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_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->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)
            {
              rese = 0;

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

      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)
{
  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)
{
  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 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;

  int restart = started;
  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;

  atfork_parent ();

  min_parallel (restart);
}

/*****************************************************************************/
/* 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 = 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
        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;

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

          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:	1.28
Committed:	Wed Aug 17 05:06:59 2005 UTC (18 years, 9 months ago) by root
Branch:	MAIN
CVS Tags:	rel-1_3
Changes since 1.27:	+30 -22 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			static pthread_cond_t reqwait = PTHREAD_COND_INITIALIZER;
68
69			static volatile aio_req reqs, reqe; /* queue start, queue end */
70			static volatile aio_req ress, rese; /* queue start, queue end */
71	root	1.1
72	root	1.26	static void free_req (aio_req req)
73			{
74			if (req->data)
75			SvREFCNT_dec (req->data);
76
77			if (req->fh)
78			SvREFCNT_dec (req->fh);
79
80	root	1.27	if (req->statdata)
81	root	1.26	Safefree (req->statdata);
82
83			if (req->callback)
84			SvREFCNT_dec (req->callback);
85
86			Safefree (req);
87			}
88
89	root	1.1	static void
90			poll_wait ()
91			{
92	root	1.11	if (nreqs && !ress)
93			{
94			fd_set rfd;
95			FD_ZERO(&rfd);
96			FD_SET(respipe [0], &rfd);
97	root	1.3
98	root	1.11	select (respipe [0] + 1, &rfd, 0, 0, 0);
99			}
100	root	1.1	}
101
102			static int
103	root	1.5	poll_cb ()
104	root	1.1	{
105			dSP;
106			int count = 0;
107	root	1.24	int do_croak = 0;
108	root	1.27	aio_req req;
109
110			for (;;)
111			{
112			pthread_mutex_lock (&reslock);
113			req = ress;
114
115			if (req)
116			{
117			ress = req->next;
118	root	1.11
119	root	1.27	if (!ress)
120			{
121			rese = 0;
122	root	1.11
123	root	1.27	/* read any signals sent by the worker threads */
124			char buf [32];
125			while (read (respipe [0], buf, 32) == 32)
126			;
127			}
128			}
129	root	1.1
130	root	1.27	pthread_mutex_unlock (&reslock);
131	root	1.3
132	root	1.27	if (!req)
133			break;
134	root	1.3
135	root	1.1	nreqs--;
136
137			if (req->type == REQ_QUIT)
138			started--;
139			else
140			{
141			int errorno = errno;
142			errno = req->errorno;
143
144			if (req->type == REQ_READ)
145	root	1.27	SvCUR_set (req->data, req->dataoffset + (req->result > 0 ? req->result : 0));
146	root	1.1
147	root	1.28	if (req->data2ptr && (req->type == REQ_READ \|\| req->type == REQ_WRITE))
148			SvREADONLY_off (req->data);
149
150	root	1.27	if (req->statdata)
151	root	1.1	{
152			PL_laststype = req->type == REQ_LSTAT ? OP_LSTAT : OP_STAT;
153			PL_laststatval = req->result;
154			PL_statcache = *(req->statdata);
155			}
156
157	root	1.13	ENTER;
158	root	1.1	PUSHMARK (SP);
159			XPUSHs (sv_2mortal (newSViv (req->result)));
160	root	1.2
161			if (req->type == REQ_OPEN)
162			{
163			/* convert fd to fh */
164			SV *fh;
165
166			PUTBACK;
167			call_pv ("IO::AIO::_fd2fh", G_SCALAR \| G_EVAL);
168			SPAGAIN;
169
170	root	1.13	fh = SvREFCNT_inc (POPs);
171	root	1.2
172			PUSHMARK (SP);
173	root	1.13	XPUSHs (sv_2mortal (fh));
174	root	1.2	}
175
176	root	1.8	if (SvOK (req->callback))
177			{
178			PUTBACK;
