4d622654bfbcc2eb10ce81d461c53a35e5f2c44d
[sfrench/cifs-2.6.git] / fs / ceph / addr.c
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
3
4 #include <linux/backing-dev.h>
5 #include <linux/fs.h>
6 #include <linux/mm.h>
7 #include <linux/pagemap.h>
8 #include <linux/writeback.h>    /* generic_writepages */
9 #include <linux/slab.h>
10 #include <linux/pagevec.h>
11 #include <linux/task_io_accounting_ops.h>
12 #include <linux/signal.h>
13
14 #include "super.h"
15 #include "mds_client.h"
16 #include "cache.h"
17 #include <linux/ceph/osd_client.h>
18
19 /*
20  * Ceph address space ops.
21  *
22  * There are a few funny things going on here.
23  *
24  * The page->private field is used to reference a struct
25  * ceph_snap_context for _every_ dirty page.  This indicates which
26  * snapshot the page was logically dirtied in, and thus which snap
27  * context needs to be associated with the osd write during writeback.
28  *
29  * Similarly, struct ceph_inode_info maintains a set of counters to
30  * count dirty pages on the inode.  In the absence of snapshots,
31  * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
32  *
33  * When a snapshot is taken (that is, when the client receives
34  * notification that a snapshot was taken), each inode with caps and
35  * with dirty pages (dirty pages implies there is a cap) gets a new
36  * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
37  * order, new snaps go to the tail).  The i_wrbuffer_ref_head count is
38  * moved to capsnap->dirty. (Unless a sync write is currently in
39  * progress.  In that case, the capsnap is said to be "pending", new
40  * writes cannot start, and the capsnap isn't "finalized" until the
41  * write completes (or fails) and a final size/mtime for the inode for
42  * that snap can be settled upon.)  i_wrbuffer_ref_head is reset to 0.
43  *
44  * On writeback, we must submit writes to the osd IN SNAP ORDER.  So,
45  * we look for the first capsnap in i_cap_snaps and write out pages in
46  * that snap context _only_.  Then we move on to the next capsnap,
47  * eventually reaching the "live" or "head" context (i.e., pages that
48  * are not yet snapped) and are writing the most recently dirtied
49  * pages.
50  *
51  * Invalidate and so forth must take care to ensure the dirty page
52  * accounting is preserved.
53  */
54
55 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
56 #define CONGESTION_OFF_THRESH(congestion_kb)                            \
57         (CONGESTION_ON_THRESH(congestion_kb) -                          \
58          (CONGESTION_ON_THRESH(congestion_kb) >> 2))
59
60 static inline struct ceph_snap_context *page_snap_context(struct page *page)
61 {
62         if (PagePrivate(page))
63                 return (void *)page->private;
64         return NULL;
65 }
66
67 /*
68  * Dirty a page.  Optimistically adjust accounting, on the assumption
69  * that we won't race with invalidate.  If we do, readjust.
70  */
71 static int ceph_set_page_dirty(struct page *page)
72 {
73         struct address_space *mapping = page->mapping;
74         struct inode *inode;
75         struct ceph_inode_info *ci;
76         struct ceph_snap_context *snapc;
77         int ret;
78
79         if (unlikely(!mapping))
80                 return !TestSetPageDirty(page);
81
82         if (PageDirty(page)) {
83                 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
84                      mapping->host, page, page->index);
85                 BUG_ON(!PagePrivate(page));
86                 return 0;
87         }
88
89         inode = mapping->host;
90         ci = ceph_inode(inode);
91
92         /* dirty the head */
93         spin_lock(&ci->i_ceph_lock);
94         BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
95         if (__ceph_have_pending_cap_snap(ci)) {
96                 struct ceph_cap_snap *capsnap =
97                                 list_last_entry(&ci->i_cap_snaps,
98                                                 struct ceph_cap_snap,
99                                                 ci_item);
100                 snapc = ceph_get_snap_context(capsnap->context);
101                 capsnap->dirty_pages++;
102         } else {
103                 BUG_ON(!ci->i_head_snapc);
104                 snapc = ceph_get_snap_context(ci->i_head_snapc);
105                 ++ci->i_wrbuffer_ref_head;
106         }
107         if (ci->i_wrbuffer_ref == 0)
108                 ihold(inode);
109         ++ci->i_wrbuffer_ref;
110         dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
111              "snapc %p seq %lld (%d snaps)\n",
112              mapping->host, page, page->index,
113              ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
114              ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
115              snapc, snapc->seq, snapc->num_snaps);
116         spin_unlock(&ci->i_ceph_lock);
117
118         /*
119          * Reference snap context in page->private.  Also set
120          * PagePrivate so that we get invalidatepage callback.
121          */
122         BUG_ON(PagePrivate(page));
123         page->private = (unsigned long)snapc;
124         SetPagePrivate(page);
125
126         ret = __set_page_dirty_nobuffers(page);
127         WARN_ON(!PageLocked(page));
128         WARN_ON(!page->mapping);
129
130         return ret;
131 }
132
133 /*
134  * If we are truncating the full page (i.e. offset == 0), adjust the
135  * dirty page counters appropriately.  Only called if there is private
136  * data on the page.
137  */
138 static void ceph_invalidatepage(struct page *page, unsigned int offset,
139                                 unsigned int length)
140 {
141         struct inode *inode;
142         struct ceph_inode_info *ci;
143         struct ceph_snap_context *snapc = page_snap_context(page);
144
145         inode = page->mapping->host;
146         ci = ceph_inode(inode);
147
148         if (offset != 0 || length != PAGE_SIZE) {
149                 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
150                      inode, page, page->index, offset, length);
151                 return;
152         }
153
154         ceph_invalidate_fscache_page(inode, page);
155
156         WARN_ON(!PageLocked(page));
157         if (!PagePrivate(page))
158                 return;
159
160         ClearPageChecked(page);
161
162         dout("%p invalidatepage %p idx %lu full dirty page\n",
163              inode, page, page->index);
164
165         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
166         ceph_put_snap_context(snapc);
167         page->private = 0;
168         ClearPagePrivate(page);
169 }
170
171 static int ceph_releasepage(struct page *page, gfp_t g)
172 {
173         dout("%p releasepage %p idx %lu (%sdirty)\n", page->mapping->host,
174              page, page->index, PageDirty(page) ? "" : "not ");
175
176         /* Can we release the page from the cache? */
177         if (!ceph_release_fscache_page(page, g))
178                 return 0;
179
180         return !PagePrivate(page);
181 }
182
183 /*
184  * read a single page, without unlocking it.
185  */
186 static int ceph_do_readpage(struct file *filp, struct page *page)
187 {
188         struct inode *inode = file_inode(filp);
189         struct ceph_inode_info *ci = ceph_inode(inode);
190         struct ceph_osd_client *osdc =
191                 &ceph_inode_to_client(inode)->client->osdc;
192         int err = 0;
193         u64 off = page_offset(page);
194         u64 len = PAGE_SIZE;
195
196         if (off >= i_size_read(inode)) {
197                 zero_user_segment(page, 0, PAGE_SIZE);
198                 SetPageUptodate(page);
199                 return 0;
200         }
201
202         if (ci->i_inline_version != CEPH_INLINE_NONE) {
203                 /*
204                  * Uptodate inline data should have been added
205                  * into page cache while getting Fcr caps.
206                  */
207                 if (off == 0)
208                         return -EINVAL;
209                 zero_user_segment(page, 0, PAGE_SIZE);
210                 SetPageUptodate(page);
211                 return 0;
212         }
213
214         err = ceph_readpage_from_fscache(inode, page);
215         if (err == 0)
216                 return -EINPROGRESS;
217
218         dout("readpage inode %p file %p page %p index %lu\n",
219              inode, filp, page, page->index);
220         err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
221                                   off, &len,
222                                   ci->i_truncate_seq, ci->i_truncate_size,
223                                   &page, 1, 0);
224         if (err == -ENOENT)
225                 err = 0;
226         if (err < 0) {
227                 SetPageError(page);
228                 ceph_fscache_readpage_cancel(inode, page);
229                 goto out;
230         }
231         if (err < PAGE_SIZE)
232                 /* zero fill remainder of page */
233                 zero_user_segment(page, err, PAGE_SIZE);
234         else
235                 flush_dcache_page(page);
236
237         SetPageUptodate(page);
238         ceph_readpage_to_fscache(inode, page);
239
240 out:
241         return err < 0 ? err : 0;
242 }
243
244 static int ceph_readpage(struct file *filp, struct page *page)
245 {
246         int r = ceph_do_readpage(filp, page);
247         if (r != -EINPROGRESS)
248                 unlock_page(page);
249         else
250                 r = 0;
251         return r;
252 }
253
254 /*
255  * Finish an async read(ahead) op.
