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 /* dirty the head */
91 spin_lock(&ci->i_ceph_lock);
92 BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
93 if (__ceph_have_pending_cap_snap(ci)) {
94 struct ceph_cap_snap *capsnap =
95 list_last_entry(&ci->i_cap_snaps,
96 struct ceph_cap_snap,
97 ci_item);
98 snapc = ceph_get_snap_context(capsnap->context);
99 capsnap->dirty_pages++;
100 } else {
101 BUG_ON(!ci->i_head_snapc);
102 snapc = ceph_get_snap_context(ci->i_head_snapc);
103 ++ci->i_wrbuffer_ref_head;
104 }
105 if (ci->i_wrbuffer_ref == 0)
106 ihold(inode);
107 ++ci->i_wrbuffer_ref;
108 dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
109 "snapc %p seq %lld (%d snaps)\n",
110 mapping->host, page, page->index,
111 ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
112 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
113 snapc, snapc->seq, snapc->num_snaps);
114 spin_unlock(&ci->i_ceph_lock);
115
116 /*
117 * Reference snap context in page->private. Also set
118 * PagePrivate so that we get invalidatepage callback.
119 */
120 BUG_ON(PagePrivate(page));
121 page->private = (unsigned long)snapc;
122 SetPagePrivate(page);
123
124 ret = __set_page_dirty_nobuffers(page);
125 WARN_ON(!PageLocked(page));
126 WARN_ON(!page->mapping);
127
128 return ret;
129}
130
131/*
132 * If we are truncating the full page (i.e. offset == 0), adjust the
133 * dirty page counters appropriately. Only called if there is private
134 * data on the page.
135 */
136static void ceph_invalidatepage(struct page *page, unsigned int offset,
137 unsigned int length)
138{
139 struct inode *inode;
140 struct ceph_inode_info *ci;
141 struct ceph_snap_context *snapc = page_snap_context(page);
142
143 inode = page->mapping->host;
144 ci = ceph_inode(inode);
145
146 if (offset != 0 || length != PAGE_SIZE) {
147 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
148 inode, page, page->index, offset, length);
149 return;
150 }
151
152 ceph_invalidate_fscache_page(inode, page);
153
154 if (!PagePrivate(page))
155 return;
156
157 /*
158 * We can get non-dirty pages here due to races between
159 * set_page_dirty and truncate_complete_page; just spit out a
160 * warning, in case we end up with accounting problems later.
161 */
162 if (!PageDirty(page))
163 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
164
165 ClearPageChecked(page);
166
167 dout("%p invalidatepage %p idx %lu full dirty page\n",
168 inode, page, page->index);
169
170 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
171 ceph_put_snap_context(snapc);
172 page->private = 0;
173 ClearPagePrivate(page);
174}
175
176static int ceph_releasepage(struct page *page, gfp_t g)
177{
178 dout("%p releasepage %p idx %lu (%sdirty)\n", page->mapping->host,
179 page, page->index, PageDirty(page) ? "" : "not ");
180
181 /* Can we release the page from the cache? */
182 if (!ceph_release_fscache_page(page, g))
183 return 0;
184
185 return !PagePrivate(page);
186}
187
188/*
189 * read a single page, without unlocking it.
190 */
191static int readpage_nounlock(struct file *filp, struct page *page)
192{
193 struct inode *inode = file_inode(filp);
194 struct ceph_inode_info *ci = ceph_inode(inode);
195 struct ceph_osd_client *osdc =
196 &ceph_inode_to_client(inode)->client->osdc;
197 int err = 0;
198 u64 off = page_offset(page);
199 u64 len = PAGE_SIZE;
200
201 if (off >= i_size_read(inode)) {
202 zero_user_segment(page, 0, PAGE_SIZE);
203 SetPageUptodate(page);
204 return 0;
205 }
206
207 if (ci->i_inline_version != CEPH_INLINE_NONE) {
208 /*
209 * Uptodate inline data should have been added
210 * into page cache while getting Fcr caps.
211 */
212 if (off == 0)
213 return -EINVAL;
214 zero_user_segment(page, 0, PAGE_SIZE);
215 SetPageUptodate(page);
216 return 0;
217 }
218
219 err = ceph_readpage_from_fscache(inode, page);
220 if (err == 0)
221 goto out;
222
223 dout("readpage inode %p file %p page %p index %lu\n",
224 inode, filp, page, page->index);
225 err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
226 off, &len,
227 ci->i_truncate_seq, ci->i_truncate_size,
228 &page, 1, 0);
229 if (err == -ENOENT)
230 err = 0;
231 if (err < 0) {
232 SetPageError(page);
233 ceph_fscache_readpage_cancel(inode, page);
234 goto out;
235 }
236 if (err < PAGE_SIZE)
237 /* zero fill remainder of page */
238 zero_user_segment(page, err, PAGE_SIZE);
239 else
240 flush_dcache_page(page);
241
242 SetPageUptodate(page);
243 ceph_readpage_to_fscache(inode, page);
244
245out:
246 return err < 0 ? err : 0;
247}
248
249static int ceph_readpage(struct file *filp, struct page *page)
250{
251 int r = readpage_nounlock(filp, page);
252 unlock_page(page);
253 return r;
254}
255
256/*
257 * Finish an async read(ahead) op.
258 */
259static void finish_read(struct ceph_osd_request *req)
260{
261 struct inode *inode = req->r_inode;
262 struct ceph_osd_data *osd_data;
263 int rc = req->r_result <= 0 ? req->r_result : 0;
264 int bytes = req->r_result >= 0 ? req->r_result : 0;
265 int num_pages;
266 int i;
267
268 dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
269
270 /* unlock all pages, zeroing any data we didn't read */
271 osd_data = osd_req_op_extent_osd_data(req, 0);
272 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
273 num_pages = calc_pages_for((u64)osd_data->alignment,
274 (u64)osd_data->length);
275 for (i = 0; i < num_pages; i++) {
276 struct page *page = osd_data->pages[i];
277
278 if (rc < 0 && rc != -ENOENT) {
279 ceph_fscache_readpage_cancel(inode, page);
280 goto unlock;
281 }
282 if (bytes < (int)PAGE_SIZE) {
283 /* zero (remainder of) page */
284 int s = bytes < 0 ? 0 : bytes;
285 zero_user_segment(page, s, PAGE_SIZE);
286 }
287 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
288 page->index);
289 flush_dcache_page(page);
290 SetPageUptodate(page);
291 ceph_readpage_to_fscache(inode, page);
292unlock:
293 unlock_page(page);
294 put_page(page);
295 bytes -= PAGE_SIZE;
296 }
297 kfree(osd_data->pages);
298}
299
300/*
301 * start an async read(ahead) operation. return nr_pages we submitted
302 * a read for on success, or negative error code.
