libeio/etp.c

/*
 * libetp implementation
 *
 * Copyright (c) 2007,2008,2009,2010,2011,2012,2013 Marc Alexander Lehmann <libetp@schmorp.de>
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without modifica-
 * tion, are permitted provided that the following conditions are met:
 *
 *   1.  Redistributions of source code must retain the above copyright notice,
 *       this list of conditions and the following disclaimer.
 *
 *   2.  Redistributions in binary form must reproduce the above copyright
 *       notice, this list of conditions and the following disclaimer in the
 *       documentation and/or other materials provided with the distribution.
 *
 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED
 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MER-
 * CHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO
 * EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPE-
 * CIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
 * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTH-
 * ERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
 * OF THE POSSIBILITY OF SUCH DAMAGE.
 *
 * Alternatively, the contents of this file may be used under the terms of
 * the GNU General Public License ("GPL") version 2 or any later version,
 * in which case the provisions of the GPL are applicable instead of
 * the above. If you wish to allow the use of your version of this file
 * only under the terms of the GPL and not to allow others to use your
 * version of this file under the BSD license, indicate your decision
 * by deleting the provisions above and replace them with the notice
 * and other provisions required by the GPL. If you do not delete the
 * provisions above, a recipient may use your version of this file under
 * either the BSD or the GPL.
 */

#ifndef ETP_API_DECL
# define ETP_API_DECL static
#endif

#ifndef ETP_PRI_MIN
# define ETP_PRI_MIN 0
# define ETP_PRI_MAX 0
#endif

#ifndef ETP_TYPE_QUIT
# define ETP_TYPE_QUIT 0
#endif

#ifndef ETP_TYPE_GROUP
# define ETP_TYPE_GROUP 1
#endif

#ifndef ETP_WANT_POLL
# define ETP_WANT_POLL(pool) pool->want_poll_cb (pool->userdata)
#endif
#ifndef ETP_DONE_POLL
# define ETP_DONE_POLL(pool) pool->done_poll_cb (pool->userdata)
#endif

#define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1)

#define ETP_TICKS ((1000000 + 1023) >> 10)

enum {
  ETP_FLAG_GROUPADD = 0x04, /* some request was added to the group */
  ETP_FLAG_DELAYED  = 0x08, /* groiup request has been delayed */
};

/* calculate time difference in ~1/ETP_TICKS of a second */
ecb_inline int
etp_tvdiff (struct timeval *tv1, struct timeval *tv2)
{
  return  (tv2->tv_sec  - tv1->tv_sec ) * ETP_TICKS
       + ((tv2->tv_usec - tv1->tv_usec) >> 10);
}

struct etp_tmpbuf
{
  void *ptr;
  int len;
};

static void *
etp_tmpbuf_get (struct etp_tmpbuf *buf, int len)
{
  if (buf->len < len)
    {
      free (buf->ptr);
      buf->ptr = malloc (buf->len = len);
    }

  return buf->ptr;
}

/*
 * a somewhat faster data structure might be nice, but
 * with 8 priorities this actually needs <20 insns
 * per shift, the most expensive operation.
 */
typedef struct
{
  ETP_REQ *qs[ETP_NUM_PRI], *qe[ETP_NUM_PRI]; /* qstart, qend */
  int size;
} etp_reqq;

struct etp_pool
{
   void *userdata;

   etp_reqq req_queue;
   etp_reqq res_queue;

   unsigned int started, idle, wanted;

   unsigned int max_poll_time;     /* pool->reslock */
   unsigned int max_poll_reqs;     /* pool->reslock */

   unsigned int nreqs;    /* pool->reqlock */
   unsigned int nready;   /* pool->reqlock */
   unsigned int npending; /* pool->reqlock */
   unsigned int max_idle;      /* maximum number of threads that can pool->idle indefinitely */
   unsigned int idle_timeout; /* number of seconds after which an pool->idle threads exit */

   void (*want_poll_cb) (void *userdata);
   void (*done_poll_cb) (void *userdata);
 
   xmutex_t wrklock;
   xmutex_t reslock;
   xmutex_t reqlock;
   xcond_t  reqwait;
};

typedef struct etp_pool *etp_pool;

typedef struct etp_worker
{
  etp_pool pool;

  struct etp_tmpbuf tmpbuf;

  /* locked by pool->wrklock */
  struct etp_worker *prev, *next;

  xthread_t tid;

#ifdef ETP_WORKER_COMMON
  ETP_WORKER_COMMON
#endif
} etp_worker;

static etp_worker wrk_first; /* NOT etp */

#define ETP_WORKER_LOCK(wrk)   X_LOCK   (pool->wrklock)
#define ETP_WORKER_UNLOCK(wrk) X_UNLOCK (pool->wrklock)

