xref: /openbmc/linux/fs/ceph/file.c (revision ae213c44)
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 
14 #include "super.h"
15 #include "mds_client.h"
16 #include "cache.h"
17 
18 static __le32 ceph_flags_sys2wire(u32 flags)
19 {
20 	u32 wire_flags = 0;
21 
22 	switch (flags & O_ACCMODE) {
23 	case O_RDONLY:
24 		wire_flags |= CEPH_O_RDONLY;
25 		break;
26 	case O_WRONLY:
27 		wire_flags |= CEPH_O_WRONLY;
28 		break;
29 	case O_RDWR:
30 		wire_flags |= CEPH_O_RDWR;
31 		break;
32 	}
33 
34 	flags &= ~O_ACCMODE;
35 
36 #define ceph_sys2wire(a) if (flags & a) { wire_flags |= CEPH_##a; flags &= ~a; }
37 
38 	ceph_sys2wire(O_CREAT);
39 	ceph_sys2wire(O_EXCL);
40 	ceph_sys2wire(O_TRUNC);
41 	ceph_sys2wire(O_DIRECTORY);
42 	ceph_sys2wire(O_NOFOLLOW);
43 
44 #undef ceph_sys2wire
45 
46 	if (flags)
47 		dout("unused open flags: %x\n", flags);
48 
49 	return cpu_to_le32(wire_flags);
50 }
51 
52 /*
53  * Ceph file operations
54  *
55  * Implement basic open/close functionality, and implement
56  * read/write.
57  *
58  * We implement three modes of file I/O:
59  *  - buffered uses the generic_file_aio_{read,write} helpers
60  *
61  *  - synchronous is used when there is multi-client read/write
62  *    sharing, avoids the page cache, and synchronously waits for an
63  *    ack from the OSD.
64  *
65  *  - direct io takes the variant of the sync path that references
66  *    user pages directly.
67  *
68  * fsync() flushes and waits on dirty pages, but just queues metadata
69  * for writeback: since the MDS can recover size and mtime there is no
70  * need to wait for MDS acknowledgement.
71  */
72 
73 /*
74  * How many pages to get in one call to iov_iter_get_pages().  This
75  * determines the size of the on-stack array used as a buffer.
76  */
77 #define ITER_GET_BVECS_PAGES	64
78 
79 static ssize_t __iter_get_bvecs(struct iov_iter *iter, size_t maxsize,
80 				struct bio_vec *bvecs)
81 {
82 	size_t size = 0;
83 	int bvec_idx = 0;
84 
85 	if (maxsize > iov_iter_count(iter))
86 		maxsize = iov_iter_count(iter);
87 
88 	while (size < maxsize) {
89 		struct page *pages[ITER_GET_BVECS_PAGES];
90 		ssize_t bytes;
91 		size_t start;
92 		int idx = 0;
93 
94 		bytes = iov_iter_get_pages(iter, pages, maxsize - size,
95 					   ITER_GET_BVECS_PAGES, &start);
96 		if (bytes < 0)
97 			return size ?: bytes;
98 
99 		iov_iter_advance(iter, bytes);
100 		size += bytes;
101 
102 		for ( ; bytes; idx++, bvec_idx++) {
103 			struct bio_vec bv = {
104 				.bv_page = pages[idx],
105 				.bv_len = min_t(int, bytes, PAGE_SIZE - start),
106 				.bv_offset = start,
107 			};
108 
109 			bvecs[bvec_idx] = bv;
110 			bytes -= bv.bv_len;
111 			start = 0;
112 		}
113 	}
114 
115 	return size;
116 }
117 
118 /*
119  * iov_iter_get_pages() only considers one iov_iter segment, no matter
120  * what maxsize or maxpages are given.  For ITER_BVEC that is a single
121  * page.
122  *
123  * Attempt to get up to @maxsize bytes worth of pages from @iter.
124  * Return the number of bytes in the created bio_vec array, or an error.
125  */
126 static ssize_t iter_get_bvecs_alloc(struct iov_iter *iter, size_t maxsize,
127 				    struct bio_vec **bvecs, int *num_bvecs)
128 {
129 	struct bio_vec *bv;
130 	size_t orig_count = iov_iter_count(iter);
131 	ssize_t bytes;
132 	int npages;
133 
134 	iov_iter_truncate(iter, maxsize);
135 	npages = iov_iter_npages(iter, INT_MAX);
136 	iov_iter_reexpand(iter, orig_count);
137 
138 	/*
139 	 * __iter_get_bvecs() may populate only part of the array -- zero it
140 	 * out.
141 	 */
142 	bv = kvmalloc_array(npages, sizeof(*bv), GFP_KERNEL | __GFP_ZERO);
143 	if (!bv)
144 		return -ENOMEM;
145 
146 	bytes = __iter_get_bvecs(iter, maxsize, bv);
147 	if (bytes < 0) {
148 		/*
149 		 * No pages were pinned -- just free the array.
150 		 */
151 		kvfree(bv);
152 		return bytes;
153 	}
154 
155 	*bvecs = bv;
156 	*num_bvecs = npages;
157 	return bytes;
158 }
159 
160 static void put_bvecs(struct bio_vec *bvecs, int num_bvecs, bool should_dirty)
161 {
162 	int i;
163 
164 	for (i = 0; i < num_bvecs; i++) {
165 		if (bvecs[i].bv_page) {
166 			if (should_dirty)
167 				set_page_dirty_lock(bvecs[i].bv_page);
168 			put_page(bvecs[i].bv_page);
169 		}
170 	}
171 	kvfree(bvecs);
172 }
173 
174 /*
175  * Prepare an open request.  Preallocate ceph_cap to avoid an
176  * inopportune ENOMEM later.
177  */
178 static struct ceph_mds_request *
179 prepare_open_request(struct super_block *sb, int flags, int create_mode)
180 {
181 	struct ceph_fs_client *fsc = ceph_sb_to_client(sb);
182 	struct ceph_mds_client *mdsc = fsc->mdsc;
183 	struct ceph_mds_request *req;
184 	int want_auth = USE_ANY_MDS;
185 	int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
186 
187 	if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
188 		want_auth = USE_AUTH_MDS;
189 
190 	req = ceph_mdsc_create_request(mdsc, op, want_auth);
191 	if (IS_ERR(req))
192 		goto out;
193 	req->r_fmode = ceph_flags_to_mode(flags);
194 	req->r_args.open.flags = ceph_flags_sys2wire(flags);
195 	req->r_args.open.mode = cpu_to_le32(create_mode);
196 out:
197 	return req;
198 }
199 
200 static int ceph_init_file_info(struct inode *inode, struct file *file,
201 					int fmode, bool isdir)
202 {
203 	struct ceph_file_info *fi;
204 
205 	dout("%s %p %p 0%o (%s)\n", __func__, inode, file,
206 			inode->i_mode, isdir ? "dir" : "regular");
207 	BUG_ON(inode->i_fop->release != ceph_release);
208 
209 	if (isdir) {
210 		struct ceph_dir_file_info *dfi =
211 			kmem_cache_zalloc(ceph_dir_file_cachep, GFP_KERNEL);
212 		if (!dfi) {
213 			ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
214 			return -ENOMEM;
215 		}
216 
217 		file->private_data = dfi;
218 		fi = &dfi->file_info;
219 		dfi->next_offset = 2;
220 		dfi->readdir_cache_idx = -1;
221 	} else {
222 		fi = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
223 		if (!fi) {
224 			ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
225 			return -ENOMEM;
226 		}
227 
228 		file->private_data = fi;
229 	}
230 
231 	fi->fmode = fmode;
232 	spin_lock_init(&fi->rw_contexts_lock);
233 	INIT_LIST_HEAD(&fi->rw_contexts);
234 
235 	return 0;
236 }
237 
238 /*
239  * initialize private struct file data.
240  * if we fail, clean up by dropping fmode reference on the ceph_inode
241  */
242 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
243 {
244 	int ret = 0;
245 
246 	switch (inode->i_mode & S_IFMT) {
247 	case S_IFREG:
248 		ceph_fscache_register_inode_cookie(inode);
249 		ceph_fscache_file_set_cookie(inode, file);
250 		/* fall through */
251 	case S_IFDIR:
252 		ret = ceph_init_file_info(inode, file, fmode,
253 						S_ISDIR(inode->i_mode));
254 		if (ret)
255 			return ret;
256 		break;
257 
258 	case S_IFLNK:
259 		dout("init_file %p %p 0%o (symlink)\n", inode, file,
260 		     inode->i_mode);
261 		ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
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 		ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
272 		BUG_ON(inode->i_fop->release == ceph_release);
273 
274 		/* call the proper open fop */
275 		ret = inode->i_fop->open(inode, file);
276 	}
277 	return ret;
278 }
279 
280 /*
281  * try renew caps after session gets killed.
282  */
283 int ceph_renew_caps(struct inode *inode)
284 {
285 	struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
286 	struct ceph_inode_info *ci = ceph_inode(inode);
287 	struct ceph_mds_request *req;
288 	int err, flags, wanted;
289 
290 	spin_lock(&ci->i_ceph_lock);
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 	req->r_fmode = -1;
325 
326 	err = ceph_mdsc_do_request(mdsc, NULL, req);
327 	ceph_mdsc_put_request(req);
328 out:
329 	dout("renew caps %p open result=%d\n", inode, err);
330 	return err < 0 ? err : 0;
331 }
332 
333 /*
334  * If we already have the requisite capabilities, we can satisfy
335  * the open request locally (no need to request new caps from the
336  * MDS).  We do, however, need to inform the MDS (asynchronously)
337  * if our wanted caps set expands.
