?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
genirq: Introduce IRQD_CAN_RESERVE flag
[linux/fpc-iii.git] / fs / ceph / file.c
blob5c17125f45c786ab65aeeaa4782faedd895525a9
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 #define ceph_sys2wire(a) if (flags & a) { wire_flags |= CEPH_##a; flags &= ~a; }
34
35 ceph_sys2wire(O_CREAT);
36 ceph_sys2wire(O_EXCL);
37 ceph_sys2wire(O_TRUNC);
38 ceph_sys2wire(O_DIRECTORY);
39 ceph_sys2wire(O_NOFOLLOW);
40
41 #undef ceph_sys2wire
42
43 if (flags)
44 dout("unused open flags: %x", flags);
45
46 return cpu_to_le32(wire_flags);
47 }
48
49 /*
50 * Ceph file operations
51 *
52 * Implement basic open/close functionality, and implement
53 * read/write.
54 *
55 * We implement three modes of file I/O:
56 * - buffered uses the generic_file_aio_{read,write} helpers
57 *
58 * - synchronous is used when there is multi-client read/write
59 * sharing, avoids the page cache, and synchronously waits for an
60 * ack from the OSD.
61 *
62 * - direct io takes the variant of the sync path that references
63 * user pages directly.
64 *
65 * fsync() flushes and waits on dirty pages, but just queues metadata
66 * for writeback: since the MDS can recover size and mtime there is no
67 * need to wait for MDS acknowledgement.
68 */
69
70 /*
71 * Calculate the length sum of direct io vectors that can
72 * be combined into one page vector.
73 */
74 static size_t dio_get_pagev_size(const struct iov_iter *it)
75 {
76 const struct iovec *iov = it->iov;
77 const struct iovec *iovend = iov + it->nr_segs;
78 size_t size;
79
80 size = iov->iov_len - it->iov_offset;
81 /*
82 * An iov can be page vectored when both the current tail
83 * and the next base are page aligned.
84 */
85 while (PAGE_ALIGNED((iov->iov_base + iov->iov_len)) &&
86 (++iov < iovend && PAGE_ALIGNED((iov->iov_base)))) {
87 size += iov->iov_len;
88 }
89 dout("dio_get_pagevlen len = %zu\n", size);
90 return size;
91 }
92
93 /*
94 * Allocate a page vector based on (@it, @nbytes).
95 * The return value is the tuple describing a page vector,
96 * that is (@pages, @page_align, @num_pages).
97 */
98 static struct page **
99 dio_get_pages_alloc(const struct iov_iter *it, size_t nbytes,
100 size_t *page_align, int *num_pages)
101 {
102 struct iov_iter tmp_it = *it;
103 size_t align;
104 struct page **pages;
105 int ret = 0, idx, npages;
106
107 align = (unsigned long)(it->iov->iov_base + it->iov_offset) &
108 (PAGE_SIZE - 1);
109 npages = calc_pages_for(align, nbytes);
110 pages = kvmalloc(sizeof(*pages) * npages, GFP_KERNEL);
111 if (!pages)
112 return ERR_PTR(-ENOMEM);
113
114 for (idx = 0; idx < npages; ) {
115 size_t start;
116 ret = iov_iter_get_pages(&tmp_it, pages + idx, nbytes,
117 npages - idx, &start);
118 if (ret < 0)
119 goto fail;
120
121 iov_iter_advance(&tmp_it, ret);
122 nbytes -= ret;
123 idx += (ret + start + PAGE_SIZE - 1) / PAGE_SIZE;
124 }
125
126 BUG_ON(nbytes != 0);
127 *num_pages = npages;
128 *page_align = align;
129 dout("dio_get_pages_alloc: got %d pages align %zu\n", npages, align);
130 return pages;
131 fail:
132 ceph_put_page_vector(pages, idx, false);
133 return ERR_PTR(ret);
134 }
135
136 /*
137 * Prepare an open request. Preallocate ceph_cap to avoid an
138 * inopportune ENOMEM later.
139 */
140 static struct ceph_mds_request *
141 prepare_open_request(struct super_block *sb, int flags, int create_mode)
142 {
143 struct ceph_fs_client *fsc = ceph_sb_to_client(sb);
144 struct ceph_mds_client *mdsc = fsc->mdsc;
145 struct ceph_mds_request *req;
146 int want_auth = USE_ANY_MDS;
147 int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
148
149 if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
150 want_auth = USE_AUTH_MDS;
151
152 req = ceph_mdsc_create_request(mdsc, op, want_auth);
153 if (IS_ERR(req))
154 goto out;
155 req->r_fmode = ceph_flags_to_mode(flags);
156 req->r_args.open.flags = ceph_flags_sys2wire(flags);
157 req->r_args.open.mode = cpu_to_le32(create_mode);
158 out:
159 return req;
160 }
161
162 /*
163 * initialize private struct file data.
164 * if we fail, clean up by dropping fmode reference on the ceph_inode
165 */
166 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
167 {
168 struct ceph_file_info *cf;
169 int ret = 0;
170
171 switch (inode->i_mode & S_IFMT) {
172 case S_IFREG:
173 ceph_fscache_register_inode_cookie(inode);
174 ceph_fscache_file_set_cookie(inode, file);
175 case S_IFDIR:
176 dout("init_file %p %p 0%o (regular)\n", inode, file,
177 inode->i_mode);
178 cf = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
179 if (!cf) {
180 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
181 return -ENOMEM;
182 }
183 cf->fmode = fmode;
184 cf->next_offset = 2;
185 cf->readdir_cache_idx = -1;
186 file->private_data = cf;
187 BUG_ON(inode->i_fop->release != ceph_release);
188 break;
189
190 case S_IFLNK:
191 dout("init_file %p %p 0%o (symlink)\n", inode, file,
192 inode->i_mode);
193 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
194 break;
195
196 default:
197 dout("init_file %p %p 0%o (special)\n", inode, file,
198 inode->i_mode);
199 /*
200 * we need to drop the open ref now, since we don't
201 * have .release set to ceph_release.
202 */
203 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
204 BUG_ON(inode->i_fop->release == ceph_release);
205
206 /* call the proper open fop */
207 ret = inode->i_fop->open(inode, file);
208 }
209 return ret;
210 }
211
212 /*
213 * try renew caps after session gets killed.
