ceph: use ceph_sb_to_client instead of ceph_client
[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_wrbuffer_ref_head == 0)
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
109                 if (mapping_cap_account_dirty(mapping)) {
110                         __inc_zone_page_state(page, NR_FILE_DIRTY);
111                         __inc_bdi_stat(mapping->backing_dev_info,
112                                         BDI_RECLAIMABLE);
113                         task_io_account_write(PAGE_CACHE_SIZE);
114                 }
115                 radix_tree_tag_set(&mapping->page_tree,
116                                 page_index(page), PAGECACHE_TAG_DIRTY);
117
118                 /*
119                  * Reference snap context in page->private.  Also set
120                  * PagePrivate so that we get invalidatepage callback.
121                  */
122                 page->private = (unsigned long)snapc;
123                 SetPagePrivate(page);
124         } else {
125                 dout("ANON set_page_dirty %p (raced truncate?)\n", page);
126                 undo = 1;
127         }
128
129         spin_unlock_irq(&mapping->tree_lock);
130
131         if (undo)
132                 /* whoops, we failed to dirty the page */
133                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
134
135         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
136
137         BUG_ON(!PageDirty(page));
138         return 1;
139 }
140
141 /*
142  * If we are truncating the full page (i.e. offset == 0), adjust the
143  * dirty page counters appropriately.  Only called if there is private
144  * data on the page.
145  */
146 static void ceph_invalidatepage(struct page *page, unsigned long offset)
147 {
148         struct inode *inode;
149         struct ceph_inode_info *ci;
150         struct ceph_snap_context *snapc = (void *)page->private;
151
152         BUG_ON(!PageLocked(page));
153         BUG_ON(!page->private);
154         BUG_ON(!PagePrivate(page));
155         BUG_ON(!page->mapping);
156
157         inode = page->mapping->host;
158
159         /*
160          * We can get non-dirty pages here due to races between
161          * set_page_dirty and truncate_complete_page; just spit out a
162          * warning, in case we end up with accounting problems later.
163          */
164         if (!PageDirty(page))
165                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
166
167         if (offset == 0)
168                 ClearPageChecked(page);
169
170         ci = ceph_inode(inode);
171         if (offset == 0) {
172                 dout("%p invalidatepage %p idx %lu full dirty page %lu\n",
173                      inode, page, page->index, offset);
174                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
175                 ceph_put_snap_context(snapc);
176                 page->private = 0;
177                 ClearPagePrivate(page);
178         } else {
179                 dout("%p invalidatepage %p idx %lu partial dirty page\n",
180                      inode, page, page->index);
181         }
182 }
183
184 /* just a sanity check */
185 static int ceph_releasepage(struct page *page, gfp_t g)
186 {
187         struct inode *inode = page->mapping ? page->mapping->host : NULL;
188         dout("%p releasepage %p idx %lu\n", inode, page, page->index);
189         WARN_ON(PageDirty(page));
190         WARN_ON(page->private);
191         WARN_ON(PagePrivate(page));
192         return 0;
193 }
194
195 /*
196  * read a single page, without unlocking it.
197  */
198 static int readpage_nounlock(struct file *filp, struct page *page)
199 {
200         struct inode *inode = filp->f_dentry->d_inode;
201         struct ceph_inode_info *ci = ceph_inode(inode);
202         struct ceph_osd_client *osdc = &ceph_inode_to_client(inode)->osdc;
203         int err = 0;
204         u64 len = PAGE_CACHE_SIZE;
205
206         dout("readpage inode %p file %p page %p index %lu\n",
207              inode, filp, page, page->index);
208         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
209                                   page->index << PAGE_CACHE_SHIFT, &len,
210                                   ci->i_truncate_seq, ci->i_truncate_size,
211                                   &page, 1);
212         if (err == -ENOENT)
213                 err = 0;
214         if (err < 0) {
215                 SetPageError(page);
216                 goto out;
217         } else if (err < PAGE_CACHE_SIZE) {
218                 /* zero fill remainder of page */
219                 zero_user_segment(page, err, PAGE_CACHE_SIZE);
220         }
221         SetPageUptodate(page);
222
223 out:
224         return err < 0 ? err : 0;
225 }
226
227 static int ceph_readpage(struct file *filp, struct page *page)
228 {
229         int r = readpage_nounlock(filp, page);
230         unlock_page(page);
231         return r;
232 }
233
234 /*
235  * Build a vector of contiguous pages from the provided page list.
