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 along
  19  * with this program; see the file COPYING; if not, write to the Free Software
  20  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
  21  */
  22
  23 #include <zebra.h>
  24 #include "thread.h"
  25 #include "memory.h"
  26 #include "workqueue.h"
  27 #include "linklist.h"
  28 #include "command.h"
  29 #include "log.h"
  30
  31 DEFINE_MTYPE(LIB, WORK_QUEUE, "Work queue")
  32 DEFINE_MTYPE_STATIC(LIB, WORK_QUEUE_ITEM, "Work queue item")
  33 DEFINE_MTYPE_STATIC(LIB, WORK_QUEUE_NAME, "Work queue name string")
  34
  35 /* master list of work_queues */
  36 static struct list _work_queues;
  37 /* pointer primarily to avoid an otherwise harmless warning on
  38  * ALL_LIST_ELEMENTS_RO
  39  */
  40 static struct list *work_queues = &_work_queues;
  41
  42 #define WORK_QUEUE_MIN_GRANULARITY 1
  43
  44 static struct work_queue_item *work_queue_item_new(struct work_queue *wq)
  45 {
  46         struct work_queue_item *item;
  47         assert(wq);
  48
  49         item = XCALLOC(MTYPE_WORK_QUEUE_ITEM, sizeof(struct work_queue_item));
  50
  51         return item;
  52 }
  53
  54 static void work_queue_item_free(struct work_queue_item *item)
  55 {
  56         XFREE(MTYPE_WORK_QUEUE_ITEM, item);
  57         return;
  58 }
  59
  60 static void work_queue_item_remove(struct work_queue *wq,
  61                                    struct work_queue_item *item)
  62 {
  63         assert(item && item->data);
  64
  65         /* call private data deletion callback if needed */
  66         if (wq->spec.del_item_data)
  67                 wq->spec.del_item_data(wq, item->data);
  68
  69         work_queue_item_dequeue(wq, item);
  70
  71         work_queue_item_free(item);
  72
  73         return;
  74 }
  75
  76 /* create new work queue */
  77 struct work_queue *work_queue_new(struct thread_master *m,
  78                                   const char *queue_name)
  79 {
  80         struct work_queue *new;
  81
  82         new = XCALLOC(MTYPE_WORK_QUEUE, sizeof(struct work_queue));
  83
  84         if (new == NULL)
  85                 return new;
  86
  87         new->name = XSTRDUP(MTYPE_WORK_QUEUE_NAME, queue_name);
  88         new->master = m;
  89         SET_FLAG(new->flags, WQ_UNPLUGGED);
  90
  91         STAILQ_INIT(&new->items);
  92
  93         listnode_add(work_queues, new);
  94
  95         new->cycles.granularity = WORK_QUEUE_MIN_GRANULARITY;
  96
  97         /* Default values, can be overriden by caller */
  98         new->spec.hold = WORK_QUEUE_DEFAULT_HOLD;
  99         new->spec.yield = THREAD_YIELD_TIME_SLOT;
 100
 101         return new;
 102 }
 103
 104 void work_queue_free_original(struct work_queue *wq)
 105 {
 106         if (wq->thread != NULL)
 107                 thread_cancel(wq->thread);
 108
 109         while (!work_queue_empty(wq)) {
 110                 struct work_queue_item *item = work_queue_last_item(wq);
 111
 112                 work_queue_item_remove(wq, item);
 113         }
 114
 115         listnode_delete(work_queues, wq);
 116
 117         XFREE(MTYPE_WORK_QUEUE_NAME, wq->name);
 118         XFREE(MTYPE_WORK_QUEUE, wq);
 119         return;
 120 }
 121
 122 void work_queue_free_and_null(struct work_queue **wq)
 123 {
 124         work_queue_free_original(*wq);
 125         *wq = NULL;
 126 }
 127
 128 bool work_queue_is_scheduled(struct work_queue *wq)
 129 {
 130         return (wq->thread != NULL);
 131 }
 132
 133 static int work_queue_schedule(struct work_queue *wq, unsigned int delay)
 134 {
 135         /* if appropriate, schedule work queue thread */
 136         if (CHECK_FLAG(wq->flags, WQ_UNPLUGGED) && (wq->thread == NULL)
 137             && !work_queue_empty(wq)) {
 138                 wq->thread = NULL;
 139                 thread_add_timer_msec(wq->master, work_queue_run, wq, delay,
 140                                       &wq->thread);
 141                 /* set thread yield time, if needed */
 142                 if (wq->thread && wq->spec.yield != THREAD_YIELD_TIME_SLOT)
 143                         thread_set_yield_time(wq->thread, wq->spec.yield);
 144                 return 1;
 145         } else
 146                 return 0;
 147 }
 148
