include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit...
[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         struct pagevec pvec;
278         loff_t offset;
279         u64 len;
280
281         dout("readpages %p file %p nr_pages %d\n",
282              inode, file, nr_pages);
283
284         pages = page_vector_from_list(page_list, &nr_pages);
285         if (IS_ERR(pages))
286                 return PTR_ERR(pages);
287
288         /* guess read extent */
289         offset = pages[0]->index << PAGE_CACHE_SHIFT;
290         len = nr_pages << PAGE_CACHE_SHIFT;
291         rc = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
292                                  offset, &len,
293                                  ci->i_truncate_seq, ci->i_truncate_size,
294                                  pages, nr_pages);
295         if (rc == -ENOENT)
296                 rc = 0;
297         if (rc < 0)
298                 goto out;
299
300         /* set uptodate and add to lru in pagevec-sized chunks */
301         pagevec_init(&pvec, 0);
302         for (; !list_empty(page_list) && len > 0;
303              rc -= PAGE_CACHE_SIZE, len -= PAGE_CACHE_SIZE) {
304                 struct page *page =
305                         list_entry(page_list->prev, struct page, lru);
306
307                 list_del(&page->lru);
308
309                 if (rc < (int)PAGE_CACHE_SIZE) {
310                         /* zero (remainder of) page */
311                         int s = rc < 0 ? 0 : rc;
312                         zero_user_segment(page, s, PAGE_CACHE_SIZE);
313                 }
314
315                 if (add_to_page_cache(page, mapping, page->index, GFP_NOFS)) {
316                         page_cache_release(page);
317                         dout("readpages %p add_to_page_cache failed %p\n",
318                              inode, page);
319                         continue;
320                 }
321                 dout("readpages %p adding %p idx %lu\n", inode, page,
322                      page->index);
323                 flush_dcache_page(page);
324                 SetPageUptodate(page);
325                 unlock_page(page);
326                 if (pagevec_add(&pvec, page) == 0)
327                         pagevec_lru_add_file(&pvec);   /* add to lru */
328         }
329         pagevec_lru_add_file(&pvec);
330         rc = 0;
331
332 out:
333         kfree(pages);
334         return rc;
335 }
336
337 /*
338  * Get ref for the oldest snapc for an inode with dirty data... that is, the
339  * only snap context we are allowed to write back.
340  *
341  * Caller holds i_lock.
342  */
343 static struct ceph_snap_context *__get_oldest_context(struct inode *inode,
344                                                       u64 *snap_size)
345 {
346         struct ceph_inode_info *ci = ceph_inode(inode);
347         struct ceph_snap_context *snapc = NULL;
348         struct ceph_cap_snap *capsnap = NULL;
349
350         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
351                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
352                      capsnap->context, capsnap->dirty_pages);
353                 if (capsnap->dirty_pages) {
354                         snapc = ceph_get_snap_context(capsnap->context);
355                         if (snap_size)
356                                 *snap_size = capsnap->size;
357                         break;
358                 }
359         }
360         if (!snapc && ci->i_snap_realm) {
361                 snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
362                 dout(" head snapc %p has %d dirty pages\n",
363                      snapc, ci->i_wrbuffer_ref_head);
364         }
365         return snapc;
366 }
367
368 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
369                                                     u64 *snap_size)
370 {
371         struct ceph_snap_context *snapc = NULL;
372
373         spin_lock(&inode->i_lock);
374         snapc = __get_oldest_context(inode, snap_size);
375         spin_unlock(&inode->i_lock);
376         return snapc;
377 }
378
379 /*
380  * Write a single page, but leave the page locked.
381  *
382  * If we get a write error, set the page error bit, but still adjust the
383  * dirty page accounting (i.e., page is no longer dirty).
