `MT_EXACT` are replaced by `0`

include/sys/zap.h

@@ -18,9 +18,11 @@
  *
  * CDDL HEADER END
  */
+
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  * Copyright (c) 2013 by Delphix. All rights reserved.
+ * Copyright 2017 Nexenta Systems, Inc.
  */
 
 #ifndef	_SYS_ZAP_H
@@ -87,22 +89,15 @@ extern "C" {
 
 /*
  * Specifies matching criteria for ZAP lookups.
+ * MT_NORMALIZE		Use ZAP normalization flags, which can include both
+ *			unicode normalization and case-insensitivity.
+ * MT_MATCH_CASE	Do case-sensitive lookups even if MT_NORMALIZE is
+ *			specified and ZAP normalization flags include
+ *			U8_TEXTPREP_TOUPPER.
  */
-typedef enum matchtype
+typedef enum matchtype {
-{
+	MT_NORMALIZE = 1 << 0,
-	/* Only find an exact match (non-normalized) */
+	MT_MATCH_CASE = 1 << 1,
-	MT_EXACT,
-	/*
-	 * If there is an exact match, find that, otherwise find the
-	 * first normalized match.
-	 */
-	MT_BEST,
-	/*
-	 * Find the "first" normalized (case and Unicode form) match;
-	 * the designated "first" match will not change as long as the
-	 * set of entries with this normalization doesn't change.
-	 */
-	MT_FIRST
 } matchtype_t;
 
 typedef enum zap_flags {
@@ -119,19 +114,6 @@ typedef enum zap_flags {
 
 /*
  * Create a new zapobj with no attributes and return its object number.
- * MT_EXACT will cause the zap object to only support MT_EXACT lookups,
- * otherwise any matchtype can be used for lookups.
- *
- * normflags specifies what normalization will be done.  values are:
- * 0: no normalization (legacy on-disk format, supports MT_EXACT matching
- *     only)
- * U8_TEXTPREP_TOLOWER: case normalization will be performed.
- *     MT_FIRST/MT_BEST matching will find entries that match without
- *     regard to case (eg. looking for "foo" can find an entry "Foo").
- * Eventually, other flags will permit unicode normalization as well.
- *
- * dnodesize specifies the on-disk size of the dnode for the new zapobj.
- * Valid values are multiples of 512 up to DNODE_MAX_SIZE.
  */
 uint64_t zap_create(objset_t *ds, dmu_object_type_t ot,
     dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx);

include/sys/zap_impl.h

@@ -18,10 +18,12 @@
  *
  * CDDL HEADER END
  */
+
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
  * Copyright (c) 2013, 2016 by Delphix. All rights reserved.
+ * Copyright 2017 Nexenta Systems, Inc.
  */
 
 #ifndef	_SYS_ZAP_IMPL_H
@@ -188,6 +190,7 @@ typedef struct zap_name {
 	int zn_key_norm_numints;
 	uint64_t zn_hash;
 	matchtype_t zn_matchtype;
+	int zn_normflags;
 	char zn_normbuf[ZAP_MAXNAMELEN];
 } zap_name_t;
 

module/zfs/dmu_objset.c

@@ -18,15 +18,16 @@
  *
  * CDDL HEADER END
  */
+
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  * Copyright (c) 2012, 2016 by Delphix. All rights reserved.
  * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
  * Copyright (c) 2013, Joyent, Inc. All rights reserved.
  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
- * Copyright (c) 2015 Nexenta Systems, Inc. All rights reserved.
  * Copyright (c) 2015, STRATO AG, Inc. All rights reserved.
  * Copyright (c) 2016 Actifio, Inc. All rights reserved.
+ * Copyright 2017 Nexenta Systems, Inc.
  */
 
 /* Portions Copyright 2010 Robert Milkowski */
@@ -1849,7 +1850,7 @@ dmu_snapshot_realname(objset_t *os, char *name, char *real, int maxlen,
 
 	return (zap_lookup_norm(ds->ds_dir->dd_pool->dp_meta_objset,
 	    dsl_dataset_phys(ds)->ds_snapnames_zapobj, name, 8, 1, &ignored,
-	    MT_FIRST, real, maxlen, conflict));
+	    MT_NORMALIZE, real, maxlen, conflict));
 }
 
 int

module/zfs/dsl_bookmark.c

@@ -12,8 +12,10 @@
  *
  * CDDL HEADER END
  */
+
 /*
  * Copyright (c) 2013, 2014 by Delphix. All rights reserved.
+ * Copyright 2017 Nexenta Systems, Inc.
  */
 
