?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
cpuset: restore sanity to cpuset_cpus_allowed_fallback()
[linux/fpc-iii.git] / fs / ceph / file.c
blob92ab2043368291e68c5c4b67b2da7e6bfdd25fcf
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
3
4 #include <linux/module.h>
5 #include <linux/sched.h>
6 #include <linux/slab.h>
7 #include <linux/file.h>
8 #include <linux/mount.h>
9 #include <linux/namei.h>
10 #include <linux/writeback.h>
11 #include <linux/falloc.h>
12
13 #include "super.h"
14 #include "mds_client.h"
15 #include "cache.h"
16
17 static __le32 ceph_flags_sys2wire(u32 flags)
18 {
19 u32 wire_flags = 0;
20
21 switch (flags & O_ACCMODE) {
22 case O_RDONLY:
23 wire_flags |= CEPH_O_RDONLY;
24 break;
25 case O_WRONLY:
26 wire_flags |= CEPH_O_WRONLY;
27 break;
28 case O_RDWR:
29 wire_flags |= CEPH_O_RDWR;
30 break;
31 }
32
33 flags &= ~O_ACCMODE;
34
35 #define ceph_sys2wire(a) if (flags & a) { wire_flags |= CEPH_##a; flags &= ~a; }
36
37 ceph_sys2wire(O_CREAT);
38 ceph_sys2wire(O_EXCL);
39 ceph_sys2wire(O_TRUNC);
40 ceph_sys2wire(O_DIRECTORY);
41 ceph_sys2wire(O_NOFOLLOW);
42
43 #undef ceph_sys2wire
44
45 if (flags)
46 dout("unused open flags: %x\n", flags);
47
48 return cpu_to_le32(wire_flags);
49 }
50
51 /*
52 * Ceph file operations
53 *
54 * Implement basic open/close functionality, and implement
55 * read/write.
56 *
57 * We implement three modes of file I/O:
58 * - buffered uses the generic_file_aio_{read,write} helpers
59 *
60 * - synchronous is used when there is multi-client read/write
61 * sharing, avoids the page cache, and synchronously waits for an
62 * ack from the OSD.
63 *
64 * - direct io takes the variant of the sync path that references
65 * user pages directly.
66 *
67 * fsync() flushes and waits on dirty pages, but just queues metadata
68 * for writeback: since the MDS can recover size and mtime there is no
69 * need to wait for MDS acknowledgement.
70 */
71
72 /*
73 * How many pages to get in one call to iov_iter_get_pages(). This
74 * determines the size of the on-stack array used as a buffer.
75 */
76 #define ITER_GET_BVECS_PAGES 64
77
78 static ssize_t __iter_get_bvecs(struct iov_iter *iter, size_t maxsize,
79 struct bio_vec *bvecs)
80 {
81 size_t size = 0;
82 int bvec_idx = 0;
83
84 if (maxsize > iov_iter_count(iter))
85 maxsize = iov_iter_count(iter);
86
87 while (size < maxsize) {
88 struct page *pages[ITER_GET_BVECS_PAGES];
89 ssize_t bytes;
90 size_t start;
91 int idx = 0;
92
93 bytes = iov_iter_get_pages(iter, pages, maxsize - size,
94 ITER_GET_BVECS_PAGES, &start);
95 if (bytes < 0)
96 return size ?: bytes;
97
98 iov_iter_advance(iter, bytes);
99 size += bytes;
100
101 for ( ; bytes; idx++, bvec_idx++) {
102 struct bio_vec bv = {
103 .bv_page = pages[idx],
104 .bv_len = min_t(int, bytes, PAGE_SIZE - start),
105 .bv_offset = start,
106 };
107
108 bvecs[bvec_idx] = bv;
109 bytes -= bv.bv_len;
110 start = 0;
111 }
112 }
113
114 return size;
115 }
116
117 /*
118 * iov_iter_get_pages() only considers one iov_iter segment, no matter
119 * what maxsize or maxpages are given. For ITER_BVEC that is a single
120 * page.
121 *
122 * Attempt to get up to @maxsize bytes worth of pages from @iter.
123 * Return the number of bytes in the created bio_vec array, or an error.
124 */
125 static ssize_t iter_get_bvecs_alloc(struct iov_iter *iter, size_t maxsize,
126 struct bio_vec **bvecs, int *num_bvecs)
127 {
128 struct bio_vec *bv;
129 size_t orig_count = iov_iter_count(iter);
130 ssize_t bytes;
131 int npages;
132
133 iov_iter_truncate(iter, maxsize);
134 npages = iov_iter_npages(iter, INT_MAX);
135 iov_iter_reexpand(iter, orig_count);
136
137 /*
138 * __iter_get_bvecs() may populate only part of the array -- zero it
139 * out.
140 */
141 bv = kvmalloc_array(npages, sizeof(*bv), GFP_KERNEL | __GFP_ZERO);
142 if (!bv)
143 return -ENOMEM;
144
145 bytes = __iter_get_bvecs(iter, maxsize, bv);
146 if (bytes < 0) {
147 /*
148 * No pages were pinned -- just free the array.
149 */
150 kvfree(bv);
151 return bytes;
152 }
153
154 *bvecs = bv;
155 *num_bvecs = npages;
156 return bytes;
157 }
158
159 static void put_bvecs(struct bio_vec *bvecs, int num_bvecs, bool should_dirty)
160 {
161 int i;
162
163 for (i = 0; i < num_bvecs; i++) {
164 if (bvecs[i].bv_page) {
165 if (should_dirty)
166 set_page_dirty_lock(bvecs[i].bv_page);
167 put_page(bvecs[i].bv_page);
168 }
169 }
170 kvfree(bvecs);
171 }
172
173 /*
174 * Prepare an open request. Preallocate ceph_cap to avoid an
175 * inopportune ENOMEM later.
176 */
177 static struct ceph_mds_request *
178 prepare_open_request(struct super_block *sb, int flags, int create_mode)
179 {
180 struct ceph_fs_client *fsc = ceph_sb_to_client(sb);
181 struct ceph_mds_client *mdsc = fsc->mdsc;
182 struct ceph_mds_request *req;
183 int want_auth = USE_ANY_MDS;
184 int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
185
186 if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
187 want_auth = USE_AUTH_MDS;
188
189 req = ceph_mdsc_create_request(mdsc, op, want_auth);
190 if (IS_ERR(req))
191 goto out;
192 req->r_fmode = ceph_flags_to_mode(flags);
193 req->r_args.open.flags = ceph_flags_sys2wire(flags);
194 req->r_args.open.mode = cpu_to_le32(create_mode);
195 out:
196 return req;
197 }
198
199 static int ceph_init_file_info(struct inode *inode, struct file *file,
200 int fmode, bool isdir)
201 {
202 struct ceph_file_info *fi;
203
204 dout("%s %p %p 0%o (%s)\n", __func__, inode, file,
205 inode->i_mode, isdir ? "dir" : "regular");
206 BUG_ON(inode->i_fop->release != ceph_release);
207
208 if (isdir) {
209 struct ceph_dir_file_info *dfi =
210 kmem_cache_zalloc(ceph_dir_file_cachep, GFP_KERNEL);
211 if (!dfi) {
212 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
213 return -ENOMEM;
214 }
215
216 file->private_data = dfi;
217 fi = &dfi->file_info;
218 dfi->next_offset = 2;
219 dfi->readdir_cache_idx = -1;
220 } else {
221 fi = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
222 if (!fi) {
223 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
224 return -ENOMEM;
225 }
226
227 file->private_data = fi;
228 }
229
230 fi->fmode = fmode;
231 spin_lock_init(&fi->rw_contexts_lock);
232 INIT_LIST_HEAD(&fi->rw_contexts);
233
234 return 0;
235 }
236
237 /*
238 * initialize private struct file data.
