IO-AIO/AIO.xs

#define _REENTRANT 1
#include <errno.h>

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

#include "autoconf/config.h"

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

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

#include <pthread.h>

typedef void *InputStream;  /* hack, but 5.6.1 is simply toooo old ;) */
typedef void *OutputStream; /* hack, but 5.6.1 is simply toooo old ;) */
typedef void *InOutStream;  /* hack, but 5.6.1 is simply toooo old ;) */

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

enum {
  REQ_QUIT,
  REQ_OPEN, REQ_CLOSE,
  REQ_READ, REQ_WRITE, REQ_READAHEAD,
  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;
  off_t offset;
  size_t length;
  ssize_t result;
  mode_t mode; /* open */
  int errorno;
  SV *data, *callback, *fh;
  void *dataptr, *data2ptr;
  STRLEN dataoffset;

  Stat_t *statdata;
} aio_cb;

typedef aio_cb *aio_req;

static int started;
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_mutex_t frklock = 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
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, prv;

  pthread_mutex_lock (&reslock);

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

  req = ress;
  ress = rese = 0;

  pthread_mutex_unlock (&reslock);

  while (req)
    {
      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->data)
            SvREFCNT_dec (req->data);

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

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

              Safefree (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;
            }

          do_croak = SvTRUE (ERRSV);

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

          errno = errorno;
          count++;
        }

      prv = req;
      req = req->next;
      Safefree (prv);

      if (do_croak)
        croak (0);
    }

  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)
{
  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);

  while (nreqs > max_outstanding)
    {
      poll_wait ();
      poll_cb ();
    }
}

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

  send_req (req);
}


static void min_parallel (int nthreads)
{
  while (nthreads > started)
    start_thread ();
}

static void max_parallel (int nthreads)
{
  int cur = started;
  while (cur > nthreads)
    {          
      end_thread ();
      cur--;
    }

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

static int fork_started;

static void atfork_prepare (void)
{
  pthread_mutex_lock (&frklock);

  fork_started = started;

  for (;;) {
    while (nreqs)
      {
        poll_wait ();
        poll_cb ();
      }

    max_parallel (0);
  
    pthread_mutex_lock (&reqlock);

    if (!nreqs && !started)
      break;

    pthread_mutex_unlock (&reqlock);

    min_parallel (fork_started);
  }

  pthread_mutex_lock (&reslock);

  assert (!started);
  assert (!nreqs);
  assert (!reqs && !reqe);
  assert (!ress && !rese);
}

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

static void atfork_child (void)
{
  reqs = reqe = 0;

  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

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

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_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 = SvREFCNT_inc (callback);
        
MODULE = IO::AIO                PACKAGE = IO::AIO

PROTOTYPES: ENABLE

BOOT:
{
        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");

        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
        IV      length
        SV *    data
        IV      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;
          req->callback = SvREFCNT_inc (callback);

          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)
          croak ("out of memory during aio_req->statdata allocation (sorry, i just leaked memory, too)");

