views:

133

answers:

3

I'm monitoring a folder using a FileSystemWatcher like this:

watcher = new FileSystemWatcher(folder);
watcher.NotifyFilter = NotifyFilters.Size;
watcher.Changed += changedCallback;

When I open a new file in notepad in that folder and save it, I get a notification. If I keep writing and then I save, I get a notification. If I close the file with saving it, I get a notification. Exactly what I wanted.

However, it turns out that if I create a file in that folder and I set its sharing mode to FileShare.Read, and then I write to it, I will not get any notifications until the file is closed. Another workaround is to open the file (e.g. in Notepad), which apparently causes its state to be updated, and then my monitoring application gets the notification. Yet another workaround is a refresh I can do in Windows Explorer, which again causes the file state to be updated.

Interestingly, if I look at Windows Explorer while I make the changes, I notice that: 1. If the file is shared for reading & writing, its size will be updated immediately in Windows Explorer as soon as I save it. 2. If the file is shared for reading only, its size will NOT be update immediately in Windows Explorer, unless I manually refresh the window.

So it seems that my monitoring application shares the same behavior as Windows Explorer. I was thinking about running a thread that will just scan the files in the folder, but I'm wondering if there's anything more elegant to be done in this case.

BTW, I'm using Win7 and I'm not sure that this problem happens on other Windows versions as well.

Thanks!

EDIT: Using ReadDirectoryChanges in C++ got me the same exact results. Implementing the thread I was talking about earlier didn't help as well. I'm wondering what is F5 in Windows Explorer actually doing, because it does cause the change to be reported.

A: 

Looks like the file data is cached and not actually written. When you write something to the file, the data is first placed to cache using certain file system IRP (driver request). Now when the data is actually written to the disk, another IRP is sent. It can be (this is a guess) that FileSystemWatcher catches only second IRP type.

The solution (if possible) is to call Flush on the file you are writing. If you are tracking changes made by other applications, things can become more complicated, of course.

Eugene Mayevski 'EldoS Corp
Thanks. Unfortunately, I'm not controlling the writing application.
Eldad Mor
I think knowing what F5 in Explorer is doing would not help a lot either, since F5 is an active action and you can't take such action just because you don't know when the file is changed.
Eugene Mayevski 'EldoS Corp
I was thinking about simulating the refresh in my own thread. My other option is to rewrite the FileSystemWatcher in a thread of my own that would check the file - but even that might not work if the data is not actually written to the disk. Refreshing Windows Explorer seems to flush the buffer to the disk, which is why I want to understand how it works.
Eldad Mor
Maybe this discussion will help: http://stackoverflow.com/questions/173560/flush-disk-write-cache
Eugene Mayevski 'EldoS Corp
Eugene, thanks for the assistance. Your insights have put me on the right track. You can see my answer for the details.
Eldad Mor
Interesting to know, thank you for pointing at your discoveries.
Eugene Mayevski 'EldoS Corp
+1  A: 

Yes, Explorer uses the exact same API as FileSystemWatcher uses. There's just one, ReadDirectoryChangesW() as you found out.

What you found strongly suggest that Win7 is optimizing the disk write that would be needed to update the directory entry for the file. Delaying it until the last possible moment, when the file is closed. There's an interesting correlation between this observation and a critical bug that a user discovered in the RTM version of Win7. The update sometimes doesn't happen at all. The bug strikes randomly but infrequently, I've seen this myself just once on my machine. Knowingly anyway.

The details are in this thread (beware of very slow server). This still fails today with all Win7 updates applied.

Well, interesting tidbit but not really germane to your question. You do need to alter your code to accommodate the OS works.

Hans Passant
Thanks for the reply, Hans. Still, given that I cannot alter the writing application, do you have any suggestions how I should proceed? I wonder how can I mimic Explorer's F5 (which does update the file size and causes my application to fire the event). I tried getting the files list and even opening them one by one, but it didn't help.
Eldad Mor
Hmm, interesting that Explorer can get your FSW event to fire. I can only think of FlushFileBuffers() but surely that's not it. Maybe you can see it back in a SysInternals' ProcMon stack trace.
Hans Passant
Turns out that Explorer reads the files, leading (this is my own theory) to the cache mechanism flushing the write to the disk, so that the reader could actually get the most updated contents. See my answer to the question for details. Also, thanks for your comments!
Eldad Mor
+1  A: 

The solution to the problem is not to open the files, but to actually READ from them. It's enough to read even one byte, and the Windows cache mechanism will write the contents of the file to the disk, thus allowing you to read them.

I ended up implementing a thread that went over all the files, opened them and read a byte from them. This caused them to change, and triggered the event in the FileSystemWatcher object.

The reason that Windows Explorer F5 works as well, is that Windows actually reads the contents of the file in order to show some extended contents (e.g., thumbnails). Once the file is being read, the cache first writes to the disk, triggering the event in the FSW.

Eldad Mor