1 /* $NetBSD: bufq_priocscan.c,v 1.21 2017/05/04 11:03:27 kamil Exp $ */
2
3 /*-
4 * Copyright (c)2004,2005,2006,2008,2009,2011,2012 YAMAMOTO Takashi,
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 *
16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26 * SUCH DAMAGE.
27 */
28
29 #include <sys/cdefs.h>
30 __KERNEL_RCSID(0, "$NetBSD: bufq_priocscan.c,v 1.21 2017/05/04 11:03:27 kamil Exp $");
31
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/buf.h>
35 #include <sys/bufq.h>
36 #include <sys/bufq_impl.h>
37 #include <sys/kmem.h>
38 #include <sys/rbtree.h>
39 #include <sys/module.h>
40
41 #undef PRIOCSCAN_USE_GLOBAL_POSITION
42
43 /*
44 * Cyclical scan (CSCAN)
45 */
46
47 struct cscan_key {
48 daddr_t k_rawblkno;
49 int k_cylinder;
50 };
51
52 struct cscan_queue {
53 rb_tree_t cq_buffers; /* ordered list of buffers */
54 #if !defined(PRIOCSCAN_USE_GLOBAL_POSITION)
55 struct cscan_key cq_lastkey; /* key of last request */
56 #endif /* !defined(PRIOCSCAN_USE_GLOBAL_POSITION) */
57 int cq_sortby; /* BUFQ_SORT_MASK */
58 rb_tree_ops_t cq_ops;
59 };
60
61 static signed int
buf_cmp(const struct buf * b1,const struct buf * b2,int sortby)62 buf_cmp(const struct buf *b1, const struct buf *b2, int sortby)
63 {
64
65 if (buf_inorder(b2, b1, sortby)) {
66 return 1; /* b1 > b2 */
67 }
68 if (buf_inorder(b1, b2, sortby)) {
69 return -1; /* b1 < b2 */
70 }
71 return 0;
72 }
73
74 /* return positive if n1 > n2 */
75 static signed int
cscan_tree_compare_nodes(void * context,const void * n1,const void * n2)76 cscan_tree_compare_nodes(void *context, const void *n1, const void *n2)
77 {
78 const struct cscan_queue * const q = context;
79 const struct buf * const b1 = n1;
80 const struct buf * const b2 = n2;
81 const int sortby = q->cq_sortby;
82 const int diff = buf_cmp(b1, b2, sortby);
83
84 /*
85 * XXX rawblkno/cylinder might not be unique. eg. unbuffered i/o
86 */
87
88 if (diff != 0) {
89 return diff;
90 }
91
92 /*
93 * XXX rawblkno/cylinder might not be unique. eg. unbuffered i/o
94 */
95 if (b1 > b2) {
96 return 1;
97 }
98 if (b1 < b2) {
99 return -1;
100 }
101 return 0;
102 }
103
104 /* return positive if n1 > k2 */
105 static signed int
cscan_tree_compare_key(void * context,const void * n1,const void * k2)106 cscan_tree_compare_key(void *context, const void *n1, const void *k2)
107 {
108 const struct cscan_queue * const q = context;
109 const struct buf * const b1 = n1;
110 const struct cscan_key * const key = k2;
111 const struct buf tmp = {
112 .b_rawblkno = key->k_rawblkno,
113 .b_cylinder = key->k_cylinder,
114 };
115 const struct buf *b2 = &tmp;
116 const int sortby = q->cq_sortby;
117
118 return buf_cmp(b1, b2, sortby);
119 }
120
121 static void __unused
cscan_dump(struct cscan_queue * cq)122 cscan_dump(struct cscan_queue *cq)
123 {
124 const int sortby = cq->cq_sortby;
125 struct buf *bp;
126
127 RB_TREE_FOREACH(bp, &cq->cq_buffers) {
128 if (sortby == BUFQ_SORT_RAWBLOCK) {
129 printf(" %jd", (intmax_t)bp->b_rawblkno);
130 } else {
131 printf(" %jd/%jd",
132 (intmax_t)bp->b_cylinder, (intmax_t)bp->b_rawblkno);
