1/*
2 * Copyright (C) 2011 STRATO.  All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18
19#include <linux/sched.h>
20#include <linux/pagemap.h>
21#include <linux/writeback.h>
22#include <linux/blkdev.h>
23#include <linux/rbtree.h>
24#include <linux/slab.h>
25#include <linux/workqueue.h>
26#include "ctree.h"
27#include "volumes.h"
28#include "disk-io.h"
29#include "transaction.h"
30#include "dev-replace.h"
31
32#undef DEBUG
33
34/*
35 * This is the implementation for the generic read ahead framework.
36 *
37 * To trigger a readahead, btrfs_reada_add must be called. It will start
38 * a read ahead for the given range [start, end) on tree root. The returned
39 * handle can either be used to wait on the readahead to finish
40 * (btrfs_reada_wait), or to send it to the background (btrfs_reada_detach).
41 *
42 * The read ahead works as follows:
43 * On btrfs_reada_add, the root of the tree is inserted into a radix_tree.
44 * reada_start_machine will then search for extents to prefetch and trigger
45 * some reads. When a read finishes for a node, all contained node/leaf
46 * pointers that lie in the given range will also be enqueued. The reads will
47 * be triggered in sequential order, thus giving a big win over a naive
48 * enumeration. It will also make use of multi-device layouts. Each disk
49 * will have its on read pointer and all disks will by utilized in parallel.
50 * Also will no two disks read both sides of a mirror simultaneously, as this
51 * would waste seeking capacity. Instead both disks will read different parts
52 * of the filesystem.
53 * Any number of readaheads can be started in parallel. The read order will be
54 * determined globally, i.e. 2 parallel readaheads will normally finish faster
55 * than the 2 started one after another.
56 */
57
58#define MAX_IN_FLIGHT 6
59
60struct reada_extctl {
61	struct list_head	list;
62	struct reada_control	*rc;
63	u64			generation;
64};
65
66struct reada_extent {
67	u64			logical;
68	struct btrfs_key	top;
69	int			err;
70	struct list_head	extctl;
71	int 			refcnt;
72	spinlock_t		lock;
73	struct reada_zone	*zones[BTRFS_MAX_MIRRORS];
74	int			nzones;
75	struct btrfs_device	*scheduled_for;
76};
77
78struct reada_zone {
79	u64			start;
80	u64			end;
81	u64			elems;
82	struct list_head	list;
83	spinlock_t		lock;
84	int			locked;
85	struct btrfs_device	*device;
86	struct btrfs_device	*devs[BTRFS_MAX_MIRRORS]; /* full list, incl
87							   * self */
88	int			ndevs;
89	struct kref		refcnt;
90};
91
92struct reada_machine_work {
93	struct btrfs_work	work;
94	struct btrfs_fs_info	*fs_info;
95};
96
97static void reada_extent_put(struct btrfs_fs_info *, struct reada_extent *);
98static void reada_control_release(struct kref *kref);
99static void reada_zone_release(struct kref *kref);
100static void reada_start_machine(struct btrfs_fs_info *fs_info);
101static void __reada_start_machine(struct btrfs_fs_info *fs_info);
102
103static int reada_add_block(struct reada_control *rc, u64 logical,
104			   struct btrfs_key *top, int level, u64 generation);
105
106/* recurses */
107/* in case of err, eb might be NULL */
108static int __readahead_hook(struct btrfs_root *root, struct extent_buffer *eb,
109			    u64 start, int err)
110{
111	int level = 0;
112	int nritems;
113	int i;
114	u64 bytenr;
115	u64 generation;
116	struct reada_extent *re;
117	struct btrfs_fs_info *fs_info = root->fs_info;
118	struct list_head list;
119	unsigned long index = start >> PAGE_CACHE_SHIFT;
120	struct btrfs_device *for_dev;
121
122	if (eb)
123		level = btrfs_header_level(eb);
124
125	/* find extent */
126	spin_lock(&fs_info->reada_lock);
127	re = radix_tree_lookup(&fs_info->reada_tree, index);
128	if (re)
129		re->refcnt++;
130	spin_unlock(&fs_info->reada_lock);
131
132	if (!re)
133		return -1;
134
135	spin_lock(&re->lock);
136	/*
137	 * just take the full list from the extent. afterwards we
138	 * don't need the lock anymore
139	 */
140	list_replace_init(&re->extctl, &list);
141	for_dev = re->scheduled_for;
142	re->scheduled_for = NULL;
143	spin_unlock(&re->lock);
144
145	if (err == 0) {
146		nritems = level ? btrfs_header_nritems(eb) : 0;
147		generation = btrfs_header_generation(eb);
148		/*
149		 * FIXME: currently we just set nritems to 0 if this is a leaf,
150		 * effectively ignoring the content. In a next step we could
151		 * trigger more readahead depending from the content, e.g.
