diff --git a/libc/calls/getcwd.greg.c b/libc/calls/getcwd.greg.c
index f7bc1801fbb..e68cd53c609 100644
--- a/libc/calls/getcwd.greg.c
+++ b/libc/calls/getcwd.greg.c
@@ -164,6 +164,6 @@ int __getcwd(char *buf, size_t size) {
   } else {
     rc = sys_getcwd_metal(buf, size);
   }
-  STRACE("__getcwd([%#hhs], %'zu) → %d% m", rc != -1 ? buf : "n/a", size, rc);
+  STRACE("getcwd([%#hhs], %'zu) → %d% m", rc != -1 ? buf : "n/a", size, rc);
   return rc;
 }
diff --git a/libc/calls/ntspawn.c b/libc/calls/ntspawn.c
index 26014b92da5..e6cb1b4058b 100644
--- a/libc/calls/ntspawn.c
+++ b/libc/calls/ntspawn.c
@@ -126,7 +126,8 @@ textwindows int ntspawn(
           if (CreateProcess(sb->path, sb->cmdline, 0, 0, true,
                             dwCreationFlags | kNtCreateUnicodeEnvironment |
                                 kNtExtendedStartupinfoPresent |
-                                kNtInheritParentAffinity,
+                                kNtInheritParentAffinity |
+                                GetPriorityClass(GetCurrentProcess()),
                             sb->envblock, opt_lpCurrentDirectory,
                             &info.StartupInfo, opt_out_lpProcessInformation)) {
             rc = 0;
diff --git a/libc/calls/park.c b/libc/calls/park.c
index 5b4890de870..e6fa668ee7b 100644
--- a/libc/calls/park.c
+++ b/libc/calls/park.c
@@ -32,11 +32,7 @@ static textwindows int _park_thread(uint32_t msdelay, sigset_t waitmask,
                                     bool restartable) {
   int sig;
   if (_check_cancel() == -1) return -1;
-  if ((sig = __sig_get(waitmask))) {
-    int handler_was_called = __sig_relay(sig, SI_KERNEL, waitmask);
-    if (_check_cancel() == -1) return -1;
-    if (!restartable || handler_was_called == 1) return eintr();
-  }
+  if ((sig = __sig_get(waitmask))) goto HandleSignal;
   int expect = 0;
   atomic_int futex = 0;
   struct PosixThread *pt = _pthread_self();
@@ -45,9 +41,12 @@ static textwindows int _park_thread(uint32_t msdelay, sigset_t waitmask,
   bool32 ok = WaitOnAddress(&futex, &expect, sizeof(int), msdelay);
   atomic_store_explicit(&pt->pt_blocker, 0, memory_order_release);
   if (ok && (sig = __sig_get(waitmask))) {
+  HandleSignal:
     int handler_was_called = __sig_relay(sig, SI_KERNEL, waitmask);
     if (_check_cancel() == -1) return -1;
-    if (!restartable || handler_was_called == 1) return eintr();
+    if (!restartable || (handler_was_called & SIG_HANDLED_NO_RESTART)) {
+      return eintr();
+    }
   }
   return 0;
 }
diff --git a/libc/calls/poll-nt.c b/libc/calls/poll-nt.c
index 3e37c8da446..d5c63ada755 100644
--- a/libc/calls/poll-nt.c
+++ b/libc/calls/poll-nt.c
@@ -55,6 +55,8 @@
 #include "libc/thread/tls.h"
 #ifdef __x86_64__
 
+#define POLL_INTERVAL_MS 10
+
 // Polls on the New Technology.
 //
 // This function is used to implement poll() and select(). You may poll
@@ -184,7 +186,7 @@ static textwindows int sys_poll_nt_impl(struct pollfd *fds, uint64_t nfds,
         remain = timespec_sub(deadline, now);
         millis = timespec_tomillis(remain);
         waitfor = MIN(millis, 0xffffffffu);
-        waitfor = MIN(waitfor, __SIG_POLL_INTERVAL_MS);
+        waitfor = MIN(waitfor, POLL_INTERVAL_MS);
         if (waitfor) {
           POLLTRACE("poll() sleeping for %'d out of %'lu ms", waitfor,
                     timespec_tomillis(remain));
diff --git a/libc/calls/read-nt.c b/libc/calls/read-nt.c
index cb3e1eadd2e..63d7dc4463b 100644
--- a/libc/calls/read-nt.c
+++ b/libc/calls/read-nt.c
@@ -723,8 +723,6 @@ static textwindows int WaitForConsole(struct Fd *f, sigset_t waitmask) {
   int sig;
   int64_t sem;
   uint32_t wi, ms = -1;
-  int handler_was_called;
-  struct PosixThread *pt;
   if (!__ttyconf.vmin) {
     if (!__ttyconf.vtime) {
       return 0;  // non-blocking w/o raising eagain
@@ -732,31 +730,24 @@ static textwindows int WaitForConsole(struct Fd *f, sigset_t waitmask) {
       ms = __ttyconf.vtime * 100;
     }
   }
-  if (f->flags & _O_NONBLOCK) {
-    return eagain();  // standard unix non-blocking
-  }
   if (_check_cancel() == -1) return -1;
-  if ((sig = __sig_get(waitmask))) {
-    handler_was_called = __sig_relay(sig, SI_KERNEL, waitmask);
-    if (_check_cancel() == -1) return -1;
-    if (handler_was_called != 1) return -2;
-    return eintr();
-  }
-  pt = _pthread_self();
-  pt->pt_semaphore = sem = CreateSemaphore(0, 0, 1, 0);
-  pthread_cleanup_push((void *)CloseHandle, (void *)sem);
+  if (f->flags & _O_NONBLOCK) return eagain();
+  if ((sig = __sig_get(waitmask))) goto DeliverSignal;
+  struct PosixThread *pt = _pthread_self();
   pt->pt_blkmask = waitmask;
+  pt->pt_semaphore = sem = CreateSemaphore(0, 0, 1, 0);
   atomic_store_explicit(&pt->pt_blocker, PT_BLOCKER_SEM, memory_order_release);
   wi = WaitForMultipleObjects(2, (int64_t[2]){__keystroke.cin, sem}, 0, ms);
   atomic_store_explicit(&pt->pt_blocker, 0, memory_order_release);
-  pthread_cleanup_pop(true);
+  CloseHandle(sem);
   if (wi == kNtWaitTimeout) return 0;  // vtime elapsed
   if (wi == 0) return -2;              // console data
   if (wi != 1) return __winerr();      // wait failed
   if (!(sig = __sig_get(waitmask))) return eintr();
-  handler_was_called = __sig_relay(sig, SI_KERNEL, waitmask);
+DeliverSignal:
+  int handler_was_called = __sig_relay(sig, SI_KERNEL, waitmask);
   if (_check_cancel() == -1) return -1;
-  if (handler_was_called != 1) return -2;
+  if (!(handler_was_called & SIG_HANDLED_NO_RESTART)) return -2;
   return eintr();
 }
 
diff --git a/libc/calls/readwrite-nt.c b/libc/calls/readwrite-nt.c
index 266dc54cf39..095503241db 100644
--- a/libc/calls/readwrite-nt.c
+++ b/libc/calls/readwrite-nt.c
@@ -16,15 +16,12 @@
 │ TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR             │
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
-#include "libc/calls/calls.h"
 #include "libc/calls/createfileflags.internal.h"
 #include "libc/calls/internal.h"
 #include "libc/calls/sig.internal.h"
 #include "libc/calls/struct/fd.internal.h"
-#include "libc/calls/struct/sigset.internal.h"
+#include "libc/calls/struct/sigset.h"
 #include "libc/calls/syscall_support-nt.internal.h"
-#include "libc/errno.h"
-#include "libc/intrin/atomic.h"
 #include "libc/nt/enum/filetype.h"
 #include "libc/nt/errors.h"
 #include "libc/nt/events.h"
@@ -34,28 +31,10 @@
 #include "libc/nt/synchronization.h"
 #include "libc/nt/thread.h"
 #include "libc/stdio/sysparam.h"
-#include "libc/str/str.h"
-#include "libc/sysv/consts/o.h"
 #include "libc/sysv/consts/sicode.h"
 #include "libc/sysv/errfuns.h"
-#include "libc/thread/posixthread.internal.h"
-#include "libc/thread/thread.h"
-#include "libc/thread/tls.h"
 #ifdef __x86_64__
 
-struct ReadwriteResources {
-  int64_t handle;
-  struct NtOverlapped *overlap;
-};
-
-static void UnwindReadwrite(void *arg) {
-  uint32_t got;
-  struct ReadwriteResources *rwc = arg;
-  CancelIoEx(rwc->handle, rwc->overlap);
-  GetOverlappedResult(rwc->handle, rwc->overlap, &got, true);
-  CloseHandle(rwc->overlap->hEvent);
-}
-
 /**
  * Runs code that's common to read/write/pread/pwrite/etc on Windows.
  *
@@ -64,17 +43,11 @@ static void UnwindReadwrite(void *arg) {
  */
 textwindows ssize_t
 sys_readwrite_nt(int fd, void *data, size_t size, ssize_t offset,
-                 int64_t handle, uint64_t waitmask,
+                 int64_t handle, sigset_t waitmask,
                  bool32 ReadOrWriteFile(int64_t, void *, uint32_t, uint32_t *,
                                         struct NtOverlapped *)) {
-  bool32 ok;
-  int sig = 0;
+  int sig;
   uint32_t exchanged;
-  int olderror = errno;
-  bool eagained = false;
-  bool canceled = false;
-  int handler_was_called;
-  struct PosixThread *pt;
   struct Fd *f = g_fds.p + fd;
 
   // win32 i/o apis generally take 32-bit values thus we implicitly
@@ -106,33 +79,26 @@ sys_readwrite_nt(int fd, void *data, size_t size, ssize_t offset,
   }
 
 RestartOperation:
+  bool eagained = false;
+  // check for signals and cancelation
+  if (_check_cancel() == -1) return -1;  // ECANCELED
+  if ((sig = __sig_get(waitmask))) goto HandleInterrupt;
+
   // signals have already been fully blocked by caller
   // perform i/o operation with atomic signal/cancel checking
   struct NtOverlapped overlap = {.hEvent = CreateEvent(0, 1, 0, 0),
                                  .Pointer = offset};
-  struct ReadwriteResources rwc = {handle, &overlap};
-  pthread_cleanup_push(UnwindReadwrite, &rwc);
-  ok = ReadOrWriteFile(handle, data, size, 0, &overlap);
+  bool32 ok = ReadOrWriteFile(handle, data, size, 0, &overlap);
   if (!ok && GetLastError() == kNtErrorIoPending) {
     // win32 says this i/o operation needs to block
     if (f->flags & _O_NONBLOCK) {
       // abort the i/o operation if file descriptor is in non-blocking mode
       CancelIoEx(handle, &overlap);
       eagained = true;
-    } else if (_check_cancel()) {
-      // _check_cancel() can go three ways:
-      // 1. it'll return 0 if we're fine and no thread cancelation happened
-      // 2. it'll pthread_exit() and cleanup, when cancelation was deferred
-      // 3. it'll return -1 and raise ECANCELED if a cancelation was masked
-      CancelIoEx(handle, &overlap);
-      canceled = true;
-    } else if ((sig = __sig_get(waitmask))) {
-      // we've dequeued a signal that was pending per caller's old sigmask
-      // we can't call the signal handler until we release win32 resources
-      CancelIoEx(handle, &overlap);
     } else {
       // wait until i/o either completes or is canceled by another thread
       // we avoid a race condition by having a second mask for unblocking
+      struct PosixThread *pt;
       pt = _pthread_self();
       pt->pt_blkmask = waitmask;
       pt->pt_iohandle = handle;
@@ -147,30 +113,13 @@ sys_readwrite_nt(int fd, void *data, size_t size, ssize_t offset,
   if (ok) {
     ok = GetOverlappedResult(handle, &overlap, &exchanged, true);
   }
-  pthread_cleanup_pop(false);
   CloseHandle(overlap.hEvent);
 
-  // if we acknowledged a pending masked mode cancelation request then
-  // we must pass it to the caller immediately now that cleanup's done
-  if (canceled) {
-    return ecanceled();
-  }
-
-  // if we removed a pending signal then we must raise it
-  // it's now safe to call a signal handler that longjmps
-  if (sig) {
-    handler_was_called = __sig_relay(sig, SI_KERNEL, waitmask);
-    if (_check_cancel() == -1) return -1;
-  } else {
-    handler_was_called = 0;
-  }
-
   // if i/o succeeded then return its result
   if (ok) {
     if (!pwriting && seekable) {
       f->pointer = offset + exchanged;
     }
-    errno = olderror;
     return exchanged;
   }
 
