FT-307 Add a 'memcmp magic' API, where users can specify that a

particular value for the first byte of a key means it may be compared
with memcmp and a length check.

FT-50 Implement toku_builtin_cmp_fun with memcmp and a length check

Author: esmet

buildheader/make_tdb.cc

@@ -545,6 +545,7 @@ static void print_db_struct (void) {
 			 "int (*change_fanout)(DB *db, uint32_t fanout)",
 			 "int (*get_fanout)(DB *db, uint32_t *fanout)",
 			 "int (*set_fanout)(DB *db, uint32_t fanout)",
+			 "int (*set_memcmp_magic)(DB *db, uint8_t magic)",
 			 "int (*set_indexer)(DB*, DB_INDEXER*)",
 			 "void (*get_indexer)(DB*, DB_INDEXER**)",
 			 "int (*verify_with_progress)(DB *, int (*progress_callback)(void *progress_extra, float progress), void *progress_extra, int verbose, int keep_going)",

ft/comparator.h

@@ -110,12 +110,16 @@ namespace toku {
     // that points may be positive or negative infinity.
 
     class comparator {
-    public:
-        void create(ft_compare_func cmp, DESCRIPTOR desc) {
+        void init(ft_compare_func cmp, DESCRIPTOR desc, uint8_t memcmp_magic) {
             _cmp = cmp;
-            XCALLOC(_fake_db);
             _fake_db->cmp_descriptor = desc;
-            _builtin = _cmp == &toku_builtin_compare_fun;
+            _memcmp_magic = memcmp_magic;
+        }
+
+    public:
+        void create(ft_compare_func cmp, DESCRIPTOR desc, uint8_t memcmp_magic = 0) {
+            XCALLOC(_fake_db);
+            init(cmp, desc, memcmp_magic);
         }
 
         // inherit the attributes of another comparator, but keep our own
@@ -124,9 +128,7 @@ namespace toku {
             invariant_notnull(_fake_db);
             invariant_notnull(cmp._cmp);
             invariant_notnull(cmp._fake_db);
-            _cmp = cmp._cmp;
-            _fake_db->cmp_descriptor = cmp._fake_db->cmp_descriptor;
-            _builtin = cmp._builtin;
+            init(cmp._cmp, cmp._fake_db->cmp_descriptor, cmp._memcmp_magic);
         }
 
         // like inherit, but doesn't require that the this comparator
@@ -148,14 +150,24 @@ namespace toku {
             return _cmp;
         }
 
+        uint8_t get_memcmp_magic() const {
+            return _memcmp_magic;
+        }
+
         bool valid() const {
             return _cmp != nullptr;
         }
 
+        inline bool dbt_has_memcmp_magic(const DBT *dbt) const {
+            return *reinterpret_cast<const char *>(dbt->data) == _memcmp_magic;
+        }
+
         int operator()(const DBT *a, const DBT *b) const {
-            if (__builtin_expect(!!(toku_dbt_is_infinite(a) || toku_dbt_is_infinite(b)), 0)) {
+            if (__builtin_expect(toku_dbt_is_infinite(a) || toku_dbt_is_infinite(b), 0)) {
                 return toku_dbt_infinite_compare(a, b);
-            } else if (_builtin) {
+            } else if (_memcmp_magic && dbt_has_memcmp_magic(a)
+                       // At this point we expect b to also have the memcmp magic
+                       && __builtin_expect(dbt_has_memcmp_magic(b), 1)) {
                 return toku_builtin_compare_fun(nullptr, a, b);
             } else {
                 // yikes, const sadness here
@@ -166,7 +178,7 @@ namespace toku {
     private:
         DB *_fake_db;
         ft_compare_func _cmp;
-        bool _builtin;
+        uint8_t _memcmp_magic;
     };
 