 #include <sys/zfs_context.h>
@@ -59,16 +61,14 @@ dsl_dataset_bmark_lookup(dsl_dataset_t *ds, const char *shortname,
 {
 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
 	uint64_t bmark_zapobj = ds->ds_bookmarks;
-	matchtype_t mt;
+	matchtype_t mt = 0;
 	int err;
 
 	if (bmark_zapobj == 0)
 		return (SET_ERROR(ESRCH));
 
 	if (dsl_dataset_phys(ds)->ds_flags & DS_FLAG_CI_DATASET)
-		mt = MT_FIRST;
+		mt = MT_NORMALIZE;
-	else
-		mt = MT_EXACT;
 
 	err = zap_lookup_norm(mos, bmark_zapobj, shortname, sizeof (uint64_t),
 	    sizeof (*bmark_phys) / sizeof (uint64_t), bmark_phys, mt,
@@ -342,12 +342,10 @@ dsl_dataset_bookmark_remove(dsl_dataset_t *ds, const char *name, dmu_tx_t *tx)
 {
 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
 	uint64_t bmark_zapobj = ds->ds_bookmarks;
-	matchtype_t mt;
+	matchtype_t mt = 0;
 
 	if (dsl_dataset_phys(ds)->ds_flags & DS_FLAG_CI_DATASET)
-		mt = MT_FIRST;
+		mt = MT_NORMALIZE;
-	else
-		mt = MT_EXACT;
 
 	return (zap_remove_norm(mos, bmark_zapobj, name, mt, tx));
 }

module/zfs/dsl_dataset.c

@@ -18,6 +18,7 @@
  *
  * CDDL HEADER END
  */
+
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  * Copyright (c) 2011, 2016 by Delphix. All rights reserved.
@@ -26,6 +27,7 @@
  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
  * Copyright (c) 2016 Actifio, Inc. All rights reserved.
  * Copyright 2016, OmniTI Computer Consulting, Inc. All rights reserved.
+ * Copyright 2017 Nexenta Systems, Inc.
  */
 
 #include <sys/dmu_objset.h>
@@ -360,17 +362,15 @@ dsl_dataset_snap_lookup(dsl_dataset_t *ds, const char *name, uint64_t *value)
 {
 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
 	uint64_t snapobj = dsl_dataset_phys(ds)->ds_snapnames_zapobj;
-	matchtype_t mt;
+	matchtype_t mt = 0;
 	int err;
 
 	if (dsl_dataset_phys(ds)->ds_flags & DS_FLAG_CI_DATASET)
-		mt = MT_FIRST;
+		mt = MT_NORMALIZE;
-	else
-		mt = MT_EXACT;
 
 	err = zap_lookup_norm(mos, snapobj, name, 8, 1,
 	    value, mt, NULL, 0, NULL);
-	if (err == ENOTSUP && mt == MT_FIRST)
+	if (err == ENOTSUP && (mt & MT_NORMALIZE))
 		err = zap_lookup(mos, snapobj, name, 8, 1, value);
 	return (err);
 }
@@ -381,18 +381,16 @@ dsl_dataset_snap_remove(dsl_dataset_t *ds, const char *name, dmu_tx_t *tx,
 {
 	objset_t *mos = ds->ds_dir->dd_pool->dp_meta_objset;
 	uint64_t snapobj = dsl_dataset_phys(ds)->ds_snapnames_zapobj;
-	matchtype_t mt;
+	matchtype_t mt = 0;
 	int err;
 
 	dsl_dir_snap_cmtime_update(ds->ds_dir);
 
 	if (dsl_dataset_phys(ds)->ds_flags & DS_FLAG_CI_DATASET)
-		mt = MT_FIRST;
+		mt = MT_NORMALIZE;
-	else
-		mt = MT_EXACT;
 
 	err = zap_remove_norm(mos, snapobj, name, mt, tx);
-	if (err == ENOTSUP && mt == MT_FIRST)
+	if (err == ENOTSUP && (mt & MT_NORMALIZE))
 		err = zap_remove(mos, snapobj, name, tx);
 
 	if (err == 0 && adj_cnt)

module/zfs/zap_leaf.c

@@ -18,9 +18,11 @@
  *
  * CDDL HEADER END
  */
+
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  * Copyright (c) 2013, 2015 by Delphix. All rights reserved.
+ * Copyright 2017 Nexenta Systems, Inc.
  */
 