        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:	1.24
Committed:	Wed Aug 17 03:01:56 2005 UTC (18 years, 9 months ago) by root
Branch:	MAIN
Changes since 1.23:	+5 -1 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	1.1	#include <sys/types.h>
11			#include <sys/stat.h>
12	root	1.10
13	root	1.1	#include <unistd.h>
14			#include <fcntl.h>
15			#include <signal.h>
16			#include <sched.h>
17
18			#include <pthread.h>
19
20			typedef void InputStream; / hack, but 5.6.1 is simply toooo old ;) */
21			typedef void OutputStream; / hack, but 5.6.1 is simply toooo old ;) */
22			typedef void InOutStream; / hack, but 5.6.1 is simply toooo old ;) */
23
24	root	1.4	#if __ia64
25			# define STACKSIZE 65536
26	root	1.1	#else
27	root	1.4	# define STACKSIZE 4096
28	root	1.1	#endif
29
30			enum {
31			REQ_QUIT,
32			REQ_OPEN, REQ_CLOSE,
33			REQ_READ, REQ_WRITE, REQ_READAHEAD,
34	root	1.22	REQ_STAT, REQ_LSTAT, REQ_FSTAT,
35	root	1.1	REQ_FSYNC, REQ_FDATASYNC,
36	root	1.23	REQ_UNLINK, REQ_RMDIR,
37	root	1.22	REQ_SYMLINK,
38	root	1.1	};
39
40			typedef struct aio_cb {
41	root	1.3	struct aio_cb *volatile next;
42	root	1.1
43			int type;
44
45			int fd;
46			off_t offset;
47			size_t length;
48			ssize_t result;
49			mode_t mode; /* open */
50			int errorno;
51	root	1.13	SV data, callback, *fh;
52	root	1.22	void dataptr, data2ptr;
53	root	1.1	STRLEN dataoffset;
54
55			Stat_t *statdata;
56			} aio_cb;
57
58			typedef aio_cb *aio_req;
59
60			static int started;
61	root	1.11	static volatile int nreqs;
62	root	1.4	static int max_outstanding = 1<<30;
63	root	1.3	static int respipe [2];
64	root	1.1
65	root	1.3	static pthread_mutex_t reslock = PTHREAD_MUTEX_INITIALIZER;
66			static pthread_mutex_t reqlock = PTHREAD_MUTEX_INITIALIZER;
67	root	1.22	static pthread_mutex_t frklock = PTHREAD_MUTEX_INITIALIZER;
68	root	1.3	static pthread_cond_t reqwait = PTHREAD_COND_INITIALIZER;
69
70			static volatile aio_req reqs, reqe; /* queue start, queue end */
71			static volatile aio_req ress, rese; /* queue start, queue end */
72	root	1.1
73			static void
74			poll_wait ()
75			{
76	root	1.11	if (nreqs && !ress)
77			{
78			fd_set rfd;
79			FD_ZERO(&rfd);
80			FD_SET(respipe [0], &rfd);
81	root	1.3
82	root	1.11	select (respipe [0] + 1, &rfd, 0, 0, 0);
83			}
84	root	1.1	}
85
86			static int
87	root	1.5	poll_cb ()
88	root	1.1	{
89			dSP;
90			int count = 0;
91	root	1.24	int do_croak = 0;
92	root	1.11	aio_req req, prv;
93
94			pthread_mutex_lock (&reslock);
95
96	root	1.3	{
97	root	1.11	/* read any signals sent by the worker threads */
98	root	1.3	char buf [32];
99	root	1.22	while (read (respipe [0], buf, 32) == 32)
100	root	1.3	;
101			}
102	root	1.1
103	root	1.11	req = ress;
104			ress = rese = 0;
105	root	1.3
106	root	1.11	pthread_mutex_unlock (&reslock);
107	root	1.3
108	root	1.11	while (req)
109			{
110	root	1.1	nreqs--;
111
112			if (req->type == REQ_QUIT)
113			started--;
114			else
115			{
116			int errorno = errno;
117			errno = req->errorno;
118
119			if (req->type == REQ_READ)
120			SvCUR_set (req->data, req->dataoffset
121			+ req->result > 0 ? req->result : 0);
122
123			if (req->data)
124			SvREFCNT_dec (req->data);
125
126	root	1.13	if (req->fh)
127			SvREFCNT_dec (req->fh);
128
129	root	1.1	if (req->type == REQ_STAT \|\| req->type == REQ_LSTAT \|\| req->type == REQ_FSTAT)
130			{
131			PL_laststype = req->type == REQ_LSTAT ? OP_LSTAT : OP_STAT;
132			PL_laststatval = req->result;
133			PL_statcache = *(req->statdata);
134
135			Safefree (req->statdata);
136			}
137
138	root	1.13	ENTER;
139	root	1.1	PUSHMARK (SP);
