]> git.ipfire.org Git - thirdparty/linux.git/blob - fs/ceph/file.c
projects / thirdparty / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
fs: use splice_copy_file_range() inline helper
[thirdparty/linux.git] / fs / ceph / file.c
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
3 #include <linux/ceph/striper.h>
4
5 #include <linux/module.h>
6 #include <linux/sched.h>
7 #include <linux/slab.h>
8 #include <linux/file.h>
9 #include <linux/mount.h>
10 #include <linux/namei.h>
11 #include <linux/writeback.h>
12 #include <linux/falloc.h>
13 #include <linux/iversion.h>
14 #include <linux/ktime.h>
15 #include <linux/splice.h>
16
17 #include "super.h"
18 #include "mds_client.h"
19 #include "cache.h"
20 #include "io.h"
21 #include "metric.h"
22
23 static __le32 ceph_flags_sys2wire(struct ceph_mds_client *mdsc, u32 flags)
24 {
25 struct ceph_client *cl = mdsc->fsc->client;
26 u32 wire_flags = 0;
27
28 switch (flags & O_ACCMODE) {
29 case O_RDONLY:
30 wire_flags |= CEPH_O_RDONLY;
31 break;
32 case O_WRONLY:
33 wire_flags |= CEPH_O_WRONLY;
34 break;
35 case O_RDWR:
36 wire_flags |= CEPH_O_RDWR;
37 break;
38 }
39
40 flags &= ~O_ACCMODE;
41
42 #define ceph_sys2wire(a) if (flags & a) { wire_flags |= CEPH_##a; flags &= ~a; }
43
44 ceph_sys2wire(O_CREAT);
45 ceph_sys2wire(O_EXCL);
46 ceph_sys2wire(O_TRUNC);
47 ceph_sys2wire(O_DIRECTORY);
48 ceph_sys2wire(O_NOFOLLOW);
49
50 #undef ceph_sys2wire
51
52 if (flags)
53 doutc(cl, "unused open flags: %x\n", flags);
54
55 return cpu_to_le32(wire_flags);
56 }
57
58 /*
59 * Ceph file operations
60 *
61 * Implement basic open/close functionality, and implement
62 * read/write.
63 *
64 * We implement three modes of file I/O:
65 * - buffered uses the generic_file_aio_{read,write} helpers
66 *
67 * - synchronous is used when there is multi-client read/write
68 * sharing, avoids the page cache, and synchronously waits for an
69 * ack from the OSD.
70 *
71 * - direct io takes the variant of the sync path that references
72 * user pages directly.
73 *
74 * fsync() flushes and waits on dirty pages, but just queues metadata
75 * for writeback: since the MDS can recover size and mtime there is no
76 * need to wait for MDS acknowledgement.
77 */
78
79 /*
80 * How many pages to get in one call to iov_iter_get_pages(). This
81 * determines the size of the on-stack array used as a buffer.
82 */
83 #define ITER_GET_BVECS_PAGES 64
84
85 static ssize_t __iter_get_bvecs(struct iov_iter *iter, size_t maxsize,
86 struct bio_vec *bvecs)
87 {
88 size_t size = 0;
89 int bvec_idx = 0;
90
91 if (maxsize > iov_iter_count(iter))
92 maxsize = iov_iter_count(iter);
93
94 while (size < maxsize) {
95 struct page *pages[ITER_GET_BVECS_PAGES];
96 ssize_t bytes;
97 size_t start;
98 int idx = 0;
99
100 bytes = iov_iter_get_pages2(iter, pages, maxsize - size,
101 ITER_GET_BVECS_PAGES, &start);
102 if (bytes < 0)
103 return size ?: bytes;
104
105 size += bytes;
106
107 for ( ; bytes; idx++, bvec_idx++) {
108 int len = min_t(int, bytes, PAGE_SIZE - start);
109
110 bvec_set_page(&bvecs[bvec_idx], pages[idx], len, start);
111 bytes -= len;
112 start = 0;
113 }
114 }
115
116 return size;
117 }
118
119 /*
120 * iov_iter_get_pages() only considers one iov_iter segment, no matter
121 * what maxsize or maxpages are given. For ITER_BVEC that is a single
122 * page.
123 *
124 * Attempt to get up to @maxsize bytes worth of pages from @iter.
125 * Return the number of bytes in the created bio_vec array, or an error.
126 */
127 static ssize_t iter_get_bvecs_alloc(struct iov_iter *iter, size_t maxsize,
128 struct bio_vec **bvecs, int *num_bvecs)
129 {
130 struct bio_vec *bv;
131 size_t orig_count = iov_iter_count(iter);
132 ssize_t bytes;
133 int npages;
134
135 iov_iter_truncate(iter, maxsize);
136 npages = iov_iter_npages(iter, INT_MAX);
137 iov_iter_reexpand(iter, orig_count);
138
139 /*
140 * __iter_get_bvecs() may populate only part of the array -- zero it
141 * out.
142 */
143 bv = kvmalloc_array(npages, sizeof(*bv), GFP_KERNEL | __GFP_ZERO);
144 if (!bv)
145 return -ENOMEM;
146
147 bytes = __iter_get_bvecs(iter, maxsize, bv);
148 if (bytes < 0) {
149 /*
150 * No pages were pinned -- just free the array.
151 */
152 kvfree(bv);
153 return bytes;
154 }
155
156 *bvecs = bv;
157 *num_bvecs = npages;
158 return bytes;
159 }
160
161 static void put_bvecs(struct bio_vec *bvecs, int num_bvecs, bool should_dirty)
162 {
163 int i;
164
165 for (i = 0; i < num_bvecs; i++) {
166 if (bvecs[i].bv_page) {
167 if (should_dirty)
168 set_page_dirty_lock(bvecs[i].bv_page);
169 put_page(bvecs[i].bv_page);
170 }
171 }
172 kvfree(bvecs);
173 }
174
175 /*
176 * Prepare an open request. Preallocate ceph_cap to avoid an
177 * inopportune ENOMEM later.
178 */
179 static struct ceph_mds_request *
180 prepare_open_request(struct super_block *sb, int flags, int create_mode)
181 {
182 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(sb);
183 struct ceph_mds_request *req;
184 int want_auth = USE_ANY_MDS;
185 int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
186
187 if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
188 want_auth = USE_AUTH_MDS;
189
190 req = ceph_mdsc_create_request(mdsc, op, want_auth);
191 if (IS_ERR(req))
192 goto out;
193 req->r_fmode = ceph_flags_to_mode(flags);
194 req->r_args.open.flags = ceph_flags_sys2wire(mdsc, flags);
195 req->r_args.open.mode = cpu_to_le32(create_mode);
196 out:
197 return req;
198 }
199
200 static int ceph_init_file_info(struct inode *inode, struct file *file,
201 int fmode, bool isdir)
202 {
203 struct ceph_inode_info *ci = ceph_inode(inode);
204 struct ceph_mount_options *opt =
205 ceph_inode_to_fs_client(&ci->netfs.inode)->mount_options;
206 struct ceph_client *cl = ceph_inode_to_client(inode);
207 struct ceph_file_info *fi;
208 int ret;
209
210 doutc(cl, "%p %llx.%llx %p 0%o (%s)\n", inode, ceph_vinop(inode),
211 file, inode->i_mode, isdir ? "dir" : "regular");
212 BUG_ON(inode->i_fop->release != ceph_release);
213
214 if (isdir) {
215 struct ceph_dir_file_info *dfi =
216 kmem_cache_zalloc(ceph_dir_file_cachep, GFP_KERNEL);
217 if (!dfi)
218 return -ENOMEM;
219
220 file->private_data = dfi;
221 fi = &dfi->file_info;
222 dfi->next_offset = 2;
223 dfi->readdir_cache_idx = -1;
224 } else {
225 fi = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
226 if (!fi)
227 return -ENOMEM;
228
229 if (opt->flags & CEPH_MOUNT_OPT_NOPAGECACHE)
230 fi->flags |= CEPH_F_SYNC;
231
232 file->private_data = fi;
233 }
234
235 ceph_get_fmode(ci, fmode, 1);
236 fi->fmode = fmode;
237
238 spin_lock_init(&fi->rw_contexts_lock);
239 INIT_LIST_HEAD(&fi->rw_contexts);
240 fi->filp_gen = READ_ONCE(ceph_inode_to_fs_client(inode)->filp_gen);
241
242 if ((file->f_mode & FMODE_WRITE) && ceph_has_inline_data(ci)) {
243 ret = ceph_uninline_data(file);
244 if (ret < 0)
245 goto error;
246 }
247
248 return 0;
249
250 error:
251 ceph_fscache_unuse_cookie(inode, file->f_mode & FMODE_WRITE);
252 ceph_put_fmode(ci, fi->fmode, 1);
253 kmem_cache_free(ceph_file_cachep, fi);
254 /* wake up anyone waiting for caps on this inode */
255 wake_up_all(&ci->i_cap_wq);
256 return ret;
257 }
258
259 /*
260 * initialize private struct file data.
261 * if we fail, clean up by dropping fmode reference on the ceph_inode
262 */
263 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
264 {
265 struct ceph_client *cl = ceph_inode_to_client(inode);
266 int ret = 0;
267
268 switch (inode->i_mode & S_IFMT) {
269 case S_IFREG:
270 ceph_fscache_use_cookie(inode, file->f_mode & FMODE_WRITE);
271 fallthrough;
272 case S_IFDIR:
273 ret = ceph_init_file_info(inode, file, fmode,
274 S_ISDIR(inode->i_mode));
275 break;
276
277 case S_IFLNK:
278 doutc(cl, "%p %llx.%llx %p 0%o (symlink)\n", inode,
279 ceph_vinop(inode), file, inode->i_mode);
280 break;
281
282 default:
283 doutc(cl, "%p %llx.%llx %p 0%o (special)\n", inode,
284 ceph_vinop(inode), file, inode->i_mode);
285 /*
286 * we need to drop the open ref now, since we don't
287 * have .release set to ceph_release.
288 */
289 BUG_ON(inode->i_fop->release == ceph_release);
290
291 /* call the proper open fop */
292 ret = inode->i_fop->open(inode, file);
293 }
294 return ret;
295 }
296
297 /*
298 * try renew caps after session gets killed.
299 */
300 int ceph_renew_caps(struct inode *inode, int fmode)
301 {
302 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
303 struct ceph_client *cl = mdsc->fsc->client;
304 struct ceph_inode_info *ci = ceph_inode(inode);
305 struct ceph_mds_request *req;
306 int err, flags, wanted;
307
308 spin_lock(&ci->i_ceph_lock);
309 __ceph_touch_fmode(ci, mdsc, fmode);
310 wanted = __ceph_caps_file_wanted(ci);
311 if (__ceph_is_any_real_caps(ci) &&
312 (!(wanted & CEPH_CAP_ANY_WR) || ci->i_auth_cap)) {
313 int issued = __ceph_caps_issued(ci, NULL);
314 spin_unlock(&ci->i_ceph_lock);
315 doutc(cl, "%p %llx.%llx want %s issued %s updating mds_wanted\n",
316 inode, ceph_vinop(inode), ceph_cap_string(wanted),
317 ceph_cap_string(issued));
318 ceph_check_caps(ci, 0);
319 return 0;
320 }
321 spin_unlock(&ci->i_ceph_lock);
322
323 flags = 0;
324 if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
325 flags = O_RDWR;
326 else if (wanted & CEPH_CAP_FILE_RD)
327 flags = O_RDONLY;
328 else if (wanted & CEPH_CAP_FILE_WR)
329 flags = O_WRONLY;
330 #ifdef O_LAZY
331 if (wanted & CEPH_CAP_FILE_LAZYIO)
332 flags |= O_LAZY;
333 #endif
334
335 req = prepare_open_request(inode->i_sb, flags, 0);
336 if (IS_ERR(req)) {
337 err = PTR_ERR(req);
338 goto out;
339 }
340
341 req->r_inode = inode;
342 ihold(inode);
343 req->r_num_caps = 1;
344
345 err = ceph_mdsc_do_request(mdsc, NULL, req);
346 ceph_mdsc_put_request(req);
347 out:
348 doutc(cl, "%p %llx.%llx open result=%d\n", inode, ceph_vinop(inode),
349 err);
350 return err < 0 ? err : 0;
351 }
352
353 /*
354 * If we already have the requisite capabilities, we can satisfy
355 * the open request locally (no need to request new caps from the
356 * MDS). We do, however, need to inform the MDS (asynchronously)
357 * if our wanted caps set expands.
358 */
359 int ceph_open(struct inode *inode, struct file *file)
360 {
361 struct ceph_inode_info *ci = ceph_inode(inode);
362 struct ceph_fs_client *fsc = ceph_sb_to_fs_client(inode->i_sb);
363 struct ceph_client *cl = fsc->client;
364 struct ceph_mds_client *mdsc = fsc->mdsc;
365 struct ceph_mds_request *req;
366 struct ceph_file_info *fi = file->private_data;
367 int err;
368 int flags, fmode, wanted;
369
370 if (fi) {
371 doutc(cl, "file %p is already opened\n", file);
372 return 0;
373 }
374
375 /* filter out O_CREAT|O_EXCL; vfs did that already. yuck. */
376 flags = file->f_flags & ~(O_CREAT|O_EXCL);
377 if (S_ISDIR(inode->i_mode)) {
378 flags = O_DIRECTORY; /* mds likes to know */
379 } else if (S_ISREG(inode->i_mode)) {
380 err = fscrypt_file_open(inode, file);
381 if (err)
382 return err;
383 }
384
385 doutc(cl, "%p %llx.%llx file %p flags %d (%d)\n", inode,
386 ceph_vinop(inode), file, flags, file->f_flags);
387 fmode = ceph_flags_to_mode(flags);
388 wanted = ceph_caps_for_mode(fmode);
389
390 /* snapped files are read-only */
391 if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
392 return -EROFS;
393
394 /* trivially open snapdir */
395 if (ceph_snap(inode) == CEPH_SNAPDIR) {
396 return ceph_init_file(inode, file, fmode);
397 }
398
399 /*
400 * No need to block if we have caps on the auth MDS (for
401 * write) or any MDS (for read). Update wanted set
402 * asynchronously.