152		 * fetch the checksums for the extents in the leaf.
153		 */
154	} else {
155		/*
156		 * this is the error case, the extent buffer has not been
157		 * read correctly. We won't access anything from it and
158		 * just cleanup our data structures. Effectively this will
159		 * cut the branch below this node from read ahead.
160		 */
161		nritems = 0;
162		generation = 0;
163	}
164
165	for (i = 0; i < nritems; i++) {
166		struct reada_extctl *rec;
167		u64 n_gen;
168		struct btrfs_key key;
169		struct btrfs_key next_key;
170
171		btrfs_node_key_to_cpu(eb, &key, i);
172		if (i + 1 < nritems)
173			btrfs_node_key_to_cpu(eb, &next_key, i + 1);
174		else
175			next_key = re->top;
176		bytenr = btrfs_node_blockptr(eb, i);
177		n_gen = btrfs_node_ptr_generation(eb, i);
178
179		list_for_each_entry(rec, &list, list) {
180			struct reada_control *rc = rec->rc;
181
182			/*
183			 * if the generation doesn't match, just ignore this
184			 * extctl. This will probably cut off a branch from
185			 * prefetch. Alternatively one could start a new (sub-)
186			 * prefetch for this branch, starting again from root.
187			 * FIXME: move the generation check out of this loop
188			 */
189#ifdef DEBUG
190			if (rec->generation != generation) {
191				btrfs_debug(root->fs_info,
192					   "generation mismatch for (%llu,%d,%llu) %llu != %llu",
193				       key.objectid, key.type, key.offset,
194				       rec->generation, generation);
195			}
196#endif
197			if (rec->generation == generation &&
198			    btrfs_comp_cpu_keys(&key, &rc->key_end) < 0 &&
199			    btrfs_comp_cpu_keys(&next_key, &rc->key_start) > 0)
200				reada_add_block(rc, bytenr, &next_key,
201						level - 1, n_gen);
202		}
203	}
204	/*
205	 * free extctl records
206	 */
207	while (!list_empty(&list)) {
208		struct reada_control *rc;
209		struct reada_extctl *rec;
210
211		rec = list_first_entry(&list, struct reada_extctl, list);
212		list_del(&rec->list);
213		rc = rec->rc;
214		kfree(rec);
215
216		kref_get(&rc->refcnt);
217		if (atomic_dec_and_test(&rc->elems)) {
218			kref_put(&rc->refcnt, reada_control_release);
219			wake_up(&rc->wait);
220		}
221		kref_put(&rc->refcnt, reada_control_release);
222
223		reada_extent_put(fs_info, re);	/* one ref for each entry */
224	}
225	reada_extent_put(fs_info, re);	/* our ref */
226	if (for_dev)
227		atomic_dec(&for_dev->reada_in_flight);
228
229	return 0;
230}
231
232/*
233 * start is passed separately in case eb in NULL, which may be the case with
234 * failed I/O
235 */
236int btree_readahead_hook(struct btrfs_root *root, struct extent_buffer *eb,
237			 u64 start, int err)
238{
239	int ret;
240
241	ret = __readahead_hook(root, eb, start, err);
242
243	reada_start_machine(root->fs_info);
244
245	return ret;
246}
247
248static struct reada_zone *reada_find_zone(struct btrfs_fs_info *fs_info,
249					  struct btrfs_device *dev, u64 logical,
250					  struct btrfs_bio *bbio)
251{
252	int ret;
253	struct reada_zone *zone;
254	struct btrfs_block_group_cache *cache = NULL;
255	u64 start;
256	u64 end;
257	int i;
258
259	zone = NULL;
260	spin_lock(&fs_info->reada_lock);
261	ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
262				     logical >> PAGE_CACHE_SHIFT, 1);