/* worker threads management */

static void
etp_worker_clear (etp_worker *wrk)
{
}

static void ecb_cold
etp_worker_free (etp_worker *wrk)
{
  free (wrk->tmpbuf.ptr);

  wrk->next->prev = wrk->prev;
  wrk->prev->next = wrk->next;

  free (wrk);
}

ETP_API_DECL unsigned int
etp_nreqs (etp_pool pool)
{
  int retval;
  if (WORDACCESS_UNSAFE) X_LOCK   (pool->reqlock);
  retval = pool->nreqs;
  if (WORDACCESS_UNSAFE) X_UNLOCK (pool->reqlock);
  return retval;
}

ETP_API_DECL unsigned int
etp_nready (etp_pool pool)
{
  unsigned int retval;

  if (WORDACCESS_UNSAFE) X_LOCK   (pool->reqlock);
  retval = pool->nready;
  if (WORDACCESS_UNSAFE) X_UNLOCK (pool->reqlock);

  return retval;
}

ETP_API_DECL unsigned int
etp_npending (etp_pool pool)
{
  unsigned int retval;

  if (WORDACCESS_UNSAFE) X_LOCK   (pool->reqlock);
  retval = pool->npending;
  if (WORDACCESS_UNSAFE) X_UNLOCK (pool->reqlock);

  return retval;
}

ETP_API_DECL unsigned int
etp_nthreads (etp_pool pool)
{
  unsigned int retval;

  if (WORDACCESS_UNSAFE) X_LOCK   (pool->reqlock);
  retval = pool->started;
  if (WORDACCESS_UNSAFE) X_UNLOCK (pool->reqlock);

  return retval;
}

static void ecb_noinline ecb_cold
reqq_init (etp_reqq *q)
{
  int pri;

  for (pri = 0; pri < ETP_NUM_PRI; ++pri)
    q->qs[pri] = q->qe[pri] = 0;

  q->size = 0;
}

static int ecb_noinline
reqq_push (etp_reqq *q, ETP_REQ *req)
{
  int pri = req->pri;
  req->next = 0;

  if (q->qe[pri])
    {
      q->qe[pri]->next = req;
      q->qe[pri] = req;
    }
  else
    q->qe[pri] = q->qs[pri] = req;

  return q->size++;
}

static ETP_REQ * ecb_noinline
reqq_shift (etp_reqq *q)
{
  int pri;

  if (!q->size)
    return 0;

  --q->size;

  for (pri = ETP_NUM_PRI; pri--; )
    {
      ETP_REQ *req = q->qs[pri];

      if (req)
        {
          if (!(q->qs[pri] = (ETP_REQ *)req->next))
            q->qe[pri] = 0;

          return req;
        }
    }

  abort ();
}

ETP_API_DECL int ecb_cold
etp_init (etp_pool pool, void (*want_poll)(void *userdata), void (*done_poll)(void *userdata))
{
  X_MUTEX_CREATE (pool->wrklock);
  X_MUTEX_CREATE (pool->reslock);
  X_MUTEX_CREATE (pool->reqlock);
  X_COND_CREATE  (pool->reqwait);

  reqq_init (&pool->req_queue);
  reqq_init (&pool->res_queue);

  wrk_first.next =
  wrk_first.prev = &wrk_first;

  pool->started  = 0;
  pool->idle     = 0;
  pool->nreqs    = 0;
  pool->nready   = 0;
  pool->npending = 0;
  pool->wanted   = 4;

  pool->max_idle = 4;      /* maximum number of threads that can pool->idle indefinitely */
  pool->idle_timeout = 10; /* number of seconds after which an pool->idle threads exit */

  pool->want_poll_cb = want_poll;
  pool->done_poll_cb = done_poll;

  return 0;
}

static void ecb_noinline ecb_cold
etp_proc_init (void)
{
#if HAVE_PRCTL_SET_NAME
  /* provide a more sensible "thread name" */
  char name[16 + 1];
  const int namelen = sizeof (name) - 1;
  int len;

  prctl (PR_GET_NAME, (unsigned long)name, 0, 0, 0);
  name [namelen] = 0;
  len = strlen (name);
  strcpy (name + (len <= namelen - 4 ? len : namelen - 4), "/eio");
  prctl (PR_SET_NAME, (unsigned long)name, 0, 0, 0);
#endif
}

X_THREAD_PROC (etp_proc)
{
  ETP_REQ *req;
  struct timespec ts;
  etp_worker *self = (etp_worker *)thr_arg;
  etp_pool pool = self->pool;

  etp_proc_init ();

  /* try to distribute timeouts somewhat evenly */
  ts.tv_nsec = ((unsigned long)self & 1023UL) * (1000000000UL / 1024UL);

  for (;;)
    {
      ts.tv_sec = 0;

