Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 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" |
Tejun Heo | bc9fcbf | 2011-10-19 14:31:18 +0200 | [diff] [blame] | 13 | #include "blk.h" |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 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 | |
Vivek Goyal | 450adcb | 2011-03-01 13:40:54 -0500 | [diff] [blame] | 24 | /* A workqueue to queue throttle related work */ |
| 25 | static struct workqueue_struct *kthrotld_workqueue; |
| 26 | static void throtl_schedule_delayed_work(struct throtl_data *td, |
| 27 | unsigned long delay); |
| 28 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 29 | struct throtl_rb_root { |
| 30 | struct rb_root rb; |
| 31 | struct rb_node *left; |
| 32 | unsigned int count; |
| 33 | unsigned long min_disptime; |
| 34 | }; |
| 35 | |
| 36 | #define THROTL_RB_ROOT (struct throtl_rb_root) { .rb = RB_ROOT, .left = NULL, \ |
| 37 | .count = 0, .min_disptime = 0} |
| 38 | |
| 39 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) |
| 40 | |
| 41 | struct throtl_grp { |
| 42 | /* List of throtl groups on the request queue*/ |
| 43 | struct hlist_node tg_node; |
| 44 | |
| 45 | /* active throtl group service_tree member */ |
| 46 | struct rb_node rb_node; |
| 47 | |
| 48 | /* |
| 49 | * Dispatch time in jiffies. This is the estimated time when group |
| 50 | * will unthrottle and is ready to dispatch more bio. It is used as |
| 51 | * key to sort active groups in service tree. |
| 52 | */ |
| 53 | unsigned long disptime; |
| 54 | |
| 55 | struct blkio_group blkg; |
| 56 | atomic_t ref; |
| 57 | unsigned int flags; |
| 58 | |
| 59 | /* Two lists for READ and WRITE */ |
| 60 | struct bio_list bio_lists[2]; |
| 61 | |
| 62 | /* Number of queued bios on READ and WRITE lists */ |
| 63 | unsigned int nr_queued[2]; |
| 64 | |
| 65 | /* bytes per second rate limits */ |
| 66 | uint64_t bps[2]; |
| 67 | |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 68 | /* IOPS limits */ |
| 69 | unsigned int iops[2]; |
| 70 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 71 | /* Number of bytes disptached in current slice */ |
| 72 | uint64_t bytes_disp[2]; |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 73 | /* Number of bio's dispatched in current slice */ |
| 74 | unsigned int io_disp[2]; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 75 | |
| 76 | /* When did we start a new slice */ |
| 77 | unsigned long slice_start[2]; |
| 78 | unsigned long slice_end[2]; |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 79 | |
| 80 | /* Some throttle limits got updated for the group */ |
Andreas Schwab | 6f03793 | 2011-03-30 12:21:56 +0200 | [diff] [blame] | 81 | int limits_changed; |
Vivek Goyal | 4843c69 | 2011-05-19 15:38:27 -0400 | [diff] [blame] | 82 | |
| 83 | struct rcu_head rcu_head; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 84 | }; |
| 85 | |
| 86 | struct throtl_data |
| 87 | { |
| 88 | /* List of throtl groups */ |
| 89 | struct hlist_head tg_list; |
| 90 | |
| 91 | /* service tree for active throtl groups */ |
| 92 | struct throtl_rb_root tg_service_tree; |
| 93 | |
Vivek Goyal | 29b1258 | 2011-05-19 15:38:24 -0400 | [diff] [blame] | 94 | struct throtl_grp *root_tg; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 95 | struct request_queue *queue; |
| 96 | |
| 97 | /* Total Number of queued bios on READ and WRITE lists */ |
| 98 | unsigned int nr_queued[2]; |
| 99 | |
| 100 | /* |
Vivek Goyal | 02977e4 | 2010-10-01 14:49:48 +0200 | [diff] [blame] | 101 | * number of total undestroyed groups |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 102 | */ |
| 103 | unsigned int nr_undestroyed_grps; |
| 104 | |
| 105 | /* Work for dispatching throttled bios */ |
| 106 | struct delayed_work throtl_work; |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 107 | |
Andreas Schwab | 6f03793 | 2011-03-30 12:21:56 +0200 | [diff] [blame] | 108 | int limits_changed; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 109 | }; |
| 110 | |
| 111 | enum tg_state_flags { |
| 112 | THROTL_TG_FLAG_on_rr = 0, /* on round-robin busy list */ |
| 113 | }; |
| 114 | |
| 115 | #define THROTL_TG_FNS(name) \ |
| 116 | static inline void throtl_mark_tg_##name(struct throtl_grp *tg) \ |
| 117 | { \ |
| 118 | (tg)->flags |= (1 << THROTL_TG_FLAG_##name); \ |
| 119 | } \ |
| 120 | static inline void throtl_clear_tg_##name(struct throtl_grp *tg) \ |
| 121 | { \ |
| 122 | (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name); \ |
| 123 | } \ |
| 124 | static inline int throtl_tg_##name(const struct throtl_grp *tg) \ |
| 125 | { \ |
| 126 | return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0; \ |
| 127 | } |
| 128 | |
| 129 | THROTL_TG_FNS(on_rr); |
| 130 | |
| 131 | #define throtl_log_tg(td, tg, fmt, args...) \ |
| 132 | blk_add_trace_msg((td)->queue, "throtl %s " fmt, \ |
| 133 | blkg_path(&(tg)->blkg), ##args); \ |
| 134 | |
| 135 | #define throtl_log(td, fmt, args...) \ |
| 136 | blk_add_trace_msg((td)->queue, "throtl " fmt, ##args) |
| 137 | |
| 138 | static inline struct throtl_grp *tg_of_blkg(struct blkio_group *blkg) |
| 139 | { |
| 140 | if (blkg) |
| 141 | return container_of(blkg, struct throtl_grp, blkg); |
| 142 | |
| 143 | return NULL; |
| 144 | } |
| 145 | |
Joe Perches | d2f31a5 | 2011-06-13 20:19:27 +0200 | [diff] [blame] | 146 | static inline unsigned int total_nr_queued(struct throtl_data *td) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 147 | { |
Joe Perches | d2f31a5 | 2011-06-13 20:19:27 +0200 | [diff] [blame] | 148 | return td->nr_queued[0] + td->nr_queued[1]; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 149 | } |
| 150 | |
| 151 | static inline struct throtl_grp *throtl_ref_get_tg(struct throtl_grp *tg) |
| 152 | { |
| 153 | atomic_inc(&tg->ref); |
| 154 | return tg; |
| 155 | } |
| 156 | |
Vivek Goyal | 4843c69 | 2011-05-19 15:38:27 -0400 | [diff] [blame] | 157 | static void throtl_free_tg(struct rcu_head *head) |
| 158 | { |
| 159 | struct throtl_grp *tg; |
| 160 | |
| 161 | tg = container_of(head, struct throtl_grp, rcu_head); |
Vivek Goyal | 5624a4e | 2011-05-19 15:38:28 -0400 | [diff] [blame] | 162 | free_percpu(tg->blkg.stats_cpu); |
Vivek Goyal | 4843c69 | 2011-05-19 15:38:27 -0400 | [diff] [blame] | 163 | kfree(tg); |
| 164 | } |
| 165 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 166 | static void throtl_put_tg(struct throtl_grp *tg) |
| 167 | { |
| 168 | BUG_ON(atomic_read(&tg->ref) <= 0); |
| 169 | if (!atomic_dec_and_test(&tg->ref)) |
| 170 | return; |
Vivek Goyal | 4843c69 | 2011-05-19 15:38:27 -0400 | [diff] [blame] | 171 | |
| 172 | /* |
| 173 | * A group is freed in rcu manner. But having an rcu lock does not |
| 174 | * mean that one can access all the fields of blkg and assume these |
| 175 | * are valid. For example, don't try to follow throtl_data and |
| 176 | * request queue links. |
| 177 | * |
| 178 | * Having a reference to blkg under an rcu allows acess to only |
| 179 | * values local to groups like group stats and group rate limits |
| 180 | */ |
| 181 | call_rcu(&tg->rcu_head, throtl_free_tg); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 182 | } |
| 183 | |
Vivek Goyal | a29a171 | 2011-05-19 15:38:19 -0400 | [diff] [blame] | 184 | static void throtl_init_group(struct throtl_grp *tg) |
| 185 | { |
| 186 | INIT_HLIST_NODE(&tg->tg_node); |
| 187 | RB_CLEAR_NODE(&tg->rb_node); |
| 188 | bio_list_init(&tg->bio_lists[0]); |
| 189 | bio_list_init(&tg->bio_lists[1]); |
| 190 | tg->limits_changed = false; |
| 191 | |
| 192 | /* Practically unlimited BW */ |
