dart:io readAsBytes will crash if the file size is greater than INT_MAX on flutter windows embedder

[jonahwilliams]

Repro:

void main() async {
  await File('path/to/huge_file').readAsBytes();
}

The file I am using is approximately 3.3 GB. INT_MAX bytes in GB is approximately 2.147483647 GB.

I'm not able to get a backtrace, but the popup that shows the assert shows the path minkernel\crts\ucrt\src\appcrt\lowio\read.cpp

[jonahwilliams]

The error message only appeared in an unopt flutter windows build. In an opt build it crashes or is killed immediately.

[jonahwilliams]

Also: yes, we should probably not ever try to open a file this big all at once.

[dnfield]

When it's this big it should probably get mmapped... whatever the equivalent of that API is on Windows, I can't remember anymore.

[jonahwilliams]

I think we can't mmap them today because dart:io APIs return mutable buffers. If we could change these APIs to return immutable buffers (#50069) then we would not only use less memory, but lots of silly patterns like the FileImage would be faster too

Edit: can't mmap

[zanderso]

Reads are happening with https://github.com/dart-lang/sdk/blob/main/runtime/platform/utils_win.cc#L83.

Docs here https://learn.microsoft.com/en-us/cpp/c-runtime-library/reference/read?view=msvc-170

say:

If buffer is NULL, or if buffer_size > INT_MAX, the invalid parameter handler is invoked. If execution is allowed to continue, the function returns -1 and errno is set to EINVAL.

The Flutter Windows embedder may need to do something like:

https://github.com/dart-lang/sdk/blob/main/runtime/bin/platform_win.cc#L37

if it isn't already, so that _read will return an error instead of crashing.

[zanderso]

cc @a-siva @brianquinlan

[jonahwilliams]

On the Dart CLI I get:

FileSystemException: read failed, path = 'C:\Users\Jonah\Downloads\supervhs-win.zip' (OS Error: The operation completed successfully.
, errno = 0)

Which ... is still wrong, but definitely better than exiting.

[zanderso]

The additional work between the error and creating the OSError object is probably clobbering errno, e.g. here: https://github.com/dart-lang/sdk/blob/main/runtime/bin/file.cc#L1302

[a-siva]

Thanks for pointing out the exact line, will fix it.

[a-siva]

I think we can't mmap them today because dart:io APIs return mutable buffers. If we could change these APIs to return immutable buffers (#50069) then we would not only use less memory, but lots of silly patterns like the FileImage would be faster too

Edit: can't mmap

See a breaking change proposal to add a mmap API to random access files in dart:io #49810 would this be useful for Flutter or is your case pretty much limited to the engine mmaping a file and creating these unmodifiable typed data buffers ?

[jonahwilliams]

That would be useful for the Flutter tool, which needs to load file bytes to compute hashes. this is generally quite slow for binary artifacts/dependencies.

This would also be useful for Image.file or similar approaches where we have no intention to modify the file bytes. That said, I think there are a lot of File implementers, which might make it hard breaking change to land.

[jonahwilliams]

I see that Platform::Initialize is what ultimately sets up the right handlers so that we can catch this error. However, I only see this method invoked from runtime/bin/main.cc and not via any of the embedder setup methods

[zanderso]

Ahh, sorry. I thought that was called from BootstrapDartIo, which the Engine does call, but it looks like it isn't:

https://github.com/dart-lang/sdk/blob/main/runtime/bin/dart_io_api_impl.cc#L22

I think the Engine may want Engine embedders to do this explicitly themselves, though. The setup that the Dart CLI uses might not be exactly the right thing for Flutter's Windows embedder, for example.