Loading...
1#include <linux/ceph/ceph_debug.h>
2
3#include <linux/module.h>
4#include <linux/sched.h>
5#include <linux/slab.h>
6#include <linux/file.h>
7#include <linux/mount.h>
8#include <linux/namei.h>
9#include <linux/writeback.h>
10#include <linux/falloc.h>
11
12#include "super.h"
13#include "mds_client.h"
14#include "cache.h"
15
16/*
17 * Ceph file operations
18 *
19 * Implement basic open/close functionality, and implement
20 * read/write.
21 *
22 * We implement three modes of file I/O:
23 * - buffered uses the generic_file_aio_{read,write} helpers
24 *
25 * - synchronous is used when there is multi-client read/write
26 * sharing, avoids the page cache, and synchronously waits for an
27 * ack from the OSD.
28 *
29 * - direct io takes the variant of the sync path that references
30 * user pages directly.
31 *
32 * fsync() flushes and waits on dirty pages, but just queues metadata
33 * for writeback: since the MDS can recover size and mtime there is no
34 * need to wait for MDS acknowledgement.
35 */
36
37/*
38 * Calculate the length sum of direct io vectors that can
39 * be combined into one page vector.
40 */
41static size_t dio_get_pagev_size(const struct iov_iter *it)
42{
43 const struct iovec *iov = it->iov;
44 const struct iovec *iovend = iov + it->nr_segs;
45 size_t size;
46
47 size = iov->iov_len - it->iov_offset;
48 /*
49 * An iov can be page vectored when both the current tail
50 * and the next base are page aligned.
51 */
52 while (PAGE_ALIGNED((iov->iov_base + iov->iov_len)) &&
53 (++iov < iovend && PAGE_ALIGNED((iov->iov_base)))) {
54 size += iov->iov_len;
55 }
56 dout("dio_get_pagevlen len = %zu\n", size);
57 return size;
58}
59
60/*
61 * Allocate a page vector based on (@it, @nbytes).
62 * The return value is the tuple describing a page vector,
63 * that is (@pages, @page_align, @num_pages).
64 */
65static struct page **
66dio_get_pages_alloc(const struct iov_iter *it, size_t nbytes,
67 size_t *page_align, int *num_pages)
68{
69 struct iov_iter tmp_it = *it;
70 size_t align;
71 struct page **pages;
72 int ret = 0, idx, npages;
73
74 align = (unsigned long)(it->iov->iov_base + it->iov_offset) &
75 (PAGE_SIZE - 1);
76 npages = calc_pages_for(align, nbytes);
77 pages = kmalloc(sizeof(*pages) * npages, GFP_KERNEL);
78 if (!pages) {
79 pages = vmalloc(sizeof(*pages) * npages);
80 if (!pages)
81 return ERR_PTR(-ENOMEM);
82 }
83
84 for (idx = 0; idx < npages; ) {
85 size_t start;
86 ret = iov_iter_get_pages(&tmp_it, pages + idx, nbytes,
87 npages - idx, &start);
88 if (ret < 0)
89 goto fail;
90
91 iov_iter_advance(&tmp_it, ret);
92 nbytes -= ret;
93 idx += (ret + start + PAGE_SIZE - 1) / PAGE_SIZE;
94 }
95
96 BUG_ON(nbytes != 0);
97 *num_pages = npages;
98 *page_align = align;
99 dout("dio_get_pages_alloc: got %d pages align %zu\n", npages, align);
100 return pages;
101fail:
102 ceph_put_page_vector(pages, idx, false);
103 return ERR_PTR(ret);
104}
105
106/*
107 * Prepare an open request. Preallocate ceph_cap to avoid an
108 * inopportune ENOMEM later.
109 */
110static struct ceph_mds_request *
111prepare_open_request(struct super_block *sb, int flags, int create_mode)
112{
113 struct ceph_fs_client *fsc = ceph_sb_to_client(sb);
114 struct ceph_mds_client *mdsc = fsc->mdsc;
115 struct ceph_mds_request *req;
116 int want_auth = USE_ANY_MDS;
117 int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
118
119 if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
120 want_auth = USE_AUTH_MDS;
121
122 req = ceph_mdsc_create_request(mdsc, op, want_auth);
123 if (IS_ERR(req))
124 goto out;
125 req->r_fmode = ceph_flags_to_mode(flags);
126 req->r_args.open.flags = cpu_to_le32(flags);
127 req->r_args.open.mode = cpu_to_le32(create_mode);
128out:
129 return req;
130}
131
132/*
133 * initialize private struct file data.
134 * if we fail, clean up by dropping fmode reference on the ceph_inode
135 */
136static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
137{
138 struct ceph_file_info *cf;
139 int ret = 0;
140 struct ceph_inode_info *ci = ceph_inode(inode);
141 struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
142 struct ceph_mds_client *mdsc = fsc->mdsc;
143
144 switch (inode->i_mode & S_IFMT) {
145 case S_IFREG:
146 /* First file open request creates the cookie, we want to keep
147 * this cookie around for the filetime of the inode as not to
148 * have to worry about fscache register / revoke / operation
149 * races.
150 *
151 * Also, if we know the operation is going to invalidate data
152 * (non readonly) just nuke the cache right away.
153 */
154 ceph_fscache_register_inode_cookie(mdsc->fsc, ci);
155 if ((fmode & CEPH_FILE_MODE_WR))
156 ceph_fscache_invalidate(inode);
157 case S_IFDIR:
158 dout("init_file %p %p 0%o (regular)\n", inode, file,
159 inode->i_mode);
160 cf = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
161 if (cf == NULL) {
162 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
163 return -ENOMEM;
164 }
165 cf->fmode = fmode;
166 cf->next_offset = 2;
167 cf->readdir_cache_idx = -1;
168 file->private_data = cf;
169 BUG_ON(inode->i_fop->release != ceph_release);
170 break;
171
172 case S_IFLNK:
173 dout("init_file %p %p 0%o (symlink)\n", inode, file,
174 inode->i_mode);
175 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
176 break;
177
178 default:
179 dout("init_file %p %p 0%o (special)\n", inode, file,
180 inode->i_mode);
181 /*
182 * we need to drop the open ref now, since we don't
183 * have .release set to ceph_release.
184 */
185 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
186 BUG_ON(inode->i_fop->release == ceph_release);
187
188 /* call the proper open fop */
189 ret = inode->i_fop->open(inode, file);
190 }
191 return ret;
192}
193
194/*
195 * If we already have the requisite capabilities, we can satisfy
196 * the open request locally (no need to request new caps from the
197 * MDS). We do, however, need to inform the MDS (asynchronously)
198 * if our wanted caps set expands.
199 */
200int ceph_open(struct inode *inode, struct file *file)
201{
202 struct ceph_inode_info *ci = ceph_inode(inode);
203 struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
204 struct ceph_mds_client *mdsc = fsc->mdsc;
205 struct ceph_mds_request *req;
206 struct ceph_file_info *cf = file->private_data;
207 int err;
208 int flags, fmode, wanted;
209
210 if (cf) {
211 dout("open file %p is already opened\n", file);
212 return 0;
213 }
214
215 /* filter out O_CREAT|O_EXCL; vfs did that already. yuck. */
216 flags = file->f_flags & ~(O_CREAT|O_EXCL);
217 if (S_ISDIR(inode->i_mode))
218 flags = O_DIRECTORY; /* mds likes to know */
219
220 dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
221 ceph_vinop(inode), file, flags, file->f_flags);
222 fmode = ceph_flags_to_mode(flags);
223 wanted = ceph_caps_for_mode(fmode);
224
225 /* snapped files are read-only */
226 if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
227 return -EROFS;
228
229 /* trivially open snapdir */
230 if (ceph_snap(inode) == CEPH_SNAPDIR) {
231 spin_lock(&ci->i_ceph_lock);
232 __ceph_get_fmode(ci, fmode);
233 spin_unlock(&ci->i_ceph_lock);
234 return ceph_init_file(inode, file, fmode);
235 }
236
237 /*
238 * No need to block if we have caps on the auth MDS (for
239 * write) or any MDS (for read). Update wanted set
240 * asynchronously.
241 */
242 spin_lock(&ci->i_ceph_lock);
243 if (__ceph_is_any_real_caps(ci) &&
244 (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
245 int mds_wanted = __ceph_caps_mds_wanted(ci);
246 int issued = __ceph_caps_issued(ci, NULL);
247
248 dout("open %p fmode %d want %s issued %s using existing\n",
249 inode, fmode, ceph_cap_string(wanted),
250 ceph_cap_string(issued));
251 __ceph_get_fmode(ci, fmode);
252 spin_unlock(&ci->i_ceph_lock);
253
254 /* adjust wanted? */
255 if ((issued & wanted) != wanted &&
256 (mds_wanted & wanted) != wanted &&
257 ceph_snap(inode) != CEPH_SNAPDIR)
258 ceph_check_caps(ci, 0, NULL);
259
260 return ceph_init_file(inode, file, fmode);
261 } else if (ceph_snap(inode) != CEPH_NOSNAP &&
262 (ci->i_snap_caps & wanted) == wanted) {
263 __ceph_get_fmode(ci, fmode);
264 spin_unlock(&ci->i_ceph_lock);
265 return ceph_init_file(inode, file, fmode);
266 }
267
268 spin_unlock(&ci->i_ceph_lock);
269
270 dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
271 req = prepare_open_request(inode->i_sb, flags, 0);
272 if (IS_ERR(req)) {
273 err = PTR_ERR(req);
274 goto out;
275 }
276 req->r_inode = inode;
277 ihold(inode);
278
279 req->r_num_caps = 1;
280 err = ceph_mdsc_do_request(mdsc, NULL, req);
281 if (!err)
282 err = ceph_init_file(inode, file, req->r_fmode);
283 ceph_mdsc_put_request(req);
284 dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
285out:
286 return err;
287}
288
289
290/*
291 * Do a lookup + open with a single request. If we get a non-existent
292 * file or symlink, return 1 so the VFS can retry.
293 */
294int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
295 struct file *file, unsigned flags, umode_t mode,
296 int *opened)
297{
298 struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
299 struct ceph_mds_client *mdsc = fsc->mdsc;
300 struct ceph_mds_request *req;
301 struct dentry *dn;
302 struct ceph_acls_info acls = {};
303 int mask;
304 int err;
305
306 dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
307 dir, dentry, dentry,
308 d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
309
310 if (dentry->d_name.len > NAME_MAX)
311 return -ENAMETOOLONG;
312
313 err = ceph_init_dentry(dentry);
314 if (err < 0)
315 return err;
316
317 if (flags & O_CREAT) {
318 err = ceph_pre_init_acls(dir, &mode, &acls);
319 if (err < 0)
320 return err;
321 }
322
323 /* do the open */
324 req = prepare_open_request(dir->i_sb, flags, mode);
325 if (IS_ERR(req)) {
326 err = PTR_ERR(req);
327 goto out_acl;
328 }
329 req->r_dentry = dget(dentry);
330 req->r_num_caps = 2;
331 if (flags & O_CREAT) {
332 req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
333 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
334 if (acls.pagelist) {
335 req->r_pagelist = acls.pagelist;
336 acls.pagelist = NULL;
337 }
338 }
339
340 mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
341 if (ceph_security_xattr_wanted(dir))
342 mask |= CEPH_CAP_XATTR_SHARED;
343 req->r_args.open.mask = cpu_to_le32(mask);
344
345 req->r_locked_dir = dir; /* caller holds dir->i_mutex */
346 err = ceph_mdsc_do_request(mdsc,
347 (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
348 req);
349 err = ceph_handle_snapdir(req, dentry, err);
350 if (err)
351 goto out_req;
352
353 if ((flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
354 err = ceph_handle_notrace_create(dir, dentry);
355
356 if (d_unhashed(dentry)) {
357 dn = ceph_finish_lookup(req, dentry, err);
358 if (IS_ERR(dn))
359 err = PTR_ERR(dn);
360 } else {
361 /* we were given a hashed negative dentry */
362 dn = NULL;
363 }
364 if (err)
365 goto out_req;
366 if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
367 /* make vfs retry on splice, ENOENT, or symlink */
368 dout("atomic_open finish_no_open on dn %p\n", dn);
369 err = finish_no_open(file, dn);
370 } else {
371 dout("atomic_open finish_open on dn %p\n", dn);
372 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
373 ceph_init_inode_acls(d_inode(dentry), &acls);
374 *opened |= FILE_CREATED;
375 }
376 err = finish_open(file, dentry, ceph_open, opened);
377 }
378out_req:
379 if (!req->r_err && req->r_target_inode)
380 ceph_put_fmode(ceph_inode(req->r_target_inode), req->r_fmode);
381 ceph_mdsc_put_request(req);
382out_acl:
383 ceph_release_acls_info(&acls);
384 dout("atomic_open result=%d\n", err);
385 return err;
386}
387
388int ceph_release(struct inode *inode, struct file *file)
389{
390 struct ceph_inode_info *ci = ceph_inode(inode);
391 struct ceph_file_info *cf = file->private_data;
392
393 dout("release inode %p file %p\n", inode, file);
394 ceph_put_fmode(ci, cf->fmode);
395 if (cf->last_readdir)
396 ceph_mdsc_put_request(cf->last_readdir);
397 kfree(cf->last_name);
398 kfree(cf->dir_info);
399 kmem_cache_free(ceph_file_cachep, cf);
400
401 /* wake up anyone waiting for caps on this inode */
402 wake_up_all(&ci->i_cap_wq);
403 return 0;
404}
405
406enum {
407 HAVE_RETRIED = 1,
408 CHECK_EOF = 2,
409 READ_INLINE = 3,
410};
411
412/*
413 * Read a range of bytes striped over one or more objects. Iterate over
414 * objects we stripe over. (That's not atomic, but good enough for now.)
415 *
416 * If we get a short result from the OSD, check against i_size; we need to
417 * only return a short read to the caller if we hit EOF.
418 */
419static int striped_read(struct inode *inode,
420 u64 off, u64 len,
421 struct page **pages, int num_pages,
422 int *checkeof)
423{
424 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
425 struct ceph_inode_info *ci = ceph_inode(inode);
426 u64 pos, this_len, left;
427 loff_t i_size;
428 int page_align, pages_left;
429 int read, ret;
430 struct page **page_pos;
431 bool hit_stripe, was_short;
432
433 /*
434 * we may need to do multiple reads. not atomic, unfortunately.
435 */
436 pos = off;
437 left = len;
438 page_pos = pages;
439 pages_left = num_pages;
440 read = 0;
441
442more:
443 page_align = pos & ~PAGE_MASK;
444 this_len = left;
445 ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode),
446 &ci->i_layout, pos, &this_len,
447 ci->i_truncate_seq,
448 ci->i_truncate_size,
449 page_pos, pages_left, page_align);
450 if (ret == -ENOENT)
451 ret = 0;
452 hit_stripe = this_len < left;
453 was_short = ret >= 0 && ret < this_len;
454 dout("striped_read %llu~%llu (read %u) got %d%s%s\n", pos, left, read,
455 ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : "");
456
457 i_size = i_size_read(inode);
458 if (ret >= 0) {
459 int didpages;
460 if (was_short && (pos + ret < i_size)) {
461 int zlen = min(this_len - ret, i_size - pos - ret);
462 int zoff = (off & ~PAGE_MASK) + read + ret;
463 dout(" zero gap %llu to %llu\n",
464 pos + ret, pos + ret + zlen);
465 ceph_zero_page_vector_range(zoff, zlen, pages);
466 ret += zlen;
467 }
468
469 didpages = (page_align + ret) >> PAGE_SHIFT;
470 pos += ret;
471 read = pos - off;
472 left -= ret;
473 page_pos += didpages;
474 pages_left -= didpages;
475
476 /* hit stripe and need continue*/
477 if (left && hit_stripe && pos < i_size)
478 goto more;
479 }
480
481 if (read > 0) {
482 ret = read;
483 /* did we bounce off eof? */
484 if (pos + left > i_size)
485 *checkeof = CHECK_EOF;
486 }
487
488 dout("striped_read returns %d\n", ret);
489 return ret;
490}
491
492/*
493 * Completely synchronous read and write methods. Direct from __user
494 * buffer to osd, or directly to user pages (if O_DIRECT).
495 *
496 * If the read spans object boundary, just do multiple reads.
497 */
498static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *i,
499 int *checkeof)
500{
501 struct file *file = iocb->ki_filp;
502 struct inode *inode = file_inode(file);
503 struct page **pages;
504 u64 off = iocb->ki_pos;
505 int num_pages, ret;
506 size_t len = iov_iter_count(i);
507
508 dout("sync_read on file %p %llu~%u %s\n", file, off,
509 (unsigned)len,
510 (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
511
512 if (!len)
513 return 0;
514 /*
515 * flush any page cache pages in this range. this
516 * will make concurrent normal and sync io slow,
517 * but it will at least behave sensibly when they are
518 * in sequence.
