1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
4 #include <linux/backing-dev.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>
18 #include "mds_client.h"
22 #include <linux/ceph/osd_client.h>
23 #include <linux/ceph/striper.h>
26 * Ceph address space ops.
28 * There are a few funny things going on here.
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.
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.
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.
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
57 * Invalidate and so forth must take care to ensure the dirty page
58 * accounting is preserved.
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))
66 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
67 struct folio **foliop, void **_fsdata);
69 static inline struct ceph_snap_context *page_snap_context(struct page *page)
71 if (PagePrivate(page))
72 return (void *)page->private;
77 * Dirty a page. Optimistically adjust accounting, on the assumption
78 * that we won't race with invalidate. If we do, readjust.
80 static bool ceph_dirty_folio(struct address_space *mapping, struct folio *folio)
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;
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);
94 ci = ceph_inode(inode);
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,
104 snapc = ceph_get_snap_context(capsnap->context);
105 capsnap->dirty_pages++;
107 BUG_ON(!ci->i_head_snapc);
108 snapc = ceph_get_snap_context(ci->i_head_snapc);
109 ++ci->i_wrbuffer_ref_head;
111 if (ci->i_wrbuffer_ref == 0)
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);
123 * Reference snap context in folio->private. Also set
124 * PagePrivate so that we get invalidate_folio callback.
126 VM_WARN_ON_FOLIO(folio->private, folio);
127 folio_attach_private(folio, snapc);
129 return ceph_fscache_dirty_folio(mapping, folio);
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
137 static void ceph_invalidate_folio(struct folio *folio, size_t offset,
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;
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);
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);
157 snapc = folio_detach_private(folio);
158 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
159 ceph_put_snap_context(snapc);
162 folio_wait_fscache(folio);
165 static bool ceph_release_folio(struct folio *folio, gfp_t gfp)
167 struct inode *inode = folio->mapping->host;
168 struct ceph_client *cl = ceph_inode_to_client(inode);
170 doutc(cl, "%llx.%llx idx %lu (%sdirty)\n", ceph_vinop(inode),
171 folio->index, folio_test_dirty(folio) ? "" : "not ");
173 if (folio_test_private(folio))
176 if (folio_test_fscache(folio)) {
177 if (current_is_kswapd() || !(gfp & __GFP_FS))
179 folio_wait_fscache(folio);
181 ceph_fscache_note_page_release(inode);
185 static void ceph_netfs_expand_readahead(struct netfs_io_request *rreq)
187 struct inode *inode = rreq->inode;
188 struct ceph_inode_info *ci = ceph_inode(inode);
189 struct ceph_file_layout *lo = &ci->i_layout;
190 unsigned long max_pages = inode->i_sb->s_bdi->ra_pages;
191 loff_t end = rreq->start + rreq->len, new_end;
192 struct ceph_netfs_request_data *priv = rreq->netfs_priv;
193 unsigned long max_len;
197 /* Readahead is disabled by posix_fadvise POSIX_FADV_RANDOM */
198 if (priv->file_ra_disabled)
201 max_pages = priv->file_ra_pages;
205 /* Readahead is disabled */
209 max_len = max_pages << PAGE_SHIFT;
212 * Try to expand the length forward by rounding up it to the next
213 * block, but do not exceed the file size, unless the original
214 * request already exceeds it.
216 new_end = min(round_up(end, lo->stripe_unit), rreq->i_size);
217 if (new_end > end && new_end <= rreq->start + max_len)
218 rreq->len = new_end - rreq->start;
220 /* Try to expand the start downward */
221 div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
222 if (rreq->len + blockoff <= max_len) {
223 rreq->start -= blockoff;
224 rreq->len += blockoff;
228 static bool ceph_netfs_clamp_length(struct netfs_io_subrequest *subreq)
230 struct inode *inode = subreq->rreq->inode;
231 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
232 struct ceph_inode_info *ci = ceph_inode(inode);
236 /* Truncate the extent at the end of the current block */
237 ceph_calc_file_object_mapping(&ci->i_layout, subreq->start, subreq->len,
238 &objno, &objoff, &xlen);
239 subreq->len = min(xlen, fsc->mount_options->rsize);
243 static void finish_netfs_read(struct ceph_osd_request *req)
245 struct inode *inode = req->r_inode;
246 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
247 struct ceph_client *cl = fsc->client;
248 struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
249 struct netfs_io_subrequest *subreq = req->r_priv;
250 struct ceph_osd_req_op *op = &req->r_ops[0];
251 int err = req->r_result;
252 bool sparse = (op->op == CEPH_OSD_OP_SPARSE_READ);
254 ceph_update_read_metrics(&fsc->mdsc->metric, req->r_start_latency,
255 req->r_end_latency, osd_data->length, err);
257 doutc(cl, "result %d subreq->len=%zu i_size=%lld\n", req->r_result,
258 subreq->len, i_size_read(req->r_inode));
260 /* no object means success but no data */
263 else if (err == -EBLOCKLISTED)
264 fsc->blocklisted = true;
267 if (sparse && err > 0)
268 err = ceph_sparse_ext_map_end(op);
269 if (err < subreq->len)
270 __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
271 if (IS_ENCRYPTED(inode) && err > 0) {
272 err = ceph_fscrypt_decrypt_extents(inode,
273 osd_data->pages, subreq->start,
274 op->extent.sparse_ext,
275 op->extent.sparse_ext_cnt);
276 if (err > subreq->len)
281 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
282 ceph_put_page_vector(osd_data->pages,
283 calc_pages_for(osd_data->alignment,
284 osd_data->length), false);
286 netfs_subreq_terminated(subreq, err, false);
288 ceph_dec_osd_stopping_blocker(fsc->mdsc);
291 static bool ceph_netfs_issue_op_inline(struct netfs_io_subrequest *subreq)
293 struct netfs_io_request *rreq = subreq->rreq;
294 struct inode *inode = rreq->inode;
295 struct ceph_mds_reply_info_parsed *rinfo;
296 struct ceph_mds_reply_info_in *iinfo;
297 struct ceph_mds_request *req;
298 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
299 struct ceph_inode_info *ci = ceph_inode(inode);
300 struct iov_iter iter;
305 __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
306 __clear_bit(NETFS_SREQ_COPY_TO_CACHE, &subreq->flags);
308 if (subreq->start >= inode->i_size)
311 /* We need to fetch the inline data. */
312 mode = ceph_try_to_choose_auth_mds(inode, CEPH_STAT_CAP_INLINE_DATA);
313 req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_GETATTR, mode);
318 req->r_ino1 = ci->i_vino;
319 req->r_args.getattr.mask = cpu_to_le32(CEPH_STAT_CAP_INLINE_DATA);
322 err = ceph_mdsc_do_request(mdsc, NULL, req);
326 rinfo = &req->r_reply_info;
327 iinfo = &rinfo->targeti;
328 if (iinfo->inline_version == CEPH_INLINE_NONE) {
329 /* The data got uninlined */
330 ceph_mdsc_put_request(req);
334 len = min_t(size_t, iinfo->inline_len - subreq->start, subreq->len);
335 iov_iter_xarray(&iter, ITER_DEST, &rreq->mapping->i_pages, subreq->start, len);
336 err = copy_to_iter(iinfo->inline_data + subreq->start, len, &iter);
340 ceph_mdsc_put_request(req);
342 netfs_subreq_terminated(subreq, err, false);
346 static void ceph_netfs_issue_read(struct netfs_io_subrequest *subreq)
348 struct netfs_io_request *rreq = subreq->rreq;
349 struct inode *inode = rreq->inode;
350 struct ceph_inode_info *ci = ceph_inode(inode);
351 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
352 struct ceph_client *cl = fsc->client;
353 struct ceph_osd_request *req = NULL;
354 struct ceph_vino vino = ceph_vino(inode);
355 struct iov_iter iter;
357 u64 len = subreq->len;
358 bool sparse = IS_ENCRYPTED(inode) || ceph_test_mount_opt(fsc, SPARSEREAD);
359 u64 off = subreq->start;
361 if (ceph_inode_is_shutdown(inode)) {
366 if (ceph_has_inline_data(ci) && ceph_netfs_issue_op_inline(subreq))
369 ceph_fscrypt_adjust_off_and_len(inode, &off, &len);
371 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, vino,
372 off, &len, 0, 1, sparse ? CEPH_OSD_OP_SPARSE_READ : CEPH_OSD_OP_READ,
373 CEPH_OSD_FLAG_READ | fsc->client->osdc.client->options->read_from_replica,
374 NULL, ci->i_truncate_seq, ci->i_truncate_size, false);
382 err = ceph_alloc_sparse_ext_map(&req->r_ops[0]);
387 doutc(cl, "%llx.%llx pos=%llu orig_len=%zu len=%llu\n",
388 ceph_vinop(inode), subreq->start, subreq->len, len);
390 iov_iter_xarray(&iter, ITER_DEST, &rreq->mapping->i_pages, subreq->start, len);
393 * FIXME: For now, use CEPH_OSD_DATA_TYPE_PAGES instead of _ITER for
394 * encrypted inodes. We'd need infrastructure that handles an iov_iter
395 * instead of page arrays, and we don't have that as of yet. Once the
396 * dust settles on the write helpers and encrypt/decrypt routines for
397 * netfs, we should be able to rework this.
