fiemap: use ranged queries in dedupe phase to avoid full-file extent scans

fiemap.c

@@ -43,6 +43,24 @@ static unsigned int fiemap_count_extents(int fd)
 	return fiemap.fm_mapped_extents;
 }
 
+static unsigned int fiemap_count_extents_range(int fd, uint64_t start,
+					       uint64_t length)
+{
+	struct fiemap fiemap = {0,};
+	int err;
+
+	fiemap.fm_start = start;
+	fiemap.fm_length = length;
+
+	err = ioctl(fd, FS_IOC_FIEMAP, &fiemap);
+	if (err < 0) {
+		perror("fiemap_count_extents_range");
+		return 0;
+	}
+
+	return fiemap.fm_mapped_extents;
+}
+
 struct fiemap_extent *get_extent(struct fiemap *fiemap, size_t loff,
 				 unsigned int *index)
 {
@@ -98,6 +116,37 @@ struct fiemap *do_fiemap(int fd)
 	return fiemap;
 }
 
+struct fiemap *do_fiemap_range(int fd, uint64_t start, uint64_t length)
+{
+	int err;
+	struct fiemap *fiemap = NULL;
+	unsigned int count = fiemap_count_extents_range(fd, start, length);
+
+	if (count == 0)
+		return NULL;
+
+	fiemap = calloc(1, sizeof(struct fiemap) +
+			count * sizeof(struct fiemap_extent));
+	if (!fiemap)
+		return NULL;
+
+	fiemap->fm_start = start;
+	fiemap->fm_length = length;
+	fiemap->fm_extent_count = count;
+
+	err = ioctl(fd, FS_IOC_FIEMAP, fiemap);
+	if (err < 0) {
+		perror("do_fiemap_range");
+		free(fiemap);
+		return NULL;
+	}
+
+	if (fiemap->fm_mapped_extents != count)
+		dprintf("do_fiemap_range: file changed between fiemap calls\n");
+
+	return fiemap;
+}
+
 int fiemap_count_shared(int fd, size_t start_off, size_t end_off, uint64_t *shared)
 {
 	_cleanup_(freep) struct fiemap *fiemap = NULL;
@@ -108,9 +157,11 @@ int fiemap_count_shared(int fd, size_t start_off, size_t end_off, uint64_t *shar
 
 	abort_on(start_off >= end_off);
 
-	fiemap = do_fiemap(fd);
-	if (!fiemap)
-		return 1;
+	fiemap = do_fiemap_range(fd, start_off, end_off - start_off);
+	if (!fiemap) {
+		*shared = 0;
+		return 0;
+	}
 
 	*shared = 0;
 

fiemap.h

@@ -23,6 +23,12 @@ struct fiemap_extent *get_extent(struct fiemap *fiemap, size_t loff,
  */
 struct fiemap *do_fiemap(int fd);
 
+/*
+ * Extract extents mapping for a specific byte range.
+ * Only returns extents overlapping [start, start+length).
+ */
+struct fiemap *do_fiemap_range(int fd, uint64_t start, uint64_t length);
+
 /*
  * Count how much of the area between start_off and end_off is shared.
  */

filerec.c

@@ -316,25 +316,19 @@ int fiemap_scan_extent(struct extent *extent)
 {
 	int ret = 0;
 	_cleanup_(freep) struct fiemap *fiemap = NULL;
-	struct fiemap_extent *result;
 
 	ret = filerec_open(extent->e_file, true);
 	if (ret)
 		return ret;
 
-	fiemap = do_fiemap(extent->e_file->fd);
-	if (!fiemap) {
+	fiemap = do_fiemap_range(extent->e_file->fd, extent->e_loff,
+				 extent_len(extent));
+	if (!fiemap || fiemap->fm_mapped_extents == 0) {
 		filerec_close(extent->e_file);
 		return -1;
 	}
 
-	result = get_extent(fiemap, extent->e_loff, NULL);
-	if (!result) {
-		filerec_close(extent->e_file);
-		return -1;
-	}
-
-	extent->e_poff = result->fe_physical;
+	extent->e_poff = fiemap->fm_extents[0].fe_physical;
 	filerec_close(extent->e_file);
 	return ret;
 }

run_dedupe.c

@@ -140,9 +140,6 @@ static void add_shared_extents(struct dupe_extents *dext, uint64_t *shared)
 
 /*
  * Fiemap the file and get our post-dedupe extent state.
- *
- * XXX: We're running fiemap too much for this. At the least we should
- * shoot for one fiemap call(s) per filerec.
  */
 static void add_shared_extents_post(struct dupe_extents *dext, uint64_t *shared)
 {