Loading...
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 "cache.h"
15#include <linux/ceph/osd_client.h>
16
17/*
18 * Ceph address space ops.
19 *
20 * There are a few funny things going on here.
21 *
22 * The page->private field is used to reference a struct
23 * ceph_snap_context for _every_ dirty page. This indicates which
24 * snapshot the page was logically dirtied in, and thus which snap
25 * context needs to be associated with the osd write during writeback.
26 *
27 * Similarly, struct ceph_inode_info maintains a set of counters to
28 * count dirty pages on the inode. In the absence of snapshots,
29 * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
30 *
31 * When a snapshot is taken (that is, when the client receives
32 * notification that a snapshot was taken), each inode with caps and
33 * with dirty pages (dirty pages implies there is a cap) gets a new
34 * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
35 * order, new snaps go to the tail). The i_wrbuffer_ref_head count is
36 * moved to capsnap->dirty. (Unless a sync write is currently in
37 * progress. In that case, the capsnap is said to be "pending", new
38 * writes cannot start, and the capsnap isn't "finalized" until the
39 * write completes (or fails) and a final size/mtime for the inode for
40 * that snap can be settled upon.) i_wrbuffer_ref_head is reset to 0.
41 *
42 * On writeback, we must submit writes to the osd IN SNAP ORDER. So,
43 * we look for the first capsnap in i_cap_snaps and write out pages in
44 * that snap context _only_. Then we move on to the next capsnap,
45 * eventually reaching the "live" or "head" context (i.e., pages that
46 * are not yet snapped) and are writing the most recently dirtied
47 * pages.
48 *
49 * Invalidate and so forth must take care to ensure the dirty page
50 * accounting is preserved.
51 */
52
53#define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
54#define CONGESTION_OFF_THRESH(congestion_kb) \
55 (CONGESTION_ON_THRESH(congestion_kb) - \
56 (CONGESTION_ON_THRESH(congestion_kb) >> 2))
57
58static inline struct ceph_snap_context *page_snap_context(struct page *page)
59{
60 if (PagePrivate(page))
61 return (void *)page->private;
62 return NULL;
63}
64
65/*
66 * Dirty a page. Optimistically adjust accounting, on the assumption
67 * that we won't race with invalidate. If we do, readjust.
68 */
69static int ceph_set_page_dirty(struct page *page)
70{
71 struct address_space *mapping = page->mapping;
72 struct inode *inode;
73 struct ceph_inode_info *ci;
74 struct ceph_snap_context *snapc;
75 int ret;
76
77 if (unlikely(!mapping))
78 return !TestSetPageDirty(page);
79
80 if (PageDirty(page)) {
81 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
82 mapping->host, page, page->index);
83 BUG_ON(!PagePrivate(page));
84 return 0;
85 }
86
87 inode = mapping->host;
88 ci = ceph_inode(inode);
89
90 /*
91 * Note that we're grabbing a snapc ref here without holding
92 * any locks!
93 */
94 snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
95
96 /* dirty the head */
97 spin_lock(&ci->i_ceph_lock);
98 if (ci->i_head_snapc == NULL)
99 ci->i_head_snapc = ceph_get_snap_context(snapc);
100 ++ci->i_wrbuffer_ref_head;
101 if (ci->i_wrbuffer_ref == 0)
102 ihold(inode);
103 ++ci->i_wrbuffer_ref;
104 dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
105 "snapc %p seq %lld (%d snaps)\n",
106 mapping->host, page, page->index,
107 ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
108 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
109 snapc, snapc->seq, snapc->num_snaps);
110 spin_unlock(&ci->i_ceph_lock);
111
112 /*
113 * Reference snap context in page->private. Also set
114 * PagePrivate so that we get invalidatepage callback.
115 */
116 BUG_ON(PagePrivate(page));
117 page->private = (unsigned long)snapc;
118 SetPagePrivate(page);
119
120 ret = __set_page_dirty_nobuffers(page);
121 WARN_ON(!PageLocked(page));
122 WARN_ON(!page->mapping);
123
124 return ret;
125}
126
127/*
128 * If we are truncating the full page (i.e. offset == 0), adjust the
129 * dirty page counters appropriately. Only called if there is private
130 * data on the page.
131 */
132static void ceph_invalidatepage(struct page *page, unsigned int offset,
133 unsigned int length)
134{
135 struct inode *inode;
136 struct ceph_inode_info *ci;
137 struct ceph_snap_context *snapc = page_snap_context(page);
138
139 inode = page->mapping->host;
140 ci = ceph_inode(inode);
141
142 if (offset != 0 || length != PAGE_CACHE_SIZE) {
143 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
144 inode, page, page->index, offset, length);
145 return;
146 }
147
148 ceph_invalidate_fscache_page(inode, page);
149
150 if (!PagePrivate(page))
151 return;
152
153 /*
154 * We can get non-dirty pages here due to races between
155 * set_page_dirty and truncate_complete_page; just spit out a
156 * warning, in case we end up with accounting problems later.
157 */
158 if (!PageDirty(page))
159 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
160
161 ClearPageChecked(page);
162
163 dout("%p invalidatepage %p idx %lu full dirty page\n",
164 inode, page, page->index);
165
166 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
167 ceph_put_snap_context(snapc);
168 page->private = 0;
169 ClearPagePrivate(page);
170}
171
172static int ceph_releasepage(struct page *page, gfp_t g)
173{
174 struct inode *inode = page->mapping ? page->mapping->host : NULL;
175 dout("%p releasepage %p idx %lu\n", inode, page, page->index);
176 WARN_ON(PageDirty(page));
177
178 /* Can we release the page from the cache? */
179 if (!ceph_release_fscache_page(page, g))
180 return 0;
181
182 return !PagePrivate(page);
183}
184
185/*
186 * read a single page, without unlocking it.
187 */
188static int readpage_nounlock(struct file *filp, struct page *page)
189{
190 struct inode *inode = file_inode(filp);
191 struct ceph_inode_info *ci = ceph_inode(inode);
192 struct ceph_osd_client *osdc =
193 &ceph_inode_to_client(inode)->client->osdc;
194 int err = 0;
195 u64 len = PAGE_CACHE_SIZE;
196
197 err = ceph_readpage_from_fscache(inode, page);
198
199 if (err == 0)
200 goto out;
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 (u64) page_offset(page), &len,
206 ci->i_truncate_seq, ci->i_truncate_size,
207 &page, 1, 0);
208 if (err == -ENOENT)
209 err = 0;
210 if (err < 0) {
211 SetPageError(page);
212 ceph_fscache_readpage_cancel(inode, page);
213 goto out;
214 } else {
215 if (err < PAGE_CACHE_SIZE) {
216 /* zero fill remainder of page */
217 zero_user_segment(page, err, PAGE_CACHE_SIZE);
218 } else {
219 flush_dcache_page(page);
220 }
221 }
222 SetPageUptodate(page);
223
224 if (err >= 0)
225 ceph_readpage_to_fscache(inode, page);
226
227out:
228 return err < 0 ? err : 0;
229}
230
231static int ceph_readpage(struct file *filp, struct page *page)
232{
233 int r = readpage_nounlock(filp, page);
234 unlock_page(page);
235 return r;
236}
237
238/*
239 * Finish an async read(ahead) op.
240 */
241static void finish_read(struct ceph_osd_request *req, struct ceph_msg *msg)
242{
243 struct inode *inode = req->r_inode;
244 struct ceph_osd_data *osd_data;
245 int rc = req->r_result;
246 int bytes = le32_to_cpu(msg->hdr.data_len);
247 int num_pages;
248 int i;
249
250 dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
251
252 /* unlock all pages, zeroing any data we didn't read */
253 osd_data = osd_req_op_extent_osd_data(req, 0);
254 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
255 num_pages = calc_pages_for((u64)osd_data->alignment,
256 (u64)osd_data->length);
257 for (i = 0; i < num_pages; i++) {
258 struct page *page = osd_data->pages[i];
259
260 if (rc < 0)
261 goto unlock;
262 if (bytes < (int)PAGE_CACHE_SIZE) {
263 /* zero (remainder of) page */
264 int s = bytes < 0 ? 0 : bytes;
265 zero_user_segment(page, s, PAGE_CACHE_SIZE);
266 }
267 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
268 page->index);
269 flush_dcache_page(page);
270 SetPageUptodate(page);
271 ceph_readpage_to_fscache(inode, page);
272unlock:
273 unlock_page(page);
274 page_cache_release(page);
275 bytes -= PAGE_CACHE_SIZE;
276 }
277 kfree(osd_data->pages);
278}
279
280static void ceph_unlock_page_vector(struct page **pages, int num_pages)
281{
282 int i;
283
284 for (i = 0; i < num_pages; i++)
285 unlock_page(pages[i]);
286}
287
288/*
289 * start an async read(ahead) operation. return nr_pages we submitted
290 * a read for on success, or negative error code.
291 */
292static int start_read(struct inode *inode, struct list_head *page_list, int max)
293{
294 struct ceph_osd_client *osdc =
295 &ceph_inode_to_client(inode)->client->osdc;
296 struct ceph_inode_info *ci = ceph_inode(inode);
297 struct page *page = list_entry(page_list->prev, struct page, lru);
298 struct ceph_vino vino;
299 struct ceph_osd_request *req;
300 u64 off;
301 u64 len;
302 int i;
303 struct page **pages;
304 pgoff_t next_index;
305 int nr_pages = 0;
306 int ret;
307
308 off = (u64) page_offset(page);
309
310 /* count pages */
311 next_index = page->index;
312 list_for_each_entry_reverse(page, page_list, lru) {
313 if (page->index != next_index)
314 break;
315 nr_pages++;
316 next_index++;
317 if (max && nr_pages == max)
318 break;
319 }
320 len = nr_pages << PAGE_CACHE_SHIFT;
321 dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
322 off, len);
323 vino = ceph_vino(inode);
324 req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
325 1, CEPH_OSD_OP_READ,
326 CEPH_OSD_FLAG_READ, NULL,
327 ci->i_truncate_seq, ci->i_truncate_size,
328 false);
329 if (IS_ERR(req))
330 return PTR_ERR(req);
331
332 /* build page vector */
333 nr_pages = calc_pages_for(0, len);
334 pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS);
335 ret = -ENOMEM;
336 if (!pages)
337 goto out;
338 for (i = 0; i < nr_pages; ++i) {
339 page = list_entry(page_list->prev, struct page, lru);
340 BUG_ON(PageLocked(page));
341 list_del(&page->lru);
342
343 dout("start_read %p adding %p idx %lu\n", inode, page,
344 page->index);
345 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
346 GFP_NOFS)) {
347 ceph_fscache_uncache_page(inode, page);
348 page_cache_release(page);
349 dout("start_read %p add_to_page_cache failed %p\n",
350 inode, page);
351 nr_pages = i;
352 goto out_pages;
353 }
354 pages[i] = page;
355 }
356 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
357 req->r_callback = finish_read;
358 req->r_inode = inode;
359
360 ceph_osdc_build_request(req, off, NULL, vino.snap, NULL);
361
362 dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
363 ret = ceph_osdc_start_request(osdc, req, false);
364 if (ret < 0)
365 goto out_pages;
366 ceph_osdc_put_request(req);
367 return nr_pages;
368
369out_pages:
370 ceph_unlock_page_vector(pages, nr_pages);
371 ceph_release_page_vector(pages, nr_pages);
372out:
373 ceph_osdc_put_request(req);
374 return ret;
375}
376
377
378/*
379 * Read multiple pages. Leave pages we don't read + unlock in page_list;
380 * the caller (VM) cleans them up.
381 */
382static int ceph_readpages(struct file *file, struct address_space *mapping,
383 struct list_head *page_list, unsigned nr_pages)
384{
385 struct inode *inode = file_inode(file);
386 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
387 int rc = 0;
388 int max = 0;
389
390 rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
391 &nr_pages);
392
393 if (rc == 0)
394 goto out;
395
396 if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE)
397 max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1)
398 >> PAGE_SHIFT;
399
400 dout("readpages %p file %p nr_pages %d max %d\n", inode,
401 file, nr_pages,
402 max);
403 while (!list_empty(page_list)) {
404 rc = start_read(inode, page_list, max);
405 if (rc < 0)
406 goto out;
407 BUG_ON(rc == 0);
408 }
409out:
410 ceph_fscache_readpages_cancel(inode, page_list);
411
412 dout("readpages %p file %p ret %d\n", inode, file, rc);
413 return rc;
414}
415
416/*
417 * Get ref for the oldest snapc for an inode with dirty data... that is, the
418 * only snap context we are allowed to write back.
419 */
420static struct ceph_snap_context *get_oldest_context(struct inode *inode,
421 u64 *snap_size)
422{
423 struct ceph_inode_info *ci = ceph_inode(inode);
424 struct ceph_snap_context *snapc = NULL;
425 struct ceph_cap_snap *capsnap = NULL;
426
427 spin_lock(&ci->i_ceph_lock);
428 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
429 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
430 capsnap->context, capsnap->dirty_pages);
431 if (capsnap->dirty_pages) {
432 snapc = ceph_get_snap_context(capsnap->context);
433 if (snap_size)
434 *snap_size = capsnap->size;
435 break;
436 }
437 }
438 if (!snapc && ci->i_wrbuffer_ref_head) {
439 snapc = ceph_get_snap_context(ci->i_head_snapc);
440 dout(" head snapc %p has %d dirty pages\n",
441 snapc, ci->i_wrbuffer_ref_head);
442 }
443 spin_unlock(&ci->i_ceph_lock);
444 return snapc;
445}
446
447/*
448 * Write a single page, but leave the page locked.