403 */
404 spin_lock(&ci->i_ceph_lock);
405 if (__ceph_is_any_real_caps(ci) &&
406 (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
407 int mds_wanted = __ceph_caps_mds_wanted(ci, true);
408 int issued = __ceph_caps_issued(ci, NULL);
409
410 doutc(cl, "open %p fmode %d want %s issued %s using existing\n",
411 inode, fmode, ceph_cap_string(wanted),
412 ceph_cap_string(issued));
413 __ceph_touch_fmode(ci, mdsc, fmode);
414 spin_unlock(&ci->i_ceph_lock);
415
416 /* adjust wanted? */
417 if ((issued & wanted) != wanted &&
418 (mds_wanted & wanted) != wanted &&
419 ceph_snap(inode) != CEPH_SNAPDIR)
420 ceph_check_caps(ci, 0);
421
422 return ceph_init_file(inode, file, fmode);
423 } else if (ceph_snap(inode) != CEPH_NOSNAP &&
424 (ci->i_snap_caps & wanted) == wanted) {
425 __ceph_touch_fmode(ci, mdsc, fmode);
426 spin_unlock(&ci->i_ceph_lock);
427 return ceph_init_file(inode, file, fmode);
428 }
429
430 spin_unlock(&ci->i_ceph_lock);
431
432 doutc(cl, "open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
433 req = prepare_open_request(inode->i_sb, flags, 0);
434 if (IS_ERR(req)) {
435 err = PTR_ERR(req);
436 goto out;
437 }
438 req->r_inode = inode;
439 ihold(inode);
440
441 req->r_num_caps = 1;
442 err = ceph_mdsc_do_request(mdsc, NULL, req);
443 if (!err)
444 err = ceph_init_file(inode, file, req->r_fmode);
445 ceph_mdsc_put_request(req);
446 doutc(cl, "open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
447 out:
448 return err;
449 }
450
451 /* Clone the layout from a synchronous create, if the dir now has Dc caps */
452 static void
453 cache_file_layout(struct inode *dst, struct inode *src)
454 {
455 struct ceph_inode_info *cdst = ceph_inode(dst);
456 struct ceph_inode_info *csrc = ceph_inode(src);
457
458 spin_lock(&cdst->i_ceph_lock);
459 if ((__ceph_caps_issued(cdst, NULL) & CEPH_CAP_DIR_CREATE) &&
460 !ceph_file_layout_is_valid(&cdst->i_cached_layout)) {
461 memcpy(&cdst->i_cached_layout, &csrc->i_layout,
462 sizeof(cdst->i_cached_layout));
463 rcu_assign_pointer(cdst->i_cached_layout.pool_ns,
464 ceph_try_get_string(csrc->i_layout.pool_ns));
465 }
466 spin_unlock(&cdst->i_ceph_lock);
467 }
468
469 /*
470 * Try to set up an async create. We need caps, a file layout, and inode number,
471 * and either a lease on the dentry or complete dir info. If any of those
472 * criteria are not satisfied, then return false and the caller can go
473 * synchronous.
474 */
475 static int try_prep_async_create(struct inode *dir, struct dentry *dentry,
476 struct ceph_file_layout *lo, u64 *pino)
477 {
478 struct ceph_inode_info *ci = ceph_inode(dir);
479 struct ceph_dentry_info *di = ceph_dentry(dentry);
480 int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_CREATE;
481 u64 ino;
482
483 spin_lock(&ci->i_ceph_lock);
484 /* No auth cap means no chance for Dc caps */
485 if (!ci->i_auth_cap)
486 goto no_async;
487
488 /* Any delegated inos? */
489 if (xa_empty(&ci->i_auth_cap->session->s_delegated_inos))
490 goto no_async;
491
492 if (!ceph_file_layout_is_valid(&ci->i_cached_layout))
493 goto no_async;
494
495 if ((__ceph_caps_issued(ci, NULL) & want) != want)
496 goto no_async;
497
498 if (d_in_lookup(dentry)) {
499 if (!__ceph_dir_is_complete(ci))
500 goto no_async;
501 spin_lock(&dentry->d_lock);
502 di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
503 spin_unlock(&dentry->d_lock);
504 } else if (atomic_read(&ci->i_shared_gen) !=
505 READ_ONCE(di->lease_shared_gen)) {
506 goto no_async;
507 }
508
509 ino = ceph_get_deleg_ino(ci->i_auth_cap->session);
510 if (!ino)
511 goto no_async;
512
513 *pino = ino;
514 ceph_take_cap_refs(ci, want, false);
515 memcpy(lo, &ci->i_cached_layout, sizeof(*lo));
516 rcu_assign_pointer(lo->pool_ns,
517 ceph_try_get_string(ci->i_cached_layout.pool_ns));
518 got = want;
519 no_async:
520 spin_unlock(&ci->i_ceph_lock);
521 return got;
522 }
523
524 static void restore_deleg_ino(struct inode *dir, u64 ino)
525 {
526 struct ceph_client *cl = ceph_inode_to_client(dir);
527 struct ceph_inode_info *ci = ceph_inode(dir);
528 struct ceph_mds_session *s = NULL;
529
530 spin_lock(&ci->i_ceph_lock);
531 if (ci->i_auth_cap)
532 s = ceph_get_mds_session(ci->i_auth_cap->session);
533 spin_unlock(&ci->i_ceph_lock);
534 if (s) {
535 int err = ceph_restore_deleg_ino(s, ino);
536 if (err)
537 pr_warn_client(cl,
538 "unable to restore delegated ino 0x%llx to session: %d\n",
539 ino, err);
540 ceph_put_mds_session(s);
541 }
542 }
543
544 static void wake_async_create_waiters(struct inode *inode,
545 struct ceph_mds_session *session)
546 {
547 struct ceph_inode_info *ci = ceph_inode(inode);
548 bool check_cap = false;
549
550 spin_lock(&ci->i_ceph_lock);
551 if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE) {
552 ci->i_ceph_flags &= ~CEPH_I_ASYNC_CREATE;
553 wake_up_bit(&ci->i_ceph_flags, CEPH_ASYNC_CREATE_BIT);
554
555 if (ci->i_ceph_flags & CEPH_I_ASYNC_CHECK_CAPS) {
556 ci->i_ceph_flags &= ~CEPH_I_ASYNC_CHECK_CAPS;
557 check_cap = true;
558 }
559 }
560 ceph_kick_flushing_inode_caps(session, ci);
561 spin_unlock(&ci->i_ceph_lock);
562
563 if (check_cap)
564 ceph_check_caps(ci, CHECK_CAPS_FLUSH);
565 }
566
567 static void ceph_async_create_cb(struct ceph_mds_client *mdsc,
568 struct ceph_mds_request *req)
569 {
570 struct ceph_client *cl = mdsc->fsc->client;
571 struct dentry *dentry = req->r_dentry;
572 struct inode *dinode = d_inode(dentry);
573 struct inode *tinode = req->r_target_inode;
574 int result = req->r_err ? req->r_err :
575 le32_to_cpu(req->r_reply_info.head->result);
576
577 WARN_ON_ONCE(dinode && tinode && dinode != tinode);
578
579 /* MDS changed -- caller must resubmit */
580 if (result == -EJUKEBOX)
581 goto out;
582
583 mapping_set_error(req->r_parent->i_mapping, result);
584
585 if (result) {
586 int pathlen = 0;
587 u64 base = 0;
588 char *path = ceph_mdsc_build_path(mdsc, req->r_dentry, &pathlen,
589 &base, 0);
590
591 pr_warn_client(cl,
592 "async create failure path=(%llx)%s result=%d!\n",
593 base, IS_ERR(path) ? "<<bad>>" : path, result);
594 ceph_mdsc_free_path(path, pathlen);
595
596 ceph_dir_clear_complete(req->r_parent);
597 if (!d_unhashed(dentry))
598 d_drop(dentry);
599
600 if (dinode) {
601 mapping_set_error(dinode->i_mapping, result);
602 ceph_inode_shutdown(dinode);
603 wake_async_create_waiters(dinode, req->r_session);
604 }
605 }
606
607 if (tinode) {
608 u64 ino = ceph_vino(tinode).ino;
609
610 if (req->r_deleg_ino != ino)
611 pr_warn_client(cl,
612 "inode number mismatch! err=%d deleg_ino=0x%llx target=0x%llx\n",
613 req->r_err, req->r_deleg_ino, ino);
614
615 mapping_set_error(tinode->i_mapping, result);
616 wake_async_create_waiters(tinode, req->r_session);
617 } else if (!result) {
618 pr_warn_client(cl, "no req->r_target_inode for 0x%llx\n",
619 req->r_deleg_ino);
620 }
621 out:
622 ceph_mdsc_release_dir_caps(req);
623 }
624
625 static int ceph_finish_async_create(struct inode *dir, struct inode *inode,
626 struct dentry *dentry,
627 struct file *file, umode_t mode,
628 struct ceph_mds_request *req,
629 struct ceph_acl_sec_ctx *as_ctx,
630 struct ceph_file_layout *lo)
631 {
632 int ret;
633 char xattr_buf[4];
634 struct ceph_mds_reply_inode in = { };
635 struct ceph_mds_reply_info_in iinfo = { .in = &in };
636 struct ceph_inode_info *ci = ceph_inode(dir);
637 struct ceph_dentry_info *di = ceph_dentry(dentry);
638 struct timespec64 now;
639 struct ceph_string *pool_ns;
640 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
641 struct ceph_client *cl = mdsc->fsc->client;
642 struct ceph_vino vino = { .ino = req->r_deleg_ino,
643 .snap = CEPH_NOSNAP };
644
645 ktime_get_real_ts64(&now);
646
647 iinfo.inline_version = CEPH_INLINE_NONE;
648 iinfo.change_attr = 1;
649 ceph_encode_timespec64(&iinfo.btime, &now);
650
651 if (req->r_pagelist) {
652 iinfo.xattr_len = req->r_pagelist->length;
653 iinfo.xattr_data = req->r_pagelist->mapped_tail;
654 } else {
655 /* fake it */
656 iinfo.xattr_len = ARRAY_SIZE(xattr_buf);
657 iinfo.xattr_data = xattr_buf;
658 memset(iinfo.xattr_data, 0, iinfo.xattr_len);
659 }
660
661 in.ino = cpu_to_le64(vino.ino);
662 in.snapid = cpu_to_le64(CEPH_NOSNAP);
663 in.version = cpu_to_le64(1); // ???
664 in.cap.caps = in.cap.wanted = cpu_to_le32(CEPH_CAP_ALL_FILE);
665 in.cap.cap_id = cpu_to_le64(1);
666 in.cap.realm = cpu_to_le64(ci->i_snap_realm->ino);
667 in.cap.flags = CEPH_CAP_FLAG_AUTH;
668 in.ctime = in.mtime = in.atime = iinfo.btime;
669 in.truncate_seq = cpu_to_le32(1);
670 in.truncate_size = cpu_to_le64(-1ULL);
671 in.xattr_version = cpu_to_le64(1);
672 in.uid = cpu_to_le32(from_kuid(&init_user_ns,
673 mapped_fsuid(req->r_mnt_idmap,
674 &init_user_ns)));
675 if (dir->i_mode & S_ISGID) {
676 in.gid = cpu_to_le32(from_kgid(&init_user_ns, dir->i_gid));
677
678 /* Directories always inherit the setgid bit. */
679 if (S_ISDIR(mode))
680 mode |= S_ISGID;
681 } else {
682 in.gid = cpu_to_le32(from_kgid(&init_user_ns,
683 mapped_fsgid(req->r_mnt_idmap,
684 &init_user_ns)));
685 }
686 in.mode = cpu_to_le32((u32)mode);
687
688 in.nlink = cpu_to_le32(1);
689 in.max_size = cpu_to_le64(lo->stripe_unit);
690
691 ceph_file_layout_to_legacy(lo, &in.layout);
692 /* lo is private, so pool_ns can't change */
693 pool_ns = rcu_dereference_raw(lo->pool_ns);
694 if (pool_ns) {
695 iinfo.pool_ns_len = pool_ns->len;
696 iinfo.pool_ns_data = pool_ns->str;
697 }
698
699 down_read(&mdsc->snap_rwsem);
700 ret = ceph_fill_inode(inode, NULL, &iinfo, NULL, req->r_session,
701 req->r_fmode, NULL);
702 up_read(&mdsc->snap_rwsem);
703 if (ret) {
704 doutc(cl, "failed to fill inode: %d\n", ret);
705 ceph_dir_clear_complete(dir);
706 if (!d_unhashed(dentry))
707 d_drop(dentry);
708 discard_new_inode(inode);
709 } else {
710 struct dentry *dn;
711
712 doutc(cl, "d_adding new inode 0x%llx to 0x%llx/%s\n",
713 vino.ino, ceph_ino(dir), dentry->d_name.name);
714 ceph_dir_clear_ordered(dir);
715 ceph_init_inode_acls(inode, as_ctx);
716 if (inode->i_state & I_NEW) {
717 /*
718 * If it's not I_NEW, then someone created this before
719 * we got here. Assume the server is aware of it at
720 * that point and don't worry about setting
721 * CEPH_I_ASYNC_CREATE.
722 */
723 ceph_inode(inode)->i_ceph_flags = CEPH_I_ASYNC_CREATE;
724 unlock_new_inode(inode);
725 }
726 if (d_in_lookup(dentry) || d_really_is_negative(dentry)) {
727 if (!d_unhashed(dentry))
728 d_drop(dentry);
729 dn = d_splice_alias(inode, dentry);
730 WARN_ON_ONCE(dn && dn != dentry);
731 }
732 file->f_mode |= FMODE_CREATED;
733 ret = finish_open(file, dentry, ceph_open);
734 }
735
736 spin_lock(&dentry->d_lock);
737 di->flags &= ~CEPH_DENTRY_ASYNC_CREATE;
738 wake_up_bit(&di->flags, CEPH_DENTRY_ASYNC_CREATE_BIT);
739 spin_unlock(&dentry->d_lock);
740
741 return ret;
742 }
743
744 /*
745 * Do a lookup + open with a single request. If we get a non-existent
746 * file or symlink, return 1 so the VFS can retry.
747 */
748 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
749 struct file *file, unsigned flags, umode_t mode)
750 {
751 struct mnt_idmap *idmap = file_mnt_idmap(file);
752 struct ceph_fs_client *fsc = ceph_sb_to_fs_client(dir->i_sb);
753 struct ceph_client *cl = fsc->client;
754 struct ceph_mds_client *mdsc = fsc->mdsc;
755 struct ceph_mds_request *req;
756 struct inode *new_inode = NULL;
757 struct dentry *dn;
758 struct ceph_acl_sec_ctx as_ctx = {};
759 bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
760 int mask;
761 int err;
762
763 doutc(cl, "%p %llx.%llx dentry %p '%pd' %s flags %d mode 0%o\n",
764 dir, ceph_vinop(dir), dentry, dentry,
765 d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
766
767 if (dentry->d_name.len > NAME_MAX)
768 return -ENAMETOOLONG;
769
770 err = ceph_wait_on_conflict_unlink(dentry);
771 if (err)
772 return err;
773 /*
774 * Do not truncate the file, since atomic_open is called before the
775 * permission check. The caller will do the truncation afterward.