@@ -180,17 +129,15 @@ sys_readwrite_nt(int fd, void *data, size_t size, ssize_t offset,
     if (eagained) {
       return eagain();
     }
-    // at this point the i/o must have been canceled due to a signal.
-    // this could be because we found the signal earlier and canceled
-    // ourself. otherwise it's due to a kill from another thread that
-    // added something to our mask and canceled our i/o, so we check.
-    if (!handler_was_called && (sig = __sig_get(waitmask))) {
-      handler_was_called = __sig_relay(sig, SI_KERNEL, waitmask);
-      if (_check_cancel() == -1) return -1;
-    }
-    // read() is @restartable unless non-SA_RESTART hands were called
-    if (handler_was_called != 1) {
-      goto RestartOperation;
+    // otherwise it must be due to a kill() via __sig_cancel()
+    if ((sig = __sig_get(waitmask))) {
+    HandleInterrupt:
+      int handler_was_called = __sig_relay(sig, SI_KERNEL, waitmask);
+      if (_check_cancel() == -1) return -1;  // possible if we SIGTHR'd
+      // read() is @restartable unless non-SA_RESTART hands were called
+      if (!(handler_was_called & SIG_HANDLED_NO_RESTART)) {
+        goto RestartOperation;
+      }
     }
     return eintr();
   }
diff --git a/libc/calls/sig.c b/libc/calls/sig.c
index 467c19939c9..8c620ee93ab 100644
--- a/libc/calls/sig.c
+++ b/libc/calls/sig.c
@@ -18,8 +18,6 @@
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/sysv/consts/sig.h"
 #include "ape/sections.internal.h"
-#include "libc/assert.h"
-#include "libc/atomic.h"
 #include "libc/calls/calls.h"
 #include "libc/calls/sig.internal.h"
 #include "libc/calls/state.internal.h"
@@ -30,34 +28,26 @@
 #include "libc/calls/ucontext.h"
 #include "libc/dce.h"
 #include "libc/errno.h"
-#include "libc/fmt/itoa.h"
 #include "libc/intrin/atomic.h"
 #include "libc/intrin/bsf.h"
 #include "libc/intrin/describebacktrace.internal.h"
-#include "libc/intrin/kprintf.h"
-#include "libc/intrin/popcnt.h"
+#include "libc/intrin/dll.h"
 #include "libc/intrin/strace.internal.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/signal.h"
 #include "libc/nt/enum/status.h"
-#include "libc/nt/errors.h"
 #include "libc/nt/runtime.h"
 #include "libc/nt/signals.h"
-#include "libc/nt/struct/context.h"
 #include "libc/nt/struct/ntexceptionpointers.h"
 #include "libc/nt/synchronization.h"
 #include "libc/nt/thread.h"
-#include "libc/runtime/runtime.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"
-#include "libc/thread/thread.h"
-#include "libc/thread/tls.h"
 #ifdef __x86_64__
 
 /**
@@ -206,16 +196,19 @@ textwindows int __sig_raise(volatile int sig, int sic) {
                                               memory_order_acquire);
 
     // call the user's signal handler
-    char ssbuf[2][128];
+    char ssbuf[128];
     siginfo_t si = {.si_signo = sig, .si_code = sic};
-    STRACE("__sig_raise(%G, %t) mask %s → %s", sig, __sig_handler(rva),
-           (DescribeSigset)(ssbuf[0], 0, &ctx.uc_sigmask),
-           (DescribeSigset)(ssbuf[1], 0, (sigset_t *)&pt->tib->tib_sigmask));
+    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);
     (void)ssbuf;
 
     // record this handler
-    handler_was_called |= (flags & SA_RESTART) ? 2 : 1;
+    if (flags & SA_RESTART) {
+      handler_was_called |= SIG_HANDLED_SA_RESTART;
+    } else {
+      handler_was_called |= SIG_HANDLED_NO_RESTART;
+    }
   }
 
   // restore sigmask
@@ -320,12 +313,18 @@ static int __sig_killer(struct PosixThread *pt, int sig, int sic) {
     __sig_terminate(sig);
   }
 
+  // ignore signals already pending
+  uintptr_t th = _pthread_syshand(pt);
+  if (atomic_load_explicit(&pt->tib->tib_sigpending, memory_order_acquire) &
+      (1ull << (sig - 1))) {
+    return 0;
+  }
+
   // take control of thread
   // suspending the thread happens asynchronously
   // however getting the context blocks until it's frozen
   static pthread_spinlock_t killer_lock;
   pthread_spin_lock(&killer_lock);
-  uintptr_t th = _pthread_syshand(pt);
   if (SuspendThread(th) == -1u) {
     STRACE("SuspendThread failed w/ %d", GetLastError());
     pthread_spin_unlock(&killer_lock);
@@ -414,6 +413,10 @@ textwindows void __sig_generate(int sig, int sic) {
     STRACE("terminating on %G due to no handler", sig);
     __sig_terminate(sig);
   }
+  if (atomic_load_explicit(&__sig.pending, memory_order_acquire) &
+      (1ull << (sig - 1))) {
+    return;
+  }
   BLOCK_SIGNALS;
   _pthread_lock();
   for (e = dll_first(_pthread_list); e; e = dll_next(_pthread_list, e)) {
@@ -428,7 +431,6 @@ textwindows void __sig_generate(int sig, int sic) {
     // choose this thread if it isn't masking sig
     if (!(atomic_load_explicit(&pt->tib->tib_sigmask, memory_order_acquire) &
           (1ull << (sig - 1)))) {
-      STRACE("generating %G by killing %d", sig, _pthread_tid(pt));
       _pthread_ref(pt);
       mark = pt;
       break;
@@ -439,7 +441,6 @@ textwindows void __sig_generate(int sig, int sic) {
     if (atomic_load_explicit(&pt->pt_blocker, memory_order_acquire) &&
         !(atomic_load_explicit(&pt->pt_blkmask, memory_order_relaxed) &
           (1ull << (sig - 1)))) {
-      STRACE("generating %G by unblocking %d", sig, _pthread_tid(pt));
       _pthread_ref(pt);
       mark = pt;
       break;
@@ -450,7 +451,6 @@ textwindows void __sig_generate(int sig, int sic) {
     __sig_killer(mark, sig, sic);
     _pthread_unref(mark);
   } else {
-    STRACE("all threads block %G so adding to pending signals of process", sig);
     atomic_fetch_or_explicit(&__sig.pending, 1ull << (sig - 1),
                              memory_order_relaxed);
   }
diff --git a/libc/calls/sig.internal.h b/libc/calls/sig.internal.h
index d6f2c1bace4..cedc38338af 100644
--- a/libc/calls/sig.internal.h
+++ b/libc/calls/sig.internal.h
@@ -3,13 +3,8 @@
 #include "libc/calls/struct/sigset.h"
 #include "libc/thread/posixthread.internal.h"
 
-#define __SIG_LOCK_INTERVAL_MS    1000 /* semaphore synchronization: solid */
-#define __SIG_SIG_INTERVAL_MS     1000 /* posix signal polyfill also solid */
-#define __SIG_PROC_INTERVAL_MS    1000 /* process waiting also pretty good */
-#define __SIG_IO_INTERVAL_MS      1000 /* read/write cancel/notify is good */
-#define __SIG_POLL_INTERVAL_MS    20   /* poll on windows is dumpster fire */
-#define __SIG_LOGGING_INTERVAL_MS 1700
-#define __SIG_QUEUE_LENGTH        32
+#define SIG_HANDLED_NO_RESTART 1
+#define SIG_HANDLED_SA_RESTART 2
 
 #if !(__ASSEMBLER__ + __LINKER__ + 0)
 COSMOPOLITAN_C_START_
diff --git a/libc/calls/struct/timespec.h b/libc/calls/struct/timespec.h
index 739cfadad68..3b8b04c314e 100644
--- a/libc/calls/struct/timespec.h
+++ b/libc/calls/struct/timespec.h
@@ -41,10 +41,10 @@ int sys_futex(int *, int, int, const struct timespec *, int *);
 static inline struct timespec timespec_fromseconds(int64_t __x) {
   return (struct timespec){__x};
 }
-static inline bool timespec_iszero(struct timespec __ts) {
+static inline int timespec_iszero(struct timespec __ts) {
   return !(__ts.tv_sec | __ts.tv_nsec);
 }
-static inline bool timespec_isvalid(struct timespec __ts) {
+static inline int timespec_isvalid(struct timespec __ts) {
   return __ts.tv_sec >= 0 && __ts.tv_nsec + 0ull < 1000000000ull;
 }
 #endif /* _COSMO_SOURCE */
diff --git a/libc/calls/struct/timeval.h b/libc/calls/struct/timeval.h
index 9c12dfa1336..85f41e222ea 100644
--- a/libc/calls/struct/timeval.h
+++ b/libc/calls/struct/timeval.h
@@ -39,10 +39,10 @@ static inline struct timeval timeval_fromseconds(int64_t __x) {
 static inline struct timespec timeval_totimespec(struct timeval __tv) {
   return (struct timespec){__tv.tv_sec, __tv.tv_usec * 1000};
 }
-static inline bool timeval_iszero(struct timeval __tv) {
+static inline int timeval_iszero(struct timeval __tv) {
   return !(__tv.tv_sec | __tv.tv_usec);
 }
-static inline bool timeval_isvalid(struct timeval __tv) {
+static inline int timeval_isvalid(struct timeval __tv) {
   return __tv.tv_sec >= 0 && __tv.tv_usec + 0ull < 1000000ull;
 }
 #endif /* _COSMO_SOURCE */
diff --git a/libc/calls/ucontext.c b/libc/calls/ucontext.c
index cb99bb669e2..66281516bd7 100644
--- a/libc/calls/ucontext.c
+++ b/libc/calls/ucontext.c
@@ -17,17 +17,14 @@
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/calls/ucontext.h"
-#include "libc/calls/struct/sigset.h"
-#include "libc/runtime/runtime.h"
+#include "libc/calls/struct/sigset.internal.h"
+#include "libc/dce.h"
+#include "libc/intrin/atomic.h"
 #include "libc/sysv/consts/sig.h"
+#include "libc/thread/tls.h"
 
 int __tailcontext(const ucontext_t *);
 
-static int __contextmask(const sigset_t *opt_set, sigset_t *opt_out_oldset) {
-  // now context switching needs to go 14x slower
-  return sigprocmask(SIG_SETMASK, opt_set, opt_out_oldset);
-}
-
 /**
  * Sets machine context.
  *
@@ -37,14 +34,31 @@ static int __contextmask(const sigset_t *opt_set, sigset_t *opt_out_oldset) {
  * @see getcontext()
  */
 int setcontext(const ucontext_t *uc) {
-  if (__contextmask(&uc->uc_sigmask, 0)) return -1;
+  if (IsWindows()) {
+    atomic_store_explicit(&__get_tls()->tib_sigmask, uc->uc_sigmask,
+                          memory_order_release);
+  } else {
+    sys_sigprocmask(SIG_SETMASK, &uc->uc_sigmask, 0);
+  }
   return __tailcontext(uc);
 }
 
 int __getcontextsig(ucontext_t *uc) {
-  return __contextmask(0, &uc->uc_sigmask);
+  if (IsWindows()) {
+    uc->uc_sigmask =
+        atomic_load_explicit(&__get_tls()->tib_sigmask, memory_order_acquire);
+    return 0;
+  } else {
+    return sys_sigprocmask(SIG_SETMASK, 0, &uc->uc_sigmask);
+  }
 }
 
 int __swapcontextsig(ucontext_t *x, const ucontext_t *y) {
-  return __contextmask(&y->uc_sigmask, &x->uc_sigmask);
+  if (IsWindows()) {
+    x->uc_sigmask = atomic_exchange_explicit(
+        &__get_tls()->tib_sigmask, y->uc_sigmask, memory_order_acquire);
+    return 0;
+  } else {
+    return sys_sigprocmask(SIG_SETMASK, &y->uc_sigmask, &x->uc_sigmask);
+  }
 }
diff --git a/libc/fmt/conv.h b/libc/fmt/conv.h
index edec8bba3d5..a2941b71551 100644
--- a/libc/fmt/conv.h
+++ b/libc/fmt/conv.h
@@ -73,7 +73,8 @@ imaxdiv_t imaxdiv(intmax_t, intmax_t) pureconst;
 #endif
 
 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 406 || defined(__llvm__)) && \
-    !defined(__STRICT_ANSI__) && defined(_COSMO_SOURCE)
+    !defined(__STRICT_ANSI__) && defined(_COSMO_SOURCE) &&           \
+    !defined(__COSMOCC__)
 int128_t i128abs(int128_t)
 libcesque pureconst;
 int128_t strtoi128(const char *, char **, int) paramsnonnull((1));
diff --git a/libc/intrin/createfile.c b/libc/intrin/createfile.c
index 183477e45c6..2aab825be18 100644
--- a/libc/intrin/createfile.c
+++ b/libc/intrin/createfile.c
@@ -71,7 +71,7 @@ CreateFile(const char16_t *lpFileName,                         //
     switch (__imp_GetLastError()) {
       case kNtErrorPipeBusy:
         if (micros >= 1024) __imp_Sleep(micros / 1024);
-        if (micros / 1024 < __SIG_IO_INTERVAL_MS) micros <<= 1;
+        if (micros < 1024 * 1024) micros <<= 1;
         goto TryAgain;
       case kNtErrorAccessDenied:
         // GetNtOpenFlags() always greedily requests execute permissions
diff --git a/libc/intrin/createnamedpipe.c b/libc/intrin/createnamedpipe.c
index aec9b2e03ec..2f23361fd3b 100644
--- a/libc/intrin/createnamedpipe.c
+++ b/libc/intrin/createnamedpipe.c
@@ -50,7 +50,7 @@ textwindows int64_t CreateNamedPipe(
                                    nDefaultTimeOutMs, opt_lpSecurityAttributes);
   if (hServer == -1 && __imp_GetLastError() == kNtErrorPipeBusy) {
     if (micros >= 1024) __imp_Sleep(micros / 1024);
-    if (micros / 1024 < __SIG_IO_INTERVAL_MS) micros <<= 1;
+    if (micros < 1024 * 1024) micros <<= 1;
     goto TryAgain;
   }
   if (hServer == -1) __winerr();
diff --git a/libc/intrin/pthread_unref.c b/libc/intrin/pthread_unref.c
index 20c40740d61..f67d157ba84 100644
--- a/libc/intrin/pthread_unref.c
+++ b/libc/intrin/pthread_unref.c
@@ -23,9 +23,7 @@
 
 static bool _pthread_deref(struct PosixThread *pt) {
   int refs = atomic_load_explicit(&pt->pt_refs, memory_order_acquire);
-  if (!refs) return true;
-  unassert(refs > 0);
-  return !atomic_fetch_sub(&pt->pt_refs, 1);
+  return !refs || !atomic_fetch_sub(&pt->pt_refs, 1);
 }
 
 void _pthread_unref(struct PosixThread *pt) {
diff --git a/libc/intrin/sigprocmask-nt.c b/libc/intrin/sigprocmask-nt.c
index 781589726d0..5acc34692cd 100644
--- a/libc/intrin/sigprocmask-nt.c
+++ b/libc/intrin/sigprocmask-nt.c
@@ -42,7 +42,7 @@ textwindows int __sig_mask(int how, const sigset_t *neu, sigset_t *old) {
     if (how == SIG_BLOCK) {
       oldmask = atomic_fetch_or_explicit(mask, *neu, memory_order_acq_rel);
     } else if (how == SIG_UNBLOCK) {
-      oldmask = atomic_fetch_and_explicit(mask, *neu, memory_order_acq_rel);
+      oldmask = atomic_fetch_and_explicit(mask, ~*neu, memory_order_acq_rel);
     } else {  // SIG_SETMASK
       oldmask = atomic_exchange_explicit(mask, *neu, memory_order_acq_rel);
     }
diff --git a/libc/intrin/wsarecv.c b/libc/intrin/wsarecv.c
index 89cbf35bf73..a35e7f7ccda 100644
--- a/libc/intrin/wsarecv.c
+++ b/libc/intrin/wsarecv.c
@@ -30,9 +30,7 @@ __msabi extern typeof(WSARecv) *const __imp_WSARecv;
 