449 *
450 * If we get a write error, set the page error bit, but still adjust the
451 * dirty page accounting (i.e., page is no longer dirty).
452 */
453static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
454{
455 struct inode *inode;
456 struct ceph_inode_info *ci;
457 struct ceph_fs_client *fsc;
458 struct ceph_osd_client *osdc;
459 struct ceph_snap_context *snapc, *oldest;
460 loff_t page_off = page_offset(page);
461 long writeback_stat;
462 u64 truncate_size, snap_size = 0;
463 u32 truncate_seq;
464 int err = 0, len = PAGE_CACHE_SIZE;
465
466 dout("writepage %p idx %lu\n", page, page->index);
467
468 if (!page->mapping || !page->mapping->host) {
469 dout("writepage %p - no mapping\n", page);
470 return -EFAULT;
471 }
472 inode = page->mapping->host;
473 ci = ceph_inode(inode);
474 fsc = ceph_inode_to_client(inode);
475 osdc = &fsc->client->osdc;
476
477 /* verify this is a writeable snap context */
478 snapc = page_snap_context(page);
479 if (snapc == NULL) {
480 dout("writepage %p page %p not dirty?\n", inode, page);
481 goto out;
482 }
483 oldest = get_oldest_context(inode, &snap_size);
484 if (snapc->seq > oldest->seq) {
485 dout("writepage %p page %p snapc %p not writeable - noop\n",
486 inode, page, snapc);
487 /* we should only noop if called by kswapd */
488 WARN_ON((current->flags & PF_MEMALLOC) == 0);
489 ceph_put_snap_context(oldest);
490 goto out;
491 }
492 ceph_put_snap_context(oldest);
493
494 spin_lock(&ci->i_ceph_lock);
495 truncate_seq = ci->i_truncate_seq;
496 truncate_size = ci->i_truncate_size;
497 if (!snap_size)
498 snap_size = i_size_read(inode);
499 spin_unlock(&ci->i_ceph_lock);
500
501 /* is this a partial page at end of file? */
502 if (page_off >= snap_size) {
503 dout("%p page eof %llu\n", page, snap_size);
504 goto out;
505 }
506 if (snap_size < page_off + len)
507 len = snap_size - page_off;
508
509 dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
510 inode, page, page->index, page_off, len, snapc);
511
512 writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
513 if (writeback_stat >
514 CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
515 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
516
517 ceph_readpage_to_fscache(inode, page);
518
519 set_page_writeback(page);
520 err = ceph_osdc_writepages(osdc, ceph_vino(inode),
521 &ci->i_layout, snapc,
522 page_off, len,
523 truncate_seq, truncate_size,
524 &inode->i_mtime, &page, 1);
525 if (err < 0) {
526 dout("writepage setting page/mapping error %d %p\n", err, page);
527 SetPageError(page);
528 mapping_set_error(&inode->i_data, err);
529 if (wbc)
530 wbc->pages_skipped++;
531 } else {
532 dout("writepage cleaned page %p\n", page);
533 err = 0; /* vfs expects us to return 0 */
534 }
535 page->private = 0;
536 ClearPagePrivate(page);
537 end_page_writeback(page);
538 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
539 ceph_put_snap_context(snapc); /* page's reference */
540out:
541 return err;
542}
543
544static int ceph_writepage(struct page *page, struct writeback_control *wbc)
545{
546 int err;
547 struct inode *inode = page->mapping->host;
548 BUG_ON(!inode);
549 ihold(inode);
550 err = writepage_nounlock(page, wbc);
551 unlock_page(page);
552 iput(inode);
553 return err;
554}
555
556
557/*
558 * lame release_pages helper. release_pages() isn't exported to
559 * modules.
560 */
561static void ceph_release_pages(struct page **pages, int num)
562{
563 struct pagevec pvec;
564 int i;
565
566 pagevec_init(&pvec, 0);
567 for (i = 0; i < num; i++) {
568 if (pagevec_add(&pvec, pages[i]) == 0)
569 pagevec_release(&pvec);
570 }
571 pagevec_release(&pvec);
572}
573
574/*
575 * async writeback completion handler.
576 *
577 * If we get an error, set the mapping error bit, but not the individual
578 * page error bits.
579 */
580static void writepages_finish(struct ceph_osd_request *req,
581 struct ceph_msg *msg)
582{
583 struct inode *inode = req->r_inode;
584 struct ceph_inode_info *ci = ceph_inode(inode);
585 struct ceph_osd_data *osd_data;
586 unsigned wrote;
587 struct page *page;
588 int num_pages;
589 int i;
590 struct ceph_snap_context *snapc = req->r_snapc;
591 struct address_space *mapping = inode->i_mapping;
592 int rc = req->r_result;
593 u64 bytes = req->r_ops[0].extent.length;
594 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
595 long writeback_stat;
596 unsigned issued = ceph_caps_issued(ci);
597
598 osd_data = osd_req_op_extent_osd_data(req, 0);
599 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
600 num_pages = calc_pages_for((u64)osd_data->alignment,
601 (u64)osd_data->length);
602 if (rc >= 0) {
603 /*
604 * Assume we wrote the pages we originally sent. The
605 * osd might reply with fewer pages if our writeback
606 * raced with a truncation and was adjusted at the osd,
607 * so don't believe the reply.
608 */
609 wrote = num_pages;
610 } else {
611 wrote = 0;
612 mapping_set_error(mapping, rc);
613 }
614 dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
615 inode, rc, bytes, wrote);
616
617 /* clean all pages */
618 for (i = 0; i < num_pages; i++) {
619 page = osd_data->pages[i];
620 BUG_ON(!page);
621 WARN_ON(!PageUptodate(page));
622
623 writeback_stat =
624 atomic_long_dec_return(&fsc->writeback_count);
625 if (writeback_stat <
626 CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
627 clear_bdi_congested(&fsc->backing_dev_info,
628 BLK_RW_ASYNC);
629
630 ceph_put_snap_context(page_snap_context(page));
631 page->private = 0;
632 ClearPagePrivate(page);
633 dout("unlocking %d %p\n", i, page);
634 end_page_writeback(page);
635
636 /*
637 * We lost the cache cap, need to truncate the page before
638 * it is unlocked, otherwise we'd truncate it later in the
639 * page truncation thread, possibly losing some data that
640 * raced its way in
641 */
642 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
643 generic_error_remove_page(inode->i_mapping, page);
644
645 unlock_page(page);
646 }
647 dout("%p wrote+cleaned %d pages\n", inode, wrote);
648 ceph_put_wrbuffer_cap_refs(ci, num_pages, snapc);
649
650 ceph_release_pages(osd_data->pages, num_pages);
651 if (osd_data->pages_from_pool)
652 mempool_free(osd_data->pages,
653 ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
654 else
655 kfree(osd_data->pages);
656 ceph_osdc_put_request(req);
657}
658
659/*
660 * initiate async writeback
661 */
662static int ceph_writepages_start(struct address_space *mapping,
663 struct writeback_control *wbc)
664{
665 struct inode *inode = mapping->host;
666 struct ceph_inode_info *ci = ceph_inode(inode);
667 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
668 struct ceph_vino vino = ceph_vino(inode);
669 pgoff_t index, start, end;
670 int range_whole = 0;
671 int should_loop = 1;
672 pgoff_t max_pages = 0, max_pages_ever = 0;
673 struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
674 struct pagevec pvec;
675 int done = 0;
676 int rc = 0;
677 unsigned wsize = 1 << inode->i_blkbits;
678 struct ceph_osd_request *req = NULL;
679 int do_sync;
680 u64 truncate_size, snap_size;
681 u32 truncate_seq;
682
683 /*
684 * Include a 'sync' in the OSD request if this is a data
685 * integrity write (e.g., O_SYNC write or fsync()), or if our
686 * cap is being revoked.
687 */
688 if ((wbc->sync_mode == WB_SYNC_ALL) ||
689 ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
690 do_sync = 1;
691 dout("writepages_start %p dosync=%d (mode=%s)\n",
692 inode, do_sync,
693 wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
694 (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
695
696 if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
697 pr_warning("writepage_start %p on forced umount\n", inode);
698 return -EIO; /* we're in a forced umount, don't write! */
699 }
700 if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
701 wsize = fsc->mount_options->wsize;
702 if (wsize < PAGE_CACHE_SIZE)
703 wsize = PAGE_CACHE_SIZE;
704 max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
705
706 pagevec_init(&pvec, 0);
707
708 /* where to start/end? */
709 if (wbc->range_cyclic) {
710 start = mapping->writeback_index; /* Start from prev offset */
711 end = -1;
712 dout(" cyclic, start at %lu\n", start);
713 } else {
714 start = wbc->range_start >> PAGE_CACHE_SHIFT;
715 end = wbc->range_end >> PAGE_CACHE_SHIFT;
716 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
717 range_whole = 1;
718 should_loop = 0;
719 dout(" not cyclic, %lu to %lu\n", start, end);
720 }
721 index = start;
722
723retry:
724 /* find oldest snap context with dirty data */
725 ceph_put_snap_context(snapc);
726 snap_size = 0;
727 snapc = get_oldest_context(inode, &snap_size);
728 if (!snapc) {
729 /* hmm, why does writepages get called when there
730 is no dirty data? */
731 dout(" no snap context with dirty data?\n");
732 goto out;
733 }
734 if (snap_size == 0)
735 snap_size = i_size_read(inode);
736 dout(" oldest snapc is %p seq %lld (%d snaps)\n",
737 snapc, snapc->seq, snapc->num_snaps);
738
739 spin_lock(&ci->i_ceph_lock);
740 truncate_seq = ci->i_truncate_seq;
741 truncate_size = ci->i_truncate_size;
742 if (!snap_size)
743 snap_size = i_size_read(inode);
744 spin_unlock(&ci->i_ceph_lock);
745
746 if (last_snapc && snapc != last_snapc) {
747 /* if we switched to a newer snapc, restart our scan at the
748 * start of the original file range. */
749 dout(" snapc differs from last pass, restarting at %lu\n",
750 index);
751 index = start;
752 }
753 last_snapc = snapc;
754
755 while (!done && index <= end) {
756 int num_ops = do_sync ? 2 : 1;
757 unsigned i;
758 int first;
759 pgoff_t next;
760 int pvec_pages, locked_pages;
761 struct page **pages = NULL;
762 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
763 struct page *page;
764 int want;
765 u64 offset, len;
766 long writeback_stat;
767
768 next = 0;
769 locked_pages = 0;
770 max_pages = max_pages_ever;
771
772get_more_pages:
773 first = -1;
774 want = min(end - index,
775 min((pgoff_t)PAGEVEC_SIZE,
776 max_pages - (pgoff_t)locked_pages) - 1)
777 + 1;
778 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
779 PAGECACHE_TAG_DIRTY,
780 want);
781 dout("pagevec_lookup_tag got %d\n", pvec_pages);
782 if (!pvec_pages && !locked_pages)
783 break;
784 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
785 page = pvec.pages[i];
786 dout("? %p idx %lu\n", page, page->index);
787 if (locked_pages == 0)
788 lock_page(page); /* first page */
789 else if (!trylock_page(page))
790 break;
791
792 /* only dirty pages, or our accounting breaks */
793 if (unlikely(!PageDirty(page)) ||
794 unlikely(page->mapping != mapping)) {
795 dout("!dirty or !mapping %p\n", page);
796 unlock_page(page);
797 break;
798 }
799 if (!wbc->range_cyclic && page->index > end) {
800 dout("end of range %p\n", page);
801 done = 1;
802 unlock_page(page);
803 break;
804 }
805 if (next && (page->index != next)) {
806 dout("not consecutive %p\n", page);
807 unlock_page(page);
808 break;
809 }
810 if (wbc->sync_mode != WB_SYNC_NONE) {
811 dout("waiting on writeback %p\n", page);
812 wait_on_page_writeback(page);
813 }
814 if (page_offset(page) >= snap_size) {
815 dout("%p page eof %llu\n", page, snap_size);
816 done = 1;
817 unlock_page(page);
818 break;
819 }
820 if (PageWriteback(page)) {
821 dout("%p under writeback\n", page);
822 unlock_page(page);
823 break;
824 }
825
826 /* only if matching snap context */
827 pgsnapc = page_snap_context(page);
828 if (pgsnapc->seq > snapc->seq) {
829 dout("page snapc %p %lld > oldest %p %lld\n",
830 pgsnapc, pgsnapc->seq, snapc, snapc->seq);
831 unlock_page(page);
832 if (!locked_pages)
833 continue; /* keep looking for snap */
834 break;
835 }
836
837 if (!clear_page_dirty_for_io(page)) {
838 dout("%p !clear_page_dirty_for_io\n", page);
839 unlock_page(page);
840 break;
841 }
842
843 /*
844 * We have something to write. If this is
845 * the first locked page this time through,
846 * allocate an osd request and a page array
847 * that it will use.