263	if (ret == 1)
264		kref_get(&zone->refcnt);
265	spin_unlock(&fs_info->reada_lock);
266
267	if (ret == 1) {
268		if (logical >= zone->start && logical < zone->end)
269			return zone;
270		spin_lock(&fs_info->reada_lock);
271		kref_put(&zone->refcnt, reada_zone_release);
272		spin_unlock(&fs_info->reada_lock);
273	}
274
275	cache = btrfs_lookup_block_group(fs_info, logical);
276	if (!cache)
277		return NULL;
278
279	start = cache->key.objectid;
280	end = start + cache->key.offset - 1;
281	btrfs_put_block_group(cache);
282
283	zone = kzalloc(sizeof(*zone), GFP_NOFS);
284	if (!zone)
285		return NULL;
286
287	zone->start = start;
288	zone->end = end;
289	INIT_LIST_HEAD(&zone->list);
290	spin_lock_init(&zone->lock);
291	zone->locked = 0;
292	kref_init(&zone->refcnt);
293	zone->elems = 0;
294	zone->device = dev; /* our device always sits at index 0 */
295	for (i = 0; i < bbio->num_stripes; ++i) {
296		/* bounds have already been checked */
297		zone->devs[i] = bbio->stripes[i].dev;
298	}
299	zone->ndevs = bbio->num_stripes;
300
301	spin_lock(&fs_info->reada_lock);
302	ret = radix_tree_insert(&dev->reada_zones,
303				(unsigned long)(zone->end >> PAGE_CACHE_SHIFT),
304				zone);
305
306	if (ret == -EEXIST) {
307		kfree(zone);
308		ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
309					     logical >> PAGE_CACHE_SHIFT, 1);
310		if (ret == 1)
311			kref_get(&zone->refcnt);
312	}
313	spin_unlock(&fs_info->reada_lock);
314
315	return zone;
316}
317
318static struct reada_extent *reada_find_extent(struct btrfs_root *root,
319					      u64 logical,
320					      struct btrfs_key *top, int level)
321{
322	int ret;
323	struct reada_extent *re = NULL;
324	struct reada_extent *re_exist = NULL;
325	struct btrfs_fs_info *fs_info = root->fs_info;
326	struct btrfs_bio *bbio = NULL;
327	struct btrfs_device *dev;
328	struct btrfs_device *prev_dev;
329	u32 blocksize;
330	u64 length;
331	int nzones = 0;
332	int i;
333	unsigned long index = logical >> PAGE_CACHE_SHIFT;
334	int dev_replace_is_ongoing;
335
336	spin_lock(&fs_info->reada_lock);
337	re = radix_tree_lookup(&fs_info->reada_tree, index);
338	if (re)
339		re->refcnt++;
340	spin_unlock(&fs_info->reada_lock);
341
342	if (re)
343		return re;
344
345	re = kzalloc(sizeof(*re), GFP_NOFS);
346	if (!re)
347		return NULL;
348
349	blocksize = root->nodesize;
350	re->logical = logical;
351	re->top = *top;
352	INIT_LIST_HEAD(&re->extctl);
353	spin_lock_init(&re->lock);
354	re->refcnt = 1;
355
356	/*
357	 * map block
358	 */
359	length = blocksize;
360	ret = btrfs_map_block(fs_info, REQ_GET_READ_MIRRORS, logical, &length,
361			      &bbio, 0);
362	if (ret || !bbio || length < blocksize)
363		goto error;
364
365	if (bbio->num_stripes > BTRFS_MAX_MIRRORS) {
366		btrfs_err(root->fs_info,
367			   "readahead: more than %d copies not supported",
368			   BTRFS_MAX_MIRRORS);
369		goto error;
370	}
371
372	for (nzones = 0; nzones < bbio->num_stripes; ++nzones) {
373		struct reada_zone *zone;
374
375		dev = bbio->stripes[nzones].dev;
376		zone = reada_find_zone(fs_info, dev, logical, bbio);
377		if (!zone)
378			break;
379
380		re->zones[nzones] = zone;
381		spin_lock(&zone->lock);
382		if (!zone->elems)
383			kref_get(&zone->refcnt);