140			XPUSHs (sv_2mortal (newSViv (req->result)));
141	root	1.2
142			if (req->type == REQ_OPEN)
143			{
144			/* convert fd to fh */
145			SV *fh;
146
147			PUTBACK;
148			call_pv ("IO::AIO::_fd2fh", G_SCALAR \| G_EVAL);
149			SPAGAIN;
150
151	root	1.13	fh = SvREFCNT_inc (POPs);
152	root	1.2
153			PUSHMARK (SP);
154	root	1.13	XPUSHs (sv_2mortal (fh));
155	root	1.2	}
156
157	root	1.8	if (SvOK (req->callback))
158			{
159			PUTBACK;
160			call_sv (req->callback, G_VOID \| G_EVAL);
161			SPAGAIN;
162			}
163	root	1.13
164	root	1.24	do_croak = SvTRUE (ERRSV);
165
166	root	1.13	LEAVE;
167	root	1.1
168			if (req->callback)
169			SvREFCNT_dec (req->callback);
170
171			errno = errorno;
172			count++;
173			}
174
175	root	1.11	prv = req;
176			req = req->next;
177			Safefree (prv);
178
179	root	1.24	if (do_croak)
180			croak (0);
181	root	1.1	}
182
183			return count;
184			}
185
186	root	1.4	static void aio_proc(void arg);
187
188			static void
189			start_thread (void)
190			{
191			sigset_t fullsigset, oldsigset;
192			pthread_t tid;
193			pthread_attr_t attr;
194
195			pthread_attr_init (&attr);
196			pthread_attr_setstacksize (&attr, STACKSIZE);
197			pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED);
198
199			sigfillset (&fullsigset);
200			sigprocmask (SIG_SETMASK, &fullsigset, &oldsigset);
201
202			if (pthread_create (&tid, &attr, aio_proc, 0) == 0)
203			started++;
204
205			sigprocmask (SIG_SETMASK, &oldsigset, 0);
206			}
207
208			static void
209			send_req (aio_req req)
210			{
211			nreqs++;
212
213			pthread_mutex_lock (&reqlock);
214
215			req->next = 0;
216
217			if (reqe)
218			{
219			reqe->next = req;
220			reqe = req;
221			}
222			else
223			reqe = reqs = req;
224
225			pthread_cond_signal (&reqwait);
226			pthread_mutex_unlock (&reqlock);
227
228			while (nreqs > max_outstanding)
229			{
230			poll_wait ();
231			poll_cb ();
232			}
233			}
234
235			static void
236			end_thread (void)
237			{
238			aio_req req;
239			New (0, req, 1, aio_cb);
240			req->type = REQ_QUIT;
241
242			send_req (req);
243			}
244
245	root	1.22
246			static void min_parallel (int nthreads)
247			{
248			while (nthreads > started)
249			start_thread ();
250			}
251
252			static void max_parallel (int nthreads)
253			{
254			int cur = started;
255			while (cur > nthreads)
256			{
257			end_thread ();
258			cur--;
259			}
260
261			while (started > nthreads)
262			{
263			poll_wait ();
264			poll_cb ();
265			}
266			}
267
268			static int fork_started;
269
270			static void atfork_prepare (void)
271			{
272			pthread_mutex_lock (&frklock);
273
274			fork_started = started;
275
276			for (;;) {
277			while (nreqs)
278			{
279			poll_wait ();
280			poll_cb ();
281			}
282
283			max_parallel (0);
284
285			pthread_mutex_lock (&reqlock);
286
287			if (!nreqs && !started)
288			break;
289
290			pthread_mutex_unlock (&reqlock);
291
292			min_parallel (fork_started);
293			}
294
295			pthread_mutex_lock (&reslock);
296
297			assert (!started);
298			assert (!nreqs);
299			assert (!reqs && !reqe);
300			assert (!ress && !rese);
301			}
302
303			static void atfork_parent (void)
304			{
305			pthread_mutex_unlock (&reslock);
306			min_parallel (fork_started);
307			pthread_mutex_unlock (&reqlock);
308			pthread_mutex_unlock (&frklock);
309			}
310
311			static void atfork_child (void)
312			{
313			reqs = reqe = 0;
314
315			atfork_parent ();
316			}
317
318			/*****************************************************************************/
319	root	1.17	/* work around various missing functions */
