aboutsummaryrefslogtreecommitdiffstats
path: root/drivers/block/cfq-iosched.c
diff options
context:
space:
mode:
Diffstat (limited to 'drivers/block/cfq-iosched.c')
-rw-r--r--drivers/block/cfq-iosched.c2097
1 files changed, 1445 insertions, 652 deletions
diff --git a/drivers/block/cfq-iosched.c b/drivers/block/cfq-iosched.c
index 2210bacad56a..cd056e7e64ec 100644
--- a/drivers/block/cfq-iosched.c
+++ b/drivers/block/cfq-iosched.c
@@ -21,22 +21,34 @@
21#include <linux/hash.h> 21#include <linux/hash.h>
22#include <linux/rbtree.h> 22#include <linux/rbtree.h>
23#include <linux/mempool.h> 23#include <linux/mempool.h>
24 24#include <linux/ioprio.h>
25static unsigned long max_elapsed_crq; 25#include <linux/writeback.h>
26static unsigned long max_elapsed_dispatch;
27 26
28/* 27/*
29 * tunables 28 * tunables
30 */ 29 */
31static int cfq_quantum = 4; /* max queue in one round of service */ 30static int cfq_quantum = 4; /* max queue in one round of service */
32static int cfq_queued = 8; /* minimum rq allocate limit per-queue*/ 31static int cfq_queued = 8; /* minimum rq allocate limit per-queue*/
33static int cfq_service = HZ; /* period over which service is avg */ 32static int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
34static int cfq_fifo_expire_r = HZ / 2; /* fifo timeout for sync requests */
35static int cfq_fifo_expire_w = 5 * HZ; /* fifo timeout for async requests */
36static int cfq_fifo_rate = HZ / 8; /* fifo expiry rate */
37static int cfq_back_max = 16 * 1024; /* maximum backwards seek, in KiB */ 33static int cfq_back_max = 16 * 1024; /* maximum backwards seek, in KiB */
38static int cfq_back_penalty = 2; /* penalty of a backwards seek */ 34static int cfq_back_penalty = 2; /* penalty of a backwards seek */
39 35
36static int cfq_slice_sync = HZ / 10;
37static int cfq_slice_async = HZ / 25;
38static int cfq_slice_async_rq = 2;
39static int cfq_slice_idle = HZ / 100;
40
41#define CFQ_IDLE_GRACE (HZ / 10)
42#define CFQ_SLICE_SCALE (5)
43
44#define CFQ_KEY_ASYNC (0)
45#define CFQ_KEY_ANY (0xffff)
46
47/*
48 * disable queueing at the driver/hardware level
49 */
50static int cfq_max_depth = 2;
51
40/* 52/*
41 * for the hash of cfqq inside the cfqd 53 * for the hash of cfqq inside the cfqd
42 */ 54 */
@@ -55,6 +67,7 @@ static int cfq_back_penalty = 2; /* penalty of a backwards seek */
55#define list_entry_hash(ptr) hlist_entry((ptr), struct cfq_rq, hash) 67#define list_entry_hash(ptr) hlist_entry((ptr), struct cfq_rq, hash)
56 68
57#define list_entry_cfqq(ptr) list_entry((ptr), struct cfq_queue, cfq_list) 69#define list_entry_cfqq(ptr) list_entry((ptr), struct cfq_queue, cfq_list)
70#define list_entry_fifo(ptr) list_entry((ptr), struct request, queuelist)
58 71
59#define RQ_DATA(rq) (rq)->elevator_private 72#define RQ_DATA(rq) (rq)->elevator_private
60 73
@@ -75,78 +88,110 @@ static int cfq_back_penalty = 2; /* penalty of a backwards seek */
75#define rb_entry_crq(node) rb_entry((node), struct cfq_rq, rb_node) 88#define rb_entry_crq(node) rb_entry((node), struct cfq_rq, rb_node)
76#define rq_rb_key(rq) (rq)->sector 89#define rq_rb_key(rq) (rq)->sector
77 90
78/*
79 * threshold for switching off non-tag accounting
80 */
81#define CFQ_MAX_TAG (4)
82
83/*
84 * sort key types and names
85 */
86enum {
87 CFQ_KEY_PGID,
88 CFQ_KEY_TGID,
89 CFQ_KEY_UID,
90 CFQ_KEY_GID,
91 CFQ_KEY_LAST,
92};
93
94static char *cfq_key_types[] = { "pgid", "tgid", "uid", "gid", NULL };
95
96static kmem_cache_t *crq_pool; 91static kmem_cache_t *crq_pool;
97static kmem_cache_t *cfq_pool; 92static kmem_cache_t *cfq_pool;
98static kmem_cache_t *cfq_ioc_pool; 93static kmem_cache_t *cfq_ioc_pool;
99 94
95#define CFQ_PRIO_LISTS IOPRIO_BE_NR
96#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
97#define cfq_class_be(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
98#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
99
100#define ASYNC (0)
101#define SYNC (1)
102
103#define cfq_cfqq_dispatched(cfqq) \
104 ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
105
106#define cfq_cfqq_class_sync(cfqq) ((cfqq)->key != CFQ_KEY_ASYNC)
107
108#define cfq_cfqq_sync(cfqq) \
109 (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
110
111/*
112 * Per block device queue structure
113 */
100struct cfq_data { 114struct cfq_data {
101 struct list_head rr_list; 115 atomic_t ref;
116 request_queue_t *queue;
117
118 /*
119 * rr list of queues with requests and the count of them
120 */
121 struct list_head rr_list[CFQ_PRIO_LISTS];
122 struct list_head busy_rr;
123 struct list_head cur_rr;
124 struct list_head idle_rr;
125 unsigned int busy_queues;
126
127 /*
128 * non-ordered list of empty cfqq's
129 */
102 struct list_head empty_list; 130 struct list_head empty_list;
103 131
132 /*
133 * cfqq lookup hash
134 */
104 struct hlist_head *cfq_hash; 135 struct hlist_head *cfq_hash;
105 struct hlist_head *crq_hash;
106 136
107 /* queues on rr_list (ie they have pending requests */ 137 /*
108 unsigned int busy_queues; 138 * global crq hash for all queues
139 */
140 struct hlist_head *crq_hash;
109 141
110 unsigned int max_queued; 142 unsigned int max_queued;
111 143
112 atomic_t ref; 144 mempool_t *crq_pool;
113 145
114 int key_type; 146 int rq_in_driver;
115 147
116 mempool_t *crq_pool; 148 /*
149 * schedule slice state info
150 */
151 /*
152 * idle window management
153 */
154 struct timer_list idle_slice_timer;
155 struct work_struct unplug_work;
117 156
118 request_queue_t *queue; 157 struct cfq_queue *active_queue;
158 struct cfq_io_context *active_cic;
159 int cur_prio, cur_end_prio;
160 unsigned int dispatch_slice;
161
162 struct timer_list idle_class_timer;
119 163
120 sector_t last_sector; 164 sector_t last_sector;
165 unsigned long last_end_request;
121 166
122 int rq_in_driver; 167 unsigned int rq_starved;
123 168
124 /* 169 /*
125 * tunables, see top of file 170 * tunables, see top of file
126 */ 171 */
127 unsigned int cfq_quantum; 172 unsigned int cfq_quantum;
128 unsigned int cfq_queued; 173 unsigned int cfq_queued;
129 unsigned int cfq_fifo_expire_r; 174 unsigned int cfq_fifo_expire[2];
130 unsigned int cfq_fifo_expire_w;
131 unsigned int cfq_fifo_batch_expire;
132 unsigned int cfq_back_penalty; 175 unsigned int cfq_back_penalty;
133 unsigned int cfq_back_max; 176 unsigned int cfq_back_max;
134 unsigned int find_best_crq; 177 unsigned int cfq_slice[2];
135 178 unsigned int cfq_slice_async_rq;
136 unsigned int cfq_tagged; 179 unsigned int cfq_slice_idle;
180 unsigned int cfq_max_depth;
137}; 181};
138 182
183/*
184 * Per process-grouping structure
185 */
139struct cfq_queue { 186struct cfq_queue {
140 /* reference count */ 187 /* reference count */
141 atomic_t ref; 188 atomic_t ref;
142 /* parent cfq_data */ 189 /* parent cfq_data */
143 struct cfq_data *cfqd; 190 struct cfq_data *cfqd;
144 /* hash of mergeable requests */ 191 /* cfqq lookup hash */
145 struct hlist_node cfq_hash; 192 struct hlist_node cfq_hash;
146 /* hash key */ 193 /* hash key */
147 unsigned long key; 194 unsigned int key;
148 /* whether queue is on rr (or empty) list */
149 int on_rr;
150 /* on either rr or empty list of cfqd */ 195 /* on either rr or empty list of cfqd */
151 struct list_head cfq_list; 196 struct list_head cfq_list;
152 /* sorted list of pending requests */ 197 /* sorted list of pending requests */
@@ -158,21 +203,22 @@ struct cfq_queue {
158 /* currently allocated requests */ 203 /* currently allocated requests */
159 int allocated[2]; 204 int allocated[2];
160 /* fifo list of requests in sort_list */ 205 /* fifo list of requests in sort_list */
161 struct list_head fifo[2]; 206 struct list_head fifo;
162 /* last time fifo expired */
163 unsigned long last_fifo_expire;
164 207
165 int key_type; 208 unsigned long slice_start;
209 unsigned long slice_end;
210 unsigned long slice_left;
211 unsigned long service_last;
166 212
167 unsigned long service_start; 213 /* number of requests that are on the dispatch list */
168 unsigned long service_used; 214 int on_dispatch[2];
169 215
170 unsigned int max_rate; 216 /* io prio of this group */
217 unsigned short ioprio, org_ioprio;
218 unsigned short ioprio_class, org_ioprio_class;
171 219
172 /* number of requests that have been handed to the driver */ 220 /* various state flags, see below */
173 int in_flight; 221 unsigned int flags;
174 /* number of currently allocated requests */
175 int alloc_limit[2];
176}; 222};
177 223
178struct cfq_rq { 224struct cfq_rq {
@@ -184,42 +230,78 @@ struct cfq_rq {
184 struct cfq_queue *cfq_queue; 230 struct cfq_queue *cfq_queue;
185 struct cfq_io_context *io_context; 231 struct cfq_io_context *io_context;
186 232
187 unsigned long service_start; 233 unsigned int crq_flags;
188 unsigned long queue_start; 234};
235
236enum cfqq_state_flags {
237 CFQ_CFQQ_FLAG_on_rr = 0,
238 CFQ_CFQQ_FLAG_wait_request,
239 CFQ_CFQQ_FLAG_must_alloc,
240 CFQ_CFQQ_FLAG_must_alloc_slice,
241 CFQ_CFQQ_FLAG_must_dispatch,
242 CFQ_CFQQ_FLAG_fifo_expire,
243 CFQ_CFQQ_FLAG_idle_window,
244 CFQ_CFQQ_FLAG_prio_changed,
245 CFQ_CFQQ_FLAG_expired,
246};
189 247
190 unsigned int in_flight : 1; 248#define CFQ_CFQQ_FNS(name) \
191 unsigned int accounted : 1; 249static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
192 unsigned int is_sync : 1; 250{ \
193 unsigned int is_write : 1; 251 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
252} \
253static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
254{ \
255 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
256} \
257static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
258{ \
259 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
260}
261
262CFQ_CFQQ_FNS(on_rr);
263CFQ_CFQQ_FNS(wait_request);
264CFQ_CFQQ_FNS(must_alloc);
265CFQ_CFQQ_FNS(must_alloc_slice);
266CFQ_CFQQ_FNS(must_dispatch);
267CFQ_CFQQ_FNS(fifo_expire);
268CFQ_CFQQ_FNS(idle_window);
269CFQ_CFQQ_FNS(prio_changed);
270CFQ_CFQQ_FNS(expired);
271#undef CFQ_CFQQ_FNS
272
273enum cfq_rq_state_flags {
274 CFQ_CRQ_FLAG_in_flight = 0,
275 CFQ_CRQ_FLAG_in_driver,
276 CFQ_CRQ_FLAG_is_sync,
277 CFQ_CRQ_FLAG_requeued,
194}; 278};
195 279
196static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned long); 280#define CFQ_CRQ_FNS(name) \
281static inline void cfq_mark_crq_##name(struct cfq_rq *crq) \
282{ \
283 crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name); \
284} \
285static inline void cfq_clear_crq_##name(struct cfq_rq *crq) \
286{ \
287 crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name); \
288} \
289static inline int cfq_crq_##name(const struct cfq_rq *crq) \
290{ \
291 return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0; \
292}
293
294CFQ_CRQ_FNS(in_flight);
295CFQ_CRQ_FNS(in_driver);
296CFQ_CRQ_FNS(is_sync);
297CFQ_CRQ_FNS(requeued);
298#undef CFQ_CRQ_FNS
299
300static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
197static void cfq_dispatch_sort(request_queue_t *, struct cfq_rq *); 301static void cfq_dispatch_sort(request_queue_t *, struct cfq_rq *);
198static void cfq_update_next_crq(struct cfq_rq *);
199static void cfq_put_cfqd(struct cfq_data *cfqd); 302static void cfq_put_cfqd(struct cfq_data *cfqd);
200 303
201/* 304#define process_sync(tsk) ((tsk)->flags & PF_SYNCWRITE)
202 * what the fairness is based on (ie how processes are grouped and
203 * differentiated)
204 */
205static inline unsigned long
206cfq_hash_key(struct cfq_data *cfqd, struct task_struct *tsk)
207{
208 /*
209 * optimize this so that ->key_type is the offset into the struct
210 */
211 switch (cfqd->key_type) {
212 case CFQ_KEY_PGID:
213 return process_group(tsk);
214 default:
215 case CFQ_KEY_TGID:
216 return tsk->tgid;
217 case CFQ_KEY_UID:
218 return tsk->uid;
219 case CFQ_KEY_GID:
220 return tsk->gid;
221 }
222}
223 305
224/* 306/*
225 * lots of deadline iosched dupes, can be abstracted later... 307 * lots of deadline iosched dupes, can be abstracted later...