399 if (IS_ENCRYPTED(inode)) {
403 err = iov_iter_get_pages_alloc2(&iter, &pages, len, &page_off);
405 doutc(cl, "%llx.%llx failed to allocate pages, %d\n",
406 ceph_vinop(inode), err);
410 /* should always give us a page-aligned read */
411 WARN_ON_ONCE(page_off);
415 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false,
418 osd_req_op_extent_osd_iter(req, 0, &iter);
420 if (!ceph_inc_osd_stopping_blocker(fsc->mdsc)) {
424 req->r_callback = finish_netfs_read;
425 req->r_priv = subreq;
426 req->r_inode = inode;
429 ceph_osdc_start_request(req->r_osdc, req);
431 ceph_osdc_put_request(req);
433 netfs_subreq_terminated(subreq, err, false);
434 doutc(cl, "%llx.%llx result %d\n", ceph_vinop(inode), err);
437 static int ceph_init_request(struct netfs_io_request *rreq, struct file *file)
439 struct inode *inode = rreq->inode;
440 struct ceph_client *cl = ceph_inode_to_client(inode);
441 int got = 0, want = CEPH_CAP_FILE_CACHE;
442 struct ceph_netfs_request_data *priv;
445 if (rreq->origin != NETFS_READAHEAD)
448 priv = kzalloc(sizeof(*priv), GFP_NOFS);
453 struct ceph_rw_context *rw_ctx;
454 struct ceph_file_info *fi = file->private_data;
456 priv->file_ra_pages = file->f_ra.ra_pages;
457 priv->file_ra_disabled = file->f_mode & FMODE_RANDOM;
459 rw_ctx = ceph_find_rw_context(fi);
461 rreq->netfs_priv = priv;
467 * readahead callers do not necessarily hold Fcb caps
468 * (e.g. fadvise, madvise).
470 ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want, true, &got);
472 doutc(cl, "%llx.%llx, error getting cap\n", ceph_vinop(inode));
477 doutc(cl, "%llx.%llx, no cache cap\n", ceph_vinop(inode));
487 rreq->netfs_priv = priv;
496 static void ceph_netfs_free_request(struct netfs_io_request *rreq)
498 struct ceph_netfs_request_data *priv = rreq->netfs_priv;
504 ceph_put_cap_refs(ceph_inode(rreq->inode), priv->caps);
506 rreq->netfs_priv = NULL;
509 const struct netfs_request_ops ceph_netfs_ops = {
510 .init_request = ceph_init_request,
511 .free_request = ceph_netfs_free_request,
512 .begin_cache_operation = ceph_begin_cache_operation,
513 .issue_read = ceph_netfs_issue_read,
514 .expand_readahead = ceph_netfs_expand_readahead,
515 .clamp_length = ceph_netfs_clamp_length,
516 .check_write_begin = ceph_netfs_check_write_begin,
519 #ifdef CONFIG_CEPH_FSCACHE
520 static void ceph_set_page_fscache(struct page *page)
522 set_page_fscache(page);
525 static void ceph_fscache_write_terminated(void *priv, ssize_t error, bool was_async)
527 struct inode *inode = priv;
529 if (IS_ERR_VALUE(error) && error != -ENOBUFS)
530 ceph_fscache_invalidate(inode, false);
533 static void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
535 struct ceph_inode_info *ci = ceph_inode(inode);
536 struct fscache_cookie *cookie = ceph_fscache_cookie(ci);
538 fscache_write_to_cache(cookie, inode->i_mapping, off, len, i_size_read(inode),
539 ceph_fscache_write_terminated, inode, caching);
542 static inline void ceph_set_page_fscache(struct page *page)
546 static inline void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
549 #endif /* CONFIG_CEPH_FSCACHE */
551 struct ceph_writeback_ctl
561 * Get ref for the oldest snapc for an inode with dirty data... that is, the
562 * only snap context we are allowed to write back.
564 static struct ceph_snap_context *
565 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
566 struct ceph_snap_context *page_snapc)
568 struct ceph_inode_info *ci = ceph_inode(inode);
569 struct ceph_client *cl = ceph_inode_to_client(inode);
570 struct ceph_snap_context *snapc = NULL;
571 struct ceph_cap_snap *capsnap = NULL;
573 spin_lock(&ci->i_ceph_lock);
574 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
575 doutc(cl, " capsnap %p snapc %p has %d dirty pages\n",
576 capsnap, capsnap->context, capsnap->dirty_pages);
577 if (!capsnap->dirty_pages)
580 /* get i_size, truncate_{seq,size} for page_snapc? */
581 if (snapc && capsnap->context != page_snapc)
585 if (capsnap->writing) {
586 ctl->i_size = i_size_read(inode);
587 ctl->size_stable = false;
589 ctl->i_size = capsnap->size;
590 ctl->size_stable = true;
592 ctl->truncate_size = capsnap->truncate_size;
593 ctl->truncate_seq = capsnap->truncate_seq;
594 ctl->head_snapc = false;
600 snapc = ceph_get_snap_context(capsnap->context);
602 page_snapc == snapc ||
603 page_snapc->seq > snapc->seq)
606 if (!snapc && ci->i_wrbuffer_ref_head) {
607 snapc = ceph_get_snap_context(ci->i_head_snapc);
608 doutc(cl, " head snapc %p has %d dirty pages\n", snapc,
609 ci->i_wrbuffer_ref_head);
611 ctl->i_size = i_size_read(inode);
612 ctl->truncate_size = ci->i_truncate_size;
613 ctl->truncate_seq = ci->i_truncate_seq;
614 ctl->size_stable = false;
615 ctl->head_snapc = true;
618 spin_unlock(&ci->i_ceph_lock);
622 static u64 get_writepages_data_length(struct inode *inode,
623 struct page *page, u64 start)
625 struct ceph_inode_info *ci = ceph_inode(inode);
626 struct ceph_snap_context *snapc;
627 struct ceph_cap_snap *capsnap = NULL;
628 u64 end = i_size_read(inode);
631 snapc = page_snap_context(ceph_fscrypt_pagecache_page(page));
632 if (snapc != ci->i_head_snapc) {
634 spin_lock(&ci->i_ceph_lock);
635 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
636 if (capsnap->context == snapc) {
637 if (!capsnap->writing)
643 spin_unlock(&ci->i_ceph_lock);
646 if (end > ceph_fscrypt_page_offset(page) + thp_size(page))
647 end = ceph_fscrypt_page_offset(page) + thp_size(page);
648 ret = end > start ? end - start : 0;
649 if (ret && fscrypt_is_bounce_page(page))
650 ret = round_up(ret, CEPH_FSCRYPT_BLOCK_SIZE);
655 * Write a single page, but leave the page locked.