256  */
257 static void finish_read(struct ceph_osd_request *req)
258 {
259         struct inode *inode = req->r_inode;
260         struct ceph_osd_data *osd_data;
261         int rc = req->r_result <= 0 ? req->r_result : 0;
262         int bytes = req->r_result >= 0 ? req->r_result : 0;
263         int num_pages;
264         int i;
265
266         dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
267
268         /* unlock all pages, zeroing any data we didn't read */
269         osd_data = osd_req_op_extent_osd_data(req, 0);
270         BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
271         num_pages = calc_pages_for((u64)osd_data->alignment,
272                                         (u64)osd_data->length);
273         for (i = 0; i < num_pages; i++) {
274                 struct page *page = osd_data->pages[i];
275
276                 if (rc < 0 && rc != -ENOENT) {
277                         ceph_fscache_readpage_cancel(inode, page);
278                         goto unlock;
279                 }
280                 if (bytes < (int)PAGE_SIZE) {
281                         /* zero (remainder of) page */
282                         int s = bytes < 0 ? 0 : bytes;
283                         zero_user_segment(page, s, PAGE_SIZE);
284                 }
285                 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
286                      page->index);
287                 flush_dcache_page(page);
288                 SetPageUptodate(page);
289                 ceph_readpage_to_fscache(inode, page);
290 unlock:
291                 unlock_page(page);
292                 put_page(page);
293                 bytes -= PAGE_SIZE;
294         }
295         kfree(osd_data->pages);
296 }
297
298 /*
299  * start an async read(ahead) operation.  return nr_pages we submitted
300  * a read for on success, or negative error code.
301  */
302 static int start_read(struct inode *inode, struct list_head *page_list, int max)
303 {
304         struct ceph_osd_client *osdc =
305                 &ceph_inode_to_client(inode)->client->osdc;
306         struct ceph_inode_info *ci = ceph_inode(inode);
307         struct page *page = list_entry(page_list->prev, struct page, lru);
308         struct ceph_vino vino;
309         struct ceph_osd_request *req;
310         u64 off;
311         u64 len;
312         int i;
313         struct page **pages;
314         pgoff_t next_index;
315         int nr_pages = 0;
316         int got = 0;
317         int ret = 0;
318
319         if (!current->journal_info) {
320                 /* caller of readpages does not hold buffer and read caps
321                  * (fadvise, madvise and readahead cases) */
322                 int want = CEPH_CAP_FILE_CACHE;
323                 ret = ceph_try_get_caps(ci, CEPH_CAP_FILE_RD, want, &got);
324                 if (ret < 0) {
325                         dout("start_read %p, error getting cap\n", inode);
326                 } else if (!(got & want)) {
327                         dout("start_read %p, no cache cap\n", inode);
328                         ret = 0;
329                 }
330                 if (ret <= 0) {
331                         if (got)
332                                 ceph_put_cap_refs(ci, got);
333                         while (!list_empty(page_list)) {
334                                 page = list_entry(page_list->prev,
335                                                   struct page, lru);
336                                 list_del(&page->lru);
337                                 put_page(page);
338                         }
339                         return ret;
340                 }
341         }
342
343         off = (u64) page_offset(page);
344
345         /* count pages */
346         next_index = page->index;
347         list_for_each_entry_reverse(page, page_list, lru) {
348                 if (page->index != next_index)
349                         break;
350                 nr_pages++;
351                 next_index++;
352                 if (max && nr_pages == max)
353                         break;
354         }
355         len = nr_pages << PAGE_SHIFT;
356         dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
357              off, len);
358         vino = ceph_vino(inode);
359         req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
360                                     0, 1, CEPH_OSD_OP_READ,
361                                     CEPH_OSD_FLAG_READ, NULL,
362                                     ci->i_truncate_seq, ci->i_truncate_size,
363                                     false);
364         if (IS_ERR(req)) {
365                 ret = PTR_ERR(req);
366                 goto out;
367         }
368
369         /* build page vector */
370         nr_pages = calc_pages_for(0, len);
371         pages = kmalloc(sizeof(*pages) * nr_pages, GFP_KERNEL);
372         if (!pages) {
373                 ret = -ENOMEM;
374                 goto out_put;
375         }
376         for (i = 0; i < nr_pages; ++i) {
377                 page = list_entry(page_list->prev, struct page, lru);
378                 BUG_ON(PageLocked(page));
379                 list_del(&page->lru);
380
381                 dout("start_read %p adding %p idx %lu\n", inode, page,
382                      page->index);
383                 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
384                                           GFP_KERNEL)) {
385                         ceph_fscache_uncache_page(inode, page);
386                         put_page(page);
387                         dout("start_read %p add_to_page_cache failed %p\n",
388                              inode, page);
389                         nr_pages = i;
390                         if (nr_pages > 0) {
391                                 len = nr_pages << PAGE_SHIFT;
392                                 osd_req_op_extent_update(req, 0, len);
393                                 break;
394                         }
395                         goto out_pages;
396                 }
397                 pages[i] = page;
398         }
399         osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
400         req->r_callback = finish_read;
401         req->r_inode = inode;
402
403         dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
404         ret = ceph_osdc_start_request(osdc, req, false);
405         if (ret < 0)
406                 goto out_pages;
407         ceph_osdc_put_request(req);
408
409         /* After adding locked pages to page cache, the inode holds cache cap.
410          * So we can drop our cap refs. */
411         if (got)
412                 ceph_put_cap_refs(ci, got);
413
414         return nr_pages;
415
416 out_pages:
417         for (i = 0; i < nr_pages; ++i) {
418                 ceph_fscache_readpage_cancel(inode, pages[i]);
419                 unlock_page(pages[i]);
420         }
421         ceph_put_page_vector(pages, nr_pages, false);
422 out_put:
423         ceph_osdc_put_request(req);
424 out:
425         if (got)
426                 ceph_put_cap_refs(ci, got);
427         return ret;
428 }
429
430
431 /*
432  * Read multiple pages.  Leave pages we don't read + unlock in page_list;
433  * the caller (VM) cleans them up.
434  */
435 static int ceph_readpages(struct file *file, struct address_space *mapping,
436                           struct list_head *page_list, unsigned nr_pages)
437 {
438         struct inode *inode = file_inode(file);
439         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
440         int rc = 0;
441         int max = 0;
442
443         if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
444                 return -EINVAL;
445
446         rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
447                                          &nr_pages);
448
449         if (rc == 0)
450                 goto out;
451
452         max = fsc->mount_options->rsize >> PAGE_SHIFT;
453         dout("readpages %p file %p nr_pages %d max %d\n",
454              inode, file, nr_pages, max);
455         while (!list_empty(page_list)) {
456                 rc = start_read(inode, page_list, max);
457                 if (rc < 0)
458                         goto out;
459         }
460 out:
461         ceph_fscache_readpages_cancel(inode, page_list);
462
463         dout("readpages %p file %p ret %d\n", inode, file, rc);
464         return rc;
465 }
466
467 struct ceph_writeback_ctl
468 {
469         loff_t i_size;
470         u64 truncate_size;
471         u32 truncate_seq;
472         bool size_stable;
473         bool head_snapc;
474 };
475
476 /*
477  * Get ref for the oldest snapc for an inode with dirty data... that is, the
478  * only snap context we are allowed to write back.