519 */
520 ret = filemap_write_and_wait_range(inode->i_mapping, off,
521 off + len);
522 if (ret < 0)
523 return ret;
524
525 num_pages = calc_pages_for(off, len);
526 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
527 if (IS_ERR(pages))
528 return PTR_ERR(pages);
529 ret = striped_read(inode, off, len, pages,
530 num_pages, checkeof);
531 if (ret > 0) {
532 int l, k = 0;
533 size_t left = ret;
534
535 while (left) {
536 size_t page_off = off & ~PAGE_MASK;
537 size_t copy = min_t(size_t, left,
538 PAGE_SIZE - page_off);
539 l = copy_page_to_iter(pages[k++], page_off, copy, i);
540 off += l;
541 left -= l;
542 if (l < copy)
543 break;
544 }
545 }
546 ceph_release_page_vector(pages, num_pages);
547
548 if (off > iocb->ki_pos) {
549 ret = off - iocb->ki_pos;
550 iocb->ki_pos = off;
551 }
552
553 dout("sync_read result %d\n", ret);
554 return ret;
555}
556
557struct ceph_aio_request {
558 struct kiocb *iocb;
559 size_t total_len;
560 int write;
561 int error;
562 struct list_head osd_reqs;
563 unsigned num_reqs;
564 atomic_t pending_reqs;
565 struct timespec mtime;
566 struct ceph_cap_flush *prealloc_cf;
567};
568
569struct ceph_aio_work {
570 struct work_struct work;
571 struct ceph_osd_request *req;
572};
573
574static void ceph_aio_retry_work(struct work_struct *work);
575
576static void ceph_aio_complete(struct inode *inode,
577 struct ceph_aio_request *aio_req)
578{
579 struct ceph_inode_info *ci = ceph_inode(inode);
580 int ret;
581
582 if (!atomic_dec_and_test(&aio_req->pending_reqs))
583 return;
584
585 ret = aio_req->error;
586 if (!ret)
587 ret = aio_req->total_len;
588
589 dout("ceph_aio_complete %p rc %d\n", inode, ret);
590
591 if (ret >= 0 && aio_req->write) {
592 int dirty;
593
594 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
595 if (endoff > i_size_read(inode)) {
596 if (ceph_inode_set_size(inode, endoff))
597 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
598 }
599
600 spin_lock(&ci->i_ceph_lock);
601 ci->i_inline_version = CEPH_INLINE_NONE;
602 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
603 &aio_req->prealloc_cf);
604 spin_unlock(&ci->i_ceph_lock);
605 if (dirty)
606 __mark_inode_dirty(inode, dirty);
607
608 }
609
610 ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
611 CEPH_CAP_FILE_RD));
612
613 aio_req->iocb->ki_complete(aio_req->iocb, ret, 0);
614
615 ceph_free_cap_flush(aio_req->prealloc_cf);
616 kfree(aio_req);
617}
618
619static void ceph_aio_complete_req(struct ceph_osd_request *req,
620 struct ceph_msg *msg)
621{
622 int rc = req->r_result;
623 struct inode *inode = req->r_inode;
624 struct ceph_aio_request *aio_req = req->r_priv;
625 struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
626 int num_pages = calc_pages_for((u64)osd_data->alignment,
627 osd_data->length);
628
629 dout("ceph_aio_complete_req %p rc %d bytes %llu\n",
630 inode, rc, osd_data->length);
631
632 if (rc == -EOLDSNAPC) {
633 struct ceph_aio_work *aio_work;
634 BUG_ON(!aio_req->write);
635
636 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
637 if (aio_work) {
638 INIT_WORK(&aio_work->work, ceph_aio_retry_work);
639 aio_work->req = req;
640 queue_work(ceph_inode_to_client(inode)->wb_wq,
641 &aio_work->work);
642 return;
643 }
644 rc = -ENOMEM;
645 } else if (!aio_req->write) {
646 if (rc == -ENOENT)
647 rc = 0;
648 if (rc >= 0 && osd_data->length > rc) {
649 int zoff = osd_data->alignment + rc;
650 int zlen = osd_data->length - rc;
651 /*
652 * If read is satisfied by single OSD request,
653 * it can pass EOF. Otherwise read is within
654 * i_size.
655 */
656 if (aio_req->num_reqs == 1) {
657 loff_t i_size = i_size_read(inode);
658 loff_t endoff = aio_req->iocb->ki_pos + rc;
659 if (endoff < i_size)
660 zlen = min_t(size_t, zlen,
661 i_size - endoff);
662 aio_req->total_len = rc + zlen;
663 }
664
665 if (zlen > 0)
666 ceph_zero_page_vector_range(zoff, zlen,
667 osd_data->pages);
668 }
669 }
670
671 ceph_put_page_vector(osd_data->pages, num_pages, false);
672 ceph_osdc_put_request(req);
673
674 if (rc < 0)
675 cmpxchg(&aio_req->error, 0, rc);
676
677 ceph_aio_complete(inode, aio_req);
678 return;
679}
680
681static void ceph_aio_retry_work(struct work_struct *work)
682{
683 struct ceph_aio_work *aio_work =
684 container_of(work, struct ceph_aio_work, work);
685 struct ceph_osd_request *orig_req = aio_work->req;
686 struct ceph_aio_request *aio_req = orig_req->r_priv;
687 struct inode *inode = orig_req->r_inode;
688 struct ceph_inode_info *ci = ceph_inode(inode);
689 struct ceph_snap_context *snapc;
690 struct ceph_osd_request *req;
691 int ret;
692
693 spin_lock(&ci->i_ceph_lock);
694 if (__ceph_have_pending_cap_snap(ci)) {
695 struct ceph_cap_snap *capsnap =
696 list_last_entry(&ci->i_cap_snaps,
697 struct ceph_cap_snap,
698 ci_item);
699 snapc = ceph_get_snap_context(capsnap->context);
700 } else {
701 BUG_ON(!ci->i_head_snapc);
702 snapc = ceph_get_snap_context(ci->i_head_snapc);
703 }
704 spin_unlock(&ci->i_ceph_lock);
705
706 req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 2,
707 false, GFP_NOFS);
708 if (!req) {
709 ret = -ENOMEM;
710 req = orig_req;
711 goto out;
712 }
713
714 req->r_flags = CEPH_OSD_FLAG_ORDERSNAP |
715 CEPH_OSD_FLAG_ONDISK |
716 CEPH_OSD_FLAG_WRITE;
717 req->r_base_oloc = orig_req->r_base_oloc;
718 req->r_base_oid = orig_req->r_base_oid;
719
720 req->r_ops[0] = orig_req->r_ops[0];
721 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC, 0);
722
723 ceph_osdc_build_request(req, req->r_ops[0].extent.offset,
724 snapc, CEPH_NOSNAP, &aio_req->mtime);
725
726 ceph_osdc_put_request(orig_req);
727
728 req->r_callback = ceph_aio_complete_req;
729 req->r_inode = inode;
730 req->r_priv = aio_req;
731
732 ret = ceph_osdc_start_request(req->r_osdc, req, false);
733out:
734 if (ret < 0) {
735 req->r_result = ret;
736 ceph_aio_complete_req(req, NULL);
737 }
738
739 ceph_put_snap_context(snapc);
740 kfree(aio_work);
741}
742
743/*
744 * Write commit request unsafe callback, called to tell us when a
745 * request is unsafe (that is, in flight--has been handed to the
746 * messenger to send to its target osd). It is called again when
747 * we've received a response message indicating the request is
748 * "safe" (its CEPH_OSD_FLAG_ONDISK flag is set), or when a request
749 * is completed early (and unsuccessfully) due to a timeout or
750 * interrupt.
751 *
752 * This is used if we requested both an ACK and ONDISK commit reply
753 * from the OSD.
754 */
755static void ceph_sync_write_unsafe(struct ceph_osd_request *req, bool unsafe)
756{
757 struct ceph_inode_info *ci = ceph_inode(req->r_inode);
758
759 dout("%s %p tid %llu %ssafe\n", __func__, req, req->r_tid,
760 unsafe ? "un" : "");
761 if (unsafe) {
762 ceph_get_cap_refs(ci, CEPH_CAP_FILE_WR);
763 spin_lock(&ci->i_unsafe_lock);
764 list_add_tail(&req->r_unsafe_item,
765 &ci->i_unsafe_writes);
766 spin_unlock(&ci->i_unsafe_lock);
767 } else {
768 spin_lock(&ci->i_unsafe_lock);
769 list_del_init(&req->r_unsafe_item);
770 spin_unlock(&ci->i_unsafe_lock);
771 ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR);
772 }
773}
774
775
776static ssize_t
777ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
778 struct ceph_snap_context *snapc,
779 struct ceph_cap_flush **pcf)
780{
781 struct file *file = iocb->ki_filp;
782 struct inode *inode = file_inode(file);
783 struct ceph_inode_info *ci = ceph_inode(inode);
784 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
785 struct ceph_vino vino;
786 struct ceph_osd_request *req;
787 struct page **pages;
788 struct ceph_aio_request *aio_req = NULL;
789 int num_pages = 0;
790 int flags;
791 int ret;
792 struct timespec mtime = current_fs_time(inode->i_sb);
793 size_t count = iov_iter_count(iter);
794 loff_t pos = iocb->ki_pos;
795 bool write = iov_iter_rw(iter) == WRITE;
796
797 if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
798 return -EROFS;
799
800 dout("sync_direct_read_write (%s) on file %p %lld~%u\n",
801 (write ? "write" : "read"), file, pos, (unsigned)count);
802
803 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
804 if (ret < 0)
805 return ret;
806
807 if (write) {
808 ret = invalidate_inode_pages2_range(inode->i_mapping,
809 pos >> PAGE_SHIFT,
810 (pos + count) >> PAGE_SHIFT);
811 if (ret < 0)
812 dout("invalidate_inode_pages2_range returned %d\n", ret);
813
814 flags = CEPH_OSD_FLAG_ORDERSNAP |
815 CEPH_OSD_FLAG_ONDISK |
816 CEPH_OSD_FLAG_WRITE;
817 } else {
818 flags = CEPH_OSD_FLAG_READ;
819 }
820
821 while (iov_iter_count(iter) > 0) {
822 u64 size = dio_get_pagev_size(iter);
823 size_t start = 0;
824 ssize_t len;
825
826 vino = ceph_vino(inode);
827 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
828 vino, pos, &size, 0,
829 /*include a 'startsync' command*/
830 write ? 2 : 1,
831 write ? CEPH_OSD_OP_WRITE :
832 CEPH_OSD_OP_READ,
833 flags, snapc,
834 ci->i_truncate_seq,
835 ci->i_truncate_size,
836 false);
837 if (IS_ERR(req)) {
838 ret = PTR_ERR(req);
839 break;
840 }
841
842 len = size;
843 pages = dio_get_pages_alloc(iter, len, &start, &num_pages);
844 if (IS_ERR(pages)) {
845 ceph_osdc_put_request(req);
846 ret = PTR_ERR(pages);
847 break;
848 }
849
850 /*
851 * To simplify error handling, allow AIO when IO within i_size
852 * or IO can be satisfied by single OSD request.
853 */
854 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
855 (len == count || pos + count <= i_size_read(inode))) {
856 aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
857 if (aio_req) {
858 aio_req->iocb = iocb;
859 aio_req->write = write;
860 INIT_LIST_HEAD(&aio_req->osd_reqs);
861 if (write) {
862 aio_req->mtime = mtime;
863 swap(aio_req->prealloc_cf, *pcf);
864 }
865 }
866 /* ignore error */
867 }
868
869 if (write) {
870 /*
871 * throw out any page cache pages in this range. this
872 * may block.
873 */
874 truncate_inode_pages_range(inode->i_mapping, pos,
875 (pos+len) | (PAGE_SIZE - 1));
876
877 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC, 0);
878 }
879
880
881 osd_req_op_extent_osd_data_pages(req, 0, pages, len, start,
882 false, false);
883
884 ceph_osdc_build_request(req, pos, snapc, vino.snap, &mtime);
885
886 if (aio_req) {
887 aio_req->total_len += len;
888 aio_req->num_reqs++;
889 atomic_inc(&aio_req->pending_reqs);
890
891 req->r_callback = ceph_aio_complete_req;
892 req->r_inode = inode;
893 req->r_priv = aio_req;
894 list_add_tail(&req->r_unsafe_item, &aio_req->osd_reqs);
895
896 pos += len;
897 iov_iter_advance(iter, len);
898 continue;
899 }
900
901 ret = ceph_osdc_start_request(req->r_osdc, req, false);
902 if (!ret)
903 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
904
905 size = i_size_read(inode);
906 if (!write) {
907 if (ret == -ENOENT)
908 ret = 0;
909 if (ret >= 0 && ret < len && pos + ret < size) {
910 int zlen = min_t(size_t, len - ret,
911 size - pos - ret);
912 ceph_zero_page_vector_range(start + ret, zlen,
913 pages);
914 ret += zlen;
915 }
916 if (ret >= 0)
917 len = ret;
918 }
919
920 ceph_put_page_vector(pages, num_pages, false);
921
922 ceph_osdc_put_request(req);
923 if (ret < 0)
924 break;
925
926 pos += len;
927 iov_iter_advance(iter, len);
928
929 if (!write && pos >= size)
930 break;
931
932 if (write && pos > size) {
933 if (ceph_inode_set_size(inode, pos))
934 ceph_check_caps(ceph_inode(inode),
935 CHECK_CAPS_AUTHONLY,
936 NULL);
937 }
938 }
939
940 if (aio_req) {
941 if (aio_req->num_reqs == 0) {
942 kfree(aio_req);
943 return ret;
944 }
945
946 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
947 CEPH_CAP_FILE_RD);
948
949 while (!list_empty(&aio_req->osd_reqs)) {
950 req = list_first_entry(&aio_req->osd_reqs,
951 struct ceph_osd_request,
952 r_unsafe_item);
953 list_del_init(&req->r_unsafe_item);
954 if (ret >= 0)
955 ret = ceph_osdc_start_request(req->r_osdc,
956 req, false);
957 if (ret < 0) {
958 req->r_result = ret;
959 ceph_aio_complete_req(req, NULL);
960 }
961 }
962 return -EIOCBQUEUED;
963 }
964
965 if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
966 ret = pos - iocb->ki_pos;
967 iocb->ki_pos = pos;
968 }
969 return ret;
970}
971
972/*
973 * Synchronous write, straight from __user pointer or user pages.
974 *
975 * If write spans object boundary, just do multiple writes. (For a
976 * correct atomic write, we should e.g. take write locks on all
977 * objects, rollback on failure, etc.)
978 */
979static ssize_t
980ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
981 struct ceph_snap_context *snapc)
982{
983 struct file *file = iocb->ki_filp;
984 struct inode *inode = file_inode(file);
985 struct ceph_inode_info *ci = ceph_inode(inode);
986 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
987 struct ceph_vino vino;
988 struct ceph_osd_request *req;
989 struct page **pages;
990 u64 len;
991 int num_pages;
992 int written = 0;
993 int flags;
994 int check_caps = 0;
995 int ret;
996 struct timespec mtime = current_fs_time(inode->i_sb);
997 size_t count = iov_iter_count(from);
998
999 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1000 return -EROFS;
1001
1002 dout("sync_write on file %p %lld~%u\n", file, pos, (unsigned)count);
1003
1004 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
1005 if (ret < 0)
1006 return ret;
1007
1008 ret = invalidate_inode_pages2_range(inode->i_mapping,
1009 pos >> PAGE_SHIFT,
1010 (pos + count) >> PAGE_SHIFT);
1011 if (ret < 0)
1012 dout("invalidate_inode_pages2_range returned %d\n", ret);
1013
1014 flags = CEPH_OSD_FLAG_ORDERSNAP |
1015 CEPH_OSD_FLAG_ONDISK |
1016 CEPH_OSD_FLAG_WRITE |
1017 CEPH_OSD_FLAG_ACK;
1018
1019 while ((len = iov_iter_count(from)) > 0) {
1020 size_t left;
1021 int n;
1022
1023 vino = ceph_vino(inode);
1024 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1025 vino, pos, &len, 0, 1,
1026 CEPH_OSD_OP_WRITE, flags, snapc,
1027 ci->i_truncate_seq,
1028 ci->i_truncate_size,
1029 false);
1030 if (IS_ERR(req)) {
1031 ret = PTR_ERR(req);
1032 break;
1033 }
1034
1035 /*
1036 * write from beginning of first page,
1037 * regardless of io alignment
1038 */
1039 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1040
1041 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1042 if (IS_ERR(pages)) {
1043 ret = PTR_ERR(pages);
1044 goto out;
1045 }
1046
1047 left = len;
1048 for (n = 0; n < num_pages; n++) {
1049 size_t plen = min_t(size_t, left, PAGE_SIZE);
1050 ret = copy_page_from_iter(pages[n], 0, plen, from);
1051 if (ret != plen) {
1052 ret = -EFAULT;
1053 break;
1054 }
1055 left -= ret;
1056 }
1057
1058 if (ret < 0) {
1059 ceph_release_page_vector(pages, num_pages);
1060 goto out;
1061 }
1062
1063 /* get a second commit callback */
1064 req->r_unsafe_callback = ceph_sync_write_unsafe;
1065 req->r_inode = inode;
1066
1067 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1068 false, true);
1069
1070 /* BUG_ON(vino.snap != CEPH_NOSNAP); */
1071 ceph_osdc_build_request(req, pos, snapc, vino.snap, &mtime);
1072
1073 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1074 if (!ret)
1075 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1076
1077out:
1078 ceph_osdc_put_request(req);
1079 if (ret == 0) {
1080 pos += len;
1081 written += len;
1082
1083 if (pos > i_size_read(inode)) {
1084 check_caps = ceph_inode_set_size(inode, pos);
1085 if (check_caps)
1086 ceph_check_caps(ceph_inode(inode),
1087 CHECK_CAPS_AUTHONLY,
1088 NULL);
1089 }
1090 } else
1091 break;
1092 }
1093
1094 if (ret != -EOLDSNAPC && written > 0) {
1095 ret = written;
1096 iocb->ki_pos = pos;
1097 }
1098 return ret;
1099}
1100
1101/*
1102 * Wrap generic_file_aio_read with checks for cap bits on the inode.
