xref: /openbmc/linux/fs/ceph/file.c (revision b18d2847)
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 
22 static __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 
83 static 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  */
130 static 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 
164 static 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  */
182 static struct ceph_mds_request *
183 prepare_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);
199 out:
200 	return req;
201 }
202 
203 static 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_mount_options *opt =
208 		ceph_inode_to_client(&ci->vfs_inode)->mount_options;
209 	struct ceph_file_info *fi;
210 	int ret;
211 
212 	dout("%s %p %p 0%o (%s)\n", __func__, inode, file,
213 			inode->i_mode, isdir ? "dir" : "regular");
214 	BUG_ON(inode->i_fop->release != ceph_release);
215 
216 	if (isdir) {
217 		struct ceph_dir_file_info *dfi =
218 			kmem_cache_zalloc(ceph_dir_file_cachep, GFP_KERNEL);
219 		if (!dfi)
220 			return -ENOMEM;
221 
222 		file->private_data = dfi;
223 		fi = &dfi->file_info;
224 		dfi->next_offset = 2;
225 		dfi->readdir_cache_idx = -1;
226 	} else {
227 		fi = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
228 		if (!fi)
229 			return -ENOMEM;
230 
231 		if (opt->flags & CEPH_MOUNT_OPT_NOPAGECACHE)
232 			fi->flags |= CEPH_F_SYNC;
233 
234 		file->private_data = fi;
235 	}
236 
237 	ceph_get_fmode(ci, fmode, 1);
238 	fi->fmode = fmode;
239 
240 	spin_lock_init(&fi->rw_contexts_lock);
241 	INIT_LIST_HEAD(&fi->rw_contexts);
242 	fi->filp_gen = READ_ONCE(ceph_inode_to_client(inode)->filp_gen);
243 
244 	if ((file->f_mode & FMODE_WRITE) &&
245 	    ci->i_inline_version != CEPH_INLINE_NONE) {
246 		ret = ceph_uninline_data(file);
247 		if (ret < 0)
248 			goto error;
249 	}
250 
251 	return 0;
252 
253 error:
254 	ceph_fscache_unuse_cookie(inode, file->f_mode & FMODE_WRITE);
255 	ceph_put_fmode(ci, fi->fmode, 1);
256 	kmem_cache_free(ceph_file_cachep, fi);
257 	/* wake up anyone waiting for caps on this inode */
258 	wake_up_all(&ci->i_cap_wq);
259 	return ret;
260 }
261 
262 /*
263  * initialize private struct file data.
264  * if we fail, clean up by dropping fmode reference on the ceph_inode
265  */
266 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
267 {
268 	int ret = 0;
269 
270 	switch (inode->i_mode & S_IFMT) {
271 	case S_IFREG:
272 		ceph_fscache_use_cookie(inode, file->f_mode & FMODE_WRITE);
273 		fallthrough;
274 	case S_IFDIR:
275 		ret = ceph_init_file_info(inode, file, fmode,
276 						S_ISDIR(inode->i_mode));
277 		break;
278 
279 	case S_IFLNK:
280 		dout("init_file %p %p 0%o (symlink)\n", inode, file,
281 		     inode->i_mode);
282 		break;
283 
284 	default:
285 		dout("init_file %p %p 0%o (special)\n", inode, file,
286 		     inode->i_mode);
287 		/*
288 		 * we need to drop the open ref now, since we don't
289 		 * have .release set to ceph_release.
290 		 */
291 		BUG_ON(inode->i_fop->release == ceph_release);
292 
293 		/* call the proper open fop */
294 		ret = inode->i_fop->open(inode, file);
295 	}
296 	return ret;
297 }
298 
299 /*
300  * try renew caps after session gets killed.
301  */
302 int ceph_renew_caps(struct inode *inode, int fmode)
303 {
304 	struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
305 	struct ceph_inode_info *ci = ceph_inode(inode);
306 	struct ceph_mds_request *req;
307 	int err, flags, wanted;
308 
309 	spin_lock(&ci->i_ceph_lock);
310 	__ceph_touch_fmode(ci, mdsc, fmode);
311 	wanted = __ceph_caps_file_wanted(ci);
312 	if (__ceph_is_any_real_caps(ci) &&
313 	    (!(wanted & CEPH_CAP_ANY_WR) || ci->i_auth_cap)) {
314 		int issued = __ceph_caps_issued(ci, NULL);
315 		spin_unlock(&ci->i_ceph_lock);
316 		dout("renew caps %p want %s issued %s updating mds_wanted\n",
317 		     inode, ceph_cap_string(wanted), ceph_cap_string(issued));
318 		ceph_check_caps(ci, 0, NULL);
319 		return 0;
320 	}
321 	spin_unlock(&ci->i_ceph_lock);
322 
323 	flags = 0;
324 	if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
325 		flags = O_RDWR;
326 	else if (wanted & CEPH_CAP_FILE_RD)
327 		flags = O_RDONLY;
328 	else if (wanted & CEPH_CAP_FILE_WR)
329 		flags = O_WRONLY;
330 #ifdef O_LAZY
331 	if (wanted & CEPH_CAP_FILE_LAZYIO)
332 		flags |= O_LAZY;
333 #endif
334 
335 	req = prepare_open_request(inode->i_sb, flags, 0);
336 	if (IS_ERR(req)) {
337 		err = PTR_ERR(req);
338 		goto out;
339 	}
340 
341 	req->r_inode = inode;
342 	ihold(inode);
343 	req->r_num_caps = 1;
344 
345 	err = ceph_mdsc_do_request(mdsc, NULL, req);
346 	ceph_mdsc_put_request(req);
347 out:
348 	dout("renew caps %p open result=%d\n", inode, err);
349 	return err < 0 ? err : 0;
350 }
351 
352 /*
353  * If we already have the requisite capabilities, we can satisfy
354  * the open request locally (no need to request new caps from the
355  * MDS).  We do, however, need to inform the MDS (asynchronously)
356  * if our wanted caps set expands.
357  */
358 int ceph_open(struct inode *inode, struct file *file)
359 {
360 	struct ceph_inode_info *ci = ceph_inode(inode);
361 	struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
362 	struct ceph_mds_client *mdsc = fsc->mdsc;
363 	struct ceph_mds_request *req;
364 	struct ceph_file_info *fi = file->private_data;
365 	int err;
366 	int flags, fmode, wanted;
367 
368 	if (fi) {
369 		dout("open file %p is already opened\n", file);
370 		return 0;
371 	}
372 
373 	/* filter out O_CREAT|O_EXCL; vfs did that already.  yuck. */
374 	flags = file->f_flags & ~(O_CREAT|O_EXCL);
375 	if (S_ISDIR(inode->i_mode))
376 		flags = O_DIRECTORY;  /* mds likes to know */
377 
378 	dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
379 	     ceph_vinop(inode), file, flags, file->f_flags);
380 	fmode = ceph_flags_to_mode(flags);
381 	wanted = ceph_caps_for_mode(fmode);
382 
383 	/* snapped files are read-only */
384 	if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
385 		return -EROFS;
386 
387 	/* trivially open snapdir */
388 	if (ceph_snap(inode) == CEPH_SNAPDIR) {
389 		return ceph_init_file(inode, file, fmode);
390 	}
391 
392 	/*
393 	 * No need to block if we have caps on the auth MDS (for
394 	 * write) or any MDS (for read).  Update wanted set
395 	 * asynchronously.
396 	 */
397 	spin_lock(&ci->i_ceph_lock);
398 	if (__ceph_is_any_real_caps(ci) &&
399 	    (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
400 		int mds_wanted = __ceph_caps_mds_wanted(ci, true);
401 		int issued = __ceph_caps_issued(ci, NULL);
402 
403 		dout("open %p fmode %d want %s issued %s using existing\n",
404 		     inode, fmode, ceph_cap_string(wanted),
405 		     ceph_cap_string(issued));
406 		__ceph_touch_fmode(ci, mdsc, fmode);
407 		spin_unlock(&ci->i_ceph_lock);
408 
409 		/* adjust wanted? */
410 		if ((issued & wanted) != wanted &&
411 		    (mds_wanted & wanted) != wanted &&
412 		    ceph_snap(inode) != CEPH_SNAPDIR)
413 			ceph_check_caps(ci, 0, NULL);
414 
415 		return ceph_init_file(inode, file, fmode);
416 	} else if (ceph_snap(inode) != CEPH_NOSNAP &&
417 		   (ci->i_snap_caps & wanted) == wanted) {
418 		__ceph_touch_fmode(ci, mdsc, fmode);
419 		spin_unlock(&ci->i_ceph_lock);
420 		return ceph_init_file(inode, file, fmode);
421 	}
422 
423 	spin_unlock(&ci->i_ceph_lock);
424 
425 	dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
426 	req = prepare_open_request(inode->i_sb, flags, 0);
427 	if (IS_ERR(req)) {
428 		err = PTR_ERR(req);
429 		goto out;
430 	}
431 	req->r_inode = inode;
432 	ihold(inode);
433 
434 	req->r_num_caps = 1;
435 	err = ceph_mdsc_do_request(mdsc, NULL, req);
436 	if (!err)
437 		err = ceph_init_file(inode, file, req->r_fmode);
438 	ceph_mdsc_put_request(req);
439 	dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
440 out:
441 	return err;
442 }
443 
444 /* Clone the layout from a synchronous create, if the dir now has Dc caps */
445 static void
446 cache_file_layout(struct inode *dst, struct inode *src)
447 {
448 	struct ceph_inode_info *cdst = ceph_inode(dst);
449 	struct ceph_inode_info *csrc = ceph_inode(src);
450 
451 	spin_lock(&cdst->i_ceph_lock);
452 	if ((__ceph_caps_issued(cdst, NULL) & CEPH_CAP_DIR_CREATE) &&
453 	    !ceph_file_layout_is_valid(&cdst->i_cached_layout)) {
454 		memcpy(&cdst->i_cached_layout, &csrc->i_layout,
455 			sizeof(cdst->i_cached_layout));
456 		rcu_assign_pointer(cdst->i_cached_layout.pool_ns,
457 				   ceph_try_get_string(csrc->i_layout.pool_ns));
458 	}
459 	spin_unlock(&cdst->i_ceph_lock);
460 }
461 
462 /*
463  * Try to set up an async create. We need caps, a file layout, and inode number,
464  * and either a lease on the dentry or complete dir info. If any of those
465  * criteria are not satisfied, then return false and the caller can go
466  * synchronous.