214 */
215 int ceph_renew_caps(struct inode *inode)
216 {
217 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
218 struct ceph_inode_info *ci = ceph_inode(inode);
219 struct ceph_mds_request *req;
220 int err, flags, wanted;
221
222 spin_lock(&ci->i_ceph_lock);
223 wanted = __ceph_caps_file_wanted(ci);
224 if (__ceph_is_any_real_caps(ci) &&
225 (!(wanted & CEPH_CAP_ANY_WR) || ci->i_auth_cap)) {
226 int issued = __ceph_caps_issued(ci, NULL);
227 spin_unlock(&ci->i_ceph_lock);
228 dout("renew caps %p want %s issued %s updating mds_wanted\n",
229 inode, ceph_cap_string(wanted), ceph_cap_string(issued));
230 ceph_check_caps(ci, 0, NULL);
231 return 0;
232 }
233 spin_unlock(&ci->i_ceph_lock);
234
235 flags = 0;
236 if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
237 flags = O_RDWR;
238 else if (wanted & CEPH_CAP_FILE_RD)
239 flags = O_RDONLY;
240 else if (wanted & CEPH_CAP_FILE_WR)
241 flags = O_WRONLY;
242 #ifdef O_LAZY
243 if (wanted & CEPH_CAP_FILE_LAZYIO)
244 flags |= O_LAZY;
245 #endif
246
247 req = prepare_open_request(inode->i_sb, flags, 0);
248 if (IS_ERR(req)) {
249 err = PTR_ERR(req);
250 goto out;
251 }
252
253 req->r_inode = inode;
254 ihold(inode);
255 req->r_num_caps = 1;
256 req->r_fmode = -1;
257
258 err = ceph_mdsc_do_request(mdsc, NULL, req);
259 ceph_mdsc_put_request(req);
260 out:
261 dout("renew caps %p open result=%d\n", inode, err);
262 return err < 0 ? err : 0;
263 }
264
265 /*
266 * If we already have the requisite capabilities, we can satisfy
267 * the open request locally (no need to request new caps from the
268 * MDS). We do, however, need to inform the MDS (asynchronously)
269 * if our wanted caps set expands.
270 */
271 int ceph_open(struct inode *inode, struct file *file)
272 {
273 struct ceph_inode_info *ci = ceph_inode(inode);
274 struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
275 struct ceph_mds_client *mdsc = fsc->mdsc;
276 struct ceph_mds_request *req;
277 struct ceph_file_info *cf = file->private_data;
278 int err;
279 int flags, fmode, wanted;
280
281 if (cf) {
282 dout("open file %p is already opened\n", file);
283 return 0;
284 }
285
286 /* filter out O_CREAT|O_EXCL; vfs did that already. yuck. */
287 flags = file->f_flags & ~(O_CREAT|O_EXCL);
288 if (S_ISDIR(inode->i_mode))
289 flags = O_DIRECTORY; /* mds likes to know */
290
291 dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
292 ceph_vinop(inode), file, flags, file->f_flags);
293 fmode = ceph_flags_to_mode(flags);
294 wanted = ceph_caps_for_mode(fmode);
295
296 /* snapped files are read-only */
297 if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
298 return -EROFS;
299
300 /* trivially open snapdir */
301 if (ceph_snap(inode) == CEPH_SNAPDIR) {
302 spin_lock(&ci->i_ceph_lock);
303 __ceph_get_fmode(ci, fmode);
304 spin_unlock(&ci->i_ceph_lock);
305 return ceph_init_file(inode, file, fmode);
306 }
307
308 /*
309 * No need to block if we have caps on the auth MDS (for
310 * write) or any MDS (for read). Update wanted set
311 * asynchronously.
312 */
313 spin_lock(&ci->i_ceph_lock);
314 if (__ceph_is_any_real_caps(ci) &&
315 (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
316 int mds_wanted = __ceph_caps_mds_wanted(ci, true);
317 int issued = __ceph_caps_issued(ci, NULL);
318
319 dout("open %p fmode %d want %s issued %s using existing\n",
320 inode, fmode, ceph_cap_string(wanted),
321 ceph_cap_string(issued));
322 __ceph_get_fmode(ci, fmode);
323 spin_unlock(&ci->i_ceph_lock);
324
325 /* adjust wanted? */
326 if ((issued & wanted) != wanted &&
327 (mds_wanted & wanted) != wanted &&
328 ceph_snap(inode) != CEPH_SNAPDIR)
329 ceph_check_caps(ci, 0, NULL);
330
331 return ceph_init_file(inode, file, fmode);
332 } else if (ceph_snap(inode) != CEPH_NOSNAP &&
333 (ci->i_snap_caps & wanted) == wanted) {
334 __ceph_get_fmode(ci, fmode);
335 spin_unlock(&ci->i_ceph_lock);
336 return ceph_init_file(inode, file, fmode);
337 }
338
339 spin_unlock(&ci->i_ceph_lock);
340
341 dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
342 req = prepare_open_request(inode->i_sb, flags, 0);
343 if (IS_ERR(req)) {
344 err = PTR_ERR(req);
345 goto out;
346 }
347 req->r_inode = inode;
348 ihold(inode);
349
350 req->r_num_caps = 1;
351 err = ceph_mdsc_do_request(mdsc, NULL, req);
352 if (!err)
353 err = ceph_init_file(inode, file, req->r_fmode);
354 ceph_mdsc_put_request(req);
355 dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
356 out:
357 return err;
358 }
359
360
361 /*
362 * Do a lookup + open with a single request. If we get a non-existent
363 * file or symlink, return 1 so the VFS can retry.