1103 * Atomically grab references, so that those bits are not released
1104 * back to the MDS mid-read.
1105 *
1106 * Hmm, the sync read case isn't actually async... should it be?
1107 */
1108static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1109{
1110 struct file *filp = iocb->ki_filp;
1111 struct ceph_file_info *fi = filp->private_data;
1112 size_t len = iov_iter_count(to);
1113 struct inode *inode = file_inode(filp);
1114 struct ceph_inode_info *ci = ceph_inode(inode);
1115 struct page *pinned_page = NULL;
1116 ssize_t ret;
1117 int want, got = 0;
1118 int retry_op = 0, read = 0;
1119
1120again:
1121 dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1122 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1123
1124 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1125 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1126 else
1127 want = CEPH_CAP_FILE_CACHE;
1128 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1129 if (ret < 0)
1130 return ret;
1131
1132 if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1133 (iocb->ki_flags & IOCB_DIRECT) ||
1134 (fi->flags & CEPH_F_SYNC)) {
1135
1136 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1137 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1138 ceph_cap_string(got));
1139
1140 if (ci->i_inline_version == CEPH_INLINE_NONE) {
1141 if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1142 ret = ceph_direct_read_write(iocb, to,
1143 NULL, NULL);
1144 if (ret >= 0 && ret < len)
1145 retry_op = CHECK_EOF;
1146 } else {
1147 ret = ceph_sync_read(iocb, to, &retry_op);
1148 }
1149 } else {
1150 retry_op = READ_INLINE;
1151 }
1152 } else {
1153 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1154 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1155 ceph_cap_string(got));
1156
1157 ret = generic_file_read_iter(iocb, to);
1158 }
1159 dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1160 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1161 if (pinned_page) {
1162 put_page(pinned_page);
1163 pinned_page = NULL;
1164 }
1165 ceph_put_cap_refs(ci, got);
1166 if (retry_op > HAVE_RETRIED && ret >= 0) {
1167 int statret;
1168 struct page *page = NULL;
1169 loff_t i_size;
1170 if (retry_op == READ_INLINE) {
1171 page = __page_cache_alloc(GFP_KERNEL);
1172 if (!page)
1173 return -ENOMEM;
1174 }
1175
1176 statret = __ceph_do_getattr(inode, page,
1177 CEPH_STAT_CAP_INLINE_DATA, !!page);
1178 if (statret < 0) {
1179 __free_page(page);
1180 if (statret == -ENODATA) {
1181 BUG_ON(retry_op != READ_INLINE);
1182 goto again;
1183 }
1184 return statret;
1185 }
1186
1187 i_size = i_size_read(inode);
1188 if (retry_op == READ_INLINE) {
1189 BUG_ON(ret > 0 || read > 0);
1190 if (iocb->ki_pos < i_size &&
1191 iocb->ki_pos < PAGE_SIZE) {
1192 loff_t end = min_t(loff_t, i_size,
1193 iocb->ki_pos + len);
1194 end = min_t(loff_t, end, PAGE_SIZE);
1195 if (statret < end)
1196 zero_user_segment(page, statret, end);
1197 ret = copy_page_to_iter(page,
1198 iocb->ki_pos & ~PAGE_MASK,
1199 end - iocb->ki_pos, to);
1200 iocb->ki_pos += ret;
1201 read += ret;
1202 }
1203 if (iocb->ki_pos < i_size && read < len) {
1204 size_t zlen = min_t(size_t, len - read,
1205 i_size - iocb->ki_pos);
1206 ret = iov_iter_zero(zlen, to);
1207 iocb->ki_pos += ret;
1208 read += ret;
1209 }
1210 __free_pages(page, 0);
1211 return read;
1212 }
1213
1214 /* hit EOF or hole? */
1215 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1216 ret < len) {
1217 dout("sync_read hit hole, ppos %lld < size %lld"
1218 ", reading more\n", iocb->ki_pos, i_size);
1219
1220 read += ret;
1221 len -= ret;
1222 retry_op = HAVE_RETRIED;
1223 goto again;
1224 }
1225 }
1226
1227 if (ret >= 0)
1228 ret += read;
1229
1230 return ret;
1231}
1232
1233/*
1234 * Take cap references to avoid releasing caps to MDS mid-write.
1235 *
1236 * If we are synchronous, and write with an old snap context, the OSD
1237 * may return EOLDSNAPC. In that case, retry the write.. _after_
1238 * dropping our cap refs and allowing the pending snap to logically
1239 * complete _before_ this write occurs.
1240 *
1241 * If we are near ENOSPC, write synchronously.
1242 */
1243static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1244{
1245 struct file *file = iocb->ki_filp;
1246 struct ceph_file_info *fi = file->private_data;
1247 struct inode *inode = file_inode(file);
1248 struct ceph_inode_info *ci = ceph_inode(inode);
1249 struct ceph_osd_client *osdc =
1250 &ceph_sb_to_client(inode->i_sb)->client->osdc;
1251 struct ceph_cap_flush *prealloc_cf;
1252 ssize_t count, written = 0;
1253 int err, want, got;
1254 loff_t pos;
1255
1256 if (ceph_snap(inode) != CEPH_NOSNAP)
1257 return -EROFS;
1258
1259 prealloc_cf = ceph_alloc_cap_flush();
1260 if (!prealloc_cf)
1261 return -ENOMEM;
1262
1263 inode_lock(inode);
1264
1265 /* We can write back this queue in page reclaim */
1266 current->backing_dev_info = inode_to_bdi(inode);
1267
1268 if (iocb->ki_flags & IOCB_APPEND) {
1269 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1270 if (err < 0)
1271 goto out;
1272 }
1273
1274 err = generic_write_checks(iocb, from);
1275 if (err <= 0)
1276 goto out;
1277
1278 pos = iocb->ki_pos;
1279 count = iov_iter_count(from);
1280 err = file_remove_privs(file);
1281 if (err)
1282 goto out;
1283
1284 err = file_update_time(file);
1285 if (err)
1286 goto out;
1287
1288 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1289 err = ceph_uninline_data(file, NULL);
1290 if (err < 0)
1291 goto out;
1292 }
1293
1294retry_snap:
1295 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) {
1296 err = -ENOSPC;
1297 goto out;
1298 }
1299
1300 dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1301 inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1302 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1303 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1304 else
1305 want = CEPH_CAP_FILE_BUFFER;
1306 got = 0;
1307 err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, pos + count,
1308 &got, NULL);
1309 if (err < 0)
1310 goto out;
1311
1312 dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1313 inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1314
1315 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1316 (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC)) {
1317 struct ceph_snap_context *snapc;
1318 struct iov_iter data;
1319 inode_unlock(inode);
1320
1321 spin_lock(&ci->i_ceph_lock);
1322 if (__ceph_have_pending_cap_snap(ci)) {
1323 struct ceph_cap_snap *capsnap =
1324 list_last_entry(&ci->i_cap_snaps,
1325 struct ceph_cap_snap,
1326 ci_item);
1327 snapc = ceph_get_snap_context(capsnap->context);
1328 } else {
1329 BUG_ON(!ci->i_head_snapc);
1330 snapc = ceph_get_snap_context(ci->i_head_snapc);
1331 }
1332 spin_unlock(&ci->i_ceph_lock);
1333
1334 /* we might need to revert back to that point */
1335 data = *from;
1336 if (iocb->ki_flags & IOCB_DIRECT)
1337 written = ceph_direct_read_write(iocb, &data, snapc,
1338 &prealloc_cf);
1339 else
1340 written = ceph_sync_write(iocb, &data, pos, snapc);
1341 if (written == -EOLDSNAPC) {
1342 dout("aio_write %p %llx.%llx %llu~%u"
1343 "got EOLDSNAPC, retrying\n",
1344 inode, ceph_vinop(inode),
1345 pos, (unsigned)count);
1346 inode_lock(inode);
1347 goto retry_snap;
1348 }
1349 if (written > 0)
1350 iov_iter_advance(from, written);
1351 ceph_put_snap_context(snapc);
1352 } else {
1353 loff_t old_size = i_size_read(inode);
1354 /*
1355 * No need to acquire the i_truncate_mutex. Because
1356 * the MDS revokes Fwb caps before sending truncate
1357 * message to us. We can't get Fwb cap while there
1358 * are pending vmtruncate. So write and vmtruncate
1359 * can not run at the same time
1360 */
1361 written = generic_perform_write(file, from, pos);
1362 if (likely(written >= 0))
1363 iocb->ki_pos = pos + written;
1364 if (i_size_read(inode) > old_size)
1365 ceph_fscache_update_objectsize(inode);
1366 inode_unlock(inode);
1367 }
1368
1369 if (written >= 0) {
1370 int dirty;
1371 spin_lock(&ci->i_ceph_lock);
1372 ci->i_inline_version = CEPH_INLINE_NONE;
1373 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1374 &prealloc_cf);
1375 spin_unlock(&ci->i_ceph_lock);
1376 if (dirty)
1377 __mark_inode_dirty(inode, dirty);
1378 }
1379
1380 dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n",
1381 inode, ceph_vinop(inode), pos, (unsigned)count,
1382 ceph_cap_string(got));
1383 ceph_put_cap_refs(ci, got);
1384
1385 if (written >= 0 &&
1386 ((file->f_flags & O_SYNC) || IS_SYNC(file->f_mapping->host) ||
1387 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_NEARFULL))) {
1388 err = vfs_fsync_range(file, pos, pos + written - 1, 1);
1389 if (err < 0)
1390 written = err;
1391 }
1392
1393 goto out_unlocked;
1394
1395out:
1396 inode_unlock(inode);
1397out_unlocked:
1398 ceph_free_cap_flush(prealloc_cf);
1399 current->backing_dev_info = NULL;
1400 return written ? written : err;
1401}
1402
1403/*
1404 * llseek. be sure to verify file size on SEEK_END.
1405 */
1406static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1407{
1408 struct inode *inode = file->f_mapping->host;
1409 loff_t i_size;
1410 int ret;
1411
1412 inode_lock(inode);
1413
1414 if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1415 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1416 if (ret < 0) {
1417 offset = ret;
1418 goto out;
1419 }
1420 }
1421
1422 i_size = i_size_read(inode);
1423 switch (whence) {
1424 case SEEK_END:
1425 offset += i_size;
1426 break;
1427 case SEEK_CUR:
1428 /*
1429 * Here we special-case the lseek(fd, 0, SEEK_CUR)
1430 * position-querying operation. Avoid rewriting the "same"
1431 * f_pos value back to the file because a concurrent read(),
1432 * write() or lseek() might have altered it
1433 */
1434 if (offset == 0) {
1435 offset = file->f_pos;
1436 goto out;
1437 }
1438 offset += file->f_pos;
1439 break;
1440 case SEEK_DATA:
1441 if (offset >= i_size) {
1442 ret = -ENXIO;
1443 goto out;
1444 }
1445 break;
1446 case SEEK_HOLE:
1447 if (offset >= i_size) {
1448 ret = -ENXIO;
1449 goto out;
1450 }
1451 offset = i_size;
1452 break;
1453 }
1454
1455 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1456
1457out:
1458 inode_unlock(inode);
1459 return offset;
1460}
1461
1462static inline void ceph_zero_partial_page(
1463 struct inode *inode, loff_t offset, unsigned size)
1464{
1465 struct page *page;
1466 pgoff_t index = offset >> PAGE_SHIFT;
1467
1468 page = find_lock_page(inode->i_mapping, index);
1469 if (page) {
1470 wait_on_page_writeback(page);
1471 zero_user(page, offset & (PAGE_SIZE - 1), size);
1472 unlock_page(page);
1473 put_page(page);
1474 }
1475}
1476
1477static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1478 loff_t length)
1479{
1480 loff_t nearly = round_up(offset, PAGE_SIZE);
1481 if (offset < nearly) {
1482 loff_t size = nearly - offset;
1483 if (length < size)
1484 size = length;
1485 ceph_zero_partial_page(inode, offset, size);
1486 offset += size;
1487 length -= size;
1488 }
1489 if (length >= PAGE_SIZE) {
1490 loff_t size = round_down(length, PAGE_SIZE);
1491 truncate_pagecache_range(inode, offset, offset + size - 1);
1492 offset += size;
1493 length -= size;
1494 }
1495 if (length)
1496 ceph_zero_partial_page(inode, offset, length);
1497}
1498
1499static int ceph_zero_partial_object(struct inode *inode,
1500 loff_t offset, loff_t *length)
1501{
1502 struct ceph_inode_info *ci = ceph_inode(inode);
1503 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1504 struct ceph_osd_request *req;
1505 int ret = 0;
1506 loff_t zero = 0;
1507 int op;
1508
1509 if (!length) {
1510 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1511 length = &zero;
1512 } else {
1513 op = CEPH_OSD_OP_ZERO;
1514 }
1515
1516 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1517 ceph_vino(inode),
1518 offset, length,
1519 0, 1, op,
1520 CEPH_OSD_FLAG_WRITE |
1521 CEPH_OSD_FLAG_ONDISK,
1522 NULL, 0, 0, false);
1523 if (IS_ERR(req)) {
1524 ret = PTR_ERR(req);
1525 goto out;
1526 }
1527
1528 ceph_osdc_build_request(req, offset, NULL, ceph_vino(inode).snap,
1529 &inode->i_mtime);
1530
1531 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1532 if (!ret) {
1533 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1534 if (ret == -ENOENT)
1535 ret = 0;
1536 }
1537 ceph_osdc_put_request(req);
1538
1539out:
1540 return ret;
1541}
1542
1543static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
1544{
1545 int ret = 0;
1546 struct ceph_inode_info *ci = ceph_inode(inode);
1547 s32 stripe_unit = ceph_file_layout_su(ci->i_layout);
1548 s32 stripe_count = ceph_file_layout_stripe_count(ci->i_layout);
1549 s32 object_size = ceph_file_layout_object_size(ci->i_layout);
1550 u64 object_set_size = object_size * stripe_count;
1551 u64 nearly, t;
1552
1553 /* round offset up to next period boundary */
1554 nearly = offset + object_set_size - 1;
1555 t = nearly;
1556 nearly -= do_div(t, object_set_size);
1557
1558 while (length && offset < nearly) {
1559 loff_t size = length;
1560 ret = ceph_zero_partial_object(inode, offset, &size);
1561 if (ret < 0)
1562 return ret;
1563 offset += size;
1564 length -= size;
1565 }
1566 while (length >= object_set_size) {
1567 int i;
1568 loff_t pos = offset;
1569 for (i = 0; i < stripe_count; ++i) {
1570 ret = ceph_zero_partial_object(inode, pos, NULL);
1571 if (ret < 0)
1572 return ret;
1573 pos += stripe_unit;
1574 }
1575 offset += object_set_size;
1576 length -= object_set_size;
1577 }
1578 while (length) {
1579 loff_t size = length;
1580 ret = ceph_zero_partial_object(inode, offset, &size);
1581 if (ret < 0)
1582 return ret;
1583 offset += size;
1584 length -= size;
1585 }
1586 return ret;
1587}
1588
1589static long ceph_fallocate(struct file *file, int mode,
1590 loff_t offset, loff_t length)
1591{
1592 struct ceph_file_info *fi = file->private_data;
1593 struct inode *inode = file_inode(file);
1594 struct ceph_inode_info *ci = ceph_inode(inode);
1595 struct ceph_osd_client *osdc =
1596 &ceph_inode_to_client(inode)->client->osdc;
1597 struct ceph_cap_flush *prealloc_cf;
1598 int want, got = 0;
1599 int dirty;
1600 int ret = 0;
1601 loff_t endoff = 0;
1602 loff_t size;
1603
1604 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1605 return -EOPNOTSUPP;
1606
1607 if (!S_ISREG(inode->i_mode))
1608 return -EOPNOTSUPP;
1609
1610 prealloc_cf = ceph_alloc_cap_flush();
1611 if (!prealloc_cf)
1612 return -ENOMEM;
1613
1614 inode_lock(inode);
1615
1616 if (ceph_snap(inode) != CEPH_NOSNAP) {
1617 ret = -EROFS;
1618 goto unlock;
1619 }
1620
1621 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) &&
1622 !(mode & FALLOC_FL_PUNCH_HOLE)) {
1623 ret = -ENOSPC;
1624 goto unlock;
1625 }
1626
1627 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1628 ret = ceph_uninline_data(file, NULL);
1629 if (ret < 0)
1630 goto unlock;
1631 }
1632
1633 size = i_size_read(inode);
1634 if (!(mode & FALLOC_FL_KEEP_SIZE))
1635 endoff = offset + length;
1636
1637 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1638 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1639 else
1640 want = CEPH_CAP_FILE_BUFFER;
1641
1642 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, endoff, &got, NULL);
1643 if (ret < 0)
1644 goto unlock;
1645
1646 if (mode & FALLOC_FL_PUNCH_HOLE) {
1647 if (offset < size)
1648 ceph_zero_pagecache_range(inode, offset, length);
1649 ret = ceph_zero_objects(inode, offset, length);
1650 } else if (endoff > size) {
1651 truncate_pagecache_range(inode, size, -1);
1652 if (ceph_inode_set_size(inode, endoff))
1653 ceph_check_caps(ceph_inode(inode),
1654 CHECK_CAPS_AUTHONLY, NULL);
1655 }
1656
1657 if (!ret) {
1658 spin_lock(&ci->i_ceph_lock);
1659 ci->i_inline_version = CEPH_INLINE_NONE;
1660 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1661 &prealloc_cf);
1662 spin_unlock(&ci->i_ceph_lock);
1663 if (dirty)
1664 __mark_inode_dirty(inode, dirty);
1665 }
1666
1667 ceph_put_cap_refs(ci, got);
1668unlock:
1669 inode_unlock(inode);
1670 ceph_free_cap_flush(prealloc_cf);
1671 return ret;
1672}
1673
1674const struct file_operations ceph_file_fops = {
1675 .open = ceph_open,
1676 .release = ceph_release,
1677 .llseek = ceph_llseek,
1678 .read_iter = ceph_read_iter,
1679 .write_iter = ceph_write_iter,
1680 .mmap = ceph_mmap,
1681 .fsync = ceph_fsync,
1682 .lock = ceph_lock,
1683 .flock = ceph_flock,
1684 .splice_read = generic_file_splice_read,
1685 .splice_write = iter_file_splice_write,
1686 .unlocked_ioctl = ceph_ioctl,
1687 .compat_ioctl = ceph_ioctl,
1688 .fallocate = ceph_fallocate,
1689};
1690
1// SPDX-License-Identifier: GPL-2.0
2#include <linux/ceph/ceph_debug.h>
3#include <linux/ceph/striper.h>
4
5#include <linux/module.h>
6#include <linux/sched.h>
7#include <linux/slab.h>
8#include <linux/file.h>
9#include <linux/mount.h>
10#include <linux/namei.h>
11#include <linux/writeback.h>
12#include <linux/falloc.h>
13#include <linux/iversion.h>
14#include <linux/ktime.h>
15
16#include "super.h"
17#include "mds_client.h"
18#include "cache.h"
19#include "io.h"
20#include "metric.h"
21
22static __le32 ceph_flags_sys2wire(u32 flags)
23{
24 u32 wire_flags = 0;
25
26 switch (flags & O_ACCMODE) {
27 case O_RDONLY:
28 wire_flags |= CEPH_O_RDONLY;
29 break;
30 case O_WRONLY:
31 wire_flags |= CEPH_O_WRONLY;
32 break;
33 case O_RDWR:
34 wire_flags |= CEPH_O_RDWR;
35 break;
36 }
37
38 flags &= ~O_ACCMODE;
39
40#define ceph_sys2wire(a) if (flags & a) { wire_flags |= CEPH_##a; flags &= ~a; }
41
42 ceph_sys2wire(O_CREAT);
43 ceph_sys2wire(O_EXCL);
44 ceph_sys2wire(O_TRUNC);
45 ceph_sys2wire(O_DIRECTORY);
46 ceph_sys2wire(O_NOFOLLOW);
47
48#undef ceph_sys2wire
49
50 if (flags)
51 dout("unused open flags: %x\n", flags);
52
53 return cpu_to_le32(wire_flags);
54}
55
56/*
57 * Ceph file operations
58 *
59 * Implement basic open/close functionality, and implement
60 * read/write.
