For immutable bio vecs, I've been auditing and removing bi_idx
references. These were harmless, but removing them will make auditing
easier.
scrub_bio_end_io_worker() was open coding a bio_reset() - but this
doesn't appear to have been needed for anything as right after it does a
bio_put(), and perusing the code it doesn't appear anything else was
holding a reference to the bio.
The other use end_bio_extent_readpage() was just for a pr_debug() -
changed it to something that might be a bit more useful.
Signed-off-by: Kent Overstreet <koverstreet@google.com>
CC: Chris Mason <chris.mason@oracle.com>
CC: Stefan Behrens <sbehrens@giantdisaster.de>
When we wrote some data by compress mode into a btrfs filesystem which was full
of the fragments, the kernel will report:
BTRFS warning (device xxx): Aborting unused transaction.
The reason is:
We can not find a long enough free space to store the compressed data because
of the fragmentary free space, and the compressed data can not be splited,
so the kernel outputed the above message.
In fact, btrfs can deal with this problem very well: it fall back to
uncompressed IO, split the uncompressed data into small ones, and then
store them into to the fragmentary free space. So we shouldn't output the
above warning message.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Wade Cline reported a problem where he was getting garbage and warnings when
writing to a preallocated range via O_DIRECT. This is because we weren't
creating our normal pinned extent_map for the range we were writing to,
which was causing all sorts of issues. This patch fixes the problem and
makes his testcase much happier. Thanks,
Reported-by: Wade Cline <clinew@linux.vnet.ibm.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Sage reported the following lockdep backtrace
=====================================
[ BUG: bad unlock balance detected! ]
3.6.0-rc2-ceph-00171-gc7ed62d #1 Not tainted
-------------------------------------
btrfs-cleaner/7607 is trying to release lock (sb_internal) at:
[<ffffffffa00422ae>] btrfs_commit_transaction+0xa6e/0xb20 [btrfs]
but there are no more locks to release!
other info that might help us debug this:
1 lock held by btrfs-cleaner/7607:
#0: (&fs_info->cleaner_mutex){+.+...}, at: [<ffffffffa003b405>] cleaner_kthread+0x95/0x120 [btrfs]
stack backtrace:
Pid: 7607, comm: btrfs-cleaner Not tainted 3.6.0-rc2-ceph-00171-gc7ed62d #1
Call Trace:
[<ffffffffa00422ae>] ? btrfs_commit_transaction+0xa6e/0xb20 [btrfs]
[<ffffffff810afa9e>] print_unlock_inbalance_bug+0xfe/0x110
[<ffffffff810b289e>] lock_release_non_nested+0x1ee/0x310
[<ffffffff81172f9b>] ? kmem_cache_free+0x7b/0x160
[<ffffffffa004106c>] ? put_transaction+0x8c/0x130 [btrfs]
[<ffffffffa00422ae>] ? btrfs_commit_transaction+0xa6e/0xb20 [btrfs]
[<ffffffff810b2a95>] lock_release+0xd5/0x220
[<ffffffff81173071>] ? kmem_cache_free+0x151/0x160
[<ffffffff8117d9ed>] __sb_end_write+0x7d/0x90
[<ffffffffa00422ae>] btrfs_commit_transaction+0xa6e/0xb20 [btrfs]
[<ffffffff81079850>] ? __init_waitqueue_head+0x60/0x60
[<ffffffff81634c6b>] ? _raw_spin_unlock+0x2b/0x40
[<ffffffffa0042758>] __btrfs_end_transaction+0x368/0x3c0 [btrfs]
[<ffffffffa0042808>] btrfs_end_transaction_throttle+0x18/0x20 [btrfs]
[<ffffffffa00318f0>] btrfs_drop_snapshot+0x410/0x600 [btrfs]
[<ffffffff8132babd>] ? do_raw_spin_unlock+0x5d/0xb0
[<ffffffffa00430ef>] btrfs_clean_old_snapshots+0xaf/0x150 [btrfs]
[<ffffffffa003b405>] ? cleaner_kthread+0x95/0x120 [btrfs]
[<ffffffffa003b419>] cleaner_kthread+0xa9/0x120 [btrfs]
[<ffffffffa003b370>] ? btrfs_destroy_delayed_refs.isra.102+0x220/0x220 [btrfs]
[<ffffffff810791ee>] kthread+0xae/0xc0
[<ffffffff810b379d>] ? trace_hardirqs_on+0xd/0x10
[<ffffffff8163e744>] kernel_thread_helper+0x4/0x10
[<ffffffff81635430>] ? retint_restore_args+0x13/0x13
[<ffffffff81079140>] ? flush_kthread_work+0x1a0/0x1a0
[<ffffffff8163e740>] ? gs_change+0x13/0x13
This is because the throttle stuff can commit the transaction, which expects to
be the one stopping the intwrite stuff, but we've already done it in the
__btrfs_end_transaction. Moving the sb_end_intewrite after this logic makes the
lockdep go away. Thanks,
Tested-by: Sage Weil <sage@inktank.com>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
This is the change of the kernel side.
