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 | |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 24 | static struct blkcg_policy blkcg_policy_throtl; |
Tejun Heo | 0381411 | 2012-03-05 13:15:14 -0800 | [diff] [blame] | 25 | |
Vivek Goyal | 450adcb | 2011-03-01 13:40:54 -0500 | [diff] [blame] | 26 | /* A workqueue to queue throttle related work */ |
| 27 | static struct workqueue_struct *kthrotld_workqueue; |
Vivek Goyal | 450adcb | 2011-03-01 13:40:54 -0500 | [diff] [blame] | 28 | |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 29 | /* |
| 30 | * To implement hierarchical throttling, throtl_grps form a tree and bios |
| 31 | * are dispatched upwards level by level until they reach the top and get |
| 32 | * issued. When dispatching bios from the children and local group at each |
| 33 | * level, if the bios are dispatched into a single bio_list, there's a risk |
| 34 | * of a local or child group which can queue many bios at once filling up |
| 35 | * the list starving others. |
| 36 | * |
| 37 | * To avoid such starvation, dispatched bios are queued separately |
| 38 | * according to where they came from. When they are again dispatched to |
| 39 | * the parent, they're popped in round-robin order so that no single source |
| 40 | * hogs the dispatch window. |
| 41 | * |
| 42 | * throtl_qnode is used to keep the queued bios separated by their sources. |
| 43 | * Bios are queued to throtl_qnode which in turn is queued to |
| 44 | * throtl_service_queue and then dispatched in round-robin order. |
| 45 | * |
| 46 | * It's also used to track the reference counts on blkg's. A qnode always |
| 47 | * belongs to a throtl_grp and gets queued on itself or the parent, so |
| 48 | * incrementing the reference of the associated throtl_grp when a qnode is |
| 49 | * queued and decrementing when dequeued is enough to keep the whole blkg |
| 50 | * tree pinned while bios are in flight. |
| 51 | */ |
| 52 | struct throtl_qnode { |
| 53 | struct list_head node; /* service_queue->queued[] */ |
| 54 | struct bio_list bios; /* queued bios */ |
| 55 | struct throtl_grp *tg; /* tg this qnode belongs to */ |
| 56 | }; |
| 57 | |
Tejun Heo | c9e0332 | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 58 | struct throtl_service_queue { |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 59 | struct throtl_service_queue *parent_sq; /* the parent service_queue */ |
| 60 | |
Tejun Heo | 73f0d49 | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 61 | /* |
| 62 | * Bios queued directly to this service_queue or dispatched from |
| 63 | * children throtl_grp's. |
| 64 | */ |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 65 | struct list_head queued[2]; /* throtl_qnode [READ/WRITE] */ |
Tejun Heo | 73f0d49 | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 66 | unsigned int nr_queued[2]; /* number of queued bios */ |
| 67 | |
| 68 | /* |
| 69 | * RB tree of active children throtl_grp's, which are sorted by |
| 70 | * their ->disptime. |
| 71 | */ |
Tejun Heo | c9e0332 | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 72 | struct rb_root pending_tree; /* RB tree of active tgs */ |
| 73 | struct rb_node *first_pending; /* first node in the tree */ |
| 74 | unsigned int nr_pending; /* # queued in the tree */ |
| 75 | unsigned long first_pending_disptime; /* disptime of the first tg */ |
Tejun Heo | 69df0ab | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 76 | struct timer_list pending_timer; /* fires on first_pending_disptime */ |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 77 | }; |
| 78 | |
Tejun Heo | 5b2c16a | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 79 | enum tg_state_flags { |
| 80 | THROTL_TG_PENDING = 1 << 0, /* on parent's pending tree */ |
Tejun Heo | 0e9f416 | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 81 | THROTL_TG_WAS_EMPTY = 1 << 1, /* bio_lists[] became non-empty */ |
Tejun Heo | 5b2c16a | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 82 | }; |
| 83 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 84 | #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) |
| 85 | |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 86 | /* Per-cpu group stats */ |
| 87 | struct tg_stats_cpu { |
| 88 | /* total bytes transferred */ |
| 89 | struct blkg_rwstat service_bytes; |
| 90 | /* total IOs serviced, post merge */ |
| 91 | struct blkg_rwstat serviced; |
| 92 | }; |
| 93 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 94 | struct throtl_grp { |
Tejun Heo | f95a04a | 2012-04-16 13:57:26 -0700 | [diff] [blame] | 95 | /* must be the first member */ |
| 96 | struct blkg_policy_data pd; |
| 97 | |
Tejun Heo | c9e0332 | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 98 | /* active throtl group service_queue member */ |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 99 | struct rb_node rb_node; |
| 100 | |
Tejun Heo | 0f3457f | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 101 | /* throtl_data this group belongs to */ |
| 102 | struct throtl_data *td; |
| 103 | |
Tejun Heo | 49a2f1e | 2013-05-14 13:52:34 -0700 | [diff] [blame] | 104 | /* this group's service queue */ |
| 105 | struct throtl_service_queue service_queue; |
| 106 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 107 | /* |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 108 | * qnode_on_self is used when bios are directly queued to this |
| 109 | * throtl_grp so that local bios compete fairly with bios |
| 110 | * dispatched from children. qnode_on_parent is used when bios are |
| 111 | * dispatched from this throtl_grp into its parent and will compete |
| 112 | * with the sibling qnode_on_parents and the parent's |
| 113 | * qnode_on_self. |
| 114 | */ |
| 115 | struct throtl_qnode qnode_on_self[2]; |
| 116 | struct throtl_qnode qnode_on_parent[2]; |
| 117 | |
| 118 | /* |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 119 | * Dispatch time in jiffies. This is the estimated time when group |
| 120 | * will unthrottle and is ready to dispatch more bio. It is used as |
| 121 | * key to sort active groups in service tree. |
| 122 | */ |
| 123 | unsigned long disptime; |
| 124 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 125 | unsigned int flags; |
| 126 | |
Tejun Heo | 693e751 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 127 | /* are there any throtl rules between this group and td? */ |
| 128 | bool has_rules[2]; |
| 129 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 130 | /* bytes per second rate limits */ |
| 131 | uint64_t bps[2]; |
| 132 | |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 133 | /* IOPS limits */ |
| 134 | unsigned int iops[2]; |
| 135 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 136 | /* Number of bytes disptached in current slice */ |
| 137 | uint64_t bytes_disp[2]; |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 138 | /* Number of bio's dispatched in current slice */ |
| 139 | unsigned int io_disp[2]; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 140 | |
| 141 | /* When did we start a new slice */ |
| 142 | unsigned long slice_start[2]; |
| 143 | unsigned long slice_end[2]; |
Vivek Goyal | fe07143 | 2010-10-01 14:49:49 +0200 | [diff] [blame] | 144 | |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 145 | /* Per cpu stats pointer */ |
| 146 | struct tg_stats_cpu __percpu *stats_cpu; |
| 147 | |
| 148 | /* List of tgs waiting for per cpu stats memory to be allocated */ |
| 149 | struct list_head stats_alloc_node; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 150 | }; |
| 151 | |
| 152 | struct throtl_data |
| 153 | { |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 154 | /* service tree for active throtl groups */ |
Tejun Heo | c9e0332 | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 155 | struct throtl_service_queue service_queue; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 156 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 157 | struct request_queue *queue; |
| 158 | |
| 159 | /* Total Number of queued bios on READ and WRITE lists */ |
| 160 | unsigned int nr_queued[2]; |
| 161 | |
| 162 | /* |
Vivek Goyal | 02977e4 | 2010-10-01 14:49:48 +0200 | [diff] [blame] | 163 | * number of total undestroyed groups |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 164 | */ |
| 165 | unsigned int nr_undestroyed_grps; |
| 166 | |
| 167 | /* Work for dispatching throttled bios */ |
Tejun Heo | 69df0ab | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 168 | struct work_struct dispatch_work; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 169 | }; |
| 170 | |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 171 | /* list and work item to allocate percpu group stats */ |
| 172 | static DEFINE_SPINLOCK(tg_stats_alloc_lock); |
| 173 | static LIST_HEAD(tg_stats_alloc_list); |
| 174 | |
| 175 | static void tg_stats_alloc_fn(struct work_struct *); |
| 176 | static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn); |
| 177 | |
Tejun Heo | 69df0ab | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 178 | static void throtl_pending_timer_fn(unsigned long arg); |
| 179 | |
Tejun Heo | f95a04a | 2012-04-16 13:57:26 -0700 | [diff] [blame] | 180 | static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd) |
| 181 | { |
| 182 | return pd ? container_of(pd, struct throtl_grp, pd) : NULL; |
| 183 | } |
| 184 | |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 185 | static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg) |
Tejun Heo | 0381411 | 2012-03-05 13:15:14 -0800 | [diff] [blame] | 186 | { |
Tejun Heo | f95a04a | 2012-04-16 13:57:26 -0700 | [diff] [blame] | 187 | return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl)); |
Tejun Heo | 0381411 | 2012-03-05 13:15:14 -0800 | [diff] [blame] | 188 | } |
| 189 | |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 190 | static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) |
Tejun Heo | 0381411 | 2012-03-05 13:15:14 -0800 | [diff] [blame] | 191 | { |
Tejun Heo | f95a04a | 2012-04-16 13:57:26 -0700 | [diff] [blame] | 192 | return pd_to_blkg(&tg->pd); |
Tejun Heo | 0381411 | 2012-03-05 13:15:14 -0800 | [diff] [blame] | 193 | } |
| 194 | |
Tejun Heo | 03d8e11 | 2012-04-13 13:11:32 -0700 | [diff] [blame] | 195 | static inline struct throtl_grp *td_root_tg(struct throtl_data *td) |
| 196 | { |
| 197 | return blkg_to_tg(td->queue->root_blkg); |
| 198 | } |
| 199 | |
Tejun Heo | fda6f27 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 200 | /** |
| 201 | * sq_to_tg - return the throl_grp the specified service queue belongs to |
| 202 | * @sq: the throtl_service_queue of interest |
| 203 | * |
| 204 | * Return the throtl_grp @sq belongs to. If @sq is the top-level one |
| 205 | * embedded in throtl_data, %NULL is returned. |
| 206 | */ |
| 207 | static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq) |
| 208 | { |
| 209 | if (sq && sq->parent_sq) |
| 210 | return container_of(sq, struct throtl_grp, service_queue); |
| 211 | else |
| 212 | return NULL; |
| 213 | } |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 214 | |
Tejun Heo | fda6f27 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 215 | /** |
| 216 | * sq_to_td - return throtl_data the specified service queue belongs to |
| 217 | * @sq: the throtl_service_queue of interest |
| 218 | * |
| 219 | * A service_queue can be embeded in either a throtl_grp or throtl_data. |
| 220 | * Determine the associated throtl_data accordingly and return it. |
| 221 | */ |
| 222 | static struct throtl_data *sq_to_td(struct throtl_service_queue *sq) |
| 223 | { |
| 224 | struct throtl_grp *tg = sq_to_tg(sq); |
| 225 | |
| 226 | if (tg) |
| 227 | return tg->td; |
| 228 | else |
| 229 | return container_of(sq, struct throtl_data, service_queue); |
| 230 | } |
| 231 | |
| 232 | /** |
| 233 | * throtl_log - log debug message via blktrace |
| 234 | * @sq: the service_queue being reported |
| 235 | * @fmt: printf format string |
| 236 | * @args: printf args |
| 237 | * |
| 238 | * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a |
| 239 | * throtl_grp; otherwise, just "throtl". |
| 240 | * |
| 241 | * TODO: this should be made a function and name formatting should happen |
| 242 | * after testing whether blktrace is enabled. |
