51bcc5ce323024a995d300b4a6035ecf8e72e94b
[linux-2.6.git] / fs / ceph / addr.c
1 #include <linux/ceph/ceph_debug.h>
2
3 #include <linux/backing-dev.h>
4 #include <linux/fs.h>
5 #include <linux/mm.h>
6 #include <linux/pagemap.h>
7 #include <linux/writeback.h>    /* generic_writepages */
8 #include <linux/slab.h>
9 #include <linux/pagevec.h>
10 #include <linux/task_io_accounting_ops.h>
11
12 #include "super.h"
13 #include "mds_client.h"
14 #include <linux/ceph/osd_client.h>
15
16 /*
17  * Ceph address space ops.
18  *
19  * There are a few funny things going on here.
20  *
21  * The page->private field is used to reference a struct
22  * ceph_snap_context for _every_ dirty page.  This indicates which
23  * snapshot the page was logically dirtied in, and thus which snap
24  * context needs to be associated with the osd write during writeback.
25  *
26  * Similarly, struct ceph_inode_info maintains a set of counters to
27  * count dirty pages on the inode.  In the absense of snapshots,
28  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
29  *
30  * When a snapshot is taken (that is, when the client receives
31  * notification that a snapshot was taken), each inode with caps and
32  * with dirty pages (dirty pages implies there is a cap) gets a new
33  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
34  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
35  * moved to capsnap->dirty. (Unless a sync write is currently in
36  * progress.  In that case, the capsnap is said to be "pending", new
37  * writes cannot start, and the capsnap isn't "finalized" until the
38  * write completes (or fails) and a final size/mtime for the inode for
39  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
40  *
41  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
42  * we look for the first capsnap in i_cap_snaps and write out pages in
43  * that snap context _only_.  Then we move on to the next capsnap,
44  * eventually reaching the "live" or "head" context (i.e., pages that
45  * are not yet snapped) and are writing the most recently dirtied
46  * pages.
47  *
48  * Invalidate and so forth must take care to ensure the dirty page
49  * accounting is preserved.
50  */
51
52 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
53 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
54         (CONGESTION_ON_THRESH(congestion_kb) -                          \
55          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
56
57
58
59 /*
60  * Dirty a page.  Optimistically adjust accounting, on the assumption
61  * that we won't race with invalidate.  If we do, readjust.
62  */
63 static int ceph_set_page_dirty(struct page *page)
64 {
65         struct address_space *mapping = page->mapping;
66         struct inode *inode;
67         struct ceph_inode_info *ci;
68         int undo = 0;
69         struct ceph_snap_context *snapc;
70
71         if (unlikely(!mapping))
72                 return !TestSetPageDirty(page);
73
74         if (TestSetPageDirty(page)) {
75                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
76                      mapping->host, page, page->index);
77                 return 0;
78         }
79
80         inode = mapping->host;
81         ci = ceph_inode(inode);
82
83         /*
84          * Note that we're grabbing a snapc ref here without holding
85          * any locks!
86          */
87         snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
88
89         /* dirty the head */
90         spin_lock(&inode->i_lock);
91         if (ci->i_head_snapc == NULL)
92                 ci->i_head_snapc = ceph_get_snap_context(snapc);
93         ++ci->i_wrbuffer_ref_head;
94         if (ci->i_wrbuffer_ref == 0)
95                 igrab(inode);
96         ++ci->i_wrbuffer_ref;
97         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
98              "snapc %p seq %lld (%d snaps)\n",
99              mapping->host, page, page->index,
100              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
101              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
102              snapc, snapc->seq, snapc->num_snaps);
103         spin_unlock(&inode->i_lock);
104
105         /* now adjust page */
106         spin_lock_irq(&mapping->tree_lock);
107         if (page->mapping) {    /* Race with truncate? */
108                 WARN_ON_ONCE(!PageUptodate(page));
109                 account_page_dirtied(page, page->mapping);
110                 radix_tree_tag_set(&mapping->page_tree,
111                                 page_index(page), PAGECACHE_TAG_DIRTY);
112
113                 /*
114                  * Reference snap context in page->private.  Also set
115                  * PagePrivate so that we get invalidatepage callback.
116                  */
117                 page->private = (unsigned long)snapc;
118                 SetPagePrivate(page);
119         } else {
120                 dout("ANON set_page_dirty %p (raced truncate?)\n", page);
121                 undo = 1;
122         }
123
124         spin_unlock_irq(&mapping->tree_lock);
125
126         if (undo)
127                 /* whoops, we failed to dirty the page */
128                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
129
130         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
131
132         BUG_ON(!PageDirty(page));
133         return 1;
134 }
135
136 /*
137  * If we are truncating the full page (i.e. offset == 0), adjust the
138  * dirty page counters appropriately.  Only called if there is private
139  * data on the page.