@@ -235,16 +317,12 @@ static void cfq_remove_merge_hints(request_queue_t *q, struct cfq_rq *crq)
235 317
236 if (q->last_merge == crq->request) 318 if (q->last_merge == crq->request)
237 q->last_merge = NULL; 319 q->last_merge = NULL;
238
239 cfq_update_next_crq(crq);
240} 320}
241 321
242static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq) 322static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq)
243{ 323{
244 const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request)); 324 const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request));
245 325
246 BUG_ON(!hlist_unhashed(&crq->hash));
247
248 hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]); 326 hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]);
249} 327}
250 328
@@ -257,8 +335,6 @@ static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
257 struct cfq_rq *crq = list_entry_hash(entry); 335 struct cfq_rq *crq = list_entry_hash(entry);
258 struct request *__rq = crq->request; 336 struct request *__rq = crq->request;
259 337
260 BUG_ON(hlist_unhashed(&crq->hash));
261
262 if (!rq_mergeable(__rq)) { 338 if (!rq_mergeable(__rq)) {
263 cfq_del_crq_hash(crq); 339 cfq_del_crq_hash(crq);
264 continue; 340 continue;
@@ -271,6 +347,28 @@ static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
271 return NULL; 347 return NULL;
272} 348}
273 349
350static inline int cfq_pending_requests(struct cfq_data *cfqd)
351{
352 return !list_empty(&cfqd->queue->queue_head) || cfqd->busy_queues;
353}
354
355/*
356 * scheduler run of queue, if there are requests pending and no one in the
357 * driver that will restart queueing
358 */
359static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
360{
361 if (!cfqd->rq_in_driver && cfq_pending_requests(cfqd))
362 kblockd_schedule_work(&cfqd->unplug_work);
363}
364
365static int cfq_queue_empty(request_queue_t *q)
366{
367 struct cfq_data *cfqd = q->elevator->elevator_data;
368
369 return !cfq_pending_requests(cfqd);
370}
371
274/* 372/*
275 * Lifted from AS - choose which of crq1 and crq2 that is best served now. 373 * Lifted from AS - choose which of crq1 and crq2 that is best served now.
276 * We choose the request that is closest to the head right now. Distance 374 * We choose the request that is closest to the head right now. Distance
@@ -288,35 +386,21 @@ cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
288 if (crq2 == NULL) 386 if (crq2 == NULL)
289 return crq1; 387 return crq1;
290 388
389 if (cfq_crq_requeued(crq1) && !cfq_crq_requeued(crq2))
390 return crq1;
391 else if (cfq_crq_requeued(crq2) && !cfq_crq_requeued(crq1))
392 return crq2;
393
394 if (cfq_crq_is_sync(crq1) && !cfq_crq_is_sync(crq2))
395 return crq1;
396 else if (cfq_crq_is_sync(crq2) && !cfq_crq_is_sync(crq1))
397 return crq2;
398
291 s1 = crq1->request->sector; 399 s1 = crq1->request->sector;
292 s2 = crq2->request->sector; 400 s2 = crq2->request->sector;
293 401
294 last = cfqd->last_sector; 402 last = cfqd->last_sector;
295 403
296#if 0
297 if (!list_empty(&cfqd->queue->queue_head)) {
298 struct list_head *entry = &cfqd->queue->queue_head;
299 unsigned long distance = ~0UL;
300 struct request *rq;
301
302 while ((entry = entry->prev) != &cfqd->queue->queue_head) {
303 rq = list_entry_rq(entry);
304
305 if (blk_barrier_rq(rq))
306 break;
307
308 if (distance < abs(s1 - rq->sector + rq->nr_sectors)) {
309 distance = abs(s1 - rq->sector +rq->nr_sectors);
310 last = rq->sector + rq->nr_sectors;
311 }
312 if (distance < abs(s2 - rq->sector + rq->nr_sectors)) {
313 distance = abs(s2 - rq->sector +rq->nr_sectors);
314 last = rq->sector + rq->nr_sectors;
315 }
316 }
317 }
318#endif
319
320 /* 404 /*
321 * by definition, 1KiB is 2 sectors 405 * by definition, 1KiB is 2 sectors
322 */ 406 */
@@ -377,11 +461,14 @@ cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
377 struct cfq_rq *crq_next = NULL, *crq_prev = NULL; 461 struct cfq_rq *crq_next = NULL, *crq_prev = NULL;
378 struct rb_node *rbnext, *rbprev; 462 struct rb_node *rbnext, *rbprev;
379 463
380 if (!ON_RB(&last->rb_node)) 464 rbnext = NULL;
381 return NULL; 465 if (ON_RB(&last->rb_node))
382 466 rbnext = rb_next(&last->rb_node);
383 if ((rbnext = rb_next(&last->rb_node)) == NULL) 467 if (!rbnext) {
384 rbnext = rb_first(&cfqq->sort_list); 468 rbnext = rb_first(&cfqq->sort_list);
469 if (rbnext == &last->rb_node)
470 rbnext = NULL;
471 }
385 472
386 rbprev = rb_prev(&last->rb_node); 473 rbprev = rb_prev(&last->rb_node);
387 474
@@ -401,67 +488,53 @@ static void cfq_update_next_crq(struct cfq_rq *crq)
401 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq); 488 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
402} 489}
403 490
404static int cfq_check_sort_rr_list(struct cfq_queue *cfqq) 491static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
405{ 492{
406 struct list_head *head = &cfqq->cfqd->rr_list; 493 struct cfq_data *cfqd = cfqq->cfqd;
407 struct list_head *next, *prev; 494 struct list_head *list, *entry;
408
409 /*
410 * list might still be ordered
411 */
412 next = cfqq->cfq_list.next;
413 if (next != head) {
414 struct cfq_queue *cnext = list_entry_cfqq(next);
415 495
416 if (cfqq->service_used > cnext->service_used) 496 BUG_ON(!cfq_cfqq_on_rr(cfqq));
417 return 1;
418 }
419 497
420 prev = cfqq->cfq_list.prev; 498 list_del(&cfqq->cfq_list);
421 if (prev != head) {
422 struct cfq_queue *cprev = list_entry_cfqq(prev);
423 499
424 if (cfqq->service_used < cprev->service_used) 500 if (cfq_class_rt(cfqq))
425 return 1; 501 list = &cfqd->cur_rr;
502 else if (cfq_class_idle(cfqq))
503 list = &cfqd->idle_rr;
504 else {
505 /*
506 * if cfqq has requests in flight, don't allow it to be
507 * found in cfq_set_active_queue before it has finished them.
508 * this is done to increase fairness between a process that
509 * has lots of io pending vs one that only generates one
510 * sporadically or synchronously
511 */
512 if (cfq_cfqq_dispatched(cfqq))
513 list = &cfqd->busy_rr;
514 else
515 list = &cfqd->rr_list[cfqq->ioprio];
426 } 516 }
427 517
428 return 0; 518 /*
429} 519 * if queue was preempted, just add to front to be fair. busy_rr
430 520 * isn't sorted.
431static void cfq_sort_rr_list(struct cfq_queue *cfqq, int new_queue) 521 */
432{ 522 if (preempted || list == &cfqd->busy_rr) {
433 struct list_head *entry = &cfqq->cfqd->rr_list; 523 list_add(&cfqq->cfq_list, list);
434
435 if (!cfqq->on_rr)
436 return;
437 if (!new_queue && !cfq_check_sort_rr_list(cfqq))
438 return; 524 return;
439 525 }
440 list_del(&cfqq->cfq_list);
441 526
442 /* 527 /*
443 * sort by our mean service_used, sub-sort by in-flight requests 528 * sort by when queue was last serviced
444 */ 529 */
445 while ((entry = entry->prev) != &cfqq->cfqd->rr_list) { 530 entry = list;
531 while ((entry = entry->prev) != list) {
446 struct cfq_queue *__cfqq = list_entry_cfqq(entry); 532 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
447 533
448 if (cfqq->service_used > __cfqq->service_used) 534 if (!__cfqq->service_last)
535 break;
536 if (time_before(__cfqq->service_last, cfqq->service_last))
449 break; 537 break;
450 else if (cfqq->service_used == __cfqq->service_used) {
451 struct list_head *prv;
452
453 while ((prv = entry->prev) != &cfqq->cfqd->rr_list) {
454 __cfqq = list_entry_cfqq(prv);
455
456 WARN_ON(__cfqq->service_used > cfqq->service_used);
457 if (cfqq->service_used != __cfqq->service_used)
458 break;
459 if (cfqq->in_flight > __cfqq->in_flight)
460 break;
461
462 entry = prv;
463 }
464 }
465 } 538 }
466 539
467 list_add(&cfqq->cfq_list, entry); 540 list_add(&cfqq->cfq_list, entry);
@@ -469,28 +542,24 @@ static void cfq_sort_rr_list(struct cfq_queue *cfqq, int new_queue)
469 542
470/* 543/*
471 * add to busy list of queues for service, trying to be fair in ordering 544 * add to busy list of queues for service, trying to be fair in ordering
472 * the pending list according to requests serviced 545 * the pending list according to last request service
473 */ 546 */
474static inline void 547static inline void
475cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) 548cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq, int requeue)
476{ 549{
477 /* 550 BUG_ON(cfq_cfqq_on_rr(cfqq));
478 * it's currently on the empty list 551 cfq_mark_cfqq_on_rr(cfqq);
479 */
480 cfqq->on_rr = 1;
481 cfqd->busy_queues++; 552 cfqd->busy_queues++;
482 553
483 if (time_after(jiffies, cfqq->service_start + cfq_service)) 554 cfq_resort_rr_list(cfqq, requeue);
484 cfqq->service_used >>= 3;
485
486 cfq_sort_rr_list(cfqq, 1);
487} 555}
488 556
489static inline void 557static inline void
490cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) 558cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
491{ 559{
560 BUG_ON(!cfq_cfqq_on_rr(cfqq));
561 cfq_clear_cfqq_on_rr(cfqq);
492 list_move(&cfqq->cfq_list, &cfqd->empty_list); 562 list_move(&cfqq->cfq_list, &cfqd->empty_list);
493 cfqq->on_rr = 0;
494 563
495 BUG_ON(!cfqd->busy_queues); 564 BUG_ON(!cfqd->busy_queues);
496 cfqd->busy_queues--; 565 cfqd->busy_queues--;
@@ -505,16 +574,17 @@ static inline void cfq_del_crq_rb(struct cfq_rq *crq)
505 574
506 if (ON_RB(&crq->rb_node)) { 575 if (ON_RB(&crq->rb_node)) {
507 struct cfq_data *cfqd = cfqq->cfqd; 576 struct cfq_data *cfqd = cfqq->cfqd;
577 const int sync = cfq_crq_is_sync(crq);
508 578
509 BUG_ON(!cfqq->queued[crq->is_sync]); 579 BUG_ON(!cfqq->queued[sync]);
580 cfqq->queued[sync]--;
510 581
511 cfq_update_next_crq(crq); 582 cfq_update_next_crq(crq);
512 583
513 cfqq->queued[crq->is_sync]--;
514 rb_erase(&crq->rb_node, &cfqq->sort_list); 584 rb_erase(&crq->rb_node, &cfqq->sort_list);
515 RB_CLEAR_COLOR(&crq->rb_node); 585 RB_CLEAR_COLOR(&crq->rb_node);
516 586
517 if (RB_EMPTY(&cfqq->sort_list) && cfqq->on_rr) 587 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY(&cfqq->sort_list))
518 cfq_del_cfqq_rr(cfqd, cfqq); 588 cfq_del_cfqq_rr(cfqd, cfqq);
519 } 589 }
520} 590}
@@ -550,7 +620,7 @@ static void cfq_add_crq_rb(struct cfq_rq *crq)
550 struct cfq_rq *__alias; 620 struct cfq_rq *__alias;
551 621
552 crq->rb_key = rq_rb_key(rq); 622 crq->rb_key = rq_rb_key(rq);
553 cfqq->queued[crq->is_sync]++; 623 cfqq->queued[cfq_crq_is_sync(crq)]++;
554 624
555 /* 625 /*
556 * looks a little odd, but the first insert might return an alias. 626 * looks a little odd, but the first insert might return an alias.