179			call_sv (req->callback, G_VOID \| G_EVAL);
180			SPAGAIN;
181	root	1.27
182			if (SvTRUE (ERRSV))
183			{
184			free_req (req);
185			croak (0);
186			}
187	root	1.8	}
188	root	1.13
189	root	1.26	LEAVE;
190
191	root	1.1	errno = errorno;
192			count++;
193			}
194
195	root	1.27	free_req (req);
196	root	1.1	}
197
198			return count;
199			}
200
201	root	1.4	static void aio_proc(void arg);
202
203			static void
204			start_thread (void)
205			{
206			sigset_t fullsigset, oldsigset;
207			pthread_t tid;
208			pthread_attr_t attr;
209
210			pthread_attr_init (&attr);
211			pthread_attr_setstacksize (&attr, STACKSIZE);
212			pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED);
213
214			sigfillset (&fullsigset);
215			sigprocmask (SIG_SETMASK, &fullsigset, &oldsigset);
216
217			if (pthread_create (&tid, &attr, aio_proc, 0) == 0)
218			started++;
219
220			sigprocmask (SIG_SETMASK, &oldsigset, 0);
221			}
222
223			static void
224			send_req (aio_req req)
225			{
226			nreqs++;
227
228			pthread_mutex_lock (&reqlock);
229
230			req->next = 0;
231
232			if (reqe)
233			{
234			reqe->next = req;
235			reqe = req;
236			}
237			else
238			reqe = reqs = req;
239
240			pthread_cond_signal (&reqwait);
241			pthread_mutex_unlock (&reqlock);
242
243	root	1.27	if (nreqs > max_outstanding)
244			for (;;)
245			{
246			poll_cb ();
247
248			if (nreqs <= max_outstanding)
249			break;
250
251			poll_wait ();
252			}
253	root	1.4	}
254
255			static void
256			end_thread (void)
257			{
258			aio_req req;
259	root	1.26	Newz (0, req, 1, aio_cb);
260	root	1.4	req->type = REQ_QUIT;
261
262			send_req (req);
263			}
264
265	root	1.22	static void min_parallel (int nthreads)
266			{
267			while (nthreads > started)
268			start_thread ();
269			}
270
271			static void max_parallel (int nthreads)
272			{
273			int cur = started;
274	root	1.27
275	root	1.22	while (cur > nthreads)
276			{
277			end_thread ();
278			cur--;
279			}
280
281			while (started > nthreads)
282			{
283			poll_wait ();
284			poll_cb ();
285			}
286			}
287
288	root	1.26	static void create_pipe ()
289	root	1.22	{
290	root	1.26	if (pipe (respipe))
291			croak ("unable to initialize result pipe");
292	root	1.22
293	root	1.26	if (fcntl (respipe [0], F_SETFL, O_NONBLOCK))
294			croak ("cannot set result pipe to nonblocking mode");
295	root	1.22
296	root	1.26	if (fcntl (respipe [1], F_SETFL, O_NONBLOCK))
297			croak ("cannot set result pipe to nonblocking mode");
298			}
299	root	1.22
300	root	1.26	static void atfork_prepare (void)
301			{
302	root	1.28	pthread_mutex_lock (&reqlock);
303	root	1.22	pthread_mutex_lock (&reslock);
304			}
305
306			static void atfork_parent (void)
307			{
308			pthread_mutex_unlock (&reslock);
309	root	1.26	pthread_mutex_unlock (&reqlock);
310			}
311
312			static void atfork_child (void)
313			{
314	root	1.28	aio_req prv;
315
316	root	1.26	int restart = started;
317			started = 0;
318
319	root	1.28	while (reqs)
320			{
321			prv = reqs;
322			reqs = prv->next;
323			free_req (prv);
324			}
325
326			reqs = reqe = 0;
327
328			while (ress)
329			{
330			prv = ress;
331			ress = prv->next;
332			free_req (prv);
333			}
334
335			ress = rese = 0;
336
337	root	1.26	atfork_parent ();
338	root	1.25
339	root	1.26	min_parallel (restart);
340	root	1.22	}
341
342			/*****************************************************************************/
343	root	1.17	/* work around various missing functions */
344
345			#if !HAVE_PREADWRITE
346			# define pread aio_pread
347			# define pwrite aio_pwrite
348
349			/*
350			* make our pread/pwrite safe against themselves, but not against
351			* normal read/write by using a mutex. slows down execution a lot,
352			* but that's your problem, not mine.
