Tatyana Brokhman | 1634906 | 2012-09-20 10:46:10 +0300 | [diff] [blame] | 1 | /* |
| 2 | * ROW (Read Over Write) I/O scheduler. |
| 3 | * |
| 4 | * Copyright (c) 2012, The Linux Foundation. All rights reserved. |
| 5 | * |
| 6 | * This program is free software; you can redistribute it and/or modify |
| 7 | * it under the terms of the GNU General Public License version 2 and |
| 8 | * only version 2 as published by the Free Software Foundation. |
| 9 | * |
| 10 | * This program is distributed in the hope that it will be useful, |
| 11 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 12 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 13 | * GNU General Public License for more details. |
| 14 | */ |
| 15 | |
| 16 | /* See Documentation/block/row-iosched.txt */ |
| 17 | |
| 18 | #include <linux/kernel.h> |
| 19 | #include <linux/fs.h> |
| 20 | #include <linux/blkdev.h> |
| 21 | #include <linux/elevator.h> |
| 22 | #include <linux/bio.h> |
| 23 | #include <linux/module.h> |
| 24 | #include <linux/slab.h> |
| 25 | #include <linux/init.h> |
| 26 | #include <linux/compiler.h> |
| 27 | #include <linux/blktrace_api.h> |
| 28 | #include <linux/jiffies.h> |
| 29 | |
| 30 | /* |
| 31 | * enum row_queue_prio - Priorities of the ROW queues |
| 32 | * |
| 33 | * This enum defines the priorities (and the number of queues) |
| 34 | * the requests will be disptributed to. The higher priority - |
| 35 | * the bigger is the dispatch quantum given to that queue. |
| 36 | * ROWQ_PRIO_HIGH_READ - is the higher priority queue. |
| 37 | * |
| 38 | */ |
| 39 | enum row_queue_prio { |
| 40 | ROWQ_PRIO_HIGH_READ = 0, |
| 41 | ROWQ_PRIO_REG_READ, |
| 42 | ROWQ_PRIO_HIGH_SWRITE, |
| 43 | ROWQ_PRIO_REG_SWRITE, |
| 44 | ROWQ_PRIO_REG_WRITE, |
| 45 | ROWQ_PRIO_LOW_READ, |
| 46 | ROWQ_PRIO_LOW_SWRITE, |
| 47 | ROWQ_MAX_PRIO, |
| 48 | }; |
| 49 | |
| 50 | /* Flags indicating whether idling is enabled on the queue */ |
| 51 | static const bool queue_idling_enabled[] = { |
| 52 | true, /* ROWQ_PRIO_HIGH_READ */ |
| 53 | true, /* ROWQ_PRIO_REG_READ */ |
| 54 | false, /* ROWQ_PRIO_HIGH_SWRITE */ |
| 55 | false, /* ROWQ_PRIO_REG_SWRITE */ |
| 56 | false, /* ROWQ_PRIO_REG_WRITE */ |
| 57 | false, /* ROWQ_PRIO_LOW_READ */ |
| 58 | false, /* ROWQ_PRIO_LOW_SWRITE */ |
| 59 | }; |
| 60 | |
Tatyana Brokhman | 0ef8143 | 2012-12-20 19:23:58 +0200 | [diff] [blame^] | 61 | /* Flags indicating whether the queue can notify on urgent requests */ |
| 62 | static const bool urgent_queues[] = { |
| 63 | true, /* ROWQ_PRIO_HIGH_READ */ |
| 64 | true, /* ROWQ_PRIO_REG_READ */ |
| 65 | false, /* ROWQ_PRIO_HIGH_SWRITE */ |
| 66 | false, /* ROWQ_PRIO_REG_SWRITE */ |
| 67 | false, /* ROWQ_PRIO_REG_WRITE */ |
| 68 | false, /* ROWQ_PRIO_LOW_READ */ |
| 69 | false, /* ROWQ_PRIO_LOW_SWRITE */ |
| 70 | }; |
| 71 | |
Tatyana Brokhman | 1634906 | 2012-09-20 10:46:10 +0300 | [diff] [blame] | 72 | /* Default values for row queues quantums in each dispatch cycle */ |
| 73 | static const int queue_quantum[] = { |
| 74 | 100, /* ROWQ_PRIO_HIGH_READ */ |
| 75 | 100, /* ROWQ_PRIO_REG_READ */ |
| 76 | 2, /* ROWQ_PRIO_HIGH_SWRITE */ |
| 77 | 1, /* ROWQ_PRIO_REG_SWRITE */ |
| 78 | 1, /* ROWQ_PRIO_REG_WRITE */ |
| 79 | 1, /* ROWQ_PRIO_LOW_READ */ |
| 80 | 1 /* ROWQ_PRIO_LOW_SWRITE */ |
| 81 | }; |
| 82 | |
| 83 | /* Default values for idling on read queues */ |
| 84 | #define ROW_IDLE_TIME 50 /* 5 msec */ |
| 85 | #define ROW_READ_FREQ 70 /* 7 msec */ |
| 86 | |
| 87 | /** |
| 88 | * struct rowq_idling_data - parameters for idling on the queue |
| 89 | * @idle_trigger_time: time (in jiffies). If a new request was |
| 90 | * inserted before this time value, idling |
| 91 | * will be enabled. |