338  */
339 int ceph_open(struct inode *inode, struct file *file)
340 {
341 	struct ceph_inode_info *ci = ceph_inode(inode);
342 	struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
343 	struct ceph_mds_client *mdsc = fsc->mdsc;
344 	struct ceph_mds_request *req;
345 	struct ceph_file_info *fi = file->private_data;
346 	int err;
347 	int flags, fmode, wanted;
348 
349 	if (fi) {
350 		dout("open file %p is already opened\n", file);
351 		return 0;
352 	}
353 
354 	/* filter out O_CREAT|O_EXCL; vfs did that already.  yuck. */
355 	flags = file->f_flags & ~(O_CREAT|O_EXCL);
356 	if (S_ISDIR(inode->i_mode))
357 		flags = O_DIRECTORY;  /* mds likes to know */
358 
359 	dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
360 	     ceph_vinop(inode), file, flags, file->f_flags);
361 	fmode = ceph_flags_to_mode(flags);
362 	wanted = ceph_caps_for_mode(fmode);
363 
364 	/* snapped files are read-only */
365 	if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
366 		return -EROFS;
367 
368 	/* trivially open snapdir */
369 	if (ceph_snap(inode) == CEPH_SNAPDIR) {
370 		spin_lock(&ci->i_ceph_lock);
371 		__ceph_get_fmode(ci, fmode);
372 		spin_unlock(&ci->i_ceph_lock);
373 		return ceph_init_file(inode, file, fmode);
374 	}
375 
376 	/*
377 	 * No need to block if we have caps on the auth MDS (for
378 	 * write) or any MDS (for read).  Update wanted set
379 	 * asynchronously.
380 	 */
381 	spin_lock(&ci->i_ceph_lock);
382 	if (__ceph_is_any_real_caps(ci) &&
383 	    (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
384 		int mds_wanted = __ceph_caps_mds_wanted(ci, true);
385 		int issued = __ceph_caps_issued(ci, NULL);
386 
387 		dout("open %p fmode %d want %s issued %s using existing\n",
388 		     inode, fmode, ceph_cap_string(wanted),
389 		     ceph_cap_string(issued));
390 		__ceph_get_fmode(ci, fmode);
391 		spin_unlock(&ci->i_ceph_lock);
392 
393 		/* adjust wanted? */
394 		if ((issued & wanted) != wanted &&
395 		    (mds_wanted & wanted) != wanted &&
396 		    ceph_snap(inode) != CEPH_SNAPDIR)
397 			ceph_check_caps(ci, 0, NULL);
398 
399 		return ceph_init_file(inode, file, fmode);
400 	} else if (ceph_snap(inode) != CEPH_NOSNAP &&
401 		   (ci->i_snap_caps & wanted) == wanted) {
402 		__ceph_get_fmode(ci, fmode);
403 		spin_unlock(&ci->i_ceph_lock);
404 		return ceph_init_file(inode, file, fmode);
405 	}
406 
407 	spin_unlock(&ci->i_ceph_lock);
408 
409 	dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
410 	req = prepare_open_request(inode->i_sb, flags, 0);
411 	if (IS_ERR(req)) {
412 		err = PTR_ERR(req);
413 		goto out;
414 	}
415 	req->r_inode = inode;
416 	ihold(inode);
417 
418 	req->r_num_caps = 1;
419 	err = ceph_mdsc_do_request(mdsc, NULL, req);
420 	if (!err)
421 		err = ceph_init_file(inode, file, req->r_fmode);
422 	ceph_mdsc_put_request(req);
423 	dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
424 out:
425 	return err;
426 }
427 
428 
429 /*
430  * Do a lookup + open with a single request.  If we get a non-existent
431  * file or symlink, return 1 so the VFS can retry.
432  */
433 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
434 		     struct file *file, unsigned flags, umode_t mode)
435 {
436 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
437 	struct ceph_mds_client *mdsc = fsc->mdsc;
438 	struct ceph_mds_request *req;
439 	struct dentry *dn;
440 	struct ceph_acls_info acls = {};
441 	int mask;
442 	int err;
443 
444 	dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
445 	     dir, dentry, dentry,
446 	     d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
447 
448 	if (dentry->d_name.len > NAME_MAX)
449 		return -ENAMETOOLONG;
450 
451 	if (flags & O_CREAT) {
452 		if (ceph_quota_is_max_files_exceeded(dir))
453 			return -EDQUOT;
454 		err = ceph_pre_init_acls(dir, &mode, &acls);
455 		if (err < 0)
456 			return err;
457 	}
458 
459 	/* do the open */
460 	req = prepare_open_request(dir->i_sb, flags, mode);
461 	if (IS_ERR(req)) {
462 		err = PTR_ERR(req);
463 		goto out_acl;
464 	}
465 	req->r_dentry = dget(dentry);
466 	req->r_num_caps = 2;
467 	if (flags & O_CREAT) {
468 		req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
469 		req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
470 		if (acls.pagelist) {
471 			req->r_pagelist = acls.pagelist;
472 			acls.pagelist = NULL;
473 		}
474 	}
475 
476        mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
477        if (ceph_security_xattr_wanted(dir))
478                mask |= CEPH_CAP_XATTR_SHARED;
479        req->r_args.open.mask = cpu_to_le32(mask);
480 
481 	req->r_parent = dir;
482 	set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
483 	err = ceph_mdsc_do_request(mdsc,
484 				   (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
485 				   req);
486 	err = ceph_handle_snapdir(req, dentry, err);
487 	if (err)
488 		goto out_req;
489 
490 	if ((flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
491 		err = ceph_handle_notrace_create(dir, dentry);
492 
493 	if (d_in_lookup(dentry)) {
494 		dn = ceph_finish_lookup(req, dentry, err);
495 		if (IS_ERR(dn))
496 			err = PTR_ERR(dn);
497 	} else {
498 		/* we were given a hashed negative dentry */
499 		dn = NULL;
500 	}
501 	if (err)
502 		goto out_req;
503 	if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
504 		/* make vfs retry on splice, ENOENT, or symlink */
505 		dout("atomic_open finish_no_open on dn %p\n", dn);
506 		err = finish_no_open(file, dn);
507 	} else {
508 		dout("atomic_open finish_open on dn %p\n", dn);
509 		if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
510 			ceph_init_inode_acls(d_inode(dentry), &acls);
511 			file->f_mode |= FMODE_CREATED;
512 		}
513 		err = finish_open(file, dentry, ceph_open);
514 	}
515 out_req:
516 	if (!req->r_err && req->r_target_inode)
517 		ceph_put_fmode(ceph_inode(req->r_target_inode), req->r_fmode);
518 	ceph_mdsc_put_request(req);
519 out_acl:
520 	ceph_release_acls_info(&acls);
521 	dout("atomic_open result=%d\n", err);
522 	return err;
523 }
524 
525 int ceph_release(struct inode *inode, struct file *file)
526 {
527 	struct ceph_inode_info *ci = ceph_inode(inode);
528 
529 	if (S_ISDIR(inode->i_mode)) {
530 		struct ceph_dir_file_info *dfi = file->private_data;
531 		dout("release inode %p dir file %p\n", inode, file);
532 		WARN_ON(!list_empty(&dfi->file_info.rw_contexts));
533 
534 		ceph_put_fmode(ci, dfi->file_info.fmode);
535 
536 		if (dfi->last_readdir)
537 			ceph_mdsc_put_request(dfi->last_readdir);
538 		kfree(dfi->last_name);
539 		kfree(dfi->dir_info);
540 		kmem_cache_free(ceph_dir_file_cachep, dfi);
541 	} else {
542 		struct ceph_file_info *fi = file->private_data;
543 		dout("release inode %p regular file %p\n", inode, file);
544 		WARN_ON(!list_empty(&fi->rw_contexts));
545 
546 		ceph_put_fmode(ci, fi->fmode);
547 		kmem_cache_free(ceph_file_cachep, fi);
548 	}
549 
550 	/* wake up anyone waiting for caps on this inode */
551 	wake_up_all(&ci->i_cap_wq);
552 	return 0;
553 }
554 
555 enum {
556 	HAVE_RETRIED = 1,
557 	CHECK_EOF =    2,
558 	READ_INLINE =  3,
559 };
560 
561 /*
562  * Completely synchronous read and write methods.  Direct from __user
563  * buffer to osd, or directly to user pages (if O_DIRECT).
564  *
565  * If the read spans object boundary, just do multiple reads.  (That's not
566  * atomic, but good enough for now.)
567  *
568  * If we get a short result from the OSD, check against i_size; we need to
569  * only return a short read to the caller if we hit EOF.
570  */
571 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
572 			      int *retry_op)
573 {
574 	struct file *file = iocb->ki_filp;
575 	struct inode *inode = file_inode(file);
576 	struct ceph_inode_info *ci = ceph_inode(inode);
577 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
578 	struct ceph_osd_client *osdc = &fsc->client->osdc;
579 	ssize_t ret;
580 	u64 off = iocb->ki_pos;
581 	u64 len = iov_iter_count(to);
582 
583 	dout("sync_read on file %p %llu~%u %s\n", file, off, (unsigned)len,
584 	     (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
585 
586 	if (!len)
587 		return 0;
588 	/*
589 	 * flush any page cache pages in this range.  this
590 	 * will make concurrent normal and sync io slow,
591 	 * but it will at least behave sensibly when they are
592 	 * in sequence.
