ceph: fix file offset wrapping at 4GB on 32-bit archs
[linux-2.6.git] / fs / ceph / addr.c
1 #include "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 "osd_client.h"
14
15 /*
16  * Ceph address space ops.
17  *
18  * There are a few funny things going on here.
19  *
20  * The page->private field is used to reference a struct
21  * ceph_snap_context for _every_ dirty page.  This indicates which
22  * snapshot the page was logically dirtied in, and thus which snap
23  * context needs to be associated with the osd write during writeback.
24  *
25  * Similarly, struct ceph_inode_info maintains a set of counters to
26  * count dirty pages on the inode.  In the absense of snapshots,
27  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
28  *
29  * When a snapshot is taken (that is, when the client receives
30  * notification that a snapshot was taken), each inode with caps and
31  * with dirty pages (dirty pages implies there is a cap) gets a new
32  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
33  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
34  * moved to capsnap->dirty. (Unless a sync write is currently in
35  * progress.  In that case, the capsnap is said to be "pending", new
36  * writes cannot start, and the capsnap isn't "finalized" until the
37  * write completes (or fails) and a final size/mtime for the inode for
38  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
39  *
40  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
41  * we look for the first capsnap in i_cap_snaps and write out pages in
42  * that snap context _only_.  Then we move on to the next capsnap,
43  * eventually reaching the "live" or "head" context (i.e., pages that
44  * are not yet snapped) and are writing the most recently dirtied
45  * pages.
46  *
47  * Invalidate and so forth must take care to ensure the dirty page
48  * accounting is preserved.
49  */
50
51 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
52 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
53         (CONGESTION_ON_THRESH(congestion_kb) -                          \
54          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
55
56
57
58 /*
59  * Dirty a page.  Optimistically adjust accounting, on the assumption
60  * that we won't race with invalidate.  If we do, readjust.
61  */
62 static int ceph_set_page_dirty(struct page *page)
63 {
64         struct address_space *mapping = page->mapping;
65         struct inode *inode;
66         struct ceph_inode_info *ci;
67         int undo = 0;
68         struct ceph_snap_context *snapc;
69
70         if (unlikely(!mapping))
71                 return !TestSetPageDirty(page);
72
73         if (TestSetPageDirty(page)) {
74                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
75                      mapping->host, page, page->index);
76                 return 0;
77         }
78
79         inode = mapping->host;
80         ci = ceph_inode(inode);
81
82         /*
83          * Note that we're grabbing a snapc ref here without holding
84          * any locks!
85          */
86         snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
87
88         /* dirty the head */
89         spin_lock(&inode->i_lock);
90         if (ci->i_head_snapc == NULL)
91                 ci->i_head_snapc = ceph_get_snap_context(snapc);
92         ++ci->i_wrbuffer_ref_head;
93         if (ci->i_wrbuffer_ref == 0)
94                 igrab(inode);
95         ++ci->i_wrbuffer_ref;
96         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
97              "snapc %p seq %lld (%d snaps)\n",
98              mapping->host, page, page->index,
99              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
100              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
101              snapc, snapc->seq, snapc->num_snaps);
102         spin_unlock(&inode->i_lock);
103
104         /* now adjust page */
105         spin_lock_irq(&mapping->tree_lock);
106         if (page->mapping) {    /* Race with truncate? */
107                 WARN_ON_ONCE(!PageUptodate(page));
108                 account_page_dirtied(page, page->mapping);
109                 radix_tree_tag_set(&mapping->page_tree,
110                                 page_index(page), PAGECACHE_TAG_DIRTY);
111
112                 /*
113                  * Reference snap context in page->private.  Also set
114                  * PagePrivate so that we get invalidatepage callback.
