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
57static inline struct ceph_snap_context *page_snap_context(struct page *page)
58{
59 if (PagePrivate(page))
60 return (void *)page->private;
61 return NULL;
62}
63
64/*
65 * Dirty a page. Optimistically adjust accounting, on the assumption
66 * that we won't race with invalidate. If we do, readjust.
67 */
68static int ceph_set_page_dirty(struct page *page)
69{
70 struct address_space *mapping = page->mapping;
71 struct inode *inode;
72 struct ceph_inode_info *ci;
73 int undo = 0;
74 struct ceph_snap_context *snapc;
75
76 if (unlikely(!mapping))
77 return !TestSetPageDirty(page);
78
79 if (TestSetPageDirty(page)) {
80 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
81 mapping->host, page, page->index);
82 return 0;
83 }
84
85 inode = mapping->host;
86 ci = ceph_inode(inode);
87
88 /*
89 * Note that we're grabbing a snapc ref here without holding
90 * any locks!
91 */
92 snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
93
94 /* dirty the head */
95 spin_lock(&ci->i_ceph_lock);
96 if (ci->i_head_snapc == NULL)
97 ci->i_head_snapc = ceph_get_snap_context(snapc);
98 ++ci->i_wrbuffer_ref_head;
99 if (ci->i_wrbuffer_ref == 0)
100 ihold(inode);
101 ++ci->i_wrbuffer_ref;
102 dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
103 "snapc %p seq %lld (%d snaps)\n",
104 mapping->host, page, page->index,
105 ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
106 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
107 snapc, snapc->seq, snapc->num_snaps);
108 spin_unlock(&ci->i_ceph_lock);
109
110 /* now adjust page */
111 spin_lock_irq(&mapping->tree_lock);
112 if (page->mapping) { /* Race with truncate? */
113 WARN_ON_ONCE(!PageUptodate(page));
114 account_page_dirtied(page, page->mapping);
115 radix_tree_tag_set(&mapping->page_tree,
116 page_index(page), PAGECACHE_TAG_DIRTY);
117
118 /*
119 * Reference snap context in page->private. Also set
120 * PagePrivate so that we get invalidatepage callback.
121 */
122 page->private = (unsigned long)snapc;
123 SetPagePrivate(page);
124 } else {
125 dout("ANON set_page_dirty %p (raced truncate?)\n", page);
126 undo = 1;
127 }
128
129 spin_unlock_irq(&mapping->tree_lock);
130
131 if (undo)
132 /* whoops, we failed to dirty the page */
133 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
134
135 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
136
137 BUG_ON(!PageDirty(page));
138 return 1;
139}
140
141/*
142 * If we are truncating the full page (i.e. offset == 0), adjust the
143 * dirty page counters appropriately. Only called if there is private
144 * data on the page.
145 */
146static void ceph_invalidatepage(struct page *page, unsigned long offset)
147{
148 struct inode *inode;
149 struct ceph_inode_info *ci;
150 struct ceph_snap_context *snapc = page_snap_context(page);
151
152 BUG_ON(!PageLocked(page));
153 BUG_ON(!PagePrivate(page));
154 BUG_ON(!page->mapping);
155
156 inode = page->mapping->host;
157
158 /*
159 * We can get non-dirty pages here due to races between
160 * set_page_dirty and truncate_complete_page; just spit out a
161 * warning, in case we end up with accounting problems later.
162 */
163 if (!PageDirty(page))
164 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
165
166 if (offset == 0)
167 ClearPageChecked(page);
168
169 ci = ceph_inode(inode);
170 if (offset == 0) {
171 dout("%p invalidatepage %p idx %lu full dirty page %lu\n",
172 inode, page, page->index, offset);
173 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
174 ceph_put_snap_context(snapc);
175 page->private = 0;
176 ClearPagePrivate(page);
177 } else {
178 dout("%p invalidatepage %p idx %lu partial dirty page\n",
179 inode, page, page->index);
180 }
181}
182
183/* just a sanity check */
184static int ceph_releasepage(struct page *page, gfp_t g)
185{
186 struct inode *inode = page->mapping ? page->mapping->host : NULL;
187 dout("%p releasepage %p idx %lu\n", inode, page, page->index);
188 WARN_ON(PageDirty(page));
189 WARN_ON(PagePrivate(page));
190 return 0;
191}
192
193/*
194 * read a single page, without unlocking it.
195 */
196static int readpage_nounlock(struct file *filp, struct page *page)
197{
198 struct inode *inode = filp->f_dentry->d_inode;
199 struct ceph_inode_info *ci = ceph_inode(inode);
200 struct ceph_osd_client *osdc =
201 &ceph_inode_to_client(inode)->client->osdc;
202 int err = 0;
203 u64 len = PAGE_CACHE_SIZE;
204
205 dout("readpage inode %p file %p page %p index %lu\n",
206 inode, filp, page, page->index);
207 err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
208 page->index << PAGE_CACHE_SHIFT, &len,
209 ci->i_truncate_seq, ci->i_truncate_size,
210 &page, 1, 0);
211 if (err == -ENOENT)
212 err = 0;
213 if (err < 0) {
214 SetPageError(page);
215 goto out;
216 } else if (err < PAGE_CACHE_SIZE) {
217 /* zero fill remainder of page */
218 zero_user_segment(page, err, PAGE_CACHE_SIZE);
219 }
220 SetPageUptodate(page);
221
222out:
223 return err < 0 ? err : 0;
224}
225
226static int ceph_readpage(struct file *filp, struct page *page)
227{
228 int r = readpage_nounlock(filp, page);
229 unlock_page(page);
230 return r;
231}
232
233/*
234 * Finish an async read(ahead) op.
235 */
236static void finish_read(struct ceph_osd_request *req, struct ceph_msg *msg)
237{
238 struct inode *inode = req->r_inode;
239 struct ceph_osd_reply_head *replyhead;
240 int rc, bytes;
241 int i;
242
243 /* parse reply */
244 replyhead = msg->front.iov_base;
245 WARN_ON(le32_to_cpu(replyhead->num_ops) == 0);
246 rc = le32_to_cpu(replyhead->result);
247 bytes = le32_to_cpu(msg->hdr.data_len);
248
249 dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
250
251 /* unlock all pages, zeroing any data we didn't read */
252 for (i = 0; i < req->r_num_pages; i++, bytes -= PAGE_CACHE_SIZE) {
253 struct page *page = req->r_pages[i];
254
255 if (bytes < (int)PAGE_CACHE_SIZE) {
256 /* zero (remainder of) page */
257 int s = bytes < 0 ? 0 : bytes;
258 zero_user_segment(page, s, PAGE_CACHE_SIZE);
259 }
260 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
261 page->index);
262 flush_dcache_page(page);
263 SetPageUptodate(page);
264 unlock_page(page);
265 page_cache_release(page);
266 }
267 kfree(req->r_pages);
268}
269
270/*
271 * start an async read(ahead) operation. return nr_pages we submitted
272 * a read for on success, or negative error code.
273 */
274static int start_read(struct inode *inode, struct list_head *page_list, int max)
275{
276 struct ceph_osd_client *osdc =
277 &ceph_inode_to_client(inode)->client->osdc;
278 struct ceph_inode_info *ci = ceph_inode(inode);
279 struct page *page = list_entry(page_list->prev, struct page, lru);
280 struct ceph_osd_request *req;
281 u64 off;
282 u64 len;
283 int i;
284 struct page **pages;
285 pgoff_t next_index;
286 int nr_pages = 0;
287 int ret;
288
289 off = page->index << PAGE_CACHE_SHIFT;
290
291 /* count pages */
292 next_index = page->index;
293 list_for_each_entry_reverse(page, page_list, lru) {
294 if (page->index != next_index)
295 break;
296 nr_pages++;
297 next_index++;
298 if (max && nr_pages == max)
299 break;
300 }
301 len = nr_pages << PAGE_CACHE_SHIFT;
302 dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
303 off, len);
304
305 req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode),
306 off, &len,
307 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
308 NULL, 0,
309 ci->i_truncate_seq, ci->i_truncate_size,
310 NULL, false, 1, 0);
311 if (!req)
312 return -ENOMEM;
313
314 /* build page vector */
315 nr_pages = len >> PAGE_CACHE_SHIFT;
316 pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS);
317 ret = -ENOMEM;
318 if (!pages)
319 goto out;
320 for (i = 0; i < nr_pages; ++i) {
321 page = list_entry(page_list->prev, struct page, lru);
322 BUG_ON(PageLocked(page));
323 list_del(&page->lru);
324
325 dout("start_read %p adding %p idx %lu\n", inode, page,
326 page->index);
327 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
328 GFP_NOFS)) {
329 page_cache_release(page);
330 dout("start_read %p add_to_page_cache failed %p\n",
331 inode, page);
332 nr_pages = i;
333 goto out_pages;
334 }
335 pages[i] = page;
336 }
337 req->r_pages = pages;
338 req->r_num_pages = nr_pages;
339 req->r_callback = finish_read;
340 req->r_inode = inode;
341
342 dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
343 ret = ceph_osdc_start_request(osdc, req, false);
344 if (ret < 0)
345 goto out_pages;
346 ceph_osdc_put_request(req);
347 return nr_pages;
348
349out_pages:
350 ceph_release_page_vector(pages, nr_pages);
351out:
352 ceph_osdc_put_request(req);
353 return ret;
354}
355
356
357/*
358 * Read multiple pages. Leave pages we don't read + unlock in page_list;
359 * the caller (VM) cleans them up.
360 */
361static int ceph_readpages(struct file *file, struct address_space *mapping,
362 struct list_head *page_list, unsigned nr_pages)
363{
364 struct inode *inode = file->f_dentry->d_inode;
365 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
366 int rc = 0;
367 int max = 0;
368
369 if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE)
370 max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1)
371 >> PAGE_SHIFT;
372
373 dout("readpages %p file %p nr_pages %d max %d\n", inode, file, nr_pages,
374 max);
375 while (!list_empty(page_list)) {
376 rc = start_read(inode, page_list, max);
377 if (rc < 0)
378 goto out;
379 BUG_ON(rc == 0);
380 }
381out:
382 dout("readpages %p file %p ret %d\n", inode, file, rc);
383 return rc;
384}
385
386/*
387 * Get ref for the oldest snapc for an inode with dirty data... that is, the
388 * only snap context we are allowed to write back.
389 */
390static struct ceph_snap_context *get_oldest_context(struct inode *inode,
391 u64 *snap_size)
392{
393 struct ceph_inode_info *ci = ceph_inode(inode);
394 struct ceph_snap_context *snapc = NULL;
395 struct ceph_cap_snap *capsnap = NULL;
396
397 spin_lock(&ci->i_ceph_lock);
398 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
399 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
400 capsnap->context, capsnap->dirty_pages);
401 if (capsnap->dirty_pages) {
402 snapc = ceph_get_snap_context(capsnap->context);
403 if (snap_size)
404 *snap_size = capsnap->size;
405 break;
406 }
407 }
408 if (!snapc && ci->i_wrbuffer_ref_head) {
409 snapc = ceph_get_snap_context(ci->i_head_snapc);
410 dout(" head snapc %p has %d dirty pages\n",
411 snapc, ci->i_wrbuffer_ref_head);
412 }
413 spin_unlock(&ci->i_ceph_lock);
414 return snapc;
415}
416
417/*
418 * Write a single page, but leave the page locked.
419 *
420 * If we get a write error, set the page error bit, but still adjust the
421 * dirty page accounting (i.e., page is no longer dirty).
422 */
423static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
424{
425 struct inode *inode;
426 struct ceph_inode_info *ci;
427 struct ceph_fs_client *fsc;
428 struct ceph_osd_client *osdc;
429 loff_t page_off = page->index << PAGE_CACHE_SHIFT;
430 int len = PAGE_CACHE_SIZE;
431 loff_t i_size;
432 int err = 0;
433 struct ceph_snap_context *snapc, *oldest;
434 u64 snap_size = 0;
435 long writeback_stat;
436
437 dout("writepage %p idx %lu\n", page, page->index);
438
439 if (!page->mapping || !page->mapping->host) {
440 dout("writepage %p - no mapping\n", page);
441 return -EFAULT;
442 }
443 inode = page->mapping->host;
444 ci = ceph_inode(inode);
445 fsc = ceph_inode_to_client(inode);
446 osdc = &fsc->client->osdc;
447
448 /* verify this is a writeable snap context */
449 snapc = page_snap_context(page);
450 if (snapc == NULL) {
451 dout("writepage %p page %p not dirty?\n", inode, page);
452 goto out;
453 }
454 oldest = get_oldest_context(inode, &snap_size);
455 if (snapc->seq > oldest->seq) {
456 dout("writepage %p page %p snapc %p not writeable - noop\n",
457 inode, page, snapc);
458 /* we should only noop if called by kswapd */
459 WARN_ON((current->flags & PF_MEMALLOC) == 0);
460 ceph_put_snap_context(oldest);
461 goto out;
462 }
463 ceph_put_snap_context(oldest);
464
465 /* is this a partial page at end of file? */
466 if (snap_size)
467 i_size = snap_size;
468 else
469 i_size = i_size_read(inode);
470 if (i_size < page_off + len)
471 len = i_size - page_off;
472
473 dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
474 inode, page, page->index, page_off, len, snapc);
475
476 writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
477 if (writeback_stat >
478 CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
479 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
480
481 set_page_writeback(page);
482 err = ceph_osdc_writepages(osdc, ceph_vino(inode),
483 &ci->i_layout, snapc,
484 page_off, len,
485 ci->i_truncate_seq, ci->i_truncate_size,
486 &inode->i_mtime,
487 &page, 1, 0, 0, true);
488 if (err < 0) {
489 dout("writepage setting page/mapping error %d %p\n", err, page);
490 SetPageError(page);
491 mapping_set_error(&inode->i_data, err);
492 if (wbc)
493 wbc->pages_skipped++;
494 } else {
495 dout("writepage cleaned page %p\n", page);
496 err = 0; /* vfs expects us to return 0 */
497 }
498 page->private = 0;
499 ClearPagePrivate(page);
500 end_page_writeback(page);
501 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
502 ceph_put_snap_context(snapc); /* page's reference */
503out:
504 return err;
505}
506
507static int ceph_writepage(struct page *page, struct writeback_control *wbc)
508{
509 int err;
510 struct inode *inode = page->mapping->host;
511 BUG_ON(!inode);
512 ihold(inode);
513 err = writepage_nounlock(page, wbc);
514 unlock_page(page);
515 iput(inode);
516 return err;
517}
518
519
520/*
521 * lame release_pages helper. release_pages() isn't exported to
522 * modules.