593 	 */
594 	ret = filemap_write_and_wait_range(inode->i_mapping,
595 					   off, off + len - 1);
596 	if (ret < 0)
597 		return ret;
598 
599 	ret = 0;
600 	while ((len = iov_iter_count(to)) > 0) {
601 		struct ceph_osd_request *req;
602 		struct page **pages;
603 		int num_pages;
604 		size_t page_off;
605 		u64 i_size;
606 		bool more;
607 
608 		req = ceph_osdc_new_request(osdc, &ci->i_layout,
609 					ci->i_vino, off, &len, 0, 1,
610 					CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
611 					NULL, ci->i_truncate_seq,
612 					ci->i_truncate_size, false);
613 		if (IS_ERR(req)) {
614 			ret = PTR_ERR(req);
615 			break;
616 		}
617 
618 		more = len < iov_iter_count(to);
619 
620 		if (unlikely(iov_iter_is_pipe(to))) {
621 			ret = iov_iter_get_pages_alloc(to, &pages, len,
622 						       &page_off);
623 			if (ret <= 0) {
624 				ceph_osdc_put_request(req);
625 				ret = -ENOMEM;
626 				break;
627 			}
628 			num_pages = DIV_ROUND_UP(ret + page_off, PAGE_SIZE);
629 			if (ret < len) {
630 				len = ret;
631 				osd_req_op_extent_update(req, 0, len);
632 				more = false;
633 			}
634 		} else {
635 			num_pages = calc_pages_for(off, len);
636 			page_off = off & ~PAGE_MASK;
637 			pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
638 			if (IS_ERR(pages)) {
639 				ceph_osdc_put_request(req);
640 				ret = PTR_ERR(pages);
641 				break;
642 			}
643 		}
644 
645 		osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_off,
646 						 false, false);
647 		ret = ceph_osdc_start_request(osdc, req, false);
648 		if (!ret)
649 			ret = ceph_osdc_wait_request(osdc, req);
650 		ceph_osdc_put_request(req);
651 
652 		i_size = i_size_read(inode);
653 		dout("sync_read %llu~%llu got %zd i_size %llu%s\n",
654 		     off, len, ret, i_size, (more ? " MORE" : ""));
655 
656 		if (ret == -ENOENT)
657 			ret = 0;
658 		if (ret >= 0 && ret < len && (off + ret < i_size)) {
659 			int zlen = min(len - ret, i_size - off - ret);
660 			int zoff = page_off + ret;
661 			dout("sync_read zero gap %llu~%llu\n",
662                              off + ret, off + ret + zlen);
663 			ceph_zero_page_vector_range(zoff, zlen, pages);
664 			ret += zlen;
665 		}
666 
667 		if (unlikely(iov_iter_is_pipe(to))) {
668 			if (ret > 0) {
669 				iov_iter_advance(to, ret);
670 				off += ret;
671 			} else {
672 				iov_iter_advance(to, 0);
673 			}
674 			ceph_put_page_vector(pages, num_pages, false);
675 		} else {
676 			int idx = 0;
677 			size_t left = ret > 0 ? ret : 0;
678 			while (left > 0) {
679 				size_t len, copied;
680 				page_off = off & ~PAGE_MASK;
681 				len = min_t(size_t, left, PAGE_SIZE - page_off);
682 				copied = copy_page_to_iter(pages[idx++],
683 							   page_off, len, to);
684 				off += copied;
685 				left -= copied;
686 				if (copied < len) {
687 					ret = -EFAULT;
688 					break;
689 				}
690 			}
691 			ceph_release_page_vector(pages, num_pages);
692 		}
693 
694 		if (ret <= 0 || off >= i_size || !more)
695 			break;
696 	}
697 
698 	if (off > iocb->ki_pos) {
699 		if (ret >= 0 &&
700 		    iov_iter_count(to) > 0 && off >= i_size_read(inode))
701 			*retry_op = CHECK_EOF;
702 		ret = off - iocb->ki_pos;
703 		iocb->ki_pos = off;
704 	}
705 
706 	dout("sync_read result %zd retry_op %d\n", ret, *retry_op);
707 	return ret;
708 }
709 
710 struct ceph_aio_request {
711 	struct kiocb *iocb;
712 	size_t total_len;
713 	bool write;
714 	bool should_dirty;
715 	int error;
716 	struct list_head osd_reqs;
717 	unsigned num_reqs;
718 	atomic_t pending_reqs;
719 	struct timespec64 mtime;
720 	struct ceph_cap_flush *prealloc_cf;
721 };
722 
723 struct ceph_aio_work {
724 	struct work_struct work;
725 	struct ceph_osd_request *req;
726 };
727 
728 static void ceph_aio_retry_work(struct work_struct *work);
729 
730 static void ceph_aio_complete(struct inode *inode,
731 			      struct ceph_aio_request *aio_req)
732 {
733 	struct ceph_inode_info *ci = ceph_inode(inode);
734 	int ret;
735 
736 	if (!atomic_dec_and_test(&aio_req->pending_reqs))
737 		return;
738 
739 	ret = aio_req->error;
740 	if (!ret)
741 		ret = aio_req->total_len;
742 
743 	dout("ceph_aio_complete %p rc %d\n", inode, ret);
744 
745 	if (ret >= 0 && aio_req->write) {
746 		int dirty;
747 
748 		loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
749 		if (endoff > i_size_read(inode)) {
750 			if (ceph_inode_set_size(inode, endoff))
751 				ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
752 		}
753 
754 		spin_lock(&ci->i_ceph_lock);
755 		ci->i_inline_version = CEPH_INLINE_NONE;
756 		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
757 					       &aio_req->prealloc_cf);
758 		spin_unlock(&ci->i_ceph_lock);
759 		if (dirty)
760 			__mark_inode_dirty(inode, dirty);
761 
762 	}
763 
764 	ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
765 						CEPH_CAP_FILE_RD));
766 
767 	aio_req->iocb->ki_complete(aio_req->iocb, ret, 0);
768 
769 	ceph_free_cap_flush(aio_req->prealloc_cf);
770 	kfree(aio_req);
771 }
772 
773 static void ceph_aio_complete_req(struct ceph_osd_request *req)
774 {
775 	int rc = req->r_result;
776 	struct inode *inode = req->r_inode;
777 	struct ceph_aio_request *aio_req = req->r_priv;
778 	struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
779 
780 	BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_BVECS);
781 	BUG_ON(!osd_data->num_bvecs);
782 
783 	dout("ceph_aio_complete_req %p rc %d bytes %u\n",
784 	     inode, rc, osd_data->bvec_pos.iter.bi_size);
785 
786 	if (rc == -EOLDSNAPC) {
787 		struct ceph_aio_work *aio_work;
788 		BUG_ON(!aio_req->write);
789 
790 		aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
791 		if (aio_work) {
792 			INIT_WORK(&aio_work->work, ceph_aio_retry_work);
793 			aio_work->req = req;
794 			queue_work(ceph_inode_to_client(inode)->wb_wq,
795 				   &aio_work->work);
796 			return;
797 		}
798 		rc = -ENOMEM;
799 	} else if (!aio_req->write) {
800 		if (rc == -ENOENT)
801 			rc = 0;
802 		if (rc >= 0 && osd_data->bvec_pos.iter.bi_size > rc) {
803 			struct iov_iter i;
804 			int zlen = osd_data->bvec_pos.iter.bi_size - rc;
805 
806 			/*
807 			 * If read is satisfied by single OSD request,
808 			 * it can pass EOF. Otherwise read is within
809 			 * i_size.