384  */
385 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
386 {
387         struct inode *inode;
388         struct ceph_inode_info *ci;
389         struct ceph_client *client;
390         struct ceph_osd_client *osdc;
391         loff_t page_off = page->index << PAGE_CACHE_SHIFT;
392         int len = PAGE_CACHE_SIZE;
393         loff_t i_size;
394         int err = 0;
395         struct ceph_snap_context *snapc;
396         u64 snap_size = 0;
397         long writeback_stat;
398
399         dout("writepage %p idx %lu\n", page, page->index);
400
401         if (!page->mapping || !page->mapping->host) {
402                 dout("writepage %p - no mapping\n", page);
403                 return -EFAULT;
404         }
405         inode = page->mapping->host;
406         ci = ceph_inode(inode);
407         client = ceph_inode_to_client(inode);
408         osdc = &client->osdc;
409
410         /* verify this is a writeable snap context */
411         snapc = (void *)page->private;
412         if (snapc == NULL) {
413                 dout("writepage %p page %p not dirty?\n", inode, page);
414                 goto out;
415         }
416         if (snapc != get_oldest_context(inode, &snap_size)) {
417                 dout("writepage %p page %p snapc %p not writeable - noop\n",
418                      inode, page, (void *)page->private);
419                 /* we should only noop if called by kswapd */
420                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
421                 goto out;
422         }
423
424         /* is this a partial page at end of file? */
425         if (snap_size)
426                 i_size = snap_size;
427         else
428                 i_size = i_size_read(inode);
429         if (i_size < page_off + len)
430                 len = i_size - page_off;
431
432         dout("writepage %p page %p index %lu on %llu~%u\n",
433              inode, page, page->index, page_off, len);
434
435         writeback_stat = atomic_long_inc_return(&client->writeback_count);
436         if (writeback_stat >
437             CONGESTION_ON_THRESH(client->mount_args->congestion_kb))
438                 set_bdi_congested(&client->backing_dev_info, BLK_RW_ASYNC);
439
440         set_page_writeback(page);
441         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
442                                    &ci->i_layout, snapc,
443                                    page_off, len,
444                                    ci->i_truncate_seq, ci->i_truncate_size,
445                                    &inode->i_mtime,
446                                    &page, 1, 0, 0, true);
447         if (err < 0) {
448                 dout("writepage setting page/mapping error %d %p\n", err, page);
449                 SetPageError(page);
450                 mapping_set_error(&inode->i_data, err);
451                 if (wbc)
452                         wbc->pages_skipped++;
453         } else {
454                 dout("writepage cleaned page %p\n", page);
455                 err = 0;  /* vfs expects us to return 0 */
456         }
457         page->private = 0;
458         ClearPagePrivate(page);
459         end_page_writeback(page);
460         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
461         ceph_put_snap_context(snapc);
462 out:
463         return err;
464 }
465
466 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
467 {
468         int err;
469         struct inode *inode = page->mapping->host;
470         BUG_ON(!inode);
471         igrab(inode);
472         err = writepage_nounlock(page, wbc);
473         unlock_page(page);
474         iput(inode);
475         return err;
476 }
477
478
479 /*
480  * lame release_pages helper.  release_pages() isn't exported to
481  * modules.
482  */
483 static void ceph_release_pages(struct page **pages, int num)
484 {
485         struct pagevec pvec;
486         int i;
487
488         pagevec_init(&pvec, 0);
489         for (i = 0; i < num; i++) {
490                 if (pagevec_add(&pvec, pages[i]) == 0)
491                         pagevec_release(&pvec);
492         }
493         pagevec_release(&pvec);
494 }
495
496
497 /*
498  * async writeback completion handler.
499  *
500  * If we get an error, set the mapping error bit, but not the individual
501  * page error bits.