| 193 | tg->bps[0] = tg->bps[1] = -1; |
| 194 | tg->iops[0] = tg->iops[1] = -1; |
| 195 | |
| 196 | /* |
| 197 | * Take the initial reference that will be released on destroy |
| 198 | * This can be thought of a joint reference by cgroup and |
| 199 | * request queue which will be dropped by either request queue |
| 200 | * exit or cgroup deletion path depending on who is exiting first. |
| 201 | */ |
| 202 | atomic_set(&tg->ref, 1); |
| 203 | } |
| 204 | |
| 205 | /* Should be called with rcu read lock held (needed for blkcg) */ |
| 206 | static void |
| 207 | throtl_add_group_to_td_list(struct throtl_data *td, struct throtl_grp *tg) |
| 208 | { |
| 209 | hlist_add_head(&tg->tg_node, &td->tg_list); |
| 210 | td->nr_undestroyed_grps++; |
| 211 | } |
| 212 | |
Vivek Goyal | 269f541 | 2011-05-19 15:38:25 -0400 | [diff] [blame] | 213 | static void |
| 214 | __throtl_tg_fill_dev_details(struct throtl_data *td, struct throtl_grp *tg) |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 215 | { |
| 216 | struct backing_dev_info *bdi = &td->queue->backing_dev_info; |
| 217 | unsigned int major, minor; |
| 218 | |
Vivek Goyal | 269f541 | 2011-05-19 15:38:25 -0400 | [diff] [blame] | 219 | if (!tg || tg->blkg.dev) |
| 220 | return; |
| 221 | |
| 222 | /* |
| 223 | * Fill in device details for a group which might not have been |
| 224 | * filled at group creation time as queue was being instantiated |
| 225 | * and driver had not attached a device yet |
| 226 | */ |
| 227 | if (bdi->dev && dev_name(bdi->dev)) { |
| 228 | sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor); |
| 229 | tg->blkg.dev = MKDEV(major, minor); |
| 230 | } |
| 231 | } |
| 232 | |
Vivek Goyal | af75cd3 | 2011-05-19 15:38:31 -0400 | [diff] [blame] | 233 | /* |
| 234 | * Should be called with without queue lock held. Here queue lock will be |
| 235 | * taken rarely. It will be taken only once during life time of a group |
| 236 | * if need be |
| 237 | */ |
| 238 | static void |
| 239 | throtl_tg_fill_dev_details(struct throtl_data *td, struct throtl_grp *tg) |
| 240 | { |
| 241 | if (!tg || tg->blkg.dev) |
| 242 | return; |
| 243 | |
| 244 | spin_lock_irq(td->queue->queue_lock); |
| 245 | __throtl_tg_fill_dev_details(td, tg); |
| 246 | spin_unlock_irq(td->queue->queue_lock); |
| 247 | } |
| 248 | |
Vivek Goyal | 269f541 | 2011-05-19 15:38:25 -0400 | [diff] [blame] | 249 | static void throtl_init_add_tg_lists(struct throtl_data *td, |
| 250 | struct throtl_grp *tg, struct blkio_cgroup *blkcg) |
| 251 | { |
| 252 | __throtl_tg_fill_dev_details(td, tg); |
| 253 | |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 254 | /* Add group onto cgroup list */ |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 255 | blkiocg_add_blkio_group(blkcg, &tg->blkg, (void *)td, |
Vivek Goyal | 269f541 | 2011-05-19 15:38:25 -0400 | [diff] [blame] | 256 | tg->blkg.dev, BLKIO_POLICY_THROTL); |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 257 | |
| 258 | tg->bps[READ] = blkcg_get_read_bps(blkcg, tg->blkg.dev); |
| 259 | tg->bps[WRITE] = blkcg_get_write_bps(blkcg, tg->blkg.dev); |
| 260 | tg->iops[READ] = blkcg_get_read_iops(blkcg, tg->blkg.dev); |
| 261 | tg->iops[WRITE] = blkcg_get_write_iops(blkcg, tg->blkg.dev); |
| 262 | |
| 263 | throtl_add_group_to_td_list(td, tg); |
| 264 | } |
| 265 | |
| 266 | /* Should be called without queue lock and outside of rcu period */ |
| 267 | static struct throtl_grp *throtl_alloc_tg(struct throtl_data *td) |
| 268 | { |
| 269 | struct throtl_grp *tg = NULL; |
Vivek Goyal | 5624a4e | 2011-05-19 15:38:28 -0400 | [diff] [blame] | 270 | int ret; |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 271 | |
| 272 | tg = kzalloc_node(sizeof(*tg), GFP_ATOMIC, td->queue->node); |
| 273 | if (!tg) |
| 274 | return NULL; |
| 275 | |
Vivek Goyal | 5624a4e | 2011-05-19 15:38:28 -0400 | [diff] [blame] | 276 | ret = blkio_alloc_blkg_stats(&tg->blkg); |
| 277 | |
| 278 | if (ret) { |
| 279 | kfree(tg); |
| 280 | return NULL; |
| 281 | } |
| 282 | |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 283 | throtl_init_group(tg); |
| 284 | return tg; |
| 285 | } |
| 286 | |
| 287 | static struct |
| 288 | throtl_grp *throtl_find_tg(struct throtl_data *td, struct blkio_cgroup *blkcg) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 289 | { |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 290 | struct throtl_grp *tg = NULL; |
| 291 | void *key = td; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 292 | |
| 293 | /* |
Vivek Goyal | be2c6b1 | 2011-01-19 08:25:02 -0700 | [diff] [blame] | 294 | * This is the common case when there are no blkio cgroups. |
| 295 | * Avoid lookup in this case |
| 296 | */ |
| 297 | if (blkcg == &blkio_root_cgroup) |
Vivek Goyal | 29b1258 | 2011-05-19 15:38:24 -0400 | [diff] [blame] | 298 | tg = td->root_tg; |
Vivek Goyal | be2c6b1 | 2011-01-19 08:25:02 -0700 | [diff] [blame] | 299 | else |
| 300 | tg = tg_of_blkg(blkiocg_lookup_group(blkcg, key)); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 301 | |
Vivek Goyal | 269f541 | 2011-05-19 15:38:25 -0400 | [diff] [blame] | 302 | __throtl_tg_fill_dev_details(td, tg); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 303 | return tg; |
| 304 | } |
| 305 | |
| 306 | static struct throtl_grp * throtl_get_tg(struct throtl_data *td) |
| 307 | { |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 308 | struct throtl_grp *tg = NULL, *__tg = NULL; |
Vivek Goyal | 70087dc | 2011-05-16 15:24:08 +0200 | [diff] [blame] | 309 | struct blkio_cgroup *blkcg; |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 310 | struct request_queue *q = td->queue; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 311 | |
Tejun Heo | c9a929d | 2011-10-19 14:42:16 +0200 | [diff] [blame] | 312 | /* no throttling for dead queue */ |
Tejun Heo | 6ecf23a | 2012-03-05 13:14:59 -0800 | [diff] [blame] | 313 | if (unlikely(blk_queue_bypass(q))) |
Tejun Heo | c9a929d | 2011-10-19 14:42:16 +0200 | [diff] [blame] | 314 | return NULL; |
| 315 | |
Vivek Goyal | 70087dc | 2011-05-16 15:24:08 +0200 | [diff] [blame] | 316 | blkcg = task_blkio_cgroup(current); |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 317 | tg = throtl_find_tg(td, blkcg); |
Tejun Heo | 2a7f124 | 2012-03-05 13:15:01 -0800 | [diff] [blame^] | 318 | if (tg) |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 319 | return tg; |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 320 | |
| 321 | /* |
| 322 | * Need to allocate a group. Allocation of group also needs allocation |
| 323 | * of per cpu stats which in-turn takes a mutex() and can block. Hence |
Tejun Heo | 315fcee | 2011-10-19 14:31:25 +0200 | [diff] [blame] | 324 | * we need to drop rcu lock and queue_lock before we call alloc. |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 325 | */ |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 326 | spin_unlock_irq(q->queue_lock); |
Tejun Heo | 2a7f124 | 2012-03-05 13:15:01 -0800 | [diff] [blame^] | 327 | rcu_read_unlock(); |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 328 | |
| 329 | tg = throtl_alloc_tg(td); |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 330 | |
| 331 | /* Group allocated and queue is still alive. take the lock */ |
Tejun Heo | 2a7f124 | 2012-03-05 13:15:01 -0800 | [diff] [blame^] | 332 | rcu_read_lock(); |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 333 | spin_lock_irq(q->queue_lock); |
| 334 | |
Tejun Heo | bc16a4f | 2011-10-19 14:33:01 +0200 | [diff] [blame] | 335 | /* Make sure @q is still alive */ |
Tejun Heo | 6ecf23a | 2012-03-05 13:14:59 -0800 | [diff] [blame] | 336 | if (unlikely(blk_queue_bypass(q))) { |
