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/swap.h>
8#include <linux/pagemap.h>
9#include <linux/slab.h>
10#include <linux/pagevec.h>
11#include <linux/task_io_accounting_ops.h>
12#include <linux/signal.h>
13#include <linux/iversion.h>
14#include <linux/ktime.h>
15#include <linux/netfs.h>
16
17#include "super.h"
18#include "mds_client.h"
19#include "cache.h"
20#include "metric.h"
21#include "crypto.h"
22#include <linux/ceph/osd_client.h>
23#include <linux/ceph/striper.h>
24
25/*
26 * Ceph address space ops.
27 *
28 * There are a few funny things going on here.
29 *
30 * The page->private field is used to reference a struct
31 * ceph_snap_context for _every_ dirty page. This indicates which
32 * snapshot the page was logically dirtied in, and thus which snap
33 * context needs to be associated with the osd write during writeback.
34 *
35 * Similarly, struct ceph_inode_info maintains a set of counters to
36 * count dirty pages on the inode. In the absence of snapshots,
37 * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
38 *
39 * When a snapshot is taken (that is, when the client receives
40 * notification that a snapshot was taken), each inode with caps and
41 * with dirty pages (dirty pages implies there is a cap) gets a new
42 * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
43 * order, new snaps go to the tail). The i_wrbuffer_ref_head count is
44 * moved to capsnap->dirty. (Unless a sync write is currently in
45 * progress. In that case, the capsnap is said to be "pending", new
46 * writes cannot start, and the capsnap isn't "finalized" until the
47 * write completes (or fails) and a final size/mtime for the inode for
48 * that snap can be settled upon.) i_wrbuffer_ref_head is reset to 0.
49 *
50 * On writeback, we must submit writes to the osd IN SNAP ORDER. So,
51 * we look for the first capsnap in i_cap_snaps and write out pages in
52 * that snap context _only_. Then we move on to the next capsnap,
53 * eventually reaching the "live" or "head" context (i.e., pages that
54 * are not yet snapped) and are writing the most recently dirtied
55 * pages.
56 *
57 * Invalidate and so forth must take care to ensure the dirty page
58 * accounting is preserved.
59 */
60
61#define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
62#define CONGESTION_OFF_THRESH(congestion_kb) \
63 (CONGESTION_ON_THRESH(congestion_kb) - \
64 (CONGESTION_ON_THRESH(congestion_kb) >> 2))
65
66static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
67 struct folio **foliop, void **_fsdata);
68
69static inline struct ceph_snap_context *page_snap_context(struct page *page)
70{
71 if (PagePrivate(page))
72 return (void *)page->private;
73 return NULL;
74}
75
76/*
77 * Dirty a page. Optimistically adjust accounting, on the assumption
78 * that we won't race with invalidate. If we do, readjust.
79 */
80static bool ceph_dirty_folio(struct address_space *mapping, struct folio *folio)
81{
82 struct inode *inode = mapping->host;
83 struct ceph_client *cl = ceph_inode_to_client(inode);
84 struct ceph_inode_info *ci;
85 struct ceph_snap_context *snapc;
86
87 if (folio_test_dirty(folio)) {
88 doutc(cl, "%llx.%llx %p idx %lu -- already dirty\n",
89 ceph_vinop(inode), folio, folio->index);
90 VM_BUG_ON_FOLIO(!folio_test_private(folio), folio);
91 return false;
92 }
93
94 ci = ceph_inode(inode);
95
96 /* dirty the head */
97 spin_lock(&ci->i_ceph_lock);
98 BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
99 if (__ceph_have_pending_cap_snap(ci)) {
100 struct ceph_cap_snap *capsnap =
101 list_last_entry(&ci->i_cap_snaps,
102 struct ceph_cap_snap,
103 ci_item);
104 snapc = ceph_get_snap_context(capsnap->context);
105 capsnap->dirty_pages++;
106 } else {
107 BUG_ON(!ci->i_head_snapc);
108 snapc = ceph_get_snap_context(ci->i_head_snapc);
109 ++ci->i_wrbuffer_ref_head;
110 }
111 if (ci->i_wrbuffer_ref == 0)
112 ihold(inode);
113 ++ci->i_wrbuffer_ref;
114 doutc(cl, "%llx.%llx %p idx %lu head %d/%d -> %d/%d "
115 "snapc %p seq %lld (%d snaps)\n",
116 ceph_vinop(inode), folio, folio->index,
117 ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
118 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
119 snapc, snapc->seq, snapc->num_snaps);
120 spin_unlock(&ci->i_ceph_lock);
121
122 /*
123 * Reference snap context in folio->private. Also set
124 * PagePrivate so that we get invalidate_folio callback.
125 */
126 VM_WARN_ON_FOLIO(folio->private, folio);
127 folio_attach_private(folio, snapc);
128
129 return ceph_fscache_dirty_folio(mapping, folio);
130}
131
132/*
133 * If we are truncating the full folio (i.e. offset == 0), adjust the
134 * dirty folio counters appropriately. Only called if there is private
135 * data on the folio.
136 */
137static void ceph_invalidate_folio(struct folio *folio, size_t offset,
138 size_t length)
139{
140 struct inode *inode = folio->mapping->host;
141 struct ceph_client *cl = ceph_inode_to_client(inode);
142 struct ceph_inode_info *ci = ceph_inode(inode);
143 struct ceph_snap_context *snapc;
144
145
146 if (offset != 0 || length != folio_size(folio)) {
147 doutc(cl, "%llx.%llx idx %lu partial dirty page %zu~%zu\n",
148 ceph_vinop(inode), folio->index, offset, length);
149 return;
150 }
151
152 WARN_ON(!folio_test_locked(folio));
153 if (folio_test_private(folio)) {
154 doutc(cl, "%llx.%llx idx %lu full dirty page\n",
155 ceph_vinop(inode), folio->index);
156
157 snapc = folio_detach_private(folio);
158 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
159 ceph_put_snap_context(snapc);
160 }
161
162 netfs_invalidate_folio(folio, offset, length);
163}
164
165static void ceph_netfs_expand_readahead(struct netfs_io_request *rreq)
166{
167 struct inode *inode = rreq->inode;
168 struct ceph_inode_info *ci = ceph_inode(inode);
169 struct ceph_file_layout *lo = &ci->i_layout;
170 unsigned long max_pages = inode->i_sb->s_bdi->ra_pages;
171 loff_t end = rreq->start + rreq->len, new_end;
172 struct ceph_netfs_request_data *priv = rreq->netfs_priv;
173 unsigned long max_len;
174 u32 blockoff;
175
176 if (priv) {
177 /* Readahead is disabled by posix_fadvise POSIX_FADV_RANDOM */
178 if (priv->file_ra_disabled)
179 max_pages = 0;
180 else
181 max_pages = priv->file_ra_pages;
182
183 }
184
185 /* Readahead is disabled */
186 if (!max_pages)
187 return;
188
189 max_len = max_pages << PAGE_SHIFT;
190
191 /*
192 * Try to expand the length forward by rounding up it to the next
193 * block, but do not exceed the file size, unless the original
194 * request already exceeds it.
