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

On Wed, Mar 3, 2021 at 1:42 PM Kenton Varda <[email protected]> wrote:

> On Wed, Mar 3, 2021 at 2:11 PM John Demme <[email protected]> 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
>>
>> On Wed, Mar 3, 2021 at 11:02 AM pepijn de vos <[email protected]>
>> wrote:
>>
>>> 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".
>>>
>>> On Wed, Mar 3, 2021 at 7:57 PM pepijn de vos <[email protected]>
>>> wrote:
>>>
>>>> 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
>>>>
>>>> On Wed, Mar 3, 2021 at 4:42 PM Kenton Varda <[email protected]>
>>>> wrote:
>>>>
>>>>> 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.
>>>>>
>>>>> KJ documentation can be found here:
>>>>> https://github.com/capnproto/capnproto/tree/master/kjdoc
>>>>>
>>>>> 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
>>>>>
>>>>> On Wed, Mar 3, 2021 at 5:54 AM [email protected] <
>>>>> [email protected]> wrote:
>>>>>
>>>>>> 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
>>>>>>
>>>>>> https://github.com/capnproto/capnproto/blob/master/c%2B%2B/src/kj/filesystem.h
>>>>>> 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
>>>>>>
>>>>>> --
>>>>>> 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 [email protected].
>>>>>> To view this discussion on the web visit
>>>>>> https://groups.google.com/d/msgid/capnproto/cd5ce66c-b12e-4612-b383-32462f047f69n%40googlegroups.com
>>>>>> <https://groups.google.com/d/msgid/capnproto/cd5ce66c-b12e-4612-b383-32462f047f69n%40googlegroups.com?utm_medium=email&utm_source=footer>
>>>>>> .
>>>>>>
>>>>> --
>>> 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 [email protected].
>>> To view this discussion on the web visit
>>> https://groups.google.com/d/msgid/capnproto/CANPfQguo2bq4G0tWPeAgzphdSpU4isMkxRUzNmxjCjyvvaxO0w%40mail.gmail.com
>>> <https://groups.google.com/d/msgid/capnproto/CANPfQguo2bq4G0tWPeAgzphdSpU4isMkxRUzNmxjCjyvvaxO0w%40mail.gmail.com?utm_medium=email&utm_source=footer>
>>> .
>>>
>>

-- 
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 [email protected].
To view this discussion on the web visit 
https://groups.google.com/d/msgid/capnproto/CAOjmg%3Du6STAueobBYoW96VMZhCBwVCGyoLQkVn9SjA36e3G_zQ%40mail.gmail.com.

Reply via email to