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.c2068
1 files changed, 1425 insertions, 643 deletions
diff --git a/drivers/block/cfq-iosched.c b/drivers/block/cfq-iosched.c
index 3ac47dde64da..ff1cc968f96d 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 = 1;
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 +230,79 @@ 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);
303static inline int cfq_pending_requests(struct cfq_data *cfqd);
200 304
201/* 305#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}
223 306
224/* 307/*
225 * lots of deadline iosched dupes, can be abstracted later... 308 * lots of deadline iosched dupes, can be abstracted later...
@@ -235,16 +318,12 @@ static void cfq_remove_merge_hints(request_queue_t *q, struct cfq_rq *crq)
235 318
236 if (q->last_merge == crq->request) 319 if (q->last_merge == crq->request)
237 q->last_merge = NULL; 320 q->last_merge = NULL;
238
239 cfq_update_next_crq(crq);
240} 321}
241 322
242static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq) 323static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq)
243{ 324{
244 const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request)); 325 const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request));
245 326
246 BUG_ON(!hlist_unhashed(&crq->hash));
247
248 hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]); 327 hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]);
249} 328}
250 329
@@ -257,8 +336,6 @@ static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
257 struct cfq_rq *crq = list_entry_hash(entry); 336 struct cfq_rq *crq = list_entry_hash(entry);
258 struct request *__rq = crq->request; 337 struct request *__rq = crq->request;
259 338
260 BUG_ON(hlist_unhashed(&crq->hash));
261
262 if (!rq_mergeable(__rq)) { 339 if (!rq_mergeable(__rq)) {
263 cfq_del_crq_hash(crq); 340 cfq_del_crq_hash(crq);
264 continue; 341 continue;
@@ -287,36 +364,16 @@ cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
287 return crq2; 364 return crq2;
288 if (crq2 == NULL) 365 if (crq2 == NULL)
289 return crq1; 366 return crq1;
367 if (cfq_crq_requeued(crq1))
368 return crq1;
369 if (cfq_crq_requeued(crq2))
370 return crq2;
290 371
291 s1 = crq1->request->sector; 372 s1 = crq1->request->sector;
292 s2 = crq2->request->sector; 373 s2 = crq2->request->sector;
293 374
294 last = cfqd->last_sector; 375 last = cfqd->last_sector;
295 376
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 /* 377 /*
321 * by definition, 1KiB is 2 sectors 378 * by definition, 1KiB is 2 sectors
322 */ 379 */
@@ -377,11 +434,14 @@ cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
377 struct cfq_rq *crq_next = NULL, *crq_prev = NULL; 434 struct cfq_rq *crq_next = NULL, *crq_prev = NULL;
378 struct rb_node *rbnext, *rbprev; 435 struct rb_node *rbnext, *rbprev;
379 436
380 if (!ON_RB(&last->rb_node)) 437 rbnext = NULL;
381 return NULL; 438 if (ON_RB(&last->rb_node))
382 439 rbnext = rb_next(&last->rb_node);
383 if ((rbnext = rb_next(&last->rb_node)) == NULL) 440 if (!rbnext) {
384 rbnext = rb_first(&cfqq->sort_list); 441 rbnext = rb_first(&cfqq->sort_list);
442 if (rbnext == &last->rb_node)
443 rbnext = NULL;
444 }
385 445
386 rbprev = rb_prev(&last->rb_node); 446 rbprev = rb_prev(&last->rb_node);
387 447
@@ -401,67 +461,53 @@ static void cfq_update_next_crq(struct cfq_rq *crq)
401 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq); 461 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
402} 462}
403 463
404static int cfq_check_sort_rr_list(struct cfq_queue *cfqq) 464static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
405{ 465{
406 struct list_head *head = &cfqq->cfqd->rr_list; 466 struct cfq_data *cfqd = cfqq->cfqd;
407 struct list_head *next, *prev; 467 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);
415 468
416 if (cfqq->service_used > cnext->service_used) 469 BUG_ON(!cfq_cfqq_on_rr(cfqq));
417 return 1;
418 }
419 470
420 prev = cfqq->cfq_list.prev; 471 list_del(&cfqq->cfq_list);
421 if (prev != head) {
422 struct cfq_queue *cprev = list_entry_cfqq(prev);
423 472
424 if (cfqq->service_used < cprev->service_used) 473 if (cfq_class_rt(cfqq))
425 return 1; 474 list = &cfqd->cur_rr;
475 else if (cfq_class_idle(cfqq))
476 list = &cfqd->idle_rr;
477 else {
478 /*
479 * if cfqq has requests in flight, don't allow it to be
480 * found in cfq_set_active_queue before it has finished them.
481 * this is done to increase fairness between a process that
482 * has lots of io pending vs one that only generates one
483 * sporadically or synchronously
484 */
485 if (cfq_cfqq_dispatched(cfqq))
486 list = &cfqd->busy_rr;
487 else
488 list = &cfqd->rr_list[cfqq->ioprio];
426 } 489 }
427 490
428 return 0; 491 /*
429} 492 * if queue was preempted, just add to front to be fair. busy_rr
430 493 * isn't sorted.
431static void cfq_sort_rr_list(struct cfq_queue *cfqq, int new_queue) 494 */
432{ 495 if (preempted || list == &cfqd->busy_rr) {
433 struct list_head *entry = &cfqq->cfqd->rr_list; 496 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; 497 return;
439 498 }
440 list_del(&cfqq->cfq_list);
441 499
442 /* 500 /*
443 * sort by our mean service_used, sub-sort by in-flight requests 501 * sort by when queue was last serviced
444 */ 502 */
445 while ((entry = entry->prev) != &cfqq->cfqd->rr_list) { 503 entry = list;
504 while ((entry = entry->prev) != list) {
446 struct cfq_queue *__cfqq = list_entry_cfqq(entry); 505 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
447 506
448 if (cfqq->service_used > __cfqq->service_used) 507 if (!__cfqq->service_last)
508 break;
509 if (time_before(__cfqq->service_last, cfqq->service_last))
449 break; 510 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 } 511 }
466 512
467 list_add(&cfqq->cfq_list, entry); 513 list_add(&cfqq->cfq_list, entry);
@@ -469,28 +515,24 @@ static void cfq_sort_rr_list(struct cfq_queue *cfqq, int new_queue)
469 515
470/* 516/*
471 * add to busy list of queues for service, trying to be fair in ordering 517 * add to busy list of queues for service, trying to be fair in ordering
472 * the pending list according to requests serviced 518 * the pending list according to last request service
473 */ 519 */
474static inline void 520static inline void
475cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) 521cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq, int requeue)
476{ 522{
477 /* 523 BUG_ON(cfq_cfqq_on_rr(cfqq));
478 * it's currently on the empty list 524 cfq_mark_cfqq_on_rr(cfqq);
479 */
480 cfqq->on_rr = 1;
481 cfqd->busy_queues++; 525 cfqd->busy_queues++;
482 526
483 if (time_after(jiffies, cfqq->service_start + cfq_service)) 527 cfq_resort_rr_list(cfqq, requeue);
484 cfqq->service_used >>= 3;
485
486 cfq_sort_rr_list(cfqq, 1);
487} 528}
488 529
489static inline void 530static inline void
490cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) 531cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
491{ 532{
533 BUG_ON(!cfq_cfqq_on_rr(cfqq));
534 cfq_clear_cfqq_on_rr(cfqq);
492 list_move(&cfqq->cfq_list, &cfqd->empty_list); 535 list_move(&cfqq->cfq_list, &cfqd->empty_list);
493 cfqq->on_rr = 0;
494 536
495 BUG_ON(!cfqd->busy_queues); 537 BUG_ON(!cfqd->busy_queues);
496 cfqd->busy_queues--; 538 cfqd->busy_queues--;
@@ -505,16 +547,17 @@ static inline void cfq_del_crq_rb(struct cfq_rq *crq)
505 547
506 if (ON_RB(&crq->rb_node)) { 548 if (ON_RB(&crq->rb_node)) {
507 struct cfq_data *cfqd = cfqq->cfqd; 549 struct cfq_data *cfqd = cfqq->cfqd;
550 const int sync = cfq_crq_is_sync(crq);
508 551
509 BUG_ON(!cfqq->queued[crq->is_sync]); 552 BUG_ON(!cfqq->queued[sync]);
553 cfqq->queued[sync]--;
510 554
511 cfq_update_next_crq(crq); 555 cfq_update_next_crq(crq);
512 556
513 cfqq->queued[crq->is_sync]--;
514 rb_erase(&crq->rb_node, &cfqq->sort_list); 557 rb_erase(&crq->rb_node, &cfqq->sort_list);
515 RB_CLEAR_COLOR(&crq->rb_node); 558 RB_CLEAR_COLOR(&crq->rb_node);
516 559
517 if (RB_EMPTY(&cfqq->sort_list) && cfqq->on_rr) 560 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY(&cfqq->sort_list))
518 cfq_del_cfqq_rr(cfqd, cfqq); 561 cfq_del_cfqq_rr(cfqd, cfqq);
519 } 562 }
520} 563}
@@ -550,7 +593,7 @@ static void cfq_add_crq_rb(struct cfq_rq *crq)
550 struct cfq_rq *__alias; 593 struct cfq_rq *__alias;
551 594
552 crq->rb_key = rq_rb_key(rq); 595 crq->rb_key = rq_rb_key(rq);
553 cfqq->queued[crq->is_sync]++; 596 cfqq->queued[cfq_crq_is_sync(crq)]++;
554 597
555 /* 598 /*
556 * looks a little odd, but the first insert might return an alias. 599 * looks a little odd, but the first insert might return an alias.