239 * if we fail, clean up by dropping fmode reference on the ceph_inode
240 */
241 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
242 {
243 int ret = 0;
244
245 switch (inode->i_mode & S_IFMT) {
246 case S_IFREG:
247 ceph_fscache_register_inode_cookie(inode);
248 ceph_fscache_file_set_cookie(inode, file);
249 case S_IFDIR:
250 ret = ceph_init_file_info(inode, file, fmode,
251 S_ISDIR(inode->i_mode));
252 if (ret)
253 return ret;
254 break;
255
256 case S_IFLNK:
257 dout("init_file %p %p 0%o (symlink)\n", inode, file,
258 inode->i_mode);
259 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
260 break;
261
262 default:
263 dout("init_file %p %p 0%o (special)\n", inode, file,
264 inode->i_mode);
265 /*
266 * we need to drop the open ref now, since we don't
267 * have .release set to ceph_release.
268 */
269 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
270 BUG_ON(inode->i_fop->release == ceph_release);
271
272 /* call the proper open fop */
273 ret = inode->i_fop->open(inode, file);
274 }
275 return ret;
276 }
277
278 /*
279 * try renew caps after session gets killed.
280 */
281 int ceph_renew_caps(struct inode *inode)
282 {
283 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
284 struct ceph_inode_info *ci = ceph_inode(inode);
285 struct ceph_mds_request *req;
286 int err, flags, wanted;
287
288 spin_lock(&ci->i_ceph_lock);
289 wanted = __ceph_caps_file_wanted(ci);
290 if (__ceph_is_any_real_caps(ci) &&
291 (!(wanted & CEPH_CAP_ANY_WR) || ci->i_auth_cap)) {
292 int issued = __ceph_caps_issued(ci, NULL);
293 spin_unlock(&ci->i_ceph_lock);
294 dout("renew caps %p want %s issued %s updating mds_wanted\n",
295 inode, ceph_cap_string(wanted), ceph_cap_string(issued));
296 ceph_check_caps(ci, 0, NULL);
297 return 0;
298 }
299 spin_unlock(&ci->i_ceph_lock);
300
301 flags = 0;
302 if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
303 flags = O_RDWR;
304 else if (wanted & CEPH_CAP_FILE_RD)
305 flags = O_RDONLY;
306 else if (wanted & CEPH_CAP_FILE_WR)
307 flags = O_WRONLY;
308 #ifdef O_LAZY
309 if (wanted & CEPH_CAP_FILE_LAZYIO)
310 flags |= O_LAZY;
311 #endif
312
313 req = prepare_open_request(inode->i_sb, flags, 0);
314 if (IS_ERR(req)) {
315 err = PTR_ERR(req);
316 goto out;
317 }
318
319 req->r_inode = inode;
320 ihold(inode);
321 req->r_num_caps = 1;
322 req->r_fmode = -1;
323
324 err = ceph_mdsc_do_request(mdsc, NULL, req);
325 ceph_mdsc_put_request(req);
326 out:
327 dout("renew caps %p open result=%d\n", inode, err);
328 return err < 0 ? err : 0;
329 }
330
331 /*
332 * If we already have the requisite capabilities, we can satisfy
333 * the open request locally (no need to request new caps from the
334 * MDS). We do, however, need to inform the MDS (asynchronously)
335 * if our wanted caps set expands.
336 */
337 int ceph_open(struct inode *inode, struct file *file)
338 {
339 struct ceph_inode_info *ci = ceph_inode(inode);
340 struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
341 struct ceph_mds_client *mdsc = fsc->mdsc;
342 struct ceph_mds_request *req;
343 struct ceph_file_info *fi = file->private_data;
344 int err;
345 int flags, fmode, wanted;
346
347 if (fi) {
348 dout("open file %p is already opened\n", file);
349 return 0;
350 }
351
352 /* filter out O_CREAT|O_EXCL; vfs did that already. yuck. */
353 flags = file->f_flags & ~(O_CREAT|O_EXCL);
354 if (S_ISDIR(inode->i_mode))
355 flags = O_DIRECTORY; /* mds likes to know */
356
357 dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
358 ceph_vinop(inode), file, flags, file->f_flags);
359 fmode = ceph_flags_to_mode(flags);
360 wanted = ceph_caps_for_mode(fmode);
361
362 /* snapped files are read-only */
363 if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
364 return -EROFS;
365
366 /* trivially open snapdir */
367 if (ceph_snap(inode) == CEPH_SNAPDIR) {
368 spin_lock(&ci->i_ceph_lock);
369 __ceph_get_fmode(ci, fmode);
370 spin_unlock(&ci->i_ceph_lock);
371 return ceph_init_file(inode, file, fmode);
372 }
373
374 /*
375 * No need to block if we have caps on the auth MDS (for
376 * write) or any MDS (for read). Update wanted set
377 * asynchronously.
378 */
379 spin_lock(&ci->i_ceph_lock);
380 if (__ceph_is_any_real_caps(ci) &&
381 (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
382 int mds_wanted = __ceph_caps_mds_wanted(ci, true);
383 int issued = __ceph_caps_issued(ci, NULL);
384
385 dout("open %p fmode %d want %s issued %s using existing\n",
386 inode, fmode, ceph_cap_string(wanted),
387 ceph_cap_string(issued));
388 __ceph_get_fmode(ci, fmode);
389 spin_unlock(&ci->i_ceph_lock);
390
391 /* adjust wanted? */
392 if ((issued & wanted) != wanted &&
393 (mds_wanted & wanted) != wanted &&
394 ceph_snap(inode) != CEPH_SNAPDIR)
395 ceph_check_caps(ci, 0, NULL);
396
397 return ceph_init_file(inode, file, fmode);
398 } else if (ceph_snap(inode) != CEPH_NOSNAP &&
399 (ci->i_snap_caps & wanted) == wanted) {
400 __ceph_get_fmode(ci, fmode);
401 spin_unlock(&ci->i_ceph_lock);
402 return ceph_init_file(inode, file, fmode);
403 }
404
405 spin_unlock(&ci->i_ceph_lock);
406
407 dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
408 req = prepare_open_request(inode->i_sb, flags, 0);
409 if (IS_ERR(req)) {
410 err = PTR_ERR(req);
411 goto out;
412 }
413 req->r_inode = inode;
414 ihold(inode);
415
416 req->r_num_caps = 1;
417 err = ceph_mdsc_do_request(mdsc, NULL, req);
418 if (!err)
419 err = ceph_init_file(inode, file, req->r_fmode);
420 ceph_mdsc_put_request(req);
421 dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
422 out:
423 return err;
424 }
425
426
427 /*
428 * Do a lookup + open with a single request. If we get a non-existent
429 * file or symlink, return 1 so the VFS can retry.