776 */
777 flags &= ~O_TRUNC;
778
779 retry:
780 if (flags & O_CREAT) {
781 if (ceph_quota_is_max_files_exceeded(dir))
782 return -EDQUOT;
783
784 new_inode = ceph_new_inode(dir, dentry, &mode, &as_ctx);
785 if (IS_ERR(new_inode)) {
786 err = PTR_ERR(new_inode);
787 goto out_ctx;
788 }
789 /* Async create can't handle more than a page of xattrs */
790 if (as_ctx.pagelist &&
791 !list_is_singular(&as_ctx.pagelist->head))
792 try_async = false;
793 } else if (!d_in_lookup(dentry)) {
794 /* If it's not being looked up, it's negative */
795 return -ENOENT;
796 }
797
798 /* do the open */
799 req = prepare_open_request(dir->i_sb, flags, mode);
800 if (IS_ERR(req)) {
801 err = PTR_ERR(req);
802 goto out_ctx;
803 }
804 req->r_dentry = dget(dentry);
805 req->r_num_caps = 2;
806 mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
807 if (ceph_security_xattr_wanted(dir))
808 mask |= CEPH_CAP_XATTR_SHARED;
809 req->r_args.open.mask = cpu_to_le32(mask);
810 req->r_parent = dir;
811 if (req->r_op == CEPH_MDS_OP_CREATE)
812 req->r_mnt_idmap = mnt_idmap_get(idmap);
813 ihold(dir);
814 if (IS_ENCRYPTED(dir)) {
815 set_bit(CEPH_MDS_R_FSCRYPT_FILE, &req->r_req_flags);
816 err = fscrypt_prepare_lookup_partial(dir, dentry);
817 if (err < 0)
818 goto out_req;
819 }
820
821 if (flags & O_CREAT) {
822 struct ceph_file_layout lo;
823
824 req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL |
825 CEPH_CAP_XATTR_EXCL;
826 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
827
828 ceph_as_ctx_to_req(req, &as_ctx);
829
830 if (try_async && (req->r_dir_caps =
831 try_prep_async_create(dir, dentry, &lo,
832 &req->r_deleg_ino))) {
833 struct ceph_vino vino = { .ino = req->r_deleg_ino,
834 .snap = CEPH_NOSNAP };
835 struct ceph_dentry_info *di = ceph_dentry(dentry);
836
837 set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
838 req->r_args.open.flags |= cpu_to_le32(CEPH_O_EXCL);
839 req->r_callback = ceph_async_create_cb;
840
841 /* Hash inode before RPC */
842 new_inode = ceph_get_inode(dir->i_sb, vino, new_inode);
843 if (IS_ERR(new_inode)) {
844 err = PTR_ERR(new_inode);
845 new_inode = NULL;
846 goto out_req;
847 }
848 WARN_ON_ONCE(!(new_inode->i_state & I_NEW));
849
850 spin_lock(&dentry->d_lock);
851 di->flags |= CEPH_DENTRY_ASYNC_CREATE;
852 spin_unlock(&dentry->d_lock);
853
854 err = ceph_mdsc_submit_request(mdsc, dir, req);
855 if (!err) {
856 err = ceph_finish_async_create(dir, new_inode,
857 dentry, file,
858 mode, req,
859 &as_ctx, &lo);
860 new_inode = NULL;
861 } else if (err == -EJUKEBOX) {
862 restore_deleg_ino(dir, req->r_deleg_ino);
863 ceph_mdsc_put_request(req);
864 discard_new_inode(new_inode);
865 ceph_release_acl_sec_ctx(&as_ctx);
866 memset(&as_ctx, 0, sizeof(as_ctx));
867 new_inode = NULL;
868 try_async = false;
869 ceph_put_string(rcu_dereference_raw(lo.pool_ns));
870 goto retry;
871 }
872 ceph_put_string(rcu_dereference_raw(lo.pool_ns));
873 goto out_req;
874 }
875 }
876
877 set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
878 req->r_new_inode = new_inode;
879 new_inode = NULL;
880 err = ceph_mdsc_do_request(mdsc, (flags & O_CREAT) ? dir : NULL, req);
881 if (err == -ENOENT) {
882 dentry = ceph_handle_snapdir(req, dentry);
883 if (IS_ERR(dentry)) {
884 err = PTR_ERR(dentry);
885 goto out_req;
886 }
887 err = 0;
888 }
889
890 if (!err && (flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
891 err = ceph_handle_notrace_create(dir, dentry);
892
893 if (d_in_lookup(dentry)) {
894 dn = ceph_finish_lookup(req, dentry, err);
895 if (IS_ERR(dn))
896 err = PTR_ERR(dn);
897 } else {
898 /* we were given a hashed negative dentry */
899 dn = NULL;
900 }
901 if (err)
902 goto out_req;
903 if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
904 /* make vfs retry on splice, ENOENT, or symlink */
905 doutc(cl, "finish_no_open on dn %p\n", dn);
906 err = finish_no_open(file, dn);
907 } else {
908 if (IS_ENCRYPTED(dir) &&
909 !fscrypt_has_permitted_context(dir, d_inode(dentry))) {
910 pr_warn_client(cl,
911 "Inconsistent encryption context (parent %llx:%llx child %llx:%llx)\n",
912 ceph_vinop(dir), ceph_vinop(d_inode(dentry)));
913 goto out_req;
914 }
915
916 doutc(cl, "finish_open on dn %p\n", dn);
917 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
918 struct inode *newino = d_inode(dentry);
919
920 cache_file_layout(dir, newino);
921 ceph_init_inode_acls(newino, &as_ctx);
922 file->f_mode |= FMODE_CREATED;
923 }
924 err = finish_open(file, dentry, ceph_open);
925 }
926 out_req:
927 ceph_mdsc_put_request(req);
928 iput(new_inode);
929 out_ctx:
930 ceph_release_acl_sec_ctx(&as_ctx);
931 doutc(cl, "result=%d\n", err);
932 return err;
933 }
934
935 int ceph_release(struct inode *inode, struct file *file)
936 {
937 struct ceph_client *cl = ceph_inode_to_client(inode);
938 struct ceph_inode_info *ci = ceph_inode(inode);
939
940 if (S_ISDIR(inode->i_mode)) {
941 struct ceph_dir_file_info *dfi = file->private_data;
942 doutc(cl, "%p %llx.%llx dir file %p\n", inode,
943 ceph_vinop(inode), file);
944 WARN_ON(!list_empty(&dfi->file_info.rw_contexts));
945
946 ceph_put_fmode(ci, dfi->file_info.fmode, 1);
947
948 if (dfi->last_readdir)
949 ceph_mdsc_put_request(dfi->last_readdir);
950 kfree(dfi->last_name);
951 kfree(dfi->dir_info);
952 kmem_cache_free(ceph_dir_file_cachep, dfi);
953 } else {
954 struct ceph_file_info *fi = file->private_data;
955 doutc(cl, "%p %llx.%llx regular file %p\n", inode,
956 ceph_vinop(inode), file);
957 WARN_ON(!list_empty(&fi->rw_contexts));
958
959 ceph_fscache_unuse_cookie(inode, file->f_mode & FMODE_WRITE);
960 ceph_put_fmode(ci, fi->fmode, 1);
961
962 kmem_cache_free(ceph_file_cachep, fi);
963 }
964
965 /* wake up anyone waiting for caps on this inode */
966 wake_up_all(&ci->i_cap_wq);
967 return 0;
968 }
969
970 enum {
971 HAVE_RETRIED = 1,
972 CHECK_EOF = 2,
973 READ_INLINE = 3,
974 };
975
976 /*
977 * Completely synchronous read and write methods. Direct from __user
978 * buffer to osd, or directly to user pages (if O_DIRECT).
979 *
980 * If the read spans object boundary, just do multiple reads. (That's not
981 * atomic, but good enough for now.)
982 *
983 * If we get a short result from the OSD, check against i_size; we need to
984 * only return a short read to the caller if we hit EOF.
985 */
986 ssize_t __ceph_sync_read(struct inode *inode, loff_t *ki_pos,
987 struct iov_iter *to, int *retry_op,
988 u64 *last_objver)
989 {
990 struct ceph_inode_info *ci = ceph_inode(inode);
991 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
992 struct ceph_client *cl = fsc->client;
993 struct ceph_osd_client *osdc = &fsc->client->osdc;
994 ssize_t ret;
995 u64 off = *ki_pos;
996 u64 len = iov_iter_count(to);
997 u64 i_size = i_size_read(inode);
998 bool sparse = IS_ENCRYPTED(inode) || ceph_test_mount_opt(fsc, SPARSEREAD);
999 u64 objver = 0;
1000
1001 doutc(cl, "on inode %p %llx.%llx %llx~%llx\n", inode,
1002 ceph_vinop(inode), *ki_pos, len);
1003
1004 if (ceph_inode_is_shutdown(inode))
1005 return -EIO;
1006
1007 if (!len)
1008 return 0;
1009 /*
1010 * flush any page cache pages in this range. this
1011 * will make concurrent normal and sync io slow,
1012 * but it will at least behave sensibly when they are
1013 * in sequence.
1014 */
1015 ret = filemap_write_and_wait_range(inode->i_mapping,
1016 off, off + len - 1);
1017 if (ret < 0)
1018 return ret;
1019
1020 ret = 0;
1021 while ((len = iov_iter_count(to)) > 0) {
1022 struct ceph_osd_request *req;
1023 struct page **pages;
1024 int num_pages;
1025 size_t page_off;
1026 bool more;
1027 int idx;
1028 size_t left;
1029 struct ceph_osd_req_op *op;
1030 u64 read_off = off;
1031 u64 read_len = len;
1032
1033 /* determine new offset/length if encrypted */
1034 ceph_fscrypt_adjust_off_and_len(inode, &read_off, &read_len);
1035
1036 doutc(cl, "orig %llu~%llu reading %llu~%llu", off, len,
1037 read_off, read_len);
1038
1039 req = ceph_osdc_new_request(osdc, &ci->i_layout,
1040 ci->i_vino, read_off, &read_len, 0, 1,
1041 sparse ? CEPH_OSD_OP_SPARSE_READ :
1042 CEPH_OSD_OP_READ,
1043 CEPH_OSD_FLAG_READ,
1044 NULL, ci->i_truncate_seq,
1045 ci->i_truncate_size, false);
1046 if (IS_ERR(req)) {
1047 ret = PTR_ERR(req);
1048 break;
1049 }
1050
1051 /* adjust len downward if the request truncated the len */
1052 if (off + len > read_off + read_len)
1053 len = read_off + read_len - off;
1054 more = len < iov_iter_count(to);
1055
1056 num_pages = calc_pages_for(read_off, read_len);
1057 page_off = offset_in_page(off);
1058 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1059 if (IS_ERR(pages)) {
1060 ceph_osdc_put_request(req);
1061 ret = PTR_ERR(pages);
1062 break;
1063 }
1064
1065 osd_req_op_extent_osd_data_pages(req, 0, pages, read_len,
1066 offset_in_page(read_off),
1067 false, false);
1068
1069 op = &req->r_ops[0];
1070 if (sparse) {
1071 ret = ceph_alloc_sparse_ext_map(op);
1072 if (ret) {
1073 ceph_osdc_put_request(req);
1074 break;
1075 }
1076 }
1077
1078 ceph_osdc_start_request(osdc, req);
1079 ret = ceph_osdc_wait_request(osdc, req);
1080
1081 ceph_update_read_metrics(&fsc->mdsc->metric,
1082 req->r_start_latency,
1083 req->r_end_latency,
1084 read_len, ret);
1085
1086 if (ret > 0)
1087 objver = req->r_version;
1088
1089 i_size = i_size_read(inode);
1090 doutc(cl, "%llu~%llu got %zd i_size %llu%s\n", off, len,
1091 ret, i_size, (more ? " MORE" : ""));
1092
1093 /* Fix it to go to end of extent map */
1094 if (sparse && ret >= 0)
1095 ret = ceph_sparse_ext_map_end(op);
1096 else if (ret == -ENOENT)
1097 ret = 0;
1098
1099 if (ret > 0 && IS_ENCRYPTED(inode)) {
1100 int fret;
1101
1102 fret = ceph_fscrypt_decrypt_extents(inode, pages,
1103 read_off, op->extent.sparse_ext,
1104 op->extent.sparse_ext_cnt);
1105 if (fret < 0) {
1106 ret = fret;
1107 ceph_osdc_put_request(req);
1108 break;
1109 }
1110
1111 /* account for any partial block at the beginning */
1112 fret -= (off - read_off);
1113
1114 /*
1115 * Short read after big offset adjustment?
1116 * Nothing is usable, just call it a zero
1117 * len read.
