IO-AIO/AIO.xs

#define _REENTRANT 1
#include <errno.h>

#include "EXTERN.h"
#include "perl.h"
#include "XSUB.h"

#include "autoconf/config.h"

#include <pthread.h>

#include <sys/types.h>
#include <sys/stat.h>

#include <unistd.h>
#include <fcntl.h>
#include <signal.h>
#include <sched.h>

#if HAVE_SENDFILE
# if __linux
#  include <sys/sendfile.h>
# elif __freebsd
#  include <sys/socket.h>
#  include <sys/uio.h>
# elif __hpux
#  include <sys/socket.h>
# endif
#endif

#if __ia64
# define STACKSIZE 65536
#else
# define STACKSIZE  4096
#endif

enum {
  REQ_QUIT,
  REQ_OPEN, REQ_CLOSE,
  REQ_READ, REQ_WRITE, REQ_READAHEAD,
  REQ_SENDFILE,
  REQ_STAT, REQ_LSTAT, REQ_FSTAT,
  REQ_FSYNC, REQ_FDATASYNC,
  REQ_UNLINK, REQ_RMDIR,
  REQ_SYMLINK,
};

typedef struct aio_cb {
  struct aio_cb *volatile next;

  int type;

  int fd, fd2;
  off_t offset;
  size_t length;
  ssize_t result;
  mode_t mode; /* open */
  int errorno;
  SV *data, *callback;
  SV *fh, *fh2;
  void *dataptr, *data2ptr;
  STRLEN dataoffset;

  Stat_t *statdata;
} aio_cb;

typedef aio_cb *aio_req;

static int started, wanted;
static volatile int nreqs;
static int max_outstanding = 1<<30;
static int respipe [2];

static pthread_mutex_t reslock = PTHREAD_MUTEX_INITIALIZER;
static pthread_mutex_t reqlock = PTHREAD_MUTEX_INITIALIZER;
static pthread_cond_t  reqwait = PTHREAD_COND_INITIALIZER;

static volatile aio_req reqs, reqe; /* queue start, queue end */
static volatile aio_req ress, rese; /* queue start, queue end */

static void free_req (aio_req req)
{
  if (req->data)
    SvREFCNT_dec (req->data);

  if (req->fh)
    SvREFCNT_dec (req->fh);

  if (req->fh2)
    SvREFCNT_dec (req->fh2);

  if (req->statdata)
    Safefree (req->statdata);

  if (req->callback)
    SvREFCNT_dec (req->callback);

  Safefree (req);
}

static void
poll_wait ()
{
  if (nreqs && !ress)
    {
      fd_set rfd;
      FD_ZERO(&rfd);
      FD_SET(respipe [0], &rfd);

      select (respipe [0] + 1, &rfd, 0, 0, 0);
    }
}

static int
poll_cb ()
{
  dSP;
  int count = 0;
  int do_croak = 0;
  aio_req req;

  for (;;)
    {
      pthread_mutex_lock (&reslock);
      req = ress;

      if (req)
        {
          ress = req->next;

          if (!ress)
            {
              /* read any signals sent by the worker threads */
              char buf [32];
              while (read (respipe [0], buf, 32) == 32)
                ;

              rese = 0;
            }
        }

      pthread_mutex_unlock (&reslock);

      if (!req)
        break;

      nreqs--;

      if (req->type == REQ_QUIT)
        started--;
      else
        {
          int errorno = errno;
          errno = req->errorno;

          if (req->type == REQ_READ)
            SvCUR_set (req->data, req->dataoffset + (req->result > 0 ? req->result : 0));

          if (req->data2ptr && (req->type == REQ_READ || req->type == REQ_WRITE))
            SvREADONLY_off (req->data);

          if (req->statdata)
            {
              PL_laststype   = req->type == REQ_LSTAT ? OP_LSTAT : OP_STAT;
              PL_laststatval = req->result;
              PL_statcache   = *(req->statdata);
            }

          ENTER;
          PUSHMARK (SP);
          XPUSHs (sv_2mortal (newSViv (req->result)));

          if (req->type == REQ_OPEN)
            {
              /* convert fd to fh */
              SV *fh;

              PUTBACK;
              call_pv ("IO::AIO::_fd2fh", G_SCALAR | G_EVAL);
              SPAGAIN;

              fh = SvREFCNT_inc (POPs);

              PUSHMARK (SP);
              XPUSHs (sv_2mortal (fh));
            }

          if (SvOK (req->callback))
            {
              PUTBACK;
              call_sv (req->callback, G_VOID | G_EVAL);
              SPAGAIN;

              if (SvTRUE (ERRSV))
                {
                  free_req (req);
                  croak (0);
                }
            }

          LEAVE;

          errno = errorno;
          count++;
        }

      free_req (req);
    }

  return count;
}

static void *aio_proc(void *arg);

static void
start_thread (void)
{
  sigset_t fullsigset, oldsigset;
  pthread_t tid;
  pthread_attr_t attr;

  pthread_attr_init (&attr);
  pthread_attr_setstacksize (&attr, STACKSIZE);
  pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED);

  sigfillset (&fullsigset);
  sigprocmask (SIG_SETMASK, &fullsigset, &oldsigset);

  if (pthread_create (&tid, &attr, aio_proc, 0) == 0)
    started++;

  sigprocmask (SIG_SETMASK, &oldsigset, 0);
}

static void
send_req (aio_req req)
{
  while (started < wanted && nreqs >= started)
    start_thread ();

  nreqs++;

  pthread_mutex_lock (&reqlock);

  req->next = 0;

  if (reqe)
    {
      reqe->next = req;
      reqe = req;
    }
  else
    reqe = reqs = req;

  pthread_cond_signal (&reqwait);
  pthread_mutex_unlock (&reqlock);

  if (nreqs > max_outstanding)
    for (;;)
      {
        poll_cb ();

        if (nreqs <= max_outstanding)
          break;