| 243 | */ |
| 244 | #define throtl_log(sq, fmt, args...) do { \ |
| 245 | struct throtl_grp *__tg = sq_to_tg((sq)); \ |
| 246 | struct throtl_data *__td = sq_to_td((sq)); \ |
| 247 | \ |
| 248 | (void)__td; \ |
| 249 | if ((__tg)) { \ |
| 250 | char __pbuf[128]; \ |
| 251 | \ |
| 252 | blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf)); \ |
| 253 | blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \ |
| 254 | } else { \ |
| 255 | blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \ |
| 256 | } \ |
| 257 | } while (0) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 258 | |
Peter Zijlstra | 90d3839 | 2013-11-12 19:42:14 -0800 | [diff] [blame] | 259 | static void tg_stats_init(struct tg_stats_cpu *tg_stats) |
| 260 | { |
| 261 | blkg_rwstat_init(&tg_stats->service_bytes); |
| 262 | blkg_rwstat_init(&tg_stats->serviced); |
| 263 | } |
| 264 | |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 265 | /* |
| 266 | * Worker for allocating per cpu stat for tgs. This is scheduled on the |
Tejun Heo | 3b07e9c | 2012-08-20 14:51:24 -0700 | [diff] [blame] | 267 | * system_wq once there are some groups on the alloc_list waiting for |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 268 | * allocation. |
| 269 | */ |
| 270 | static void tg_stats_alloc_fn(struct work_struct *work) |
| 271 | { |
| 272 | static struct tg_stats_cpu *stats_cpu; /* this fn is non-reentrant */ |
| 273 | struct delayed_work *dwork = to_delayed_work(work); |
| 274 | bool empty = false; |
| 275 | |
| 276 | alloc_stats: |
| 277 | if (!stats_cpu) { |
Peter Zijlstra | 90d3839 | 2013-11-12 19:42:14 -0800 | [diff] [blame] | 278 | int cpu; |
| 279 | |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 280 | stats_cpu = alloc_percpu(struct tg_stats_cpu); |
| 281 | if (!stats_cpu) { |
| 282 | /* allocation failed, try again after some time */ |
Tejun Heo | 3b07e9c | 2012-08-20 14:51:24 -0700 | [diff] [blame] | 283 | schedule_delayed_work(dwork, msecs_to_jiffies(10)); |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 284 | return; |
| 285 | } |
Peter Zijlstra | 90d3839 | 2013-11-12 19:42:14 -0800 | [diff] [blame] | 286 | for_each_possible_cpu(cpu) |
| 287 | tg_stats_init(per_cpu_ptr(stats_cpu, cpu)); |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 288 | } |
| 289 | |
| 290 | spin_lock_irq(&tg_stats_alloc_lock); |
| 291 | |
| 292 | if (!list_empty(&tg_stats_alloc_list)) { |
| 293 | struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list, |
| 294 | struct throtl_grp, |
| 295 | stats_alloc_node); |
| 296 | swap(tg->stats_cpu, stats_cpu); |
| 297 | list_del_init(&tg->stats_alloc_node); |
| 298 | } |
| 299 | |
| 300 | empty = list_empty(&tg_stats_alloc_list); |
| 301 | spin_unlock_irq(&tg_stats_alloc_lock); |
| 302 | if (!empty) |
| 303 | goto alloc_stats; |
| 304 | } |
| 305 | |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 306 | static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg) |
| 307 | { |
| 308 | INIT_LIST_HEAD(&qn->node); |
| 309 | bio_list_init(&qn->bios); |
| 310 | qn->tg = tg; |
| 311 | } |
| 312 | |
| 313 | /** |
| 314 | * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it |
| 315 | * @bio: bio being added |
| 316 | * @qn: qnode to add bio to |
| 317 | * @queued: the service_queue->queued[] list @qn belongs to |
| 318 | * |
| 319 | * Add @bio to @qn and put @qn on @queued if it's not already on. |
| 320 | * @qn->tg's reference count is bumped when @qn is activated. See the |
| 321 | * comment on top of throtl_qnode definition for details. |
| 322 | */ |
| 323 | static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn, |
| 324 | struct list_head *queued) |
| 325 | { |
| 326 | bio_list_add(&qn->bios, bio); |
| 327 | if (list_empty(&qn->node)) { |
| 328 | list_add_tail(&qn->node, queued); |
| 329 | blkg_get(tg_to_blkg(qn->tg)); |
| 330 | } |
| 331 | } |
| 332 | |
| 333 | /** |
| 334 | * throtl_peek_queued - peek the first bio on a qnode list |
| 335 | * @queued: the qnode list to peek |
| 336 | */ |
| 337 | static struct bio *throtl_peek_queued(struct list_head *queued) |
| 338 | { |
| 339 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); |
| 340 | struct bio *bio; |
| 341 | |
| 342 | if (list_empty(queued)) |
| 343 | return NULL; |
| 344 | |
| 345 | bio = bio_list_peek(&qn->bios); |
| 346 | WARN_ON_ONCE(!bio); |
| 347 | return bio; |
| 348 | } |
| 349 | |
| 350 | /** |
| 351 | * throtl_pop_queued - pop the first bio form a qnode list |
| 352 | * @queued: the qnode list to pop a bio from |
| 353 | * @tg_to_put: optional out argument for throtl_grp to put |
| 354 | * |
| 355 | * Pop the first bio from the qnode list @queued. After popping, the first |
| 356 | * qnode is removed from @queued if empty or moved to the end of @queued so |
| 357 | * that the popping order is round-robin. |
| 358 | * |
| 359 | * When the first qnode is removed, its associated throtl_grp should be put |
| 360 | * too. If @tg_to_put is NULL, this function automatically puts it; |
| 361 | * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is |
| 362 | * responsible for putting it. |
| 363 | */ |
| 364 | static struct bio *throtl_pop_queued(struct list_head *queued, |
| 365 | struct throtl_grp **tg_to_put) |
| 366 | { |
| 367 | struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node); |
| 368 | struct bio *bio; |
| 369 | |
| 370 | if (list_empty(queued)) |
| 371 | return NULL; |
| 372 | |
| 373 | bio = bio_list_pop(&qn->bios); |
| 374 | WARN_ON_ONCE(!bio); |
| 375 | |
| 376 | if (bio_list_empty(&qn->bios)) { |
| 377 | list_del_init(&qn->node); |
| 378 | if (tg_to_put) |
| 379 | *tg_to_put = qn->tg; |
| 380 | else |
| 381 | blkg_put(tg_to_blkg(qn->tg)); |
| 382 | } else { |
| 383 | list_move_tail(&qn->node, queued); |
| 384 | } |
| 385 | |
| 386 | return bio; |
| 387 | } |
| 388 | |
Tejun Heo | 49a2f1e | 2013-05-14 13:52:34 -0700 | [diff] [blame] | 389 | /* init a service_queue, assumes the caller zeroed it */ |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 390 | static void throtl_service_queue_init(struct throtl_service_queue *sq, |
| 391 | struct throtl_service_queue *parent_sq) |
Tejun Heo | 49a2f1e | 2013-05-14 13:52:34 -0700 | [diff] [blame] | 392 | { |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 393 | INIT_LIST_HEAD(&sq->queued[0]); |
| 394 | INIT_LIST_HEAD(&sq->queued[1]); |
Tejun Heo | 49a2f1e | 2013-05-14 13:52:34 -0700 | [diff] [blame] | 395 | sq->pending_tree = RB_ROOT; |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 396 | sq->parent_sq = parent_sq; |
Tejun Heo | 69df0ab | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 397 | setup_timer(&sq->pending_timer, throtl_pending_timer_fn, |
| 398 | (unsigned long)sq); |
| 399 | } |
| 400 | |
| 401 | static void throtl_service_queue_exit(struct throtl_service_queue *sq) |
| 402 | { |
| 403 | del_timer_sync(&sq->pending_timer); |
Tejun Heo | 49a2f1e | 2013-05-14 13:52:34 -0700 | [diff] [blame] | 404 | } |
| 405 | |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 406 | static void throtl_pd_init(struct blkcg_gq *blkg) |
Vivek Goyal | a29a171 | 2011-05-19 15:38:19 -0400 | [diff] [blame] | 407 | { |
Tejun Heo | 0381411 | 2012-03-05 13:15:14 -0800 | [diff] [blame] | 408 | struct throtl_grp *tg = blkg_to_tg(blkg); |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 409 | struct throtl_data *td = blkg->q->td; |
Tejun Heo | 9138125 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 410 | struct throtl_service_queue *parent_sq; |
Tejun Heo | ff26eaa | 2012-05-23 12:16:21 +0200 | [diff] [blame] | 411 | unsigned long flags; |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 412 | int rw; |
Tejun Heo | cd1604f | 2012-03-05 13:15:06 -0800 | [diff] [blame] | 413 | |
Tejun Heo | 9138125 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 414 | /* |
| 415 | * If sane_hierarchy is enabled, we switch to properly hierarchical |
| 416 | * behavior where limits on a given throtl_grp are applied to the |
| 417 | * whole subtree rather than just the group itself. e.g. If 16M |
| 418 | * read_bps limit is set on the root group, the whole system can't |
| 419 | * exceed 16M for the device. |
| 420 | * |
| 421 | * If sane_hierarchy is not enabled, the broken flat hierarchy |
| 422 | * behavior is retained where all throtl_grps are treated as if |
| 423 | * they're all separate root groups right below throtl_data. |
| 424 | * Limits of a group don't interact with limits of other groups |
| 425 | * regardless of the position of the group in the hierarchy. |
| 426 | */ |
| 427 | parent_sq = &td->service_queue; |
| 428 | |
| 429 | if (cgroup_sane_behavior(blkg->blkcg->css.cgroup) && blkg->parent) |
| 430 | parent_sq = &blkg_to_tg(blkg->parent)->service_queue; |
| 431 | |
| 432 | throtl_service_queue_init(&tg->service_queue, parent_sq); |
| 433 | |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 434 | for (rw = READ; rw <= WRITE; rw++) { |
| 435 | throtl_qnode_init(&tg->qnode_on_self[rw], tg); |
| 436 | throtl_qnode_init(&tg->qnode_on_parent[rw], tg); |
| 437 | } |
| 438 | |
Vivek Goyal | a29a171 | 2011-05-19 15:38:19 -0400 | [diff] [blame] | 439 | RB_CLEAR_NODE(&tg->rb_node); |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 440 | tg->td = td; |
Vivek Goyal | a29a171 | 2011-05-19 15:38:19 -0400 | [diff] [blame] | 441 | |
Tejun Heo | e56da7e | 2012-03-05 13:15:07 -0800 | [diff] [blame] | 442 | tg->bps[READ] = -1; |
| 443 | tg->bps[WRITE] = -1; |
| 444 | tg->iops[READ] = -1; |
| 445 | tg->iops[WRITE] = -1; |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 446 | |
| 447 | /* |
| 448 | * Ugh... We need to perform per-cpu allocation for tg->stats_cpu |
| 449 | * but percpu allocator can't be called from IO path. Queue tg on |
| 450 | * tg_stats_alloc_list and allocate from work item. |
| 451 | */ |
Tejun Heo | ff26eaa | 2012-05-23 12:16:21 +0200 | [diff] [blame] | 452 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 453 | list_add(&tg->stats_alloc_node, &tg_stats_alloc_list); |
Tejun Heo | 3b07e9c | 2012-08-20 14:51:24 -0700 | [diff] [blame] | 454 | schedule_delayed_work(&tg_stats_alloc_work, 0); |
Tejun Heo | ff26eaa | 2012-05-23 12:16:21 +0200 | [diff] [blame] | 455 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 456 | } |
| 457 | |
Tejun Heo | 693e751 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 458 | /* |
| 459 | * Set has_rules[] if @tg or any of its parents have limits configured. |
| 460 | * This doesn't require walking up to the top of the hierarchy as the |
| 461 | * parent's has_rules[] is guaranteed to be correct. |
| 462 | */ |
| 463 | static void tg_update_has_rules(struct throtl_grp *tg) |
| 464 | { |
| 465 | struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq); |
| 466 | int rw; |
| 467 | |
| 468 | for (rw = READ; rw <= WRITE; rw++) |
| 469 | tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) || |
| 470 | (tg->bps[rw] != -1 || tg->iops[rw] != -1); |
| 471 | } |
| 472 | |
| 473 | static void throtl_pd_online(struct blkcg_gq *blkg) |
| 474 | { |
| 475 | /* |
| 476 | * We don't want new groups to escape the limits of its ancestors. |
| 477 | * Update has_rules[] after a new group is brought online. |
| 478 | */ |
| 479 | tg_update_has_rules(blkg_to_tg(blkg)); |
| 480 | } |
| 481 | |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 482 | static void throtl_pd_exit(struct blkcg_gq *blkg) |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 483 | { |
| 484 | struct throtl_grp *tg = blkg_to_tg(blkg); |
Tejun Heo | ff26eaa | 2012-05-23 12:16:21 +0200 | [diff] [blame] | 485 | unsigned long flags; |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 486 | |
Tejun Heo | ff26eaa | 2012-05-23 12:16:21 +0200 | [diff] [blame] | 487 | spin_lock_irqsave(&tg_stats_alloc_lock, flags); |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 488 | list_del_init(&tg->stats_alloc_node); |
Tejun Heo | ff26eaa | 2012-05-23 12:16:21 +0200 | [diff] [blame] | 489 | spin_unlock_irqrestore(&tg_stats_alloc_lock, flags); |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 490 | |
| 491 | free_percpu(tg->stats_cpu); |
Tejun Heo | 69df0ab | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 492 | |
| 493 | throtl_service_queue_exit(&tg->service_queue); |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 494 | } |