1118 */
1119 fret = max(fret, 0);
1120
1121 /* account for partial block at the end */
1122 ret = min_t(ssize_t, fret, len);
1123 }
1124
1125 ceph_osdc_put_request(req);
1126
1127 /* Short read but not EOF? Zero out the remainder. */
1128 if (ret >= 0 && ret < len && (off + ret < i_size)) {
1129 int zlen = min(len - ret, i_size - off - ret);
1130 int zoff = page_off + ret;
1131
1132 doutc(cl, "zero gap %llu~%llu\n", off + ret,
1133 off + ret + zlen);
1134 ceph_zero_page_vector_range(zoff, zlen, pages);
1135 ret += zlen;
1136 }
1137
1138 idx = 0;
1139 left = ret > 0 ? ret : 0;
1140 while (left > 0) {
1141 size_t plen, copied;
1142
1143 plen = min_t(size_t, left, PAGE_SIZE - page_off);
1144 SetPageUptodate(pages[idx]);
1145 copied = copy_page_to_iter(pages[idx++],
1146 page_off, plen, to);
1147 off += copied;
1148 left -= copied;
1149 page_off = 0;
1150 if (copied < plen) {
1151 ret = -EFAULT;
1152 break;
1153 }
1154 }
1155 ceph_release_page_vector(pages, num_pages);
1156
1157 if (ret < 0) {
1158 if (ret == -EBLOCKLISTED)
1159 fsc->blocklisted = true;
1160 break;
1161 }
1162
1163 if (off >= i_size || !more)
1164 break;
1165 }
1166
1167 if (ret > 0) {
1168 if (off > *ki_pos) {
1169 if (off >= i_size) {
1170 *retry_op = CHECK_EOF;
1171 ret = i_size - *ki_pos;
1172 *ki_pos = i_size;
1173 } else {
1174 ret = off - *ki_pos;
1175 *ki_pos = off;
1176 }
1177 }
1178
1179 if (last_objver)
1180 *last_objver = objver;
1181 }
1182 doutc(cl, "result %zd retry_op %d\n", ret, *retry_op);
1183 return ret;
1184 }
1185
1186 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
1187 int *retry_op)
1188 {
1189 struct file *file = iocb->ki_filp;
1190 struct inode *inode = file_inode(file);
1191 struct ceph_client *cl = ceph_inode_to_client(inode);
1192
1193 doutc(cl, "on file %p %llx~%zx %s\n", file, iocb->ki_pos,
1194 iov_iter_count(to),
1195 (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
1196
1197 return __ceph_sync_read(inode, &iocb->ki_pos, to, retry_op, NULL);
1198 }
1199
1200 struct ceph_aio_request {
1201 struct kiocb *iocb;
1202 size_t total_len;
1203 bool write;
1204 bool should_dirty;
1205 int error;
1206 struct list_head osd_reqs;
1207 unsigned num_reqs;
1208 atomic_t pending_reqs;
1209 struct timespec64 mtime;
1210 struct ceph_cap_flush *prealloc_cf;
1211 };
1212
1213 struct ceph_aio_work {
1214 struct work_struct work;
1215 struct ceph_osd_request *req;
1216 };
1217
1218 static void ceph_aio_retry_work(struct work_struct *work);
1219
1220 static void ceph_aio_complete(struct inode *inode,
1221 struct ceph_aio_request *aio_req)
1222 {
1223 struct ceph_client *cl = ceph_inode_to_client(inode);
1224 struct ceph_inode_info *ci = ceph_inode(inode);
1225 int ret;
1226
1227 if (!atomic_dec_and_test(&aio_req->pending_reqs))
1228 return;
1229
1230 if (aio_req->iocb->ki_flags & IOCB_DIRECT)
1231 inode_dio_end(inode);
1232
1233 ret = aio_req->error;
1234 if (!ret)
1235 ret = aio_req->total_len;
1236
1237 doutc(cl, "%p %llx.%llx rc %d\n", inode, ceph_vinop(inode), ret);
1238
1239 if (ret >= 0 && aio_req->write) {
1240 int dirty;
1241
1242 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
1243 if (endoff > i_size_read(inode)) {
1244 if (ceph_inode_set_size(inode, endoff))
1245 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY);
1246 }
1247
1248 spin_lock(&ci->i_ceph_lock);
1249 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1250 &aio_req->prealloc_cf);
1251 spin_unlock(&ci->i_ceph_lock);
1252 if (dirty)
1253 __mark_inode_dirty(inode, dirty);
1254
1255 }
1256
1257 ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
1258 CEPH_CAP_FILE_RD));
1259
1260 aio_req->iocb->ki_complete(aio_req->iocb, ret);
1261
1262 ceph_free_cap_flush(aio_req->prealloc_cf);
1263 kfree(aio_req);
1264 }
1265
1266 static void ceph_aio_complete_req(struct ceph_osd_request *req)
1267 {
1268 int rc = req->r_result;
1269 struct inode *inode = req->r_inode;
1270 struct ceph_aio_request *aio_req = req->r_priv;
1271 struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
1272 struct ceph_osd_req_op *op = &req->r_ops[0];
1273 struct ceph_client_metric *metric = &ceph_sb_to_mdsc(inode->i_sb)->metric;
1274 unsigned int len = osd_data->bvec_pos.iter.bi_size;
1275 bool sparse = (op->op == CEPH_OSD_OP_SPARSE_READ);
1276 struct ceph_client *cl = ceph_inode_to_client(inode);
1277
1278 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_BVECS);
1279 BUG_ON(!osd_data->num_bvecs);
1280
1281 doutc(cl, "req %p inode %p %llx.%llx, rc %d bytes %u\n", req,
1282 inode, ceph_vinop(inode), rc, len);
1283
1284 if (rc == -EOLDSNAPC) {
1285 struct ceph_aio_work *aio_work;
1286 BUG_ON(!aio_req->write);
1287
1288 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
1289 if (aio_work) {
1290 INIT_WORK(&aio_work->work, ceph_aio_retry_work);
1291 aio_work->req = req;
1292 queue_work(ceph_inode_to_fs_client(inode)->inode_wq,
1293 &aio_work->work);
1294 return;
1295 }
1296 rc = -ENOMEM;
1297 } else if (!aio_req->write) {
1298 if (sparse && rc >= 0)
1299 rc = ceph_sparse_ext_map_end(op);
1300 if (rc == -ENOENT)
1301 rc = 0;
1302 if (rc >= 0 && len > rc) {
1303 struct iov_iter i;
1304 int zlen = len - rc;
1305
1306 /*
1307 * If read is satisfied by single OSD request,
1308 * it can pass EOF. Otherwise read is within
1309 * i_size.
1310 */
1311 if (aio_req->num_reqs == 1) {
1312 loff_t i_size = i_size_read(inode);
1313 loff_t endoff = aio_req->iocb->ki_pos + rc;
1314 if (endoff < i_size)
1315 zlen = min_t(size_t, zlen,
1316 i_size - endoff);
1317 aio_req->total_len = rc + zlen;
1318 }
1319
1320 iov_iter_bvec(&i, ITER_DEST, osd_data->bvec_pos.bvecs,
1321 osd_data->num_bvecs, len);
1322 iov_iter_advance(&i, rc);
1323 iov_iter_zero(zlen, &i);
1324 }
1325 }
1326
1327 /* r_start_latency == 0 means the request was not submitted */
1328 if (req->r_start_latency) {
1329 if (aio_req->write)
1330 ceph_update_write_metrics(metric, req->r_start_latency,
1331 req->r_end_latency, len, rc);
1332 else
1333 ceph_update_read_metrics(metric, req->r_start_latency,
1334 req->r_end_latency, len, rc);
1335 }
1336
1337 put_bvecs(osd_data->bvec_pos.bvecs, osd_data->num_bvecs,
1338 aio_req->should_dirty);
1339 ceph_osdc_put_request(req);
1340
1341 if (rc < 0)
1342 cmpxchg(&aio_req->error, 0, rc);
1343
1344 ceph_aio_complete(inode, aio_req);
1345 return;
1346 }
1347
1348 static void ceph_aio_retry_work(struct work_struct *work)
1349 {
1350 struct ceph_aio_work *aio_work =
1351 container_of(work, struct ceph_aio_work, work);
1352 struct ceph_osd_request *orig_req = aio_work->req;
1353 struct ceph_aio_request *aio_req = orig_req->r_priv;
1354 struct inode *inode = orig_req->r_inode;
1355 struct ceph_inode_info *ci = ceph_inode(inode);
1356 struct ceph_snap_context *snapc;
1357 struct ceph_osd_request *req;
1358 int ret;
1359
1360 spin_lock(&ci->i_ceph_lock);
1361 if (__ceph_have_pending_cap_snap(ci)) {
1362 struct ceph_cap_snap *capsnap =
1363 list_last_entry(&ci->i_cap_snaps,
1364 struct ceph_cap_snap,
1365 ci_item);
1366 snapc = ceph_get_snap_context(capsnap->context);
1367 } else {
1368 BUG_ON(!ci->i_head_snapc);
1369 snapc = ceph_get_snap_context(ci->i_head_snapc);
1370 }
1371 spin_unlock(&ci->i_ceph_lock);
1372
1373 req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 1,
1374 false, GFP_NOFS);
1375 if (!req) {
1376 ret = -ENOMEM;
1377 req = orig_req;
1378 goto out;
1379 }
1380
1381 req->r_flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1382 ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
1383 ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
1384
1385 req->r_ops[0] = orig_req->r_ops[0];
1386
1387 req->r_mtime = aio_req->mtime;
1388 req->r_data_offset = req->r_ops[0].extent.offset;
1389
1390 ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
1391 if (ret) {
1392 ceph_osdc_put_request(req);
1393 req = orig_req;
1394 goto out;
1395 }
1396
1397 ceph_osdc_put_request(orig_req);
1398
1399 req->r_callback = ceph_aio_complete_req;
1400 req->r_inode = inode;
1401 req->r_priv = aio_req;
1402
1403 ceph_osdc_start_request(req->r_osdc, req);
1404 out:
1405 if (ret < 0) {
1406 req->r_result = ret;
1407 ceph_aio_complete_req(req);
1408 }
1409
1410 ceph_put_snap_context(snapc);
1411 kfree(aio_work);
1412 }
1413
1414 static ssize_t
1415 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
1416 struct ceph_snap_context *snapc,
1417 struct ceph_cap_flush **pcf)
1418 {
1419 struct file *file = iocb->ki_filp;
1420 struct inode *inode = file_inode(file);
1421 struct ceph_inode_info *ci = ceph_inode(inode);
1422 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
1423 struct ceph_client *cl = fsc->client;
1424 struct ceph_client_metric *metric = &fsc->mdsc->metric;
1425 struct ceph_vino vino;
1426 struct ceph_osd_request *req;
1427 struct bio_vec *bvecs;
1428 struct ceph_aio_request *aio_req = NULL;
1429 int num_pages = 0;
1430 int flags;
1431 int ret = 0;
1432 struct timespec64 mtime = current_time(inode);
1433 size_t count = iov_iter_count(iter);
1434 loff_t pos = iocb->ki_pos;
1435 bool write = iov_iter_rw(iter) == WRITE;
1436 bool should_dirty = !write && user_backed_iter(iter);
1437 bool sparse = ceph_test_mount_opt(fsc, SPARSEREAD);
1438
1439 if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1440 return -EROFS;
1441
1442 doutc(cl, "sync_direct_%s on file %p %lld~%u snapc %p seq %lld\n",
1443 (write ? "write" : "read"), file, pos, (unsigned)count,
1444 snapc, snapc ? snapc->seq : 0);
1445
1446 if (write) {
1447 int ret2;
1448
1449 ceph_fscache_invalidate(inode, true);
1450
1451 ret2 = invalidate_inode_pages2_range(inode->i_mapping,
1452 pos >> PAGE_SHIFT,
1453 (pos + count - 1) >> PAGE_SHIFT);
1454 if (ret2 < 0)
1455 doutc(cl, "invalidate_inode_pages2_range returned %d\n",
1456 ret2);
1457
1458 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1459 } else {
1460 flags = CEPH_OSD_FLAG_READ;
1461 }
1462
1463 while (iov_iter_count(iter) > 0) {
1464 u64 size = iov_iter_count(iter);
1465 ssize_t len;
1466 struct ceph_osd_req_op *op;
1467 int readop = sparse ? CEPH_OSD_OP_SPARSE_READ : CEPH_OSD_OP_READ;
1468
1469 if (write)
1470 size = min_t(u64, size, fsc->mount_options->wsize);
1471 else
1472 size = min_t(u64, size, fsc->mount_options->rsize);
1473
1474 vino = ceph_vino(inode);
1475 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1476 vino, pos, &size, 0,
1477 1,
1478 write ? CEPH_OSD_OP_WRITE : readop,
1479 flags, snapc,
1480 ci->i_truncate_seq,
1481 ci->i_truncate_size,
1482 false);
1483 if (IS_ERR(req)) {
1484 ret = PTR_ERR(req);
1485 break;
1486 }
1487
1488 len = iter_get_bvecs_alloc(iter, size, &bvecs, &num_pages);
1489 if (len < 0) {
1490 ceph_osdc_put_request(req);
1491 ret = len;
1492 break;
1493 }
1494 if (len != size)
1495 osd_req_op_extent_update(req, 0, len);
1496
1497 /*
1498 * To simplify error handling, allow AIO when IO within i_size
1499 * or IO can be satisfied by single OSD request.
1500 */
1501 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
1502 (len == count || pos + count <= i_size_read(inode))) {
1503 aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
1504 if (aio_req) {
1505 aio_req->iocb = iocb;
1506 aio_req->write = write;
1507 aio_req->should_dirty = should_dirty;
1508 INIT_LIST_HEAD(&aio_req->osd_reqs);
1509 if (write) {
1510 aio_req->mtime = mtime;
1511 swap(aio_req->prealloc_cf, *pcf);
1512 }
1513 }
1514 /* ignore error */
1515 }
1516
1517 if (write) {
1518 /*
1519 * throw out any page cache pages in this range. this
1520 * may block.