657 * If we get a write error, mark the mapping for error, but still adjust the
658 * dirty page accounting (i.e., page is no longer dirty).
660 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
662 struct folio *folio = page_folio(page);
663 struct inode *inode = page->mapping->host;
664 struct ceph_inode_info *ci = ceph_inode(inode);
665 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
666 struct ceph_client *cl = fsc->client;
667 struct ceph_snap_context *snapc, *oldest;
668 loff_t page_off = page_offset(page);
670 loff_t len = thp_size(page);
672 struct ceph_writeback_ctl ceph_wbc;
673 struct ceph_osd_client *osdc = &fsc->client->osdc;
674 struct ceph_osd_request *req;
675 bool caching = ceph_is_cache_enabled(inode);
676 struct page *bounce_page = NULL;
678 doutc(cl, "%llx.%llx page %p idx %lu\n", ceph_vinop(inode), page,
681 if (ceph_inode_is_shutdown(inode))
684 /* verify this is a writeable snap context */
685 snapc = page_snap_context(page);
687 doutc(cl, "%llx.%llx page %p not dirty?\n", ceph_vinop(inode),
691 oldest = get_oldest_context(inode, &ceph_wbc, snapc);
692 if (snapc->seq > oldest->seq) {
693 doutc(cl, "%llx.%llx page %p snapc %p not writeable - noop\n",
694 ceph_vinop(inode), page, snapc);
695 /* we should only noop if called by kswapd */
696 WARN_ON(!(current->flags & PF_MEMALLOC));
697 ceph_put_snap_context(oldest);
698 redirty_page_for_writepage(wbc, page);
701 ceph_put_snap_context(oldest);
703 /* is this a partial page at end of file? */
704 if (page_off >= ceph_wbc.i_size) {
705 doutc(cl, "%llx.%llx folio at %lu beyond eof %llu\n",
706 ceph_vinop(inode), folio->index, ceph_wbc.i_size);
707 folio_invalidate(folio, 0, folio_size(folio));
711 if (ceph_wbc.i_size < page_off + len)
712 len = ceph_wbc.i_size - page_off;
714 wlen = IS_ENCRYPTED(inode) ? round_up(len, CEPH_FSCRYPT_BLOCK_SIZE) : len;
715 doutc(cl, "%llx.%llx page %p index %lu on %llu~%llu snapc %p seq %lld\n",
716 ceph_vinop(inode), page, page->index, page_off, wlen, snapc,
719 if (atomic_long_inc_return(&fsc->writeback_count) >
720 CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
721 fsc->write_congested = true;
723 req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode),
724 page_off, &wlen, 0, 1, CEPH_OSD_OP_WRITE,
725 CEPH_OSD_FLAG_WRITE, snapc,
726 ceph_wbc.truncate_seq,
727 ceph_wbc.truncate_size, true);
729 redirty_page_for_writepage(wbc, page);
736 set_page_writeback(page);
738 ceph_set_page_fscache(page);
739 ceph_fscache_write_to_cache(inode, page_off, len, caching);
741 if (IS_ENCRYPTED(inode)) {
742 bounce_page = fscrypt_encrypt_pagecache_blocks(page,
743 CEPH_FSCRYPT_BLOCK_SIZE, 0,
745 if (IS_ERR(bounce_page)) {
746 redirty_page_for_writepage(wbc, page);
747 end_page_writeback(page);
748 ceph_osdc_put_request(req);
749 return PTR_ERR(bounce_page);
753 /* it may be a short write due to an object boundary */
754 WARN_ON_ONCE(len > thp_size(page));
755 osd_req_op_extent_osd_data_pages(req, 0,
756 bounce_page ? &bounce_page : &page, wlen, 0,
758 doutc(cl, "%llx.%llx %llu~%llu (%llu bytes, %sencrypted)\n",
759 ceph_vinop(inode), page_off, len, wlen,
760 IS_ENCRYPTED(inode) ? "" : "not ");
762 req->r_mtime = inode_get_mtime(inode);
763 ceph_osdc_start_request(osdc, req);
764 err = ceph_osdc_wait_request(osdc, req);
766 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
767 req->r_end_latency, len, err);
768 fscrypt_free_bounce_page(bounce_page);
769 ceph_osdc_put_request(req);
774 struct writeback_control tmp_wbc;
777 if (err == -ERESTARTSYS) {
778 /* killed by SIGKILL */
779 doutc(cl, "%llx.%llx interrupted page %p\n",
780 ceph_vinop(inode), page);
781 redirty_page_for_writepage(wbc, page);
782 end_page_writeback(page);
785 if (err == -EBLOCKLISTED)
786 fsc->blocklisted = true;
787 doutc(cl, "%llx.%llx setting page/mapping error %d %p\n",
788 ceph_vinop(inode), err, page);
789 mapping_set_error(&inode->i_data, err);
790 wbc->pages_skipped++;
792 doutc(cl, "%llx.%llx cleaned page %p\n",
793 ceph_vinop(inode), page);
794 err = 0; /* vfs expects us to return 0 */
796 oldest = detach_page_private(page);
797 WARN_ON_ONCE(oldest != snapc);
798 end_page_writeback(page);
799 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
800 ceph_put_snap_context(snapc); /* page's reference */
802 if (atomic_long_dec_return(&fsc->writeback_count) <
803 CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
804 fsc->write_congested = false;
809 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
812 struct inode *inode = page->mapping->host;
816 if (wbc->sync_mode == WB_SYNC_NONE &&
817 ceph_inode_to_fs_client(inode)->write_congested)
818 return AOP_WRITEPAGE_ACTIVATE;
820 wait_on_page_fscache(page);
822 err = writepage_nounlock(page, wbc);
823 if (err == -ERESTARTSYS) {
824 /* direct memory reclaimer was killed by SIGKILL. return 0
825 * to prevent caller from setting mapping/page error */
834 * async writeback completion handler.