| 495 | |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 496 | static void throtl_pd_reset_stats(struct blkcg_gq *blkg) |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 497 | { |
| 498 | struct throtl_grp *tg = blkg_to_tg(blkg); |
| 499 | int cpu; |
| 500 | |
| 501 | if (tg->stats_cpu == NULL) |
| 502 | return; |
| 503 | |
| 504 | for_each_possible_cpu(cpu) { |
| 505 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); |
| 506 | |
| 507 | blkg_rwstat_reset(&sc->service_bytes); |
| 508 | blkg_rwstat_reset(&sc->serviced); |
| 509 | } |
Vivek Goyal | a29a171 | 2011-05-19 15:38:19 -0400 | [diff] [blame] | 510 | } |
| 511 | |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 512 | static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td, |
| 513 | struct blkcg *blkcg) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 514 | { |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 515 | /* |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 516 | * This is the common case when there are no blkcgs. Avoid lookup |
| 517 | * in this case |
Tejun Heo | cd1604f | 2012-03-05 13:15:06 -0800 | [diff] [blame] | 518 | */ |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 519 | if (blkcg == &blkcg_root) |
Tejun Heo | 03d8e11 | 2012-04-13 13:11:32 -0700 | [diff] [blame] | 520 | return td_root_tg(td); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 521 | |
Tejun Heo | e8989fa | 2012-03-05 13:15:20 -0800 | [diff] [blame] | 522 | return blkg_to_tg(blkg_lookup(blkcg, td->queue)); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 523 | } |
| 524 | |
Tejun Heo | cd1604f | 2012-03-05 13:15:06 -0800 | [diff] [blame] | 525 | static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td, |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 526 | struct blkcg *blkcg) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 527 | { |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 528 | struct request_queue *q = td->queue; |
Tejun Heo | cd1604f | 2012-03-05 13:15:06 -0800 | [diff] [blame] | 529 | struct throtl_grp *tg = NULL; |
Tejun Heo | 0a5a7d0 | 2012-03-05 13:15:02 -0800 | [diff] [blame] | 530 | |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 531 | /* |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 532 | * This is the common case when there are no blkcgs. Avoid lookup |
| 533 | * in this case |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 534 | */ |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 535 | if (blkcg == &blkcg_root) { |
Tejun Heo | 03d8e11 | 2012-04-13 13:11:32 -0700 | [diff] [blame] | 536 | tg = td_root_tg(td); |
Tejun Heo | cd1604f | 2012-03-05 13:15:06 -0800 | [diff] [blame] | 537 | } else { |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 538 | struct blkcg_gq *blkg; |
Tejun Heo | cd1604f | 2012-03-05 13:15:06 -0800 | [diff] [blame] | 539 | |
Tejun Heo | 3c96cb3 | 2012-04-13 13:11:34 -0700 | [diff] [blame] | 540 | blkg = blkg_lookup_create(blkcg, q); |
Tejun Heo | cd1604f | 2012-03-05 13:15:06 -0800 | [diff] [blame] | 541 | |
| 542 | /* if %NULL and @q is alive, fall back to root_tg */ |
| 543 | if (!IS_ERR(blkg)) |
Tejun Heo | 0381411 | 2012-03-05 13:15:14 -0800 | [diff] [blame] | 544 | tg = blkg_to_tg(blkg); |
Bart Van Assche | 3f3299d | 2012-11-28 13:42:38 +0100 | [diff] [blame] | 545 | else if (!blk_queue_dying(q)) |
Tejun Heo | 03d8e11 | 2012-04-13 13:11:32 -0700 | [diff] [blame] | 546 | tg = td_root_tg(td); |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 547 | } |
| 548 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 549 | return tg; |
| 550 | } |
| 551 | |
Tejun Heo | 0049af7 | 2013-05-14 13:52:33 -0700 | [diff] [blame] | 552 | static struct throtl_grp * |
| 553 | throtl_rb_first(struct throtl_service_queue *parent_sq) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 554 | { |
| 555 | /* Service tree is empty */ |
Tejun Heo | 0049af7 | 2013-05-14 13:52:33 -0700 | [diff] [blame] | 556 | if (!parent_sq->nr_pending) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 557 | return NULL; |
| 558 | |
Tejun Heo | 0049af7 | 2013-05-14 13:52:33 -0700 | [diff] [blame] | 559 | if (!parent_sq->first_pending) |
| 560 | parent_sq->first_pending = rb_first(&parent_sq->pending_tree); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 561 | |
Tejun Heo | 0049af7 | 2013-05-14 13:52:33 -0700 | [diff] [blame] | 562 | if (parent_sq->first_pending) |
| 563 | return rb_entry_tg(parent_sq->first_pending); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 564 | |
| 565 | return NULL; |
| 566 | } |
| 567 | |
| 568 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) |
| 569 | { |
| 570 | rb_erase(n, root); |
| 571 | RB_CLEAR_NODE(n); |
| 572 | } |
| 573 | |
Tejun Heo | 0049af7 | 2013-05-14 13:52:33 -0700 | [diff] [blame] | 574 | static void throtl_rb_erase(struct rb_node *n, |
| 575 | struct throtl_service_queue *parent_sq) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 576 | { |
Tejun Heo | 0049af7 | 2013-05-14 13:52:33 -0700 | [diff] [blame] | 577 | if (parent_sq->first_pending == n) |
| 578 | parent_sq->first_pending = NULL; |
| 579 | rb_erase_init(n, &parent_sq->pending_tree); |
| 580 | --parent_sq->nr_pending; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 581 | } |
| 582 | |
Tejun Heo | 0049af7 | 2013-05-14 13:52:33 -0700 | [diff] [blame] | 583 | static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 584 | { |
| 585 | struct throtl_grp *tg; |
| 586 | |
Tejun Heo | 0049af7 | 2013-05-14 13:52:33 -0700 | [diff] [blame] | 587 | tg = throtl_rb_first(parent_sq); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 588 | if (!tg) |
| 589 | return; |
| 590 | |
Tejun Heo | 0049af7 | 2013-05-14 13:52:33 -0700 | [diff] [blame] | 591 | parent_sq->first_pending_disptime = tg->disptime; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 592 | } |
| 593 | |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 594 | static void tg_service_queue_add(struct throtl_grp *tg) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 595 | { |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 596 | struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq; |
Tejun Heo | 0049af7 | 2013-05-14 13:52:33 -0700 | [diff] [blame] | 597 | struct rb_node **node = &parent_sq->pending_tree.rb_node; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 598 | struct rb_node *parent = NULL; |
| 599 | struct throtl_grp *__tg; |
| 600 | unsigned long key = tg->disptime; |
| 601 | int left = 1; |
| 602 | |
| 603 | while (*node != NULL) { |
| 604 | parent = *node; |
| 605 | __tg = rb_entry_tg(parent); |
| 606 | |
| 607 | if (time_before(key, __tg->disptime)) |
| 608 | node = &parent->rb_left; |
| 609 | else { |
| 610 | node = &parent->rb_right; |
| 611 | left = 0; |
| 612 | } |
| 613 | } |
| 614 | |
| 615 | if (left) |
Tejun Heo | 0049af7 | 2013-05-14 13:52:33 -0700 | [diff] [blame] | 616 | parent_sq->first_pending = &tg->rb_node; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 617 | |
| 618 | rb_link_node(&tg->rb_node, parent, node); |
Tejun Heo | 0049af7 | 2013-05-14 13:52:33 -0700 | [diff] [blame] | 619 | rb_insert_color(&tg->rb_node, &parent_sq->pending_tree); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 620 | } |
| 621 | |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 622 | static void __throtl_enqueue_tg(struct throtl_grp *tg) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 623 | { |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 624 | tg_service_queue_add(tg); |
Tejun Heo | 5b2c16a | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 625 | tg->flags |= THROTL_TG_PENDING; |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 626 | tg->service_queue.parent_sq->nr_pending++; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 627 | } |
| 628 | |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 629 | static void throtl_enqueue_tg(struct throtl_grp *tg) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 630 | { |
Tejun Heo | 5b2c16a | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 631 | if (!(tg->flags & THROTL_TG_PENDING)) |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 632 | __throtl_enqueue_tg(tg); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 633 | } |
| 634 | |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 635 | static void __throtl_dequeue_tg(struct throtl_grp *tg) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 636 | { |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 637 | throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq); |
Tejun Heo | 5b2c16a | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 638 | tg->flags &= ~THROTL_TG_PENDING; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 639 | } |
| 640 | |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 641 | static void throtl_dequeue_tg(struct throtl_grp *tg) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 642 | { |
Tejun Heo | 5b2c16a | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 643 | if (tg->flags & THROTL_TG_PENDING) |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 644 | __throtl_dequeue_tg(tg); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 645 | } |
| 646 | |
Tejun Heo | a9131a2 | 2013-05-14 13:52:31 -0700 | [diff] [blame] | 647 | /* Call with queue lock held */ |
Tejun Heo | 69df0ab | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 648 | static void throtl_schedule_pending_timer(struct throtl_service_queue *sq, |
| 649 | unsigned long expires) |
Tejun Heo | a9131a2 | 2013-05-14 13:52:31 -0700 | [diff] [blame] | 650 | { |
Tejun Heo | 69df0ab | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 651 | mod_timer(&sq->pending_timer, expires); |
| 652 | throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu", |
| 653 | expires - jiffies, jiffies); |
Tejun Heo | a9131a2 | 2013-05-14 13:52:31 -0700 | [diff] [blame] | 654 | } |
| 655 | |
Tejun Heo | 7f52f98 | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 656 | /** |
| 657 | * throtl_schedule_next_dispatch - schedule the next dispatch cycle |
| 658 | * @sq: the service_queue to schedule dispatch for |
| 659 | * @force: force scheduling |
| 660 | * |
| 661 | * Arm @sq->pending_timer so that the next dispatch cycle starts on the |
| 662 | * dispatch time of the first pending child. Returns %true if either timer |
| 663 | * is armed or there's no pending child left. %false if the current |
| 664 | * dispatch window is still open and the caller should continue |
| 665 | * dispatching. |
| 666 | * |
| 667 | * If @force is %true, the dispatch timer is always scheduled and this |
| 668 | * function is guaranteed to return %true. This is to be used when the |
| 669 | * caller can't dispatch itself and needs to invoke pending_timer |
| 670 | * unconditionally. Note that forced scheduling is likely to induce short |
| 671 | * delay before dispatch starts even if @sq->first_pending_disptime is not |
| 672 | * in the future and thus shouldn't be used in hot paths. |
| 673 | */ |
| 674 | static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq, |
| 675 | bool force) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 676 | { |
Tejun Heo | 6a52560 | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 677 | /* any pending children left? */ |
Tejun Heo | c9e0332 | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 678 | if (!sq->nr_pending) |
Tejun Heo | 7f52f98 | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 679 | return true; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 680 | |
Tejun Heo | c9e0332 | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 681 | update_min_dispatch_time(sq); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 682 | |
Tejun Heo | 69df0ab | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 683 | /* is the next dispatch time in the future? */ |
Tejun Heo | 7f52f98 | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 684 | if (force || time_after(sq->first_pending_disptime, jiffies)) { |
Tejun Heo | 69df0ab | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 685 | throtl_schedule_pending_timer(sq, sq->first_pending_disptime); |
Tejun Heo | 7f52f98 | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 686 | return true; |
