Loading...
1#include <linux/ceph/ceph_debug.h>
2
3#include <linux/backing-dev.h>
4#include <linux/fs.h>
5#include <linux/mm.h>
6#include <linux/pagemap.h>
7#include <linux/writeback.h> /* generic_writepages */
8#include <linux/slab.h>
9#include <linux/pagevec.h>
10#include <linux/task_io_accounting_ops.h>
11
12#include "super.h"
13#include "mds_client.h"
14#include <linux/ceph/osd_client.h>
15
16/*
17 * Ceph address space ops.
18 *
19 * There are a few funny things going on here.
20 *
21 * The page->private field is used to reference a struct
22 * ceph_snap_context for _every_ dirty page. This indicates which
23 * snapshot the page was logically dirtied in, and thus which snap
24 * context needs to be associated with the osd write during writeback.
25 *
26 * Similarly, struct ceph_inode_info maintains a set of counters to
27 * count dirty pages on the inode. In the absence of snapshots,
28 * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
29 *
30 * When a snapshot is taken (that is, when the client receives
31 * notification that a snapshot was taken), each inode with caps and
32 * with dirty pages (dirty pages implies there is a cap) gets a new
33 * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
34 * order, new snaps go to the tail). The i_wrbuffer_ref_head count is
35 * moved to capsnap->dirty. (Unless a sync write is currently in
36 * progress. In that case, the capsnap is said to be "pending", new
37 * writes cannot start, and the capsnap isn't "finalized" until the
38 * write completes (or fails) and a final size/mtime for the inode for
39 * that snap can be settled upon.) i_wrbuffer_ref_head is reset to 0.
40 *
41 * On writeback, we must submit writes to the osd IN SNAP ORDER. So,
42 * we look for the first capsnap in i_cap_snaps and write out pages in
43 * that snap context _only_. Then we move on to the next capsnap,
44 * eventually reaching the "live" or "head" context (i.e., pages that
45 * are not yet snapped) and are writing the most recently dirtied
46 * pages.
47 *
48 * Invalidate and so forth must take care to ensure the dirty page
49 * accounting is preserved.
50 */
51
52#define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
53#define CONGESTION_OFF_THRESH(congestion_kb) \
54 (CONGESTION_ON_THRESH(congestion_kb) - \
55 (CONGESTION_ON_THRESH(congestion_kb) >> 2))
56
57
58
59/*
60 * Dirty a page. Optimistically adjust accounting, on the assumption
61 * that we won't race with invalidate. If we do, readjust.
62 */
63static int ceph_set_page_dirty(struct page *page)
64{
65 struct address_space *mapping = page->mapping;
66 struct inode *inode;
67 struct ceph_inode_info *ci;
68 int undo = 0;
69 struct ceph_snap_context *snapc;
70
71 if (unlikely(!mapping))
72 return !TestSetPageDirty(page);
73
74 if (TestSetPageDirty(page)) {
75 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
76 mapping->host, page, page->index);
77 return 0;
78 }
79
80 inode = mapping->host;
81 ci = ceph_inode(inode);
82
83 /*
84 * Note that we're grabbing a snapc ref here without holding
85 * any locks!
86 */
87 snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
88
89 /* dirty the head */
90 spin_lock(&inode->i_lock);
91 if (ci->i_head_snapc == NULL)
92 ci->i_head_snapc = ceph_get_snap_context(snapc);
93 ++ci->i_wrbuffer_ref_head;
94 if (ci->i_wrbuffer_ref == 0)
95 ihold(inode);
96 ++ci->i_wrbuffer_ref;
97 dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
98 "snapc %p seq %lld (%d snaps)\n",
99 mapping->host, page, page->index,
100 ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
101 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
102 snapc, snapc->seq, snapc->num_snaps);
103 spin_unlock(&inode->i_lock);
104
105 /* now adjust page */
106 spin_lock_irq(&mapping->tree_lock);
107 if (page->mapping) { /* Race with truncate? */
108 WARN_ON_ONCE(!PageUptodate(page));
109 account_page_dirtied(page, page->mapping);
110 radix_tree_tag_set(&mapping->page_tree,
111 page_index(page), PAGECACHE_TAG_DIRTY);
112
113 /*
114 * Reference snap context in page->private. Also set
115 * PagePrivate so that we get invalidatepage callback.
116 */
117 page->private = (unsigned long)snapc;
118 SetPagePrivate(page);
119 } else {
120 dout("ANON set_page_dirty %p (raced truncate?)\n", page);
121 undo = 1;
122 }
123
124 spin_unlock_irq(&mapping->tree_lock);
125
126 if (undo)
127 /* whoops, we failed to dirty the page */
128 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
129
130 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
131
132 BUG_ON(!PageDirty(page));
133 return 1;
134}
135
136/*
137 * If we are truncating the full page (i.e. offset == 0), adjust the
138 * dirty page counters appropriately. Only called if there is private
139 * data on the page.
140 */
141static void ceph_invalidatepage(struct page *page, unsigned long offset)
142{
143 struct inode *inode;
144 struct ceph_inode_info *ci;
145 struct ceph_snap_context *snapc = (void *)page->private;
146
147 BUG_ON(!PageLocked(page));
148 BUG_ON(!page->private);
149 BUG_ON(!PagePrivate(page));
150 BUG_ON(!page->mapping);
151
152 inode = page->mapping->host;
153
154 /*
155 * We can get non-dirty pages here due to races between
156 * set_page_dirty and truncate_complete_page; just spit out a
157 * warning, in case we end up with accounting problems later.
158 */
159 if (!PageDirty(page))
160 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
161
162 if (offset == 0)
163 ClearPageChecked(page);
164
165 ci = ceph_inode(inode);
166 if (offset == 0) {
167 dout("%p invalidatepage %p idx %lu full dirty page %lu\n",
168 inode, page, page->index, offset);
169 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
170 ceph_put_snap_context(snapc);
171 page->private = 0;
172 ClearPagePrivate(page);
173 } else {
174 dout("%p invalidatepage %p idx %lu partial dirty page\n",
175 inode, page, page->index);
176 }
177}
178
179/* just a sanity check */
180static int ceph_releasepage(struct page *page, gfp_t g)
181{
182 struct inode *inode = page->mapping ? page->mapping->host : NULL;
183 dout("%p releasepage %p idx %lu\n", inode, page, page->index);
184 WARN_ON(PageDirty(page));
185 WARN_ON(page->private);
186 WARN_ON(PagePrivate(page));
187 return 0;
188}
189
190/*
191 * read a single page, without unlocking it.
192 */
193static int readpage_nounlock(struct file *filp, struct page *page)
194{
195 struct inode *inode = filp->f_dentry->d_inode;
196 struct ceph_inode_info *ci = ceph_inode(inode);
197 struct ceph_osd_client *osdc =
198 &ceph_inode_to_client(inode)->client->osdc;
199 int err = 0;
200 u64 len = PAGE_CACHE_SIZE;
201
202 dout("readpage inode %p file %p page %p index %lu\n",
203 inode, filp, page, page->index);
204 err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
205 page->index << PAGE_CACHE_SHIFT, &len,
206 ci->i_truncate_seq, ci->i_truncate_size,
207 &page, 1, 0);
208 if (err == -ENOENT)
209 err = 0;
210 if (err < 0) {
211 SetPageError(page);
212 goto out;
213 } else if (err < PAGE_CACHE_SIZE) {
214 /* zero fill remainder of page */
215 zero_user_segment(page, err, PAGE_CACHE_SIZE);
216 }
217 SetPageUptodate(page);
218
219out:
220 return err < 0 ? err : 0;
221}
222
223static int ceph_readpage(struct file *filp, struct page *page)
224{
225 int r = readpage_nounlock(filp, page);
226 unlock_page(page);
227 return r;
228}
229
230/*
231 * Build a vector of contiguous pages from the provided page list.
232 */
233static struct page **page_vector_from_list(struct list_head *page_list,
234 unsigned *nr_pages)
235{
236 struct page **pages;
237 struct page *page;
238 int next_index, contig_pages = 0;
239
240 /* build page vector */
241 pages = kmalloc(sizeof(*pages) * *nr_pages, GFP_NOFS);
242 if (!pages)
243 return ERR_PTR(-ENOMEM);
244
245 BUG_ON(list_empty(page_list));
246 next_index = list_entry(page_list->prev, struct page, lru)->index;
247 list_for_each_entry_reverse(page, page_list, lru) {
248 if (page->index == next_index) {
249 dout("readpages page %d %p\n", contig_pages, page);
250 pages[contig_pages] = page;
251 contig_pages++;
252 next_index++;
253 } else {
254 break;
255 }
256 }
257 *nr_pages = contig_pages;
258 return pages;
259}
260
261/*
262 * Read multiple pages. Leave pages we don't read + unlock in page_list;
263 * the caller (VM) cleans them up.
264 */
265static int ceph_readpages(struct file *file, struct address_space *mapping,
266 struct list_head *page_list, unsigned nr_pages)
267{
268 struct inode *inode = file->f_dentry->d_inode;
269 struct ceph_inode_info *ci = ceph_inode(inode);
270 struct ceph_osd_client *osdc =
271 &ceph_inode_to_client(inode)->client->osdc;
272 int rc = 0;
273 struct page **pages;
274 loff_t offset;
275 u64 len;
276
277 dout("readpages %p file %p nr_pages %d\n",
278 inode, file, nr_pages);
279
280 pages = page_vector_from_list(page_list, &nr_pages);
281 if (IS_ERR(pages))
282 return PTR_ERR(pages);
283
284 /* guess read extent */
285 offset = pages[0]->index << PAGE_CACHE_SHIFT;
286 len = nr_pages << PAGE_CACHE_SHIFT;
287 rc = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
288 offset, &len,
289 ci->i_truncate_seq, ci->i_truncate_size,
290 pages, nr_pages, 0);
291 if (rc == -ENOENT)
292 rc = 0;
293 if (rc < 0)
294 goto out;
295
296 for (; !list_empty(page_list) && len > 0;
297 rc -= PAGE_CACHE_SIZE, len -= PAGE_CACHE_SIZE) {
298 struct page *page =
299 list_entry(page_list->prev, struct page, lru);
300
301 list_del(&page->lru);
302
303 if (rc < (int)PAGE_CACHE_SIZE) {
304 /* zero (remainder of) page */
305 int s = rc < 0 ? 0 : rc;
306 zero_user_segment(page, s, PAGE_CACHE_SIZE);
307 }
308
309 if (add_to_page_cache_lru(page, mapping, page->index,
310 GFP_NOFS)) {
311 page_cache_release(page);
312 dout("readpages %p add_to_page_cache failed %p\n",
313 inode, page);
314 continue;
315 }
316 dout("readpages %p adding %p idx %lu\n", inode, page,
317 page->index);
318 flush_dcache_page(page);
319 SetPageUptodate(page);
320 unlock_page(page);
321 page_cache_release(page);
322 }
323 rc = 0;
324
325out:
326 kfree(pages);
327 return rc;
328}
329
330/*
331 * Get ref for the oldest snapc for an inode with dirty data... that is, the
332 * only snap context we are allowed to write back.