1521 */
1522 truncate_inode_pages_range(inode->i_mapping, pos,
1523 PAGE_ALIGN(pos + len) - 1);
1524
1525 req->r_mtime = mtime;
1526 }
1527
1528 osd_req_op_extent_osd_data_bvecs(req, 0, bvecs, num_pages, len);
1529 op = &req->r_ops[0];
1530 if (sparse) {
1531 ret = ceph_alloc_sparse_ext_map(op);
1532 if (ret) {
1533 ceph_osdc_put_request(req);
1534 break;
1535 }
1536 }
1537
1538 if (aio_req) {
1539 aio_req->total_len += len;
1540 aio_req->num_reqs++;
1541 atomic_inc(&aio_req->pending_reqs);
1542
1543 req->r_callback = ceph_aio_complete_req;
1544 req->r_inode = inode;
1545 req->r_priv = aio_req;
1546 list_add_tail(&req->r_private_item, &aio_req->osd_reqs);
1547
1548 pos += len;
1549 continue;
1550 }
1551
1552 ceph_osdc_start_request(req->r_osdc, req);
1553 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1554
1555 if (write)
1556 ceph_update_write_metrics(metric, req->r_start_latency,
1557 req->r_end_latency, len, ret);
1558 else
1559 ceph_update_read_metrics(metric, req->r_start_latency,
1560 req->r_end_latency, len, ret);
1561
1562 size = i_size_read(inode);
1563 if (!write) {
1564 if (sparse && ret >= 0)
1565 ret = ceph_sparse_ext_map_end(op);
1566 else if (ret == -ENOENT)
1567 ret = 0;
1568
1569 if (ret >= 0 && ret < len && pos + ret < size) {
1570 struct iov_iter i;
1571 int zlen = min_t(size_t, len - ret,
1572 size - pos - ret);
1573
1574 iov_iter_bvec(&i, ITER_DEST, bvecs, num_pages, len);
1575 iov_iter_advance(&i, ret);
1576 iov_iter_zero(zlen, &i);
1577 ret += zlen;
1578 }
1579 if (ret >= 0)
1580 len = ret;
1581 }
1582
1583 put_bvecs(bvecs, num_pages, should_dirty);
1584 ceph_osdc_put_request(req);
1585 if (ret < 0)
1586 break;
1587
1588 pos += len;
1589 if (!write && pos >= size)
1590 break;
1591
1592 if (write && pos > size) {
1593 if (ceph_inode_set_size(inode, pos))
1594 ceph_check_caps(ceph_inode(inode),
1595 CHECK_CAPS_AUTHONLY);
1596 }
1597 }
1598
1599 if (aio_req) {
1600 LIST_HEAD(osd_reqs);
1601
1602 if (aio_req->num_reqs == 0) {
1603 kfree(aio_req);
1604 return ret;
1605 }
1606
1607 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1608 CEPH_CAP_FILE_RD);
1609
1610 list_splice(&aio_req->osd_reqs, &osd_reqs);
1611 inode_dio_begin(inode);
1612 while (!list_empty(&osd_reqs)) {
1613 req = list_first_entry(&osd_reqs,
1614 struct ceph_osd_request,
1615 r_private_item);
1616 list_del_init(&req->r_private_item);
1617 if (ret >= 0)
1618 ceph_osdc_start_request(req->r_osdc, req);
1619 if (ret < 0) {
1620 req->r_result = ret;
1621 ceph_aio_complete_req(req);
1622 }
1623 }
1624 return -EIOCBQUEUED;
1625 }
1626
1627 if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1628 ret = pos - iocb->ki_pos;
1629 iocb->ki_pos = pos;
1630 }
1631 return ret;
1632 }
1633
1634 /*
1635 * Synchronous write, straight from __user pointer or user pages.
1636 *
1637 * If write spans object boundary, just do multiple writes. (For a
1638 * correct atomic write, we should e.g. take write locks on all
1639 * objects, rollback on failure, etc.)
1640 */
1641 static ssize_t
1642 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1643 struct ceph_snap_context *snapc)
1644 {
1645 struct file *file = iocb->ki_filp;
1646 struct inode *inode = file_inode(file);
1647 struct ceph_inode_info *ci = ceph_inode(inode);
1648 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
1649 struct ceph_client *cl = fsc->client;
1650 struct ceph_osd_client *osdc = &fsc->client->osdc;
1651 struct ceph_osd_request *req;
1652 struct page **pages;
1653 u64 len;
1654 int num_pages;
1655 int written = 0;
1656 int ret;
1657 bool check_caps = false;
1658 struct timespec64 mtime = current_time(inode);
1659 size_t count = iov_iter_count(from);
1660
1661 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1662 return -EROFS;
1663
1664 doutc(cl, "on file %p %lld~%u snapc %p seq %lld\n", file, pos,
1665 (unsigned)count, snapc, snapc->seq);
1666
1667 ret = filemap_write_and_wait_range(inode->i_mapping,
1668 pos, pos + count - 1);
1669 if (ret < 0)
1670 return ret;
1671
1672 ceph_fscache_invalidate(inode, false);
1673
1674 while ((len = iov_iter_count(from)) > 0) {
1675 size_t left;
1676 int n;
1677 u64 write_pos = pos;
1678 u64 write_len = len;
1679 u64 objnum, objoff;
1680 u32 xlen;
1681 u64 assert_ver = 0;
1682 bool rmw;
1683 bool first, last;
1684 struct iov_iter saved_iter = *from;
1685 size_t off;
1686
1687 ceph_fscrypt_adjust_off_and_len(inode, &write_pos, &write_len);
1688
1689 /* clamp the length to the end of first object */
1690 ceph_calc_file_object_mapping(&ci->i_layout, write_pos,
1691 write_len, &objnum, &objoff,
1692 &xlen);
1693 write_len = xlen;
1694
1695 /* adjust len downward if it goes beyond current object */
1696 if (pos + len > write_pos + write_len)
1697 len = write_pos + write_len - pos;
1698
1699 /*
1700 * If we had to adjust the length or position to align with a
1701 * crypto block, then we must do a read/modify/write cycle. We
1702 * use a version assertion to redrive the thing if something
1703 * changes in between.
1704 */
1705 first = pos != write_pos;
1706 last = (pos + len) != (write_pos + write_len);
1707 rmw = first || last;
1708
1709 doutc(cl, "ino %llx %lld~%llu adjusted %lld~%llu -- %srmw\n",
1710 ci->i_vino.ino, pos, len, write_pos, write_len,
1711 rmw ? "" : "no ");
1712
1713 /*
1714 * The data is emplaced into the page as it would be if it were
1715 * in an array of pagecache pages.
1716 */
1717 num_pages = calc_pages_for(write_pos, write_len);
1718 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1719 if (IS_ERR(pages)) {
1720 ret = PTR_ERR(pages);
1721 break;
1722 }
1723
1724 /* Do we need to preload the pages? */
1725 if (rmw) {
1726 u64 first_pos = write_pos;
1727 u64 last_pos = (write_pos + write_len) - CEPH_FSCRYPT_BLOCK_SIZE;
1728 u64 read_len = CEPH_FSCRYPT_BLOCK_SIZE;
1729 struct ceph_osd_req_op *op;
1730
1731 /* We should only need to do this for encrypted inodes */
1732 WARN_ON_ONCE(!IS_ENCRYPTED(inode));
1733
1734 /* No need to do two reads if first and last blocks are same */
1735 if (first && last_pos == first_pos)
1736 last = false;
1737
1738 /*
1739 * Allocate a read request for one or two extents,
1740 * depending on how the request was aligned.
1741 */
1742 req = ceph_osdc_new_request(osdc, &ci->i_layout,
1743 ci->i_vino, first ? first_pos : last_pos,
1744 &read_len, 0, (first && last) ? 2 : 1,
1745 CEPH_OSD_OP_SPARSE_READ, CEPH_OSD_FLAG_READ,
1746 NULL, ci->i_truncate_seq,
1747 ci->i_truncate_size, false);
1748 if (IS_ERR(req)) {
1749 ceph_release_page_vector(pages, num_pages);
1750 ret = PTR_ERR(req);
1751 break;
1752 }
1753
1754 /* Something is misaligned! */
1755 if (read_len != CEPH_FSCRYPT_BLOCK_SIZE) {
1756 ceph_osdc_put_request(req);
1757 ceph_release_page_vector(pages, num_pages);
1758 ret = -EIO;
1759 break;
1760 }
1761
1762 /* Add extent for first block? */
1763 op = &req->r_ops[0];
1764
1765 if (first) {
1766 osd_req_op_extent_osd_data_pages(req, 0, pages,
1767 CEPH_FSCRYPT_BLOCK_SIZE,
1768 offset_in_page(first_pos),
1769 false, false);
1770 /* We only expect a single extent here */
1771 ret = __ceph_alloc_sparse_ext_map(op, 1);
1772 if (ret) {
1773 ceph_osdc_put_request(req);
1774 ceph_release_page_vector(pages, num_pages);
1775 break;
1776 }
1777 }
1778
1779 /* Add extent for last block */
1780 if (last) {
1781 /* Init the other extent if first extent has been used */
1782 if (first) {
1783 op = &req->r_ops[1];
1784 osd_req_op_extent_init(req, 1,
1785 CEPH_OSD_OP_SPARSE_READ,
1786 last_pos, CEPH_FSCRYPT_BLOCK_SIZE,
1787 ci->i_truncate_size,
1788 ci->i_truncate_seq);
1789 }
1790
1791 ret = __ceph_alloc_sparse_ext_map(op, 1);
1792 if (ret) {
1793 ceph_osdc_put_request(req);
1794 ceph_release_page_vector(pages, num_pages);
1795 break;
1796 }
1797
1798 osd_req_op_extent_osd_data_pages(req, first ? 1 : 0,
1799 &pages[num_pages - 1],
1800 CEPH_FSCRYPT_BLOCK_SIZE,
1801 offset_in_page(last_pos),
1802 false, false);
1803 }
1804
1805 ceph_osdc_start_request(osdc, req);
1806 ret = ceph_osdc_wait_request(osdc, req);
1807
1808 /* FIXME: length field is wrong if there are 2 extents */
1809 ceph_update_read_metrics(&fsc->mdsc->metric,
1810 req->r_start_latency,
1811 req->r_end_latency,
1812 read_len, ret);
1813
1814 /* Ok if object is not already present */
1815 if (ret == -ENOENT) {
1816 /*
1817 * If there is no object, then we can't assert
1818 * on its version. Set it to 0, and we'll use an
1819 * exclusive create instead.
1820 */
1821 ceph_osdc_put_request(req);
1822 ret = 0;
1823
1824 /*
1825 * zero out the soon-to-be uncopied parts of the
1826 * first and last pages.
1827 */
1828 if (first)
1829 zero_user_segment(pages[0], 0,
1830 offset_in_page(first_pos));
1831 if (last)
1832 zero_user_segment(pages[num_pages - 1],
1833 offset_in_page(last_pos),
1834 PAGE_SIZE);
1835 } else {
1836 if (ret < 0) {
1837 ceph_osdc_put_request(req);
1838 ceph_release_page_vector(pages, num_pages);
1839 break;
1840 }
1841
1842 op = &req->r_ops[0];
1843 if (op->extent.sparse_ext_cnt == 0) {
1844 if (first)
1845 zero_user_segment(pages[0], 0,
1846 offset_in_page(first_pos));
1847 else
1848 zero_user_segment(pages[num_pages - 1],
1849 offset_in_page(last_pos),
1850 PAGE_SIZE);
1851 } else if (op->extent.sparse_ext_cnt != 1 ||
1852 ceph_sparse_ext_map_end(op) !=
1853 CEPH_FSCRYPT_BLOCK_SIZE) {
1854 ret = -EIO;
1855 ceph_osdc_put_request(req);
1856 ceph_release_page_vector(pages, num_pages);
1857 break;
1858 }
1859
1860 if (first && last) {
1861 op = &req->r_ops[1];
1862 if (op->extent.sparse_ext_cnt == 0) {
1863 zero_user_segment(pages[num_pages - 1],
1864 offset_in_page(last_pos),
1865 PAGE_SIZE);
1866 } else if (op->extent.sparse_ext_cnt != 1 ||
1867 ceph_sparse_ext_map_end(op) !=
1868 CEPH_FSCRYPT_BLOCK_SIZE) {
1869 ret = -EIO;
1870 ceph_osdc_put_request(req);
1871 ceph_release_page_vector(pages, num_pages);
1872 break;
1873 }
1874 }
1875
1876 /* Grab assert version. It must be non-zero. */
1877 assert_ver = req->r_version;
1878 WARN_ON_ONCE(ret > 0 && assert_ver == 0);
1879
1880 ceph_osdc_put_request(req);
1881 if (first) {
1882 ret = ceph_fscrypt_decrypt_block_inplace(inode,
1883 pages[0], CEPH_FSCRYPT_BLOCK_SIZE,
1884 offset_in_page(first_pos),
1885 first_pos >> CEPH_FSCRYPT_BLOCK_SHIFT);
1886 if (ret < 0) {
1887 ceph_release_page_vector(pages, num_pages);
1888 break;
1889 }
1890 }
1891 if (last) {
1892 ret = ceph_fscrypt_decrypt_block_inplace(inode,
1893 pages[num_pages - 1],
1894 CEPH_FSCRYPT_BLOCK_SIZE,
1895 offset_in_page(last_pos),
1896 last_pos >> CEPH_FSCRYPT_BLOCK_SHIFT);
1897 if (ret < 0) {
1898 ceph_release_page_vector(pages, num_pages);
1899 break;
1900 }
1901 }
1902 }
1903 }
1904
1905 left = len;
1906 off = offset_in_page(pos);
1907 for (n = 0; n < num_pages; n++) {
1908 size_t plen = min_t(size_t, left, PAGE_SIZE - off);
1909
1910 /* copy the data */
1911 ret = copy_page_from_iter(pages[n], off, plen, from);
1912 if (ret != plen) {
1913 ret = -EFAULT;
1914 break;
1915 }
1916 off = 0;
1917 left -= ret;
1918 }
1919 if (ret < 0) {
1920 doutc(cl, "write failed with %d\n", ret);
1921 ceph_release_page_vector(pages, num_pages);
1922 break;
1923 }
1924
1925 if (IS_ENCRYPTED(inode)) {
1926 ret = ceph_fscrypt_encrypt_pages(inode, pages,
1927 write_pos, write_len,
1928 GFP_KERNEL);
1929 if (ret < 0) {
1930 doutc(cl, "encryption failed with %d\n", ret);
1931 ceph_release_page_vector(pages, num_pages);
1932 break;
1933 }
1934 }
1935
1936 req = ceph_osdc_new_request(osdc, &ci->i_layout,
1937 ci->i_vino, write_pos, &write_len,
1938 rmw ? 1 : 0, rmw ? 2 : 1,
1939 CEPH_OSD_OP_WRITE,
1940 CEPH_OSD_FLAG_WRITE,
1941 snapc, ci->i_truncate_seq,
1942 ci->i_truncate_size, false);
1943 if (IS_ERR(req)) {
1944 ret = PTR_ERR(req);
1945 ceph_release_page_vector(pages, num_pages);
1946 break;
1947 }
1948
1949 doutc(cl, "write op %lld~%llu\n", write_pos, write_len);
1950 osd_req_op_extent_osd_data_pages(req, rmw ? 1 : 0, pages, write_len,
1951 offset_in_page(write_pos), false,
1952 true);
1953 req->r_inode = inode;
1954 req->r_mtime = mtime;
1955
1956 /* Set up the assertion */
1957 if (rmw) {
1958 /*
1959 * Set up the assertion. If we don't have a version
1960 * number, then the object doesn't exist yet. Use an
1961 * exclusive create instead of a version assertion in
1962 * that case.