195 */
196 new_end = min(round_up(end, lo->stripe_unit), rreq->i_size);
197 if (new_end > end && new_end <= rreq->start + max_len)
198 rreq->len = new_end - rreq->start;
199
200 /* Try to expand the start downward */
201 div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
202 if (rreq->len + blockoff <= max_len) {
203 rreq->start -= blockoff;
204 rreq->len += blockoff;
205 }
206}
207
208static bool ceph_netfs_clamp_length(struct netfs_io_subrequest *subreq)
209{
210 struct inode *inode = subreq->rreq->inode;
211 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
212 struct ceph_inode_info *ci = ceph_inode(inode);
213 u64 objno, objoff;
214 u32 xlen;
215
216 /* Truncate the extent at the end of the current block */
217 ceph_calc_file_object_mapping(&ci->i_layout, subreq->start, subreq->len,
218 &objno, &objoff, &xlen);
219 subreq->len = min(xlen, fsc->mount_options->rsize);
220 return true;
221}
222
223static void finish_netfs_read(struct ceph_osd_request *req)
224{
225 struct inode *inode = req->r_inode;
226 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
227 struct ceph_client *cl = fsc->client;
228 struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
229 struct netfs_io_subrequest *subreq = req->r_priv;
230 struct ceph_osd_req_op *op = &req->r_ops[0];
231 int err = req->r_result;
232 bool sparse = (op->op == CEPH_OSD_OP_SPARSE_READ);
233
234 ceph_update_read_metrics(&fsc->mdsc->metric, req->r_start_latency,
235 req->r_end_latency, osd_data->length, err);
236
237 doutc(cl, "result %d subreq->len=%zu i_size=%lld\n", req->r_result,
238 subreq->len, i_size_read(req->r_inode));
239
240 /* no object means success but no data */
241 if (err == -ENOENT)
242 err = 0;
243 else if (err == -EBLOCKLISTED)
244 fsc->blocklisted = true;
245
246 if (err >= 0) {
247 if (sparse && err > 0)
248 err = ceph_sparse_ext_map_end(op);
249 if (err < subreq->len)
250 __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
251 if (IS_ENCRYPTED(inode) && err > 0) {
252 err = ceph_fscrypt_decrypt_extents(inode,
253 osd_data->pages, subreq->start,
254 op->extent.sparse_ext,
255 op->extent.sparse_ext_cnt);
256 if (err > subreq->len)
257 err = subreq->len;
258 }
259 }
260
261 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
262 ceph_put_page_vector(osd_data->pages,
263 calc_pages_for(osd_data->alignment,
264 osd_data->length), false);
265 }
266 netfs_subreq_terminated(subreq, err, false);
267 iput(req->r_inode);
268 ceph_dec_osd_stopping_blocker(fsc->mdsc);
269}
270
271static bool ceph_netfs_issue_op_inline(struct netfs_io_subrequest *subreq)
272{
273 struct netfs_io_request *rreq = subreq->rreq;
274 struct inode *inode = rreq->inode;
275 struct ceph_mds_reply_info_parsed *rinfo;
276 struct ceph_mds_reply_info_in *iinfo;
277 struct ceph_mds_request *req;
278 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
279 struct ceph_inode_info *ci = ceph_inode(inode);
280 struct iov_iter iter;
281 ssize_t err = 0;
282 size_t len;
283 int mode;
284
285 __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
286 __clear_bit(NETFS_SREQ_COPY_TO_CACHE, &subreq->flags);
287
288 if (subreq->start >= inode->i_size)
289 goto out;
290
291 /* We need to fetch the inline data. */
292 mode = ceph_try_to_choose_auth_mds(inode, CEPH_STAT_CAP_INLINE_DATA);
293 req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_GETATTR, mode);
294 if (IS_ERR(req)) {
295 err = PTR_ERR(req);
296 goto out;
297 }
298 req->r_ino1 = ci->i_vino;
299 req->r_args.getattr.mask = cpu_to_le32(CEPH_STAT_CAP_INLINE_DATA);
300 req->r_num_caps = 2;
301
302 err = ceph_mdsc_do_request(mdsc, NULL, req);
303 if (err < 0)
304 goto out;
305
306 rinfo = &req->r_reply_info;
307 iinfo = &rinfo->targeti;
308 if (iinfo->inline_version == CEPH_INLINE_NONE) {
309 /* The data got uninlined */
310 ceph_mdsc_put_request(req);
311 return false;
312 }
313
314 len = min_t(size_t, iinfo->inline_len - subreq->start, subreq->len);
315 iov_iter_xarray(&iter, ITER_DEST, &rreq->mapping->i_pages, subreq->start, len);
316 err = copy_to_iter(iinfo->inline_data + subreq->start, len, &iter);
317 if (err == 0)
318 err = -EFAULT;
319
320 ceph_mdsc_put_request(req);
321out:
322 netfs_subreq_terminated(subreq, err, false);
323 return true;
324}
325
326static void ceph_netfs_issue_read(struct netfs_io_subrequest *subreq)
327{
328 struct netfs_io_request *rreq = subreq->rreq;
329 struct inode *inode = rreq->inode;
330 struct ceph_inode_info *ci = ceph_inode(inode);
331 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
332 struct ceph_client *cl = fsc->client;
333 struct ceph_osd_request *req = NULL;
334 struct ceph_vino vino = ceph_vino(inode);
335 struct iov_iter iter;
336 int err = 0;
337 u64 len = subreq->len;
338 bool sparse = IS_ENCRYPTED(inode) || ceph_test_mount_opt(fsc, SPARSEREAD);
339 u64 off = subreq->start;
340 int extent_cnt;
341
342 if (ceph_inode_is_shutdown(inode)) {
343 err = -EIO;
344 goto out;
345 }
346
347 if (ceph_has_inline_data(ci) && ceph_netfs_issue_op_inline(subreq))
348 return;
349
350 ceph_fscrypt_adjust_off_and_len(inode, &off, &len);
351
352 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, vino,
353 off, &len, 0, 1, sparse ? CEPH_OSD_OP_SPARSE_READ : CEPH_OSD_OP_READ,
354 CEPH_OSD_FLAG_READ, NULL, ci->i_truncate_seq,
355 ci->i_truncate_size, false);
356 if (IS_ERR(req)) {
357 err = PTR_ERR(req);
358 req = NULL;
359 goto out;
360 }
361
362 if (sparse) {
363 extent_cnt = __ceph_sparse_read_ext_count(inode, len);
364 err = ceph_alloc_sparse_ext_map(&req->r_ops[0], extent_cnt);
365 if (err)
366 goto out;
367 }
368
369 doutc(cl, "%llx.%llx pos=%llu orig_len=%zu len=%llu\n",
370 ceph_vinop(inode), subreq->start, subreq->len, len);
371
372 iov_iter_xarray(&iter, ITER_DEST, &rreq->mapping->i_pages, subreq->start, len);
373
374 /*
375 * FIXME: For now, use CEPH_OSD_DATA_TYPE_PAGES instead of _ITER for
376 * encrypted inodes. We'd need infrastructure that handles an iov_iter
377 * instead of page arrays, and we don't have that as of yet. Once the
378 * dust settles on the write helpers and encrypt/decrypt routines for
379 * netfs, we should be able to rework this.
380 */
381 if (IS_ENCRYPTED(inode)) {
382 struct page **pages;
383 size_t page_off;
384
385 err = iov_iter_get_pages_alloc2(&iter, &pages, len, &page_off);
386 if (err < 0) {
387 doutc(cl, "%llx.%llx failed to allocate pages, %d\n",
388 ceph_vinop(inode), err);
389 goto out;
390 }
391
392 /* should always give us a page-aligned read */
393 WARN_ON_ONCE(page_off);
394 len = err;
395 err = 0;
396
397 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false,
398 false);
399 } else {
400 osd_req_op_extent_osd_iter(req, 0, &iter);
401 }
402 if (!ceph_inc_osd_stopping_blocker(fsc->mdsc)) {
403 err = -EIO;
404 goto out;
405 }
406 req->r_callback = finish_netfs_read;
407 req->r_priv = subreq;
408 req->r_inode = inode;
409 ihold(inode);
410
411 ceph_osdc_start_request(req->r_osdc, req);
412out:
413 ceph_osdc_put_request(req);
414 if (err)
415 netfs_subreq_terminated(subreq, err, false);
416 doutc(cl, "%llx.%llx result %d\n", ceph_vinop(inode), err);
417}
418
419static int ceph_init_request(struct netfs_io_request *rreq, struct file *file)
420{
421 struct inode *inode = rreq->inode;
422 struct ceph_client *cl = ceph_inode_to_client(inode);
423 int got = 0, want = CEPH_CAP_FILE_CACHE;
424 struct ceph_netfs_request_data *priv;
425 int ret = 0;
426
427 if (rreq->origin != NETFS_READAHEAD)
428 return 0;
429
430 priv = kzalloc(sizeof(*priv), GFP_NOFS);
431 if (!priv)
432 return -ENOMEM;
433
434 if (file) {
435 struct ceph_rw_context *rw_ctx;
436 struct ceph_file_info *fi = file->private_data;
437
438 priv->file_ra_pages = file->f_ra.ra_pages;
439 priv->file_ra_disabled = file->f_mode & FMODE_RANDOM;
440
441 rw_ctx = ceph_find_rw_context(fi);
442 if (rw_ctx) {
443 rreq->netfs_priv = priv;
444 return 0;
445 }
446 }
447
448 /*
449 * readahead callers do not necessarily hold Fcb caps
450 * (e.g. fadvise, madvise).
451 */
452 ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want, true, &got);
453 if (ret < 0) {
454 doutc(cl, "%llx.%llx, error getting cap\n", ceph_vinop(inode));
455 goto out;
456 }
457
458 if (!(got & want)) {
459 doutc(cl, "%llx.%llx, no cache cap\n", ceph_vinop(inode));
460 ret = -EACCES;
461 goto out;
462 }
463 if (ret == 0) {
464 ret = -EACCES;
465 goto out;
466 }
467
468 priv->caps = got;
469 rreq->netfs_priv = priv;
470
471out:
472 if (ret < 0)
473 kfree(priv);
474
475 return ret;
476}
477
478static void ceph_netfs_free_request(struct netfs_io_request *rreq)
479{
480 struct ceph_netfs_request_data *priv = rreq->netfs_priv;
481
482 if (!priv)
483 return;
484
485 if (priv->caps)
486 ceph_put_cap_refs(ceph_inode(rreq->inode), priv->caps);
487 kfree(priv);
488 rreq->netfs_priv = NULL;
489}
490
491const struct netfs_request_ops ceph_netfs_ops = {
492 .init_request = ceph_init_request,
493 .free_request = ceph_netfs_free_request,
494 .issue_read = ceph_netfs_issue_read,
495 .expand_readahead = ceph_netfs_expand_readahead,
496 .clamp_length = ceph_netfs_clamp_length,
497 .check_write_begin = ceph_netfs_check_write_begin,
498};
499
500#ifdef CONFIG_CEPH_FSCACHE
501static void ceph_set_page_fscache(struct page *page)
502{
503 set_page_fscache(page);
504}
505
506static void ceph_fscache_write_terminated(void *priv, ssize_t error, bool was_async)
507{
508 struct inode *inode = priv;
509
510 if (IS_ERR_VALUE(error) && error != -ENOBUFS)
511 ceph_fscache_invalidate(inode, false);
512}
513
514static void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
515{
516 struct ceph_inode_info *ci = ceph_inode(inode);
517 struct fscache_cookie *cookie = ceph_fscache_cookie(ci);
518
519 fscache_write_to_cache(cookie, inode->i_mapping, off, len, i_size_read(inode),
520 ceph_fscache_write_terminated, inode, caching);
521}
522#else
523static inline void ceph_set_page_fscache(struct page *page)
524{
525}
526
527static inline void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
528{
529}
530#endif /* CONFIG_CEPH_FSCACHE */
531
532struct ceph_writeback_ctl
533{
534 loff_t i_size;
535 u64 truncate_size;
536 u32 truncate_seq;
537 bool size_stable;
538 bool head_snapc;
539};
540
541/*
542 * Get ref for the oldest snapc for an inode with dirty data... that is, the
543 * only snap context we are allowed to write back.