        poll_wait ();
      }
}

static void
end_thread (void)
{
  aio_req req;
  Newz (0, req, 1, aio_cb);
  req->type = REQ_QUIT;

  send_req (req);
}

static void min_parallel (int nthreads)
{
  if (wanted < nthreads)
    wanted = nthreads;
}

static void max_parallel (int nthreads)
{
  int cur = started;

  if (wanted > nthreads)
    wanted = nthreads;

  while (cur > wanted)
    {
      end_thread ();
      cur--;
    }

  while (started > wanted)
    {
      poll_wait ();
      poll_cb ();
    }
}

static void create_pipe ()
{
  if (pipe (respipe))
    croak ("unable to initialize result pipe");

  if (fcntl (respipe [0], F_SETFL, O_NONBLOCK))
    croak ("cannot set result pipe to nonblocking mode");

  if (fcntl (respipe [1], F_SETFL, O_NONBLOCK))
    croak ("cannot set result pipe to nonblocking mode");
}

static void atfork_prepare (void)
{
  pthread_mutex_lock (&reqlock);
  pthread_mutex_lock (&reslock);
}

static void atfork_parent (void)
{
  pthread_mutex_unlock (&reslock);
  pthread_mutex_unlock (&reqlock);
}

static void atfork_child (void)
{
  aio_req prv;

  started = 0;

  while (reqs)
    {
      prv = reqs;
      reqs = prv->next;
      free_req (prv);
    }

  reqs = reqe = 0;
      
  while (ress)
    {
      prv = ress;
      ress = prv->next;
      free_req (prv);
    }
      
  ress = rese = 0;

  close (respipe [0]);
  close (respipe [1]);
  create_pipe ();

  atfork_parent ();
}

/*****************************************************************************/
/* work around various missing functions */

#if !HAVE_PREADWRITE
# define pread  aio_pread
# define pwrite aio_pwrite

/*
 * make our pread/pwrite safe against themselves, but not against
 * normal read/write by using a mutex. slows down execution a lot,
 * but that's your problem, not mine.
 */
static pthread_mutex_t iolock = PTHREAD_MUTEX_INITIALIZER;

static ssize_t
pread (int fd, void *buf, size_t count, off_t offset)
{
  ssize_t res;
  off_t ooffset;

  pthread_mutex_lock (&iolock);
  ooffset = lseek (fd, 0, SEEK_CUR);
  lseek (fd, offset, SEEK_SET);
  res = read (fd, buf, count);
  lseek (fd, ooffset, SEEK_SET);
  pthread_mutex_unlock (&iolock);

  return res;
}

static ssize_t
pwrite (int fd, void *buf, size_t count, off_t offset)
{
  ssize_t res;
  off_t ooffset;

  pthread_mutex_lock (&iolock);
  ooffset = lseek (fd, 0, SEEK_CUR);
  lseek (fd, offset, SEEK_SET);
  res = write (fd, buf, count);
  lseek (fd, offset, SEEK_SET);
  pthread_mutex_unlock (&iolock);

  return res;
}
#endif

#if !HAVE_FDATASYNC
# define fdatasync fsync
#endif

#if !HAVE_READAHEAD
# define readahead aio_readahead

static char readahead_buf[4096];

static ssize_t
readahead (int fd, off_t offset, size_t count)
{
  while (count > 0)
    {
      size_t len = count < sizeof (readahead_buf) ? count : sizeof (readahead_buf);

      pread (fd, readahead_buf, len, offset);
      offset += len;
      count  -= len;
    }

  errno = 0;
}
#endif

/* sendfile always needs emulation */
static ssize_t
sendfile_ (int ofd, int ifd, off_t offset, size_t count)
{
   ssize_t res;

   if (!count)
     return 0;

#if __linux
   res = sendfile (ofd, ifd, &offset, count);

#elif __freebsd
   /*
    * Of course, the freebsd sendfile is a dire hack with no thoughts
    * wasted on making it similar to other i/o functions.
    */
   {
     off_t sbytes;
     res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);

     if (!res && errno == EAGAIN)
       res = sbytes;
   }

#elif __hpux
   res = sendfile (ofd, ifd, offset, count, 0, 0);

#else
   res = -1;
   errno = ENOSYS;
#endif

   if (res < 0 && (errno == ENOSYS || errno == EINVAL || errno == ENOTSOCK))
     {
       /* emulate sendfile. this is a major pain in the ass */
       char *buf = malloc (4096);
       res = 0;

       for (;;)
         {
           ssize_t cnt;
           
           cnt = pread (ifd, buf, 4096, offset);

           if (cnt <= 0)
             {
               if (cnt && !res) res = -1;
               break;
             }

           cnt = write (ofd, buf, cnt);

           if (cnt <= 0)
             {
               if (cnt && !res) res = -1;
               break;
             }

           offset += cnt;
           res    += cnt;
         }

       {
         int errorno = errno;
         free (buf);
         errno = errorno;
       }
     }

   return res;
}

/*****************************************************************************/

static void *
aio_proc (void *thr_arg)
{
  aio_req req;
  int type;

  do
    {
      pthread_mutex_lock (&reqlock);

      for (;;)
        {
          req = reqs;

          if (reqs)
            {
              reqs = reqs->next;
              if (!reqs) reqe = 0;
            }

          if (req)
            break;

          pthread_cond_wait (&reqwait, &reqlock);
        }

      pthread_mutex_unlock (&reqlock);
     
      errno = 0; /* strictly unnecessary */

      type = req->type;

      switch (type)
        {
          case REQ_READ:      req->result = pread     (req->fd, req->dataptr, req->length, req->offset); break;
          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_STAT:      req->result = stat      (req->dataptr, req->statdata); break;
          case REQ_LSTAT:     req->result = lstat     (req->dataptr, req->statdata); break;
          case REQ_FSTAT:     req->result = fstat     (req->fd     , req->statdata); break;

          case REQ_OPEN:      req->result = open      (req->dataptr, req->fd, req->mode); break;
          case REQ_CLOSE:     req->result = close     (req->fd); break;
          case REQ_UNLINK:    req->result = unlink    (req->dataptr); break;
          case REQ_RMDIR:     req->result = rmdir     (req->dataptr); break;
          case REQ_SYMLINK:   req->result = symlink   (req->data2ptr, req->dataptr); break;

          case REQ_FDATASYNC: req->result = fdatasync (req->fd); break;
          case REQ_FSYNC:     req->result = fsync     (req->fd); break;

          case REQ_QUIT:
            break;

          default:
            req->result = ENOSYS;
            break;
        }

      req->errorno = errno;

      pthread_mutex_lock (&reslock);

      req->next = 0;

      if (rese)
        {
          rese->next = req;
          rese = req;
        }
      else
        {
          rese = ress = req;

