fs::remove_dir_all rarely succeeds for large directories on windows

[Diggsey]

I've been trying to track this one down for a while with little success. So far I've been able to confirm that no external programs are holding long-standing locks on any of the files in the directory, and I've confirmed that it happens with trivial rust programs such as the following:

use std::fs;

fn main() {
    println!("{:?}", fs::remove_dir_all("<path>"));
}

I've also confirmed that deleting the folder from explorer (equivalent to using SHFileOperation), or using rmdir on the command line will always succeed.

Currently my best guess is that either windows or other programs are holding transient locks on some of the files in the directory, possibly as part of the directory indexing windows performs to enable efficient searching.

Several factors led me to think this:

• fs::remove_dir_all will pretty much always succeed on the second or third invocations.
• it seems to happen most commonly when dealing with large numbers of text files
• unlocker and other tools show no active handles to any files in the directory

Maybe there should be some sort of automated retry system, such that temporary locks do not hinder progress?

edit:
The errors returned seem somewhat random: sometimes it gives a "directory is not empty" error, while sometimes it will show up as an "access denied" error.

[retep998]

I think this is just race conditions at work. When you delete a file another program might get a notification and check out what happened and that might hold a temporary lock on something. Or it might be that the previous deletions weren't actually done by the time you tried to delete the directory.

[petrochenkov]

Hm, https://github.com/CppCon/CppCon2015/blob/master/Tutorials/Racing%20the%20Filesystem/Racing%20the%20Filesystem%20-%20Niall%20Douglas%20-%20CppCon%202015.pdf slide "3. Deleting a directory tree" may be related. I accidentally watched this presentation yesterday. If I understood correctly, DeleteFileW used in remove_dir_all doesn't actually delete a file, but only marks it for deletion, it still can be alive when we try to delete the parent directory causing the "directory is not empty" error.

[Diggsey]

@petrochenkov Great link - that seems like it's exactly the problem I'm experiencing.

It would be great to see some of these portable race-free idioms implemented in std.

[alexcrichton]

I'd be interested to dig a bit more into this to see what's going on. According to the DeleteFile documentation the only reason the file would be flagged for deletion would be if there are active open handles, which seems like a legitimate race condition? Note though that the same docs also recommend using SHFileOperation for recursively deleting a tree.

Those slides are certainly an interesting read though! I'd be a little wary of putting it into the standard library as the "transactional semantics" are a little weird. For example if you successfully rename a file to a temporary directory and then fail to delete it, what happens?

It may be the case that the best implementation for this function on Windows is to just use SHFileOperation, and that may end up just taking care of these sorts of issues on its own.

[pitdicker]

I will give fixing this a try.

On Windows XP we could use SHFileOperation, and on Vista and up IFileOperation. For Windows store applications there is no easy function. But I would still like to try implementing it by walking the directories and deleting recursively.

Current problems with remove_dir_all:

• can not remove contents if the path becomes longer than MAX_PATH
• files may not be deleted immediately, causing remove_dir to fail
• unable to remove read-only files

And what I hope will fix it:

• use canonicalise() on the base dir to get support for long paths
• generate hash
• read-only files (possibly with hard links) TODO: can the read-only flag be stale?
  • open file
  • remove read-only flag
  • ReOpenFile with FILE_FLAG_DELETE_ON_CLOSE
  • set read-only flag
  • continue as with normal files
• normal files/dirs
  • open with FILE_FLAG_BACKUP_SEMANTICS (for opening dirs)
    FILE_FLAG_OPEN_REPARSE_POINT (for opening reparse points)
    FILE_FLAG_DELETE_ON_CLOSE
  • atomically rename with SetFileInformationByHandle and FILE_RENAME_INFO
    (normal rename on xp)
  • move all files and subdirectories to the parent of the dir to remove
    (not to %TEMP%, may not be on the same drive)
    rename them to rm-[hash]-[nr]

I think we know deleting will fail when opening the file with FILE_FLAG_DELETE_ON_CLOSE fails, so before moving.