Tejun Heo | 69df0ab | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 687 | } |
| 688 | |
Tejun Heo | 7f52f98 | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 689 | /* tell the caller to continue dispatching */ |
| 690 | return false; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 691 | } |
| 692 | |
Vivek Goyal | 32ee5bc | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 693 | static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg, |
| 694 | bool rw, unsigned long start) |
| 695 | { |
| 696 | tg->bytes_disp[rw] = 0; |
| 697 | tg->io_disp[rw] = 0; |
| 698 | |
| 699 | /* |
| 700 | * Previous slice has expired. We must have trimmed it after last |
| 701 | * bio dispatch. That means since start of last slice, we never used |
| 702 | * that bandwidth. Do try to make use of that bandwidth while giving |
| 703 | * credit. |
| 704 | */ |
| 705 | if (time_after_eq(start, tg->slice_start[rw])) |
| 706 | tg->slice_start[rw] = start; |
| 707 | |
| 708 | tg->slice_end[rw] = jiffies + throtl_slice; |
| 709 | throtl_log(&tg->service_queue, |
| 710 | "[%c] new slice with credit start=%lu end=%lu jiffies=%lu", |
| 711 | rw == READ ? 'R' : 'W', tg->slice_start[rw], |
| 712 | tg->slice_end[rw], jiffies); |
| 713 | } |
| 714 | |
Tejun Heo | 0f3457f | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 715 | static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 716 | { |
| 717 | tg->bytes_disp[rw] = 0; |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 718 | tg->io_disp[rw] = 0; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 719 | tg->slice_start[rw] = jiffies; |
| 720 | tg->slice_end[rw] = jiffies + throtl_slice; |
Tejun Heo | fda6f27 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 721 | throtl_log(&tg->service_queue, |
| 722 | "[%c] new slice start=%lu end=%lu jiffies=%lu", |
| 723 | rw == READ ? 'R' : 'W', tg->slice_start[rw], |
| 724 | tg->slice_end[rw], jiffies); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 725 | } |
| 726 | |
Tejun Heo | 0f3457f | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 727 | static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, |
| 728 | unsigned long jiffy_end) |
Vivek Goyal | d1ae8ff | 2010-12-01 19:34:46 +0100 | [diff] [blame] | 729 | { |
| 730 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); |
| 731 | } |
| 732 | |
Tejun Heo | 0f3457f | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 733 | static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, |
| 734 | unsigned long jiffy_end) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 735 | { |
| 736 | tg->slice_end[rw] = roundup(jiffy_end, throtl_slice); |
Tejun Heo | fda6f27 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 737 | throtl_log(&tg->service_queue, |
| 738 | "[%c] extend slice start=%lu end=%lu jiffies=%lu", |
| 739 | rw == READ ? 'R' : 'W', tg->slice_start[rw], |
| 740 | tg->slice_end[rw], jiffies); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 741 | } |
| 742 | |
| 743 | /* Determine if previously allocated or extended slice is complete or not */ |
Tejun Heo | 0f3457f | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 744 | static bool throtl_slice_used(struct throtl_grp *tg, bool rw) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 745 | { |
| 746 | if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) |
| 747 | return 0; |
| 748 | |
| 749 | return 1; |
| 750 | } |
| 751 | |
| 752 | /* Trim the used slices and adjust slice start accordingly */ |
Tejun Heo | 0f3457f | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 753 | static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 754 | { |
Vivek Goyal | 3aad5d3 | 2010-10-01 14:51:14 +0200 | [diff] [blame] | 755 | unsigned long nr_slices, time_elapsed, io_trim; |
| 756 | u64 bytes_trim, tmp; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 757 | |
| 758 | BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); |
| 759 | |
| 760 | /* |
| 761 | * If bps are unlimited (-1), then time slice don't get |
| 762 | * renewed. Don't try to trim the slice if slice is used. A new |
| 763 | * slice will start when appropriate. |
| 764 | */ |
Tejun Heo | 0f3457f | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 765 | if (throtl_slice_used(tg, rw)) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 766 | return; |
| 767 | |
Vivek Goyal | d1ae8ff | 2010-12-01 19:34:46 +0100 | [diff] [blame] | 768 | /* |
| 769 | * A bio has been dispatched. Also adjust slice_end. It might happen |
| 770 | * that initially cgroup limit was very low resulting in high |
| 771 | * slice_end, but later limit was bumped up and bio was dispached |
| 772 | * sooner, then we need to reduce slice_end. A high bogus slice_end |
| 773 | * is bad because it does not allow new slice to start. |
| 774 | */ |
| 775 | |
Tejun Heo | 0f3457f | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 776 | throtl_set_slice_end(tg, rw, jiffies + throtl_slice); |
Vivek Goyal | d1ae8ff | 2010-12-01 19:34:46 +0100 | [diff] [blame] | 777 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 778 | time_elapsed = jiffies - tg->slice_start[rw]; |
| 779 | |
| 780 | nr_slices = time_elapsed / throtl_slice; |
| 781 | |
| 782 | if (!nr_slices) |
| 783 | return; |
Vivek Goyal | 3aad5d3 | 2010-10-01 14:51:14 +0200 | [diff] [blame] | 784 | tmp = tg->bps[rw] * throtl_slice * nr_slices; |
| 785 | do_div(tmp, HZ); |
| 786 | bytes_trim = tmp; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 787 | |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 788 | io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 789 | |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 790 | if (!bytes_trim && !io_trim) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 791 | return; |
| 792 | |
| 793 | if (tg->bytes_disp[rw] >= bytes_trim) |
| 794 | tg->bytes_disp[rw] -= bytes_trim; |
| 795 | else |
| 796 | tg->bytes_disp[rw] = 0; |
| 797 | |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 798 | if (tg->io_disp[rw] >= io_trim) |
| 799 | tg->io_disp[rw] -= io_trim; |
| 800 | else |
| 801 | tg->io_disp[rw] = 0; |
| 802 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 803 | tg->slice_start[rw] += nr_slices * throtl_slice; |
| 804 | |
Tejun Heo | fda6f27 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 805 | throtl_log(&tg->service_queue, |
| 806 | "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu", |
| 807 | rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim, |
| 808 | tg->slice_start[rw], tg->slice_end[rw], jiffies); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 809 | } |
| 810 | |
Tejun Heo | 0f3457f | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 811 | static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio, |
| 812 | unsigned long *wait) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 813 | { |
| 814 | bool rw = bio_data_dir(bio); |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 815 | unsigned int io_allowed; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 816 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
Vivek Goyal | c49c06e | 2010-10-01 21:16:42 +0200 | [diff] [blame] | 817 | u64 tmp; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 818 | |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 819 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 820 | |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 821 | /* Slice has just started. Consider one slice interval */ |
| 822 | if (!jiffy_elapsed) |
| 823 | jiffy_elapsed_rnd = throtl_slice; |
| 824 | |
| 825 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); |
| 826 | |
Vivek Goyal | c49c06e | 2010-10-01 21:16:42 +0200 | [diff] [blame] | 827 | /* |
| 828 | * jiffy_elapsed_rnd should not be a big value as minimum iops can be |
| 829 | * 1 then at max jiffy elapsed should be equivalent of 1 second as we |
| 830 | * will allow dispatch after 1 second and after that slice should |
| 831 | * have been trimmed. |
| 832 | */ |
| 833 | |
| 834 | tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd; |
| 835 | do_div(tmp, HZ); |
| 836 | |
| 837 | if (tmp > UINT_MAX) |
| 838 | io_allowed = UINT_MAX; |
| 839 | else |
| 840 | io_allowed = tmp; |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 841 | |
| 842 | if (tg->io_disp[rw] + 1 <= io_allowed) { |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 843 | if (wait) |
| 844 | *wait = 0; |
| 845 | return 1; |
| 846 | } |
| 847 | |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 848 | /* Calc approx time to dispatch */ |
| 849 | jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1; |
| 850 | |
| 851 | if (jiffy_wait > jiffy_elapsed) |
| 852 | jiffy_wait = jiffy_wait - jiffy_elapsed; |
| 853 | else |
| 854 | jiffy_wait = 1; |
| 855 | |
| 856 | if (wait) |
| 857 | *wait = jiffy_wait; |
| 858 | return 0; |
| 859 | } |
| 860 | |
Tejun Heo | 0f3457f | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 861 | static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio, |
| 862 | unsigned long *wait) |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 863 | { |
| 864 | bool rw = bio_data_dir(bio); |
Vivek Goyal | 3aad5d3 | 2010-10-01 14:51:14 +0200 | [diff] [blame] | 865 | u64 bytes_allowed, extra_bytes, tmp; |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 866 | unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 867 | |
| 868 | jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; |
| 869 | |
| 870 | /* Slice has just started. Consider one slice interval */ |
| 871 | if (!jiffy_elapsed) |
| 872 | jiffy_elapsed_rnd = throtl_slice; |
| 873 | |
| 874 | jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice); |
| 875 | |
Vivek Goyal | 5e901a2 | 2010-10-01 21:16:38 +0200 | [diff] [blame] | 876 | tmp = tg->bps[rw] * jiffy_elapsed_rnd; |
| 877 | do_div(tmp, HZ); |
Vivek Goyal | 3aad5d3 | 2010-10-01 14:51:14 +0200 | [diff] [blame] | 878 | bytes_allowed = tmp; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 879 | |
| 880 | if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) { |
| 881 | if (wait) |
| 882 | *wait = 0; |
| 883 | return 1; |
| 884 | } |
| 885 | |
| 886 | /* Calc approx time to dispatch */ |
| 887 | extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed; |
| 888 | jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]); |
| 889 | |
| 890 | if (!jiffy_wait) |
| 891 | jiffy_wait = 1; |
| 892 | |
| 893 | /* |
| 894 | * This wait time is without taking into consideration the rounding |
| 895 | * up we did. Add that time also. |
| 896 | */ |
| 897 | jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 898 | if (wait) |
| 899 | *wait = jiffy_wait; |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 900 | return 0; |
| 901 | } |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 902 | |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 903 | /* |
| 904 | * Returns whether one can dispatch a bio or not. Also returns approx number |
| 905 | * of jiffies to wait before this bio is with-in IO rate and can be dispatched |
| 906 | */ |
Tejun Heo | 0f3457f | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 907 | static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio, |
| 908 | unsigned long *wait) |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 909 | { |
| 910 | bool rw = bio_data_dir(bio); |
| 911 | unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0; |
| 912 | |
| 913 | /* |
| 914 | * Currently whole state machine of group depends on first bio |
| 915 | * queued in the group bio list. So one should not be calling |
| 916 | * this function with a different bio if there are other bios |
| 917 | * queued. |
| 918 | */ |
Tejun Heo | 73f0d49 | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 919 | BUG_ON(tg->service_queue.nr_queued[rw] && |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 920 | bio != throtl_peek_queued(&tg->service_queue.queued[rw])); |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 921 | |
| 922 | /* If tg->bps = -1, then BW is unlimited */ |
| 923 | if (tg->bps[rw] == -1 && tg->iops[rw] == -1) { |
| 924 | if (wait) |
| 925 | *wait = 0; |
| 926 | return 1; |
| 927 | } |
| 928 | |
| 929 | /* |
| 930 | * If previous slice expired, start a new one otherwise renew/extend |
| 931 | * existing slice to make sure it is at least throtl_slice interval |