140  */
141 static void ceph_invalidatepage(struct page *page, unsigned long offset)
142 {
143         struct inode *inode;
144         struct ceph_inode_info *ci;
145         struct ceph_snap_context *snapc = (void *)page->private;
146
147         BUG_ON(!PageLocked(page));
148         BUG_ON(!page->private);
149         BUG_ON(!PagePrivate(page));
150         BUG_ON(!page->mapping);
151
152         inode = page->mapping->host;
153
154         /*
155          * We can get non-dirty pages here due to races between
156          * set_page_dirty and truncate_complete_page; just spit out a
157          * warning, in case we end up with accounting problems later.
158          */
159         if (!PageDirty(page))
160                 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
161
162         if (offset == 0)
163                 ClearPageChecked(page);
164
165         ci = ceph_inode(inode);
166         if (offset == 0) {
167                 dout("%p invalidatepage %p idx %lu full dirty page %lu\n",
168                      inode, page, page->index, offset);
169                 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
170                 ceph_put_snap_context(snapc);
171                 page->private = 0;
172                 ClearPagePrivate(page);
173         } else {
174                 dout("%p invalidatepage %p idx %lu partial dirty page\n",
175                      inode, page, page->index);
176         }
177 }
178
179 /* just a sanity check */
180 static int ceph_releasepage(struct page *page, gfp_t g)
181 {
182         struct inode *inode = page->mapping ? page->mapping->host : NULL;
183         dout("%p releasepage %p idx %lu\n", inode, page, page->index);
184         WARN_ON(PageDirty(page));
185         WARN_ON(page->private);
186         WARN_ON(PagePrivate(page));
187         return 0;
188 }
189
190 /*
191  * read a single page, without unlocking it.
192  */
193 static int readpage_nounlock(struct file *filp, struct page *page)
194 {
195         struct inode *inode = filp->f_dentry->d_inode;
196         struct ceph_inode_info *ci = ceph_inode(inode);
197         struct ceph_osd_client *osdc = 
198                 &ceph_inode_to_client(inode)->client->osdc;
199         int err = 0;
200         u64 len = PAGE_CACHE_SIZE;
201
202         dout("readpage inode %p file %p page %p index %lu\n",
203              inode, filp, page, page->index);
204         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
205                                   page->index << PAGE_CACHE_SHIFT, &len,
206                                   ci->i_truncate_seq, ci->i_truncate_size,
207                                   &page, 1);
208         if (err == -ENOENT)
209                 err = 0;
210         if (err < 0) {
211                 SetPageError(page);
212                 goto out;
213         } else if (err < PAGE_CACHE_SIZE) {
214                 /* zero fill remainder of page */
215                 zero_user_segment(page, err, PAGE_CACHE_SIZE);
216         }
217         SetPageUptodate(page);
218
219 out:
220         return err < 0 ? err : 0;
221 }
222
223 static int ceph_readpage(struct file *filp, struct page *page)
224 {
225         int r = readpage_nounlock(filp, page);
226         unlock_page(page);
227         return r;
228 }
229
230 /*
231  * Build a vector of contiguous pages from the provided page list.
232  */
233 static struct page **page_vector_from_list(struct list_head *page_list,
234                                            unsigned *nr_pages)
235 {
236         struct page **pages;
237         struct page *page;
238         int next_index, contig_pages = 0;
239
240         /* build page vector */
241         pages = kmalloc(sizeof(*pages) * *nr_pages, GFP_NOFS);
242         if (!pages)
243                 return ERR_PTR(-ENOMEM);
244
245         BUG_ON(list_empty(page_list));
246         next_index = list_entry(page_list->prev, struct page, lru)->index;
247         list_for_each_entry_reverse(page, page_list, lru) {
248                 if (page->index == next_index) {
249                         dout("readpages page %d %p\n", contig_pages, page);
250                         pages[contig_pages] = page;
251                         contig_pages++;
252                         next_index++;
253                 } else {
254                         break;
255                 }
256         }
257         *nr_pages = contig_pages;
258         return pages;
259 }
260
261 /*
262  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
263  * the caller (VM) cleans them up.