 /*
@@ -364,7 +366,7 @@ zap_leaf_array_match(zap_leaf_t *l, zap_name_t *zn,
 	}
 
 	ASSERT(zn->zn_key_intlen == 1);
-	if (zn->zn_matchtype == MT_FIRST) {
+	if (zn->zn_matchtype & MT_NORMALIZE) {
 		char *thisname = kmem_alloc(array_numints, KM_SLEEP);
 		boolean_t match;
 
@@ -406,7 +408,6 @@ zap_leaf_lookup(zap_leaf_t *l, zap_name_t *zn, zap_entry_handle_t *zeh)
 
 	ASSERT3U(zap_leaf_phys(l)->l_hdr.lh_magic, ==, ZAP_LEAF_MAGIC);
 
-again:
 	for (chunkp = LEAF_HASH_ENTPTR(l, zn->zn_hash);
 	    *chunkp != CHAIN_END; chunkp = &le->le_next) {
 		uint16_t chunk = *chunkp;
@@ -421,9 +422,9 @@ again:
 		/*
 		 * NB: the entry chain is always sorted by cd on
 		 * normalized zap objects, so this will find the
-		 * lowest-cd match for MT_FIRST.
+		 * lowest-cd match for MT_NORMALIZE.
 		 */
-		ASSERT(zn->zn_matchtype == MT_EXACT ||
+		ASSERT((zn->zn_matchtype == 0) ||
 		    (zap_leaf_phys(l)->l_hdr.lh_flags & ZLF_ENTRIES_CDSORTED));
 		if (zap_leaf_array_match(l, zn, le->le_name_chunk,
 		    le->le_name_numints)) {
@@ -437,15 +438,6 @@ again:
 		}
 	}
 
-	/*
-	 * NB: we could of course do this in one pass, but that would be
-	 * a pain.  We'll see if MT_BEST is even used much.
-	 */
-	if (zn->zn_matchtype == MT_BEST) {
-		zn->zn_matchtype = MT_FIRST;
-		goto again;
-	}
-
 	return (SET_ERROR(ENOENT));
 }
 
@@ -700,7 +692,7 @@ zap_entry_normalization_conflict(zap_entry_handle_t *zeh, zap_name_t *zn,
 			continue;
 
 		if (zn == NULL) {
-			zn = zap_name_alloc(zap, name, MT_FIRST);
+			zn = zap_name_alloc(zap, name, MT_NORMALIZE);
 			allocdzn = B_TRUE;
 		}
 		if (zap_leaf_array_match(zeh->zeh_leaf, zn,

module/zfs/zap_micro.c

@@ -18,10 +18,12 @@
  *
  * CDDL HEADER END
  */
+
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  * Copyright (c) 2011, 2016 by Delphix. All rights reserved.
  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
+ * Copyright 2017 Nexenta Systems, Inc.
  */
 
 #include <sys/zio.h>
@@ -132,7 +134,7 @@ zap_hash(zap_name_t *zn)
 }
 
 static int
-zap_normalize(zap_t *zap, const char *name, char *namenorm)
+zap_normalize(zap_t *zap, const char *name, char *namenorm, int normflags)
 {
 	size_t inlen, outlen;
 	int err;
@@ -144,8 +146,8 @@ zap_normalize(zap_t *zap, const char *name, char *namenorm)
 
 	err = 0;
 	(void) u8_textprep_str((char *)name, &inlen, namenorm, &outlen,
-	    zap->zap_normflags | U8_TEXTPREP_IGNORE_NULL |
+	    normflags | U8_TEXTPREP_IGNORE_NULL | U8_TEXTPREP_IGNORE_INVALID,
-	    U8_TEXTPREP_IGNORE_INVALID, U8_UNICODE_LATEST, &err);
+	    U8_UNICODE_LATEST, &err);
 
 	return (err);
 }
@@ -155,15 +157,15 @@ zap_match(zap_name_t *zn, const char *matchname)
 {
 	ASSERT(!(zap_getflags(zn->zn_zap) & ZAP_FLAG_UINT64_KEY));
 