523 */
524static void ceph_release_pages(struct page **pages, int num)
525{
526 struct pagevec pvec;
527 int i;
528
529 pagevec_init(&pvec, 0);
530 for (i = 0; i < num; i++) {
531 if (pagevec_add(&pvec, pages[i]) == 0)
532 pagevec_release(&pvec);
533 }
534 pagevec_release(&pvec);
535}
536
537
538/*
539 * async writeback completion handler.
540 *
541 * If we get an error, set the mapping error bit, but not the individual
542 * page error bits.
543 */
544static void writepages_finish(struct ceph_osd_request *req,
545 struct ceph_msg *msg)
546{
547 struct inode *inode = req->r_inode;
548 struct ceph_osd_reply_head *replyhead;
549 struct ceph_osd_op *op;
550 struct ceph_inode_info *ci = ceph_inode(inode);
551 unsigned wrote;
552 struct page *page;
553 int i;
554 struct ceph_snap_context *snapc = req->r_snapc;
555 struct address_space *mapping = inode->i_mapping;
556 __s32 rc = -EIO;
557 u64 bytes = 0;
558 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
559 long writeback_stat;
560 unsigned issued = ceph_caps_issued(ci);
561
562 /* parse reply */
563 replyhead = msg->front.iov_base;
564 WARN_ON(le32_to_cpu(replyhead->num_ops) == 0);
565 op = (void *)(replyhead + 1);
566 rc = le32_to_cpu(replyhead->result);
567 bytes = le64_to_cpu(op->extent.length);
568
569 if (rc >= 0) {
570 /*
571 * Assume we wrote the pages we originally sent. The
572 * osd might reply with fewer pages if our writeback
573 * raced with a truncation and was adjusted at the osd,
574 * so don't believe the reply.
575 */
576 wrote = req->r_num_pages;
577 } else {
578 wrote = 0;
579 mapping_set_error(mapping, rc);
580 }
581 dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
582 inode, rc, bytes, wrote);
583
584 /* clean all pages */
585 for (i = 0; i < req->r_num_pages; i++) {
586 page = req->r_pages[i];
587 BUG_ON(!page);
588 WARN_ON(!PageUptodate(page));
589
590 writeback_stat =
591 atomic_long_dec_return(&fsc->writeback_count);
592 if (writeback_stat <
593 CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
594 clear_bdi_congested(&fsc->backing_dev_info,
595 BLK_RW_ASYNC);
596
597 ceph_put_snap_context(page_snap_context(page));
598 page->private = 0;
599 ClearPagePrivate(page);
600 dout("unlocking %d %p\n", i, page);
601 end_page_writeback(page);
602
603 /*
604 * We lost the cache cap, need to truncate the page before
605 * it is unlocked, otherwise we'd truncate it later in the
606 * page truncation thread, possibly losing some data that
607 * raced its way in
608 */
609 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
610 generic_error_remove_page(inode->i_mapping, page);
611
612 unlock_page(page);
613 }
614 dout("%p wrote+cleaned %d pages\n", inode, wrote);
615 ceph_put_wrbuffer_cap_refs(ci, req->r_num_pages, snapc);
616
617 ceph_release_pages(req->r_pages, req->r_num_pages);
618 if (req->r_pages_from_pool)
619 mempool_free(req->r_pages,
620 ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
621 else
622 kfree(req->r_pages);
623 ceph_osdc_put_request(req);
624}
625
626/*
627 * allocate a page vec, either directly, or if necessary, via a the
628 * mempool. we avoid the mempool if we can because req->r_num_pages
629 * may be less than the maximum write size.
630 */
631static void alloc_page_vec(struct ceph_fs_client *fsc,
632 struct ceph_osd_request *req)
633{
634 req->r_pages = kmalloc(sizeof(struct page *) * req->r_num_pages,
635 GFP_NOFS);
636 if (!req->r_pages) {
637 req->r_pages = mempool_alloc(fsc->wb_pagevec_pool, GFP_NOFS);
638 req->r_pages_from_pool = 1;
639 WARN_ON(!req->r_pages);
640 }
641}
642
643/*
644 * initiate async writeback
645 */
646static int ceph_writepages_start(struct address_space *mapping,
647 struct writeback_control *wbc)
648{
649 struct inode *inode = mapping->host;
650 struct ceph_inode_info *ci = ceph_inode(inode);
651 struct ceph_fs_client *fsc;
652 pgoff_t index, start, end;
653 int range_whole = 0;
654 int should_loop = 1;
655 pgoff_t max_pages = 0, max_pages_ever = 0;
656 struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
657 struct pagevec pvec;
658 int done = 0;
659 int rc = 0;
660 unsigned wsize = 1 << inode->i_blkbits;
661 struct ceph_osd_request *req = NULL;
662 int do_sync;
663 u64 snap_size = 0;
664
665 /*
666 * Include a 'sync' in the OSD request if this is a data
667 * integrity write (e.g., O_SYNC write or fsync()), or if our
668 * cap is being revoked.
669 */
670 do_sync = wbc->sync_mode == WB_SYNC_ALL;
671 if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
672 do_sync = 1;
673 dout("writepages_start %p dosync=%d (mode=%s)\n",
674 inode, do_sync,
675 wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
676 (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
677
678 fsc = ceph_inode_to_client(inode);
679 if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
680 pr_warning("writepage_start %p on forced umount\n", inode);
681 return -EIO; /* we're in a forced umount, don't write! */
682 }
683 if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
684 wsize = fsc->mount_options->wsize;
685 if (wsize < PAGE_CACHE_SIZE)
686 wsize = PAGE_CACHE_SIZE;
687 max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
688
689 pagevec_init(&pvec, 0);
690
691 /* where to start/end? */
692 if (wbc->range_cyclic) {
693 start = mapping->writeback_index; /* Start from prev offset */
694 end = -1;
695 dout(" cyclic, start at %lu\n", start);
696 } else {
697 start = wbc->range_start >> PAGE_CACHE_SHIFT;
698 end = wbc->range_end >> PAGE_CACHE_SHIFT;
699 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
700 range_whole = 1;
701 should_loop = 0;
702 dout(" not cyclic, %lu to %lu\n", start, end);
703 }
704 index = start;
705
706retry:
707 /* find oldest snap context with dirty data */
708 ceph_put_snap_context(snapc);
709 snapc = get_oldest_context(inode, &snap_size);
710 if (!snapc) {
711 /* hmm, why does writepages get called when there
712 is no dirty data? */
713 dout(" no snap context with dirty data?\n");
714 goto out;
715 }
716 dout(" oldest snapc is %p seq %lld (%d snaps)\n",
717 snapc, snapc->seq, snapc->num_snaps);
718 if (last_snapc && snapc != last_snapc) {
719 /* if we switched to a newer snapc, restart our scan at the
720 * start of the original file range. */
721 dout(" snapc differs from last pass, restarting at %lu\n",
722 index);
723 index = start;
724 }
725 last_snapc = snapc;
726
727 while (!done && index <= end) {
728 unsigned i;
729 int first;
730 pgoff_t next;
731 int pvec_pages, locked_pages;
732 struct page *page;
733 int want;
734 u64 offset, len;
735 struct ceph_osd_request_head *reqhead;
736 struct ceph_osd_op *op;
737 long writeback_stat;
738
739 next = 0;
740 locked_pages = 0;
741 max_pages = max_pages_ever;
742
743get_more_pages:
744 first = -1;
745 want = min(end - index,
746 min((pgoff_t)PAGEVEC_SIZE,
747 max_pages - (pgoff_t)locked_pages) - 1)
748 + 1;
749 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
750 PAGECACHE_TAG_DIRTY,
751 want);
752 dout("pagevec_lookup_tag got %d\n", pvec_pages);
753 if (!pvec_pages && !locked_pages)
754 break;
755 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
756 page = pvec.pages[i];
757 dout("? %p idx %lu\n", page, page->index);
758 if (locked_pages == 0)
759 lock_page(page); /* first page */
760 else if (!trylock_page(page))
761 break;
762
763 /* only dirty pages, or our accounting breaks */
764 if (unlikely(!PageDirty(page)) ||
765 unlikely(page->mapping != mapping)) {
766 dout("!dirty or !mapping %p\n", page);
767 unlock_page(page);
768 break;
769 }
770 if (!wbc->range_cyclic && page->index > end) {
771 dout("end of range %p\n", page);
772 done = 1;
773 unlock_page(page);
774 break;
775 }
776 if (next && (page->index != next)) {
777 dout("not consecutive %p\n", page);
778 unlock_page(page);
779 break;
780 }
781 if (wbc->sync_mode != WB_SYNC_NONE) {
782 dout("waiting on writeback %p\n", page);
783 wait_on_page_writeback(page);
784 }
785 if ((snap_size && page_offset(page) > snap_size) ||
786 (!snap_size &&
787 page_offset(page) > i_size_read(inode))) {
788 dout("%p page eof %llu\n", page, snap_size ?
789 snap_size : i_size_read(inode));
790 done = 1;
791 unlock_page(page);
792 break;
793 }
794 if (PageWriteback(page)) {
795 dout("%p under writeback\n", page);
796 unlock_page(page);
797 break;
798 }
799
800 /* only if matching snap context */
801 pgsnapc = page_snap_context(page);
802 if (pgsnapc->seq > snapc->seq) {
803 dout("page snapc %p %lld > oldest %p %lld\n",
804 pgsnapc, pgsnapc->seq, snapc, snapc->seq);
805 unlock_page(page);
806 if (!locked_pages)
807 continue; /* keep looking for snap */
808 break;
809 }
810
811 if (!clear_page_dirty_for_io(page)) {
812 dout("%p !clear_page_dirty_for_io\n", page);
813 unlock_page(page);
814 break;
815 }
816
817 /* ok */
818 if (locked_pages == 0) {
819 /* prepare async write request */
820 offset = (unsigned long long)page->index
821 << PAGE_CACHE_SHIFT;
822 len = wsize;
823 req = ceph_osdc_new_request(&fsc->client->osdc,
824 &ci->i_layout,
825 ceph_vino(inode),
826 offset, &len,
827 CEPH_OSD_OP_WRITE,
828 CEPH_OSD_FLAG_WRITE |
829 CEPH_OSD_FLAG_ONDISK,
830 snapc, do_sync,
831 ci->i_truncate_seq,
832 ci->i_truncate_size,
833 &inode->i_mtime, true, 1, 0);
834
835 if (!req) {
836 rc = -ENOMEM;
837 unlock_page(page);
838 break;
839 }
840
841 max_pages = req->r_num_pages;
842
843 alloc_page_vec(fsc, req);
844 req->r_callback = writepages_finish;
845 req->r_inode = inode;
846 }
847
848 /* note position of first page in pvec */
849 if (first < 0)
850 first = i;
851 dout("%p will write page %p idx %lu\n",
852 inode, page, page->index);
853
854 writeback_stat =
855 atomic_long_inc_return(&fsc->writeback_count);
856 if (writeback_stat > CONGESTION_ON_THRESH(
857 fsc->mount_options->congestion_kb)) {
858 set_bdi_congested(&fsc->backing_dev_info,
859 BLK_RW_ASYNC);
860 }
861
862 set_page_writeback(page);
863 req->r_pages[locked_pages] = page;
864 locked_pages++;
865 next = page->index + 1;
866 }
867
868 /* did we get anything? */
869 if (!locked_pages)
870 goto release_pvec_pages;
871 if (i) {
872 int j;
873 BUG_ON(!locked_pages || first < 0);
874
875 if (pvec_pages && i == pvec_pages &&
876 locked_pages < max_pages) {
877 dout("reached end pvec, trying for more\n");
878 pagevec_reinit(&pvec);
879 goto get_more_pages;
880 }
881
882 /* shift unused pages over in the pvec... we
883 * will need to release them below. */
884 for (j = i; j < pvec_pages; j++) {
885 dout(" pvec leftover page %p\n",
886 pvec.pages[j]);
887 pvec.pages[j-i+first] = pvec.pages[j];
888 }
889 pvec.nr -= i-first;
890 }
891
892 /* submit the write */
893 offset = req->r_pages[0]->index << PAGE_CACHE_SHIFT;
894 len = min((snap_size ? snap_size : i_size_read(inode)) - offset,
895 (u64)locked_pages << PAGE_CACHE_SHIFT);
896 dout("writepages got %d pages at %llu~%llu\n",
897 locked_pages, offset, len);
898
899 /* revise final length, page count */
900 req->r_num_pages = locked_pages;
901 reqhead = req->r_request->front.iov_base;
902 op = (void *)(reqhead + 1);
903 op->extent.length = cpu_to_le64(len);
904 op->payload_len = cpu_to_le32(len);
905 req->r_request->hdr.data_len = cpu_to_le32(len);
906
907 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
908 BUG_ON(rc);
909 req = NULL;
910
911 /* continue? */
912 index = next;
913 wbc->nr_to_write -= locked_pages;
914 if (wbc->nr_to_write <= 0)
915 done = 1;
916
917release_pvec_pages:
918 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
919 pvec.nr ? pvec.pages[0] : NULL);
920 pagevec_release(&pvec);
921
922 if (locked_pages && !done)
923 goto retry;
924 }
925
926 if (should_loop && !done) {
927 /* more to do; loop back to beginning of file */
928 dout("writepages looping back to beginning of file\n");
929 should_loop = 0;
930 index = 0;
931 goto retry;
932 }
933
934 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
935 mapping->writeback_index = index;
936
937out:
938 if (req)
939 ceph_osdc_put_request(req);
940 ceph_put_snap_context(snapc);
941 dout("writepages done, rc = %d\n", rc);
942 return rc;
943}
944
945
946
947/*
948 * See if a given @snapc is either writeable, or already written.
949 */
950static int context_is_writeable_or_written(struct inode *inode,
951 struct ceph_snap_context *snapc)
952{
953 struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
954 int ret = !oldest || snapc->seq <= oldest->seq;
955
956 ceph_put_snap_context(oldest);
957 return ret;
958}
959
960/*
961 * We are only allowed to write into/dirty the page if the page is
962 * clean, or already dirty within the same snap context.
963 *
964 * called with page locked.
965 * return success with page locked,
966 * or any failure (incl -EAGAIN) with page unlocked.