848 */
849 if (locked_pages == 0) {
850 BUG_ON(pages);
851 /* prepare async write request */
852 offset = (u64)page_offset(page);
853 len = wsize;
854 req = ceph_osdc_new_request(&fsc->client->osdc,
855 &ci->i_layout, vino,
856 offset, &len, num_ops,
857 CEPH_OSD_OP_WRITE,
858 CEPH_OSD_FLAG_WRITE |
859 CEPH_OSD_FLAG_ONDISK,
860 snapc, truncate_seq,
861 truncate_size, true);
862 if (IS_ERR(req)) {
863 rc = PTR_ERR(req);
864 unlock_page(page);
865 break;
866 }
867
868 req->r_callback = writepages_finish;
869 req->r_inode = inode;
870
871 max_pages = calc_pages_for(0, (u64)len);
872 pages = kmalloc(max_pages * sizeof (*pages),
873 GFP_NOFS);
874 if (!pages) {
875 pool = fsc->wb_pagevec_pool;
876 pages = mempool_alloc(pool, GFP_NOFS);
877 BUG_ON(!pages);
878 }
879 }
880
881 /* note position of first page in pvec */
882 if (first < 0)
883 first = i;
884 dout("%p will write page %p idx %lu\n",
885 inode, page, page->index);
886
887 writeback_stat =
888 atomic_long_inc_return(&fsc->writeback_count);
889 if (writeback_stat > CONGESTION_ON_THRESH(
890 fsc->mount_options->congestion_kb)) {
891 set_bdi_congested(&fsc->backing_dev_info,
892 BLK_RW_ASYNC);
893 }
894
895 set_page_writeback(page);
896 pages[locked_pages] = page;
897 locked_pages++;
898 next = page->index + 1;
899 }
900
901 /* did we get anything? */
902 if (!locked_pages)
903 goto release_pvec_pages;
904 if (i) {
905 int j;
906 BUG_ON(!locked_pages || first < 0);
907
908 if (pvec_pages && i == pvec_pages &&
909 locked_pages < max_pages) {
910 dout("reached end pvec, trying for more\n");
911 pagevec_reinit(&pvec);
912 goto get_more_pages;
913 }
914
915 /* shift unused pages over in the pvec... we
916 * will need to release them below. */
917 for (j = i; j < pvec_pages; j++) {
918 dout(" pvec leftover page %p\n",
919 pvec.pages[j]);
920 pvec.pages[j-i+first] = pvec.pages[j];
921 }
922 pvec.nr -= i-first;
923 }
924
925 /* Format the osd request message and submit the write */
926
927 offset = page_offset(pages[0]);
928 len = min(snap_size - offset,
929 (u64)locked_pages << PAGE_CACHE_SHIFT);
930 dout("writepages got %d pages at %llu~%llu\n",
931 locked_pages, offset, len);
932
933 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
934 !!pool, false);
935
936 pages = NULL; /* request message now owns the pages array */
937 pool = NULL;
938
939 /* Update the write op length in case we changed it */
940
941 osd_req_op_extent_update(req, 0, len);
942
943 vino = ceph_vino(inode);
944 ceph_osdc_build_request(req, offset, snapc, vino.snap,
945 &inode->i_mtime);
946
947 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
948 BUG_ON(rc);
949 req = NULL;
950
951 /* continue? */
952 index = next;
953 wbc->nr_to_write -= locked_pages;
954 if (wbc->nr_to_write <= 0)
955 done = 1;
956
957release_pvec_pages:
958 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
959 pvec.nr ? pvec.pages[0] : NULL);
960 pagevec_release(&pvec);
961
962 if (locked_pages && !done)
963 goto retry;
964 }
965
966 if (should_loop && !done) {
967 /* more to do; loop back to beginning of file */
968 dout("writepages looping back to beginning of file\n");
969 should_loop = 0;
970 index = 0;
971 goto retry;
972 }
973
974 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
975 mapping->writeback_index = index;
976
977out:
978 if (req)
979 ceph_osdc_put_request(req);
980 ceph_put_snap_context(snapc);
981 dout("writepages done, rc = %d\n", rc);
982 return rc;
983}
984
985
986
987/*
988 * See if a given @snapc is either writeable, or already written.
989 */
990static int context_is_writeable_or_written(struct inode *inode,
991 struct ceph_snap_context *snapc)
992{
993 struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
994 int ret = !oldest || snapc->seq <= oldest->seq;
995
996 ceph_put_snap_context(oldest);
997 return ret;
998}
999
1000/*
1001 * We are only allowed to write into/dirty the page if the page is
1002 * clean, or already dirty within the same snap context.
1003 *
1004 * called with page locked.
1005 * return success with page locked,
1006 * or any failure (incl -EAGAIN) with page unlocked.
1007 */
1008static int ceph_update_writeable_page(struct file *file,
1009 loff_t pos, unsigned len,
1010 struct page *page)
1011{
1012 struct inode *inode = file_inode(file);
1013 struct ceph_inode_info *ci = ceph_inode(inode);
1014 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1015 loff_t page_off = pos & PAGE_CACHE_MASK;
1016 int pos_in_page = pos & ~PAGE_CACHE_MASK;
1017 int end_in_page = pos_in_page + len;
1018 loff_t i_size;
1019 int r;
1020 struct ceph_snap_context *snapc, *oldest;
1021
1022retry_locked:
1023 /* writepages currently holds page lock, but if we change that later, */
1024 wait_on_page_writeback(page);
1025
1026 /* check snap context */
1027 BUG_ON(!ci->i_snap_realm);
1028 down_read(&mdsc->snap_rwsem);
1029 BUG_ON(!ci->i_snap_realm->cached_context);
1030 snapc = page_snap_context(page);
1031 if (snapc && snapc != ci->i_head_snapc) {
1032 /*
1033 * this page is already dirty in another (older) snap
1034 * context! is it writeable now?
1035 */
1036 oldest = get_oldest_context(inode, NULL);
1037 up_read(&mdsc->snap_rwsem);
1038
1039 if (snapc->seq > oldest->seq) {
1040 ceph_put_snap_context(oldest);
1041 dout(" page %p snapc %p not current or oldest\n",
1042 page, snapc);
1043 /*
1044 * queue for writeback, and wait for snapc to
1045 * be writeable or written
1046 */
1047 snapc = ceph_get_snap_context(snapc);
1048 unlock_page(page);
1049 ceph_queue_writeback(inode);
1050 r = wait_event_interruptible(ci->i_cap_wq,
1051 context_is_writeable_or_written(inode, snapc));
1052 ceph_put_snap_context(snapc);
1053 if (r == -ERESTARTSYS)
1054 return r;
1055 return -EAGAIN;
1056 }
1057 ceph_put_snap_context(oldest);
1058
1059 /* yay, writeable, do it now (without dropping page lock) */
1060 dout(" page %p snapc %p not current, but oldest\n",
1061 page, snapc);
1062 if (!clear_page_dirty_for_io(page))
1063 goto retry_locked;
1064 r = writepage_nounlock(page, NULL);
1065 if (r < 0)
1066 goto fail_nosnap;
1067 goto retry_locked;
1068 }
1069
1070 if (PageUptodate(page)) {
1071 dout(" page %p already uptodate\n", page);
1072 return 0;
1073 }
1074
1075 /* full page? */
1076 if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
1077 return 0;
1078
1079 /* past end of file? */
1080 i_size = inode->i_size; /* caller holds i_mutex */
1081
1082 if (i_size + len > inode->i_sb->s_maxbytes) {
1083 /* file is too big */
1084 r = -EINVAL;
1085 goto fail;
1086 }
1087
1088 if (page_off >= i_size ||
1089 (pos_in_page == 0 && (pos+len) >= i_size &&
1090 end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1091 dout(" zeroing %p 0 - %d and %d - %d\n",
1092 page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1093 zero_user_segments(page,
1094 0, pos_in_page,
1095 end_in_page, PAGE_CACHE_SIZE);
1096 return 0;
1097 }
1098
1099 /* we need to read it. */
1100 up_read(&mdsc->snap_rwsem);
1101 r = readpage_nounlock(file, page);
1102 if (r < 0)
1103 goto fail_nosnap;
1104 goto retry_locked;
1105
1106fail:
1107 up_read(&mdsc->snap_rwsem);
1108fail_nosnap:
1109 unlock_page(page);
1110 return r;
1111}
1112
1113/*
1114 * We are only allowed to write into/dirty the page if the page is
1115 * clean, or already dirty within the same snap context.
1116 */
1117static int ceph_write_begin(struct file *file, struct address_space *mapping,
1118 loff_t pos, unsigned len, unsigned flags,
1119 struct page **pagep, void **fsdata)
1120{
1121 struct inode *inode = file_inode(file);
1122 struct page *page;
1123 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1124 int r;
1125
1126 do {
1127 /* get a page */
1128 page = grab_cache_page_write_begin(mapping, index, 0);
1129 if (!page)
1130 return -ENOMEM;
1131 *pagep = page;
1132
1133 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1134 inode, page, (int)pos, (int)len);
1135
1136 r = ceph_update_writeable_page(file, pos, len, page);
1137 } while (r == -EAGAIN);
1138
1139 return r;
1140}
1141
1142/*
1143 * we don't do anything in here that simple_write_end doesn't do
1144 * except adjust dirty page accounting and drop read lock on
1145 * mdsc->snap_rwsem.
1146 */
1147static int ceph_write_end(struct file *file, struct address_space *mapping,
1148 loff_t pos, unsigned len, unsigned copied,
1149 struct page *page, void *fsdata)
1150{
1151 struct inode *inode = file_inode(file);
1152 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1153 struct ceph_mds_client *mdsc = fsc->mdsc;
1154 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1155 int check_cap = 0;
1156
1157 dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1158 inode, page, (int)pos, (int)copied, (int)len);
1159
1160 /* zero the stale part of the page if we did a short copy */
1161 if (copied < len)
1162 zero_user_segment(page, from+copied, len);
1163
1164 /* did file size increase? */
1165 /* (no need for i_size_read(); we caller holds i_mutex */
1166 if (pos+copied > inode->i_size)
1167 check_cap = ceph_inode_set_size(inode, pos+copied);
1168
1169 if (!PageUptodate(page))
1170 SetPageUptodate(page);
1171
1172 set_page_dirty(page);
1173
1174 unlock_page(page);
1175 up_read(&mdsc->snap_rwsem);
1176 page_cache_release(page);
1177
1178 if (check_cap)
1179 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1180
1181 return copied;
1182}
1183
1184/*
1185 * we set .direct_IO to indicate direct io is supported, but since we
1186 * intercept O_DIRECT reads and writes early, this function should
1187 * never get called.
1188 */
1189static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1190 const struct iovec *iov,
1191 loff_t pos, unsigned long nr_segs)
1192{
1193 WARN_ON(1);
1194 return -EINVAL;
1195}
1196
1197const struct address_space_operations ceph_aops = {
1198 .readpage = ceph_readpage,
1199 .readpages = ceph_readpages,
1200 .writepage = ceph_writepage,
1201 .writepages = ceph_writepages_start,
1202 .write_begin = ceph_write_begin,
1203 .write_end = ceph_write_end,
1204 .set_page_dirty = ceph_set_page_dirty,
1205 .invalidatepage = ceph_invalidatepage,
1206 .releasepage = ceph_releasepage,
1207 .direct_IO = ceph_direct_io,
1208};
1209
1210
1211/*
1212 * vm ops
1213 */
1214static int ceph_filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1215{
1216 struct inode *inode = file_inode(vma->vm_file);
1217 struct ceph_inode_info *ci = ceph_inode(inode);
1218 struct ceph_file_info *fi = vma->vm_file->private_data;
1219 loff_t off = vmf->pgoff << PAGE_CACHE_SHIFT;
1220 int want, got, ret;
1221
1222 dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1223 inode, ceph_vinop(inode), off, (size_t)PAGE_CACHE_SIZE);
1224 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1225 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1226 else
1227 want = CEPH_CAP_FILE_CACHE;
1228 while (1) {
1229 got = 0;
1230 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, &got, -1);
1231 if (ret == 0)
1232 break;
1233 if (ret != -ERESTARTSYS) {
1234 WARN_ON(1);
1235 return VM_FAULT_SIGBUS;
1236 }
1237 }
1238 dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1239 inode, off, (size_t)PAGE_CACHE_SIZE, ceph_cap_string(got));
1240
1241 ret = filemap_fault(vma, vmf);
1242
1243 dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1244 inode, off, (size_t)PAGE_CACHE_SIZE, ceph_cap_string(got), ret);
1245 ceph_put_cap_refs(ci, got);
1246
1247 return ret;
1248}
1249
1250/*
1251 * Reuse write_begin here for simplicity.