@@ -561,8 +631,8 @@ static void cfq_add_crq_rb(struct cfq_rq *crq)
561 631
562 rb_insert_color(&crq->rb_node, &cfqq->sort_list); 632 rb_insert_color(&crq->rb_node, &cfqq->sort_list);
563 633
564 if (!cfqq->on_rr) 634 if (!cfq_cfqq_on_rr(cfqq))
565 cfq_add_cfqq_rr(cfqd, cfqq); 635 cfq_add_cfqq_rr(cfqd, cfqq, cfq_crq_requeued(crq));
566 636
567 /* 637 /*
568 * check if this request is a better next-serve candidate 638 * check if this request is a better next-serve candidate
@@ -575,17 +645,16 @@ cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
575{ 645{
576 if (ON_RB(&crq->rb_node)) { 646 if (ON_RB(&crq->rb_node)) {
577 rb_erase(&crq->rb_node, &cfqq->sort_list); 647 rb_erase(&crq->rb_node, &cfqq->sort_list);
578 cfqq->queued[crq->is_sync]--; 648 cfqq->queued[cfq_crq_is_sync(crq)]--;
579 } 649 }
580 650
581 cfq_add_crq_rb(crq); 651 cfq_add_crq_rb(crq);
582} 652}
583 653
584static struct request * 654static struct request *cfq_find_rq_rb(struct cfq_data *cfqd, sector_t sector)
585cfq_find_rq_rb(struct cfq_data *cfqd, sector_t sector) 655
586{ 656{
587 const unsigned long key = cfq_hash_key(cfqd, current); 657 struct cfq_queue *cfqq = cfq_find_cfq_hash(cfqd, current->pid, CFQ_KEY_ANY);
588 struct cfq_queue *cfqq = cfq_find_cfq_hash(cfqd, key);
589 struct rb_node *n; 658 struct rb_node *n;
590 659
591 if (!cfqq) 660 if (!cfqq)
@@ -609,20 +678,25 @@ out:
609 678
610static void cfq_deactivate_request(request_queue_t *q, struct request *rq) 679static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
611{ 680{
681 struct cfq_data *cfqd = q->elevator->elevator_data;
612 struct cfq_rq *crq = RQ_DATA(rq); 682 struct cfq_rq *crq = RQ_DATA(rq);
613 683
614 if (crq) { 684 if (crq) {
615 struct cfq_queue *cfqq = crq->cfq_queue; 685 struct cfq_queue *cfqq = crq->cfq_queue;
616 686
617 if (cfqq->cfqd->cfq_tagged) { 687 if (cfq_crq_in_driver(crq)) {
618 cfqq->service_used--; 688 cfq_clear_crq_in_driver(crq);
619 cfq_sort_rr_list(cfqq, 0); 689 WARN_ON(!cfqd->rq_in_driver);
690 cfqd->rq_in_driver--;
620 } 691 }
692 if (cfq_crq_in_flight(crq)) {
693 const int sync = cfq_crq_is_sync(crq);
621 694
622 if (crq->accounted) { 695 cfq_clear_crq_in_flight(crq);
623 crq->accounted = 0; 696 WARN_ON(!cfqq->on_dispatch[sync]);
624 cfqq->cfqd->rq_in_driver--; 697 cfqq->on_dispatch[sync]--;
625 } 698 }
699 cfq_mark_crq_requeued(crq);
626 } 700 }
627} 701}
628 702
@@ -640,11 +714,10 @@ static void cfq_remove_request(request_queue_t *q, struct request *rq)
640 struct cfq_rq *crq = RQ_DATA(rq); 714 struct cfq_rq *crq = RQ_DATA(rq);
641 715
642 if (crq) { 716 if (crq) {
643 cfq_remove_merge_hints(q, crq);
644 list_del_init(&rq->queuelist); 717 list_del_init(&rq->queuelist);
718 cfq_del_crq_rb(crq);
719 cfq_remove_merge_hints(q, crq);
645 720
646 if (crq->cfq_queue)
647 cfq_del_crq_rb(crq);
648 } 721 }
649} 722}
650 723
@@ -662,21 +735,15 @@ cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
662 } 735 }
663 736
664 __rq = cfq_find_rq_hash(cfqd, bio->bi_sector); 737 __rq = cfq_find_rq_hash(cfqd, bio->bi_sector);
665 if (__rq) { 738 if (__rq && elv_rq_merge_ok(__rq, bio)) {
666 BUG_ON(__rq->sector + __rq->nr_sectors != bio->bi_sector); 739 ret = ELEVATOR_BACK_MERGE;
667 740 goto out;
668 if (elv_rq_merge_ok(__rq, bio)) {
669 ret = ELEVATOR_BACK_MERGE;
670 goto out;
671 }
672 } 741 }
673 742
674 __rq = cfq_find_rq_rb(cfqd, bio->bi_sector + bio_sectors(bio)); 743 __rq = cfq_find_rq_rb(cfqd, bio->bi_sector + bio_sectors(bio));
675 if (__rq) { 744 if (__rq && elv_rq_merge_ok(__rq, bio)) {
676 if (elv_rq_merge_ok(__rq, bio)) { 745 ret = ELEVATOR_FRONT_MERGE;
677 ret = ELEVATOR_FRONT_MERGE; 746 goto out;
678 goto out;
679 }
680 } 747 }
681 748
682 return ELEVATOR_NO_MERGE; 749 return ELEVATOR_NO_MERGE;
@@ -709,20 +776,220 @@ static void
709cfq_merged_requests(request_queue_t *q, struct request *rq, 776cfq_merged_requests(request_queue_t *q, struct request *rq,
710 struct request *next) 777 struct request *next)
711{ 778{
712 struct cfq_rq *crq = RQ_DATA(rq);
713 struct cfq_rq *cnext = RQ_DATA(next);
714
715 cfq_merged_request(q, rq); 779 cfq_merged_request(q, rq);
716 780
717 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist)) { 781 /*
718 if (time_before(cnext->queue_start, crq->queue_start)) { 782 * reposition in fifo if next is older than rq
719 list_move(&rq->queuelist, &next->queuelist); 783 */
720 crq->queue_start = cnext->queue_start; 784 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
785 time_before(next->start_time, rq->start_time))
786 list_move(&rq->queuelist, &next->queuelist);
787
788 cfq_remove_request(q, next);
789}
790
791static inline void
792__cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
793{
794 if (cfqq) {
795 /*
796 * stop potential idle class queues waiting service
797 */
798 del_timer(&cfqd->idle_class_timer);
799
800 cfqq->slice_start = jiffies;
801 cfqq->slice_end = 0;
802 cfqq->slice_left = 0;
803 cfq_clear_cfqq_must_alloc_slice(cfqq);
804 cfq_clear_cfqq_fifo_expire(cfqq);
805 cfq_clear_cfqq_expired(cfqq);
806 }
807
808 cfqd->active_queue = cfqq;
809}
810
811/*
812 * 0
813 * 0,1
814 * 0,1,2
815 * 0,1,2,3
816 * 0,1,2,3,4
817 * 0,1,2,3,4,5
818 * 0,1,2,3,4,5,6
819 * 0,1,2,3,4,5,6,7
820 */
821static int cfq_get_next_prio_level(struct cfq_data *cfqd)
822{
823 int prio, wrap;
824
825 prio = -1;
826 wrap = 0;
827 do {
828 int p;
829
830 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
831 if (!list_empty(&cfqd->rr_list[p])) {
832 prio = p;
833 break;
834 }
835 }
836
837 if (prio != -1)
838 break;
839 cfqd->cur_prio = 0;
840 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
841 cfqd->cur_end_prio = 0;
842 if (wrap)
843 break;
844 wrap = 1;
721 } 845 }
846 } while (1);
847
848 if (unlikely(prio == -1))
849 return -1;
850
851 BUG_ON(prio >= CFQ_PRIO_LISTS);
852
853 list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
854
855 cfqd->cur_prio = prio + 1;
856 if (cfqd->cur_prio > cfqd->cur_end_prio) {
857 cfqd->cur_end_prio = cfqd->cur_prio;
858 cfqd->cur_prio = 0;
859 }
860 if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
861 cfqd->cur_prio = 0;
862 cfqd->cur_end_prio = 0;
722 } 863 }
723 864
724 cfq_update_next_crq(cnext); 865 return prio;
725 cfq_remove_request(q, next); 866}
867
868static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
869{
870 struct cfq_queue *cfqq;
871
872 /*
873 * if current queue is expired but not done with its requests yet,
874 * wait for that to happen
875 */
876 if ((cfqq = cfqd->active_queue) != NULL) {
877 if (cfq_cfqq_expired(cfqq) && cfq_cfqq_dispatched(cfqq))
878 return NULL;
879 }
880
881 /*
882 * if current list is non-empty, grab first entry. if it is empty,
883 * get next prio level and grab first entry then if any are spliced
884 */
885 if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
886 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
887
888 /*
889 * if we have idle queues and no rt or be queues had pending
890 * requests, either allow immediate service if the grace period
891 * has passed or arm the idle grace timer
892 */
893 if (!cfqq && !list_empty(&cfqd->idle_rr)) {
894 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
895
896 if (time_after_eq(jiffies, end))
897 cfqq = list_entry_cfqq(cfqd->idle_rr.next);
898 else
899 mod_timer(&cfqd->idle_class_timer, end);
900 }
901
902 __cfq_set_active_queue(cfqd, cfqq);
903 return cfqq;
904}
905
906/*
907 * current cfqq expired its slice (or was too idle), select new one
908 */
909static void
910__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
911 int preempted)
912{
913 unsigned long now = jiffies;
914
915 if (cfq_cfqq_wait_request(cfqq))
916 del_timer(&cfqd->idle_slice_timer);
917
918 if (!preempted && !cfq_cfqq_dispatched(cfqq))
919 cfqq->service_last = now;
920
921 cfq_clear_cfqq_must_dispatch(cfqq);
922 cfq_clear_cfqq_wait_request(cfqq);
923
924 /*
925 * store what was left of this slice, if the queue idled out
926 * or was preempted
927 */
928 if (time_after(now, cfqq->slice_end))
929 cfqq->slice_left = now - cfqq->slice_end;
930 else
931 cfqq->slice_left = 0;
932
933 if (cfq_cfqq_on_rr(cfqq))
934 cfq_resort_rr_list(cfqq, preempted);
935
936 if (cfqq == cfqd->active_queue)
937 cfqd->active_queue = NULL;
938
939 if (cfqd->active_cic) {
940 put_io_context(cfqd->active_cic->ioc);
941 cfqd->active_cic = NULL;
942 }
943
944 cfqd->dispatch_slice = 0;
945}
946
947static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
948{
949 struct cfq_queue *cfqq = cfqd->active_queue;
950
951 if (cfqq) {
952 /*
953 * use deferred expiry, if there are requests in progress as
954 * not to disturb the slice of the next queue
955 */
956 if (cfq_cfqq_dispatched(cfqq))
957 cfq_mark_cfqq_expired(cfqq);
958 else
959 __cfq_slice_expired(cfqd, cfqq, preempted);
960 }
961}
962
963static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
964
965{
966 WARN_ON(!RB_EMPTY(&cfqq->sort_list));
967 WARN_ON(cfqq != cfqd->active_queue);
968
969 /*
970 * idle is disabled, either manually or by past process history
971 */
972 if (!cfqd->cfq_slice_idle)
973 return 0;
974 if (!cfq_cfqq_idle_window(cfqq))
975 return 0;
976 /*
977 * task has exited, don't wait
978 */
979 if (cfqd->active_cic && !cfqd->active_cic->ioc->task)
980 return 0;
981
982 cfq_mark_cfqq_must_dispatch(cfqq);
983 cfq_mark_cfqq_wait_request(cfqq);
984
985 if (!timer_pending(&cfqd->idle_slice_timer)) {
986 unsigned long slice_left = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
987
988 cfqd->idle_slice_timer.expires = jiffies + slice_left;
989 add_timer(&cfqd->idle_slice_timer);
990 }
991
992 return 1;
726} 993}
727 994
728/* 995/*
@@ -738,31 +1005,40 @@ static void cfq_dispatch_sort(request_queue_t *q, struct cfq_rq *crq)
738 struct request *__rq; 1005 struct request *__rq;
739 sector_t last; 1006 sector_t last;
740 1007
741 cfq_del_crq_rb(crq);
742 cfq_remove_merge_hints(q, crq);
743 list_del(&crq->request->queuelist); 1008 list_del(&crq->request->queuelist);
744 1009
745 last = cfqd->last_sector; 1010 last = cfqd->last_sector;
746 while ((entry = entry->prev) != head) { 1011 list_for_each_entry_reverse(__rq, head, queuelist) {
747 __rq = list_entry_rq(entry); 1012 struct cfq_rq *__crq = RQ_DATA(__rq);
748 1013
749 if (blk_barrier_rq(crq->request)) 1014 if (blk_barrier_rq(__rq))
750 break; 1015 break;
751 if (!blk_fs_request(crq->request)) 1016 if (!blk_fs_request(__rq))
1017 break;
1018 if (cfq_crq_requeued(__crq))
752 break; 1019 break;
753 1020
754 if (crq->request->sector > __rq->sector) 1021 if (__rq->sector <= crq->request->sector)
755 break; 1022 break;
756 if (__rq->sector > last && crq->request->sector < last) { 1023 if (__rq->sector > last && crq->request->sector < last) {
757 last = crq->request->sector; 1024 last = crq->request->sector + crq->request->nr_sectors;
758 break; 1025 break;
759 } 1026 }
1027 entry = &__rq->queuelist;
760 } 1028 }
761 1029
762 cfqd->last_sector = last; 1030 cfqd->last_sector = last;
763 crq->in_flight = 1; 1031
764 cfqq->in_flight++; 1032 cfqq->next_crq = cfq_find_next_crq(cfqd, cfqq, crq);
765 list_add(&crq->request->queuelist, entry); 1033
1034 cfq_del_crq_rb(crq);
1035 cfq_remove_merge_hints(q, crq);
1036
1037 cfq_mark_crq_in_flight(crq);
1038 cfq_clear_crq_requeued(crq);
1039
1040 cfqq->on_dispatch[cfq_crq_is_sync(crq)]++;
1041 list_add_tail(&crq->request->queuelist, entry);
766} 1042}
767 1043
768/* 1044/*
@@ -771,173 +1047,225 @@ static void cfq_dispatch_sort(request_queue_t *q, struct cfq_rq *crq)
771static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq) 1047static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
772{ 1048{
773 struct cfq_data *cfqd = cfqq->cfqd; 1049 struct cfq_data *cfqd = cfqq->cfqd;
774 const int reads = !list_empty(&cfqq->fifo[0]); 1050 struct request *rq;
775 const int writes = !list_empty(&cfqq->fifo[1]);
776 unsigned long now = jiffies;
777 struct cfq_rq *crq; 1051 struct cfq_rq *crq;
778 1052
779 if (time_before(now, cfqq->last_fifo_expire + cfqd->cfq_fifo_batch_expire)) 1053 if (cfq_cfqq_fifo_expire(cfqq))
780 return NULL; 1054 return NULL;
781 1055
782 crq = RQ_DATA(list_entry(cfqq->fifo[0].next, struct request, queuelist)); 1056 if (!list_empty(&cfqq->fifo)) {
783 if (reads && time_after(now, crq->queue_start + cfqd->cfq_fifo_expire_r)) { 1057 int fifo = cfq_cfqq_class_sync(cfqq);
784 cfqq->last_fifo_expire = now;
785 return crq;
786 }
787 1058
788 crq = RQ_DATA(list_entry(cfqq->fifo[1].next, struct request, queuelist)); 1059 crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next));
789 if (writes && time_after(now, crq->queue_start + cfqd->cfq_fifo_expire_w)) { 1060 rq = crq->request;
790 cfqq->last_fifo_expire = now; 1061 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
791 return crq; 1062 cfq_mark_cfqq_fifo_expire(cfqq);
1063 return crq;
1064 }
792 } 1065 }
793 1066
794 return NULL; 1067 return NULL;
795} 1068}
796 1069
797/* 1070/*
798 * dispatch a single request from given queue 1071 * Scale schedule slice based on io priority. Use the sync time slice only
1072 * if a queue is marked sync and has sync io queued. A sync queue with async
1073 * io only, should not get full sync slice length.