967 */
968static int ceph_update_writeable_page(struct file *file,
969 loff_t pos, unsigned len,
970 struct page *page)
971{
972 struct inode *inode = file->f_dentry->d_inode;
973 struct ceph_inode_info *ci = ceph_inode(inode);
974 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
975 loff_t page_off = pos & PAGE_CACHE_MASK;
976 int pos_in_page = pos & ~PAGE_CACHE_MASK;
977 int end_in_page = pos_in_page + len;
978 loff_t i_size;
979 int r;
980 struct ceph_snap_context *snapc, *oldest;
981
982retry_locked:
983 /* writepages currently holds page lock, but if we change that later, */
984 wait_on_page_writeback(page);
985
986 /* check snap context */
987 BUG_ON(!ci->i_snap_realm);
988 down_read(&mdsc->snap_rwsem);
989 BUG_ON(!ci->i_snap_realm->cached_context);
990 snapc = page_snap_context(page);
991 if (snapc && snapc != ci->i_head_snapc) {
992 /*
993 * this page is already dirty in another (older) snap
994 * context! is it writeable now?
995 */
996 oldest = get_oldest_context(inode, NULL);
997 up_read(&mdsc->snap_rwsem);
998
999 if (snapc->seq > oldest->seq) {
1000 ceph_put_snap_context(oldest);
1001 dout(" page %p snapc %p not current or oldest\n",
1002 page, snapc);
1003 /*
1004 * queue for writeback, and wait for snapc to
1005 * be writeable or written
1006 */
1007 snapc = ceph_get_snap_context(snapc);
1008 unlock_page(page);
1009 ceph_queue_writeback(inode);
1010 r = wait_event_interruptible(ci->i_cap_wq,
1011 context_is_writeable_or_written(inode, snapc));
1012 ceph_put_snap_context(snapc);
1013 if (r == -ERESTARTSYS)
1014 return r;
1015 return -EAGAIN;
1016 }
1017 ceph_put_snap_context(oldest);
1018
1019 /* yay, writeable, do it now (without dropping page lock) */
1020 dout(" page %p snapc %p not current, but oldest\n",
1021 page, snapc);
1022 if (!clear_page_dirty_for_io(page))
1023 goto retry_locked;
1024 r = writepage_nounlock(page, NULL);
1025 if (r < 0)
1026 goto fail_nosnap;
1027 goto retry_locked;
1028 }
1029
1030 if (PageUptodate(page)) {
1031 dout(" page %p already uptodate\n", page);
1032 return 0;
1033 }
1034
1035 /* full page? */
1036 if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
1037 return 0;
1038
1039 /* past end of file? */
1040 i_size = inode->i_size; /* caller holds i_mutex */
1041
1042 if (i_size + len > inode->i_sb->s_maxbytes) {
1043 /* file is too big */
1044 r = -EINVAL;
1045 goto fail;
1046 }
1047
1048 if (page_off >= i_size ||
1049 (pos_in_page == 0 && (pos+len) >= i_size &&
1050 end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1051 dout(" zeroing %p 0 - %d and %d - %d\n",
1052 page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1053 zero_user_segments(page,
1054 0, pos_in_page,
1055 end_in_page, PAGE_CACHE_SIZE);
1056 return 0;
1057 }
1058
1059 /* we need to read it. */
1060 up_read(&mdsc->snap_rwsem);
1061 r = readpage_nounlock(file, page);
1062 if (r < 0)
1063 goto fail_nosnap;
1064 goto retry_locked;
1065
1066fail:
1067 up_read(&mdsc->snap_rwsem);
1068fail_nosnap:
1069 unlock_page(page);
1070 return r;
1071}
1072
1073/*
1074 * We are only allowed to write into/dirty the page if the page is
1075 * clean, or already dirty within the same snap context.
1076 */
1077static int ceph_write_begin(struct file *file, struct address_space *mapping,
1078 loff_t pos, unsigned len, unsigned flags,
1079 struct page **pagep, void **fsdata)
1080{
1081 struct inode *inode = file->f_dentry->d_inode;
1082 struct page *page;
1083 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1084 int r;
1085
1086 do {
1087 /* get a page */
1088 page = grab_cache_page_write_begin(mapping, index, 0);
1089 if (!page)
1090 return -ENOMEM;
1091 *pagep = page;
1092
1093 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1094 inode, page, (int)pos, (int)len);
1095
1096 r = ceph_update_writeable_page(file, pos, len, page);
1097 } while (r == -EAGAIN);
1098
1099 return r;
1100}
1101
1102/*
1103 * we don't do anything in here that simple_write_end doesn't do
1104 * except adjust dirty page accounting and drop read lock on
1105 * mdsc->snap_rwsem.
1106 */
1107static int ceph_write_end(struct file *file, struct address_space *mapping,
1108 loff_t pos, unsigned len, unsigned copied,
1109 struct page *page, void *fsdata)
1110{
1111 struct inode *inode = file->f_dentry->d_inode;
1112 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1113 struct ceph_mds_client *mdsc = fsc->mdsc;
1114 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1115 int check_cap = 0;
1116
1117 dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1118 inode, page, (int)pos, (int)copied, (int)len);
1119
1120 /* zero the stale part of the page if we did a short copy */
1121 if (copied < len)
1122 zero_user_segment(page, from+copied, len);
1123
1124 /* did file size increase? */
1125 /* (no need for i_size_read(); we caller holds i_mutex */
1126 if (pos+copied > inode->i_size)
1127 check_cap = ceph_inode_set_size(inode, pos+copied);
1128
1129 if (!PageUptodate(page))
1130 SetPageUptodate(page);
1131
1132 set_page_dirty(page);
1133
1134 unlock_page(page);
1135 up_read(&mdsc->snap_rwsem);
1136 page_cache_release(page);
1137
1138 if (check_cap)
1139 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1140
1141 return copied;
1142}
1143
1144/*
1145 * we set .direct_IO to indicate direct io is supported, but since we
1146 * intercept O_DIRECT reads and writes early, this function should
1147 * never get called.
1148 */
1149static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1150 const struct iovec *iov,
1151 loff_t pos, unsigned long nr_segs)
1152{
1153 WARN_ON(1);
1154 return -EINVAL;
1155}
1156
1157const struct address_space_operations ceph_aops = {
1158 .readpage = ceph_readpage,
1159 .readpages = ceph_readpages,
1160 .writepage = ceph_writepage,
1161 .writepages = ceph_writepages_start,
1162 .write_begin = ceph_write_begin,
1163 .write_end = ceph_write_end,
1164 .set_page_dirty = ceph_set_page_dirty,
1165 .invalidatepage = ceph_invalidatepage,
1166 .releasepage = ceph_releasepage,
1167 .direct_IO = ceph_direct_io,
1168};
1169
1170
1171/*
1172 * vm ops
1173 */
1174
1175/*
1176 * Reuse write_begin here for simplicity.
1177 */
1178static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1179{
1180 struct inode *inode = vma->vm_file->f_dentry->d_inode;
1181 struct page *page = vmf->page;
1182 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1183 loff_t off = page->index << PAGE_CACHE_SHIFT;
1184 loff_t size, len;
1185 int ret;
1186
1187 size = i_size_read(inode);
1188 if (off + PAGE_CACHE_SIZE <= size)
1189 len = PAGE_CACHE_SIZE;
1190 else
1191 len = size & ~PAGE_CACHE_MASK;
1192
1193 dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1194 off, len, page, page->index);
1195
1196 lock_page(page);
1197
1198 ret = VM_FAULT_NOPAGE;
1199 if ((off > size) ||
1200 (page->mapping != inode->i_mapping))
1201 goto out;
1202
1203 ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1204 if (ret == 0) {
1205 /* success. we'll keep the page locked. */
1206 set_page_dirty(page);
1207 up_read(&mdsc->snap_rwsem);
1208 ret = VM_FAULT_LOCKED;
1209 } else {
1210 if (ret == -ENOMEM)
1211 ret = VM_FAULT_OOM;
1212 else
1213 ret = VM_FAULT_SIGBUS;
1214 }
1215out:
1216 dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1217 if (ret != VM_FAULT_LOCKED)
1218 unlock_page(page);
1219 return ret;
1220}
1221
1222static struct vm_operations_struct ceph_vmops = {
1223 .fault = filemap_fault,
1224 .page_mkwrite = ceph_page_mkwrite,
1225};
1226
1227int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1228{
1229 struct address_space *mapping = file->f_mapping;
1230
1231 if (!mapping->a_ops->readpage)
1232 return -ENOEXEC;
1233 file_accessed(file);
1234 vma->vm_ops = &ceph_vmops;
1235 vma->vm_flags |= VM_CAN_NONLINEAR;
1236 return 0;
1237}
1// SPDX-License-Identifier: GPL-2.0
2#include <linux/ceph/ceph_debug.h>
3
4#include <linux/backing-dev.h>
5#include <linux/fs.h>
6#include <linux/mm.h>
7#include <linux/swap.h>
8#include <linux/pagemap.h>
9#include <linux/slab.h>
10#include <linux/pagevec.h>
11#include <linux/task_io_accounting_ops.h>
12#include <linux/signal.h>
13#include <linux/iversion.h>
14#include <linux/ktime.h>
15#include <linux/netfs.h>
16
17#include "super.h"
18#include "mds_client.h"
19#include "cache.h"
20#include "metric.h"
21#include "crypto.h"
22#include <linux/ceph/osd_client.h>
23#include <linux/ceph/striper.h>
24
25/*
26 * Ceph address space ops.
27 *
28 * There are a few funny things going on here.
29 *
30 * The page->private field is used to reference a struct
31 * ceph_snap_context for _every_ dirty page. This indicates which
32 * snapshot the page was logically dirtied in, and thus which snap
33 * context needs to be associated with the osd write during writeback.
34 *
35 * Similarly, struct ceph_inode_info maintains a set of counters to
36 * count dirty pages on the inode. In the absence of snapshots,
37 * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
38 *
39 * When a snapshot is taken (that is, when the client receives
40 * notification that a snapshot was taken), each inode with caps and
41 * with dirty pages (dirty pages implies there is a cap) gets a new
42 * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
43 * order, new snaps go to the tail). The i_wrbuffer_ref_head count is
44 * moved to capsnap->dirty. (Unless a sync write is currently in
45 * progress. In that case, the capsnap is said to be "pending", new
46 * writes cannot start, and the capsnap isn't "finalized" until the
47 * write completes (or fails) and a final size/mtime for the inode for
48 * that snap can be settled upon.) i_wrbuffer_ref_head is reset to 0.
49 *
50 * On writeback, we must submit writes to the osd IN SNAP ORDER. So,
51 * we look for the first capsnap in i_cap_snaps and write out pages in
52 * that snap context _only_. Then we move on to the next capsnap,
53 * eventually reaching the "live" or "head" context (i.e., pages that
54 * are not yet snapped) and are writing the most recently dirtied
55 * pages.
56 *
57 * Invalidate and so forth must take care to ensure the dirty page
58 * accounting is preserved.
59 */
60
61#define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
62#define CONGESTION_OFF_THRESH(congestion_kb) \
63 (CONGESTION_ON_THRESH(congestion_kb) - \
64 (CONGESTION_ON_THRESH(congestion_kb) >> 2))
65
66static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
67 struct folio **foliop, void **_fsdata);
68
69static inline struct ceph_snap_context *page_snap_context(struct page *page)
70{
71 if (PagePrivate(page))
72 return (void *)page->private;
73 return NULL;
74}
75
76/*
77 * Dirty a page. Optimistically adjust accounting, on the assumption
78 * that we won't race with invalidate. If we do, readjust.
79 */
80static bool ceph_dirty_folio(struct address_space *mapping, struct folio *folio)
81{
82 struct inode *inode = mapping->host;
83 struct ceph_client *cl = ceph_inode_to_client(inode);
84 struct ceph_inode_info *ci;
85 struct ceph_snap_context *snapc;
86
87 if (folio_test_dirty(folio)) {
88 doutc(cl, "%llx.%llx %p idx %lu -- already dirty\n",
89 ceph_vinop(inode), folio, folio->index);
90 VM_BUG_ON_FOLIO(!folio_test_private(folio), folio);
91 return false;
92 }
93
94 ci = ceph_inode(inode);
95
96 /* dirty the head */
97 spin_lock(&ci->i_ceph_lock);
98 BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
99 if (__ceph_have_pending_cap_snap(ci)) {
100 struct ceph_cap_snap *capsnap =
101 list_last_entry(&ci->i_cap_snaps,
102 struct ceph_cap_snap,
103 ci_item);
104 snapc = ceph_get_snap_context(capsnap->context);
105 capsnap->dirty_pages++;
106 } else {
107 BUG_ON(!ci->i_head_snapc);
108 snapc = ceph_get_snap_context(ci->i_head_snapc);
109 ++ci->i_wrbuffer_ref_head;
110 }
111 if (ci->i_wrbuffer_ref == 0)
112 ihold(inode);
113 ++ci->i_wrbuffer_ref;
114 doutc(cl, "%llx.%llx %p idx %lu head %d/%d -> %d/%d "
115 "snapc %p seq %lld (%d snaps)\n",
116 ceph_vinop(inode), folio, folio->index,
117 ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
118 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
119 snapc, snapc->seq, snapc->num_snaps);
120 spin_unlock(&ci->i_ceph_lock);
121
122 /*
123 * Reference snap context in folio->private. Also set
124 * PagePrivate so that we get invalidate_folio callback.
125 */
126 VM_WARN_ON_FOLIO(folio->private, folio);
127 folio_attach_private(folio, snapc);
128
129 return ceph_fscache_dirty_folio(mapping, folio);
130}
131
132/*
133 * If we are truncating the full folio (i.e. offset == 0), adjust the
134 * dirty folio counters appropriately. Only called if there is private
135 * data on the folio.