467  */
468 static int try_prep_async_create(struct inode *dir, struct dentry *dentry,
469 				 struct ceph_file_layout *lo, u64 *pino)
470 {
471 	struct ceph_inode_info *ci = ceph_inode(dir);
472 	struct ceph_dentry_info *di = ceph_dentry(dentry);
473 	int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_CREATE;
474 	u64 ino;
475 
476 	spin_lock(&ci->i_ceph_lock);
477 	/* No auth cap means no chance for Dc caps */
478 	if (!ci->i_auth_cap)
479 		goto no_async;
480 
481 	/* Any delegated inos? */
482 	if (xa_empty(&ci->i_auth_cap->session->s_delegated_inos))
483 		goto no_async;
484 
485 	if (!ceph_file_layout_is_valid(&ci->i_cached_layout))
486 		goto no_async;
487 
488 	if ((__ceph_caps_issued(ci, NULL) & want) != want)
489 		goto no_async;
490 
491 	if (d_in_lookup(dentry)) {
492 		if (!__ceph_dir_is_complete(ci))
493 			goto no_async;
494 		spin_lock(&dentry->d_lock);
495 		di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
496 		spin_unlock(&dentry->d_lock);
497 	} else if (atomic_read(&ci->i_shared_gen) !=
498 		   READ_ONCE(di->lease_shared_gen)) {
499 		goto no_async;
500 	}
501 
502 	ino = ceph_get_deleg_ino(ci->i_auth_cap->session);
503 	if (!ino)
504 		goto no_async;
505 
506 	*pino = ino;
507 	ceph_take_cap_refs(ci, want, false);
508 	memcpy(lo, &ci->i_cached_layout, sizeof(*lo));
509 	rcu_assign_pointer(lo->pool_ns,
510 			   ceph_try_get_string(ci->i_cached_layout.pool_ns));
511 	got = want;
512 no_async:
513 	spin_unlock(&ci->i_ceph_lock);
514 	return got;
515 }
516 
517 static void restore_deleg_ino(struct inode *dir, u64 ino)
518 {
519 	struct ceph_inode_info *ci = ceph_inode(dir);
520 	struct ceph_mds_session *s = NULL;
521 
522 	spin_lock(&ci->i_ceph_lock);
523 	if (ci->i_auth_cap)
524 		s = ceph_get_mds_session(ci->i_auth_cap->session);
525 	spin_unlock(&ci->i_ceph_lock);
526 	if (s) {
527 		int err = ceph_restore_deleg_ino(s, ino);
528 		if (err)
529 			pr_warn("ceph: unable to restore delegated ino 0x%llx to session: %d\n",
530 				ino, err);
531 		ceph_put_mds_session(s);
532 	}
533 }
534 
535 static void wake_async_create_waiters(struct inode *inode,
536 				      struct ceph_mds_session *session)
537 {
538 	struct ceph_inode_info *ci = ceph_inode(inode);
539 
540 	spin_lock(&ci->i_ceph_lock);
541 	if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE) {
542 		ci->i_ceph_flags &= ~CEPH_I_ASYNC_CREATE;
543 		wake_up_bit(&ci->i_ceph_flags, CEPH_ASYNC_CREATE_BIT);
544 	}
545 	ceph_kick_flushing_inode_caps(session, ci);
546 	spin_unlock(&ci->i_ceph_lock);
547 }
548 
549 static void ceph_async_create_cb(struct ceph_mds_client *mdsc,
550                                  struct ceph_mds_request *req)
551 {
552 	struct dentry *dentry = req->r_dentry;
553 	struct inode *dinode = d_inode(dentry);
554 	struct inode *tinode = req->r_target_inode;
555 	int result = req->r_err ? req->r_err :
556 			le32_to_cpu(req->r_reply_info.head->result);
557 
558 	WARN_ON_ONCE(dinode && tinode && dinode != tinode);
559 
560 	/* MDS changed -- caller must resubmit */
561 	if (result == -EJUKEBOX)
562 		goto out;
563 
564 	mapping_set_error(req->r_parent->i_mapping, result);
565 
566 	if (result) {
567 		int pathlen = 0;
568 		u64 base = 0;
569 		char *path = ceph_mdsc_build_path(req->r_dentry, &pathlen,
570 						  &base, 0);
571 
572 		pr_warn("ceph: async create failure path=(%llx)%s result=%d!\n",
573 			base, IS_ERR(path) ? "<<bad>>" : path, result);
574 		ceph_mdsc_free_path(path, pathlen);
575 
576 		ceph_dir_clear_complete(req->r_parent);
577 		if (!d_unhashed(dentry))
578 			d_drop(dentry);
579 
580 		if (dinode) {
581 			mapping_set_error(dinode->i_mapping, result);
582 			ceph_inode_shutdown(dinode);
583 			wake_async_create_waiters(dinode, req->r_session);
584 		}
585 	}
586 
587 	if (tinode) {
588 		u64 ino = ceph_vino(tinode).ino;
589 
590 		if (req->r_deleg_ino != ino)
591 			pr_warn("%s: inode number mismatch! err=%d deleg_ino=0x%llx target=0x%llx\n",
592 				__func__, req->r_err, req->r_deleg_ino, ino);
593 
594 		mapping_set_error(tinode->i_mapping, result);
595 		wake_async_create_waiters(tinode, req->r_session);
596 	} else if (!result) {
597 		pr_warn("%s: no req->r_target_inode for 0x%llx\n", __func__,
598 			req->r_deleg_ino);
599 	}
600 out:
601 	ceph_mdsc_release_dir_caps(req);
602 }
603 
604 static int ceph_finish_async_create(struct inode *dir, struct dentry *dentry,
605 				    struct file *file, umode_t mode,
606 				    struct ceph_mds_request *req,
607 				    struct ceph_acl_sec_ctx *as_ctx,
608 				    struct ceph_file_layout *lo)
609 {
610 	int ret;
611 	char xattr_buf[4];
612 	struct ceph_mds_reply_inode in = { };
613 	struct ceph_mds_reply_info_in iinfo = { .in = &in };
614 	struct ceph_inode_info *ci = ceph_inode(dir);
615 	struct inode *inode;
616 	struct timespec64 now;
617 	struct ceph_string *pool_ns;
618 	struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
619 	struct ceph_vino vino = { .ino = req->r_deleg_ino,
620 				  .snap = CEPH_NOSNAP };
621 
622 	ktime_get_real_ts64(&now);
623 
624 	inode = ceph_get_inode(dentry->d_sb, vino);
625 	if (IS_ERR(inode))
626 		return PTR_ERR(inode);
627 
628 	iinfo.inline_version = CEPH_INLINE_NONE;
629 	iinfo.change_attr = 1;
630 	ceph_encode_timespec64(&iinfo.btime, &now);
631 
632 	iinfo.xattr_len = ARRAY_SIZE(xattr_buf);
633 	iinfo.xattr_data = xattr_buf;
634 	memset(iinfo.xattr_data, 0, iinfo.xattr_len);
635 
636 	in.ino = cpu_to_le64(vino.ino);
637 	in.snapid = cpu_to_le64(CEPH_NOSNAP);
638 	in.version = cpu_to_le64(1);	// ???
639 	in.cap.caps = in.cap.wanted = cpu_to_le32(CEPH_CAP_ALL_FILE);
640 	in.cap.cap_id = cpu_to_le64(1);
641 	in.cap.realm = cpu_to_le64(ci->i_snap_realm->ino);
642 	in.cap.flags = CEPH_CAP_FLAG_AUTH;
643 	in.ctime = in.mtime = in.atime = iinfo.btime;
644 	in.truncate_seq = cpu_to_le32(1);
645 	in.truncate_size = cpu_to_le64(-1ULL);
646 	in.xattr_version = cpu_to_le64(1);
647 	in.uid = cpu_to_le32(from_kuid(&init_user_ns, current_fsuid()));
648 	if (dir->i_mode & S_ISGID) {
649 		in.gid = cpu_to_le32(from_kgid(&init_user_ns, dir->i_gid));
650 
651 		/* Directories always inherit the setgid bit. */
652 		if (S_ISDIR(mode))
653 			mode |= S_ISGID;
654 		else if ((mode & (S_ISGID | S_IXGRP)) == (S_ISGID | S_IXGRP) &&
655 			 !in_group_p(dir->i_gid) &&
656 			 !capable_wrt_inode_uidgid(&init_user_ns, dir, CAP_FSETID))
657 			mode &= ~S_ISGID;
658 	} else {
659 		in.gid = cpu_to_le32(from_kgid(&init_user_ns, current_fsgid()));
660 	}
661 	in.mode = cpu_to_le32((u32)mode);
662 
663 	in.nlink = cpu_to_le32(1);
664 	in.max_size = cpu_to_le64(lo->stripe_unit);
665 
666 	ceph_file_layout_to_legacy(lo, &in.layout);
667 	/* lo is private, so pool_ns can't change */
668 	pool_ns = rcu_dereference_raw(lo->pool_ns);
669 	if (pool_ns) {
670 		iinfo.pool_ns_len = pool_ns->len;
671 		iinfo.pool_ns_data = pool_ns->str;
672 	}
673 
674 	down_read(&mdsc->snap_rwsem);
675 	ret = ceph_fill_inode(inode, NULL, &iinfo, NULL, req->r_session,
676 			      req->r_fmode, NULL);
677 	up_read(&mdsc->snap_rwsem);
678 	if (ret) {
679 		dout("%s failed to fill inode: %d\n", __func__, ret);
680 		ceph_dir_clear_complete(dir);
681 		if (!d_unhashed(dentry))
682 			d_drop(dentry);
683 		if (inode->i_state & I_NEW)
684 			discard_new_inode(inode);
685 	} else {
686 		struct dentry *dn;
687 
688 		dout("%s d_adding new inode 0x%llx to 0x%llx/%s\n", __func__,
689 			vino.ino, ceph_ino(dir), dentry->d_name.name);
690 		ceph_dir_clear_ordered(dir);
691 		ceph_init_inode_acls(inode, as_ctx);
692 		if (inode->i_state & I_NEW) {
693 			/*
694 			 * If it's not I_NEW, then someone created this before
695 			 * we got here. Assume the server is aware of it at
696 			 * that point and don't worry about setting
697 			 * CEPH_I_ASYNC_CREATE.
698 			 */
699 			ceph_inode(inode)->i_ceph_flags = CEPH_I_ASYNC_CREATE;
700 			unlock_new_inode(inode);
701 		}
702 		if (d_in_lookup(dentry) || d_really_is_negative(dentry)) {
703 			if (!d_unhashed(dentry))
704 				d_drop(dentry);
705 			dn = d_splice_alias(inode, dentry);
706 			WARN_ON_ONCE(dn && dn != dentry);
707 		}
708 		file->f_mode |= FMODE_CREATED;
709 		ret = finish_open(file, dentry, ceph_open);
710 	}
711 	return ret;
712 }
713 
714 /*
715  * Do a lookup + open with a single request.  If we get a non-existent
716  * file or symlink, return 1 so the VFS can retry.
