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)
83 struct ceph_inode_info *ci;
84 struct ceph_snap_context *snapc;
86 if (folio_test_dirty(folio)) {
87 dout("%p dirty_folio %p idx %lu -- already dirty\n",
88 mapping->host, folio, folio->index);
89 VM_BUG_ON_FOLIO(!folio_test_private(folio), folio);
93 inode = mapping->host;
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 dout("%p dirty_folio %p idx %lu head %d/%d -> %d/%d "
115 "snapc %p seq %lld (%d snaps)\n",
116 mapping->host, 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,
141 struct ceph_inode_info *ci;
142 struct ceph_snap_context *snapc;
144 inode = folio->mapping->host;
145 ci = ceph_inode(inode);
147 if (offset != 0 || length != folio_size(folio)) {
148 dout("%p invalidate_folio idx %lu partial dirty page %zu~%zu\n",
149 inode, folio->index, offset, length);
153 WARN_ON(!folio_test_locked(folio));
154 if (folio_test_private(folio)) {
155 dout("%p invalidate_folio idx %lu full dirty page\n",
156 inode, folio->index);
158 snapc = folio_detach_private(folio);
159 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
160 ceph_put_snap_context(snapc);
163 folio_wait_fscache(folio);
166 static bool ceph_release_folio(struct folio *folio, gfp_t gfp)
168 struct inode *inode = folio->mapping->host;
170 dout("%llx:%llx release_folio idx %lu (%sdirty)\n",
172 folio->index, folio_test_dirty(folio) ? "" : "not ");
174 if (folio_test_private(folio))
177 if (folio_test_fscache(folio)) {
178 if (current_is_kswapd() || !(gfp & __GFP_FS))
180 folio_wait_fscache(folio);
182 ceph_fscache_note_page_release(inode);
186 static void ceph_netfs_expand_readahead(struct netfs_io_request *rreq)
188 struct inode *inode = rreq->inode;
189 struct ceph_inode_info *ci = ceph_inode(inode);
190 struct ceph_file_layout *lo = &ci->i_layout;
191 unsigned long max_pages = inode->i_sb->s_bdi->ra_pages;
192 loff_t end = rreq->start + rreq->len, new_end;
193 struct ceph_netfs_request_data *priv = rreq->netfs_priv;
194 unsigned long max_len;
198 /* Readahead is disabled by posix_fadvise POSIX_FADV_RANDOM */
199 if (priv->file_ra_disabled)
202 max_pages = priv->file_ra_pages;
206 /* Readahead is disabled */
210 max_len = max_pages << PAGE_SHIFT;
213 * Try to expand the length forward by rounding up it to the next
214 * block, but do not exceed the file size, unless the original
215 * request already exceeds it.
217 new_end = min(round_up(end, lo->stripe_unit), rreq->i_size);
218 if (new_end > end && new_end <= rreq->start + max_len)
219 rreq->len = new_end - rreq->start;
221 /* Try to expand the start downward */
222 div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
223 if (rreq->len + blockoff <= max_len) {
224 rreq->start -= blockoff;
225 rreq->len += blockoff;
229 static bool ceph_netfs_clamp_length(struct netfs_io_subrequest *subreq)
231 struct inode *inode = subreq->rreq->inode;
232 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
233 struct ceph_inode_info *ci = ceph_inode(inode);
237 /* Truncate the extent at the end of the current block */
238 ceph_calc_file_object_mapping(&ci->i_layout, subreq->start, subreq->len,
239 &objno, &objoff, &xlen);
240 subreq->len = min(xlen, fsc->mount_options->rsize);
244 static void finish_netfs_read(struct ceph_osd_request *req)
246 struct inode *inode = req->r_inode;
247 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
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 dout("%s: result %d subreq->len=%zu i_size=%lld\n", __func__, 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);
290 static bool ceph_netfs_issue_op_inline(struct netfs_io_subrequest *subreq)
292 struct netfs_io_request *rreq = subreq->rreq;
293 struct inode *inode = rreq->inode;
294 struct ceph_mds_reply_info_parsed *rinfo;
295 struct ceph_mds_reply_info_in *iinfo;
296 struct ceph_mds_request *req;
297 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
298 struct ceph_inode_info *ci = ceph_inode(inode);
299 struct iov_iter iter;
304 __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
305 __clear_bit(NETFS_SREQ_COPY_TO_CACHE, &subreq->flags);
307 if (subreq->start >= inode->i_size)
310 /* We need to fetch the inline data. */
311 mode = ceph_try_to_choose_auth_mds(inode, CEPH_STAT_CAP_INLINE_DATA);
312 req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_GETATTR, mode);
317 req->r_ino1 = ci->i_vino;
318 req->r_args.getattr.mask = cpu_to_le32(CEPH_STAT_CAP_INLINE_DATA);
321 err = ceph_mdsc_do_request(mdsc, NULL, req);
325 rinfo = &req->r_reply_info;
326 iinfo = &rinfo->targeti;
327 if (iinfo->inline_version == CEPH_INLINE_NONE) {
328 /* The data got uninlined */
329 ceph_mdsc_put_request(req);
333 len = min_t(size_t, iinfo->inline_len - subreq->start, subreq->len);
334 iov_iter_xarray(&iter, ITER_DEST, &rreq->mapping->i_pages, subreq->start, len);
335 err = copy_to_iter(iinfo->inline_data + subreq->start, len, &iter);
339 ceph_mdsc_put_request(req);
341 netfs_subreq_terminated(subreq, err, false);
345 static void ceph_netfs_issue_read(struct netfs_io_subrequest *subreq)
347 struct netfs_io_request *rreq = subreq->rreq;
348 struct inode *inode = rreq->inode;
349 struct ceph_inode_info *ci = ceph_inode(inode);
350 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
351 struct ceph_osd_request *req = NULL;
352 struct ceph_vino vino = ceph_vino(inode);
353 struct iov_iter iter;
355 u64 len = subreq->len;
356 bool sparse = IS_ENCRYPTED(inode) || ceph_test_mount_opt(fsc, SPARSEREAD);
357 u64 off = subreq->start;
359 if (ceph_inode_is_shutdown(inode)) {
364 if (ceph_has_inline_data(ci) && ceph_netfs_issue_op_inline(subreq))
367 ceph_fscrypt_adjust_off_and_len(inode, &off, &len);
369 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, vino,
370 off, &len, 0, 1, sparse ? CEPH_OSD_OP_SPARSE_READ : CEPH_OSD_OP_READ,
371 CEPH_OSD_FLAG_READ | fsc->client->osdc.client->options->read_from_replica,
372 NULL, ci->i_truncate_seq, ci->i_truncate_size, false);
380 err = ceph_alloc_sparse_ext_map(&req->r_ops[0]);
385 dout("%s: pos=%llu orig_len=%zu len=%llu\n", __func__, subreq->start, subreq->len, len);
387 iov_iter_xarray(&iter, ITER_DEST, &rreq->mapping->i_pages, subreq->start, len);
390 * FIXME: For now, use CEPH_OSD_DATA_TYPE_PAGES instead of _ITER for
391 * encrypted inodes. We'd need infrastructure that handles an iov_iter
392 * instead of page arrays, and we don't have that as of yet. Once the
393 * dust settles on the write helpers and encrypt/decrypt routines for
394 * netfs, we should be able to rework this.