1252 */
1253static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1254{
1255 struct inode *inode = file_inode(vma->vm_file);
1256 struct ceph_inode_info *ci = ceph_inode(inode);
1257 struct ceph_file_info *fi = vma->vm_file->private_data;
1258 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1259 struct page *page = vmf->page;
1260 loff_t off = page_offset(page);
1261 loff_t size = i_size_read(inode);
1262 size_t len;
1263 int want, got, ret;
1264
1265 if (off + PAGE_CACHE_SIZE <= size)
1266 len = PAGE_CACHE_SIZE;
1267 else
1268 len = size & ~PAGE_CACHE_MASK;
1269
1270 dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1271 inode, ceph_vinop(inode), off, len, size);
1272 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1273 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1274 else
1275 want = CEPH_CAP_FILE_BUFFER;
1276 while (1) {
1277 got = 0;
1278 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, &got, off + len);
1279 if (ret == 0)
1280 break;
1281 if (ret != -ERESTARTSYS) {
1282 WARN_ON(1);
1283 return VM_FAULT_SIGBUS;
1284 }
1285 }
1286 dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1287 inode, off, len, ceph_cap_string(got));
1288
1289 /* Update time before taking page lock */
1290 file_update_time(vma->vm_file);
1291
1292 lock_page(page);
1293
1294 ret = VM_FAULT_NOPAGE;
1295 if ((off > size) ||
1296 (page->mapping != inode->i_mapping))
1297 goto out;
1298
1299 ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1300 if (ret == 0) {
1301 /* success. we'll keep the page locked. */
1302 set_page_dirty(page);
1303 up_read(&mdsc->snap_rwsem);
1304 ret = VM_FAULT_LOCKED;
1305 } else {
1306 if (ret == -ENOMEM)
1307 ret = VM_FAULT_OOM;
1308 else
1309 ret = VM_FAULT_SIGBUS;
1310 }
1311out:
1312 if (ret != VM_FAULT_LOCKED) {
1313 unlock_page(page);
1314 } else {
1315 int dirty;
1316 spin_lock(&ci->i_ceph_lock);
1317 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR);
1318 spin_unlock(&ci->i_ceph_lock);
1319 if (dirty)
1320 __mark_inode_dirty(inode, dirty);
1321 }
1322
1323 dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1324 inode, off, len, ceph_cap_string(got), ret);
1325 ceph_put_cap_refs(ci, got);
1326
1327 return ret;
1328}
1329
1330static struct vm_operations_struct ceph_vmops = {
1331 .fault = ceph_filemap_fault,
1332 .page_mkwrite = ceph_page_mkwrite,
1333 .remap_pages = generic_file_remap_pages,
1334};
1335
1336int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1337{
1338 struct address_space *mapping = file->f_mapping;
1339
1340 if (!mapping->a_ops->readpage)
1341 return -ENOEXEC;
1342 file_accessed(file);
1343 vma->vm_ops = &ceph_vmops;
1344 return 0;
1345}
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#include <linux/ceph/striper.h>
19
20/*
21 * Ceph address space ops.
22 *
23 * There are a few funny things going on here.
24 *
25 * The page->private field is used to reference a struct
26 * ceph_snap_context for _every_ dirty page. This indicates which
27 * snapshot the page was logically dirtied in, and thus which snap
28 * context needs to be associated with the osd write during writeback.
29 *
30 * Similarly, struct ceph_inode_info maintains a set of counters to
31 * count dirty pages on the inode. In the absence of snapshots,
32 * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
33 *
34 * When a snapshot is taken (that is, when the client receives
35 * notification that a snapshot was taken), each inode with caps and
36 * with dirty pages (dirty pages implies there is a cap) gets a new
37 * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
38 * order, new snaps go to the tail). The i_wrbuffer_ref_head count is
39 * moved to capsnap->dirty. (Unless a sync write is currently in
40 * progress. In that case, the capsnap is said to be "pending", new
41 * writes cannot start, and the capsnap isn't "finalized" until the
42 * write completes (or fails) and a final size/mtime for the inode for
43 * that snap can be settled upon.) i_wrbuffer_ref_head is reset to 0.
44 *
45 * On writeback, we must submit writes to the osd IN SNAP ORDER. So,
46 * we look for the first capsnap in i_cap_snaps and write out pages in
47 * that snap context _only_. Then we move on to the next capsnap,
48 * eventually reaching the "live" or "head" context (i.e., pages that
49 * are not yet snapped) and are writing the most recently dirtied
50 * pages.
51 *
52 * Invalidate and so forth must take care to ensure the dirty page
53 * accounting is preserved.
54 */
55
56#define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
57#define CONGESTION_OFF_THRESH(congestion_kb) \
58 (CONGESTION_ON_THRESH(congestion_kb) - \
59 (CONGESTION_ON_THRESH(congestion_kb) >> 2))
60
61static inline struct ceph_snap_context *page_snap_context(struct page *page)
62{
63 if (PagePrivate(page))
64 return (void *)page->private;
65 return NULL;
66}
67
68/*
69 * Dirty a page. Optimistically adjust accounting, on the assumption
70 * that we won't race with invalidate. If we do, readjust.
71 */
72static int ceph_set_page_dirty(struct page *page)
73{
74 struct address_space *mapping = page->mapping;
75 struct inode *inode;
76 struct ceph_inode_info *ci;
77 struct ceph_snap_context *snapc;
78 int ret;
79
80 if (unlikely(!mapping))
81 return !TestSetPageDirty(page);
82
83 if (PageDirty(page)) {
84 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
85 mapping->host, page, page->index);
86 BUG_ON(!PagePrivate(page));
87 return 0;
88 }
89
90 inode = mapping->host;
91 ci = ceph_inode(inode);
92
93 /* dirty the head */
94 spin_lock(&ci->i_ceph_lock);
95 BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
96 if (__ceph_have_pending_cap_snap(ci)) {
97 struct ceph_cap_snap *capsnap =
98 list_last_entry(&ci->i_cap_snaps,
99 struct ceph_cap_snap,
100 ci_item);
101 snapc = ceph_get_snap_context(capsnap->context);
102 capsnap->dirty_pages++;
103 } else {
104 BUG_ON(!ci->i_head_snapc);
105 snapc = ceph_get_snap_context(ci->i_head_snapc);
106 ++ci->i_wrbuffer_ref_head;
107 }
108 if (ci->i_wrbuffer_ref == 0)
109 ihold(inode);
110 ++ci->i_wrbuffer_ref;
111 dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
112 "snapc %p seq %lld (%d snaps)\n",
113 mapping->host, page, page->index,
114 ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
115 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
116 snapc, snapc->seq, snapc->num_snaps);
117 spin_unlock(&ci->i_ceph_lock);
118
119 /*
120 * Reference snap context in page->private. Also set
121 * PagePrivate so that we get invalidatepage callback.
122 */
123 BUG_ON(PagePrivate(page));
124 page->private = (unsigned long)snapc;
125 SetPagePrivate(page);
126
127 ret = __set_page_dirty_nobuffers(page);
128 WARN_ON(!PageLocked(page));
129 WARN_ON(!page->mapping);
130
131 return ret;
132}
133
134/*
135 * If we are truncating the full page (i.e. offset == 0), adjust the
136 * dirty page counters appropriately. Only called if there is private
137 * data on the page.
138 */
139static void ceph_invalidatepage(struct page *page, unsigned int offset,
140 unsigned int length)
141{
142 struct inode *inode;
143 struct ceph_inode_info *ci;
144 struct ceph_snap_context *snapc = page_snap_context(page);
145
146 inode = page->mapping->host;
147 ci = ceph_inode(inode);
148
149 if (offset != 0 || length != PAGE_SIZE) {
150 dout("%p invalidatepage %p idx %lu partial dirty page %u~%u\n",
151 inode, page, page->index, offset, length);
152 return;
153 }
154
155 ceph_invalidate_fscache_page(inode, page);
156
157 WARN_ON(!PageLocked(page));
158 if (!PagePrivate(page))
159 return;
160
161 ClearPageChecked(page);
162
163 dout("%p invalidatepage %p idx %lu full dirty page\n",
164 inode, page, page->index);
165
166 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
167 ceph_put_snap_context(snapc);
168 page->private = 0;
169 ClearPagePrivate(page);
170}
171
172static int ceph_releasepage(struct page *page, gfp_t g)
173{
174 dout("%p releasepage %p idx %lu (%sdirty)\n", page->mapping->host,
175 page, page->index, PageDirty(page) ? "" : "not ");
176
177 /* Can we release the page from the cache? */
178 if (!ceph_release_fscache_page(page, g))
179 return 0;
180
181 return !PagePrivate(page);
182}
183
184/*
185 * read a single page, without unlocking it.
186 */
187static int ceph_do_readpage(struct file *filp, struct page *page)
188{
189 struct inode *inode = file_inode(filp);
190 struct ceph_inode_info *ci = ceph_inode(inode);
191 struct ceph_osd_client *osdc =
192 &ceph_inode_to_client(inode)->client->osdc;
193 int err = 0;
194 u64 off = page_offset(page);
195 u64 len = PAGE_SIZE;
196
197 if (off >= i_size_read(inode)) {
198 zero_user_segment(page, 0, PAGE_SIZE);
199 SetPageUptodate(page);
200 return 0;
201 }
202
203 if (ci->i_inline_version != CEPH_INLINE_NONE) {
204 /*
205 * Uptodate inline data should have been added
206 * into page cache while getting Fcr caps.
207 */
208 if (off == 0)
209 return -EINVAL;
210 zero_user_segment(page, 0, PAGE_SIZE);
211 SetPageUptodate(page);
212 return 0;
213 }
214
215 err = ceph_readpage_from_fscache(inode, page);
216 if (err == 0)
217 return -EINPROGRESS;
218
219 dout("readpage inode %p file %p page %p index %lu\n",
220 inode, filp, page, page->index);
221 err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
222 off, &len,
223 ci->i_truncate_seq, ci->i_truncate_size,
224 &page, 1, 0);
225 if (err == -ENOENT)
226 err = 0;
227 if (err < 0) {
228 SetPageError(page);
229 ceph_fscache_readpage_cancel(inode, page);
230 goto out;
231 }
232 if (err < PAGE_SIZE)
233 /* zero fill remainder of page */
234 zero_user_segment(page, err, PAGE_SIZE);
235 else
236 flush_dcache_page(page);
237
238 SetPageUptodate(page);
239 ceph_readpage_to_fscache(inode, page);
240
241out:
242 return err < 0 ? err : 0;
243}
244
245static int ceph_readpage(struct file *filp, struct page *page)
246{
247 int r = ceph_do_readpage(filp, page);
248 if (r != -EINPROGRESS)
249 unlock_page(page);
250 else
251 r = 0;
252 return r;
253}
254
255/*
256 * Finish an async read(ahead) op.
257 */
258static void finish_read(struct ceph_osd_request *req)
259{
260 struct inode *inode = req->r_inode;
261 struct ceph_osd_data *osd_data;
262 int rc = req->r_result <= 0 ? req->r_result : 0;
263 int bytes = req->r_result >= 0 ? req->r_result : 0;
264 int num_pages;
265 int i;
266
267 dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
268
269 /* unlock all pages, zeroing any data we didn't read */
270 osd_data = osd_req_op_extent_osd_data(req, 0);
271 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
272 num_pages = calc_pages_for((u64)osd_data->alignment,
273 (u64)osd_data->length);
274 for (i = 0; i < num_pages; i++) {
275 struct page *page = osd_data->pages[i];
276
277 if (rc < 0 && rc != -ENOENT) {
278 ceph_fscache_readpage_cancel(inode, page);
279 goto unlock;
280 }
281 if (bytes < (int)PAGE_SIZE) {
282 /* zero (remainder of) page */
283 int s = bytes < 0 ? 0 : bytes;
284 zero_user_segment(page, s, PAGE_SIZE);
285 }
286 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
287 page->index);
288 flush_dcache_page(page);
289 SetPageUptodate(page);
290 ceph_readpage_to_fscache(inode, page);
291unlock:
292 unlock_page(page);
293 put_page(page);
294 bytes -= PAGE_SIZE;
295 }
296 kfree(osd_data->pages);
297}
298
299/*
300 * start an async read(ahead) operation. return nr_pages we submitted
301 * a read for on success, or negative error code.
302 */
303static int start_read(struct inode *inode, struct ceph_rw_context *rw_ctx,
304 struct list_head *page_list, int max)
305{
306 struct ceph_osd_client *osdc =
307 &ceph_inode_to_client(inode)->client->osdc;
308 struct ceph_inode_info *ci = ceph_inode(inode);
309 struct page *page = list_entry(page_list->prev, struct page, lru);
310 struct ceph_vino vino;
311 struct ceph_osd_request *req;
312 u64 off;
313 u64 len;
314 int i;
315 struct page **pages;
316 pgoff_t next_index;
317 int nr_pages = 0;
318 int got = 0;
319 int ret = 0;
320
321 if (!rw_ctx) {
322 /* caller of readpages does not hold buffer and read caps
323 * (fadvise, madvise and readahead cases) */
324 int want = CEPH_CAP_FILE_CACHE;
325 ret = ceph_try_get_caps(ci, CEPH_CAP_FILE_RD, want, &got);
326 if (ret < 0) {
327 dout("start_read %p, error getting cap\n", inode);
328 } else if (!(got & want)) {
329 dout("start_read %p, no cache cap\n", inode);
330 ret = 0;
331 }
332 if (ret <= 0) {
333 if (got)
334 ceph_put_cap_refs(ci, got);
335 while (!list_empty(page_list)) {
336 page = list_entry(page_list->prev,
337 struct page, lru);
338 list_del(&page->lru);
339 put_page(page);
340 }
341 return ret;
342 }
343 }
344
345 off = (u64) page_offset(page);
346
347 /* count pages */
348 next_index = page->index;
349 list_for_each_entry_reverse(page, page_list, lru) {
350 if (page->index != next_index)
351 break;
352 nr_pages++;
353 next_index++;
354 if (max && nr_pages == max)
355 break;
356 }
357 len = nr_pages << PAGE_SHIFT;
358 dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
359 off, len);
360 vino = ceph_vino(inode);
361 req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
362 0, 1, CEPH_OSD_OP_READ,
363 CEPH_OSD_FLAG_READ, NULL,
364 ci->i_truncate_seq, ci->i_truncate_size,
365 false);
366 if (IS_ERR(req)) {
367 ret = PTR_ERR(req);
368 goto out;
369 }
370
371 /* build page vector */
372 nr_pages = calc_pages_for(0, len);
373 pages = kmalloc(sizeof(*pages) * nr_pages, GFP_KERNEL);
374 if (!pages) {
375 ret = -ENOMEM;
376 goto out_put;
377 }
378 for (i = 0; i < nr_pages; ++i) {
379 page = list_entry(page_list->prev, struct page, lru);
380 BUG_ON(PageLocked(page));
381 list_del(&page->lru);
382
383 dout("start_read %p adding %p idx %lu\n", inode, page,
384 page->index);
385 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
386 GFP_KERNEL)) {
387 ceph_fscache_uncache_page(inode, page);
388 put_page(page);
389 dout("start_read %p add_to_page_cache failed %p\n",
390 inode, page);
391 nr_pages = i;
392 if (nr_pages > 0) {
393 len = nr_pages << PAGE_SHIFT;
394 osd_req_op_extent_update(req, 0, len);
395 break;
396 }
397 goto out_pages;
398 }
399 pages[i] = page;
400 }
401 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false, false);
402 req->r_callback = finish_read;
403 req->r_inode = inode;
404
405 dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
406 ret = ceph_osdc_start_request(osdc, req, false);
407 if (ret < 0)
408 goto out_pages;
409 ceph_osdc_put_request(req);
410
411 /* After adding locked pages to page cache, the inode holds cache cap.