264  */
265 static int ceph_readpages(struct file *file, struct address_space *mapping,
266                           struct list_head *page_list, unsigned nr_pages)
267 {
268         struct inode *inode = file->f_dentry->d_inode;
269         struct ceph_inode_info *ci = ceph_inode(inode);
270         struct ceph_osd_client *osdc =
271                 &ceph_inode_to_client(inode)->client->osdc;
272         int rc = 0;
273         struct page **pages;
274         loff_t offset;
275         u64 len;
276
277         dout("readpages %p file %p nr_pages %d\n",
278              inode, file, nr_pages);
279
280         pages = page_vector_from_list(page_list, &nr_pages);
281         if (IS_ERR(pages))
282                 return PTR_ERR(pages);
283
284         /* guess read extent */
285         offset = pages[0]->index << PAGE_CACHE_SHIFT;
286         len = nr_pages << PAGE_CACHE_SHIFT;
287         rc = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
288                                  offset, &len,
289                                  ci->i_truncate_seq, ci->i_truncate_size,
290                                  pages, nr_pages);
291         if (rc == -ENOENT)
292                 rc = 0;
293         if (rc < 0)
294                 goto out;
295
296         for (; !list_empty(page_list) && len > 0;
297              rc -= PAGE_CACHE_SIZE, len -= PAGE_CACHE_SIZE) {
298                 struct page *page =
299                         list_entry(page_list->prev, struct page, lru);
300
301                 list_del(&page->lru);
302
303                 if (rc < (int)PAGE_CACHE_SIZE) {
304                         /* zero (remainder of) page */
305                         int s = rc < 0 ? 0 : rc;
306                         zero_user_segment(page, s, PAGE_CACHE_SIZE);
307                 }
308
309                 if (add_to_page_cache_lru(page, mapping, page->index,
310                                           GFP_NOFS)) {
311                         page_cache_release(page);
312                         dout("readpages %p add_to_page_cache failed %p\n",
313                              inode, page);
314                         continue;
315                 }
316                 dout("readpages %p adding %p idx %lu\n", inode, page,
317                      page->index);
318                 flush_dcache_page(page);
319                 SetPageUptodate(page);
320                 unlock_page(page);
321                 page_cache_release(page);
322         }
323         rc = 0;
324
325 out:
326         kfree(pages);
327         return rc;
328 }
329
330 /*
331  * Get ref for the oldest snapc for an inode with dirty data... that is, the
332  * only snap context we are allowed to write back.
333  */
334 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
335                                                     u64 *snap_size)
336 {
337         struct ceph_inode_info *ci = ceph_inode(inode);
338         struct ceph_snap_context *snapc = NULL;
339         struct ceph_cap_snap *capsnap = NULL;
340
341         spin_lock(&inode->i_lock);
342         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
343                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
344                      capsnap->context, capsnap->dirty_pages);
345                 if (capsnap->dirty_pages) {
346                         snapc = ceph_get_snap_context(capsnap->context);
347                         if (snap_size)
348                                 *snap_size = capsnap->size;
349                         break;
350                 }
351         }
352         if (!snapc && ci->i_wrbuffer_ref_head) {
353                 snapc = ceph_get_snap_context(ci->i_head_snapc);
354                 dout(" head snapc %p has %d dirty pages\n",
355                      snapc, ci->i_wrbuffer_ref_head);
356         }
357         spin_unlock(&inode->i_lock);
358         return snapc;
359 }
360
361 /*
362  * Write a single page, but leave the page locked.
363  *
364  * If we get a write error, set the page error bit, but still adjust the
365  * dirty page accounting (i.e., page is no longer dirty).
366  */
367 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
368 {
369         struct inode *inode;
370         struct ceph_inode_info *ci;
371         struct ceph_fs_client *fsc;
372         struct ceph_osd_client *osdc;
373         loff_t page_off = page->index << PAGE_CACHE_SHIFT;
374         int len = PAGE_CACHE_SIZE;
375         loff_t i_size;
376         int err = 0;
377         struct ceph_snap_context *snapc, *oldest;
378         u64 snap_size = 0;
379         long writeback_stat;
380
381         dout("writepage %p idx %lu\n", page, page->index);
382
383         if (!page->mapping || !page->mapping->host) {
384                 dout("writepage %p - no mapping\n", page);
385                 return -EFAULT;
386         }
387         inode = page->mapping->host;
388         ci = ceph_inode(inode);
389         fsc = ceph_inode_to_client(inode);
390         osdc = &fsc->client->osdc;
391
392         /* verify this is a writeable snap context */
393         snapc = (void *)page->private;
394         if (snapc == NULL) {
395                 dout("writepage %p page %p not dirty?\n", inode, page);
396                 goto out;
397         }
398         oldest = get_oldest_context(inode, &snap_size);
399         if (snapc->seq > oldest->seq) {
400                 dout("writepage %p page %p snapc %p not writeable - noop\n",
401                      inode, page, (void *)page->private);
402                 /* we should only noop if called by kswapd */
403                 WARN_ON((current->flags & PF_MEMALLOC) == 0);
404                 ceph_put_snap_context(oldest);
405                 goto out;
406         }
407         ceph_put_snap_context(oldest);
408
409         /* is this a partial page at end of file? */
410         if (snap_size)
411                 i_size = snap_size;
412         else
413                 i_size = i_size_read(inode);
414         if (i_size < page_off + len)
415                 len = i_size - page_off;
416
417         dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
418              inode, page, page->index, page_off, len, snapc);
419
420         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
421         if (writeback_stat >
422             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
423                 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
424
425         set_page_writeback(page);
426         err = ceph_osdc_writepages(osdc, ceph_vino(inode),
427                                    &ci->i_layout, snapc,
428                                    page_off, len,
429                                    ci->i_truncate_seq, ci->i_truncate_size,
430                                    &inode->i_mtime,
431                                    &page, 1, 0, 0, true);
432         if (err < 0) {
433                 dout("writepage setting page/mapping error %d %p\n", err, page);
434                 SetPageError(page);
435                 mapping_set_error(&inode->i_data, err);
436                 if (wbc)
437                         wbc->pages_skipped++;
438         } else {
439                 dout("writepage cleaned page %p\n", page);
440                 err = 0;  /* vfs expects us to return 0 */
441         }
442         page->private = 0;
443         ClearPagePrivate(page);
444         end_page_writeback(page);
445         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
446         ceph_put_snap_context(snapc);  /* page's reference */
447 out:
448         return err;
449 }
450
451 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
452 {
453         int err;
454         struct inode *inode = page->mapping->host;
455         BUG_ON(!inode);
456         igrab(inode);
457         err = writepage_nounlock(page, wbc);
458         unlock_page(page);
459         iput(inode);
460         return err;
461 }
462
463
464 /*
465  * lame release_pages helper.  release_pages() isn't exported to
466  * modules.