          /* write a dummy byte to the pipe so fh becomes ready */
          write (respipe [1], &respipe, 1);
        }

      pthread_mutex_unlock (&reslock);
    }
  while (type != REQ_QUIT);

  return 0;
}

#define dREQ                                                    \
  aio_req req;                                                  \
                                                                \
  if (SvOK (callback) && !SvROK (callback))                     \
    croak ("clalback must be undef or of reference type");      \
                                                                \
  Newz (0, req, 1, aio_cb);                                     \
  if (!req)                                                     \
    croak ("out of memory during aio_req allocation");          \
                                                                \
  req->callback = newSVsv (callback);
        
MODULE = IO::AIO                PACKAGE = IO::AIO

PROTOTYPES: ENABLE

BOOT:
{
        create_pipe ();
        pthread_atfork (atfork_prepare, atfork_parent, atfork_child);
}

void
min_parallel(nthreads)
        int     nthreads
        PROTOTYPE: $

void
max_parallel(nthreads)
        int     nthreads
        PROTOTYPE: $

int
max_outstanding(nreqs)
        int nreqs
        PROTOTYPE: $
        CODE:
        RETVAL = max_outstanding;
        max_outstanding = nreqs;

void
aio_open(pathname,flags,mode,callback=&PL_sv_undef)
        SV *    pathname
        int     flags
        int     mode
        SV *    callback
        PROTOTYPE: $$$;$
        CODE:
{
        dREQ;

        req->type = REQ_OPEN;
        req->data = newSVsv (pathname);
        req->dataptr = SvPVbyte_nolen (req->data);
        req->fd = flags;
        req->mode = mode;

        send_req (req);
}

void
aio_close(fh,callback=&PL_sv_undef)
        SV *    fh
        SV *    callback
        PROTOTYPE: $;$
        ALIAS:
           aio_close     = REQ_CLOSE
           aio_fsync     = REQ_FSYNC
           aio_fdatasync = REQ_FDATASYNC
        CODE:
{
        dREQ;

        req->type = ix;
        req->fh = newSVsv (fh);
        req->fd = PerlIO_fileno (IoIFP (sv_2io (fh)));

        send_req (req);
}

void
aio_read(fh,offset,length,data,dataoffset,callback=&PL_sv_undef)
        SV *    fh
        UV      offset
        UV      length
        SV *    data
        UV      dataoffset
        SV *    callback
        ALIAS:
           aio_read  = REQ_READ
           aio_write = REQ_WRITE
        PROTOTYPE: $$$$$;$
        CODE:
{
        aio_req req;
        STRLEN svlen;
        char *svptr = SvPVbyte (data, svlen);

        SvUPGRADE (data, SVt_PV);
        SvPOK_on (data);

        if (dataoffset < 0)
          dataoffset += svlen;

        if (dataoffset < 0 || dataoffset > svlen)
          croak ("data offset outside of string");

        if (ix == REQ_WRITE)
          {
            /* write: check length and adjust. */
            if (length < 0 || length + dataoffset > svlen)
              length = svlen - dataoffset;
          }
        else
          {
            /* read: grow scalar as necessary */
            svptr = SvGROW (data, length + dataoffset);
          }

        if (length < 0)
          croak ("length must not be negative");

        {
          dREQ;

          req->type = ix;
          req->fh = newSVsv (fh);
          req->fd = PerlIO_fileno (ix == REQ_READ ? IoIFP (sv_2io (fh))
                                                  : IoOFP (sv_2io (fh)));
          req->offset = offset;
          req->length = length;
          req->data = SvREFCNT_inc (data);
          req->dataptr = (char *)svptr + dataoffset;

          if (!SvREADONLY (data))
            {
              SvREADONLY_on (data);
              req->data2ptr = (void *)data;
            }

          send_req (req);
        }
}

void
aio_sendfile(out_fh,in_fh,in_offset,length,callback=&PL_sv_undef)
        SV *    out_fh
        SV *    in_fh
        UV      in_offset
        UV      length
        SV *    callback
        PROTOTYPE: $$$$;$
        CODE:
{
        dREQ;

        req->type = REQ_SENDFILE;
        req->fh = newSVsv (out_fh);
        req->fd = PerlIO_fileno (IoIFP (sv_2io (out_fh)));
        req->fh2 = newSVsv (in_fh);
        req->fd2 = PerlIO_fileno (IoIFP (sv_2io (in_fh)));
        req->offset = in_offset;
        req->length = length;

        send_req (req);
}

void
aio_readahead(fh,offset,length,callback=&PL_sv_undef)
        SV *    fh
        UV      offset
        IV      length
        SV *    callback
        PROTOTYPE: $$$;$
        CODE:
{
        dREQ;

        req->type = REQ_READAHEAD;
        req->fh = newSVsv (fh);
        req->fd = PerlIO_fileno (IoIFP (sv_2io (fh)));
        req->offset = offset;
        req->length = length;

        send_req (req);
}

void
aio_stat(fh_or_path,callback=&PL_sv_undef)
        SV *            fh_or_path
        SV *            callback
        ALIAS:
           aio_stat  = REQ_STAT
           aio_lstat = REQ_LSTAT
        CODE:
{
        dREQ;