364 */
365 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
366 struct file *file, unsigned flags, umode_t mode,
367 int *opened)
368 {
369 struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
370 struct ceph_mds_client *mdsc = fsc->mdsc;
371 struct ceph_mds_request *req;
372 struct dentry *dn;
373 struct ceph_acls_info acls = {};
374 int mask;
375 int err;
376
377 dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
378 dir, dentry, dentry,
379 d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
380
381 if (dentry->d_name.len > NAME_MAX)
382 return -ENAMETOOLONG;
383
384 if (flags & O_CREAT) {
385 err = ceph_pre_init_acls(dir, &mode, &acls);
386 if (err < 0)
387 return err;
388 }
389
390 /* do the open */
391 req = prepare_open_request(dir->i_sb, flags, mode);
392 if (IS_ERR(req)) {
393 err = PTR_ERR(req);
394 goto out_acl;
395 }
396 req->r_dentry = dget(dentry);
397 req->r_num_caps = 2;
398 if (flags & O_CREAT) {
399 req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
400 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
401 if (acls.pagelist) {
402 req->r_pagelist = acls.pagelist;
403 acls.pagelist = NULL;
404 }
405 }
406
407 mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
408 if (ceph_security_xattr_wanted(dir))
409 mask |= CEPH_CAP_XATTR_SHARED;
410 req->r_args.open.mask = cpu_to_le32(mask);
411
412 req->r_parent = dir;
413 set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
414 err = ceph_mdsc_do_request(mdsc,
415 (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
416 req);
417 err = ceph_handle_snapdir(req, dentry, err);
418 if (err)
419 goto out_req;
420
421 if ((flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
422 err = ceph_handle_notrace_create(dir, dentry);
423
424 if (d_in_lookup(dentry)) {
425 dn = ceph_finish_lookup(req, dentry, err);
426 if (IS_ERR(dn))
427 err = PTR_ERR(dn);
428 } else {
429 /* we were given a hashed negative dentry */
430 dn = NULL;
431 }
432 if (err)
433 goto out_req;
434 if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
435 /* make vfs retry on splice, ENOENT, or symlink */
436 dout("atomic_open finish_no_open on dn %p\n", dn);
437 err = finish_no_open(file, dn);
438 } else {
439 dout("atomic_open finish_open on dn %p\n", dn);
440 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
441 ceph_init_inode_acls(d_inode(dentry), &acls);
442 *opened |= FILE_CREATED;
443 }
444 err = finish_open(file, dentry, ceph_open, opened);
445 }
446 out_req:
447 if (!req->r_err && req->r_target_inode)
448 ceph_put_fmode(ceph_inode(req->r_target_inode), req->r_fmode);
449 ceph_mdsc_put_request(req);
450 out_acl:
451 ceph_release_acls_info(&acls);
452 dout("atomic_open result=%d\n", err);
453 return err;
454 }
455
456 int ceph_release(struct inode *inode, struct file *file)
457 {
458 struct ceph_inode_info *ci = ceph_inode(inode);
459 struct ceph_file_info *cf = file->private_data;
460
461 dout("release inode %p file %p\n", inode, file);
462 ceph_put_fmode(ci, cf->fmode);
463 if (cf->last_readdir)
464 ceph_mdsc_put_request(cf->last_readdir);
465 kfree(cf->last_name);
466 kfree(cf->dir_info);
467 kmem_cache_free(ceph_file_cachep, cf);
468
469 /* wake up anyone waiting for caps on this inode */
470 wake_up_all(&ci->i_cap_wq);
471 return 0;
472 }
473
474 enum {
475 HAVE_RETRIED = 1,
476 CHECK_EOF = 2,
477 READ_INLINE = 3,
478 };
479
480 /*
481 * Read a range of bytes striped over one or more objects. Iterate over
482 * objects we stripe over. (That's not atomic, but good enough for now.)
483 *
484 * If we get a short result from the OSD, check against i_size; we need to
485 * only return a short read to the caller if we hit EOF.
486 */
487 static int striped_read(struct inode *inode,
488 u64 pos, u64 len,
489 struct page **pages, int num_pages,
490 int page_align, int *checkeof)
491 {
492 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
493 struct ceph_inode_info *ci = ceph_inode(inode);
494 u64 this_len;
495 loff_t i_size;
496 int page_idx;
497 int ret, read = 0;
498 bool hit_stripe, was_short;
499
500 /*
501 * we may need to do multiple reads. not atomic, unfortunately.
502 */
503 more:
504 this_len = len;
505 page_idx = (page_align + read) >> PAGE_SHIFT;
506 ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode),
507 &ci->i_layout, pos, &this_len,
508 ci->i_truncate_seq, ci->i_truncate_size,
509 pages + page_idx, num_pages - page_idx,
510 ((page_align + read) & ~PAGE_MASK));
511 if (ret == -ENOENT)
512 ret = 0;
513 hit_stripe = this_len < len;
514 was_short = ret >= 0 && ret < this_len;
515 dout("striped_read %llu~%llu (read %u) got %d%s%s\n", pos, len, read,
516 ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : "");
517
518 i_size = i_size_read(inode);
519 if (ret >= 0) {
520 if (was_short && (pos + ret < i_size)) {
521 int zlen = min(this_len - ret, i_size - pos - ret);
522 int zoff = page_align + read + ret;
523 dout(" zero gap %llu to %llu\n",
524 pos + ret, pos + ret + zlen);
525 ceph_zero_page_vector_range(zoff, zlen, pages);
526 ret += zlen;
527 }
528
529 read += ret;
530 pos += ret;
531 len -= ret;
532
533 /* hit stripe and need continue*/
534 if (len && hit_stripe && pos < i_size)
535 goto more;
536 }
537
538 if (read > 0) {
539 ret = read;
540 /* did we bounce off eof? */
541 if (pos + len > i_size)
542 *checkeof = CHECK_EOF;
543 }
544
545 dout("striped_read returns %d\n", ret);
546 return ret;
547 }
548
549 /*
550 * Completely synchronous read and write methods. Direct from __user
551 * buffer to osd, or directly to user pages (if O_DIRECT).
552 *
553 * If the read spans object boundary, just do multiple reads.
554 */
555 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
556 int *checkeof)
557 {
558 struct file *file = iocb->ki_filp;
559 struct inode *inode = file_inode(file);
560 struct page **pages;
561 u64 off = iocb->ki_pos;
562 int num_pages;
563 ssize_t ret;
564 size_t len = iov_iter_count(to);
565
566 dout("sync_read on file %p %llu~%u %s\n", file, off, (unsigned)len,
567 (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
568
569 if (!len)
570 return 0;
571 /*
572 * flush any page cache pages in this range. this
573 * will make concurrent normal and sync io slow,
574 * but it will at least behave sensibly when they are
575 * in sequence.