      X_LOCK (pool->reqlock);

      for (;;)
        {
          req = reqq_shift (&pool->req_queue);

          if (req)
            break;

          if (ts.tv_sec == 1) /* no request, but timeout detected, let's quit */
            {
              X_UNLOCK (pool->reqlock);
              X_LOCK (pool->wrklock);
              --pool->started;
              X_UNLOCK (pool->wrklock);
              goto quit;
            }

          ++pool->idle;

          if (pool->idle <= pool->max_idle)
            /* we are allowed to pool->idle, so do so without any timeout */
            X_COND_WAIT (pool->reqwait, pool->reqlock);
          else
            {
              /* initialise timeout once */
              if (!ts.tv_sec)
                ts.tv_sec = time (0) + pool->idle_timeout;

              if (X_COND_TIMEDWAIT (pool->reqwait, pool->reqlock, ts) == ETIMEDOUT)
                ts.tv_sec = 1; /* assuming this is not a value computed above.,.. */
            }

          --pool->idle;
        }

      --pool->nready;

      X_UNLOCK (pool->reqlock);
     
      if (req->type == ETP_TYPE_QUIT)
        goto quit;

      ETP_EXECUTE (self, req);

      X_LOCK (pool->reslock);

      ++pool->npending;

      if (!reqq_push (&pool->res_queue, req))
        ETP_WANT_POLL (poll);

      etp_worker_clear (self);

      X_UNLOCK (pool->reslock);
    }

quit:
  free (req);

  X_LOCK (pool->wrklock);
  etp_worker_free (self);
  X_UNLOCK (pool->wrklock);

  return 0;
}

static void ecb_cold
etp_start_thread (etp_pool pool)
{
  etp_worker *wrk = calloc (1, sizeof (etp_worker));

  /*TODO*/
  assert (("unable to allocate worker thread data", wrk));

  wrk->pool = pool;

  X_LOCK (pool->wrklock);

  if (xthread_create (&wrk->tid, etp_proc, (void *)wrk))
    {
      wrk->prev = &wrk_first;
      wrk->next = wrk_first.next;
      wrk_first.next->prev = wrk;
      wrk_first.next = wrk;
      ++pool->started;
    }
  else
    free (wrk);

  X_UNLOCK (pool->wrklock);
}

static void
etp_maybe_start_thread (etp_pool pool)
{
  if (ecb_expect_true (etp_nthreads (pool) >= pool->wanted))
    return;
  
  /* todo: maybe use pool->idle here, but might be less exact */
  if (ecb_expect_true (0 <= (int)etp_nthreads (pool) + (int)etp_npending (pool) - (int)etp_nreqs (pool)))
    return;

  etp_start_thread (pool);
}

static void ecb_cold
etp_end_thread (etp_pool pool)
{
  ETP_REQ *req = calloc (1, sizeof (ETP_REQ)); /* will be freed by worker */

  req->type = ETP_TYPE_QUIT;
  req->pri  = ETP_PRI_MAX - ETP_PRI_MIN;

  X_LOCK (pool->reqlock);
  reqq_push (&pool->req_queue, req);
  X_COND_SIGNAL (pool->reqwait);
  X_UNLOCK (pool->reqlock);

  X_LOCK (pool->wrklock);
  --pool->started;
  X_UNLOCK (pool->wrklock);
}

ETP_API_DECL int
etp_poll (etp_pool pool)
{
  unsigned int maxreqs;
  unsigned int maxtime;
  struct timeval tv_start, tv_now;

  X_LOCK (pool->reslock);
  maxreqs = pool->max_poll_reqs;
  maxtime = pool->max_poll_time;
  X_UNLOCK (pool->reslock);

  if (maxtime)
    gettimeofday (&tv_start, 0);

  for (;;)
    {
      ETP_REQ *req;

      etp_maybe_start_thread (pool);

      X_LOCK (pool->reslock);
      req = reqq_shift (&pool->res_queue);

      if (req)
        {
          --pool->npending;

          if (!pool->res_queue.size)
            ETP_DONE_POLL (pool->userdata);
        }

      X_UNLOCK (pool->reslock);

      if (!req)
        return 0;

      X_LOCK (pool->reqlock);
      --pool->nreqs;
      X_UNLOCK (pool->reqlock);

      if (ecb_expect_false (req->type == ETP_TYPE_GROUP && req->size))
        {
          req->flags |= ETP_FLAG_DELAYED; /* mark request as delayed */
          continue;
        }
      else
        {
          int res = ETP_FINISH (req);
          if (ecb_expect_false (res))
            return res;
        }

      if (ecb_expect_false (maxreqs && !--maxreqs))
        break;

      if (maxtime)
        {
          gettimeofday (&tv_now, 0);

          if (etp_tvdiff (&tv_start, &tv_now) >= maxtime)
            break;
        }
    }

  errno = EAGAIN;
  return -1;
}

ETP_API_DECL void
etp_grp_cancel (etp_pool pool, ETP_REQ *grp);

