rsync - Apr 2017 - [Bug 12732] hard links can cause rsync to block or to silently skip files

https://bugzilla.samba.org/show_bug.cgi?id=12732

            Bug ID: 12732
           Summary: hard links can cause rsync to block or to silently
                    skip files
           Product: rsync
           Version: 3.1.2
          Hardware: x64
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P5
         Component: core
          Assignee: wayned at samba.org
          Reporter: samba at hlipp.de
        QA Contact: rsync-qa at samba.org

Overview
=======
Hard link handling seems to be broken when using "rsync -aH
--compare-dest". I
found two possible scenarios:

1) rsync completes without error message and exit code 0, although some files
are missing from the backup
2) rsync blocks and must be interrupted/killed

I found this bug when tracking down random hangs of rsync. It turned out that
most, but not all of these hangs occur when hard links are present. Therefore,
I hope the latter case might give some hints to a larger problem that might be
triggered by this hard link bug.


How to reproduce (1)
===================
[ Using Linux on e.g. ext4 ]
############################
mkdir srclt
cd srclt
echo x > a
ln a b
echo x > c
ln c d
cd ..

cp -aix srclt dstlt
rm dstlt/{b,c}

mkdir baklt

rsync -aHvv --compare-dest=$PWD/dstlt/. $PWD/srclt/. $PWD/baklt/. >>
testlt.log
2>&1
############################


Actual Results (1)
=================
cat testlt.log
#####
sending incremental file list
delta-transmission disabled for local transfer or --whole-file
./
a is uptodate
d is uptodate
b
a => b
total: matches=0  hash_hits=0  false_alarms=0 data=2

sent 173 bytes  received 160 bytes  666.00 bytes/sec
total size is 8  speedup is 0.02
#####
ls -il srclt dstlt baklt
#####
baklt:
total 8
3249818642 -rw-r--r-- 2 X X 2 2017-04-04 X a
3249818642 -rw-r--r-- 2 X X 2 2017-04-04 X b

dstlt:
total 8
2205741698 -rw-r--r-- 1 X X 2 2017-04-04 X a
2205741699 -rw-r--r-- 1 X X 2 2017-04-04 X d

srclt:
total 16
1138988347 -rw-r--r-- 2 X X 2 2017-04-04 X a
1138988347 -rw-r--r-- 2 X X 2 2017-04-04 X b
1138988348 -rw-r--r-- 2 X X 2 2017-04-04 X c
1138988348 -rw-r--r-- 2 X X 2 2017-04-04 X d
#####


Expected Results (1)
===================
The directory baklt should contain the entries b and c.
Entry c is completely ignored and does not show up in the log.
Entry a does not need to appear in the backup, as it is present in both
directories srclt and dstlt.


How to reproduce (2)
===================
[ Using Linux on e.g. ext4 ]
############################
mkdir srclt2
cd srclt2
echo x > a
ln a b
cd ..

cp -aix srclt2 dstlt2
rm dstlt2/b

mkdir baklt2

rsync -aHvv --compare-dest=$PWD/dstlt2/. $PWD/srclt2/. $PWD/baklt2/. >>
testlt2.log 2>&1
############################


Actual Results (2)
=================
=> rsync hangs and must be interrupted/killed.

cat testlt2.log
#####
sending incremental file list
delta-transmission disabled for local transfer or --whole-file
./
a is uptodate
b
a => b
rsync error: received SIGINT, SIGTERM, or SIGHUP (code 20) at rsync.c(636)
[sender=3.1.2]
rsync error: received SIGINT, SIGTERM, or SIGHUP (code 20) at io.c(504)
[generator=3.1.2]
#####
ls -il srclt2 dstlt2 baklt2
#####
baklt2:
total 8
2191211 -rw-r--r-- 2 X X 2 2017-04-04 X a
2191211 -rw-r--r-- 2 X X 2 2017-04-04 X b

dstlt2:
total 4
2191208 -rw-r--r-- 1 X X 2 2017-04-04 X a

srclt2:
total 8
2191206 -rw-r--r-- 2 X X 2 2017-04-04 X a
2191206 -rw-r--r-- 2 X X 2 2017-04-04 X b
#####


Expected Results (2)
===================
rsync should not block.
Entry a does not need to appear in the backup, as it is present in both
directories srclt and dstlt.