353			*/
354			static pthread_mutex_t iolock = PTHREAD_MUTEX_INITIALIZER;
355
356			static ssize_t
357			pread (int fd, void *buf, size_t count, off_t offset)
358			{
359			ssize_t res;
360			off_t ooffset;
361
362			pthread_mutex_lock (&iolock);
363			ooffset = lseek (fd, 0, SEEK_CUR);
364			lseek (fd, offset, SEEK_SET);
365			res = read (fd, buf, count);
366			lseek (fd, ooffset, SEEK_SET);
367			pthread_mutex_unlock (&iolock);
368
369			return res;
370			}
371
372			static ssize_t
373			pwrite (int fd, void *buf, size_t count, off_t offset)
374			{
375			ssize_t res;
376			off_t ooffset;
377
378			pthread_mutex_lock (&iolock);
379			ooffset = lseek (fd, 0, SEEK_CUR);
380			lseek (fd, offset, SEEK_SET);
381			res = write (fd, buf, count);
382			lseek (fd, offset, SEEK_SET);
383			pthread_mutex_unlock (&iolock);
384
385			return res;
386			}
387			#endif
388
389			#if !HAVE_FDATASYNC
390			# define fdatasync fsync
391			#endif
392
393			#if !HAVE_READAHEAD
394			# define readahead aio_readahead
395
396			static char readahead_buf[4096];
397
398			static ssize_t
399			readahead (int fd, off_t offset, size_t count)
400			{
401			while (count > 0)
402			{
403			size_t len = count < sizeof (readahead_buf) ? count : sizeof (readahead_buf);
404
405			pread (fd, readahead_buf, len, offset);
406			offset += len;
407			count -= len;
408			}
409
410			errno = 0;
411			}
412			#endif
413
414	root	1.22	/*****************************************************************************/
415
416	root	1.1	static void *
417			aio_proc (void *thr_arg)
418			{
419			aio_req req;
420	root	1.3	int type;
421	root	1.1
422	root	1.3	do
423	root	1.1	{
424	root	1.3	pthread_mutex_lock (&reqlock);
425
426			for (;;)
427			{
428			req = reqs;
429
430			if (reqs)
431			{
432			reqs = reqs->next;
433			if (!reqs) reqe = 0;
434			}
435
436			if (req)
437			break;
438
439			pthread_cond_wait (&reqwait, &reqlock);
440			}
441
442			pthread_mutex_unlock (&reqlock);
443
444	root	1.1	errno = 0; /* strictly unnecessary */
445
446	root	1.3	type = req->type;
447
448			switch (type)
449	root	1.1	{
450	root	1.10	case REQ_READ: req->result = pread (req->fd, req->dataptr, req->length, req->offset); break;
451			case REQ_WRITE: req->result = pwrite (req->fd, req->dataptr, req->length, req->offset); break;
452	root	1.16
453	root	1.1	case REQ_READAHEAD: req->result = readahead (req->fd, req->offset, req->length); break;
454
455			case REQ_STAT: req->result = stat (req->dataptr, req->statdata); break;
456			case REQ_LSTAT: req->result = lstat (req->dataptr, req->statdata); break;
457			case REQ_FSTAT: req->result = fstat (req->fd , req->statdata); break;
458
459			case REQ_OPEN: req->result = open (req->dataptr, req->fd, req->mode); break;
460			case REQ_CLOSE: req->result = close (req->fd); break;
461			case REQ_UNLINK: req->result = unlink (req->dataptr); break;
462	root	1.22	case REQ_RMDIR: req->result = rmdir (req->dataptr); break;
463			case REQ_SYMLINK: req->result = symlink (req->data2ptr, req->dataptr); break;
464	root	1.1
465	root	1.17	case REQ_FDATASYNC: req->result = fdatasync (req->fd); break;
466	root	1.1	case REQ_FSYNC: req->result = fsync (req->fd); break;
467
468			case REQ_QUIT:
469	root	1.3	break;
470	root	1.1
471			default:
472			req->result = ENOSYS;
473			break;
474			}
475
476			req->errorno = errno;
477	root	1.3
478			pthread_mutex_lock (&reslock);
