X-Git-Url: https://git.ozlabs.org/?a=blobdiff_plain;f=ccan%2Ftdb2%2Ffree.c;h=49880f18dd51517601eca798d163e5e98b10b514;hb=4e185ad8ab5a7e01edbbe12d11eb2f1577de7e8b;hp=a854aadfb6a3b25f763062ba8acea18a20a5d7f1;hpb=a5b66d7072af7dd5c17c1f4b1ede85f5cd4b7ca3;p=ccan

diff --git a/ccan/tdb2/free.c b/ccan/tdb2/free.c
index a854aadf..49880f18 100644
--- a/ccan/tdb2/free.c
+++ b/ccan/tdb2/free.c
@@ -105,7 +105,7 @@ static tdb_off_t find_free_head(struct tdb_context *tdb,
 /* Remove from free bucket. */
 static int remove_from_list(struct tdb_context *tdb,
 			    tdb_off_t b_off, tdb_off_t r_off,
-			    struct tdb_free_record *r)
+			    const struct tdb_free_record *r)
 {
 	tdb_off_t off;
 
@@ -118,8 +118,8 @@ static int remove_from_list(struct tdb_context *tdb,
 
 #ifdef DEBUG
 	if (tdb_read_off(tdb, off) != r_off) {
-		tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
-			 "remove_from_list: %llu bad prev in list %llu\n",
+		tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
+			 "remove_from_list: %llu bad prev in list %llu",
 			 (long long)r_off, (long long)b_off);
 		return -1;
 	}
@@ -136,9 +136,9 @@ static int remove_from_list(struct tdb_context *tdb,
 
 #ifdef DEBUG
 		if (tdb_read_off(tdb, off) & TDB_OFF_MASK != r_off) {
-			tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
-				 "remove_from_list: %llu bad list %llu\n",
-				 (long long)r_off, (long long)b_off);
+			tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
+				   "remove_from_list: %llu bad list %llu",
+				   (long long)r_off, (long long)b_off);
 			return -1;
 		}
 #endif
@@ -176,9 +176,10 @@ static int enqueue_in_free(struct tdb_context *tdb,
 				 new.next + offsetof(struct tdb_free_record,
 						     magic_and_prev))
 		    != magic) {
-			tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
-				 "enqueue_in_free: %llu bad head prev %llu\n",
-				 (long long)new.next, (long long)b_off);
+			tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
+				   "enqueue_in_free: %llu bad head"
+				   " prev %llu",
+				   (long long)new.next, (long long)b_off);
 			return -1;
 		}
 #endif
@@ -244,12 +245,15 @@ static size_t record_leftover(size_t keylen, size_t datalen,
 	return leftover;
 }
 
-/* FIXME: Shortcut common case where tdb->flist == flist */
 static tdb_off_t flist_offset(struct tdb_context *tdb, unsigned int flist)
 {
-	tdb_off_t off = first_flist(tdb);
+	tdb_off_t off;
 	unsigned int i;
 
+	if (likely(tdb->flist == flist))
+		return tdb->flist_off;
+
+	off = first_flist(tdb);
 	for (i = 0; i < flist; i++)
 		off = next_flist(tdb, off);
 	return off;
@@ -259,90 +263,94 @@ static tdb_off_t flist_offset(struct tdb_context *tdb, unsigned int flist)
 static int coalesce(struct tdb_context *tdb,
 		    tdb_off_t off, tdb_off_t b_off, tdb_len_t data_len)
 {
-	struct tdb_free_record pad, *r;
 	tdb_off_t end;
+	struct tdb_free_record rec;
 
+	add_stat(tdb, alloc_coalesce_tried, 1);
 	end = off + sizeof(struct tdb_used_record) + data_len;
 
 	while (end < tdb->map_size) {
+		const struct tdb_free_record *r;
 		tdb_off_t nb_off;
 		unsigned flist, bucket;
 
-		/* FIXME: do tdb_get here and below really win? */
-		r = tdb_get(tdb, end, &pad, sizeof(pad));
+		r = tdb_access_read(tdb, end, sizeof(*r), true);
 		if (!r)
 			goto err;
 
-		if (frec_magic(r) != TDB_FREE_MAGIC)
+		if (frec_magic(r) != TDB_FREE_MAGIC) {
+			tdb_access_release(tdb, r);
 			break;
+		}
 
 		flist = frec_flist(r);
 		bucket = size_to_bucket(frec_len(r));
 		nb_off = bucket_off(flist_offset(tdb, flist), bucket);
+		tdb_access_release(tdb, r);
 