396 if (IS_ENCRYPTED(inode)) {
400 err = iov_iter_get_pages_alloc2(&iter, &pages, len, &page_off);
402 dout("%s: iov_ter_get_pages_alloc returned %d\n",
407 /* should always give us a page-aligned read */
408 WARN_ON_ONCE(page_off);
412 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false,
415 osd_req_op_extent_osd_iter(req, 0, &iter);
417 req->r_callback = finish_netfs_read;
418 req->r_priv = subreq;
419 req->r_inode = inode;
422 ceph_osdc_start_request(req->r_osdc, req);
424 ceph_osdc_put_request(req);
426 netfs_subreq_terminated(subreq, err, false);
427 dout("%s: result %d\n", __func__, err);
430 static int ceph_init_request(struct netfs_io_request *rreq, struct file *file)
432 struct inode *inode = rreq->inode;
433 int got = 0, want = CEPH_CAP_FILE_CACHE;
434 struct ceph_netfs_request_data *priv;
437 if (rreq->origin != NETFS_READAHEAD)
440 priv = kzalloc(sizeof(*priv), GFP_NOFS);
445 struct ceph_rw_context *rw_ctx;
446 struct ceph_file_info *fi = file->private_data;
448 priv->file_ra_pages = file->f_ra.ra_pages;
449 priv->file_ra_disabled = file->f_mode & FMODE_RANDOM;
451 rw_ctx = ceph_find_rw_context(fi);
453 rreq->netfs_priv = priv;
459 * readahead callers do not necessarily hold Fcb caps
460 * (e.g. fadvise, madvise).
462 ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want, true, &got);
464 dout("start_read %p, error getting cap\n", inode);
469 dout("start_read %p, no cache cap\n", inode);
479 rreq->netfs_priv = priv;
488 static void ceph_netfs_free_request(struct netfs_io_request *rreq)
490 struct ceph_netfs_request_data *priv = rreq->netfs_priv;
496 ceph_put_cap_refs(ceph_inode(rreq->inode), priv->caps);
498 rreq->netfs_priv = NULL;
501 const struct netfs_request_ops ceph_netfs_ops = {
502 .init_request = ceph_init_request,
503 .free_request = ceph_netfs_free_request,
504 .begin_cache_operation = ceph_begin_cache_operation,
505 .issue_read = ceph_netfs_issue_read,
506 .expand_readahead = ceph_netfs_expand_readahead,
507 .clamp_length = ceph_netfs_clamp_length,
508 .check_write_begin = ceph_netfs_check_write_begin,
511 #ifdef CONFIG_CEPH_FSCACHE
512 static void ceph_set_page_fscache(struct page *page)
514 set_page_fscache(page);
517 static void ceph_fscache_write_terminated(void *priv, ssize_t error, bool was_async)
519 struct inode *inode = priv;
521 if (IS_ERR_VALUE(error) && error != -ENOBUFS)
522 ceph_fscache_invalidate(inode, false);
525 static void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
527 struct ceph_inode_info *ci = ceph_inode(inode);
528 struct fscache_cookie *cookie = ceph_fscache_cookie(ci);
530 fscache_write_to_cache(cookie, inode->i_mapping, off, len, i_size_read(inode),
531 ceph_fscache_write_terminated, inode, caching);
534 static inline void ceph_set_page_fscache(struct page *page)
538 static inline void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
541 #endif /* CONFIG_CEPH_FSCACHE */
543 struct ceph_writeback_ctl
553 * Get ref for the oldest snapc for an inode with dirty data... that is, the
554 * only snap context we are allowed to write back.
556 static struct ceph_snap_context *
557 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
558 struct ceph_snap_context *page_snapc)
560 struct ceph_inode_info *ci = ceph_inode(inode);
561 struct ceph_snap_context *snapc = NULL;
562 struct ceph_cap_snap *capsnap = NULL;
564 spin_lock(&ci->i_ceph_lock);
565 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
566 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
567 capsnap->context, capsnap->dirty_pages);
568 if (!capsnap->dirty_pages)
571 /* get i_size, truncate_{seq,size} for page_snapc? */
572 if (snapc && capsnap->context != page_snapc)
576 if (capsnap->writing) {
577 ctl->i_size = i_size_read(inode);
578 ctl->size_stable = false;
580 ctl->i_size = capsnap->size;
581 ctl->size_stable = true;
583 ctl->truncate_size = capsnap->truncate_size;
584 ctl->truncate_seq = capsnap->truncate_seq;
585 ctl->head_snapc = false;
591 snapc = ceph_get_snap_context(capsnap->context);
593 page_snapc == snapc ||
594 page_snapc->seq > snapc->seq)
597 if (!snapc && ci->i_wrbuffer_ref_head) {
598 snapc = ceph_get_snap_context(ci->i_head_snapc);
599 dout(" head snapc %p has %d dirty pages\n",
600 snapc, ci->i_wrbuffer_ref_head);
602 ctl->i_size = i_size_read(inode);
603 ctl->truncate_size = ci->i_truncate_size;
604 ctl->truncate_seq = ci->i_truncate_seq;
605 ctl->size_stable = false;
606 ctl->head_snapc = true;
609 spin_unlock(&ci->i_ceph_lock);
613 static u64 get_writepages_data_length(struct inode *inode,
614 struct page *page, u64 start)
616 struct ceph_inode_info *ci = ceph_inode(inode);
617 struct ceph_snap_context *snapc;
618 struct ceph_cap_snap *capsnap = NULL;
619 u64 end = i_size_read(inode);
622 snapc = page_snap_context(ceph_fscrypt_pagecache_page(page));
623 if (snapc != ci->i_head_snapc) {
625 spin_lock(&ci->i_ceph_lock);
626 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
627 if (capsnap->context == snapc) {
628 if (!capsnap->writing)
634 spin_unlock(&ci->i_ceph_lock);
637 if (end > ceph_fscrypt_page_offset(page) + thp_size(page))
638 end = ceph_fscrypt_page_offset(page) + thp_size(page);
639 ret = end > start ? end - start : 0;
640 if (ret && fscrypt_is_bounce_page(page))
641 ret = round_up(ret, CEPH_FSCRYPT_BLOCK_SIZE);
646 * Write a single page, but leave the page locked.
648 * If we get a write error, mark the mapping for error, but still adjust the
649 * dirty page accounting (i.e., page is no longer dirty).