467  */
468 static void ceph_release_pages(struct page **pages, int num)
469 {
470         struct pagevec pvec;
471         int i;
472
473         pagevec_init(&pvec, 0);
474         for (i = 0; i < num; i++) {
475                 if (pagevec_add(&pvec, pages[i]) == 0)
476                         pagevec_release(&pvec);
477         }
478         pagevec_release(&pvec);
479 }
480
481
482 /*
483  * async writeback completion handler.
484  *
485  * If we get an error, set the mapping error bit, but not the individual
486  * page error bits.
487  */
488 static void writepages_finish(struct ceph_osd_request *req,
489                               struct ceph_msg *msg)
490 {
491         struct inode *inode = req->r_inode;
492         struct ceph_osd_reply_head *replyhead;
493         struct ceph_osd_op *op;
494         struct ceph_inode_info *ci = ceph_inode(inode);
495         unsigned wrote;
496         struct page *page;
497         int i;
498         struct ceph_snap_context *snapc = req->r_snapc;
499         struct address_space *mapping = inode->i_mapping;
500         __s32 rc = -EIO;
501         u64 bytes = 0;
502         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
503         long writeback_stat;
504         unsigned issued = ceph_caps_issued(ci);
505
506         /* parse reply */
507         replyhead = msg->front.iov_base;
508         WARN_ON(le32_to_cpu(replyhead->num_ops) == 0);
509         op = (void *)(replyhead + 1);
510         rc = le32_to_cpu(replyhead->result);
511         bytes = le64_to_cpu(op->extent.length);
512
513         if (rc >= 0) {
514                 /*
515                  * Assume we wrote the pages we originally sent.  The
516                  * osd might reply with fewer pages if our writeback
517                  * raced with a truncation and was adjusted at the osd,
518                  * so don't believe the reply.
519                  */
520                 wrote = req->r_num_pages;
521         } else {
522                 wrote = 0;
523                 mapping_set_error(mapping, rc);
524         }
525         dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
526              inode, rc, bytes, wrote);
527
528         /* clean all pages */
529         for (i = 0; i < req->r_num_pages; i++) {
530                 page = req->r_pages[i];
531                 BUG_ON(!page);
532                 WARN_ON(!PageUptodate(page));
533
534                 writeback_stat =
535                         atomic_long_dec_return(&fsc->writeback_count);
536                 if (writeback_stat <
537                     CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
538                         clear_bdi_congested(&fsc->backing_dev_info,
539                                             BLK_RW_ASYNC);
540
541                 ceph_put_snap_context((void *)page->private);
542                 page->private = 0;
543                 ClearPagePrivate(page);
544                 dout("unlocking %d %p\n", i, page);
545                 end_page_writeback(page);
546
547                 /*
548                  * We lost the cache cap, need to truncate the page before
549                  * it is unlocked, otherwise we'd truncate it later in the
550                  * page truncation thread, possibly losing some data that
551                  * raced its way in
552                  */
553                 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
554                         generic_error_remove_page(inode->i_mapping, page);
555
556                 unlock_page(page);
557         }
558         dout("%p wrote+cleaned %d pages\n", inode, wrote);
559         ceph_put_wrbuffer_cap_refs(ci, req->r_num_pages, snapc);
560
561         ceph_release_pages(req->r_pages, req->r_num_pages);
562         if (req->r_pages_from_pool)
563                 mempool_free(req->r_pages,
564                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
565         else
566                 kfree(req->r_pages);
567         ceph_osdc_put_request(req);
568 }
569
570 /*
571  * allocate a page vec, either directly, or if necessary, via a the
572  * mempool.  we avoid the mempool if we can because req->r_num_pages
573  * may be less than the maximum write size.