320
321			#if !HAVE_PREADWRITE
322			# define pread aio_pread
323			# define pwrite aio_pwrite
324
325			/*
326			* make our pread/pwrite safe against themselves, but not against
327			* normal read/write by using a mutex. slows down execution a lot,
328			* but that's your problem, not mine.
329			*/
330			static pthread_mutex_t iolock = PTHREAD_MUTEX_INITIALIZER;
331
332			static ssize_t
333			pread (int fd, void *buf, size_t count, off_t offset)
334			{
335			ssize_t res;
336			off_t ooffset;
337
338			pthread_mutex_lock (&iolock);
339			ooffset = lseek (fd, 0, SEEK_CUR);
340			lseek (fd, offset, SEEK_SET);
341			res = read (fd, buf, count);
342			lseek (fd, ooffset, SEEK_SET);
343			pthread_mutex_unlock (&iolock);
344
345			return res;
346			}
347
348			static ssize_t
349			pwrite (int fd, void *buf, size_t count, off_t offset)
350			{
351			ssize_t res;
352			off_t ooffset;
353
354			pthread_mutex_lock (&iolock);
355			ooffset = lseek (fd, 0, SEEK_CUR);
356			lseek (fd, offset, SEEK_SET);
357			res = write (fd, buf, count);
358			lseek (fd, offset, SEEK_SET);
359			pthread_mutex_unlock (&iolock);
360
361			return res;
362			}
363			#endif
364
365			#if !HAVE_FDATASYNC
366			# define fdatasync fsync
367			#endif
368
369			#if !HAVE_READAHEAD
370			# define readahead aio_readahead
371
372			static char readahead_buf[4096];
373
374			static ssize_t
375			readahead (int fd, off_t offset, size_t count)
376			{
377			while (count > 0)
378			{
379			size_t len = count < sizeof (readahead_buf) ? count : sizeof (readahead_buf);
380
381			pread (fd, readahead_buf, len, offset);
382			offset += len;
383			count -= len;
384			}
385
386			errno = 0;
387			}
388			#endif
389
390	root	1.22	/*****************************************************************************/
391
392	root	1.1	static void *
393			aio_proc (void *thr_arg)
394			{
395			aio_req req;
396	root	1.3	int type;
397	root	1.1
398	root	1.3	do
399	root	1.1	{
400	root	1.3	pthread_mutex_lock (&reqlock);
401
402			for (;;)
403			{
404			req = reqs;
405
406			if (reqs)
407			{
408			reqs = reqs->next;
409			if (!reqs) reqe = 0;
410			}
411
412			if (req)
413			break;
414
415			pthread_cond_wait (&reqwait, &reqlock);
416			}
417
418			pthread_mutex_unlock (&reqlock);
419
420	root	1.1	errno = 0; /* strictly unnecessary */
421
422	root	1.3	type = req->type;
423
424			switch (type)
425	root	1.1	{
426	root	1.10	case REQ_READ: req->result = pread (req->fd, req->dataptr, req->length, req->offset); break;
427			case REQ_WRITE: req->result = pwrite (req->fd, req->dataptr, req->length, req->offset); break;
428	root	1.16
429	root	1.1	case REQ_READAHEAD: req->result = readahead (req->fd, req->offset, req->length); break;
430
431			case REQ_STAT: req->result = stat (req->dataptr, req->statdata); break;
432			case REQ_LSTAT: req->result = lstat (req->dataptr, req->statdata); break;
433			case REQ_FSTAT: req->result = fstat (req->fd , req->statdata); break;
434
435			case REQ_OPEN: req->result = open (req->dataptr, req->fd, req->mode); break;
436			case REQ_CLOSE: req->result = close (req->fd); break;
437			case REQ_UNLINK: req->result = unlink (req->dataptr); break;
438	root	1.22	case REQ_RMDIR: req->result = rmdir (req->dataptr); break;
439			case REQ_SYMLINK: req->result = symlink (req->data2ptr, req->dataptr); break;
440	root	1.1
441	root	1.17	case REQ_FDATASYNC: req->result = fdatasync (req->fd); break;
442	root	1.1	case REQ_FSYNC: req->result = fsync (req->fd); break;
443
444			case REQ_QUIT:
445	root	1.3	break;
446	root	1.1
447			default:
448			req->result = ENOSYS;
449			break;
450			}
451
452			req->errorno = errno;
453	root	1.3
454			pthread_mutex_lock (&reslock);
455
456			req->next = 0;
457
458			if (rese)
459			{
460			rese->next = req;
461			rese = req;
462			}
463			else
464			{
465			rese = ress = req;
466
467			/* write a dummy byte to the pipe so fh becomes ready */
468			write (respipe [1], &respipe, 1);
469			}
470
471			pthread_mutex_unlock (&reslock);
472	root	1.1	}
473	root	1.3	while (type != REQ_QUIT);
474	root	1.1
475			return 0;
476			}
477
478	root	1.22	#define dREQ \
479			aio_req req; \
480			\
481			if (SvOK (callback) && !SvROK (callback)) \
482			croak ("clalback must be undef or of reference type"); \
483			\
484			Newz (0, req, 1, aio_cb); \
485			if (!req) \
486			croak ("out of memory during aio_req allocation"); \
487			\
488			req->callback = SvREFCNT_inc (callback);
489
490	root	1.1	MODULE = IO::AIO PACKAGE = IO::AIO
491
492	root	1.8	PROTOTYPES: ENABLE
493
494	root	1.1	BOOT:
495			{
496	root	1.3	if (pipe (respipe))
497			croak ("unable to initialize result pipe");
498	root	1.1
499	root	1.3	if (fcntl (respipe [0], F_SETFL, O_NONBLOCK))
500	root	1.1	croak ("cannot set result pipe to nonblocking mode");
501
502	root	1.3	if (fcntl (respipe [1], F_SETFL, O_NONBLOCK))
503	root	1.1	croak ("cannot set result pipe to nonblocking mode");
504	root	1.22
505			pthread_atfork (atfork_prepare, atfork_parent, atfork_child);
506	root	1.1	}
507
508			void
509			min_parallel(nthreads)
510			int nthreads
511			PROTOTYPE: $
512
513			void
514			max_parallel(nthreads)
515			int nthreads
516			PROTOTYPE: $
517
518	root	1.4	int
519			max_outstanding(nreqs)
520			int nreqs
521			PROTOTYPE: $
522			CODE:
523			RETVAL = max_outstanding;
524			max_outstanding = nreqs;
525
526	root	1.1	void
527	root	1.8	aio_open(pathname,flags,mode,callback=&PL_sv_undef)
528	root	1.1	SV * pathname
529			int flags
530			int mode
531			SV * callback
532	root	1.8	PROTOTYPE: $$$;$
533	root	1.1	CODE:
534			{
535	root	1.22	dREQ;
536	root	1.1
537			req->type = REQ_OPEN;
538			req->data = newSVsv (pathname);
539	root	1.22	req->dataptr = SvPVbyte_nolen (req->data);
540	root	1.1	req->fd = flags;
541			req->mode = mode;
542
543			send_req (req);
544			}
545
546			void
547	root	1.8	aio_close(fh,callback=&PL_sv_undef)
548	root	1.13	SV * fh
549			SV * callback
550	root	1.8	PROTOTYPE: $;$
551	root	1.1	ALIAS:
552			aio_close = REQ_CLOSE
553			aio_fsync = REQ_FSYNC
554			aio_fdatasync = REQ_FDATASYNC
555			CODE:
556			{
557	root	1.22	dREQ;
558	root	1.1
559			req->type = ix;
560	root	1.13	req->fh = newSVsv (fh);
561			req->fd = PerlIO_fileno (IoIFP (sv_2io (fh)));
562	root	1.1
563			send_req (req);
564			}
565
566			void
567	root	1.8	aio_read(fh,offset,length,data,dataoffset,callback=&PL_sv_undef)
568	root	1.13	SV * fh
569			UV offset
570			IV length
571			SV * data
572			IV dataoffset
573			SV * callback
574			ALIAS:
575			aio_read = REQ_READ
576			aio_write = REQ_WRITE
577	root	1.8	PROTOTYPE: $$$$$;$
578	root	1.1	CODE:
579	root	1.13	{
580			aio_req req;
581			STRLEN svlen;
582	root	1.21	char *svptr = SvPVbyte (data, svlen);
583	root	1.13
584			SvUPGRADE (data, SVt_PV);
585			SvPOK_on (data);
586	root	1.1
587	root	1.13	if (dataoffset < 0)
588			dataoffset += svlen;
589
590			if (dataoffset < 0 \|\| dataoffset > svlen)
591			croak ("data offset outside of string");
592
593			if (ix == REQ_WRITE)
594			{
595			/* write: check length and adjust. */
596			if (length < 0 \|\| length + dataoffset > svlen)
597			length = svlen - dataoffset;
598			}
599			else
600			{