136 */
137static void ceph_invalidate_folio(struct folio *folio, size_t offset,
138 size_t length)
139{
140 struct inode *inode = folio->mapping->host;
141 struct ceph_client *cl = ceph_inode_to_client(inode);
142 struct ceph_inode_info *ci = ceph_inode(inode);
143 struct ceph_snap_context *snapc;
144
145
146 if (offset != 0 || length != folio_size(folio)) {
147 doutc(cl, "%llx.%llx idx %lu partial dirty page %zu~%zu\n",
148 ceph_vinop(inode), folio->index, offset, length);
149 return;
150 }
151
152 WARN_ON(!folio_test_locked(folio));
153 if (folio_test_private(folio)) {
154 doutc(cl, "%llx.%llx idx %lu full dirty page\n",
155 ceph_vinop(inode), folio->index);
156
157 snapc = folio_detach_private(folio);
158 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
159 ceph_put_snap_context(snapc);
160 }
161
162 netfs_invalidate_folio(folio, offset, length);
163}
164
165static void ceph_netfs_expand_readahead(struct netfs_io_request *rreq)
166{
167 struct inode *inode = rreq->inode;
168 struct ceph_inode_info *ci = ceph_inode(inode);
169 struct ceph_file_layout *lo = &ci->i_layout;
170 unsigned long max_pages = inode->i_sb->s_bdi->ra_pages;
171 loff_t end = rreq->start + rreq->len, new_end;
172 struct ceph_netfs_request_data *priv = rreq->netfs_priv;
173 unsigned long max_len;
174 u32 blockoff;
175
176 if (priv) {
177 /* Readahead is disabled by posix_fadvise POSIX_FADV_RANDOM */
178 if (priv->file_ra_disabled)
179 max_pages = 0;
180 else
181 max_pages = priv->file_ra_pages;
182
183 }
184
185 /* Readahead is disabled */
186 if (!max_pages)
187 return;
188
189 max_len = max_pages << PAGE_SHIFT;
190
191 /*
192 * Try to expand the length forward by rounding up it to the next
193 * block, but do not exceed the file size, unless the original
194 * request already exceeds it.
195 */
196 new_end = min(round_up(end, lo->stripe_unit), rreq->i_size);
197 if (new_end > end && new_end <= rreq->start + max_len)
198 rreq->len = new_end - rreq->start;
199
200 /* Try to expand the start downward */
201 div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
202 if (rreq->len + blockoff <= max_len) {
203 rreq->start -= blockoff;
204 rreq->len += blockoff;
205 }
206}
207
208static bool ceph_netfs_clamp_length(struct netfs_io_subrequest *subreq)
209{
210 struct inode *inode = subreq->rreq->inode;
211 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
212 struct ceph_inode_info *ci = ceph_inode(inode);
213 u64 objno, objoff;
214 u32 xlen;
215
216 /* Truncate the extent at the end of the current block */
217 ceph_calc_file_object_mapping(&ci->i_layout, subreq->start, subreq->len,
218 &objno, &objoff, &xlen);
219 subreq->len = min(xlen, fsc->mount_options->rsize);
220 return true;
221}
222
223static void finish_netfs_read(struct ceph_osd_request *req)
224{
225 struct inode *inode = req->r_inode;
226 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
227 struct ceph_client *cl = fsc->client;
228 struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
229 struct netfs_io_subrequest *subreq = req->r_priv;
230 struct ceph_osd_req_op *op = &req->r_ops[0];
231 int err = req->r_result;
232 bool sparse = (op->op == CEPH_OSD_OP_SPARSE_READ);
233
234 ceph_update_read_metrics(&fsc->mdsc->metric, req->r_start_latency,
235 req->r_end_latency, osd_data->length, err);
236
237 doutc(cl, "result %d subreq->len=%zu i_size=%lld\n", req->r_result,
238 subreq->len, i_size_read(req->r_inode));
239
240 /* no object means success but no data */
241 if (err == -ENOENT)
242 err = 0;
243 else if (err == -EBLOCKLISTED)
244 fsc->blocklisted = true;
245
246 if (err >= 0) {
247 if (sparse && err > 0)
248 err = ceph_sparse_ext_map_end(op);
249 if (err < subreq->len)
250 __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
251 if (IS_ENCRYPTED(inode) && err > 0) {
252 err = ceph_fscrypt_decrypt_extents(inode,
253 osd_data->pages, subreq->start,
254 op->extent.sparse_ext,
255 op->extent.sparse_ext_cnt);
256 if (err > subreq->len)
257 err = subreq->len;
258 }
259 }
260
261 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
262 ceph_put_page_vector(osd_data->pages,
263 calc_pages_for(osd_data->alignment,
264 osd_data->length), false);
265 }
266 netfs_subreq_terminated(subreq, err, false);
267 iput(req->r_inode);
268 ceph_dec_osd_stopping_blocker(fsc->mdsc);
269}
270
271static bool ceph_netfs_issue_op_inline(struct netfs_io_subrequest *subreq)
272{
273 struct netfs_io_request *rreq = subreq->rreq;
274 struct inode *inode = rreq->inode;
275 struct ceph_mds_reply_info_parsed *rinfo;
276 struct ceph_mds_reply_info_in *iinfo;
277 struct ceph_mds_request *req;
278 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
279 struct ceph_inode_info *ci = ceph_inode(inode);
280 struct iov_iter iter;
281 ssize_t err = 0;
282 size_t len;
283 int mode;
284
285 __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
286 __clear_bit(NETFS_SREQ_COPY_TO_CACHE, &subreq->flags);
287
288 if (subreq->start >= inode->i_size)
289 goto out;
290
291 /* We need to fetch the inline data. */
292 mode = ceph_try_to_choose_auth_mds(inode, CEPH_STAT_CAP_INLINE_DATA);
293 req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_GETATTR, mode);
294 if (IS_ERR(req)) {
295 err = PTR_ERR(req);
296 goto out;
297 }
298 req->r_ino1 = ci->i_vino;
299 req->r_args.getattr.mask = cpu_to_le32(CEPH_STAT_CAP_INLINE_DATA);
300 req->r_num_caps = 2;
301
302 err = ceph_mdsc_do_request(mdsc, NULL, req);
303 if (err < 0)
304 goto out;
305
306 rinfo = &req->r_reply_info;
307 iinfo = &rinfo->targeti;
308 if (iinfo->inline_version == CEPH_INLINE_NONE) {
309 /* The data got uninlined */
310 ceph_mdsc_put_request(req);
311 return false;
312 }
313
314 len = min_t(size_t, iinfo->inline_len - subreq->start, subreq->len);
315 iov_iter_xarray(&iter, ITER_DEST, &rreq->mapping->i_pages, subreq->start, len);
316 err = copy_to_iter(iinfo->inline_data + subreq->start, len, &iter);
317 if (err == 0)
318 err = -EFAULT;
319
320 ceph_mdsc_put_request(req);
321out:
322 netfs_subreq_terminated(subreq, err, false);
323 return true;
324}
325
326static void ceph_netfs_issue_read(struct netfs_io_subrequest *subreq)
327{
328 struct netfs_io_request *rreq = subreq->rreq;
329 struct inode *inode = rreq->inode;
330 struct ceph_inode_info *ci = ceph_inode(inode);
331 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
332 struct ceph_client *cl = fsc->client;
333 struct ceph_osd_request *req = NULL;
334 struct ceph_vino vino = ceph_vino(inode);
335 struct iov_iter iter;
336 int err = 0;
337 u64 len = subreq->len;
338 bool sparse = IS_ENCRYPTED(inode) || ceph_test_mount_opt(fsc, SPARSEREAD);
339 u64 off = subreq->start;
340 int extent_cnt;
341
342 if (ceph_inode_is_shutdown(inode)) {
343 err = -EIO;
344 goto out;
345 }
346
347 if (ceph_has_inline_data(ci) && ceph_netfs_issue_op_inline(subreq))
348 return;
349
350 ceph_fscrypt_adjust_off_and_len(inode, &off, &len);
351
352 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, vino,
353 off, &len, 0, 1, sparse ? CEPH_OSD_OP_SPARSE_READ : CEPH_OSD_OP_READ,
354 CEPH_OSD_FLAG_READ, NULL, ci->i_truncate_seq,
355 ci->i_truncate_size, false);
356 if (IS_ERR(req)) {
357 err = PTR_ERR(req);
358 req = NULL;
359 goto out;
360 }
361
362 if (sparse) {
363 extent_cnt = __ceph_sparse_read_ext_count(inode, len);
364 err = ceph_alloc_sparse_ext_map(&req->r_ops[0], extent_cnt);
365 if (err)
366 goto out;
367 }
368
369 doutc(cl, "%llx.%llx pos=%llu orig_len=%zu len=%llu\n",
370 ceph_vinop(inode), subreq->start, subreq->len, len);
371
372 iov_iter_xarray(&iter, ITER_DEST, &rreq->mapping->i_pages, subreq->start, len);
373
374 /*
375 * FIXME: For now, use CEPH_OSD_DATA_TYPE_PAGES instead of _ITER for
376 * encrypted inodes. We'd need infrastructure that handles an iov_iter
377 * instead of page arrays, and we don't have that as of yet. Once the
378 * dust settles on the write helpers and encrypt/decrypt routines for
379 * netfs, we should be able to rework this.
380 */
381 if (IS_ENCRYPTED(inode)) {
382 struct page **pages;
383 size_t page_off;
384
385 err = iov_iter_get_pages_alloc2(&iter, &pages, len, &page_off);
386 if (err < 0) {
387 doutc(cl, "%llx.%llx failed to allocate pages, %d\n",
388 ceph_vinop(inode), err);
389 goto out;
390 }
391
392 /* should always give us a page-aligned read */
393 WARN_ON_ONCE(page_off);
394 len = err;
395 err = 0;
396
397 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false,
398 false);
399 } else {
400 osd_req_op_extent_osd_iter(req, 0, &iter);
401 }
402 if (!ceph_inc_osd_stopping_blocker(fsc->mdsc)) {
403 err = -EIO;
404 goto out;
405 }
406 req->r_callback = finish_netfs_read;
407 req->r_priv = subreq;
408 req->r_inode = inode;
409 ihold(inode);
410
411 ceph_osdc_start_request(req->r_osdc, req);
412out:
413 ceph_osdc_put_request(req);
414 if (err)
415 netfs_subreq_terminated(subreq, err, false);
416 doutc(cl, "%llx.%llx result %d\n", ceph_vinop(inode), err);
417}
418
419static int ceph_init_request(struct netfs_io_request *rreq, struct file *file)
420{
421 struct inode *inode = rreq->inode;
422 struct ceph_client *cl = ceph_inode_to_client(inode);
423 int got = 0, want = CEPH_CAP_FILE_CACHE;
424 struct ceph_netfs_request_data *priv;
425 int ret = 0;
426
427 if (rreq->origin != NETFS_READAHEAD)
428 return 0;
429
430 priv = kzalloc(sizeof(*priv), GFP_NOFS);
431 if (!priv)
432 return -ENOMEM;
433
434 if (file) {
435 struct ceph_rw_context *rw_ctx;
436 struct ceph_file_info *fi = file->private_data;
437
438 priv->file_ra_pages = file->f_ra.ra_pages;
439 priv->file_ra_disabled = file->f_mode & FMODE_RANDOM;
440
441 rw_ctx = ceph_find_rw_context(fi);
442 if (rw_ctx) {
443 rreq->netfs_priv = priv;
444 return 0;
445 }
446 }
447
448 /*
449 * readahead callers do not necessarily hold Fcb caps
450 * (e.g. fadvise, madvise).
451 */
452 ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want, true, &got);
453 if (ret < 0) {
454 doutc(cl, "%llx.%llx, error getting cap\n", ceph_vinop(inode));
455 goto out;
456 }
457
458 if (!(got & want)) {
459 doutc(cl, "%llx.%llx, no cache cap\n", ceph_vinop(inode));
460 ret = -EACCES;
461 goto out;
462 }
463 if (ret == 0) {
464 ret = -EACCES;
465 goto out;
466 }
467
468 priv->caps = got;
469 rreq->netfs_priv = priv;
470
471out:
472 if (ret < 0)
473 kfree(priv);
474
475 return ret;
476}
477
478static void ceph_netfs_free_request(struct netfs_io_request *rreq)
479{
480 struct ceph_netfs_request_data *priv = rreq->netfs_priv;
481
482 if (!priv)
483 return;
484
485 if (priv->caps)
486 ceph_put_cap_refs(ceph_inode(rreq->inode), priv->caps);
487 kfree(priv);
488 rreq->netfs_priv = NULL;
489}
490
491const struct netfs_request_ops ceph_netfs_ops = {
492 .init_request = ceph_init_request,
493 .free_request = ceph_netfs_free_request,
494 .issue_read = ceph_netfs_issue_read,
495 .expand_readahead = ceph_netfs_expand_readahead,
496 .clamp_length = ceph_netfs_clamp_length,
497 .check_write_begin = ceph_netfs_check_write_begin,
498};
499
500#ifdef CONFIG_CEPH_FSCACHE
501static void ceph_set_page_fscache(struct page *page)
502{
503 set_page_fscache(page);
504}
505
506static void ceph_fscache_write_terminated(void *priv, ssize_t error, bool was_async)
507{
508 struct inode *inode = priv;
509
510 if (IS_ERR_VALUE(error) && error != -ENOBUFS)
511 ceph_fscache_invalidate(inode, false);
512}
513
514static void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
515{
516 struct ceph_inode_info *ci = ceph_inode(inode);
517 struct fscache_cookie *cookie = ceph_fscache_cookie(ci);
518
519 fscache_write_to_cache(cookie, inode->i_mapping, off, len, i_size_read(inode),
520 ceph_fscache_write_terminated, inode, caching);
521}
522#else
523static inline void ceph_set_page_fscache(struct page *page)
524{
525}
526
527static inline void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
528{
529}
530#endif /* CONFIG_CEPH_FSCACHE */
531
532struct ceph_writeback_ctl
533{
534 loff_t i_size;
535 u64 truncate_size;
536 u32 truncate_seq;
537 bool size_stable;
538 bool head_snapc;
539};
540
541/*
542 * Get ref for the oldest snapc for an inode with dirty data... that is, the
543 * only snap context we are allowed to write back.