430 */
431 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
432 struct file *file, unsigned flags, umode_t mode)
433 {
434 struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
435 struct ceph_mds_client *mdsc = fsc->mdsc;
436 struct ceph_mds_request *req;
437 struct dentry *dn;
438 struct ceph_acls_info acls = {};
439 int mask;
440 int err;
441
442 dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
443 dir, dentry, dentry,
444 d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
445
446 if (dentry->d_name.len > NAME_MAX)
447 return -ENAMETOOLONG;
448
449 if (flags & O_CREAT) {
450 if (ceph_quota_is_max_files_exceeded(dir))
451 return -EDQUOT;
452 err = ceph_pre_init_acls(dir, &mode, &acls);
453 if (err < 0)
454 return err;
455 }
456
457 /* do the open */
458 req = prepare_open_request(dir->i_sb, flags, mode);
459 if (IS_ERR(req)) {
460 err = PTR_ERR(req);
461 goto out_acl;
462 }
463 req->r_dentry = dget(dentry);
464 req->r_num_caps = 2;
465 if (flags & O_CREAT) {
466 req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
467 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
468 if (acls.pagelist) {
469 req->r_pagelist = acls.pagelist;
470 acls.pagelist = NULL;
471 }
472 }
473
474 mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
475 if (ceph_security_xattr_wanted(dir))
476 mask |= CEPH_CAP_XATTR_SHARED;
477 req->r_args.open.mask = cpu_to_le32(mask);
478
479 req->r_parent = dir;
480 set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
481 err = ceph_mdsc_do_request(mdsc,
482 (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
483 req);
484 err = ceph_handle_snapdir(req, dentry, err);
485 if (err)
486 goto out_req;
487
488 if ((flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
489 err = ceph_handle_notrace_create(dir, dentry);
490
491 if (d_in_lookup(dentry)) {
492 dn = ceph_finish_lookup(req, dentry, err);
493 if (IS_ERR(dn))
494 err = PTR_ERR(dn);
495 } else {
496 /* we were given a hashed negative dentry */
497 dn = NULL;
498 }
499 if (err)
500 goto out_req;
501 if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
502 /* make vfs retry on splice, ENOENT, or symlink */
503 dout("atomic_open finish_no_open on dn %p\n", dn);
504 err = finish_no_open(file, dn);
505 } else {
506 dout("atomic_open finish_open on dn %p\n", dn);
507 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
508 ceph_init_inode_acls(d_inode(dentry), &acls);
509 file->f_mode |= FMODE_CREATED;
510 }
511 err = finish_open(file, dentry, ceph_open);
512 }
513 out_req:
514 if (!req->r_err && req->r_target_inode)
515 ceph_put_fmode(ceph_inode(req->r_target_inode), req->r_fmode);
516 ceph_mdsc_put_request(req);
517 out_acl:
518 ceph_release_acls_info(&acls);
519 dout("atomic_open result=%d\n", err);
520 return err;
521 }
522
523 int ceph_release(struct inode *inode, struct file *file)
524 {
525 struct ceph_inode_info *ci = ceph_inode(inode);
526
527 if (S_ISDIR(inode->i_mode)) {
528 struct ceph_dir_file_info *dfi = file->private_data;
529 dout("release inode %p dir file %p\n", inode, file);
530 WARN_ON(!list_empty(&dfi->file_info.rw_contexts));
531
532 ceph_put_fmode(ci, dfi->file_info.fmode);
533
534 if (dfi->last_readdir)
535 ceph_mdsc_put_request(dfi->last_readdir);
536 kfree(dfi->last_name);
537 kfree(dfi->dir_info);
538 kmem_cache_free(ceph_dir_file_cachep, dfi);
539 } else {
540 struct ceph_file_info *fi = file->private_data;
541 dout("release inode %p regular file %p\n", inode, file);
542 WARN_ON(!list_empty(&fi->rw_contexts));
543
544 ceph_put_fmode(ci, fi->fmode);
545 kmem_cache_free(ceph_file_cachep, fi);
546 }
547
548 /* wake up anyone waiting for caps on this inode */
549 wake_up_all(&ci->i_cap_wq);
550 return 0;
551 }
552
553 enum {
554 HAVE_RETRIED = 1,
555 CHECK_EOF = 2,
556 READ_INLINE = 3,
557 };
558
559 /*
560 * Read a range of bytes striped over one or more objects. Iterate over
561 * objects we stripe over. (That's not atomic, but good enough for now.)
562 *
563 * If we get a short result from the OSD, check against i_size; we need to
564 * only return a short read to the caller if we hit EOF.
565 */
566 static int striped_read(struct inode *inode,
567 u64 pos, u64 len,
568 struct page **pages, int num_pages,
569 int page_align, int *checkeof)
570 {
571 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
572 struct ceph_inode_info *ci = ceph_inode(inode);
573 u64 this_len;
574 loff_t i_size;
575 int page_idx;
576 int ret, read = 0;
577 bool hit_stripe, was_short;
578
579 /*
580 * we may need to do multiple reads. not atomic, unfortunately.
581 */
582 more:
583 this_len = len;
584 page_idx = (page_align + read) >> PAGE_SHIFT;
585 ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode),
586 &ci->i_layout, pos, &this_len,
587 ci->i_truncate_seq, ci->i_truncate_size,
588 pages + page_idx, num_pages - page_idx,
589 ((page_align + read) & ~PAGE_MASK));
590 if (ret == -ENOENT)
591 ret = 0;
592 hit_stripe = this_len < len;
593 was_short = ret >= 0 && ret < this_len;
594 dout("striped_read %llu~%llu (read %u) got %d%s%s\n", pos, len, read,
595 ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : "");
596
597 i_size = i_size_read(inode);
598 if (ret >= 0) {
599 if (was_short && (pos + ret < i_size)) {
600 int zlen = min(this_len - ret, i_size - pos - ret);
601 int zoff = page_align + read + ret;
602 dout(" zero gap %llu to %llu\n",
603 pos + ret, pos + ret + zlen);
604 ceph_zero_page_vector_range(zoff, zlen, pages);
605 ret += zlen;
606 }
607
608 read += ret;
609 pos += ret;
610 len -= ret;
611
612 /* hit stripe and need continue*/
613 if (len && hit_stripe && pos < i_size)
614 goto more;
615 }
616
617 if (read > 0) {
618 ret = read;
619 /* did we bounce off eof? */
620 if (pos + len > i_size)
621 *checkeof = CHECK_EOF;
622 }
623
624 dout("striped_read returns %d\n", ret);
625 return ret;
626 }
627
628 /*
629 * Completely synchronous read and write methods. Direct from __user
630 * buffer to osd, or directly to user pages (if O_DIRECT).
631 *
632 * If the read spans object boundary, just do multiple reads.
633 */
634 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
635 int *checkeof)
636 {
637 struct file *file = iocb->ki_filp;
638 struct inode *inode = file_inode(file);
639 struct page **pages;
640 u64 off = iocb->ki_pos;
641 int num_pages;
642 ssize_t ret;
643 size_t len = iov_iter_count(to);
644
645 dout("sync_read on file %p %llu~%u %s\n", file, off, (unsigned)len,
646 (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
647
648 if (!len)
649 return 0;
650 /*
651 * flush any page cache pages in this range. this
652 * will make concurrent normal and sync io slow,
653 * but it will at least behave sensibly when they are
654 * in sequence.
