Back to home page

LXR

 
 

    


0001 /*
0002  * Interface for controlling IO bandwidth on a request queue
0003  *
0004  * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
0005  */
0006 
0007 #include <linux/module.h>
0008 #include <linux/slab.h>
0009 #include <linux/blkdev.h>
0010 #include <linux/bio.h>
0011 #include <linux/blktrace_api.h>
0012 #include <linux/blk-cgroup.h>
0013 #include "blk.h"
0014 
0015 /* Max dispatch from a group in 1 round */
0016 static int throtl_grp_quantum = 8;
0017 
0018 /* Total max dispatch from all groups in one round */
0019 static int throtl_quantum = 32;
0020 
0021 /* Throttling is performed over 100ms slice and after that slice is renewed */
0022 static unsigned long throtl_slice = HZ/10;  /* 100 ms */
0023 
0024 static struct blkcg_policy blkcg_policy_throtl;
0025 
0026 /* A workqueue to queue throttle related work */
0027 static struct workqueue_struct *kthrotld_workqueue;
0028 
0029 /*
0030  * To implement hierarchical throttling, throtl_grps form a tree and bios
0031  * are dispatched upwards level by level until they reach the top and get
0032  * issued.  When dispatching bios from the children and local group at each
0033  * level, if the bios are dispatched into a single bio_list, there's a risk
0034  * of a local or child group which can queue many bios at once filling up
0035  * the list starving others.
0036  *
0037  * To avoid such starvation, dispatched bios are queued separately
0038  * according to where they came from.  When they are again dispatched to
0039  * the parent, they're popped in round-robin order so that no single source
0040  * hogs the dispatch window.
0041  *
0042  * throtl_qnode is used to keep the queued bios separated by their sources.
0043  * Bios are queued to throtl_qnode which in turn is queued to
0044  * throtl_service_queue and then dispatched in round-robin order.
0045  *
0046  * It's also used to track the reference counts on blkg's.  A qnode always
0047  * belongs to a throtl_grp and gets queued on itself or the parent, so
0048  * incrementing the reference of the associated throtl_grp when a qnode is
0049  * queued and decrementing when dequeued is enough to keep the whole blkg
0050  * tree pinned while bios are in flight.
0051  */
0052 struct throtl_qnode {
0053     struct list_head    node;       /* service_queue->queued[] */
0054     struct bio_list     bios;       /* queued bios */
0055     struct throtl_grp   *tg;        /* tg this qnode belongs to */
0056 };
0057 
0058 struct throtl_service_queue {
0059     struct throtl_service_queue *parent_sq; /* the parent service_queue */
0060 
0061     /*
0062      * Bios queued directly to this service_queue or dispatched from
0063      * children throtl_grp's.
0064      */
0065     struct list_head    queued[2];  /* throtl_qnode [READ/WRITE] */
0066     unsigned int        nr_queued[2];   /* number of queued bios */
0067 
0068     /*
0069      * RB tree of active children throtl_grp's, which are sorted by
0070      * their ->disptime.
0071      */
0072     struct rb_root      pending_tree;   /* RB tree of active tgs */
0073     struct rb_node      *first_pending; /* first node in the tree */
0074     unsigned int        nr_pending; /* # queued in the tree */
0075     unsigned long       first_pending_disptime; /* disptime of the first tg */
0076     struct timer_list   pending_timer;  /* fires on first_pending_disptime */
0077 };
0078 
0079 enum tg_state_flags {
0080     THROTL_TG_PENDING   = 1 << 0,   /* on parent's pending tree */
0081     THROTL_TG_WAS_EMPTY = 1 << 1,   /* bio_lists[] became non-empty */
0082 };
0083 
0084 #define rb_entry_tg(node)   rb_entry((node), struct throtl_grp, rb_node)
0085 
0086 struct throtl_grp {
0087     /* must be the first member */
0088     struct blkg_policy_data pd;
0089 
0090     /* active throtl group service_queue member */
0091     struct rb_node rb_node;
0092 
0093     /* throtl_data this group belongs to */
0094     struct throtl_data *td;
0095 
0096     /* this group's service queue */
0097     struct throtl_service_queue service_queue;
0098 
0099     /*
0100      * qnode_on_self is used when bios are directly queued to this
0101      * throtl_grp so that local bios compete fairly with bios
0102      * dispatched from children.  qnode_on_parent is used when bios are
0103      * dispatched from this throtl_grp into its parent and will compete
0104      * with the sibling qnode_on_parents and the parent's
0105      * qnode_on_self.
0106      */
0107     struct throtl_qnode qnode_on_self[2];
0108     struct throtl_qnode qnode_on_parent[2];
0109 
0110     /*
0111      * Dispatch time in jiffies. This is the estimated time when group
0112      * will unthrottle and is ready to dispatch more bio. It is used as
0113      * key to sort active groups in service tree.
0114      */
0115     unsigned long disptime;
0116 
0117     unsigned int flags;
0118 
0119     /* are there any throtl rules between this group and td? */
0120     bool has_rules[2];
0121 
0122     /* bytes per second rate limits */
0123     uint64_t bps[2];
0124 
0125     /* IOPS limits */
0126     unsigned int iops[2];
0127 
0128     /* Number of bytes disptached in current slice */
0129     uint64_t bytes_disp[2];
0130     /* Number of bio's dispatched in current slice */
0131     unsigned int io_disp[2];
0132 
0133     /* When did we start a new slice */
0134     unsigned long slice_start[2];
0135     unsigned long slice_end[2];
0136 };
0137 
0138 struct throtl_data
0139 {
0140     /* service tree for active throtl groups */
0141     struct throtl_service_queue service_queue;
0142 
0143     struct request_queue *queue;
0144 
0145     /* Total Number of queued bios on READ and WRITE lists */
0146     unsigned int nr_queued[2];
0147 
0148     /* Work for dispatching throttled bios */
0149     struct work_struct dispatch_work;
0150 };
0151 
0152 static void throtl_pending_timer_fn(unsigned long arg);
0153 
0154 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
0155 {
0156     return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
0157 }
0158 
0159 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
0160 {
0161     return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
0162 }
0163 
0164 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
0165 {
0166     return pd_to_blkg(&tg->pd);
0167 }
0168 
0169 /**
0170  * sq_to_tg - return the throl_grp the specified service queue belongs to
0171  * @sq: the throtl_service_queue of interest
0172  *
0173  * Return the throtl_grp @sq belongs to.  If @sq is the top-level one
0174  * embedded in throtl_data, %NULL is returned.
0175  */
0176 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
0177 {
0178     if (sq && sq->parent_sq)
0179         return container_of(sq, struct throtl_grp, service_queue);
0180     else
0181         return NULL;
0182 }
0183 
0184 /**
0185  * sq_to_td - return throtl_data the specified service queue belongs to
0186  * @sq: the throtl_service_queue of interest
0187  *
0188  * A service_queue can be embeded in either a throtl_grp or throtl_data.