Translation of logical to inode used to have an upper limit 4k on
inode container's size, but the limit is not large enough for a data
with a great many of refs, so when resolving logical address,
we can end up with
"ioctl ret=0, bytes_left=0, bytes_missing=19944, cnt=510, missed=2493"
This changes to regard 64k as the upper limit and use vmalloc instead of
kmalloc to get memory more easily.
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
In logical resolve, we parse extent_from_logical()'s 'ret' as a kind of flag.
It is possible to lose our errors because
(-EXXXX & BTRFS_EXTENT_FLAG_TREE_BLOCK) is true.
I'm not sure if it is on purpose, it just looks too hacky if it is.
I'd rather use a real flag and a 'ret' to catch errors.
Acked-by: Jan Schmidt <list.btrfs@jan-o-sch.net>
Signed-off-by: Liu Bo <liub.liubo@gmail.com>
As ref cache has been removed from btrfs, there is no user on
its lock and its check.
Signed-off-by: Liu Bo <liubo2009@cn.fujitsu.com>
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
When we delete a inode, we will remove all the delayed items including delayed
inode update, and then truncate all the relative metadata. If there is lots of
metadata, we will end the current transaction, and start a new transaction to
truncate the left metadata. In this way, we will leave a inode item that its
link counter is > 0, and also may leave some directory index items in fs/file tree
after the current transaction ends. In other words, the metadata in this fs/file tree
is inconsistent. If we create a snapshot for this tree now, we will find a inode with
corrupted metadata in the new snapshot, and we won't continue to drop the left metadata,
because its link counter is not 0.
We fix this problem by updating the inode item before the current transaction ends.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
I noticed I was seeing large lags when running my torrent test in a vm on my
laptop. While trying to make it lag less I noticed that our overcommit math
was taking into account the number of bytes we wanted to reclaim, not the
number of bytes we actually wanted to allocate, which means we wouldn't
overcommit as often. This patch fixes the overcommit math and makes
shrink_delalloc() use that logic so that it will stop looping faster. We
still have pretty high spikes of latency, but the test now takes 3 minutes
less time (about 5% faster). Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Mitch reported a problem where you could get an ENOSPC error when untarring
a kernel git tree onto a 16gb file system with compress-force=zlib. This is
because compression is a huge pain, it will return from ->writepages()
without having actually created any ordered extents. To get around this we
check to see if the async submit counter is up, and if it is wait until it
drops to 0 before doing our normal ordered wait dance. With this patch I
can now untar a kernel git tree onto a 16gb file system without getting
ENOSPC errors. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
We're going to use this flag EXTENT_DEFRAG to indicate which range
belongs to defragment so that we can implement snapshow-aware defrag:
We set the EXTENT_DEFRAG flag when dirtying the extents that need
defragmented, so later on writeback thread can differentiate between
normal writeback and writeback started by defragmentation.