[pitdicker]

O, and I walked against this bug when rustbuild (is that the current name) failed after make clean because of this.

[alexcrichton]

@pitdicker thanks for taking this on! We may want to discuss this a bit before moving much more on a solution. One question is how principled should something like fs::remove_dir_all be. For example if fs::remove_dir_all automatically handled read-only files, should fs::remove_file also do the same? Currently we have this "duality" where some fs functions are straight counterparts to the underlying system operations, but some are somewhat fancier like create_dir_all and remove_dir_all. We'd just want to make sure to have an explicitly drawn line here. Also note I'm not really thinking of an RFC here, just some more discussion before jumping to a PR.

I personally feel that remove_dir_all is fine to be fancy and do lots of implicit operations. I would continue to want, however, that remove_file and remove_dir are as straightforward as possible (e.g. don't mirror what's happening here unless it's a similarly small set of operations as to what happens today). Your proposed strategy seems reasonable to me at first glance, and I wouldn't mind reviewing more over a PR.

Curious what others think, though? cc @rust-lang/libs

[Diggsey]

@alexcrichton Hmm, I would prefer remove_file and remove_dir to try to give the same guarantees they have on linux if possible. Specifically that if they succeed, the file or directory should be completely gone.

It's inevitable that people will build their own abstractions over the top of these primitive operations, and I'd very much like to avoid these kind of bugs where it's sufficiently rare that it's almost impossible to track down the true cause of the problem.

[retep998]

Specifically that if they succeed, the file or directory should be completely gone.

This wouldn't really work because files don't necessarily get deleted right away. You can't just move the file somewhere else and then delete it because there isn't always a reasonable place on the same volume to move the file to. Sitting around and indefinitely waiting for the file to vanish isn't ideal either. I'd personally just use IFileOperation for remove_dir_all and call it a day.

[Diggsey]

What if the file is opened with FILE_FLAG_DELETE_ON_CLOSE, no share flags specified, and then is immediately closed? Successfully opening the file means that you have exclusive access, and so closing it should delete the file.

[retep998]

No share flags means nobody else can open it to read/write/delete it. However another handle can still be opened without read/write/delete permissions. Also what happens if you fail to open the file with no share flags? Does it error or fallback to the normal deletion method that provides no guarantees?

[pitdicker]

Great to see so many comments. Just what I hoped for!

@alexcrichton I completely agree to keep remove_file and remove_dir simple, and only let remove_dir_all do magic. Actually removing read-only files is pretty difficult when you also don't want to break hard links. http://stackoverflow.com/questions/3055668/delete-link-to-file-without-clearing-readonly-bit is an example, but I am sure I found some better way somewhere.

@Diggsey I would also like cross-platform consistency, but sometimes Windows is just to different... I think the standard library should at least expose primitives that only need one system call. Otherwise we lose performance and maybe get more vulnerable to race conditions. Maybe we can add a remove_no_matter_what in the future :)

@retep998 I have not yet written the code, so I don't know if this will really solve the problem... But if this doesn't work out I am all for IFileOperation. Meanwhile it is a nice way for me and maybe the standard library to get better at avoiding filesystem races.

I would have to test what happens with FILE_FLAG_DELETE_ON_CLOSE, but in theory if the file opens successfully we only have to move it to avoid a race of a couple of milliseconds. Directly after opening the file no one else can open the file anymore (it is not a share flag).

What is difficult is where to move a file / dir to. I think the parent of the dir we are removing is good enough. It should be on the same volume, and I think we also have write permission to it because otherwise deleting a dir is not possible (wild speculation here, it is late :))

[brson]

I feel ok about going to extra effort to make remove_dir_all work as expected.

[alexcrichton]

@pitdicker perhaps you could draw up a list of the possible implementation strategies for these functions? That'd also help in terms of weighing the pros/cons of each strategy and we could get a good idea of which terminology is associated with what. So far it sounds like IFileOperation may be the best option here, but I'm personally at least not following what FILE_FLAG_DELETE_ON_CLOSE would mean, especially in the case that you can't open it as someone else does.