]> git.ozlabs.org Git - ccan/blobdiff - ccan/htable/htable.c
htable: handle v. unlikely case where entries look deleted/empty.
[ccan] / ccan / htable / htable.c
index a15c54d7958a6b779c7e3b7c7b951eb360ef0409..0371e81d0fac4837c221fcb7cb57fff5a12b8fdf 100644 (file)
@@ -1,27 +1,42 @@
+/* Licensed under LGPLv2+ - see LICENSE file for details */
 #include <ccan/htable/htable.h>
 #include <ccan/compiler/compiler.h>
-#include <stdint.h>
 #include <stdlib.h>
+#include <stdio.h>
 #include <limits.h>
 #include <stdbool.h>
 #include <assert.h>
-
-/* This means a struct htable takes at least 512 bytes / 1k (32/64 bits). */
-#define HTABLE_BASE_BITS 7
+#include <string.h>
 
 /* We use 0x1 as deleted marker. */
 #define HTABLE_DELETED (0x1)
 
-struct htable {
-       size_t (*rehash)(const void *elem, void *priv);
-       void *priv;
-       unsigned int bits;
-       size_t elems, deleted, max, max_with_deleted;
-       /* These are the bits which are the same in all pointers. */
-       uintptr_t common_mask, common_bits;
-       uintptr_t perfect_bit;
-       uintptr_t *table;
-};
+/* perfect_bitnum 63 means there's no perfect bitnum */
+#define NO_PERFECT_BIT (sizeof(uintptr_t) * CHAR_BIT - 1)
+
+static void *htable_default_alloc(struct htable *ht, size_t len)
+{
+       return calloc(len, 1);
+}
+
+static void htable_default_free(struct htable *ht, void *p)
+{
+       free(p);
+}
+
+static void *(*htable_alloc)(struct htable *, size_t) = htable_default_alloc;
+static void (*htable_free)(struct htable *, void *) = htable_default_free;
+
+void htable_set_allocator(void *(*alloc)(struct htable *, size_t len),
+                         void (*free)(struct htable *, void *p))
+{
+       if (!alloc)
+               alloc = htable_default_alloc;
+       if (!free)
+               free = htable_default_free;
+       htable_alloc = alloc;
+       htable_free = free;
+}
 
 /* We clear out the bits which are always the same, and put metadata there. */
 static inline uintptr_t get_extra_ptr_bits(const struct htable *ht,
@@ -41,9 +56,73 @@ static inline uintptr_t make_hval(const struct htable *ht,
        return ((uintptr_t)p & ~ht->common_mask) | bits;
 }
 
-static inline bool entry_is_valid(uintptr_t e)
+static inline uintptr_t *actually_valid_pair(const struct htable *ht)
+{
+       return ht->table + ((size_t)1 << ht->bits);
+}
+
+/* We have have two entries which look deleted, but we remember
+ * they are not! */
+static inline bool entry_actually_valid(const struct htable *ht, size_t off)
 {
-       return e > HTABLE_DELETED;
+       const uintptr_t *valid = actually_valid_pair(ht);
+       /* Empty table looks like this! */
+       if (valid == &ht->common_bits + 1)
+               return false;
+       return valid[0] == off || valid[1] == off;
+}
+
+/* Initialize the "actually valid" pair. */
+static inline void init_actually_valid(struct htable *ht)
+{
+       uintptr_t *valid = actually_valid_pair(ht);
+       valid[0] = valid[1] = ((size_t)1 << ht->bits);
+}
+
+/* Add to the "actually valid" pair: there can only ever be two! */
+static COLD void add_actually_valid(struct htable *ht, size_t off)
+{
+       uintptr_t *valid = actually_valid_pair(ht);
+       if (valid[0] == ((size_t)1 << ht->bits))
+               valid[0] = off;
+       else {
+               assert(valid[1] == ((size_t)1 << ht->bits));
+               valid[1] = off;
+       }
+}
+
+static COLD void del_actually_valid(struct htable *ht, size_t off)
+{
+       uintptr_t *validpair = actually_valid_pair(ht);
+       if (validpair[0] == off)
+               validpair[0] = ((size_t)1 << ht->bits);
+       else {
+               assert(validpair[1] == off);
+               validpair[1] = ((size_t)1 << ht->bits);
+       }
+}
+
+/* If this entry looks invalid, check entry_actually_valid! */
+static inline bool entry_looks_invalid(const struct htable *ht, size_t off)
+{
+       return ht->table[off] <= HTABLE_DELETED;
+}
+
+static inline bool entry_is_valid(const struct htable *ht, size_t off)
+{
+       if (!entry_looks_invalid(ht, off))
+               return true;
+       return entry_actually_valid(ht, off);
+}
+
+static inline bool entry_is_deleted(const struct htable *ht, size_t off)
+{
+       return ht->table[off] == HTABLE_DELETED && !entry_actually_valid(ht, off);
+}
+
+static inline uintptr_t ht_perfect_mask(const struct htable *ht)
+{
+       return (uintptr_t)2 << ht->perfect_bitnum;
 }
 