 /**
  * Receives data from Windows socket.
- *
  * @return 0 on success, or -1 on failure
- * @note this wrapper takes care of ABI, STRACE(), and __winerr()
  */
 textwindows int WSARecv(
     uint64_t s, const struct NtIovec *inout_lpBuffers, uint32_t dwBufferCount,
@@ -57,9 +55,6 @@ textwindows int WSARecv(
   if (opt_out_lpNumberOfBytesRecvd) {
     *opt_out_lpNumberOfBytesRecvd = NumberOfBytesRecvd;
   }
-  if (rc == -1) {
-    __winsockerr();
-  }
   if (UNLIKELY(__strace > 0) && strace_enabled(0) > 0) {
     kprintf(STRACE_PROLOGUE "WSARecv(%lu, [", s);
     DescribeIovNt(inout_lpBuffers, dwBufferCount,
@@ -73,9 +68,6 @@ textwindows int WSARecv(
   rc = __imp_WSARecv(s, inout_lpBuffers, dwBufferCount,
                      opt_out_lpNumberOfBytesRecvd, inout_lpFlags,
                      opt_inout_lpOverlapped, opt_lpCompletionRoutine);
-  if (rc == -1) {
-    __winsockerr();
-  }
 #endif
   return rc;
 }
diff --git a/libc/intrin/wsarecvfrom.c b/libc/intrin/wsarecvfrom.c
index 74d881dce48..49f70e736e0 100644
--- a/libc/intrin/wsarecvfrom.c
+++ b/libc/intrin/wsarecvfrom.c
@@ -29,9 +29,7 @@ __msabi extern typeof(WSARecvFrom) *const __imp_WSARecvFrom;
 
 /**
  * Receives data from Windows socket.
- *
  * @return 0 on success, or -1 on failure
- * @note this wrapper takes care of ABI, STRACE(), and __winerr()
  */
 textwindows int WSARecvFrom(
     uint64_t s, const struct NtIovec *inout_lpBuffers, uint32_t dwBufferCount,
@@ -52,9 +50,6 @@ textwindows int WSARecvFrom(
   if (opt_out_lpNumberOfBytesRecvd) {
     *opt_out_lpNumberOfBytesRecvd = NumberOfBytesRecvd;
   }
-  if (rc == -1) {
-    __winerr();
-  }
   if (UNLIKELY(__strace > 0) && strace_enabled(0) > 0) {
     kprintf(STRACE_PROLOGUE "WSARecvFrom(%lu, [", s);
     DescribeIovNt(inout_lpBuffers, dwBufferCount,
@@ -69,9 +64,6 @@ textwindows int WSARecvFrom(
                          opt_out_lpNumberOfBytesRecvd, inout_lpFlags,
                          opt_out_fromsockaddr, opt_inout_fromsockaddrlen,
                          opt_inout_lpOverlapped, opt_lpCompletionRoutine);
-  if (rc == -1) {
-    __winerr();
-  }
 #endif
   return rc;
 }
diff --git a/libc/proc/execve-nt.greg.c b/libc/proc/execve-nt.greg.c
index 9894b6484ee..a8799cda65b 100644
--- a/libc/proc/execve-nt.greg.c
+++ b/libc/proc/execve-nt.greg.c
@@ -22,6 +22,7 @@
 #include "libc/calls/struct/sigset.internal.h"
 #include "libc/calls/syscall-nt.internal.h"
 #include "libc/errno.h"
+#include "libc/fmt/itoa.h"
 #include "libc/intrin/kprintf.h"
 #include "libc/mem/mem.h"
 #include "libc/nt/enum/processaccess.h"
@@ -46,7 +47,7 @@ textwindows int sys_execve_nt(const char *program, char *const argv[],
                               char *const envp[]) {
 
   // execve() needs to be @asyncsignalsafe
-  sigset_t m = __sig_block();
+  sigset_t sigmask = __sig_block();
   _pthread_lock();
 
   // new process should be a child of our parent
@@ -55,10 +56,14 @@ textwindows int sys_execve_nt(const char *program, char *const argv[],
   if (!(hParentProcess = OpenProcess(
             kNtProcessDupHandle | kNtProcessCreateProcess, false, ppid))) {
     _pthread_unlock();
-    __sig_unblock(m);
+    __sig_unblock(sigmask);
     return -1;
   }
 
+  // inherit signal mask
+  char maskvar[6 + 21];
+  FormatUint64(stpcpy(maskvar, "_MASK="), sigmask);
+
   // define stdio handles for the spawned subprocess
   struct NtStartupInfo si = {
       .cb = sizeof(struct NtStartupInfo),
@@ -80,21 +85,21 @@ textwindows int sys_execve_nt(const char *program, char *const argv[],
                                 &lpExplicitHandles, &dwExplicitHandleCount))) {
     CloseHandle(hParentProcess);
     _pthread_unlock();
-    __sig_unblock(m);
+    __sig_unblock(sigmask);
     return -1;
   }
 
   // launch the process
   struct NtProcessInformation pi;
-  int rc = ntspawn(AT_FDCWD, program, argv, envp, (char *[]){fdspec, 0}, 0, 0,
-                   hParentProcess, lpExplicitHandles, dwExplicitHandleCount,
-                   &si, &pi);
+  int rc = ntspawn(AT_FDCWD, program, argv, envp,
+                   (char *[]){fdspec, maskvar, 0}, 0, 0, hParentProcess,
+                   lpExplicitHandles, dwExplicitHandleCount, &si, &pi);
   __undescribe_fds(hParentProcess, lpExplicitHandles, dwExplicitHandleCount);
   if (rc == -1) {
     free(fdspec);
     CloseHandle(hParentProcess);
     _pthread_unlock();
-    __sig_unblock(m);
+    __sig_unblock(sigmask);
     if (GetLastError() == kNtErrorSharingViolation) {
       return etxtbsy();
     } else {
diff --git a/libc/proc/fork-nt.c b/libc/proc/fork-nt.c
index 529004c8964..1be406dd1f4 100644
--- a/libc/proc/fork-nt.c
+++ b/libc/proc/fork-nt.c
@@ -401,7 +401,7 @@ textwindows int sys_fork_nt(uint32_t dwCreationFlags) {
     }
   }
   if (rc == -1) {
-    __proc_free(proc);
+    dll_make_first(&__proc.free, &proc->elem);
   }
   ftrace_enabled(+1);
   strace_enabled(+1);
diff --git a/libc/proc/posix_spawn.c b/libc/proc/posix_spawn.c
index bd9466daeeb..f2327a34003 100644
--- a/libc/proc/posix_spawn.c
+++ b/libc/proc/posix_spawn.c
@@ -248,9 +248,9 @@ static textwindows errno_t posix_spawn_nt_impl(
   struct SpawnFds fds = {0};
   int64_t dirhand = AT_FDCWD;
   int64_t *lpExplicitHandles = 0;
+  sigset_t sigmask = __sig_block();
   uint32_t dwExplicitHandleCount = 0;
   int64_t hCreatorProcess = GetCurrentProcess();
-  sigset_t m = __sig_block();
 
   // reserve process tracking object
   __proc_lock();
@@ -266,11 +266,11 @@ static textwindows errno_t posix_spawn_nt_impl(
     free(fdspec);
     if (proc) {
       __proc_lock();
-      __proc_free(proc);
+      dll_make_first(&__proc.free, &proc->elem);
       __proc_unlock();
     }
     spawnfds_destroy(&fds);
-    __sig_unblock(m);
+    __sig_unblock(sigmask);
     errno = e;
     return err;
   }
@@ -360,13 +360,17 @@ static textwindows errno_t posix_spawn_nt_impl(
     }
   }
 
+  // inherit signal mask
+  char maskvar[6 + 21];
+  FormatUint64(stpcpy(maskvar, "_MASK="), sigmask);
+
   // launch process
   int rc = -1;
   struct NtProcessInformation procinfo;
   if (!envp) envp = environ;
   if ((fdspec = __describe_fds(fds.p, fds.n, &startinfo, hCreatorProcess,
                                &lpExplicitHandles, &dwExplicitHandleCount))) {
-    rc = ntspawn(dirhand, path, argv, envp, (char *[]){fdspec, 0},
+    rc = ntspawn(dirhand, path, argv, envp, (char *[]){fdspec, maskvar, 0},
                  dwCreationFlags, lpCurrentDirectory, 0, lpExplicitHandles,
                  dwExplicitHandleCount, &startinfo, &procinfo);
   }
diff --git a/libc/proc/posix_spawnattr_setflags.c b/libc/proc/posix_spawnattr_setflags.c
index 9521bb9b2cd..cccc017f6bc 100644
--- a/libc/proc/posix_spawnattr_setflags.c
+++ b/libc/proc/posix_spawnattr_setflags.c
@@ -32,15 +32,20 @@
  *     - `POSIX_SPAWN_SETSCHEDPARAM`
  *     - `POSIX_SPAWN_SETSCHEDULER`
  *     - `POSIX_SPAWN_SETSID`
+ *     - `POSIX_SPAWN_SETRLIMIT`
  * @return 0 on success, or errno on error
  * @raise EINVAL if `flags` has invalid bits
  */
 int posix_spawnattr_setflags(posix_spawnattr_t *attr, short flags) {
-  if (flags &
-      ~(POSIX_SPAWN_USEVFORK | POSIX_SPAWN_RESETIDS | POSIX_SPAWN_SETPGROUP |
-        POSIX_SPAWN_SETSIGDEF | POSIX_SPAWN_SETSIGMASK |
-        POSIX_SPAWN_SETSCHEDPARAM | POSIX_SPAWN_SETSCHEDULER |
-        POSIX_SPAWN_SETSID | POSIX_SPAWN_SETRLIMIT)) {
+  if (flags & ~(POSIX_SPAWN_USEVFORK |       //
+                POSIX_SPAWN_RESETIDS |       //
+                POSIX_SPAWN_SETPGROUP |      //
+                POSIX_SPAWN_SETSIGDEF |      //
+                POSIX_SPAWN_SETSIGMASK |     //
+                POSIX_SPAWN_SETSCHEDPARAM |  //
+                POSIX_SPAWN_SETSCHEDULER |   //
+                POSIX_SPAWN_SETSID |         //
+                POSIX_SPAWN_SETRLIMIT)) {
     return EINVAL;
   }
   (*attr)->flags = flags;
diff --git a/libc/proc/posix_spawnattr_setpgroup.c b/libc/proc/posix_spawnattr_setpgroup.c
index 2d6263d611b..9e1d53d8534 100644
--- a/libc/proc/posix_spawnattr_setpgroup.c
+++ b/libc/proc/posix_spawnattr_setpgroup.c
@@ -16,7 +16,6 @@
 │ TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR             │
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
-#include "libc/proc/posix_spawn.h"
 #include "libc/proc/posix_spawn.internal.h"
 
