2 * linux/fs/nfs/blocklayout/blocklayout.h
4 * Module for the NFSv4.1 pNFS block layout driver.
6 * Copyright (c) 2006 The Regents of the University of Michigan.
9 * Andy Adamson <andros@citi.umich.edu>
10 * Fred Isaman <iisaman@umich.edu>
12 * permission is granted to use, copy, create derivative works and
13 * redistribute this software and such derivative works for any purpose,
14 * so long as the name of the university of michigan is not used in
15 * any advertising or publicity pertaining to the use or distribution
16 * of this software without specific, written prior authorization. if
17 * the above copyright notice or any other identification of the
18 * university of michigan is included in any copy of any portion of
19 * this software, then the disclaimer below must also be included.
21 * this software is provided as is, without representation from the
22 * university of michigan as to its fitness for any purpose, and without
23 * warranty by the university of michigan of any kind, either express
24 * or implied, including without limitation the implied warranties of
25 * merchantability and fitness for a particular purpose. the regents
26 * of the university of michigan shall not be liable for any damages,
27 * including special, indirect, incidental, or consequential damages,
28 * with respect to any claim arising out or in connection with the use
29 * of the software, even if it has been or is hereafter advised of the
30 * possibility of such damages.
33 #include "blocklayout.h"
34 #define NFSDBG_FACILITY NFSDBG_PNFS_LD
37 #define EXTENT_INITIALIZED 0
38 #define EXTENT_WRITTEN 1
39 #define EXTENT_IN_COMMIT 2
40 #define INTERNAL_EXISTS MY_MAX_TAGS
41 #define INTERNAL_MASK ((1 << INTERNAL_EXISTS) - 1)
43 /* Returns largest t<=s s.t. t%base==0 */
44 static inline sector_t normalize(sector_t s, int base)
46 sector_t tmp = s; /* Since do_div modifies its argument */
47 return s - do_div(tmp, base);
50 static inline sector_t normalize_up(sector_t s, int base)
52 return normalize(s + base - 1, base);
55 /* Complete stub using list while determine API wanted */
57 /* Returns tags, or negative */
58 static int32_t _find_entry(struct my_tree *tree, u64 s)
60 struct pnfs_inval_tracking *pos;
62 dprintk("%s(%llu) enter\n", __func__, s);
63 list_for_each_entry_reverse(pos, &tree->mtt_stub, it_link) {
64 if (pos->it_sector > s)
66 else if (pos->it_sector == s)
67 return pos->it_tags & INTERNAL_MASK;
75 int _has_tag(struct my_tree *tree, u64 s, int32_t tag)
79 dprintk("%s(%llu, %i) enter\n", __func__, s, tag);
80 s = normalize(s, tree->mtt_step_size);
81 tags = _find_entry(tree, s);
82 if ((tags < 0) || !(tags & (1 << tag)))
88 /* Creates entry with tag, or if entry already exists, unions tag to it.
89 * If storage is not NULL, newly created entry will use it.
90 * Returns number of entries added, or negative on error.