544 */
545static struct ceph_snap_context *
546get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
547 struct ceph_snap_context *page_snapc)
548{
549 struct ceph_inode_info *ci = ceph_inode(inode);
550 struct ceph_client *cl = ceph_inode_to_client(inode);
551 struct ceph_snap_context *snapc = NULL;
552 struct ceph_cap_snap *capsnap = NULL;
553
554 spin_lock(&ci->i_ceph_lock);
555 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
556 doutc(cl, " capsnap %p snapc %p has %d dirty pages\n",
557 capsnap, capsnap->context, capsnap->dirty_pages);
558 if (!capsnap->dirty_pages)
559 continue;
560
561 /* get i_size, truncate_{seq,size} for page_snapc? */
562 if (snapc && capsnap->context != page_snapc)
563 continue;
564
565 if (ctl) {
566 if (capsnap->writing) {
567 ctl->i_size = i_size_read(inode);
568 ctl->size_stable = false;
569 } else {
570 ctl->i_size = capsnap->size;
571 ctl->size_stable = true;
572 }
573 ctl->truncate_size = capsnap->truncate_size;
574 ctl->truncate_seq = capsnap->truncate_seq;
575 ctl->head_snapc = false;
576 }
577
578 if (snapc)
579 break;
580
581 snapc = ceph_get_snap_context(capsnap->context);
582 if (!page_snapc ||
583 page_snapc == snapc ||
584 page_snapc->seq > snapc->seq)
585 break;
586 }
587 if (!snapc && ci->i_wrbuffer_ref_head) {
588 snapc = ceph_get_snap_context(ci->i_head_snapc);
589 doutc(cl, " head snapc %p has %d dirty pages\n", snapc,
590 ci->i_wrbuffer_ref_head);
591 if (ctl) {
592 ctl->i_size = i_size_read(inode);
593 ctl->truncate_size = ci->i_truncate_size;
594 ctl->truncate_seq = ci->i_truncate_seq;
595 ctl->size_stable = false;
596 ctl->head_snapc = true;
597 }
598 }
599 spin_unlock(&ci->i_ceph_lock);
600 return snapc;
601}
602
603static u64 get_writepages_data_length(struct inode *inode,
604 struct page *page, u64 start)
605{
606 struct ceph_inode_info *ci = ceph_inode(inode);
607 struct ceph_snap_context *snapc;
608 struct ceph_cap_snap *capsnap = NULL;
609 u64 end = i_size_read(inode);
610 u64 ret;
611
612 snapc = page_snap_context(ceph_fscrypt_pagecache_page(page));
613 if (snapc != ci->i_head_snapc) {
614 bool found = false;
615 spin_lock(&ci->i_ceph_lock);
616 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
617 if (capsnap->context == snapc) {
618 if (!capsnap->writing)
619 end = capsnap->size;
620 found = true;
621 break;
622 }
623 }
624 spin_unlock(&ci->i_ceph_lock);
625 WARN_ON(!found);
626 }
627 if (end > ceph_fscrypt_page_offset(page) + thp_size(page))
628 end = ceph_fscrypt_page_offset(page) + thp_size(page);
629 ret = end > start ? end - start : 0;
630 if (ret && fscrypt_is_bounce_page(page))
631 ret = round_up(ret, CEPH_FSCRYPT_BLOCK_SIZE);
632 return ret;
633}
634
635/*
636 * Write a single page, but leave the page locked.
637 *
638 * If we get a write error, mark the mapping for error, but still adjust the
639 * dirty page accounting (i.e., page is no longer dirty).
640 */
641static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
642{
643 struct folio *folio = page_folio(page);
644 struct inode *inode = page->mapping->host;
645 struct ceph_inode_info *ci = ceph_inode(inode);
646 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
647 struct ceph_client *cl = fsc->client;
648 struct ceph_snap_context *snapc, *oldest;
649 loff_t page_off = page_offset(page);
650 int err;
651 loff_t len = thp_size(page);
652 loff_t wlen;
653 struct ceph_writeback_ctl ceph_wbc;
654 struct ceph_osd_client *osdc = &fsc->client->osdc;
655 struct ceph_osd_request *req;
656 bool caching = ceph_is_cache_enabled(inode);
657 struct page *bounce_page = NULL;
658
659 doutc(cl, "%llx.%llx page %p idx %lu\n", ceph_vinop(inode), page,
660 page->index);
661
662 if (ceph_inode_is_shutdown(inode))
663 return -EIO;
664
665 /* verify this is a writeable snap context */
666 snapc = page_snap_context(page);
667 if (!snapc) {
668 doutc(cl, "%llx.%llx page %p not dirty?\n", ceph_vinop(inode),
669 page);
670 return 0;
671 }
672 oldest = get_oldest_context(inode, &ceph_wbc, snapc);
673 if (snapc->seq > oldest->seq) {
674 doutc(cl, "%llx.%llx page %p snapc %p not writeable - noop\n",
675 ceph_vinop(inode), page, snapc);
676 /* we should only noop if called by kswapd */
677 WARN_ON(!(current->flags & PF_MEMALLOC));
678 ceph_put_snap_context(oldest);
679 redirty_page_for_writepage(wbc, page);
680 return 0;
681 }
682 ceph_put_snap_context(oldest);
683
684 /* is this a partial page at end of file? */
685 if (page_off >= ceph_wbc.i_size) {
686 doutc(cl, "%llx.%llx folio at %lu beyond eof %llu\n",
687 ceph_vinop(inode), folio->index, ceph_wbc.i_size);
688 folio_invalidate(folio, 0, folio_size(folio));
689 return 0;
690 }
691
692 if (ceph_wbc.i_size < page_off + len)
693 len = ceph_wbc.i_size - page_off;
694
695 wlen = IS_ENCRYPTED(inode) ? round_up(len, CEPH_FSCRYPT_BLOCK_SIZE) : len;
696 doutc(cl, "%llx.%llx page %p index %lu on %llu~%llu snapc %p seq %lld\n",
697 ceph_vinop(inode), page, page->index, page_off, wlen, snapc,
698 snapc->seq);
699
700 if (atomic_long_inc_return(&fsc->writeback_count) >
701 CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
702 fsc->write_congested = true;
703
704 req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode),
705 page_off, &wlen, 0, 1, CEPH_OSD_OP_WRITE,
706 CEPH_OSD_FLAG_WRITE, snapc,
707 ceph_wbc.truncate_seq,
708 ceph_wbc.truncate_size, true);
709 if (IS_ERR(req)) {
710 redirty_page_for_writepage(wbc, page);
711 return PTR_ERR(req);
712 }
713
714 if (wlen < len)
715 len = wlen;
716
717 set_page_writeback(page);
718 if (caching)
719 ceph_set_page_fscache(page);
720 ceph_fscache_write_to_cache(inode, page_off, len, caching);
721
722 if (IS_ENCRYPTED(inode)) {
723 bounce_page = fscrypt_encrypt_pagecache_blocks(page,
724 CEPH_FSCRYPT_BLOCK_SIZE, 0,
725 GFP_NOFS);
726 if (IS_ERR(bounce_page)) {
727 redirty_page_for_writepage(wbc, page);
728 end_page_writeback(page);
729 ceph_osdc_put_request(req);
730 return PTR_ERR(bounce_page);
731 }
732 }
733
734 /* it may be a short write due to an object boundary */
735 WARN_ON_ONCE(len > thp_size(page));
736 osd_req_op_extent_osd_data_pages(req, 0,
737 bounce_page ? &bounce_page : &page, wlen, 0,
738 false, false);
739 doutc(cl, "%llx.%llx %llu~%llu (%llu bytes, %sencrypted)\n",
740 ceph_vinop(inode), page_off, len, wlen,
741 IS_ENCRYPTED(inode) ? "" : "not ");
742
743 req->r_mtime = inode_get_mtime(inode);
744 ceph_osdc_start_request(osdc, req);
745 err = ceph_osdc_wait_request(osdc, req);
746
747 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
748 req->r_end_latency, len, err);
749 fscrypt_free_bounce_page(bounce_page);
750 ceph_osdc_put_request(req);
751 if (err == 0)
752 err = len;
753
754 if (err < 0) {
755 struct writeback_control tmp_wbc;
756 if (!wbc)
757 wbc = &tmp_wbc;
758 if (err == -ERESTARTSYS) {
759 /* killed by SIGKILL */
760 doutc(cl, "%llx.%llx interrupted page %p\n",
761 ceph_vinop(inode), page);
762 redirty_page_for_writepage(wbc, page);
763 end_page_writeback(page);
764 return err;
765 }
766 if (err == -EBLOCKLISTED)
767 fsc->blocklisted = true;
768 doutc(cl, "%llx.%llx setting page/mapping error %d %p\n",
769 ceph_vinop(inode), err, page);
770 mapping_set_error(&inode->i_data, err);
771 wbc->pages_skipped++;
772 } else {
773 doutc(cl, "%llx.%llx cleaned page %p\n",
774 ceph_vinop(inode), page);
775 err = 0; /* vfs expects us to return 0 */
776 }
777 oldest = detach_page_private(page);
778 WARN_ON_ONCE(oldest != snapc);
779 end_page_writeback(page);
780 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
781 ceph_put_snap_context(snapc); /* page's reference */
782
783 if (atomic_long_dec_return(&fsc->writeback_count) <
784 CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
785 fsc->write_congested = false;
786
787 return err;
788}
789
790static int ceph_writepage(struct page *page, struct writeback_control *wbc)
791{
792 int err;
793 struct inode *inode = page->mapping->host;
794 BUG_ON(!inode);
795 ihold(inode);
796
797 if (wbc->sync_mode == WB_SYNC_NONE &&
798 ceph_inode_to_fs_client(inode)->write_congested) {
799 redirty_page_for_writepage(wbc, page);
800 return AOP_WRITEPAGE_ACTIVATE;
801 }
802
803 wait_on_page_fscache(page);
804
805 err = writepage_nounlock(page, wbc);
806 if (err == -ERESTARTSYS) {
807 /* direct memory reclaimer was killed by SIGKILL. return 0
808 * to prevent caller from setting mapping/page error */
809 err = 0;
810 }
811 unlock_page(page);
812 iput(inode);
813 return err;
814}
815
816/*
817 * async writeback completion handler.
818 *
819 * If we get an error, set the mapping error bit, but not the individual
820 * page error bits.
821 */
822static void writepages_finish(struct ceph_osd_request *req)
823{
824 struct inode *inode = req->r_inode;
825 struct ceph_inode_info *ci = ceph_inode(inode);
826 struct ceph_client *cl = ceph_inode_to_client(inode);
827 struct ceph_osd_data *osd_data;
828 struct page *page;
829 int num_pages, total_pages = 0;
830 int i, j;
831 int rc = req->r_result;
832 struct ceph_snap_context *snapc = req->r_snapc;
833 struct address_space *mapping = inode->i_mapping;
834 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
835 unsigned int len = 0;
836 bool remove_page;
837
838 doutc(cl, "%llx.%llx rc %d\n", ceph_vinop(inode), rc);
839 if (rc < 0) {
840 mapping_set_error(mapping, rc);
841 ceph_set_error_write(ci);
842 if (rc == -EBLOCKLISTED)
843 fsc->blocklisted = true;
844 } else {
845 ceph_clear_error_write(ci);
846 }
847
848 /*
849 * We lost the cache cap, need to truncate the page before
850 * it is unlocked, otherwise we'd truncate it later in the
851 * page truncation thread, possibly losing some data that
852 * raced its way in
853 */
854 remove_page = !(ceph_caps_issued(ci) &
855 (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
856
857 /* clean all pages */
858 for (i = 0; i < req->r_num_ops; i++) {
859 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE) {
860 pr_warn_client(cl,
861 "%llx.%llx incorrect op %d req %p index %d tid %llu\n",
862 ceph_vinop(inode), req->r_ops[i].op, req, i,
863 req->r_tid);
864 break;
865 }
866
867 osd_data = osd_req_op_extent_osd_data(req, i);
868 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
869 len += osd_data->length;
870 num_pages = calc_pages_for((u64)osd_data->alignment,
871 (u64)osd_data->length);
872 total_pages += num_pages;
873 for (j = 0; j < num_pages; j++) {
874 page = osd_data->pages[j];
875 if (fscrypt_is_bounce_page(page)) {
876 page = fscrypt_pagecache_page(page);
877 fscrypt_free_bounce_page(osd_data->pages[j]);
878 osd_data->pages[j] = page;
879 }
880 BUG_ON(!page);
881 WARN_ON(!PageUptodate(page));
882
883 if (atomic_long_dec_return(&fsc->writeback_count) <
884 CONGESTION_OFF_THRESH(
885 fsc->mount_options->congestion_kb))
886 fsc->write_congested = false;
887
888 ceph_put_snap_context(detach_page_private(page));
889 end_page_writeback(page);
890 doutc(cl, "unlocking %p\n", page);
891
892 if (remove_page)
893 generic_error_remove_folio(inode->i_mapping,
894 page_folio(page));
895
896 unlock_page(page);
897 }
898 doutc(cl, "%llx.%llx wrote %llu bytes cleaned %d pages\n",
899 ceph_vinop(inode), osd_data->length,
900 rc >= 0 ? num_pages : 0);
901
902 release_pages(osd_data->pages, num_pages);
903 }
904
905 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
906 req->r_end_latency, len, rc);
907
908 ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
909
910 osd_data = osd_req_op_extent_osd_data(req, 0);
911 if (osd_data->pages_from_pool)
912 mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
913 else
914 kfree(osd_data->pages);
915 ceph_osdc_put_request(req);
916 ceph_dec_osd_stopping_blocker(fsc->mdsc);
917}
918
919/*
920 * initiate async writeback
921 */
922static int ceph_writepages_start(struct address_space *mapping,
923 struct writeback_control *wbc)
924{
925 struct inode *inode = mapping->host;
926 struct ceph_inode_info *ci = ceph_inode(inode);
927 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
928 struct ceph_client *cl = fsc->client;
929 struct ceph_vino vino = ceph_vino(inode);
930 pgoff_t index, start_index, end = -1;
931 struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
932 struct folio_batch fbatch;
933 int rc = 0;
934 unsigned int wsize = i_blocksize(inode);
935 struct ceph_osd_request *req = NULL;
936 struct ceph_writeback_ctl ceph_wbc;
937 bool should_loop, range_whole = false;
938 bool done = false;
939 bool caching = ceph_is_cache_enabled(inode);
940 xa_mark_t tag;
941
942 if (wbc->sync_mode == WB_SYNC_NONE &&
943 fsc->write_congested)
944 return 0;
945
946 doutc(cl, "%llx.%llx (mode=%s)\n", ceph_vinop(inode),
947 wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
948 (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
949
950 if (ceph_inode_is_shutdown(inode)) {
951 if (ci->i_wrbuffer_ref > 0) {
952 pr_warn_ratelimited_client(cl,
953 "%llx.%llx %lld forced umount\n",
954 ceph_vinop(inode), ceph_ino(inode));
955 }
956 mapping_set_error(mapping, -EIO);
957 return -EIO; /* we're in a forced umount, don't write! */
958 }
959 if (fsc->mount_options->wsize < wsize)
960 wsize = fsc->mount_options->wsize;
961
962 folio_batch_init(&fbatch);
963
964 start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
965 index = start_index;
966
967 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
968 tag = PAGECACHE_TAG_TOWRITE;
969 } else {
970 tag = PAGECACHE_TAG_DIRTY;
971 }
972retry:
973 /* find oldest snap context with dirty data */
974 snapc = get_oldest_context(inode, &ceph_wbc, NULL);
975 if (!snapc) {
976 /* hmm, why does writepages get called when there
977 is no dirty data? */
978 doutc(cl, " no snap context with dirty data?\n");
979 goto out;
980 }
981 doutc(cl, " oldest snapc is %p seq %lld (%d snaps)\n", snapc,
982 snapc->seq, snapc->num_snaps);
983
984 should_loop = false;
985 if (ceph_wbc.head_snapc && snapc != last_snapc) {
986 /* where to start/end? */
987 if (wbc->range_cyclic) {
988 index = start_index;
989 end = -1;
990 if (index > 0)
991 should_loop = true;
992 doutc(cl, " cyclic, start at %lu\n", index);
993 } else {
994 index = wbc->range_start >> PAGE_SHIFT;
995 end = wbc->range_end >> PAGE_SHIFT;
996 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
997 range_whole = true;
998 doutc(cl, " not cyclic, %lu to %lu\n", index, end);
999 }
1000 } else if (!ceph_wbc.head_snapc) {
1001 /* Do not respect wbc->range_{start,end}. Dirty pages
1002 * in that range can be associated with newer snapc.
