iosched: prevent aliased requests from starving other I/O
[linux-3.10.git] / block / deadline-iosched.c
1 /*
2  *  Deadline i/o scheduler.
3  *
4  *  Copyright (C) 2002 Jens Axboe <axboe@kernel.dk>
5  */
6 #include <linux/kernel.h>
7 #include <linux/fs.h>
8 #include <linux/blkdev.h>
9 #include <linux/elevator.h>
10 #include <linux/bio.h>
11 #include <linux/module.h>
12 #include <linux/slab.h>
13 #include <linux/init.h>
14 #include <linux/compiler.h>
15 #include <linux/rbtree.h>
16
17 /*
18  * See Documentation/block/deadline-iosched.txt
19  */
20 static const int read_expire = HZ / 2;  /* max time before a read is submitted. */
21 static const int write_expire = 5 * HZ; /* ditto for writes, these limits are SOFT! */
22 static const int writes_starved = 2;    /* max times reads can starve a write */
23 static const int fifo_batch = 16;       /* # of sequential requests treated as one
24                                      by the above parameters. For throughput. */
25
26 struct deadline_data {
27         /*
28          * run time data
29          */
30
31         /*
32          * requests (deadline_rq s) are present on both sort_list and fifo_list
33          */
34         struct rb_root sort_list[2];    
35         struct list_head fifo_list[2];
36
37         /*
38          * next in sort order. read, write or both are NULL
39          */
40         struct request *next_rq[2];
41         unsigned int batching;          /* number of sequential requests made */
42         sector_t last_sector;           /* head position */
43         unsigned int starved;           /* times reads have starved writes */
44
45         /*
46          * settings that change how the i/o scheduler behaves
47          */
48         int fifo_expire[2];
49         int fifo_batch;
50         int writes_starved;
51         int front_merges;
52 };
53
54 static void deadline_move_request(struct deadline_data *, struct request *);
55
56 static inline struct rb_root *
57 deadline_rb_root(struct deadline_data *dd, struct request *rq)
58 {
59         return &dd->sort_list[rq_data_dir(rq)];
60 }
61
62 /*
63  * get the request after `rq' in sector-sorted order
64  */
65 static inline struct request *
66 deadline_latter_request(struct request *rq)
67 {
68         struct rb_node *node = rb_next(&rq->rb_node);
69
70         if (node)
71                 return rb_entry_rq(node);
72
73         return NULL;
74 }
75
76 static void
77 deadline_add_rq_rb(struct deadline_data *dd, struct request *rq)
78 {
79         struct rb_root *root = deadline_rb_root(dd, rq);
80
81         elv_rb_add(root, rq);
82 }
83
84 static inline void
85 deadline_del_rq_rb(struct deadline_data *dd, struct request *rq)
86 {
87         const int data_dir = rq_data_dir(rq);
88
89         if (dd->next_rq[data_dir] == rq)
90                 dd->next_rq[data_dir] = deadline_latter_request(rq);
91
92         elv_rb_del(deadline_rb_root(dd, rq), rq);
93 }
94
95 /*
96  * add rq to rbtree and fifo
97  */
98 static void
99 deadline_add_request(struct request_queue *q, struct request *rq)
100 {
101         struct deadline_data *dd = q->elevator->elevator_data;
102         const int data_dir = rq_data_dir(rq);
103
104         deadline_add_rq_rb(dd, rq);
105
106         /*
107          * set expire time and add to fifo list
108          */
109         rq_set_fifo_time(rq, jiffies + dd->fifo_expire[data_dir]);
110         list_add_tail(&rq->queuelist, &dd->fifo_list[data_dir]);
111 }
112
113 /*
114  * remove rq from rbtree and fifo.