92 static int _add_entry(struct my_tree *tree, u64 s, int32_t tag,
93 struct pnfs_inval_tracking *storage)
96 struct pnfs_inval_tracking *pos;
98 dprintk("%s(%llu, %i, %p) enter\n", __func__, s, tag, storage);
99 list_for_each_entry_reverse(pos, &tree->mtt_stub, it_link) {
100 if (pos->it_sector > s)
102 else if (pos->it_sector == s) {
109 pos->it_tags |= (1 << tag);
112 struct pnfs_inval_tracking *new;
116 new = kmalloc(sizeof(*new), GFP_NOFS);
121 new->it_tags = (1 << tag);
122 list_add(&new->it_link, &pos->it_link);
127 /* XXXX Really want option to not create */
128 /* Over range, unions tag with existing entries, else creates entry with tag */
129 static int _set_range(struct my_tree *tree, int32_t tag, u64 s, u64 length)
133 dprintk("%s(%i, %llu, %llu) enter\n", __func__, tag, s, length);
134 for (i = normalize(s, tree->mtt_step_size); i < s + length;
135 i += tree->mtt_step_size)
136 if (_add_entry(tree, i, tag, NULL))
141 /* Ensure that future operations on given range of tree will not malloc */
142 static int _preload_range(struct my_tree *tree, u64 offset, u64 length)
145 int count, i, used = 0, status = -ENOMEM;
146 struct pnfs_inval_tracking **storage;
148 dprintk("%s(%llu, %llu) enter\n", __func__, offset, length);
149 start = normalize(offset, tree->mtt_step_size);
150 end = normalize_up(offset + length, tree->mtt_step_size);
151 count = (int)(end - start) / (int)tree->mtt_step_size;
153 /* Pre-malloc what memory we might need */
154 storage = kmalloc(sizeof(*storage) * count, GFP_NOFS);
157 for (i = 0; i < count; i++) {
158 storage[i] = kmalloc(sizeof(struct pnfs_inval_tracking),
164 /* Now need lock - HOW??? */
166 for (s = start; s < end; s += tree->mtt_step_size)
167 used += _add_entry(tree, s, INTERNAL_EXISTS, storage[used]);
169 /* Unlock - HOW??? */
173 for (i = used; i < count; i++) {
182 static void set_needs_init(sector_t *array, sector_t offset)
186 dprintk("%s enter\n", __func__);
199 dprintk("%s Adding %llu\n", __func__, (u64)offset);
203 memmove(save + 1, save, (char *)p - (char *)save);
209 /* We are relying on page lock to serialize this */
210 int bl_is_sector_init(struct pnfs_inval_markings *marks, sector_t isect)
214 spin_lock(&marks->im_lock);
215 rv = _has_tag(&marks->im_tree, isect, EXTENT_INITIALIZED);
216 spin_unlock(&marks->im_lock);
220 /* Marks sectors in [offest, offset_length) as having been initialized.
221 * All lengths are step-aligned, where step is min(pagesize, blocksize).
222 * Notes where partial block is initialized, and helps prepare it for
223 * complete initialization later.
225 /* Currently assumes offset is page-aligned */
226 int bl_mark_sectors_init(struct pnfs_inval_markings *marks,
227 sector_t offset, sector_t length,
230 sector_t s, start, end;
231 sector_t *array = NULL; /* Pages to mark */
233 dprintk("%s(offset=%llu,len=%llu) enter\n",
234 __func__, (u64)offset, (u64)length);
235 s = max((sector_t) 3,
236 2 * (marks->im_block_size / (PAGE_CACHE_SECTORS)));
237 dprintk("%s set max=%llu\n", __func__, (u64)s);
239 array = kmalloc(s * sizeof(sector_t), GFP_NOFS);
245 start = normalize(offset, marks->im_block_size);
246 end = normalize_up(offset + length, marks->im_block_size);
247 if (_preload_range(&marks->im_tree, start, end - start))
250 spin_lock(&marks->im_lock);
252 for (s = normalize_up(start, PAGE_CACHE_SECTORS);
253 s < offset; s += PAGE_CACHE_SECTORS) {
254 dprintk("%s pre-area pages\n", __func__);
255 /* Portion of used block is not initialized */
256 if (!_has_tag(&marks->im_tree, s, EXTENT_INITIALIZED))
257 set_needs_init(array, s);