 static inline uintptr_t get_hash_ptr_bits(const struct htable *ht,
@@ -53,38 +132,75 @@ static inline uintptr_t get_hash_ptr_bits(const struct htable *ht,
         * end is quite expensive.  But the lower bits are redundant, so
         * we fold the value first. */
        return (hash ^ (hash >> ht->bits))
-               & ht->common_mask & ~ht->perfect_bit;
-}
-
-struct htable *htable_new(size_t (*rehash)(const void *elem, void *priv),
-                         void *priv)
-{
-       struct htable *ht = malloc(sizeof(struct htable));
-       if (ht) {
-               ht->bits = HTABLE_BASE_BITS;
-               ht->rehash = rehash;
-               ht->priv = priv;
-               ht->elems = 0;
-               ht->deleted = 0;
-               ht->max = ((size_t)1 << ht->bits) * 3 / 4;
-               ht->max_with_deleted = ((size_t)1 << ht->bits) * 9 / 10;
-               /* This guarantees we enter update_common first add. */
-               ht->common_mask = -1;
-               ht->common_bits = 0;
-               ht->perfect_bit = 0;
-               ht->table = calloc(1 << ht->bits, sizeof(uintptr_t));
-               if (!ht->table) {
-                       free(ht);
-                       ht = NULL;
-               }
+               & ht->common_mask & ~ht_perfect_mask(ht);
+}
+
+void htable_init(struct htable *ht,
+                size_t (*rehash)(const void *elem, void *priv), void *priv)
+{
+       struct htable empty = HTABLE_INITIALIZER(empty, NULL, NULL);
+       *ht = empty;
+       ht->rehash = rehash;
+       ht->priv = priv;
+       ht->table = &ht->common_bits;
+}
+
+static inline size_t ht_max(const struct htable *ht)
+{
+       return ((size_t)3 << ht->bits) / 4;
+}
+
+static inline size_t ht_max_with_deleted(const struct htable *ht)
+{
+       return ((size_t)9 << ht->bits) / 10;
+}
+
+/* Includes the two trailing "not-deleted" entries */
+static size_t htable_alloc_size(size_t bits)
+{
+       return (sizeof(size_t) << bits) + 2 * sizeof(size_t);
+}
+
+bool htable_init_sized(struct htable *ht,
+                      size_t (*rehash)(const void *, void *),
+                      void *priv, size_t expect)
+{
+       htable_init(ht, rehash, priv);
+
+       /* Don't go insane with sizing. */
+       for (ht->bits = 1; ((size_t)3 << ht->bits) / 4 < expect; ht->bits++) {
+               if (ht->bits == 30)
+                       break;
        }
-       return ht;
+
+       ht->table = htable_alloc(ht, htable_alloc_size(ht->bits));
+       if (!ht->table) {
+               ht->table = &ht->common_bits;
+               return false;
+       }
+       init_actually_valid(ht);
+       (void)htable_debug(ht, HTABLE_LOC);
+       return true;
+}
+       
+void htable_clear(struct htable *ht)
+{
+       if (ht->table != &ht->common_bits)
+               htable_free(ht, (void *)ht->table);
+       htable_init(ht, ht->rehash, ht->priv);
 }
 
-void htable_free(const struct htable *ht)
+bool htable_copy_(struct htable *dst, const struct htable *src)
 {
-       free((void *)ht->table);
-       free((void *)ht);
+       uintptr_t *htable = htable_alloc(dst, htable_alloc_size(src->bits));
+
+       if (!htable)
+               return false;
+
+       *dst = *src;
+       dst->table = htable;
+       memcpy(dst->table, src->table, htable_alloc_size(src->bits));
+       return true;
 }
 
 static size_t hash_bucket(const struct htable *ht, size_t h)
@@ -97,8 +213,8 @@ static void *htable_val(const struct htable *ht,
 {
        uintptr_t h2 = get_hash_ptr_bits(ht, hash) | perfect;
 
