lib/workqueue.c

   1 /*
   2  * Quagga Work Queue Support.
   3  *
   4  * Copyright (C) 2005 Sun Microsystems, Inc.
   5  *
   6  * This file is part of GNU Zebra.
   7  *
   8  * Quagga is free software; you can redistribute it and/or modify it
   9  * under the terms of the GNU General Public License as published by the
  10  * Free Software Foundation; either version 2, or (at your option) any
  11  * later version.
  12  *
  13  * Quagga is distributed in the hope that it will be useful, but
  14  * WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * along with Quagga; see the file COPYING.  If not, write to the Free
  20  * Software Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA
  21  * 02111-1307, USA.
  22  */
  23
  24 #include <zebra.h>
  25 #include "thread.h"
  26 #include "memory.h"
  27 #include "workqueue.h"
  28 #include "linklist.h"
  29 #include "command.h"
  30 #include "log.h"
  31
  32 DEFINE_MTYPE(LIB, WORK_QUEUE,             "Work queue")
  33 DEFINE_MTYPE_STATIC(LIB, WORK_QUEUE_ITEM, "Work queue item")
  34 DEFINE_MTYPE_STATIC(LIB, WORK_QUEUE_NAME, "Work queue name string")
  35
  36 /* master list of work_queues */
  37 static struct list _work_queues;
  38 /* pointer primarily to avoid an otherwise harmless warning on
  39  * ALL_LIST_ELEMENTS_RO
  40  */
  41 static struct list *work_queues = &_work_queues;
  42
  43 #define WORK_QUEUE_MIN_GRANULARITY 1
  44
  45 static struct work_queue_item *
  46 work_queue_item_new (struct work_queue *wq)
  47 {
  48   struct work_queue_item *item;
  49   assert (wq);
  50
  51   item = XCALLOC (MTYPE_WORK_QUEUE_ITEM,
  52                   sizeof (struct work_queue_item));
  53
  54   return item;
  55 }
  56
  57 static void
  58 work_queue_item_free (struct work_queue_item *item)
  59 {
  60   XFREE (MTYPE_WORK_QUEUE_ITEM, item);
  61   return;
  62 }
  63
  64 /* create new work queue */
  65 struct work_queue *
  66 work_queue_new (struct thread_master *m, const char *queue_name)
  67 {
  68   struct work_queue *new;
  69
  70   new = XCALLOC (MTYPE_WORK_QUEUE, sizeof (struct work_queue));
  71
  72   if (new == NULL)
  73     return new;
  74
  75   new->name = XSTRDUP (MTYPE_WORK_QUEUE_NAME, queue_name);
  76   new->master = m;
  77   SET_FLAG (new->flags, WQ_UNPLUGGED);
  78
  79   if ( (new->items = list_new ()) == NULL)
  80     {
  81       XFREE (MTYPE_WORK_QUEUE_NAME, new->name);
  82       XFREE (MTYPE_WORK_QUEUE, new);
  83
  84       return NULL;
  85     }
  86
  87   new->items->del = (void (*)(void *)) work_queue_item_free;
  88
  89   listnode_add (work_queues, new);
  90
  91   new->cycles.granularity = WORK_QUEUE_MIN_GRANULARITY;
  92
  93   /* Default values, can be overriden by caller */
  94   new->spec.hold = WORK_QUEUE_DEFAULT_HOLD;
  95   new->spec.yield = THREAD_YIELD_TIME_SLOT;
  96
  97   return new;
  98 }
  99
 100 void
 101 work_queue_free (struct work_queue *wq)
 102 {
 103   if (wq->thread != NULL)
 104     thread_cancel(wq->thread);
 105
 106   /* list_delete frees items via callback */
 107   list_delete (wq->items);
 108   listnode_delete (work_queues, wq);
 109
 110   XFREE (MTYPE_WORK_QUEUE_NAME, wq->name);
 111   XFREE (MTYPE_WORK_QUEUE, wq);
 112   return;
 113 }
 114
 115 bool
 116 work_queue_is_scheduled (struct work_queue *wq)
 117 {
 118   return (wq->thread != NULL);
 119 }
 120
 121 static int
 122 work_queue_schedule (struct work_queue *wq, unsigned int delay)
 123 {
 124   /* if appropriate, schedule work queue thread */
 125   if ( CHECK_FLAG (wq->flags, WQ_UNPLUGGED)
 126        && (wq->thread == NULL)
 127        && (listcount (wq->items) > 0) )
 128     {
 129       wq->thread = thread_add_background (wq->master, work_queue_run,
 130                                           wq, delay);
 131       /* set thread yield time, if needed */
 132       if (wq->thread && wq->spec.yield != THREAD_YIELD_TIME_SLOT)
 133         thread_set_yield_time (wq->thread, wq->spec.yield);
 134       return 1;
 135     }