810 			 */
811 			if (aio_req->num_reqs == 1) {
812 				loff_t i_size = i_size_read(inode);
813 				loff_t endoff = aio_req->iocb->ki_pos + rc;
814 				if (endoff < i_size)
815 					zlen = min_t(size_t, zlen,
816 						     i_size - endoff);
817 				aio_req->total_len = rc + zlen;
818 			}
819 
820 			iov_iter_bvec(&i, READ, osd_data->bvec_pos.bvecs,
821 				      osd_data->num_bvecs,
822 				      osd_data->bvec_pos.iter.bi_size);
823 			iov_iter_advance(&i, rc);
824 			iov_iter_zero(zlen, &i);
825 		}
826 	}
827 
828 	put_bvecs(osd_data->bvec_pos.bvecs, osd_data->num_bvecs,
829 		  aio_req->should_dirty);
830 	ceph_osdc_put_request(req);
831 
832 	if (rc < 0)
833 		cmpxchg(&aio_req->error, 0, rc);
834 
835 	ceph_aio_complete(inode, aio_req);
836 	return;
837 }
838 
839 static void ceph_aio_retry_work(struct work_struct *work)
840 {
841 	struct ceph_aio_work *aio_work =
842 		container_of(work, struct ceph_aio_work, work);
843 	struct ceph_osd_request *orig_req = aio_work->req;
844 	struct ceph_aio_request *aio_req = orig_req->r_priv;
845 	struct inode *inode = orig_req->r_inode;
846 	struct ceph_inode_info *ci = ceph_inode(inode);
847 	struct ceph_snap_context *snapc;
848 	struct ceph_osd_request *req;
849 	int ret;
850 
851 	spin_lock(&ci->i_ceph_lock);
852 	if (__ceph_have_pending_cap_snap(ci)) {
853 		struct ceph_cap_snap *capsnap =
854 			list_last_entry(&ci->i_cap_snaps,
855 					struct ceph_cap_snap,
856 					ci_item);
857 		snapc = ceph_get_snap_context(capsnap->context);
858 	} else {
859 		BUG_ON(!ci->i_head_snapc);
860 		snapc = ceph_get_snap_context(ci->i_head_snapc);
861 	}
862 	spin_unlock(&ci->i_ceph_lock);
863 
864 	req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 1,
865 			false, GFP_NOFS);
866 	if (!req) {
867 		ret = -ENOMEM;
868 		req = orig_req;
869 		goto out;
870 	}
871 
872 	req->r_flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
873 	ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
874 	ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
875 
876 	req->r_ops[0] = orig_req->r_ops[0];
877 
878 	req->r_mtime = aio_req->mtime;
879 	req->r_data_offset = req->r_ops[0].extent.offset;
880 
881 	ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
882 	if (ret) {
883 		ceph_osdc_put_request(req);
884 		req = orig_req;
885 		goto out;
886 	}
887 
888 	ceph_osdc_put_request(orig_req);
889 
890 	req->r_callback = ceph_aio_complete_req;
891 	req->r_inode = inode;
892 	req->r_priv = aio_req;
893 
894 	ret = ceph_osdc_start_request(req->r_osdc, req, false);
895 out:
896 	if (ret < 0) {
897 		req->r_result = ret;
898 		ceph_aio_complete_req(req);
899 	}
900 
901 	ceph_put_snap_context(snapc);
902 	kfree(aio_work);
903 }
904 
905 static ssize_t
906 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
907 		       struct ceph_snap_context *snapc,
908 		       struct ceph_cap_flush **pcf)
909 {
910 	struct file *file = iocb->ki_filp;
911 	struct inode *inode = file_inode(file);
912 	struct ceph_inode_info *ci = ceph_inode(inode);
913 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
914 	struct ceph_vino vino;
915 	struct ceph_osd_request *req;
916 	struct bio_vec *bvecs;
917 	struct ceph_aio_request *aio_req = NULL;
918 	int num_pages = 0;
919 	int flags;
920 	int ret;
921 	struct timespec64 mtime = current_time(inode);
922 	size_t count = iov_iter_count(iter);
923 	loff_t pos = iocb->ki_pos;
924 	bool write = iov_iter_rw(iter) == WRITE;
925 	bool should_dirty = !write && iter_is_iovec(iter);
926 
927 	if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
928 		return -EROFS;
929 
930 	dout("sync_direct_%s on file %p %lld~%u snapc %p seq %lld\n",
931 	     (write ? "write" : "read"), file, pos, (unsigned)count,
932 	     snapc, snapc ? snapc->seq : 0);
933 
934 	ret = filemap_write_and_wait_range(inode->i_mapping,
935 					   pos, pos + count - 1);
936 	if (ret < 0)
937 		return ret;
938 
939 	if (write) {
940 		int ret2 = invalidate_inode_pages2_range(inode->i_mapping,
941 					pos >> PAGE_SHIFT,
942 					(pos + count - 1) >> PAGE_SHIFT);
943 		if (ret2 < 0)
944 			dout("invalidate_inode_pages2_range returned %d\n", ret2);
945 
946 		flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
947 	} else {
948 		flags = CEPH_OSD_FLAG_READ;
949 	}
950 
951 	while (iov_iter_count(iter) > 0) {
952 		u64 size = iov_iter_count(iter);
953 		ssize_t len;
954 
955 		if (write)
956 			size = min_t(u64, size, fsc->mount_options->wsize);
957 		else
958 			size = min_t(u64, size, fsc->mount_options->rsize);
959 
960 		vino = ceph_vino(inode);
961 		req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
962 					    vino, pos, &size, 0,
963 					    1,
964 					    write ? CEPH_OSD_OP_WRITE :
965 						    CEPH_OSD_OP_READ,
966 					    flags, snapc,
967 					    ci->i_truncate_seq,
968 					    ci->i_truncate_size,
969 					    false);
970 		if (IS_ERR(req)) {
971 			ret = PTR_ERR(req);
972 			break;
973 		}
974 
975 		len = iter_get_bvecs_alloc(iter, size, &bvecs, &num_pages);
976 		if (len < 0) {
977 			ceph_osdc_put_request(req);
978 			ret = len;
979 			break;
980 		}
981 		if (len != size)
982 			osd_req_op_extent_update(req, 0, len);
983 
984 		/*
985 		 * To simplify error handling, allow AIO when IO within i_size
986 		 * or IO can be satisfied by single OSD request.
987 		 */
988 		if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
989 		    (len == count || pos + count <= i_size_read(inode))) {
990 			aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
991 			if (aio_req) {
992 				aio_req->iocb = iocb;
993 				aio_req->write = write;
994 				aio_req->should_dirty = should_dirty;
995 				INIT_LIST_HEAD(&aio_req->osd_reqs);
996 				if (write) {
997 					aio_req->mtime = mtime;
998 					swap(aio_req->prealloc_cf, *pcf);
999 				}
1000 			}
1001 			/* ignore error */
1002 		}
1003 
1004 		if (write) {
1005 			/*
1006 			 * throw out any page cache pages in this range. this
1007 			 * may block.
1008 			 */
1009 			truncate_inode_pages_range(inode->i_mapping, pos,
1010 					(pos+len) | (PAGE_SIZE - 1));
1011 
1012 			req->r_mtime = mtime;
1013 		}
1014 
1015 		osd_req_op_extent_osd_data_bvecs(req, 0, bvecs, num_pages, len);
1016 
1017 		if (aio_req) {
1018 			aio_req->total_len += len;
1019 			aio_req->num_reqs++;
1020 			atomic_inc(&aio_req->pending_reqs);
1021 
1022 			req->r_callback = ceph_aio_complete_req;
1023 			req->r_inode = inode;
1024 			req->r_priv = aio_req;
1025 			list_add_tail(&req->r_unsafe_item, &aio_req->osd_reqs);
1026 
1027 			pos += len;
1028 			continue;
1029 		}
1030 
1031 		ret = ceph_osdc_start_request(req->r_osdc, req, false);
1032 		if (!ret)
1033 			ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1034 
1035 		size = i_size_read(inode);
1036 		if (!write) {
1037 			if (ret == -ENOENT)
1038 				ret = 0;
1039 			if (ret >= 0 && ret < len && pos + ret < size) {
1040 				struct iov_iter i;
1041 				int zlen = min_t(size_t, len - ret,
1042 						 size - pos - ret);
1043 
1044 				iov_iter_bvec(&i, READ, bvecs, num_pages, len);
1045 				iov_iter_advance(&i, ret);
1046 				iov_iter_zero(zlen, &i);
1047 				ret += zlen;
1048 			}
1049 			if (ret >= 0)
1050 				len = ret;
1051 		}
1052 
1053 		put_bvecs(bvecs, num_pages, should_dirty);
1054 		ceph_osdc_put_request(req);
1055 		if (ret < 0)
1056 			break;
1057 
1058 		pos += len;
1059 		if (!write && pos >= size)
1060 			break;
1061 
1062 		if (write && pos > size) {
1063 			if (ceph_inode_set_size(inode, pos))
1064 				ceph_check_caps(ceph_inode(inode),
1065 						CHECK_CAPS_AUTHONLY,
1066 						NULL);
1067 		}
1068 	}
1069 
1070 	if (aio_req) {
1071 		LIST_HEAD(osd_reqs);
1072 
1073 		if (aio_req->num_reqs == 0) {
1074 			kfree(aio_req);
1075 			return ret;
1076 		}
1077 
1078 		ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1079 					      CEPH_CAP_FILE_RD);
1080 
1081 		list_splice(&aio_req->osd_reqs, &osd_reqs);
1082 		while (!list_empty(&osd_reqs)) {
1083 			req = list_first_entry(&osd_reqs,
1084 					       struct ceph_osd_request,
1085 					       r_unsafe_item);
1086 			list_del_init(&req->r_unsafe_item);
1087 			if (ret >= 0)
1088 				ret = ceph_osdc_start_request(req->r_osdc,
1089 							      req, false);
1090 			if (ret < 0) {
1091 				req->r_result = ret;
1092 				ceph_aio_complete_req(req);
1093 			}
1094 		}
1095 		return -EIOCBQUEUED;
1096 	}
1097 
1098 	if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1099 		ret = pos - iocb->ki_pos;
1100 		iocb->ki_pos = pos;
1101 	}
1102 	return ret;
1103 }
1104 
1105 /*
1106  * Synchronous write, straight from __user pointer or user pages.
1107  *
1108  * If write spans object boundary, just do multiple writes.  (For a
1109  * correct atomic write, we should e.g. take write locks on all
1110  * objects, rollback on failure, etc.)
