tdb: Make tdb_find circular-safe
[samba.git] / lib / tdb / common / freelist.c
index 6f8f812d416a73e1916b921b185f05280026c4a5..86fac2ff078ba87c82969e7731f58ced7f713e8a 100644 (file)
@@ -97,10 +97,217 @@ static int update_tailer(struct tdb_context *tdb, tdb_off_t offset,
                         &totalsize);
 }
 
-/* Add an element into the freelist. Merge adjacent records if
-   necessary. */
+/**
+ * Read the record directly on the left.
+ * Fail if there is no record on the left.
+ */
+static int read_record_on_left(struct tdb_context *tdb, tdb_off_t rec_ptr,
+                              tdb_off_t *left_p,
+                              struct tdb_record *left_r)
+{
+       tdb_off_t left_ptr;
+       tdb_off_t left_size;
+       struct tdb_record left_rec;
+       int ret;
+
+       left_ptr = rec_ptr - sizeof(tdb_off_t);
+
+       if (left_ptr <= TDB_DATA_START(tdb->hash_size)) {
+               /* no record on the left */
+               return -1;
+       }
+
+       /* Read in tailer and jump back to header */
+       ret = tdb_ofs_read(tdb, left_ptr, &left_size);
+       if (ret == -1) {
+               TDB_LOG((tdb, TDB_DEBUG_FATAL,
+                       "tdb_free: left offset read failed at %u\n", left_ptr));
+               return -1;
+       }
+
+       /* it could be uninitialised data */
+       if (left_size == 0 || left_size == TDB_PAD_U32) {
+               return -1;
+       }
+
+       if (left_size > rec_ptr) {
+               return -1;
+       }
+
+       left_ptr = rec_ptr - left_size;
+
+       if (left_ptr < TDB_DATA_START(tdb->hash_size)) {
+               return -1;
+       }
+
+       /* Now read in the left record */
+       ret = tdb->methods->tdb_read(tdb, left_ptr, &left_rec,
+                                    sizeof(left_rec), DOCONV());
+       if (ret == -1) {
+               TDB_LOG((tdb, TDB_DEBUG_FATAL,
+                        "tdb_free: left read failed at %u (%u)\n",
+                        left_ptr, left_size));
+               return -1;
+       }
+
+       *left_p = left_ptr;
+       *left_r = left_rec;
+
+       return 0;
+}
+
+/**
+ * Merge new freelist record with the direct left neighbour.
+ * This assumes that left_rec represents the record
+ * directly to the left of right_rec and that this is
+ * a freelist record.
+ */
+static int merge_with_left_record(struct tdb_context *tdb,
+                                 tdb_off_t left_ptr,
+                                 struct tdb_record *left_rec,
+                                 struct tdb_record *right_rec)
+{
+       int ret;
+
+       left_rec->rec_len += sizeof(*right_rec) + right_rec->rec_len;
+
+       ret = tdb_rec_write(tdb, left_ptr, left_rec);
+       if (ret == -1) {
+               TDB_LOG((tdb, TDB_DEBUG_FATAL,
+                        "merge_with_left_record: update_left failed at %u\n",
+                        left_ptr));
+               return -1;
+       }
+
+       ret = update_tailer(tdb, left_ptr, left_rec);
+       if (ret == -1) {
+               TDB_LOG((tdb, TDB_DEBUG_FATAL,
+                        "merge_with_left_record: update_tailer failed at %u\n",
+                        left_ptr));
+               return -1;
+       }
+
+       return 0;
+}
+
+/**
+ * Check whether the record left of a given freelist record is
+ * also a freelist record, and if so, merge the two records.
+ *
+ * Return code:
+ *  -1 upon error
+ *   0 if left was not a free record
+ *   1 if left was free and successfully merged.
+ *
+ * The currend record is handed in with pointer and fully read record.
+ *
+ * The left record pointer and struct can be retrieved as result
+ * in lp and lr;
+ */
+static int check_merge_with_left_record(struct tdb_context *tdb,
+                                       tdb_off_t rec_ptr,
+                                       struct tdb_record *rec,
+                                       tdb_off_t *lp,
+                                       struct tdb_record *lr)
+{
+       tdb_off_t left_ptr;
+       struct tdb_record left_rec;
+       int ret;
+
+       ret = read_record_on_left(tdb, rec_ptr, &left_ptr, &left_rec);
+       if (ret != 0) {
+               return 0;
+       }
+
+       if (left_rec.magic != TDB_FREE_MAGIC) {
+               return 0;
+       }
+
+       /* It's free - expand to include it. */
+       ret = merge_with_left_record(tdb, left_ptr, &left_rec, rec);
+       if (ret != 0) {
+               return -1;
+       }
+
+       if (lp != NULL) {
+               *lp = left_ptr;
+       }
+
+       if (lr != NULL) {
+               *lr = left_rec;
+       }
+
+       return 1;
+}
+
+/**
+ * Check whether the record left of a given freelist record is
+ * also a freelist record, and if so, merge the two records.
+ *
+ * Return code:
+ *  -1 upon error
+ *   0 if left was not a free record
+ *   1 if left was free and successfully merged.
+ *
+ * In this variant, the input record is specified just as the pointer
+ * and is read from the database if needed.
+ *
+ * next_ptr will contain the original record's next pointer after
+ * successful merging (which will be lost after merging), so that
+ * the caller can update the last pointer.
+ */
+static int check_merge_ptr_with_left_record(struct tdb_context *tdb,
+                                           tdb_off_t rec_ptr,
+                                           tdb_off_t *next_ptr)
+{
+       tdb_off_t left_ptr;
+       struct tdb_record rec, left_rec;
+       int ret;
+
+       ret = read_record_on_left(tdb, rec_ptr, &left_ptr, &left_rec);
+       if (ret != 0) {
+               return 0;
+       }
+
+       if (left_rec.magic != TDB_FREE_MAGIC) {
+               return 0;
+       }
+
+       /* It's free - expand to include it. */
+
+       ret = tdb->methods->tdb_read(tdb, rec_ptr, &rec,
+                                    sizeof(rec), DOCONV());
+       if (ret != 0) {
+               return -1;
+       }
+
+       ret = merge_with_left_record(tdb, left_ptr, &left_rec, &rec);
+       if (ret != 0) {
+               return -1;
+       }
+
+       if (next_ptr != NULL) {
+               *next_ptr = rec.next;
+       }
+
+       return 1;
+}
+
+/**
+ * Add an element into the freelist.
+ *
+ * We merge the new record into the left record if it is also a
+ * free record, but not with the right one. This makes the
+ * operation O(1) instead of O(n): merging with the right record
+ * requires a traverse of the freelist to find the previous
+ * record in the free list.
+ *
+ * This prevents db traverses from being O(n^2) after a lot of deletes.
+ */
 int tdb_free(struct tdb_context *tdb, tdb_off_t offset, struct tdb_record *rec)
 {
+       int ret;
+
        /* Allocation and tailer lock */
        if (tdb_lock(tdb, -1, F_WRLCK) != 0)
                return -1;
@@ -138,58 +345,17 @@ int tdb_free(struct tdb_context *tdb, tdb_off_t offset, struct tdb_record *rec)
 left:
 #endif
 