544 */
545static struct ceph_snap_context *
546get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
547 struct ceph_snap_context *page_snapc)
548{
549 struct ceph_inode_info *ci = ceph_inode(inode);
550 struct ceph_client *cl = ceph_inode_to_client(inode);
551 struct ceph_snap_context *snapc = NULL;
552 struct ceph_cap_snap *capsnap = NULL;
553
554 spin_lock(&ci->i_ceph_lock);
555 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
556 doutc(cl, " capsnap %p snapc %p has %d dirty pages\n",
557 capsnap, capsnap->context, capsnap->dirty_pages);
558 if (!capsnap->dirty_pages)
559 continue;
560
561 /* get i_size, truncate_{seq,size} for page_snapc? */
562 if (snapc && capsnap->context != page_snapc)
563 continue;
564
565 if (ctl) {
566 if (capsnap->writing) {
567 ctl->i_size = i_size_read(inode);
568 ctl->size_stable = false;
569 } else {
570 ctl->i_size = capsnap->size;
571 ctl->size_stable = true;
572 }
573 ctl->truncate_size = capsnap->truncate_size;
574 ctl->truncate_seq = capsnap->truncate_seq;
575 ctl->head_snapc = false;
576 }
577
578 if (snapc)
579 break;
580
581 snapc = ceph_get_snap_context(capsnap->context);
582 if (!page_snapc ||
583 page_snapc == snapc ||
584 page_snapc->seq > snapc->seq)
585 break;
586 }
587 if (!snapc && ci->i_wrbuffer_ref_head) {
588 snapc = ceph_get_snap_context(ci->i_head_snapc);
589 doutc(cl, " head snapc %p has %d dirty pages\n", snapc,
590 ci->i_wrbuffer_ref_head);
591 if (ctl) {
592 ctl->i_size = i_size_read(inode);
593 ctl->truncate_size = ci->i_truncate_size;
594 ctl->truncate_seq = ci->i_truncate_seq;
595 ctl->size_stable = false;
596 ctl->head_snapc = true;
597 }
598 }
599 spin_unlock(&ci->i_ceph_lock);
600 return snapc;
601}
602
603static u64 get_writepages_data_length(struct inode *inode,
604 struct page *page, u64 start)
605{
606 struct ceph_inode_info *ci = ceph_inode(inode);
607 struct ceph_snap_context *snapc;
608 struct ceph_cap_snap *capsnap = NULL;
609 u64 end = i_size_read(inode);
610 u64 ret;
611
612 snapc = page_snap_context(ceph_fscrypt_pagecache_page(page));
613 if (snapc != ci->i_head_snapc) {
614 bool found = false;
615 spin_lock(&ci->i_ceph_lock);
616 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
617 if (capsnap->context == snapc) {
618 if (!capsnap->writing)
619 end = capsnap->size;
620 found = true;
621 break;
622 }
623 }
624 spin_unlock(&ci->i_ceph_lock);
625 WARN_ON(!found);
626 }
627 if (end > ceph_fscrypt_page_offset(page) + thp_size(page))
628 end = ceph_fscrypt_page_offset(page) + thp_size(page);
629 ret = end > start ? end - start : 0;
630 if (ret && fscrypt_is_bounce_page(page))
631 ret = round_up(ret, CEPH_FSCRYPT_BLOCK_SIZE);
632 return ret;
633}
634
635/*
636 * Write a single page, but leave the page locked.
637 *
638 * If we get a write error, mark the mapping for error, but still adjust the
639 * dirty page accounting (i.e., page is no longer dirty).
640 */
641static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
642{
643 struct folio *folio = page_folio(page);
644 struct inode *inode = page->mapping->host;
645 struct ceph_inode_info *ci = ceph_inode(inode);
646 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
647 struct ceph_client *cl = fsc->client;
648 struct ceph_snap_context *snapc, *oldest;
649 loff_t page_off = page_offset(page);
650 int err;
651 loff_t len = thp_size(page);
652 loff_t wlen;
653 struct ceph_writeback_ctl ceph_wbc;
654 struct ceph_osd_client *osdc = &fsc->client->osdc;
655 struct ceph_osd_request *req;
656 bool caching = ceph_is_cache_enabled(inode);
657 struct page *bounce_page = NULL;
658
659 doutc(cl, "%llx.%llx page %p idx %lu\n", ceph_vinop(inode), page,
660 page->index);
661
662 if (ceph_inode_is_shutdown(inode))
663 return -EIO;
664
665 /* verify this is a writeable snap context */
666 snapc = page_snap_context(page);
667 if (!snapc) {
668 doutc(cl, "%llx.%llx page %p not dirty?\n", ceph_vinop(inode),
669 page);
670 return 0;
671 }
672 oldest = get_oldest_context(inode, &ceph_wbc, snapc);
673 if (snapc->seq > oldest->seq) {
674 doutc(cl, "%llx.%llx page %p snapc %p not writeable - noop\n",
675 ceph_vinop(inode), page, snapc);
676 /* we should only noop if called by kswapd */
677 WARN_ON(!(current->flags & PF_MEMALLOC));
678 ceph_put_snap_context(oldest);
679 redirty_page_for_writepage(wbc, page);
680 return 0;
681 }
682 ceph_put_snap_context(oldest);
683
684 /* is this a partial page at end of file? */
685 if (page_off >= ceph_wbc.i_size) {
686 doutc(cl, "%llx.%llx folio at %lu beyond eof %llu\n",
687 ceph_vinop(inode), folio->index, ceph_wbc.i_size);
688 folio_invalidate(folio, 0, folio_size(folio));
689 return 0;
690 }
691
692 if (ceph_wbc.i_size < page_off + len)
693 len = ceph_wbc.i_size - page_off;
694
695 wlen = IS_ENCRYPTED(inode) ? round_up(len, CEPH_FSCRYPT_BLOCK_SIZE) : len;
696 doutc(cl, "%llx.%llx page %p index %lu on %llu~%llu snapc %p seq %lld\n",
697 ceph_vinop(inode), page, page->index, page_off, wlen, snapc,
698 snapc->seq);
699
700 if (atomic_long_inc_return(&fsc->writeback_count) >
701 CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
702 fsc->write_congested = true;
703
704 req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode),
705 page_off, &wlen, 0, 1, CEPH_OSD_OP_WRITE,
706 CEPH_OSD_FLAG_WRITE, snapc,
707 ceph_wbc.truncate_seq,
708 ceph_wbc.truncate_size, true);
709 if (IS_ERR(req)) {
710 redirty_page_for_writepage(wbc, page);
711 return PTR_ERR(req);
712 }
713
714 if (wlen < len)
715 len = wlen;
716
717 set_page_writeback(page);
718 if (caching)
719 ceph_set_page_fscache(page);
720 ceph_fscache_write_to_cache(inode, page_off, len, caching);
721
722 if (IS_ENCRYPTED(inode)) {
723 bounce_page = fscrypt_encrypt_pagecache_blocks(page,
724 CEPH_FSCRYPT_BLOCK_SIZE, 0,
725 GFP_NOFS);
726 if (IS_ERR(bounce_page)) {
727 redirty_page_for_writepage(wbc, page);
728 end_page_writeback(page);
729 ceph_osdc_put_request(req);
730 return PTR_ERR(bounce_page);
731 }
732 }
733
734 /* it may be a short write due to an object boundary */
735 WARN_ON_ONCE(len > thp_size(page));
736 osd_req_op_extent_osd_data_pages(req, 0,
737 bounce_page ? &bounce_page : &page, wlen, 0,
738 false, false);
739 doutc(cl, "%llx.%llx %llu~%llu (%llu bytes, %sencrypted)\n",
740 ceph_vinop(inode), page_off, len, wlen,
741 IS_ENCRYPTED(inode) ? "" : "not ");
742
743 req->r_mtime = inode_get_mtime(inode);
744 ceph_osdc_start_request(osdc, req);
745 err = ceph_osdc_wait_request(osdc, req);
746
747 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
748 req->r_end_latency, len, err);
749 fscrypt_free_bounce_page(bounce_page);
750 ceph_osdc_put_request(req);
751 if (err == 0)
752 err = len;
753
754 if (err < 0) {
755 struct writeback_control tmp_wbc;
756 if (!wbc)
757 wbc = &tmp_wbc;
758 if (err == -ERESTARTSYS) {
759 /* killed by SIGKILL */
760 doutc(cl, "%llx.%llx interrupted page %p\n",
761 ceph_vinop(inode), page);
762 redirty_page_for_writepage(wbc, page);
763 end_page_writeback(page);
764 return err;
765 }
766 if (err == -EBLOCKLISTED)
767 fsc->blocklisted = true;
768 doutc(cl, "%llx.%llx setting page/mapping error %d %p\n",
769 ceph_vinop(inode), err, page);
770 mapping_set_error(&inode->i_data, err);
771 wbc->pages_skipped++;
772 } else {
773 doutc(cl, "%llx.%llx cleaned page %p\n",
774 ceph_vinop(inode), page);
775 err = 0; /* vfs expects us to return 0 */
776 }
777 oldest = detach_page_private(page);
778 WARN_ON_ONCE(oldest != snapc);
779 end_page_writeback(page);
780 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
781 ceph_put_snap_context(snapc); /* page's reference */
782
783 if (atomic_long_dec_return(&fsc->writeback_count) <
784 CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
785 fsc->write_congested = false;
786
787 return err;
788}
789
790static int ceph_writepage(struct page *page, struct writeback_control *wbc)
791{
792 int err;
793 struct inode *inode = page->mapping->host;
794 BUG_ON(!inode);
795 ihold(inode);
796
797 if (wbc->sync_mode == WB_SYNC_NONE &&
798 ceph_inode_to_fs_client(inode)->write_congested)
799 return AOP_WRITEPAGE_ACTIVATE;
800
801 wait_on_page_fscache(page);
802
803 err = writepage_nounlock(page, wbc);
804 if (err == -ERESTARTSYS) {
805 /* direct memory reclaimer was killed by SIGKILL. return 0
806 * to prevent caller from setting mapping/page error */
807 err = 0;
808 }
809 unlock_page(page);
810 iput(inode);
811 return err;
812}
813
814/*
815 * async writeback completion handler.
816 *
817 * If we get an error, set the mapping error bit, but not the individual
818 * page error bits.