412 * So we can drop our cap refs. */
413 if (got)
414 ceph_put_cap_refs(ci, got);
415
416 return nr_pages;
417
418out_pages:
419 for (i = 0; i < nr_pages; ++i) {
420 ceph_fscache_readpage_cancel(inode, pages[i]);
421 unlock_page(pages[i]);
422 }
423 ceph_put_page_vector(pages, nr_pages, false);
424out_put:
425 ceph_osdc_put_request(req);
426out:
427 if (got)
428 ceph_put_cap_refs(ci, got);
429 return ret;
430}
431
432
433/*
434 * Read multiple pages. Leave pages we don't read + unlock in page_list;
435 * the caller (VM) cleans them up.
436 */
437static int ceph_readpages(struct file *file, struct address_space *mapping,
438 struct list_head *page_list, unsigned nr_pages)
439{
440 struct inode *inode = file_inode(file);
441 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
442 struct ceph_file_info *fi = file->private_data;
443 struct ceph_rw_context *rw_ctx;
444 int rc = 0;
445 int max = 0;
446
447 if (ceph_inode(inode)->i_inline_version != CEPH_INLINE_NONE)
448 return -EINVAL;
449
450 rc = ceph_readpages_from_fscache(mapping->host, mapping, page_list,
451 &nr_pages);
452
453 if (rc == 0)
454 goto out;
455
456 rw_ctx = ceph_find_rw_context(fi);
457 max = fsc->mount_options->rsize >> PAGE_SHIFT;
458 dout("readpages %p file %p ctx %p nr_pages %d max %d\n",
459 inode, file, rw_ctx, nr_pages, max);
460 while (!list_empty(page_list)) {
461 rc = start_read(inode, rw_ctx, page_list, max);
462 if (rc < 0)
463 goto out;
464 }
465out:
466 ceph_fscache_readpages_cancel(inode, page_list);
467
468 dout("readpages %p file %p ret %d\n", inode, file, rc);
469 return rc;
470}
471
472struct ceph_writeback_ctl
473{
474 loff_t i_size;
475 u64 truncate_size;
476 u32 truncate_seq;
477 bool size_stable;
478 bool head_snapc;
479};
480
481/*
482 * Get ref for the oldest snapc for an inode with dirty data... that is, the
483 * only snap context we are allowed to write back.
484 */
485static struct ceph_snap_context *
486get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
487 struct ceph_snap_context *page_snapc)
488{
489 struct ceph_inode_info *ci = ceph_inode(inode);
490 struct ceph_snap_context *snapc = NULL;
491 struct ceph_cap_snap *capsnap = NULL;
492
493 spin_lock(&ci->i_ceph_lock);
494 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
495 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
496 capsnap->context, capsnap->dirty_pages);
497 if (!capsnap->dirty_pages)
498 continue;
499
500 /* get i_size, truncate_{seq,size} for page_snapc? */
501 if (snapc && capsnap->context != page_snapc)
502 continue;
503
504 if (ctl) {
505 if (capsnap->writing) {
506 ctl->i_size = i_size_read(inode);
507 ctl->size_stable = false;
508 } else {
509 ctl->i_size = capsnap->size;
510 ctl->size_stable = true;
511 }
512 ctl->truncate_size = capsnap->truncate_size;
513 ctl->truncate_seq = capsnap->truncate_seq;
514 ctl->head_snapc = false;
515 }
516
517 if (snapc)
518 break;
519
520 snapc = ceph_get_snap_context(capsnap->context);
521 if (!page_snapc ||
522 page_snapc == snapc ||
523 page_snapc->seq > snapc->seq)
524 break;
525 }
526 if (!snapc && ci->i_wrbuffer_ref_head) {
527 snapc = ceph_get_snap_context(ci->i_head_snapc);
528 dout(" head snapc %p has %d dirty pages\n",
529 snapc, ci->i_wrbuffer_ref_head);
530 if (ctl) {
531 ctl->i_size = i_size_read(inode);
532 ctl->truncate_size = ci->i_truncate_size;
533 ctl->truncate_seq = ci->i_truncate_seq;
534 ctl->size_stable = false;
535 ctl->head_snapc = true;
536 }
537 }
538 spin_unlock(&ci->i_ceph_lock);
539 return snapc;
540}
541
542static u64 get_writepages_data_length(struct inode *inode,
543 struct page *page, u64 start)
544{
545 struct ceph_inode_info *ci = ceph_inode(inode);
546 struct ceph_snap_context *snapc = page_snap_context(page);
547 struct ceph_cap_snap *capsnap = NULL;
548 u64 end = i_size_read(inode);
549
550 if (snapc != ci->i_head_snapc) {
551 bool found = false;
552 spin_lock(&ci->i_ceph_lock);
553 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
554 if (capsnap->context == snapc) {
555 if (!capsnap->writing)
556 end = capsnap->size;
557 found = true;
558 break;
559 }
560 }
561 spin_unlock(&ci->i_ceph_lock);
562 WARN_ON(!found);
563 }
564 if (end > page_offset(page) + PAGE_SIZE)
565 end = page_offset(page) + PAGE_SIZE;
566 return end > start ? end - start : 0;
567}
568
569/*
570 * Write a single page, but leave the page locked.
571 *
572 * If we get a write error, set the page error bit, but still adjust the
573 * dirty page accounting (i.e., page is no longer dirty).
574 */
575static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
576{
577 struct inode *inode;
578 struct ceph_inode_info *ci;
579 struct ceph_fs_client *fsc;
580 struct ceph_snap_context *snapc, *oldest;
581 loff_t page_off = page_offset(page);
582 int err, len = PAGE_SIZE;
583 struct ceph_writeback_ctl ceph_wbc;
584
585 dout("writepage %p idx %lu\n", page, page->index);
586
587 inode = page->mapping->host;
588 ci = ceph_inode(inode);
589 fsc = ceph_inode_to_client(inode);
590
591 /* verify this is a writeable snap context */
592 snapc = page_snap_context(page);
593 if (!snapc) {
594 dout("writepage %p page %p not dirty?\n", inode, page);
595 return 0;
596 }
597 oldest = get_oldest_context(inode, &ceph_wbc, snapc);
598 if (snapc->seq > oldest->seq) {
599 dout("writepage %p page %p snapc %p not writeable - noop\n",
600 inode, page, snapc);
601 /* we should only noop if called by kswapd */
602 WARN_ON(!(current->flags & PF_MEMALLOC));
603 ceph_put_snap_context(oldest);
604 redirty_page_for_writepage(wbc, page);
605 return 0;
606 }
607 ceph_put_snap_context(oldest);
608
609 /* is this a partial page at end of file? */
610 if (page_off >= ceph_wbc.i_size) {
611 dout("%p page eof %llu\n", page, ceph_wbc.i_size);
612 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
613 return 0;
614 }
615
616 if (ceph_wbc.i_size < page_off + len)
617 len = ceph_wbc.i_size - page_off;
618
619 dout("writepage %p page %p index %lu on %llu~%u snapc %p seq %lld\n",
620 inode, page, page->index, page_off, len, snapc, snapc->seq);
621
622 if (atomic_long_inc_return(&fsc->writeback_count) >
623 CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
624 set_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
625
626 set_page_writeback(page);
627 err = ceph_osdc_writepages(&fsc->client->osdc, ceph_vino(inode),
628 &ci->i_layout, snapc, page_off, len,
629 ceph_wbc.truncate_seq,
630 ceph_wbc.truncate_size,
631 &inode->i_mtime, &page, 1);
632 if (err < 0) {
633 struct writeback_control tmp_wbc;
634 if (!wbc)
635 wbc = &tmp_wbc;
636 if (err == -ERESTARTSYS) {
637 /* killed by SIGKILL */
638 dout("writepage interrupted page %p\n", page);
639 redirty_page_for_writepage(wbc, page);
640 end_page_writeback(page);
641 return err;
642 }
643 dout("writepage setting page/mapping error %d %p\n",
644 err, page);
645 SetPageError(page);
646 mapping_set_error(&inode->i_data, err);
647 wbc->pages_skipped++;
648 } else {
649 dout("writepage cleaned page %p\n", page);
650 err = 0; /* vfs expects us to return 0 */
651 }
652 page->private = 0;
653 ClearPagePrivate(page);
654 end_page_writeback(page);
655 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
656 ceph_put_snap_context(snapc); /* page's reference */
657
658 if (atomic_long_dec_return(&fsc->writeback_count) <
659 CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
660 clear_bdi_congested(inode_to_bdi(inode), BLK_RW_ASYNC);
661
662 return err;
663}
664
665static int ceph_writepage(struct page *page, struct writeback_control *wbc)
666{
667 int err;
668 struct inode *inode = page->mapping->host;
669 BUG_ON(!inode);
670 ihold(inode);
671 err = writepage_nounlock(page, wbc);
672 if (err == -ERESTARTSYS) {
673 /* direct memory reclaimer was killed by SIGKILL. return 0
674 * to prevent caller from setting mapping/page error */
675 err = 0;
676 }
677 unlock_page(page);
678 iput(inode);
679 return err;
680}
681
682/*
683 * lame release_pages helper. release_pages() isn't exported to
684 * modules.
685 */
686static void ceph_release_pages(struct page **pages, int num)
687{
688 struct pagevec pvec;
689 int i;
690
691 pagevec_init(&pvec);
692 for (i = 0; i < num; i++) {
693 if (pagevec_add(&pvec, pages[i]) == 0)
694 pagevec_release(&pvec);
695 }
696 pagevec_release(&pvec);
697}
698
699/*
700 * async writeback completion handler.
701 *
702 * If we get an error, set the mapping error bit, but not the individual
703 * page error bits.
704 */
705static void writepages_finish(struct ceph_osd_request *req)
706{
707 struct inode *inode = req->r_inode;
708 struct ceph_inode_info *ci = ceph_inode(inode);
709 struct ceph_osd_data *osd_data;
710 struct page *page;
711 int num_pages, total_pages = 0;
712 int i, j;
713 int rc = req->r_result;
714 struct ceph_snap_context *snapc = req->r_snapc;
715 struct address_space *mapping = inode->i_mapping;
716 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
717 bool remove_page;
718
719 dout("writepages_finish %p rc %d\n", inode, rc);
720 if (rc < 0) {
721 mapping_set_error(mapping, rc);
722 ceph_set_error_write(ci);
723 } else {
724 ceph_clear_error_write(ci);
725 }
726
727 /*
728 * We lost the cache cap, need to truncate the page before
729 * it is unlocked, otherwise we'd truncate it later in the
730 * page truncation thread, possibly losing some data that
731 * raced its way in
732 */
733 remove_page = !(ceph_caps_issued(ci) &
734 (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
735
736 /* clean all pages */
737 for (i = 0; i < req->r_num_ops; i++) {
738 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE)
739 break;
740
741 osd_data = osd_req_op_extent_osd_data(req, i);
742 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
743 num_pages = calc_pages_for((u64)osd_data->alignment,
744 (u64)osd_data->length);
745 total_pages += num_pages;
746 for (j = 0; j < num_pages; j++) {
747 page = osd_data->pages[j];
748 BUG_ON(!page);
749 WARN_ON(!PageUptodate(page));
750
751 if (atomic_long_dec_return(&fsc->writeback_count) <
752 CONGESTION_OFF_THRESH(
753 fsc->mount_options->congestion_kb))
754 clear_bdi_congested(inode_to_bdi(inode),
755 BLK_RW_ASYNC);
756
757 ceph_put_snap_context(page_snap_context(page));
758 page->private = 0;
759 ClearPagePrivate(page);
760 dout("unlocking %p\n", page);
761 end_page_writeback(page);
762
763 if (remove_page)
764 generic_error_remove_page(inode->i_mapping,
765 page);
766
767 unlock_page(page);
768 }
769 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
770 inode, osd_data->length, rc >= 0 ? num_pages : 0);
771
772 ceph_release_pages(osd_data->pages, num_pages);
773 }
774
775 ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
776
777 osd_data = osd_req_op_extent_osd_data(req, 0);
778 if (osd_data->pages_from_pool)
779 mempool_free(osd_data->pages,
780 ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
781 else
782 kfree(osd_data->pages);
783 ceph_osdc_put_request(req);
784}
785
786/*
787 * initiate async writeback
788 */
789static int ceph_writepages_start(struct address_space *mapping,
790 struct writeback_control *wbc)
791{
792 struct inode *inode = mapping->host;
793 struct ceph_inode_info *ci = ceph_inode(inode);
794 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
795 struct ceph_vino vino = ceph_vino(inode);
796 pgoff_t index, start_index, end = -1;
797 struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
798 struct pagevec pvec;
799 int rc = 0;
800 unsigned int wsize = i_blocksize(inode);
801 struct ceph_osd_request *req = NULL;
802 struct ceph_writeback_ctl ceph_wbc;
803 bool should_loop, range_whole = false;
804 bool done = false;
805
806 dout("writepages_start %p (mode=%s)\n", inode,
807 wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
808 (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
809
810 if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
811 if (ci->i_wrbuffer_ref > 0) {
812 pr_warn_ratelimited(
813 "writepage_start %p %lld forced umount\n",
814 inode, ceph_ino(inode));
815 }
816 mapping_set_error(mapping, -EIO);
817 return -EIO; /* we're in a forced umount, don't write! */
818 }
819 if (fsc->mount_options->wsize < wsize)
820 wsize = fsc->mount_options->wsize;
821
822 pagevec_init(&pvec);
823
824 start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
825 index = start_index;
826
827retry:
828 /* find oldest snap context with dirty data */
829 snapc = get_oldest_context(inode, &ceph_wbc, NULL);
830 if (!snapc) {
831 /* hmm, why does writepages get called when there
832 is no dirty data? */
833 dout(" no snap context with dirty data?\n");
834 goto out;
835 }
836 dout(" oldest snapc is %p seq %lld (%d snaps)\n",
837 snapc, snapc->seq, snapc->num_snaps);
838
839 should_loop = false;
840 if (ceph_wbc.head_snapc && snapc != last_snapc) {
841 /* where to start/end? */
842 if (wbc->range_cyclic) {
843 index = start_index;
844 end = -1;
845 if (index > 0)
846 should_loop = true;
847 dout(" cyclic, start at %lu\n", index);
848 } else {
849 index = wbc->range_start >> PAGE_SHIFT;
850 end = wbc->range_end >> PAGE_SHIFT;
851 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
852 range_whole = true;
853 dout(" not cyclic, %lu to %lu\n", index, end);
854 }
855 } else if (!ceph_wbc.head_snapc) {
856 /* Do not respect wbc->range_{start,end}. Dirty pages
857 * in that range can be associated with newer snapc.