@@ -561,8 +604,8 @@ static void cfq_add_crq_rb(struct cfq_rq *crq)
561 604
562 rb_insert_color(&crq->rb_node, &cfqq->sort_list); 605 rb_insert_color(&crq->rb_node, &cfqq->sort_list);
563 606
564 if (!cfqq->on_rr) 607 if (!cfq_cfqq_on_rr(cfqq))
565 cfq_add_cfqq_rr(cfqd, cfqq); 608 cfq_add_cfqq_rr(cfqd, cfqq, cfq_crq_requeued(crq));
566 609
567 /* 610 /*
568 * check if this request is a better next-serve candidate 611 * check if this request is a better next-serve candidate
@@ -575,17 +618,16 @@ cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
575{ 618{
576 if (ON_RB(&crq->rb_node)) { 619 if (ON_RB(&crq->rb_node)) {
577 rb_erase(&crq->rb_node, &cfqq->sort_list); 620 rb_erase(&crq->rb_node, &cfqq->sort_list);
578 cfqq->queued[crq->is_sync]--; 621 cfqq->queued[cfq_crq_is_sync(crq)]--;
579 } 622 }
580 623
581 cfq_add_crq_rb(crq); 624 cfq_add_crq_rb(crq);
582} 625}
583 626
584static struct request * 627static struct request *cfq_find_rq_rb(struct cfq_data *cfqd, sector_t sector)
585cfq_find_rq_rb(struct cfq_data *cfqd, sector_t sector) 628
586{ 629{
587 const unsigned long key = cfq_hash_key(cfqd, current); 630 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; 631 struct rb_node *n;
590 632
591 if (!cfqq) 633 if (!cfqq)
@@ -609,20 +651,25 @@ out:
609 651
610static void cfq_deactivate_request(request_queue_t *q, struct request *rq) 652static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
611{ 653{
654 struct cfq_data *cfqd = q->elevator->elevator_data;
612 struct cfq_rq *crq = RQ_DATA(rq); 655 struct cfq_rq *crq = RQ_DATA(rq);
613 656
614 if (crq) { 657 if (crq) {
615 struct cfq_queue *cfqq = crq->cfq_queue; 658 struct cfq_queue *cfqq = crq->cfq_queue;
616 659
617 if (cfqq->cfqd->cfq_tagged) { 660 if (cfq_crq_in_driver(crq)) {
618 cfqq->service_used--; 661 cfq_clear_crq_in_driver(crq);
619 cfq_sort_rr_list(cfqq, 0); 662 WARN_ON(!cfqd->rq_in_driver);
663 cfqd->rq_in_driver--;
620 } 664 }
665 if (cfq_crq_in_flight(crq)) {
666 const int sync = cfq_crq_is_sync(crq);
621 667
622 if (crq->accounted) { 668 cfq_clear_crq_in_flight(crq);
623 crq->accounted = 0; 669 WARN_ON(!cfqq->on_dispatch[sync]);
624 cfqq->cfqd->rq_in_driver--; 670 cfqq->on_dispatch[sync]--;
625 } 671 }
672 cfq_mark_crq_requeued(crq);
626 } 673 }
627} 674}
628 675
@@ -640,11 +687,10 @@ static void cfq_remove_request(request_queue_t *q, struct request *rq)
640 struct cfq_rq *crq = RQ_DATA(rq); 687 struct cfq_rq *crq = RQ_DATA(rq);
641 688
642 if (crq) { 689 if (crq) {
643 cfq_remove_merge_hints(q, crq);
644 list_del_init(&rq->queuelist); 690 list_del_init(&rq->queuelist);
691 cfq_del_crq_rb(crq);
692 cfq_remove_merge_hints(q, crq);
645 693
646 if (crq->cfq_queue)
647 cfq_del_crq_rb(crq);
648 } 694 }
649} 695}
650 696
@@ -662,21 +708,15 @@ cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
662 } 708 }
663 709
664 __rq = cfq_find_rq_hash(cfqd, bio->bi_sector); 710 __rq = cfq_find_rq_hash(cfqd, bio->bi_sector);
665 if (__rq) { 711 if (__rq && elv_rq_merge_ok(__rq, bio)) {
666 BUG_ON(__rq->sector + __rq->nr_sectors != bio->bi_sector); 712 ret = ELEVATOR_BACK_MERGE;
667 713 goto out;
668 if (elv_rq_merge_ok(__rq, bio)) {
669 ret = ELEVATOR_BACK_MERGE;
670 goto out;
671 }
672 } 714 }
673 715
674 __rq = cfq_find_rq_rb(cfqd, bio->bi_sector + bio_sectors(bio)); 716 __rq = cfq_find_rq_rb(cfqd, bio->bi_sector + bio_sectors(bio));
675 if (__rq) { 717 if (__rq && elv_rq_merge_ok(__rq, bio)) {
676 if (elv_rq_merge_ok(__rq, bio)) { 718 ret = ELEVATOR_FRONT_MERGE;
677 ret = ELEVATOR_FRONT_MERGE; 719 goto out;
678 goto out;
679 }
680 } 720 }
681 721
682 return ELEVATOR_NO_MERGE; 722 return ELEVATOR_NO_MERGE;
@@ -709,20 +749,220 @@ static void
709cfq_merged_requests(request_queue_t *q, struct request *rq, 749cfq_merged_requests(request_queue_t *q, struct request *rq,
710 struct request *next) 750 struct request *next)
711{ 751{
712 struct cfq_rq *crq = RQ_DATA(rq);
713 struct cfq_rq *cnext = RQ_DATA(next);
714
715 cfq_merged_request(q, rq); 752 cfq_merged_request(q, rq);
716 753
717 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist)) { 754 /*
718 if (time_before(cnext->queue_start, crq->queue_start)) { 755 * reposition in fifo if next is older than rq
719 list_move(&rq->queuelist, &next->queuelist); 756 */
720 crq->queue_start = cnext->queue_start; 757 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
758 time_before(next->start_time, rq->start_time))
759 list_move(&rq->queuelist, &next->queuelist);
760
761 cfq_remove_request(q, next);
762}
763
764static inline void
765__cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
766{
767 if (cfqq) {
768 /*
769 * stop potential idle class queues waiting service
770 */
771 del_timer(&cfqd->idle_class_timer);
772
773 cfqq->slice_start = jiffies;
774 cfqq->slice_end = 0;
775 cfqq->slice_left = 0;
776 cfq_clear_cfqq_must_alloc_slice(cfqq);
777 cfq_clear_cfqq_fifo_expire(cfqq);
778 cfq_clear_cfqq_expired(cfqq);
779 }
780
781 cfqd->active_queue = cfqq;
782}
783
784/*
785 * 0
786 * 0,1
787 * 0,1,2
788 * 0,1,2,3
789 * 0,1,2,3,4
790 * 0,1,2,3,4,5
791 * 0,1,2,3,4,5,6
792 * 0,1,2,3,4,5,6,7
793 */
794static int cfq_get_next_prio_level(struct cfq_data *cfqd)
795{
796 int prio, wrap;
797
798 prio = -1;
799 wrap = 0;
800 do {
801 int p;
802
803 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
804 if (!list_empty(&cfqd->rr_list[p])) {
805 prio = p;
806 break;
807 }
808 }
809
810 if (prio != -1)
811 break;
812 cfqd->cur_prio = 0;
813 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
814 cfqd->cur_end_prio = 0;
815 if (wrap)
816 break;
817 wrap = 1;
721 } 818 }
819 } while (1);
820
821 if (unlikely(prio == -1))
822 return -1;
823
824 BUG_ON(prio >= CFQ_PRIO_LISTS);
825
826 list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
827
828 cfqd->cur_prio = prio + 1;
829 if (cfqd->cur_prio > cfqd->cur_end_prio) {
830 cfqd->cur_end_prio = cfqd->cur_prio;
831 cfqd->cur_prio = 0;
832 }
833 if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
834 cfqd->cur_prio = 0;
835 cfqd->cur_end_prio = 0;
722 } 836 }
723 837
724 cfq_update_next_crq(cnext); 838 return prio;
725 cfq_remove_request(q, next); 839}
840
841static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
842{
843 struct cfq_queue *cfqq;
844
845 /*
846 * if current queue is expired but not done with its requests yet,
847 * wait for that to happen
848 */
849 if ((cfqq = cfqd->active_queue) != NULL) {
850 if (cfq_cfqq_expired(cfqq) && cfq_cfqq_dispatched(cfqq))
851 return NULL;
852 }
853
854 /*
855 * if current list is non-empty, grab first entry. if it is empty,
856 * get next prio level and grab first entry then if any are spliced
857 */
858 if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
859 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
860
861 /*
862 * if we have idle queues and no rt or be queues had pending
863 * requests, either allow immediate service if the grace period
864 * has passed or arm the idle grace timer
865 */
866 if (!cfqq && !list_empty(&cfqd->idle_rr)) {
867 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
868
869 if (time_after_eq(jiffies, end))
870 cfqq = list_entry_cfqq(cfqd->idle_rr.next);
871 else
872 mod_timer(&cfqd->idle_class_timer, end);
873 }
874
875 __cfq_set_active_queue(cfqd, cfqq);
876 return cfqq;
877}
878
879/*
880 * current cfqq expired its slice (or was too idle), select new one
881 */
882static void
883__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
884 int preempted)
885{
886 unsigned long now = jiffies;
887
888 if (cfq_cfqq_wait_request(cfqq))
889 del_timer(&cfqd->idle_slice_timer);
890
891 if (!preempted && !cfq_cfqq_dispatched(cfqq))
892 cfqq->service_last = now;
893
894 cfq_clear_cfqq_must_dispatch(cfqq);
895 cfq_clear_cfqq_wait_request(cfqq);
896
897 /*
898 * store what was left of this slice, if the queue idled out
899 * or was preempted
900 */
901 if (time_after(now, cfqq->slice_end))
902 cfqq->slice_left = now - cfqq->slice_end;
903 else
904 cfqq->slice_left = 0;
905
906 if (cfq_cfqq_on_rr(cfqq))
907 cfq_resort_rr_list(cfqq, preempted);
908
909 if (cfqq == cfqd->active_queue)
910 cfqd->active_queue = NULL;
911
912 if (cfqd->active_cic) {
913 put_io_context(cfqd->active_cic->ioc);
914 cfqd->active_cic = NULL;
915 }
916
917 cfqd->dispatch_slice = 0;
918}
919
920static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
921{
922 struct cfq_queue *cfqq = cfqd->active_queue;
923
924 if (cfqq) {
925 /*
926 * use deferred expiry, if there are requests in progress as
927 * not to disturb the slice of the next queue
928 */
929 if (cfq_cfqq_dispatched(cfqq))
930 cfq_mark_cfqq_expired(cfqq);
931 else
932 __cfq_slice_expired(cfqd, cfqq, preempted);
933 }
934}
935
936static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
937
938{
939 WARN_ON(!RB_EMPTY(&cfqq->sort_list));
940 WARN_ON(cfqq != cfqd->active_queue);
941
942 /*
943 * idle is disabled, either manually or by past process history
944 */
945 if (!cfqd->cfq_slice_idle)
946 return 0;
947 if (!cfq_cfqq_idle_window(cfqq))
948 return 0;
949 /*
950 * task has exited, don't wait
951 */
952 if (cfqd->active_cic && !cfqd->active_cic->ioc->task)
953 return 0;
954
955 cfq_mark_cfqq_must_dispatch(cfqq);
956 cfq_mark_cfqq_wait_request(cfqq);
957
958 if (!timer_pending(&cfqd->idle_slice_timer)) {
959 unsigned long slice_left = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
960
961 cfqd->idle_slice_timer.expires = jiffies + slice_left;
962 add_timer(&cfqd->idle_slice_timer);
963 }
964
965 return 1;
726} 966}
727 967
728/* 968/*
@@ -738,31 +978,40 @@ static void cfq_dispatch_sort(request_queue_t *q, struct cfq_rq *crq)
738 struct request *__rq; 978 struct request *__rq;
739 sector_t last; 979 sector_t last;
740 980
741 cfq_del_crq_rb(crq);
742 cfq_remove_merge_hints(q, crq);
743 list_del(&crq->request->queuelist); 981 list_del(&crq->request->queuelist);
744 982
745 last = cfqd->last_sector; 983 last = cfqd->last_sector;
746 while ((entry = entry->prev) != head) { 984 list_for_each_entry_reverse(__rq, head, queuelist) {
747 __rq = list_entry_rq(entry); 985 struct cfq_rq *__crq = RQ_DATA(__rq);
748 986
749 if (blk_barrier_rq(crq->request)) 987 if (blk_barrier_rq(__rq))
988 break;
989 if (!blk_fs_request(__rq))
750 break; 990 break;
751 if (!blk_fs_request(crq->request)) 991 if (cfq_crq_requeued(__crq))
752 break; 992 break;
753 993
754 if (crq->request->sector > __rq->sector) 994 if (__rq->sector <= crq->request->sector)
755 break; 995 break;
756 if (__rq->sector > last && crq->request->sector < last) { 996 if (__rq->sector > last && crq->request->sector < last) {
757 last = crq->request->sector; 997 last = crq->request->sector + crq->request->nr_sectors;
758 break; 998 break;
759 } 999 }
1000 entry = &__rq->queuelist;
760 } 1001 }
761 1002
762 cfqd->last_sector = last; 1003 cfqd->last_sector = last;
763 crq->in_flight = 1; 1004
764 cfqq->in_flight++; 1005 cfqq->next_crq = cfq_find_next_crq(cfqd, cfqq, crq);
765 list_add(&crq->request->queuelist, entry); 1006
1007 cfq_del_crq_rb(crq);
1008 cfq_remove_merge_hints(q, crq);
1009
1010 cfq_mark_crq_in_flight(crq);
1011 cfq_clear_crq_requeued(crq);
1012
1013 cfqq->on_dispatch[cfq_crq_is_sync(crq)]++;
1014 list_add_tail(&crq->request->queuelist, entry);
766} 1015}
767 1016
768/* 1017/*
@@ -771,173 +1020,235 @@ 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) 1020static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
772{ 1021{
773 struct cfq_data *cfqd = cfqq->cfqd; 1022 struct cfq_data *cfqd = cfqq->cfqd;
774 const int reads = !list_empty(&cfqq->fifo[0]); 1023 struct request *rq;
775 const int writes = !list_empty(&cfqq->fifo[1]);
776 unsigned long now = jiffies;
777 struct cfq_rq *crq; 1024 struct cfq_rq *crq;
778 1025
779 if (time_before(now, cfqq->last_fifo_expire + cfqd->cfq_fifo_batch_expire)) 1026 if (cfq_cfqq_fifo_expire(cfqq))
780 return NULL; 1027 return NULL;
781 1028
782 crq = RQ_DATA(list_entry(cfqq->fifo[0].next, struct request, queuelist)); 1029 if (!list_empty(&cfqq->fifo)) {
783 if (reads && time_after(now, crq->queue_start + cfqd->cfq_fifo_expire_r)) { 1030 int fifo = cfq_cfqq_class_sync(cfqq);
784 cfqq->last_fifo_expire = now;
785 return crq;
786 }
787 1031
788 crq = RQ_DATA(list_entry(cfqq->fifo[1].next, struct request, queuelist)); 1032 crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next));
789 if (writes && time_after(now, crq->queue_start + cfqd->cfq_fifo_expire_w)) { 1033 rq = crq->request;
790 cfqq->last_fifo_expire = now; 1034 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
791 return crq; 1035 cfq_mark_cfqq_fifo_expire(cfqq);
1036 return crq;
1037 }
792 } 1038 }
793 1039
794 return NULL; 1040 return NULL;
795} 1041}
796 1042
797/* 1043/*
798 * dispatch a single request from given queue 1044 * Scale schedule slice based on io priority. Use the sync time slice only
1045 * if a queue is marked sync and has sync io queued. A sync queue with async
1046 * io only, should not get full sync slice length.