 136   else
 137     return 0;
 138 }
 139
 140 void
 141 work_queue_add (struct work_queue *wq, void *data)
 142 {
 143   struct work_queue_item *item;
 144
 145   assert (wq);
 146
 147   if (!(item = work_queue_item_new (wq)))
 148     {
 149       zlog_err ("%s: unable to get new queue item", __func__);
 150       return;
 151     }
 152
 153   item->data = data;
 154   listnode_add (wq->items, item);
 155
 156   work_queue_schedule (wq, wq->spec.hold);
 157
 158   return;
 159 }
 160
 161 static void
 162 work_queue_item_remove (struct work_queue *wq, struct listnode *ln)
 163 {
 164   struct work_queue_item *item = listgetdata (ln);
 165
 166   assert (item && item->data);
 167
 168   /* call private data deletion callback if needed */
 169   if (wq->spec.del_item_data)
 170     wq->spec.del_item_data (wq, item->data);
 171
 172   list_delete_node (wq->items, ln);
 173   work_queue_item_free (item);
 174
 175   return;
 176 }
 177
 178 static void
 179 work_queue_item_requeue (struct work_queue *wq, struct listnode *ln)
 180 {
 181   LISTNODE_DETACH (wq->items, ln);
 182   LISTNODE_ATTACH (wq->items, ln); /* attach to end of list */
 183 }
 184
 185 DEFUN (show_work_queues,
 186        show_work_queues_cmd,
 187        "show work-queues",
 188        SHOW_STR
 189        "Work Queue information\n")
 190 {
 191   struct listnode *node;
 192   struct work_queue *wq;
 193
 194   vty_out (vty,
 195            "%c %8s %5s %8s %8s %21s%s",
 196            ' ', "List","(ms) ","Q. Runs","Yields","Cycle Counts   ",
 197            VTY_NEWLINE);
 198   vty_out (vty,
 199            "%c %8s %5s %8s %8s %7s %6s %8s %6s %s%s",
 200            'P',
 201            "Items",
 202            "Hold",
 203            "Total","Total",
 204            "Best","Gran.","Total","Avg.",
 205            "Name",
 206            VTY_NEWLINE);
 207
 208   for (ALL_LIST_ELEMENTS_RO (work_queues, node, wq))
 209     {
 210       vty_out (vty,"%c %8d %5d %8ld %8ld %7d %6d %8ld %6u %s%s",
 211                (CHECK_FLAG (wq->flags, WQ_UNPLUGGED) ? ' ' : 'P'),
 212                listcount (wq->items),
 213                wq->spec.hold,
 214                wq->runs, wq->yields,
 215                wq->cycles.best, wq->cycles.granularity, wq->cycles.total,
 216                  (wq->runs) ?
 217                    (unsigned int) (wq->cycles.total / wq->runs) : 0,
 218                wq->name,
 219                VTY_NEWLINE);
 220     }
 221
 222   return CMD_SUCCESS;
 223 }
 224
 225 /* 'plug' a queue: Stop it from being scheduled,
 226  * ie: prevent the queue from draining.
 227  */
 228 void
 229 work_queue_plug (struct work_queue *wq)
 230 {
 231   if (wq->thread)
 232     thread_cancel (wq->thread);
 233
 234   wq->thread = NULL;
 235
 236   UNSET_FLAG (wq->flags, WQ_UNPLUGGED);
 237 }
 238
 239 /* unplug queue, schedule it again, if appropriate
 240  * Ie: Allow the queue to be drained again
 241  */
 242 void
 243 work_queue_unplug (struct work_queue *wq)
 244 {
 245   SET_FLAG (wq->flags, WQ_UNPLUGGED);
 246
 247   /* if thread isnt already waiting, add one */
 248   work_queue_schedule (wq, wq->spec.hold);
 249 }
 250
 251 /* timer thread to process a work queue
 252  * will reschedule itself if required,
 253  * otherwise work_queue_item_add
 254  */
 255 int
 256 work_queue_run (struct thread *thread)
 257 {
 258   struct work_queue *wq;
 259   struct work_queue_item *item;
 260   wq_item_status ret;
 261   unsigned int cycles = 0;
 262   struct listnode *node, *nnode;
 263   char yielded = 0;
 264
 265   wq = THREAD_ARG (thread);
 266   wq->thread = NULL;
 267
 268   assert (wq && wq->items);
 269
 270   /* calculate cycle granularity:
 271    * list iteration == 1 run
 272    * listnode processing == 1 cycle
 273    * granularity == # cycles between checks whether we should yield.
 274    *
 275    * granularity should be > 0, and can increase slowly after each run to
 276    * provide some hysteris, but not past cycles.best or 2*cycles.