        New (0, req->statdata, 1, Stat_t);
        if (!req->statdata)
          {
            free_req (req);
            croak ("out of memory during aio_req->statdata allocation");
          }

        if (SvPOK (fh_or_path))
          {
            req->type = ix;
            req->data = newSVsv (fh_or_path);
            req->dataptr = SvPVbyte_nolen (req->data);
          }
        else
          {
            req->type = REQ_FSTAT;
            req->fh = newSVsv (fh_or_path);
            req->fd = PerlIO_fileno (IoIFP (sv_2io (fh_or_path)));
          }

        send_req (req);
}

void
aio_unlink(pathname,callback=&PL_sv_undef)
        SV * pathname
        SV * callback
        ALIAS:
           aio_unlink = REQ_UNLINK
           aio_rmdir  = REQ_RMDIR
        CODE:
{
        dREQ;
        
        req->type = ix;
        req->data = newSVsv (pathname);
        req->dataptr = SvPVbyte_nolen (req->data);
        
        send_req (req);
}

void
aio_symlink(oldpath,newpath,callback=&PL_sv_undef)
        SV * oldpath
        SV * newpath
        SV * callback
        CODE:
{
        dREQ;
        
        req->type = REQ_SYMLINK;
        req->fh = newSVsv (oldpath);
        req->data2ptr = SvPVbyte_nolen (req->fh);
        req->data = newSVsv (newpath);
        req->dataptr = SvPVbyte_nolen (req->data);
        
        send_req (req);
}

void
flush()
        PROTOTYPE:
        CODE:
        while (nreqs)
          {
            poll_wait ();
            poll_cb ();
          }

void
poll()
        PROTOTYPE:
        CODE:
        if (nreqs)
          {
            poll_wait ();
            poll_cb ();
          }

int
poll_fileno()
        PROTOTYPE:
        CODE:
        RETVAL = respipe [0];
        OUTPUT:
        RETVAL

int
poll_cb(...)
        PROTOTYPE:
        CODE:
        RETVAL = poll_cb ();
        OUTPUT:
        RETVAL

void
poll_wait()
        PROTOTYPE:
        CODE:
        if (nreqs)
          poll_wait ();

int
nreqs()
        PROTOTYPE:
        CODE:
        RETVAL = nreqs;
        OUTPUT:
        RETVAL