574  */
575 static void alloc_page_vec(struct ceph_fs_client *fsc,
576                            struct ceph_osd_request *req)
577 {
578         req->r_pages = kmalloc(sizeof(struct page *) * req->r_num_pages,
579                                GFP_NOFS);
580         if (!req->r_pages) {
581                 req->r_pages = mempool_alloc(fsc->wb_pagevec_pool, GFP_NOFS);
582                 req->r_pages_from_pool = 1;
583                 WARN_ON(!req->r_pages);
584         }
585 }
586
587 /*
588  * initiate async writeback
589  */
590 static int ceph_writepages_start(struct address_space *mapping,
591                                  struct writeback_control *wbc)
592 {
593         struct inode *inode = mapping->host;
594         struct backing_dev_info *bdi = mapping->backing_dev_info;
595         struct ceph_inode_info *ci = ceph_inode(inode);
596         struct ceph_fs_client *fsc;
597         pgoff_t index, start, end;
598         int range_whole = 0;
599         int should_loop = 1;
600         pgoff_t max_pages = 0, max_pages_ever = 0;
601         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
602         struct pagevec pvec;
603         int done = 0;
604         int rc = 0;
605         unsigned wsize = 1 << inode->i_blkbits;
606         struct ceph_osd_request *req = NULL;
607         int do_sync;
608         u64 snap_size = 0;
609
610         /*
611          * Include a 'sync' in the OSD request if this is a data
612          * integrity write (e.g., O_SYNC write or fsync()), or if our
613          * cap is being revoked.
614          */
615         do_sync = wbc->sync_mode == WB_SYNC_ALL;
616         if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
617                 do_sync = 1;
618         dout("writepages_start %p dosync=%d (mode=%s)\n",
619              inode, do_sync,
620              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
621              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
622
623         fsc = ceph_inode_to_client(inode);
624         if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
625                 pr_warning("writepage_start %p on forced umount\n", inode);
626                 return -EIO; /* we're in a forced umount, don't write! */
627         }
628         if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
629                 wsize = fsc->mount_options->wsize;
630         if (wsize < PAGE_CACHE_SIZE)
631                 wsize = PAGE_CACHE_SIZE;
632         max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
633
634         pagevec_init(&pvec, 0);
635
636         /* ?? */
637         if (wbc->nonblocking && bdi_write_congested(bdi)) {
638                 dout(" writepages congested\n");
639                 wbc->encountered_congestion = 1;
640                 goto out_final;
641         }
642
643         /* where to start/end? */
644         if (wbc->range_cyclic) {
645                 start = mapping->writeback_index; /* Start from prev offset */
646                 end = -1;
647                 dout(" cyclic, start at %lu\n", start);
648         } else {
649                 start = wbc->range_start >> PAGE_CACHE_SHIFT;
650                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
651                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
652                         range_whole = 1;
653                 should_loop = 0;
654                 dout(" not cyclic, %lu to %lu\n", start, end);
655         }
656         index = start;
657
658 retry:
659         /* find oldest snap context with dirty data */
660         ceph_put_snap_context(snapc);
661         snapc = get_oldest_context(inode, &snap_size);
662         if (!snapc) {
663                 /* hmm, why does writepages get called when there
664                    is no dirty data? */
665                 dout(" no snap context with dirty data?\n");
666                 goto out;
667         }
668         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
669              snapc, snapc->seq, snapc->num_snaps);
670         if (last_snapc && snapc != last_snapc) {
671                 /* if we switched to a newer snapc, restart our scan at the
672                  * start of the original file range. */
673                 dout("  snapc differs from last pass, restarting at %lu\n",
674                      index);
675                 index = start;
676         }
677         last_snapc = snapc;
678
679         while (!done && index <= end) {
680                 unsigned i;
681                 int first;
682                 pgoff_t next;
683                 int pvec_pages, locked_pages;
684                 struct page *page;
685                 int want;
686                 u64 offset, len;
687                 struct ceph_osd_request_head *reqhead;
688                 struct ceph_osd_op *op;
689                 long writeback_stat;
690
691                 next = 0;
692                 locked_pages = 0;
693                 max_pages = max_pages_ever;
694
695 get_more_pages:
696                 first = -1;
697                 want = min(end - index,
698                            min((pgoff_t)PAGEVEC_SIZE,
699                                max_pages - (pgoff_t)locked_pages) - 1)
700                         + 1;
701                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
702                                                 PAGECACHE_TAG_DIRTY,
703                                                 want);
704                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
705                 if (!pvec_pages && !locked_pages)
706                         break;
707                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
708                         page = pvec.pages[i];
709                         dout("? %p idx %lu\n", page, page->index);
710                         if (locked_pages == 0)
711                                 lock_page(page);  /* first page */
712                         else if (!trylock_page(page))
713                                 break;
714
715                         /* only dirty pages, or our accounting breaks */
716                         if (unlikely(!PageDirty(page)) ||
717                             unlikely(page->mapping != mapping)) {
718                                 dout("!dirty or !mapping %p\n", page);
719                                 unlock_page(page);
720                                 break;
721                         }
722                         if (!wbc->range_cyclic && page->index > end) {
723                                 dout("end of range %p\n", page);
724                                 done = 1;
725                                 unlock_page(page);
726                                 break;
727                         }
728                         if (next && (page->index != next)) {
729                                 dout("not consecutive %p\n", page);
730                                 unlock_page(page);
731                                 break;
732                         }
733                         if (wbc->sync_mode != WB_SYNC_NONE) {
734                                 dout("waiting on writeback %p\n", page);
735                                 wait_on_page_writeback(page);
736                         }
737                         if ((snap_size && page_offset(page) > snap_size) ||
738                             (!snap_size &&
739                              page_offset(page) > i_size_read(inode))) {
740                                 dout("%p page eof %llu\n", page, snap_size ?