651 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
653 struct folio *folio = page_folio(page);
654 struct inode *inode = page->mapping->host;
655 struct ceph_inode_info *ci = ceph_inode(inode);
656 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
657 struct ceph_snap_context *snapc, *oldest;
658 loff_t page_off = page_offset(page);
660 loff_t len = thp_size(page);
662 struct ceph_writeback_ctl ceph_wbc;
663 struct ceph_osd_client *osdc = &fsc->client->osdc;
664 struct ceph_osd_request *req;
665 bool caching = ceph_is_cache_enabled(inode);
666 struct page *bounce_page = NULL;
668 dout("writepage %p idx %lu\n", page, page->index);
670 if (ceph_inode_is_shutdown(inode))
673 /* verify this is a writeable snap context */
674 snapc = page_snap_context(page);
676 dout("writepage %p page %p not dirty?\n", inode, page);
679 oldest = get_oldest_context(inode, &ceph_wbc, snapc);
680 if (snapc->seq > oldest->seq) {
681 dout("writepage %p page %p snapc %p not writeable - noop\n",
683 /* we should only noop if called by kswapd */
684 WARN_ON(!(current->flags & PF_MEMALLOC));
685 ceph_put_snap_context(oldest);
686 redirty_page_for_writepage(wbc, page);
689 ceph_put_snap_context(oldest);
691 /* is this a partial page at end of file? */
692 if (page_off >= ceph_wbc.i_size) {
693 dout("folio at %lu beyond eof %llu\n", folio->index,
695 folio_invalidate(folio, 0, folio_size(folio));
699 if (ceph_wbc.i_size < page_off + len)
700 len = ceph_wbc.i_size - page_off;
702 wlen = IS_ENCRYPTED(inode) ? round_up(len, CEPH_FSCRYPT_BLOCK_SIZE) : len;
703 dout("writepage %p page %p index %lu on %llu~%llu snapc %p seq %lld\n",
704 inode, page, page->index, page_off, wlen, snapc, snapc->seq);
706 if (atomic_long_inc_return(&fsc->writeback_count) >
707 CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
708 fsc->write_congested = true;
710 req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode),
711 page_off, &wlen, 0, 1, CEPH_OSD_OP_WRITE,
712 CEPH_OSD_FLAG_WRITE, snapc,
713 ceph_wbc.truncate_seq,
714 ceph_wbc.truncate_size, true);
716 redirty_page_for_writepage(wbc, page);
723 set_page_writeback(page);
725 ceph_set_page_fscache(page);
726 ceph_fscache_write_to_cache(inode, page_off, len, caching);
728 if (IS_ENCRYPTED(inode)) {
729 bounce_page = fscrypt_encrypt_pagecache_blocks(page,
730 CEPH_FSCRYPT_BLOCK_SIZE, 0,
732 if (IS_ERR(bounce_page)) {
733 redirty_page_for_writepage(wbc, page);
734 end_page_writeback(page);
735 ceph_osdc_put_request(req);
736 return PTR_ERR(bounce_page);
740 /* it may be a short write due to an object boundary */
741 WARN_ON_ONCE(len > thp_size(page));
742 osd_req_op_extent_osd_data_pages(req, 0,
743 bounce_page ? &bounce_page : &page, wlen, 0,
745 dout("writepage %llu~%llu (%llu bytes, %sencrypted)\n",
746 page_off, len, wlen, IS_ENCRYPTED(inode) ? "" : "not ");
748 req->r_mtime = inode->i_mtime;
749 ceph_osdc_start_request(osdc, req);
750 err = ceph_osdc_wait_request(osdc, req);
752 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
753 req->r_end_latency, len, err);
754 fscrypt_free_bounce_page(bounce_page);
755 ceph_osdc_put_request(req);
760 struct writeback_control tmp_wbc;
763 if (err == -ERESTARTSYS) {
764 /* killed by SIGKILL */
765 dout("writepage interrupted page %p\n", page);
766 redirty_page_for_writepage(wbc, page);
767 end_page_writeback(page);
770 if (err == -EBLOCKLISTED)
771 fsc->blocklisted = true;
772 dout("writepage setting page/mapping error %d %p\n",
774 mapping_set_error(&inode->i_data, err);
775 wbc->pages_skipped++;
777 dout("writepage cleaned page %p\n", page);
778 err = 0; /* vfs expects us to return 0 */
780 oldest = detach_page_private(page);
781 WARN_ON_ONCE(oldest != snapc);
782 end_page_writeback(page);
783 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
784 ceph_put_snap_context(snapc); /* page's reference */
786 if (atomic_long_dec_return(&fsc->writeback_count) <
787 CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
788 fsc->write_congested = false;
793 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
796 struct inode *inode = page->mapping->host;
800 if (wbc->sync_mode == WB_SYNC_NONE &&
801 ceph_inode_to_client(inode)->write_congested)
802 return AOP_WRITEPAGE_ACTIVATE;
804 wait_on_page_fscache(page);
806 err = writepage_nounlock(page, wbc);
807 if (err == -ERESTARTSYS) {
808 /* direct memory reclaimer was killed by SIGKILL. return 0
809 * to prevent caller from setting mapping/page error */
818 * async writeback completion handler.