655 */
656 ret = filemap_write_and_wait_range(inode->i_mapping, off,
657 off + len);
658 if (ret < 0)
659 return ret;
660
661 if (unlikely(to->type & ITER_PIPE)) {
662 size_t page_off;
663 ret = iov_iter_get_pages_alloc(to, &pages, len,
664 &page_off);
665 if (ret <= 0)
666 return -ENOMEM;
667 num_pages = DIV_ROUND_UP(ret + page_off, PAGE_SIZE);
668
669 ret = striped_read(inode, off, ret, pages, num_pages,
670 page_off, checkeof);
671 if (ret > 0) {
672 iov_iter_advance(to, ret);
673 off += ret;
674 } else {
675 iov_iter_advance(to, 0);
676 }
677 ceph_put_page_vector(pages, num_pages, false);
678 } else {
679 num_pages = calc_pages_for(off, len);
680 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
681 if (IS_ERR(pages))
682 return PTR_ERR(pages);
683
684 ret = striped_read(inode, off, len, pages, num_pages,
685 (off & ~PAGE_MASK), checkeof);
686 if (ret > 0) {
687 int l, k = 0;
688 size_t left = ret;
689
690 while (left) {
691 size_t page_off = off & ~PAGE_MASK;
692 size_t copy = min_t(size_t, left,
693 PAGE_SIZE - page_off);
694 l = copy_page_to_iter(pages[k++], page_off,
695 copy, to);
696 off += l;
697 left -= l;
698 if (l < copy)
699 break;
700 }
701 }
702 ceph_release_page_vector(pages, num_pages);
703 }
704
705 if (off > iocb->ki_pos) {
706 ret = off - iocb->ki_pos;
707 iocb->ki_pos = off;
708 }
709
710 dout("sync_read result %zd\n", ret);
711 return ret;
712 }
713
714 struct ceph_aio_request {
715 struct kiocb *iocb;
716 size_t total_len;
717 bool write;
718 bool should_dirty;
719 int error;
720 struct list_head osd_reqs;
721 unsigned num_reqs;
722 atomic_t pending_reqs;
723 struct timespec64 mtime;
724 struct ceph_cap_flush *prealloc_cf;
725 };
726
727 struct ceph_aio_work {
728 struct work_struct work;
729 struct ceph_osd_request *req;
730 };
731
732 static void ceph_aio_retry_work(struct work_struct *work);
733
734 static void ceph_aio_complete(struct inode *inode,
735 struct ceph_aio_request *aio_req)
736 {
737 struct ceph_inode_info *ci = ceph_inode(inode);
738 int ret;
739
740 if (!atomic_dec_and_test(&aio_req->pending_reqs))
741 return;
742
743 ret = aio_req->error;
744 if (!ret)
745 ret = aio_req->total_len;
746
747 dout("ceph_aio_complete %p rc %d\n", inode, ret);
748
749 if (ret >= 0 && aio_req->write) {
750 int dirty;
751
752 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
753 if (endoff > i_size_read(inode)) {
754 if (ceph_inode_set_size(inode, endoff))
755 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
756 }
757
758 spin_lock(&ci->i_ceph_lock);
759 ci->i_inline_version = CEPH_INLINE_NONE;
760 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
761 &aio_req->prealloc_cf);
762 spin_unlock(&ci->i_ceph_lock);
763 if (dirty)
764 __mark_inode_dirty(inode, dirty);
765
766 }
767
768 ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
769 CEPH_CAP_FILE_RD));
770
771 aio_req->iocb->ki_complete(aio_req->iocb, ret, 0);
772
773 ceph_free_cap_flush(aio_req->prealloc_cf);
774 kfree(aio_req);
775 }
776
777 static void ceph_aio_complete_req(struct ceph_osd_request *req)
778 {
779 int rc = req->r_result;
780 struct inode *inode = req->r_inode;
781 struct ceph_aio_request *aio_req = req->r_priv;
782 struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
783
784 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_BVECS);
785 BUG_ON(!osd_data->num_bvecs);
786
787 dout("ceph_aio_complete_req %p rc %d bytes %u\n",
788 inode, rc, osd_data->bvec_pos.iter.bi_size);
789
790 if (rc == -EOLDSNAPC) {
791 struct ceph_aio_work *aio_work;
792 BUG_ON(!aio_req->write);
793
794 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
795 if (aio_work) {
796 INIT_WORK(&aio_work->work, ceph_aio_retry_work);
797 aio_work->req = req;
798 queue_work(ceph_inode_to_client(inode)->wb_wq,
799 &aio_work->work);
800 return;
801 }
802 rc = -ENOMEM;
803 } else if (!aio_req->write) {
804 if (rc == -ENOENT)
805 rc = 0;
806 if (rc >= 0 && osd_data->bvec_pos.iter.bi_size > rc) {
807 struct iov_iter i;
808 int zlen = osd_data->bvec_pos.iter.bi_size - rc;
809
810 /*
811 * If read is satisfied by single OSD request,
812 * it can pass EOF. Otherwise read is within
813 * i_size.
814 */
815 if (aio_req->num_reqs == 1) {
816 loff_t i_size = i_size_read(inode);
817 loff_t endoff = aio_req->iocb->ki_pos + rc;
818 if (endoff < i_size)
819 zlen = min_t(size_t, zlen,
820 i_size - endoff);
821 aio_req->total_len = rc + zlen;
822 }
823
824 iov_iter_bvec(&i, ITER_BVEC, osd_data->bvec_pos.bvecs,
825 osd_data->num_bvecs,
826 osd_data->bvec_pos.iter.bi_size);
827 iov_iter_advance(&i, rc);
828 iov_iter_zero(zlen, &i);
829 }
830 }
831
832 put_bvecs(osd_data->bvec_pos.bvecs, osd_data->num_bvecs,
833 aio_req->should_dirty);
834 ceph_osdc_put_request(req);
835
836 if (rc < 0)
837 cmpxchg(&aio_req->error, 0, rc);
838
839 ceph_aio_complete(inode, aio_req);
840 return;
841 }
842
843 static void ceph_aio_retry_work(struct work_struct *work)
844 {
845 struct ceph_aio_work *aio_work =
846 container_of(work, struct ceph_aio_work, work);
847 struct ceph_osd_request *orig_req = aio_work->req;
848 struct ceph_aio_request *aio_req = orig_req->r_priv;
849 struct inode *inode = orig_req->r_inode;
850 struct ceph_inode_info *ci = ceph_inode(inode);
851 struct ceph_snap_context *snapc;
852 struct ceph_osd_request *req;
853 int ret;
854
855 spin_lock(&ci->i_ceph_lock);
856 if (__ceph_have_pending_cap_snap(ci)) {
857 struct ceph_cap_snap *capsnap =
858 list_last_entry(&ci->i_cap_snaps,
859 struct ceph_cap_snap,
860 ci_item);
861 snapc = ceph_get_snap_context(capsnap->context);
862 } else {
863 BUG_ON(!ci->i_head_snapc);
864 snapc = ceph_get_snap_context(ci->i_head_snapc);
865 }
866 spin_unlock(&ci->i_ceph_lock);
867
868 req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 2,
869 false, GFP_NOFS);
870 if (!req) {
871 ret = -ENOMEM;
872 req = orig_req;
873 goto out;
874 }
875
876 req->r_flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
877 ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
878 ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
879
880 ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