 /**
@@ -25,14 +24,14 @@
  * Setting `pgroup` to zero will ensure newly created processes are
  * placed within their own brand new process group.
  *
- * This setter also sets the `POSIX_SPAWN_SETPGROUP` flag.
+ * You also need to pass `POSIX_SPAWN_SETPGROUP` to
+ * posix_spawnattr_setflags() for it to take effect.
  *
  * @param attr was initialized by posix_spawnattr_init()
  * @param pgroup is the process group id, or 0 for self
  * @return 0 on success, or errno on error
  */
 int posix_spawnattr_setpgroup(posix_spawnattr_t *attr, int pgroup) {
-  (*attr)->flags |= POSIX_SPAWN_SETPGROUP;
   (*attr)->pgroup = pgroup;
   return 0;
 }
diff --git a/libc/proc/posix_spawnattr_setrlimit.c b/libc/proc/posix_spawnattr_setrlimit.c
index 90be67c431d..7c7202edbdc 100644
--- a/libc/proc/posix_spawnattr_setrlimit.c
+++ b/libc/proc/posix_spawnattr_setrlimit.c
@@ -26,7 +26,8 @@
 /**
  * Sets resource limit on spawned process.
  *
- * Using this setter automatically sets `POSIX_SPAWN_SETRLIMIT`.
+ * You also need to pass `POSIX_SPAWN_SETRLIMIT` to
+ * posix_spawnattr_setflags() for it to take effect.
  *
  * @return 0 on success, or errno on error
  * @raise EINVAL if resource is invalid
@@ -34,7 +35,6 @@
 int posix_spawnattr_setrlimit(posix_spawnattr_t *attr, int resource,
                               const struct rlimit *rlim) {
   if (0 <= resource && resource < MIN(RLIM_NLIMITS, ARRAYLEN((*attr)->rlim))) {
-    (*attr)->flags |= POSIX_SPAWN_SETRLIMIT;
     (*attr)->rlimset |= 1u << resource;
     (*attr)->rlim[resource] = *rlim;
     return 0;
diff --git a/libc/proc/posix_spawnattr_setschedparam.c b/libc/proc/posix_spawnattr_setschedparam.c
index e73fdce0155..b7628d49d21 100644
--- a/libc/proc/posix_spawnattr_setschedparam.c
+++ b/libc/proc/posix_spawnattr_setschedparam.c
@@ -16,14 +16,13 @@
 │ TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR             │
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
-#include "libc/calls/struct/sched_param.h"
-#include "libc/proc/posix_spawn.h"
 #include "libc/proc/posix_spawn.internal.h"
 
 /**
  * Specifies scheduler parameter override for spawned process.
  *
- * Using this setter automatically sets `POSIX_SPAWN_SETSCHEDPARAM`.
+ * You also need to pass `POSIX_SPAWN_SETSCHEDPARAM` to
+ * posix_spawnattr_setflags() for it to take effect.
  *
  * @param attr was initialized by posix_spawnattr_init()
  * @param schedparam receives the result
@@ -31,7 +30,6 @@
  */
 int posix_spawnattr_setschedparam(posix_spawnattr_t *attr,
                                   const struct sched_param *schedparam) {
-  (*attr)->flags |= POSIX_SPAWN_SETSCHEDPARAM;
   (*attr)->schedparam = *schedparam;
   return 0;
 }
diff --git a/libc/proc/posix_spawnattr_setschedpolicy.c b/libc/proc/posix_spawnattr_setschedpolicy.c
index dd1ed863cf7..e52c9ca93fe 100644
--- a/libc/proc/posix_spawnattr_setschedpolicy.c
+++ b/libc/proc/posix_spawnattr_setschedpolicy.c
@@ -16,19 +16,18 @@
 │ TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR             │
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
-#include "libc/proc/posix_spawn.h"
 #include "libc/proc/posix_spawn.internal.h"
 
 /**
  * Specifies scheduler policy override for spawned process.
  *
- * Using this setter automatically sets `POSIX_SPAWN_SETSCHEDULER`.
+ * You also need to pass `POSIX_SPAWN_SETSCHEDULER` to
+ * posix_spawnattr_setflags() for it to take effect.
  *
  * @param attr was initialized by posix_spawnattr_init()
  * @return 0 on success, or errno on error
  */
 int posix_spawnattr_setschedpolicy(posix_spawnattr_t *attr, int schedpolicy) {
-  (*attr)->flags |= POSIX_SPAWN_SETSCHEDULER;
   (*attr)->schedpolicy = schedpolicy;
   return 0;
 }
diff --git a/libc/proc/posix_spawnattr_setsigdefault.c b/libc/proc/posix_spawnattr_setsigdefault.c
index 7f277d83997..7c55e6494b4 100644
--- a/libc/proc/posix_spawnattr_setsigdefault.c
+++ b/libc/proc/posix_spawnattr_setsigdefault.c
@@ -16,8 +16,6 @@
 │ TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR             │
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
-#include "libc/calls/struct/sigset.h"
-#include "libc/proc/posix_spawn.h"
 #include "libc/proc/posix_spawn.internal.h"
 
 /**
@@ -32,13 +30,13 @@
  * set the signals to `SIG_IGN` earlier (since posix_spawn() will not
  * issue O(128) system calls just to be totally pedantic about that).
  *
- * Using this setter automatically sets `POSIX_SPAWN_SETSIGDEF`.
+ * You also need to pass `POSIX_SPAWN_SETSIGDEF` to
+ * posix_spawnattr_setflags() for it to take effect.
  *
  * @return 0 on success, or errno on error
  */
 int posix_spawnattr_setsigdefault(posix_spawnattr_t *attr,
                                   const sigset_t *sigdefault) {
-  (*attr)->flags |= POSIX_SPAWN_SETSIGDEF;
   (*attr)->sigdefault = *sigdefault;
   return 0;
 }
diff --git a/libc/proc/posix_spawnattr_setsigmask.c b/libc/proc/posix_spawnattr_setsigmask.c
index 842fcd8f177..a882f79a0cf 100644
--- a/libc/proc/posix_spawnattr_setsigmask.c
+++ b/libc/proc/posix_spawnattr_setsigmask.c
@@ -16,20 +16,18 @@
 │ TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR             │
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
-#include "libc/calls/struct/sigset.h"
-#include "libc/proc/posix_spawn.h"
 #include "libc/proc/posix_spawn.internal.h"
 
 /**
  * Specifies signal mask for sigprocmask() in child process.
  *
- * This setter also sets the `POSIX_SPAWN_SETSIGMASK` flag.
+ * You also need to pass `POSIX_SPAWN_SETSIGMASK` to
+ * posix_spawnattr_setflags() for it to take effect.
  *
  * @return 0 on success, or errno on error
  */
 int posix_spawnattr_setsigmask(posix_spawnattr_t *attr,
                                const sigset_t *sigmask) {
-  (*attr)->flags |= POSIX_SPAWN_SETSIGMASK;
   (*attr)->sigmask = *sigmask;
   return 0;
 }
diff --git a/libc/proc/proc.c b/libc/proc/proc.c
index 4c3cbef1e31..10a8a80ee53 100644
--- a/libc/proc/proc.c
+++ b/libc/proc/proc.c
@@ -16,7 +16,6 @@
 │ TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR             │
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
-#include "libc/assert.h"
 #include "libc/calls/calls.h"
 #include "libc/calls/internal.h"
 #include "libc/calls/sig.internal.h"
@@ -37,6 +36,7 @@
 #include "libc/nt/enum/processcreationflags.h"
 #include "libc/nt/enum/status.h"
 #include "libc/nt/enum/wait.h"
+#include "libc/nt/events.h"
 #include "libc/nt/process.h"
 #include "libc/nt/runtime.h"
 #include "libc/nt/struct/filetime.h"
@@ -59,95 +59,167 @@
 
 struct Procs __proc;
 
-static textwindows void GetProcessStats(int64_t h, struct rusage *ru) {
+static textwindows void __proc_stats(int64_t h, struct rusage *ru) {
   bzero(ru, sizeof(*ru));
   struct NtProcessMemoryCountersEx memcount = {sizeof(memcount)};
-  unassert(GetProcessMemoryInfo(h, &memcount, sizeof(memcount)));
+  GetProcessMemoryInfo(h, &memcount, sizeof(memcount));
   ru->ru_maxrss = memcount.PeakWorkingSetSize / 1024;
   ru->ru_majflt = memcount.PageFaultCount;
   struct NtFileTime createtime, exittime;
   struct NtFileTime kerneltime, usertime;
-  unassert(GetProcessTimes(h, &createtime, &exittime, &kerneltime, &usertime));
+  GetProcessTimes(h, &createtime, &exittime, &kerneltime, &usertime);
   ru->ru_utime = WindowsDurationToTimeVal(ReadFileTime(usertime));
   ru->ru_stime = WindowsDurationToTimeVal(ReadFileTime(kerneltime));
   struct NtIoCounters iocount;
-  unassert(GetProcessIoCounters(h, &iocount));
+  GetProcessIoCounters(h, &iocount);
   ru->ru_inblock = iocount.ReadOperationCount;
   ru->ru_oublock = iocount.WriteOperationCount;
 }
 
+// performs accounting on exited process
+// multiple threads can wait on a process
+// it's important that only one calls this
+textwindows int __proc_harvest(struct Proc *pr, bool iswait4) {
+  int sic = 0;
+  uint32_t status;
+  struct rusage ru;
+  GetExitCodeProcess(pr->handle, &status);
+  if (status == kNtStillActive) return 0;
+  __proc_stats(pr->handle, &ru);
+  rusage_add(&pr->ru, &ru);
+  rusage_add(&__proc.ruchlds, &ru);
+  if ((status & 0xFF000000u) == 0x23000000u) {
+    // handle child execve()
+    CloseHandle(pr->handle);
+    pr->handle = status & 0x00FFFFFF;
+  } else {
+    // handle child _Exit()
+    if (status == 0xc9af3d51u) {
+      status = kNtStillActive;
+    }
+    pr->wstatus = status;
+    if (!iswait4 && !pr->waiters && !__proc.waiters &&
+        (__sighandrvas[SIGCHLD] == (uintptr_t)SIG_IGN ||
+         (__sighandflags[SIGCHLD] & SA_NOCLDWAIT))) {
+      // perform automatic zombie reaping
+      dll_remove(&__proc.list, &pr->elem);
+      dll_make_first(&__proc.free, &pr->elem);
+      CloseHandle(pr->handle);
+    } else {
+      // transitions process to zombie state
+      // wait4 is responsible for reaping it
+      pr->status = PROC_ZOMBIE;
+      dll_remove(&__proc.list, &pr->elem);
+      dll_make_first(&__proc.zombies, &pr->elem);
+      SetEvent(__proc.haszombies);
+      if (!pr->waiters && !__proc.waiters) {
+        if (WIFSIGNALED(status)) {
+          sic = CLD_KILLED;
+        } else {
+          sic = CLD_EXITED;
+        }
+      }
+    }
+  }
+  return sic;
+}
+
 static textwindows dontinstrument uint32_t __proc_worker(void *arg) {
-  __bootstrap_tls(&__proc.tls, __builtin_frame_address(0));
+  struct CosmoTib tls;
+  __bootstrap_tls(&tls, __builtin_frame_address(0));
   for (;;) {
 
     // assemble a group of processes to wait on. if more than 64
     // children exist, then we'll use a small timeout and select
     // processes with a shifting window via a double linked list
-    struct rusage ru;
+    // if fewer than 64 processes exist then we'll also wait for
+    // process birth notifications, and wait on them immediately
     int64_t handles[64];
-    int sic, dosignal = 0;
-    struct Proc *pr, *objects[64];
+    struct Proc *objects[64];
+    uint32_t millis, i, n = 0;
     struct Dll *e, *e2, *samples = 0;
-    uint32_t millis, status, i, n = 1;
     __proc_lock();
-    handles[0] = __proc.onstart;
-    for (e = dll_first(__proc.list); e && n < 64; ++n, e = e2) {
-      pr = PROC_CONTAINER(e);
+    for (e = dll_first(__proc.list); e && n < 64; e = e2) {
+      struct Proc *pr = PROC_CONTAINER(e);
       e2 = dll_next(__proc.list, e);
+      // cycle process to end of list
       dll_remove(&__proc.list, e);
       dll_make_last(&samples, e);
-      handles[n] = pr->handle;
-      objects[n] = pr;
+      // don't bother waiting if it's already awaited
+      if (!pr->waiters) {
+        handles[n] = pr->handle;
+        objects[n] = pr;
+        ++pr->waiters;
+        ++n;
+      }
     }
     dll_make_last(&__proc.list, samples);
     __proc_unlock();
 
