xref: /openbmc/linux/block/blk-merge.c (revision a1e58bbd)
1 /*
2  * Functions related to segment and merge handling
3  */
4 #include <linux/kernel.h>
5 #include <linux/module.h>
6 #include <linux/bio.h>
7 #include <linux/blkdev.h>
8 #include <linux/scatterlist.h>
9 
10 #include "blk.h"
11 
12 void blk_recalc_rq_sectors(struct request *rq, int nsect)
13 {
14 	if (blk_fs_request(rq)) {
15 		rq->hard_sector += nsect;
16 		rq->hard_nr_sectors -= nsect;
17 
18 		/*
19 		 * Move the I/O submission pointers ahead if required.
20 		 */
21 		if ((rq->nr_sectors >= rq->hard_nr_sectors) &&
22 		    (rq->sector <= rq->hard_sector)) {
23 			rq->sector = rq->hard_sector;
24 			rq->nr_sectors = rq->hard_nr_sectors;
25 			rq->hard_cur_sectors = bio_cur_sectors(rq->bio);
26 			rq->current_nr_sectors = rq->hard_cur_sectors;
27 			rq->buffer = bio_data(rq->bio);
28 		}
29 
30 		/*
31 		 * if total number of sectors is less than the first segment
32 		 * size, something has gone terribly wrong
33 		 */
34 		if (rq->nr_sectors < rq->current_nr_sectors) {
35 			printk(KERN_ERR "blk: request botched\n");
36 			rq->nr_sectors = rq->current_nr_sectors;
37 		}
38 	}
39 }
40 
41 void blk_recalc_rq_segments(struct request *rq)
42 {
43 	int nr_phys_segs;
44 	int nr_hw_segs;
45 	unsigned int phys_size;
46 	unsigned int hw_size;
47 	struct bio_vec *bv, *bvprv = NULL;
48 	int seg_size;
49 	int hw_seg_size;
50 	int cluster;
51 	struct req_iterator iter;
52 	int high, highprv = 1;
53 	struct request_queue *q = rq->q;
54 
55 	if (!rq->bio)
56 		return;
57 
58 	cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
59 	hw_seg_size = seg_size = 0;
60 	phys_size = hw_size = nr_phys_segs = nr_hw_segs = 0;
61 	rq_for_each_segment(bv, rq, iter) {
62 		/*
63 		 * the trick here is making sure that a high page is never
64 		 * considered part of another segment, since that might
65 		 * change with the bounce page.
66 		 */
67 		high = page_to_pfn(bv->bv_page) > q->bounce_pfn;
68 		if (high || highprv)
69 			goto new_hw_segment;
70 		if (cluster) {
71 			if (seg_size + bv->bv_len > q->max_segment_size)
72 				goto new_segment;
73 			if (!BIOVEC_PHYS_MERGEABLE(bvprv, bv))
74 				goto new_segment;
75 			if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bv))
76 				goto new_segment;
77 			if (BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
78 				goto new_hw_segment;
79 
80 			seg_size += bv->bv_len;
81 			hw_seg_size += bv->bv_len;
82 			bvprv = bv;
83 			continue;
84 		}
85 new_segment:
86 		if (BIOVEC_VIRT_MERGEABLE(bvprv, bv) &&
87 		    !BIOVEC_VIRT_OVERSIZE(hw_seg_size + bv->bv_len))
88 			hw_seg_size += bv->bv_len;
89 		else {
90 new_hw_segment:
91 			if (nr_hw_segs == 1 &&
92 			    hw_seg_size > rq->bio->bi_hw_front_size)
93 				rq->bio->bi_hw_front_size = hw_seg_size;
94 			hw_seg_size = BIOVEC_VIRT_START_SIZE(bv) + bv->bv_len;
95 			nr_hw_segs++;
96 		}
97 
98 		nr_phys_segs++;
99 		bvprv = bv;
100 		seg_size = bv->bv_len;
101 		highprv = high;
102 	}
103 
104 	if (nr_hw_segs == 1 &&
105 	    hw_seg_size > rq->bio->bi_hw_front_size)
106 		rq->bio->bi_hw_front_size = hw_seg_size;
107 	if (hw_seg_size > rq->biotail->bi_hw_back_size)
108 		rq->biotail->bi_hw_back_size = hw_seg_size;
109 	rq->nr_phys_segments = nr_phys_segs;
110 	rq->nr_hw_segments = nr_hw_segs;
111 }
112 
113 void blk_recount_segments(struct request_queue *q, struct bio *bio)
114 {
115 	struct request rq;
116 	struct bio *nxt = bio->bi_next;
117 	rq.q = q;
118 	rq.bio = rq.biotail = bio;
119 	bio->bi_next = NULL;
120 	blk_recalc_rq_segments(&rq);
121 	bio->bi_next = nxt;
122 	bio->bi_phys_segments = rq.nr_phys_segments;
123 	bio->bi_hw_segments = rq.nr_hw_segments;
124 	bio->bi_flags |= (1 << BIO_SEG_VALID);
125 }
126 EXPORT_SYMBOL(blk_recount_segments);
127 
128 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