741                                      snap_size : i_size_read(inode));
742                                 done = 1;
743                                 unlock_page(page);
744                                 break;
745                         }
746                         if (PageWriteback(page)) {
747                                 dout("%p under writeback\n", page);
748                                 unlock_page(page);
749                                 break;
750                         }
751
752                         /* only if matching snap context */
753                         pgsnapc = (void *)page->private;
754                         if (pgsnapc->seq > snapc->seq) {
755                                 dout("page snapc %p %lld > oldest %p %lld\n",
756                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
757                                 unlock_page(page);
758                                 if (!locked_pages)
759                                         continue; /* keep looking for snap */
760                                 break;
761                         }
762
763                         if (!clear_page_dirty_for_io(page)) {
764                                 dout("%p !clear_page_dirty_for_io\n", page);
765                                 unlock_page(page);
766                                 break;
767                         }
768
769                         /* ok */
770                         if (locked_pages == 0) {
771                                 /* prepare async write request */
772                                 offset = (unsigned long long)page->index
773                                         << PAGE_CACHE_SHIFT;
774                                 len = wsize;
775                                 req = ceph_osdc_new_request(&fsc->client->osdc,
776                                             &ci->i_layout,
777                                             ceph_vino(inode),
778                                             offset, &len,
779                                             CEPH_OSD_OP_WRITE,
780                                             CEPH_OSD_FLAG_WRITE |
781                                                     CEPH_OSD_FLAG_ONDISK,
782                                             snapc, do_sync,
783                                             ci->i_truncate_seq,
784                                             ci->i_truncate_size,
785                                             &inode->i_mtime, true, 1);
786                                 max_pages = req->r_num_pages;
787
788                                 alloc_page_vec(fsc, req);
789                                 req->r_callback = writepages_finish;
790                                 req->r_inode = inode;
791                         }
792
793                         /* note position of first page in pvec */
794                         if (first < 0)
795                                 first = i;
796                         dout("%p will write page %p idx %lu\n",
797                              inode, page, page->index);
798
799                         writeback_stat =
800                                atomic_long_inc_return(&fsc->writeback_count);
801                         if (writeback_stat > CONGESTION_ON_THRESH(
802                                     fsc->mount_options->congestion_kb)) {
803                                 set_bdi_congested(&fsc->backing_dev_info,
804                                                   BLK_RW_ASYNC);
805                         }
806
807                         set_page_writeback(page);
808                         req->r_pages[locked_pages] = page;
809                         locked_pages++;
810                         next = page->index + 1;
811                 }
812
813                 /* did we get anything? */
814                 if (!locked_pages)
815                         goto release_pvec_pages;
816                 if (i) {
817                         int j;
818                         BUG_ON(!locked_pages || first < 0);
819
820                         if (pvec_pages && i == pvec_pages &&
821                             locked_pages < max_pages) {
822                                 dout("reached end pvec, trying for more\n");
823                                 pagevec_reinit(&pvec);
824                                 goto get_more_pages;
825                         }
826
827                         /* shift unused pages over in the pvec...  we
828                          * will need to release them below. */
829                         for (j = i; j < pvec_pages; j++) {
830                                 dout(" pvec leftover page %p\n",
831                                      pvec.pages[j]);
832                                 pvec.pages[j-i+first] = pvec.pages[j];
833                         }
834                         pvec.nr -= i-first;
835                 }
836
837                 /* submit the write */
838                 offset = req->r_pages[0]->index << PAGE_CACHE_SHIFT;
839                 len = min((snap_size ? snap_size : i_size_read(inode)) - offset,
840                           (u64)locked_pages << PAGE_CACHE_SHIFT);
841                 dout("writepages got %d pages at %llu~%llu\n",
842                      locked_pages, offset, len);
843
844                 /* revise final length, page count */
845                 req->r_num_pages = locked_pages;
846                 reqhead = req->r_request->front.iov_base;
847                 op = (void *)(reqhead + 1);
848                 op->extent.length = cpu_to_le64(len);
849                 op->payload_len = cpu_to_le32(len);
850                 req->r_request->hdr.data_len = cpu_to_le32(len);
851
852                 ceph_osdc_start_request(&fsc->client->osdc, req, true);
853                 req = NULL;
854
855                 /* continue? */
856                 index = next;
857                 wbc->nr_to_write -= locked_pages;
858                 if (wbc->nr_to_write <= 0)
859                         done = 1;
860
861 release_pvec_pages:
862                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
863                      pvec.nr ? pvec.pages[0] : NULL);
864                 pagevec_release(&pvec);
865
866                 if (locked_pages && !done)
867                         goto retry;
868         }
869
870         if (should_loop && !done) {
871                 /* more to do; loop back to beginning of file */
872                 dout("writepages looping back to beginning of file\n");
873                 should_loop = 0;
874                 index = 0;
875                 goto retry;
876         }
877
878         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
879                 mapping->writeback_index = index;
880
881 out:
882         if (req)
883                 ceph_osdc_put_request(req);
884         if (rc > 0)
885                 rc = 0;  /* vfs expects us to return 0 */
886         ceph_put_snap_context(snapc);
887         dout("writepages done, rc = %d\n", rc);
888 out_final:
889         return rc;
890 }
891
892
893
894 /*
895  * See if a given @snapc is either writeable, or already written.
