{:toc}
Extent tree is a well used structure in filesystem, it is to replace the legacy indirect mapping method to look for a physical block mapped by a logical block. Here I'll take ext4 filesystem as an instance to begin the story. If you are not familiar with the legacy indirect mapping method, you can refer to this article: dentry_inode.md
Firstly, let's begin with two questions, with the indirect mapping method:
- how many extra blocks (for storing indirect blocks) we need for a 4TB file?
- how many times do we do the logical block-->physical block translation when we write to a 4TB file.
Assume the block size is 4KB. For the first one, we know the inode->i_block[14] support 3 layers of indirect blocks. So the number of indirect blocks are 1 + 2^10 + 2^10 * 2^10 ~= 2^20 blocks. That's a huge storage overhead. For the second one, we have to do the translation for each logical block, so around 2^30 times in total.
Hmm, seems not very good, right? Think about a case, all the physical blocks of this 4TB file are all contiguous. In such a case, we just need one info like: (lblk, pblk, len) where lblk = logical block number, pblk = physical block number and len is the length of contiguous blocks. Then we just need one this kind of struct to describe the relationship between logical block and physical block, which is (0, x, 4TB), where x is a specific physical block number. And for the translation, it becomes very simple, we can just do one visit to get all the mapping details.
Here the key is we leverage the truth that for a file, contiguous logical blocks can be contiguous in physical blocks too. For those logical blocks, we don't need to store the mapping data separately. On the contrary, we use a single info (lblk, pblk, len) to describe the relationship. And this is called a extent.
The extents are organized as a B+ tree on disk. the root is in inode->i_block[]. A node of this B tree is made up of a header and a body. The body consists of several entries, either is ext4_extent_idx or ext4_extent. The former is for non-leaf nodes, and the latter is for leaf nodes. Both the header and an entry is 12Bytes. Here let's recall the inode->i_block[], it has 15 items with each one 4Bytes, so 15 * 4 = 60Bytes in total. It can store 60 / 12 = 5 entries. Minus the header, it can store 4 entries there. So if a file has <= 4 extents, we just put the extents in inode->i_block[].
Otherwise, the entries in i_block[] are ext4_extent_idx, and point to extent blocks. extent blocks are data blocks that are full of entries. here please notice onething: a data block(extent block) is a B+ tree node, and entries in it all belong to this node.
Let's first have a overlook on it:
+-------->+---extent block-----+ +-->+---data blocks--+
| | ext4_extent_header | / | |
| +--------------------+ / | |
| | ext4_extent_idx | / +----------------+
| +--------------------| / | |
| | ext4_extent_idx | / | |
| +--------------------| / +----------------+
| | ............... | +---extent block-----+ / | |
| +--------------------+ +--->| ext4_extent_header | / | |
| | ext4_extent_idx | | +--------------------+ / +----------------+
| +--------------------+ | | ext4_extent |---------+ | |
| | ext4_extent_tail | | +--------------------+ \ | |
| +--------------------+ | | ext4_extent | \ +----------------+
| | +--------------------+ \ | |
| | | ............... | \ | |
+------------------------+ | +---->+---extent block-----+ | +--------------------+ +------->+----------------+
| ext4_inode | | | | ext4_extent_header | | | ext4_extent |-------+ | |
+------------------------+ | | +--------------------+ | +--------------------+ |\ | |
| +------i_block[]-----+ | | | | ext4_extent_idx |----+ | ext4_extent_tail | | +------------>+----------------+
| | ext4_extent_header | | | | +--------------------+ +--------------------+ | | |
| +--------------------+ | | | | ext4_extent_idx | | | |
| | ext4_extent_idx |-|--------+ | +--------------------+ \ +----------------+
| +--------------------+ | | | ............... | \ | |
| | ext4_extent_idx |-|------------+ +--------------------+ \ | |
| +--------------------+ | | ext4_extent_idx |-------->+---extent block-----+ +----------->+----------------+