115                  */
116                 page->private = (unsigned long)snapc;
117                 SetPagePrivate(page);
118         } else {
119                 dout("ANON set_page_dirty %p (raced truncate?)\n", page);
120                 undo = 1;
121         }
122
123         spin_unlock_irq(&mapping->tree_lock);
124
125         if (undo)
126                 /* whoops, we failed to dirty the page */
127                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
128
129         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
130
131         BUG_ON(!PageDirty(page));
132         return 1;
133 }
134
135 /*
136  * If we are truncating the full page (i.e. offset == 0), adjust the
137  * dirty page counters appropriately.  Only called if there is private
138  * data on the page.
139  */
140 static void ceph_invalidatepage(struct page *page, unsigned long offset)
141 {
142         struct inode *inode;
143         struct ceph_inode_info *ci;
144         struct ceph_snap_context *snapc = (void *)page->private;
145
146         BUG_ON(!PageLocked(page));
147         BUG_ON(!page->private);
148         BUG_ON(!PagePrivate(page));
149         BUG_ON(!page->mapping);
150
151         inode = page->mapping->host;
152
153         /*
154          * We can get non-dirty pages here due to races between
155          * set_page_dirty and truncate_complete_page; just spit out a
156          * warning, in case we end up with accounting problems later.
157          */
158         if (!PageDirty(page))
159                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
160
161         if (offset == 0)
162                 ClearPageChecked(page);
163
164         ci = ceph_inode(inode);
165         if (offset == 0) {
166                 dout("%p invalidatepage %p idx %lu full dirty page %lu\n",
167                      inode, page, page->index, offset);
168                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
169                 ceph_put_snap_context(snapc);
170                 page->private = 0;
171                 ClearPagePrivate(page);
172         } else {
173                 dout("%p invalidatepage %p idx %lu partial dirty page\n",
174                      inode, page, page->index);
175         }
176 }
177
178 /* just a sanity check */
179 static int ceph_releasepage(struct page *page, gfp_t g)
180 {
181         struct inode *inode = page->mapping ? page->mapping->host : NULL;
182         dout("%p releasepage %p idx %lu\n", inode, page, page->index);
183         WARN_ON(PageDirty(page));
184         WARN_ON(page->private);
185         WARN_ON(PagePrivate(page));
186         return 0;
187 }
188
189 /*
190  * read a single page, without unlocking it.
191  */
192 static int readpage_nounlock(struct file *filp, struct page *page)
193 {
194         struct inode *inode = filp->f_dentry->d_inode;
195         struct ceph_inode_info *ci = ceph_inode(inode);
196         struct ceph_osd_client *osdc = &ceph_inode_to_client(inode)->osdc;
197         int err = 0;
198         u64 len = PAGE_CACHE_SIZE;
199
200         dout("readpage inode %p file %p page %p index %lu\n",
201              inode, filp, page, page->index);
202         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
203                                   page->index << PAGE_CACHE_SHIFT, &len,
204                                   ci->i_truncate_seq, ci->i_truncate_size,
205                                   &page, 1);
206         if (err == -ENOENT)
207                 err = 0;
208         if (err < 0) {
209                 SetPageError(page);
210                 goto out;
211         } else if (err < PAGE_CACHE_SIZE) {
212                 /* zero fill remainder of page */
213                 zero_user_segment(page, err, PAGE_CACHE_SIZE);
214         }
215         SetPageUptodate(page);
216
217 out:
218         return err < 0 ? err : 0;
219 }
220
221 static int ceph_readpage(struct file *filp, struct page *page)
222 {
223         int r = readpage_nounlock(filp, page);
224         unlock_page(page);
225         return r;
226 }
227
228 /*
229  * Build a vector of contiguous pages from the provided page list.
