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 <linux/ceph/osd_client.h>
15
16/*
17 * Ceph address space ops.
18 *
19 * There are a few funny things going on here.
20 *
21 * The page->private field is used to reference a struct
22 * ceph_snap_context for _every_ dirty page. This indicates which
23 * snapshot the page was logically dirtied in, and thus which snap
24 * context needs to be associated with the osd write during writeback.
25 *
26 * Similarly, struct ceph_inode_info maintains a set of counters to
27 * count dirty pages on the inode. In the absence of snapshots,
28 * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
29 *
30 * When a snapshot is taken (that is, when the client receives
31 * notification that a snapshot was taken), each inode with caps and
32 * with dirty pages (dirty pages implies there is a cap) gets a new
33 * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
34 * order, new snaps go to the tail). The i_wrbuffer_ref_head count is
35 * moved to capsnap->dirty. (Unless a sync write is currently in
36 * progress. In that case, the capsnap is said to be "pending", new
37 * writes cannot start, and the capsnap isn't "finalized" until the
38 * write completes (or fails) and a final size/mtime for the inode for
39 * that snap can be settled upon.) i_wrbuffer_ref_head is reset to 0.
40 *
41 * On writeback, we must submit writes to the osd IN SNAP ORDER. So,
42 * we look for the first capsnap in i_cap_snaps and write out pages in
43 * that snap context _only_. Then we move on to the next capsnap,
44 * eventually reaching the "live" or "head" context (i.e., pages that
45 * are not yet snapped) and are writing the most recently dirtied
46 * pages.
47 *
48 * Invalidate and so forth must take care to ensure the dirty page
49 * accounting is preserved.
50 */
51
52#define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
53#define CONGESTION_OFF_THRESH(congestion_kb) \
54 (CONGESTION_ON_THRESH(congestion_kb) - \
55 (CONGESTION_ON_THRESH(congestion_kb) >> 2))
56
57
58
59/*
60 * Dirty a page. Optimistically adjust accounting, on the assumption
61 * that we won't race with invalidate. If we do, readjust.
62 */
63static int ceph_set_page_dirty(struct page *page)
64{
65 struct address_space *mapping = page->mapping;
66 struct inode *inode;
67 struct ceph_inode_info *ci;
68 int undo = 0;
69 struct ceph_snap_context *snapc;
70
71 if (unlikely(!mapping))
72 return !TestSetPageDirty(page);
73
74 if (TestSetPageDirty(page)) {
75 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
76 mapping->host, page, page->index);
77 return 0;
78 }
79
80 inode = mapping->host;
81 ci = ceph_inode(inode);
82
83 /*
84 * Note that we're grabbing a snapc ref here without holding
85 * any locks!
86 */
87 snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
88
89 /* dirty the head */
90 spin_lock(&inode->i_lock);
91 if (ci->i_head_snapc == NULL)
92 ci->i_head_snapc = ceph_get_snap_context(snapc);
93 ++ci->i_wrbuffer_ref_head;
94 if (ci->i_wrbuffer_ref == 0)
95 ihold(inode);
96 ++ci->i_wrbuffer_ref;
97 dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
98 "snapc %p seq %lld (%d snaps)\n",
99 mapping->host, page, page->index,
100 ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
101 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
102 snapc, snapc->seq, snapc->num_snaps);
103 spin_unlock(&inode->i_lock);
104
105 /* now adjust page */
106 spin_lock_irq(&mapping->tree_lock);
107 if (page->mapping) { /* Race with truncate? */
108 WARN_ON_ONCE(!PageUptodate(page));
109 account_page_dirtied(page, page->mapping);
110 radix_tree_tag_set(&mapping->page_tree,
111 page_index(page), PAGECACHE_TAG_DIRTY);
112
113 /*
114 * Reference snap context in page->private. Also set
115 * PagePrivate so that we get invalidatepage callback.
116 */
117 page->private = (unsigned long)snapc;
118 SetPagePrivate(page);
119 } else {
120 dout("ANON set_page_dirty %p (raced truncate?)\n", page);
121 undo = 1;
122 }
123
124 spin_unlock_irq(&mapping->tree_lock);
125
126 if (undo)
127 /* whoops, we failed to dirty the page */
128 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
129
130 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
131
132 BUG_ON(!PageDirty(page));
133 return 1;
134}
135
136/*
137 * If we are truncating the full page (i.e. offset == 0), adjust the
138 * dirty page counters appropriately. Only called if there is private
139 * data on the page.
140 */
141static void ceph_invalidatepage(struct page *page, unsigned long offset)
142{
143 struct inode *inode;
144 struct ceph_inode_info *ci;
145 struct ceph_snap_context *snapc = (void *)page->private;
146
147 BUG_ON(!PageLocked(page));
148 BUG_ON(!page->private);
149 BUG_ON(!PagePrivate(page));
150 BUG_ON(!page->mapping);
151
152 inode = page->mapping->host;
153
154 /*
155 * We can get non-dirty pages here due to races between
156 * set_page_dirty and truncate_complete_page; just spit out a
157 * warning, in case we end up with accounting problems later.
158 */
159 if (!PageDirty(page))
160 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
161
162 if (offset == 0)
163 ClearPageChecked(page);
164
165 ci = ceph_inode(inode);
166 if (offset == 0) {
167 dout("%p invalidatepage %p idx %lu full dirty page %lu\n",
168 inode, page, page->index, offset);
169 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
170 ceph_put_snap_context(snapc);
171 page->private = 0;
172 ClearPagePrivate(page);
173 } else {
174 dout("%p invalidatepage %p idx %lu partial dirty page\n",
175 inode, page, page->index);
176 }
177}
178
179/* just a sanity check */
180static int ceph_releasepage(struct page *page, gfp_t g)
181{
182 struct inode *inode = page->mapping ? page->mapping->host : NULL;
183 dout("%p releasepage %p idx %lu\n", inode, page, page->index);
184 WARN_ON(PageDirty(page));
185 WARN_ON(page->private);
186 WARN_ON(PagePrivate(page));
187 return 0;
188}
189
190/*
191 * read a single page, without unlocking it.
192 */
193static int readpage_nounlock(struct file *filp, struct page *page)
194{
195 struct inode *inode = filp->f_dentry->d_inode;
196 struct ceph_inode_info *ci = ceph_inode(inode);
197 struct ceph_osd_client *osdc =
198 &ceph_inode_to_client(inode)->client->osdc;
199 int err = 0;
200 u64 len = PAGE_CACHE_SIZE;
201
202 dout("readpage inode %p file %p page %p index %lu\n",
203 inode, filp, page, page->index);
204 err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
205 page->index << PAGE_CACHE_SHIFT, &len,
206 ci->i_truncate_seq, ci->i_truncate_size,
207 &page, 1, 0);
208 if (err == -ENOENT)
209 err = 0;
210 if (err < 0) {
211 SetPageError(page);
212 goto out;
213 } else if (err < PAGE_CACHE_SIZE) {
214 /* zero fill remainder of page */
215 zero_user_segment(page, err, PAGE_CACHE_SIZE);
216 }
217 SetPageUptodate(page);
218
219out:
220 return err < 0 ? err : 0;
221}
222
223static int ceph_readpage(struct file *filp, struct page *page)
224{
225 int r = readpage_nounlock(filp, page);
226 unlock_page(page);
227 return r;
228}
229
230/*
231 * Build a vector of contiguous pages from the provided page list.
232 */
233static struct page **page_vector_from_list(struct list_head *page_list,
234 unsigned *nr_pages)
235{
236 struct page **pages;
237 struct page *page;
238 int next_index, contig_pages = 0;
239
240 /* build page vector */
241 pages = kmalloc(sizeof(*pages) * *nr_pages, GFP_NOFS);
242 if (!pages)
243 return ERR_PTR(-ENOMEM);
244
245 BUG_ON(list_empty(page_list));
246 next_index = list_entry(page_list->prev, struct page, lru)->index;
247 list_for_each_entry_reverse(page, page_list, lru) {
248 if (page->index == next_index) {
249 dout("readpages page %d %p\n", contig_pages, page);
250 pages[contig_pages] = page;
251 contig_pages++;
252 next_index++;
253 } else {
254 break;
255 }
256 }
257 *nr_pages = contig_pages;
258 return pages;
259}
260
261/*
262 * Read multiple pages. Leave pages we don't read + unlock in page_list;
263 * the caller (VM) cleans them up.
264 */
265static int ceph_readpages(struct file *file, struct address_space *mapping,
266 struct list_head *page_list, unsigned nr_pages)
267{
268 struct inode *inode = file->f_dentry->d_inode;
269 struct ceph_inode_info *ci = ceph_inode(inode);
270 struct ceph_osd_client *osdc =
271 &ceph_inode_to_client(inode)->client->osdc;
272 int rc = 0;
273 struct page **pages;
274 loff_t offset;
275 u64 len;
276
277 dout("readpages %p file %p nr_pages %d\n",
278 inode, file, nr_pages);
279
280 pages = page_vector_from_list(page_list, &nr_pages);
281 if (IS_ERR(pages))
282 return PTR_ERR(pages);
283
284 /* guess read extent */
285 offset = pages[0]->index << PAGE_CACHE_SHIFT;
286 len = nr_pages << PAGE_CACHE_SHIFT;
287 rc = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
288 offset, &len,
289 ci->i_truncate_seq, ci->i_truncate_size,
290 pages, nr_pages, 0);
291 if (rc == -ENOENT)
292 rc = 0;
293 if (rc < 0)
294 goto out;
295
296 for (; !list_empty(page_list) && len > 0;
297 rc -= PAGE_CACHE_SIZE, len -= PAGE_CACHE_SIZE) {
298 struct page *page =
299 list_entry(page_list->prev, struct page, lru);
300
301 list_del(&page->lru);
302
303 if (rc < (int)PAGE_CACHE_SIZE) {
304 /* zero (remainder of) page */
305 int s = rc < 0 ? 0 : rc;
306 zero_user_segment(page, s, PAGE_CACHE_SIZE);
307 }
308
309 if (add_to_page_cache_lru(page, mapping, page->index,
310 GFP_NOFS)) {
311 page_cache_release(page);
312 dout("readpages %p add_to_page_cache failed %p\n",
313 inode, page);
314 continue;
315 }
316 dout("readpages %p adding %p idx %lu\n", inode, page,
317 page->index);
318 flush_dcache_page(page);
319 SetPageUptodate(page);
320 unlock_page(page);
321 page_cache_release(page);
322 }
323 rc = 0;
324
325out:
326 kfree(pages);
327 return rc;
328}
329
330/*
331 * Get ref for the oldest snapc for an inode with dirty data... that is, the
332 * only snap context we are allowed to write back.
333 */
334static struct ceph_snap_context *get_oldest_context(struct inode *inode,
335 u64 *snap_size)
336{
337 struct ceph_inode_info *ci = ceph_inode(inode);
338 struct ceph_snap_context *snapc = NULL;
339 struct ceph_cap_snap *capsnap = NULL;
340
341 spin_lock(&inode->i_lock);
342 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
343 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
344 capsnap->context, capsnap->dirty_pages);
345 if (capsnap->dirty_pages) {
346 snapc = ceph_get_snap_context(capsnap->context);
347 if (snap_size)
348 *snap_size = capsnap->size;
349 break;
350 }
351 }
352 if (!snapc && ci->i_wrbuffer_ref_head) {
353 snapc = ceph_get_snap_context(ci->i_head_snapc);
354 dout(" head snapc %p has %d dirty pages\n",
355 snapc, ci->i_wrbuffer_ref_head);
356 }
357 spin_unlock(&inode->i_lock);
358 return snapc;
359}
360
361/*
362 * Write a single page, but leave the page locked.
363 *
364 * If we get a write error, set the page error bit, but still adjust the
365 * dirty page accounting (i.e., page is no longer dirty).
366 */
367static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
368{
369 struct inode *inode;
370 struct ceph_inode_info *ci;
371 struct ceph_fs_client *fsc;
372 struct ceph_osd_client *osdc;
373 loff_t page_off = page->index << PAGE_CACHE_SHIFT;
374 int len = PAGE_CACHE_SIZE;
375 loff_t i_size;
376 int err = 0;
377 struct ceph_snap_context *snapc, *oldest;
378 u64 snap_size = 0;
379 long writeback_stat;
380
381 dout("writepage %p idx %lu\n", page, page->index);
382
383 if (!page->mapping || !page->mapping->host) {
384 dout("writepage %p - no mapping\n", page);
385 return -EFAULT;
386 }
387 inode = page->mapping->host;
388 ci = ceph_inode(inode);
389 fsc = ceph_inode_to_client(inode);
390 osdc = &fsc->client->osdc;
391
392 /* verify this is a writeable snap context */
393 snapc = (void *)page->private;
394 if (snapc == NULL) {
395 dout("writepage %p page %p not dirty?\n", inode, page);
396 goto out;
397 }
398 oldest = get_oldest_context(inode, &snap_size);
399 if (snapc->seq > oldest->seq) {
400 dout("writepage %p page %p snapc %p not writeable - noop\n",
401 inode, page, (void *)page->private);
402 /* we should only noop if called by kswapd */
403 WARN_ON((current->flags & PF_MEMALLOC) == 0);
404 ceph_put_snap_context(oldest);
405 goto out;
406 }
407 ceph_put_snap_context(oldest);
408
409 /* is this a partial page at end of file? */
410 if (snap_size)
411 i_size = snap_size;
412 else
413 i_size = i_size_read(inode);
414 if (i_size < page_off + len)
415 len = i_size - page_off;
416
417 dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
418 inode, page, page->index, page_off, len, snapc);
419
420 writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
421 if (writeback_stat >
422 CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
423 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
424
425 set_page_writeback(page);
426 err = ceph_osdc_writepages(osdc, ceph_vino(inode),
427 &ci->i_layout, snapc,
428 page_off, len,
429 ci->i_truncate_seq, ci->i_truncate_size,
430 &inode->i_mtime,
431 &page, 1, 0, 0, true);
432 if (err < 0) {
433 dout("writepage setting page/mapping error %d %p\n", err, page);
434 SetPageError(page);
435 mapping_set_error(&inode->i_data, err);
436 if (wbc)
437 wbc->pages_skipped++;
438 } else {
439 dout("writepage cleaned page %p\n", page);
440 err = 0; /* vfs expects us to return 0 */
441 }
442 page->private = 0;
443 ClearPagePrivate(page);
444 end_page_writeback(page);
445 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
446 ceph_put_snap_context(snapc); /* page's reference */
447out:
448 return err;
449}
450
451static int ceph_writepage(struct page *page, struct writeback_control *wbc)
452{
453 int err;
454 struct inode *inode = page->mapping->host;
455 BUG_ON(!inode);
456 ihold(inode);
457 err = writepage_nounlock(page, wbc);
458 unlock_page(page);
459 iput(inode);
460 return err;
461}
462
463
464/*
465 * lame release_pages helper. release_pages() isn't exported to
466 * modules.