717  */
718 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
719 		     struct file *file, unsigned flags, umode_t mode)
720 {
721 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
722 	struct ceph_mds_client *mdsc = fsc->mdsc;
723 	struct ceph_mds_request *req;
724 	struct dentry *dn;
725 	struct ceph_acl_sec_ctx as_ctx = {};
726 	bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
727 	int mask;
728 	int err;
729 
730 	dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
731 	     dir, dentry, dentry,
732 	     d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
733 
734 	if (dentry->d_name.len > NAME_MAX)
735 		return -ENAMETOOLONG;
736 
737 	if (flags & O_CREAT) {
738 		if (ceph_quota_is_max_files_exceeded(dir))
739 			return -EDQUOT;
740 		err = ceph_pre_init_acls(dir, &mode, &as_ctx);
741 		if (err < 0)
742 			return err;
743 		err = ceph_security_init_secctx(dentry, mode, &as_ctx);
744 		if (err < 0)
745 			goto out_ctx;
746 	} else if (!d_in_lookup(dentry)) {
747 		/* If it's not being looked up, it's negative */
748 		return -ENOENT;
749 	}
750 retry:
751 	/* do the open */
752 	req = prepare_open_request(dir->i_sb, flags, mode);
753 	if (IS_ERR(req)) {
754 		err = PTR_ERR(req);
755 		goto out_ctx;
756 	}
757 	req->r_dentry = dget(dentry);
758 	req->r_num_caps = 2;
759 	mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
760 	if (ceph_security_xattr_wanted(dir))
761 		mask |= CEPH_CAP_XATTR_SHARED;
762 	req->r_args.open.mask = cpu_to_le32(mask);
763 	req->r_parent = dir;
764 	ihold(dir);
765 
766 	if (flags & O_CREAT) {
767 		struct ceph_file_layout lo;
768 
769 		req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
770 		req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
771 		if (as_ctx.pagelist) {
772 			req->r_pagelist = as_ctx.pagelist;
773 			as_ctx.pagelist = NULL;
774 		}
775 		if (try_async &&
776 		    (req->r_dir_caps =
777 		      try_prep_async_create(dir, dentry, &lo,
778 					    &req->r_deleg_ino))) {
779 			set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
780 			req->r_args.open.flags |= cpu_to_le32(CEPH_O_EXCL);
781 			req->r_callback = ceph_async_create_cb;
782 			err = ceph_mdsc_submit_request(mdsc, dir, req);
783 			if (!err) {
784 				err = ceph_finish_async_create(dir, dentry,
785 							file, mode, req,
786 							&as_ctx, &lo);
787 			} else if (err == -EJUKEBOX) {
788 				restore_deleg_ino(dir, req->r_deleg_ino);
789 				ceph_mdsc_put_request(req);
790 				try_async = false;
791 				ceph_put_string(rcu_dereference_raw(lo.pool_ns));
792 				goto retry;
793 			}
794 			ceph_put_string(rcu_dereference_raw(lo.pool_ns));
795 			goto out_req;
796 		}
797 	}
798 
799 	set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
800 	err = ceph_mdsc_do_request(mdsc,
801 				   (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
802 				   req);
803 	if (err == -ENOENT) {
804 		dentry = ceph_handle_snapdir(req, dentry);
805 		if (IS_ERR(dentry)) {
806 			err = PTR_ERR(dentry);
807 			goto out_req;
808 		}
809 		err = 0;
810 	}
811 
812 	if (!err && (flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
813 		err = ceph_handle_notrace_create(dir, dentry);
814 
815 	if (d_in_lookup(dentry)) {
816 		dn = ceph_finish_lookup(req, dentry, err);
817 		if (IS_ERR(dn))
818 			err = PTR_ERR(dn);
819 	} else {
820 		/* we were given a hashed negative dentry */
821 		dn = NULL;
822 	}
823 	if (err)
824 		goto out_req;
825 	if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
826 		/* make vfs retry on splice, ENOENT, or symlink */
827 		dout("atomic_open finish_no_open on dn %p\n", dn);
828 		err = finish_no_open(file, dn);
829 	} else {
830 		dout("atomic_open finish_open on dn %p\n", dn);
831 		if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
832 			struct inode *newino = d_inode(dentry);
833 
834 			cache_file_layout(dir, newino);
835 			ceph_init_inode_acls(newino, &as_ctx);
836 			file->f_mode |= FMODE_CREATED;
837 		}
838 		err = finish_open(file, dentry, ceph_open);
839 	}
840 out_req:
841 	ceph_mdsc_put_request(req);
842 out_ctx:
843 	ceph_release_acl_sec_ctx(&as_ctx);
844 	dout("atomic_open result=%d\n", err);
845 	return err;
846 }
847 
848 int ceph_release(struct inode *inode, struct file *file)
849 {
850 	struct ceph_inode_info *ci = ceph_inode(inode);
851 
852 	if (S_ISDIR(inode->i_mode)) {
853 		struct ceph_dir_file_info *dfi = file->private_data;
854 		dout("release inode %p dir file %p\n", inode, file);
855 		WARN_ON(!list_empty(&dfi->file_info.rw_contexts));
856 
857 		ceph_put_fmode(ci, dfi->file_info.fmode, 1);
858 
859 		if (dfi->last_readdir)
860 			ceph_mdsc_put_request(dfi->last_readdir);
861 		kfree(dfi->last_name);
862 		kfree(dfi->dir_info);
863 		kmem_cache_free(ceph_dir_file_cachep, dfi);
864 	} else {
865 		struct ceph_file_info *fi = file->private_data;
866 		dout("release inode %p regular file %p\n", inode, file);
867 		WARN_ON(!list_empty(&fi->rw_contexts));
868 
869 		ceph_fscache_unuse_cookie(inode, file->f_mode & FMODE_WRITE);
870 		ceph_put_fmode(ci, fi->fmode, 1);
871 
872 		kmem_cache_free(ceph_file_cachep, fi);
873 	}
874 
875 	/* wake up anyone waiting for caps on this inode */
876 	wake_up_all(&ci->i_cap_wq);
877 	return 0;
878 }
879 
880 enum {
881 	HAVE_RETRIED = 1,
882 	CHECK_EOF =    2,
883 	READ_INLINE =  3,
884 };
885 
886 /*
887  * Completely synchronous read and write methods.  Direct from __user
888  * buffer to osd, or directly to user pages (if O_DIRECT).
889  *
890  * If the read spans object boundary, just do multiple reads.  (That's not
891  * atomic, but good enough for now.)
892  *
893  * If we get a short result from the OSD, check against i_size; we need to
894  * only return a short read to the caller if we hit EOF.
895  */
896 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
897 			      int *retry_op)
898 {
899 	struct file *file = iocb->ki_filp;
900 	struct inode *inode = file_inode(file);
901 	struct ceph_inode_info *ci = ceph_inode(inode);
902 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
903 	struct ceph_osd_client *osdc = &fsc->client->osdc;
904 	ssize_t ret;
905 	u64 off = iocb->ki_pos;
906 	u64 len = iov_iter_count(to);
907 	u64 i_size = i_size_read(inode);
908 
909 	dout("sync_read on file %p %llu~%u %s\n", file, off, (unsigned)len,
910 	     (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
911 
912 	if (!len)
913 		return 0;
914 	/*
915 	 * flush any page cache pages in this range.  this
916 	 * will make concurrent normal and sync io slow,
917 	 * but it will at least behave sensibly when they are
918 	 * in sequence.
919 	 */
920 	ret = filemap_write_and_wait_range(inode->i_mapping,
921 					   off, off + len - 1);
922 	if (ret < 0)
923 		return ret;
924 
925 	ret = 0;
926 	while ((len = iov_iter_count(to)) > 0) {
927 		struct ceph_osd_request *req;
928 		struct page **pages;
929 		int num_pages;
930 		size_t page_off;
931 		bool more;
932 		int idx;
933 		size_t left;
934 
935 		req = ceph_osdc_new_request(osdc, &ci->i_layout,
936 					ci->i_vino, off, &len, 0, 1,
937 					CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
938 					NULL, ci->i_truncate_seq,
939 					ci->i_truncate_size, false);
940 		if (IS_ERR(req)) {
941 			ret = PTR_ERR(req);
942 			break;
943 		}
944 
945 		more = len < iov_iter_count(to);
946 
947 		num_pages = calc_pages_for(off, len);
948 		page_off = off & ~PAGE_MASK;
949 		pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
950 		if (IS_ERR(pages)) {
951 			ceph_osdc_put_request(req);
952 			ret = PTR_ERR(pages);
953 			break;
954 		}
955 
956 		osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_off,
957 						 false, false);
958 		ret = ceph_osdc_start_request(osdc, req, false);
959 		if (!ret)
960 			ret = ceph_osdc_wait_request(osdc, req);
961 
962 		ceph_update_read_metrics(&fsc->mdsc->metric,
963 					 req->r_start_latency,
964 					 req->r_end_latency,
965 					 len, ret);
966 
967 		ceph_osdc_put_request(req);
968 
969 		i_size = i_size_read(inode);
970 		dout("sync_read %llu~%llu got %zd i_size %llu%s\n",
971 		     off, len, ret, i_size, (more ? " MORE" : ""));
972 
973 		if (ret == -ENOENT)
974 			ret = 0;
975 		if (ret >= 0 && ret < len && (off + ret < i_size)) {
976 			int zlen = min(len - ret, i_size - off - ret);
977 			int zoff = page_off + ret;
978 			dout("sync_read zero gap %llu~%llu\n",
979                              off + ret, off + ret + zlen);
980 			ceph_zero_page_vector_range(zoff, zlen, pages);
981 			ret += zlen;
982 		}
983 
984 		idx = 0;
985 		left = ret > 0 ? ret : 0;
986 		while (left > 0) {
987 			size_t len, copied;
988 			page_off = off & ~PAGE_MASK;
989 			len = min_t(size_t, left, PAGE_SIZE - page_off);
990 			SetPageUptodate(pages[idx]);
991 			copied = copy_page_to_iter(pages[idx++],
992 						   page_off, len, to);
993 			off += copied;
994 			left -= copied;
995 			if (copied < len) {
996 				ret = -EFAULT;
997 				break;
998 			}
999 		}
1000 		ceph_release_page_vector(pages, num_pages);
1001 
1002 		if (ret < 0) {
1003 			if (ret == -EBLOCKLISTED)
1004 				fsc->blocklisted = true;
1005 			break;
1006 		}
1007 
1008 		if (off >= i_size || !more)
1009 			break;
1010 	}
1011 
1012 	if (off > iocb->ki_pos) {
1013 		if (off >= i_size) {
1014 			*retry_op = CHECK_EOF;
1015 			ret = i_size - iocb->ki_pos;
1016 			iocb->ki_pos = i_size;
1017 		} else {
1018 			ret = off - iocb->ki_pos;
1019 			iocb->ki_pos = off;
1020 		}
1021 	}
1022 
1023 	dout("sync_read result %zd retry_op %d\n", ret, *retry_op);
1024 	return ret;
1025 }
1026 
1027 struct ceph_aio_request {
1028 	struct kiocb *iocb;
1029 	size_t total_len;
1030 	bool write;
1031 	bool should_dirty;
1032 	int error;
1033 	struct list_head osd_reqs;
1034 	unsigned num_reqs;
1035 	atomic_t pending_reqs;
1036 	struct timespec64 mtime;
1037 	struct ceph_cap_flush *prealloc_cf;
1038 };
1039 
1040 struct ceph_aio_work {
1041 	struct work_struct work;
1042 	struct ceph_osd_request *req;
1043 };
1044 
1045 static void ceph_aio_retry_work(struct work_struct *work);
1046 
1047 static void ceph_aio_complete(struct inode *inode,
1048 			      struct ceph_aio_request *aio_req)
1049 {
1050 	struct ceph_inode_info *ci = ceph_inode(inode);
1051 	int ret;
1052 
1053 	if (!atomic_dec_and_test(&aio_req->pending_reqs))
1054 		return;
1055 
1056 	if (aio_req->iocb->ki_flags & IOCB_DIRECT)
1057 		inode_dio_end(inode);
1058 
1059 	ret = aio_req->error;
1060 	if (!ret)
1061 		ret = aio_req->total_len;
1062 
1063 	dout("ceph_aio_complete %p rc %d\n", inode, ret);
1064 
1065 	if (ret >= 0 && aio_req->write) {
1066 		int dirty;
1067 
1068 		loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
1069 		if (endoff > i_size_read(inode)) {
1070 			if (ceph_inode_set_size(inode, endoff))
1071 				ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
1072 		}
1073 
1074 		spin_lock(&ci->i_ceph_lock);
1075 		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1076 					       &aio_req->prealloc_cf);
1077 		spin_unlock(&ci->i_ceph_lock);
1078 		if (dirty)
1079 			__mark_inode_dirty(inode, dirty);
1080 
1081 	}
1082 
1083 	ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
1084 						CEPH_CAP_FILE_RD));
1085 
1086 	aio_req->iocb->ki_complete(aio_req->iocb, ret);
1087 
1088 	ceph_free_cap_flush(aio_req->prealloc_cf);
1089 	kfree(aio_req);
1090 }
1091 
1092 static void ceph_aio_complete_req(struct ceph_osd_request *req)
1093 {
1094 	int rc = req->r_result;
1095 	struct inode *inode = req->r_inode;
1096 	struct ceph_aio_request *aio_req = req->r_priv;
1097 	struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
1098 	struct ceph_client_metric *metric = &ceph_sb_to_mdsc(inode->i_sb)->metric;
1099 	unsigned int len = osd_data->bvec_pos.iter.bi_size;
1100 
1101 	BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_BVECS);
1102 	BUG_ON(!osd_data->num_bvecs);
1103 
1104 	dout("ceph_aio_complete_req %p rc %d bytes %u\n", inode, rc, len);
1105 
1106 	if (rc == -EOLDSNAPC) {
1107 		struct ceph_aio_work *aio_work;
1108 		BUG_ON(!aio_req->write);
1109 
1110 		aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
1111 		if (aio_work) {
1112 			INIT_WORK(&aio_work->work, ceph_aio_retry_work);
1113 			aio_work->req = req;
1114 			queue_work(ceph_inode_to_client(inode)->inode_wq,
1115 				   &aio_work->work);
1116 			return;
1117 		}
1118 		rc = -ENOMEM;
1119 	} else if (!aio_req->write) {
1120 		if (rc == -ENOENT)
1121 			rc = 0;
1122 		if (rc >= 0 && len > rc) {
1123 			struct iov_iter i;
1124 			int zlen = len - rc;
1125 
1126 			/*
1127 			 * If read is satisfied by single OSD request,
1128 			 * it can pass EOF. Otherwise read is within
1129 			 * i_size.