333 */
334static struct ceph_snap_context *get_oldest_context(struct inode *inode,
335 u64 *snap_size)
336{
337 struct ceph_inode_info *ci = ceph_inode(inode);
338 struct ceph_snap_context *snapc = NULL;
339 struct ceph_cap_snap *capsnap = NULL;
340
341 spin_lock(&inode->i_lock);
342 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
343 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
344 capsnap->context, capsnap->dirty_pages);
345 if (capsnap->dirty_pages) {
346 snapc = ceph_get_snap_context(capsnap->context);
347 if (snap_size)
348 *snap_size = capsnap->size;
349 break;
350 }
351 }
352 if (!snapc && ci->i_wrbuffer_ref_head) {
353 snapc = ceph_get_snap_context(ci->i_head_snapc);
354 dout(" head snapc %p has %d dirty pages\n",
355 snapc, ci->i_wrbuffer_ref_head);
356 }
357 spin_unlock(&inode->i_lock);
358 return snapc;
359}
360
361/*
362 * Write a single page, but leave the page locked.
363 *
364 * If we get a write error, set the page error bit, but still adjust the
365 * dirty page accounting (i.e., page is no longer dirty).
366 */
367static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
368{
369 struct inode *inode;
370 struct ceph_inode_info *ci;
371 struct ceph_fs_client *fsc;
372 struct ceph_osd_client *osdc;
373 loff_t page_off = page->index << PAGE_CACHE_SHIFT;
374 int len = PAGE_CACHE_SIZE;
375 loff_t i_size;
376 int err = 0;
377 struct ceph_snap_context *snapc, *oldest;
378 u64 snap_size = 0;
379 long writeback_stat;
380
381 dout("writepage %p idx %lu\n", page, page->index);
382
383 if (!page->mapping || !page->mapping->host) {
384 dout("writepage %p - no mapping\n", page);
385 return -EFAULT;
386 }
387 inode = page->mapping->host;
388 ci = ceph_inode(inode);
389 fsc = ceph_inode_to_client(inode);
390 osdc = &fsc->client->osdc;
391
392 /* verify this is a writeable snap context */
393 snapc = (void *)page->private;
394 if (snapc == NULL) {
395 dout("writepage %p page %p not dirty?\n", inode, page);
396 goto out;
397 }
398 oldest = get_oldest_context(inode, &snap_size);
399 if (snapc->seq > oldest->seq) {
400 dout("writepage %p page %p snapc %p not writeable - noop\n",
401 inode, page, (void *)page->private);
402 /* we should only noop if called by kswapd */
403 WARN_ON((current->flags & PF_MEMALLOC) == 0);
404 ceph_put_snap_context(oldest);
405 goto out;
406 }
407 ceph_put_snap_context(oldest);
408
409 /* is this a partial page at end of file? */
410 if (snap_size)
411 i_size = snap_size;
412 else
413 i_size = i_size_read(inode);
414 if (i_size < page_off + len)
415 len = i_size - page_off;
416
417 dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
418 inode, page, page->index, page_off, len, snapc);
419
420 writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
421 if (writeback_stat >
422 CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
423 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
424
425 set_page_writeback(page);
426 err = ceph_osdc_writepages(osdc, ceph_vino(inode),
427 &ci->i_layout, snapc,
428 page_off, len,
429 ci->i_truncate_seq, ci->i_truncate_size,
430 &inode->i_mtime,
431 &page, 1, 0, 0, true);
432 if (err < 0) {
433 dout("writepage setting page/mapping error %d %p\n", err, page);
434 SetPageError(page);
435 mapping_set_error(&inode->i_data, err);
436 if (wbc)
437 wbc->pages_skipped++;
438 } else {
439 dout("writepage cleaned page %p\n", page);
440 err = 0; /* vfs expects us to return 0 */
441 }
442 page->private = 0;
443 ClearPagePrivate(page);
444 end_page_writeback(page);
445 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
446 ceph_put_snap_context(snapc); /* page's reference */
447out:
448 return err;
449}
450
451static int ceph_writepage(struct page *page, struct writeback_control *wbc)
452{
453 int err;
454 struct inode *inode = page->mapping->host;
455 BUG_ON(!inode);
456 ihold(inode);
457 err = writepage_nounlock(page, wbc);
458 unlock_page(page);
459 iput(inode);
460 return err;
461}
462
463
464/*
465 * lame release_pages helper. release_pages() isn't exported to
466 * modules.
467 */
468static void ceph_release_pages(struct page **pages, int num)
469{
470 struct pagevec pvec;
471 int i;
472
473 pagevec_init(&pvec, 0);
474 for (i = 0; i < num; i++) {
475 if (pagevec_add(&pvec, pages[i]) == 0)
476 pagevec_release(&pvec);
477 }
478 pagevec_release(&pvec);
479}
480
481
482/*
483 * async writeback completion handler.
484 *
485 * If we get an error, set the mapping error bit, but not the individual
486 * page error bits.
487 */
488static void writepages_finish(struct ceph_osd_request *req,
489 struct ceph_msg *msg)
490{
491 struct inode *inode = req->r_inode;
492 struct ceph_osd_reply_head *replyhead;
493 struct ceph_osd_op *op;
494 struct ceph_inode_info *ci = ceph_inode(inode);
495 unsigned wrote;
496 struct page *page;
497 int i;
498 struct ceph_snap_context *snapc = req->r_snapc;
499 struct address_space *mapping = inode->i_mapping;
500 __s32 rc = -EIO;
501 u64 bytes = 0;
502 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
503 long writeback_stat;
504 unsigned issued = ceph_caps_issued(ci);
505
506 /* parse reply */
507 replyhead = msg->front.iov_base;
508 WARN_ON(le32_to_cpu(replyhead->num_ops) == 0);
509 op = (void *)(replyhead + 1);
510 rc = le32_to_cpu(replyhead->result);
511 bytes = le64_to_cpu(op->extent.length);
512
513 if (rc >= 0) {
514 /*
515 * Assume we wrote the pages we originally sent. The
516 * osd might reply with fewer pages if our writeback
517 * raced with a truncation and was adjusted at the osd,
518 * so don't believe the reply.
519 */
520 wrote = req->r_num_pages;
521 } else {
522 wrote = 0;
523 mapping_set_error(mapping, rc);
524 }
525 dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
526 inode, rc, bytes, wrote);
527
528 /* clean all pages */
529 for (i = 0; i < req->r_num_pages; i++) {
530 page = req->r_pages[i];
531 BUG_ON(!page);
532 WARN_ON(!PageUptodate(page));
533
534 writeback_stat =
535 atomic_long_dec_return(&fsc->writeback_count);
536 if (writeback_stat <
537 CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
538 clear_bdi_congested(&fsc->backing_dev_info,
539 BLK_RW_ASYNC);
540
541 ceph_put_snap_context((void *)page->private);
542 page->private = 0;
543 ClearPagePrivate(page);
544 dout("unlocking %d %p\n", i, page);
545 end_page_writeback(page);
546
547 /*
548 * We lost the cache cap, need to truncate the page before
549 * it is unlocked, otherwise we'd truncate it later in the
550 * page truncation thread, possibly losing some data that
551 * raced its way in
552 */
553 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
554 generic_error_remove_page(inode->i_mapping, page);
555
556 unlock_page(page);
557 }
558 dout("%p wrote+cleaned %d pages\n", inode, wrote);
559 ceph_put_wrbuffer_cap_refs(ci, req->r_num_pages, snapc);
560
561 ceph_release_pages(req->r_pages, req->r_num_pages);
562 if (req->r_pages_from_pool)
563 mempool_free(req->r_pages,
564 ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
565 else
566 kfree(req->r_pages);
567 ceph_osdc_put_request(req);
568}
569
570/*
571 * allocate a page vec, either directly, or if necessary, via a the
572 * mempool. we avoid the mempool if we can because req->r_num_pages
573 * may be less than the maximum write size.
574 */
575static void alloc_page_vec(struct ceph_fs_client *fsc,
576 struct ceph_osd_request *req)
577{
578 req->r_pages = kmalloc(sizeof(struct page *) * req->r_num_pages,
579 GFP_NOFS);
580 if (!req->r_pages) {
581 req->r_pages = mempool_alloc(fsc->wb_pagevec_pool, GFP_NOFS);
582 req->r_pages_from_pool = 1;
583 WARN_ON(!req->r_pages);
584 }
585}
586
587/*
588 * initiate async writeback
589 */
590static int ceph_writepages_start(struct address_space *mapping,
591 struct writeback_control *wbc)
592{
593 struct inode *inode = mapping->host;
594 struct ceph_inode_info *ci = ceph_inode(inode);
595 struct ceph_fs_client *fsc;
596 pgoff_t index, start, end;
597 int range_whole = 0;
598 int should_loop = 1;
599 pgoff_t max_pages = 0, max_pages_ever = 0;
600 struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
601 struct pagevec pvec;
602 int done = 0;
603 int rc = 0;
604 unsigned wsize = 1 << inode->i_blkbits;
605 struct ceph_osd_request *req = NULL;
606 int do_sync;
607 u64 snap_size = 0;
608
609 /*
610 * Include a 'sync' in the OSD request if this is a data
611 * integrity write (e.g., O_SYNC write or fsync()), or if our
612 * cap is being revoked.