467 */
468static void ceph_release_pages(struct page **pages, int num)
469{
470 struct pagevec pvec;
471 int i;
472
473 pagevec_init(&pvec, 0);
474 for (i = 0; i < num; i++) {
475 if (pagevec_add(&pvec, pages[i]) == 0)
476 pagevec_release(&pvec);
477 }
478 pagevec_release(&pvec);
479}
480
481
482/*
483 * async writeback completion handler.
484 *
485 * If we get an error, set the mapping error bit, but not the individual
486 * page error bits.
487 */
488static void writepages_finish(struct ceph_osd_request *req,
489 struct ceph_msg *msg)
490{
491 struct inode *inode = req->r_inode;
492 struct ceph_osd_reply_head *replyhead;
493 struct ceph_osd_op *op;
494 struct ceph_inode_info *ci = ceph_inode(inode);
495 unsigned wrote;
496 struct page *page;
497 int i;
498 struct ceph_snap_context *snapc = req->r_snapc;
499 struct address_space *mapping = inode->i_mapping;
500 __s32 rc = -EIO;
501 u64 bytes = 0;
502 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
503 long writeback_stat;
504 unsigned issued = ceph_caps_issued(ci);
505
506 /* parse reply */
507 replyhead = msg->front.iov_base;
508 WARN_ON(le32_to_cpu(replyhead->num_ops) == 0);
509 op = (void *)(replyhead + 1);
510 rc = le32_to_cpu(replyhead->result);
511 bytes = le64_to_cpu(op->extent.length);
512
513 if (rc >= 0) {
514 /*
515 * Assume we wrote the pages we originally sent. The
516 * osd might reply with fewer pages if our writeback
517 * raced with a truncation and was adjusted at the osd,
518 * so don't believe the reply.
519 */
520 wrote = req->r_num_pages;
521 } else {
522 wrote = 0;
523 mapping_set_error(mapping, rc);
524 }
525 dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
526 inode, rc, bytes, wrote);
527
528 /* clean all pages */
529 for (i = 0; i < req->r_num_pages; i++) {
530 page = req->r_pages[i];
531 BUG_ON(!page);
532 WARN_ON(!PageUptodate(page));
533
534 writeback_stat =
535 atomic_long_dec_return(&fsc->writeback_count);
536 if (writeback_stat <
537 CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
538 clear_bdi_congested(&fsc->backing_dev_info,
539 BLK_RW_ASYNC);
540
541 ceph_put_snap_context((void *)page->private);
542 page->private = 0;
543 ClearPagePrivate(page);
544 dout("unlocking %d %p\n", i, page);
545 end_page_writeback(page);
546
547 /*
548 * We lost the cache cap, need to truncate the page before
549 * it is unlocked, otherwise we'd truncate it later in the
550 * page truncation thread, possibly losing some data that
551 * raced its way in
552 */
553 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
554 generic_error_remove_page(inode->i_mapping, page);
555
556 unlock_page(page);
557 }
558 dout("%p wrote+cleaned %d pages\n", inode, wrote);
559 ceph_put_wrbuffer_cap_refs(ci, req->r_num_pages, snapc);
560
561 ceph_release_pages(req->r_pages, req->r_num_pages);
562 if (req->r_pages_from_pool)
563 mempool_free(req->r_pages,
564 ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
565 else
566 kfree(req->r_pages);
567 ceph_osdc_put_request(req);
568}
569
570/*
571 * allocate a page vec, either directly, or if necessary, via a the
572 * mempool. we avoid the mempool if we can because req->r_num_pages
573 * may be less than the maximum write size.
574 */
575static void alloc_page_vec(struct ceph_fs_client *fsc,
576 struct ceph_osd_request *req)
577{
578 req->r_pages = kmalloc(sizeof(struct page *) * req->r_num_pages,
579 GFP_NOFS);
580 if (!req->r_pages) {
581 req->r_pages = mempool_alloc(fsc->wb_pagevec_pool, GFP_NOFS);
582 req->r_pages_from_pool = 1;
583 WARN_ON(!req->r_pages);
584 }
585}
586
587/*
588 * initiate async writeback
589 */
590static int ceph_writepages_start(struct address_space *mapping,
591 struct writeback_control *wbc)
592{
593 struct inode *inode = mapping->host;
594 struct ceph_inode_info *ci = ceph_inode(inode);
595 struct ceph_fs_client *fsc;
596 pgoff_t index, start, end;
597 int range_whole = 0;
598 int should_loop = 1;
599 pgoff_t max_pages = 0, max_pages_ever = 0;
600 struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
601 struct pagevec pvec;
602 int done = 0;
603 int rc = 0;
604 unsigned wsize = 1 << inode->i_blkbits;
605 struct ceph_osd_request *req = NULL;
606 int do_sync;
607 u64 snap_size = 0;
608
609 /*
610 * Include a 'sync' in the OSD request if this is a data
611 * integrity write (e.g., O_SYNC write or fsync()), or if our
612 * cap is being revoked.
613 */
614 do_sync = wbc->sync_mode == WB_SYNC_ALL;
615 if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
616 do_sync = 1;
617 dout("writepages_start %p dosync=%d (mode=%s)\n",
618 inode, do_sync,
619 wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
620 (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
621
622 fsc = ceph_inode_to_client(inode);
623 if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
624 pr_warning("writepage_start %p on forced umount\n", inode);
625 return -EIO; /* we're in a forced umount, don't write! */
626 }
627 if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
628 wsize = fsc->mount_options->wsize;
629 if (wsize < PAGE_CACHE_SIZE)
630 wsize = PAGE_CACHE_SIZE;
631 max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
632
633 pagevec_init(&pvec, 0);
634
635 /* where to start/end? */
636 if (wbc->range_cyclic) {
637 start = mapping->writeback_index; /* Start from prev offset */
638 end = -1;
639 dout(" cyclic, start at %lu\n", start);
640 } else {
641 start = wbc->range_start >> PAGE_CACHE_SHIFT;
642 end = wbc->range_end >> PAGE_CACHE_SHIFT;
643 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
644 range_whole = 1;
645 should_loop = 0;
646 dout(" not cyclic, %lu to %lu\n", start, end);
647 }
648 index = start;
649
650retry:
651 /* find oldest snap context with dirty data */
652 ceph_put_snap_context(snapc);
653 snapc = get_oldest_context(inode, &snap_size);
654 if (!snapc) {
655 /* hmm, why does writepages get called when there
656 is no dirty data? */
657 dout(" no snap context with dirty data?\n");
658 goto out;
659 }
660 dout(" oldest snapc is %p seq %lld (%d snaps)\n",
661 snapc, snapc->seq, snapc->num_snaps);
662 if (last_snapc && snapc != last_snapc) {
663 /* if we switched to a newer snapc, restart our scan at the
664 * start of the original file range. */
665 dout(" snapc differs from last pass, restarting at %lu\n",
666 index);
667 index = start;
668 }
669 last_snapc = snapc;
670
671 while (!done && index <= end) {
672 unsigned i;
673 int first;
674 pgoff_t next;
675 int pvec_pages, locked_pages;
676 struct page *page;
677 int want;
678 u64 offset, len;
679 struct ceph_osd_request_head *reqhead;
680 struct ceph_osd_op *op;
681 long writeback_stat;
682
683 next = 0;
684 locked_pages = 0;
685 max_pages = max_pages_ever;
686
687get_more_pages:
688 first = -1;
689 want = min(end - index,
690 min((pgoff_t)PAGEVEC_SIZE,
691 max_pages - (pgoff_t)locked_pages) - 1)
692 + 1;
693 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
694 PAGECACHE_TAG_DIRTY,
695 want);
696 dout("pagevec_lookup_tag got %d\n", pvec_pages);
697 if (!pvec_pages && !locked_pages)
698 break;
699 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
700 page = pvec.pages[i];
701 dout("? %p idx %lu\n", page, page->index);
702 if (locked_pages == 0)
703 lock_page(page); /* first page */
704 else if (!trylock_page(page))
705 break;
706
707 /* only dirty pages, or our accounting breaks */
708 if (unlikely(!PageDirty(page)) ||
709 unlikely(page->mapping != mapping)) {
710 dout("!dirty or !mapping %p\n", page);
711 unlock_page(page);
712 break;
713 }
714 if (!wbc->range_cyclic && page->index > end) {
715 dout("end of range %p\n", page);
716 done = 1;
717 unlock_page(page);
718 break;
719 }
720 if (next && (page->index != next)) {
721 dout("not consecutive %p\n", page);
722 unlock_page(page);
723 break;
724 }
725 if (wbc->sync_mode != WB_SYNC_NONE) {
726 dout("waiting on writeback %p\n", page);
727 wait_on_page_writeback(page);
728 }
729 if ((snap_size && page_offset(page) > snap_size) ||
730 (!snap_size &&
731 page_offset(page) > i_size_read(inode))) {
732 dout("%p page eof %llu\n", page, snap_size ?
733 snap_size : i_size_read(inode));
734 done = 1;
735 unlock_page(page);
736 break;
737 }
738 if (PageWriteback(page)) {
739 dout("%p under writeback\n", page);
740 unlock_page(page);
741 break;
742 }
743
744 /* only if matching snap context */
745 pgsnapc = (void *)page->private;
746 if (pgsnapc->seq > snapc->seq) {
747 dout("page snapc %p %lld > oldest %p %lld\n",
748 pgsnapc, pgsnapc->seq, snapc, snapc->seq);
749 unlock_page(page);
750 if (!locked_pages)
751 continue; /* keep looking for snap */
752 break;
753 }
754
755 if (!clear_page_dirty_for_io(page)) {
756 dout("%p !clear_page_dirty_for_io\n", page);
757 unlock_page(page);
758 break;
759 }
760
761 /* ok */
762 if (locked_pages == 0) {
763 /* prepare async write request */
764 offset = (unsigned long long)page->index
765 << PAGE_CACHE_SHIFT;
766 len = wsize;
767 req = ceph_osdc_new_request(&fsc->client->osdc,
768 &ci->i_layout,
769 ceph_vino(inode),
770 offset, &len,
771 CEPH_OSD_OP_WRITE,
772 CEPH_OSD_FLAG_WRITE |
773 CEPH_OSD_FLAG_ONDISK,
774 snapc, do_sync,
775 ci->i_truncate_seq,
776 ci->i_truncate_size,
777 &inode->i_mtime, true, 1, 0);
778
779 if (!req) {
780 rc = -ENOMEM;
781 unlock_page(page);
782 break;
783 }
784
785 max_pages = req->r_num_pages;
786
787 alloc_page_vec(fsc, req);
788 req->r_callback = writepages_finish;
789 req->r_inode = inode;
790 }
791
792 /* note position of first page in pvec */
793 if (first < 0)
794 first = i;
795 dout("%p will write page %p idx %lu\n",
796 inode, page, page->index);
797
798 writeback_stat =
799 atomic_long_inc_return(&fsc->writeback_count);
800 if (writeback_stat > CONGESTION_ON_THRESH(
801 fsc->mount_options->congestion_kb)) {
802 set_bdi_congested(&fsc->backing_dev_info,
803 BLK_RW_ASYNC);
804 }
805
806 set_page_writeback(page);
807 req->r_pages[locked_pages] = page;
808 locked_pages++;
809 next = page->index + 1;
810 }
811
812 /* did we get anything? */
813 if (!locked_pages)
814 goto release_pvec_pages;
815 if (i) {
816 int j;
817 BUG_ON(!locked_pages || first < 0);
818
819 if (pvec_pages && i == pvec_pages &&
820 locked_pages < max_pages) {
821 dout("reached end pvec, trying for more\n");
822 pagevec_reinit(&pvec);
823 goto get_more_pages;
824 }
825
826 /* shift unused pages over in the pvec... we
827 * will need to release them below. */
828 for (j = i; j < pvec_pages; j++) {
829 dout(" pvec leftover page %p\n",
830 pvec.pages[j]);
831 pvec.pages[j-i+first] = pvec.pages[j];
832 }
833 pvec.nr -= i-first;
834 }
835
836 /* submit the write */
837 offset = req->r_pages[0]->index << PAGE_CACHE_SHIFT;
838 len = min((snap_size ? snap_size : i_size_read(inode)) - offset,
839 (u64)locked_pages << PAGE_CACHE_SHIFT);
840 dout("writepages got %d pages at %llu~%llu\n",
841 locked_pages, offset, len);
842
843 /* revise final length, page count */
844 req->r_num_pages = locked_pages;
845 reqhead = req->r_request->front.iov_base;
846 op = (void *)(reqhead + 1);
847 op->extent.length = cpu_to_le64(len);
848 op->payload_len = cpu_to_le32(len);
849 req->r_request->hdr.data_len = cpu_to_le32(len);
850
851 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
852 BUG_ON(rc);
853 req = NULL;
854
855 /* continue? */
856 index = next;
857 wbc->nr_to_write -= locked_pages;
858 if (wbc->nr_to_write <= 0)
859 done = 1;
860
861release_pvec_pages:
862 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
863 pvec.nr ? pvec.pages[0] : NULL);
864 pagevec_release(&pvec);
865
866 if (locked_pages && !done)
867 goto retry;
868 }
869
870 if (should_loop && !done) {
871 /* more to do; loop back to beginning of file */
872 dout("writepages looping back to beginning of file\n");
873 should_loop = 0;
874 index = 0;
875 goto retry;
876 }
877
878 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
879 mapping->writeback_index = index;
880
881out:
882 if (req)
883 ceph_osdc_put_request(req);
884 ceph_put_snap_context(snapc);
885 dout("writepages done, rc = %d\n", rc);
886 return rc;
887}
888
889
890
891/*
892 * See if a given @snapc is either writeable, or already written.