Revision:	1.32
Committed:	Mon Aug 22 23:20:37 2005 UTC (18 years, 8 months ago) by root
Branch:	MAIN
Changes since 1.31:	+118 -9 lines
Log Message:	* empty log message *
#	User	Rev	Content
1	root	1.20	#define _REENTRANT 1
2	root	1.19	#include <errno.h>
3
4	root	1.15	#include "EXTERN.h"
5	root	1.1	#include "perl.h"
6			#include "XSUB.h"
7
8	root	1.16	#include "autoconf/config.h"
9
10	root	1.32	#include <pthread.h>
11
12	root	1.1	#include <sys/types.h>
13			#include <sys/stat.h>
14	root	1.10
15	root	1.1	#include <unistd.h>
16			#include <fcntl.h>
17			#include <signal.h>
18			#include <sched.h>
19
20	root	1.32	#if HAVE_SENDFILE
21			# if __linux
22			# include <sys/sendfile.h>
23			# elif __freebsd
24			# include <sys/socket.h>
25			# include <sys/uio.h>
26			# elif __hpux
27			# include <sys/socket.h>
28			# endif
29			#endif
30	root	1.1
31	root	1.4	#if __ia64
32			# define STACKSIZE 65536
33	root	1.1	#else
34	root	1.4	# define STACKSIZE 4096
35	root	1.1	#endif
36
37			enum {
38			REQ_QUIT,
39			REQ_OPEN, REQ_CLOSE,
40			REQ_READ, REQ_WRITE, REQ_READAHEAD,
41	root	1.32	REQ_SENDFILE,
42	root	1.22	REQ_STAT, REQ_LSTAT, REQ_FSTAT,
43	root	1.1	REQ_FSYNC, REQ_FDATASYNC,
44	root	1.23	REQ_UNLINK, REQ_RMDIR,
45	root	1.22	REQ_SYMLINK,
46	root	1.1	};
47
48			typedef struct aio_cb {
49	root	1.3	struct aio_cb *volatile next;
50	root	1.1
51			int type;
52
53	root	1.32	int fd, fd2;
54	root	1.1	off_t offset;
55			size_t length;
56			ssize_t result;
57			mode_t mode; /* open */
58			int errorno;
59	root	1.32	SV data, callback;
60			SV fh, fh2;
61	root	1.22	void dataptr, data2ptr;
62	root	1.1	STRLEN dataoffset;
63
64			Stat_t *statdata;
65			} aio_cb;
66
67			typedef aio_cb *aio_req;
68
69	root	1.30	static int started, wanted;
70	root	1.11	static volatile int nreqs;
71	root	1.4	static int max_outstanding = 1<<30;
72	root	1.3	static int respipe [2];
73	root	1.1
74	root	1.3	static pthread_mutex_t reslock = PTHREAD_MUTEX_INITIALIZER;
75			static pthread_mutex_t reqlock = PTHREAD_MUTEX_INITIALIZER;
76			static pthread_cond_t reqwait = PTHREAD_COND_INITIALIZER;
77
78			static volatile aio_req reqs, reqe; /* queue start, queue end */
79			static volatile aio_req ress, rese; /* queue start, queue end */
80	root	1.1
81	root	1.26	static void free_req (aio_req req)
82			{
83			if (req->data)
84			SvREFCNT_dec (req->data);
85
86			if (req->fh)
87			SvREFCNT_dec (req->fh);
88
89	root	1.32	if (req->fh2)
90			SvREFCNT_dec (req->fh2);
91
92	root	1.27	if (req->statdata)
93	root	1.26	Safefree (req->statdata);
94
95			if (req->callback)
96			SvREFCNT_dec (req->callback);
97
98			Safefree (req);
99			}
100
101	root	1.1	static void
102			poll_wait ()
103			{
104	root	1.11	if (nreqs && !ress)
105			{
106			fd_set rfd;
107			FD_ZERO(&rfd);
108			FD_SET(respipe [0], &rfd);
109	root	1.3
110	root	1.11	select (respipe [0] + 1, &rfd, 0, 0, 0);
111			}
112	root	1.1	}
113
114			static int
115	root	1.5	poll_cb ()
116	root	1.1	{
117			dSP;
118			int count = 0;
119	root	1.24	int do_croak = 0;
120	root	1.27	aio_req req;
121
122			for (;;)
123			{
124			pthread_mutex_lock (&reslock);
125			req = ress;
126
127			if (req)
128			{
129			ress = req->next;
130	root	1.11
131	root	1.27	if (!ress)
132			{
133			/* read any signals sent by the worker threads */
134			char buf [32];
135			while (read (respipe [0], buf, 32) == 32)
136			;
137	root	1.30
138			rese = 0;
139	root	1.27	}
140			}
141	root	1.1
142	root	1.27	pthread_mutex_unlock (&reslock);
143	root	1.3
144	root	1.27	if (!req)
145			break;
146	root	1.3
147	root	1.1	nreqs--;
148
149			if (req->type == REQ_QUIT)
150			started--;
151			else
152			{
153			int errorno = errno;
154			errno = req->errorno;
155
156			if (req->type == REQ_READ)
157	root	1.27	SvCUR_set (req->data, req->dataoffset + (req->result > 0 ? req->result : 0));
158	root	1.1
159	root	1.28	if (req->data2ptr && (req->type == REQ_READ \|\| req->type == REQ_WRITE))
160			SvREADONLY_off (req->data);
161
162	root	1.27	if (req->statdata)
163	root	1.1	{
164			PL_laststype = req->type == REQ_LSTAT ? OP_LSTAT : OP_STAT;
165			PL_laststatval = req->result;
166			PL_statcache = *(req->statdata);
167			}
168
169	root	1.13	ENTER;
170	root	1.1	PUSHMARK (SP);
171			XPUSHs (sv_2mortal (newSViv (req->result)));
172	root	1.2
173			if (req->type == REQ_OPEN)
174			{
175			/* convert fd to fh */
176			SV *fh;
177
178			PUTBACK;
179			call_pv ("IO::AIO::_fd2fh", G_SCALAR \| G_EVAL);
180			SPAGAIN;
181
182	root	1.13	fh = SvREFCNT_inc (POPs);
183	root	1.2
184			PUSHMARK (SP);
185	root	1.13	XPUSHs (sv_2mortal (fh));
186	root	1.2	}
187
188	root	1.8	if (SvOK (req->callback))
189			{
190			PUTBACK;
191			call_sv (req->callback, G_VOID \| G_EVAL);
192			SPAGAIN;
193	root	1.27
194			if (SvTRUE (ERRSV))
195			{
196			free_req (req);
197			croak (0);
198			}
199	root	1.8	}
200	root	1.13
201	root	1.26	LEAVE;
202
203	root	1.1	errno = errorno;
204			count++;
205			}
206
207	root	1.27	free_req (req);
208	root	1.1	}
209
210			return count;
211			}
212
213	root	1.4	static void aio_proc(void arg);
214
215			static void
216			start_thread (void)
217			{
218			sigset_t fullsigset, oldsigset;
219			pthread_t tid;
220			pthread_attr_t attr;
221
222			pthread_attr_init (&attr);
