Introduce cosmo_futex_wait and cosmo_futex_wake

Cosmopolitan Futexes are now exposed as a public API.
2025-02-07 06:53:33 +00:00 · 2024-11-22 11:08:29 -08:00 · 2024-11-22 11:08:29 -08:00 · 9ddbfd921e
commit 9ddbfd921e
parent 729f7045e3
66 changed files with 886 additions and 917 deletions
--- a/libc/calls/sig.c
+++ b/libc/calls/sig.c
@ -1,717 +0,0 @@
 /*-*- mode:c;indent-tabs-mode:nil;c-basic-offset:2;tab-width:8;coding:utf-8 -*-│
 │ vi: set et ft=c ts=2 sts=2 sw=2 fenc=utf-8                               :vi │
 ╞══════════════════════════════════════════════════════════════════════════════╡
 │ Copyright 2022 Justine Alexandra Roberts Tunney                              │
 │                                                                              │
 │ Permission to use, copy, modify, and/or distribute this software for         │
 │ any purpose with or without fee is hereby granted, provided that the         │
 │ above copyright notice and this permission notice appear in all copies.      │
 │                                                                              │
 │ THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL                │
 │ WARRANTIES WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED                │
 │ WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE             │
 │ AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL         │
 │ DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR        │
 │ PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER               │
 │ TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR             │
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/sysv/consts/sig.h"
 #include "ape/sections.internal.h"
 #include "libc/calls/calls.h"
 #include "libc/calls/sig.internal.h"
 #include "libc/calls/state.internal.h"
 #include "libc/calls/struct/sigaction.h"
 #include "libc/calls/struct/siginfo.h"
 #include "libc/calls/struct/sigset.internal.h"
 #include "libc/calls/struct/ucontext.internal.h"
 #include "libc/calls/syscall_support-nt.internal.h"
 #include "libc/calls/ucontext.h"
 #include "libc/dce.h"
 #include "libc/errno.h"
 #include "libc/intrin/atomic.h"
 #include "libc/intrin/bsf.h"
 #include "libc/intrin/describebacktrace.h"
 #include "libc/intrin/dll.h"
 #include "libc/intrin/maps.h"
 #include "libc/intrin/strace.h"
 #include "libc/intrin/weaken.h"
 #include "libc/nt/console.h"
 #include "libc/nt/enum/context.h"
 #include "libc/nt/enum/exceptionhandleractions.h"
 #include "libc/nt/enum/processcreationflags.h"
 #include "libc/nt/enum/signal.h"
 #include "libc/nt/enum/status.h"
 #include "libc/nt/events.h"
 #include "libc/nt/runtime.h"
 #include "libc/nt/signals.h"
 #include "libc/nt/struct/ntexceptionpointers.h"
 #include "libc/nt/synchronization.h"
 #include "libc/nt/thread.h"
 #include "libc/runtime/internal.h"
 #include "libc/runtime/symbols.internal.h"
 #include "libc/str/str.h"
 #include "libc/sysv/consts/sa.h"
 #include "libc/sysv/consts/sicode.h"
 #include "libc/sysv/consts/ss.h"
 #include "libc/thread/posixthread.internal.h"
 #ifdef __x86_64__
 /**
 * @fileoverview Cosmopolitan Signals for Windows.
 */
 #define STKSZ 65536
 struct SignalFrame {
  unsigned rva;
  unsigned flags;
  siginfo_t si;
  ucontext_t ctx;
 };
 static textwindows bool __sig_ignored_by_default(int sig) {
  return sig == SIGURG ||   //
         sig == SIGCONT ||  //
         sig == SIGCHLD ||  //
         sig == SIGWINCH;
 }
 textwindows bool __sig_ignored(int sig) {
  return __sighandrvas[sig] == (intptr_t)SIG_IGN ||
         (__sighandrvas[sig] == (intptr_t)SIG_DFL &&
          __sig_ignored_by_default(sig));
 }
 textwindows void __sig_delete(int sig) {
  struct Dll *e;
  atomic_fetch_and_explicit(__sig.process, ~(1ull << (sig - 1)),
                            memory_order_relaxed);
  _pthread_lock();
  for (e = dll_last(_pthread_list); e; e = dll_prev(_pthread_list, e))
    atomic_fetch_and_explicit(&POSIXTHREAD_CONTAINER(e)->tib->tib_sigpending,
                              ~(1ull << (sig - 1)), memory_order_relaxed);
  _pthread_unlock();
 }
 static textwindows int __sig_getter(atomic_ulong *sigs, sigset_t masked) {
  int sig;
  sigset_t bit, pending, deliverable;
  for (;;) {
    pending = atomic_load_explicit(sigs, memory_order_acquire);
    if ((deliverable = pending & ~masked)) {
      sig = bsfl(deliverable) + 1;
      bit = 1ull << (sig - 1);
      if (atomic_fetch_and_explicit(sigs, ~bit, memory_order_acq_rel) & bit)
        return sig;
    } else {
      return 0;
    }
  }
 }
 textwindows int __sig_get(sigset_t masked) {
  int sig;
  if (!(sig = __sig_getter(&__get_tls()->tib_sigpending, masked)))
    sig = __sig_getter(__sig.process, masked);
  return sig;
 }
 static textwindows bool __sig_should_use_altstack(unsigned flags,
                                                  struct CosmoTib *tib) {
  if (!(flags & SA_ONSTACK))
    return false;  // signal handler didn't enable it
  if (!tib->tib_sigstack_size)
    return false;  // sigaltstack() wasn't installed on this thread
  if (tib->tib_sigstack_flags & SS_DISABLE)
    return false;  // sigaltstack() on this thread was disabled by user
  char *bp = __builtin_frame_address(0);
  if (tib->tib_sigstack_addr <= bp &&
      bp <= tib->tib_sigstack_addr + tib->tib_sigstack_size)
    return false;  // we're already on the alternate stack
  return true;
 }
 static textwindows wontreturn void __sig_terminate(int sig) {
  TerminateThisProcess(sig);
 }
 textwindows static bool __sig_wake(struct PosixThread *pt, int sig) {
  atomic_int *blocker;
  blocker = atomic_load_explicit(&pt->pt_blocker, memory_order_acquire);
  if (!blocker)
    return false;
  // threads can create semaphores on an as-needed basis
  if (blocker == PT_BLOCKER_EVENT) {
    STRACE("%G set %d's event object", sig, _pthread_tid(pt));
    SetEvent(pt->pt_event);
    return !!atomic_load_explicit(&pt->pt_blocker, memory_order_acquire);
  }
  // all other blocking ops that aren't overlap should use futexes
  // we force restartable futexes to churn by waking w/o releasing
  STRACE("%G waking %d's futex", sig, _pthread_tid(pt));
  WakeByAddressSingle(blocker);
  return !!atomic_load_explicit(&pt->pt_blocker, memory_order_acquire);
 }
 textwindows static bool __sig_start(struct PosixThread *pt, int sig,
                                    unsigned *rva, unsigned *flags) {
  *rva = __sighandrvas[sig];
  *flags = __sighandflags[sig];
  if (*rva == (intptr_t)SIG_IGN ||
      (*rva == (intptr_t)SIG_DFL && __sig_ignored_by_default(sig))) {
    STRACE("ignoring %G", sig);
    return false;
  }
  if (atomic_load_explicit(&pt->tib->tib_sigmask, memory_order_acquire) &
      (1ull << (sig - 1))) {
    STRACE("enqueing %G on %d", sig, _pthread_tid(pt));
    atomic_fetch_or_explicit(&pt->tib->tib_sigpending, 1ull << (sig - 1),
                             memory_order_relaxed);
    __sig_wake(pt, sig);
    return false;
  }
  if (*rva == (intptr_t)SIG_DFL) {
    STRACE("terminating on %G due to no handler", sig);
    __sig_terminate(sig);
  }
  return true;
 }
 textwindows static sigaction_f __sig_handler(unsigned rva) {
  atomic_fetch_add_explicit(&__sig.count, 1, memory_order_relaxed);
  return (sigaction_f)(__executable_start + rva);
 }
 textwindows int __sig_raise(volatile int sig, int sic) {
  // bitset of kinds of handlers called
  volatile int handler_was_called = 0;
  // loop over pending signals
  ucontext_t ctx;
  getcontext(&ctx);
  if (!sig) {
    if ((sig = __sig_get(ctx.uc_sigmask))) {
      sic = SI_KERNEL;
    } else {
      return handler_was_called;
    }
  }
  // process signal(s)
  unsigned rva, flags;
  struct PosixThread *pt = _pthread_self();
  if (__sig_start(pt, sig, &rva, &flags)) {
    if (flags & SA_RESETHAND) {
      STRACE("resetting %G handler", sig);
      __sighandrvas[sig] = (int32_t)(intptr_t)SIG_DFL;
    }
    // update the signal mask in preparation for signal handller
    sigset_t blocksigs = __sighandmask[sig];
    if (!(flags & SA_NODEFER))
      blocksigs |= 1ull << (sig - 1);
    ctx.uc_sigmask = atomic_fetch_or_explicit(&pt->tib->tib_sigmask, blocksigs,
                                              memory_order_acquire);
    // call the user's signal handler
    char ssbuf[128];
    siginfo_t si = {.si_signo = sig, .si_code = sic};
    STRACE("__sig_raise(%G, %t) mask %s", sig, __sig_handler(rva),
           _DescribeSigset(ssbuf, 0, (sigset_t *)&pt->tib->tib_sigmask));
    __sig_handler(rva)(sig, &si, &ctx);
    // record this handler
    if (flags & SA_RESTART) {
      handler_was_called |= SIG_HANDLED_SA_RESTART;
    } else {
      handler_was_called |= SIG_HANDLED_NO_RESTART;
    }
  }
  // restore sigmask
  // loop back to top
  // jump where handler says
  sig = 0;
  return setcontext(&ctx);
 }
 textwindows int __sig_relay(int sig, int sic, sigset_t waitmask) {
  sigset_t m;
  int handler_was_called;
  m = atomic_exchange_explicit(&__get_tls()->tib_sigmask, waitmask,
                               memory_order_acquire);
  handler_was_called = __sig_raise(sig, SI_KERNEL);
  atomic_store_explicit(&__get_tls()->tib_sigmask, m, memory_order_release);
  return handler_was_called;
 }
 // the user's signal handler callback is wrapped with this trampoline
 static textwindows wontreturn void __sig_tramp(struct SignalFrame *sf) {
  int sig = sf->si.si_signo;
  struct CosmoTib *tib = __get_tls();
  struct PosixThread *pt = (struct PosixThread *)tib->tib_pthread;
  atomic_store_explicit(&pt->pt_intoff, 0, memory_order_release);
  for (;;) {
    // update the signal mask in preparation for signal handler
    sigset_t blocksigs = __sighandmask[sig];
    if (!(sf->flags & SA_NODEFER))
      blocksigs |= 1ull << (sig - 1);
    sf->ctx.uc_sigmask = atomic_fetch_or_explicit(&tib->tib_sigmask, blocksigs,
                                                  memory_order_acquire);
    // call the user's signal handler
    char ssbuf[2][128];
    STRACE("__sig_tramp(%G, %t) mask %s → %s", sig, __sig_handler(sf->rva),
           _DescribeSigset(ssbuf[0], 0, &sf->ctx.uc_sigmask),
           _DescribeSigset(ssbuf[1], 0, (sigset_t *)&tib->tib_sigmask));
    __sig_handler(sf->rva)(sig, &sf->si, &sf->ctx);
    // restore the signal mask that was used by the interrupted code
    // this may have been modified by the signal handler in the callback
    atomic_store_explicit(&tib->tib_sigmask, sf->ctx.uc_sigmask,
                          memory_order_release);
    // jump back into original code if there aren't any pending signals
    do {
      if (!(sig = __sig_get(sf->ctx.uc_sigmask)))
        __sig_restore(&sf->ctx);
    } while (!__sig_start(pt, sig, &sf->rva, &sf->flags));
    // tail recurse into another signal handler
    sf->si.si_signo = sig;
    sf->si.si_code = SI_KERNEL;
    if (sf->flags & SA_RESETHAND) {
      STRACE("resetting %G handler", sig);
      __sighandrvas[sig] = (int32_t)(intptr_t)SIG_DFL;
    }
  }
 }
 // sends signal to another specific thread which is ref'd
 static textwindows int __sig_killer(struct PosixThread *pt, int sig, int sic) {
  unsigned rva = __sighandrvas[sig];
  unsigned flags = __sighandflags[sig];
  // do nothing if signal is ignored
  if (rva == (intptr_t)SIG_IGN ||
      (rva == (intptr_t)SIG_DFL && __sig_ignored_by_default(sig))) {
    STRACE("ignoring %G", sig);
    return 0;
  }
  // we can't preempt threads that masked sigs or are blocked on i/o
  while ((atomic_load_explicit(&pt->tib->tib_sigmask, memory_order_acquire) &
          (1ull << (sig - 1)))) {
    if (atomic_fetch_or_explicit(&pt->tib->tib_sigpending, 1ull << (sig - 1),