799 */ 1074 */
1075static inline int
1076cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1077{
1078 const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
1079
1080 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1081
1082 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
1083}
1084
800static inline void 1085static inline void
801cfq_dispatch_request(request_queue_t *q, struct cfq_data *cfqd, 1086cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
802 struct cfq_queue *cfqq)
803{ 1087{
804 struct cfq_rq *crq; 1088 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
1089}
1090
1091static inline int
1092cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1093{
1094 const int base_rq = cfqd->cfq_slice_async_rq;
1095
1096 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1097
1098 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1099}
1100
1101/*
1102 * get next queue for service
1103 */
1104static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd, int force)
1105{
1106 unsigned long now = jiffies;
1107 struct cfq_queue *cfqq;
1108
1109 cfqq = cfqd->active_queue;
1110 if (!cfqq)
1111 goto new_queue;
1112
1113 if (cfq_cfqq_expired(cfqq))
1114 goto new_queue;
805 1115
806 /* 1116 /*
807 * follow expired path, else get first next available 1117 * slice has expired
808 */ 1118 */
809 if ((crq = cfq_check_fifo(cfqq)) == NULL) { 1119 if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
810 if (cfqd->find_best_crq) 1120 goto expire;
811 crq = cfqq->next_crq;
812 else
813 crq = rb_entry_crq(rb_first(&cfqq->sort_list));
814 }
815
816 cfqd->last_sector = crq->request->sector + crq->request->nr_sectors;
817 1121
818 /* 1122 /*
819 * finally, insert request into driver list 1123 * if queue has requests, dispatch one. if not, check if
1124 * enough slice is left to wait for one
820 */ 1125 */
821 cfq_dispatch_sort(q, crq); 1126 if (!RB_EMPTY(&cfqq->sort_list))
1127 goto keep_queue;
1128 else if (!force && cfq_cfqq_class_sync(cfqq) &&
1129 time_before(now, cfqq->slice_end)) {
1130 if (cfq_arm_slice_timer(cfqd, cfqq))
1131 return NULL;
1132 }
1133
1134expire:
1135 cfq_slice_expired(cfqd, 0);
1136new_queue:
1137 cfqq = cfq_set_active_queue(cfqd);
1138keep_queue:
1139 return cfqq;
822} 1140}
823 1141
824static int cfq_dispatch_requests(request_queue_t *q, int max_dispatch) 1142static int
1143__cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1144 int max_dispatch)
825{ 1145{
826 struct cfq_data *cfqd = q->elevator->elevator_data; 1146 int dispatched = 0;
827 struct cfq_queue *cfqq;
828 struct list_head *entry, *tmp;
829 int queued, busy_queues, first_round;
830 1147
831 if (list_empty(&cfqd->rr_list)) 1148 BUG_ON(RB_EMPTY(&cfqq->sort_list));
832 return 0;
833 1149
834 queued = 0; 1150 do {
835 first_round = 1; 1151 struct cfq_rq *crq;
836restart:
837 busy_queues = 0;
838 list_for_each_safe(entry, tmp, &cfqd->rr_list) {
839 cfqq = list_entry_cfqq(entry);
840 1152
841 BUG_ON(RB_EMPTY(&cfqq->sort_list)); 1153 /*
1154 * follow expired path, else get first next available
1155 */
1156 if ((crq = cfq_check_fifo(cfqq)) == NULL)
1157 crq = cfqq->next_crq;
842 1158
843 /* 1159 /*
844 * first round of queueing, only select from queues that 1160 * finally, insert request into driver dispatch list
845 * don't already have io in-flight
846 */ 1161 */
847 if (first_round && cfqq->in_flight) 1162 cfq_dispatch_sort(cfqd->queue, crq);
848 continue;
849 1163
850 cfq_dispatch_request(q, cfqd, cfqq); 1164 cfqd->dispatch_slice++;
1165 dispatched++;
851 1166
852 if (!RB_EMPTY(&cfqq->sort_list)) 1167 if (!cfqd->active_cic) {
853 busy_queues++; 1168 atomic_inc(&crq->io_context->ioc->refcount);
1169 cfqd->active_cic = crq->io_context;
1170 }
854 1171
855 queued++; 1172 if (RB_EMPTY(&cfqq->sort_list))
856 } 1173 break;
1174
1175 } while (dispatched < max_dispatch);
1176
1177 /*
1178 * if slice end isn't set yet, set it. if at least one request was
1179 * sync, use the sync time slice value
1180 */
1181 if (!cfqq->slice_end)
1182 cfq_set_prio_slice(cfqd, cfqq);
1183
1184 /*
1185 * expire an async queue immediately if it has used up its slice. idle
1186 * queue always expire after 1 dispatch round.
1187 */
1188 if ((!cfq_cfqq_sync(cfqq) &&
1189 cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1190 cfq_class_idle(cfqq))
1191 cfq_slice_expired(cfqd, 0);
1192
1193 return dispatched;
1194}
1195
1196static int
1197cfq_dispatch_requests(request_queue_t *q, int max_dispatch, int force)
1198{
1199 struct cfq_data *cfqd = q->elevator->elevator_data;
1200 struct cfq_queue *cfqq;
1201
1202 if (!cfqd->busy_queues)
1203 return 0;
1204
1205 cfqq = cfq_select_queue(cfqd, force);
1206 if (cfqq) {
1207 cfq_clear_cfqq_must_dispatch(cfqq);
1208 cfq_clear_cfqq_wait_request(cfqq);
1209 del_timer(&cfqd->idle_slice_timer);
857 1210
858 if ((queued < max_dispatch) && (busy_queues || first_round)) { 1211 if (cfq_class_idle(cfqq))
859 first_round = 0; 1212 max_dispatch = 1;
860 goto restart; 1213
1214 return __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
861 } 1215 }
862 1216
863 return queued; 1217 return 0;
864} 1218}
865 1219
866static inline void cfq_account_dispatch(struct cfq_rq *crq) 1220static inline void cfq_account_dispatch(struct cfq_rq *crq)
867{ 1221{
868 struct cfq_queue *cfqq = crq->cfq_queue; 1222 struct cfq_queue *cfqq = crq->cfq_queue;
869 struct cfq_data *cfqd = cfqq->cfqd; 1223 struct cfq_data *cfqd = cfqq->cfqd;
870 unsigned long now, elapsed;
871 1224
872 if (!blk_fs_request(crq->request)) 1225 if (unlikely(!blk_fs_request(crq->request)))
873 return; 1226 return;
874 1227
875 /* 1228 /*
876 * accounted bit is necessary since some drivers will call 1229 * accounted bit is necessary since some drivers will call
877 * elv_next_request() many times for the same request (eg ide) 1230 * elv_next_request() many times for the same request (eg ide)
878 */ 1231 */
879 if (crq->accounted) 1232 if (cfq_crq_in_driver(crq))
880 return; 1233 return;
881 1234
882 now = jiffies; 1235 cfq_mark_crq_in_driver(crq);
883 if (cfqq->service_start == ~0UL) 1236 cfqd->rq_in_driver++;
884 cfqq->service_start = now;
885
886 /*
887 * on drives with tagged command queueing, command turn-around time
888 * doesn't necessarily reflect the time spent processing this very
889 * command inside the drive. so do the accounting differently there,
890 * by just sorting on the number of requests
891 */
892 if (cfqd->cfq_tagged) {
893 if (time_after(now, cfqq->service_start + cfq_service)) {
894 cfqq->service_start = now;
895 cfqq->service_used /= 10;
896 }
897
898 cfqq->service_used++;
899 cfq_sort_rr_list(cfqq, 0);
900 }
901
902 elapsed = now - crq->queue_start;
903 if (elapsed > max_elapsed_dispatch)
904 max_elapsed_dispatch = elapsed;
905
906 crq->accounted = 1;
907 crq->service_start = now;
908
909 if (++cfqd->rq_in_driver >= CFQ_MAX_TAG && !cfqd->cfq_tagged) {
910 cfqq->cfqd->cfq_tagged = 1;
911 printk("cfq: depth %d reached, tagging now on\n", CFQ_MAX_TAG);
912 }
913} 1237}
914 1238
915static inline void 1239static inline void
916cfq_account_completion(struct cfq_queue *cfqq, struct cfq_rq *crq) 1240cfq_account_completion(struct cfq_queue *cfqq, struct cfq_rq *crq)
917{ 1241{
918 struct cfq_data *cfqd = cfqq->cfqd; 1242 struct cfq_data *cfqd = cfqq->cfqd;
1243 unsigned long now;
919 1244
920 if (!crq->accounted) 1245 if (!cfq_crq_in_driver(crq))
921 return; 1246 return;
922 1247
1248 now = jiffies;
1249
923 WARN_ON(!cfqd->rq_in_driver); 1250 WARN_ON(!cfqd->rq_in_driver);
924 cfqd->rq_in_driver--; 1251 cfqd->rq_in_driver--;
925 1252
926 if (!cfqd->cfq_tagged) { 1253 if (!cfq_class_idle(cfqq))
927 unsigned long now = jiffies; 1254 cfqd->last_end_request = now;
928 unsigned long duration = now - crq->service_start;
929 1255
930 if (time_after(now, cfqq->service_start + cfq_service)) { 1256 if (!cfq_cfqq_dispatched(cfqq)) {
931 cfqq->service_start = now; 1257 if (cfq_cfqq_on_rr(cfqq)) {
932 cfqq->service_used >>= 3; 1258 cfqq->service_last = now;
1259 cfq_resort_rr_list(cfqq, 0);
1260 }
1261 if (cfq_cfqq_expired(cfqq)) {
1262 __cfq_slice_expired(cfqd, cfqq, 0);
1263 cfq_schedule_dispatch(cfqd);
933 } 1264 }
934
935 cfqq->service_used += duration;
936 cfq_sort_rr_list(cfqq, 0);
937
938 if (duration > max_elapsed_crq)
939 max_elapsed_crq = duration;
940 } 1265 }
1266
1267 if (cfq_crq_is_sync(crq))
1268 crq->io_context->last_end_request = now;
941} 1269}
942 1270
943static struct request *cfq_next_request(request_queue_t *q) 1271static struct request *cfq_next_request(request_queue_t *q)
@@ -950,7 +1278,19 @@ static struct request *cfq_next_request(request_queue_t *q)
950dispatch: 1278dispatch:
951 rq = list_entry_rq(q->queue_head.next); 1279 rq = list_entry_rq(q->queue_head.next);
952 1280
953 if ((crq = RQ_DATA(rq)) != NULL) { 1281 crq = RQ_DATA(rq);
1282 if (crq) {
1283 struct cfq_queue *cfqq = crq->cfq_queue;
1284
1285 /*
1286 * if idle window is disabled, allow queue buildup
1287 */
1288 if (!cfq_crq_in_driver(crq) &&
1289 !cfq_cfqq_idle_window(cfqq) &&
1290 !blk_barrier_rq(rq) &&
1291 cfqd->rq_in_driver >= cfqd->cfq_max_depth)
1292 return NULL;
1293
954 cfq_remove_merge_hints(q, crq); 1294 cfq_remove_merge_hints(q, crq);
955 cfq_account_dispatch(crq); 1295 cfq_account_dispatch(crq);
956 } 1296 }
@@ -958,7 +1298,7 @@ dispatch:
958 return rq; 1298 return rq;
959 } 1299 }
960 1300
961 if (cfq_dispatch_requests(q, cfqd->cfq_quantum)) 1301 if (cfq_dispatch_requests(q, cfqd->cfq_quantum, 0))
962 goto dispatch; 1302 goto dispatch;
963 1303
964 return NULL; 1304 return NULL;
@@ -972,13 +1312,21 @@ dispatch:
972 */ 1312 */
973static void cfq_put_queue(struct cfq_queue *cfqq) 1313static void cfq_put_queue(struct cfq_queue *cfqq)
974{ 1314{
975 BUG_ON(!atomic_read(&cfqq->ref)); 1315 struct cfq_data *cfqd = cfqq->cfqd;
1316
1317 BUG_ON(atomic_read(&cfqq->ref) <= 0);
976 1318
977 if (!atomic_dec_and_test(&cfqq->ref)) 1319 if (!atomic_dec_and_test(&cfqq->ref))
978 return; 1320 return;
979 1321
980 BUG_ON(rb_first(&cfqq->sort_list)); 1322 BUG_ON(rb_first(&cfqq->sort_list));
981 BUG_ON(cfqq->on_rr); 1323 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1324 BUG_ON(cfq_cfqq_on_rr(cfqq));
1325
1326 if (unlikely(cfqd->active_queue == cfqq)) {
1327 __cfq_slice_expired(cfqd, cfqq, 0);
1328 cfq_schedule_dispatch(cfqd);
1329 }
982 1330
983 cfq_put_cfqd(cfqq->cfqd); 1331 cfq_put_cfqd(cfqq->cfqd);
984 1332
@@ -991,15 +1339,17 @@ static void cfq_put_queue(struct cfq_queue *cfqq)
991} 1339}
992 1340
993static inline struct cfq_queue * 1341static inline struct cfq_queue *
994__cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned long key, const int hashval) 1342__cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1343 const int hashval)
995{ 1344{
996 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval]; 1345 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
997 struct hlist_node *entry, *next; 1346 struct hlist_node *entry, *next;
998 1347
999 hlist_for_each_safe(entry, next, hash_list) { 1348 hlist_for_each_safe(entry, next, hash_list) {
1000 struct cfq_queue *__cfqq = list_entry_qhash(entry); 1349 struct cfq_queue *__cfqq = list_entry_qhash(entry);
1350 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->ioprio_class, __cfqq->ioprio);
1001 1351
1002 if (__cfqq->key == key) 1352 if (__cfqq->key == key && (__p == prio || prio == CFQ_KEY_ANY))
1003 return __cfqq; 1353 return __cfqq;
1004 } 1354 }
1005 1355
@@ -1007,94 +1357,220 @@ __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned long key, const int hashval)
1007} 1357}
1008 1358
1009static struct cfq_queue * 1359static struct cfq_queue *
1010cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned long key) 1360cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1011{ 1361{
1012 return __cfq_find_cfq_hash(cfqd, key, hash_long(key, CFQ_QHASH_SHIFT)); 1362 return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1013} 1363}
1014 1364
1015static inline void 1365static void cfq_free_io_context(struct cfq_io_context *cic)
1016cfq_rehash_cfqq(struct cfq_data *cfqd, struct cfq_queue **cfqq,
1017 struct cfq_io_context *cic)
1018{ 1366{
1019 unsigned long hashkey = cfq_hash_key(cfqd, current); 1367 struct cfq_io_context *__cic;
1020 unsigned long hashval = hash_long(hashkey, CFQ_QHASH_SHIFT); 1368 struct list_head *entry, *next;
1021 struct cfq_queue *__cfqq;
1022 unsigned long flags;
1023
1024 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1025 1369
1026 hlist_del(&(*cfqq)->cfq_hash); 1370 list_for_each_safe(entry, next, &cic->list) {
1027 1371 __cic = list_entry(entry, struct cfq_io_context, list);
1028 __cfqq = __cfq_find_cfq_hash(cfqd, hashkey, hashval); 1372 kmem_cache_free(cfq_ioc_pool, __cic);
1029 if (!__cfqq || __cfqq == *cfqq) {
1030 __cfqq = *cfqq;
1031 hlist_add_head(&__cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1032 __cfqq->key_type = cfqd->key_type;
1033 } else {
1034 atomic_inc(&__cfqq->ref);
1035 cic->cfqq = __cfqq;
1036 cfq_put_queue(*cfqq);
1037 *cfqq = __cfqq;
1038 } 1373 }
1039 1374
1040 cic->cfqq = __cfqq; 1375 kmem_cache_free(cfq_ioc_pool, cic);
1041 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1042} 1376}
1043 1377
1044static void cfq_free_io_context(struct cfq_io_context *cic) 1378/*
1379 * Called with interrupts disabled
1380 */
1381static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1045{ 1382{
1046 kmem_cache_free(cfq_ioc_pool, cic); 1383 struct cfq_data *cfqd = cic->cfqq->cfqd;
1384 request_queue_t *q = cfqd->queue;
1385
1386 WARN_ON(!irqs_disabled());
1387
1388 spin_lock(q->queue_lock);
1389
1390 if (unlikely(cic->cfqq == cfqd->active_queue)) {
1391 __cfq_slice_expired(cfqd, cic->cfqq, 0);