259 if (_set_range(&marks->im_tree, EXTENT_INITIALIZED, offset, length))
261 for (s = normalize_up(offset + length, PAGE_CACHE_SECTORS);
262 s < end; s += PAGE_CACHE_SECTORS) {
263 dprintk("%s post-area pages\n", __func__);
264 if (!_has_tag(&marks->im_tree, s, EXTENT_INITIALIZED))
265 set_needs_init(array, s);
268 spin_unlock(&marks->im_lock);
271 if (array[0] == ~0) {
280 spin_unlock(&marks->im_lock);
289 static void print_bl_extent(struct pnfs_block_extent *be)
291 dprintk("PRINT EXTENT extent %p\n", be);
293 dprintk(" be_f_offset %llu\n", (u64)be->be_f_offset);
294 dprintk(" be_length %llu\n", (u64)be->be_length);
295 dprintk(" be_v_offset %llu\n", (u64)be->be_v_offset);
296 dprintk(" be_state %d\n", be->be_state);
301 destroy_extent(struct kref *kref)
303 struct pnfs_block_extent *be;
305 be = container_of(kref, struct pnfs_block_extent, be_refcnt);
306 dprintk("%s be=%p\n", __func__, be);
311 bl_put_extent(struct pnfs_block_extent *be)
314 dprintk("%s enter %p (%i)\n", __func__, be,
315 atomic_read(&be->be_refcnt.refcount));
316 kref_put(&be->be_refcnt, destroy_extent);
320 struct pnfs_block_extent *bl_alloc_extent(void)
322 struct pnfs_block_extent *be;
324 be = kmalloc(sizeof(struct pnfs_block_extent), GFP_NOFS);
327 INIT_LIST_HEAD(&be->be_node);
328 kref_init(&be->be_refcnt);
333 static void print_elist(struct list_head *list)
335 struct pnfs_block_extent *be;
336 dprintk("****************\n");
337 dprintk("Extent list looks like:\n");
338 list_for_each_entry(be, list, be_node) {
341 dprintk("****************\n");
345 extents_consistent(struct pnfs_block_extent *old, struct pnfs_block_extent *new)
347 /* Note this assumes new->be_f_offset >= old->be_f_offset */
348 return (new->be_state == old->be_state) &&
349 ((new->be_state == PNFS_BLOCK_NONE_DATA) ||
350 ((new->be_v_offset - old->be_v_offset ==
351 new->be_f_offset - old->be_f_offset) &&
352 new->be_mdev == old->be_mdev));
355 /* Adds new to appropriate list in bl, modifying new and removing existing
356 * extents as appropriate to deal with overlaps.
358 * See bl_find_get_extent for list constraints.
360 * Refcount on new is already set. If end up not using it, or error out,
361 * need to put the reference.
363 * bl->bl_ext_lock is held by caller.
366 bl_add_merge_extent(struct pnfs_block_layout *bl,
367 struct pnfs_block_extent *new)
369 struct pnfs_block_extent *be, *tmp;
370 sector_t end = new->be_f_offset + new->be_length;
371 struct list_head *list;
373 dprintk("%s enter with be=%p\n", __func__, new);
374 print_bl_extent(new);
375 list = &bl->bl_extents[bl_choose_list(new->be_state)];
378 /* Scan for proper place to insert, extending new to the left
379 * as much as possible.
381 list_for_each_entry_safe(be, tmp, list, be_node) {
382 if (new->be_f_offset < be->be_f_offset)
384 if (end <= be->be_f_offset + be->be_length) {
385 /* new is a subset of existing be*/
386 if (extents_consistent(be, new)) {
387 dprintk("%s: new is subset, ignoring\n",
393 } else if (new->be_f_offset <=
394 be->be_f_offset + be->be_length) {
395 /* new overlaps or abuts existing be */
396 if (extents_consistent(be, new)) {
397 /* extend new to fully replace be */
398 new->be_length += new->be_f_offset -
400 new->be_f_offset = be->be_f_offset;
401 new->be_v_offset = be->be_v_offset;
402 dprintk("%s: removing %p\n", __func__, be);
403 list_del(&be->be_node);
405 } else if (new->be_f_offset !=
406 be->be_f_offset + be->be_length)
410 /* Note that if we never hit the above break, be will not point to a
411 * valid extent. However, in that case &be->be_node==list.
