~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/block/blk-throttle.c

Version: ~ [ linux-5.10-rc5 ] ~ [ linux-5.9.10 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.79 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.159 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.208 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.245 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.245 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.85 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * Interface for controlling IO bandwidth on a request queue
  3  *
  4  * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
  5  */
  6 
  7 #include <linux/module.h>
  8 #include <linux/slab.h>
  9 #include <linux/blkdev.h>
 10 #include <linux/bio.h>
 11 #include <linux/blktrace_api.h>
 12 #include "blk-cgroup.h"
 13 #include "blk.h"
 14 
 15 /* Max dispatch from a group in 1 round */
 16 static int throtl_grp_quantum = 8;
 17 
 18 /* Total max dispatch from all groups in one round */
 19 static int throtl_quantum = 32;
 20 
 21 /* Throttling is performed over 100ms slice and after that slice is renewed */
 22 static unsigned long throtl_slice = HZ/10;      /* 100 ms */
 23 
 24 static struct blkcg_policy blkcg_policy_throtl;
 25 
 26 /* A workqueue to queue throttle related work */
 27 static struct workqueue_struct *kthrotld_workqueue;
 28 
 29 /*
 30  * To implement hierarchical throttling, throtl_grps form a tree and bios
 31  * are dispatched upwards level by level until they reach the top and get
 32  * issued.  When dispatching bios from the children and local group at each
 33  * level, if the bios are dispatched into a single bio_list, there's a risk
 34  * of a local or child group which can queue many bios at once filling up
 35  * the list starving others.
 36  *
 37  * To avoid such starvation, dispatched bios are queued separately
 38  * according to where they came from.  When they are again dispatched to
 39  * the parent, they're popped in round-robin order so that no single source
 40  * hogs the dispatch window.
 41  *
 42  * throtl_qnode is used to keep the queued bios separated by their sources.
 43  * Bios are queued to throtl_qnode which in turn is queued to
 44  * throtl_service_queue and then dispatched in round-robin order.
 45  *
 46  * It's also used to track the reference counts on blkg's.  A qnode always
 47  * belongs to a throtl_grp and gets queued on itself or the parent, so
 48  * incrementing the reference of the associated throtl_grp when a qnode is
 49  * queued and decrementing when dequeued is enough to keep the whole blkg
 50  * tree pinned while bios are in flight.
 51  */
 52 struct throtl_qnode {
 53         struct list_head        node;           /* service_queue->queued[] */
 54         struct bio_list         bios;           /* queued bios */
 55         struct throtl_grp       *tg;            /* tg this qnode belongs to */
 56 };
 57 
 58 struct throtl_service_queue {
 59         struct throtl_service_queue *parent_sq; /* the parent service_queue */
 60 
 61         /*
 62          * Bios queued directly to this service_queue or dispatched from
 63          * children throtl_grp's.
 64          */
 65         struct list_head        queued[2];      /* throtl_qnode [READ/WRITE] */
 66         unsigned int            nr_queued[2];   /* number of queued bios */
 67 
 68         /*
 69          * RB tree of active children throtl_grp's, which are sorted by
 70          * their ->disptime.
 71          */
 72         struct rb_root          pending_tree;   /* RB tree of active tgs */
 73         struct rb_node          *first_pending; /* first node in the tree */
 74         unsigned int            nr_pending;     /* # queued in the tree */
 75         unsigned long           first_pending_disptime; /* disptime of the first tg */
 76         struct timer_list       pending_timer;  /* fires on first_pending_disptime */
 77 };
 78 
 79 enum tg_state_flags {
 80         THROTL_TG_PENDING       = 1 << 0,       /* on parent's pending tree */
 81         THROTL_TG_WAS_EMPTY     = 1 << 1,       /* bio_lists[] became non-empty */
 82 };
 83 
 84 #define rb_entry_tg(node)       rb_entry((node), struct throtl_grp, rb_node)
 85 
 86 /* Per-cpu group stats */
 87 struct tg_stats_cpu {
 88         /* total bytes transferred */
 89         struct blkg_rwstat              service_bytes;
 90         /* total IOs serviced, post merge */
 91         struct blkg_rwstat              serviced;
 92 };
 93 
 94 struct throtl_grp {
 95         /* must be the first member */
 96         struct blkg_policy_data pd;
 97 
 98         /* active throtl group service_queue member */
 99         struct rb_node rb_node;
100 
101         /* throtl_data this group belongs to */
102         struct throtl_data *td;
103 
104         /* this group's service queue */
105         struct throtl_service_queue service_queue;
106 
107         /*
108          * qnode_on_self is used when bios are directly queued to this
109          * throtl_grp so that local bios compete fairly with bios
110          * dispatched from children.  qnode_on_parent is used when bios are
111          * dispatched from this throtl_grp into its parent and will compete
112          * with the sibling qnode_on_parents and the parent's
113          * qnode_on_self.
114          */
115         struct throtl_qnode qnode_on_self[2];
116         struct throtl_qnode qnode_on_parent[2];
117 
118         /*
119          * Dispatch time in jiffies. This is the estimated time when group
120          * will unthrottle and is ready to dispatch more bio. It is used as
121          * key to sort active groups in service tree.
122          */
123         unsigned long disptime;
124 
125         unsigned int flags;
126 
127         /* are there any throtl rules between this group and td? */
128         bool has_rules[2];
129 
130         /* bytes per second rate limits */
131         uint64_t bps[2];
132 
133         /* IOPS limits */
134         unsigned int iops[2];
135 
136         /* Number of bytes disptached in current slice */
137         uint64_t bytes_disp[2];
138         /* Number of bio's dispatched in current slice */
139         unsigned int io_disp[2];
140 
141         /* When did we start a new slice */
142         unsigned long slice_start[2];
143         unsigned long slice_end[2];
144 
145         /* Per cpu stats pointer */
146         struct tg_stats_cpu __percpu *stats_cpu;
147 
148         /* List of tgs waiting for per cpu stats memory to be allocated */
149         struct list_head stats_alloc_node;
150 };
151 
152 struct throtl_data
153 {
154         /* service tree for active throtl groups */
155         struct throtl_service_queue service_queue;
156 
157         struct request_queue *queue;
158 
159         /* Total Number of queued bios on READ and WRITE lists */
160         unsigned int nr_queued[2];
161 
162         /*
163          * number of total undestroyed groups
164          */
165         unsigned int nr_undestroyed_grps;
166 
167         /* Work for dispatching throttled bios */
168         struct work_struct dispatch_work;
169 };
170 
171 /* list and work item to allocate percpu group stats */
172 static DEFINE_SPINLOCK(tg_stats_alloc_lock);
173 static LIST_HEAD(tg_stats_alloc_list);
174 
175 static void tg_stats_alloc_fn(struct work_struct *);
176 static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn);
177 
178 static void throtl_pending_timer_fn(unsigned long arg);
179 
180 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
181 {
182         return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
183 }
184 
185 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
186 {
187         return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
188 }
189 
190 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
191 {
192         return pd_to_blkg(&tg->pd);
193 }
194 
195 static inline struct throtl_grp *td_root_tg(struct throtl_data *td)
196 {
197         return blkg_to_tg(td->queue->root_blkg);
198 }
199 
200 /**
201  * sq_to_tg - return the throl_grp the specified service queue belongs to
202  * @sq: the throtl_service_queue of interest
203  *
204  * Return the throtl_grp @sq belongs to.  If @sq is the top-level one
205  * embedded in throtl_data, %NULL is returned.