-       while (ht->table[i->off]) {
-               if (ht->table[i->off] != HTABLE_DELETED) {
+       while (ht->table[i->off] || entry_actually_valid(ht, i->off)) {
+               if (!entry_is_deleted(ht, i->off)) {
                        if (get_extra_ptr_bits(ht, ht->table[i->off]) == h2)
                                return get_raw_ptr(ht, ht->table[i->off]);
                }
@@ -108,112 +224,143 @@ static void *htable_val(const struct htable *ht,
        return NULL;
 }
 
-void *htable_firstval(const struct htable *ht,
-                     struct htable_iter *i, size_t hash)
+void *htable_firstval_(const struct htable *ht,
+                      struct htable_iter *i, size_t hash)
 {
        i->off = hash_bucket(ht, hash);
-       return htable_val(ht, i, hash, ht->perfect_bit);
+       return htable_val(ht, i, hash, ht_perfect_mask(ht));
 }
 
-void *htable_nextval(const struct htable *ht,
-                    struct htable_iter *i, size_t hash)
+void *htable_nextval_(const struct htable *ht,
+                     struct htable_iter *i, size_t hash)
 {
        i->off = (i->off + 1) & ((1 << ht->bits)-1);
        return htable_val(ht, i, hash, 0);
 }
 
-void *htable_first(const struct htable *ht, struct htable_iter *i)
+void *htable_first_(const struct htable *ht, struct htable_iter *i)
 {
        for (i->off = 0; i->off < (size_t)1 << ht->bits; i->off++) {
-               if (entry_is_valid(ht->table[i->off]))
+               if (entry_is_valid(ht, i->off))
                        return get_raw_ptr(ht, ht->table[i->off]);
        }
        return NULL;
 }
 
-void *htable_next(const struct htable *ht, struct htable_iter *i)
+void *htable_next_(const struct htable *ht, struct htable_iter *i)
 {
        for (i->off++; i->off < (size_t)1 << ht->bits; i->off++) {
-               if (entry_is_valid(ht->table[i->off]))
+               if (entry_is_valid(ht, i->off))
                        return get_raw_ptr(ht, ht->table[i->off]);
        }
        return NULL;
 }
 
+void *htable_prev_(const struct htable *ht, struct htable_iter *i)
+{
+       for (;;) {
+               if (!i->off)
+                       return NULL;
+               i->off--;
+               if (entry_is_valid(ht, i->off))
+                       return get_raw_ptr(ht, ht->table[i->off]);
+       }
+}
+
 /* This does not expand the hash table, that's up to caller. */
 static void ht_add(struct htable *ht, const void *new, size_t h)
 {
        size_t i;
-       uintptr_t perfect = ht->perfect_bit;
+       uintptr_t perfect = ht_perfect_mask(ht);
 
        i = hash_bucket(ht, h);
 
-       while (entry_is_valid(ht->table[i])) {
+       while (entry_is_valid(ht, i)) {
                perfect = 0;
                i = (i + 1) & ((1 << ht->bits)-1);
        }
        ht->table[i] = make_hval(ht, new, get_hash_ptr_bits(ht, h)|perfect);
+
+       /* If it looks invalid, add it to exceptions */
+       if (ht->table[i] <= HTABLE_DELETED)
+               add_actually_valid(ht, i);
 }
 
 static COLD bool double_table(struct htable *ht)
 {
-       unsigned int i;
-       size_t oldnum = (size_t)1 << ht->bits;
-       uintptr_t *oldtable, e;
+       size_t i;
+       struct htable oldht = *ht;
 
-       oldtable = ht->table;
-       ht->table = calloc(1 << (ht->bits+1), sizeof(size_t));
+       ht->table = htable_alloc(ht, htable_alloc_size(ht->bits+1));
        if (!ht->table) {
-               ht->table = oldtable;
+               ht->table = oldht.table;
                return false;
        }
        ht->bits++;
-       ht->max *= 2;
-       ht->max_with_deleted *= 2;
+       init_actually_valid(ht);
 