613 */
614 do_sync = wbc->sync_mode == WB_SYNC_ALL;
615 if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
616 do_sync = 1;
617 dout("writepages_start %p dosync=%d (mode=%s)\n",
618 inode, do_sync,
619 wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
620 (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
621
622 fsc = ceph_inode_to_client(inode);
623 if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
624 pr_warning("writepage_start %p on forced umount\n", inode);
625 return -EIO; /* we're in a forced umount, don't write! */
626 }
627 if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
628 wsize = fsc->mount_options->wsize;
629 if (wsize < PAGE_CACHE_SIZE)
630 wsize = PAGE_CACHE_SIZE;
631 max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
632
633 pagevec_init(&pvec, 0);
634
635 /* where to start/end? */
636 if (wbc->range_cyclic) {
637 start = mapping->writeback_index; /* Start from prev offset */
638 end = -1;
639 dout(" cyclic, start at %lu\n", start);
640 } else {
641 start = wbc->range_start >> PAGE_CACHE_SHIFT;
642 end = wbc->range_end >> PAGE_CACHE_SHIFT;
643 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
644 range_whole = 1;
645 should_loop = 0;
646 dout(" not cyclic, %lu to %lu\n", start, end);
647 }
648 index = start;
649
650retry:
651 /* find oldest snap context with dirty data */
652 ceph_put_snap_context(snapc);
653 snapc = get_oldest_context(inode, &snap_size);
654 if (!snapc) {
655 /* hmm, why does writepages get called when there
656 is no dirty data? */
657 dout(" no snap context with dirty data?\n");
658 goto out;
659 }
660 dout(" oldest snapc is %p seq %lld (%d snaps)\n",
661 snapc, snapc->seq, snapc->num_snaps);
662 if (last_snapc && snapc != last_snapc) {
663 /* if we switched to a newer snapc, restart our scan at the
664 * start of the original file range. */
665 dout(" snapc differs from last pass, restarting at %lu\n",
666 index);
667 index = start;
668 }
669 last_snapc = snapc;
670
671 while (!done && index <= end) {
672 unsigned i;
673 int first;
674 pgoff_t next;
675 int pvec_pages, locked_pages;
676 struct page *page;
677 int want;
678 u64 offset, len;
679 struct ceph_osd_request_head *reqhead;
680 struct ceph_osd_op *op;
681 long writeback_stat;
682
683 next = 0;
684 locked_pages = 0;
685 max_pages = max_pages_ever;
686
687get_more_pages:
688 first = -1;
689 want = min(end - index,
690 min((pgoff_t)PAGEVEC_SIZE,
691 max_pages - (pgoff_t)locked_pages) - 1)
692 + 1;
693 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
694 PAGECACHE_TAG_DIRTY,
695 want);
696 dout("pagevec_lookup_tag got %d\n", pvec_pages);
697 if (!pvec_pages && !locked_pages)
698 break;
699 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
700 page = pvec.pages[i];
701 dout("? %p idx %lu\n", page, page->index);
702 if (locked_pages == 0)
703 lock_page(page); /* first page */
704 else if (!trylock_page(page))
705 break;
706
707 /* only dirty pages, or our accounting breaks */
708 if (unlikely(!PageDirty(page)) ||
709 unlikely(page->mapping != mapping)) {
710 dout("!dirty or !mapping %p\n", page);
711 unlock_page(page);
712 break;
713 }
714 if (!wbc->range_cyclic && page->index > end) {
715 dout("end of range %p\n", page);
716 done = 1;
717 unlock_page(page);
718 break;
719 }
720 if (next && (page->index != next)) {
721 dout("not consecutive %p\n", page);
722 unlock_page(page);
723 break;
724 }
725 if (wbc->sync_mode != WB_SYNC_NONE) {
726 dout("waiting on writeback %p\n", page);
727 wait_on_page_writeback(page);
728 }
729 if ((snap_size && page_offset(page) > snap_size) ||
730 (!snap_size &&
731 page_offset(page) > i_size_read(inode))) {
732 dout("%p page eof %llu\n", page, snap_size ?
733 snap_size : i_size_read(inode));
734 done = 1;
735 unlock_page(page);
736 break;
737 }
738 if (PageWriteback(page)) {
739 dout("%p under writeback\n", page);
740 unlock_page(page);
741 break;
742 }
743
744 /* only if matching snap context */
745 pgsnapc = (void *)page->private;
746 if (pgsnapc->seq > snapc->seq) {
747 dout("page snapc %p %lld > oldest %p %lld\n",
748 pgsnapc, pgsnapc->seq, snapc, snapc->seq);
749 unlock_page(page);
750 if (!locked_pages)
751 continue; /* keep looking for snap */
752 break;
753 }
754
755 if (!clear_page_dirty_for_io(page)) {
756 dout("%p !clear_page_dirty_for_io\n", page);
757 unlock_page(page);
758 break;
759 }
760
761 /* ok */
762 if (locked_pages == 0) {
763 /* prepare async write request */
764 offset = (unsigned long long)page->index
765 << PAGE_CACHE_SHIFT;
766 len = wsize;
767 req = ceph_osdc_new_request(&fsc->client->osdc,
768 &ci->i_layout,
769 ceph_vino(inode),
770 offset, &len,
771 CEPH_OSD_OP_WRITE,
772 CEPH_OSD_FLAG_WRITE |
773 CEPH_OSD_FLAG_ONDISK,
774 snapc, do_sync,
775 ci->i_truncate_seq,
776 ci->i_truncate_size,
777 &inode->i_mtime, true, 1, 0);
778
779 if (!req) {
780 rc = -ENOMEM;
781 unlock_page(page);
782 break;
783 }
784
785 max_pages = req->r_num_pages;
786
787 alloc_page_vec(fsc, req);
788 req->r_callback = writepages_finish;
789 req->r_inode = inode;
790 }
791
792 /* note position of first page in pvec */
793 if (first < 0)
794 first = i;
795 dout("%p will write page %p idx %lu\n",
796 inode, page, page->index);
797
798 writeback_stat =
799 atomic_long_inc_return(&fsc->writeback_count);
800 if (writeback_stat > CONGESTION_ON_THRESH(
801 fsc->mount_options->congestion_kb)) {
802 set_bdi_congested(&fsc->backing_dev_info,
803 BLK_RW_ASYNC);
804 }
805
806 set_page_writeback(page);
807 req->r_pages[locked_pages] = page;
808 locked_pages++;
809 next = page->index + 1;
810 }
811
812 /* did we get anything? */
813 if (!locked_pages)
814 goto release_pvec_pages;
815 if (i) {
816 int j;
817 BUG_ON(!locked_pages || first < 0);
818
819 if (pvec_pages && i == pvec_pages &&
820 locked_pages < max_pages) {
821 dout("reached end pvec, trying for more\n");
822 pagevec_reinit(&pvec);
823 goto get_more_pages;
824 }
825
826 /* shift unused pages over in the pvec... we
827 * will need to release them below. */
828 for (j = i; j < pvec_pages; j++) {
829 dout(" pvec leftover page %p\n",
830 pvec.pages[j]);
831 pvec.pages[j-i+first] = pvec.pages[j];
832 }
833 pvec.nr -= i-first;
834 }
835
836 /* submit the write */
837 offset = req->r_pages[0]->index << PAGE_CACHE_SHIFT;
838 len = min((snap_size ? snap_size : i_size_read(inode)) - offset,
839 (u64)locked_pages << PAGE_CACHE_SHIFT);
840 dout("writepages got %d pages at %llu~%llu\n",
841 locked_pages, offset, len);
842
843 /* revise final length, page count */
844 req->r_num_pages = locked_pages;
845 reqhead = req->r_request->front.iov_base;
846 op = (void *)(reqhead + 1);
847 op->extent.length = cpu_to_le64(len);
848 op->payload_len = cpu_to_le32(len);
849 req->r_request->hdr.data_len = cpu_to_le32(len);
850
851 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
852 BUG_ON(rc);
853 req = NULL;
854
855 /* continue? */
856 index = next;
857 wbc->nr_to_write -= locked_pages;
858 if (wbc->nr_to_write <= 0)
859 done = 1;
860
861release_pvec_pages:
862 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
863 pvec.nr ? pvec.pages[0] : NULL);
864 pagevec_release(&pvec);
865
866 if (locked_pages && !done)
867 goto retry;
868 }
869
870 if (should_loop && !done) {
871 /* more to do; loop back to beginning of file */
872 dout("writepages looping back to beginning of file\n");
873 should_loop = 0;
874 index = 0;
875 goto retry;
876 }
877
878 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
879 mapping->writeback_index = index;
880
881out:
882 if (req)
883 ceph_osdc_put_request(req);
884 ceph_put_snap_context(snapc);
885 dout("writepages done, rc = %d\n", rc);
886 return rc;
887}
888
889
890
891/*
892 * See if a given @snapc is either writeable, or already written.
893 */
894static int context_is_writeable_or_written(struct inode *inode,
895 struct ceph_snap_context *snapc)
896{
897 struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
898 int ret = !oldest || snapc->seq <= oldest->seq;
899
900 ceph_put_snap_context(oldest);
901 return ret;
902}
903
904/*
905 * We are only allowed to write into/dirty the page if the page is
906 * clean, or already dirty within the same snap context.
907 *
908 * called with page locked.
909 * return success with page locked,
910 * or any failure (incl -EAGAIN) with page unlocked.
911 */
912static int ceph_update_writeable_page(struct file *file,
913 loff_t pos, unsigned len,
914 struct page *page)
915{
916 struct inode *inode = file->f_dentry->d_inode;
917 struct ceph_inode_info *ci = ceph_inode(inode);
918 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
919 loff_t page_off = pos & PAGE_CACHE_MASK;
920 int pos_in_page = pos & ~PAGE_CACHE_MASK;
921 int end_in_page = pos_in_page + len;
922 loff_t i_size;
923 int r;
924 struct ceph_snap_context *snapc, *oldest;
925
926retry_locked:
927 /* writepages currently holds page lock, but if we change that later, */
928 wait_on_page_writeback(page);
929
930 /* check snap context */
931 BUG_ON(!ci->i_snap_realm);
932 down_read(&mdsc->snap_rwsem);
933 BUG_ON(!ci->i_snap_realm->cached_context);
934 snapc = (void *)page->private;
935 if (snapc && snapc != ci->i_head_snapc) {
936 /*
937 * this page is already dirty in another (older) snap
938 * context! is it writeable now?