384		++zone->elems;
385		spin_unlock(&zone->lock);
386		spin_lock(&fs_info->reada_lock);
387		kref_put(&zone->refcnt, reada_zone_release);
388		spin_unlock(&fs_info->reada_lock);
389	}
390	re->nzones = nzones;
391	if (nzones == 0) {
392		/* not a single zone found, error and out */
393		goto error;
394	}
395
396	/* insert extent in reada_tree + all per-device trees, all or nothing */
397	btrfs_dev_replace_lock(&fs_info->dev_replace);
398	spin_lock(&fs_info->reada_lock);
399	ret = radix_tree_insert(&fs_info->reada_tree, index, re);
400	if (ret == -EEXIST) {
401		re_exist = radix_tree_lookup(&fs_info->reada_tree, index);
402		BUG_ON(!re_exist);
403		re_exist->refcnt++;
404		spin_unlock(&fs_info->reada_lock);
405		btrfs_dev_replace_unlock(&fs_info->dev_replace);
406		goto error;
407	}
408	if (ret) {
409		spin_unlock(&fs_info->reada_lock);
410		btrfs_dev_replace_unlock(&fs_info->dev_replace);
411		goto error;
412	}
413	prev_dev = NULL;
414	dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(
415			&fs_info->dev_replace);
416	for (i = 0; i < nzones; ++i) {
417		dev = bbio->stripes[i].dev;
418		if (dev == prev_dev) {
419			/*
420			 * in case of DUP, just add the first zone. As both
421			 * are on the same device, there's nothing to gain
422			 * from adding both.
423			 * Also, it wouldn't work, as the tree is per device
424			 * and adding would fail with EEXIST
425			 */
426			continue;
427		}
428		if (!dev->bdev) {
429			/*
430			 * cannot read ahead on missing device, but for RAID5/6,
431			 * REQ_GET_READ_MIRRORS return 1. So don't skip missing
432			 * device for such case.
433			 */
434			if (nzones > 1)
435				continue;
436		}
437		if (dev_replace_is_ongoing &&
438		    dev == fs_info->dev_replace.tgtdev) {
439			/*
440			 * as this device is selected for reading only as
441			 * a last resort, skip it for read ahead.
442			 */
443			continue;
444		}
445		prev_dev = dev;
446		ret = radix_tree_insert(&dev->reada_extents, index, re);
447		if (ret) {
448			while (--i >= 0) {
449				dev = bbio->stripes[i].dev;
450				BUG_ON(dev == NULL);
451				/* ignore whether the entry was inserted */
452				radix_tree_delete(&dev->reada_extents, index);
453			}
454			BUG_ON(fs_info == NULL);
455			radix_tree_delete(&fs_info->reada_tree, index);
456			spin_unlock(&fs_info->reada_lock);
457			btrfs_dev_replace_unlock(&fs_info->dev_replace);
458			goto error;
459		}
460	}
461	spin_unlock(&fs_info->reada_lock);
462	btrfs_dev_replace_unlock(&fs_info->dev_replace);
463
464	btrfs_put_bbio(bbio);
465	return re;
466
467error:
468	while (nzones) {
469		struct reada_zone *zone;
470
471		--nzones;
472		zone = re->zones[nzones];
473		kref_get(&zone->refcnt);
474		spin_lock(&zone->lock);
475		--zone->elems;
476		if (zone->elems == 0) {
477			/*
478			 * no fs_info->reada_lock needed, as this can't be
479			 * the last ref
480			 */
481			kref_put(&zone->refcnt, reada_zone_release);
482		}
483		spin_unlock(&zone->lock);
484
485		spin_lock(&fs_info->reada_lock);
486		kref_put(&zone->refcnt, reada_zone_release);
487		spin_unlock(&fs_info->reada_lock);
488	}
489	btrfs_put_bbio(bbio);
490	kfree(re);
491	return re_exist;
492}
493
494static void reada_extent_put(struct btrfs_fs_info *fs_info,