                                 memory_order_acq_rel) &
        (1ull << (sig - 1)))
      // we believe signal was already enqueued
      return 0;
    if (__sig_wake(pt, sig))
      // we believe i/o routine will handle signal
      return 0;
    if (atomic_load_explicit(&pt->tib->tib_sigmask, memory_order_acquire) &
        (1ull << (sig - 1)))
      // we believe ALLOW_SIGNALS will handle signal
      return 0;
    if (!(atomic_fetch_and_explicit(&pt->tib->tib_sigpending,
                                    ~(1ull << (sig - 1)),
                                    memory_order_acq_rel) &
          (1ull << (sig - 1))))
      // we believe another thread sniped our signal
      return 0;
    break;
  }
  // avoid race conditions and deadlocks with thread suspend process
  if (atomic_exchange_explicit(&pt->pt_intoff, 1, memory_order_acquire)) {
    // we believe another thread is asynchronously waking the mark
    if (atomic_fetch_or_explicit(&pt->tib->tib_sigpending, 1ull << (sig - 1),
                                 memory_order_acq_rel) &
        (1ull << (sig - 1)))
      // we believe our signal is already being delivered
      return 0;
    if (atomic_load_explicit(&pt->pt_intoff, memory_order_acquire) ||
        atomic_exchange_explicit(&pt->pt_intoff, 1, memory_order_acquire))
      // we believe __sig_tramp will deliver our signal
      return 0;
    if (!(atomic_fetch_and_explicit(&pt->tib->tib_sigpending,
                                    ~(1ull << (sig - 1)),
                                    memory_order_acq_rel) &
          (1ull << (sig - 1))))
      // we believe another thread sniped our signal
      return 0;
  }
  // if there's no handler then killing a thread kills the process
  if (rva == (intptr_t)SIG_DFL) {
    STRACE("terminating on %G due to no handler", sig);
    __sig_terminate(sig);
  }
  // take control of thread
  // suspending the thread happens asynchronously
  // however getting the context blocks until it's frozen
  uintptr_t th = _pthread_syshand(pt);
  if (SuspendThread(th) == -1u) {
    STRACE("SuspendThread failed w/ %d", GetLastError());
    atomic_store_explicit(&pt->pt_intoff, 0, memory_order_release);
    return ESRCH;
  }
  struct NtContext nc;
  nc.ContextFlags = kNtContextFull;
  if (!GetThreadContext(th, &nc)) {
    STRACE("GetThreadContext failed w/ %d", GetLastError());
    ResumeThread(th);
    atomic_store_explicit(&pt->pt_intoff, 0, memory_order_release);
    return ESRCH;
  }
  // we can't preempt threads that masked sig or are blocked
  // we can't preempt threads that are running in win32 code
  // so we shall unblock the thread and let it signal itself
  if (!((uintptr_t)__executable_start <= nc.Rip &&
        nc.Rip < (uintptr_t)__privileged_start)) {
    atomic_fetch_or_explicit(&pt->tib->tib_sigpending, 1ull << (sig - 1),
                             memory_order_relaxed);
    ResumeThread(th);
    atomic_store_explicit(&pt->pt_intoff, 0, memory_order_release);
    __sig_wake(pt, sig);
    return 0;
  }
  // preferring to live dangerously
  // the thread will be signaled asynchronously
  if (flags & SA_RESETHAND) {
    STRACE("resetting %G handler", sig);
    __sighandrvas[sig] = (int32_t)(intptr_t)SIG_DFL;
  }
  // inject call to trampoline function into thread
  uintptr_t sp;
  if (__sig_should_use_altstack(flags, pt->tib)) {
    sp = (uintptr_t)pt->tib->tib_sigstack_addr + pt->tib->tib_sigstack_size;
  } else {
    sp = nc.Rsp;
  }
  sp -= sizeof(struct SignalFrame);
  sp &= -16;
  struct SignalFrame *sf = (struct SignalFrame *)sp;
  _ntcontext2linux(&sf->ctx, &nc);
  bzero(&sf->si, sizeof(sf->si));
  sf->rva = rva;
  sf->flags = flags;
  sf->si.si_code = sic;
  sf->si.si_signo = sig;
  *(uintptr_t *)(sp -= sizeof(uintptr_t)) = nc.Rip;
  nc.Rip = (intptr_t)__sig_tramp;
  nc.Rdi = (intptr_t)sf;
  nc.Rsp = sp;
  if (!SetThreadContext(th, &nc)) {
    STRACE("SetThreadContext failed w/ %d", GetLastError());
    atomic_store_explicit(&pt->pt_intoff, 0, memory_order_release);
    return ESRCH;
  }
  ResumeThread(th);
  __sig_wake(pt, sig);
  return 0;
 }
 // sends signal to another specific thread
 textwindows int __sig_kill(struct PosixThread *pt, int sig, int sic) {
  int rc;
  BLOCK_SIGNALS;
  rc = __sig_killer(pt, sig, sic);
  ALLOW_SIGNALS;
  return rc;
 }
 // sends signal to any other thread
 // this should only be called by non-posix threads
 textwindows void __sig_generate(int sig, int sic) {
  struct Dll *e;
  struct PosixThread *pt, *mark = 0;
  if (__sig_ignored(sig)) {
    STRACE("ignoring %G", sig);
    return;
  }
  if (__sighandrvas[sig] == (intptr_t)SIG_DFL) {
    STRACE("terminating on %G due to no handler", sig);
    __sig_terminate(sig);
  }
  if (atomic_load_explicit(__sig.process, memory_order_acquire) &
      (1ull << (sig - 1))) {
    return;
  }
  _pthread_lock();
  for (e = dll_first(_pthread_list); e; e = dll_next(_pthread_list, e)) {
    pt = POSIXTHREAD_CONTAINER(e);
    // we don't want to signal ourself
    if (pt == _pthread_self())
      continue;
    // we don't want to signal a thread that isn't running
    if (atomic_load_explicit(&pt->pt_status, memory_order_acquire) >=
        kPosixThreadTerminated) {
      continue;
    }
    // choose this thread if it isn't masking sig
    if (!(atomic_load_explicit(&pt->tib->tib_sigmask, memory_order_acquire) &
          (1ull << (sig - 1)))) {
      _pthread_ref(pt);
      mark = pt;
      break;
    }
    // if a thread is blocking then we check to see if it's planning
    // to unblock our sig once the wait operation is completed; when
    // that's the case we can cancel the thread's i/o to deliver sig
    if (atomic_load_explicit(&pt->pt_blocker, memory_order_acquire) &&
        !(pt->pt_blkmask & (1ull << (sig - 1)))) {
      _pthread_ref(pt);
      mark = pt;
      break;
    }
  }
  _pthread_unlock();
  if (mark) {
    // no lock needed since current thread is nameless and formless
    __sig_killer(mark, sig, sic);
    _pthread_unref(mark);
  } else {
    atomic_fetch_or_explicit(__sig.process, 1ull << (sig - 1),
                             memory_order_relaxed);
  }
 }
 static textwindows char *__sig_stpcpy(char *d, const char *s) {
  size_t i;
  for (i = 0;; ++i)
    if (!(d[i] = s[i]))
      return d + i;
 }
 static textwindows wontreturn void __sig_death(int sig, const char *thing) {
 #ifndef TINY
  intptr_t hStderr;
  char sigbuf[21], s[128], *p;
  hStderr = GetStdHandle(kNtStdErrorHandle);
  p = __sig_stpcpy(s, "Terminating on ");
  p = __sig_stpcpy(p, thing);
  p = __sig_stpcpy(p, strsignal_r(sig, sigbuf));
  p = __sig_stpcpy(p,
                   ". Pass --strace and/or ShowCrashReports() for details.\n");
  WriteFile(hStderr, s, p - s, 0, 0);
 #endif
  __sig_terminate(sig);
 }
 static textwindows void __sig_unmaskable(struct NtExceptionPointers *ep,
                                         int code, int sig,
                                         struct CosmoTib *tib) {
  // log vital crash information reliably for --strace before doing much
  // we don't print this without the flag since raw numbers scare people
  // this needs at least one page of stack memory in order to get logged
  // otherwise it'll print a warning message about the lack of stack mem
  STRACE("win32 vectored exception 0x%08Xu raising %G "
         "cosmoaddr2line %s %lx %s",
         ep->ExceptionRecord->ExceptionCode, sig,
         _weaken(FindDebugBinary) ? _weaken(FindDebugBinary)()
                                  : program_invocation_name,
         ep->ContextRecord->Rip,
         DescribeBacktrace((struct StackFrame *)ep->ContextRecord->Rbp));
  // if the user didn't install a signal handler for this unmaskable
  // exception, then print a friendly helpful hint message to stderr
  unsigned rva = __sighandrvas[sig];
  if (rva == (intptr_t)SIG_DFL || rva == (intptr_t)SIG_IGN)
    __sig_death(sig, "uncaught ");
  // if this signal handler is configured to auto-reset to the default
  // then that reset needs to happen before the user handler is called
  unsigned flags = __sighandflags[sig];
  if (flags & SA_RESETHAND) {
    STRACE("resetting %G handler", sig);
    __sighandrvas[sig] = (int32_t)(intptr_t)SIG_DFL;
  }
  // determine the true memory address at which fault occurred
  // if this is a stack overflow then reapply guard protection
  void *si_addr;
  if (ep->ExceptionRecord->ExceptionCode == kNtSignalGuardPage) {
    si_addr = (void *)ep->ExceptionRecord->ExceptionInformation[1];
  } else {
    si_addr = ep->ExceptionRecord->ExceptionAddress;
  }
  // call the user signal handler
  // and a modifiable view of the faulting code's cpu state
  // temporarily replace signal mask while calling crash handler
  // abort process if sig is already blocked to avoid crash loop
  // note ucontext_t is a hefty data structures on top of NtContext
  ucontext_t ctx = {0};
  siginfo_t si = {.si_signo = sig, .si_code = code, .si_addr = si_addr};
  _ntcontext2linux(&ctx, ep->ContextRecord);
  sigset_t blocksigs = __sighandmask[sig];
  if (!(flags & SA_NODEFER))
    blocksigs |= 1ull << (sig - 1);
  ctx.uc_sigmask = atomic_fetch_or_explicit(&tib->tib_sigmask, blocksigs,
                                            memory_order_acquire);
  if (ctx.uc_sigmask & (1ull << (sig - 1))) {
    __sig_death(sig, "masked ");
    __sig_terminate(sig);
  }
  __sig_handler(rva)(sig, &si, &ctx);
  atomic_store_explicit(&tib->tib_sigmask, ctx.uc_sigmask,
                        memory_order_release);
  _ntlinux2context(ep->ContextRecord, &ctx);
 }
 void __stack_call(struct NtExceptionPointers *, int, int, struct CosmoTib *,
                  void (*)(struct NtExceptionPointers *, int, int,
                           struct CosmoTib *),
                  void *);
 //                         abashed the devil stood
 //                      and felt how awful goodness is
 __msabi dontinstrument unsigned __sig_crash(struct NtExceptionPointers *ep) {
  // translate win32 to unix si_signo and si_code
  int code, sig = __sig_crash_sig(ep->ExceptionRecord->ExceptionCode, &code);
  // advance the instruction pointer to skip over debugger breakpoints
  // this behavior is consistent with how unix kernels are implemented
  if (sig == SIGTRAP) {
    ep->ContextRecord->Rip++;
    if (__sig_ignored(sig))
      return kNtExceptionContinueExecution;
  }
  // win32 stack overflow detection executes INSIDE the guard page
  // thus switch to the alternate signal stack as soon as possible
  struct CosmoTib *tib = __get_tls();
  unsigned flags = __sighandflags[sig];
  if (__sig_should_use_altstack(flags, tib)) {
    __stack_call(ep, code, sig, tib, __sig_unmaskable,
                 tib->tib_sigstack_addr + tib->tib_sigstack_size);
  } else {
    __sig_unmaskable(ep, code, sig, tib);
  }
  // resume running user program
  // hopefully the user fixed the cpu state
  // otherwise the crash will keep happening
  return kNtExceptionContinueExecution;
 }
 static textwindows int __sig_console_sig(uint32_t dwCtrlType) {
  switch (dwCtrlType) {
    case kNtCtrlCEvent:
      return SIGINT;
    case kNtCtrlBreakEvent:
      return SIGQUIT;
    case kNtCtrlCloseEvent:
    case kNtCtrlLogoffEvent:    // only received by services
    case kNtCtrlShutdownEvent:  // only received by services
      return SIGHUP;
    default:
      return SIGSTKFLT;
  }
 }
 __msabi textwindows dontinstrument bool32 __sig_console(uint32_t dwCtrlType) {
  // win32 launches a thread to deliver ctrl-c and ctrl-break when typed
  // it only happens when kNtEnableProcessedInput is in play on console.
