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         new->name = XSTRDUP(MTYPE_WORK_QUEUE_NAME, queue_name);
  85         new->master = m;
  86         SET_FLAG(new->flags, WQ_UNPLUGGED);
  87
  88         STAILQ_INIT(&new->items);
  89
  90         listnode_add(work_queues, new);
  91
  92         new->cycles.granularity = WORK_QUEUE_MIN_GRANULARITY;
  93
  94         /* Default values, can be overridden by caller */
  95         new->spec.hold = WORK_QUEUE_DEFAULT_HOLD;
  96         new->spec.yield = THREAD_YIELD_TIME_SLOT;
  97         new->spec.retry = WORK_QUEUE_DEFAULT_RETRY;
  98
  99         return new;
 100 }
 101
 102 void work_queue_free_and_null(struct work_queue **wqp)
 103 {
 104         struct work_queue *wq = *wqp;
 105
 106         THREAD_OFF(wq->thread);
 107
 108         while (!work_queue_empty(wq)) {
 109                 struct work_queue_item *item = work_queue_last_item(wq);
 110
 111                 work_queue_item_remove(wq, item);
 112         }
 113
 114         listnode_delete(work_queues, wq);
 115
 116         XFREE(MTYPE_WORK_QUEUE_NAME, wq->name);
 117         XFREE(MTYPE_WORK_QUEUE, wq);
 118
 119         *wqp = NULL;
 120 }
 121
 122 bool work_queue_is_scheduled(struct work_queue *wq)
 123 {
 124         return thread_is_scheduled(wq->thread);
 125 }
 126
 127 static int work_queue_schedule(struct work_queue *wq, unsigned int delay)
 128 {
 129         /* if appropriate, schedule work queue thread */
 130         if (CHECK_FLAG(wq->flags, WQ_UNPLUGGED) &&
 131             !thread_is_scheduled(wq->thread) && !work_queue_empty(wq)) {
 132                 /* Schedule timer if there's a delay, otherwise just schedule
 133                  * as an 'event'
 134                  */
 135                 if (delay > 0) {
 136                         thread_add_timer_msec(wq->master, work_queue_run, wq,
 137                                               delay, &wq->thread);
 138                         thread_ignore_late_timer(wq->thread);
 139                 } else
 140                         thread_add_event(wq->master, work_queue_run, wq, 0,
 141                                          &wq->thread);
 142
 143                 /* set thread yield time, if needed */
 144                 if (thread_is_scheduled(wq->thread) &&
 145                     wq->spec.yield != THREAD_YIELD_TIME_SLOT)
 146                         thread_set_yield_time(wq->thread, wq->spec.yield);
 147                 return 1;
 148         } else
 149                 return 0;
 150 }
 151
 152 void work_queue_add(struct work_queue *wq, void *data)
 153 {
 154         struct work_queue_item *item;
 155
 156         assert(wq);
 157
 158         item = work_queue_item_new(wq);
 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         THREAD_OFF(wq->thread);
 217
 218         UNSET_FLAG(wq->flags, WQ_UNPLUGGED);
 219 }
 220
 221 /* unplug queue, schedule it again, if appropriate
 222  * Ie: Allow the queue to be drained again
 223  */
 224 void work_queue_unplug(struct work_queue *wq)
 225 {
 226         SET_FLAG(wq->flags, WQ_UNPLUGGED);
 227
 228         /* if thread isnt already waiting, add one */
 229         work_queue_schedule(wq, wq->spec.hold);
 230 }
 231
 232 /* timer thread to process a work queue
 233  * will reschedule itself if required,
 234  * otherwise work_queue_item_add
 235  */
 236 void work_queue_run(struct thread *thread)
 237 {
 238         struct work_queue *wq;
 239         struct work_queue_item *item, *titem;
 240         wq_item_status ret = WQ_SUCCESS;
 241         unsigned int cycles = 0;
 242         char yielded = 0;
 243
 244         wq = THREAD_ARG(thread);
 245
 246         assert(wq);
 247
 248         /* calculate cycle granularity:
 249          * list iteration == 1 run
 250          * listnode processing == 1 cycle
 251          * granularity == # cycles between checks whether we should yield.
 252          *
 253          * granularity should be > 0, and can increase slowly after each run to
 254          * provide some hysteris, but not past cycles.best or 2*cycles.
 255          *
 256          * Best: starts low, can only increase
 257          *
 258          * Granularity: starts at WORK_QUEUE_MIN_GRANULARITY, can be decreased
 259          *              if we run to end of time slot, can increase otherwise
 260          *              by a small factor.