495			     struct reada_extent *re)
496{
497	int i;
498	unsigned long index = re->logical >> PAGE_CACHE_SHIFT;
499
500	spin_lock(&fs_info->reada_lock);
501	if (--re->refcnt) {
502		spin_unlock(&fs_info->reada_lock);
503		return;
504	}
505
506	radix_tree_delete(&fs_info->reada_tree, index);
507	for (i = 0; i < re->nzones; ++i) {
508		struct reada_zone *zone = re->zones[i];
509
510		radix_tree_delete(&zone->device->reada_extents, index);
511	}
512
513	spin_unlock(&fs_info->reada_lock);
514
515	for (i = 0; i < re->nzones; ++i) {
516		struct reada_zone *zone = re->zones[i];
517
518		kref_get(&zone->refcnt);
519		spin_lock(&zone->lock);
520		--zone->elems;
521		if (zone->elems == 0) {
522			/* no fs_info->reada_lock needed, as this can't be
523			 * the last ref */
524			kref_put(&zone->refcnt, reada_zone_release);
525		}
526		spin_unlock(&zone->lock);
527
528		spin_lock(&fs_info->reada_lock);
529		kref_put(&zone->refcnt, reada_zone_release);
530		spin_unlock(&fs_info->reada_lock);
531	}
532	if (re->scheduled_for)
533		atomic_dec(&re->scheduled_for->reada_in_flight);
534
535	kfree(re);
536}
537
538static void reada_zone_release(struct kref *kref)
539{
540	struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
541
542	radix_tree_delete(&zone->device->reada_zones,
543			  zone->end >> PAGE_CACHE_SHIFT);
544
545	kfree(zone);
546}
547
548static void reada_control_release(struct kref *kref)
549{
550	struct reada_control *rc = container_of(kref, struct reada_control,
551						refcnt);
552
553	kfree(rc);
554}
555
556static int reada_add_block(struct reada_control *rc, u64 logical,
557			   struct btrfs_key *top, int level, u64 generation)
558{
559	struct btrfs_root *root = rc->root;
560	struct reada_extent *re;
561	struct reada_extctl *rec;
562
563	re = reada_find_extent(root, logical, top, level); /* takes one ref */
564	if (!re)
565		return -1;
566
567	rec = kzalloc(sizeof(*rec), GFP_NOFS);
568	if (!rec) {
569		reada_extent_put(root->fs_info, re);
570		return -1;
571	}
572
573	rec->rc = rc;
574	rec->generation = generation;
575	atomic_inc(&rc->elems);
576
577	spin_lock(&re->lock);
578	list_add_tail(&rec->list, &re->extctl);
579	spin_unlock(&re->lock);
580
581	/* leave the ref on the extent */
582
583	return 0;
584}
585
586/*
587 * called with fs_info->reada_lock held
588 */
589static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
590{
591	int i;
592	unsigned long index = zone->end >> PAGE_CACHE_SHIFT;
593
594	for (i = 0; i < zone->ndevs; ++i) {
595		struct reada_zone *peer;
596		peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
597		if (peer && peer->device != zone->device)
598			peer->locked = lock;
599	}
600}
601
602/*
603 * called with fs_info->reada_lock held
604 */
605static int reada_pick_zone(struct btrfs_device *dev)
606{
607	struct reada_zone *top_zone = NULL;
608	struct reada_zone *top_locked_zone = NULL;
609	u64 top_elems = 0;
610	u64 top_locked_elems = 0;
611	unsigned long index = 0;
612	int ret;
613
614	if (dev->reada_curr_zone) {
615		reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
616		kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
617		dev->reada_curr_zone = NULL;
618	}
619	/* pick the zone with the most elements */