1392 cfq_schedule_dispatch(cfqd);
1393 }
1394
1395 cfq_put_queue(cic->cfqq);
1396 cic->cfqq = NULL;
1397 spin_unlock(q->queue_lock);
1047} 1398}
1048 1399
1049/* 1400/*
1050 * locking hierarchy is: io_context lock -> queue locks 1401 * Another task may update the task cic list, if it is doing a queue lookup
1402 * on its behalf. cfq_cic_lock excludes such concurrent updates
1051 */ 1403 */
1052static void cfq_exit_io_context(struct cfq_io_context *cic) 1404static void cfq_exit_io_context(struct cfq_io_context *cic)
1053{ 1405{
1054 struct cfq_queue *cfqq = cic->cfqq; 1406 struct cfq_io_context *__cic;
1055 struct list_head *entry = &cic->list; 1407 struct list_head *entry;
1056 request_queue_t *q;
1057 unsigned long flags; 1408 unsigned long flags;
1058 1409
1410 local_irq_save(flags);
1411
1059 /* 1412 /*
1060 * put the reference this task is holding to the various queues 1413 * put the reference this task is holding to the various queues
1061 */ 1414 */
1062 spin_lock_irqsave(&cic->ioc->lock, flags); 1415 list_for_each(entry, &cic->list) {
1063 while ((entry = cic->list.next) != &cic->list) {
1064 struct cfq_io_context *__cic;
1065
1066 __cic = list_entry(entry, struct cfq_io_context, list); 1416 __cic = list_entry(entry, struct cfq_io_context, list);
1067 list_del(entry); 1417 cfq_exit_single_io_context(__cic);
1068
1069 q = __cic->cfqq->cfqd->queue;
1070 spin_lock(q->queue_lock);
1071 cfq_put_queue(__cic->cfqq);
1072 spin_unlock(q->queue_lock);
1073 } 1418 }
1074 1419
1075 q = cfqq->cfqd->queue; 1420 cfq_exit_single_io_context(cic);
1076 spin_lock(q->queue_lock); 1421 local_irq_restore(flags);
1077 cfq_put_queue(cfqq);
1078 spin_unlock(q->queue_lock);
1079
1080 cic->cfqq = NULL;
1081 spin_unlock_irqrestore(&cic->ioc->lock, flags);
1082} 1422}
1083 1423
1084static struct cfq_io_context *cfq_alloc_io_context(int gfp_flags) 1424static struct cfq_io_context *
1425cfq_alloc_io_context(struct cfq_data *cfqd, int gfp_mask)
1085{ 1426{
1086 struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_flags); 1427 struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1087 1428
1088 if (cic) { 1429 if (cic) {
1089 cic->dtor = cfq_free_io_context;
1090 cic->exit = cfq_exit_io_context;
1091 INIT_LIST_HEAD(&cic->list); 1430 INIT_LIST_HEAD(&cic->list);
1092 cic->cfqq = NULL; 1431 cic->cfqq = NULL;
1432 cic->key = NULL;
1433 cic->last_end_request = jiffies;
1434 cic->ttime_total = 0;
1435 cic->ttime_samples = 0;
1436 cic->ttime_mean = 0;
1437 cic->dtor = cfq_free_io_context;
1438 cic->exit = cfq_exit_io_context;
1093 } 1439 }
1094 1440
1095 return cic; 1441 return cic;
1096} 1442}
1097 1443
1444static void cfq_init_prio_data(struct cfq_queue *cfqq)
1445{
1446 struct task_struct *tsk = current;
1447 int ioprio_class;
1448
1449 if (!cfq_cfqq_prio_changed(cfqq))
1450 return;
1451
1452 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1453 switch (ioprio_class) {
1454 default:
1455 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1456 case IOPRIO_CLASS_NONE:
1457 /*
1458 * no prio set, place us in the middle of the BE classes
1459 */
1460 cfqq->ioprio = task_nice_ioprio(tsk);
1461 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1462 break;
1463 case IOPRIO_CLASS_RT:
1464 cfqq->ioprio = task_ioprio(tsk);
1465 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1466 break;
1467 case IOPRIO_CLASS_BE:
1468 cfqq->ioprio = task_ioprio(tsk);
1469 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1470 break;
1471 case IOPRIO_CLASS_IDLE:
1472 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1473 cfqq->ioprio = 7;
1474 cfq_clear_cfqq_idle_window(cfqq);
1475 break;
1476 }
1477
1478 /*
1479 * keep track of original prio settings in case we have to temporarily
1480 * elevate the priority of this queue
1481 */
1482 cfqq->org_ioprio = cfqq->ioprio;
1483 cfqq->org_ioprio_class = cfqq->ioprio_class;
1484
1485 if (cfq_cfqq_on_rr(cfqq))
1486 cfq_resort_rr_list(cfqq, 0);
1487
1488 cfq_clear_cfqq_prio_changed(cfqq);
1489}
1490
1491static inline void changed_ioprio(struct cfq_queue *cfqq)
1492{
1493 if (cfqq) {
1494 struct cfq_data *cfqd = cfqq->cfqd;
1495
1496 spin_lock(cfqd->queue->queue_lock);
1497 cfq_mark_cfqq_prio_changed(cfqq);
1498 cfq_init_prio_data(cfqq);
1499 spin_unlock(cfqd->queue->queue_lock);
1500 }
1501}
1502
1503/*
1504 * callback from sys_ioprio_set, irqs are disabled
1505 */
1506static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
1507{
1508 struct cfq_io_context *cic = ioc->cic;
1509
1510 changed_ioprio(cic->cfqq);
1511
1512 list_for_each_entry(cic, &cic->list, list)
1513 changed_ioprio(cic->cfqq);
1514
1515 return 0;
1516}
1517
1518static struct cfq_queue *
1519cfq_get_queue(struct cfq_data *cfqd, unsigned int key, unsigned short ioprio,
1520 int gfp_mask)
1521{
1522 const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1523 struct cfq_queue *cfqq, *new_cfqq = NULL;
1524
1525retry:
1526 cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1527
1528 if (!cfqq) {
1529 if (new_cfqq) {
1530 cfqq = new_cfqq;
1531 new_cfqq = NULL;
1532 } else if (gfp_mask & __GFP_WAIT) {
1533 spin_unlock_irq(cfqd->queue->queue_lock);
1534 new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1535 spin_lock_irq(cfqd->queue->queue_lock);
1536 goto retry;
1537 } else {
1538 cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1539 if (!cfqq)
1540 goto out;
1541 }
1542
1543 memset(cfqq, 0, sizeof(*cfqq));
1544
1545 INIT_HLIST_NODE(&cfqq->cfq_hash);
1546 INIT_LIST_HEAD(&cfqq->cfq_list);
1547 RB_CLEAR_ROOT(&cfqq->sort_list);
1548 INIT_LIST_HEAD(&cfqq->fifo);
1549
1550 cfqq->key = key;
1551 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1552 atomic_set(&cfqq->ref, 0);
1553 cfqq->cfqd = cfqd;
1554 atomic_inc(&cfqd->ref);
1555 cfqq->service_last = 0;
1556 /*
1557 * set ->slice_left to allow preemption for a new process
1558 */
1559 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1560 cfq_mark_cfqq_idle_window(cfqq);
1561 cfq_mark_cfqq_prio_changed(cfqq);
1562 cfq_init_prio_data(cfqq);
1563 }
1564
1565 if (new_cfqq)
1566 kmem_cache_free(cfq_pool, new_cfqq);
1567
1568 atomic_inc(&cfqq->ref);
1569out:
1570 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1571 return cfqq;
1572}
1573
1098/* 1574/*
1099 * Setup general io context and cfq io context. There can be several cfq 1575 * Setup general io context and cfq io context. There can be several cfq
1100 * io contexts per general io context, if this process is doing io to more 1576 * io contexts per general io context, if this process is doing io to more
@@ -1102,39 +1578,39 @@ static struct cfq_io_context *cfq_alloc_io_context(int gfp_flags)
1102 * cfqq, so we don't need to worry about it disappearing 1578 * cfqq, so we don't need to worry about it disappearing
1103 */ 1579 */
1104static struct cfq_io_context * 1580static struct cfq_io_context *
1105cfq_get_io_context(struct cfq_queue **cfqq, int gfp_flags) 1581cfq_get_io_context(struct cfq_data *cfqd, pid_t pid, int gfp_mask)
1106{ 1582{
1107 struct cfq_data *cfqd = (*cfqq)->cfqd; 1583 struct io_context *ioc = NULL;
1108 struct cfq_queue *__cfqq = *cfqq;
1109 struct cfq_io_context *cic; 1584 struct cfq_io_context *cic;
1110 struct io_context *ioc;
1111 1585
1112 might_sleep_if(gfp_flags & __GFP_WAIT); 1586 might_sleep_if(gfp_mask & __GFP_WAIT);
1113 1587
1114 ioc = get_io_context(gfp_flags); 1588 ioc = get_io_context(gfp_mask);
1115 if (!ioc) 1589 if (!ioc)
1116 return NULL; 1590 return NULL;
1117 1591
1118 if ((cic = ioc->cic) == NULL) { 1592 if ((cic = ioc->cic) == NULL) {
1119 cic = cfq_alloc_io_context(gfp_flags); 1593 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1120 1594
1121 if (cic == NULL) 1595 if (cic == NULL)
1122 goto err; 1596 goto err;
1123 1597
1598 /*
1599 * manually increment generic io_context usage count, it
1600 * cannot go away since we are already holding one ref to it
1601 */
1124 ioc->cic = cic; 1602 ioc->cic = cic;
1603 ioc->set_ioprio = cfq_ioc_set_ioprio;
1125 cic->ioc = ioc; 1604 cic->ioc = ioc;
1126 cic->cfqq = __cfqq; 1605 cic->key = cfqd;
1127 atomic_inc(&__cfqq->ref); 1606 atomic_inc(&cfqd->ref);
1128 } else { 1607 } else {
1129 struct cfq_io_context *__cic; 1608 struct cfq_io_context *__cic;
1130 unsigned long flags;
1131 1609
1132 /* 1610 /*
1133 * since the first cic on the list is actually the head 1611 * the first cic on the list is actually the head itself
1134 * itself, need to check this here or we'll duplicate an
1135 * cic per ioc for no reason
1136 */ 1612 */
1137 if (cic->cfqq == __cfqq) 1613 if (cic->key == cfqd)
1138 goto out; 1614 goto out;
1139 1615
1140 /* 1616 /*
@@ -1142,152 +1618,255 @@ cfq_get_io_context(struct cfq_queue **cfqq, int gfp_flags)
1142 * should be ok here, the list will usually not be more than 1618 * should be ok here, the list will usually not be more than
1143 * 1 or a few entries long 1619 * 1 or a few entries long
1144 */ 1620 */
1145 spin_lock_irqsave(&ioc->lock, flags);
1146 list_for_each_entry(__cic, &cic->list, list) { 1621 list_for_each_entry(__cic, &cic->list, list) {
1147 /* 1622 /*
1148 * this process is already holding a reference to 1623 * this process is already holding a reference to
1149 * this queue, so no need to get one more 1624 * this queue, so no need to get one more
1150 */ 1625 */
1151 if (__cic->cfqq == __cfqq) { 1626 if (__cic->key == cfqd) {
1152 cic = __cic; 1627 cic = __cic;
1153 spin_unlock_irqrestore(&ioc->lock, flags);
1154 goto out; 1628 goto out;
1155 } 1629 }
1156 } 1630 }
1157 spin_unlock_irqrestore(&ioc->lock, flags);
1158 1631
1159 /* 1632 /*
1160 * nope, process doesn't have a cic assoicated with this 1633 * nope, process doesn't have a cic assoicated with this
1161 * cfqq yet. get a new one and add to list 1634 * cfqq yet. get a new one and add to list
1162 */ 1635 */
1163 __cic = cfq_alloc_io_context(gfp_flags); 1636 __cic = cfq_alloc_io_context(cfqd, gfp_mask);
1164 if (__cic == NULL) 1637 if (__cic == NULL)
1165 goto err; 1638 goto err;
1166 1639
1167 __cic->ioc = ioc; 1640 __cic->ioc = ioc;
1168 __cic->cfqq = __cfqq; 1641 __cic->key = cfqd;
1169 atomic_inc(&__cfqq->ref); 1642 atomic_inc(&cfqd->ref);
1170 spin_lock_irqsave(&ioc->lock, flags);
1171 list_add(&__cic->list, &cic->list); 1643 list_add(&__cic->list, &cic->list);
1172 spin_unlock_irqrestore(&ioc->lock, flags);
1173
1174 cic = __cic; 1644 cic = __cic;
1175 *cfqq = __cfqq;
1176 } 1645 }
1177 1646
1178out: 1647out:
1179 /*
1180 * if key_type has been changed on the fly, we lazily rehash
1181 * each queue at lookup time
1182 */
1183 if ((*cfqq)->key_type != cfqd->key_type)
1184 cfq_rehash_cfqq(cfqd, cfqq, cic);
1185
1186 return cic; 1648 return cic;
1187err: 1649err:
1188 put_io_context(ioc); 1650 put_io_context(ioc);
1189 return NULL; 1651 return NULL;
1190} 1652}
1191 1653
1192static struct cfq_queue * 1654static void
1193__cfq_get_queue(struct cfq_data *cfqd, unsigned long key, int gfp_mask) 1655cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1194{ 1656{
1195 const int hashval = hash_long(key, CFQ_QHASH_SHIFT); 1657 unsigned long elapsed, ttime;
1196 struct cfq_queue *cfqq, *new_cfqq = NULL;
1197
1198retry:
1199 cfqq = __cfq_find_cfq_hash(cfqd, key, hashval);
1200 1658
1201 if (!cfqq) { 1659 /*
1202 if (new_cfqq) { 1660 * if this context already has stuff queued, thinktime is from
1203 cfqq = new_cfqq; 1661 * last queue not last end
1204 new_cfqq = NULL; 1662 */
1205 } else { 1663#if 0
1206 spin_unlock_irq(cfqd->queue->queue_lock); 1664 if (time_after(cic->last_end_request, cic->last_queue))
1207 new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask); 1665 elapsed = jiffies - cic->last_end_request;
1208 spin_lock_irq(cfqd->queue->queue_lock); 1666 else
1667 elapsed = jiffies - cic->last_queue;
1668#else
1669 elapsed = jiffies - cic->last_end_request;
1670#endif
1209 1671
1210 if (!new_cfqq && !(gfp_mask & __GFP_WAIT)) 1672 ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1211 goto out;
1212 1673
1213 goto retry; 1674 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1214 } 1675 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1676 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1677}
1215 1678
1216 memset(cfqq, 0, sizeof(*cfqq)); 1679#define sample_valid(samples) ((samples) > 80)
1217 1680
1218 INIT_HLIST_NODE(&cfqq->cfq_hash); 1681/*
1219 INIT_LIST_HEAD(&cfqq->cfq_list); 1682 * Disable idle window if the process thinks too long or seeks so much that
1220 RB_CLEAR_ROOT(&cfqq->sort_list); 1683 * it doesn't matter
1221 INIT_LIST_HEAD(&cfqq->fifo[0]); 1684 */
1222 INIT_LIST_HEAD(&cfqq->fifo[1]); 1685static void
1686cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1687 struct cfq_io_context *cic)
1688{
1689 int enable_idle = cfq_cfqq_idle_window(cfqq);
1223 1690
1224 cfqq->key = key; 1691 if (!cic->ioc->task || !cfqd->cfq_slice_idle)
1225 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]); 1692 enable_idle = 0;
1226 atomic_set(&cfqq->ref, 0); 1693 else if (sample_valid(cic->ttime_samples)) {
1227 cfqq->cfqd = cfqd; 1694 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1228 atomic_inc(&cfqd->ref); 1695 enable_idle = 0;
1229 cfqq->key_type = cfqd->key_type; 1696 else
1230 cfqq->service_start = ~0UL; 1697 enable_idle = 1;
1231 } 1698 }
1232 1699
1233 if (new_cfqq) 1700 if (enable_idle)
1234 kmem_cache_free(cfq_pool, new_cfqq); 1701 cfq_mark_cfqq_idle_window(cfqq);
1702 else
1703 cfq_clear_cfqq_idle_window(cfqq);
1704}
1235 1705
1236 atomic_inc(&cfqq->ref); 1706
1237out: 1707/*
1238 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq); 1708 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1239 return cfqq; 1709 * no or if we aren't sure, a 1 will cause a preempt.