576 */
577 ret = filemap_write_and_wait_range(inode->i_mapping, off,
578 off + len);
579 if (ret < 0)
580 return ret;
581
582 if (unlikely(to->type & ITER_PIPE)) {
583 size_t page_off;
584 ret = iov_iter_get_pages_alloc(to, &pages, len,
585 &page_off);
586 if (ret <= 0)
587 return -ENOMEM;
588 num_pages = DIV_ROUND_UP(ret + page_off, PAGE_SIZE);
589
590 ret = striped_read(inode, off, ret, pages, num_pages,
591 page_off, checkeof);
592 if (ret > 0) {
593 iov_iter_advance(to, ret);
594 off += ret;
595 } else {
596 iov_iter_advance(to, 0);
597 }
598 ceph_put_page_vector(pages, num_pages, false);
599 } else {
600 num_pages = calc_pages_for(off, len);
601 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
602 if (IS_ERR(pages))
603 return PTR_ERR(pages);
604
605 ret = striped_read(inode, off, len, pages, num_pages,
606 (off & ~PAGE_MASK), checkeof);
607 if (ret > 0) {
608 int l, k = 0;
609 size_t left = ret;
610
611 while (left) {
612 size_t page_off = off & ~PAGE_MASK;
613 size_t copy = min_t(size_t, left,
614 PAGE_SIZE - page_off);
615 l = copy_page_to_iter(pages[k++], page_off,
616 copy, to);
617 off += l;
618 left -= l;
619 if (l < copy)
620 break;
621 }
622 }
623 ceph_release_page_vector(pages, num_pages);
624 }
625
626 if (off > iocb->ki_pos) {
627 ret = off - iocb->ki_pos;
628 iocb->ki_pos = off;
629 }
630
631 dout("sync_read result %zd\n", ret);
632 return ret;
633 }
634
635 struct ceph_aio_request {
636 struct kiocb *iocb;
637 size_t total_len;
638 int write;
639 int error;
640 struct list_head osd_reqs;
641 unsigned num_reqs;
642 atomic_t pending_reqs;
643 struct timespec mtime;
644 struct ceph_cap_flush *prealloc_cf;
645 };
646
647 struct ceph_aio_work {
648 struct work_struct work;
649 struct ceph_osd_request *req;
650 };
651
652 static void ceph_aio_retry_work(struct work_struct *work);
653
654 static void ceph_aio_complete(struct inode *inode,
655 struct ceph_aio_request *aio_req)
656 {
657 struct ceph_inode_info *ci = ceph_inode(inode);
658 int ret;
659
660 if (!atomic_dec_and_test(&aio_req->pending_reqs))
661 return;
662
663 ret = aio_req->error;
664 if (!ret)
665 ret = aio_req->total_len;
666
667 dout("ceph_aio_complete %p rc %d\n", inode, ret);
668
669 if (ret >= 0 && aio_req->write) {
670 int dirty;
671
672 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
673 if (endoff > i_size_read(inode)) {
674 if (ceph_inode_set_size(inode, endoff))
675 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
676 }
677
678 spin_lock(&ci->i_ceph_lock);
679 ci->i_inline_version = CEPH_INLINE_NONE;
680 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
681 &aio_req->prealloc_cf);
682 spin_unlock(&ci->i_ceph_lock);
683 if (dirty)
684 __mark_inode_dirty(inode, dirty);
685
686 }
687
688 ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
689 CEPH_CAP_FILE_RD));
690
691 aio_req->iocb->ki_complete(aio_req->iocb, ret, 0);
692
693 ceph_free_cap_flush(aio_req->prealloc_cf);
694 kfree(aio_req);
695 }
696
697 static void ceph_aio_complete_req(struct ceph_osd_request *req)
698 {
699 int rc = req->r_result;
700 struct inode *inode = req->r_inode;
701 struct ceph_aio_request *aio_req = req->r_priv;
702 struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
703 int num_pages = calc_pages_for((u64)osd_data->alignment,
704 osd_data->length);
705
706 dout("ceph_aio_complete_req %p rc %d bytes %llu\n",
707 inode, rc, osd_data->length);
708
709 if (rc == -EOLDSNAPC) {
710 struct ceph_aio_work *aio_work;
711 BUG_ON(!aio_req->write);
712
713 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
714 if (aio_work) {
715 INIT_WORK(&aio_work->work, ceph_aio_retry_work);
716 aio_work->req = req;
717 queue_work(ceph_inode_to_client(inode)->wb_wq,
718 &aio_work->work);
719 return;
720 }
721 rc = -ENOMEM;
722 } else if (!aio_req->write) {
723 if (rc == -ENOENT)
724 rc = 0;
725 if (rc >= 0 && osd_data->length > rc) {
726 int zoff = osd_data->alignment + rc;
727 int zlen = osd_data->length - rc;
728 /*
729 * If read is satisfied by single OSD request,
730 * it can pass EOF. Otherwise read is within
731 * i_size.
732 */
733 if (aio_req->num_reqs == 1) {
734 loff_t i_size = i_size_read(inode);
735 loff_t endoff = aio_req->iocb->ki_pos + rc;
736 if (endoff < i_size)
737 zlen = min_t(size_t, zlen,
738 i_size - endoff);
739 aio_req->total_len = rc + zlen;
740 }
741
742 if (zlen > 0)
743 ceph_zero_page_vector_range(zoff, zlen,
744 osd_data->pages);
745 }
746 }
747
748 ceph_put_page_vector(osd_data->pages, num_pages, !aio_req->write);
749 ceph_osdc_put_request(req);
750
751 if (rc < 0)
752 cmpxchg(&aio_req->error, 0, rc);
753
754 ceph_aio_complete(inode, aio_req);
755 return;
756 }
757
758 static void ceph_aio_retry_work(struct work_struct *work)
759 {
760 struct ceph_aio_work *aio_work =
761 container_of(work, struct ceph_aio_work, work);
762 struct ceph_osd_request *orig_req = aio_work->req;
763 struct ceph_aio_request *aio_req = orig_req->r_priv;
764 struct inode *inode = orig_req->r_inode;
765 struct ceph_inode_info *ci = ceph_inode(inode);
766 struct ceph_snap_context *snapc;
767 struct ceph_osd_request *req;
768 int ret;
769
770 spin_lock(&ci->i_ceph_lock);
771 if (__ceph_have_pending_cap_snap(ci)) {
772 struct ceph_cap_snap *capsnap =
773 list_last_entry(&ci->i_cap_snaps,
774 struct ceph_cap_snap,
775 ci_item);
776 snapc = ceph_get_snap_context(capsnap->context);
777 } else {
778 BUG_ON(!ci->i_head_snapc);
779 snapc = ceph_get_snap_context(ci->i_head_snapc);
780 }
781 spin_unlock(&ci->i_ceph_lock);
782
783 req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 2,
784 false, GFP_NOFS);
785 if (!req) {
786 ret = -ENOMEM;
787 req = orig_req;
788 goto out;
789 }
790
791 req->r_flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
792 ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
793 ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
794
795 ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
796 if (ret) {
797 ceph_osdc_put_request(req);
798 req = orig_req;
799 goto out;
800 }
801
802 req->r_ops[0] = orig_req->r_ops[0];
803
804 req->r_mtime = aio_req->mtime;
805 req->r_data_offset = req->r_ops[0].extent.offset;
806
807 ceph_osdc_put_request(orig_req);
808
809 req->r_callback = ceph_aio_complete_req;
810 req->r_inode = inode;
811 req->r_priv = aio_req;
812 req->r_abort_on_full = true;
813
814 ret = ceph_osdc_start_request(req->r_osdc, req, false);
815 out:
816 if (ret < 0) {
817 req->r_result = ret;
818 ceph_aio_complete_req(req);
819 }
820
821 ceph_put_snap_context(snapc);
822 kfree(aio_work);
823 }
824
825 static ssize_t
826 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
827 struct ceph_snap_context *snapc,
828 struct ceph_cap_flush **pcf)
829 {
830 struct file *file = iocb->ki_filp;
831 struct inode *inode = file_inode(file);
832 struct ceph_inode_info *ci = ceph_inode(inode);
833 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
834 struct ceph_vino vino;
835 struct ceph_osd_request *req;
836 struct page **pages;
837 struct ceph_aio_request *aio_req = NULL;
838 int num_pages = 0;
839 int flags;
840 int ret;
841 struct timespec mtime = current_time(inode);
842 size_t count = iov_iter_count(iter);
843 loff_t pos = iocb->ki_pos;
844 bool write = iov_iter_rw(iter) == WRITE;
845
846 if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
847 return -EROFS;
848
849 dout("sync_direct_%s on file %p %lld~%u snapc %p seq %lld\n",
850 (write ? "write" : "read"), file, pos, (unsigned)count,
851 snapc, snapc->seq);
852
853 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
854 if (ret < 0)
855 return ret;
856
857 if (write) {
858 int ret2 = invalidate_inode_pages2_range(inode->i_mapping,
859 pos >> PAGE_SHIFT,
860 (pos + count) >> PAGE_SHIFT);
861 if (ret2 < 0)
862 dout("invalidate_inode_pages2_range returned %d\n", ret2);
863
864 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
865 } else {
866 flags = CEPH_OSD_FLAG_READ;
867 }
868
869 while (iov_iter_count(iter) > 0) {
870 u64 size = dio_get_pagev_size(iter);
871 size_t start = 0;
872 ssize_t len;
873
874 vino = ceph_vino(inode);
875 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
876 vino, pos, &size, 0,
877 1,
878 write ? CEPH_OSD_OP_WRITE :
879 CEPH_OSD_OP_READ,
880 flags, snapc,
881 ci->i_truncate_seq,
882 ci->i_truncate_size,
883 false);
884 if (IS_ERR(req)) {
885 ret = PTR_ERR(req);
886 break;
887 }
888
889 if (write)
890 size = min_t(u64, size, fsc->mount_options->wsize);
891 else
892 size = min_t(u64, size, fsc->mount_options->rsize);
893
894 len = size;
895 pages = dio_get_pages_alloc(iter, len, &start, &num_pages);
896 if (IS_ERR(pages)) {
897 ceph_osdc_put_request(req);
898 ret = PTR_ERR(pages);
899 break;
900 }
901
902 /*
903 * To simplify error handling, allow AIO when IO within i_size
904 * or IO can be satisfied by single OSD request.