1130 			 */
1131 			if (aio_req->num_reqs == 1) {
1132 				loff_t i_size = i_size_read(inode);
1133 				loff_t endoff = aio_req->iocb->ki_pos + rc;
1134 				if (endoff < i_size)
1135 					zlen = min_t(size_t, zlen,
1136 						     i_size - endoff);
1137 				aio_req->total_len = rc + zlen;
1138 			}
1139 
1140 			iov_iter_bvec(&i, READ, osd_data->bvec_pos.bvecs,
1141 				      osd_data->num_bvecs, len);
1142 			iov_iter_advance(&i, rc);
1143 			iov_iter_zero(zlen, &i);
1144 		}
1145 	}
1146 
1147 	/* r_start_latency == 0 means the request was not submitted */
1148 	if (req->r_start_latency) {
1149 		if (aio_req->write)
1150 			ceph_update_write_metrics(metric, req->r_start_latency,
1151 						  req->r_end_latency, len, rc);
1152 		else
1153 			ceph_update_read_metrics(metric, req->r_start_latency,
1154 						 req->r_end_latency, len, rc);
1155 	}
1156 
1157 	put_bvecs(osd_data->bvec_pos.bvecs, osd_data->num_bvecs,
1158 		  aio_req->should_dirty);
1159 	ceph_osdc_put_request(req);
1160 
1161 	if (rc < 0)
1162 		cmpxchg(&aio_req->error, 0, rc);
1163 
1164 	ceph_aio_complete(inode, aio_req);
1165 	return;
1166 }
1167 
1168 static void ceph_aio_retry_work(struct work_struct *work)
1169 {
1170 	struct ceph_aio_work *aio_work =
1171 		container_of(work, struct ceph_aio_work, work);
1172 	struct ceph_osd_request *orig_req = aio_work->req;
1173 	struct ceph_aio_request *aio_req = orig_req->r_priv;
1174 	struct inode *inode = orig_req->r_inode;
1175 	struct ceph_inode_info *ci = ceph_inode(inode);
1176 	struct ceph_snap_context *snapc;
1177 	struct ceph_osd_request *req;
1178 	int ret;
1179 
1180 	spin_lock(&ci->i_ceph_lock);
1181 	if (__ceph_have_pending_cap_snap(ci)) {
1182 		struct ceph_cap_snap *capsnap =
1183 			list_last_entry(&ci->i_cap_snaps,
1184 					struct ceph_cap_snap,
1185 					ci_item);
1186 		snapc = ceph_get_snap_context(capsnap->context);
1187 	} else {
1188 		BUG_ON(!ci->i_head_snapc);
1189 		snapc = ceph_get_snap_context(ci->i_head_snapc);
1190 	}
1191 	spin_unlock(&ci->i_ceph_lock);
1192 
1193 	req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 1,
1194 			false, GFP_NOFS);
1195 	if (!req) {
1196 		ret = -ENOMEM;
1197 		req = orig_req;
1198 		goto out;
1199 	}
1200 
1201 	req->r_flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1202 	ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
1203 	ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
1204 
1205 	req->r_ops[0] = orig_req->r_ops[0];
1206 
1207 	req->r_mtime = aio_req->mtime;
1208 	req->r_data_offset = req->r_ops[0].extent.offset;
1209 
1210 	ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
1211 	if (ret) {
1212 		ceph_osdc_put_request(req);
1213 		req = orig_req;
1214 		goto out;
1215 	}
1216 
1217 	ceph_osdc_put_request(orig_req);
1218 
1219 	req->r_callback = ceph_aio_complete_req;
1220 	req->r_inode = inode;
1221 	req->r_priv = aio_req;
1222 
1223 	ret = ceph_osdc_start_request(req->r_osdc, req, false);
1224 out:
1225 	if (ret < 0) {
1226 		req->r_result = ret;
1227 		ceph_aio_complete_req(req);
1228 	}
1229 
1230 	ceph_put_snap_context(snapc);
1231 	kfree(aio_work);
1232 }
1233 
1234 static ssize_t
1235 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
1236 		       struct ceph_snap_context *snapc,
1237 		       struct ceph_cap_flush **pcf)
1238 {
1239 	struct file *file = iocb->ki_filp;
1240 	struct inode *inode = file_inode(file);
1241 	struct ceph_inode_info *ci = ceph_inode(inode);
1242 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1243 	struct ceph_client_metric *metric = &fsc->mdsc->metric;
1244 	struct ceph_vino vino;
1245 	struct ceph_osd_request *req;
1246 	struct bio_vec *bvecs;
1247 	struct ceph_aio_request *aio_req = NULL;
1248 	int num_pages = 0;
1249 	int flags;
1250 	int ret = 0;
1251 	struct timespec64 mtime = current_time(inode);
1252 	size_t count = iov_iter_count(iter);
1253 	loff_t pos = iocb->ki_pos;
1254 	bool write = iov_iter_rw(iter) == WRITE;
1255 	bool should_dirty = !write && iter_is_iovec(iter);
1256 
1257 	if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1258 		return -EROFS;
1259 
1260 	dout("sync_direct_%s on file %p %lld~%u snapc %p seq %lld\n",
1261 	     (write ? "write" : "read"), file, pos, (unsigned)count,
1262 	     snapc, snapc ? snapc->seq : 0);
1263 
1264 	if (write) {
1265 		int ret2;
1266 
1267 		ceph_fscache_invalidate(inode, true);
1268 
1269 		ret2 = invalidate_inode_pages2_range(inode->i_mapping,
1270 					pos >> PAGE_SHIFT,
1271 					(pos + count - 1) >> PAGE_SHIFT);
1272 		if (ret2 < 0)
1273 			dout("invalidate_inode_pages2_range returned %d\n", ret2);
1274 
1275 		flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1276 	} else {
1277 		flags = CEPH_OSD_FLAG_READ;
1278 	}
1279 
1280 	while (iov_iter_count(iter) > 0) {
1281 		u64 size = iov_iter_count(iter);
1282 		ssize_t len;
1283 
1284 		if (write)
1285 			size = min_t(u64, size, fsc->mount_options->wsize);
1286 		else
1287 			size = min_t(u64, size, fsc->mount_options->rsize);
1288 
1289 		vino = ceph_vino(inode);
1290 		req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1291 					    vino, pos, &size, 0,
1292 					    1,
1293 					    write ? CEPH_OSD_OP_WRITE :
1294 						    CEPH_OSD_OP_READ,
1295 					    flags, snapc,
1296 					    ci->i_truncate_seq,
1297 					    ci->i_truncate_size,
1298 					    false);
1299 		if (IS_ERR(req)) {
1300 			ret = PTR_ERR(req);
1301 			break;
1302 		}
1303 
1304 		len = iter_get_bvecs_alloc(iter, size, &bvecs, &num_pages);
1305 		if (len < 0) {
1306 			ceph_osdc_put_request(req);
1307 			ret = len;
1308 			break;
1309 		}
1310 		if (len != size)
1311 			osd_req_op_extent_update(req, 0, len);
1312 
1313 		/*
1314 		 * To simplify error handling, allow AIO when IO within i_size
1315 		 * or IO can be satisfied by single OSD request.
1316 		 */
1317 		if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
1318 		    (len == count || pos + count <= i_size_read(inode))) {
1319 			aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
1320 			if (aio_req) {
1321 				aio_req->iocb = iocb;
1322 				aio_req->write = write;
1323 				aio_req->should_dirty = should_dirty;
1324 				INIT_LIST_HEAD(&aio_req->osd_reqs);
1325 				if (write) {
1326 					aio_req->mtime = mtime;
1327 					swap(aio_req->prealloc_cf, *pcf);
1328 				}
1329 			}
1330 			/* ignore error */
1331 		}
1332 
1333 		if (write) {
1334 			/*
1335 			 * throw out any page cache pages in this range. this
1336 			 * may block.
1337 			 */
1338 			truncate_inode_pages_range(inode->i_mapping, pos,
1339 						   PAGE_ALIGN(pos + len) - 1);
1340 
1341 			req->r_mtime = mtime;
1342 		}
1343 
1344 		osd_req_op_extent_osd_data_bvecs(req, 0, bvecs, num_pages, len);
1345 
1346 		if (aio_req) {
1347 			aio_req->total_len += len;
1348 			aio_req->num_reqs++;
1349 			atomic_inc(&aio_req->pending_reqs);
1350 
1351 			req->r_callback = ceph_aio_complete_req;
1352 			req->r_inode = inode;
1353 			req->r_priv = aio_req;
1354 			list_add_tail(&req->r_private_item, &aio_req->osd_reqs);
1355 
1356 			pos += len;
1357 			continue;
1358 		}
1359 
1360 		ret = ceph_osdc_start_request(req->r_osdc, req, false);
1361 		if (!ret)
1362 			ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1363 
1364 		if (write)
1365 			ceph_update_write_metrics(metric, req->r_start_latency,
1366 						  req->r_end_latency, len, ret);
1367 		else
1368 			ceph_update_read_metrics(metric, req->r_start_latency,
1369 						 req->r_end_latency, len, ret);
1370 
1371 		size = i_size_read(inode);
1372 		if (!write) {
1373 			if (ret == -ENOENT)
1374 				ret = 0;
1375 			if (ret >= 0 && ret < len && pos + ret < size) {
1376 				struct iov_iter i;
1377 				int zlen = min_t(size_t, len - ret,
1378 						 size - pos - ret);
1379 
1380 				iov_iter_bvec(&i, READ, bvecs, num_pages, len);
1381 				iov_iter_advance(&i, ret);
1382 				iov_iter_zero(zlen, &i);
1383 				ret += zlen;
1384 			}
1385 			if (ret >= 0)
1386 				len = ret;
1387 		}
1388 
1389 		put_bvecs(bvecs, num_pages, should_dirty);
1390 		ceph_osdc_put_request(req);
1391 		if (ret < 0)
1392 			break;
1393 
1394 		pos += len;
1395 		if (!write && pos >= size)
1396 			break;
1397 
1398 		if (write && pos > size) {
1399 			if (ceph_inode_set_size(inode, pos))
1400 				ceph_check_caps(ceph_inode(inode),
1401 						CHECK_CAPS_AUTHONLY,
1402 						NULL);
1403 		}
1404 	}
1405 
1406 	if (aio_req) {
1407 		LIST_HEAD(osd_reqs);
1408 
1409 		if (aio_req->num_reqs == 0) {
1410 			kfree(aio_req);
1411 			return ret;
1412 		}
1413 
1414 		ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1415 					      CEPH_CAP_FILE_RD);
1416 
1417 		list_splice(&aio_req->osd_reqs, &osd_reqs);
1418 		inode_dio_begin(inode);
1419 		while (!list_empty(&osd_reqs)) {
1420 			req = list_first_entry(&osd_reqs,
1421 					       struct ceph_osd_request,
1422 					       r_private_item);
1423 			list_del_init(&req->r_private_item);
1424 			if (ret >= 0)
1425 				ret = ceph_osdc_start_request(req->r_osdc,
1426 							      req, false);
1427 			if (ret < 0) {
1428 				req->r_result = ret;
1429 				ceph_aio_complete_req(req);
1430 			}
1431 		}
1432 		return -EIOCBQUEUED;
1433 	}
1434 
1435 	if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1436 		ret = pos - iocb->ki_pos;
1437 		iocb->ki_pos = pos;
1438 	}
1439 	return ret;
1440 }
1441 
1442 /*
1443  * Synchronous write, straight from __user pointer or user pages.