893 */
894static int context_is_writeable_or_written(struct inode *inode,
895 struct ceph_snap_context *snapc)
896{
897 struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
898 int ret = !oldest || snapc->seq <= oldest->seq;
899
900 ceph_put_snap_context(oldest);
901 return ret;
902}
903
904/*
905 * We are only allowed to write into/dirty the page if the page is
906 * clean, or already dirty within the same snap context.
907 *
908 * called with page locked.
909 * return success with page locked,
910 * or any failure (incl -EAGAIN) with page unlocked.
911 */
912static int ceph_update_writeable_page(struct file *file,
913 loff_t pos, unsigned len,
914 struct page *page)
915{
916 struct inode *inode = file->f_dentry->d_inode;
917 struct ceph_inode_info *ci = ceph_inode(inode);
918 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
919 loff_t page_off = pos & PAGE_CACHE_MASK;
920 int pos_in_page = pos & ~PAGE_CACHE_MASK;
921 int end_in_page = pos_in_page + len;
922 loff_t i_size;
923 int r;
924 struct ceph_snap_context *snapc, *oldest;
925
926retry_locked:
927 /* writepages currently holds page lock, but if we change that later, */
928 wait_on_page_writeback(page);
929
930 /* check snap context */
931 BUG_ON(!ci->i_snap_realm);
932 down_read(&mdsc->snap_rwsem);
933 BUG_ON(!ci->i_snap_realm->cached_context);
934 snapc = (void *)page->private;
935 if (snapc && snapc != ci->i_head_snapc) {
936 /*
937 * this page is already dirty in another (older) snap
938 * context! is it writeable now?
939 */
940 oldest = get_oldest_context(inode, NULL);
941 up_read(&mdsc->snap_rwsem);
942
943 if (snapc->seq > oldest->seq) {
944 ceph_put_snap_context(oldest);
945 dout(" page %p snapc %p not current or oldest\n",
946 page, snapc);
947 /*
948 * queue for writeback, and wait for snapc to
949 * be writeable or written
950 */
951 snapc = ceph_get_snap_context(snapc);
952 unlock_page(page);
953 ceph_queue_writeback(inode);
954 r = wait_event_interruptible(ci->i_cap_wq,
955 context_is_writeable_or_written(inode, snapc));
956 ceph_put_snap_context(snapc);
957 if (r == -ERESTARTSYS)
958 return r;
959 return -EAGAIN;
960 }
961 ceph_put_snap_context(oldest);
962
963 /* yay, writeable, do it now (without dropping page lock) */
964 dout(" page %p snapc %p not current, but oldest\n",
965 page, snapc);
966 if (!clear_page_dirty_for_io(page))
967 goto retry_locked;
968 r = writepage_nounlock(page, NULL);
969 if (r < 0)
970 goto fail_nosnap;
971 goto retry_locked;
972 }
973
974 if (PageUptodate(page)) {
975 dout(" page %p already uptodate\n", page);
976 return 0;
977 }
978
979 /* full page? */
980 if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
981 return 0;
982
983 /* past end of file? */
984 i_size = inode->i_size; /* caller holds i_mutex */
985
986 if (i_size + len > inode->i_sb->s_maxbytes) {
987 /* file is too big */
988 r = -EINVAL;
989 goto fail;
990 }
991
992 if (page_off >= i_size ||
993 (pos_in_page == 0 && (pos+len) >= i_size &&
994 end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
995 dout(" zeroing %p 0 - %d and %d - %d\n",
996 page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
997 zero_user_segments(page,
998 0, pos_in_page,
999 end_in_page, PAGE_CACHE_SIZE);
1000 return 0;
1001 }
1002
1003 /* we need to read it. */
1004 up_read(&mdsc->snap_rwsem);
1005 r = readpage_nounlock(file, page);
1006 if (r < 0)
1007 goto fail_nosnap;
1008 goto retry_locked;
1009
1010fail:
1011 up_read(&mdsc->snap_rwsem);
1012fail_nosnap:
1013 unlock_page(page);
1014 return r;
1015}
1016
1017/*
1018 * We are only allowed to write into/dirty the page if the page is
1019 * clean, or already dirty within the same snap context.
1020 */
1021static int ceph_write_begin(struct file *file, struct address_space *mapping,
1022 loff_t pos, unsigned len, unsigned flags,
1023 struct page **pagep, void **fsdata)
1024{
1025 struct inode *inode = file->f_dentry->d_inode;
1026 struct page *page;
1027 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1028 int r;
1029
1030 do {
1031 /* get a page */
1032 page = grab_cache_page_write_begin(mapping, index, 0);
1033 if (!page)
1034 return -ENOMEM;
1035 *pagep = page;
1036
1037 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1038 inode, page, (int)pos, (int)len);
1039
1040 r = ceph_update_writeable_page(file, pos, len, page);
1041 } while (r == -EAGAIN);
1042
1043 return r;
1044}
1045
1046/*
1047 * we don't do anything in here that simple_write_end doesn't do
1048 * except adjust dirty page accounting and drop read lock on
1049 * mdsc->snap_rwsem.
1050 */
1051static int ceph_write_end(struct file *file, struct address_space *mapping,
1052 loff_t pos, unsigned len, unsigned copied,
1053 struct page *page, void *fsdata)
1054{
1055 struct inode *inode = file->f_dentry->d_inode;
1056 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1057 struct ceph_mds_client *mdsc = fsc->mdsc;
1058 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1059 int check_cap = 0;
1060
1061 dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1062 inode, page, (int)pos, (int)copied, (int)len);
1063
1064 /* zero the stale part of the page if we did a short copy */
1065 if (copied < len)
1066 zero_user_segment(page, from+copied, len);
1067
1068 /* did file size increase? */
1069 /* (no need for i_size_read(); we caller holds i_mutex */
1070 if (pos+copied > inode->i_size)
1071 check_cap = ceph_inode_set_size(inode, pos+copied);
1072
1073 if (!PageUptodate(page))
1074 SetPageUptodate(page);
1075
1076 set_page_dirty(page);
1077
1078 unlock_page(page);
1079 up_read(&mdsc->snap_rwsem);
1080 page_cache_release(page);
1081
1082 if (check_cap)
1083 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1084
1085 return copied;
1086}
1087
1088/*
1089 * we set .direct_IO to indicate direct io is supported, but since we
1090 * intercept O_DIRECT reads and writes early, this function should
1091 * never get called.
1092 */
1093static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1094 const struct iovec *iov,
1095 loff_t pos, unsigned long nr_segs)
1096{
1097 WARN_ON(1);
1098 return -EINVAL;
1099}
1100
1101const struct address_space_operations ceph_aops = {
1102 .readpage = ceph_readpage,
1103 .readpages = ceph_readpages,
1104 .writepage = ceph_writepage,
1105 .writepages = ceph_writepages_start,
1106 .write_begin = ceph_write_begin,
1107 .write_end = ceph_write_end,
1108 .set_page_dirty = ceph_set_page_dirty,
1109 .invalidatepage = ceph_invalidatepage,
1110 .releasepage = ceph_releasepage,
1111 .direct_IO = ceph_direct_io,
1112};
1113
1114
1115/*
1116 * vm ops
1117 */
1118
1119/*
1120 * Reuse write_begin here for simplicity.
1121 */
1122static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1123{
1124 struct inode *inode = vma->vm_file->f_dentry->d_inode;
1125 struct page *page = vmf->page;
1126 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1127 loff_t off = page->index << PAGE_CACHE_SHIFT;
1128 loff_t size, len;
1129 int ret;
1130
1131 size = i_size_read(inode);
1132 if (off + PAGE_CACHE_SIZE <= size)
1133 len = PAGE_CACHE_SIZE;
1134 else
1135 len = size & ~PAGE_CACHE_MASK;
1136
1137 dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1138 off, len, page, page->index);
1139
1140 lock_page(page);
1141
1142 ret = VM_FAULT_NOPAGE;
1143 if ((off > size) ||
1144 (page->mapping != inode->i_mapping))
1145 goto out;
1146
1147 ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1148 if (ret == 0) {
1149 /* success. we'll keep the page locked. */
1150 set_page_dirty(page);
1151 up_read(&mdsc->snap_rwsem);
1152 ret = VM_FAULT_LOCKED;
1153 } else {
1154 if (ret == -ENOMEM)
1155 ret = VM_FAULT_OOM;
1156 else
1157 ret = VM_FAULT_SIGBUS;
1158 }
1159out:
1160 dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1161 if (ret != VM_FAULT_LOCKED)
1162 unlock_page(page);
1163 return ret;
1164}
1165
1166static struct vm_operations_struct ceph_vmops = {
1167 .fault = filemap_fault,
1168 .page_mkwrite = ceph_page_mkwrite,
1169};
1170
1171int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1172{
1173 struct address_space *mapping = file->f_mapping;
1174
1175 if (!mapping->a_ops->readpage)
1176 return -ENOEXEC;
1177 file_accessed(file);
1178 vma->vm_ops = &ceph_vmops;
1179 vma->vm_flags |= VM_CAN_NONLINEAR;
1180 return 0;
1181}
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#include <trace/events/netfs.h>
17
18#include "super.h"
19#include "mds_client.h"
20#include "cache.h"
21#include "metric.h"
22#include "crypto.h"
23#include <linux/ceph/osd_client.h>
24#include <linux/ceph/striper.h>
25
26/*
27 * Ceph address space ops.
28 *
29 * There are a few funny things going on here.
30 *
31 * The page->private field is used to reference a struct
32 * ceph_snap_context for _every_ dirty page. This indicates which
33 * snapshot the page was logically dirtied in, and thus which snap
34 * context needs to be associated with the osd write during writeback.
35 *
36 * Similarly, struct ceph_inode_info maintains a set of counters to
37 * count dirty pages on the inode. In the absence of snapshots,
38 * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
39 *
40 * When a snapshot is taken (that is, when the client receives
41 * notification that a snapshot was taken), each inode with caps and
42 * with dirty pages (dirty pages implies there is a cap) gets a new
43 * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
44 * order, new snaps go to the tail). The i_wrbuffer_ref_head count is
45 * moved to capsnap->dirty. (Unless a sync write is currently in
46 * progress. In that case, the capsnap is said to be "pending", new
47 * writes cannot start, and the capsnap isn't "finalized" until the
48 * write completes (or fails) and a final size/mtime for the inode for
49 * that snap can be settled upon.) i_wrbuffer_ref_head is reset to 0.
50 *
51 * On writeback, we must submit writes to the osd IN SNAP ORDER. So,
52 * we look for the first capsnap in i_cap_snaps and write out pages in
53 * that snap context _only_. Then we move on to the next capsnap,
54 * eventually reaching the "live" or "head" context (i.e., pages that
55 * are not yet snapped) and are writing the most recently dirtied
56 * pages.
57 *
58 * Invalidate and so forth must take care to ensure the dirty page
59 * accounting is preserved.
60 */
61
62#define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
63#define CONGESTION_OFF_THRESH(congestion_kb) \
64 (CONGESTION_ON_THRESH(congestion_kb) - \
65 (CONGESTION_ON_THRESH(congestion_kb) >> 2))
66
67static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
68 struct folio **foliop, void **_fsdata);
69
70static inline struct ceph_snap_context *page_snap_context(struct page *page)
71{
72 if (PagePrivate(page))
73 return (void *)page->private;
74 return NULL;
75}
76
77/*
78 * Dirty a page. Optimistically adjust accounting, on the assumption
79 * that we won't race with invalidate. If we do, readjust.