Tejun Heo | bc16a4f | 2011-10-19 14:33:01 +0200 | [diff] [blame] | 337 | kfree(tg); |
| 338 | return NULL; |
| 339 | } |
| 340 | |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 341 | /* |
| 342 | * Initialize the new group. After sleeping, read the blkcg again. |
| 343 | */ |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 344 | blkcg = task_blkio_cgroup(current); |
| 345 | |
| 346 | /* |
| 347 | * If some other thread already allocated the group while we were |
| 348 | * not holding queue lock, free up the group |
| 349 | */ |
| 350 | __tg = throtl_find_tg(td, blkcg); |
| 351 | |
| 352 | if (__tg) { |
| 353 | kfree(tg); |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 354 | return __tg; |
| 355 | } |
| 356 | |
| 357 | /* Group allocation failed. Account the IO to root group */ |
| 358 | if (!tg) { |
Vivek Goyal | 29b1258 | 2011-05-19 15:38:24 -0400 | [diff] [blame] | 359 | tg = td->root_tg; |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 360 | return tg; |
| 361 | } |
| 362 | |
| 363 | throtl_init_add_tg_lists(td, tg, blkcg); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 364 | return tg; |
| 365 | } |
| 366 | |
| 367 | static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root) |
| 368 | { |
| 369 | /* Service tree is empty */ |
| 370 | if (!root->count) |
| 371 | return NULL; |
| 372 | |
| 373 | if (!root->left) |
| 374 | root->left = rb_first(&root->rb); |
| 375 | |
| 376 | if (root->left) |
| 377 | return rb_entry_tg(root->left); |
| 378 | |
| 379 | return NULL; |
| 380 | } |
| 381 | |
| 382 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) |
| 383 | { |
| 384 | rb_erase(n, root); |
| 385 | RB_CLEAR_NODE(n); |
| 386 | } |
| 387 | |
| 388 | static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root) |
| 389 | { |
| 390 | if (root->left == n) |
| 391 | root->left = NULL; |
| 392 | rb_erase_init(n, &root->rb); |
| 393 | --root->count; |
| 394 | } |
| 395 | |
| 396 | static void update_min_dispatch_time(struct throtl_rb_root *st) |
| 397 | { |
| 398 | struct throtl_grp *tg; |
| 399 | |
| 400 | tg = throtl_rb_first(st); |
| 401 | if (!tg) |
| 402 | return; |
| 403 | |
| 404 | st->min_disptime = tg->disptime; |
| 405 | } |
| 406 | |
| 407 | static void |
| 408 | tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg) |
| 409 | { |
| 410 | struct rb_node **node = &st->rb.rb_node; |
| 411 | struct rb_node *parent = NULL; |
| 412 | struct throtl_grp *__tg; |
| 413 | unsigned long key = tg->disptime; |
| 414 | int left = 1; |
| 415 | |
| 416 | while (*node != NULL) { |
| 417 | parent = *node; |
| 418 | __tg = rb_entry_tg(parent); |
| 419 | |
| 420 | if (time_before(key, __tg->disptime)) |
| 421 | node = &parent->rb_left; |
| 422 | else { |
| 423 | node = &parent->rb_right; |
| 424 | left = 0; |
| 425 | } |
| 426 | } |
| 427 | |
| 428 | if (left) |
| 429 | st->left = &tg->rb_node; |
| 430 | |
| 431 | rb_link_node(&tg->rb_node, parent, node); |
| 432 | rb_insert_color(&tg->rb_node, &st->rb); |
| 433 | } |
| 434 | |
| 435 | static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) |
| 436 | { |
| 437 | struct throtl_rb_root *st = &td->tg_service_tree; |
| 438 | |
| 439 | tg_service_tree_add(st, tg); |
| 440 | throtl_mark_tg_on_rr(tg); |
| 441 | st->count++; |
| 442 | } |
| 443 | |
| 444 | static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg) |
| 445 | { |
| 446 | if (!throtl_tg_on_rr(tg)) |
| 447 | __throtl_enqueue_tg(td, tg); |
| 448 | } |
| 449 | |
| 450 | static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) |
| 451 | { |
| 452 | throtl_rb_erase(&tg->rb_node, &td->tg_service_tree); |
| 453 | throtl_clear_tg_on_rr(tg); |
| 454 | } |
| 455 | |
| 456 | static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg) |
| 457 | { |
| 458 | if (throtl_tg_on_rr(tg)) |
| 459 | __throtl_dequeue_tg(td, tg); |
| 460 | } |
| 461 | |
| 462 | static void throtl_schedule_next_dispatch(struct throtl_data *td) |
| 463 | { |
| 464 | struct throtl_rb_root *st = &td->tg_service_tree; |
| 465 | |
| 466 | /* |
| 467 | * If there are more bios pending, schedule more work. |
| 468 | */ |
| 469 | if (!total_nr_queued(td)) |
| 470 | return; |
| 471 | |
| 472 | BUG_ON(!st->count); |
| 473 | |
| 474 | update_min_dispatch_time(st); |
| 475 | |
| 476 | if (time_before_eq(st->min_disptime, jiffies)) |
Vivek Goyal | 450adcb | 2011-03-01 13:40:54 -0500 | [diff] [blame] | 477 | throtl_schedule_delayed_work(td, 0); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 478 | else |
Vivek Goyal | 450adcb | 2011-03-01 13:40:54 -0500 | [diff] [blame] | 479 | throtl_schedule_delayed_work(td, (st->min_disptime - jiffies)); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 480 | } |
| 481 | |
| 482 | static inline void |
| 483 | throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) |
| 484 | { |
| 485 | tg->bytes_disp[rw] = 0; |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 486 | tg->io_disp[rw] = 0; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 487 | tg->slice_start[rw] = jiffies; |
| 488 | tg->slice_end[rw] = jiffies + throtl_slice; |
| 489 | throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu", |
| 490 | rw == READ ? 'R' : 'W', tg->slice_start[rw], |
| 491 | tg->slice_end[rw], jiffies); |
| 492 | } |
| 493 | |
Vivek Goyal | d1ae8ff | 2010-12-01 19:34:46 +0100 | [diff] [blame] | 494 | static inline void throtl_set_slice_end(struct throtl_data *td, |
| 495 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) |
| 496 | { |
| 497 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); |
| 498 | } |
| 499 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 500 | static inline void throtl_extend_slice(struct throtl_data *td, |
| 501 | struct throtl_grp *tg, bool rw, unsigned long jiffy_end) |
| 502 | { |
| 503 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); |
| 504 | throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu", |
| 505 | rw == READ ? 'R' : 'W', tg->slice_start[rw], |
| 506 | tg->slice_end[rw], jiffies); |
| 507 | } |
| 508 | |
| 509 | /* Determine if previously allocated or extended slice is complete or not */ |
| 510 | static bool |
| 511 | throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw) |
| 512 | { |
| 513 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) |
| 514 | return 0; |
| 515 | |
| 516 | return 1; |
| 517 | } |
| 518 | |
| 519 | /* Trim the used slices and adjust slice start accordingly */ |
| 520 | static inline void |
| 521 | throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw) |
| 522 | { |
Vivek Goyal | 3aad5d3 | 2010-10-01 14:51:14 +0200 | [diff] [blame] | 523 | unsigned long nr_slices, time_elapsed, io_trim; |
| 524 | u64 bytes_trim, tmp; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 525 | |
| 526 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); |
| 527 | |
| 528 | /* |
| 529 | * If bps are unlimited (-1), then time slice don't get |
| 530 | * renewed. Don't try to trim the slice if slice is used. A new |
| 531 | * slice will start when appropriate. |
| 532 | */ |
| 533 | if (throtl_slice_used(td, tg, rw)) |
| 534 | return; |
| 535 | |
Vivek Goyal | d1ae8ff | 2010-12-01 19:34:46 +0100 | [diff] [blame] | 536 | /* |
| 537 | * A bio has been dispatched. Also adjust slice_end. It might happen |
| 538 | * that initially cgroup limit was very low resulting in high |
| 539 | * slice_end, but later limit was bumped up and bio was dispached |
| 540 | * sooner, then we need to reduce slice_end. A high bogus slice_end |
| 541 | * is bad because it does not allow new slice to start. |