905 */
906 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
907 (len == count || pos + count <= i_size_read(inode))) {
908 aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
909 if (aio_req) {
910 aio_req->iocb = iocb;
911 aio_req->write = write;
912 INIT_LIST_HEAD(&aio_req->osd_reqs);
913 if (write) {
914 aio_req->mtime = mtime;
915 swap(aio_req->prealloc_cf, *pcf);
916 }
917 }
918 /* ignore error */
919 }
920
921 if (write) {
922 /*
923 * throw out any page cache pages in this range. this
924 * may block.
925 */
926 truncate_inode_pages_range(inode->i_mapping, pos,
927 (pos+len) | (PAGE_SIZE - 1));
928
929 req->r_mtime = mtime;
930 }
931
932 osd_req_op_extent_osd_data_pages(req, 0, pages, len, start,
933 false, false);
934
935 if (aio_req) {
936 aio_req->total_len += len;
937 aio_req->num_reqs++;
938 atomic_inc(&aio_req->pending_reqs);
939
940 req->r_callback = ceph_aio_complete_req;
941 req->r_inode = inode;
942 req->r_priv = aio_req;
943 list_add_tail(&req->r_unsafe_item, &aio_req->osd_reqs);
944
945 pos += len;
946 iov_iter_advance(iter, len);
947 continue;
948 }
949
950 ret = ceph_osdc_start_request(req->r_osdc, req, false);
951 if (!ret)
952 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
953
954 size = i_size_read(inode);
955 if (!write) {
956 if (ret == -ENOENT)
957 ret = 0;
958 if (ret >= 0 && ret < len && pos + ret < size) {
959 int zlen = min_t(size_t, len - ret,
960 size - pos - ret);
961 ceph_zero_page_vector_range(start + ret, zlen,
962 pages);
963 ret += zlen;
964 }
965 if (ret >= 0)
966 len = ret;
967 }
968
969 ceph_put_page_vector(pages, num_pages, !write);
970
971 ceph_osdc_put_request(req);
972 if (ret < 0)
973 break;
974
975 pos += len;
976 iov_iter_advance(iter, len);
977
978 if (!write && pos >= size)
979 break;
980
981 if (write && pos > size) {
982 if (ceph_inode_set_size(inode, pos))
983 ceph_check_caps(ceph_inode(inode),
984 CHECK_CAPS_AUTHONLY,
985 NULL);
986 }
987 }
988
989 if (aio_req) {
990 LIST_HEAD(osd_reqs);
991
992 if (aio_req->num_reqs == 0) {
993 kfree(aio_req);
994 return ret;
995 }
996
997 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
998 CEPH_CAP_FILE_RD);
999
1000 list_splice(&aio_req->osd_reqs, &osd_reqs);
1001 while (!list_empty(&osd_reqs)) {
1002 req = list_first_entry(&osd_reqs,
1003 struct ceph_osd_request,
1004 r_unsafe_item);
1005 list_del_init(&req->r_unsafe_item);
1006 if (ret >= 0)
1007 ret = ceph_osdc_start_request(req->r_osdc,
1008 req, false);
1009 if (ret < 0) {
1010 req->r_result = ret;
1011 ceph_aio_complete_req(req);
1012 }
1013 }
1014 return -EIOCBQUEUED;
1015 }
1016
1017 if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1018 ret = pos - iocb->ki_pos;
1019 iocb->ki_pos = pos;
1020 }
1021 return ret;
1022 }
1023
1024 /*
1025 * Synchronous write, straight from __user pointer or user pages.
1026 *
1027 * If write spans object boundary, just do multiple writes. (For a
1028 * correct atomic write, we should e.g. take write locks on all
1029 * objects, rollback on failure, etc.)