1710 */
1711static int
1712cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1713 struct cfq_rq *crq)
1714{
1715 struct cfq_queue *cfqq = cfqd->active_queue;
1716
1717 if (cfq_class_idle(new_cfqq))
1718 return 0;
1719
1720 if (!cfqq)
1721 return 1;
1722
1723 if (cfq_class_idle(cfqq))
1724 return 1;
1725 if (!cfq_cfqq_wait_request(new_cfqq))
1726 return 0;
1727 /*
1728 * if it doesn't have slice left, forget it
1729 */
1730 if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1731 return 0;
1732 if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq))
1733 return 1;
1734
1735 return 0;
1736}
1737
1738/*
1739 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1740 * let it have half of its nominal slice.
1741 */
1742static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1743{
1744 struct cfq_queue *__cfqq, *next;
1745
1746 list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1747 cfq_resort_rr_list(__cfqq, 1);
1748
1749 if (!cfqq->slice_left)
1750 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1751
1752 cfqq->slice_end = cfqq->slice_left + jiffies;
1753 __cfq_slice_expired(cfqd, cfqq, 1);
1754 __cfq_set_active_queue(cfqd, cfqq);
1755}
1756
1757/*
1758 * should really be a ll_rw_blk.c helper
1759 */
1760static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1761{
1762 request_queue_t *q = cfqd->queue;
1763
1764 if (!blk_queue_plugged(q))
1765 q->request_fn(q);
1766 else
1767 __generic_unplug_device(q);
1768}
1769
1770/*
1771 * Called when a new fs request (crq) is added (to cfqq). Check if there's
1772 * something we should do about it
1773 */
1774static void
1775cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1776 struct cfq_rq *crq)
1777{
1778 struct cfq_io_context *cic;
1779
1780 cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
1781
1782 /*
1783 * we never wait for an async request and we don't allow preemption
1784 * of an async request. so just return early
1785 */
1786 if (!cfq_crq_is_sync(crq))
1787 return;
1788
1789 cic = crq->io_context;
1790
1791 cfq_update_io_thinktime(cfqd, cic);
1792 cfq_update_idle_window(cfqd, cfqq, cic);
1793
1794 cic->last_queue = jiffies;
1795
1796 if (cfqq == cfqd->active_queue) {
1797 /*
1798 * if we are waiting for a request for this queue, let it rip
1799 * immediately and flag that we must not expire this queue
1800 * just now
1801 */
1802 if (cfq_cfqq_wait_request(cfqq)) {
1803 cfq_mark_cfqq_must_dispatch(cfqq);
1804 del_timer(&cfqd->idle_slice_timer);
1805 cfq_start_queueing(cfqd, cfqq);
1806 }
1807 } else if (cfq_should_preempt(cfqd, cfqq, crq)) {
1808 /*
1809 * not the active queue - expire current slice if it is
1810 * idle and has expired it's mean thinktime or this new queue
1811 * has some old slice time left and is of higher priority
1812 */
1813 cfq_preempt_queue(cfqd, cfqq);
1814 cfq_mark_cfqq_must_dispatch(cfqq);
1815 cfq_start_queueing(cfqd, cfqq);
1816 }
1240} 1817}
1241 1818
1242static void cfq_enqueue(struct cfq_data *cfqd, struct cfq_rq *crq) 1819static void cfq_enqueue(struct cfq_data *cfqd, struct request *rq)
1243{ 1820{
1244 crq->is_sync = 0; 1821 struct cfq_rq *crq = RQ_DATA(rq);
1245 if (rq_data_dir(crq->request) == READ || current->flags & PF_SYNCWRITE) 1822 struct cfq_queue *cfqq = crq->cfq_queue;
1246 crq->is_sync = 1; 1823
1824 cfq_init_prio_data(cfqq);
1247 1825
1248 cfq_add_crq_rb(crq); 1826 cfq_add_crq_rb(crq);
1249 crq->queue_start = jiffies;
1250 1827
1251 list_add_tail(&crq->request->queuelist, &crq->cfq_queue->fifo[crq->is_sync]); 1828 list_add_tail(&rq->queuelist, &cfqq->fifo);
1829
1830 if (rq_mergeable(rq)) {
1831 cfq_add_crq_hash(cfqd, crq);
1832
1833 if (!cfqd->queue->last_merge)
1834 cfqd->queue->last_merge = rq;
1835 }
1836
1837 cfq_crq_enqueued(cfqd, cfqq, crq);
1252} 1838}
1253 1839
1254static void 1840static void
1255cfq_insert_request(request_queue_t *q, struct request *rq, int where) 1841cfq_insert_request(request_queue_t *q, struct request *rq, int where)
1256{ 1842{
1257 struct cfq_data *cfqd = q->elevator->elevator_data; 1843 struct cfq_data *cfqd = q->elevator->elevator_data;
1258 struct cfq_rq *crq = RQ_DATA(rq);
1259 1844
1260 switch (where) { 1845 switch (where) {
1261 case ELEVATOR_INSERT_BACK: 1846 case ELEVATOR_INSERT_BACK:
1262 while (cfq_dispatch_requests(q, cfqd->cfq_quantum)) 1847 while (cfq_dispatch_requests(q, INT_MAX, 1))
1263 ; 1848 ;
1264 list_add_tail(&rq->queuelist, &q->queue_head); 1849 list_add_tail(&rq->queuelist, &q->queue_head);
1850 /*
1851 * If we were idling with pending requests on
1852 * inactive cfqqs, force dispatching will
1853 * remove the idle timer and the queue won't
1854 * be kicked by __make_request() afterward.
1855 * Kick it here.
1856 */
1857 cfq_schedule_dispatch(cfqd);
1265 break; 1858 break;
1266 case ELEVATOR_INSERT_FRONT: 1859 case ELEVATOR_INSERT_FRONT:
1267 list_add(&rq->queuelist, &q->queue_head); 1860 list_add(&rq->queuelist, &q->queue_head);
1268 break; 1861 break;
1269 case ELEVATOR_INSERT_SORT: 1862 case ELEVATOR_INSERT_SORT:
1270 BUG_ON(!blk_fs_request(rq)); 1863 BUG_ON(!blk_fs_request(rq));
1271 cfq_enqueue(cfqd, crq); 1864 cfq_enqueue(cfqd, rq);
1272 break; 1865 break;
1273 default: 1866 default:
1274 printk("%s: bad insert point %d\n", __FUNCTION__,where); 1867 printk("%s: bad insert point %d\n", __FUNCTION__,where);
1275 return; 1868 return;
1276 } 1869 }
1277
1278 if (rq_mergeable(rq)) {
1279 cfq_add_crq_hash(cfqd, crq);
1280
1281 if (!q->last_merge)
1282 q->last_merge = rq;
1283 }
1284}
1285
1286static int cfq_queue_empty(request_queue_t *q)
1287{
1288 struct cfq_data *cfqd = q->elevator->elevator_data;
1289
1290 return list_empty(&q->queue_head) && list_empty(&cfqd->rr_list);
1291} 1870}
1292 1871
1293static void cfq_completed_request(request_queue_t *q, struct request *rq) 1872static void cfq_completed_request(request_queue_t *q, struct request *rq)
@@ -1300,9 +1879,11 @@ static void cfq_completed_request(request_queue_t *q, struct request *rq)
1300 1879
1301 cfqq = crq->cfq_queue; 1880 cfqq = crq->cfq_queue;
1302 1881
1303 if (crq->in_flight) { 1882 if (cfq_crq_in_flight(crq)) {
1304 WARN_ON(!cfqq->in_flight); 1883 const int sync = cfq_crq_is_sync(crq);
1305 cfqq->in_flight--; 1884
1885 WARN_ON(!cfqq->on_dispatch[sync]);
1886 cfqq->on_dispatch[sync]--;
1306 } 1887 }
1307 1888
1308 cfq_account_completion(cfqq, crq); 1889 cfq_account_completion(cfqq, crq);
@@ -1332,51 +1913,136 @@ cfq_latter_request(request_queue_t *q, struct request *rq)
1332 return NULL; 1913 return NULL;
1333} 1914}
1334 1915
1335static int cfq_may_queue(request_queue_t *q, int rw) 1916/*
1917 * we temporarily boost lower priority queues if they are holding fs exclusive
1918 * resources. they are boosted to normal prio (CLASS_BE/4)
1919 */
1920static void cfq_prio_boost(struct cfq_queue *cfqq)
1336{ 1921{
1337 struct cfq_data *cfqd = q->elevator->elevator_data; 1922 const int ioprio_class = cfqq->ioprio_class;
1338 struct cfq_queue *cfqq; 1923 const int ioprio = cfqq->ioprio;
1339 int ret = ELV_MQUEUE_MAY;
1340 1924
1341 if (current->flags & PF_MEMALLOC) 1925 if (has_fs_excl()) {
1342 return ELV_MQUEUE_MAY; 1926 /*
1927 * boost idle prio on transactions that would lock out other
1928 * users of the filesystem
1929 */
1930 if (cfq_class_idle(cfqq))
1931 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1932 if (cfqq->ioprio > IOPRIO_NORM)
1933 cfqq->ioprio = IOPRIO_NORM;
1934 } else {
1935 /*
1936 * check if we need to unboost the queue
1937 */
1938 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1939 cfqq->ioprio_class = cfqq->org_ioprio_class;
1940 if (cfqq->ioprio != cfqq->org_ioprio)
1941 cfqq->ioprio = cfqq->org_ioprio;
1942 }
1343 1943
1344 cfqq = cfq_find_cfq_hash(cfqd, cfq_hash_key(cfqd, current)); 1944 /*
1345 if (cfqq) { 1945 * refile between round-robin lists if we moved the priority class
1346 int limit = cfqd->max_queued; 1946 */
1947 if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1948 cfq_cfqq_on_rr(cfqq))
1949 cfq_resort_rr_list(cfqq, 0);
1950}
1347 1951
1348 if (cfqq->allocated[rw] < cfqd->cfq_queued) 1952static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
1349 return ELV_MQUEUE_MUST; 1953{
1954 if (rw == READ || process_sync(task))
1955 return task->pid;
1350 1956
1351 if (cfqd->busy_queues) 1957 return CFQ_KEY_ASYNC;
1352 limit = q->nr_requests / cfqd->busy_queues; 1958}
1353 1959
1354 if (limit < cfqd->cfq_queued) 1960static inline int
1355 limit = cfqd->cfq_queued; 1961__cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1356 else if (limit > cfqd->max_queued) 1962 struct task_struct *task, int rw)
1357 limit = cfqd->max_queued; 1963{
1964#if 1
1965 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1966 !cfq_cfqq_must_alloc_slice(cfqq)) {
1967 cfq_mark_cfqq_must_alloc_slice(cfqq);
1968 return ELV_MQUEUE_MUST;
1969 }
1358 1970
1359 if (cfqq->allocated[rw] >= limit) { 1971 return ELV_MQUEUE_MAY;
1360 if (limit > cfqq->alloc_limit[rw]) 1972#else
1361 cfqq->alloc_limit[rw] = limit; 1973 if (!cfqq || task->flags & PF_MEMALLOC)
1974 return ELV_MQUEUE_MAY;
1975 if (!cfqq->allocated[rw] || cfq_cfqq_must_alloc(cfqq)) {
1976 if (cfq_cfqq_wait_request(cfqq))
1977 return ELV_MQUEUE_MUST;
1362 1978
1363 ret = ELV_MQUEUE_NO; 1979 /*
1980 * only allow 1 ELV_MQUEUE_MUST per slice, otherwise we
1981 * can quickly flood the queue with writes from a single task
1982 */
1983 if (rw == READ || !cfq_cfqq_must_alloc_slice(cfqq)) {
1984 cfq_mark_cfqq_must_alloc_slice(cfqq);
1985 return ELV_MQUEUE_MUST;
1364 } 1986 }
1987
1988 return ELV_MQUEUE_MAY;
1365 } 1989 }
1990 if (cfq_class_idle(cfqq))
1991 return ELV_MQUEUE_NO;
1992 if (cfqq->allocated[rw] >= cfqd->max_queued) {
1993 struct io_context *ioc = get_io_context(GFP_ATOMIC);
1994 int ret = ELV_MQUEUE_NO;
1366 1995
1367 return ret; 1996 if (ioc && ioc->nr_batch_requests)
1997 ret = ELV_MQUEUE_MAY;
1998
1999 put_io_context(ioc);
2000 return ret;
2001 }
2002
2003 return ELV_MQUEUE_MAY;
2004#endif
2005}
2006
2007static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
2008{
2009 struct cfq_data *cfqd = q->elevator->elevator_data;
2010 struct task_struct *tsk = current;
2011 struct cfq_queue *cfqq;
2012
2013 /*
2014 * don't force setup of a queue from here, as a call to may_queue
2015 * does not necessarily imply that a request actually will be queued.
