xref: /linux/fs/ceph/file.c (revision 4b132aacb0768ac1e652cf517097ea6f237214b9)
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 #include <linux/splice.h>
16 
17 #include "super.h"
18 #include "mds_client.h"
19 #include "cache.h"
20 #include "io.h"
21 #include "metric.h"
22 
23 static __le32 ceph_flags_sys2wire(struct ceph_mds_client *mdsc, u32 flags)
24 {
25 	struct ceph_client *cl = mdsc->fsc->client;
26 	u32 wire_flags = 0;
27 
28 	switch (flags & O_ACCMODE) {
29 	case O_RDONLY:
30 		wire_flags |= CEPH_O_RDONLY;
31 		break;
32 	case O_WRONLY:
33 		wire_flags |= CEPH_O_WRONLY;
34 		break;
35 	case O_RDWR:
36 		wire_flags |= CEPH_O_RDWR;
37 		break;
38 	}
39 
40 	flags &= ~O_ACCMODE;
41 
42 #define ceph_sys2wire(a) if (flags & a) { wire_flags |= CEPH_##a; flags &= ~a; }
43 
44 	ceph_sys2wire(O_CREAT);
45 	ceph_sys2wire(O_EXCL);
46 	ceph_sys2wire(O_TRUNC);
47 	ceph_sys2wire(O_DIRECTORY);
48 	ceph_sys2wire(O_NOFOLLOW);
49 
50 #undef ceph_sys2wire
51 
52 	if (flags)
53 		doutc(cl, "unused open flags: %x\n", flags);
54 
55 	return cpu_to_le32(wire_flags);
56 }
57 
58 /*
59  * Ceph file operations
60  *
61  * Implement basic open/close functionality, and implement
62  * read/write.
63  *
64  * We implement three modes of file I/O:
65  *  - buffered uses the generic_file_aio_{read,write} helpers
66  *
67  *  - synchronous is used when there is multi-client read/write
68  *    sharing, avoids the page cache, and synchronously waits for an
69  *    ack from the OSD.
70  *
71  *  - direct io takes the variant of the sync path that references
72  *    user pages directly.
73  *
74  * fsync() flushes and waits on dirty pages, but just queues metadata
75  * for writeback: since the MDS can recover size and mtime there is no
76  * need to wait for MDS acknowledgement.
77  */
78 
79 /*
80  * How many pages to get in one call to iov_iter_get_pages().  This
81  * determines the size of the on-stack array used as a buffer.
82  */
83 #define ITER_GET_BVECS_PAGES	64
84 
85 static ssize_t __iter_get_bvecs(struct iov_iter *iter, size_t maxsize,
86 				struct bio_vec *bvecs)
87 {
88 	size_t size = 0;
89 	int bvec_idx = 0;
90 
91 	if (maxsize > iov_iter_count(iter))
92 		maxsize = iov_iter_count(iter);
93 
94 	while (size < maxsize) {
95 		struct page *pages[ITER_GET_BVECS_PAGES];
96 		ssize_t bytes;
97 		size_t start;
98 		int idx = 0;
99 
100 		bytes = iov_iter_get_pages2(iter, pages, maxsize - size,
101 					   ITER_GET_BVECS_PAGES, &start);
102 		if (bytes < 0)
103 			return size ?: bytes;
104 
105 		size += bytes;
106 
107 		for ( ; bytes; idx++, bvec_idx++) {
108 			int len = min_t(int, bytes, PAGE_SIZE - start);
109 
110 			bvec_set_page(&bvecs[bvec_idx], pages[idx], len, start);
111 			bytes -= len;
112 			start = 0;
113 		}
114 	}
115 
116 	return size;
117 }
118 
119 /*
120  * iov_iter_get_pages() only considers one iov_iter segment, no matter
121  * what maxsize or maxpages are given.  For ITER_BVEC that is a single
122  * page.
123  *
124  * Attempt to get up to @maxsize bytes worth of pages from @iter.
125  * Return the number of bytes in the created bio_vec array, or an error.