479  */
480 static struct ceph_snap_context *
481 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
482                    struct ceph_snap_context *page_snapc)
483 {
484         struct ceph_inode_info *ci = ceph_inode(inode);
485         struct ceph_snap_context *snapc = NULL;
486         struct ceph_cap_snap *capsnap = NULL;
487
488         spin_lock(&ci->i_ceph_lock);
489         list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
490                 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
491                      capsnap->context, capsnap->dirty_pages);
492                 if (!capsnap->dirty_pages)
493                         continue;
494
495                 /* get i_size, truncate_{seq,size} for page_snapc? */
496                 if (snapc && capsnap->context != page_snapc)
497                         continue;
498
499                 if (ctl) {
500                         if (capsnap->writing) {
501                                 ctl->i_size = i_size_read(inode);
502                                 ctl->size_stable = false;
503                         } else {
504                                 ctl->i_size = capsnap->size;
505                                 ctl->size_stable = true;
506                         }
507                         ctl->truncate_size = capsnap->truncate_size;
508                         ctl->truncate_seq = capsnap->truncate_seq;
509                         ctl->head_snapc = false;
510                 }
511
512                 if (snapc)
513                         break;
514
515                 snapc = ceph_get_snap_context(capsnap->context);
516                 if (!page_snapc ||
517                     page_snapc == snapc ||
518                     page_snapc->seq > snapc->seq)
519                         break;
520         }
521         if (!snapc && ci->i_wrbuffer_ref_head) {
522                 snapc = ceph_get_snap_context(ci->i_head_snapc);
523                 dout(" head snapc %p has %d dirty pages\n",
524                      snapc, ci->i_wrbuffer_ref_head);
525                 if (ctl) {
526                         ctl->i_size = i_size_read(inode);
527                         ctl->truncate_size = ci->i_truncate_size;
528                         ctl->truncate_seq = ci->i_truncate_seq;
529                         ctl->size_stable = false;
530                         ctl->head_snapc = true;
531                 }
532         }
533         spin_unlock(&ci->i_ceph_lock);
534         return snapc;
535 }
536
537 static u64 get_writepages_data_length(struct inode *inode,
538                                       struct page *page, u64 start)
539 {
540         struct ceph_inode_info *ci = ceph_inode(inode);
541         struct ceph_snap_context *snapc = page_snap_context(page);
542         struct ceph_cap_snap *capsnap = NULL;
543         u64 end = i_size_read(inode);
544
545         if (snapc != ci->i_head_snapc) {
546                 bool found = false;
547                 spin_lock(&ci->i_ceph_lock);
548                 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
549                         if (capsnap->context == snapc) {
550                                 if (!capsnap->writing)
551                                         end = capsnap->size;
552                                 found = true;
553                                 break;
554                         }
555                 }
556                 spin_unlock(&ci->i_ceph_lock);
557                 WARN_ON(!found);
558         }
559         if (end > page_offset(page) + PAGE_SIZE)
560                 end = page_offset(page) + PAGE_SIZE;
561         return end > start ? end - start : 0;
562 }
563
564 /*
565  * Write a single page, but leave the page locked.
566  *
567  * If we get a write error, set the page error bit, but still adjust the
568  * dirty page accounting (i.e., page is no longer dirty).
569  */
570 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
571 {
572         struct inode *inode;
573         struct ceph_inode_info *ci;
574         struct ceph_fs_client *fsc;
575         struct ceph_snap_context *snapc, *oldest;
576         loff_t page_off = page_offset(page);
577         long writeback_stat;
578         int err, len = PAGE_SIZE;
579         struct ceph_writeback_ctl ceph_wbc;
580
581         dout("writepage %p idx %lu\n", page, page->index);
582
583         inode = page->mapping->host;
584         ci = ceph_inode(inode);
585         fsc = ceph_inode_to_client(inode);
586
587         /* verify this is a writeable snap context */
588         snapc = page_snap_context(page);
589         if (!snapc) {
590                 dout("writepage %p page %p not dirty?\n", inode, page);
591                 return 0;
592         }
593         oldest = get_oldest_context(inode, &ceph_wbc, snapc);
594         if (snapc->seq > oldest->seq) {
595                 dout("writepage %p page %p snapc %p not writeable - noop\n",
596                      inode, page, snapc);
597                 /* we should only noop if called by kswapd */
598                 WARN_ON(!(current->flags & PF_MEMALLOC));
599                 ceph_put_snap_context(oldest);
600                 redirty_page_for_writepage(wbc, page);
601                 return 0;
602         }
603         ceph_put_snap_context(oldest);
604
605         /* is this a partial page at end of file? */
606         if (page_off >= ceph_wbc.i_size) {
607                 dout("%p page eof %llu\n", page, ceph_wbc.i_size);
608                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
609                 return 0;
610         }
611
612         if (ceph_wbc.i_size < page_off + len)
613                 len = ceph_wbc.i_size - page_off;
614
615         dout("writepage %p page %p index %lu on %llu~%u snapc %p seq %lld\n",
616              inode, page, page->index, page_off, len, snapc, snapc->seq);
617
618         writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
619         if (writeback_stat >
620             CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
621                 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
622
623         set_page_writeback(page);
624         err = ceph_osdc_writepages(&fsc->client->osdc, ceph_vino(inode),
625                                    &ci->i_layout, snapc, page_off, len,
626                                    ceph_wbc.truncate_seq,
627                                    ceph_wbc.truncate_size,
628                                    &inode->i_mtime, &page, 1);
629         if (err < 0) {
630                 struct writeback_control tmp_wbc;
631                 if (!wbc)
632                         wbc = &tmp_wbc;
633                 if (err == -ERESTARTSYS) {
634                         /* killed by SIGKILL */
635                         dout("writepage interrupted page %p\n", page);
636                         redirty_page_for_writepage(wbc, page);
637                         end_page_writeback(page);
638                         return err;
639                 }
640                 dout("writepage setting page/mapping error %d %p\n",
641                      err, page);
642                 SetPageError(page);
643                 mapping_set_error(&inode->i_data, err);
644                 wbc->pages_skipped++;
645         } else {
646                 dout("writepage cleaned page %p\n", page);
647                 err = 0;  /* vfs expects us to return 0 */
648         }
649         page->private = 0;
650         ClearPagePrivate(page);
651         end_page_writeback(page);
652         ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
653         ceph_put_snap_context(snapc);  /* page's reference */
654         return err;
655 }
656
657 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
658 {
659         int err;
660         struct inode *inode = page->mapping->host;
661         BUG_ON(!inode);
662         ihold(inode);
663         err = writepage_nounlock(page, wbc);
664         if (err == -ERESTARTSYS) {
665                 /* direct memory reclaimer was killed by SIGKILL. return 0
666                  * to prevent caller from setting mapping/page error */
667                 err = 0;
668         }
669         unlock_page(page);
670         iput(inode);
671         return err;
672 }
673
674 /*
675  * lame release_pages helper.  release_pages() isn't exported to
676  * modules.
677  */
678 static void ceph_release_pages(struct page **pages, int num)
679 {
680         struct pagevec pvec;
681         int i;
682
683         pagevec_init(&pvec, 0);
684         for (i = 0; i < num; i++) {
685                 if (pagevec_add(&pvec, pages[i]) == 0)
686                         pagevec_release(&pvec);
687         }
688         pagevec_release(&pvec);
689 }
690
691 /*
692  * async writeback completion handler.
693  *
694  * If we get an error, set the mapping error bit, but not the individual
695  * page error bits.