| 92 | * @begin_idling: flag indicating wether we should idle |
| 93 | * |
| 94 | */ |
| 95 | struct rowq_idling_data { |
| 96 | unsigned long idle_trigger_time; |
| 97 | bool begin_idling; |
| 98 | }; |
| 99 | |
| 100 | /** |
| 101 | * struct row_queue - requests grouping structure |
| 102 | * @rdata: parent row_data structure |
| 103 | * @fifo: fifo of requests |
| 104 | * @prio: queue priority (enum row_queue_prio) |
| 105 | * @nr_dispatched: number of requests already dispatched in |
| 106 | * the current dispatch cycle |
| 107 | * @slice: number of requests to dispatch in a cycle |
| 108 | * @idle_data: data for idling on queues |
| 109 | * |
| 110 | */ |
| 111 | struct row_queue { |
| 112 | struct row_data *rdata; |
| 113 | struct list_head fifo; |
| 114 | enum row_queue_prio prio; |
| 115 | |
| 116 | unsigned int nr_dispatched; |
| 117 | unsigned int slice; |
| 118 | |
| 119 | /* used only for READ queues */ |
| 120 | struct rowq_idling_data idle_data; |
| 121 | }; |
| 122 | |
| 123 | /** |
| 124 | * struct idling_data - data for idling on empty rqueue |
| 125 | * @idle_time: idling duration (msec) |
| 126 | * @freq: min time between two requests that |
| 127 | * triger idling (msec) |
| 128 | * @idle_work: pointer to struct delayed_work |
| 129 | * |
| 130 | */ |
| 131 | struct idling_data { |
| 132 | unsigned long idle_time; |
| 133 | unsigned long freq; |
| 134 | |
| 135 | struct workqueue_struct *idle_workqueue; |
| 136 | struct delayed_work idle_work; |
| 137 | }; |
| 138 | |
| 139 | /** |
| 140 | * struct row_queue - Per block device rqueue structure |
| 141 | * @dispatch_queue: dispatch rqueue |
| 142 | * @row_queues: array of priority request queues with |
| 143 | * dispatch quantum per rqueue |
| 144 | * @curr_queue: index in the row_queues array of the |
| 145 | * currently serviced rqueue |
| 146 | * @read_idle: data for idling after READ request |
| 147 | * @nr_reqs: nr_reqs[0] holds the number of all READ requests in |
| 148 | * scheduler, nr_reqs[1] holds the number of all WRITE |
| 149 | * requests in scheduler |
| 150 | * @cycle_flags: used for marking unserved queueus |
| 151 | * |
| 152 | */ |
| 153 | struct row_data { |
| 154 | struct request_queue *dispatch_queue; |
| 155 | |
| 156 | struct { |
| 157 | struct row_queue rqueue; |
| 158 | int disp_quantum; |
| 159 | } row_queues[ROWQ_MAX_PRIO]; |
| 160 | |
| 161 | enum row_queue_prio curr_queue; |
| 162 | |
| 163 | struct idling_data read_idle; |
| 164 | unsigned int nr_reqs[2]; |
| 165 | |
| 166 | unsigned int cycle_flags; |
| 167 | }; |
| 168 | |
| 169 | #define RQ_ROWQ(rq) ((struct row_queue *) ((rq)->elv.priv[0])) |
| 170 | |
| 171 | #define row_log(q, fmt, args...) \ |
| 172 | blk_add_trace_msg(q, "%s():" fmt , __func__, ##args) |
| 173 | #define row_log_rowq(rdata, rowq_id, fmt, args...) \ |
| 174 | blk_add_trace_msg(rdata->dispatch_queue, "rowq%d " fmt, \ |
| 175 | rowq_id, ##args) |
| 176 | |
| 177 | static inline void row_mark_rowq_unserved(struct row_data *rd, |
| 178 | enum row_queue_prio qnum) |
| 179 | { |
| 180 | rd->cycle_flags |= (1 << qnum); |
| 181 | } |
| 182 | |
| 183 | static inline void row_clear_rowq_unserved(struct row_data *rd, |
| 184 | enum row_queue_prio qnum) |
| 185 | { |
| 186 | rd->cycle_flags &= ~(1 << qnum); |
| 187 | } |
| 188 | |
| 189 | static inline int row_rowq_unserved(struct row_data *rd, |
| 190 | enum row_queue_prio qnum) |
| 191 | { |
| 192 | return rd->cycle_flags & (1 << qnum); |
| 193 | } |
| 194 | |
| 195 | /******************** Static helper functions ***********************/ |
| 196 | /* |
| 197 | * kick_queue() - Wake up device driver queue thread |
| 198 | * @work: pointer to struct work_struct |
| 199 | * |