ETP_API_DECL void
etp_cancel (etp_pool pool, ETP_REQ *req)
{
  req->cancelled = 1;

  etp_grp_cancel (pool, req);
}

ETP_API_DECL void
etp_grp_cancel (etp_pool pool, ETP_REQ *grp)
{
  for (grp = grp->grp_first; grp; grp = grp->grp_next)
    etp_cancel (pool, grp);
}

ETP_API_DECL void
etp_submit (etp_pool pool, ETP_REQ *req)
{
  req->pri -= ETP_PRI_MIN;

  if (ecb_expect_false (req->pri < ETP_PRI_MIN - ETP_PRI_MIN)) req->pri = ETP_PRI_MIN - ETP_PRI_MIN;
  if (ecb_expect_false (req->pri > ETP_PRI_MAX - ETP_PRI_MIN)) req->pri = ETP_PRI_MAX - ETP_PRI_MIN;

  if (ecb_expect_false (req->type == ETP_TYPE_GROUP))
    {
      /* I hope this is worth it :/ */
      X_LOCK (pool->reqlock);
      ++pool->nreqs;
      X_UNLOCK (pool->reqlock);

      X_LOCK (pool->reslock);

      ++pool->npending;

      if (!reqq_push (&pool->res_queue, req))
        ETP_WANT_POLL (pool);

      X_UNLOCK (pool->reslock);
    }
  else
    {
      X_LOCK (pool->reqlock);
      ++pool->nreqs;
      ++pool->nready;
      reqq_push (&pool->req_queue, req);
      X_COND_SIGNAL (pool->reqwait);
      X_UNLOCK (pool->reqlock);

      etp_maybe_start_thread (pool);
    }
}

ETP_API_DECL void ecb_cold
etp_set_max_poll_time (etp_pool pool, double nseconds)
{
  if (WORDACCESS_UNSAFE) X_LOCK   (pool->reslock);
  pool->max_poll_time = nseconds * ETP_TICKS;
  if (WORDACCESS_UNSAFE) X_UNLOCK (pool->reslock);
}

ETP_API_DECL void ecb_cold
etp_set_max_poll_reqs (etp_pool pool, unsigned int maxreqs)
{
  if (WORDACCESS_UNSAFE) X_LOCK   (pool->reslock);
  pool->max_poll_reqs = maxreqs;
  if (WORDACCESS_UNSAFE) X_UNLOCK (pool->reslock);
}

ETP_API_DECL void ecb_cold
etp_set_max_idle (etp_pool pool, unsigned int nthreads)
{
  if (WORDACCESS_UNSAFE) X_LOCK   (pool->reqlock);
  pool->max_idle = nthreads;
  if (WORDACCESS_UNSAFE) X_UNLOCK (pool->reqlock);
}

ETP_API_DECL void ecb_cold
etp_set_idle_timeout (etp_pool pool, unsigned int seconds)
{
  if (WORDACCESS_UNSAFE) X_LOCK   (pool->reqlock);
  pool->idle_timeout = seconds;
  if (WORDACCESS_UNSAFE) X_UNLOCK (pool->reqlock);
}

ETP_API_DECL void ecb_cold
etp_set_min_parallel (etp_pool pool, unsigned int nthreads)
{
  if (pool->wanted < nthreads)
    pool->wanted = nthreads;
}

ETP_API_DECL void ecb_cold
etp_set_max_parallel (etp_pool pool, unsigned int nthreads)
{
  if (pool->wanted > nthreads)
    pool->wanted = nthreads;