 		/* We may be violating lock order here, so best effort. */
-		if (tdb_lock_free_bucket(tdb, nb_off, TDB_LOCK_NOWAIT) == -1)
+		if (tdb_lock_free_bucket(tdb, nb_off, TDB_LOCK_NOWAIT) == -1) {
+			add_stat(tdb, alloc_coalesce_lockfail, 1);
 			break;
+		}
 
 		/* Now we have lock, re-check. */
-		r = tdb_get(tdb, end, &pad, sizeof(pad));
-		if (!r) {
+		if (tdb_read_convert(tdb, end, &rec, sizeof(rec))) {
 			tdb_unlock_free_bucket(tdb, nb_off);
 			goto err;
 		}
 
-		if (unlikely(frec_magic(r) != TDB_FREE_MAGIC)) {
+		if (unlikely(frec_magic(&rec) != TDB_FREE_MAGIC)) {
+			add_stat(tdb, alloc_coalesce_race, 1);
 			tdb_unlock_free_bucket(tdb, nb_off);
 			break;
 		}
 
-		if (unlikely(frec_flist(r) != flist)
-		    || unlikely(size_to_bucket(frec_len(r)) != bucket)) {
+		if (unlikely(frec_flist(&rec) != flist)
+		    || unlikely(size_to_bucket(frec_len(&rec)) != bucket)) {
+			add_stat(tdb, alloc_coalesce_race, 1);
 			tdb_unlock_free_bucket(tdb, nb_off);
 			break;
 		}
 
-		if (remove_from_list(tdb, nb_off, end, r) == -1) {
+		if (remove_from_list(tdb, nb_off, end, &rec) == -1) {
 			tdb_unlock_free_bucket(tdb, nb_off);
 			goto err;
 		}
 
-		end += sizeof(struct tdb_used_record) + frec_len(r);
+		end += sizeof(struct tdb_used_record) + frec_len(&rec);
 		tdb_unlock_free_bucket(tdb, nb_off);
+		add_stat(tdb, alloc_coalesce_num_merged, 1);
 	}
 
 	/* Didn't find any adjacent free? */
 	if (end == off + sizeof(struct tdb_used_record) + data_len)
 		return 0;
 
-	/* OK, expand record */
-	r = tdb_get(tdb, off, &pad, sizeof(pad));
-	if (!r)
+	/* OK, expand initial record */
+	if (tdb_read_convert(tdb, off, &rec, sizeof(rec)))
 		goto err;
 
-	if (frec_len(r) != data_len) {
-		tdb->ecode = TDB_ERR_CORRUPT;
-		tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
-			 "coalesce: expected data len %llu not %llu\n",
-			 (long long)data_len, (long long)frec_len(r));
+	if (frec_len(&rec) != data_len) {
+		tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
+			   "coalesce: expected data len %zu not %zu",
+			   (size_t)data_len, (size_t)frec_len(&rec));
 		goto err;
 	}
 
-	if (remove_from_list(tdb, b_off, off, r) == -1)
-		goto err;
-
-	r = tdb_access_write(tdb, off, sizeof(*r), true);
-	if (!r)
+	if (remove_from_list(tdb, b_off, off, &rec) == -1)
 		goto err;
 
 	/* We have to drop this to avoid deadlocks, so make sure record
 	 * doesn't get coalesced by someone else! */
-	r->magic_and_prev = TDB_COALESCING_MAGIC << (64 - TDB_OFF_UPPER_STEAL);
+	rec.magic_and_prev = TDB_COALESCING_MAGIC
+		<< (64 - TDB_OFF_UPPER_STEAL);
 	/* FIXME: Use 255 as invalid free list? */
-	r->flist_and_len = end - off - sizeof(struct tdb_used_record);
-	if (tdb_access_commit(tdb, r) != 0)
+	rec.flist_and_len = end - off - sizeof(struct tdb_used_record);
+	if (tdb_write_convert(tdb, off, &rec, sizeof(rec)) != 0)
 		goto err;
 
+	add_stat(tdb, alloc_coalesce_succeeded, 1);
 	tdb_unlock_free_bucket(tdb, b_off);
 
 	if (add_free_record(tdb, off, end - off) == -1)
@@ -364,10 +372,11 @@ static tdb_off_t lock_and_alloc(struct tdb_context *tdb,
 				unsigned hashlow)
 {
 	tdb_off_t off, b_off,best_off;
-	struct tdb_free_record pad, best = { 0 }, *r;
+	struct tdb_free_record best = { 0 };
 	double multiplier;
 	size_t size = adjust_size(keylen, datalen);
 
+	add_stat(tdb, allocs, 1);
 again:
 	b_off = bucket_off(flist_off, bucket);
 