799 */ 1047 */
1048static inline int
1049cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1050{
1051 const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
1052
1053 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1054
1055 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
1056}
1057
800static inline void 1058static inline void
801cfq_dispatch_request(request_queue_t *q, struct cfq_data *cfqd, 1059cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
802 struct cfq_queue *cfqq)
803{ 1060{
804 struct cfq_rq *crq; 1061 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
1062}
1063
1064static inline int
1065cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1066{
1067 const int base_rq = cfqd->cfq_slice_async_rq;
1068
1069 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1070
1071 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1072}
1073
1074/*
1075 * scheduler run of queue, if there are requests pending and no one in the
1076 * driver that will restart queueing
1077 */
1078static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
1079{
1080 if (!cfqd->rq_in_driver && cfq_pending_requests(cfqd))
1081 kblockd_schedule_work(&cfqd->unplug_work);
1082}
1083
1084/*
1085 * get next queue for service
1086 */
1087static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd, int force)
1088{
1089 unsigned long now = jiffies;
1090 struct cfq_queue *cfqq;
1091
1092 cfqq = cfqd->active_queue;
1093 if (!cfqq)
1094 goto new_queue;
1095
1096 if (cfq_cfqq_expired(cfqq))
1097 goto new_queue;
805 1098
806 /* 1099 /*
807 * follow expired path, else get first next available 1100 * slice has expired
808 */ 1101 */
809 if ((crq = cfq_check_fifo(cfqq)) == NULL) { 1102 if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
810 if (cfqd->find_best_crq) 1103 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 1104
818 /* 1105 /*
819 * finally, insert request into driver list 1106 * if queue has requests, dispatch one. if not, check if
1107 * enough slice is left to wait for one
820 */ 1108 */
821 cfq_dispatch_sort(q, crq); 1109 if (!RB_EMPTY(&cfqq->sort_list))
1110 goto keep_queue;
1111 else if (!force && cfq_cfqq_class_sync(cfqq) &&
1112 time_before(now, cfqq->slice_end)) {
1113 if (cfq_arm_slice_timer(cfqd, cfqq))
1114 return NULL;
1115 }
1116
1117expire:
1118 cfq_slice_expired(cfqd, 0);
1119new_queue:
1120 cfqq = cfq_set_active_queue(cfqd);
1121keep_queue:
1122 return cfqq;
822} 1123}
823 1124
824static int cfq_dispatch_requests(request_queue_t *q, int max_dispatch) 1125static int
1126__cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1127 int max_dispatch)
825{ 1128{
826 struct cfq_data *cfqd = q->elevator->elevator_data; 1129 int dispatched = 0;
827 struct cfq_queue *cfqq;
828 struct list_head *entry, *tmp;
829 int queued, busy_queues, first_round;
830 1130
831 if (list_empty(&cfqd->rr_list)) 1131 BUG_ON(RB_EMPTY(&cfqq->sort_list));
832 return 0;
833 1132
834 queued = 0; 1133 do {
835 first_round = 1; 1134 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 1135
841 BUG_ON(RB_EMPTY(&cfqq->sort_list)); 1136 /*
1137 * follow expired path, else get first next available
1138 */
1139 if ((crq = cfq_check_fifo(cfqq)) == NULL)
1140 crq = cfqq->next_crq;
842 1141
843 /* 1142 /*
844 * first round of queueing, only select from queues that 1143 * finally, insert request into driver dispatch list
845 * don't already have io in-flight
846 */ 1144 */
847 if (first_round && cfqq->in_flight) 1145 cfq_dispatch_sort(cfqd->queue, crq);
848 continue; 1146
1147 cfqd->dispatch_slice++;
1148 dispatched++;
849 1149
850 cfq_dispatch_request(q, cfqd, cfqq); 1150 if (!cfqd->active_cic) {
1151 atomic_inc(&crq->io_context->ioc->refcount);
1152 cfqd->active_cic = crq->io_context;
1153 }
851 1154
852 if (!RB_EMPTY(&cfqq->sort_list)) 1155 if (RB_EMPTY(&cfqq->sort_list))
853 busy_queues++; 1156 break;
854 1157
855 queued++; 1158 } while (dispatched < max_dispatch);
856 } 1159
1160 /*
1161 * if slice end isn't set yet, set it. if at least one request was
1162 * sync, use the sync time slice value
1163 */
1164 if (!cfqq->slice_end)
1165 cfq_set_prio_slice(cfqd, cfqq);
1166
1167 /*
1168 * expire an async queue immediately if it has used up its slice. idle
1169 * queue always expire after 1 dispatch round.
1170 */
1171 if ((!cfq_cfqq_sync(cfqq) &&
1172 cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1173 cfq_class_idle(cfqq))
1174 cfq_slice_expired(cfqd, 0);
1175
1176 return dispatched;
1177}
1178
1179static int
1180cfq_dispatch_requests(request_queue_t *q, int max_dispatch, int force)
1181{
1182 struct cfq_data *cfqd = q->elevator->elevator_data;
1183 struct cfq_queue *cfqq;
1184
1185 if (!cfqd->busy_queues)
1186 return 0;
857 1187
858 if ((queued < max_dispatch) && (busy_queues || first_round)) { 1188 cfqq = cfq_select_queue(cfqd, force);
859 first_round = 0; 1189 if (cfqq) {
860 goto restart; 1190 cfq_clear_cfqq_must_dispatch(cfqq);
1191 cfq_clear_cfqq_wait_request(cfqq);
1192 del_timer(&cfqd->idle_slice_timer);
1193
1194 if (cfq_class_idle(cfqq))
1195 max_dispatch = 1;
1196
1197 return __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
861 } 1198 }
862 1199
863 return queued; 1200 return 0;
864} 1201}
865 1202
866static inline void cfq_account_dispatch(struct cfq_rq *crq) 1203static inline void cfq_account_dispatch(struct cfq_rq *crq)
867{ 1204{
868 struct cfq_queue *cfqq = crq->cfq_queue; 1205 struct cfq_queue *cfqq = crq->cfq_queue;
869 struct cfq_data *cfqd = cfqq->cfqd; 1206 struct cfq_data *cfqd = cfqq->cfqd;
870 unsigned long now, elapsed;
871 1207
872 if (!blk_fs_request(crq->request)) 1208 if (unlikely(!blk_fs_request(crq->request)))
873 return; 1209 return;
874 1210
875 /* 1211 /*
876 * accounted bit is necessary since some drivers will call 1212 * accounted bit is necessary since some drivers will call
877 * elv_next_request() many times for the same request (eg ide) 1213 * elv_next_request() many times for the same request (eg ide)
878 */ 1214 */
879 if (crq->accounted) 1215 if (cfq_crq_in_driver(crq))
880 return; 1216 return;
881 1217
882 now = jiffies; 1218 cfq_mark_crq_in_driver(crq);
883 if (cfqq->service_start == ~0UL) 1219 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} 1220}
914 1221
915static inline void 1222static inline void
916cfq_account_completion(struct cfq_queue *cfqq, struct cfq_rq *crq) 1223cfq_account_completion(struct cfq_queue *cfqq, struct cfq_rq *crq)
917{ 1224{
918 struct cfq_data *cfqd = cfqq->cfqd; 1225 struct cfq_data *cfqd = cfqq->cfqd;
1226 unsigned long now;
919 1227
920 if (!crq->accounted) 1228 if (!cfq_crq_in_driver(crq))
921 return; 1229 return;
922 1230
1231 now = jiffies;
1232
923 WARN_ON(!cfqd->rq_in_driver); 1233 WARN_ON(!cfqd->rq_in_driver);
924 cfqd->rq_in_driver--; 1234 cfqd->rq_in_driver--;
925 1235
926 if (!cfqd->cfq_tagged) { 1236 if (!cfq_class_idle(cfqq))
927 unsigned long now = jiffies; 1237 cfqd->last_end_request = now;
928 unsigned long duration = now - crq->service_start;
929 1238
930 if (time_after(now, cfqq->service_start + cfq_service)) { 1239 if (!cfq_cfqq_dispatched(cfqq)) {
931 cfqq->service_start = now; 1240 if (cfq_cfqq_on_rr(cfqq)) {
932 cfqq->service_used >>= 3; 1241 cfqq->service_last = now;
1242 cfq_resort_rr_list(cfqq, 0);
1243 }
1244 if (cfq_cfqq_expired(cfqq)) {
1245 __cfq_slice_expired(cfqd, cfqq, 0);
1246 cfq_schedule_dispatch(cfqd);
933 } 1247 }
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 } 1248 }
1249
1250 if (cfq_crq_is_sync(crq))
1251 crq->io_context->last_end_request = now;
941} 1252}
942 1253
943static struct request *cfq_next_request(request_queue_t *q) 1254static struct request *cfq_next_request(request_queue_t *q)
@@ -950,7 +1261,18 @@ static struct request *cfq_next_request(request_queue_t *q)
950dispatch: 1261dispatch:
951 rq = list_entry_rq(q->queue_head.next); 1262 rq = list_entry_rq(q->queue_head.next);
952 1263
953 if ((crq = RQ_DATA(rq)) != NULL) { 1264 crq = RQ_DATA(rq);
1265 if (crq) {
1266 struct cfq_queue *cfqq = crq->cfq_queue;
1267
1268 /*
1269 * if idle window is disabled, allow queue buildup
1270 */
1271 if (!cfq_crq_in_driver(crq) &&
1272 !cfq_cfqq_idle_window(cfqq) &&
1273 cfqd->rq_in_driver >= cfqd->cfq_max_depth)
1274 return NULL;
1275
954 cfq_remove_merge_hints(q, crq); 1276 cfq_remove_merge_hints(q, crq);
955 cfq_account_dispatch(crq); 1277 cfq_account_dispatch(crq);
956 } 1278 }
@@ -958,7 +1280,7 @@ dispatch:
958 return rq; 1280 return rq;
959 } 1281 }
960 1282
961 if (cfq_dispatch_requests(q, cfqd->cfq_quantum)) 1283 if (cfq_dispatch_requests(q, cfqd->cfq_quantum, 0))
962 goto dispatch; 1284 goto dispatch;
963 1285
964 return NULL; 1286 return NULL;
