--- IO-AIO/AIO.xs	2006/10/24 03:17:39	1.67
+++ IO-AIO/AIO.xs	2006/10/24 17:22:17	1.72
@@ -1,3 +1,6 @@
+/* solaris */
+#define _POSIX_PTHREAD_SEMANTICS 1
+
 #if __linux
 # define _GNU_SOURCE
 #endif
@@ -58,12 +61,13 @@
 #define AIO_BUFSIZE 65536
 
 #define dBUF	 				\
-  char *aio_buf = malloc (AIO_BUFSIZE);		\
+  char *aio_buf;				\
+  LOCK (wrklock);				\
+  self->dbuf = aio_buf = malloc (AIO_BUFSIZE);	\
+  UNLOCK (wrklock);				\
   if (!aio_buf)					\
     return -1;
 
-#define fBUF free (aio_buf)
-
 enum {
   REQ_QUIT,
   REQ_OPEN, REQ_CLOSE,
@@ -75,7 +79,7 @@
   REQ_READDIR,
   REQ_LINK, REQ_SYMLINK,
   REQ_GROUP, REQ_NOP,
-  REQ_SLEEP,
+  REQ_BUSY,
 };
 
 #define AIO_REQ_KLASS "IO::AIO::REQ"
@@ -135,6 +139,49 @@
 # define AIO_MUTEX_INIT PTHREAD_MUTEX_INITIALIZER
 #endif
 
+#define LOCK(mutex)   pthread_mutex_lock   (&(mutex))
+#define UNLOCK(mutex) pthread_mutex_unlock (&(mutex))
+
+/* worker threasd management */
+static pthread_mutex_t wrklock = AIO_MUTEX_INIT;
+
+typedef struct worker {
+  /* locked by wrklock */
+  struct worker *prev, *next;
+
+  pthread_t tid;
+
+  /* locked by reslock, reqlock or wrklock */
+  aio_req req; /* currently processed request */
+  void *dbuf;
+  DIR *dirp;
+} worker;
+
+static worker wrk_first = { &wrk_first, &wrk_first, 0 };
+
+static void worker_clear (worker *wrk)
+{
+  if (wrk->dirp)
+    {
+      closedir (wrk->dirp);
+      wrk->dirp = 0;
+    }
+
+  if (wrk->dbuf)
+    {
+      free (wrk->dbuf);
+      wrk->dbuf = 0;
+    }
+}
+
+static void worker_free (worker *wrk)
+{
+  wrk->next->prev = wrk->prev;
+  wrk->prev->next = wrk->next;
+
+  free (wrk);
+}
+
 static pthread_mutex_t reslock = AIO_MUTEX_INIT;
 static pthread_mutex_t reqlock = AIO_MUTEX_INIT;
 static pthread_cond_t  reqwait = PTHREAD_COND_INITIALIZER;
@@ -195,6 +242,7 @@
 
 static void req_invoke (aio_req req);
 static void req_free (aio_req req);
+static void req_cancel (aio_req req);
 
 /* must be called at most once */
 static SV *req_sv (aio_req req, const char *klass)
@@ -273,12 +321,12 @@
   while (nreqs)
     {
       int size;
-#if !(__i386 || __x86_64) /* safe without sempahore on this archs */
-      pthread_mutex_lock (&reslock);
+#if !(__i386 || __x86_64) /* safe without sempahore on these archs */
+      LOCK (reslock);
 #endif
       size = res_queue.size;
-#if !(__i386 || __x86_64) /* safe without sempahore on this archs */
-      pthread_mutex_unlock (&reslock);
+#if !(__i386 || __x86_64) /* safe without sempahore on these archs */
+      UNLOCK (reslock);
 #endif
 
       if (size)
@@ -312,16 +360,18 @@
 
               if (req->result >= 0)
                 {
+                  int i;
                   char *buf = req->data2ptr;
                   AV *av = newAV ();
 
-                  while (req->result)
+                  av_extend (av, req->result - 1);
+
+                  for (i = 0; i < req->result; ++i)
                     {
                       SV *sv = newSVpv (buf, 0);
 