502  */
503 static void writepages_finish(struct ceph_osd_request *req,
504                               struct ceph_msg *msg)
505 {
506         struct inode *inode = req->r_inode;
507         struct ceph_osd_reply_head *replyhead;
508         struct ceph_osd_op *op;
509         struct ceph_inode_info *ci = ceph_inode(inode);
510         unsigned wrote;
511         struct page *page;
512         int i;
513         struct ceph_snap_context *snapc = req->r_snapc;
514         struct address_space *mapping = inode->i_mapping;
515         struct writeback_control *wbc = req->r_wbc;
516         __s32 rc = -EIO;
517         u64 bytes = 0;
518         struct ceph_client *client = ceph_inode_to_client(inode);
519         long writeback_stat;
520         unsigned issued = __ceph_caps_issued(ci, NULL);
521
522         /* parse reply */
523         replyhead = msg->front.iov_base;
524         WARN_ON(le32_to_cpu(replyhead->num_ops) == 0);
525         op = (void *)(replyhead + 1);
526         rc = le32_to_cpu(replyhead->result);
527         bytes = le64_to_cpu(op->extent.length);
528
529         if (rc >= 0) {
530                 /*
531                  * Assume we wrote the pages we originally sent.  The
532                  * osd might reply with fewer pages if our writeback
533                  * raced with a truncation and was adjusted at the osd,
534                  * so don't believe the reply.
535                  */
536                 wrote = req->r_num_pages;
537         } else {
538                 wrote = 0;
539                 mapping_set_error(mapping, rc);
540         }
541         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
542              inode, rc, bytes, wrote);
543
544         /* clean all pages */
545         for (i = 0; i < req->r_num_pages; i++) {
546                 page = req->r_pages[i];
547                 BUG_ON(!page);
548                 WARN_ON(!PageUptodate(page));
549
550                 writeback_stat =
551                         atomic_long_dec_return(&client->writeback_count);
552                 if (writeback_stat <
553                     CONGESTION_OFF_THRESH(client->mount_args->congestion_kb))
554                         clear_bdi_congested(&client->backing_dev_info,
555                                             BLK_RW_ASYNC);
556
557                 if (i >= wrote) {
558                         dout("inode %p skipping page %p\n", inode, page);
559                         wbc->pages_skipped++;
560                 }
561                 page->private = 0;
562                 ClearPagePrivate(page);
563                 ceph_put_snap_context(snapc);
564                 dout("unlocking %d %p\n", i, page);
565                 end_page_writeback(page);
566
567                 /*
568                  * We lost the cache cap, need to truncate the page before
569                  * it is unlocked, otherwise we'd truncate it later in the
570                  * page truncation thread, possibly losing some data that
571                  * raced its way in
572                  */
573                 if ((issued & CEPH_CAP_FILE_CACHE) == 0)
574                         generic_error_remove_page(inode->i_mapping, page);
575
576                 unlock_page(page);
577         }
578         dout("%p wrote+cleaned %d pages\n", inode, wrote);
579         ceph_put_wrbuffer_cap_refs(ci, req->r_num_pages, snapc);
580
581         ceph_release_pages(req->r_pages, req->r_num_pages);
582         if (req->r_pages_from_pool)
583                 mempool_free(req->r_pages,
584                              ceph_client(inode->i_sb)->wb_pagevec_pool);
585         else
586                 kfree(req->r_pages);
587         ceph_osdc_put_request(req);
588 }
589
590 /*
591  * allocate a page vec, either directly, or if necessary, via a the
592  * mempool.  we avoid the mempool if we can because req->r_num_pages
593  * may be less than the maximum write size.
594  */
595 static void alloc_page_vec(struct ceph_client *client,
596                            struct ceph_osd_request *req)
597 {
598         req->r_pages = kmalloc(sizeof(struct page *) * req->r_num_pages,
599                                GFP_NOFS);
600         if (!req->r_pages) {
601                 req->r_pages = mempool_alloc(client->wb_pagevec_pool, GFP_NOFS);
602                 req->r_pages_from_pool = 1;
603                 WARN_ON(!req->r_pages);
604         }
605 }
606
607 /*
608  * initiate async writeback
609  */
610 static int ceph_writepages_start(struct address_space *mapping,
611                                  struct writeback_control *wbc)
612 {
613         struct inode *inode = mapping->host;
614         struct backing_dev_info *bdi = mapping->backing_dev_info;
615         struct ceph_inode_info *ci = ceph_inode(inode);
616         struct ceph_client *client;
617         pgoff_t index, start, end;
618         int range_whole = 0;
619         int should_loop = 1;
620         pgoff_t max_pages = 0, max_pages_ever = 0;
621         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL;
622         struct pagevec pvec;
623         int done = 0;
624         int rc = 0;
625         unsigned wsize = 1 << inode->i_blkbits;
626         struct ceph_osd_request *req = NULL;
627         int do_sync;
628         u64 snap_size = 0;
629
630         /*
631          * Include a 'sync' in the OSD request if this is a data
632          * integrity write (e.g., O_SYNC write or fsync()), or if our
633          * cap is being revoked.
