Hi Alexander, your solution is simple and elegant. I this this issue is solved now. Thanks! Two minor issues: 1) /* * We need some special handling for inodes that get processed before the parent * directory got created. See process_all_refs for details. * This function does the check if we already created the dir out of order. */ /* * Only creates the inode if it is: * 1. Not a directory * 2. Or a directory which was not created already due to out of order * directories. See did_create_dir and process_all_refs for details. */
These comments tell to look at process_all_refs(), while we should
look at process_recorded_refs().
2)
* We can however not delete the orphan in case the inode relies
* in a directory that was not created yet (see
* __record_new_ref)
*/
This part should be removed, because your new solution does not work this way.
If you find, time, pls look at the two attached scripts.
btrfs_test_1.sh:
it tries to explore the is_first_ref() issue and founds a problem.
Proposed patch - compare also the (dir,gen) tuple and only the name:
diff --git a/fs/btrfs/send.c b/fs/btrfs/send.c
index 68e504c..b83ec5f 100644
--- a/fs/btrfs/send.c
+++ b/fs/btrfs/send.c
@@ -1597,7 +1597,7 @@ out:
static int is_first_ref(struct send_ctx *sctx,
struct btrfs_root *root,
- u64 ino, u64 dir,
+ u64 ino, u64 dir, u64 dir_gen,
const char *name, int name_len)
{
int ret;
@@ -1613,6 +1613,11 @@ static int is_first_ref(struct send_ctx *sctx,
if (ret < 0)
goto out;
+ if (dir != tmp_dir || dir_gen != tmp_dir_gen) {
+ ret = 0;
+ goto out;
+ }
+
if (name_len != fs_path_len(tmp_name)) {
ret = 0;
goto out;
@@ -2834,7 +2839,7 @@ verbose_printk("btrfs: process_recorded_refs
%llu\n", sctx->cur_ino);
goto out;
if (ret) {
ret = is_first_ref(sctx, sctx->parent_root,
- ow_inode, cur->dir, cur->name,
+ ow_inode, cur->dir,
cur->dir_gen, cur->name,
cur->name_len);
if (ret < 0)
goto out;
btrfs_test_2.sh
The last test exposes an interesting issue: when a directory has a
deleted reference recorded, this deleted reference is not added to the
'check_dirs' list. As a result, the upper directory (already
orphanized) is not rmdir'd.
Thanks,
Alex.
On Fri, Jul 27, 2012 at 12:42 AM, Alexander Block
<abloc...@googlemail.com> wrote:
> I have pushed a for-alex branch to github with a new approach for the
> whole problem. Can you test this?
>
> On Thu, Jul 26, 2012 at 4:07 PM, Alexander Block
> <abloc...@googlemail.com> wrote:
>> I'm currently working on another solution for the initial problem. I
>> will create a for-alex branch for you to test later.
>>
>> On Thu, Jul 26, 2012 at 4:04 PM, Alex Lyakas
>> <alex.bolshoy.bt...@gmail.com> wrote:
>>> Alexander,
>>> (pls let me know when this gets annoying:).
>>>
>>> Parent:
>>> /mnt/src/v2_snap0/
>>> └── [ 257] file1
>>>
>>> Send:
>>> /mnt/src/v2_snap1
>>> └── [ 259] dir1
>>> └── [ 258] dir2
>>> └── [ 257] file1
>>>
>>> I encountered two problems:
>>> 1) process_recorded_refs_if_needed() if needed does not call
>>> process_recorded_refs() if both new_refs and deleted_refs() are empty.
>>> But in this case, we need to get to finish_outoforder_dir() by dir2 to
>>> move file1 under it (this is before dir1 is created).
>>>
>>> @@ -4199,8 +4227,25 @@ static int
>>> process_recorded_refs_if_needed(struct send_ctx *sctx, int at_end)
>>> if (!at_end && sctx->cur_ino == sctx->cmp_key->objectid &&
>>> sctx->cmp_key->type <= BTRFS_INODE_REF_KEY)
>>> goto out;
>>> - if (list_empty(&sctx->new_refs) && list_empty(&sctx->deleted_refs))
>>> - goto out;
>>> + if (list_empty(&sctx->new_refs) && list_empty(&sctx->deleted_refs)
>>> &&
>>> + /*
>>> + * If this is a new directory, still do the
>>> finish_outoforder_dir() thing,
>>> + * even though it has no references recorded. This
>>> means that the directory's
>>> + * parent has higher inode and was not created yet
>>> (thus we should have
>>> + * sctx->cur_inode_first_ref_orphan flag set).
>>> + * Note that after a call to process_recorded_refs(),
>>> new_refs and deleted_refs
>>> + * become empty, which prevents further calls to
>>> process_recorded_refs(),
>>> + * but here we need something else to prevent it, so
>>> look at send_progress too.