303 */
304static int start_read(struct inode *inode, struct list_head *page_list, int max)
305{
306 struct ceph_osd_client *osdc =
307 &ceph_inode_to_client(inode)->client->osdc;
308 struct ceph_inode_info *ci = ceph_inode(inode);
309 struct page *page = list_entry(page_list->prev, struct page, lru);
310 struct ceph_vino vino;
311 struct ceph_osd_request *req;
312 u64 off;
313 u64 len;
314 int i;
315 struct page **pages;
316 pgoff_t next_index;
317 int nr_pages = 0;
318 int got = 0;
319 int ret = 0;
320
321 if (!current->journal_info) {
322 /* caller of readpages does not hold buffer and read caps
323 * (fadvise, madvise and readahead cases) */
324 int want = CEPH_CAP_FILE_CACHE;
325 ret = ceph_try_get_caps(ci, CEPH_CAP_FILE_RD, want, &got);
326 if (ret < 0) {
327 dout("start_read %p, error getting cap\n", inode);
328 } else if (!(got & want)) {
329 dout("start_read %p, no cache cap\n", inode);
330 ret = 0;
331 }
332 if (ret <= 0) {
333 if (got)
334 ceph_put_cap_refs(ci, got);
335 while (!list_empty(page_list)) {
336 page = list_entry(page_list->prev,
337 struct page, lru);
338 list_del(&page->lru);
339 put_page(page);
340 }
341 return ret;
342 }
343 }
344
345 off = (u64) page_offset(page);
346
347 /* count pages */
348 next_index = page->index;
349 list_for_each_entry_reverse(page, page_list, lru) {
350 if (page->index != next_index)
351 break;
352 nr_pages++;
353 next_index++;
354 if (max && nr_pages == max)
355 break;
356 }
357 len = nr_pages << PAGE_SHIFT;
358 dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
359 off, len);
360 vino = ceph_vino(inode);
361 req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
362 0, 1, CEPH_OSD_OP_READ,
363 CEPH_OSD_FLAG_READ, NULL,
364 ci->i_truncate_seq, ci->i_truncate_size,
365 false);
366 if (IS_ERR(req)) {
367 ret = PTR_ERR(req);
368 goto out;
369 }
370
371 /* build page vector */
372 nr_pages = calc_pages_for(0, len);
373 pages = kmalloc(sizeof(*pages) * nr_pages, GFP_KERNEL);
374 if (!pages) {
375 ret = -ENOMEM;
376 goto out_put;
377 }
378 for (i = 0; i < nr_pages; ++i) {
379 page = list_entry(page_list->prev, struct page, lru);
380 BUG_ON(PageLocked(page));
381 list_del(&page->lru);
382
383 dout("start_read %p adding %p idx %lu\n", inode, page,
384 page->index);
385 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
386 GFP_KERNEL)) {
387 ceph_fscache_uncache_page(inode, page);
388 put_page(page);
389 dout("start_read %p add_to_page_cache failed %p\n",
390 inode, page);
391 nr_pages = i;
392 if (nr_pages > 0) {
393 len = nr_pages << PAGE_SHIFT;
394 osd_req_op_extent_update(req, 0, len);
395 break;
396 }
397 goto out_pages;
398 }
399 pages[i] = page;
400 }
401 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
402 req->r_callback = finish_read;
403 req->r_inode = inode;
404
405 dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
406 ret = ceph_osdc_start_request(osdc, req, false);
407 if (ret < 0)
408 goto out_pages;
409 ceph_osdc_put_request(req);
410
411 /* After adding locked pages to page cache, the inode holds cache cap.
412 * So we can drop our cap refs. */
413 if (got)
414 ceph_put_cap_refs(ci, got);
415
416 return nr_pages;
417
418out_pages:
419 for (i = 0; i < nr_pages; ++i) {
420 ceph_fscache_readpage_cancel(inode, pages[i]);
421 unlock_page(pages[i]);
422 }
423 ceph_put_page_vector(pages, nr_pages, false);
424out_put:
425 ceph_osdc_put_request(req);
426out:
427 if (got)
428 ceph_put_cap_refs(ci, got);
429 return ret;
430}
431
432
433/*
434 * Read multiple pages. Leave pages we don't read + unlock in page_list;
435 * the caller (VM) cleans them up.
436 */
437static int ceph_readpages(struct file *file, struct address_space *mapping,
438 struct list_head *page_list, unsigned nr_pages)
439{
440 struct inode *inode = file_inode(file);
441 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
442 int rc = 0;
443 int max = 0;
444
445 if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
446 return -EINVAL;
447
448 rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
449 &nr_pages);
450
451 if (rc == 0)
452 goto out;
453
454 if (fsc->mount_options->rsize >= PAGE_SIZE)
455 max = (fsc->mount_options->rsize + PAGE_SIZE - 1)
456 >> PAGE_SHIFT;
457
458 dout("readpages %p file %p nr_pages %d max %d\n", inode,
459 file, nr_pages,
460 max);
461 while (!list_empty(page_list)) {
462 rc = start_read(inode, page_list, max);
463 if (rc < 0)
464 goto out;
465 }
466out:
467 ceph_fscache_readpages_cancel(inode, page_list);
468
469 dout("readpages %p file %p ret %d\n", inode, file, rc);
470 return rc;
471}
472
473/*
474 * Get ref for the oldest snapc for an inode with dirty data... that is, the
475 * only snap context we are allowed to write back.
476 */
477static struct ceph_snap_context *get_oldest_context(struct inode *inode,
478 loff_t *snap_size,
479 u64 *truncate_size,
480 u32 *truncate_seq)
481{
482 struct ceph_inode_info *ci = ceph_inode(inode);
483 struct ceph_snap_context *snapc = NULL;
484 struct ceph_cap_snap *capsnap = NULL;
485
486 spin_lock(&ci->i_ceph_lock);
487 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
488 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
489 capsnap->context, capsnap->dirty_pages);
490 if (capsnap->dirty_pages) {
491 snapc = ceph_get_snap_context(capsnap->context);
492 if (snap_size)
493 *snap_size = capsnap->size;
494 if (truncate_size)
495 *truncate_size = capsnap->truncate_size;
496 if (truncate_seq)
497 *truncate_seq = capsnap->truncate_seq;
498 break;
499 }
500 }
501 if (!snapc && ci->i_wrbuffer_ref_head) {
502 snapc = ceph_get_snap_context(ci->i_head_snapc);
503 dout(" head snapc %p has %d dirty pages\n",
504 snapc, ci->i_wrbuffer_ref_head);
505 if (truncate_size)
506 *truncate_size = ci->i_truncate_size;
507 if (truncate_seq)
508 *truncate_seq = ci->i_truncate_seq;
509 }
510 spin_unlock(&ci->i_ceph_lock);
511 return snapc;
512}
513
514/*
515 * Write a single page, but leave the page locked.
516 *
517 * If we get a write error, set the page error bit, but still adjust the
518 * dirty page accounting (i.e., page is no longer dirty).