  while (pool->started > pool->wanted)
    etp_end_thread (pool);
}

Revision:	1.6
Committed:	Thu Jun 25 18:08:47 2015 UTC (8 years, 11 months ago) by root
Content type:	text/plain
Branch:	MAIN
Changes since 1.5:	+133 -120 lines
Log Message:	* empty log message *
#	Content
1	/*
2	* libetp implementation
3	*
4	* Copyright (c) 2007,2008,2009,2010,2011,2012,2013 Marc Alexander Lehmann <libetp@schmorp.de>
5	* All rights reserved.
6	*
7	* Redistribution and use in source and binary forms, with or without modifica-
8	* tion, are permitted provided that the following conditions are met:
9	*
10	* 1. Redistributions of source code must retain the above copyright notice,
11	* this list of conditions and the following disclaimer.
12	*
13	* 2. Redistributions in binary form must reproduce the above copyright
14	* notice, this list of conditions and the following disclaimer in the
15	* documentation and/or other materials provided with the distribution.
16	*
17	* THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED
18	* WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MER-
19	* CHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO
20	* EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPE-
21	* CIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
22	* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
23	* OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
24	* WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTH-
25	* ERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
26	* OF THE POSSIBILITY OF SUCH DAMAGE.
27	*
28	* Alternatively, the contents of this file may be used under the terms of
29	* the GNU General Public License ("GPL") version 2 or any later version,
30	* in which case the provisions of the GPL are applicable instead of
31	* the above. If you wish to allow the use of your version of this file
32	* only under the terms of the GPL and not to allow others to use your
33	* version of this file under the BSD license, indicate your decision
34	* by deleting the provisions above and replace them with the notice
35	* and other provisions required by the GPL. If you do not delete the
36	* provisions above, a recipient may use your version of this file under
37	* either the BSD or the GPL.
38	*/
39
40	#ifndef ETP_API_DECL
41	# define ETP_API_DECL static
42	#endif
43
44	#ifndef ETP_PRI_MIN
45	# define ETP_PRI_MIN 0
46	# define ETP_PRI_MAX 0
47	#endif
48
49	#ifndef ETP_TYPE_QUIT
50	# define ETP_TYPE_QUIT 0
51	#endif
52
53	#ifndef ETP_TYPE_GROUP
54	# define ETP_TYPE_GROUP 1
55	#endif
56
57	#ifndef ETP_WANT_POLL
58	# define ETP_WANT_POLL(pool) pool->want_poll_cb (pool->userdata)
59	#endif
60	#ifndef ETP_DONE_POLL
61	# define ETP_DONE_POLL(pool) pool->done_poll_cb (pool->userdata)
62	#endif
63
64	#define ETP_NUM_PRI (ETP_PRI_MAX - ETP_PRI_MIN + 1)
65
66	#define ETP_TICKS ((1000000 + 1023) >> 10)
67
68	enum {
69	ETP_FLAG_GROUPADD = 0x04, /* some request was added to the group */
70	ETP_FLAG_DELAYED = 0x08, /* groiup request has been delayed */
71	};
72
73	/* calculate time difference in ~1/ETP_TICKS of a second */
74	ecb_inline int
75	etp_tvdiff (struct timeval tv1, struct timeval tv2)
76	{
77	return (tv2->tv_sec - tv1->tv_sec ) * ETP_TICKS
78	+ ((tv2->tv_usec - tv1->tv_usec) >> 10);
79	}
80
81	struct etp_tmpbuf
82	{
83	void *ptr;
84	int len;
85	};
86
87	static void *
88	etp_tmpbuf_get (struct etp_tmpbuf *buf, int len)
89	{
90	if (buf->len < len)
91	{
92	free (buf->ptr);
93	buf->ptr = malloc (buf->len = len);
94	}
95
96	return buf->ptr;
97	}
98
99	/*
100	* a somewhat faster data structure might be nice, but
101	* with 8 priorities this actually needs <20 insns
102	* per shift, the most expensive operation.
103	*/
104	typedef struct
105	{
106	ETP_REQ qs[ETP_NUM_PRI], qe[ETP_NUM_PRI]; /* qstart, qend */
107	int size;
108	} etp_reqq;
109
110	struct etp_pool
111	{
112	void *userdata;
113
114	etp_reqq req_queue;