>>> + */
>>> + !(S_ISDIR(sctx->cur_inode_mode) && sctx->cur_inode_new &&
>>> + sctx->cur_inode_first_ref_orphan &&
>>> sctx->send_progress == sctx->cur_ino))
>>> + goto out;
>>>
>>> ret = process_recorded_refs(sctx);
>>>
>>> Then I encountered another problem that finish_outoforder_dir() does
>>> not check for itself the cur_inode_first_ref_orphan flag:
>>> @@ -2736,7 +2754,17 @@ static int finish_outoforder_dir(struct
>>> send_ctx *sctx, u64 dir, u64 dir_gen)
>>> }
>>> fctx.dir_ino = dir;
>>>
>>> - ret = get_cur_path(sctx, dir, dir_gen, fctx.dir_path,
>>> 1/*do_print*/);
>>> + /*
>>> + * If the current directory itself has a parent, which was not
>>> + * created yet, we need to use gen_unique_name().
>>> + */
>>> + BUG_ON(sctx->cur_ino != dir || sctx->cur_inode_gen != dir_gen);
>>> + if (sctx->cur_inode_first_ref_orphan)
>>> + ret = gen_unique_name(sctx, dir, dir_gen, fctx.dir_path);
>>> + else
>>> + ret = get_cur_path(sctx, dir, dir_gen, fctx.dir_path);
>>>
>>> Finally, the send_truncate(), send_chmod(), send_chown(),send_utimes()
>>> need the following check:
>>>
>>> if (sctx->cur_ino == ino && sctx->cur_inode_first_ref_orphan) {
>>> WARN_ON(sctx->cur_inode_gen != gen);
>>> ret = gen_unique_name(sctx, ino, gen, p);
>>> } else {
>>> ret = get_cur_path(sctx, ino, gen, p);
>>> }
>>>
>>> All of them except utimes() are used with cur_ino only, so for those
>>> this check is redundant (and probably makes sense to drop ino/gen
>>> parameters of them?).
>>>
>>> Thanks,
>>> Alex.
>>>
>>>
>>> On Thu, Jul 26, 2012 at 1:52 PM, Alex Lyakas
>>> <alex.bolshoy.bt...@gmail.com> wrote:
>>>> Hi Alexander,
>>>> I did some very initial testing, and there is still an issue.
>>>> The logic of finish_outoforder_dir works as expected. But then problem
>>>> is that later, when we process xattr/extents or finish the inode, the
>>>> code still uses get_cur_path(), which brings an incorrect name.
>>>>
>>>> Consider the following simple scenario:
>>>>
>>>> Parent tree:
>>>> /mnt/src/v2
>>>> └── [ 260] file1
>>>>
>>>> Send tree:
>>>> /mnt/src/v2
>>>> └── [ 262] dir1
>>>> └── [ 260] file1
>>>>
>>>> So when file1 is being processed, it is first renamed, as expected:
>>>> C_RENAME: A_PATH=file1, A_PATH_TO=o260-511-0
>>>> But then, when we finish it, we do:
>>>> C_TRUNCATE: A_PATH=o262-517-0/file1, A_SIZE=16
>>>>
>>>> So in some functions like send_truncate(), send_write(), send_utimes()
>>>> etc, we need:
>>>>
>>>> - ret = get_cur_path(sctx, ino, gen, p, 0/*do_print*/);
>>>> + if (sctx->cur_inode_first_ref_orphan)
>>>> + ret = gen_unique_name(sctx, ino, gen, p);
>>>> + else
>>>> + ret = get_cur_path(sctx, ino, gen, p, 0/*do_print*/);
>>>> if (ret < 0)
>>>> goto out;
>>>>
>>>> I will continue testing more complicated cases now.
>>>>
>>>> Thanks,
>>>> Alex.
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>> On Tue, Jul 24, 2012 at 11:26 PM, Alexander Block
>>>> <abloc...@googlemail.com> wrote:
>>>>> On Wed, Jul 18, 2012 at 7:45 PM, Alex Lyakas
>>>>> <alex.bolshoy.bt...@gmail.com> wrote:
>>>>>> Hi Alexander,
>>>>>> I am testing different scenarios in order to better understand the
>>>>>> non-trivial magic of
>>>>>> get_cur_path()/will_overwrite_ref()/did_overwrite_ref()/did_overwrite_first_ref().
>>>>>> I hit the following issue, when testing full-send:
>>>>>>
>>>>>> This is my source subvolume (inode numbers are written):
>>>>>> tree -A --inodes --noreport /mnt/src/tmp/
>>>>>> /mnt/src/tmp/
>>>>>> └── [ 270] dir2
>>>>>> └── [ 268] file1_nod
>>>>>>
>>>>>> As you see, the ino(file1_nod) < ino(dir2). It is very easy to
>>>>>> achieve: first create the file, then the dir, and then move the file
>>>>>> to dir.