  // otherwise we need to wait until read-nt.c discovers that keystroke.
  struct CosmoTib tls;
  __bootstrap_tls(&tls, __builtin_frame_address(0));
  __sig_generate(__sig_console_sig(dwCtrlType), SI_KERNEL);
  return true;
 }
 // returns 0 if no signal handlers were called, otherwise a bitmask
 // consisting of `1` which means a signal handler was invoked which
 // didn't have the SA_RESTART flag, and `2`, which means SA_RESTART
 // handlers were called (or `3` if both were the case).
 textwindows int __sig_check(void) {
  int sig, res = 0;
  while ((sig = __sig_get(atomic_load_explicit(&__get_tls()->tib_sigmask,
                                               memory_order_acquire))))
    res |= __sig_raise(sig, SI_KERNEL);
  return res;
 }
 // background thread for delivering inter-process signals asynchronously
 // this checks for undelivered process-wide signals, once per scheduling
 // quantum, which on windows should be every ~15ms or so, unless somehow
 // the process was tuned to have more fine-grained event timing. we want
 // signals to happen faster when possible; that happens when cancelation
 // points, e.g. read need to wait on i/o; they too check for new signals
 textwindows dontinstrument static uint32_t __sig_worker(void *arg) {
  struct CosmoTib tls;
  __bootstrap_tls(&tls, __builtin_frame_address(0));
  char *sp = __builtin_frame_address(0);
  __maps_track((char *)(((uintptr_t)sp + __pagesize - 1) & -__pagesize) - STKSZ,
               STKSZ);
  for (;;) {
    // dequeue all pending signals and fire them off. if there's no
    // thread that can handle them then __sig_generate will requeue
    // those signals back to __sig.process; hence the need for xchg
    unsigned long sigs =
        atomic_exchange_explicit(__sig.process, 0, memory_order_acq_rel);
    while (sigs) {
      int sig = bsfl(sigs) + 1;
      sigs &= ~(1ull << (sig - 1));
      __sig_generate(sig, SI_KERNEL);
    }
    // unblock stalled asynchronous signals in threads
    _pthread_lock();
    for (struct Dll *e = dll_first(_pthread_list); e;
         e = dll_next(_pthread_list, e)) {
      struct PosixThread *pt = POSIXTHREAD_CONTAINER(e);
      if (atomic_load_explicit(&pt->pt_status, memory_order_acquire) >=
          kPosixThreadTerminated) {
        break;
      }
      sigset_t pending =
          atomic_load_explicit(&pt->tib->tib_sigpending, memory_order_acquire);
      sigset_t mask =
          atomic_load_explicit(&pt->tib->tib_sigmask, memory_order_acquire);
      if (pending & ~mask) {
        _pthread_ref(pt);
        _pthread_unlock();
        while (!atomic_compare_exchange_weak_explicit(
            &pt->tib->tib_sigpending, &pending, pending & ~mask,
            memory_order_acq_rel, memory_order_relaxed)) {
        }
        while ((pending = pending & ~mask)) {
          int sig = bsfl(pending) + 1;
          pending &= ~(1ull << (sig - 1));
          __sig_killer(pt, sig, SI_KERNEL);
        }
        _pthread_lock();
        _pthread_unref(pt);
      }
    }
    _pthread_unlock();
    // wait until next scheduler quantum
    Sleep(POLL_INTERVAL_MS);
  }
  return 0;
 }
 __attribute__((__constructor__(10))) textstartup void __sig_init(void) {
  if (!IsWindows())
    return;
  AddVectoredExceptionHandler(true, (void *)__sig_crash);
  SetConsoleCtrlHandler((void *)__sig_console, true);
  CreateThread(0, STKSZ, __sig_worker, 0, kNtStackSizeParamIsAReservation, 0);
 }
 #endif /* __x86_64__ */
--- a/libc/cosmo.h
+++ b/libc/cosmo.h
@ -1,5 +1,6 @@
 #ifndef COSMOPOLITAN_LIBC_COSMO_H_
 #define COSMOPOLITAN_LIBC_COSMO_H_
 #include "libc/calls/struct/timespec.h"
 COSMOPOLITAN_C_START_
 #ifndef __cplusplus
@ -17,6 +18,9 @@ int __is_mangled(const char *) libcesque;
 bool32 IsLinuxModern(void) libcesque;
 int LoadZipArgs(int *, char ***) libcesque;
 int cosmo_args(const char *, char ***) libcesque;
 int cosmo_futex_wake(_COSMO_ATOMIC(int) *, int, char);
 int cosmo_futex_wait(_COSMO_ATOMIC(int) *, int, char, int,
                     const struct timespec *);
 COSMOPOLITAN_C_END_
 #endif /* COSMOPOLITAN_LIBC_COSMO_H_ */
--- a/libc/intrin/BUILD.mk
+++ b/libc/intrin/BUILD.mk
@ -30,9 +30,11 @@ LIBC_INTRIN_A_CHECKS =					\
 LIBC_INTRIN_A_DIRECTDEPS =				\
 	LIBC_NEXGEN32E					\
 	LIBC_NT_KERNEL32				\
 	LIBC_NT_REALTIME				\
 	LIBC_NT_SYNCHRONIZATION				\
 	LIBC_NT_WS2_32					\
 	LIBC_SYSV					\
-	LIBC_SYSV_CALLS
+	LIBC_SYSV_CALLS					\
 LIBC_INTRIN_A_DEPS :=					\
 	$(call uniq,$(foreach x,$(LIBC_INTRIN_A_DIRECTDEPS),$($(x))))
@ -106,6 +108,16 @@ o//libc/intrin/demangle.o: private			\
 		CFLAGS +=				\
 			-mgeneral-regs-only
 # ensure that division is optimized
 o/$(MODE)/libc/intrin/windowsdurationtotimeval.o	\
 o/$(MODE)/libc/intrin/windowsdurationtotimespec.o	\
 o/$(MODE)/libc/intrin/timevaltowindowstime.o		\
 o/$(MODE)/libc/intrin/timespectowindowstime.o		\
 o/$(MODE)/libc/intrin/windowstimetotimeval.o		\
 o/$(MODE)/libc/intrin/windowstimetotimespec.o: private	\
 		CFLAGS +=				\
 			-O2
 # these assembly files are safe to build on aarch64
 o/$(MODE)/libc/intrin/aarch64/%.o: libc/intrin/aarch64/%.S
 	@$(COMPILE) -AOBJECTIFY.S $(OBJECTIFY.S) $(OUTPUT_OPTION) -c $<
--- a/libc/intrin/checkcancel.c
+++ b/libc/intrin/checkcancel.c
--- a/libc/intrin/clock_gettime-mono.c
+++ b/libc/intrin/clock_gettime-mono.c
--- a/libc/intrin/clock_gettime-nt.c
+++ b/libc/intrin/clock_gettime-nt.c
--- a/libc/intrin/clock_gettime-sysv.c
+++ b/libc/intrin/clock_gettime-sysv.c
--- a/libc/intrin/clock_gettime-xnu.c
+++ b/libc/intrin/clock_gettime-xnu.c
--- a/libc/intrin/clock_gettime.c
+++ b/libc/intrin/clock_gettime.c
--- a/libc/intrin/clock_gettime_monotonic_nt.c
+++ b/libc/intrin/clock_gettime_monotonic_nt.c
--- a/libc/intrin/cosmo_futex.c
+++ b/libc/intrin/cosmo_futex.c
@ -16,18 +16,14 @@
 │ TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR             │
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/sysv/consts/futex.h"
 #include "libc/assert.h"
 #include "libc/atomic.h"
 #include "libc/calls/calls.h"
 #include "libc/calls/internal.h"
 #include "libc/calls/sig.internal.h"
 #include "libc/calls/state.internal.h"
 #include "libc/calls/struct/sigset.h"
 #include "libc/calls/struct/sigset.internal.h"
 #include "libc/calls/struct/timespec.h"
 #include "libc/calls/struct/timespec.internal.h"
 #include "libc/calls/syscall_support-nt.internal.h"
 #include "libc/cosmo.h"
 #include "libc/dce.h"
 #include "libc/errno.h"
@ -37,62 +33,56 @@
 #include "libc/intrin/ulock.h"
 #include "libc/intrin/weaken.h"
 #include "libc/limits.h"
 #include "libc/nexgen32e/vendor.internal.h"
 #include "libc/nt/runtime.h"
 #include "libc/nt/synchronization.h"
-#include "libc/runtime/clktck.h"
+#include "libc/sysv/consts/clock.h"
 #include "libc/sysv/consts/futex.h"
 #include "libc/sysv/consts/sicode.h"
 #include "libc/sysv/consts/timer.h"
 #include "libc/sysv/errfuns.h"
 #include "libc/thread/freebsd.internal.h"
 #include "libc/thread/posixthread.internal.h"
 #include "libc/thread/thread.h"
-#include "libc/thread/tls.h"
+// clang-format off
 #include "third_party/nsync/atomic.h"
 #include "third_party/nsync/time.h"
 #include "third_party/nsync/common.internal.h"
 #include "third_party/nsync/futex.internal.h"
 #include "third_party/nsync/time.h"
 #define FUTEX_WAIT_BITS_ FUTEX_BITSET_MATCH_ANY
-errno_t _futex (atomic_int *, int, int, const struct timespec *, int *, int);
+errno_t cosmo_futex_thunk (atomic_int *, int, int, const struct timespec *, int *, int);
-errno_t _futex_wake (atomic_int *, int, int) asm ("_futex");
+errno_t _futex_wake (atomic_int *, int, int) asm ("cosmo_futex_thunk");
 int sys_futex_cp (atomic_int *, int, int, const struct timespec *, int *, int);
-static struct NsyncFutex {
+static struct CosmoFutex {
 	atomic_uint once;
 	int FUTEX_WAIT_;
 	int FUTEX_PRIVATE_FLAG_;
 	int FUTEX_CLOCK_REALTIME_;
 	bool is_supported;
 	bool timeout_is_relative;
-} nsync_futex_;
+} g_cosmo_futex;
-static void nsync_futex_init_ (void) {
+static void cosmo_futex_init (void) {
 	int e;
 	atomic_int x;
-	nsync_futex_.FUTEX_WAIT_ = FUTEX_WAIT;
+	g_cosmo_futex.FUTEX_WAIT_ = FUTEX_WAIT;
 	if (IsWindows ()) {
-		nsync_futex_.is_supported = true;
+		g_cosmo_futex.is_supported = true;
 		return;
 	}
 	if (IsXnu ()) {
-		nsync_futex_.is_supported = true;
+		g_cosmo_futex.is_supported = true;
-		nsync_futex_.timeout_is_relative = true;
+		g_cosmo_futex.timeout_is_relative = true;
 		return;
 	}
 	if (IsFreebsd ()) {
-		nsync_futex_.is_supported = true;
+		g_cosmo_futex.is_supported = true;
-		nsync_futex_.FUTEX_PRIVATE_FLAG_ = FUTEX_PRIVATE_FLAG;
+		g_cosmo_futex.FUTEX_PRIVATE_FLAG_ = FUTEX_PRIVATE_FLAG;
 		return;
 	}
-	if (!(nsync_futex_.is_supported = IsLinux () || IsOpenbsd ()))
+	if (!(g_cosmo_futex.is_supported = IsLinux () || IsOpenbsd ()))
 		return;
 	// In our testing, we found that the monotonic clock on various
@ -100,7 +90,7 @@ static void nsync_futex_init_ (void) {
 	// better behaved than the realtime clock, and routinely took
 	// large steps backwards, especially on multiprocessors. Given
 	// that "monotonic" doesn't seem to mean what it says,
-	// implementers of nsync_time might consider retaining the
+	// implementers of cosmo_time might consider retaining the
 	// simplicity of a single epoch within an address space, by
 	// configuring any time synchronization mechanism (like ntp) to
 	// adjust for leap seconds by adjusting the rate, rather than
@ -108,31 +98,32 @@ static void nsync_futex_init_ (void) {
 	e = errno;
 	atomic_store_explicit (&x, 0, memory_order_relaxed);
 	if (IsLinux () &&
-	    _futex (&x, FUTEX_WAIT_BITSET | FUTEX_CLOCK_REALTIME,
+	    cosmo_futex_thunk (&x, FUTEX_WAIT_BITSET | FUTEX_CLOCK_REALTIME,
 			       1, 0, 0, FUTEX_BITSET_MATCH_ANY) == -EAGAIN) {
-		nsync_futex_.FUTEX_WAIT_ = FUTEX_WAIT_BITSET;
+		g_cosmo_futex.FUTEX_WAIT_ = FUTEX_WAIT_BITSET;
-		nsync_futex_.FUTEX_PRIVATE_FLAG_ = FUTEX_PRIVATE_FLAG;
+		g_cosmo_futex.FUTEX_PRIVATE_FLAG_ = FUTEX_PRIVATE_FLAG;
-		nsync_futex_.FUTEX_CLOCK_REALTIME_ = FUTEX_CLOCK_REALTIME;
+		g_cosmo_futex.FUTEX_CLOCK_REALTIME_ = FUTEX_CLOCK_REALTIME;
 	} else if (IsOpenbsd () ||
 		   (IsLinux () &&
 		    !_futex_wake (&x, FUTEX_WAKE_PRIVATE, 1))) {
-		nsync_futex_.FUTEX_WAIT_ = FUTEX_WAIT;
+		g_cosmo_futex.FUTEX_WAIT_ = FUTEX_WAIT;
-		nsync_futex_.FUTEX_PRIVATE_FLAG_ = FUTEX_PRIVATE_FLAG;
+		g_cosmo_futex.FUTEX_PRIVATE_FLAG_ = FUTEX_PRIVATE_FLAG;
-		nsync_futex_.timeout_is_relative = true;
+		g_cosmo_futex.timeout_is_relative = true;
 	} else {
-		nsync_futex_.FUTEX_WAIT_ = FUTEX_WAIT;
+		g_cosmo_futex.FUTEX_WAIT_ = FUTEX_WAIT;
-		nsync_futex_.timeout_is_relative = true;
+		g_cosmo_futex.timeout_is_relative = true;
 	}
 	errno = e;
 }
-static uint32_t nsync_time_64to32u (uint64_t duration) {
+static uint32_t cosmo_time_64to32u (uint64_t duration) {
 	if (duration <= -1u)
 		return duration;
 	return -1u;
 }
-static int nsync_futex_polyfill_ (atomic_int *w, int expect, int clock, struct timespec *abstime) {
+static int cosmo_futex_polyfill (atomic_int *w, int expect, int clock,
 				 struct timespec *abstime) {
 	for (;;) {
 		if (atomic_load_explicit (w, memory_order_acquire) != expect)
 			return 0;
@ -148,7 +139,7 @@ static int nsync_futex_polyfill_ (atomic_int *w, int expect, int clock, struct t
 	}
 }
-static int nsync_futex_wait_win32_ (atomic_int *w, int expect, char pshare,
+static int cosmo_futex_wait_win32 (atomic_int *w, int expect, char pshare,
 				   int clock, const struct timespec *timeout,
 				   struct PosixThread *pt,
 				   sigset_t waitmask) {
@ -183,7 +174,7 @@ static int nsync_futex_wait_win32_ (atomic_int *w, int expect, char pshare,
 			pt->pt_blkmask = waitmask;
 			atomic_store_explicit (&pt->pt_blocker, w, memory_order_release);
 		}
-		ok = WaitOnAddress (w, &expect, sizeof(int), nsync_time_64to32u (timespec_tomillis (wait)));
+		ok = WaitOnAddress (w, &expect, sizeof(int), cosmo_time_64to32u (timespec_tomillis (wait)));
 		if (pt) {
 			/* __sig_wake wakes our futex without changing `w` after enqueing signals */
 			atomic_store_explicit (&pt->pt_blocker, 0, memory_order_release);
@ -197,7 +188,7 @@ static int nsync_futex_wait_win32_ (atomic_int *w, int expect, char pshare,
 		if (ok) {
 			return 0;
 		} else {
-			ASSERT (GetLastError () == ETIMEDOUT);
+			unassert (GetLastError () == ETIMEDOUT);
 		}
 	}
 #else
@ -205,14 +196,14 @@ static int nsync_futex_wait_win32_ (atomic_int *w, int expect, char pshare,
 #endif /* __x86_64__ */
 }
-static int nsync_futex_fix_timeout_ (struct timespec *memory, int clock,
+static int cosmo_futex_fix_timeout (struct timespec *memory, int clock,
 				    const struct timespec *abstime,
 				    struct timespec **result) {
 	struct timespec now;
 	if (!abstime) {
 		*result = 0;
 		return 0;
-	} else if (!nsync_futex_.timeout_is_relative) {
+	} else if (!g_cosmo_futex.timeout_is_relative) {
 		*memory = *abstime;
 		*result = memory;
 		return 0;
@ -225,7 +216,24 @@ static int nsync_futex_fix_timeout_ (struct timespec *memory, int clock,
 	}
 }
-int nsync_futex_wait_ (atomic_int *w, int expect, char pshare,
+/**
 * Waits on futex.