206  */
207 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
208 {
209         if (sq && sq->parent_sq)
210                 return container_of(sq, struct throtl_grp, service_queue);
211         else
212                 return NULL;
213 }
214 
215 /**
216  * sq_to_td - return throtl_data the specified service queue belongs to
217  * @sq: the throtl_service_queue of interest
218  *
219  * A service_queue can be embeded in either a throtl_grp or throtl_data.
220  * Determine the associated throtl_data accordingly and return it.
221  */
222 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
223 {
224         struct throtl_grp *tg = sq_to_tg(sq);
225 
226         if (tg)
227                 return tg->td;
228         else
229                 return container_of(sq, struct throtl_data, service_queue);
230 }
231 
232 /**
233  * throtl_log - log debug message via blktrace
234  * @sq: the service_queue being reported
235  * @fmt: printf format string
236  * @args: printf args
237  *
238  * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
239  * throtl_grp; otherwise, just "throtl".
240  *
241  * TODO: this should be made a function and name formatting should happen
242  * after testing whether blktrace is enabled.
243  */
244 #define throtl_log(sq, fmt, args...)    do {                            \
245         struct throtl_grp *__tg = sq_to_tg((sq));                       \
246         struct throtl_data *__td = sq_to_td((sq));                      \
247                                                                         \
248         (void)__td;                                                     \
249         if ((__tg)) {                                                   \
250                 char __pbuf[128];                                       \
251                                                                         \
252                 blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf));    \
253                 blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \
254         } else {                                                        \
255                 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args);  \
256         }                                                               \
257 } while (0)
258 
259 static void tg_stats_init(struct tg_stats_cpu *tg_stats)
260 {
261         blkg_rwstat_init(&tg_stats->service_bytes);
262         blkg_rwstat_init(&tg_stats->serviced);
263 }
264 
265 /*
266  * Worker for allocating per cpu stat for tgs. This is scheduled on the
267  * system_wq once there are some groups on the alloc_list waiting for
268  * allocation.
269  */
270 static void tg_stats_alloc_fn(struct work_struct *work)
271 {
272         static struct tg_stats_cpu *stats_cpu;  /* this fn is non-reentrant */
273         struct delayed_work *dwork = to_delayed_work(work);
274         bool empty = false;
275 
276 alloc_stats:
277         if (!stats_cpu) {
278                 int cpu;
279 
280                 stats_cpu = alloc_percpu(struct tg_stats_cpu);
281                 if (!stats_cpu) {
282                         /* allocation failed, try again after some time */
283                         schedule_delayed_work(dwork, msecs_to_jiffies(10));
284                         return;
285                 }
286                 for_each_possible_cpu(cpu)
287                         tg_stats_init(per_cpu_ptr(stats_cpu, cpu));
288         }
289 
290         spin_lock_irq(&tg_stats_alloc_lock);
291 
292         if (!list_empty(&tg_stats_alloc_list)) {
293                 struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list,
294                                                          struct throtl_grp,
295                                                          stats_alloc_node);
296                 swap(tg->stats_cpu, stats_cpu);
297                 list_del_init(&tg->stats_alloc_node);
298         }
299 
300         empty = list_empty(&tg_stats_alloc_list);
301         spin_unlock_irq(&tg_stats_alloc_lock);
302         if (!empty)
303                 goto alloc_stats;
304 }
305 
306 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
307 {
308         INIT_LIST_HEAD(&qn->node);
309         bio_list_init(&qn->bios);
310         qn->tg = tg;
311 }
312 
313 /**
314  * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
315  * @bio: bio being added
316  * @qn: qnode to add bio to
317  * @queued: the service_queue->queued[] list @qn belongs to
318  *
319  * Add @bio to @qn and put @qn on @queued if it's not already on.
320  * @qn->tg's reference count is bumped when @qn is activated.  See the
321  * comment on top of throtl_qnode definition for details.
322  */
323 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
324                                  struct list_head *queued)
325 {
326         bio_list_add(&qn->bios, bio);
327         if (list_empty(&qn->node)) {
328                 list_add_tail(&qn->node, queued);
329                 blkg_get(tg_to_blkg(qn->tg));
330         }
331 }
332 
333 /**
334  * throtl_peek_queued - peek the first bio on a qnode list
335  * @queued: the qnode list to peek
336  */
337 static struct bio *throtl_peek_queued(struct list_head *queued)
338 {
339         struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
340         struct bio *bio;
341 
342         if (list_empty(queued))
343                 return NULL;
344 
345         bio = bio_list_peek(&qn->bios);
346         WARN_ON_ONCE(!bio);
347         return bio;
348 }
349 
350 /**
351  * throtl_pop_queued - pop the first bio form a qnode list
352  * @queued: the qnode list to pop a bio from
353  * @tg_to_put: optional out argument for throtl_grp to put
354  *
355  * Pop the first bio from the qnode list @queued.  After popping, the first
356  * qnode is removed from @queued if empty or moved to the end of @queued so
357  * that the popping order is round-robin.
358  *
359  * When the first qnode is removed, its associated throtl_grp should be put
360  * too.  If @tg_to_put is NULL, this function automatically puts it;
361  * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
362  * responsible for putting it.
363  */
364 static struct bio *throtl_pop_queued(struct list_head *queued,
365                                      struct throtl_grp **tg_to_put)
366 {
367         struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
368         struct bio *bio;
369 
370         if (list_empty(queued))
371                 return NULL;
372 
373         bio = bio_list_pop(&qn->bios);
374         WARN_ON_ONCE(!bio);
375 
376         if (bio_list_empty(&qn->bios)) {
377                 list_del_init(&qn->node);
378                 if (tg_to_put)
379                         *tg_to_put = qn->tg;
380                 else
381                         blkg_put(tg_to_blkg(qn->tg));
382         } else {
383                 list_move_tail(&qn->node, queued);
384         }
385 
386         return bio;
387 }
388 
389 /* init a service_queue, assumes the caller zeroed it */
390 static void throtl_service_queue_init(struct throtl_service_queue *sq,
391                                       struct throtl_service_queue *parent_sq)
392 {
393         INIT_LIST_HEAD(&sq->queued[0]);
394         INIT_LIST_HEAD(&sq->queued[1]);
395         sq->pending_tree = RB_ROOT;
396         sq->parent_sq = parent_sq;
397         setup_timer(&sq->pending_timer, throtl_pending_timer_fn,
398                     (unsigned long)sq);
399 }
400 
401 static void throtl_service_queue_exit(struct throtl_service_queue *sq)
402 {
403         del_timer_sync(&sq->pending_timer);
404 }
405 
406 static void throtl_pd_init(struct blkcg_gq *blkg)
407 {
408         struct throtl_grp *tg = blkg_to_tg(blkg);
409         struct throtl_data *td = blkg->q->td;
410         struct throtl_service_queue *parent_sq;
411         unsigned long flags;
412         int rw;
413 
414         /*
415          * If on the default hierarchy, we switch to properly hierarchical
416          * behavior where limits on a given throtl_grp are applied to the
417          * whole subtree rather than just the group itself.  e.g. If 16M
418          * read_bps limit is set on the root group, the whole system can't
419          * exceed 16M for the device.