0189  * Determine the associated throtl_data accordingly and return it.
0190  */
0191 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
0192 {
0193     struct throtl_grp *tg = sq_to_tg(sq);
0194 
0195     if (tg)
0196         return tg->td;
0197     else
0198         return container_of(sq, struct throtl_data, service_queue);
0199 }
0200 
0201 /**
0202  * throtl_log - log debug message via blktrace
0203  * @sq: the service_queue being reported
0204  * @fmt: printf format string
0205  * @args: printf args
0206  *
0207  * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
0208  * throtl_grp; otherwise, just "throtl".
0209  */
0210 #define throtl_log(sq, fmt, args...)    do {                \
0211     struct throtl_grp *__tg = sq_to_tg((sq));           \
0212     struct throtl_data *__td = sq_to_td((sq));          \
0213                                     \
0214     (void)__td;                         \
0215     if (likely(!blk_trace_note_message_enabled(__td->queue)))   \
0216         break;                          \
0217     if ((__tg)) {                           \
0218         char __pbuf[128];                   \
0219                                     \
0220         blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf));    \
0221         blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \
0222     } else {                            \
0223         blk_add_trace_msg(__td->queue, "throtl " fmt, ##args);  \
0224     }                               \
0225 } while (0)
0226 
0227 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
0228 {
0229     INIT_LIST_HEAD(&qn->node);
0230     bio_list_init(&qn->bios);
0231     qn->tg = tg;
0232 }
0233 
0234 /**
0235  * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
0236  * @bio: bio being added
0237  * @qn: qnode to add bio to
0238  * @queued: the service_queue->queued[] list @qn belongs to
0239  *
0240  * Add @bio to @qn and put @qn on @queued if it's not already on.
0241  * @qn->tg's reference count is bumped when @qn is activated.  See the
0242  * comment on top of throtl_qnode definition for details.
0243  */
0244 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
0245                  struct list_head *queued)
0246 {
0247     bio_list_add(&qn->bios, bio);
0248     if (list_empty(&qn->node)) {
0249         list_add_tail(&qn->node, queued);
0250         blkg_get(tg_to_blkg(qn->tg));
0251     }
0252 }
0253 
0254 /**
0255  * throtl_peek_queued - peek the first bio on a qnode list
0256  * @queued: the qnode list to peek
0257  */
0258 static struct bio *throtl_peek_queued(struct list_head *queued)
0259 {
0260     struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
0261     struct bio *bio;
0262 
0263     if (list_empty(queued))
0264         return NULL;
0265 
0266     bio = bio_list_peek(&qn->bios);
0267     WARN_ON_ONCE(!bio);
0268     return bio;
0269 }
0270 
0271 /**
0272  * throtl_pop_queued - pop the first bio form a qnode list
0273  * @queued: the qnode list to pop a bio from
0274  * @tg_to_put: optional out argument for throtl_grp to put
0275  *
0276  * Pop the first bio from the qnode list @queued.  After popping, the first
0277  * qnode is removed from @queued if empty or moved to the end of @queued so
0278  * that the popping order is round-robin.
0279  *
0280  * When the first qnode is removed, its associated throtl_grp should be put
0281  * too.  If @tg_to_put is NULL, this function automatically puts it;
0282  * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
0283  * responsible for putting it.
0284  */
0285 static struct bio *throtl_pop_queued(struct list_head *queued,
0286                      struct throtl_grp **tg_to_put)
0287 {
0288     struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
0289     struct bio *bio;
0290 
0291     if (list_empty(queued))
0292         return NULL;
0293 
0294     bio = bio_list_pop(&qn->bios);
0295     WARN_ON_ONCE(!bio);
0296 
0297     if (bio_list_empty(&qn->bios)) {
0298         list_del_init(&qn->node);
0299         if (tg_to_put)
0300             *tg_to_put = qn->tg;
0301         else
0302             blkg_put(tg_to_blkg(qn->tg));
0303     } else {
0304         list_move_tail(&qn->node, queued);
0305     }
0306 
0307     return bio;
0308 }
0309 
0310 /* init a service_queue, assumes the caller zeroed it */
0311 static void throtl_service_queue_init(struct throtl_service_queue *sq)
0312 {
0313     INIT_LIST_HEAD(&sq->queued[0]);
0314     INIT_LIST_HEAD(&sq->queued[1]);
0315     sq->pending_tree = RB_ROOT;
0316     setup_timer(&sq->pending_timer, throtl_pending_timer_fn,
0317             (unsigned long)sq);
0318 }
0319 
0320 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node)
0321 {
0322     struct throtl_grp *tg;
0323     int rw;
0324 
0325     tg = kzalloc_node(sizeof(*tg), gfp, node);
0326     if (!tg)
0327         return NULL;
0328 
0329     throtl_service_queue_init(&tg->service_queue);
0330 
0331     for (rw = READ; rw <= WRITE; rw++) {
0332         throtl_qnode_init(&tg->qnode_on_self[rw], tg);
0333         throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
0334     }
0335 
0336     RB_CLEAR_NODE(&tg->rb_node);
0337     tg->bps[READ] = -1;
0338     tg->bps[WRITE] = -1;
0339     tg->iops[READ] = -1;
0340     tg->iops[WRITE] = -1;
0341 
0342     return &tg->pd;
0343 }
0344 
0345 static void throtl_pd_init(struct blkg_policy_data *pd)
0346 {
0347     struct throtl_grp *tg = pd_to_tg(pd);
0348     struct blkcg_gq *blkg = tg_to_blkg(tg);
0349     struct throtl_data *td = blkg->q->td;
0350     struct throtl_service_queue *sq = &tg->service_queue;
0351 
0352     /*
0353      * If on the default hierarchy, we switch to properly hierarchical
0354      * behavior where limits on a given throtl_grp are applied to the
0355      * whole subtree rather than just the group itself.  e.g. If 16M
0356      * read_bps limit is set on the root group, the whole system can't
0357      * exceed 16M for the device.
0358      *
0359      * If not on the default hierarchy, the broken flat hierarchy
0360      * behavior is retained where all throtl_grps are treated as if
0361      * they're all separate root groups right below throtl_data.
0362      * Limits of a group don't interact with limits of other groups
0363      * regardless of the position of the group in the hierarchy.
0364      */
0365     sq->parent_sq = &td->service_queue;
0366     if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
0367         sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
0368     tg->td = td;
0369 }
0370 
0371 /*
0372  * Set has_rules[] if @tg or any of its parents have limits configured.
0373  * This doesn't require walking up to the top of the hierarchy as the
0374  * parent's has_rules[] is guaranteed to be correct.