1444  *
1445  * If write spans object boundary, just do multiple writes.  (For a
1446  * correct atomic write, we should e.g. take write locks on all
1447  * objects, rollback on failure, etc.)
1448  */
1449 static ssize_t
1450 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1451 		struct ceph_snap_context *snapc)
1452 {
1453 	struct file *file = iocb->ki_filp;
1454 	struct inode *inode = file_inode(file);
1455 	struct ceph_inode_info *ci = ceph_inode(inode);
1456 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1457 	struct ceph_vino vino;
1458 	struct ceph_osd_request *req;
1459 	struct page **pages;
1460 	u64 len;
1461 	int num_pages;
1462 	int written = 0;
1463 	int flags;
1464 	int ret;
1465 	bool check_caps = false;
1466 	struct timespec64 mtime = current_time(inode);
1467 	size_t count = iov_iter_count(from);
1468 
1469 	if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1470 		return -EROFS;
1471 
1472 	dout("sync_write on file %p %lld~%u snapc %p seq %lld\n",
1473 	     file, pos, (unsigned)count, snapc, snapc->seq);
1474 
1475 	ret = filemap_write_and_wait_range(inode->i_mapping,
1476 					   pos, pos + count - 1);
1477 	if (ret < 0)
1478 		return ret;
1479 
1480 	ceph_fscache_invalidate(inode, false);
1481 	ret = invalidate_inode_pages2_range(inode->i_mapping,
1482 					    pos >> PAGE_SHIFT,
1483 					    (pos + count - 1) >> PAGE_SHIFT);
1484 	if (ret < 0)
1485 		dout("invalidate_inode_pages2_range returned %d\n", ret);
1486 
1487 	flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1488 
1489 	while ((len = iov_iter_count(from)) > 0) {
1490 		size_t left;
1491 		int n;
1492 
1493 		vino = ceph_vino(inode);
1494 		req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1495 					    vino, pos, &len, 0, 1,
1496 					    CEPH_OSD_OP_WRITE, flags, snapc,
1497 					    ci->i_truncate_seq,
1498 					    ci->i_truncate_size,
1499 					    false);
1500 		if (IS_ERR(req)) {
1501 			ret = PTR_ERR(req);
1502 			break;
1503 		}
1504 
1505 		/*
1506 		 * write from beginning of first page,
1507 		 * regardless of io alignment
1508 		 */
1509 		num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1510 
1511 		pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1512 		if (IS_ERR(pages)) {
1513 			ret = PTR_ERR(pages);
1514 			goto out;
1515 		}
1516 
1517 		left = len;
1518 		for (n = 0; n < num_pages; n++) {
1519 			size_t plen = min_t(size_t, left, PAGE_SIZE);
1520 			ret = copy_page_from_iter(pages[n], 0, plen, from);
1521 			if (ret != plen) {
1522 				ret = -EFAULT;
1523 				break;
1524 			}
1525 			left -= ret;
1526 		}
1527 
1528 		if (ret < 0) {
1529 			ceph_release_page_vector(pages, num_pages);
1530 			goto out;
1531 		}
1532 
1533 		req->r_inode = inode;
1534 
1535 		osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1536 						false, true);
1537 
1538 		req->r_mtime = mtime;
1539 		ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1540 		if (!ret)
1541 			ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1542 
1543 		ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1544 					  req->r_end_latency, len, ret);
1545 out:
1546 		ceph_osdc_put_request(req);
1547 		if (ret != 0) {
1548 			ceph_set_error_write(ci);
1549 			break;
1550 		}
1551 
1552 		ceph_clear_error_write(ci);
1553 		pos += len;
1554 		written += len;
1555 		if (pos > i_size_read(inode)) {
1556 			check_caps = ceph_inode_set_size(inode, pos);
1557 			if (check_caps)
1558 				ceph_check_caps(ceph_inode(inode),
1559 						CHECK_CAPS_AUTHONLY,
1560 						NULL);
1561 		}
1562 
1563 	}
1564 
1565 	if (ret != -EOLDSNAPC && written > 0) {
1566 		ret = written;
1567 		iocb->ki_pos = pos;
1568 	}
1569 	return ret;
1570 }
1571 
1572 /*
1573  * Wrap generic_file_aio_read with checks for cap bits on the inode.
1574  * Atomically grab references, so that those bits are not released
1575  * back to the MDS mid-read.
1576  *
1577  * Hmm, the sync read case isn't actually async... should it be?
1578  */
1579 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1580 {
1581 	struct file *filp = iocb->ki_filp;
1582 	struct ceph_file_info *fi = filp->private_data;
1583 	size_t len = iov_iter_count(to);
1584 	struct inode *inode = file_inode(filp);
1585 	struct ceph_inode_info *ci = ceph_inode(inode);
1586 	bool direct_lock = iocb->ki_flags & IOCB_DIRECT;
1587 	ssize_t ret;
1588 	int want = 0, got = 0;
1589 	int retry_op = 0, read = 0;
1590 
1591 again:
1592 	dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1593 	     inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1594 
1595 	if (ceph_inode_is_shutdown(inode))
1596 		return -ESTALE;
1597 
1598 	if (direct_lock)
1599 		ceph_start_io_direct(inode);
1600 	else
1601 		ceph_start_io_read(inode);
1602 
1603 	if (!(fi->flags & CEPH_F_SYNC) && !direct_lock)
1604 		want |= CEPH_CAP_FILE_CACHE;
1605 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
1606 		want |= CEPH_CAP_FILE_LAZYIO;
1607 
1608 	ret = ceph_get_caps(filp, CEPH_CAP_FILE_RD, want, -1, &got);
1609 	if (ret < 0) {
1610 		if (direct_lock)
1611 			ceph_end_io_direct(inode);
1612 		else
1613 			ceph_end_io_read(inode);
1614 		return ret;
1615 	}
1616 
1617 	if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1618 	    (iocb->ki_flags & IOCB_DIRECT) ||
1619 	    (fi->flags & CEPH_F_SYNC)) {
1620 
1621 		dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1622 		     inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1623 		     ceph_cap_string(got));
1624 
1625 		if (ci->i_inline_version == CEPH_INLINE_NONE) {
1626 			if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1627 				ret = ceph_direct_read_write(iocb, to,
1628 							     NULL, NULL);
1629 				if (ret >= 0 && ret < len)
1630 					retry_op = CHECK_EOF;
1631 			} else {
1632 				ret = ceph_sync_read(iocb, to, &retry_op);
1633 			}
1634 		} else {
1635 			retry_op = READ_INLINE;
1636 		}
1637 	} else {
1638 		CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1639 		dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1640 		     inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1641 		     ceph_cap_string(got));
1642 		ceph_add_rw_context(fi, &rw_ctx);
1643 		ret = generic_file_read_iter(iocb, to);
1644 		ceph_del_rw_context(fi, &rw_ctx);
1645 	}
1646 
1647 	dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1648 	     inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1649 	ceph_put_cap_refs(ci, got);
1650 
1651 	if (direct_lock)
1652 		ceph_end_io_direct(inode);
1653 	else
1654 		ceph_end_io_read(inode);
1655 
1656 	if (retry_op > HAVE_RETRIED && ret >= 0) {
1657 		int statret;
1658 		struct page *page = NULL;
1659 		loff_t i_size;
1660 		if (retry_op == READ_INLINE) {
1661 			page = __page_cache_alloc(GFP_KERNEL);
1662 			if (!page)
1663 				return -ENOMEM;
1664 		}
1665 
1666 		statret = __ceph_do_getattr(inode, page,
1667 					    CEPH_STAT_CAP_INLINE_DATA, !!page);
1668 		if (statret < 0) {
1669 			if (page)
1670 				__free_page(page);
1671 			if (statret == -ENODATA) {
1672 				BUG_ON(retry_op != READ_INLINE);
1673 				goto again;
1674 			}
1675 			return statret;
1676 		}
1677 
1678 		i_size = i_size_read(inode);
1679 		if (retry_op == READ_INLINE) {
1680 			BUG_ON(ret > 0 || read > 0);
1681 			if (iocb->ki_pos < i_size &&
1682 			    iocb->ki_pos < PAGE_SIZE) {
1683 				loff_t end = min_t(loff_t, i_size,
1684 						   iocb->ki_pos + len);
1685 				end = min_t(loff_t, end, PAGE_SIZE);
1686 				if (statret < end)
1687 					zero_user_segment(page, statret, end);
1688 				ret = copy_page_to_iter(page,
1689 						iocb->ki_pos & ~PAGE_MASK,
1690 						end - iocb->ki_pos, to);
1691 				iocb->ki_pos += ret;
1692 				read += ret;
1693 			}
1694 			if (iocb->ki_pos < i_size && read < len) {
1695 				size_t zlen = min_t(size_t, len - read,
1696 						    i_size - iocb->ki_pos);
1697 				ret = iov_iter_zero(zlen, to);
1698 				iocb->ki_pos += ret;
1699 				read += ret;
1700 			}
1701 			__free_pages(page, 0);
1702 			return read;
1703 		}
1704 
1705 		/* hit EOF or hole? */
1706 		if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1707 		    ret < len) {
1708 			dout("sync_read hit hole, ppos %lld < size %lld"
1709 			     ", reading more\n", iocb->ki_pos, i_size);
1710 
1711 			read += ret;
1712 			len -= ret;
1713 			retry_op = HAVE_RETRIED;
1714 			goto again;
1715 		}
1716 	}
1717 
1718 	if (ret >= 0)
1719 		ret += read;
1720 
1721 	return ret;
1722 }
1723 
1724 /*
1725  * Take cap references to avoid releasing caps to MDS mid-write.
1726  *
1727  * If we are synchronous, and write with an old snap context, the OSD
1728  * may return EOLDSNAPC.  In that case, retry the write.. _after_
1729  * dropping our cap refs and allowing the pending snap to logically
1730  * complete _before_ this write occurs.
1731  *
1732  * If we are near ENOSPC, write synchronously.