 } /* namespace toku */

ft/ft-internal.h

@@ -278,6 +278,7 @@ struct ft_options {
     enum toku_compression_method compression_method;
     unsigned int fanout;
     unsigned int flags;
+    uint8_t memcmp_magic;
     ft_compare_func compare_fun;
     ft_update_func update_fun;
 };

ft/ft-ops.cc

@@ -2887,6 +2887,17 @@ toku_ft_handle_get_fanout(FT_HANDLE ft_handle, unsigned int *fanout)
         *fanout = ft_handle->options.fanout;
     }
 }
+
+void toku_ft_handle_set_memcmp_magic(FT_HANDLE ft_handle, uint8_t magic) {
+    invariant(magic != 0);
+    if (ft_handle->ft) {
+        // handle is already open, application bug if memcmp magic changes
+        invariant(ft_handle->ft->cmp.get_memcmp_magic() == magic);
+    } else {
+        ft_handle->options.memcmp_magic = magic;
+    }
+}
+
 static int
 verify_builtin_comparisons_consistent(FT_HANDLE t, uint32_t flags) {
     if ((flags & TOKU_DB_KEYCMP_BUILTIN) && (t->options.compare_fun != toku_builtin_compare_fun))
@@ -2955,6 +2966,7 @@ toku_ft_handle_inherit_options(FT_HANDLE t, FT ft) {
         .compression_method = ft->h->compression_method,
         .fanout = ft->h->fanout,
         .flags = ft->h->flags,
+        .memcmp_magic = ft->cmp.get_memcmp_magic(),
         .compare_fun = ft->cmp.get_compare_func(),
         .update_fun = ft->update_fun
     };
@@ -4721,32 +4733,23 @@ int toku_ft_strerror_r(int error, char *buf, size_t buflen)
     }
 }
 
-// when a and b are chars, return a-b is safe here because return type is int.  No over/underflow possible.
-int toku_keycompare (const void *key1, uint32_t key1len, const void *key2, uint32_t key2len) {
-    int comparelen = key1len<key2len ? key1len : key2len;
-    const unsigned char *k1;
-    const unsigned char *k2;
-    for (CAST_FROM_VOIDP(k1, key1), CAST_FROM_VOIDP(k2, key2);
-	 comparelen>4;
-	 k1+=4, k2+=4, comparelen-=4) {
-	{ int v1=k1[0], v2=k2[0]; if (v1!=v2) return v1-v2; }
-	{ int v1=k1[1], v2=k2[1]; if (v1!=v2) return v1-v2; }
-	{ int v1=k1[2], v2=k2[2]; if (v1!=v2) return v1-v2; }
-	{ int v1=k1[3], v2=k2[3]; if (v1!=v2) return v1-v2; }
-    }
-    for (;
-	 comparelen>0;
-	 k1++, k2++, comparelen--) {
-	if (*k1 != *k2) {
-	    return (int)*k1-(int)*k2;
-	}
-    }
-    if (key1len<key2len) return -1;
-    if (key1len>key2len) return 1;
-    return 0;
+int toku_keycompare(const void *key1, uint32_t key1len, const void *key2, uint32_t key2len) {
+    int comparelen = key1len < key2len ? key1len : key2len;
+    int c = memcmp(key1, key2, comparelen);
+    if (__builtin_expect(c != 0, 1)) {
+        return c;
+    } else {
+        if (key1len < key2len) {
+            return -1;
+        } else if (key1len > key2len) { 
+            return 1;
+        } else {
+            return 0;
+        }
+    }
 }
 