1111  */
1112 static ssize_t
1113 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1114 		struct ceph_snap_context *snapc)
1115 {
1116 	struct file *file = iocb->ki_filp;
1117 	struct inode *inode = file_inode(file);
1118 	struct ceph_inode_info *ci = ceph_inode(inode);
1119 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1120 	struct ceph_vino vino;
1121 	struct ceph_osd_request *req;
1122 	struct page **pages;
1123 	u64 len;
1124 	int num_pages;
1125 	int written = 0;
1126 	int flags;
1127 	int ret;
1128 	bool check_caps = false;
1129 	struct timespec64 mtime = current_time(inode);
1130 	size_t count = iov_iter_count(from);
1131 
1132 	if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1133 		return -EROFS;
1134 
1135 	dout("sync_write on file %p %lld~%u snapc %p seq %lld\n",
1136 	     file, pos, (unsigned)count, snapc, snapc->seq);
1137 
1138 	ret = filemap_write_and_wait_range(inode->i_mapping,
1139 					   pos, pos + count - 1);
1140 	if (ret < 0)
1141 		return ret;
1142 
1143 	ret = invalidate_inode_pages2_range(inode->i_mapping,
1144 					    pos >> PAGE_SHIFT,
1145 					    (pos + count - 1) >> PAGE_SHIFT);
1146 	if (ret < 0)
1147 		dout("invalidate_inode_pages2_range returned %d\n", ret);
1148 
1149 	flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1150 
1151 	while ((len = iov_iter_count(from)) > 0) {
1152 		size_t left;
1153 		int n;
1154 
1155 		vino = ceph_vino(inode);
1156 		req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1157 					    vino, pos, &len, 0, 1,
1158 					    CEPH_OSD_OP_WRITE, flags, snapc,
1159 					    ci->i_truncate_seq,
1160 					    ci->i_truncate_size,
1161 					    false);
1162 		if (IS_ERR(req)) {
1163 			ret = PTR_ERR(req);
1164 			break;
1165 		}
1166 
1167 		/*
1168 		 * write from beginning of first page,
1169 		 * regardless of io alignment
1170 		 */
1171 		num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1172 
1173 		pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1174 		if (IS_ERR(pages)) {
1175 			ret = PTR_ERR(pages);
1176 			goto out;
1177 		}
1178 
1179 		left = len;
1180 		for (n = 0; n < num_pages; n++) {
1181 			size_t plen = min_t(size_t, left, PAGE_SIZE);
1182 			ret = copy_page_from_iter(pages[n], 0, plen, from);
1183 			if (ret != plen) {
1184 				ret = -EFAULT;
1185 				break;
1186 			}
1187 			left -= ret;
1188 		}
1189 
1190 		if (ret < 0) {
1191 			ceph_release_page_vector(pages, num_pages);
1192 			goto out;
1193 		}
1194 
1195 		req->r_inode = inode;
1196 
1197 		osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1198 						false, true);
1199 
1200 		req->r_mtime = mtime;
1201 		ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1202 		if (!ret)
1203 			ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1204 
1205 out:
1206 		ceph_osdc_put_request(req);
1207 		if (ret != 0) {
1208 			ceph_set_error_write(ci);
1209 			break;
1210 		}
1211 
1212 		ceph_clear_error_write(ci);
1213 		pos += len;
1214 		written += len;
1215 		if (pos > i_size_read(inode)) {
1216 			check_caps = ceph_inode_set_size(inode, pos);
1217 			if (check_caps)
1218 				ceph_check_caps(ceph_inode(inode),
1219 						CHECK_CAPS_AUTHONLY,
1220 						NULL);
1221 		}
1222 
1223 	}
1224 
1225 	if (ret != -EOLDSNAPC && written > 0) {
1226 		ret = written;
1227 		iocb->ki_pos = pos;
1228 	}
1229 	return ret;
1230 }
1231 
1232 /*
1233  * Wrap generic_file_aio_read with checks for cap bits on the inode.
1234  * Atomically grab references, so that those bits are not released
1235  * back to the MDS mid-read.
1236  *
1237  * Hmm, the sync read case isn't actually async... should it be?
1238  */
1239 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1240 {
1241 	struct file *filp = iocb->ki_filp;
1242 	struct ceph_file_info *fi = filp->private_data;
1243 	size_t len = iov_iter_count(to);
1244 	struct inode *inode = file_inode(filp);
1245 	struct ceph_inode_info *ci = ceph_inode(inode);
1246 	struct page *pinned_page = NULL;
1247 	ssize_t ret;
1248 	int want, got = 0;
1249 	int retry_op = 0, read = 0;
1250 
1251 again:
1252 	dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1253 	     inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1254 
1255 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
1256 		want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1257 	else
1258 		want = CEPH_CAP_FILE_CACHE;
1259 	ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1260 	if (ret < 0)
1261 		return ret;
1262 
1263 	if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1264 	    (iocb->ki_flags & IOCB_DIRECT) ||
1265 	    (fi->flags & CEPH_F_SYNC)) {
1266 
1267 		dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1268 		     inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1269 		     ceph_cap_string(got));
1270 
1271 		if (ci->i_inline_version == CEPH_INLINE_NONE) {
1272 			if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1273 				ret = ceph_direct_read_write(iocb, to,
1274 							     NULL, NULL);
1275 				if (ret >= 0 && ret < len)
1276 					retry_op = CHECK_EOF;
1277 			} else {
1278 				ret = ceph_sync_read(iocb, to, &retry_op);
1279 			}
1280 		} else {
1281 			retry_op = READ_INLINE;
1282 		}
1283 	} else {
1284 		CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1285 		dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1286 		     inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1287 		     ceph_cap_string(got));
1288 		ceph_add_rw_context(fi, &rw_ctx);
1289 		ret = generic_file_read_iter(iocb, to);
1290 		ceph_del_rw_context(fi, &rw_ctx);
1291 	}
1292 	dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1293 	     inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1294 	if (pinned_page) {
1295 		put_page(pinned_page);
1296 		pinned_page = NULL;
1297 	}
1298 	ceph_put_cap_refs(ci, got);
1299 	if (retry_op > HAVE_RETRIED && ret >= 0) {
1300 		int statret;
1301 		struct page *page = NULL;
1302 		loff_t i_size;
1303 		if (retry_op == READ_INLINE) {
1304 			page = __page_cache_alloc(GFP_KERNEL);
1305 			if (!page)
1306 				return -ENOMEM;
1307 		}
1308 
1309 		statret = __ceph_do_getattr(inode, page,
1310 					    CEPH_STAT_CAP_INLINE_DATA, !!page);
1311 		if (statret < 0) {
1312 			if (page)
1313 				__free_page(page);
1314 			if (statret == -ENODATA) {
1315 				BUG_ON(retry_op != READ_INLINE);
1316 				goto again;
1317 			}
1318 			return statret;
1319 		}
1320 
1321 		i_size = i_size_read(inode);
1322 		if (retry_op == READ_INLINE) {
1323 			BUG_ON(ret > 0 || read > 0);
1324 			if (iocb->ki_pos < i_size &&
1325 			    iocb->ki_pos < PAGE_SIZE) {
1326 				loff_t end = min_t(loff_t, i_size,
1327 						   iocb->ki_pos + len);
1328 				end = min_t(loff_t, end, PAGE_SIZE);
1329 				if (statret < end)
1330 					zero_user_segment(page, statret, end);
1331 				ret = copy_page_to_iter(page,
1332 						iocb->ki_pos & ~PAGE_MASK,
1333 						end - iocb->ki_pos, to);
1334 				iocb->ki_pos += ret;
1335 				read += ret;
1336 			}
1337 			if (iocb->ki_pos < i_size && read < len) {
1338 				size_t zlen = min_t(size_t, len - read,
1339 						    i_size - iocb->ki_pos);
1340 				ret = iov_iter_zero(zlen, to);
1341 				iocb->ki_pos += ret;
1342 				read += ret;
1343 			}
1344 			__free_pages(page, 0);
1345 			return read;
1346 		}
1347 
1348 		/* hit EOF or hole? */
1349 		if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1350 		    ret < len) {
1351 			dout("sync_read hit hole, ppos %lld < size %lld"
1352 			     ", reading more\n", iocb->ki_pos, i_size);
1353 
1354 			read += ret;
1355 			len -= ret;
1356 			retry_op = HAVE_RETRIED;
1357 			goto again;
1358 		}
1359 	}
1360 
1361 	if (ret >= 0)
1362 		ret += read;
1363 
1364 	return ret;
1365 }
1366 
1367 /*
1368  * Take cap references to avoid releasing caps to MDS mid-write.
1369  *
1370  * If we are synchronous, and write with an old snap context, the OSD
1371  * may return EOLDSNAPC.  In that case, retry the write.. _after_
1372  * dropping our cap refs and allowing the pending snap to logically
1373  * complete _before_ this write occurs.
1374  *
1375  * If we are near ENOSPC, write synchronously.