1963 */
1964 if (assert_ver) {
1965 osd_req_op_init(req, 0, CEPH_OSD_OP_ASSERT_VER, 0);
1966 req->r_ops[0].assert_ver.ver = assert_ver;
1967 } else {
1968 osd_req_op_init(req, 0, CEPH_OSD_OP_CREATE,
1969 CEPH_OSD_OP_FLAG_EXCL);
1970 }
1971 }
1972
1973 ceph_osdc_start_request(osdc, req);
1974 ret = ceph_osdc_wait_request(osdc, req);
1975
1976 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1977 req->r_end_latency, len, ret);
1978 ceph_osdc_put_request(req);
1979 if (ret != 0) {
1980 doutc(cl, "osd write returned %d\n", ret);
1981 /* Version changed! Must re-do the rmw cycle */
1982 if ((assert_ver && (ret == -ERANGE || ret == -EOVERFLOW)) ||
1983 (!assert_ver && ret == -EEXIST)) {
1984 /* We should only ever see this on a rmw */
1985 WARN_ON_ONCE(!rmw);
1986
1987 /* The version should never go backward */
1988 WARN_ON_ONCE(ret == -EOVERFLOW);
1989
1990 *from = saved_iter;
1991
1992 /* FIXME: limit number of times we loop? */
1993 continue;
1994 }
1995 ceph_set_error_write(ci);
1996 break;
1997 }
1998
1999 ceph_clear_error_write(ci);
2000
2001 /*
2002 * We successfully wrote to a range of the file. Declare
2003 * that region of the pagecache invalid.
2004 */
2005 ret = invalidate_inode_pages2_range(
2006 inode->i_mapping,
2007 pos >> PAGE_SHIFT,
2008 (pos + len - 1) >> PAGE_SHIFT);
2009 if (ret < 0) {
2010 doutc(cl, "invalidate_inode_pages2_range returned %d\n",
2011 ret);
2012 ret = 0;
2013 }
2014 pos += len;
2015 written += len;
2016 doutc(cl, "written %d\n", written);
2017 if (pos > i_size_read(inode)) {
2018 check_caps = ceph_inode_set_size(inode, pos);
2019 if (check_caps)
2020 ceph_check_caps(ceph_inode(inode),
2021 CHECK_CAPS_AUTHONLY);
2022 }
2023
2024 }
2025
2026 if (ret != -EOLDSNAPC && written > 0) {
2027 ret = written;
2028 iocb->ki_pos = pos;
2029 }
2030 doutc(cl, "returning %d\n", ret);
2031 return ret;
2032 }
2033
2034 /*
2035 * Wrap generic_file_aio_read with checks for cap bits on the inode.
2036 * Atomically grab references, so that those bits are not released
2037 * back to the MDS mid-read.
2038 *
2039 * Hmm, the sync read case isn't actually async... should it be?
2040 */
2041 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
2042 {
2043 struct file *filp = iocb->ki_filp;
2044 struct ceph_file_info *fi = filp->private_data;
2045 size_t len = iov_iter_count(to);
2046 struct inode *inode = file_inode(filp);
2047 struct ceph_inode_info *ci = ceph_inode(inode);
2048 bool direct_lock = iocb->ki_flags & IOCB_DIRECT;
2049 struct ceph_client *cl = ceph_inode_to_client(inode);
2050 ssize_t ret;
2051 int want = 0, got = 0;
2052 int retry_op = 0, read = 0;
2053
2054 again:
2055 doutc(cl, "%llu~%u trying to get caps on %p %llx.%llx\n",
2056 iocb->ki_pos, (unsigned)len, inode, ceph_vinop(inode));
2057
2058 if (ceph_inode_is_shutdown(inode))
2059 return -ESTALE;
2060
2061 if (direct_lock)
2062 ceph_start_io_direct(inode);
2063 else
2064 ceph_start_io_read(inode);
2065
2066 if (!(fi->flags & CEPH_F_SYNC) && !direct_lock)
2067 want |= CEPH_CAP_FILE_CACHE;
2068 if (fi->fmode & CEPH_FILE_MODE_LAZY)
2069 want |= CEPH_CAP_FILE_LAZYIO;
2070
2071 ret = ceph_get_caps(filp, CEPH_CAP_FILE_RD, want, -1, &got);
2072 if (ret < 0) {
2073 if (direct_lock)
2074 ceph_end_io_direct(inode);
2075 else
2076 ceph_end_io_read(inode);
2077 return ret;
2078 }
2079
2080 if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
2081 (iocb->ki_flags & IOCB_DIRECT) ||
2082 (fi->flags & CEPH_F_SYNC)) {
2083
2084 doutc(cl, "sync %p %llx.%llx %llu~%u got cap refs on %s\n",
2085 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
2086 ceph_cap_string(got));
2087
2088 if (!ceph_has_inline_data(ci)) {
2089 if (!retry_op &&
2090 (iocb->ki_flags & IOCB_DIRECT) &&
2091 !IS_ENCRYPTED(inode)) {
2092 ret = ceph_direct_read_write(iocb, to,
2093 NULL, NULL);
2094 if (ret >= 0 && ret < len)
2095 retry_op = CHECK_EOF;
2096 } else {
2097 ret = ceph_sync_read(iocb, to, &retry_op);
2098 }
2099 } else {
2100 retry_op = READ_INLINE;
2101 }
2102 } else {
2103 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
2104 doutc(cl, "async %p %llx.%llx %llu~%u got cap refs on %s\n",
2105 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
2106 ceph_cap_string(got));
2107 ceph_add_rw_context(fi, &rw_ctx);
2108 ret = generic_file_read_iter(iocb, to);
2109 ceph_del_rw_context(fi, &rw_ctx);
2110 }
2111
2112 doutc(cl, "%p %llx.%llx dropping cap refs on %s = %d\n",
2113 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
2114 ceph_put_cap_refs(ci, got);
2115
2116 if (direct_lock)
2117 ceph_end_io_direct(inode);
2118 else
2119 ceph_end_io_read(inode);
2120
2121 if (retry_op > HAVE_RETRIED && ret >= 0) {
2122 int statret;
2123 struct page *page = NULL;
2124 loff_t i_size;
2125 if (retry_op == READ_INLINE) {
2126 page = __page_cache_alloc(GFP_KERNEL);
2127 if (!page)
2128 return -ENOMEM;
2129 }
2130
2131 statret = __ceph_do_getattr(inode, page,
2132 CEPH_STAT_CAP_INLINE_DATA, !!page);
2133 if (statret < 0) {
2134 if (page)
2135 __free_page(page);
2136 if (statret == -ENODATA) {
2137 BUG_ON(retry_op != READ_INLINE);
2138 goto again;
2139 }
2140 return statret;
2141 }
2142
2143 i_size = i_size_read(inode);
2144 if (retry_op == READ_INLINE) {
2145 BUG_ON(ret > 0 || read > 0);
2146 if (iocb->ki_pos < i_size &&
2147 iocb->ki_pos < PAGE_SIZE) {
2148 loff_t end = min_t(loff_t, i_size,
2149 iocb->ki_pos + len);
2150 end = min_t(loff_t, end, PAGE_SIZE);
2151 if (statret < end)
2152 zero_user_segment(page, statret, end);
2153 ret = copy_page_to_iter(page,
2154 iocb->ki_pos & ~PAGE_MASK,
2155 end - iocb->ki_pos, to);
2156 iocb->ki_pos += ret;
2157 read += ret;
2158 }
2159 if (iocb->ki_pos < i_size && read < len) {
2160 size_t zlen = min_t(size_t, len - read,
2161 i_size - iocb->ki_pos);
2162 ret = iov_iter_zero(zlen, to);
2163 iocb->ki_pos += ret;
2164 read += ret;
2165 }
2166 __free_pages(page, 0);
2167 return read;
2168 }
2169
2170 /* hit EOF or hole? */
2171 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
2172 ret < len) {
2173 doutc(cl, "hit hole, ppos %lld < size %lld, reading more\n",
2174 iocb->ki_pos, i_size);
2175
2176 read += ret;
2177 len -= ret;
2178 retry_op = HAVE_RETRIED;
2179 goto again;
2180 }
2181 }
2182
2183 if (ret >= 0)
2184 ret += read;
2185
2186 return ret;
2187 }
2188
2189 /*
2190 * Wrap filemap_splice_read with checks for cap bits on the inode.
2191 * Atomically grab references, so that those bits are not released
2192 * back to the MDS mid-read.
2193 */
2194 static ssize_t ceph_splice_read(struct file *in, loff_t *ppos,
2195 struct pipe_inode_info *pipe,
2196 size_t len, unsigned int flags)
2197 {
2198 struct ceph_file_info *fi = in->private_data;
2199 struct inode *inode = file_inode(in);
2200 struct ceph_inode_info *ci = ceph_inode(inode);
2201 ssize_t ret;
2202 int want = 0, got = 0;
2203 CEPH_DEFINE_RW_CONTEXT(rw_ctx, 0);
2204
2205 dout("splice_read %p %llx.%llx %llu~%zu trying to get caps on %p\n",
2206 inode, ceph_vinop(inode), *ppos, len, inode);
2207
2208 if (ceph_inode_is_shutdown(inode))
2209 return -ESTALE;
2210
2211 if (ceph_has_inline_data(ci) ||
2212 (fi->flags & CEPH_F_SYNC))
2213 return copy_splice_read(in, ppos, pipe, len, flags);
2214
2215 ceph_start_io_read(inode);
2216
2217 want = CEPH_CAP_FILE_CACHE;
2218 if (fi->fmode & CEPH_FILE_MODE_LAZY)
2219 want |= CEPH_CAP_FILE_LAZYIO;
2220
2221 ret = ceph_get_caps(in, CEPH_CAP_FILE_RD, want, -1, &got);
2222 if (ret < 0)
2223 goto out_end;
2224
2225 if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) == 0) {
2226 dout("splice_read/sync %p %llx.%llx %llu~%zu got cap refs on %s\n",
2227 inode, ceph_vinop(inode), *ppos, len,
2228 ceph_cap_string(got));
2229
2230 ceph_put_cap_refs(ci, got);
2231 ceph_end_io_read(inode);
2232 return copy_splice_read(in, ppos, pipe, len, flags);
2233 }
2234
2235 dout("splice_read %p %llx.%llx %llu~%zu got cap refs on %s\n",
2236 inode, ceph_vinop(inode), *ppos, len, ceph_cap_string(got));
2237
2238 rw_ctx.caps = got;
2239 ceph_add_rw_context(fi, &rw_ctx);
2240 ret = filemap_splice_read(in, ppos, pipe, len, flags);
2241 ceph_del_rw_context(fi, &rw_ctx);
2242
2243 dout("splice_read %p %llx.%llx dropping cap refs on %s = %zd\n",
2244 inode, ceph_vinop(inode), ceph_cap_string(got), ret);
2245
2246 ceph_put_cap_refs(ci, got);
2247 out_end:
2248 ceph_end_io_read(inode);
2249 return ret;
2250 }
2251
2252 /*
2253 * Take cap references to avoid releasing caps to MDS mid-write.
2254 *
2255 * If we are synchronous, and write with an old snap context, the OSD
2256 * may return EOLDSNAPC. In that case, retry the write.. _after_
2257 * dropping our cap refs and allowing the pending snap to logically
2258 * complete _before_ this write occurs.
2259 *
2260 * If we are near ENOSPC, write synchronously.
2261 */
2262 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
2263 {
2264 struct file *file = iocb->ki_filp;
2265 struct ceph_file_info *fi = file->private_data;
2266 struct inode *inode = file_inode(file);
2267 struct ceph_inode_info *ci = ceph_inode(inode);
2268 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
2269 struct ceph_client *cl = fsc->client;
2270 struct ceph_osd_client *osdc = &fsc->client->osdc;
2271 struct ceph_cap_flush *prealloc_cf;
2272 ssize_t count, written = 0;
2273 int err, want = 0, got;
2274 bool direct_lock = false;
2275 u32 map_flags;
2276 u64 pool_flags;
2277 loff_t pos;
2278 loff_t limit = max(i_size_read(inode), fsc->max_file_size);
2279
2280 if (ceph_inode_is_shutdown(inode))
2281 return -ESTALE;
2282
2283 if (ceph_snap(inode) != CEPH_NOSNAP)
2284 return -EROFS;
2285
2286 prealloc_cf = ceph_alloc_cap_flush();
2287 if (!prealloc_cf)
2288 return -ENOMEM;
2289
2290 if ((iocb->ki_flags & (IOCB_DIRECT | IOCB_APPEND)) == IOCB_DIRECT)
2291 direct_lock = true;
2292
2293 retry_snap:
2294 if (direct_lock)
2295 ceph_start_io_direct(inode);
2296 else
2297 ceph_start_io_write(inode);
2298
2299 if (iocb->ki_flags & IOCB_APPEND) {
2300 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
2301 if (err < 0)
2302 goto out;
2303 }
2304
2305 err = generic_write_checks(iocb, from);
2306 if (err <= 0)
2307 goto out;
2308
2309 pos = iocb->ki_pos;
2310 if (unlikely(pos >= limit)) {
2311 err = -EFBIG;
2312 goto out;
2313 } else {
2314 iov_iter_truncate(from, limit - pos);
2315 }
2316
2317 count = iov_iter_count(from);
2318 if (ceph_quota_is_max_bytes_exceeded(inode, pos + count)) {
2319 err = -EDQUOT;
2320 goto out;
2321 }
2322
2323 down_read(&osdc->lock);
2324 map_flags = osdc->osdmap->flags;
2325 pool_flags = ceph_pg_pool_flags(osdc->osdmap, ci->i_layout.pool_id);
2326 up_read(&osdc->lock);
2327 if ((map_flags & CEPH_OSDMAP_FULL) ||
2328 (pool_flags & CEPH_POOL_FLAG_FULL)) {
2329 err = -ENOSPC;
2330 goto out;
2331 }
2332
2333 err = file_remove_privs(file);
2334 if (err)
2335 goto out;
2336
2337 doutc(cl, "%p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
2338 inode, ceph_vinop(inode), pos, count,
2339 i_size_read(inode));
2340 if (!(fi->flags & CEPH_F_SYNC) && !direct_lock)
2341 want |= CEPH_CAP_FILE_BUFFER;
2342 if (fi->fmode & CEPH_FILE_MODE_LAZY)
2343 want |= CEPH_CAP_FILE_LAZYIO;
2344 got = 0;
2345 err = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, pos + count, &got);
2346 if (err < 0)
2347 goto out;
2348
2349 err = file_update_time(file);
2350 if (err)
2351 goto out_caps;
2352
2353 inode_inc_iversion_raw(inode);
2354
2355 doutc(cl, "%p %llx.%llx %llu~%zd got cap refs on %s\n",
2356 inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
2357
2358 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
2359 (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
2360 (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
2361 struct ceph_snap_context *snapc;
2362 struct iov_iter data;
2363
2364 spin_lock(&ci->i_ceph_lock);
2365 if (__ceph_have_pending_cap_snap(ci)) {
2366 struct ceph_cap_snap *capsnap =
2367 list_last_entry(&ci->i_cap_snaps,
2368 struct ceph_cap_snap,
2369 ci_item);
2370 snapc = ceph_get_snap_context(capsnap->context);
2371 } else {
2372 BUG_ON(!ci->i_head_snapc);
2373 snapc = ceph_get_snap_context(ci->i_head_snapc);
2374 }
2375 spin_unlock(&ci->i_ceph_lock);
2376
2377 /* we might need to revert back to that point */
2378 data = *from;
2379 if ((iocb->ki_flags & IOCB_DIRECT) && !IS_ENCRYPTED(inode))
2380 written = ceph_direct_read_write(iocb, &data, snapc,
2381 &prealloc_cf);
2382 else
2383 written = ceph_sync_write(iocb, &data, pos, snapc);
2384 if (direct_lock)
2385 ceph_end_io_direct(inode);
2386 else
2387 ceph_end_io_write(inode);
2388 if (written > 0)
2389 iov_iter_advance(from, written);
2390 ceph_put_snap_context(snapc);
2391 } else {
2392 /*
2393 * No need to acquire the i_truncate_mutex. Because
2394 * the MDS revokes Fwb caps before sending truncate
2395 * message to us. We can't get Fwb cap while there
2396 * are pending vmtruncate. So write and vmtruncate
2397 * can not run at the same time
2398 */
2399 written = generic_perform_write(iocb, from);
2400 ceph_end_io_write(inode);
2401 }
2402
2403 if (written >= 0) {
2404 int dirty;
2405
2406 spin_lock(&ci->i_ceph_lock);
2407 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
2408 &prealloc_cf);
2409 spin_unlock(&ci->i_ceph_lock);
2410 if (dirty)
2411 __mark_inode_dirty(inode, dirty);
2412 if (ceph_quota_is_max_bytes_approaching(inode, iocb->ki_pos))
2413 ceph_check_caps(ci, CHECK_CAPS_FLUSH);
2414 }
2415
2416 doutc(cl, "%p %llx.%llx %llu~%u dropping cap refs on %s\n",
2417 inode, ceph_vinop(inode), pos, (unsigned)count,
2418 ceph_cap_string(got));
2419 ceph_put_cap_refs(ci, got);
2420
2421 if (written == -EOLDSNAPC) {
2422 doutc(cl, "%p %llx.%llx %llu~%u" "got EOLDSNAPC, retrying\n",
2423 inode, ceph_vinop(inode), pos, (unsigned)count);
2424 goto retry_snap;
2425 }
2426
2427 if (written >= 0) {
2428 if ((map_flags & CEPH_OSDMAP_NEARFULL) ||
2429 (pool_flags & CEPH_POOL_FLAG_NEARFULL))
2430 iocb->ki_flags |= IOCB_DSYNC;
2431 written = generic_write_sync(iocb, written);
2432 }
2433
2434 goto out_unlocked;
2435 out_caps:
2436 ceph_put_cap_refs(ci, got);
2437 out:
2438 if (direct_lock)
2439 ceph_end_io_direct(inode);
2440 else
2441 ceph_end_io_write(inode);
2442 out_unlocked:
2443 ceph_free_cap_flush(prealloc_cf);
2444 return written ? written : err;
2445 }
2446
2447 /*
2448 * llseek. be sure to verify file size on SEEK_END.
