- Finally implement fixups and use them in the detailer and updater
thread as appropriate. We now longer crash/hang when there is an MD5
checksum error, but request a fixup (or fail, if the checksum error is
from a fixup).
- Portability fix: don't use SIZE_T_MAX when ~0 will do fine.
- Plug a memory leak in keyword_prepare().
- Fix the build without assertions.
- Properly check for the success of the asynchronous connect() when we've
been interrupted by a signal. Move the logic into proto_waitconnect().
- Properly print transport layer addresses with getnameinfo() and the
NI_NUMERICHOST flag. We were printing garbage...
- Assert that we're never printing a NULL string in proto_printf().
- Make sure we disallow 0-length fields in proto_get_ascii() or all sorts
of bad things will happen. Also remove a useless check.
- In statusrec_cook(), there is nothing left to parse for DirDown
entries, however we have to check that there was nothing left in the
line, or return an error if it's not the case.
- Restore the original string upon error in fattr_scanattr(), called by
fattr_encode() only.
- The struct diff used by diff_apply() and keyword_expand() has been
rename to struct diffinfo, and it now only contains the metadata
of a diff. This changes the prototype for the two aforementioned
functions, so update the code and the consumers appropriately.
- Create the worker threads in the detached state since we don't use
pthread_join() to wait for them but have our own API for that (which
allows us to wait for multiple threads).
- Move the fattr_init() and fattr_fini() calls earlier to avoid calling
them several times during a run.
- When printing the "Connected to" message, print the actual address we
are connected. This makes us deviate slightly from CVSup, but since
csup tries any address returned by a host (including IPv6 addresses),
we really need to know where we connected.
- Make the errors/error messages handling much nicer in the status file
API. Nearly all the asprintf() calls are centralized now.
- Before entering multi-threaded mode, starts a "killer" thread that
will spend most of his time blocking in sigwait() and will call
mux_shutdown() to nicely abort the run in case we get a fatal signal.
- Remove the need for the "closing" condition variable in
mux_shutdown(), we are now handling the race it protected against
much more sanely. We just disable cancellation in the "killer" thread
before calling mux_shutdown() and re-enable it afterwards. This way,
we can stop the killer thread at any time and after having joined it
we know it is safe to call mux_close() since there are no more
references to it but us.
- Update the lister, detailer and updater threads to correctly check for
error on read/write/parsing. Generate a proper error message in each
case and return it back to the main thread, along with a status code
indicating either success, failure or a transient failure.
- Always call status_close() in the updater thread, to ensure that the
status file is properly updated even if we are being interrupted by an
error.
- Slightly tweak the threads API to make it match our needs more closely.
- Add a few useful comments here and there.
- Rename proto_init() to the more correct proto_run() name.
- Use the status code returned by the worker thread to only retry a run
when we had a transient error, and to return a proper exit code at the
end of the program's execution.
- Many minor stuff.
All those changes allow csup to properly handle any synchronous or
asynchronous error and to print meaningul messages nearly all the time,
without duplicated messages. Hangs should not happen anymore, even in
case of an error. We also correctly handle being sent some signals such
as SIGINT, by correctly updating the status file and cleaning after us,
so ^C is usable. Finally, csup now returns a proper exit code: 0 in case
of success or 1 in case of an error, similarly to CVSup. Oh, and since
fixups are now supported, I'd call csup "production ready", module the
bugs I have introduced.
Please give this version as much testing as you can!
This snapshot only addresses one problem introduced in the 20060211
snapshot. Because of what seems to be a bug in CVSup, fixing the
keyword expansion code so that it works correctly with OpenBSD CVSup
servers broke updates from PostgreSQL CVSup servers. This snapshot
"fixes" this issue by being bug-to-bug compatible with CVSup.
- Add support for retries when the connection is rejected by the server
and the associated -1 and -r maxRetries options.
- Add missing description for the -4 and -6 options in csup.1.
- Ignore the return value of a chflags() call in fattr_install() to
match CVSup. This fixes csup over NFS.
- Correctly handle any locking error with assertions.
- Make the multiplexer code fully dynamic and cancelable.
- Handle errors in the sender and receiver threads correctly by closing
the multilpexer and waking up all the threads blocked on any of the
channels. This means we don't hang when being disconnected for
instance.
- Make several functions of the chan API, most notably chan_read() and
chan_write() take a struct chan * instead of an id. This saves a
mutex lock and unlock for each call to these functions, and also
reduces the contention on the multiplexer lock.
- Change the stream API so that we can associate a stream with a void *
to support the previous change. Update all the consumers.
- Optimize the scheduling of the sender thread so that it's not
possible to have some channels starve others.
- Optimize mkdirhier() so that it saves many access() calls on average.
- Always set the "no rsync" option to the collections since we don't
support the rsync updating algorithm yet. I have yet to see a CVSup
server trying to send me rsync updates of files in checkout mode but
better safe than sorry.
- Fix the RCS keyword handling. Updating the OpenBSD-src collection
from an OpenBSD CVSup server now works.
- Correctly handle deletion requests for directories.
- And various minor bugfixes.
- Don't use the C99 %zd format in fattr_encode(), it fails with a
pre-C99 compiler and just prints "zd", which caused csup to send
corrupted file attributes and get kicked by the server. I have no
idea why I didn't see this when testing on ref4.FreeBSD.org but csup
now really works on RELENG_4.
- Completely revamp the file attributes computation when checking out
new files and when updating them via diffs. We now create the files
with correct attributes. If you've seen "SetAttrs" commands when you
didn't use to see with CVSup, that fixes it.
- Rewrite mkdirhier() so that it works as intended and respects the
umask setting of the collection.
- Fix a bad bug in proto_get_ascii() that made csup crash when there
were no more tokens to eat.
- Rework the status file API so that it's always possible to determine
if there was an error and to supply appropriate error messages with
status_errmsg(). Use this new function in the lister, detailer and
updater threads.
- Implement support for two new formats in proto_printf(), %f and %F,
that are used to print encoded file attributes. Use them everywhere
appropriate to simplify code.
And probably a few things more that I forgot.
- The -4 and -6 switches have been addede to force usage of IPv4 or IPv6,
respectively,
- The "list=suffix" option in the supfile is now supported.
- And many bugfixes.
- Many bugfixes and code improvements.
- There is now a csup(1) manual page.
- We now use the Z_DEFAULT_COMPRESSION zlib compression level instead of
Z_BEST_SPEED (the lowest). As a consequence of this, less data needs to
be send and csup is slightly faster than CVSup in my tests.
While it still has some rough edges, it has reached a state where
it's pretty much usable and this port will get me a wider audience for
testing.
Reviewed by: ahze