620	while (1) {
621		struct reada_zone *zone;
622
623		ret = radix_tree_gang_lookup(&dev->reada_zones,
624					     (void **)&zone, index, 1);
625		if (ret == 0)
626			break;
627		index = (zone->end >> PAGE_CACHE_SHIFT) + 1;
628		if (zone->locked) {
629			if (zone->elems > top_locked_elems) {
630				top_locked_elems = zone->elems;
631				top_locked_zone = zone;
632			}
633		} else {
634			if (zone->elems > top_elems) {
635				top_elems = zone->elems;
636				top_zone = zone;
637			}
638		}
639	}
640	if (top_zone)
641		dev->reada_curr_zone = top_zone;
642	else if (top_locked_zone)
643		dev->reada_curr_zone = top_locked_zone;
644	else
645		return 0;
646
647	dev->reada_next = dev->reada_curr_zone->start;
648	kref_get(&dev->reada_curr_zone->refcnt);
649	reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
650
651	return 1;
652}
653
654static int reada_start_machine_dev(struct btrfs_fs_info *fs_info,
655				   struct btrfs_device *dev)
656{
657	struct reada_extent *re = NULL;
658	int mirror_num = 0;
659	struct extent_buffer *eb = NULL;
660	u64 logical;
661	int ret;
662	int i;
663	int need_kick = 0;
664
665	spin_lock(&fs_info->reada_lock);
666	if (dev->reada_curr_zone == NULL) {
667		ret = reada_pick_zone(dev);
668		if (!ret) {
669			spin_unlock(&fs_info->reada_lock);
670			return 0;
671		}
672	}
673	/*
674	 * FIXME currently we issue the reads one extent at a time. If we have
675	 * a contiguous block of extents, we could also coagulate them or use
676	 * plugging to speed things up
677	 */
678	ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
679				     dev->reada_next >> PAGE_CACHE_SHIFT, 1);
680	if (ret == 0 || re->logical >= dev->reada_curr_zone->end) {
681		ret = reada_pick_zone(dev);
682		if (!ret) {
683			spin_unlock(&fs_info->reada_lock);
684			return 0;
685		}
686		re = NULL;
687		ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
688					dev->reada_next >> PAGE_CACHE_SHIFT, 1);
689	}
690	if (ret == 0) {
691		spin_unlock(&fs_info->reada_lock);
692		return 0;
693	}
694	dev->reada_next = re->logical + fs_info->tree_root->nodesize;
695	re->refcnt++;
696
697	spin_unlock(&fs_info->reada_lock);
698
699	/*
700	 * find mirror num
701	 */
702	for (i = 0; i < re->nzones; ++i) {
703		if (re->zones[i]->device == dev) {
704			mirror_num = i + 1;
705			break;
706		}
707	}
708	logical = re->logical;
709
710	spin_lock(&re->lock);
711	if (re->scheduled_for == NULL) {
712		re->scheduled_for = dev;
713		need_kick = 1;
714	}
715	spin_unlock(&re->lock);
716
717	reada_extent_put(fs_info, re);
718
719	if (!need_kick)
720		return 0;
721
722	atomic_inc(&dev->reada_in_flight);
723	ret = reada_tree_block_flagged(fs_info->extent_root, logical,
724			mirror_num, &eb);
725	if (ret)
726		__readahead_hook(fs_info->extent_root, NULL, logical, ret);
727	else if (eb)
728		__readahead_hook(fs_info->extent_root, eb, eb->start, ret);
729
730	if (eb)
731		free_extent_buffer(eb);
732
733	return 1;
734
735}
736
737static void reada_start_machine_worker(struct btrfs_work *work)
738{
739	struct reada_machine_work *rmw;
740	struct btrfs_fs_info *fs_info;
741	int old_ioprio;
742
743	rmw = container_of(work, struct reada_machine_work, work);
744	fs_info = rmw->fs_info;
745
746	kfree(rmw);
747