| 542 | */ |
| 543 | |
| 544 | throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice); |
| 545 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 546 | time_elapsed = jiffies - tg->slice_start[rw]; |
| 547 | |
| 548 | nr_slices = time_elapsed / throtl_slice; |
| 549 | |
| 550 | if (!nr_slices) |
| 551 | return; |
Vivek Goyal | 3aad5d3 | 2010-10-01 14:51:14 +0200 | [diff] [blame] | 552 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
| 553 | do_div(tmp, HZ); |
| 554 | bytes_trim = tmp; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 555 | |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 556 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 557 | |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 558 | if (!bytes_trim && !io_trim) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 559 | return; |
| 560 | |
| 561 | if (tg->bytes_disp[rw] >= bytes_trim) |
| 562 | tg->bytes_disp[rw] -= bytes_trim; |
| 563 | else |
| 564 | tg->bytes_disp[rw] = 0; |
| 565 | |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 566 | if (tg->io_disp[rw] >= io_trim) |
| 567 | tg->io_disp[rw] -= io_trim; |
| 568 | else |
| 569 | tg->io_disp[rw] = 0; |
| 570 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 571 | tg->slice_start[rw] += nr_slices * throtl_slice; |
| 572 | |
Vivek Goyal | 3aad5d3 | 2010-10-01 14:51:14 +0200 | [diff] [blame] | 573 | throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu" |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 574 | " start=%lu end=%lu jiffies=%lu", |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 575 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 576 | tg->slice_start[rw], tg->slice_end[rw], jiffies); |
| 577 | } |
| 578 | |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 579 | static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg, |
| 580 | struct bio *bio, unsigned long *wait) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 581 | { |
| 582 | bool rw = bio_data_dir(bio); |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 583 | unsigned int io_allowed; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 584 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
Vivek Goyal | c49c06e | 2010-10-01 21:16:42 +0200 | [diff] [blame] | 585 | u64 tmp; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 586 | |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 587 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 588 | |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 589 | /* Slice has just started. Consider one slice interval */ |
| 590 | if (!jiffy_elapsed) |
| 591 | jiffy_elapsed_rnd = throtl_slice; |
| 592 | |
| 593 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); |
| 594 | |
Vivek Goyal | c49c06e | 2010-10-01 21:16:42 +0200 | [diff] [blame] | 595 | /* |
| 596 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be |
| 597 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we |
| 598 | * will allow dispatch after 1 second and after that slice should |
| 599 | * have been trimmed. |
| 600 | */ |
| 601 | |
| 602 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; |
| 603 | do_div(tmp, HZ); |
| 604 | |
| 605 | if (tmp > UINT_MAX) |
| 606 | io_allowed = UINT_MAX; |
| 607 | else |
| 608 | io_allowed = tmp; |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 609 | |
| 610 | if (tg->io_disp[rw] + 1 <= io_allowed) { |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 611 | if (wait) |
| 612 | *wait = 0; |
| 613 | return 1; |
| 614 | } |
| 615 | |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 616 | /* Calc approx time to dispatch */ |
| 617 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; |
| 618 | |
| 619 | if (jiffy_wait > jiffy_elapsed) |
| 620 | jiffy_wait = jiffy_wait - jiffy_elapsed; |
| 621 | else |
| 622 | jiffy_wait = 1; |
| 623 | |
| 624 | if (wait) |
| 625 | *wait = jiffy_wait; |
| 626 | return 0; |
| 627 | } |
| 628 | |
| 629 | static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg, |
| 630 | struct bio *bio, unsigned long *wait) |
| 631 | { |
| 632 | bool rw = bio_data_dir(bio); |
Vivek Goyal | 3aad5d3 | 2010-10-01 14:51:14 +0200 | [diff] [blame] | 633 | u64 bytes_allowed, extra_bytes, tmp; |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 634 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 635 | |
| 636 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
| 637 | |
| 638 | /* Slice has just started. Consider one slice interval */ |
| 639 | if (!jiffy_elapsed) |
| 640 | jiffy_elapsed_rnd = throtl_slice; |
| 641 | |
| 642 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); |
| 643 | |
Vivek Goyal | 5e901a2 | 2010-10-01 21:16:38 +0200 | [diff] [blame] | 644 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
| 645 | do_div(tmp, HZ); |
Vivek Goyal | 3aad5d3 | 2010-10-01 14:51:14 +0200 | [diff] [blame] | 646 | bytes_allowed = tmp; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 647 | |
| 648 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { |
| 649 | if (wait) |
| 650 | *wait = 0; |
| 651 | return 1; |
| 652 | } |
| 653 | |
| 654 | /* Calc approx time to dispatch */ |
| 655 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; |
| 656 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); |
| 657 | |
| 658 | if (!jiffy_wait) |
| 659 | jiffy_wait = 1; |
| 660 | |
| 661 | /* |
| 662 | * This wait time is without taking into consideration the rounding |
| 663 | * up we did. Add that time also. |
| 664 | */ |
| 665 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 666 | if (wait) |
| 667 | *wait = jiffy_wait; |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 668 | return 0; |
| 669 | } |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 670 | |
Vivek Goyal | af75cd3 | 2011-05-19 15:38:31 -0400 | [diff] [blame] | 671 | static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) { |
| 672 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) |
| 673 | return 1; |
| 674 | return 0; |
| 675 | } |
| 676 | |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 677 | /* |
| 678 | * Returns whether one can dispatch a bio or not. Also returns approx number |
| 679 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched |
| 680 | */ |
| 681 | static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg, |
| 682 | struct bio *bio, unsigned long *wait) |
| 683 | { |
| 684 | bool rw = bio_data_dir(bio); |
| 685 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; |
| 686 | |
| 687 | /* |
| 688 | * Currently whole state machine of group depends on first bio |
| 689 | * queued in the group bio list. So one should not be calling |
| 690 | * this function with a different bio if there are other bios |
| 691 | * queued. |
| 692 | */ |
| 693 | BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw])); |
| 694 | |
| 695 | /* If tg->bps = -1, then BW is unlimited */ |
| 696 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { |
| 697 | if (wait) |
| 698 | *wait = 0; |
| 699 | return 1; |
| 700 | } |
| 701 | |
| 702 | /* |
| 703 | * If previous slice expired, start a new one otherwise renew/extend |
| 704 | * existing slice to make sure it is at least throtl_slice interval |
| 705 | * long since now. |
| 706 | */ |
| 707 | if (throtl_slice_used(td, tg, rw)) |
| 708 | throtl_start_new_slice(td, tg, rw); |
| 709 | else { |
| 710 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) |
| 711 | throtl_extend_slice(td, tg, rw, jiffies + throtl_slice); |
| 712 | } |
| 713 | |
| 714 | if (tg_with_in_bps_limit(td, tg, bio, &bps_wait) |