896  */
897 static int context_is_writeable_or_written(struct inode *inode,
898                                            struct ceph_snap_context *snapc)
899 {
900         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
901         int ret = !oldest || snapc->seq <= oldest->seq;
902
903         ceph_put_snap_context(oldest);
904         return ret;
905 }
906
907 /*
908  * We are only allowed to write into/dirty the page if the page is
909  * clean, or already dirty within the same snap context.
910  *
911  * called with page locked.
912  * return success with page locked,
913  * or any failure (incl -EAGAIN) with page unlocked.
914  */
915 static int ceph_update_writeable_page(struct file *file,
916                             loff_t pos, unsigned len,
917                             struct page *page)
918 {
919         struct inode *inode = file->f_dentry->d_inode;
920         struct ceph_inode_info *ci = ceph_inode(inode);
921         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
922         loff_t page_off = pos & PAGE_CACHE_MASK;
923         int pos_in_page = pos & ~PAGE_CACHE_MASK;
924         int end_in_page = pos_in_page + len;
925         loff_t i_size;
926         int r;
927         struct ceph_snap_context *snapc, *oldest;
928
929 retry_locked:
930         /* writepages currently holds page lock, but if we change that later, */
931         wait_on_page_writeback(page);
932
933         /* check snap context */
934         BUG_ON(!ci->i_snap_realm);
935         down_read(&mdsc->snap_rwsem);
936         BUG_ON(!ci->i_snap_realm->cached_context);
937         snapc = (void *)page->private;
938         if (snapc && snapc != ci->i_head_snapc) {
939                 /*
940                  * this page is already dirty in another (older) snap
941                  * context!  is it writeable now?
942                  */
943                 oldest = get_oldest_context(inode, NULL);
944                 up_read(&mdsc->snap_rwsem);
945
946                 if (snapc->seq > oldest->seq) {
947                         ceph_put_snap_context(oldest);
948                         dout(" page %p snapc %p not current or oldest\n",
949                              page, snapc);
950                         /*
951                          * queue for writeback, and wait for snapc to
952                          * be writeable or written
953                          */
954                         snapc = ceph_get_snap_context(snapc);
955                         unlock_page(page);
956                         ceph_queue_writeback(inode);
957                         r = wait_event_interruptible(ci->i_cap_wq,
958                                context_is_writeable_or_written(inode, snapc));
959                         ceph_put_snap_context(snapc);
960                         if (r == -ERESTARTSYS)
961                                 return r;
962                         return -EAGAIN;
963                 }
964                 ceph_put_snap_context(oldest);
965
966                 /* yay, writeable, do it now (without dropping page lock) */
967                 dout(" page %p snapc %p not current, but oldest\n",
968                      page, snapc);
969                 if (!clear_page_dirty_for_io(page))
970                         goto retry_locked;
971                 r = writepage_nounlock(page, NULL);
972                 if (r < 0)
973                         goto fail_nosnap;
974                 goto retry_locked;
975         }
976
977         if (PageUptodate(page)) {
978                 dout(" page %p already uptodate\n", page);
979                 return 0;
980         }
981
982         /* full page? */
983         if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
984                 return 0;
985
986         /* past end of file? */
987         i_size = inode->i_size;   /* caller holds i_mutex */
988
989         if (i_size + len > inode->i_sb->s_maxbytes) {
990                 /* file is too big */
991                 r = -EINVAL;
992                 goto fail;
993         }
994
995         if (page_off >= i_size ||
996             (pos_in_page == 0 && (pos+len) >= i_size &&
997              end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
998                 dout(" zeroing %p 0 - %d and %d - %d\n",
999                      page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1000                 zero_user_segments(page,
1001                                    0, pos_in_page,
1002                                    end_in_page, PAGE_CACHE_SIZE);
1003                 return 0;
1004         }
1005
1006         /* we need to read it. */
1007         up_read(&mdsc->snap_rwsem);
1008         r = readpage_nounlock(file, page);
1009         if (r < 0)
1010                 goto fail_nosnap;
1011         goto retry_locked;
1012
1013 fail:
1014         up_read(&mdsc->snap_rwsem);
1015 fail_nosnap:
1016         unlock_page(page);
1017         return r;
1018 }
1019
1020 /*
1021  * We are only allowed to write into/dirty the page if the page is
1022  * clean, or already dirty within the same snap context.
