Protocol: 28 (unchanged)
Changes since 2.6.2:
SECURITY FIXES:
- A bug in the sanitize_path routine (which affects a non-chrooted
rsync daemon) could allow a user to craft a pathname that would get
transformed into an absolute path for certain options (but not for
file-transfer names). If you're running an rsync daemon with chroot
disabled, *please upgrade*, ESPECIALLY if the user privs you run
rsync under is anything above "nobody".
OUTPUT CHANGES (ATTN: those using a script to parse the verbose output):
- Please note that the 2-line footer (output when verbose) now uses the
term "sent" instead of "wrote" and "received" instead of "read". If
you are not parsing the numeric values out of this footer, a script
would be better off using the empty line prior to the footer as the
indicator that the verbose output is over.
- The output from the --stats option was similarly affected to change
"written" to "sent" and "read" to "received".
- Rsync ensures that a filename that contains a newline gets mentioned
with each newline transformed into a question mark (which prevents a
filename from causing an empty line to be output).
- The "backed up ..." message that is output when at least 2 --verbose
options are specified is now the same both with and without the
--backup-dir option.
BUG FIXES:
- Fixed a crash bug that might appear when --delete was used and
multiple source directories were specified.
- Fixed a 32-bit truncation of the file length when generating the
checksums.
- The --backup code no longer attempts to create some directories
over and over again (generating warnings along the way).
- Fixed a bug in the reading of the secrets file (by the daemon) and
the password file (by the client): the files no longer need to be
terminated by a newline for their content to be read in.
- If a file has a read error on the sending side or the reconstructed
data doesn't match the expected checksum (perhaps due to the basis
file changing during the transfer), the receiver will no longer
retain the resulting file unless the --partial option was specified.
(Note: for the read-error detection to work, neither side can be
older than 2.6.3 -- older receivers will always retain the file, and
older senders don't tell the receiver that the file had a read
error.)
- If a file gets resent in a single transfer and the --backup option
is enabled, rsync no longer performs a duplicate backup (it used to
overwrite the original file in the backup area).
- Files specified in the daemon's "exclude" or "exclude from" config
items are now excluded from being uploaded (assuming that the module
allows uploading at all) in addition to the old download exclusion.
- Got rid of a potential hang in the receiver when near the end of a
phase.
- When using --backup without a --backup-dir, rsync no longer preserves
the modify time on directories. This avoids confusing NFS.
- When --copy-links (-L) is specified, we now output a separate error
for a symlink that has no referent instead of claiming that a file
"vanished".
- The --copy-links (-L) option no longer has the side-effect of telling
the receiving side to follow symlinks. See the --keep-dirlinks
option (mentioned below) for a way to specify that behavior.
- Error messages from the daemon server's option-parsing (such as
refused options) are now successfully transferred back to the client
(the server used to fail to send the message because the socket
wasn't in the right state for the message to get through).
- Most transfer errors that occur during a daemon transfer are now
returned to the user in addition to being logged (some messages are
intended to be daemon-only and are not affected by this).
- Fixed a bug in the daemon authentication code when using one of the
batch-processing options.
- We try to work around some buggy IPv6 implementations that fail to
implement IPV6_V6ONLY. This should fix the "address in use" error
that some daemons get when running on an OS with a buggy IPv6
implementation. Also, if the new code gets this error, we might
suggest that the user specify --ipv4 or --ipv6 (if we think it will
help).
- When the remote rsync dies, make a better effort to recover any error
messages it may have sent before dying (the local rsync used to just
die with a socket-write error).
- When using --delete and a --backup-dir that contains files that are
hard-linked to their destination equivalents, rsync now makes sure
that removed files really get removed (avoids a really weird rename()
behavior).
- Avoid a bogus run-time complaint about a lack of 64-bit integers when
the int64 type is defined as an off_t and it actually has 64-bits.
- Added a configure check for open64() without mkstemp64() so that we
can avoid using mkstemp() when such a combination is encountered.
This bypasses a problem writing out large temp files on OSes such as
AIX and HP-UX.
- Fixed an age-old crash problem with --read-batch on a local copy
(rsync was improperly assuming --whole-file for the local copy).
- When --dry-run (-n) is used and the destination directory does not
exist, rsync now produces a correct report of files that would be
sent instead of dying with a chdir() error.
- Fixed a bug that could cause a slow-to-connect rsync daemon to die
with an error instead of waiting for the connection to finish.