129 				   struct bio *nxt)
130 {
131 	if (!(q->queue_flags & (1 << QUEUE_FLAG_CLUSTER)))
132 		return 0;
133 
134 	if (!BIOVEC_PHYS_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)))
135 		return 0;
136 	if (bio->bi_size + nxt->bi_size > q->max_segment_size)
137 		return 0;
138 
139 	/*
140 	 * bio and nxt are contigous in memory, check if the queue allows
141 	 * these two to be merged into one
142 	 */
143 	if (BIO_SEG_BOUNDARY(q, bio, nxt))
144 		return 1;
145 
146 	return 0;
147 }
148 
149 static int blk_hw_contig_segment(struct request_queue *q, struct bio *bio,
150 				 struct bio *nxt)
151 {
152 	if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
153 		blk_recount_segments(q, bio);
154 	if (unlikely(!bio_flagged(nxt, BIO_SEG_VALID)))
155 		blk_recount_segments(q, nxt);
156 	if (!BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(nxt)) ||
157 	    BIOVEC_VIRT_OVERSIZE(bio->bi_hw_back_size + nxt->bi_hw_front_size))
158 		return 0;
159 	if (bio->bi_hw_back_size + nxt->bi_hw_front_size > q->max_segment_size)
160 		return 0;
161 
162 	return 1;
163 }
164 
165 /*
166  * map a request to scatterlist, return number of sg entries setup. Caller
167  * must make sure sg can hold rq->nr_phys_segments entries
168  */
169 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
170 		  struct scatterlist *sglist)
171 {
172 	struct bio_vec *bvec, *bvprv;
173 	struct req_iterator iter;
174 	struct scatterlist *sg;
175 	int nsegs, cluster;
176 
177 	nsegs = 0;
178 	cluster = q->queue_flags & (1 << QUEUE_FLAG_CLUSTER);
179 
180 	/*
181 	 * for each bio in rq
182 	 */
183 	bvprv = NULL;
184 	sg = NULL;
185 	rq_for_each_segment(bvec, rq, iter) {
186 		int nbytes = bvec->bv_len;
187 
188 		if (bvprv && cluster) {
189 			if (sg->length + nbytes > q->max_segment_size)
190 				goto new_segment;
191 
192 			if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
193 				goto new_segment;
194 			if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
195 				goto new_segment;
196 
197 			sg->length += nbytes;
198 		} else {
199 new_segment:
200 			if (!sg)
201 				sg = sglist;
202 			else {
203 				/*
204 				 * If the driver previously mapped a shorter
205 				 * list, we could see a termination bit
206 				 * prematurely unless it fully inits the sg
207 				 * table on each mapping. We KNOW that there
208 				 * must be more entries here or the driver
209 				 * would be buggy, so force clear the
210 				 * termination bit to avoid doing a full
211 				 * sg_init_table() in drivers for each command.
212 				 */
213 				sg->page_link &= ~0x02;
214 				sg = sg_next(sg);
215 			}
216 
217 			sg_set_page(sg, bvec->bv_page, nbytes, bvec->bv_offset);
218 			nsegs++;
219 		}
220 		bvprv = bvec;
221 	} /* segments in rq */
222 
223 	if (q->dma_drain_size && q->dma_drain_needed(rq)) {
224 		if (rq->cmd_flags & REQ_RW)
225 			memset(q->dma_drain_buffer, 0, q->dma_drain_size);
226 
227 		sg->page_link &= ~0x02;
228 		sg = sg_next(sg);
229 		sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
230 			    q->dma_drain_size,
231 			    ((unsigned long)q->dma_drain_buffer) &
232 			    (PAGE_SIZE - 1));
233 		nsegs++;
234 		rq->extra_len += q->dma_drain_size;
235 	}
236 
237 	if (sg)
238 		sg_mark_end(sg);
239 
240 	return nsegs;
241 }
242 EXPORT_SYMBOL(blk_rq_map_sg);
243 
244 static inline int ll_new_mergeable(struct request_queue *q,
245 				   struct request *req,
246 				   struct bio *bio)
247 {
248 	int nr_phys_segs = bio_phys_segments(q, bio);
249 
250 	if (req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
251 		req->cmd_flags |= REQ_NOMERGE;
252 		if (req == q->last_merge)
253 			q->last_merge = NULL;
254 		return 0;
255 	}
256 
257 	/*
258 	 * A hw segment is just getting larger, bump just the phys
259 	 * counter.