819 */
820static void writepages_finish(struct ceph_osd_request *req)
821{
822 struct inode *inode = req->r_inode;
823 struct ceph_inode_info *ci = ceph_inode(inode);
824 struct ceph_client *cl = ceph_inode_to_client(inode);
825 struct ceph_osd_data *osd_data;
826 struct page *page;
827 int num_pages, total_pages = 0;
828 int i, j;
829 int rc = req->r_result;
830 struct ceph_snap_context *snapc = req->r_snapc;
831 struct address_space *mapping = inode->i_mapping;
832 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
833 unsigned int len = 0;
834 bool remove_page;
835
836 doutc(cl, "%llx.%llx rc %d\n", ceph_vinop(inode), rc);
837 if (rc < 0) {
838 mapping_set_error(mapping, rc);
839 ceph_set_error_write(ci);
840 if (rc == -EBLOCKLISTED)
841 fsc->blocklisted = true;
842 } else {
843 ceph_clear_error_write(ci);
844 }
845
846 /*
847 * We lost the cache cap, need to truncate the page before
848 * it is unlocked, otherwise we'd truncate it later in the
849 * page truncation thread, possibly losing some data that
850 * raced its way in
851 */
852 remove_page = !(ceph_caps_issued(ci) &
853 (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
854
855 /* clean all pages */
856 for (i = 0; i < req->r_num_ops; i++) {
857 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE) {
858 pr_warn_client(cl,
859 "%llx.%llx incorrect op %d req %p index %d tid %llu\n",
860 ceph_vinop(inode), req->r_ops[i].op, req, i,
861 req->r_tid);
862 break;
863 }
864
865 osd_data = osd_req_op_extent_osd_data(req, i);
866 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
867 len += osd_data->length;
868 num_pages = calc_pages_for((u64)osd_data->alignment,
869 (u64)osd_data->length);
870 total_pages += num_pages;
871 for (j = 0; j < num_pages; j++) {
872 page = osd_data->pages[j];
873 if (fscrypt_is_bounce_page(page)) {
874 page = fscrypt_pagecache_page(page);
875 fscrypt_free_bounce_page(osd_data->pages[j]);
876 osd_data->pages[j] = page;
877 }
878 BUG_ON(!page);
879 WARN_ON(!PageUptodate(page));
880
881 if (atomic_long_dec_return(&fsc->writeback_count) <
882 CONGESTION_OFF_THRESH(
883 fsc->mount_options->congestion_kb))
884 fsc->write_congested = false;
885
886 ceph_put_snap_context(detach_page_private(page));
887 end_page_writeback(page);
888 doutc(cl, "unlocking %p\n", page);
889
890 if (remove_page)
891 generic_error_remove_folio(inode->i_mapping,
892 page_folio(page));
893
894 unlock_page(page);
895 }
896 doutc(cl, "%llx.%llx wrote %llu bytes cleaned %d pages\n",
897 ceph_vinop(inode), osd_data->length,
898 rc >= 0 ? num_pages : 0);
899
900 release_pages(osd_data->pages, num_pages);
901 }
902
903 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
904 req->r_end_latency, len, rc);
905
906 ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
907
908 osd_data = osd_req_op_extent_osd_data(req, 0);
909 if (osd_data->pages_from_pool)
910 mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
911 else
912 kfree(osd_data->pages);
913 ceph_osdc_put_request(req);
914 ceph_dec_osd_stopping_blocker(fsc->mdsc);
915}
916
917/*
918 * initiate async writeback
919 */
920static int ceph_writepages_start(struct address_space *mapping,
921 struct writeback_control *wbc)
922{
923 struct inode *inode = mapping->host;
924 struct ceph_inode_info *ci = ceph_inode(inode);
925 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
926 struct ceph_client *cl = fsc->client;
927 struct ceph_vino vino = ceph_vino(inode);
928 pgoff_t index, start_index, end = -1;
929 struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
930 struct folio_batch fbatch;
931 int rc = 0;
932 unsigned int wsize = i_blocksize(inode);
933 struct ceph_osd_request *req = NULL;
934 struct ceph_writeback_ctl ceph_wbc;
935 bool should_loop, range_whole = false;
936 bool done = false;
937 bool caching = ceph_is_cache_enabled(inode);
938 xa_mark_t tag;
939
940 if (wbc->sync_mode == WB_SYNC_NONE &&
941 fsc->write_congested)
942 return 0;
943
944 doutc(cl, "%llx.%llx (mode=%s)\n", ceph_vinop(inode),
945 wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
946 (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
947
948 if (ceph_inode_is_shutdown(inode)) {
949 if (ci->i_wrbuffer_ref > 0) {
950 pr_warn_ratelimited_client(cl,
951 "%llx.%llx %lld forced umount\n",
952 ceph_vinop(inode), ceph_ino(inode));
953 }
954 mapping_set_error(mapping, -EIO);
955 return -EIO; /* we're in a forced umount, don't write! */
956 }
957 if (fsc->mount_options->wsize < wsize)
958 wsize = fsc->mount_options->wsize;
959
960 folio_batch_init(&fbatch);
961
962 start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
963 index = start_index;
964
965 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
966 tag = PAGECACHE_TAG_TOWRITE;
967 } else {
968 tag = PAGECACHE_TAG_DIRTY;
969 }
970retry:
971 /* find oldest snap context with dirty data */
972 snapc = get_oldest_context(inode, &ceph_wbc, NULL);
973 if (!snapc) {
974 /* hmm, why does writepages get called when there
975 is no dirty data? */
976 doutc(cl, " no snap context with dirty data?\n");
977 goto out;
978 }
979 doutc(cl, " oldest snapc is %p seq %lld (%d snaps)\n", snapc,
980 snapc->seq, snapc->num_snaps);
981
982 should_loop = false;
983 if (ceph_wbc.head_snapc && snapc != last_snapc) {
984 /* where to start/end? */
985 if (wbc->range_cyclic) {
986 index = start_index;
987 end = -1;
988 if (index > 0)
989 should_loop = true;
990 doutc(cl, " cyclic, start at %lu\n", index);
991 } else {
992 index = wbc->range_start >> PAGE_SHIFT;
993 end = wbc->range_end >> PAGE_SHIFT;
994 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
995 range_whole = true;
996 doutc(cl, " not cyclic, %lu to %lu\n", index, end);
997 }
998 } else if (!ceph_wbc.head_snapc) {
999 /* Do not respect wbc->range_{start,end}. Dirty pages
1000 * in that range can be associated with newer snapc.
1001 * They are not writeable until we write all dirty pages
1002 * associated with 'snapc' get written */
1003 if (index > 0)
1004 should_loop = true;
1005 doutc(cl, " non-head snapc, range whole\n");
1006 }
1007
1008 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1009 tag_pages_for_writeback(mapping, index, end);
1010
1011 ceph_put_snap_context(last_snapc);
1012 last_snapc = snapc;
1013
1014 while (!done && index <= end) {
1015 int num_ops = 0, op_idx;
1016 unsigned i, nr_folios, max_pages, locked_pages = 0;
1017 struct page **pages = NULL, **data_pages;
1018 struct page *page;
1019 pgoff_t strip_unit_end = 0;
1020 u64 offset = 0, len = 0;
1021 bool from_pool = false;
1022
1023 max_pages = wsize >> PAGE_SHIFT;
1024
1025get_more_pages:
1026 nr_folios = filemap_get_folios_tag(mapping, &index,
1027 end, tag, &fbatch);
1028 doutc(cl, "pagevec_lookup_range_tag got %d\n", nr_folios);
1029 if (!nr_folios && !locked_pages)
1030 break;
1031 for (i = 0; i < nr_folios && locked_pages < max_pages; i++) {
1032 page = &fbatch.folios[i]->page;
1033 doutc(cl, "? %p idx %lu\n", page, page->index);
1034 if (locked_pages == 0)
1035 lock_page(page); /* first page */
1036 else if (!trylock_page(page))
1037 break;
1038
1039 /* only dirty pages, or our accounting breaks */
1040 if (unlikely(!PageDirty(page)) ||
1041 unlikely(page->mapping != mapping)) {
1042 doutc(cl, "!dirty or !mapping %p\n", page);
1043 unlock_page(page);
1044 continue;
1045 }
1046 /* only if matching snap context */
1047 pgsnapc = page_snap_context(page);
1048 if (pgsnapc != snapc) {
1049 doutc(cl, "page snapc %p %lld != oldest %p %lld\n",
1050 pgsnapc, pgsnapc->seq, snapc, snapc->seq);
1051 if (!should_loop &&
1052 !ceph_wbc.head_snapc &&
1053 wbc->sync_mode != WB_SYNC_NONE)
1054 should_loop = true;
1055 unlock_page(page);
1056 continue;
1057 }
1058 if (page_offset(page) >= ceph_wbc.i_size) {
1059 struct folio *folio = page_folio(page);
1060
1061 doutc(cl, "folio at %lu beyond eof %llu\n",
1062 folio->index, ceph_wbc.i_size);
1063 if ((ceph_wbc.size_stable ||
