Allow `weakly_canonical` to be used with Windows long paths

[emmenlau]

Currently I'm unable to call weakly_canonical() with a long path on Windows. Long path in this context means the Microsoft Windows extension described at https://docs.microsoft.com/en-us/windows/win32/fileio/maximum-file-path-limitation. A path can be prefixed by \\?\ and then can have more than 260 characters (with a number of restrictions).

I could trace the problem of using long paths in weakly_canonical() to the method create_file_handle() in

filesystem/src/windows_tools.hpp

Line 259 in ea22e76

return ::CreateFileW(p.c_str(), dwDesiredAccess, dwShareMode, lpSecurityAttributes, dwCreationDisposition, dwFlagsAndAttributes, hTemplateFile);

This is called via:

filesystem/src/operations.cpp

Line 4437 in ea22e76

file_status head_status = detail::status_impl(head, &local_ec);

But the call fails for long paths. Since in weakly_canonical(), the handle is only needed for the root item, I could work around this by adding some extra handling to create_file_handle(). My added logic is:

• If the given path has a long-path prefix but is less than 260 characters long, remove the long path prefix and rely on Windows to do the right thing.

Is that reasonable, and could something like this be added to the library?

[Lastique]

If you need to operate on long paths on Windows, it is up to the application to prepend the path with \\?\ prefix before passing to Boost.Filesystem. Boost.Filesystem does not do that for you, and I don't think I want to make it do so. Mostly because that would introduce unnecessary inefficiency and complicate implementation quite substantially.

If you're on a recent enough Windows, I would also recommend to enable long paths by default system-wide.

[emmenlau]

Dear @Lastique , thanks for the quick reply!

My problem is actually the opposite. I do have prepend the path with \\?\ prefix before passing to Boost.Filesystem. This is causing the problem in

filesystem/src/windows_tools.hpp

Line 259 in ea22e76

return ::CreateFileW(p.c_str(), dwDesiredAccess, dwShareMode, lpSecurityAttributes, dwCreationDisposition, dwFlagsAndAttributes, hTemplateFile);

[Lastique]

I don't understand, what problem? Does the call fail with the \\?\ prefix? With what error code? Can you prepare a small reproducer code?

Note that the prefix is only valid for absolute paths, so make sure you're not adding it if the path is relative.

[emmenlau]

I don't understand, what problem? Does the call fail with the \\?\ prefix? With what error code? Can you prepare a small reproducer code?

I assumed that my description was more clear, sorry if it was not!

Here are more details: The call of ::CreateFileW() fails, if the path has a prefix \\?\. If the path has the prefix, then ::CreateFileW() does return a handle, but the handle is all zeros. I'm not knowledgeable about Windows methods, so I do not understand if a handle of zero is correct. I just assume it is wrong. I assume this because subsequently file_status head_status = detail::status_impl(head, &local_ec); showed an error status.

As a test, I added code above ::CreateFileW() that removes the prefix \\?\, and the error was gone. Therefore I conclude that ::CreateFileW() does not support \\?\.

[emmenlau]

So here is my workaround that fixes the problem for me:

diff --git a/libs/filesystem/src/windows_tools.hpp b/libs/filesystem/src/windows_tools.hpp
index 469be041e..e49ecd6a4 100644
--- a/libs/filesystem/src/windows_tools.hpp
+++ b/libs/filesystem/src/windows_tools.hpp
@@ -179,7 +179,11 @@ struct handle_wrapper
 //! Creates a file handle
 inline HANDLE create_file_handle(boost::filesystem::path const& p, DWORD dwDesiredAccess, DWORD dwShareMode, LPSECURITY_ATTRIBUTES lpSecurityAttributes, DWORD dwCreationDisposition, DWORD dwFlagsAndAttributes, HANDLE hTemplateFile = NULL)
 {
-    return ::CreateFileW(p.c_str(), dwDesiredAccess, dwShareMode, lpSecurityAttributes, dwCreationDisposition, dwFlagsAndAttributes, hTemplateFile);
+    std::wstring p_str = p.wstring();
+    if ((p_str.size() < 260) && (p_str.size() > 4) && (p_str[0] == L'\\') && (p_str[1] == L'\\') && (p_str[2] == L'?') && (p_str[3] == L'\\')) {
+        p_str = p_str.substr(4);
+    }
+    return ::CreateFileW(p_str.c_str(), dwDesiredAccess, dwShareMode, lpSecurityAttributes, dwCreationDisposition, dwFlagsAndAttributes, hTemplateFile);
 }
 
 } // namespace detail

I'm not saying that this is beautiful. But it just shows that there would be a way to avoid the problem.

[Lastique]

CreateFileW returns INVALID_HANDLE_VALUE on error, which is (HANDLE)-1 (i.e. not zero). A zero returned handle is not considered an error.

Also, I asked about the error code and reproducer because this could give a clue as to what the problem is.

[emmenlau]

Also, I asked about the error code and reproducer because this could give a clue as to what the problem is.

Yes, I understand that. I do not currently have access to a Windows machine, but I'll try to send more information when I have.

[Lastique]

CreateFileW is listed among the APIs that do support long paths. I have added a test to verify this.

I suspect the problem is with the path you're using. The \\?\ prefix is not simply allowing for longer paths, it actually disables a great deal of path processing in Windows API. This means that the path must be absolute, must use backslashes as separators, and must not contain dot (".") and dot-dot ("..") elements. If you're using an UNC path then the prefix must be \\?\UNC\ instead of \\. Please verify that the path you're passing to weakly_canonical meets these requirements.

[emmenlau]

Thanks a lot for the follow-up!

What I can say so from my debugger session yesterday is that weakly_canonical was calling CreateFileW with only the head of the path in file_status head_status = detail::status_impl(head, &local_ec);. From what I recall from the debugger, the head of the path was something like \\?\C: and nothing else, maybe with one more backslash.

Please verify that the path you're passing to weakly_canonical meets these requirements.

I have a method that takes care of these requirements before passing the path to weakly_canonical. It's possible that the method is buggy, but it did work for thousands of paths with canonical, and just failed for weakly_canonical so far. But I don't want to guess, so I'll try to reproduce the issue and let you know what I find!

[Lastique]

From what I recall from the debugger, the head of the path was something like \\?\C: and nothing else, maybe with one more backslash.

Yes, that was part of how weakly_canonical works, and this turned out to be the problem because the "C:" path is not absolute. I have committed a fix that should resolve this problem.

Thank you for reporting this.

[emmenlau]

Ok, I'm sitting at the Windows computer and can reproduce the issue now.

The path I pass to weakly_canonical is \\\\?\\C:\\data\\aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa. The double backslashes in the string are copied from source code, but the actual string should have single backslashes.

With this path, I get exception boost::filesystem::filesystem_error with description boost::filesystem::weakly_canonical: Incorrect function [system:1]: "\\?\C:"

When I execute the test in the debugger, it stops with an exception in operations.cpp line 4076. In my version of boost (1.79.0) that line reads:

BOOST_FILESYSTEM_THROW(filesystem_error("boost::filesystem::weakly_canonical", head, local_ec));

The debugger says about the exception:

Exception thrown at 0x00007FFBC778474C in BaseFileTest.exe: Microsoft C++ exception: boost::filesystem::filesystem_error at memory location 0x0000001AA913EDE0.

[emmenlau]

But I see its fixed now, thanks!