1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
3
4 #include <linux/backing-dev.h>
5 #include <linux/fs.h>
6 #include <linux/mm.h>
7 #include <linux/swap.h>
8 #include <linux/pagemap.h>
9 #include <linux/slab.h>
10 #include <linux/pagevec.h>
11 #include <linux/task_io_accounting_ops.h>
12 #include <linux/signal.h>
13 #include <linux/iversion.h>
14 #include <linux/ktime.h>
15 #include <linux/netfs.h>
16 #include <trace/events/netfs.h>
17
18 #include "super.h"
19 #include "mds_client.h"
20 #include "cache.h"
21 #include "metric.h"
22 #include "crypto.h"
23 #include <linux/ceph/osd_client.h>
24 #include <linux/ceph/striper.h>
25
26 /*
27 * Ceph address space ops.
28 *
29 * There are a few funny things going on here.
30 *
31 * The page->private field is used to reference a struct
32 * ceph_snap_context for _every_ dirty page. This indicates which
33 * snapshot the page was logically dirtied in, and thus which snap
34 * context needs to be associated with the osd write during writeback.
35 *
36 * Similarly, struct ceph_inode_info maintains a set of counters to
37 * count dirty pages on the inode. In the absence of snapshots,
38 * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
39 *
40 * When a snapshot is taken (that is, when the client receives
41 * notification that a snapshot was taken), each inode with caps and
42 * with dirty pages (dirty pages implies there is a cap) gets a new
43 * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
44 * order, new snaps go to the tail). The i_wrbuffer_ref_head count is
45 * moved to capsnap->dirty. (Unless a sync write is currently in
46 * progress. In that case, the capsnap is said to be "pending", new
47 * writes cannot start, and the capsnap isn't "finalized" until the
48 * write completes (or fails) and a final size/mtime for the inode for
49 * that snap can be settled upon.) i_wrbuffer_ref_head is reset to 0.
50 *
51 * On writeback, we must submit writes to the osd IN SNAP ORDER. So,
52 * we look for the first capsnap in i_cap_snaps and write out pages in
53 * that snap context _only_. Then we move on to the next capsnap,
54 * eventually reaching the "live" or "head" context (i.e., pages that
55 * are not yet snapped) and are writing the most recently dirtied
56 * pages.
57 *
58 * Invalidate and so forth must take care to ensure the dirty page
59 * accounting is preserved.
60 */
61
62 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
63 #define CONGESTION_OFF_THRESH(congestion_kb) \
64 (CONGESTION_ON_THRESH(congestion_kb) - \
65 (CONGESTION_ON_THRESH(congestion_kb) >> 2))
66
67 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
68 struct folio **foliop, void **_fsdata);
69
page_snap_context(struct page * page)70 static inline struct ceph_snap_context *page_snap_context(struct page *page)
71 {
72 if (PagePrivate(page))
73 return (void *)page->private;
74 return NULL;
75 }
76
77 /*
78 * Dirty a page. Optimistically adjust accounting, on the assumption
79 * that we won't race with invalidate. If we do, readjust.
80 */
ceph_dirty_folio(struct address_space * mapping,struct folio * folio)81 static bool ceph_dirty_folio(struct address_space *mapping, struct folio *folio)
82 {
83 struct inode *inode = mapping->host;
84 struct ceph_client *cl = ceph_inode_to_client(inode);
85 struct ceph_inode_info *ci;
86 struct ceph_snap_context *snapc;
87
88 if (folio_test_dirty(folio)) {
89 doutc(cl, "%llx.%llx %p idx %lu -- already dirty\n",
90 ceph_vinop(inode), folio, folio->index);
91 VM_BUG_ON_FOLIO(!folio_test_private(folio), folio);
92 return false;
93 }
94
95 ci = ceph_inode(inode);
96
97 /* dirty the head */
98 spin_lock(&ci->i_ceph_lock);
99 if (__ceph_have_pending_cap_snap(ci)) {
100 struct ceph_cap_snap *capsnap =
101 list_last_entry(&ci->i_cap_snaps,
102 struct ceph_cap_snap,
103 ci_item);
104 snapc = ceph_get_snap_context(capsnap->context);
105 capsnap->dirty_pages++;
106 } else {
107 BUG_ON(!ci->i_head_snapc);
108 snapc = ceph_get_snap_context(ci->i_head_snapc);
109 ++ci->i_wrbuffer_ref_head;
110 }
111 if (ci->i_wrbuffer_ref == 0)
112 ihold(inode);
113 ++ci->i_wrbuffer_ref;
114 doutc(cl, "%llx.%llx %p idx %lu head %d/%d -> %d/%d "
115 "snapc %p seq %lld (%d snaps)\n",
116 ceph_vinop(inode), folio, folio->index,
117 ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
118 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
119 snapc, snapc->seq, snapc->num_snaps);
120 spin_unlock(&ci->i_ceph_lock);
121
122 /*
123 * Reference snap context in folio->private. Also set
124 * PagePrivate so that we get invalidate_folio callback.
125 */
126 VM_WARN_ON_FOLIO(folio->private, folio);
127 folio_attach_private(folio, snapc);
128
129 return ceph_fscache_dirty_folio(mapping, folio);
130 }
131
132 /*
133 * If we are truncating the full folio (i.e. offset == 0), adjust the
134 * dirty folio counters appropriately. Only called if there is private
135 * data on the folio.
136 */
ceph_invalidate_folio(struct folio * folio,size_t offset,size_t length)137 static void ceph_invalidate_folio(struct folio *folio, size_t offset,
138 size_t length)
139 {
140 struct inode *inode = folio->mapping->host;
141 struct ceph_client *cl = ceph_inode_to_client(inode);
142 struct ceph_inode_info *ci = ceph_inode(inode);
143 struct ceph_snap_context *snapc;
144
145
146 if (offset != 0 || length != folio_size(folio)) {
147 doutc(cl, "%llx.%llx idx %lu partial dirty page %zu~%zu\n",
148 ceph_vinop(inode), folio->index, offset, length);
149 return;
150 }
151
152 WARN_ON(!folio_test_locked(folio));
153 if (folio_test_private(folio)) {
154 doutc(cl, "%llx.%llx idx %lu full dirty page\n",
155 ceph_vinop(inode), folio->index);
156
157 snapc = folio_detach_private(folio);
158 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
159 ceph_put_snap_context(snapc);
160 }
161
162 netfs_invalidate_folio(folio, offset, length);
163 }
164
ceph_netfs_expand_readahead(struct netfs_io_request * rreq)165 static void ceph_netfs_expand_readahead(struct netfs_io_request *rreq)
166 {
167 struct inode *inode = rreq->inode;
168 struct ceph_inode_info *ci = ceph_inode(inode);
169 struct ceph_file_layout *lo = &ci->i_layout;
170 unsigned long max_pages = inode->i_sb->s_bdi->ra_pages;
171 loff_t end = rreq->start + rreq->len, new_end;
172 struct ceph_netfs_request_data *priv = rreq->netfs_priv;
173 unsigned long max_len;
174 u32 blockoff;
175
176 if (priv) {
177 /* Readahead is disabled by posix_fadvise POSIX_FADV_RANDOM */
178 if (priv->file_ra_disabled)
179 max_pages = 0;
180 else
181 max_pages = priv->file_ra_pages;
182
183 }
184
185 /* Readahead is disabled */
186 if (!max_pages)
187 return;
188
189 max_len = max_pages << PAGE_SHIFT;
190
191 /*
192 * Try to expand the length forward by rounding up it to the next
193 * block, but do not exceed the file size, unless the original
194 * request already exceeds it.
195 */
196 new_end = umin(round_up(end, lo->stripe_unit), rreq->i_size);
197 if (new_end > end && new_end <= rreq->start + max_len)
198 rreq->len = new_end - rreq->start;
199
200 /* Try to expand the start downward */
201 div_u64_rem(rreq->start, lo->stripe_unit, &blockoff);
202 if (rreq->len + blockoff <= max_len) {
203 rreq->start -= blockoff;
204 rreq->len += blockoff;
205 }
206 }
207
finish_netfs_read(struct ceph_osd_request * req)208 static void finish_netfs_read(struct ceph_osd_request *req)
209 {
210 struct inode *inode = req->r_inode;
211 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
212 struct ceph_client *cl = fsc->client;
213 struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
214 struct netfs_io_subrequest *subreq = req->r_priv;
215 struct ceph_osd_req_op *op = &req->r_ops[0];
216 int err = req->r_result;
217 bool sparse = (op->op == CEPH_OSD_OP_SPARSE_READ);
218
219 ceph_update_read_metrics(&fsc->mdsc->metric, req->r_start_latency,
220 req->r_end_latency, osd_data->length, err);
221
222 doutc(cl, "result %d subreq->len=%zu i_size=%lld\n", req->r_result,
223 subreq->len, i_size_read(req->r_inode));
224
225 /* no object means success but no data */
226 if (err == -ENOENT)
227 err = 0;
228 else if (err == -EBLOCKLISTED)
229 fsc->blocklisted = true;
230
231 if (err >= 0) {
232 if (sparse && err > 0)
233 err = ceph_sparse_ext_map_end(op);
234 if (err < subreq->len &&
235 subreq->rreq->origin != NETFS_DIO_READ)
236 __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
237 if (IS_ENCRYPTED(inode) && err > 0) {
238 err = ceph_fscrypt_decrypt_extents(inode,
239 osd_data->pages, subreq->start,
240 op->extent.sparse_ext,
241 op->extent.sparse_ext_cnt);
242 if (err > subreq->len)
243 err = subreq->len;
244 }
245 }
246
247 if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
248 ceph_put_page_vector(osd_data->pages,
249 calc_pages_for(osd_data->alignment,
250 osd_data->length), false);
251 }
252 if (err > 0) {
253 subreq->transferred = err;
254 err = 0;
255 }
256 trace_netfs_sreq(subreq, netfs_sreq_trace_io_progress);
257 netfs_read_subreq_terminated(subreq, err, false);
258 iput(req->r_inode);
259 ceph_dec_osd_stopping_blocker(fsc->mdsc);
260 }
261
ceph_netfs_issue_op_inline(struct netfs_io_subrequest * subreq)262 static bool ceph_netfs_issue_op_inline(struct netfs_io_subrequest *subreq)
263 {
264 struct netfs_io_request *rreq = subreq->rreq;
265 struct inode *inode = rreq->inode;
266 struct ceph_mds_reply_info_parsed *rinfo;
267 struct ceph_mds_reply_info_in *iinfo;
268 struct ceph_mds_request *req;
