Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[NativeAOT-LLVM] support System.Net.Http.HttpClient on WASIp2 #2614

Closed
wants to merge 20 commits into from

Conversation

dicej
Copy link

@dicej dicej commented Jun 13, 2024

This adds WasiHttpHandler, a new implementation of HttpMessageHandler based on
wasi:http/outgoing-handler, plus tweaks to System.Threading to allow async Tasks to work in a single-threaded context, with ThreadPool work items dispatched from an application-provided event loop.

WASIp2 supports asynchronous I/O and timers via wasi:io/poll/pollable resource handles. One or more of those handles may be passed to wasi:io/poll/poll, which will block until at least one of them is ready. In order to make this model play nice with C#'s async/await and Task features, we need to reconcile several constraints:

  • WASI is currently single-threaded, and will continue to be that way for a while.
  • C#'s async/await and Task features require a working ThreadPool implementation capable of deferring work.
  • A WASI component can export an arbitrary number of functions to the host, and though they will always be called synchronously from the host, they need to be able to perform asynchronous operations before returning.
  • WASIp3 (currently in the design and prototype phase) will support asynchronous exports, with the top level event loop running in the host instead of the guest, and wasi:io/poll/pollable will no longer exist. Therefore, we don't want to add any temporary public APIs to the .NET runtime which will become obsolete when WASIp3 arrives.

The solution we arrived at looks something like this:

  • Tweak the existing ThreadPool implementation for WASI so that methods such as RequestWorkerThread don't throw PlatformNotSupportedExceptions (i.e. allow work items to be queued even though the "worker thread" is always the same one that is queuing the work)
  • Add two new methods to Thread:
    • internal static void Dispatch: Runs an iteration of the event loop, draining the ThreadPool queue of ready work items and calling wasi:io/poll/poll with any accumulated pollable handles
    • internal static Task Register(int pollableHandle): Registers the specified pollable handle to be polled during the current or next call to Dispatch
    • Note that these methods are internal because they're temporary and should not be part of the public API, but they are intended to be called via UnsafeAccessor by application code (or more precisely, code generated by wit-bindgen for the application)

The upshot is that application code can use wit-bindgen (either directly or via the new componentize-dotnet package) to generate async export bindings which will provide an event loop backed by Thread.Dispatch. Additionally, wit-bindgen can transparently convert any pollable handles returned by WASI imports into Tasks via Thread.Register, allowing the component to await them, pass them to a combinator such as Task.WhenEach, etc.

Later, when WASIp3 arrives and we update the .NET runtime to target it, we'll be able to remove some of this code (and the corresponding code in wit-bindgen) without requiring significant changes to the application developer's experience.

This PR contains a few C# source files that were generated by wit-bindgen from the official WASI WIT files, plus scripts to regenerate them as desired.

@dicej dicej changed the title support System.Net.Http.HttpClient on WASIp2 [NativeAOT-LLVM] support System.Net.Http.HttpClient on WASIp2 Jun 13, 2024
@dicej
Copy link
Author

dicej commented Jun 13, 2024

TODO items:

  • Upstream wit-bindgen changes so we can update the code generation scripts to use an official release
  • Determine the correct way to sort response headers into HttpResponseHeaders and HttpContentHeaders
  • Run applicable existing tests and/or new tests on a WASIp2-capable runtime such as Wasmtime

@jkotas jkotas added the area-NativeAOT-LLVM LLVM generation for Native AOT compilation (including Web Assembly) label Jun 13, 2024
@dicej
Copy link
Author

dicej commented Jun 13, 2024

Ah, that CI failure was useful; now I see how WASI tests are currently run. Looks like I'll need to make some changes to convert the test module(s) into component(s) and run them on a WASIp2-compatible runtime.

@dicej
Copy link
Author

dicej commented Jun 28, 2024

Status update on this: CI is green, and I've manually tested WasiHttpHandler by way of HttpClient to verify it works. However, I could use some guidance on how to automate this testing as part of CI.

AFAICT, CI currently only runs the library smoke tests for both the WASI and browser targets. I can see that e.g. ResponseStreamTest has methods like BrowserHttpHandler_Streaming, so it looks like there is test coverage for TARGET_BROWSER, but I'm not seeing how or where those tests are run, nor how to replicate them for WASI.

I've tried running the System.Net.Http.Functional.Tests suite, but have not had much success with that (many of the tests rely directly or indirectly on TaskAwaiter.GetResult, which leads to an infinite busy wait), and I'm not sure if debugging that is a good use of time.

So my question is: What's the best (read: most practical and efficient) way to add test coverage for WasiHttpHandler such that it can be run as part of CI (or elsewhere if there's some other automated testing infrastructure I'm missing)?

Meanwhile, I'm going to mark this "ready for review" since I believe it's ready for feedback even with the testing question unresolved.

@dicej dicej marked this pull request as ready for review June 28, 2024 15:59
@@ -165,7 +165,7 @@ public static int Main(string[] args)
if (passed && CoreFXTestLibrary.Internal.Runner.NumPassedTests > 0)
{
CoreFXTestLibrary.Logger.LogInformation("All tests PASSED.");
return 100;
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Using 100 to indicate success is a convention used by all tests under src/tests. Changing this convention in the wasm branch is going to be perpetual conflict with upstream. Is this change really necessary?