230  */
231 static struct page **page_vector_from_list(struct list_head *page_list,
232                                            unsigned *nr_pages)
233 {
234         struct page **pages;
235         struct page *page;
236         int next_index, contig_pages = 0;
237
238         /* build page vector */
239         pages = kmalloc(sizeof(*pages) * *nr_pages, GFP_NOFS);
240         if (!pages)
241                 return ERR_PTR(-ENOMEM);
242
243         BUG_ON(list_empty(page_list));
244         next_index = list_entry(page_list->prev, struct page, lru)->index;
245         list_for_each_entry_reverse(page, page_list, lru) {
246                 if (page->index == next_index) {
247                         dout("readpages page %d %p\n", contig_pages, page);
248                         pages[contig_pages] = page;
249                         contig_pages++;
250                         next_index++;
251                 } else {
252                         break;
253                 }
254         }
255         *nr_pages = contig_pages;
256         return pages;
257 }
258
259 /*
260  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
261  * the caller (VM) cleans them up.
262  */
263 static int ceph_readpages(struct file *file, struct address_space *mapping,
264                           struct list_head *page_list, unsigned nr_pages)
265 {
266         struct inode *inode = file->f_dentry->d_inode;
267         struct ceph_inode_info *ci = ceph_inode(inode);
268         struct ceph_osd_client *osdc = &ceph_inode_to_client(inode)->osdc;
269         int rc = 0;
270         struct page **pages;
271         loff_t offset;
272         u64 len;
273
274         dout("readpages %p file %p nr_pages %d\n",
275              inode, file, nr_pages);
276
277         pages = page_vector_from_list(page_list, &nr_pages);
278         if (IS_ERR(pages))
279                 return PTR_ERR(pages);
280
281         /* guess read extent */
282         offset = pages[0]->index << PAGE_CACHE_SHIFT;
283         len = nr_pages << PAGE_CACHE_SHIFT;
284         rc = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
285                                  offset, &len,
286                                  ci->i_truncate_seq, ci->i_truncate_size,
287                                  pages, nr_pages);
288         if (rc == -ENOENT)
289                 rc = 0;
290         if (rc < 0)
291                 goto out;
292
293         for (; !list_empty(page_list) && len > 0;
294              rc -= PAGE_CACHE_SIZE, len -= PAGE_CACHE_SIZE) {
295                 struct page *page =
296                         list_entry(page_list->prev, struct page, lru);
297
298                 list_del(&page->lru);
299
300                 if (rc < (int)PAGE_CACHE_SIZE) {
301                         /* zero (remainder of) page */
302                         int s = rc < 0 ? 0 : rc;
303                         zero_user_segment(page, s, PAGE_CACHE_SIZE);
304                 }
305
306                 if (add_to_page_cache_lru(page, mapping, page->index,
307                                           GFP_NOFS)) {
308                         page_cache_release(page);
309                         dout("readpages %p add_to_page_cache failed %p\n",
310                              inode, page);
311                         continue;
312                 }
313                 dout("readpages %p adding %p idx %lu\n", inode, page,
314                      page->index);
315                 flush_dcache_page(page);
316                 SetPageUptodate(page);
317                 unlock_page(page);
318                 page_cache_release(page);
319         }
320         rc = 0;
321
322 out:
323         kfree(pages);
324         return rc;
325 }
326
327 /*
328  * Get ref for the oldest snapc for an inode with dirty data... that is, the
329  * only snap context we are allowed to write back.
330  */
331 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
332                                                     u64 *snap_size)
333 {
334         struct ceph_inode_info *ci = ceph_inode(inode);
335         struct ceph_snap_context *snapc = NULL;
336         struct ceph_cap_snap *capsnap = NULL;
337
338         spin_lock(&inode->i_lock);
339         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
340                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
341                      capsnap->context, capsnap->dirty_pages);
342                 if (capsnap->dirty_pages) {
343                         snapc = ceph_get_snap_context(capsnap->context);
344                         if (snap_size)
345                                 *snap_size = capsnap->size;
346                         break;
347                 }
348         }
349         if (!snapc && ci->i_wrbuffer_ref_head) {
350                 snapc = ceph_get_snap_context(ci->i_head_snapc);
351                 dout(" head snapc %p has %d dirty pages\n",
352                      snapc, ci->i_wrbuffer_ref_head);
353         }
354         spin_unlock(&inode->i_lock);
355         return snapc;
356 }
357
358 /*
359  * Write a single page, but leave the page locked.