939 */
940 oldest = get_oldest_context(inode, NULL);
941 up_read(&mdsc->snap_rwsem);
942
943 if (snapc->seq > oldest->seq) {
944 ceph_put_snap_context(oldest);
945 dout(" page %p snapc %p not current or oldest\n",
946 page, snapc);
947 /*
948 * queue for writeback, and wait for snapc to
949 * be writeable or written
950 */
951 snapc = ceph_get_snap_context(snapc);
952 unlock_page(page);
953 ceph_queue_writeback(inode);
954 r = wait_event_interruptible(ci->i_cap_wq,
955 context_is_writeable_or_written(inode, snapc));
956 ceph_put_snap_context(snapc);
957 if (r == -ERESTARTSYS)
958 return r;
959 return -EAGAIN;
960 }
961 ceph_put_snap_context(oldest);
962
963 /* yay, writeable, do it now (without dropping page lock) */
964 dout(" page %p snapc %p not current, but oldest\n",
965 page, snapc);
966 if (!clear_page_dirty_for_io(page))
967 goto retry_locked;
968 r = writepage_nounlock(page, NULL);
969 if (r < 0)
970 goto fail_nosnap;
971 goto retry_locked;
972 }
973
974 if (PageUptodate(page)) {
975 dout(" page %p already uptodate\n", page);
976 return 0;
977 }
978
979 /* full page? */
980 if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
981 return 0;
982
983 /* past end of file? */
984 i_size = inode->i_size; /* caller holds i_mutex */
985
986 if (i_size + len > inode->i_sb->s_maxbytes) {
987 /* file is too big */
988 r = -EINVAL;
989 goto fail;
990 }
991
992 if (page_off >= i_size ||
993 (pos_in_page == 0 && (pos+len) >= i_size &&
994 end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
995 dout(" zeroing %p 0 - %d and %d - %d\n",
996 page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
997 zero_user_segments(page,
998 0, pos_in_page,
999 end_in_page, PAGE_CACHE_SIZE);
1000 return 0;
1001 }
1002
1003 /* we need to read it. */
1004 up_read(&mdsc->snap_rwsem);
1005 r = readpage_nounlock(file, page);
1006 if (r < 0)
1007 goto fail_nosnap;
1008 goto retry_locked;
1009
1010fail:
1011 up_read(&mdsc->snap_rwsem);
1012fail_nosnap:
1013 unlock_page(page);
1014 return r;
1015}
1016
1017/*
1018 * We are only allowed to write into/dirty the page if the page is
1019 * clean, or already dirty within the same snap context.
1020 */
1021static int ceph_write_begin(struct file *file, struct address_space *mapping,
1022 loff_t pos, unsigned len, unsigned flags,
1023 struct page **pagep, void **fsdata)
1024{
1025 struct inode *inode = file->f_dentry->d_inode;
1026 struct page *page;
1027 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1028 int r;
1029
1030 do {
1031 /* get a page */
1032 page = grab_cache_page_write_begin(mapping, index, 0);
1033 if (!page)
1034 return -ENOMEM;
1035 *pagep = page;
1036
1037 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1038 inode, page, (int)pos, (int)len);
1039
1040 r = ceph_update_writeable_page(file, pos, len, page);
1041 } while (r == -EAGAIN);
1042
1043 return r;
1044}
1045
1046/*
1047 * we don't do anything in here that simple_write_end doesn't do
1048 * except adjust dirty page accounting and drop read lock on
1049 * mdsc->snap_rwsem.
1050 */
1051static int ceph_write_end(struct file *file, struct address_space *mapping,
1052 loff_t pos, unsigned len, unsigned copied,
1053 struct page *page, void *fsdata)
1054{
1055 struct inode *inode = file->f_dentry->d_inode;
1056 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1057 struct ceph_mds_client *mdsc = fsc->mdsc;
1058 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1059 int check_cap = 0;
1060
1061 dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1062 inode, page, (int)pos, (int)copied, (int)len);
1063
1064 /* zero the stale part of the page if we did a short copy */
1065 if (copied < len)
1066 zero_user_segment(page, from+copied, len);
1067
1068 /* did file size increase? */
1069 /* (no need for i_size_read(); we caller holds i_mutex */
1070 if (pos+copied > inode->i_size)
1071 check_cap = ceph_inode_set_size(inode, pos+copied);
1072
1073 if (!PageUptodate(page))
1074 SetPageUptodate(page);
1075
1076 set_page_dirty(page);
1077
1078 unlock_page(page);
1079 up_read(&mdsc->snap_rwsem);
1080 page_cache_release(page);
1081
1082 if (check_cap)
1083 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1084
1085 return copied;
1086}
1087
1088/*
1089 * we set .direct_IO to indicate direct io is supported, but since we
1090 * intercept O_DIRECT reads and writes early, this function should
1091 * never get called.
1092 */
1093static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1094 const struct iovec *iov,
1095 loff_t pos, unsigned long nr_segs)
1096{
1097 WARN_ON(1);
1098 return -EINVAL;
1099}
1100
1101const struct address_space_operations ceph_aops = {
1102 .readpage = ceph_readpage,
1103 .readpages = ceph_readpages,
1104 .writepage = ceph_writepage,
1105 .writepages = ceph_writepages_start,
1106 .write_begin = ceph_write_begin,
1107 .write_end = ceph_write_end,
1108 .set_page_dirty = ceph_set_page_dirty,
1109 .invalidatepage = ceph_invalidatepage,
1110 .releasepage = ceph_releasepage,
1111 .direct_IO = ceph_direct_io,
1112};
1113
1114
1115/*
1116 * vm ops
1117 */
1118
1119/*
1120 * Reuse write_begin here for simplicity.
1121 */
1122static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1123{
1124 struct inode *inode = vma->vm_file->f_dentry->d_inode;
1125 struct page *page = vmf->page;
1126 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1127 loff_t off = page->index << PAGE_CACHE_SHIFT;
1128 loff_t size, len;
1129 int ret;
1130
1131 size = i_size_read(inode);
1132 if (off + PAGE_CACHE_SIZE <= size)
1133 len = PAGE_CACHE_SIZE;
1134 else
1135 len = size & ~PAGE_CACHE_MASK;
1136
1137 dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1138 off, len, page, page->index);
1139
1140 lock_page(page);
1141
1142 ret = VM_FAULT_NOPAGE;
1143 if ((off > size) ||
1144 (page->mapping != inode->i_mapping))
1145 goto out;
1146
1147 ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1148 if (ret == 0) {
1149 /* success. we'll keep the page locked. */
1150 set_page_dirty(page);
1151 up_read(&mdsc->snap_rwsem);
1152 ret = VM_FAULT_LOCKED;
1153 } else {
1154 if (ret == -ENOMEM)
1155 ret = VM_FAULT_OOM;
1156 else
1157 ret = VM_FAULT_SIGBUS;
1158 }
1159out:
1160 dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1161 if (ret != VM_FAULT_LOCKED)
1162 unlock_page(page);
1163 return ret;
1164}
1165
1166static struct vm_operations_struct ceph_vmops = {
1167 .fault = filemap_fault,
1168 .page_mkwrite = ceph_page_mkwrite,
1169};
1170
1171int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1172{
1173 struct address_space *mapping = file->f_mapping;
1174
1175 if (!mapping->a_ops->readpage)
1176 return -ENOEXEC;
1177 file_accessed(file);
1178 vma->vm_ops = &ceph_vmops;
1179 vma->vm_flags |= VM_CAN_NONLINEAR;
1180 return 0;
1181}
1#include <linux/ceph/ceph_debug.h>
2
3#include <linux/backing-dev.h>
4#include <linux/fs.h>
5#include <linux/mm.h>
6#include <linux/pagemap.h>
7#include <linux/writeback.h> /* generic_writepages */
8#include <linux/slab.h>
9#include <linux/pagevec.h>
10#include <linux/task_io_accounting_ops.h>
11
12#include "super.h"
13#include "mds_client.h"
14#include "cache.h"
15#include <linux/ceph/osd_client.h>
16
17/*
18 * Ceph address space ops.
19 *
20 * There are a few funny things going on here.
21 *
22 * The page->private field is used to reference a struct
23 * ceph_snap_context for _every_ dirty page. This indicates which
24 * snapshot the page was logically dirtied in, and thus which snap
25 * context needs to be associated with the osd write during writeback.
26 *
27 * Similarly, struct ceph_inode_info maintains a set of counters to
28 * count dirty pages on the inode. In the absence of snapshots,
29 * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
30 *
31 * When a snapshot is taken (that is, when the client receives
32 * notification that a snapshot was taken), each inode with caps and
33 * with dirty pages (dirty pages implies there is a cap) gets a new
34 * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
35 * order, new snaps go to the tail). The i_wrbuffer_ref_head count is
36 * moved to capsnap->dirty. (Unless a sync write is currently in
37 * progress. In that case, the capsnap is said to be "pending", new
38 * writes cannot start, and the capsnap isn't "finalized" until the
39 * write completes (or fails) and a final size/mtime for the inode for
40 * that snap can be settled upon.) i_wrbuffer_ref_head is reset to 0.
41 *
42 * On writeback, we must submit writes to the osd IN SNAP ORDER. So,
43 * we look for the first capsnap in i_cap_snaps and write out pages in
44 * that snap context _only_. Then we move on to the next capsnap,
45 * eventually reaching the "live" or "head" context (i.e., pages that
46 * are not yet snapped) and are writing the most recently dirtied
47 * pages.
48 *
49 * Invalidate and so forth must take care to ensure the dirty page
50 * accounting is preserved.