| | ext4_extent_idx |-|------------+ +--------------------+ | ext4_extent_header | | |
| +--------------------+ | | | ext4_extent_tail | +--------------------+ | |
| | ext4_extent_idx |-|--------+ | +--------------------+ | ext4_extent | +----------------+
| +--------------------+ | | | +--------------------+ | |
+------------------------+ | +---->+---extent block-----+ | ext4_extent | | |
| | ext4_extent_header | +--------------------+ +----------------+
| +--------------------+ | ............... |
| | ext4_extent_idx | +--------------------+
| +--------------------+ | ext4_extent |
| | ext4_extent_idx | +--------------------+
| +--------------------+ | ext4_extent_tail |
| | ............... | +--------------------+
| +--------------------+
| | ext4_extent_idx |
| +--------------------+
| | ext4_extent_tail |
| +--------------------+
|
+-------->+---extent block-----+
| ext4_extent_header |
+--------------------+
| ext4_extent_idx |
+--------------------+
| ext4_extent_idx |
+--------------------+
| ............... |
+--------------------+
| ext4_extent_idx |
+--------------------+
| ext4_extent_tail |
+--------------------+All structures:
struct ext4_extent_tail {
__le32 et_checksum; /* crc32c(uuid+inum+extent_block) */
};
struct ext4_extent {
__le32 ee_block; /* first logical block extent covers */
__le16 ee_len; /* number of blocks covered by extent */
__le16 ee_start_hi; /* high 16 bits of physical block */
__le32 ee_start_lo; /* low 32 bits of physical block */
};
struct ext4_extent_idx {
__le32 ei_block; /* index covers logical blocks from 'block' */
__le32 ei_leaf_lo; /* pointer to the physical block of the next *
* level. leaf or next index could be there */
__le16 ei_leaf_hi; /* high 16 bits of physical block */
__u16 ei_unused;
};
struct ext4_extent_header {
__le16 eh_magic; /* probably will support different formats */
__le16 eh_entries; /* number of valid entries */
__le16 eh_max; /* capacity of store in entries */
__le16 eh_depth; /* has tree real underlying blocks? */
__le32 eh_generation; /* generation of the tree */
};-
ext4_extent_header
eh_depthis counted from 0 and from leaf to root, so leaf->eh_depth = 0 and root->eh_depth is the maximum one. -
ext4_extent_idx ei_leaf_lo, ei_leaf_hi together stand for the physical block number of next layer B+ tree node. ei_block is the logical block number. Given the current logical block is x, how can we find leaf from it? For example: the current non-leaf node where we are is:
[lblk0, ..][lblk1, ..][lblk2, ..][lblk3, ..]We can compare x with lblk0, lbk1,2 and 3 to see which idx it locates and then go to the next layer, to be more accurate, we can do binary search here (this is basic knowledge of B+ tree, you can refer to wiki or some data structure materials) -
ext4_extent_tail Just a checksum. Note: we don't need this for i_block in inode, since inode is checked somewhere else.
-
ext4_extent This is the entry in leaf node.
Now that we are clear about the ondisk layout, let's turn to the memory. Like inode and dentry, extent also has memory cache, it's called extent status tree.
/*
* fourth extended file system inode data in memory
*/
struct ext4_inode_info {
struct rw_semaphore i_data_sem;
struct inode vfs_inode;
struct jbd2_inode *jinode;
/* extents status tree */
struct ext4_es_tree i_es_tree;
rwlock_t i_es_lock;
struct list_head i_es_list;
unsigned int i_es_all_nr; /* protected by i_es_lock */
unsigned int i_es_shk_nr; /* protected by i_es_lock */
ext4_lblk_t i_es_shrink_lblk; /* Offset where we start searching for
extents to shrink. Protected by
i_es_lock */
};ext4_inode_info is the memory version ext4_inode. i_es_tree represents the extent status tree.
struct ext4_es_tree {
struct rb_root root;
struct extent_status *cache_es; /* recently accessed extent */
};The es tree is organized as a red black tree. cache_es is the recent visited item, AKA a cache of cache. A node in this rb tree is:
struct extent_status {
struct rb_node rb_node;
ext4_lblk_t es_lblk; /* first logical block extent covers */
ext4_lblk_t es_len; /* length of extent in block */
ext4_fsblk_t es_pblk; /* first physical block */
};Each node is a cache of a disk extent. The rb tree is sorted by es_lblk----the start block address this extent covers.