-                      av_push (av, sv);
+                      av_store (av, i, sv);
                       buf += SvCUR (sv) + 1;
-                      req->result--;
                     }
 
                   rv = sv_2mortal (newRV_noinc ((SV *)av));
@@ -363,7 +413,7 @@
             break;
 
           case REQ_NOP:
-          case REQ_SLEEP:
+          case REQ_BUSY:
             break;
 
           default:
@@ -415,23 +465,31 @@
   SvREFCNT_dec (req->callback);
   Safefree (req->statdata);
 
-  if (req->type == REQ_READDIR && req->result >= 0)
+  if (req->type == REQ_READDIR)
     free (req->data2ptr);
 
   Safefree (req);
 }
 
+static void req_cancel_subs (aio_req grp)
+{
+  aio_req sub;
+
+  if (grp->type != REQ_GROUP)
+    return;
+
+  SvREFCNT_dec (grp->fh2);
+  grp->fh2 = 0;
+
+  for (sub = grp->grp_first; sub; sub = sub->grp_next)
+    req_cancel (sub);
+}
+
 static void req_cancel (aio_req req)
 {
   req->flags |= FLAG_CANCELLED;
 
-  if (req->type == REQ_GROUP)
-    {
-      aio_req sub;
-
-      for (sub = req->grp_first; sub; sub = sub->grp_next)
-        req_cancel (sub);
-    }
+  req_cancel_subs (req);
 }
 
 static int poll_cb ()
@@ -443,7 +501,7 @@
 
   for (;;)
     {
-      pthread_mutex_lock (&reslock);
+      LOCK (reslock);
       req = reqq_shift (&res_queue);
 
       if (req)
@@ -457,7 +515,7 @@
             }
         }
 
-      pthread_mutex_unlock (&reslock);
+      UNLOCK (reslock);
 
       if (!req)
         break;
@@ -501,8 +559,12 @@
 
 static void start_thread (void)
 {
+  worker *wrk = calloc (1, sizeof (worker));
+
+  if (!wrk)
+    croak ("unable to allocate worker thread data");
+
   sigset_t fullsigset, oldsigset;
-  pthread_t tid;
   pthread_attr_t attr;
 
   pthread_attr_init (&attr);
@@ -510,12 +572,23 @@
   pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED);
 
   sigfillset (&fullsigset);
+
+  LOCK (wrklock);
   sigprocmask (SIG_SETMASK, &fullsigset, &oldsigset);
 
-  if (pthread_create (&tid, &attr, aio_proc, 0) == 0)
-    started++;
+  if (pthread_create (&wrk->tid, &attr, aio_proc, (void *)wrk) == 0)
+    {
+      wrk->prev = &wrk_first;
+      wrk->next = wrk_first.next;
+      wrk_first.next->prev = wrk;
+      wrk_first.next = wrk;
+      started++;
+    }
+  else
+    free (wrk);
 
   sigprocmask (SIG_SETMASK, &oldsigset, 0);
+  UNLOCK (wrklock);
 }
 
 static void req_send (aio_req req)
@@ -525,10 +598,10 @@
 
   ++nreqs;
 
-  pthread_mutex_lock (&reqlock);
+  LOCK (reqlock);
   reqq_push (&req_queue, req);
   pthread_cond_signal (&reqwait);
-  pthread_mutex_unlock (&reqlock);
+  UNLOCK (reqlock);
 
   if (nreqs > max_outstanding)
     for (;;)
@@ -611,12 +684,12 @@
   ssize_t res;
   off_t ooffset;
 
-  pthread_mutex_lock (&preadwritelock);
+  LOCK (preadwritelock);
   ooffset = lseek (fd, 0, SEEK_CUR);
   lseek (fd, offset, SEEK_SET);
   res = read (fd, buf, count);
   lseek (fd, ooffset, SEEK_SET);
-  pthread_mutex_unlock (&preadwritelock);
+  UNLOCK (preadwritelock);
 
   return res;
 }
@@ -626,12 +699,12 @@
   ssize_t res;
   off_t ooffset;
 
-  pthread_mutex_lock (&preadwritelock);
+  LOCK (preadwritelock);
   ooffset = lseek (fd, 0, SEEK_CUR);
   lseek (fd, offset, SEEK_SET);
   res = write (fd, buf, count);
   lseek (fd, offset, SEEK_SET);
-  pthread_mutex_unlock (&preadwritelock);
+  UNLOCK (preadwritelock);
 
   return res;
 }
@@ -657,8 +730,6 @@
       count  -= len;
     }
 
-  fBUF;
-
   errno = 0;
 }
 #endif
@@ -673,7 +744,7 @@
   struct dirent *e;
   int errorno;
 