820 * If we get an error, set the mapping error bit, but not the individual
823 static void writepages_finish(struct ceph_osd_request *req)
825 struct inode *inode = req->r_inode;
826 struct ceph_inode_info *ci = ceph_inode(inode);
827 struct ceph_osd_data *osd_data;
829 int num_pages, total_pages = 0;
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_client(inode);
835 unsigned int len = 0;
838 dout("writepages_finish %p rc %d\n", inode, rc);
840 mapping_set_error(mapping, rc);
841 ceph_set_error_write(ci);
842 if (rc == -EBLOCKLISTED)
843 fsc->blocklisted = true;
845 ceph_clear_error_write(ci);
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
854 remove_page = !(ceph_caps_issued(ci) &
855 (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
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("%s incorrect op %d req %p index %d tid %llu\n",
861 __func__, req->r_ops[i].op, req, i, req->r_tid);
865 osd_data = osd_req_op_extent_osd_data(req, i);
866 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
867 len += osd_data->length;
868 num_pages = calc_pages_for((u64)osd_data->alignment,
869 (u64)osd_data->length);
870 total_pages += num_pages;
871 for (j = 0; j < num_pages; j++) {
872 page = osd_data->pages[j];
873 if (fscrypt_is_bounce_page(page)) {
874 page = fscrypt_pagecache_page(page);
875 fscrypt_free_bounce_page(osd_data->pages[j]);
876 osd_data->pages[j] = page;
879 WARN_ON(!PageUptodate(page));
881 if (atomic_long_dec_return(&fsc->writeback_count) <
882 CONGESTION_OFF_THRESH(
883 fsc->mount_options->congestion_kb))
884 fsc->write_congested = false;
886 ceph_put_snap_context(detach_page_private(page));
887 end_page_writeback(page);
888 dout("unlocking %p\n", page);
891 generic_error_remove_page(inode->i_mapping,
896 dout("writepages_finish %p wrote %llu bytes cleaned %d pages\n",
897 inode, osd_data->length, rc >= 0 ? num_pages : 0);
899 release_pages(osd_data->pages, num_pages);
902 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
903 req->r_end_latency, len, rc);
905 ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
907 osd_data = osd_req_op_extent_osd_data(req, 0);
908 if (osd_data->pages_from_pool)
909 mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
911 kfree(osd_data->pages);
912 ceph_osdc_put_request(req);
916 * initiate async writeback
918 static int ceph_writepages_start(struct address_space *mapping,
919 struct writeback_control *wbc)
921 struct inode *inode = mapping->host;
922 struct ceph_inode_info *ci = ceph_inode(inode);
923 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
924 struct ceph_vino vino = ceph_vino(inode);
925 pgoff_t index, start_index, end = -1;
926 struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
927 struct folio_batch fbatch;
929 unsigned int wsize = i_blocksize(inode);
930 struct ceph_osd_request *req = NULL;
931 struct ceph_writeback_ctl ceph_wbc;
932 bool should_loop, range_whole = false;
934 bool caching = ceph_is_cache_enabled(inode);
937 if (wbc->sync_mode == WB_SYNC_NONE &&
938 fsc->write_congested)
941 dout("writepages_start %p (mode=%s)\n", inode,
942 wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
943 (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
945 if (ceph_inode_is_shutdown(inode)) {
946 if (ci->i_wrbuffer_ref > 0) {
948 "writepage_start %p %lld forced umount\n",
949 inode, ceph_ino(inode));
951 mapping_set_error(mapping, -EIO);
952 return -EIO; /* we're in a forced umount, don't write! */
954 if (fsc->mount_options->wsize < wsize)
955 wsize = fsc->mount_options->wsize;
957 folio_batch_init(&fbatch);
959 start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
962 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
963 tag = PAGECACHE_TAG_TOWRITE;
965 tag = PAGECACHE_TAG_DIRTY;
968 /* find oldest snap context with dirty data */
969 snapc = get_oldest_context(inode, &ceph_wbc, NULL);
971 /* hmm, why does writepages get called when there
973 dout(" no snap context with dirty data?\n");
976 dout(" oldest snapc is %p seq %lld (%d snaps)\n",
977 snapc, snapc->seq, snapc->num_snaps);
980 if (ceph_wbc.head_snapc && snapc != last_snapc) {
981 /* where to start/end? */
982 if (wbc->range_cyclic) {
987 dout(" cyclic, start at %lu\n", index);
989 index = wbc->range_start >> PAGE_SHIFT;
990 end = wbc->range_end >> PAGE_SHIFT;
991 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
993 dout(" not cyclic, %lu to %lu\n", index, end);
995 } else if (!ceph_wbc.head_snapc) {
996 /* Do not respect wbc->range_{start,end}. Dirty pages
997 * in that range can be associated with newer snapc.
998 * They are not writeable until we write all dirty pages
999 * associated with 'snapc' get written */
1002 dout(" non-head snapc, range whole\n");
1005 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1006 tag_pages_for_writeback(mapping, index, end);
1008 ceph_put_snap_context(last_snapc);
1011 while (!done && index <= end) {
1012 int num_ops = 0, op_idx;
1013 unsigned i, nr_folios, max_pages, locked_pages = 0;
1014 struct page **pages = NULL, **data_pages;
1016 pgoff_t strip_unit_end = 0;
1017 u64 offset = 0, len = 0;
1018 bool from_pool = false;
1020 max_pages = wsize >> PAGE_SHIFT;
1023 nr_folios = filemap_get_folios_tag(mapping, &index,
1025 dout("pagevec_lookup_range_tag got %d\n", nr_folios);
1026 if (!nr_folios && !locked_pages)
1028 for (i = 0; i < nr_folios && locked_pages < max_pages; i++) {
1029 page = &fbatch.folios[i]->page;
1030 dout("? %p idx %lu\n", page, page->index);
1031 if (locked_pages == 0)
1032 lock_page(page); /* first page */
1033 else if (!trylock_page(page))
1036 /* only dirty pages, or our accounting breaks */
1037 if (unlikely(!PageDirty(page)) ||
1038 unlikely(page->mapping != mapping)) {
1039 dout("!dirty or !mapping %p\n", page);
1043 /* only if matching snap context */
1044 pgsnapc = page_snap_context(page);
1045 if (pgsnapc != snapc) {
1046 dout("page snapc %p %lld != oldest %p %lld\n",
1047 pgsnapc, pgsnapc->seq, snapc, snapc->seq);
1049 !ceph_wbc.head_snapc &&
1050 wbc->sync_mode != WB_SYNC_NONE)
1055 if (page_offset(page) >= ceph_wbc.i_size) {
1056 struct folio *folio = page_folio(page);
1058 dout("folio at %lu beyond eof %llu\n",
1059 folio->index, ceph_wbc.i_size);
1060 if ((ceph_wbc.size_stable ||
1061 folio_pos(folio) >= i_size_read(inode)) &&
1062 folio_clear_dirty_for_io(folio))
1063 folio_invalidate(folio, 0,
1065 folio_unlock(folio);
1068 if (strip_unit_end && (page->index > strip_unit_end)) {
1069 dout("end of strip unit %p\n", page);
1073 if (PageWriteback(page) || PageFsCache(page)) {
1074 if (wbc->sync_mode == WB_SYNC_NONE) {
1075 dout("%p under writeback\n", page);
1079 dout("waiting on writeback %p\n", page);
1080 wait_on_page_writeback(page);
1081 wait_on_page_fscache(page);
1084 if (!clear_page_dirty_for_io(page)) {