634          */
635         do_sync = wbc->sync_mode == WB_SYNC_ALL;
636         if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
637                 do_sync = 1;
638         dout("writepages_start %p dosync=%d (mode=%s)\n",
639              inode, do_sync,
640              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
641              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
642
643         client = ceph_inode_to_client(inode);
644         if (client->mount_state == CEPH_MOUNT_SHUTDOWN) {
645                 pr_warning("writepage_start %p on forced umount\n", inode);
646                 return -EIO; /* we're in a forced umount, don't write! */
647         }
648         if (client->mount_args->wsize && client->mount_args->wsize < wsize)
649                 wsize = client->mount_args->wsize;
650         if (wsize < PAGE_CACHE_SIZE)
651                 wsize = PAGE_CACHE_SIZE;
652         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
653
654         pagevec_init(&pvec, 0);
655
656         /* ?? */
657         if (wbc->nonblocking && bdi_write_congested(bdi)) {
658                 dout(" writepages congested\n");
659                 wbc->encountered_congestion = 1;
660                 goto out_final;
661         }
662
663         /* where to start/end? */
664         if (wbc->range_cyclic) {
665                 start = mapping->writeback_index; /* Start from prev offset */
666                 end = -1;
667                 dout(" cyclic, start at %lu\n", start);
668         } else {
669                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
670                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
671                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
672                         range_whole = 1;
673                 should_loop = 0;
674                 dout(" not cyclic, %lu to %lu\n", start, end);
675         }
676         index = start;
677
678 retry:
679         /* find oldest snap context with dirty data */
680         ceph_put_snap_context(snapc);
681         snapc = get_oldest_context(inode, &snap_size);
682         if (!snapc) {
683                 /* hmm, why does writepages get called when there
684                    is no dirty data? */
685                 dout(" no snap context with dirty data?\n");
686                 goto out;
687         }
688         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
689              snapc, snapc->seq, snapc->num_snaps);
690         if (last_snapc && snapc != last_snapc) {
691                 /* if we switched to a newer snapc, restart our scan at the
692                  * start of the original file range. */
693                 dout("  snapc differs from last pass, restarting at %lu\n",
694                      index);
695                 index = start;
696         }
697         last_snapc = snapc;
698
699         while (!done && index <= end) {
700                 unsigned i;
701                 int first;
702                 pgoff_t next;
703                 int pvec_pages, locked_pages;
704                 struct page *page;
705                 int want;
706                 u64 offset, len;
707                 struct ceph_osd_request_head *reqhead;
708                 struct ceph_osd_op *op;
709                 long writeback_stat;
710
711                 next = 0;
712                 locked_pages = 0;
713                 max_pages = max_pages_ever;
714
715 get_more_pages:
716                 first = -1;
717                 want = min(end - index,
718                            min((pgoff_t)PAGEVEC_SIZE,
719                                max_pages - (pgoff_t)locked_pages) - 1)
720                         + 1;
721                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
722                                                 PAGECACHE_TAG_DIRTY,
723                                                 want);
724                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
725                 if (!pvec_pages && !locked_pages)
726                         break;
727                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
728                         page = pvec.pages[i];
729                         dout("? %p idx %lu\n", page, page->index);
730                         if (locked_pages == 0)
731                                 lock_page(page);  /* first page */
732                         else if (!trylock_page(page))
733                                 break;
734
735                         /* only dirty pages, or our accounting breaks */
736                         if (unlikely(!PageDirty(page)) ||
737                             unlikely(page->mapping != mapping)) {
738                                 dout("!dirty or !mapping %p\n", page);
739                                 unlock_page(page);
740                                 break;
741                         }
742                         if (!wbc->range_cyclic && page->index > end) {
743                                 dout("end of range %p\n", page);
744                                 done = 1;
745                                 unlock_page(page);
746                                 break;
747                         }
748                         if (next && (page->index != next)) {
749                                 dout("not consecutive %p\n", page);
750                                 unlock_page(page);
751                                 break;
752                         }
753                         if (wbc->sync_mode != WB_SYNC_NONE) {
754                                 dout("waiting on writeback %p\n", page);
755                                 wait_on_page_writeback(page);
756                         }
757                         if ((snap_size && page_offset(page) > snap_size) ||
758                             (!snap_size &&
759                              page_offset(page) > i_size_read(inode))) {
760                                 dout("%p page eof %llu\n", page, snap_size ?
