free diskblocks. Addresses #1195.

git-svn-id: file:///svn/toku/tokudb.1195@7679 c7de825b-a66e-492c-adef-691d508d4ae1

newbrt/brt-internal.h

@@ -115,8 +115,8 @@ struct remembered_hash {
 };
 
 struct block_translation_pair {
-    DISKOFF diskoff;
+    DISKOFF diskoff; // When in free list, set to the next free block.  In this case it's really a BLOCKNUM.
-    DISKOFF size;
+    DISKOFF size;    // set to 0xFFFFFFFFFFFFFFFF for free
 };
 
 // The brt_header is not managed by the cachetable.  Instead, it hangs off the cachefile as userdata.

newbrt/brt-serialize.c

@@ -294,7 +294,7 @@ void toku_serialize_brtnode_to (int fd, BLOCKNUM blocknum, BRTNODE node, struct
 	assert(r==Z_OK);
     }
 
-    if (0) printf("Size before compressing %u, after compression %lu\n", calculated_size-uncompressed_magic_len, compressed_len);
+    if (0) printf("Block %" PRId64 " Size before compressing %u, after compression %lu\n", blocknum.b, calculated_size-uncompressed_magic_len, compressed_len);
 
     ((int32_t*)(compressed_buf+uncompressed_magic_len))[0] = htonl(compressed_len);
     ((int32_t*)(compressed_buf+uncompressed_magic_len))[1] = htonl(uncompressed_len);
@@ -308,18 +308,7 @@ void toku_serialize_brtnode_to (int fd, BLOCKNUM blocknum, BRTNODE node, struct
 	//printf("%s:%d translated_blocknum_limit=%lu blocknum.b=%lu\n", __FILE__, __LINE__, h->translated_blocknum_limit, blocknum.b);
 	//printf("%s:%d allocator=%p\n", __FILE__, __LINE__, h->block_allocator);
 	//printf("%s:%d bt=%p\n", __FILE__, __LINE__, h->block_translation);
-	if (h->translated_blocknum_limit <= (u_int64_t)blocknum.b) {
+	extend_block_translation(blocknum, h);
-	    if (h->block_translation == 0) assert(h->translated_blocknum_limit==0);
-	    u_int64_t new_limit = blocknum.b + 1;
-	    u_int64_t old_limit = h->translated_blocknum_limit;
-	    u_int64_t j;
-	    XREALLOC_N(new_limit, h->block_translation);
-	    for (j=old_limit; j<new_limit; j++) {
-		h->block_translation[j].diskoff = 0;
-		h->block_translation[j].size    = 0;
-	    }
-	    h->translated_blocknum_limit = new_limit;
-	}
 	if (h->block_translation[blocknum.b].size > 0) {
 	    block_allocator_free_block(h->block_allocator, h->block_translation[blocknum.b].diskoff);
 	    h->block_translation[blocknum.b].diskoff = 0;
@@ -344,6 +333,7 @@ void toku_serialize_brtnode_to (int fd, BLOCKNUM blocknum, BRTNODE node, struct
 }
 
 int toku_deserialize_brtnode_from (int fd, BLOCKNUM blocknum, u_int32_t fullhash, BRTNODE *brtnode, struct brt_header *h) {
+    if (0) printf("Deserializing Block %" PRId64 "\n", blocknum.b);
     assert(0 <= blocknum.b && (u_int64_t)blocknum.b < h->translated_blocknum_limit);
     DISKOFF offset = h->block_translation[blocknum.b].diskoff;
     TAGMALLOC(BRTNODE, result);
@@ -371,8 +361,8 @@ int toku_deserialize_brtnode_from (int fd, BLOCKNUM blocknum, u_int32_t fullhash
 	compressed_size   = ntohl(*(u_int32_t*)(&uncompressed_header[uncompressed_magic_len]));
 	if (compressed_size<=0   || compressed_size>(1<<30)) { r = DB_BADFORMAT; goto died0; }
 	uncompressed_size = ntohl(*(u_int32_t*)(&uncompressed_header[uncompressed_magic_len+4]));
+	if (0) printf("Block %" PRId64 " Compressed size = %u, uncompressed size=%u\n", blocknum.b, compressed_size, uncompressed_size);
 	if (uncompressed_size<=0 || uncompressed_size>(1<<30)) { r = DB_BADFORMAT; goto died0; }
-	if (0) printf("Compressed size = %u, uncompressed size=%u\n", compressed_size, uncompressed_size);
     }
     
     //printf("%s:%d serializing %" PRIu64 " size=%d\n", __FILE__, __LINE__, blocknum.b, uncompressed_size);

newbrt/brt.c

@@ -546,13 +546,59 @@ void toku_brtheader_free (struct brt_header *h) {
     toku_free(h);
 }
 