0375  */
0376 static void tg_update_has_rules(struct throtl_grp *tg)
0377 {
0378     struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
0379     int rw;
0380 
0381     for (rw = READ; rw <= WRITE; rw++)
0382         tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
0383                     (tg->bps[rw] != -1 || tg->iops[rw] != -1);
0384 }
0385 
0386 static void throtl_pd_online(struct blkg_policy_data *pd)
0387 {
0388     /*
0389      * We don't want new groups to escape the limits of its ancestors.
0390      * Update has_rules[] after a new group is brought online.
0391      */
0392     tg_update_has_rules(pd_to_tg(pd));
0393 }
0394 
0395 static void throtl_pd_free(struct blkg_policy_data *pd)
0396 {
0397     struct throtl_grp *tg = pd_to_tg(pd);
0398 
0399     del_timer_sync(&tg->service_queue.pending_timer);
0400     kfree(tg);
0401 }
0402 
0403 static struct throtl_grp *
0404 throtl_rb_first(struct throtl_service_queue *parent_sq)
0405 {
0406     /* Service tree is empty */
0407     if (!parent_sq->nr_pending)
0408         return NULL;
0409 
0410     if (!parent_sq->first_pending)
0411         parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
0412 
0413     if (parent_sq->first_pending)
0414         return rb_entry_tg(parent_sq->first_pending);
0415 
0416     return NULL;
0417 }
0418 
0419 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
0420 {
0421     rb_erase(n, root);
0422     RB_CLEAR_NODE(n);
0423 }
0424 
0425 static void throtl_rb_erase(struct rb_node *n,
0426                 struct throtl_service_queue *parent_sq)
0427 {
0428     if (parent_sq->first_pending == n)
0429         parent_sq->first_pending = NULL;
0430     rb_erase_init(n, &parent_sq->pending_tree);
0431     --parent_sq->nr_pending;
0432 }
0433 
0434 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
0435 {
0436     struct throtl_grp *tg;
0437 
0438     tg = throtl_rb_first(parent_sq);
0439     if (!tg)
0440         return;
0441 
0442     parent_sq->first_pending_disptime = tg->disptime;
0443 }
0444 
0445 static void tg_service_queue_add(struct throtl_grp *tg)
0446 {
0447     struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
0448     struct rb_node **node = &parent_sq->pending_tree.rb_node;
0449     struct rb_node *parent = NULL;
0450     struct throtl_grp *__tg;
0451     unsigned long key = tg->disptime;
0452     int left = 1;
0453 
0454     while (*node != NULL) {
0455         parent = *node;
0456         __tg = rb_entry_tg(parent);
0457 
0458         if (time_before(key, __tg->disptime))
0459             node = &parent->rb_left;
0460         else {
0461             node = &parent->rb_right;
0462             left = 0;
0463         }
0464     }
0465 
0466     if (left)
0467         parent_sq->first_pending = &tg->rb_node;
0468 
0469     rb_link_node(&tg->rb_node, parent, node);
0470     rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
0471 }
0472 
0473 static void __throtl_enqueue_tg(struct throtl_grp *tg)
0474 {
0475     tg_service_queue_add(tg);
0476     tg->flags |= THROTL_TG_PENDING;
0477     tg->service_queue.parent_sq->nr_pending++;
0478 }
0479 
0480 static void throtl_enqueue_tg(struct throtl_grp *tg)
0481 {
0482     if (!(tg->flags & THROTL_TG_PENDING))
0483         __throtl_enqueue_tg(tg);
0484 }
0485 
0486 static void __throtl_dequeue_tg(struct throtl_grp *tg)
0487 {
0488     throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
0489     tg->flags &= ~THROTL_TG_PENDING;
0490 }
0491 
0492 static void throtl_dequeue_tg(struct throtl_grp *tg)
0493 {
0494     if (tg->flags & THROTL_TG_PENDING)
0495         __throtl_dequeue_tg(tg);
0496 }
0497 
0498 /* Call with queue lock held */
0499 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
0500                       unsigned long expires)
0501 {
0502     mod_timer(&sq->pending_timer, expires);
0503     throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
0504            expires - jiffies, jiffies);
0505 }
0506 
0507 /**
0508  * throtl_schedule_next_dispatch - schedule the next dispatch cycle
0509  * @sq: the service_queue to schedule dispatch for
0510  * @force: force scheduling
0511  *
0512  * Arm @sq->pending_timer so that the next dispatch cycle starts on the
0513  * dispatch time of the first pending child.  Returns %true if either timer
0514  * is armed or there's no pending child left.  %false if the current
0515  * dispatch window is still open and the caller should continue
0516  * dispatching.
0517  *
0518  * If @force is %true, the dispatch timer is always scheduled and this
0519  * function is guaranteed to return %true.  This is to be used when the
0520  * caller can't dispatch itself and needs to invoke pending_timer
0521  * unconditionally.  Note that forced scheduling is likely to induce short
0522  * delay before dispatch starts even if @sq->first_pending_disptime is not
0523  * in the future and thus shouldn't be used in hot paths.
0524  */
0525 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
0526                       bool force)
0527 {
0528     /* any pending children left? */
0529     if (!sq->nr_pending)
0530         return true;
0531 
0532     update_min_dispatch_time(sq);
0533 
0534     /* is the next dispatch time in the future? */
0535     if (force || time_after(sq->first_pending_disptime, jiffies)) {
0536         throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
0537         return true;
0538     }
0539 
0540     /* tell the caller to continue dispatching */
0541     return false;
0542 }
0543 
0544 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
0545         bool rw, unsigned long start)
0546 {
0547     tg->bytes_disp[rw] = 0;
0548     tg->io_disp[rw] = 0;
0549 
0550     /*
0551      * Previous slice has expired. We must have trimmed it after last
0552      * bio dispatch. That means since start of last slice, we never used
0553      * that bandwidth. Do try to make use of that bandwidth while giving
0554      * credit.
0555      */
0556     if (time_after_eq(start, tg->slice_start[rw]))
0557         tg->slice_start[rw] = start;
0558 
0559     tg->slice_end[rw] = jiffies + throtl_slice;
0560     throtl_log(&tg->service_queue,
0561            "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
0562            rw == READ ? 'R' : 'W', tg->slice_start[rw],
0563            tg->slice_end[rw], jiffies);
0564 }
0565 
0566 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
0567 {
0568     tg->bytes_disp[rw] = 0;
0569     tg->io_disp[rw] = 0;
0570     tg->slice_start[rw] = jiffies;
0571     tg->slice_end[rw] = jiffies + throtl_slice;
0572     throtl_log(&tg->service_queue,
0573            "[%c] new slice start=%lu end=%lu jiffies=%lu",
0574            rw == READ ? 'R' : 'W', tg->slice_start[rw],
0575            tg->slice_end[rw], jiffies);
0576 }
0577 
0578 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
0579                     unsigned long jiffy_end)
0580 {
0581     tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
0582 }
0583 
0584 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
0585                        unsigned long jiffy_end)
0586 {
0587     tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
0588     throtl_log(&tg->service_queue,
0589            "[%c] extend slice start=%lu end=%lu jiffies=%lu",
0590            rw == READ ? 'R' : 'W', tg->slice_start[rw],
0591            tg->slice_end[rw], jiffies);
0592 }
0593 
0594 /* Determine if previously allocated or extended slice is complete or not */
0595 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
0596 {
0597     if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
0598         return false;
0599 
0600     return 1;
0601 }
0602 
0603 /* Trim the used slices and adjust slice start accordingly */
0604 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
0605 {
0606     unsigned long nr_slices, time_elapsed, io_trim;
0607     u64 bytes_trim, tmp;
0608 
0609     BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
0610 
0611     /*
0612      * If bps are unlimited (-1), then time slice don't get
0613      * renewed. Don't try to trim the slice if slice is used. A new
0614      * slice will start when appropriate.