881 if (ret) {
882 ceph_osdc_put_request(req);
883 req = orig_req;
884 goto out;
885 }
886
887 req->r_ops[0] = orig_req->r_ops[0];
888
889 req->r_mtime = aio_req->mtime;
890 req->r_data_offset = req->r_ops[0].extent.offset;
891
892 ceph_osdc_put_request(orig_req);
893
894 req->r_callback = ceph_aio_complete_req;
895 req->r_inode = inode;
896 req->r_priv = aio_req;
897
898 ret = ceph_osdc_start_request(req->r_osdc, req, false);
899 out:
900 if (ret < 0) {
901 req->r_result = ret;
902 ceph_aio_complete_req(req);
903 }
904
905 ceph_put_snap_context(snapc);
906 kfree(aio_work);
907 }
908
909 static ssize_t
910 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
911 struct ceph_snap_context *snapc,
912 struct ceph_cap_flush **pcf)
913 {
914 struct file *file = iocb->ki_filp;
915 struct inode *inode = file_inode(file);
916 struct ceph_inode_info *ci = ceph_inode(inode);
917 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
918 struct ceph_vino vino;
919 struct ceph_osd_request *req;
920 struct bio_vec *bvecs;
921 struct ceph_aio_request *aio_req = NULL;
922 int num_pages = 0;
923 int flags;
924 int ret;
925 struct timespec64 mtime = current_time(inode);
926 size_t count = iov_iter_count(iter);
927 loff_t pos = iocb->ki_pos;
928 bool write = iov_iter_rw(iter) == WRITE;
929 bool should_dirty = !write && iter_is_iovec(iter);
930
931 if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
932 return -EROFS;
933
934 dout("sync_direct_%s on file %p %lld~%u snapc %p seq %lld\n",
935 (write ? "write" : "read"), file, pos, (unsigned)count,
936 snapc, snapc->seq);
937
938 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
939 if (ret < 0)
940 return ret;
941
942 if (write) {
943 int ret2 = invalidate_inode_pages2_range(inode->i_mapping,
944 pos >> PAGE_SHIFT,
945 (pos + count) >> PAGE_SHIFT);
946 if (ret2 < 0)
947 dout("invalidate_inode_pages2_range returned %d\n", ret2);
948
949 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
950 } else {
951 flags = CEPH_OSD_FLAG_READ;
952 }
953
954 while (iov_iter_count(iter) > 0) {
955 u64 size = iov_iter_count(iter);
956 ssize_t len;
957
958 if (write)
959 size = min_t(u64, size, fsc->mount_options->wsize);
960 else
961 size = min_t(u64, size, fsc->mount_options->rsize);
962
963 vino = ceph_vino(inode);
964 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
965 vino, pos, &size, 0,
966 1,
967 write ? CEPH_OSD_OP_WRITE :
968 CEPH_OSD_OP_READ,
969 flags, snapc,
970 ci->i_truncate_seq,
971 ci->i_truncate_size,
972 false);
973 if (IS_ERR(req)) {
974 ret = PTR_ERR(req);
975 break;
976 }
977
978 len = iter_get_bvecs_alloc(iter, size, &bvecs, &num_pages);
979 if (len < 0) {
980 ceph_osdc_put_request(req);
981 ret = len;
982 break;
983 }
984 if (len != size)
985 osd_req_op_extent_update(req, 0, len);
986
987 /*
988 * To simplify error handling, allow AIO when IO within i_size
989 * or IO can be satisfied by single OSD request.
990 */
991 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
992 (len == count || pos + count <= i_size_read(inode))) {
993 aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
994 if (aio_req) {
995 aio_req->iocb = iocb;
996 aio_req->write = write;
997 aio_req->should_dirty = should_dirty;
998 INIT_LIST_HEAD(&aio_req->osd_reqs);
999 if (write) {
1000 aio_req->mtime = mtime;
1001 swap(aio_req->prealloc_cf, *pcf);
1002 }
1003 }
1004 /* ignore error */
1005 }
1006
1007 if (write) {
1008 /*
1009 * throw out any page cache pages in this range. this
1010 * may block.
1011 */
1012 truncate_inode_pages_range(inode->i_mapping, pos,
1013 (pos+len) | (PAGE_SIZE - 1));
1014
1015 req->r_mtime = mtime;
1016 }
1017
1018 osd_req_op_extent_osd_data_bvecs(req, 0, bvecs, num_pages, len);
1019
1020 if (aio_req) {
1021 aio_req->total_len += len;
1022 aio_req->num_reqs++;
1023 atomic_inc(&aio_req->pending_reqs);
1024
1025 req->r_callback = ceph_aio_complete_req;
1026 req->r_inode = inode;
1027 req->r_priv = aio_req;
1028 list_add_tail(&req->r_unsafe_item, &aio_req->osd_reqs);
1029
1030 pos += len;
1031 continue;
1032 }
1033
1034 ret = ceph_osdc_start_request(req->r_osdc, req, false);
1035 if (!ret)
1036 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1037
1038 size = i_size_read(inode);
1039 if (!write) {
1040 if (ret == -ENOENT)
1041 ret = 0;
1042 if (ret >= 0 && ret < len && pos + ret < size) {
1043 struct iov_iter i;
1044 int zlen = min_t(size_t, len - ret,
1045 size - pos - ret);
1046
1047 iov_iter_bvec(&i, ITER_BVEC, bvecs, num_pages,
1048 len);
1049 iov_iter_advance(&i, ret);
1050 iov_iter_zero(zlen, &i);
1051 ret += zlen;
1052 }
1053 if (ret >= 0)
1054 len = ret;
1055 }
1056
1057 put_bvecs(bvecs, num_pages, should_dirty);
1058 ceph_osdc_put_request(req);
1059 if (ret < 0)
1060 break;
1061
1062 pos += len;
1063 if (!write && pos >= size)
1064 break;
1065
1066 if (write && pos > size) {
1067 if (ceph_inode_set_size(inode, pos))
1068 ceph_check_caps(ceph_inode(inode),
1069 CHECK_CAPS_AUTHONLY,
1070 NULL);
1071 }
1072 }
1073
1074 if (aio_req) {
1075 LIST_HEAD(osd_reqs);
1076
1077 if (aio_req->num_reqs == 0) {
1078 kfree(aio_req);
1079 return ret;
1080 }
1081
1082 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1083 CEPH_CAP_FILE_RD);
1084
1085 list_splice(&aio_req->osd_reqs, &osd_reqs);
1086 while (!list_empty(&osd_reqs)) {
1087 req = list_first_entry(&osd_reqs,
1088 struct ceph_osd_request,
1089 r_unsafe_item);
1090 list_del_init(&req->r_unsafe_item);
1091 if (ret >= 0)
1092 ret = ceph_osdc_start_request(req->r_osdc,
1093 req, false);
1094 if (ret < 0) {
1095 req->r_result = ret;
1096 ceph_aio_complete_req(req);
1097 }
1098 }
1099 return -EIOCBQUEUED;
1100 }
1101
1102 if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1103 ret = pos - iocb->ki_pos;
1104 iocb->ki_pos = pos;
1105 }
1106 return ret;
1107 }
1108
1109 /*
1110 * Synchronous write, straight from __user pointer or user pages.
1111 *
1112 * If write spans object boundary, just do multiple writes. (For a
1113 * correct atomic write, we should e.g. take write locks on all
1114 * objects, rollback on failure, etc.)