 149 void work_queue_add(struct work_queue *wq, void *data)
 150 {
 151         struct work_queue_item *item;
 152
 153         assert(wq);
 154
 155         if (!(item = work_queue_item_new(wq))) {
 156                 zlog_err("%s: unable to get new queue item", __func__);
 157                 return;
 158         }
 159
 160         item->data = data;
 161         work_queue_item_enqueue(wq, item);
 162
 163         work_queue_schedule(wq, wq->spec.hold);
 164
 165         return;
 166 }
 167
 168 static void work_queue_item_requeue(struct work_queue *wq,
 169                                     struct work_queue_item *item)
 170 {
 171         work_queue_item_dequeue(wq, item);
 172
 173         /* attach to end of list */
 174         work_queue_item_enqueue(wq, item);
 175 }
 176
 177 DEFUN (show_work_queues,
 178        show_work_queues_cmd,
 179        "show work-queues",
 180        SHOW_STR
 181        "Work Queue information\n")
 182 {
 183         struct listnode *node;
 184         struct work_queue *wq;
 185
 186         vty_out(vty, "%c %8s %5s %8s %8s %21s\n", ' ', "List", "(ms) ",
 187                 "Q. Runs", "Yields", "Cycle Counts   ");
 188         vty_out(vty, "%c %8s %5s %8s %8s %7s %6s %8s %6s %s\n", 'P', "Items",
 189                 "Hold", "Total", "Total", "Best", "Gran.", "Total", "Avg.",
 190                 "Name");
 191
 192         for (ALL_LIST_ELEMENTS_RO(work_queues, node, wq)) {
 193                 vty_out(vty, "%c %8d %5d %8ld %8ld %7d %6d %8ld %6u %s\n",
 194                         (CHECK_FLAG(wq->flags, WQ_UNPLUGGED) ? ' ' : 'P'),
 195                         work_queue_item_count(wq), wq->spec.hold, wq->runs,
 196                         wq->yields, wq->cycles.best, wq->cycles.granularity,
 197                         wq->cycles.total,
 198                         (wq->runs) ? (unsigned int)(wq->cycles.total / wq->runs)
 199                                    : 0,
 200                         wq->name);
 201         }
 202
 203         return CMD_SUCCESS;
 204 }
 205
 206 void workqueue_cmd_init(void)
 207 {
 208         install_element(VIEW_NODE, &show_work_queues_cmd);
 209 }
 210
 211 /* 'plug' a queue: Stop it from being scheduled,
 212  * ie: prevent the queue from draining.
 213  */
 214 void work_queue_plug(struct work_queue *wq)
 215 {
 216         if (wq->thread)
 217                 thread_cancel(wq->thread);
 218
 219         wq->thread = NULL;
 220
 221         UNSET_FLAG(wq->flags, WQ_UNPLUGGED);
 222 }
 223
 224 /* unplug queue, schedule it again, if appropriate
 225  * Ie: Allow the queue to be drained again
 226  */
 227 void work_queue_unplug(struct work_queue *wq)
 228 {
 229         SET_FLAG(wq->flags, WQ_UNPLUGGED);
 230
 231         /* if thread isnt already waiting, add one */
 232         work_queue_schedule(wq, wq->spec.hold);
 233 }
 234
 235 /* timer thread to process a work queue
 236  * will reschedule itself if required,
 237  * otherwise work_queue_item_add
 238  */
 239 int work_queue_run(struct thread *thread)
 240 {
 241         struct work_queue *wq;
 242         struct work_queue_item *item, *titem;
 243         wq_item_status ret;
 244         unsigned int cycles = 0;
 245         char yielded = 0;
 246
 247         wq = THREAD_ARG(thread);
 248         wq->thread = NULL;
 249
 250         assert(wq);
 251
 252         /* calculate cycle granularity:
 253          * list iteration == 1 run
 254          * listnode processing == 1 cycle
 255          * granularity == # cycles between checks whether we should yield.
 256          *
 257          * granularity should be > 0, and can increase slowly after each run to
 258          * provide some hysteris, but not past cycles.best or 2*cycles.
 259          *
 260          * Best: starts low, can only increase
 261          *
 262          * Granularity: starts at WORK_QUEUE_MIN_GRANULARITY, can be decreased
 263          *              if we run to end of time slot, can increase otherwise
 264          *              by a small factor.
 265          *
 266          * We could use just the average and save some work, however we want to
 267          * be
 268          * able to adjust quickly to CPU pressure. Average wont shift much if
 269          * daemon has been running a long time.