2016 * so just lookup a possibly existing queue, or return 'may queue'
2017 * if that fails
2018 */
2019 cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
2020 if (cfqq) {
2021 cfq_init_prio_data(cfqq);
2022 cfq_prio_boost(cfqq);
2023
2024 return __cfq_may_queue(cfqd, cfqq, tsk, rw);
2025 }
2026
2027 return ELV_MQUEUE_MAY;
1368} 2028}
1369 2029
1370static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq) 2030static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
1371{ 2031{
2032 struct cfq_data *cfqd = q->elevator->elevator_data;
1372 struct request_list *rl = &q->rq; 2033 struct request_list *rl = &q->rq;
1373 const int write = waitqueue_active(&rl->wait[WRITE]);
1374 const int read = waitqueue_active(&rl->wait[READ]);
1375 2034
1376 if (read && cfqq->allocated[READ] < cfqq->alloc_limit[READ]) 2035 if (cfqq->allocated[READ] <= cfqd->max_queued || cfqd->rq_starved) {
1377 wake_up(&rl->wait[READ]); 2036 smp_mb();
1378 if (write && cfqq->allocated[WRITE] < cfqq->alloc_limit[WRITE]) 2037 if (waitqueue_active(&rl->wait[READ]))
1379 wake_up(&rl->wait[WRITE]); 2038 wake_up(&rl->wait[READ]);
2039 }
2040
2041 if (cfqq->allocated[WRITE] <= cfqd->max_queued || cfqd->rq_starved) {
2042 smp_mb();
2043 if (waitqueue_active(&rl->wait[WRITE]))
2044 wake_up(&rl->wait[WRITE]);
2045 }
1380} 2046}
1381 2047
1382/* 2048/*
@@ -1389,69 +2055,61 @@ static void cfq_put_request(request_queue_t *q, struct request *rq)
1389 2055
1390 if (crq) { 2056 if (crq) {
1391 struct cfq_queue *cfqq = crq->cfq_queue; 2057 struct cfq_queue *cfqq = crq->cfq_queue;
2058 const int rw = rq_data_dir(rq);
1392 2059
1393 BUG_ON(q->last_merge == rq); 2060 BUG_ON(!cfqq->allocated[rw]);
1394 BUG_ON(!hlist_unhashed(&crq->hash)); 2061 cfqq->allocated[rw]--;
1395 2062
1396 if (crq->io_context) 2063 put_io_context(crq->io_context->ioc);
1397 put_io_context(crq->io_context->ioc);
1398
1399 BUG_ON(!cfqq->allocated[crq->is_write]);
1400 cfqq->allocated[crq->is_write]--;
1401 2064
1402 mempool_free(crq, cfqd->crq_pool); 2065 mempool_free(crq, cfqd->crq_pool);
1403 rq->elevator_private = NULL; 2066 rq->elevator_private = NULL;
1404 2067
1405 smp_mb();
1406 cfq_check_waiters(q, cfqq); 2068 cfq_check_waiters(q, cfqq);
1407 cfq_put_queue(cfqq); 2069 cfq_put_queue(cfqq);
1408 } 2070 }
1409} 2071}
1410 2072
1411/* 2073/*
1412 * Allocate cfq data structures associated with this request. A queue and 2074 * Allocate cfq data structures associated with this request.
1413 */ 2075 */
1414static int cfq_set_request(request_queue_t *q, struct request *rq, int gfp_mask) 2076static int
2077cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
2078 int gfp_mask)
1415{ 2079{
1416 struct cfq_data *cfqd = q->elevator->elevator_data; 2080 struct cfq_data *cfqd = q->elevator->elevator_data;
2081 struct task_struct *tsk = current;
1417 struct cfq_io_context *cic; 2082 struct cfq_io_context *cic;
1418 const int rw = rq_data_dir(rq); 2083 const int rw = rq_data_dir(rq);
1419 struct cfq_queue *cfqq, *saved_cfqq; 2084 pid_t key = cfq_queue_pid(tsk, rw);
2085 struct cfq_queue *cfqq;
1420 struct cfq_rq *crq; 2086 struct cfq_rq *crq;
1421 unsigned long flags; 2087 unsigned long flags;
1422 2088
1423 might_sleep_if(gfp_mask & __GFP_WAIT); 2089 might_sleep_if(gfp_mask & __GFP_WAIT);
1424 2090
2091 cic = cfq_get_io_context(cfqd, key, gfp_mask);
2092
1425 spin_lock_irqsave(q->queue_lock, flags); 2093 spin_lock_irqsave(q->queue_lock, flags);
1426 2094
1427 cfqq = __cfq_get_queue(cfqd, cfq_hash_key(cfqd, current), gfp_mask); 2095 if (!cic)
1428 if (!cfqq) 2096 goto queue_fail;
1429 goto out_lock; 2097
2098 if (!cic->cfqq) {
2099 cfqq = cfq_get_queue(cfqd, key, tsk->ioprio, gfp_mask);
2100 if (!cfqq)
2101 goto queue_fail;
1430 2102
1431repeat: 2103 cic->cfqq = cfqq;
1432 if (cfqq->allocated[rw] >= cfqd->max_queued) 2104 } else
1433 goto out_lock; 2105 cfqq = cic->cfqq;
1434 2106
1435 cfqq->allocated[rw]++; 2107 cfqq->allocated[rw]++;
2108 cfq_clear_cfqq_must_alloc(cfqq);
2109 cfqd->rq_starved = 0;
2110 atomic_inc(&cfqq->ref);
1436 spin_unlock_irqrestore(q->queue_lock, flags); 2111 spin_unlock_irqrestore(q->queue_lock, flags);
1437 2112
1438 /*
1439 * if hashing type has changed, the cfq_queue might change here.
1440 */
1441 saved_cfqq = cfqq;
1442 cic = cfq_get_io_context(&cfqq, gfp_mask);
1443 if (!cic)
1444 goto err;
1445
1446 /*
1447 * repeat allocation checks on queue change
1448 */
1449 if (unlikely(saved_cfqq != cfqq)) {
1450 spin_lock_irqsave(q->queue_lock, flags);
1451 saved_cfqq->allocated[rw]--;
1452 goto repeat;
1453 }
1454
1455 crq = mempool_alloc(cfqd->crq_pool, gfp_mask); 2113 crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
1456 if (crq) { 2114 if (crq) {
1457 RB_CLEAR(&crq->rb_node); 2115 RB_CLEAR(&crq->rb_node);
@@ -1460,24 +2118,141 @@ repeat:
1460 INIT_HLIST_NODE(&crq->hash); 2118 INIT_HLIST_NODE(&crq->hash);
1461 crq->cfq_queue = cfqq; 2119 crq->cfq_queue = cfqq;
1462 crq->io_context = cic; 2120 crq->io_context = cic;
1463 crq->service_start = crq->queue_start = 0; 2121 cfq_clear_crq_in_flight(crq);
1464 crq->in_flight = crq->accounted = crq->is_sync = 0; 2122 cfq_clear_crq_in_driver(crq);
1465 crq->is_write = rw; 2123 cfq_clear_crq_requeued(crq);
2124
2125 if (rw == READ || process_sync(tsk))
2126 cfq_mark_crq_is_sync(crq);
2127 else
2128 cfq_clear_crq_is_sync(crq);
2129
1466 rq->elevator_private = crq; 2130 rq->elevator_private = crq;
1467 cfqq->alloc_limit[rw] = 0;
1468 return 0; 2131 return 0;
1469 } 2132 }
1470 2133
1471 put_io_context(cic->ioc);
1472err:
1473 spin_lock_irqsave(q->queue_lock, flags); 2134 spin_lock_irqsave(q->queue_lock, flags);
1474 cfqq->allocated[rw]--; 2135 cfqq->allocated[rw]--;
2136 if (!(cfqq->allocated[0] + cfqq->allocated[1]))
2137 cfq_mark_cfqq_must_alloc(cfqq);
1475 cfq_put_queue(cfqq); 2138 cfq_put_queue(cfqq);
1476out_lock: 2139queue_fail:
2140 if (cic)
2141 put_io_context(cic->ioc);
2142 /*
2143 * mark us rq allocation starved. we need to kickstart the process
2144 * ourselves if there are no pending requests that can do it for us.
2145 * that would be an extremely rare OOM situation
2146 */
2147 cfqd->rq_starved = 1;
2148 cfq_schedule_dispatch(cfqd);
1477 spin_unlock_irqrestore(q->queue_lock, flags); 2149 spin_unlock_irqrestore(q->queue_lock, flags);
1478 return 1; 2150 return 1;
1479} 2151}
1480 2152
2153static void cfq_kick_queue(void *data)
2154{
2155 request_queue_t *q = data;
2156 struct cfq_data *cfqd = q->elevator->elevator_data;
2157 unsigned long flags;
2158
2159 spin_lock_irqsave(q->queue_lock, flags);
2160
2161 if (cfqd->rq_starved) {
2162 struct request_list *rl = &q->rq;
2163
2164 /*
2165 * we aren't guaranteed to get a request after this, but we
2166 * have to be opportunistic
2167 */
2168 smp_mb();
2169 if (waitqueue_active(&rl->wait[READ]))
2170 wake_up(&rl->wait[READ]);
2171 if (waitqueue_active(&rl->wait[WRITE]))
2172 wake_up(&rl->wait[WRITE]);
2173 }
2174
2175 blk_remove_plug(q);
2176 q->request_fn(q);
2177 spin_unlock_irqrestore(q->queue_lock, flags);
2178}
2179
2180/*
2181 * Timer running if the active_queue is currently idling inside its time slice
2182 */
2183static void cfq_idle_slice_timer(unsigned long data)
2184{
2185 struct cfq_data *cfqd = (struct cfq_data *) data;
2186 struct cfq_queue *cfqq;
2187 unsigned long flags;
2188
2189 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2190
2191 if ((cfqq = cfqd->active_queue) != NULL) {
2192 unsigned long now = jiffies;
2193
2194 /*
2195 * expired
2196 */
2197 if (time_after(now, cfqq->slice_end))
2198 goto expire;
2199
2200 /*
2201 * only expire and reinvoke request handler, if there are
2202 * other queues with pending requests
2203 */
2204 if (!cfq_pending_requests(cfqd)) {
2205 cfqd->idle_slice_timer.expires = min(now + cfqd->cfq_slice_idle, cfqq->slice_end);
2206 add_timer(&cfqd->idle_slice_timer);
2207 goto out_cont;
2208 }
2209
2210 /*
2211 * not expired and it has a request pending, let it dispatch
2212 */
2213 if (!RB_EMPTY(&cfqq->sort_list)) {
2214 cfq_mark_cfqq_must_dispatch(cfqq);
2215 goto out_kick;
2216 }
2217 }
2218expire:
2219 cfq_slice_expired(cfqd, 0);
2220out_kick:
2221 cfq_schedule_dispatch(cfqd);
2222out_cont:
2223 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2224}
2225
2226/*
2227 * Timer running if an idle class queue is waiting for service
2228 */
2229static void cfq_idle_class_timer(unsigned long data)
2230{
2231 struct cfq_data *cfqd = (struct cfq_data *) data;
2232 unsigned long flags, end;
2233
2234 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2235
2236 /*
2237 * race with a non-idle queue, reset timer
2238 */
2239 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2240 if (!time_after_eq(jiffies, end)) {
2241 cfqd->idle_class_timer.expires = end;
2242 add_timer(&cfqd->idle_class_timer);
2243 } else
2244 cfq_schedule_dispatch(cfqd);
2245
2246 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2247}
2248
2249static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2250{
2251 del_timer_sync(&cfqd->idle_slice_timer);
2252 del_timer_sync(&cfqd->idle_class_timer);
2253 blk_sync_queue(cfqd->queue);
2254}
2255
1481static void cfq_put_cfqd(struct cfq_data *cfqd) 2256static void cfq_put_cfqd(struct cfq_data *cfqd)
1482{ 2257{
1483 request_queue_t *q = cfqd->queue; 2258 request_queue_t *q = cfqd->queue;
@@ -1487,6 +2262,9 @@ static void cfq_put_cfqd(struct cfq_data *cfqd)
1487 2262
1488 blk_put_queue(q); 2263 blk_put_queue(q);
1489 2264
2265 cfq_shutdown_timer_wq(cfqd);
2266 q->elevator->elevator_data = NULL;
2267
1490 mempool_destroy(cfqd->crq_pool); 2268 mempool_destroy(cfqd->crq_pool);
1491 kfree(cfqd->crq_hash); 2269 kfree(cfqd->crq_hash);
1492 kfree(cfqd->cfq_hash); 2270 kfree(cfqd->cfq_hash);
@@ -1495,7 +2273,10 @@ static void cfq_put_cfqd(struct cfq_data *cfqd)
1495 2273
1496static void cfq_exit_queue(elevator_t *e) 2274static void cfq_exit_queue(elevator_t *e)
1497{ 2275{
1498 cfq_put_cfqd(e->elevator_data); 2276 struct cfq_data *cfqd = e->elevator_data;
2277
2278 cfq_shutdown_timer_wq(cfqd);
2279 cfq_put_cfqd(cfqd);
1499} 2280}
1500 2281
1501static int cfq_init_queue(request_queue_t *q, elevator_t *e) 2282static int cfq_init_queue(request_queue_t *q, elevator_t *e)
@@ -1508,7 +2289,13 @@ static int cfq_init_queue(request_queue_t *q, elevator_t *e)
1508 return -ENOMEM; 2289 return -ENOMEM;
1509 2290
1510 memset(cfqd, 0, sizeof(*cfqd)); 2291 memset(cfqd, 0, sizeof(*cfqd));
1511 INIT_LIST_HEAD(&cfqd->rr_list); 2292
2293 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2294 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2295
2296 INIT_LIST_HEAD(&cfqd->busy_rr);
2297 INIT_LIST_HEAD(&cfqd->cur_rr);
2298 INIT_LIST_HEAD(&cfqd->idle_rr);
1512 INIT_LIST_HEAD(&cfqd->empty_list); 2299 INIT_LIST_HEAD(&cfqd->empty_list);
1513 2300
1514 cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL); 2301 cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL);
@@ -1533,24 +2320,32 @@ static int cfq_init_queue(request_queue_t *q, elevator_t *e)
1533 cfqd->queue = q; 2320 cfqd->queue = q;
1534 atomic_inc(&q->refcnt); 2321 atomic_inc(&q->refcnt);
1535 2322
1536 /* 2323 cfqd->max_queued = q->nr_requests / 4;
1537 * just set it to some high value, we want anyone to be able to queue
1538 * some requests. fairness is handled differently
1539 */
1540 q->nr_requests = 1024;
1541 cfqd->max_queued = q->nr_requests / 16;
1542 q->nr_batching = cfq_queued; 2324 q->nr_batching = cfq_queued;
1543 cfqd->key_type = CFQ_KEY_TGID; 2325
1544 cfqd->find_best_crq = 1; 2326 init_timer(&cfqd->idle_slice_timer);
2327 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2328 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2329
2330 init_timer(&cfqd->idle_class_timer);
2331 cfqd->idle_class_timer.function = cfq_idle_class_timer;
2332 cfqd->idle_class_timer.data = (unsigned long) cfqd;
2333
2334 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2335
1545 atomic_set(&cfqd->ref, 1); 2336 atomic_set(&cfqd->ref, 1);
1546 2337
1547 cfqd->cfq_queued = cfq_queued; 2338 cfqd->cfq_queued = cfq_queued;