@@ -972,13 +1294,21 @@ dispatch:
972 */ 1294 */
973static void cfq_put_queue(struct cfq_queue *cfqq) 1295static void cfq_put_queue(struct cfq_queue *cfqq)
974{ 1296{
975 BUG_ON(!atomic_read(&cfqq->ref)); 1297 struct cfq_data *cfqd = cfqq->cfqd;
1298
1299 BUG_ON(atomic_read(&cfqq->ref) <= 0);
976 1300
977 if (!atomic_dec_and_test(&cfqq->ref)) 1301 if (!atomic_dec_and_test(&cfqq->ref))
978 return; 1302 return;
979 1303
980 BUG_ON(rb_first(&cfqq->sort_list)); 1304 BUG_ON(rb_first(&cfqq->sort_list));
981 BUG_ON(cfqq->on_rr); 1305 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1306 BUG_ON(cfq_cfqq_on_rr(cfqq));
1307
1308 if (unlikely(cfqd->active_queue == cfqq)) {
1309 __cfq_slice_expired(cfqd, cfqq, 0);
1310 cfq_schedule_dispatch(cfqd);
1311 }
982 1312
983 cfq_put_cfqd(cfqq->cfqd); 1313 cfq_put_cfqd(cfqq->cfqd);
984 1314
@@ -991,15 +1321,17 @@ static void cfq_put_queue(struct cfq_queue *cfqq)
991} 1321}
992 1322
993static inline struct cfq_queue * 1323static inline struct cfq_queue *
994__cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned long key, const int hashval) 1324__cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1325 const int hashval)
995{ 1326{
996 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval]; 1327 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
997 struct hlist_node *entry, *next; 1328 struct hlist_node *entry, *next;
998 1329
999 hlist_for_each_safe(entry, next, hash_list) { 1330 hlist_for_each_safe(entry, next, hash_list) {
1000 struct cfq_queue *__cfqq = list_entry_qhash(entry); 1331 struct cfq_queue *__cfqq = list_entry_qhash(entry);
1332 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->ioprio_class, __cfqq->ioprio);
1001 1333
1002 if (__cfqq->key == key) 1334 if (__cfqq->key == key && (__p == prio || prio == CFQ_KEY_ANY))
1003 return __cfqq; 1335 return __cfqq;
1004 } 1336 }
1005 1337
@@ -1007,94 +1339,220 @@ __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned long key, const int hashval)
1007} 1339}
1008 1340
1009static struct cfq_queue * 1341static struct cfq_queue *
1010cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned long key) 1342cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1011{ 1343{
1012 return __cfq_find_cfq_hash(cfqd, key, hash_long(key, CFQ_QHASH_SHIFT)); 1344 return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1013} 1345}
1014 1346
1015static inline void 1347static 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{ 1348{
1019 unsigned long hashkey = cfq_hash_key(cfqd, current); 1349 struct cfq_io_context *__cic;
1020 unsigned long hashval = hash_long(hashkey, CFQ_QHASH_SHIFT); 1350 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 1351
1026 hlist_del(&(*cfqq)->cfq_hash); 1352 list_for_each_safe(entry, next, &cic->list) {
1027 1353 __cic = list_entry(entry, struct cfq_io_context, list);
1028 __cfqq = __cfq_find_cfq_hash(cfqd, hashkey, hashval); 1354 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 } 1355 }
1039 1356
1040 cic->cfqq = __cfqq; 1357 kmem_cache_free(cfq_ioc_pool, cic);
1041 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1042} 1358}
1043 1359
1044static void cfq_free_io_context(struct cfq_io_context *cic) 1360/*
1361 * Called with interrupts disabled
1362 */
1363static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1045{ 1364{
1046 kmem_cache_free(cfq_ioc_pool, cic); 1365 struct cfq_data *cfqd = cic->cfqq->cfqd;
1366 request_queue_t *q = cfqd->queue;
1367
1368 WARN_ON(!irqs_disabled());
1369
1370 spin_lock(q->queue_lock);
1371
1372 if (unlikely(cic->cfqq == cfqd->active_queue)) {
1373 __cfq_slice_expired(cfqd, cic->cfqq, 0);
1374 cfq_schedule_dispatch(cfqd);
1375 }
1376
1377 cfq_put_queue(cic->cfqq);
1378 cic->cfqq = NULL;
1379 spin_unlock(q->queue_lock);
1047} 1380}
1048 1381
1049/* 1382/*
1050 * locking hierarchy is: io_context lock -> queue locks 1383 * Another task may update the task cic list, if it is doing a queue lookup
1384 * on its behalf. cfq_cic_lock excludes such concurrent updates
1051 */ 1385 */
1052static void cfq_exit_io_context(struct cfq_io_context *cic) 1386static void cfq_exit_io_context(struct cfq_io_context *cic)
1053{ 1387{
1054 struct cfq_queue *cfqq = cic->cfqq; 1388 struct cfq_io_context *__cic;
1055 struct list_head *entry = &cic->list; 1389 struct list_head *entry;
1056 request_queue_t *q;
1057 unsigned long flags; 1390 unsigned long flags;
1058 1391
1392 local_irq_save(flags);
1393
1059 /* 1394 /*
1060 * put the reference this task is holding to the various queues 1395 * put the reference this task is holding to the various queues
1061 */ 1396 */
1062 spin_lock_irqsave(&cic->ioc->lock, flags); 1397 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); 1398 __cic = list_entry(entry, struct cfq_io_context, list);
1067 list_del(entry); 1399 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 } 1400 }
1074 1401
1075 q = cfqq->cfqd->queue; 1402 cfq_exit_single_io_context(cic);
1076 spin_lock(q->queue_lock); 1403 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} 1404}
1083 1405
1084static struct cfq_io_context *cfq_alloc_io_context(int gfp_flags) 1406static struct cfq_io_context *
1407cfq_alloc_io_context(struct cfq_data *cfqd, int gfp_mask)
1085{ 1408{
1086 struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_flags); 1409 struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1087 1410
1088 if (cic) { 1411 if (cic) {
1089 cic->dtor = cfq_free_io_context;
1090 cic->exit = cfq_exit_io_context;
1091 INIT_LIST_HEAD(&cic->list); 1412 INIT_LIST_HEAD(&cic->list);
1092 cic->cfqq = NULL; 1413 cic->cfqq = NULL;
1414 cic->key = NULL;
1415 cic->last_end_request = jiffies;
1416 cic->ttime_total = 0;
1417 cic->ttime_samples = 0;
1418 cic->ttime_mean = 0;
1419 cic->dtor = cfq_free_io_context;
1420 cic->exit = cfq_exit_io_context;
1093 } 1421 }
1094 1422
1095 return cic; 1423 return cic;
1096} 1424}
1097 1425
1426static void cfq_init_prio_data(struct cfq_queue *cfqq)
1427{
1428 struct task_struct *tsk = current;
1429 int ioprio_class;
1430
1431 if (!cfq_cfqq_prio_changed(cfqq))
1432 return;
1433
1434 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1435 switch (ioprio_class) {
1436 default:
1437 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1438 case IOPRIO_CLASS_NONE:
1439 /*
1440 * no prio set, place us in the middle of the BE classes
1441 */
1442 cfqq->ioprio = task_nice_ioprio(tsk);
1443 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1444 break;
1445 case IOPRIO_CLASS_RT:
1446 cfqq->ioprio = task_ioprio(tsk);
1447 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1448 break;
1449 case IOPRIO_CLASS_BE:
1450 cfqq->ioprio = task_ioprio(tsk);
1451 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1452 break;
1453 case IOPRIO_CLASS_IDLE:
1454 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1455 cfqq->ioprio = 7;
1456 cfq_clear_cfqq_idle_window(cfqq);
1457 break;
1458 }
1459
1460 /*
1461 * keep track of original prio settings in case we have to temporarily
1462 * elevate the priority of this queue
1463 */
1464 cfqq->org_ioprio = cfqq->ioprio;
1465 cfqq->org_ioprio_class = cfqq->ioprio_class;
1466
1467 if (cfq_cfqq_on_rr(cfqq))
1468 cfq_resort_rr_list(cfqq, 0);
1469
1470 cfq_clear_cfqq_prio_changed(cfqq);
1471}
1472
1473static inline void changed_ioprio(struct cfq_queue *cfqq)
1474{
1475 if (cfqq) {
1476 struct cfq_data *cfqd = cfqq->cfqd;
1477
1478 spin_lock(cfqd->queue->queue_lock);
1479 cfq_mark_cfqq_prio_changed(cfqq);
1480 cfq_init_prio_data(cfqq);
1481 spin_unlock(cfqd->queue->queue_lock);
1482 }
1483}
1484
1485/*
1486 * callback from sys_ioprio_set, irqs are disabled
1487 */
1488static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
1489{
1490 struct cfq_io_context *cic = ioc->cic;
1491
1492 changed_ioprio(cic->cfqq);
1493
1494 list_for_each_entry(cic, &cic->list, list)
1495 changed_ioprio(cic->cfqq);
1496
1497 return 0;
1498}
1499
1500static struct cfq_queue *
1501cfq_get_queue(struct cfq_data *cfqd, unsigned int key, unsigned short ioprio,
1502 int gfp_mask)
1503{
1504 const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1505 struct cfq_queue *cfqq, *new_cfqq = NULL;
1506
1507retry:
1508 cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1509
1510 if (!cfqq) {
1511 if (new_cfqq) {
1512 cfqq = new_cfqq;
1513 new_cfqq = NULL;
1514 } else if (gfp_mask & __GFP_WAIT) {
1515 spin_unlock_irq(cfqd->queue->queue_lock);
1516 new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1517 spin_lock_irq(cfqd->queue->queue_lock);
1518 goto retry;
1519 } else {
1520 cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1521 if (!cfqq)
1522 goto out;
1523 }
1524
1525 memset(cfqq, 0, sizeof(*cfqq));