1003 * They are not writeable until we write all dirty pages
1004 * associated with 'snapc' get written */
1005 if (index > 0)
1006 should_loop = true;
1007 doutc(cl, " non-head snapc, range whole\n");
1008 }
1009
1010 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1011 tag_pages_for_writeback(mapping, index, end);
1012
1013 ceph_put_snap_context(last_snapc);
1014 last_snapc = snapc;
1015
1016 while (!done && index <= end) {
1017 int num_ops = 0, op_idx;
1018 unsigned i, nr_folios, max_pages, locked_pages = 0;
1019 struct page **pages = NULL, **data_pages;
1020 struct page *page;
1021 pgoff_t strip_unit_end = 0;
1022 u64 offset = 0, len = 0;
1023 bool from_pool = false;
1024
1025 max_pages = wsize >> PAGE_SHIFT;
1026
1027get_more_pages:
1028 nr_folios = filemap_get_folios_tag(mapping, &index,
1029 end, tag, &fbatch);
1030 doutc(cl, "pagevec_lookup_range_tag got %d\n", nr_folios);
1031 if (!nr_folios && !locked_pages)
1032 break;
1033 for (i = 0; i < nr_folios && locked_pages < max_pages; i++) {
1034 page = &fbatch.folios[i]->page;
1035 doutc(cl, "? %p idx %lu\n", page, page->index);
1036 if (locked_pages == 0)
1037 lock_page(page); /* first page */
1038 else if (!trylock_page(page))
1039 break;
1040
1041 /* only dirty pages, or our accounting breaks */
1042 if (unlikely(!PageDirty(page)) ||
1043 unlikely(page->mapping != mapping)) {
1044 doutc(cl, "!dirty or !mapping %p\n", page);
1045 unlock_page(page);
1046 continue;
1047 }
1048 /* only if matching snap context */
1049 pgsnapc = page_snap_context(page);
1050 if (pgsnapc != snapc) {
1051 doutc(cl, "page snapc %p %lld != oldest %p %lld\n",
1052 pgsnapc, pgsnapc->seq, snapc, snapc->seq);
1053 if (!should_loop &&
1054 !ceph_wbc.head_snapc &&
1055 wbc->sync_mode != WB_SYNC_NONE)
1056 should_loop = true;
1057 unlock_page(page);
1058 continue;
1059 }
1060 if (page_offset(page) >= ceph_wbc.i_size) {
1061 struct folio *folio = page_folio(page);
1062
1063 doutc(cl, "folio at %lu beyond eof %llu\n",
1064 folio->index, ceph_wbc.i_size);
1065 if ((ceph_wbc.size_stable ||
1066 folio_pos(folio) >= i_size_read(inode)) &&
1067 folio_clear_dirty_for_io(folio))
1068 folio_invalidate(folio, 0,
1069 folio_size(folio));
1070 folio_unlock(folio);
1071 continue;
1072 }
1073 if (strip_unit_end && (page->index > strip_unit_end)) {
1074 doutc(cl, "end of strip unit %p\n", page);
1075 unlock_page(page);
1076 break;
1077 }
1078 if (PageWriteback(page) || PageFsCache(page)) {
1079 if (wbc->sync_mode == WB_SYNC_NONE) {
1080 doutc(cl, "%p under writeback\n", page);
1081 unlock_page(page);
1082 continue;
1083 }
1084 doutc(cl, "waiting on writeback %p\n", page);
1085 wait_on_page_writeback(page);
1086 wait_on_page_fscache(page);
1087 }
1088
1089 if (!clear_page_dirty_for_io(page)) {
1090 doutc(cl, "%p !clear_page_dirty_for_io\n", page);
1091 unlock_page(page);
1092 continue;
1093 }
1094
1095 /*
1096 * We have something to write. If this is
1097 * the first locked page this time through,
1098 * calculate max possinle write size and
1099 * allocate a page array
1100 */
1101 if (locked_pages == 0) {
1102 u64 objnum;
1103 u64 objoff;
1104 u32 xlen;
1105
1106 /* prepare async write request */
1107 offset = (u64)page_offset(page);
1108 ceph_calc_file_object_mapping(&ci->i_layout,
1109 offset, wsize,
1110 &objnum, &objoff,
1111 &xlen);
1112 len = xlen;
1113
1114 num_ops = 1;
1115 strip_unit_end = page->index +
1116 ((len - 1) >> PAGE_SHIFT);
1117
1118 BUG_ON(pages);
1119 max_pages = calc_pages_for(0, (u64)len);
1120 pages = kmalloc_array(max_pages,
1121 sizeof(*pages),
1122 GFP_NOFS);
1123 if (!pages) {
1124 from_pool = true;
1125 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1126 BUG_ON(!pages);
1127 }
1128
1129 len = 0;
1130 } else if (page->index !=
1131 (offset + len) >> PAGE_SHIFT) {
1132 if (num_ops >= (from_pool ? CEPH_OSD_SLAB_OPS :
1133 CEPH_OSD_MAX_OPS)) {
1134 redirty_page_for_writepage(wbc, page);
1135 unlock_page(page);
1136 break;
1137 }
1138
1139 num_ops++;
1140 offset = (u64)page_offset(page);
1141 len = 0;
1142 }
1143
1144 /* note position of first page in fbatch */
1145 doutc(cl, "%llx.%llx will write page %p idx %lu\n",
1146 ceph_vinop(inode), page, page->index);
1147
1148 if (atomic_long_inc_return(&fsc->writeback_count) >
1149 CONGESTION_ON_THRESH(
1150 fsc->mount_options->congestion_kb))
1151 fsc->write_congested = true;
1152
1153 if (IS_ENCRYPTED(inode)) {
1154 pages[locked_pages] =
1155 fscrypt_encrypt_pagecache_blocks(page,
1156 PAGE_SIZE, 0,
1157 locked_pages ? GFP_NOWAIT : GFP_NOFS);
1158 if (IS_ERR(pages[locked_pages])) {
1159 if (PTR_ERR(pages[locked_pages]) == -EINVAL)
1160 pr_err_client(cl,
1161 "inode->i_blkbits=%hhu\n",
1162 inode->i_blkbits);
1163 /* better not fail on first page! */
1164 BUG_ON(locked_pages == 0);
1165 pages[locked_pages] = NULL;
1166 redirty_page_for_writepage(wbc, page);
1167 unlock_page(page);
1168 break;
1169 }
1170 ++locked_pages;
1171 } else {
1172 pages[locked_pages++] = page;
1173 }
1174
1175 fbatch.folios[i] = NULL;
1176 len += thp_size(page);
1177 }
1178
1179 /* did we get anything? */
1180 if (!locked_pages)
1181 goto release_folios;
1182 if (i) {
1183 unsigned j, n = 0;
1184 /* shift unused page to beginning of fbatch */
1185 for (j = 0; j < nr_folios; j++) {
1186 if (!fbatch.folios[j])
1187 continue;
1188 if (n < j)
1189 fbatch.folios[n] = fbatch.folios[j];
1190 n++;
1191 }
1192 fbatch.nr = n;
1193
1194 if (nr_folios && i == nr_folios &&
1195 locked_pages < max_pages) {
1196 doutc(cl, "reached end fbatch, trying for more\n");
1197 folio_batch_release(&fbatch);
1198 goto get_more_pages;
1199 }
1200 }
1201
1202new_request:
1203 offset = ceph_fscrypt_page_offset(pages[0]);
1204 len = wsize;
1205
1206 req = ceph_osdc_new_request(&fsc->client->osdc,
1207 &ci->i_layout, vino,
1208 offset, &len, 0, num_ops,
1209 CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1210 snapc, ceph_wbc.truncate_seq,
1211 ceph_wbc.truncate_size, false);
1212 if (IS_ERR(req)) {
1213 req = ceph_osdc_new_request(&fsc->client->osdc,
1214 &ci->i_layout, vino,
1215 offset, &len, 0,
1216 min(num_ops,
1217 CEPH_OSD_SLAB_OPS),
1218 CEPH_OSD_OP_WRITE,
1219 CEPH_OSD_FLAG_WRITE,
1220 snapc, ceph_wbc.truncate_seq,
1221 ceph_wbc.truncate_size, true);
1222 BUG_ON(IS_ERR(req));
1223 }
1224 BUG_ON(len < ceph_fscrypt_page_offset(pages[locked_pages - 1]) +
1225 thp_size(pages[locked_pages - 1]) - offset);
1226
1227 if (!ceph_inc_osd_stopping_blocker(fsc->mdsc)) {
1228 rc = -EIO;
1229 goto release_folios;
1230 }
1231 req->r_callback = writepages_finish;
1232 req->r_inode = inode;
1233
1234 /* Format the osd request message and submit the write */
1235 len = 0;
1236 data_pages = pages;
1237 op_idx = 0;
1238 for (i = 0; i < locked_pages; i++) {
1239 struct page *page = ceph_fscrypt_pagecache_page(pages[i]);
1240
1241 u64 cur_offset = page_offset(page);
1242 /*
1243 * Discontinuity in page range? Ceph can handle that by just passing
1244 * multiple extents in the write op.
1245 */
1246 if (offset + len != cur_offset) {
1247 /* If it's full, stop here */
1248 if (op_idx + 1 == req->r_num_ops)
1249 break;
1250
1251 /* Kick off an fscache write with what we have so far. */
1252 ceph_fscache_write_to_cache(inode, offset, len, caching);
1253
1254 /* Start a new extent */
1255 osd_req_op_extent_dup_last(req, op_idx,
1256 cur_offset - offset);
1257 doutc(cl, "got pages at %llu~%llu\n", offset,
1258 len);
1259 osd_req_op_extent_osd_data_pages(req, op_idx,
1260 data_pages, len, 0,
1261 from_pool, false);
1262 osd_req_op_extent_update(req, op_idx, len);
1263
1264 len = 0;
1265 offset = cur_offset;
1266 data_pages = pages + i;
1267 op_idx++;
1268 }
1269
1270 set_page_writeback(page);
1271 if (caching)
1272 ceph_set_page_fscache(page);
1273 len += thp_size(page);
1274 }
1275 ceph_fscache_write_to_cache(inode, offset, len, caching);
1276
1277 if (ceph_wbc.size_stable) {
1278 len = min(len, ceph_wbc.i_size - offset);
1279 } else if (i == locked_pages) {
1280 /* writepages_finish() clears writeback pages
1281 * according to the data length, so make sure
1282 * data length covers all locked pages */
1283 u64 min_len = len + 1 - thp_size(page);
1284 len = get_writepages_data_length(inode, pages[i - 1],
1285 offset);
1286 len = max(len, min_len);
1287 }
1288 if (IS_ENCRYPTED(inode))
1289 len = round_up(len, CEPH_FSCRYPT_BLOCK_SIZE);
1290
1291 doutc(cl, "got pages at %llu~%llu\n", offset, len);
1292
1293 if (IS_ENCRYPTED(inode) &&
1294 ((offset | len) & ~CEPH_FSCRYPT_BLOCK_MASK))
1295 pr_warn_client(cl,
1296 "bad encrypted write offset=%lld len=%llu\n",
1297 offset, len);
1298
1299 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1300 0, from_pool, false);
1301 osd_req_op_extent_update(req, op_idx, len);
1302
1303 BUG_ON(op_idx + 1 != req->r_num_ops);
1304
1305 from_pool = false;
1306 if (i < locked_pages) {
1307 BUG_ON(num_ops <= req->r_num_ops);
1308 num_ops -= req->r_num_ops;
1309 locked_pages -= i;
1310
1311 /* allocate new pages array for next request */
1312 data_pages = pages;
1313 pages = kmalloc_array(locked_pages, sizeof(*pages),
1314 GFP_NOFS);
1315 if (!pages) {
1316 from_pool = true;
1317 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1318 BUG_ON(!pages);
1319 }
1320 memcpy(pages, data_pages + i,
1321 locked_pages * sizeof(*pages));
1322 memset(data_pages + i, 0,
1323 locked_pages * sizeof(*pages));
1324 } else {
1325 BUG_ON(num_ops != req->r_num_ops);
1326 index = pages[i - 1]->index + 1;
1327 /* request message now owns the pages array */
1328 pages = NULL;
1329 }
1330
1331 req->r_mtime = inode_get_mtime(inode);
1332 ceph_osdc_start_request(&fsc->client->osdc, req);
1333 req = NULL;
1334
1335 wbc->nr_to_write -= i;
1336 if (pages)
1337 goto new_request;
1338
1339 /*
1340 * We stop writing back only if we are not doing
1341 * integrity sync. In case of integrity sync we have to
1342 * keep going until we have written all the pages
1343 * we tagged for writeback prior to entering this loop.