0615      */
0616     if (throtl_slice_used(tg, rw))
0617         return;
0618 
0619     /*
0620      * A bio has been dispatched. Also adjust slice_end. It might happen
0621      * that initially cgroup limit was very low resulting in high
0622      * slice_end, but later limit was bumped up and bio was dispached
0623      * sooner, then we need to reduce slice_end. A high bogus slice_end
0624      * is bad because it does not allow new slice to start.
0625      */
0626 
0627     throtl_set_slice_end(tg, rw, jiffies + throtl_slice);
0628 
0629     time_elapsed = jiffies - tg->slice_start[rw];
0630 
0631     nr_slices = time_elapsed / throtl_slice;
0632 
0633     if (!nr_slices)
0634         return;
0635     tmp = tg->bps[rw] * throtl_slice * nr_slices;
0636     do_div(tmp, HZ);
0637     bytes_trim = tmp;
0638 
0639     io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
0640 
0641     if (!bytes_trim && !io_trim)
0642         return;
0643 
0644     if (tg->bytes_disp[rw] >= bytes_trim)
0645         tg->bytes_disp[rw] -= bytes_trim;
0646     else
0647         tg->bytes_disp[rw] = 0;
0648 
0649     if (tg->io_disp[rw] >= io_trim)
0650         tg->io_disp[rw] -= io_trim;
0651     else
0652         tg->io_disp[rw] = 0;
0653 
0654     tg->slice_start[rw] += nr_slices * throtl_slice;
0655 
0656     throtl_log(&tg->service_queue,
0657            "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
0658            rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
0659            tg->slice_start[rw], tg->slice_end[rw], jiffies);
0660 }
0661 
0662 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
0663                   unsigned long *wait)
0664 {
0665     bool rw = bio_data_dir(bio);
0666     unsigned int io_allowed;
0667     unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
0668     u64 tmp;
0669 
0670     jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
0671 
0672     /* Slice has just started. Consider one slice interval */
0673     if (!jiffy_elapsed)
0674         jiffy_elapsed_rnd = throtl_slice;
0675 
0676     jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
0677 
0678     /*
0679      * jiffy_elapsed_rnd should not be a big value as minimum iops can be
0680      * 1 then at max jiffy elapsed should be equivalent of 1 second as we
0681      * will allow dispatch after 1 second and after that slice should
0682      * have been trimmed.
0683      */
0684 
0685     tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
0686     do_div(tmp, HZ);
0687 
0688     if (tmp > UINT_MAX)
0689         io_allowed = UINT_MAX;
0690     else
0691         io_allowed = tmp;
0692 
0693     if (tg->io_disp[rw] + 1 <= io_allowed) {
0694         if (wait)
0695             *wait = 0;
0696         return true;
0697     }
0698 
0699     /* Calc approx time to dispatch */
0700     jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
0701 
0702     if (jiffy_wait > jiffy_elapsed)
0703         jiffy_wait = jiffy_wait - jiffy_elapsed;
0704     else
0705         jiffy_wait = 1;
0706 
0707     if (wait)
0708         *wait = jiffy_wait;
0709     return 0;
0710 }
0711 
0712 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
0713                  unsigned long *wait)
0714 {
0715     bool rw = bio_data_dir(bio);
0716     u64 bytes_allowed, extra_bytes, tmp;
0717     unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
0718 
0719     jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
0720 
0721     /* Slice has just started. Consider one slice interval */
0722     if (!jiffy_elapsed)
0723         jiffy_elapsed_rnd = throtl_slice;
0724 
0725     jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
0726 
0727     tmp = tg->bps[rw] * jiffy_elapsed_rnd;
0728     do_div(tmp, HZ);
0729     bytes_allowed = tmp;
0730 
0731     if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) {
0732         if (wait)
0733             *wait = 0;
0734         return true;
0735     }
0736 
0737     /* Calc approx time to dispatch */
0738     extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed;
0739     jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
0740 
0741     if (!jiffy_wait)
0742         jiffy_wait = 1;
0743 
0744     /*
0745      * This wait time is without taking into consideration the rounding
0746      * up we did. Add that time also.
0747      */
0748     jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
0749     if (wait)
0750         *wait = jiffy_wait;
0751     return 0;
0752 }
0753 
0754 /*
0755  * Returns whether one can dispatch a bio or not. Also returns approx number
0756  * of jiffies to wait before this bio is with-in IO rate and can be dispatched
0757  */
0758 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
0759                 unsigned long *wait)
0760 {
0761     bool rw = bio_data_dir(bio);
0762     unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
0763 
0764     /*
0765      * Currently whole state machine of group depends on first bio
0766      * queued in the group bio list. So one should not be calling
0767      * this function with a different bio if there are other bios
0768      * queued.
0769      */
0770     BUG_ON(tg->service_queue.nr_queued[rw] &&
0771            bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
0772 
0773     /* If tg->bps = -1, then BW is unlimited */
0774     if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
0775         if (wait)
0776             *wait = 0;
0777         return true;
0778     }
0779 
0780     /*
0781      * If previous slice expired, start a new one otherwise renew/extend
0782      * existing slice to make sure it is at least throtl_slice interval
0783      * long since now. New slice is started only for empty throttle group.
0784      * If there is queued bio, that means there should be an active
0785      * slice and it should be extended instead.
0786      */
0787     if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
0788         throtl_start_new_slice(tg, rw);
0789     else {
0790         if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
0791             throtl_extend_slice(tg, rw, jiffies + throtl_slice);
0792     }
0793 
0794     if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
0795         tg_with_in_iops_limit(tg, bio, &iops_wait)) {
0796         if (wait)
0797             *wait = 0;
0798         return 1;
0799     }
0800 
0801     max_wait = max(bps_wait, iops_wait);
0802 
0803     if (wait)
0804         *wait = max_wait;
0805 
0806     if (time_before(tg->slice_end[rw], jiffies + max_wait))
0807         throtl_extend_slice(tg, rw, jiffies + max_wait);
0808 
0809     return 0;
0810 }
0811 
0812 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
0813 {
0814     bool rw = bio_data_dir(bio);
0815 
0816     /* Charge the bio to the group */
0817     tg->bytes_disp[rw] += bio->bi_iter.bi_size;
0818     tg->io_disp[rw]++;
0819 
0820     /*
0821      * BIO_THROTTLED is used to prevent the same bio to be throttled
0822      * more than once as a throttled bio will go through blk-throtl the
0823      * second time when it eventually gets issued.  Set it when a bio
0824      * is being charged to a tg.