1376  */
1377 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1378 {
1379 	struct file *file = iocb->ki_filp;
1380 	struct ceph_file_info *fi = file->private_data;
1381 	struct inode *inode = file_inode(file);
1382 	struct ceph_inode_info *ci = ceph_inode(inode);
1383 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1384 	struct ceph_cap_flush *prealloc_cf;
1385 	ssize_t count, written = 0;
1386 	int err, want, got;
1387 	loff_t pos;
1388 	loff_t limit = max(i_size_read(inode), fsc->max_file_size);
1389 
1390 	if (ceph_snap(inode) != CEPH_NOSNAP)
1391 		return -EROFS;
1392 
1393 	prealloc_cf = ceph_alloc_cap_flush();
1394 	if (!prealloc_cf)
1395 		return -ENOMEM;
1396 
1397 retry_snap:
1398 	inode_lock(inode);
1399 
1400 	/* We can write back this queue in page reclaim */
1401 	current->backing_dev_info = inode_to_bdi(inode);
1402 
1403 	if (iocb->ki_flags & IOCB_APPEND) {
1404 		err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1405 		if (err < 0)
1406 			goto out;
1407 	}
1408 
1409 	err = generic_write_checks(iocb, from);
1410 	if (err <= 0)
1411 		goto out;
1412 
1413 	pos = iocb->ki_pos;
1414 	if (unlikely(pos >= limit)) {
1415 		err = -EFBIG;
1416 		goto out;
1417 	} else {
1418 		iov_iter_truncate(from, limit - pos);
1419 	}
1420 
1421 	count = iov_iter_count(from);
1422 	if (ceph_quota_is_max_bytes_exceeded(inode, pos + count)) {
1423 		err = -EDQUOT;
1424 		goto out;
1425 	}
1426 
1427 	err = file_remove_privs(file);
1428 	if (err)
1429 		goto out;
1430 
1431 	err = file_update_time(file);
1432 	if (err)
1433 		goto out;
1434 
1435 	if (ci->i_inline_version != CEPH_INLINE_NONE) {
1436 		err = ceph_uninline_data(file, NULL);
1437 		if (err < 0)
1438 			goto out;
1439 	}
1440 
1441 	/* FIXME: not complete since it doesn't account for being at quota */
1442 	if (ceph_osdmap_flag(&fsc->client->osdc, CEPH_OSDMAP_FULL)) {
1443 		err = -ENOSPC;
1444 		goto out;
1445 	}
1446 
1447 	dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1448 	     inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1449 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
1450 		want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1451 	else
1452 		want = CEPH_CAP_FILE_BUFFER;
1453 	got = 0;
1454 	err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, pos + count,
1455 			    &got, NULL);
1456 	if (err < 0)
1457 		goto out;
1458 
1459 	dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1460 	     inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1461 
1462 	if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1463 	    (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
1464 	    (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
1465 		struct ceph_snap_context *snapc;
1466 		struct iov_iter data;
1467 		inode_unlock(inode);
1468 
1469 		spin_lock(&ci->i_ceph_lock);
1470 		if (__ceph_have_pending_cap_snap(ci)) {
1471 			struct ceph_cap_snap *capsnap =
1472 					list_last_entry(&ci->i_cap_snaps,
1473 							struct ceph_cap_snap,
1474 							ci_item);
1475 			snapc = ceph_get_snap_context(capsnap->context);
1476 		} else {
1477 			BUG_ON(!ci->i_head_snapc);
1478 			snapc = ceph_get_snap_context(ci->i_head_snapc);
1479 		}
1480 		spin_unlock(&ci->i_ceph_lock);
1481 
1482 		/* we might need to revert back to that point */
1483 		data = *from;
1484 		if (iocb->ki_flags & IOCB_DIRECT)
1485 			written = ceph_direct_read_write(iocb, &data, snapc,
1486 							 &prealloc_cf);
1487 		else
1488 			written = ceph_sync_write(iocb, &data, pos, snapc);
1489 		if (written > 0)
1490 			iov_iter_advance(from, written);
1491 		ceph_put_snap_context(snapc);
1492 	} else {
1493 		/*
1494 		 * No need to acquire the i_truncate_mutex. Because
1495 		 * the MDS revokes Fwb caps before sending truncate
1496 		 * message to us. We can't get Fwb cap while there
1497 		 * are pending vmtruncate. So write and vmtruncate
1498 		 * can not run at the same time
1499 		 */
1500 		written = generic_perform_write(file, from, pos);
1501 		if (likely(written >= 0))
1502 			iocb->ki_pos = pos + written;
1503 		inode_unlock(inode);
1504 	}
1505 
1506 	if (written >= 0) {
1507 		int dirty;
1508 
1509 		spin_lock(&ci->i_ceph_lock);
1510 		ci->i_inline_version = CEPH_INLINE_NONE;
1511 		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1512 					       &prealloc_cf);
1513 		spin_unlock(&ci->i_ceph_lock);
1514 		if (dirty)
1515 			__mark_inode_dirty(inode, dirty);
1516 		if (ceph_quota_is_max_bytes_approaching(inode, iocb->ki_pos))
1517 			ceph_check_caps(ci, CHECK_CAPS_NODELAY, NULL);
1518 	}
1519 
1520 	dout("aio_write %p %llx.%llx %llu~%u  dropping cap refs on %s\n",
1521 	     inode, ceph_vinop(inode), pos, (unsigned)count,
1522 	     ceph_cap_string(got));
1523 	ceph_put_cap_refs(ci, got);
1524 
1525 	if (written == -EOLDSNAPC) {
1526 		dout("aio_write %p %llx.%llx %llu~%u" "got EOLDSNAPC, retrying\n",
1527 		     inode, ceph_vinop(inode), pos, (unsigned)count);
1528 		goto retry_snap;
1529 	}
1530 
1531 	if (written >= 0) {
1532 		if (ceph_osdmap_flag(&fsc->client->osdc, CEPH_OSDMAP_NEARFULL))
1533 			iocb->ki_flags |= IOCB_DSYNC;
1534 		written = generic_write_sync(iocb, written);
1535 	}
1536 
1537 	goto out_unlocked;
1538 
1539 out:
1540 	inode_unlock(inode);
1541 out_unlocked:
1542 	ceph_free_cap_flush(prealloc_cf);
1543 	current->backing_dev_info = NULL;
1544 	return written ? written : err;
1545 }
1546 
1547 /*
1548  * llseek.  be sure to verify file size on SEEK_END.
1549  */
1550 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1551 {
1552 	struct inode *inode = file->f_mapping->host;
1553 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1554 	loff_t i_size;
1555 	loff_t ret;
1556 
1557 	inode_lock(inode);
1558 
1559 	if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1560 		ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1561 		if (ret < 0)
1562 			goto out;
1563 	}
1564 
1565 	i_size = i_size_read(inode);
1566 	switch (whence) {
1567 	case SEEK_END:
1568 		offset += i_size;
1569 		break;
1570 	case SEEK_CUR:
1571 		/*
1572 		 * Here we special-case the lseek(fd, 0, SEEK_CUR)
1573 		 * position-querying operation.  Avoid rewriting the "same"
1574 		 * f_pos value back to the file because a concurrent read(),
1575 		 * write() or lseek() might have altered it
1576 		 */
1577 		if (offset == 0) {
1578 			ret = file->f_pos;
1579 			goto out;
1580 		}
1581 		offset += file->f_pos;
1582 		break;
1583 	case SEEK_DATA:
1584 		if (offset < 0 || offset >= i_size) {
1585 			ret = -ENXIO;
1586 			goto out;
1587 		}
1588 		break;
1589 	case SEEK_HOLE:
1590 		if (offset < 0 || offset >= i_size) {
1591 			ret = -ENXIO;
1592 			goto out;
1593 		}
1594 		offset = i_size;
1595 		break;
1596 	}
1597 
1598 	ret = vfs_setpos(file, offset, max(i_size, fsc->max_file_size));
1599 
1600 out:
1601 	inode_unlock(inode);
1602 	return ret;
1603 }
1604 
1605 static inline void ceph_zero_partial_page(
1606 	struct inode *inode, loff_t offset, unsigned size)
1607 {
1608 	struct page *page;
1609 	pgoff_t index = offset >> PAGE_SHIFT;
1610 
1611 	page = find_lock_page(inode->i_mapping, index);
1612 	if (page) {
1613 		wait_on_page_writeback(page);
1614 		zero_user(page, offset & (PAGE_SIZE - 1), size);
1615 		unlock_page(page);
1616 		put_page(page);
1617 	}
1618 }
1619 
1620 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1621 				      loff_t length)
1622 {
1623 	loff_t nearly = round_up(offset, PAGE_SIZE);
1624 	if (offset < nearly) {
1625 		loff_t size = nearly - offset;
1626 		if (length < size)
1627 			size = length;
1628 		ceph_zero_partial_page(inode, offset, size);
1629 		offset += size;
1630 		length -= size;
1631 	}
1632 	if (length >= PAGE_SIZE) {
1633 		loff_t size = round_down(length, PAGE_SIZE);
1634 		truncate_pagecache_range(inode, offset, offset + size - 1);
1635 		offset += size;
1636 		length -= size;
1637 	}
1638 	if (length)
1639 		ceph_zero_partial_page(inode, offset, length);
1640 }
1641 
1642 static int ceph_zero_partial_object(struct inode *inode,
1643 				    loff_t offset, loff_t *length)
1644 {
1645 	struct ceph_inode_info *ci = ceph_inode(inode);
1646 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1647 	struct ceph_osd_request *req;
1648 	int ret = 0;
1649 	loff_t zero = 0;
1650 	int op;
1651 
1652 	if (!length) {
1653 		op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1654 		length = &zero;