51 */
52
53#define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
54#define CONGESTION_OFF_THRESH(congestion_kb) \
55 (CONGESTION_ON_THRESH(congestion_kb) - \
56 (CONGESTION_ON_THRESH(congestion_kb) >> 2))
57
58static inline struct ceph_snap_context *page_snap_context(struct page *page)
59{
60 if (PagePrivate(page))
61 return (void *)page->private;
62 return NULL;
63}
64
65/*
66 * Dirty a page. Optimistically adjust accounting, on the assumption
67 * that we won't race with invalidate. If we do, readjust.
68 */
69static int ceph_set_page_dirty(struct page *page)
70{
71 struct address_space *mapping = page->mapping;
72 struct inode *inode;
73 struct ceph_inode_info *ci;
74 struct ceph_snap_context *snapc;
75 int ret;
76
77 if (unlikely(!mapping))
78 return !TestSetPageDirty(page);
79
80 if (PageDirty(page)) {
81 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
82 mapping->host, page, page->index);
83 BUG_ON(!PagePrivate(page));
84 return 0;
85 }
86
87 inode = mapping->host;
88 ci = ceph_inode(inode);
89
90 /*
91 * Note that we're grabbing a snapc ref here without holding
92 * any locks!
93 */
94 snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
95
96 /* dirty the head */
97 spin_lock(&ci->i_ceph_lock);
98 if (ci->i_head_snapc == NULL)
99 ci->i_head_snapc = ceph_get_snap_context(snapc);
100 ++ci->i_wrbuffer_ref_head;
101 if (ci->i_wrbuffer_ref == 0)
102 ihold(inode);
103 ++ci->i_wrbuffer_ref;
104 dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
105 "snapc %p seq %lld (%d snaps)\n",
106 mapping->host, page, page->index,
107 ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
108 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
109 snapc, snapc->seq, snapc->num_snaps);
110 spin_unlock(&ci->i_ceph_lock);
111
112 /*
113 * Reference snap context in page->private. Also set
114 * PagePrivate so that we get invalidatepage callback.
115 */
116 BUG_ON(PagePrivate(page));
117 page->private = (unsigned long)snapc;
118 SetPagePrivate(page);
119
120 ret = __set_page_dirty_nobuffers(page);
121 WARN_ON(!PageLocked(page));
122 WARN_ON(!page->mapping);
123
124 return ret;
125}
126
127/*
128 * If we are truncating the full page (i.e. offset == 0), adjust the
129 * dirty page counters appropriately. Only called if there is private
130 * data on the page.
131 */
132static void ceph_invalidatepage(struct page *page, unsigned int offset,
133 unsigned int length)
134{
135 struct inode *inode;
136 struct ceph_inode_info *ci;
137 struct ceph_snap_context *snapc = page_snap_context(page);
138
139 inode = page->mapping->host;
140 ci = ceph_inode(inode);
141
142 if (offset != 0 || length != PAGE_CACHE_SIZE) {
143 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
144 inode, page, page->index, offset, length);
145 return;
146 }
147
148 ceph_invalidate_fscache_page(inode, page);
149
150 if (!PagePrivate(page))
151 return;
152
153 /*
154 * We can get non-dirty pages here due to races between
155 * set_page_dirty and truncate_complete_page; just spit out a
156 * warning, in case we end up with accounting problems later.
157 */
158 if (!PageDirty(page))
159 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
160
161 ClearPageChecked(page);
162
163 dout("%p invalidatepage %p idx %lu full dirty page\n",
164 inode, page, page->index);
165
166 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
167 ceph_put_snap_context(snapc);
168 page->private = 0;
169 ClearPagePrivate(page);
170}
171
172static int ceph_releasepage(struct page *page, gfp_t g)
173{
174 struct inode *inode = page->mapping ? page->mapping->host : NULL;
175 dout("%p releasepage %p idx %lu\n", inode, page, page->index);
176 WARN_ON(PageDirty(page));
177
178 /* Can we release the page from the cache? */
179 if (!ceph_release_fscache_page(page, g))
180 return 0;
181
182 return !PagePrivate(page);
183}
184
185/*
186 * read a single page, without unlocking it.
187 */
188static int readpage_nounlock(struct file *filp, struct page *page)
189{
190 struct inode *inode = file_inode(filp);
191 struct ceph_inode_info *ci = ceph_inode(inode);
192 struct ceph_osd_client *osdc =
193 &ceph_inode_to_client(inode)->client->osdc;
194 int err = 0;
195 u64 len = PAGE_CACHE_SIZE;
196
197 err = ceph_readpage_from_fscache(inode, page);
198
199 if (err == 0)
200 goto out;
201
202 dout("readpage inode %p file %p page %p index %lu\n",
203 inode, filp, page, page->index);
204 err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
205 (u64) page_offset(page), &len,
206 ci->i_truncate_seq, ci->i_truncate_size,
207 &page, 1, 0);
208 if (err == -ENOENT)
209 err = 0;
210 if (err < 0) {
211 SetPageError(page);
212 ceph_fscache_readpage_cancel(inode, page);
213 goto out;
214 } else {
215 if (err < PAGE_CACHE_SIZE) {
216 /* zero fill remainder of page */
217 zero_user_segment(page, err, PAGE_CACHE_SIZE);
218 } else {
219 flush_dcache_page(page);
220 }
221 }
222 SetPageUptodate(page);
223
224 if (err >= 0)
225 ceph_readpage_to_fscache(inode, page);
226
227out:
228 return err < 0 ? err : 0;
229}
230
231static int ceph_readpage(struct file *filp, struct page *page)
232{
233 int r = readpage_nounlock(filp, page);
234 unlock_page(page);
235 return r;
236}
237
238/*
239 * Finish an async read(ahead) op.
240 */
241static void finish_read(struct ceph_osd_request *req, struct ceph_msg *msg)
242{
243 struct inode *inode = req->r_inode;
244 struct ceph_osd_data *osd_data;
245 int rc = req->r_result;
246 int bytes = le32_to_cpu(msg->hdr.data_len);
247 int num_pages;
248 int i;
249
250 dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
251
252 /* unlock all pages, zeroing any data we didn't read */
253 osd_data = osd_req_op_extent_osd_data(req, 0);
254 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
255 num_pages = calc_pages_for((u64)osd_data->alignment,
256 (u64)osd_data->length);
257 for (i = 0; i < num_pages; i++) {
258 struct page *page = osd_data->pages[i];
259
260 if (rc < 0)
261 goto unlock;
262 if (bytes < (int)PAGE_CACHE_SIZE) {
263 /* zero (remainder of) page */
264 int s = bytes < 0 ? 0 : bytes;
265 zero_user_segment(page, s, PAGE_CACHE_SIZE);
266 }
267 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
268 page->index);
269 flush_dcache_page(page);
270 SetPageUptodate(page);
271 ceph_readpage_to_fscache(inode, page);
272unlock:
273 unlock_page(page);
274 page_cache_release(page);
275 bytes -= PAGE_CACHE_SIZE;
276 }
277 kfree(osd_data->pages);
278}
279
280static void ceph_unlock_page_vector(struct page **pages, int num_pages)
281{
282 int i;
283
284 for (i = 0; i < num_pages; i++)
285 unlock_page(pages[i]);
286}
287
288/*
289 * start an async read(ahead) operation. return nr_pages we submitted
290 * a read for on success, or negative error code.
291 */
292static int start_read(struct inode *inode, struct list_head *page_list, int max)
293{
294 struct ceph_osd_client *osdc =
295 &ceph_inode_to_client(inode)->client->osdc;
296 struct ceph_inode_info *ci = ceph_inode(inode);
297 struct page *page = list_entry(page_list->prev, struct page, lru);
298 struct ceph_vino vino;
299 struct ceph_osd_request *req;
300 u64 off;
301 u64 len;
302 int i;
303 struct page **pages;
304 pgoff_t next_index;
305 int nr_pages = 0;
306 int ret;
307
308 off = (u64) page_offset(page);
309
310 /* count pages */
311 next_index = page->index;
312 list_for_each_entry_reverse(page, page_list, lru) {
313 if (page->index != next_index)
314 break;
315 nr_pages++;
316 next_index++;
317 if (max && nr_pages == max)
318 break;
319 }
320 len = nr_pages << PAGE_CACHE_SHIFT;
321 dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
322 off, len);
323 vino = ceph_vino(inode);
324 req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
325 1, CEPH_OSD_OP_READ,
326 CEPH_OSD_FLAG_READ, NULL,
327 ci->i_truncate_seq, ci->i_truncate_size,
328 false);
329 if (IS_ERR(req))
330 return PTR_ERR(req);
331
332 /* build page vector */
333 nr_pages = calc_pages_for(0, len);
334 pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS);
335 ret = -ENOMEM;
336 if (!pages)
337 goto out;
338 for (i = 0; i < nr_pages; ++i) {
339 page = list_entry(page_list->prev, struct page, lru);
340 BUG_ON(PageLocked(page));
341 list_del(&page->lru);
342
343 dout("start_read %p adding %p idx %lu\n", inode, page,
344 page->index);
345 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
346 GFP_NOFS)) {
347 ceph_fscache_uncache_page(inode, page);
348 page_cache_release(page);
349 dout("start_read %p add_to_page_cache failed %p\n",
350 inode, page);
351 nr_pages = i;
352 goto out_pages;
353 }
354 pages[i] = page;
355 }
356 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
357 req->r_callback = finish_read;
358 req->r_inode = inode;
359
360 ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
361
362 dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
363 ret = ceph_osdc_start_request(osdc, req, false);
364 if (ret < 0)
365 goto out_pages;
366 ceph_osdc_put_request(req);
367 return nr_pages;
368
369out_pages:
370 ceph_unlock_page_vector(pages, nr_pages);
371 ceph_release_page_vector(pages, nr_pages);
372out:
373 ceph_osdc_put_request(req);
374 return ret;
375}
376
377
378/*
379 * Read multiple pages. Leave pages we don't read + unlock in page_list;
380 * the caller (VM) cleans them up.
