Fixed a few bugs.

[rsync/rsync-patches.git] / dynamic_hash.diff

This patch makes the processing of large really large files more efficient
by making sure that the sender's hash table is large enough to hold all the
checksum entries without being overloaded.  It also makes the hashing of
normal sized files use slightly less memory than before.

An extended version of a patch by Shachar Shemesh.

--- old/match.c
+++ new/match.c
@@ -26,63 +26,47 @@ extern int append_mode;
 
 int updating_basis_file;
 
-typedef unsigned short tag;
-
-#define TABLESIZE (1<<16)
-#define NULL_TAG (-1)
-
 static int false_alarms;
-static int tag_hits;
+static int hash_hits;
 static int matches;
 static int64 data_transfer;
 
 static int total_false_alarms;
-static int total_tag_hits;
+static int total_hash_hits;
 static int total_matches;
 
 extern struct stats stats;
 
-struct target {
-	tag t;
-	int32 i;
-};
-
-static struct target *targets;
-
-static int32 *tag_table;
-
-#define gettag2(s1,s2) (((s1) + (s2)) & 0xFFFF)
-#define gettag(sum) gettag2((sum)&0xFFFF,(sum)>>16)
-
-static int compare_targets(struct target *t1,struct target *t2)
-{
-	return (int)t1->t - (int)t2->t;
-}
+static uint32 tablesize;
+static int32 *sum_table;
 
+#define GETTAG(sum) ((sum)%tablesize)
 
 static void build_hash_table(struct sum_struct *s)
 {
 	int32 i;
+	uint32 prior_size = tablesize;
 
-	if (!tag_table)
-		tag_table = new_array(int32, TABLESIZE);
+	/* Dynamically calculate the hash table size so that the hash load
+	 * for big files is about 80%.  This number must be odd or s2 will
+	 * not be able to span the entire set. */
+	tablesize = (uint32)(s->count/8) * 10 + 11;
+	if (tablesize < 65537)
+		tablesize = 65537; /* a prime number */
+	if (tablesize != prior_size) {
+		free(sum_table);
+		sum_table = new_array(int32, tablesize);
+		if (!sum_table)
+			out_of_memory("build_hash_table");
+	}
 
-	targets = new_array(struct target, s->count);
-	if (!tag_table || !targets)
-		out_of_memory("build_hash_table");
+	memset(sum_table, 0xFF, tablesize * sizeof sum_table[0]);
 
 	for (i = 0; i < s->count; i++) {
-		targets[i].i = i;
-		targets[i].t = gettag(s->sums[i].sum1);
+		uint32 t = GETTAG(s->sums[i].sum1);
+		s->sums[i].chain = sum_table[t];
+		sum_table[t] = i;
 	}
-
-	qsort(targets,s->count,sizeof(targets[0]),(int (*)())compare_targets);
-
-	for (i = 0; i < TABLESIZE; i++)
-		tag_table[i] = NULL_TAG;
-
-	for (i = s->count; i-- > 0; )
-		tag_table[targets[i].t] = i;
 }
 
 
@@ -176,20 +160,21 @@ static void hash_search(int f,struct sum
 	}
 
 	do {
-		tag t = gettag2(s1,s2);
 		int done_csum2 = 0;
-		int32 j = tag_table[t];
+		int32 i;
+		uint32 t;
 
 		if (verbose > 4)
 			rprintf(FINFO,"offset=%.0f sum=%08x\n",(double)offset,sum);
 
-		if (j == NULL_TAG)
-			goto null_tag;
+		t = GETTAG(sum);
+		i = sum_table[t];
+		if (i < 0)
+			goto null_hash;
 
-		sum = (s1 & 0xffff) | (s2 << 16);
-		tag_hits++;
+		hash_hits++;
 		do {
-			int32 l, i = targets[j].i;
+			int32 l;
 
 			if (sum != s->sums[i].sum1)
 				continue;
@@ -205,9 +190,11 @@ static void hash_search(int f,struct sum
 			    && !(s->sums[i].flags & SUMFLG_SAME_OFFSET))
 				continue;
 