1655 	} else {
1656 		op = CEPH_OSD_OP_ZERO;
1657 	}
1658 
1659 	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1660 					ceph_vino(inode),
1661 					offset, length,
1662 					0, 1, op,
1663 					CEPH_OSD_FLAG_WRITE,
1664 					NULL, 0, 0, false);
1665 	if (IS_ERR(req)) {
1666 		ret = PTR_ERR(req);
1667 		goto out;
1668 	}
1669 
1670 	req->r_mtime = inode->i_mtime;
1671 	ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1672 	if (!ret) {
1673 		ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1674 		if (ret == -ENOENT)
1675 			ret = 0;
1676 	}
1677 	ceph_osdc_put_request(req);
1678 
1679 out:
1680 	return ret;
1681 }
1682 
1683 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
1684 {
1685 	int ret = 0;
1686 	struct ceph_inode_info *ci = ceph_inode(inode);
1687 	s32 stripe_unit = ci->i_layout.stripe_unit;
1688 	s32 stripe_count = ci->i_layout.stripe_count;
1689 	s32 object_size = ci->i_layout.object_size;
1690 	u64 object_set_size = object_size * stripe_count;
1691 	u64 nearly, t;
1692 
1693 	/* round offset up to next period boundary */
1694 	nearly = offset + object_set_size - 1;
1695 	t = nearly;
1696 	nearly -= do_div(t, object_set_size);
1697 
1698 	while (length && offset < nearly) {
1699 		loff_t size = length;
1700 		ret = ceph_zero_partial_object(inode, offset, &size);
1701 		if (ret < 0)
1702 			return ret;
1703 		offset += size;
1704 		length -= size;
1705 	}
1706 	while (length >= object_set_size) {
1707 		int i;
1708 		loff_t pos = offset;
1709 		for (i = 0; i < stripe_count; ++i) {
1710 			ret = ceph_zero_partial_object(inode, pos, NULL);
1711 			if (ret < 0)
1712 				return ret;
1713 			pos += stripe_unit;
1714 		}
1715 		offset += object_set_size;
1716 		length -= object_set_size;
1717 	}
1718 	while (length) {
1719 		loff_t size = length;
1720 		ret = ceph_zero_partial_object(inode, offset, &size);
1721 		if (ret < 0)
1722 			return ret;
1723 		offset += size;
1724 		length -= size;
1725 	}
1726 	return ret;
1727 }
1728 
1729 static long ceph_fallocate(struct file *file, int mode,
1730 				loff_t offset, loff_t length)
1731 {
1732 	struct ceph_file_info *fi = file->private_data;
1733 	struct inode *inode = file_inode(file);
1734 	struct ceph_inode_info *ci = ceph_inode(inode);
1735 	struct ceph_cap_flush *prealloc_cf;
1736 	int want, got = 0;
1737 	int dirty;
1738 	int ret = 0;
1739 	loff_t endoff = 0;
1740 	loff_t size;
1741 
1742 	if (mode != (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1743 		return -EOPNOTSUPP;
1744 
1745 	if (!S_ISREG(inode->i_mode))
1746 		return -EOPNOTSUPP;
1747 
1748 	prealloc_cf = ceph_alloc_cap_flush();
1749 	if (!prealloc_cf)
1750 		return -ENOMEM;
1751 
1752 	inode_lock(inode);
1753 
1754 	if (ceph_snap(inode) != CEPH_NOSNAP) {
1755 		ret = -EROFS;
1756 		goto unlock;
1757 	}
1758 
1759 	if (ci->i_inline_version != CEPH_INLINE_NONE) {
1760 		ret = ceph_uninline_data(file, NULL);
1761 		if (ret < 0)
1762 			goto unlock;
1763 	}
1764 
1765 	size = i_size_read(inode);
1766 
1767 	/* Are we punching a hole beyond EOF? */
1768 	if (offset >= size)
1769 		goto unlock;
1770 	if ((offset + length) > size)
1771 		length = size - offset;
1772 
1773 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
1774 		want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1775 	else
1776 		want = CEPH_CAP_FILE_BUFFER;
1777 
1778 	ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, endoff, &got, NULL);
1779 	if (ret < 0)
1780 		goto unlock;
1781 
1782 	ceph_zero_pagecache_range(inode, offset, length);
1783 	ret = ceph_zero_objects(inode, offset, length);
1784 
1785 	if (!ret) {
1786 		spin_lock(&ci->i_ceph_lock);
1787 		ci->i_inline_version = CEPH_INLINE_NONE;
1788 		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1789 					       &prealloc_cf);
1790 		spin_unlock(&ci->i_ceph_lock);
1791 		if (dirty)
1792 			__mark_inode_dirty(inode, dirty);
1793 	}
1794 
1795 	ceph_put_cap_refs(ci, got);
1796 unlock:
1797 	inode_unlock(inode);
1798 	ceph_free_cap_flush(prealloc_cf);
1799 	return ret;
1800 }
1801 
1802 /*
1803  * This function tries to get FILE_WR capabilities for dst_ci and FILE_RD for
1804  * src_ci.  Two attempts are made to obtain both caps, and an error is return if
1805  * this fails; zero is returned on success.
1806  */
1807 static int get_rd_wr_caps(struct ceph_inode_info *src_ci,
1808 			  loff_t src_endoff, int *src_got,
1809 			  struct ceph_inode_info *dst_ci,
1810 			  loff_t dst_endoff, int *dst_got)
1811 {
1812 	int ret = 0;
1813 	bool retrying = false;
1814 
1815 retry_caps:
1816 	ret = ceph_get_caps(dst_ci, CEPH_CAP_FILE_WR, CEPH_CAP_FILE_BUFFER,
1817 			    dst_endoff, dst_got, NULL);
1818 	if (ret < 0)
1819 		return ret;
1820 
1821 	/*
1822 	 * Since we're already holding the FILE_WR capability for the dst file,
1823 	 * we would risk a deadlock by using ceph_get_caps.  Thus, we'll do some
1824 	 * retry dance instead to try to get both capabilities.
1825 	 */
1826 	ret = ceph_try_get_caps(src_ci, CEPH_CAP_FILE_RD, CEPH_CAP_FILE_SHARED,
1827 				false, src_got);
1828 	if (ret <= 0) {
1829 		/* Start by dropping dst_ci caps and getting src_ci caps */
1830 		ceph_put_cap_refs(dst_ci, *dst_got);
1831 		if (retrying) {
1832 			if (!ret)
1833 				/* ceph_try_get_caps masks EAGAIN */
1834 				ret = -EAGAIN;
1835 			return ret;
1836 		}
1837 		ret = ceph_get_caps(src_ci, CEPH_CAP_FILE_RD,
1838 				    CEPH_CAP_FILE_SHARED, src_endoff,
1839 				    src_got, NULL);
1840 		if (ret < 0)
1841 			return ret;
1842 		/*... drop src_ci caps too, and retry */
1843 		ceph_put_cap_refs(src_ci, *src_got);
1844 		retrying = true;
1845 		goto retry_caps;
1846 	}
1847 	return ret;
1848 }
1849 
1850 static void put_rd_wr_caps(struct ceph_inode_info *src_ci, int src_got,
1851 			   struct ceph_inode_info *dst_ci, int dst_got)
1852 {
1853 	ceph_put_cap_refs(src_ci, src_got);
1854 	ceph_put_cap_refs(dst_ci, dst_got);
1855 }
1856 
1857 /*
1858  * This function does several size-related checks, returning an error if:
1859  *  - source file is smaller than off+len
1860  *  - destination file size is not OK (inode_newsize_ok())
1861  *  - max bytes quotas is exceeded
1862  */
1863 static int is_file_size_ok(struct inode *src_inode, struct inode *dst_inode,
1864 			   loff_t src_off, loff_t dst_off, size_t len)
1865 {
1866 	loff_t size, endoff;
1867 
1868 	size = i_size_read(src_inode);
1869 	/*
1870 	 * Don't copy beyond source file EOF.  Instead of simply setting length
1871 	 * to (size - src_off), just drop to VFS default implementation, as the
1872 	 * local i_size may be stale due to other clients writing to the source
1873 	 * inode.
1874 	 */
1875 	if (src_off + len > size) {
1876 		dout("Copy beyond EOF (%llu + %zu > %llu)\n",
1877 		     src_off, len, size);
1878 		return -EOPNOTSUPP;
1879 	}
1880 	size = i_size_read(dst_inode);
1881 
1882 	endoff = dst_off + len;
1883 	if (inode_newsize_ok(dst_inode, endoff))
1884 		return -EOPNOTSUPP;
1885 
1886 	if (ceph_quota_is_max_bytes_exceeded(dst_inode, endoff))
1887 		return -EDQUOT;
1888 
1889 	return 0;
1890 }
1891 
1892 static ssize_t ceph_copy_file_range(struct file *src_file, loff_t src_off,
1893 				    struct file *dst_file, loff_t dst_off,
1894 				    size_t len, unsigned int flags)
1895 {
1896 	struct inode *src_inode = file_inode(src_file);
1897 	struct inode *dst_inode = file_inode(dst_file);
1898 	struct ceph_inode_info *src_ci = ceph_inode(src_inode);
1899 	struct ceph_inode_info *dst_ci = ceph_inode(dst_inode);
1900 	struct ceph_cap_flush *prealloc_cf;
1901 	struct ceph_object_locator src_oloc, dst_oloc;
1902 	struct ceph_object_id src_oid, dst_oid;
1903 	loff_t endoff = 0, size;
1904 	ssize_t ret = -EIO;
1905 	u64 src_objnum, dst_objnum, src_objoff, dst_objoff;
1906 	u32 src_objlen, dst_objlen, object_size;
1907 	int src_got = 0, dst_got = 0, err, dirty;
1908 	bool do_final_copy = false;
1909 
1910 	if (src_inode == dst_inode)
1911 		return -EINVAL;
1912 	if (ceph_snap(dst_inode) != CEPH_NOSNAP)
1913 		return -EROFS;
1914 
1915 	/*
1916 	 * Some of the checks below will return -EOPNOTSUPP, which will force a
1917 	 * fallback to the default VFS copy_file_range implementation.  This is
1918 	 * desirable in several cases (for ex, the 'len' is smaller than the
1919 	 * size of the objects, or in cases where that would be more
1920 	 * efficient).