1030 */
1031 static ssize_t
1032 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1033 struct ceph_snap_context *snapc)
1034 {
1035 struct file *file = iocb->ki_filp;
1036 struct inode *inode = file_inode(file);
1037 struct ceph_inode_info *ci = ceph_inode(inode);
1038 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1039 struct ceph_vino vino;
1040 struct ceph_osd_request *req;
1041 struct page **pages;
1042 u64 len;
1043 int num_pages;
1044 int written = 0;
1045 int flags;
1046 int ret;
1047 bool check_caps = false;
1048 struct timespec mtime = current_time(inode);
1049 size_t count = iov_iter_count(from);
1050
1051 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1052 return -EROFS;
1053
1054 dout("sync_write on file %p %lld~%u snapc %p seq %lld\n",
1055 file, pos, (unsigned)count, snapc, snapc->seq);
1056
1057 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
1058 if (ret < 0)
1059 return ret;
1060
1061 ret = invalidate_inode_pages2_range(inode->i_mapping,
1062 pos >> PAGE_SHIFT,
1063 (pos + count) >> PAGE_SHIFT);
1064 if (ret < 0)
1065 dout("invalidate_inode_pages2_range returned %d\n", ret);
1066
1067 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1068
1069 while ((len = iov_iter_count(from)) > 0) {
1070 size_t left;
1071 int n;
1072
1073 vino = ceph_vino(inode);
1074 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1075 vino, pos, &len, 0, 1,
1076 CEPH_OSD_OP_WRITE, flags, snapc,
1077 ci->i_truncate_seq,
1078 ci->i_truncate_size,
1079 false);
1080 if (IS_ERR(req)) {
1081 ret = PTR_ERR(req);
1082 break;
1083 }
1084
1085 /*
1086 * write from beginning of first page,
1087 * regardless of io alignment
1088 */
1089 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1090
1091 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1092 if (IS_ERR(pages)) {
1093 ret = PTR_ERR(pages);
1094 goto out;
1095 }
1096
1097 left = len;
1098 for (n = 0; n < num_pages; n++) {
1099 size_t plen = min_t(size_t, left, PAGE_SIZE);
1100 ret = copy_page_from_iter(pages[n], 0, plen, from);
1101 if (ret != plen) {
1102 ret = -EFAULT;
1103 break;
1104 }
1105 left -= ret;
1106 }
1107
1108 if (ret < 0) {
1109 ceph_release_page_vector(pages, num_pages);
1110 goto out;
1111 }
1112
1113 req->r_inode = inode;
1114
1115 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1116 false, true);
1117
1118 req->r_mtime = mtime;
1119 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1120 if (!ret)
1121 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1122
1123 out:
1124 ceph_osdc_put_request(req);
1125 if (ret != 0) {
1126 ceph_set_error_write(ci);
1127 break;
1128 }
1129
1130 ceph_clear_error_write(ci);
1131 pos += len;
1132 written += len;
1133 if (pos > i_size_read(inode)) {
1134 check_caps = ceph_inode_set_size(inode, pos);
1135 if (check_caps)
1136 ceph_check_caps(ceph_inode(inode),
1137 CHECK_CAPS_AUTHONLY,
1138 NULL);
1139 }
1140
1141 }
1142
1143 if (ret != -EOLDSNAPC && written > 0) {
1144 ret = written;
1145 iocb->ki_pos = pos;
1146 }
1147 return ret;
1148 }
1149
1150 /*
1151 * Wrap generic_file_aio_read with checks for cap bits on the inode.
1152 * Atomically grab references, so that those bits are not released
1153 * back to the MDS mid-read.
1154 *
1155 * Hmm, the sync read case isn't actually async... should it be?
1156 */
1157 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1158 {
1159 struct file *filp = iocb->ki_filp;
1160 struct ceph_file_info *fi = filp->private_data;
1161 size_t len = iov_iter_count(to);
1162 struct inode *inode = file_inode(filp);
1163 struct ceph_inode_info *ci = ceph_inode(inode);
1164 struct page *pinned_page = NULL;
1165 ssize_t ret;
1166 int want, got = 0;
1167 int retry_op = 0, read = 0;
1168
1169 again:
1170 dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1171 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1172
1173 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1174 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1175 else
1176 want = CEPH_CAP_FILE_CACHE;
1177 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1178 if (ret < 0)
1179 return ret;
1180
1181 if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1182 (iocb->ki_flags & IOCB_DIRECT) ||
1183 (fi->flags & CEPH_F_SYNC)) {
1184
1185 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1186 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1187 ceph_cap_string(got));
1188
1189 if (ci->i_inline_version == CEPH_INLINE_NONE) {
1190 if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1191 ret = ceph_direct_read_write(iocb, to,
1192 NULL, NULL);
1193 if (ret >= 0 && ret < len)
1194 retry_op = CHECK_EOF;
1195 } else {
1196 ret = ceph_sync_read(iocb, to, &retry_op);
1197 }
1198 } else {
1199 retry_op = READ_INLINE;
1200 }
1201 } else {
1202 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1203 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1204 ceph_cap_string(got));
1205 current->journal_info = filp;
1206 ret = generic_file_read_iter(iocb, to);
1207 current->journal_info = NULL;
1208 }
1209 dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1210 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1211 if (pinned_page) {
1212 put_page(pinned_page);
1213 pinned_page = NULL;
1214 }
1215 ceph_put_cap_refs(ci, got);
1216 if (retry_op > HAVE_RETRIED && ret >= 0) {
1217 int statret;
1218 struct page *page = NULL;
1219 loff_t i_size;
1220 if (retry_op == READ_INLINE) {
1221 page = __page_cache_alloc(GFP_KERNEL);
1222 if (!page)
1223 return -ENOMEM;
1224 }
1225
1226 statret = __ceph_do_getattr(inode, page,
1227 CEPH_STAT_CAP_INLINE_DATA, !!page);
1228 if (statret < 0) {
1229 if (page)
1230 __free_page(page);
1231 if (statret == -ENODATA) {
1232 BUG_ON(retry_op != READ_INLINE);
1233 goto again;
1234 }
1235 return statret;
1236 }
1237
1238 i_size = i_size_read(inode);
1239 if (retry_op == READ_INLINE) {
1240 BUG_ON(ret > 0 || read > 0);
1241 if (iocb->ki_pos < i_size &&
1242 iocb->ki_pos < PAGE_SIZE) {
1243 loff_t end = min_t(loff_t, i_size,
1244 iocb->ki_pos + len);
1245 end = min_t(loff_t, end, PAGE_SIZE);
1246 if (statret < end)
1247 zero_user_segment(page, statret, end);
1248 ret = copy_page_to_iter(page,
1249 iocb->ki_pos & ~PAGE_MASK,
1250 end - iocb->ki_pos, to);
1251 iocb->ki_pos += ret;
1252 read += ret;
1253 }
1254 if (iocb->ki_pos < i_size && read < len) {
1255 size_t zlen = min_t(size_t, len - read,
1256 i_size - iocb->ki_pos);
1257 ret = iov_iter_zero(zlen, to);
1258 iocb->ki_pos += ret;
1259 read += ret;
1260 }
1261 __free_pages(page, 0);
1262 return read;
1263 }
1264
1265 /* hit EOF or hole? */
1266 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1267 ret < len) {
1268 dout("sync_read hit hole, ppos %lld < size %lld"
1269 ", reading more\n", iocb->ki_pos, i_size);
1270
1271 read += ret;
1272 len -= ret;
1273 retry_op = HAVE_RETRIED;
1274 goto again;
1275 }
1276 }
1277
1278 if (ret >= 0)
1279 ret += read;
1280
1281 return ret;
1282 }
1283
1284 /*
1285 * Take cap references to avoid releasing caps to MDS mid-write.
1286 *
1287 * If we are synchronous, and write with an old snap context, the OSD
1288 * may return EOLDSNAPC. In that case, retry the write.. _after_
1289 * dropping our cap refs and allowing the pending snap to logically
1290 * complete _before_ this write occurs.
1291 *
1292 * If we are near ENOSPC, write synchronously.