1064 folio_pos(folio) >= i_size_read(inode)) &&
1065 folio_clear_dirty_for_io(folio))
1066 folio_invalidate(folio, 0,
1067 folio_size(folio));
1068 folio_unlock(folio);
1069 continue;
1070 }
1071 if (strip_unit_end && (page->index > strip_unit_end)) {
1072 doutc(cl, "end of strip unit %p\n", page);
1073 unlock_page(page);
1074 break;
1075 }
1076 if (PageWriteback(page) || PageFsCache(page)) {
1077 if (wbc->sync_mode == WB_SYNC_NONE) {
1078 doutc(cl, "%p under writeback\n", page);
1079 unlock_page(page);
1080 continue;
1081 }
1082 doutc(cl, "waiting on writeback %p\n", page);
1083 wait_on_page_writeback(page);
1084 wait_on_page_fscache(page);
1085 }
1086
1087 if (!clear_page_dirty_for_io(page)) {
1088 doutc(cl, "%p !clear_page_dirty_for_io\n", page);
1089 unlock_page(page);
1090 continue;
1091 }
1092
1093 /*
1094 * We have something to write. If this is
1095 * the first locked page this time through,
1096 * calculate max possinle write size and
1097 * allocate a page array
1098 */
1099 if (locked_pages == 0) {
1100 u64 objnum;
1101 u64 objoff;
1102 u32 xlen;
1103
1104 /* prepare async write request */
1105 offset = (u64)page_offset(page);
1106 ceph_calc_file_object_mapping(&ci->i_layout,
1107 offset, wsize,
1108 &objnum, &objoff,
1109 &xlen);
1110 len = xlen;
1111
1112 num_ops = 1;
1113 strip_unit_end = page->index +
1114 ((len - 1) >> PAGE_SHIFT);
1115
1116 BUG_ON(pages);
1117 max_pages = calc_pages_for(0, (u64)len);
1118 pages = kmalloc_array(max_pages,
1119 sizeof(*pages),
1120 GFP_NOFS);
1121 if (!pages) {
1122 from_pool = true;
1123 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1124 BUG_ON(!pages);
1125 }
1126
1127 len = 0;
1128 } else if (page->index !=
1129 (offset + len) >> PAGE_SHIFT) {
1130 if (num_ops >= (from_pool ? CEPH_OSD_SLAB_OPS :
1131 CEPH_OSD_MAX_OPS)) {
1132 redirty_page_for_writepage(wbc, page);
1133 unlock_page(page);
1134 break;
1135 }
1136
1137 num_ops++;
1138 offset = (u64)page_offset(page);
1139 len = 0;
1140 }
1141
1142 /* note position of first page in fbatch */
1143 doutc(cl, "%llx.%llx will write page %p idx %lu\n",
1144 ceph_vinop(inode), page, page->index);
1145
1146 if (atomic_long_inc_return(&fsc->writeback_count) >
1147 CONGESTION_ON_THRESH(
1148 fsc->mount_options->congestion_kb))
1149 fsc->write_congested = true;
1150
1151 if (IS_ENCRYPTED(inode)) {
1152 pages[locked_pages] =
1153 fscrypt_encrypt_pagecache_blocks(page,
1154 PAGE_SIZE, 0,
1155 locked_pages ? GFP_NOWAIT : GFP_NOFS);
1156 if (IS_ERR(pages[locked_pages])) {
1157 if (PTR_ERR(pages[locked_pages]) == -EINVAL)
1158 pr_err_client(cl,
1159 "inode->i_blkbits=%hhu\n",
1160 inode->i_blkbits);
1161 /* better not fail on first page! */
1162 BUG_ON(locked_pages == 0);
1163 pages[locked_pages] = NULL;
1164 redirty_page_for_writepage(wbc, page);
1165 unlock_page(page);
1166 break;
1167 }
1168 ++locked_pages;
1169 } else {
1170 pages[locked_pages++] = page;
1171 }
1172
1173 fbatch.folios[i] = NULL;
1174 len += thp_size(page);
1175 }
1176
1177 /* did we get anything? */
1178 if (!locked_pages)
1179 goto release_folios;
1180 if (i) {
1181 unsigned j, n = 0;
1182 /* shift unused page to beginning of fbatch */
1183 for (j = 0; j < nr_folios; j++) {
1184 if (!fbatch.folios[j])
1185 continue;
1186 if (n < j)
1187 fbatch.folios[n] = fbatch.folios[j];
1188 n++;
1189 }
1190 fbatch.nr = n;
1191
1192 if (nr_folios && i == nr_folios &&
1193 locked_pages < max_pages) {
1194 doutc(cl, "reached end fbatch, trying for more\n");
1195 folio_batch_release(&fbatch);
1196 goto get_more_pages;
1197 }
1198 }
1199
1200new_request:
1201 offset = ceph_fscrypt_page_offset(pages[0]);
1202 len = wsize;
1203
1204 req = ceph_osdc_new_request(&fsc->client->osdc,
1205 &ci->i_layout, vino,
1206 offset, &len, 0, num_ops,
1207 CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1208 snapc, ceph_wbc.truncate_seq,
1209 ceph_wbc.truncate_size, false);
1210 if (IS_ERR(req)) {
1211 req = ceph_osdc_new_request(&fsc->client->osdc,
1212 &ci->i_layout, vino,
1213 offset, &len, 0,
1214 min(num_ops,
1215 CEPH_OSD_SLAB_OPS),
1216 CEPH_OSD_OP_WRITE,
1217 CEPH_OSD_FLAG_WRITE,
1218 snapc, ceph_wbc.truncate_seq,
1219 ceph_wbc.truncate_size, true);
1220 BUG_ON(IS_ERR(req));
1221 }
1222 BUG_ON(len < ceph_fscrypt_page_offset(pages[locked_pages - 1]) +
1223 thp_size(pages[locked_pages - 1]) - offset);
1224
1225 if (!ceph_inc_osd_stopping_blocker(fsc->mdsc)) {
1226 rc = -EIO;
1227 goto release_folios;
1228 }
1229 req->r_callback = writepages_finish;
1230 req->r_inode = inode;
1231
1232 /* Format the osd request message and submit the write */
1233 len = 0;
1234 data_pages = pages;
1235 op_idx = 0;
1236 for (i = 0; i < locked_pages; i++) {
1237 struct page *page = ceph_fscrypt_pagecache_page(pages[i]);
1238
1239 u64 cur_offset = page_offset(page);
1240 /*
1241 * Discontinuity in page range? Ceph can handle that by just passing
1242 * multiple extents in the write op.
1243 */
1244 if (offset + len != cur_offset) {
1245 /* If it's full, stop here */
1246 if (op_idx + 1 == req->r_num_ops)
1247 break;
1248
1249 /* Kick off an fscache write with what we have so far. */
1250 ceph_fscache_write_to_cache(inode, offset, len, caching);
1251
1252 /* Start a new extent */
1253 osd_req_op_extent_dup_last(req, op_idx,
1254 cur_offset - offset);
1255 doutc(cl, "got pages at %llu~%llu\n", offset,
1256 len);
1257 osd_req_op_extent_osd_data_pages(req, op_idx,
1258 data_pages, len, 0,
1259 from_pool, false);
1260 osd_req_op_extent_update(req, op_idx, len);
1261
1262 len = 0;
1263 offset = cur_offset;
1264 data_pages = pages + i;
1265 op_idx++;
1266 }
1267
1268 set_page_writeback(page);
1269 if (caching)
1270 ceph_set_page_fscache(page);
1271 len += thp_size(page);
1272 }
1273 ceph_fscache_write_to_cache(inode, offset, len, caching);
1274
1275 if (ceph_wbc.size_stable) {
1276 len = min(len, ceph_wbc.i_size - offset);
1277 } else if (i == locked_pages) {
1278 /* writepages_finish() clears writeback pages
1279 * according to the data length, so make sure
1280 * data length covers all locked pages */
1281 u64 min_len = len + 1 - thp_size(page);
1282 len = get_writepages_data_length(inode, pages[i - 1],
1283 offset);
1284 len = max(len, min_len);
1285 }
1286 if (IS_ENCRYPTED(inode))
1287 len = round_up(len, CEPH_FSCRYPT_BLOCK_SIZE);
1288
1289 doutc(cl, "got pages at %llu~%llu\n", offset, len);
1290
1291 if (IS_ENCRYPTED(inode) &&
1292 ((offset | len) & ~CEPH_FSCRYPT_BLOCK_MASK))
1293 pr_warn_client(cl,
1294 "bad encrypted write offset=%lld len=%llu\n",
1295 offset, len);
1296
1297 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1298 0, from_pool, false);
1299 osd_req_op_extent_update(req, op_idx, len);
1300
1301 BUG_ON(op_idx + 1 != req->r_num_ops);
1302
1303 from_pool = false;
1304 if (i < locked_pages) {
1305 BUG_ON(num_ops <= req->r_num_ops);
1306 num_ops -= req->r_num_ops;
1307 locked_pages -= i;
1308
1309 /* allocate new pages array for next request */
1310 data_pages = pages;
1311 pages = kmalloc_array(locked_pages, sizeof(*pages),
1312 GFP_NOFS);
1313 if (!pages) {
1314 from_pool = true;
1315 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1316 BUG_ON(!pages);
1317 }
1318 memcpy(pages, data_pages + i,
1319 locked_pages * sizeof(*pages));
1320 memset(data_pages + i, 0,
1321 locked_pages * sizeof(*pages));
1322 } else {
1323 BUG_ON(num_ops != req->r_num_ops);
1324 index = pages[i - 1]->index + 1;
1325 /* request message now owns the pages array */
1326 pages = NULL;
1327 }
1328
1329 req->r_mtime = inode_get_mtime(inode);
1330 ceph_osdc_start_request(&fsc->client->osdc, req);
1331 req = NULL;
1332
1333 wbc->nr_to_write -= i;
1334 if (pages)
1335 goto new_request;
1336
1337 /*
1338 * We stop writing back only if we are not doing
1339 * integrity sync. In case of integrity sync we have to
1340 * keep going until we have written all the pages
1341 * we tagged for writeback prior to entering this loop.