420          *
421          * If not on the default hierarchy, the broken flat hierarchy
422          * behavior is retained where all throtl_grps are treated as if
423          * they're all separate root groups right below throtl_data.
424          * Limits of a group don't interact with limits of other groups
425          * regardless of the position of the group in the hierarchy.
426          */
427         parent_sq = &td->service_queue;
428 
429         if (cgroup_on_dfl(blkg->blkcg->css.cgroup) && blkg->parent)
430                 parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
431 
432         throtl_service_queue_init(&tg->service_queue, parent_sq);
433 
434         for (rw = READ; rw <= WRITE; rw++) {
435                 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
436                 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
437         }
438 
439         RB_CLEAR_NODE(&tg->rb_node);
440         tg->td = td;
441 
442         tg->bps[READ] = -1;
443         tg->bps[WRITE] = -1;
444         tg->iops[READ] = -1;
445         tg->iops[WRITE] = -1;
446 
447         /*
448          * Ugh... We need to perform per-cpu allocation for tg->stats_cpu
449          * but percpu allocator can't be called from IO path.  Queue tg on
450          * tg_stats_alloc_list and allocate from work item.
451          */
452         spin_lock_irqsave(&tg_stats_alloc_lock, flags);
453         list_add(&tg->stats_alloc_node, &tg_stats_alloc_list);
454         schedule_delayed_work(&tg_stats_alloc_work, 0);
455         spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
456 }
457 
458 /*
459  * Set has_rules[] if @tg or any of its parents have limits configured.
460  * This doesn't require walking up to the top of the hierarchy as the
461  * parent's has_rules[] is guaranteed to be correct.
462  */
463 static void tg_update_has_rules(struct throtl_grp *tg)
464 {
465         struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
466         int rw;
467 
468         for (rw = READ; rw <= WRITE; rw++)
469                 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
470                                     (tg->bps[rw] != -1 || tg->iops[rw] != -1);
471 }
472 
473 static void throtl_pd_online(struct blkcg_gq *blkg)
474 {
475         /*
476          * We don't want new groups to escape the limits of its ancestors.
477          * Update has_rules[] after a new group is brought online.
478          */
479         tg_update_has_rules(blkg_to_tg(blkg));
480 }
481 
482 static void throtl_pd_exit(struct blkcg_gq *blkg)
483 {
484         struct throtl_grp *tg = blkg_to_tg(blkg);
485         unsigned long flags;
486 
487         spin_lock_irqsave(&tg_stats_alloc_lock, flags);
488         list_del_init(&tg->stats_alloc_node);
489         spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
490 
491         free_percpu(tg->stats_cpu);
492 
493         throtl_service_queue_exit(&tg->service_queue);
494 }
495 
496 static void throtl_pd_reset_stats(struct blkcg_gq *blkg)
497 {
498         struct throtl_grp *tg = blkg_to_tg(blkg);
499         int cpu;
500 
501         if (tg->stats_cpu == NULL)
502                 return;
503 
504         for_each_possible_cpu(cpu) {
505                 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
506 
507                 blkg_rwstat_reset(&sc->service_bytes);
508                 blkg_rwstat_reset(&sc->serviced);
509         }
510 }
511 
512 static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td,
513                                            struct blkcg *blkcg)
514 {
515         /*
516          * This is the common case when there are no blkcgs.  Avoid lookup
517          * in this case
518          */
519         if (blkcg == &blkcg_root)
520                 return td_root_tg(td);
521 
522         return blkg_to_tg(blkg_lookup(blkcg, td->queue));
523 }
524 
525 static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td,
526                                                   struct blkcg *blkcg)
527 {
528         struct request_queue *q = td->queue;
529         struct throtl_grp *tg = NULL;
530 
531         /*
532          * This is the common case when there are no blkcgs.  Avoid lookup
533          * in this case
534          */
535         if (blkcg == &blkcg_root) {
536                 tg = td_root_tg(td);
537         } else {
538                 struct blkcg_gq *blkg;
539 
540                 blkg = blkg_lookup_create(blkcg, q);
541 
542                 /* if %NULL and @q is alive, fall back to root_tg */
543                 if (!IS_ERR(blkg))
544                         tg = blkg_to_tg(blkg);
545                 else if (!blk_queue_dying(q))
546                         tg = td_root_tg(td);
547         }
548 
549         return tg;
550 }
551 
552 static struct throtl_grp *
553 throtl_rb_first(struct throtl_service_queue *parent_sq)
554 {
555         /* Service tree is empty */
556         if (!parent_sq->nr_pending)
557                 return NULL;
558 
559         if (!parent_sq->first_pending)
560                 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
561 
562         if (parent_sq->first_pending)
563                 return rb_entry_tg(parent_sq->first_pending);
564 
565         return NULL;
566 }
567 
568 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
569 {
570         rb_erase(n, root);
571         RB_CLEAR_NODE(n);
572 }
573 
574 static void throtl_rb_erase(struct rb_node *n,
575                             struct throtl_service_queue *parent_sq)
576 {
577         if (parent_sq->first_pending == n)
578                 parent_sq->first_pending = NULL;
579         rb_erase_init(n, &parent_sq->pending_tree);
580         --parent_sq->nr_pending;
581 }
582 
583 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
584 {
585         struct throtl_grp *tg;
586 
587         tg = throtl_rb_first(parent_sq);
588         if (!tg)
589                 return;
590 
591         parent_sq->first_pending_disptime = tg->disptime;
592 }
593 
594 static void tg_service_queue_add(struct throtl_grp *tg)
595 {
596         struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
597         struct rb_node **node = &parent_sq->pending_tree.rb_node;
598         struct rb_node *parent = NULL;
599         struct throtl_grp *__tg;
600         unsigned long key = tg->disptime;
601         int left = 1;
602 
603         while (*node != NULL) {
604                 parent = *node;
605                 __tg = rb_entry_tg(parent);
606 
607                 if (time_before(key, __tg->disptime))
608                         node = &parent->rb_left;
609                 else {
610                         node = &parent->rb_right;
611                         left = 0;
612                 }
613         }
614 
615         if (left)
616                 parent_sq->first_pending = &tg->rb_node;
617 
618         rb_link_node(&tg->rb_node, parent, node);
619         rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
620 }
621 
622 static void __throtl_enqueue_tg(struct throtl_grp *tg)
623 {
624         tg_service_queue_add(tg);
625         tg->flags |= THROTL_TG_PENDING;
626         tg->service_queue.parent_sq->nr_pending++;
627 }
628 
629 static void throtl_enqueue_tg(struct throtl_grp *tg)
630 {
631         if (!(tg->flags & THROTL_TG_PENDING))
632                 __throtl_enqueue_tg(tg);
633 }
634 
635 static void __throtl_dequeue_tg(struct throtl_grp *tg)
636 {
637         throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
638         tg->flags &= ~THROTL_TG_PENDING;
639 }
640 
641 static void throtl_dequeue_tg(struct throtl_grp *tg)
642 {
643         if (tg->flags & THROTL_TG_PENDING)
644                 __throtl_dequeue_tg(tg);
645 }
646 
647 /* Call with queue lock held */
648 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
649                                           unsigned long expires)
650 {
651         unsigned long max_expire = jiffies + 8 * throtl_slice;
652 
653         /*
654          * Since we are adjusting the throttle limit dynamically, the sleep
655          * time calculated according to previous limit might be invalid. It's
656          * possible the cgroup sleep time is very long and no other cgroups
657          * have IO running so notify the limit changes. Make sure the cgroup
658          * doesn't sleep too long to avoid the missed notification.