| 200 | * This is a idling delayed work function. It's purpose is to wake up the |
| 201 | * device driver in order for it to start fetching requests. |
| 202 | * |
| 203 | */ |
| 204 | static void kick_queue(struct work_struct *work) |
| 205 | { |
| 206 | struct delayed_work *idle_work = to_delayed_work(work); |
| 207 | struct idling_data *read_data = |
| 208 | container_of(idle_work, struct idling_data, idle_work); |
| 209 | struct row_data *rd = |
| 210 | container_of(read_data, struct row_data, read_idle); |
| 211 | |
| 212 | row_log_rowq(rd, rd->curr_queue, "Performing delayed work"); |
| 213 | /* Mark idling process as done */ |
| 214 | rd->row_queues[rd->curr_queue].rqueue.idle_data.begin_idling = false; |
| 215 | |
| 216 | if (!(rd->nr_reqs[0] + rd->nr_reqs[1])) |
| 217 | row_log(rd->dispatch_queue, "No requests in scheduler"); |
| 218 | else { |
| 219 | spin_lock_irq(rd->dispatch_queue->queue_lock); |
| 220 | __blk_run_queue(rd->dispatch_queue); |
| 221 | spin_unlock_irq(rd->dispatch_queue->queue_lock); |
| 222 | } |
| 223 | } |
| 224 | |
| 225 | /* |
| 226 | * row_restart_disp_cycle() - Restart the dispatch cycle |
| 227 | * @rd: pointer to struct row_data |
| 228 | * |
| 229 | * This function restarts the dispatch cycle by: |
| 230 | * - Setting current queue to ROWQ_PRIO_HIGH_READ |
| 231 | * - For each queue: reset the number of requests dispatched in |
| 232 | * the cycle |
| 233 | */ |
| 234 | static inline void row_restart_disp_cycle(struct row_data *rd) |
| 235 | { |
| 236 | int i; |
| 237 | |
| 238 | for (i = 0; i < ROWQ_MAX_PRIO; i++) |
| 239 | rd->row_queues[i].rqueue.nr_dispatched = 0; |
| 240 | |
| 241 | rd->curr_queue = ROWQ_PRIO_HIGH_READ; |
| 242 | row_log(rd->dispatch_queue, "Restarting cycle"); |
| 243 | } |
| 244 | |
| 245 | static inline void row_get_next_queue(struct row_data *rd) |
| 246 | { |
| 247 | rd->curr_queue++; |
| 248 | if (rd->curr_queue == ROWQ_MAX_PRIO) |
| 249 | row_restart_disp_cycle(rd); |
| 250 | } |
| 251 | |
| 252 | /******************* Elevator callback functions *********************/ |
| 253 | |
| 254 | /* |
| 255 | * row_add_request() - Add request to the scheduler |
| 256 | * @q: requests queue |
| 257 | * @rq: request to add |
| 258 | * |
| 259 | */ |
| 260 | static void row_add_request(struct request_queue *q, |
| 261 | struct request *rq) |
| 262 | { |
| 263 | struct row_data *rd = (struct row_data *)q->elevator->elevator_data; |
| 264 | struct row_queue *rqueue = RQ_ROWQ(rq); |
| 265 | |
| 266 | list_add_tail(&rq->queuelist, &rqueue->fifo); |
| 267 | rd->nr_reqs[rq_data_dir(rq)]++; |
| 268 | rq_set_fifo_time(rq, jiffies); /* for statistics*/ |
| 269 | |
| 270 | if (queue_idling_enabled[rqueue->prio]) { |
| 271 | if (delayed_work_pending(&rd->read_idle.idle_work)) |
| 272 | (void)cancel_delayed_work( |
| 273 | &rd->read_idle.idle_work); |
| 274 | if (time_before(jiffies, rqueue->idle_data.idle_trigger_time)) { |
| 275 | rqueue->idle_data.begin_idling = true; |
| 276 | row_log_rowq(rd, rqueue->prio, "Enable idling"); |
| 277 | } else |
| 278 | rqueue->idle_data.begin_idling = false; |
| 279 | |
| 280 | rqueue->idle_data.idle_trigger_time = |
| 281 | jiffies + msecs_to_jiffies(rd->read_idle.freq); |
| 282 | } |
Tatyana Brokhman | 0ef8143 | 2012-12-20 19:23:58 +0200 | [diff] [blame^] | 283 | if (urgent_queues[rqueue->prio] && |
| 284 | row_rowq_unserved(rd, rqueue->prio)) { |
| 285 | row_log_rowq(rd, rqueue->prio, |
| 286 | "added urgent req curr_queue = %d", |
| 287 | rd->curr_queue); |
| 288 | } else |
| 289 | row_log_rowq(rd, rqueue->prio, "added request"); |
Tatyana Brokhman | 1634906 | 2012-09-20 10:46:10 +0300 | [diff] [blame] | 290 | } |
| 291 | |
Tatyana Brokhman | b7bf9ac | 2012-10-30 08:33:06 +0200 | [diff] [blame] | 292 | /** |