236  */
237 static struct page **page_vector_from_list(struct list_head *page_list,
238                                            unsigned *nr_pages)
239 {
240         struct page **pages;
241         struct page *page;
242         int next_index, contig_pages = 0;
243
244         /* build page vector */
245         pages = kmalloc(sizeof(*pages) * *nr_pages, GFP_NOFS);
246         if (!pages)
247                 return ERR_PTR(-ENOMEM);
248
249         BUG_ON(list_empty(page_list));
250         next_index = list_entry(page_list->prev, struct page, lru)->index;
251         list_for_each_entry_reverse(page, page_list, lru) {
252                 if (page->index == next_index) {
253                         dout("readpages page %d %p\n", contig_pages, page);
254                         pages[contig_pages] = page;
255                         contig_pages++;
256                         next_index++;
257                 } else {
258                         break;
259                 }
260         }
261         *nr_pages = contig_pages;
262         return pages;
263 }
264
265 /*
266  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
267  * the caller (VM) cleans them up.
268  */
269 static int ceph_readpages(struct file *file, struct address_space *mapping,
270                           struct list_head *page_list, unsigned nr_pages)
271 {
272         struct inode *inode = file->f_dentry->d_inode;
273         struct ceph_inode_info *ci = ceph_inode(inode);
274         struct ceph_osd_client *osdc = &ceph_inode_to_client(inode)->osdc;
275         int rc = 0;
276         struct page **pages;
277         loff_t offset;
278         u64 len;
279
280         dout("readpages %p file %p nr_pages %d\n",
281              inode, file, nr_pages);
282
283         pages = page_vector_from_list(page_list, &nr_pages);
284         if (IS_ERR(pages))
285                 return PTR_ERR(pages);
286
287         /* guess read extent */
288         offset = pages[0]->index << PAGE_CACHE_SHIFT;
289         len = nr_pages << PAGE_CACHE_SHIFT;
290         rc = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
291                                  offset, &len,
292                                  ci->i_truncate_seq, ci->i_truncate_size,
293                                  pages, nr_pages);
294         if (rc == -ENOENT)
295                 rc = 0;
296         if (rc < 0)
297                 goto out;
298
299         for (; !list_empty(page_list) && len > 0;
300              rc -= PAGE_CACHE_SIZE, len -= PAGE_CACHE_SIZE) {
301                 struct page *page =
302                         list_entry(page_list->prev, struct page, lru);
303
304                 list_del(&page->lru);
305
306                 if (rc < (int)PAGE_CACHE_SIZE) {
307                         /* zero (remainder of) page */
308                         int s = rc < 0 ? 0 : rc;
309                         zero_user_segment(page, s, PAGE_CACHE_SIZE);
310                 }
311
312                 if (add_to_page_cache_lru(page, mapping, page->index, GFP_NOFS)) {
313                         page_cache_release(page);
314                         dout("readpages %p add_to_page_cache failed %p\n",
315                              inode, page);
316                         continue;
317                 }
318                 dout("readpages %p adding %p idx %lu\n", inode, page,
319                      page->index);
320                 flush_dcache_page(page);
321                 SetPageUptodate(page);
322                 unlock_page(page);
323                 page_cache_release(page);
324         }
325         rc = 0;
326
327 out:
328         kfree(pages);
329         return rc;
330 }
331
332 /*
333  * Get ref for the oldest snapc for an inode with dirty data... that is, the
334  * only snap context we are allowed to write back.
335  */
336 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
337                                                     u64 *snap_size)
338 {
339         struct ceph_inode_info *ci = ceph_inode(inode);
340         struct ceph_snap_context *snapc = NULL;
341         struct ceph_cap_snap *capsnap = NULL;
342
343         spin_lock(&inode->i_lock);
344         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
345                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
346                      capsnap->context, capsnap->dirty_pages);
347                 if (capsnap->dirty_pages) {
348                         snapc = ceph_get_snap_context(capsnap->context);
349                         if (snap_size)
350                                 *snap_size = capsnap->size;
351                         break;
352                 }
353         }
354         if (!snapc && ci->i_head_snapc) {
355                 snapc = ceph_get_snap_context(ci->i_head_snapc);
356                 dout(" head snapc %p has %d dirty pages\n",
357                      snapc, ci->i_wrbuffer_ref_head);
358         }
359         spin_unlock(&inode->i_lock);
360         return snapc;
361 }
362
363 /*
364  * Write a single page, but leave the page locked.
365  *
366  * If we get a write error, set the page error bit, but still adjust the
367  * dirty page accounting (i.e., page is no longer dirty).