1115 */
1116 static ssize_t
1117 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1118 struct ceph_snap_context *snapc)
1119 {
1120 struct file *file = iocb->ki_filp;
1121 struct inode *inode = file_inode(file);
1122 struct ceph_inode_info *ci = ceph_inode(inode);
1123 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1124 struct ceph_vino vino;
1125 struct ceph_osd_request *req;
1126 struct page **pages;
1127 u64 len;
1128 int num_pages;
1129 int written = 0;
1130 int flags;
1131 int ret;
1132 bool check_caps = false;
1133 struct timespec64 mtime = current_time(inode);
1134 size_t count = iov_iter_count(from);
1135
1136 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1137 return -EROFS;
1138
1139 dout("sync_write on file %p %lld~%u snapc %p seq %lld\n",
1140 file, pos, (unsigned)count, snapc, snapc->seq);
1141
1142 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
1143 if (ret < 0)
1144 return ret;
1145
1146 ret = invalidate_inode_pages2_range(inode->i_mapping,
1147 pos >> PAGE_SHIFT,
1148 (pos + count) >> PAGE_SHIFT);
1149 if (ret < 0)
1150 dout("invalidate_inode_pages2_range returned %d\n", ret);
1151
1152 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1153
1154 while ((len = iov_iter_count(from)) > 0) {
1155 size_t left;
1156 int n;
1157
1158 vino = ceph_vino(inode);
1159 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1160 vino, pos, &len, 0, 1,
1161 CEPH_OSD_OP_WRITE, flags, snapc,
1162 ci->i_truncate_seq,
1163 ci->i_truncate_size,
1164 false);
1165 if (IS_ERR(req)) {
1166 ret = PTR_ERR(req);
1167 break;
1168 }
1169
1170 /*
1171 * write from beginning of first page,
1172 * regardless of io alignment
1173 */
1174 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1175
1176 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1177 if (IS_ERR(pages)) {
1178 ret = PTR_ERR(pages);
1179 goto out;
1180 }
1181
1182 left = len;
1183 for (n = 0; n < num_pages; n++) {
1184 size_t plen = min_t(size_t, left, PAGE_SIZE);
1185 ret = copy_page_from_iter(pages[n], 0, plen, from);
1186 if (ret != plen) {
1187 ret = -EFAULT;
1188 break;
1189 }
1190 left -= ret;
1191 }
1192
1193 if (ret < 0) {
1194 ceph_release_page_vector(pages, num_pages);
1195 goto out;
1196 }
1197
1198 req->r_inode = inode;
1199
1200 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1201 false, true);
1202
1203 req->r_mtime = mtime;
1204 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1205 if (!ret)
1206 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1207
1208 out:
1209 ceph_osdc_put_request(req);
1210 if (ret != 0) {
1211 ceph_set_error_write(ci);
1212 break;
1213 }
1214
1215 ceph_clear_error_write(ci);
1216 pos += len;
1217 written += len;
1218 if (pos > i_size_read(inode)) {
1219 check_caps = ceph_inode_set_size(inode, pos);
1220 if (check_caps)
1221 ceph_check_caps(ceph_inode(inode),
1222 CHECK_CAPS_AUTHONLY,
1223 NULL);
1224 }
1225
1226 }
1227
1228 if (ret != -EOLDSNAPC && written > 0) {
1229 ret = written;
1230 iocb->ki_pos = pos;
1231 }
1232 return ret;
1233 }
1234
1235 /*
1236 * Wrap generic_file_aio_read with checks for cap bits on the inode.
1237 * Atomically grab references, so that those bits are not released
1238 * back to the MDS mid-read.
1239 *
1240 * Hmm, the sync read case isn't actually async... should it be?
1241 */
1242 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1243 {
1244 struct file *filp = iocb->ki_filp;
1245 struct ceph_file_info *fi = filp->private_data;
1246 size_t len = iov_iter_count(to);
1247 struct inode *inode = file_inode(filp);
1248 struct ceph_inode_info *ci = ceph_inode(inode);
1249 struct page *pinned_page = NULL;
1250 ssize_t ret;
1251 int want, got = 0;
1252 int retry_op = 0, read = 0;
1253
1254 again:
1255 dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1256 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1257
1258 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1259 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1260 else
1261 want = CEPH_CAP_FILE_CACHE;
1262 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1263 if (ret < 0)
1264 return ret;
1265
1266 if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1267 (iocb->ki_flags & IOCB_DIRECT) ||
1268 (fi->flags & CEPH_F_SYNC)) {
1269
1270 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1271 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1272 ceph_cap_string(got));
1273
1274 if (ci->i_inline_version == CEPH_INLINE_NONE) {
1275 if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1276 ret = ceph_direct_read_write(iocb, to,
1277 NULL, NULL);
1278 if (ret >= 0 && ret < len)
1279 retry_op = CHECK_EOF;
1280 } else {
1281 ret = ceph_sync_read(iocb, to, &retry_op);
1282 }
1283 } else {
1284 retry_op = READ_INLINE;
1285 }
1286 } else {
1287 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1288 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1289 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1290 ceph_cap_string(got));
1291 ceph_add_rw_context(fi, &rw_ctx);
1292 ret = generic_file_read_iter(iocb, to);
1293 ceph_del_rw_context(fi, &rw_ctx);
1294 }
1295 dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1296 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1297 if (pinned_page) {
1298 put_page(pinned_page);
1299 pinned_page = NULL;
1300 }
1301 ceph_put_cap_refs(ci, got);
1302 if (retry_op > HAVE_RETRIED && ret >= 0) {
1303 int statret;
1304 struct page *page = NULL;
1305 loff_t i_size;
1306 if (retry_op == READ_INLINE) {
1307 page = __page_cache_alloc(GFP_KERNEL);
1308 if (!page)
1309 return -ENOMEM;
1310 }
1311
1312 statret = __ceph_do_getattr(inode, page,
1313 CEPH_STAT_CAP_INLINE_DATA, !!page);
1314 if (statret < 0) {
1315 if (page)
1316 __free_page(page);
1317 if (statret == -ENODATA) {
1318 BUG_ON(retry_op != READ_INLINE);
1319 goto again;
1320 }
1321 return statret;
1322 }
1323
1324 i_size = i_size_read(inode);
1325 if (retry_op == READ_INLINE) {
1326 BUG_ON(ret > 0 || read > 0);
1327 if (iocb->ki_pos < i_size &&
1328 iocb->ki_pos < PAGE_SIZE) {
1329 loff_t end = min_t(loff_t, i_size,
1330 iocb->ki_pos + len);
1331 end = min_t(loff_t, end, PAGE_SIZE);
1332 if (statret < end)
1333 zero_user_segment(page, statret, end);
1334 ret = copy_page_to_iter(page,
1335 iocb->ki_pos & ~PAGE_MASK,
1336 end - iocb->ki_pos, to);
1337 iocb->ki_pos += ret;
1338 read += ret;
1339 }
1340 if (iocb->ki_pos < i_size && read < len) {
1341 size_t zlen = min_t(size_t, len - read,
1342 i_size - iocb->ki_pos);
1343 ret = iov_iter_zero(zlen, to);
1344 iocb->ki_pos += ret;
1345 read += ret;
1346 }
1347 __free_pages(page, 0);
1348 return read;
1349 }
1350
1351 /* hit EOF or hole? */
1352 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1353 ret < len) {
1354 dout("sync_read hit hole, ppos %lld < size %lld"
1355 ", reading more\n", iocb->ki_pos, i_size);
1356
1357 read += ret;
1358 len -= ret;
1359 retry_op = HAVE_RETRIED;
1360 goto again;
1361 }
1362 }
1363
1364 if (ret >= 0)
1365 ret += read;
1366
1367 return ret;
1368 }
1369
1370 /*
1371 * Take cap references to avoid releasing caps to MDS mid-write.
1372 *
1373 * If we are synchronous, and write with an old snap context, the OSD
1374 * may return EOLDSNAPC. In that case, retry the write.. _after_
1375 * dropping our cap refs and allowing the pending snap to logically
1376 * complete _before_ this write occurs.
1377 *
1378 * If we are near ENOSPC, write synchronously.