61 *
62 * We implement three modes of file I/O:
63 * - buffered uses the generic_file_aio_{read,write} helpers
64 *
65 * - synchronous is used when there is multi-client read/write
66 * sharing, avoids the page cache, and synchronously waits for an
67 * ack from the OSD.
68 *
69 * - direct io takes the variant of the sync path that references
70 * user pages directly.
71 *
72 * fsync() flushes and waits on dirty pages, but just queues metadata
73 * for writeback: since the MDS can recover size and mtime there is no
74 * need to wait for MDS acknowledgement.
75 */
76
77/*
78 * How many pages to get in one call to iov_iter_get_pages(). This
79 * determines the size of the on-stack array used as a buffer.
80 */
81#define ITER_GET_BVECS_PAGES 64
82
83static ssize_t __iter_get_bvecs(struct iov_iter *iter, size_t maxsize,
84 struct bio_vec *bvecs)
85{
86 size_t size = 0;
87 int bvec_idx = 0;
88
89 if (maxsize > iov_iter_count(iter))
90 maxsize = iov_iter_count(iter);
91
92 while (size < maxsize) {
93 struct page *pages[ITER_GET_BVECS_PAGES];
94 ssize_t bytes;
95 size_t start;
96 int idx = 0;
97
98 bytes = iov_iter_get_pages(iter, pages, maxsize - size,
99 ITER_GET_BVECS_PAGES, &start);
100 if (bytes < 0)
101 return size ?: bytes;
102
103 iov_iter_advance(iter, bytes);
104 size += bytes;
105
106 for ( ; bytes; idx++, bvec_idx++) {
107 struct bio_vec bv = {
108 .bv_page = pages[idx],
109 .bv_len = min_t(int, bytes, PAGE_SIZE - start),
110 .bv_offset = start,
111 };
112
113 bvecs[bvec_idx] = bv;
114 bytes -= bv.bv_len;
115 start = 0;
116 }
117 }
118
119 return size;
120}
121
122/*
123 * iov_iter_get_pages() only considers one iov_iter segment, no matter
124 * what maxsize or maxpages are given. For ITER_BVEC that is a single
125 * page.
126 *
127 * Attempt to get up to @maxsize bytes worth of pages from @iter.
128 * Return the number of bytes in the created bio_vec array, or an error.
129 */
130static ssize_t iter_get_bvecs_alloc(struct iov_iter *iter, size_t maxsize,
131 struct bio_vec **bvecs, int *num_bvecs)
132{
133 struct bio_vec *bv;
134 size_t orig_count = iov_iter_count(iter);
135 ssize_t bytes;
136 int npages;
137
138 iov_iter_truncate(iter, maxsize);
139 npages = iov_iter_npages(iter, INT_MAX);
140 iov_iter_reexpand(iter, orig_count);
141
142 /*
143 * __iter_get_bvecs() may populate only part of the array -- zero it
144 * out.
145 */
146 bv = kvmalloc_array(npages, sizeof(*bv), GFP_KERNEL | __GFP_ZERO);
147 if (!bv)
148 return -ENOMEM;
149
150 bytes = __iter_get_bvecs(iter, maxsize, bv);
151 if (bytes < 0) {
152 /*
153 * No pages were pinned -- just free the array.
154 */
155 kvfree(bv);
156 return bytes;
157 }
158
159 *bvecs = bv;
160 *num_bvecs = npages;
161 return bytes;
162}
163
164static void put_bvecs(struct bio_vec *bvecs, int num_bvecs, bool should_dirty)
165{
166 int i;
167
168 for (i = 0; i < num_bvecs; i++) {
169 if (bvecs[i].bv_page) {
170 if (should_dirty)
171 set_page_dirty_lock(bvecs[i].bv_page);
172 put_page(bvecs[i].bv_page);
173 }
174 }
175 kvfree(bvecs);
176}
177
178/*
179 * Prepare an open request. Preallocate ceph_cap to avoid an
180 * inopportune ENOMEM later.
181 */
182static struct ceph_mds_request *
183prepare_open_request(struct super_block *sb, int flags, int create_mode)
184{
185 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(sb);
186 struct ceph_mds_request *req;
187 int want_auth = USE_ANY_MDS;
188 int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
189
190 if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
191 want_auth = USE_AUTH_MDS;
192
193 req = ceph_mdsc_create_request(mdsc, op, want_auth);
194 if (IS_ERR(req))
195 goto out;
196 req->r_fmode = ceph_flags_to_mode(flags);
197 req->r_args.open.flags = ceph_flags_sys2wire(flags);
198 req->r_args.open.mode = cpu_to_le32(create_mode);
199out:
200 return req;
201}
202
203static int ceph_init_file_info(struct inode *inode, struct file *file,
204 int fmode, bool isdir)
205{
206 struct ceph_inode_info *ci = ceph_inode(inode);
207 struct ceph_file_info *fi;
208
209 dout("%s %p %p 0%o (%s)\n", __func__, inode, file,
210 inode->i_mode, isdir ? "dir" : "regular");
211 BUG_ON(inode->i_fop->release != ceph_release);
212
213 if (isdir) {
214 struct ceph_dir_file_info *dfi =
215 kmem_cache_zalloc(ceph_dir_file_cachep, GFP_KERNEL);
216 if (!dfi)
217 return -ENOMEM;
218
219 file->private_data = dfi;
220 fi = &dfi->file_info;
221 dfi->next_offset = 2;
222 dfi->readdir_cache_idx = -1;
223 } else {
224 fi = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
225 if (!fi)
226 return -ENOMEM;
227
228 file->private_data = fi;
229 }
230
231 ceph_get_fmode(ci, fmode, 1);
232 fi->fmode = fmode;
233
234 spin_lock_init(&fi->rw_contexts_lock);
235 INIT_LIST_HEAD(&fi->rw_contexts);
236 fi->filp_gen = READ_ONCE(ceph_inode_to_client(inode)->filp_gen);
237
238 return 0;
239}
240
241/*
242 * initialize private struct file data.
243 * if we fail, clean up by dropping fmode reference on the ceph_inode
244 */
245static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
246{
247 int ret = 0;
248
249 switch (inode->i_mode & S_IFMT) {
250 case S_IFREG:
251 ceph_fscache_register_inode_cookie(inode);
252 ceph_fscache_file_set_cookie(inode, file);
253 fallthrough;
254 case S_IFDIR:
255 ret = ceph_init_file_info(inode, file, fmode,
256 S_ISDIR(inode->i_mode));
257 break;
258
259 case S_IFLNK:
260 dout("init_file %p %p 0%o (symlink)\n", inode, file,
261 inode->i_mode);
262 break;
263
264 default:
265 dout("init_file %p %p 0%o (special)\n", inode, file,
266 inode->i_mode);
267 /*
268 * we need to drop the open ref now, since we don't
269 * have .release set to ceph_release.
270 */
271 BUG_ON(inode->i_fop->release == ceph_release);
272
273 /* call the proper open fop */
274 ret = inode->i_fop->open(inode, file);
275 }
276 return ret;
277}
278
279/*
280 * try renew caps after session gets killed.
281 */
282int ceph_renew_caps(struct inode *inode, int fmode)
283{
284 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
285 struct ceph_inode_info *ci = ceph_inode(inode);
286 struct ceph_mds_request *req;
287 int err, flags, wanted;
288
289 spin_lock(&ci->i_ceph_lock);
290 __ceph_touch_fmode(ci, mdsc, fmode);
291 wanted = __ceph_caps_file_wanted(ci);
292 if (__ceph_is_any_real_caps(ci) &&
293 (!(wanted & CEPH_CAP_ANY_WR) || ci->i_auth_cap)) {
294 int issued = __ceph_caps_issued(ci, NULL);
295 spin_unlock(&ci->i_ceph_lock);
296 dout("renew caps %p want %s issued %s updating mds_wanted\n",
297 inode, ceph_cap_string(wanted), ceph_cap_string(issued));
298 ceph_check_caps(ci, 0, NULL);
299 return 0;
300 }
301 spin_unlock(&ci->i_ceph_lock);
302
303 flags = 0;
304 if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
305 flags = O_RDWR;
306 else if (wanted & CEPH_CAP_FILE_RD)
307 flags = O_RDONLY;
308 else if (wanted & CEPH_CAP_FILE_WR)
309 flags = O_WRONLY;
310#ifdef O_LAZY
311 if (wanted & CEPH_CAP_FILE_LAZYIO)
312 flags |= O_LAZY;
313#endif
314
315 req = prepare_open_request(inode->i_sb, flags, 0);
316 if (IS_ERR(req)) {
317 err = PTR_ERR(req);
318 goto out;
319 }
320
321 req->r_inode = inode;
322 ihold(inode);
323 req->r_num_caps = 1;
324
325 err = ceph_mdsc_do_request(mdsc, NULL, req);
326 ceph_mdsc_put_request(req);
327out:
328 dout("renew caps %p open result=%d\n", inode, err);
329 return err < 0 ? err : 0;
330}
331
332/*
333 * If we already have the requisite capabilities, we can satisfy
334 * the open request locally (no need to request new caps from the
335 * MDS). We do, however, need to inform the MDS (asynchronously)
336 * if our wanted caps set expands.
337 */
338int ceph_open(struct inode *inode, struct file *file)
339{
340 struct ceph_inode_info *ci = ceph_inode(inode);
341 struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
342 struct ceph_mds_client *mdsc = fsc->mdsc;
343 struct ceph_mds_request *req;
344 struct ceph_file_info *fi = file->private_data;
345 int err;
346 int flags, fmode, wanted;
347
348 if (fi) {
349 dout("open file %p is already opened\n", file);
350 return 0;
351 }
352
353 /* filter out O_CREAT|O_EXCL; vfs did that already. yuck. */
354 flags = file->f_flags & ~(O_CREAT|O_EXCL);
355 if (S_ISDIR(inode->i_mode))
356 flags = O_DIRECTORY; /* mds likes to know */
357
358 dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
359 ceph_vinop(inode), file, flags, file->f_flags);
360 fmode = ceph_flags_to_mode(flags);
361 wanted = ceph_caps_for_mode(fmode);
362
363 /* snapped files are read-only */
364 if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
365 return -EROFS;
366
367 /* trivially open snapdir */
368 if (ceph_snap(inode) == CEPH_SNAPDIR) {
369 return ceph_init_file(inode, file, fmode);
370 }
371
372 /*
373 * No need to block if we have caps on the auth MDS (for
374 * write) or any MDS (for read). Update wanted set
375 * asynchronously.
376 */
377 spin_lock(&ci->i_ceph_lock);
378 if (__ceph_is_any_real_caps(ci) &&
379 (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
380 int mds_wanted = __ceph_caps_mds_wanted(ci, true);
381 int issued = __ceph_caps_issued(ci, NULL);
382
383 dout("open %p fmode %d want %s issued %s using existing\n",
384 inode, fmode, ceph_cap_string(wanted),
385 ceph_cap_string(issued));
386 __ceph_touch_fmode(ci, mdsc, fmode);
387 spin_unlock(&ci->i_ceph_lock);
388
389 /* adjust wanted? */
390 if ((issued & wanted) != wanted &&
391 (mds_wanted & wanted) != wanted &&
392 ceph_snap(inode) != CEPH_SNAPDIR)
393 ceph_check_caps(ci, 0, NULL);
394
395 return ceph_init_file(inode, file, fmode);
396 } else if (ceph_snap(inode) != CEPH_NOSNAP &&
397 (ci->i_snap_caps & wanted) == wanted) {
398 __ceph_touch_fmode(ci, mdsc, fmode);
399 spin_unlock(&ci->i_ceph_lock);
400 return ceph_init_file(inode, file, fmode);
401 }
402
403 spin_unlock(&ci->i_ceph_lock);
404
405 dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
406 req = prepare_open_request(inode->i_sb, flags, 0);
407 if (IS_ERR(req)) {
408 err = PTR_ERR(req);
409 goto out;
410 }
411 req->r_inode = inode;
412 ihold(inode);
413
414 req->r_num_caps = 1;
415 err = ceph_mdsc_do_request(mdsc, NULL, req);
416 if (!err)
417 err = ceph_init_file(inode, file, req->r_fmode);
418 ceph_mdsc_put_request(req);
419 dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
420out:
421 return err;
422}
423
424/* Clone the layout from a synchronous create, if the dir now has Dc caps */
425static void
426cache_file_layout(struct inode *dst, struct inode *src)
427{
428 struct ceph_inode_info *cdst = ceph_inode(dst);
429 struct ceph_inode_info *csrc = ceph_inode(src);
430
431 spin_lock(&cdst->i_ceph_lock);
432 if ((__ceph_caps_issued(cdst, NULL) & CEPH_CAP_DIR_CREATE) &&
433 !ceph_file_layout_is_valid(&cdst->i_cached_layout)) {
434 memcpy(&cdst->i_cached_layout, &csrc->i_layout,
435 sizeof(cdst->i_cached_layout));
436 rcu_assign_pointer(cdst->i_cached_layout.pool_ns,
437 ceph_try_get_string(csrc->i_layout.pool_ns));
438 }
439 spin_unlock(&cdst->i_ceph_lock);
440}
441
442/*
443 * Try to set up an async create. We need caps, a file layout, and inode number,
444 * and either a lease on the dentry or complete dir info. If any of those
445 * criteria are not satisfied, then return false and the caller can go
446 * synchronous.