1526
1527 INIT_HLIST_NODE(&cfqq->cfq_hash);
1528 INIT_LIST_HEAD(&cfqq->cfq_list);
1529 RB_CLEAR_ROOT(&cfqq->sort_list);
1530 INIT_LIST_HEAD(&cfqq->fifo);
1531
1532 cfqq->key = key;
1533 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1534 atomic_set(&cfqq->ref, 0);
1535 cfqq->cfqd = cfqd;
1536 atomic_inc(&cfqd->ref);
1537 cfqq->service_last = 0;
1538 /*
1539 * set ->slice_left to allow preemption for a new process
1540 */
1541 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1542 cfq_mark_cfqq_idle_window(cfqq);
1543 cfq_mark_cfqq_prio_changed(cfqq);
1544 cfq_init_prio_data(cfqq);
1545 }
1546
1547 if (new_cfqq)
1548 kmem_cache_free(cfq_pool, new_cfqq);
1549
1550 atomic_inc(&cfqq->ref);
1551out:
1552 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1553 return cfqq;
1554}
1555
1098/* 1556/*
1099 * Setup general io context and cfq io context. There can be several cfq 1557 * 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 1558 * io contexts per general io context, if this process is doing io to more
@@ -1102,39 +1560,39 @@ static struct cfq_io_context *cfq_alloc_io_context(int gfp_flags)
1102 * cfqq, so we don't need to worry about it disappearing 1560 * cfqq, so we don't need to worry about it disappearing
1103 */ 1561 */
1104static struct cfq_io_context * 1562static struct cfq_io_context *
1105cfq_get_io_context(struct cfq_queue **cfqq, int gfp_flags) 1563cfq_get_io_context(struct cfq_data *cfqd, pid_t pid, int gfp_mask)
1106{ 1564{
1107 struct cfq_data *cfqd = (*cfqq)->cfqd; 1565 struct io_context *ioc = NULL;
1108 struct cfq_queue *__cfqq = *cfqq;
1109 struct cfq_io_context *cic; 1566 struct cfq_io_context *cic;
1110 struct io_context *ioc;
1111 1567
1112 might_sleep_if(gfp_flags & __GFP_WAIT); 1568 might_sleep_if(gfp_mask & __GFP_WAIT);
1113 1569
1114 ioc = get_io_context(gfp_flags); 1570 ioc = get_io_context(gfp_mask);
1115 if (!ioc) 1571 if (!ioc)
1116 return NULL; 1572 return NULL;
1117 1573
1118 if ((cic = ioc->cic) == NULL) { 1574 if ((cic = ioc->cic) == NULL) {
1119 cic = cfq_alloc_io_context(gfp_flags); 1575 cic = cfq_alloc_io_context(cfqd, gfp_mask);
1120 1576
1121 if (cic == NULL) 1577 if (cic == NULL)
1122 goto err; 1578 goto err;
1123 1579
1580 /*
1581 * manually increment generic io_context usage count, it
1582 * cannot go away since we are already holding one ref to it
1583 */
1124 ioc->cic = cic; 1584 ioc->cic = cic;
1585 ioc->set_ioprio = cfq_ioc_set_ioprio;
1125 cic->ioc = ioc; 1586 cic->ioc = ioc;
1126 cic->cfqq = __cfqq; 1587 cic->key = cfqd;
1127 atomic_inc(&__cfqq->ref); 1588 atomic_inc(&cfqd->ref);
1128 } else { 1589 } else {
1129 struct cfq_io_context *__cic; 1590 struct cfq_io_context *__cic;
1130 unsigned long flags;
1131 1591
1132 /* 1592 /*
1133 * since the first cic on the list is actually the head 1593 * 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 */ 1594 */
1137 if (cic->cfqq == __cfqq) 1595 if (cic->key == cfqd)
1138 goto out; 1596 goto out;
1139 1597
1140 /* 1598 /*
@@ -1142,152 +1600,262 @@ cfq_get_io_context(struct cfq_queue **cfqq, int gfp_flags)
1142 * should be ok here, the list will usually not be more than 1600 * should be ok here, the list will usually not be more than
1143 * 1 or a few entries long 1601 * 1 or a few entries long
1144 */ 1602 */
1145 spin_lock_irqsave(&ioc->lock, flags);
1146 list_for_each_entry(__cic, &cic->list, list) { 1603 list_for_each_entry(__cic, &cic->list, list) {
1147 /* 1604 /*
1148 * this process is already holding a reference to 1605 * this process is already holding a reference to
1149 * this queue, so no need to get one more 1606 * this queue, so no need to get one more
1150 */ 1607 */
1151 if (__cic->cfqq == __cfqq) { 1608 if (__cic->key == cfqd) {
1152 cic = __cic; 1609 cic = __cic;
1153 spin_unlock_irqrestore(&ioc->lock, flags);
1154 goto out; 1610 goto out;
1155 } 1611 }
1156 } 1612 }
1157 spin_unlock_irqrestore(&ioc->lock, flags);
1158 1613
1159 /* 1614 /*
1160 * nope, process doesn't have a cic assoicated with this 1615 * nope, process doesn't have a cic assoicated with this
1161 * cfqq yet. get a new one and add to list 1616 * cfqq yet. get a new one and add to list
1162 */ 1617 */
1163 __cic = cfq_alloc_io_context(gfp_flags); 1618 __cic = cfq_alloc_io_context(cfqd, gfp_mask);
1164 if (__cic == NULL) 1619 if (__cic == NULL)
1165 goto err; 1620 goto err;
1166 1621
1167 __cic->ioc = ioc; 1622 __cic->ioc = ioc;
1168 __cic->cfqq = __cfqq; 1623 __cic->key = cfqd;
1169 atomic_inc(&__cfqq->ref); 1624 atomic_inc(&cfqd->ref);
1170 spin_lock_irqsave(&ioc->lock, flags);
1171 list_add(&__cic->list, &cic->list); 1625 list_add(&__cic->list, &cic->list);
1172 spin_unlock_irqrestore(&ioc->lock, flags);
1173
1174 cic = __cic; 1626 cic = __cic;
1175 *cfqq = __cfqq;
1176 } 1627 }
1177 1628
1178out: 1629out:
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; 1630 return cic;
1187err: 1631err:
1188 put_io_context(ioc); 1632 put_io_context(ioc);
1189 return NULL; 1633 return NULL;
1190} 1634}
1191 1635
1192static struct cfq_queue * 1636static void
1193__cfq_get_queue(struct cfq_data *cfqd, unsigned long key, int gfp_mask) 1637cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1194{ 1638{
1195 const int hashval = hash_long(key, CFQ_QHASH_SHIFT); 1639 unsigned long elapsed, ttime;
1196 struct cfq_queue *cfqq, *new_cfqq = NULL;
1197 1640
1198retry: 1641 /*
1199 cfqq = __cfq_find_cfq_hash(cfqd, key, hashval); 1642 * if this context already has stuff queued, thinktime is from
1200 1643 * last queue not last end
1201 if (!cfqq) { 1644 */
1202 if (new_cfqq) { 1645#if 0
1203 cfqq = new_cfqq; 1646 if (time_after(cic->last_end_request, cic->last_queue))
1204 new_cfqq = NULL; 1647 elapsed = jiffies - cic->last_end_request;
1205 } else { 1648 else
1206 spin_unlock_irq(cfqd->queue->queue_lock); 1649 elapsed = jiffies - cic->last_queue;
1207 new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask); 1650#else
1208 spin_lock_irq(cfqd->queue->queue_lock); 1651 elapsed = jiffies - cic->last_end_request;
1652#endif
1209 1653
1210 if (!new_cfqq && !(gfp_mask & __GFP_WAIT)) 1654 ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1211 goto out;
1212 1655
1213 goto retry; 1656 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1214 } 1657 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1658 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1659}
1215 1660
1216 memset(cfqq, 0, sizeof(*cfqq)); 1661#define sample_valid(samples) ((samples) > 80)
1217 1662
1218 INIT_HLIST_NODE(&cfqq->cfq_hash); 1663/*
1219 INIT_LIST_HEAD(&cfqq->cfq_list); 1664 * Disable idle window if the process thinks too long or seeks so much that
1220 RB_CLEAR_ROOT(&cfqq->sort_list); 1665 * it doesn't matter
1221 INIT_LIST_HEAD(&cfqq->fifo[0]); 1666 */
1222 INIT_LIST_HEAD(&cfqq->fifo[1]); 1667static void
1668cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1669 struct cfq_io_context *cic)
1670{
1671 int enable_idle = cfq_cfqq_idle_window(cfqq);
1223 1672
1224 cfqq->key = key; 1673 if (!cic->ioc->task || !cfqd->cfq_slice_idle)
1225 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]); 1674 enable_idle = 0;
1226 atomic_set(&cfqq->ref, 0); 1675 else if (sample_valid(cic->ttime_samples)) {
1227 cfqq->cfqd = cfqd; 1676 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1228 atomic_inc(&cfqd->ref); 1677 enable_idle = 0;
1229 cfqq->key_type = cfqd->key_type; 1678 else
1230 cfqq->service_start = ~0UL; 1679 enable_idle = 1;
1231 } 1680 }
1232 1681
1233 if (new_cfqq) 1682 if (enable_idle)
1234 kmem_cache_free(cfq_pool, new_cfqq); 1683 cfq_mark_cfqq_idle_window(cfqq);
1684 else
1685 cfq_clear_cfqq_idle_window(cfqq);
1686}
1235 1687
1236 atomic_inc(&cfqq->ref); 1688
1237out: 1689/*
1238 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq); 1690 * Check if new_cfqq should preempt the currently active queue. Return 0 for
1239 return cfqq; 1691 * no or if we aren't sure, a 1 will cause a preempt.
1692 */
1693static int
1694cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1695 struct cfq_rq *crq)
1696{
1697 struct cfq_queue *cfqq = cfqd->active_queue;
1698
1699 if (cfq_class_idle(new_cfqq))
1700 return 0;
1701
1702 if (!cfqq)
1703 return 1;
1704
1705 if (cfq_class_idle(cfqq))
1706 return 1;
1707 if (!cfq_cfqq_wait_request(new_cfqq))
1708 return 0;
1709 /*
1710 * if it doesn't have slice left, forget it
1711 */
1712 if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1713 return 0;
1714 if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq))
1715 return 1;
1716
1717 return 0;
1718}
1719
1720/*
1721 * cfqq preempts the active queue. if we allowed preempt with no slice left,
1722 * let it have half of its nominal slice.