381 */
382static int ceph_readpages(struct file *file, struct address_space *mapping,
383 struct list_head *page_list, unsigned nr_pages)
384{
385 struct inode *inode = file_inode(file);
386 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
387 int rc = 0;
388 int max = 0;
389
390 rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
391 &nr_pages);
392
393 if (rc == 0)
394 goto out;
395
396 if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE)
397 max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1)
398 >> PAGE_SHIFT;
399
400 dout("readpages %p file %p nr_pages %d max %d\n", inode,
401 file, nr_pages,
402 max);
403 while (!list_empty(page_list)) {
404 rc = start_read(inode, page_list, max);
405 if (rc < 0)
406 goto out;
407 BUG_ON(rc == 0);
408 }
409out:
410 ceph_fscache_readpages_cancel(inode, page_list);
411
412 dout("readpages %p file %p ret %d\n", inode, file, rc);
413 return rc;
414}
415
416/*
417 * Get ref for the oldest snapc for an inode with dirty data... that is, the
418 * only snap context we are allowed to write back.
419 */
420static struct ceph_snap_context *get_oldest_context(struct inode *inode,
421 u64 *snap_size)
422{
423 struct ceph_inode_info *ci = ceph_inode(inode);
424 struct ceph_snap_context *snapc = NULL;
425 struct ceph_cap_snap *capsnap = NULL;
426
427 spin_lock(&ci->i_ceph_lock);
428 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
429 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
430 capsnap->context, capsnap->dirty_pages);
431 if (capsnap->dirty_pages) {
432 snapc = ceph_get_snap_context(capsnap->context);
433 if (snap_size)
434 *snap_size = capsnap->size;
435 break;
436 }
437 }
438 if (!snapc && ci->i_wrbuffer_ref_head) {
439 snapc = ceph_get_snap_context(ci->i_head_snapc);
440 dout(" head snapc %p has %d dirty pages\n",
441 snapc, ci->i_wrbuffer_ref_head);
442 }
443 spin_unlock(&ci->i_ceph_lock);
444 return snapc;
445}
446
447/*
448 * Write a single page, but leave the page locked.
449 *
450 * If we get a write error, set the page error bit, but still adjust the
451 * dirty page accounting (i.e., page is no longer dirty).
452 */
453static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
454{
455 struct inode *inode;
456 struct ceph_inode_info *ci;
457 struct ceph_fs_client *fsc;
458 struct ceph_osd_client *osdc;
459 struct ceph_snap_context *snapc, *oldest;
460 loff_t page_off = page_offset(page);
461 long writeback_stat;
462 u64 truncate_size, snap_size = 0;
463 u32 truncate_seq;
464 int err = 0, len = PAGE_CACHE_SIZE;
465
466 dout("writepage %p idx %lu\n", page, page->index);
467
468 if (!page->mapping || !page->mapping->host) {
469 dout("writepage %p - no mapping\n", page);
470 return -EFAULT;
471 }
472 inode = page->mapping->host;
473 ci = ceph_inode(inode);
474 fsc = ceph_inode_to_client(inode);
475 osdc = &fsc->client->osdc;
476
477 /* verify this is a writeable snap context */
478 snapc = page_snap_context(page);
479 if (snapc == NULL) {
480 dout("writepage %p page %p not dirty?\n", inode, page);
481 goto out;
482 }
483 oldest = get_oldest_context(inode, &snap_size);
484 if (snapc->seq > oldest->seq) {
485 dout("writepage %p page %p snapc %p not writeable - noop\n",
486 inode, page, snapc);
487 /* we should only noop if called by kswapd */
488 WARN_ON((current->flags & PF_MEMALLOC) == 0);
489 ceph_put_snap_context(oldest);
490 goto out;
491 }
492 ceph_put_snap_context(oldest);
493
494 spin_lock(&ci->i_ceph_lock);
495 truncate_seq = ci->i_truncate_seq;
496 truncate_size = ci->i_truncate_size;
497 if (!snap_size)
498 snap_size = i_size_read(inode);
499 spin_unlock(&ci->i_ceph_lock);
500
501 /* is this a partial page at end of file? */
502 if (page_off >= snap_size) {
503 dout("%p page eof %llu\n", page, snap_size);
504 goto out;
505 }
506 if (snap_size < page_off + len)
507 len = snap_size - page_off;
508
509 dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
510 inode, page, page->index, page_off, len, snapc);
511
512 writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
513 if (writeback_stat >
514 CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
515 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
516
517 ceph_readpage_to_fscache(inode, page);
518
519 set_page_writeback(page);
520 err = ceph_osdc_writepages(osdc, ceph_vino(inode),
521 &ci->i_layout, snapc,
522 page_off, len,
523 truncate_seq, truncate_size,
524 &inode->i_mtime, &page, 1);
525 if (err < 0) {
526 dout("writepage setting page/mapping error %d %p\n", err, page);
527 SetPageError(page);
528 mapping_set_error(&inode->i_data, err);
529 if (wbc)
530 wbc->pages_skipped++;
531 } else {
532 dout("writepage cleaned page %p\n", page);
533 err = 0; /* vfs expects us to return 0 */
534 }
535 page->private = 0;
536 ClearPagePrivate(page);
537 end_page_writeback(page);
538 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
539 ceph_put_snap_context(snapc); /* page's reference */
540out:
541 return err;
542}
543
544static int ceph_writepage(struct page *page, struct writeback_control *wbc)
545{
546 int err;
547 struct inode *inode = page->mapping->host;
548 BUG_ON(!inode);
549 ihold(inode);
550 err = writepage_nounlock(page, wbc);
551 unlock_page(page);
552 iput(inode);
553 return err;
554}
555
556
557/*
558 * lame release_pages helper. release_pages() isn't exported to
559 * modules.
560 */
561static void ceph_release_pages(struct page **pages, int num)
562{
563 struct pagevec pvec;
564 int i;
565
566 pagevec_init(&pvec, 0);
567 for (i = 0; i < num; i++) {
568 if (pagevec_add(&pvec, pages[i]) == 0)
569 pagevec_release(&pvec);
570 }
571 pagevec_release(&pvec);
572}
573
574/*
575 * async writeback completion handler.
576 *
577 * If we get an error, set the mapping error bit, but not the individual
578 * page error bits.
579 */
580static void writepages_finish(struct ceph_osd_request *req,
581 struct ceph_msg *msg)
582{
583 struct inode *inode = req->r_inode;
584 struct ceph_inode_info *ci = ceph_inode(inode);
585 struct ceph_osd_data *osd_data;
586 unsigned wrote;
587 struct page *page;
588 int num_pages;
589 int i;
590 struct ceph_snap_context *snapc = req->r_snapc;
591 struct address_space *mapping = inode->i_mapping;
592 int rc = req->r_result;
593 u64 bytes = req->r_ops[0].extent.length;
594 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
595 long writeback_stat;
596 unsigned issued = ceph_caps_issued(ci);
597
598 osd_data = osd_req_op_extent_osd_data(req, 0);
599 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
600 num_pages = calc_pages_for((u64)osd_data->alignment,
601 (u64)osd_data->length);
602 if (rc >= 0) {
603 /*
604 * Assume we wrote the pages we originally sent. The
605 * osd might reply with fewer pages if our writeback
606 * raced with a truncation and was adjusted at the osd,
607 * so don't believe the reply.
608 */
609 wrote = num_pages;
610 } else {
611 wrote = 0;
612 mapping_set_error(mapping, rc);
613 }
614 dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
615 inode, rc, bytes, wrote);
616
617 /* clean all pages */
618 for (i = 0; i < num_pages; i++) {
619 page = osd_data->pages[i];
620 BUG_ON(!page);
621 WARN_ON(!PageUptodate(page));
622
623 writeback_stat =
624 atomic_long_dec_return(&fsc->writeback_count);
625 if (writeback_stat <
626 CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
627 clear_bdi_congested(&fsc->backing_dev_info,
628 BLK_RW_ASYNC);
629
630 ceph_put_snap_context(page_snap_context(page));
631 page->private = 0;
632 ClearPagePrivate(page);
633 dout("unlocking %d %p\n", i, page);
634 end_page_writeback(page);
635
636 /*
637 * We lost the cache cap, need to truncate the page before
638 * it is unlocked, otherwise we'd truncate it later in the
639 * page truncation thread, possibly losing some data that
640 * raced its way in
641 */
642 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
643 generic_error_remove_page(inode->i_mapping, page);
644
645 unlock_page(page);
646 }
647 dout("%p wrote+cleaned %d pages\n", inode, wrote);
648 ceph_put_wrbuffer_cap_refs(ci, num_pages, snapc);
649
650 ceph_release_pages(osd_data->pages, num_pages);
651 if (osd_data->pages_from_pool)
652 mempool_free(osd_data->pages,
653 ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
654 else
655 kfree(osd_data->pages);
656 ceph_osdc_put_request(req);
657}
658
659/*
660 * initiate async writeback
661 */
662static int ceph_writepages_start(struct address_space *mapping,
663 struct writeback_control *wbc)
664{
665 struct inode *inode = mapping->host;
666 struct ceph_inode_info *ci = ceph_inode(inode);
667 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
668 struct ceph_vino vino = ceph_vino(inode);
669 pgoff_t index, start, end;
670 int range_whole = 0;
671 int should_loop = 1;
672 pgoff_t max_pages = 0, max_pages_ever = 0;
673 struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
674 struct pagevec pvec;
675 int done = 0;
676 int rc = 0;
677 unsigned wsize = 1 << inode->i_blkbits;
678 struct ceph_osd_request *req = NULL;
679 int do_sync;
680 u64 truncate_size, snap_size;
681 u32 truncate_seq;
682
683 /*
684 * Include a 'sync' in the OSD request if this is a data
685 * integrity write (e.g., O_SYNC write or fsync()), or if our
686 * cap is being revoked.