659          */
660         if (time_after(expires, max_expire))
661                 expires = max_expire;
662         mod_timer(&sq->pending_timer, expires);
663         throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
664                    expires - jiffies, jiffies);
665 }
666 
667 /**
668  * throtl_schedule_next_dispatch - schedule the next dispatch cycle
669  * @sq: the service_queue to schedule dispatch for
670  * @force: force scheduling
671  *
672  * Arm @sq->pending_timer so that the next dispatch cycle starts on the
673  * dispatch time of the first pending child.  Returns %true if either timer
674  * is armed or there's no pending child left.  %false if the current
675  * dispatch window is still open and the caller should continue
676  * dispatching.
677  *
678  * If @force is %true, the dispatch timer is always scheduled and this
679  * function is guaranteed to return %true.  This is to be used when the
680  * caller can't dispatch itself and needs to invoke pending_timer
681  * unconditionally.  Note that forced scheduling is likely to induce short
682  * delay before dispatch starts even if @sq->first_pending_disptime is not
683  * in the future and thus shouldn't be used in hot paths.
684  */
685 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
686                                           bool force)
687 {
688         /* any pending children left? */
689         if (!sq->nr_pending)
690                 return true;
691 
692         update_min_dispatch_time(sq);
693 
694         /* is the next dispatch time in the future? */
695         if (force || time_after(sq->first_pending_disptime, jiffies)) {
696                 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
697                 return true;
698         }
699 
700         /* tell the caller to continue dispatching */
701         return false;
702 }
703 
704 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
705                 bool rw, unsigned long start)
706 {
707         tg->bytes_disp[rw] = 0;
708         tg->io_disp[rw] = 0;
709 
710         /*
711          * Previous slice has expired. We must have trimmed it after last
712          * bio dispatch. That means since start of last slice, we never used
713          * that bandwidth. Do try to make use of that bandwidth while giving
714          * credit.
715          */
716         if (time_after_eq(start, tg->slice_start[rw]))
717                 tg->slice_start[rw] = start;
718 
719         tg->slice_end[rw] = jiffies + throtl_slice;
720         throtl_log(&tg->service_queue,
721                    "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
722                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
723                    tg->slice_end[rw], jiffies);
724 }
725 
726 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
727 {
728         tg->bytes_disp[rw] = 0;
729         tg->io_disp[rw] = 0;
730         tg->slice_start[rw] = jiffies;
731         tg->slice_end[rw] = jiffies + throtl_slice;
732         throtl_log(&tg->service_queue,
733                    "[%c] new slice start=%lu end=%lu jiffies=%lu",
734                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
735                    tg->slice_end[rw], jiffies);
736 }
737 
738 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
739                                         unsigned long jiffy_end)
740 {
741         tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
742 }
743 
744 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
745                                        unsigned long jiffy_end)
746 {
747         tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
748         throtl_log(&tg->service_queue,
749                    "[%c] extend slice start=%lu end=%lu jiffies=%lu",
750                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
751                    tg->slice_end[rw], jiffies);
752 }
753 
754 /* Determine if previously allocated or extended slice is complete or not */
755 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
756 {
757         if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
758                 return false;
759 
760         return 1;
761 }
762 
763 /* Trim the used slices and adjust slice start accordingly */
764 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
765 {
766         unsigned long nr_slices, time_elapsed, io_trim;
767         u64 bytes_trim, tmp;
768 
769         BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
770 
771         /*
772          * If bps are unlimited (-1), then time slice don't get
773          * renewed. Don't try to trim the slice if slice is used. A new
774          * slice will start when appropriate.
775          */
776         if (throtl_slice_used(tg, rw))
777                 return;
778 
779         /*
780          * A bio has been dispatched. Also adjust slice_end. It might happen
781          * that initially cgroup limit was very low resulting in high
782          * slice_end, but later limit was bumped up and bio was dispached
783          * sooner, then we need to reduce slice_end. A high bogus slice_end
784          * is bad because it does not allow new slice to start.
785          */
786 
787         throtl_set_slice_end(tg, rw, jiffies + throtl_slice);
788 
789         time_elapsed = jiffies - tg->slice_start[rw];
790 
791         nr_slices = time_elapsed / throtl_slice;
792 
793         if (!nr_slices)
794                 return;
795         tmp = tg->bps[rw] * throtl_slice * nr_slices;
796         do_div(tmp, HZ);
797         bytes_trim = tmp;
798 
799         io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
800 
801         if (!bytes_trim && !io_trim)
802                 return;
803 
804         if (tg->bytes_disp[rw] >= bytes_trim)
805                 tg->bytes_disp[rw] -= bytes_trim;
806         else
807                 tg->bytes_disp[rw] = 0;
808 
809         if (tg->io_disp[rw] >= io_trim)
810                 tg->io_disp[rw] -= io_trim;
811         else
812                 tg->io_disp[rw] = 0;
813 
814         tg->slice_start[rw] += nr_slices * throtl_slice;
815 
816         throtl_log(&tg->service_queue,
817                    "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
818                    rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
819                    tg->slice_start[rw], tg->slice_end[rw], jiffies);
820 }
821 
822 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
823                                   unsigned long *wait)
824 {
825         bool rw = bio_data_dir(bio);
826         unsigned int io_allowed;
827         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
828         u64 tmp;
829 
830         jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
831 
832         /* Slice has just started. Consider one slice interval */
833         if (!jiffy_elapsed)
834                 jiffy_elapsed_rnd = throtl_slice;
835 
836         jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
837 
838         /*
839          * jiffy_elapsed_rnd should not be a big value as minimum iops can be
840          * 1 then at max jiffy elapsed should be equivalent of 1 second as we
841          * will allow dispatch after 1 second and after that slice should
842          * have been trimmed.