413 list_add_tail(&new->be_node, &be->be_node);
414 dprintk("%s: inserting new\n", __func__);
416 /* Scan forward for overlaps. If we find any, extend new and
417 * remove the overlapped extent.
419 be = list_prepare_entry(new, list, be_node);
420 list_for_each_entry_safe_continue(be, tmp, list, be_node) {
421 if (end < be->be_f_offset)
423 /* new overlaps or abuts existing be */
424 if (extents_consistent(be, new)) {
425 if (end < be->be_f_offset + be->be_length) {
426 /* extend new to fully cover be */
427 end = be->be_f_offset + be->be_length;
428 new->be_length = end - new->be_f_offset;
430 dprintk("%s: removing %p\n", __func__, be);
431 list_del(&be->be_node);
433 } else if (end != be->be_f_offset) {
434 list_del(&new->be_node);
438 dprintk("%s: after merging\n", __func__);
440 /* FIXME - The per-list consistency checks have all been done,
441 * should now check cross-list consistency.
450 /* Returns extent, or NULL. If a second READ extent exists, it is returned
451 * in cow_read, if given.
453 * The extents are kept in two seperate ordered lists, one for READ and NONE,
454 * one for READWRITE and INVALID. Within each list, we assume:
455 * 1. Extents are ordered by file offset.
456 * 2. For any given isect, there is at most one extents that matches.
458 struct pnfs_block_extent *
459 bl_find_get_extent(struct pnfs_block_layout *bl, sector_t isect,
460 struct pnfs_block_extent **cow_read)
462 struct pnfs_block_extent *be, *cow, *ret;
465 dprintk("%s enter with isect %llu\n", __func__, (u64)isect);
467 spin_lock(&bl->bl_ext_lock);
468 for (i = 0; i < EXTENT_LISTS; i++) {
469 list_for_each_entry_reverse(be, &bl->bl_extents[i], be_node) {
470 if (isect >= be->be_f_offset + be->be_length)
472 if (isect >= be->be_f_offset) {
473 /* We have found an extent */
474 dprintk("%s Get %p (%i)\n", __func__, be,
475 atomic_read(&be->be_refcnt.refcount));
476 kref_get(&be->be_refcnt);
479 else if (be->be_state != PNFS_BLOCK_READ_DATA)
487 (!cow_read || ret->be_state != PNFS_BLOCK_INVALID_DATA))
490 spin_unlock(&bl->bl_ext_lock);
493 print_bl_extent(ret);
497 /* Helper function to set_to_rw that initialize a new extent */
499 _prep_new_extent(struct pnfs_block_extent *new,
500 struct pnfs_block_extent *orig,
501 sector_t offset, sector_t length, int state)
503 kref_init(&new->be_refcnt);
504 /* don't need to INIT_LIST_HEAD(&new->be_node) */
505 memcpy(&new->be_devid, &orig->be_devid, sizeof(struct nfs4_deviceid));
506 new->be_mdev = orig->be_mdev;
507 new->be_f_offset = offset;
508 new->be_length = length;
509 new->be_v_offset = orig->be_v_offset - orig->be_f_offset + offset;
510 new->be_state = state;
511 new->be_inval = orig->be_inval;
514 /* Tries to merge be with extent in front of it in list.
515 * Frees storage if not used.
517 static struct pnfs_block_extent *
518 _front_merge(struct pnfs_block_extent *be, struct list_head *head,
519 struct pnfs_block_extent *storage)
521 struct pnfs_block_extent *prev;
525 if (&be->be_node == head || be->be_node.prev == head)
527 prev = list_entry(be->be_node.prev, struct pnfs_block_extent, be_node);
528 if ((prev->be_f_offset + prev->be_length != be->be_f_offset) ||
529 !extents_consistent(prev, be))
531 _prep_new_extent(storage, prev, prev->be_f_offset,
532 prev->be_length + be->be_length, prev->be_state);
533 list_replace(&prev->be_node, &storage->be_node);
535 list_del(&be->be_node);