1342 */
1343 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1344 done = true;
1345
1346release_folios:
1347 doutc(cl, "folio_batch release on %d folios (%p)\n",
1348 (int)fbatch.nr, fbatch.nr ? fbatch.folios[0] : NULL);
1349 folio_batch_release(&fbatch);
1350 }
1351
1352 if (should_loop && !done) {
1353 /* more to do; loop back to beginning of file */
1354 doutc(cl, "looping back to beginning of file\n");
1355 end = start_index - 1; /* OK even when start_index == 0 */
1356
1357 /* to write dirty pages associated with next snapc,
1358 * we need to wait until current writes complete */
1359 if (wbc->sync_mode != WB_SYNC_NONE &&
1360 start_index == 0 && /* all dirty pages were checked */
1361 !ceph_wbc.head_snapc) {
1362 struct page *page;
1363 unsigned i, nr;
1364 index = 0;
1365 while ((index <= end) &&
1366 (nr = filemap_get_folios_tag(mapping, &index,
1367 (pgoff_t)-1,
1368 PAGECACHE_TAG_WRITEBACK,
1369 &fbatch))) {
1370 for (i = 0; i < nr; i++) {
1371 page = &fbatch.folios[i]->page;
1372 if (page_snap_context(page) != snapc)
1373 continue;
1374 wait_on_page_writeback(page);
1375 }
1376 folio_batch_release(&fbatch);
1377 cond_resched();
1378 }
1379 }
1380
1381 start_index = 0;
1382 index = 0;
1383 goto retry;
1384 }
1385
1386 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1387 mapping->writeback_index = index;
1388
1389out:
1390 ceph_osdc_put_request(req);
1391 ceph_put_snap_context(last_snapc);
1392 doutc(cl, "%llx.%llx dend - startone, rc = %d\n", ceph_vinop(inode),
1393 rc);
1394 return rc;
1395}
1396
1397
1398
1399/*
1400 * See if a given @snapc is either writeable, or already written.
1401 */
1402static int context_is_writeable_or_written(struct inode *inode,
1403 struct ceph_snap_context *snapc)
1404{
1405 struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1406 int ret = !oldest || snapc->seq <= oldest->seq;
1407
1408 ceph_put_snap_context(oldest);
1409 return ret;
1410}
1411
1412/**
1413 * ceph_find_incompatible - find an incompatible context and return it
1414 * @page: page being dirtied
1415 *
1416 * We are only allowed to write into/dirty a page if the page is
1417 * clean, or already dirty within the same snap context. Returns a
1418 * conflicting context if there is one, NULL if there isn't, or a
1419 * negative error code on other errors.
1420 *
1421 * Must be called with page lock held.
1422 */
1423static struct ceph_snap_context *
1424ceph_find_incompatible(struct page *page)
1425{
1426 struct inode *inode = page->mapping->host;
1427 struct ceph_client *cl = ceph_inode_to_client(inode);
1428 struct ceph_inode_info *ci = ceph_inode(inode);
1429
1430 if (ceph_inode_is_shutdown(inode)) {
1431 doutc(cl, " %llx.%llx page %p is shutdown\n",
1432 ceph_vinop(inode), page);
1433 return ERR_PTR(-ESTALE);
1434 }
1435
1436 for (;;) {
1437 struct ceph_snap_context *snapc, *oldest;
1438
1439 wait_on_page_writeback(page);
1440
1441 snapc = page_snap_context(page);
1442 if (!snapc || snapc == ci->i_head_snapc)
1443 break;
1444
1445 /*
1446 * this page is already dirty in another (older) snap
1447 * context! is it writeable now?
1448 */
1449 oldest = get_oldest_context(inode, NULL, NULL);
1450 if (snapc->seq > oldest->seq) {
1451 /* not writeable -- return it for the caller to deal with */
1452 ceph_put_snap_context(oldest);
1453 doutc(cl, " %llx.%llx page %p snapc %p not current or oldest\n",
1454 ceph_vinop(inode), page, snapc);
1455 return ceph_get_snap_context(snapc);
1456 }
1457 ceph_put_snap_context(oldest);
1458
1459 /* yay, writeable, do it now (without dropping page lock) */
1460 doutc(cl, " %llx.%llx page %p snapc %p not current, but oldest\n",
1461 ceph_vinop(inode), page, snapc);
1462 if (clear_page_dirty_for_io(page)) {
1463 int r = writepage_nounlock(page, NULL);
1464 if (r < 0)
1465 return ERR_PTR(r);
1466 }
1467 }
1468 return NULL;
1469}
1470
1471static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
1472 struct folio **foliop, void **_fsdata)
1473{
1474 struct inode *inode = file_inode(file);
1475 struct ceph_inode_info *ci = ceph_inode(inode);
1476 struct ceph_snap_context *snapc;
1477
1478 snapc = ceph_find_incompatible(folio_page(*foliop, 0));
1479 if (snapc) {
1480 int r;
1481
1482 folio_unlock(*foliop);
1483 folio_put(*foliop);
1484 *foliop = NULL;
1485 if (IS_ERR(snapc))
1486 return PTR_ERR(snapc);
1487
1488 ceph_queue_writeback(inode);
1489 r = wait_event_killable(ci->i_cap_wq,
1490 context_is_writeable_or_written(inode, snapc));
1491 ceph_put_snap_context(snapc);
1492 return r == 0 ? -EAGAIN : r;
1493 }
1494 return 0;
1495}
1496
1497/*
1498 * We are only allowed to write into/dirty the page if the page is
1499 * clean, or already dirty within the same snap context.
1500 */
1501static int ceph_write_begin(struct file *file, struct address_space *mapping,
1502 loff_t pos, unsigned len,
1503 struct page **pagep, void **fsdata)
1504{
1505 struct inode *inode = file_inode(file);
1506 struct ceph_inode_info *ci = ceph_inode(inode);
1507 struct folio *folio = NULL;
1508 int r;
1509
1510 r = netfs_write_begin(&ci->netfs, file, inode->i_mapping, pos, len, &folio, NULL);
1511 if (r < 0)
1512 return r;
1513
1514 folio_wait_fscache(folio);
1515 WARN_ON_ONCE(!folio_test_locked(folio));
1516 *pagep = &folio->page;
1517 return 0;
1518}
1519
1520/*
1521 * we don't do anything in here that simple_write_end doesn't do
1522 * except adjust dirty page accounting
1523 */
1524static int ceph_write_end(struct file *file, struct address_space *mapping,
1525 loff_t pos, unsigned len, unsigned copied,
1526 struct page *subpage, void *fsdata)
1527{
1528 struct folio *folio = page_folio(subpage);
1529 struct inode *inode = file_inode(file);
1530 struct ceph_client *cl = ceph_inode_to_client(inode);
1531 bool check_cap = false;
1532
1533 doutc(cl, "%llx.%llx file %p folio %p %d~%d (%d)\n", ceph_vinop(inode),
1534 file, folio, (int)pos, (int)copied, (int)len);
1535
1536 if (!folio_test_uptodate(folio)) {
1537 /* just return that nothing was copied on a short copy */
1538 if (copied < len) {
1539 copied = 0;
1540 goto out;
1541 }
1542 folio_mark_uptodate(folio);
1543 }
1544
1545 /* did file size increase? */
1546 if (pos+copied > i_size_read(inode))
1547 check_cap = ceph_inode_set_size(inode, pos+copied);
1548
1549 folio_mark_dirty(folio);
1550
1551out:
1552 folio_unlock(folio);
1553 folio_put(folio);
1554
1555 if (check_cap)
1556 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY);
1557
1558 return copied;
1559}
1560
1561const struct address_space_operations ceph_aops = {
1562 .read_folio = netfs_read_folio,
1563 .readahead = netfs_readahead,
1564 .writepage = ceph_writepage,
1565 .writepages = ceph_writepages_start,
1566 .write_begin = ceph_write_begin,
1567 .write_end = ceph_write_end,
1568 .dirty_folio = ceph_dirty_folio,
1569 .invalidate_folio = ceph_invalidate_folio,
1570 .release_folio = netfs_release_folio,
1571 .direct_IO = noop_direct_IO,
1572};
1573
1574static void ceph_block_sigs(sigset_t *oldset)
1575{
1576 sigset_t mask;
1577 siginitsetinv(&mask, sigmask(SIGKILL));
1578 sigprocmask(SIG_BLOCK, &mask, oldset);
1579}
1580
1581static void ceph_restore_sigs(sigset_t *oldset)
1582{
1583 sigprocmask(SIG_SETMASK, oldset, NULL);
1584}
1585
1586/*
1587 * vm ops
1588 */
1589static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1590{
1591 struct vm_area_struct *vma = vmf->vma;
1592 struct inode *inode = file_inode(vma->vm_file);
1593 struct ceph_inode_info *ci = ceph_inode(inode);
1594 struct ceph_client *cl = ceph_inode_to_client(inode);
1595 struct ceph_file_info *fi = vma->vm_file->private_data;
1596 loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1597 int want, got, err;
1598 sigset_t oldset;
1599 vm_fault_t ret = VM_FAULT_SIGBUS;
1600
1601 if (ceph_inode_is_shutdown(inode))
1602 return ret;
1603
1604 ceph_block_sigs(&oldset);
1605
1606 doutc(cl, "%llx.%llx %llu trying to get caps\n",
1607 ceph_vinop(inode), off);
1608 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1609 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1610 else
1611 want = CEPH_CAP_FILE_CACHE;
1612
1613 got = 0;
1614 err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1, &got);
1615 if (err < 0)
1616 goto out_restore;
1617
1618 doutc(cl, "%llx.%llx %llu got cap refs on %s\n", ceph_vinop(inode),
1619 off, ceph_cap_string(got));
1620
1621 if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1622 !ceph_has_inline_data(ci)) {
1623 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1624 ceph_add_rw_context(fi, &rw_ctx);