479
480			req->next = 0;
481
482			if (rese)
483			{
484			rese->next = req;
485			rese = req;
486			}
487			else
488			{
489			rese = ress = req;
490
491			/* write a dummy byte to the pipe so fh becomes ready */
492			write (respipe [1], &respipe, 1);
493			}
494
495			pthread_mutex_unlock (&reslock);
496	root	1.1	}
497	root	1.3	while (type != REQ_QUIT);
498	root	1.1
499			return 0;
500			}
501
502	root	1.22	#define dREQ \
503			aio_req req; \
504			\
505			if (SvOK (callback) && !SvROK (callback)) \
506			croak ("clalback must be undef or of reference type"); \
507			\
508			Newz (0, req, 1, aio_cb); \
509			if (!req) \
510			croak ("out of memory during aio_req allocation"); \
511			\
512	root	1.27	req->callback = newSVsv (callback);
513	root	1.22
514	root	1.1	MODULE = IO::AIO PACKAGE = IO::AIO
515
516	root	1.8	PROTOTYPES: ENABLE
517
518	root	1.1	BOOT:
519			{
520	root	1.26	create_pipe ();
521	root	1.22	pthread_atfork (atfork_prepare, atfork_parent, atfork_child);
522	root	1.1	}
523
524			void
525			min_parallel(nthreads)
526			int nthreads
527			PROTOTYPE: $
528
529			void
530			max_parallel(nthreads)
531			int nthreads
532			PROTOTYPE: $
533
534	root	1.4	int
535			max_outstanding(nreqs)
536			int nreqs
537			PROTOTYPE: $
538			CODE:
539			RETVAL = max_outstanding;
540			max_outstanding = nreqs;
541
542	root	1.1	void
543	root	1.8	aio_open(pathname,flags,mode,callback=&PL_sv_undef)
544	root	1.1	SV * pathname
545			int flags
546			int mode
547			SV * callback
548	root	1.8	PROTOTYPE: $$$;$
549	root	1.1	CODE:
550			{
551	root	1.22	dREQ;
552	root	1.1
553			req->type = REQ_OPEN;
554			req->data = newSVsv (pathname);
555	root	1.22	req->dataptr = SvPVbyte_nolen (req->data);
556	root	1.1	req->fd = flags;
557			req->mode = mode;
558
559			send_req (req);
560			}
561
562			void
563	root	1.8	aio_close(fh,callback=&PL_sv_undef)
564	root	1.13	SV * fh
565			SV * callback
566	root	1.8	PROTOTYPE: $;$
567	root	1.1	ALIAS:
568			aio_close = REQ_CLOSE
569			aio_fsync = REQ_FSYNC
570			aio_fdatasync = REQ_FDATASYNC
571			CODE:
572			{
573	root	1.22	dREQ;
574	root	1.1
575			req->type = ix;
576	root	1.13	req->fh = newSVsv (fh);
577			req->fd = PerlIO_fileno (IoIFP (sv_2io (fh)));
578	root	1.1
579			send_req (req);
580			}
581
582			void
583	root	1.8	aio_read(fh,offset,length,data,dataoffset,callback=&PL_sv_undef)
584	root	1.13	SV * fh
585			UV offset
586			IV length
587			SV * data
588			IV dataoffset
589			SV * callback
590			ALIAS:
591			aio_read = REQ_READ
592			aio_write = REQ_WRITE
593	root	1.8	PROTOTYPE: $$$$$;$
594	root	1.1	CODE:
595	root	1.13	{
596			aio_req req;
597			STRLEN svlen;
598	root	1.21	char *svptr = SvPVbyte (data, svlen);
599	root	1.13
600			SvUPGRADE (data, SVt_PV);
601			SvPOK_on (data);
602	root	1.1
603	root	1.13	if (dataoffset < 0)
604			dataoffset += svlen;
605
606			if (dataoffset < 0 \|\| dataoffset > svlen)
607			croak ("data offset outside of string");
608
609			if (ix == REQ_WRITE)
610			{
611			/* write: check length and adjust. */
612			if (length < 0 \|\| length + dataoffset > svlen)
613			length = svlen - dataoffset;
614			}
615			else
616			{
617			/* read: grow scalar as necessary */
618			svptr = SvGROW (data, length + dataoffset);
619			}
620
621			if (length < 0)
622			croak ("length must not be negative");
623
624	root	1.22	{
625			dREQ;
626	root	1.13
627	root	1.22	req->type = ix;