| 715 | && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) { |
| 716 | if (wait) |
| 717 | *wait = 0; |
| 718 | return 1; |
| 719 | } |
| 720 | |
| 721 | max_wait = max(bps_wait, iops_wait); |
| 722 | |
| 723 | if (wait) |
| 724 | *wait = max_wait; |
| 725 | |
| 726 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) |
| 727 | throtl_extend_slice(td, tg, rw, jiffies + max_wait); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 728 | |
| 729 | return 0; |
| 730 | } |
| 731 | |
| 732 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) |
| 733 | { |
| 734 | bool rw = bio_data_dir(bio); |
Shaohua Li | e5a94f5 | 2011-08-01 10:31:06 +0200 | [diff] [blame] | 735 | bool sync = rw_is_sync(bio->bi_rw); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 736 | |
| 737 | /* Charge the bio to the group */ |
| 738 | tg->bytes_disp[rw] += bio->bi_size; |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 739 | tg->io_disp[rw]++; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 740 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 741 | blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size, rw, sync); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 742 | } |
| 743 | |
| 744 | static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg, |
| 745 | struct bio *bio) |
| 746 | { |
| 747 | bool rw = bio_data_dir(bio); |
| 748 | |
| 749 | bio_list_add(&tg->bio_lists[rw], bio); |
| 750 | /* Take a bio reference on tg */ |
| 751 | throtl_ref_get_tg(tg); |
| 752 | tg->nr_queued[rw]++; |
| 753 | td->nr_queued[rw]++; |
| 754 | throtl_enqueue_tg(td, tg); |
| 755 | } |
| 756 | |
| 757 | static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg) |
| 758 | { |
| 759 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; |
| 760 | struct bio *bio; |
| 761 | |
| 762 | if ((bio = bio_list_peek(&tg->bio_lists[READ]))) |
| 763 | tg_may_dispatch(td, tg, bio, &read_wait); |
| 764 | |
| 765 | if ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) |
| 766 | tg_may_dispatch(td, tg, bio, &write_wait); |
| 767 | |
| 768 | min_wait = min(read_wait, write_wait); |
| 769 | disptime = jiffies + min_wait; |
| 770 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 771 | /* Update dispatch time */ |
| 772 | throtl_dequeue_tg(td, tg); |
| 773 | tg->disptime = disptime; |
| 774 | throtl_enqueue_tg(td, tg); |
| 775 | } |
| 776 | |
| 777 | static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg, |
| 778 | bool rw, struct bio_list *bl) |
| 779 | { |
| 780 | struct bio *bio; |
| 781 | |
| 782 | bio = bio_list_pop(&tg->bio_lists[rw]); |
| 783 | tg->nr_queued[rw]--; |
| 784 | /* Drop bio reference on tg */ |
| 785 | throtl_put_tg(tg); |
| 786 | |
| 787 | BUG_ON(td->nr_queued[rw] <= 0); |
| 788 | td->nr_queued[rw]--; |
| 789 | |
| 790 | throtl_charge_bio(tg, bio); |
| 791 | bio_list_add(bl, bio); |
| 792 | bio->bi_rw |= REQ_THROTTLED; |
| 793 | |
| 794 | throtl_trim_slice(td, tg, rw); |
| 795 | } |
| 796 | |
| 797 | static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg, |
| 798 | struct bio_list *bl) |
| 799 | { |
| 800 | unsigned int nr_reads = 0, nr_writes = 0; |
| 801 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; |
Vivek Goyal | c2f6805 | 2010-11-15 19:32:42 +0100 | [diff] [blame] | 802 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 803 | struct bio *bio; |
| 804 | |
| 805 | /* Try to dispatch 75% READS and 25% WRITES */ |
| 806 | |
| 807 | while ((bio = bio_list_peek(&tg->bio_lists[READ])) |
| 808 | && tg_may_dispatch(td, tg, bio, NULL)) { |
| 809 | |
| 810 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); |
| 811 | nr_reads++; |
| 812 | |
| 813 | if (nr_reads >= max_nr_reads) |
| 814 | break; |
| 815 | } |
| 816 | |
| 817 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE])) |
| 818 | && tg_may_dispatch(td, tg, bio, NULL)) { |
| 819 | |
| 820 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl); |
| 821 | nr_writes++; |
| 822 | |
| 823 | if (nr_writes >= max_nr_writes) |
| 824 | break; |
| 825 | } |
| 826 | |
| 827 | return nr_reads + nr_writes; |
| 828 | } |
| 829 | |
| 830 | static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl) |
| 831 | { |
| 832 | unsigned int nr_disp = 0; |
| 833 | struct throtl_grp *tg; |
| 834 | struct throtl_rb_root *st = &td->tg_service_tree; |
| 835 | |
| 836 | while (1) { |
| 837 | tg = throtl_rb_first(st); |
| 838 | |
| 839 | if (!tg) |
| 840 | break; |
| 841 | |
| 842 | if (time_before(jiffies, tg->disptime)) |
| 843 | break; |
| 844 | |
| 845 | throtl_dequeue_tg(td, tg); |
| 846 | |
| 847 | nr_disp += throtl_dispatch_tg(td, tg, bl); |
| 848 | |
| 849 | if (tg->nr_queued[0] || tg->nr_queued[1]) { |
| 850 | tg_update_disptime(td, tg); |
| 851 | throtl_enqueue_tg(td, tg); |
| 852 | } |
| 853 | |
| 854 | if (nr_disp >= throtl_quantum) |
| 855 | break; |
| 856 | } |
| 857 | |
| 858 | return nr_disp; |
| 859 | } |
| 860 | |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 861 | static void throtl_process_limit_change(struct throtl_data *td) |
| 862 | { |
| 863 | struct throtl_grp *tg; |
| 864 | struct hlist_node *pos, *n; |
| 865 | |
Vivek Goyal | de701c7 | 2011-03-07 21:09:32 +0100 | [diff] [blame] | 866 | if (!td->limits_changed) |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 867 | return; |
| 868 | |
Vivek Goyal | de701c7 | 2011-03-07 21:09:32 +0100 | [diff] [blame] | 869 | xchg(&td->limits_changed, false); |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 870 | |
Vivek Goyal | de701c7 | 2011-03-07 21:09:32 +0100 | [diff] [blame] | 871 | throtl_log(td, "limits changed"); |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 872 | |
Vivek Goyal | 04a6b51 | 2010-12-01 19:34:52 +0100 | [diff] [blame] | 873 | hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) { |
Vivek Goyal | de701c7 | 2011-03-07 21:09:32 +0100 | [diff] [blame] | 874 | if (!tg->limits_changed) |
| 875 | continue; |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 876 | |
Vivek Goyal | de701c7 | 2011-03-07 21:09:32 +0100 | [diff] [blame] | 877 | if (!xchg(&tg->limits_changed, false)) |
| 878 | continue; |
| 879 | |
| 880 | throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu" |
| 881 | " riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE], |
| 882 | tg->iops[READ], tg->iops[WRITE]); |
| 883 | |
Vivek Goyal | 04521db | 2011-03-22 21:54:29 +0100 | [diff] [blame] | 884 | /* |
| 885 | * Restart the slices for both READ and WRITES. It |
| 886 | * might happen that a group's limit are dropped |
| 887 | * suddenly and we don't want to account recently |
| 888 | * dispatched IO with new low rate |
| 889 | */ |
| 890 | throtl_start_new_slice(td, tg, 0); |
| 891 | throtl_start_new_slice(td, tg, 1); |
| 892 | |
Vivek Goyal | de701c7 | 2011-03-07 21:09:32 +0100 | [diff] [blame] | 893 | if (throtl_tg_on_rr(tg)) |
| 894 | tg_update_disptime(td, tg); |
| 895 | } |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 896 | } |
| 897 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 898 | /* Dispatch throttled bios. Should be called without queue lock held. */ |
| 899 | static int throtl_dispatch(struct request_queue *q) |
| 900 | { |
| 901 | struct throtl_data *td = q->td; |
| 902 | unsigned int nr_disp = 0; |
| 903 | struct bio_list bio_list_on_stack; |
| 904 | struct bio *bio; |
Vivek Goyal | 69d60eb | 2011-03-09 08:27:37 +0100 | [diff] [blame] | 905 | struct blk_plug plug; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 906 | |
| 907 | spin_lock_irq(q->queue_lock); |
| 908 | |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 909 | throtl_process_limit_change(td); |