269 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
270 struct ceph_inode_info *ci = ceph_inode(inode);
271 ssize_t err = 0;
272 size_t len;
273 int mode;
274
275 if (rreq->origin != NETFS_DIO_READ)
276 __set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
277 __clear_bit(NETFS_SREQ_COPY_TO_CACHE, &subreq->flags);
278
279 if (subreq->start >= inode->i_size)
280 goto out;
281
282 /* We need to fetch the inline data. */
283 mode = ceph_try_to_choose_auth_mds(inode, CEPH_STAT_CAP_INLINE_DATA);
284 req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_GETATTR, mode);
285 if (IS_ERR(req)) {
286 err = PTR_ERR(req);
287 goto out;
288 }
289 req->r_ino1 = ci->i_vino;
290 req->r_args.getattr.mask = cpu_to_le32(CEPH_STAT_CAP_INLINE_DATA);
291 req->r_num_caps = 2;
292
293 trace_netfs_sreq(subreq, netfs_sreq_trace_submit);
294 err = ceph_mdsc_do_request(mdsc, NULL, req);
295 if (err < 0)
296 goto out;
297
298 rinfo = &req->r_reply_info;
299 iinfo = &rinfo->targeti;
300 if (iinfo->inline_version == CEPH_INLINE_NONE) {
301 /* The data got uninlined */
302 ceph_mdsc_put_request(req);
303 return false;
304 }
305
306 len = min_t(size_t, iinfo->inline_len - subreq->start, subreq->len);
307 err = copy_to_iter(iinfo->inline_data + subreq->start, len, &subreq->io_iter);
308 if (err == 0) {
309 err = -EFAULT;
310 } else {
311 subreq->transferred += err;
312 err = 0;
313 }
314
315 ceph_mdsc_put_request(req);
316 out:
317 netfs_read_subreq_terminated(subreq, err, false);
318 return true;
319 }
320
ceph_netfs_prepare_read(struct netfs_io_subrequest * subreq)321 static int ceph_netfs_prepare_read(struct netfs_io_subrequest *subreq)
322 {
323 struct netfs_io_request *rreq = subreq->rreq;
324 struct inode *inode = rreq->inode;
325 struct ceph_inode_info *ci = ceph_inode(inode);
326 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
327 u64 objno, objoff;
328 u32 xlen;
329
330 /* Truncate the extent at the end of the current block */
331 ceph_calc_file_object_mapping(&ci->i_layout, subreq->start, subreq->len,
332 &objno, &objoff, &xlen);
333 rreq->io_streams[0].sreq_max_len = umin(xlen, fsc->mount_options->rsize);
334 return 0;
335 }
336
ceph_netfs_issue_read(struct netfs_io_subrequest * subreq)337 static void ceph_netfs_issue_read(struct netfs_io_subrequest *subreq)
338 {
339 struct netfs_io_request *rreq = subreq->rreq;
340 struct inode *inode = rreq->inode;
341 struct ceph_inode_info *ci = ceph_inode(inode);
342 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
343 struct ceph_client *cl = fsc->client;
344 struct ceph_osd_request *req = NULL;
345 struct ceph_vino vino = ceph_vino(inode);
346 int err;
347 u64 len;
348 bool sparse = IS_ENCRYPTED(inode) || ceph_test_mount_opt(fsc, SPARSEREAD);
349 u64 off = subreq->start;
350 int extent_cnt;
351
352 if (ceph_inode_is_shutdown(inode)) {
353 err = -EIO;
354 goto out;
355 }
356
357 if (ceph_has_inline_data(ci) && ceph_netfs_issue_op_inline(subreq))
358 return;
359
360 // TODO: This rounding here is slightly dodgy. It *should* work, for
361 // now, as the cache only deals in blocks that are a multiple of
362 // PAGE_SIZE and fscrypt blocks are at most PAGE_SIZE. What needs to
363 // happen is for the fscrypt driving to be moved into netfslib and the
364 // data in the cache also to be stored encrypted.
365 len = subreq->len;
366 ceph_fscrypt_adjust_off_and_len(inode, &off, &len);
367
368 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, vino,
369 off, &len, 0, 1, sparse ? CEPH_OSD_OP_SPARSE_READ : CEPH_OSD_OP_READ,
370 CEPH_OSD_FLAG_READ, NULL, ci->i_truncate_seq,
371 ci->i_truncate_size, false);
372 if (IS_ERR(req)) {
373 err = PTR_ERR(req);
374 req = NULL;
375 goto out;
376 }
377
378 if (sparse) {
379 extent_cnt = __ceph_sparse_read_ext_count(inode, len);
380 err = ceph_alloc_sparse_ext_map(&req->r_ops[0], extent_cnt);
381 if (err)
382 goto out;
383 }
384
385 doutc(cl, "%llx.%llx pos=%llu orig_len=%zu len=%llu\n",
386 ceph_vinop(inode), subreq->start, subreq->len, len);
387
388 /*
389 * FIXME: For now, use CEPH_OSD_DATA_TYPE_PAGES instead of _ITER for
390 * encrypted inodes. We'd need infrastructure that handles an iov_iter
391 * instead of page arrays, and we don't have that as of yet. Once the
392 * dust settles on the write helpers and encrypt/decrypt routines for
393 * netfs, we should be able to rework this.
394 */
395 if (IS_ENCRYPTED(inode)) {
396 struct page **pages;
397 size_t page_off;
398
399 err = iov_iter_get_pages_alloc2(&subreq->io_iter, &pages, len, &page_off);
400 if (err < 0) {
401 doutc(cl, "%llx.%llx failed to allocate pages, %d\n",
402 ceph_vinop(inode), err);
403 goto out;
404 }
405
406 /* should always give us a page-aligned read */
407 WARN_ON_ONCE(page_off);
408 len = err;
409 err = 0;
410
411 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false,
412 false);
413 } else {
414 osd_req_op_extent_osd_iter(req, 0, &subreq->io_iter);
415 }
416 if (!ceph_inc_osd_stopping_blocker(fsc->mdsc)) {
417 err = -EIO;
418 goto out;
419 }
420 req->r_callback = finish_netfs_read;
421 req->r_priv = subreq;
422 req->r_inode = inode;
423 ihold(inode);
424
425 trace_netfs_sreq(subreq, netfs_sreq_trace_submit);
426 ceph_osdc_start_request(req->r_osdc, req);
427 out:
428 ceph_osdc_put_request(req);
429 if (err)
430 netfs_read_subreq_terminated(subreq, err, false);
431 doutc(cl, "%llx.%llx result %d\n", ceph_vinop(inode), err);
432 }
433
ceph_init_request(struct netfs_io_request * rreq,struct file * file)434 static int ceph_init_request(struct netfs_io_request *rreq, struct file *file)
435 {
436 struct inode *inode = rreq->inode;
437 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
438 struct ceph_client *cl = ceph_inode_to_client(inode);
439 int got = 0, want = CEPH_CAP_FILE_CACHE;
440 struct ceph_netfs_request_data *priv;
441 int ret = 0;
442
443 /* [DEPRECATED] Use PG_private_2 to mark folio being written to the cache. */
444 __set_bit(NETFS_RREQ_USE_PGPRIV2, &rreq->flags);
445
446 if (rreq->origin != NETFS_READAHEAD)
447 return 0;
448
449 priv = kzalloc(sizeof(*priv), GFP_NOFS);
450 if (!priv)
451 return -ENOMEM;
452
453 if (file) {
454 struct ceph_rw_context *rw_ctx;
455 struct ceph_file_info *fi = file->private_data;
456
457 priv->file_ra_pages = file->f_ra.ra_pages;
458 priv->file_ra_disabled = file->f_mode & FMODE_RANDOM;
459
460 rw_ctx = ceph_find_rw_context(fi);
461 if (rw_ctx) {
462 rreq->netfs_priv = priv;
463 return 0;
464 }
465 }
466
467 /*
468 * readahead callers do not necessarily hold Fcb caps
469 * (e.g. fadvise, madvise).
470 */
471 ret = ceph_try_get_caps(inode, CEPH_CAP_FILE_RD, want, true, &got);
472 if (ret < 0) {
473 doutc(cl, "%llx.%llx, error getting cap\n", ceph_vinop(inode));
474 goto out;
475 }
476
477 if (!(got & want)) {
478 doutc(cl, "%llx.%llx, no cache cap\n", ceph_vinop(inode));
479 ret = -EACCES;
480 goto out;
481 }
482 if (ret == 0) {
483 ret = -EACCES;
484 goto out;
485 }
486
487 priv->caps = got;
488 rreq->netfs_priv = priv;
489 rreq->io_streams[0].sreq_max_len = fsc->mount_options->rsize;
490
491 out:
492 if (ret < 0)
493 kfree(priv);
494
495 return ret;
496 }
497
ceph_netfs_free_request(struct netfs_io_request * rreq)498 static void ceph_netfs_free_request(struct netfs_io_request *rreq)
499 {
500 struct ceph_netfs_request_data *priv = rreq->netfs_priv;
501
502 if (!priv)
503 return;
504
505 if (priv->caps)
506 ceph_put_cap_refs(ceph_inode(rreq->inode), priv->caps);
507 kfree(priv);
508 rreq->netfs_priv = NULL;
509 }
510
511 const struct netfs_request_ops ceph_netfs_ops = {
512 .init_request = ceph_init_request,
513 .free_request = ceph_netfs_free_request,
514 .prepare_read = ceph_netfs_prepare_read,
515 .issue_read = ceph_netfs_issue_read,
516 .expand_readahead = ceph_netfs_expand_readahead,
517 .check_write_begin = ceph_netfs_check_write_begin,
518 };
519
520 #ifdef CONFIG_CEPH_FSCACHE
ceph_set_page_fscache(struct page * page)521 static void ceph_set_page_fscache(struct page *page)
522 {
523 folio_start_private_2(page_folio(page)); /* [DEPRECATED] */
524 }
525
ceph_fscache_write_terminated(void * priv,ssize_t error,bool was_async)526 static void ceph_fscache_write_terminated(void *priv, ssize_t error, bool was_async)
527 {
528 struct inode *inode = priv;
529
530 if (IS_ERR_VALUE(error) && error != -ENOBUFS)
531 ceph_fscache_invalidate(inode, false);
532 }
533
ceph_fscache_write_to_cache(struct inode * inode,u64 off,u64 len,bool caching)534 static void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
535 {
536 struct ceph_inode_info *ci = ceph_inode(inode);
537 struct fscache_cookie *cookie = ceph_fscache_cookie(ci);
538
539 fscache_write_to_cache(cookie, inode->i_mapping, off, len, i_size_read(inode),
540 ceph_fscache_write_terminated, inode, true, caching);
541 }
542 #else
ceph_set_page_fscache(struct page * page)543 static inline void ceph_set_page_fscache(struct page *page)
544 {
545 }
546
ceph_fscache_write_to_cache(struct inode * inode,u64 off,u64 len,bool caching)547 static inline void ceph_fscache_write_to_cache(struct inode *inode, u64 off, u64 len, bool caching)
548 {
549 }
550 #endif /* CONFIG_CEPH_FSCACHE */
551
552 struct ceph_writeback_ctl
553 {
554 loff_t i_size;
555 u64 truncate_size;
556 u32 truncate_seq;
557 bool size_stable;
558 bool head_snapc;
559 };
560
561 /*
562 * Get ref for the oldest snapc for an inode with dirty data... that is, the
563 * only snap context we are allowed to write back.