| 932 | * long since now. |
| 933 | */ |
Tejun Heo | 0f3457f | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 934 | if (throtl_slice_used(tg, rw)) |
| 935 | throtl_start_new_slice(tg, rw); |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 936 | else { |
| 937 | if (time_before(tg->slice_end[rw], jiffies + throtl_slice)) |
Tejun Heo | 0f3457f | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 938 | throtl_extend_slice(tg, rw, jiffies + throtl_slice); |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 939 | } |
| 940 | |
Tejun Heo | 0f3457f | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 941 | if (tg_with_in_bps_limit(tg, bio, &bps_wait) && |
| 942 | tg_with_in_iops_limit(tg, bio, &iops_wait)) { |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 943 | if (wait) |
| 944 | *wait = 0; |
| 945 | return 1; |
| 946 | } |
| 947 | |
| 948 | max_wait = max(bps_wait, iops_wait); |
| 949 | |
| 950 | if (wait) |
| 951 | *wait = max_wait; |
| 952 | |
| 953 | if (time_before(tg->slice_end[rw], jiffies + max_wait)) |
Tejun Heo | 0f3457f | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 954 | throtl_extend_slice(tg, rw, jiffies + max_wait); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 955 | |
| 956 | return 0; |
| 957 | } |
| 958 | |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 959 | static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes, |
Tejun Heo | 629ed0b | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 960 | int rw) |
| 961 | { |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 962 | struct throtl_grp *tg = blkg_to_tg(blkg); |
| 963 | struct tg_stats_cpu *stats_cpu; |
Tejun Heo | 629ed0b | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 964 | unsigned long flags; |
| 965 | |
| 966 | /* If per cpu stats are not allocated yet, don't do any accounting. */ |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 967 | if (tg->stats_cpu == NULL) |
Tejun Heo | 629ed0b | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 968 | return; |
| 969 | |
| 970 | /* |
| 971 | * Disabling interrupts to provide mutual exclusion between two |
| 972 | * writes on same cpu. It probably is not needed for 64bit. Not |
| 973 | * optimizing that case yet. |
| 974 | */ |
| 975 | local_irq_save(flags); |
| 976 | |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 977 | stats_cpu = this_cpu_ptr(tg->stats_cpu); |
Tejun Heo | 629ed0b | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 978 | |
Tejun Heo | 629ed0b | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 979 | blkg_rwstat_add(&stats_cpu->serviced, rw, 1); |
| 980 | blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes); |
| 981 | |
| 982 | local_irq_restore(flags); |
| 983 | } |
| 984 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 985 | static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio) |
| 986 | { |
| 987 | bool rw = bio_data_dir(bio); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 988 | |
| 989 | /* Charge the bio to the group */ |
| 990 | tg->bytes_disp[rw] += bio->bi_size; |
Vivek Goyal | 8e89d13 | 2010-09-15 17:06:37 -0400 | [diff] [blame] | 991 | tg->io_disp[rw]++; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 992 | |
Tejun Heo | 2a0f61e | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 993 | /* |
| 994 | * REQ_THROTTLED is used to prevent the same bio to be throttled |
| 995 | * more than once as a throttled bio will go through blk-throtl the |
| 996 | * second time when it eventually gets issued. Set it when a bio |
| 997 | * is being charged to a tg. |
| 998 | * |
| 999 | * Dispatch stats aren't recursive and each @bio should only be |
| 1000 | * accounted by the @tg it was originally associated with. Let's |
| 1001 | * update the stats when setting REQ_THROTTLED for the first time |
| 1002 | * which is guaranteed to be for the @bio's original tg. |
| 1003 | */ |
| 1004 | if (!(bio->bi_rw & REQ_THROTTLED)) { |
| 1005 | bio->bi_rw |= REQ_THROTTLED; |
| 1006 | throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, |
| 1007 | bio->bi_rw); |
| 1008 | } |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1009 | } |
| 1010 | |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1011 | /** |
| 1012 | * throtl_add_bio_tg - add a bio to the specified throtl_grp |
| 1013 | * @bio: bio to add |
| 1014 | * @qn: qnode to use |
| 1015 | * @tg: the target throtl_grp |
| 1016 | * |
| 1017 | * Add @bio to @tg's service_queue using @qn. If @qn is not specified, |
| 1018 | * tg->qnode_on_self[] is used. |
| 1019 | */ |
| 1020 | static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn, |
| 1021 | struct throtl_grp *tg) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1022 | { |
Tejun Heo | 73f0d49 | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 1023 | struct throtl_service_queue *sq = &tg->service_queue; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1024 | bool rw = bio_data_dir(bio); |
| 1025 | |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1026 | if (!qn) |
| 1027 | qn = &tg->qnode_on_self[rw]; |
| 1028 | |
Tejun Heo | 0e9f416 | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 1029 | /* |
| 1030 | * If @tg doesn't currently have any bios queued in the same |
| 1031 | * direction, queueing @bio can change when @tg should be |
| 1032 | * dispatched. Mark that @tg was empty. This is automatically |
| 1033 | * cleaered on the next tg_update_disptime(). |
| 1034 | */ |
| 1035 | if (!sq->nr_queued[rw]) |
| 1036 | tg->flags |= THROTL_TG_WAS_EMPTY; |
| 1037 | |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1038 | throtl_qnode_add_bio(bio, qn, &sq->queued[rw]); |
| 1039 | |
Tejun Heo | 73f0d49 | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 1040 | sq->nr_queued[rw]++; |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1041 | throtl_enqueue_tg(tg); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1042 | } |
| 1043 | |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1044 | static void tg_update_disptime(struct throtl_grp *tg) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1045 | { |
Tejun Heo | 73f0d49 | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 1046 | struct throtl_service_queue *sq = &tg->service_queue; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1047 | unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime; |
| 1048 | struct bio *bio; |
| 1049 | |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1050 | if ((bio = throtl_peek_queued(&sq->queued[READ]))) |
Tejun Heo | 0f3457f | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 1051 | tg_may_dispatch(tg, bio, &read_wait); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1052 | |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1053 | if ((bio = throtl_peek_queued(&sq->queued[WRITE]))) |
Tejun Heo | 0f3457f | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 1054 | tg_may_dispatch(tg, bio, &write_wait); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1055 | |
| 1056 | min_wait = min(read_wait, write_wait); |
| 1057 | disptime = jiffies + min_wait; |
| 1058 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1059 | /* Update dispatch time */ |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1060 | throtl_dequeue_tg(tg); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1061 | tg->disptime = disptime; |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1062 | throtl_enqueue_tg(tg); |
Tejun Heo | 0e9f416 | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 1063 | |
| 1064 | /* see throtl_add_bio_tg() */ |
| 1065 | tg->flags &= ~THROTL_TG_WAS_EMPTY; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1066 | } |
| 1067 | |
Vivek Goyal | 32ee5bc | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1068 | static void start_parent_slice_with_credit(struct throtl_grp *child_tg, |
| 1069 | struct throtl_grp *parent_tg, bool rw) |
| 1070 | { |
| 1071 | if (throtl_slice_used(parent_tg, rw)) { |
| 1072 | throtl_start_new_slice_with_credit(parent_tg, rw, |
| 1073 | child_tg->slice_start[rw]); |
| 1074 | } |
| 1075 | |
| 1076 | } |
| 1077 | |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1078 | static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1079 | { |
Tejun Heo | 73f0d49 | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 1080 | struct throtl_service_queue *sq = &tg->service_queue; |
Tejun Heo | 6bc9c2b | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1081 | struct throtl_service_queue *parent_sq = sq->parent_sq; |
| 1082 | struct throtl_grp *parent_tg = sq_to_tg(parent_sq); |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1083 | struct throtl_grp *tg_to_put = NULL; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1084 | struct bio *bio; |
| 1085 | |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1086 | /* |
| 1087 | * @bio is being transferred from @tg to @parent_sq. Popping a bio |
| 1088 | * from @tg may put its reference and @parent_sq might end up |
| 1089 | * getting released prematurely. Remember the tg to put and put it |
| 1090 | * after @bio is transferred to @parent_sq. |
| 1091 | */ |
| 1092 | bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put); |
Tejun Heo | 73f0d49 | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 1093 | sq->nr_queued[rw]--; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1094 | |
| 1095 | throtl_charge_bio(tg, bio); |
Tejun Heo | 6bc9c2b | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1096 | |
| 1097 | /* |
| 1098 | * If our parent is another tg, we just need to transfer @bio to |
| 1099 | * the parent using throtl_add_bio_tg(). If our parent is |
| 1100 | * @td->service_queue, @bio is ready to be issued. Put it on its |
| 1101 | * bio_lists[] and decrease total number queued. The caller is |
| 1102 | * responsible for issuing these bios. |
| 1103 | */ |
| 1104 | if (parent_tg) { |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1105 | throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg); |
Vivek Goyal | 32ee5bc | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1106 | start_parent_slice_with_credit(tg, parent_tg, rw); |
Tejun Heo | 6bc9c2b | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1107 | } else { |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1108 | throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw], |
| 1109 | &parent_sq->queued[rw]); |
Tejun Heo | 6bc9c2b | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1110 | BUG_ON(tg->td->nr_queued[rw] <= 0); |
| 1111 | tg->td->nr_queued[rw]--; |
| 1112 | } |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1113 | |
Tejun Heo | 0f3457f | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 1114 | throtl_trim_slice(tg, rw); |
Tejun Heo | 6bc9c2b | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1115 | |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1116 | if (tg_to_put) |
| 1117 | blkg_put(tg_to_blkg(tg_to_put)); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1118 | } |
| 1119 | |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1120 | static int throtl_dispatch_tg(struct throtl_grp *tg) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1121 | { |
Tejun Heo | 73f0d49 | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 1122 | struct throtl_service_queue *sq = &tg->service_queue; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1123 | unsigned int nr_reads = 0, nr_writes = 0; |
| 1124 | unsigned int max_nr_reads = throtl_grp_quantum*3/4; |
Vivek Goyal | c2f6805 | 2010-11-15 19:32:42 +0100 | [diff] [blame] | 1125 | unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1126 | struct bio *bio; |
| 1127 | |
| 1128 | /* Try to dispatch 75% READS and 25% WRITES */ |
| 1129 | |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1130 | while ((bio = throtl_peek_queued(&sq->queued[READ])) && |
Tejun Heo | 0f3457f | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 1131 | tg_may_dispatch(tg, bio, NULL)) { |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1132 | |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1133 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1134 | nr_reads++; |
| 1135 | |
| 1136 | if (nr_reads >= max_nr_reads) |
| 1137 | break; |
| 1138 | } |
