Add missing calls to close() to many places where files are
opened. Relying on reference counting to catch them soon-ish is not
portable and fails in environments with a proper GC, such as PyPy.
When path is too long to be an Unix socket address, we create a socket in a
temporary directory and link from the long path to the shorter one.
But checks in server code at startup were insufficient in this case, and used
to raise an unclear "tried linking .hg/inotify.sock to a temporary socket but
.hg/inotify.sock already exists"
This makes it consistent with dirstate.status(), which is important if
there are other extensions messing with the output of status(). Those
extensions can safely assume that dirstate.status() returns a tuple of
lists, because its docstring says it does. But
inotifystatus.dirstate() returns a list of lists, which can break
those other extensions.
With inotify enabled, files that should be ignored could be detected as
untracked by mercurial. This behavior was wrong because inotify's filestatus
implementation only matched filenames against ignore patterns, instead of
checking if other elements of their paths matched them. This patch fixes the
behavior by checking the file paths against the ignore patterns.
A new test has also been added to the main inotify test to prevent any
regressions.
This patch accomplishes the port of the inotify C module to py3k by #including
mercurial's util.h file, and by defining the necessary boilerplate code
required by py3k through conditional compilation.
This patch reimplements the event_repr function. It got mostly rewritten to
eliminate the need for conditional compilation of the module when building in
py3k. The trick there (thanks to Antoine Pitrou) is to use the % operator to
let the python interpreter format the string to be returned.
In older python versions, it was ok to access an object's type by accessing its
ob_type "member". With python 2.6+, the proper way of accessing it is via
Py_TYPE(object). This patch implements the correct call for the inotify
extension. When under python < 2.6, this macro is defined in mercurial's
util.h.
this helps users to know what kind of option is:
- no value is required(flag option)
- value is required
- value is required, and multiple occurrences are allowed
each kinds are shown as below:
-f --force force push
-e --ssh CMD specify ssh command to use
-b --branch BRANCH [+] a specific branch you would like to push
if one or more 3rd type options are shown, explanation for '[+]' mark
is also shown as footnote.
hg inserve was ignoring and miscomputing the --timeout-idle option (seconds
vs. minutes).
Thanks to Jesse Glick for the bugreport and the initial patch.
On Windows, Mercurial can be run from the python script of from a frozen
executable. In the first case, we have to call the python interpreter since the
script is not executable. Frozen executable can be called directly.
Fix 3/3 for issue421
* prefix messages by inotify-(client|server)
* make sure that all warning and abort messages use the same format.
* in the case where inotify.sock is an old broken symlink, say so and abort
instead of trying to overwrite the already existing link
This will mainly help us in our tests to log pids of inotify servers
started implicitely, to make sure that unkilled inotify daemons do not clutter
the output of unrelated tests.
Also desactivate the workaround introduced in 37824a274d63
Original patch was provided by Simon Heimberg
It delegates dirstate computation to dirstate.status when dirstate is dirty:
better be slow from time to time instead of using wrong data.
This solves issue1719. As the last component, issue1810, is still not solved,
test-inotify-dirty-dirstate will fail for now. It emphasizes a regression due
to e8efd88001e7:
changeset: 9515:e8efd88001e7
user: Nicolas Dumazet <nicdumz.commits@gmail.com>
date: Sun Aug 16 11:11:37 2009 +0900
summary: inotify: use cmdutil.service instead of local daemonizing code
Ancestors of 7c01599dd30 are passing the test, when applied this patch.
Regression has to be investigated, but this patch is important since it affects
often mq operations.
Emulate the match.dir calls that are made in dirstate.walk:
* first mark the visited directories on the server side
* then extend the transmitted response to include this directory list
* and lastly call match.dir on each directory
We are not currently using that API function in inotify, hence the reason
for the "silent" bug. But returning NULL here causes the interpreter to crash.
Let's keep code clean for reusers :)
The whole "bail" logic was unneeded here.
self.wprefix = self.repo.root + '/' = self.repo.wjoin('') + '/'
Since we'll eventually get rid of self.repo, keep only the first form.
use self.prefixlen to compute only once the root prefix size.
The biggest problem was the data structure, which did not allow changing
a file into a directory or vice versa. This problem is fixed by 47d29dcb7266.
The walk() method also had an issue in this case:
- we know 'path' as a directory. inotify server sleeps.
- 'path' is deleted
- 'path' is recreated as a file
- the server catches up here, and see the deletion. it instantiates a scan(),
which in its turn calls for walk(repo, path).
- walk() then assumes that 'path' is a directory and calls os.listdir on it,
which raises an OSError(errno.ENOTDIR)
Catch the error, and yield the file instead of the directory contents.
== Rationale for the new structure ==
Current structure was a dictionary tree. One directory was tracked
as a dictionary:
- keys: file/subdir name
- values:
- for a file, the status (a/r/m/...)
- for a subdir, the directory representing the subdir
It allowed efficient lookups, no matter of the type of the terminal leaf:
for part in path.split('/'): tree = tree[part]
However, there is no way to represent a directory and a file with the same name
because keys are conflicting in the dictionary. Concrete example:
Initial state:
root dir
|- foo (file)
|- bar (file)
# data state is: {'foo': 'n', 'bar': 'n'}
Remove foo:
root dir
|- bar (file)
# Data becomes {'foo': 'r'} until next commit.
Add foo, as a directory, and foo/barbar file:
root dir
|- bar (file)
|-> foo (dir)
|- barbar (file)
# New state should be represented as:
{'foo': {'barbar': 'a'}, 'bar': 'n'}
however, the key "foo" is already used and represents the old file.
The dirstate:
D foo
A foo/barbar
cannot be represented, hence the need for a new structure.
== The new structure ==
'directory' class. Represents one directory level.
* Notable attributes:
Two dictionaries:
- 'files' Maps filename -> status for the current dir.
- 'dirs' Maps subdir's name -> directory object representing the subdir
* methods
- walk(), formerly server.walk
- lookup(), old server.lookup
- dir(), old server.dir
This new class allows embedding all the tree walks/lookups in its own class,
instead of having everything mixed together in server.
Incidently, since files and directories are not stored in the same
dictionaries, we are solving the previous key conflict problem.
The small drawback is that lookup operation is a bit more complex:
for a path a/b/c/d/e we have to check twice the leaf, if e is a directory or a
file.
Trying as much as possible to consistently:
- use a present tense predicate followed by a direct object
- verb referring directly to the functionality provided
(ie. not "add command that does this" but simple "do that")
- keep simple and to the point, leaving details for the long help
(width is tight, possibly even more so for translations)
Thanks to timeless, Martin Geisler, Rafael Villar Burke, Dan Villiom
Podlaski Christiansen and others for the helpful suggestions.