-    // wait for win32 to report any status change
-    millis = n == 64 ? 20 : -1u;
-    unassert((i = WaitForMultipleObjects(n, handles, false, millis)) != -1u);
-    i &= ~kNtWaitAbandoned;
-    if (!i || i == kNtWaitTimeout) continue;
-    GetExitCodeProcess(handles[i], &status);
-    if (status == kNtStillActive) continue;
-    GetProcessStats(handles[i], &ru);
-
-    // update data structures and notify folks
-    __proc_lock();
-    pr = objects[i];
-    rusage_add(&pr->ru, &ru);
-    rusage_add(&__proc.ruchlds, &ru);
-    if ((status & 0xFF000000u) == 0x23000000u) {
-      // handle child execve()
-      CloseHandle(pr->handle);
-      pr->handle = status & 0x00FFFFFF;
+    // wait for something to happen
+    if (n == 64) {
+      millis = 5;
     } else {
-      // handle child _exit()
-      CloseHandle(pr->handle);
-      if (status == 0xc9af3d51u) {
-        status = kNtStillActive;
+      millis = -1u;
+      handles[n++] = __proc.onbirth;
+    }
+    i = WaitForMultipleObjects(n, handles, false, millis);
+    if (i == -1u) {
+      STRACE("proc wait panic %d", GetLastError());
+      _Exit(157);
+    }
+    if (i & kNtWaitAbandoned) {
+      i &= ~kNtWaitAbandoned;
+      STRACE("proc %u handle %ld abandoned", i, handles[i]);
+    }
+    __proc_lock();
+
+    // release our waiter status
+    for (int j = 0; j < n; ++j) {
+      if (handles[j] == __proc.onbirth) continue;
+      if (j == i) continue;
+      if (!--objects[j]->waiters && objects[j]->status == PROC_UNDEAD) {
+        __proc_free(objects[j]);
       }
-      pr->wstatus = status;
-      if ((__sighandrvas[SIGCHLD] == (uintptr_t)SIG_IGN ||
-           (__sighandflags[SIGCHLD] & SA_NOCLDWAIT)) &&
-          (!pr->waiters && !__proc.waiters)) {
-        dll_remove(&__proc.list, &pr->elem);
-        dll_make_first(&__proc.free, &pr->elem);
-      } else {
-        pr->iszombie = 1;
-        dll_remove(&__proc.list, &pr->elem);
-        dll_make_first(&__proc.zombies, &pr->elem);
-        if (pr->waiters) {
-          nsync_cv_broadcast(&pr->onexit);
-        } else if (__proc.waiters) {
-          nsync_cv_signal(&__proc.onexit);
-        } else {
-          dosignal = 1;
-          sic = WIFSIGNALED(status) ? CLD_KILLED : CLD_EXITED;
+    }
+
+    // check if we need to churn due to >64 processes
+    if (i == kNtWaitTimeout) {
+      __proc_unlock();
+      continue;
+    }
+
+    // churn on new process birth
+    if (handles[i] == __proc.onbirth) {
+      __proc_unlock();
+      continue;
+    }
+
+    // handle process status change
+    int sic = 0;
+    --objects[i]->waiters;
+    switch (objects[i]->status) {
+      case PROC_ALIVE:
+        sic = __proc_harvest(objects[i], false);
+        break;
+      case PROC_ZOMBIE:
+        break;
+      case PROC_UNDEAD:
+        if (!objects[i]->waiters) {
+          __proc_free(objects[i]);
         }
-      }
+        break;
+      default:
+        __builtin_unreachable();
     }
+
     __proc_unlock();
-    if (dosignal) {
+
+    // don't raise SIGCHLD if
+    // 1. wait4() is being used
+    // 2. SIGCHLD has SIG_IGN handler
+    // 3. SIGCHLD has SA_NOCLDWAIT flag
+    if (sic) {
       __sig_generate(SIGCHLD, sic);
     }
   }
@@ -158,7 +230,8 @@ static textwindows dontinstrument uint32_t __proc_worker(void *arg) {
  * Lazy initializes process tracker data structures and worker.
  */
 static textwindows void __proc_setup(void) {
-  __proc.onstart = CreateSemaphore(0, 0, 1, 0);
+  __proc.onbirth = CreateEvent(0, 0, 0, 0);     // auto reset
+  __proc.haszombies = CreateEvent(0, 1, 0, 0);  // manual reset
   __proc.thread = CreateThread(0, 65536, __proc_worker, 0,
                                kNtStackSizeParamIsAReservation, 0);
 }
@@ -191,7 +264,7 @@ textwindows void __proc_wipe(void) {
  * The returned memory is not tracked by any list. It must be filled in
  * with system process information and then added back to the system by
  * calling __proc_add(). If process creation fails, then it needs to be
- * released using __proc_free().
+ * added back to the __proc.free list by caller.
  */
 textwindows struct Proc *__proc_new(void) {
   struct Dll *e;
@@ -230,16 +303,13 @@ IGNORE_LEAKS(__proc_new)
  */
 textwindows void __proc_add(struct Proc *proc) {
   dll_make_first(&__proc.list, &proc->elem);
-  ReleaseSemaphore(__proc.onstart, 1, 0);
+  SetEvent(__proc.onbirth);
 }
 
-/**
- * Frees process allocation.
- *
- * Process must not be currently tracked in the active or zombies list.
- */
-textwindows void __proc_free(struct Proc *proc) {
-  dll_make_first(&__proc.free, &proc->elem);
+textwindows void __proc_free(struct Proc *pr) {
+  dll_remove(&__proc.undead, &pr->elem);
+  dll_make_first(&__proc.free, &pr->elem);
+  CloseHandle(pr->handle);
 }
 
 // returns owned handle of direct child process
diff --git a/libc/proc/proc.internal.h b/libc/proc/proc.internal.h
index c9ae58507da..1d9417d10a2 100644
--- a/libc/proc/proc.internal.h
+++ b/libc/proc/proc.internal.h
@@ -2,25 +2,27 @@
 #define COSMOPOLITAN_LIBC_PROC_H_
 #include "libc/atomic.h"
 #include "libc/calls/struct/rusage.h"
+#include "libc/intrin/atomic.h"
 #include "libc/intrin/dll.h"
 #include "libc/thread/thread.h"
-#include "libc/thread/tls.h"
-#include "third_party/nsync/cv.h"
 #include "third_party/nsync/mu.h"
 #if !(__ASSEMBLER__ + __LINKER__ + 0)
 COSMOPOLITAN_C_START_
 
+#define PROC_ALIVE  0
+#define PROC_ZOMBIE 1
+#define PROC_UNDEAD 2
+
 #define PROC_CONTAINER(e) DLL_CONTAINER(struct Proc, elem, e)
 
 struct Proc {
   int pid;
+  int status;
   int waiters;
-  bool iszombie;
   bool wasforked;
   uint32_t wstatus;
   int64_t handle;
   struct Dll elem;
-  nsync_cv onexit;
   struct rusage ru;
 };
 
@@ -28,15 +30,15 @@ struct Procs {
   int waiters;
   atomic_uint once;
   nsync_mu lock;
-  nsync_cv onexit;
   intptr_t thread;
-  intptr_t onstart;
+  intptr_t onbirth;
+  intptr_t haszombies;
   struct Dll *list;
   struct Dll *free;
+  struct Dll *undead;
   struct Dll *zombies;
   struct Proc pool[8];
   unsigned allocated;
-  struct CosmoTib tls;
   struct rusage ruchlds;
 };
 
@@ -50,6 +52,7 @@ int64_t __proc_search(int);
 struct Proc *__proc_new(void);
 void __proc_add(struct Proc *);
 void __proc_free(struct Proc *);
+int __proc_harvest(struct Proc *, bool);
 int sys_wait4_nt(int, int *, int, struct rusage *);
 
 COSMOPOLITAN_C_END_
diff --git a/libc/proc/proc.mk b/libc/proc/proc.mk
index 420ac911f5d..b8d7ef9265c 100644
--- a/libc/proc/proc.mk
+++ b/libc/proc/proc.mk
@@ -35,8 +35,7 @@ LIBC_PROC_A_DIRECTDEPS =				\
 	LIBC_STR					\
 	LIBC_SYSV					\
 	LIBC_SYSV_CALLS					\
-	THIRD_PARTY_NSYNC				\
-	THIRD_PARTY_NSYNC_MEM
+	THIRD_PARTY_NSYNC
 
 LIBC_PROC_A_DEPS :=					\
 	$(call uniq,$(foreach x,$(LIBC_PROC_A_DIRECTDEPS),$($(x))))
diff --git a/libc/proc/verynice.c b/libc/proc/verynice.c
index 602c3c65bb1..75a9832e260 100644
--- a/libc/proc/verynice.c
+++ b/libc/proc/verynice.c
@@ -19,6 +19,7 @@
 #include "libc/calls/calls.h"
 #include "libc/calls/struct/sched_param.h"
 #include "libc/errno.h"
+#include "libc/limits.h"
 #include "libc/sysv/consts/ioprio.h"
 #include "libc/sysv/consts/prio.h"
 #include "libc/sysv/consts/sched.h"
@@ -31,7 +32,7 @@
  */
 int verynice(void) {
   int e = errno;
-  setpriority(PRIO_PROCESS, 0, 10);
+  setpriority(PRIO_PROCESS, 0, NZERO);
   sys_ioprio_set(IOPRIO_WHO_PROCESS, 0,
                  IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
   struct sched_param param = {sched_get_priority_min(SCHED_IDLE)};
diff --git a/libc/proc/wait4-nt.c b/libc/proc/wait4-nt.c
index f5459712311..07fc4252615 100644
--- a/libc/proc/wait4-nt.c
+++ b/libc/proc/wait4-nt.c
@@ -16,139 +16,191 @@
 │ TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR             │
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
-#include "libc/assert.h"
-#include "libc/calls/calls.h"
 #include "libc/calls/internal.h"
 #include "libc/calls/sig.internal.h"
-#include "libc/calls/struct/rusage.h"
+#include "libc/calls/struct/sigset.h"
 #include "libc/calls/struct/sigset.internal.h"
-#include "libc/calls/struct/timespec.h"
-#include "libc/cosmo.h"
-#include "libc/errno.h"
-#include "libc/fmt/wintime.internal.h"
-#include "libc/intrin/atomic.h"
 #include "libc/intrin/dll.h"
 #include "libc/intrin/strace.internal.h"
-#include "libc/macros.internal.h"
-#include "libc/nt/accounting.h"
-#include "libc/nt/process.h"
+#include "libc/nt/enum/wait.h"
+#include "libc/nt/events.h"
 #include "libc/nt/runtime.h"
-#include "libc/nt/struct/filetime.h"
-#include "libc/nt/struct/processmemorycounters.h"
+#include "libc/nt/synchronization.h"
 #include "libc/proc/proc.internal.h"
-#include "libc/str/str.h"
+#include "libc/sysv/consts/sicode.h"
 #include "libc/sysv/consts/sig.h"
 #include "libc/sysv/consts/w.h"
 #include "libc/sysv/errfuns.h"
-#include "libc/thread/thread.h"
-#include "libc/thread/tls.h"
 #ifdef __x86_64__
 
-static textwindows struct timespec GetNextDeadline(struct timespec deadline) {
-  if (timespec_iszero(deadline)) deadline = timespec_real();
-  struct timespec delay = timespec_frommillis(__SIG_PROC_INTERVAL_MS);
-  return timespec_add(deadline, delay);
-}
-
-static textwindows int ReapZombie(struct Proc *pr, int *wstatus,
-                                  struct rusage *opt_out_rusage) {
+static textwindows int __proc_reap(struct Proc *pr, int *wstatus,
+                                   struct rusage *opt_out_rusage) {
   if (wstatus) {
     *wstatus = pr->wstatus;
   }
   if (opt_out_rusage) {
     *opt_out_rusage = pr->ru;
   }
-  if (!pr->waiters) {
-    dll_remove(&__proc.zombies, &pr->elem);
+  dll_remove(&__proc.zombies, &pr->elem);
+  if (dll_is_empty(__proc.zombies)) {
+    ResetEvent(__proc.haszombies);
+  }
+  if (pr->waiters) {
+    pr->status = PROC_UNDEAD;
+    dll_make_first(&__proc.undead, &pr->elem);
+  } else {
     dll_make_first(&__proc.free, &pr->elem);
+    CloseHandle(pr->handle);
   }
   return pr->pid;
 }
 
-static textwindows int CheckZombies(int pid, int *wstatus,
+static textwindows int __proc_check(int pid, int *wstatus,
                                     struct rusage *opt_out_rusage) {
   struct Dll *e;
   for (e = dll_first(__proc.zombies); e; e = dll_next(__proc.zombies, e)) {
     struct Proc *pr = PROC_CONTAINER(e);
-    if (pid == -1 && pr->waiters) {
-      continue;  // this zombie has been claimed
-    }
     if (pid == -1 || pid == pr->pid) {
-      return ReapZombie(pr, wstatus, opt_out_rusage);
+      return __proc_reap(pr, wstatus, opt_out_rusage);
     }
   }
   return 0;
 }
 
-static textwindows void UnwindWaiterCount(void *arg) {
-  int *waiters = arg;
-  --*waiters;
-}
+static textwindows int __proc_wait(int pid, int *wstatus, int options,
+                                   struct rusage *rusage, sigset_t waitmask) {
+  for (;;) {
 
-static textwindows int WaitForProcess(int pid, int *wstatus, int options,
-                                      struct rusage *rusage,
-                                      uint64_t waitmask) {
-  uint64_t m;
-  int rc, *wv;
-  nsync_cv *cv;
-  struct Dll *e;
-  struct Proc *pr;
-  struct timespec deadline = timespec_zero;
+    // check for signals and cancelation
+    int sig, handler_was_called;
+    if (_check_cancel() == -1) {
+      return -1;
+    }
+    if ((sig = __sig_get(waitmask))) {
+      handler_was_called = __sig_relay(sig, SI_KERNEL, waitmask);
+      if (_check_cancel() == -1) {
+        return -1;  // ECANCELED because SIGTHR was just handled
+      }
+      if (handler_was_called & SIG_HANDLED_NO_RESTART) {
+        return eintr();  // a non-SA_RESTART handler was called
+      }
+    }
 
-  // check list of processes that've already exited
-  if ((rc = CheckZombies(pid, wstatus, rusage))) {
-    return rc;
-  }
+    // check for zombie to harvest
+    __proc_lock();
+  CheckForZombies:
+    int rc = __proc_check(pid, wstatus, rusage);
+    if (rc || (options & WNOHANG)) {
+      __proc_unlock();
+      return rc;
+    }
 