519 */
520static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
521{
522 struct inode *inode;
523 struct ceph_inode_info *ci;
524 struct ceph_fs_client *fsc;
525 struct ceph_osd_client *osdc;
526 struct ceph_snap_context *snapc, *oldest;
527 loff_t page_off = page_offset(page);
528 loff_t snap_size = -1;
529 long writeback_stat;
530 u64 truncate_size;
531 u32 truncate_seq;
532 int err = 0, len = PAGE_SIZE;
533
534 dout("writepage %p idx %lu\n", page, page->index);
535
536 if (!page->mapping || !page->mapping->host) {
537 dout("writepage %p - no mapping\n", page);
538 return -EFAULT;
539 }
540 inode = page->mapping->host;
541 ci = ceph_inode(inode);
542 fsc = ceph_inode_to_client(inode);
543 osdc = &fsc->client->osdc;
544
545 /* verify this is a writeable snap context */
546 snapc = page_snap_context(page);
547 if (snapc == NULL) {
548 dout("writepage %p page %p not dirty?\n", inode, page);
549 goto out;
550 }
551 oldest = get_oldest_context(inode, &snap_size,
552 &truncate_size, &truncate_seq);
553 if (snapc->seq > oldest->seq) {
554 dout("writepage %p page %p snapc %p not writeable - noop\n",
555 inode, page, snapc);
556 /* we should only noop if called by kswapd */
557 WARN_ON((current->flags & PF_MEMALLOC) == 0);
558 ceph_put_snap_context(oldest);
559 goto out;
560 }
561 ceph_put_snap_context(oldest);
562
563 if (snap_size == -1)
564 snap_size = i_size_read(inode);
565
566 /* is this a partial page at end of file? */
567 if (page_off >= snap_size) {
568 dout("%p page eof %llu\n", page, snap_size);
569 goto out;
570 }
571 if (snap_size < page_off + len)
572 len = snap_size - page_off;
573
574 dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
575 inode, page, page->index, page_off, len, snapc);
576
577 writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
578 if (writeback_stat >
579 CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
580 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
581
582 set_page_writeback(page);
583 err = ceph_osdc_writepages(osdc, ceph_vino(inode),
584 &ci->i_layout, snapc,
585 page_off, len,
586 truncate_seq, truncate_size,
587 &inode->i_mtime, &page, 1);
588 if (err < 0) {
589 struct writeback_control tmp_wbc;
590 if (!wbc)
591 wbc = &tmp_wbc;
592 if (err == -ERESTARTSYS) {
593 /* killed by SIGKILL */
594 dout("writepage interrupted page %p\n", page);
595 redirty_page_for_writepage(wbc, page);
596 end_page_writeback(page);
597 goto out;
598 }
599 dout("writepage setting page/mapping error %d %p\n",
600 err, page);
601 SetPageError(page);
602 mapping_set_error(&inode->i_data, err);
603 wbc->pages_skipped++;
604 } else {
605 dout("writepage cleaned page %p\n", page);
606 err = 0; /* vfs expects us to return 0 */
607 }
608 page->private = 0;
609 ClearPagePrivate(page);
610 end_page_writeback(page);
611 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
612 ceph_put_snap_context(snapc); /* page's reference */
613out:
614 return err;
615}
616
617static int ceph_writepage(struct page *page, struct writeback_control *wbc)
618{
619 int err;
620 struct inode *inode = page->mapping->host;
621 BUG_ON(!inode);
622 ihold(inode);
623 err = writepage_nounlock(page, wbc);
624 if (err == -ERESTARTSYS) {
625 /* direct memory reclaimer was killed by SIGKILL. return 0
626 * to prevent caller from setting mapping/page error */
627 err = 0;
628 }
629 unlock_page(page);
630 iput(inode);
631 return err;
632}
633
634/*
635 * lame release_pages helper. release_pages() isn't exported to
636 * modules.
637 */
638static void ceph_release_pages(struct page **pages, int num)
639{
640 struct pagevec pvec;
641 int i;
642
643 pagevec_init(&pvec, 0);
644 for (i = 0; i < num; i++) {
645 if (pagevec_add(&pvec, pages[i]) == 0)
646 pagevec_release(&pvec);
647 }
648 pagevec_release(&pvec);
649}
650
651/*
652 * async writeback completion handler.
653 *
654 * If we get an error, set the mapping error bit, but not the individual
655 * page error bits.
656 */
657static void writepages_finish(struct ceph_osd_request *req)
658{
659 struct inode *inode = req->r_inode;
660 struct ceph_inode_info *ci = ceph_inode(inode);
661 struct ceph_osd_data *osd_data;
662 struct page *page;
663 int num_pages, total_pages = 0;
664 int i, j;
665 int rc = req->r_result;
666 struct ceph_snap_context *snapc = req->r_snapc;
667 struct address_space *mapping = inode->i_mapping;
668 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
669 bool remove_page;
670
671 dout("writepages_finish %p rc %d\n", inode, rc);
672 if (rc < 0)
673 mapping_set_error(mapping, rc);
674
675 /*
676 * We lost the cache cap, need to truncate the page before
677 * it is unlocked, otherwise we'd truncate it later in the
678 * page truncation thread, possibly losing some data that
679 * raced its way in
680 */
681 remove_page = !(ceph_caps_issued(ci) &
682 (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
683
684 /* clean all pages */
685 for (i = 0; i < req->r_num_ops; i++) {
686 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
687 break;
688
689 osd_data = osd_req_op_extent_osd_data(req, i);
690 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
691 num_pages = calc_pages_for((u64)osd_data->alignment,
692 (u64)osd_data->length);
693 total_pages += num_pages;
694 for (j = 0; j < num_pages; j++) {
695 page = osd_data->pages[j];
696 BUG_ON(!page);
697 WARN_ON(!PageUptodate(page));
698
699 if (atomic_long_dec_return(&fsc->writeback_count) <
700 CONGESTION_OFF_THRESH(
701 fsc->mount_options->congestion_kb))
702 clear_bdi_congested(&fsc->backing_dev_info,
703 BLK_RW_ASYNC);
704
705 if (rc < 0)
706 SetPageError(page);
707
708 ceph_put_snap_context(page_snap_context(page));
709 page->private = 0;
710 ClearPagePrivate(page);
711 dout("unlocking %p\n", page);
712 end_page_writeback(page);
713
714 if (remove_page)
715 generic_error_remove_page(inode->i_mapping,
716 page);
717
718 unlock_page(page);
719 }
720 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
721 inode, osd_data->length, rc >= 0 ? num_pages : 0);
722
723 ceph_release_pages(osd_data->pages, num_pages);
724 }
725
726 ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
727
728 osd_data = osd_req_op_extent_osd_data(req, 0);
729 if (osd_data->pages_from_pool)
730 mempool_free(osd_data->pages,
731 ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
732 else
733 kfree(osd_data->pages);
734 ceph_osdc_put_request(req);
735}
736
737/*
738 * initiate async writeback
739 */
740static int ceph_writepages_start(struct address_space *mapping,
741 struct writeback_control *wbc)
742{
743 struct inode *inode = mapping->host;
744 struct ceph_inode_info *ci = ceph_inode(inode);
745 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
746 struct ceph_vino vino = ceph_vino(inode);
747 pgoff_t index, start, end;
748 int range_whole = 0;
749 int should_loop = 1;
750 pgoff_t max_pages = 0, max_pages_ever = 0;
751 struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
752 struct pagevec pvec;
753 int done = 0;
754 int rc = 0;
755 unsigned wsize = 1 << inode->i_blkbits;
756 struct ceph_osd_request *req = NULL;
757 int do_sync = 0;
758 loff_t snap_size, i_size;
759 u64 truncate_size;
760 u32 truncate_seq;
761
762 /*
763 * Include a 'sync' in the OSD request if this is a data
764 * integrity write (e.g., O_SYNC write or fsync()), or if our
765 * cap is being revoked.