564 */
565 static struct ceph_snap_context *
get_oldest_context(struct inode * inode,struct ceph_writeback_ctl * ctl,struct ceph_snap_context * page_snapc)566 get_oldest_context(struct inode *inode, struct ceph_writeback_ctl *ctl,
567 struct ceph_snap_context *page_snapc)
568 {
569 struct ceph_inode_info *ci = ceph_inode(inode);
570 struct ceph_client *cl = ceph_inode_to_client(inode);
571 struct ceph_snap_context *snapc = NULL;
572 struct ceph_cap_snap *capsnap = NULL;
573
574 spin_lock(&ci->i_ceph_lock);
575 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
576 doutc(cl, " capsnap %p snapc %p has %d dirty pages\n",
577 capsnap, capsnap->context, capsnap->dirty_pages);
578 if (!capsnap->dirty_pages)
579 continue;
580
581 /* get i_size, truncate_{seq,size} for page_snapc? */
582 if (snapc && capsnap->context != page_snapc)
583 continue;
584
585 if (ctl) {
586 if (capsnap->writing) {
587 ctl->i_size = i_size_read(inode);
588 ctl->size_stable = false;
589 } else {
590 ctl->i_size = capsnap->size;
591 ctl->size_stable = true;
592 }
593 ctl->truncate_size = capsnap->truncate_size;
594 ctl->truncate_seq = capsnap->truncate_seq;
595 ctl->head_snapc = false;
596 }
597
598 if (snapc)
599 break;
600
601 snapc = ceph_get_snap_context(capsnap->context);
602 if (!page_snapc ||
603 page_snapc == snapc ||
604 page_snapc->seq > snapc->seq)
605 break;
606 }
607 if (!snapc && ci->i_wrbuffer_ref_head) {
608 snapc = ceph_get_snap_context(ci->i_head_snapc);
609 doutc(cl, " head snapc %p has %d dirty pages\n", snapc,
610 ci->i_wrbuffer_ref_head);
611 if (ctl) {
612 ctl->i_size = i_size_read(inode);
613 ctl->truncate_size = ci->i_truncate_size;
614 ctl->truncate_seq = ci->i_truncate_seq;
615 ctl->size_stable = false;
616 ctl->head_snapc = true;
617 }
618 }
619 spin_unlock(&ci->i_ceph_lock);
620 return snapc;
621 }
622
get_writepages_data_length(struct inode * inode,struct page * page,u64 start)623 static u64 get_writepages_data_length(struct inode *inode,
624 struct page *page, u64 start)
625 {
626 struct ceph_inode_info *ci = ceph_inode(inode);
627 struct ceph_snap_context *snapc;
628 struct ceph_cap_snap *capsnap = NULL;
629 u64 end = i_size_read(inode);
630 u64 ret;
631
632 snapc = page_snap_context(ceph_fscrypt_pagecache_page(page));
633 if (snapc != ci->i_head_snapc) {
634 bool found = false;
635 spin_lock(&ci->i_ceph_lock);
636 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
637 if (capsnap->context == snapc) {
638 if (!capsnap->writing)
639 end = capsnap->size;
640 found = true;
641 break;
642 }
643 }
644 spin_unlock(&ci->i_ceph_lock);
645 WARN_ON(!found);
646 }
647 if (end > ceph_fscrypt_page_offset(page) + thp_size(page))
648 end = ceph_fscrypt_page_offset(page) + thp_size(page);
649 ret = end > start ? end - start : 0;
650 if (ret && fscrypt_is_bounce_page(page))
651 ret = round_up(ret, CEPH_FSCRYPT_BLOCK_SIZE);
652 return ret;
653 }
654
655 /*
656 * Write a single page, but leave the page locked.
657 *
658 * If we get a write error, mark the mapping for error, but still adjust the
659 * dirty page accounting (i.e., page is no longer dirty).
660 */
writepage_nounlock(struct page * page,struct writeback_control * wbc)661 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
662 {
663 struct folio *folio = page_folio(page);
664 struct inode *inode = page->mapping->host;
665 struct ceph_inode_info *ci = ceph_inode(inode);
666 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
667 struct ceph_client *cl = fsc->client;
668 struct ceph_snap_context *snapc, *oldest;
669 loff_t page_off = page_offset(page);
670 int err;
671 loff_t len = thp_size(page);
672 loff_t wlen;
673 struct ceph_writeback_ctl ceph_wbc;
674 struct ceph_osd_client *osdc = &fsc->client->osdc;
675 struct ceph_osd_request *req;
676 bool caching = ceph_is_cache_enabled(inode);
677 struct page *bounce_page = NULL;
678
679 doutc(cl, "%llx.%llx page %p idx %lu\n", ceph_vinop(inode), page,
680 page->index);
681
682 if (ceph_inode_is_shutdown(inode))
683 return -EIO;
684
685 /* verify this is a writeable snap context */
686 snapc = page_snap_context(page);
687 if (!snapc) {
688 doutc(cl, "%llx.%llx page %p not dirty?\n", ceph_vinop(inode),
689 page);
690 return 0;
691 }
692 oldest = get_oldest_context(inode, &ceph_wbc, snapc);
693 if (snapc->seq > oldest->seq) {
694 doutc(cl, "%llx.%llx page %p snapc %p not writeable - noop\n",
695 ceph_vinop(inode), page, snapc);
696 /* we should only noop if called by kswapd */
697 WARN_ON(!(current->flags & PF_MEMALLOC));
698 ceph_put_snap_context(oldest);
699 redirty_page_for_writepage(wbc, page);
700 return 0;
701 }
702 ceph_put_snap_context(oldest);
703
704 /* is this a partial page at end of file? */
705 if (page_off >= ceph_wbc.i_size) {
706 doutc(cl, "%llx.%llx folio at %lu beyond eof %llu\n",
707 ceph_vinop(inode), folio->index, ceph_wbc.i_size);
708 folio_invalidate(folio, 0, folio_size(folio));
709 return 0;
710 }
711
712 if (ceph_wbc.i_size < page_off + len)
713 len = ceph_wbc.i_size - page_off;
714
715 wlen = IS_ENCRYPTED(inode) ? round_up(len, CEPH_FSCRYPT_BLOCK_SIZE) : len;
716 doutc(cl, "%llx.%llx page %p index %lu on %llu~%llu snapc %p seq %lld\n",
717 ceph_vinop(inode), page, page->index, page_off, wlen, snapc,
718 snapc->seq);
719
720 if (atomic_long_inc_return(&fsc->writeback_count) >
721 CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
722 fsc->write_congested = true;
723
724 req = ceph_osdc_new_request(osdc, &ci->i_layout, ceph_vino(inode),
725 page_off, &wlen, 0, 1, CEPH_OSD_OP_WRITE,
726 CEPH_OSD_FLAG_WRITE, snapc,
727 ceph_wbc.truncate_seq,
728 ceph_wbc.truncate_size, true);
729 if (IS_ERR(req)) {
730 redirty_page_for_writepage(wbc, page);
731 return PTR_ERR(req);
732 }
733
734 if (wlen < len)
735 len = wlen;
736
737 set_page_writeback(page);
738 if (caching)
739 ceph_set_page_fscache(page);
740 ceph_fscache_write_to_cache(inode, page_off, len, caching);
741
742 if (IS_ENCRYPTED(inode)) {
743 bounce_page = fscrypt_encrypt_pagecache_blocks(page,
744 CEPH_FSCRYPT_BLOCK_SIZE, 0,
745 GFP_NOFS);
746 if (IS_ERR(bounce_page)) {
747 redirty_page_for_writepage(wbc, page);
748 end_page_writeback(page);
749 ceph_osdc_put_request(req);
750 return PTR_ERR(bounce_page);
751 }
752 }
753
754 /* it may be a short write due to an object boundary */
755 WARN_ON_ONCE(len > thp_size(page));
756 osd_req_op_extent_osd_data_pages(req, 0,
757 bounce_page ? &bounce_page : &page, wlen, 0,
758 false, false);
759 doutc(cl, "%llx.%llx %llu~%llu (%llu bytes, %sencrypted)\n",
760 ceph_vinop(inode), page_off, len, wlen,
761 IS_ENCRYPTED(inode) ? "" : "not ");
762
763 req->r_mtime = inode_get_mtime(inode);
764 ceph_osdc_start_request(osdc, req);
765 err = ceph_osdc_wait_request(osdc, req);
766
767 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
768 req->r_end_latency, len, err);
769 fscrypt_free_bounce_page(bounce_page);
770 ceph_osdc_put_request(req);
771 if (err == 0)
772 err = len;
773
774 if (err < 0) {
775 struct writeback_control tmp_wbc;
776 if (!wbc)
777 wbc = &tmp_wbc;
778 if (err == -ERESTARTSYS) {
779 /* killed by SIGKILL */
780 doutc(cl, "%llx.%llx interrupted page %p\n",
781 ceph_vinop(inode), page);
782 redirty_page_for_writepage(wbc, page);
783 end_page_writeback(page);
784 return err;
785 }
786 if (err == -EBLOCKLISTED)
787 fsc->blocklisted = true;
788 doutc(cl, "%llx.%llx setting page/mapping error %d %p\n",
789 ceph_vinop(inode), err, page);
790 mapping_set_error(&inode->i_data, err);
791 wbc->pages_skipped++;
792 } else {
793 doutc(cl, "%llx.%llx cleaned page %p\n",
794 ceph_vinop(inode), page);
795 err = 0; /* vfs expects us to return 0 */
796 }
797 oldest = detach_page_private(page);
798 WARN_ON_ONCE(oldest != snapc);
799 end_page_writeback(page);
800 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
801 ceph_put_snap_context(snapc); /* page's reference */
802
803 if (atomic_long_dec_return(&fsc->writeback_count) <
804 CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
805 fsc->write_congested = false;
806
807 return err;
808 }
809
ceph_writepage(struct page * page,struct writeback_control * wbc)810 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
811 {
812 int err;
813 struct inode *inode = page->mapping->host;
814 BUG_ON(!inode);
815 ihold(inode);
816
817 if (wbc->sync_mode == WB_SYNC_NONE &&
818 ceph_inode_to_fs_client(inode)->write_congested) {
819 redirty_page_for_writepage(wbc, page);
820 return AOP_WRITEPAGE_ACTIVATE;
821 }
822
823 folio_wait_private_2(page_folio(page)); /* [DEPRECATED] */
824
825 err = writepage_nounlock(page, wbc);
826 if (err == -ERESTARTSYS) {
827 /* direct memory reclaimer was killed by SIGKILL. return 0
828 * to prevent caller from setting mapping/page error */
829 err = 0;
830 }
831 unlock_page(page);
832 iput(inode);
833 return err;
834 }
835
836 /*
837 * async writeback completion handler.