836 * If we get an error, set the mapping error bit, but not the individual
839 static void writepages_finish(struct ceph_osd_request *req)
841 struct inode *inode = req->r_inode;
842 struct ceph_inode_info *ci = ceph_inode(inode);
843 struct ceph_client *cl = ceph_inode_to_client(inode);
844 struct ceph_osd_data *osd_data;
846 int num_pages, total_pages = 0;
848 int rc = req->r_result;
849 struct ceph_snap_context *snapc = req->r_snapc;
850 struct address_space *mapping = inode->i_mapping;
851 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
852 unsigned int len = 0;
855 doutc(cl, "%llx.%llx rc %d\n", ceph_vinop(inode), rc);
857 mapping_set_error(mapping, rc);
858 ceph_set_error_write(ci);
859 if (rc == -EBLOCKLISTED)
860 fsc->blocklisted = true;
862 ceph_clear_error_write(ci);
866 * We lost the cache cap, need to truncate the page before
867 * it is unlocked, otherwise we'd truncate it later in the
868 * page truncation thread, possibly losing some data that
871 remove_page = !(ceph_caps_issued(ci) &
872 (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
874 /* clean all pages */
875 for (i = 0; i < req->r_num_ops; i++) {
876 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE) {
878 "%llx.%llx incorrect op %d req %p index %d tid %llu\n",
879 ceph_vinop(inode), req->r_ops[i].op, req, i,
884 osd_data = osd_req_op_extent_osd_data(req, i);
885 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
886 len += osd_data->length;
887 num_pages = calc_pages_for((u64)osd_data->alignment,
888 (u64)osd_data->length);
889 total_pages += num_pages;
890 for (j = 0; j < num_pages; j++) {
891 page = osd_data->pages[j];
892 if (fscrypt_is_bounce_page(page)) {
893 page = fscrypt_pagecache_page(page);
894 fscrypt_free_bounce_page(osd_data->pages[j]);
895 osd_data->pages[j] = page;
898 WARN_ON(!PageUptodate(page));
900 if (atomic_long_dec_return(&fsc->writeback_count) <
901 CONGESTION_OFF_THRESH(
902 fsc->mount_options->congestion_kb))
903 fsc->write_congested = false;
905 ceph_put_snap_context(detach_page_private(page));
906 end_page_writeback(page);
907 doutc(cl, "unlocking %p\n", page);
910 generic_error_remove_page(inode->i_mapping,
915 doutc(cl, "%llx.%llx wrote %llu bytes cleaned %d pages\n",
916 ceph_vinop(inode), osd_data->length,
917 rc >= 0 ? num_pages : 0);
919 release_pages(osd_data->pages, num_pages);
922 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
923 req->r_end_latency, len, rc);
925 ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
927 osd_data = osd_req_op_extent_osd_data(req, 0);
928 if (osd_data->pages_from_pool)
929 mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
931 kfree(osd_data->pages);
932 ceph_osdc_put_request(req);
933 ceph_dec_osd_stopping_blocker(fsc->mdsc);
937 * initiate async writeback
939 static int ceph_writepages_start(struct address_space *mapping,
940 struct writeback_control *wbc)
942 struct inode *inode = mapping->host;
943 struct ceph_inode_info *ci = ceph_inode(inode);
944 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
945 struct ceph_client *cl = fsc->client;
946 struct ceph_vino vino = ceph_vino(inode);
947 pgoff_t index, start_index, end = -1;
948 struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
949 struct folio_batch fbatch;
951 unsigned int wsize = i_blocksize(inode);
952 struct ceph_osd_request *req = NULL;
953 struct ceph_writeback_ctl ceph_wbc;
954 bool should_loop, range_whole = false;
956 bool caching = ceph_is_cache_enabled(inode);
959 if (wbc->sync_mode == WB_SYNC_NONE &&
960 fsc->write_congested)
963 doutc(cl, "%llx.%llx (mode=%s)\n", ceph_vinop(inode),
964 wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
965 (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
967 if (ceph_inode_is_shutdown(inode)) {
968 if (ci->i_wrbuffer_ref > 0) {
969 pr_warn_ratelimited_client(cl,
970 "%llx.%llx %lld forced umount\n",
971 ceph_vinop(inode), ceph_ino(inode));
973 mapping_set_error(mapping, -EIO);
974 return -EIO; /* we're in a forced umount, don't write! */
976 if (fsc->mount_options->wsize < wsize)
977 wsize = fsc->mount_options->wsize;
979 folio_batch_init(&fbatch);
981 start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
984 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
985 tag = PAGECACHE_TAG_TOWRITE;
987 tag = PAGECACHE_TAG_DIRTY;
990 /* find oldest snap context with dirty data */
991 snapc = get_oldest_context(inode, &ceph_wbc, NULL);
993 /* hmm, why does writepages get called when there
995 doutc(cl, " no snap context with dirty data?\n");
998 doutc(cl, " oldest snapc is %p seq %lld (%d snaps)\n", snapc,
999 snapc->seq, snapc->num_snaps);
1001 should_loop = false;
1002 if (ceph_wbc.head_snapc && snapc != last_snapc) {
1003 /* where to start/end? */
1004 if (wbc->range_cyclic) {
1005 index = start_index;
1009 doutc(cl, " cyclic, start at %lu\n", index);
1011 index = wbc->range_start >> PAGE_SHIFT;
1012 end = wbc->range_end >> PAGE_SHIFT;
1013 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1015 doutc(cl, " not cyclic, %lu to %lu\n", index, end);
1017 } else if (!ceph_wbc.head_snapc) {
1018 /* Do not respect wbc->range_{start,end}. Dirty pages
1019 * in that range can be associated with newer snapc.
1020 * They are not writeable until we write all dirty pages
1021 * associated with 'snapc' get written */
1024 doutc(cl, " non-head snapc, range whole\n");
1027 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1028 tag_pages_for_writeback(mapping, index, end);
1030 ceph_put_snap_context(last_snapc);
1033 while (!done && index <= end) {
1034 int num_ops = 0, op_idx;
1035 unsigned i, nr_folios, max_pages, locked_pages = 0;
1036 struct page **pages = NULL, **data_pages;
1038 pgoff_t strip_unit_end = 0;
1039 u64 offset = 0, len = 0;
1040 bool from_pool = false;
1042 max_pages = wsize >> PAGE_SHIFT;
1045 nr_folios = filemap_get_folios_tag(mapping, &index,
1047 doutc(cl, "pagevec_lookup_range_tag got %d\n", nr_folios);
1048 if (!nr_folios && !locked_pages)
1050 for (i = 0; i < nr_folios && locked_pages < max_pages; i++) {
1051 page = &fbatch.folios[i]->page;
1052 doutc(cl, "? %p idx %lu\n", page, page->index);
1053 if (locked_pages == 0)
1054 lock_page(page); /* first page */
1055 else if (!trylock_page(page))
1058 /* only dirty pages, or our accounting breaks */
1059 if (unlikely(!PageDirty(page)) ||
1060 unlikely(page->mapping != mapping)) {
1061 doutc(cl, "!dirty or !mapping %p\n", page);
1065 /* only if matching snap context */
1066 pgsnapc = page_snap_context(page);
1067 if (pgsnapc != snapc) {