 *
 * This function may be used to ask the OS to park the calling thread
 * until cosmo_futex_wake() is called on the memory address `w`.
 *
 * @param w is your futex
 * @param expect is the value `*w` is expected to have on entry
 * @param pshare is `PTHREAD_PROCESS_PRIVATE` / `PTHREAD_PROCESS_SHARED`
 * @param clock is `CLOCK_MONOTONIC`, `CLOCK_REALTIME`, etc.
 * @param abstime is null to wait forever or absolute timestamp to stop
 * @return 0 on success, or -errno on error
 * @raise EINVAL on bad parameter
 * @raise EAGAIN if `*w` wasn't `expect`
 * @raise EINTR if a signal handler was called while waiting
 * @raise ECANCELED if calling thread was canceled while waiting
 */
 int cosmo_futex_wait (atomic_int *w, int expect, char pshare,
 		      int clock, const struct timespec *abstime) {
 	int e, rc, op;
 	struct CosmoTib *tib;
@ -233,14 +241,14 @@ int nsync_futex_wait_ (atomic_int *w, int expect, char pshare,
 	struct timespec tsmem;
 	struct timespec *timeout = 0;
-	cosmo_once (&nsync_futex_.once, nsync_futex_init_);
+	cosmo_once (&g_cosmo_futex.once, cosmo_futex_init);
-	op = nsync_futex_.FUTEX_WAIT_;
+	op = g_cosmo_futex.FUTEX_WAIT_;
 	if (pshare == PTHREAD_PROCESS_PRIVATE)
-		op |= nsync_futex_.FUTEX_PRIVATE_FLAG_;
+		op |= g_cosmo_futex.FUTEX_PRIVATE_FLAG_;
 	if (clock == CLOCK_REALTIME ||
 	    clock == CLOCK_REALTIME_COARSE)
-		op |= nsync_futex_.FUTEX_CLOCK_REALTIME_;
+		op |= g_cosmo_futex.FUTEX_CLOCK_REALTIME_;
 	if (abstime && timespec_cmp (*abstime, timespec_zero) <= 0) {
 		rc = -ETIMEDOUT;
@ -252,7 +260,7 @@ int nsync_futex_wait_ (atomic_int *w, int expect, char pshare,
 		goto Finished;
 	}
-	if ((rc = nsync_futex_fix_timeout_ (&tsmem, clock, abstime, &timeout)))
+	if ((rc = cosmo_futex_fix_timeout (&tsmem, clock, abstime, &timeout)))
 		goto Finished;
 	LOCKTRACE ("futex(%t [%d], %s, %#x, %s) → ...",
@ -263,13 +271,13 @@ int nsync_futex_wait_ (atomic_int *w, int expect, char pshare,
 	tib = __get_tls();
 	pt = (struct PosixThread *)tib->tib_pthread;
-	if (nsync_futex_.is_supported) {
+	if (g_cosmo_futex.is_supported) {
 		e = errno;
 		if (IsWindows ()) {
 			// Windows 8 futexes don't support multiple processes :(
 			if (pshare) goto Polyfill;
 			sigset_t m = __sig_block ();
-			rc = nsync_futex_wait_win32_ (w, expect, pshare, clock, timeout, pt, m);
+			rc = cosmo_futex_wait_win32 (w, expect, pshare, clock, timeout, pt, m);
 			__sig_unblock (m);
 		} else if (IsXnu ()) {
@ -293,7 +301,7 @@ int nsync_futex_wait_ (atomic_int *w, int expect, char pshare,
 				op = UL_COMPARE_AND_WAIT;
 			}
 			if (timeout) {
-				us = nsync_time_64to32u (timespec_tomicros (*timeout));
+				us = cosmo_time_64to32u (timespec_tomicros (*timeout));
 			} else {
 				us = -1u;
 			}
@ -333,7 +341,7 @@ int nsync_futex_wait_ (atomic_int *w, int expect, char pshare,
 		}
 	} else {
 	Polyfill:
-		rc = nsync_futex_polyfill_ (w, expect, clock, timeout);
+		rc = cosmo_futex_polyfill (w, expect, clock, timeout);
 	}
 Finished:
@ -346,18 +354,24 @@ Finished:
 	return rc;
 }
-int nsync_futex_wake_ (atomic_int *w, int count, char pshare) {
+/**
 * Wakes futex.
 *
 * @param w is your futex
 * @param count is number of threads to wake (usually 1 or `INT_MAX`)
 * @param pshare is `PTHREAD_PROCESS_PRIVATE` / `PTHREAD_PROCESS_SHARED`
 * @return number of threads woken on success, or -errno on error
 */
 int cosmo_futex_wake (atomic_int *w, int count, char pshare) {
 	int rc, op, fop;
-	ASSERT (count == 1 || count == INT_MAX);
+	cosmo_once (&g_cosmo_futex.once, cosmo_futex_init);
 	cosmo_once (&nsync_futex_.once, nsync_futex_init_);
 	op = FUTEX_WAKE;
 	if (pshare == PTHREAD_PROCESS_PRIVATE)
-		op |= nsync_futex_.FUTEX_PRIVATE_FLAG_;
+		op |= g_cosmo_futex.FUTEX_PRIVATE_FLAG_;
-	if (nsync_futex_.is_supported) {
+	if (g_cosmo_futex.is_supported) {
 		if (IsWindows ()) {
 			if (pshare) {
 				goto Polyfill;
@ -379,7 +393,7 @@ int nsync_futex_wake_ (atomic_int *w, int count, char pshare) {
 				op |= ULF_WAKE_ALL;
 			}
 			rc = ulock_wake (op, w, 0);
-			ASSERT (!rc || rc == -ENOENT);
+			unassert (!rc || rc == -ENOENT);
 			if (!rc) {
 				rc = 1;
 			} else if (rc == -ENOENT) {
--- a/libc/intrin/cosmo_futex_thunk.S
+++ b/libc/intrin/cosmo_futex_thunk.S
@ -20,7 +20,7 @@
 #include "libc/macros.h"
 .privileged
-_futex:
+cosmo_futex_thunk:
 #ifdef __x86_64__
 	push	%rbp
 	mov	%rsp,%rbp
@ -47,4 +47,4 @@ _futex:
 #error "unsupported architecture"
 #endif /* __x86_64__ */
 1:	ret
-	.endfn	_futex,globl,hidden
+	.endfn	cosmo_futex_thunk,globl,hidden
--- a/libc/intrin/getcontext.S
+++ b/libc/intrin/getcontext.S
@ -27,6 +27,6 @@
 	.ftrace1
 getcontext:
 	.ftrace2
-#include "libc/calls/getcontext.inc"
+#include "libc/intrin/getcontext.inc"
 	jmp	__getcontextsig
 	.endfn	getcontext,globl
--- a/libc/intrin/getcontext.inc
+++ b/libc/intrin/getcontext.inc
--- a/libc/intrin/pthread_mutex_lock.c
+++ b/libc/intrin/pthread_mutex_lock.c
@ -19,6 +19,7 @@
 #include "libc/calls/blockcancel.internal.h"
 #include "libc/calls/calls.h"
 #include "libc/calls/state.internal.h"
 #include "libc/cosmo.h"
 #include "libc/dce.h"
 #include "libc/errno.h"
 #include "libc/intrin/atomic.h"
@ -28,25 +29,8 @@
 #include "libc/runtime/internal.h"
 #include "libc/thread/lock.h"
 #include "libc/thread/thread.h"
 #include "third_party/nsync/futex.internal.h"
 #include "third_party/nsync/mu.h"
 static void pthread_mutex_lock_spin(atomic_int *word) {
  int backoff = 0;
  if (atomic_exchange_explicit(word, 1, memory_order_acquire)) {
    LOCKTRACE("acquiring pthread_mutex_lock_spin(%t)...", word);
    for (;;) {
      for (;;) {
        if (!atomic_load_explicit(word, memory_order_relaxed))
          break;
        backoff = pthread_delay_np(word, backoff);
      }
      if (!atomic_exchange_explicit(word, 1, memory_order_acquire))
        break;
    }
  }
 }
 // see "take 3" algorithm in "futexes are tricky" by ulrich drepper
 // slightly improved to attempt acquiring multiple times b4 syscall
 static void pthread_mutex_lock_drepper(atomic_int *futex, char pshare) {
@ -59,7 +43,7 @@ static void pthread_mutex_lock_drepper(atomic_int *futex, char pshare) {
    word = atomic_exchange_explicit(futex, 2, memory_order_acquire);
  BLOCK_CANCELATION;
  while (word > 0) {
-    _weaken(nsync_futex_wait_)(futex, 2, pshare, 0, 0);
+    cosmo_futex_wait(futex, 2, pshare, 0, 0);
    word = atomic_exchange_explicit(futex, 2, memory_order_acquire);
  }
  ALLOW_CANCELATION;
@ -164,11 +148,7 @@ static errno_t pthread_mutex_lock_impl(pthread_mutex_t *mutex) {
  // handle normal mutexes
  if (MUTEX_TYPE(word) == PTHREAD_MUTEX_NORMAL) {
    if (_weaken(nsync_futex_wait_)) {
    pthread_mutex_lock_drepper(&mutex->_futex, MUTEX_PSHARED(word));
    } else {
      pthread_mutex_lock_spin(&mutex->_futex);
    }
    return 0;
  }
--- a/libc/intrin/pthread_mutex_trylock.c
+++ b/libc/intrin/pthread_mutex_trylock.c
@ -24,15 +24,8 @@
 #include "libc/runtime/internal.h"
 #include "libc/thread/lock.h"
 #include "libc/thread/thread.h"
 #include "third_party/nsync/futex.internal.h"
 #include "third_party/nsync/mu.h"
 static errno_t pthread_mutex_trylock_spin(atomic_int *word) {
  if (!atomic_exchange_explicit(word, 1, memory_order_acquire))
    return 0;
  return EBUSY;
 }
 static errno_t pthread_mutex_trylock_drepper(atomic_int *futex) {
  int word = 0;
  if (atomic_compare_exchange_strong_explicit(
@ -142,13 +135,8 @@ errno_t pthread_mutex_trylock(pthread_mutex_t *mutex) {
 #endif
  // handle normal mutexes
-  if (MUTEX_TYPE(word) == PTHREAD_MUTEX_NORMAL) {
+  if (MUTEX_TYPE(word) == PTHREAD_MUTEX_NORMAL)
    if (_weaken(nsync_futex_wait_)) {
    return pthread_mutex_trylock_drepper(&mutex->_futex);
    } else {
      return pthread_mutex_trylock_spin(&mutex->_futex);
    }
  }
  // handle recursive and error checking mutexes
 #if PTHREAD_USE_NSYNC
--- a/libc/intrin/pthread_mutex_unlock.c
+++ b/libc/intrin/pthread_mutex_unlock.c
@ -18,6 +18,7 @@
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/calls/calls.h"
 #include "libc/calls/state.internal.h"
 #include "libc/cosmo.h"
 #include "libc/dce.h"
 #include "libc/errno.h"
 #include "libc/intrin/atomic.h"
@ -26,19 +27,14 @@
 #include "libc/runtime/internal.h"
 #include "libc/thread/lock.h"
 #include "libc/thread/thread.h"
 #include "third_party/nsync/futex.internal.h"
 #include "third_party/nsync/mu.h"
 static void pthread_mutex_unlock_spin(atomic_int *word) {
  atomic_store_explicit(word, 0, memory_order_release);
 }
 // see "take 3" algorithm in "futexes are tricky" by ulrich drepper
 static void pthread_mutex_unlock_drepper(atomic_int *futex, char pshare) {
  int word = atomic_fetch_sub_explicit(futex, 1, memory_order_release);
  if (word == 2) {
    atomic_store_explicit(futex, 0, memory_order_release);
-    _weaken(nsync_futex_wake_)(futex, 1, pshare);
+    cosmo_futex_wake(futex, 1, pshare);
  }
 }
@ -137,11 +133,7 @@ errno_t pthread_mutex_unlock(pthread_mutex_t *mutex) {
  // implement barebones normal mutexes
  if (MUTEX_TYPE(word) == PTHREAD_MUTEX_NORMAL) {
    if (_weaken(nsync_futex_wake_)) {
    pthread_mutex_unlock_drepper(&mutex->_futex, MUTEX_PSHARED(word));
    } else {
      pthread_mutex_unlock_spin(&mutex->_futex);
    }
    return 0;
  }
--- a/libc/intrin/restore.S
+++ b/libc/intrin/restore.S
--- a/libc/intrin/sig.c
+++ b/libc/intrin/sig.c
@ -1,7 +1,7 @@
 /*-*- mode:c;indent-tabs-mode:nil;c-basic-offset:2;tab-width:8;coding:utf-8 -*-│
 │ vi: set et ft=c ts=2 sts=2 sw=2 fenc=utf-8                               :vi │
 ╞══════════════════════════════════════════════════════════════════════════════╡
-│ Copyright 2023 Justine Alexandra Roberts Tunney                              │
+│ Copyright 2022 Justine Alexandra Roberts Tunney                              │
 │                                                                              │
 │ Permission to use, copy, modify, and/or distribute this software for         │
 │ any purpose with or without fee is hereby granted, provided that the         │
@ -17,37 +17,701 @@
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/sysv/consts/sig.h"
 #include "ape/sections.internal.h"
 #include "libc/calls/calls.h"
 #include "libc/calls/sig.internal.h"
 #include "libc/calls/state.internal.h"
 #include "libc/calls/struct/sigaction.h"
 #include "libc/calls/struct/siginfo.h"
 #include "libc/calls/struct/sigset.internal.h"
 #include "libc/calls/struct/ucontext.internal.h"
 #include "libc/calls/syscall_support-nt.internal.h"
 #include "libc/calls/ucontext.h"
 #include "libc/dce.h"
 #include "libc/errno.h"
 #include "libc/intrin/atomic.h"
 #include "libc/intrin/bsf.h"
 #include "libc/intrin/describebacktrace.h"
 #include "libc/intrin/dll.h"
 #include "libc/intrin/maps.h"
 #include "libc/intrin/strace.h"
 #include "libc/intrin/weaken.h"
-#include "libc/thread/tls.h"
+#include "libc/nt/console.h"
 #include "libc/nt/enum/context.h"
 #include "libc/nt/enum/exceptionhandleractions.h"
 #include "libc/nt/enum/processcreationflags.h"
 #include "libc/nt/enum/signal.h"
 #include "libc/nt/enum/status.h"
 #include "libc/nt/events.h"
 #include "libc/nt/runtime.h"
 #include "libc/nt/signals.h"
 #include "libc/nt/struct/ntexceptionpointers.h"
 #include "libc/nt/synchronization.h"
 #include "libc/nt/thread.h"
 #include "libc/runtime/internal.h"
 #include "libc/runtime/symbols.internal.h"
 #include "libc/str/str.h"
 #include "libc/sysv/consts/sa.h"
 #include "libc/sysv/consts/sicode.h"
 #include "libc/sysv/consts/ss.h"
 #include "libc/thread/posixthread.internal.h"
 #ifdef __x86_64__
-struct Signals __sig;
+/**
 * @fileoverview Cosmopolitan Signals for Windows.