80 */
81static bool ceph_dirty_folio(struct address_space *mapping, struct folio *folio)
82{
83 struct inode *inode = mapping->host;
84 struct ceph_client *cl = ceph_inode_to_client(inode);
85 struct ceph_inode_info *ci;
86 struct ceph_snap_context *snapc;
87
88 if (folio_test_dirty(folio)) {
89 doutc(cl, "%llx.%llx %p idx %lu -- already dirty\n",
90 ceph_vinop(inode), folio, folio->index);
91 VM_BUG_ON_FOLIO(!folio_test_private(folio), folio);
92 return false;
93 }
94
95 ci = ceph_inode(inode);
96
97 /* dirty the head */
98 spin_lock(&ci->i_ceph_lock);
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 = umin(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 void finish_netfs_read(struct ceph_osd_request *req)
209{
210 struct inode *inode = req->r_inode;
211 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
212 struct ceph_client *cl = fsc->client;
213 struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
214 struct netfs_io_subrequest *subreq = req->r_priv;
215 struct ceph_osd_req_op *op = &req->r_ops[0];
216 int err = req->r_result;
217 bool sparse = (op->op == CEPH_OSD_OP_SPARSE_READ);
218
219 ceph_update_read_metrics(&fsc->mdsc->metric, req->r_start_latency,
220 req->r_end_latency, osd_data->length, err);
221
222 doutc(cl, "result %d subreq->len=%zu i_size=%lld\n", req->r_result,
223 subreq->len, i_size_read(req->r_inode));
224
225 /* no object means success but no data */
226 if (err == -ENOENT)
227 err = 0;
228 else if (err == -EBLOCKLISTED)
229 fsc->blocklisted = true;
230
231 if (err >= 0) {
232 if (sparse && err > 0)
233 err = ceph_sparse_ext_map_end(op);
234 if (err < subreq->len &&
235 subreq->rreq->origin != NETFS_DIO_READ)
236 __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
237 if (IS_ENCRYPTED(inode) && err > 0) {
238 err = ceph_fscrypt_decrypt_extents(inode,
239 osd_data->pages, subreq->start,
240 op->extent.sparse_ext,
241 op->extent.sparse_ext_cnt);
242 if (err > subreq->len)
243 err = subreq->len;
244 }
245 }
246
247 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
248 ceph_put_page_vector(osd_data->pages,
249 calc_pages_for(osd_data->alignment,
250 osd_data->length), false);
251 }
252 if (err > 0) {
253 subreq->transferred = err;
254 err = 0;
255 }
256 trace_netfs_sreq(subreq, netfs_sreq_trace_io_progress);
257 netfs_read_subreq_terminated(subreq, err, false);
258 iput(req->r_inode);
259 ceph_dec_osd_stopping_blocker(fsc->mdsc);
260}
261
262static bool ceph_netfs_issue_op_inline(struct netfs_io_subrequest *subreq)
263{
264 struct netfs_io_request *rreq = subreq->rreq;
265 struct inode *inode = rreq->inode;
266 struct ceph_mds_reply_info_parsed *rinfo;
267 struct ceph_mds_reply_info_in *iinfo;
268 struct ceph_mds_request *req;
269 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
270 struct ceph_inode_info *ci = ceph_inode(inode);
271 ssize_t err = 0;
272 size_t len;
273 int mode;
274
275 if (rreq->origin != NETFS_DIO_READ)
276 __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
277 __clear_bit(NETFS_SREQ_COPY_TO_CACHE, &subreq->flags);
278
279 if (subreq->start >= inode->i_size)
280 goto out;
281
282 /* We need to fetch the inline data. */
283 mode = ceph_try_to_choose_auth_mds(inode, CEPH_STAT_CAP_INLINE_DATA);
284 req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_GETATTR, mode);
285 if (IS_ERR(req)) {
286 err = PTR_ERR(req);
287 goto out;
288 }
289 req->r_ino1 = ci->i_vino;
290 req->r_args.getattr.mask = cpu_to_le32(CEPH_STAT_CAP_INLINE_DATA);
291 req->r_num_caps = 2;
292
293 trace_netfs_sreq(subreq, netfs_sreq_trace_submit);
294 err = ceph_mdsc_do_request(mdsc, NULL, req);
295 if (err < 0)
296 goto out;
297
298 rinfo = &req->r_reply_info;
299 iinfo = &rinfo->targeti;
300 if (iinfo->inline_version == CEPH_INLINE_NONE) {
301 /* The data got uninlined */
302 ceph_mdsc_put_request(req);
303 return false;
304 }
305
306 len = min_t(size_t, iinfo->inline_len - subreq->start, subreq->len);
307 err = copy_to_iter(iinfo->inline_data + subreq->start, len, &subreq->io_iter);
308 if (err == 0) {
309 err = -EFAULT;
310 } else {
311 subreq->transferred += err;
312 err = 0;
313 }
314
315 ceph_mdsc_put_request(req);
316out:
317 netfs_read_subreq_terminated(subreq, err, false);
318 return true;
319}
320
321static int ceph_netfs_prepare_read(struct netfs_io_subrequest *subreq)
322{
323 struct netfs_io_request *rreq = subreq->rreq;
324 struct inode *inode = rreq->inode;
325 struct ceph_inode_info *ci = ceph_inode(inode);
326 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
327 u64 objno, objoff;
328 u32 xlen;
329
330 /* Truncate the extent at the end of the current block */
331 ceph_calc_file_object_mapping(&ci->i_layout, subreq->start, subreq->len,
332 &objno, &objoff, &xlen);
333 rreq->io_streams[0].sreq_max_len = umin(xlen, fsc->mount_options->rsize);
334 return 0;
335}
336
337static void ceph_netfs_issue_read(struct netfs_io_subrequest *subreq)
338{
339 struct netfs_io_request *rreq = subreq->rreq;
340 struct inode *inode = rreq->inode;
341 struct ceph_inode_info *ci = ceph_inode(inode);
342 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
343 struct ceph_client *cl = fsc->client;
344 struct ceph_osd_request *req = NULL;
345 struct ceph_vino vino = ceph_vino(inode);
346 int err;
347 u64 len;
348 bool sparse = IS_ENCRYPTED(inode) || ceph_test_mount_opt(fsc, SPARSEREAD);
349 u64 off = subreq->start;
350 int extent_cnt;
351
352 if (ceph_inode_is_shutdown(inode)) {
353 err = -EIO;
354 goto out;
355 }
356
357 if (ceph_has_inline_data(ci) && ceph_netfs_issue_op_inline(subreq))
358 return;
359
360 // TODO: This rounding here is slightly dodgy. It *should* work, for
361 // now, as the cache only deals in blocks that are a multiple of
362 // PAGE_SIZE and fscrypt blocks are at most PAGE_SIZE. What needs to
363 // happen is for the fscrypt driving to be moved into netfslib and the
364 // data in the cache also to be stored encrypted.
365 len = subreq->len;
366 ceph_fscrypt_adjust_off_and_len(inode, &off, &len);
367
368 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, vino,
369 off, &len, 0, 1, sparse ? CEPH_OSD_OP_SPARSE_READ : CEPH_OSD_OP_READ,
370 CEPH_OSD_FLAG_READ, NULL, ci->i_truncate_seq,
371 ci->i_truncate_size, false);
372 if (IS_ERR(req)) {
373 err = PTR_ERR(req);
374 req = NULL;
375 goto out;
376 }
377
378 if (sparse) {
379 extent_cnt = __ceph_sparse_read_ext_count(inode, len);
380 err = ceph_alloc_sparse_ext_map(&req->r_ops[0], extent_cnt);
381 if (err)
382 goto out;
383 }
384
385 doutc(cl, "%llx.%llx pos=%llu orig_len=%zu len=%llu\n",
386 ceph_vinop(inode), subreq->start, subreq->len, len);
387
388 /*
389 * FIXME: For now, use CEPH_OSD_DATA_TYPE_PAGES instead of _ITER for
390 * encrypted inodes. We'd need infrastructure that handles an iov_iter
391 * instead of page arrays, and we don't have that as of yet. Once the
392 * dust settles on the write helpers and encrypt/decrypt routines for
393 * netfs, we should be able to rework this.
394 */
395 if (IS_ENCRYPTED(inode)) {
396 struct page **pages;
397 size_t page_off;
398
399 err = iov_iter_get_pages_alloc2(&subreq->io_iter, &pages, len, &page_off);
400 if (err < 0) {
401 doutc(cl, "%llx.%llx failed to allocate pages, %d\n",
402 ceph_vinop(inode), err);
403 goto out;
404 }
405
406 /* should always give us a page-aligned read */
407 WARN_ON_ONCE(page_off);
408 len = err;
409 err = 0;
410
411 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false,
412 false);
413 } else {
414 osd_req_op_extent_osd_iter(req, 0, &subreq->io_iter);
415 }
416 if (!ceph_inc_osd_stopping_blocker(fsc->mdsc)) {
417 err = -EIO;
418 goto out;
419 }
420 req->r_callback = finish_netfs_read;
421 req->r_priv = subreq;
422 req->r_inode = inode;
423 ihold(inode);
424
425 trace_netfs_sreq(subreq, netfs_sreq_trace_submit);
426 ceph_osdc_start_request(req->r_osdc, req);
427out:
428 ceph_osdc_put_request(req);
429 if (err)
430 netfs_read_subreq_terminated(subreq, err, false);
431 doutc(cl, "%llx.%llx result %d\n", ceph_vinop(inode), err);
432}
433
434static int ceph_init_request(struct netfs_io_request *rreq, struct file *file)
435{
436 struct inode *inode = rreq->inode;
437 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
438 struct ceph_client *cl = ceph_inode_to_client(inode);
439 int got = 0, want = CEPH_CAP_FILE_CACHE;
440 struct ceph_netfs_request_data *priv;
441 int ret = 0;
442
443 /* [DEPRECATED] Use PG_private_2 to mark folio being written to the cache. */
444 __set_bit(NETFS_RREQ_USE_PGPRIV2, &rreq->flags);
445
446 if (rreq->origin != NETFS_READAHEAD)
447 return 0;
448
449 priv = kzalloc(sizeof(*priv), GFP_NOFS);
450 if (!priv)
451 return -ENOMEM;
452
453 if (file) {
454 struct ceph_rw_context *rw_ctx;
455 struct ceph_file_info *fi = file->private_data;
456
457 priv->file_ra_pages = file->f_ra.ra_pages;
458 priv->file_ra_disabled = file->f_mode & FMODE_RANDOM;
459
460 rw_ctx = ceph_find_rw_context(fi);
461 if (rw_ctx) {
462 rreq->netfs_priv = priv;
463 return 0;
464 }
465 }
466
467 /*
468 * readahead callers do not necessarily hold Fcb caps
469 * (e.g. fadvise, madvise).
470 */
471 ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want, true, &got);
472 if (ret < 0) {
473 doutc(cl, "%llx.%llx, error getting cap\n", ceph_vinop(inode));
474 goto out;
475 }
476
477 if (!(got & want)) {
478 doutc(cl, "%llx.%llx, no cache cap\n", ceph_vinop(inode));
479 ret = -EACCES;
480 goto out;
481 }
482 if (ret == 0) {
483 ret = -EACCES;
484 goto out;
485 }
486
487 priv->caps = got;
488 rreq->netfs_priv = priv;
489 rreq->io_streams[0].sreq_max_len = fsc->mount_options->rsize;
490
491out:
492 if (ret < 0) {
493 if (got)
494 ceph_put_cap_refs(ceph_inode(inode), got);
495 kfree(priv);
496 }
497
498 return ret;
499}
500
501static void ceph_netfs_free_request(struct netfs_io_request *rreq)
502{
503 struct ceph_netfs_request_data *priv = rreq->netfs_priv;
504
505 if (!priv)
506 return;
507
508 if (priv->caps)
509 ceph_put_cap_refs(ceph_inode(rreq->inode), priv->caps);
510 kfree(priv);
511 rreq->netfs_priv = NULL;
512}
513
514const struct netfs_request_ops ceph_netfs_ops = {
515 .init_request = ceph_init_request,
516 .free_request = ceph_netfs_free_request,
517 .prepare_read = ceph_netfs_prepare_read,
518 .issue_read = ceph_netfs_issue_read,
519 .expand_readahead = ceph_netfs_expand_readahead,
520 .check_write_begin = ceph_netfs_check_write_begin,
521};
522
523#ifdef CONFIG_CEPH_FSCACHE
524static void ceph_set_page_fscache(struct page *page)
525{
526 folio_start_private_2(page_folio(page)); /* [DEPRECATED] */
527}
528
529static void ceph_fscache_write_terminated(void *priv, ssize_t error, bool was_async)
530{
531 struct inode *inode = priv;
532
533 if (IS_ERR_VALUE(error) && error != -ENOBUFS)
534 ceph_fscache_invalidate(inode, false);
535}
536
537static void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
538{
539 struct ceph_inode_info *ci = ceph_inode(inode);
540 struct fscache_cookie *cookie = ceph_fscache_cookie(ci);
541
542 fscache_write_to_cache(cookie, inode->i_mapping, off, len, i_size_read(inode),
543 ceph_fscache_write_terminated, inode, true, caching);
544}
545#else
546static inline void ceph_set_page_fscache(struct page *page)
547{
548}
549
550static inline void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
551{
552}
553#endif /* CONFIG_CEPH_FSCACHE */
554
555struct ceph_writeback_ctl
556{
557 loff_t i_size;
558 u64 truncate_size;
559 u32 truncate_seq;
560 bool size_stable;
561 bool head_snapc;
562};
563
564/*
565 * Get ref for the oldest snapc for an inode with dirty data... that is, the
566 * only snap context we are allowed to write back.
567 */
568static struct ceph_snap_context *
569get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
570 struct ceph_snap_context *page_snapc)
571{
572 struct ceph_inode_info *ci = ceph_inode(inode);
573 struct ceph_client *cl = ceph_inode_to_client(inode);
574 struct ceph_snap_context *snapc = NULL;
575 struct ceph_cap_snap *capsnap = NULL;
576
577 spin_lock(&ci->i_ceph_lock);
578 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
579 doutc(cl, " capsnap %p snapc %p has %d dirty pages\n",
580 capsnap, capsnap->context, capsnap->dirty_pages);
581 if (!capsnap->dirty_pages)
582 continue;
583
584 /* get i_size, truncate_{seq,size} for page_snapc? */
585 if (snapc && capsnap->context != page_snapc)
586 continue;
587
588 if (ctl) {
589 if (capsnap->writing) {
590 ctl->i_size = i_size_read(inode);
591 ctl->size_stable = false;
592 } else {
593 ctl->i_size = capsnap->size;
594 ctl->size_stable = true;
595 }
596 ctl->truncate_size = capsnap->truncate_size;
597 ctl->truncate_seq = capsnap->truncate_seq;
598 ctl->head_snapc = false;
599 }
600
601 if (snapc)
602 break;
603
604 snapc = ceph_get_snap_context(capsnap->context);
605 if (!page_snapc ||
606 page_snapc == snapc ||
607 page_snapc->seq > snapc->seq)
608 break;
609 }
610 if (!snapc && ci->i_wrbuffer_ref_head) {
611 snapc = ceph_get_snap_context(ci->i_head_snapc);
612 doutc(cl, " head snapc %p has %d dirty pages\n", snapc,
613 ci->i_wrbuffer_ref_head);
614 if (ctl) {
615 ctl->i_size = i_size_read(inode);
616 ctl->truncate_size = ci->i_truncate_size;
617 ctl->truncate_seq = ci->i_truncate_seq;
618 ctl->size_stable = false;
619 ctl->head_snapc = true;
620 }
621 }
622 spin_unlock(&ci->i_ceph_lock);
623 return snapc;
624}
625
626static u64 get_writepages_data_length(struct inode *inode,
627 struct page *page, u64 start)
628{
629 struct ceph_inode_info *ci = ceph_inode(inode);
630 struct ceph_snap_context *snapc;
631 struct ceph_cap_snap *capsnap = NULL;
632 u64 end = i_size_read(inode);
633 u64 ret;
634
635 snapc = page_snap_context(ceph_fscrypt_pagecache_page(page));
636 if (snapc != ci->i_head_snapc) {
637 bool found = false;
638 spin_lock(&ci->i_ceph_lock);
639 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
640 if (capsnap->context == snapc) {
641 if (!capsnap->writing)
642 end = capsnap->size;
643 found = true;
644 break;
645 }
646 }
647 spin_unlock(&ci->i_ceph_lock);
648 WARN_ON(!found);
649 }
650 if (end > ceph_fscrypt_page_offset(page) + thp_size(page))
651 end = ceph_fscrypt_page_offset(page) + thp_size(page);
652 ret = end > start ? end - start : 0;
653 if (ret && fscrypt_is_bounce_page(page))
654 ret = round_up(ret, CEPH_FSCRYPT_BLOCK_SIZE);
655 return ret;
656}
657
658/*
659 * Write a single page, but leave the page locked.