858 * They are not writeable until we write all dirty pages
859 * associated with 'snapc' get written */
860 if (index > 0)
861 should_loop = true;
862 dout(" non-head snapc, range whole\n");
863 }
864
865 ceph_put_snap_context(last_snapc);
866 last_snapc = snapc;
867
868 while (!done && index <= end) {
869 int num_ops = 0, op_idx;
870 unsigned i, pvec_pages, max_pages, locked_pages = 0;
871 struct page **pages = NULL, **data_pages;
872 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
873 struct page *page;
874 pgoff_t strip_unit_end = 0;
875 u64 offset = 0, len = 0;
876
877 max_pages = wsize >> PAGE_SHIFT;
878
879get_more_pages:
880 pvec_pages = pagevec_lookup_range_nr_tag(&pvec, mapping, &index,
881 end, PAGECACHE_TAG_DIRTY,
882 max_pages - locked_pages);
883 dout("pagevec_lookup_range_tag got %d\n", pvec_pages);
884 if (!pvec_pages && !locked_pages)
885 break;
886 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
887 page = pvec.pages[i];
888 dout("? %p idx %lu\n", page, page->index);
889 if (locked_pages == 0)
890 lock_page(page); /* first page */
891 else if (!trylock_page(page))
892 break;
893
894 /* only dirty pages, or our accounting breaks */
895 if (unlikely(!PageDirty(page)) ||
896 unlikely(page->mapping != mapping)) {
897 dout("!dirty or !mapping %p\n", page);
898 unlock_page(page);
899 continue;
900 }
901 /* only if matching snap context */
902 pgsnapc = page_snap_context(page);
903 if (pgsnapc != snapc) {
904 dout("page snapc %p %lld != oldest %p %lld\n",
905 pgsnapc, pgsnapc->seq, snapc, snapc->seq);
906 if (!should_loop &&
907 !ceph_wbc.head_snapc &&
908 wbc->sync_mode != WB_SYNC_NONE)
909 should_loop = true;
910 unlock_page(page);
911 continue;
912 }
913 if (page_offset(page) >= ceph_wbc.i_size) {
914 dout("%p page eof %llu\n",
915 page, ceph_wbc.i_size);
916 if (ceph_wbc.size_stable ||
917 page_offset(page) >= i_size_read(inode))
918 mapping->a_ops->invalidatepage(page,
919 0, PAGE_SIZE);
920 unlock_page(page);
921 continue;
922 }
923 if (strip_unit_end && (page->index > strip_unit_end)) {
924 dout("end of strip unit %p\n", page);
925 unlock_page(page);
926 break;
927 }
928 if (PageWriteback(page)) {
929 if (wbc->sync_mode == WB_SYNC_NONE) {
930 dout("%p under writeback\n", page);
931 unlock_page(page);
932 continue;
933 }
934 dout("waiting on writeback %p\n", page);
935 wait_on_page_writeback(page);
936 }
937
938 if (!clear_page_dirty_for_io(page)) {
939 dout("%p !clear_page_dirty_for_io\n", page);
940 unlock_page(page);
941 continue;
942 }
943
944 /*
945 * We have something to write. If this is
946 * the first locked page this time through,
947 * calculate max possinle write size and
948 * allocate a page array
949 */
950 if (locked_pages == 0) {
951 u64 objnum;
952 u64 objoff;
953 u32 xlen;
954
955 /* prepare async write request */
956 offset = (u64)page_offset(page);
957 ceph_calc_file_object_mapping(&ci->i_layout,
958 offset, wsize,
959 &objnum, &objoff,
960 &xlen);
961 len = xlen;
962
963 num_ops = 1;
964 strip_unit_end = page->index +
965 ((len - 1) >> PAGE_SHIFT);
966
967 BUG_ON(pages);
968 max_pages = calc_pages_for(0, (u64)len);
969 pages = kmalloc(max_pages * sizeof (*pages),
970 GFP_NOFS);
971 if (!pages) {
972 pool = fsc->wb_pagevec_pool;
973 pages = mempool_alloc(pool, GFP_NOFS);
974 BUG_ON(!pages);
975 }
976
977 len = 0;
978 } else if (page->index !=
979 (offset + len) >> PAGE_SHIFT) {
980 if (num_ops >= (pool ? CEPH_OSD_SLAB_OPS :
981 CEPH_OSD_MAX_OPS)) {
982 redirty_page_for_writepage(wbc, page);
983 unlock_page(page);
984 break;
985 }
986
987 num_ops++;
988 offset = (u64)page_offset(page);
989 len = 0;
990 }
991
992 /* note position of first page in pvec */
993 dout("%p will write page %p idx %lu\n",
994 inode, page, page->index);
995
996 if (atomic_long_inc_return(&fsc->writeback_count) >
997 CONGESTION_ON_THRESH(
998 fsc->mount_options->congestion_kb)) {
999 set_bdi_congested(inode_to_bdi(inode),
1000 BLK_RW_ASYNC);
1001 }
1002
1003
1004 pages[locked_pages++] = page;
1005 pvec.pages[i] = NULL;
1006
1007 len += PAGE_SIZE;
1008 }
1009
1010 /* did we get anything? */
1011 if (!locked_pages)
1012 goto release_pvec_pages;
1013 if (i) {
1014 unsigned j, n = 0;
1015 /* shift unused page to beginning of pvec */
1016 for (j = 0; j < pvec_pages; j++) {
1017 if (!pvec.pages[j])
1018 continue;
1019 if (n < j)
1020 pvec.pages[n] = pvec.pages[j];
1021 n++;
1022 }
1023 pvec.nr = n;
1024
1025 if (pvec_pages && i == pvec_pages &&
1026 locked_pages < max_pages) {
1027 dout("reached end pvec, trying for more\n");
1028 pagevec_release(&pvec);
1029 goto get_more_pages;
1030 }
1031 }
1032
1033new_request:
1034 offset = page_offset(pages[0]);
1035 len = wsize;
1036
1037 req = ceph_osdc_new_request(&fsc->client->osdc,
1038 &ci->i_layout, vino,
1039 offset, &len, 0, num_ops,
1040 CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1041 snapc, ceph_wbc.truncate_seq,
1042 ceph_wbc.truncate_size, false);
1043 if (IS_ERR(req)) {
1044 req = ceph_osdc_new_request(&fsc->client->osdc,
1045 &ci->i_layout, vino,
1046 offset, &len, 0,
1047 min(num_ops,
1048 CEPH_OSD_SLAB_OPS),
1049 CEPH_OSD_OP_WRITE,
1050 CEPH_OSD_FLAG_WRITE,
1051 snapc, ceph_wbc.truncate_seq,
1052 ceph_wbc.truncate_size, true);
1053 BUG_ON(IS_ERR(req));
1054 }
1055 BUG_ON(len < page_offset(pages[locked_pages - 1]) +
1056 PAGE_SIZE - offset);
1057
1058 req->r_callback = writepages_finish;
1059 req->r_inode = inode;
1060
1061 /* Format the osd request message and submit the write */
1062 len = 0;
1063 data_pages = pages;
1064 op_idx = 0;
1065 for (i = 0; i < locked_pages; i++) {
1066 u64 cur_offset = page_offset(pages[i]);
1067 if (offset + len != cur_offset) {
1068 if (op_idx + 1 == req->r_num_ops)
1069 break;
1070 osd_req_op_extent_dup_last(req, op_idx,
1071 cur_offset - offset);
1072 dout("writepages got pages at %llu~%llu\n",
1073 offset, len);
1074 osd_req_op_extent_osd_data_pages(req, op_idx,
1075 data_pages, len, 0,
1076 !!pool, false);
1077 osd_req_op_extent_update(req, op_idx, len);
1078
1079 len = 0;
1080 offset = cur_offset;
1081 data_pages = pages + i;
1082 op_idx++;
1083 }
1084
1085 set_page_writeback(pages[i]);
1086 len += PAGE_SIZE;
1087 }
1088
1089 if (ceph_wbc.size_stable) {
1090 len = min(len, ceph_wbc.i_size - offset);
1091 } else if (i == locked_pages) {
1092 /* writepages_finish() clears writeback pages
1093 * according to the data length, so make sure
1094 * data length covers all locked pages */
1095 u64 min_len = len + 1 - PAGE_SIZE;
1096 len = get_writepages_data_length(inode, pages[i - 1],
1097 offset);
1098 len = max(len, min_len);
1099 }
1100 dout("writepages got pages at %llu~%llu\n", offset, len);
1101
1102 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1103 0, !!pool, false);
1104 osd_req_op_extent_update(req, op_idx, len);
1105
1106 BUG_ON(op_idx + 1 != req->r_num_ops);
1107
1108 pool = NULL;
1109 if (i < locked_pages) {
1110 BUG_ON(num_ops <= req->r_num_ops);
1111 num_ops -= req->r_num_ops;
1112 locked_pages -= i;
1113
1114 /* allocate new pages array for next request */
1115 data_pages = pages;
1116 pages = kmalloc(locked_pages * sizeof (*pages),
1117 GFP_NOFS);
1118 if (!pages) {
1119 pool = fsc->wb_pagevec_pool;
1120 pages = mempool_alloc(pool, GFP_NOFS);
1121 BUG_ON(!pages);
1122 }
1123 memcpy(pages, data_pages + i,
1124 locked_pages * sizeof(*pages));
1125 memset(data_pages + i, 0,
1126 locked_pages * sizeof(*pages));
1127 } else {
1128 BUG_ON(num_ops != req->r_num_ops);
1129 index = pages[i - 1]->index + 1;
1130 /* request message now owns the pages array */
1131 pages = NULL;
1132 }
1133
1134 req->r_mtime = inode->i_mtime;
1135 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
1136 BUG_ON(rc);
1137 req = NULL;
1138
1139 wbc->nr_to_write -= i;
1140 if (pages)
1141 goto new_request;
1142
1143 /*
1144 * We stop writing back only if we are not doing
1145 * integrity sync. In case of integrity sync we have to
1146 * keep going until we have written all the pages
1147 * we tagged for writeback prior to entering this loop.
1148 */
1149 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1150 done = true;
1151
1152release_pvec_pages:
1153 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
1154 pvec.nr ? pvec.pages[0] : NULL);
1155 pagevec_release(&pvec);
1156 }
1157
1158 if (should_loop && !done) {
1159 /* more to do; loop back to beginning of file */
1160 dout("writepages looping back to beginning of file\n");
1161 end = start_index - 1; /* OK even when start_index == 0 */
1162
1163 /* to write dirty pages associated with next snapc,
1164 * we need to wait until current writes complete */
1165 if (wbc->sync_mode != WB_SYNC_NONE &&
1166 start_index == 0 && /* all dirty pages were checked */
1167 !ceph_wbc.head_snapc) {
1168 struct page *page;
1169 unsigned i, nr;
1170 index = 0;
1171 while ((index <= end) &&
1172 (nr = pagevec_lookup_tag(&pvec, mapping, &index,
1173 PAGECACHE_TAG_WRITEBACK))) {
1174 for (i = 0; i < nr; i++) {
1175 page = pvec.pages[i];
1176 if (page_snap_context(page) != snapc)
1177 continue;
1178 wait_on_page_writeback(page);
1179 }
1180 pagevec_release(&pvec);
1181 cond_resched();
1182 }
1183 }
1184
1185 start_index = 0;
1186 index = 0;
1187 goto retry;
1188 }
1189
1190 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1191 mapping->writeback_index = index;
1192
1193out:
1194 ceph_osdc_put_request(req);
1195 ceph_put_snap_context(last_snapc);
1196 dout("writepages dend - startone, rc = %d\n", rc);
1197 return rc;
1198}
1199
1200
1201
1202/*
1203 * See if a given @snapc is either writeable, or already written.