Further information
==================
This problem exists at least for rsync versions 3.1.0, 3.1.1, and 3.1.2 for
different Linux varieties using various file systems:
https://lists.samba.org/archive/rsync/2015-April/030092.html

Latest test on openSUSE 42.2 (x86_64) on ext4 + on nfs with
rsync --version
#####
rsync  version 3.1.2  protocol version 31
Copyright (C) 1996-2015 by Andrew Tridgell, Wayne Davison, and others.
Web site: http://rsync.samba.org/
Capabilities:
    64-bit files, 64-bit inums, 64-bit timestamps, 64-bit long ints,
    socketpairs, hardlinks, symlinks, IPv6, batchfiles, inplace,
    append, ACLs, xattrs, iconv, symtimes, prealloc
#####

-- 
You are receiving this mail because:
You are the QA Contact for the bug.

just subscribed for rsync-qa from bugzilla via rsync
wrote:
> Hard link handling seems to be broken when using "rsync -aH
--compare-dest". I
> found two possible scenarios:
>
> 1) rsync completes without error message and exit code 0, although some
files
> are missing from the backup
> 2) rsync blocks and must be interrupted/killed
> ....
>
> Further information
> ==================>
> This problem exists at least for rsync versions 3.1.0, 3.1.1, and 3.1.2 for
> different Linux varieties using various file systems:
> https://lists.samba.org/archive/rsync/2015-April/030092.html
>   
---

    I ran rsync 3.1.1 for over a year to help generate
snapshots.  I can't say if it copied all the files or not, as
it was backing up a large "/home" partition, BUT, it never hung.
It did take 45min to a few hours to do the compare, but it
was comparing a large amount of data (>750G) w/a snapshot
(another 750G) to dump diffs to a third, and my /home partion
has a *very* large number of hard links.

    So I know that hardlinks are handled 'fine' on comparing
'xfs' to 'xfs'. 
> Latest test on openSUSE 42.2 (x86_64) on ext4 + on nfs with
>   
----
    Ah...  I'd suspect nfs...
   
    Why are you using nfs?  rsync was designed to compare
against local file systems.  You should try running rsync
directly from the nfs-host machine to the client and bypassing
NFS.  I.e. -- you need to bypass NFS, since local->local
with hardlinks works.

    Just checked my /home partition.
    find shows 9295431 names (of any type), but du shows
(using du --inodes) shows 4407458 inodes.  That means over
half of the filenames are hard linked.  While my home
partition takes up 60% more space now, even cutting
those counts in half would still a large number of
hard links -- and rsync didn't crash doing an
rsync of the partition to an empty one, but first comparing
to a previous snapshot (the empty partition ended up
with differences between the main partition & the snapshot.

    I'd remove NFS...

Am 05.04.2017 um 22:05 schrieb L A Walsh via rsync:
>    I ran rsync 3.1.1 for over a year to help generate
> snapshots.  I can't say if it copied all the files or not, as
> it was backing up a large "/home" partition, BUT, it never hung.
> It did take 45min to a few hours to do the compare, but it
> was comparing a large amount of data (>750G) w/a snapshot
> (another 750G) to dump diffs to a third, and my /home partion
> has a *very* large number of hard links.
I've been using rsync for many years and it works fine most of the time.
I'm not sure if all of the occasional hangs have the same reason, these
are really hard to track down as they usually occur during large
transfers (e.g. when synchronizing large backup disks). That's why I was
happy that I could find a small test case which triggers this problem.

Does your rsync hang after the sequence of commands described in section
"How to reproduce (2)"?
>> Latest test on openSUSE 42.2 (x86_64) on ext4 + on nfs with
>>   
> ----
>    Ah...  I'd suspect nfs...
>      Why are you using nfs?
In order to find out if there is a difference when using another file
system type. The most recent tests were on ext4 and on nfs
(independently), older tests were on at least ext3 and xfs. IIRC I only
tested on different OpenSUSE and Debian versions on x86_64 systems, though.
>    Just checked my /home partition.
>    find shows 9295431 names (of any type), but du shows
> (using du --inodes) shows 4407458 inodes.  That means over
> half of the filenames are hard linked.  While my home
> partition takes up 60% more space now, even cutting
> those counts in half would still a large number of
> hard links -- and rsync didn't crash doing an
> rsync of the partition to an empty one, but first comparing
> to a previous snapshot (the empty partition ended up
> with differences between the main partition & the snapshot.
Probably using different options? Can this be some sort of Heisenbug,
nobody can reproduce? Do the two sequences of shell commands work for
you as expected? Please note that both rsync commands in the mail
generated by bugzilla are split into two lines (each): Both rsync
commands should read
   rsync PARAMS DIRS >> XXX.log 2>&1