| 1139 | |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1140 | while ((bio = throtl_peek_queued(&sq->queued[WRITE])) && |
Tejun Heo | 0f3457f | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 1141 | tg_may_dispatch(tg, bio, NULL)) { |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1142 | |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1143 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1144 | nr_writes++; |
| 1145 | |
| 1146 | if (nr_writes >= max_nr_writes) |
| 1147 | break; |
| 1148 | } |
| 1149 | |
| 1150 | return nr_reads + nr_writes; |
| 1151 | } |
| 1152 | |
Tejun Heo | 651930b | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 1153 | static int throtl_select_dispatch(struct throtl_service_queue *parent_sq) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1154 | { |
| 1155 | unsigned int nr_disp = 0; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1156 | |
| 1157 | while (1) { |
Tejun Heo | 73f0d49 | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 1158 | struct throtl_grp *tg = throtl_rb_first(parent_sq); |
| 1159 | struct throtl_service_queue *sq = &tg->service_queue; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1160 | |
| 1161 | if (!tg) |
| 1162 | break; |
| 1163 | |
| 1164 | if (time_before(jiffies, tg->disptime)) |
| 1165 | break; |
| 1166 | |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1167 | throtl_dequeue_tg(tg); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1168 | |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1169 | nr_disp += throtl_dispatch_tg(tg); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1170 | |
Tejun Heo | 73f0d49 | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 1171 | if (sq->nr_queued[0] || sq->nr_queued[1]) |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1172 | tg_update_disptime(tg); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1173 | |
| 1174 | if (nr_disp >= throtl_quantum) |
| 1175 | break; |
| 1176 | } |
| 1177 | |
| 1178 | return nr_disp; |
| 1179 | } |
| 1180 | |
Tejun Heo | 6e1a570 | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 1181 | /** |
| 1182 | * throtl_pending_timer_fn - timer function for service_queue->pending_timer |
| 1183 | * @arg: the throtl_service_queue being serviced |
| 1184 | * |
| 1185 | * This timer is armed when a child throtl_grp with active bio's become |
| 1186 | * pending and queued on the service_queue's pending_tree and expires when |
| 1187 | * the first child throtl_grp should be dispatched. This function |
Tejun Heo | 2e48a53 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1188 | * dispatches bio's from the children throtl_grps to the parent |
| 1189 | * service_queue. |
| 1190 | * |
| 1191 | * If the parent's parent is another throtl_grp, dispatching is propagated |
| 1192 | * by either arming its pending_timer or repeating dispatch directly. If |
| 1193 | * the top-level service_tree is reached, throtl_data->dispatch_work is |
| 1194 | * kicked so that the ready bio's are issued. |
Tejun Heo | 6e1a570 | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 1195 | */ |
Tejun Heo | 69df0ab | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1196 | static void throtl_pending_timer_fn(unsigned long arg) |
| 1197 | { |
| 1198 | struct throtl_service_queue *sq = (void *)arg; |
Tejun Heo | 2e48a53 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1199 | struct throtl_grp *tg = sq_to_tg(sq); |
Tejun Heo | 69df0ab | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1200 | struct throtl_data *td = sq_to_td(sq); |
Tejun Heo | cb76199 | 2013-05-14 13:52:31 -0700 | [diff] [blame] | 1201 | struct request_queue *q = td->queue; |
Tejun Heo | 2e48a53 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1202 | struct throtl_service_queue *parent_sq; |
| 1203 | bool dispatched; |
Tejun Heo | 6e1a570 | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 1204 | int ret; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1205 | |
| 1206 | spin_lock_irq(q->queue_lock); |
Tejun Heo | 2e48a53 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1207 | again: |
| 1208 | parent_sq = sq->parent_sq; |
| 1209 | dispatched = false; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1210 | |
Tejun Heo | 7f52f98 | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 1211 | while (true) { |
| 1212 | throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u", |
Tejun Heo | 2e48a53 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1213 | sq->nr_queued[READ] + sq->nr_queued[WRITE], |
| 1214 | sq->nr_queued[READ], sq->nr_queued[WRITE]); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1215 | |
Tejun Heo | 7f52f98 | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 1216 | ret = throtl_select_dispatch(sq); |
| 1217 | if (ret) { |
Tejun Heo | 7f52f98 | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 1218 | throtl_log(sq, "bios disp=%u", ret); |
| 1219 | dispatched = true; |
Tejun Heo | 651930b | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 1220 | } |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1221 | |
Tejun Heo | 7f52f98 | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 1222 | if (throtl_schedule_next_dispatch(sq, false)) |
| 1223 | break; |
| 1224 | |
| 1225 | /* this dispatch windows is still open, relax and repeat */ |
| 1226 | spin_unlock_irq(q->queue_lock); |
| 1227 | cpu_relax(); |
| 1228 | spin_lock_irq(q->queue_lock); |
| 1229 | } |
Tejun Heo | 6a52560 | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 1230 | |
Tejun Heo | 2e48a53 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1231 | if (!dispatched) |
| 1232 | goto out_unlock; |
Tejun Heo | 6e1a570 | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 1233 | |
Tejun Heo | 2e48a53 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1234 | if (parent_sq) { |
| 1235 | /* @parent_sq is another throl_grp, propagate dispatch */ |
| 1236 | if (tg->flags & THROTL_TG_WAS_EMPTY) { |
| 1237 | tg_update_disptime(tg); |
| 1238 | if (!throtl_schedule_next_dispatch(parent_sq, false)) { |
| 1239 | /* window is already open, repeat dispatching */ |
| 1240 | sq = parent_sq; |
| 1241 | tg = sq_to_tg(sq); |
| 1242 | goto again; |
| 1243 | } |
| 1244 | } |
| 1245 | } else { |
| 1246 | /* reached the top-level, queue issueing */ |
| 1247 | queue_work(kthrotld_workqueue, &td->dispatch_work); |
| 1248 | } |
| 1249 | out_unlock: |
Tejun Heo | 6e1a570 | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 1250 | spin_unlock_irq(q->queue_lock); |
| 1251 | } |
| 1252 | |
| 1253 | /** |
| 1254 | * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work |
| 1255 | * @work: work item being executed |
| 1256 | * |
| 1257 | * This function is queued for execution when bio's reach the bio_lists[] |
| 1258 | * of throtl_data->service_queue. Those bio's are ready and issued by this |
| 1259 | * function. |
| 1260 | */ |
| 1261 | void blk_throtl_dispatch_work_fn(struct work_struct *work) |
| 1262 | { |
| 1263 | struct throtl_data *td = container_of(work, struct throtl_data, |
| 1264 | dispatch_work); |
| 1265 | struct throtl_service_queue *td_sq = &td->service_queue; |
| 1266 | struct request_queue *q = td->queue; |
| 1267 | struct bio_list bio_list_on_stack; |
| 1268 | struct bio *bio; |
| 1269 | struct blk_plug plug; |
| 1270 | int rw; |
| 1271 | |
| 1272 | bio_list_init(&bio_list_on_stack); |
| 1273 | |
| 1274 | spin_lock_irq(q->queue_lock); |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1275 | for (rw = READ; rw <= WRITE; rw++) |
| 1276 | while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL))) |
| 1277 | bio_list_add(&bio_list_on_stack, bio); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1278 | spin_unlock_irq(q->queue_lock); |
| 1279 | |
Tejun Heo | 6e1a570 | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 1280 | if (!bio_list_empty(&bio_list_on_stack)) { |
Vivek Goyal | 69d60eb | 2011-03-09 08:27:37 +0100 | [diff] [blame] | 1281 | blk_start_plug(&plug); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1282 | while((bio = bio_list_pop(&bio_list_on_stack))) |
| 1283 | generic_make_request(bio); |
Vivek Goyal | 69d60eb | 2011-03-09 08:27:37 +0100 | [diff] [blame] | 1284 | blk_finish_plug(&plug); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1285 | } |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1286 | } |
| 1287 | |
Tejun Heo | f95a04a | 2012-04-16 13:57:26 -0700 | [diff] [blame] | 1288 | static u64 tg_prfill_cpu_rwstat(struct seq_file *sf, |
| 1289 | struct blkg_policy_data *pd, int off) |
Tejun Heo | 41b38b6 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1290 | { |
Tejun Heo | f95a04a | 2012-04-16 13:57:26 -0700 | [diff] [blame] | 1291 | struct throtl_grp *tg = pd_to_tg(pd); |
Tejun Heo | 41b38b6 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1292 | struct blkg_rwstat rwstat = { }, tmp; |
| 1293 | int i, cpu; |
| 1294 | |
| 1295 | for_each_possible_cpu(cpu) { |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1296 | struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu); |
Tejun Heo | 41b38b6 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1297 | |
| 1298 | tmp = blkg_rwstat_read((void *)sc + off); |
| 1299 | for (i = 0; i < BLKG_RWSTAT_NR; i++) |
| 1300 | rwstat.cnt[i] += tmp.cnt[i]; |
| 1301 | } |
| 1302 | |
Tejun Heo | f95a04a | 2012-04-16 13:57:26 -0700 | [diff] [blame] | 1303 | return __blkg_prfill_rwstat(sf, pd, &rwstat); |
Tejun Heo | 41b38b6 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1304 | } |
| 1305 | |
Tejun Heo | 182446d | 2013-08-08 20:11:24 -0400 | [diff] [blame] | 1306 | static int tg_print_cpu_rwstat(struct cgroup_subsys_state *css, |
| 1307 | struct cftype *cft, struct seq_file *sf) |
Tejun Heo | 41b38b6 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1308 | { |
Tejun Heo | 182446d | 2013-08-08 20:11:24 -0400 | [diff] [blame] | 1309 | struct blkcg *blkcg = css_to_blkcg(css); |
Tejun Heo | 41b38b6 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1310 | |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 1311 | blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl, |
Tejun Heo | 5bc4afb1 | 2012-04-01 14:38:45 -0700 | [diff] [blame] | 1312 | cft->private, true); |
Tejun Heo | 41b38b6 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1313 | return 0; |
| 1314 | } |
| 1315 | |
Tejun Heo | f95a04a | 2012-04-16 13:57:26 -0700 | [diff] [blame] | 1316 | static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, |
| 1317 | int off) |
Tejun Heo | 60c2bc2 | 2012-04-01 14:38:43 -0700 | [diff] [blame] | 1318 | { |
Tejun Heo | f95a04a | 2012-04-16 13:57:26 -0700 | [diff] [blame] | 1319 | struct throtl_grp *tg = pd_to_tg(pd); |
| 1320 | u64 v = *(u64 *)((void *)tg + off); |
Tejun Heo | 60c2bc2 | 2012-04-01 14:38:43 -0700 | [diff] [blame] | 1321 | |
Tejun Heo | af133ce | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1322 | if (v == -1) |
Tejun Heo | 60c2bc2 | 2012-04-01 14:38:43 -0700 | [diff] [blame] | 1323 | return 0; |
Tejun Heo | f95a04a | 2012-04-16 13:57:26 -0700 | [diff] [blame] | 1324 | return __blkg_prfill_u64(sf, pd, v); |
Tejun Heo | 60c2bc2 | 2012-04-01 14:38:43 -0700 | [diff] [blame] | 1325 | } |
| 1326 | |
Tejun Heo | f95a04a | 2012-04-16 13:57:26 -0700 | [diff] [blame] | 1327 | static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, |
| 1328 | int off) |
Tejun Heo | 60c2bc2 | 2012-04-01 14:38:43 -0700 | [diff] [blame] | 1329 | { |
Tejun Heo | f95a04a | 2012-04-16 13:57:26 -0700 | [diff] [blame] | 1330 | struct throtl_grp *tg = pd_to_tg(pd); |
| 1331 | unsigned int v = *(unsigned int *)((void *)tg + off); |
Tejun Heo | af133ce | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1332 | |
| 1333 | if (v == -1) |
| 1334 | return 0; |
Tejun Heo | f95a04a | 2012-04-16 13:57:26 -0700 | [diff] [blame] | 1335 | return __blkg_prfill_u64(sf, pd, v); |
Tejun Heo | af133ce | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1336 | } |
| 1337 | |
Tejun Heo | 182446d | 2013-08-08 20:11:24 -0400 | [diff] [blame] | 1338 | static int tg_print_conf_u64(struct cgroup_subsys_state *css, |
| 1339 | struct cftype *cft, struct seq_file *sf) |
Tejun Heo | af133ce | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1340 | { |