761                                      snap_size : i_size_read(inode));
762                                 done = 1;
763                                 unlock_page(page);
764                                 break;
765                         }
766                         if (PageWriteback(page)) {
767                                 dout("%p under writeback\n", page);
768                                 unlock_page(page);
769                                 break;
770                         }
771
772                         /* only if matching snap context */
773                         if (snapc != (void *)page->private) {
774                                 dout("page snapc %p != oldest %p\n",
775                                      (void *)page->private, snapc);
776                                 unlock_page(page);
777                                 if (!locked_pages)
778                                         continue; /* keep looking for snap */
779                                 break;
780                         }
781
782                         if (!clear_page_dirty_for_io(page)) {
783                                 dout("%p !clear_page_dirty_for_io\n", page);
784                                 unlock_page(page);
785                                 break;
786                         }
787
788                         /* ok */
789                         if (locked_pages == 0) {
790                                 /* prepare async write request */
791                                 offset = page->index << PAGE_CACHE_SHIFT;
792                                 len = wsize;
793                                 req = ceph_osdc_new_request(&client->osdc,
794                                             &ci->i_layout,
795                                             ceph_vino(inode),
796                                             offset, &len,
797                                             CEPH_OSD_OP_WRITE,
798                                             CEPH_OSD_FLAG_WRITE |
799                                                     CEPH_OSD_FLAG_ONDISK,
800                                             snapc, do_sync,
801                                             ci->i_truncate_seq,
802                                             ci->i_truncate_size,
803                                             &inode->i_mtime, true, 1);
804                                 max_pages = req->r_num_pages;
805
806                                 alloc_page_vec(client, req);
807                                 req->r_callback = writepages_finish;
808                                 req->r_inode = inode;
809                                 req->r_wbc = wbc;
810                         }
811
812                         /* note position of first page in pvec */
813                         if (first < 0)
814                                 first = i;
815                         dout("%p will write page %p idx %lu\n",
816                              inode, page, page->index);
817
818                         writeback_stat = atomic_long_inc_return(&client->writeback_count);
819                         if (writeback_stat > CONGESTION_ON_THRESH(client->mount_args->congestion_kb)) {
820                                 set_bdi_congested(&client->backing_dev_info, BLK_RW_ASYNC);
821                         }
822
823                         set_page_writeback(page);
824                         req->r_pages[locked_pages] = page;
825                         locked_pages++;
826                         next = page->index + 1;
827                 }
828
829                 /* did we get anything? */
830                 if (!locked_pages)
831                         goto release_pvec_pages;
832                 if (i) {
833                         int j;
834                         BUG_ON(!locked_pages || first < 0);
835
836                         if (pvec_pages && i == pvec_pages &&
837                             locked_pages < max_pages) {
838                                 dout("reached end pvec, trying for more\n");
839                                 pagevec_reinit(&pvec);
840                                 goto get_more_pages;
841                         }
842
843                         /* shift unused pages over in the pvec...  we
844                          * will need to release them below. */