447 */
448static int try_prep_async_create(struct inode *dir, struct dentry *dentry,
449 struct ceph_file_layout *lo, u64 *pino)
450{
451 struct ceph_inode_info *ci = ceph_inode(dir);
452 struct ceph_dentry_info *di = ceph_dentry(dentry);
453 int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_CREATE;
454 u64 ino;
455
456 spin_lock(&ci->i_ceph_lock);
457 /* No auth cap means no chance for Dc caps */
458 if (!ci->i_auth_cap)
459 goto no_async;
460
461 /* Any delegated inos? */
462 if (xa_empty(&ci->i_auth_cap->session->s_delegated_inos))
463 goto no_async;
464
465 if (!ceph_file_layout_is_valid(&ci->i_cached_layout))
466 goto no_async;
467
468 if ((__ceph_caps_issued(ci, NULL) & want) != want)
469 goto no_async;
470
471 if (d_in_lookup(dentry)) {
472 if (!__ceph_dir_is_complete(ci))
473 goto no_async;
474 spin_lock(&dentry->d_lock);
475 di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
476 spin_unlock(&dentry->d_lock);
477 } else if (atomic_read(&ci->i_shared_gen) !=
478 READ_ONCE(di->lease_shared_gen)) {
479 goto no_async;
480 }
481
482 ino = ceph_get_deleg_ino(ci->i_auth_cap->session);
483 if (!ino)
484 goto no_async;
485
486 *pino = ino;
487 ceph_take_cap_refs(ci, want, false);
488 memcpy(lo, &ci->i_cached_layout, sizeof(*lo));
489 rcu_assign_pointer(lo->pool_ns,
490 ceph_try_get_string(ci->i_cached_layout.pool_ns));
491 got = want;
492no_async:
493 spin_unlock(&ci->i_ceph_lock);
494 return got;
495}
496
497static void restore_deleg_ino(struct inode *dir, u64 ino)
498{
499 struct ceph_inode_info *ci = ceph_inode(dir);
500 struct ceph_mds_session *s = NULL;
501
502 spin_lock(&ci->i_ceph_lock);
503 if (ci->i_auth_cap)
504 s = ceph_get_mds_session(ci->i_auth_cap->session);
505 spin_unlock(&ci->i_ceph_lock);
506 if (s) {
507 int err = ceph_restore_deleg_ino(s, ino);
508 if (err)
509 pr_warn("ceph: unable to restore delegated ino 0x%llx to session: %d\n",
510 ino, err);
511 ceph_put_mds_session(s);
512 }
513}
514
515static void ceph_async_create_cb(struct ceph_mds_client *mdsc,
516 struct ceph_mds_request *req)
517{
518 int result = req->r_err ? req->r_err :
519 le32_to_cpu(req->r_reply_info.head->result);
520
521 if (result == -EJUKEBOX)
522 goto out;
523
524 mapping_set_error(req->r_parent->i_mapping, result);
525
526 if (result) {
527 struct dentry *dentry = req->r_dentry;
528 int pathlen = 0;
529 u64 base = 0;
530 char *path = ceph_mdsc_build_path(req->r_dentry, &pathlen,
531 &base, 0);
532
533 ceph_dir_clear_complete(req->r_parent);
534 if (!d_unhashed(dentry))
535 d_drop(dentry);
536
537 /* FIXME: start returning I/O errors on all accesses? */
538 pr_warn("ceph: async create failure path=(%llx)%s result=%d!\n",
539 base, IS_ERR(path) ? "<<bad>>" : path, result);
540 ceph_mdsc_free_path(path, pathlen);
541 }
542
543 if (req->r_target_inode) {
544 struct ceph_inode_info *ci = ceph_inode(req->r_target_inode);
545 u64 ino = ceph_vino(req->r_target_inode).ino;
546
547 if (req->r_deleg_ino != ino)
548 pr_warn("%s: inode number mismatch! err=%d deleg_ino=0x%llx target=0x%llx\n",
549 __func__, req->r_err, req->r_deleg_ino, ino);
550 mapping_set_error(req->r_target_inode->i_mapping, result);
551
552 spin_lock(&ci->i_ceph_lock);
553 if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE) {
554 ci->i_ceph_flags &= ~CEPH_I_ASYNC_CREATE;
555 wake_up_bit(&ci->i_ceph_flags, CEPH_ASYNC_CREATE_BIT);
556 }
557 ceph_kick_flushing_inode_caps(req->r_session, ci);
558 spin_unlock(&ci->i_ceph_lock);
559 } else {
560 pr_warn("%s: no req->r_target_inode for 0x%llx\n", __func__,
561 req->r_deleg_ino);
562 }
563out:
564 ceph_mdsc_release_dir_caps(req);
565}
566
567static int ceph_finish_async_create(struct inode *dir, struct dentry *dentry,
568 struct file *file, umode_t mode,
569 struct ceph_mds_request *req,
570 struct ceph_acl_sec_ctx *as_ctx,
571 struct ceph_file_layout *lo)
572{
573 int ret;
574 char xattr_buf[4];
575 struct ceph_mds_reply_inode in = { };
576 struct ceph_mds_reply_info_in iinfo = { .in = &in };
577 struct ceph_inode_info *ci = ceph_inode(dir);
578 struct inode *inode;
579 struct timespec64 now;
580 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
581 struct ceph_vino vino = { .ino = req->r_deleg_ino,
582 .snap = CEPH_NOSNAP };
583
584 ktime_get_real_ts64(&now);
585
586 inode = ceph_get_inode(dentry->d_sb, vino);
587 if (IS_ERR(inode))
588 return PTR_ERR(inode);
589
590 iinfo.inline_version = CEPH_INLINE_NONE;
591 iinfo.change_attr = 1;
592 ceph_encode_timespec64(&iinfo.btime, &now);
593
594 iinfo.xattr_len = ARRAY_SIZE(xattr_buf);
595 iinfo.xattr_data = xattr_buf;
596 memset(iinfo.xattr_data, 0, iinfo.xattr_len);
597
598 in.ino = cpu_to_le64(vino.ino);
599 in.snapid = cpu_to_le64(CEPH_NOSNAP);
600 in.version = cpu_to_le64(1); // ???
601 in.cap.caps = in.cap.wanted = cpu_to_le32(CEPH_CAP_ALL_FILE);
602 in.cap.cap_id = cpu_to_le64(1);
603 in.cap.realm = cpu_to_le64(ci->i_snap_realm->ino);
604 in.cap.flags = CEPH_CAP_FLAG_AUTH;
605 in.ctime = in.mtime = in.atime = iinfo.btime;
606 in.mode = cpu_to_le32((u32)mode);
607 in.truncate_seq = cpu_to_le32(1);
608 in.truncate_size = cpu_to_le64(-1ULL);
609 in.xattr_version = cpu_to_le64(1);
610 in.uid = cpu_to_le32(from_kuid(&init_user_ns, current_fsuid()));
611 in.gid = cpu_to_le32(from_kgid(&init_user_ns, dir->i_mode & S_ISGID ?
612 dir->i_gid : current_fsgid()));
613 in.nlink = cpu_to_le32(1);
614 in.max_size = cpu_to_le64(lo->stripe_unit);
615
616 ceph_file_layout_to_legacy(lo, &in.layout);
617
618 down_read(&mdsc->snap_rwsem);
619 ret = ceph_fill_inode(inode, NULL, &iinfo, NULL, req->r_session,
620 req->r_fmode, NULL);
621 up_read(&mdsc->snap_rwsem);
622 if (ret) {
623 dout("%s failed to fill inode: %d\n", __func__, ret);
624 ceph_dir_clear_complete(dir);
625 if (!d_unhashed(dentry))
626 d_drop(dentry);
627 if (inode->i_state & I_NEW)
628 discard_new_inode(inode);
629 } else {
630 struct dentry *dn;
631
632 dout("%s d_adding new inode 0x%llx to 0x%llx/%s\n", __func__,
633 vino.ino, ceph_ino(dir), dentry->d_name.name);
634 ceph_dir_clear_ordered(dir);
635 ceph_init_inode_acls(inode, as_ctx);
636 if (inode->i_state & I_NEW) {
637 /*
638 * If it's not I_NEW, then someone created this before
639 * we got here. Assume the server is aware of it at
640 * that point and don't worry about setting
641 * CEPH_I_ASYNC_CREATE.
642 */
643 ceph_inode(inode)->i_ceph_flags = CEPH_I_ASYNC_CREATE;
644 unlock_new_inode(inode);
645 }
646 if (d_in_lookup(dentry) || d_really_is_negative(dentry)) {
647 if (!d_unhashed(dentry))
648 d_drop(dentry);
649 dn = d_splice_alias(inode, dentry);
650 WARN_ON_ONCE(dn && dn != dentry);
651 }
652 file->f_mode |= FMODE_CREATED;
653 ret = finish_open(file, dentry, ceph_open);
654 }
655 return ret;
656}
657
658/*
659 * Do a lookup + open with a single request. If we get a non-existent
660 * file or symlink, return 1 so the VFS can retry.
661 */
662int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
663 struct file *file, unsigned flags, umode_t mode)
664{
665 struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
666 struct ceph_mds_client *mdsc = fsc->mdsc;
667 struct ceph_mds_request *req;
668 struct dentry *dn;
669 struct ceph_acl_sec_ctx as_ctx = {};
670 bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
671 int mask;
672 int err;
673
674 dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
675 dir, dentry, dentry,
676 d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
677
678 if (dentry->d_name.len > NAME_MAX)
679 return -ENAMETOOLONG;
680
681 if (flags & O_CREAT) {
682 if (ceph_quota_is_max_files_exceeded(dir))
683 return -EDQUOT;
684 err = ceph_pre_init_acls(dir, &mode, &as_ctx);
685 if (err < 0)
686 return err;
687 err = ceph_security_init_secctx(dentry, mode, &as_ctx);
688 if (err < 0)
689 goto out_ctx;
690 } else if (!d_in_lookup(dentry)) {
691 /* If it's not being looked up, it's negative */
692 return -ENOENT;
693 }
694retry:
695 /* do the open */
696 req = prepare_open_request(dir->i_sb, flags, mode);
697 if (IS_ERR(req)) {
698 err = PTR_ERR(req);
699 goto out_ctx;
700 }
701 req->r_dentry = dget(dentry);
702 req->r_num_caps = 2;
703 mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
704 if (ceph_security_xattr_wanted(dir))
705 mask |= CEPH_CAP_XATTR_SHARED;
706 req->r_args.open.mask = cpu_to_le32(mask);
707 req->r_parent = dir;
708 ihold(dir);
709
710 if (flags & O_CREAT) {
711 struct ceph_file_layout lo;
712
713 req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
714 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
715 if (as_ctx.pagelist) {
716 req->r_pagelist = as_ctx.pagelist;
717 as_ctx.pagelist = NULL;
718 }
719 if (try_async &&
720 (req->r_dir_caps =
721 try_prep_async_create(dir, dentry, &lo,
722 &req->r_deleg_ino))) {
723 set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
724 req->r_args.open.flags |= cpu_to_le32(CEPH_O_EXCL);
725 req->r_callback = ceph_async_create_cb;
726 err = ceph_mdsc_submit_request(mdsc, dir, req);
727 if (!err) {
728 err = ceph_finish_async_create(dir, dentry,
729 file, mode, req,
730 &as_ctx, &lo);
731 } else if (err == -EJUKEBOX) {
732 restore_deleg_ino(dir, req->r_deleg_ino);
733 ceph_mdsc_put_request(req);
734 try_async = false;
735 goto retry;
736 }
737 goto out_req;
738 }
739 }
740
741 set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
742 err = ceph_mdsc_do_request(mdsc,
743 (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
744 req);
745 if (err == -ENOENT) {
746 dentry = ceph_handle_snapdir(req, dentry);
747 if (IS_ERR(dentry)) {
748 err = PTR_ERR(dentry);
749 goto out_req;
750 }
751 err = 0;
752 }
753
754 if (!err && (flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
755 err = ceph_handle_notrace_create(dir, dentry);
756
757 if (d_in_lookup(dentry)) {
758 dn = ceph_finish_lookup(req, dentry, err);
759 if (IS_ERR(dn))
760 err = PTR_ERR(dn);
761 } else {
762 /* we were given a hashed negative dentry */
763 dn = NULL;
764 }
765 if (err)
766 goto out_req;
767 if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
768 /* make vfs retry on splice, ENOENT, or symlink */
769 dout("atomic_open finish_no_open on dn %p\n", dn);
770 err = finish_no_open(file, dn);
771 } else {
772 dout("atomic_open finish_open on dn %p\n", dn);
773 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
774 struct inode *newino = d_inode(dentry);
775
776 cache_file_layout(dir, newino);
777 ceph_init_inode_acls(newino, &as_ctx);
778 file->f_mode |= FMODE_CREATED;
779 }
780 err = finish_open(file, dentry, ceph_open);
781 }
782out_req:
783 ceph_mdsc_put_request(req);
784out_ctx:
785 ceph_release_acl_sec_ctx(&as_ctx);
786 dout("atomic_open result=%d\n", err);
787 return err;
788}
789
790int ceph_release(struct inode *inode, struct file *file)
791{
792 struct ceph_inode_info *ci = ceph_inode(inode);
793
794 if (S_ISDIR(inode->i_mode)) {
795 struct ceph_dir_file_info *dfi = file->private_data;
796 dout("release inode %p dir file %p\n", inode, file);
797 WARN_ON(!list_empty(&dfi->file_info.rw_contexts));
798
799 ceph_put_fmode(ci, dfi->file_info.fmode, 1);
800
801 if (dfi->last_readdir)
802 ceph_mdsc_put_request(dfi->last_readdir);
803 kfree(dfi->last_name);
804 kfree(dfi->dir_info);
805 kmem_cache_free(ceph_dir_file_cachep, dfi);
806 } else {
807 struct ceph_file_info *fi = file->private_data;
808 dout("release inode %p regular file %p\n", inode, file);
809 WARN_ON(!list_empty(&fi->rw_contexts));
810
811 ceph_put_fmode(ci, fi->fmode, 1);
812
813 kmem_cache_free(ceph_file_cachep, fi);
814 }
815
816 /* wake up anyone waiting for caps on this inode */
817 wake_up_all(&ci->i_cap_wq);
818 return 0;
819}
820
821enum {
822 HAVE_RETRIED = 1,
823 CHECK_EOF = 2,
824 READ_INLINE = 3,
825};
826
827/*
828 * Completely synchronous read and write methods. Direct from __user
829 * buffer to osd, or directly to user pages (if O_DIRECT).
830 *
831 * If the read spans object boundary, just do multiple reads. (That's not
832 * atomic, but good enough for now.)
833 *
834 * If we get a short result from the OSD, check against i_size; we need to
835 * only return a short read to the caller if we hit EOF.
836 */
837static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
838 int *retry_op)
839{
840 struct file *file = iocb->ki_filp;
841 struct inode *inode = file_inode(file);
842 struct ceph_inode_info *ci = ceph_inode(inode);
843 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
844 struct ceph_osd_client *osdc = &fsc->client->osdc;
845 ssize_t ret;
846 u64 off = iocb->ki_pos;
847 u64 len = iov_iter_count(to);
848
849 dout("sync_read on file %p %llu~%u %s\n", file, off, (unsigned)len,
850 (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
851
852 if (!len)
853 return 0;
854 /*
855 * flush any page cache pages in this range. this
856 * will make concurrent normal and sync io slow,
857 * but it will at least behave sensibly when they are
858 * in sequence.
859 */
860 ret = filemap_write_and_wait_range(inode->i_mapping,
861 off, off + len - 1);
862 if (ret < 0)
863 return ret;
864
865 ret = 0;
866 while ((len = iov_iter_count(to)) > 0) {
867 struct ceph_osd_request *req;
868 struct page **pages;
869 int num_pages;
870 size_t page_off;
871 u64 i_size;
872 bool more;
873 int idx;
874 size_t left;
875
876 req = ceph_osdc_new_request(osdc, &ci->i_layout,
877 ci->i_vino, off, &len, 0, 1,
878 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
879 NULL, ci->i_truncate_seq,
880 ci->i_truncate_size, false);
881 if (IS_ERR(req)) {
882 ret = PTR_ERR(req);
883 break;
884 }
885
886 more = len < iov_iter_count(to);
887
888 num_pages = calc_pages_for(off, len);
889 page_off = off & ~PAGE_MASK;
890 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
891 if (IS_ERR(pages)) {
892 ceph_osdc_put_request(req);
893 ret = PTR_ERR(pages);
894 break;
895 }
896
897 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_off,
898 false, false);
899 ret = ceph_osdc_start_request(osdc, req, false);
900 if (!ret)
901 ret = ceph_osdc_wait_request(osdc, req);
902
903 ceph_update_read_metrics(&fsc->mdsc->metric,
904 req->r_start_latency,
905 req->r_end_latency,
906 len, ret);
907
908 ceph_osdc_put_request(req);
909
910 i_size = i_size_read(inode);
911 dout("sync_read %llu~%llu got %zd i_size %llu%s\n",
912 off, len, ret, i_size, (more ? " MORE" : ""));
913
914 if (ret == -ENOENT)
915 ret = 0;
916 if (ret >= 0 && ret < len && (off + ret < i_size)) {
917 int zlen = min(len - ret, i_size - off - ret);
918 int zoff = page_off + ret;
919 dout("sync_read zero gap %llu~%llu\n",
920 off + ret, off + ret + zlen);
921 ceph_zero_page_vector_range(zoff, zlen, pages);
922 ret += zlen;
923 }
924
925 idx = 0;
926 left = ret > 0 ? ret : 0;
927 while (left > 0) {
928 size_t len, copied;
929 page_off = off & ~PAGE_MASK;
930 len = min_t(size_t, left, PAGE_SIZE - page_off);
931 SetPageUptodate(pages[idx]);
932 copied = copy_page_to_iter(pages[idx++],
933 page_off, len, to);
934 off += copied;
935 left -= copied;
936 if (copied < len) {
937 ret = -EFAULT;
938 break;
939 }
940 }
941 ceph_release_page_vector(pages, num_pages);
942
943 if (ret < 0) {
944 if (ret == -EBLOCKLISTED)
945 fsc->blocklisted = true;
946 break;
947 }
948
949 if (off >= i_size || !more)
950 break;
951 }
952
953 if (off > iocb->ki_pos) {
954 if (ret >= 0 &&
955 iov_iter_count(to) > 0 && off >= i_size_read(inode))
956 *retry_op = CHECK_EOF;
957 ret = off - iocb->ki_pos;
958 iocb->ki_pos = off;
959 }
960
961 dout("sync_read result %zd retry_op %d\n", ret, *retry_op);
962 return ret;
963}
964
965struct ceph_aio_request {
966 struct kiocb *iocb;
967 size_t total_len;
968 bool write;
969 bool should_dirty;
970 int error;
971 struct list_head osd_reqs;
972 unsigned num_reqs;
973 atomic_t pending_reqs;
974 struct timespec64 mtime;
975 struct ceph_cap_flush *prealloc_cf;
976};
977
978struct ceph_aio_work {
979 struct work_struct work;
980 struct ceph_osd_request *req;
981};
982
983static void ceph_aio_retry_work(struct work_struct *work);
984
985static void ceph_aio_complete(struct inode *inode,
986 struct ceph_aio_request *aio_req)
987{
988 struct ceph_inode_info *ci = ceph_inode(inode);
989 int ret;
990
991 if (!atomic_dec_and_test(&aio_req->pending_reqs))
992 return;
993
994 if (aio_req->iocb->ki_flags & IOCB_DIRECT)
995 inode_dio_end(inode);
996
997 ret = aio_req->error;
998 if (!ret)
999 ret = aio_req->total_len;
1000
1001 dout("ceph_aio_complete %p rc %d\n", inode, ret);
1002
1003 if (ret >= 0 && aio_req->write) {
1004 int dirty;
1005
1006 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
1007 if (endoff > i_size_read(inode)) {
1008 if (ceph_inode_set_size(inode, endoff))
1009 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
1010 }
1011
1012 spin_lock(&ci->i_ceph_lock);
1013 ci->i_inline_version = CEPH_INLINE_NONE;
1014 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1015 &aio_req->prealloc_cf);
1016 spin_unlock(&ci->i_ceph_lock);
1017 if (dirty)
1018 __mark_inode_dirty(inode, dirty);
1019
1020 }
1021
1022 ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
1023 CEPH_CAP_FILE_RD));
1024
1025 aio_req->iocb->ki_complete(aio_req->iocb, ret, 0);
1026
1027 ceph_free_cap_flush(aio_req->prealloc_cf);
1028 kfree(aio_req);
1029}
1030
1031static void ceph_aio_complete_req(struct ceph_osd_request *req)
1032{
1033 int rc = req->r_result;
1034 struct inode *inode = req->r_inode;
1035 struct ceph_aio_request *aio_req = req->r_priv;
1036 struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
1037 struct ceph_client_metric *metric = &ceph_sb_to_mdsc(inode->i_sb)->metric;
1038 unsigned int len = osd_data->bvec_pos.iter.bi_size;
1039
1040 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_BVECS);
1041 BUG_ON(!osd_data->num_bvecs);
1042
1043 dout("ceph_aio_complete_req %p rc %d bytes %u\n", inode, rc, len);
1044
1045 if (rc == -EOLDSNAPC) {
1046 struct ceph_aio_work *aio_work;
1047 BUG_ON(!aio_req->write);
1048
1049 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
1050 if (aio_work) {
1051 INIT_WORK(&aio_work->work, ceph_aio_retry_work);
1052 aio_work->req = req;
1053 queue_work(ceph_inode_to_client(inode)->inode_wq,
1054 &aio_work->work);
1055 return;
1056 }
1057 rc = -ENOMEM;
1058 } else if (!aio_req->write) {
1059 if (rc == -ENOENT)
1060 rc = 0;
1061 if (rc >= 0 && len > rc) {
1062 struct iov_iter i;
1063 int zlen = len - rc;
1064
1065 /*
1066 * If read is satisfied by single OSD request,
1067 * it can pass EOF. Otherwise read is within
1068 * i_size.