Tejun Heo | 182446d | 2013-08-08 20:11:24 -0400 | [diff] [blame] | 1341 | blkcg_print_blkgs(sf, css_to_blkcg(css), tg_prfill_conf_u64, |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 1342 | &blkcg_policy_throtl, cft->private, false); |
Tejun Heo | 60c2bc2 | 2012-04-01 14:38:43 -0700 | [diff] [blame] | 1343 | return 0; |
| 1344 | } |
| 1345 | |
Tejun Heo | 182446d | 2013-08-08 20:11:24 -0400 | [diff] [blame] | 1346 | static int tg_print_conf_uint(struct cgroup_subsys_state *css, |
| 1347 | struct cftype *cft, struct seq_file *sf) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1348 | { |
Tejun Heo | 182446d | 2013-08-08 20:11:24 -0400 | [diff] [blame] | 1349 | blkcg_print_blkgs(sf, css_to_blkcg(css), tg_prfill_conf_uint, |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 1350 | &blkcg_policy_throtl, cft->private, false); |
Tejun Heo | af133ce | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1351 | return 0; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1352 | } |
| 1353 | |
Tejun Heo | 182446d | 2013-08-08 20:11:24 -0400 | [diff] [blame] | 1354 | static int tg_set_conf(struct cgroup_subsys_state *css, struct cftype *cft, |
| 1355 | const char *buf, bool is_u64) |
Tejun Heo | 60c2bc2 | 2012-04-01 14:38:43 -0700 | [diff] [blame] | 1356 | { |
Tejun Heo | 182446d | 2013-08-08 20:11:24 -0400 | [diff] [blame] | 1357 | struct blkcg *blkcg = css_to_blkcg(css); |
Tejun Heo | 60c2bc2 | 2012-04-01 14:38:43 -0700 | [diff] [blame] | 1358 | struct blkg_conf_ctx ctx; |
Tejun Heo | af133ce | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1359 | struct throtl_grp *tg; |
Tejun Heo | 69df0ab | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1360 | struct throtl_service_queue *sq; |
Tejun Heo | 693e751 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1361 | struct blkcg_gq *blkg; |
Tejun Heo | 492eb21 | 2013-08-08 20:11:25 -0400 | [diff] [blame] | 1362 | struct cgroup_subsys_state *pos_css; |
Tejun Heo | 60c2bc2 | 2012-04-01 14:38:43 -0700 | [diff] [blame] | 1363 | int ret; |
| 1364 | |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 1365 | ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx); |
Tejun Heo | 60c2bc2 | 2012-04-01 14:38:43 -0700 | [diff] [blame] | 1366 | if (ret) |
| 1367 | return ret; |
| 1368 | |
Tejun Heo | af133ce | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1369 | tg = blkg_to_tg(ctx.blkg); |
Tejun Heo | 69df0ab | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1370 | sq = &tg->service_queue; |
Tejun Heo | af133ce | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1371 | |
Tejun Heo | a2b1693 | 2012-04-13 13:11:33 -0700 | [diff] [blame] | 1372 | if (!ctx.v) |
| 1373 | ctx.v = -1; |
Tejun Heo | af133ce | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1374 | |
Tejun Heo | a2b1693 | 2012-04-13 13:11:33 -0700 | [diff] [blame] | 1375 | if (is_u64) |
| 1376 | *(u64 *)((void *)tg + cft->private) = ctx.v; |
| 1377 | else |
| 1378 | *(unsigned int *)((void *)tg + cft->private) = ctx.v; |
Tejun Heo | af133ce | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1379 | |
Tejun Heo | fda6f27 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1380 | throtl_log(&tg->service_queue, |
| 1381 | "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", |
| 1382 | tg->bps[READ], tg->bps[WRITE], |
| 1383 | tg->iops[READ], tg->iops[WRITE]); |
Tejun Heo | 632b449 | 2013-05-14 13:52:31 -0700 | [diff] [blame] | 1384 | |
| 1385 | /* |
Tejun Heo | 693e751 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1386 | * Update has_rules[] flags for the updated tg's subtree. A tg is |
| 1387 | * considered to have rules if either the tg itself or any of its |
| 1388 | * ancestors has rules. This identifies groups without any |
| 1389 | * restrictions in the whole hierarchy and allows them to bypass |
| 1390 | * blk-throttle. |
| 1391 | */ |
Tejun Heo | 492eb21 | 2013-08-08 20:11:25 -0400 | [diff] [blame] | 1392 | blkg_for_each_descendant_pre(blkg, pos_css, ctx.blkg) |
Tejun Heo | 693e751 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1393 | tg_update_has_rules(blkg_to_tg(blkg)); |
| 1394 | |
| 1395 | /* |
Tejun Heo | 632b449 | 2013-05-14 13:52:31 -0700 | [diff] [blame] | 1396 | * We're already holding queue_lock and know @tg is valid. Let's |
| 1397 | * apply the new config directly. |
| 1398 | * |
| 1399 | * Restart the slices for both READ and WRITES. It might happen |
| 1400 | * that a group's limit are dropped suddenly and we don't want to |
| 1401 | * account recently dispatched IO with new low rate. |
| 1402 | */ |
Tejun Heo | 0f3457f | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 1403 | throtl_start_new_slice(tg, 0); |
| 1404 | throtl_start_new_slice(tg, 1); |
Tejun Heo | 632b449 | 2013-05-14 13:52:31 -0700 | [diff] [blame] | 1405 | |
Tejun Heo | 5b2c16a | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 1406 | if (tg->flags & THROTL_TG_PENDING) { |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1407 | tg_update_disptime(tg); |
Tejun Heo | 7f52f98 | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 1408 | throtl_schedule_next_dispatch(sq->parent_sq, true); |
Tejun Heo | 632b449 | 2013-05-14 13:52:31 -0700 | [diff] [blame] | 1409 | } |
Tejun Heo | 60c2bc2 | 2012-04-01 14:38:43 -0700 | [diff] [blame] | 1410 | |
| 1411 | blkg_conf_finish(&ctx); |
Tejun Heo | a2b1693 | 2012-04-13 13:11:33 -0700 | [diff] [blame] | 1412 | return 0; |
Tejun Heo | 60c2bc2 | 2012-04-01 14:38:43 -0700 | [diff] [blame] | 1413 | } |
| 1414 | |
Tejun Heo | 182446d | 2013-08-08 20:11:24 -0400 | [diff] [blame] | 1415 | static int tg_set_conf_u64(struct cgroup_subsys_state *css, struct cftype *cft, |
Tejun Heo | af133ce | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1416 | const char *buf) |
Tejun Heo | 60c2bc2 | 2012-04-01 14:38:43 -0700 | [diff] [blame] | 1417 | { |
Tejun Heo | 182446d | 2013-08-08 20:11:24 -0400 | [diff] [blame] | 1418 | return tg_set_conf(css, cft, buf, true); |
Tejun Heo | 60c2bc2 | 2012-04-01 14:38:43 -0700 | [diff] [blame] | 1419 | } |
| 1420 | |
Tejun Heo | 182446d | 2013-08-08 20:11:24 -0400 | [diff] [blame] | 1421 | static int tg_set_conf_uint(struct cgroup_subsys_state *css, struct cftype *cft, |
Tejun Heo | af133ce | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1422 | const char *buf) |
Tejun Heo | 60c2bc2 | 2012-04-01 14:38:43 -0700 | [diff] [blame] | 1423 | { |
Tejun Heo | 182446d | 2013-08-08 20:11:24 -0400 | [diff] [blame] | 1424 | return tg_set_conf(css, cft, buf, false); |
Tejun Heo | 60c2bc2 | 2012-04-01 14:38:43 -0700 | [diff] [blame] | 1425 | } |
| 1426 | |
| 1427 | static struct cftype throtl_files[] = { |
| 1428 | { |
| 1429 | .name = "throttle.read_bps_device", |
Tejun Heo | af133ce | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1430 | .private = offsetof(struct throtl_grp, bps[READ]), |
| 1431 | .read_seq_string = tg_print_conf_u64, |
| 1432 | .write_string = tg_set_conf_u64, |
Tejun Heo | 60c2bc2 | 2012-04-01 14:38:43 -0700 | [diff] [blame] | 1433 | .max_write_len = 256, |
| 1434 | }, |
| 1435 | { |
| 1436 | .name = "throttle.write_bps_device", |
Tejun Heo | af133ce | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1437 | .private = offsetof(struct throtl_grp, bps[WRITE]), |
| 1438 | .read_seq_string = tg_print_conf_u64, |
| 1439 | .write_string = tg_set_conf_u64, |
Tejun Heo | 60c2bc2 | 2012-04-01 14:38:43 -0700 | [diff] [blame] | 1440 | .max_write_len = 256, |
| 1441 | }, |
| 1442 | { |
| 1443 | .name = "throttle.read_iops_device", |
Tejun Heo | af133ce | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1444 | .private = offsetof(struct throtl_grp, iops[READ]), |
| 1445 | .read_seq_string = tg_print_conf_uint, |
| 1446 | .write_string = tg_set_conf_uint, |
Tejun Heo | 60c2bc2 | 2012-04-01 14:38:43 -0700 | [diff] [blame] | 1447 | .max_write_len = 256, |
| 1448 | }, |
| 1449 | { |
| 1450 | .name = "throttle.write_iops_device", |
Tejun Heo | af133ce | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1451 | .private = offsetof(struct throtl_grp, iops[WRITE]), |
| 1452 | .read_seq_string = tg_print_conf_uint, |
| 1453 | .write_string = tg_set_conf_uint, |
Tejun Heo | 60c2bc2 | 2012-04-01 14:38:43 -0700 | [diff] [blame] | 1454 | .max_write_len = 256, |
| 1455 | }, |
| 1456 | { |
| 1457 | .name = "throttle.io_service_bytes", |
Tejun Heo | 5bc4afb1 | 2012-04-01 14:38:45 -0700 | [diff] [blame] | 1458 | .private = offsetof(struct tg_stats_cpu, service_bytes), |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1459 | .read_seq_string = tg_print_cpu_rwstat, |
Tejun Heo | 60c2bc2 | 2012-04-01 14:38:43 -0700 | [diff] [blame] | 1460 | }, |
| 1461 | { |
| 1462 | .name = "throttle.io_serviced", |
Tejun Heo | 5bc4afb1 | 2012-04-01 14:38:45 -0700 | [diff] [blame] | 1463 | .private = offsetof(struct tg_stats_cpu, serviced), |
Tejun Heo | 8a3d261 | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1464 | .read_seq_string = tg_print_cpu_rwstat, |
Tejun Heo | 60c2bc2 | 2012-04-01 14:38:43 -0700 | [diff] [blame] | 1465 | }, |
| 1466 | { } /* terminate */ |
| 1467 | }; |
| 1468 | |
Vivek Goyal | da52777 | 2011-03-02 19:05:33 -0500 | [diff] [blame] | 1469 | static void throtl_shutdown_wq(struct request_queue *q) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1470 | { |
| 1471 | struct throtl_data *td = q->td; |
| 1472 | |
Tejun Heo | 69df0ab | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1473 | cancel_work_sync(&td->dispatch_work); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1474 | } |
| 1475 | |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 1476 | static struct blkcg_policy blkcg_policy_throtl = { |
Tejun Heo | f9fcc2d | 2012-04-16 13:57:27 -0700 | [diff] [blame] | 1477 | .pd_size = sizeof(struct throtl_grp), |
| 1478 | .cftypes = throtl_files, |
| 1479 | |
| 1480 | .pd_init_fn = throtl_pd_init, |
Tejun Heo | 693e751 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1481 | .pd_online_fn = throtl_pd_online, |
Tejun Heo | f9fcc2d | 2012-04-16 13:57:27 -0700 | [diff] [blame] | 1482 | .pd_exit_fn = throtl_pd_exit, |
| 1483 | .pd_reset_stats_fn = throtl_pd_reset_stats, |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1484 | }; |
| 1485 | |
Tejun Heo | bc16a4f | 2011-10-19 14:33:01 +0200 | [diff] [blame] | 1486 | bool blk_throtl_bio(struct request_queue *q, struct bio *bio) |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1487 | { |
| 1488 | struct throtl_data *td = q->td; |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1489 | struct throtl_qnode *qn = NULL; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1490 | struct throtl_grp *tg; |
Tejun Heo | 73f0d49 | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 1491 | struct throtl_service_queue *sq; |
Tejun Heo | 0e9f416 | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 1492 | bool rw = bio_data_dir(bio); |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 1493 | struct blkcg *blkcg; |
Tejun Heo | bc16a4f | 2011-10-19 14:33:01 +0200 | [diff] [blame] | 1494 | bool throttled = false; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1495 | |
Tejun Heo | 2a0f61e | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1496 | /* see throtl_charge_bio() */ |
| 1497 | if (bio->bi_rw & REQ_THROTTLED) |
Tejun Heo | bc16a4f | 2011-10-19 14:33:01 +0200 | [diff] [blame] | 1498 | goto out; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1499 | |
Vivek Goyal | af75cd3 | 2011-05-19 15:38:31 -0400 | [diff] [blame] | 1500 | /* |
| 1501 | * A throtl_grp pointer retrieved under rcu can be used to access |
| 1502 | * basic fields like stats and io rates. If a group has no rules, |
| 1503 | * just update the dispatch stats in lockless manner and return. |
| 1504 | */ |
Vivek Goyal | af75cd3 | 2011-05-19 15:38:31 -0400 | [diff] [blame] | 1505 | rcu_read_lock(); |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 1506 | blkcg = bio_blkcg(bio); |
Tejun Heo | cd1604f | 2012-03-05 13:15:06 -0800 | [diff] [blame] | 1507 | tg = throtl_lookup_tg(td, blkcg); |
Vivek Goyal | af75cd3 | 2011-05-19 15:38:31 -0400 | [diff] [blame] | 1508 | if (tg) { |
Tejun Heo | 693e751 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1509 | if (!tg->has_rules[rw]) { |
Tejun Heo | 629ed0b | 2012-04-01 14:38:44 -0700 | [diff] [blame] | 1510 | throtl_update_dispatch_stats(tg_to_blkg(tg), |
| 1511 | bio->bi_size, bio->bi_rw); |
Tejun Heo | 2a7f124 | 2012-03-05 13:15:01 -0800 | [diff] [blame] | 1512 | goto out_unlock_rcu; |