| 293 | * row_reinsert_req() - Reinsert request back to the scheduler |
| 294 | * @q: requests queue |
| 295 | * @rq: request to add |
| 296 | * |
| 297 | * Reinsert the given request back to the queue it was |
| 298 | * dispatched from as if it was never dispatched. |
| 299 | * |
| 300 | * Returns 0 on success, error code otherwise |
| 301 | */ |
| 302 | static int row_reinsert_req(struct request_queue *q, |
| 303 | struct request *rq) |
| 304 | { |
| 305 | struct row_data *rd = q->elevator->elevator_data; |
| 306 | struct row_queue *rqueue = RQ_ROWQ(rq); |
| 307 | |
| 308 | /* Verify rqueue is legitimate */ |
| 309 | if (rqueue->prio >= ROWQ_MAX_PRIO) { |
| 310 | pr_err("\n\nROW BUG: row_reinsert_req() rqueue->prio = %d\n", |
| 311 | rqueue->prio); |
| 312 | blk_dump_rq_flags(rq, ""); |
| 313 | return -EIO; |
| 314 | } |
| 315 | |
| 316 | list_add(&rq->queuelist, &rqueue->fifo); |
| 317 | rd->nr_reqs[rq_data_dir(rq)]++; |
| 318 | |
| 319 | row_log_rowq(rd, rqueue->prio, "request reinserted"); |
| 320 | |
| 321 | return 0; |
| 322 | } |
| 323 | |
| 324 | /** |
Tatyana Brokhman | 0ef8143 | 2012-12-20 19:23:58 +0200 | [diff] [blame^] | 325 | * row_urgent_pending() - Return TRUE if there is an urgent |
| 326 | * request on scheduler |
| 327 | * @q: requests queue |
| 328 | */ |
| 329 | static bool row_urgent_pending(struct request_queue *q) |
| 330 | { |
| 331 | struct row_data *rd = q->elevator->elevator_data; |
| 332 | int i; |
| 333 | |
| 334 | for (i = 0; i < ROWQ_MAX_PRIO; i++) |
| 335 | if (urgent_queues[i] && row_rowq_unserved(rd, i) && |
| 336 | !list_empty(&rd->row_queues[i].rqueue.fifo)) { |
| 337 | row_log_rowq(rd, i, |
| 338 | "Urgent request pending (curr=%i)", |
| 339 | rd->curr_queue); |
| 340 | return true; |
| 341 | } |
| 342 | |
| 343 | return false; |
| 344 | } |
| 345 | |
| 346 | /** |
Tatyana Brokhman | 1634906 | 2012-09-20 10:46:10 +0300 | [diff] [blame] | 347 | * row_remove_request() - Remove given request from scheduler |
| 348 | * @q: requests queue |
| 349 | * @rq: request to remove |
| 350 | * |
| 351 | */ |
| 352 | static void row_remove_request(struct request_queue *q, |
| 353 | struct request *rq) |
| 354 | { |
| 355 | struct row_data *rd = (struct row_data *)q->elevator->elevator_data; |
| 356 | |
| 357 | rq_fifo_clear(rq); |
| 358 | rd->nr_reqs[rq_data_dir(rq)]--; |
| 359 | } |
| 360 | |
| 361 | /* |
| 362 | * row_dispatch_insert() - move request to dispatch queue |
| 363 | * @rd: pointer to struct row_data |
| 364 | * |
| 365 | * This function moves the next request to dispatch from |
| 366 | * rd->curr_queue to the dispatch queue |
| 367 | * |
| 368 | */ |
| 369 | static void row_dispatch_insert(struct row_data *rd) |
| 370 | { |
| 371 | struct request *rq; |
| 372 | |
| 373 | rq = rq_entry_fifo(rd->row_queues[rd->curr_queue].rqueue.fifo.next); |
| 374 | row_remove_request(rd->dispatch_queue, rq); |
| 375 | elv_dispatch_add_tail(rd->dispatch_queue, rq); |
| 376 | rd->row_queues[rd->curr_queue].rqueue.nr_dispatched++; |
| 377 | row_clear_rowq_unserved(rd, rd->curr_queue); |
| 378 | row_log_rowq(rd, rd->curr_queue, " Dispatched request nr_disp = %d", |
| 379 | rd->row_queues[rd->curr_queue].rqueue.nr_dispatched); |
| 380 | } |
| 381 | |
| 382 | /* |
| 383 | * row_choose_queue() - choose the next queue to dispatch from |
| 384 | * @rd: pointer to struct row_data |
| 385 | * |
| 386 | * Updates rd->curr_queue. Returns 1 if there are requests to |
| 387 | * dispatch, 0 if there are no requests in scheduler |
| 388 | * |
| 389 | */ |
| 390 | static int row_choose_queue(struct row_data *rd) |
| 391 | { |
| 392 | int prev_curr_queue = rd->curr_queue; |
| 393 | |
