How to write to a file from an RPC server?

13 views
Skip to first unread message

pepij...@gmail.com

unread,
Mar 3, 2021, 6:54:33 AMMar 3
to Cap'n Proto
Hey,

I'm just getting started and pretty confused.
The RPC server is all async and stuff, so no blocking IO in there.
Then there are some vague mentions that you should use KJ for most things.
But KJ documentation is nowhere to be found...

All I want to do is just open a file and write to it.
After resorting to reading the source code, I found
So I managed to create a path so far.
Then I started to look for how to open a file.
I found some methods on Directory.
So then I started to look how to make a directory.
Then I found a Filesystem, which is abstract.
Then I found newDiskFilesystem which tells me
DO NOT CALL THIS except in main()
Well **** how am I supposed to open a file inside my server handler then??

Pepijn

Kenton Varda

unread,
Mar 3, 2021, 10:42:02 AMMar 3
to pepij...@gmail.com, Cap'n Proto
Hi Pepijn,

The full comment you refer to says:

// DO NOT CALL THIS except at the top level of your program, e.g. in main(). Anywhere else, you
// should instead have your caller pass in a Filesystem object, or a specific Directory object,
// or whatever it is that your code needs. This ensures that your code supports dependency
// injection, which makes it more reusable and testable.

As the comment says, the idea is that you construct your Filesystem object at the top level of your program, then you pass it (or specific File or Directory objects) down to whatever parts of your program need it. That ensures that those parts of your program can easily be unit-tested against a fake in-memory filesystem, by swapping out the objects that you pass down. This is trying to help you make your code more flexible, but you can ignore it if you want.

You also don't have to use the KJ filesystem API. In fact, nothing else in KJ or Cap'n Proto cares what you use to open files. KJ is a toolkit, meaning it provides a bunch of tools, but doesn't force you to use them if you don't want to.


Please note that Cap'n Proto and KJ are open source software provided for free. I'd like them to be useful but I don't get any actual benefit from you using them, and I'm not interested in helping people who are rude. So, if you have further questions, please try to tone it down a bit. Thanks.

-Kenton

--
You received this message because you are subscribed to the Google Groups "Cap'n Proto" group.
To unsubscribe from this group and stop receiving emails from it, send an email to capnproto+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/capnproto/cd5ce66c-b12e-4612-b383-32462f047f69n%40googlegroups.com.

pepijn de vos

unread,
Mar 3, 2021, 1:57:16 PMMar 3
to Kenton Varda, Cap'n Proto
Hey Kenton,

Thanks for the answer, and sorry my frustration in trying to find out how to do async file IO in cap'n proto came across rude.

What puzzles me is your suggestion that I don't need to use KJ.
What puzzles me even more is that the KJ file I just obtained seems to not support async operations at all.
Are you suggesting it's actually fine to just do blocking IO in the RPC eventloop?
Is there some KJ thing to do async file IO, or a way to use other libraries for async file IO with the KJ eventloop?

Thanks to all the people working on cap'n proto, it seems pretty neat :)

Cheers,
Pepijn

pepijn de vos

unread,
Mar 3, 2021, 2:02:37 PMMar 3
to Kenton Varda, Cap'n Proto
Oh I just went hunting for the asynch bits and in the header it actually say there is no such thing as async file IO

// =======================================================================================
// The filesystem API
//
// This API is strictly synchronous because, unfortunately, there's no such thing as asynchronous
// filesystem access in practice. The filesystem drivers on Linux are written to assume they can
// block. The AIO API is only actually asynchronous for reading/writing the raw file blocks, but if
// the filesystem needs to be involved (to allocate blocks, update metadata, etc.) that will block.
// It's best to imagine that the filesystem is just another tier of memory that happens to be
// slower than RAM (which is slower than L3 cache, which is slower than L2, which is slower than
// L1). You can't do asynchronous RAM access so why asynchronous filesystem? The only way to
// parallelize these is using threads.
//
// All KJ filesystem objects are thread-safe, and so all methods are marked "const" (even write
// methods). Of course, if you concurrently write the same bytes of a file from multiple threads,
// it's unspecified which write will "win".

John Demme

unread,
Mar 3, 2021, 3:11:07 PMMar 3
to pepijn de vos, Kenton Varda, Cap'n Proto
Some operating systems (*cough* Windows *cough*) don't support non-blocking file (or inter-process pipe) I/O without some unreliable hacks. AFAICT.

~John

Kenton Varda

unread,
Mar 3, 2021, 4:42:09 PMMar 3
to John Demme, pepijn de vos, Cap'n Proto
On Wed, Mar 3, 2021 at 2:11 PM John Demme <teqd...@gmail.com> wrote:
Some operating systems (*cough* Windows *cough*) don't support non-blocking file (or inter-process pipe) I/O without some unreliable hacks. AFAICT.

Hate to say it, but my understanding is that Windows has much more robust async filesystem support than Linux does. The comment that Pepijn found is about Linux, not Windows.

Linux filesystem implementations are not async (at least for metadata / directory walking), therefore the only way to access them in an async way is to use threads -- either in the kernel, or in userspace. The kernel isn't very excited about spawning kernel threads transparently -- it would rather than you create userspace threads and make blocking calls.

That said, this is changing a bit with io_uring, where kernel threads running asynchronously from userspace threads are becoming more of a thing. Still, on the kernel side, there are threads, even if it looks like "async" I/O from the userspace side.

Disclaimer: I'm not a kernel developer, this is my second-hand understanding. Also, I know almost nothing about Windows, except that async file I/O has featured much more prominently in the Win32 API going back decades, so I'm guessing it's also baked deeper into the kernel.

-Kenton

John Demme

unread,
Mar 3, 2021, 4:54:36 PMMar 3
to Kenton Varda, pepijn de vos, Cap'n Proto
I think in Windows, async IO is different from non-blocking IO. I haven't used Windows async IO, other than briefly trying to use C# async IO, getting very confused, and (perhaps prematurely) that C# async is intended to be async everywhere (without some really nasty hacks) and you can only use async IO in async code.

~John

Kenton Varda

unread,
Mar 3, 2021, 5:02:47 PMMar 3
to John Demme, pepijn de vos, Cap'n Proto
Hmm when I say "async I/O" I mean the same thing as "non-blocking I/O".

I haven't worked in C#, but have done some stuff with the Win32 C API, where they call it "overlapped" I/O.

-Kenton

John Demme

unread,
Mar 3, 2021, 5:20:26 PMMar 3
to Kenton Varda, pepijn de vos, Cap'n Proto
I could easily be confusing the C# API with the Win32 API, but I recall seeing some StackOverflow posts which indicated there was no way to do what I wanted to do. This was some time ago and my memory ain't so great. My hack-y solution was to spin up a thread for each file/pipe I needed to read from and send all the data down a thread safe queue, which could then implement a 'select()' like mechanism, which is what I really wanted.

... but I no longer have to care about that. At least until I do again.

~John
Reply all
Reply to author
Forward
0 new messages