2449 */
2450 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
2451 {
2452 if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
2453 struct inode *inode = file_inode(file);
2454 int ret;
2455
2456 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
2457 if (ret < 0)
2458 return ret;
2459 }
2460 return generic_file_llseek(file, offset, whence);
2461 }
2462
2463 static inline void ceph_zero_partial_page(
2464 struct inode *inode, loff_t offset, unsigned size)
2465 {
2466 struct page *page;
2467 pgoff_t index = offset >> PAGE_SHIFT;
2468
2469 page = find_lock_page(inode->i_mapping, index);
2470 if (page) {
2471 wait_on_page_writeback(page);
2472 zero_user(page, offset & (PAGE_SIZE - 1), size);
2473 unlock_page(page);
2474 put_page(page);
2475 }
2476 }
2477
2478 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
2479 loff_t length)
2480 {
2481 loff_t nearly = round_up(offset, PAGE_SIZE);
2482 if (offset < nearly) {
2483 loff_t size = nearly - offset;
2484 if (length < size)
2485 size = length;
2486 ceph_zero_partial_page(inode, offset, size);
2487 offset += size;
2488 length -= size;
2489 }
2490 if (length >= PAGE_SIZE) {
2491 loff_t size = round_down(length, PAGE_SIZE);
2492 truncate_pagecache_range(inode, offset, offset + size - 1);
2493 offset += size;
2494 length -= size;
2495 }
2496 if (length)
2497 ceph_zero_partial_page(inode, offset, length);
2498 }
2499
2500 static int ceph_zero_partial_object(struct inode *inode,
2501 loff_t offset, loff_t *length)
2502 {
2503 struct ceph_inode_info *ci = ceph_inode(inode);
2504 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
2505 struct ceph_osd_request *req;
2506 int ret = 0;
2507 loff_t zero = 0;
2508 int op;
2509
2510 if (ceph_inode_is_shutdown(inode))
2511 return -EIO;
2512
2513 if (!length) {
2514 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
2515 length = &zero;
2516 } else {
2517 op = CEPH_OSD_OP_ZERO;
2518 }
2519
2520 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
2521 ceph_vino(inode),
2522 offset, length,
2523 0, 1, op,
2524 CEPH_OSD_FLAG_WRITE,
2525 NULL, 0, 0, false);
2526 if (IS_ERR(req)) {
2527 ret = PTR_ERR(req);
2528 goto out;
2529 }
2530
2531 req->r_mtime = inode_get_mtime(inode);
2532 ceph_osdc_start_request(&fsc->client->osdc, req);
2533 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
2534 if (ret == -ENOENT)
2535 ret = 0;
2536 ceph_osdc_put_request(req);
2537
2538 out:
2539 return ret;
2540 }
2541
2542 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
2543 {
2544 int ret = 0;
2545 struct ceph_inode_info *ci = ceph_inode(inode);
2546 s32 stripe_unit = ci->i_layout.stripe_unit;
2547 s32 stripe_count = ci->i_layout.stripe_count;
2548 s32 object_size = ci->i_layout.object_size;
2549 u64 object_set_size = object_size * stripe_count;
2550 u64 nearly, t;
2551
2552 /* round offset up to next period boundary */
2553 nearly = offset + object_set_size - 1;
2554 t = nearly;
2555 nearly -= do_div(t, object_set_size);
2556
2557 while (length && offset < nearly) {
2558 loff_t size = length;
2559 ret = ceph_zero_partial_object(inode, offset, &size);
2560 if (ret < 0)
2561 return ret;
2562 offset += size;
2563 length -= size;
2564 }
2565 while (length >= object_set_size) {
2566 int i;
2567 loff_t pos = offset;
2568 for (i = 0; i < stripe_count; ++i) {
2569 ret = ceph_zero_partial_object(inode, pos, NULL);
2570 if (ret < 0)
2571 return ret;
2572 pos += stripe_unit;
2573 }
2574 offset += object_set_size;
2575 length -= object_set_size;
2576 }
2577 while (length) {
2578 loff_t size = length;
2579 ret = ceph_zero_partial_object(inode, offset, &size);
2580 if (ret < 0)
2581 return ret;
2582 offset += size;
2583 length -= size;
2584 }
2585 return ret;
2586 }
2587
2588 static long ceph_fallocate(struct file *file, int mode,
2589 loff_t offset, loff_t length)
2590 {
2591 struct ceph_file_info *fi = file->private_data;
2592 struct inode *inode = file_inode(file);
2593 struct ceph_inode_info *ci = ceph_inode(inode);
2594 struct ceph_cap_flush *prealloc_cf;
2595 struct ceph_client *cl = ceph_inode_to_client(inode);
2596 int want, got = 0;
2597 int dirty;
2598 int ret = 0;
2599 loff_t endoff = 0;
2600 loff_t size;
2601
2602 doutc(cl, "%p %llx.%llx mode %x, offset %llu length %llu\n",
2603 inode, ceph_vinop(inode), mode, offset, length);
2604
2605 if (mode != (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2606 return -EOPNOTSUPP;
2607
2608 if (!S_ISREG(inode->i_mode))
2609 return -EOPNOTSUPP;
2610
2611 if (IS_ENCRYPTED(inode))
2612 return -EOPNOTSUPP;
2613
2614 prealloc_cf = ceph_alloc_cap_flush();
2615 if (!prealloc_cf)
2616 return -ENOMEM;
2617
2618 inode_lock(inode);
2619
2620 if (ceph_snap(inode) != CEPH_NOSNAP) {
2621 ret = -EROFS;
2622 goto unlock;
2623 }
2624
2625 size = i_size_read(inode);
2626
2627 /* Are we punching a hole beyond EOF? */
2628 if (offset >= size)
2629 goto unlock;
2630 if ((offset + length) > size)
2631 length = size - offset;
2632
2633 if (fi->fmode & CEPH_FILE_MODE_LAZY)
2634 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
2635 else
2636 want = CEPH_CAP_FILE_BUFFER;
2637
2638 ret = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, endoff, &got);
2639 if (ret < 0)
2640 goto unlock;
2641
2642 ret = file_modified(file);
2643 if (ret)
2644 goto put_caps;
2645
2646 filemap_invalidate_lock(inode->i_mapping);
2647 ceph_fscache_invalidate(inode, false);
2648 ceph_zero_pagecache_range(inode, offset, length);
2649 ret = ceph_zero_objects(inode, offset, length);
2650
2651 if (!ret) {
2652 spin_lock(&ci->i_ceph_lock);
2653 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
2654 &prealloc_cf);
2655 spin_unlock(&ci->i_ceph_lock);
2656 if (dirty)
2657 __mark_inode_dirty(inode, dirty);
2658 }
2659 filemap_invalidate_unlock(inode->i_mapping);
2660
2661 put_caps:
2662 ceph_put_cap_refs(ci, got);
2663 unlock:
2664 inode_unlock(inode);
2665 ceph_free_cap_flush(prealloc_cf);
2666 return ret;
2667 }
2668
2669 /*
2670 * This function tries to get FILE_WR capabilities for dst_ci and FILE_RD for
2671 * src_ci. Two attempts are made to obtain both caps, and an error is return if
2672 * this fails; zero is returned on success.
2673 */
2674 static int get_rd_wr_caps(struct file *src_filp, int *src_got,
2675 struct file *dst_filp,
2676 loff_t dst_endoff, int *dst_got)
2677 {
2678 int ret = 0;
2679 bool retrying = false;
2680
2681 retry_caps:
2682 ret = ceph_get_caps(dst_filp, CEPH_CAP_FILE_WR, CEPH_CAP_FILE_BUFFER,
2683 dst_endoff, dst_got);
2684 if (ret < 0)
2685 return ret;
2686
2687 /*
2688 * Since we're already holding the FILE_WR capability for the dst file,
2689 * we would risk a deadlock by using ceph_get_caps. Thus, we'll do some
2690 * retry dance instead to try to get both capabilities.
2691 */
2692 ret = ceph_try_get_caps(file_inode(src_filp),
2693 CEPH_CAP_FILE_RD, CEPH_CAP_FILE_SHARED,
2694 false, src_got);
2695 if (ret <= 0) {
2696 /* Start by dropping dst_ci caps and getting src_ci caps */
2697 ceph_put_cap_refs(ceph_inode(file_inode(dst_filp)), *dst_got);
2698 if (retrying) {
2699 if (!ret)
2700 /* ceph_try_get_caps masks EAGAIN */
2701 ret = -EAGAIN;
2702 return ret;
2703 }
2704 ret = ceph_get_caps(src_filp, CEPH_CAP_FILE_RD,
2705 CEPH_CAP_FILE_SHARED, -1, src_got);
2706 if (ret < 0)
2707 return ret;
2708 /*... drop src_ci caps too, and retry */
2709 ceph_put_cap_refs(ceph_inode(file_inode(src_filp)), *src_got);
2710 retrying = true;
2711 goto retry_caps;
2712 }
2713 return ret;
2714 }
2715
2716 static void put_rd_wr_caps(struct ceph_inode_info *src_ci, int src_got,
2717 struct ceph_inode_info *dst_ci, int dst_got)
2718 {
2719 ceph_put_cap_refs(src_ci, src_got);
2720 ceph_put_cap_refs(dst_ci, dst_got);
2721 }
2722
2723 /*
2724 * This function does several size-related checks, returning an error if:
2725 * - source file is smaller than off+len
2726 * - destination file size is not OK (inode_newsize_ok())
2727 * - max bytes quotas is exceeded
2728 */
2729 static int is_file_size_ok(struct inode *src_inode, struct inode *dst_inode,
2730 loff_t src_off, loff_t dst_off, size_t len)
2731 {
2732 struct ceph_client *cl = ceph_inode_to_client(src_inode);
2733 loff_t size, endoff;
2734
2735 size = i_size_read(src_inode);
2736 /*
2737 * Don't copy beyond source file EOF. Instead of simply setting length
2738 * to (size - src_off), just drop to VFS default implementation, as the
2739 * local i_size may be stale due to other clients writing to the source
2740 * inode.