| 394 | if (!(rd->nr_reqs[0] + rd->nr_reqs[1])) { |
| 395 | row_log(rd->dispatch_queue, "No more requests in scheduler"); |
| 396 | return 0; |
| 397 | } |
| 398 | |
| 399 | row_get_next_queue(rd); |
| 400 | |
| 401 | /* |
| 402 | * Loop over all queues to find the next queue that is not empty. |
| 403 | * Stop when you get back to curr_queue |
| 404 | */ |
| 405 | while (list_empty(&rd->row_queues[rd->curr_queue].rqueue.fifo) |
| 406 | && rd->curr_queue != prev_curr_queue) { |
| 407 | /* Mark rqueue as unserved */ |
| 408 | row_mark_rowq_unserved(rd, rd->curr_queue); |
| 409 | row_get_next_queue(rd); |
| 410 | } |
| 411 | |
| 412 | return 1; |
| 413 | } |
| 414 | |
| 415 | /* |
| 416 | * row_dispatch_requests() - selects the next request to dispatch |
| 417 | * @q: requests queue |
| 418 | * @force: ignored |
| 419 | * |
| 420 | * Return 0 if no requests were moved to the dispatch queue. |
| 421 | * 1 otherwise |
| 422 | * |
| 423 | */ |
| 424 | static int row_dispatch_requests(struct request_queue *q, int force) |
| 425 | { |
| 426 | struct row_data *rd = (struct row_data *)q->elevator->elevator_data; |
| 427 | int ret = 0, currq, i; |
| 428 | |
| 429 | currq = rd->curr_queue; |
| 430 | |
| 431 | /* |
| 432 | * Find the first unserved queue (with higher priority then currq) |
| 433 | * that is not empty |
| 434 | */ |
| 435 | for (i = 0; i < currq; i++) { |
| 436 | if (row_rowq_unserved(rd, i) && |
| 437 | !list_empty(&rd->row_queues[i].rqueue.fifo)) { |
| 438 | row_log_rowq(rd, currq, |
| 439 | " Preemting for unserved rowq%d", i); |
| 440 | rd->curr_queue = i; |
| 441 | row_dispatch_insert(rd); |
| 442 | ret = 1; |
| 443 | goto done; |
| 444 | } |
| 445 | } |
| 446 | |
| 447 | if (rd->row_queues[currq].rqueue.nr_dispatched >= |
| 448 | rd->row_queues[currq].disp_quantum) { |
| 449 | rd->row_queues[currq].rqueue.nr_dispatched = 0; |
| 450 | row_log_rowq(rd, currq, "Expiring rqueue"); |
| 451 | ret = row_choose_queue(rd); |
| 452 | if (ret) |
| 453 | row_dispatch_insert(rd); |
| 454 | goto done; |
| 455 | } |
| 456 | |
| 457 | /* Dispatch from curr_queue */ |
| 458 | if (list_empty(&rd->row_queues[currq].rqueue.fifo)) { |
| 459 | /* check idling */ |
| 460 | if (delayed_work_pending(&rd->read_idle.idle_work)) { |
Tatyana Brokhman | 25f3988 | 2012-10-15 20:56:02 +0200 | [diff] [blame] | 461 | if (force) { |
| 462 | (void)cancel_delayed_work( |
| 463 | &rd->read_idle.idle_work); |
| 464 | row_log_rowq(rd, currq, |
| 465 | "Canceled delayed work - forced dispatch"); |
| 466 | } else { |
| 467 | row_log_rowq(rd, currq, |
| 468 | "Delayed work pending. Exiting"); |
| 469 | goto done; |
| 470 | } |
Tatyana Brokhman | 1634906 | 2012-09-20 10:46:10 +0300 | [diff] [blame] | 471 | } |
| 472 | |
Tatyana Brokhman | 25f3988 | 2012-10-15 20:56:02 +0200 | [diff] [blame] | 473 | if (!force && queue_idling_enabled[currq] && |
Tatyana Brokhman | 1634906 | 2012-09-20 10:46:10 +0300 | [diff] [blame] | 474 | rd->row_queues[currq].rqueue.idle_data.begin_idling) { |
| 475 | if (!queue_delayed_work(rd->read_idle.idle_workqueue, |
| 476 | &rd->read_idle.idle_work, |
| 477 | jiffies + |
| 478 | msecs_to_jiffies(rd->read_idle.idle_time))) { |
| 479 | row_log_rowq(rd, currq, |
| 480 | "Work already on queue!"); |
| 481 | pr_err("ROW_BUG: Work already on queue!"); |
| 482 | } else |
| 483 | row_log_rowq(rd, currq, |
| 484 | "Scheduled delayed work. exiting"); |
| 485 | goto done; |
| 486 | } else { |
| 487 | row_log_rowq(rd, currq, |
| 488 | "Currq empty. Choose next queue"); |
| 489 | ret = row_choose_queue(rd); |
| 490 | if (!ret) |
| 491 | goto done; |
| 492 | } |
| 493 | } |
| 494 | |
| 495 | ret = 1; |
| 496 | row_dispatch_insert(rd); |
| 497 | |
| 498 | done: |
| 499 | return ret; |