360  *
361  * If we get a write error, set the page error bit, but still adjust the
362  * dirty page accounting (i.e., page is no longer dirty).
363  */
364 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
365 {
366         struct inode *inode;
367         struct ceph_inode_info *ci;
368         struct ceph_client *client;
369         struct ceph_osd_client *osdc;
370         loff_t page_off = page->index << PAGE_CACHE_SHIFT;
371         int len = PAGE_CACHE_SIZE;
372         loff_t i_size;
373         int err = 0;
374         struct ceph_snap_context *snapc, *oldest;
375         u64 snap_size = 0;
376         long writeback_stat;
377
378         dout("writepage %p idx %lu\n", page, page->index);
379
380         if (!page->mapping || !page->mapping->host) {
381                 dout("writepage %p - no mapping\n", page);
382                 return -EFAULT;
383         }
384         inode = page->mapping->host;
385         ci = ceph_inode(inode);
386         client = ceph_inode_to_client(inode);
387         osdc = &client->osdc;
388
389         /* verify this is a writeable snap context */
390         snapc = (void *)page->private;
391         if (snapc == NULL) {
392                 dout("writepage %p page %p not dirty?\n", inode, page);
393                 goto out;
394         }
395         oldest = get_oldest_context(inode, &snap_size);
396         if (snapc->seq > oldest->seq) {
397                 dout("writepage %p page %p snapc %p not writeable - noop\n",
398                      inode, page, (void *)page->private);
399                 /* we should only noop if called by kswapd */
400                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
401                 ceph_put_snap_context(oldest);
402                 goto out;
403         }
404         ceph_put_snap_context(oldest);
405
406         /* is this a partial page at end of file? */
407         if (snap_size)
408                 i_size = snap_size;
409         else
410                 i_size = i_size_read(inode);
411         if (i_size < page_off + len)
412                 len = i_size - page_off;
413
414         dout("writepage %p page %p index %lu on %llu~%u\n",
415              inode, page, page->index, page_off, len);
416
417         writeback_stat = atomic_long_inc_return(&client->writeback_count);
418         if (writeback_stat >
419             CONGESTION_ON_THRESH(client->mount_args->congestion_kb))
420                 set_bdi_congested(&client->backing_dev_info, BLK_RW_ASYNC);
421
422         set_page_writeback(page);
423         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
424                                    &ci->i_layout, snapc,
425                                    page_off, len,
426                                    ci->i_truncate_seq, ci->i_truncate_size,
427                                    &inode->i_mtime,
428                                    &page, 1, 0, 0, true);
429         if (err < 0) {
430                 dout("writepage setting page/mapping error %d %p\n", err, page);
431                 SetPageError(page);
432                 mapping_set_error(&inode->i_data, err);
433                 if (wbc)
434                         wbc->pages_skipped++;
435         } else {
436                 dout("writepage cleaned page %p\n", page);
437                 err = 0;  /* vfs expects us to return 0 */
438         }
439         page->private = 0;
440         ClearPagePrivate(page);
441         end_page_writeback(page);
442         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
443         ceph_put_snap_context(snapc);  /* page's reference */
444 out:
445         return err;
446 }
447
448 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
449 {
450         int err;
451         struct inode *inode = page->mapping->host;
452         BUG_ON(!inode);
453         igrab(inode);
454         err = writepage_nounlock(page, wbc);
455         unlock_page(page);
456         iput(inode);
457         return err;
458 }
459
460
461 /*
462  * lame release_pages helper.  release_pages() isn't exported to
463  * modules.