843          */
844 
845         tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
846         do_div(tmp, HZ);
847 
848         if (tmp > UINT_MAX)
849                 io_allowed = UINT_MAX;
850         else
851                 io_allowed = tmp;
852 
853         if (tg->io_disp[rw] + 1 <= io_allowed) {
854                 if (wait)
855                         *wait = 0;
856                 return true;
857         }
858 
859         /* Calc approx time to dispatch */
860         jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
861 
862         if (jiffy_wait > jiffy_elapsed)
863                 jiffy_wait = jiffy_wait - jiffy_elapsed;
864         else
865                 jiffy_wait = 1;
866 
867         if (wait)
868                 *wait = jiffy_wait;
869         return 0;
870 }
871 
872 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
873                                  unsigned long *wait)
874 {
875         bool rw = bio_data_dir(bio);
876         u64 bytes_allowed, extra_bytes, tmp;
877         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
878 
879         jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
880 
881         /* Slice has just started. Consider one slice interval */
882         if (!jiffy_elapsed)
883                 jiffy_elapsed_rnd = throtl_slice;
884 
885         jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
886 
887         tmp = tg->bps[rw] * jiffy_elapsed_rnd;
888         do_div(tmp, HZ);
889         bytes_allowed = tmp;
890 
891         if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) {
892                 if (wait)
893                         *wait = 0;
894                 return true;
895         }
896 
897         /* Calc approx time to dispatch */
898         extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed;
899         jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
900 
901         if (!jiffy_wait)
902                 jiffy_wait = 1;
903 
904         /*
905          * This wait time is without taking into consideration the rounding
906          * up we did. Add that time also.
907          */
908         jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
909         if (wait)
910                 *wait = jiffy_wait;
911         return 0;
912 }
913 
914 /*
915  * Returns whether one can dispatch a bio or not. Also returns approx number
916  * of jiffies to wait before this bio is with-in IO rate and can be dispatched
917  */
918 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
919                             unsigned long *wait)
920 {
921         bool rw = bio_data_dir(bio);
922         unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
923 
924         /*
925          * Currently whole state machine of group depends on first bio
926          * queued in the group bio list. So one should not be calling
927          * this function with a different bio if there are other bios
928          * queued.
929          */
930         BUG_ON(tg->service_queue.nr_queued[rw] &&
931                bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
932 
933         /* If tg->bps = -1, then BW is unlimited */
934         if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
935                 if (wait)
936                         *wait = 0;
937                 return true;
938         }
939 
940         /*
941          * If previous slice expired, start a new one otherwise renew/extend
942          * existing slice to make sure it is at least throtl_slice interval
943          * long since now.
944          */
945         if (throtl_slice_used(tg, rw))
946                 throtl_start_new_slice(tg, rw);
947         else {
948                 if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
949                         throtl_extend_slice(tg, rw, jiffies + throtl_slice);
950         }
951 
952         if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
953             tg_with_in_iops_limit(tg, bio, &iops_wait)) {
954                 if (wait)
955                         *wait = 0;
956                 return 1;
957         }
958 
959         max_wait = max(bps_wait, iops_wait);
960 
961         if (wait)
962                 *wait = max_wait;
963 
964         if (time_before(tg->slice_end[rw], jiffies + max_wait))
965                 throtl_extend_slice(tg, rw, jiffies + max_wait);
966 
967         return 0;
968 }
969 
970 static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes,
971                                          int rw)
972 {
973         struct throtl_grp *tg = blkg_to_tg(blkg);
974         struct tg_stats_cpu *stats_cpu;
975         unsigned long flags;
976 
977         /* If per cpu stats are not allocated yet, don't do any accounting. */
978         if (tg->stats_cpu == NULL)
979                 return;
980 
981         /*
982          * Disabling interrupts to provide mutual exclusion between two
983          * writes on same cpu. It probably is not needed for 64bit. Not
984          * optimizing that case yet.
985          */
986         local_irq_save(flags);
987 
988         stats_cpu = this_cpu_ptr(tg->stats_cpu);
989 
990         blkg_rwstat_add(&stats_cpu->serviced, rw, 1);
991         blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes);
992 
993         local_irq_restore(flags);
994 }
995 
996 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
997 {
998         bool rw = bio_data_dir(bio);
999 
1000         /* Charge the bio to the group */
1001         tg->bytes_disp[rw] += bio->bi_iter.bi_size;
1002         tg->io_disp[rw]++;
1003 
1004         /*
1005          * REQ_THROTTLED is used to prevent the same bio to be throttled
1006          * more than once as a throttled bio will go through blk-throtl the
1007          * second time when it eventually gets issued.  Set it when a bio
1008          * is being charged to a tg.
1009          *
1010          * Dispatch stats aren't recursive and each @bio should only be
1011          * accounted by the @tg it was originally associated with.  Let's
1012          * update the stats when setting REQ_THROTTLED for the first time
1013          * which is guaranteed to be for the @bio's original tg.
1014          */
1015         if (!(bio->bi_rw & REQ_THROTTLED)) {
1016                 bio->bi_rw |= REQ_THROTTLED;
1017                 throtl_update_dispatch_stats(tg_to_blkg(tg),
1018                                              bio->bi_iter.bi_size, bio->bi_rw);
1019         }
1020 }
1021 
1022 /**
1023  * throtl_add_bio_tg - add a bio to the specified throtl_grp
1024  * @bio: bio to add
1025  * @qn: qnode to use
1026  * @tg: the target throtl_grp
1027  *
1028  * Add @bio to @tg's service_queue using @qn.  If @qn is not specified,
1029  * tg->qnode_on_self[] is used.
1030  */
1031 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
1032                               struct throtl_grp *tg)
1033 {
1034         struct throtl_service_queue *sq = &tg->service_queue;
1035         bool rw = bio_data_dir(bio);
1036 
1037         if (!qn)
1038                 qn = &tg->qnode_on_self[rw];
1039 
1040         /*
1041          * If @tg doesn't currently have any bios queued in the same
1042          * direction, queueing @bio can change when @tg should be
1043          * dispatched.  Mark that @tg was empty.  This is automatically
1044          * cleaered on the next tg_update_disptime().
1045          */
1046         if (!sq->nr_queued[rw])
1047                 tg->flags |= THROTL_TG_WAS_EMPTY;
1048 
1049         throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1050 
1051         sq->nr_queued[rw]++;
1052         throtl_enqueue_tg(tg);
1053 }
1054 
1055 static void tg_update_disptime(struct throtl_grp *tg)
1056 {
1057         struct throtl_service_queue *sq = &tg->service_queue;
1058         unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1059         struct bio *bio;
1060 
1061         if ((bio = throtl_peek_queued(&sq->queued[READ])))
1062                 tg_may_dispatch(tg, bio, &read_wait);
1063 
1064         if ((bio = throtl_peek_queued(&sq->queued[WRITE])))
1065                 tg_may_dispatch(tg, bio, &write_wait);
1066 
1067         min_wait = min(read_wait, write_wait);
1068         disptime = jiffies + min_wait;
1069 
1070         /* Update dispatch time */
1071         throtl_dequeue_tg(tg);
1072         tg->disptime = disptime;
1073         throtl_enqueue_tg(tg);
1074 
1075         /* see throtl_add_bio_tg() */
1076         tg->flags &= ~THROTL_TG_WAS_EMPTY;
1077 }
1078 
1079 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1080                                         struct throtl_grp *parent_tg, bool rw)
1081 {
1082         if (throtl_slice_used(parent_tg, rw)) {
1083                 throtl_start_new_slice_with_credit(parent_tg, rw,
1084                                 child_tg->slice_start[rw]);
1085         }
1086 
1087 }
1088 
1089 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1090 {
1091         struct throtl_service_queue *sq = &tg->service_queue;
1092         struct throtl_service_queue *parent_sq = sq->parent_sq;
1093         struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1094         struct throtl_grp *tg_to_put = NULL;
1095         struct bio *bio;
1096 
1097         /*
1098          * @bio is being transferred from @tg to @parent_sq.  Popping a bio
1099          * from @tg may put its reference and @parent_sq might end up
1100          * getting released prematurely.  Remember the tg to put and put it
1101          * after @bio is transferred to @parent_sq.