260 	 */
261 	req->nr_phys_segments += nr_phys_segs;
262 	return 1;
263 }
264 
265 static inline int ll_new_hw_segment(struct request_queue *q,
266 				    struct request *req,
267 				    struct bio *bio)
268 {
269 	int nr_hw_segs = bio_hw_segments(q, bio);
270 	int nr_phys_segs = bio_phys_segments(q, bio);
271 
272 	if (req->nr_hw_segments + nr_hw_segs > q->max_hw_segments
273 	    || req->nr_phys_segments + nr_phys_segs > q->max_phys_segments) {
274 		req->cmd_flags |= REQ_NOMERGE;
275 		if (req == q->last_merge)
276 			q->last_merge = NULL;
277 		return 0;
278 	}
279 
280 	/*
281 	 * This will form the start of a new hw segment.  Bump both
282 	 * counters.
283 	 */
284 	req->nr_hw_segments += nr_hw_segs;
285 	req->nr_phys_segments += nr_phys_segs;
286 	return 1;
287 }
288 
289 int ll_back_merge_fn(struct request_queue *q, struct request *req,
290 		     struct bio *bio)
291 {
292 	unsigned short max_sectors;
293 	int len;
294 
295 	if (unlikely(blk_pc_request(req)))
296 		max_sectors = q->max_hw_sectors;
297 	else
298 		max_sectors = q->max_sectors;
299 
300 	if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
301 		req->cmd_flags |= REQ_NOMERGE;
302 		if (req == q->last_merge)
303 			q->last_merge = NULL;
304 		return 0;
305 	}
306 	if (unlikely(!bio_flagged(req->biotail, BIO_SEG_VALID)))
307 		blk_recount_segments(q, req->biotail);
308 	if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
309 		blk_recount_segments(q, bio);
310 	len = req->biotail->bi_hw_back_size + bio->bi_hw_front_size;
311 	if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(req->biotail), __BVEC_START(bio))
312 	    && !BIOVEC_VIRT_OVERSIZE(len)) {
313 		int mergeable =  ll_new_mergeable(q, req, bio);
314 
315 		if (mergeable) {
316 			if (req->nr_hw_segments == 1)
317 				req->bio->bi_hw_front_size = len;
318 			if (bio->bi_hw_segments == 1)
319 				bio->bi_hw_back_size = len;
320 		}
321 		return mergeable;
322 	}
323 
324 	return ll_new_hw_segment(q, req, bio);
325 }
326 
327 int ll_front_merge_fn(struct request_queue *q, struct request *req,
328 		      struct bio *bio)
329 {
330 	unsigned short max_sectors;
331 	int len;
332 
333 	if (unlikely(blk_pc_request(req)))
334 		max_sectors = q->max_hw_sectors;
335 	else
336 		max_sectors = q->max_sectors;
337 
338 
339 	if (req->nr_sectors + bio_sectors(bio) > max_sectors) {
340 		req->cmd_flags |= REQ_NOMERGE;
341 		if (req == q->last_merge)
342 			q->last_merge = NULL;
343 		return 0;
344 	}
345 	len = bio->bi_hw_back_size + req->bio->bi_hw_front_size;
346 	if (unlikely(!bio_flagged(bio, BIO_SEG_VALID)))
347 		blk_recount_segments(q, bio);
348 	if (unlikely(!bio_flagged(req->bio, BIO_SEG_VALID)))
349 		blk_recount_segments(q, req->bio);
350 	if (BIOVEC_VIRT_MERGEABLE(__BVEC_END(bio), __BVEC_START(req->bio)) &&
351 	    !BIOVEC_VIRT_OVERSIZE(len)) {
352 		int mergeable =  ll_new_mergeable(q, req, bio);
353 
354 		if (mergeable) {
355 			if (bio->bi_hw_segments == 1)
356 				bio->bi_hw_front_size = len;
357 			if (req->nr_hw_segments == 1)
358 				req->biotail->bi_hw_back_size = len;
359 		}
360 		return mergeable;
361 	}
362 
363 	return ll_new_hw_segment(q, req, bio);
364 }
365 
366 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
367 				struct request *next)
368 {
369 	int total_phys_segments;
370 	int total_hw_segments;
371 
372 	/*
373 	 * First check if the either of the requests are re-queued
374 	 * requests.  Can't merge them if they are.