1068 doutc(cl, "page snapc %p %lld != oldest %p %lld\n",
1069 pgsnapc, pgsnapc->seq, snapc, snapc->seq);
1071 !ceph_wbc.head_snapc &&
1072 wbc->sync_mode != WB_SYNC_NONE)
1077 if (page_offset(page) >= ceph_wbc.i_size) {
1078 struct folio *folio = page_folio(page);
1080 doutc(cl, "folio at %lu beyond eof %llu\n",
1081 folio->index, ceph_wbc.i_size);
1082 if ((ceph_wbc.size_stable ||
1083 folio_pos(folio) >= i_size_read(inode)) &&
1084 folio_clear_dirty_for_io(folio))
1085 folio_invalidate(folio, 0,
1087 folio_unlock(folio);
1090 if (strip_unit_end && (page->index > strip_unit_end)) {
1091 doutc(cl, "end of strip unit %p\n", page);
1095 if (PageWriteback(page) || PageFsCache(page)) {
1096 if (wbc->sync_mode == WB_SYNC_NONE) {
1097 doutc(cl, "%p under writeback\n", page);
1101 doutc(cl, "waiting on writeback %p\n", page);
1102 wait_on_page_writeback(page);
1103 wait_on_page_fscache(page);
1106 if (!clear_page_dirty_for_io(page)) {
1107 doutc(cl, "%p !clear_page_dirty_for_io\n", page);
1113 * We have something to write. If this is
1114 * the first locked page this time through,
1115 * calculate max possinle write size and
1116 * allocate a page array
1118 if (locked_pages == 0) {
1123 /* prepare async write request */
1124 offset = (u64)page_offset(page);
1125 ceph_calc_file_object_mapping(&ci->i_layout,
1132 strip_unit_end = page->index +
1133 ((len - 1) >> PAGE_SHIFT);
1136 max_pages = calc_pages_for(0, (u64)len);
1137 pages = kmalloc_array(max_pages,
1142 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1147 } else if (page->index !=
1148 (offset + len) >> PAGE_SHIFT) {
1149 if (num_ops >= (from_pool ? CEPH_OSD_SLAB_OPS :
1150 CEPH_OSD_MAX_OPS)) {
1151 redirty_page_for_writepage(wbc, page);
1157 offset = (u64)page_offset(page);
1161 /* note position of first page in fbatch */
1162 doutc(cl, "%llx.%llx will write page %p idx %lu\n",
1163 ceph_vinop(inode), page, page->index);
1165 if (atomic_long_inc_return(&fsc->writeback_count) >
1166 CONGESTION_ON_THRESH(
1167 fsc->mount_options->congestion_kb))
1168 fsc->write_congested = true;
1170 if (IS_ENCRYPTED(inode)) {
1171 pages[locked_pages] =
1172 fscrypt_encrypt_pagecache_blocks(page,
1174 locked_pages ? GFP_NOWAIT : GFP_NOFS);
1175 if (IS_ERR(pages[locked_pages])) {
1176 if (PTR_ERR(pages[locked_pages]) == -EINVAL)
1178 "inode->i_blkbits=%hhu\n",
1180 /* better not fail on first page! */
1181 BUG_ON(locked_pages == 0);
1182 pages[locked_pages] = NULL;
1183 redirty_page_for_writepage(wbc, page);
1189 pages[locked_pages++] = page;
1192 fbatch.folios[i] = NULL;
1193 len += thp_size(page);
1196 /* did we get anything? */
1198 goto release_folios;
1201 /* shift unused page to beginning of fbatch */
1202 for (j = 0; j < nr_folios; j++) {
1203 if (!fbatch.folios[j])
1206 fbatch.folios[n] = fbatch.folios[j];
1211 if (nr_folios && i == nr_folios &&
1212 locked_pages < max_pages) {
1213 doutc(cl, "reached end fbatch, trying for more\n");
1214 folio_batch_release(&fbatch);
1215 goto get_more_pages;
1220 offset = ceph_fscrypt_page_offset(pages[0]);
1223 req = ceph_osdc_new_request(&fsc->client->osdc,
1224 &ci->i_layout, vino,
1225 offset, &len, 0, num_ops,
1226 CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1227 snapc, ceph_wbc.truncate_seq,
1228 ceph_wbc.truncate_size, false);
1230 req = ceph_osdc_new_request(&fsc->client->osdc,
1231 &ci->i_layout, vino,
1236 CEPH_OSD_FLAG_WRITE,
1237 snapc, ceph_wbc.truncate_seq,
1238 ceph_wbc.truncate_size, true);
1239 BUG_ON(IS_ERR(req));
1241 BUG_ON(len < ceph_fscrypt_page_offset(pages[locked_pages - 1]) +
1242 thp_size(pages[locked_pages - 1]) - offset);
1244 if (!ceph_inc_osd_stopping_blocker(fsc->mdsc)) {
1246 goto release_folios;
1248 req->r_callback = writepages_finish;
1249 req->r_inode = inode;
1251 /* Format the osd request message and submit the write */
1255 for (i = 0; i < locked_pages; i++) {
1256 struct page *page = ceph_fscrypt_pagecache_page(pages[i]);
1258 u64 cur_offset = page_offset(page);
1260 * Discontinuity in page range? Ceph can handle that by just passing
1261 * multiple extents in the write op.
1263 if (offset + len != cur_offset) {
1264 /* If it's full, stop here */
1265 if (op_idx + 1 == req->r_num_ops)
1268 /* Kick off an fscache write with what we have so far. */
1269 ceph_fscache_write_to_cache(inode, offset, len, caching);
1271 /* Start a new extent */
1272 osd_req_op_extent_dup_last(req, op_idx,
1273 cur_offset - offset);
1274 doutc(cl, "got pages at %llu~%llu\n", offset,
1276 osd_req_op_extent_osd_data_pages(req, op_idx,
1279 osd_req_op_extent_update(req, op_idx, len);
1282 offset = cur_offset;
1283 data_pages = pages + i;
1287 set_page_writeback(page);
1289 ceph_set_page_fscache(page);
1290 len += thp_size(page);
1292 ceph_fscache_write_to_cache(inode, offset, len, caching);
1294 if (ceph_wbc.size_stable) {
1295 len = min(len, ceph_wbc.i_size - offset);
1296 } else if (i == locked_pages) {
1297 /* writepages_finish() clears writeback pages
1298 * according to the data length, so make sure
1299 * data length covers all locked pages */
1300 u64 min_len = len + 1 - thp_size(page);
1301 len = get_writepages_data_length(inode, pages[i - 1],
1303 len = max(len, min_len);
1305 if (IS_ENCRYPTED(inode))
1306 len = round_up(len, CEPH_FSCRYPT_BLOCK_SIZE);
1308 doutc(cl, "got pages at %llu~%llu\n", offset, len);
1310 if (IS_ENCRYPTED(inode) &&
1311 ((offset | len) & ~CEPH_FSCRYPT_BLOCK_MASK))
1313 "bad encrypted write offset=%lld len=%llu\n",
1316 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1317 0, from_pool, false);
1318 osd_req_op_extent_update(req, op_idx, len);
1320 BUG_ON(op_idx + 1 != req->r_num_ops);
1323 if (i < locked_pages) {
1324 BUG_ON(num_ops <= req->r_num_ops);
1325 num_ops -= req->r_num_ops;
1328 /* allocate new pages array for next request */
1330 pages = kmalloc_array(locked_pages, sizeof(*pages),
1334 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1337 memcpy(pages, data_pages + i,
1338 locked_pages * sizeof(*pages));
1339 memset(data_pages + i, 0,
1340 locked_pages * sizeof(*pages));
1342 BUG_ON(num_ops != req->r_num_ops);
1343 index = pages[i - 1]->index + 1;
1344 /* request message now owns the pages array */
1348 req->r_mtime = inode_get_mtime(inode);
1349 ceph_osdc_start_request(&fsc->client->osdc, req);
1352 wbc->nr_to_write -= i;
1357 * We stop writing back only if we are not doing
1358 * integrity sync. In case of integrity sync we have to
1359 * keep going until we have written all the pages
1360 * we tagged for writeback prior to entering this loop.
