Land pthread_attr_* libc

[jserv]

Close the last narrow gap in the PSE51-oriented userspace surface. A portable libc that promises pthread_attr_t needs both the attribute accessors and a way to honor PTHREAD_EXPLICIT_SCHED without the SYS_THREAD_SETSCHEDPARAM race window; ship both in one branch.

Library lives at include/mazu/pthread.h as a header-only set of static inline functions: pthread_attr_init / _destroy, _set/_get for detachstate, inheritsched, schedpolicy, schedparam, stacksize, stack. All entry points NULL-guard, validate inputs, and return positive errno (POSIX convention). pthread_attr_destroy stamps every field so use-after-destroy is debugger-visible.

Stack accessors are constrained by the shared address-space model: the kernel assigns each thread slot a fixed VA band, so a caller-supplied stack cannot be honored. pthread_attr_setstack always returns ENOTSUP, and pthread_attr_setstacksize succeeds only for USER_STACK_SIZE; both signal the constraint up front instead of letting a portable caller observe a stack contract the kernel will not satisfy. pthread_attr_getstack / _getstacksize still round-trip the init-time defaults so feature probes work.

Kernel ABI adds a dedicated syscall, SYS_THREAD_CREATE_EXPLICIT (slot 101), rather than repurposing SYS_THREAD_CREATE's a2 register. Reusing the original syscall risked misinterpreting garbage a2 from pre-existing callers who never cleared unused registers. The new syscall reads a (prio + 1) encoding from a2 (0 inherits; 1..CONFIG_SCHED_NPRIO sets explicit prio = a2 - 1; anything else returns EINVAL), enforces the same privilege bound as pthread_setschedparam (cannot raise above creator's base priority, EPERM otherwise), and shares its body with sys_thread_create_h via sys_thread_create_common(use_explicit_prio_abi). td_base_prio is snapshotted with __atomic_load_n(__ATOMIC_RELAXED) so a cross-hart setschedparam against the creator cannot make the EPERM gate disagree with the inherit-default fallback.

pthread_attr_resolve_create_syscall and _resolve_prio_arg let a future pthread_create wrapper hand the attr straight to the right syscall. The prio resolver bounds-checks before the (u64) cast: a negative i32 cast to u64 plus 1 wraps to 0, which would have mimicked the inherit encoding and silently demoted an EXPLICIT_SCHED request from a caller who bypassed the setters. The resolver returns a sentinel above CONFIG_SCHED_NPRIO for any out-of-range value so the kernel's own bound check surfaces EINVAL.

ENOTSUP joins the rest of the errno constants in
include/mazu/errordef.h (value 95, matching the Linux EOPNOTSUPP numeric) with a matching error_code_str branch.

---

Summary by cubic

Implements a header-only pthread_attr_* library and adds SYS_THREAD_CREATE_EXPLICIT to support PTHREAD_EXPLICIT_SCHED without the post-create race. This closes the last gap in the PSE51 userspace surface.

• New Features

  • Header-only pthread_attr_* at include/mazu/pthread.h (init/destroy; get/set for detachstate, inheritsched, schedpolicy, schedparam, stacksize, stack). All calls NULL-guarded and return positive errno.
  • Stack semantics: pthread_attr_setstack returns ENOTSUP; pthread_attr_setstacksize accepts only USER_STACK_SIZE. Getters round-trip defaults.
  • New SYS_THREAD_CREATE_EXPLICIT (slot 101): a2=0 inherits, 1..CONFIG_SCHED_NPRIO sets prio (a2−1), else EINVAL; EPERM if above creator’s base. Legacy SYS_THREAD_CREATE unchanged.
  • Adds ENOTSUP (95) and error string; updates docs (PSE51 matrix) and selftests for attrs and the new syscall.

• Migration

  • For a pthread_create wrapper, use pthread_attr_resolve_create_syscall and pthread_attr_resolve_prio_arg to pick the syscall and a2 encoding.
  • Do not rely on caller-supplied stacks; accept kernel-assigned stacks and handle pthread_attr_setstack as unsupported.

^{Written for commit 6b5ed23. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 7 files