Re: [RFC PATCH 0/2] Add a futex interface to glibc (sys/futex.h)

On 25/08/26 20:58, Carlos O'Donell wrote:
> On 8/25/26 12:37 PM, Adhemerval Zanella wrote:
>> This patchset adds wrappers for the futex system call through a new
>> <sys/futex.h> header.  They are futex_wait, futex_timedwait,
>> futex_wake, futex_requeue, and futex_waitv.
> 
> I worked through a prototype for this last year after my LPC talk about
> creating wrappers for all syscalls.
> 
> I think this is the right direction to go and I can review the wrappers.
> 
> I think we should expose as much as is not deprecated.
> 
>> Exporting a futex interface from libc is a long-standing request, as
>> the futex syscall is the blocking building block for essentially
>> every custom synchronization primitive on Linux.  And since glibc
>> provides no wrapper, every project ends up hand-rolling it; with
>> per-architecture syscall assembly, ad-hoc time64 syscall selection,
>> and workarounds for the syscall's historical quirks duplicated in
>> each of them (libstdc++, libc++, libgomp, libitm, the Go runtime,
>> OpenMP, allocators, and so on).
> 
> Right.
> 
>> It has also become relevant as the kernel expands the futex interface,
>> where futex_waitv and the futex2 syscall family only make sense to
>> expose from libc if there is a consistent futex API to build on.
> 
> Agreed.
> 
>> This patchset provides that base operations plus futex_waitv, sharing
>> one set of conventions (flags following the kernel FUTEX_PRIVATE_FLAG,
>> absolute timeouts with an explicit clockid, -1/errno error reporting,
>> no cancellation points), so future futex2 additions are incremental
>> rather than a new design.
>>
>> Only the well-defined subset of the kernel interface is exposed (the
>> racy unchecked FUTEX_REQUEUE, the removed FUTEX_FD, FUTEX_WAKE_OP,
>> bitset matching, and the PI operations are left out).  The
>> futex_timedwait supports both time_t sizes with the usual TIME_BITS=64
>> redirection; futex_waitv mirrors the kernel ABI directly (struct layout,
>> argument order, FUTEX2* per-waiter flags) and is 64-bit time_t only.
>> The wrappers compile down to the same code as a hand-rolled syscall
>> (the waits and requeue are a flags check plus a tail call; wake
>> inlines the syscall), so there is no cost to adopting them.
>>
>> To verify the interface's usability, I built a proof of concept that
>> converts the two futex users in GCC (libitm and libgomp) to the new
>> functions.  The required work is minimal: libitm's wrappers become
>> three-line calls, and libgomp only needs an asm-label binding because
>> its internal wrappers reuse the  utex_wait/futex_wake names.
>>
>> In both cases the per-architecture syscall assembly and the obsolete
>> private-to-shared runtime fallbacks become unnecessary.  A survey of
>> the remaining futex users in GCC and LLVM (libstdc++, libc++, libgo,
>> OpenMP, scudo) showed that all of their usage should map onto this
>> interface as well.
> 
> Great! I'll wait for a full non-RFC to review.
There you go [1]. I am still not fully sure about providing futex_timedwait
for 32-bit time_t, although it *might* simplify adoption in some cases [2].

[1] https://patchwork.sourceware.org/project/glibc/list/?series=65564
[2] https://sourceware.org/pipermail/libc-alpha/2026-August/180012.html