1293 */
1294 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1295 {
1296 struct file *file = iocb->ki_filp;
1297 struct ceph_file_info *fi = file->private_data;
1298 struct inode *inode = file_inode(file);
1299 struct ceph_inode_info *ci = ceph_inode(inode);
1300 struct ceph_osd_client *osdc =
1301 &ceph_sb_to_client(inode->i_sb)->client->osdc;
1302 struct ceph_cap_flush *prealloc_cf;
1303 ssize_t count, written = 0;
1304 int err, want, got;
1305 loff_t pos;
1306
1307 if (ceph_snap(inode) != CEPH_NOSNAP)
1308 return -EROFS;
1309
1310 prealloc_cf = ceph_alloc_cap_flush();
1311 if (!prealloc_cf)
1312 return -ENOMEM;
1313
1314 retry_snap:
1315 inode_lock(inode);
1316
1317 /* We can write back this queue in page reclaim */
1318 current->backing_dev_info = inode_to_bdi(inode);
1319
1320 if (iocb->ki_flags & IOCB_APPEND) {
1321 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1322 if (err < 0)
1323 goto out;
1324 }
1325
1326 err = generic_write_checks(iocb, from);
1327 if (err <= 0)
1328 goto out;
1329
1330 pos = iocb->ki_pos;
1331 count = iov_iter_count(from);
1332 err = file_remove_privs(file);
1333 if (err)
1334 goto out;
1335
1336 err = file_update_time(file);
1337 if (err)
1338 goto out;
1339
1340 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1341 err = ceph_uninline_data(file, NULL);
1342 if (err < 0)
1343 goto out;
1344 }
1345
1346 /* FIXME: not complete since it doesn't account for being at quota */
1347 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL)) {
1348 err = -ENOSPC;
1349 goto out;
1350 }
1351
1352 dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1353 inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1354 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1355 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1356 else
1357 want = CEPH_CAP_FILE_BUFFER;
1358 got = 0;
1359 err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, pos + count,
1360 &got, NULL);
1361 if (err < 0)
1362 goto out;
1363
1364 dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1365 inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1366
1367 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1368 (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
1369 (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
1370 struct ceph_snap_context *snapc;
1371 struct iov_iter data;
1372 inode_unlock(inode);
1373
1374 spin_lock(&ci->i_ceph_lock);
1375 if (__ceph_have_pending_cap_snap(ci)) {
1376 struct ceph_cap_snap *capsnap =
1377 list_last_entry(&ci->i_cap_snaps,
1378 struct ceph_cap_snap,
1379 ci_item);
1380 snapc = ceph_get_snap_context(capsnap->context);
1381 } else {
1382 BUG_ON(!ci->i_head_snapc);
1383 snapc = ceph_get_snap_context(ci->i_head_snapc);
1384 }
1385 spin_unlock(&ci->i_ceph_lock);
1386
1387 /* we might need to revert back to that point */
1388 data = *from;
1389 if (iocb->ki_flags & IOCB_DIRECT)
1390 written = ceph_direct_read_write(iocb, &data, snapc,
1391 &prealloc_cf);
1392 else
1393 written = ceph_sync_write(iocb, &data, pos, snapc);
1394 if (written > 0)
1395 iov_iter_advance(from, written);
1396 ceph_put_snap_context(snapc);
1397 } else {
1398 /*
1399 * No need to acquire the i_truncate_mutex. Because
1400 * the MDS revokes Fwb caps before sending truncate
1401 * message to us. We can't get Fwb cap while there
1402 * are pending vmtruncate. So write and vmtruncate
1403 * can not run at the same time
1404 */
1405 written = generic_perform_write(file, from, pos);
1406 if (likely(written >= 0))
1407 iocb->ki_pos = pos + written;
1408 inode_unlock(inode);
1409 }
1410
1411 if (written >= 0) {
1412 int dirty;
1413 spin_lock(&ci->i_ceph_lock);
1414 ci->i_inline_version = CEPH_INLINE_NONE;
1415 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1416 &prealloc_cf);
1417 spin_unlock(&ci->i_ceph_lock);
1418 if (dirty)
1419 __mark_inode_dirty(inode, dirty);
1420 }
1421
1422 dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n",
1423 inode, ceph_vinop(inode), pos, (unsigned)count,
1424 ceph_cap_string(got));
1425 ceph_put_cap_refs(ci, got);
1426
1427 if (written == -EOLDSNAPC) {
1428 dout("aio_write %p %llx.%llx %llu~%u" "got EOLDSNAPC, retrying\n",
1429 inode, ceph_vinop(inode), pos, (unsigned)count);
1430 goto retry_snap;
1431 }
1432
1433 if (written >= 0) {
1434 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_NEARFULL))
1435 iocb->ki_flags |= IOCB_DSYNC;
1436 written = generic_write_sync(iocb, written);
1437 }
1438
1439 goto out_unlocked;
1440
1441 out:
1442 inode_unlock(inode);
1443 out_unlocked:
1444 ceph_free_cap_flush(prealloc_cf);
1445 current->backing_dev_info = NULL;
1446 return written ? written : err;
1447 }
1448
1449 /*
1450 * llseek. be sure to verify file size on SEEK_END.