 */
-sigset_t __sig_block(void) {
+#define STKSZ 65536
-  if (IsWindows() || IsMetal()) {
+
-    if (__tls_enabled)
+struct SignalFrame {
-      return atomic_exchange_explicit(&__get_tls()->tib_sigmask, -1,
+  unsigned rva;
-                                      memory_order_acquire);
+  unsigned flags;
-    else
+  siginfo_t si;
  ucontext_t ctx;
 };
 static textwindows bool __sig_ignored_by_default(int sig) {
  return sig == SIGURG ||   //
         sig == SIGCONT ||  //
         sig == SIGCHLD ||  //
         sig == SIGWINCH;
 }
 textwindows bool __sig_ignored(int sig) {
  return __sighandrvas[sig] == (intptr_t)SIG_IGN ||
         (__sighandrvas[sig] == (intptr_t)SIG_DFL &&
          __sig_ignored_by_default(sig));
 }
 textwindows void __sig_delete(int sig) {
  struct Dll *e;
  atomic_fetch_and_explicit(__sig.process, ~(1ull << (sig - 1)),
                            memory_order_relaxed);
  _pthread_lock();
  for (e = dll_last(_pthread_list); e; e = dll_prev(_pthread_list, e))
    atomic_fetch_and_explicit(&POSIXTHREAD_CONTAINER(e)->tib->tib_sigpending,
                              ~(1ull << (sig - 1)), memory_order_relaxed);
  _pthread_unlock();
 }
 static textwindows int __sig_getter(atomic_ulong *sigs, sigset_t masked) {
  int sig;
  sigset_t bit, pending, deliverable;
  for (;;) {
    pending = atomic_load_explicit(sigs, memory_order_acquire);
    if ((deliverable = pending & ~masked)) {
      sig = bsfl(deliverable) + 1;
      bit = 1ull << (sig - 1);
      if (atomic_fetch_and_explicit(sigs, ~bit, memory_order_acq_rel) & bit)
        return sig;
    } else {
      return 0;
-  } else {
+    }
    sigset_t res, neu = -1;
    sys_sigprocmask(SIG_SETMASK, &neu, &res);
    return res;
  }
 }
-void __sig_unblock(sigset_t m) {
+textwindows int __sig_get(sigset_t masked) {
-  if (IsWindows() || IsMetal()) {
+  int sig;
-    if (__tls_enabled) {
+  if (!(sig = __sig_getter(&__get_tls()->tib_sigpending, masked)))
-      atomic_store_explicit(&__get_tls()->tib_sigmask, m, memory_order_release);
+    sig = __sig_getter(__sig.process, masked);
-      if (_weaken(__sig_check))
+  return sig;
-        _weaken(__sig_check)();
+}
 static textwindows bool __sig_should_use_altstack(unsigned flags,
                                                  struct CosmoTib *tib) {
  if (!(flags & SA_ONSTACK))
    return false;  // signal handler didn't enable it
  if (!tib->tib_sigstack_size)
    return false;  // sigaltstack() wasn't installed on this thread
  if (tib->tib_sigstack_flags & SS_DISABLE)
    return false;  // sigaltstack() on this thread was disabled by user
  char *bp = __builtin_frame_address(0);
  if (tib->tib_sigstack_addr <= bp &&
      bp <= tib->tib_sigstack_addr + tib->tib_sigstack_size)
    return false;  // we're already on the alternate stack
  return true;
 }
 static textwindows wontreturn void __sig_terminate(int sig) {
  TerminateThisProcess(sig);
 }
 textwindows static bool __sig_wake(struct PosixThread *pt, int sig) {
  atomic_int *blocker;
  blocker = atomic_load_explicit(&pt->pt_blocker, memory_order_acquire);
  if (!blocker)
    return false;
  // threads can create semaphores on an as-needed basis
  if (blocker == PT_BLOCKER_EVENT) {
    STRACE("%G set %d's event object", sig, _pthread_tid(pt));
    SetEvent(pt->pt_event);
    return !!atomic_load_explicit(&pt->pt_blocker, memory_order_acquire);
  }
  // all other blocking ops that aren't overlap should use futexes
  // we force restartable futexes to churn by waking w/o releasing
  STRACE("%G waking %d's futex", sig, _pthread_tid(pt));
  WakeByAddressSingle(blocker);
  return !!atomic_load_explicit(&pt->pt_blocker, memory_order_acquire);
 }
 textwindows static bool __sig_start(struct PosixThread *pt, int sig,
                                    unsigned *rva, unsigned *flags) {
  *rva = __sighandrvas[sig];
  *flags = __sighandflags[sig];
  if (*rva == (intptr_t)SIG_IGN ||
      (*rva == (intptr_t)SIG_DFL && __sig_ignored_by_default(sig))) {
    STRACE("ignoring %G", sig);
    return false;
  }
  if (atomic_load_explicit(&pt->tib->tib_sigmask, memory_order_acquire) &
      (1ull << (sig - 1))) {
    STRACE("enqueing %G on %d", sig, _pthread_tid(pt));
    atomic_fetch_or_explicit(&pt->tib->tib_sigpending, 1ull << (sig - 1),
                             memory_order_relaxed);
    __sig_wake(pt, sig);
    return false;
  }
  if (*rva == (intptr_t)SIG_DFL) {
    STRACE("terminating on %G due to no handler", sig);
    __sig_terminate(sig);
  }
  return true;
 }
 textwindows static sigaction_f __sig_handler(unsigned rva) {
  atomic_fetch_add_explicit(&__sig.count, 1, memory_order_relaxed);
  return (sigaction_f)(__executable_start + rva);
 }
 textwindows int __sig_raise(volatile int sig, int sic) {
  // bitset of kinds of handlers called
  volatile int handler_was_called = 0;
  // loop over pending signals
  ucontext_t ctx;
  getcontext(&ctx);
  if (!sig) {
    if ((sig = __sig_get(ctx.uc_sigmask))) {
      sic = SI_KERNEL;
    } else {
-    sys_sigprocmask(SIG_SETMASK, &m, 0);
+      return handler_was_called;
    }
  }
  // process signal(s)
  unsigned rva, flags;
  struct PosixThread *pt = _pthread_self();
  if (__sig_start(pt, sig, &rva, &flags)) {
    if (flags & SA_RESETHAND) {
      STRACE("resetting %G handler", sig);
      __sighandrvas[sig] = (int32_t)(intptr_t)SIG_DFL;
    }
    // update the signal mask in preparation for signal handller
    sigset_t blocksigs = __sighandmask[sig];
    if (!(flags & SA_NODEFER))
      blocksigs |= 1ull << (sig - 1);
    ctx.uc_sigmask = atomic_fetch_or_explicit(&pt->tib->tib_sigmask, blocksigs,
                                              memory_order_acquire);
    // call the user's signal handler
    char ssbuf[128];
    siginfo_t si = {.si_signo = sig, .si_code = sic};
    STRACE("__sig_raise(%G, %t) mask %s", sig, __sig_handler(rva),
           _DescribeSigset(ssbuf, 0, (sigset_t *)&pt->tib->tib_sigmask));
    __sig_handler(rva)(sig, &si, &ctx);
    // record this handler
    if (flags & SA_RESTART) {
      handler_was_called |= SIG_HANDLED_SA_RESTART;
    } else {
      handler_was_called |= SIG_HANDLED_NO_RESTART;
    }
  }
  // restore sigmask
  // loop back to top
  // jump where handler says
  sig = 0;
  return setcontext(&ctx);
 }
 textwindows int __sig_relay(int sig, int sic, sigset_t waitmask) {
  sigset_t m;
  int handler_was_called;
  m = atomic_exchange_explicit(&__get_tls()->tib_sigmask, waitmask,
                               memory_order_acquire);
  handler_was_called = __sig_raise(sig, SI_KERNEL);
  atomic_store_explicit(&__get_tls()->tib_sigmask, m, memory_order_release);
  return handler_was_called;
 }
 // the user's signal handler callback is wrapped with this trampoline
 static textwindows wontreturn void __sig_tramp(struct SignalFrame *sf) {
  int sig = sf->si.si_signo;
  struct CosmoTib *tib = __get_tls();
  struct PosixThread *pt = (struct PosixThread *)tib->tib_pthread;
  atomic_store_explicit(&pt->pt_intoff, 0, memory_order_release);
  for (;;) {
    // update the signal mask in preparation for signal handler
    sigset_t blocksigs = __sighandmask[sig];
    if (!(sf->flags & SA_NODEFER))
      blocksigs |= 1ull << (sig - 1);
    sf->ctx.uc_sigmask = atomic_fetch_or_explicit(&tib->tib_sigmask, blocksigs,
                                                  memory_order_acquire);
    // call the user's signal handler
    char ssbuf[2][128];
    STRACE("__sig_tramp(%G, %t) mask %s → %s", sig, __sig_handler(sf->rva),
           _DescribeSigset(ssbuf[0], 0, &sf->ctx.uc_sigmask),
           _DescribeSigset(ssbuf[1], 0, (sigset_t *)&tib->tib_sigmask));
    __sig_handler(sf->rva)(sig, &sf->si, &sf->ctx);
    // restore the signal mask that was used by the interrupted code
    // this may have been modified by the signal handler in the callback
    atomic_store_explicit(&tib->tib_sigmask, sf->ctx.uc_sigmask,
                          memory_order_release);
    // jump back into original code if there aren't any pending signals
    do {
      if (!(sig = __sig_get(sf->ctx.uc_sigmask)))
        __sig_restore(&sf->ctx);
    } while (!__sig_start(pt, sig, &sf->rva, &sf->flags));
    // tail recurse into another signal handler
    sf->si.si_signo = sig;
    sf->si.si_code = SI_KERNEL;
    if (sf->flags & SA_RESETHAND) {
      STRACE("resetting %G handler", sig);
      __sighandrvas[sig] = (int32_t)(intptr_t)SIG_DFL;
    }
  }
 }
 // sends signal to another specific thread which is ref'd
 static textwindows int __sig_killer(struct PosixThread *pt, int sig, int sic) {
  unsigned rva = __sighandrvas[sig];
  unsigned flags = __sighandflags[sig];
  // do nothing if signal is ignored
  if (rva == (intptr_t)SIG_IGN ||
      (rva == (intptr_t)SIG_DFL && __sig_ignored_by_default(sig))) {
    STRACE("ignoring %G", sig);
    return 0;
  }
  // we can't preempt threads that masked sigs or are blocked on i/o
  while ((atomic_load_explicit(&pt->tib->tib_sigmask, memory_order_acquire) &
          (1ull << (sig - 1)))) {
    if (atomic_fetch_or_explicit(&pt->tib->tib_sigpending, 1ull << (sig - 1),
                                 memory_order_acq_rel) &
        (1ull << (sig - 1)))
      // we believe signal was already enqueued
      return 0;
    if (__sig_wake(pt, sig))
      // we believe i/o routine will handle signal
      return 0;
    if (atomic_load_explicit(&pt->tib->tib_sigmask, memory_order_acquire) &
        (1ull << (sig - 1)))
      // we believe ALLOW_SIGNALS will handle signal
      return 0;
    if (!(atomic_fetch_and_explicit(&pt->tib->tib_sigpending,
                                    ~(1ull << (sig - 1)),
                                    memory_order_acq_rel) &
          (1ull << (sig - 1))))
      // we believe another thread sniped our signal
      return 0;
    break;
  }
  // avoid race conditions and deadlocks with thread suspend process
  if (atomic_exchange_explicit(&pt->pt_intoff, 1, memory_order_acquire)) {
    // we believe another thread is asynchronously waking the mark
    if (atomic_fetch_or_explicit(&pt->tib->tib_sigpending, 1ull << (sig - 1),
                                 memory_order_acq_rel) &
        (1ull << (sig - 1)))
      // we believe our signal is already being delivered
      return 0;
    if (atomic_load_explicit(&pt->pt_intoff, memory_order_acquire) ||