-	if (zn->zn_matchtype == MT_FIRST) {
+	if (zn->zn_matchtype & MT_NORMALIZE) {
 		char norm[ZAP_MAXNAMELEN];
 
-		if (zap_normalize(zn->zn_zap, matchname, norm) != 0)
+		if (zap_normalize(zn->zn_zap, matchname, norm,
+		    zn->zn_normflags) != 0)
 			return (B_FALSE);
 
 		return (strcmp(zn->zn_key_norm, norm) == 0);
 	} else {
-		/* MT_BEST or MT_EXACT */
 		return (strcmp(zn->zn_key_orig, matchname) == 0);
 	}
 }
@@ -184,15 +186,30 @@ zap_name_alloc(zap_t *zap, const char *key, matchtype_t mt)
 	zn->zn_key_orig = key;
 	zn->zn_key_orig_numints = strlen(zn->zn_key_orig) + 1;
 	zn->zn_matchtype = mt;
+	zn->zn_normflags = zap->zap_normflags;
+
+	/*
+	 * If we're dealing with a case sensitive lookup on a mixed or
+	 * insensitive fs, remove U8_TEXTPREP_TOUPPER or the lookup
+	 * will fold case to all caps overriding the lookup request.
+	 */
+	if (mt & MT_MATCH_CASE)
+		zn->zn_normflags &= ~U8_TEXTPREP_TOUPPER;
+
 	if (zap->zap_normflags) {
-		if (zap_normalize(zap, key, zn->zn_normbuf) != 0) {
+		/*
+		 * We *must* use zap_normflags because this normalization is
+		 * what the hash is computed from.
+		 */
+		if (zap_normalize(zap, key, zn->zn_normbuf,
+		    zap->zap_normflags) != 0) {
 			zap_name_free(zn);
 			return (NULL);
 		}
 		zn->zn_key_norm = zn->zn_normbuf;
 		zn->zn_key_norm_numints = strlen(zn->zn_key_norm) + 1;
 	} else {
-		if (mt != MT_EXACT) {
+		if (mt != 0) {
 			zap_name_free(zn);
 			return (NULL);
 		}
@@ -201,6 +218,20 @@ zap_name_alloc(zap_t *zap, const char *key, matchtype_t mt)
 	}
 
 	zn->zn_hash = zap_hash(zn);
+
+	if (zap->zap_normflags != zn->zn_normflags) {
+		/*
+		 * We *must* use zn_normflags because this normalization is
+		 * what the matching is based on.  (Not the hash!)
+		 */
+		if (zap_normalize(zap, key, zn->zn_normbuf,
+		    zn->zn_normflags) != 0) {
+			zap_name_free(zn);
+			return (NULL);
+		}
+		zn->zn_key_norm_numints = strlen(zn->zn_key_norm) + 1;
+	}
+
 	return (zn);
 }
 
@@ -214,7 +245,7 @@ zap_name_alloc_uint64(zap_t *zap, const uint64_t *key, int numints)
 	zn->zn_key_intlen = sizeof (*key);
 	zn->zn_key_orig = zn->zn_key_norm = key;
 	zn->zn_key_orig_numints = zn->zn_key_norm_numints = numints;
-	zn->zn_matchtype = MT_EXACT;
+	zn->zn_matchtype = 0;
 
 	zn->zn_hash = zap_hash(zn);
 	return (zn);
@@ -294,7 +325,6 @@ mze_find(zap_name_t *zn)
 	mze_tofind.mze_hash = zn->zn_hash;
 	mze_tofind.mze_cd = 0;
 
-again:
 	mze = avl_find(avl, &mze_tofind, &idx);
 	if (mze == NULL)
 		mze = avl_nearest(avl, idx, AVL_AFTER);
@@ -303,10 +333,7 @@ again:
 		if (zap_match(zn, MZE_PHYS(zn->zn_zap, mze)->mze_name))
 			return (mze);
 	}
-	if (zn->zn_matchtype == MT_BEST) {
+
-		zn->zn_matchtype = MT_FIRST;
-		goto again;
-	}
 	return (NULL);
 }
 