Ok, the name <extent status> may be confusing, the word 'status' here means one of:
#define EXTENT_STATUS_WRITTEN (1 << ES_WRITTEN_B)
#define EXTENT_STATUS_UNWRITTEN (1 << ES_UNWRITTEN_B)
#define EXTENT_STATUS_DELAYED (1 << ES_DELAYED_B)
#define EXTENT_STATUS_HOLE (1 << ES_HOLE_B)
#define EXTENT_STATUS_REFERENCED (1 << ES_REFERENCED_B)The status is represented by . Now let's look into each of them, this is related with the initative of the es tree.
-
WRITTENpages exists in page cache, ondisk blocks exists too.this is the normal case, an es node is a cache version of the ondisk extent
-
UNWRITTENpages exists in page cache, ondisk blocks exists but not filled (they contains junk data).this is also cache version of an ondisk extent. The difference is the ondisk data blocks hasn't been touched yet. This is useful for syscall like fallocate(). fallocate() allows you to allocate blocks to a file without filling zeroes. Think about the situation when there is no es tree: if you want to generate a file of 4GB, you have to fill all the 4GB blocks to zeroes. With es tree, you can just set up metadata like the extent and extent status, and mark it as UNWRITTEN, and allocate data blocks in block map, no need to do any data IO.
-
HOLEno pages in page cache, and no ondisk blocks either.stands for a hole in a file, in this case, the pblk is meaningless. The file logical range doesn't have it's physical range ondisk.
-
DELAYEDpages exits in page cache, no ondisk blocks. (This is why es tree is invented)This is used by a feature called delay allocation. Normally, when a buffered write request writes some data to the page cache and expands the file size or fill a hole area, it should allocate blocks ondisk immediately. But actually we can delay the allocation to the writeback period, this will improve the latency. Think about the situation of no es tree: to distinguish HOLE and DELAYED extent, we have to look into the page cache to see if there are pages in the range. That makes things complicated and may cause bugs. (comments are welcome here: why it's complicated? I just get this from mail list) With this flag, we can do all these by just checking this flag. The advantage of this feature is we can gether neighbor logical blocks and in the writeback period we try to allocate contiguous physical blocks to get better performance.
-
REFERENCEDundiscovered. TOBEDONE.
The above flags are stored in the highest (five) bits in extent_status->es_pblk.
A write request is like write(fd, offset, buf, len), means writing data [buf, buf+len) to position [offset, offset+len) in file fd. From [offset, offset+len) we can get the logical blocks range [offset/blk_sz, (offset+len)/blk_sz). After we get the inode of this file, we have to translate the logical range to physical ranges.
Pick up ext4 direct write as an example, let's see how the translation goes. As this is highly related with iomap, please read IOMAP first, there is a big picture which describes the whole process of ext4 dio write. Here let's focus on the extent part. As being said in IOMAP, iomap_begin() does the translation, which is ext4_iomap_begin() in ext4 filesystem.
ext4_iomap_begin() is simple:
- it defines a struct ext4_map_blocks and initializes its members, then calls ext4_map_blocks().
struct ext4_map_blocks {
ext4_fsblk_t m_pblk;
ext4_lblk_t m_lblk;
unsigned int m_len;
unsigned int m_flags;
};struct ext4_map_blocks is used to store a mapping, you can see it as 'ext4's private struct iomap'. And it indeed will deliver its info to struct iomap at last.
-
ext4_map_blocks() this is the main dish, we'll dive into it later.