115  */
116 static void deadline_remove_request(struct request_queue *q, struct request *rq)
117 {
118         struct deadline_data *dd = q->elevator->elevator_data;
119
120         rq_fifo_clear(rq);
121         deadline_del_rq_rb(dd, rq);
122 }
123
124 static int
125 deadline_merge(struct request_queue *q, struct request **req, struct bio *bio)
126 {
127         struct deadline_data *dd = q->elevator->elevator_data;
128         struct request *__rq;
129         int ret;
130
131         /*
132          * check for front merge
133          */
134         if (dd->front_merges) {
135                 sector_t sector = bio->bi_sector + bio_sectors(bio);
136
137                 __rq = elv_rb_find(&dd->sort_list[bio_data_dir(bio)], sector);
138                 if (__rq) {
139                         BUG_ON(sector != blk_rq_pos(__rq));
140
141                         if (elv_rq_merge_ok(__rq, bio)) {
142                                 ret = ELEVATOR_FRONT_MERGE;
143                                 goto out;
144                         }
145                 }
146         }
147
148         return ELEVATOR_NO_MERGE;
149 out:
150         *req = __rq;
151         return ret;
152 }
153
154 static void deadline_merged_request(struct request_queue *q,
155                                     struct request *req, int type)
156 {
157         struct deadline_data *dd = q->elevator->elevator_data;
158
159         /*
160          * if the merge was a front merge, we need to reposition request
161          */
162         if (type == ELEVATOR_FRONT_MERGE) {
163                 elv_rb_del(deadline_rb_root(dd, req), req);
164                 deadline_add_rq_rb(dd, req);
165         }
166 }
167
168 static void
169 deadline_merged_requests(struct request_queue *q, struct request *req,
170                          struct request *next)
171 {
172         /*
173          * if next expires before rq, assign its expire time to rq
174          * and move into next position (next will be deleted) in fifo
175          */
176         if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) {
177                 if (time_before(rq_fifo_time(next), rq_fifo_time(req))) {
178                         list_move(&req->queuelist, &next->queuelist);
179                         rq_set_fifo_time(req, rq_fifo_time(next));
180                 }
181         }
182
183         /*
184          * kill knowledge of next, this one is a goner
185          */
186         deadline_remove_request(q, next);
187 }
188
189 /*
190  * move request from sort list to dispatch queue.
191  */
192 static inline void
193 deadline_move_to_dispatch(struct deadline_data *dd, struct request *rq)
194 {
195         struct request_queue *q = rq->q;
196
197         deadline_remove_request(q, rq);
198         elv_dispatch_add_tail(q, rq);
199 }
200
201 /*
202  * move an entry to dispatch queue
203  */
204 static void
205 deadline_move_request(struct deadline_data *dd, struct request *rq)
206 {
207         const int data_dir = rq_data_dir(rq);
208
209         dd->next_rq[READ] = NULL;
210         dd->next_rq[WRITE] = NULL;
211         dd->next_rq[data_dir] = deadline_latter_request(rq);
212
213         dd->last_sector = rq_end_sector(rq);
214
215         /*
216          * take it off the sort and fifo list, move
217          * to dispatch queue
218          */
219         deadline_move_to_dispatch(dd, rq);
220 }
221
222 /*
223  * deadline_check_fifo returns 0 if there are no expired requests on the fifo,
224  * 1 otherwise. Requires !list_empty(&dd->fifo_list[data_dir])
225  */
226 static inline int deadline_check_fifo(struct deadline_data *dd, int ddir)
227 {
228         struct request *rq = rq_entry_fifo(dd->fifo_list[ddir].next);
229
230         /*
231          * rq is expired!
