dect
/
linux-2.6
Archived
13
0
Fork 0

Btrfs: Fixup reference counting on cows

Signed-off-by: Chris Mason <chris.mason@oracle.com>
This commit is contained in:
Chris Mason 2007-03-06 20:08:01 -05:00 committed by David Woodhouse
parent 02217ed299
commit a28ec19775
9 changed files with 274 additions and 134 deletions

View File

@ -1,7 +1,7 @@
CC=gcc
CFLAGS = -g -Wall
headers = radix-tree.h ctree.h disk-io.h kerncompat.h print-tree.h
headers = radix-tree.h ctree.h disk-io.h kerncompat.h print-tree.h list.h
objects = ctree.o disk-io.o radix-tree.o mkfs.o extent-tree.o print-tree.o
# if you don't have sparse installed, use ls instead

View File

@ -50,16 +50,17 @@ int btrfs_cow_block(struct ctree_root *root,
memcpy(&cow->node, &buf->node, sizeof(buf->node));
cow->node.header.blocknr = cow->blocknr;
*cow_ret = cow;
btrfs_inc_ref(root, buf);
if (buf == root->node) {
root->node = cow;
cow->count++;
if (buf != root->commit_root)
free_extent(root, buf->blocknr, 1);
tree_block_release(root, buf);
} else {
parent->node.blockptrs[parent_slot] = cow->blocknr;
BUG_ON(list_empty(&parent->dirty));
}
if (0 && root != root->extent_root && !is_leaf(cow->node.header.flags)) {
btrfs_inc_ref(root, cow);
free_extent(root, buf->blocknr, 1);
}
tree_block_release(root, buf);
return 0;
@ -1018,7 +1019,6 @@ static int split_leaf(struct ctree_root *root, struct ctree_path *path,
slot = path->slots[0];
nritems = l->header.nritems;
mid = (nritems + 1)/ 2;
right_buffer = alloc_free_block(root);
BUG_ON(!right_buffer);
BUG_ON(mid == nritems);
@ -1170,7 +1170,6 @@ static int del_ptr(struct ctree_root *root, struct ctree_path *path, int level,
node = &parent->node;
nritems = node->header.nritems;
if (slot != nritems -1) {
memmove(node->keys + slot, node->keys + slot + 1,
sizeof(struct key) * (nritems - slot - 1));

View File

@ -51,10 +51,12 @@ struct tree_buffer;
*/
struct ctree_root {
struct tree_buffer *node;
struct tree_buffer *commit_root;
struct ctree_root *extent_root;
struct key current_insert;
int fp;
struct radix_tree_root cache_radix;
struct radix_tree_root pinned_radix;
struct list_head trans;
struct list_head cache;
int cache_size;
@ -151,4 +153,6 @@ int del_item(struct ctree_root *root, struct ctree_path *path);
int insert_item(struct ctree_root *root, struct key *key, void *data, int data_size);
int next_leaf(struct ctree_root *root, struct ctree_path *path);
int leaf_free_space(struct leaf *leaf);
int btrfs_drop_snapshot(struct ctree_root *root, struct tree_buffer *snap);
int btrfs_finish_extent_commit(struct ctree_root *root);
#endif