0825      */
0826     if (!bio_flagged(bio, BIO_THROTTLED))
0827         bio_set_flag(bio, BIO_THROTTLED);
0828 }
0829 
0830 /**
0831  * throtl_add_bio_tg - add a bio to the specified throtl_grp
0832  * @bio: bio to add
0833  * @qn: qnode to use
0834  * @tg: the target throtl_grp
0835  *
0836  * Add @bio to @tg's service_queue using @qn.  If @qn is not specified,
0837  * tg->qnode_on_self[] is used.
0838  */
0839 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
0840                   struct throtl_grp *tg)
0841 {
0842     struct throtl_service_queue *sq = &tg->service_queue;
0843     bool rw = bio_data_dir(bio);
0844 
0845     if (!qn)
0846         qn = &tg->qnode_on_self[rw];
0847 
0848     /*
0849      * If @tg doesn't currently have any bios queued in the same
0850      * direction, queueing @bio can change when @tg should be
0851      * dispatched.  Mark that @tg was empty.  This is automatically
0852      * cleaered on the next tg_update_disptime().
0853      */
0854     if (!sq->nr_queued[rw])
0855         tg->flags |= THROTL_TG_WAS_EMPTY;
0856 
0857     throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
0858 
0859     sq->nr_queued[rw]++;
0860     throtl_enqueue_tg(tg);
0861 }
0862 
0863 static void tg_update_disptime(struct throtl_grp *tg)
0864 {
0865     struct throtl_service_queue *sq = &tg->service_queue;
0866     unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
0867     struct bio *bio;
0868 
0869     if ((bio = throtl_peek_queued(&sq->queued[READ])))
0870         tg_may_dispatch(tg, bio, &read_wait);
0871 
0872     if ((bio = throtl_peek_queued(&sq->queued[WRITE])))
0873         tg_may_dispatch(tg, bio, &write_wait);
0874 
0875     min_wait = min(read_wait, write_wait);
0876     disptime = jiffies + min_wait;
0877 
0878     /* Update dispatch time */
0879     throtl_dequeue_tg(tg);
0880     tg->disptime = disptime;
0881     throtl_enqueue_tg(tg);
0882 
0883     /* see throtl_add_bio_tg() */
0884     tg->flags &= ~THROTL_TG_WAS_EMPTY;
0885 }
0886 
0887 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
0888                     struct throtl_grp *parent_tg, bool rw)
0889 {
0890     if (throtl_slice_used(parent_tg, rw)) {
0891         throtl_start_new_slice_with_credit(parent_tg, rw,
0892                 child_tg->slice_start[rw]);
0893     }
0894 
0895 }
0896 
0897 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
0898 {
0899     struct throtl_service_queue *sq = &tg->service_queue;
0900     struct throtl_service_queue *parent_sq = sq->parent_sq;
0901     struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
0902     struct throtl_grp *tg_to_put = NULL;
0903     struct bio *bio;
0904 
0905     /*
0906      * @bio is being transferred from @tg to @parent_sq.  Popping a bio
0907      * from @tg may put its reference and @parent_sq might end up
0908      * getting released prematurely.  Remember the tg to put and put it
0909      * after @bio is transferred to @parent_sq.
0910      */
0911     bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
0912     sq->nr_queued[rw]--;
0913 
0914     throtl_charge_bio(tg, bio);
0915 
0916     /*
0917      * If our parent is another tg, we just need to transfer @bio to
0918      * the parent using throtl_add_bio_tg().  If our parent is
0919      * @td->service_queue, @bio is ready to be issued.  Put it on its
0920      * bio_lists[] and decrease total number queued.  The caller is
0921      * responsible for issuing these bios.
0922      */
0923     if (parent_tg) {
0924         throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
0925         start_parent_slice_with_credit(tg, parent_tg, rw);
0926     } else {
0927         throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
0928                      &parent_sq->queued[rw]);
0929         BUG_ON(tg->td->nr_queued[rw] <= 0);
0930         tg->td->nr_queued[rw]--;
0931     }
0932 
0933     throtl_trim_slice(tg, rw);
0934 
0935     if (tg_to_put)
0936         blkg_put(tg_to_blkg(tg_to_put));
0937 }
0938 
0939 static int throtl_dispatch_tg(struct throtl_grp *tg)
0940 {
0941     struct throtl_service_queue *sq = &tg->service_queue;
0942     unsigned int nr_reads = 0, nr_writes = 0;
0943     unsigned int max_nr_reads = throtl_grp_quantum*3/4;
0944     unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
0945     struct bio *bio;
0946 
0947     /* Try to dispatch 75% READS and 25% WRITES */
0948 
0949     while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
0950            tg_may_dispatch(tg, bio, NULL)) {
0951 
0952         tg_dispatch_one_bio(tg, bio_data_dir(bio));
0953         nr_reads++;
0954 
0955         if (nr_reads >= max_nr_reads)
0956             break;
0957     }
0958 
0959     while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
0960            tg_may_dispatch(tg, bio, NULL)) {
0961 
0962         tg_dispatch_one_bio(tg, bio_data_dir(bio));
0963         nr_writes++;
0964 
0965         if (nr_writes >= max_nr_writes)
0966             break;
0967     }
0968 
0969     return nr_reads + nr_writes;
0970 }
0971 
0972 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
0973 {
0974     unsigned int nr_disp = 0;
0975 
0976     while (1) {
0977         struct throtl_grp *tg = throtl_rb_first(parent_sq);
0978         struct throtl_service_queue *sq = &tg->service_queue;
0979 
0980         if (!tg)
0981             break;
0982 
0983         if (time_before(jiffies, tg->disptime))
0984             break;
0985 
0986         throtl_dequeue_tg(tg);
0987 
0988         nr_disp += throtl_dispatch_tg(tg);
0989 
0990         if (sq->nr_queued[0] || sq->nr_queued[1])
0991             tg_update_disptime(tg);
0992 
0993         if (nr_disp >= throtl_quantum)
0994             break;
0995     }
0996 
0997     return nr_disp;
0998 }
0999 
1000 /**
1001  * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1002  * @arg: the throtl_service_queue being serviced
1003  *
1004  * This timer is armed when a child throtl_grp with active bio's become
1005  * pending and queued on the service_queue's pending_tree and expires when
1006  * the first child throtl_grp should be dispatched.  This function
1007  * dispatches bio's from the children throtl_grps to the parent
1008  * service_queue.