766 */
767 if ((wbc->sync_mode == WB_SYNC_ALL) ||
768 ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
769 do_sync = 1;
770 dout("writepages_start %p dosync=%d (mode=%s)\n",
771 inode, do_sync,
772 wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
773 (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
774
775 if (ACCESS_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
776 if (ci->i_wrbuffer_ref > 0) {
777 pr_warn_ratelimited(
778 "writepage_start %p %lld forced umount\n",
779 inode, ceph_ino(inode));
780 }
781 mapping_set_error(mapping, -EIO);
782 return -EIO; /* we're in a forced umount, don't write! */
783 }
784 if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
785 wsize = fsc->mount_options->wsize;
786 if (wsize < PAGE_SIZE)
787 wsize = PAGE_SIZE;
788 max_pages_ever = wsize >> PAGE_SHIFT;
789
790 pagevec_init(&pvec, 0);
791
792 /* where to start/end? */
793 if (wbc->range_cyclic) {
794 start = mapping->writeback_index; /* Start from prev offset */
795 end = -1;
796 dout(" cyclic, start at %lu\n", start);
797 } else {
798 start = wbc->range_start >> PAGE_SHIFT;
799 end = wbc->range_end >> PAGE_SHIFT;
800 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
801 range_whole = 1;
802 should_loop = 0;
803 dout(" not cyclic, %lu to %lu\n", start, end);
804 }
805 index = start;
806
807retry:
808 /* find oldest snap context with dirty data */
809 ceph_put_snap_context(snapc);
810 snap_size = -1;
811 snapc = get_oldest_context(inode, &snap_size,
812 &truncate_size, &truncate_seq);
813 if (!snapc) {
814 /* hmm, why does writepages get called when there
815 is no dirty data? */
816 dout(" no snap context with dirty data?\n");
817 goto out;
818 }
819 dout(" oldest snapc is %p seq %lld (%d snaps)\n",
820 snapc, snapc->seq, snapc->num_snaps);
821
822 i_size = i_size_read(inode);
823
824 if (last_snapc && snapc != last_snapc) {
825 /* if we switched to a newer snapc, restart our scan at the
826 * start of the original file range. */
827 dout(" snapc differs from last pass, restarting at %lu\n",
828 index);
829 index = start;
830 }
831 last_snapc = snapc;
832
833 while (!done && index <= end) {
834 unsigned i;
835 int first;
836 pgoff_t strip_unit_end = 0;
837 int num_ops = 0, op_idx;
838 int pvec_pages, locked_pages = 0;
839 struct page **pages = NULL, **data_pages;
840 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
841 struct page *page;
842 int want;
843 u64 offset = 0, len = 0;
844
845 max_pages = max_pages_ever;
846
847get_more_pages:
848 first = -1;
849 want = min(end - index,
850 min((pgoff_t)PAGEVEC_SIZE,
851 max_pages - (pgoff_t)locked_pages) - 1)
852 + 1;
853 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
854 PAGECACHE_TAG_DIRTY,
855 want);
856 dout("pagevec_lookup_tag got %d\n", pvec_pages);
857 if (!pvec_pages && !locked_pages)
858 break;
859 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
860 page = pvec.pages[i];
861 dout("? %p idx %lu\n", page, page->index);
862 if (locked_pages == 0)
863 lock_page(page); /* first page */
864 else if (!trylock_page(page))
865 break;
866
867 /* only dirty pages, or our accounting breaks */
868 if (unlikely(!PageDirty(page)) ||
869 unlikely(page->mapping != mapping)) {
870 dout("!dirty or !mapping %p\n", page);
871 unlock_page(page);
872 break;
873 }
874 if (!wbc->range_cyclic && page->index > end) {
875 dout("end of range %p\n", page);
876 done = 1;
877 unlock_page(page);
878 break;
879 }
880 if (strip_unit_end && (page->index > strip_unit_end)) {
881 dout("end of strip unit %p\n", page);
882 unlock_page(page);
883 break;
884 }
885 if (wbc->sync_mode != WB_SYNC_NONE) {
886 dout("waiting on writeback %p\n", page);
887 wait_on_page_writeback(page);
888 }
889 if (page_offset(page) >=
890 (snap_size == -1 ? i_size : snap_size)) {
891 dout("%p page eof %llu\n", page,
892 (snap_size == -1 ? i_size : snap_size));
893 done = 1;
894 unlock_page(page);
895 break;
896 }
897 if (PageWriteback(page)) {
898 dout("%p under writeback\n", page);
899 unlock_page(page);
900 break;
901 }
902
903 /* only if matching snap context */
904 pgsnapc = page_snap_context(page);
905 if (pgsnapc->seq > snapc->seq) {
906 dout("page snapc %p %lld > oldest %p %lld\n",
907 pgsnapc, pgsnapc->seq, snapc, snapc->seq);
908 unlock_page(page);
909 if (!locked_pages)
910 continue; /* keep looking for snap */
911 break;
912 }
913
914 if (!clear_page_dirty_for_io(page)) {
915 dout("%p !clear_page_dirty_for_io\n", page);
916 unlock_page(page);
917 break;
918 }
919
920 /*
921 * We have something to write. If this is
922 * the first locked page this time through,
923 * calculate max possinle write size and
924 * allocate a page array
925 */
926 if (locked_pages == 0) {
927 u64 objnum;
928 u64 objoff;
929
930 /* prepare async write request */
931 offset = (u64)page_offset(page);
932 len = wsize;
933
934 rc = ceph_calc_file_object_mapping(&ci->i_layout,
935 offset, len,
936 &objnum, &objoff,
937 &len);
938 if (rc < 0) {
939 unlock_page(page);
940 break;
941 }
942
943 num_ops = 1 + do_sync;
944 strip_unit_end = page->index +
945 ((len - 1) >> PAGE_SHIFT);
946
947 BUG_ON(pages);
948 max_pages = calc_pages_for(0, (u64)len);
949 pages = kmalloc(max_pages * sizeof (*pages),
950 GFP_NOFS);
951 if (!pages) {
952 pool = fsc->wb_pagevec_pool;
953 pages = mempool_alloc(pool, GFP_NOFS);
954 BUG_ON(!pages);
955 }
956
957 len = 0;
958 } else if (page->index !=
959 (offset + len) >> PAGE_SHIFT) {
960 if (num_ops >= (pool ? CEPH_OSD_SLAB_OPS :
961 CEPH_OSD_MAX_OPS)) {
962 redirty_page_for_writepage(wbc, page);
963 unlock_page(page);
964 break;
965 }
966
967 num_ops++;
968 offset = (u64)page_offset(page);
969 len = 0;
970 }
971
972 /* note position of first page in pvec */
973 if (first < 0)
974 first = i;
975 dout("%p will write page %p idx %lu\n",
976 inode, page, page->index);
977
978 if (atomic_long_inc_return(&fsc->writeback_count) >
979 CONGESTION_ON_THRESH(
980 fsc->mount_options->congestion_kb)) {
981 set_bdi_congested(&fsc->backing_dev_info,
982 BLK_RW_ASYNC);
983 }
984
985 pages[locked_pages] = page;
986 locked_pages++;
987 len += PAGE_SIZE;
988 }
989
990 /* did we get anything? */
991 if (!locked_pages)
992 goto release_pvec_pages;
993 if (i) {
994 int j;
995 BUG_ON(!locked_pages || first < 0);
996
997 if (pvec_pages && i == pvec_pages &&
998 locked_pages < max_pages) {
999 dout("reached end pvec, trying for more\n");
1000 pagevec_reinit(&pvec);
1001 goto get_more_pages;
1002 }
1003
1004 /* shift unused pages over in the pvec... we
1005 * will need to release them below. */
1006 for (j = i; j < pvec_pages; j++) {
1007 dout(" pvec leftover page %p\n", pvec.pages[j]);
1008 pvec.pages[j-i+first] = pvec.pages[j];
1009 }
1010 pvec.nr -= i-first;
1011 }
1012
1013new_request:
1014 offset = page_offset(pages[0]);
1015 len = wsize;
1016
1017 req = ceph_osdc_new_request(&fsc->client->osdc,
1018 &ci->i_layout, vino,
1019 offset, &len, 0, num_ops,
1020 CEPH_OSD_OP_WRITE,
1021 CEPH_OSD_FLAG_WRITE |
1022 CEPH_OSD_FLAG_ONDISK,
1023 snapc, truncate_seq,
1024 truncate_size, false);
1025 if (IS_ERR(req)) {
1026 req = ceph_osdc_new_request(&fsc->client->osdc,
1027 &ci->i_layout, vino,
1028 offset, &len, 0,
1029 min(num_ops,