View File

@ -153,13 +153,24 @@ static int __commit_transaction(struct ctree_root *root)
return ret;
}
int commit_transaction(struct ctree_root *root)
int commit_transaction(struct ctree_root *root, struct ctree_super_block *s)
{
int ret;
int ret = 0;
ret = __commit_transaction(root);
if (!ret && root != root->extent_root)
ret = __commit_transaction(root->extent_root);
BUG_ON(ret);
if (root->commit_root != root->node) {
struct tree_buffer *snap = root->commit_root;
root->commit_root = root->node;
root->node->count++;
ret = btrfs_drop_snapshot(root, snap);
BUG_ON(ret);
tree_block_release(root, snap);
}
write_ctree_super(root, s);
btrfs_finish_extent_commit(root);
return ret;
}
@ -168,10 +179,13 @@ static int __setup_root(struct ctree_root *root, struct ctree_root *extent_root,
{
INIT_LIST_HEAD(&root->trans);
INIT_LIST_HEAD(&root->cache);
root->cache_size = 0;
root->fp = fp;
root->node = NULL;
root->node = read_tree_block(root, info->tree_root);
root->extent_root = extent_root;
root->commit_root = NULL;
root->node = read_tree_block(root, info->tree_root);
memset(&root->current_insert, 0, sizeof(root->current_insert));
return 0;
}
@ -188,6 +202,8 @@ struct ctree_root *open_ctree(char *filename, struct ctree_super_block *super)
return NULL;
}
INIT_RADIX_TREE(&root->cache_radix, GFP_KERNEL);
INIT_RADIX_TREE(&root->pinned_radix, GFP_KERNEL);
INIT_RADIX_TREE(&extent_root->pinned_radix, GFP_KERNEL);
INIT_RADIX_TREE(&extent_root->cache_radix, GFP_KERNEL);
ret = pread(fp, super, sizeof(struct ctree_super_block),
CTREE_SUPER_INFO_OFFSET(CTREE_BLOCKSIZE));
@ -204,6 +220,8 @@ struct ctree_root *open_ctree(char *filename, struct ctree_super_block *super)
BUG_ON(ret < 0);
__setup_root(root, extent_root, &super->root_info, fp);
__setup_root(extent_root, extent_root, &super->extent_info, fp);
root->commit_root = root->node;
root->node->count++;
return root;
}
@ -236,9 +254,11 @@ static int drop_cache(struct ctree_root *root)
}
return 0;
}
int close_ctree(struct ctree_root *root)
int close_ctree(struct ctree_root *root, struct ctree_super_block *s)
{
commit_transaction(root);
commit_transaction(root, s);
__commit_transaction(root->extent_root);
write_ctree_super(root, s);
drop_cache(root->extent_root);
drop_cache(root);
BUG_ON(!list_empty(&root->trans));
@ -249,6 +269,7 @@ int close_ctree(struct ctree_root *root)
tree_block_release(root, root->node);
if (root->extent_root->node)
tree_block_release(root->extent_root, root->extent_root->node);
tree_block_release(root, root->commit_root);
free(root);
printf("on close %d blocks are allocated\n", allocated_blocks);
return 0;

View File

@ -18,9 +18,9 @@ struct tree_buffer *find_tree_block(struct ctree_root *root, u64 blocknr);
int write_tree_block(struct ctree_root *root, struct tree_buffer *buf);
int dirty_tree_block(struct ctree_root *root, struct tree_buffer *buf);
int clean_tree_block(struct ctree_root *root, struct tree_buffer *buf);
int commit_transaction(struct ctree_root *root);
int commit_transaction(struct ctree_root *root, struct ctree_super_block *s);
struct ctree_root *open_ctree(char *filename, struct ctree_super_block *s);
int close_ctree(struct ctree_root *root);
int close_ctree(struct ctree_root *root, struct ctree_super_block *s);
void tree_block_release(struct ctree_root *root, struct tree_buffer *buf);
int write_ctree_super(struct ctree_root *root, struct ctree_super_block *s);
int mkfs(int fd);