748	old_ioprio = IOPRIO_PRIO_VALUE(task_nice_ioclass(current),
749				       task_nice_ioprio(current));
750	set_task_ioprio(current, BTRFS_IOPRIO_READA);
751	__reada_start_machine(fs_info);
752	set_task_ioprio(current, old_ioprio);
753}
754
755static void __reada_start_machine(struct btrfs_fs_info *fs_info)
756{
757	struct btrfs_device *device;
758	struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
759	u64 enqueued;
760	u64 total = 0;
761	int i;
762
763	do {
764		enqueued = 0;
765		list_for_each_entry(device, &fs_devices->devices, dev_list) {
766			if (atomic_read(&device->reada_in_flight) <
767			    MAX_IN_FLIGHT)
768				enqueued += reada_start_machine_dev(fs_info,
769								    device);
770		}
771		total += enqueued;
772	} while (enqueued && total < 10000);
773
774	if (enqueued == 0)
775		return;
776
777	/*
778	 * If everything is already in the cache, this is effectively single
779	 * threaded. To a) not hold the caller for too long and b) to utilize
780	 * more cores, we broke the loop above after 10000 iterations and now
781	 * enqueue to workers to finish it. This will distribute the load to
782	 * the cores.
783	 */
784	for (i = 0; i < 2; ++i)
785		reada_start_machine(fs_info);
786}
787
788static void reada_start_machine(struct btrfs_fs_info *fs_info)
789{
790	struct reada_machine_work *rmw;
791
792	rmw = kzalloc(sizeof(*rmw), GFP_NOFS);
793	if (!rmw) {
794		/* FIXME we cannot handle this properly right now */
795		BUG();
796	}
797	btrfs_init_work(&rmw->work, btrfs_readahead_helper,
798			reada_start_machine_worker, NULL, NULL);
799	rmw->fs_info = fs_info;
800
801	btrfs_queue_work(fs_info->readahead_workers, &rmw->work);
802}
803
804#ifdef DEBUG
805static void dump_devs(struct btrfs_fs_info *fs_info, int all)
806{
807	struct btrfs_device *device;
808	struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
809	unsigned long index;
810	int ret;
811	int i;
812	int j;
813	int cnt;
814
815	spin_lock(&fs_info->reada_lock);
816	list_for_each_entry(device, &fs_devices->devices, dev_list) {
817		printk(KERN_DEBUG "dev %lld has %d in flight\n", device->devid,
818			atomic_read(&device->reada_in_flight));
819		index = 0;
820		while (1) {
821			struct reada_zone *zone;
822			ret = radix_tree_gang_lookup(&device->reada_zones,
823						     (void **)&zone, index, 1);
824			if (ret == 0)
825				break;
826			printk(KERN_DEBUG "  zone %llu-%llu elems %llu locked "
827				"%d devs", zone->start, zone->end, zone->elems,
828				zone->locked);
829			for (j = 0; j < zone->ndevs; ++j) {
830				printk(KERN_CONT " %lld",
831					zone->devs[j]->devid);
832			}
833			if (device->reada_curr_zone == zone)
834				printk(KERN_CONT " curr off %llu",
835					device->reada_next - zone->start);
836			printk(KERN_CONT "\n");
837			index = (zone->end >> PAGE_CACHE_SHIFT) + 1;
838		}
839		cnt = 0;
840		index = 0;
841		while (all) {
842			struct reada_extent *re = NULL;
843
844			ret = radix_tree_gang_lookup(&device->reada_extents,
845						     (void **)&re, index, 1);
846			if (ret == 0)
847				break;
848			printk(KERN_DEBUG
849				"  re: logical %llu size %u empty %d for %lld",
850				re->logical, fs_info->tree_root->nodesize,
851				list_empty(&re->extctl), re->scheduled_for ?