133 }
134 }
135 }
136
137 static inline bool
cscan_empty(struct cscan_queue * q)138 cscan_empty(struct cscan_queue *q)
139 {
140
141 /* XXX this might do more work than necessary */
142 return rb_tree_iterate(&q->cq_buffers, NULL, RB_DIR_LEFT) == NULL;
143 }
144
145 static void
cscan_put(struct cscan_queue * q,struct buf * bp)146 cscan_put(struct cscan_queue *q, struct buf *bp)
147 {
148 struct buf *obp __diagused;
149
150 obp = rb_tree_insert_node(&q->cq_buffers, bp);
151 KASSERT(obp == bp); /* see cscan_tree_compare_nodes */
152 }
153
154 static struct buf *
cscan_get(struct cscan_queue * q,int remove,struct cscan_key * key)155 cscan_get(struct cscan_queue *q, int remove, struct cscan_key *key)
156 {
157 struct buf *bp;
158
159 bp = rb_tree_find_node_geq(&q->cq_buffers, key);
160 KDASSERT(bp == NULL || cscan_tree_compare_key(q, bp, key) >= 0);
161 if (bp == NULL) {
162 bp = rb_tree_iterate(&q->cq_buffers, NULL, RB_DIR_LEFT);
163 KDASSERT(cscan_tree_compare_key(q, bp, key) < 0);
164 }
165 if (bp != NULL && remove) {
166 #if defined(DEBUG)
167 struct buf *nbp;
168 #endif /* defined(DEBUG) */
169
170 rb_tree_remove_node(&q->cq_buffers, bp);
171 /*
172 * remember the head position.
173 */
174 key->k_cylinder = bp->b_cylinder;
175 key->k_rawblkno = bp->b_rawblkno + (bp->b_bcount >> DEV_BSHIFT);
176 #if defined(DEBUG)
177 nbp = rb_tree_find_node_geq(&q->cq_buffers, key);
178 if (nbp != NULL && cscan_tree_compare_nodes(q, nbp, bp) < 0) {
179 panic("%s: wrong order %p < %p\n", __func__,
180 nbp, bp);
181 }
182 #endif /* defined(DEBUG) */
183 }
184 return bp;
185 }
186
187 static void
cscan_init(struct cscan_queue * q,int sortby)188 cscan_init(struct cscan_queue *q, int sortby)
189 {
190 static const rb_tree_ops_t cscan_tree_ops = {
191 .rbto_compare_nodes = cscan_tree_compare_nodes,
192 .rbto_compare_key = cscan_tree_compare_key,
193 .rbto_node_offset = offsetof(struct buf, b_u.u_rbnode),
194 .rbto_context = NULL,
195 };
196
197 q->cq_sortby = sortby;
198 /* XXX copy ops to workaround rbtree.h API limitation */
199 q->cq_ops = cscan_tree_ops;
200 q->cq_ops.rbto_context = q;
201 rb_tree_init(&q->cq_buffers, &q->cq_ops);
202 }
203
204 /*
205 * Per-prioritiy CSCAN.
206 *
207 * XXX probably we should have a way to raise
208 * priority of the on-queue requests.
209 */
210 #define PRIOCSCAN_NQUEUE 3
211
212 struct priocscan_queue {
213 struct cscan_queue q_queue;
214 unsigned int q_burst;
215 };
216
217 struct bufq_priocscan {
218 struct priocscan_queue bq_queue[PRIOCSCAN_NQUEUE];
219
220 #if defined(PRIOCSCAN_USE_GLOBAL_POSITION)
221 /*
222 * XXX using "global" head position can reduce positioning time
223 * when switching between queues.
224 * although it might affect against fairness.
225 */
226 struct cscan_key bq_lastkey;
227 #endif
228 };
229
230 /*
231 * how many requests to serve when having pending requests on other queues.
232 *
233 * XXX tune
234 * be careful: while making these values larger likely
235 * increases the total throughput, it can also increase latencies
236 * for some workloads.