601			/* read: grow scalar as necessary */
602			svptr = SvGROW (data, length + dataoffset);
603			}
604
605			if (length < 0)
606			croak ("length must not be negative");
607
608	root	1.22	{
609			dREQ;
610	root	1.13
611	root	1.22	req->type = ix;
612			req->fh = newSVsv (fh);
613			req->fd = PerlIO_fileno (ix == REQ_READ ? IoIFP (sv_2io (fh))
614			: IoOFP (sv_2io (fh)));
615			req->offset = offset;
616			req->length = length;
617			req->data = SvREFCNT_inc (data);
618			req->dataptr = (char *)svptr + dataoffset;
619			req->callback = SvREFCNT_inc (callback);
620	root	1.13
621	root	1.22	send_req (req);
622			}
623	root	1.13	}
624	root	1.1
625			void
626	root	1.8	aio_readahead(fh,offset,length,callback=&PL_sv_undef)
627	root	1.13	SV * fh
628			UV offset
629			IV length
630			SV * callback
631	root	1.8	PROTOTYPE: $$$;$
632	root	1.1	CODE:
633			{
634	root	1.22	dREQ;
635	root	1.1
636			req->type = REQ_READAHEAD;
637	root	1.13	req->fh = newSVsv (fh);
638			req->fd = PerlIO_fileno (IoIFP (sv_2io (fh)));
639	root	1.1	req->offset = offset;
640			req->length = length;
641
642			send_req (req);
643			}
644
645			void
646	root	1.8	aio_stat(fh_or_path,callback=&PL_sv_undef)
647	root	1.1	SV * fh_or_path
648			SV * callback
649			ALIAS:
650	root	1.8	aio_stat = REQ_STAT
651			aio_lstat = REQ_LSTAT
652	root	1.1	CODE:
653			{
654	root	1.22	dREQ;
655	root	1.1
656			New (0, req->statdata, 1, Stat_t);
657			if (!req->statdata)
658	root	1.22	croak ("out of memory during aio_req->statdata allocation (sorry, i just leaked memory, too)");
659	root	1.1
660			if (SvPOK (fh_or_path))
661			{
662	root	1.8	req->type = ix;
663	root	1.1	req->data = newSVsv (fh_or_path);
664	root	1.22	req->dataptr = SvPVbyte_nolen (req->data);
665	root	1.1	}
666			else
667			{
668			req->type = REQ_FSTAT;
669	root	1.13	req->fh = newSVsv (fh_or_path);
670	root	1.1	req->fd = PerlIO_fileno (IoIFP (sv_2io (fh_or_path)));
671			}
672
673			send_req (req);
674			}
675
676			void
677	root	1.8	aio_unlink(pathname,callback=&PL_sv_undef)
678	root	1.1	SV * pathname
679			SV * callback
680	root	1.22	ALIAS:
681			aio_unlink = REQ_UNLINK
682			aio_rmdir = REQ_RMDIR
683	root	1.1	CODE:
684			{
685	root	1.22	dREQ;
686	root	1.1
687	root	1.22	req->type = ix;
688			req->data = newSVsv (pathname);
689			req->dataptr = SvPVbyte_nolen (req->data);
690	root	1.1
691	root	1.22	send_req (req);
692			}
693
694			void
695			aio_symlink(oldpath,newpath,callback=&PL_sv_undef)
696			SV * oldpath
697			SV * newpath
698			SV * callback
699			CODE:
700			{
701			dREQ;
702	root	1.1
703	root	1.22	req->type = REQ_SYMLINK;
704			req->fh = newSVsv (oldpath);
705			req->data2ptr = SvPVbyte_nolen (req->fh);
706			req->data = newSVsv (newpath);
707			req->dataptr = SvPVbyte_nolen (req->data);
708	root	1.1
709			send_req (req);
710			}
711
712	root	1.6	void
713			flush()
714			PROTOTYPE:
715			CODE:
716			while (nreqs)
717			{
718			poll_wait ();
719			poll_cb ();
720			}
721
722	root	1.7	void
723			poll()
724			PROTOTYPE:
725			CODE:
726			if (nreqs)
727			{
728			poll_wait ();
729			poll_cb ();
730			}
731
732	root	1.1	int
733			poll_fileno()
734			PROTOTYPE:
735			CODE:
736	root	1.3	RETVAL = respipe [0];
737	root	1.1	OUTPUT:
738			RETVAL
739
740			int
741			poll_cb(...)
742			PROTOTYPE:
743			CODE:
744	root	1.5	RETVAL = poll_cb ();
745	root	1.1	OUTPUT:
746			RETVAL
747
748			void
749			poll_wait()
750			PROTOTYPE:
751			CODE:
752	root	1.3	if (nreqs)
753			poll_wait ();
754	root	1.1
755			int
756			nreqs()
757			PROTOTYPE:
758			CODE:
759			RETVAL = nreqs;
760			OUTPUT:
761			RETVAL
762