1102          */
1103         bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1104         sq->nr_queued[rw]--;
1105 
1106         throtl_charge_bio(tg, bio);
1107 
1108         /*
1109          * If our parent is another tg, we just need to transfer @bio to
1110          * the parent using throtl_add_bio_tg().  If our parent is
1111          * @td->service_queue, @bio is ready to be issued.  Put it on its
1112          * bio_lists[] and decrease total number queued.  The caller is
1113          * responsible for issuing these bios.
1114          */
1115         if (parent_tg) {
1116                 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1117                 start_parent_slice_with_credit(tg, parent_tg, rw);
1118         } else {
1119                 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1120                                      &parent_sq->queued[rw]);
1121                 BUG_ON(tg->td->nr_queued[rw] <= 0);
1122                 tg->td->nr_queued[rw]--;
1123         }
1124 
1125         throtl_trim_slice(tg, rw);
1126 
1127         if (tg_to_put)
1128                 blkg_put(tg_to_blkg(tg_to_put));
1129 }
1130 
1131 static int throtl_dispatch_tg(struct throtl_grp *tg)
1132 {
1133         struct throtl_service_queue *sq = &tg->service_queue;
1134         unsigned int nr_reads = 0, nr_writes = 0;
1135         unsigned int max_nr_reads = throtl_grp_quantum*3/4;
1136         unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
1137         struct bio *bio;
1138 
1139         /* Try to dispatch 75% READS and 25% WRITES */
1140 
1141         while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1142                tg_may_dispatch(tg, bio, NULL)) {
1143 
1144                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1145                 nr_reads++;
1146 
1147                 if (nr_reads >= max_nr_reads)
1148                         break;
1149         }
1150 
1151         while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1152                tg_may_dispatch(tg, bio, NULL)) {
1153 
1154                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1155                 nr_writes++;
1156 
1157                 if (nr_writes >= max_nr_writes)
1158                         break;
1159         }
1160 
1161         return nr_reads + nr_writes;
1162 }
1163 
1164 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1165 {
1166         unsigned int nr_disp = 0;
1167 
1168         while (1) {
1169                 struct throtl_grp *tg = throtl_rb_first(parent_sq);
1170                 struct throtl_service_queue *sq = &tg->service_queue;
1171 
1172                 if (!tg)
1173                         break;
1174 
1175                 if (time_before(jiffies, tg->disptime))
1176                         break;
1177 
1178                 throtl_dequeue_tg(tg);
1179 
1180                 nr_disp += throtl_dispatch_tg(tg);
1181 
1182                 if (sq->nr_queued[0] || sq->nr_queued[1])
1183                         tg_update_disptime(tg);
1184 
1185                 if (nr_disp >= throtl_quantum)
1186                         break;
1187         }
1188 
1189         return nr_disp;
1190 }
1191 
1192 /**
1193  * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1194  * @arg: the throtl_service_queue being serviced
1195  *
1196  * This timer is armed when a child throtl_grp with active bio's become
1197  * pending and queued on the service_queue's pending_tree and expires when
1198  * the first child throtl_grp should be dispatched.  This function
1199  * dispatches bio's from the children throtl_grps to the parent
1200  * service_queue.
1201  *
1202  * If the parent's parent is another throtl_grp, dispatching is propagated
1203  * by either arming its pending_timer or repeating dispatch directly.  If
1204  * the top-level service_tree is reached, throtl_data->dispatch_work is
1205  * kicked so that the ready bio's are issued.
1206  */
1207 static void throtl_pending_timer_fn(unsigned long arg)
1208 {
1209         struct throtl_service_queue *sq = (void *)arg;
1210         struct throtl_grp *tg = sq_to_tg(sq);
1211         struct throtl_data *td = sq_to_td(sq);
1212         struct request_queue *q = td->queue;
1213         struct throtl_service_queue *parent_sq;
1214         bool dispatched;
1215         int ret;
1216 
1217         spin_lock_irq(q->queue_lock);
1218 again:
1219         parent_sq = sq->parent_sq;
1220         dispatched = false;
1221 
1222         while (true) {
1223                 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1224                            sq->nr_queued[READ] + sq->nr_queued[WRITE],
1225                            sq->nr_queued[READ], sq->nr_queued[WRITE]);
1226 
1227                 ret = throtl_select_dispatch(sq);
1228                 if (ret) {
1229                         throtl_log(sq, "bios disp=%u", ret);
1230                         dispatched = true;
1231                 }
1232 
1233                 if (throtl_schedule_next_dispatch(sq, false))
1234                         break;
1235 
1236                 /* this dispatch windows is still open, relax and repeat */
1237                 spin_unlock_irq(q->queue_lock);
1238                 cpu_relax();
1239                 spin_lock_irq(q->queue_lock);
1240         }
1241 
1242         if (!dispatched)
1243                 goto out_unlock;
1244 
1245         if (parent_sq) {
1246                 /* @parent_sq is another throl_grp, propagate dispatch */
1247                 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1248                         tg_update_disptime(tg);
1249                         if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1250                                 /* window is already open, repeat dispatching */
1251                                 sq = parent_sq;
1252                                 tg = sq_to_tg(sq);
1253                                 goto again;
1254                         }
1255                 }
1256         } else {
1257                 /* reached the top-level, queue issueing */
1258                 queue_work(kthrotld_workqueue, &td->dispatch_work);
1259         }
1260 out_unlock:
1261         spin_unlock_irq(q->queue_lock);
1262 }
1263 
1264 /**
1265  * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1266  * @work: work item being executed
1267  *
1268  * This function is queued for execution when bio's reach the bio_lists[]
1269  * of throtl_data->service_queue.  Those bio's are ready and issued by this
1270  * function.