1344 */
1345 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1346 done = true;
1347
1348release_folios:
1349 doutc(cl, "folio_batch release on %d folios (%p)\n",
1350 (int)fbatch.nr, fbatch.nr ? fbatch.folios[0] : NULL);
1351 folio_batch_release(&fbatch);
1352 }
1353
1354 if (should_loop && !done) {
1355 /* more to do; loop back to beginning of file */
1356 doutc(cl, "looping back to beginning of file\n");
1357 end = start_index - 1; /* OK even when start_index == 0 */
1358
1359 /* to write dirty pages associated with next snapc,
1360 * we need to wait until current writes complete */
1361 if (wbc->sync_mode != WB_SYNC_NONE &&
1362 start_index == 0 && /* all dirty pages were checked */
1363 !ceph_wbc.head_snapc) {
1364 struct page *page;
1365 unsigned i, nr;
1366 index = 0;
1367 while ((index <= end) &&
1368 (nr = filemap_get_folios_tag(mapping, &index,
1369 (pgoff_t)-1,
1370 PAGECACHE_TAG_WRITEBACK,
1371 &fbatch))) {
1372 for (i = 0; i < nr; i++) {
1373 page = &fbatch.folios[i]->page;
1374 if (page_snap_context(page) != snapc)
1375 continue;
1376 wait_on_page_writeback(page);
1377 }
1378 folio_batch_release(&fbatch);
1379 cond_resched();
1380 }
1381 }
1382
1383 start_index = 0;
1384 index = 0;
1385 goto retry;
1386 }
1387
1388 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1389 mapping->writeback_index = index;
1390
1391out:
1392 ceph_osdc_put_request(req);
1393 ceph_put_snap_context(last_snapc);
1394 doutc(cl, "%llx.%llx dend - startone, rc = %d\n", ceph_vinop(inode),
1395 rc);
1396 return rc;
1397}
1398
1399
1400
1401/*
1402 * See if a given @snapc is either writeable, or already written.
1403 */
1404static int context_is_writeable_or_written(struct inode *inode,
1405 struct ceph_snap_context *snapc)
1406{
1407 struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1408 int ret = !oldest || snapc->seq <= oldest->seq;
1409
1410 ceph_put_snap_context(oldest);
1411 return ret;
1412}
1413
1414/**
1415 * ceph_find_incompatible - find an incompatible context and return it
1416 * @page: page being dirtied
1417 *
1418 * We are only allowed to write into/dirty a page if the page is
1419 * clean, or already dirty within the same snap context. Returns a
1420 * conflicting context if there is one, NULL if there isn't, or a
1421 * negative error code on other errors.
1422 *
1423 * Must be called with page lock held.
1424 */
1425static struct ceph_snap_context *
1426ceph_find_incompatible(struct page *page)
1427{
1428 struct inode *inode = page->mapping->host;
1429 struct ceph_client *cl = ceph_inode_to_client(inode);
1430 struct ceph_inode_info *ci = ceph_inode(inode);
1431
1432 if (ceph_inode_is_shutdown(inode)) {
1433 doutc(cl, " %llx.%llx page %p is shutdown\n",
1434 ceph_vinop(inode), page);
1435 return ERR_PTR(-ESTALE);
1436 }
1437
1438 for (;;) {
1439 struct ceph_snap_context *snapc, *oldest;
1440
1441 wait_on_page_writeback(page);
1442
1443 snapc = page_snap_context(page);
1444 if (!snapc || snapc == ci->i_head_snapc)
1445 break;
1446
1447 /*
1448 * this page is already dirty in another (older) snap
1449 * context! is it writeable now?
1450 */
1451 oldest = get_oldest_context(inode, NULL, NULL);
1452 if (snapc->seq > oldest->seq) {
1453 /* not writeable -- return it for the caller to deal with */
1454 ceph_put_snap_context(oldest);
1455 doutc(cl, " %llx.%llx page %p snapc %p not current or oldest\n",
1456 ceph_vinop(inode), page, snapc);
1457 return ceph_get_snap_context(snapc);
1458 }
1459 ceph_put_snap_context(oldest);
1460
1461 /* yay, writeable, do it now (without dropping page lock) */
1462 doutc(cl, " %llx.%llx page %p snapc %p not current, but oldest\n",
1463 ceph_vinop(inode), page, snapc);
1464 if (clear_page_dirty_for_io(page)) {
1465 int r = writepage_nounlock(page, NULL);
1466 if (r < 0)
1467 return ERR_PTR(r);
1468 }
1469 }
1470 return NULL;
1471}
1472
1473static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
1474 struct folio **foliop, void **_fsdata)
1475{
1476 struct inode *inode = file_inode(file);
1477 struct ceph_inode_info *ci = ceph_inode(inode);
1478 struct ceph_snap_context *snapc;
1479
1480 snapc = ceph_find_incompatible(folio_page(*foliop, 0));
1481 if (snapc) {
1482 int r;
1483
1484 folio_unlock(*foliop);
1485 folio_put(*foliop);
1486 *foliop = NULL;
1487 if (IS_ERR(snapc))
1488 return PTR_ERR(snapc);
1489
1490 ceph_queue_writeback(inode);
1491 r = wait_event_killable(ci->i_cap_wq,
1492 context_is_writeable_or_written(inode, snapc));
1493 ceph_put_snap_context(snapc);
1494 return r == 0 ? -EAGAIN : r;
1495 }
1496 return 0;
1497}
1498
1499/*
1500 * We are only allowed to write into/dirty the page if the page is
1501 * clean, or already dirty within the same snap context.
1502 */
1503static int ceph_write_begin(struct file *file, struct address_space *mapping,
1504 loff_t pos, unsigned len,
1505 struct page **pagep, void **fsdata)
1506{
1507 struct inode *inode = file_inode(file);
1508 struct ceph_inode_info *ci = ceph_inode(inode);
1509 struct folio *folio = NULL;
1510 int r;
1511
1512 r = netfs_write_begin(&ci->netfs, file, inode->i_mapping, pos, len, &folio, NULL);
1513 if (r < 0)
1514 return r;
1515
1516 folio_wait_fscache(folio);
1517 WARN_ON_ONCE(!folio_test_locked(folio));
1518 *pagep = &folio->page;
1519 return 0;
1520}
1521
1522/*
1523 * we don't do anything in here that simple_write_end doesn't do
1524 * except adjust dirty page accounting
1525 */
1526static int ceph_write_end(struct file *file, struct address_space *mapping,
1527 loff_t pos, unsigned len, unsigned copied,
1528 struct page *subpage, void *fsdata)
1529{
1530 struct folio *folio = page_folio(subpage);
1531 struct inode *inode = file_inode(file);
1532 struct ceph_client *cl = ceph_inode_to_client(inode);
1533 bool check_cap = false;
1534
1535 doutc(cl, "%llx.%llx file %p folio %p %d~%d (%d)\n", ceph_vinop(inode),
1536 file, folio, (int)pos, (int)copied, (int)len);
1537
1538 if (!folio_test_uptodate(folio)) {
1539 /* just return that nothing was copied on a short copy */
1540 if (copied < len) {
1541 copied = 0;
1542 goto out;
1543 }
1544 folio_mark_uptodate(folio);
1545 }
1546
1547 /* did file size increase? */
1548 if (pos+copied > i_size_read(inode))
1549 check_cap = ceph_inode_set_size(inode, pos+copied);
1550
1551 folio_mark_dirty(folio);
1552
1553out:
1554 folio_unlock(folio);
1555 folio_put(folio);
1556
1557 if (check_cap)
1558 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY);
1559
1560 return copied;
1561}
1562
1563const struct address_space_operations ceph_aops = {
1564 .read_folio = netfs_read_folio,
1565 .readahead = netfs_readahead,
1566 .writepage = ceph_writepage,
1567 .writepages = ceph_writepages_start,
1568 .write_begin = ceph_write_begin,
1569 .write_end = ceph_write_end,
1570 .dirty_folio = ceph_dirty_folio,
1571 .invalidate_folio = ceph_invalidate_folio,
1572 .release_folio = netfs_release_folio,
1573 .direct_IO = noop_direct_IO,
1574};
1575
1576static void ceph_block_sigs(sigset_t *oldset)
1577{
1578 sigset_t mask;
1579 siginitsetinv(&mask, sigmask(SIGKILL));
1580 sigprocmask(SIG_BLOCK, &mask, oldset);
1581}
1582
1583static void ceph_restore_sigs(sigset_t *oldset)
1584{
1585 sigprocmask(SIG_SETMASK, oldset, NULL);
1586}
1587
1588/*
1589 * vm ops
1590 */
1591static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1592{
1593 struct vm_area_struct *vma = vmf->vma;
1594 struct inode *inode = file_inode(vma->vm_file);
1595 struct ceph_inode_info *ci = ceph_inode(inode);
1596 struct ceph_client *cl = ceph_inode_to_client(inode);
1597 struct ceph_file_info *fi = vma->vm_file->private_data;
1598 loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1599 int want, got, err;
1600 sigset_t oldset;
1601 vm_fault_t ret = VM_FAULT_SIGBUS;
1602
1603 if (ceph_inode_is_shutdown(inode))
1604 return ret;
1605
1606 ceph_block_sigs(&oldset);
1607
1608 doutc(cl, "%llx.%llx %llu trying to get caps\n",
1609 ceph_vinop(inode), off);
1610 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1611 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1612 else
1613 want = CEPH_CAP_FILE_CACHE;
1614
1615 got = 0;
1616 err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1, &got);
1617 if (err < 0)
1618 goto out_restore;
1619
1620 doutc(cl, "%llx.%llx %llu got cap refs on %s\n", ceph_vinop(inode),
1621 off, ceph_cap_string(got));
1622
1623 if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1624 !ceph_has_inline_data(ci)) {
1625 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1626 ceph_add_rw_context(fi, &rw_ctx);
1627 ret = filemap_fault(vmf);
1628 ceph_del_rw_context(fi, &rw_ctx);
1629 doutc(cl, "%llx.%llx %llu drop cap refs %s ret %x\n",
1630 ceph_vinop(inode), off, ceph_cap_string(got), ret);
1631 } else
1632 err = -EAGAIN;
1633
1634 ceph_put_cap_refs(ci, got);
1635
1636 if (err != -EAGAIN)
1637 goto out_restore;
1638
1639 /* read inline data */
1640 if (off >= PAGE_SIZE) {
1641 /* does not support inline data > PAGE_SIZE */
1642 ret = VM_FAULT_SIGBUS;
1643 } else {
1644 struct address_space *mapping = inode->i_mapping;
1645 struct page *page;
1646
1647 filemap_invalidate_lock_shared(mapping);
1648 page = find_or_create_page(mapping, 0,
1649 mapping_gfp_constraint(mapping, ~__GFP_FS));
1650 if (!page) {
1651 ret = VM_FAULT_OOM;
1652 goto out_inline;
1653 }
1654 err = __ceph_do_getattr(inode, page,
1655 CEPH_STAT_CAP_INLINE_DATA, true);
1656 if (err < 0 || off >= i_size_read(inode)) {
1657 unlock_page(page);
1658 put_page(page);
1659 ret = vmf_error(err);
1660 goto out_inline;
1661 }
1662 if (err < PAGE_SIZE)
1663 zero_user_segment(page, err, PAGE_SIZE);
1664 else
1665 flush_dcache_page(page);
1666 SetPageUptodate(page);
1667 vmf->page = page;
1668 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1669out_inline:
1670 filemap_invalidate_unlock_shared(mapping);
1671 doutc(cl, "%llx.%llx %llu read inline data ret %x\n",
1672 ceph_vinop(inode), off, ret);
1673 }
1674out_restore:
1675 ceph_restore_sigs(&oldset);
1676 if (err < 0)
1677 ret = vmf_error(err);
1678
1679 return ret;
1680}
1681
1682static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1683{
1684 struct vm_area_struct *vma = vmf->vma;
1685 struct inode *inode = file_inode(vma->vm_file);
1686 struct ceph_client *cl = ceph_inode_to_client(inode);
1687 struct ceph_inode_info *ci = ceph_inode(inode);
1688 struct ceph_file_info *fi = vma->vm_file->private_data;
1689 struct ceph_cap_flush *prealloc_cf;
1690 struct page *page = vmf->page;
1691 loff_t off = page_offset(page);
1692 loff_t size = i_size_read(inode);
1693 size_t len;
1694 int want, got, err;
1695 sigset_t oldset;
1696 vm_fault_t ret = VM_FAULT_SIGBUS;
1697
1698 if (ceph_inode_is_shutdown(inode))
1699 return ret;
1700
1701 prealloc_cf = ceph_alloc_cap_flush();
1702 if (!prealloc_cf)
1703 return VM_FAULT_OOM;
1704
1705 sb_start_pagefault(inode->i_sb);
1706 ceph_block_sigs(&oldset);
1707
1708 if (off + thp_size(page) <= size)
1709 len = thp_size(page);
1710 else
1711 len = offset_in_thp(page, size);
1712
1713 doutc(cl, "%llx.%llx %llu~%zd getting caps i_size %llu\n",
1714 ceph_vinop(inode), off, len, size);
1715 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1716 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1717 else
1718 want = CEPH_CAP_FILE_BUFFER;
1719
1720 got = 0;
1721 err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len, &got);
1722 if (err < 0)
1723 goto out_free;
1724
1725 doutc(cl, "%llx.%llx %llu~%zd got cap refs on %s\n", ceph_vinop(inode),
1726 off, len, ceph_cap_string(got));
1727
1728 /* Update time before taking page lock */
1729 file_update_time(vma->vm_file);
1730 inode_inc_iversion_raw(inode);
1731
1732 do {
1733 struct ceph_snap_context *snapc;
1734
1735 lock_page(page);
1736
1737 if (page_mkwrite_check_truncate(page, inode) < 0) {
1738 unlock_page(page);
1739 ret = VM_FAULT_NOPAGE;
1740 break;
1741 }
1742
1743 snapc = ceph_find_incompatible(page);
1744 if (!snapc) {
1745 /* success. we'll keep the page locked. */
1746 set_page_dirty(page);
1747 ret = VM_FAULT_LOCKED;
1748 break;
1749 }
1750
1751 unlock_page(page);
1752
1753 if (IS_ERR(snapc)) {
1754 ret = VM_FAULT_SIGBUS;
1755 break;
1756 }
1757
1758 ceph_queue_writeback(inode);
1759 err = wait_event_killable(ci->i_cap_wq,
1760 context_is_writeable_or_written(inode, snapc));
1761 ceph_put_snap_context(snapc);
1762 } while (err == 0);
1763
1764 if (ret == VM_FAULT_LOCKED) {
1765 int dirty;
1766 spin_lock(&ci->i_ceph_lock);
1767 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1768 &prealloc_cf);
1769 spin_unlock(&ci->i_ceph_lock);
1770 if (dirty)
1771 __mark_inode_dirty(inode, dirty);
1772 }
1773
1774 doutc(cl, "%llx.%llx %llu~%zd dropping cap refs on %s ret %x\n",
1775 ceph_vinop(inode), off, len, ceph_cap_string(got), ret);