223			pthread_attr_setstacksize (&attr, STACKSIZE);
224	root	1.31	pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED);
225	root	1.4
226			sigfillset (&fullsigset);
227			sigprocmask (SIG_SETMASK, &fullsigset, &oldsigset);
228
229			if (pthread_create (&tid, &attr, aio_proc, 0) == 0)
230			started++;
231
232			sigprocmask (SIG_SETMASK, &oldsigset, 0);
233			}
234
235			static void
236			send_req (aio_req req)
237			{
238	root	1.30	while (started < wanted && nreqs >= started)
239			start_thread ();
240
241	root	1.4	nreqs++;
242
243			pthread_mutex_lock (&reqlock);
244
245			req->next = 0;
246
247			if (reqe)
248			{
249			reqe->next = req;
250			reqe = req;
251			}
252			else
253			reqe = reqs = req;
254
255			pthread_cond_signal (&reqwait);
256			pthread_mutex_unlock (&reqlock);
257
258	root	1.27	if (nreqs > max_outstanding)
259			for (;;)
260			{
261			poll_cb ();
262
263			if (nreqs <= max_outstanding)
264			break;
265
266			poll_wait ();
267			}
268	root	1.4	}
269
270			static void
271			end_thread (void)
272			{
273			aio_req req;
274	root	1.26	Newz (0, req, 1, aio_cb);
275	root	1.4	req->type = REQ_QUIT;
276
277			send_req (req);
278			}
279
280	root	1.22	static void min_parallel (int nthreads)
281			{
282	root	1.30	if (wanted < nthreads)
283			wanted = nthreads;
284	root	1.22	}
285
286			static void max_parallel (int nthreads)
287			{
288			int cur = started;
289	root	1.27
290	root	1.30	if (wanted > nthreads)
291			wanted = nthreads;
292
293			while (cur > wanted)
294			{
295	root	1.22	end_thread ();
296			cur--;
297			}
298
299	root	1.30	while (started > wanted)
300	root	1.22	{
301			poll_wait ();
302			poll_cb ();
303			}
304			}
305
306	root	1.26	static void create_pipe ()
307	root	1.22	{
308	root	1.26	if (pipe (respipe))
309			croak ("unable to initialize result pipe");
310	root	1.22
311	root	1.26	if (fcntl (respipe [0], F_SETFL, O_NONBLOCK))
312			croak ("cannot set result pipe to nonblocking mode");
313	root	1.22
314	root	1.26	if (fcntl (respipe [1], F_SETFL, O_NONBLOCK))
315			croak ("cannot set result pipe to nonblocking mode");
316			}
317	root	1.22
318	root	1.26	static void atfork_prepare (void)
319			{
320	root	1.28	pthread_mutex_lock (&reqlock);
321	root	1.22	pthread_mutex_lock (&reslock);
322			}
323
324			static void atfork_parent (void)
325			{
326			pthread_mutex_unlock (&reslock);
327	root	1.26	pthread_mutex_unlock (&reqlock);
328			}
329
330			static void atfork_child (void)
331			{
332	root	1.28	aio_req prv;
333
334	root	1.26	started = 0;
335
336	root	1.28	while (reqs)
337			{
338			prv = reqs;
339			reqs = prv->next;
340			free_req (prv);
341			}
342
343			reqs = reqe = 0;
344
345			while (ress)
346			{
347			prv = ress;
348			ress = prv->next;
349			free_req (prv);
350			}
351
352			ress = rese = 0;
353
354	root	1.29	close (respipe [0]);
355			close (respipe [1]);
356			create_pipe ();
357
358	root	1.26	atfork_parent ();
359	root	1.22	}
360
361			/*****************************************************************************/
362	root	1.17	/* work around various missing functions */
363
364			#if !HAVE_PREADWRITE
365			# define pread aio_pread
366			# define pwrite aio_pwrite
367
368			/*
369			* make our pread/pwrite safe against themselves, but not against
370			* normal read/write by using a mutex. slows down execution a lot,
371			* but that's your problem, not mine.
372			*/
373			static pthread_mutex_t iolock = PTHREAD_MUTEX_INITIALIZER;
374
375			static ssize_t
376			pread (int fd, void *buf, size_t count, off_t offset)
377			{
378			ssize_t res;
379			off_t ooffset;
380
381			pthread_mutex_lock (&iolock);
382			ooffset = lseek (fd, 0, SEEK_CUR);
383			lseek (fd, offset, SEEK_SET);
384			res = read (fd, buf, count);
385			lseek (fd, ooffset, SEEK_SET);
386			pthread_mutex_unlock (&iolock);
387
388			return res;
389			}
390
391			static ssize_t
392			pwrite (int fd, void *buf, size_t count, off_t offset)
393			{
394			ssize_t res;
395			off_t ooffset;
396
397			pthread_mutex_lock (&iolock);
398			ooffset = lseek (fd, 0, SEEK_CUR);
399			lseek (fd, offset, SEEK_SET);
400			res = write (fd, buf, count);
401			lseek (fd, offset, SEEK_SET);
402			pthread_mutex_unlock (&iolock);
403
404			return res;
405			}
406			#endif
407
408			#if !HAVE_FDATASYNC
409			# define fdatasync fsync
410			#endif
411
412			#if !HAVE_READAHEAD
413			# define readahead aio_readahead
414
415			static char readahead_buf[4096];
416
417			static ssize_t
418			readahead (int fd, off_t offset, size_t count)
419			{
420			while (count > 0)
421			{
422			size_t len = count < sizeof (readahead_buf) ? count : sizeof (readahead_buf);
423
424			pread (fd, readahead_buf, len, offset);
425			offset += len;
426			count -= len;
427			}
428
429			errno = 0;
430			}
431			#endif
432
433	root	1.32	/* sendfile always needs emulation */
434			static ssize_t
435			sendfile_ (int ofd, int ifd, off_t offset, size_t count)
436			{
437			ssize_t res;
438
439			if (!count)
440			return 0;
441
442			#if __linux
443			res = sendfile (ofd, ifd, &offset, count);
444
445			#elif __freebsd
446			/*
447			* Of course, the freebsd sendfile is a dire hack with no thoughts
448			* wasted on making it similar to other i/o functions.
449			*/
450			{
