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(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 bool work_queue_is_scheduled(struct work_queue *wq)
 123 {
 124         return (wq->thread != NULL);
 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) && (wq->thread == NULL) &&
 131             !work_queue_empty(wq)) {
 132                 wq->thread = NULL;
 133                 thread_add_timer_msec(wq->master, work_queue_run, wq, delay,
 134                                       &wq->thread);
 135                 /* set thread yield time, if needed */
 136                 if (wq->thread && wq->spec.yield != THREAD_YIELD_TIME_SLOT)
 137                         thread_set_yield_time(wq->thread, wq->spec.yield);
 138                 return 1;
 139         } else
 140                 return 0;
 141 }
 142
 143 void work_queue_add(struct work_queue *wq, void *data)
 144 {
 145         struct work_queue_item *item;
 146
 147         assert(wq);
 148
 149         if (!(item = work_queue_item_new(wq))) {
 150                 zlog_err("%s: unable to get new queue item", __func__);
 151                 return;
 152         }
 153
 154         item->data = data;
 155         work_queue_item_enqueue(wq, item);
 156
 157         work_queue_schedule(wq, wq->spec.hold);
 158
 159         return;
 160 }
 161
 162 static void work_queue_item_requeue(struct work_queue *wq, struct work_queue_item *item)
 163 {
 164         work_queue_item_dequeue(wq, item);
 165
 166         /* attach to end of list */
 167         work_queue_item_enqueue(wq, item);
 168 }
 169
 170 DEFUN (show_work_queues,
 171        show_work_queues_cmd,
 172        "show work-queues",
 173        SHOW_STR
 174        "Work Queue information\n")
 175 {
 176         struct listnode *node;
 177         struct work_queue *wq;
 178
 179         vty_out(vty, "%c %8s %5s %8s %8s %21s\n", ' ', "List", "(ms) ",
 180                 "Q. Runs", "Yields", "Cycle Counts   ");
 181         vty_out(vty, "%c %8s %5s %8s %8s %7s %6s %8s %6s %s\n", 'P', "Items",
 182                 "Hold", "Total", "Total", "Best", "Gran.", "Total", "Avg.",
 183                 "Name");
 184
 185         for (ALL_LIST_ELEMENTS_RO(work_queues, node, wq)) {
 186                 vty_out(vty, "%c %8d %5d %8ld %8ld %7d %6d %8ld %6u %s\n",
 187                         (CHECK_FLAG(wq->flags, WQ_UNPLUGGED) ? ' ' : 'P'),
 188                         work_queue_item_count(wq), wq->spec.hold, wq->runs,
 189                         wq->yields, wq->cycles.best, wq->cycles.granularity,
 190                         wq->cycles.total,
 191                         (wq->runs) ? (unsigned int)(wq->cycles.total / wq->runs)
 192                                    : 0,
 193                         wq->name);
 194         }
 195
 196         return CMD_SUCCESS;
 197 }
 198
 199 void workqueue_cmd_init(void)
 200 {
 201         install_element(VIEW_NODE, &show_work_queues_cmd);
 202 }
 203
 204 /* 'plug' a queue: Stop it from being scheduled,
 205  * ie: prevent the queue from draining.
 206  */
 207 void work_queue_plug(struct work_queue *wq)
 208 {
 209         if (wq->thread)
 210                 thread_cancel(wq->thread);
 211
 212         wq->thread = NULL;
 213
 214         UNSET_FLAG(wq->flags, WQ_UNPLUGGED);
 215 }
 216
 217 /* unplug queue, schedule it again, if appropriate
 218  * Ie: Allow the queue to be drained again
 219  */
 220 void work_queue_unplug(struct work_queue *wq)
 221 {
 222         SET_FLAG(wq->flags, WQ_UNPLUGGED);
 223
 224         /* if thread isnt already waiting, add one */
 225         work_queue_schedule(wq, wq->spec.hold);
 226 }
 227
 228 /* timer thread to process a work queue
 229  * will reschedule itself if required,
 230  * otherwise work_queue_item_add
 231  */
 232 int work_queue_run(struct thread *thread)
 233 {
 234         struct work_queue *wq;
 235         struct work_queue_item *item, *titem;
 236         wq_item_status ret;
 237         unsigned int cycles = 0;
 238         char yielded = 0;
 239
 240         wq = THREAD_ARG(thread);
 241         wq->thread = NULL;
 242
 243         assert(wq);
 244
 245         /* calculate cycle granularity:
 246          * list iteration == 1 run
 247          * listnode processing == 1 cycle
 248          * granularity == # cycles between checks whether we should yield.