-       /* Look left */
-       if (offset - sizeof(tdb_off_t) > TDB_DATA_START(tdb->hash_size)) {
-               tdb_off_t left = offset - sizeof(tdb_off_t);
-               struct tdb_record l;
-               tdb_off_t leftsize;
-
-               /* Read in tailer and jump back to header */
-               if (tdb_ofs_read(tdb, left, &leftsize) == -1) {
-                       TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: left offset read failed at %u\n", left));
-                       goto update;
-               }
-
-               /* it could be uninitialised data */
-               if (leftsize == 0 || leftsize == TDB_PAD_U32) {
-                       goto update;
-               }
-
-               left = offset - leftsize;
-
-               if (leftsize > offset ||
-                   left < TDB_DATA_START(tdb->hash_size)) {
-                       goto update;
-               }
-
-               /* Now read in the left record */
-               if (tdb->methods->tdb_read(tdb, left, &l, sizeof(l), DOCONV()) == -1) {
-                       TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: left read failed at %u (%u)\n", left, leftsize));
-                       goto update;
-               }
-
-               /* If it's free, expand to include it. */
-               if (l.magic == TDB_FREE_MAGIC) {
-                       /* we now merge the new record into the left record, rather than the other
-                          way around. This makes the operation O(1) instead of O(n). This change
-                          prevents traverse from being O(n^2) after a lot of deletes */
-                       l.rec_len += sizeof(*rec) + rec->rec_len;
-                       if (tdb_rec_write(tdb, left, &l) == -1) {
-                               TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: update_left failed at %u\n", left));
-                               goto fail;
-                       }
-                       if (update_tailer(tdb, left, &l) == -1) {
-                               TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: update_tailer failed at %u\n", offset));
-                               goto fail;
-                       }
-                       tdb_unlock(tdb, -1, F_WRLCK);
-                       return 0;
-               }
+       ret = check_merge_with_left_record(tdb, offset, rec, NULL, NULL);
+       if (ret == -1) {
+               goto fail;
+       }
+       if (ret == 1) {
+               /* merged */
+               goto done;
        }
 