1085 dout("%p !clear_page_dirty_for_io\n", page);
1091 * We have something to write. If this is
1092 * the first locked page this time through,
1093 * calculate max possinle write size and
1094 * allocate a page array
1096 if (locked_pages == 0) {
1101 /* prepare async write request */
1102 offset = (u64)page_offset(page);
1103 ceph_calc_file_object_mapping(&ci->i_layout,
1110 strip_unit_end = page->index +
1111 ((len - 1) >> PAGE_SHIFT);
1114 max_pages = calc_pages_for(0, (u64)len);
1115 pages = kmalloc_array(max_pages,
1120 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1125 } else if (page->index !=
1126 (offset + len) >> PAGE_SHIFT) {
1127 if (num_ops >= (from_pool ? CEPH_OSD_SLAB_OPS :
1128 CEPH_OSD_MAX_OPS)) {
1129 redirty_page_for_writepage(wbc, page);
1135 offset = (u64)page_offset(page);
1139 /* note position of first page in fbatch */
1140 dout("%p will write page %p idx %lu\n",
1141 inode, page, page->index);
1143 if (atomic_long_inc_return(&fsc->writeback_count) >
1144 CONGESTION_ON_THRESH(
1145 fsc->mount_options->congestion_kb))
1146 fsc->write_congested = true;
1148 if (IS_ENCRYPTED(inode)) {
1149 pages[locked_pages] =
1150 fscrypt_encrypt_pagecache_blocks(page,
1152 locked_pages ? GFP_NOWAIT : GFP_NOFS);
1153 if (IS_ERR(pages[locked_pages])) {
1154 if (PTR_ERR(pages[locked_pages]) == -EINVAL)
1155 pr_err("%s: inode->i_blkbits=%hhu\n",
1156 __func__, inode->i_blkbits);
1157 /* better not fail on first page! */
1158 BUG_ON(locked_pages == 0);
1159 pages[locked_pages] = NULL;
1160 redirty_page_for_writepage(wbc, page);
1166 pages[locked_pages++] = page;
1169 fbatch.folios[i] = NULL;
1170 len += thp_size(page);
1173 /* did we get anything? */
1175 goto release_folios;
1178 /* shift unused page to beginning of fbatch */
1179 for (j = 0; j < nr_folios; j++) {
1180 if (!fbatch.folios[j])
1183 fbatch.folios[n] = fbatch.folios[j];
1188 if (nr_folios && i == nr_folios &&
1189 locked_pages < max_pages) {
1190 dout("reached end fbatch, trying for more\n");
1191 folio_batch_release(&fbatch);
1192 goto get_more_pages;
1197 offset = ceph_fscrypt_page_offset(pages[0]);
1200 req = ceph_osdc_new_request(&fsc->client->osdc,
1201 &ci->i_layout, vino,
1202 offset, &len, 0, num_ops,
1203 CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1204 snapc, ceph_wbc.truncate_seq,
1205 ceph_wbc.truncate_size, false);
1207 req = ceph_osdc_new_request(&fsc->client->osdc,
1208 &ci->i_layout, vino,
1213 CEPH_OSD_FLAG_WRITE,
1214 snapc, ceph_wbc.truncate_seq,
1215 ceph_wbc.truncate_size, true);
1216 BUG_ON(IS_ERR(req));
1218 BUG_ON(len < ceph_fscrypt_page_offset(pages[locked_pages - 1]) +
1219 thp_size(pages[locked_pages - 1]) - offset);
1221 req->r_callback = writepages_finish;
1222 req->r_inode = inode;
1224 /* Format the osd request message and submit the write */
1228 for (i = 0; i < locked_pages; i++) {
1229 struct page *page = ceph_fscrypt_pagecache_page(pages[i]);
1231 u64 cur_offset = page_offset(page);
1233 * Discontinuity in page range? Ceph can handle that by just passing
1234 * multiple extents in the write op.
1236 if (offset + len != cur_offset) {
1237 /* If it's full, stop here */
1238 if (op_idx + 1 == req->r_num_ops)
1241 /* Kick off an fscache write with what we have so far. */
1242 ceph_fscache_write_to_cache(inode, offset, len, caching);
1244 /* Start a new extent */
1245 osd_req_op_extent_dup_last(req, op_idx,
1246 cur_offset - offset);
1247 dout("writepages got pages at %llu~%llu\n",
1249 osd_req_op_extent_osd_data_pages(req, op_idx,
1252 osd_req_op_extent_update(req, op_idx, len);
1255 offset = cur_offset;
1256 data_pages = pages + i;
1260 set_page_writeback(page);
1262 ceph_set_page_fscache(page);
1263 len += thp_size(page);
1265 ceph_fscache_write_to_cache(inode, offset, len, caching);
1267 if (ceph_wbc.size_stable) {
1268 len = min(len, ceph_wbc.i_size - offset);
1269 } else if (i == locked_pages) {
1270 /* writepages_finish() clears writeback pages
1271 * according to the data length, so make sure
1272 * data length covers all locked pages */
1273 u64 min_len = len + 1 - thp_size(page);
1274 len = get_writepages_data_length(inode, pages[i - 1],
1276 len = max(len, min_len);
1278 if (IS_ENCRYPTED(inode))
1279 len = round_up(len, CEPH_FSCRYPT_BLOCK_SIZE);
1281 dout("writepages got pages at %llu~%llu\n", offset, len);
1283 if (IS_ENCRYPTED(inode) &&
1284 ((offset | len) & ~CEPH_FSCRYPT_BLOCK_MASK))
1285 pr_warn("%s: bad encrypted write offset=%lld len=%llu\n",
1286 __func__, offset, len);
1288 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1289 0, from_pool, false);
1290 osd_req_op_extent_update(req, op_idx, len);
1292 BUG_ON(op_idx + 1 != req->r_num_ops);
1295 if (i < locked_pages) {
1296 BUG_ON(num_ops <= req->r_num_ops);
1297 num_ops -= req->r_num_ops;
1300 /* allocate new pages array for next request */
1302 pages = kmalloc_array(locked_pages, sizeof(*pages),
1306 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1309 memcpy(pages, data_pages + i,
1310 locked_pages * sizeof(*pages));
1311 memset(data_pages + i, 0,
1312 locked_pages * sizeof(*pages));
1314 BUG_ON(num_ops != req->r_num_ops);
1315 index = pages[i - 1]->index + 1;
1316 /* request message now owns the pages array */
1320 req->r_mtime = inode->i_mtime;
1321 ceph_osdc_start_request(&fsc->client->osdc, req);
1324 wbc->nr_to_write -= i;
1329 * We stop writing back only if we are not doing
1330 * integrity sync. In case of integrity sync we have to
1331 * keep going until we have written all the pages
1332 * we tagged for writeback prior to entering this loop.
1334 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1338 dout("folio_batch release on %d folios (%p)\n", (int)fbatch.nr,
1339 fbatch.nr ? fbatch.folios[0] : NULL);
1340 folio_batch_release(&fbatch);
1343 if (should_loop && !done) {
1344 /* more to do; loop back to beginning of file */
1345 dout("writepages looping back to beginning of file\n");
1346 end = start_index - 1; /* OK even when start_index == 0 */
1348 /* to write dirty pages associated with next snapc,
1349 * we need to wait until current writes complete */
1350 if (wbc->sync_mode != WB_SYNC_NONE &&
1351 start_index == 0 && /* all dirty pages were checked */
1352 !ceph_wbc.head_snapc) {
1356 while ((index <= end) &&
1357 (nr = filemap_get_folios_tag(mapping, &index,
1359 PAGECACHE_TAG_WRITEBACK,
1361 for (i = 0; i < nr; i++) {
1362 page = &fbatch.folios[i]->page;
1363 if (page_snap_context(page) != snapc)
1365 wait_on_page_writeback(page);
1367 folio_batch_release(&fbatch);
1377 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1378 mapping->writeback_index = index;
1381 ceph_osdc_put_request(req);
1382 ceph_put_snap_context(last_snapc);
1383 dout("writepages dend - startone, rc = %d\n", rc);
1390 * See if a given @snapc is either writeable, or already written.