1030 CEPH_OSD_SLAB_OPS),
1031 CEPH_OSD_OP_WRITE,
1032 CEPH_OSD_FLAG_WRITE |
1033 CEPH_OSD_FLAG_ONDISK,
1034 snapc, truncate_seq,
1035 truncate_size, true);
1036 BUG_ON(IS_ERR(req));
1037 }
1038 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1039 PAGE_SIZE - offset);
1040
1041 req->r_callback = writepages_finish;
1042 req->r_inode = inode;
1043
1044 /* Format the osd request message and submit the write */
1045 len = 0;
1046 data_pages = pages;
1047 op_idx = 0;
1048 for (i = 0; i < locked_pages; i++) {
1049 u64 cur_offset = page_offset(pages[i]);
1050 if (offset + len != cur_offset) {
1051 if (op_idx + do_sync + 1 == req->r_num_ops)
1052 break;
1053 osd_req_op_extent_dup_last(req, op_idx,
1054 cur_offset - offset);
1055 dout("writepages got pages at %llu~%llu\n",
1056 offset, len);
1057 osd_req_op_extent_osd_data_pages(req, op_idx,
1058 data_pages, len, 0,
1059 !!pool, false);
1060 osd_req_op_extent_update(req, op_idx, len);
1061
1062 len = 0;
1063 offset = cur_offset;
1064 data_pages = pages + i;
1065 op_idx++;
1066 }
1067
1068 set_page_writeback(pages[i]);
1069 len += PAGE_SIZE;
1070 }
1071
1072 if (snap_size != -1) {
1073 len = min(len, snap_size - offset);
1074 } else if (i == locked_pages) {
1075 /* writepages_finish() clears writeback pages
1076 * according to the data length, so make sure
1077 * data length covers all locked pages */
1078 u64 min_len = len + 1 - PAGE_SIZE;
1079 len = min(len, (u64)i_size_read(inode) - offset);
1080 len = max(len, min_len);
1081 }
1082 dout("writepages got pages at %llu~%llu\n", offset, len);
1083
1084 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1085 0, !!pool, false);
1086 osd_req_op_extent_update(req, op_idx, len);
1087
1088 if (do_sync) {
1089 op_idx++;
1090 osd_req_op_init(req, op_idx, CEPH_OSD_OP_STARTSYNC, 0);
1091 }
1092 BUG_ON(op_idx + 1 != req->r_num_ops);
1093
1094 pool = NULL;
1095 if (i < locked_pages) {
1096 BUG_ON(num_ops <= req->r_num_ops);
1097 num_ops -= req->r_num_ops;
1098 num_ops += do_sync;
1099 locked_pages -= i;
1100
1101 /* allocate new pages array for next request */
1102 data_pages = pages;
1103 pages = kmalloc(locked_pages * sizeof (*pages),
1104 GFP_NOFS);
1105 if (!pages) {
1106 pool = fsc->wb_pagevec_pool;
1107 pages = mempool_alloc(pool, GFP_NOFS);
1108 BUG_ON(!pages);
1109 }
1110 memcpy(pages, data_pages + i,
1111 locked_pages * sizeof(*pages));
1112 memset(data_pages + i, 0,
1113 locked_pages * sizeof(*pages));
1114 } else {
1115 BUG_ON(num_ops != req->r_num_ops);
1116 index = pages[i - 1]->index + 1;
1117 /* request message now owns the pages array */
1118 pages = NULL;
1119 }
1120
1121 req->r_mtime = inode->i_mtime;
1122 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1123 BUG_ON(rc);
1124 req = NULL;
1125
1126 wbc->nr_to_write -= i;
1127 if (pages)
1128 goto new_request;
1129
1130 if (wbc->nr_to_write <= 0)
1131 done = 1;
1132
1133release_pvec_pages:
1134 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1135 pvec.nr ? pvec.pages[0] : NULL);
1136 pagevec_release(&pvec);
1137
1138 if (locked_pages && !done)
1139 goto retry;
1140 }
1141
1142 if (should_loop && !done) {
1143 /* more to do; loop back to beginning of file */
1144 dout("writepages looping back to beginning of file\n");
1145 should_loop = 0;
1146 index = 0;
1147 goto retry;
1148 }
1149
1150 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1151 mapping->writeback_index = index;
1152
1153out:
1154 ceph_osdc_put_request(req);
1155 ceph_put_snap_context(snapc);
1156 dout("writepages done, rc = %d\n", rc);
1157 return rc;
1158}
1159
1160
1161
1162/*
1163 * See if a given @snapc is either writeable, or already written.
1164 */
1165static int context_is_writeable_or_written(struct inode *inode,
1166 struct ceph_snap_context *snapc)
1167{
1168 struct ceph_snap_context *oldest = get_oldest_context(inode, NULL,
1169 NULL, NULL);
1170 int ret = !oldest || snapc->seq <= oldest->seq;
1171
1172 ceph_put_snap_context(oldest);
1173 return ret;
1174}
1175
1176/*
1177 * We are only allowed to write into/dirty the page if the page is
1178 * clean, or already dirty within the same snap context.
1179 *
1180 * called with page locked.
1181 * return success with page locked,
1182 * or any failure (incl -EAGAIN) with page unlocked.
1183 */
1184static int ceph_update_writeable_page(struct file *file,
1185 loff_t pos, unsigned len,
1186 struct page *page)
1187{
1188 struct inode *inode = file_inode(file);
1189 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1190 struct ceph_inode_info *ci = ceph_inode(inode);
1191 loff_t page_off = pos & PAGE_MASK;
1192 int pos_in_page = pos & ~PAGE_MASK;
1193 int end_in_page = pos_in_page + len;
1194 loff_t i_size;
1195 int r;
1196 struct ceph_snap_context *snapc, *oldest;
1197
1198 if (ACCESS_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1199 dout(" page %p forced umount\n", page);
1200 unlock_page(page);
1201 return -EIO;
1202 }
1203
1204retry_locked:
1205 /* writepages currently holds page lock, but if we change that later, */
1206 wait_on_page_writeback(page);
1207
1208 snapc = page_snap_context(page);
1209 if (snapc && snapc != ci->i_head_snapc) {
1210 /*
1211 * this page is already dirty in another (older) snap
1212 * context! is it writeable now?
1213 */
1214 oldest = get_oldest_context(inode, NULL, NULL, NULL);
1215
1216 if (snapc->seq > oldest->seq) {
1217 ceph_put_snap_context(oldest);
1218 dout(" page %p snapc %p not current or oldest\n",
1219 page, snapc);
1220 /*
1221 * queue for writeback, and wait for snapc to
1222 * be writeable or written
1223 */
1224 snapc = ceph_get_snap_context(snapc);
1225 unlock_page(page);
1226 ceph_queue_writeback(inode);
1227 r = wait_event_killable(ci->i_cap_wq,
1228 context_is_writeable_or_written(inode, snapc));
1229 ceph_put_snap_context(snapc);
1230 if (r == -ERESTARTSYS)
1231 return r;
1232 return -EAGAIN;
1233 }
1234 ceph_put_snap_context(oldest);
1235
1236 /* yay, writeable, do it now (without dropping page lock) */
1237 dout(" page %p snapc %p not current, but oldest\n",
1238 page, snapc);
1239 if (!clear_page_dirty_for_io(page))
1240 goto retry_locked;
1241 r = writepage_nounlock(page, NULL);
1242 if (r < 0)
1243 goto fail_nosnap;
1244 goto retry_locked;
1245 }
1246
1247 if (PageUptodate(page)) {
1248 dout(" page %p already uptodate\n", page);
1249 return 0;
1250 }
1251
1252 /* full page? */
1253 if (pos_in_page == 0 && len == PAGE_SIZE)
1254 return 0;
1255
1256 /* past end of file? */
1257 i_size = i_size_read(inode);
1258
1259 if (page_off >= i_size ||
1260 (pos_in_page == 0 && (pos+len) >= i_size &&
1261 end_in_page - pos_in_page != PAGE_SIZE)) {
1262 dout(" zeroing %p 0 - %d and %d - %d\n",
1263 page, pos_in_page, end_in_page, (int)PAGE_SIZE);
1264 zero_user_segments(page,
1265 0, pos_in_page,
1266 end_in_page, PAGE_SIZE);
1267 return 0;
1268 }
1269
1270 /* we need to read it. */
1271 r = readpage_nounlock(file, page);
1272 if (r < 0)
1273 goto fail_nosnap;
1274 goto retry_locked;
1275fail_nosnap:
1276 unlock_page(page);
1277 return r;
1278}
1279
1280/*
1281 * We are only allowed to write into/dirty the page if the page is
1282 * clean, or already dirty within the same snap context.