 249          *
 250          * granularity should be > 0, and can increase slowly after each run to
 251          * provide some hysteris, but not past cycles.best or 2*cycles.
 252          *
 253          * Best: starts low, can only increase
 254          *
 255          * Granularity: starts at WORK_QUEUE_MIN_GRANULARITY, can be decreased
 256          *              if we run to end of time slot, can increase otherwise
 257          *              by a small factor.
 258          *
 259          * We could use just the average and save some work, however we want to
 260          * be
 261          * able to adjust quickly to CPU pressure. Average wont shift much if
 262          * daemon has been running a long time.
 263          */
 264         if (wq->cycles.granularity == 0)
 265                 wq->cycles.granularity = WORK_QUEUE_MIN_GRANULARITY;
 266
 267         STAILQ_FOREACH_SAFE (item, &wq->items, wq, titem) {
 268                 assert(item && item->data);
 269
 270                 /* dont run items which are past their allowed retries */
 271                 if (item->ran > wq->spec.max_retries) {
 272                         /* run error handler, if any */
 273                         if (wq->spec.errorfunc)
 274                                 wq->spec.errorfunc(wq, item->data);
 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                 case WQ_ERROR: {
 318                         if (wq->spec.errorfunc)
 319                                 wq->spec.errorfunc(wq, item);
 320                 }
 321                 /* fallthru */
 322                 case WQ_SUCCESS:
 323                 default: {
 324                         work_queue_item_remove(wq, item);
 325                         break;
 326                 }
 327                 }
 328
 329                 /* completed cycle */
 330                 cycles++;
 331
 332                 /* test if we should yield */
 333                 if (!(cycles % wq->cycles.granularity)
 334                     && thread_should_yield(thread)) {
 335                         yielded = 1;
 336                         goto stats;
 337                 }
 338         }
 339
 340 stats:
 341
 342 #define WQ_HYSTERESIS_FACTOR 4
 343
 344         /* we yielded, check whether granularity should be reduced */
 345         if (yielded && (cycles < wq->cycles.granularity)) {
 346                 wq->cycles.granularity =
 347                         ((cycles > 0) ? cycles : WORK_QUEUE_MIN_GRANULARITY);
 348         }
 349         /* otherwise, should granularity increase? */
 350         else if (cycles >= (wq->cycles.granularity)) {
 351                 if (cycles > wq->cycles.best)
 352                         wq->cycles.best = cycles;
 353
 354                 /* along with yielded check, provides hysteresis for granularity
 355                  */
 356                 if (cycles > (wq->cycles.granularity * WQ_HYSTERESIS_FACTOR
 357                               * WQ_HYSTERESIS_FACTOR))
 358                         wq->cycles.granularity *=
 359                                 WQ_HYSTERESIS_FACTOR; /* quick ramp-up */
 360                 else if (cycles
 361                          > (wq->cycles.granularity * WQ_HYSTERESIS_FACTOR))
 362                         wq->cycles.granularity += WQ_HYSTERESIS_FACTOR;
 363         }
 364 #undef WQ_HYSTERIS_FACTOR
 365
 366         wq->runs++;
 367         wq->cycles.total += cycles;
 368         if (yielded)
 369                 wq->yields++;
 370
 371 #if 0
 372   printf ("%s: cycles %d, new: best %d, worst %d\n",
 373             __func__, cycles, wq->cycles.best, wq->cycles.granularity);
 374 #endif
 375
 376         /* Is the queue done yet? If it is, call the completion callback. */
 377         if (!work_queue_empty(wq))
 378                 work_queue_schedule(wq, 0);
 379         else if (wq->spec.completion_func)
 380                 wq->spec.completion_func(wq);
 381
 382         return 0;
 383 }