View File

@ -13,7 +13,8 @@
* other allocations are done. The pending tag is also used in the same
* manner for deletes.
*/
#define CTREE_EXTENT_PENDING 0
#define CTREE_EXTENT_PENDING_ADD 0
#define CTREE_EXTENT_PENDING_DEL 1
static int inc_block_ref(struct ctree_root *root, u64 blocknr)
{
@ -27,20 +28,51 @@ static int inc_block_ref(struct ctree_root *root, u64 blocknr)
key.flags = 0;
key.offset = 1;
ret = search_slot(root->extent_root, &key, &path, 0, 1);
if (ret != 0)
BUG();
BUG_ON(ret != 0);
l = &path.nodes[0]->leaf;
item = (struct extent_item *)(l->data +
l->items[path.slots[0]].offset);
item->refs++;
BUG_ON(list_empty(&path.nodes[0]->dirty));
release_path(root->extent_root, &path);
return 0;
}
static int lookup_block_ref(struct ctree_root *root, u64 blocknr, int *refs)
{
struct ctree_path path;
int ret;
struct key key;
struct leaf *l;
struct extent_item *item;
init_path(&path);
key.objectid = blocknr;
key.flags = 0;
key.offset = 1;
ret = search_slot(root->extent_root, &key, &path, 0, 0);
if (ret != 0)
BUG();
l = &path.nodes[0]->leaf;
item = (struct extent_item *)(l->data +
l->items[path.slots[0]].offset);
*refs = item->refs;
release_path(root->extent_root, &path);
return 0;
}
int btrfs_inc_ref(struct ctree_root *root, struct tree_buffer *buf)
{
u64 blocknr;
int i;
if (root == root->extent_root)
return 0;
if (is_leaf(buf->node.header.flags))
return 0;
for (i = 0; i < buf->node.header.nritems; i++) {
blocknr = buf->node.blockptrs[i];
inc_block_ref(root, blocknr);
@ -48,6 +80,113 @@ int btrfs_inc_ref(struct ctree_root *root, struct tree_buffer *buf)
return 0;
}
int btrfs_finish_extent_commit(struct ctree_root *root)
{
struct ctree_root *extent_root = root->extent_root;
unsigned long gang[8];
int ret;
int i;
while(1) {
ret = radix_tree_gang_lookup(&extent_root->pinned_radix,
(void **)gang, 0,
ARRAY_SIZE(gang));
if (!ret)
break;
for (i = 0; i < ret; i++)
radix_tree_delete(&extent_root->pinned_radix, gang[i]);
}
return 0;
}
/*
* remove an extent from the root, returns 0 on success
*/
int __free_extent(struct ctree_root *root, u64 blocknr, u64 num_blocks)
{
struct ctree_path path;
struct key key;
struct ctree_root *extent_root = root->extent_root;
int ret;
struct item *item;
struct extent_item *ei;
key.objectid = blocknr;
key.flags = 0;
key.offset = num_blocks;
init_path(&path);
ret = search_slot(extent_root, &key, &path, -1, 1);
if (ret) {
printf("failed to find %Lu\n", key.objectid);
print_tree(extent_root, extent_root->node);
printf("failed to find %Lu\n", key.objectid);
BUG();
}
item = path.nodes[0]->leaf.items + path.slots[0];
ei = (struct extent_item *)(path.nodes[0]->leaf.data + item->offset);
BUG_ON(ei->refs == 0);
ei->refs--;
if (ei->refs == 0) {
if (root == extent_root) {
int err;
radix_tree_preload(GFP_KERNEL);
err = radix_tree_insert(&extent_root->pinned_radix,
blocknr, (void *)blocknr);
BUG_ON(err);
radix_tree_preload_end();
}
ret = del_item(extent_root, &path);
if (ret)
BUG();
}
release_path(extent_root, &path);
return ret;
}
/*
* insert all of the pending extents reserved during the original
* allocation. (CTREE_EXTENT_PENDING). Returns zero if it all worked out
*/
static int insert_pending_extents(struct ctree_root *extent_root)
{
int ret;
struct key key;
struct extent_item item;
struct tree_buffer *gang[4];
int i;
// FIXME -ENOSPC
item.owner = extent_root->node->node.header.parentid;
item.refs = 1;
while(1) {
ret = radix_tree_gang_lookup_tag(&extent_root->cache_radix,
(void **)gang, 0,
