Hi Nathan, Hi Daniel, Thank you both for your replies.
As a workaround I ended up doing exactly the things you mentioned. After the cancellation of the operation I call additional code to "cleanup". This code goes through open file handles held by the current process and closes the ones left over in .svn\tmp. That being said, I don't know how easy/hard this is to do on Linux, but doing this kind of stuff on Windows is a giant pain in the a**. It involves calling a bunch of poorly documented, low-level native methods from Windows' internal api. Luckily, I found some code online and modified it in order to work on both 32 and 64 bit apps. For reference, you can find this code on my github <https://github.com/urosjovanovic/MceController/blob/master/VmcServices/DetectOpenFiles.cs> . So the final process is as follows: 1. Perform checkout 2. Cancel checkout 3. If present, force-close the open file handle in .svn\tmp 4. Perform a regular svn cleanup to break locks etc. It works, but now I have an additional ~700 LOC to maintain just in order to detect and close those rogue file handles, hoping it will not crash the app. :) If you want, I can provide a test case written in C# using SharpSvn which will work exactly as Daniel wrote, but on Windows. I can also use the repo of your choice, if you have a preferred one for testing. Regards, Uros On Wed, Sep 9, 2020 at 8:18 AM Daniel Sahlberg <daniel.l.sahlb...@gmail.com> wrote: > Den ons 9 sep. 2020 kl 06:44 skrev Nathan Hartman < > hartman.nat...@gmail.com>: > >> On Tue, Sep 8, 2020 at 6:08 AM Uroš Jovanović <uros...@gmail.com> wrote: >> > >> > Then, mid downloading some of the larger files a temp file will appear >> in .svn\tmp. Once that happens, hit the Cancel button. >> > It will signal the cancellation to the svn client and it will throw the >> SvnOperationCanceledException, the SvnClient gets disposed BUT an open file >> handle remains on ".svn\tmp\svn-XYZ123" file. >> > If you try to delete it, Windows will complain that it is used by our >> test app. :( >> >> Moving this to the dev@ list... >> >> Potentially long-running APIs such as 'checkout' allow the client to >> provide a 'cancel_func' callback, which is called at various strategic >> places to ask the client whether the operation should be canceled. >> >> It sounds to me like one of those places sees a cancel request and >> returns to its caller, forgetting to do some cleanup. >> >> Last night I tried to find such a place by reading code. >> >> The 'checkout' command sets up a working copy (if necessary) and then >> calls the 'update' logic to do the heavy lifting. >> >> The 'update' logic is quite involved as it handles all sorts of >> possibilities, which means the number of branches of the call tree >> that need to be checked are too numerous for my code reading approach >> to be sensible. >> >> My thoughts for an automated approach, provided there is a way for a >> process to inquire how many open file handles it has (I assume there >> is a way; I've just never had to do this): The idea is to write a >> minimal client that does the following (on a ramdrive): >> >> 1. Check out a working copy of a repository, giving a cancel_func 'A' >> that increments a global variable 'n' each time it is called and >> always returns "don't cancel." >> >> 2. Loop n times, the loop counter being a global variable 'x': >> >> 2.1: Delete the working copy. >> >> 2.2: Check out a working copy of the same repository, giving a >> different cancel_func 'B' that returns "don't cancel" the >> first (x - 1) times it is called, and returns "cancel" the >> x-th time it is called. >> >> 2.3: Test whether there are open file handles. If there are, we >> know at which iteration the cleanup is not done, and we break >> out of the loop. >> >> 3. If x >= n, quit; we didn't find the problem. >> >> 4. Delete the working copy. >> >> 5. Check out a working copy of the same repository, giving a different >> cancel_func 'C' that returns "don't cancel" the first (x - 1) times >> it is called, and traps the x-th time it is called, allowing the >> call stack to be examined. >> >> Notes and caveats: >> >> 1. This could run for days (or years). >> >> 2. Then again, if it can be exposed pretty reliably by a user hitting >> a Cancel button in a GUI, that means cancel_func is called >> frequently enough from the offending location that it should >> (hopefully) be caught relatively soon in the process. >> >> 3. I think a huge repository isn't needed. The Greek Tree used by the >> test suite may suffice. If it doesn't expose the bug, I'd retry >> with a larger file thrown in. If that doesn't expose it, add >> increasing complexity such as externals, etc. >> >> 4. This relies on the logic being executed identically for each >> checkout (i.e., cancel_func is called the same number of times from >> the same call sites). >> >> 5. No idea how this could be turned into a regression test. >> >> 6. If there's a better way, I'd love to hear it! >> > > For a regression test (as well as trying to pinpoint what goes wrong), > wouldn't it be enough if the cancel_func check for the presence of a file > in .svn/tmp (maybe even checking if it is open - in Linux that should be > easy enough to check in /proc/$PID/fd) and then signal to cancel. That > would "only" need a repository/file that is large enough to trigger calling > the cancel_func. > > I checked quickly and I also see the open file when checking out using > TortoiseSVN and cancelling and it seems to occur all the time. > > Kind regards > Daniel >
