std.Io.Threaded: fix incorrect alignment of trailing data in closure

[Techatrix]

When calling async, concurrent or std.Io.Group.async on std.Io.Threaded with a function that has parameters or a return value with extra alignment, it may fail because the internal async closure did not properly create well aligned storage for them.

conststd=@import("std");
pubfnmain() !void {
// Depending on which allocator is used, the code may work "by accident". // Using a fixed buffer allocator did consistently reproduce the issue for mevarbuffer: [1024]u8align(64) =undefined;
varfba: std.heap.FixedBufferAllocator= .init(buffer[1..]);
varthreaded: std.Io.Threaded= .init(fba.allocator());
deferthreaded.deinit();
constio=threaded.io();
varfuture=io.async(paramWithExtraAlignment, .{.{}});
future.await(io);
}
fnparamWithExtraAlignment(unused: Align64) void {
_=unused;
}
constAlign64=struct {
unused: u8align(64) =0,
};

zig run sample.zig thread 5824 panic: incorrect alignment
/home/techatrix/repos/zig/lib/std/Io.zig:1534:46: 0x1163409 in start (std.zig)
const args_casted: *const Args = @ptrCast(@alignCast(context));
^
/home/techatrix/repos/zig/lib/std/Io/Threaded.zig:405:16: 0x106af5e in start (std.zig)
ac.func(ac.contextPointer(), ac.resultPointer());
^
/home/techatrix/repos/zig/lib/std/Io/Threaded.zig:188:26: 0x10d9b06 in worker (std.zig)
closure.start(closure);
^
/home/techatrix/repos/zig/lib/std/Thread.zig:558:13: 0x10b3100 in callFn__anon_16055 (std.zig)
@call(.auto, f, args);
^
/home/techatrix/repos/zig/lib/std/Thread.zig:1534:30: 0x108f070 in entryFn (std.zig)
return callFn(f, self.fn_args);
^
/home/techatrix/repos/zig/lib/std/os/linux/x86_64.zig:105:5: 0x10b31b5 in clone (std.zig)
asm volatile (
^
Aborted (core dumped)

[squeek502]

Overlaps with #25817 (cc @castholm)

[castholm]

Aside from the problem I point out, the differences between this PR and #25817 is that this PR is slightly more efficient with the sizes of closure fields (2 * u8 + 2 * usize vs 3 * usize) but allocates a few more wasted bytes when the context alignment is <= the closure alignemnt.

#25817 also deduplicates the allocation code to a common init() function to avoid having the code paths going out of sync.

I could try to adapt the test to my PR if you think it'd be useful.

[castholm]

This looks wrong to me, the code is not aware of the context length. resultPointer() will return a pointer into the middle of the context. AsyncClosure needs to remember either the computed result offset, or the context alignment + context length + result alignment to recompute it.

[Techatrix]

Good catch. I adjusted it to result offset instead of the total allocated memory for the closure.

[Techatrix]

I could try to adapt the test to my PR if you think it'd be useful.

Feel free to reuse the tests in your PR.

[squeek502]

Superseded by #25817