X-Git-Url: https://mattmccutchen.net/rsync/rsync.git/blobdiff_plain/ba582f753a4c3eb2dd793904fb6a378ed8ffd90a..beab3078d637de643825a3b8f0c246a7c2db042d:/match.c

diff --git a/match.c b/match.c
index 1eecf3c3..4fe23fd5 100644
--- a/match.c
+++ b/match.c
@@ -20,14 +20,11 @@
 #include "rsync.h"
 
 extern int verbose;
-extern int am_server;
 extern int do_progress;
 extern int checksum_seed;
+extern int append_mode;
 
-typedef unsigned short tag;
-
-#define TABLESIZE (1<<16)
-#define NULL_TAG ((size_t)-1)
+int updating_basis_file;
 
 static int false_alarms;
 static int tag_hits;
@@ -40,47 +37,37 @@ static int total_matches;
 
 extern struct stats stats;
 
-struct target {
-	tag t;
-	size_t i;
-};
-
-static struct target *targets;
-
-static size_t *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 gettag2(s1,s2) gettag((s1) + ((s2)<<16))
+#define gettag(sum) ((sum)%tablesize)
 
 static void build_hash_table(struct sum_struct *s)
 {
-	size_t i;
-
-	if (!tag_table)
-		tag_table = new_array(size_t, TABLESIZE);
+	int32 i;
+	uint32 prior_size = 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;
 }
 
 
@@ -99,17 +86,20 @@ static OFF_T last_match;
  * @param i If >0, the number of a matched token.  If 0, indicates we
  * have only literal data.
  **/
-static void matched(int f,struct sum_struct *s,struct map_struct *buf,
-		    OFF_T offset,int i)
+static void matched(int f, struct sum_struct *s, struct map_struct *buf,
+		    OFF_T offset, int32 i)
 {
-	OFF_T n = offset - last_match;
-	OFF_T j;
-
-	if (verbose > 2 && i >= 0)
-		rprintf(FINFO,"match at %.0f last_match=%.0f j=%d len=%u n=%.0f\n",
-			(double)offset,(double)last_match,i,s->sums[i].len,(double)n);
+	int32 n = offset - last_match; /* max value: block_size (int32) */
+	int32 j;
+
+	if (verbose > 2 && i >= 0) {
+		rprintf(FINFO,
+			"match at %.0f last_match=%.0f j=%d len=%ld n=%ld\n",
+			(double)offset, (double)last_match, i,
+			(long)s->sums[i].len, (long)n);
+	}
 
-	send_token(f,i,buf,last_match,n,i<0?0:s->sums[i].len);
+	send_token(f, i, buf, last_match, n, i < 0 ? 0 : s->sums[i].len);
 	data_transfer += n;
 
 	if (i >= 0) {
@@ -118,45 +108,40 @@ static void matched(int f,struct sum_struct *s,struct map_struct *buf,
 	}
 
 	for (j = 0; j < n; j += CHUNK_SIZE) {
-		int n1 = MIN(CHUNK_SIZE,n-j);
-		sum_update(map_ptr(buf,last_match+j,n1),n1);
+		int32 n1 = MIN(CHUNK_SIZE, n - j);
+		sum_update(map_ptr(buf, last_match + j, n1), n1);
 	}
 
-
 	if (i >= 0)
 		last_match = offset + s->sums[i].len;
 	else
 		last_match = offset;
 
-	if (buf && do_progress) {
+	if (buf && do_progress)
 		show_progress(last_match, buf->file_size);
-
-		if (i == -1)
-			end_progress(buf->file_size);
-	}
 }
 
 
 static void hash_search(int f,struct sum_struct *s,
 			struct map_struct *buf, OFF_T len)
 {
-	OFF_T offset, end;
-	unsigned int k;
-	size_t last_i;
+	OFF_T offset, end, backup;
+	int32 k, want_i;
 	char sum2[SUM_LENGTH];
 	uint32 s1, s2, sum;
+	int more;
 	schar *map;
 
-	/* last_i is used to encourage adjacent matches, allowing the RLL coding of the
-	   output to work more efficiently */
-	last_i = (size_t)-1;
+	/* want_i is used to encourage adjacent matches, allowing the RLL
+	 * coding of the output to work more efficiently. */
+	want_i = 0;
 
 	if (verbose > 2) {
-		rprintf(FINFO,"hash search b=%u len=%.0f\n",
-			s->blength, (double)len);
+		rprintf(FINFO, "hash search b=%ld len=%.0f\n",
+			(long)s->blength, (double)len);
 	}
 
