Put more thought into new signaling code

libc/calls/sig.c

@@ -596,6 +596,9 @@ static textwindows int __sig_console_sig(uint32_t dwCtrlType) {
 }
 
 __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);
@@ -616,7 +619,12 @@ textwindows int __sig_check(void) {
   }
 }
 
-// delivers signals from other processes asynchronously
+// 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));
@@ -624,9 +632,16 @@ textwindows dontinstrument static uint32_t __sig_worker(void *arg) {
   __maps_track((char *)(((uintptr_t)sp + __pagesize - 1) & -__pagesize) - STKSZ,
                STKSZ);
   for (;;) {
-    int sig;
-    if ((sig = __sig_getter(__sig.process, 0)))
+    // 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);
+    }
     Sleep(1);
   }
   return 0;

libc/calls/sig.internal.h

@@ -2,6 +2,7 @@
 #define COSMOPOLITAN_LIBC_CALLS_SIGNALS_INTERNAL_H_
 #include "libc/atomic.h"
 #include "libc/calls/struct/sigset.h"
+#include "libc/nt/thunk/msabi.h"
 #include "libc/thread/posixthread.internal.h"
 
 #define SIG_HANDLED_NO_RESTART 1
@@ -28,8 +29,8 @@ void __sig_delete(int);
 void __sig_generate(int, int);
 void __sig_init(void);
 
-char16_t *__sig_process_path(char16_t *, uint32_t);
-atomic_ulong *__sig_map_process(int, int);
+__msabi char16_t *__sig_process_path(char16_t *, uint32_t, int);
+__msabi atomic_ulong *__sig_map_process(int, int);
 
 COSMOPOLITAN_C_END_
 #endif /* COSMOPOLITAN_LIBC_CALLS_SIGNALS_INTERNAL_H_ */

libc/fmt/internal.h

@@ -2,6 +2,7 @@
 #define COSMOPOLITAN_LIBC_FMT_STRTOL_H_
 #include "libc/ctype.h"
 #include "libc/errno.h"
+#include "libc/nt/thunk/msabi.h"
 #include "libc/str/str.h"
 
 #define CONSUME_SPACES(t, s, c) \
@@ -48,6 +49,6 @@
 int __vcscanf(int (*)(void *), int (*)(int, void *), void *, const char *,
               va_list);
 int __fmt(void *, void *, const char *, va_list, int *);
-char16_t *__itoa16(char16_t[21], uint64_t);
+__msabi char16_t *__itoa16(char16_t[21], uint64_t);
 
 #endif /* COSMOPOLITAN_LIBC_FMT_STRTOL_H_ */

libc/intrin/itoa16.c

@@ -18,7 +18,7 @@
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/fmt/internal.h"
 
-textwindows char16_t *__itoa16(char16_t p[21], uint64_t x) {
+__msabi textwindows char16_t *__itoa16(char16_t p[21], uint64_t x) {
   char t;
   size_t a, b, i = 0;
   do {

libc/intrin/sigproc.c

@@ -21,6 +21,7 @@
 #include "libc/fmt/internal.h"
 #include "libc/nt/createfile.h"
 #include "libc/nt/enum/accessmask.h"
+#include "libc/nt/enum/creationdisposition.h"
 #include "libc/nt/enum/fileflagandattributes.h"
 #include "libc/nt/enum/filemapflags.h"
 #include "libc/nt/enum/filemovemethod.h"
@@ -29,28 +30,60 @@
 #include "libc/nt/files.h"
 #include "libc/nt/memory.h"
 #include "libc/nt/runtime.h"
+#include "libc/nt/thunk/msabi.h"
 #ifdef __x86_64__
 