@@ -412,8 +439,7 @@ mzap_open(objset_t *os, uint64_t obj, dmu_buf_t *db)
 				zap_name_t *zn;
 
 				zap->zap_m.zap_num_entries++;
-				zn = zap_name_alloc(zap, mze->mze_name,
+				zn = zap_name_alloc(zap, mze->mze_name, 0);
-				    MT_EXACT);
 				mze_insert(zap, i, zn->zn_hash);
 				zap_name_free(zn);
 			}
@@ -611,7 +637,7 @@ mzap_upgrade(zap_t **zapp, void *tag, dmu_tx_t *tx, zap_flags_t flags)
 			continue;
 		dprintf("adding %s=%llu\n",
 		    mze->mze_name, mze->mze_value);
-		zn = zap_name_alloc(zap, mze->mze_name, MT_EXACT);
+		zn = zap_name_alloc(zap, mze->mze_name, 0);
 		err = fzap_add_cd(zn, 8, 1, &mze->mze_value, mze->mze_cd,
 		    tag, tx);
 		zap = zn->zn_zap;	/* fzap_add_cd() may change zap */
@@ -624,6 +650,23 @@ mzap_upgrade(zap_t **zapp, void *tag, dmu_tx_t *tx, zap_flags_t flags)
 	return (err);
 }
 
+/*
+ * The "normflags" determine the behavior of the matchtype_t which is
+ * passed to zap_lookup_norm().  Names which have the same normalized
+ * version will be stored with the same hash value, and therefore we can
+ * perform normalization-insensitive lookups.  We can be Unicode form-
+ * insensitive and/or case-insensitive.  The following flags are valid for
+ * "normflags":
+ *
+ * U8_TEXTPREP_NFC
+ * U8_TEXTPREP_NFD
+ * U8_TEXTPREP_NFKC
+ * U8_TEXTPREP_NFKD
+ * U8_TEXTPREP_TOUPPER
+ *
+ * The *_NF* (Normalization Form) flags are mutually exclusive; at most one
+ * of them may be supplied.
+ */
 void
 mzap_create_impl(objset_t *os, uint64_t obj, int normflags, zap_flags_t flags,
     dmu_tx_t *tx)
@@ -827,7 +870,7 @@ again:
 
 		if (zn == NULL) {
 			zn = zap_name_alloc(zap, MZE_PHYS(zap, mze)->mze_name,
-			    MT_FIRST);
+			    MT_NORMALIZE);
 			allocdzn = B_TRUE;
 		}
 		if (zap_match(zn, MZE_PHYS(zap, other)->mze_name)) {
@@ -856,7 +899,7 @@ zap_lookup(objset_t *os, uint64_t zapobj, const char *name,
     uint64_t integer_size, uint64_t num_integers, void *buf)
 {
 	return (zap_lookup_norm(os, zapobj, name, integer_size,
-	    num_integers, buf, MT_EXACT, NULL, 0, NULL));
+	    num_integers, buf, 0, NULL, 0, NULL));
 }
 
 static int
@@ -929,7 +972,7 @@ zap_prefetch(objset_t *os, uint64_t zapobj, const char *name)
 	err = zap_lockdir(os, zapobj, NULL, RW_READER, TRUE, FALSE, FTAG, &zap);
 	if (err)
 		return (err);
-	zn = zap_name_alloc(zap, name, MT_EXACT);
+	zn = zap_name_alloc(zap, name, 0);
 	if (zn == NULL) {
 		zap_unlockdir(zap, FTAG);
 		return (SET_ERROR(ENOTSUP));
@@ -946,7 +989,7 @@ zap_lookup_by_dnode(dnode_t *dn, const char *name,
     uint64_t integer_size, uint64_t num_integers, void *buf)
 {
 	return (zap_lookup_norm_by_dnode(dn, name, integer_size,
-	    num_integers, buf, MT_EXACT, NULL, 0, NULL));
+	    num_integers, buf, 0, NULL, 0, NULL));
 }
 