ARRAY_SIZE(gang),
CTREE_EXTENT_PENDING_ADD);
if (!ret)
break;
for (i = 0; i < ret; i++) {
key.objectid = gang[i]->blocknr;
key.flags = 0;
key.offset = 1;
ret = insert_item(extent_root, &key, &item,
sizeof(item));
if (ret) {
printf("%Lu already in tree\n", key.objectid);
print_tree(extent_root, extent_root->node);
BUG();
// FIXME undo it and return sane
return ret;
}
radix_tree_tag_clear(&extent_root->cache_radix,
gang[i]->blocknr,
CTREE_EXTENT_PENDING_ADD);
tree_block_release(extent_root, gang[i]);
}
}
return 0;
}
/*
* find all the blocks marked as pending in the radix tree and remove
* them from the extent map
@ -55,78 +194,73 @@ int btrfs_inc_ref(struct ctree_root *root, struct tree_buffer *buf)
static int del_pending_extents(struct ctree_root *extent_root)
{
int ret;
struct key key;
struct tree_buffer *gang[4];
int i;
struct ctree_path path;
while(1) {
ret = radix_tree_gang_lookup_tag(&extent_root->cache_radix,
(void **)gang, 0,
ARRAY_SIZE(gang),
CTREE_EXTENT_PENDING);
CTREE_EXTENT_PENDING_DEL);
if (!ret)
break;
for (i = 0; i < ret; i++) {
key.objectid = gang[i]->blocknr;
key.flags = 0;
key.offset = 1;
init_path(&path);
ret = search_slot(extent_root, &key, &path, -1, 1);
if (ret) {
print_tree(extent_root, extent_root->node);
printf("unable to find %Lu\n", key.objectid);
BUG();
// FIXME undo it and return sane
return ret;
}
ret = del_item(extent_root, &path);
if (ret) {
BUG();
return ret;
}
release_path(extent_root, &path);
ret = __free_extent(extent_root, gang[i]->blocknr, 1);
radix_tree_tag_clear(&extent_root->cache_radix,
gang[i]->blocknr,
CTREE_EXTENT_PENDING);
CTREE_EXTENT_PENDING_DEL);
tree_block_release(extent_root, gang[i]);
}
}
return 0;
}
static int run_pending(struct ctree_root *extent_root)
{
while(radix_tree_tagged(&extent_root->cache_radix,
CTREE_EXTENT_PENDING_DEL) ||
radix_tree_tagged(&extent_root->cache_radix,
CTREE_EXTENT_PENDING_ADD)) {
insert_pending_extents(extent_root);
del_pending_extents(extent_root);
}
return 0;
}
/*
* remove an extent from the root, returns 0 on success
*/
int free_extent(struct ctree_root *root, u64 blocknr, u64 num_blocks)
{
struct ctree_path path;
struct key key;
struct ctree_root *extent_root = root->extent_root;
struct tree_buffer *t;
int pending_ret;
int ret;
if (root == extent_root) {
t = find_tree_block(root, blocknr);
if (radix_tree_tag_get(&root->cache_radix, blocknr,
CTREE_EXTENT_PENDING_ADD)) {
radix_tree_tag_clear(&root->cache_radix,
blocknr,
CTREE_EXTENT_PENDING_ADD);
/* once for us */
tree_block_release(root, t);
/* once for the pending add */
tree_block_release(root, t);
} else {
radix_tree_tag_set(&root->cache_radix, blocknr,
CTREE_EXTENT_PENDING_DEL);
}
return 0;
}
key.objectid = blocknr;
key.flags = 0;
key.offset = num_blocks;
if (root == extent_root) {
t = read_tree_block(root, key.objectid);
radix_tree_tag_set(&root->cache_radix, key.objectid,
CTREE_EXTENT_PENDING);
return 0;
}
init_path(&path);
ret = search_slot(extent_root, &key, &path, -1, 1);
if (ret) {
print_tree(extent_root, extent_root->node);
printf("failed to find %Lu\n", key.objectid);
BUG();
}
ret = del_item(extent_root, &path);
if (ret)
BUG();
release_path(extent_root, &path);
pending_ret = del_pending_extents(root->extent_root);
ret = __free_extent(root, blocknr, num_blocks);
pending_ret = run_pending(root->extent_root);
return ret ? ret : pending_ret;
}
@ -203,7 +337,7 @@ check_pending:
*/
release_path(root, &path);
BUG_ON(ins->objectid < search_start);
if (orig_root->extent_root == orig_root) {