1362 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1366 doutc(cl, "folio_batch release on %d folios (%p)\n",
1367 (int)fbatch.nr, fbatch.nr ? fbatch.folios[0] : NULL);
1368 folio_batch_release(&fbatch);
1371 if (should_loop && !done) {
1372 /* more to do; loop back to beginning of file */
1373 doutc(cl, "looping back to beginning of file\n");
1374 end = start_index - 1; /* OK even when start_index == 0 */
1376 /* to write dirty pages associated with next snapc,
1377 * we need to wait until current writes complete */
1378 if (wbc->sync_mode != WB_SYNC_NONE &&
1379 start_index == 0 && /* all dirty pages were checked */
1380 !ceph_wbc.head_snapc) {
1384 while ((index <= end) &&
1385 (nr = filemap_get_folios_tag(mapping, &index,
1387 PAGECACHE_TAG_WRITEBACK,
1389 for (i = 0; i < nr; i++) {
1390 page = &fbatch.folios[i]->page;
1391 if (page_snap_context(page) != snapc)
1393 wait_on_page_writeback(page);
1395 folio_batch_release(&fbatch);
1405 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1406 mapping->writeback_index = index;
1409 ceph_osdc_put_request(req);
1410 ceph_put_snap_context(last_snapc);
1411 doutc(cl, "%llx.%llx dend - startone, rc = %d\n", ceph_vinop(inode),
1419 * See if a given @snapc is either writeable, or already written.
1421 static int context_is_writeable_or_written(struct inode *inode,
1422 struct ceph_snap_context *snapc)
1424 struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1425 int ret = !oldest || snapc->seq <= oldest->seq;
1427 ceph_put_snap_context(oldest);
1432 * ceph_find_incompatible - find an incompatible context and return it
1433 * @page: page being dirtied
1435 * We are only allowed to write into/dirty a page if the page is
1436 * clean, or already dirty within the same snap context. Returns a
1437 * conflicting context if there is one, NULL if there isn't, or a
1438 * negative error code on other errors.
1440 * Must be called with page lock held.
1442 static struct ceph_snap_context *
1443 ceph_find_incompatible(struct page *page)
1445 struct inode *inode = page->mapping->host;
1446 struct ceph_client *cl = ceph_inode_to_client(inode);
1447 struct ceph_inode_info *ci = ceph_inode(inode);
1449 if (ceph_inode_is_shutdown(inode)) {
1450 doutc(cl, " %llx.%llx page %p is shutdown\n",
1451 ceph_vinop(inode), page);
1452 return ERR_PTR(-ESTALE);
1456 struct ceph_snap_context *snapc, *oldest;
1458 wait_on_page_writeback(page);
1460 snapc = page_snap_context(page);
1461 if (!snapc || snapc == ci->i_head_snapc)
1465 * this page is already dirty in another (older) snap
1466 * context! is it writeable now?
1468 oldest = get_oldest_context(inode, NULL, NULL);
1469 if (snapc->seq > oldest->seq) {
1470 /* not writeable -- return it for the caller to deal with */
1471 ceph_put_snap_context(oldest);
1472 doutc(cl, " %llx.%llx page %p snapc %p not current or oldest\n",
1473 ceph_vinop(inode), page, snapc);
1474 return ceph_get_snap_context(snapc);
1476 ceph_put_snap_context(oldest);
1478 /* yay, writeable, do it now (without dropping page lock) */
1479 doutc(cl, " %llx.%llx page %p snapc %p not current, but oldest\n",
1480 ceph_vinop(inode), page, snapc);
1481 if (clear_page_dirty_for_io(page)) {
1482 int r = writepage_nounlock(page, NULL);
1490 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
1491 struct folio **foliop, void **_fsdata)
1493 struct inode *inode = file_inode(file);
1494 struct ceph_inode_info *ci = ceph_inode(inode);
1495 struct ceph_snap_context *snapc;
1497 snapc = ceph_find_incompatible(folio_page(*foliop, 0));
1501 folio_unlock(*foliop);
1505 return PTR_ERR(snapc);
1507 ceph_queue_writeback(inode);
1508 r = wait_event_killable(ci->i_cap_wq,
1509 context_is_writeable_or_written(inode, snapc));
1510 ceph_put_snap_context(snapc);
1511 return r == 0 ? -EAGAIN : r;
1517 * We are only allowed to write into/dirty the page if the page is
1518 * clean, or already dirty within the same snap context.
1520 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1521 loff_t pos, unsigned len,
1522 struct page **pagep, void **fsdata)
1524 struct inode *inode = file_inode(file);
1525 struct ceph_inode_info *ci = ceph_inode(inode);
1526 struct folio *folio = NULL;
1529 r = netfs_write_begin(&ci->netfs, file, inode->i_mapping, pos, len, &folio, NULL);
1533 folio_wait_fscache(folio);
1534 WARN_ON_ONCE(!folio_test_locked(folio));
1535 *pagep = &folio->page;
1540 * we don't do anything in here that simple_write_end doesn't do
1541 * except adjust dirty page accounting
1543 static int ceph_write_end(struct file *file, struct address_space *mapping,
1544 loff_t pos, unsigned len, unsigned copied,
1545 struct page *subpage, void *fsdata)
1547 struct folio *folio = page_folio(subpage);
1548 struct inode *inode = file_inode(file);
1549 struct ceph_client *cl = ceph_inode_to_client(inode);
1550 bool check_cap = false;
1552 doutc(cl, "%llx.%llx file %p folio %p %d~%d (%d)\n", ceph_vinop(inode),
1553 file, folio, (int)pos, (int)copied, (int)len);
1555 if (!folio_test_uptodate(folio)) {
1556 /* just return that nothing was copied on a short copy */
1561 folio_mark_uptodate(folio);
1564 /* did file size increase? */
1565 if (pos+copied > i_size_read(inode))
1566 check_cap = ceph_inode_set_size(inode, pos+copied);
1568 folio_mark_dirty(folio);
1571 folio_unlock(folio);
1575 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY);
1580 const struct address_space_operations ceph_aops = {
1581 .read_folio = netfs_read_folio,
1582 .readahead = netfs_readahead,
1583 .writepage = ceph_writepage,
1584 .writepages = ceph_writepages_start,
1585 .write_begin = ceph_write_begin,
1586 .write_end = ceph_write_end,
1587 .dirty_folio = ceph_dirty_folio,
1588 .invalidate_folio = ceph_invalidate_folio,
1589 .release_folio = ceph_release_folio,
1590 .direct_IO = noop_direct_IO,
1593 static void ceph_block_sigs(sigset_t *oldset)
1596 siginitsetinv(&mask, sigmask(SIGKILL));
1597 sigprocmask(SIG_BLOCK, &mask, oldset);
1600 static void ceph_restore_sigs(sigset_t *oldset)
1602 sigprocmask(SIG_SETMASK, oldset, NULL);
1608 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1610 struct vm_area_struct *vma = vmf->vma;
1611 struct inode *inode = file_inode(vma->vm_file);
1612 struct ceph_inode_info *ci = ceph_inode(inode);
1613 struct ceph_client *cl = ceph_inode_to_client(inode);
1614 struct ceph_file_info *fi = vma->vm_file->private_data;
1615 loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1618 vm_fault_t ret = VM_FAULT_SIGBUS;
1620 if (ceph_inode_is_shutdown(inode))
1623 ceph_block_sigs(&oldset);
1625 doutc(cl, "%llx.%llx %llu trying to get caps\n",
1626 ceph_vinop(inode), off);