-textwindows char16_t *__sig_process_path(char16_t *path, uint32_t pid) {
+// cut back on code size and avoid setting errno
+// this code is a mandatory dependency of winmain
+__msabi extern typeof(CloseHandle) *const __imp_CloseHandle;
+__msabi extern typeof(CreateDirectory) *const __imp_CreateDirectoryW;
+__msabi extern typeof(CreateFile) *const __imp_CreateFileW;
+__msabi extern typeof(CreateFileMapping) *const __imp_CreateFileMappingW;
+__msabi extern typeof(MapViewOfFileEx) *const __imp_MapViewOfFileEx;
+__msabi extern typeof(SetEndOfFile) *const __imp_SetEndOfFile;
+__msabi extern typeof(SetFilePointer) *const __imp_SetFilePointer;
+
+__msabi textwindows char16_t *__sig_process_path(char16_t *path, uint32_t pid,
+                                                 int create_directories) {
   char16_t *p = path;
-  *p++ = 'C';
+  *p++ = 'C';  // C:\ProgramData\cosmo\sig\x\y.pid
   *p++ = ':';
   *p++ = '\\';
-  *p++ = 'v';
-  *p++ = 'a';
+  *p++ = 'P';
   *p++ = 'r';
+  *p++ = 'o';
+  *p++ = 'g';
+  *p++ = 'r';
+  *p++ = 'a';
+  *p++ = 'm';
+  *p++ = 'D';
+  *p++ = 'a';
+  *p++ = 't';
+  *p++ = 'a';
   *p = 0;
-  CreateDirectory(path, 0);
+  if (create_directories)
+    __imp_CreateDirectoryW(path, 0);
+  *p++ = '\\';
+  *p++ = 'c';
+  *p++ = 'o';
+  *p++ = 's';
+  *p++ = 'm';
+  *p++ = 'o';
+  *p = 0;
+  if (create_directories)
+    __imp_CreateDirectoryW(path, 0);
   *p++ = '\\';
   *p++ = 's';
   *p++ = 'i';
   *p++ = 'g';
   *p = 0;
-  CreateDirectory(path, 0);
+  if (create_directories)
+    __imp_CreateDirectoryW(path, 0);
   *p++ = '\\';
-  p = __itoa16(p, (pid & 0x000fff00) >> 8);
+  p = __itoa16(p, (pid & 0x000ff800) >> 11);
   *p = 0;
-  CreateDirectory(path, 0);
+  if (create_directories)
+    __imp_CreateDirectoryW(path, 0);
   *p++ = '\\';
   p = __itoa16(p, pid);
   *p++ = '.';
@@ -61,33 +94,25 @@ textwindows char16_t *__sig_process_path(char16_t *path, uint32_t pid) {
   return path;
 }
 
-textwindows static atomic_ulong *__sig_map_process_impl(int pid,
-                                                        int disposition) {
+__msabi textwindows atomic_ulong *__sig_map_process(int pid, int disposition) {
   char16_t path[128];
-  intptr_t hand = CreateFile(__sig_process_path(path, pid),
-                             kNtGenericRead | kNtGenericWrite,
-                             kNtFileShareRead | kNtFileShareWrite, 0,
-                             disposition, kNtFileAttributeNormal, 0);
+  __sig_process_path(path, pid, disposition == kNtOpenAlways);
+  intptr_t hand = __imp_CreateFileW(path, kNtGenericRead | kNtGenericWrite,
+                                    kNtFileShareRead | kNtFileShareWrite, 0,
+                                    disposition, kNtFileAttributeNormal, 0);
   if (hand == -1)
     return 0;
-  SetFilePointer(hand, 8, 0, kNtFileBegin);
-  SetEndOfFile(hand);
-  intptr_t map = CreateFileMapping(hand, 0, kNtPageReadwrite, 0, 8, 0);
+  __imp_SetFilePointer(hand, 8, 0, kNtFileBegin);
+  __imp_SetEndOfFile(hand);
+  intptr_t map = __imp_CreateFileMappingW(hand, 0, kNtPageReadwrite, 0, 8, 0);
   if (!map) {
-    CloseHandle(hand);
+    __imp_CloseHandle(hand);
     return 0;
   }
-  atomic_ulong *sigs = MapViewOfFileEx(map, kNtFileMapWrite, 0, 0, 8, 0);
-  CloseHandle(map);
-  CloseHandle(hand);
+  atomic_ulong *sigs = __imp_MapViewOfFileEx(map, kNtFileMapWrite, 0, 0, 8, 0);
+  __imp_CloseHandle(map);
+  __imp_CloseHandle(hand);
   return sigs;
 }
 
-textwindows atomic_ulong *__sig_map_process(int pid, int disposition) {
-  int e = errno;
-  atomic_ulong *res = __sig_map_process_impl(pid, disposition);
-  errno = e;
-  return res;
-}
-
 #endif /* __x86_64__ */

libc/intrin/terminatethisprocess.c

@@ -27,7 +27,9 @@
 #include "libc/runtime/internal.h"
 #ifdef __x86_64__
 
+__msabi extern typeof(DeleteFile) *const __imp_DeleteFileW;
 __msabi extern typeof(TerminateProcess) *const __imp_TerminateProcess;
+__msabi extern typeof(UnmapViewOfFile) *const __imp_UnmapViewOfFile;
 
 /**
  * Terminates the calling process and all of its threads.
@@ -40,8 +42,8 @@ textwindows dontinstrument void TerminateThisProcess(uint32_t dwWaitStatus) {
   atomic_ulong fake = 0;
   real = __sig.process;
   __sig.process = &fake;
-  UnmapViewOfFile(real);
-  DeleteFile(__sig_process_path(path, __pid));
+  __imp_UnmapViewOfFile(real);
+  __imp_DeleteFileW(__sig_process_path(path, __pid, false));
 