-  // find the mark
-  pr = 0;
-  if (pid == -1) {
+    // there's no zombies left
+    // check if there's any living processes
     if (dll_is_empty(__proc.list)) {
+      __proc_unlock();
       return echild();
     }
-    cv = &__proc.onexit;
-    wv = &__proc.waiters;
-  } else {
-    for (e = dll_first(__proc.list); e; e = dll_next(__proc.list, e)) {
-      if (pid == PROC_CONTAINER(e)->pid) {
+
+    // get appropriate wait object
+    // register ourself as waiting
+    struct Proc *pr = 0;
+    uintptr_t hWaitObject;
+    if (pid == -1) {
+      // wait for any status change
+      hWaitObject = __proc.haszombies;
+      ++__proc.waiters;
+    } else {
+      // wait on specific child
+      for (struct Dll *e = dll_first(__proc.list); e;
+           e = dll_next(__proc.list, e)) {
         pr = PROC_CONTAINER(e);
+        if (pid == pr->pid) break;
+      }
+      if (pr) {
+        // by making the waiter count non-zero, the proc daemon stops
+        // being obligated to monitor this process. this means we may
+        // need to assume responsibility later on for zombifying this
+        ++pr->waiters;
+        hWaitObject = pr->handle;
+      } else {
+        __proc_unlock();
+        return echild();
       }
     }
-    if (pr) {
-      unassert(!pr->iszombie);
-      cv = &pr->onexit;
-      wv = &pr->waiters;
-    } else {
-      return echild();
+    __proc_unlock();
+
+    // perform blocking operation
+    uint32_t wi;
+    uintptr_t sem;
+    struct PosixThread *pt = _pthread_self();
+    pt->pt_blkmask = waitmask;
+    pt->pt_semaphore = sem = CreateSemaphore(0, 0, 1, 0);
+    atomic_store_explicit(&pt->pt_blocker, PT_BLOCKER_SEM,
+                          memory_order_release);
+    wi = WaitForMultipleObjects(2, (intptr_t[2]){hWaitObject, sem}, 0, -1u);
+    atomic_store_explicit(&pt->pt_blocker, 0, memory_order_release);
+    CloseHandle(sem);
+
+    // log warning if handle unexpectedly closed
+    if (wi & kNtWaitAbandoned) {
+      wi &= ~kNtWaitAbandoned;
+      STRACE("wait4 abandoned %u", wi);
     }
-  }
 
-  // wait for status change
-  if (options & WNOHANG) return 0;
-WaitMore:
-  deadline = GetNextDeadline(deadline);
-SpuriousWakeup:
-  ++*wv;
-  pthread_cleanup_push(UnwindWaiterCount, wv);
-  m = __sig_begin(waitmask);
-  if ((rc = _check_signal(true)) != -1) {
-    rc = nsync_cv_wait_with_deadline(cv, &__proc.lock, deadline, 0);
+    // check for wait() style wakeup
+    __proc_lock();
+    if (!wi && !pr) {
+      --__proc.waiters;
+      goto CheckForZombies;
+    }
+
+    // check if killed or win32 error
+    if (wi) {
+      if (pr && --pr->waiters && pr->status == PROC_UNDEAD) {
+        __proc_free(pr);
+      }
+      __proc_unlock();
+      if (wi == 1) {
+        // __sig_cancel() woke our semaphore
+        continue;
+      } else {
+        // neither posix or win32 define i/o error conditions for
+        // generic wait. failure should only be due to api misuse
+        return einval();
+      }
+    }
+
+    // handle process exit notification
+    --pr->waiters;
+    if (pr->status == PROC_ALIVE) {
+      __proc_harvest(pr, true);
+    }
+    switch (pr->status) {
+      case PROC_ALIVE:
+        // exit caused by execve() reparenting
+        __proc_unlock();
+        break;
+      case PROC_ZOMBIE:
+        // exit happened and we're the first to know
+        rc = __proc_reap(pr, wstatus, rusage);
+        __proc_unlock();
+        return rc;
+      case PROC_UNDEAD:
+        // exit happened but another thread waited first
+        if (!pr->waiters) {
+          __proc_free(pr);
+        }
+        __proc_unlock();
+        return echild();
+      default:
+        __builtin_unreachable();
+    }
   }
-  __sig_finish(m);
-  pthread_cleanup_pop(true);
-  if (rc == -1) return -1;
-  if (rc == ETIMEDOUT) goto WaitMore;
-  if (rc == ECANCELED) return ecanceled();
-  if (pr && pr->iszombie) return ReapZombie(pr, wstatus, rusage);
-  unassert(!rc);  // i have to follow my dreams however crazy they seem
-  if (!pr && (rc = CheckZombies(pid, wstatus, rusage))) return rc;
-  goto SpuriousWakeup;
 }
 
 textwindows int sys_wait4_nt(int pid, int *opt_out_wstatus, int options,
                              struct rusage *opt_out_rusage) {
-  int rc;
-  uint64_t m;
   // no support for WCONTINUED and WUNTRACED yet
   if (options & ~WNOHANG) return einval();
   // XXX: NT doesn't really have process groups. For instance the
@@ -156,12 +208,9 @@ textwindows int sys_wait4_nt(int pid, int *opt_out_wstatus, int options,
   //      just does an "ignore ctrl-c" internally.
   if (pid == 0) pid = -1;
   if (pid < -1) pid = -pid;
-  m = __sig_block();
-  __proc_lock();
-  pthread_cleanup_push((void *)__proc_unlock, 0);
-  rc = WaitForProcess(pid, opt_out_wstatus, options, opt_out_rusage,
-                      m | 1ull << (SIGCHLD - 1));
-  pthread_cleanup_pop(true);
+  sigset_t m = __sig_block();
+  int rc = __proc_wait(pid, opt_out_wstatus, options, opt_out_rusage,
+                       m | 1ull << (SIGCHLD - 1));
   __sig_unblock(m);
   return rc;
 }
diff --git a/libc/runtime/enable_tls.c b/libc/runtime/enable_tls.c
index ccc96fb990e..54dafe1537b 100644
--- a/libc/runtime/enable_tls.c
+++ b/libc/runtime/enable_tls.c
@@ -24,6 +24,7 @@
 #include "libc/intrin/asancodes.h"
 #include "libc/intrin/atomic.h"
 #include "libc/intrin/dll.h"
+#include "libc/intrin/getenv.internal.h"
 #include "libc/intrin/weaken.h"
 #include "libc/macros.internal.h"
 #include "libc/nt/files.h"
@@ -48,6 +49,18 @@ extern unsigned char __tls_add_nt_rax[];
 
 _Alignas(TLS_ALIGNMENT) static char __static_tls[6016];
 
+static unsigned long ParseMask(const char *str) {
+  int c;
+  unsigned long x = 0;
+  if (str) {
+    while ((c = *str++)) {
+      x *= 10;
+      x += c - '0';
+    }
+  }
+  return x;
+}
+
 /**
  * Enables thread local storage for main process.
  *
@@ -210,6 +223,13 @@ textstartup void __enable_tls(void) {
   atomic_store_explicit(&tib->tib_tid, tid, memory_order_relaxed);
   // TODO(jart): set_tid_address?
 
+  // inherit signal mask
+  if (IsWindows()) {
+    atomic_store_explicit(&tib->tib_sigmask,
+                          ParseMask(__getenv(environ, "_MASK").s),
+                          memory_order_relaxed);
+  }
+
   // initialize posix threads
   _pthread_static.tib = tib;
   _pthread_static.pt_flags = PT_STATIC;
diff --git a/libc/sock/accept-nt.c b/libc/sock/accept-nt.c
index aaa870d4822..d1aa29cd93b 100644
--- a/libc/sock/accept-nt.c
+++ b/libc/sock/accept-nt.c
@@ -109,7 +109,7 @@ textwindows int sys_accept_nt(struct Fd *f, struct sockaddr_storage *addr,
   if ((resources.handle = WSASocket(f->family, f->type, f->protocol, 0, 0,
                                     kNtWsaFlagOverlapped)) == -1) {
     client = __winsockerr();
-    goto WeFailed;
+    goto Finish;
   }
 
   // accept network connection
@@ -118,7 +118,10 @@ textwindows int sys_accept_nt(struct Fd *f, struct sockaddr_storage *addr,
   ssize_t bytes_received = __winsock_block(
       resources.handle, 0, !!(f->flags & O_NONBLOCK), f->rcvtimeo, m,
       sys_accept_nt_start, &(struct AcceptArgs){f->handle, &buffer});
-  if (bytes_received == -1) goto WeFailed;
+  if (bytes_received == -1) {
+    __imp_closesocket(resources.handle);
+    goto Finish;
+  }
 
   // create file descriptor for new socket
   // don't inherit the file open mode bits
@@ -138,7 +141,7 @@ textwindows int sys_accept_nt(struct Fd *f, struct sockaddr_storage *addr,
   memcpy(addr, &buffer.remote.addr, sizeof(*addr));
   g_fds.p[client].kind = kFdSocket;
 
-WeFailed:
+Finish:
   pthread_cleanup_pop(false);
   __sig_unblock(m);
   if (client == -1 && errno == ECONNRESET) {
diff --git a/libc/sock/recv-nt.c b/libc/sock/recv-nt.c
index 8c728255332..8d812d17e7a 100644
--- a/libc/sock/recv-nt.c
+++ b/libc/sock/recv-nt.c
@@ -25,8 +25,14 @@
 #include "libc/sock/internal.h"
 #include "libc/sock/syscall_fd.internal.h"
 #include "libc/sysv/consts/o.h"
+#include "libc/sysv/errfuns.h"
 #ifdef __x86_64__
 
+#define _MSG_OOB      1
+#define _MSG_PEEK     2
+#define _MSG_WAITALL  8
+#define _MSG_DONTWAIT 64
+
 struct RecvArgs {
   const struct iovec *iov;
   size_t iovlen;
@@ -44,11 +50,17 @@ static textwindows int sys_recv_nt_start(int64_t handle,
 
 textwindows ssize_t sys_recv_nt(int fd, const struct iovec *iov, size_t iovlen,
                                 uint32_t flags) {
+  if (flags & ~(_MSG_DONTWAIT | _MSG_OOB | _MSG_PEEK | _MSG_WAITALL)) {
+    return einval();
+  }
   ssize_t rc;
   struct Fd *f = g_fds.p + fd;
   sigset_t m = __sig_block();
-  rc = __winsock_block(f->handle, flags, !!(f->flags & O_NONBLOCK), f->rcvtimeo,
-                       m, sys_recv_nt_start, &(struct RecvArgs){iov, iovlen});
+  bool nonblock = !(flags & _MSG_WAITALL) &&
+                  ((f->flags & O_NONBLOCK) || (flags & _MSG_DONTWAIT));
+  flags &= ~_MSG_DONTWAIT;
+  rc = __winsock_block(f->handle, flags, nonblock, f->rcvtimeo, m,
+                       sys_recv_nt_start, &(struct RecvArgs){iov, iovlen});
   __sig_unblock(m);
   return rc;
 }
diff --git a/libc/sock/recv.c b/libc/sock/recv.c
index 39dd14f2ca8..10dbe7473a1 100644
--- a/libc/sock/recv.c
+++ b/libc/sock/recv.c
@@ -33,7 +33,7 @@
  * @param fd is the file descriptor returned by socket()
  * @param buf is where received network data gets copied
  * @param size is the byte capacity of buf
- * @param flags can have MSG_{WAITALL,PEEK,OOB}, etc.
+ * @param flags can have `MSG_OOB`, `MSG_PEEK`, `MSG_DONTWAIT`, `MSG_WAITALL`
  * @return number of bytes received, 0 on remote close, or -1 w/ errno
  * @error EINTR, EHOSTUNREACH, ECONNRESET (UDP ICMP Port Unreachable),
  *     EPIPE (if MSG_NOSIGNAL), EMSGSIZE, ENOTSOCK, EFAULT, etc.
diff --git a/libc/sock/recvfrom-nt.c b/libc/sock/recvfrom-nt.c
index c7170767236..ab2dbd5cd36 100644
--- a/libc/sock/recvfrom-nt.c
+++ b/libc/sock/recvfrom-nt.c
@@ -24,9 +24,15 @@
 #include "libc/nt/winsock.h"
 #include "libc/sock/internal.h"
 #include "libc/sock/syscall_fd.internal.h"
+#include "libc/sysv/consts/msg.h"
 #include "libc/sysv/consts/o.h"
+#include "libc/sysv/errfuns.h"
 #ifdef __x86_64__
 
+#define _MSG_OOB      1
+#define _MSG_PEEK     2
+#define _MSG_DONTWAIT 64
+
 struct RecvFromArgs {
   const struct iovec *iov;
   size_t iovlen;
@@ -48,11 +54,14 @@ textwindows ssize_t sys_recvfrom_nt(int fd, const struct iovec *iov,
                                     size_t iovlen, uint32_t flags,
                                     void *opt_out_srcaddr,
                                     uint32_t *opt_inout_srcaddrsize) {
+  if (flags & ~(_MSG_DONTWAIT | _MSG_OOB | _MSG_PEEK)) return einval();
   ssize_t rc;
   struct Fd *f = g_fds.p + fd;
   sigset_t m = __sig_block();
-  rc = __winsock_block(f->handle, flags, !!(f->flags & O_NONBLOCK), f->rcvtimeo,
-                       m, sys_recvfrom_nt_start,
+  bool nonblock = (f->flags & O_NONBLOCK) || (flags & _MSG_DONTWAIT);
+  flags &= ~_MSG_DONTWAIT;
+  rc = __winsock_block(f->handle, flags, nonblock, f->rcvtimeo, m,
+                       sys_recvfrom_nt_start,
                        &(struct RecvFromArgs){iov, iovlen, opt_out_srcaddr,
                                               opt_inout_srcaddrsize});
   __sig_unblock(m);
diff --git a/libc/sock/recvfrom.c b/libc/sock/recvfrom.c
index 2b096c60d42..65587bd648a 100644
--- a/libc/sock/recvfrom.c
+++ b/libc/sock/recvfrom.c
@@ -34,15 +34,15 @@
 /**
  * Receives data from network.
  *
+ * This function blocks unless MSG_DONTWAIT is passed. In that case, the
+ * non-error EWOULDBLOCK might be returned. It basically means we didn't
+ * wait around to learn an amount of bytes were written that we know in
+ * advance are guaranteed to be atomic.
+ *
  * @param fd is the file descriptor returned by socket()
  * @param buf is where received network data gets copied
  * @param size is the byte capacity of buf
- * @param flags is a bitmask which may contain any of the following:
- *     - `MSG_DONTWAIT` to force `O_NONBLOCK` behavior for this call
- *     - `MSG_OOB` is broadly supported (untested by cosmo)
- *     - `MSG_PEEK` is broadly supported (untested by cosmo)
- *     - `MSG_WAITALL` is broadly supported (untested by cosmo)
- *     - `MSG_DONTROUTE` is broadly supported (untested by cosmo)
+ * @param flags can have `MSG_OOB`, `MSG_PEEK`, and `MSG_DONTWAIT`
  * @param opt_out_srcaddr receives the binary ip:port of the data's origin
  * @param opt_inout_srcaddrsize is srcaddr capacity which gets updated
  * @return number of bytes received, 0 on remote close, or -1 w/ errno
diff --git a/libc/sock/send-nt.c b/libc/sock/send-nt.c
index 1fb48c8a3bd..b11f1eba7e5 100644
--- a/libc/sock/send-nt.c
+++ b/libc/sock/send-nt.c
@@ -25,8 +25,13 @@
 #include "libc/sock/internal.h"
 #include "libc/sock/syscall_fd.internal.h"
 #include "libc/sysv/consts/o.h"
+#include "libc/sysv/errfuns.h"
 #ifdef __x86_64__
 