>>>>>>
>>>>>> During send the following happens (I augmented the send code with many
>>>>>> prints):
>>>>>>
>>>>>> file1_nod is sent first. Since its a new inode, it is sent as an
>>>>>> orphan. When recording its reference, __record_new_ref() calls
>>>>>> get_cur_path() for its parent (270). Then __get_cur_name_and_parent()
>>>>>> is called on 270, which calls is_inode_existent(), which calls
>>>>>> get_cur_inode_state(), and the state of the parent is "will_create".
>>>>>> So __get_cur_name_and_parent() creates an orphan name for it, and
>>>>>> finally the new reference for 268 is recorded as:
>>>>>> o270-136-0/file1_nod:
>>>>>>
>>>>>> [changed_cb:4102] key(256 INODE_ITEM 0) : NEW
>>>>>> [changed_cb:4102] key(256 INODE_REF 256) : NEW
>>>>>> [changed_cb:4102] key(268 INODE_ITEM 0) : NEW
>>>>>> [send_create_inode:2407] NEW ino(268,135) type=0100000, path=[o268-135-0]
>>>>>> [changed_cb:4102] key(268 INODE_REF 270) : NEW
>>>>>> [get_cur_inode_state:1475] (270,136): L(EX,136)
>>>>>> R(NE,18446744072099047770) sp=268 ==> will_create
>>>>>> [is_inode_existent:1498] (270,136): NOT existent
>>>>>> [__get_cur_name_and_parent:1918] ino(270,136) not existent => unique
>>>>>> name [o270-136-0]
>>>>>> [get_cur_path:2051] ino(0,0) cur_path=[o270-136-0]
>>>>>> [__record_new_ref:2911] record new ref [o270-136-0/file1_nod]
>>>>>>
>>>>>> Then process_recorded_refs() sees that 268 is still orphan, so it
>>>>>> sends "rename" to its valid place, but the problem is that its parent
>>>>>> dir was not sent yet (and its parent dir is also an orphan):
>>>>>> [process_recorded_refs:2601] ino(268,135): start with refs
>>>>>> [28118.347602] [process_recorded_refs:2651] ino(268,135): new=1,
>>>>>> did_overwrite_first_ref=0, is_orphan=1, valid_path=[o268-135-0]
>>>>>> [28118.347605] [process_recorded_refs:2701] ino(268,135): is orphan,
>>>>>> move it: [o268-135-0]=>[o270-136-0/file1_nod]
>>>>>> [28118.347610] [process_recorded_refs:2837] checking dir(270,136)
>>>>>> [28118.347612] [process_recorded_refs:2869] ino(268,135) done with refs
>>>>>>
>>>>>> Now the parent dir is processed:
>>>>>> [changed_cb:4102] key(270 INODE_ITEM 0) : NEW
>>>>>> [send_create_inode:2407] NEW ino(270,136) type=040000, path=[o270-136-0]
>>>>>> [changed_cb:4102] key(270 INODE_REF 256) : NEW
>>>>>> [get_cur_path:2051] ino(256,133) cur_path=[]
>>>>>> [__record_new_ref:2911] record new ref [dir2]
>>>>>> [process_recorded_refs:2601] ino(270,136): start with refs
>>>>>> [process_recorded_refs:2651] ino(270,136): new=1,
>>>>>> did_overwrite_first_ref=0, is_orphan=1, valid_path=[o270-136-0]
>>>>>> [process_recorded_refs:2701] ino(270,136): is orphan, move it:
>>>>>> [o270-136-0]=>[dir2]
>>>>>> [process_recorded_refs:2837] checking dir(256,133)
>>>>>> [get_cur_inode_state:1475] (256,133): L(EX,133)
>>>>>> R(NE,18446612135413283512) sp=270 ==> did_create
>>>>>> [process_recorded_refs:2869] ino(270,136) done with refs
>>>>>>
>>>>>> Nothing special here, the parent is first sent as an orphan, and then
>>>>>> renamed to its valid name, but it's too late.
>>>>>>
>>>>>> During receive:
>>>>>> ERROR: rename o268-135-0 -> o270-136-0/file1_nod failed. No such file
>>>>>> or directory
>>>>>>
>>>>>> I am not yet sure where is the proper place to fix this, I just wanted
>>>>>> to report it first. Basically, I think that when sending any kind of
>>>>>> A_PATH, it is needed to ensure that path components exist, either as
>>>>>> orphan or real path (by sending them out-of-order if needed?). But I
>>>>>> am not yet sure where is the core place that should ensure this.
>>>>>>
>>>>>> Thanks,
>>>>>> Alex.
>>>>>
>>>>> I have pushed a fix for this case. Basically, the solution is to
>>>>> postpone the processing of refs in not created dirs until the dir is
>>>>> created. Big thanks for investigating this one.
btrfs_init_tests.sh
Description: Bourne shell script
btrfs_test_1.sh
Description: Bourne shell script
btrfs_test_2.sh
Description: Bourne shell script