115	etp_reqq res_queue;
116
117	unsigned int started, idle, wanted;
118
119	unsigned int max_poll_time; /* pool->reslock */
120	unsigned int max_poll_reqs; /* pool->reslock */
121
122	unsigned int nreqs; /* pool->reqlock */
123	unsigned int nready; /* pool->reqlock */
124	unsigned int npending; /* pool->reqlock */
125	unsigned int max_idle; /* maximum number of threads that can pool->idle indefinitely */
126	unsigned int idle_timeout; /* number of seconds after which an pool->idle threads exit */
127
128	void (want_poll_cb) (void userdata);
129	void (done_poll_cb) (void userdata);
130
131	xmutex_t wrklock;
132	xmutex_t reslock;
133	xmutex_t reqlock;
134	xcond_t reqwait;
135	};
136
137	typedef struct etp_pool *etp_pool;
138
139	typedef struct etp_worker
140	{
141	etp_pool pool;
142
143	struct etp_tmpbuf tmpbuf;
144
145	/* locked by pool->wrklock */
146	struct etp_worker prev, next;
147
148	xthread_t tid;
149
150	#ifdef ETP_WORKER_COMMON
151	ETP_WORKER_COMMON
152	#endif
153	} etp_worker;
154
155	static etp_worker wrk_first; /* NOT etp */
156
157	#define ETP_WORKER_LOCK(wrk) X_LOCK (pool->wrklock)
158	#define ETP_WORKER_UNLOCK(wrk) X_UNLOCK (pool->wrklock)
159
160	/* worker threads management */
161
162	static void
163	etp_worker_clear (etp_worker *wrk)
164	{
165	}
166
167	static void ecb_cold
168	etp_worker_free (etp_worker *wrk)
169	{
170	free (wrk->tmpbuf.ptr);
171
172	wrk->next->prev = wrk->prev;
173	wrk->prev->next = wrk->next;
174
175	free (wrk);
176	}
177
178	ETP_API_DECL unsigned int
179	etp_nreqs (etp_pool pool)
180	{
181	int retval;
182	if (WORDACCESS_UNSAFE) X_LOCK (pool->reqlock);
183	retval = pool->nreqs;
184	if (WORDACCESS_UNSAFE) X_UNLOCK (pool->reqlock);
185	return retval;
186	}
187
188	ETP_API_DECL unsigned int
189	etp_nready (etp_pool pool)
190	{
191	unsigned int retval;
192
193	if (WORDACCESS_UNSAFE) X_LOCK (pool->reqlock);
194	retval = pool->nready;
195	if (WORDACCESS_UNSAFE) X_UNLOCK (pool->reqlock);
196
197	return retval;
198	}
199
200	ETP_API_DECL unsigned int
201	etp_npending (etp_pool pool)
202	{
203	unsigned int retval;
204
205	if (WORDACCESS_UNSAFE) X_LOCK (pool->reqlock);
206	retval = pool->npending;
207	if (WORDACCESS_UNSAFE) X_UNLOCK (pool->reqlock);
208
209	return retval;
210	}
211
212	ETP_API_DECL unsigned int
213	etp_nthreads (etp_pool pool)
214	{
215	unsigned int retval;
216
217	if (WORDACCESS_UNSAFE) X_LOCK (pool->reqlock);
218	retval = pool->started;
219	if (WORDACCESS_UNSAFE) X_UNLOCK (pool->reqlock);
220
221	return retval;
222	}
223
224	static void ecb_noinline ecb_cold
225	reqq_init (etp_reqq *q)
226	{
227	int pri;
228
229	for (pri = 0; pri < ETP_NUM_PRI; ++pri)
230	q->qs[pri] = q->qe[pri] = 0;
231
232	q->size = 0;
233	}
234
235	static int ecb_noinline
236	reqq_push (etp_reqq q, ETP_REQ req)
237	{
238	int pri = req->pri;
239	req->next = 0;
240
241	if (q->qe[pri])
242	{
243	q->qe[pri]->next = req;
244	q->qe[pri] = req;
245	}
246	else
247	q->qe[pri] = q->qs[pri] = req;
248
249	return q->size++;
250	}
251
252	static ETP_REQ * ecb_noinline
253	reqq_shift (etp_reqq *q)
254	{
255	int pri;
256
257	if (!q->size)
258	return 0;
259
260	--q->size;
261
262	for (pri = ETP_NUM_PRI; pri--; )
263	{
264	ETP_REQ *req = q->qs[pri];
265
266	if (req)
267	{
268	if (!(q->qs[pri] = (ETP_REQ *)req->next))
269	q->qe[pri] = 0;
270
271	return req;
272	}
273	}
274
275	abort ();
276	}
277
278	ETP_API_DECL int ecb_cold
279	etp_init (etp_pool pool, void (want_poll)(void userdata), void (done_poll)(void userdata))
280	{
281	X_MUTEX_CREATE (pool->wrklock);
282	X_MUTEX_CREATE (pool->reslock);
283	X_MUTEX_CREATE (pool->reqlock);
284	X_COND_CREATE (pool->reqwait);
285
286	reqq_init (&pool->req_queue);
287	reqq_init (&pool->res_queue);
288
289	wrk_first.next =
290	wrk_first.prev = &wrk_first;
291
292	pool->started = 0;
293	pool->idle = 0;
294	pool->nreqs = 0;
295	pool->nready = 0;
296	pool->npending = 0;
297	pool->wanted = 4;
