thr3ads.net - rsync - rsync-2.6.2: NFS clients confused after an rsync [May 2004]

If this information is useful, please help other people find it:
Share via:

Brian Childs

2004-May-06 22:22 UTC

rsync-2.6.2: NFS clients confused after an rsync

We use rsync to update an nfs server.  After an update, we noticed that
a large number of clients didn't see the updated data.

It took me a while to be able to reliably reproduce this problem, but it
happens on old and new versions of rysnc.  It also happens across all
the platforms we use here (sun/linux/netapp).

This shows the problem: [Note my home directory is NFS mounted]

***************************************************
rm -f /tmp/testpath/* $HOME/testpath/*

#Prime it
echo some data > /tmp/testpath/testfile
sleep 1
rsync -a /tmp/testpath/ $HOME/testpath/
ssh $otherbox ls -li $HOME/testpath

#Break it
echo additional data >> /tmp/testpath/testfile
sleep 1
rsync -a -b /tmp/testpath/ $HOME/testpath/
ssh $otherbox ls -li $HOME/testpath

#Fix it
touch $HOME/testpath
ssh $otherbox ls -li $HOME/testpath
***************************************************

Here's the output:
total 0
    483079 -rw-r--r--   1 brian    source        10 May  6 18:05 testfile
total 0
    483079 -rw-r--r--   1 brian    source        10 May  6 18:05 testfile
total 0
    483080 -rw-r--r--   1 brian    source        26 May  6 18:05 testfile
    483079 -rw-r--r--   1 brian    source        10 May  6 18:05 testfile~

The output from the "break it" section matches "prime it",
but it should
match the "fix it" output.

What I think is happening:
--------------------------

* testfile is updated in place, so the mtime of testpath
  isn't udpated.
* The nfs client is caching the mtime of "testpath"
* The seconds rsync modifies the contents of testpath, but
  when finished, it resets the mtime on "testpath"
* The client believes it's cache is valid, so I doesn't refresh, and
  therefore it misses the update

NFS Clients tested: sun (solaris 8 & 9), linux (2.4.18 & 2.4.24)
NFS Servers tested: sun, linux, and netapp.

They all behaved the same.

Proposed fix:
-------------
Change rsync so it doesn't reset mtimes on directories if --backup is
in effect.

Thanks,
Brian

Brian Childs

2004-May-07 15:19 UTC

head link

rsync-2.6.2: NFS clients confused after an rsync [PATCH included]

As a response to my original post, here's a patch that implements my
proposed solution.

I've tested it, and it fixes the problem, but I'm afraid there may be
some hidden consequences of doing this.

If anyone can think of anything, please let me know.

Thanks,
Brian

--- rsync-2.6.2/util.c.orig	2004-05-07 10:06:08.000000000 -0400
+++ rsync-2.6.2/util.c	2004-05-07 11:06:24.000000000 -0400
@@ -127,9 +127,17 @@
 int set_modtime(char *fname, time_t modtime)
 {
 	extern int dry_run;
+	extern int make_backups;
+
 	if (dry_run)
 		return 0;
 
+	if(make_backups) {
+		struct stat sb;
+		if(!lstat(fname, &sb) && S_ISDIR(sb.st_mode))
+		  return 0;
+	}
+
 	if (verbose > 2) {
 		rprintf(FINFO, "set modtime of %s to (%ld) %s",
 			fname, (long) modtime,