| 910 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 911 | if (!total_nr_queued(td)) |
| 912 | goto out; |
| 913 | |
| 914 | bio_list_init(&bio_list_on_stack); |
| 915 | |
Joe Perches | d2f31a5 | 2011-06-13 20:19:27 +0200 | [diff] [blame] | 916 | throtl_log(td, "dispatch nr_queued=%u read=%u write=%u", |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 917 | total_nr_queued(td), td->nr_queued[READ], |
| 918 | td->nr_queued[WRITE]); |
| 919 | |
| 920 | nr_disp = throtl_select_dispatch(td, &bio_list_on_stack); |
| 921 | |
| 922 | if (nr_disp) |
| 923 | throtl_log(td, "bios disp=%u", nr_disp); |
| 924 | |
| 925 | throtl_schedule_next_dispatch(td); |
| 926 | out: |
| 927 | spin_unlock_irq(q->queue_lock); |
| 928 | |
| 929 | /* |
| 930 | * If we dispatched some requests, unplug the queue to make sure |
| 931 | * immediate dispatch |
| 932 | */ |
| 933 | if (nr_disp) { |
Vivek Goyal | 69d60eb | 2011-03-09 08:27:37 +0100 | [diff] [blame] | 934 | blk_start_plug(&plug); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 935 | while((bio = bio_list_pop(&bio_list_on_stack))) |
| 936 | generic_make_request(bio); |
Vivek Goyal | 69d60eb | 2011-03-09 08:27:37 +0100 | [diff] [blame] | 937 | blk_finish_plug(&plug); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 938 | } |
| 939 | return nr_disp; |
| 940 | } |
| 941 | |
| 942 | void blk_throtl_work(struct work_struct *work) |
| 943 | { |
| 944 | struct throtl_data *td = container_of(work, struct throtl_data, |
| 945 | throtl_work.work); |
| 946 | struct request_queue *q = td->queue; |
| 947 | |
| 948 | throtl_dispatch(q); |
| 949 | } |
| 950 | |
| 951 | /* Call with queue lock held */ |
Vivek Goyal | 450adcb | 2011-03-01 13:40:54 -0500 | [diff] [blame] | 952 | static void |
| 953 | throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 954 | { |
| 955 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 956 | struct delayed_work *dwork = &td->throtl_work; |
| 957 | |
Vivek Goyal | 04521db | 2011-03-22 21:54:29 +0100 | [diff] [blame] | 958 | /* schedule work if limits changed even if no bio is queued */ |
Joe Perches | d2f31a5 | 2011-06-13 20:19:27 +0200 | [diff] [blame] | 959 | if (total_nr_queued(td) || td->limits_changed) { |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 960 | /* |
| 961 | * We might have a work scheduled to be executed in future. |
| 962 | * Cancel that and schedule a new one. |
| 963 | */ |
| 964 | __cancel_delayed_work(dwork); |
Vivek Goyal | 450adcb | 2011-03-01 13:40:54 -0500 | [diff] [blame] | 965 | queue_delayed_work(kthrotld_workqueue, dwork, delay); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 966 | throtl_log(td, "schedule work. delay=%lu jiffies=%lu", |
| 967 | delay, jiffies); |
| 968 | } |
| 969 | } |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 970 | |
| 971 | static void |
| 972 | throtl_destroy_tg(struct throtl_data *td, struct throtl_grp *tg) |
| 973 | { |
| 974 | /* Something wrong if we are trying to remove same group twice */ |
| 975 | BUG_ON(hlist_unhashed(&tg->tg_node)); |
| 976 | |
| 977 | hlist_del_init(&tg->tg_node); |
| 978 | |
| 979 | /* |
| 980 | * Put the reference taken at the time of creation so that when all |
| 981 | * queues are gone, group can be destroyed. |
| 982 | */ |
| 983 | throtl_put_tg(tg); |
| 984 | td->nr_undestroyed_grps--; |
| 985 | } |
| 986 | |
Tejun Heo | 72e06c2 | 2012-03-05 13:15:00 -0800 | [diff] [blame] | 987 | static bool throtl_release_tgs(struct throtl_data *td, bool release_root) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 988 | { |
| 989 | struct hlist_node *pos, *n; |
| 990 | struct throtl_grp *tg; |
Tejun Heo | 72e06c2 | 2012-03-05 13:15:00 -0800 | [diff] [blame] | 991 | bool empty = true; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 992 | |
| 993 | hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) { |
Tejun Heo | 72e06c2 | 2012-03-05 13:15:00 -0800 | [diff] [blame] | 994 | /* skip root? */ |
| 995 | if (!release_root && tg == td->root_tg) |
| 996 | continue; |
| 997 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 998 | /* |
| 999 | * If cgroup removal path got to blk_group first and removed |
| 1000 | * it from cgroup list, then it will take care of destroying |
| 1001 | * cfqg also. |
| 1002 | */ |
| 1003 | if (!blkiocg_del_blkio_group(&tg->blkg)) |
| 1004 | throtl_destroy_tg(td, tg); |
Tejun Heo | 72e06c2 | 2012-03-05 13:15:00 -0800 | [diff] [blame] | 1005 | else |
| 1006 | empty = false; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1007 | } |
Tejun Heo | 72e06c2 | 2012-03-05 13:15:00 -0800 | [diff] [blame] | 1008 | return empty; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1009 | } |
| 1010 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1011 | /* |
| 1012 | * Blk cgroup controller notification saying that blkio_group object is being |
| 1013 | * delinked as associated cgroup object is going away. That also means that |
| 1014 | * no new IO will come in this group. So get rid of this group as soon as |
| 1015 | * any pending IO in the group is finished. |
| 1016 | * |
| 1017 | * This function is called under rcu_read_lock(). key is the rcu protected |
| 1018 | * pointer. That means "key" is a valid throtl_data pointer as long as we are |
| 1019 | * rcu read lock. |
| 1020 | * |
| 1021 | * "key" was fetched from blkio_group under blkio_cgroup->lock. That means |
| 1022 | * it should not be NULL as even if queue was going away, cgroup deltion |
| 1023 | * path got to it first. |
| 1024 | */ |
| 1025 | void throtl_unlink_blkio_group(void *key, struct blkio_group *blkg) |
| 1026 | { |
| 1027 | unsigned long flags; |
| 1028 | struct throtl_data *td = key; |
| 1029 | |
| 1030 | spin_lock_irqsave(td->queue->queue_lock, flags); |
| 1031 | throtl_destroy_tg(td, tg_of_blkg(blkg)); |
| 1032 | spin_unlock_irqrestore(td->queue->queue_lock, flags); |
| 1033 | } |
| 1034 | |
Tejun Heo | 72e06c2 | 2012-03-05 13:15:00 -0800 | [diff] [blame] | 1035 | static bool throtl_clear_queue(struct request_queue *q) |
| 1036 | { |
| 1037 | lockdep_assert_held(q->queue_lock); |
| 1038 | |
| 1039 | /* |
| 1040 | * Clear tgs but leave the root one alone. This is necessary |
| 1041 | * because root_tg is expected to be persistent and safe because |
| 1042 | * blk-throtl can never be disabled while @q is alive. This is a |
| 1043 | * kludge to prepare for unified blkg. This whole function will be |
| 1044 | * removed soon. |
| 1045 | */ |
| 1046 | return throtl_release_tgs(q->td, false); |
| 1047 | } |
| 1048 | |
Vivek Goyal | de701c7 | 2011-03-07 21:09:32 +0100 | [diff] [blame] | 1049 | static void throtl_update_blkio_group_common(struct throtl_data *td, |
| 1050 | struct throtl_grp *tg) |
| 1051 | { |
| 1052 | xchg(&tg->limits_changed, true); |
| 1053 | xchg(&td->limits_changed, true); |
| 1054 | /* Schedule a work now to process the limit change */ |
| 1055 | throtl_schedule_delayed_work(td, 0); |
| 1056 | } |
| 1057 | |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 1058 | /* |
| 1059 | * For all update functions, key should be a valid pointer because these |
| 1060 | * update functions are called under blkcg_lock, that means, blkg is |
Lucas De Marchi | 25985ed | 2011-03-30 22:57:33 -0300 | [diff] [blame] | 1061 | * valid and in turn key is valid. queue exit path can not race because |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 1062 | * of blkcg_lock |
| 1063 | * |
| 1064 | * Can not take queue lock in update functions as queue lock under blkcg_lock |
| 1065 | * is not allowed. Under other paths we take blkcg_lock under queue_lock. |
| 1066 | */ |