1069 */
1070 if (aio_req->num_reqs == 1) {
1071 loff_t i_size = i_size_read(inode);
1072 loff_t endoff = aio_req->iocb->ki_pos + rc;
1073 if (endoff < i_size)
1074 zlen = min_t(size_t, zlen,
1075 i_size - endoff);
1076 aio_req->total_len = rc + zlen;
1077 }
1078
1079 iov_iter_bvec(&i, READ, osd_data->bvec_pos.bvecs,
1080 osd_data->num_bvecs, len);
1081 iov_iter_advance(&i, rc);
1082 iov_iter_zero(zlen, &i);
1083 }
1084 }
1085
1086 /* r_start_latency == 0 means the request was not submitted */
1087 if (req->r_start_latency) {
1088 if (aio_req->write)
1089 ceph_update_write_metrics(metric, req->r_start_latency,
1090 req->r_end_latency, len, rc);
1091 else
1092 ceph_update_read_metrics(metric, req->r_start_latency,
1093 req->r_end_latency, len, rc);
1094 }
1095
1096 put_bvecs(osd_data->bvec_pos.bvecs, osd_data->num_bvecs,
1097 aio_req->should_dirty);
1098 ceph_osdc_put_request(req);
1099
1100 if (rc < 0)
1101 cmpxchg(&aio_req->error, 0, rc);
1102
1103 ceph_aio_complete(inode, aio_req);
1104 return;
1105}
1106
1107static void ceph_aio_retry_work(struct work_struct *work)
1108{
1109 struct ceph_aio_work *aio_work =
1110 container_of(work, struct ceph_aio_work, work);
1111 struct ceph_osd_request *orig_req = aio_work->req;
1112 struct ceph_aio_request *aio_req = orig_req->r_priv;
1113 struct inode *inode = orig_req->r_inode;
1114 struct ceph_inode_info *ci = ceph_inode(inode);
1115 struct ceph_snap_context *snapc;
1116 struct ceph_osd_request *req;
1117 int ret;
1118
1119 spin_lock(&ci->i_ceph_lock);
1120 if (__ceph_have_pending_cap_snap(ci)) {
1121 struct ceph_cap_snap *capsnap =
1122 list_last_entry(&ci->i_cap_snaps,
1123 struct ceph_cap_snap,
1124 ci_item);
1125 snapc = ceph_get_snap_context(capsnap->context);
1126 } else {
1127 BUG_ON(!ci->i_head_snapc);
1128 snapc = ceph_get_snap_context(ci->i_head_snapc);
1129 }
1130 spin_unlock(&ci->i_ceph_lock);
1131
1132 req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 1,
1133 false, GFP_NOFS);
1134 if (!req) {
1135 ret = -ENOMEM;
1136 req = orig_req;
1137 goto out;
1138 }
1139
1140 req->r_flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1141 ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
1142 ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
1143
1144 req->r_ops[0] = orig_req->r_ops[0];
1145
1146 req->r_mtime = aio_req->mtime;
1147 req->r_data_offset = req->r_ops[0].extent.offset;
1148
1149 ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
1150 if (ret) {
1151 ceph_osdc_put_request(req);
1152 req = orig_req;
1153 goto out;
1154 }
1155
1156 ceph_osdc_put_request(orig_req);
1157
1158 req->r_callback = ceph_aio_complete_req;
1159 req->r_inode = inode;
1160 req->r_priv = aio_req;
1161
1162 ret = ceph_osdc_start_request(req->r_osdc, req, false);
1163out:
1164 if (ret < 0) {
1165 req->r_result = ret;
1166 ceph_aio_complete_req(req);
1167 }
1168
1169 ceph_put_snap_context(snapc);
1170 kfree(aio_work);
1171}
1172
1173static ssize_t
1174ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
1175 struct ceph_snap_context *snapc,
1176 struct ceph_cap_flush **pcf)
1177{
1178 struct file *file = iocb->ki_filp;
1179 struct inode *inode = file_inode(file);
1180 struct ceph_inode_info *ci = ceph_inode(inode);
1181 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1182 struct ceph_client_metric *metric = &fsc->mdsc->metric;
1183 struct ceph_vino vino;
1184 struct ceph_osd_request *req;
1185 struct bio_vec *bvecs;
1186 struct ceph_aio_request *aio_req = NULL;
1187 int num_pages = 0;
1188 int flags;
1189 int ret = 0;
1190 struct timespec64 mtime = current_time(inode);
1191 size_t count = iov_iter_count(iter);
1192 loff_t pos = iocb->ki_pos;
1193 bool write = iov_iter_rw(iter) == WRITE;
1194 bool should_dirty = !write && iter_is_iovec(iter);
1195
1196 if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1197 return -EROFS;
1198
1199 dout("sync_direct_%s on file %p %lld~%u snapc %p seq %lld\n",
1200 (write ? "write" : "read"), file, pos, (unsigned)count,
1201 snapc, snapc ? snapc->seq : 0);
1202
1203 if (write) {
1204 int ret2 = invalidate_inode_pages2_range(inode->i_mapping,
1205 pos >> PAGE_SHIFT,
1206 (pos + count - 1) >> PAGE_SHIFT);
1207 if (ret2 < 0)
1208 dout("invalidate_inode_pages2_range returned %d\n", ret2);
1209
1210 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1211 } else {
1212 flags = CEPH_OSD_FLAG_READ;
1213 }
1214
1215 while (iov_iter_count(iter) > 0) {
1216 u64 size = iov_iter_count(iter);
1217 ssize_t len;
1218
1219 if (write)
1220 size = min_t(u64, size, fsc->mount_options->wsize);
1221 else
1222 size = min_t(u64, size, fsc->mount_options->rsize);
1223
1224 vino = ceph_vino(inode);
1225 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1226 vino, pos, &size, 0,
1227 1,
1228 write ? CEPH_OSD_OP_WRITE :
1229 CEPH_OSD_OP_READ,
1230 flags, snapc,
1231 ci->i_truncate_seq,
1232 ci->i_truncate_size,
1233 false);
1234 if (IS_ERR(req)) {
1235 ret = PTR_ERR(req);
1236 break;
1237 }
1238
1239 len = iter_get_bvecs_alloc(iter, size, &bvecs, &num_pages);
1240 if (len < 0) {
1241 ceph_osdc_put_request(req);
1242 ret = len;
1243 break;
1244 }
1245 if (len != size)
1246 osd_req_op_extent_update(req, 0, len);
1247
1248 /*
1249 * To simplify error handling, allow AIO when IO within i_size
1250 * or IO can be satisfied by single OSD request.
1251 */
1252 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
1253 (len == count || pos + count <= i_size_read(inode))) {
1254 aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
1255 if (aio_req) {
1256 aio_req->iocb = iocb;
1257 aio_req->write = write;
1258 aio_req->should_dirty = should_dirty;
1259 INIT_LIST_HEAD(&aio_req->osd_reqs);
1260 if (write) {
1261 aio_req->mtime = mtime;
1262 swap(aio_req->prealloc_cf, *pcf);
1263 }
1264 }
1265 /* ignore error */
1266 }
1267
1268 if (write) {
1269 /*
1270 * throw out any page cache pages in this range. this
1271 * may block.
1272 */
1273 truncate_inode_pages_range(inode->i_mapping, pos,
1274 PAGE_ALIGN(pos + len) - 1);
1275
1276 req->r_mtime = mtime;
1277 }
1278
1279 osd_req_op_extent_osd_data_bvecs(req, 0, bvecs, num_pages, len);
1280
1281 if (aio_req) {
1282 aio_req->total_len += len;
1283 aio_req->num_reqs++;
1284 atomic_inc(&aio_req->pending_reqs);
1285
1286 req->r_callback = ceph_aio_complete_req;
1287 req->r_inode = inode;
1288 req->r_priv = aio_req;
1289 list_add_tail(&req->r_private_item, &aio_req->osd_reqs);
1290
1291 pos += len;
1292 continue;
1293 }
1294
1295 ret = ceph_osdc_start_request(req->r_osdc, req, false);
1296 if (!ret)
1297 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1298
1299 if (write)
1300 ceph_update_write_metrics(metric, req->r_start_latency,
1301 req->r_end_latency, len, ret);
1302 else
1303 ceph_update_read_metrics(metric, req->r_start_latency,
1304 req->r_end_latency, len, ret);
1305
1306 size = i_size_read(inode);
1307 if (!write) {
1308 if (ret == -ENOENT)
1309 ret = 0;
1310 if (ret >= 0 && ret < len && pos + ret < size) {
1311 struct iov_iter i;
1312 int zlen = min_t(size_t, len - ret,
1313 size - pos - ret);
1314
1315 iov_iter_bvec(&i, READ, bvecs, num_pages, len);
1316 iov_iter_advance(&i, ret);
1317 iov_iter_zero(zlen, &i);
1318 ret += zlen;
1319 }
1320 if (ret >= 0)
1321 len = ret;
1322 }
1323
1324 put_bvecs(bvecs, num_pages, should_dirty);
1325 ceph_osdc_put_request(req);
1326 if (ret < 0)
1327 break;
1328
1329 pos += len;
1330 if (!write && pos >= size)
1331 break;
1332
1333 if (write && pos > size) {
1334 if (ceph_inode_set_size(inode, pos))
1335 ceph_check_caps(ceph_inode(inode),
1336 CHECK_CAPS_AUTHONLY,
1337 NULL);
1338 }
1339 }
1340
1341 if (aio_req) {
1342 LIST_HEAD(osd_reqs);
1343
1344 if (aio_req->num_reqs == 0) {
1345 kfree(aio_req);
1346 return ret;
1347 }
1348
1349 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1350 CEPH_CAP_FILE_RD);
1351
1352 list_splice(&aio_req->osd_reqs, &osd_reqs);
1353 inode_dio_begin(inode);
1354 while (!list_empty(&osd_reqs)) {
1355 req = list_first_entry(&osd_reqs,
1356 struct ceph_osd_request,
1357 r_private_item);
1358 list_del_init(&req->r_private_item);
1359 if (ret >= 0)
1360 ret = ceph_osdc_start_request(req->r_osdc,
1361 req, false);
1362 if (ret < 0) {
1363 req->r_result = ret;
1364 ceph_aio_complete_req(req);
1365 }
1366 }
1367 return -EIOCBQUEUED;
1368 }
1369
1370 if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1371 ret = pos - iocb->ki_pos;
1372 iocb->ki_pos = pos;
1373 }
1374 return ret;
1375}
1376
1377/*
1378 * Synchronous write, straight from __user pointer or user pages.
1379 *
1380 * If write spans object boundary, just do multiple writes. (For a
1381 * correct atomic write, we should e.g. take write locks on all
1382 * objects, rollback on failure, etc.)
1383 */
1384static ssize_t
1385ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1386 struct ceph_snap_context *snapc)
1387{
1388 struct file *file = iocb->ki_filp;
1389 struct inode *inode = file_inode(file);
1390 struct ceph_inode_info *ci = ceph_inode(inode);
1391 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1392 struct ceph_vino vino;
1393 struct ceph_osd_request *req;
1394 struct page **pages;
1395 u64 len;
1396 int num_pages;
1397 int written = 0;
1398 int flags;
1399 int ret;
1400 bool check_caps = false;
1401 struct timespec64 mtime = current_time(inode);
1402 size_t count = iov_iter_count(from);
1403
1404 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1405 return -EROFS;
1406
1407 dout("sync_write on file %p %lld~%u snapc %p seq %lld\n",
1408 file, pos, (unsigned)count, snapc, snapc->seq);
1409
1410 ret = filemap_write_and_wait_range(inode->i_mapping,
1411 pos, pos + count - 1);
1412 if (ret < 0)
1413 return ret;
1414
1415 ret = invalidate_inode_pages2_range(inode->i_mapping,
1416 pos >> PAGE_SHIFT,
1417 (pos + count - 1) >> PAGE_SHIFT);
1418 if (ret < 0)
1419 dout("invalidate_inode_pages2_range returned %d\n", ret);
1420
1421 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1422
1423 while ((len = iov_iter_count(from)) > 0) {
1424 size_t left;
1425 int n;
1426
1427 vino = ceph_vino(inode);
1428 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1429 vino, pos, &len, 0, 1,
1430 CEPH_OSD_OP_WRITE, flags, snapc,
1431 ci->i_truncate_seq,
1432 ci->i_truncate_size,
1433 false);
1434 if (IS_ERR(req)) {
1435 ret = PTR_ERR(req);
1436 break;
1437 }
1438
1439 /*
1440 * write from beginning of first page,
1441 * regardless of io alignment
1442 */
1443 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1444
1445 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1446 if (IS_ERR(pages)) {
1447 ret = PTR_ERR(pages);
1448 goto out;
1449 }
1450
1451 left = len;
1452 for (n = 0; n < num_pages; n++) {
1453 size_t plen = min_t(size_t, left, PAGE_SIZE);
1454 ret = copy_page_from_iter(pages[n], 0, plen, from);
1455 if (ret != plen) {
1456 ret = -EFAULT;
1457 break;
1458 }
1459 left -= ret;
1460 }
1461
1462 if (ret < 0) {
1463 ceph_release_page_vector(pages, num_pages);
1464 goto out;
1465 }
1466
1467 req->r_inode = inode;
1468
1469 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1470 false, true);
1471
1472 req->r_mtime = mtime;
1473 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1474 if (!ret)
1475 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1476
1477 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1478 req->r_end_latency, len, ret);
1479out:
1480 ceph_osdc_put_request(req);
1481 if (ret != 0) {
1482 ceph_set_error_write(ci);
1483 break;
1484 }
1485
1486 ceph_clear_error_write(ci);
1487 pos += len;
1488 written += len;
1489 if (pos > i_size_read(inode)) {
1490 check_caps = ceph_inode_set_size(inode, pos);
1491 if (check_caps)
1492 ceph_check_caps(ceph_inode(inode),
1493 CHECK_CAPS_AUTHONLY,
1494 NULL);
1495 }
1496
1497 }
1498
1499 if (ret != -EOLDSNAPC && written > 0) {
1500 ret = written;
1501 iocb->ki_pos = pos;
1502 }
1503 return ret;
1504}
1505
1506/*
1507 * Wrap generic_file_aio_read with checks for cap bits on the inode.
1508 * Atomically grab references, so that those bits are not released
1509 * back to the MDS mid-read.
1510 *
1511 * Hmm, the sync read case isn't actually async... should it be?