464  */
465 static void ceph_release_pages(struct page **pages, int num)
466 {
467         struct pagevec pvec;
468         int i;
469
470         pagevec_init(&pvec, 0);
471         for (i = 0; i < num; i++) {
472                 if (pagevec_add(&pvec, pages[i]) == 0)
473                         pagevec_release(&pvec);
474         }
475         pagevec_release(&pvec);
476 }
477
478
479 /*
480  * async writeback completion handler.
481  *
482  * If we get an error, set the mapping error bit, but not the individual
483  * page error bits.
484  */
485 static void writepages_finish(struct ceph_osd_request *req,
486                               struct ceph_msg *msg)
487 {
488         struct inode *inode = req->r_inode;
489         struct ceph_osd_reply_head *replyhead;
490         struct ceph_osd_op *op;
491         struct ceph_inode_info *ci = ceph_inode(inode);
492         unsigned wrote;
493         struct page *page;
494         int i;
495         struct ceph_snap_context *snapc = req->r_snapc;
496         struct address_space *mapping = inode->i_mapping;
497         __s32 rc = -EIO;
498         u64 bytes = 0;
499         struct ceph_client *client = ceph_inode_to_client(inode);
500         long writeback_stat;
501         unsigned issued = ceph_caps_issued(ci);
502
503         /* parse reply */
504         replyhead = msg->front.iov_base;
505         WARN_ON(le32_to_cpu(replyhead->num_ops) == 0);
506         op = (void *)(replyhead + 1);
507         rc = le32_to_cpu(replyhead->result);
508         bytes = le64_to_cpu(op->extent.length);
509
510         if (rc >= 0) {
511                 /*
512                  * Assume we wrote the pages we originally sent.  The
513                  * osd might reply with fewer pages if our writeback
514                  * raced with a truncation and was adjusted at the osd,
515                  * so don't believe the reply.
516                  */
517                 wrote = req->r_num_pages;
518         } else {
519                 wrote = 0;
520                 mapping_set_error(mapping, rc);
521         }
522         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
523              inode, rc, bytes, wrote);
524
525         /* clean all pages */
526         for (i = 0; i < req->r_num_pages; i++) {
527                 page = req->r_pages[i];
528                 BUG_ON(!page);
529                 WARN_ON(!PageUptodate(page));
530
531                 writeback_stat =
532                         atomic_long_dec_return(&client->writeback_count);
533                 if (writeback_stat <
534                     CONGESTION_OFF_THRESH(client->mount_args->congestion_kb))
535                         clear_bdi_congested(&client->backing_dev_info,
536                                             BLK_RW_ASYNC);
537
538                 ceph_put_snap_context((void *)page->private);
539                 page->private = 0;
540                 ClearPagePrivate(page);
541                 dout("unlocking %d %p\n", i, page);
542                 end_page_writeback(page);
543
544                 /*
545                  * We lost the cache cap, need to truncate the page before
546                  * it is unlocked, otherwise we'd truncate it later in the
547                  * page truncation thread, possibly losing some data that
548                  * raced its way in
549                  */
550                 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
551                         generic_error_remove_page(inode->i_mapping, page);
552
553                 unlock_page(page);
554         }
555         dout("%p wrote+cleaned %d pages\n", inode, wrote);
556         ceph_put_wrbuffer_cap_refs(ci, req->r_num_pages, snapc);
557
558         ceph_release_pages(req->r_pages, req->r_num_pages);
559         if (req->r_pages_from_pool)
560                 mempool_free(req->r_pages,
561                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
562         else
563                 kfree(req->r_pages);
564         ceph_osdc_put_request(req);
565 }
566
567 /*
568  * allocate a page vec, either directly, or if necessary, via a the
569  * mempool.  we avoid the mempool if we can because req->r_num_pages
570  * may be less than the maximum write size.