1625 ret = filemap_fault(vmf);
1626 ceph_del_rw_context(fi, &rw_ctx);
1627 doutc(cl, "%llx.%llx %llu drop cap refs %s ret %x\n",
1628 ceph_vinop(inode), off, ceph_cap_string(got), ret);
1629 } else
1630 err = -EAGAIN;
1631
1632 ceph_put_cap_refs(ci, got);
1633
1634 if (err != -EAGAIN)
1635 goto out_restore;
1636
1637 /* read inline data */
1638 if (off >= PAGE_SIZE) {
1639 /* does not support inline data > PAGE_SIZE */
1640 ret = VM_FAULT_SIGBUS;
1641 } else {
1642 struct address_space *mapping = inode->i_mapping;
1643 struct page *page;
1644
1645 filemap_invalidate_lock_shared(mapping);
1646 page = find_or_create_page(mapping, 0,
1647 mapping_gfp_constraint(mapping, ~__GFP_FS));
1648 if (!page) {
1649 ret = VM_FAULT_OOM;
1650 goto out_inline;
1651 }
1652 err = __ceph_do_getattr(inode, page,
1653 CEPH_STAT_CAP_INLINE_DATA, true);
1654 if (err < 0 || off >= i_size_read(inode)) {
1655 unlock_page(page);
1656 put_page(page);
1657 ret = vmf_error(err);
1658 goto out_inline;
1659 }
1660 if (err < PAGE_SIZE)
1661 zero_user_segment(page, err, PAGE_SIZE);
1662 else
1663 flush_dcache_page(page);
1664 SetPageUptodate(page);
1665 vmf->page = page;
1666 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1667out_inline:
1668 filemap_invalidate_unlock_shared(mapping);
1669 doutc(cl, "%llx.%llx %llu read inline data ret %x\n",
1670 ceph_vinop(inode), off, ret);
1671 }
1672out_restore:
1673 ceph_restore_sigs(&oldset);
1674 if (err < 0)
1675 ret = vmf_error(err);
1676
1677 return ret;
1678}
1679
1680static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1681{
1682 struct vm_area_struct *vma = vmf->vma;
1683 struct inode *inode = file_inode(vma->vm_file);
1684 struct ceph_client *cl = ceph_inode_to_client(inode);
1685 struct ceph_inode_info *ci = ceph_inode(inode);
1686 struct ceph_file_info *fi = vma->vm_file->private_data;
1687 struct ceph_cap_flush *prealloc_cf;
1688 struct page *page = vmf->page;
1689 loff_t off = page_offset(page);
1690 loff_t size = i_size_read(inode);
1691 size_t len;
1692 int want, got, err;
1693 sigset_t oldset;
1694 vm_fault_t ret = VM_FAULT_SIGBUS;
1695
1696 if (ceph_inode_is_shutdown(inode))
1697 return ret;
1698
1699 prealloc_cf = ceph_alloc_cap_flush();
1700 if (!prealloc_cf)
1701 return VM_FAULT_OOM;
1702
1703 sb_start_pagefault(inode->i_sb);
1704 ceph_block_sigs(&oldset);
1705
1706 if (off + thp_size(page) <= size)
1707 len = thp_size(page);
1708 else
1709 len = offset_in_thp(page, size);
1710
1711 doutc(cl, "%llx.%llx %llu~%zd getting caps i_size %llu\n",
1712 ceph_vinop(inode), off, len, size);
1713 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1714 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1715 else
1716 want = CEPH_CAP_FILE_BUFFER;
1717
1718 got = 0;
1719 err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len, &got);
1720 if (err < 0)
1721 goto out_free;
1722
1723 doutc(cl, "%llx.%llx %llu~%zd got cap refs on %s\n", ceph_vinop(inode),
1724 off, len, ceph_cap_string(got));
1725
1726 /* Update time before taking page lock */
1727 file_update_time(vma->vm_file);
1728 inode_inc_iversion_raw(inode);
1729
1730 do {
1731 struct ceph_snap_context *snapc;
1732
1733 lock_page(page);
1734
1735 if (page_mkwrite_check_truncate(page, inode) < 0) {
1736 unlock_page(page);
1737 ret = VM_FAULT_NOPAGE;
1738 break;
1739 }
1740
1741 snapc = ceph_find_incompatible(page);
1742 if (!snapc) {
1743 /* success. we'll keep the page locked. */
1744 set_page_dirty(page);
1745 ret = VM_FAULT_LOCKED;
1746 break;
1747 }
1748
1749 unlock_page(page);
1750
1751 if (IS_ERR(snapc)) {
1752 ret = VM_FAULT_SIGBUS;
1753 break;
1754 }
1755
1756 ceph_queue_writeback(inode);
1757 err = wait_event_killable(ci->i_cap_wq,
1758 context_is_writeable_or_written(inode, snapc));
1759 ceph_put_snap_context(snapc);
1760 } while (err == 0);
1761
1762 if (ret == VM_FAULT_LOCKED) {
1763 int dirty;
1764 spin_lock(&ci->i_ceph_lock);
1765 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1766 &prealloc_cf);
1767 spin_unlock(&ci->i_ceph_lock);
1768 if (dirty)
1769 __mark_inode_dirty(inode, dirty);
1770 }
1771
1772 doutc(cl, "%llx.%llx %llu~%zd dropping cap refs on %s ret %x\n",
1773 ceph_vinop(inode), off, len, ceph_cap_string(got), ret);
1774 ceph_put_cap_refs_async(ci, got);
1775out_free:
1776 ceph_restore_sigs(&oldset);
1777 sb_end_pagefault(inode->i_sb);
1778 ceph_free_cap_flush(prealloc_cf);
1779 if (err < 0)
1780 ret = vmf_error(err);
1781 return ret;
1782}
1783
1784void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1785 char *data, size_t len)
1786{
1787 struct ceph_client *cl = ceph_inode_to_client(inode);
1788 struct address_space *mapping = inode->i_mapping;
1789 struct page *page;
1790
1791 if (locked_page) {
1792 page = locked_page;
1793 } else {
1794 if (i_size_read(inode) == 0)
1795 return;
1796 page = find_or_create_page(mapping, 0,
1797 mapping_gfp_constraint(mapping,
1798 ~__GFP_FS));
1799 if (!page)
1800 return;
1801 if (PageUptodate(page)) {
1802 unlock_page(page);
1803 put_page(page);
1804 return;
1805 }
1806 }
1807
1808 doutc(cl, "%p %llx.%llx len %zu locked_page %p\n", inode,
1809 ceph_vinop(inode), len, locked_page);
1810
1811 if (len > 0) {
1812 void *kaddr = kmap_atomic(page);
1813 memcpy(kaddr, data, len);
1814 kunmap_atomic(kaddr);
1815 }
1816
1817 if (page != locked_page) {
1818 if (len < PAGE_SIZE)
1819 zero_user_segment(page, len, PAGE_SIZE);
1820 else
1821 flush_dcache_page(page);
1822
1823 SetPageUptodate(page);
1824 unlock_page(page);
1825 put_page(page);
1826 }
1827}
1828
1829int ceph_uninline_data(struct file *file)
1830{
1831 struct inode *inode = file_inode(file);
1832 struct ceph_inode_info *ci = ceph_inode(inode);
1833 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
1834 struct ceph_client *cl = fsc->client;
1835 struct ceph_osd_request *req = NULL;
1836 struct ceph_cap_flush *prealloc_cf = NULL;
1837 struct folio *folio = NULL;
1838 u64 inline_version = CEPH_INLINE_NONE;
1839 struct page *pages[1];
1840 int err = 0;
1841 u64 len;
1842
1843 spin_lock(&ci->i_ceph_lock);
1844 inline_version = ci->i_inline_version;
1845 spin_unlock(&ci->i_ceph_lock);
1846
1847 doutc(cl, "%llx.%llx inline_version %llu\n", ceph_vinop(inode),
1848 inline_version);
1849
1850 if (ceph_inode_is_shutdown(inode)) {
1851 err = -EIO;
1852 goto out;
1853 }
1854
1855 if (inline_version == CEPH_INLINE_NONE)
1856 return 0;
1857
1858 prealloc_cf = ceph_alloc_cap_flush();
1859 if (!prealloc_cf)
1860 return -ENOMEM;
1861
1862 if (inline_version == 1) /* initial version, no data */
1863 goto out_uninline;
1864
1865 folio = read_mapping_folio(inode->i_mapping, 0, file);
1866 if (IS_ERR(folio)) {
1867 err = PTR_ERR(folio);
1868 goto out;
1869 }
1870
1871 folio_lock(folio);
1872
1873 len = i_size_read(inode);
1874 if (len > folio_size(folio))
1875 len = folio_size(folio);
1876
1877 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1878 ceph_vino(inode), 0, &len, 0, 1,
1879 CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1880 NULL, 0, 0, false);
1881 if (IS_ERR(req)) {
1882 err = PTR_ERR(req);
1883 goto out_unlock;
1884 }
1885
1886 req->r_mtime = inode_get_mtime(inode);
1887 ceph_osdc_start_request(&fsc->client->osdc, req);
1888 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1889 ceph_osdc_put_request(req);
1890 if (err < 0)
1891 goto out_unlock;
1892
1893 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1894 ceph_vino(inode), 0, &len, 1, 3,
1895 CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1896 NULL, ci->i_truncate_seq,
1897 ci->i_truncate_size, false);
1898 if (IS_ERR(req)) {
1899 err = PTR_ERR(req);
1900 goto out_unlock;
1901 }
1902
1903 pages[0] = folio_page(folio, 0);
1904 osd_req_op_extent_osd_data_pages(req, 1, pages, len, 0, false, false);
1905
1906 {
1907 __le64 xattr_buf = cpu_to_le64(inline_version);
1908 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1909 "inline_version", &xattr_buf,
1910 sizeof(xattr_buf),
1911 CEPH_OSD_CMPXATTR_OP_GT,
1912 CEPH_OSD_CMPXATTR_MODE_U64);
1913 if (err)
1914 goto out_put_req;
1915 }
1916
1917 {
1918 char xattr_buf[32];
1919 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1920 "%llu", inline_version);
1921 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1922 "inline_version",
1923 xattr_buf, xattr_len, 0, 0);
1924 if (err)
1925 goto out_put_req;
1926 }
1927
1928 req->r_mtime = inode_get_mtime(inode);
1929 ceph_osdc_start_request(&fsc->client->osdc, req);
1930 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1931