1392 static int context_is_writeable_or_written(struct inode *inode,
1393 struct ceph_snap_context *snapc)
1395 struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1396 int ret = !oldest || snapc->seq <= oldest->seq;
1398 ceph_put_snap_context(oldest);
1403 * ceph_find_incompatible - find an incompatible context and return it
1404 * @page: page being dirtied
1406 * We are only allowed to write into/dirty a page if the page is
1407 * clean, or already dirty within the same snap context. Returns a
1408 * conflicting context if there is one, NULL if there isn't, or a
1409 * negative error code on other errors.
1411 * Must be called with page lock held.
1413 static struct ceph_snap_context *
1414 ceph_find_incompatible(struct page *page)
1416 struct inode *inode = page->mapping->host;
1417 struct ceph_inode_info *ci = ceph_inode(inode);
1419 if (ceph_inode_is_shutdown(inode)) {
1420 dout(" page %p %llx:%llx is shutdown\n", page,
1422 return ERR_PTR(-ESTALE);
1426 struct ceph_snap_context *snapc, *oldest;
1428 wait_on_page_writeback(page);
1430 snapc = page_snap_context(page);
1431 if (!snapc || snapc == ci->i_head_snapc)
1435 * this page is already dirty in another (older) snap
1436 * context! is it writeable now?
1438 oldest = get_oldest_context(inode, NULL, NULL);
1439 if (snapc->seq > oldest->seq) {
1440 /* not writeable -- return it for the caller to deal with */
1441 ceph_put_snap_context(oldest);
1442 dout(" page %p snapc %p not current or oldest\n", page, snapc);
1443 return ceph_get_snap_context(snapc);
1445 ceph_put_snap_context(oldest);
1447 /* yay, writeable, do it now (without dropping page lock) */
1448 dout(" page %p snapc %p not current, but oldest\n", page, snapc);
1449 if (clear_page_dirty_for_io(page)) {
1450 int r = writepage_nounlock(page, NULL);
1458 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
1459 struct folio **foliop, void **_fsdata)
1461 struct inode *inode = file_inode(file);
1462 struct ceph_inode_info *ci = ceph_inode(inode);
1463 struct ceph_snap_context *snapc;
1465 snapc = ceph_find_incompatible(folio_page(*foliop, 0));
1469 folio_unlock(*foliop);
1473 return PTR_ERR(snapc);
1475 ceph_queue_writeback(inode);
1476 r = wait_event_killable(ci->i_cap_wq,
1477 context_is_writeable_or_written(inode, snapc));
1478 ceph_put_snap_context(snapc);
1479 return r == 0 ? -EAGAIN : r;
1485 * We are only allowed to write into/dirty the page if the page is
1486 * clean, or already dirty within the same snap context.
1488 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1489 loff_t pos, unsigned len,
1490 struct page **pagep, void **fsdata)
1492 struct inode *inode = file_inode(file);
1493 struct ceph_inode_info *ci = ceph_inode(inode);
1494 struct folio *folio = NULL;
1497 r = netfs_write_begin(&ci->netfs, file, inode->i_mapping, pos, len, &folio, NULL);
1501 folio_wait_fscache(folio);
1502 WARN_ON_ONCE(!folio_test_locked(folio));
1503 *pagep = &folio->page;
1508 * we don't do anything in here that simple_write_end doesn't do
1509 * except adjust dirty page accounting
1511 static int ceph_write_end(struct file *file, struct address_space *mapping,
1512 loff_t pos, unsigned len, unsigned copied,
1513 struct page *subpage, void *fsdata)
1515 struct folio *folio = page_folio(subpage);
1516 struct inode *inode = file_inode(file);
1517 bool check_cap = false;
1519 dout("write_end file %p inode %p folio %p %d~%d (%d)\n", file,
1520 inode, folio, (int)pos, (int)copied, (int)len);
1522 if (!folio_test_uptodate(folio)) {
1523 /* just return that nothing was copied on a short copy */
1528 folio_mark_uptodate(folio);
1531 /* did file size increase? */
1532 if (pos+copied > i_size_read(inode))
1533 check_cap = ceph_inode_set_size(inode, pos+copied);
1535 folio_mark_dirty(folio);
1538 folio_unlock(folio);
1542 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY);
1547 const struct address_space_operations ceph_aops = {
1548 .read_folio = netfs_read_folio,
1549 .readahead = netfs_readahead,
1550 .writepage = ceph_writepage,
1551 .writepages = ceph_writepages_start,
1552 .write_begin = ceph_write_begin,
1553 .write_end = ceph_write_end,
1554 .dirty_folio = ceph_dirty_folio,
1555 .invalidate_folio = ceph_invalidate_folio,
1556 .release_folio = ceph_release_folio,
1557 .direct_IO = noop_direct_IO,
1560 static void ceph_block_sigs(sigset_t *oldset)
1563 siginitsetinv(&mask, sigmask(SIGKILL));
1564 sigprocmask(SIG_BLOCK, &mask, oldset);
1567 static void ceph_restore_sigs(sigset_t *oldset)
1569 sigprocmask(SIG_SETMASK, oldset, NULL);
1575 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1577 struct vm_area_struct *vma = vmf->vma;
1578 struct inode *inode = file_inode(vma->vm_file);
1579 struct ceph_inode_info *ci = ceph_inode(inode);
1580 struct ceph_file_info *fi = vma->vm_file->private_data;
1581 loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1584 vm_fault_t ret = VM_FAULT_SIGBUS;
1586 if (ceph_inode_is_shutdown(inode))
1589 ceph_block_sigs(&oldset);
1591 dout("filemap_fault %p %llx.%llx %llu trying to get caps\n",
1592 inode, ceph_vinop(inode), off);
1593 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1594 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1596 want = CEPH_CAP_FILE_CACHE;
1599 err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1, &got);
1603 dout("filemap_fault %p %llu got cap refs on %s\n",
1604 inode, off, ceph_cap_string(got));
1606 if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1607 !ceph_has_inline_data(ci)) {
1608 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1609 ceph_add_rw_context(fi, &rw_ctx);
1610 ret = filemap_fault(vmf);
1611 ceph_del_rw_context(fi, &rw_ctx);
1612 dout("filemap_fault %p %llu drop cap refs %s ret %x\n",
1613 inode, off, ceph_cap_string(got), ret);
1617 ceph_put_cap_refs(ci, got);
1622 /* read inline data */
1623 if (off >= PAGE_SIZE) {
1624 /* does not support inline data > PAGE_SIZE */
1625 ret = VM_FAULT_SIGBUS;
1627 struct address_space *mapping = inode->i_mapping;
1630 filemap_invalidate_lock_shared(mapping);