571  */
572 static void alloc_page_vec(struct ceph_client *client,
573                            struct ceph_osd_request *req)
574 {
575         req->r_pages = kmalloc(sizeof(struct page *) * req->r_num_pages,
576                                GFP_NOFS);
577         if (!req->r_pages) {
578                 req->r_pages = mempool_alloc(client->wb_pagevec_pool, GFP_NOFS);
579                 req->r_pages_from_pool = 1;
580                 WARN_ON(!req->r_pages);
581         }
582 }
583
584 /*
585  * initiate async writeback
586  */
587 static int ceph_writepages_start(struct address_space *mapping,
588                                  struct writeback_control *wbc)
589 {
590         struct inode *inode = mapping->host;
591         struct backing_dev_info *bdi = mapping->backing_dev_info;
592         struct ceph_inode_info *ci = ceph_inode(inode);
593         struct ceph_client *client;
594         pgoff_t index, start, end;
595         int range_whole = 0;
596         int should_loop = 1;
597         pgoff_t max_pages = 0, max_pages_ever = 0;
598         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
599         struct pagevec pvec;
600         int done = 0;
601         int rc = 0;
602         unsigned wsize = 1 << inode->i_blkbits;
603         struct ceph_osd_request *req = NULL;
604         int do_sync;
605         u64 snap_size = 0;
606
607         /*
608          * Include a 'sync' in the OSD request if this is a data
609          * integrity write (e.g., O_SYNC write or fsync()), or if our
610          * cap is being revoked.
611          */
612         do_sync = wbc->sync_mode == WB_SYNC_ALL;
613         if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
614                 do_sync = 1;
615         dout("writepages_start %p dosync=%d (mode=%s)\n",
616              inode, do_sync,
617              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
618              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
619
620         client = ceph_inode_to_client(inode);
621         if (client->mount_state == CEPH_MOUNT_SHUTDOWN) {
622                 pr_warning("writepage_start %p on forced umount\n", inode);
623                 return -EIO; /* we're in a forced umount, don't write! */
624         }
625         if (client->mount_args->wsize && client->mount_args->wsize < wsize)
626                 wsize = client->mount_args->wsize;
627         if (wsize < PAGE_CACHE_SIZE)
628                 wsize = PAGE_CACHE_SIZE;
629         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
630
631         pagevec_init(&pvec, 0);
632
633         /* ?? */
634         if (wbc->nonblocking && bdi_write_congested(bdi)) {
635                 dout(" writepages congested\n");
636                 wbc->encountered_congestion = 1;
637                 goto out_final;
638         }
639
640         /* where to start/end? */
641         if (wbc->range_cyclic) {
642                 start = mapping->writeback_index; /* Start from prev offset */
643                 end = -1;
644                 dout(" cyclic, start at %lu\n", start);
645         } else {
646                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
647                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
648                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
649                         range_whole = 1;
650                 should_loop = 0;
651                 dout(" not cyclic, %lu to %lu\n", start, end);
652         }
653         index = start;
654
655 retry:
656         /* find oldest snap context with dirty data */
657         ceph_put_snap_context(snapc);
658         snapc = get_oldest_context(inode, &snap_size);
659         if (!snapc) {
660                 /* hmm, why does writepages get called when there
661                    is no dirty data? */
662                 dout(" no snap context with dirty data?\n");
663                 goto out;
664         }
665         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
666              snapc, snapc->seq, snapc->num_snaps);
667         if (last_snapc && snapc != last_snapc) {
668                 /* if we switched to a newer snapc, restart our scan at the
669                  * start of the original file range. */
670                 dout("  snapc differs from last pass, restarting at %lu\n",
671                      index);
672                 index = start;
673         }
674         last_snapc = snapc;
675
676         while (!done && index <= end) {
677                 unsigned i;
678                 int first;
679                 pgoff_t next;
680                 int pvec_pages, locked_pages;
681                 struct page *page;
682                 int want;
683                 u64 offset, len;
684                 struct ceph_osd_request_head *reqhead;
685                 struct ceph_osd_op *op;
686                 long writeback_stat;
687
688                 next = 0;
689                 locked_pages = 0;
690                 max_pages = max_pages_ever;
691
692 get_more_pages:
693                 first = -1;
694                 want = min(end - index,
695                            min((pgoff_t)PAGEVEC_SIZE,
696                                max_pages - (pgoff_t)locked_pages) - 1)
697                         + 1;
