<filesystem>: exists() fails with "Access denied" for certain type of symlinks

[alexlipa91]

Describe the bug

We use std::filesystem::exists to check existence of paths on the system. Since few days we started to see an error where this function would fail with Access Denied. Unfortunately I could not get a fully reproducible example but I am gonna describe what I noticed.

This is the logged error

Exists D:\some_path\WebKit.framework = 0 with ec 5 and msg Access is denied.

This is the list of paths inside some_path

d-----        11/15/2024   4:51 AM                _SceneKit_SwiftUI.framework
d-----        11/15/2024   4:51 AM                _StoreKit_SwiftUI.framework
d-----        11/15/2024   4:51 AM                _SwiftData_SwiftUI.framework
d-----        11/15/2024   4:51 AM                _Translation_SwiftUI.framework
-a---l        11/15/2024   4:51 AM              0 AuthenticationServices.framework
-a---l        11/15/2024   4:51 AM              0 JavaScriptCore.framework
-a---l        11/15/2024   4:51 AM              0 SafariServices.framework
-a---l        11/15/2024   4:51 AM              0 WebKit.framework

The function fails on all the paths having the a and l bit on while it succeeds on the others.
Permissions should not be the problem: I tried with different users and with all of them I can read the files using powershell (with a simple ls).

Other things worth mentioning: these are symlinks that point to other symlinks

$ stat Modules
  File: Modules -> Versions/Current/Modules
  Size: 24              Blocks: 0          IO Block: 65536  symbolic link

$ stat Versions/Current/Modules
  File: Versions/Current/Modules
  Size: 0               Blocks: 0          IO Block: 65536  directory

$ stat Versions/Current
  File: Versions/Current -> A
  Size: 1               Blocks: 0          IO Block: 65536  symbolic link

$ stat Versions/A/
  File: Versions/A/
  Size: 0               Blocks: 0          IO Block: 65536  directory

again, I tried to repro by re-building a folder structure of this type but I couldn't reproduce the error

Any advice?

[CaseyCarter]

I can reproduce the behavior by creating a non-directory symlink to a directory:

C:\tmp>type exists.cpp
#include <filesystem>
#include <print>

int main(int argc, const char* argv[]) {
    int error_count = 0;
    for (int i = 1; i < argc; ++i) {
        std::error_code ec;
        std::filesystem::path p{argv[i]};
        auto const result = std::filesystem::exists(p, ec);
        if (ec) ++error_count;
        std::println("{:?} {}: {}\n", p.string(), (result ? "exists" : "does not exist"), ec.message());
    }
}

C:\tmp>cl /nologo /std:c++latest /EHsc exists.cpp
exists.cpp

C:\tmp>md target

C:\tmp>mklink /d source target
symbolic link created for source <<===>> target

C:\tmp>exists target source
"target" exists: The operation completed successfully.

"source" exists: The operation completed successfully.

C:\tmp>rd source

C:\tmp>mklink source target
symbolic link created for source <<===>> target

C:\tmp>exists target source
"target" exists: The operation completed successfully.

"source" does not exist: Access is denied.

I'm not sure why EACCESS is the error code for this condition, but I didn't implement the OS so I assume it's intended. Is this your scenario?

[StephanTLavavej]

We talked about this at the weekly maintainer meeting and @CaseyCarter's repro needs more investigation to determine whether it's a bug or by design according to Windows filesystem rules.

[YexuanXiao]

Three years ago, I encountered the same issue and concluded that it's due to the design of std::filesystem and Windows. In fact, I think std::filesystem does not accurately reflect the design of both Windows and POSIX, whether it's the good parts or the quirks.

[CaseyCarter]

@YexuanXiao Thanks for the archaeology! I agree completely with @strega-nil-ms' assertion that we want to report an error for this case. Is there a more appropriate error than EACCESS, or should we continue to return the error code we get from the underlying Win32 API?

[YexuanXiao]

[fs.dir.entry.obs] points out that exist uses status, and status will follow symlinks. In the STL, this falls back to CreateFileW, and when the current issue occurs, it gets INVALID_HANDLE_VALUE, then retrieves the error code from GetLastError.

I also noticed a discrepancy: when a symlink pointing to a file actually points to a directory, GetLastError returns 5 (Access Denied), causing an exception to be thrown; but when a symlink pointing to a directory actually points to a file, GetLastError returns 267 (Dictionary name is invalid) and no exception is thrown. #4844 changed the previous behavior, and it may need to be reconsidered. If we maintain the behavior of throwing an exception, I think we could reinvent the message while retaining the original error code to inform the user about what really happened, rather than the current unintelligible message. @CaseyCarter @StephanTLavavej