845                         for (j = i; j < pvec_pages; j++) {
846                                 dout(" pvec leftover page %p\n",
847                                      pvec.pages[j]);
848                                 pvec.pages[j-i+first] = pvec.pages[j];
849                         }
850                         pvec.nr -= i-first;
851                 }
852
853                 /* submit the write */
854                 offset = req->r_pages[0]->index << PAGE_CACHE_SHIFT;
855                 len = min((snap_size ? snap_size : i_size_read(inode)) - offset,
856                           (u64)locked_pages << PAGE_CACHE_SHIFT);
857                 dout("writepages got %d pages at %llu~%llu\n",
858                      locked_pages, offset, len);
859
860                 /* revise final length, page count */
861                 req->r_num_pages = locked_pages;
862                 reqhead = req->r_request->front.iov_base;
863                 op = (void *)(reqhead + 1);
864                 op->extent.length = cpu_to_le64(len);
865                 op->payload_len = cpu_to_le32(len);
866                 req->r_request->hdr.data_len = cpu_to_le32(len);
867
868                 ceph_osdc_start_request(&client->osdc, req, true);
869                 req = NULL;
870
871                 /* continue? */
872                 index = next;
873                 wbc->nr_to_write -= locked_pages;
874                 if (wbc->nr_to_write <= 0)
875                         done = 1;
876
877 release_pvec_pages:
878                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
879                      pvec.nr ? pvec.pages[0] : NULL);
880                 pagevec_release(&pvec);
881
882                 if (locked_pages && !done)
883                         goto retry;
884         }
885
886         if (should_loop && !done) {
887                 /* more to do; loop back to beginning of file */
888                 dout("writepages looping back to beginning of file\n");
889                 should_loop = 0;
890                 index = 0;
891                 goto retry;
892         }
893
894         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
895                 mapping->writeback_index = index;
896
897 out:
898         if (req)
899                 ceph_osdc_put_request(req);
900         if (rc > 0)
901                 rc = 0;  /* vfs expects us to return 0 */
902         ceph_put_snap_context(snapc);
903         dout("writepages done, rc = %d\n", rc);
904 out_final:
905         return rc;
906 }
907
908
909
910 /*
911  * See if a given @snapc is either writeable, or already written.
912  */
913 static int context_is_writeable_or_written(struct inode *inode,
914                                            struct ceph_snap_context *snapc)
915 {
916         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
917         return !oldest || snapc->seq <= oldest->seq;
918 }
919
920 /*
921  * We are only allowed to write into/dirty the page if the page is
922  * clean, or already dirty within the same snap context.
923  *
924  * called with page locked.
925  * return success with page locked,
926  * or any failure (incl -EAGAIN) with page unlocked.
927  */
928 static int ceph_update_writeable_page(struct file *file,
929                             loff_t pos, unsigned len,
930                             struct page *page)
931 {
932         struct inode *inode = file->f_dentry->d_inode;
933         struct ceph_inode_info *ci = ceph_inode(inode);
934         struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
935         loff_t page_off = pos & PAGE_CACHE_MASK;
936         int pos_in_page = pos & ~PAGE_CACHE_MASK;
937         int end_in_page = pos_in_page + len;
938         loff_t i_size;
939         struct ceph_snap_context *snapc;
940         int r;
941
942 retry_locked:
943         /* writepages currently holds page lock, but if we change that later, */
944         wait_on_page_writeback(page);
945
946         /* check snap context */
947         BUG_ON(!ci->i_snap_realm);
948         down_read(&mdsc->snap_rwsem);
949         BUG_ON(!ci->i_snap_realm->cached_context);
950         if (page->private &&
951             (void *)page->private != ci->i_snap_realm->cached_context) {
952                 /*
953                  * this page is already dirty in another (older) snap
954                  * context!  is it writeable now?