838 *
839 * If we get an error, set the mapping error bit, but not the individual
840 * page error bits.
841 */
writepages_finish(struct ceph_osd_request * req)842 static void writepages_finish(struct ceph_osd_request *req)
843 {
844 struct inode *inode = req->r_inode;
845 struct ceph_inode_info *ci = ceph_inode(inode);
846 struct ceph_client *cl = ceph_inode_to_client(inode);
847 struct ceph_osd_data *osd_data;
848 struct page *page;
849 int num_pages, total_pages = 0;
850 int i, j;
851 int rc = req->r_result;
852 struct ceph_snap_context *snapc = req->r_snapc;
853 struct address_space *mapping = inode->i_mapping;
854 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
855 unsigned int len = 0;
856 bool remove_page;
857
858 doutc(cl, "%llx.%llx rc %d\n", ceph_vinop(inode), rc);
859 if (rc < 0) {
860 mapping_set_error(mapping, rc);
861 ceph_set_error_write(ci);
862 if (rc == -EBLOCKLISTED)
863 fsc->blocklisted = true;
864 } else {
865 ceph_clear_error_write(ci);
866 }
867
868 /*
869 * We lost the cache cap, need to truncate the page before
870 * it is unlocked, otherwise we'd truncate it later in the
871 * page truncation thread, possibly losing some data that
872 * raced its way in
873 */
874 remove_page = !(ceph_caps_issued(ci) &
875 (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO));
876
877 /* clean all pages */
878 for (i = 0; i < req->r_num_ops; i++) {
879 if (req->r_ops[i].op != CEPH_OSD_OP_WRITE) {
880 pr_warn_client(cl,
881 "%llx.%llx incorrect op %d req %p index %d tid %llu\n",
882 ceph_vinop(inode), req->r_ops[i].op, req, i,
883 req->r_tid);
884 break;
885 }
886
887 osd_data = osd_req_op_extent_osd_data(req, i);
888 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
889 len += osd_data->length;
890 num_pages = calc_pages_for((u64)osd_data->alignment,
891 (u64)osd_data->length);
892 total_pages += num_pages;
893 for (j = 0; j < num_pages; j++) {
894 page = osd_data->pages[j];
895 if (fscrypt_is_bounce_page(page)) {
896 page = fscrypt_pagecache_page(page);
897 fscrypt_free_bounce_page(osd_data->pages[j]);
898 osd_data->pages[j] = page;
899 }
900 BUG_ON(!page);
901 WARN_ON(!PageUptodate(page));
902
903 if (atomic_long_dec_return(&fsc->writeback_count) <
904 CONGESTION_OFF_THRESH(
905 fsc->mount_options->congestion_kb))
906 fsc->write_congested = false;
907
908 ceph_put_snap_context(detach_page_private(page));
909 end_page_writeback(page);
910 doutc(cl, "unlocking %p\n", page);
911
912 if (remove_page)
913 generic_error_remove_folio(inode->i_mapping,
914 page_folio(page));
915
916 unlock_page(page);
917 }
918 doutc(cl, "%llx.%llx wrote %llu bytes cleaned %d pages\n",
919 ceph_vinop(inode), osd_data->length,
920 rc >= 0 ? num_pages : 0);
921
922 release_pages(osd_data->pages, num_pages);
923 }
924
925 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
926 req->r_end_latency, len, rc);
927
928 ceph_put_wrbuffer_cap_refs(ci, total_pages, snapc);
929
930 osd_data = osd_req_op_extent_osd_data(req, 0);
931 if (osd_data->pages_from_pool)
932 mempool_free(osd_data->pages, ceph_wb_pagevec_pool);
933 else
934 kfree(osd_data->pages);
935 ceph_osdc_put_request(req);
936 ceph_dec_osd_stopping_blocker(fsc->mdsc);
937 }
938
939 /*
940 * initiate async writeback
941 */
ceph_writepages_start(struct address_space * mapping,struct writeback_control * wbc)942 static int ceph_writepages_start(struct address_space *mapping,
943 struct writeback_control *wbc)
944 {
945 struct inode *inode = mapping->host;
946 struct ceph_inode_info *ci = ceph_inode(inode);
947 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
948 struct ceph_client *cl = fsc->client;
949 struct ceph_vino vino = ceph_vino(inode);
950 pgoff_t index, start_index, end = -1;
951 struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
952 struct folio_batch fbatch;
953 int rc = 0;
954 unsigned int wsize = i_blocksize(inode);
955 struct ceph_osd_request *req = NULL;
956 struct ceph_writeback_ctl ceph_wbc;
957 bool should_loop, range_whole = false;
958 bool done = false;
959 bool caching = ceph_is_cache_enabled(inode);
960 xa_mark_t tag;
961
962 if (wbc->sync_mode == WB_SYNC_NONE &&
963 fsc->write_congested)
964 return 0;
965
966 doutc(cl, "%llx.%llx (mode=%s)\n", ceph_vinop(inode),
967 wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
968 (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
969
970 if (ceph_inode_is_shutdown(inode)) {
971 if (ci->i_wrbuffer_ref > 0) {
972 pr_warn_ratelimited_client(cl,
973 "%llx.%llx %lld forced umount\n",
974 ceph_vinop(inode), ceph_ino(inode));
975 }
976 mapping_set_error(mapping, -EIO);
977 return -EIO; /* we're in a forced umount, don't write! */
978 }
979 if (fsc->mount_options->wsize < wsize)
980 wsize = fsc->mount_options->wsize;
981
982 folio_batch_init(&fbatch);
983
984 start_index = wbc->range_cyclic ? mapping->writeback_index : 0;
985 index = start_index;
986
987 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
988 tag = PAGECACHE_TAG_TOWRITE;
989 } else {
990 tag = PAGECACHE_TAG_DIRTY;
991 }
992 retry:
993 /* find oldest snap context with dirty data */
994 snapc = get_oldest_context(inode, &ceph_wbc, NULL);
995 if (!snapc) {
996 /* hmm, why does writepages get called when there
997 is no dirty data? */
998 doutc(cl, " no snap context with dirty data?\n");
999 goto out;
1000 }
1001 doutc(cl, " oldest snapc is %p seq %lld (%d snaps)\n", snapc,
1002 snapc->seq, snapc->num_snaps);
1003
1004 should_loop = false;
1005 if (ceph_wbc.head_snapc && snapc != last_snapc) {
1006 /* where to start/end? */
1007 if (wbc->range_cyclic) {
1008 index = start_index;
1009 end = -1;
1010 if (index > 0)
1011 should_loop = true;
1012 doutc(cl, " cyclic, start at %lu\n", index);
1013 } else {
1014 index = wbc->range_start >> PAGE_SHIFT;
1015 end = wbc->range_end >> PAGE_SHIFT;
1016 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1017 range_whole = true;
1018 doutc(cl, " not cyclic, %lu to %lu\n", index, end);
1019 }
1020 } else if (!ceph_wbc.head_snapc) {
1021 /* Do not respect wbc->range_{start,end}. Dirty pages
1022 * in that range can be associated with newer snapc.
1023 * They are not writeable until we write all dirty pages
1024 * associated with 'snapc' get written */
1025 if (index > 0)
1026 should_loop = true;
1027 doutc(cl, " non-head snapc, range whole\n");
1028 }
1029
1030 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1031 tag_pages_for_writeback(mapping, index, end);
1032
1033 ceph_put_snap_context(last_snapc);
1034 last_snapc = snapc;
1035
1036 while (!done && index <= end) {
1037 int num_ops = 0, op_idx;
1038 unsigned i, nr_folios, max_pages, locked_pages = 0;
1039 struct page **pages = NULL, **data_pages;
1040 struct page *page;
1041 pgoff_t strip_unit_end = 0;
1042 u64 offset = 0, len = 0;
1043 bool from_pool = false;
1044
1045 max_pages = wsize >> PAGE_SHIFT;
1046
1047 get_more_pages:
1048 nr_folios = filemap_get_folios_tag(mapping, &index,
1049 end, tag, &fbatch);
1050 doutc(cl, "pagevec_lookup_range_tag got %d\n", nr_folios);
1051 if (!nr_folios && !locked_pages)
1052 break;
1053 for (i = 0; i < nr_folios && locked_pages < max_pages; i++) {
1054 page = &fbatch.folios[i]->page;
1055 doutc(cl, "? %p idx %lu\n", page, page->index);
1056 if (locked_pages == 0)
1057 lock_page(page); /* first page */
1058 else if (!trylock_page(page))
1059 break;
1060
1061 /* only dirty pages, or our accounting breaks */
1062 if (unlikely(!PageDirty(page)) ||
1063 unlikely(page->mapping != mapping)) {
1064 doutc(cl, "!dirty or !mapping %p\n", page);
1065 unlock_page(page);
1066 continue;
1067 }
1068 /* only if matching snap context */
1069 pgsnapc = page_snap_context(page);
1070 if (pgsnapc != snapc) {
1071 doutc(cl, "page snapc %p %lld != oldest %p %lld\n",
1072 pgsnapc, pgsnapc->seq, snapc, snapc->seq);
1073 if (!should_loop &&
1074 !ceph_wbc.head_snapc &&
1075 wbc->sync_mode != WB_SYNC_NONE)
1076 should_loop = true;
1077 unlock_page(page);
1078 continue;
1079 }
1080 if (page_offset(page) >= ceph_wbc.i_size) {
1081 struct folio *folio = page_folio(page);
1082
1083 doutc(cl, "folio at %lu beyond eof %llu\n",
1084 folio->index, ceph_wbc.i_size);
1085 if ((ceph_wbc.size_stable ||
1086 folio_pos(folio) >= i_size_read(inode)) &&
1087 folio_clear_dirty_for_io(folio))
1088 folio_invalidate(folio, 0,
1089 folio_size(folio));
1090 folio_unlock(folio);
1091 continue;
1092 }
1093 if (strip_unit_end && (page->index > strip_unit_end)) {
1094 doutc(cl, "end of strip unit %p\n", page);
1095 unlock_page(page);
1096 break;
1097 }
1098 if (PageWriteback(page) ||
1099 PagePrivate2(page) /* [DEPRECATED] */) {
1100 if (wbc->sync_mode == WB_SYNC_NONE) {
1101 doutc(cl, "%p under writeback\n", page);
1102 unlock_page(page);
1103 continue;
1104 }
1105 doutc(cl, "waiting on writeback %p\n", page);
1106 wait_on_page_writeback(page);
1107 folio_wait_private_2(page_folio(page)); /* [DEPRECATED] */
1108 }
1109
