Add support for FUTEX_{WAIT,WAKE}_BITSET #2054
Conversation
| assert_eq!(libc::syscall( | ||
| libc::SYS_futex, | ||
| &futex as *const i32, | ||
| libc::FUTEX_WAIT_BITSET, |
There was a problem hiding this comment.
This test also passes if I put FUTEX_WAIT here, as it seems that the monotonic clock starts at zero in the tests. It'd be a better test if it was run with a monotonic clock that does not start at zero.
There was a problem hiding this comment.
Specifically, our monotone clock is relative to the time_anchor which is initialized via Instant::now() when the interpreter starts.
I guess we could add an arbitrary offset to that? It also should make a difference if the interpreter has already been running for a bit (e.g. because of a previous timeout test).
It was pretty easy to add full support for these syscalls. I've updated the PR. |
m-ou-se
changed the title
Lol, I think I know that feeling. ;) Thanks a ton! |
|
Could you put |
|
I like green CI. :) |
|
📌 Commit 4fdda31 has been approved by |
|
☀️ Test successful - checks-actions |
| libc::FUTEX_WAIT_BITSET, | ||
| 123, | ||
| &timeout, | ||
| 0, |
There was a problem hiding this comment.
This turns out to be a subtle bug, since 0 becomes 0i32 but this argument should be pointer-sized. #2058 makes Miri detect that problem.
(It'd probably still work on x86 due to details of the calling convention, but e.g. if arguments are passed on the stack then va_arg would probably do the wrong thing here.)
FUTEX_WAIT_BITSET and FUTEX_WAKE_BITSET are extensions of FUTEX_WAIT and FUTEX_WAKE that allow tagging each waiting thread with up to 32 'labels', and then only wake up threads that match certain labels. The non-bitset operations behave like their bitset was fully set (u32::MAX), meaning that they'll wait for anything, and wake up anything.
The only other difference is that FUTEX_WAIT_BITSET uses an absolute timeout instead of an relative timeout like FUTEX_WAIT.
Often, FUTEX_WAIT_BITSET is used not for its bitset functionality, but only for its absolute timeout functionality. It is then used with a bitset of u32::MAX.
This adds support for only that use case to Miri, as that's allstdcurrently needs. Any other bitset is still unsupported.Update: This adds full support for both these syscalls.