232          */
233         if (time_after(jiffies, rq_fifo_time(rq)))
234                 return 1;
235
236         return 0;
237 }
238
239 /*
240  * deadline_dispatch_requests selects the best request according to
241  * read/write expire, fifo_batch, etc
242  */
243 static int deadline_dispatch_requests(struct request_queue *q, int force)
244 {
245         struct deadline_data *dd = q->elevator->elevator_data;
246         const int reads = !list_empty(&dd->fifo_list[READ]);
247         const int writes = !list_empty(&dd->fifo_list[WRITE]);
248         struct request *rq;
249         int data_dir;
250
251         /*
252          * batches are currently reads XOR writes
253          */
254         if (dd->next_rq[WRITE])
255                 rq = dd->next_rq[WRITE];
256         else
257                 rq = dd->next_rq[READ];
258
259         if (rq && dd->batching < dd->fifo_batch)
260                 /* we have a next request are still entitled to batch */
261                 goto dispatch_request;
262
263         /*
264          * at this point we are not running a batch. select the appropriate
265          * data direction (read / write)
266          */
267
268         if (reads) {
269                 BUG_ON(RB_EMPTY_ROOT(&dd->sort_list[READ]));
270
271                 if (writes && (dd->starved++ >= dd->writes_starved))
272                         goto dispatch_writes;
273
274                 data_dir = READ;
275
276                 goto dispatch_find_request;
277         }
278
279         /*
280          * there are either no reads or writes have been starved
281          */
282
283         if (writes) {
284 dispatch_writes:
285                 BUG_ON(RB_EMPTY_ROOT(&dd->sort_list[WRITE]));
286
287                 dd->starved = 0;
288
289                 data_dir = WRITE;
290
291                 goto dispatch_find_request;
292         }
293
294         return 0;
295
296 dispatch_find_request:
297         /*
298          * we are not running a batch, find best request for selected data_dir
299          */
300         if (deadline_check_fifo(dd, data_dir) || !dd->next_rq[data_dir]) {
301                 /*
302                  * A deadline has expired, the last request was in the other
303                  * direction, or we have run out of higher-sectored requests.
304                  * Start again from the request with the earliest expiry time.
305                  */
306                 rq = rq_entry_fifo(dd->fifo_list[data_dir].next);
307         } else {
308                 /*
309                  * The last req was the same dir and we have a next request in
310                  * sort order. No expired requests so continue on from here.
311                  */
312                 rq = dd->next_rq[data_dir];
313         }
314
315         dd->batching = 0;
316
317 dispatch_request:
318         /*
319          * rq is the selected appropriate request.
320          */
321         dd->batching++;
322         deadline_move_request(dd, rq);
323
324         return 1;
325 }
326
327 static void deadline_exit_queue(struct elevator_queue *e)
328 {
329         struct deadline_data *dd = e->elevator_data;
330
331         BUG_ON(!list_empty(&dd->fifo_list[READ]));
332         BUG_ON(!list_empty(&dd->fifo_list[WRITE]));
333
334         kfree(dd);
335 }
336
337 /*
338  * initialize elevator private data (deadline_data).
339  */
340 static void *deadline_init_queue(struct request_queue *q)
341 {
342         struct deadline_data *dd;
343
344         dd = kmalloc_node(sizeof(*dd), GFP_KERNEL | __GFP_ZERO, q->node);
345         if (!dd)
346                 return NULL;
347
348         INIT_LIST_HEAD(&dd->fifo_list[READ]);
349         INIT_LIST_HEAD(&dd->fifo_list[WRITE]);
350         dd->sort_list[READ] = RB_ROOT;
351         dd->sort_list[WRITE] = RB_ROOT;
352         dd->fifo_expire[READ] = read_expire;
353         dd->fifo_expire[WRITE] = write_expire;
354         dd->writes_starved = writes_starved;
355         dd->front_merges = 1;
356         dd->fifo_batch = fifo_batch;
357         return dd;
358 }
359
360 /*
361  * sysfs parts below
362  */
363
364 static ssize_t
365 deadline_var_show(int var, char *page)
366 {
367         return sprintf(page, "%d\n", var);
368 }
369
370 static ssize_t
371 deadline_var_store(int *var, const char *page, size_t count)
372 {
373         char *p = (char *) page;
374