+#define _MSG_OOB       1
+#define _MSG_DONTROUTE 4
+#define _MSG_DONTWAIT  64
+
 struct SendArgs {
   const struct iovec *iov;
   size_t iovlen;
@@ -44,11 +49,14 @@ static textwindows int sys_send_nt_start(int64_t handle,
 
 textwindows ssize_t sys_send_nt(int fd, const struct iovec *iov, size_t iovlen,
                                 uint32_t flags) {
+  if (flags & ~(_MSG_DONTWAIT | _MSG_OOB | _MSG_DONTROUTE)) return einval();
   ssize_t rc;
   struct Fd *f = g_fds.p + fd;
   sigset_t m = __sig_block();
-  rc = __winsock_block(f->handle, flags, !!(f->flags & O_NONBLOCK), f->sndtimeo,
-                       m, sys_send_nt_start, &(struct SendArgs){iov, iovlen});
+  bool nonblock = (f->flags & O_NONBLOCK) || (flags & _MSG_DONTWAIT);
+  flags &= ~_MSG_DONTWAIT;
+  rc = __winsock_block(f->handle, flags, nonblock, f->sndtimeo, m,
+                       sys_send_nt_start, &(struct SendArgs){iov, iovlen});
   __sig_unblock(m);
   return rc;
 }
diff --git a/libc/sock/send.c b/libc/sock/send.c
index 31d871953d0..fbc3b7eebd7 100644
--- a/libc/sock/send.c
+++ b/libc/sock/send.c
@@ -34,7 +34,7 @@
  * @param fd is the file descriptor returned by socket()
  * @param buf is the data to send, which we'll copy if necessary
  * @param size is the byte-length of buf
- * @param flags MSG_OOB, MSG_DONTROUTE, MSG_PARTIAL, MSG_NOSIGNAL, etc.
+ * @param flags can have `MSG_OOB`, `MSG_DONTROUTE`, and `MSG_DONTWAIT`
  * @return number of bytes transmitted, or -1 w/ errno
  * @error EINTR, EHOSTUNREACH, ECONNRESET (UDP ICMP Port Unreachable),
  *     EPIPE (if MSG_NOSIGNAL), EMSGSIZE, ENOTSOCK, EFAULT, etc.
diff --git a/libc/sock/sendto-nt.c b/libc/sock/sendto-nt.c
index bd6c2a3c9a9..1d2325616a9 100644
--- a/libc/sock/sendto-nt.c
+++ b/libc/sock/sendto-nt.c
@@ -25,8 +25,13 @@
 #include "libc/sock/internal.h"
 #include "libc/sock/syscall_fd.internal.h"
 #include "libc/sysv/consts/o.h"
+#include "libc/sysv/errfuns.h"
 #ifdef __x86_64__
 
+#define _MSG_OOB       1
+#define _MSG_DONTROUTE 4
+#define _MSG_DONTWAIT  64
+
 struct SendToArgs {
   const struct iovec *iov;
   size_t iovlen;
@@ -47,12 +52,14 @@ static textwindows int sys_sendto_nt_start(int64_t handle,
 textwindows ssize_t sys_sendto_nt(int fd, const struct iovec *iov,
                                   size_t iovlen, uint32_t flags,
                                   void *opt_in_addr, uint32_t in_addrsize) {
+  if (flags & ~(_MSG_DONTWAIT | _MSG_OOB | _MSG_DONTROUTE)) return einval();
   ssize_t rc;
   struct Fd *f = g_fds.p + fd;
   sigset_t m = __sig_block();
+  bool nonblock = (f->flags & O_NONBLOCK) || (flags & _MSG_DONTWAIT);
+  flags &= ~_MSG_DONTWAIT;
   rc = __winsock_block(
-      f->handle, flags, !!(f->flags & O_NONBLOCK), f->sndtimeo, m,
-      sys_sendto_nt_start,
+      f->handle, flags, nonblock, f->sndtimeo, m, sys_sendto_nt_start,
       &(struct SendToArgs){iov, iovlen, opt_in_addr, in_addrsize});
   __sig_unblock(m);
   return rc;
diff --git a/libc/sock/sendto.c b/libc/sock/sendto.c
index 3206be8b12d..a740c714c95 100644
--- a/libc/sock/sendto.c
+++ b/libc/sock/sendto.c
@@ -43,7 +43,7 @@
  * @param fd is the file descriptor returned by socket()
  * @param buf is the data to send, which we'll copy if necessary
  * @param size is the byte-length of buf
- * @param flags MSG_OOB, MSG_DONTROUTE, MSG_PARTIAL, MSG_NOSIGNAL, etc.
+ * @param flags can have `MSG_OOB`, `MSG_DONTROUTE`, and `MSG_DONTWAIT`
  * @param opt_addr is a binary ip:port destination override, which is
  *     mandatory for UDP if connect() wasn't called
  * @param addrsize is the byte-length of addr's true polymorphic form
diff --git a/libc/sock/winsockblock.c b/libc/sock/winsockblock.c
index c519669871b..522d9245e05 100644
--- a/libc/sock/winsockblock.c
+++ b/libc/sock/winsockblock.c
@@ -16,50 +16,21 @@
 │ TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR             │
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
-#include "libc/assert.h"
 #include "libc/calls/internal.h"
 #include "libc/calls/sig.internal.h"
-#include "libc/calls/struct/fd.internal.h"
-#include "libc/calls/struct/sigset.internal.h"
+#include "libc/calls/struct/sigset.h"
 #include "libc/errno.h"
-#include "libc/intrin/atomic.h"
-#include "libc/intrin/kprintf.h"
-#include "libc/intrin/strace.internal.h"
 #include "libc/nt/enum/wait.h"
 #include "libc/nt/errors.h"
-#include "libc/nt/runtime.h"
-#include "libc/nt/struct/iovec.h"
 #include "libc/nt/struct/overlapped.h"
 #include "libc/nt/thread.h"
 #include "libc/nt/winsock.h"
-#include "libc/runtime/runtime.h"
 #include "libc/sock/internal.h"
-#include "libc/sock/syscall_fd.internal.h"
-#include "libc/sysv/consts/o.h"
 #include "libc/sysv/consts/sicode.h"
 #include "libc/sysv/errfuns.h"
 #include "libc/thread/posixthread.internal.h"
 #ifdef __x86_64__
 
-struct WinsockBlockResources {
-  int64_t handle;
-  struct NtOverlapped *overlap;
-};
-
-static void UnwindWinsockBlock(void *arg) {
-  struct WinsockBlockResources *wbr = arg;
-  uint32_t got, flags;
-  CancelIoEx(wbr->handle, wbr->overlap);
-  WSAGetOverlappedResult(wbr->handle, wbr->overlap, &got, true, &flags);
-  WSACloseEvent(wbr->overlap->hEvent);
-}
-
-static void CancelWinsockBlock(int64_t handle, struct NtOverlapped *overlap) {
-  if (!CancelIoEx(handle, overlap)) {
-    unassert(WSAGetLastError() == kNtErrorNotFound);
-  }
-}
-
 textwindows ssize_t
 __winsock_block(int64_t handle, uint32_t flags, bool nonblock,
                 uint32_t srwtimeout, sigset_t waitmask,
@@ -67,31 +38,20 @@ __winsock_block(int64_t handle, uint32_t flags, bool nonblock,
                                   uint32_t *flags, void *arg),
                 void *arg) {
 
-  int rc;
-  int sig = 0;
-  uint32_t status;
-  uint32_t exchanged;
-  int olderror = errno;
-  bool eagained = false;
-  bool canceled = false;
-  int handler_was_called;
-  struct PosixThread *pt;
-
 RestartOperation:
+  int rc, sig, reason = 0;
+  uint32_t status, exchanged;
+  if (_check_cancel() == -1) return -1;  // ECANCELED
+  if ((sig = __sig_get(waitmask))) goto HandleInterrupt;
+
   struct NtOverlapped overlap = {.hEvent = WSACreateEvent()};
-  struct WinsockBlockResources wbr = {handle, &overlap};
-  pthread_cleanup_push(UnwindWinsockBlock, &wbr);
   rc = StartSocketOp(handle, &overlap, &flags, arg);
   if (rc && WSAGetLastError() == kNtErrorIoPending) {
     if (nonblock) {
-      CancelWinsockBlock(handle, &overlap);
-      eagained = true;
-    } else if (_check_cancel()) {
-      CancelWinsockBlock(handle, &overlap);
-      canceled = true;
-    } else if ((sig = __sig_get(waitmask))) {
-      CancelWinsockBlock(handle, &overlap);
+      CancelIoEx(handle, &overlap);
+      reason = EAGAIN;
     } else {
+      struct PosixThread *pt;
       pt = _pthread_self();
       pt->pt_blkmask = waitmask;
       pt->pt_iohandle = handle;
@@ -101,10 +61,13 @@ __winsock_block(int64_t handle, uint32_t flags, bool nonblock,
       status = WSAWaitForMultipleEvents(1, &overlap.hEvent, 0,
                                         srwtimeout ? srwtimeout : -1u, 0);
       atomic_store_explicit(&pt->pt_blocker, 0, memory_order_release);
-      if (status == kNtWaitTimeout) {
-        // SO_RCVTIMEO or SO_SNDTIMEO elapsed
-        CancelWinsockBlock(handle, &overlap);
-        eagained = true;
+      if (status) {
+        if (status == kNtWaitTimeout) {
+          reason = EAGAIN;  // SO_RCVTIMEO or SO_SNDTIMEO elapsed
+        } else {
+          reason = WSAGetLastError();  // ENETDOWN or ENOBUFS
+        }
+        CancelIoEx(handle, &overlap);
       }
     }
     rc = 0;
@@ -114,30 +77,21 @@ __winsock_block(int64_t handle, uint32_t flags, bool nonblock,
              ? 0
              : -1;
   }
-  pthread_cleanup_pop(false);
   WSACloseEvent(overlap.hEvent);
 
-  if (canceled) {
-    return ecanceled();
-  }
-  if (sig) {
-    handler_was_called = __sig_relay(sig, SI_KERNEL, waitmask);
-    if (_check_cancel() == -1) return -1;
-  } else {
-    handler_was_called = 0;
-  }
   if (!rc) {
-    errno = olderror;
     return exchanged;
   }
   if (WSAGetLastError() == kNtErrorOperationAborted) {
-    if (eagained) return eagain();
-    if (!handler_was_called && (sig = __sig_get(waitmask))) {
-      handler_was_called = __sig_relay(sig, SI_KERNEL, waitmask);
-      if (_check_cancel() == -1) return -1;
+    if (reason) {
+      errno = reason;
+      return -1;
     }
-    if (handler_was_called != 1) {
-      goto RestartOperation;
+    if ((sig = __sig_get(waitmask))) {
+    HandleInterrupt:
+      int handler_was_called = __sig_relay(sig, SI_KERNEL, waitmask);
+      if (_check_cancel() == -1) return -1;
+      if (handler_was_called != 1) goto RestartOperation;
     }
     return eintr();
   }
diff --git a/libc/stdio/printargs.c b/libc/stdio/printargs.c
index 476c457a3e5..1565111a202 100644
--- a/libc/stdio/printargs.c
+++ b/libc/stdio/printargs.c
@@ -309,9 +309,9 @@ textstartup void __printargs(const char *prologue) {
     PRINT("");
     PRINT("SIGNAL MASK %#lx", ss);
     if (ss) {
-      for (i = 0; i < 32; ++i) {
-        if (ss & (1u << i)) {
-          PRINT(" ☼ %G (%d) is masked", i + 1, i + 1);
+      for (i = 1; i <= NSIG; ++i) {
+        if (ss & (1ull << (i - 1))) {
+          PRINT(" ☼ %G (%d) is masked", i, i);
         }
       }
     } else {
diff --git a/libc/str/str.h b/libc/str/str.h
index 7891a647a3e..797db1e93d5 100644
--- a/libc/str/str.h
+++ b/libc/str/str.h
@@ -56,17 +56,14 @@ int iswseparator(wint_t);
 wint_t towlower(wint_t);
 wint_t towupper(wint_t);
 