-update:
+       /* Nothing to merge, prepend to free list */
 
-       /* Now, prepend to free list */
        rec->magic = TDB_FREE_MAGIC;
 
        if (tdb_ofs_read(tdb, FREELIST_TOP, &rec->next) == -1 ||
@@ -199,6 +365,7 @@ update:
                goto fail;
        }
 
+done:
        /* And we're done. */
        tdb_unlock(tdb, -1, F_WRLCK);
        return 0;
@@ -282,6 +449,7 @@ static tdb_off_t tdb_allocate_from_freelist(
                tdb_len_t rec_len;
        } bestfit;
        float multiplier = 1.0;
+       bool merge_created_candidate;
 
        /* over-allocate to reduce fragmentation */
        length *= 1.25;
@@ -291,6 +459,7 @@ static tdb_off_t tdb_allocate_from_freelist(
        length = TDB_ALIGN(length, TDB_ALIGNMENT);
 
  again:
+       merge_created_candidate = false;
        last_ptr = FREELIST_TOP;
 
        /* read in the freelist top */
@@ -307,10 +476,59 @@ static tdb_off_t tdb_allocate_from_freelist(
           issues when faced with a slowly increasing record size.
         */
        while (rec_ptr) {
+               int ret;
+               tdb_off_t left_ptr;
+               struct tdb_record left_rec;
+
                if (tdb_rec_free_read(tdb, rec_ptr, rec) == -1) {
                        return 0;
                }
 
+               ret = check_merge_with_left_record(tdb, rec_ptr, rec,
+                                                  &left_ptr, &left_rec);
+               if (ret == -1) {
+                       return 0;
+               }
+               if (ret == 1) {
+                       /* merged */
+                       rec_ptr = rec->next;
+                       ret = tdb_ofs_write(tdb, last_ptr, &rec->next);
+                       if (ret == -1) {
+                               return 0;
+                       }
+
+                       /*
+                        * We have merged the current record into the left
+                        * neighbour. So our traverse of the freelist will
+                        * skip it and consider the next record in the chain.
+                        *
+                        * But the enlarged left neighbour may be a candidate.
+                        * If it is, we can not directly use it, though.
+                        * The only thing we can do and have to do here is to
+                        * update the current best fit size in the chain if the
+                        * current best fit is the left record. (By that we may
+                        * worsen the best fit we already had, bit this is not a
+                        * problem.)
+                        *
+                        * If the current best fit is not the left record,
+                        * all we can do is remember the fact that a merge
+                        * created a new candidate so that we can trigger
+                        * a second walk of the freelist if at the end of
+                        * the first walk we have not found any fit.
+                        * This way we can avoid expanding the database.
+                        */
+
+                       if (bestfit.rec_ptr == left_ptr) {
+                               bestfit.rec_len = left_rec.rec_len;
+                       }
+
+                       if (left_rec.rec_len > length) {
+                               merge_created_candidate = true;
+                       }
+
+                       continue;
+               }
+
                if (rec->rec_len >= length) {
                        if (bestfit.rec_ptr == 0 ||
                            rec->rec_len < bestfit.rec_len) {
@@ -350,6 +568,10 @@ static tdb_off_t tdb_allocate_from_freelist(
                return newrec_ptr;
        }
 
+       if (merge_created_candidate) {
+               goto again;
+       }
+
        /* we didn't find enough space. See if we can expand the
           database and if we can then try again */
        if (tdb_expand(tdb, length + sizeof(*rec)) == 0)
@@ -404,22 +626,10 @@ tdb_off_t tdb_allocate(struct tdb_context *tdb, int hash, tdb_len_t length,
         * hash chains, something which is pretty bad normally)
         */
 