687 */
688 if ((wbc->sync_mode == WB_SYNC_ALL) ||
689 ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
690 do_sync = 1;
691 dout("writepages_start %p dosync=%d (mode=%s)\n",
692 inode, do_sync,
693 wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
694 (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
695
696 if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
697 pr_warning("writepage_start %p on forced umount\n", inode);
698 return -EIO; /* we're in a forced umount, don't write! */
699 }
700 if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
701 wsize = fsc->mount_options->wsize;
702 if (wsize < PAGE_CACHE_SIZE)
703 wsize = PAGE_CACHE_SIZE;
704 max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
705
706 pagevec_init(&pvec, 0);
707
708 /* where to start/end? */
709 if (wbc->range_cyclic) {
710 start = mapping->writeback_index; /* Start from prev offset */
711 end = -1;
712 dout(" cyclic, start at %lu\n", start);
713 } else {
714 start = wbc->range_start >> PAGE_CACHE_SHIFT;
715 end = wbc->range_end >> PAGE_CACHE_SHIFT;
716 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
717 range_whole = 1;
718 should_loop = 0;
719 dout(" not cyclic, %lu to %lu\n", start, end);
720 }
721 index = start;
722
723retry:
724 /* find oldest snap context with dirty data */
725 ceph_put_snap_context(snapc);
726 snap_size = 0;
727 snapc = get_oldest_context(inode, &snap_size);
728 if (!snapc) {
729 /* hmm, why does writepages get called when there
730 is no dirty data? */
731 dout(" no snap context with dirty data?\n");
732 goto out;
733 }
734 if (snap_size == 0)
735 snap_size = i_size_read(inode);
736 dout(" oldest snapc is %p seq %lld (%d snaps)\n",
737 snapc, snapc->seq, snapc->num_snaps);
738
739 spin_lock(&ci->i_ceph_lock);
740 truncate_seq = ci->i_truncate_seq;
741 truncate_size = ci->i_truncate_size;
742 if (!snap_size)
743 snap_size = i_size_read(inode);
744 spin_unlock(&ci->i_ceph_lock);
745
746 if (last_snapc && snapc != last_snapc) {
747 /* if we switched to a newer snapc, restart our scan at the
748 * start of the original file range. */
749 dout(" snapc differs from last pass, restarting at %lu\n",
750 index);
751 index = start;
752 }
753 last_snapc = snapc;
754
755 while (!done && index <= end) {
756 int num_ops = do_sync ? 2 : 1;
757 unsigned i;
758 int first;
759 pgoff_t next;
760 int pvec_pages, locked_pages;
761 struct page **pages = NULL;
762 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
763 struct page *page;
764 int want;
765 u64 offset, len;
766 long writeback_stat;
767
768 next = 0;
769 locked_pages = 0;
770 max_pages = max_pages_ever;
771
772get_more_pages:
773 first = -1;
774 want = min(end - index,
775 min((pgoff_t)PAGEVEC_SIZE,
776 max_pages - (pgoff_t)locked_pages) - 1)
777 + 1;
778 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
779 PAGECACHE_TAG_DIRTY,
780 want);
781 dout("pagevec_lookup_tag got %d\n", pvec_pages);
782 if (!pvec_pages && !locked_pages)
783 break;
784 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
785 page = pvec.pages[i];
786 dout("? %p idx %lu\n", page, page->index);
787 if (locked_pages == 0)
788 lock_page(page); /* first page */
789 else if (!trylock_page(page))
790 break;
791
792 /* only dirty pages, or our accounting breaks */
793 if (unlikely(!PageDirty(page)) ||
794 unlikely(page->mapping != mapping)) {
795 dout("!dirty or !mapping %p\n", page);
796 unlock_page(page);
797 break;
798 }
799 if (!wbc->range_cyclic && page->index > end) {
800 dout("end of range %p\n", page);
801 done = 1;
802 unlock_page(page);
803 break;
804 }
805 if (next && (page->index != next)) {
806 dout("not consecutive %p\n", page);
807 unlock_page(page);
808 break;
809 }
810 if (wbc->sync_mode != WB_SYNC_NONE) {
811 dout("waiting on writeback %p\n", page);
812 wait_on_page_writeback(page);
813 }
814 if (page_offset(page) >= snap_size) {
815 dout("%p page eof %llu\n", page, snap_size);
816 done = 1;
817 unlock_page(page);
818 break;
819 }
820 if (PageWriteback(page)) {
821 dout("%p under writeback\n", page);
822 unlock_page(page);
823 break;
824 }
825
826 /* only if matching snap context */
827 pgsnapc = page_snap_context(page);
828 if (pgsnapc->seq > snapc->seq) {
829 dout("page snapc %p %lld > oldest %p %lld\n",
830 pgsnapc, pgsnapc->seq, snapc, snapc->seq);
831 unlock_page(page);
832 if (!locked_pages)
833 continue; /* keep looking for snap */
834 break;
835 }
836
837 if (!clear_page_dirty_for_io(page)) {
838 dout("%p !clear_page_dirty_for_io\n", page);
839 unlock_page(page);
840 break;
841 }
842
843 /*
844 * We have something to write. If this is
845 * the first locked page this time through,
846 * allocate an osd request and a page array
847 * that it will use.
848 */
849 if (locked_pages == 0) {
850 BUG_ON(pages);
851 /* prepare async write request */
852 offset = (u64)page_offset(page);
853 len = wsize;
854 req = ceph_osdc_new_request(&fsc->client->osdc,
855 &ci->i_layout, vino,
856 offset, &len, num_ops,
857 CEPH_OSD_OP_WRITE,
858 CEPH_OSD_FLAG_WRITE |
859 CEPH_OSD_FLAG_ONDISK,
860 snapc, truncate_seq,
861 truncate_size, true);
862 if (IS_ERR(req)) {
863 rc = PTR_ERR(req);
864 unlock_page(page);
865 break;
866 }
867
868 req->r_callback = writepages_finish;
869 req->r_inode = inode;
870
871 max_pages = calc_pages_for(0, (u64)len);
872 pages = kmalloc(max_pages * sizeof (*pages),
873 GFP_NOFS);
874 if (!pages) {
875 pool = fsc->wb_pagevec_pool;
876 pages = mempool_alloc(pool, GFP_NOFS);
877 BUG_ON(!pages);
878 }
879 }
880
881 /* note position of first page in pvec */
882 if (first < 0)
883 first = i;
884 dout("%p will write page %p idx %lu\n",
885 inode, page, page->index);
886
887 writeback_stat =
888 atomic_long_inc_return(&fsc->writeback_count);
889 if (writeback_stat > CONGESTION_ON_THRESH(
890 fsc->mount_options->congestion_kb)) {
891 set_bdi_congested(&fsc->backing_dev_info,
892 BLK_RW_ASYNC);
893 }
894
895 set_page_writeback(page);
896 pages[locked_pages] = page;
897 locked_pages++;
898 next = page->index + 1;
899 }
900
901 /* did we get anything? */
902 if (!locked_pages)
903 goto release_pvec_pages;
904 if (i) {
905 int j;
906 BUG_ON(!locked_pages || first < 0);
907
908 if (pvec_pages && i == pvec_pages &&
909 locked_pages < max_pages) {
910 dout("reached end pvec, trying for more\n");
911 pagevec_reinit(&pvec);
912 goto get_more_pages;
913 }
914
915 /* shift unused pages over in the pvec... we
916 * will need to release them below. */
917 for (j = i; j < pvec_pages; j++) {
918 dout(" pvec leftover page %p\n",
919 pvec.pages[j]);
920 pvec.pages[j-i+first] = pvec.pages[j];
921 }
922 pvec.nr -= i-first;
923 }
924
925 /* Format the osd request message and submit the write */
926
927 offset = page_offset(pages[0]);
928 len = min(snap_size - offset,
929 (u64)locked_pages << PAGE_CACHE_SHIFT);
930 dout("writepages got %d pages at %llu~%llu\n",
931 locked_pages, offset, len);
932
933 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
934 !!pool, false);
935
936 pages = NULL; /* request message now owns the pages array */
937 pool = NULL;
938
939 /* Update the write op length in case we changed it */
940
941 osd_req_op_extent_update(req, 0, len);
942
943 vino = ceph_vino(inode);
944 ceph_osdc_build_request(req, offset, snapc, vino.snap,
945 &inode->i_mtime);
946
947 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
948 BUG_ON(rc);
949 req = NULL;
950
951 /* continue? */
952 index = next;
953 wbc->nr_to_write -= locked_pages;
954 if (wbc->nr_to_write <= 0)
955 done = 1;
956
957release_pvec_pages:
958 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
959 pvec.nr ? pvec.pages[0] : NULL);
960 pagevec_release(&pvec);
961
962 if (locked_pages && !done)
963 goto retry;
964 }
965
966 if (should_loop && !done) {
967 /* more to do; loop back to beginning of file */
968 dout("writepages looping back to beginning of file\n");
969 should_loop = 0;
970 index = 0;
971 goto retry;
972 }
973
974 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
975 mapping->writeback_index = index;
976
977out:
978 if (req)
979 ceph_osdc_put_request(req);
980 ceph_put_snap_context(snapc);
981 dout("writepages done, rc = %d\n", rc);
982 return rc;
983}
984
985
986
987/*
988 * See if a given @snapc is either writeable, or already written.
989 */
990static int context_is_writeable_or_written(struct inode *inode,
991 struct ceph_snap_context *snapc)
992{
993 struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
994 int ret = !oldest || snapc->seq <= oldest->seq;
995
996 ceph_put_snap_context(oldest);
997 return ret;
998}
999
1000/*
1001 * We are only allowed to write into/dirty the page if the page is
1002 * clean, or already dirty within the same snap context.
1003 *
1004 * called with page locked.
1005 * return success with page locked,
1006 * or any failure (incl -EAGAIN) with page unlocked.
1007 */
1008static int ceph_update_writeable_page(struct file *file,
1009 loff_t pos, unsigned len,
1010 struct page *page)
1011{
1012 struct inode *inode = file_inode(file);
1013 struct ceph_inode_info *ci = ceph_inode(inode);
1014 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1015 loff_t page_off = pos & PAGE_CACHE_MASK;
1016 int pos_in_page = pos & ~PAGE_CACHE_MASK;
1017 int end_in_page = pos_in_page + len;
1018 loff_t i_size;
1019 int r;
1020 struct ceph_snap_context *snapc, *oldest;
1021
1022retry_locked:
1023 /* writepages currently holds page lock, but if we change that later, */
1024 wait_on_page_writeback(page);
1025
1026 /* check snap context */
1027 BUG_ON(!ci->i_snap_realm);
1028 down_read(&mdsc->snap_rwsem);
1029 BUG_ON(!ci->i_snap_realm->cached_context);
1030 snapc = page_snap_context(page);
1031 if (snapc && snapc != ci->i_head_snapc) {
1032 /*
1033 * this page is already dirty in another (older) snap
1034 * context! is it writeable now?