696  */
697 static void writepages_finish(struct ceph_osd_request *req)
698 {
699         struct inode *inode = req->r_inode;
700         struct ceph_inode_info *ci = ceph_inode(inode);
701         struct ceph_osd_data *osd_data;
702         struct page *page;
703         int num_pages, total_pages = 0;
704         int i, j;
705         int rc = req->r_result;
706         struct ceph_snap_context *snapc = req->r_snapc;
707         struct address_space *mapping = inode->i_mapping;
708         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
709         bool remove_page;
710
711         dout("writepages_finish %p rc %d\n", inode, rc);
712         if (rc < 0) {
713                 mapping_set_error(mapping, rc);
714                 ceph_set_error_write(ci);
715         } else {
716                 ceph_clear_error_write(ci);
717         }
718
719         /*
720          * We lost the cache cap, need to truncate the page before
721          * it is unlocked, otherwise we'd truncate it later in the
722          * page truncation thread, possibly losing some data that
723          * raced its way in
724          */
725         remove_page = !(ceph_caps_issued(ci) &
726                         (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
727
728         /* clean all pages */
729         for (i = 0; i < req->r_num_ops; i++) {
730                 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
731                         break;
732
733                 osd_data = osd_req_op_extent_osd_data(req, i);
734                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
735                 num_pages = calc_pages_for((u64)osd_data->alignment,
736                                            (u64)osd_data->length);
737                 total_pages += num_pages;
738                 for (j = 0; j < num_pages; j++) {
739                         page = osd_data->pages[j];
740                         BUG_ON(!page);
741                         WARN_ON(!PageUptodate(page));
742
743                         if (atomic_long_dec_return(&fsc->writeback_count) <
744                              CONGESTION_OFF_THRESH(
745                                         fsc->mount_options->congestion_kb))
746                                 clear_bdi_congested(inode_to_bdi(inode),
747                                                     BLK_RW_ASYNC);
748
749                         ceph_put_snap_context(page_snap_context(page));
750                         page->private = 0;
751                         ClearPagePrivate(page);
752                         dout("unlocking %p\n", page);
753                         end_page_writeback(page);
754
755                         if (remove_page)
756                                 generic_error_remove_page(inode->i_mapping,
757                                                           page);
758
759                         unlock_page(page);
760                 }
761                 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
762                      inode, osd_data->length, rc >= 0 ? num_pages : 0);
763
764                 ceph_release_pages(osd_data->pages, num_pages);
765         }
766
767         ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
768
769         osd_data = osd_req_op_extent_osd_data(req, 0);
770         if (osd_data->pages_from_pool)
771                 mempool_free(osd_data->pages,
772                              ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
773         else
774                 kfree(osd_data->pages);
775         ceph_osdc_put_request(req);
776 }
777
778 /*
779  * initiate async writeback
780  */
781 static int ceph_writepages_start(struct address_space *mapping,
782                                  struct writeback_control *wbc)
783 {
784         struct inode *inode = mapping->host;
785         struct ceph_inode_info *ci = ceph_inode(inode);
786         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
787         struct ceph_vino vino = ceph_vino(inode);
788         pgoff_t index, start_index, end = -1;
789         struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
790         struct pagevec pvec;
791         int rc = 0;
792         unsigned int wsize = i_blocksize(inode);
793         struct ceph_osd_request *req = NULL;
794         struct ceph_writeback_ctl ceph_wbc;
795         bool should_loop, range_whole = false;
796         bool stop, done = false;
797
798         dout("writepages_start %p (mode=%s)\n", inode,
799              wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
800              (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
801
802         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
803                 if (ci->i_wrbuffer_ref > 0) {
804                         pr_warn_ratelimited(
805                                 "writepage_start %p %lld forced umount\n",
806                                 inode, ceph_ino(inode));
807                 }
808                 mapping_set_error(mapping, -EIO);
809                 return -EIO; /* we're in a forced umount, don't write! */
810         }
811         if (fsc->mount_options->wsize < wsize)
812                 wsize = fsc->mount_options->wsize;
813
814         pagevec_init(&pvec, 0);
815
816         start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
817         index = start_index;
818
819 retry:
820         /* find oldest snap context with dirty data */
821         snapc = get_oldest_context(inode, &ceph_wbc, NULL);
822         if (!snapc) {
823                 /* hmm, why does writepages get called when there
824                    is no dirty data? */
825                 dout(" no snap context with dirty data?\n");
826                 goto out;
827         }
828         dout(" oldest snapc is %p seq %lld (%d snaps)\n",
829              snapc, snapc->seq, snapc->num_snaps);
830
831         should_loop = false;
832         if (ceph_wbc.head_snapc && snapc != last_snapc) {
833                 /* where to start/end? */
834                 if (wbc->range_cyclic) {
835                         index = start_index;
836                         end = -1;
837                         if (index > 0)
838                                 should_loop = true;
839                         dout(" cyclic, start at %lu\n", index);
840                 } else {
841                         index = wbc->range_start >> PAGE_SHIFT;
842                         end = wbc->range_end >> PAGE_SHIFT;
843                         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
844                                 range_whole = true;
845                         dout(" not cyclic, %lu to %lu\n", index, end);
846                 }
847         } else if (!ceph_wbc.head_snapc) {
848                 /* Do not respect wbc->range_{start,end}. Dirty pages
849                  * in that range can be associated with newer snapc.
850                  * They are not writeable until we write all dirty pages
851                  * associated with 'snapc' get written */
852                 if (index > 0 || wbc->sync_mode != WB_SYNC_NONE)
853                         should_loop = true;
854                 dout(" non-head snapc, range whole\n");
855         }
856
857         ceph_put_snap_context(last_snapc);
858         last_snapc = snapc;
859
860         stop = false;
861         while (!stop && index <= end) {
862                 int num_ops = 0, op_idx;
863                 unsigned i, pvec_pages, max_pages, locked_pages = 0;
864                 struct page **pages = NULL, **data_pages;
865                 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
866                 struct page *page;
867                 pgoff_t strip_unit_end = 0;
868                 u64 offset = 0, len = 0;
869
870                 max_pages = wsize >> PAGE_SHIFT;
871
872 get_more_pages:
873                 pvec_pages = min_t(unsigned, PAGEVEC_SIZE,
874                                    max_pages - locked_pages);
875                 if (end - index < (u64)(pvec_pages - 1))
876                         pvec_pages = (unsigned)(end - index) + 1;
877
878                 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
879                                                 PAGECACHE_TAG_DIRTY,
880                                                 pvec_pages);
881                 dout("pagevec_lookup_tag got %d\n", pvec_pages);
882                 if (!pvec_pages && !locked_pages)
883                         break;
884                 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
885                         page = pvec.pages[i];
886                         dout("? %p idx %lu\n", page, page->index);
887                         if (locked_pages == 0)
888                                 lock_page(page);  /* first page */
889                         else if (!trylock_page(page))
890                                 break;
891
892                         /* only dirty pages, or our accounting breaks */
893                         if (unlikely(!PageDirty(page)) ||
894                             unlikely(page->mapping != mapping)) {
895                                 dout("!dirty or !mapping %p\n", page);
896                                 unlock_page(page);
897                                 continue;
898                         }
899                         if (page->index > end) {
900                                 dout("end of range %p\n", page);
901                                 /* can't be range_cyclic (1st pass) because
902                                  * end == -1 in that case. */
903                                 stop = true;
904                                 if (ceph_wbc.head_snapc)
905                                         done = true;
906                                 unlock_page(page);
907                                 break;
908                         }
909                         if (strip_unit_end && (page->index > strip_unit_end)) {
910                                 dout("end of strip unit %p\n", page);
911                                 unlock_page(page);
912                                 break;
913                         }
914                         if (page_offset(page) >= ceph_wbc.i_size) {
915                                 dout("%p page eof %llu\n",
916                                      page, ceph_wbc.i_size);
917                                 /* not done if range_cyclic */
918                                 stop = true;
919                                 unlock_page(page);
920                                 break;
921                         }
922                         if (PageWriteback(page)) {
923                                 if (wbc->sync_mode == WB_SYNC_NONE) {
924                                         dout("%p under writeback\n", page);
925                                         unlock_page(page);
926                                         continue;
927                                 }
928                                 dout("waiting on writeback %p\n", page);
929                                 wait_on_page_writeback(page);
930                         }
931
932                         /* only if matching snap context */
933                         pgsnapc = page_snap_context(page);
934                         if (pgsnapc != snapc) {
935                                 dout("page snapc %p %lld != oldest %p %lld\n",
936                                      pgsnapc, pgsnapc->seq, snapc, snapc->seq);
937                                 unlock_page(page);
938                                 continue;
939                         }
940
941                         if (!clear_page_dirty_for_io(page)) {
942                                 dout("%p !clear_page_dirty_for_io\n", page);
943                                 unlock_page(page);
944                                 continue;
945                         }
946
947                         /*
948                          * We have something to write.  If this is
949                          * the first locked page this time through,
950                          * calculate max possinle write size and
951                          * allocate a page array
952                          */
953                         if (locked_pages == 0) {
954                                 u64 objnum;
955                                 u64 objoff;
956
957                                 /* prepare async write request */
958                                 offset = (u64)page_offset(page);
959                                 len = wsize;
960
961                                 rc = ceph_calc_file_object_mapping(&ci->i_layout,
962                                                                 offset, len,
963                                                                 &objnum, &objoff,
964                                                                 &len);
965                                 if (rc < 0) {
966                                         unlock_page(page);
967                                         break;
968                                 }
969
970                                 num_ops = 1;
971                                 strip_unit_end = page->index +
972                                         ((len - 1) >> PAGE_SHIFT);
973
974                                 BUG_ON(pages);
975                                 max_pages = calc_pages_for(0, (u64)len);
976                                 pages = kmalloc(max_pages * sizeof (*pages),
977                                                 GFP_NOFS);