| 500 | } |
| 501 | |
| 502 | /* |
| 503 | * row_init_queue() - Init scheduler data structures |
| 504 | * @q: requests queue |
| 505 | * |
| 506 | * Return pointer to struct row_data to be saved in elevator for |
| 507 | * this dispatch queue |
| 508 | * |
| 509 | */ |
| 510 | static void *row_init_queue(struct request_queue *q) |
| 511 | { |
| 512 | |
| 513 | struct row_data *rdata; |
| 514 | int i; |
| 515 | |
| 516 | rdata = kmalloc_node(sizeof(*rdata), |
| 517 | GFP_KERNEL | __GFP_ZERO, q->node); |
| 518 | if (!rdata) |
| 519 | return NULL; |
| 520 | |
| 521 | for (i = 0; i < ROWQ_MAX_PRIO; i++) { |
| 522 | INIT_LIST_HEAD(&rdata->row_queues[i].rqueue.fifo); |
| 523 | rdata->row_queues[i].disp_quantum = queue_quantum[i]; |
| 524 | rdata->row_queues[i].rqueue.rdata = rdata; |
| 525 | rdata->row_queues[i].rqueue.prio = i; |
| 526 | rdata->row_queues[i].rqueue.idle_data.begin_idling = false; |
| 527 | } |
| 528 | |
| 529 | /* |
| 530 | * Currently idling is enabled only for READ queues. If we want to |
| 531 | * enable it for write queues also, note that idling frequency will |
| 532 | * be the same in both cases |
| 533 | */ |
| 534 | rdata->read_idle.idle_time = ROW_IDLE_TIME; |
| 535 | rdata->read_idle.freq = ROW_READ_FREQ; |
| 536 | rdata->read_idle.idle_workqueue = alloc_workqueue("row_idle_work", |
| 537 | WQ_MEM_RECLAIM | WQ_HIGHPRI, 0); |
| 538 | if (!rdata->read_idle.idle_workqueue) |
| 539 | panic("Failed to create idle workqueue\n"); |
| 540 | INIT_DELAYED_WORK(&rdata->read_idle.idle_work, kick_queue); |
| 541 | |
| 542 | rdata->curr_queue = ROWQ_PRIO_HIGH_READ; |
| 543 | rdata->dispatch_queue = q; |
| 544 | |
| 545 | rdata->nr_reqs[READ] = rdata->nr_reqs[WRITE] = 0; |
| 546 | |
| 547 | return rdata; |
| 548 | } |
| 549 | |
| 550 | /* |
| 551 | * row_exit_queue() - called on unloading the RAW scheduler |
| 552 | * @e: poiner to struct elevator_queue |
| 553 | * |
| 554 | */ |
| 555 | static void row_exit_queue(struct elevator_queue *e) |
| 556 | { |
| 557 | struct row_data *rd = (struct row_data *)e->elevator_data; |
| 558 | int i; |
| 559 | |
| 560 | for (i = 0; i < ROWQ_MAX_PRIO; i++) |
| 561 | BUG_ON(!list_empty(&rd->row_queues[i].rqueue.fifo)); |
| 562 | (void)cancel_delayed_work_sync(&rd->read_idle.idle_work); |
| 563 | kfree(rd); |
| 564 | } |
| 565 | |
| 566 | /* |
| 567 | * row_merged_requests() - Called when 2 requests are merged |
| 568 | * @q: requests queue |
| 569 | * @rq: request the two requests were merged into |
| 570 | * @next: request that was merged |
| 571 | */ |
| 572 | static void row_merged_requests(struct request_queue *q, struct request *rq, |
| 573 | struct request *next) |
| 574 | { |
| 575 | struct row_queue *rqueue = RQ_ROWQ(next); |
| 576 | |
| 577 | list_del_init(&next->queuelist); |
| 578 | |
| 579 | rqueue->rdata->nr_reqs[rq_data_dir(rq)]--; |
| 580 | } |
| 581 | |
| 582 | /* |
| 583 | * get_queue_type() - Get queue type for a given request |
| 584 | * |
| 585 | * This is a helping function which purpose is to determine what |
| 586 | * ROW queue the given request should be added to (and |
| 587 | * dispatched from leter on) |
| 588 | * |
| 589 | * TODO: Right now only 3 queues are used REG_READ, REG_WRITE |
| 590 | * and REG_SWRITE |
| 591 | */ |
| 592 | static enum row_queue_prio get_queue_type(struct request *rq) |
| 593 | { |
| 594 | const int data_dir = rq_data_dir(rq); |
| 595 | const bool is_sync = rq_is_sync(rq); |
| 596 | |
| 597 | if (data_dir == READ) |
| 598 | return ROWQ_PRIO_REG_READ; |
| 599 | else if (is_sync) |
| 600 | return ROWQ_PRIO_REG_SWRITE; |
| 601 | else |
| 602 | return ROWQ_PRIO_REG_WRITE; |
| 603 | } |
| 604 | |
| 605 | /* |