-void bzero(void *, size_t) memcpyesque;
 void *memset(void *, int, size_t) memcpyesque;
 void *memmove(void *, const void *, size_t) memcpyesque;
 void *memcpy(void *restrict, const void *restrict, size_t) memcpyesque;
 void *mempcpy(void *restrict, const void *restrict, size_t) memcpyesque;
 char *hexpcpy(char *restrict, const void *restrict, size_t) memcpyesque;
 void *memccpy(void *restrict, const void *restrict, int, size_t) memcpyesque;
-void bcopy(const void *, void *, size_t) memcpyesque;
 void explicit_bzero(void *, size_t);
 
-int bcmp(const void *, const void *, size_t) strlenesque;
 int memcmp(const void *, const void *, size_t) strlenesque;
 int timingsafe_bcmp(const void *, const void *, size_t);
 int timingsafe_memcmp(const void *, const void *, size_t);
@@ -75,7 +72,6 @@ size_t strlen(const char *) strlenesque;
 size_t strnlen(const char *, size_t) strlenesque;
 size_t strnlen_s(const char *, size_t);
 char *strchr(const char *, int) strlenesque;
-char *index(const char *, int) strlenesque;
 void *memchr(const void *, int, size_t) strlenesque;
 char *strchrnul(const char *, int) strlenesque returnsnonnull;
 void *rawmemchr(const void *, int) strlenesque returnsnonnull;
@@ -220,6 +216,17 @@ bool wcsendswith(const wchar_t *, const wchar_t *) strlenesque;
 char *__join_paths(char *, size_t, const char *, const char *) __wur;
 #endif /* _COSMO_SOURCE */
 
+#if defined(_COSMO_SOURCE) || defined(_GNU_SOURCE) || defined(_BSD_SOURCE) || \
+    defined(_POSIX_SOURCE) ||                                                 \
+    (defined(_POSIX_C_SOURCE) && _POSIX_C_SOURCE + 0 < 200809L) ||            \
+    (defined(_XOPEN_SOURCE) && _XOPEN_SOURCE + 0 < 700)
+int bcmp(const void *, const void *, size_t) strlenesque;
+void bcopy(const void *, void *, size_t) memcpyesque;
+void bzero(void *, size_t) memcpyesque;
+char *index(const char *, int) strlenesque;
+char *rindex(const char *, int) strlenesque;
+#endif
+
 COSMOPOLITAN_C_END_
 #endif /* !(__ASSEMBLER__ + __LINKER__ + 0) */
 #endif /* COSMOPOLITAN_LIBC_STR_STR_H_ */
diff --git a/libc/thread/pthread_create.c b/libc/thread/pthread_create.c
index 5fece617d62..f7c436d1ae4 100644
--- a/libc/thread/pthread_create.c
+++ b/libc/thread/pthread_create.c
@@ -115,7 +115,12 @@ static int PosixThread(void *arg, int tid) {
   }
   // set long jump handler so pthread_exit can bring control back here
   if (!setjmp(pt->pt_exiter)) {
-    pthread_sigmask(SIG_SETMASK, &pt->pt_attr.__sigmask, 0);
+    if (IsWindows()) {
+      atomic_store_explicit(&__get_tls()->tib_sigmask, pt->pt_attr.__sigmask,
+                            memory_order_release);
+    } else {
+      sys_sigprocmask(SIG_SETMASK, &pt->pt_attr.__sigmask, 0);
+    }
     rc = pt->pt_start(pt->pt_arg);
     // ensure pthread_cleanup_pop(), and pthread_exit() popped cleanup
     unassert(!pt->pt_cleanup);
diff --git a/test/libc/calls/stackoverflow1_test.c b/test/libc/calls/stackoverflow1_test.c
index d0205797897..6bf8e0e0b0f 100644
--- a/test/libc/calls/stackoverflow1_test.c
+++ b/test/libc/calls/stackoverflow1_test.c
@@ -27,6 +27,7 @@
 #include "libc/mem/mem.h"
 #include "libc/runtime/runtime.h"
 #include "libc/runtime/sysconf.h"
+#include "libc/stdio/sysparam.h"
 #include "libc/sysv/consts/rlimit.h"
 #include "libc/sysv/consts/sa.h"
 #include "libc/sysv/consts/sig.h"
@@ -58,8 +59,10 @@ void SetUp(void) {
   // tune down the main process's stack size to a reasonable amount
   // some operating systems, e.g. freebsd, will do things like have
   // 500mb RLIMIT_STACK by default, even on machines with 400mb RAM
-  struct rlimit rl = {2 * 1024 * 1024, 2 * 1024 * 1024};
   if (!IsWindows() && !IsXnu()) {
+    struct rlimit rl;
+    getrlimit(RLIMIT_STACK, &rl);
+    rl.rlim_cur = MIN(rl.rlim_cur, 2 * 1024 * 1024);
     ASSERT_SYS(0, 0, setrlimit(RLIMIT_STACK, &rl));
   }
 
diff --git a/examples/stackoverflow.c b/test/libc/calls/stackoverflow3_test.c
similarity index 56%
rename from examples/stackoverflow.c
rename to test/libc/calls/stackoverflow3_test.c
index 4e3ad91fa7a..c5dc6c65761 100644
--- a/examples/stackoverflow.c
+++ b/test/libc/calls/stackoverflow3_test.c
@@ -1,12 +1,21 @@
-#if 0
-/*─────────────────────────────────────────────────────────────────╗
-│ To the extent possible under law, Justine Tunney has waived      │
-│ all copyright and related or neighboring rights to this file,    │
-│ as it is written in the following disclaimers:                   │
-│   • http://unlicense.org/                                        │
-│   • http://creativecommons.org/publicdomain/zero/1.0/            │
-╚─────────────────────────────────────────────────────────────────*/
-#endif
+/*-*- mode:c;indent-tabs-mode:nil;c-basic-offset:2;tab-width:8;coding:utf-8 -*-│
+│vi: set net 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/calls/struct/sigaction.h"
 #include "libc/calls/struct/sigaltstack.h"
 #include "libc/calls/struct/siginfo.h"
@@ -29,10 +38,6 @@
  * rewrite thread cpu state to call pthread_exit
  * this method returns gracefully from signal handlers
  * unfortunately it relies on cpu architecture knowledge
- *
- * @see test/libc/thread/stackoverflow1_test.c
- * @see test/libc/thread/stackoverflow2_test.c
- * @see test/libc/thread/stackoverflow3_test.c
  */
 
 volatile bool smashed_stack;
@@ -47,11 +52,12 @@ void Exiter(void *rc) {
 void CrashHandler(int sig, siginfo_t *si, void *arg) {
   ucontext_t *ctx = arg;
   struct sigaltstack ss;
+  ASSERT_FALSE(smashed_stack);
   ASSERT_SYS(0, 0, sigaltstack(0, &ss));
   ASSERT_EQ(SS_ONSTACK, ss.ss_flags);
   kprintf("kprintf avoids overflowing %G %p\n", si->si_signo, si->si_addr);
   smashed_stack = true;
-  ASSERT_TRUE(__is_stack_overflow(si, ctx));
+  // EXPECT_TRUE(__is_stack_overflow(si, ctx));
   //
   // the backtrace will look like this
   //
@@ -67,7 +73,15 @@ void CrashHandler(int sig, siginfo_t *si, void *arg) {
   ctx->uc_mcontext.PC = (long)Exiter;
   ctx->uc_mcontext.SP += 32768;
   ctx->uc_mcontext.SP &= -16;
-  ctx->uc_mcontext.SP -= 8;
+#ifdef __x86_64__
+  *(long *)(ctx->uc_mcontext.SP -= 8) = ctx->uc_mcontext.PC;
+#elif defined(__aarch64__)
+  *(long *)(ctx->uc_mcontext.SP -= 8) = ctx->uc_mcontext.regs[30];
+  *(long *)(ctx->uc_mcontext.SP -= 8) = ctx->uc_mcontext.regs[29];
+  ctx->uc_mcontext.BP = ctx->uc_mcontext.SP;
+#else
+#error "unsupported architecture"
+#endif
 }
 
 int StackOverflow(int f(), int n) {
@@ -96,7 +110,7 @@ void *MyPosixThread(void *arg) {
   return 0;
 }
 
-int main(int argc, char *argv[]) {
+TEST(stackoverflow, standardStack_altStack_thread_teleport) {
   void *res;
   pthread_t th;
   struct sigaltstack ss;
@@ -105,6 +119,7 @@ int main(int argc, char *argv[]) {
   pthread_join(th, &res);
   ASSERT_EQ((void *)123L, res);
   ASSERT_TRUE(smashed_stack);
+  // this should be SS_DISABLE but ShowCrashReports() creates an alt stack
   ASSERT_SYS(0, 0, sigaltstack(0, &ss));
-  ASSERT_EQ(SS_DISABLE, ss.ss_flags);
+  ASSERT_EQ(0, ss.ss_flags);
 }
diff --git a/test/libc/calls/stackoverflow4_test.c b/test/libc/calls/stackoverflow4_test.c
new file mode 100644
index 00000000000..5cd73988f48
--- /dev/null
+++ b/test/libc/calls/stackoverflow4_test.c
@@ -0,0 +1,86 @@
+/*-*- mode:c;indent-tabs-mode:nil;c-basic-offset:2;tab-width:8;coding:utf-8 -*-│
+│vi: set net 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/calls/struct/sigaction.h"
+#include "libc/calls/struct/sigaltstack.h"
+#include "libc/calls/struct/siginfo.h"
+#include "libc/calls/struct/ucontext.internal.h"
+#include "libc/calls/ucontext.h"
+#include "libc/intrin/kprintf.h"
+#include "libc/limits.h"
+#include "libc/mem/gc.internal.h"
+#include "libc/mem/mem.h"
+#include "libc/runtime/runtime.h"
+#include "libc/runtime/sysconf.h"
+#include "libc/sysv/consts/sa.h"
+#include "libc/sysv/consts/sig.h"
+#include "libc/sysv/consts/ss.h"
+#include "libc/testlib/testlib.h"
+#include "libc/thread/thread.h"
+
+/**
+ * stack overflow recovery technique #4
+ * just call pthread_exit() and let the thread die
+ */
+
+volatile bool smashed_stack;
+
+void CrashHandler(int sig) {
+  smashed_stack = true;
+  pthread_exit((void *)123L);
+}
+
+int StackOverflow(int f(), int n) {
+  if (n < INT_MAX) {
+    return f(f, n + 1) - 1;
+  } else {
+    return INT_MAX;
+  }
+}
+
+int (*pStackOverflow)(int (*)(), int) = StackOverflow;
+
+void *MyPosixThread(void *arg) {
+  struct sigaction sa;
+  struct sigaltstack ss;
+  ss.ss_flags = 0;
+  ss.ss_size = sysconf(_SC_MINSIGSTKSZ) + 4096;
+  ss.ss_sp = gc(malloc(ss.ss_size));
+  ASSERT_SYS(0, 0, sigaltstack(&ss, 0));
+  sa.sa_flags = SA_SIGINFO | SA_ONSTACK;  // <-- important
+  sigemptyset(&sa.sa_mask);
+  sa.sa_handler = CrashHandler;
+  sigaction(SIGBUS, &sa, 0);
+  sigaction(SIGSEGV, &sa, 0);
+  exit(pStackOverflow(pStackOverflow, 0));
+  return 0;
+}
+
+TEST(stackoverflow, standardStack_altStack_thread_teleport) {
+  void *res;
+  pthread_t th;
+  struct sigaltstack ss;
+  smashed_stack = false;
+  pthread_create(&th, 0, MyPosixThread, 0);
+  pthread_join(th, &res);
+  ASSERT_EQ((void *)123L, res);
+  ASSERT_TRUE(smashed_stack);
+  // this should be SS_DISABLE but ShowCrashReports() creates an alt stack
+  ASSERT_SYS(0, 0, sigaltstack(0, &ss));
+  ASSERT_EQ(0, ss.ss_flags);
+}
diff --git a/tool/build/compile.c b/tool/build/compile.c
index abdf7b3a2f9..0382b32d388 100644
--- a/tool/build/compile.c
+++ b/tool/build/compile.c
@@ -637,6 +637,8 @@ int Launch(void) {
 
   posix_spawnattr_init(&spawnattr);
   posix_spawnattr_setsigmask(&spawnattr, &savemask);
+  posix_spawnattr_setflags(&spawnattr,
+                           POSIX_SPAWN_SETSIGMASK | POSIX_SPAWN_SETRLIMIT);
   SetCpuLimit(cpuquota);
   SetFszLimit(fszquota);
   SetMemLimit(memquota);
diff --git a/tool/build/runitd.c b/tool/build/runitd.c
index e0e14f12218..0885c4b18e5 100644
--- a/tool/build/runitd.c
+++ b/tool/build/runitd.c
@@ -552,7 +552,8 @@ void *ClientWorker(void *arg) {
   started = timespec_real();
   pipe2(client->pipe, O_CLOEXEC);
   posix_spawnattr_init(&spawnattr);
-  posix_spawnattr_setflags(&spawnattr, POSIX_SPAWN_SETPGROUP);
+  posix_spawnattr_setflags(&spawnattr,
+                           POSIX_SPAWN_SETSIGMASK | POSIX_SPAWN_SETPGROUP);
   posix_spawnattr_setsigmask(&spawnattr, &sigmask);
   posix_spawn_file_actions_init(&spawnfila);
   posix_spawn_file_actions_adddup2(&spawnfila, g_bogusfd, 0);