978                                 if (!pages) {
979                                         pool = fsc->wb_pagevec_pool;
980                                         pages = mempool_alloc(pool, GFP_NOFS);
981                                         BUG_ON(!pages);
982                                 }
983
984                                 len = 0;
985                         } else if (page->index !=
986                                    (offset + len) >> PAGE_SHIFT) {
987                                 if (num_ops >= (pool ?  CEPH_OSD_SLAB_OPS :
988                                                         CEPH_OSD_MAX_OPS)) {
989                                         redirty_page_for_writepage(wbc, page);
990                                         unlock_page(page);
991                                         break;
992                                 }
993
994                                 num_ops++;
995                                 offset = (u64)page_offset(page);
996                                 len = 0;
997                         }
998
999                         /* note position of first page in pvec */
1000                         dout("%p will write page %p idx %lu\n",
1001                              inode, page, page->index);
1002
1003                         if (atomic_long_inc_return(&fsc->writeback_count) >
1004                             CONGESTION_ON_THRESH(
1005                                     fsc->mount_options->congestion_kb)) {
1006                                 set_bdi_congested(inode_to_bdi(inode),
1007                                                   BLK_RW_ASYNC);
1008                         }
1009
1010
1011                         pages[locked_pages++] = page;
1012                         pvec.pages[i] = NULL;
1013
1014                         len += PAGE_SIZE;
1015                 }
1016
1017                 /* did we get anything? */
1018                 if (!locked_pages)
1019                         goto release_pvec_pages;
1020                 if (i) {
1021                         unsigned j, n = 0;
1022                         /* shift unused page to beginning of pvec */
1023                         for (j = 0; j < pvec_pages; j++) {
1024                                 if (!pvec.pages[j])
1025                                         continue;
1026                                 if (n < j)
1027                                         pvec.pages[n] = pvec.pages[j];
1028                                 n++;
1029                         }
1030                         pvec.nr = n;
1031
1032                         if (pvec_pages && i == pvec_pages &&
1033                             locked_pages < max_pages) {
1034                                 dout("reached end pvec, trying for more\n");
1035                                 pagevec_release(&pvec);
1036                                 goto get_more_pages;
1037                         }
1038                 }
1039
1040 new_request:
1041                 offset = page_offset(pages[0]);
1042                 len = wsize;
1043
1044                 req = ceph_osdc_new_request(&fsc->client->osdc,
1045                                         &ci->i_layout, vino,
1046                                         offset, &len, 0, num_ops,
1047                                         CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1048                                         snapc, ceph_wbc.truncate_seq,
1049                                         ceph_wbc.truncate_size, false);
1050                 if (IS_ERR(req)) {
1051                         req = ceph_osdc_new_request(&fsc->client->osdc,
1052                                                 &ci->i_layout, vino,
1053                                                 offset, &len, 0,
1054                                                 min(num_ops,
1055                                                     CEPH_OSD_SLAB_OPS),
1056                                                 CEPH_OSD_OP_WRITE,
1057                                                 CEPH_OSD_FLAG_WRITE,
1058                                                 snapc, ceph_wbc.truncate_seq,
1059                                                 ceph_wbc.truncate_size, true);
1060                         BUG_ON(IS_ERR(req));
1061                 }
1062                 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1063                              PAGE_SIZE - offset);
1064
1065                 req->r_callback = writepages_finish;
1066                 req->r_inode = inode;
1067
1068                 /* Format the osd request message and submit the write */
1069                 len = 0;
1070                 data_pages = pages;
1071                 op_idx = 0;
1072                 for (i = 0; i < locked_pages; i++) {
1073                         u64 cur_offset = page_offset(pages[i]);
1074                         if (offset + len != cur_offset) {
1075                                 if (op_idx + 1 == req->r_num_ops)
1076                                         break;
1077                                 osd_req_op_extent_dup_last(req, op_idx,
1078                                                            cur_offset - offset);
1079                                 dout("writepages got pages at %llu~%llu\n",
1080                                      offset, len);
1081                                 osd_req_op_extent_osd_data_pages(req, op_idx,
1082                                                         data_pages, len, 0,
1083                                                         !!pool, false);
1084                                 osd_req_op_extent_update(req, op_idx, len);
1085
1086                                 len = 0;
1087                                 offset = cur_offset; 
1088                                 data_pages = pages + i;
1089                                 op_idx++;
1090                         }
1091
1092                         set_page_writeback(pages[i]);
1093                         len += PAGE_SIZE;
1094                 }
1095
1096                 if (ceph_wbc.size_stable) {
1097                         len = min(len, ceph_wbc.i_size - offset);
1098                 } else if (i == locked_pages) {
1099                         /* writepages_finish() clears writeback pages
1100                          * according to the data length, so make sure
1101                          * data length covers all locked pages */
1102                         u64 min_len = len + 1 - PAGE_SIZE;
1103                         len = get_writepages_data_length(inode, pages[i - 1],
1104                                                          offset);
1105                         len = max(len, min_len);
1106                 }
1107                 dout("writepages got pages at %llu~%llu\n", offset, len);
1108
1109                 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1110                                                  0, !!pool, false);
1111                 osd_req_op_extent_update(req, op_idx, len);
1112
1113                 BUG_ON(op_idx + 1 != req->r_num_ops);
1114
1115                 pool = NULL;
1116                 if (i < locked_pages) {
1117                         BUG_ON(num_ops <= req->r_num_ops);
1118                         num_ops -= req->r_num_ops;
1119                         locked_pages -= i;
1120
1121                         /* allocate new pages array for next request */
1122                         data_pages = pages;
1123                         pages = kmalloc(locked_pages * sizeof (*pages),
1124                                         GFP_NOFS);
1125                         if (!pages) {
1126                                 pool = fsc->wb_pagevec_pool;
1127                                 pages = mempool_alloc(pool, GFP_NOFS);
1128                                 BUG_ON(!pages);
1129                         }
1130                         memcpy(pages, data_pages + i,
1131                                locked_pages * sizeof(*pages));
1132                         memset(data_pages + i, 0,
1133                                locked_pages * sizeof(*pages));
1134                 } else {
1135                         BUG_ON(num_ops != req->r_num_ops);
1136                         index = pages[i - 1]->index + 1;
1137                         /* request message now owns the pages array */
1138                         pages = NULL;
1139                 }
1140
1141                 req->r_mtime = inode->i_mtime;
1142                 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1143                 BUG_ON(rc);
1144                 req = NULL;
1145
1146                 wbc->nr_to_write -= i;
1147                 if (pages)
1148                         goto new_request;
1149
1150                 /*
1151                  * We stop writing back only if we are not doing
1152                  * integrity sync. In case of integrity sync we have to
1153                  * keep going until we have written all the pages
1154                  * we tagged for writeback prior to entering this loop.
1155                  */
1156                 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1157                         done = stop = true;
1158
1159 release_pvec_pages:
1160                 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1161                      pvec.nr ? pvec.pages[0] : NULL);
1162                 pagevec_release(&pvec);
1163         }
1164
1165         if (should_loop && !done) {
1166                 /* more to do; loop back to beginning of file */
1167                 dout("writepages looping back to beginning of file\n");
1168                 end = start_index - 1; /* OK even when start_index == 0 */
1169
1170                 /* to write dirty pages associated with next snapc,
1171                  * we need to wait until current writes complete */
1172                 if (wbc->sync_mode != WB_SYNC_NONE &&
1173                     start_index == 0 && /* all dirty pages were checked */
1174                     !ceph_wbc.head_snapc) {
1175                         struct page *page;
1176                         unsigned i, nr;
1177                         index = 0;
1178                         while ((index <= end) &&
1179                                (nr = pagevec_lookup_tag(&pvec, mapping, &index,
1180                                                         PAGECACHE_TAG_WRITEBACK,
1181                                                         PAGEVEC_SIZE))) {
1182                                 for (i = 0; i < nr; i++) {
1183                                         page = pvec.pages[i];
1184                                         if (page_snap_context(page) != snapc)
1185                                                 continue;
1186                                         wait_on_page_writeback(page);
1187                                 }
1188                                 pagevec_release(&pvec);
1189                                 cond_resched();
1190                         }
1191                 }
1192
1193                 start_index = 0;
1194                 index = 0;
1195                 goto retry;
1196         }
1197
1198         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1199                 mapping->writeback_index = index;
1200
1201 out:
1202         ceph_osdc_put_request(req);
1203         ceph_put_snap_context(last_snapc);
1204         dout("writepages dend - startone, rc = %d\n", rc);
1205         return rc;
1206 }
1207
1208
1209
1210 /*
1211  * See if a given @snapc is either writeable, or already written.
1212  */
1213 static int context_is_writeable_or_written(struct inode *inode,
1214                                            struct ceph_snap_context *snapc)
1215 {
1216         struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1217         int ret = !oldest || snapc->seq <= oldest->seq;
1218
1219         ceph_put_snap_context(oldest);
1220         return ret;
1221 }
1222
1223 /*
1224  * We are only allowed to write into/dirty the page if the page is
1225  * clean, or already dirty within the same snap context.
1226  *
1227  * called with page locked.
1228  * return success with page locked,
1229  * or any failure (incl -EAGAIN) with page unlocked.
1230  */
1231 static int ceph_update_writeable_page(struct file *file,
1232                             loff_t pos, unsigned len,
1233                             struct page *page)
1234 {
1235         struct inode *inode = file_inode(file);
1236         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1237         struct ceph_inode_info *ci = ceph_inode(inode);
1238         loff_t page_off = pos & PAGE_MASK;
1239         int pos_in_page = pos & ~PAGE_MASK;
1240         int end_in_page = pos_in_page + len;
1241         loff_t i_size;
1242         int r;
1243         struct ceph_snap_context *snapc, *oldest;
1244
1245         if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1246                 dout(" page %p forced umount\n", page);
1247                 unlock_page(page);
1248                 return -EIO;
1249         }
1250
1251 retry_locked:
1252         /* writepages currently holds page lock, but if we change that later, */
1253         wait_on_page_writeback(page);
1254
1255         snapc = page_snap_context(page);
1256         if (snapc && snapc != ci->i_head_snapc) {
1257                 /*
1258                  * this page is already dirty in another (older) snap
1259                  * context!  is it writeable now?