if (1 || orig_root->extent_root == orig_root) {
BUG_ON(num_blocks != 1);
if ((root->current_insert.objectid <= ins->objectid &&
root->current_insert.objectid +
@ -211,8 +345,9 @@ check_pending:
(root->current_insert.objectid > ins->objectid &&
root->current_insert.objectid <= ins->objectid +
ins->offset) ||
radix_tree_lookup(&root->pinned_radix, ins->objectid) ||
radix_tree_tag_get(&root->cache_radix, ins->objectid,
CTREE_EXTENT_PENDING)) {
CTREE_EXTENT_PENDING_ADD)) {
search_start = ins->objectid + 1;
goto check_failed;
}
@ -225,51 +360,6 @@ error:
return ret;
}
/*
* insert all of the pending extents reserved during the original
* allocation. (CTREE_EXTENT_PENDING). Returns zero if it all worked out
*/
static int insert_pending_extents(struct ctree_root *extent_root)
{
int ret;
struct key key;
struct extent_item item;
struct tree_buffer *gang[4];
int i;
// FIXME -ENOSPC
item.refs = 1;
item.owner = extent_root->node->node.header.parentid;
while(1) {
ret = radix_tree_gang_lookup_tag(&extent_root->cache_radix,
(void **)gang, 0,
ARRAY_SIZE(gang),
CTREE_EXTENT_PENDING);
if (!ret)
break;
for (i = 0; i < ret; i++) {
key.objectid = gang[i]->blocknr;
key.flags = 0;
key.offset = 1;
ret = insert_item(extent_root, &key, &item,
sizeof(item));
if (ret) {
printf("%Lu already in tree\n", key.objectid);
print_tree(extent_root, extent_root->node);
BUG();
// FIXME undo it and return sane
return ret;
}
radix_tree_tag_clear(&extent_root->cache_radix,
gang[i]->blocknr,
CTREE_EXTENT_PENDING);
printf("%Lu is not pending\n", gang[i]->blocknr);
tree_block_release(extent_root, gang[i]);
}
}
return 0;
}
/*
* finds a free extent and does all the dirty work required for allocation
* returns the key for the extent through ins, and a tree buffer for
@ -296,7 +386,7 @@ int alloc_extent(struct ctree_root *root, u64 num_blocks, u64 search_start,
sizeof(extent_item));
memset(&root->extent_root->current_insert, 0,
sizeof(struct key));
pending_ret = insert_pending_extents(root->extent_root);
pending_ret = run_pending(root->extent_root);
if (ret)
return ret;
if (pending_ret)
@ -309,9 +399,8 @@ int alloc_extent(struct ctree_root *root, u64 num_blocks, u64 search_start,
BUG_ON(ins->offset != 1);
*buf = find_tree_block(root, ins->objectid);
BUG_ON(!*buf);
printf("%Lu is pending\n", ins->objectid);
radix_tree_tag_set(&root->cache_radix, ins->objectid,
CTREE_EXTENT_PENDING);
CTREE_EXTENT_PENDING_ADD);
(*buf)->count++;
dirty_tree_block(root, *buf);
return 0;
@ -331,13 +420,41 @@ struct tree_buffer *alloc_free_block(struct ctree_root *root)
ret = alloc_extent(root, 1, 0, (unsigned long)-1,
root->node->node.header.parentid,
&ins, &buf);
if (ret) {
BUG();
return NULL;
}
if (root != root->extent_root)
BUG_ON(radix_tree_tag_get(&root->extent_root->cache_radix,
buf->blocknr, CTREE_EXTENT_PENDING));
buf->blocknr,
CTREE_EXTENT_PENDING_ADD));
return buf;
}
int btrfs_drop_snapshot(struct ctree_root *root, struct tree_buffer *snap)
{
int ret;
int level;
int refs;
u64 blocknr = snap->blocknr;
level = node_level(snap->node.header.flags);
ret = lookup_block_ref(root, snap->blocknr, &refs);
BUG_ON(ret);
if (refs == 1 && level != 0) {
struct node *n = &snap->node;
struct tree_buffer *b;
int i;
for (i = 0; i < n->header.nritems; i++) {
b = read_tree_block(root, n->blockptrs[i]);
/* FIXME, don't recurse here */
ret = btrfs_drop_snapshot(root, b);
BUG_ON(ret);
tree_block_release(root, b);
}
}
ret = free_extent(root, blocknr, 1);
BUG_ON(ret);
return 0;
}