| 606 | * row_set_request() - Set ROW data structures associated with this request. |
| 607 | * @q: requests queue |
| 608 | * @rq: pointer to the request |
| 609 | * @gfp_mask: ignored |
| 610 | * |
| 611 | */ |
| 612 | static int |
| 613 | row_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask) |
| 614 | { |
| 615 | struct row_data *rd = (struct row_data *)q->elevator->elevator_data; |
| 616 | unsigned long flags; |
| 617 | |
| 618 | spin_lock_irqsave(q->queue_lock, flags); |
| 619 | rq->elv.priv[0] = |
| 620 | (void *)(&rd->row_queues[get_queue_type(rq)]); |
| 621 | spin_unlock_irqrestore(q->queue_lock, flags); |
| 622 | |
| 623 | return 0; |
| 624 | } |
| 625 | |
| 626 | /********** Helping sysfs functions/defenitions for ROW attributes ******/ |
| 627 | static ssize_t row_var_show(int var, char *page) |
| 628 | { |
| 629 | return snprintf(page, 100, "%d\n", var); |
| 630 | } |
| 631 | |
| 632 | static ssize_t row_var_store(int *var, const char *page, size_t count) |
| 633 | { |
| 634 | int err; |
| 635 | err = kstrtoul(page, 10, (unsigned long *)var); |
| 636 | |
| 637 | return count; |
| 638 | } |
| 639 | |
| 640 | #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ |
| 641 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ |
| 642 | { \ |
| 643 | struct row_data *rowd = e->elevator_data; \ |
| 644 | int __data = __VAR; \ |
| 645 | if (__CONV) \ |
| 646 | __data = jiffies_to_msecs(__data); \ |
| 647 | return row_var_show(__data, (page)); \ |
| 648 | } |
| 649 | SHOW_FUNCTION(row_hp_read_quantum_show, |
| 650 | rowd->row_queues[ROWQ_PRIO_HIGH_READ].disp_quantum, 0); |
| 651 | SHOW_FUNCTION(row_rp_read_quantum_show, |
| 652 | rowd->row_queues[ROWQ_PRIO_REG_READ].disp_quantum, 0); |
| 653 | SHOW_FUNCTION(row_hp_swrite_quantum_show, |
| 654 | rowd->row_queues[ROWQ_PRIO_HIGH_SWRITE].disp_quantum, 0); |
| 655 | SHOW_FUNCTION(row_rp_swrite_quantum_show, |
| 656 | rowd->row_queues[ROWQ_PRIO_REG_SWRITE].disp_quantum, 0); |
| 657 | SHOW_FUNCTION(row_rp_write_quantum_show, |
| 658 | rowd->row_queues[ROWQ_PRIO_REG_WRITE].disp_quantum, 0); |
| 659 | SHOW_FUNCTION(row_lp_read_quantum_show, |
| 660 | rowd->row_queues[ROWQ_PRIO_LOW_READ].disp_quantum, 0); |
| 661 | SHOW_FUNCTION(row_lp_swrite_quantum_show, |
| 662 | rowd->row_queues[ROWQ_PRIO_LOW_SWRITE].disp_quantum, 0); |
| 663 | SHOW_FUNCTION(row_read_idle_show, rowd->read_idle.idle_time, 1); |
| 664 | SHOW_FUNCTION(row_read_idle_freq_show, rowd->read_idle.freq, 1); |
| 665 | #undef SHOW_FUNCTION |
| 666 | |
| 667 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ |
| 668 | static ssize_t __FUNC(struct elevator_queue *e, \ |
| 669 | const char *page, size_t count) \ |
| 670 | { \ |
| 671 | struct row_data *rowd = e->elevator_data; \ |
| 672 | int __data; \ |
| 673 | int ret = row_var_store(&__data, (page), count); \ |
| 674 | if (__CONV) \ |
| 675 | __data = (int)msecs_to_jiffies(__data); \ |
| 676 | if (__data < (MIN)) \ |
| 677 | __data = (MIN); \ |
| 678 | else if (__data > (MAX)) \ |
| 679 | __data = (MAX); \ |
| 680 | *(__PTR) = __data; \ |
| 681 | return ret; \ |
| 682 | } |
| 683 | STORE_FUNCTION(row_hp_read_quantum_store, |
Tatyana Brokhman | 0a0345a | 2012-10-15 20:50:54 +0200 | [diff] [blame] | 684 | &rowd->row_queues[ROWQ_PRIO_HIGH_READ].disp_quantum, 1, INT_MAX, 0); |
Tatyana Brokhman | 1634906 | 2012-09-20 10:46:10 +0300 | [diff] [blame] | 685 | STORE_FUNCTION(row_rp_read_quantum_store, |
Tatyana Brokhman | 0a0345a | 2012-10-15 20:50:54 +0200 | [diff] [blame] | 686 | &rowd->row_queues[ROWQ_PRIO_REG_READ].disp_quantum, |
| 687 | 1, INT_MAX, 0); |
Tatyana Brokhman | 1634906 | 2012-09-20 10:46:10 +0300 | [diff] [blame] | 688 | STORE_FUNCTION(row_hp_swrite_quantum_store, |