1035 */
1036 oldest = get_oldest_context(inode, NULL);
1037 up_read(&mdsc->snap_rwsem);
1038
1039 if (snapc->seq > oldest->seq) {
1040 ceph_put_snap_context(oldest);
1041 dout(" page %p snapc %p not current or oldest\n",
1042 page, snapc);
1043 /*
1044 * queue for writeback, and wait for snapc to
1045 * be writeable or written
1046 */
1047 snapc = ceph_get_snap_context(snapc);
1048 unlock_page(page);
1049 ceph_queue_writeback(inode);
1050 r = wait_event_interruptible(ci->i_cap_wq,
1051 context_is_writeable_or_written(inode, snapc));
1052 ceph_put_snap_context(snapc);
1053 if (r == -ERESTARTSYS)
1054 return r;
1055 return -EAGAIN;
1056 }
1057 ceph_put_snap_context(oldest);
1058
1059 /* yay, writeable, do it now (without dropping page lock) */
1060 dout(" page %p snapc %p not current, but oldest\n",
1061 page, snapc);
1062 if (!clear_page_dirty_for_io(page))
1063 goto retry_locked;
1064 r = writepage_nounlock(page, NULL);
1065 if (r < 0)
1066 goto fail_nosnap;
1067 goto retry_locked;
1068 }
1069
1070 if (PageUptodate(page)) {
1071 dout(" page %p already uptodate\n", page);
1072 return 0;
1073 }
1074
1075 /* full page? */
1076 if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
1077 return 0;
1078
1079 /* past end of file? */
1080 i_size = inode->i_size; /* caller holds i_mutex */
1081
1082 if (i_size + len > inode->i_sb->s_maxbytes) {
1083 /* file is too big */
1084 r = -EINVAL;
1085 goto fail;
1086 }
1087
1088 if (page_off >= i_size ||
1089 (pos_in_page == 0 && (pos+len) >= i_size &&
1090 end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1091 dout(" zeroing %p 0 - %d and %d - %d\n",
1092 page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1093 zero_user_segments(page,
1094 0, pos_in_page,
1095 end_in_page, PAGE_CACHE_SIZE);
1096 return 0;
1097 }
1098
1099 /* we need to read it. */
1100 up_read(&mdsc->snap_rwsem);
1101 r = readpage_nounlock(file, page);
1102 if (r < 0)
1103 goto fail_nosnap;
1104 goto retry_locked;
1105
1106fail:
1107 up_read(&mdsc->snap_rwsem);
1108fail_nosnap:
1109 unlock_page(page);
1110 return r;
1111}
1112
1113/*
1114 * We are only allowed to write into/dirty the page if the page is
1115 * clean, or already dirty within the same snap context.
1116 */
1117static int ceph_write_begin(struct file *file, struct address_space *mapping,
1118 loff_t pos, unsigned len, unsigned flags,
1119 struct page **pagep, void **fsdata)
1120{
1121 struct inode *inode = file_inode(file);
1122 struct page *page;
1123 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1124 int r;
1125
1126 do {
1127 /* get a page */
1128 page = grab_cache_page_write_begin(mapping, index, 0);
1129 if (!page)
1130 return -ENOMEM;
1131 *pagep = page;
1132
1133 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1134 inode, page, (int)pos, (int)len);
1135
1136 r = ceph_update_writeable_page(file, pos, len, page);
1137 } while (r == -EAGAIN);
1138
1139 return r;
1140}
1141
1142/*
1143 * we don't do anything in here that simple_write_end doesn't do
1144 * except adjust dirty page accounting and drop read lock on
1145 * mdsc->snap_rwsem.
1146 */
1147static int ceph_write_end(struct file *file, struct address_space *mapping,
1148 loff_t pos, unsigned len, unsigned copied,
1149 struct page *page, void *fsdata)
1150{
1151 struct inode *inode = file_inode(file);
1152 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1153 struct ceph_mds_client *mdsc = fsc->mdsc;
1154 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1155 int check_cap = 0;
1156
1157 dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1158 inode, page, (int)pos, (int)copied, (int)len);
1159
1160 /* zero the stale part of the page if we did a short copy */
1161 if (copied < len)
1162 zero_user_segment(page, from+copied, len);
1163
1164 /* did file size increase? */
1165 /* (no need for i_size_read(); we caller holds i_mutex */
1166 if (pos+copied > inode->i_size)
1167 check_cap = ceph_inode_set_size(inode, pos+copied);
1168
1169 if (!PageUptodate(page))
1170 SetPageUptodate(page);
1171
1172 set_page_dirty(page);
1173
1174 unlock_page(page);
1175 up_read(&mdsc->snap_rwsem);
1176 page_cache_release(page);
1177
1178 if (check_cap)
1179 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1180
1181 return copied;
1182}
1183
1184/*
1185 * we set .direct_IO to indicate direct io is supported, but since we
1186 * intercept O_DIRECT reads and writes early, this function should
1187 * never get called.
1188 */
1189static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1190 const struct iovec *iov,
1191 loff_t pos, unsigned long nr_segs)
1192{
1193 WARN_ON(1);
1194 return -EINVAL;
1195}
1196
1197const struct address_space_operations ceph_aops = {
1198 .readpage = ceph_readpage,
1199 .readpages = ceph_readpages,
1200 .writepage = ceph_writepage,
1201 .writepages = ceph_writepages_start,
1202 .write_begin = ceph_write_begin,
1203 .write_end = ceph_write_end,
1204 .set_page_dirty = ceph_set_page_dirty,
1205 .invalidatepage = ceph_invalidatepage,
1206 .releasepage = ceph_releasepage,
1207 .direct_IO = ceph_direct_io,
1208};
1209
1210
1211/*
1212 * vm ops
1213 */
1214static int ceph_filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1215{
1216 struct inode *inode = file_inode(vma->vm_file);
1217 struct ceph_inode_info *ci = ceph_inode(inode);
1218 struct ceph_file_info *fi = vma->vm_file->private_data;
1219 loff_t off = vmf->pgoff << PAGE_CACHE_SHIFT;
1220 int want, got, ret;
1221
1222 dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1223 inode, ceph_vinop(inode), off, (size_t)PAGE_CACHE_SIZE);
1224 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1225 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1226 else
1227 want = CEPH_CAP_FILE_CACHE;
1228 while (1) {
1229 got = 0;
1230 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, &got, -1);
1231 if (ret == 0)
1232 break;
1233 if (ret != -ERESTARTSYS) {
1234 WARN_ON(1);
1235 return VM_FAULT_SIGBUS;
1236 }
1237 }
1238 dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1239 inode, off, (size_t)PAGE_CACHE_SIZE, ceph_cap_string(got));
1240
1241 ret = filemap_fault(vma, vmf);
1242
1243 dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1244 inode, off, (size_t)PAGE_CACHE_SIZE, ceph_cap_string(got), ret);
1245 ceph_put_cap_refs(ci, got);
1246
1247 return ret;
1248}
1249
1250/*
1251 * Reuse write_begin here for simplicity.
1252 */
1253static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1254{
1255 struct inode *inode = file_inode(vma->vm_file);
1256 struct ceph_inode_info *ci = ceph_inode(inode);
1257 struct ceph_file_info *fi = vma->vm_file->private_data;
1258 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1259 struct page *page = vmf->page;
1260 loff_t off = page_offset(page);
1261 loff_t size = i_size_read(inode);
1262 size_t len;
1263 int want, got, ret;
1264
1265 if (off + PAGE_CACHE_SIZE <= size)
1266 len = PAGE_CACHE_SIZE;
1267 else
1268 len = size & ~PAGE_CACHE_MASK;
1269
1270 dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1271 inode, ceph_vinop(inode), off, len, size);
1272 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1273 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1274 else
1275 want = CEPH_CAP_FILE_BUFFER;
1276 while (1) {
1277 got = 0;
1278 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, &got, off + len);
1279 if (ret == 0)
1280 break;
1281 if (ret != -ERESTARTSYS) {
1282 WARN_ON(1);
1283 return VM_FAULT_SIGBUS;
1284 }
1285 }
1286 dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1287 inode, off, len, ceph_cap_string(got));
1288
1289 /* Update time before taking page lock */
1290 file_update_time(vma->vm_file);
1291
1292 lock_page(page);
1293
1294 ret = VM_FAULT_NOPAGE;
1295 if ((off > size) ||
1296 (page->mapping != inode->i_mapping))
1297 goto out;
1298
1299 ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1300 if (ret == 0) {
1301 /* success. we'll keep the page locked. */
1302 set_page_dirty(page);
1303 up_read(&mdsc->snap_rwsem);
1304 ret = VM_FAULT_LOCKED;
1305 } else {
1306 if (ret == -ENOMEM)
1307 ret = VM_FAULT_OOM;
1308 else
1309 ret = VM_FAULT_SIGBUS;
1310 }
1311out:
1312 if (ret != VM_FAULT_LOCKED) {
1313 unlock_page(page);
1314 } else {
1315 int dirty;
1316 spin_lock(&ci->i_ceph_lock);
1317 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR);
1318 spin_unlock(&ci->i_ceph_lock);
1319 if (dirty)
1320 __mark_inode_dirty(inode, dirty);
1321 }
1322
1323 dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1324 inode, off, len, ceph_cap_string(got), ret);
1325 ceph_put_cap_refs(ci, got);
1326
1327 return ret;
1328}
1329
1330static struct vm_operations_struct ceph_vmops = {
1331 .fault = ceph_filemap_fault,
1332 .page_mkwrite = ceph_page_mkwrite,
1333 .remap_pages = generic_file_remap_pages,
1334};
1335
1336int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1337{
1338 struct address_space *mapping = file->f_mapping;
1339
1340 if (!mapping->a_ops->readpage)
1341 return -ENOEXEC;
1342 file_accessed(file);
1343 vma->vm_ops = &ceph_vmops;
1344 return 0;
1345}