660 *
661 * If we get a write error, mark the mapping for error, but still adjust the
662 * dirty page accounting (i.e., page is no longer dirty).
663 */
664static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
665{
666 struct folio *folio = page_folio(page);
667 struct inode *inode = page->mapping->host;
668 struct ceph_inode_info *ci = ceph_inode(inode);
669 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
670 struct ceph_client *cl = fsc->client;
671 struct ceph_snap_context *snapc, *oldest;
672 loff_t page_off = page_offset(page);
673 int err;
674 loff_t len = thp_size(page);
675 loff_t wlen;
676 struct ceph_writeback_ctl ceph_wbc;
677 struct ceph_osd_client *osdc = &fsc->client->osdc;
678 struct ceph_osd_request *req;
679 bool caching = ceph_is_cache_enabled(inode);
680 struct page *bounce_page = NULL;
681
682 doutc(cl, "%llx.%llx page %p idx %lu\n", ceph_vinop(inode), page,
683 page->index);
684
685 if (ceph_inode_is_shutdown(inode))
686 return -EIO;
687
688 /* verify this is a writeable snap context */
689 snapc = page_snap_context(page);
690 if (!snapc) {
691 doutc(cl, "%llx.%llx page %p not dirty?\n", ceph_vinop(inode),
692 page);
693 return 0;
694 }
695 oldest = get_oldest_context(inode, &ceph_wbc, snapc);
696 if (snapc->seq > oldest->seq) {
697 doutc(cl, "%llx.%llx page %p snapc %p not writeable - noop\n",
698 ceph_vinop(inode), page, snapc);
699 /* we should only noop if called by kswapd */
700 WARN_ON(!(current->flags & PF_MEMALLOC));
701 ceph_put_snap_context(oldest);
702 redirty_page_for_writepage(wbc, page);
703 return 0;
704 }
705 ceph_put_snap_context(oldest);
706
707 /* is this a partial page at end of file? */
708 if (page_off >= ceph_wbc.i_size) {
709 doutc(cl, "%llx.%llx folio at %lu beyond eof %llu\n",
710 ceph_vinop(inode), folio->index, ceph_wbc.i_size);
711 folio_invalidate(folio, 0, folio_size(folio));
712 return 0;
713 }
714
715 if (ceph_wbc.i_size < page_off + len)
716 len = ceph_wbc.i_size - page_off;
717
718 wlen = IS_ENCRYPTED(inode) ? round_up(len, CEPH_FSCRYPT_BLOCK_SIZE) : len;
719 doutc(cl, "%llx.%llx page %p index %lu on %llu~%llu snapc %p seq %lld\n",
720 ceph_vinop(inode), page, page->index, page_off, wlen, snapc,
721 snapc->seq);
722
723 if (atomic_long_inc_return(&fsc->writeback_count) >
724 CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
725 fsc->write_congested = true;
726
727 req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode),
728 page_off, &wlen, 0, 1, CEPH_OSD_OP_WRITE,
729 CEPH_OSD_FLAG_WRITE, snapc,
730 ceph_wbc.truncate_seq,
731 ceph_wbc.truncate_size, true);
732 if (IS_ERR(req)) {
733 redirty_page_for_writepage(wbc, page);
734 return PTR_ERR(req);
735 }
736
737 if (wlen < len)
738 len = wlen;
739
740 set_page_writeback(page);
741 if (caching)
742 ceph_set_page_fscache(page);
743 ceph_fscache_write_to_cache(inode, page_off, len, caching);
744
745 if (IS_ENCRYPTED(inode)) {
746 bounce_page = fscrypt_encrypt_pagecache_blocks(page,
747 CEPH_FSCRYPT_BLOCK_SIZE, 0,
748 GFP_NOFS);
749 if (IS_ERR(bounce_page)) {
750 redirty_page_for_writepage(wbc, page);
751 end_page_writeback(page);
752 ceph_osdc_put_request(req);
753 return PTR_ERR(bounce_page);
754 }
755 }
756
757 /* it may be a short write due to an object boundary */
758 WARN_ON_ONCE(len > thp_size(page));
759 osd_req_op_extent_osd_data_pages(req, 0,
760 bounce_page ? &bounce_page : &page, wlen, 0,
761 false, false);
762 doutc(cl, "%llx.%llx %llu~%llu (%llu bytes, %sencrypted)\n",
763 ceph_vinop(inode), page_off, len, wlen,
764 IS_ENCRYPTED(inode) ? "" : "not ");
765
766 req->r_mtime = inode_get_mtime(inode);
767 ceph_osdc_start_request(osdc, req);
768 err = ceph_osdc_wait_request(osdc, req);
769
770 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
771 req->r_end_latency, len, err);
772 fscrypt_free_bounce_page(bounce_page);
773 ceph_osdc_put_request(req);
774 if (err == 0)
775 err = len;
776
777 if (err < 0) {
778 struct writeback_control tmp_wbc;
779 if (!wbc)
780 wbc = &tmp_wbc;
781 if (err == -ERESTARTSYS) {
782 /* killed by SIGKILL */
783 doutc(cl, "%llx.%llx interrupted page %p\n",
784 ceph_vinop(inode), page);
785 redirty_page_for_writepage(wbc, page);
786 end_page_writeback(page);
787 return err;
788 }
789 if (err == -EBLOCKLISTED)
790 fsc->blocklisted = true;
791 doutc(cl, "%llx.%llx setting page/mapping error %d %p\n",
792 ceph_vinop(inode), err, page);
793 mapping_set_error(&inode->i_data, err);
794 wbc->pages_skipped++;
795 } else {
796 doutc(cl, "%llx.%llx cleaned page %p\n",
797 ceph_vinop(inode), page);
798 err = 0; /* vfs expects us to return 0 */
799 }
800 oldest = detach_page_private(page);
801 WARN_ON_ONCE(oldest != snapc);
802 end_page_writeback(page);
803 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
804 ceph_put_snap_context(snapc); /* page's reference */
805
806 if (atomic_long_dec_return(&fsc->writeback_count) <
807 CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
808 fsc->write_congested = false;
809
810 return err;
811}
812
813static int ceph_writepage(struct page *page, struct writeback_control *wbc)
814{
815 int err;
816 struct inode *inode = page->mapping->host;
817 BUG_ON(!inode);
818 ihold(inode);
819
820 if (wbc->sync_mode == WB_SYNC_NONE &&
821 ceph_inode_to_fs_client(inode)->write_congested) {
822 redirty_page_for_writepage(wbc, page);
823 return AOP_WRITEPAGE_ACTIVATE;
824 }
825
826 folio_wait_private_2(page_folio(page)); /* [DEPRECATED] */
827
828 err = writepage_nounlock(page, wbc);
829 if (err == -ERESTARTSYS) {
830 /* direct memory reclaimer was killed by SIGKILL. return 0
831 * to prevent caller from setting mapping/page error */
832 err = 0;
833 }
834 unlock_page(page);
835 iput(inode);
836 return err;
837}
838
839/*
840 * async writeback completion handler.
841 *
842 * If we get an error, set the mapping error bit, but not the individual
843 * page error bits.
844 */
845static void writepages_finish(struct ceph_osd_request *req)
846{
847 struct inode *inode = req->r_inode;
848 struct ceph_inode_info *ci = ceph_inode(inode);
849 struct ceph_client *cl = ceph_inode_to_client(inode);
850 struct ceph_osd_data *osd_data;
851 struct page *page;
852 int num_pages, total_pages = 0;
853 int i, j;
854 int rc = req->r_result;
855 struct ceph_snap_context *snapc = req->r_snapc;
856 struct address_space *mapping = inode->i_mapping;
857 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
858 unsigned int len = 0;
859 bool remove_page;
860
861 doutc(cl, "%llx.%llx rc %d\n", ceph_vinop(inode), rc);
862 if (rc < 0) {
863 mapping_set_error(mapping, rc);
864 ceph_set_error_write(ci);
865 if (rc == -EBLOCKLISTED)
866 fsc->blocklisted = true;
867 } else {
868 ceph_clear_error_write(ci);
869 }
870
871 /*
872 * We lost the cache cap, need to truncate the page before
873 * it is unlocked, otherwise we'd truncate it later in the
874 * page truncation thread, possibly losing some data that
875 * raced its way in
876 */
877 remove_page = !(ceph_caps_issued(ci) &
878 (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
879
880 /* clean all pages */
881 for (i = 0; i < req->r_num_ops; i++) {
882 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE) {
883 pr_warn_client(cl,
884 "%llx.%llx incorrect op %d req %p index %d tid %llu\n",
885 ceph_vinop(inode), req->r_ops[i].op, req, i,
886 req->r_tid);
887 break;
888 }
889
890 osd_data = osd_req_op_extent_osd_data(req, i);
891 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
892 len += osd_data->length;
893 num_pages = calc_pages_for((u64)osd_data->alignment,
894 (u64)osd_data->length);
895 total_pages += num_pages;
896 for (j = 0; j < num_pages; j++) {
897 page = osd_data->pages[j];
898 if (fscrypt_is_bounce_page(page)) {
899 page = fscrypt_pagecache_page(page);
900 fscrypt_free_bounce_page(osd_data->pages[j]);
901 osd_data->pages[j] = page;
902 }
903 BUG_ON(!page);
904 WARN_ON(!PageUptodate(page));
905
906 if (atomic_long_dec_return(&fsc->writeback_count) <
907 CONGESTION_OFF_THRESH(
908 fsc->mount_options->congestion_kb))
909 fsc->write_congested = false;
910
911 ceph_put_snap_context(detach_page_private(page));
912 end_page_writeback(page);
913 doutc(cl, "unlocking %p\n", page);
914
915 if (remove_page)
916 generic_error_remove_folio(inode->i_mapping,
917 page_folio(page));
918
919 unlock_page(page);
920 }
921 doutc(cl, "%llx.%llx wrote %llu bytes cleaned %d pages\n",
922 ceph_vinop(inode), osd_data->length,
923 rc >= 0 ? num_pages : 0);
924
925 release_pages(osd_data->pages, num_pages);
926 }
927
928 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
929 req->r_end_latency, len, rc);
930
931 ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
932
933 osd_data = osd_req_op_extent_osd_data(req, 0);
934 if (osd_data->pages_from_pool)
935 mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
936 else
937 kfree(osd_data->pages);
938 ceph_osdc_put_request(req);
939 ceph_dec_osd_stopping_blocker(fsc->mdsc);
940}
941
942/*
943 * initiate async writeback
944 */
945static int ceph_writepages_start(struct address_space *mapping,
946 struct writeback_control *wbc)
947{
948 struct inode *inode = mapping->host;
949 struct ceph_inode_info *ci = ceph_inode(inode);
950 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
951 struct ceph_client *cl = fsc->client;
952 struct ceph_vino vino = ceph_vino(inode);
953 pgoff_t index, start_index, end = -1;
954 struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
955 struct folio_batch fbatch;
956 int rc = 0;
957 unsigned int wsize = i_blocksize(inode);
958 struct ceph_osd_request *req = NULL;
959 struct ceph_writeback_ctl ceph_wbc;
960 bool should_loop, range_whole = false;
961 bool done = false;
962 bool caching = ceph_is_cache_enabled(inode);
963 xa_mark_t tag;
964
965 if (wbc->sync_mode == WB_SYNC_NONE &&
966 fsc->write_congested)
967 return 0;
968
969 doutc(cl, "%llx.%llx (mode=%s)\n", ceph_vinop(inode),
970 wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
971 (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
972
973 if (ceph_inode_is_shutdown(inode)) {
974 if (ci->i_wrbuffer_ref > 0) {
975 pr_warn_ratelimited_client(cl,
976 "%llx.%llx %lld forced umount\n",
977 ceph_vinop(inode), ceph_ino(inode));
978 }
979 mapping_set_error(mapping, -EIO);
980 return -EIO; /* we're in a forced umount, don't write! */
981 }
982 if (fsc->mount_options->wsize < wsize)
983 wsize = fsc->mount_options->wsize;
984
985 folio_batch_init(&fbatch);
986
987 start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
988 index = start_index;
989
990 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
991 tag = PAGECACHE_TAG_TOWRITE;
992 } else {
993 tag = PAGECACHE_TAG_DIRTY;
994 }
995retry:
996 /* find oldest snap context with dirty data */
997 snapc = get_oldest_context(inode, &ceph_wbc, NULL);
998 if (!snapc) {
999 /* hmm, why does writepages get called when there
1000 is no dirty data? */
1001 doutc(cl, " no snap context with dirty data?\n");
1002 goto out;
1003 }
1004 doutc(cl, " oldest snapc is %p seq %lld (%d snaps)\n", snapc,
1005 snapc->seq, snapc->num_snaps);
1006
1007 should_loop = false;
1008 if (ceph_wbc.head_snapc && snapc != last_snapc) {
1009 /* where to start/end? */
1010 if (wbc->range_cyclic) {
1011 index = start_index;
1012 end = -1;
1013 if (index > 0)
1014 should_loop = true;
1015 doutc(cl, " cyclic, start at %lu\n", index);
1016 } else {
1017 index = wbc->range_start >> PAGE_SHIFT;
1018 end = wbc->range_end >> PAGE_SHIFT;
1019 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1020 range_whole = true;
1021 doutc(cl, " not cyclic, %lu to %lu\n", index, end);
1022 }
1023 } else if (!ceph_wbc.head_snapc) {
1024 /* Do not respect wbc->range_{start,end}. Dirty pages
1025 * in that range can be associated with newer snapc.