2741 */
2742 if (src_off + len > size) {
2743 doutc(cl, "Copy beyond EOF (%llu + %zu > %llu)\n", src_off,
2744 len, size);
2745 return -EOPNOTSUPP;
2746 }
2747 size = i_size_read(dst_inode);
2748
2749 endoff = dst_off + len;
2750 if (inode_newsize_ok(dst_inode, endoff))
2751 return -EOPNOTSUPP;
2752
2753 if (ceph_quota_is_max_bytes_exceeded(dst_inode, endoff))
2754 return -EDQUOT;
2755
2756 return 0;
2757 }
2758
2759 static struct ceph_osd_request *
2760 ceph_alloc_copyfrom_request(struct ceph_osd_client *osdc,
2761 u64 src_snapid,
2762 struct ceph_object_id *src_oid,
2763 struct ceph_object_locator *src_oloc,
2764 struct ceph_object_id *dst_oid,
2765 struct ceph_object_locator *dst_oloc,
2766 u32 truncate_seq, u64 truncate_size)
2767 {
2768 struct ceph_osd_request *req;
2769 int ret;
2770 u32 src_fadvise_flags =
2771 CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2772 CEPH_OSD_OP_FLAG_FADVISE_NOCACHE;
2773 u32 dst_fadvise_flags =
2774 CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2775 CEPH_OSD_OP_FLAG_FADVISE_DONTNEED;
2776
2777 req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_KERNEL);
2778 if (!req)
2779 return ERR_PTR(-ENOMEM);
2780
2781 req->r_flags = CEPH_OSD_FLAG_WRITE;
2782
2783 ceph_oloc_copy(&req->r_t.base_oloc, dst_oloc);
2784 ceph_oid_copy(&req->r_t.base_oid, dst_oid);
2785
2786 ret = osd_req_op_copy_from_init(req, src_snapid, 0,
2787 src_oid, src_oloc,
2788 src_fadvise_flags,
2789 dst_fadvise_flags,
2790 truncate_seq,
2791 truncate_size,
2792 CEPH_OSD_COPY_FROM_FLAG_TRUNCATE_SEQ);
2793 if (ret)
2794 goto out;
2795
2796 ret = ceph_osdc_alloc_messages(req, GFP_KERNEL);
2797 if (ret)
2798 goto out;
2799
2800 return req;
2801
2802 out:
2803 ceph_osdc_put_request(req);
2804 return ERR_PTR(ret);
2805 }
2806
2807 static ssize_t ceph_do_objects_copy(struct ceph_inode_info *src_ci, u64 *src_off,
2808 struct ceph_inode_info *dst_ci, u64 *dst_off,
2809 struct ceph_fs_client *fsc,
2810 size_t len, unsigned int flags)
2811 {
2812 struct ceph_object_locator src_oloc, dst_oloc;
2813 struct ceph_object_id src_oid, dst_oid;
2814 struct ceph_osd_client *osdc;
2815 struct ceph_osd_request *req;
2816 size_t bytes = 0;
2817 u64 src_objnum, src_objoff, dst_objnum, dst_objoff;
2818 u32 src_objlen, dst_objlen;
2819 u32 object_size = src_ci->i_layout.object_size;
2820 struct ceph_client *cl = fsc->client;
2821 int ret;
2822
2823 src_oloc.pool = src_ci->i_layout.pool_id;
2824 src_oloc.pool_ns = ceph_try_get_string(src_ci->i_layout.pool_ns);
2825 dst_oloc.pool = dst_ci->i_layout.pool_id;
2826 dst_oloc.pool_ns = ceph_try_get_string(dst_ci->i_layout.pool_ns);
2827 osdc = &fsc->client->osdc;
2828
2829 while (len >= object_size) {
2830 ceph_calc_file_object_mapping(&src_ci->i_layout, *src_off,
2831 object_size, &src_objnum,
2832 &src_objoff, &src_objlen);
2833 ceph_calc_file_object_mapping(&dst_ci->i_layout, *dst_off,
2834 object_size, &dst_objnum,
2835 &dst_objoff, &dst_objlen);
2836 ceph_oid_init(&src_oid);
2837 ceph_oid_printf(&src_oid, "%llx.%08llx",
2838 src_ci->i_vino.ino, src_objnum);
2839 ceph_oid_init(&dst_oid);
2840 ceph_oid_printf(&dst_oid, "%llx.%08llx",
2841 dst_ci->i_vino.ino, dst_objnum);
2842 /* Do an object remote copy */
2843 req = ceph_alloc_copyfrom_request(osdc, src_ci->i_vino.snap,
2844 &src_oid, &src_oloc,
2845 &dst_oid, &dst_oloc,
2846 dst_ci->i_truncate_seq,
2847 dst_ci->i_truncate_size);
2848 if (IS_ERR(req))
2849 ret = PTR_ERR(req);
2850 else {
2851 ceph_osdc_start_request(osdc, req);
2852 ret = ceph_osdc_wait_request(osdc, req);
2853 ceph_update_copyfrom_metrics(&fsc->mdsc->metric,
2854 req->r_start_latency,
2855 req->r_end_latency,
2856 object_size, ret);
2857 ceph_osdc_put_request(req);
2858 }
2859 if (ret) {
2860 if (ret == -EOPNOTSUPP) {
2861 fsc->have_copy_from2 = false;
2862 pr_notice_client(cl,
2863 "OSDs don't support copy-from2; disabling copy offload\n");
2864 }
2865 doutc(cl, "returned %d\n", ret);
2866 if (!bytes)
2867 bytes = ret;
2868 goto out;
2869 }
2870 len -= object_size;
2871 bytes += object_size;
2872 *src_off += object_size;
2873 *dst_off += object_size;
2874 }
2875
2876 out:
2877 ceph_oloc_destroy(&src_oloc);
2878 ceph_oloc_destroy(&dst_oloc);
2879 return bytes;
2880 }
2881
2882 static ssize_t __ceph_copy_file_range(struct file *src_file, loff_t src_off,
2883 struct file *dst_file, loff_t dst_off,
2884 size_t len, unsigned int flags)
2885 {
2886 struct inode *src_inode = file_inode(src_file);
2887 struct inode *dst_inode = file_inode(dst_file);
2888 struct ceph_inode_info *src_ci = ceph_inode(src_inode);
2889 struct ceph_inode_info *dst_ci = ceph_inode(dst_inode);
2890 struct ceph_cap_flush *prealloc_cf;
2891 struct ceph_fs_client *src_fsc = ceph_inode_to_fs_client(src_inode);
2892 struct ceph_client *cl = src_fsc->client;
2893 loff_t size;
2894 ssize_t ret = -EIO, bytes;
2895 u64 src_objnum, dst_objnum, src_objoff, dst_objoff;
2896 u32 src_objlen, dst_objlen;
2897 int src_got = 0, dst_got = 0, err, dirty;
2898
2899 if (src_inode->i_sb != dst_inode->i_sb) {
2900 struct ceph_fs_client *dst_fsc = ceph_inode_to_fs_client(dst_inode);
2901
2902 if (ceph_fsid_compare(&src_fsc->client->fsid,
2903 &dst_fsc->client->fsid)) {
2904 dout("Copying files across clusters: src: %pU dst: %pU\n",
2905 &src_fsc->client->fsid, &dst_fsc->client->fsid);
2906 return -EXDEV;
2907 }
2908 }
2909 if (ceph_snap(dst_inode) != CEPH_NOSNAP)
2910 return -EROFS;
2911
2912 /*
2913 * Some of the checks below will return -EOPNOTSUPP, which will force a
2914 * fallback to the default VFS copy_file_range implementation. This is
2915 * desirable in several cases (for ex, the 'len' is smaller than the
2916 * size of the objects, or in cases where that would be more
2917 * efficient).
2918 */
2919
2920 if (ceph_test_mount_opt(src_fsc, NOCOPYFROM))
2921 return -EOPNOTSUPP;
2922
2923 if (!src_fsc->have_copy_from2)
2924 return -EOPNOTSUPP;
2925
2926 /*
2927 * Striped file layouts require that we copy partial objects, but the
2928 * OSD copy-from operation only supports full-object copies. Limit
2929 * this to non-striped file layouts for now.
2930 */
2931 if ((src_ci->i_layout.stripe_unit != dst_ci->i_layout.stripe_unit) ||
2932 (src_ci->i_layout.stripe_count != 1) ||
2933 (dst_ci->i_layout.stripe_count != 1) ||
2934 (src_ci->i_layout.object_size != dst_ci->i_layout.object_size)) {
2935 doutc(cl, "Invalid src/dst files layout\n");
2936 return -EOPNOTSUPP;
2937 }
2938
2939 /* Every encrypted inode gets its own key, so we can't offload them */
2940 if (IS_ENCRYPTED(src_inode) || IS_ENCRYPTED(dst_inode))
2941 return -EOPNOTSUPP;
2942
2943 if (len < src_ci->i_layout.object_size)
2944 return -EOPNOTSUPP; /* no remote copy will be done */
2945
2946 prealloc_cf = ceph_alloc_cap_flush();
2947 if (!prealloc_cf)
2948 return -ENOMEM;
2949
2950 /* Start by sync'ing the source and destination files */
2951 ret = file_write_and_wait_range(src_file, src_off, (src_off + len));
2952 if (ret < 0) {
2953 doutc(cl, "failed to write src file (%zd)\n", ret);
2954 goto out;
2955 }
2956 ret = file_write_and_wait_range(dst_file, dst_off, (dst_off + len));
2957 if (ret < 0) {
2958 doutc(cl, "failed to write dst file (%zd)\n", ret);
2959 goto out;
2960 }
2961
2962 /*
2963 * We need FILE_WR caps for dst_ci and FILE_RD for src_ci as other
2964 * clients may have dirty data in their caches. And OSDs know nothing
2965 * about caps, so they can't safely do the remote object copies.
2966 */
2967 err = get_rd_wr_caps(src_file, &src_got,
2968 dst_file, (dst_off + len), &dst_got);
2969 if (err < 0) {
2970 doutc(cl, "get_rd_wr_caps returned %d\n", err);
2971 ret = -EOPNOTSUPP;
2972 goto out;
2973 }
2974
2975 ret = is_file_size_ok(src_inode, dst_inode, src_off, dst_off, len);
2976 if (ret < 0)
2977 goto out_caps;
2978
2979 /* Drop dst file cached pages */
2980 ceph_fscache_invalidate(dst_inode, false);
2981 ret = invalidate_inode_pages2_range(dst_inode->i_mapping,
2982 dst_off >> PAGE_SHIFT,
2983 (dst_off + len) >> PAGE_SHIFT);
2984 if (ret < 0) {
2985 doutc(cl, "Failed to invalidate inode pages (%zd)\n",
2986 ret);
2987 ret = 0; /* XXX */
2988 }
2989 ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
2990 src_ci->i_layout.object_size,
2991 &src_objnum, &src_objoff, &src_objlen);
2992 ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
2993 dst_ci->i_layout.object_size,
2994 &dst_objnum, &dst_objoff, &dst_objlen);
2995 /* object-level offsets need to the same */
2996 if (src_objoff != dst_objoff) {
2997 ret = -EOPNOTSUPP;
2998 goto out_caps;
2999 }
3000
3001 /*
3002 * Do a manual copy if the object offset isn't object aligned.
3003 * 'src_objlen' contains the bytes left until the end of the object,
3004 * starting at the src_off
3005 */
3006 if (src_objoff) {
3007 doutc(cl, "Initial partial copy of %u bytes\n", src_objlen);
3008
3009 /*
3010 * we need to temporarily drop all caps as we'll be calling
3011 * {read,write}_iter, which will get caps again.
3012 */
3013 put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
3014 ret = splice_file_range(src_file, &src_off, dst_file, &dst_off,
3015 src_objlen);
3016 /* Abort on short copies or on error */
3017 if (ret < (long)src_objlen) {
3018 doutc(cl, "Failed partial copy (%zd)\n", ret);
3019 goto out;
3020 }
3021 len -= ret;
3022 err = get_rd_wr_caps(src_file, &src_got,
3023 dst_file, (dst_off + len), &dst_got);
3024 if (err < 0)
3025 goto out;
3026 err = is_file_size_ok(src_inode, dst_inode,
3027 src_off, dst_off, len);
3028 if (err < 0)
3029 goto out_caps;
3030 }
3031
3032 size = i_size_read(dst_inode);
3033 bytes = ceph_do_objects_copy(src_ci, &src_off, dst_ci, &dst_off,
3034 src_fsc, len, flags);
3035 if (bytes <= 0) {
3036 if (!ret)
3037 ret = bytes;
3038 goto out_caps;
3039 }
3040 doutc(cl, "Copied %zu bytes out of %zu\n", bytes, len);
3041 len -= bytes;
3042 ret += bytes;
3043
3044 file_update_time(dst_file);
3045 inode_inc_iversion_raw(dst_inode);
3046
3047 if (dst_off > size) {
3048 /* Let the MDS know about dst file size change */
3049 if (ceph_inode_set_size(dst_inode, dst_off) ||
3050 ceph_quota_is_max_bytes_approaching(dst_inode, dst_off))
3051 ceph_check_caps(dst_ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_FLUSH);
3052 }
3053 /* Mark Fw dirty */
3054 spin_lock(&dst_ci->i_ceph_lock);
3055 dirty = __ceph_mark_dirty_caps(dst_ci, CEPH_CAP_FILE_WR, &prealloc_cf);
3056 spin_unlock(&dst_ci->i_ceph_lock);
3057 if (dirty)
3058 __mark_inode_dirty(dst_inode, dirty);
3059
3060 out_caps:
3061 put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
3062
3063 /*
3064 * Do the final manual copy if we still have some bytes left, unless
3065 * there were errors in remote object copies (len >= object_size).
3066 */
3067 if (len && (len < src_ci->i_layout.object_size)) {
3068 doutc(cl, "Final partial copy of %zu bytes\n", len);
3069 bytes = splice_file_range(src_file, &src_off, dst_file,
3070 &dst_off, len);
3071 if (bytes > 0)
3072 ret += bytes;
3073 else
3074 doutc(cl, "Failed partial copy (%zd)\n", bytes);
3075 }
3076
3077 out:
3078 ceph_free_cap_flush(prealloc_cf);
3079
3080 return ret;
3081 }
3082
3083 static ssize_t ceph_copy_file_range(struct file *src_file, loff_t src_off,
3084 struct file *dst_file, loff_t dst_off,
3085 size_t len, unsigned int flags)
3086 {
3087 ssize_t ret;
3088
3089 ret = __ceph_copy_file_range(src_file, src_off, dst_file, dst_off,
3090 len, flags);
3091
3092 if (ret == -EOPNOTSUPP || ret == -EXDEV)
3093 ret = splice_copy_file_range(src_file, src_off, dst_file,
3094 dst_off, len);
3095 return ret;
3096 }
3097
3098 const struct file_operations ceph_file_fops = {
3099 .open = ceph_open,
3100 .release = ceph_release,
3101 .llseek = ceph_llseek,
3102 .read_iter = ceph_read_iter,
3103 .write_iter = ceph_write_iter,
3104 .mmap = ceph_mmap,
3105 .fsync = ceph_fsync,
3106 .lock = ceph_lock,
3107 .setlease = simple_nosetlease,
3108 .flock = ceph_flock,
3109 .splice_read = ceph_splice_read,
3110 .splice_write = iter_file_splice_write,
3111 .unlocked_ioctl = ceph_ioctl,
3112 .compat_ioctl = compat_ptr_ioctl,
3113 .fallocate = ceph_fallocate,
3114 .copy_file_range = ceph_copy_file_range,
3115 };
A mirror of Linus' kernel repository