1627 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1628 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1630 want = CEPH_CAP_FILE_CACHE;
1633 err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1, &got);
1637 doutc(cl, "%llx.%llx %llu got cap refs on %s\n", ceph_vinop(inode),
1638 off, ceph_cap_string(got));
1640 if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1641 !ceph_has_inline_data(ci)) {
1642 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1643 ceph_add_rw_context(fi, &rw_ctx);
1644 ret = filemap_fault(vmf);
1645 ceph_del_rw_context(fi, &rw_ctx);
1646 doutc(cl, "%llx.%llx %llu drop cap refs %s ret %x\n",
1647 ceph_vinop(inode), off, ceph_cap_string(got), ret);
1651 ceph_put_cap_refs(ci, got);
1656 /* read inline data */
1657 if (off >= PAGE_SIZE) {
1658 /* does not support inline data > PAGE_SIZE */
1659 ret = VM_FAULT_SIGBUS;
1661 struct address_space *mapping = inode->i_mapping;
1664 filemap_invalidate_lock_shared(mapping);
1665 page = find_or_create_page(mapping, 0,
1666 mapping_gfp_constraint(mapping, ~__GFP_FS));
1671 err = __ceph_do_getattr(inode, page,
1672 CEPH_STAT_CAP_INLINE_DATA, true);
1673 if (err < 0 || off >= i_size_read(inode)) {
1676 ret = vmf_error(err);
1679 if (err < PAGE_SIZE)
1680 zero_user_segment(page, err, PAGE_SIZE);
1682 flush_dcache_page(page);
1683 SetPageUptodate(page);
1685 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1687 filemap_invalidate_unlock_shared(mapping);
1688 doutc(cl, "%llx.%llx %llu read inline data ret %x\n",
1689 ceph_vinop(inode), off, ret);
1692 ceph_restore_sigs(&oldset);
1694 ret = vmf_error(err);
1699 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1701 struct vm_area_struct *vma = vmf->vma;
1702 struct inode *inode = file_inode(vma->vm_file);
1703 struct ceph_client *cl = ceph_inode_to_client(inode);
1704 struct ceph_inode_info *ci = ceph_inode(inode);
1705 struct ceph_file_info *fi = vma->vm_file->private_data;
1706 struct ceph_cap_flush *prealloc_cf;
1707 struct page *page = vmf->page;
1708 loff_t off = page_offset(page);
1709 loff_t size = i_size_read(inode);
1713 vm_fault_t ret = VM_FAULT_SIGBUS;
1715 if (ceph_inode_is_shutdown(inode))
1718 prealloc_cf = ceph_alloc_cap_flush();
1720 return VM_FAULT_OOM;
1722 sb_start_pagefault(inode->i_sb);
1723 ceph_block_sigs(&oldset);
1725 if (off + thp_size(page) <= size)
1726 len = thp_size(page);
1728 len = offset_in_thp(page, size);
1730 doutc(cl, "%llx.%llx %llu~%zd getting caps i_size %llu\n",
1731 ceph_vinop(inode), off, len, size);
1732 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1733 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1735 want = CEPH_CAP_FILE_BUFFER;
1738 err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len, &got);
1742 doutc(cl, "%llx.%llx %llu~%zd got cap refs on %s\n", ceph_vinop(inode),
1743 off, len, ceph_cap_string(got));
1745 /* Update time before taking page lock */
1746 file_update_time(vma->vm_file);
1747 inode_inc_iversion_raw(inode);
1750 struct ceph_snap_context *snapc;
1754 if (page_mkwrite_check_truncate(page, inode) < 0) {
1756 ret = VM_FAULT_NOPAGE;
1760 snapc = ceph_find_incompatible(page);
1762 /* success. we'll keep the page locked. */
1763 set_page_dirty(page);
1764 ret = VM_FAULT_LOCKED;
1770 if (IS_ERR(snapc)) {
1771 ret = VM_FAULT_SIGBUS;
1775 ceph_queue_writeback(inode);
1776 err = wait_event_killable(ci->i_cap_wq,
1777 context_is_writeable_or_written(inode, snapc));
1778 ceph_put_snap_context(snapc);
1781 if (ret == VM_FAULT_LOCKED) {
1783 spin_lock(&ci->i_ceph_lock);
1784 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1786 spin_unlock(&ci->i_ceph_lock);
1788 __mark_inode_dirty(inode, dirty);
1791 doutc(cl, "%llx.%llx %llu~%zd dropping cap refs on %s ret %x\n",
1792 ceph_vinop(inode), off, len, ceph_cap_string(got), ret);
1793 ceph_put_cap_refs_async(ci, got);
1795 ceph_restore_sigs(&oldset);
1796 sb_end_pagefault(inode->i_sb);
1797 ceph_free_cap_flush(prealloc_cf);
1799 ret = vmf_error(err);
1803 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1804 char *data, size_t len)
1806 struct ceph_client *cl = ceph_inode_to_client(inode);
1807 struct address_space *mapping = inode->i_mapping;
1813 if (i_size_read(inode) == 0)
1815 page = find_or_create_page(mapping, 0,
1816 mapping_gfp_constraint(mapping,
1820 if (PageUptodate(page)) {
1827 doutc(cl, "%p %llx.%llx len %zu locked_page %p\n", inode,
1828 ceph_vinop(inode), len, locked_page);
1831 void *kaddr = kmap_atomic(page);
1832 memcpy(kaddr, data, len);
1833 kunmap_atomic(kaddr);
1836 if (page != locked_page) {
1837 if (len < PAGE_SIZE)
1838 zero_user_segment(page, len, PAGE_SIZE);
1840 flush_dcache_page(page);
1842 SetPageUptodate(page);
1848 int ceph_uninline_data(struct file *file)
1850 struct inode *inode = file_inode(file);
1851 struct ceph_inode_info *ci = ceph_inode(inode);
1852 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
1853 struct ceph_client *cl = fsc->client;
1854 struct ceph_osd_request *req = NULL;
1855 struct ceph_cap_flush *prealloc_cf = NULL;
1856 struct folio *folio = NULL;
1857 u64 inline_version = CEPH_INLINE_NONE;
1858 struct page *pages[1];
1862 spin_lock(&ci->i_ceph_lock);
1863 inline_version = ci->i_inline_version;
1864 spin_unlock(&ci->i_ceph_lock);
1866 doutc(cl, "%llx.%llx inline_version %llu\n", ceph_vinop(inode),
1869 if (ceph_inode_is_shutdown(inode)) {
1874 if (inline_version == CEPH_INLINE_NONE)
1877 prealloc_cf = ceph_alloc_cap_flush();
1881 if (inline_version == 1) /* initial version, no data */
1884 folio = read_mapping_folio(inode->i_mapping, 0, file);
1885 if (IS_ERR(folio)) {
1886 err = PTR_ERR(folio);
1892 len = i_size_read(inode);
1893 if (len > folio_size(folio))
1894 len = folio_size(folio);
1896 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1897 ceph_vino(inode), 0, &len, 0, 1,
1898 CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1905 req->r_mtime = inode_get_mtime(inode);
1906 ceph_osdc_start_request(&fsc->client->osdc, req);
1907 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1908 ceph_osdc_put_request(req);
1912 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1913 ceph_vino(inode), 0, &len, 1, 3,
1914 CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1915 NULL, ci->i_truncate_seq,
1916 ci->i_truncate_size, false);
1922 pages[0] = folio_page(folio, 0);
1923 osd_req_op_extent_osd_data_pages(req, 1, pages, len, 0, false, false);
1926 __le64 xattr_buf = cpu_to_le64(inline_version);
1927 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1928 "inline_version", &xattr_buf,
1930 CEPH_OSD_CMPXATTR_OP_GT,
1931 CEPH_OSD_CMPXATTR_MODE_U64);