1379 */
1380 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1381 {
1382 struct file *file = iocb->ki_filp;
1383 struct ceph_file_info *fi = file->private_data;
1384 struct inode *inode = file_inode(file);
1385 struct ceph_inode_info *ci = ceph_inode(inode);
1386 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1387 struct ceph_cap_flush *prealloc_cf;
1388 ssize_t count, written = 0;
1389 int err, want, got;
1390 loff_t pos;
1391 loff_t limit = max(i_size_read(inode), fsc->max_file_size);
1392
1393 if (ceph_snap(inode) != CEPH_NOSNAP)
1394 return -EROFS;
1395
1396 prealloc_cf = ceph_alloc_cap_flush();
1397 if (!prealloc_cf)
1398 return -ENOMEM;
1399
1400 retry_snap:
1401 inode_lock(inode);
1402
1403 /* We can write back this queue in page reclaim */
1404 current->backing_dev_info = inode_to_bdi(inode);
1405
1406 if (iocb->ki_flags & IOCB_APPEND) {
1407 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1408 if (err < 0)
1409 goto out;
1410 }
1411
1412 err = generic_write_checks(iocb, from);
1413 if (err <= 0)
1414 goto out;
1415
1416 pos = iocb->ki_pos;
1417 if (unlikely(pos >= limit)) {
1418 err = -EFBIG;
1419 goto out;
1420 } else {
1421 iov_iter_truncate(from, limit - pos);
1422 }
1423
1424 count = iov_iter_count(from);
1425 if (ceph_quota_is_max_bytes_exceeded(inode, pos + count)) {
1426 err = -EDQUOT;
1427 goto out;
1428 }
1429
1430 err = file_remove_privs(file);
1431 if (err)
1432 goto out;
1433
1434 err = file_update_time(file);
1435 if (err)
1436 goto out;
1437
1438 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1439 err = ceph_uninline_data(file, NULL);
1440 if (err < 0)
1441 goto out;
1442 }
1443
1444 /* FIXME: not complete since it doesn't account for being at quota */
1445 if (ceph_osdmap_flag(&fsc->client->osdc, CEPH_OSDMAP_FULL)) {
1446 err = -ENOSPC;
1447 goto out;
1448 }
1449
1450 dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1451 inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1452 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1453 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1454 else
1455 want = CEPH_CAP_FILE_BUFFER;
1456 got = 0;
1457 err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, pos + count,
1458 &got, NULL);
1459 if (err < 0)
1460 goto out;
1461
1462 dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1463 inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1464
1465 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1466 (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
1467 (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
1468 struct ceph_snap_context *snapc;
1469 struct iov_iter data;
1470 inode_unlock(inode);
1471
1472 spin_lock(&ci->i_ceph_lock);
1473 if (__ceph_have_pending_cap_snap(ci)) {
1474 struct ceph_cap_snap *capsnap =
1475 list_last_entry(&ci->i_cap_snaps,
1476 struct ceph_cap_snap,
1477 ci_item);
1478 snapc = ceph_get_snap_context(capsnap->context);
1479 } else {
1480 BUG_ON(!ci->i_head_snapc);
1481 snapc = ceph_get_snap_context(ci->i_head_snapc);
1482 }
1483 spin_unlock(&ci->i_ceph_lock);
1484
1485 /* we might need to revert back to that point */
1486 data = *from;
1487 if (iocb->ki_flags & IOCB_DIRECT)
1488 written = ceph_direct_read_write(iocb, &data, snapc,
1489 &prealloc_cf);
1490 else
1491 written = ceph_sync_write(iocb, &data, pos, snapc);
1492 if (written > 0)
1493 iov_iter_advance(from, written);
1494 ceph_put_snap_context(snapc);
1495 } else {
1496 /*
1497 * No need to acquire the i_truncate_mutex. Because
1498 * the MDS revokes Fwb caps before sending truncate
1499 * message to us. We can't get Fwb cap while there
1500 * are pending vmtruncate. So write and vmtruncate
1501 * can not run at the same time
1502 */
1503 written = generic_perform_write(file, from, pos);
1504 if (likely(written >= 0))
1505 iocb->ki_pos = pos + written;
1506 inode_unlock(inode);
1507 }
1508
1509 if (written >= 0) {
1510 int dirty;
1511
1512 spin_lock(&ci->i_ceph_lock);
1513 ci->i_inline_version = CEPH_INLINE_NONE;
1514 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1515 &prealloc_cf);
1516 spin_unlock(&ci->i_ceph_lock);
1517 if (dirty)
1518 __mark_inode_dirty(inode, dirty);
1519 if (ceph_quota_is_max_bytes_approaching(inode, iocb->ki_pos))
1520 ceph_check_caps(ci, CHECK_CAPS_NODELAY, NULL);
1521 }
1522
1523 dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n",
1524 inode, ceph_vinop(inode), pos, (unsigned)count,
1525 ceph_cap_string(got));
1526 ceph_put_cap_refs(ci, got);
1527
1528 if (written == -EOLDSNAPC) {
1529 dout("aio_write %p %llx.%llx %llu~%u" "got EOLDSNAPC, retrying\n",
1530 inode, ceph_vinop(inode), pos, (unsigned)count);
1531 goto retry_snap;
1532 }
1533
1534 if (written >= 0) {
1535 if (ceph_osdmap_flag(&fsc->client->osdc, CEPH_OSDMAP_NEARFULL))
1536 iocb->ki_flags |= IOCB_DSYNC;
1537 written = generic_write_sync(iocb, written);
1538 }
1539
1540 goto out_unlocked;
1541
1542 out:
1543 inode_unlock(inode);
1544 out_unlocked:
1545 ceph_free_cap_flush(prealloc_cf);
1546 current->backing_dev_info = NULL;
1547 return written ? written : err;
1548 }
1549
1550 /*
1551 * llseek. be sure to verify file size on SEEK_END.