1260                  */
1261                 oldest = get_oldest_context(inode, NULL, NULL);
1262                 if (snapc->seq > oldest->seq) {
1263                         ceph_put_snap_context(oldest);
1264                         dout(" page %p snapc %p not current or oldest\n",
1265                              page, snapc);
1266                         /*
1267                          * queue for writeback, and wait for snapc to
1268                          * be writeable or written
1269                          */
1270                         snapc = ceph_get_snap_context(snapc);
1271                         unlock_page(page);
1272                         ceph_queue_writeback(inode);
1273                         r = wait_event_killable(ci->i_cap_wq,
1274                                context_is_writeable_or_written(inode, snapc));
1275                         ceph_put_snap_context(snapc);
1276                         if (r == -ERESTARTSYS)
1277                                 return r;
1278                         return -EAGAIN;
1279                 }
1280                 ceph_put_snap_context(oldest);
1281
1282                 /* yay, writeable, do it now (without dropping page lock) */
1283                 dout(" page %p snapc %p not current, but oldest\n",
1284                      page, snapc);
1285                 if (!clear_page_dirty_for_io(page))
1286                         goto retry_locked;
1287                 r = writepage_nounlock(page, NULL);
1288                 if (r < 0)
1289                         goto fail_unlock;
1290                 goto retry_locked;
1291         }
1292
1293         if (PageUptodate(page)) {
1294                 dout(" page %p already uptodate\n", page);
1295                 return 0;
1296         }
1297
1298         /* full page? */
1299         if (pos_in_page == 0 && len == PAGE_SIZE)
1300                 return 0;
1301
1302         /* past end of file? */
1303         i_size = i_size_read(inode);
1304
1305         if (page_off >= i_size ||
1306             (pos_in_page == 0 && (pos+len) >= i_size &&
1307              end_in_page - pos_in_page != PAGE_SIZE)) {
1308                 dout(" zeroing %p 0 - %d and %d - %d\n",
1309                      page, pos_in_page, end_in_page, (int)PAGE_SIZE);
1310                 zero_user_segments(page,
1311                                    0, pos_in_page,
1312                                    end_in_page, PAGE_SIZE);
1313                 return 0;
1314         }
1315
1316         /* we need to read it. */
1317         r = ceph_do_readpage(file, page);
1318         if (r < 0) {
1319                 if (r == -EINPROGRESS)
1320                         return -EAGAIN;
1321                 goto fail_unlock;
1322         }
1323         goto retry_locked;
1324 fail_unlock:
1325         unlock_page(page);
1326         return r;
1327 }
1328
1329 /*
1330  * We are only allowed to write into/dirty the page if the page is
1331  * clean, or already dirty within the same snap context.
1332  */
1333 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1334                             loff_t pos, unsigned len, unsigned flags,
1335                             struct page **pagep, void **fsdata)
1336 {
1337         struct inode *inode = file_inode(file);
1338         struct page *page;
1339         pgoff_t index = pos >> PAGE_SHIFT;
1340         int r;
1341
1342         do {
1343                 /* get a page */
1344                 page = grab_cache_page_write_begin(mapping, index, 0);
1345                 if (!page)
1346                         return -ENOMEM;
1347
1348                 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1349                      inode, page, (int)pos, (int)len);
1350
1351                 r = ceph_update_writeable_page(file, pos, len, page);
1352                 if (r < 0)
1353                         put_page(page);
1354                 else
1355                         *pagep = page;
1356         } while (r == -EAGAIN);
1357
1358         return r;
1359 }
1360
1361 /*
1362  * we don't do anything in here that simple_write_end doesn't do
1363  * except adjust dirty page accounting
1364  */
1365 static int ceph_write_end(struct file *file, struct address_space *mapping,
1366                           loff_t pos, unsigned len, unsigned copied,
1367                           struct page *page, void *fsdata)
1368 {
1369         struct inode *inode = file_inode(file);
1370         bool check_cap = false;
1371
1372         dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1373              inode, page, (int)pos, (int)copied, (int)len);
1374
1375         /* zero the stale part of the page if we did a short copy */
1376         if (!PageUptodate(page)) {
1377                 if (copied < len) {
1378                         copied = 0;
1379                         goto out;
1380                 }
1381                 SetPageUptodate(page);
1382         }
1383
1384         /* did file size increase? */
1385         if (pos+copied > i_size_read(inode))
1386                 check_cap = ceph_inode_set_size(inode, pos+copied);
1387
1388         set_page_dirty(page);
1389
1390 out:
1391         unlock_page(page);
1392         put_page(page);
1393
1394         if (check_cap)
1395                 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1396
1397         return copied;
1398 }
1399
1400 /*
1401  * we set .direct_IO to indicate direct io is supported, but since we
1402  * intercept O_DIRECT reads and writes early, this function should
1403  * never get called.
1404  */
1405 static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
1406 {
1407         WARN_ON(1);
1408         return -EINVAL;
1409 }
1410
1411 const struct address_space_operations ceph_aops = {
1412         .readpage = ceph_readpage,
1413         .readpages = ceph_readpages,
1414         .writepage = ceph_writepage,
1415         .writepages = ceph_writepages_start,
1416         .write_begin = ceph_write_begin,
1417         .write_end = ceph_write_end,
1418         .set_page_dirty = ceph_set_page_dirty,
1419         .invalidatepage = ceph_invalidatepage,
1420         .releasepage = ceph_releasepage,
1421         .direct_IO = ceph_direct_io,
1422 };
1423
1424 static void ceph_block_sigs(sigset_t *oldset)
1425 {
1426         sigset_t mask;
1427         siginitsetinv(&mask, sigmask(SIGKILL));
1428         sigprocmask(SIG_BLOCK, &mask, oldset);
1429 }
1430
1431 static void ceph_restore_sigs(sigset_t *oldset)
1432 {
1433         sigprocmask(SIG_SETMASK, oldset, NULL);
1434 }
1435
1436 /*
1437  * vm ops
1438  */
1439 static int ceph_filemap_fault(struct vm_fault *vmf)
1440 {
1441         struct vm_area_struct *vma = vmf->vma;
1442         struct inode *inode = file_inode(vma->vm_file);
1443         struct ceph_inode_info *ci = ceph_inode(inode);
1444         struct ceph_file_info *fi = vma->vm_file->private_data;
1445         struct page *pinned_page = NULL;
1446         loff_t off = vmf->pgoff << PAGE_SHIFT;
1447         int want, got, ret;
1448         sigset_t oldset;
1449
1450         ceph_block_sigs(&oldset);
1451
1452         dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1453              inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1454         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1455                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1456         else
1457                 want = CEPH_CAP_FILE_CACHE;
1458
1459         got = 0;
1460         ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1461         if (ret < 0)
1462                 goto out_restore;
1463
1464         dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1465              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1466
1467         if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1468             ci->i_inline_version == CEPH_INLINE_NONE) {
1469                 current->journal_info = vma->vm_file;
1470                 ret = filemap_fault(vmf);
1471                 current->journal_info = NULL;
1472         } else
1473                 ret = -EAGAIN;
1474
1475         dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1476              inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got), ret);
1477         if (pinned_page)
1478                 put_page(pinned_page);
1479         ceph_put_cap_refs(ci, got);
1480
1481         if (ret != -EAGAIN)
1482                 goto out_restore;
1483
1484         /* read inline data */
1485         if (off >= PAGE_SIZE) {
1486                 /* does not support inline data > PAGE_SIZE */
1487                 ret = VM_FAULT_SIGBUS;
1488         } else {
1489                 int ret1;
1490                 struct address_space *mapping = inode->i_mapping;
1491                 struct page *page = find_or_create_page(mapping, 0,
1492                                                 mapping_gfp_constraint(mapping,
1493                                                 ~__GFP_FS));
1494                 if (!page) {
1495                         ret = VM_FAULT_OOM;
1496                         goto out_inline;
1497                 }
1498                 ret1 = __ceph_do_getattr(inode, page,
1499                                          CEPH_STAT_CAP_INLINE_DATA, true);
1500                 if (ret1 < 0 || off >= i_size_read(inode)) {
1501                         unlock_page(page);
1502                         put_page(page);
1503                         if (ret1 < 0)
1504                                 ret = ret1;
1505                         else
1506                                 ret = VM_FAULT_SIGBUS;
1507                         goto out_inline;
1508                 }
1509                 if (ret1 < PAGE_SIZE)
1510                         zero_user_segment(page, ret1, PAGE_SIZE);
1511                 else
1512                         flush_dcache_page(page);
1513                 SetPageUptodate(page);
1514                 vmf->page = page;
1515                 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1516 out_inline:
1517                 dout("filemap_fault %p %llu~%zd read inline data ret %d\n",
1518                      inode, off, (size_t)PAGE_SIZE, ret);
1519         }
1520 out_restore:
1521         ceph_restore_sigs(&oldset);
1522         if (ret < 0)
1523                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1524
1525         return ret;
1526 }
1527
1528 /*
1529  * Reuse write_begin here for simplicity.