 int
@@ -1019,7 +1062,7 @@ int
 zap_contains(objset_t *os, uint64_t zapobj, const char *name)
 {
 	int err = zap_lookup_norm(os, zapobj, name, 0,
-	    0, NULL, MT_EXACT, NULL, 0, NULL);
+	    0, NULL, 0, NULL, 0, NULL);
 	if (err == EOVERFLOW || err == EINVAL)
 		err = 0; /* found, but skipped reading the value */
 	return (err);
@@ -1037,7 +1080,7 @@ zap_length(objset_t *os, uint64_t zapobj, const char *name,
 	err = zap_lockdir(os, zapobj, NULL, RW_READER, TRUE, FALSE, FTAG, &zap);
 	if (err)
 		return (err);
-	zn = zap_name_alloc(zap, name, MT_EXACT);
+	zn = zap_name_alloc(zap, name, 0);
 	if (zn == NULL) {
 		zap_unlockdir(zap, FTAG);
 		return (SET_ERROR(ENOTSUP));
@@ -1138,7 +1181,7 @@ zap_add_impl(zap_t *zap, const char *key,
 	const uint64_t *intval = val;
 	zap_name_t *zn;
 
-	zn = zap_name_alloc(zap, key, MT_EXACT);
+	zn = zap_name_alloc(zap, key, 0);
 	if (zn == NULL) {
 		zap_unlockdir(zap, tag);
 		return (SET_ERROR(ENOTSUP));
@@ -1249,7 +1292,7 @@ zap_update(objset_t *os, uint64_t zapobj, const char *name,
 	err = zap_lockdir(os, zapobj, tx, RW_WRITER, TRUE, TRUE, FTAG, &zap);
 	if (err)
 		return (err);
-	zn = zap_name_alloc(zap, name, MT_EXACT);
+	zn = zap_name_alloc(zap, name, 0);
 	if (zn == NULL) {
 		zap_unlockdir(zap, FTAG);
 		return (SET_ERROR(ENOTSUP));
@@ -1312,7 +1355,7 @@ zap_update_uint64(objset_t *os, uint64_t zapobj, const uint64_t *key,
 int
 zap_remove(objset_t *os, uint64_t zapobj, const char *name, dmu_tx_t *tx)
 {
-	return (zap_remove_norm(os, zapobj, name, MT_EXACT, tx));
+	return (zap_remove_norm(os, zapobj, name, 0, tx));
 }
 
 static int
@@ -1367,7 +1410,7 @@ zap_remove_by_dnode(dnode_t *dn, const char *name, dmu_tx_t *tx)
 	err = zap_lockdir_by_dnode(dn, tx, RW_WRITER, TRUE, FALSE, FTAG, &zap);
 	if (err)
 		return (err);
-	err = zap_remove_impl(zap, name, MT_EXACT, tx);
+	err = zap_remove_impl(zap, name, 0, tx);
 	zap_unlockdir(zap, FTAG);
 	return (err);
 }
@@ -1586,7 +1629,7 @@ zap_count_write_by_dnode(dnode_t *dn, const char *name, int add,
 		return (err);
 
 	if (!zap->zap_ismicro) {
-		zap_name_t *zn = zap_name_alloc(zap, name, MT_EXACT);
+		zap_name_t *zn = zap_name_alloc(zap, name, 0);
 		if (zn) {
 			err = fzap_count_write(zn, add, towrite,
 			    tooverwrite);

module/zfs/zfs_dir.c

@@ -18,9 +18,11 @@
  *
  * CDDL HEADER END
  */
+
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  * Copyright (c) 2013, 2015 by Delphix. All rights reserved.
+ * Copyright 2017 Nexenta Systems, Inc.
  */
 
 
@@ -64,14 +66,13 @@
  * of names after deciding which is the appropriate lookup interface.
  */
 static int
-zfs_match_find(zfs_sb_t *zsb, znode_t *dzp, char *name, boolean_t exact,
+zfs_match_find(zfs_sb_t *zsb, znode_t *dzp, char *name, matchtype_t mt,
     boolean_t update, int *deflags, pathname_t *rpnp, uint64_t *zoid)
 {
 	boolean_t conflict = B_FALSE;
 	int error;
 
 	if (zsb->z_norm) {
-		matchtype_t mt = MT_FIRST;
 		size_t bufsz = 0;
 		char *buf = NULL;
 