1026 * They are not writeable until we write all dirty pages
1027 * associated with 'snapc' get written */
1028 if (index > 0)
1029 should_loop = true;
1030 doutc(cl, " non-head snapc, range whole\n");
1031 }
1032
1033 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1034 tag_pages_for_writeback(mapping, index, end);
1035
1036 ceph_put_snap_context(last_snapc);
1037 last_snapc = snapc;
1038
1039 while (!done && index <= end) {
1040 int num_ops = 0, op_idx;
1041 unsigned i, nr_folios, max_pages, locked_pages = 0;
1042 struct page **pages = NULL, **data_pages;
1043 struct page *page;
1044 pgoff_t strip_unit_end = 0;
1045 u64 offset = 0, len = 0;
1046 bool from_pool = false;
1047
1048 max_pages = wsize >> PAGE_SHIFT;
1049
1050get_more_pages:
1051 nr_folios = filemap_get_folios_tag(mapping, &index,
1052 end, tag, &fbatch);
1053 doutc(cl, "pagevec_lookup_range_tag got %d\n", nr_folios);
1054 if (!nr_folios && !locked_pages)
1055 break;
1056 for (i = 0; i < nr_folios && locked_pages < max_pages; i++) {
1057 struct folio *folio = fbatch.folios[i];
1058
1059 page = &folio->page;
1060 doutc(cl, "? %p idx %lu\n", page, page->index);
1061 if (locked_pages == 0)
1062 lock_page(page); /* first page */
1063 else if (!trylock_page(page))
1064 break;
1065
1066 /* only dirty pages, or our accounting breaks */
1067 if (unlikely(!PageDirty(page)) ||
1068 unlikely(page->mapping != mapping)) {
1069 doutc(cl, "!dirty or !mapping %p\n", page);
1070 unlock_page(page);
1071 continue;
1072 }
1073 /* only if matching snap context */
1074 pgsnapc = page_snap_context(page);
1075 if (pgsnapc != snapc) {
1076 doutc(cl, "page snapc %p %lld != oldest %p %lld\n",
1077 pgsnapc, pgsnapc->seq, snapc, snapc->seq);
1078 if (!should_loop &&
1079 !ceph_wbc.head_snapc &&
1080 wbc->sync_mode != WB_SYNC_NONE)
1081 should_loop = true;
1082 unlock_page(page);
1083 continue;
1084 }
1085 if (page_offset(page) >= ceph_wbc.i_size) {
1086 doutc(cl, "folio at %lu beyond eof %llu\n",
1087 folio->index, ceph_wbc.i_size);
1088 if ((ceph_wbc.size_stable ||
1089 folio_pos(folio) >= i_size_read(inode)) &&
1090 folio_clear_dirty_for_io(folio))
1091 folio_invalidate(folio, 0,
1092 folio_size(folio));
1093 folio_unlock(folio);
1094 continue;
1095 }
1096 if (strip_unit_end && (page->index > strip_unit_end)) {
1097 doutc(cl, "end of strip unit %p\n", page);
1098 unlock_page(page);
1099 break;
1100 }
1101 if (folio_test_writeback(folio) ||
1102 folio_test_private_2(folio) /* [DEPRECATED] */) {
1103 if (wbc->sync_mode == WB_SYNC_NONE) {
1104 doutc(cl, "%p under writeback\n", folio);
1105 folio_unlock(folio);
1106 continue;
1107 }
1108 doutc(cl, "waiting on writeback %p\n", folio);
1109 folio_wait_writeback(folio);
1110 folio_wait_private_2(folio); /* [DEPRECATED] */
1111 }
1112
1113 if (!clear_page_dirty_for_io(page)) {
1114 doutc(cl, "%p !clear_page_dirty_for_io\n", page);
1115 unlock_page(page);
1116 continue;
1117 }
1118
1119 /*
1120 * We have something to write. If this is
1121 * the first locked page this time through,
1122 * calculate max possinle write size and
1123 * allocate a page array
1124 */
1125 if (locked_pages == 0) {
1126 u64 objnum;
1127 u64 objoff;
1128 u32 xlen;
1129
1130 /* prepare async write request */
1131 offset = (u64)page_offset(page);
1132 ceph_calc_file_object_mapping(&ci->i_layout,
1133 offset, wsize,
1134 &objnum, &objoff,
1135 &xlen);
1136 len = xlen;
1137
1138 num_ops = 1;
1139 strip_unit_end = page->index +
1140 ((len - 1) >> PAGE_SHIFT);
1141
1142 BUG_ON(pages);
1143 max_pages = calc_pages_for(0, (u64)len);
1144 pages = kmalloc_array(max_pages,
1145 sizeof(*pages),
1146 GFP_NOFS);
1147 if (!pages) {
1148 from_pool = true;
1149 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1150 BUG_ON(!pages);
1151 }
1152
1153 len = 0;
1154 } else if (page->index !=
1155 (offset + len) >> PAGE_SHIFT) {
1156 if (num_ops >= (from_pool ? CEPH_OSD_SLAB_OPS :
1157 CEPH_OSD_MAX_OPS)) {
1158 redirty_page_for_writepage(wbc, page);
1159 unlock_page(page);
1160 break;
1161 }
1162
1163 num_ops++;
1164 offset = (u64)page_offset(page);
1165 len = 0;
1166 }
1167
1168 /* note position of first page in fbatch */
1169 doutc(cl, "%llx.%llx will write page %p idx %lu\n",
1170 ceph_vinop(inode), page, page->index);
1171
1172 if (atomic_long_inc_return(&fsc->writeback_count) >
1173 CONGESTION_ON_THRESH(
1174 fsc->mount_options->congestion_kb))
1175 fsc->write_congested = true;
1176
1177 if (IS_ENCRYPTED(inode)) {
1178 pages[locked_pages] =
1179 fscrypt_encrypt_pagecache_blocks(page,
1180 PAGE_SIZE, 0,
1181 locked_pages ? GFP_NOWAIT : GFP_NOFS);
1182 if (IS_ERR(pages[locked_pages])) {
1183 if (PTR_ERR(pages[locked_pages]) == -EINVAL)
1184 pr_err_client(cl,
1185 "inode->i_blkbits=%hhu\n",
1186 inode->i_blkbits);
1187 /* better not fail on first page! */
1188 BUG_ON(locked_pages == 0);
1189 pages[locked_pages] = NULL;
1190 redirty_page_for_writepage(wbc, page);
1191 unlock_page(page);
1192 break;
1193 }
1194 ++locked_pages;
1195 } else {
1196 pages[locked_pages++] = page;
1197 }
1198
1199 fbatch.folios[i] = NULL;
1200 len += thp_size(page);
1201 }
1202
1203 /* did we get anything? */
1204 if (!locked_pages)
1205 goto release_folios;
1206 if (i) {
1207 unsigned j, n = 0;
1208 /* shift unused page to beginning of fbatch */
1209 for (j = 0; j < nr_folios; j++) {
1210 if (!fbatch.folios[j])
1211 continue;
1212 if (n < j)
1213 fbatch.folios[n] = fbatch.folios[j];
1214 n++;
1215 }
1216 fbatch.nr = n;
1217
1218 if (nr_folios && i == nr_folios &&
1219 locked_pages < max_pages) {
1220 doutc(cl, "reached end fbatch, trying for more\n");
1221 folio_batch_release(&fbatch);
1222 goto get_more_pages;
1223 }
1224 }
1225
1226new_request:
1227 offset = ceph_fscrypt_page_offset(pages[0]);
1228 len = wsize;
1229
1230 req = ceph_osdc_new_request(&fsc->client->osdc,
1231 &ci->i_layout, vino,
1232 offset, &len, 0, num_ops,
1233 CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1234 snapc, ceph_wbc.truncate_seq,
1235 ceph_wbc.truncate_size, false);
1236 if (IS_ERR(req)) {
1237 req = ceph_osdc_new_request(&fsc->client->osdc,
1238 &ci->i_layout, vino,
1239 offset, &len, 0,
1240 min(num_ops,
1241 CEPH_OSD_SLAB_OPS),
1242 CEPH_OSD_OP_WRITE,
1243 CEPH_OSD_FLAG_WRITE,
1244 snapc, ceph_wbc.truncate_seq,
1245 ceph_wbc.truncate_size, true);
1246 BUG_ON(IS_ERR(req));
1247 }
1248 BUG_ON(len < ceph_fscrypt_page_offset(pages[locked_pages - 1]) +
1249 thp_size(pages[locked_pages - 1]) - offset);
1250
1251 if (!ceph_inc_osd_stopping_blocker(fsc->mdsc)) {
1252 rc = -EIO;
1253 goto release_folios;
1254 }
1255 req->r_callback = writepages_finish;
1256 req->r_inode = inode;
1257
1258 /* Format the osd request message and submit the write */
1259 len = 0;
1260 data_pages = pages;
1261 op_idx = 0;
1262 for (i = 0; i < locked_pages; i++) {
1263 struct page *page = ceph_fscrypt_pagecache_page(pages[i]);
1264
1265 u64 cur_offset = page_offset(page);
1266 /*
1267 * Discontinuity in page range? Ceph can handle that by just passing
1268 * multiple extents in the write op.
1269 */
1270 if (offset + len != cur_offset) {
1271 /* If it's full, stop here */
1272 if (op_idx + 1 == req->r_num_ops)
1273 break;
1274
1275 /* Kick off an fscache write with what we have so far. */
1276 ceph_fscache_write_to_cache(inode, offset, len, caching);
1277
1278 /* Start a new extent */
1279 osd_req_op_extent_dup_last(req, op_idx,
1280 cur_offset - offset);
1281 doutc(cl, "got pages at %llu~%llu\n", offset,
1282 len);
1283 osd_req_op_extent_osd_data_pages(req, op_idx,
1284 data_pages, len, 0,
1285 from_pool, false);
1286 osd_req_op_extent_update(req, op_idx, len);
1287
1288 len = 0;
1289 offset = cur_offset;
1290 data_pages = pages + i;
1291 op_idx++;
1292 }
1293
1294 set_page_writeback(page);
1295 if (caching)
1296 ceph_set_page_fscache(page);
1297 len += thp_size(page);
1298 }
1299 ceph_fscache_write_to_cache(inode, offset, len, caching);
1300
1301 if (ceph_wbc.size_stable) {
1302 len = min(len, ceph_wbc.i_size - offset);
1303 } else if (i == locked_pages) {
1304 /* writepages_finish() clears writeback pages
1305 * according to the data length, so make sure
1306 * data length covers all locked pages */
1307 u64 min_len = len + 1 - thp_size(page);
1308 len = get_writepages_data_length(inode, pages[i - 1],
1309 offset);
1310 len = max(len, min_len);
1311 }
1312 if (IS_ENCRYPTED(inode))
1313 len = round_up(len, CEPH_FSCRYPT_BLOCK_SIZE);
1314
1315 doutc(cl, "got pages at %llu~%llu\n", offset, len);
1316
1317 if (IS_ENCRYPTED(inode) &&
1318 ((offset | len) & ~CEPH_FSCRYPT_BLOCK_MASK))
1319 pr_warn_client(cl,
1320 "bad encrypted write offset=%lld len=%llu\n",
1321 offset, len);
1322
1323 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1324 0, from_pool, false);
1325 osd_req_op_extent_update(req, op_idx, len);
1326
1327 BUG_ON(op_idx + 1 != req->r_num_ops);
1328
1329 from_pool = false;
1330 if (i < locked_pages) {
1331 BUG_ON(num_ops <= req->r_num_ops);
1332 num_ops -= req->r_num_ops;
1333 locked_pages -= i;
1334
1335 /* allocate new pages array for next request */
1336 data_pages = pages;
1337 pages = kmalloc_array(locked_pages, sizeof(*pages),
1338 GFP_NOFS);
1339 if (!pages) {
1340 from_pool = true;
1341 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1342 BUG_ON(!pages);
1343 }
1344 memcpy(pages, data_pages + i,
1345 locked_pages * sizeof(*pages));
1346 memset(data_pages + i, 0,
1347 locked_pages * sizeof(*pages));
1348 } else {
1349 BUG_ON(num_ops != req->r_num_ops);
1350 index = pages[i - 1]->index + 1;
1351 /* request message now owns the pages array */
1352 pages = NULL;
1353 }
1354
1355 req->r_mtime = inode_get_mtime(inode);
1356 ceph_osdc_start_request(&fsc->client->osdc, req);
1357 req = NULL;
1358
1359 wbc->nr_to_write -= i;
1360 if (pages)
1361 goto new_request;
1362
1363 /*
1364 * We stop writing back only if we are not doing
1365 * integrity sync. In case of integrity sync we have to
1366 * keep going until we have written all the pages
1367 * we tagged for writeback prior to entering this loop.
1368 */
1369 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1370 done = true;
1371
1372release_folios:
1373 doutc(cl, "folio_batch release on %d folios (%p)\n",
1374 (int)fbatch.nr, fbatch.nr ? fbatch.folios[0] : NULL);
1375 folio_batch_release(&fbatch);
1376 }
1377
1378 if (should_loop && !done) {
1379 /* more to do; loop back to beginning of file */
1380 doutc(cl, "looping back to beginning of file\n");
1381 end = start_index - 1; /* OK even when start_index == 0 */
1382
1383 /* to write dirty pages associated with next snapc,
1384 * we need to wait until current writes complete */
1385 if (wbc->sync_mode != WB_SYNC_NONE &&
1386 start_index == 0 && /* all dirty pages were checked */
1387 !ceph_wbc.head_snapc) {
1388 struct page *page;
1389 unsigned i, nr;
1390 index = 0;
1391 while ((index <= end) &&
1392 (nr = filemap_get_folios_tag(mapping, &index,
1393 (pgoff_t)-1,
1394 PAGECACHE_TAG_WRITEBACK,
1395 &fbatch))) {
1396 for (i = 0; i < nr; i++) {
1397 page = &fbatch.folios[i]->page;
1398 if (page_snap_context(page) != snapc)
1399 continue;
1400 wait_on_page_writeback(page);
1401 }
1402 folio_batch_release(&fbatch);
1403 cond_resched();
1404 }
1405 }
1406
1407 start_index = 0;
1408 index = 0;
1409 goto retry;
1410 }
1411
1412 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1413 mapping->writeback_index = index;
1414
1415out:
1416 ceph_osdc_put_request(req);
1417 ceph_put_snap_context(last_snapc);
1418 doutc(cl, "%llx.%llx dend - startone, rc = %d\n", ceph_vinop(inode),
1419 rc);
1420 return rc;
1421}
1422
1423
1424
1425/*
1426 * See if a given @snapc is either writeable, or already written.