-  pthread_mutex_lock (&readdirlock);
+  LOCK (readdirlock);
 
   e = readdir (dirp);
   errorno = errno;
@@ -686,7 +757,7 @@
   else
     *res = 0;
 
-  pthread_mutex_unlock (&readdirlock);
+  UNLOCK (readdirlock);
 
   errno = errorno;
   return e ? 0 : -1;
@@ -694,7 +765,7 @@
 #endif
 
 /* sendfile always needs emulation */
-static ssize_t sendfile_ (int ofd, int ifd, off_t offset, size_t count)
+static ssize_t sendfile_ (int ofd, int ifd, off_t offset, size_t count, worker *self)
 {
   ssize_t res;
 
@@ -781,15 +852,13 @@
           res    += cnt;
           count  -= cnt;
         }
-
-      fBUF;
     }
 
   return res;
 }
 
 /* read a full directory */
-static int scandir_ (const char *path, void **namesp)
+static void scandir_ (aio_req req, worker *self)
 {
   DIR *dirp;
   union
@@ -804,14 +873,14 @@
   int res = 0;
   int errorno;
 
-  dirp = opendir (path);
-  if (!dirp)
-    return -1;
+  LOCK (wrklock);
+  self->dirp = dirp = opendir (req->dataptr);
+  self->dbuf = u = malloc (sizeof (*u));
+  UNLOCK (wrklock);
 
-  u = malloc (sizeof (*u));
-  names = malloc (memlen);
+  req->data2ptr = names = malloc (memlen);
 
-  if (u && names)
+  if (dirp && u && names)
     for (;;)
       {
         errno = 0;
@@ -831,7 +900,10 @@
             while (memofs + len > memlen)
               {
                 memlen *= 2;
-                names = realloc (names, memlen);
+                LOCK (wrklock);
+                req->data2ptr = names = realloc (names, memlen);
+                UNLOCK (wrklock);
+
                 if (!names)
                   break;
               }
@@ -841,19 +913,10 @@
           }
       }
 
-  errorno = errno;
-  free (u);
-  closedir (dirp);
-
-  if (errorno)
-    {
-      free (names);
-      errno = errorno;
-      res = -1;
-    }
-
-  *namesp = (void *)names;
-  return res;
+  if (errno)
+    res = -1;
+  
+  req->result = res;
 }
 
 /*****************************************************************************/
@@ -862,14 +925,15 @@
 {
   aio_req req;
   int type;
+  worker *self = (worker *)thr_arg;
 
   do
     {
-      pthread_mutex_lock (&reqlock);
+      LOCK (reqlock);
 
       for (;;)
         {
-          req = reqq_shift (&req_queue);
+          self->req = req = reqq_shift (&req_queue);
 
           if (req)
             break;
@@ -877,7 +941,7 @@
           pthread_cond_wait (&reqwait, &reqlock);
         }
 
-      pthread_mutex_unlock (&reqlock);
+      UNLOCK (reqlock);
      
       errno = 0; /* strictly unnecessary */
       type = req->type; /* remember type for QUIT check */
@@ -889,7 +953,7 @@
             case REQ_WRITE:     req->result = pwrite    (req->fd, req->dataptr, req->length, req->offset); break;
 
             case REQ_READAHEAD: req->result = readahead (req->fd, req->offset, req->length); break;
-            case REQ_SENDFILE:  req->result = sendfile_ (req->fd, req->fd2, req->offset, req->length); break;
+            case REQ_SENDFILE:  req->result = sendfile_ (req->fd, req->fd2, req->offset, req->length, self); break;
 
             case REQ_STAT:      req->result = stat      (req->dataptr, req->statdata); break;
             case REQ_LSTAT:     req->result = lstat     (req->dataptr, req->statdata); break;
@@ -905,9 +969,9 @@
 
             case REQ_FDATASYNC: req->result = fdatasync (req->fd); break;
             case REQ_FSYNC:     req->result = fsync     (req->fd); break;
-            case REQ_READDIR:   req->result = scandir_  (req->dataptr, &req->data2ptr); break;
+            case REQ_READDIR:   scandir_ (req, self); break;
 