1938 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1939 "%llu", inline_version);
1940 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1942 xattr_buf, xattr_len, 0, 0);
1947 req->r_mtime = inode_get_mtime(inode);
1948 ceph_osdc_start_request(&fsc->client->osdc, req);
1949 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1951 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1952 req->r_end_latency, len, err);
1958 /* Set to CAP_INLINE_NONE and dirty the caps */
1959 down_read(&fsc->mdsc->snap_rwsem);
1960 spin_lock(&ci->i_ceph_lock);
1961 ci->i_inline_version = CEPH_INLINE_NONE;
1962 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR, &prealloc_cf);
1963 spin_unlock(&ci->i_ceph_lock);
1964 up_read(&fsc->mdsc->snap_rwsem);
1966 __mark_inode_dirty(inode, dirty);
1969 ceph_osdc_put_request(req);
1970 if (err == -ECANCELED)
1974 folio_unlock(folio);
1978 ceph_free_cap_flush(prealloc_cf);
1979 doutc(cl, "%llx.%llx inline_version %llu = %d\n",
1980 ceph_vinop(inode), inline_version, err);
1984 static const struct vm_operations_struct ceph_vmops = {
1985 .fault = ceph_filemap_fault,
1986 .page_mkwrite = ceph_page_mkwrite,
1989 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1991 struct address_space *mapping = file->f_mapping;
1993 if (!mapping->a_ops->read_folio)
1995 vma->vm_ops = &ceph_vmops;
2004 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
2005 s64 pool, struct ceph_string *pool_ns)
2007 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(&ci->netfs.inode);
2008 struct ceph_mds_client *mdsc = fsc->mdsc;
2009 struct ceph_client *cl = fsc->client;
2010 struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
2011 struct rb_node **p, *parent;
2012 struct ceph_pool_perm *perm;
2013 struct page **pages;
2015 int err = 0, err2 = 0, have = 0;
2017 down_read(&mdsc->pool_perm_rwsem);
2018 p = &mdsc->pool_perm_tree.rb_node;
2020 perm = rb_entry(*p, struct ceph_pool_perm, node);
2021 if (pool < perm->pool)
2023 else if (pool > perm->pool)
2024 p = &(*p)->rb_right;
2026 int ret = ceph_compare_string(pool_ns,
2032 p = &(*p)->rb_right;
2039 up_read(&mdsc->pool_perm_rwsem);
2044 doutc(cl, "pool %lld ns %.*s no perm cached\n", pool,
2045 (int)pool_ns->len, pool_ns->str);
2047 doutc(cl, "pool %lld no perm cached\n", pool);
2049 down_write(&mdsc->pool_perm_rwsem);
2050 p = &mdsc->pool_perm_tree.rb_node;
2054 perm = rb_entry(parent, struct ceph_pool_perm, node);
2055 if (pool < perm->pool)
2057 else if (pool > perm->pool)
2058 p = &(*p)->rb_right;
2060 int ret = ceph_compare_string(pool_ns,
2066 p = &(*p)->rb_right;
2074 up_write(&mdsc->pool_perm_rwsem);
2078 rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2079 1, false, GFP_NOFS);
2085 rd_req->r_flags = CEPH_OSD_FLAG_READ;
2086 osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
2087 rd_req->r_base_oloc.pool = pool;
2089 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
2090 ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
2092 err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
2096 wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2097 1, false, GFP_NOFS);
2103 wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
2104 osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
2105 ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
2106 ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
2108 err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
2112 /* one page should be large enough for STAT data */
2113 pages = ceph_alloc_page_vector(1, GFP_KERNEL);
2114 if (IS_ERR(pages)) {
2115 err = PTR_ERR(pages);
2119 osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
2121 ceph_osdc_start_request(&fsc->client->osdc, rd_req);
2123 wr_req->r_mtime = inode_get_mtime(&ci->netfs.inode);
2124 ceph_osdc_start_request(&fsc->client->osdc, wr_req);
2126 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
2127 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
2129 if (err >= 0 || err == -ENOENT)
2131 else if (err != -EPERM) {
2132 if (err == -EBLOCKLISTED)
2133 fsc->blocklisted = true;
2137 if (err2 == 0 || err2 == -EEXIST)
2139 else if (err2 != -EPERM) {
2140 if (err2 == -EBLOCKLISTED)
2141 fsc->blocklisted = true;
2146 pool_ns_len = pool_ns ? pool_ns->len : 0;
2147 perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
2155 perm->pool_ns_len = pool_ns_len;
2156 if (pool_ns_len > 0)
2157 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
2158 perm->pool_ns[pool_ns_len] = 0;
2160 rb_link_node(&perm->node, parent, p);
2161 rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
2164 up_write(&mdsc->pool_perm_rwsem);
2166 ceph_osdc_put_request(rd_req);
2167 ceph_osdc_put_request(wr_req);
2172 doutc(cl, "pool %lld ns %.*s result = %d\n", pool,
2173 (int)pool_ns->len, pool_ns->str, err);
2175 doutc(cl, "pool %lld result = %d\n", pool, err);
2179 int ceph_pool_perm_check(struct inode *inode, int need)
2181 struct ceph_client *cl = ceph_inode_to_client(inode);
2182 struct ceph_inode_info *ci = ceph_inode(inode);
2183 struct ceph_string *pool_ns;
2187 /* Only need to do this for regular files */
2188 if (!S_ISREG(inode->i_mode))
2191 if (ci->i_vino.snap != CEPH_NOSNAP) {
2193 * Pool permission check needs to write to the first object.
2194 * But for snapshot, head of the first object may have alread
2195 * been deleted. Skip check to avoid creating orphan object.
2200 if (ceph_test_mount_opt(ceph_inode_to_fs_client(inode),
2204 spin_lock(&ci->i_ceph_lock);
2205 flags = ci->i_ceph_flags;
2206 pool = ci->i_layout.pool_id;
2207 spin_unlock(&ci->i_ceph_lock);
2209 if (flags & CEPH_I_POOL_PERM) {
2210 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2211 doutc(cl, "pool %lld no read perm\n", pool);
2214 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2215 doutc(cl, "pool %lld no write perm\n", pool);
2221 pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2222 ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2223 ceph_put_string(pool_ns);
2227 flags = CEPH_I_POOL_PERM;
2228 if (ret & POOL_READ)
2229 flags |= CEPH_I_POOL_RD;
2230 if (ret & POOL_WRITE)
2231 flags |= CEPH_I_POOL_WR;
2233 spin_lock(&ci->i_ceph_lock);
2234 if (pool == ci->i_layout.pool_id &&
2235 pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2236 ci->i_ceph_flags |= flags;
2238 pool = ci->i_layout.pool_id;
2239 flags = ci->i_ceph_flags;
2241 spin_unlock(&ci->i_ceph_lock);
2245 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2247 struct ceph_pool_perm *perm;
2250 while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2251 n = rb_first(&mdsc->pool_perm_tree);
2252 perm = rb_entry(n, struct ceph_pool_perm, node);
2253 rb_erase(n, &mdsc->pool_perm_tree);