1723 */
1724static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1725{
1726 struct cfq_queue *__cfqq, *next;
1727
1728 list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1729 cfq_resort_rr_list(__cfqq, 1);
1730
1731 if (!cfqq->slice_left)
1732 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1733
1734 cfqq->slice_end = cfqq->slice_left + jiffies;
1735 __cfq_slice_expired(cfqd, cfqq, 1);
1736 __cfq_set_active_queue(cfqd, cfqq);
1737}
1738
1739/*
1740 * should really be a ll_rw_blk.c helper
1741 */
1742static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1743{
1744 request_queue_t *q = cfqd->queue;
1745
1746 if (!blk_queue_plugged(q))
1747 q->request_fn(q);
1748 else
1749 __generic_unplug_device(q);
1240} 1750}
1241 1751
1242static void cfq_enqueue(struct cfq_data *cfqd, struct cfq_rq *crq) 1752/*
1753 * Called when a new fs request (crq) is added (to cfqq). Check if there's
1754 * something we should do about it
1755 */
1756static void
1757cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1758 struct cfq_rq *crq)
1243{ 1759{
1244 crq->is_sync = 0; 1760 const int sync = cfq_crq_is_sync(crq);
1245 if (rq_data_dir(crq->request) == READ || current->flags & PF_SYNCWRITE) 1761
1246 crq->is_sync = 1; 1762 cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
1763
1764 if (sync) {
1765 struct cfq_io_context *cic = crq->io_context;
1766
1767 cfq_update_io_thinktime(cfqd, cic);
1768 cfq_update_idle_window(cfqd, cfqq, cic);
1769
1770 cic->last_queue = jiffies;
1771 }
1772
1773 if (cfqq == cfqd->active_queue) {
1774 /*
1775 * if we are waiting for a request for this queue, let it rip
1776 * immediately and flag that we must not expire this queue
1777 * just now
1778 */
1779 if (cfq_cfqq_wait_request(cfqq)) {
1780 cfq_mark_cfqq_must_dispatch(cfqq);
1781 del_timer(&cfqd->idle_slice_timer);
1782 cfq_start_queueing(cfqd, cfqq);
1783 }
1784 } else if (cfq_should_preempt(cfqd, cfqq, crq)) {
1785 /*
1786 * not the active queue - expire current slice if it is
1787 * idle and has expired it's mean thinktime or this new queue
1788 * has some old slice time left and is of higher priority
1789 */
1790 cfq_preempt_queue(cfqd, cfqq);
1791 cfq_mark_cfqq_must_dispatch(cfqq);
1792 cfq_start_queueing(cfqd, cfqq);
1793 }
1794}
1795
1796static void cfq_enqueue(struct cfq_data *cfqd, struct request *rq)
1797{
1798 struct cfq_rq *crq = RQ_DATA(rq);
1799 struct cfq_queue *cfqq = crq->cfq_queue;
1800
1801 cfq_init_prio_data(cfqq);
1247 1802
1248 cfq_add_crq_rb(crq); 1803 cfq_add_crq_rb(crq);
1249 crq->queue_start = jiffies;
1250 1804
1251 list_add_tail(&crq->request->queuelist, &crq->cfq_queue->fifo[crq->is_sync]); 1805 list_add_tail(&rq->queuelist, &cfqq->fifo);
1806
1807 if (rq_mergeable(rq)) {
1808 cfq_add_crq_hash(cfqd, crq);
1809
1810 if (!cfqd->queue->last_merge)
1811 cfqd->queue->last_merge = rq;
1812 }
1813
1814 cfq_crq_enqueued(cfqd, cfqq, crq);
1252} 1815}
1253 1816
1254static void 1817static void
1255cfq_insert_request(request_queue_t *q, struct request *rq, int where) 1818cfq_insert_request(request_queue_t *q, struct request *rq, int where)
1256{ 1819{
1257 struct cfq_data *cfqd = q->elevator->elevator_data; 1820 struct cfq_data *cfqd = q->elevator->elevator_data;
1258 struct cfq_rq *crq = RQ_DATA(rq);
1259 1821
1260 switch (where) { 1822 switch (where) {
1261 case ELEVATOR_INSERT_BACK: 1823 case ELEVATOR_INSERT_BACK:
1262 while (cfq_dispatch_requests(q, cfqd->cfq_quantum)) 1824 while (cfq_dispatch_requests(q, INT_MAX, 1))
1263 ; 1825 ;
1264 list_add_tail(&rq->queuelist, &q->queue_head); 1826 list_add_tail(&rq->queuelist, &q->queue_head);
1827 /*
1828 * If we were idling with pending requests on
1829 * inactive cfqqs, force dispatching will
1830 * remove the idle timer and the queue won't
1831 * be kicked by __make_request() afterward.
1832 * Kick it here.
1833 */
1834 cfq_schedule_dispatch(cfqd);
1265 break; 1835 break;
1266 case ELEVATOR_INSERT_FRONT: 1836 case ELEVATOR_INSERT_FRONT:
1267 list_add(&rq->queuelist, &q->queue_head); 1837 list_add(&rq->queuelist, &q->queue_head);
1268 break; 1838 break;
1269 case ELEVATOR_INSERT_SORT: 1839 case ELEVATOR_INSERT_SORT:
1270 BUG_ON(!blk_fs_request(rq)); 1840 BUG_ON(!blk_fs_request(rq));
1271 cfq_enqueue(cfqd, crq); 1841 cfq_enqueue(cfqd, rq);
1272 break; 1842 break;
1273 default: 1843 default:
1274 printk("%s: bad insert point %d\n", __FUNCTION__,where); 1844 printk("%s: bad insert point %d\n", __FUNCTION__,where);
1275 return; 1845 return;
1276 } 1846 }
1847}
1277 1848
1278 if (rq_mergeable(rq)) { 1849static inline int cfq_pending_requests(struct cfq_data *cfqd)
1279 cfq_add_crq_hash(cfqd, crq); 1850{
1280 1851 return !list_empty(&cfqd->queue->queue_head) || cfqd->busy_queues;
1281 if (!q->last_merge)
1282 q->last_merge = rq;
1283 }
1284} 1852}
1285 1853
1286static int cfq_queue_empty(request_queue_t *q) 1854static int cfq_queue_empty(request_queue_t *q)
1287{ 1855{
1288 struct cfq_data *cfqd = q->elevator->elevator_data; 1856 struct cfq_data *cfqd = q->elevator->elevator_data;
1289 1857
1290 return list_empty(&q->queue_head) && list_empty(&cfqd->rr_list); 1858 return !cfq_pending_requests(cfqd);
1291} 1859}
1292 1860
1293static void cfq_completed_request(request_queue_t *q, struct request *rq) 1861static void cfq_completed_request(request_queue_t *q, struct request *rq)
@@ -1300,9 +1868,11 @@ static void cfq_completed_request(request_queue_t *q, struct request *rq)
1300 1868
1301 cfqq = crq->cfq_queue; 1869 cfqq = crq->cfq_queue;
1302 1870
1303 if (crq->in_flight) { 1871 if (cfq_crq_in_flight(crq)) {
1304 WARN_ON(!cfqq->in_flight); 1872 const int sync = cfq_crq_is_sync(crq);
1305 cfqq->in_flight--; 1873
1874 WARN_ON(!cfqq->on_dispatch[sync]);
1875 cfqq->on_dispatch[sync]--;
1306 } 1876 }
1307 1877
1308 cfq_account_completion(cfqq, crq); 1878 cfq_account_completion(cfqq, crq);
@@ -1332,51 +1902,136 @@ cfq_latter_request(request_queue_t *q, struct request *rq)
1332 return NULL; 1902 return NULL;
1333} 1903}
1334 1904
1335static int cfq_may_queue(request_queue_t *q, int rw) 1905/*
1906 * we temporarily boost lower priority queues if they are holding fs exclusive
1907 * resources. they are boosted to normal prio (CLASS_BE/4)
1908 */
1909static void cfq_prio_boost(struct cfq_queue *cfqq)
1336{ 1910{
1337 struct cfq_data *cfqd = q->elevator->elevator_data; 1911 const int ioprio_class = cfqq->ioprio_class;
1338 struct cfq_queue *cfqq; 1912 const int ioprio = cfqq->ioprio;
1339 int ret = ELV_MQUEUE_MAY;
1340 1913
1341 if (current->flags & PF_MEMALLOC) 1914 if (has_fs_excl()) {
1342 return ELV_MQUEUE_MAY; 1915 /*
1916 * boost idle prio on transactions that would lock out other
1917 * users of the filesystem
1918 */
1919 if (cfq_class_idle(cfqq))
1920 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1921 if (cfqq->ioprio > IOPRIO_NORM)
1922 cfqq->ioprio = IOPRIO_NORM;
1923 } else {
1924 /*
1925 * check if we need to unboost the queue
1926 */
1927 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1928 cfqq->ioprio_class = cfqq->org_ioprio_class;
1929 if (cfqq->ioprio != cfqq->org_ioprio)
1930 cfqq->ioprio = cfqq->org_ioprio;
1931 }
1343 1932
1344 cfqq = cfq_find_cfq_hash(cfqd, cfq_hash_key(cfqd, current)); 1933 /*
1345 if (cfqq) { 1934 * refile between round-robin lists if we moved the priority class
1346 int limit = cfqd->max_queued; 1935 */
1936 if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1937 cfq_cfqq_on_rr(cfqq))
1938 cfq_resort_rr_list(cfqq, 0);
1939}
1347 1940
1348 if (cfqq->allocated[rw] < cfqd->cfq_queued) 1941static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
1349 return ELV_MQUEUE_MUST; 1942{
1943 if (rw == READ || process_sync(task))
1944 return task->pid;
1350 1945
1351 if (cfqd->busy_queues) 1946 return CFQ_KEY_ASYNC;
1352 limit = q->nr_requests / cfqd->busy_queues; 1947}
1353 1948
1354 if (limit < cfqd->cfq_queued) 1949static inline int
1355 limit = cfqd->cfq_queued; 1950__cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1356 else if (limit > cfqd->max_queued) 1951 struct task_struct *task, int rw)
1357 limit = cfqd->max_queued; 1952{
1953#if 1
1954 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1955 !cfq_cfqq_must_alloc_slice) {
1956 cfq_mark_cfqq_must_alloc_slice(cfqq);
1957 return ELV_MQUEUE_MUST;
1958 }
1358 1959
1359 if (cfqq->allocated[rw] >= limit) { 1960 return ELV_MQUEUE_MAY;
1360 if (limit > cfqq->alloc_limit[rw]) 1961#else
1361 cfqq->alloc_limit[rw] = limit; 1962 if (!cfqq || task->flags & PF_MEMALLOC)
1963 return ELV_MQUEUE_MAY;
1964 if (!cfqq->allocated[rw] || cfq_cfqq_must_alloc(cfqq)) {
1965 if (cfq_cfqq_wait_request(cfqq))
1966 return ELV_MQUEUE_MUST;
1362 1967
1363 ret = ELV_MQUEUE_NO; 1968 /*
1969 * only allow 1 ELV_MQUEUE_MUST per slice, otherwise we
1970 * can quickly flood the queue with writes from a single task
1971 */
1972 if (rw == READ || !cfq_cfqq_must_alloc_slice) {
1973 cfq_mark_cfqq_must_alloc_slice(cfqq);
1974 return ELV_MQUEUE_MUST;
1364 } 1975 }
1976
1977 return ELV_MQUEUE_MAY;
1365 } 1978 }
1979 if (cfq_class_idle(cfqq))
1980 return ELV_MQUEUE_NO;
1981 if (cfqq->allocated[rw] >= cfqd->max_queued) {
1982 struct io_context *ioc = get_io_context(GFP_ATOMIC);
1983 int ret = ELV_MQUEUE_NO;
1366 1984
1367 return ret; 1985 if (ioc && ioc->nr_batch_requests)
1986 ret = ELV_MQUEUE_MAY;
1987
1988 put_io_context(ioc);
1989 return ret;
1990 }
1991
1992 return ELV_MQUEUE_MAY;
1993#endif
1994}
1995
1996static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
1997{
1998 struct cfq_data *cfqd = q->elevator->elevator_data;
1999 struct task_struct *tsk = current;
2000 struct cfq_queue *cfqq;
2001
2002 /*
2003 * don't force setup of a queue from here, as a call to may_queue
2004 * does not necessarily imply that a request actually will be queued.