 277    *
 278    * Best: starts low, can only increase
 279    *
 280    * Granularity: starts at WORK_QUEUE_MIN_GRANULARITY, can be decreased
 281    *              if we run to end of time slot, can increase otherwise
 282    *              by a small factor.
 283    *
 284    * We could use just the average and save some work, however we want to be
 285    * able to adjust quickly to CPU pressure. Average wont shift much if
 286    * daemon has been running a long time.
 287    */
 288    if (wq->cycles.granularity == 0)
 289      wq->cycles.granularity = WORK_QUEUE_MIN_GRANULARITY;
 290
 291   for (ALL_LIST_ELEMENTS (wq->items, node, nnode, item))
 292   {
 293     assert (item && item->data);
 294
 295     /* dont run items which are past their allowed retries */
 296     if (item->ran > wq->spec.max_retries)
 297       {
 298         /* run error handler, if any */
 299         if (wq->spec.errorfunc)
 300           wq->spec.errorfunc (wq, item->data);
 301         work_queue_item_remove (wq, node);
 302         continue;
 303       }
 304
 305     /* run and take care of items that want to be retried immediately */
 306     do
 307       {
 308         ret = wq->spec.workfunc (wq, item->data);
 309         item->ran++;
 310       }
 311     while ((ret == WQ_RETRY_NOW)
 312            && (item->ran < wq->spec.max_retries));
 313
 314     switch (ret)
 315       {
 316       case WQ_QUEUE_BLOCKED:
 317         {
 318           /* decrement item->ran again, cause this isn't an item
 319            * specific error, and fall through to WQ_RETRY_LATER
 320            */
 321           item->ran--;
 322         }
 323       case WQ_RETRY_LATER:
 324         {
 325           goto stats;
 326         }
 327       case WQ_REQUEUE:
 328         {
 329           item->ran--;
 330           work_queue_item_requeue (wq, node);
 331       /* If a single node is being used with a meta-queue (e.g., zebra),
 332        * update the next node as we don't want to exit the thread and
 333        * reschedule it after every node. By definition, WQ_REQUEUE is
 334        * meant to continue the processing; the yield logic will kick in
 335        * to terminate the thread when time has exceeded.
 336        */
 337       if (nnode == NULL)
 338         nnode = node;
 339           break;
 340         }
 341       case WQ_RETRY_NOW:
 342         /* a RETRY_NOW that gets here has exceeded max_tries, same as ERROR */
 343       case WQ_ERROR:
 344         {
 345           if (wq->spec.errorfunc)
 346             wq->spec.errorfunc (wq, item);
 347         }
 348         /* fall through here is deliberate */
 349       case WQ_SUCCESS:
 350       default:
 351         {
 352           work_queue_item_remove (wq, node);
 353           break;
 354         }
 355       }
 356
 357     /* completed cycle */
 358     cycles++;
 359
 360     /* test if we should yield */
 361     if ( !(cycles % wq->cycles.granularity)
 362         && thread_should_yield (thread))
 363       {
 364         yielded = 1;
 365         goto stats;
 366       }
 367   }
 368
 369 stats:
 370
 371 #define WQ_HYSTERESIS_FACTOR 4
 372
 373   /* we yielded, check whether granularity should be reduced */
 374   if (yielded && (cycles < wq->cycles.granularity))
 375     {
 376       wq->cycles.granularity = ((cycles > 0) ? cycles
 377                                              : WORK_QUEUE_MIN_GRANULARITY);
 378     }
 379   /* otherwise, should granularity increase? */
 380   else if (cycles >= (wq->cycles.granularity))
 381     {
 382       if (cycles > wq->cycles.best)
 383         wq->cycles.best = cycles;
 384
 385       /* along with yielded check, provides hysteresis for granularity */
 386       if (cycles > (wq->cycles.granularity * WQ_HYSTERESIS_FACTOR
 387                                            * WQ_HYSTERESIS_FACTOR))
 388         wq->cycles.granularity *= WQ_HYSTERESIS_FACTOR; /* quick ramp-up */
 389       else if (cycles > (wq->cycles.granularity * WQ_HYSTERESIS_FACTOR))
 390         wq->cycles.granularity += WQ_HYSTERESIS_FACTOR;
 391     }
 392 #undef WQ_HYSTERIS_FACTOR
 393
 394   wq->runs++;
 395   wq->cycles.total += cycles;
 396   if (yielded)
 397     wq->yields++;
 398
 399 #if 0
 400   printf ("%s: cycles %d, new: best %d, worst %d\n",
 401             __func__, cycles, wq->cycles.best, wq->cycles.granularity);
 402 #endif
 403
 404   /* Is the queue done yet? If it is, call the completion callback. */
 405   if (listcount (wq->items) > 0)
 406     work_queue_schedule (wq, 0);
 407   else if (wq->spec.completion_func)
 408     wq->spec.completion_func (wq);
 409
 410   return 0;
 411 }