368  */
369 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
370 {
371         struct inode *inode;
372         struct ceph_inode_info *ci;
373         struct ceph_client *client;
374         struct ceph_osd_client *osdc;
375         loff_t page_off = page->index << PAGE_CACHE_SHIFT;
376         int len = PAGE_CACHE_SIZE;
377         loff_t i_size;
378         int err = 0;
379         struct ceph_snap_context *snapc, *oldest;
380         u64 snap_size = 0;
381         long writeback_stat;
382
383         dout("writepage %p idx %lu\n", page, page->index);
384
385         if (!page->mapping || !page->mapping->host) {
386                 dout("writepage %p - no mapping\n", page);
387                 return -EFAULT;
388         }
389         inode = page->mapping->host;
390         ci = ceph_inode(inode);
391         client = ceph_inode_to_client(inode);
392         osdc = &client->osdc;
393
394         /* verify this is a writeable snap context */
395         snapc = (void *)page->private;
396         if (snapc == NULL) {
397                 dout("writepage %p page %p not dirty?\n", inode, page);
398                 goto out;
399         }
400         oldest = get_oldest_context(inode, &snap_size);
401         if (snapc->seq > oldest->seq) {
402                 dout("writepage %p page %p snapc %p not writeable - noop\n",
403                      inode, page, (void *)page->private);
404                 /* we should only noop if called by kswapd */
405                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
406                 ceph_put_snap_context(oldest);
407                 goto out;
408         }
409         ceph_put_snap_context(oldest);
410
411         /* is this a partial page at end of file? */
412         if (snap_size)
413                 i_size = snap_size;
414         else
415                 i_size = i_size_read(inode);
416         if (i_size < page_off + len)
417                 len = i_size - page_off;
418
419         dout("writepage %p page %p index %lu on %llu~%u\n",
420              inode, page, page->index, page_off, len);
421
422         writeback_stat = atomic_long_inc_return(&client->writeback_count);
423         if (writeback_stat >
424             CONGESTION_ON_THRESH(client->mount_args->congestion_kb))
425                 set_bdi_congested(&client->backing_dev_info, BLK_RW_ASYNC);
426
427         set_page_writeback(page);
428         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
429                                    &ci->i_layout, snapc,
430                                    page_off, len,
431                                    ci->i_truncate_seq, ci->i_truncate_size,
432                                    &inode->i_mtime,
433                                    &page, 1, 0, 0, true);
434         if (err < 0) {
435                 dout("writepage setting page/mapping error %d %p\n", err, page);
436                 SetPageError(page);
437                 mapping_set_error(&inode->i_data, err);
438                 if (wbc)
439                         wbc->pages_skipped++;
440         } else {
441                 dout("writepage cleaned page %p\n", page);
442                 err = 0;  /* vfs expects us to return 0 */
443         }
444         page->private = 0;
445         ClearPagePrivate(page);
446         end_page_writeback(page);
447         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
448         ceph_put_snap_context(snapc);  /* page's reference */
449 out:
450         return err;
451 }
452
453 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
454 {
455         int err;
456         struct inode *inode = page->mapping->host;
457         BUG_ON(!inode);
458         igrab(inode);
459         err = writepage_nounlock(page, wbc);
460         unlock_page(page);
461         iput(inode);
462         return err;
463 }
464
465
466 /*
467  * lame release_pages helper.  release_pages() isn't exported to
468  * modules.
469  */
470 static void ceph_release_pages(struct page **pages, int num)
471 {
472         struct pagevec pvec;
473         int i;
474
475         pagevec_init(&pvec, 0);
476         for (i = 0; i < num; i++) {
477                 if (pagevec_add(&pvec, pages[i]) == 0)
478                         pagevec_release(&pvec);
479         }
480         pagevec_release(&pvec);
481 }
482
483
484 /*
485  * async writeback completion handler.
486  *
487  * If we get an error, set the mapping error bit, but not the individual
488  * page error bits.
489  */
490 static void writepages_finish(struct ceph_osd_request *req,
491                               struct ceph_msg *msg)
492 {
493         struct inode *inode = req->r_inode;
494         struct ceph_osd_reply_head *replyhead;
495         struct ceph_osd_op *op;
496         struct ceph_inode_info *ci = ceph_inode(inode);
497         unsigned wrote;
498         struct page *page;
499         int i;
500         struct ceph_snap_context *snapc = req->r_snapc;
501         struct address_space *mapping = inode->i_mapping;
502         __s32 rc = -EIO;
503         u64 bytes = 0;
504         struct ceph_client *client = ceph_inode_to_client(inode);
505         long writeback_stat;
506         unsigned issued = ceph_caps_issued(ci);
507
508         /* parse reply */
509         replyhead = msg->front.iov_base;
510         WARN_ON(le32_to_cpu(replyhead->num_ops) == 0);
511         op = (void *)(replyhead + 1);
512         rc = le32_to_cpu(replyhead->result);
513         bytes = le64_to_cpu(op->extent.length);
514
515         if (rc >= 0) {
516                 /*
517                  * Assume we wrote the pages we originally sent.  The
518                  * osd might reply with fewer pages if our writeback
519                  * raced with a truncation and was adjusted at the osd,
520                  * so don't believe the reply.