 261          *
 262          * We could use just the average and save some work, however we want to
 263          * be
 264          * able to adjust quickly to CPU pressure. Average wont shift much if
 265          * daemon has been running a long time.
 266          */
 267         if (wq->cycles.granularity == 0)
 268                 wq->cycles.granularity = WORK_QUEUE_MIN_GRANULARITY;
 269
 270         STAILQ_FOREACH_SAFE (item, &wq->items, wq, titem) {
 271                 assert(item->data);
 272
 273                 /* dont run items which are past their allowed retries */
 274                 if (item->ran > wq->spec.max_retries) {
 275                         work_queue_item_remove(wq, item);
 276                         continue;
 277                 }
 278
 279                 /* run and take care of items that want to be retried
 280                  * immediately */
 281                 do {
 282                         ret = wq->spec.workfunc(wq, item->data);
 283                         item->ran++;
 284                 } while ((ret == WQ_RETRY_NOW)
 285                          && (item->ran < wq->spec.max_retries));
 286
 287                 switch (ret) {
 288                 case WQ_QUEUE_BLOCKED: {
 289                         /* decrement item->ran again, cause this isn't an item
 290                          * specific error, and fall through to WQ_RETRY_LATER
 291                          */
 292                         item->ran--;
 293                 }
 294                 case WQ_RETRY_LATER: {
 295                         goto stats;
 296                 }
 297                 case WQ_REQUEUE: {
 298                         item->ran--;
 299                         work_queue_item_requeue(wq, item);
 300                         /* If a single node is being used with a meta-queue
 301                          * (e.g., zebra),
 302                          * update the next node as we don't want to exit the
 303                          * thread and
 304                          * reschedule it after every node. By definition,
 305                          * WQ_REQUEUE is
 306                          * meant to continue the processing; the yield logic
 307                          * will kick in
 308                          * to terminate the thread when time has exceeded.
 309                          */
 310                         if (titem == NULL)
 311                                 titem = item;
 312                         break;
 313                 }
 314                 case WQ_RETRY_NOW:
 315                 /* a RETRY_NOW that gets here has exceeded max_tries, same as
 316                  * ERROR */
 317                 /* fallthru */
 318                 case WQ_SUCCESS:
 319                 default: {
 320                         work_queue_item_remove(wq, item);
 321                         break;
 322                 }
 323                 }
 324
 325                 /* completed cycle */
 326                 cycles++;
 327
 328                 /* test if we should yield */
 329                 if (!(cycles % wq->cycles.granularity)
 330                     && thread_should_yield(thread)) {
 331                         yielded = 1;
 332                         goto stats;
 333                 }
 334         }
 335
 336 stats:
 337
 338 #define WQ_HYSTERESIS_FACTOR 4
 339
 340         /* we yielded, check whether granularity should be reduced */
 341         if (yielded && (cycles < wq->cycles.granularity)) {
 342                 wq->cycles.granularity =
 343                         ((cycles > 0) ? cycles : WORK_QUEUE_MIN_GRANULARITY);
 344         }
 345         /* otherwise, should granularity increase? */
 346         else if (cycles >= (wq->cycles.granularity)) {
 347                 if (cycles > wq->cycles.best)
 348                         wq->cycles.best = cycles;
 349
 350                 /* along with yielded check, provides hysteresis for granularity
 351                  */
 352                 if (cycles > (wq->cycles.granularity * WQ_HYSTERESIS_FACTOR
 353                               * WQ_HYSTERESIS_FACTOR))
 354                         wq->cycles.granularity *=
 355                                 WQ_HYSTERESIS_FACTOR; /* quick ramp-up */
 356                 else if (cycles
 357                          > (wq->cycles.granularity * WQ_HYSTERESIS_FACTOR))
 358                         wq->cycles.granularity += WQ_HYSTERESIS_FACTOR;
 359         }
 360 #undef WQ_HYSTERIS_FACTOR
 361
 362         wq->runs++;
 363         wq->cycles.total += cycles;
 364         if (yielded)
 365                 wq->yields++;
 366
 367         /* Is the queue done yet? If it is, call the completion callback. */
 368         if (!work_queue_empty(wq)) {
 369                 if (ret == WQ_RETRY_LATER ||
 370                     ret == WQ_QUEUE_BLOCKED)
 371                         work_queue_schedule(wq, wq->spec.retry);
 372                 else
 373                         work_queue_schedule(wq, 0);
 374
 375         } else if (wq->spec.completion_func)
 376                 wq->spec.completion_func(wq);
 377 }