RFR: 8293067: (fs) Implement WatchService using system library (macOS) [v11]
Maxim Kartashev
mkartashev at openjdk.org
Thu Jul 20 09:16:20 UTC 2023
On Tue, 29 Nov 2022 10:47:16 GMT, Maxim Kartashev <mkartashev at openjdk.org> wrote:
>> This is an implementation of `WatchService` based on File System Events API that is capable of generating events whenever a change occurs in an interesting directory or underneath it. Since the API naturally supports "recursive" watch, the `FILE_TREE` is supported by the watch service.
>>
>> Some things of note:
>> * There's one "service" thread per `WatchService` instance that is inactive unless changes occur in the watched directory. The changes are grouped by introducing a time delay between when they occurred and when they are reported, which is controlled by the sensitivity modifier of the watch service.
>> * Since FSEvents API reports directories only, the watch service keeps a snapshot (hierarchical if necessary) of the files in the directory being watched. The snapshot gets updated when an event in that directory or underneath it gets delivered. File changes are detected by comparing "last modified" time with a millisecond precision (`BasicFileAttributes.lastModifiedTime()`).
>> * There is a slight complication with the move of an entire directory hierarchy: FSEvents API only reports about the containing directory of that move and not about any of the directories actually moved. There's a separate test for that (`Move.java`).
>> * The code is careful not to do any I/O (such as reading the contents of a directory or attributes of a file) unless unavoidable. Any deviation from this line should be considered a bug (of, arguably, low priority).
>> * The native part consists mostly of API wrappers with one exception of the callback function invoked by the system to report the events that occurred. The sole task of the function is to convert from C strings to Java strings and pass the array of affected directories to Java code. This can be rewritten if desired to make the code more future-proof.
>>
>> This commit leaves `PollingWatchService` unused. I'm not sure if I should/can do anything about it. Any advice is welcomed.
>>
>> ### Testing
>>
>> * Tested by running `test/jdk/java/nio/file` and `test/jdk/jdk/nio` on MacOS 10.15.7 (x64) and `test/jdk/java/nio/` plus `test/jdk/jdk/nio` on MacOS 12.5.1 (aarch64).
>> * Also verified that new tests pass on Linux and Windows.
>> * This code (albeit in a slightly modified form) has been in use at JetBrains for around half a year and a few bugs have been found and fixed during that time period.
>
> Maxim Kartashev has updated the pull request incrementally with one additional commit since the last revision:
>
> Check FSEventStreamStart() return value
For the record: I finally got around to trying the new dispatch queues API. This reduces the complexity of the current implementation quite a bit and passes all tests except one. The obstacle I have not been able to overcome is this: the new implementation consistently fails the test `test/jdk/java/nio/file/WatchService/LotsOfCancels.java`, which creates many watch services and constantly starts/stops FSEventStream's.
I have not been able to find any solution to this; perhaps we are hitting some OS resource limit, even though not that many streams are active at any given moment (sometimes you can also see the "too many open files" exception from that test when it creates new temporary files).
So far it loos like this new API is no go.
-------------
PR Comment: https://git.openjdk.org/jdk/pull/10140#issuecomment-1643564943
More information about the nio-dev
mailing list