1552 */
1553 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1554 {
1555 struct inode *inode = file->f_mapping->host;
1556 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1557 loff_t i_size;
1558 loff_t ret;
1559
1560 inode_lock(inode);
1561
1562 if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1563 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1564 if (ret < 0)
1565 goto out;
1566 }
1567
1568 i_size = i_size_read(inode);
1569 switch (whence) {
1570 case SEEK_END:
1571 offset += i_size;
1572 break;
1573 case SEEK_CUR:
1574 /*
1575 * Here we special-case the lseek(fd, 0, SEEK_CUR)
1576 * position-querying operation. Avoid rewriting the "same"
1577 * f_pos value back to the file because a concurrent read(),
1578 * write() or lseek() might have altered it
1579 */
1580 if (offset == 0) {
1581 ret = file->f_pos;
1582 goto out;
1583 }
1584 offset += file->f_pos;
1585 break;
1586 case SEEK_DATA:
1587 if (offset < 0 || offset >= i_size) {
1588 ret = -ENXIO;
1589 goto out;
1590 }
1591 break;
1592 case SEEK_HOLE:
1593 if (offset < 0 || offset >= i_size) {
1594 ret = -ENXIO;
1595 goto out;
1596 }
1597 offset = i_size;
1598 break;
1599 }
1600
1601 ret = vfs_setpos(file, offset, max(i_size, fsc->max_file_size));
1602
1603 out:
1604 inode_unlock(inode);
1605 return ret;
1606 }
1607
1608 static inline void ceph_zero_partial_page(
1609 struct inode *inode, loff_t offset, unsigned size)
1610 {
1611 struct page *page;
1612 pgoff_t index = offset >> PAGE_SHIFT;
1613
1614 page = find_lock_page(inode->i_mapping, index);
1615 if (page) {
1616 wait_on_page_writeback(page);
1617 zero_user(page, offset & (PAGE_SIZE - 1), size);
1618 unlock_page(page);
1619 put_page(page);
1620 }
1621 }
1622
1623 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1624 loff_t length)
1625 {
1626 loff_t nearly = round_up(offset, PAGE_SIZE);
1627 if (offset < nearly) {
1628 loff_t size = nearly - offset;
1629 if (length < size)
1630 size = length;
1631 ceph_zero_partial_page(inode, offset, size);
1632 offset += size;
1633 length -= size;
1634 }
1635 if (length >= PAGE_SIZE) {
1636 loff_t size = round_down(length, PAGE_SIZE);
1637 truncate_pagecache_range(inode, offset, offset + size - 1);
1638 offset += size;
1639 length -= size;
1640 }
1641 if (length)
1642 ceph_zero_partial_page(inode, offset, length);
1643 }
1644
1645 static int ceph_zero_partial_object(struct inode *inode,
1646 loff_t offset, loff_t *length)
1647 {
1648 struct ceph_inode_info *ci = ceph_inode(inode);
1649 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1650 struct ceph_osd_request *req;
1651 int ret = 0;
1652 loff_t zero = 0;
1653 int op;
1654
1655 if (!length) {
1656 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1657 length = &zero;
1658 } else {
1659 op = CEPH_OSD_OP_ZERO;
1660 }
1661
1662 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1663 ceph_vino(inode),
1664 offset, length,
1665 0, 1, op,
1666 CEPH_OSD_FLAG_WRITE,
1667 NULL, 0, 0, false);
1668 if (IS_ERR(req)) {
1669 ret = PTR_ERR(req);
1670 goto out;
1671 }
1672
1673 req->r_mtime = inode->i_mtime;
1674 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1675 if (!ret) {
1676 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1677 if (ret == -ENOENT)
1678 ret = 0;
1679 }
1680 ceph_osdc_put_request(req);
1681
1682 out:
1683 return ret;
1684 }
1685
1686 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
1687 {
1688 int ret = 0;
1689 struct ceph_inode_info *ci = ceph_inode(inode);
1690 s32 stripe_unit = ci->i_layout.stripe_unit;
1691 s32 stripe_count = ci->i_layout.stripe_count;
1692 s32 object_size = ci->i_layout.object_size;
1693 u64 object_set_size = object_size * stripe_count;
1694 u64 nearly, t;
1695
1696 /* round offset up to next period boundary */
1697 nearly = offset + object_set_size - 1;
1698 t = nearly;
1699 nearly -= do_div(t, object_set_size);
1700
1701 while (length && offset < nearly) {
1702 loff_t size = length;
1703 ret = ceph_zero_partial_object(inode, offset, &size);
1704 if (ret < 0)
1705 return ret;
1706 offset += size;
1707 length -= size;
1708 }
1709 while (length >= object_set_size) {
1710 int i;
1711 loff_t pos = offset;
1712 for (i = 0; i < stripe_count; ++i) {
1713 ret = ceph_zero_partial_object(inode, pos, NULL);
1714 if (ret < 0)
1715 return ret;
1716 pos += stripe_unit;
1717 }
1718 offset += object_set_size;
1719 length -= object_set_size;
1720 }
1721 while (length) {
1722 loff_t size = length;
1723 ret = ceph_zero_partial_object(inode, offset, &size);
1724 if (ret < 0)
1725 return ret;
1726 offset += size;
1727 length -= size;
1728 }
1729 return ret;
1730 }
1731
1732 static long ceph_fallocate(struct file *file, int mode,
1733 loff_t offset, loff_t length)
1734 {
1735 struct ceph_file_info *fi = file->private_data;
1736 struct inode *inode = file_inode(file);
1737 struct ceph_inode_info *ci = ceph_inode(inode);
1738 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1739 struct ceph_cap_flush *prealloc_cf;
1740 int want, got = 0;
1741 int dirty;
1742 int ret = 0;
1743 loff_t endoff = 0;
1744 loff_t size;
1745
1746 if ((offset + length) > max(i_size_read(inode), fsc->max_file_size))
1747 return -EFBIG;
1748
1749 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1750 return -EOPNOTSUPP;
1751
1752 if (!S_ISREG(inode->i_mode))
1753 return -EOPNOTSUPP;
1754
1755 prealloc_cf = ceph_alloc_cap_flush();
1756 if (!prealloc_cf)
1757 return -ENOMEM;
1758
1759 inode_lock(inode);
1760
1761 if (ceph_snap(inode) != CEPH_NOSNAP) {
1762 ret = -EROFS;
1763 goto unlock;
1764 }
1765
1766 if (!(mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE)) &&
1767 ceph_quota_is_max_bytes_exceeded(inode, offset + length)) {
1768 ret = -EDQUOT;
1769 goto unlock;
1770 }
1771
1772 if (ceph_osdmap_flag(&fsc->client->osdc, CEPH_OSDMAP_FULL) &&
1773 !(mode & FALLOC_FL_PUNCH_HOLE)) {
1774 ret = -ENOSPC;
1775 goto unlock;
1776 }
1777
1778 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1779 ret = ceph_uninline_data(file, NULL);
1780 if (ret < 0)
1781 goto unlock;
1782 }
1783
1784 size = i_size_read(inode);
1785 if (!(mode & FALLOC_FL_KEEP_SIZE)) {
1786 endoff = offset + length;
1787 ret = inode_newsize_ok(inode, endoff);
1788 if (ret)
1789 goto unlock;
1790 }
1791
1792 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1793 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1794 else
1795 want = CEPH_CAP_FILE_BUFFER;
1796
1797 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, endoff, &got, NULL);
1798 if (ret < 0)
1799 goto unlock;
1800
1801 if (mode & FALLOC_FL_PUNCH_HOLE) {
1802 if (offset < size)
1803 ceph_zero_pagecache_range(inode, offset, length);
1804 ret = ceph_zero_objects(inode, offset, length);
1805 } else if (endoff > size) {
1806 truncate_pagecache_range(inode, size, -1);
1807 if (ceph_inode_set_size(inode, endoff))
1808 ceph_check_caps(ceph_inode(inode),
1809 CHECK_CAPS_AUTHONLY, NULL);
1810 }
1811
1812 if (!ret) {
1813 spin_lock(&ci->i_ceph_lock);
1814 ci->i_inline_version = CEPH_INLINE_NONE;
1815 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1816 &prealloc_cf);
1817 spin_unlock(&ci->i_ceph_lock);
1818 if (dirty)
1819 __mark_inode_dirty(inode, dirty);
1820 if ((endoff > size) &&
1821 ceph_quota_is_max_bytes_approaching(inode, endoff))
1822 ceph_check_caps(ci, CHECK_CAPS_NODELAY, NULL);
1823 }
1824
1825 ceph_put_cap_refs(ci, got);
1826 unlock:
1827 inode_unlock(inode);
1828 ceph_free_cap_flush(prealloc_cf);
1829 return ret;
1830 }
1831
1832 const struct file_operations ceph_file_fops = {
1833 .open = ceph_open,
1834 .release = ceph_release,
1835 .llseek = ceph_llseek,
1836 .read_iter = ceph_read_iter,
1837 .write_iter = ceph_write_iter,
1838 .mmap = ceph_mmap,
1839 .fsync = ceph_fsync,
1840 .lock = ceph_lock,
1841 .flock = ceph_flock,
1842 .splice_read = generic_file_splice_read,
1843 .splice_write = iter_file_splice_write,
1844 .unlocked_ioctl = ceph_ioctl,
1845 .compat_ioctl = ceph_ioctl,
1846 .fallocate = ceph_fallocate,
1847 };
1848
Linux with added FPC-III support