-            case REQ_SLEEP:
+            case REQ_BUSY:
               {
                 struct timeval tv;
 
@@ -929,16 +993,23 @@
 
       req->errorno = errno;
 
-      pthread_mutex_lock (&reslock);
+      LOCK (reslock);
 
       if (!reqq_push (&res_queue, req))
         /* write a dummy byte to the pipe so fh becomes ready */
         write (respipe [1], &respipe, 1);
 
-      pthread_mutex_unlock (&reslock);
+      self->req = 0;
+      worker_clear (self);
+
+      UNLOCK (reslock);
     }
   while (type != REQ_QUIT);
 
+  LOCK (wrklock);
+  worker_free (self);
+  UNLOCK (wrklock);
+
   return 0;
 }
 
@@ -946,40 +1017,54 @@
 
 static void atfork_prepare (void)
 {
-  pthread_mutex_lock (&reqlock);
-  pthread_mutex_lock (&reslock);
+  LOCK (wrklock);
+  LOCK (reqlock);
+  LOCK (reslock);
 #if !HAVE_PREADWRITE
-  pthread_mutex_lock (&preadwritelock);
+  LOCK (preadwritelock);
 #endif
 #if !HAVE_READDIR_R
-  pthread_mutex_lock (&readdirlock);
+  LOCK (readdirlock);
 #endif
 }
 
 static void atfork_parent (void)
 {
 #if !HAVE_READDIR_R
-  pthread_mutex_unlock (&readdirlock);
+  UNLOCK (readdirlock);
 #endif
 #if !HAVE_PREADWRITE
-  pthread_mutex_unlock (&preadwritelock);
+  UNLOCK (preadwritelock);
 #endif
-  pthread_mutex_unlock (&reslock);
-  pthread_mutex_unlock (&reqlock);
+  UNLOCK (reslock);
+  UNLOCK (reqlock);
+  UNLOCK (wrklock);
 }
 
 static void atfork_child (void)
 {
   aio_req prv;
 
-  started = 0;
-
   while (prv = reqq_shift (&req_queue))
     req_free (prv);
 
   while (prv = reqq_shift (&res_queue))
     req_free (prv);
- 
+
+  while (wrk_first.next != &wrk_first)
+    {
+      worker *wrk = wrk_first.next;
+
+      if (wrk->req)
+        req_free (wrk->req);
+
+      worker_clear (wrk);
+      worker_free (wrk);
+    }
+
+  started = 0;
+  nreqs = 0;
+
   close (respipe [0]);
   close (respipe [1]);
   create_pipe ();
@@ -1264,14 +1349,14 @@
 }
 
 void
-aio_sleep (delay,callback=&PL_sv_undef)
+aio_busy (delay,callback=&PL_sv_undef)
 	double delay
 	SV * callback
 	PPCODE:
 {
 	dREQ;
 
-        req->type = REQ_SLEEP;
+        req->type = REQ_BUSY;
         req->fd  = delay < 0. ? 0 : delay;
         req->fd2 = delay < 0. ? 0 : 1000. * (delay - req->fd);
 
@@ -1306,10 +1391,18 @@
 
 void
 aioreq_pri (int pri = DEFAULT_PRI)
-	CODE:
-        if (pri < PRI_MIN) pri = PRI_MIN;
-        if (pri > PRI_MAX) pri = PRI_MAX;
-        next_pri = pri + PRI_BIAS;
+      CODE:
+      if (pri < PRI_MIN) pri = PRI_MIN;
+      if (pri > PRI_MAX) pri = PRI_MAX;
+      next_pri = pri + PRI_BIAS;
+
+void
+aioreq_nice (int nice = 0)
+      CODE:
+      nice = next_pri - nice;
+      if (nice < PRI_MIN) nice = PRI_MIN;
+      if (nice > PRI_MAX) nice = PRI_MAX;
+      next_pri = nice + PRI_BIAS;
 
 void
 flush ()
@@ -1413,6 +1506,11 @@
 }
 
 void
+cancel_subs (aio_req_ornot req)
+	CODE:
+        req_cancel_subs (req);
+
+void
 result (aio_req grp, ...)
         CODE:
 {