        /* If we lost our "perfect bit", get it back now. */
-       if (!ht->perfect_bit && ht->common_mask) {
+       if (ht->perfect_bitnum == NO_PERFECT_BIT && ht->common_mask) {
                for (i = 0; i < sizeof(ht->common_mask) * CHAR_BIT; i++) {
-                       if (ht->common_mask & ((size_t)1 << i)) {
-                               ht->perfect_bit = (size_t)1 << i;
+                       if (ht->common_mask & ((size_t)2 << i)) {
+                               ht->perfect_bitnum = i;
                                break;
                        }
                }
        }
 
-       for (i = 0; i < oldnum; i++) {
-               if (entry_is_valid(e = oldtable[i])) {
-                       void *p = get_raw_ptr(ht, e);
-                       ht_add(ht, p, ht->rehash(p, ht->priv));
+       if (oldht.table != &ht->common_bits) {
+               for (i = 0; i < (size_t)1 << oldht.bits; i++) {
+                       if (entry_is_valid(&oldht, i)) {
+                               void *p = get_raw_ptr(&oldht, oldht.table[i]);
+                               ht_add(ht, p, ht->rehash(p, ht->priv));
+                       }
                }
+               /* Pass ht here to callback: oldht is an internal figment */
+               htable_free(ht, oldht.table);
        }
        ht->deleted = 0;
-       free(oldtable);
+
+       (void)htable_debug(ht, HTABLE_LOC);
        return true;
 }
 
 static COLD void rehash_table(struct htable *ht)
 {
        size_t start, i;
-       uintptr_t e;
+       uintptr_t e, perfect = ht_perfect_mask(ht);
+       uintptr_t *validpair = actually_valid_pair(ht);
 
        /* Beware wrap cases: we need to start from first empty bucket. */
        for (start = 0; ht->table[start]; start++);
 
        for (i = 0; i < (size_t)1 << ht->bits; i++) {
                size_t h = (i + start) & ((1 << ht->bits)-1);
+               uintptr_t *actually = NULL;
                e = ht->table[h];
-               if (!e)
-                       continue;
-               if (e == HTABLE_DELETED)
-                       ht->table[h] = 0;
-               else if (!(e & ht->perfect_bit)) {
+               if (e <= HTABLE_DELETED) {
+                       /* If it's actually valid, remember in case we move it! */
+                       if (validpair[0] == h) {
+                               actually = &validpair[0];
+                       } else if (validpair[1] == h) {
+                               actually = &validpair[1];
+                       } else {
+                               ht->table[h] = 0;
+                               continue;
+                       }
+               }
+
+               if (!(e & perfect)) {
                        void *p = get_raw_ptr(ht, e);
                        ht->table[h] = 0;
+                       /* Clear actuallyvalid, let ht_add refill */
+                       if (actually)
+                               *actually = ((size_t)1 << ht->bits);
                        ht_add(ht, p, ht->rehash(p, ht->priv));
                }
        }
        ht->deleted = 0;
+       (void)htable_debug(ht, HTABLE_LOC);
 }
 
 /* We stole some bits, now we need to put them back... */
@@ -224,8 +371,9 @@ static COLD void update_common(struct htable *ht, const void *p)
 
        if (ht->elems == 0) {
                ht->common_mask = -1;
-               ht->common_bits = (uintptr_t)p;
-               ht->perfect_bit = 1;
+               ht->common_bits = ((uintptr_t)p & ht->common_mask);
+               ht->perfect_bitnum = 0;
+               (void)htable_debug(ht, HTABLE_LOC);
                return;
        }
 
@@ -236,25 +384,31 @@ static COLD void update_common(struct htable *ht, const void *p)
        bitsdiff = ht->common_bits & maskdiff;
 
        for (i = 0; i < (size_t)1 << ht->bits; i++) {
-               if (!entry_is_valid(ht->table[i]))
+               if (!entry_is_valid(ht, i))
                        continue;
                /* Clear the bits no longer in the mask, set them as
                 * expected. */
                ht->table[i] &= ~maskdiff;
                ht->table[i] |= bitsdiff;
+
+               /* Make sure it's not newly falsely invalid */
+               if (ht->table[i] <= HTABLE_DELETED && !entry_actually_valid(ht, i))
+                       add_actually_valid(ht, i);
        }
 
        /* Take away those bits from our mask, bits and perfect bit. */
        ht->common_mask &= ~maskdiff;
        ht->common_bits &= ~maskdiff;
-       ht->perfect_bit &= ~maskdiff;
+       if (ht_perfect_mask(ht) & maskdiff)
+               ht->perfect_bitnum = NO_PERFECT_BIT;
+       (void)htable_debug(ht, HTABLE_LOC);
 }
 