-	k = MIN(len, s->blength);
+	k = (int32)MIN(len, (OFF_T)s->blength);
 
 	map = (schar *)map_ptr(buf, 0, k);
 
@@ -164,45 +149,49 @@ static void hash_search(int f,struct sum_struct *s,
 	s1 = sum & 0xFFFF;
 	s2 = sum >> 16;
 	if (verbose > 3)
-		rprintf(FINFO, "sum=%.8x k=%u\n", sum, k);
+		rprintf(FINFO, "sum=%.8x k=%ld\n", sum, (long)k);
 
 	offset = 0;
 
 	end = len + 1 - s->sums[s->count-1].len;
 
 	if (verbose > 3) {
-		rprintf(FINFO, "hash search s->blength=%u len=%.0f count=%.0f\n",
-			s->blength, (double)len, (double)s->count);
+		rprintf(FINFO, "hash search s->blength=%ld len=%.0f count=%.0f\n",
+			(long)s->blength, (double)len, (double)s->count);
 	}
 
 	do {
-		tag t = gettag2(s1,s2);
+		uint32 t = gettag2(s1,s2);
 		int done_csum2 = 0;
-		size_t j = tag_table[t];
+		int32 i;
 
 		if (verbose > 4)
 			rprintf(FINFO,"offset=%.0f sum=%08x\n",(double)offset,sum);
 
-		if (j == NULL_TAG)
-			goto null_tag;
-
 		sum = (s1 & 0xffff) | (s2 << 16);
 		tag_hits++;
-		for (; j < s->count && targets[j].t == t; j++) {
-			unsigned int l;
-			size_t i = targets[j].i;
+		for (i = sum_table[t]; i >= 0; i = s->sums[i].chain) {
+			int32 l;
 
 			if (sum != s->sums[i].sum1)
 				continue;
 
 			/* also make sure the two blocks are the same length */
-			l = MIN((OFF_T)s->blength, len-offset);
+			l = (int32)MIN((OFF_T)s->blength, len-offset);
 			if (l != s->sums[i].len)
 				continue;
 
-			if (verbose > 3)
-				rprintf(FINFO,"potential match at %.0f target=%.0f %.0f sum=%08x\n",
-					(double)offset,(double)j,(double)i,sum);
+			/* in-place: ensure chunk's offset is either >= our
+			 * offset or that the data didn't move. */
+			if (updating_basis_file && s->sums[i].offset < offset
+			    && !(s->sums[i].flags & SUMFLG_SAME_OFFSET))
+				continue;
+
+			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);
@@ -215,27 +204,46 @@ static void hash_search(int f,struct sum_struct *s,
 				continue;
 			}
 
-			/* we've found a match, but now check to see
-			 * if last_i can hint at a better match */
-			for (j++; j < s->count && targets[j].t == t; j++) {
-				size_t i2 = targets[j].i;
-				if (i2 == last_i + 1) {
-					if (sum != s->sums[i2].sum1)
-						break;
-					if (memcmp(sum2,s->sums[i2].sum2,s->s2length) != 0)
-						break;
-					/* we've found an adjacent match - the RLL coder
-					 * will be happy */
-					i = i2;
-					break;
+			/* When updating in-place, the best possible match is
+			 * one with an identical offset, so we prefer that over
+			 * the following want_i optimization. */
+			if (updating_basis_file) {
+				int32 i2;
+				for (i2 = i; i2 >= 0; i2 = s->sums[i2].chain) {
+					if (s->sums[i2].offset != offset)
+						continue;
+					if (i2 != i) {
+						if (sum != s->sums[i2].sum1)
+							break;
+						if (memcmp(sum2, s->sums[i2].sum2,
+							   s->s2length) != 0)
+							break;
+						i = i2;
+					}
+					/* This chunk was at the same offset on
+					 * both the sender and the receiver. */
+					s->sums[i].flags |= SUMFLG_SAME_OFFSET;
+					goto set_want_i;
 				}
 			}
 
-			last_i = i;
+			/* we've found a match, but now check to see
+			 * if want_i can hint at a better match. */
+			if (i != want_i && want_i < s->count
+			    && (!updating_basis_file || s->sums[want_i].offset >= offset
+			     || s->sums[want_i].flags & SUMFLG_SAME_OFFSET)
+			    && sum == s->sums[want_i].sum1
+			    && memcmp(sum2, s->sums[want_i].sum2, s->s2length) == 0) {
+				/* we've found an adjacent match - the RLL coder
+				 * will be happy */
+				i = want_i;
+			}
+		    set_want_i:
+			want_i = i + 1;
 
 			matched(f,s,buf,offset,i);
 			offset += s->sums[i].len - 1;
-			k = MIN(s->blength, len-offset);
+			k = (int32)MIN((OFF_T)s->blength, len-offset);
 			map = (schar *)map_ptr(buf, offset, k);
 			sum = get_checksum1((char *)map, k);
 			s1 = sum & 0xFFFF;
@@ -244,15 +252,21 @@ static void hash_search(int f,struct sum_struct *s,
 			break;
 		}
 