698                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
699                                                 PAGECACHE_TAG_DIRTY,
700                                                 want);
701                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
702                 if (!pvec_pages && !locked_pages)
703                         break;
704                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
705                         page = pvec.pages[i];
706                         dout("? %p idx %lu\n", page, page->index);
707                         if (locked_pages == 0)
708                                 lock_page(page);  /* first page */
709                         else if (!trylock_page(page))
710                                 break;
711
712                         /* only dirty pages, or our accounting breaks */
713                         if (unlikely(!PageDirty(page)) ||
714                             unlikely(page->mapping != mapping)) {
715                                 dout("!dirty or !mapping %p\n", page);
716                                 unlock_page(page);
717                                 break;
718                         }
719                         if (!wbc->range_cyclic && page->index > end) {
720                                 dout("end of range %p\n", page);
721                                 done = 1;
722                                 unlock_page(page);
723                                 break;
724                         }
725                         if (next && (page->index != next)) {
726                                 dout("not consecutive %p\n", page);
727                                 unlock_page(page);
728                                 break;
729                         }
730                         if (wbc->sync_mode != WB_SYNC_NONE) {
731                                 dout("waiting on writeback %p\n", page);
732                                 wait_on_page_writeback(page);
733                         }
734                         if ((snap_size && page_offset(page) > snap_size) ||
735                             (!snap_size &&
736                              page_offset(page) > i_size_read(inode))) {
737                                 dout("%p page eof %llu\n", page, snap_size ?
738                                      snap_size : i_size_read(inode));
739                                 done = 1;
740                                 unlock_page(page);
741                                 break;
742                         }
743                         if (PageWriteback(page)) {
744                                 dout("%p under writeback\n", page);
745                                 unlock_page(page);
746                                 break;
747                         }
748
749                         /* only if matching snap context */
750                         pgsnapc = (void *)page->private;
751                         if (pgsnapc->seq > snapc->seq) {
752                                 dout("page snapc %p %lld > oldest %p %lld\n",
753                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
754                                 unlock_page(page);
755                                 if (!locked_pages)
756                                         continue; /* keep looking for snap */
757                                 break;
758                         }
759
760                         if (!clear_page_dirty_for_io(page)) {
761                                 dout("%p !clear_page_dirty_for_io\n", page);
762                                 unlock_page(page);
763                                 break;
764                         }
765
766                         /* ok */
767                         if (locked_pages == 0) {
768                                 /* prepare async write request */
769                                 offset = (unsigned long long)page->index
770                                         << PAGE_CACHE_SHIFT;
771                                 len = wsize;
772                                 req = ceph_osdc_new_request(&client->osdc,
773                                             &ci->i_layout,
774                                             ceph_vino(inode),
775                                             offset, &len,
776                                             CEPH_OSD_OP_WRITE,
777                                             CEPH_OSD_FLAG_WRITE |
778                                                     CEPH_OSD_FLAG_ONDISK,
779                                             snapc, do_sync,
780                                             ci->i_truncate_seq,
781                                             ci->i_truncate_size,
782                                             &inode->i_mtime, true, 1);
783                                 max_pages = req->r_num_pages;
784
785                                 alloc_page_vec(client, req);
786                                 req->r_callback = writepages_finish;
787                                 req->r_inode = inode;
788                         }
789
790                         /* note position of first page in pvec */
791                         if (first < 0)
792                                 first = i;
793                         dout("%p will write page %p idx %lu\n",