1427 */
1428static int context_is_writeable_or_written(struct inode *inode,
1429 struct ceph_snap_context *snapc)
1430{
1431 struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1432 int ret = !oldest || snapc->seq <= oldest->seq;
1433
1434 ceph_put_snap_context(oldest);
1435 return ret;
1436}
1437
1438/**
1439 * ceph_find_incompatible - find an incompatible context and return it
1440 * @page: page being dirtied
1441 *
1442 * We are only allowed to write into/dirty a page if the page is
1443 * clean, or already dirty within the same snap context. Returns a
1444 * conflicting context if there is one, NULL if there isn't, or a
1445 * negative error code on other errors.
1446 *
1447 * Must be called with page lock held.
1448 */
1449static struct ceph_snap_context *
1450ceph_find_incompatible(struct page *page)
1451{
1452 struct inode *inode = page->mapping->host;
1453 struct ceph_client *cl = ceph_inode_to_client(inode);
1454 struct ceph_inode_info *ci = ceph_inode(inode);
1455
1456 if (ceph_inode_is_shutdown(inode)) {
1457 doutc(cl, " %llx.%llx page %p is shutdown\n",
1458 ceph_vinop(inode), page);
1459 return ERR_PTR(-ESTALE);
1460 }
1461
1462 for (;;) {
1463 struct ceph_snap_context *snapc, *oldest;
1464
1465 wait_on_page_writeback(page);
1466
1467 snapc = page_snap_context(page);
1468 if (!snapc || snapc == ci->i_head_snapc)
1469 break;
1470
1471 /*
1472 * this page is already dirty in another (older) snap
1473 * context! is it writeable now?
1474 */
1475 oldest = get_oldest_context(inode, NULL, NULL);
1476 if (snapc->seq > oldest->seq) {
1477 /* not writeable -- return it for the caller to deal with */
1478 ceph_put_snap_context(oldest);
1479 doutc(cl, " %llx.%llx page %p snapc %p not current or oldest\n",
1480 ceph_vinop(inode), page, snapc);
1481 return ceph_get_snap_context(snapc);
1482 }
1483 ceph_put_snap_context(oldest);
1484
1485 /* yay, writeable, do it now (without dropping page lock) */
1486 doutc(cl, " %llx.%llx page %p snapc %p not current, but oldest\n",
1487 ceph_vinop(inode), page, snapc);
1488 if (clear_page_dirty_for_io(page)) {
1489 int r = writepage_nounlock(page, NULL);
1490 if (r < 0)
1491 return ERR_PTR(r);
1492 }
1493 }
1494 return NULL;
1495}
1496
1497static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
1498 struct folio **foliop, void **_fsdata)
1499{
1500 struct inode *inode = file_inode(file);
1501 struct ceph_inode_info *ci = ceph_inode(inode);
1502 struct ceph_snap_context *snapc;
1503
1504 snapc = ceph_find_incompatible(folio_page(*foliop, 0));
1505 if (snapc) {
1506 int r;
1507
1508 folio_unlock(*foliop);
1509 folio_put(*foliop);
1510 *foliop = NULL;
1511 if (IS_ERR(snapc))
1512 return PTR_ERR(snapc);
1513
1514 ceph_queue_writeback(inode);
1515 r = wait_event_killable(ci->i_cap_wq,
1516 context_is_writeable_or_written(inode, snapc));
1517 ceph_put_snap_context(snapc);
1518 return r == 0 ? -EAGAIN : r;
1519 }
1520 return 0;
1521}
1522
1523/*
1524 * We are only allowed to write into/dirty the page if the page is
1525 * clean, or already dirty within the same snap context.
1526 */
1527static int ceph_write_begin(struct file *file, struct address_space *mapping,
1528 loff_t pos, unsigned len,
1529 struct folio **foliop, void **fsdata)
1530{
1531 struct inode *inode = file_inode(file);
1532 struct ceph_inode_info *ci = ceph_inode(inode);
1533 int r;
1534
1535 r = netfs_write_begin(&ci->netfs, file, inode->i_mapping, pos, len, foliop, NULL);
1536 if (r < 0)
1537 return r;
1538
1539 folio_wait_private_2(*foliop); /* [DEPRECATED] */
1540 WARN_ON_ONCE(!folio_test_locked(*foliop));
1541 return 0;
1542}
1543
1544/*
1545 * we don't do anything in here that simple_write_end doesn't do
1546 * except adjust dirty page accounting
1547 */
1548static int ceph_write_end(struct file *file, struct address_space *mapping,
1549 loff_t pos, unsigned len, unsigned copied,
1550 struct folio *folio, void *fsdata)
1551{
1552 struct inode *inode = file_inode(file);
1553 struct ceph_client *cl = ceph_inode_to_client(inode);
1554 bool check_cap = false;
1555
1556 doutc(cl, "%llx.%llx file %p folio %p %d~%d (%d)\n", ceph_vinop(inode),
1557 file, folio, (int)pos, (int)copied, (int)len);
1558
1559 if (!folio_test_uptodate(folio)) {
1560 /* just return that nothing was copied on a short copy */
1561 if (copied < len) {
1562 copied = 0;
1563 goto out;
1564 }
1565 folio_mark_uptodate(folio);
1566 }
1567
1568 /* did file size increase? */
1569 if (pos+copied > i_size_read(inode))
1570 check_cap = ceph_inode_set_size(inode, pos+copied);
1571
1572 folio_mark_dirty(folio);
1573
1574out:
1575 folio_unlock(folio);
1576 folio_put(folio);
1577
1578 if (check_cap)
1579 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY);
1580
1581 return copied;
1582}
1583
1584const struct address_space_operations ceph_aops = {
1585 .read_folio = netfs_read_folio,
1586 .readahead = netfs_readahead,
1587 .writepage = ceph_writepage,
1588 .writepages = ceph_writepages_start,
1589 .write_begin = ceph_write_begin,
1590 .write_end = ceph_write_end,
1591 .dirty_folio = ceph_dirty_folio,
1592 .invalidate_folio = ceph_invalidate_folio,
1593 .release_folio = netfs_release_folio,
1594 .direct_IO = noop_direct_IO,
1595};
1596
1597static void ceph_block_sigs(sigset_t *oldset)
1598{
1599 sigset_t mask;
1600 siginitsetinv(&mask, sigmask(SIGKILL));
1601 sigprocmask(SIG_BLOCK, &mask, oldset);
1602}
1603
1604static void ceph_restore_sigs(sigset_t *oldset)
1605{
1606 sigprocmask(SIG_SETMASK, oldset, NULL);
1607}
1608
1609/*
1610 * vm ops
1611 */
1612static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1613{
1614 struct vm_area_struct *vma = vmf->vma;
1615 struct inode *inode = file_inode(vma->vm_file);
1616 struct ceph_inode_info *ci = ceph_inode(inode);
1617 struct ceph_client *cl = ceph_inode_to_client(inode);
1618 struct ceph_file_info *fi = vma->vm_file->private_data;
1619 loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1620 int want, got, err;
1621 sigset_t oldset;
1622 vm_fault_t ret = VM_FAULT_SIGBUS;
1623
1624 if (ceph_inode_is_shutdown(inode))
1625 return ret;
1626
1627 ceph_block_sigs(&oldset);
1628
1629 doutc(cl, "%llx.%llx %llu trying to get caps\n",
1630 ceph_vinop(inode), off);
1631 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1632 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1633 else
1634 want = CEPH_CAP_FILE_CACHE;
1635
1636 got = 0;
1637 err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1, &got);
1638 if (err < 0)
1639 goto out_restore;
1640
1641 doutc(cl, "%llx.%llx %llu got cap refs on %s\n", ceph_vinop(inode),
1642 off, ceph_cap_string(got));
1643
1644 if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1645 !ceph_has_inline_data(ci)) {
1646 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1647 ceph_add_rw_context(fi, &rw_ctx);
1648 ret = filemap_fault(vmf);
1649 ceph_del_rw_context(fi, &rw_ctx);
1650 doutc(cl, "%llx.%llx %llu drop cap refs %s ret %x\n",
1651 ceph_vinop(inode), off, ceph_cap_string(got), ret);
1652 } else
1653 err = -EAGAIN;
1654
1655 ceph_put_cap_refs(ci, got);
1656
1657 if (err != -EAGAIN)
1658 goto out_restore;
1659
1660 /* read inline data */
1661 if (off >= PAGE_SIZE) {
1662 /* does not support inline data > PAGE_SIZE */
1663 ret = VM_FAULT_SIGBUS;
1664 } else {
1665 struct address_space *mapping = inode->i_mapping;
1666 struct page *page;
1667
1668 filemap_invalidate_lock_shared(mapping);
1669 page = find_or_create_page(mapping, 0,
1670 mapping_gfp_constraint(mapping, ~__GFP_FS));
1671 if (!page) {
1672 ret = VM_FAULT_OOM;
1673 goto out_inline;
1674 }
1675 err = __ceph_do_getattr(inode, page,
1676 CEPH_STAT_CAP_INLINE_DATA, true);
1677 if (err < 0 || off >= i_size_read(inode)) {
1678 unlock_page(page);
1679 put_page(page);
1680 ret = vmf_error(err);
1681 goto out_inline;
1682 }
1683 if (err < PAGE_SIZE)
1684 zero_user_segment(page, err, PAGE_SIZE);
1685 else
1686 flush_dcache_page(page);
1687 SetPageUptodate(page);
1688 vmf->page = page;
1689 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1690out_inline:
1691 filemap_invalidate_unlock_shared(mapping);
1692 doutc(cl, "%llx.%llx %llu read inline data ret %x\n",
1693 ceph_vinop(inode), off, ret);
1694 }
1695out_restore:
1696 ceph_restore_sigs(&oldset);
1697 if (err < 0)
1698 ret = vmf_error(err);
1699
1700 return ret;
1701}
1702
1703static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1704{
1705 struct vm_area_struct *vma = vmf->vma;
1706 struct inode *inode = file_inode(vma->vm_file);
1707 struct ceph_client *cl = ceph_inode_to_client(inode);
1708 struct ceph_inode_info *ci = ceph_inode(inode);
1709 struct ceph_file_info *fi = vma->vm_file->private_data;
1710 struct ceph_cap_flush *prealloc_cf;
1711 struct page *page = vmf->page;
1712 loff_t off = page_offset(page);
1713 loff_t size = i_size_read(inode);
1714 size_t len;
1715 int want, got, err;
1716 sigset_t oldset;
1717 vm_fault_t ret = VM_FAULT_SIGBUS;
1718
1719 if (ceph_inode_is_shutdown(inode))
1720 return ret;
1721
1722 prealloc_cf = ceph_alloc_cap_flush();
1723 if (!prealloc_cf)
1724 return VM_FAULT_OOM;
1725
1726 sb_start_pagefault(inode->i_sb);
1727 ceph_block_sigs(&oldset);
1728
1729 if (off + thp_size(page) <= size)
1730 len = thp_size(page);
1731 else
1732 len = offset_in_thp(page, size);
1733
1734 doutc(cl, "%llx.%llx %llu~%zd getting caps i_size %llu\n",
1735 ceph_vinop(inode), off, len, size);
1736 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1737 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1738 else
1739 want = CEPH_CAP_FILE_BUFFER;
1740
1741 got = 0;
1742 err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len, &got);
1743 if (err < 0)
1744 goto out_free;
1745
1746 doutc(cl, "%llx.%llx %llu~%zd got cap refs on %s\n", ceph_vinop(inode),
1747 off, len, ceph_cap_string(got));
1748
1749 /* Update time before taking page lock */
1750 file_update_time(vma->vm_file);
1751 inode_inc_iversion_raw(inode);
1752
1753 do {
1754 struct ceph_snap_context *snapc;
1755
1756 lock_page(page);
1757
1758 if (page_mkwrite_check_truncate(page, inode) < 0) {
1759 unlock_page(page);
1760 ret = VM_FAULT_NOPAGE;
1761 break;
1762 }
1763
1764 snapc = ceph_find_incompatible(page);
1765 if (!snapc) {
1766 /* success. we'll keep the page locked. */
1767 set_page_dirty(page);
1768 ret = VM_FAULT_LOCKED;
1769 break;
1770 }
1771
1772 unlock_page(page);
1773
1774 if (IS_ERR(snapc)) {
1775 ret = VM_FAULT_SIGBUS;
1776 break;
1777 }
1778
1779 ceph_queue_writeback(inode);
1780 err = wait_event_killable(ci->i_cap_wq,
1781 context_is_writeable_or_written(inode, snapc));
1782 ceph_put_snap_context(snapc);
1783 } while (err == 0);
1784
1785 if (ret == VM_FAULT_LOCKED) {
1786 int dirty;
1787 spin_lock(&ci->i_ceph_lock);
1788 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1789 &prealloc_cf);
1790 spin_unlock(&ci->i_ceph_lock);
1791 if (dirty)
1792 __mark_inode_dirty(inode, dirty);
1793 }
1794
1795 doutc(cl, "%llx.%llx %llu~%zd dropping cap refs on %s ret %x\n",
1796 ceph_vinop(inode), off, len, ceph_cap_string(got), ret);
1797 ceph_put_cap_refs_async(ci, got);
1798out_free:
1799 ceph_restore_sigs(&oldset);
1800 sb_end_pagefault(inode->i_sb);
1801 ceph_free_cap_flush(prealloc_cf);
1802 if (err < 0)
1803 ret = vmf_error(err);
1804 return ret;
1805}
1806
1807void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1808 char *data, size_t len)
1809{
1810 struct ceph_client *cl = ceph_inode_to_client(inode);
1811 struct address_space *mapping = inode->i_mapping;
1812 struct page *page;
1813
1814 if (locked_page) {
1815 page = locked_page;
1816 } else {
1817 if (i_size_read(inode) == 0)
1818 return;
1819 page = find_or_create_page(mapping, 0,
1820 mapping_gfp_constraint(mapping,
1821 ~__GFP_FS));
1822 if (!page)
1823 return;
1824 if (PageUptodate(page)) {
1825 unlock_page(page);
1826 put_page(page);
1827 return;
1828 }
1829 }
1830
1831 doutc(cl, "%p %llx.%llx len %zu locked_page %p\n", inode,