1631 page = find_or_create_page(mapping, 0,
1632 mapping_gfp_constraint(mapping, ~__GFP_FS));
1637 err = __ceph_do_getattr(inode, page,
1638 CEPH_STAT_CAP_INLINE_DATA, true);
1639 if (err < 0 || off >= i_size_read(inode)) {
1642 ret = vmf_error(err);
1645 if (err < PAGE_SIZE)
1646 zero_user_segment(page, err, PAGE_SIZE);
1648 flush_dcache_page(page);
1649 SetPageUptodate(page);
1651 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1653 filemap_invalidate_unlock_shared(mapping);
1654 dout("filemap_fault %p %llu read inline data ret %x\n",
1658 ceph_restore_sigs(&oldset);
1660 ret = vmf_error(err);
1665 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1667 struct vm_area_struct *vma = vmf->vma;
1668 struct inode *inode = file_inode(vma->vm_file);
1669 struct ceph_inode_info *ci = ceph_inode(inode);
1670 struct ceph_file_info *fi = vma->vm_file->private_data;
1671 struct ceph_cap_flush *prealloc_cf;
1672 struct page *page = vmf->page;
1673 loff_t off = page_offset(page);
1674 loff_t size = i_size_read(inode);
1678 vm_fault_t ret = VM_FAULT_SIGBUS;
1680 if (ceph_inode_is_shutdown(inode))
1683 prealloc_cf = ceph_alloc_cap_flush();
1685 return VM_FAULT_OOM;
1687 sb_start_pagefault(inode->i_sb);
1688 ceph_block_sigs(&oldset);
1690 if (off + thp_size(page) <= size)
1691 len = thp_size(page);
1693 len = offset_in_thp(page, size);
1695 dout("page_mkwrite %p %llx.%llx %llu~%zd getting caps i_size %llu\n",
1696 inode, ceph_vinop(inode), off, len, size);
1697 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1698 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1700 want = CEPH_CAP_FILE_BUFFER;
1703 err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len, &got);
1707 dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
1708 inode, off, len, ceph_cap_string(got));
1710 /* Update time before taking page lock */
1711 file_update_time(vma->vm_file);
1712 inode_inc_iversion_raw(inode);
1715 struct ceph_snap_context *snapc;
1719 if (page_mkwrite_check_truncate(page, inode) < 0) {
1721 ret = VM_FAULT_NOPAGE;
1725 snapc = ceph_find_incompatible(page);
1727 /* success. we'll keep the page locked. */
1728 set_page_dirty(page);
1729 ret = VM_FAULT_LOCKED;
1735 if (IS_ERR(snapc)) {
1736 ret = VM_FAULT_SIGBUS;
1740 ceph_queue_writeback(inode);
1741 err = wait_event_killable(ci->i_cap_wq,
1742 context_is_writeable_or_written(inode, snapc));
1743 ceph_put_snap_context(snapc);
1746 if (ret == VM_FAULT_LOCKED) {
1748 spin_lock(&ci->i_ceph_lock);
1749 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1751 spin_unlock(&ci->i_ceph_lock);
1753 __mark_inode_dirty(inode, dirty);
1756 dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %x\n",
1757 inode, off, len, ceph_cap_string(got), ret);
1758 ceph_put_cap_refs_async(ci, got);
1760 ceph_restore_sigs(&oldset);
1761 sb_end_pagefault(inode->i_sb);
1762 ceph_free_cap_flush(prealloc_cf);
1764 ret = vmf_error(err);
1768 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1769 char *data, size_t len)
1771 struct address_space *mapping = inode->i_mapping;
1777 if (i_size_read(inode) == 0)
1779 page = find_or_create_page(mapping, 0,
1780 mapping_gfp_constraint(mapping,
1784 if (PageUptodate(page)) {
1791 dout("fill_inline_data %p %llx.%llx len %zu locked_page %p\n",
1792 inode, ceph_vinop(inode), len, locked_page);
1795 void *kaddr = kmap_atomic(page);
1796 memcpy(kaddr, data, len);
1797 kunmap_atomic(kaddr);
1800 if (page != locked_page) {
1801 if (len < PAGE_SIZE)
1802 zero_user_segment(page, len, PAGE_SIZE);
1804 flush_dcache_page(page);
1806 SetPageUptodate(page);
1812 int ceph_uninline_data(struct file *file)
1814 struct inode *inode = file_inode(file);
1815 struct ceph_inode_info *ci = ceph_inode(inode);
1816 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1817 struct ceph_osd_request *req = NULL;
1818 struct ceph_cap_flush *prealloc_cf = NULL;
1819 struct folio *folio = NULL;
1820 u64 inline_version = CEPH_INLINE_NONE;
1821 struct page *pages[1];
1825 spin_lock(&ci->i_ceph_lock);
1826 inline_version = ci->i_inline_version;
1827 spin_unlock(&ci->i_ceph_lock);
1829 dout("uninline_data %p %llx.%llx inline_version %llu\n",
1830 inode, ceph_vinop(inode), inline_version);
1832 if (ceph_inode_is_shutdown(inode)) {
1837 if (inline_version == CEPH_INLINE_NONE)
1840 prealloc_cf = ceph_alloc_cap_flush();
1844 if (inline_version == 1) /* initial version, no data */
1847 folio = read_mapping_folio(inode->i_mapping, 0, file);
1848 if (IS_ERR(folio)) {
1849 err = PTR_ERR(folio);
1855 len = i_size_read(inode);
1856 if (len > folio_size(folio))
1857 len = folio_size(folio);
1859 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1860 ceph_vino(inode), 0, &len, 0, 1,
1861 CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1868 req->r_mtime = inode->i_mtime;
1869 ceph_osdc_start_request(&fsc->client->osdc, req);
1870 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1871 ceph_osdc_put_request(req);
1875 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1876 ceph_vino(inode), 0, &len, 1, 3,
1877 CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1878 NULL, ci->i_truncate_seq,
1879 ci->i_truncate_size, false);
1885 pages[0] = folio_page(folio, 0);
1886 osd_req_op_extent_osd_data_pages(req, 1, pages, len, 0, false, false);
1889 __le64 xattr_buf = cpu_to_le64(inline_version);
1890 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1891 "inline_version", &xattr_buf,
1893 CEPH_OSD_CMPXATTR_OP_GT,
1894 CEPH_OSD_CMPXATTR_MODE_U64);
1901 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1902 "%llu", inline_version);
1903 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1905 xattr_buf, xattr_len, 0, 0);
1910 req->r_mtime = inode->i_mtime;
1911 ceph_osdc_start_request(&fsc->client->osdc, req);
1912 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1914 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1915 req->r_end_latency, len, err);
1921 /* Set to CAP_INLINE_NONE and dirty the caps */
1922 down_read(&fsc->mdsc->snap_rwsem);
1923 spin_lock(&ci->i_ceph_lock);
1924 ci->i_inline_version = CEPH_INLINE_NONE;