View File

@ -21,9 +21,11 @@ void print_leaf(struct leaf *l)
item->key.objectid, item->key.flags, item->key.offset,
item->offset, item->size);
fflush(stdout);
printf("\t\titem data %.*s\n", item->size, l->data+item->offset);
printf("\t\titem data %.*s\n", item->size,
l->data+item->offset);
ei = (struct extent_item *)(l->data + item->offset);
printf("\t\textent data %u %Lu\n", ei->refs, ei->owner);
printf("\t\textent data refs %u owner %Lu\n", ei->refs,
ei->owner);
fflush(stdout);
}
}

View File

@ -19,7 +19,7 @@ int main(int ac, char **av) {
int i;
int num;
int ret;
int run_size = 1024;
int run_size = 100000;
int max_key = 100000000;
int tree_size = 0;
struct ctree_path path;
@ -44,9 +44,9 @@ int main(int ac, char **av) {
if (!ret)
tree_size++;
free(buf);
}
write_ctree_super(root, &super);
close_ctree(root);
close_ctree(root, &super);
root = open_ctree("dbfile", &super);
printf("starting search\n");
@ -65,8 +65,7 @@ int main(int ac, char **av) {
}
release_path(root, &path);
}
write_ctree_super(root, &super);
close_ctree(root);
close_ctree(root, &super);
root = open_ctree("dbfile", &super);
printf("node %p level %d total ptrs %d free spc %lu\n", root->node,
node_level(root->node->node.header.flags),
@ -90,8 +89,7 @@ int main(int ac, char **av) {
}
release_path(root, &path);
}
write_ctree_super(root, &super);
close_ctree(root);
close_ctree(root, &super);
root = open_ctree("dbfile", &super);
srand(128);
for (i = 0; i < run_size; i++) {
@ -106,8 +104,7 @@ int main(int ac, char **av) {
tree_size++;
free(buf);
}
write_ctree_super(root, &super);
close_ctree(root);
close_ctree(root, &super);
root = open_ctree("dbfile", &super);
srand(128);
printf("starting search2\n");
@ -156,10 +153,17 @@ int main(int ac, char **av) {
}
release_path(root, &path);
}
/*
printf("previous tree:\n");
print_tree(root, root->commit_root);
printf("map before commit\n");
print_tree(root->extent_root, root->extent_root->node);
*/
commit_transaction(root, &super);
printf("tree size is now %d\n", tree_size);
printf("root %p commit root %p\n", root->node, root->commit_root);
printf("map tree\n");
print_tree(root->extent_root, root->extent_root->node);
write_ctree_super(root, &super);
close_ctree(root);
close_ctree(root, &super);
return 0;
}

View File

@ -8,6 +8,7 @@
#include "print-tree.h"
int keep_running = 1;
struct ctree_super_block super;
static int setup_key(struct radix_tree_root *root, struct key *key, int exists)
{
@ -59,11 +60,6 @@ error:
return -1;
}
static int run_commit(struct ctree_root *root, struct radix_tree_root *radix)
{
return commit_transaction(root);
}
static int insert_dup(struct ctree_root *root, struct radix_tree_root *radix)
{
struct ctree_path path;
@ -210,7 +206,7 @@ static int fill_tree(struct ctree_root *root, struct radix_tree_root *radix,
goto out;
}
if (i % 1000 == 0) {
ret = commit_transaction(root);
ret = commit_transaction(root, &super);
if (ret) {
fprintf(stderr, "fill commit failed\n");
return ret;
@ -229,7 +225,7 @@ out:
static int bulk_op(struct ctree_root *root, struct radix_tree_root *radix)
{
int ret;
int nr = rand() % 20000;
int nr = rand() % 5000;
static int run_nr = 0;
/* do the bulk op much less frequently */
@ -247,7 +243,7 @@ static int bulk_op(struct ctree_root *root, struct radix_tree_root *radix)
int (*ops[])(struct ctree_root *root, struct radix_tree_root *radix) =
{ ins_one, insert_dup, del_one, lookup_item,
lookup_enoent, bulk_op, run_commit };
lookup_enoent, bulk_op };
static int fill_radix(struct ctree_root *root, struct radix_tree_root *radix)
{
@ -314,7 +310,6 @@ int print_usage(void)
int main(int ac, char **av)
{
RADIX_TREE(radix, GFP_KERNEL);
struct ctree_super_block super;
struct ctree_root *root;
int i;
int ret;
@ -365,8 +360,7 @@ int main(int ac, char **av)
printf("open & close, root level %d nritems %d\n",
node_level(root->node->node.header.flags),
root->node->node.header.nritems);
write_ctree_super(root, &super);
close_ctree(root);
close_ctree(root, &super);
root = open_ctree("dbfile", &super);
}
while(count--) {
@ -380,7 +374,7 @@ int main(int ac, char **av)
err = ret;
goto out;
}
if (ops[op] == bulk_op || ops[op] == run_commit)
if (ops[op] == bulk_op)
break;
if (keep_running == 0) {
err = 0;
@ -389,8 +383,7 @@ int main(int ac, char **av)
}
}
out:
write_ctree_super(root, &super);
close_ctree(root);
close_ctree(root, &super);
return err;
}