-int toku_builtin_compare_fun (DB *db __attribute__((__unused__)), const DBT *a, const DBT*b) {
+int toku_builtin_compare_fun(DB *db __attribute__((__unused__)), const DBT *a, const DBT*b) {
     return toku_keycompare(a->data, a->size, b->data, b->size);
 }
 

ft/ft-ops.h

@@ -126,6 +126,7 @@ void toku_ft_handle_set_compression_method(FT_HANDLE, enum toku_compression_meth
 void toku_ft_handle_get_compression_method(FT_HANDLE, enum toku_compression_method *);
 void toku_ft_handle_set_fanout(FT_HANDLE, unsigned int fanout);
 void toku_ft_handle_get_fanout(FT_HANDLE, unsigned int *fanout);
+void toku_ft_handle_set_memcmp_magic(FT_HANDLE, uint8_t magic);
 
 void toku_ft_set_bt_compare(FT_HANDLE ft_handle, ft_compare_func cmp_func);
 const toku::comparator &toku_ft_get_comparator(FT_HANDLE ft_handle);

ft/ft.cc

@@ -384,7 +384,7 @@ static void ft_init(FT ft, FT_OPTIONS options, CACHEFILE cf) {
     toku_list_init(&ft->live_ft_handles);
 
     // intuitively, the comparator points to the FT's cmp descriptor
-    ft->cmp.create(options->compare_fun, &ft->cmp_descriptor);
+    ft->cmp.create(options->compare_fun, &ft->cmp_descriptor, options->memcmp_magic);
     ft->update_fun = options->update_fun;
 
     if (ft->cf != NULL) {
@@ -486,7 +486,7 @@ int toku_read_ft_and_store_in_cachefile (FT_HANDLE ft_handle, CACHEFILE cf, LSN
 
     invariant_notnull(ft);
     // intuitively, the comparator points to the FT's cmp descriptor
-    ft->cmp.create(ft_handle->options.compare_fun, &ft->cmp_descriptor);
+    ft->cmp.create(ft_handle->options.compare_fun, &ft->cmp_descriptor, ft_handle->options.memcmp_magic);
     ft->update_fun = ft_handle->options.update_fun;
     ft->cf = cf;
     toku_cachefile_set_userdata(cf,
@@ -611,6 +611,7 @@ toku_ft_init(FT ft,
         .compression_method = compression_method,
         .fanout = fanout,
         .flags = 0,
+        .memcmp_magic = 0,
         .compare_fun = NULL,
         .update_fun = NULL
     };

src/ydb.cc

@@ -465,27 +465,6 @@ needs_recovery (DB_ENV *env) {
 
 static int toku_env_txn_checkpoint(DB_ENV * env, uint32_t kbyte, uint32_t min, uint32_t flags);
 
-// Instruct db to use the default (built-in) key comparison function
-// by setting the flag bits in the db and ft structs
-static int
-db_use_builtin_key_cmp(DB *db) {
-    HANDLE_PANICKED_DB(db);
-    int r = 0;
-    if (db_opened(db))
-        r = toku_ydb_do_error(db->dbenv, EINVAL, "Comparison functions cannot be set after DB open.\n");
-    else if (db->i->key_compare_was_set)
-        r = toku_ydb_do_error(db->dbenv, EINVAL, "Key comparison function already set.\n");
-    else {
-        uint32_t tflags;
-        toku_ft_get_flags(db->i->ft_handle, &tflags);
-
-        tflags |= TOKU_DB_KEYCMP_BUILTIN;
-        toku_ft_set_flags(db->i->ft_handle, tflags);
-        db->i->key_compare_was_set = true;
-    }
-    return r;
-}
-
 // Keys used in persistent environment dictionary:
 // Following keys added in version 12
 static const char * orig_env_ver_key = "original_version";
@@ -1025,7 +1004,7 @@ env_open(DB_ENV * env, const char *home, uint32_t flags, int mode) {
     {
         r = toku_db_create(&env->i->persistent_environment, env, 0);
         assert_zero(r);
-        r = db_use_builtin_key_cmp(env->i->persistent_environment);
+        r = toku_db_use_builtin_key_cmp(env->i->persistent_environment);
         assert_zero(r);
         r = toku_db_open_iname(env->i->persistent_environment, txn, toku_product_name_strings.environmentdictionary, DB_CREATE, mode);
         if (r != 0) {
@@ -1063,7 +1042,7 @@ env_open(DB_ENV * env, const char *home, uint32_t flags, int mode) {
     {
         r = toku_db_create(&env->i->directory, env, 0);
         assert_zero(r);
-        r = db_use_builtin_key_cmp(env->i->directory);
+        r = toku_db_use_builtin_key_cmp(env->i->directory);
         assert_zero(r);
         r = toku_db_open_iname(env->i->directory, txn, toku_product_name_strings.fileopsdirectory, DB_CREATE, mode);
         if (r != 0) {

