]> asedeno.scripts.mit.edu Git - linux.git/commitdiff
Btrfs: heuristic: implement sampling logic
authorTimofey Titovets <nefelim4ag@gmail.com>
Thu, 28 Sep 2017 14:33:38 +0000 (17:33 +0300)
committerDavid Sterba <dsterba@suse.com>
Wed, 1 Nov 2017 19:45:36 +0000 (20:45 +0100)
Copy sample data from the input data range to sample buffer then
calculate byte value count for that sample into bucket.

Signed-off-by: Timofey Titovets <nefelim4ag@gmail.com>
[ minor comment updates ]
Signed-off-by: David Sterba <dsterba@suse.com>
fs/btrfs/compression.c

index 4f5d958c71ad8fdeb7d4871f71c67da822896493..0e1561cc9578b9c395c7c67052ec97c19652bfab 100644 (file)
@@ -746,6 +746,7 @@ struct bucket_item {
 struct heuristic_ws {
        /* Partial copy of input data */
        u8 *sample;
+       u32 sample_size;
        /* Buckets store counters for each byte value */
        struct bucket_item *bucket;
        struct list_head list;
@@ -1221,6 +1222,58 @@ int btrfs_decompress_buf2page(const char *buf, unsigned long buf_start,
        return 1;
 }
 
+static void heuristic_collect_sample(struct inode *inode, u64 start, u64 end,
+                                    struct heuristic_ws *ws)
+{
+       struct page *page;
+       u64 index, index_end;
+       u32 i, curr_sample_pos;
+       u8 *in_data;
+
+       /*
+        * Compression handles the input data by chunks of 128KiB
+        * (defined by BTRFS_MAX_UNCOMPRESSED)
+        *
+        * We do the same for the heuristic and loop over the whole range.
+        *
+        * MAX_SAMPLE_SIZE - calculated under assumption that heuristic will
+        * process no more than BTRFS_MAX_UNCOMPRESSED at a time.
+        */
+       if (end - start > BTRFS_MAX_UNCOMPRESSED)
+               end = start + BTRFS_MAX_UNCOMPRESSED;
+
+       index = start >> PAGE_SHIFT;
+       index_end = end >> PAGE_SHIFT;
+
+       /* Don't miss unaligned end */
+       if (!IS_ALIGNED(end, PAGE_SIZE))
+               index_end++;
+
+       curr_sample_pos = 0;
+       while (index < index_end) {
+               page = find_get_page(inode->i_mapping, index);
+               in_data = kmap(page);
+               /* Handle case where the start is not aligned to PAGE_SIZE */
+               i = start % PAGE_SIZE;
+               while (i < PAGE_SIZE - SAMPLING_READ_SIZE) {
+                       /* Don't sample any garbage from the last page */
+                       if (start > end - SAMPLING_READ_SIZE)
+                               break;
+                       memcpy(&ws->sample[curr_sample_pos], &in_data[i],
+                                       SAMPLING_READ_SIZE);
+                       i += SAMPLING_INTERVAL;
+                       start += SAMPLING_INTERVAL;
+                       curr_sample_pos += SAMPLING_READ_SIZE;
+               }
+               kunmap(page);
+               put_page(page);
+
+               index++;
+       }
+
+       ws->sample_size = curr_sample_pos;
+}
+
 /*
  * Compression heuristic.
  *
@@ -1240,19 +1293,19 @@ int btrfs_compress_heuristic(struct inode *inode, u64 start, u64 end)
 {
        struct list_head *ws_list = __find_workspace(0, true);
        struct heuristic_ws *ws;
-       u64 index = start >> PAGE_SHIFT;
-       u64 end_index = end >> PAGE_SHIFT;
-       struct page *page;
+       u32 i;
+       u8 byte;
        int ret = 1;
 
        ws = list_entry(ws_list, struct heuristic_ws, list);
 
-       while (index <= end_index) {
-               page = find_get_page(inode->i_mapping, index);
-               kmap(page);
-               kunmap(page);
-               put_page(page);
-               index++;
+       heuristic_collect_sample(inode, start, end, ws);
+
+       memset(ws->bucket, 0, sizeof(*ws->bucket)*BUCKET_SIZE);
+
+       for (i = 0; i < ws->sample_size; i++) {
+               byte = ws->sample[i];
+               ws->bucket[byte].count++;
        }
 
        __free_workspace(0, ws_list, true);