-			if (verbose > 3)
-				rprintf(FINFO,"potential match at %.0f target=%.0f %.0f sum=%08x\n",
-					(double)offset,(double)j,(double)i,sum);
+			if (verbose > 3) {
+				rprintf(FINFO,
+					"potential match at %.0f i=%ld sum=%08x\n",
+					(double)offset, (long)i, sum);
+			}
 
 			if (!done_csum2) {
 				map = (schar *)map_ptr(buf,offset,l);
@@ -224,8 +211,8 @@ static void hash_search(int f,struct sum
 			 * one with an identical offset, so we prefer that over
 			 * the following want_i optimization. */
 			if (updating_basis_file) {
-				do {
-					int32 i2 = targets[j].i;
+				int32 i2;
+				for (i2 = i; i2 >= 0; i2 = s->sums[i2].chain) {
 					if (s->sums[i2].offset != offset)
 						continue;
 					if (i2 != i) {
@@ -240,7 +227,7 @@ static void hash_search(int f,struct sum
 					 * both the sender and the receiver. */
 					s->sums[i].flags |= SUMFLG_SAME_OFFSET;
 					goto set_want_i;
-				} while (++j < s->count && targets[j].t == t);
+				}
 			}
 
 			/* we've found a match, but now check to see
@@ -266,9 +253,9 @@ static void hash_search(int f,struct sum
 			s2 = sum >> 16;
 			matches++;
 			break;
-		} while (++j < s->count && targets[j].t == t);
+		} while ((i = s->sums[i].chain) >= 0);
 
-	null_tag:
+	  null_hash:
 		backup = offset - last_match;
 		/* We sometimes read 1 byte prior to last_match... */
 		if (backup < 0)
@@ -287,6 +274,7 @@ static void hash_search(int f,struct sum
 			s2 += s1;
 		} else
 			--k;
+		sum = (s1 & 0xffff) | (s2 << 16);
 
 		/* By matching early we avoid re-reading the
 		   data 3 times in the case where a token
@@ -323,7 +311,7 @@ void match_sums(int f, struct sum_struct
 
 	last_match = 0;
 	false_alarms = 0;
-	tag_hits = 0;
+	hash_hits = 0;
 	matches = 0;
 	data_transfer = 0;
 
@@ -375,16 +363,11 @@ void match_sums(int f, struct sum_struct
 		rprintf(FINFO,"sending file_sum\n");
 	write_buf(f,file_sum,MD4_SUM_LENGTH);
 
-	if (targets) {
-		free(targets);
-		targets=NULL;
-	}
-
 	if (verbose > 2)
-		rprintf(FINFO, "false_alarms=%d tag_hits=%d matches=%d\n",
-			false_alarms, tag_hits, matches);
+		rprintf(FINFO, "false_alarms=%d hash_hits=%d matches=%d\n",
+			false_alarms, hash_hits, matches);
 
-	total_tag_hits += tag_hits;
+	total_hash_hits += hash_hits;
 	total_false_alarms += false_alarms;
 	total_matches += matches;
 	stats.literal_data += data_transfer;
@@ -396,8 +379,7 @@ void match_report(void)
 		return;
 
 	rprintf(FINFO,
-		"total: matches=%d  tag_hits=%d  false_alarms=%d data=%.0f\n",
-		total_matches,total_tag_hits,
-		total_false_alarms,
+		"total: matches=%d  hash_hits=%d  false_alarms=%d data=%.0f\n",
+		total_matches, total_hash_hits, total_false_alarms,
 		(double)stats.literal_data);
 }
--- old/rsync.h
+++ new/rsync.h
@@ -560,6 +560,7 @@ struct sum_buf {
 	OFF_T offset;		/**< offset in file of this chunk */
 	int32 len;		/**< length of chunk of file */
 	uint32 sum1;	        /**< simple checksum */
+	int32 chain;		/**< next hash-table collision */
 	short flags;		/**< flag bits */
 	char sum2[SUM_LENGTH];	/**< checksum  */
 };