955                  */
956                 snapc = get_oldest_context(inode, NULL);
957                 up_read(&mdsc->snap_rwsem);
958
959                 if (snapc != (void *)page->private) {
960                         dout(" page %p snapc %p not current or oldest\n",
961                              page, (void *)page->private);
962                         /*
963                          * queue for writeback, and wait for snapc to
964                          * be writeable or written
965                          */
966                         snapc = ceph_get_snap_context((void *)page->private);
967                         unlock_page(page);
968                         ceph_queue_writeback(inode);
969                         r = wait_event_interruptible(ci->i_cap_wq,
970                                context_is_writeable_or_written(inode, snapc));
971                         ceph_put_snap_context(snapc);
972                         if (r == -ERESTARTSYS)
973                                 return r;
974                         return -EAGAIN;
975                 }
976
977                 /* yay, writeable, do it now (without dropping page lock) */
978                 dout(" page %p snapc %p not current, but oldest\n",
979                      page, snapc);
980                 if (!clear_page_dirty_for_io(page))
981                         goto retry_locked;
982                 r = writepage_nounlock(page, NULL);
983                 if (r < 0)
984                         goto fail_nosnap;
985                 goto retry_locked;
986         }
987
988         if (PageUptodate(page)) {
989                 dout(" page %p already uptodate\n", page);
990                 return 0;
991         }
992
993         /* full page? */
994         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
995                 return 0;
996
997         /* past end of file? */
998         i_size = inode->i_size;   /* caller holds i_mutex */
999
1000         if (i_size + len > inode->i_sb->s_maxbytes) {
1001                 /* file is too big */
1002                 r = -EINVAL;
1003                 goto fail;
1004         }
1005
1006         if (page_off >= i_size ||
1007             (pos_in_page == 0 && (pos+len) >= i_size &&
1008              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1009                 dout(" zeroing %p 0 - %d and %d - %d\n",
1010                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1011                 zero_user_segments(page,
1012                                    0, pos_in_page,
1013                                    end_in_page, PAGE_CACHE_SIZE);
1014                 return 0;
1015         }
1016
1017         /* we need to read it. */
1018         up_read(&mdsc->snap_rwsem);
1019         r = readpage_nounlock(file, page);
1020         if (r < 0)
1021                 goto fail_nosnap;
1022         goto retry_locked;
1023
1024 fail:
1025         up_read(&mdsc->snap_rwsem);
1026 fail_nosnap:
1027         unlock_page(page);
1028         return r;
1029 }
1030
1031 /*
1032  * We are only allowed to write into/dirty the page if the page is
1033  * clean, or already dirty within the same snap context.
1034  */
1035 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1036                             loff_t pos, unsigned len, unsigned flags,
1037                             struct page **pagep, void **fsdata)
1038 {
1039         struct inode *inode = file->f_dentry->d_inode;
1040         struct page *page;
1041         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1042         int r;
1043
1044         do {
1045                 /* get a page */
1046                 page = grab_cache_page_write_begin(mapping, index, 0);
1047                 if (!page)
1048                         return -ENOMEM;
1049                 *pagep = page;
1050
1051                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1052                 inode, page, (int)pos, (int)len);
1053
1054                 r = ceph_update_writeable_page(file, pos, len, page);
1055         } while (r == -EAGAIN);
1056
1057         return r;
1058 }
1059
1060 /*
1061  * we don't do anything in here that simple_write_end doesn't do
1062  * except adjust dirty page accounting and drop read lock on
1063  * mdsc->snap_rwsem.