1548 cfqd->cfq_quantum = cfq_quantum; 2339 cfqd->cfq_quantum = cfq_quantum;
1549 cfqd->cfq_fifo_expire_r = cfq_fifo_expire_r; 2340 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
1550 cfqd->cfq_fifo_expire_w = cfq_fifo_expire_w; 2341 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
1551 cfqd->cfq_fifo_batch_expire = cfq_fifo_rate;
1552 cfqd->cfq_back_max = cfq_back_max; 2342 cfqd->cfq_back_max = cfq_back_max;
1553 cfqd->cfq_back_penalty = cfq_back_penalty; 2343 cfqd->cfq_back_penalty = cfq_back_penalty;
2344 cfqd->cfq_slice[0] = cfq_slice_async;
2345 cfqd->cfq_slice[1] = cfq_slice_sync;
2346 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2347 cfqd->cfq_slice_idle = cfq_slice_idle;
2348 cfqd->cfq_max_depth = cfq_max_depth;
1554 2349
1555 return 0; 2350 return 0;
1556out_crqpool: 2351out_crqpool:
@@ -1595,7 +2390,6 @@ fail:
1595 return -ENOMEM; 2390 return -ENOMEM;
1596} 2391}
1597 2392
1598
1599/* 2393/*
1600 * sysfs parts below --> 2394 * sysfs parts below -->
1601 */ 2395 */
@@ -1620,45 +2414,6 @@ cfq_var_store(unsigned int *var, const char *page, size_t count)
1620 return count; 2414 return count;
1621} 2415}
1622 2416
1623static ssize_t
1624cfq_clear_elapsed(struct cfq_data *cfqd, const char *page, size_t count)
1625{
1626 max_elapsed_dispatch = max_elapsed_crq = 0;
1627 return count;
1628}
1629
1630static ssize_t
1631cfq_set_key_type(struct cfq_data *cfqd, const char *page, size_t count)
1632{
1633 spin_lock_irq(cfqd->queue->queue_lock);
1634 if (!strncmp(page, "pgid", 4))
1635 cfqd->key_type = CFQ_KEY_PGID;
1636 else if (!strncmp(page, "tgid", 4))
1637 cfqd->key_type = CFQ_KEY_TGID;
1638 else if (!strncmp(page, "uid", 3))
1639 cfqd->key_type = CFQ_KEY_UID;
1640 else if (!strncmp(page, "gid", 3))
1641 cfqd->key_type = CFQ_KEY_GID;
1642 spin_unlock_irq(cfqd->queue->queue_lock);
1643 return count;
1644}
1645
1646static ssize_t
1647cfq_read_key_type(struct cfq_data *cfqd, char *page)
1648{
1649 ssize_t len = 0;
1650 int i;
1651
1652 for (i = CFQ_KEY_PGID; i < CFQ_KEY_LAST; i++) {
1653 if (cfqd->key_type == i)
1654 len += sprintf(page+len, "[%s] ", cfq_key_types[i]);
1655 else
1656 len += sprintf(page+len, "%s ", cfq_key_types[i]);
1657 }
1658 len += sprintf(page+len, "\n");
1659 return len;
1660}
1661
1662#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ 2417#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
1663static ssize_t __FUNC(struct cfq_data *cfqd, char *page) \ 2418static ssize_t __FUNC(struct cfq_data *cfqd, char *page) \
1664{ \ 2419{ \
@@ -1669,12 +2424,15 @@ static ssize_t __FUNC(struct cfq_data *cfqd, char *page) \
1669} 2424}
1670SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); 2425SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
1671SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0); 2426SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
1672SHOW_FUNCTION(cfq_fifo_expire_r_show, cfqd->cfq_fifo_expire_r, 1); 2427SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
1673SHOW_FUNCTION(cfq_fifo_expire_w_show, cfqd->cfq_fifo_expire_w, 1); 2428SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
1674SHOW_FUNCTION(cfq_fifo_batch_expire_show, cfqd->cfq_fifo_batch_expire, 1);
1675SHOW_FUNCTION(cfq_find_best_show, cfqd->find_best_crq, 0);
1676SHOW_FUNCTION(cfq_back_max_show, cfqd->cfq_back_max, 0); 2429SHOW_FUNCTION(cfq_back_max_show, cfqd->cfq_back_max, 0);
1677SHOW_FUNCTION(cfq_back_penalty_show, cfqd->cfq_back_penalty, 0); 2430SHOW_FUNCTION(cfq_back_penalty_show, cfqd->cfq_back_penalty, 0);
2431SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2432SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2433SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2434SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2435SHOW_FUNCTION(cfq_max_depth_show, cfqd->cfq_max_depth, 0);
1678#undef SHOW_FUNCTION 2436#undef SHOW_FUNCTION
1679 2437
1680#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ 2438#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
@@ -1694,12 +2452,15 @@ static ssize_t __FUNC(struct cfq_data *cfqd, const char *page, size_t count) \
1694} 2452}
1695STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); 2453STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
1696STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0); 2454STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
1697STORE_FUNCTION(cfq_fifo_expire_r_store, &cfqd->cfq_fifo_expire_r, 1, UINT_MAX, 1); 2455STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
1698STORE_FUNCTION(cfq_fifo_expire_w_store, &cfqd->cfq_fifo_expire_w, 1, UINT_MAX, 1); 2456STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
1699STORE_FUNCTION(cfq_fifo_batch_expire_store, &cfqd->cfq_fifo_batch_expire, 0, UINT_MAX, 1);
1700STORE_FUNCTION(cfq_find_best_store, &cfqd->find_best_crq, 0, 1, 0);
1701STORE_FUNCTION(cfq_back_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); 2457STORE_FUNCTION(cfq_back_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
1702STORE_FUNCTION(cfq_back_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0); 2458STORE_FUNCTION(cfq_back_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2459STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2460STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2461STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2462STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2463STORE_FUNCTION(cfq_max_depth_store, &cfqd->cfq_max_depth, 1, UINT_MAX, 0);
1703#undef STORE_FUNCTION 2464#undef STORE_FUNCTION
1704 2465
1705static struct cfq_fs_entry cfq_quantum_entry = { 2466static struct cfq_fs_entry cfq_quantum_entry = {
@@ -1712,25 +2473,15 @@ static struct cfq_fs_entry cfq_queued_entry = {
1712 .show = cfq_queued_show, 2473 .show = cfq_queued_show,
1713 .store = cfq_queued_store, 2474 .store = cfq_queued_store,
1714}; 2475};
1715static struct cfq_fs_entry cfq_fifo_expire_r_entry = { 2476static struct cfq_fs_entry cfq_fifo_expire_sync_entry = {
1716 .attr = {.name = "fifo_expire_sync", .mode = S_IRUGO | S_IWUSR }, 2477 .attr = {.name = "fifo_expire_sync", .mode = S_IRUGO | S_IWUSR },
1717 .show = cfq_fifo_expire_r_show, 2478 .show = cfq_fifo_expire_sync_show,
1718 .store = cfq_fifo_expire_r_store, 2479 .store = cfq_fifo_expire_sync_store,
1719}; 2480};
1720static struct cfq_fs_entry cfq_fifo_expire_w_entry = { 2481static struct cfq_fs_entry cfq_fifo_expire_async_entry = {
1721 .attr = {.name = "fifo_expire_async", .mode = S_IRUGO | S_IWUSR }, 2482 .attr = {.name = "fifo_expire_async", .mode = S_IRUGO | S_IWUSR },
1722 .show = cfq_fifo_expire_w_show, 2483 .show = cfq_fifo_expire_async_show,
1723 .store = cfq_fifo_expire_w_store, 2484 .store = cfq_fifo_expire_async_store,
1724};
1725static struct cfq_fs_entry cfq_fifo_batch_expire_entry = {
1726 .attr = {.name = "fifo_batch_expire", .mode = S_IRUGO | S_IWUSR },
1727 .show = cfq_fifo_batch_expire_show,
1728 .store = cfq_fifo_batch_expire_store,
1729};
1730static struct cfq_fs_entry cfq_find_best_entry = {
1731 .attr = {.name = "find_best_crq", .mode = S_IRUGO | S_IWUSR },
1732 .show = cfq_find_best_show,
1733 .store = cfq_find_best_store,
1734}; 2485};
1735static struct cfq_fs_entry cfq_back_max_entry = { 2486static struct cfq_fs_entry cfq_back_max_entry = {
1736 .attr = {.name = "back_seek_max", .mode = S_IRUGO | S_IWUSR }, 2487 .attr = {.name = "back_seek_max", .mode = S_IRUGO | S_IWUSR },
@@ -1742,27 +2493,44 @@ static struct cfq_fs_entry cfq_back_penalty_entry = {
1742 .show = cfq_back_penalty_show, 2493 .show = cfq_back_penalty_show,
1743 .store = cfq_back_penalty_store, 2494 .store = cfq_back_penalty_store,
1744}; 2495};
1745static struct cfq_fs_entry cfq_clear_elapsed_entry = { 2496static struct cfq_fs_entry cfq_slice_sync_entry = {
1746 .attr = {.name = "clear_elapsed", .mode = S_IWUSR }, 2497 .attr = {.name = "slice_sync", .mode = S_IRUGO | S_IWUSR },
1747 .store = cfq_clear_elapsed, 2498 .show = cfq_slice_sync_show,
2499 .store = cfq_slice_sync_store,
2500};
2501static struct cfq_fs_entry cfq_slice_async_entry = {
2502 .attr = {.name = "slice_async", .mode = S_IRUGO | S_IWUSR },
2503 .show = cfq_slice_async_show,
2504 .store = cfq_slice_async_store,
1748}; 2505};
1749static struct cfq_fs_entry cfq_key_type_entry = { 2506static struct cfq_fs_entry cfq_slice_async_rq_entry = {
1750 .attr = {.name = "key_type", .mode = S_IRUGO | S_IWUSR }, 2507 .attr = {.name = "slice_async_rq", .mode = S_IRUGO | S_IWUSR },
1751 .show = cfq_read_key_type, 2508 .show = cfq_slice_async_rq_show,
1752 .store = cfq_set_key_type, 2509 .store = cfq_slice_async_rq_store,
2510};
2511static struct cfq_fs_entry cfq_slice_idle_entry = {
2512 .attr = {.name = "slice_idle", .mode = S_IRUGO | S_IWUSR },
2513 .show = cfq_slice_idle_show,
2514 .store = cfq_slice_idle_store,
2515};
2516static struct cfq_fs_entry cfq_max_depth_entry = {
2517 .attr = {.name = "max_depth", .mode = S_IRUGO | S_IWUSR },
2518 .show = cfq_max_depth_show,
2519 .store = cfq_max_depth_store,
1753}; 2520};
1754 2521
1755static struct attribute *default_attrs[] = { 2522static struct attribute *default_attrs[] = {
1756 &cfq_quantum_entry.attr, 2523 &cfq_quantum_entry.attr,
1757 &cfq_queued_entry.attr, 2524 &cfq_queued_entry.attr,
1758 &cfq_fifo_expire_r_entry.attr, 2525 &cfq_fifo_expire_sync_entry.attr,
1759 &cfq_fifo_expire_w_entry.attr, 2526 &cfq_fifo_expire_async_entry.attr,
1760 &cfq_fifo_batch_expire_entry.attr,
1761 &cfq_key_type_entry.attr,
1762 &cfq_find_best_entry.attr,
1763 &cfq_back_max_entry.attr, 2527 &cfq_back_max_entry.attr,
1764 &cfq_back_penalty_entry.attr, 2528 &cfq_back_penalty_entry.attr,
1765 &cfq_clear_elapsed_entry.attr, 2529 &cfq_slice_sync_entry.attr,
2530 &cfq_slice_async_entry.attr,
2531 &cfq_slice_async_rq_entry.attr,
2532 &cfq_slice_idle_entry.attr,
2533 &cfq_max_depth_entry.attr,
1766 NULL, 2534 NULL,
1767}; 2535};
1768 2536
@@ -1775,7 +2543,7 @@ cfq_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
1775 struct cfq_fs_entry *entry = to_cfq(attr); 2543 struct cfq_fs_entry *entry = to_cfq(attr);
1776 2544
1777 if (!entry->show) 2545 if (!entry->show)
1778 return 0; 2546 return -EIO;
1779 2547
1780 return entry->show(e->elevator_data, page); 2548 return entry->show(e->elevator_data, page);
1781} 2549}
@@ -1788,7 +2556,7 @@ cfq_attr_store(struct kobject *kobj, struct attribute *attr,
1788 struct cfq_fs_entry *entry = to_cfq(attr); 2556 struct cfq_fs_entry *entry = to_cfq(attr);
1789 2557
1790 if (!entry->store) 2558 if (!entry->store)
1791 return -EINVAL; 2559 return -EIO;
1792 2560
1793 return entry->store(e->elevator_data, page, length); 2561 return entry->store(e->elevator_data, page, length);
1794} 2562}
@@ -1832,21 +2600,46 @@ static int __init cfq_init(void)
1832{ 2600{
1833 int ret; 2601 int ret;
1834 2602
2603 /*
2604 * could be 0 on HZ < 1000 setups
2605 */
2606 if (!cfq_slice_async)
2607 cfq_slice_async = 1;
2608 if (!cfq_slice_idle)
2609 cfq_slice_idle = 1;
2610
1835 if (cfq_slab_setup()) 2611 if (cfq_slab_setup())
1836 return -ENOMEM; 2612 return -ENOMEM;
1837 2613
1838 ret = elv_register(&iosched_cfq); 2614 ret = elv_register(&iosched_cfq);
1839 if (!ret) { 2615 if (ret)
1840 __module_get(THIS_MODULE); 2616 cfq_slab_kill();
1841 return 0;
1842 }
1843 2617
1844 cfq_slab_kill();
1845 return ret; 2618 return ret;
1846} 2619}
1847 2620
1848static void __exit cfq_exit(void) 2621static void __exit cfq_exit(void)
1849{ 2622{
2623 struct task_struct *g, *p;
2624 unsigned long flags;
2625
2626 read_lock_irqsave(&tasklist_lock, flags);
2627
2628 /*
2629 * iterate each process in the system, removing our io_context
2630 */
2631 do_each_thread(g, p) {
2632 struct io_context *ioc = p->io_context;
2633
2634 if (ioc && ioc->cic) {
2635 ioc->cic->exit(ioc->cic);
2636 cfq_free_io_context(ioc->cic);
2637 ioc->cic = NULL;
2638 }
2639 } while_each_thread(g, p);
2640
2641 read_unlock_irqrestore(&tasklist_lock, flags);
2642
1850 cfq_slab_kill(); 2643 cfq_slab_kill();
1851 elv_unregister(&iosched_cfq); 2644 elv_unregister(&iosched_cfq);
1852} 2645}
generated by cgit v1.2.2 (git 2.25.0) at 2024-11-23 07:25:42 -0500