521                  */
522                 wrote = req->r_num_pages;
523         } else {
524                 wrote = 0;
525                 mapping_set_error(mapping, rc);
526         }
527         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
528              inode, rc, bytes, wrote);
529
530         /* clean all pages */
531         for (i = 0; i < req->r_num_pages; i++) {
532                 page = req->r_pages[i];
533                 BUG_ON(!page);
534                 WARN_ON(!PageUptodate(page));
535
536                 writeback_stat =
537                         atomic_long_dec_return(&client->writeback_count);
538                 if (writeback_stat <
539                     CONGESTION_OFF_THRESH(client->mount_args->congestion_kb))
540                         clear_bdi_congested(&client->backing_dev_info,
541                                             BLK_RW_ASYNC);
542
543                 ceph_put_snap_context((void *)page->private);
544                 page->private = 0;
545                 ClearPagePrivate(page);
546                 dout("unlocking %d %p\n", i, page);
547                 end_page_writeback(page);
548
549                 /*
550                  * We lost the cache cap, need to truncate the page before
551                  * it is unlocked, otherwise we'd truncate it later in the
552                  * page truncation thread, possibly losing some data that
553                  * raced its way in
554                  */
555                 if ((issued & CEPH_CAP_FILE_CACHE) == 0)
556                         generic_error_remove_page(inode->i_mapping, page);
557
558                 unlock_page(page);
559         }
560         dout("%p wrote+cleaned %d pages\n", inode, wrote);
561         ceph_put_wrbuffer_cap_refs(ci, req->r_num_pages, snapc);
562
563         ceph_release_pages(req->r_pages, req->r_num_pages);
564         if (req->r_pages_from_pool)
565                 mempool_free(req->r_pages,
566                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
567         else
568                 kfree(req->r_pages);
569         ceph_osdc_put_request(req);
570 }
571
572 /*
573  * allocate a page vec, either directly, or if necessary, via a the
574  * mempool.  we avoid the mempool if we can because req->r_num_pages
575  * may be less than the maximum write size.
576  */
577 static void alloc_page_vec(struct ceph_client *client,
578                            struct ceph_osd_request *req)
579 {
580         req->r_pages = kmalloc(sizeof(struct page *) * req->r_num_pages,
581                                GFP_NOFS);
582         if (!req->r_pages) {
583                 req->r_pages = mempool_alloc(client->wb_pagevec_pool, GFP_NOFS);
584                 req->r_pages_from_pool = 1;
585                 WARN_ON(!req->r_pages);
586         }
587 }
588
589 /*
590  * initiate async writeback
591  */
592 static int ceph_writepages_start(struct address_space *mapping,
593                                  struct writeback_control *wbc)
594 {
595         struct inode *inode = mapping->host;
596         struct backing_dev_info *bdi = mapping->backing_dev_info;
597         struct ceph_inode_info *ci = ceph_inode(inode);
598         struct ceph_client *client;
599         pgoff_t index, start, end;
600         int range_whole = 0;
601         int should_loop = 1;
602         pgoff_t max_pages = 0, max_pages_ever = 0;
603         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
604         struct pagevec pvec;
605         int done = 0;
606         int rc = 0;
607         unsigned wsize = 1 << inode->i_blkbits;
608         struct ceph_osd_request *req = NULL;
609         int do_sync;
610         u64 snap_size = 0;
611
612         /*
613          * Include a 'sync' in the OSD request if this is a data
614          * integrity write (e.g., O_SYNC write or fsync()), or if our
615          * cap is being revoked.