375 	 */
376 	if (req->special || next->special)
377 		return 0;
378 
379 	/*
380 	 * Will it become too large?
381 	 */
382 	if ((req->nr_sectors + next->nr_sectors) > q->max_sectors)
383 		return 0;
384 
385 	total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
386 	if (blk_phys_contig_segment(q, req->biotail, next->bio))
387 		total_phys_segments--;
388 
389 	if (total_phys_segments > q->max_phys_segments)
390 		return 0;
391 
392 	total_hw_segments = req->nr_hw_segments + next->nr_hw_segments;
393 	if (blk_hw_contig_segment(q, req->biotail, next->bio)) {
394 		int len = req->biotail->bi_hw_back_size +
395 				next->bio->bi_hw_front_size;
396 		/*
397 		 * propagate the combined length to the end of the requests
398 		 */
399 		if (req->nr_hw_segments == 1)
400 			req->bio->bi_hw_front_size = len;
401 		if (next->nr_hw_segments == 1)
402 			next->biotail->bi_hw_back_size = len;
403 		total_hw_segments--;
404 	}
405 
406 	if (total_hw_segments > q->max_hw_segments)
407 		return 0;
408 
409 	/* Merge is OK... */
410 	req->nr_phys_segments = total_phys_segments;
411 	req->nr_hw_segments = total_hw_segments;
412 	return 1;
413 }
414 
415 /*
416  * Has to be called with the request spinlock acquired
417  */
418 static int attempt_merge(struct request_queue *q, struct request *req,
419 			  struct request *next)
420 {
421 	if (!rq_mergeable(req) || !rq_mergeable(next))
422 		return 0;
423 
424 	/*
425 	 * not contiguous
426 	 */
427 	if (req->sector + req->nr_sectors != next->sector)
428 		return 0;
429 
430 	if (rq_data_dir(req) != rq_data_dir(next)
431 	    || req->rq_disk != next->rq_disk
432 	    || next->special)
433 		return 0;
434 
435 	/*
436 	 * If we are allowed to merge, then append bio list
437 	 * from next to rq and release next. merge_requests_fn
438 	 * will have updated segment counts, update sector
439 	 * counts here.
440 	 */
441 	if (!ll_merge_requests_fn(q, req, next))
442 		return 0;
443 
444 	/*
445 	 * At this point we have either done a back merge
446 	 * or front merge. We need the smaller start_time of
447 	 * the merged requests to be the current request
448 	 * for accounting purposes.
449 	 */
450 	if (time_after(req->start_time, next->start_time))
451 		req->start_time = next->start_time;
452 
453 	req->biotail->bi_next = next->bio;
454 	req->biotail = next->biotail;
455 
456 	req->nr_sectors = req->hard_nr_sectors += next->hard_nr_sectors;
457 
458 	elv_merge_requests(q, req, next);
459 
460 	if (req->rq_disk) {
461 		struct hd_struct *part
462 			= get_part(req->rq_disk, req->sector);
463 		disk_round_stats(req->rq_disk);
464 		req->rq_disk->in_flight--;
465 		if (part) {
466 			part_round_stats(part);
467 			part->in_flight--;
468 		}
469 	}
470 
471 	req->ioprio = ioprio_best(req->ioprio, next->ioprio);
472 
473 	__blk_put_request(q, next);
474 	return 1;
475 }
476 
477 int attempt_back_merge(struct request_queue *q, struct request *rq)
478 {
479 	struct request *next = elv_latter_request(q, rq);
480 
481 	if (next)
482 		return attempt_merge(q, rq, next);
483 
484 	return 0;
485 }
486 
487 int attempt_front_merge(struct request_queue *q, struct request *rq)
488 {
489 	struct request *prev = elv_former_request(q, rq);
490 
491 	if (prev)
492 		return attempt_merge(q, prev, rq);
493 
494 	return 0;
495 }
496