852				re->scheduled_for->devid : -1);
853
854			for (i = 0; i < re->nzones; ++i) {
855				printk(KERN_CONT " zone %llu-%llu devs",
856					re->zones[i]->start,
857					re->zones[i]->end);
858				for (j = 0; j < re->zones[i]->ndevs; ++j) {
859					printk(KERN_CONT " %lld",
860						re->zones[i]->devs[j]->devid);
861				}
862			}
863			printk(KERN_CONT "\n");
864			index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
865			if (++cnt > 15)
866				break;
867		}
868	}
869
870	index = 0;
871	cnt = 0;
872	while (all) {
873		struct reada_extent *re = NULL;
874
875		ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
876					     index, 1);
877		if (ret == 0)
878			break;
879		if (!re->scheduled_for) {
880			index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
881			continue;
882		}
883		printk(KERN_DEBUG
884			"re: logical %llu size %u list empty %d for %lld",
885			re->logical, fs_info->tree_root->nodesize,
886			list_empty(&re->extctl),
887			re->scheduled_for ? re->scheduled_for->devid : -1);
888		for (i = 0; i < re->nzones; ++i) {
889			printk(KERN_CONT " zone %llu-%llu devs",
890				re->zones[i]->start,
891				re->zones[i]->end);
892			for (i = 0; i < re->nzones; ++i) {
893				printk(KERN_CONT " zone %llu-%llu devs",
894					re->zones[i]->start,
895					re->zones[i]->end);
896				for (j = 0; j < re->zones[i]->ndevs; ++j) {
897					printk(KERN_CONT " %lld",
898						re->zones[i]->devs[j]->devid);
899				}
900			}
901		}
902		printk(KERN_CONT "\n");
903		index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
904	}
905	spin_unlock(&fs_info->reada_lock);
906}
907#endif
908
909/*
910 * interface
911 */
912struct reada_control *btrfs_reada_add(struct btrfs_root *root,
913			struct btrfs_key *key_start, struct btrfs_key *key_end)
914{
915	struct reada_control *rc;
916	u64 start;
917	u64 generation;
918	int level;
919	struct extent_buffer *node;
920	static struct btrfs_key max_key = {
921		.objectid = (u64)-1,
922		.type = (u8)-1,
923		.offset = (u64)-1
924	};
925
926	rc = kzalloc(sizeof(*rc), GFP_NOFS);
927	if (!rc)
928		return ERR_PTR(-ENOMEM);
929
930	rc->root = root;
931	rc->key_start = *key_start;
932	rc->key_end = *key_end;
933	atomic_set(&rc->elems, 0);
934	init_waitqueue_head(&rc->wait);
935	kref_init(&rc->refcnt);
936	kref_get(&rc->refcnt); /* one ref for having elements */
937
938	node = btrfs_root_node(root);
939	start = node->start;
940	level = btrfs_header_level(node);
941	generation = btrfs_header_generation(node);
942	free_extent_buffer(node);
943
944	if (reada_add_block(rc, start, &max_key, level, generation)) {
945		kfree(rc);
946		return ERR_PTR(-ENOMEM);
947	}
948
949	reada_start_machine(root->fs_info);
950
951	return rc;
952}
953
954#ifdef DEBUG
955int btrfs_reada_wait(void *handle)
956{
957	struct reada_control *rc = handle;
958
959	while (atomic_read(&rc->elems)) {
960		wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
961				   5 * HZ);
962		dump_devs(rc->root->fs_info,
963			  atomic_read(&rc->elems) < 10 ? 1 : 0);
964	}
965
966	dump_devs(rc->root->fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
967
968	kref_put(&rc->refcnt, reada_control_release);
969
970	return 0;
971}
972#else
973int btrfs_reada_wait(void *handle)
974{
975	struct reada_control *rc = handle;
976
977	while (atomic_read(&rc->elems)) {
978		wait_event(rc->wait, atomic_read(&rc->elems) == 0);
979	}
980
981	kref_put(&rc->refcnt, reada_control_release);
982
983	return 0;
984}
985#endif
986
987void btrfs_reada_detach(void *handle)
988{
989	struct reada_control *rc = handle;
990
991	kref_put(&rc->refcnt, reada_control_release);
992}
993