-bool htable_add(struct htable *ht, size_t hash, const void *p)
+bool htable_add_(struct htable *ht, size_t hash, const void *p)
 {
-       if (ht->elems+1 > ht->max && !double_table(ht))
+       if (ht->elems+1 > ht_max(ht) && !double_table(ht))
                return false;
-       if (ht->elems+1 + ht->deleted > ht->max_with_deleted)
+       if (ht->elems+1 + ht->deleted > ht_max_with_deleted(ht))
                rehash_table(ht);
        assert(p);
        if (((uintptr_t)p & ht->common_mask) != ht->common_bits)
@@ -265,7 +419,7 @@ bool htable_add(struct htable *ht, size_t hash, const void *p)
        return true;
 }
 
-bool htable_del(struct htable *ht, size_t h, const void *p)
+bool htable_del_(struct htable *ht, size_t h, const void *p)
 {
        struct htable_iter i;
        void *c;
@@ -279,12 +433,109 @@ bool htable_del(struct htable *ht, size_t h, const void *p)
        return false;
 }
 
-void htable_delval(struct htable *ht, struct htable_iter *i)
+void htable_delval_(struct htable *ht, struct htable_iter *i)
 {
        assert(i->off < (size_t)1 << ht->bits);
-       assert(entry_is_valid(ht->table[i->off]));
+
+       if (entry_looks_invalid(ht, i->off))
+               del_actually_valid(ht, i->off);
 
        ht->elems--;
        ht->table[i->off] = HTABLE_DELETED;
        ht->deleted++;
 }
+
+void *htable_pick_(const struct htable *ht, size_t seed, struct htable_iter *i)
+{
+       void *e;
+       struct htable_iter unwanted;
+
+       if (!i)
+               i = &unwanted;
+       i->off = seed % ((size_t)1 << ht->bits);
+       e = htable_next(ht, i);
+       if (!e)
+               e = htable_first(ht, i);
+       return e;
+}
+
+struct htable *htable_check(const struct htable *ht, const char *abortstr)
+{
+       void *p;
+       struct htable_iter i;
+       size_t n = 0;
+       const uintptr_t *validpair = actually_valid_pair(ht);
+
+       /* Use non-DEBUG versions here, to avoid infinite recursion with
+        * CCAN_HTABLE_DEBUG! */
+       for (p = htable_first_(ht, &i); p; p = htable_next_(ht, &i)) {
+               struct htable_iter i2;
+               void *c;
+               size_t h = ht->rehash(p, ht->priv);
+               bool found = false;
+
+               n++;
+
+               /* Open-code htable_get to avoid CCAN_HTABLE_DEBUG */
+               for (c = htable_firstval_(ht, &i2, h);
+                    c;
+                    c = htable_nextval_(ht, &i2, h)) {
+                       if (c == p) {
+                               found = true;
+                               break;
+                       }
+               }
+
+               if (!found) {
+                       if (abortstr) {
+                               fprintf(stderr,
+                                       "%s: element %p in position %zu"
+                                       " cannot find itself\n",
+                                       abortstr, p, i.off);
+                               abort();
+                       }
+                       return NULL;
+               }
+       }
+       if (n != ht->elems) {
+               if (abortstr) {
+                       fprintf(stderr,
+                               "%s: found %zu elems, expected %zu\n",
+                               abortstr, n, ht->elems);
+                       abort();
+               }
+               return NULL;
+       }
+
+       /* Check validpair does actually override invalid-looking entries! */
+       if (ht->table != &ht->common_bits) {
+               size_t i;
+               for (i = 0; i < 2; i++) {
+                       if (validpair[i] == ((size_t)1 << ht->bits))
+                               continue;
+                       if (validpair[i] > ((size_t)1 << ht->bits)) {
+                               if (abortstr) {
+                                       fprintf(stderr,
+                                               "%s: validpair[%zu] points at %zu"
+                                               " which is out of bounds\n",
+                                               abortstr, i, validpair[i]);
+                                       abort();
+                               }
+                               return NULL;
+                       }
+                       if (entry_looks_invalid(ht, validpair[i]))
+                               continue;
+
+                       if (abortstr) {
+                               fprintf(stderr,
+                                       "%s: validpair[%zu] points at %zu"
+                                       " which seems valid\n",
+                                       abortstr, i, validpair[i]);
+                               abort();
+                       }
+                       return NULL;
+               }
+       }
+
+       return (struct htable *)ht;
+}