1023  */
1024 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1025                             loff_t pos, unsigned len, unsigned flags,
1026                             struct page **pagep, void **fsdata)
1027 {
1028         struct inode *inode = file->f_dentry->d_inode;
1029         struct page *page;
1030         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1031         int r;
1032
1033         do {
1034                 /* get a page */
1035                 page = grab_cache_page_write_begin(mapping, index, 0);
1036                 if (!page)
1037                         return -ENOMEM;
1038                 *pagep = page;
1039
1040                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1041                      inode, page, (int)pos, (int)len);
1042
1043                 r = ceph_update_writeable_page(file, pos, len, page);
1044         } while (r == -EAGAIN);
1045
1046         return r;
1047 }
1048
1049 /*
1050  * we don't do anything in here that simple_write_end doesn't do
1051  * except adjust dirty page accounting and drop read lock on
1052  * mdsc->snap_rwsem.
1053  */
1054 static int ceph_write_end(struct file *file, struct address_space *mapping,
1055                           loff_t pos, unsigned len, unsigned copied,
1056                           struct page *page, void *fsdata)
1057 {
1058         struct inode *inode = file->f_dentry->d_inode;
1059         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1060         struct ceph_mds_client *mdsc = fsc->mdsc;
1061         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1062         int check_cap = 0;
1063
1064         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1065              inode, page, (int)pos, (int)copied, (int)len);
1066
1067         /* zero the stale part of the page if we did a short copy */
1068         if (copied < len)
1069                 zero_user_segment(page, from+copied, len);
1070
1071         /* did file size increase? */
1072         /* (no need for i_size_read(); we caller holds i_mutex */
1073         if (pos+copied > inode->i_size)
1074                 check_cap = ceph_inode_set_size(inode, pos+copied);
1075
1076         if (!PageUptodate(page))
1077                 SetPageUptodate(page);
1078
1079         set_page_dirty(page);
1080
1081         unlock_page(page);
1082         up_read(&mdsc->snap_rwsem);
1083         page_cache_release(page);
1084
1085         if (check_cap)
1086                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1087
1088         return copied;
1089 }
1090
1091 /*
1092  * we set .direct_IO to indicate direct io is supported, but since we
1093  * intercept O_DIRECT reads and writes early, this function should
1094  * never get called.
1095  */
1096 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1097                               const struct iovec *iov,
1098                               loff_t pos, unsigned long nr_segs)
1099 {
1100         WARN_ON(1);
1101         return -EINVAL;
1102 }
1103
1104 const struct address_space_operations ceph_aops = {
1105         .readpage = ceph_readpage,
1106         .readpages = ceph_readpages,
1107         .writepage = ceph_writepage,
1108         .writepages = ceph_writepages_start,
1109         .write_begin = ceph_write_begin,
1110         .write_end = ceph_write_end,
1111         .set_page_dirty = ceph_set_page_dirty,
1112         .invalidatepage = ceph_invalidatepage,
1113         .releasepage = ceph_releasepage,
1114         .direct_IO = ceph_direct_io,
1115 };
1116
1117
1118 /*
1119  * vm ops
1120  */
1121
1122 /*
1123  * Reuse write_begin here for simplicity.
1124  */
1125 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1126 {
1127         struct inode *inode = vma->vm_file->f_dentry->d_inode;
1128         struct page *page = vmf->page;
1129         struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1130         loff_t off = page->index << PAGE_CACHE_SHIFT;
1131         loff_t size, len;
1132         int ret;
1133
1134         size = i_size_read(inode);
1135         if (off + PAGE_CACHE_SIZE <= size)
1136                 len = PAGE_CACHE_SIZE;
1137         else
1138                 len = size & ~PAGE_CACHE_MASK;
1139
1140         dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1141              off, len, page, page->index);
1142
1143         lock_page(page);
1144
1145         ret = VM_FAULT_NOPAGE;
1146         if ((off > size) ||
1147             (page->mapping != inode->i_mapping))
1148                 goto out;
1149
1150         ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1151         if (ret == 0) {
1152                 /* success.  we'll keep the page locked. */
1153                 set_page_dirty(page);
1154                 up_read(&mdsc->snap_rwsem);
1155                 ret = VM_FAULT_LOCKED;
1156         } else {
1157                 if (ret == -ENOMEM)
1158                         ret = VM_FAULT_OOM;
1159                 else
1160                         ret = VM_FAULT_SIGBUS;
1161         }
1162 out:
1163         dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1164         if (ret != VM_FAULT_LOCKED)
1165                 unlock_page(page);
1166         return ret;
1167 }
1168
1169 static struct vm_operations_struct ceph_vmops = {
1170         .fault          = filemap_fault,
1171         .page_mkwrite   = ceph_page_mkwrite,
1172 };
1173
1174 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1175 {
1176         struct address_space *mapping = file->f_mapping;
1177
1178         if (!mapping->a_ops->readpage)
1179                 return -ENOEXEC;
1180         file_accessed(file);
1181         vma->vm_ops = &ceph_vmops;
1182         vma->vm_flags |= VM_CAN_NONLINEAR;
1183         return 0;
1184 }