| 1067 | static void throtl_update_blkio_group_read_bps(void *key, |
| 1068 | struct blkio_group *blkg, u64 read_bps) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1069 | { |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 1070 | struct throtl_data *td = key; |
Vivek Goyal | de701c7 | 2011-03-07 21:09:32 +0100 | [diff] [blame] | 1071 | struct throtl_grp *tg = tg_of_blkg(blkg); |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 1072 | |
Vivek Goyal | de701c7 | 2011-03-07 21:09:32 +0100 | [diff] [blame] | 1073 | tg->bps[READ] = read_bps; |
| 1074 | throtl_update_blkio_group_common(td, tg); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1075 | } |
| 1076 | |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 1077 | static void throtl_update_blkio_group_write_bps(void *key, |
| 1078 | struct blkio_group *blkg, u64 write_bps) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1079 | { |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 1080 | struct throtl_data *td = key; |
Vivek Goyal | de701c7 | 2011-03-07 21:09:32 +0100 | [diff] [blame] | 1081 | struct throtl_grp *tg = tg_of_blkg(blkg); |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 1082 | |
Vivek Goyal | de701c7 | 2011-03-07 21:09:32 +0100 | [diff] [blame] | 1083 | tg->bps[WRITE] = write_bps; |
| 1084 | throtl_update_blkio_group_common(td, tg); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1085 | } |
| 1086 | |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 1087 | static void throtl_update_blkio_group_read_iops(void *key, |
| 1088 | struct blkio_group *blkg, unsigned int read_iops) |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 1089 | { |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 1090 | struct throtl_data *td = key; |
Vivek Goyal | de701c7 | 2011-03-07 21:09:32 +0100 | [diff] [blame] | 1091 | struct throtl_grp *tg = tg_of_blkg(blkg); |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 1092 | |
Vivek Goyal | de701c7 | 2011-03-07 21:09:32 +0100 | [diff] [blame] | 1093 | tg->iops[READ] = read_iops; |
| 1094 | throtl_update_blkio_group_common(td, tg); |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 1095 | } |
| 1096 | |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 1097 | static void throtl_update_blkio_group_write_iops(void *key, |
| 1098 | struct blkio_group *blkg, unsigned int write_iops) |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 1099 | { |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 1100 | struct throtl_data *td = key; |
Vivek Goyal | de701c7 | 2011-03-07 21:09:32 +0100 | [diff] [blame] | 1101 | struct throtl_grp *tg = tg_of_blkg(blkg); |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 1102 | |
Vivek Goyal | de701c7 | 2011-03-07 21:09:32 +0100 | [diff] [blame] | 1103 | tg->iops[WRITE] = write_iops; |
| 1104 | throtl_update_blkio_group_common(td, tg); |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 1105 | } |
| 1106 | |
Vivek Goyal | da52777 | 2011-03-02 19:05:33 -0500 | [diff] [blame] | 1107 | static void throtl_shutdown_wq(struct request_queue *q) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1108 | { |
| 1109 | struct throtl_data *td = q->td; |
| 1110 | |
| 1111 | cancel_delayed_work_sync(&td->throtl_work); |
| 1112 | } |
| 1113 | |
| 1114 | static struct blkio_policy_type blkio_policy_throtl = { |
| 1115 | .ops = { |
| 1116 | .blkio_unlink_group_fn = throtl_unlink_blkio_group, |
Tejun Heo | 72e06c2 | 2012-03-05 13:15:00 -0800 | [diff] [blame] | 1117 | .blkio_clear_queue_fn = throtl_clear_queue, |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1118 | .blkio_update_group_read_bps_fn = |
| 1119 | throtl_update_blkio_group_read_bps, |
| 1120 | .blkio_update_group_write_bps_fn = |
| 1121 | throtl_update_blkio_group_write_bps, |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 1122 | .blkio_update_group_read_iops_fn = |
| 1123 | throtl_update_blkio_group_read_iops, |
| 1124 | .blkio_update_group_write_iops_fn = |
| 1125 | throtl_update_blkio_group_write_iops, |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1126 | }, |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 1127 | .plid = BLKIO_POLICY_THROTL, |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1128 | }; |
| 1129 | |
Tejun Heo | bc16a4f | 2011-10-19 14:33:01 +0200 | [diff] [blame] | 1130 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1131 | { |
| 1132 | struct throtl_data *td = q->td; |
| 1133 | struct throtl_grp *tg; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1134 | bool rw = bio_data_dir(bio), update_disptime = true; |
Vivek Goyal | af75cd3 | 2011-05-19 15:38:31 -0400 | [diff] [blame] | 1135 | struct blkio_cgroup *blkcg; |
Tejun Heo | bc16a4f | 2011-10-19 14:33:01 +0200 | [diff] [blame] | 1136 | bool throttled = false; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1137 | |
| 1138 | if (bio->bi_rw & REQ_THROTTLED) { |
| 1139 | bio->bi_rw &= ~REQ_THROTTLED; |
Tejun Heo | bc16a4f | 2011-10-19 14:33:01 +0200 | [diff] [blame] | 1140 | goto out; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1141 | } |
| 1142 | |
Vivek Goyal | af75cd3 | 2011-05-19 15:38:31 -0400 | [diff] [blame] | 1143 | /* |
| 1144 | * A throtl_grp pointer retrieved under rcu can be used to access |
| 1145 | * basic fields like stats and io rates. If a group has no rules, |
| 1146 | * just update the dispatch stats in lockless manner and return. |
| 1147 | */ |
Vivek Goyal | af75cd3 | 2011-05-19 15:38:31 -0400 | [diff] [blame] | 1148 | rcu_read_lock(); |
| 1149 | blkcg = task_blkio_cgroup(current); |
| 1150 | tg = throtl_find_tg(td, blkcg); |
| 1151 | if (tg) { |
| 1152 | throtl_tg_fill_dev_details(td, tg); |
| 1153 | |
| 1154 | if (tg_no_rule_group(tg, rw)) { |
| 1155 | blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size, |
Shaohua Li | e5a94f5 | 2011-08-01 10:31:06 +0200 | [diff] [blame] | 1156 | rw, rw_is_sync(bio->bi_rw)); |
Tejun Heo | 2a7f124 | 2012-03-05 13:15:01 -0800 | [diff] [blame^] | 1157 | goto out_unlock_rcu; |
Vivek Goyal | af75cd3 | 2011-05-19 15:38:31 -0400 | [diff] [blame] | 1158 | } |
| 1159 | } |
Vivek Goyal | af75cd3 | 2011-05-19 15:38:31 -0400 | [diff] [blame] | 1160 | |
| 1161 | /* |
| 1162 | * Either group has not been allocated yet or it is not an unlimited |
| 1163 | * IO group |
| 1164 | */ |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1165 | spin_lock_irq(q->queue_lock); |
| 1166 | tg = throtl_get_tg(td); |
Tejun Heo | bc16a4f | 2011-10-19 14:33:01 +0200 | [diff] [blame] | 1167 | if (unlikely(!tg)) |
| 1168 | goto out_unlock; |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 1169 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1170 | if (tg->nr_queued[rw]) { |
| 1171 | /* |
| 1172 | * There is already another bio queued in same dir. No |
| 1173 | * need to update dispatch time. |
| 1174 | */ |
Vivek Goyal | 231d704 | 2011-03-07 21:05:14 +0100 | [diff] [blame] | 1175 | update_disptime = false; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1176 | goto queue_bio; |
Vivek Goyal | de701c7 | 2011-03-07 21:09:32 +0100 | [diff] [blame] | 1177 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1178 | } |
| 1179 | |
| 1180 | /* Bio is with-in rate limit of group */ |
| 1181 | if (tg_may_dispatch(td, tg, bio, NULL)) { |
| 1182 | throtl_charge_bio(tg, bio); |
Vivek Goyal | 04521db | 2011-03-22 21:54:29 +0100 | [diff] [blame] | 1183 | |
| 1184 | /* |
| 1185 | * We need to trim slice even when bios are not being queued |
| 1186 | * otherwise it might happen that a bio is not queued for |
| 1187 | * a long time and slice keeps on extending and trim is not |