1283 */
1284static int ceph_write_begin(struct file *file, struct address_space *mapping,
1285 loff_t pos, unsigned len, unsigned flags,
1286 struct page **pagep, void **fsdata)
1287{
1288 struct inode *inode = file_inode(file);
1289 struct page *page;
1290 pgoff_t index = pos >> PAGE_SHIFT;
1291 int r;
1292
1293 do {
1294 /* get a page */
1295 page = grab_cache_page_write_begin(mapping, index, 0);
1296 if (!page)
1297 return -ENOMEM;
1298
1299 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1300 inode, page, (int)pos, (int)len);
1301
1302 r = ceph_update_writeable_page(file, pos, len, page);
1303 if (r < 0)
1304 put_page(page);
1305 else
1306 *pagep = page;
1307 } while (r == -EAGAIN);
1308
1309 return r;
1310}
1311
1312/*
1313 * we don't do anything in here that simple_write_end doesn't do
1314 * except adjust dirty page accounting
1315 */
1316static int ceph_write_end(struct file *file, struct address_space *mapping,
1317 loff_t pos, unsigned len, unsigned copied,
1318 struct page *page, void *fsdata)
1319{
1320 struct inode *inode = file_inode(file);
1321 int check_cap = 0;
1322
1323 dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1324 inode, page, (int)pos, (int)copied, (int)len);
1325
1326 /* zero the stale part of the page if we did a short copy */
1327 if (!PageUptodate(page)) {
1328 if (copied < len) {
1329 copied = 0;
1330 goto out;
1331 }
1332 SetPageUptodate(page);
1333 }
1334
1335 /* did file size increase? */
1336 if (pos+copied > i_size_read(inode))
1337 check_cap = ceph_inode_set_size(inode, pos+copied);
1338
1339 set_page_dirty(page);
1340
1341out:
1342 unlock_page(page);
1343 put_page(page);
1344
1345 if (check_cap)
1346 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1347
1348 return copied;
1349}
1350
1351/*
1352 * we set .direct_IO to indicate direct io is supported, but since we
1353 * intercept O_DIRECT reads and writes early, this function should
1354 * never get called.
1355 */
1356static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
1357{
1358 WARN_ON(1);
1359 return -EINVAL;
1360}
1361
1362const struct address_space_operations ceph_aops = {
1363 .readpage = ceph_readpage,
1364 .readpages = ceph_readpages,
1365 .writepage = ceph_writepage,
1366 .writepages = ceph_writepages_start,
1367 .write_begin = ceph_write_begin,
1368 .write_end = ceph_write_end,
1369 .set_page_dirty = ceph_set_page_dirty,
1370 .invalidatepage = ceph_invalidatepage,
1371 .releasepage = ceph_releasepage,
1372 .direct_IO = ceph_direct_io,
1373};
1374
1375static void ceph_block_sigs(sigset_t *oldset)
1376{
1377 sigset_t mask;
1378 siginitsetinv(&mask, sigmask(SIGKILL));
1379 sigprocmask(SIG_BLOCK, &mask, oldset);
1380}
1381
1382static void ceph_restore_sigs(sigset_t *oldset)
1383{
1384 sigprocmask(SIG_SETMASK, oldset, NULL);
1385}
1386
1387/*
1388 * vm ops
1389 */
1390static int ceph_filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1391{
1392 struct inode *inode = file_inode(vma->vm_file);
1393 struct ceph_inode_info *ci = ceph_inode(inode);
1394 struct ceph_file_info *fi = vma->vm_file->private_data;
1395 struct page *pinned_page = NULL;
1396 loff_t off = vmf->pgoff << PAGE_SHIFT;
1397 int want, got, ret;
1398 sigset_t oldset;
1399
1400 ceph_block_sigs(&oldset);
1401
1402 dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1403 inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1404 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1405 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1406 else
1407 want = CEPH_CAP_FILE_CACHE;
1408
1409 got = 0;
1410 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1411 if (ret < 0)
1412 goto out_restore;
1413
1414 dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1415 inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1416
1417 if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1418 ci->i_inline_version == CEPH_INLINE_NONE) {
1419 current->journal_info = vma->vm_file;
1420 ret = filemap_fault(vma, vmf);
1421 current->journal_info = NULL;
1422 } else
1423 ret = -EAGAIN;
1424
1425 dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1426 inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got), ret);
1427 if (pinned_page)
1428 put_page(pinned_page);
1429 ceph_put_cap_refs(ci, got);
1430
1431 if (ret != -EAGAIN)
1432 goto out_restore;
1433
1434 /* read inline data */
1435 if (off >= PAGE_SIZE) {
1436 /* does not support inline data > PAGE_SIZE */
1437 ret = VM_FAULT_SIGBUS;
1438 } else {
1439 int ret1;
1440 struct address_space *mapping = inode->i_mapping;
1441 struct page *page = find_or_create_page(mapping, 0,
1442 mapping_gfp_constraint(mapping,
1443 ~__GFP_FS));
1444 if (!page) {
1445 ret = VM_FAULT_OOM;
1446 goto out_inline;
1447 }
1448 ret1 = __ceph_do_getattr(inode, page,
1449 CEPH_STAT_CAP_INLINE_DATA, true);
1450 if (ret1 < 0 || off >= i_size_read(inode)) {
1451 unlock_page(page);
1452 put_page(page);
1453 if (ret1 < 0)
1454 ret = ret1;
1455 else
1456 ret = VM_FAULT_SIGBUS;
1457 goto out_inline;
1458 }
1459 if (ret1 < PAGE_SIZE)
1460 zero_user_segment(page, ret1, PAGE_SIZE);
1461 else
1462 flush_dcache_page(page);
1463 SetPageUptodate(page);
1464 vmf->page = page;
1465 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1466out_inline:
1467 dout("filemap_fault %p %llu~%zd read inline data ret %d\n",
1468 inode, off, (size_t)PAGE_SIZE, ret);
1469 }
1470out_restore:
1471 ceph_restore_sigs(&oldset);
1472 if (ret < 0)
1473 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1474
1475 return ret;
1476}
1477
1478/*
1479 * Reuse write_begin here for simplicity.