1776 ceph_put_cap_refs_async(ci, got);
1777out_free:
1778 ceph_restore_sigs(&oldset);
1779 sb_end_pagefault(inode->i_sb);
1780 ceph_free_cap_flush(prealloc_cf);
1781 if (err < 0)
1782 ret = vmf_error(err);
1783 return ret;
1784}
1785
1786void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1787 char *data, size_t len)
1788{
1789 struct ceph_client *cl = ceph_inode_to_client(inode);
1790 struct address_space *mapping = inode->i_mapping;
1791 struct page *page;
1792
1793 if (locked_page) {
1794 page = locked_page;
1795 } else {
1796 if (i_size_read(inode) == 0)
1797 return;
1798 page = find_or_create_page(mapping, 0,
1799 mapping_gfp_constraint(mapping,
1800 ~__GFP_FS));
1801 if (!page)
1802 return;
1803 if (PageUptodate(page)) {
1804 unlock_page(page);
1805 put_page(page);
1806 return;
1807 }
1808 }
1809
1810 doutc(cl, "%p %llx.%llx len %zu locked_page %p\n", inode,
1811 ceph_vinop(inode), len, locked_page);
1812
1813 if (len > 0) {
1814 void *kaddr = kmap_atomic(page);
1815 memcpy(kaddr, data, len);
1816 kunmap_atomic(kaddr);
1817 }
1818
1819 if (page != locked_page) {
1820 if (len < PAGE_SIZE)
1821 zero_user_segment(page, len, PAGE_SIZE);
1822 else
1823 flush_dcache_page(page);
1824
1825 SetPageUptodate(page);
1826 unlock_page(page);
1827 put_page(page);
1828 }
1829}
1830
1831int ceph_uninline_data(struct file *file)
1832{
1833 struct inode *inode = file_inode(file);
1834 struct ceph_inode_info *ci = ceph_inode(inode);
1835 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
1836 struct ceph_client *cl = fsc->client;
1837 struct ceph_osd_request *req = NULL;
1838 struct ceph_cap_flush *prealloc_cf = NULL;
1839 struct folio *folio = NULL;
1840 u64 inline_version = CEPH_INLINE_NONE;
1841 struct page *pages[1];
1842 int err = 0;
1843 u64 len;
1844
1845 spin_lock(&ci->i_ceph_lock);
1846 inline_version = ci->i_inline_version;
1847 spin_unlock(&ci->i_ceph_lock);
1848
1849 doutc(cl, "%llx.%llx inline_version %llu\n", ceph_vinop(inode),
1850 inline_version);
1851
1852 if (ceph_inode_is_shutdown(inode)) {
1853 err = -EIO;
1854 goto out;
1855 }
1856
1857 if (inline_version == CEPH_INLINE_NONE)
1858 return 0;
1859
1860 prealloc_cf = ceph_alloc_cap_flush();
1861 if (!prealloc_cf)
1862 return -ENOMEM;
1863
1864 if (inline_version == 1) /* initial version, no data */
1865 goto out_uninline;
1866
1867 folio = read_mapping_folio(inode->i_mapping, 0, file);
1868 if (IS_ERR(folio)) {
1869 err = PTR_ERR(folio);
1870 goto out;
1871 }
1872
1873 folio_lock(folio);
1874
1875 len = i_size_read(inode);
1876 if (len > folio_size(folio))
1877 len = folio_size(folio);
1878
1879 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1880 ceph_vino(inode), 0, &len, 0, 1,
1881 CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1882 NULL, 0, 0, false);
1883 if (IS_ERR(req)) {
1884 err = PTR_ERR(req);
1885 goto out_unlock;
1886 }
1887
1888 req->r_mtime = inode_get_mtime(inode);
1889 ceph_osdc_start_request(&fsc->client->osdc, req);
1890 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1891 ceph_osdc_put_request(req);
1892 if (err < 0)
1893 goto out_unlock;
1894
1895 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1896 ceph_vino(inode), 0, &len, 1, 3,
1897 CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1898 NULL, ci->i_truncate_seq,
1899 ci->i_truncate_size, false);
1900 if (IS_ERR(req)) {
1901 err = PTR_ERR(req);
1902 goto out_unlock;
1903 }
1904
1905 pages[0] = folio_page(folio, 0);
1906 osd_req_op_extent_osd_data_pages(req, 1, pages, len, 0, false, false);
1907
1908 {
1909 __le64 xattr_buf = cpu_to_le64(inline_version);
1910 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1911 "inline_version", &xattr_buf,
1912 sizeof(xattr_buf),
1913 CEPH_OSD_CMPXATTR_OP_GT,
1914 CEPH_OSD_CMPXATTR_MODE_U64);
1915 if (err)
1916 goto out_put_req;
1917 }
1918
1919 {
1920 char xattr_buf[32];
1921 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1922 "%llu", inline_version);
1923 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1924 "inline_version",
1925 xattr_buf, xattr_len, 0, 0);
1926 if (err)
1927 goto out_put_req;
1928 }
1929
1930 req->r_mtime = inode_get_mtime(inode);
1931 ceph_osdc_start_request(&fsc->client->osdc, req);
1932 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1933
1934 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1935 req->r_end_latency, len, err);
1936
1937out_uninline:
1938 if (!err) {
1939 int dirty;
1940
1941 /* Set to CAP_INLINE_NONE and dirty the caps */
1942 down_read(&fsc->mdsc->snap_rwsem);
1943 spin_lock(&ci->i_ceph_lock);
1944 ci->i_inline_version = CEPH_INLINE_NONE;
1945 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR, &prealloc_cf);
1946 spin_unlock(&ci->i_ceph_lock);
1947 up_read(&fsc->mdsc->snap_rwsem);
1948 if (dirty)
1949 __mark_inode_dirty(inode, dirty);
1950 }
1951out_put_req:
1952 ceph_osdc_put_request(req);
1953 if (err == -ECANCELED)
1954 err = 0;
1955out_unlock:
1956 if (folio) {
1957 folio_unlock(folio);
1958 folio_put(folio);
1959 }
1960out:
1961 ceph_free_cap_flush(prealloc_cf);
1962 doutc(cl, "%llx.%llx inline_version %llu = %d\n",
1963 ceph_vinop(inode), inline_version, err);
1964 return err;
1965}
1966
1967static const struct vm_operations_struct ceph_vmops = {
1968 .fault = ceph_filemap_fault,
1969 .page_mkwrite = ceph_page_mkwrite,
1970};
1971
1972int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1973{
1974 struct address_space *mapping = file->f_mapping;
1975
1976 if (!mapping->a_ops->read_folio)
1977 return -ENOEXEC;
1978 vma->vm_ops = &ceph_vmops;
1979 return 0;
1980}
1981
1982enum {
1983 POOL_READ = 1,
1984 POOL_WRITE = 2,
1985};
1986
1987static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1988 s64 pool, struct ceph_string *pool_ns)
1989{
1990 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(&ci->netfs.inode);
1991 struct ceph_mds_client *mdsc = fsc->mdsc;
1992 struct ceph_client *cl = fsc->client;
1993 struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1994 struct rb_node **p, *parent;
1995 struct ceph_pool_perm *perm;
1996 struct page **pages;
1997 size_t pool_ns_len;
1998 int err = 0, err2 = 0, have = 0;
1999
2000 down_read(&mdsc->pool_perm_rwsem);
2001 p = &mdsc->pool_perm_tree.rb_node;
2002 while (*p) {
2003 perm = rb_entry(*p, struct ceph_pool_perm, node);
2004 if (pool < perm->pool)
2005 p = &(*p)->rb_left;
2006 else if (pool > perm->pool)
2007 p = &(*p)->rb_right;
2008 else {
2009 int ret = ceph_compare_string(pool_ns,
2010 perm->pool_ns,
2011 perm->pool_ns_len);
2012 if (ret < 0)
2013 p = &(*p)->rb_left;
2014 else if (ret > 0)
2015 p = &(*p)->rb_right;
2016 else {
2017 have = perm->perm;
2018 break;
2019 }
2020 }
2021 }
2022 up_read(&mdsc->pool_perm_rwsem);
2023 if (*p)
2024 goto out;
2025
2026 if (pool_ns)
2027 doutc(cl, "pool %lld ns %.*s no perm cached\n", pool,
2028 (int)pool_ns->len, pool_ns->str);
2029 else
2030 doutc(cl, "pool %lld no perm cached\n", pool);
2031
2032 down_write(&mdsc->pool_perm_rwsem);
2033 p = &mdsc->pool_perm_tree.rb_node;
2034 parent = NULL;
2035 while (*p) {
2036 parent = *p;
2037 perm = rb_entry(parent, struct ceph_pool_perm, node);
2038 if (pool < perm->pool)
2039 p = &(*p)->rb_left;
2040 else if (pool > perm->pool)
2041 p = &(*p)->rb_right;
2042 else {
2043 int ret = ceph_compare_string(pool_ns,
2044 perm->pool_ns,
2045 perm->pool_ns_len);
2046 if (ret < 0)
2047 p = &(*p)->rb_left;
2048 else if (ret > 0)
2049 p = &(*p)->rb_right;
2050 else {
2051 have = perm->perm;
2052 break;
2053 }
2054 }
2055 }
2056 if (*p) {
2057 up_write(&mdsc->pool_perm_rwsem);
2058 goto out;
2059 }
2060
2061 rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2062 1, false, GFP_NOFS);
2063 if (!rd_req) {
2064 err = -ENOMEM;
2065 goto out_unlock;
2066 }
2067
2068 rd_req->r_flags = CEPH_OSD_FLAG_READ;
2069 osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
2070 rd_req->r_base_oloc.pool = pool;
2071 if (pool_ns)
2072 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
2073 ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
2074
2075 err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
2076 if (err)
2077 goto out_unlock;
2078
2079 wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2080 1, false, GFP_NOFS);
2081 if (!wr_req) {
2082 err = -ENOMEM;
2083 goto out_unlock;
2084 }
2085
2086 wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
2087 osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
2088 ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
2089 ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
2090
2091 err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
2092 if (err)
2093 goto out_unlock;
2094
2095 /* one page should be large enough for STAT data */
2096 pages = ceph_alloc_page_vector(1, GFP_KERNEL);
2097 if (IS_ERR(pages)) {
2098 err = PTR_ERR(pages);
2099 goto out_unlock;
2100 }
2101
2102 osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
2103 0, false, true);
2104 ceph_osdc_start_request(&fsc->client->osdc, rd_req);
2105
2106 wr_req->r_mtime = inode_get_mtime(&ci->netfs.inode);
2107 ceph_osdc_start_request(&fsc->client->osdc, wr_req);
2108
2109 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
2110 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
2111
2112 if (err >= 0 || err == -ENOENT)
2113 have |= POOL_READ;
2114 else if (err != -EPERM) {
2115 if (err == -EBLOCKLISTED)
2116 fsc->blocklisted = true;
2117 goto out_unlock;
2118 }
2119
2120 if (err2 == 0 || err2 == -EEXIST)
2121 have |= POOL_WRITE;
2122 else if (err2 != -EPERM) {
2123 if (err2 == -EBLOCKLISTED)
2124 fsc->blocklisted = true;
2125 err = err2;
2126 goto out_unlock;
2127 }
2128
2129 pool_ns_len = pool_ns ? pool_ns->len : 0;
2130 perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
2131 if (!perm) {
2132 err = -ENOMEM;
2133 goto out_unlock;
2134 }
2135
2136 perm->pool = pool;
2137 perm->perm = have;
2138 perm->pool_ns_len = pool_ns_len;
2139 if (pool_ns_len > 0)
2140 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
2141 perm->pool_ns[pool_ns_len] = 0;
2142
2143 rb_link_node(&perm->node, parent, p);
2144 rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
2145 err = 0;
2146out_unlock:
2147 up_write(&mdsc->pool_perm_rwsem);
2148
2149 ceph_osdc_put_request(rd_req);
2150 ceph_osdc_put_request(wr_req);
2151out:
2152 if (!err)
2153 err = have;
2154 if (pool_ns)
2155 doutc(cl, "pool %lld ns %.*s result = %d\n", pool,
2156 (int)pool_ns->len, pool_ns->str, err);
2157 else
2158 doutc(cl, "pool %lld result = %d\n", pool, err);
2159 return err;
2160}
2161
2162int ceph_pool_perm_check(struct inode *inode, int need)
2163{
2164 struct ceph_client *cl = ceph_inode_to_client(inode);
2165 struct ceph_inode_info *ci = ceph_inode(inode);
2166 struct ceph_string *pool_ns;
2167 s64 pool;
2168 int ret, flags;
2169
2170 /* Only need to do this for regular files */
2171 if (!S_ISREG(inode->i_mode))
2172 return 0;
2173
2174 if (ci->i_vino.snap != CEPH_NOSNAP) {
2175 /*
2176 * Pool permission check needs to write to the first object.
2177 * But for snapshot, head of the first object may have alread
2178 * been deleted. Skip check to avoid creating orphan object.
2179 */
2180 return 0;
2181 }
2182
2183 if (ceph_test_mount_opt(ceph_inode_to_fs_client(inode),
2184 NOPOOLPERM))
2185 return 0;
2186
2187 spin_lock(&ci->i_ceph_lock);
2188 flags = ci->i_ceph_flags;
2189 pool = ci->i_layout.pool_id;
2190 spin_unlock(&ci->i_ceph_lock);
2191check:
2192 if (flags & CEPH_I_POOL_PERM) {
2193 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2194 doutc(cl, "pool %lld no read perm\n", pool);
2195 return -EPERM;
2196 }
2197 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2198 doutc(cl, "pool %lld no write perm\n", pool);
2199 return -EPERM;
2200 }
2201 return 0;
2202 }
2203
2204 pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2205 ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2206 ceph_put_string(pool_ns);
2207 if (ret < 0)
2208 return ret;
2209
2210 flags = CEPH_I_POOL_PERM;
2211 if (ret & POOL_READ)
2212 flags |= CEPH_I_POOL_RD;
2213 if (ret & POOL_WRITE)
2214 flags |= CEPH_I_POOL_WR;
2215
2216 spin_lock(&ci->i_ceph_lock);
2217 if (pool == ci->i_layout.pool_id &&
2218 pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2219 ci->i_ceph_flags |= flags;
2220 } else {
2221 pool = ci->i_layout.pool_id;
2222 flags = ci->i_ceph_flags;
2223 }
2224 spin_unlock(&ci->i_ceph_lock);
2225 goto check;
2226}
2227
2228void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2229{
2230 struct ceph_pool_perm *perm;
2231 struct rb_node *n;
2232
2233 while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2234 n = rb_first(&mdsc->pool_perm_tree);
2235 perm = rb_entry(n, struct ceph_pool_perm, node);
2236 rb_erase(n, &mdsc->pool_perm_tree);
2237 kfree(perm);
2238 }
2239}