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Initially, I followed @akoeplinger 's advice to print the exit code to standard out, but then @SingleAccretion advised me to change the codes instead. I'm happy to do whatever you all think is best here.

For context: the reason we need to do something special for WASIp2 here is WebAssembly/wasi-cli#11

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

+1 for xharness solution rather than editing 100 in too many places

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I do not see how you can avoid modifying the tests that are Main-based. There is no code to be hooked there (unless you start thinking of some really involved workaround like making everything use CustomMain).

If we need to modify the tests regardless, the simplest thing is to change the exit code.

(Actually, I wanted to probe grounds for doing this upstream as well, but it may be problematic.)

Copy link
Member

@pavelsavara pavelsavara Jul 1, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

could we just modify the native C main to printf exit code ?
Perhaps only when the application is build with some flags on, so that production applications don't do it.

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

could we just modify the native C main to printf exit code ?
Perhaps only when the application is build with some flags on, so that production applications don't so it.

I have not seen a precedent for adding such test hooks into production code.

It wouldn't work for cases that exit abnormally (Environment.Exit) - we have one such test.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Especially for NAOT the application is recompiled every time, so #ifdef seems OK to me.

For C# exit, we could have hook in PAL. But I'm not sure what to do about abort() somewhere in native code.

Copy link

@SingleAccretion SingleAccretion Jul 1, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Especially for NAOT the application is recompiled every time, so #ifdef seems OK to me. For C# exit, we could have hook in PAL

The bootstrapper, which is where the native main is defined, is only compiled once, as part of the runtime build. I agree it is technically possible to make it work (e. g. via clever use of weak symbols, or compiler intrinsics, or something else).

My point is that it would not be to the scale of the problem, which is just running these smoke tests. If we don't convert all tests to use the zero exit code (unlikely), or don't wait enough for WASI P2.1 to add support for exit codes (possible but not terribly likely), then, when the question of how to run all of the runtime tests comes up - which will be upstream, it will be solved in some more involved manner. At that time, the same solution, if technically possible, will be adopted downstream too.

Edit: we'll be getting functional exit codes in a few months: WebAssembly/wasi-cli#44.

@pavelsavara
Copy link
Member

pavelsavara commented Jul 1, 2024

Meanwhile, I'm going to mark this "ready for review" since I believe it's ready for feedback even with the testing question unresolved.

Yeah, enabling more tests should be independent PR.

It takes non-trivial effort in the first pass.

Also because at the same time it will test host's HTTP stack and interop, including HTTP edge cases.

I've tried running the System.Net.Http.Functional.Tests suite

That's one we want to use but we only use tests which are fully async.
Meaning tests which don't contain blocking Task.Wait and yield to the host/browser.

There is TARGET_BROWSER and PlatformDetection.IsBrowser with variations.
We will need to use the same/similar conditional tests.

[ConditionalTheory(typeof(PlatformDetection), nameof(PlatformDetection.IsBrowserDomSupportedOrNotBrowser))]

[ConditionalFact(typeof(PlatformDetection), nameof(PlatformDetection.IsChromium))]

[ConditionalFact(typeof(PlatformDetection), nameof(PlatformDetection.IsBrowser))]

[ConditionalClass(typeof(PlatformDetection), nameof(PlatformDetection.IsBrowserDomSupportedOrNotBrowser))]

https://github.com/dotnet/runtime/blob/f8bcb89796d3c83264fffd913e5196d29d8d730e/src/libraries/Common/tests/TestUtilities/System/PlatformDetection.cs#L50-L53

https://github.com/dotnet/runtime/blob/f8bcb89796d3c83264fffd913e5196d29d8d730e/src/libraries/Common/tests/TestUtilities/System/PlatformDetection.cs#L147-L151

The HTTP server role is played by NetCoreServer for browser.
https://github.com/dotnet/runtime/tree/main/src/libraries/Common/tests/System/Net/Prerequisites/NetCoreServer

https://github.com/dotnet/runtime/blob/f8bcb89796d3c83264fffd913e5196d29d8d730e/src/libraries/Common/tests/System/Net/Prerequisites/LocalEchoServer.props#L15-L19

https://github.com/dotnet/runtime/blob/main/src/libraries/Common/tests/System/Net/Configuration.Http.cs

Also note, that some of the tests assume HTTP server in the same process.
With browser that's not possible and so we offload the HTTP server role to xharness (helper which is driving the browser).
They communicate via WebSocket, which we don't have for WASI yet.
See https://github.com/dotnet/runtime/tree/main/src/libraries/Common/tests/System/Net/Prerequisites/RemoteLoopServer

Things which are broken should be marked with filter for broken scenario and a link to GH issue for it.