2005 * so just lookup a possibly existing queue, or return 'may queue'
2006 * if that fails
2007 */
2008 cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
2009 if (cfqq) {
2010 cfq_init_prio_data(cfqq);
2011 cfq_prio_boost(cfqq);
2012
2013 return __cfq_may_queue(cfqd, cfqq, tsk, rw);
2014 }
2015
2016 return ELV_MQUEUE_MAY;
1368} 2017}
1369 2018
1370static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq) 2019static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
1371{ 2020{
2021 struct cfq_data *cfqd = q->elevator->elevator_data;
1372 struct request_list *rl = &q->rq; 2022 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 2023
1376 if (read && cfqq->allocated[READ] < cfqq->alloc_limit[READ]) 2024 if (cfqq->allocated[READ] <= cfqd->max_queued || cfqd->rq_starved) {
1377 wake_up(&rl->wait[READ]); 2025 smp_mb();
1378 if (write && cfqq->allocated[WRITE] < cfqq->alloc_limit[WRITE]) 2026 if (waitqueue_active(&rl->wait[READ]))
1379 wake_up(&rl->wait[WRITE]); 2027 wake_up(&rl->wait[READ]);
2028 }
2029
2030 if (cfqq->allocated[WRITE] <= cfqd->max_queued || cfqd->rq_starved) {
2031 smp_mb();
2032 if (waitqueue_active(&rl->wait[WRITE]))
2033 wake_up(&rl->wait[WRITE]);
2034 }
1380} 2035}
1381 2036
1382/* 2037/*
@@ -1389,69 +2044,61 @@ static void cfq_put_request(request_queue_t *q, struct request *rq)
1389 2044
1390 if (crq) { 2045 if (crq) {
1391 struct cfq_queue *cfqq = crq->cfq_queue; 2046 struct cfq_queue *cfqq = crq->cfq_queue;
2047 const int rw = rq_data_dir(rq);
1392 2048
1393 BUG_ON(q->last_merge == rq); 2049 BUG_ON(!cfqq->allocated[rw]);
1394 BUG_ON(!hlist_unhashed(&crq->hash)); 2050 cfqq->allocated[rw]--;
1395 2051
1396 if (crq->io_context) 2052 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 2053
1402 mempool_free(crq, cfqd->crq_pool); 2054 mempool_free(crq, cfqd->crq_pool);
1403 rq->elevator_private = NULL; 2055 rq->elevator_private = NULL;
1404 2056
1405 smp_mb();
1406 cfq_check_waiters(q, cfqq); 2057 cfq_check_waiters(q, cfqq);
1407 cfq_put_queue(cfqq); 2058 cfq_put_queue(cfqq);
1408 } 2059 }
1409} 2060}
1410 2061
1411/* 2062/*
1412 * Allocate cfq data structures associated with this request. A queue and 2063 * Allocate cfq data structures associated with this request.
1413 */ 2064 */
1414static int cfq_set_request(request_queue_t *q, struct request *rq, int gfp_mask) 2065static int
2066cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
2067 int gfp_mask)
1415{ 2068{
1416 struct cfq_data *cfqd = q->elevator->elevator_data; 2069 struct cfq_data *cfqd = q->elevator->elevator_data;
2070 struct task_struct *tsk = current;
1417 struct cfq_io_context *cic; 2071 struct cfq_io_context *cic;
1418 const int rw = rq_data_dir(rq); 2072 const int rw = rq_data_dir(rq);
1419 struct cfq_queue *cfqq, *saved_cfqq; 2073 pid_t key = cfq_queue_pid(tsk, rw);
2074 struct cfq_queue *cfqq;
1420 struct cfq_rq *crq; 2075 struct cfq_rq *crq;
1421 unsigned long flags; 2076 unsigned long flags;
1422 2077
1423 might_sleep_if(gfp_mask & __GFP_WAIT); 2078 might_sleep_if(gfp_mask & __GFP_WAIT);
1424 2079
2080 cic = cfq_get_io_context(cfqd, key, gfp_mask);
2081
1425 spin_lock_irqsave(q->queue_lock, flags); 2082 spin_lock_irqsave(q->queue_lock, flags);
1426 2083
1427 cfqq = __cfq_get_queue(cfqd, cfq_hash_key(cfqd, current), gfp_mask); 2084 if (!cic)
1428 if (!cfqq) 2085 goto queue_fail;
1429 goto out_lock; 2086
2087 if (!cic->cfqq) {
2088 cfqq = cfq_get_queue(cfqd, key, tsk->ioprio, gfp_mask);
2089 if (!cfqq)
2090 goto queue_fail;
1430 2091
1431repeat: 2092 cic->cfqq = cfqq;
1432 if (cfqq->allocated[rw] >= cfqd->max_queued) 2093 } else
1433 goto out_lock; 2094 cfqq = cic->cfqq;
1434 2095
1435 cfqq->allocated[rw]++; 2096 cfqq->allocated[rw]++;
2097 cfq_clear_cfqq_must_alloc(cfqq);
2098 cfqd->rq_starved = 0;
2099 atomic_inc(&cfqq->ref);
1436 spin_unlock_irqrestore(q->queue_lock, flags); 2100 spin_unlock_irqrestore(q->queue_lock, flags);
1437 2101
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); 2102 crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
1456 if (crq) { 2103 if (crq) {
1457 RB_CLEAR(&crq->rb_node); 2104 RB_CLEAR(&crq->rb_node);
@@ -1460,24 +2107,141 @@ repeat:
1460 INIT_HLIST_NODE(&crq->hash); 2107 INIT_HLIST_NODE(&crq->hash);
1461 crq->cfq_queue = cfqq; 2108 crq->cfq_queue = cfqq;
1462 crq->io_context = cic; 2109 crq->io_context = cic;
1463 crq->service_start = crq->queue_start = 0; 2110 cfq_clear_crq_in_flight(crq);
1464 crq->in_flight = crq->accounted = crq->is_sync = 0; 2111 cfq_clear_crq_in_driver(crq);
1465 crq->is_write = rw; 2112 cfq_clear_crq_requeued(crq);
2113
2114 if (rw == READ || process_sync(tsk))
2115 cfq_mark_crq_is_sync(crq);
2116 else
2117 cfq_clear_crq_is_sync(crq);
2118
1466 rq->elevator_private = crq; 2119 rq->elevator_private = crq;
1467 cfqq->alloc_limit[rw] = 0;
1468 return 0; 2120 return 0;
1469 } 2121 }
1470 2122
1471 put_io_context(cic->ioc);
1472err:
1473 spin_lock_irqsave(q->queue_lock, flags); 2123 spin_lock_irqsave(q->queue_lock, flags);
1474 cfqq->allocated[rw]--; 2124 cfqq->allocated[rw]--;
2125 if (!(cfqq->allocated[0] + cfqq->allocated[1]))
2126 cfq_mark_cfqq_must_alloc(cfqq);
1475 cfq_put_queue(cfqq); 2127 cfq_put_queue(cfqq);
1476out_lock: 2128queue_fail:
2129 if (cic)
2130 put_io_context(cic->ioc);
2131 /*
2132 * mark us rq allocation starved. we need to kickstart the process
2133 * ourselves if there are no pending requests that can do it for us.
2134 * that would be an extremely rare OOM situation
2135 */
2136 cfqd->rq_starved = 1;
2137 cfq_schedule_dispatch(cfqd);
1477 spin_unlock_irqrestore(q->queue_lock, flags); 2138 spin_unlock_irqrestore(q->queue_lock, flags);
1478 return 1; 2139 return 1;
1479} 2140}
1480 2141
2142static void cfq_kick_queue(void *data)
2143{
2144 request_queue_t *q = data;
2145 struct cfq_data *cfqd = q->elevator->elevator_data;
2146 unsigned long flags;
2147
2148 spin_lock_irqsave(q->queue_lock, flags);
2149
2150 if (cfqd->rq_starved) {
2151 struct request_list *rl = &q->rq;
2152
2153 /*
2154 * we aren't guaranteed to get a request after this, but we
2155 * have to be opportunistic
2156 */
2157 smp_mb();
2158 if (waitqueue_active(&rl->wait[READ]))
2159 wake_up(&rl->wait[READ]);
2160 if (waitqueue_active(&rl->wait[WRITE]))
2161 wake_up(&rl->wait[WRITE]);
2162 }
2163
2164 blk_remove_plug(q);
2165 q->request_fn(q);
2166 spin_unlock_irqrestore(q->queue_lock, flags);
2167}
2168
2169/*
2170 * Timer running if the active_queue is currently idling inside its time slice
2171 */
2172static void cfq_idle_slice_timer(unsigned long data)
2173{
2174 struct cfq_data *cfqd = (struct cfq_data *) data;
2175 struct cfq_queue *cfqq;
2176 unsigned long flags;
2177
2178 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2179
2180 if ((cfqq = cfqd->active_queue) != NULL) {
2181 unsigned long now = jiffies;
2182
2183 /*
2184 * expired
2185 */
2186 if (time_after(now, cfqq->slice_end))
2187 goto expire;
2188
2189 /*
2190 * only expire and reinvoke request handler, if there are
2191 * other queues with pending requests
2192 */
2193 if (!cfq_pending_requests(cfqd)) {
2194 cfqd->idle_slice_timer.expires = min(now + cfqd->cfq_slice_idle, cfqq->slice_end);
2195 add_timer(&cfqd->idle_slice_timer);
2196 goto out_cont;
2197 }
2198
2199 /*
2200 * not expired and it has a request pending, let it dispatch
2201 */
2202 if (!RB_EMPTY(&cfqq->sort_list)) {
2203 cfq_mark_cfqq_must_dispatch(cfqq);
2204 goto out_kick;
2205 }
2206 }
2207expire:
2208 cfq_slice_expired(cfqd, 0);
2209out_kick:
2210 cfq_schedule_dispatch(cfqd);
2211out_cont:
2212 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2213}
2214
2215/*
2216 * Timer running if an idle class queue is waiting for service
2217 */
2218static void cfq_idle_class_timer(unsigned long data)
2219{
2220 struct cfq_data *cfqd = (struct cfq_data *) data;
2221 unsigned long flags, end;
2222
2223 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2224
2225 /*
2226 * race with a non-idle queue, reset timer
2227 */
2228 end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2229 if (!time_after_eq(jiffies, end)) {
2230 cfqd->idle_class_timer.expires = end;
2231 add_timer(&cfqd->idle_class_timer);
2232 } else
2233 cfq_schedule_dispatch(cfqd);
2234
2235 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2236}
2237
2238static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2239{
2240 del_timer_sync(&cfqd->idle_slice_timer);
2241 del_timer_sync(&cfqd->idle_class_timer);
2242 blk_sync_queue(cfqd->queue);
2243}
2244
1481static void cfq_put_cfqd(struct cfq_data *cfqd) 2245static void cfq_put_cfqd(struct cfq_data *cfqd)
1482{ 2246{
1483 request_queue_t *q = cfqd->queue; 2247 request_queue_t *q = cfqd->queue;
@@ -1487,6 +2251,9 @@ static void cfq_put_cfqd(struct cfq_data *cfqd)
1487 2251
1488 blk_put_queue(q); 2252 blk_put_queue(q);
1489 2253
2254 cfq_shutdown_timer_wq(cfqd);
2255 q->elevator->elevator_data = NULL;
2256
1490 mempool_destroy(cfqd->crq_pool); 2257 mempool_destroy(cfqd->crq_pool);
1491 kfree(cfqd->crq_hash); 2258 kfree(cfqd->crq_hash);
1492 kfree(cfqd->cfq_hash); 2259 kfree(cfqd->cfq_hash);
@@ -1495,7 +2262,10 @@ static void cfq_put_cfqd(struct cfq_data *cfqd)
1495 2262
1496static void cfq_exit_queue(elevator_t *e) 2263static void cfq_exit_queue(elevator_t *e)
1497{ 2264{
1498 cfq_put_cfqd(e->elevator_data); 2265 struct cfq_data *cfqd = e->elevator_data;
2266
2267 cfq_shutdown_timer_wq(cfqd);
2268 cfq_put_cfqd(cfqd);
1499} 2269}
1500 2270
1501static int cfq_init_queue(request_queue_t *q, elevator_t *e) 2271static int cfq_init_queue(request_queue_t *q, elevator_t *e)
@@ -1508,7 +2278,13 @@ static int cfq_init_queue(request_queue_t *q, elevator_t *e)
1508 return -ENOMEM; 2278 return -ENOMEM;
1509 2279
1510 memset(cfqd, 0, sizeof(*cfqd)); 2280 memset(cfqd, 0, sizeof(*cfqd));
1511 INIT_LIST_HEAD(&cfqd->rr_list); 2281
2282 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2283 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2284
2285 INIT_LIST_HEAD(&cfqd->busy_rr);