- Fixed an ssh interaction that could cause output to be lost when the
user chose to combine the output of rsync's stdout and stderr (e.g.
using the "2>&1").
ENHANCEMENTS:
- Added the --partial-dir=DIR option that lets you specify where to
(temporarily) put a partially transferred file (instead of over-
writing the destination file). E.g. --partial-dir=.rsync-partial
Also added support for the RSYNC_PARTIAL_DIR environment variable
that, when found, transforms a regular --partial option (such as
the convenient -P option) into one that also specifies a directory.
- Added --keep-dirlinks (-K), which allows you to symlink a directory
onto another partition on the receiving side and have rsync treat it
as matching a normal directory from the sender.
- Added the --inplace option that tells rsync to write each destination
file without using a temporary file. The matching of existing data
in the destination file can be severely limited by this, but there
are also cases where this is more efficient (such as appending data).
Use only when needed (see the man page for more details).
- Added the "write only" option for the daemon's config file.
- Added long-option names for -4 and -6 (namely --ipv4 and --ipv6)
and documented all these options in the man page.
- Improved the handling of the --bwlimit option so that it's less
bursty, more accurate, and works properly over a larger range of
values.
- The rsync daemon-over-ssh code now looks for SSH_CONNECTION and
SSH2_CLIENT in addition to SSH_CLIENT to figure out the IP address.
- Added the --checksum-seed=N option for advanced users.
- Batch writing/reading has a brand-new implementation that is simpler,
fixes a few weird problems with the old code (such as no longer
sprinkling the batch files into different dirs or even onto different
systems), and is much less intrusive into the code (making it easier
to maintain for the future). The new code generates just one data
file instead of three, which makes it possible to read the batch on
stdin via a remote shell. Also, the old requirement of forcing the
same fixed checksum-seed for all batch processing has been removed.
- If an rsync daemon has a module set with "list = no" (which hides its
presence in the list of available modules), a user that fails to
authenticate gets the same "unknown module" error that they would get
if the module were actually unknown (while still logging the real
error to the daemon's log file). This prevents fishing for module
names.
- The daemon's "refuse options" config item now allows you to match
option names using wildcards and/or the single-letter option names.
- Each transferred file now gets its permissions and modified-time
updated before the temp-file gets moved into place. Previously, the
finished file would have a very brief window where its permissions
disallowed all group and world access.
- Added the ability to parse a literal IPv6 address in an "rsync:" URL
(e.g. rsync://[2001:638:500:101::21]:873/module/dir).
- The daemon's wildcard expanding code can now handle more than 1000
filenames (it's now limited by memory instead of having a hard-wired
limit).
INTERNAL:
- Some cleanup in the exclude code has saved some per-exclude memory
and made the code easier to maintain.
- Improved the argv-overflow checking for a remote command that has a
lot of args.
- Use rsyserr() in the various places that were still calling rprintf()
with strerror() as an arg.
- If an rsync daemon is listening on multiple sockets (to handle both
IPv4 and IPv6 to a single port), we now close all the unneeded file
handles after we accept a connection (we used to close just one of
them).
- Optimized the handling of larger block sizes (rsync used to slow to a
crawl if the block size got too large).
- Optimized away a loop in hash_search().
- Some improvements to the sanitize_path() and clean_fname() functions
makes them more efficient and produce better results (while still
being compatible with the file-name cleaning that gets done on both
sides when sending the file-list).
- Got rid of alloc_sanitize_path() after adding a destination-buffer
arg to sanitize_path() made it possible to put all the former's
functionality into the latter.
- The file-list that is output when at least 4 verbose options are
specified reports the uid value on the sender even when rsync is
not running as root (since we might be sending to a root receiver).
BUILD CHANGES:
- Added a "gen" target to rebuild most of the generated files,
including configure, config.h.in, the man pages, and proto.h.
- If "make proto" doesn't find some changes in the prototypes, the
proto.h file is left untouched (its time-stamp used to always be
updated).
- The variable $STRIP (that is optionally set by the install-strip
target's rule) was changed to $INSTALL_STRIP because some systems
have $STRIP already set in the environment.
- Fixed a build problem when SUPPORT_HARD_LINKS isn't defined.
- When cross-compiling, the gettimeofday() function is now assumed to
be a modern version that takes two-args (since we can't test it).
DEVELOPER RELATED:
- The scripts in the testsuite dir were cleaned up a bit and a few
new tests added.
- Some new diffs were added to the patches dir, and some accepted
ones were removed.
4e70e5635f