1204 */
1205static int context_is_writeable_or_written(struct inode *inode,
1206 struct ceph_snap_context *snapc)
1207{
1208 struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1209 int ret = !oldest || snapc->seq <= oldest->seq;
1210
1211 ceph_put_snap_context(oldest);
1212 return ret;
1213}
1214
1215/*
1216 * We are only allowed to write into/dirty the page if the page is
1217 * clean, or already dirty within the same snap context.
1218 *
1219 * called with page locked.
1220 * return success with page locked,
1221 * or any failure (incl -EAGAIN) with page unlocked.
1222 */
1223static int ceph_update_writeable_page(struct file *file,
1224 loff_t pos, unsigned len,
1225 struct page *page)
1226{
1227 struct inode *inode = file_inode(file);
1228 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1229 struct ceph_inode_info *ci = ceph_inode(inode);
1230 loff_t page_off = pos & PAGE_MASK;
1231 int pos_in_page = pos & ~PAGE_MASK;
1232 int end_in_page = pos_in_page + len;
1233 loff_t i_size;
1234 int r;
1235 struct ceph_snap_context *snapc, *oldest;
1236
1237 if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
1238 dout(" page %p forced umount\n", page);
1239 unlock_page(page);
1240 return -EIO;
1241 }
1242
1243retry_locked:
1244 /* writepages currently holds page lock, but if we change that later, */
1245 wait_on_page_writeback(page);
1246
1247 snapc = page_snap_context(page);
1248 if (snapc && snapc != ci->i_head_snapc) {
1249 /*
1250 * this page is already dirty in another (older) snap
1251 * context! is it writeable now?
1252 */
1253 oldest = get_oldest_context(inode, NULL, NULL);
1254 if (snapc->seq > oldest->seq) {
1255 ceph_put_snap_context(oldest);
1256 dout(" page %p snapc %p not current or oldest\n",
1257 page, snapc);
1258 /*
1259 * queue for writeback, and wait for snapc to
1260 * be writeable or written
1261 */
1262 snapc = ceph_get_snap_context(snapc);
1263 unlock_page(page);
1264 ceph_queue_writeback(inode);
1265 r = wait_event_killable(ci->i_cap_wq,
1266 context_is_writeable_or_written(inode, snapc));
1267 ceph_put_snap_context(snapc);
1268 if (r == -ERESTARTSYS)
1269 return r;
1270 return -EAGAIN;
1271 }
1272 ceph_put_snap_context(oldest);
1273
1274 /* yay, writeable, do it now (without dropping page lock) */
1275 dout(" page %p snapc %p not current, but oldest\n",
1276 page, snapc);
1277 if (!clear_page_dirty_for_io(page))
1278 goto retry_locked;
1279 r = writepage_nounlock(page, NULL);
1280 if (r < 0)
1281 goto fail_unlock;
1282 goto retry_locked;
1283 }
1284
1285 if (PageUptodate(page)) {
1286 dout(" page %p already uptodate\n", page);
1287 return 0;
1288 }
1289
1290 /* full page? */
1291 if (pos_in_page == 0 && len == PAGE_SIZE)
1292 return 0;
1293
1294 /* past end of file? */
1295 i_size = i_size_read(inode);
1296
1297 if (page_off >= i_size ||
1298 (pos_in_page == 0 && (pos+len) >= i_size &&
1299 end_in_page - pos_in_page != PAGE_SIZE)) {
1300 dout(" zeroing %p 0 - %d and %d - %d\n",
1301 page, pos_in_page, end_in_page, (int)PAGE_SIZE);
1302 zero_user_segments(page,
1303 0, pos_in_page,
1304 end_in_page, PAGE_SIZE);
1305 return 0;
1306 }
1307
1308 /* we need to read it. */
1309 r = ceph_do_readpage(file, page);
1310 if (r < 0) {
1311 if (r == -EINPROGRESS)
1312 return -EAGAIN;
1313 goto fail_unlock;
1314 }
1315 goto retry_locked;
1316fail_unlock:
1317 unlock_page(page);
1318 return r;
1319}
1320
1321/*
1322 * We are only allowed to write into/dirty the page if the page is
1323 * clean, or already dirty within the same snap context.
1324 */
1325static int ceph_write_begin(struct file *file, struct address_space *mapping,
1326 loff_t pos, unsigned len, unsigned flags,
1327 struct page **pagep, void **fsdata)
1328{
1329 struct inode *inode = file_inode(file);
1330 struct page *page;
1331 pgoff_t index = pos >> PAGE_SHIFT;
1332 int r;
1333
1334 do {
1335 /* get a page */
1336 page = grab_cache_page_write_begin(mapping, index, 0);
1337 if (!page)
1338 return -ENOMEM;
1339
1340 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1341 inode, page, (int)pos, (int)len);
1342
1343 r = ceph_update_writeable_page(file, pos, len, page);
1344 if (r < 0)
1345 put_page(page);
1346 else
1347 *pagep = page;
1348 } while (r == -EAGAIN);
1349
1350 return r;
1351}
1352
1353/*
1354 * we don't do anything in here that simple_write_end doesn't do
1355 * except adjust dirty page accounting
1356 */
1357static int ceph_write_end(struct file *file, struct address_space *mapping,
1358 loff_t pos, unsigned len, unsigned copied,
1359 struct page *page, void *fsdata)
1360{
1361 struct inode *inode = file_inode(file);
1362 bool check_cap = false;
1363
1364 dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1365 inode, page, (int)pos, (int)copied, (int)len);
1366
1367 /* zero the stale part of the page if we did a short copy */
1368 if (!PageUptodate(page)) {
1369 if (copied < len) {
1370 copied = 0;
1371 goto out;
1372 }
1373 SetPageUptodate(page);
1374 }
1375
1376 /* did file size increase? */
1377 if (pos+copied > i_size_read(inode))
1378 check_cap = ceph_inode_set_size(inode, pos+copied);
1379
1380 set_page_dirty(page);
1381
1382out:
1383 unlock_page(page);
1384 put_page(page);
1385
1386 if (check_cap)
1387 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1388
1389 return copied;
1390}
1391
1392/*
1393 * we set .direct_IO to indicate direct io is supported, but since we
1394 * intercept O_DIRECT reads and writes early, this function should
1395 * never get called.
1396 */
1397static ssize_t ceph_direct_io(struct kiocb *iocb, struct iov_iter *iter)
1398{
1399 WARN_ON(1);
1400 return -EINVAL;
1401}
1402
1403const struct address_space_operations ceph_aops = {
1404 .readpage = ceph_readpage,
1405 .readpages = ceph_readpages,
1406 .writepage = ceph_writepage,
1407 .writepages = ceph_writepages_start,
1408 .write_begin = ceph_write_begin,
1409 .write_end = ceph_write_end,
1410 .set_page_dirty = ceph_set_page_dirty,
1411 .invalidatepage = ceph_invalidatepage,
1412 .releasepage = ceph_releasepage,
1413 .direct_IO = ceph_direct_io,
1414};
1415
1416static void ceph_block_sigs(sigset_t *oldset)
1417{
1418 sigset_t mask;
1419 siginitsetinv(&mask, sigmask(SIGKILL));
1420 sigprocmask(SIG_BLOCK, &mask, oldset);
1421}
1422
1423static void ceph_restore_sigs(sigset_t *oldset)
1424{
1425 sigprocmask(SIG_SETMASK, oldset, NULL);
1426}
1427
1428/*
1429 * vm ops
1430 */
1431static int ceph_filemap_fault(struct vm_fault *vmf)
1432{
1433 struct vm_area_struct *vma = vmf->vma;
1434 struct inode *inode = file_inode(vma->vm_file);
1435 struct ceph_inode_info *ci = ceph_inode(inode);
1436 struct ceph_file_info *fi = vma->vm_file->private_data;
1437 struct page *pinned_page = NULL;
1438 loff_t off = vmf->pgoff << PAGE_SHIFT;
1439 int want, got, ret;
1440 sigset_t oldset;
1441
1442 ceph_block_sigs(&oldset);
1443
1444 dout("filemap_fault %p %llx.%llx %llu~%zd trying to get caps\n",
1445 inode, ceph_vinop(inode), off, (size_t)PAGE_SIZE);
1446 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1447 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1448 else
1449 want = CEPH_CAP_FILE_CACHE;
1450
1451 got = 0;
1452 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1453 if (ret < 0)
1454 goto out_restore;
1455
1456 dout("filemap_fault %p %llu~%zd got cap refs on %s\n",
1457 inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got));
1458
1459 if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1460 ci->i_inline_version == CEPH_INLINE_NONE) {
1461 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1462 ceph_add_rw_context(fi, &rw_ctx);
1463 ret = filemap_fault(vmf);
1464 ceph_del_rw_context(fi, &rw_ctx);
1465 } else
1466 ret = -EAGAIN;
1467
1468 dout("filemap_fault %p %llu~%zd dropping cap refs on %s ret %d\n",
1469 inode, off, (size_t)PAGE_SIZE, ceph_cap_string(got), ret);
1470 if (pinned_page)
1471 put_page(pinned_page);
1472 ceph_put_cap_refs(ci, got);
1473
1474 if (ret != -EAGAIN)
1475 goto out_restore;
1476
1477 /* read inline data */
1478 if (off >= PAGE_SIZE) {
1479 /* does not support inline data > PAGE_SIZE */
1480 ret = VM_FAULT_SIGBUS;
1481 } else {
1482 int ret1;
1483 struct address_space *mapping = inode->i_mapping;
1484 struct page *page = find_or_create_page(mapping, 0,
1485 mapping_gfp_constraint(mapping,
1486 ~__GFP_FS));
1487 if (!page) {
1488 ret = VM_FAULT_OOM;
1489 goto out_inline;
1490 }
1491 ret1 = __ceph_do_getattr(inode, page,
1492 CEPH_STAT_CAP_INLINE_DATA, true);
1493 if (ret1 < 0 || off >= i_size_read(inode)) {
1494 unlock_page(page);
1495 put_page(page);
1496 if (ret1 < 0)
1497 ret = ret1;
1498 else
1499 ret = VM_FAULT_SIGBUS;
1500 goto out_inline;
1501 }
1502 if (ret1 < PAGE_SIZE)
1503 zero_user_segment(page, ret1, PAGE_SIZE);
1504 else
1505 flush_dcache_page(page);
1506 SetPageUptodate(page);
1507 vmf->page = page;
1508 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1509out_inline:
1510 dout("filemap_fault %p %llu~%zd read inline data ret %d\n",
1511 inode, off, (size_t)PAGE_SIZE, ret);
1512 }
1513out_restore:
1514 ceph_restore_sigs(&oldset);
1515 if (ret < 0)
1516 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1517
1518 return ret;
1519}
1520
1521/*
1522 * Reuse write_begin here for simplicity.