451			off_t sbytes;
452			res = sendfile (ifd, ofd, offset, count, 0, &sbytes, 0);
453
454			if (!res && errno == EAGAIN)
455			res = sbytes;
456			}
457
458			#elif __hpux
459			res = sendfile (ofd, ifd, offset, count, 0, 0);
460
461			#else
462			res = -1;
463			errno = ENOSYS;
464			#endif
465
466			if (res < 0 && (errno == ENOSYS \|\| errno == EINVAL \|\| errno == ENOTSOCK))
467			{
468			/* emulate sendfile. this is a major pain in the ass */
469			char *buf = malloc (4096);
470			res = 0;
471
472			for (;;)
473			{
474			ssize_t cnt;
475
476			cnt = pread (ifd, buf, 4096, offset);
477
478			if (cnt <= 0)
479			{
480			if (cnt && !res) res = -1;
481			break;
482			}
483
484			cnt = write (ofd, buf, cnt);
485
486			if (cnt <= 0)
487			{
488			if (cnt && !res) res = -1;
489			break;
490			}
491
492			offset += cnt;
493			res += cnt;
494			}
495
496			{
497			int errorno = errno;
498			free (buf);
499			errno = errorno;
500			}
501			}
502
503			return res;
504			}
505
506	root	1.22	/*****************************************************************************/
507
508	root	1.1	static void *
509			aio_proc (void *thr_arg)
510			{
511			aio_req req;
512	root	1.3	int type;
513	root	1.1
514	root	1.3	do
515	root	1.1	{
516	root	1.3	pthread_mutex_lock (&reqlock);
517
518			for (;;)
519			{
520			req = reqs;
521
522			if (reqs)
523			{
524			reqs = reqs->next;
525			if (!reqs) reqe = 0;
526			}
527
528			if (req)
529			break;
530
531			pthread_cond_wait (&reqwait, &reqlock);
532			}
533
534			pthread_mutex_unlock (&reqlock);
535
536	root	1.1	errno = 0; /* strictly unnecessary */
537
538	root	1.3	type = req->type;
539
540			switch (type)
541	root	1.1	{
542	root	1.10	case REQ_READ: req->result = pread (req->fd, req->dataptr, req->length, req->offset); break;
543			case REQ_WRITE: req->result = pwrite (req->fd, req->dataptr, req->length, req->offset); break;
544	root	1.16
545	root	1.1	case REQ_READAHEAD: req->result = readahead (req->fd, req->offset, req->length); break;
546	root	1.32	case REQ_SENDFILE: req->result = sendfile_ (req->fd, req->fd2, req->offset, req->length); break;
547	root	1.1
548			case REQ_STAT: req->result = stat (req->dataptr, req->statdata); break;
549			case REQ_LSTAT: req->result = lstat (req->dataptr, req->statdata); break;
550			case REQ_FSTAT: req->result = fstat (req->fd , req->statdata); break;
551
552			case REQ_OPEN: req->result = open (req->dataptr, req->fd, req->mode); break;
553			case REQ_CLOSE: req->result = close (req->fd); break;
554			case REQ_UNLINK: req->result = unlink (req->dataptr); break;
555	root	1.22	case REQ_RMDIR: req->result = rmdir (req->dataptr); break;
556			case REQ_SYMLINK: req->result = symlink (req->data2ptr, req->dataptr); break;
557	root	1.1
558	root	1.17	case REQ_FDATASYNC: req->result = fdatasync (req->fd); break;
559	root	1.1	case REQ_FSYNC: req->result = fsync (req->fd); break;
560
561			case REQ_QUIT:
562	root	1.3	break;
563	root	1.1
564			default:
565			req->result = ENOSYS;
566			break;
567			}
568
569			req->errorno = errno;
570	root	1.3
571			pthread_mutex_lock (&reslock);
572
573			req->next = 0;
574
575			if (rese)
576			{
577			rese->next = req;
578			rese = req;
579			}
580			else
581			{
582			rese = ress = req;
583
584			/* write a dummy byte to the pipe so fh becomes ready */
585			write (respipe [1], &respipe, 1);
586			}
587
588			pthread_mutex_unlock (&reslock);
589	root	1.1	}
590	root	1.3	while (type != REQ_QUIT);
591	root	1.1
592			return 0;
593			}
594
595	root	1.22	#define dREQ \
596			aio_req req; \
597			\
598			if (SvOK (callback) && !SvROK (callback)) \
599			croak ("clalback must be undef or of reference type"); \
600			\
601			Newz (0, req, 1, aio_cb); \
602			if (!req) \
603			croak ("out of memory during aio_req allocation"); \
604			\
605	root	1.27	req->callback = newSVsv (callback);
606	root	1.22
607	root	1.1	MODULE = IO::AIO PACKAGE = IO::AIO
608
609	root	1.8	PROTOTYPES: ENABLE
610
611	root	1.1	BOOT:
612			{
613	root	1.26	create_pipe ();
614	root	1.22	pthread_atfork (atfork_prepare, atfork_parent, atfork_child);
615	root	1.1	}
616
617			void
618			min_parallel(nthreads)
619			int nthreads
620			PROTOTYPE: $
621
622			void
623			max_parallel(nthreads)
624			int nthreads
625			PROTOTYPE: $
626
627	root	1.4	int
628			max_outstanding(nreqs)
629			int nreqs
630			PROTOTYPE: $
631			CODE:
632			RETVAL = max_outstanding;
633			max_outstanding = nreqs;
634
635	root	1.1	void
636	root	1.8	aio_open(pathname,flags,mode,callback=&PL_sv_undef)
637	root	1.1	SV * pathname
638			int flags
639			int mode
640			SV * callback
641	root	1.8	PROTOTYPE: $$$;$
642	root	1.1	CODE:
643			{
644	root	1.22	dREQ;
645	root	1.1
646			req->type = REQ_OPEN;
647			req->data = newSVsv (pathname);
648	root	1.22	req->dataptr = SvPVbyte_nolen (req->data);
649	root	1.1	req->fd = flags;
650			req->mode = mode;
651
652			send_req (req);
653			}
654
655			void
656	root	1.8	aio_close(fh,callback=&PL_sv_undef)
657	root	1.13	SV * fh
658			SV * callback
659	root	1.8	PROTOTYPE: $;$
660	root	1.1	ALIAS:
661			aio_close = REQ_CLOSE
662			aio_fsync = REQ_FSYNC
663			aio_fdatasync = REQ_FDATASYNC
664			CODE:
665			{
666	root	1.22	dREQ;
667	root	1.1
668			req->type = ix;
669	root	1.13	req->fh = newSVsv (fh);