1480 */
1481static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1482{
1483 struct inode *inode = file_inode(vma->vm_file);
1484 struct ceph_inode_info *ci = ceph_inode(inode);
1485 struct ceph_file_info *fi = vma->vm_file->private_data;
1486 struct ceph_cap_flush *prealloc_cf;
1487 struct page *page = vmf->page;
1488 loff_t off = page_offset(page);
1489 loff_t size = i_size_read(inode);
1490 size_t len;
1491 int want, got, ret;
1492 sigset_t oldset;
1493
1494 prealloc_cf = ceph_alloc_cap_flush();
1495 if (!prealloc_cf)
1496 return VM_FAULT_OOM;
1497
1498 ceph_block_sigs(&oldset);
1499
1500 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1501 struct page *locked_page = NULL;
1502 if (off == 0) {
1503 lock_page(page);
1504 locked_page = page;
1505 }
1506 ret = ceph_uninline_data(vma->vm_file, locked_page);
1507 if (locked_page)
1508 unlock_page(locked_page);
1509 if (ret < 0)
1510 goto out_free;
1511 }
1512
1513 if (off + PAGE_SIZE <= size)
1514 len = PAGE_SIZE;
1515 else
1516 len = size & ~PAGE_MASK;
1517
1518 dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1519 inode, ceph_vinop(inode), off, len, size);
1520 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1521 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1522 else
1523 want = CEPH_CAP_FILE_BUFFER;
1524
1525 got = 0;
1526 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1527 &got, NULL);
1528 if (ret < 0)
1529 goto out_free;
1530
1531 dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1532 inode, off, len, ceph_cap_string(got));
1533
1534 /* Update time before taking page lock */
1535 file_update_time(vma->vm_file);
1536
1537 do {
1538 lock_page(page);
1539
1540 if ((off > size) || (page->mapping != inode->i_mapping)) {
1541 unlock_page(page);
1542 ret = VM_FAULT_NOPAGE;
1543 break;
1544 }
1545
1546 ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1547 if (ret >= 0) {
1548 /* success. we'll keep the page locked. */
1549 set_page_dirty(page);
1550 ret = VM_FAULT_LOCKED;
1551 }
1552 } while (ret == -EAGAIN);
1553
1554 if (ret == VM_FAULT_LOCKED ||
1555 ci->i_inline_version != CEPH_INLINE_NONE) {
1556 int dirty;
1557 spin_lock(&ci->i_ceph_lock);
1558 ci->i_inline_version = CEPH_INLINE_NONE;
1559 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1560 &prealloc_cf);
1561 spin_unlock(&ci->i_ceph_lock);
1562 if (dirty)
1563 __mark_inode_dirty(inode, dirty);
1564 }
1565
1566 dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1567 inode, off, len, ceph_cap_string(got), ret);
1568 ceph_put_cap_refs(ci, got);
1569out_free:
1570 ceph_restore_sigs(&oldset);
1571 ceph_free_cap_flush(prealloc_cf);
1572 if (ret < 0)
1573 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1574 return ret;
1575}
1576
1577void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1578 char *data, size_t len)
1579{
1580 struct address_space *mapping = inode->i_mapping;
1581 struct page *page;
1582
1583 if (locked_page) {
1584 page = locked_page;
1585 } else {
1586 if (i_size_read(inode) == 0)
1587 return;
1588 page = find_or_create_page(mapping, 0,
1589 mapping_gfp_constraint(mapping,
1590 ~__GFP_FS));
1591 if (!page)
1592 return;
1593 if (PageUptodate(page)) {
1594 unlock_page(page);
1595 put_page(page);
1596 return;
1597 }
1598 }
1599
1600 dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1601 inode, ceph_vinop(inode), len, locked_page);
1602
1603 if (len > 0) {
1604 void *kaddr = kmap_atomic(page);
1605 memcpy(kaddr, data, len);
1606 kunmap_atomic(kaddr);
1607 }
1608
1609 if (page != locked_page) {
1610 if (len < PAGE_SIZE)
1611 zero_user_segment(page, len, PAGE_SIZE);
1612 else
1613 flush_dcache_page(page);
1614
1615 SetPageUptodate(page);
1616 unlock_page(page);
1617 put_page(page);
1618 }
1619}
1620
1621int ceph_uninline_data(struct file *filp, struct page *locked_page)
1622{
1623 struct inode *inode = file_inode(filp);
1624 struct ceph_inode_info *ci = ceph_inode(inode);
1625 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1626 struct ceph_osd_request *req;
1627 struct page *page = NULL;
1628 u64 len, inline_version;
1629 int err = 0;
1630 bool from_pagecache = false;
1631
1632 spin_lock(&ci->i_ceph_lock);
1633 inline_version = ci->i_inline_version;
1634 spin_unlock(&ci->i_ceph_lock);
1635
1636 dout("uninline_data %p %llx.%llx inline_version %llu\n",
1637 inode, ceph_vinop(inode), inline_version);
1638
1639 if (inline_version == 1 || /* initial version, no data */
1640 inline_version == CEPH_INLINE_NONE)
1641 goto out;
1642
1643 if (locked_page) {
1644 page = locked_page;
1645 WARN_ON(!PageUptodate(page));
1646 } else if (ceph_caps_issued(ci) &
1647 (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1648 page = find_get_page(inode->i_mapping, 0);
1649 if (page) {
1650 if (PageUptodate(page)) {
1651 from_pagecache = true;
1652 lock_page(page);
1653 } else {
1654 put_page(page);
1655 page = NULL;
1656 }
1657 }
1658 }
1659
1660 if (page) {
1661 len = i_size_read(inode);
1662 if (len > PAGE_SIZE)
1663 len = PAGE_SIZE;
1664 } else {
1665 page = __page_cache_alloc(GFP_NOFS);
1666 if (!page) {
1667 err = -ENOMEM;
1668 goto out;
1669 }
1670 err = __ceph_do_getattr(inode, page,
1671 CEPH_STAT_CAP_INLINE_DATA, true);
1672 if (err < 0) {
1673 /* no inline data */
1674 if (err == -ENODATA)
1675 err = 0;
1676 goto out;
1677 }
1678 len = err;
1679 }
1680
1681 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1682 ceph_vino(inode), 0, &len, 0, 1,
1683 CEPH_OSD_OP_CREATE,
1684 CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1685 NULL, 0, 0, false);
1686 if (IS_ERR(req)) {
1687 err = PTR_ERR(req);
1688 goto out;
1689 }
1690
1691 req->r_mtime = inode->i_mtime;
1692 err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1693 if (!err)
1694 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1695 ceph_osdc_put_request(req);
1696 if (err < 0)
1697 goto out;
1698
1699 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1700 ceph_vino(inode), 0, &len, 1, 3,
1701 CEPH_OSD_OP_WRITE,
1702 CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1703 NULL, ci->i_truncate_seq,
1704 ci->i_truncate_size, false);
1705 if (IS_ERR(req)) {
1706 err = PTR_ERR(req);
1707 goto out;
1708 }
1709
1710 osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1711
1712 {
1713 __le64 xattr_buf = cpu_to_le64(inline_version);
1714 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1715 "inline_version", &xattr_buf,
1716 sizeof(xattr_buf),
1717 CEPH_OSD_CMPXATTR_OP_GT,
1718 CEPH_OSD_CMPXATTR_MODE_U64);
1719 if (err)
1720 goto out_put;
1721 }
1722
1723 {
1724 char xattr_buf[32];
1725 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1726 "%llu", inline_version);
1727 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1728 "inline_version",
1729 xattr_buf, xattr_len, 0, 0);
1730 if (err)
1731 goto out_put;
1732 }
1733
1734 req->r_mtime = inode->i_mtime;
1735 err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1736 if (!err)
1737 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1738out_put:
1739 ceph_osdc_put_request(req);
1740 if (err == -ECANCELED)
1741 err = 0;
1742out:
1743 if (page && page != locked_page) {
1744 if (from_pagecache) {
1745 unlock_page(page);
1746 put_page(page);
1747 } else
1748 __free_pages(page, 0);
1749 }
1750
1751 dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1752 inode, ceph_vinop(inode), inline_version, err);