1932 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1933 req->r_end_latency, len, err);
1934
1935out_uninline:
1936 if (!err) {
1937 int dirty;
1938
1939 /* Set to CAP_INLINE_NONE and dirty the caps */
1940 down_read(&fsc->mdsc->snap_rwsem);
1941 spin_lock(&ci->i_ceph_lock);
1942 ci->i_inline_version = CEPH_INLINE_NONE;
1943 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR, &prealloc_cf);
1944 spin_unlock(&ci->i_ceph_lock);
1945 up_read(&fsc->mdsc->snap_rwsem);
1946 if (dirty)
1947 __mark_inode_dirty(inode, dirty);
1948 }
1949out_put_req:
1950 ceph_osdc_put_request(req);
1951 if (err == -ECANCELED)
1952 err = 0;
1953out_unlock:
1954 if (folio) {
1955 folio_unlock(folio);
1956 folio_put(folio);
1957 }
1958out:
1959 ceph_free_cap_flush(prealloc_cf);
1960 doutc(cl, "%llx.%llx inline_version %llu = %d\n",
1961 ceph_vinop(inode), inline_version, err);
1962 return err;
1963}
1964
1965static const struct vm_operations_struct ceph_vmops = {
1966 .fault = ceph_filemap_fault,
1967 .page_mkwrite = ceph_page_mkwrite,
1968};
1969
1970int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1971{
1972 struct address_space *mapping = file->f_mapping;
1973
1974 if (!mapping->a_ops->read_folio)
1975 return -ENOEXEC;
1976 vma->vm_ops = &ceph_vmops;
1977 return 0;
1978}
1979
1980enum {
1981 POOL_READ = 1,
1982 POOL_WRITE = 2,
1983};
1984
1985static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1986 s64 pool, struct ceph_string *pool_ns)
1987{
1988 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(&ci->netfs.inode);
1989 struct ceph_mds_client *mdsc = fsc->mdsc;
1990 struct ceph_client *cl = fsc->client;
1991 struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1992 struct rb_node **p, *parent;
1993 struct ceph_pool_perm *perm;
1994 struct page **pages;
1995 size_t pool_ns_len;
1996 int err = 0, err2 = 0, have = 0;
1997
1998 down_read(&mdsc->pool_perm_rwsem);
1999 p = &mdsc->pool_perm_tree.rb_node;
2000 while (*p) {
2001 perm = rb_entry(*p, struct ceph_pool_perm, node);
2002 if (pool < perm->pool)
2003 p = &(*p)->rb_left;
2004 else if (pool > perm->pool)
2005 p = &(*p)->rb_right;
2006 else {
2007 int ret = ceph_compare_string(pool_ns,
2008 perm->pool_ns,
2009 perm->pool_ns_len);
2010 if (ret < 0)
2011 p = &(*p)->rb_left;
2012 else if (ret > 0)
2013 p = &(*p)->rb_right;
2014 else {
2015 have = perm->perm;
2016 break;
2017 }
2018 }
2019 }
2020 up_read(&mdsc->pool_perm_rwsem);
2021 if (*p)
2022 goto out;
2023
2024 if (pool_ns)
2025 doutc(cl, "pool %lld ns %.*s no perm cached\n", pool,
2026 (int)pool_ns->len, pool_ns->str);
2027 else
2028 doutc(cl, "pool %lld no perm cached\n", pool);
2029
2030 down_write(&mdsc->pool_perm_rwsem);
2031 p = &mdsc->pool_perm_tree.rb_node;
2032 parent = NULL;
2033 while (*p) {
2034 parent = *p;
2035 perm = rb_entry(parent, struct ceph_pool_perm, node);
2036 if (pool < perm->pool)
2037 p = &(*p)->rb_left;
2038 else if (pool > perm->pool)
2039 p = &(*p)->rb_right;
2040 else {
2041 int ret = ceph_compare_string(pool_ns,
2042 perm->pool_ns,
2043 perm->pool_ns_len);
2044 if (ret < 0)
2045 p = &(*p)->rb_left;
2046 else if (ret > 0)
2047 p = &(*p)->rb_right;
2048 else {
2049 have = perm->perm;
2050 break;
2051 }
2052 }
2053 }
2054 if (*p) {
2055 up_write(&mdsc->pool_perm_rwsem);
2056 goto out;
2057 }
2058
2059 rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2060 1, false, GFP_NOFS);
2061 if (!rd_req) {
2062 err = -ENOMEM;
2063 goto out_unlock;
2064 }
2065
2066 rd_req->r_flags = CEPH_OSD_FLAG_READ;
2067 osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
2068 rd_req->r_base_oloc.pool = pool;
2069 if (pool_ns)
2070 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
2071 ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
2072
2073 err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
2074 if (err)
2075 goto out_unlock;
2076
2077 wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2078 1, false, GFP_NOFS);
2079 if (!wr_req) {
2080 err = -ENOMEM;
2081 goto out_unlock;
2082 }
2083
2084 wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
2085 osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
2086 ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
2087 ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
2088
2089 err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
2090 if (err)
2091 goto out_unlock;
2092
2093 /* one page should be large enough for STAT data */
2094 pages = ceph_alloc_page_vector(1, GFP_KERNEL);
2095 if (IS_ERR(pages)) {
2096 err = PTR_ERR(pages);
2097 goto out_unlock;
2098 }
2099
2100 osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
2101 0, false, true);
2102 ceph_osdc_start_request(&fsc->client->osdc, rd_req);
2103
2104 wr_req->r_mtime = inode_get_mtime(&ci->netfs.inode);
2105 ceph_osdc_start_request(&fsc->client->osdc, wr_req);
2106
2107 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
2108 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
2109
2110 if (err >= 0 || err == -ENOENT)
2111 have |= POOL_READ;
2112 else if (err != -EPERM) {
2113 if (err == -EBLOCKLISTED)
2114 fsc->blocklisted = true;
2115 goto out_unlock;
2116 }
2117
2118 if (err2 == 0 || err2 == -EEXIST)
2119 have |= POOL_WRITE;
2120 else if (err2 != -EPERM) {
2121 if (err2 == -EBLOCKLISTED)
2122 fsc->blocklisted = true;
2123 err = err2;
2124 goto out_unlock;
2125 }
2126
2127 pool_ns_len = pool_ns ? pool_ns->len : 0;
2128 perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
2129 if (!perm) {
2130 err = -ENOMEM;
2131 goto out_unlock;
2132 }
2133
2134 perm->pool = pool;
2135 perm->perm = have;
2136 perm->pool_ns_len = pool_ns_len;
2137 if (pool_ns_len > 0)
2138 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
2139 perm->pool_ns[pool_ns_len] = 0;
2140
2141 rb_link_node(&perm->node, parent, p);
2142 rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
2143 err = 0;
2144out_unlock:
2145 up_write(&mdsc->pool_perm_rwsem);
2146
2147 ceph_osdc_put_request(rd_req);
2148 ceph_osdc_put_request(wr_req);
2149out:
2150 if (!err)
2151 err = have;
2152 if (pool_ns)
2153 doutc(cl, "pool %lld ns %.*s result = %d\n", pool,
2154 (int)pool_ns->len, pool_ns->str, err);
2155 else
2156 doutc(cl, "pool %lld result = %d\n", pool, err);
2157 return err;
2158}
2159
2160int ceph_pool_perm_check(struct inode *inode, int need)
2161{
2162 struct ceph_client *cl = ceph_inode_to_client(inode);
2163 struct ceph_inode_info *ci = ceph_inode(inode);
2164 struct ceph_string *pool_ns;
2165 s64 pool;
2166 int ret, flags;
2167
2168 /* Only need to do this for regular files */
2169 if (!S_ISREG(inode->i_mode))
2170 return 0;
2171
2172 if (ci->i_vino.snap != CEPH_NOSNAP) {
2173 /*
2174 * Pool permission check needs to write to the first object.
2175 * But for snapshot, head of the first object may have alread
2176 * been deleted. Skip check to avoid creating orphan object.
2177 */
2178 return 0;
2179 }
2180
2181 if (ceph_test_mount_opt(ceph_inode_to_fs_client(inode),
2182 NOPOOLPERM))
2183 return 0;
2184
2185 spin_lock(&ci->i_ceph_lock);
2186 flags = ci->i_ceph_flags;
2187 pool = ci->i_layout.pool_id;
2188 spin_unlock(&ci->i_ceph_lock);
2189check:
2190 if (flags & CEPH_I_POOL_PERM) {
2191 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2192 doutc(cl, "pool %lld no read perm\n", pool);
2193 return -EPERM;
2194 }
2195 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2196 doutc(cl, "pool %lld no write perm\n", pool);
2197 return -EPERM;
2198 }
2199 return 0;
2200 }
2201
2202 pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2203 ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2204 ceph_put_string(pool_ns);
2205 if (ret < 0)
2206 return ret;
2207
2208 flags = CEPH_I_POOL_PERM;
2209 if (ret & POOL_READ)
2210 flags |= CEPH_I_POOL_RD;
2211 if (ret & POOL_WRITE)
2212 flags |= CEPH_I_POOL_WR;
2213
2214 spin_lock(&ci->i_ceph_lock);
2215 if (pool == ci->i_layout.pool_id &&
2216 pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2217 ci->i_ceph_flags |= flags;
2218 } else {
2219 pool = ci->i_layout.pool_id;
2220 flags = ci->i_ceph_flags;
2221 }
2222 spin_unlock(&ci->i_ceph_lock);
2223 goto check;
2224}
2225
2226void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2227{
2228 struct ceph_pool_perm *perm;
2229 struct rb_node *n;
2230
2231 while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2232 n = rb_first(&mdsc->pool_perm_tree);
2233 perm = rb_entry(n, struct ceph_pool_perm, node);
2234 rb_erase(n, &mdsc->pool_perm_tree);
2235 kfree(perm);
2236 }
2237}