1271  */
1272 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1273 {
1274         struct throtl_data *td = container_of(work, struct throtl_data,
1275                                               dispatch_work);
1276         struct throtl_service_queue *td_sq = &td->service_queue;
1277         struct request_queue *q = td->queue;
1278         struct bio_list bio_list_on_stack;
1279         struct bio *bio;
1280         struct blk_plug plug;
1281         int rw;
1282 
1283         bio_list_init(&bio_list_on_stack);
1284 
1285         spin_lock_irq(q->queue_lock);
1286         for (rw = READ; rw <= WRITE; rw++)
1287                 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1288                         bio_list_add(&bio_list_on_stack, bio);
1289         spin_unlock_irq(q->queue_lock);
1290 
1291         if (!bio_list_empty(&bio_list_on_stack)) {
1292                 blk_start_plug(&plug);
1293                 while((bio = bio_list_pop(&bio_list_on_stack)))
1294                         generic_make_request(bio);
1295                 blk_finish_plug(&plug);
1296         }
1297 }
1298 
1299 static u64 tg_prfill_cpu_rwstat(struct seq_file *sf,
1300                                 struct blkg_policy_data *pd, int off)
1301 {
1302         struct throtl_grp *tg = pd_to_tg(pd);
1303         struct blkg_rwstat rwstat = { }, tmp;
1304         int i, cpu;
1305 
1306         if (tg->stats_cpu == NULL)
1307                 return 0;
1308 
1309         for_each_possible_cpu(cpu) {
1310                 struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
1311 
1312                 tmp = blkg_rwstat_read((void *)sc + off);
1313                 for (i = 0; i < BLKG_RWSTAT_NR; i++)
1314                         rwstat.cnt[i] += tmp.cnt[i];
1315         }
1316 
1317         return __blkg_prfill_rwstat(sf, pd, &rwstat);
1318 }
1319 
1320 static int tg_print_cpu_rwstat(struct seq_file *sf, void *v)
1321 {
1322         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_cpu_rwstat,
1323                           &blkcg_policy_throtl, seq_cft(sf)->private, true);
1324         return 0;
1325 }
1326 
1327 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1328                               int off)
1329 {
1330         struct throtl_grp *tg = pd_to_tg(pd);
1331         u64 v = *(u64 *)((void *)tg + off);
1332 
1333         if (v == -1)
1334                 return 0;
1335         return __blkg_prfill_u64(sf, pd, v);
1336 }
1337 
1338 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1339                                int off)
1340 {
1341         struct throtl_grp *tg = pd_to_tg(pd);
1342         unsigned int v = *(unsigned int *)((void *)tg + off);
1343 
1344         if (v == -1)
1345                 return 0;
1346         return __blkg_prfill_u64(sf, pd, v);
1347 }
1348 
1349 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1350 {
1351         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1352                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1353         return 0;
1354 }
1355 
1356 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1357 {
1358         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1359                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1360         return 0;
1361 }
1362 
1363 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1364                            char *buf, size_t nbytes, loff_t off, bool is_u64)
1365 {
1366         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1367         struct blkg_conf_ctx ctx;
1368         struct throtl_grp *tg;
1369         struct throtl_service_queue *sq;
1370         struct blkcg_gq *blkg;
1371         struct cgroup_subsys_state *pos_css;
1372         int ret;
1373 
1374         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1375         if (ret)
1376                 return ret;
1377 
1378         tg = blkg_to_tg(ctx.blkg);
1379         sq = &tg->service_queue;
1380 
1381         if (!ctx.v)
1382                 ctx.v = -1;
1383 
1384         if (is_u64)
1385                 *(u64 *)((void *)tg + of_cft(of)->private) = ctx.v;
1386         else
1387                 *(unsigned int *)((void *)tg + of_cft(of)->private) = ctx.v;
1388 
1389         throtl_log(&tg->service_queue,
1390                    "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1391                    tg->bps[READ], tg->bps[WRITE],
1392                    tg->iops[READ], tg->iops[WRITE]);
1393 
1394         /*
1395          * Update has_rules[] flags for the updated tg's subtree.  A tg is
1396          * considered to have rules if either the tg itself or any of its
1397          * ancestors has rules.  This identifies groups without any
1398          * restrictions in the whole hierarchy and allows them to bypass
1399          * blk-throttle.
1400          */
1401         blkg_for_each_descendant_pre(blkg, pos_css, ctx.blkg)
1402                 tg_update_has_rules(blkg_to_tg(blkg));
1403 
1404         /*
1405          * We're already holding queue_lock and know @tg is valid.  Let's
1406          * apply the new config directly.
1407          *
1408          * Restart the slices for both READ and WRITES. It might happen
1409          * that a group's limit are dropped suddenly and we don't want to
1410          * account recently dispatched IO with new low rate.
1411          */
1412         throtl_start_new_slice(tg, 0);
1413         throtl_start_new_slice(tg, 1);
1414 
1415         if (tg->flags & THROTL_TG_PENDING) {
1416                 tg_update_disptime(tg);
1417                 throtl_schedule_next_dispatch(sq->parent_sq, true);
1418         }
1419 
1420         blkg_conf_finish(&ctx);
1421         return nbytes;
1422 }
1423 
1424 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1425                                char *buf, size_t nbytes, loff_t off)
1426 {
1427         return tg_set_conf(of, buf, nbytes, off, true);
1428 }
1429 
1430 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1431                                 char *buf, size_t nbytes, loff_t off)
1432 {
1433         return tg_set_conf(of, buf, nbytes, off, false);
1434 }
1435 
1436 static struct cftype throtl_files[] = {
1437         {
1438                 .name = "throttle.read_bps_device",
1439                 .private = offsetof(struct throtl_grp, bps[READ]),
1440                 .seq_show = tg_print_conf_u64,
1441                 .write = tg_set_conf_u64,
1442         },
1443         {
1444                 .name = "throttle.write_bps_device",
1445                 .private = offsetof(struct throtl_grp, bps[WRITE]),
1446                 .seq_show = tg_print_conf_u64,
1447                 .write = tg_set_conf_u64,
1448         },
1449         {
1450                 .name = "throttle.read_iops_device",
1451                 .private = offsetof(struct throtl_grp, iops[READ]),
1452                 .seq_show = tg_print_conf_uint,
1453                 .write = tg_set_conf_uint,
1454         },
1455         {
1456                 .name = "throttle.write_iops_device",
1457                 .private = offsetof(struct throtl_grp, iops[WRITE]),
1458                 .seq_show = tg_print_conf_uint,
1459                 .write = tg_set_conf_uint,
1460         },
1461         {
1462                 .name = "throttle.io_service_bytes",
1463                 .private = offsetof(struct tg_stats_cpu, service_bytes),
1464                 .seq_show = tg_print_cpu_rwstat,
1465         },
1466         {
1467                 .name = "throttle.io_serviced",
1468                 .private = offsetof(struct tg_stats_cpu, serviced),
1469                 .seq_show = tg_print_cpu_rwstat,
1470         },
1471         { }     /* terminate */
1472 };
1473 
1474 static void throtl_shutdown_wq(struct request_queue *q)
1475 {
1476         struct throtl_data *td = q->td;
1477 
1478         cancel_work_sync(&td->dispatch_work);
1479 }
1480 
1481 static struct blkcg_policy blkcg_policy_throtl = {
1482         .pd_size                = sizeof(struct throtl_grp),
1483         .cftypes                = throtl_files,
1484 
1485         .pd_init_fn             = throtl_pd_init,
1486         .pd_online_fn           = throtl_pd_online,
1487         .pd_exit_fn             = throtl_pd_exit,
1488         .pd_reset_stats_fn      = throtl_pd_reset_stats,
1489 };
1490 
1491 bool blk_throtl_bio(struct request_queue *q, struct bio *bio)
1492 {
1493         struct throtl_data *td = q->td;
1494         struct throtl_qnode *qn = NULL;
1495         struct throtl_grp *tg;
1496         struct throtl_service_queue *sq;
1497         bool rw = bio_data_dir(bio);
1498         struct blkcg *blkcg;
1499         bool throttled = false;
1500 
1501         /* see throtl_charge_bio() */
1502         if (bio->bi_rw & REQ_THROTTLED)
1503                 goto out;
1504 
1505         /*
1506          * A throtl_grp pointer retrieved under rcu can be used to access
1507          * basic fields like stats and io rates. If a group has no rules,
1508          * just update the dispatch stats in lockless manner and return.