1832 ceph_vinop(inode), len, locked_page);
1833
1834 if (len > 0) {
1835 void *kaddr = kmap_atomic(page);
1836 memcpy(kaddr, data, len);
1837 kunmap_atomic(kaddr);
1838 }
1839
1840 if (page != locked_page) {
1841 if (len < PAGE_SIZE)
1842 zero_user_segment(page, len, PAGE_SIZE);
1843 else
1844 flush_dcache_page(page);
1845
1846 SetPageUptodate(page);
1847 unlock_page(page);
1848 put_page(page);
1849 }
1850}
1851
1852int ceph_uninline_data(struct file *file)
1853{
1854 struct inode *inode = file_inode(file);
1855 struct ceph_inode_info *ci = ceph_inode(inode);
1856 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
1857 struct ceph_client *cl = fsc->client;
1858 struct ceph_osd_request *req = NULL;
1859 struct ceph_cap_flush *prealloc_cf = NULL;
1860 struct folio *folio = NULL;
1861 u64 inline_version = CEPH_INLINE_NONE;
1862 struct page *pages[1];
1863 int err = 0;
1864 u64 len;
1865
1866 spin_lock(&ci->i_ceph_lock);
1867 inline_version = ci->i_inline_version;
1868 spin_unlock(&ci->i_ceph_lock);
1869
1870 doutc(cl, "%llx.%llx inline_version %llu\n", ceph_vinop(inode),
1871 inline_version);
1872
1873 if (ceph_inode_is_shutdown(inode)) {
1874 err = -EIO;
1875 goto out;
1876 }
1877
1878 if (inline_version == CEPH_INLINE_NONE)
1879 return 0;
1880
1881 prealloc_cf = ceph_alloc_cap_flush();
1882 if (!prealloc_cf)
1883 return -ENOMEM;
1884
1885 if (inline_version == 1) /* initial version, no data */
1886 goto out_uninline;
1887
1888 folio = read_mapping_folio(inode->i_mapping, 0, file);
1889 if (IS_ERR(folio)) {
1890 err = PTR_ERR(folio);
1891 goto out;
1892 }
1893
1894 folio_lock(folio);
1895
1896 len = i_size_read(inode);
1897 if (len > folio_size(folio))
1898 len = folio_size(folio);
1899
1900 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1901 ceph_vino(inode), 0, &len, 0, 1,
1902 CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1903 NULL, 0, 0, false);
1904 if (IS_ERR(req)) {
1905 err = PTR_ERR(req);
1906 goto out_unlock;
1907 }
1908
1909 req->r_mtime = inode_get_mtime(inode);
1910 ceph_osdc_start_request(&fsc->client->osdc, req);
1911 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1912 ceph_osdc_put_request(req);
1913 if (err < 0)
1914 goto out_unlock;
1915
1916 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1917 ceph_vino(inode), 0, &len, 1, 3,
1918 CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1919 NULL, ci->i_truncate_seq,
1920 ci->i_truncate_size, false);
1921 if (IS_ERR(req)) {
1922 err = PTR_ERR(req);
1923 goto out_unlock;
1924 }
1925
1926 pages[0] = folio_page(folio, 0);
1927 osd_req_op_extent_osd_data_pages(req, 1, pages, len, 0, false, false);
1928
1929 {
1930 __le64 xattr_buf = cpu_to_le64(inline_version);
1931 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1932 "inline_version", &xattr_buf,
1933 sizeof(xattr_buf),
1934 CEPH_OSD_CMPXATTR_OP_GT,
1935 CEPH_OSD_CMPXATTR_MODE_U64);
1936 if (err)
1937 goto out_put_req;
1938 }
1939
1940 {
1941 char xattr_buf[32];
1942 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1943 "%llu", inline_version);
1944 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1945 "inline_version",
1946 xattr_buf, xattr_len, 0, 0);
1947 if (err)
1948 goto out_put_req;
1949 }
1950
1951 req->r_mtime = inode_get_mtime(inode);
1952 ceph_osdc_start_request(&fsc->client->osdc, req);
1953 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1954
1955 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1956 req->r_end_latency, len, err);
1957
1958out_uninline:
1959 if (!err) {
1960 int dirty;
1961
1962 /* Set to CAP_INLINE_NONE and dirty the caps */
1963 down_read(&fsc->mdsc->snap_rwsem);
1964 spin_lock(&ci->i_ceph_lock);
1965 ci->i_inline_version = CEPH_INLINE_NONE;
1966 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR, &prealloc_cf);
1967 spin_unlock(&ci->i_ceph_lock);
1968 up_read(&fsc->mdsc->snap_rwsem);
1969 if (dirty)
1970 __mark_inode_dirty(inode, dirty);
1971 }
1972out_put_req:
1973 ceph_osdc_put_request(req);
1974 if (err == -ECANCELED)
1975 err = 0;
1976out_unlock:
1977 if (folio) {
1978 folio_unlock(folio);
1979 folio_put(folio);
1980 }
1981out:
1982 ceph_free_cap_flush(prealloc_cf);
1983 doutc(cl, "%llx.%llx inline_version %llu = %d\n",
1984 ceph_vinop(inode), inline_version, err);
1985 return err;
1986}
1987
1988static const struct vm_operations_struct ceph_vmops = {
1989 .fault = ceph_filemap_fault,
1990 .page_mkwrite = ceph_page_mkwrite,
1991};
1992
1993int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1994{
1995 struct address_space *mapping = file->f_mapping;
1996
1997 if (!mapping->a_ops->read_folio)
1998 return -ENOEXEC;
1999 vma->vm_ops = &ceph_vmops;
2000 return 0;
2001}
2002
2003enum {
2004 POOL_READ = 1,
2005 POOL_WRITE = 2,
2006};
2007
2008static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
2009 s64 pool, struct ceph_string *pool_ns)
2010{
2011 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(&ci->netfs.inode);
2012 struct ceph_mds_client *mdsc = fsc->mdsc;
2013 struct ceph_client *cl = fsc->client;
2014 struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
2015 struct rb_node **p, *parent;
2016 struct ceph_pool_perm *perm;
2017 struct page **pages;
2018 size_t pool_ns_len;
2019 int err = 0, err2 = 0, have = 0;
2020
2021 down_read(&mdsc->pool_perm_rwsem);
2022 p = &mdsc->pool_perm_tree.rb_node;
2023 while (*p) {
2024 perm = rb_entry(*p, struct ceph_pool_perm, node);
2025 if (pool < perm->pool)
2026 p = &(*p)->rb_left;
2027 else if (pool > perm->pool)
2028 p = &(*p)->rb_right;
2029 else {
2030 int ret = ceph_compare_string(pool_ns,
2031 perm->pool_ns,
2032 perm->pool_ns_len);
2033 if (ret < 0)
2034 p = &(*p)->rb_left;
2035 else if (ret > 0)
2036 p = &(*p)->rb_right;
2037 else {
2038 have = perm->perm;
2039 break;
2040 }
2041 }
2042 }
2043 up_read(&mdsc->pool_perm_rwsem);
2044 if (*p)
2045 goto out;
2046
2047 if (pool_ns)
2048 doutc(cl, "pool %lld ns %.*s no perm cached\n", pool,
2049 (int)pool_ns->len, pool_ns->str);
2050 else
2051 doutc(cl, "pool %lld no perm cached\n", pool);
2052
2053 down_write(&mdsc->pool_perm_rwsem);
2054 p = &mdsc->pool_perm_tree.rb_node;
2055 parent = NULL;
2056 while (*p) {
2057 parent = *p;
2058 perm = rb_entry(parent, struct ceph_pool_perm, node);
2059 if (pool < perm->pool)
2060 p = &(*p)->rb_left;
2061 else if (pool > perm->pool)
2062 p = &(*p)->rb_right;
2063 else {
2064 int ret = ceph_compare_string(pool_ns,
2065 perm->pool_ns,
2066 perm->pool_ns_len);
2067 if (ret < 0)
2068 p = &(*p)->rb_left;
2069 else if (ret > 0)
2070 p = &(*p)->rb_right;
2071 else {
2072 have = perm->perm;
2073 break;
2074 }
2075 }
2076 }
2077 if (*p) {
2078 up_write(&mdsc->pool_perm_rwsem);
2079 goto out;
2080 }
2081
2082 rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2083 1, false, GFP_NOFS);
2084 if (!rd_req) {
2085 err = -ENOMEM;
2086 goto out_unlock;
2087 }
2088
2089 rd_req->r_flags = CEPH_OSD_FLAG_READ;
2090 osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
2091 rd_req->r_base_oloc.pool = pool;
2092 if (pool_ns)
2093 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
2094 ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
2095
2096 err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
2097 if (err)
2098 goto out_unlock;
2099
2100 wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2101 1, false, GFP_NOFS);
2102 if (!wr_req) {
2103 err = -ENOMEM;
2104 goto out_unlock;
2105 }
2106
2107 wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
2108 osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
2109 ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
2110 ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
2111
2112 err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
2113 if (err)
2114 goto out_unlock;
2115
2116 /* one page should be large enough for STAT data */
2117 pages = ceph_alloc_page_vector(1, GFP_KERNEL);
2118 if (IS_ERR(pages)) {
2119 err = PTR_ERR(pages);
2120 goto out_unlock;
2121 }
2122
2123 osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
2124 0, false, true);
2125 ceph_osdc_start_request(&fsc->client->osdc, rd_req);
2126
2127 wr_req->r_mtime = inode_get_mtime(&ci->netfs.inode);
2128 ceph_osdc_start_request(&fsc->client->osdc, wr_req);
2129
2130 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
2131 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
2132
2133 if (err >= 0 || err == -ENOENT)
2134 have |= POOL_READ;
2135 else if (err != -EPERM) {
2136 if (err == -EBLOCKLISTED)
2137 fsc->blocklisted = true;
2138 goto out_unlock;
2139 }
2140
2141 if (err2 == 0 || err2 == -EEXIST)
2142 have |= POOL_WRITE;
2143 else if (err2 != -EPERM) {
2144 if (err2 == -EBLOCKLISTED)
2145 fsc->blocklisted = true;
2146 err = err2;
2147 goto out_unlock;
2148 }
2149
2150 pool_ns_len = pool_ns ? pool_ns->len : 0;
2151 perm = kmalloc(struct_size(perm, pool_ns, pool_ns_len + 1), GFP_NOFS);
2152 if (!perm) {
2153 err = -ENOMEM;
2154 goto out_unlock;
2155 }
2156
2157 perm->pool = pool;
2158 perm->perm = have;
2159 perm->pool_ns_len = pool_ns_len;
2160 if (pool_ns_len > 0)
2161 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
2162 perm->pool_ns[pool_ns_len] = 0;
2163
2164 rb_link_node(&perm->node, parent, p);
2165 rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
2166 err = 0;
2167out_unlock:
2168 up_write(&mdsc->pool_perm_rwsem);
2169
2170 ceph_osdc_put_request(rd_req);
2171 ceph_osdc_put_request(wr_req);
2172out:
2173 if (!err)
2174 err = have;
2175 if (pool_ns)
2176 doutc(cl, "pool %lld ns %.*s result = %d\n", pool,
2177 (int)pool_ns->len, pool_ns->str, err);
2178 else
2179 doutc(cl, "pool %lld result = %d\n", pool, err);
2180 return err;
2181}
2182
2183int ceph_pool_perm_check(struct inode *inode, int need)
2184{
2185 struct ceph_client *cl = ceph_inode_to_client(inode);
2186 struct ceph_inode_info *ci = ceph_inode(inode);
2187 struct ceph_string *pool_ns;
2188 s64 pool;
2189 int ret, flags;
2190
2191 /* Only need to do this for regular files */
2192 if (!S_ISREG(inode->i_mode))
2193 return 0;
2194
2195 if (ci->i_vino.snap != CEPH_NOSNAP) {
2196 /*
2197 * Pool permission check needs to write to the first object.
2198 * But for snapshot, head of the first object may have already
2199 * been deleted. Skip check to avoid creating orphan object.
2200 */
2201 return 0;
2202 }
2203
2204 if (ceph_test_mount_opt(ceph_inode_to_fs_client(inode),
2205 NOPOOLPERM))
2206 return 0;
2207
2208 spin_lock(&ci->i_ceph_lock);
2209 flags = ci->i_ceph_flags;
2210 pool = ci->i_layout.pool_id;
2211 spin_unlock(&ci->i_ceph_lock);
2212check:
2213 if (flags & CEPH_I_POOL_PERM) {
2214 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2215 doutc(cl, "pool %lld no read perm\n", pool);
2216 return -EPERM;
2217 }
2218 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2219 doutc(cl, "pool %lld no write perm\n", pool);
2220 return -EPERM;
2221 }
2222 return 0;
2223 }
2224
2225 pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2226 ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2227 ceph_put_string(pool_ns);
2228 if (ret < 0)
2229 return ret;
2230
2231 flags = CEPH_I_POOL_PERM;
2232 if (ret & POOL_READ)
2233 flags |= CEPH_I_POOL_RD;
2234 if (ret & POOL_WRITE)
2235 flags |= CEPH_I_POOL_WR;
2236
2237 spin_lock(&ci->i_ceph_lock);
2238 if (pool == ci->i_layout.pool_id &&
2239 pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2240 ci->i_ceph_flags |= flags;
2241 } else {
2242 pool = ci->i_layout.pool_id;
2243 flags = ci->i_ceph_flags;
2244 }
2245 spin_unlock(&ci->i_ceph_lock);
2246 goto check;
2247}
2248
2249void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2250{
2251 struct ceph_pool_perm *perm;
2252 struct rb_node *n;
2253
2254 while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2255 n = rb_first(&mdsc->pool_perm_tree);
2256 perm = rb_entry(n, struct ceph_pool_perm, node);
2257 rb_erase(n, &mdsc->pool_perm_tree);
2258 kfree(perm);
2259 }
2260}