        atomic_exchange_explicit(&pt->pt_intoff, 1, memory_order_acquire))
      // we believe __sig_tramp will deliver our signal
      return 0;
    if (!(atomic_fetch_and_explicit(&pt->tib->tib_sigpending,
                                    ~(1ull << (sig - 1)),
                                    memory_order_acq_rel) &
          (1ull << (sig - 1))))
      // we believe another thread sniped our signal
      return 0;
  }
  // if there's no handler then killing a thread kills the process
  if (rva == (intptr_t)SIG_DFL) {
    STRACE("terminating on %G due to no handler", sig);
    __sig_terminate(sig);
  }
  // take control of thread
  // suspending the thread happens asynchronously
  // however getting the context blocks until it's frozen
  uintptr_t th = _pthread_syshand(pt);
  if (SuspendThread(th) == -1u) {
    STRACE("SuspendThread failed w/ %d", GetLastError());
    atomic_store_explicit(&pt->pt_intoff, 0, memory_order_release);
    return ESRCH;
  }
  struct NtContext nc;
  nc.ContextFlags = kNtContextFull;
  if (!GetThreadContext(th, &nc)) {
    STRACE("GetThreadContext failed w/ %d", GetLastError());
    ResumeThread(th);
    atomic_store_explicit(&pt->pt_intoff, 0, memory_order_release);
    return ESRCH;
  }
  // we can't preempt threads that masked sig or are blocked
  // we can't preempt threads that are running in win32 code
  // so we shall unblock the thread and let it signal itself
  if (!((uintptr_t)__executable_start <= nc.Rip &&
        nc.Rip < (uintptr_t)__privileged_start)) {
    atomic_fetch_or_explicit(&pt->tib->tib_sigpending, 1ull << (sig - 1),
                             memory_order_relaxed);
    ResumeThread(th);
    atomic_store_explicit(&pt->pt_intoff, 0, memory_order_release);
    __sig_wake(pt, sig);
    return 0;
  }
  // preferring to live dangerously
  // the thread will be signaled asynchronously
  if (flags & SA_RESETHAND) {
    STRACE("resetting %G handler", sig);
    __sighandrvas[sig] = (int32_t)(intptr_t)SIG_DFL;
  }
  // inject call to trampoline function into thread
  uintptr_t sp;
  if (__sig_should_use_altstack(flags, pt->tib)) {
    sp = (uintptr_t)pt->tib->tib_sigstack_addr + pt->tib->tib_sigstack_size;
  } else {
    sp = nc.Rsp;
  }
  sp -= sizeof(struct SignalFrame);
  sp &= -16;
  struct SignalFrame *sf = (struct SignalFrame *)sp;
  _ntcontext2linux(&sf->ctx, &nc);
  bzero(&sf->si, sizeof(sf->si));
  sf->rva = rva;
  sf->flags = flags;
  sf->si.si_code = sic;
  sf->si.si_signo = sig;
  *(uintptr_t *)(sp -= sizeof(uintptr_t)) = nc.Rip;
  nc.Rip = (intptr_t)__sig_tramp;
  nc.Rdi = (intptr_t)sf;
  nc.Rsp = sp;
  if (!SetThreadContext(th, &nc)) {
    STRACE("SetThreadContext failed w/ %d", GetLastError());
    atomic_store_explicit(&pt->pt_intoff, 0, memory_order_release);
    return ESRCH;
  }
  ResumeThread(th);
  __sig_wake(pt, sig);
  return 0;
 }
 // sends signal to another specific thread
 textwindows int __sig_kill(struct PosixThread *pt, int sig, int sic) {
  int rc;
  BLOCK_SIGNALS;
  rc = __sig_killer(pt, sig, sic);
  ALLOW_SIGNALS;
  return rc;
 }
 // sends signal to any other thread
 // this should only be called by non-posix threads
 textwindows void __sig_generate(int sig, int sic) {
  struct Dll *e;
  struct PosixThread *pt, *mark = 0;
  if (__sig_ignored(sig)) {
    STRACE("ignoring %G", sig);
    return;
  }
  if (__sighandrvas[sig] == (intptr_t)SIG_DFL) {
    STRACE("terminating on %G due to no handler", sig);
    __sig_terminate(sig);
  }
  if (atomic_load_explicit(__sig.process, memory_order_acquire) &
      (1ull << (sig - 1))) {
    return;
  }
  _pthread_lock();
  for (e = dll_first(_pthread_list); e; e = dll_next(_pthread_list, e)) {
    pt = POSIXTHREAD_CONTAINER(e);
    // we don't want to signal ourself
    if (pt == _pthread_self())
      continue;
    // we don't want to signal a thread that isn't running
    if (atomic_load_explicit(&pt->pt_status, memory_order_acquire) >=
        kPosixThreadTerminated) {
      continue;
    }
    // choose this thread if it isn't masking sig
    if (!(atomic_load_explicit(&pt->tib->tib_sigmask, memory_order_acquire) &
          (1ull << (sig - 1)))) {
      _pthread_ref(pt);
      mark = pt;
      break;
    }
    // if a thread is blocking then we check to see if it's planning
    // to unblock our sig once the wait operation is completed; when
    // that's the case we can cancel the thread's i/o to deliver sig
    if (atomic_load_explicit(&pt->pt_blocker, memory_order_acquire) &&
        !(pt->pt_blkmask & (1ull << (sig - 1)))) {
      _pthread_ref(pt);
      mark = pt;
      break;
    }
  }
  _pthread_unlock();
  if (mark) {
    // no lock needed since current thread is nameless and formless
    __sig_killer(mark, sig, sic);
    _pthread_unref(mark);
  } else {
    atomic_fetch_or_explicit(__sig.process, 1ull << (sig - 1),
                             memory_order_relaxed);
  }
 }
 static textwindows char *__sig_stpcpy(char *d, const char *s) {
  size_t i;
  for (i = 0;; ++i)
    if (!(d[i] = s[i]))
      return d + i;
 }
 static textwindows wontreturn void __sig_death(int sig, const char *thing) {
 #ifndef TINY
  intptr_t hStderr;
  char sigbuf[21], s[128], *p;
  hStderr = GetStdHandle(kNtStdErrorHandle);
  p = __sig_stpcpy(s, "Terminating on ");
  p = __sig_stpcpy(p, thing);
  p = __sig_stpcpy(p, strsignal_r(sig, sigbuf));
  p = __sig_stpcpy(p,
                   ". Pass --strace and/or ShowCrashReports() for details.\n");
  WriteFile(hStderr, s, p - s, 0, 0);
 #endif
  __sig_terminate(sig);
 }
 static textwindows void __sig_unmaskable(struct NtExceptionPointers *ep,
                                         int code, int sig,
                                         struct CosmoTib *tib) {
  // log vital crash information reliably for --strace before doing much
  // we don't print this without the flag since raw numbers scare people
  // this needs at least one page of stack memory in order to get logged
  // otherwise it'll print a warning message about the lack of stack mem
  STRACE("win32 vectored exception 0x%08Xu raising %G "
         "cosmoaddr2line %s %lx %s",
         ep->ExceptionRecord->ExceptionCode, sig,
         _weaken(FindDebugBinary) ? _weaken(FindDebugBinary)()
                                  : program_invocation_name,
         ep->ContextRecord->Rip,
         DescribeBacktrace((struct StackFrame *)ep->ContextRecord->Rbp));
  // if the user didn't install a signal handler for this unmaskable
  // exception, then print a friendly helpful hint message to stderr
  unsigned rva = __sighandrvas[sig];
  if (rva == (intptr_t)SIG_DFL || rva == (intptr_t)SIG_IGN)
    __sig_death(sig, "uncaught ");
  // if this signal handler is configured to auto-reset to the default
  // then that reset needs to happen before the user handler is called
  unsigned flags = __sighandflags[sig];
  if (flags & SA_RESETHAND) {
    STRACE("resetting %G handler", sig);
    __sighandrvas[sig] = (int32_t)(intptr_t)SIG_DFL;
  }
  // determine the true memory address at which fault occurred
  // if this is a stack overflow then reapply guard protection
  void *si_addr;
  if (ep->ExceptionRecord->ExceptionCode == kNtSignalGuardPage) {
    si_addr = (void *)ep->ExceptionRecord->ExceptionInformation[1];
  } else {
    si_addr = ep->ExceptionRecord->ExceptionAddress;
  }
  // call the user signal handler
  // and a modifiable view of the faulting code's cpu state
  // temporarily replace signal mask while calling crash handler
  // abort process if sig is already blocked to avoid crash loop
  // note ucontext_t is a hefty data structures on top of NtContext
  ucontext_t ctx = {0};
  siginfo_t si = {.si_signo = sig, .si_code = code, .si_addr = si_addr};
  _ntcontext2linux(&ctx, ep->ContextRecord);
  sigset_t blocksigs = __sighandmask[sig];
  if (!(flags & SA_NODEFER))
    blocksigs |= 1ull << (sig - 1);
  ctx.uc_sigmask = atomic_fetch_or_explicit(&tib->tib_sigmask, blocksigs,
                                            memory_order_acquire);
  if (ctx.uc_sigmask & (1ull << (sig - 1))) {
    __sig_death(sig, "masked ");
    __sig_terminate(sig);
  }
  __sig_handler(rva)(sig, &si, &ctx);
  atomic_store_explicit(&tib->tib_sigmask, ctx.uc_sigmask,
                        memory_order_release);
  _ntlinux2context(ep->ContextRecord, &ctx);
 }
 void __stack_call(struct NtExceptionPointers *, int, int, struct CosmoTib *,
                  void (*)(struct NtExceptionPointers *, int, int,
                           struct CosmoTib *),
                  void *);
 //                         abashed the devil stood
 //                      and felt how awful goodness is
 __msabi dontinstrument unsigned __sig_crash(struct NtExceptionPointers *ep) {
  // translate win32 to unix si_signo and si_code
  int code, sig = __sig_crash_sig(ep->ExceptionRecord->ExceptionCode, &code);
  // advance the instruction pointer to skip over debugger breakpoints
  // this behavior is consistent with how unix kernels are implemented
  if (sig == SIGTRAP) {
    ep->ContextRecord->Rip++;
    if (__sig_ignored(sig))
      return kNtExceptionContinueExecution;
  }
  // win32 stack overflow detection executes INSIDE the guard page
  // thus switch to the alternate signal stack as soon as possible
  struct CosmoTib *tib = __get_tls();
  unsigned flags = __sighandflags[sig];
  if (__sig_should_use_altstack(flags, tib)) {
    __stack_call(ep, code, sig, tib, __sig_unmaskable,
                 tib->tib_sigstack_addr + tib->tib_sigstack_size);
  } else {
    __sig_unmaskable(ep, code, sig, tib);
  }
  // resume running user program
  // hopefully the user fixed the cpu state
  // otherwise the crash will keep happening
  return kNtExceptionContinueExecution;
 }
 static textwindows int __sig_console_sig(uint32_t dwCtrlType) {
  switch (dwCtrlType) {
    case kNtCtrlCEvent:
      return SIGINT;
    case kNtCtrlBreakEvent:
      return SIGQUIT;
    case kNtCtrlCloseEvent:
    case kNtCtrlLogoffEvent:    // only received by services
    case kNtCtrlShutdownEvent:  // only received by services
      return SIGHUP;
    default:
      return SIGSTKFLT;
  }
 }
 __msabi textwindows dontinstrument bool32 __sig_console(uint32_t dwCtrlType) {
  // win32 launches a thread to deliver ctrl-c and ctrl-break when typed
  // it only happens when kNtEnableProcessedInput is in play on console.