1110 if (!clear_page_dirty_for_io(page)) {
1111 doutc(cl, "%p !clear_page_dirty_for_io\n", page);
1112 unlock_page(page);
1113 continue;
1114 }
1115
1116 /*
1117 * We have something to write. If this is
1118 * the first locked page this time through,
1119 * calculate max possinle write size and
1120 * allocate a page array
1121 */
1122 if (locked_pages == 0) {
1123 u64 objnum;
1124 u64 objoff;
1125 u32 xlen;
1126
1127 /* prepare async write request */
1128 offset = (u64)page_offset(page);
1129 ceph_calc_file_object_mapping(&ci->i_layout,
1130 offset, wsize,
1131 &objnum, &objoff,
1132 &xlen);
1133 len = xlen;
1134
1135 num_ops = 1;
1136 strip_unit_end = page->index +
1137 ((len - 1) >> PAGE_SHIFT);
1138
1139 BUG_ON(pages);
1140 max_pages = calc_pages_for(0, (u64)len);
1141 pages = kmalloc_array(max_pages,
1142 sizeof(*pages),
1143 GFP_NOFS);
1144 if (!pages) {
1145 from_pool = true;
1146 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1147 BUG_ON(!pages);
1148 }
1149
1150 len = 0;
1151 } else if (page->index !=
1152 (offset + len) >> PAGE_SHIFT) {
1153 if (num_ops >= (from_pool ? CEPH_OSD_SLAB_OPS :
1154 CEPH_OSD_MAX_OPS)) {
1155 redirty_page_for_writepage(wbc, page);
1156 unlock_page(page);
1157 break;
1158 }
1159
1160 num_ops++;
1161 offset = (u64)page_offset(page);
1162 len = 0;
1163 }
1164
1165 /* note position of first page in fbatch */
1166 doutc(cl, "%llx.%llx will write page %p idx %lu\n",
1167 ceph_vinop(inode), page, page->index);
1168
1169 if (atomic_long_inc_return(&fsc->writeback_count) >
1170 CONGESTION_ON_THRESH(
1171 fsc->mount_options->congestion_kb))
1172 fsc->write_congested = true;
1173
1174 if (IS_ENCRYPTED(inode)) {
1175 pages[locked_pages] =
1176 fscrypt_encrypt_pagecache_blocks(page,
1177 PAGE_SIZE, 0,
1178 locked_pages ? GFP_NOWAIT : GFP_NOFS);
1179 if (IS_ERR(pages[locked_pages])) {
1180 if (PTR_ERR(pages[locked_pages]) == -EINVAL)
1181 pr_err_client(cl,
1182 "inode->i_blkbits=%hhu\n",
1183 inode->i_blkbits);
1184 /* better not fail on first page! */
1185 BUG_ON(locked_pages == 0);
1186 pages[locked_pages] = NULL;
1187 redirty_page_for_writepage(wbc, page);
1188 unlock_page(page);
1189 break;
1190 }
1191 ++locked_pages;
1192 } else {
1193 pages[locked_pages++] = page;
1194 }
1195
1196 fbatch.folios[i] = NULL;
1197 len += thp_size(page);
1198 }
1199
1200 /* did we get anything? */
1201 if (!locked_pages)
1202 goto release_folios;
1203 if (i) {
1204 unsigned j, n = 0;
1205 /* shift unused page to beginning of fbatch */
1206 for (j = 0; j < nr_folios; j++) {
1207 if (!fbatch.folios[j])
1208 continue;
1209 if (n < j)
1210 fbatch.folios[n] = fbatch.folios[j];
1211 n++;
1212 }
1213 fbatch.nr = n;
1214
1215 if (nr_folios && i == nr_folios &&
1216 locked_pages < max_pages) {
1217 doutc(cl, "reached end fbatch, trying for more\n");
1218 folio_batch_release(&fbatch);
1219 goto get_more_pages;
1220 }
1221 }
1222
1223 new_request:
1224 offset = ceph_fscrypt_page_offset(pages[0]);
1225 len = wsize;
1226
1227 req = ceph_osdc_new_request(&fsc->client->osdc,
1228 &ci->i_layout, vino,
1229 offset, &len, 0, num_ops,
1230 CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1231 snapc, ceph_wbc.truncate_seq,
1232 ceph_wbc.truncate_size, false);
1233 if (IS_ERR(req)) {
1234 req = ceph_osdc_new_request(&fsc->client->osdc,
1235 &ci->i_layout, vino,
1236 offset, &len, 0,
1237 min(num_ops,
1238 CEPH_OSD_SLAB_OPS),
1239 CEPH_OSD_OP_WRITE,
1240 CEPH_OSD_FLAG_WRITE,
1241 snapc, ceph_wbc.truncate_seq,
1242 ceph_wbc.truncate_size, true);
1243 BUG_ON(IS_ERR(req));
1244 }
1245 BUG_ON(len < ceph_fscrypt_page_offset(pages[locked_pages - 1]) +
1246 thp_size(pages[locked_pages - 1]) - offset);
1247
1248 if (!ceph_inc_osd_stopping_blocker(fsc->mdsc)) {
1249 rc = -EIO;
1250 goto release_folios;
1251 }
1252 req->r_callback = writepages_finish;
1253 req->r_inode = inode;
1254
1255 /* Format the osd request message and submit the write */
1256 len = 0;
1257 data_pages = pages;
1258 op_idx = 0;
1259 for (i = 0; i < locked_pages; i++) {
1260 struct page *page = ceph_fscrypt_pagecache_page(pages[i]);
1261
1262 u64 cur_offset = page_offset(page);
1263 /*
1264 * Discontinuity in page range? Ceph can handle that by just passing
1265 * multiple extents in the write op.
1266 */
1267 if (offset + len != cur_offset) {
1268 /* If it's full, stop here */
1269 if (op_idx + 1 == req->r_num_ops)
1270 break;
1271
1272 /* Kick off an fscache write with what we have so far. */
1273 ceph_fscache_write_to_cache(inode, offset, len, caching);
1274
1275 /* Start a new extent */
1276 osd_req_op_extent_dup_last(req, op_idx,
1277 cur_offset - offset);
1278 doutc(cl, "got pages at %llu~%llu\n", offset,
1279 len);
1280 osd_req_op_extent_osd_data_pages(req, op_idx,
1281 data_pages, len, 0,
1282 from_pool, false);
1283 osd_req_op_extent_update(req, op_idx, len);
1284
1285 len = 0;
1286 offset = cur_offset;
1287 data_pages = pages + i;
1288 op_idx++;
1289 }
1290
1291 set_page_writeback(page);
1292 if (caching)
1293 ceph_set_page_fscache(page);
1294 len += thp_size(page);
1295 }
1296 ceph_fscache_write_to_cache(inode, offset, len, caching);
1297
1298 if (ceph_wbc.size_stable) {
1299 len = min(len, ceph_wbc.i_size - offset);
1300 } else if (i == locked_pages) {
1301 /* writepages_finish() clears writeback pages
1302 * according to the data length, so make sure
1303 * data length covers all locked pages */
1304 u64 min_len = len + 1 - thp_size(page);
1305 len = get_writepages_data_length(inode, pages[i - 1],
1306 offset);
1307 len = max(len, min_len);
1308 }
1309 if (IS_ENCRYPTED(inode))
1310 len = round_up(len, CEPH_FSCRYPT_BLOCK_SIZE);
1311
1312 doutc(cl, "got pages at %llu~%llu\n", offset, len);
1313
1314 if (IS_ENCRYPTED(inode) &&
1315 ((offset | len) & ~CEPH_FSCRYPT_BLOCK_MASK))
1316 pr_warn_client(cl,
1317 "bad encrypted write offset=%lld len=%llu\n",
1318 offset, len);
1319
1320 osd_req_op_extent_osd_data_pages(req, op_idx, data_pages, len,
1321 0, from_pool, false);
1322 osd_req_op_extent_update(req, op_idx, len);
1323
1324 BUG_ON(op_idx + 1 != req->r_num_ops);
1325
1326 from_pool = false;
1327 if (i < locked_pages) {
1328 BUG_ON(num_ops <= req->r_num_ops);
1329 num_ops -= req->r_num_ops;
1330 locked_pages -= i;
1331
1332 /* allocate new pages array for next request */
1333 data_pages = pages;
1334 pages = kmalloc_array(locked_pages, sizeof(*pages),
1335 GFP_NOFS);
1336 if (!pages) {
1337 from_pool = true;
1338 pages = mempool_alloc(ceph_wb_pagevec_pool, GFP_NOFS);
1339 BUG_ON(!pages);
1340 }
1341 memcpy(pages, data_pages + i,
1342 locked_pages * sizeof(*pages));
1343 memset(data_pages + i, 0,
1344 locked_pages * sizeof(*pages));
1345 } else {
1346 BUG_ON(num_ops != req->r_num_ops);
1347 index = pages[i - 1]->index + 1;
1348 /* request message now owns the pages array */
1349 pages = NULL;
1350 }
1351
1352 req->r_mtime = inode_get_mtime(inode);
1353 ceph_osdc_start_request(&fsc->client->osdc, req);
1354 req = NULL;
1355
1356 wbc->nr_to_write -= i;
1357 if (pages)
1358 goto new_request;
1359
1360 /*
1361 * We stop writing back only if we are not doing
1362 * integrity sync. In case of integrity sync we have to
1363 * keep going until we have written all the pages
1364 * we tagged for writeback prior to entering this loop.
1365 */
1366 if (wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE)
1367 done = true;
1368
1369 release_folios:
1370 doutc(cl, "folio_batch release on %d folios (%p)\n",
1371 (int)fbatch.nr, fbatch.nr ? fbatch.folios[0] : NULL);
1372 folio_batch_release(&fbatch);
1373 }
1374
1375 if (should_loop && !done) {
1376 /* more to do; loop back to beginning of file */
1377 doutc(cl, "looping back to beginning of file\n");
1378 end = start_index - 1; /* OK even when start_index == 0 */
1379
1380 /* to write dirty pages associated with next snapc,
1381 * we need to wait until current writes complete */
1382 if (wbc->sync_mode != WB_SYNC_NONE &&
1383 start_index == 0 && /* all dirty pages were checked */
1384 !ceph_wbc.head_snapc) {
1385 struct page *page;
1386 unsigned i, nr;
1387 index = 0;
1388 while ((index <= end) &&
1389 (nr = filemap_get_folios_tag(mapping, &index,
1390 (pgoff_t)-1,
1391 PAGECACHE_TAG_WRITEBACK,
1392 &fbatch))) {
1393 for (i = 0; i < nr; i++) {
1394 page = &fbatch.folios[i]->page;
1395 if (page_snap_context(page) != snapc)
1396 continue;
1397 wait_on_page_writeback(page);
1398 }
1399 folio_batch_release(&fbatch);
1400 cond_resched();
1401 }
1402 }
1403
1404 start_index = 0;
1405 index = 0;
1406 goto retry;
1407 }
1408
1409 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1410 mapping->writeback_index = index;
1411
1412 out:
1413 ceph_osdc_put_request(req);
1414 ceph_put_snap_context(last_snapc);
1415 doutc(cl, "%llx.%llx dend - startone, rc = %d\n", ceph_vinop(inode),
1416 rc);
1417 return rc;
1418 }
1419
1420
1421
1422 /*
1423 * See if a given @snapc is either writeable, or already written.