 270          */
 271         if (wq->cycles.granularity == 0)
 272                 wq->cycles.granularity = WORK_QUEUE_MIN_GRANULARITY;
 273
 274         STAILQ_FOREACH_SAFE (item, &wq->items, wq, titem) {
 275                 assert(item && item->data);
 276
 277                 /* dont run items which are past their allowed retries */
 278                 if (item->ran > wq->spec.max_retries) {
 279                         /* run error handler, if any */
 280                         if (wq->spec.errorfunc)
 281                                 wq->spec.errorfunc(wq, item->data);
 282                         work_queue_item_remove(wq, item);
 283                         continue;
 284                 }
 285
 286                 /* run and take care of items that want to be retried
 287                  * immediately */
 288                 do {
 289                         ret = wq->spec.workfunc(wq, item->data);
 290                         item->ran++;
 291                 } while ((ret == WQ_RETRY_NOW)
 292                          && (item->ran < wq->spec.max_retries));
 293
 294                 switch (ret) {
 295                 case WQ_QUEUE_BLOCKED: {
 296                         /* decrement item->ran again, cause this isn't an item
 297                          * specific error, and fall through to WQ_RETRY_LATER
 298                          */
 299                         item->ran--;
 300                 }
 301                 case WQ_RETRY_LATER: {
 302                         goto stats;
 303                 }
 304                 case WQ_REQUEUE: {
 305                         item->ran--;
 306                         work_queue_item_requeue(wq, item);
 307                         /* If a single node is being used with a meta-queue
 308                          * (e.g., zebra),
 309                          * update the next node as we don't want to exit the
 310                          * thread and
 311                          * reschedule it after every node. By definition,
 312                          * WQ_REQUEUE is
 313                          * meant to continue the processing; the yield logic
 314                          * will kick in
 315                          * to terminate the thread when time has exceeded.
 316                          */
 317                         if (titem == NULL)
 318                                 titem = item;
 319                         break;
 320                 }
 321                 case WQ_RETRY_NOW:
 322                 /* a RETRY_NOW that gets here has exceeded max_tries, same as
 323                  * ERROR */
 324                 case WQ_ERROR: {
 325                         if (wq->spec.errorfunc)
 326                                 wq->spec.errorfunc(wq, item);
 327                 }
 328                 /* fallthru */
 329                 case WQ_SUCCESS:
 330                 default: {
 331                         work_queue_item_remove(wq, item);
 332                         break;
 333                 }
 334                 }
 335
 336                 /* completed cycle */
 337                 cycles++;
 338
 339                 /* test if we should yield */
 340                 if (!(cycles % wq->cycles.granularity)
 341                     && thread_should_yield(thread)) {
 342                         yielded = 1;
 343                         goto stats;
 344                 }
 345         }
 346
 347 stats:
 348
 349 #define WQ_HYSTERESIS_FACTOR 4
 350
 351         /* we yielded, check whether granularity should be reduced */
 352         if (yielded && (cycles < wq->cycles.granularity)) {
 353                 wq->cycles.granularity =
 354                         ((cycles > 0) ? cycles : WORK_QUEUE_MIN_GRANULARITY);
 355         }
 356         /* otherwise, should granularity increase? */
 357         else if (cycles >= (wq->cycles.granularity)) {
 358                 if (cycles > wq->cycles.best)
 359                         wq->cycles.best = cycles;
 360
 361                 /* along with yielded check, provides hysteresis for granularity
 362                  */
 363                 if (cycles > (wq->cycles.granularity * WQ_HYSTERESIS_FACTOR
 364                               * WQ_HYSTERESIS_FACTOR))
 365                         wq->cycles.granularity *=
 366                                 WQ_HYSTERESIS_FACTOR; /* quick ramp-up */
 367                 else if (cycles
 368                          > (wq->cycles.granularity * WQ_HYSTERESIS_FACTOR))
 369                         wq->cycles.granularity += WQ_HYSTERESIS_FACTOR;
 370         }
 371 #undef WQ_HYSTERIS_FACTOR
 372
 373         wq->runs++;
 374         wq->cycles.total += cycles;
 375         if (yielded)
 376                 wq->yields++;
 377
 378 #if 0
 379   printf ("%s: cycles %d, new: best %d, worst %d\n",
 380             __func__, cycles, wq->cycles.best, wq->cycles.granularity);
 381 #endif
 382
 383         /* Is the queue done yet? If it is, call the completion callback. */
 384         if (!work_queue_empty(wq))
 385                 work_queue_schedule(wq, 0);
 386         else if (wq->spec.completion_func)
 387                 wq->spec.completion_func(wq);
 388
 389         return 0;
 390 }