1733  */
1734 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1735 {
1736 	struct file *file = iocb->ki_filp;
1737 	struct ceph_file_info *fi = file->private_data;
1738 	struct inode *inode = file_inode(file);
1739 	struct ceph_inode_info *ci = ceph_inode(inode);
1740 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1741 	struct ceph_osd_client *osdc = &fsc->client->osdc;
1742 	struct ceph_cap_flush *prealloc_cf;
1743 	ssize_t count, written = 0;
1744 	int err, want = 0, got;
1745 	bool direct_lock = false;
1746 	u32 map_flags;
1747 	u64 pool_flags;
1748 	loff_t pos;
1749 	loff_t limit = max(i_size_read(inode), fsc->max_file_size);
1750 
1751 	if (ceph_inode_is_shutdown(inode))
1752 		return -ESTALE;
1753 
1754 	if (ceph_snap(inode) != CEPH_NOSNAP)
1755 		return -EROFS;
1756 
1757 	prealloc_cf = ceph_alloc_cap_flush();
1758 	if (!prealloc_cf)
1759 		return -ENOMEM;
1760 
1761 	if ((iocb->ki_flags & (IOCB_DIRECT | IOCB_APPEND)) == IOCB_DIRECT)
1762 		direct_lock = true;
1763 
1764 retry_snap:
1765 	if (direct_lock)
1766 		ceph_start_io_direct(inode);
1767 	else
1768 		ceph_start_io_write(inode);
1769 
1770 	/* We can write back this queue in page reclaim */
1771 	current->backing_dev_info = inode_to_bdi(inode);
1772 
1773 	if (iocb->ki_flags & IOCB_APPEND) {
1774 		err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1775 		if (err < 0)
1776 			goto out;
1777 	}
1778 
1779 	err = generic_write_checks(iocb, from);
1780 	if (err <= 0)
1781 		goto out;
1782 
1783 	pos = iocb->ki_pos;
1784 	if (unlikely(pos >= limit)) {
1785 		err = -EFBIG;
1786 		goto out;
1787 	} else {
1788 		iov_iter_truncate(from, limit - pos);
1789 	}
1790 
1791 	count = iov_iter_count(from);
1792 	if (ceph_quota_is_max_bytes_exceeded(inode, pos + count)) {
1793 		err = -EDQUOT;
1794 		goto out;
1795 	}
1796 
1797 	down_read(&osdc->lock);
1798 	map_flags = osdc->osdmap->flags;
1799 	pool_flags = ceph_pg_pool_flags(osdc->osdmap, ci->i_layout.pool_id);
1800 	up_read(&osdc->lock);
1801 	if ((map_flags & CEPH_OSDMAP_FULL) ||
1802 	    (pool_flags & CEPH_POOL_FLAG_FULL)) {
1803 		err = -ENOSPC;
1804 		goto out;
1805 	}
1806 
1807 	err = file_remove_privs(file);
1808 	if (err)
1809 		goto out;
1810 
1811 	dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1812 	     inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1813 	if (!(fi->flags & CEPH_F_SYNC) && !direct_lock)
1814 		want |= CEPH_CAP_FILE_BUFFER;
1815 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
1816 		want |= CEPH_CAP_FILE_LAZYIO;
1817 	got = 0;
1818 	err = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, pos + count, &got);
1819 	if (err < 0)
1820 		goto out;
1821 
1822 	err = file_update_time(file);
1823 	if (err)
1824 		goto out_caps;
1825 
1826 	inode_inc_iversion_raw(inode);
1827 
1828 	dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1829 	     inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1830 
1831 	if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1832 	    (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
1833 	    (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
1834 		struct ceph_snap_context *snapc;
1835 		struct iov_iter data;
1836 
1837 		spin_lock(&ci->i_ceph_lock);
1838 		if (__ceph_have_pending_cap_snap(ci)) {
1839 			struct ceph_cap_snap *capsnap =
1840 					list_last_entry(&ci->i_cap_snaps,
1841 							struct ceph_cap_snap,
1842 							ci_item);
1843 			snapc = ceph_get_snap_context(capsnap->context);
1844 		} else {
1845 			BUG_ON(!ci->i_head_snapc);
1846 			snapc = ceph_get_snap_context(ci->i_head_snapc);
1847 		}
1848 		spin_unlock(&ci->i_ceph_lock);
1849 
1850 		/* we might need to revert back to that point */
1851 		data = *from;
1852 		if (iocb->ki_flags & IOCB_DIRECT)
1853 			written = ceph_direct_read_write(iocb, &data, snapc,
1854 							 &prealloc_cf);
1855 		else
1856 			written = ceph_sync_write(iocb, &data, pos, snapc);
1857 		if (direct_lock)
1858 			ceph_end_io_direct(inode);
1859 		else
1860 			ceph_end_io_write(inode);
1861 		if (written > 0)
1862 			iov_iter_advance(from, written);
1863 		ceph_put_snap_context(snapc);
1864 	} else {
1865 		/*
1866 		 * No need to acquire the i_truncate_mutex. Because
1867 		 * the MDS revokes Fwb caps before sending truncate
1868 		 * message to us. We can't get Fwb cap while there
1869 		 * are pending vmtruncate. So write and vmtruncate
1870 		 * can not run at the same time
1871 		 */
1872 		written = generic_perform_write(iocb, from);
1873 		if (likely(written >= 0))
1874 			iocb->ki_pos = pos + written;
1875 		ceph_end_io_write(inode);
1876 	}
1877 
1878 	if (written >= 0) {
1879 		int dirty;
1880 
1881 		spin_lock(&ci->i_ceph_lock);
1882 		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1883 					       &prealloc_cf);
1884 		spin_unlock(&ci->i_ceph_lock);
1885 		if (dirty)
1886 			__mark_inode_dirty(inode, dirty);
1887 		if (ceph_quota_is_max_bytes_approaching(inode, iocb->ki_pos))
1888 			ceph_check_caps(ci, 0, NULL);
1889 	}
1890 
1891 	dout("aio_write %p %llx.%llx %llu~%u  dropping cap refs on %s\n",
1892 	     inode, ceph_vinop(inode), pos, (unsigned)count,
1893 	     ceph_cap_string(got));
1894 	ceph_put_cap_refs(ci, got);
1895 
1896 	if (written == -EOLDSNAPC) {
1897 		dout("aio_write %p %llx.%llx %llu~%u" "got EOLDSNAPC, retrying\n",
1898 		     inode, ceph_vinop(inode), pos, (unsigned)count);
1899 		goto retry_snap;
1900 	}
1901 
1902 	if (written >= 0) {
1903 		if ((map_flags & CEPH_OSDMAP_NEARFULL) ||
1904 		    (pool_flags & CEPH_POOL_FLAG_NEARFULL))
1905 			iocb->ki_flags |= IOCB_DSYNC;
1906 		written = generic_write_sync(iocb, written);
1907 	}
1908 
1909 	goto out_unlocked;
1910 out_caps:
1911 	ceph_put_cap_refs(ci, got);
1912 out:
1913 	if (direct_lock)
1914 		ceph_end_io_direct(inode);
1915 	else
1916 		ceph_end_io_write(inode);
1917 out_unlocked:
1918 	ceph_free_cap_flush(prealloc_cf);
1919 	current->backing_dev_info = NULL;
1920 	return written ? written : err;
1921 }
1922 
1923 /*
1924  * llseek.  be sure to verify file size on SEEK_END.
1925  */
1926 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1927 {
1928 	struct inode *inode = file->f_mapping->host;
1929 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1930 	loff_t i_size;
1931 	loff_t ret;
1932 
1933 	inode_lock(inode);
1934 
1935 	if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1936 		ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1937 		if (ret < 0)
1938 			goto out;
1939 	}
1940 
1941 	i_size = i_size_read(inode);
1942 	switch (whence) {
1943 	case SEEK_END:
1944 		offset += i_size;
1945 		break;
1946 	case SEEK_CUR:
1947 		/*
1948 		 * Here we special-case the lseek(fd, 0, SEEK_CUR)
1949 		 * position-querying operation.  Avoid rewriting the "same"
1950 		 * f_pos value back to the file because a concurrent read(),
1951 		 * write() or lseek() might have altered it
1952 		 */
1953 		if (offset == 0) {
1954 			ret = file->f_pos;
1955 			goto out;
1956 		}
1957 		offset += file->f_pos;
1958 		break;
1959 	case SEEK_DATA:
1960 		if (offset < 0 || offset >= i_size) {
1961 			ret = -ENXIO;
1962 			goto out;
1963 		}
1964 		break;
1965 	case SEEK_HOLE:
1966 		if (offset < 0 || offset >= i_size) {
1967 			ret = -ENXIO;
1968 			goto out;
1969 		}
1970 		offset = i_size;
1971 		break;
1972 	}
1973 
1974 	ret = vfs_setpos(file, offset, max(i_size, fsc->max_file_size));
1975 
1976 out:
1977 	inode_unlock(inode);
1978 	return ret;
1979 }
1980 
1981 static inline void ceph_zero_partial_page(
1982 	struct inode *inode, loff_t offset, unsigned size)
1983 {
1984 	struct page *page;
1985 	pgoff_t index = offset >> PAGE_SHIFT;
1986 
1987 	page = find_lock_page(inode->i_mapping, index);
1988 	if (page) {
1989 		wait_on_page_writeback(page);
1990 		zero_user(page, offset & (PAGE_SIZE - 1), size);
1991 		unlock_page(page);
1992 		put_page(page);
1993 	}
1994 }
1995 
1996 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1997 				      loff_t length)
1998 {
1999 	loff_t nearly = round_up(offset, PAGE_SIZE);
2000 	if (offset < nearly) {
2001 		loff_t size = nearly - offset;
2002 		if (length < size)
2003 			size = length;
2004 		ceph_zero_partial_page(inode, offset, size);
2005 		offset += size;
2006 		length -= size;
2007 	}
2008 	if (length >= PAGE_SIZE) {
2009 		loff_t size = round_down(length, PAGE_SIZE);
2010 		truncate_pagecache_range(inode, offset, offset + size - 1);
2011 		offset += size;
2012 		length -= size;
2013 	}
2014 	if (length)
2015 		ceph_zero_partial_page(inode, offset, length);
2016 }
2017 
2018 static int ceph_zero_partial_object(struct inode *inode,
2019 				    loff_t offset, loff_t *length)
2020 {
2021 	struct ceph_inode_info *ci = ceph_inode(inode);
2022 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
2023 	struct ceph_osd_request *req;
2024 	int ret = 0;
2025 	loff_t zero = 0;
2026 	int op;
2027 
2028 	if (!length) {
2029 		op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
2030 		length = &zero;
2031 	} else {
2032 		op = CEPH_OSD_OP_ZERO;
2033 	}
2034 
2035 	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
2036 					ceph_vino(inode),
2037 					offset, length,
2038 					0, 1, op,
2039 					CEPH_OSD_FLAG_WRITE,
2040 					NULL, 0, 0, false);
2041 	if (IS_ERR(req)) {
2042 		ret = PTR_ERR(req);
2043 		goto out;
2044 	}
2045 
2046 	req->r_mtime = inode->i_mtime;
2047 	ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
2048 	if (!ret) {
2049 		ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
2050 		if (ret == -ENOENT)
2051 			ret = 0;
2052 	}
2053 	ceph_osdc_put_request(req);
2054 
2055 out:
2056 	return ret;
2057 }
2058 
2059 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
2060 {
2061 	int ret = 0;
2062 	struct ceph_inode_info *ci = ceph_inode(inode);
2063 	s32 stripe_unit = ci->i_layout.stripe_unit;
2064 	s32 stripe_count = ci->i_layout.stripe_count;
2065 	s32 object_size = ci->i_layout.object_size;
2066 	u64 object_set_size = object_size * stripe_count;
2067 	u64 nearly, t;
2068 
2069 	/* round offset up to next period boundary */
2070 	nearly = offset + object_set_size - 1;
2071 	t = nearly;
2072 	nearly -= do_div(t, object_set_size);
2073 
2074 	while (length && offset < nearly) {
2075 		loff_t size = length;
2076 		ret = ceph_zero_partial_object(inode, offset, &size);
2077 		if (ret < 0)
2078 			return ret;
2079 		offset += size;
2080 		length -= size;
2081 	}
2082 	while (length >= object_set_size) {
2083 		int i;
2084 		loff_t pos = offset;
2085 		for (i = 0; i < stripe_count; ++i) {
2086 			ret = ceph_zero_partial_object(inode, pos, NULL);
2087 			if (ret < 0)
2088 				return ret;
2089 			pos += stripe_unit;
2090 		}
2091 		offset += object_set_size;
2092 		length -= object_set_size;
2093 	}
2094 	while (length) {
2095 		loff_t size = length;
2096 		ret = ceph_zero_partial_object(inode, offset, &size);
2097 		if (ret < 0)
2098 			return ret;
2099 		offset += size;
2100 		length -= size;
2101 	}
2102 	return ret;
2103 }
2104 
2105 static long ceph_fallocate(struct file *file, int mode,
2106 				loff_t offset, loff_t length)
2107 {
2108 	struct ceph_file_info *fi = file->private_data;
2109 	struct inode *inode = file_inode(file);
2110 	struct ceph_inode_info *ci = ceph_inode(inode);
2111 	struct ceph_cap_flush *prealloc_cf;
2112 	int want, got = 0;
2113 	int dirty;
2114 	int ret = 0;
2115 	loff_t endoff = 0;
2116 	loff_t size;
2117 
2118 	if (mode != (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2119 		return -EOPNOTSUPP;
2120 
2121 	if (!S_ISREG(inode->i_mode))
2122 		return -EOPNOTSUPP;
2123 
2124 	prealloc_cf = ceph_alloc_cap_flush();
2125 	if (!prealloc_cf)
2126 		return -ENOMEM;
2127 
2128 	inode_lock(inode);
2129 
2130 	if (ceph_snap(inode) != CEPH_NOSNAP) {
2131 		ret = -EROFS;
2132 		goto unlock;
2133 	}
2134 
2135 	size = i_size_read(inode);
2136 
2137 	/* Are we punching a hole beyond EOF? */
2138 	if (offset >= size)
2139 		goto unlock;
2140 	if ((offset + length) > size)
2141 		length = size - offset;
2142 
2143 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
2144 		want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
2145 	else
2146 		want = CEPH_CAP_FILE_BUFFER;
2147 
2148 	ret = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, endoff, &got);
2149 	if (ret < 0)
2150 		goto unlock;
2151 
2152 	filemap_invalidate_lock(inode->i_mapping);
2153 	ceph_fscache_invalidate(inode, false);
2154 	ceph_zero_pagecache_range(inode, offset, length);
2155 	ret = ceph_zero_objects(inode, offset, length);
2156 
2157 	if (!ret) {
2158 		spin_lock(&ci->i_ceph_lock);
2159 		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
2160 					       &prealloc_cf);
2161 		spin_unlock(&ci->i_ceph_lock);
2162 		if (dirty)
2163 			__mark_inode_dirty(inode, dirty);
2164 	}
2165 	filemap_invalidate_unlock(inode->i_mapping);
2166 
2167 	ceph_put_cap_refs(ci, got);
2168 unlock:
2169 	inode_unlock(inode);
2170 	ceph_free_cap_flush(prealloc_cf);
2171 	return ret;
2172 }
2173 
2174 /*
2175  * This function tries to get FILE_WR capabilities for dst_ci and FILE_RD for
2176  * src_ci.  Two attempts are made to obtain both caps, and an error is return if
2177  * this fails; zero is returned on success.
2178  */
2179 static int get_rd_wr_caps(struct file *src_filp, int *src_got,
2180 			  struct file *dst_filp,
2181 			  loff_t dst_endoff, int *dst_got)
2182 {
2183 	int ret = 0;
2184 	bool retrying = false;
2185 
2186 retry_caps:
2187 	ret = ceph_get_caps(dst_filp, CEPH_CAP_FILE_WR, CEPH_CAP_FILE_BUFFER,
2188 			    dst_endoff, dst_got);
2189 	if (ret < 0)
2190 		return ret;
2191 
2192 	/*
2193 	 * Since we're already holding the FILE_WR capability for the dst file,
2194 	 * we would risk a deadlock by using ceph_get_caps.  Thus, we'll do some
2195 	 * retry dance instead to try to get both capabilities.
2196 	 */
2197 	ret = ceph_try_get_caps(file_inode(src_filp),
2198 				CEPH_CAP_FILE_RD, CEPH_CAP_FILE_SHARED,
2199 				false, src_got);
2200 	if (ret <= 0) {
2201 		/* Start by dropping dst_ci caps and getting src_ci caps */
2202 		ceph_put_cap_refs(ceph_inode(file_inode(dst_filp)), *dst_got);
2203 		if (retrying) {
2204 			if (!ret)
2205 				/* ceph_try_get_caps masks EAGAIN */
2206 				ret = -EAGAIN;
2207 			return ret;
2208 		}
2209 		ret = ceph_get_caps(src_filp, CEPH_CAP_FILE_RD,
2210 				    CEPH_CAP_FILE_SHARED, -1, src_got);
2211 		if (ret < 0)
2212 			return ret;
2213 		/*... drop src_ci caps too, and retry */
2214 		ceph_put_cap_refs(ceph_inode(file_inode(src_filp)), *src_got);
2215 		retrying = true;
2216 		goto retry_caps;
2217 	}
2218 	return ret;
2219 }
2220 
2221 static void put_rd_wr_caps(struct ceph_inode_info *src_ci, int src_got,
2222 			   struct ceph_inode_info *dst_ci, int dst_got)
2223 {
2224 	ceph_put_cap_refs(src_ci, src_got);
2225 	ceph_put_cap_refs(dst_ci, dst_got);
2226 }
2227 
2228 /*
2229  * This function does several size-related checks, returning an error if:
2230  *  - source file is smaller than off+len
2231  *  - destination file size is not OK (inode_newsize_ok())
2232  *  - max bytes quotas is exceeded
2233  */
2234 static int is_file_size_ok(struct inode *src_inode, struct inode *dst_inode,
2235 			   loff_t src_off, loff_t dst_off, size_t len)
2236 {
2237 	loff_t size, endoff;
2238 
2239 	size = i_size_read(src_inode);
2240 	/*
2241 	 * Don't copy beyond source file EOF.  Instead of simply setting length
2242 	 * to (size - src_off), just drop to VFS default implementation, as the
2243 	 * local i_size may be stale due to other clients writing to the source
2244 	 * inode.
2245 	 */
2246 	if (src_off + len > size) {
2247 		dout("Copy beyond EOF (%llu + %zu > %llu)\n",
2248 		     src_off, len, size);
2249 		return -EOPNOTSUPP;
2250 	}
2251 	size = i_size_read(dst_inode);
2252 
2253 	endoff = dst_off + len;
2254 	if (inode_newsize_ok(dst_inode, endoff))
2255 		return -EOPNOTSUPP;
2256 
2257 	if (ceph_quota_is_max_bytes_exceeded(dst_inode, endoff))
2258 		return -EDQUOT;
2259 
2260 	return 0;
2261 }
2262 
2263 static struct ceph_osd_request *
2264 ceph_alloc_copyfrom_request(struct ceph_osd_client *osdc,
2265 			    u64 src_snapid,
2266 			    struct ceph_object_id *src_oid,
2267 			    struct ceph_object_locator *src_oloc,
2268 			    struct ceph_object_id *dst_oid,
2269 			    struct ceph_object_locator *dst_oloc,
2270 			    u32 truncate_seq, u64 truncate_size)
2271 {
2272 	struct ceph_osd_request *req;
2273 	int ret;
2274 	u32 src_fadvise_flags =
2275 		CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2276 		CEPH_OSD_OP_FLAG_FADVISE_NOCACHE;
2277 	u32 dst_fadvise_flags =
2278 		CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2279 		CEPH_OSD_OP_FLAG_FADVISE_DONTNEED;
2280 
2281 	req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_KERNEL);
2282 	if (!req)
2283 		return ERR_PTR(-ENOMEM);
2284 
2285 	req->r_flags = CEPH_OSD_FLAG_WRITE;
2286 
2287 	ceph_oloc_copy(&req->r_t.base_oloc, dst_oloc);
2288 	ceph_oid_copy(&req->r_t.base_oid, dst_oid);
2289 
2290 	ret = osd_req_op_copy_from_init(req, src_snapid, 0,
2291 					src_oid, src_oloc,
2292 					src_fadvise_flags,
2293 					dst_fadvise_flags,
2294 					truncate_seq,
2295 					truncate_size,
2296 					CEPH_OSD_COPY_FROM_FLAG_TRUNCATE_SEQ);
2297 	if (ret)
2298 		goto out;
2299 
2300 	ret = ceph_osdc_alloc_messages(req, GFP_KERNEL);
2301 	if (ret)
2302 		goto out;
2303 
2304 	return req;
2305 
2306 out:
2307 	ceph_osdc_put_request(req);
2308 	return ERR_PTR(ret);
2309 }
2310 
2311 static ssize_t ceph_do_objects_copy(struct ceph_inode_info *src_ci, u64 *src_off,
2312 				    struct ceph_inode_info *dst_ci, u64 *dst_off,
2313 				    struct ceph_fs_client *fsc,
2314 				    size_t len, unsigned int flags)
2315 {
2316 	struct ceph_object_locator src_oloc, dst_oloc;
2317 	struct ceph_object_id src_oid, dst_oid;
2318 	struct ceph_osd_client *osdc;
2319 	struct ceph_osd_request *req;
2320 	size_t bytes = 0;
2321 	u64 src_objnum, src_objoff, dst_objnum, dst_objoff;
2322 	u32 src_objlen, dst_objlen;
2323 	u32 object_size = src_ci->i_layout.object_size;
2324 	int ret;
2325 
2326 	src_oloc.pool = src_ci->i_layout.pool_id;
2327 	src_oloc.pool_ns = ceph_try_get_string(src_ci->i_layout.pool_ns);
2328 	dst_oloc.pool = dst_ci->i_layout.pool_id;
2329 	dst_oloc.pool_ns = ceph_try_get_string(dst_ci->i_layout.pool_ns);
2330 	osdc = &fsc->client->osdc;
2331 
2332 	while (len >= object_size) {
2333 		ceph_calc_file_object_mapping(&src_ci->i_layout, *src_off,
2334 					      object_size, &src_objnum,
2335 					      &src_objoff, &src_objlen);
2336 		ceph_calc_file_object_mapping(&dst_ci->i_layout, *dst_off,
2337 					      object_size, &dst_objnum,
2338 					      &dst_objoff, &dst_objlen);
2339 		ceph_oid_init(&src_oid);
2340 		ceph_oid_printf(&src_oid, "%llx.%08llx",
2341 				src_ci->i_vino.ino, src_objnum);
2342 		ceph_oid_init(&dst_oid);
2343 		ceph_oid_printf(&dst_oid, "%llx.%08llx",
2344 				dst_ci->i_vino.ino, dst_objnum);
2345 		/* Do an object remote copy */
2346 		req = ceph_alloc_copyfrom_request(osdc, src_ci->i_vino.snap,
2347 						  &src_oid, &src_oloc,
2348 						  &dst_oid, &dst_oloc,
2349 						  dst_ci->i_truncate_seq,
2350 						  dst_ci->i_truncate_size);
2351 		if (IS_ERR(req))
2352 			ret = PTR_ERR(req);
2353 		else {
2354 			ceph_osdc_start_request(osdc, req, false);
2355 			ret = ceph_osdc_wait_request(osdc, req);
2356 			ceph_update_copyfrom_metrics(&fsc->mdsc->metric,
2357 						     req->r_start_latency,
2358 						     req->r_end_latency,
2359 						     object_size, ret);
2360 			ceph_osdc_put_request(req);
2361 		}
2362 		if (ret) {
2363 			if (ret == -EOPNOTSUPP) {
2364 				fsc->have_copy_from2 = false;
2365 				pr_notice("OSDs don't support copy-from2; disabling copy offload\n");
2366 			}
2367 			dout("ceph_osdc_copy_from returned %d\n", ret);
2368 			if (!bytes)
2369 				bytes = ret;
2370 			goto out;
2371 		}
2372 		len -= object_size;
2373 		bytes += object_size;
2374 		*src_off += object_size;
2375 		*dst_off += object_size;
2376 	}
2377 
2378 out:
2379 	ceph_oloc_destroy(&src_oloc);
2380 	ceph_oloc_destroy(&dst_oloc);
2381 	return bytes;
2382 }
2383 
2384 static ssize_t __ceph_copy_file_range(struct file *src_file, loff_t src_off,
2385 				      struct file *dst_file, loff_t dst_off,
2386 				      size_t len, unsigned int flags)
2387 {
2388 	struct inode *src_inode = file_inode(src_file);
2389 	struct inode *dst_inode = file_inode(dst_file);
2390 	struct ceph_inode_info *src_ci = ceph_inode(src_inode);
2391 	struct ceph_inode_info *dst_ci = ceph_inode(dst_inode);
2392 	struct ceph_cap_flush *prealloc_cf;
2393 	struct ceph_fs_client *src_fsc = ceph_inode_to_client(src_inode);
2394 	loff_t size;
2395 	ssize_t ret = -EIO, bytes;
2396 	u64 src_objnum, dst_objnum, src_objoff, dst_objoff;
2397 	u32 src_objlen, dst_objlen;
2398 	int src_got = 0, dst_got = 0, err, dirty;
2399 
2400 	if (src_inode->i_sb != dst_inode->i_sb) {
2401 		struct ceph_fs_client *dst_fsc = ceph_inode_to_client(dst_inode);
2402 
2403 		if (ceph_fsid_compare(&src_fsc->client->fsid,
2404 				      &dst_fsc->client->fsid)) {
2405 			dout("Copying files across clusters: src: %pU dst: %pU\n",
2406 			     &src_fsc->client->fsid, &dst_fsc->client->fsid);
2407 			return -EXDEV;
2408 		}
2409 	}
2410 	if (ceph_snap(dst_inode) != CEPH_NOSNAP)
2411 		return -EROFS;
2412 
2413 	/*
2414 	 * Some of the checks below will return -EOPNOTSUPP, which will force a
2415 	 * fallback to the default VFS copy_file_range implementation.  This is
2416 	 * desirable in several cases (for ex, the 'len' is smaller than the
2417 	 * size of the objects, or in cases where that would be more
2418 	 * efficient).
2419 	 */
2420 
2421 	if (ceph_test_mount_opt(src_fsc, NOCOPYFROM))
2422 		return -EOPNOTSUPP;
2423 
2424 	if (!src_fsc->have_copy_from2)
2425 		return -EOPNOTSUPP;
2426 
2427 	/*
2428 	 * Striped file layouts require that we copy partial objects, but the
2429 	 * OSD copy-from operation only supports full-object copies.  Limit
2430 	 * this to non-striped file layouts for now.
2431 	 */
2432 	if ((src_ci->i_layout.stripe_unit != dst_ci->i_layout.stripe_unit) ||
2433 	    (src_ci->i_layout.stripe_count != 1) ||
2434 	    (dst_ci->i_layout.stripe_count != 1) ||
2435 	    (src_ci->i_layout.object_size != dst_ci->i_layout.object_size)) {
2436 		dout("Invalid src/dst files layout\n");
2437 		return -EOPNOTSUPP;
2438 	}
2439 
2440 	if (len < src_ci->i_layout.object_size)
2441 		return -EOPNOTSUPP; /* no remote copy will be done */
2442 
2443 	prealloc_cf = ceph_alloc_cap_flush();
2444 	if (!prealloc_cf)
2445 		return -ENOMEM;
2446 
2447 	/* Start by sync'ing the source and destination files */
2448 	ret = file_write_and_wait_range(src_file, src_off, (src_off + len));
2449 	if (ret < 0) {
2450 		dout("failed to write src file (%zd)\n", ret);
2451 		goto out;
2452 	}
2453 	ret = file_write_and_wait_range(dst_file, dst_off, (dst_off + len));
2454 	if (ret < 0) {
2455 		dout("failed to write dst file (%zd)\n", ret);
2456 		goto out;
2457 	}
2458 
2459 	/*
2460 	 * We need FILE_WR caps for dst_ci and FILE_RD for src_ci as other
2461 	 * clients may have dirty data in their caches.  And OSDs know nothing
2462 	 * about caps, so they can't safely do the remote object copies.
2463 	 */
2464 	err = get_rd_wr_caps(src_file, &src_got,
2465 			     dst_file, (dst_off + len), &dst_got);
2466 	if (err < 0) {
2467 		dout("get_rd_wr_caps returned %d\n", err);
2468 		ret = -EOPNOTSUPP;
2469 		goto out;
2470 	}
2471 
2472 	ret = is_file_size_ok(src_inode, dst_inode, src_off, dst_off, len);
2473 	if (ret < 0)
2474 		goto out_caps;
2475 
2476 	/* Drop dst file cached pages */
2477 	ceph_fscache_invalidate(dst_inode, false);
2478 	ret = invalidate_inode_pages2_range(dst_inode->i_mapping,
2479 					    dst_off >> PAGE_SHIFT,
2480 					    (dst_off + len) >> PAGE_SHIFT);
2481 	if (ret < 0) {
2482 		dout("Failed to invalidate inode pages (%zd)\n", ret);
2483 		ret = 0; /* XXX */
2484 	}
2485 	ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
2486 				      src_ci->i_layout.object_size,
2487 				      &src_objnum, &src_objoff, &src_objlen);
2488 	ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
2489 				      dst_ci->i_layout.object_size,
2490 				      &dst_objnum, &dst_objoff, &dst_objlen);
2491 	/* object-level offsets need to the same */
2492 	if (src_objoff != dst_objoff) {
2493 		ret = -EOPNOTSUPP;
2494 		goto out_caps;
2495 	}
2496 
2497 	/*
2498 	 * Do a manual copy if the object offset isn't object aligned.
2499 	 * 'src_objlen' contains the bytes left until the end of the object,
2500 	 * starting at the src_off
2501 	 */
2502 	if (src_objoff) {
2503 		dout("Initial partial copy of %u bytes\n", src_objlen);
2504 
2505 		/*
2506 		 * we need to temporarily drop all caps as we'll be calling
2507 		 * {read,write}_iter, which will get caps again.
2508 		 */
2509 		put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2510 		ret = do_splice_direct(src_file, &src_off, dst_file,
2511 				       &dst_off, src_objlen, flags);
2512 		/* Abort on short copies or on error */
2513 		if (ret < src_objlen) {
2514 			dout("Failed partial copy (%zd)\n", ret);
2515 			goto out;
2516 		}
2517 		len -= ret;
2518 		err = get_rd_wr_caps(src_file, &src_got,
2519 				     dst_file, (dst_off + len), &dst_got);
2520 		if (err < 0)
2521 			goto out;
2522 		err = is_file_size_ok(src_inode, dst_inode,
2523 				      src_off, dst_off, len);
2524 		if (err < 0)
2525 			goto out_caps;
2526 	}
2527 
2528 	size = i_size_read(dst_inode);
2529 	bytes = ceph_do_objects_copy(src_ci, &src_off, dst_ci, &dst_off,
2530 				     src_fsc, len, flags);
2531 	if (bytes <= 0) {
2532 		if (!ret)
2533 			ret = bytes;
2534 		goto out_caps;
2535 	}
2536 	dout("Copied %zu bytes out of %zu\n", bytes, len);
2537 	len -= bytes;
2538 	ret += bytes;
2539 
2540 	file_update_time(dst_file);
2541 	inode_inc_iversion_raw(dst_inode);
2542 
2543 	if (dst_off > size) {
2544 		/* Let the MDS know about dst file size change */
2545 		if (ceph_inode_set_size(dst_inode, dst_off) ||
2546 		    ceph_quota_is_max_bytes_approaching(dst_inode, dst_off))
2547 			ceph_check_caps(dst_ci, CHECK_CAPS_AUTHONLY, NULL);
2548 	}
2549 	/* Mark Fw dirty */
2550 	spin_lock(&dst_ci->i_ceph_lock);
2551 	dirty = __ceph_mark_dirty_caps(dst_ci, CEPH_CAP_FILE_WR, &prealloc_cf);
2552 	spin_unlock(&dst_ci->i_ceph_lock);
2553 	if (dirty)
2554 		__mark_inode_dirty(dst_inode, dirty);
2555 
2556 out_caps:
2557 	put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2558 
2559 	/*
2560 	 * Do the final manual copy if we still have some bytes left, unless
2561 	 * there were errors in remote object copies (len >= object_size).
2562 	 */
2563 	if (len && (len < src_ci->i_layout.object_size)) {
2564 		dout("Final partial copy of %zu bytes\n", len);
2565 		bytes = do_splice_direct(src_file, &src_off, dst_file,
2566 					 &dst_off, len, flags);
2567 		if (bytes > 0)
2568 			ret += bytes;
2569 		else
2570 			dout("Failed partial copy (%zd)\n", bytes);
2571 	}
2572 
2573 out:
2574 	ceph_free_cap_flush(prealloc_cf);
2575 
2576 	return ret;
2577 }
2578 
2579 static ssize_t ceph_copy_file_range(struct file *src_file, loff_t src_off,
2580 				    struct file *dst_file, loff_t dst_off,
2581 				    size_t len, unsigned int flags)
2582 {
2583 	ssize_t ret;
2584 
2585 	ret = __ceph_copy_file_range(src_file, src_off, dst_file, dst_off,
2586 				     len, flags);
2587 
2588 	if (ret == -EOPNOTSUPP || ret == -EXDEV)
2589 		ret = generic_copy_file_range(src_file, src_off, dst_file,
2590 					      dst_off, len, flags);
2591 	return ret;
2592 }
2593 
2594 const struct file_operations ceph_file_fops = {
2595 	.open = ceph_open,
2596 	.release = ceph_release,
2597 	.llseek = ceph_llseek,
2598 	.read_iter = ceph_read_iter,
2599 	.write_iter = ceph_write_iter,
2600 	.mmap = ceph_mmap,
2601 	.fsync = ceph_fsync,
2602 	.lock = ceph_lock,
2603 	.setlease = simple_nosetlease,
2604 	.flock = ceph_flock,
2605 	.splice_read = generic_file_splice_read,
2606 	.splice_write = iter_file_splice_write,
2607 	.unlocked_ioctl = ceph_ioctl,
2608 	.compat_ioctl = compat_ptr_ioctl,
2609 	.fallocate	= ceph_fallocate,
2610 	.copy_file_range = ceph_copy_file_range,
2611 };
2612