Vivek Goyal | af75cd3 | 2011-05-19 15:38:31 -0400 | [diff] [blame] | 1513 | } |
| 1514 | } |
Vivek Goyal | af75cd3 | 2011-05-19 15:38:31 -0400 | [diff] [blame] | 1515 | |
| 1516 | /* |
| 1517 | * Either group has not been allocated yet or it is not an unlimited |
| 1518 | * IO group |
| 1519 | */ |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1520 | spin_lock_irq(q->queue_lock); |
Tejun Heo | cd1604f | 2012-03-05 13:15:06 -0800 | [diff] [blame] | 1521 | tg = throtl_lookup_create_tg(td, blkcg); |
Tejun Heo | bc16a4f | 2011-10-19 14:33:01 +0200 | [diff] [blame] | 1522 | if (unlikely(!tg)) |
| 1523 | goto out_unlock; |
Vivek Goyal | f469a7b | 2011-05-19 15:38:23 -0400 | [diff] [blame] | 1524 | |
Tejun Heo | 73f0d49 | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 1525 | sq = &tg->service_queue; |
| 1526 | |
Tejun Heo | 9e660ac | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1527 | while (true) { |
| 1528 | /* throtl is FIFO - if bios are already queued, should queue */ |
| 1529 | if (sq->nr_queued[rw]) |
| 1530 | break; |
Vivek Goyal | de701c7 | 2011-03-07 21:09:32 +0100 | [diff] [blame] | 1531 | |
Tejun Heo | 9e660ac | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1532 | /* if above limits, break to queue */ |
| 1533 | if (!tg_may_dispatch(tg, bio, NULL)) |
| 1534 | break; |
| 1535 | |
| 1536 | /* within limits, let's charge and dispatch directly */ |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1537 | throtl_charge_bio(tg, bio); |
Vivek Goyal | 04521db | 2011-03-22 21:54:29 +0100 | [diff] [blame] | 1538 | |
| 1539 | /* |
| 1540 | * We need to trim slice even when bios are not being queued |
| 1541 | * otherwise it might happen that a bio is not queued for |
| 1542 | * a long time and slice keeps on extending and trim is not |
| 1543 | * called for a long time. Now if limits are reduced suddenly |
| 1544 | * we take into account all the IO dispatched so far at new |
| 1545 | * low rate and * newly queued IO gets a really long dispatch |
| 1546 | * time. |
| 1547 | * |
| 1548 | * So keep on trimming slice even if bio is not queued. |
| 1549 | */ |
Tejun Heo | 0f3457f | 2013-05-14 13:52:32 -0700 | [diff] [blame] | 1550 | throtl_trim_slice(tg, rw); |
Tejun Heo | 9e660ac | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1551 | |
| 1552 | /* |
| 1553 | * @bio passed through this layer without being throttled. |
| 1554 | * Climb up the ladder. If we''re already at the top, it |
| 1555 | * can be executed directly. |
| 1556 | */ |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1557 | qn = &tg->qnode_on_parent[rw]; |
Tejun Heo | 9e660ac | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1558 | sq = sq->parent_sq; |
| 1559 | tg = sq_to_tg(sq); |
| 1560 | if (!tg) |
| 1561 | goto out_unlock; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1562 | } |
| 1563 | |
Tejun Heo | 9e660ac | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1564 | /* out-of-limit, queue to @tg */ |
Tejun Heo | fda6f27 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1565 | throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d", |
| 1566 | rw == READ ? 'R' : 'W', |
| 1567 | tg->bytes_disp[rw], bio->bi_size, tg->bps[rw], |
| 1568 | tg->io_disp[rw], tg->iops[rw], |
| 1569 | sq->nr_queued[READ], sq->nr_queued[WRITE]); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1570 | |
Tejun Heo | 671058f | 2012-03-05 13:15:29 -0800 | [diff] [blame] | 1571 | bio_associate_current(bio); |
Tejun Heo | 6bc9c2b | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1572 | tg->td->nr_queued[rw]++; |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1573 | throtl_add_bio_tg(bio, qn, tg); |
Tejun Heo | bc16a4f | 2011-10-19 14:33:01 +0200 | [diff] [blame] | 1574 | throttled = true; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1575 | |
Tejun Heo | 7f52f98 | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 1576 | /* |
| 1577 | * Update @tg's dispatch time and force schedule dispatch if @tg |
| 1578 | * was empty before @bio. The forced scheduling isn't likely to |
| 1579 | * cause undue delay as @bio is likely to be dispatched directly if |
| 1580 | * its @tg's disptime is not in the future. |
| 1581 | */ |
Tejun Heo | 0e9f416 | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 1582 | if (tg->flags & THROTL_TG_WAS_EMPTY) { |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1583 | tg_update_disptime(tg); |
Tejun Heo | 7f52f98 | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 1584 | throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1585 | } |
| 1586 | |
Tejun Heo | bc16a4f | 2011-10-19 14:33:01 +0200 | [diff] [blame] | 1587 | out_unlock: |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1588 | spin_unlock_irq(q->queue_lock); |
Tejun Heo | 2a7f124 | 2012-03-05 13:15:01 -0800 | [diff] [blame] | 1589 | out_unlock_rcu: |
| 1590 | rcu_read_unlock(); |
Tejun Heo | bc16a4f | 2011-10-19 14:33:01 +0200 | [diff] [blame] | 1591 | out: |
Tejun Heo | 2a0f61e | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1592 | /* |
| 1593 | * As multiple blk-throtls may stack in the same issue path, we |
| 1594 | * don't want bios to leave with the flag set. Clear the flag if |
| 1595 | * being issued. |
| 1596 | */ |
| 1597 | if (!throttled) |
| 1598 | bio->bi_rw &= ~REQ_THROTTLED; |
Tejun Heo | bc16a4f | 2011-10-19 14:33:01 +0200 | [diff] [blame] | 1599 | return throttled; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1600 | } |
| 1601 | |
Tejun Heo | 2a12f0d | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 1602 | /* |
| 1603 | * Dispatch all bios from all children tg's queued on @parent_sq. On |
| 1604 | * return, @parent_sq is guaranteed to not have any active children tg's |
| 1605 | * and all bios from previously active tg's are on @parent_sq->bio_lists[]. |
| 1606 | */ |
| 1607 | static void tg_drain_bios(struct throtl_service_queue *parent_sq) |
| 1608 | { |
| 1609 | struct throtl_grp *tg; |
| 1610 | |
| 1611 | while ((tg = throtl_rb_first(parent_sq))) { |
| 1612 | struct throtl_service_queue *sq = &tg->service_queue; |
| 1613 | struct bio *bio; |
| 1614 | |
| 1615 | throtl_dequeue_tg(tg); |
| 1616 | |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1617 | while ((bio = throtl_peek_queued(&sq->queued[READ]))) |
Tejun Heo | 2a12f0d | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 1618 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1619 | while ((bio = throtl_peek_queued(&sq->queued[WRITE]))) |
Tejun Heo | 2a12f0d | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 1620 | tg_dispatch_one_bio(tg, bio_data_dir(bio)); |
| 1621 | } |
| 1622 | } |
| 1623 | |
Tejun Heo | c9a929d | 2011-10-19 14:42:16 +0200 | [diff] [blame] | 1624 | /** |
| 1625 | * blk_throtl_drain - drain throttled bios |
| 1626 | * @q: request_queue to drain throttled bios for |
| 1627 | * |
| 1628 | * Dispatch all currently throttled bios on @q through ->make_request_fn(). |
| 1629 | */ |
| 1630 | void blk_throtl_drain(struct request_queue *q) |
| 1631 | __releases(q->queue_lock) __acquires(q->queue_lock) |
| 1632 | { |
| 1633 | struct throtl_data *td = q->td; |
Tejun Heo | 2a12f0d | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 1634 | struct blkcg_gq *blkg; |
Tejun Heo | 492eb21 | 2013-08-08 20:11:25 -0400 | [diff] [blame] | 1635 | struct cgroup_subsys_state *pos_css; |
Tejun Heo | c9a929d | 2011-10-19 14:42:16 +0200 | [diff] [blame] | 1636 | struct bio *bio; |
Tejun Heo | 651930b | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 1637 | int rw; |
Tejun Heo | c9a929d | 2011-10-19 14:42:16 +0200 | [diff] [blame] | 1638 | |
Andi Kleen | 8bcb6c7 | 2012-03-30 12:33:28 +0200 | [diff] [blame] | 1639 | queue_lockdep_assert_held(q); |
Tejun Heo | 2a12f0d | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 1640 | rcu_read_lock(); |
Tejun Heo | c9a929d | 2011-10-19 14:42:16 +0200 | [diff] [blame] | 1641 | |
Tejun Heo | 2a12f0d | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 1642 | /* |
| 1643 | * Drain each tg while doing post-order walk on the blkg tree, so |
| 1644 | * that all bios are propagated to td->service_queue. It'd be |
| 1645 | * better to walk service_queue tree directly but blkg walk is |
| 1646 | * easier. |
| 1647 | */ |
Tejun Heo | 492eb21 | 2013-08-08 20:11:25 -0400 | [diff] [blame] | 1648 | blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) |
Tejun Heo | 2a12f0d | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 1649 | tg_drain_bios(&blkg_to_tg(blkg)->service_queue); |
Tejun Heo | 73f0d49 | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 1650 | |
Tejun Heo | 2a12f0d | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 1651 | /* finally, transfer bios from top-level tg's into the td */ |
| 1652 | tg_drain_bios(&td->service_queue); |
| 1653 | |
| 1654 | rcu_read_unlock(); |
Tejun Heo | c9a929d | 2011-10-19 14:42:16 +0200 | [diff] [blame] | 1655 | spin_unlock_irq(q->queue_lock); |
| 1656 | |
Tejun Heo | 2a12f0d | 2013-05-14 13:52:37 -0700 | [diff] [blame] | 1657 | /* all bios now should be in td->service_queue, issue them */ |
Tejun Heo | 651930b | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 1658 | for (rw = READ; rw <= WRITE; rw++) |
Tejun Heo | c5cc207 | 2013-05-14 13:52:38 -0700 | [diff] [blame] | 1659 | while ((bio = throtl_pop_queued(&td->service_queue.queued[rw], |
| 1660 | NULL))) |
Tejun Heo | 651930b | 2013-05-14 13:52:35 -0700 | [diff] [blame] | 1661 | generic_make_request(bio); |
Tejun Heo | c9a929d | 2011-10-19 14:42:16 +0200 | [diff] [blame] | 1662 | |
| 1663 | spin_lock_irq(q->queue_lock); |
| 1664 | } |
| 1665 | |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1666 | int blk_throtl_init(struct request_queue *q) |
| 1667 | { |
| 1668 | struct throtl_data *td; |
Tejun Heo | a2b1693 | 2012-04-13 13:11:33 -0700 | [diff] [blame] | 1669 | int ret; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1670 | |
| 1671 | td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); |
| 1672 | if (!td) |
| 1673 | return -ENOMEM; |
| 1674 | |
Tejun Heo | 69df0ab | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1675 | INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); |
Tejun Heo | 77216b0 | 2013-05-14 13:52:36 -0700 | [diff] [blame] | 1676 | throtl_service_queue_init(&td->service_queue, NULL); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1677 | |
Tejun Heo | cd1604f | 2012-03-05 13:15:06 -0800 | [diff] [blame] | 1678 | q->td = td; |
Vivek Goyal | 29b1258 | 2011-05-19 15:38:24 -0400 | [diff] [blame] | 1679 | td->queue = q; |
Vivek Goyal | 02977e4 | 2010-10-01 14:49:48 +0200 | [diff] [blame] | 1680 | |
Tejun Heo | a2b1693 | 2012-04-13 13:11:33 -0700 | [diff] [blame] | 1681 | /* activate policy */ |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 1682 | ret = blkcg_activate_policy(q, &blkcg_policy_throtl); |
Tejun Heo | a2b1693 | 2012-04-13 13:11:33 -0700 | [diff] [blame] | 1683 | if (ret) |
Vivek Goyal | 29b1258 | 2011-05-19 15:38:24 -0400 | [diff] [blame] | 1684 | kfree(td); |
Tejun Heo | a2b1693 | 2012-04-13 13:11:33 -0700 | [diff] [blame] | 1685 | return ret; |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1686 | } |
| 1687 | |
| 1688 | void blk_throtl_exit(struct request_queue *q) |
| 1689 | { |
Tejun Heo | c875f4d | 2012-03-05 13:15:22 -0800 | [diff] [blame] | 1690 | BUG_ON(!q->td); |
Vivek Goyal | da52777 | 2011-03-02 19:05:33 -0500 | [diff] [blame] | 1691 | throtl_shutdown_wq(q); |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 1692 | blkcg_deactivate_policy(q, &blkcg_policy_throtl); |
Tejun Heo | c9a929d | 2011-10-19 14:42:16 +0200 | [diff] [blame] | 1693 | kfree(q->td); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1694 | } |
| 1695 | |
| 1696 | static int __init throtl_init(void) |
| 1697 | { |
Vivek Goyal | 450adcb | 2011-03-01 13:40:54 -0500 | [diff] [blame] | 1698 | kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); |
| 1699 | if (!kthrotld_workqueue) |
| 1700 | panic("Failed to create kthrotld\n"); |
| 1701 | |
Tejun Heo | 3c79839 | 2012-04-16 13:57:25 -0700 | [diff] [blame] | 1702 | return blkcg_policy_register(&blkcg_policy_throtl); |
Vivek Goyal | e43473b | 2010-09-15 17:06:35 -0400 | [diff] [blame] | 1703 | } |
| 1704 | |
| 1705 | module_init(throtl_init); |