Kind regards,
Hansjoerg

https://bugzilla.samba.org/show_bug.cgi?id=12732

--- Comment #1 from Hansjoerg Lipp <samba at hlipp.de> ---
Am 05.04.2017 um 22:05 schrieb L A Walsh via rsync:
>    I ran rsync 3.1.1 for over a year to help generate
> snapshots.  I can't say if it copied all the files or not, as
> it was backing up a large "/home" partition, BUT, it never hung.
> It did take 45min to a few hours to do the compare, but it
> was comparing a large amount of data (>750G) w/a snapshot
> (another 750G) to dump diffs to a third, and my /home partion
> has a *very* large number of hard links.
I've been using rsync for many years and it works fine most of the time.
I'm not sure if all of the occasional hangs have the same reason, these
are really hard to track down as they usually occur during large
transfers (e.g. when synchronizing large backup disks). That's why I was
happy that I could find a small test case which triggers this problem.

Does your rsync hang after the sequence of commands described in section
"How to reproduce (2)"?
>> Latest test on openSUSE 42.2 (x86_64) on ext4 + on nfs with
>>   
> ----
>    Ah...  I'd suspect nfs...
>      Why are you using nfs?
In order to find out if there is a difference when using another file
system type. The most recent tests were on ext4 and on nfs
(independently), older tests were on at least ext3 and xfs. IIRC I only
tested on different OpenSUSE and Debian versions on x86_64 systems, though.
>    Just checked my /home partition.
>    find shows 9295431 names (of any type), but du shows
> (using du --inodes) shows 4407458 inodes.  That means over
> half of the filenames are hard linked.  While my home
> partition takes up 60% more space now, even cutting
> those counts in half would still a large number of
> hard links -- and rsync didn't crash doing an
> rsync of the partition to an empty one, but first comparing
> to a previous snapshot (the empty partition ended up
> with differences between the main partition & the snapshot.
Probably using different options? Can this be some sort of Heisenbug,
nobody can reproduce? Do the two sequences of shell commands work for
you as expected? Please note that both rsync commands in the mail
generated by bugzilla are split into two lines (each): Both rsync
commands should read
   rsync PARAMS DIRS >> XXX.log 2>&1

Kind regards,
Hansjoerg

-- 
You are receiving this mail because:
You are the QA Contact for the bug.

just subscribed for rsync-qa from bugzilla via rsync
wrote:
> Probably using different options? Can this be some sort of Heisenbug,
> nobody can reproduce? Do the two sequences of shell commands work for
> you as expected? Please note that both rsync commands in the mail
> generated by bugzilla are split into two lines (each): Both rsync
> commands should read
>   
---
Actually, I am reproducing your bug exactly as you entered
it. 

And it is very sensitive to options.

When I take out the $PWD it works.