298
299	pool->max_idle = 4; /* maximum number of threads that can pool->idle indefinitely */
300	pool->idle_timeout = 10; /* number of seconds after which an pool->idle threads exit */
301
302	pool->want_poll_cb = want_poll;
303	pool->done_poll_cb = done_poll;
304
305	return 0;
306	}
307
308	static void ecb_noinline ecb_cold
309	etp_proc_init (void)
310	{
311	#if HAVE_PRCTL_SET_NAME
312	/* provide a more sensible "thread name" */
313	char name[16 + 1];
314	const int namelen = sizeof (name) - 1;
315	int len;
316
317	prctl (PR_GET_NAME, (unsigned long)name, 0, 0, 0);
318	name [namelen] = 0;
319	len = strlen (name);
320	strcpy (name + (len <= namelen - 4 ? len : namelen - 4), "/eio");
321	prctl (PR_SET_NAME, (unsigned long)name, 0, 0, 0);
322	#endif
323	}
324
325	X_THREAD_PROC (etp_proc)
326	{
327	ETP_REQ *req;
328	struct timespec ts;
329	etp_worker self = (etp_worker )thr_arg;
330	etp_pool pool = self->pool;
331
332	etp_proc_init ();
333
334	/* try to distribute timeouts somewhat evenly */
335	ts.tv_nsec = ((unsigned long)self & 1023UL) * (1000000000UL / 1024UL);
336
337	for (;;)
338	{
339	ts.tv_sec = 0;
340
341	X_LOCK (pool->reqlock);
342
343	for (;;)
344	{
345	req = reqq_shift (&pool->req_queue);
346
347	if (req)
348	break;
349
350	if (ts.tv_sec == 1) /* no request, but timeout detected, let's quit */
351	{
352	X_UNLOCK (pool->reqlock);
353	X_LOCK (pool->wrklock);
354	--pool->started;
355	X_UNLOCK (pool->wrklock);
356	goto quit;
357	}
358
359	++pool->idle;
360
361	if (pool->idle <= pool->max_idle)
362	/* we are allowed to pool->idle, so do so without any timeout */
363	X_COND_WAIT (pool->reqwait, pool->reqlock);
364	else
365	{
366	/* initialise timeout once */
367	if (!ts.tv_sec)
368	ts.tv_sec = time (0) + pool->idle_timeout;
369
370	if (X_COND_TIMEDWAIT (pool->reqwait, pool->reqlock, ts) == ETIMEDOUT)
371	ts.tv_sec = 1; /* assuming this is not a value computed above.,.. */
372	}
373
374	--pool->idle;
375	}
376
377	--pool->nready;
378
379	X_UNLOCK (pool->reqlock);
380
381	if (req->type == ETP_TYPE_QUIT)
382	goto quit;
383
384	ETP_EXECUTE (self, req);
385
386	X_LOCK (pool->reslock);
387
388	++pool->npending;
389
390	if (!reqq_push (&pool->res_queue, req))
391	ETP_WANT_POLL (poll);
392
393	etp_worker_clear (self);
394
395	X_UNLOCK (pool->reslock);
396	}
397
398	quit:
399	free (req);
400
401	X_LOCK (pool->wrklock);
402	etp_worker_free (self);
403	X_UNLOCK (pool->wrklock);
404
405	return 0;
406	}
407
408	static void ecb_cold
409	etp_start_thread (etp_pool pool)
410	{
411	etp_worker *wrk = calloc (1, sizeof (etp_worker));
412
413	/TODO/
414	assert (("unable to allocate worker thread data", wrk));
415
416	wrk->pool = pool;
417
418	X_LOCK (pool->wrklock);
419
420	if (xthread_create (&wrk->tid, etp_proc, (void *)wrk))
421	{
422	wrk->prev = &wrk_first;
423	wrk->next = wrk_first.next;
424	wrk_first.next->prev = wrk;
425	wrk_first.next = wrk;
426	++pool->started;
427	}
428	else
429	free (wrk);
430
431	X_UNLOCK (pool->wrklock);
432	}
433
434	static void
435	etp_maybe_start_thread (etp_pool pool)
436	{
437	if (ecb_expect_true (etp_nthreads (pool) >= pool->wanted))
438	return;
439
440	/* todo: maybe use pool->idle here, but might be less exact */
441	if (ecb_expect_true (0 <= (int)etp_nthreads (pool) + (int)etp_npending (pool) - (int)etp_nreqs (pool)))
442	return;
443
444	etp_start_thread (pool);
445	}
446
447	static void ecb_cold
448	etp_end_thread (etp_pool pool)
449	{
450	ETP_REQ req = calloc (1, sizeof (ETP_REQ)); / will be freed by worker */
451
452	req->type = ETP_TYPE_QUIT;
453	req->pri = ETP_PRI_MAX - ETP_PRI_MIN;
454
455	X_LOCK (pool->reqlock);
456	reqq_push (&pool->req_queue, req);
457	X_COND_SIGNAL (pool->reqwait);
458	X_UNLOCK (pool->reqlock);
459
460	X_LOCK (pool->wrklock);
461	--pool->started;
462	X_UNLOCK (pool->wrklock);
463	}
464
465	ETP_API_DECL int
466	etp_poll (etp_pool pool)
467	{
468	unsigned int maxreqs;