[ActiveIssue("https://github.com/dotnet/runtime/issues/50957", typeof(PlatformDetection), nameof(PlatformDetection.IsBrowser), nameof(PlatformDetection.IsMonoAOT))]

pavelsavara added a commit to pavelsavara/runtime that referenced this pull request Jul 1, 2024
@dicej dicej force-pushed the system-net-http-wasi branch from aecff25 to 28d758f Compare August 26, 2024 20:09
dicej added 3 commits August 26, 2024 16:42
Sometimes there are two Wasm files in Jco's output; sometimes three.  In any
case, hard-coding the number won't fly.

Signed-off-by: Joel Dice <[email protected]>
Signed-off-by: Joel Dice <[email protected]>
Comment on lines +166 to +167
// For WASI/browser/iOS/tvOS we will proactively fallback to polling since FileSystemWatcher is not supported.
if (OperatingSystem.IsWasi() || OperatingSystem.IsBrowser() || (OperatingSystem.IsIOS() && !OperatingSystem.IsMacCatalyst()) || OperatingSystem.IsTvOS())

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Most library changes here should be upstreamed (we will merge them here before that, but it should only be temporary).

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I have many such changes in messy PR here dotnet/runtime#105838
I don't expect to merge it in current form.
The problem is that attributes are visible on the runtime API, so we better do it right on the first attempt.
But I don't know exactly what would be PNSE and what would eventually work. And that will change with next WASI preview.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'll admit I've been using this PR as a dumping ground for miscellaneous WASI fixes as I build demo apps (this one came up when testing AspNetCore). Happy to open separate, upstream PRs as appropriate, but I figured I'd collect them all here first, at least temporarily.

@pavelsavara: yeah, I've been avoiding attributes for the time being since I also don't know which features might be supported in the future. I'm pretty sure WASI will never support signal handling, but I could imagine it might support file notification someday.

@@ -462,7 +462,7 @@
</ItemGroup>

<!-- TODO-LLVM: This is not upstreamable because it makes the build runtime-specific. -->

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
<!-- TODO-LLVM: This is not upstreamable because it makes the build runtime-specific. -->

There is another one above (line 27) that needs to be deleted.

Comment on lines 13 to +14
// TODO-LLVM: This is not upstreamable and should be deleted when https://github.com/dotnet/runtimelab/pull/2614 is merged
#if TARGET_WASI && !NATIVE_AOT
#if TARGET_WASI

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I presume the TODOs here should be removed?

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm rewriting the whole file upstream to respect child resources and more.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I wasn't sure if the TODO-LLVM comments were still relevant (i.e. did they refer to just the && !NATIVE_AOT part, or the whole #if conditional?) Sounds like they're no longer relevant, so I'll remove them.

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

There were introduced very recently in https://github.com/dotnet/runtimelab/pull/2605/files. I assumed it was a fix for some build break.

dicej and others added 10 commits August 28, 2024 09:29
Otherwise, we end up overwriting `response.Content.Headers`.

Signed-off-by: Joel Dice <[email protected]>
If one or more tasks have been canceled during the call to
`ThreadPoolWorkQueue.Dispatch`, one or more tasks of interest to the application
may have completed, so we return control immediately without polling, allowing
the app to exit if it chooses.

A practical example of this is in the SharedLibrary smoke test.  Without this
patch, that test will take over 100 seconds to complete, whereas with this patch
it completes in under a second.

Signed-off-by: Joel Dice <[email protected]>
@dicej dicej force-pushed the system-net-http-wasi branch from c1fd2a7 to 98bf4dc Compare August 29, 2024 18:49
...and hopefully make CI happier.

Signed-off-by: Joel Dice <[email protected]>
I'm still investigating whether this actually _is_ a `wasmtime-wasi-http` bug;
stay tuned.

Signed-off-by: Joel Dice <[email protected]>
@pavelsavara
Copy link
Member

I rebased this PR into #2758

@pavelsavara pavelsavara closed this Nov 8, 2024
dicej added a commit to dicej/runtime that referenced this pull request Nov 14, 2024
This change was extracted from dotnet/runtimelab#2614,
which includes various fixes to enable building on non-Windows systems.  We're
in the process of upstreaming the parts of that PR which are not specific to
NativeAOT-LLVM, and this is the latest such change.

Signed-off-by: Joel Dice <[email protected]>
jakobbotsch pushed a commit to dotnet/runtime that referenced this pull request Nov 15, 2024
…9830)

This change was extracted from dotnet/runtimelab#2614,
which includes various fixes to enable building on non-Windows systems.  We're
in the process of upstreaming the parts of that PR which are not specific to
NativeAOT-LLVM, and this is the latest such change.
mikelle-rogers pushed a commit to mikelle-rogers/runtime that referenced this pull request Dec 10, 2024
…net#109830)

This change was extracted from dotnet/runtimelab#2614,
which includes various fixes to enable building on non-Windows systems.  We're
in the process of upstreaming the parts of that PR which are not specific to
NativeAOT-LLVM, and this is the latest such change.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
area-NativeAOT-LLVM LLVM generation for Native AOT compilation (including Web Assembly)
Projects
None yet
Development

Successfully merging this pull request may close these issues.

7 participants