1424 */
context_is_writeable_or_written(struct inode * inode,struct ceph_snap_context * snapc)1425 static int context_is_writeable_or_written(struct inode *inode,
1426 struct ceph_snap_context *snapc)
1427 {
1428 struct ceph_snap_context *oldest = get_oldest_context(inode, NULL, NULL);
1429 int ret = !oldest || snapc->seq <= oldest->seq;
1430
1431 ceph_put_snap_context(oldest);
1432 return ret;
1433 }
1434
1435 /**
1436 * ceph_find_incompatible - find an incompatible context and return it
1437 * @page: page being dirtied
1438 *
1439 * We are only allowed to write into/dirty a page if the page is
1440 * clean, or already dirty within the same snap context. Returns a
1441 * conflicting context if there is one, NULL if there isn't, or a
1442 * negative error code on other errors.
1443 *
1444 * Must be called with page lock held.
1445 */
1446 static struct ceph_snap_context *
ceph_find_incompatible(struct page * page)1447 ceph_find_incompatible(struct page *page)
1448 {
1449 struct inode *inode = page->mapping->host;
1450 struct ceph_client *cl = ceph_inode_to_client(inode);
1451 struct ceph_inode_info *ci = ceph_inode(inode);
1452
1453 if (ceph_inode_is_shutdown(inode)) {
1454 doutc(cl, " %llx.%llx page %p is shutdown\n",
1455 ceph_vinop(inode), page);
1456 return ERR_PTR(-ESTALE);
1457 }
1458
1459 for (;;) {
1460 struct ceph_snap_context *snapc, *oldest;
1461
1462 wait_on_page_writeback(page);
1463
1464 snapc = page_snap_context(page);
1465 if (!snapc || snapc == ci->i_head_snapc)
1466 break;
1467
1468 /*
1469 * this page is already dirty in another (older) snap
1470 * context! is it writeable now?
1471 */
1472 oldest = get_oldest_context(inode, NULL, NULL);
1473 if (snapc->seq > oldest->seq) {
1474 /* not writeable -- return it for the caller to deal with */
1475 ceph_put_snap_context(oldest);
1476 doutc(cl, " %llx.%llx page %p snapc %p not current or oldest\n",
1477 ceph_vinop(inode), page, snapc);
1478 return ceph_get_snap_context(snapc);
1479 }
1480 ceph_put_snap_context(oldest);
1481
1482 /* yay, writeable, do it now (without dropping page lock) */
1483 doutc(cl, " %llx.%llx page %p snapc %p not current, but oldest\n",
1484 ceph_vinop(inode), page, snapc);
1485 if (clear_page_dirty_for_io(page)) {
1486 int r = writepage_nounlock(page, NULL);
1487 if (r < 0)
1488 return ERR_PTR(r);
1489 }
1490 }
1491 return NULL;
1492 }
1493
ceph_netfs_check_write_begin(struct file * file,loff_t pos,unsigned int len,struct folio ** foliop,void ** _fsdata)1494 static int ceph_netfs_check_write_begin(struct file *file, loff_t pos, unsigned int len,
1495 struct folio **foliop, void **_fsdata)
1496 {
1497 struct inode *inode = file_inode(file);
1498 struct ceph_inode_info *ci = ceph_inode(inode);
1499 struct ceph_snap_context *snapc;
1500
1501 snapc = ceph_find_incompatible(folio_page(*foliop, 0));
1502 if (snapc) {
1503 int r;
1504
1505 folio_unlock(*foliop);
1506 folio_put(*foliop);
1507 *foliop = NULL;
1508 if (IS_ERR(snapc))
1509 return PTR_ERR(snapc);
1510
1511 ceph_queue_writeback(inode);
1512 r = wait_event_killable(ci->i_cap_wq,
1513 context_is_writeable_or_written(inode, snapc));
1514 ceph_put_snap_context(snapc);
1515 return r == 0 ? -EAGAIN : r;
1516 }
1517 return 0;
1518 }
1519
1520 /*
1521 * We are only allowed to write into/dirty the page if the page is
1522 * clean, or already dirty within the same snap context.
1523 */
ceph_write_begin(struct file * file,struct address_space * mapping,loff_t pos,unsigned len,struct folio ** foliop,void ** fsdata)1524 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1525 loff_t pos, unsigned len,
1526 struct folio **foliop, void **fsdata)
1527 {
1528 struct inode *inode = file_inode(file);
1529 struct ceph_inode_info *ci = ceph_inode(inode);
1530 int r;
1531
1532 r = netfs_write_begin(&ci->netfs, file, inode->i_mapping, pos, len, foliop, NULL);
1533 if (r < 0)
1534 return r;
1535
1536 folio_wait_private_2(*foliop); /* [DEPRECATED] */
1537 WARN_ON_ONCE(!folio_test_locked(*foliop));
1538 return 0;
1539 }
1540
1541 /*
1542 * we don't do anything in here that simple_write_end doesn't do
1543 * except adjust dirty page accounting
1544 */
ceph_write_end(struct file * file,struct address_space * mapping,loff_t pos,unsigned len,unsigned copied,struct folio * folio,void * fsdata)1545 static int ceph_write_end(struct file *file, struct address_space *mapping,
1546 loff_t pos, unsigned len, unsigned copied,
1547 struct folio *folio, void *fsdata)
1548 {
1549 struct inode *inode = file_inode(file);
1550 struct ceph_client *cl = ceph_inode_to_client(inode);
1551 bool check_cap = false;
1552
1553 doutc(cl, "%llx.%llx file %p folio %p %d~%d (%d)\n", ceph_vinop(inode),
1554 file, folio, (int)pos, (int)copied, (int)len);
1555
1556 if (!folio_test_uptodate(folio)) {
1557 /* just return that nothing was copied on a short copy */
1558 if (copied < len) {
1559 copied = 0;
1560 goto out;
1561 }
1562 folio_mark_uptodate(folio);
1563 }
1564
1565 /* did file size increase? */
1566 if (pos+copied > i_size_read(inode))
1567 check_cap = ceph_inode_set_size(inode, pos+copied);
1568
1569 folio_mark_dirty(folio);
1570
1571 out:
1572 folio_unlock(folio);
1573 folio_put(folio);
1574
1575 if (check_cap)
1576 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY);
1577
1578 return copied;
1579 }
1580
1581 const struct address_space_operations ceph_aops = {
1582 .read_folio = netfs_read_folio,
1583 .readahead = netfs_readahead,
1584 .writepage = ceph_writepage,
1585 .writepages = ceph_writepages_start,
1586 .write_begin = ceph_write_begin,
1587 .write_end = ceph_write_end,
1588 .dirty_folio = ceph_dirty_folio,
1589 .invalidate_folio = ceph_invalidate_folio,
1590 .release_folio = netfs_release_folio,
1591 .direct_IO = noop_direct_IO,
1592 };
1593
ceph_block_sigs(sigset_t * oldset)1594 static void ceph_block_sigs(sigset_t *oldset)
1595 {
1596 sigset_t mask;
1597 siginitsetinv(&mask, sigmask(SIGKILL));
1598 sigprocmask(SIG_BLOCK, &mask, oldset);
1599 }
1600
ceph_restore_sigs(sigset_t * oldset)1601 static void ceph_restore_sigs(sigset_t *oldset)
1602 {
1603 sigprocmask(SIG_SETMASK, oldset, NULL);
1604 }
1605
1606 /*
1607 * vm ops
1608 */
ceph_filemap_fault(struct vm_fault * vmf)1609 static vm_fault_t ceph_filemap_fault(struct vm_fault *vmf)
1610 {
1611 struct vm_area_struct *vma = vmf->vma;
1612 struct inode *inode = file_inode(vma->vm_file);
1613 struct ceph_inode_info *ci = ceph_inode(inode);
1614 struct ceph_client *cl = ceph_inode_to_client(inode);
1615 struct ceph_file_info *fi = vma->vm_file->private_data;
1616 loff_t off = (loff_t)vmf->pgoff << PAGE_SHIFT;
1617 int want, got, err;
1618 sigset_t oldset;
1619 vm_fault_t ret = VM_FAULT_SIGBUS;
1620
1621 if (ceph_inode_is_shutdown(inode))
1622 return ret;
1623
1624 ceph_block_sigs(&oldset);
1625
1626 doutc(cl, "%llx.%llx %llu trying to get caps\n",
1627 ceph_vinop(inode), off);
1628 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1629 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1630 else
1631 want = CEPH_CAP_FILE_CACHE;
1632
1633 got = 0;
1634 err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_RD, want, -1, &got);
1635 if (err < 0)
1636 goto out_restore;
1637
1638 doutc(cl, "%llx.%llx %llu got cap refs on %s\n", ceph_vinop(inode),
1639 off, ceph_cap_string(got));
1640
1641 if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) ||
1642 !ceph_has_inline_data(ci)) {
1643 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1644 ceph_add_rw_context(fi, &rw_ctx);
1645 ret = filemap_fault(vmf);
1646 ceph_del_rw_context(fi, &rw_ctx);
1647 doutc(cl, "%llx.%llx %llu drop cap refs %s ret %x\n",
1648 ceph_vinop(inode), off, ceph_cap_string(got), ret);
1649 } else
1650 err = -EAGAIN;
1651
1652 ceph_put_cap_refs(ci, got);
1653
1654 if (err != -EAGAIN)
1655 goto out_restore;
1656
1657 /* read inline data */
1658 if (off >= PAGE_SIZE) {
1659 /* does not support inline data > PAGE_SIZE */
1660 ret = VM_FAULT_SIGBUS;
1661 } else {
1662 struct address_space *mapping = inode->i_mapping;
1663 struct page *page;
1664
1665 filemap_invalidate_lock_shared(mapping);
1666 page = find_or_create_page(mapping, 0,
1667 mapping_gfp_constraint(mapping, ~__GFP_FS));
1668 if (!page) {
1669 ret = VM_FAULT_OOM;
1670 goto out_inline;
1671 }
1672 err = __ceph_do_getattr(inode, page,