@@ -393,14 +402,18 @@ again:
 		goto unlock_err;
 
 	while (off) {
-		/* FIXME: Does tdb_get win anything here? */
-		r = tdb_get(tdb, off, &pad, sizeof(*r));
+		const struct tdb_free_record *r;
+		tdb_len_t len;
+		tdb_off_t next;
+
+		r = tdb_access_read(tdb, off, sizeof(*r), true);
 		if (!r)
 			goto unlock_err;
 
 		if (frec_magic(r) != TDB_FREE_MAGIC) {
-			tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-				 "lock_and_alloc: %llu non-free 0x%llx\n",
+			tdb_access_release(tdb, r);
+			tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
+				 "lock_and_alloc: %llu non-free 0x%llx",
 				 (long long)off, (long long)r->magic_and_prev);
 			goto unlock_err;
 		}
@@ -410,13 +423,19 @@ again:
 			best = *r;
 		}
 
-		if (frec_len(&best) < size * multiplier && best_off)
+		if (frec_len(&best) < size * multiplier && best_off) {
+			tdb_access_release(tdb, r);
 			break;
+		}
 
 		multiplier *= 1.01;
 
+		next = r->next;
+		len = frec_len(r);
+		tdb_access_release(tdb, r);
+
 		/* Since we're going slow anyway, try coalescing here. */
-		switch (coalesce(tdb, off, b_off, frec_len(r))) {
+		switch (coalesce(tdb, off, b_off, len)) {
 		case -1:
 			/* This has already unlocked on error. */
 			return -1;
@@ -424,7 +443,7 @@ again:
 			/* This has unlocked list, restart. */
 			goto again;
 		}
-		off = r->next;
+		off = next;
 	}
 
 	/* If we found anything at all, use it. */
@@ -453,6 +472,7 @@ again:
 		/* Bucket of leftover will be <= current bucket, so nested
 		 * locking is allowed. */
 		if (leftover) {
+			add_stat(tdb, alloc_leftover, 1);
 			if (add_free_record(tdb,
 					    best_off + sizeof(rec)
 					    + frec_len(&best) - leftover,
@@ -502,6 +522,10 @@ static tdb_off_t get_free(struct tdb_context *tdb,
 			if (off == TDB_OFF_ERR)
 				return TDB_OFF_ERR;
 			if (off != 0) {
+				if (b == start_b)
+					add_stat(tdb, alloc_bucket_exact, 1);
+				if (b == TDB_FREE_BUCKETS - 1)
+					add_stat(tdb, alloc_bucket_max, 1);
 				/* Worked?  Stay using this list. */
 				tdb->flist_off = flist_off;
 				tdb->flist = flist;
@@ -513,6 +537,7 @@ static tdb_off_t get_free(struct tdb_context *tdb,
 		/* Hmm, try next list. */
 		flist_off = next_flist(tdb, flist_off);
 		flist++;
+
 		if (flist_off == 0) {
 			wrapped = true;
 			flist_off = first_flist(tdb);
@@ -541,9 +566,8 @@ int set_used_header(struct tdb_context *tdb,
 	if (rec_key_length(rec) != keylen
 	    || rec_data_length(rec) != datalen
 	    || rec_extra_padding(rec) != actuallen - (keylen + datalen)) {
-		tdb->ecode = TDB_ERR_IO;
-		tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-			 "Could not encode k=%llu,d=%llu,a=%llu\n",
+		tdb_logerr(tdb, TDB_ERR_IO, TDB_DEBUG_ERROR,
+			 "Could not encode k=%llu,d=%llu,a=%llu",
 			 (long long)keylen, (long long)datalen,
 			 (long long)actuallen);
 		return -1;
@@ -563,8 +587,8 @@ static int tdb_expand(struct tdb_context *tdb, tdb_len_t size)
 	/* Need to hold a hash lock to expand DB: transactions rely on it. */
 	if (!(tdb->flags & TDB_NOLOCK)
 	    && !tdb->allrecord_lock.count && !tdb_has_hash_locks(tdb)) {
-		tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
-			 "tdb_expand: must hold lock during expand\n");
+		tdb_logerr(tdb, TDB_ERR_LOCK, TDB_DEBUG_ERROR,
+			   "tdb_expand: must hold lock during expand");
 		return -1;
 	}
 
@@ -596,6 +620,7 @@ static int tdb_expand(struct tdb_context *tdb, tdb_len_t size)
 	/* We need to drop this lock before adding free record. */
 	tdb_unlock_expand(tdb, F_WRLCK);
 
+	add_stat(tdb, expands, 1);
 	return add_free_record(tdb, old_size, wanted);
 }