1009  *
1010  * If the parent's parent is another throtl_grp, dispatching is propagated
1011  * by either arming its pending_timer or repeating dispatch directly.  If
1012  * the top-level service_tree is reached, throtl_data->dispatch_work is
1013  * kicked so that the ready bio's are issued.
1014  */
1015 static void throtl_pending_timer_fn(unsigned long arg)
1016 {
1017     struct throtl_service_queue *sq = (void *)arg;
1018     struct throtl_grp *tg = sq_to_tg(sq);
1019     struct throtl_data *td = sq_to_td(sq);
1020     struct request_queue *q = td->queue;
1021     struct throtl_service_queue *parent_sq;
1022     bool dispatched;
1023     int ret;
1024 
1025     spin_lock_irq(q->queue_lock);
1026 again:
1027     parent_sq = sq->parent_sq;
1028     dispatched = false;
1029 
1030     while (true) {
1031         throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1032                sq->nr_queued[READ] + sq->nr_queued[WRITE],
1033                sq->nr_queued[READ], sq->nr_queued[WRITE]);
1034 
1035         ret = throtl_select_dispatch(sq);
1036         if (ret) {
1037             throtl_log(sq, "bios disp=%u", ret);
1038             dispatched = true;
1039         }
1040 
1041         if (throtl_schedule_next_dispatch(sq, false))
1042             break;
1043 
1044         /* this dispatch windows is still open, relax and repeat */
1045         spin_unlock_irq(q->queue_lock);
1046         cpu_relax();
1047         spin_lock_irq(q->queue_lock);
1048     }
1049 
1050     if (!dispatched)
1051         goto out_unlock;
1052 
1053     if (parent_sq) {
1054         /* @parent_sq is another throl_grp, propagate dispatch */
1055         if (tg->flags & THROTL_TG_WAS_EMPTY) {
1056             tg_update_disptime(tg);
1057             if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1058                 /* window is already open, repeat dispatching */
1059                 sq = parent_sq;
1060                 tg = sq_to_tg(sq);
1061                 goto again;
1062             }
1063         }
1064     } else {
1065         /* reached the top-level, queue issueing */
1066         queue_work(kthrotld_workqueue, &td->dispatch_work);
1067     }
1068 out_unlock:
1069     spin_unlock_irq(q->queue_lock);
1070 }
1071 
1072 /**
1073  * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1074  * @work: work item being executed
1075  *
1076  * This function is queued for execution when bio's reach the bio_lists[]
1077  * of throtl_data->service_queue.  Those bio's are ready and issued by this
1078  * function.
1079  */
1080 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1081 {
1082     struct throtl_data *td = container_of(work, struct throtl_data,
1083                           dispatch_work);
1084     struct throtl_service_queue *td_sq = &td->service_queue;
1085     struct request_queue *q = td->queue;
1086     struct bio_list bio_list_on_stack;
1087     struct bio *bio;
1088     struct blk_plug plug;
1089     int rw;
1090 
1091     bio_list_init(&bio_list_on_stack);
1092 
1093     spin_lock_irq(q->queue_lock);
1094     for (rw = READ; rw <= WRITE; rw++)
1095         while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1096             bio_list_add(&bio_list_on_stack, bio);
1097     spin_unlock_irq(q->queue_lock);
1098 
1099     if (!bio_list_empty(&bio_list_on_stack)) {
1100         blk_start_plug(&plug);
1101         while((bio = bio_list_pop(&bio_list_on_stack)))
1102             generic_make_request(bio);
1103         blk_finish_plug(&plug);
1104     }
1105 }
1106 
1107 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1108                   int off)
1109 {
1110     struct throtl_grp *tg = pd_to_tg(pd);
1111     u64 v = *(u64 *)((void *)tg + off);
1112 
1113     if (v == -1)
1114         return 0;
1115     return __blkg_prfill_u64(sf, pd, v);
1116 }
1117 
1118 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1119                    int off)
1120 {
1121     struct throtl_grp *tg = pd_to_tg(pd);
1122     unsigned int v = *(unsigned int *)((void *)tg + off);
1123 
1124     if (v == -1)
1125         return 0;
1126     return __blkg_prfill_u64(sf, pd, v);
1127 }
1128 
1129 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1130 {
1131     blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1132               &blkcg_policy_throtl, seq_cft(sf)->private, false);
1133     return 0;
1134 }
1135 
1136 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1137 {
1138     blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1139               &blkcg_policy_throtl, seq_cft(sf)->private, false);
1140     return 0;
1141 }
1142 
1143 static void tg_conf_updated(struct throtl_grp *tg)
1144 {
1145     struct throtl_service_queue *sq = &tg->service_queue;
1146     struct cgroup_subsys_state *pos_css;
1147     struct blkcg_gq *blkg;
1148 
1149     throtl_log(&tg->service_queue,
1150            "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1151            tg->bps[READ], tg->bps[WRITE],
1152            tg->iops[READ], tg->iops[WRITE]);
1153 
1154     /*
1155      * Update has_rules[] flags for the updated tg's subtree.  A tg is
1156      * considered to have rules if either the tg itself or any of its
1157      * ancestors has rules.  This identifies groups without any
1158      * restrictions in the whole hierarchy and allows them to bypass
1159      * blk-throttle.
1160      */
1161     blkg_for_each_descendant_pre(blkg, pos_css, tg_to_blkg(tg))
1162         tg_update_has_rules(blkg_to_tg(blkg));
1163 
1164     /*
1165      * We're already holding queue_lock and know @tg is valid.  Let's
1166      * apply the new config directly.
1167      *
1168      * Restart the slices for both READ and WRITES. It might happen
1169      * that a group's limit are dropped suddenly and we don't want to
1170      * account recently dispatched IO with new low rate.