1921 	 */
1922 
1923 	if (ceph_test_mount_opt(ceph_inode_to_client(src_inode), NOCOPYFROM))
1924 		return -EOPNOTSUPP;
1925 
1926 	if ((src_ci->i_layout.stripe_unit != dst_ci->i_layout.stripe_unit) ||
1927 	    (src_ci->i_layout.stripe_count != dst_ci->i_layout.stripe_count) ||
1928 	    (src_ci->i_layout.object_size != dst_ci->i_layout.object_size))
1929 		return -EOPNOTSUPP;
1930 
1931 	if (len < src_ci->i_layout.object_size)
1932 		return -EOPNOTSUPP; /* no remote copy will be done */
1933 
1934 	prealloc_cf = ceph_alloc_cap_flush();
1935 	if (!prealloc_cf)
1936 		return -ENOMEM;
1937 
1938 	/* Start by sync'ing the source and destination files */
1939 	ret = file_write_and_wait_range(src_file, src_off, (src_off + len));
1940 	if (ret < 0) {
1941 		dout("failed to write src file (%zd)\n", ret);
1942 		goto out;
1943 	}
1944 	ret = file_write_and_wait_range(dst_file, dst_off, (dst_off + len));
1945 	if (ret < 0) {
1946 		dout("failed to write dst file (%zd)\n", ret);
1947 		goto out;
1948 	}
1949 
1950 	/*
1951 	 * We need FILE_WR caps for dst_ci and FILE_RD for src_ci as other
1952 	 * clients may have dirty data in their caches.  And OSDs know nothing
1953 	 * about caps, so they can't safely do the remote object copies.
1954 	 */
1955 	err = get_rd_wr_caps(src_ci, (src_off + len), &src_got,
1956 			     dst_ci, (dst_off + len), &dst_got);
1957 	if (err < 0) {
1958 		dout("get_rd_wr_caps returned %d\n", err);
1959 		ret = -EOPNOTSUPP;
1960 		goto out;
1961 	}
1962 
1963 	ret = is_file_size_ok(src_inode, dst_inode, src_off, dst_off, len);
1964 	if (ret < 0)
1965 		goto out_caps;
1966 
1967 	size = i_size_read(dst_inode);
1968 	endoff = dst_off + len;
1969 
1970 	/* Drop dst file cached pages */
1971 	ret = invalidate_inode_pages2_range(dst_inode->i_mapping,
1972 					    dst_off >> PAGE_SHIFT,
1973 					    endoff >> PAGE_SHIFT);
1974 	if (ret < 0) {
1975 		dout("Failed to invalidate inode pages (%zd)\n", ret);
1976 		ret = 0; /* XXX */
1977 	}
1978 	src_oloc.pool = src_ci->i_layout.pool_id;
1979 	src_oloc.pool_ns = ceph_try_get_string(src_ci->i_layout.pool_ns);
1980 	dst_oloc.pool = dst_ci->i_layout.pool_id;
1981 	dst_oloc.pool_ns = ceph_try_get_string(dst_ci->i_layout.pool_ns);
1982 
1983 	ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
1984 				      src_ci->i_layout.object_size,
1985 				      &src_objnum, &src_objoff, &src_objlen);
1986 	ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
1987 				      dst_ci->i_layout.object_size,
1988 				      &dst_objnum, &dst_objoff, &dst_objlen);
1989 	/* object-level offsets need to the same */
1990 	if (src_objoff != dst_objoff) {
1991 		ret = -EOPNOTSUPP;
1992 		goto out_caps;
1993 	}
1994 
1995 	/*
1996 	 * Do a manual copy if the object offset isn't object aligned.
1997 	 * 'src_objlen' contains the bytes left until the end of the object,
1998 	 * starting at the src_off
1999 	 */
2000 	if (src_objoff) {
2001 		/*
2002 		 * we need to temporarily drop all caps as we'll be calling
2003 		 * {read,write}_iter, which will get caps again.
2004 		 */
2005 		put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2006 		ret = do_splice_direct(src_file, &src_off, dst_file,
2007 				       &dst_off, src_objlen, flags);
2008 		if (ret < 0) {
2009 			dout("do_splice_direct returned %d\n", err);
2010 			goto out;
2011 		}
2012 		len -= ret;
2013 		err = get_rd_wr_caps(src_ci, (src_off + len),
2014 				     &src_got, dst_ci,
2015 				     (dst_off + len), &dst_got);
2016 		if (err < 0)
2017 			goto out;
2018 		err = is_file_size_ok(src_inode, dst_inode,
2019 				      src_off, dst_off, len);
2020 		if (err < 0)
2021 			goto out_caps;
2022 	}
2023 	object_size = src_ci->i_layout.object_size;
2024 	while (len >= object_size) {
2025 		ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
2026 					      object_size, &src_objnum,
2027 					      &src_objoff, &src_objlen);
2028 		ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
2029 					      object_size, &dst_objnum,
2030 					      &dst_objoff, &dst_objlen);
2031 		ceph_oid_init(&src_oid);
2032 		ceph_oid_printf(&src_oid, "%llx.%08llx",
2033 				src_ci->i_vino.ino, src_objnum);
2034 		ceph_oid_init(&dst_oid);
2035 		ceph_oid_printf(&dst_oid, "%llx.%08llx",
2036 				dst_ci->i_vino.ino, dst_objnum);
2037 		/* Do an object remote copy */
2038 		err = ceph_osdc_copy_from(
2039 			&ceph_inode_to_client(src_inode)->client->osdc,
2040 			src_ci->i_vino.snap, 0,
2041 			&src_oid, &src_oloc,
2042 			CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2043 			CEPH_OSD_OP_FLAG_FADVISE_NOCACHE,
2044 			&dst_oid, &dst_oloc,
2045 			CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2046 			CEPH_OSD_OP_FLAG_FADVISE_DONTNEED, 0);
2047 		if (err) {
2048 			dout("ceph_osdc_copy_from returned %d\n", err);
2049 			if (!ret)
2050 				ret = err;
2051 			goto out_caps;
2052 		}
2053 		len -= object_size;
2054 		src_off += object_size;
2055 		dst_off += object_size;
2056 		ret += object_size;
2057 	}
2058 
2059 	if (len)
2060 		/* We still need one final local copy */
2061 		do_final_copy = true;
2062 
2063 	file_update_time(dst_file);
2064 	if (endoff > size) {
2065 		int caps_flags = 0;
2066 
2067 		/* Let the MDS know about dst file size change */
2068 		if (ceph_quota_is_max_bytes_approaching(dst_inode, endoff))
2069 			caps_flags |= CHECK_CAPS_NODELAY;
2070 		if (ceph_inode_set_size(dst_inode, endoff))
2071 			caps_flags |= CHECK_CAPS_AUTHONLY;
2072 		if (caps_flags)
2073 			ceph_check_caps(dst_ci, caps_flags, NULL);
2074 	}
2075 	/* Mark Fw dirty */
2076 	spin_lock(&dst_ci->i_ceph_lock);
2077 	dst_ci->i_inline_version = CEPH_INLINE_NONE;
2078 	dirty = __ceph_mark_dirty_caps(dst_ci, CEPH_CAP_FILE_WR, &prealloc_cf);
2079 	spin_unlock(&dst_ci->i_ceph_lock);
2080 	if (dirty)
2081 		__mark_inode_dirty(dst_inode, dirty);
2082 
2083 out_caps:
2084 	put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2085 
2086 	if (do_final_copy) {
2087 		err = do_splice_direct(src_file, &src_off, dst_file,
2088 				       &dst_off, len, flags);
2089 		if (err < 0) {
2090 			dout("do_splice_direct returned %d\n", err);
2091 			goto out;
2092 		}
2093 		len -= err;
2094 		ret += err;
2095 	}
2096 
2097 out:
2098 	ceph_free_cap_flush(prealloc_cf);
2099 
2100 	return ret;
2101 }
2102 
2103 const struct file_operations ceph_file_fops = {
2104 	.open = ceph_open,
2105 	.release = ceph_release,
2106 	.llseek = ceph_llseek,
2107 	.read_iter = ceph_read_iter,
2108 	.write_iter = ceph_write_iter,
2109 	.mmap = ceph_mmap,
2110 	.fsync = ceph_fsync,
2111 	.lock = ceph_lock,
2112 	.flock = ceph_flock,
2113 	.splice_read = generic_file_splice_read,
2114 	.splice_write = iter_file_splice_write,
2115 	.unlocked_ioctl = ceph_ioctl,
2116 	.compat_ioctl	= ceph_ioctl,
2117 	.fallocate	= ceph_fallocate,
2118 	.copy_file_range = ceph_copy_file_range,
2119 };
2120