| 1188 | * called for a long time. Now if limits are reduced suddenly |
| 1189 | * we take into account all the IO dispatched so far at new |
| 1190 | * low rate and * newly queued IO gets a really long dispatch |
| 1191 | * time. |
| 1192 | * |
| 1193 | * So keep on trimming slice even if bio is not queued. |
| 1194 | */ |
| 1195 | throtl_trim_slice(td, tg, rw); |
Tejun Heo | bc16a4f | 2011-10-19 14:33:01 +0200 | [diff] [blame] | 1196 | goto out_unlock; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1197 | } |
| 1198 | |
| 1199 | queue_bio: |
Joe Perches | fd16d26 | 2011-06-13 10:42:49 +0200 | [diff] [blame] | 1200 | throtl_log_tg(td, tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu" |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 1201 | " iodisp=%u iops=%u queued=%d/%d", |
| 1202 | rw == READ ? 'R' : 'W', |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1203 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 1204 | tg->io_disp[rw], tg->iops[rw], |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1205 | tg->nr_queued[READ], tg->nr_queued[WRITE]); |
| 1206 | |
| 1207 | throtl_add_bio_tg(q->td, tg, bio); |
Tejun Heo | bc16a4f | 2011-10-19 14:33:01 +0200 | [diff] [blame] | 1208 | throttled = true; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1209 | |
| 1210 | if (update_disptime) { |
| 1211 | tg_update_disptime(td, tg); |
| 1212 | throtl_schedule_next_dispatch(td); |
| 1213 | } |
| 1214 | |
Tejun Heo | bc16a4f | 2011-10-19 14:33:01 +0200 | [diff] [blame] | 1215 | out_unlock: |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1216 | spin_unlock_irq(q->queue_lock); |
Tejun Heo | 2a7f124 | 2012-03-05 13:15:01 -0800 | [diff] [blame^] | 1217 | out_unlock_rcu: |
| 1218 | rcu_read_unlock(); |
Tejun Heo | bc16a4f | 2011-10-19 14:33:01 +0200 | [diff] [blame] | 1219 | out: |
| 1220 | return throttled; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1221 | } |
| 1222 | |
Tejun Heo | c9a929d | 2011-10-19 14:42:16 +0200 | [diff] [blame] | 1223 | /** |
| 1224 | * blk_throtl_drain - drain throttled bios |
| 1225 | * @q: request_queue to drain throttled bios for |
| 1226 | * |
| 1227 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). |
| 1228 | */ |
| 1229 | void blk_throtl_drain(struct request_queue *q) |
| 1230 | __releases(q->queue_lock) __acquires(q->queue_lock) |
| 1231 | { |
| 1232 | struct throtl_data *td = q->td; |
| 1233 | struct throtl_rb_root *st = &td->tg_service_tree; |
| 1234 | struct throtl_grp *tg; |
| 1235 | struct bio_list bl; |
| 1236 | struct bio *bio; |
| 1237 | |
Jens Axboe | 334c2b0 | 2011-10-25 15:51:48 +0200 | [diff] [blame] | 1238 | WARN_ON_ONCE(!queue_is_locked(q)); |
Tejun Heo | c9a929d | 2011-10-19 14:42:16 +0200 | [diff] [blame] | 1239 | |
| 1240 | bio_list_init(&bl); |
| 1241 | |
| 1242 | while ((tg = throtl_rb_first(st))) { |
| 1243 | throtl_dequeue_tg(td, tg); |
| 1244 | |
| 1245 | while ((bio = bio_list_peek(&tg->bio_lists[READ]))) |
| 1246 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl); |
| 1247 | while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))) |
| 1248 | tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl); |
| 1249 | } |
| 1250 | spin_unlock_irq(q->queue_lock); |
| 1251 | |
| 1252 | while ((bio = bio_list_pop(&bl))) |
| 1253 | generic_make_request(bio); |
| 1254 | |
| 1255 | spin_lock_irq(q->queue_lock); |
| 1256 | } |
| 1257 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1258 | int blk_throtl_init(struct request_queue *q) |
| 1259 | { |
| 1260 | struct throtl_data *td; |
| 1261 | struct throtl_grp *tg; |
| 1262 | |
| 1263 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); |
| 1264 | if (!td) |
| 1265 | return -ENOMEM; |
| 1266 | |
| 1267 | INIT_HLIST_HEAD(&td->tg_list); |
| 1268 | td->tg_service_tree = THROTL_RB_ROOT; |
Vivek Goyal | de701c7 | 2011-03-07 21:09:32 +0100 | [diff] [blame] | 1269 | td->limits_changed = false; |
Vivek Goyal | a29a171 | 2011-05-19 15:38:19 -0400 | [diff] [blame] | 1270 | INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1271 | |
Vivek Goyal | 29b1258 | 2011-05-19 15:38:24 -0400 | [diff] [blame] | 1272 | /* alloc and Init root group. */ |
| 1273 | td->queue = q; |
| 1274 | tg = throtl_alloc_tg(td); |
Vivek Goyal | 02977e4 | 2010-10-01 14:49:48 +0200 | [diff] [blame] | 1275 | |
Vivek Goyal | 29b1258 | 2011-05-19 15:38:24 -0400 | [diff] [blame] | 1276 | if (!tg) { |
| 1277 | kfree(td); |
| 1278 | return -ENOMEM; |
| 1279 | } |
| 1280 | |
| 1281 | td->root_tg = tg; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1282 | |
| 1283 | rcu_read_lock(); |
Vivek Goyal | 5617cbe | 2011-05-19 15:38:26 -0400 | [diff] [blame] | 1284 | throtl_init_add_tg_lists(td, tg, &blkio_root_cgroup); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1285 | rcu_read_unlock(); |
| 1286 | |
| 1287 | /* Attach throtl data to request queue */ |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1288 | q->td = td; |
| 1289 | return 0; |
| 1290 | } |
| 1291 | |
| 1292 | void blk_throtl_exit(struct request_queue *q) |
| 1293 | { |
| 1294 | struct throtl_data *td = q->td; |
| 1295 | bool wait = false; |
| 1296 | |
| 1297 | BUG_ON(!td); |
| 1298 | |
Vivek Goyal | da52777 | 2011-03-02 19:05:33 -0500 | [diff] [blame] | 1299 | throtl_shutdown_wq(q); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1300 | |
| 1301 | spin_lock_irq(q->queue_lock); |
Tejun Heo | 72e06c2 | 2012-03-05 13:15:00 -0800 | [diff] [blame] | 1302 | throtl_release_tgs(td, true); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1303 | |
| 1304 | /* If there are other groups */ |
Vivek Goyal | 02977e4 | 2010-10-01 14:49:48 +0200 | [diff] [blame] | 1305 | if (td->nr_undestroyed_grps > 0) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1306 | wait = true; |
| 1307 | |
| 1308 | spin_unlock_irq(q->queue_lock); |
| 1309 | |
| 1310 | /* |
| 1311 | * Wait for tg->blkg->key accessors to exit their grace periods. |
| 1312 | * Do this wait only if there are other undestroyed groups out |
| 1313 | * there (other than root group). This can happen if cgroup deletion |
| 1314 | * path claimed the responsibility of cleaning up a group before |
| 1315 | * queue cleanup code get to the group. |
| 1316 | * |
| 1317 | * Do not call synchronize_rcu() unconditionally as there are drivers |
| 1318 | * which create/delete request queue hundreds of times during scan/boot |
| 1319 | * and synchronize_rcu() can take significant time and slow down boot. |
| 1320 | */ |
| 1321 | if (wait) |
| 1322 | synchronize_rcu(); |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 1323 | |
| 1324 | /* |
| 1325 | * Just being safe to make sure after previous flush if some body did |
| 1326 | * update limits through cgroup and another work got queued, cancel |
| 1327 | * it. |
| 1328 | */ |
Vivek Goyal | da52777 | 2011-03-02 19:05:33 -0500 | [diff] [blame] | 1329 | throtl_shutdown_wq(q); |
Tejun Heo | c9a929d | 2011-10-19 14:42:16 +0200 | [diff] [blame] | 1330 | } |
| 1331 | |
| 1332 | void blk_throtl_release(struct request_queue *q) |
| 1333 | { |
| 1334 | kfree(q->td); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1335 | } |
| 1336 | |
| 1337 | static int __init throtl_init(void) |
| 1338 | { |
Vivek Goyal | 450adcb | 2011-03-01 13:40:54 -0500 | [diff] [blame] | 1339 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
| 1340 | if (!kthrotld_workqueue) |
| 1341 | panic("Failed to create kthrotld\n"); |
| 1342 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1343 | blkio_policy_register(&blkio_policy_throtl); |
| 1344 | return 0; |
| 1345 | } |
| 1346 | |
| 1347 | module_init(throtl_init); |