616          */
617         do_sync = wbc->sync_mode == WB_SYNC_ALL;
618         if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
619                 do_sync = 1;
620         dout("writepages_start %p dosync=%d (mode=%s)\n",
621              inode, do_sync,
622              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
623              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
624
625         client = ceph_inode_to_client(inode);
626         if (client->mount_state == CEPH_MOUNT_SHUTDOWN) {
627                 pr_warning("writepage_start %p on forced umount\n", inode);
628                 return -EIO; /* we're in a forced umount, don't write! */
629         }
630         if (client->mount_args->wsize && client->mount_args->wsize < wsize)
631                 wsize = client->mount_args->wsize;
632         if (wsize < PAGE_CACHE_SIZE)
633                 wsize = PAGE_CACHE_SIZE;
634         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
635
636         pagevec_init(&pvec, 0);
637
638         /* ?? */
639         if (wbc->nonblocking && bdi_write_congested(bdi)) {
640                 dout(" writepages congested\n");
641                 wbc->encountered_congestion = 1;
642                 goto out_final;
643         }
644
645         /* where to start/end? */
646         if (wbc->range_cyclic) {
647                 start = mapping->writeback_index; /* Start from prev offset */
648                 end = -1;
649                 dout(" cyclic, start at %lu\n", start);
650         } else {
651                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
652                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
653                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
654                         range_whole = 1;
655                 should_loop = 0;
656                 dout(" not cyclic, %lu to %lu\n", start, end);
657         }
658         index = start;
659
660 retry:
661         /* find oldest snap context with dirty data */
662         ceph_put_snap_context(snapc);
663         snapc = get_oldest_context(inode, &snap_size);
664         if (!snapc) {
665                 /* hmm, why does writepages get called when there
666                    is no dirty data? */
667                 dout(" no snap context with dirty data?\n");
668                 goto out;
669         }
670         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
671              snapc, snapc->seq, snapc->num_snaps);
672         if (last_snapc && snapc != last_snapc) {
673                 /* if we switched to a newer snapc, restart our scan at the
674                  * start of the original file range. */
675                 dout("  snapc differs from last pass, restarting at %lu\n",
676                      index);
677                 index = start;
678         }
679         last_snapc = snapc;
680
681         while (!done && index <= end) {
682                 unsigned i;
683                 int first;
684                 pgoff_t next;
685                 int pvec_pages, locked_pages;
686                 struct page *page;
687                 int want;
688                 u64 offset, len;
689                 struct ceph_osd_request_head *reqhead;
690                 struct ceph_osd_op *op;
691                 long writeback_stat;
692
693                 next = 0;
694                 locked_pages = 0;
695                 max_pages = max_pages_ever;
696
697 get_more_pages:
698                 first = -1;
699                 want = min(end - index,
700                            min((pgoff_t)PAGEVEC_SIZE,
701                                max_pages - (pgoff_t)locked_pages) - 1)
702                         + 1;
703                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
704                                                 PAGECACHE_TAG_DIRTY,
705                                                 want);
706                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
707                 if (!pvec_pages && !locked_pages)
708                         break;
709                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
710                         page = pvec.pages[i];
711                         dout("? %p idx %lu\n", page, page->index);
712                         if (locked_pages == 0)
713                                 lock_page(page);  /* first page */
714                         else if (!trylock_page(page))
715                                 break;
716
717                         /* only dirty pages, or our accounting breaks */
718                         if (unlikely(!PageDirty(page)) ||
719                             unlikely(page->mapping != mapping)) {
720                                 dout("!dirty or !mapping %p\n", page);
721                                 unlock_page(page);
722                                 break;
723                         }
724                         if (!wbc->range_cyclic && page->index > end) {
725                                 dout("end of range %p\n", page);
726                                 done = 1;
727                                 unlock_page(page);
728                                 break;
729                         }
730                         if (next && (page->index != next)) {
731                                 dout("not consecutive %p\n", page);
732                                 unlock_page(page);
733                                 break;
734                         }
735                         if (wbc->sync_mode != WB_SYNC_NONE) {
736                                 dout("waiting on writeback %p\n", page);
737                                 wait_on_page_writeback(page);
738                         }
739                         if ((snap_size && page_offset(page) > snap_size) ||
740                             (!snap_size &&
741                              page_offset(page) > i_size_read(inode))) {
742                                 dout("%p page eof %llu\n", page, snap_size ?