1753 return err;
1754}
1755
1756static const struct vm_operations_struct ceph_vmops = {
1757 .fault = ceph_filemap_fault,
1758 .page_mkwrite = ceph_page_mkwrite,
1759};
1760
1761int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1762{
1763 struct address_space *mapping = file->f_mapping;
1764
1765 if (!mapping->a_ops->readpage)
1766 return -ENOEXEC;
1767 file_accessed(file);
1768 vma->vm_ops = &ceph_vmops;
1769 return 0;
1770}
1771
1772enum {
1773 POOL_READ = 1,
1774 POOL_WRITE = 2,
1775};
1776
1777static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1778 s64 pool, struct ceph_string *pool_ns)
1779{
1780 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1781 struct ceph_mds_client *mdsc = fsc->mdsc;
1782 struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1783 struct rb_node **p, *parent;
1784 struct ceph_pool_perm *perm;
1785 struct page **pages;
1786 size_t pool_ns_len;
1787 int err = 0, err2 = 0, have = 0;
1788
1789 down_read(&mdsc->pool_perm_rwsem);
1790 p = &mdsc->pool_perm_tree.rb_node;
1791 while (*p) {
1792 perm = rb_entry(*p, struct ceph_pool_perm, node);
1793 if (pool < perm->pool)
1794 p = &(*p)->rb_left;
1795 else if (pool > perm->pool)
1796 p = &(*p)->rb_right;
1797 else {
1798 int ret = ceph_compare_string(pool_ns,
1799 perm->pool_ns,
1800 perm->pool_ns_len);
1801 if (ret < 0)
1802 p = &(*p)->rb_left;
1803 else if (ret > 0)
1804 p = &(*p)->rb_right;
1805 else {
1806 have = perm->perm;
1807 break;
1808 }
1809 }
1810 }
1811 up_read(&mdsc->pool_perm_rwsem);
1812 if (*p)
1813 goto out;
1814
1815 if (pool_ns)
1816 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1817 pool, (int)pool_ns->len, pool_ns->str);
1818 else
1819 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1820
1821 down_write(&mdsc->pool_perm_rwsem);
1822 p = &mdsc->pool_perm_tree.rb_node;
1823 parent = NULL;
1824 while (*p) {
1825 parent = *p;
1826 perm = rb_entry(parent, struct ceph_pool_perm, node);
1827 if (pool < perm->pool)
1828 p = &(*p)->rb_left;
1829 else if (pool > perm->pool)
1830 p = &(*p)->rb_right;
1831 else {
1832 int ret = ceph_compare_string(pool_ns,
1833 perm->pool_ns,
1834 perm->pool_ns_len);
1835 if (ret < 0)
1836 p = &(*p)->rb_left;
1837 else if (ret > 0)
1838 p = &(*p)->rb_right;
1839 else {
1840 have = perm->perm;
1841 break;
1842 }
1843 }
1844 }
1845 if (*p) {
1846 up_write(&mdsc->pool_perm_rwsem);
1847 goto out;
1848 }
1849
1850 rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1851 1, false, GFP_NOFS);
1852 if (!rd_req) {
1853 err = -ENOMEM;
1854 goto out_unlock;
1855 }
1856
1857 rd_req->r_flags = CEPH_OSD_FLAG_READ;
1858 osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1859 rd_req->r_base_oloc.pool = pool;
1860 if (pool_ns)
1861 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1862 ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1863
1864 err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1865 if (err)
1866 goto out_unlock;
1867
1868 wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1869 1, false, GFP_NOFS);
1870 if (!wr_req) {
1871 err = -ENOMEM;
1872 goto out_unlock;
1873 }
1874
1875 wr_req->r_flags = CEPH_OSD_FLAG_WRITE | CEPH_OSD_FLAG_ACK;
1876 osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1877 ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1878 ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1879
1880 err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1881 if (err)
1882 goto out_unlock;
1883
1884 /* one page should be large enough for STAT data */
1885 pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1886 if (IS_ERR(pages)) {
1887 err = PTR_ERR(pages);
1888 goto out_unlock;
1889 }
1890
1891 osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1892 0, false, true);
1893 err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1894
1895 wr_req->r_mtime = ci->vfs_inode.i_mtime;
1896 err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1897
1898 if (!err)
1899 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1900 if (!err2)
1901 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1902
1903 if (err >= 0 || err == -ENOENT)
1904 have |= POOL_READ;
1905 else if (err != -EPERM)
1906 goto out_unlock;
1907
1908 if (err2 == 0 || err2 == -EEXIST)
1909 have |= POOL_WRITE;
1910 else if (err2 != -EPERM) {
1911 err = err2;
1912 goto out_unlock;
1913 }
1914
1915 pool_ns_len = pool_ns ? pool_ns->len : 0;
1916 perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1917 if (!perm) {
1918 err = -ENOMEM;
1919 goto out_unlock;
1920 }
1921
1922 perm->pool = pool;
1923 perm->perm = have;
1924 perm->pool_ns_len = pool_ns_len;
1925 if (pool_ns_len > 0)
1926 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1927 perm->pool_ns[pool_ns_len] = 0;
1928
1929 rb_link_node(&perm->node, parent, p);
1930 rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1931 err = 0;
1932out_unlock:
1933 up_write(&mdsc->pool_perm_rwsem);
1934
1935 ceph_osdc_put_request(rd_req);
1936 ceph_osdc_put_request(wr_req);
1937out:
1938 if (!err)
1939 err = have;
1940 if (pool_ns)
1941 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1942 pool, (int)pool_ns->len, pool_ns->str, err);
1943 else
1944 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1945 return err;
1946}
1947
1948int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1949{
1950 s64 pool;
1951 struct ceph_string *pool_ns;
1952 int ret, flags;
1953
1954 if (ci->i_vino.snap != CEPH_NOSNAP) {
1955 /*
1956 * Pool permission check needs to write to the first object.
1957 * But for snapshot, head of the first object may have alread
1958 * been deleted. Skip check to avoid creating orphan object.
1959 */
1960 return 0;
1961 }
1962
1963 if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
1964 NOPOOLPERM))
1965 return 0;
1966
1967 spin_lock(&ci->i_ceph_lock);
1968 flags = ci->i_ceph_flags;
1969 pool = ci->i_layout.pool_id;
1970 spin_unlock(&ci->i_ceph_lock);
1971check:
1972 if (flags & CEPH_I_POOL_PERM) {
1973 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
1974 dout("ceph_pool_perm_check pool %lld no read perm\n",
1975 pool);
1976 return -EPERM;
1977 }
1978 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
1979 dout("ceph_pool_perm_check pool %lld no write perm\n",
1980 pool);
1981 return -EPERM;
1982 }
1983 return 0;
1984 }
1985
1986 pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
1987 ret = __ceph_pool_perm_get(ci, pool, pool_ns);
1988 ceph_put_string(pool_ns);
1989 if (ret < 0)
1990 return ret;
1991
1992 flags = CEPH_I_POOL_PERM;
1993 if (ret & POOL_READ)
1994 flags |= CEPH_I_POOL_RD;
1995 if (ret & POOL_WRITE)
1996 flags |= CEPH_I_POOL_WR;
1997
1998 spin_lock(&ci->i_ceph_lock);
1999 if (pool == ci->i_layout.pool_id &&
2000 pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2001 ci->i_ceph_flags |= flags;
2002 } else {
2003 pool = ci->i_layout.pool_id;
2004 flags = ci->i_ceph_flags;
2005 }
2006 spin_unlock(&ci->i_ceph_lock);
2007 goto check;
2008}
2009
2010void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2011{
2012 struct ceph_pool_perm *perm;
2013 struct rb_node *n;
2014
2015 while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2016 n = rb_first(&mdsc->pool_perm_tree);
2017 perm = rb_entry(n, struct ceph_pool_perm, node);
2018 rb_erase(n, &mdsc->pool_perm_tree);
2019 kfree(perm);
2020 }
2021}