1512 */
1513static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1514{
1515 struct file *filp = iocb->ki_filp;
1516 struct ceph_file_info *fi = filp->private_data;
1517 size_t len = iov_iter_count(to);
1518 struct inode *inode = file_inode(filp);
1519 struct ceph_inode_info *ci = ceph_inode(inode);
1520 bool direct_lock = iocb->ki_flags & IOCB_DIRECT;
1521 ssize_t ret;
1522 int want, got = 0;
1523 int retry_op = 0, read = 0;
1524
1525again:
1526 dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1527 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1528
1529 if (direct_lock)
1530 ceph_start_io_direct(inode);
1531 else
1532 ceph_start_io_read(inode);
1533
1534 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1535 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1536 else
1537 want = CEPH_CAP_FILE_CACHE;
1538 ret = ceph_get_caps(filp, CEPH_CAP_FILE_RD, want, -1, &got);
1539 if (ret < 0) {
1540 if (iocb->ki_flags & IOCB_DIRECT)
1541 ceph_end_io_direct(inode);
1542 else
1543 ceph_end_io_read(inode);
1544 return ret;
1545 }
1546
1547 if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1548 (iocb->ki_flags & IOCB_DIRECT) ||
1549 (fi->flags & CEPH_F_SYNC)) {
1550
1551 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1552 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1553 ceph_cap_string(got));
1554
1555 if (ci->i_inline_version == CEPH_INLINE_NONE) {
1556 if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1557 ret = ceph_direct_read_write(iocb, to,
1558 NULL, NULL);
1559 if (ret >= 0 && ret < len)
1560 retry_op = CHECK_EOF;
1561 } else {
1562 ret = ceph_sync_read(iocb, to, &retry_op);
1563 }
1564 } else {
1565 retry_op = READ_INLINE;
1566 }
1567 } else {
1568 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1569 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1570 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1571 ceph_cap_string(got));
1572 ceph_add_rw_context(fi, &rw_ctx);
1573 ret = generic_file_read_iter(iocb, to);
1574 ceph_del_rw_context(fi, &rw_ctx);
1575 }
1576
1577 dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1578 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1579 ceph_put_cap_refs(ci, got);
1580
1581 if (direct_lock)
1582 ceph_end_io_direct(inode);
1583 else
1584 ceph_end_io_read(inode);
1585
1586 if (retry_op > HAVE_RETRIED && ret >= 0) {
1587 int statret;
1588 struct page *page = NULL;
1589 loff_t i_size;
1590 if (retry_op == READ_INLINE) {
1591 page = __page_cache_alloc(GFP_KERNEL);
1592 if (!page)
1593 return -ENOMEM;
1594 }
1595
1596 statret = __ceph_do_getattr(inode, page,
1597 CEPH_STAT_CAP_INLINE_DATA, !!page);
1598 if (statret < 0) {
1599 if (page)
1600 __free_page(page);
1601 if (statret == -ENODATA) {
1602 BUG_ON(retry_op != READ_INLINE);
1603 goto again;
1604 }
1605 return statret;
1606 }
1607
1608 i_size = i_size_read(inode);
1609 if (retry_op == READ_INLINE) {
1610 BUG_ON(ret > 0 || read > 0);
1611 if (iocb->ki_pos < i_size &&
1612 iocb->ki_pos < PAGE_SIZE) {
1613 loff_t end = min_t(loff_t, i_size,
1614 iocb->ki_pos + len);
1615 end = min_t(loff_t, end, PAGE_SIZE);
1616 if (statret < end)
1617 zero_user_segment(page, statret, end);
1618 ret = copy_page_to_iter(page,
1619 iocb->ki_pos & ~PAGE_MASK,
1620 end - iocb->ki_pos, to);
1621 iocb->ki_pos += ret;
1622 read += ret;
1623 }
1624 if (iocb->ki_pos < i_size && read < len) {
1625 size_t zlen = min_t(size_t, len - read,
1626 i_size - iocb->ki_pos);
1627 ret = iov_iter_zero(zlen, to);
1628 iocb->ki_pos += ret;
1629 read += ret;
1630 }
1631 __free_pages(page, 0);
1632 return read;
1633 }
1634
1635 /* hit EOF or hole? */
1636 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1637 ret < len) {
1638 dout("sync_read hit hole, ppos %lld < size %lld"
1639 ", reading more\n", iocb->ki_pos, i_size);
1640
1641 read += ret;
1642 len -= ret;
1643 retry_op = HAVE_RETRIED;
1644 goto again;
1645 }
1646 }
1647
1648 if (ret >= 0)
1649 ret += read;
1650
1651 return ret;
1652}
1653
1654/*
1655 * Take cap references to avoid releasing caps to MDS mid-write.
1656 *
1657 * If we are synchronous, and write with an old snap context, the OSD
1658 * may return EOLDSNAPC. In that case, retry the write.. _after_
1659 * dropping our cap refs and allowing the pending snap to logically
1660 * complete _before_ this write occurs.
1661 *
1662 * If we are near ENOSPC, write synchronously.
1663 */
1664static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1665{
1666 struct file *file = iocb->ki_filp;
1667 struct ceph_file_info *fi = file->private_data;
1668 struct inode *inode = file_inode(file);
1669 struct ceph_inode_info *ci = ceph_inode(inode);
1670 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1671 struct ceph_osd_client *osdc = &fsc->client->osdc;
1672 struct ceph_cap_flush *prealloc_cf;
1673 ssize_t count, written = 0;
1674 int err, want, got;
1675 bool direct_lock = false;
1676 u32 map_flags;
1677 u64 pool_flags;
1678 loff_t pos;
1679 loff_t limit = max(i_size_read(inode), fsc->max_file_size);
1680
1681 if (ceph_snap(inode) != CEPH_NOSNAP)
1682 return -EROFS;
1683
1684 prealloc_cf = ceph_alloc_cap_flush();
1685 if (!prealloc_cf)
1686 return -ENOMEM;
1687
1688 if ((iocb->ki_flags & (IOCB_DIRECT | IOCB_APPEND)) == IOCB_DIRECT)
1689 direct_lock = true;
1690
1691retry_snap:
1692 if (direct_lock)
1693 ceph_start_io_direct(inode);
1694 else
1695 ceph_start_io_write(inode);
1696
1697 /* We can write back this queue in page reclaim */
1698 current->backing_dev_info = inode_to_bdi(inode);
1699
1700 if (iocb->ki_flags & IOCB_APPEND) {
1701 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1702 if (err < 0)
1703 goto out;
1704 }
1705
1706 err = generic_write_checks(iocb, from);
1707 if (err <= 0)
1708 goto out;
1709
1710 pos = iocb->ki_pos;
1711 if (unlikely(pos >= limit)) {
1712 err = -EFBIG;
1713 goto out;
1714 } else {
1715 iov_iter_truncate(from, limit - pos);
1716 }
1717
1718 count = iov_iter_count(from);
1719 if (ceph_quota_is_max_bytes_exceeded(inode, pos + count)) {
1720 err = -EDQUOT;
1721 goto out;
1722 }
1723
1724 down_read(&osdc->lock);
1725 map_flags = osdc->osdmap->flags;
1726 pool_flags = ceph_pg_pool_flags(osdc->osdmap, ci->i_layout.pool_id);
1727 up_read(&osdc->lock);
1728 if ((map_flags & CEPH_OSDMAP_FULL) ||
1729 (pool_flags & CEPH_POOL_FLAG_FULL)) {
1730 err = -ENOSPC;
1731 goto out;
1732 }
1733
1734 err = file_remove_privs(file);
1735 if (err)
1736 goto out;
1737
1738 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1739 err = ceph_uninline_data(file, NULL);
1740 if (err < 0)
1741 goto out;
1742 }
1743
1744 dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1745 inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1746 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1747 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1748 else
1749 want = CEPH_CAP_FILE_BUFFER;
1750 got = 0;
1751 err = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, pos + count, &got);
1752 if (err < 0)
1753 goto out;
1754
1755 err = file_update_time(file);
1756 if (err)
1757 goto out_caps;
1758
1759 inode_inc_iversion_raw(inode);
1760
1761 dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1762 inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1763
1764 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1765 (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
1766 (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
1767 struct ceph_snap_context *snapc;
1768 struct iov_iter data;
1769
1770 spin_lock(&ci->i_ceph_lock);
1771 if (__ceph_have_pending_cap_snap(ci)) {
1772 struct ceph_cap_snap *capsnap =
1773 list_last_entry(&ci->i_cap_snaps,
1774 struct ceph_cap_snap,
1775 ci_item);
1776 snapc = ceph_get_snap_context(capsnap->context);
1777 } else {
1778 BUG_ON(!ci->i_head_snapc);
1779 snapc = ceph_get_snap_context(ci->i_head_snapc);
1780 }
1781 spin_unlock(&ci->i_ceph_lock);
1782
1783 /* we might need to revert back to that point */
1784 data = *from;
1785 if (iocb->ki_flags & IOCB_DIRECT)
1786 written = ceph_direct_read_write(iocb, &data, snapc,
1787 &prealloc_cf);
1788 else
1789 written = ceph_sync_write(iocb, &data, pos, snapc);
1790 if (direct_lock)
1791 ceph_end_io_direct(inode);
1792 else
1793 ceph_end_io_write(inode);
1794 if (written > 0)
1795 iov_iter_advance(from, written);
1796 ceph_put_snap_context(snapc);
1797 } else {
1798 /*
1799 * No need to acquire the i_truncate_mutex. Because
1800 * the MDS revokes Fwb caps before sending truncate
1801 * message to us. We can't get Fwb cap while there
1802 * are pending vmtruncate. So write and vmtruncate
1803 * can not run at the same time
1804 */
1805 written = generic_perform_write(file, from, pos);
1806 if (likely(written >= 0))
1807 iocb->ki_pos = pos + written;
1808 ceph_end_io_write(inode);
1809 }
1810
1811 if (written >= 0) {
1812 int dirty;
1813
1814 spin_lock(&ci->i_ceph_lock);
1815 ci->i_inline_version = CEPH_INLINE_NONE;
1816 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1817 &prealloc_cf);
1818 spin_unlock(&ci->i_ceph_lock);
1819 if (dirty)
1820 __mark_inode_dirty(inode, dirty);
1821 if (ceph_quota_is_max_bytes_approaching(inode, iocb->ki_pos))
1822 ceph_check_caps(ci, 0, NULL);
1823 }
1824
1825 dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n",
1826 inode, ceph_vinop(inode), pos, (unsigned)count,
1827 ceph_cap_string(got));
1828 ceph_put_cap_refs(ci, got);
1829
1830 if (written == -EOLDSNAPC) {
1831 dout("aio_write %p %llx.%llx %llu~%u" "got EOLDSNAPC, retrying\n",
1832 inode, ceph_vinop(inode), pos, (unsigned)count);
1833 goto retry_snap;
1834 }
1835
1836 if (written >= 0) {
1837 if ((map_flags & CEPH_OSDMAP_NEARFULL) ||
1838 (pool_flags & CEPH_POOL_FLAG_NEARFULL))
1839 iocb->ki_flags |= IOCB_DSYNC;
1840 written = generic_write_sync(iocb, written);
1841 }
1842
1843 goto out_unlocked;
1844out_caps:
1845 ceph_put_cap_refs(ci, got);
1846out:
1847 if (direct_lock)
1848 ceph_end_io_direct(inode);
1849 else
1850 ceph_end_io_write(inode);
1851out_unlocked:
1852 ceph_free_cap_flush(prealloc_cf);
1853 current->backing_dev_info = NULL;
1854 return written ? written : err;
1855}
1856
1857/*
1858 * llseek. be sure to verify file size on SEEK_END.
1859 */
1860static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1861{
1862 struct inode *inode = file->f_mapping->host;
1863 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1864 loff_t i_size;
1865 loff_t ret;
1866
1867 inode_lock(inode);
1868
1869 if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1870 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1871 if (ret < 0)
1872 goto out;
1873 }
1874
1875 i_size = i_size_read(inode);
1876 switch (whence) {
1877 case SEEK_END:
1878 offset += i_size;
1879 break;
1880 case SEEK_CUR:
1881 /*
1882 * Here we special-case the lseek(fd, 0, SEEK_CUR)
1883 * position-querying operation. Avoid rewriting the "same"
1884 * f_pos value back to the file because a concurrent read(),
1885 * write() or lseek() might have altered it
1886 */
1887 if (offset == 0) {
1888 ret = file->f_pos;
1889 goto out;
1890 }
1891 offset += file->f_pos;
1892 break;
1893 case SEEK_DATA:
1894 if (offset < 0 || offset >= i_size) {
1895 ret = -ENXIO;
1896 goto out;
1897 }
1898 break;
1899 case SEEK_HOLE:
1900 if (offset < 0 || offset >= i_size) {
1901 ret = -ENXIO;
1902 goto out;
1903 }
1904 offset = i_size;
1905 break;
1906 }
1907
1908 ret = vfs_setpos(file, offset, max(i_size, fsc->max_file_size));
1909
1910out:
1911 inode_unlock(inode);
1912 return ret;
1913}
1914
1915static inline void ceph_zero_partial_page(
1916 struct inode *inode, loff_t offset, unsigned size)
1917{
1918 struct page *page;
1919 pgoff_t index = offset >> PAGE_SHIFT;
1920
1921 page = find_lock_page(inode->i_mapping, index);
1922 if (page) {
1923 wait_on_page_writeback(page);
1924 zero_user(page, offset & (PAGE_SIZE - 1), size);
1925 unlock_page(page);
1926 put_page(page);
1927 }
1928}
1929
1930static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1931 loff_t length)
1932{
1933 loff_t nearly = round_up(offset, PAGE_SIZE);
1934 if (offset < nearly) {
1935 loff_t size = nearly - offset;
1936 if (length < size)
1937 size = length;
1938 ceph_zero_partial_page(inode, offset, size);
1939 offset += size;
1940 length -= size;
1941 }
1942 if (length >= PAGE_SIZE) {
1943 loff_t size = round_down(length, PAGE_SIZE);
1944 truncate_pagecache_range(inode, offset, offset + size - 1);
1945 offset += size;
1946 length -= size;
1947 }
1948 if (length)
1949 ceph_zero_partial_page(inode, offset, length);
1950}
1951
1952static int ceph_zero_partial_object(struct inode *inode,
1953 loff_t offset, loff_t *length)
1954{
1955 struct ceph_inode_info *ci = ceph_inode(inode);
1956 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1957 struct ceph_osd_request *req;
1958 int ret = 0;
1959 loff_t zero = 0;
1960 int op;
1961
1962 if (!length) {
1963 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1964 length = &zero;
1965 } else {
1966 op = CEPH_OSD_OP_ZERO;
1967 }
1968
1969 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1970 ceph_vino(inode),
1971 offset, length,
1972 0, 1, op,
1973 CEPH_OSD_FLAG_WRITE,
1974 NULL, 0, 0, false);
1975 if (IS_ERR(req)) {
1976 ret = PTR_ERR(req);
1977 goto out;
1978 }
1979
1980 req->r_mtime = inode->i_mtime;
1981 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1982 if (!ret) {
1983 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1984 if (ret == -ENOENT)
1985 ret = 0;
1986 }
1987 ceph_osdc_put_request(req);
1988
1989out:
1990 return ret;
1991}
1992
1993static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
1994{
1995 int ret = 0;
1996 struct ceph_inode_info *ci = ceph_inode(inode);
1997 s32 stripe_unit = ci->i_layout.stripe_unit;
1998 s32 stripe_count = ci->i_layout.stripe_count;
1999 s32 object_size = ci->i_layout.object_size;
2000 u64 object_set_size = object_size * stripe_count;
2001 u64 nearly, t;
2002
2003 /* round offset up to next period boundary */
2004 nearly = offset + object_set_size - 1;
2005 t = nearly;
2006 nearly -= do_div(t, object_set_size);
2007
2008 while (length && offset < nearly) {
2009 loff_t size = length;
2010 ret = ceph_zero_partial_object(inode, offset, &size);
2011 if (ret < 0)
2012 return ret;
2013 offset += size;
2014 length -= size;
2015 }
2016 while (length >= object_set_size) {
2017 int i;
2018 loff_t pos = offset;
2019 for (i = 0; i < stripe_count; ++i) {
2020 ret = ceph_zero_partial_object(inode, pos, NULL);
2021 if (ret < 0)
2022 return ret;
2023 pos += stripe_unit;
2024 }
2025 offset += object_set_size;
2026 length -= object_set_size;
2027 }
2028 while (length) {
2029 loff_t size = length;
2030 ret = ceph_zero_partial_object(inode, offset, &size);
2031 if (ret < 0)
2032 return ret;
2033 offset += size;
2034 length -= size;
2035 }
2036 return ret;
2037}
2038
2039static long ceph_fallocate(struct file *file, int mode,
2040 loff_t offset, loff_t length)
2041{
2042 struct ceph_file_info *fi = file->private_data;
2043 struct inode *inode = file_inode(file);
2044 struct ceph_inode_info *ci = ceph_inode(inode);
2045 struct ceph_cap_flush *prealloc_cf;
2046 int want, got = 0;
2047 int dirty;
2048 int ret = 0;
2049 loff_t endoff = 0;
2050 loff_t size;
2051
2052 if (mode != (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2053 return -EOPNOTSUPP;
2054
2055 if (!S_ISREG(inode->i_mode))
2056 return -EOPNOTSUPP;
2057
2058 prealloc_cf = ceph_alloc_cap_flush();
2059 if (!prealloc_cf)
2060 return -ENOMEM;
2061
2062 inode_lock(inode);
2063
2064 if (ceph_snap(inode) != CEPH_NOSNAP) {
2065 ret = -EROFS;
2066 goto unlock;
2067 }
2068
2069 if (ci->i_inline_version != CEPH_INLINE_NONE) {
2070 ret = ceph_uninline_data(file, NULL);
2071 if (ret < 0)
2072 goto unlock;
2073 }
2074
2075 size = i_size_read(inode);
2076
2077 /* Are we punching a hole beyond EOF? */
2078 if (offset >= size)
2079 goto unlock;
2080 if ((offset + length) > size)
2081 length = size - offset;
2082
2083 if (fi->fmode & CEPH_FILE_MODE_LAZY)
2084 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
2085 else
2086 want = CEPH_CAP_FILE_BUFFER;
2087
2088 ret = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, endoff, &got);
2089 if (ret < 0)
2090 goto unlock;
2091
2092 ceph_zero_pagecache_range(inode, offset, length);
2093 ret = ceph_zero_objects(inode, offset, length);
2094
2095 if (!ret) {
2096 spin_lock(&ci->i_ceph_lock);
2097 ci->i_inline_version = CEPH_INLINE_NONE;
2098 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
2099 &prealloc_cf);
2100 spin_unlock(&ci->i_ceph_lock);
2101 if (dirty)
2102 __mark_inode_dirty(inode, dirty);
2103 }
2104
2105 ceph_put_cap_refs(ci, got);
2106unlock:
2107 inode_unlock(inode);
2108 ceph_free_cap_flush(prealloc_cf);
2109 return ret;
2110}
2111
2112/*
2113 * This function tries to get FILE_WR capabilities for dst_ci and FILE_RD for
2114 * src_ci. Two attempts are made to obtain both caps, and an error is return if
2115 * this fails; zero is returned on success.