1451 */
1452 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1453 {
1454 struct inode *inode = file->f_mapping->host;
1455 loff_t i_size;
1456 loff_t ret;
1457
1458 inode_lock(inode);
1459
1460 if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1461 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1462 if (ret < 0)
1463 goto out;
1464 }
1465
1466 i_size = i_size_read(inode);
1467 switch (whence) {
1468 case SEEK_END:
1469 offset += i_size;
1470 break;
1471 case SEEK_CUR:
1472 /*
1473 * Here we special-case the lseek(fd, 0, SEEK_CUR)
1474 * position-querying operation. Avoid rewriting the "same"
1475 * f_pos value back to the file because a concurrent read(),
1476 * write() or lseek() might have altered it
1477 */
1478 if (offset == 0) {
1479 ret = file->f_pos;
1480 goto out;
1481 }
1482 offset += file->f_pos;
1483 break;
1484 case SEEK_DATA:
1485 if (offset < 0 || offset >= i_size) {
1486 ret = -ENXIO;
1487 goto out;
1488 }
1489 break;
1490 case SEEK_HOLE:
1491 if (offset < 0 || offset >= i_size) {
1492 ret = -ENXIO;
1493 goto out;
1494 }
1495 offset = i_size;
1496 break;
1497 }
1498
1499 ret = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1500
1501 out:
1502 inode_unlock(inode);
1503 return ret;
1504 }
1505
1506 static inline void ceph_zero_partial_page(
1507 struct inode *inode, loff_t offset, unsigned size)
1508 {
1509 struct page *page;
1510 pgoff_t index = offset >> PAGE_SHIFT;
1511
1512 page = find_lock_page(inode->i_mapping, index);
1513 if (page) {
1514 wait_on_page_writeback(page);
1515 zero_user(page, offset & (PAGE_SIZE - 1), size);
1516 unlock_page(page);
1517 put_page(page);
1518 }
1519 }
1520
1521 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1522 loff_t length)
1523 {
1524 loff_t nearly = round_up(offset, PAGE_SIZE);
1525 if (offset < nearly) {
1526 loff_t size = nearly - offset;
1527 if (length < size)
1528 size = length;
1529 ceph_zero_partial_page(inode, offset, size);
1530 offset += size;
1531 length -= size;
1532 }
1533 if (length >= PAGE_SIZE) {
1534 loff_t size = round_down(length, PAGE_SIZE);
1535 truncate_pagecache_range(inode, offset, offset + size - 1);
1536 offset += size;
1537 length -= size;
1538 }
1539 if (length)
1540 ceph_zero_partial_page(inode, offset, length);
1541 }
1542
1543 static int ceph_zero_partial_object(struct inode *inode,
1544 loff_t offset, loff_t *length)
1545 {
1546 struct ceph_inode_info *ci = ceph_inode(inode);
1547 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1548 struct ceph_osd_request *req;
1549 int ret = 0;
1550 loff_t zero = 0;
1551 int op;
1552
1553 if (!length) {
1554 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1555 length = &zero;
1556 } else {
1557 op = CEPH_OSD_OP_ZERO;
1558 }
1559
1560 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1561 ceph_vino(inode),
1562 offset, length,
1563 0, 1, op,
1564 CEPH_OSD_FLAG_WRITE,
1565 NULL, 0, 0, false);
1566 if (IS_ERR(req)) {
1567 ret = PTR_ERR(req);
1568 goto out;
1569 }
1570
1571 req->r_mtime = inode->i_mtime;
1572 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1573 if (!ret) {
1574 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1575 if (ret == -ENOENT)
1576 ret = 0;
1577 }
1578 ceph_osdc_put_request(req);
1579
1580 out:
1581 return ret;
1582 }
1583
1584 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
1585 {
1586 int ret = 0;
1587 struct ceph_inode_info *ci = ceph_inode(inode);
1588 s32 stripe_unit = ci->i_layout.stripe_unit;
1589 s32 stripe_count = ci->i_layout.stripe_count;
1590 s32 object_size = ci->i_layout.object_size;
1591 u64 object_set_size = object_size * stripe_count;
1592 u64 nearly, t;
1593
1594 /* round offset up to next period boundary */
1595 nearly = offset + object_set_size - 1;
1596 t = nearly;
1597 nearly -= do_div(t, object_set_size);
1598
1599 while (length && offset < nearly) {
1600 loff_t size = length;
1601 ret = ceph_zero_partial_object(inode, offset, &size);
1602 if (ret < 0)
1603 return ret;
1604 offset += size;
1605 length -= size;
1606 }
1607 while (length >= object_set_size) {
1608 int i;
1609 loff_t pos = offset;
1610 for (i = 0; i < stripe_count; ++i) {
1611 ret = ceph_zero_partial_object(inode, pos, NULL);
1612 if (ret < 0)
1613 return ret;
1614 pos += stripe_unit;
1615 }
1616 offset += object_set_size;
1617 length -= object_set_size;
1618 }
1619 while (length) {
1620 loff_t size = length;
1621 ret = ceph_zero_partial_object(inode, offset, &size);
1622 if (ret < 0)
1623 return ret;
1624 offset += size;
1625 length -= size;
1626 }
1627 return ret;
1628 }
1629
1630 static long ceph_fallocate(struct file *file, int mode,
1631 loff_t offset, loff_t length)
1632 {
1633 struct ceph_file_info *fi = file->private_data;
1634 struct inode *inode = file_inode(file);
1635 struct ceph_inode_info *ci = ceph_inode(inode);
1636 struct ceph_osd_client *osdc =
1637 &ceph_inode_to_client(inode)->client->osdc;
1638 struct ceph_cap_flush *prealloc_cf;
1639 int want, got = 0;
1640 int dirty;
1641 int ret = 0;
1642 loff_t endoff = 0;
1643 loff_t size;
1644
1645 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1646 return -EOPNOTSUPP;
1647
1648 if (!S_ISREG(inode->i_mode))
1649 return -EOPNOTSUPP;
1650
1651 prealloc_cf = ceph_alloc_cap_flush();
1652 if (!prealloc_cf)
1653 return -ENOMEM;
1654
1655 inode_lock(inode);
1656
1657 if (ceph_snap(inode) != CEPH_NOSNAP) {
1658 ret = -EROFS;
1659 goto unlock;
1660 }
1661
1662 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) &&
1663 !(mode & FALLOC_FL_PUNCH_HOLE)) {
1664 ret = -ENOSPC;
1665 goto unlock;
1666 }
1667
1668 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1669 ret = ceph_uninline_data(file, NULL);
1670 if (ret < 0)
1671 goto unlock;
1672 }
1673
1674 size = i_size_read(inode);
1675 if (!(mode & FALLOC_FL_KEEP_SIZE)) {
1676 endoff = offset + length;
1677 ret = inode_newsize_ok(inode, endoff);
1678 if (ret)
1679 goto unlock;
1680 }
1681
1682 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1683 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1684 else
1685 want = CEPH_CAP_FILE_BUFFER;
1686
1687 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, endoff, &got, NULL);
1688 if (ret < 0)
1689 goto unlock;
1690
1691 if (mode & FALLOC_FL_PUNCH_HOLE) {
1692 if (offset < size)
1693 ceph_zero_pagecache_range(inode, offset, length);
1694 ret = ceph_zero_objects(inode, offset, length);
1695 } else if (endoff > size) {
1696 truncate_pagecache_range(inode, size, -1);
1697 if (ceph_inode_set_size(inode, endoff))
1698 ceph_check_caps(ceph_inode(inode),
1699 CHECK_CAPS_AUTHONLY, NULL);
1700 }
1701
1702 if (!ret) {
1703 spin_lock(&ci->i_ceph_lock);
1704 ci->i_inline_version = CEPH_INLINE_NONE;
1705 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1706 &prealloc_cf);
1707 spin_unlock(&ci->i_ceph_lock);
1708 if (dirty)
1709 __mark_inode_dirty(inode, dirty);
1710 }
1711
1712 ceph_put_cap_refs(ci, got);
1713 unlock:
1714 inode_unlock(inode);
1715 ceph_free_cap_flush(prealloc_cf);
1716 return ret;
1717 }
1718
1719 const struct file_operations ceph_file_fops = {
1720 .open = ceph_open,
1721 .release = ceph_release,
1722 .llseek = ceph_llseek,
1723 .read_iter = ceph_read_iter,
1724 .write_iter = ceph_write_iter,
1725 .mmap = ceph_mmap,
1726 .fsync = ceph_fsync,
1727 .lock = ceph_lock,
1728 .flock = ceph_flock,
1729 .splice_read = generic_file_splice_read,
1730 .splice_write = iter_file_splice_write,
1731 .unlocked_ioctl = ceph_ioctl,
1732 .compat_ioctl = ceph_ioctl,
1733 .fallocate = ceph_fallocate,
1734 };
1735
Linux with added FPC-III support