  // otherwise we need to wait until read-nt.c discovers that keystroke.
  struct CosmoTib tls;
  __bootstrap_tls(&tls, __builtin_frame_address(0));
  __sig_generate(__sig_console_sig(dwCtrlType), SI_KERNEL);
  return true;
 }
 // returns 0 if no signal handlers were called, otherwise a bitmask
 // consisting of `1` which means a signal handler was invoked which
 // didn't have the SA_RESTART flag, and `2`, which means SA_RESTART
 // handlers were called (or `3` if both were the case).
 textwindows int __sig_check(void) {
  int sig, res = 0;
  while ((sig = __sig_get(atomic_load_explicit(&__get_tls()->tib_sigmask,
                                               memory_order_acquire))))
    res |= __sig_raise(sig, SI_KERNEL);
  return res;
 }
 // background thread for delivering inter-process signals asynchronously
 // this checks for undelivered process-wide signals, once per scheduling
 // quantum, which on windows should be every ~15ms or so, unless somehow
 // the process was tuned to have more fine-grained event timing. we want
 // signals to happen faster when possible; that happens when cancelation
 // points, e.g. read need to wait on i/o; they too check for new signals
 textwindows dontinstrument static uint32_t __sig_worker(void *arg) {
  struct CosmoTib tls;
  __bootstrap_tls(&tls, __builtin_frame_address(0));
  char *sp = __builtin_frame_address(0);
  __maps_track((char *)(((uintptr_t)sp + __pagesize - 1) & -__pagesize) - STKSZ,
               STKSZ);
  for (;;) {
    // dequeue all pending signals and fire them off. if there's no
    // thread that can handle them then __sig_generate will requeue
    // those signals back to __sig.process; hence the need for xchg
    unsigned long sigs =
        atomic_exchange_explicit(__sig.process, 0, memory_order_acq_rel);
    while (sigs) {
      int sig = bsfl(sigs) + 1;
      sigs &= ~(1ull << (sig - 1));
      __sig_generate(sig, SI_KERNEL);
    }
    // unblock stalled asynchronous signals in threads
    _pthread_lock();
    for (struct Dll *e = dll_first(_pthread_list); e;
         e = dll_next(_pthread_list, e)) {
      struct PosixThread *pt = POSIXTHREAD_CONTAINER(e);
      if (atomic_load_explicit(&pt->pt_status, memory_order_acquire) >=
          kPosixThreadTerminated) {
        break;
      }
      sigset_t pending =
          atomic_load_explicit(&pt->tib->tib_sigpending, memory_order_acquire);
      sigset_t mask =
          atomic_load_explicit(&pt->tib->tib_sigmask, memory_order_acquire);
      if (pending & ~mask) {
        _pthread_ref(pt);
        _pthread_unlock();
        while (!atomic_compare_exchange_weak_explicit(
            &pt->tib->tib_sigpending, &pending, pending & ~mask,
            memory_order_acq_rel, memory_order_relaxed)) {
        }
        while ((pending = pending & ~mask)) {
          int sig = bsfl(pending) + 1;
          pending &= ~(1ull << (sig - 1));
          __sig_killer(pt, sig, SI_KERNEL);
        }
        _pthread_lock();
        _pthread_unref(pt);
      }
    }
    _pthread_unlock();
    // wait until next scheduler quantum
    Sleep(POLL_INTERVAL_MS);
  }
  return 0;
 }
 __attribute__((__constructor__(10))) textstartup void __sig_init(void) {
  if (!IsWindows())
    return;
  AddVectoredExceptionHandler(true, (void *)__sig_crash);
  SetConsoleCtrlHandler((void *)__sig_console, true);
  CreateThread(0, STKSZ, __sig_worker, 0, kNtStackSizeParamIsAReservation, 0);
 }
 #endif /* __x86_64__ */
--- a/libc/intrin/sigblock.c
+++ b/libc/intrin/sigblock.c
@ -0,0 +1,53 @@
 /*-*- mode:c;indent-tabs-mode:nil;c-basic-offset:2;tab-width:8;coding:utf-8 -*-│
 │ vi: set et ft=c ts=2 sts=2 sw=2 fenc=utf-8                               :vi │
 ╞══════════════════════════════════════════════════════════════════════════════╡
 │ Copyright 2023 Justine Alexandra Roberts Tunney                              │
 │                                                                              │
 │ Permission to use, copy, modify, and/or distribute this software for         │
 │ any purpose with or without fee is hereby granted, provided that the         │
 │ above copyright notice and this permission notice appear in all copies.      │
 │                                                                              │
 │ THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL                │
 │ WARRANTIES WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED                │
 │ WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE             │
 │ AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL         │
 │ DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR        │
 │ PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER               │
 │ TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR             │
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/sysv/consts/sig.h"
 #include "libc/calls/sig.internal.h"
 #include "libc/calls/struct/sigset.internal.h"
 #include "libc/dce.h"
 #include "libc/intrin/atomic.h"
 #include "libc/intrin/weaken.h"
 #include "libc/thread/tls.h"
 struct Signals __sig;
 sigset_t __sig_block(void) {
  if (IsWindows() || IsMetal()) {
    if (__tls_enabled)
      return atomic_exchange_explicit(&__get_tls()->tib_sigmask, -1,
                                      memory_order_acquire);
    else
      return 0;
  } else {
    sigset_t res, neu = -1;
    sys_sigprocmask(SIG_SETMASK, &neu, &res);
    return res;
  }
 }
 void __sig_unblock(sigset_t m) {
  if (IsWindows() || IsMetal()) {
    if (__tls_enabled) {
      atomic_store_explicit(&__get_tls()->tib_sigmask, m, memory_order_release);
      if (_weaken(__sig_check))
        _weaken(__sig_check)();
    }
  } else {
    sys_sigprocmask(SIG_SETMASK, &m, 0);
  }
 }
--- a/libc/intrin/sigcrashsig.c
+++ b/libc/intrin/sigcrashsig.c
--- a/libc/intrin/swapcontext.S
+++ b/libc/intrin/swapcontext.S
@ -32,7 +32,7 @@
 	.ftrace1
 swapcontext:
 	.ftrace2
-#include "libc/calls/getcontext.inc"
+#include "libc/intrin/getcontext.inc"
 #ifdef __x86_64__
 	push	%rbp
 	mov	%rsp,%rbp
--- a/libc/intrin/tailcontext.S
+++ b/libc/intrin/tailcontext.S
--- a/libc/intrin/timespec_add.c
+++ b/libc/intrin/timespec_add.c
--- a/libc/intrin/timespec_cmp.c
+++ b/libc/intrin/timespec_cmp.c
--- a/libc/intrin/timespec_frommicros.c
+++ b/libc/intrin/timespec_frommicros.c
--- a/libc/intrin/timespec_frommillis.c
+++ b/libc/intrin/timespec_frommillis.c
--- a/libc/intrin/timespec_fromnanos.c
+++ b/libc/intrin/timespec_fromnanos.c
--- a/libc/intrin/timespec_sub.c
+++ b/libc/intrin/timespec_sub.c
--- a/libc/intrin/timespec_subz.c
+++ b/libc/intrin/timespec_subz.c
--- a/libc/intrin/timespec_tomicros.c
+++ b/libc/intrin/timespec_tomicros.c
--- a/libc/intrin/timespec_tomillis.c
+++ b/libc/intrin/timespec_tomillis.c
--- a/libc/intrin/timespec_tonanos.c
+++ b/libc/intrin/timespec_tonanos.c
--- a/libc/intrin/timespec_totimeval.c
+++ b/libc/intrin/timespec_totimeval.c
--- a/libc/intrin/timespectowindowstime.c
+++ b/libc/intrin/timespectowindowstime.c
--- a/libc/intrin/timeval_add.c
+++ b/libc/intrin/timeval_add.c
--- a/libc/intrin/timeval_cmp.c
+++ b/libc/intrin/timeval_cmp.c
--- a/libc/intrin/timeval_frommicros.c
+++ b/libc/intrin/timeval_frommicros.c
--- a/libc/intrin/timeval_frommillis.c
+++ b/libc/intrin/timeval_frommillis.c
--- a/libc/intrin/timeval_sub.c
+++ b/libc/intrin/timeval_sub.c
--- a/libc/intrin/timeval_subz.c
+++ b/libc/intrin/timeval_subz.c
--- a/libc/intrin/timeval_tomicros.c
+++ b/libc/intrin/timeval_tomicros.c
--- a/libc/intrin/timeval_tomillis.c
+++ b/libc/intrin/timeval_tomillis.c
--- a/libc/intrin/timeval_toseconds.c
+++ b/libc/intrin/timeval_toseconds.c
--- a/libc/intrin/timevaltowindowstime.c
+++ b/libc/intrin/timevaltowindowstime.c
--- a/libc/intrin/ucontext.c
+++ b/libc/intrin/ucontext.c
--- a/libc/calls/vdsofunc.greg.c
+++ b/libc/calls/vdsofunc.greg.c
--- a/libc/intrin/windowsdurationtotimespec.c
+++ b/libc/intrin/windowsdurationtotimespec.c
--- a/libc/intrin/windowsdurationtotimeval.c
+++ b/libc/intrin/windowsdurationtotimeval.c
--- a/libc/intrin/windowstimetotimespec.c
+++ b/libc/intrin/windowstimetotimespec.c
--- a/libc/intrin/windowstimetotimeval.c
+++ b/libc/intrin/windowstimetotimeval.c
--- a/libc/str/BUILD.mk
+++ b/libc/str/BUILD.mk
@ -77,13 +77,7 @@ o/$(MODE)/libc/str/iswseparator.o: private			\
 # ensure that division is optimized
 o/$(MODE)/libc/str/bcmp.o					\
-o/$(MODE)/libc/str/strcmp.o					\
+o/$(MODE)/libc/str/strcmp.o: private				\
 o/$(MODE)/libc/str/windowsdurationtotimeval.o			\
 o/$(MODE)/libc/str/windowsdurationtotimespec.o			\
 o/$(MODE)/libc/str/timevaltowindowstime.o			\
 o/$(MODE)/libc/str/timespectowindowstime.o			\
 o/$(MODE)/libc/str/windowstimetotimeval.o			\
 o/$(MODE)/libc/str/windowstimetotimespec.o: private		\
 		CFLAGS +=					\
 			-O2
--- a/libc/thread/pthread_barrier_wait.c
+++ b/libc/thread/pthread_barrier_wait.c
@ -17,11 +17,11 @@
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/calls/blockcancel.internal.h"
 #include "libc/cosmo.h"
 #include "libc/errno.h"
 #include "libc/intrin/atomic.h"
 #include "libc/limits.h"
 #include "libc/thread/thread.h"
 #include "third_party/nsync/futex.internal.h"
 /**
 * Waits for all threads to arrive at barrier.