469	unsigned int maxtime;
470	struct timeval tv_start, tv_now;
471
472	X_LOCK (pool->reslock);
473	maxreqs = pool->max_poll_reqs;
474	maxtime = pool->max_poll_time;
475	X_UNLOCK (pool->reslock);
476
477	if (maxtime)
478	gettimeofday (&tv_start, 0);
479
480	for (;;)
481	{
482	ETP_REQ *req;
483
484	etp_maybe_start_thread (pool);
485
486	X_LOCK (pool->reslock);
487	req = reqq_shift (&pool->res_queue);
488
489	if (req)
490	{
491	--pool->npending;
492
493	if (!pool->res_queue.size)
494	ETP_DONE_POLL (pool->userdata);
495	}
496
497	X_UNLOCK (pool->reslock);
498
499	if (!req)
500	return 0;
501
502	X_LOCK (pool->reqlock);
503	--pool->nreqs;
504	X_UNLOCK (pool->reqlock);
505
506	if (ecb_expect_false (req->type == ETP_TYPE_GROUP && req->size))
507	{
508	req->flags \|= ETP_FLAG_DELAYED; /* mark request as delayed */
509	continue;
510	}
511	else
512	{
513	int res = ETP_FINISH (req);
514	if (ecb_expect_false (res))
515	return res;
516	}
517
518	if (ecb_expect_false (maxreqs && !--maxreqs))
519	break;
520
521	if (maxtime)
522	{
523	gettimeofday (&tv_now, 0);
524
525	if (etp_tvdiff (&tv_start, &tv_now) >= maxtime)
526	break;
527	}
528	}
529
530	errno = EAGAIN;
531	return -1;
532	}
533
534	ETP_API_DECL void
535	etp_grp_cancel (etp_pool pool, ETP_REQ *grp);
536
537	ETP_API_DECL void
538	etp_cancel (etp_pool pool, ETP_REQ *req)
539	{
540	req->cancelled = 1;
541
542	etp_grp_cancel (pool, req);
543	}
544
545	ETP_API_DECL void
546	etp_grp_cancel (etp_pool pool, ETP_REQ *grp)
547	{
548	for (grp = grp->grp_first; grp; grp = grp->grp_next)
549	etp_cancel (pool, grp);
550	}
551
552	ETP_API_DECL void
553	etp_submit (etp_pool pool, ETP_REQ *req)
554	{
555	req->pri -= ETP_PRI_MIN;
556
557	if (ecb_expect_false (req->pri < ETP_PRI_MIN - ETP_PRI_MIN)) req->pri = ETP_PRI_MIN - ETP_PRI_MIN;
558	if (ecb_expect_false (req->pri > ETP_PRI_MAX - ETP_PRI_MIN)) req->pri = ETP_PRI_MAX - ETP_PRI_MIN;
559
560	if (ecb_expect_false (req->type == ETP_TYPE_GROUP))
561	{
562	/* I hope this is worth it :/ */
563	X_LOCK (pool->reqlock);
564	++pool->nreqs;
565	X_UNLOCK (pool->reqlock);
566
567	X_LOCK (pool->reslock);
568
569	++pool->npending;
570
571	if (!reqq_push (&pool->res_queue, req))
572	ETP_WANT_POLL (pool);
573
574	X_UNLOCK (pool->reslock);
575	}
576	else
577	{
578	X_LOCK (pool->reqlock);
579	++pool->nreqs;
580	++pool->nready;
581	reqq_push (&pool->req_queue, req);
582	X_COND_SIGNAL (pool->reqwait);
583	X_UNLOCK (pool->reqlock);
584
585	etp_maybe_start_thread (pool);
586	}
587	}
588
589	ETP_API_DECL void ecb_cold
590	etp_set_max_poll_time (etp_pool pool, double nseconds)
591	{
592	if (WORDACCESS_UNSAFE) X_LOCK (pool->reslock);
593	pool->max_poll_time = nseconds * ETP_TICKS;
594	if (WORDACCESS_UNSAFE) X_UNLOCK (pool->reslock);
595	}
596
597	ETP_API_DECL void ecb_cold
598	etp_set_max_poll_reqs (etp_pool pool, unsigned int maxreqs)
599	{
600	if (WORDACCESS_UNSAFE) X_LOCK (pool->reslock);
601	pool->max_poll_reqs = maxreqs;
602	if (WORDACCESS_UNSAFE) X_UNLOCK (pool->reslock);
603	}
604
605	ETP_API_DECL void ecb_cold
606	etp_set_max_idle (etp_pool pool, unsigned int nthreads)
607	{
608	if (WORDACCESS_UNSAFE) X_LOCK (pool->reqlock);
609	pool->max_idle = nthreads;
610	if (WORDACCESS_UNSAFE) X_UNLOCK (pool->reqlock);
611	}
612
613	ETP_API_DECL void ecb_cold
614	etp_set_idle_timeout (etp_pool pool, unsigned int seconds)
615	{
616	if (WORDACCESS_UNSAFE) X_LOCK (pool->reqlock);
617	pool->idle_timeout = seconds;
618	if (WORDACCESS_UNSAFE) X_UNLOCK (pool->reqlock);
619	}
620
621	ETP_API_DECL void ecb_cold
622	etp_set_min_parallel (etp_pool pool, unsigned int nthreads)
623	{
624	if (pool->wanted < nthreads)
625	pool->wanted = nthreads;
626	}
627
628	ETP_API_DECL void ecb_cold
629	etp_set_max_parallel (etp_pool pool, unsigned int nthreads)
630	{
631	if (pool->wanted > nthreads)
632	pool->wanted = nthreads;
633
634	while (pool->started > pool->wanted)
635	etp_end_thread (pool);
636	}
637