743                                      snap_size : i_size_read(inode));
744                                 done = 1;
745                                 unlock_page(page);
746                                 break;
747                         }
748                         if (PageWriteback(page)) {
749                                 dout("%p under writeback\n", page);
750                                 unlock_page(page);
751                                 break;
752                         }
753
754                         /* only if matching snap context */
755                         pgsnapc = (void *)page->private;
756                         if (pgsnapc->seq > snapc->seq) {
757                                 dout("page snapc %p %lld > oldest %p %lld\n",
758                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
759                                 unlock_page(page);
760                                 if (!locked_pages)
761                                         continue; /* keep looking for snap */
762                                 break;
763                         }
764
765                         if (!clear_page_dirty_for_io(page)) {
766                                 dout("%p !clear_page_dirty_for_io\n", page);
767                                 unlock_page(page);
768                                 break;
769                         }
770
771                         /* ok */
772                         if (locked_pages == 0) {
773                                 /* prepare async write request */
774                                 offset = page->index << PAGE_CACHE_SHIFT;
775                                 len = wsize;
776                                 req = ceph_osdc_new_request(&client->osdc,
777                                             &ci->i_layout,
778                                             ceph_vino(inode),
779                                             offset, &len,
780                                             CEPH_OSD_OP_WRITE,
781                                             CEPH_OSD_FLAG_WRITE |
782                                                     CEPH_OSD_FLAG_ONDISK,
783                                             snapc, do_sync,
784                                             ci->i_truncate_seq,
785                                             ci->i_truncate_size,
786                                             &inode->i_mtime, true, 1);
787                                 max_pages = req->r_num_pages;
788
789                                 alloc_page_vec(client, req);
790                                 req->r_callback = writepages_finish;
791                                 req->r_inode = inode;
792                         }
793
794                         /* note position of first page in pvec */
795                         if (first < 0)
796                                 first = i;
797                         dout("%p will write page %p idx %lu\n",
798                              inode, page, page->index);
799
800                         writeback_stat = atomic_long_inc_return(&client->writeback_count);
801                         if (writeback_stat > CONGESTION_ON_THRESH(client->mount_args->congestion_kb)) {
802                                 set_bdi_congested(&client->backing_dev_info, BLK_RW_ASYNC);
803                         }
804
805                         set_page_writeback(page);
806                         req->r_pages[locked_pages] = page;
807                         locked_pages++;
808                         next = page->index + 1;
809                 }
810
811                 /* did we get anything? */
812                 if (!locked_pages)
813                         goto release_pvec_pages;
814                 if (i) {
815                         int j;
816                         BUG_ON(!locked_pages || first < 0);
817
818                         if (pvec_pages && i == pvec_pages &&
819                             locked_pages < max_pages) {
820                                 dout("reached end pvec, trying for more\n");
821                                 pagevec_reinit(&pvec);
822                                 goto get_more_pages;
823                         }
824
825                         /* shift unused pages over in the pvec...  we
826                          * will need to release them below. */
827                         for (j = i; j < pvec_pages; j++) {
828                                 dout(" pvec leftover page %p\n",
829                                      pvec.pages[j]);
830                                 pvec.pages[j-i+first] = pvec.pages[j];
831                         }
832                         pvec.nr -= i-first;
833                 }
834
835                 /* submit the write */
836                 offset = req->r_pages[0]->index << PAGE_CACHE_SHIFT;
837                 len = min((snap_size ? snap_size : i_size_read(inode)) - offset,
838                           (u64)locked_pages << PAGE_CACHE_SHIFT);
839                 dout("writepages got %d pages at %llu~%llu\n",
840                      locked_pages, offset, len);
841
842                 /* revise final length, page count */
843                 req->r_num_pages = locked_pages;
844                 reqhead = req->r_request->front.iov_base;
845                 op = (void *)(reqhead + 1);
846                 op->extent.length = cpu_to_le64(len);
847                 op->payload_len = cpu_to_le32(len);
848                 req->r_request->hdr.data_len = cpu_to_le32(len);
849
850                 ceph_osdc_start_request(&client->osdc, req, true);
851                 req = NULL;
852
853                 /* continue? */
854                 index = next;
855                 wbc->nr_to_write -= locked_pages;
856                 if (wbc->nr_to_write <= 0)
857                         done = 1;
858
859 release_pvec_pages:
860                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
861                      pvec.nr ? pvec.pages[0] : NULL);
862                 pagevec_release(&pvec);
863
864                 if (locked_pages && !done)
865                         goto retry;
866         }
867
868         if (should_loop && !done) {
869                 /* more to do; loop back to beginning of file */
870                 dout("writepages looping back to beginning of file\n");
871                 should_loop = 0;
872                 index = 0;
873                 goto retry;
874         }
875
876         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
877                 mapping->writeback_index = index;
878
879 out:
880         if (req)
881                 ceph_osdc_put_request(req);
882         if (rc > 0)
883                 rc = 0;  /* vfs expects us to return 0 */
884         ceph_put_snap_context(snapc);
885         dout("writepages done, rc = %d\n", rc);
886 out_final:
887         return rc;
888 }
889
890
891
892 /*
893  * See if a given @snapc is either writeable, or already written.
894  */
895 static int context_is_writeable_or_written(struct inode *inode,
896                                            struct ceph_snap_context *snapc)
897 {
898         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
899         int ret = !oldest || snapc->seq <= oldest->seq;
900
901         ceph_put_snap_context(oldest);
902         return ret;
903 }
904
905 /*
906  * We are only allowed to write into/dirty the page if the page is
907  * clean, or already dirty within the same snap context.