1673 CEPH_STAT_CAP_INLINE_DATA, true);
1674 if (err < 0 || off >= i_size_read(inode)) {
1675 unlock_page(page);
1676 put_page(page);
1677 ret = vmf_error(err);
1678 goto out_inline;
1679 }
1680 if (err < PAGE_SIZE)
1681 zero_user_segment(page, err, PAGE_SIZE);
1682 else
1683 flush_dcache_page(page);
1684 SetPageUptodate(page);
1685 vmf->page = page;
1686 ret = VM_FAULT_MAJOR | VM_FAULT_LOCKED;
1687 out_inline:
1688 filemap_invalidate_unlock_shared(mapping);
1689 doutc(cl, "%llx.%llx %llu read inline data ret %x\n",
1690 ceph_vinop(inode), off, ret);
1691 }
1692 out_restore:
1693 ceph_restore_sigs(&oldset);
1694 if (err < 0)
1695 ret = vmf_error(err);
1696
1697 return ret;
1698 }
1699
ceph_page_mkwrite(struct vm_fault * vmf)1700 static vm_fault_t ceph_page_mkwrite(struct vm_fault *vmf)
1701 {
1702 struct vm_area_struct *vma = vmf->vma;
1703 struct inode *inode = file_inode(vma->vm_file);
1704 struct ceph_client *cl = ceph_inode_to_client(inode);
1705 struct ceph_inode_info *ci = ceph_inode(inode);
1706 struct ceph_file_info *fi = vma->vm_file->private_data;
1707 struct ceph_cap_flush *prealloc_cf;
1708 struct page *page = vmf->page;
1709 loff_t off = page_offset(page);
1710 loff_t size = i_size_read(inode);
1711 size_t len;
1712 int want, got, err;
1713 sigset_t oldset;
1714 vm_fault_t ret = VM_FAULT_SIGBUS;
1715
1716 if (ceph_inode_is_shutdown(inode))
1717 return ret;
1718
1719 prealloc_cf = ceph_alloc_cap_flush();
1720 if (!prealloc_cf)
1721 return VM_FAULT_OOM;
1722
1723 sb_start_pagefault(inode->i_sb);
1724 ceph_block_sigs(&oldset);
1725
1726 if (off + thp_size(page) <= size)
1727 len = thp_size(page);
1728 else
1729 len = offset_in_thp(page, size);
1730
1731 doutc(cl, "%llx.%llx %llu~%zd getting caps i_size %llu\n",
1732 ceph_vinop(inode), off, len, size);
1733 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1734 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1735 else
1736 want = CEPH_CAP_FILE_BUFFER;
1737
1738 got = 0;
1739 err = ceph_get_caps(vma->vm_file, CEPH_CAP_FILE_WR, want, off + len, &got);
1740 if (err < 0)
1741 goto out_free;
1742
1743 doutc(cl, "%llx.%llx %llu~%zd got cap refs on %s\n", ceph_vinop(inode),
1744 off, len, ceph_cap_string(got));
1745
1746 /* Update time before taking page lock */
1747 file_update_time(vma->vm_file);
1748 inode_inc_iversion_raw(inode);
1749
1750 do {
1751 struct ceph_snap_context *snapc;
1752
1753 lock_page(page);
1754
1755 if (page_mkwrite_check_truncate(page, inode) < 0) {
1756 unlock_page(page);
1757 ret = VM_FAULT_NOPAGE;
1758 break;
1759 }
1760
1761 snapc = ceph_find_incompatible(page);
1762 if (!snapc) {
1763 /* success. we'll keep the page locked. */
1764 set_page_dirty(page);
1765 ret = VM_FAULT_LOCKED;
1766 break;
1767 }
1768
1769 unlock_page(page);
1770
1771 if (IS_ERR(snapc)) {
1772 ret = VM_FAULT_SIGBUS;
1773 break;
1774 }
1775
1776 ceph_queue_writeback(inode);
1777 err = wait_event_killable(ci->i_cap_wq,
1778 context_is_writeable_or_written(inode, snapc));
1779 ceph_put_snap_context(snapc);
1780 } while (err == 0);
1781
1782 if (ret == VM_FAULT_LOCKED) {
1783 int dirty;
1784 spin_lock(&ci->i_ceph_lock);
1785 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1786 &prealloc_cf);
1787 spin_unlock(&ci->i_ceph_lock);
1788 if (dirty)
1789 __mark_inode_dirty(inode, dirty);
1790 }
1791
1792 doutc(cl, "%llx.%llx %llu~%zd dropping cap refs on %s ret %x\n",
1793 ceph_vinop(inode), off, len, ceph_cap_string(got), ret);
1794 ceph_put_cap_refs_async(ci, got);
1795 out_free:
1796 ceph_restore_sigs(&oldset);
1797 sb_end_pagefault(inode->i_sb);
1798 ceph_free_cap_flush(prealloc_cf);
1799 if (err < 0)
1800 ret = vmf_error(err);
1801 return ret;
1802 }
1803
ceph_fill_inline_data(struct inode * inode,struct page * locked_page,char * data,size_t len)1804 void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
1805 char *data, size_t len)
1806 {
1807 struct ceph_client *cl = ceph_inode_to_client(inode);
1808 struct address_space *mapping = inode->i_mapping;
1809 struct page *page;
1810
1811 if (locked_page) {
1812 page = locked_page;
1813 } else {
1814 if (i_size_read(inode) == 0)
1815 return;
1816 page = find_or_create_page(mapping, 0,
1817 mapping_gfp_constraint(mapping,
1818 ~__GFP_FS));
1819 if (!page)
1820 return;
1821 if (PageUptodate(page)) {
1822 unlock_page(page);
1823 put_page(page);
1824 return;
1825 }
1826 }
1827
1828 doutc(cl, "%p %llx.%llx len %zu locked_page %p\n", inode,
1829 ceph_vinop(inode), len, locked_page);
1830
1831 if (len > 0) {
1832 void *kaddr = kmap_atomic(page);
1833 memcpy(kaddr, data, len);
1834 kunmap_atomic(kaddr);
1835 }
1836
1837 if (page != locked_page) {
1838 if (len < PAGE_SIZE)
1839 zero_user_segment(page, len, PAGE_SIZE);
1840 else
1841 flush_dcache_page(page);
1842
1843 SetPageUptodate(page);
1844 unlock_page(page);
1845 put_page(page);
1846 }
1847 }
1848
ceph_uninline_data(struct file * file)1849 int ceph_uninline_data(struct file *file)
1850 {
1851 struct inode *inode = file_inode(file);
1852 struct ceph_inode_info *ci = ceph_inode(inode);
1853 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
1854 struct ceph_client *cl = fsc->client;
1855 struct ceph_osd_request *req = NULL;
1856 struct ceph_cap_flush *prealloc_cf = NULL;
1857 struct folio *folio = NULL;
1858 u64 inline_version = CEPH_INLINE_NONE;
1859 struct page *pages[1];
1860 int err = 0;
1861 u64 len;
1862
1863 spin_lock(&ci->i_ceph_lock);
1864 inline_version = ci->i_inline_version;
1865 spin_unlock(&ci->i_ceph_lock);
1866
1867 doutc(cl, "%llx.%llx inline_version %llu\n", ceph_vinop(inode),
1868 inline_version);
1869
1870 if (ceph_inode_is_shutdown(inode)) {
1871 err = -EIO;
1872 goto out;
1873 }
1874
1875 if (inline_version == CEPH_INLINE_NONE)
1876 return 0;
1877
1878 prealloc_cf = ceph_alloc_cap_flush();
1879 if (!prealloc_cf)
1880 return -ENOMEM;
1881
1882 if (inline_version == 1) /* initial version, no data */
1883 goto out_uninline;
1884
1885 folio = read_mapping_folio(inode->i_mapping, 0, file);
1886 if (IS_ERR(folio)) {
1887 err = PTR_ERR(folio);
1888 goto out;
1889 }
1890
1891 folio_lock(folio);
1892
1893 len = i_size_read(inode);
1894 if (len > folio_size(folio))
1895 len = folio_size(folio);
1896
1897 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1898 ceph_vino(inode), 0, &len, 0, 1,
1899 CEPH_OSD_OP_CREATE, CEPH_OSD_FLAG_WRITE,
1900 NULL, 0, 0, false);
1901 if (IS_ERR(req)) {
1902 err = PTR_ERR(req);
1903 goto out_unlock;
1904 }
1905
1906 req->r_mtime = inode_get_mtime(inode);
1907 ceph_osdc_start_request(&fsc->client->osdc, req);
1908 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1909 ceph_osdc_put_request(req);
1910 if (err < 0)
1911 goto out_unlock;
1912
1913 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1914 ceph_vino(inode), 0, &len, 1, 3,
1915 CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE,
1916 NULL, ci->i_truncate_seq,
1917 ci->i_truncate_size, false);
1918 if (IS_ERR(req)) {
1919 err = PTR_ERR(req);
1920 goto out_unlock;
1921 }
1922
1923 pages[0] = folio_page(folio, 0);
1924 osd_req_op_extent_osd_data_pages(req, 1, pages, len, 0, false, false);
1925
1926 {
1927 __le64 xattr_buf = cpu_to_le64(inline_version);
1928 err = osd_req_op_xattr_init(req, 0, CEPH_OSD_OP_CMPXATTR,
1929 "inline_version", &xattr_buf,
1930 sizeof(xattr_buf),
1931 CEPH_OSD_CMPXATTR_OP_GT,
1932 CEPH_OSD_CMPXATTR_MODE_U64);
1933 if (err)
1934 goto out_put_req;
1935 }
1936
1937 {
1938 char xattr_buf[32];
1939 int xattr_len = snprintf(xattr_buf, sizeof(xattr_buf),
1940 "%llu", inline_version);
1941 err = osd_req_op_xattr_init(req, 2, CEPH_OSD_OP_SETXATTR,
1942 "inline_version",
1943 xattr_buf, xattr_len, 0, 0);
1944 if (err)
1945 goto out_put_req;
1946 }
1947
1948 req->r_mtime = inode_get_mtime(inode);
1949 ceph_osdc_start_request(&fsc->client->osdc, req);
1950 err = ceph_osdc_wait_request(&fsc->client->osdc, req);
1951
1952 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1953 req->r_end_latency, len, err);
1954
1955 out_uninline:
1956 if (!err) {
1957 int dirty;
1958
1959 /* Set to CAP_INLINE_NONE and dirty the caps */
1960 down_read(&fsc->mdsc->snap_rwsem);
1961 spin_lock(&ci->i_ceph_lock);
1962 ci->i_inline_version = CEPH_INLINE_NONE;
1963 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR, &prealloc_cf);
1964 spin_unlock(&ci->i_ceph_lock);