1064  */
1065 static int ceph_write_end(struct file *file, struct address_space *mapping,
1066                           loff_t pos, unsigned len, unsigned copied,
1067                           struct page *page, void *fsdata)
1068 {
1069         struct inode *inode = file->f_dentry->d_inode;
1070         struct ceph_client *client = ceph_inode_to_client(inode);
1071         struct ceph_mds_client *mdsc = &client->mdsc;
1072         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1073         int check_cap = 0;
1074
1075         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1076              inode, page, (int)pos, (int)copied, (int)len);
1077
1078         /* zero the stale part of the page if we did a short copy */
1079         if (copied < len)
1080                 zero_user_segment(page, from+copied, len);
1081
1082         /* did file size increase? */
1083         /* (no need for i_size_read(); we caller holds i_mutex */
1084         if (pos+copied > inode->i_size)
1085                 check_cap = ceph_inode_set_size(inode, pos+copied);
1086
1087         if (!PageUptodate(page))
1088                 SetPageUptodate(page);
1089
1090         set_page_dirty(page);
1091
1092         unlock_page(page);
1093         up_read(&mdsc->snap_rwsem);
1094         page_cache_release(page);
1095
1096         if (check_cap)
1097                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1098
1099         return copied;
1100 }
1101
1102 /*
1103  * we set .direct_IO to indicate direct io is supported, but since we
1104  * intercept O_DIRECT reads and writes early, this function should
1105  * never get called.
1106  */
1107 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1108                               const struct iovec *iov,
1109                               loff_t pos, unsigned long nr_segs)
1110 {
1111         WARN_ON(1);
1112         return -EINVAL;
1113 }
1114
1115 const struct address_space_operations ceph_aops = {
1116         .readpage = ceph_readpage,
1117         .readpages = ceph_readpages,
1118         .writepage = ceph_writepage,
1119         .writepages = ceph_writepages_start,
1120         .write_begin = ceph_write_begin,
1121         .write_end = ceph_write_end,
1122         .set_page_dirty = ceph_set_page_dirty,
1123         .invalidatepage = ceph_invalidatepage,
1124         .releasepage = ceph_releasepage,
1125         .direct_IO = ceph_direct_io,
1126 };
1127
1128
1129 /*
1130  * vm ops
1131  */
1132
1133 /*
1134  * Reuse write_begin here for simplicity.
1135  */
1136 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1137 {
1138         struct inode *inode = vma->vm_file->f_dentry->d_inode;
1139         struct page *page = vmf->page;
1140         struct ceph_mds_client *mdsc = &ceph_inode_to_client(inode)->mdsc;
1141         loff_t off = page->index << PAGE_CACHE_SHIFT;
1142         loff_t size, len;
1143         int ret;
1144
1145         size = i_size_read(inode);
1146         if (off + PAGE_CACHE_SIZE <= size)
1147                 len = PAGE_CACHE_SIZE;
1148         else
1149                 len = size & ~PAGE_CACHE_MASK;
1150
1151         dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1152              off, len, page, page->index);
1153
1154         lock_page(page);
1155
1156         ret = VM_FAULT_NOPAGE;
1157         if ((off > size) ||
1158             (page->mapping != inode->i_mapping))
1159                 goto out;
1160
1161         ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1162         if (ret == 0) {
1163                 /* success.  we'll keep the page locked. */
1164                 set_page_dirty(page);
1165                 up_read(&mdsc->snap_rwsem);
1166                 ret = VM_FAULT_LOCKED;
1167         } else {
1168                 if (ret == -ENOMEM)
1169                         ret = VM_FAULT_OOM;
1170                 else
1171                         ret = VM_FAULT_SIGBUS;
1172         }
1173 out:
1174         dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1175         if (ret != VM_FAULT_LOCKED)
1176                 unlock_page(page);
1177         return ret;
1178 }
1179
1180 static struct vm_operations_struct ceph_vmops = {
1181         .fault          = filemap_fault,
1182         .page_mkwrite   = ceph_page_mkwrite,
1183 };
1184
1185 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1186 {
1187         struct address_space *mapping = file->f_mapping;
1188
1189         if (!mapping->a_ops->readpage)
1190                 return -ENOEXEC;
1191         file_accessed(file);
1192         vma->vm_ops = &ceph_vmops;
1193         vma->vm_flags |= VM_CAN_NONLINEAR;
1194         return 0;
1195 }