@@ -79,8 +80,7 @@ zfs_match_find(zfs_sb_t *zsb, znode_t *dzp, char *name, boolean_t exact,
 			buf = rpnp->pn_buf;
 			bufsz = rpnp->pn_bufsize;
 		}
-		if (exact)
+
-			mt = MT_EXACT;
 		/*
 		 * In the non-mixed case we only expect there would ever
 		 * be one match, but we need to use the normalizing lookup.
@@ -156,7 +156,7 @@ zfs_dirent_lock(zfs_dirlock_t **dlpp, znode_t *dzp, char *name, znode_t **zpp,
 	zfs_sb_t	*zsb = ZTOZSB(dzp);
 	zfs_dirlock_t	*dl;
 	boolean_t	update;
-	boolean_t	exact;
+	matchtype_t	mt = 0;
 	uint64_t	zoid;
 #ifdef HAVE_DNLC
 	vnode_t		*vp = NULL;
@@ -193,13 +193,29 @@ zfs_dirent_lock(zfs_dirlock_t **dlpp, znode_t *dzp, char *name, znode_t **zpp,
 	 */
 
 	/*
-	 * Decide if exact matches should be requested when performing
+	 * When matching we may need to normalize & change case according to
-	 * a zap lookup on file systems supporting case-insensitive
+	 * FS settings.
-	 * access.
+	 *
+	 * Note that a normalized match is necessary for a case insensitive
+	 * filesystem when the lookup request is not exact because normalization
+	 * can fold case independent of normalizing code point sequences.
+	 *
+	 * See the table above zfs_dropname().
 	 */
-	exact =
+	if (zsb->z_norm != 0) {
-	    ((zsb->z_case == ZFS_CASE_INSENSITIVE) && (flag & ZCIEXACT)) ||
+		mt = MT_NORMALIZE;
-	    ((zsb->z_case == ZFS_CASE_MIXED) && !(flag & ZCILOOK));
+
+		/*
+		 * Determine if the match needs to honor the case specified in
+		 * lookup, and if so keep track of that so that during
+		 * normalization we don't fold case.
+		 */
+		if ((zsb->z_case == ZFS_CASE_INSENSITIVE &&
+		    (flag & ZCIEXACT)) ||
+		    (zsb->z_case == ZFS_CASE_MIXED && !(flag & ZCILOOK))) {
+			mt |= MT_MATCH_CASE;
+		}
+	}
 
 	/*
 	 * Only look in or update the DNLC if we are looking for the
@@ -212,7 +228,7 @@ zfs_dirent_lock(zfs_dirlock_t **dlpp, znode_t *dzp, char *name, znode_t **zpp,
 	 * case for performance improvement?
 	 */
 	update = !zsb->z_norm ||
-	    ((zsb->z_case == ZFS_CASE_MIXED) &&
+	    (zsb->z_case == ZFS_CASE_MIXED &&
 	    !(zsb->z_norm & ~U8_TEXTPREP_TOUPPER) && !(flag & ZCILOOK));
 
 	/*
@@ -327,11 +343,11 @@ zfs_dirent_lock(zfs_dirlock_t **dlpp, znode_t *dzp, char *name, znode_t **zpp,
 			*zpp = VTOZ(vp);
 			return (0);
 		} else {
-			error = zfs_match_find(zsb, dzp, name, exact,
+			error = zfs_match_find(zsb, dzp, name, mt,
 			    update, direntflags, realpnp, &zoid);
 		}
 #else
-		error = zfs_match_find(zsb, dzp, name, exact,
+		error = zfs_match_find(zsb, dzp, name, mt,
 		    update, direntflags, realpnp, &zoid);
 #endif /* HAVE_DNLC */
 	}
@@ -805,6 +821,28 @@ zfs_link_create(zfs_dirlock_t *dl, znode_t *zp, dmu_tx_t *tx, int flag)
 	return (0);
 }
 