-
ext4_set_iomap() After we get the mapping, we need to fill the struct iomap by struct ext4_map_blocks. This basically involves setting up:
- iomap->flags
/* * Writes that span EOF might trigger an I/O size update on completion, * so consider them to be dirty for the purpose of O_DSYNC, even if * there is no other metadata changes being made or are pending. */ iomap->flags = 0; if (ext4_inode_datasync_dirty(inode) || offset + length > i_size_read(inode)) iomap->flags |= IOMAP_F_DIRTY; if (map->m_flags & EXT4_MAP_NEW) iomap->flags |= IOMAP_F_NEW;
TOBEDONE: analyze IOMAP_F_DIRTY and IOMAP_F_NEW.- iomap->bdev
if (flags & IOMAP_DAX) iomap->dax_dev = EXT4_SB(inode->i_sb)->s_daxdev; else iomap->bdev = inode->i_sb->s_bdev;
- iomap->offset && iomap->length
iomap->offset = (u64) map->m_lblk << blkbits; iomap->length = (u64) map->m_len << blkbits;
- iomap->type && iomap->addr
if (map->m_flags & EXT4_MAP_UNWRITTEN) { iomap->type = IOMAP_UNWRITTEN; iomap->addr = (u64) map->m_pblk << blkbits; if (flags & IOMAP_DAX) iomap->addr += EXT4_SB(inode->i_sb)->s_dax_part_off; } else if (map->m_flags & EXT4_MAP_MAPPED) { iomap->type = IOMAP_MAPPED; iomap->addr = (u64) map->m_pblk << blkbits; if (flags & IOMAP_DAX) iomap->addr += EXT4_SB(inode->i_sb)->s_dax_part_off; } else { iomap->type = IOMAP_HOLE; iomap->addr = IOMAP_NULL_ADDR; }
I have to say that I haven't figured out all the detail of it. I'll explain it as possible as I can. Generally speaking, ext4_map_blocks() does two things:
- look up the mapping
- fast path: look it up in es tree
- slow path: look it up in extent tree
- look up needed blocks (extent node) in per cpu array bh_lrus
- look up needed blocks in the page cache of the device file
- page cache miss, load needed blocks from disk(I'm not sure here)
- handle the mapping
This one is super simple, we give it the desired logical block number, it search the rb tree for the right node. I'm not gona do code analysis for it here, the code explains itself well.
After get the extent, we can update struct ext4_map_blocks now. This includes
map->pblk, map->m_flags and map->m_len. Here we must be clear about the m_flags:
EXTENT_STATUS_WRITTEN --> EXT4_MAP_MAPPED
EXTENT_STATUS_UNWRITTEN --> EXT4_MAP_UNWRITTEN
EXTENT_STATUS_DELAYED --> ?
EXTENT_STATUS_HOLE --> ?
If we fail to find the target extent in es tree, we have to search the extent tree. And we have to hold the rw semaphore ext4_inode_info->i_data_sem.
in ext4_map_blocks():
down_read(&EXT4_I(inode)->i_data_sem);
if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
retval = ext4_ext_map_blocks(handle, inode, map, 0);
} else {
retval = ext4_ind_map_blocks(handle, inode, map, 0);
}
One thing we can obviously see is there are two ways to organize the mapping. One is the legacy mode----indirect mapping, the other one is the extent tree. And a flag EXT4_INODE_EXTENTS in inode controls which way to use. (So theoretically a filesystem can have two methods in use at the same time?)
Ok, let's jump into the function. I'll list the main operation it does in order.
path = ext4_find_extent(inode, map->m_lblk, NULL, 0);the first thing ext4_ext_map_blocks() does is searching the target extent in the extent tree.the return value is an arraystruct ext4_ext_path { ext4_fsblk_t p_block; // physical block number __u16 p_depth; __u16 p_maxdepth; struct ext4_extent *p_ext; // to the extent if it's leaf struct ext4_extent_idx *p_idx; // to the idx node if it's non-leaf struct ext4_extent_header *p_hdr; // to the extent header struct buffer_head *p_bh; // to the block };
path[]struct ext4_ext_path *pathrecords the path from root to leaf during the walking. Meaning of members are clear, here I want to mentionstruct buffer_head. It stands for a disk block, and is used to manage the block. Here it points to the corresponding block which is a extent node. Notice that this block is in a page in the blkdev page cache. The detail of buffer_head is complex, I've wrote another article for it. buffer_head It's ok not to know the detail of buffer_head, you just need to know that page cache consists of pages, and those pages each contains one or more blocks(if block size < page size). These blocks are caches of filesystem's ondisk blocks.