908  *
909  * called with page locked.
910  * return success with page locked,
911  * or any failure (incl -EAGAIN) with page unlocked.
912  */
913 static int ceph_update_writeable_page(struct file *file,
914                             loff_t pos, unsigned len,
915                             struct page *page)
916 {
917         struct inode *inode = file->f_dentry->d_inode;
918         struct ceph_inode_info *ci = ceph_inode(inode);
919         struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
920         loff_t page_off = pos & PAGE_CACHE_MASK;
921         int pos_in_page = pos & ~PAGE_CACHE_MASK;
922         int end_in_page = pos_in_page + len;
923         loff_t i_size;
924         int r;
925         struct ceph_snap_context *snapc, *oldest;
926
927 retry_locked:
928         /* writepages currently holds page lock, but if we change that later, */
929         wait_on_page_writeback(page);
930
931         /* check snap context */
932         BUG_ON(!ci->i_snap_realm);
933         down_read(&mdsc->snap_rwsem);
934         BUG_ON(!ci->i_snap_realm->cached_context);
935         snapc = (void *)page->private;
936         if (snapc && snapc != ci->i_head_snapc) {
937                 /*
938                  * this page is already dirty in another (older) snap
939                  * context!  is it writeable now?
940                  */
941                 oldest = get_oldest_context(inode, NULL);
942                 up_read(&mdsc->snap_rwsem);
943
944                 if (snapc->seq > oldest->seq) {
945                         ceph_put_snap_context(oldest);
946                         dout(" page %p snapc %p not current or oldest\n",
947                              page, snapc);
948                         /*
949                          * queue for writeback, and wait for snapc to
950                          * be writeable or written
951                          */
952                         snapc = ceph_get_snap_context(snapc);
953                         unlock_page(page);
954                         ceph_queue_writeback(inode);
955                         r = wait_event_interruptible(ci->i_cap_wq,
956                                context_is_writeable_or_written(inode, snapc));
957                         ceph_put_snap_context(snapc);
958                         if (r == -ERESTARTSYS)
959                                 return r;
960                         return -EAGAIN;
961                 }
962                 ceph_put_snap_context(oldest);
963
964                 /* yay, writeable, do it now (without dropping page lock) */
965                 dout(" page %p snapc %p not current, but oldest\n",
966                      page, snapc);
967                 if (!clear_page_dirty_for_io(page))
968                         goto retry_locked;
969                 r = writepage_nounlock(page, NULL);
970                 if (r < 0)
971                         goto fail_nosnap;
972                 goto retry_locked;
973         }
974
975         if (PageUptodate(page)) {
976                 dout(" page %p already uptodate\n", page);
977                 return 0;
978         }
979
980         /* full page? */
981         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
982                 return 0;
983
984         /* past end of file? */
985         i_size = inode->i_size;   /* caller holds i_mutex */
986
987         if (i_size + len > inode->i_sb->s_maxbytes) {
988                 /* file is too big */
989                 r = -EINVAL;
990                 goto fail;
991         }
992
993         if (page_off >= i_size ||
994             (pos_in_page == 0 && (pos+len) >= i_size &&
995              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
996                 dout(" zeroing %p 0 - %d and %d - %d\n",
997                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
998                 zero_user_segments(page,
999                                    0, pos_in_page,
1000                                    end_in_page, PAGE_CACHE_SIZE);
1001                 return 0;
1002         }
1003
1004         /* we need to read it. */
1005         up_read(&mdsc->snap_rwsem);
1006         r = readpage_nounlock(file, page);
1007         if (r < 0)
1008                 goto fail_nosnap;
1009         goto retry_locked;
1010
1011 fail:
1012         up_read(&mdsc->snap_rwsem);
1013 fail_nosnap:
1014         unlock_page(page);
1015         return r;
1016 }
1017
1018 /*
1019  * We are only allowed to write into/dirty the page if the page is
1020  * clean, or already dirty within the same snap context.
1021  */
1022 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1023                             loff_t pos, unsigned len, unsigned flags,
1024                             struct page **pagep, void **fsdata)
1025 {
1026         struct inode *inode = file->f_dentry->d_inode;
1027         struct page *page;
1028         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1029         int r;
1030
1031         do {
1032                 /* get a page */
1033                 page = grab_cache_page_write_begin(mapping, index, 0);
1034                 if (!page)
1035                         return -ENOMEM;
1036                 *pagep = page;
1037
1038                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1039                 inode, page, (int)pos, (int)len);
1040
1041                 r = ceph_update_writeable_page(file, pos, len, page);
1042         } while (r == -EAGAIN);
1043
1044         return r;
1045 }
1046
1047 /*
1048  * we don't do anything in here that simple_write_end doesn't do
1049  * except adjust dirty page accounting and drop read lock on
1050  * mdsc->snap_rwsem.