1925 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR, &prealloc_cf);
1926 spin_unlock(&ci->i_ceph_lock);
1927 up_read(&fsc->mdsc->snap_rwsem);
1929 __mark_inode_dirty(inode, dirty);
1932 ceph_osdc_put_request(req);
1933 if (err == -ECANCELED)
1937 folio_unlock(folio);
1941 ceph_free_cap_flush(prealloc_cf);
1942 dout("uninline_data %p %llx.%llx inline_version %llu = %d\n",
1943 inode, ceph_vinop(inode), inline_version, err);
1947 static const struct vm_operations_struct ceph_vmops = {
1948 .fault = ceph_filemap_fault,
1949 .page_mkwrite = ceph_page_mkwrite,
1952 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1954 struct address_space *mapping = file->f_mapping;
1956 if (!mapping->a_ops->read_folio)
1958 vma->vm_ops = &ceph_vmops;
1967 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
1968 s64 pool, struct ceph_string *pool_ns)
1970 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->netfs.inode);
1971 struct ceph_mds_client *mdsc = fsc->mdsc;
1972 struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1973 struct rb_node **p, *parent;
1974 struct ceph_pool_perm *perm;
1975 struct page **pages;
1977 int err = 0, err2 = 0, have = 0;
1979 down_read(&mdsc->pool_perm_rwsem);
1980 p = &mdsc->pool_perm_tree.rb_node;
1982 perm = rb_entry(*p, struct ceph_pool_perm, node);
1983 if (pool < perm->pool)
1985 else if (pool > perm->pool)
1986 p = &(*p)->rb_right;
1988 int ret = ceph_compare_string(pool_ns,
1994 p = &(*p)->rb_right;
2001 up_read(&mdsc->pool_perm_rwsem);
2006 dout("__ceph_pool_perm_get pool %lld ns %.*s no perm cached\n",
2007 pool, (int)pool_ns->len, pool_ns->str);
2009 dout("__ceph_pool_perm_get pool %lld no perm cached\n", pool);
2011 down_write(&mdsc->pool_perm_rwsem);
2012 p = &mdsc->pool_perm_tree.rb_node;
2016 perm = rb_entry(parent, struct ceph_pool_perm, node);
2017 if (pool < perm->pool)
2019 else if (pool > perm->pool)
2020 p = &(*p)->rb_right;
2022 int ret = ceph_compare_string(pool_ns,
2028 p = &(*p)->rb_right;
2036 up_write(&mdsc->pool_perm_rwsem);
2040 rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2041 1, false, GFP_NOFS);
2047 rd_req->r_flags = CEPH_OSD_FLAG_READ;
2048 osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
2049 rd_req->r_base_oloc.pool = pool;
2051 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
2052 ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
2054 err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
2058 wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2059 1, false, GFP_NOFS);
2065 wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
2066 osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
2067 ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
2068 ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
2070 err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
2074 /* one page should be large enough for STAT data */
2075 pages = ceph_alloc_page_vector(1, GFP_KERNEL);
2076 if (IS_ERR(pages)) {
2077 err = PTR_ERR(pages);
2081 osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
2083 ceph_osdc_start_request(&fsc->client->osdc, rd_req);
2085 wr_req->r_mtime = ci->netfs.inode.i_mtime;
2086 ceph_osdc_start_request(&fsc->client->osdc, wr_req);
2088 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
2089 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
2091 if (err >= 0 || err == -ENOENT)
2093 else if (err != -EPERM) {
2094 if (err == -EBLOCKLISTED)
2095 fsc->blocklisted = true;
2099 if (err2 == 0 || err2 == -EEXIST)
2101 else if (err2 != -EPERM) {
2102 if (err2 == -EBLOCKLISTED)
2103 fsc->blocklisted = true;
2108 pool_ns_len = pool_ns ? pool_ns->len : 0;
2109 perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
2117 perm->pool_ns_len = pool_ns_len;
2118 if (pool_ns_len > 0)
2119 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
2120 perm->pool_ns[pool_ns_len] = 0;
2122 rb_link_node(&perm->node, parent, p);
2123 rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
2126 up_write(&mdsc->pool_perm_rwsem);
2128 ceph_osdc_put_request(rd_req);
2129 ceph_osdc_put_request(wr_req);
2134 dout("__ceph_pool_perm_get pool %lld ns %.*s result = %d\n",
2135 pool, (int)pool_ns->len, pool_ns->str, err);
2137 dout("__ceph_pool_perm_get pool %lld result = %d\n", pool, err);
2141 int ceph_pool_perm_check(struct inode *inode, int need)
2143 struct ceph_inode_info *ci = ceph_inode(inode);
2144 struct ceph_string *pool_ns;
2148 /* Only need to do this for regular files */
2149 if (!S_ISREG(inode->i_mode))
2152 if (ci->i_vino.snap != CEPH_NOSNAP) {
2154 * Pool permission check needs to write to the first object.
2155 * But for snapshot, head of the first object may have alread
2156 * been deleted. Skip check to avoid creating orphan object.
2161 if (ceph_test_mount_opt(ceph_inode_to_client(inode),
2165 spin_lock(&ci->i_ceph_lock);
2166 flags = ci->i_ceph_flags;
2167 pool = ci->i_layout.pool_id;
2168 spin_unlock(&ci->i_ceph_lock);
2170 if (flags & CEPH_I_POOL_PERM) {
2171 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2172 dout("ceph_pool_perm_check pool %lld no read perm\n",
2176 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2177 dout("ceph_pool_perm_check pool %lld no write perm\n",
2184 pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2185 ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2186 ceph_put_string(pool_ns);
2190 flags = CEPH_I_POOL_PERM;
2191 if (ret & POOL_READ)
2192 flags |= CEPH_I_POOL_RD;
2193 if (ret & POOL_WRITE)
2194 flags |= CEPH_I_POOL_WR;
2196 spin_lock(&ci->i_ceph_lock);
2197 if (pool == ci->i_layout.pool_id &&
2198 pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2199 ci->i_ceph_flags |= flags;
2201 pool = ci->i_layout.pool_id;
2202 flags = ci->i_ceph_flags;
2204 spin_unlock(&ci->i_ceph_lock);
2208 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2210 struct ceph_pool_perm *perm;
2213 while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2214 n = rb_first(&mdsc->pool_perm_tree);
2215 perm = rb_entry(n, struct ceph_pool_perm, node);
2216 rb_erase(n, &mdsc->pool_perm_tree);