1171      */
1172     throtl_start_new_slice(tg, 0);
1173     throtl_start_new_slice(tg, 1);
1174 
1175     if (tg->flags & THROTL_TG_PENDING) {
1176         tg_update_disptime(tg);
1177         throtl_schedule_next_dispatch(sq->parent_sq, true);
1178     }
1179 }
1180 
1181 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1182                char *buf, size_t nbytes, loff_t off, bool is_u64)
1183 {
1184     struct blkcg *blkcg = css_to_blkcg(of_css(of));
1185     struct blkg_conf_ctx ctx;
1186     struct throtl_grp *tg;
1187     int ret;
1188     u64 v;
1189 
1190     ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1191     if (ret)
1192         return ret;
1193 
1194     ret = -EINVAL;
1195     if (sscanf(ctx.body, "%llu", &v) != 1)
1196         goto out_finish;
1197     if (!v)
1198         v = -1;
1199 
1200     tg = blkg_to_tg(ctx.blkg);
1201 
1202     if (is_u64)
1203         *(u64 *)((void *)tg + of_cft(of)->private) = v;
1204     else
1205         *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1206 
1207     tg_conf_updated(tg);
1208     ret = 0;
1209 out_finish:
1210     blkg_conf_finish(&ctx);
1211     return ret ?: nbytes;
1212 }
1213 
1214 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1215                    char *buf, size_t nbytes, loff_t off)
1216 {
1217     return tg_set_conf(of, buf, nbytes, off, true);
1218 }
1219 
1220 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1221                 char *buf, size_t nbytes, loff_t off)
1222 {
1223     return tg_set_conf(of, buf, nbytes, off, false);
1224 }
1225 
1226 static struct cftype throtl_legacy_files[] = {
1227     {
1228         .name = "throttle.read_bps_device",
1229         .private = offsetof(struct throtl_grp, bps[READ]),
1230         .seq_show = tg_print_conf_u64,
1231         .write = tg_set_conf_u64,
1232     },
1233     {
1234         .name = "throttle.write_bps_device",
1235         .private = offsetof(struct throtl_grp, bps[WRITE]),
1236         .seq_show = tg_print_conf_u64,
1237         .write = tg_set_conf_u64,
1238     },
1239     {
1240         .name = "throttle.read_iops_device",
1241         .private = offsetof(struct throtl_grp, iops[READ]),
1242         .seq_show = tg_print_conf_uint,
1243         .write = tg_set_conf_uint,
1244     },
1245     {
1246         .name = "throttle.write_iops_device",
1247         .private = offsetof(struct throtl_grp, iops[WRITE]),
1248         .seq_show = tg_print_conf_uint,
1249         .write = tg_set_conf_uint,
1250     },
1251     {
1252         .name = "throttle.io_service_bytes",
1253         .private = (unsigned long)&blkcg_policy_throtl,
1254         .seq_show = blkg_print_stat_bytes,
1255     },
1256     {
1257         .name = "throttle.io_serviced",
1258         .private = (unsigned long)&blkcg_policy_throtl,
1259         .seq_show = blkg_print_stat_ios,
1260     },
1261     { } /* terminate */
1262 };
1263 
1264 static u64 tg_prfill_max(struct seq_file *sf, struct blkg_policy_data *pd,
1265              int off)
1266 {
1267     struct throtl_grp *tg = pd_to_tg(pd);
1268     const char *dname = blkg_dev_name(pd->blkg);
1269     char bufs[4][21] = { "max", "max", "max", "max" };
1270 
1271     if (!dname)
1272         return 0;
1273     if (tg->bps[READ] == -1 && tg->bps[WRITE] == -1 &&
1274         tg->iops[READ] == -1 && tg->iops[WRITE] == -1)
1275         return 0;
1276 
1277     if (tg->bps[READ] != -1)
1278         snprintf(bufs[0], sizeof(bufs[0]), "%llu", tg->bps[READ]);
1279     if (tg->bps[WRITE] != -1)
1280         snprintf(bufs[1], sizeof(bufs[1]), "%llu", tg->bps[WRITE]);
1281     if (tg->iops[READ] != -1)
1282         snprintf(bufs[2], sizeof(bufs[2]), "%u", tg->iops[READ]);
1283     if (tg->iops[WRITE] != -1)
1284         snprintf(bufs[3], sizeof(bufs[3]), "%u", tg->iops[WRITE]);
1285 
1286     seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s\n",
1287            dname, bufs[0], bufs[1], bufs[2], bufs[3]);
1288     return 0;
1289 }
1290 
1291 static int tg_print_max(struct seq_file *sf, void *v)
1292 {
1293     blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_max,
1294               &blkcg_policy_throtl, seq_cft(sf)->private, false);
1295     return 0;
1296 }
1297 
1298 static ssize_t tg_set_max(struct kernfs_open_file *of,
1299               char *buf, size_t nbytes, loff_t off)
1300 {
1301     struct blkcg *blkcg = css_to_blkcg(of_css(of));
1302     struct blkg_conf_ctx ctx;
1303     struct throtl_grp *tg;
1304     u64 v[4];
1305     int ret;
1306 
1307     ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1308     if (ret)
1309         return ret;
1310 
1311     tg = blkg_to_tg(ctx.blkg);
1312 
1313     v[0] = tg->bps[READ];
1314     v[1] = tg->bps[WRITE];
1315     v[2] = tg->iops[READ];
1316     v[3] = tg->iops[WRITE];
1317 
1318     while (true) {
1319         char tok[27];   /* wiops=18446744073709551616 */
1320         char *p;
1321         u64 val = -1;
1322         int len;
1323 
1324         if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1325             break;
1326         if (tok[0] == '\0')
1327             break;
1328         ctx.body += len;
1329 
1330         ret = -EINVAL;
1331         p = tok;
1332         strsep(&p, "=");
1333         if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1334             goto out_finish;
1335 
1336         ret = -ERANGE;
1337         if (!val)
1338             goto out_finish;
1339 
1340         ret = -EINVAL;
1341         if (!strcmp(tok, "rbps"))
1342             v[0] = val;
1343         else if (!strcmp(tok, "wbps"))
1344             v[1] = val;
1345         else if (!strcmp(tok, "riops"))
1346             v[2] = min_t(u64, val, UINT_MAX);
1347         else if (!strcmp(tok, "wiops"))
1348             v[3] = min_t(u64, val, UINT_MAX);
1349         else
1350             goto out_finish;
1351     }
1352 
1353     tg->bps[READ] = v[0];
1354     tg->bps[WRITE] = v[1];
1355     tg->iops[READ] = v[2];
1356     tg->iops[WRITE] = v[3];
1357 
1358     tg_conf_updated(tg);
1359     ret = 0;
1360 out_finish:
1361     blkg_conf_finish(&ctx);
1362     return ret ?: nbytes;
1363 }
1364 
1365 static struct cftype throtl_files[] = {
1366     {
1367         .name = "max",
1368         .flags = CFTYPE_NOT_ON_ROOT,
1369         .seq_show = tg_print_max,
1370         .write = tg_set_max,
1371     },
1372     { } /* terminate */
1373 };
1374 
1375 static void throtl_shutdown_wq(struct request_queue *q)
1376 {
1377     struct throtl_data *td = q->td;
1378 
1379     cancel_work_sync(&td->dispatch_work);
1380 }
1381 
1382 static struct blkcg_policy blkcg_policy_throtl = {
1383     .dfl_cftypes        = throtl_files,
1384     .legacy_cftypes     = throtl_legacy_files,
1385 
1386     .pd_alloc_fn        = throtl_pd_alloc,
1387     .pd_init_fn     = throtl_pd_init,
1388     .pd_online_fn       = throtl_pd_online,
1389     .pd_free_fn     = throtl_pd_free,
1390 };
1391 
1392 bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg,
1393             struct bio *bio)
1394 {
1395     struct throtl_qnode *qn = NULL;
1396     struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg);
1397     struct throtl_service_queue *sq;
1398     bool rw = bio_data_dir(bio);
1399     bool throttled = false;
1400 
1401     WARN_ON_ONCE(!rcu_read_lock_held());
1402 
1403     /* see throtl_charge_bio() */
1404     if (bio_flagged(bio, BIO_THROTTLED) || !tg->has_rules[rw])
1405         goto out;
1406 
1407     spin_lock_irq(q->queue_lock);
1408 
1409     if (unlikely(blk_queue_bypass(q)))
1410         goto out_unlock;
1411 
1412     sq = &tg->service_queue;
1413 
1414     while (true) {
1415         /* throtl is FIFO - if bios are already queued, should queue */
1416         if (sq->nr_queued[rw])
1417             break;
1418 
1419         /* if above limits, break to queue */
1420         if (!tg_may_dispatch(tg, bio, NULL))
1421             break;
1422 
1423         /* within limits, let's charge and dispatch directly */
1424         throtl_charge_bio(tg, bio);
1425 
1426         /*
1427          * We need to trim slice even when bios are not being queued
1428          * otherwise it might happen that a bio is not queued for
1429          * a long time and slice keeps on extending and trim is not
1430          * called for a long time. Now if limits are reduced suddenly
1431          * we take into account all the IO dispatched so far at new
1432          * low rate and * newly queued IO gets a really long dispatch
1433          * time.