On Thu, May 06, 2004 at 06:21:55PM -0400, Brian Childs
wrote:> We use rsync to update an nfs server.  After an update, we noticed that
> a large number of clients didn't see the updated data.
> 
> It took me a while to be able to reliably reproduce this problem, but it
> happens on old and new versions of rysnc.  It also happens across all
> the platforms we use here (sun/linux/netapp).
> 
> This shows the problem: [Note my home directory is NFS mounted]
> 
> ***************************************************
> rm -f /tmp/testpath/* $HOME/testpath/*
> 
> #Prime it
> echo some data > /tmp/testpath/testfile
> sleep 1
> rsync -a /tmp/testpath/ $HOME/testpath/
> ssh $otherbox ls -li $HOME/testpath
> 
> #Break it
> echo additional data >> /tmp/testpath/testfile
> sleep 1
> rsync -a -b /tmp/testpath/ $HOME/testpath/
> ssh $otherbox ls -li $HOME/testpath
> 
> #Fix it
> touch $HOME/testpath
> ssh $otherbox ls -li $HOME/testpath
> ***************************************************
> 
> Here's the output:
> total 0
>     483079 -rw-r--r--   1 brian    source        10 May  6 18:05 testfile
> total 0
>     483079 -rw-r--r--   1 brian    source        10 May  6 18:05 testfile
> total 0
>     483080 -rw-r--r--   1 brian    source        26 May  6 18:05 testfile
>     483079 -rw-r--r--   1 brian    source        10 May  6 18:05 testfile~
> 
> The output from the "break it" section matches "prime
it", but it should
> match the "fix it" output.
> 
> What I think is happening:
> --------------------------
> 
> * testfile is updated in place, so the mtime of testpath
>   isn't udpated.
> * The nfs client is caching the mtime of "testpath"
> * The seconds rsync modifies the contents of testpath, but
>   when finished, it resets the mtime on "testpath"
> * The client believes it's cache is valid, so I doesn't refresh,
and
>   therefore it misses the update
> 
> NFS Clients tested: sun (solaris 8 & 9), linux (2.4.18 & 2.4.24)
> NFS Servers tested: sun, linux, and netapp.
> 
> They all behaved the same.
> 
> Proposed fix:
> -------------
> Change rsync so it doesn't reset mtimes on directories if --backup is
> in effect.
> 
> Thanks,
> Brian
> -- 
> To unsubscribe or change options:
http://lists.samba.org/mailman/listinfo/rsync
> Before posting, read: http://www.catb.org/~esr/faqs/smart-questions.html

Wayne Davison

2004-May-11 17:02 UTC

head link

rsync-2.6.2: NFS clients confused after an rsync

On Thu, May 06, 2004 at 06:21:55PM -0400, Brian Childs
wrote:> rm -f /tmp/testpath/* $HOME/testpath/*
> 
> #Prime it
> echo some data > /tmp/testpath/testfile
> sleep 1
> rsync -a /tmp/testpath/ $HOME/testpath/
> ssh $otherbox ls -li $HOME/testpath
> 
> #Break it
> echo additional data >> /tmp/testpath/testfile
> sleep 1
> rsync -a -b /tmp/testpath/ $HOME/testpath/
> ssh $otherbox ls -li $HOME/testpath
Do you actually need to use the -b option for the breakage to happen?
In the output you show, not only is the new testfile~ not yet found,
but the extra data in testfile is not seen (including showing the wrong
inode).  Since rsync isn't actually updating any files in-place, the
file gets a new inode while keeping the directory time unchanged, so
that may be enough to confuse NFS.

I had been considering making the preservation of directory times an
optional occurance with rsync.  The appended patch implements this.
Comments?

..wayne..
-------------- next part --------------
--- options.c	6 May 2004 21:08:01 -0000	1.148
+++ options.c	11 May 2004 16:57:18 -0000
@@ -46,6 +46,7 @@ int preserve_devices = 0;
 int preserve_uid = 0;
 int preserve_gid = 0;
 int preserve_times = 0;
+int preserve_dir_times = 0;
 int update_only = 0;
 int cvs_exclude = 0;
 int dry_run = 0;
@@ -240,7 +241,8 @@ void usage(enum logcode F)
   rprintf(F," -o, --owner                 preserve owner (root
only)\n");
   rprintf(F," -g, --group                 preserve group\n");
   rprintf(F," -D, --devices               preserve devices (root
only)\n");
-  rprintf(F," -t, --times                 preserve times\n");
+  rprintf(F," -t, --times                 preserve times on
non-directories\n");
+  rprintf(F," -d, --dir-times             preserve times on
directories\n");
   rprintf(F," -S, --sparse                handle sparse files
efficiently\n");
   rprintf(F," -n, --dry-run               show what would have been
transferred\n");
   rprintf(F," -W, --whole-file            copy whole files, no incremental
checks\n");
@@ -346,6 +348,7 @@ static struct poptOption long_options[] 
   {"group",           'g', POPT_ARG_NONE,  
&preserve_gid, 0, 0, 0 },
   {"devices",         'D', POPT_ARG_NONE,  
&preserve_devices, 0, 0, 0 },
   {"times",           't', POPT_ARG_NONE,  
&preserve_times, 0, 0, 0 },
+  {"dir-times",       'd', POPT_ARG_NONE,  
&preserve_dir_times, 0, 0, 0 },
   {"checksum",        'c', POPT_ARG_NONE,  
&always_checksum, 0, 0, 0 },
   {"verbose",         'v', POPT_ARG_NONE,   0,              
'v', 0, 0 },
   {"quiet",           'q', POPT_ARG_NONE,   0,              
'q', 0, 0 },
@@ -823,6 +826,8 @@ void server_options(char **args,int *arg
 		argstr[x++] = 'D';
 	if (preserve_times)
 		argstr[x++] = 't';
+	if (preserve_dir_times && am_sender)
+		argstr[x++] = 'd';
 	if (preserve_perms)
 		argstr[x++] = 'p';
 	if (recurse)
--- rsync.c	23 Mar 2004 16:16:15 -0000	1.135
+++ rsync.c	11 May 2004 16:57:19 -0000
@@ -25,6 +25,7 @@
 extern int verbose;
 extern int dry_run;
 extern int preserve_times;
+extern int preserve_dir_times;
 extern int am_root;
 extern int am_server;
 extern int am_sender;
@@ -125,9 +126,9 @@ int delete_file(char *fname)
 int set_perms(char *fname,struct file_struct *file,STRUCT_STAT *st,
 		int report)
 {
-	int updated = 0;
 	STRUCT_STAT st2;
 	int change_uid, change_gid;
+	int keep_time, updated = 0;
 
 	if (dry_run) return 0;
 
@@ -140,12 +141,10 @@ int set_perms(char *fname,struct file_st
 		st = &st2;
 	}
 
-	if (preserve_times && !S_ISLNK(st->st_mode) &&
-	    cmp_modtime(st->st_mtime, file->modtime) != 0) {
-		/* don't complain about not setting times on directories
-		 * because some filesystems can't do it */
-		if (set_modtime(fname,file->modtime) != 0 &&
-		    !S_ISDIR(st->st_mode)) {
+	keep_time = S_ISDIR(st->st_mode) ? preserve_dir_times
+		  : preserve_times && !S_ISLNK(st->st_mode);
+	if (keep_time && cmp_modtime(st->st_mtime, file->modtime) != 0)
{
+		if (set_modtime(fname,file->modtime) != 0) {
 			rprintf(FERROR, "failed to set times on %s: %s\n",
 				full_fname(fname), strerror(errno));
 			return 0;
--- rsync.yo	7 May 2004 00:18:37 -0000	1.169
+++ rsync.yo	11 May 2004 16:57:19 -0000
@@ -298,7 +298,8 @@ verb(
  -o, --owner                 preserve owner (root only)
  -g, --group                 preserve group
  -D, --devices               preserve devices (root only)
- -t, --times                 preserve times
+ -t, --times                 preserve times on non-directories
+ -d, --dir-times             preserve times on directories
  -S, --sparse                handle sparse files efficiently
  -n, --dry-run               show what would have been transferred
  -W, --whole-file            copy whole files, no incremental checks
@@ -538,13 +539,22 @@ dit(bf(-D, --devices)) This option cause
 block device information to the remote system to recreate these
 devices. This option is only available to the super-user.
 
-dit(bf(-t, --times)) This tells rsync to transfer modification times along
-with the files and update them on the remote system.  Note that if this
+dit(bf(-t, --times)) This tells rsync to preserve modification times of
+non-directories transferred to the remote system.  Note that if this
 option is not used, the optimization that excludes files that have not been
 modified cannot be effective; in other words, a missing -t or -a will
 cause the next transfer to behave as if it used -I, and all files will have
 their checksums compared and show up in log messages even if they haven't
 changed.
+
+dit(bf(-d, --dir-times)) This tells rsync to preserve the modification
+times of directories transferred to the remote system.  On a modern
+rsync, these are left unpreserved by default to avoid causing problems
+for NFS.
+
+Note: when sending files to an older rsync, the --times option will
+imply --dir-times (if the option causes an error on the receiving
+system, omit it and use --times to preserve all file/directory times).
 
 dit(bf(-n, --dry-run)) This tells rsync to not do any file transfers,
 instead it will just report the actions it would have taken.

Possibly Parallel Threads

Search for more reasonably related threads

rsync - May 2004 - rsync-2.6.2: NFS clients confused after an rsync

rsync-2.6.2: NFS clients confused after an rsync

rsync-2.6.2: NFS clients confused after an rsync [PATCH included]

rsync-2.6.2: NFS clients confused after an rsync

Possibly Parallel Threads