1051  */
1052 static int ceph_write_end(struct file *file, struct address_space *mapping,
1053                           loff_t pos, unsigned len, unsigned copied,
1054                           struct page *page, void *fsdata)
1055 {
1056         struct inode *inode = file->f_dentry->d_inode;
1057         struct ceph_client *client = ceph_inode_to_client(inode);
1058         struct ceph_mds_client *mdsc = &client->mdsc;
1059         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1060         int check_cap = 0;
1061
1062         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1063              inode, page, (int)pos, (int)copied, (int)len);
1064
1065         /* zero the stale part of the page if we did a short copy */
1066         if (copied < len)
1067                 zero_user_segment(page, from+copied, len);
1068
1069         /* did file size increase? */
1070         /* (no need for i_size_read(); we caller holds i_mutex */
1071         if (pos+copied > inode->i_size)
1072                 check_cap = ceph_inode_set_size(inode, pos+copied);
1073
1074         if (!PageUptodate(page))
1075                 SetPageUptodate(page);
1076
1077         set_page_dirty(page);
1078
1079         unlock_page(page);
1080         up_read(&mdsc->snap_rwsem);
1081         page_cache_release(page);
1082
1083         if (check_cap)
1084                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1085
1086         return copied;
1087 }
1088
1089 /*
1090  * we set .direct_IO to indicate direct io is supported, but since we
1091  * intercept O_DIRECT reads and writes early, this function should
1092  * never get called.
1093  */
1094 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1095                               const struct iovec *iov,
1096                               loff_t pos, unsigned long nr_segs)
1097 {
1098         WARN_ON(1);
1099         return -EINVAL;
1100 }
1101
1102 const struct address_space_operations ceph_aops = {
1103         .readpage = ceph_readpage,
1104         .readpages = ceph_readpages,
1105         .writepage = ceph_writepage,
1106         .writepages = ceph_writepages_start,
1107         .write_begin = ceph_write_begin,
1108         .write_end = ceph_write_end,
1109         .set_page_dirty = ceph_set_page_dirty,
1110         .invalidatepage = ceph_invalidatepage,
1111         .releasepage = ceph_releasepage,
1112         .direct_IO = ceph_direct_io,
1113 };
1114
1115
1116 /*
1117  * vm ops
1118  */
1119
1120 /*
1121  * Reuse write_begin here for simplicity.
1122  */
1123 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1124 {
1125         struct inode *inode = vma->vm_file->f_dentry->d_inode;
1126         struct page *page = vmf->page;
1127         struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
1128         loff_t off = page->index << PAGE_CACHE_SHIFT;
1129         loff_t size, len;
1130         int ret;
1131
1132         size = i_size_read(inode);
1133         if (off + PAGE_CACHE_SIZE <= size)
1134                 len = PAGE_CACHE_SIZE;
1135         else
1136                 len = size & ~PAGE_CACHE_MASK;
1137
1138         dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1139              off, len, page, page->index);
1140
1141         lock_page(page);
1142
1143         ret = VM_FAULT_NOPAGE;
1144         if ((off > size) ||
1145             (page->mapping != inode->i_mapping))
1146                 goto out;
1147
1148         ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1149         if (ret == 0) {
1150                 /* success.  we'll keep the page locked. */
1151                 set_page_dirty(page);
1152                 up_read(&mdsc->snap_rwsem);
1153                 ret = VM_FAULT_LOCKED;
1154         } else {
1155                 if (ret == -ENOMEM)
1156                         ret = VM_FAULT_OOM;
1157                 else
1158                         ret = VM_FAULT_SIGBUS;
1159         }
1160 out:
1161         dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1162         if (ret != VM_FAULT_LOCKED)
1163                 unlock_page(page);
1164         return ret;
1165 }
1166
1167 static struct vm_operations_struct ceph_vmops = {
1168         .fault          = filemap_fault,
1169         .page_mkwrite   = ceph_page_mkwrite,
1170 };
1171
1172 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1173 {
1174         struct address_space *mapping = file->f_mapping;
1175
1176         if (!mapping->a_ops->readpage)
1177                 return -ENOEXEC;
1178         file_accessed(file);
1179         vma->vm_ops = &ceph_vmops;
1180         vma->vm_flags |= VM_CAN_NONLINEAR;
1181         return 0;
1182 }