-       for (i=1; i<tdb->hash_size; i++) {
+       for (i=0; i<tdb->hash_size; i++) {
 
                int list;
 
-               if (tdb_lock_nonblock(tdb, -1, F_WRLCK) == 0) {
-                       /*
-                        * Under the freelist lock take the chance to give
-                        * back our dead records.
-                        */
-                       tdb_purge_dead(tdb, hash);
-
-                       ret = tdb_allocate_from_freelist(tdb, length, rec);
-                       tdb_unlock(tdb, -1, F_WRLCK);
-                       return ret;
-               }
-
                list = BUCKET(hash+i);
 
                if (tdb_lock_nonblock(tdb, list, F_WRLCK) == 0) {
@@ -432,6 +642,18 @@ tdb_off_t tdb_allocate(struct tdb_context *tdb, int hash, tdb_len_t length,
                                return ret;
                        }
                }
+
+               if (tdb_lock_nonblock(tdb, -1, F_WRLCK) == 0) {
+                       /*
+                        * Under the freelist lock take the chance to give
+                        * back our dead records.
+                        */
+                       tdb_purge_dead(tdb, hash);
+
+                       ret = tdb_allocate_from_freelist(tdb, length, rec);
+                       tdb_unlock(tdb, -1, F_WRLCK);
+                       return ret;
+               }
        }
 
 blocking_freelist_allocate:
@@ -444,10 +666,69 @@ blocking_freelist_allocate:
        return ret;
 }
 
-/*
-   return the size of the freelist - used to decide if we should repack
-*/
-_PUBLIC_ int tdb_freelist_size(struct tdb_context *tdb)
+/**
+ * Merge adjacent records in the freelist.
+ */
+static int tdb_freelist_merge_adjacent(struct tdb_context *tdb,
+                                      int *count_records, int *count_merged)
+{
+       tdb_off_t cur, next;
+       int count = 0;
+       int merged = 0;
+       int ret;
+
+       ret = tdb_lock(tdb, -1, F_RDLCK);
+       if (ret == -1) {
+               return -1;
+       }
+
+       cur = FREELIST_TOP;
+       while (tdb_ofs_read(tdb, cur, &next) == 0 && next != 0) {
+               tdb_off_t next2;
+
+               count++;
+
+               ret = check_merge_ptr_with_left_record(tdb, next, &next2);
+               if (ret == -1) {
+                       goto done;
+               }
+               if (ret == 1) {
+                       /*
+                        * merged:
+                        * now let cur->next point to next2 instead of next
+                        */
+
+                       ret = tdb_ofs_write(tdb, cur, &next2);
+                       if (ret != 0) {
+                               goto done;
+                       }
+
+                       next = next2;
+                       merged++;
+               }
+
+               cur = next;
+       }
+
+       if (count_records != NULL) {
+               *count_records = count;
+       }
+
+       if (count_merged != NULL) {
+               *count_merged = merged;
+       }
+
+       ret = 0;
+
+done:
+       tdb_unlock(tdb, -1, F_RDLCK);
+       return ret;
+}
+
+/**
+ * return the size of the freelist - no merging done
+ */
+static int tdb_freelist_size_no_merge(struct tdb_context *tdb)
 {
        tdb_off_t ptr;
        int count=0;
@@ -464,3 +745,28 @@ _PUBLIC_ int tdb_freelist_size(struct tdb_context *tdb)
        tdb_unlock(tdb, -1, F_RDLCK);
        return count;
 }
+
+/**
+ * return the size of the freelist - used to decide if we should repack
+ *
+ * As a side effect, adjacent records are merged unless the
+ * database is read-only, in order to reduce the fragmentation
+ * without repacking.
+ */
+_PUBLIC_ int tdb_freelist_size(struct tdb_context *tdb)
+{
+
+       int count = 0;
+
+       if (tdb->read_only) {
+               count = tdb_freelist_size_no_merge(tdb);
+       } else {
+               int ret;
+               ret = tdb_freelist_merge_adjacent(tdb, &count, NULL);
+               if (ret != 0) {
+                       return -1;
+               }
+       }
+
+       return count;
+}