2116 */
2117static int get_rd_wr_caps(struct file *src_filp, int *src_got,
2118 struct file *dst_filp,
2119 loff_t dst_endoff, int *dst_got)
2120{
2121 int ret = 0;
2122 bool retrying = false;
2123
2124retry_caps:
2125 ret = ceph_get_caps(dst_filp, CEPH_CAP_FILE_WR, CEPH_CAP_FILE_BUFFER,
2126 dst_endoff, dst_got);
2127 if (ret < 0)
2128 return ret;
2129
2130 /*
2131 * Since we're already holding the FILE_WR capability for the dst file,
2132 * we would risk a deadlock by using ceph_get_caps. Thus, we'll do some
2133 * retry dance instead to try to get both capabilities.
2134 */
2135 ret = ceph_try_get_caps(file_inode(src_filp),
2136 CEPH_CAP_FILE_RD, CEPH_CAP_FILE_SHARED,
2137 false, src_got);
2138 if (ret <= 0) {
2139 /* Start by dropping dst_ci caps and getting src_ci caps */
2140 ceph_put_cap_refs(ceph_inode(file_inode(dst_filp)), *dst_got);
2141 if (retrying) {
2142 if (!ret)
2143 /* ceph_try_get_caps masks EAGAIN */
2144 ret = -EAGAIN;
2145 return ret;
2146 }
2147 ret = ceph_get_caps(src_filp, CEPH_CAP_FILE_RD,
2148 CEPH_CAP_FILE_SHARED, -1, src_got);
2149 if (ret < 0)
2150 return ret;
2151 /*... drop src_ci caps too, and retry */
2152 ceph_put_cap_refs(ceph_inode(file_inode(src_filp)), *src_got);
2153 retrying = true;
2154 goto retry_caps;
2155 }
2156 return ret;
2157}
2158
2159static void put_rd_wr_caps(struct ceph_inode_info *src_ci, int src_got,
2160 struct ceph_inode_info *dst_ci, int dst_got)
2161{
2162 ceph_put_cap_refs(src_ci, src_got);
2163 ceph_put_cap_refs(dst_ci, dst_got);
2164}
2165
2166/*
2167 * This function does several size-related checks, returning an error if:
2168 * - source file is smaller than off+len
2169 * - destination file size is not OK (inode_newsize_ok())
2170 * - max bytes quotas is exceeded
2171 */
2172static int is_file_size_ok(struct inode *src_inode, struct inode *dst_inode,
2173 loff_t src_off, loff_t dst_off, size_t len)
2174{
2175 loff_t size, endoff;
2176
2177 size = i_size_read(src_inode);
2178 /*
2179 * Don't copy beyond source file EOF. Instead of simply setting length
2180 * to (size - src_off), just drop to VFS default implementation, as the
2181 * local i_size may be stale due to other clients writing to the source
2182 * inode.
2183 */
2184 if (src_off + len > size) {
2185 dout("Copy beyond EOF (%llu + %zu > %llu)\n",
2186 src_off, len, size);
2187 return -EOPNOTSUPP;
2188 }
2189 size = i_size_read(dst_inode);
2190
2191 endoff = dst_off + len;
2192 if (inode_newsize_ok(dst_inode, endoff))
2193 return -EOPNOTSUPP;
2194
2195 if (ceph_quota_is_max_bytes_exceeded(dst_inode, endoff))
2196 return -EDQUOT;
2197
2198 return 0;
2199}
2200
2201static ssize_t ceph_do_objects_copy(struct ceph_inode_info *src_ci, u64 *src_off,
2202 struct ceph_inode_info *dst_ci, u64 *dst_off,
2203 struct ceph_fs_client *fsc,
2204 size_t len, unsigned int flags)
2205{
2206 struct ceph_object_locator src_oloc, dst_oloc;
2207 struct ceph_object_id src_oid, dst_oid;
2208 size_t bytes = 0;
2209 u64 src_objnum, src_objoff, dst_objnum, dst_objoff;
2210 u32 src_objlen, dst_objlen;
2211 u32 object_size = src_ci->i_layout.object_size;
2212 int ret;
2213
2214 src_oloc.pool = src_ci->i_layout.pool_id;
2215 src_oloc.pool_ns = ceph_try_get_string(src_ci->i_layout.pool_ns);
2216 dst_oloc.pool = dst_ci->i_layout.pool_id;
2217 dst_oloc.pool_ns = ceph_try_get_string(dst_ci->i_layout.pool_ns);
2218
2219 while (len >= object_size) {
2220 ceph_calc_file_object_mapping(&src_ci->i_layout, *src_off,
2221 object_size, &src_objnum,
2222 &src_objoff, &src_objlen);
2223 ceph_calc_file_object_mapping(&dst_ci->i_layout, *dst_off,
2224 object_size, &dst_objnum,
2225 &dst_objoff, &dst_objlen);
2226 ceph_oid_init(&src_oid);
2227 ceph_oid_printf(&src_oid, "%llx.%08llx",
2228 src_ci->i_vino.ino, src_objnum);
2229 ceph_oid_init(&dst_oid);
2230 ceph_oid_printf(&dst_oid, "%llx.%08llx",
2231 dst_ci->i_vino.ino, dst_objnum);
2232 /* Do an object remote copy */
2233 ret = ceph_osdc_copy_from(&fsc->client->osdc,
2234 src_ci->i_vino.snap, 0,
2235 &src_oid, &src_oloc,
2236 CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2237 CEPH_OSD_OP_FLAG_FADVISE_NOCACHE,
2238 &dst_oid, &dst_oloc,
2239 CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2240 CEPH_OSD_OP_FLAG_FADVISE_DONTNEED,
2241 dst_ci->i_truncate_seq,
2242 dst_ci->i_truncate_size,
2243 CEPH_OSD_COPY_FROM_FLAG_TRUNCATE_SEQ);
2244 if (ret) {
2245 if (ret == -EOPNOTSUPP) {
2246 fsc->have_copy_from2 = false;
2247 pr_notice("OSDs don't support copy-from2; disabling copy offload\n");
2248 }
2249 dout("ceph_osdc_copy_from returned %d\n", ret);
2250 if (!bytes)
2251 bytes = ret;
2252 goto out;
2253 }
2254 len -= object_size;
2255 bytes += object_size;
2256 *src_off += object_size;
2257 *dst_off += object_size;
2258 }
2259
2260out:
2261 ceph_oloc_destroy(&src_oloc);
2262 ceph_oloc_destroy(&dst_oloc);
2263 return bytes;
2264}
2265
2266static ssize_t __ceph_copy_file_range(struct file *src_file, loff_t src_off,
2267 struct file *dst_file, loff_t dst_off,
2268 size_t len, unsigned int flags)
2269{
2270 struct inode *src_inode = file_inode(src_file);
2271 struct inode *dst_inode = file_inode(dst_file);
2272 struct ceph_inode_info *src_ci = ceph_inode(src_inode);
2273 struct ceph_inode_info *dst_ci = ceph_inode(dst_inode);
2274 struct ceph_cap_flush *prealloc_cf;
2275 struct ceph_fs_client *src_fsc = ceph_inode_to_client(src_inode);
2276 loff_t size;
2277 ssize_t ret = -EIO, bytes;
2278 u64 src_objnum, dst_objnum, src_objoff, dst_objoff;
2279 u32 src_objlen, dst_objlen;
2280 int src_got = 0, dst_got = 0, err, dirty;
2281
2282 if (src_inode->i_sb != dst_inode->i_sb) {
2283 struct ceph_fs_client *dst_fsc = ceph_inode_to_client(dst_inode);
2284
2285 if (ceph_fsid_compare(&src_fsc->client->fsid,
2286 &dst_fsc->client->fsid)) {
2287 dout("Copying files across clusters: src: %pU dst: %pU\n",
2288 &src_fsc->client->fsid, &dst_fsc->client->fsid);
2289 return -EXDEV;
2290 }
2291 }
2292 if (ceph_snap(dst_inode) != CEPH_NOSNAP)
2293 return -EROFS;
2294
2295 /*
2296 * Some of the checks below will return -EOPNOTSUPP, which will force a
2297 * fallback to the default VFS copy_file_range implementation. This is
2298 * desirable in several cases (for ex, the 'len' is smaller than the
2299 * size of the objects, or in cases where that would be more
2300 * efficient).
2301 */
2302
2303 if (ceph_test_mount_opt(src_fsc, NOCOPYFROM))
2304 return -EOPNOTSUPP;
2305
2306 if (!src_fsc->have_copy_from2)
2307 return -EOPNOTSUPP;
2308
2309 /*
2310 * Striped file layouts require that we copy partial objects, but the
2311 * OSD copy-from operation only supports full-object copies. Limit
2312 * this to non-striped file layouts for now.
2313 */
2314 if ((src_ci->i_layout.stripe_unit != dst_ci->i_layout.stripe_unit) ||
2315 (src_ci->i_layout.stripe_count != 1) ||
2316 (dst_ci->i_layout.stripe_count != 1) ||
2317 (src_ci->i_layout.object_size != dst_ci->i_layout.object_size)) {
2318 dout("Invalid src/dst files layout\n");
2319 return -EOPNOTSUPP;
2320 }
2321
2322 if (len < src_ci->i_layout.object_size)
2323 return -EOPNOTSUPP; /* no remote copy will be done */
2324
2325 prealloc_cf = ceph_alloc_cap_flush();
2326 if (!prealloc_cf)
2327 return -ENOMEM;
2328
2329 /* Start by sync'ing the source and destination files */
2330 ret = file_write_and_wait_range(src_file, src_off, (src_off + len));
2331 if (ret < 0) {
2332 dout("failed to write src file (%zd)\n", ret);
2333 goto out;
2334 }
2335 ret = file_write_and_wait_range(dst_file, dst_off, (dst_off + len));
2336 if (ret < 0) {
2337 dout("failed to write dst file (%zd)\n", ret);
2338 goto out;
2339 }
2340
2341 /*
2342 * We need FILE_WR caps for dst_ci and FILE_RD for src_ci as other
2343 * clients may have dirty data in their caches. And OSDs know nothing
2344 * about caps, so they can't safely do the remote object copies.
2345 */
2346 err = get_rd_wr_caps(src_file, &src_got,
2347 dst_file, (dst_off + len), &dst_got);
2348 if (err < 0) {
2349 dout("get_rd_wr_caps returned %d\n", err);
2350 ret = -EOPNOTSUPP;
2351 goto out;
2352 }
2353
2354 ret = is_file_size_ok(src_inode, dst_inode, src_off, dst_off, len);
2355 if (ret < 0)
2356 goto out_caps;
2357
2358 /* Drop dst file cached pages */
2359 ret = invalidate_inode_pages2_range(dst_inode->i_mapping,
2360 dst_off >> PAGE_SHIFT,
2361 (dst_off + len) >> PAGE_SHIFT);
2362 if (ret < 0) {
2363 dout("Failed to invalidate inode pages (%zd)\n", ret);
2364 ret = 0; /* XXX */
2365 }
2366 ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
2367 src_ci->i_layout.object_size,
2368 &src_objnum, &src_objoff, &src_objlen);
2369 ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
2370 dst_ci->i_layout.object_size,
2371 &dst_objnum, &dst_objoff, &dst_objlen);
2372 /* object-level offsets need to the same */
2373 if (src_objoff != dst_objoff) {
2374 ret = -EOPNOTSUPP;
2375 goto out_caps;
2376 }
2377
2378 /*
2379 * Do a manual copy if the object offset isn't object aligned.
2380 * 'src_objlen' contains the bytes left until the end of the object,
2381 * starting at the src_off
2382 */
2383 if (src_objoff) {
2384 dout("Initial partial copy of %u bytes\n", src_objlen);
2385
2386 /*
2387 * we need to temporarily drop all caps as we'll be calling
2388 * {read,write}_iter, which will get caps again.
2389 */
2390 put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2391 ret = do_splice_direct(src_file, &src_off, dst_file,
2392 &dst_off, src_objlen, flags);
2393 /* Abort on short copies or on error */
2394 if (ret < src_objlen) {
2395 dout("Failed partial copy (%zd)\n", ret);
2396 goto out;
2397 }
2398 len -= ret;
2399 err = get_rd_wr_caps(src_file, &src_got,
2400 dst_file, (dst_off + len), &dst_got);
2401 if (err < 0)
2402 goto out;
2403 err = is_file_size_ok(src_inode, dst_inode,
2404 src_off, dst_off, len);
2405 if (err < 0)
2406 goto out_caps;
2407 }
2408
2409 size = i_size_read(dst_inode);
2410 bytes = ceph_do_objects_copy(src_ci, &src_off, dst_ci, &dst_off,
2411 src_fsc, len, flags);
2412 if (bytes <= 0) {
2413 if (!ret)
2414 ret = bytes;
2415 goto out_caps;
2416 }
2417 dout("Copied %zu bytes out of %zu\n", bytes, len);
2418 len -= bytes;
2419 ret += bytes;
2420
2421 file_update_time(dst_file);
2422 inode_inc_iversion_raw(dst_inode);
2423
2424 if (dst_off > size) {
2425 /* Let the MDS know about dst file size change */
2426 if (ceph_inode_set_size(dst_inode, dst_off) ||
2427 ceph_quota_is_max_bytes_approaching(dst_inode, dst_off))
2428 ceph_check_caps(dst_ci, CHECK_CAPS_AUTHONLY, NULL);
2429 }
2430 /* Mark Fw dirty */
2431 spin_lock(&dst_ci->i_ceph_lock);
2432 dst_ci->i_inline_version = CEPH_INLINE_NONE;
2433 dirty = __ceph_mark_dirty_caps(dst_ci, CEPH_CAP_FILE_WR, &prealloc_cf);
2434 spin_unlock(&dst_ci->i_ceph_lock);
2435 if (dirty)
2436 __mark_inode_dirty(dst_inode, dirty);
2437
2438out_caps:
2439 put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2440
2441 /*
2442 * Do the final manual copy if we still have some bytes left, unless
2443 * there were errors in remote object copies (len >= object_size).
2444 */
2445 if (len && (len < src_ci->i_layout.object_size)) {
2446 dout("Final partial copy of %zu bytes\n", len);
2447 bytes = do_splice_direct(src_file, &src_off, dst_file,
2448 &dst_off, len, flags);
2449 if (bytes > 0)
2450 ret += bytes;
2451 else
2452 dout("Failed partial copy (%zd)\n", bytes);
2453 }
2454
2455out:
2456 ceph_free_cap_flush(prealloc_cf);
2457
2458 return ret;
2459}
2460
2461static ssize_t ceph_copy_file_range(struct file *src_file, loff_t src_off,
2462 struct file *dst_file, loff_t dst_off,
2463 size_t len, unsigned int flags)
2464{
2465 ssize_t ret;
2466
2467 ret = __ceph_copy_file_range(src_file, src_off, dst_file, dst_off,
2468 len, flags);
2469
2470 if (ret == -EOPNOTSUPP || ret == -EXDEV)
2471 ret = generic_copy_file_range(src_file, src_off, dst_file,
2472 dst_off, len, flags);
2473 return ret;
2474}
2475
2476const struct file_operations ceph_file_fops = {
2477 .open = ceph_open,
2478 .release = ceph_release,
2479 .llseek = ceph_llseek,
2480 .read_iter = ceph_read_iter,
2481 .write_iter = ceph_write_iter,
2482 .mmap = ceph_mmap,
2483 .fsync = ceph_fsync,
2484 .lock = ceph_lock,
2485 .setlease = simple_nosetlease,
2486 .flock = ceph_flock,
2487 .splice_read = generic_file_splice_read,
2488 .splice_write = iter_file_splice_write,
2489 .unlocked_ioctl = ceph_ioctl,
2490 .compat_ioctl = compat_ptr_ioctl,
2491 .fallocate = ceph_fallocate,
2492 .copy_file_range = ceph_copy_file_range,
2493};