1530  */
1531 static int ceph_page_mkwrite(struct vm_fault *vmf)
1532 {
1533         struct vm_area_struct *vma = vmf->vma;
1534         struct inode *inode = file_inode(vma->vm_file);
1535         struct ceph_inode_info *ci = ceph_inode(inode);
1536         struct ceph_file_info *fi = vma->vm_file->private_data;
1537         struct ceph_cap_flush *prealloc_cf;
1538         struct page *page = vmf->page;
1539         loff_t off = page_offset(page);
1540         loff_t size = i_size_read(inode);
1541         size_t len;
1542         int want, got, ret;
1543         sigset_t oldset;
1544
1545         prealloc_cf = ceph_alloc_cap_flush();
1546         if (!prealloc_cf)
1547                 return VM_FAULT_OOM;
1548
1549         ceph_block_sigs(&oldset);
1550
1551         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1552                 struct page *locked_page = NULL;
1553                 if (off == 0) {
1554                         lock_page(page);
1555                         locked_page = page;
1556                 }
1557                 ret = ceph_uninline_data(vma->vm_file, locked_page);
1558                 if (locked_page)
1559                         unlock_page(locked_page);
1560                 if (ret < 0)
1561                         goto out_free;
1562         }
1563
1564         if (off + PAGE_SIZE <= size)
1565                 len = PAGE_SIZE;
1566         else
1567                 len = size & ~PAGE_MASK;
1568
1569         dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1570              inode, ceph_vinop(inode), off, len, size);
1571         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1572                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1573         else
1574                 want = CEPH_CAP_FILE_BUFFER;
1575
1576         got = 0;
1577         ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1578                             &got, NULL);
1579         if (ret < 0)
1580                 goto out_free;
1581
1582         dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1583              inode, off, len, ceph_cap_string(got));
1584
1585         /* Update time before taking page lock */
1586         file_update_time(vma->vm_file);
1587
1588         do {
1589                 lock_page(page);
1590
1591                 if ((off > size) || (page->mapping != inode->i_mapping)) {
1592                         unlock_page(page);
1593                         ret = VM_FAULT_NOPAGE;
1594                         break;
1595                 }
1596
1597                 ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1598                 if (ret >= 0) {
1599                         /* success.  we'll keep the page locked. */
1600                         set_page_dirty(page);
1601                         ret = VM_FAULT_LOCKED;
1602                 }
1603         } while (ret == -EAGAIN);
1604
1605         if (ret == VM_FAULT_LOCKED ||
1606             ci->i_inline_version != CEPH_INLINE_NONE) {
1607                 int dirty;
1608                 spin_lock(&ci->i_ceph_lock);
1609                 ci->i_inline_version = CEPH_INLINE_NONE;
1610                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1611                                                &prealloc_cf);
1612                 spin_unlock(&ci->i_ceph_lock);
1613                 if (dirty)
1614                         __mark_inode_dirty(inode, dirty);
1615         }
1616
1617         dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1618              inode, off, len, ceph_cap_string(got), ret);
1619         ceph_put_cap_refs(ci, got);
1620 out_free:
1621         ceph_restore_sigs(&oldset);
1622         ceph_free_cap_flush(prealloc_cf);
1623         if (ret < 0)
1624                 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1625         return ret;
1626 }
1627
1628 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1629                            char *data, size_t len)
1630 {
1631         struct address_space *mapping = inode->i_mapping;
1632         struct page *page;
1633
1634         if (locked_page) {
1635                 page = locked_page;
1636         } else {
1637                 if (i_size_read(inode) == 0)
1638                         return;
1639                 page = find_or_create_page(mapping, 0,
1640                                            mapping_gfp_constraint(mapping,
1641                                            ~__GFP_FS));
1642                 if (!page)
1643                         return;
1644                 if (PageUptodate(page)) {
1645                         unlock_page(page);
1646                         put_page(page);
1647                         return;
1648                 }
1649         }
1650
1651         dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1652              inode, ceph_vinop(inode), len, locked_page);
1653
1654         if (len > 0) {
1655                 void *kaddr = kmap_atomic(page);
1656                 memcpy(kaddr, data, len);
1657                 kunmap_atomic(kaddr);
1658         }
1659
1660         if (page != locked_page) {
1661                 if (len < PAGE_SIZE)
1662                         zero_user_segment(page, len, PAGE_SIZE);
1663                 else
1664                         flush_dcache_page(page);
1665
1666                 SetPageUptodate(page);
1667                 unlock_page(page);
1668                 put_page(page);
1669         }
1670 }
1671
1672 int ceph_uninline_data(struct file *filp, struct page *locked_page)
1673 {
1674         struct inode *inode = file_inode(filp);
1675         struct ceph_inode_info *ci = ceph_inode(inode);
1676         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1677         struct ceph_osd_request *req;
1678         struct page *page = NULL;
1679         u64 len, inline_version;
1680         int err = 0;
1681         bool from_pagecache = false;
1682
1683         spin_lock(&ci->i_ceph_lock);
1684         inline_version = ci->i_inline_version;
1685         spin_unlock(&ci->i_ceph_lock);
1686
1687         dout("uninline_data %p %llx.%llx inline_version %llu\n",
1688              inode, ceph_vinop(inode), inline_version);
1689
1690         if (inline_version == 1 || /* initial version, no data */
1691             inline_version == CEPH_INLINE_NONE)
1692                 goto out;
1693
1694         if (locked_page) {
1695                 page = locked_page;
1696                 WARN_ON(!PageUptodate(page));
1697         } else if (ceph_caps_issued(ci) &
1698                    (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1699                 page = find_get_page(inode->i_mapping, 0);
1700                 if (page) {
1701                         if (PageUptodate(page)) {
1702                                 from_pagecache = true;
1703                                 lock_page(page);
1704                         } else {
1705                                 put_page(page);
1706                                 page = NULL;
1707                         }
1708                 }
1709         }
1710
1711         if (page) {
1712                 len = i_size_read(inode);
1713                 if (len > PAGE_SIZE)
1714                         len = PAGE_SIZE;
1715         } else {
1716                 page = __page_cache_alloc(GFP_NOFS);
1717                 if (!page) {
1718                         err = -ENOMEM;
1719                         goto out;
1720                 }
1721                 err = __ceph_do_getattr(inode, page,
1722                                         CEPH_STAT_CAP_INLINE_DATA, true);
1723                 if (err < 0) {
1724                         /* no inline data */
1725                         if (err == -ENODATA)
1726                                 err = 0;
1727                         goto out;
1728                 }
1729                 len = err;
1730         }
1731
1732         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1733                                     ceph_vino(inode), 0, &len, 0, 1,
1734                                     CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1735                                     NULL, 0, 0, false);
1736         if (IS_ERR(req)) {
1737                 err = PTR_ERR(req);
1738                 goto out;
1739         }
1740
1741         req->r_mtime = inode->i_mtime;
1742         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1743         if (!err)
1744                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1745         ceph_osdc_put_request(req);
1746         if (err < 0)
1747                 goto out;
1748
1749         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1750                                     ceph_vino(inode), 0, &len, 1, 3,
1751                                     CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1752                                     NULL, ci->i_truncate_seq,
1753                                     ci->i_truncate_size, false);
1754         if (IS_ERR(req)) {
1755                 err = PTR_ERR(req);
1756                 goto out;
1757         }
1758
1759         osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1760
1761         {
1762                 __le64 xattr_buf = cpu_to_le64(inline_version);
1763                 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1764                                             "inline_version", &xattr_buf,
1765                                             sizeof(xattr_buf),
1766                                             CEPH_OSD_CMPXATTR_OP_GT,
1767                                             CEPH_OSD_CMPXATTR_MODE_U64);
1768                 if (err)
1769                         goto out_put;
1770         }
1771
1772         {
1773                 char xattr_buf[32];
1774                 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1775                                          "%llu", inline_version);
1776                 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1777                                             "inline_version",
1778                                             xattr_buf, xattr_len, 0, 0);
1779                 if (err)
1780                         goto out_put;
1781         }
1782
1783         req->r_mtime = inode->i_mtime;
1784         err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1785         if (!err)
1786                 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1787 out_put:
1788         ceph_osdc_put_request(req);
1789         if (err == -ECANCELED)
1790                 err = 0;
1791 out:
1792         if (page && page != locked_page) {
1793                 if (from_pagecache) {
1794                         unlock_page(page);