794                              inode, page, page->index);
795
796                         writeback_stat =
797                                atomic_long_inc_return(&client->writeback_count);
798                         if (writeback_stat > CONGESTION_ON_THRESH(
799                                     client->mount_args->congestion_kb)) {
800                                 set_bdi_congested(&client->backing_dev_info,
801                                                   BLK_RW_ASYNC);
802                         }
803
804                         set_page_writeback(page);
805                         req->r_pages[locked_pages] = page;
806                         locked_pages++;
807                         next = page->index + 1;
808                 }
809
810                 /* did we get anything? */
811                 if (!locked_pages)
812                         goto release_pvec_pages;
813                 if (i) {
814                         int j;
815                         BUG_ON(!locked_pages || first < 0);
816
817                         if (pvec_pages && i == pvec_pages &&
818                             locked_pages < max_pages) {
819                                 dout("reached end pvec, trying for more\n");
820                                 pagevec_reinit(&pvec);
821                                 goto get_more_pages;
822                         }
823
824                         /* shift unused pages over in the pvec...  we
825                          * will need to release them below. */
826                         for (j = i; j < pvec_pages; j++) {
827                                 dout(" pvec leftover page %p\n",
828                                      pvec.pages[j]);
829                                 pvec.pages[j-i+first] = pvec.pages[j];
830                         }
831                         pvec.nr -= i-first;
832                 }
833
834                 /* submit the write */
835                 offset = req->r_pages[0]->index << PAGE_CACHE_SHIFT;
836                 len = min((snap_size ? snap_size : i_size_read(inode)) - offset,
837                           (u64)locked_pages << PAGE_CACHE_SHIFT);
838                 dout("writepages got %d pages at %llu~%llu\n",
839                      locked_pages, offset, len);
840
841                 /* revise final length, page count */
842                 req->r_num_pages = locked_pages;
843                 reqhead = req->r_request->front.iov_base;
844                 op = (void *)(reqhead + 1);
845                 op->extent.length = cpu_to_le64(len);
846                 op->payload_len = cpu_to_le32(len);
847                 req->r_request->hdr.data_len = cpu_to_le32(len);
848
849                 ceph_osdc_start_request(&client->osdc, req, true);
850                 req = NULL;
851
852                 /* continue? */
853                 index = next;
854                 wbc->nr_to_write -= locked_pages;
855                 if (wbc->nr_to_write <= 0)
856                         done = 1;
857
858 release_pvec_pages:
859                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
860                      pvec.nr ? pvec.pages[0] : NULL);
861                 pagevec_release(&pvec);
862
863                 if (locked_pages && !done)
864                         goto retry;
865         }
866
867         if (should_loop && !done) {
868                 /* more to do; loop back to beginning of file */
869                 dout("writepages looping back to beginning of file\n");
870                 should_loop = 0;
871                 index = 0;
872                 goto retry;
873         }
874
875         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
876                 mapping->writeback_index = index;
877
878 out:
879         if (req)
880                 ceph_osdc_put_request(req);
881         if (rc > 0)
882                 rc = 0;  /* vfs expects us to return 0 */
883         ceph_put_snap_context(snapc);
884         dout("writepages done, rc = %d\n", rc);
885 out_final:
886         return rc;
887 }
888
889
890
891 /*
892  * See if a given @snapc is either writeable, or already written.
893  */
894 static 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  */
912 static 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
926 retry_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
1010 fail:
1011         up_read(&mdsc->snap_rwsem);
1012 fail_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  */
1021 static 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  */
1051 static 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_client *client = ceph_inode_to_client(inode);
1057         struct ceph_mds_client *mdsc = &client->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  */
1093 static 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
1101 const 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  */
1122 static 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         }
1159 out:
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
1166 static struct vm_operations_struct ceph_vmops = {
1167         .fault          = filemap_fault,
1168         .page_mkwrite   = ceph_page_mkwrite,
1169 };
1170
1171 int 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 }