This function searches the ondisk extent tree to get the desired extent. But it doesn't load blocks from disk directly. Remember in bdev filesystem we know the blkdev file also has page cache.
-
path = kcalloc(depth + 2, sizeof(struct ext4_ext_path), gfp_flags);The first thing is allocating memory for path[]. Notice the size isn't depth+1 but depth+2 since the B+ tree node may be splited and finally get a new layer
-
path[0].p_hdr = eh;Init p_hdr for path[0], and it iseh = ext_inode_hdr(inode);It's easy to get extent header from inode, see [Ondisk struture](## Ondisk structure) section. -
i = depthi is used as an iterator in the later loop which is the main logic of this function depth isdepth = ext_depth(inode);which finally come fromext4_extent_header->eh_depth(remember the leaf has depth = 0 and the root has the max depth, so hdr->eh_depth means the depth of the tree whose root is the node where hdr locates)This and the previous step are to init the start point
path[0]of the tree walking -
A loop walking though the extent tree
while (i) { ext_debug(inode, "depth %d: num %d, max %d\n", ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max)); ext4_ext_binsearch_idx(inode, path + ppos, block); path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx); path[ppos].p_depth = i; path[ppos].p_ext = NULL; bh = read_extent_tree_block(inode, path[ppos].p_idx, --i, flags); if (IS_ERR(bh)) { ret = PTR_ERR(bh); goto err; } eh = ext_block_hdr(bh); ppos++; path[ppos].p_bh = bh; path[ppos].p_hdr = eh; }
Ok, this is a typical B+ tree walking, the idea is to do binary search at each node to find the right child to go, and then go to that node and loop again.
-
ext4_ext_binsearch_idx()This is just a simple binary search. -
path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx);Just like struct ext4_extent, path[ppos].p_block points to the idx node. -
bh = read_extent_tree_block(inode, path[ppos].p_idx, --i, flags);In the current nodepath[ppos].p_bh(for root, it is inode of course), we've found the child node we should go is path[ppos].p_idx but the block of that node may not be in memory, if that's the case we have to load it from disk first,read_extent_tree_block()is for that. -
The left code Now we get the buffer_head of next extent node, we should init path[++ppos], that is what the left code does.
-
It is a marco, which actually calls function __read_extent_tree_block, and the latter
is the same as the former, just add two parameters __func__ and __LINE__ for debugging.
static struct buffer_head *
__read_extent_tree_block(const char *function, unsigned int line,
struct inode *inode, struct ext4_extent_idx *idx,
int depth, int flags)
{
struct buffer_head *bh;
int err;
gfp_t gfp_flags = __GFP_MOVABLE | GFP_NOFS;
ext4_fsblk_t pblk;
if (flags & EXT4_EX_NOFAIL)
gfp_flags |= __GFP_NOFAIL;
pblk = ext4_idx_pblock(idx);
bh = sb_getblk_gfp(inode->i_sb, pblk, gfp_flags);
if (unlikely(!bh))
return ERR_PTR(-ENOMEM);
if (!bh_uptodate_or_lock(bh)) {
trace_ext4_ext_load_extent(inode, pblk, _RET_IP_);
err = ext4_read_bh(bh, 0, NULL);
if (err < 0)
goto errout;
}
if (buffer_verified(bh) && !(flags & EXT4_EX_FORCE_CACHE))
return bh;
err = __ext4_ext_check(function, line, inode, ext_block_hdr(bh),
depth, pblk, le32_to_cpu(idx->ei_block));
if (err)
goto errout;
set_buffer_verified(bh);
/*
* If this is a leaf block, cache all of its entries
*/
if (!(flags & EXT4_EX_NOCACHE) && depth == 0) {
struct ext4_extent_header *eh = ext_block_hdr(bh);
ext4_cache_extents(inode, eh);
}
return bh;
errout:
put_bh(bh);
return ERR_PTR(err);
}bh = sb_getblk_gfp(inode->i_sb, pblk, gfp_flags);Get the buffer_head associated with physical block pblk. The way is searching the corresbonding blkdev file page cache. Once we get the page, we get all the buffer_head associated with it by page->private. If the page is not in page cache, we have to create it. The detail is in another article, please refer to buffer_head