1509          */
1510         rcu_read_lock();
1511         blkcg = bio_blkcg(bio);
1512         tg = throtl_lookup_tg(td, blkcg);
1513         if (tg) {
1514                 if (!tg->has_rules[rw]) {
1515                         throtl_update_dispatch_stats(tg_to_blkg(tg),
1516                                         bio->bi_iter.bi_size, bio->bi_rw);
1517                         goto out_unlock_rcu;
1518                 }
1519         }
1520 
1521         /*
1522          * Either group has not been allocated yet or it is not an unlimited
1523          * IO group
1524          */
1525         spin_lock_irq(q->queue_lock);
1526         tg = throtl_lookup_create_tg(td, blkcg);
1527         if (unlikely(!tg))
1528                 goto out_unlock;
1529 
1530         sq = &tg->service_queue;
1531 
1532         while (true) {
1533                 /* throtl is FIFO - if bios are already queued, should queue */
1534                 if (sq->nr_queued[rw])
1535                         break;
1536 
1537                 /* if above limits, break to queue */
1538                 if (!tg_may_dispatch(tg, bio, NULL))
1539                         break;
1540 
1541                 /* within limits, let's charge and dispatch directly */
1542                 throtl_charge_bio(tg, bio);
1543 
1544                 /*
1545                  * We need to trim slice even when bios are not being queued
1546                  * otherwise it might happen that a bio is not queued for
1547                  * a long time and slice keeps on extending and trim is not
1548                  * called for a long time. Now if limits are reduced suddenly
1549                  * we take into account all the IO dispatched so far at new
1550                  * low rate and * newly queued IO gets a really long dispatch
1551                  * time.
1552                  *
1553                  * So keep on trimming slice even if bio is not queued.
1554                  */
1555                 throtl_trim_slice(tg, rw);
1556 
1557                 /*
1558                  * @bio passed through this layer without being throttled.
1559                  * Climb up the ladder.  If we''re already at the top, it
1560                  * can be executed directly.
1561                  */
1562                 qn = &tg->qnode_on_parent[rw];
1563                 sq = sq->parent_sq;
1564                 tg = sq_to_tg(sq);
1565                 if (!tg)
1566                         goto out_unlock;
1567         }
1568 
1569         /* out-of-limit, queue to @tg */
1570         throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1571                    rw == READ ? 'R' : 'W',
1572                    tg->bytes_disp[rw], bio->bi_iter.bi_size, tg->bps[rw],
1573                    tg->io_disp[rw], tg->iops[rw],
1574                    sq->nr_queued[READ], sq->nr_queued[WRITE]);
1575 
1576         bio_associate_current(bio);
1577         tg->td->nr_queued[rw]++;
1578         throtl_add_bio_tg(bio, qn, tg);
1579         throttled = true;
1580 
1581         /*
1582          * Update @tg's dispatch time and force schedule dispatch if @tg
1583          * was empty before @bio.  The forced scheduling isn't likely to
1584          * cause undue delay as @bio is likely to be dispatched directly if
1585          * its @tg's disptime is not in the future.
1586          */
1587         if (tg->flags & THROTL_TG_WAS_EMPTY) {
1588                 tg_update_disptime(tg);
1589                 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
1590         }
1591 
1592 out_unlock:
1593         spin_unlock_irq(q->queue_lock);
1594 out_unlock_rcu:
1595         rcu_read_unlock();
1596 out:
1597         /*
1598          * As multiple blk-throtls may stack in the same issue path, we
1599          * don't want bios to leave with the flag set.  Clear the flag if
1600          * being issued.
1601          */
1602         if (!throttled)
1603                 bio->bi_rw &= ~REQ_THROTTLED;
1604         return throttled;
1605 }
1606 
1607 /*
1608  * Dispatch all bios from all children tg's queued on @parent_sq.  On
1609  * return, @parent_sq is guaranteed to not have any active children tg's
1610  * and all bios from previously active tg's are on @parent_sq->bio_lists[].
1611  */
1612 static void tg_drain_bios(struct throtl_service_queue *parent_sq)
1613 {
1614         struct throtl_grp *tg;
1615 
1616         while ((tg = throtl_rb_first(parent_sq))) {
1617                 struct throtl_service_queue *sq = &tg->service_queue;
1618                 struct bio *bio;
1619 
1620                 throtl_dequeue_tg(tg);
1621 
1622                 while ((bio = throtl_peek_queued(&sq->queued[READ])))
1623                         tg_dispatch_one_bio(tg, bio_data_dir(bio));
1624                 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
1625                         tg_dispatch_one_bio(tg, bio_data_dir(bio));
1626         }
1627 }
1628 
1629 /**
1630  * blk_throtl_drain - drain throttled bios
1631  * @q: request_queue to drain throttled bios for
1632  *
1633  * Dispatch all currently throttled bios on @q through ->make_request_fn().
1634  */
1635 void blk_throtl_drain(struct request_queue *q)
1636         __releases(q->queue_lock) __acquires(q->queue_lock)
1637 {
1638         struct throtl_data *td = q->td;
1639         struct blkcg_gq *blkg;
1640         struct cgroup_subsys_state *pos_css;
1641         struct bio *bio;
1642         int rw;
1643 
1644         queue_lockdep_assert_held(q);
1645         rcu_read_lock();
1646 
1647         /*
1648          * Drain each tg while doing post-order walk on the blkg tree, so
1649          * that all bios are propagated to td->service_queue.  It'd be
1650          * better to walk service_queue tree directly but blkg walk is
1651          * easier.
1652          */
1653         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
1654                 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
1655 
1656         /* finally, transfer bios from top-level tg's into the td */
1657         tg_drain_bios(&td->service_queue);
1658 
1659         rcu_read_unlock();
1660         spin_unlock_irq(q->queue_lock);
1661 
1662         /* all bios now should be in td->service_queue, issue them */
1663         for (rw = READ; rw <= WRITE; rw++)
1664                 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
1665                                                 NULL)))
1666                         generic_make_request(bio);
1667 
1668         spin_lock_irq(q->queue_lock);
1669 }
1670 
1671 int blk_throtl_init(struct request_queue *q)
1672 {
1673         struct throtl_data *td;
1674         int ret;
1675 
1676         td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1677         if (!td)
1678                 return -ENOMEM;
1679 
1680         INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
1681         throtl_service_queue_init(&td->service_queue, NULL);
1682 
1683         q->td = td;
1684         td->queue = q;
1685 
1686         /* activate policy */
1687         ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1688         if (ret)
1689                 kfree(td);
1690         return ret;
1691 }
1692 
1693 void blk_throtl_exit(struct request_queue *q)
1694 {
1695         BUG_ON(!q->td);
1696         throtl_shutdown_wq(q);
1697         blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1698         kfree(q->td);
1699 }
1700 
1701 static int __init throtl_init(void)
1702 {
1703         kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1704         if (!kthrotld_workqueue)
1705                 panic("Failed to create kthrotld\n");
1706 
1707         return blkcg_policy_register(&blkcg_policy_throtl);
1708 }
1709 
1710 module_init(throtl_init);
1711 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp