Draft of WASI Preview 2 support, featuring `wasi-sockets` #449

dicej · 2023-12-12T21:05:41Z

Per #447, this is an (in-progress) implementation of wasi-sockets, including a new wasm32-wasi-preview2 target.

Currently, TCP, UDP, and name resolution are mostly implemented (minus a few sockopt options, e.g. SO_RCVTIMEO).

Note that I've changed SYSROOT_INC from $(SYSROOT)/include to $(SYSROOT)/include/$(TARGET_TRIPLE) since the capabilities of Preview 2 are much wider than Preview 1, and the header files reflect that. Also, wasm32-wasi and wasm32-wasi-threads sharing the same include directory was already awkward (e.g. the Makefile was ambivalent about whether pthread.h should be present or not).

See https://github.com/dicej/wasi-sockets-tests for examples of using wasi-sockets in Rust and Python via their respective standard libraries, both of which are built on top of wasi-libc.

Big thanks to @badeend for doing most of the work here!

Signed-off-by: Joel Dice <[email protected]>

…-sockets

…re, now that sockets are fully based on preview2

Co-authored-by: Joel Dice <[email protected]>

Tcp sockets

Signed-off-by: Joel Dice <[email protected]>

…tion of wit-bindgen to generate shorter names. The full command used to generate the bindings was: wit-bindgen c --world imports \ --rename wasi:clocks/[email protected]=monotonic_clock \ --rename wasi:clocks/[email protected]=wall_clock \ --rename wasi:filesystem/[email protected]=filesystem_preopens \ --rename wasi:filesystem/[email protected]=filesystem \ --rename wasi:io/[email protected]=io_error \ --rename wasi:io/[email protected]=poll \ --rename wasi:io/[email protected]=streams \ --rename wasi:random/[email protected]=random_insecure_seed \ --rename wasi:random/[email protected]=random_insecure \ --rename wasi:random/[email protected]=random \ --rename wasi:sockets/[email protected]=instance_network \ --rename wasi:sockets/[email protected]=ip_name_lookup \ --rename wasi:sockets/[email protected]=network \ --rename wasi:sockets/[email protected]=tcp_create_socket \ --rename wasi:sockets/[email protected]=tcp \ --rename wasi:sockets/[email protected]=udp_create_socket \ --rename wasi:sockets/[email protected]=udp \ --rename wasi:cli/[email protected]=environment \ --rename wasi:cli/[email protected]=exit \ --rename wasi:cli/[email protected]=stdin \ --rename wasi:cli/[email protected]=stdout \ --rename wasi:cli/[email protected]=stderr \ ./wit

…ate-preview2-version

The tag and union always need to be updated in tandem. Having them close together helps to clarify the relationship between the two types.

…te often for input validation.

- On input addresses: check against the expected address family and return EAFNOSUPPORT when mismatch - On output addresses: perform validations _before_ making the actual syscall.

Signed-off-by: Joel Dice <[email protected]>

dicej · 2023-12-18T23:21:25Z

Is there some way we can make the including of pthread.h dependent on the current target perhaps?

It looks like libc++ has _LIBCPP_HAS_THREAD_API_PTHREAD .. can that depend on the target triple?

Yeah, that was my initial thought when starting this project, but when I ran e.g. clang --target=wasm32-wasi-preview2 -dM -E -x c /dev/null I couldn't find any preprocessor definitions that told me which triple I was building for, so I abandoned that approach. Did I miss something? Or are you thinking we should add something to clang?

sbc100 · 2023-12-18T23:38:41Z

Is there some way we can make the including of pthread.h dependent on the current target perhaps?
It looks like libc++ has _LIBCPP_HAS_THREAD_API_PTHREAD .. can that depend on the target triple?

Yeah, that was my initial thought when starting this project, but when I ran e.g. clang --target=wasm32-wasi-preview2 -dM -E -x c /dev/null I couldn't find any preprocessor definitions that told me which triple I was building for, so I abandoned that approach. Did I miss something? Or are you thinking we should add something to clang?

I though we were talking about pthread.h? Does clang --target=wasm32-wasi-threads also not have a way to detect it?

dicej · 2023-12-18T23:49:49Z

I though we were talking about pthread.h? Does clang --target=wasm32-wasi-threads also not have a way to detect it?

Same story with clang --target=wasm32-wasi-threads -dM -E -x c /dev/null. Again, I could be missing something.

And, yes, we're talking about pthread.h, but I'm also thinking of headers like netdb.h which the Preview 1 targets should not have but the Preview 2 target(s) should. Not that libc++ uses netdb.h, but I'm thinking of what would be required to switch back to a completely shared include dir.

Signed-off-by: Joel Dice <[email protected]>

Also, stub out other netdb functions so they return errors instead of aborting. Signed-off-by: Joel Dice <[email protected]>

Signed-off-by: Joel Dice <[email protected]>

Currently, this is identical to the `wasm32-wasi` in all but name. See WebAssembly#449 for the next step, which is to incrementally add Preview 2 features, e.g. `wasi-sockets`. Per the discussion in that PR, I've split the `wasi-sysroot/include` directory into per-target directories. Eventually, we'll want to build a separate sysroot for each target, but there's currently uncertainty about how to configure the default sysroot for e.g. clang, so we're not tackling that yet. See also WebAssembly#447 for further details. Signed-off-by: Joel Dice <[email protected]>

Currently, this is identical to the `wasm32-wasi` in all but name. See #449 for the next step, which is to incrementally add Preview 2 features, e.g. `wasi-sockets`. Per the discussion in that PR, I've split the `wasi-sysroot/include` directory into per-target directories. Eventually, we'll want to build a separate sysroot for each target, but there's currently uncertainty about how to configure the default sysroot for e.g. clang, so we're not tackling that yet. See also #447 for further details. Signed-off-by: Joel Dice <[email protected]>

yamt · 2024-01-23T08:26:59Z

do you have any idea how these stuff would work when we add multi thread support later?
adapting code with bold single thread assumptions is a heavy task. it's better to do it "properly" from the start.

Signed-off-by: Joel Dice <[email protected]>

sunfishcode · 2024-01-26T22:36:06Z

.gitignore

@@ -1,2 +1,3 @@
 sysroot
 build
+.vscode


Please put editor-specific paths in your own global git ignore file, rather than in repository .gitignore files.

sunfishcode · 2024-01-26T22:41:59Z

libc-bottom-half/cloudlibc/src/libc/sys/socket/accept.c

@@ -0,0 +1,173 @@
+// SPDX-License-Identifier: BSD-2-Clause


Are files like this one derived from cloudlibc, or are they new? We should ideally put new code somewhere outside of the cloudlibc directory, to keep it clear when we're using and modifying upstream code which may have upstream assumptions, and when we're writing our own code.

sunfishcode · 2024-01-26T22:49:22Z

libc-bottom-half/headers/private/descriptor_table.h

+typedef struct {} tcp_socket_state_unbound_t;
+typedef struct {} tcp_socket_state_bound_t;
+typedef struct {} tcp_socket_state_connecting_t;
+typedef struct {} tcp_socket_state_listening_t;


In this context, at this time: if clang is ok with it, it's fine.

sunfishcode · 2024-01-26T22:51:43Z

libc-bottom-half/headers/private/preview2.h

@@ -0,0 +1,2481 @@
+// Generated by `wit-bindgen` 0.16.0. DO NOT EDIT!


Unfortunately, #include <wasi/api.h> is public interface for preview1; we'd break users if we renamed it.

badeend · 2024-01-29T19:14:43Z

do you have any idea how these stuff would work when we add multi thread support later? adapting code with bold single thread assumptions is a heavy task. it's better to do it "properly" from the start.

The component model does not support multi-threading across component boundaries, and probably won't be doing so for a while. I agree it wouldn't hurt to have at least a rough idea before leaning too much into it, but I don't think it's worth blocking this PR on.

As for such a rough idea: add a mutex per file descriptor and perform everything behind that lock. For sockets specifically; state changes happen only during socket initialization. They spend the bulk of the time in their end states (TCP_SOCKET_STATE_CONNECTED or TCP_SOCKET_STATE_LISTENING). So there might be optimizations available where we don't need to lock at all in those states.

ghost · 2024-02-02T03:15:26Z

The Alpine Linux wasi-sdk package literally just reexports a few packages and adds a .cfg file that Clang uses in order for --target=... to work properly (without the need to specify --sysroot=... explicitly). Could a similar approach be used here too? I.e. having the Clang in wasi-sdk tarball read this a .cfg relative to the extracted directory. (It seems to use the target triple to resolve this .cfg file, so it should be possible.)

More concretely, I think having something like this should work:

wasi-sdk/etc/clangXX/wasm32-unknown-wasi.cfg
wasi-sdk/etc/clangXX/wasm32-unknown-wasi-preview2.cfg

I guess in Alpine, they can specify the an absolute --sysroot=... file name, which would not be possible in this case. 🤔 Hmm.

ghost · 2024-02-02T03:18:22Z

make-bindings.sh

+    --rename wasi:cli/[email protected]=terminal_stderr \
+    ./wit
+mv imports.h libc-bottom-half/headers/private/preview2.h
+sed 's/#include "imports.h"/#include "preview2.h"/' < imports.c > libc-bottom-half/cloudlibc/src/libc/sys/wasi_preview2/preview2.c


A very small point: The . in the regex for sed matches any character (rather than a single dot). It might be sensible to escape it:

- sed 's/#include "imports.h"/#include "preview2.h"/' < ... + sed 's/#include "imports\.h"/#include "preview2.h"/' < ...

ghost · 2024-02-02T04:18:55Z

test/Makefile

-	$(CC) $(CFLAGS) $(LDFLAGS) -o $@ $^ 
+	$(CC) $(CFLAGS) $(LDFLAGS) -o $@ $^
+ifeq ($(WASI_SNAPSHOT), preview2)
+	$(WASM_TOOLS) component new --adapt $(ADAPTER) $@ -o $@


This is not very good practice. If wasm-tools fails here (for whatever reason), the Wasm file will not be changed, but it will be considered up to date for make, so this rule will not run on the next invocation to make. Ideally, you’d have something like a temporary intermediate file, so that the input and output to wasm-tools is a different file.

Something like this:

$(CC) $(CFLAGS) $(LDFLAGS) -o tmp.wasm $^ $(WASM_TOOLS) component new --adapt $(ADAPTER) tmp.wasm -o $@ $(RM) tmp.wasm

dicej · 2024-02-12T17:01:22Z

Thanks to everyone who has provided feedback on this PR.

FYI, I am planning to address that feedback, but not necessarily in this PR, since @sbc100 has requested that I break it up into smaller PRs, which I agree is a good idea, and I've been addressing any relevant feedback in each of those PRs. The quicker we can get those reviewed, the quicker we'll get everything merged.

Speaking of which: this one could use some attention: #464

I'm planning to open another small one today.

dicej · 2024-04-03T14:53:01Z

Everything that was in this PR has now been merged into main via a series of smaller PRs, so we can close this.

dicej and others added 30 commits November 27, 2023 10:24

first draft of socket support

6da9ef8

Signed-off-by: Joel Dice <[email protected]>

support non-blocking sockets

95454e8

Signed-off-by: Joel Dice <[email protected]>

close Preview 2 handles in close(2)

e961d20

Signed-off-by: Joel Dice <[email protected]>

add getaddrinfo and setsockopt, etc.

9d98fe3

Signed-off-by: Joel Dice <[email protected]>

Implement TCP based on preview2 wasi-sockets

1937e94

Ignore .vscode

c7eda81

Merge branch 'sockets' of https://github.com/dicej/wasi-libc into tcp…

a14b30d

…-sockets

Fix errors after merge

f3a1106

Perform implicit bind in Listen

ae41e77

Implement MSG_DONTWAIT flag

a00da21

"Implement" MSG_NOSIGNAL

b120ad4

No need for these constant to match their preview1 counterparts anymo…

ee04f9e

…re, now that sockets are fully based on preview2

Make my own TODOs easier to find

ea2fcf0

Decode IP addresses

6cf2c3b

Replace assert with abort

76160c3

Update libc-bottom-half/cloudlibc/src/libc/sys/socket/recv.c

6bb179a

Co-authored-by: Joel Dice <[email protected]>

Merge pull request #1 from badeend/tcp-sockets

5af19d5

Tcp sockets

add TCP_NODELAY support

74b703e

Signed-off-by: Joel Dice <[email protected]>

Merge branch 'sockets' of https://github.com/dicej/wasi-libc into upd…

7112e67

…ate-preview2-version

TCP_NODELAY was removed in rc-2023-11-10

879d533

Update to rc-2023-12-05

9ab49f3

Implement remaining socket options introduced in rc-2023-11-10

eb3602d

Represent tagged unions as a single type instead of three.

91b0a92

The tag and union always need to be updated in tandem. Having them close together helps to clarify the relationship between the two types.

Open & close UDP sockets

1e97c2e

UDP socket options

e0f4314

Keep track of the address family ourselves, because we'll need it qui…

c11533a

…te often for input validation.

Refactor sockaddr validation:

4fc628c

- On input addresses: check against the expected address family and return EAFNOSUPPORT when mismatch - On output addresses: perform validations _before_ making the actual syscall.

add basic TCP_NODELAY support to setsockopt

ebb3e3f

Signed-off-by: Joel Dice <[email protected]>

add netdb.h-related stubs

253f895

Signed-off-by: Joel Dice <[email protected]>

dicej added 9 commits December 19, 2023 13:41

support TCP_SOCKET_STATE_LISTENING sockets in poll

0b50784

Signed-off-by: Joel Dice <[email protected]>

return getaddrinfo results in order

1db6e1a

Also, stub out other netdb functions so they return errors instead of aborting. Signed-off-by: Joel Dice <[email protected]>

honor ai_family hint in getaddrinfo

5b1ae5e

Signed-off-by: Joel Dice <[email protected]>

add UDP support to ioctl

9c4d23a

Signed-off-by: Joel Dice <[email protected]>

add fake SO_REUSEADDR support

7c28169

Signed-off-by: Joel Dice <[email protected]>

add UDP support to poll

69842d6

Signed-off-by: Joel Dice <[email protected]>

re-implement pselect using poll

c7eeb21

Signed-off-by: Joel Dice <[email protected]>

convert nanoseconds to milliseconds in pselect

32f3a06

Signed-off-by: Joel Dice <[email protected]>

return zero from recv on STREAMS_STREAM_ERROR_CLOSED

f3adffe

Signed-off-by: Joel Dice <[email protected]>

dicej mentioned this pull request Dec 20, 2023

add wasm32-wasi-preview2 target #457

Merged

dicej mentioned this pull request Jan 17, 2024

wit-component: IndexMap: key not found bytecodealliance/wasm-tools#1375

Closed

dicej added 2 commits January 26, 2024 07:53

update bindings to WASI 0.2.0

a3fb845

Signed-off-by: Joel Dice <[email protected]>

update tests to use Wasmtime 17

60e3428

Signed-off-by: Joel Dice <[email protected]>

sunfishcode reviewed Jan 26, 2024

View reviewed changes

dicej mentioned this pull request Feb 1, 2024

meta: Stabilize WASI support tokio-rs/tokio#4827

Open

6 tasks

ghost reviewed Feb 2, 2024

View reviewed changes

dicej mentioned this pull request Feb 26, 2024

[wasi] Support wasi-sockets and wasi-http dotnet/runtime#98957

Closed

dicej closed this Apr 3, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Draft of WASI Preview 2 support, featuring `wasi-sockets` #449

Draft of WASI Preview 2 support, featuring `wasi-sockets` #449

dicej commented Dec 12, 2023 •

edited

Loading

dicej commented Dec 18, 2023

sbc100 commented Dec 18, 2023

dicej commented Dec 18, 2023

yamt commented Jan 23, 2024

sunfishcode Jan 26, 2024

sunfishcode Jan 26, 2024

sunfishcode Jan 26, 2024

sunfishcode Jan 26, 2024

badeend commented Jan 29, 2024

ghost commented Feb 2, 2024

ghost Feb 2, 2024

ghost Feb 2, 2024

dicej commented Feb 12, 2024

dicej commented Apr 3, 2024

@@ @@ -1,2 +1,3 @@ @@
               sysroot
               build
+              .vscode

		@@ -0,0 +1,2481 @@
		// Generated by `wit-bindgen` 0.16.0. DO NOT EDIT!

Draft of WASI Preview 2 support, featuring wasi-sockets #449

Draft of WASI Preview 2 support, featuring wasi-sockets #449

Conversation

dicej commented Dec 12, 2023 • edited Loading

dicej commented Dec 18, 2023

sbc100 commented Dec 18, 2023

dicej commented Dec 18, 2023

yamt commented Jan 23, 2024

sunfishcode Jan 26, 2024

Choose a reason for hiding this comment

sunfishcode Jan 26, 2024

Choose a reason for hiding this comment

sunfishcode Jan 26, 2024

Choose a reason for hiding this comment

sunfishcode Jan 26, 2024

Choose a reason for hiding this comment

badeend commented Jan 29, 2024

ghost commented Feb 2, 2024

ghost Feb 2, 2024

Choose a reason for hiding this comment

ghost Feb 2, 2024

Choose a reason for hiding this comment

dicej commented Feb 12, 2024

dicej commented Apr 3, 2024

Draft of WASI Preview 2 support, featuring `wasi-sockets` #449

Draft of WASI Preview 2 support, featuring `wasi-sockets` #449

dicej commented Dec 12, 2023 •

edited

Loading