+/*
+ * The match type in the code for this function should conform to:
+ *
+ * ------------------------------------------------------------------------
+ * fs type  | z_norm      | lookup type | match type
+ * ---------|-------------|-------------|----------------------------------
+ * CS !norm | 0           |           0 | 0 (exact)
+ * CS  norm | formX       |           0 | MT_NORMALIZE
+ * CI !norm | upper       |   !ZCIEXACT | MT_NORMALIZE
+ * CI !norm | upper       |    ZCIEXACT | MT_NORMALIZE | MT_MATCH_CASE
+ * CI  norm | upper|formX |   !ZCIEXACT | MT_NORMALIZE
+ * CI  norm | upper|formX |    ZCIEXACT | MT_NORMALIZE | MT_MATCH_CASE
+ * CM !norm | upper       |    !ZCILOOK | MT_NORMALIZE | MT_MATCH_CASE
+ * CM !norm | upper       |     ZCILOOK | MT_NORMALIZE
+ * CM  norm | upper|formX |    !ZCILOOK | MT_NORMALIZE | MT_MATCH_CASE
+ * CM  norm | upper|formX |     ZCILOOK | MT_NORMALIZE
+ *
+ * Abbreviations:
+ *    CS = Case Sensitive, CI = Case Insensitive, CM = Case Mixed
+ *    upper = case folding set by fs type on creation (U8_TEXTPREP_TOUPPER)
+ *    formX = unicode normalization form set on fs creation
+ */
 static int
 zfs_dropname(zfs_dirlock_t *dl, znode_t *zp, znode_t *dzp, dmu_tx_t *tx,
     int flag)
@@ -812,18 +850,20 @@ zfs_dropname(zfs_dirlock_t *dl, znode_t *zp, znode_t *dzp, dmu_tx_t *tx,
 	int error;
 
 	if (ZTOZSB(zp)->z_norm) {
-		if (((ZTOZSB(zp)->z_case == ZFS_CASE_INSENSITIVE) &&
+		matchtype_t mt = MT_NORMALIZE;
+
+		if ((ZTOZSB(zp)->z_case == ZFS_CASE_INSENSITIVE &&
 		    (flag & ZCIEXACT)) ||
-		    ((ZTOZSB(zp)->z_case == ZFS_CASE_MIXED) &&
+		    (ZTOZSB(zp)->z_case == ZFS_CASE_MIXED &&
-		    !(flag & ZCILOOK)))
+		    !(flag & ZCILOOK))) {
-			error = zap_remove_norm(ZTOZSB(zp)->z_os,
+			mt |= MT_MATCH_CASE;
-			    dzp->z_id, dl->dl_name, MT_EXACT, tx);
+		}
-		else
+
-			error = zap_remove_norm(ZTOZSB(zp)->z_os,
+		error = zap_remove_norm(ZTOZSB(zp)->z_os, dzp->z_id,
-			    dzp->z_id, dl->dl_name, MT_FIRST, tx);
+		    dl->dl_name, mt, tx);
 	} else {
-		error = zap_remove(ZTOZSB(zp)->z_os,
+		error = zap_remove(ZTOZSB(zp)->z_os, dzp->z_id, dl->dl_name,
-		    dzp->z_id, dl->dl_name, tx);
+		    tx);
 	}
 
 	return (error);

module/zfs/zfs_vnops.c

@@ -18,11 +18,12 @@
  *
  * CDDL HEADER END
  */
+
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  * Copyright (c) 2012, 2015 by Delphix. All rights reserved.
- * Copyright 2014 Nexenta Systems, Inc.  All rights reserved.
  * Copyright (c) 2015 by Chunwei Chen. All rights reserved.
+ * Copyright 2017 Nexenta Systems, Inc.
  */
 
 /* Portions Copyright 2007 Jeremy Teo */
@@ -1157,7 +1158,15 @@ zfs_lookup(struct inode *dip, char *nm, struct inode **ipp, int flags,
 	zfs_sb_t *zsb = ITOZSB(dip);
 	int error = 0;
 
-	/* fast path */
+	/*
+	 * Fast path lookup, however we must skip DNLC lookup
+	 * for case folding or normalizing lookups because the
+	 * DNLC code only stores the passed in name.  This means
+	 * creating 'a' and removing 'A' on a case insensitive
+	 * file system would work, but DNLC still thinks 'a'
+	 * exists and won't let you create it again on the next
+	 * pass through fast path.
+	 */
 	if (!(flags & (LOOKUP_XATTR | FIGNORECASE))) {
 
 		if (!S_ISDIR(dip->i_mode)) {
@@ -1175,7 +1184,9 @@ zfs_lookup(struct inode *dip, char *nm, struct inode **ipp, int flags,
 			}
 			return (error);
 #ifdef HAVE_DNLC
-		} else {
+		} else if (!zdp->z_zfsvfs->z_norm &&
+		    (zdp->z_zfsvfs->z_case == ZFS_CASE_SENSITIVE)) {
+
 			vnode_t *tvp = dnlc_lookup(dvp, nm);
 
 			if (tvp) {