-	null_tag:
+		backup = offset - last_match;
+		/* We sometimes read 1 byte prior to last_match... */
+		if (backup < 0)
+			backup = 0;
+
 		/* Trim off the first byte from the checksum */
-		map = (schar *)map_ptr(buf, offset, k+1);
+		more = offset + k < len;
+		map = (schar *)map_ptr(buf, offset - backup, k + more + backup)
+		    + backup;
 		s1 -= map[0] + CHAR_OFFSET;
 		s2 -= k * (map[0]+CHAR_OFFSET);
 
 		/* Add on the next byte (if there is one) to the checksum */
-		if (k < (len-offset)) {
-			s1 += (map[k]+CHAR_OFFSET);
+		if (more) {
+			s1 += map[k] + CHAR_OFFSET;
 			s2 += s1;
 		} else
 			--k;
@@ -263,15 +277,12 @@ static void hash_search(int f,struct sum_struct *s,
 		   match. The 3 reads are caused by the
 		   running match, the checksum update and the
 		   literal send. */
-		if (offset > last_match
-		 && offset-last_match >= CHUNK_SIZE+s->blength
-		 && end-offset > CHUNK_SIZE) {
-			matched(f,s,buf,offset - s->blength, -2);
-		}
+		if (backup >= s->blength+CHUNK_SIZE && end-offset > CHUNK_SIZE)
+			matched(f, s, buf, offset - s->blength, -2);
 	} while (++offset < end);
 
-	matched(f,s,buf,len,-1);
-	map_ptr(buf,len-1,1);
+	matched(f, s, buf, len, -1);
+	map_ptr(buf, len-1, 1);
 }
 
 
@@ -292,7 +303,6 @@ static void hash_search(int f,struct sum_struct *s,
 void match_sums(int f, struct sum_struct *s, struct map_struct *buf, OFF_T len)
 {
 	char file_sum[MD4_SUM_LENGTH];
-	extern int write_batch;
 
 	last_match = 0;
 	false_alarms = 0;
@@ -302,7 +312,26 @@ void match_sums(int f, struct sum_struct *s, struct map_struct *buf, OFF_T len)
 
 	sum_init(checksum_seed);
 
-	if (len > 0 && s->count>0) {
+	if (append_mode) {
+		OFF_T j = 0;
+		for (j = CHUNK_SIZE; j < s->flength; j += CHUNK_SIZE) {
+			if (buf && do_progress)
+				show_progress(last_match, buf->file_size);
+			sum_update(map_ptr(buf, last_match, CHUNK_SIZE),
+				   CHUNK_SIZE);
+			last_match = j;
+		}
+		if (last_match < s->flength) {
+			int32 len = s->flength - last_match;
+			if (buf && do_progress)
+				show_progress(last_match, buf->file_size);
+			sum_update(map_ptr(buf, last_match, len), len);
+			last_match = s->flength;
+		}
+		s->count = 0;
+	}
+
+	if (len > 0 && s->count > 0) {
 		build_hash_table(s);
 
 		if (verbose > 2)
@@ -315,11 +344,9 @@ void match_sums(int f, struct sum_struct *s, struct map_struct *buf, OFF_T len)
 	} else {
 		OFF_T j;
 		/* by doing this in pieces we avoid too many seeks */
-		for (j = 0; j < len-CHUNK_SIZE; j += CHUNK_SIZE) {
-			int n1 = MIN(CHUNK_SIZE,(len-CHUNK_SIZE)-j);
-			matched(f,s,buf,j+n1,-2);
-		}
-		matched(f,s,buf,len,-1);
+		for (j = last_match + CHUNK_SIZE; j < len; j += CHUNK_SIZE)
+			matched(f, s, buf, j, -2);
+		matched(f, s, buf, len, -1);
 	}
 
 	sum_end(file_sum);
@@ -330,13 +357,6 @@ void match_sums(int f, struct sum_struct *s, struct map_struct *buf, OFF_T len)
 	if (verbose > 2)
 		rprintf(FINFO,"sending file_sum\n");
 	write_buf(f,file_sum,MD4_SUM_LENGTH);
-	if (write_batch)
-		write_batch_delta_file(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",