@ -54,14 +54,14 @@ errno_t pthread_barrier_wait(pthread_barrier_t *barrier) {
    atomic_store_explicit(&barrier->_counter, barrier->_count,
                          memory_order_release);
    atomic_store_explicit(&barrier->_waiters, 0, memory_order_release);
-    nsync_futex_wake_(&barrier->_waiters, INT_MAX, barrier->_pshared);
+    cosmo_futex_wake(&barrier->_waiters, INT_MAX, barrier->_pshared);
    return PTHREAD_BARRIER_SERIAL_THREAD;
  }
  // wait for everyone else to arrive at barrier
  BLOCK_CANCELATION;
  while ((n = atomic_load_explicit(&barrier->_waiters, memory_order_acquire)))
-    nsync_futex_wait_(&barrier->_waiters, n, barrier->_pshared, 0, 0);
+    cosmo_futex_wait(&barrier->_waiters, n, barrier->_pshared, 0, 0);
  ALLOW_CANCELATION;
  return 0;
--- a/libc/thread/pthread_cond_broadcast.c
+++ b/libc/thread/pthread_cond_broadcast.c
@ -16,12 +16,12 @@
 │ TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR             │
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/cosmo.h"
 #include "libc/dce.h"
 #include "libc/intrin/atomic.h"
 #include "libc/limits.h"
 #include "libc/thread/thread.h"
 #include "third_party/nsync/cv.h"
 #include "third_party/nsync/futex.internal.h"
 __static_yoink("nsync_mu_lock");
 __static_yoink("nsync_mu_unlock");
@ -63,6 +63,6 @@ errno_t pthread_cond_broadcast(pthread_cond_t *cond) {
  // roll forward the monotonic sequence
  atomic_fetch_add_explicit(&cond->_sequence, 1, memory_order_acq_rel);
  if (atomic_load_explicit(&cond->_waiters, memory_order_acquire))
-    nsync_futex_wake_((atomic_int *)&cond->_sequence, INT_MAX, cond->_pshared);
+    cosmo_futex_wake((atomic_int *)&cond->_sequence, INT_MAX, cond->_pshared);
  return 0;
 }
--- a/libc/thread/pthread_cond_signal.c
+++ b/libc/thread/pthread_cond_signal.c
@ -16,11 +16,11 @@
 │ TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR             │
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/cosmo.h"
 #include "libc/dce.h"
 #include "libc/intrin/atomic.h"
 #include "libc/thread/thread.h"
 #include "third_party/nsync/cv.h"
 #include "third_party/nsync/futex.internal.h"
 __static_yoink("nsync_mu_lock");
 __static_yoink("nsync_mu_unlock");
@ -62,6 +62,6 @@ errno_t pthread_cond_signal(pthread_cond_t *cond) {
  // roll forward the monotonic sequence
  atomic_fetch_add_explicit(&cond->_sequence, 1, memory_order_acq_rel);
  if (atomic_load_explicit(&cond->_waiters, memory_order_acquire))
-    nsync_futex_wake_((atomic_int *)&cond->_sequence, 1, cond->_pshared);
+    cosmo_futex_wake((atomic_int *)&cond->_sequence, 1, cond->_pshared);
  return 0;
 }
--- a/libc/thread/pthread_cond_timedwait.c
+++ b/libc/thread/pthread_cond_timedwait.c
@ -18,6 +18,8 @@
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/calls/calls.h"
 #include "libc/calls/cp.internal.h"
 #include "libc/calls/struct/timespec.h"
 #include "libc/cosmo.h"
 #include "libc/dce.h"
 #include "libc/errno.h"
 #include "libc/intrin/atomic.h"
@ -28,7 +30,6 @@
 #include "libc/thread/thread2.h"
 #include "third_party/nsync/common.internal.h"
 #include "third_party/nsync/cv.h"
 #include "third_party/nsync/futex.internal.h"
 #include "third_party/nsync/time.h"
 __static_yoink("nsync_mu_lock");
@ -74,7 +75,7 @@ static errno_t pthread_cond_timedwait_impl(pthread_cond_t *cond,
  int rc;
  struct PthreadWait waiter = {cond, mutex};
  pthread_cleanup_push(pthread_cond_leave, &waiter);
-  rc = nsync_futex_wait_((atomic_int *)&cond->_sequence, seq1, cond->_pshared,
+  rc = cosmo_futex_wait((atomic_int *)&cond->_sequence, seq1, cond->_pshared,
                        cond->_clock, abstime);
  pthread_cleanup_pop(true);
  if (rc == -EAGAIN)
--- a/libc/thread/pthread_exit.c
+++ b/libc/thread/pthread_exit.c
@ -18,6 +18,7 @@
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/assert.h"
 #include "libc/atomic.h"
 #include "libc/cosmo.h"
 #include "libc/cxxabi.h"
 #include "libc/dce.h"
 #include "libc/intrin/atomic.h"
@ -33,7 +34,6 @@
 #include "libc/thread/posixthread.internal.h"
 #include "libc/thread/thread.h"
 #include "libc/thread/tls.h"
 #include "third_party/nsync/futex.internal.h"
 #include "third_party/nsync/wait_s.internal.h"
 /**
@ -137,7 +137,7 @@ wontreturn void pthread_exit(void *rc) {
  // note that the main thread is joinable by child threads
  if (pt->pt_flags & PT_STATIC) {
    atomic_store_explicit(&tib->tib_tid, 0, memory_order_release);
-    nsync_futex_wake_((atomic_int *)&tib->tib_tid, INT_MAX,
+    cosmo_futex_wake((atomic_int *)&tib->tib_tid, INT_MAX,
                     !IsWindows() && !IsXnu());
    _Exit1(0);
  }
--- a/libc/thread/pthread_timedjoin_np.c
+++ b/libc/thread/pthread_timedjoin_np.c
@ -20,6 +20,8 @@
 #include "libc/calls/cp.internal.h"
 #include "libc/calls/struct/timespec.h"
 #include "libc/calls/struct/timespec.internal.h"
 #include "libc/cosmo.h"
 #include "libc/dce.h"
 #include "libc/errno.h"
 #include "libc/fmt/itoa.h"
 #include "libc/intrin/atomic.h"
@ -30,7 +32,6 @@
 #include "libc/thread/posixthread.internal.h"
 #include "libc/thread/thread2.h"
 #include "libc/thread/tls.h"
 #include "third_party/nsync/futex.internal.h"
 static const char *DescribeReturnValue(char buf[30], int err, void **value) {
  char *p = buf;
@ -75,7 +76,7 @@ static errno_t _pthread_wait(atomic_int *ctid, struct timespec *abstime) {
    if (!(err = pthread_testcancel_np())) {
      BEGIN_CANCELATION_POINT;
      while ((x = atomic_load_explicit(ctid, memory_order_acquire))) {
-        e = nsync_futex_wait_(ctid, x, !IsWindows() && !IsXnu(), CLOCK_REALTIME,
+        e = cosmo_futex_wait(ctid, x, !IsWindows() && !IsXnu(), CLOCK_REALTIME,
                             abstime);
        if (e == -ECANCELED) {
          err = ECANCELED;
--- a/libc/thread/sem_post.c
+++ b/libc/thread/sem_post.c
@ -19,6 +19,7 @@
 #include "libc/assert.h"
 #include "libc/calls/calls.h"
 #include "libc/calls/syscall-sysv.internal.h"
 #include "libc/cosmo.h"
 #include "libc/dce.h"
 #include "libc/errno.h"
 #include "libc/intrin/atomic.h"
@ -26,7 +27,6 @@
 #include "libc/runtime/syslib.internal.h"
 #include "libc/sysv/errfuns.h"
 #include "libc/thread/semaphore.h"
 #include "third_party/nsync/futex.internal.h"
 /**
 * Unlocks semaphore.
@ -46,7 +46,7 @@ int sem_post(sem_t *sem) {
  old = atomic_fetch_add_explicit(&sem->sem_value, 1, memory_order_acq_rel);
  unassert(old > INT_MIN);
  if (old >= 0) {
-    wakeups = nsync_futex_wake_(&sem->sem_value, 1, sem->sem_pshared);
+    wakeups = cosmo_futex_wake(&sem->sem_value, 1, sem->sem_pshared);
    npassert(wakeups >= 0);
    rc = 0;
  } else {
--- a/libc/thread/sem_timedwait.c
+++ b/libc/thread/sem_timedwait.c
@ -22,6 +22,7 @@
 #include "libc/calls/struct/timespec.h"
 #include "libc/calls/struct/timespec.internal.h"
 #include "libc/calls/syscall-sysv.internal.h"
 #include "libc/cosmo.h"
 #include "libc/dce.h"
 #include "libc/errno.h"
 #include "libc/intrin/atomic.h"
@ -32,7 +33,6 @@
 #include "libc/sysv/errfuns.h"
 #include "libc/thread/semaphore.h"
 #include "libc/thread/thread.h"
 #include "third_party/nsync/futex.internal.h"
 static void sem_delay(int n) {
  volatile int i;
@ -119,7 +119,7 @@ int sem_timedwait(sem_t *sem, const struct timespec *abstime) {
  do {
    if (!(v = atomic_load_explicit(&sem->sem_value, memory_order_relaxed))) {
-      rc = nsync_futex_wait_(&sem->sem_value, v, sem->sem_pshared,
+      rc = cosmo_futex_wait(&sem->sem_value, v, sem->sem_pshared,
                            CLOCK_REALTIME, abstime);
      if (rc == -EINTR || rc == -ECANCELED) {
        errno = -rc;
--- a/test/libc/thread/footek_test.c
+++ b/test/libc/thread/footek_test.c
@ -9,8 +9,6 @@
 #ifdef __COSMOPOLITAN__
 #include <cosmo.h>
 #include "libc/thread/thread.h"
 #include "third_party/nsync/futex.internal.h"
 #endif
 #include <assert.h>
--- a/third_party/lua/lunix.c
+++ b/third_party/lua/lunix.c
@ -109,8 +109,9 @@
 #include "third_party/lua/lgc.h"
 #include "third_party/lua/lua.h"
 #include "third_party/lua/luaconf.h"
 #include "third_party/nsync/futex.internal.h"
 #include "libc/sysv/consts/clock.h"
 #include "libc/cosmo.h"
 #include "libc/cosmo.h"
 #include "tool/net/luacheck.h"
 #define DNS_NAME_MAX  253
@ -2855,7 +2856,7 @@ static int LuaUnixMemoryWait(lua_State *L) {
    deadline = &ts;
  }
  BEGIN_CANCELATION_POINT;
-  rc = nsync_futex_wait_((atomic_int *)GetWord(L), expect,
+  rc = cosmo_futex_wait((atomic_int *)GetWord(L), expect,
                         PTHREAD_PROCESS_SHARED, CLOCK_REALTIME, deadline);
  END_CANCELATION_POINT;
  if (rc < 0) errno = -rc, rc = -1;
@ -2867,7 +2868,7 @@ static int LuaUnixMemoryWait(lua_State *L) {
 static int LuaUnixMemoryWake(lua_State *L) {
  int count, woken;
  count = luaL_optinteger(L, 3, INT_MAX);
-  woken = nsync_futex_wake_((atomic_int *)GetWord(L), count,
+  woken = cosmo_futex_wake((atomic_int *)GetWord(L), count,
                            PTHREAD_PROCESS_SHARED);
  npassert(woken >= 0);
  return ReturnInteger(L, woken);
--- a/third_party/nsync/BUILD.mk
+++ b/third_party/nsync/BUILD.mk
@ -27,7 +27,6 @@ THIRD_PARTY_NSYNC_A_DIRECTDEPS =			\
 	LIBC_INTRIN					\
 	LIBC_NEXGEN32E					\
 	LIBC_NT_KERNEL32				\
 	LIBC_NT_SYNCHRONIZATION				\
 	LIBC_STR					\
 	LIBC_SYSV					\
 	LIBC_SYSV_CALLS
--- a/third_party/nsync/futex.internal.h
+++ b/third_party/nsync/futex.internal.h
@ -1,17 +0,0 @@
 #ifndef NSYNC_FUTEX_INTERNAL_H_
 #define NSYNC_FUTEX_INTERNAL_H_
 #include "libc/calls/struct/timespec.h"
 #include "libc/dce.h"
 COSMOPOLITAN_C_START_
 #ifndef __cplusplus
 #define _FUTEX_ATOMIC(x) _Atomic(x)
 #else
 #define _FUTEX_ATOMIC(x) x
 #endif
 int nsync_futex_wake_(_FUTEX_ATOMIC(int) *, int, char);
 int nsync_futex_wait_(_FUTEX_ATOMIC(int) *, int, char, int, const struct timespec *);
 COSMOPOLITAN_C_END_
 #endif /* NSYNC_FUTEX_INTERNAL_H_ */
--- a/third_party/nsync/mu_semaphore_futex.c
+++ b/third_party/nsync/mu_semaphore_futex.c
@ -21,7 +21,9 @@
 #include "libc/thread/thread.h"
 #include "third_party/nsync/atomic.h"
 #include "third_party/nsync/atomic.internal.h"
-#include "third_party/nsync/futex.internal.h"
+#include "libc/cosmo.h"
 #include "libc/calls/struct/timespec.h"
 #include "libc/cosmo.h"
 #include "third_party/nsync/mu_semaphore.internal.h"
 /**
@ -61,7 +63,7 @@ errno_t nsync_mu_semaphore_p_futex (nsync_semaphore *s) {
 		i = ATM_LOAD ((nsync_atomic_uint32_ *) &f->i);
 		if (i == 0) {
 			int futex_result;
-			futex_result = -nsync_futex_wait_ (
+			futex_result = -cosmo_futex_wait (
 				(atomic_int *)&f->i, i,
 				PTHREAD_PROCESS_PRIVATE, 0, 0);
 			ASSERT (futex_result == 0 ||
@ -100,7 +102,7 @@ errno_t nsync_mu_semaphore_p_with_deadline_futex (nsync_semaphore *s, int clock,
 				ts_buf.tv_nsec = NSYNC_TIME_NSEC (abs_deadline);
 				ts = &ts_buf;
 			}
-			futex_result = nsync_futex_wait_ ((atomic_int *)&f->i, i,
+			futex_result = cosmo_futex_wait ((atomic_int *)&f->i, i,
 							 PTHREAD_PROCESS_PRIVATE,
 							 clock, ts);
 			ASSERT (futex_result == 0 ||
@ -136,5 +138,5 @@ void nsync_mu_semaphore_v_futex (nsync_semaphore *s) {
 		       (nsync_atomic_uint32_ *) &f->i, &old_value, old_value+1,
 		       memory_order_release, memory_order_relaxed)) {
 	}
-	ASSERT (nsync_futex_wake_ ((atomic_int *)&f->i, 1, PTHREAD_PROCESS_PRIVATE) >= 0);
+	ASSERT (cosmo_futex_wake ((atomic_int *)&f->i, 1, PTHREAD_PROCESS_PRIVATE) >= 0);
 }
--- a/third_party/openmp/kmp_lock.cpp
+++ b/third_party/openmp/kmp_lock.cpp
@ -23,7 +23,7 @@
 #if KMP_USE_FUTEX
 #ifdef __COSMOPOLITAN__
-#include "third_party/nsync/futex.internal.h"
+#include <cosmo.h>
 #else
 #include <sys/syscall.h>
 #include <unistd.h>
@ -380,7 +380,7 @@ __kmp_acquire_futex_lock_timed_template(kmp_futex_lock_t *lck, kmp_int32 gtid) {
    long rc;
 #ifdef __COSMOPOLITAN__
-    if ((rc = nsync_futex_wait_((int *)&(lck->lk.poll), poll_val, false, 0, NULL)) != 0) {
+    if ((rc = cosmo_futex_wait((int *)&(lck->lk.poll), poll_val, false, 0, NULL)) != 0) {
 #else
    if ((rc = syscall(__NR_futex, (int *)&(lck->lk.poll), FUTEX_WAIT, poll_val, NULL,
                      NULL, 0)) != 0) {
@ -462,7 +462,7 @@ int __kmp_release_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
             ("__kmp_release_futex_lock: lck:%p, T#%d futex_wake 1 thread\n",
              lck, gtid));
 #ifdef __COSMOPOLITAN__
-    nsync_futex_wake_((int *)&(lck->lk.poll), 1, false);
+    cosmo_futex_wake((int *)&(lck->lk.poll), 1, false);
 #else
    syscall(__NR_futex, &(lck->lk.poll), FUTEX_WAKE, KMP_LOCK_BUSY(1, futex),
            NULL, NULL, 0);