Tatyana Brokhman | 0a0345a | 2012-10-15 20:50:54 +0200 | [diff] [blame] | 689 | &rowd->row_queues[ROWQ_PRIO_HIGH_SWRITE].disp_quantum, |
| 690 | 1, INT_MAX, 0); |
Tatyana Brokhman | 1634906 | 2012-09-20 10:46:10 +0300 | [diff] [blame] | 691 | STORE_FUNCTION(row_rp_swrite_quantum_store, |
Tatyana Brokhman | 0a0345a | 2012-10-15 20:50:54 +0200 | [diff] [blame] | 692 | &rowd->row_queues[ROWQ_PRIO_REG_SWRITE].disp_quantum, |
| 693 | 1, INT_MAX, 0); |
Tatyana Brokhman | 1634906 | 2012-09-20 10:46:10 +0300 | [diff] [blame] | 694 | STORE_FUNCTION(row_rp_write_quantum_store, |
Tatyana Brokhman | 0a0345a | 2012-10-15 20:50:54 +0200 | [diff] [blame] | 695 | &rowd->row_queues[ROWQ_PRIO_REG_WRITE].disp_quantum, |
| 696 | 1, INT_MAX, 0); |
Tatyana Brokhman | 1634906 | 2012-09-20 10:46:10 +0300 | [diff] [blame] | 697 | STORE_FUNCTION(row_lp_read_quantum_store, |
Tatyana Brokhman | 0a0345a | 2012-10-15 20:50:54 +0200 | [diff] [blame] | 698 | &rowd->row_queues[ROWQ_PRIO_LOW_READ].disp_quantum, |
| 699 | 1, INT_MAX, 0); |
Tatyana Brokhman | 1634906 | 2012-09-20 10:46:10 +0300 | [diff] [blame] | 700 | STORE_FUNCTION(row_lp_swrite_quantum_store, |
Tatyana Brokhman | 0a0345a | 2012-10-15 20:50:54 +0200 | [diff] [blame] | 701 | &rowd->row_queues[ROWQ_PRIO_LOW_SWRITE].disp_quantum, |
| 702 | 1, INT_MAX, 1); |
Tatyana Brokhman | 1634906 | 2012-09-20 10:46:10 +0300 | [diff] [blame] | 703 | STORE_FUNCTION(row_read_idle_store, &rowd->read_idle.idle_time, 1, INT_MAX, 1); |
Tatyana Brokhman | 0a0345a | 2012-10-15 20:50:54 +0200 | [diff] [blame] | 704 | STORE_FUNCTION(row_read_idle_freq_store, &rowd->read_idle.freq, 1, INT_MAX, 1); |
Tatyana Brokhman | 1634906 | 2012-09-20 10:46:10 +0300 | [diff] [blame] | 705 | |
| 706 | #undef STORE_FUNCTION |
| 707 | |
| 708 | #define ROW_ATTR(name) \ |
| 709 | __ATTR(name, S_IRUGO|S_IWUSR, row_##name##_show, \ |
| 710 | row_##name##_store) |
| 711 | |
| 712 | static struct elv_fs_entry row_attrs[] = { |
| 713 | ROW_ATTR(hp_read_quantum), |
| 714 | ROW_ATTR(rp_read_quantum), |
| 715 | ROW_ATTR(hp_swrite_quantum), |
| 716 | ROW_ATTR(rp_swrite_quantum), |
| 717 | ROW_ATTR(rp_write_quantum), |
| 718 | ROW_ATTR(lp_read_quantum), |
| 719 | ROW_ATTR(lp_swrite_quantum), |
| 720 | ROW_ATTR(read_idle), |
| 721 | ROW_ATTR(read_idle_freq), |
| 722 | __ATTR_NULL |
| 723 | }; |
| 724 | |
| 725 | static struct elevator_type iosched_row = { |
| 726 | .ops = { |
| 727 | .elevator_merge_req_fn = row_merged_requests, |
| 728 | .elevator_dispatch_fn = row_dispatch_requests, |
| 729 | .elevator_add_req_fn = row_add_request, |
Tatyana Brokhman | b7bf9ac | 2012-10-30 08:33:06 +0200 | [diff] [blame] | 730 | .elevator_reinsert_req_fn = row_reinsert_req, |
Tatyana Brokhman | 0ef8143 | 2012-12-20 19:23:58 +0200 | [diff] [blame^] | 731 | .elevator_is_urgent_fn = row_urgent_pending, |
Tatyana Brokhman | 1634906 | 2012-09-20 10:46:10 +0300 | [diff] [blame] | 732 | .elevator_former_req_fn = elv_rb_former_request, |
| 733 | .elevator_latter_req_fn = elv_rb_latter_request, |
| 734 | .elevator_set_req_fn = row_set_request, |
| 735 | .elevator_init_fn = row_init_queue, |
| 736 | .elevator_exit_fn = row_exit_queue, |
| 737 | }, |
| 738 | |
| 739 | .elevator_attrs = row_attrs, |
| 740 | .elevator_name = "row", |
| 741 | .elevator_owner = THIS_MODULE, |
| 742 | }; |
| 743 | |
| 744 | static int __init row_init(void) |
| 745 | { |
| 746 | elv_register(&iosched_row); |
| 747 | return 0; |
| 748 | } |
| 749 | |
| 750 | static void __exit row_exit(void) |
| 751 | { |
| 752 | elv_unregister(&iosched_row); |
| 753 | } |
| 754 | |
| 755 | module_init(row_init); |
| 756 | module_exit(row_exit); |
| 757 | |
| 758 | MODULE_LICENSE("GPLv2"); |
| 759 | MODULE_DESCRIPTION("Read Over Write IO scheduler"); |