+void
+extend_block_translation (BLOCKNUM blocknum, struct brt_header *h)
+// Effect: Record a block translation.  This means extending the translation table, and setting the diskoff and size to zero in any of the unused spots.
+{
+    if (h->translated_blocknum_limit <= (u_int64_t)blocknum.b) {
+	if (h->block_translation == 0) assert(h->translated_blocknum_limit==0);
+	u_int64_t new_limit = blocknum.b + 1;
+	u_int64_t old_limit = h->translated_blocknum_limit;
+	u_int64_t j;
+	XREALLOC_N(new_limit, h->block_translation);
+	for (j=old_limit; j<new_limit; j++) {
+	    h->block_translation[j].diskoff = 0;
+	    h->block_translation[j].size    = 0;
+	}
+	h->translated_blocknum_limit = new_limit;
+    }
+}
+
+const DISKOFF diskoff_is_null = (DISKOFF)-1; // in a freelist, this indicates end of list
+const DISKOFF size_is_free = (DISKOFF)-1;
+
 static int
 allocate_diskblocknumber (BLOCKNUM *res, BRT brt, TOKULOGGER logger __attribute__((__unused__))) {
-    assert(brt->h->free_blocks.b == -1); // no blocks in the free list
+    BLOCKNUM result;
-    BLOCKNUM result = brt->h->unused_blocks;
+    if (brt->h->free_blocks.b == diskoff_is_null) {
+	// no blocks in the free list
+	result = brt->h->unused_blocks;
 	brt->h->unused_blocks.b++;
-    brt->h->dirty = 1;
+    } else {
+	result = brt->h->free_blocks;
+	assert(brt->h->block_translation[result.b].size = size_is_free);
+	brt->h->block_translation[result.b].size = 0;
+	brt->h->free_blocks.b = brt->h->block_translation[result.b].diskoff; // pop the freelist
+    }
+    assert(result.b>0);
     *res = result;
+    brt->h->dirty = 1;
+    return 0;
+}
+
+static int
+free_diskblocknumber (BLOCKNUM *b, struct brt_header *h, TOKULOGGER logger __attribute__((__unused__)))
+// Effect: Free a diskblock
+//  Watch out for the case where the disk block was never yet written to disk and is beyond the translated_blocknum_limit.
+{
+    extend_block_translation(*b, h);
+    assert((u_int64_t)b->b <= h->translated_blocknum_limit);
+    assert(h->block_translation[b->b].size != size_is_free);
+    h->block_translation[b->b].size = size_is_free;
+    h->block_translation[b->b].diskoff = h->free_blocks.b;
+    h->free_blocks.b = b->b;
+    b->b = 0;
+    h->dirty = 1;
     return 0;
 }
 
@@ -2080,14 +2126,18 @@ brt_merge_child (BRT t, BRTNODE node, int childnum_to_merge, BOOL *did_io, TOKUL
     // Unpin both, and return the first nonzero error code that is found
     assert(node->dirty);
     {
+	int rrb1 = 0;
 	int rra = toku_unpin_brtnode(t, childa);
 	int rrb;
 	if (did_merge) {
-	    rrb = toku_cachetable_unpin_and_remove(t->cf, childb->thisnodename);
+	    BLOCKNUM bn = childb->thisnodename;
+	    rrb = toku_cachetable_unpin_and_remove(t->cf, bn);
+	    rrb1 = free_diskblocknumber(&bn, t->h, logger);
 	} else {
 	    rrb = toku_unpin_brtnode(t, childb);
 	}
 
+	if (rrb1) return rrb1;
 	if (rra) return rra;
 	if (rrb) return rrb;
     }
@@ -4125,6 +4175,12 @@ int toku_dump_brt (FILE *f, BRT brt) {
     CACHEKEY *rootp;
     assert(brt->h);
     u_int32_t fullhash;
+    u_int64_t i;
+    fprintf(f, "Block translation:");
+    for (i=0; i<brt->h->translated_blocknum_limit; i++) {
+        fprintf(f, " %"PRIu64": %"PRId64" %"PRId64"", i, brt->h->block_translation[i].diskoff, brt->h->block_translation[i].size);
+    }
+    fprintf(f, "\n");
     rootp = toku_calculate_root_offset_pointer(brt, &fullhash);
     return toku_dump_brtnode(f, brt, *rootp, 0, 0, 0, 0, 0);
 }

newbrt/brt.h

@@ -113,4 +113,6 @@ enum brt_header_flags {
 
 int toku_brt_keyrange (BRT brt, DBT *key, u_int64_t *less,  u_int64_t *equal,  u_int64_t *greater);
 
+void extend_block_translation (BLOCKNUM blocknum, struct brt_header *h);
+
 #endif

newbrt/brtdump.c

@@ -291,7 +291,14 @@ main (int argc, const char *argv[]) {
         }
     } else {
 	BLOCKNUM blocknum;
+	printf("Block translation:");
+	for (blocknum.b=0; blocknum.b<h->unused_blocks.b; blocknum.b++) {
+	    printf(" %" PRIu64 ":", blocknum.b);
+	    if (h->block_translation[blocknum.b].size == -1) printf("free");
+	    else printf("%" PRIu64 ":%" PRIu64, h->block_translation[blocknum.b].diskoff, h->block_translation[blocknum.b].size);
+	}
 	for (blocknum.b=1; blocknum.b<h->unused_blocks.b; blocknum.b++) {
+	    if (h->block_translation[blocknum.b].size != -1)
 		dump_node(f, blocknum, h);
         }
     }

newbrt/tests/brt-test.c

@@ -1089,6 +1089,7 @@ static void test_new_brt_cursor_last(int n, int dup_mode) {
         memcpy(&vv, val.data, val.size);
         assert(vv == (int) htonl(i));
 
+	//if (n==512 && i<=360) { printf("i=%d\n", i); toku_dump_brt(stdout, t); }
         r = toku_brt_cursor_delete(cursor, 0, null_txn); assert(r == 0);
     }
     assert(i == -1);