375         *var = simple_strtol(p, &p, 10);
376         return count;
377 }
378
379 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
380 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
381 {                                                                       \
382         struct deadline_data *dd = e->elevator_data;                    \
383         int __data = __VAR;                                             \
384         if (__CONV)                                                     \
385                 __data = jiffies_to_msecs(__data);                      \
386         return deadline_var_show(__data, (page));                       \
387 }
388 SHOW_FUNCTION(deadline_read_expire_show, dd->fifo_expire[READ], 1);
389 SHOW_FUNCTION(deadline_write_expire_show, dd->fifo_expire[WRITE], 1);
390 SHOW_FUNCTION(deadline_writes_starved_show, dd->writes_starved, 0);
391 SHOW_FUNCTION(deadline_front_merges_show, dd->front_merges, 0);
392 SHOW_FUNCTION(deadline_fifo_batch_show, dd->fifo_batch, 0);
393 #undef SHOW_FUNCTION
394
395 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
396 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
397 {                                                                       \
398         struct deadline_data *dd = e->elevator_data;                    \
399         int __data;                                                     \
400         int ret = deadline_var_store(&__data, (page), count);           \
401         if (__data < (MIN))                                             \
402                 __data = (MIN);                                         \
403         else if (__data > (MAX))                                        \
404                 __data = (MAX);                                         \
405         if (__CONV)                                                     \
406                 *(__PTR) = msecs_to_jiffies(__data);                    \
407         else                                                            \
408                 *(__PTR) = __data;                                      \
409         return ret;                                                     \
410 }
411 STORE_FUNCTION(deadline_read_expire_store, &dd->fifo_expire[READ], 0, INT_MAX, 1);
412 STORE_FUNCTION(deadline_write_expire_store, &dd->fifo_expire[WRITE], 0, INT_MAX, 1);
413 STORE_FUNCTION(deadline_writes_starved_store, &dd->writes_starved, INT_MIN, INT_MAX, 0);
414 STORE_FUNCTION(deadline_front_merges_store, &dd->front_merges, 0, 1, 0);
415 STORE_FUNCTION(deadline_fifo_batch_store, &dd->fifo_batch, 0, INT_MAX, 0);
416 #undef STORE_FUNCTION
417
418 #define DD_ATTR(name) \
419         __ATTR(name, S_IRUGO|S_IWUSR, deadline_##name##_show, \
420                                       deadline_##name##_store)
421
422 static struct elv_fs_entry deadline_attrs[] = {
423         DD_ATTR(read_expire),
424         DD_ATTR(write_expire),
425         DD_ATTR(writes_starved),
426         DD_ATTR(front_merges),
427         DD_ATTR(fifo_batch),
428         __ATTR_NULL
429 };
430
431 static struct elevator_type iosched_deadline = {
432         .ops = {
433                 .elevator_merge_fn =            deadline_merge,
434                 .elevator_merged_fn =           deadline_merged_request,
435                 .elevator_merge_req_fn =        deadline_merged_requests,
436                 .elevator_dispatch_fn =         deadline_dispatch_requests,
437                 .elevator_add_req_fn =          deadline_add_request,
438                 .elevator_former_req_fn =       elv_rb_former_request,
439                 .elevator_latter_req_fn =       elv_rb_latter_request,
440                 .elevator_init_fn =             deadline_init_queue,
441                 .elevator_exit_fn =             deadline_exit_queue,
442         },
443
444         .elevator_attrs = deadline_attrs,
445         .elevator_name = "deadline",
446         .elevator_owner = THIS_MODULE,
447 };
448
449 static int __init deadline_init(void)
450 {
451         elv_register(&iosched_deadline);
452
453         return 0;
454 }
455
456 static void __exit deadline_exit(void)
457 {
458         elv_unregister(&iosched_deadline);
459 }
460
461 module_init(deadline_init);
462 module_exit(deadline_exit);
463
464 MODULE_AUTHOR("Jens Axboe");
465 MODULE_LICENSE("GPL");
466 MODULE_DESCRIPTION("deadline IO scheduler");