   // "When a process terminates itself, TerminateProcess stops execution
   // of the calling thread and does not return." -Quoth MSDN

libc/proc/fork-nt.c

@@ -465,7 +465,6 @@ textwindows int sys_fork_nt(uint32_t dwCreationFlags) {
     __morph_tls();
     __tls_enabled = true;
     // the child's pending signals is initially empty
-    atomic_store_explicit(__sig.process, 0, memory_order_relaxed);
     atomic_store_explicit(&tib->tib_sigpending, 0, memory_order_relaxed);
     // re-apply code morphing for function tracing
     if (ftrace_stackdigs)

libc/proc/kill-nt.c

@@ -59,11 +59,22 @@ textwindows int sys_kill_nt(int pid, int sig) {
   if (pid <= 0 || pid == getpid()) {
     if (sig) {
       if (pid <= 0) {
+        // if pid is 0 or -1 then kill the processes beneath us too.
+        // this isn't entirely right but it's closer to being right.
+        // having this behavior is helpful for servers like redbean.
         struct Dll *e;
         BLOCK_SIGNALS;
         __proc_lock();
-        for (e = dll_first(__proc.list); e; e = dll_next(__proc.list, e))
-          TerminateProcess(PROC_CONTAINER(e)->handle, sig);
+        for (e = dll_first(__proc.list); e; e = dll_next(__proc.list, e)) {
+          atomic_ulong *sigproc;
+          struct Proc *pr = PROC_CONTAINER(e);
+          if (sig != 9 && (sigproc = __sig_map_process(pid, kNtOpenExisting))) {
+            atomic_fetch_or_explicit(sigproc, 1ull << (sig - 1),
+                                     memory_order_release);
+          } else {
+            TerminateProcess(pr->handle, sig);
+          }
+        }
         __proc_unlock();
         ALLOW_SIGNALS;
       }
@@ -73,7 +84,25 @@ textwindows int sys_kill_nt(int pid, int sig) {
     }
   }
 
-  // attempt to signal via /var/sig shared memory file
+  // find existing handle we own for process
+  //
+  // this step should come first to verify process existence. this is
+  // because there's no guarantee that just because the shared memory
+  // file exists, the process actually exists.
+  int64_t handle, closeme = 0;
+  if (!(handle = __proc_handle(pid))) {
+    if ((handle = OpenProcess(kNtProcessTerminate, false, pid))) {
+      closeme = handle;
+    } else {
+      goto OnError;
+    }
+  }
+
+  // attempt to signal via shared memory file
+  //
+  // now that we know the process exists, if it has a shared memory file
+  // then we can be reasonably certain it's a cosmo process which should
+  // be trusted to deliver its signal, unless it's a nine exterminations
   if (pid > 0 && sig != 9) {
     atomic_ulong *sigproc;
     if ((sigproc = __sig_map_process(pid, kNtOpenExisting))) {
@@ -81,20 +110,12 @@ textwindows int sys_kill_nt(int pid, int sig) {
         atomic_fetch_or_explicit(sigproc, 1ull << (sig - 1),
                                  memory_order_release);
       UnmapViewOfFile(sigproc);
+      if (closeme)
+        CloseHandle(closeme);
       return 0;
     }
   }
 
-  // find existing handle we own for process
-  int64_t handle, closeme = 0;
-  if (!(handle = __proc_handle(pid))) {
-    if ((handle = OpenProcess(kNtProcessTerminate, false, pid))) {
-      closeme = handle;
-    } else {
-      goto OnError;
-    }
-  }
-
   // perform actual kill
   // process will report WIFSIGNALED with WTERMSIG(sig)
   bool32 ok = TerminateProcess(handle, sig);

libc/proc/kill.c

@@ -29,6 +29,17 @@
  * The impact of this action can be terminating the process, or
  * interrupting it to request something happen.
  *
+ * On Windows, signals are delivered between processes using shared
+ * memory files stored in C:\ProgramData\cosmo\sig\x\y.pid which hold
+ * the process signal mask. Any process that can access these files can
+ * signal a cosmo process. The targeting process will then notice that a
+ * signal has been added and delivers to any thread as soon as possible.
+ *
+ * On Windows, the concept of a process group isn't fully implemented.
+ * Saying `kill(0, sig)` will deliver `sig` to all direct descendent
+ * processes. Saying `kill(-pid, sig)` will be the same as saying
+ * `kill(pid, sig)`.
+ *
  * @param pid can be:
  *      >0 signals one process by id
  *      =0 signals all processes in current process group

libc/proc/proc.c

@@ -323,6 +323,7 @@ textwindows int64_t __proc_search(int pid) {
   int64_t handle = 0;
   BLOCK_SIGNALS;
   __proc_lock();
+  // TODO(jart): we should increment a reference count when returning
   for (e = dll_first(__proc.list); e; e = dll_next(__proc.list, e)) {
     if (pid == PROC_CONTAINER(e)->pid) {
       handle = PROC_CONTAINER(e)->handle;