src/ydb_db.cc

@@ -431,8 +431,27 @@ void toku_db_lt_on_destroy_callback(toku::locktree *lt) {
     toku_ft_handle_close(ft_handle);
 }
 
-int 
-toku_db_open_iname(DB * db, DB_TXN * txn, const char *iname_in_env, uint32_t flags, int mode) {
+// Instruct db to use the default (built-in) key comparison function
+// by setting the flag bits in the db and ft structs
+int toku_db_use_builtin_key_cmp(DB *db) {
+    HANDLE_PANICKED_DB(db);
+    int r = 0;
+    if (db_opened(db)) {
+        r = toku_ydb_do_error(db->dbenv, EINVAL, "Comparison functions cannot be set after DB open.\n");
+    } else if (db->i->key_compare_was_set) {
+        r = toku_ydb_do_error(db->dbenv, EINVAL, "Key comparison function already set.\n");
+    } else {
+        uint32_t tflags;
+        toku_ft_get_flags(db->i->ft_handle, &tflags);
+
+        tflags |= TOKU_DB_KEYCMP_BUILTIN;
+        toku_ft_set_flags(db->i->ft_handle, tflags);
+        db->i->key_compare_was_set = true;
+    }
+    return r;
+}
+
+int toku_db_open_iname(DB * db, DB_TXN * txn, const char *iname_in_env, uint32_t flags, int mode) {
     //Set comparison functions if not yet set.
     HANDLE_READ_ONLY_TXN(txn);
     if (!db->i->key_compare_was_set && db->dbenv->i->bt_compare) {
@@ -704,6 +723,19 @@ toku_db_get_fanout(DB *db, unsigned int *fanout) {
     return 0;
 }
 
+static int
+toku_db_set_memcmp_magic(DB *db, uint8_t magic) {
+    HANDLE_PANICKED_DB(db);
+    if (db_opened(db)) {
+        return EINVAL;
+    }
+    if (magic == 0) {
+        return EINVAL;
+    }
+    toku_ft_handle_set_memcmp_magic(db->i->ft_handle, magic);
+    return 0;
+}
+
 static int
 toku_db_get_fractal_tree_info64(DB *db, uint64_t *num_blocks_allocated, uint64_t *num_blocks_in_use, uint64_t *size_allocated, uint64_t *size_in_use) {
     HANDLE_PANICKED_DB(db);
@@ -1101,6 +1133,7 @@ toku_db_create(DB ** db, DB_ENV * env, uint32_t flags) {
     USDB(change_compression_method);
     USDB(set_fanout);
     USDB(get_fanout);
+    USDB(set_memcmp_magic);
     USDB(change_fanout);
     USDB(set_flags);
     USDB(get_flags);

src/ydb_db.h

@@ -131,6 +131,7 @@ static inline const toku::comparator &toku_db_get_comparator(DB *db) {
     return toku_ft_get_comparator(db->i->ft_handle);
 }
 
+int toku_db_use_builtin_key_cmp(DB *db);
 int toku_db_pre_acquire_fileops_lock(DB *db, DB_TXN *txn);
 int toku_db_open_iname(DB * db, DB_TXN * txn, const char *iname, uint32_t flags, int mode);
 int toku_db_pre_acquire_table_lock(DB *db, DB_TXN *txn);