1795                         put_page(page);
1796                 } else
1797                         __free_pages(page, 0);
1798         }
1799
1800         dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1801              inode, ceph_vinop(inode), inline_version, err);
1802         return err;
1803 }
1804
1805 static const struct vm_operations_struct ceph_vmops = {
1806         .fault          = ceph_filemap_fault,
1807         .page_mkwrite   = ceph_page_mkwrite,
1808 };
1809
1810 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1811 {
1812         struct address_space *mapping = file->f_mapping;
1813
1814         if (!mapping->a_ops->readpage)
1815                 return -ENOEXEC;
1816         file_accessed(file);
1817         vma->vm_ops = &ceph_vmops;
1818         return 0;
1819 }
1820
1821 enum {
1822         POOL_READ       = 1,
1823         POOL_WRITE      = 2,
1824 };
1825
1826 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1827                                 s64 pool, struct ceph_string *pool_ns)
1828 {
1829         struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1830         struct ceph_mds_client *mdsc = fsc->mdsc;
1831         struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1832         struct rb_node **p, *parent;
1833         struct ceph_pool_perm *perm;
1834         struct page **pages;
1835         size_t pool_ns_len;
1836         int err = 0, err2 = 0, have = 0;
1837
1838         down_read(&mdsc->pool_perm_rwsem);
1839         p = &mdsc->pool_perm_tree.rb_node;
1840         while (*p) {
1841                 perm = rb_entry(*p, struct ceph_pool_perm, node);
1842                 if (pool < perm->pool)
1843                         p = &(*p)->rb_left;
1844                 else if (pool > perm->pool)
1845                         p = &(*p)->rb_right;
1846                 else {
1847                         int ret = ceph_compare_string(pool_ns,
1848                                                 perm->pool_ns,
1849                                                 perm->pool_ns_len);
1850                         if (ret < 0)
1851                                 p = &(*p)->rb_left;
1852                         else if (ret > 0)
1853                                 p = &(*p)->rb_right;
1854                         else {
1855                                 have = perm->perm;
1856                                 break;
1857                         }
1858                 }
1859         }
1860         up_read(&mdsc->pool_perm_rwsem);
1861         if (*p)
1862                 goto out;
1863
1864         if (pool_ns)
1865                 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1866                      pool, (int)pool_ns->len, pool_ns->str);
1867         else
1868                 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1869
1870         down_write(&mdsc->pool_perm_rwsem);
1871         p = &mdsc->pool_perm_tree.rb_node;
1872         parent = NULL;
1873         while (*p) {
1874                 parent = *p;
1875                 perm = rb_entry(parent, struct ceph_pool_perm, node);
1876                 if (pool < perm->pool)
1877                         p = &(*p)->rb_left;
1878                 else if (pool > perm->pool)
1879                         p = &(*p)->rb_right;
1880                 else {
1881                         int ret = ceph_compare_string(pool_ns,
1882                                                 perm->pool_ns,
1883                                                 perm->pool_ns_len);
1884                         if (ret < 0)
1885                                 p = &(*p)->rb_left;
1886                         else if (ret > 0)
1887                                 p = &(*p)->rb_right;
1888                         else {
1889                                 have = perm->perm;
1890                                 break;
1891                         }
1892                 }
1893         }
1894         if (*p) {
1895                 up_write(&mdsc->pool_perm_rwsem);
1896                 goto out;
1897         }
1898
1899         rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1900                                          1, false, GFP_NOFS);
1901         if (!rd_req) {
1902                 err = -ENOMEM;
1903                 goto out_unlock;
1904         }
1905
1906         rd_req->r_flags = CEPH_OSD_FLAG_READ;
1907         osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1908         rd_req->r_base_oloc.pool = pool;
1909         if (pool_ns)
1910                 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1911         ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1912
1913         err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1914         if (err)
1915                 goto out_unlock;
1916
1917         wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1918                                          1, false, GFP_NOFS);
1919         if (!wr_req) {
1920                 err = -ENOMEM;
1921                 goto out_unlock;
1922         }
1923
1924         wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1925         osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1926         ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1927         ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1928
1929         err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1930         if (err)
1931                 goto out_unlock;
1932
1933         /* one page should be large enough for STAT data */
1934         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1935         if (IS_ERR(pages)) {
1936                 err = PTR_ERR(pages);
1937                 goto out_unlock;
1938         }
1939
1940         osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1941                                      0, false, true);
1942         err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1943
1944         wr_req->r_mtime = ci->vfs_inode.i_mtime;
1945         wr_req->r_abort_on_full = true;
1946         err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1947
1948         if (!err)
1949                 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1950         if (!err2)
1951                 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1952
1953         if (err >= 0 || err == -ENOENT)
1954                 have |= POOL_READ;
1955         else if (err != -EPERM)
1956                 goto out_unlock;
1957
1958         if (err2 == 0 || err2 == -EEXIST)
1959                 have |= POOL_WRITE;
1960         else if (err2 != -EPERM) {
1961                 err = err2;
1962                 goto out_unlock;
1963         }
1964
1965         pool_ns_len = pool_ns ? pool_ns->len : 0;
1966         perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1967         if (!perm) {
1968                 err = -ENOMEM;
1969                 goto out_unlock;
1970         }
1971
1972         perm->pool = pool;
1973         perm->perm = have;
1974         perm->pool_ns_len = pool_ns_len;
1975         if (pool_ns_len > 0)
1976                 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1977         perm->pool_ns[pool_ns_len] = 0;
1978
1979         rb_link_node(&perm->node, parent, p);
1980         rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1981         err = 0;
1982 out_unlock:
1983         up_write(&mdsc->pool_perm_rwsem);
1984
1985         ceph_osdc_put_request(rd_req);
1986         ceph_osdc_put_request(wr_req);
1987 out:
1988         if (!err)
1989                 err = have;
1990         if (pool_ns)
1991                 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1992                      pool, (int)pool_ns->len, pool_ns->str, err);
1993         else
1994                 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1995         return err;
1996 }
1997
1998 int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1999 {
2000         s64 pool;
2001         struct ceph_string *pool_ns;
2002         int ret, flags;
2003
2004         if (ci->i_vino.snap != CEPH_NOSNAP) {
2005                 /*
2006                  * Pool permission check needs to write to the first object.
2007                  * But for snapshot, head of the first object may have alread
2008                  * been deleted. Skip check to avoid creating orphan object.
2009                  */
2010                 return 0;
2011         }
2012
2013         if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
2014                                 NOPOOLPERM))
2015                 return 0;
2016
2017         spin_lock(&ci->i_ceph_lock);
2018         flags = ci->i_ceph_flags;
2019         pool = ci->i_layout.pool_id;
2020         spin_unlock(&ci->i_ceph_lock);
2021 check:
2022         if (flags & CEPH_I_POOL_PERM) {
2023                 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2024                         dout("ceph_pool_perm_check pool %lld no read perm\n",
2025                              pool);
2026                         return -EPERM;
2027                 }
2028                 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2029                         dout("ceph_pool_perm_check pool %lld no write perm\n",
2030                              pool);
2031                         return -EPERM;
2032                 }
2033                 return 0;
2034         }
2035
2036         pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2037         ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2038         ceph_put_string(pool_ns);
2039         if (ret < 0)
2040                 return ret;
2041
2042         flags = CEPH_I_POOL_PERM;
2043         if (ret & POOL_READ)
2044                 flags |= CEPH_I_POOL_RD;
2045         if (ret & POOL_WRITE)
2046                 flags |= CEPH_I_POOL_WR;
2047
2048         spin_lock(&ci->i_ceph_lock);
2049         if (pool == ci->i_layout.pool_id &&
2050             pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2051                 ci->i_ceph_flags |= flags;
2052         } else {
2053                 pool = ci->i_layout.pool_id;
2054                 flags = ci->i_ceph_flags;
2055         }
2056         spin_unlock(&ci->i_ceph_lock);
2057         goto check;
2058 }
2059
2060 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2061 {
2062         struct ceph_pool_perm *perm;
2063         struct rb_node *n;
2064
2065         while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2066                 n = rb_first(&mdsc->pool_perm_tree);
2067                 perm = rb_entry(n, struct ceph_pool_perm, node);
2068                 rb_erase(n, &mdsc->pool_perm_tree);
2069                 kfree(perm);
2070         }
2071 }