First your case:
/tmp> rsync -aHvv --compare-dest=$PWD/dstlt/. $PWD/srclt/. $PWD/baklt/. 
|& tee testlt.log
sending incremental file list
delta-transmission disabled for local transfer or --whole-file
./
a is uptodate
d is uptodate
b
a => b
/tmp> ll -i srclt dstlt baklt
baklt:
total 8
17305770 -rw-rw-r-- 2 2 Apr  5 15:46 a
17305770 -rw-rw-r-- 2 2 Apr  5 15:46 b
dstlt:
total 8
8474888 -rw-rw-r-- 1 2 Apr  5 15:46 a
8474896 -rw-rw-r-- 1 2 Apr  5 15:46 d
srclt:
total 16
8866 -rw-rw-r-- 2 2 Apr  5 15:46 a
8866 -rw-rw-r-- 2 2 Apr  5 15:46 b
8867 -rw-rw-r-- 2 2 Apr  5 15:46 c
8867 -rw-rw-r-- 2 2 Apr  5 15:46 d
----
Then w/o $PWD:
/tmp> rsync -aHvv --compare-dest=dstlt/. srclt/. baklt/. |& tee
testlt.log
sending incremental file list
--compare-dest arg does not exist: dstlt/.
delta-transmission disabled for local transfer or --whole-file
./
b
d
a => b
c => d
/tmp> ll -i srclt dstlt baklt
baklt:
total 16
17305770 -rw-rw-r-- 2 2 Apr  5 15:46 a
17305770 -rw-rw-r-- 2 2 Apr  5 15:46 b
17305782 -rw-rw-r-- 2 2 Apr  5 15:46 c
17305782 -rw-rw-r-- 2 2 Apr  5 15:46 d
dstlt:
total 8
8474888 -rw-rw-r-- 1 2 Apr  5 15:46 a
8474896 -rw-rw-r-- 1 2 Apr  5 15:46 d
srclt:
total 16
8866 -rw-rw-r-- 2 2 Apr  5 15:46 a
8866 -rw-rw-r-- 2 2 Apr  5 15:46 b
8867 -rw-rw-r-- 2 2 Apr  5 15:46 c
8867 -rw-rw-r-- 2 2 Apr  5 15:46 d
--------------

However, if "dstl" is a different file system, it
works:
>  rsync -aHv --compare-dest=~/dstlt/. $PWD/srclt/. $PWD/baklt/. |& tee 
testlt.log  
sending incremental file list
--compare-dest arg does not exist: ~/dstlt/.
./
b
d
a => b
c => d

In my usage both the compare-dest & baklt were 'technically'
different file systems (though the compare-dest was a snap
shot).
-----------
ARG: if baklt is on a different FS, but src+dst are same:
/tmp> rsync -aHv --compare-dest=$PWD/dstlt/. $PWD/srclt/. ~/baklt/. |& 
tee testlt.log  
sending incremental file list
./
b
a => b
/tmp> ll -i srclt dstlt ~/baklt
/home/law/baklt:
total 8
3677653445 -rw-rw-r-- 2 2 Apr  5 15:46 a
3677653445 -rw-rw-r-- 2 2 Apr  5 15:46 b

dstlt:
total 8
8474888 -rw-rw-r-- 1 2 Apr  5 15:46 a
8474896 -rw-rw-r-- 1 2 Apr  5 15:46 d

srclt:
total 16
8866 -rw-rw-r-- 2 2 Apr  5 15:46 a
8866 -rw-rw-r-- 2 2 Apr  5 15:46 b
8867 -rw-rw-r-- 2 2 Apr  5 15:46 c
8867 -rw-rw-r-- 2 2 Apr  5 15:46 d

----
Ok... something's screwy here.  But no hangs.

Like I said in my 1st response, if some hard links
were missing in the bak directory, I really would be
unlikely to notice, as I didn't even try to check
them (way too many files), but I never got a hang --
and that's the part I'm thinking might be nfs
related, since I've seen several issues with nfs not
working the same way as a local fs.

I usually use smbfs -- in my usage, it's faster and I
usually have fewer compatibility problems.
(faster meaning ~ Reads~162MB/s, W~220MB/s,
though i've seen explorer hit over 400MB/s).
(using a 10Gb link).

But the above testing shows some unexplained behaviors
out of rsync that sure look like a bug.

Good test case!
:-)

https://bugzilla.samba.org/show_bug.cgi?id=12732

--- Comment #2 from devurandom at gmx.net ---
There are a lot of bugreports related to rsync hanging mysteriously, some of
which may be duplicates of each other:

https://bugzilla.samba.org/show_bug.cgi?id=1442
https://bugzilla.samba.org/show_bug.cgi?id=2957
https://bugzilla.samba.org/show_bug.cgi?id=9164
https://bugzilla.samba.org/show_bug.cgi?id=10035
https://bugzilla.samba.org/show_bug.cgi?id=10092
https://bugzilla.samba.org/show_bug.cgi?id=10518
https://bugzilla.samba.org/show_bug.cgi?id=10950
https://bugzilla.samba.org/show_bug.cgi?id=11166
https://bugzilla.samba.org/show_bug.cgi?id=12732
https://bugzilla.samba.org/show_bug.cgi?id=13109

-- 
You are receiving this mail because:
You are the QA Contact for the bug.