Original-Signed-off-by: Li Zefan <lizf@cn.fujitsu.com>
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
ulist_alloc() has the possibility of returning NULL.
So, it is necessary to check the return value.
Signed-off-by: Tsutomu Itoh <t-itoh@jp.fujitsu.com>
When we ran fsstress(a program in xfstests), the filesystem hung up when it
is full. It was because the space reserved in btrfs_fallocate() was wrong,
btrfs_fallocate() just used the size of the pre-allocation to reserve the
space, didn't took the block size aligning into account, so the size of
the reserved space was less than the allocated space, it caused the over
reserve problem and made the filesystem hung up when invoking cow_file_range().
Fix it.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Though we dump the stack information when aborting a unused transaction
handle, we don't know the correct place where we decide to abort the
transaction handle if one function has several place where the transaction
abort function is invoked and jumps to the same place after this call.
And beside that we also don't know the reason why we jump to abort
the current handle. So I modify the transaction abort function and make
it output the function name, line and error information.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
We forget to protect ->log_batch when syncing a file, this patch fix
this problem by atomic operation. And ->log_batch is used to check
if there are parallel sync operations or not, so it is unnecessary to
reset it to 0 after the sync operation of the current log tree complete.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
We should insert/update 6 items(root ref, root backref, dir item, dir index,
root item and parent inode) when creating a snapshot, not 5 items, fix it.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
The snapshot should be the image of the fs tree before it was created,
so the metadata of the snapshot should not exist in the its tree. But now, we
found the directory item and directory name index is in both the snapshot tree
and the fs tree. It introduces some problems and makes the users feel strange:
# mkfs.btrfs /dev/sda1
# mount /dev/sda1 /mnt
# mkdir /mnt/1
# cd /mnt/1
# btrfs subvolume snapshot /mnt snap0
# ls -a /mnt/1/snap0/1
. .. [no other file/dir]
# ll /mnt/1/snap0/
total 0
drwxr-xr-x 1 root root 10 Ju1 24 12:11 1
^^^
There is no file/dir in it, but it's size is 10
# cd /mnt/1/snap0/1/snap0
[Enter a unexisted directory successfully...]
There is nothing in the directory 1 in snap0, but btrfs told the length of
this directory is 10. Beside that, we can enter an unexisted directory, it is
very strange to the users.
# btrfs subvolume snapshot /mnt/1/snap0 /mnt/snap1
# ll /mnt/1/snap0/1/
total 0
[None]
# ll /mnt/snap1/1/
total 0
drwxr-xr-x 1 root root 0 Ju1 24 12:14 snap0
And the source of snap1 did have any directory in Directory 1, but snap1 have
a snap0, it is different between the source and the snapshot.
So I think we should insert directory item and directory name index and update
the parent inode as the last step of snapshot creation, and do not leave the
useless metadata in the file tree.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Sometimes we need choose the method of the reservation according to the type
of the block reservation, such as the reservation for the delayed inode update.
Now we identify the type just by comparing the address of the reservation
variants, it is very ugly if it is a temporary one because we need compare it
with all the common reservation variants. So we add a new "type" field to keep
the type the reservation variants.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
The ordered extent allocation is in the fast path of the IO, so use a slab
to improve the speed of the allocation.
"Size of the struct is 280, so this will fall into the size-512 bucket,
giving 8 objects per page, while own slab will pack 14 objects into a page.
Another benefit I see is to check for leaked objects when the module is
removed (and the cache destroy takes place)."