1965 up_read(&fsc->mdsc->snap_rwsem);
1966 if (dirty)
1967 __mark_inode_dirty(inode, dirty);
1968 }
1969 out_put_req:
1970 ceph_osdc_put_request(req);
1971 if (err == -ECANCELED)
1972 err = 0;
1973 out_unlock:
1974 if (folio) {
1975 folio_unlock(folio);
1976 folio_put(folio);
1977 }
1978 out:
1979 ceph_free_cap_flush(prealloc_cf);
1980 doutc(cl, "%llx.%llx inline_version %llu = %d\n",
1981 ceph_vinop(inode), inline_version, err);
1982 return err;
1983 }
1984
1985 static const struct vm_operations_struct ceph_vmops = {
1986 .fault = ceph_filemap_fault,
1987 .page_mkwrite = ceph_page_mkwrite,
1988 };
1989
ceph_mmap(struct file * file,struct vm_area_struct * vma)1990 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1991 {
1992 struct address_space *mapping = file->f_mapping;
1993
1994 if (!mapping->a_ops->read_folio)
1995 return -ENOEXEC;
1996 vma->vm_ops = &ceph_vmops;
1997 return 0;
1998 }
1999
2000 enum {
2001 POOL_READ = 1,
2002 POOL_WRITE = 2,
2003 };
2004
__ceph_pool_perm_get(struct ceph_inode_info * ci,s64 pool,struct ceph_string * pool_ns)2005 static int __ceph_pool_perm_get(struct ceph_inode_info *ci,
2006 s64 pool, struct ceph_string *pool_ns)
2007 {
2008 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(&ci->netfs.inode);
2009 struct ceph_mds_client *mdsc = fsc->mdsc;
2010 struct ceph_client *cl = fsc->client;
2011 struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
2012 struct rb_node **p, *parent;
2013 struct ceph_pool_perm *perm;
2014 struct page **pages;
2015 size_t pool_ns_len;
2016 int err = 0, err2 = 0, have = 0;
2017
2018 down_read(&mdsc->pool_perm_rwsem);
2019 p = &mdsc->pool_perm_tree.rb_node;
2020 while (*p) {
2021 perm = rb_entry(*p, struct ceph_pool_perm, node);
2022 if (pool < perm->pool)
2023 p = &(*p)->rb_left;
2024 else if (pool > perm->pool)
2025 p = &(*p)->rb_right;
2026 else {
2027 int ret = ceph_compare_string(pool_ns,
2028 perm->pool_ns,
2029 perm->pool_ns_len);
2030 if (ret < 0)
2031 p = &(*p)->rb_left;
2032 else if (ret > 0)
2033 p = &(*p)->rb_right;
2034 else {
2035 have = perm->perm;
2036 break;
2037 }
2038 }
2039 }
2040 up_read(&mdsc->pool_perm_rwsem);
2041 if (*p)
2042 goto out;
2043
2044 if (pool_ns)
2045 doutc(cl, "pool %lld ns %.*s no perm cached\n", pool,
2046 (int)pool_ns->len, pool_ns->str);
2047 else
2048 doutc(cl, "pool %lld no perm cached\n", pool);
2049
2050 down_write(&mdsc->pool_perm_rwsem);
2051 p = &mdsc->pool_perm_tree.rb_node;
2052 parent = NULL;
2053 while (*p) {
2054 parent = *p;
2055 perm = rb_entry(parent, struct ceph_pool_perm, node);
2056 if (pool < perm->pool)
2057 p = &(*p)->rb_left;
2058 else if (pool > perm->pool)
2059 p = &(*p)->rb_right;
2060 else {
2061 int ret = ceph_compare_string(pool_ns,
2062 perm->pool_ns,
2063 perm->pool_ns_len);
2064 if (ret < 0)
2065 p = &(*p)->rb_left;
2066 else if (ret > 0)
2067 p = &(*p)->rb_right;
2068 else {
2069 have = perm->perm;
2070 break;
2071 }
2072 }
2073 }
2074 if (*p) {
2075 up_write(&mdsc->pool_perm_rwsem);
2076 goto out;
2077 }
2078
2079 rd_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2080 1, false, GFP_NOFS);
2081 if (!rd_req) {
2082 err = -ENOMEM;
2083 goto out_unlock;
2084 }
2085
2086 rd_req->r_flags = CEPH_OSD_FLAG_READ;
2087 osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
2088 rd_req->r_base_oloc.pool = pool;
2089 if (pool_ns)
2090 rd_req->r_base_oloc.pool_ns = ceph_get_string(pool_ns);
2091 ceph_oid_printf(&rd_req->r_base_oid, "%llx.00000000", ci->i_vino.ino);
2092
2093 err = ceph_osdc_alloc_messages(rd_req, GFP_NOFS);
2094 if (err)
2095 goto out_unlock;
2096
2097 wr_req = ceph_osdc_alloc_request(&fsc->client->osdc, NULL,
2098 1, false, GFP_NOFS);
2099 if (!wr_req) {
2100 err = -ENOMEM;
2101 goto out_unlock;
2102 }
2103
2104 wr_req->r_flags = CEPH_OSD_FLAG_WRITE;
2105 osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
2106 ceph_oloc_copy(&wr_req->r_base_oloc, &rd_req->r_base_oloc);
2107 ceph_oid_copy(&wr_req->r_base_oid, &rd_req->r_base_oid);
2108
2109 err = ceph_osdc_alloc_messages(wr_req, GFP_NOFS);
2110 if (err)
2111 goto out_unlock;
2112
2113 /* one page should be large enough for STAT data */
2114 pages = ceph_alloc_page_vector(1, GFP_KERNEL);
2115 if (IS_ERR(pages)) {
2116 err = PTR_ERR(pages);
2117 goto out_unlock;
2118 }
2119
2120 osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
2121 0, false, true);
2122 ceph_osdc_start_request(&fsc->client->osdc, rd_req);
2123
2124 wr_req->r_mtime = inode_get_mtime(&ci->netfs.inode);
2125 ceph_osdc_start_request(&fsc->client->osdc, wr_req);
2126
2127 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
2128 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
2129
2130 if (err >= 0 || err == -ENOENT)
2131 have |= POOL_READ;
2132 else if (err != -EPERM) {
2133 if (err == -EBLOCKLISTED)
2134 fsc->blocklisted = true;
2135 goto out_unlock;
2136 }
2137
2138 if (err2 == 0 || err2 == -EEXIST)
2139 have |= POOL_WRITE;
2140 else if (err2 != -EPERM) {
2141 if (err2 == -EBLOCKLISTED)
2142 fsc->blocklisted = true;
2143 err = err2;
2144 goto out_unlock;
2145 }
2146
2147 pool_ns_len = pool_ns ? pool_ns->len : 0;
2148 perm = kmalloc(sizeof(*perm) + pool_ns_len + 1, GFP_NOFS);
2149 if (!perm) {
2150 err = -ENOMEM;
2151 goto out_unlock;
2152 }
2153
2154 perm->pool = pool;
2155 perm->perm = have;
2156 perm->pool_ns_len = pool_ns_len;
2157 if (pool_ns_len > 0)
2158 memcpy(perm->pool_ns, pool_ns->str, pool_ns_len);
2159 perm->pool_ns[pool_ns_len] = 0;
2160
2161 rb_link_node(&perm->node, parent, p);
2162 rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
2163 err = 0;
2164 out_unlock:
2165 up_write(&mdsc->pool_perm_rwsem);
2166
2167 ceph_osdc_put_request(rd_req);
2168 ceph_osdc_put_request(wr_req);
2169 out:
2170 if (!err)
2171 err = have;
2172 if (pool_ns)
2173 doutc(cl, "pool %lld ns %.*s result = %d\n", pool,
2174 (int)pool_ns->len, pool_ns->str, err);
2175 else
2176 doutc(cl, "pool %lld result = %d\n", pool, err);
2177 return err;
2178 }
2179
ceph_pool_perm_check(struct inode * inode,int need)2180 int ceph_pool_perm_check(struct inode *inode, int need)
2181 {
2182 struct ceph_client *cl = ceph_inode_to_client(inode);
2183 struct ceph_inode_info *ci = ceph_inode(inode);
2184 struct ceph_string *pool_ns;
2185 s64 pool;
2186 int ret, flags;
2187
2188 /* Only need to do this for regular files */
2189 if (!S_ISREG(inode->i_mode))
2190 return 0;
2191
2192 if (ci->i_vino.snap != CEPH_NOSNAP) {
2193 /*
2194 * Pool permission check needs to write to the first object.
2195 * But for snapshot, head of the first object may have alread
2196 * been deleted. Skip check to avoid creating orphan object.
2197 */
2198 return 0;
2199 }
2200
2201 if (ceph_test_mount_opt(ceph_inode_to_fs_client(inode),
2202 NOPOOLPERM))
2203 return 0;
2204
2205 spin_lock(&ci->i_ceph_lock);
2206 flags = ci->i_ceph_flags;
2207 pool = ci->i_layout.pool_id;
2208 spin_unlock(&ci->i_ceph_lock);
2209 check:
2210 if (flags & CEPH_I_POOL_PERM) {
2211 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
2212 doutc(cl, "pool %lld no read perm\n", pool);
2213 return -EPERM;
2214 }
2215 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
2216 doutc(cl, "pool %lld no write perm\n", pool);
2217 return -EPERM;
2218 }
2219 return 0;
2220 }
2221
2222 pool_ns = ceph_try_get_string(ci->i_layout.pool_ns);
2223 ret = __ceph_pool_perm_get(ci, pool, pool_ns);
2224 ceph_put_string(pool_ns);
2225 if (ret < 0)
2226 return ret;
2227
2228 flags = CEPH_I_POOL_PERM;
2229 if (ret & POOL_READ)
2230 flags |= CEPH_I_POOL_RD;
2231 if (ret & POOL_WRITE)
2232 flags |= CEPH_I_POOL_WR;
2233
2234 spin_lock(&ci->i_ceph_lock);
2235 if (pool == ci->i_layout.pool_id &&
2236 pool_ns == rcu_dereference_raw(ci->i_layout.pool_ns)) {
2237 ci->i_ceph_flags |= flags;
2238 } else {
2239 pool = ci->i_layout.pool_id;
2240 flags = ci->i_ceph_flags;
2241 }
2242 spin_unlock(&ci->i_ceph_lock);
2243 goto check;
2244 }
2245
ceph_pool_perm_destroy(struct ceph_mds_client * mdsc)2246 void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
2247 {
2248 struct ceph_pool_perm *perm;
2249 struct rb_node *n;
2250
2251 while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
2252 n = rb_first(&mdsc->pool_perm_tree);
2253 perm = rb_entry(n, struct ceph_pool_perm, node);
2254 rb_erase(n, &mdsc->pool_perm_tree);
2255 kfree(perm);
2256 }
2257 }
2258