628			req->fh = newSVsv (fh);
629			req->fd = PerlIO_fileno (ix == REQ_READ ? IoIFP (sv_2io (fh))
630			: IoOFP (sv_2io (fh)));
631			req->offset = offset;
632			req->length = length;
633			req->data = SvREFCNT_inc (data);
634			req->dataptr = (char *)svptr + dataoffset;
635	root	1.13
636	root	1.28	if (!SvREADONLY (data))
637			{
638			SvREADONLY_on (data);
639			req->data2ptr = (void *)data;
640			}
641
642	root	1.22	send_req (req);
643			}
644	root	1.13	}
645	root	1.1
646			void
647	root	1.8	aio_readahead(fh,offset,length,callback=&PL_sv_undef)
648	root	1.13	SV * fh
649			UV offset
650			IV length
651			SV * callback
652	root	1.8	PROTOTYPE: $$$;$
653	root	1.1	CODE:
654			{
655	root	1.22	dREQ;
656	root	1.1
657			req->type = REQ_READAHEAD;
658	root	1.13	req->fh = newSVsv (fh);
659			req->fd = PerlIO_fileno (IoIFP (sv_2io (fh)));
660	root	1.1	req->offset = offset;
661			req->length = length;
662
663			send_req (req);
664			}
665
666			void
667	root	1.8	aio_stat(fh_or_path,callback=&PL_sv_undef)
668	root	1.1	SV * fh_or_path
669			SV * callback
670			ALIAS:
671	root	1.8	aio_stat = REQ_STAT
672			aio_lstat = REQ_LSTAT
673	root	1.1	CODE:
674			{
675	root	1.22	dREQ;
676	root	1.1
677			New (0, req->statdata, 1, Stat_t);
678			if (!req->statdata)
679	root	1.27	{
680			free_req (req);
681			croak ("out of memory during aio_req->statdata allocation");
682			}
683	root	1.1
684			if (SvPOK (fh_or_path))
685			{
686	root	1.8	req->type = ix;
687	root	1.1	req->data = newSVsv (fh_or_path);
688	root	1.22	req->dataptr = SvPVbyte_nolen (req->data);
689	root	1.1	}
690			else
691			{
692			req->type = REQ_FSTAT;
693	root	1.13	req->fh = newSVsv (fh_or_path);
694	root	1.1	req->fd = PerlIO_fileno (IoIFP (sv_2io (fh_or_path)));
695			}
696
697			send_req (req);
698			}
699
700			void
701	root	1.8	aio_unlink(pathname,callback=&PL_sv_undef)
702	root	1.1	SV * pathname
703			SV * callback
704	root	1.22	ALIAS:
705			aio_unlink = REQ_UNLINK
706			aio_rmdir = REQ_RMDIR
707	root	1.1	CODE:
708			{
709	root	1.22	dREQ;
710	root	1.1
711	root	1.22	req->type = ix;
712			req->data = newSVsv (pathname);
713			req->dataptr = SvPVbyte_nolen (req->data);
714	root	1.1
715	root	1.22	send_req (req);
716			}
717
718			void
719			aio_symlink(oldpath,newpath,callback=&PL_sv_undef)
720			SV * oldpath
721			SV * newpath
722			SV * callback
723			CODE:
724			{
725			dREQ;
726	root	1.1
727	root	1.22	req->type = REQ_SYMLINK;
728			req->fh = newSVsv (oldpath);
729			req->data2ptr = SvPVbyte_nolen (req->fh);
730			req->data = newSVsv (newpath);
731			req->dataptr = SvPVbyte_nolen (req->data);
732	root	1.1
733			send_req (req);
734			}
735
736	root	1.6	void
737			flush()
738			PROTOTYPE:
739			CODE:
740			while (nreqs)
741			{
742			poll_wait ();
743			poll_cb ();
744			}
745
746	root	1.7	void
747			poll()
748			PROTOTYPE:
749			CODE:
750			if (nreqs)
751			{
752			poll_wait ();
753			poll_cb ();
754			}
755
756	root	1.1	int
757			poll_fileno()
758			PROTOTYPE:
759			CODE:
760	root	1.3	RETVAL = respipe [0];
761	root	1.1	OUTPUT:
762			RETVAL
763
764			int
765			poll_cb(...)
766			PROTOTYPE:
767			CODE:
768	root	1.5	RETVAL = poll_cb ();
769	root	1.1	OUTPUT:
770			RETVAL
771
772			void
773			poll_wait()
774			PROTOTYPE:
775			CODE:
776	root	1.3	if (nreqs)
777			poll_wait ();
778	root	1.1
779			int
780			nreqs()
781			PROTOTYPE:
782			CODE:
783			RETVAL = nreqs;
784			OUTPUT:
785			RETVAL
786