1523 */
1524static int ceph_page_mkwrite(struct vm_fault *vmf)
1525{
1526 struct vm_area_struct *vma = vmf->vma;
1527 struct inode *inode = file_inode(vma->vm_file);
1528 struct ceph_inode_info *ci = ceph_inode(inode);
1529 struct ceph_file_info *fi = vma->vm_file->private_data;
1530 struct ceph_cap_flush *prealloc_cf;
1531 struct page *page = vmf->page;
1532 loff_t off = page_offset(page);
1533 loff_t size = i_size_read(inode);
1534 size_t len;
1535 int want, got, ret;
1536 sigset_t oldset;
1537
1538 prealloc_cf = ceph_alloc_cap_flush();
1539 if (!prealloc_cf)
1540 return VM_FAULT_OOM;
1541
1542 ceph_block_sigs(&oldset);
1543
1544 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1545 struct page *locked_page = NULL;
1546 if (off == 0) {
1547 lock_page(page);
1548 locked_page = page;
1549 }
1550 ret = ceph_uninline_data(vma->vm_file, locked_page);
1551 if (locked_page)
1552 unlock_page(locked_page);
1553 if (ret < 0)
1554 goto out_free;
1555 }
1556
1557 if (off + PAGE_SIZE <= size)
1558 len = PAGE_SIZE;
1559 else
1560 len = size & ~PAGE_MASK;
1561
1562 dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1563 inode, ceph_vinop(inode), off, len, size);
1564 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1565 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1566 else
1567 want = CEPH_CAP_FILE_BUFFER;
1568
1569 got = 0;
1570 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, off + len,
1571 &got, NULL);
1572 if (ret < 0)
1573 goto out_free;
1574
1575 dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1576 inode, off, len, ceph_cap_string(got));
1577
1578 /* Update time before taking page lock */
1579 file_update_time(vma->vm_file);
1580
1581 do {
1582 lock_page(page);
1583
1584 if ((off > size) || (page->mapping != inode->i_mapping)) {
1585 unlock_page(page);
1586 ret = VM_FAULT_NOPAGE;
1587 break;
1588 }
1589
1590 ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1591 if (ret >= 0) {
1592 /* success. we'll keep the page locked. */
1593 set_page_dirty(page);
1594 ret = VM_FAULT_LOCKED;
1595 }
1596 } while (ret == -EAGAIN);
1597
1598 if (ret == VM_FAULT_LOCKED ||
1599 ci->i_inline_version != CEPH_INLINE_NONE) {
1600 int dirty;
1601 spin_lock(&ci->i_ceph_lock);
1602 ci->i_inline_version = CEPH_INLINE_NONE;
1603 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1604 &prealloc_cf);
1605 spin_unlock(&ci->i_ceph_lock);
1606 if (dirty)
1607 __mark_inode_dirty(inode, dirty);
1608 }
1609
1610 dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1611 inode, off, len, ceph_cap_string(got), ret);
1612 ceph_put_cap_refs(ci, got);
1613out_free:
1614 ceph_restore_sigs(&oldset);
1615 ceph_free_cap_flush(prealloc_cf);
1616 if (ret < 0)
1617 ret = (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS;
1618 return ret;
1619}
1620
1621void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1622 char *data, size_t len)
1623{
1624 struct address_space *mapping = inode->i_mapping;
1625 struct page *page;
1626
1627 if (locked_page) {
1628 page = locked_page;
1629 } else {
1630 if (i_size_read(inode) == 0)
1631 return;
1632 page = find_or_create_page(mapping, 0,
1633 mapping_gfp_constraint(mapping,
1634 ~__GFP_FS));
1635 if (!page)
1636 return;
1637 if (PageUptodate(page)) {
1638 unlock_page(page);
1639 put_page(page);
1640 return;
1641 }
1642 }
1643
1644 dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1645 inode, ceph_vinop(inode), len, locked_page);
1646
1647 if (len > 0) {
1648 void *kaddr = kmap_atomic(page);
1649 memcpy(kaddr, data, len);
1650 kunmap_atomic(kaddr);
1651 }
1652
1653 if (page != locked_page) {
1654 if (len < PAGE_SIZE)
1655 zero_user_segment(page, len, PAGE_SIZE);
1656 else
1657 flush_dcache_page(page);
1658
1659 SetPageUptodate(page);
1660 unlock_page(page);
1661 put_page(page);
1662 }
1663}
1664
1665int ceph_uninline_data(struct file *filp, struct page *locked_page)
1666{
1667 struct inode *inode = file_inode(filp);
1668 struct ceph_inode_info *ci = ceph_inode(inode);
1669 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1670 struct ceph_osd_request *req;
1671 struct page *page = NULL;
1672 u64 len, inline_version;
1673 int err = 0;
1674 bool from_pagecache = false;
1675
1676 spin_lock(&ci->i_ceph_lock);
1677 inline_version = ci->i_inline_version;
1678 spin_unlock(&ci->i_ceph_lock);
1679
1680 dout("uninline_data %p %llx.%llx inline_version %llu\n",
1681 inode, ceph_vinop(inode), inline_version);
1682
1683 if (inline_version == 1 || /* initial version, no data */
1684 inline_version == CEPH_INLINE_NONE)
1685 goto out;
1686
1687 if (locked_page) {
1688 page = locked_page;
1689 WARN_ON(!PageUptodate(page));
1690 } else if (ceph_caps_issued(ci) &
1691 (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) {
1692 page = find_get_page(inode->i_mapping, 0);
1693 if (page) {
1694 if (PageUptodate(page)) {
1695 from_pagecache = true;
1696 lock_page(page);
1697 } else {
1698 put_page(page);
1699 page = NULL;
1700 }
1701 }
1702 }
1703
1704 if (page) {
1705 len = i_size_read(inode);
1706 if (len > PAGE_SIZE)
1707 len = PAGE_SIZE;
1708 } else {
1709 page = __page_cache_alloc(GFP_NOFS);
1710 if (!page) {
1711 err = -ENOMEM;
1712 goto out;
1713 }
1714 err = __ceph_do_getattr(inode, page,
1715 CEPH_STAT_CAP_INLINE_DATA, true);
1716 if (err < 0) {
1717 /* no inline data */
1718 if (err == -ENODATA)
1719 err = 0;
1720 goto out;
1721 }
1722 len = err;
1723 }
1724
1725 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1726 ceph_vino(inode), 0, &len, 0, 1,
1727 CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1728 NULL, 0, 0, false);
1729 if (IS_ERR(req)) {
1730 err = PTR_ERR(req);
1731 goto out;
1732 }
1733
1734 req->r_mtime = inode->i_mtime;
1735 err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1736 if (!err)
1737 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1738 ceph_osdc_put_request(req);
1739 if (err < 0)
1740 goto out;
1741
1742 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1743 ceph_vino(inode), 0, &len, 1, 3,
1744 CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1745 NULL, ci->i_truncate_seq,
1746 ci->i_truncate_size, false);
1747 if (IS_ERR(req)) {
1748 err = PTR_ERR(req);
1749 goto out;
1750 }
1751
1752 osd_req_op_extent_osd_data_pages(req, 1, &page, len, 0, false, false);
1753
1754 {
1755 __le64 xattr_buf = cpu_to_le64(inline_version);
1756 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1757 "inline_version", &xattr_buf,
1758 sizeof(xattr_buf),
1759 CEPH_OSD_CMPXATTR_OP_GT,
1760 CEPH_OSD_CMPXATTR_MODE_U64);
1761 if (err)
1762 goto out_put;
1763 }
1764
1765 {
1766 char xattr_buf[32];
1767 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1768 "%llu", inline_version);
1769 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1770 "inline_version",
1771 xattr_buf, xattr_len, 0, 0);
1772 if (err)
1773 goto out_put;
1774 }
1775
1776 req->r_mtime = inode->i_mtime;
1777 err = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1778 if (!err)
1779 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1780out_put:
1781 ceph_osdc_put_request(req);
1782 if (err == -ECANCELED)
1783 err = 0;
1784out:
1785 if (page && page != locked_page) {
1786 if (from_pagecache) {
1787 unlock_page(page);
1788 put_page(page);
1789 } else
1790 __free_pages(page, 0);
1791 }
1792
1793 dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1794 inode, ceph_vinop(inode), inline_version, err);
1795 return err;
1796}
1797
1798static const struct vm_operations_struct ceph_vmops = {
1799 .fault = ceph_filemap_fault,
1800 .page_mkwrite = ceph_page_mkwrite,
1801};
1802
1803int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1804{
1805 struct address_space *mapping = file->f_mapping;
1806
1807 if (!mapping->a_ops->readpage)
1808 return -ENOEXEC;
1809 file_accessed(file);
1810 vma->vm_ops = &ceph_vmops;
1811 return 0;
1812}
1813
1814enum {
1815 POOL_READ = 1,
1816 POOL_WRITE = 2,
1817};
1818
1819static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1820 s64 pool, struct ceph_string *pool_ns)
1821{
1822 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1823 struct ceph_mds_client *mdsc = fsc->mdsc;
1824 struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1825 struct rb_node **p, *parent;
1826 struct ceph_pool_perm *perm;
1827 struct page **pages;
1828 size_t pool_ns_len;
1829 int err = 0, err2 = 0, have = 0;
1830
1831 down_read(&mdsc->pool_perm_rwsem);
1832 p = &mdsc->pool_perm_tree.rb_node;
1833 while (*p) {
1834 perm = rb_entry(*p, struct ceph_pool_perm, node);
1835 if (pool < perm->pool)
1836 p = &(*p)->rb_left;
1837 else if (pool > perm->pool)
1838 p = &(*p)->rb_right;
1839 else {
1840 int ret = ceph_compare_string(pool_ns,
1841 perm->pool_ns,
1842 perm->pool_ns_len);
1843 if (ret < 0)
1844 p = &(*p)->rb_left;
1845 else if (ret > 0)
1846 p = &(*p)->rb_right;
1847 else {
1848 have = perm->perm;
1849 break;
1850 }
1851 }
1852 }
1853 up_read(&mdsc->pool_perm_rwsem);
1854 if (*p)
1855 goto out;
1856
1857 if (pool_ns)
1858 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
1859 pool, (int)pool_ns->len, pool_ns->str);
1860 else
1861 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
1862
1863 down_write(&mdsc->pool_perm_rwsem);
1864 p = &mdsc->pool_perm_tree.rb_node;
1865 parent = NULL;
1866 while (*p) {
1867 parent = *p;
1868 perm = rb_entry(parent, struct ceph_pool_perm, node);
1869 if (pool < perm->pool)
1870 p = &(*p)->rb_left;
1871 else if (pool > perm->pool)
1872 p = &(*p)->rb_right;
1873 else {
1874 int ret = ceph_compare_string(pool_ns,
1875 perm->pool_ns,
1876 perm->pool_ns_len);
1877 if (ret < 0)
1878 p = &(*p)->rb_left;
1879 else if (ret > 0)
1880 p = &(*p)->rb_right;
1881 else {
1882 have = perm->perm;
1883 break;
1884 }
1885 }
1886 }
1887 if (*p) {
1888 up_write(&mdsc->pool_perm_rwsem);
1889 goto out;
1890 }
1891
1892 rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1893 1, false, GFP_NOFS);
1894 if (!rd_req) {
1895 err = -ENOMEM;
1896 goto out_unlock;
1897 }
1898
1899 rd_req->r_flags = CEPH_OSD_FLAG_READ;
1900 osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1901 rd_req->r_base_oloc.pool = pool;
1902 if (pool_ns)
1903 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
1904 ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
1905
1906 err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
1907 if (err)
1908 goto out_unlock;
1909
1910 wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
1911 1, false, GFP_NOFS);
1912 if (!wr_req) {
1913 err = -ENOMEM;
1914 goto out_unlock;
1915 }
1916
1917 wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
1918 osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1919 ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
1920 ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
1921
1922 err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
1923 if (err)
1924 goto out_unlock;
1925
1926 /* one page should be large enough for STAT data */
1927 pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1928 if (IS_ERR(pages)) {
1929 err = PTR_ERR(pages);
1930 goto out_unlock;
1931 }
1932
1933 osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1934 0, false, true);
1935 err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1936
1937 wr_req->r_mtime = ci->vfs_inode.i_mtime;
1938 wr_req->r_abort_on_full = true;
1939 err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1940
1941 if (!err)
1942 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1943 if (!err2)
1944 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1945
1946 if (err >= 0 || err == -ENOENT)
1947 have |= POOL_READ;
1948 else if (err != -EPERM)
1949 goto out_unlock;
1950
1951 if (err2 == 0 || err2 == -EEXIST)
1952 have |= POOL_WRITE;
1953 else if (err2 != -EPERM) {
1954 err = err2;
1955 goto out_unlock;
1956 }
1957
1958 pool_ns_len = pool_ns ? pool_ns->len : 0;
1959 perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
1960 if (!perm) {
1961 err = -ENOMEM;
1962 goto out_unlock;
1963 }
1964
1965 perm->pool = pool;
1966 perm->perm = have;
1967 perm->pool_ns_len = pool_ns_len;
1968 if (pool_ns_len > 0)
1969 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
1970 perm->pool_ns[pool_ns_len] = 0;
1971
1972 rb_link_node(&perm->node, parent, p);
1973 rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1974 err = 0;
1975out_unlock:
1976 up_write(&mdsc->pool_perm_rwsem);
1977
1978 ceph_osdc_put_request(rd_req);
1979 ceph_osdc_put_request(wr_req);
1980out:
1981 if (!err)
1982 err = have;
1983 if (pool_ns)
1984 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
1985 pool, (int)pool_ns->len, pool_ns->str, err);
1986 else
1987 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
1988 return err;
1989}
1990
1991int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1992{
1993 s64 pool;
1994 struct ceph_string *pool_ns;
1995 int ret, flags;
1996
1997 if (ci->i_vino.snap != CEPH_NOSNAP) {
1998 /*
1999 * Pool permission check needs to write to the first object.
2000 * But for snapshot, head of the first object may have alread
2001 * been deleted. Skip check to avoid creating orphan object.
2002 */
2003 return 0;
2004 }
2005
2006 if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
2007 NOPOOLPERM))
2008 return 0;
2009
2010 spin_lock(&ci->i_ceph_lock);
2011 flags = ci->i_ceph_flags;
2012 pool = ci->i_layout.pool_id;
2013 spin_unlock(&ci->i_ceph_lock);
2014check:
2015 if (flags & CEPH_I_POOL_PERM) {
2016 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2017 dout("ceph_pool_perm_check pool %lld no read perm\n",
2018 pool);
2019 return -EPERM;
2020 }
2021 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2022 dout("ceph_pool_perm_check pool %lld no write perm\n",
2023 pool);
2024 return -EPERM;
2025 }
2026 return 0;
2027 }
2028
2029 pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2030 ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2031 ceph_put_string(pool_ns);
2032 if (ret < 0)
2033 return ret;
2034
2035 flags = CEPH_I_POOL_PERM;
2036 if (ret & POOL_READ)
2037 flags |= CEPH_I_POOL_RD;
2038 if (ret & POOL_WRITE)
2039 flags |= CEPH_I_POOL_WR;
2040
2041 spin_lock(&ci->i_ceph_lock);
2042 if (pool == ci->i_layout.pool_id &&
2043 pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2044 ci->i_ceph_flags |= flags;
2045 } else {
2046 pool = ci->i_layout.pool_id;
2047 flags = ci->i_ceph_flags;
2048 }
2049 spin_unlock(&ci->i_ceph_lock);
2050 goto check;
2051}
2052
2053void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2054{
2055 struct ceph_pool_perm *perm;
2056 struct rb_node *n;
2057
2058 while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2059 n = rb_first(&mdsc->pool_perm_tree);
2060 perm = rb_entry(n, struct ceph_pool_perm, node);
2061 rb_erase(n, &mdsc->pool_perm_tree);
2062 kfree(perm);
2063 }
2064}