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