237 */
238 const int priocscan_burst[] = {
239 64, 16, 4
240 };
241
242 static void bufq_priocscan_init(struct bufq_state *);
243 static void bufq_priocscan_put(struct bufq_state *, struct buf *);
244 static struct buf *bufq_priocscan_get(struct bufq_state *, int);
245
246 BUFQ_DEFINE(priocscan, 40, bufq_priocscan_init);
247
248 static inline struct cscan_queue *bufq_priocscan_selectqueue(
249 struct bufq_priocscan *, const struct buf *);
250
251 static inline struct cscan_queue *
bufq_priocscan_selectqueue(struct bufq_priocscan * q,const struct buf * bp)252 bufq_priocscan_selectqueue(struct bufq_priocscan *q, const struct buf *bp)
253 {
254 static const int priocscan_priomap[] = {
255 [BPRIO_TIMENONCRITICAL] = 2,
256 [BPRIO_TIMELIMITED] = 1,
257 [BPRIO_TIMECRITICAL] = 0
258 };
259
260 return &q->bq_queue[priocscan_priomap[BIO_GETPRIO(bp)]].q_queue;
261 }
262
263 static void
bufq_priocscan_put(struct bufq_state * bufq,struct buf * bp)264 bufq_priocscan_put(struct bufq_state *bufq, struct buf *bp)
265 {
266 struct bufq_priocscan *q = bufq_private(bufq);
267 struct cscan_queue *cq;
268
269 cq = bufq_priocscan_selectqueue(q, bp);
270 cscan_put(cq, bp);
271 }
272
273 static struct buf *
bufq_priocscan_get(struct bufq_state * bufq,int remove)274 bufq_priocscan_get(struct bufq_state *bufq, int remove)
275 {
276 struct bufq_priocscan *q = bufq_private(bufq);
277 struct priocscan_queue *pq, *npq;
278 struct priocscan_queue *first; /* highest priority non-empty queue */
279 const struct priocscan_queue *epq;
280 struct buf *bp;
281 bool single; /* true if there's only one non-empty queue */
282
283 /*
284 * find the highest priority non-empty queue.
285 */
286 pq = &q->bq_queue[0];
287 epq = pq + PRIOCSCAN_NQUEUE;
288 for (; pq < epq; pq++) {
289 if (!cscan_empty(&pq->q_queue)) {
290 break;
291 }
292 }
293 if (pq == epq) {
294 /*
295 * all our queues are empty. there's nothing to serve.
296 */
297 return NULL;
298 }
299 first = pq;
300
301 /*
302 * scan the rest of queues.
303 *
304 * if we have two or more non-empty queues, we serve the highest
305 * priority one with non-zero burst count.
306 */
307 single = true;
308 for (npq = pq + 1; npq < epq; npq++) {
309 if (!cscan_empty(&npq->q_queue)) {
310 /*
311 * we found another non-empty queue.
312 * it means that a queue needs to consume its burst
313 * count to be served.
314 */
315 single = false;
316
317 /*
318 * check if our current candidate queue has already
319 * exhausted its burst count.
320 */
321 if (pq->q_burst > 0) {
322 break;
323 }
324 pq = npq;
325 }
326 }
327 if (single) {
328 /*
329 * there's only a non-empty queue.
330 * just serve it without consuming its burst count.
331 */
332 KASSERT(pq == first);
333 } else {
334 /*
335 * there are two or more non-empty queues.
336 */
337 if (pq->q_burst == 0) {
338 /*
339 * no queues can be served because they have already
340 * exhausted their burst count.
341 */
342 unsigned int i;
343 #ifdef DEBUG
344 for (i = 0; i < PRIOCSCAN_NQUEUE; i++) {
345 pq = &q->bq_queue[i];
346 if (!cscan_empty(&pq->q_queue) && pq->q_burst) {
347 panic("%s: inconsist", __func__);
348 }
349 }
350 #endif /* DEBUG */
351 /*
352 * reset burst counts.