2286 INIT_LIST_HEAD(&cfqd->cur_rr);
2287 INIT_LIST_HEAD(&cfqd->idle_rr);
1512 INIT_LIST_HEAD(&cfqd->empty_list); 2288 INIT_LIST_HEAD(&cfqd->empty_list);
1513 2289
1514 cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL); 2290 cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL);
@@ -1533,24 +2309,32 @@ static int cfq_init_queue(request_queue_t *q, elevator_t *e)
1533 cfqd->queue = q; 2309 cfqd->queue = q;
1534 atomic_inc(&q->refcnt); 2310 atomic_inc(&q->refcnt);
1535 2311
1536 /* 2312 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; 2313 q->nr_batching = cfq_queued;
1543 cfqd->key_type = CFQ_KEY_TGID; 2314
1544 cfqd->find_best_crq = 1; 2315 init_timer(&cfqd->idle_slice_timer);
2316 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2317 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2318
2319 init_timer(&cfqd->idle_class_timer);
2320 cfqd->idle_class_timer.function = cfq_idle_class_timer;
2321 cfqd->idle_class_timer.data = (unsigned long) cfqd;
2322
2323 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2324
1545 atomic_set(&cfqd->ref, 1); 2325 atomic_set(&cfqd->ref, 1);
1546 2326
1547 cfqd->cfq_queued = cfq_queued; 2327 cfqd->cfq_queued = cfq_queued;
1548 cfqd->cfq_quantum = cfq_quantum; 2328 cfqd->cfq_quantum = cfq_quantum;
1549 cfqd->cfq_fifo_expire_r = cfq_fifo_expire_r; 2329 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
1550 cfqd->cfq_fifo_expire_w = cfq_fifo_expire_w; 2330 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; 2331 cfqd->cfq_back_max = cfq_back_max;
1553 cfqd->cfq_back_penalty = cfq_back_penalty; 2332 cfqd->cfq_back_penalty = cfq_back_penalty;
2333 cfqd->cfq_slice[0] = cfq_slice_async;
2334 cfqd->cfq_slice[1] = cfq_slice_sync;
2335 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2336 cfqd->cfq_slice_idle = cfq_slice_idle;
2337 cfqd->cfq_max_depth = cfq_max_depth;
1554 2338
1555 return 0; 2339 return 0;
1556out_crqpool: 2340out_crqpool:
@@ -1595,7 +2379,6 @@ fail:
1595 return -ENOMEM; 2379 return -ENOMEM;
1596} 2380}
1597 2381
1598
1599/* 2382/*
1600 * sysfs parts below --> 2383 * sysfs parts below -->
1601 */ 2384 */
@@ -1620,45 +2403,6 @@ cfq_var_store(unsigned int *var, const char *page, size_t count)
1620 return count; 2403 return count;
1621} 2404}
1622 2405
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) \ 2406#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
1663static ssize_t __FUNC(struct cfq_data *cfqd, char *page) \ 2407static ssize_t __FUNC(struct cfq_data *cfqd, char *page) \
1664{ \ 2408{ \
@@ -1669,12 +2413,15 @@ static ssize_t __FUNC(struct cfq_data *cfqd, char *page) \
1669} 2413}
1670SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); 2414SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
1671SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0); 2415SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
1672SHOW_FUNCTION(cfq_fifo_expire_r_show, cfqd->cfq_fifo_expire_r, 1); 2416SHOW_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); 2417SHOW_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); 2418SHOW_FUNCTION(cfq_back_max_show, cfqd->cfq_back_max, 0);
1677SHOW_FUNCTION(cfq_back_penalty_show, cfqd->cfq_back_penalty, 0); 2419SHOW_FUNCTION(cfq_back_penalty_show, cfqd->cfq_back_penalty, 0);
2420SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2421SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2422SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2423SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2424SHOW_FUNCTION(cfq_max_depth_show, cfqd->cfq_max_depth, 0);
1678#undef SHOW_FUNCTION 2425#undef SHOW_FUNCTION
1679 2426
1680#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ 2427#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
@@ -1694,12 +2441,15 @@ static ssize_t __FUNC(struct cfq_data *cfqd, const char *page, size_t count) \
1694} 2441}
1695STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); 2442STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
1696STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0); 2443STORE_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); 2444STORE_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); 2445STORE_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); 2446STORE_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); 2447STORE_FUNCTION(cfq_back_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2448STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2449STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2450STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2451STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2452STORE_FUNCTION(cfq_max_depth_store, &cfqd->cfq_max_depth, 1, UINT_MAX, 0);
1703#undef STORE_FUNCTION 2453#undef STORE_FUNCTION
1704 2454
1705static struct cfq_fs_entry cfq_quantum_entry = { 2455static struct cfq_fs_entry cfq_quantum_entry = {
@@ -1712,25 +2462,15 @@ static struct cfq_fs_entry cfq_queued_entry = {
1712 .show = cfq_queued_show, 2462 .show = cfq_queued_show,
1713 .store = cfq_queued_store, 2463 .store = cfq_queued_store,
1714}; 2464};
1715static struct cfq_fs_entry cfq_fifo_expire_r_entry = { 2465static struct cfq_fs_entry cfq_fifo_expire_sync_entry = {
1716 .attr = {.name = "fifo_expire_sync", .mode = S_IRUGO | S_IWUSR }, 2466 .attr = {.name = "fifo_expire_sync", .mode = S_IRUGO | S_IWUSR },
1717 .show = cfq_fifo_expire_r_show, 2467 .show = cfq_fifo_expire_sync_show,
1718 .store = cfq_fifo_expire_r_store, 2468 .store = cfq_fifo_expire_sync_store,
1719}; 2469};
1720static struct cfq_fs_entry cfq_fifo_expire_w_entry = { 2470static struct cfq_fs_entry cfq_fifo_expire_async_entry = {
1721 .attr = {.name = "fifo_expire_async", .mode = S_IRUGO | S_IWUSR }, 2471 .attr = {.name = "fifo_expire_async", .mode = S_IRUGO | S_IWUSR },
1722 .show = cfq_fifo_expire_w_show, 2472 .show = cfq_fifo_expire_async_show,
1723 .store = cfq_fifo_expire_w_store, 2473 .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}; 2474};
1735static struct cfq_fs_entry cfq_back_max_entry = { 2475static struct cfq_fs_entry cfq_back_max_entry = {
1736 .attr = {.name = "back_seek_max", .mode = S_IRUGO | S_IWUSR }, 2476 .attr = {.name = "back_seek_max", .mode = S_IRUGO | S_IWUSR },
@@ -1742,27 +2482,44 @@ static struct cfq_fs_entry cfq_back_penalty_entry = {
1742 .show = cfq_back_penalty_show, 2482 .show = cfq_back_penalty_show,
1743 .store = cfq_back_penalty_store, 2483 .store = cfq_back_penalty_store,
1744}; 2484};
1745static struct cfq_fs_entry cfq_clear_elapsed_entry = { 2485static struct cfq_fs_entry cfq_slice_sync_entry = {
1746 .attr = {.name = "clear_elapsed", .mode = S_IWUSR }, 2486 .attr = {.name = "slice_sync", .mode = S_IRUGO | S_IWUSR },
1747 .store = cfq_clear_elapsed, 2487 .show = cfq_slice_sync_show,
2488 .store = cfq_slice_sync_store,
2489};
2490static struct cfq_fs_entry cfq_slice_async_entry = {
2491 .attr = {.name = "slice_async", .mode = S_IRUGO | S_IWUSR },
2492 .show = cfq_slice_async_show,
2493 .store = cfq_slice_async_store,
1748}; 2494};
1749static struct cfq_fs_entry cfq_key_type_entry = { 2495static struct cfq_fs_entry cfq_slice_async_rq_entry = {
1750 .attr = {.name = "key_type", .mode = S_IRUGO | S_IWUSR }, 2496 .attr = {.name = "slice_async_rq", .mode = S_IRUGO | S_IWUSR },
1751 .show = cfq_read_key_type, 2497 .show = cfq_slice_async_rq_show,
1752 .store = cfq_set_key_type, 2498 .store = cfq_slice_async_rq_store,
2499};
2500static struct cfq_fs_entry cfq_slice_idle_entry = {
2501 .attr = {.name = "slice_idle", .mode = S_IRUGO | S_IWUSR },
2502 .show = cfq_slice_idle_show,
2503 .store = cfq_slice_idle_store,
2504};
2505static struct cfq_fs_entry cfq_max_depth_entry = {
2506 .attr = {.name = "max_depth", .mode = S_IRUGO | S_IWUSR },
2507 .show = cfq_max_depth_show,
2508 .store = cfq_max_depth_store,
1753}; 2509};
1754 2510
1755static struct attribute *default_attrs[] = { 2511static struct attribute *default_attrs[] = {
1756 &cfq_quantum_entry.attr, 2512 &cfq_quantum_entry.attr,
1757 &cfq_queued_entry.attr, 2513 &cfq_queued_entry.attr,
1758 &cfq_fifo_expire_r_entry.attr, 2514 &cfq_fifo_expire_sync_entry.attr,
1759 &cfq_fifo_expire_w_entry.attr, 2515 &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, 2516 &cfq_back_max_entry.attr,
1764 &cfq_back_penalty_entry.attr, 2517 &cfq_back_penalty_entry.attr,
1765 &cfq_clear_elapsed_entry.attr, 2518 &cfq_slice_sync_entry.attr,
2519 &cfq_slice_async_entry.attr,
2520 &cfq_slice_async_rq_entry.attr,
2521 &cfq_slice_idle_entry.attr,
2522 &cfq_max_depth_entry.attr,
1766 NULL, 2523 NULL,
1767}; 2524};
1768 2525
@@ -1832,21 +2589,46 @@ static int __init cfq_init(void)
1832{ 2589{
1833 int ret; 2590 int ret;
1834 2591
2592 /*
2593 * could be 0 on HZ < 1000 setups
2594 */
2595 if (!cfq_slice_async)
2596 cfq_slice_async = 1;
2597 if (!cfq_slice_idle)
2598 cfq_slice_idle = 1;
2599
1835 if (cfq_slab_setup()) 2600 if (cfq_slab_setup())
1836 return -ENOMEM; 2601 return -ENOMEM;
1837 2602
1838 ret = elv_register(&iosched_cfq); 2603 ret = elv_register(&iosched_cfq);
1839 if (!ret) { 2604 if (ret)
1840 __module_get(THIS_MODULE); 2605 cfq_slab_kill();
1841 return 0;
1842 }
1843 2606
1844 cfq_slab_kill();
1845 return ret; 2607 return ret;
1846} 2608}
1847 2609
1848static void __exit cfq_exit(void) 2610static void __exit cfq_exit(void)
1849{ 2611{
2612 struct task_struct *g, *p;
2613 unsigned long flags;
2614
2615 read_lock_irqsave(&tasklist_lock, flags);
2616
2617 /*
2618 * iterate each process in the system, removing our io_context
2619 */
2620 do_each_thread(g, p) {
2621 struct io_context *ioc = p->io_context;
2622
2623 if (ioc && ioc->cic) {
2624 ioc->cic->exit(ioc->cic);
2625 cfq_free_io_context(ioc->cic);
2626 ioc->cic = NULL;
2627 }
2628 } while_each_thread(g, p);
2629
2630 read_unlock_irqrestore(&tasklist_lock, flags);
2631
1850 cfq_slab_kill(); 2632 cfq_slab_kill();
1851 elv_unregister(&iosched_cfq); 2633 elv_unregister(&iosched_cfq);
1852} 2634}
generated by cgit v1.2.2 (git 2.25.0) at 2024-11-25 04:30:26 -0500