670			req->fd = PerlIO_fileno (IoIFP (sv_2io (fh)));
671	root	1.1
672			send_req (req);
673			}
674
675			void
676	root	1.8	aio_read(fh,offset,length,data,dataoffset,callback=&PL_sv_undef)
677	root	1.13	SV * fh
678			UV offset
679	root	1.32	UV length
680	root	1.13	SV * data
681	root	1.32	UV dataoffset
682	root	1.13	SV * callback
683			ALIAS:
684			aio_read = REQ_READ
685			aio_write = REQ_WRITE
686	root	1.8	PROTOTYPE: $$$$$;$
687	root	1.1	CODE:
688	root	1.13	{
689			aio_req req;
690			STRLEN svlen;
691	root	1.21	char *svptr = SvPVbyte (data, svlen);
692	root	1.13
693			SvUPGRADE (data, SVt_PV);
694			SvPOK_on (data);
695	root	1.1
696	root	1.13	if (dataoffset < 0)
697			dataoffset += svlen;
698
699			if (dataoffset < 0 \|\| dataoffset > svlen)
700			croak ("data offset outside of string");
701
702			if (ix == REQ_WRITE)
703			{
704			/* write: check length and adjust. */
705			if (length < 0 \|\| length + dataoffset > svlen)
706			length = svlen - dataoffset;
707			}
708			else
709			{
710			/* read: grow scalar as necessary */
711			svptr = SvGROW (data, length + dataoffset);
712			}
713
714			if (length < 0)
715			croak ("length must not be negative");
716
717	root	1.22	{
718			dREQ;
719	root	1.13
720	root	1.22	req->type = ix;
721			req->fh = newSVsv (fh);
722			req->fd = PerlIO_fileno (ix == REQ_READ ? IoIFP (sv_2io (fh))
723			: IoOFP (sv_2io (fh)));
724			req->offset = offset;
725			req->length = length;
726			req->data = SvREFCNT_inc (data);
727			req->dataptr = (char *)svptr + dataoffset;
728	root	1.13
729	root	1.28	if (!SvREADONLY (data))
730			{
731			SvREADONLY_on (data);
732			req->data2ptr = (void *)data;
733			}
734
735	root	1.22	send_req (req);
736			}
737	root	1.13	}
738	root	1.1
739			void
740	root	1.32	aio_sendfile(out_fh,in_fh,in_offset,length,callback=&PL_sv_undef)
741			SV * out_fh
742			SV * in_fh
743			UV in_offset
744			UV length
745			SV * callback
746			PROTOTYPE: $$$$;$
747			CODE:
748			{
749			dREQ;
750
751			req->type = REQ_SENDFILE;
752			req->fh = newSVsv (out_fh);
753			req->fd = PerlIO_fileno (IoIFP (sv_2io (out_fh)));
754			req->fh2 = newSVsv (in_fh);
755			req->fd2 = PerlIO_fileno (IoIFP (sv_2io (in_fh)));
756			req->offset = in_offset;
757			req->length = length;
758
759			send_req (req);
760			}
761
762			void
763	root	1.8	aio_readahead(fh,offset,length,callback=&PL_sv_undef)
764	root	1.13	SV * fh
765			UV offset
766			IV length
767			SV * callback
768	root	1.8	PROTOTYPE: $$$;$
769	root	1.1	CODE:
770			{
771	root	1.22	dREQ;
772	root	1.1
773			req->type = REQ_READAHEAD;
774	root	1.13	req->fh = newSVsv (fh);
775			req->fd = PerlIO_fileno (IoIFP (sv_2io (fh)));
776	root	1.1	req->offset = offset;
777			req->length = length;
778
779			send_req (req);
780			}
781
782			void
783	root	1.8	aio_stat(fh_or_path,callback=&PL_sv_undef)
784	root	1.1	SV * fh_or_path
785			SV * callback
786			ALIAS:
787	root	1.8	aio_stat = REQ_STAT
788			aio_lstat = REQ_LSTAT
789	root	1.1	CODE:
790			{
791	root	1.22	dREQ;
792	root	1.1
793			New (0, req->statdata, 1, Stat_t);
794			if (!req->statdata)
795	root	1.27	{
796			free_req (req);
797			croak ("out of memory during aio_req->statdata allocation");
798			}
799	root	1.1
800			if (SvPOK (fh_or_path))
801			{
802	root	1.8	req->type = ix;
803	root	1.1	req->data = newSVsv (fh_or_path);
804	root	1.22	req->dataptr = SvPVbyte_nolen (req->data);
805	root	1.1	}
806			else
807			{
808			req->type = REQ_FSTAT;
809	root	1.13	req->fh = newSVsv (fh_or_path);
810	root	1.1	req->fd = PerlIO_fileno (IoIFP (sv_2io (fh_or_path)));
811			}
812
813			send_req (req);
814			}
815
816			void
817	root	1.8	aio_unlink(pathname,callback=&PL_sv_undef)
818	root	1.1	SV * pathname
819			SV * callback
820	root	1.22	ALIAS:
821			aio_unlink = REQ_UNLINK
822			aio_rmdir = REQ_RMDIR
823	root	1.1	CODE:
824			{
825	root	1.22	dREQ;
826	root	1.1
827	root	1.22	req->type = ix;
828			req->data = newSVsv (pathname);
829			req->dataptr = SvPVbyte_nolen (req->data);
830	root	1.1
831	root	1.22	send_req (req);
832			}
833
834			void
835			aio_symlink(oldpath,newpath,callback=&PL_sv_undef)
836			SV * oldpath
837			SV * newpath
838			SV * callback
839			CODE:
840			{
841			dREQ;
842	root	1.1
843	root	1.22	req->type = REQ_SYMLINK;
844			req->fh = newSVsv (oldpath);
845			req->data2ptr = SvPVbyte_nolen (req->fh);
846			req->data = newSVsv (newpath);
847			req->dataptr = SvPVbyte_nolen (req->data);
848	root	1.1
849			send_req (req);
850			}
851
852	root	1.6	void
853			flush()
854			PROTOTYPE:
855			CODE:
856			while (nreqs)
857			{
858			poll_wait ();
859			poll_cb ();
860			}
861
862	root	1.7	void
863			poll()
864			PROTOTYPE:
865			CODE:
866			if (nreqs)
867			{
868			poll_wait ();
869			poll_cb ();
870			}
871
872	root	1.1	int
873			poll_fileno()
874			PROTOTYPE:
875			CODE:
876	root	1.3	RETVAL = respipe [0];
877	root	1.1	OUTPUT:
878			RETVAL
879
880			int
881			poll_cb(...)
882			PROTOTYPE:
883			CODE:
884	root	1.5	RETVAL = poll_cb ();
885	root	1.1	OUTPUT:
886			RETVAL
887
888			void
889			poll_wait()
890			PROTOTYPE:
891			CODE:
892	root	1.3	if (nreqs)
893			poll_wait ();
894	root	1.1
895			int
896			nreqs()
897			PROTOTYPE:
898			CODE:
899			RETVAL = nreqs;
900			OUTPUT:
901			RETVAL
902