353 */
354 if (remove) {
355 for (i = 0; i < PRIOCSCAN_NQUEUE; i++) {
356 pq = &q->bq_queue[i];
357 pq->q_burst = priocscan_burst[i];
358 }
359 }
360
361 /*
362 * serve the highest priority non-empty queue.
363 */
364 pq = first;
365 }
366 /*
367 * consume the burst count.
368 *
369 * XXX account only by number of requests. is it good enough?
370 */
371 if (remove) {
372 KASSERT(pq->q_burst > 0);
373 pq->q_burst--;
374 }
375 }
376
377 /*
378 * finally, get a request from the selected queue.
379 */
380 KDASSERT(!cscan_empty(&pq->q_queue));
381 bp = cscan_get(&pq->q_queue, remove,
382 #if defined(PRIOCSCAN_USE_GLOBAL_POSITION)
383 &q->bq_lastkey
384 #else /* defined(PRIOCSCAN_USE_GLOBAL_POSITION) */
385 &pq->q_queue.cq_lastkey
386 #endif /* defined(PRIOCSCAN_USE_GLOBAL_POSITION) */
387 );
388 KDASSERT(bp != NULL);
389 KDASSERT(&pq->q_queue == bufq_priocscan_selectqueue(q, bp));
390
391 return bp;
392 }
393
394 static struct buf *
bufq_priocscan_cancel(struct bufq_state * bufq,struct buf * bp)395 bufq_priocscan_cancel(struct bufq_state *bufq, struct buf *bp)
396 {
397 struct bufq_priocscan * const q = bufq_private(bufq);
398 unsigned int i;
399
400 for (i = 0; i < PRIOCSCAN_NQUEUE; i++) {
401 struct cscan_queue * const cq = &q->bq_queue[i].q_queue;
402 struct buf *it;
403
404 /*
405 * XXX probably could be faster but the cancel functionality
406 * is not widely used anyway.
407 */
408 RB_TREE_FOREACH(it, &cq->cq_buffers) {
409 if (it == bp) {
410 rb_tree_remove_node(&cq->cq_buffers, bp);
411 return bp;
412 }
413 }
414 }
415 return NULL;
416 }
417
418 static void
bufq_priocscan_fini(struct bufq_state * bufq)419 bufq_priocscan_fini(struct bufq_state *bufq)
420 {
421
422 KASSERT(bufq->bq_private != NULL);
423 kmem_free(bufq->bq_private, sizeof(struct bufq_priocscan));
424 }
425
426 static void
bufq_priocscan_init(struct bufq_state * bufq)427 bufq_priocscan_init(struct bufq_state *bufq)
428 {
429 struct bufq_priocscan *q;
430 const int sortby = bufq->bq_flags & BUFQ_SORT_MASK;
431 unsigned int i;
432
433 bufq->bq_get = bufq_priocscan_get;
434 bufq->bq_put = bufq_priocscan_put;
435 bufq->bq_cancel = bufq_priocscan_cancel;
436 bufq->bq_fini = bufq_priocscan_fini;
437 bufq->bq_private = kmem_zalloc(sizeof(struct bufq_priocscan), KM_SLEEP);
438
439 q = bufq->bq_private;
440 for (i = 0; i < PRIOCSCAN_NQUEUE; i++) {
441 struct cscan_queue *cq = &q->bq_queue[i].q_queue;
442
443 cscan_init(cq, sortby);
444 }
445 }
446
447 MODULE(MODULE_CLASS_BUFQ, bufq_priocscan, NULL);
448
449 static int
bufq_priocscan_modcmd(modcmd_t cmd,void * opaque)450 bufq_priocscan_modcmd(modcmd_t cmd, void *opaque)
451 {
452
453 switch (cmd) {
454 case MODULE_CMD_INIT:
455 return bufq_register(&bufq_strat_priocscan);
456 case MODULE_CMD_FINI:
457 return bufq_unregister(&bufq_strat_priocscan);
458 default:
459 return ENOTTY;
460 }
461 }
462