1434          *
1435          * So keep on trimming slice even if bio is not queued.
1436          */
1437         throtl_trim_slice(tg, rw);
1438 
1439         /*
1440          * @bio passed through this layer without being throttled.
1441          * Climb up the ladder.  If we''re already at the top, it
1442          * can be executed directly.
1443          */
1444         qn = &tg->qnode_on_parent[rw];
1445         sq = sq->parent_sq;
1446         tg = sq_to_tg(sq);
1447         if (!tg)
1448             goto out_unlock;
1449     }
1450 
1451     /* out-of-limit, queue to @tg */
1452     throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1453            rw == READ ? 'R' : 'W',
1454            tg->bytes_disp[rw], bio->bi_iter.bi_size, tg->bps[rw],
1455            tg->io_disp[rw], tg->iops[rw],
1456            sq->nr_queued[READ], sq->nr_queued[WRITE]);
1457 
1458     bio_associate_current(bio);
1459     tg->td->nr_queued[rw]++;
1460     throtl_add_bio_tg(bio, qn, tg);
1461     throttled = true;
1462 
1463     /*
1464      * Update @tg's dispatch time and force schedule dispatch if @tg
1465      * was empty before @bio.  The forced scheduling isn't likely to
1466      * cause undue delay as @bio is likely to be dispatched directly if
1467      * its @tg's disptime is not in the future.
1468      */
1469     if (tg->flags & THROTL_TG_WAS_EMPTY) {
1470         tg_update_disptime(tg);
1471         throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
1472     }
1473 
1474 out_unlock:
1475     spin_unlock_irq(q->queue_lock);
1476 out:
1477     /*
1478      * As multiple blk-throtls may stack in the same issue path, we
1479      * don't want bios to leave with the flag set.  Clear the flag if
1480      * being issued.
1481      */
1482     if (!throttled)
1483         bio_clear_flag(bio, BIO_THROTTLED);
1484     return throttled;
1485 }
1486 
1487 /*
1488  * Dispatch all bios from all children tg's queued on @parent_sq.  On
1489  * return, @parent_sq is guaranteed to not have any active children tg's
1490  * and all bios from previously active tg's are on @parent_sq->bio_lists[].
1491  */
1492 static void tg_drain_bios(struct throtl_service_queue *parent_sq)
1493 {
1494     struct throtl_grp *tg;
1495 
1496     while ((tg = throtl_rb_first(parent_sq))) {
1497         struct throtl_service_queue *sq = &tg->service_queue;
1498         struct bio *bio;
1499 
1500         throtl_dequeue_tg(tg);
1501 
1502         while ((bio = throtl_peek_queued(&sq->queued[READ])))
1503             tg_dispatch_one_bio(tg, bio_data_dir(bio));
1504         while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
1505             tg_dispatch_one_bio(tg, bio_data_dir(bio));
1506     }
1507 }
1508 
1509 /**
1510  * blk_throtl_drain - drain throttled bios
1511  * @q: request_queue to drain throttled bios for
1512  *
1513  * Dispatch all currently throttled bios on @q through ->make_request_fn().
1514  */
1515 void blk_throtl_drain(struct request_queue *q)
1516     __releases(q->queue_lock) __acquires(q->queue_lock)
1517 {
1518     struct throtl_data *td = q->td;
1519     struct blkcg_gq *blkg;
1520     struct cgroup_subsys_state *pos_css;
1521     struct bio *bio;
1522     int rw;
1523 
1524     queue_lockdep_assert_held(q);
1525     rcu_read_lock();
1526 
1527     /*
1528      * Drain each tg while doing post-order walk on the blkg tree, so
1529      * that all bios are propagated to td->service_queue.  It'd be
1530      * better to walk service_queue tree directly but blkg walk is
1531      * easier.
1532      */
1533     blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
1534         tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
1535 
1536     /* finally, transfer bios from top-level tg's into the td */
1537     tg_drain_bios(&td->service_queue);
1538 
1539     rcu_read_unlock();
1540     spin_unlock_irq(q->queue_lock);
1541 
1542     /* all bios now should be in td->service_queue, issue them */
1543     for (rw = READ; rw <= WRITE; rw++)
1544         while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
1545                         NULL)))
1546             generic_make_request(bio);
1547 
1548     spin_lock_irq(q->queue_lock);
1549 }
1550 
1551 int blk_throtl_init(struct request_queue *q)
1552 {
1553     struct throtl_data *td;
1554     int ret;
1555 
1556     td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1557     if (!td)
1558         return -ENOMEM;
1559 
1560     INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
1561     throtl_service_queue_init(&td->service_queue);
1562 
1563     q->td = td;
1564     td->queue = q;
1565 
1566     /* activate policy */
1567     ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1568     if (ret)
1569         kfree(td);
1570     return ret;
1571 }
1572 
1573 void blk_throtl_exit(struct request_queue *q)
1574 {
1575     BUG_ON(!q->td);
1576     throtl_shutdown_wq(q);
1577     blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1578     kfree(q->td);
1579 }
1580 
1581 static int __init throtl_init(void)
1582 {
1583     kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1584     if (!kthrotld_workqueue)
1585         panic("Failed to create kthrotld\n");
1586 
1587     return blkcg_policy_register(&blkcg_policy_throtl);
1588 }
1589 
1590 module_init(throtl_init);