-- David Sterba
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
If a snapshot is created while we are writing some data into the file,
the i_size of the corresponding file in the snapshot will be wrong, it will
be beyond the end of the last file extent. And btrfsck will report:
root 256 inode 257 errors 100
Steps to reproduce:
# mkfs.btrfs <partition>
# mount <partition> <mnt>
# cd <mnt>
# dd if=/dev/zero of=tmpfile bs=4M count=1024 &
# for ((i=0; i<4; i++))
> do
> btrfs sub snap . $i
> done
This because the algorithm of disk_i_size update is wrong. Though there are
some ordered extents behind the current one which we use to update disk_i_size,
it doesn't mean those extents will be dealt with in the same transaction. So
We shouldn't use the offset of those extents to update disk_i_size. Or we will
get the wrong i_size in the snapshot.
We fix this problem by recording the max real i_size. If we find there is a
ordered extent which is in front of the current one and doesn't complete, we
will record the end of the current one into that ordered extent. Surely, if
the current extent holds the end of other extent(it must be greater than
the current one because it is behind the current one), we will record the
number that the current extent holds. In this way, we can exclude the ordered
extents that may not be dealth with in the same transaction, and be easy to
know the real disk_i_size.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
If we create several snapshots at the same time, the following BUG_ON() will be
triggered.
kernel BUG at fs/btrfs/extent-tree.c:6047!
Steps to reproduce:
# mkfs.btrfs <partition>
# mount <partition> <mnt>
# cd <mnt>
# for ((i=0;i<2400;i++)); do touch long_name_to_make_tree_more_deep$i; done
# for ((i=0; i<4; i++))
> do
> mkdir $i
> for ((j=0; j<200; j++))
> do
> btrfs sub snap . $i/$j
> done &
> done
The reason is:
Before transaction commit, some operations changed the fs tree and new tree
blocks were allocated because of COW. We used the implicit non-shared back
reference for those newly allocated tree blocks because they were not shared by
two or more trees.
And then we created the first snapshot for the fs tree, according to the back
reference rules, we also used implicit back refs for the child tree blocks of
the root node of the fs tree, now those child nodes/leaves were shared by two
trees.
Then We didn't deal with the delayed references, and continued to change the fs
tree(created the second snapshot and inserted the dir item of the new snapshot
into the fs tree). According to the rules of the back reference, we added full
back refs for those tree blocks whose parents have be shared by two trees.
Now some newly allocated tree blocks had two types of the references.
As we know, the delayed reference system handles these delayed references from
back to front, and the full delayed reference is inserted after the implicit
ones. So when we dealt with the back references of those newly allocated tree
blocks, the full references was dealt with at first. And if the first reference
is a shared back reference and the tree block that the reference points to is
newly allocated, It would be considered as a tree block which is shared by two
or more trees when it is allocated and should be a full back reference not a
implicit one, the flag of its reference also should be set to FULL_BACKREF.
But in fact, it was a non-shared tree block with a implicit reference at
beginning, so it was not compulsory to set the flags to FULL_BACKREF. So BUG_ON
was triggered.
We have several methods to fix this bug:
1. deal with delayed references after the snapshot is created and before we
change the source tree of the snapshot. This is the easiest and safest way.
2. modify the sort method of the delayed reference tree, make the full delayed
references be inserted before the implicit ones. It is also very easy, but
I don't know if it will introduce some problems or not.
3. modify select_delayed_ref() and make it select the implicit delayed reference
at first. This way is not so good because it may wastes CPU time if we have
lots of delayed references.
4. set the flags to FULL_BACKREF, this method is a little complex comparing with
the 1st way.
I chose the 1st way to fix it.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
This patch fixes the following problem:
- If we failed to deal with the delayed dir items, we should abort transaction,
just as its comment said. Fix it.
- If root reference or root back reference insertion failed, we should
abort transaction. Fix it.
- Fix the double free problem of pending->inherit.
- Do not restore the trans->rsv if we doesn't change it.
- make the error path more clearly.
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
bbio has been malloced in btrfs_map_block() and should be
freed before leaving from the error handling cases.
spatch with a semantic match is used to found this problem.
(http://coccinelle.lip6.fr/)
Signed-off-by: Wei Yongjun <yongjun_wei@trendmicro.com.cn>
I noticed this when I was doing the fsync stuff, we allocate split extents if we
drop an extent range that is in the middle of an existing extent. This BUG()'s
if we fail to allocate memory, but the fact is this is just a cache, we will
just regenerate the cache if we need it, the important part is that we free the
range we are given. This can be done without allocations, so if we fail to
allocate splits just skip the splitting stage and free our em and look for more
extents to drop. This also makes btrfs_drop_extent_cache a void since nobody
was checking the return value anyway. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
The freeze rwsem is taken by sb_start_intwrite() and dropped during the
commit_ or end_transaction(). In the async case, that happens in a worker
thread. Tell lockdep the calling thread is releasing ownership of the
rwsem and the async thread is picking it up.
XFS plays the same trick in fs/xfs/xfs_aops.c.
Signed-off-by: Sage Weil <sage@inktank.com>
I audited all users of btrfs_drop_extents and found that nobody actually uses
the hint_byte argument. I'm sure it was used for something at some point but
it's not used now, and the way the pinning works the disk bytenr would never be
immediately useful anyway so lets just remove it. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
This is based on Josef's "Btrfs: turbo charge fsync".
If an inode is a BTRFS_INODE_NODATASUM one, we don't need to look for csum
items any more.
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
This is based on Josef's "Btrfs: turbo charge fsync".
The current btrfs checks if an inode is in log by comparing
root's last_log_commit to inode's last_sub_trans[2].
But the problem is that this root->last_log_commit is shared among
inodes.
Say we have N inodes to be logged, after the first inode,
root's last_log_commit is updated and the N-1 remained files will
be skipped.
This fixes the bug by keeping a local copy of root's last_log_commit
inside each inode and this local copy will be maintained itself.
[1]: we regard each log transaction as a subset of btrfs's transaction,
i.e. sub_trans
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
This is based on Josef's "Btrfs: turbo charge fsync".
The above Josef's patch performs very good in random sync write test,
because we won't have too much extents to merge.
However, it does not performs good on the test:
dd if=/dev/zero of=foobar bs=4k count=12500 oflag=sync
The reason is when we do sequencial sync write, we need to merge the
current extent just with the previous one, so that we can get accumulated
extents to log:
A(4k) --> AA(8k) --> AAA(12k) --> AAAA(16k) ...
So we'll have to flush more and more checksum into log tree, which is the
bottleneck according to my tests.
But we can avoid this by telling fsync the real extents that are needed
to be logged.
With this, I did the above dd sync write test (size=50m),
w/o (orig) w/ (josef's) w/ (this)
SATA 104KB/s 109KB/s 121KB/s
ramdisk 1.5MB/s 1.5MB/s 10.7MB/s (613%)
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
We will stop and restart a transaction every time we move to a different leaf
when truncating a file. This is for enospc reasons, but really we could
probably get away with doing this a little better by actually working until we
hit an ENOSPC. So add a ->failfast flag to the block_rsv and set it when we do
truncates which will fail as soon as the block rsv runs out of space, and then
at that point we can stop and restart the transaction and refill the block rsv
and carry on. This will make rm'ing of a file with lots of extents a bit
faster. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
This is based on Josef's "Btrfs: turbo charge fsync".
We should cleanup those extents after we've finished logging inode,
otherwise we may do redundant work on them.
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
I hit this a couple times while working on my fsync patch (all my bugs, not
normal operation), but with my new stuff we could have new errors from cases
I have not encountered, so instead of BUG()'ing we should be WARN()'ing so
that we are notified there is a problem but the user doesn't lose their
data. We can easily commit the transaction in the case that the tree
logging fails and still be fine, so let's try and be as nice to the user as
possible. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
At least for the vm workload. Currently on fsync we will
1) Truncate all items in the log tree for the given inode if they exist
and
2) Copy all items for a given inode into the log
The problem with this is that for things like VMs you can have lots of
extents from the fragmented writing behavior, and worst yet you may have
only modified a few extents, not the entire thing. This patch fixes this
problem by tracking which transid modified our extent, and then when we do
the tree logging we find all of the extents we've modified in our current
transaction, sort them and commit them. We also only truncate up to the
xattrs of the inode and copy that stuff in normally, and then just drop any
extents in the range we have that exist in the log already. Here are some
numbers of a 50 meg fio job that does random writes and fsync()s after every
write
Original Patched
SATA drive 82KB/s 140KB/s
Fusion drive 431KB/s 2532KB/s
So around 2-6 times faster depending on your hardware. There are a few
corner cases, for example if you truncate at all we have to do it the old
way since there is no way to be sure what is in the log is ok. This
probably could be done smarter, but if you write-fsync-truncate-write-fsync
you deserve what you get. All this work is in RAM of course so if your
inode gets evicted from cache and you read it in and fsync it we'll do it
the slow way if we are still in the same transaction that we last modified
the inode in.
The biggest cool part of this is that it requires no changes to the recovery
code, so if you fsync with this patch and crash and load an old kernel, it
will run the recovery and be a-ok. I have tested this pretty thoroughly
with an fsync tester and everything comes back fine, as well as xfstests.
Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
While working on my fsync patch my fsync tester kept hitting mismatching
md5sums when I would randomly write to a prealloc'ed region, syncfs() and
then write to the prealloced region some more and then fsync() and then
immediately reboot. This is because the tree logging code will skip writing
csums for file extents who's generation is less than the current running
transaction. When we mark extents as written we haven't been updating their
generation so they were always being skipped. This wouldn't happen if you
were to preallocate and then write in the same transaction, but if you for
example prealloced a VM you could definitely run into this problem. This
patch makes my fsync tester happy again. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Swinging this pendulum back the other way. We've been allocating chunks up
to 2% of the disk no matter how much we actually have allocated. So instead
fix this calculation to only allocate chunks if we have more than 80% of the
space available allocated. Please test this as it will likely cause all
sorts of ENOSPC problems to pop up suddenly. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
There is a completely impossible situation to hit where you can preallocate
a file, fsync it, write into the preallocated region, have the transaction
commit twice and then fsync and then immediately lose power and lose all of
the contents of the write. This patch fixes this just so I feel better
about the situation and because it is lightweight, we just update the
last_trans when we finish an ordered IO and we don't update the inode
itself. This way we are completely safe and I feel better. Thanks,
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
The btrfs send code was assuming the offset of the file item into the
extent translated to bytes on disk. If we're compressed, this isn't
true, and so it was off into extents owned by other files.
It was also improperly handling inline extents. This solves a crash
where we may have gone past the end of the file extent item by not
testing early enough for an inline extent. It also solves problems
where we have a whole between the end of the inline item and the start
of the full extent.
Signed-off-by: Chris Mason <chris.mason@fusionio.com>
We can't do the deleted/reused logic for top/root inodes as it would
create a stream that tries to delete and recreate the root dir.
Reported-by: Alex Lyakas <alex.bolshoy.btrfs@gmail.com>
Signed-off-by: Alexander Block <ablock84@googlemail.com>
We have to ignore inode/space cache objects in send/receive.
Reported-by: Alex Lyakas <alex.bolshoy.btrfs@gmail.com>
Signed-off-by: Alexander Block <ablock84@googlemail.com>
We need to pass the root that we determined earlier to iterate_inode_ref.
Reported-by: Alex Lyakas <alex.bolshoy.btrfs@gmail.com>
Signed-off-by: Alexander Block <ablock84@googlemail.com>
The previous check was working fine, but this check should be
easier to read. Also, we could theoritically have some exotic
bugs with the previous checks.
Signed-off-by: Alexander Block <ablock84@googlemail.com>
Kent Overstreet
Miao Xie
Josef Bacik
Liu Bo
Liu Bo
liubo
David Sterba
Tsutomu Itoh
Wei Yongjun
Sage Weil
Jan Schmidt
Chris Mason
Alexander Block