126  */
127 static ssize_t iter_get_bvecs_alloc(struct iov_iter *iter, size_t maxsize,
128 				    struct bio_vec **bvecs, int *num_bvecs)
129 {
130 	struct bio_vec *bv;
131 	size_t orig_count = iov_iter_count(iter);
132 	ssize_t bytes;
133 	int npages;
134 
135 	iov_iter_truncate(iter, maxsize);
136 	npages = iov_iter_npages(iter, INT_MAX);
137 	iov_iter_reexpand(iter, orig_count);
138 
139 	/*
140 	 * __iter_get_bvecs() may populate only part of the array -- zero it
141 	 * out.
142 	 */
143 	bv = kvmalloc_array(npages, sizeof(*bv), GFP_KERNEL | __GFP_ZERO);
144 	if (!bv)
145 		return -ENOMEM;
146 
147 	bytes = __iter_get_bvecs(iter, maxsize, bv);
148 	if (bytes < 0) {
149 		/*
150 		 * No pages were pinned -- just free the array.
151 		 */
152 		kvfree(bv);
153 		return bytes;
154 	}
155 
156 	*bvecs = bv;
157 	*num_bvecs = npages;
158 	return bytes;
159 }
160 
161 static void put_bvecs(struct bio_vec *bvecs, int num_bvecs, bool should_dirty)
162 {
163 	int i;
164 
165 	for (i = 0; i < num_bvecs; i++) {
166 		if (bvecs[i].bv_page) {
167 			if (should_dirty)
168 				set_page_dirty_lock(bvecs[i].bv_page);
169 			put_page(bvecs[i].bv_page);
170 		}
171 	}
172 	kvfree(bvecs);
173 }
174 
175 /*
176  * Prepare an open request.  Preallocate ceph_cap to avoid an
177  * inopportune ENOMEM later.
178  */
179 static struct ceph_mds_request *
180 prepare_open_request(struct super_block *sb, int flags, int create_mode)
181 {
182 	struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(sb);
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(mdsc, 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_inode_info *ci = ceph_inode(inode);
204 	struct ceph_mount_options *opt =
205 		ceph_inode_to_fs_client(&ci->netfs.inode)->mount_options;
206 	struct ceph_client *cl = ceph_inode_to_client(inode);
207 	struct ceph_file_info *fi;
208 	int ret;
209 
210 	doutc(cl, "%p %llx.%llx %p 0%o (%s)\n", inode, ceph_vinop(inode),
211 	      file, inode->i_mode, isdir ? "dir" : "regular");
212 	BUG_ON(inode->i_fop->release != ceph_release);
213 
214 	if (isdir) {
215 		struct ceph_dir_file_info *dfi =
216 			kmem_cache_zalloc(ceph_dir_file_cachep, GFP_KERNEL);
217 		if (!dfi)
218 			return -ENOMEM;
219 
220 		file->private_data = dfi;
221 		fi = &dfi->file_info;
222 		dfi->next_offset = 2;
223 		dfi->readdir_cache_idx = -1;
224 	} else {
225 		fi = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
226 		if (!fi)
227 			return -ENOMEM;
228 
229 		if (opt->flags & CEPH_MOUNT_OPT_NOPAGECACHE)
230 			fi->flags |= CEPH_F_SYNC;
231 
232 		file->private_data = fi;
233 	}
234 
235 	ceph_get_fmode(ci, fmode, 1);
236 	fi->fmode = fmode;
237 
238 	spin_lock_init(&fi->rw_contexts_lock);
239 	INIT_LIST_HEAD(&fi->rw_contexts);
240 	fi->filp_gen = READ_ONCE(ceph_inode_to_fs_client(inode)->filp_gen);
241 
242 	if ((file->f_mode & FMODE_WRITE) && ceph_has_inline_data(ci)) {
243 		ret = ceph_uninline_data(file);
244 		if (ret < 0)
245 			goto error;
246 	}
247 
248 	return 0;
249 
250 error:
251 	ceph_fscache_unuse_cookie(inode, file->f_mode & FMODE_WRITE);
252 	ceph_put_fmode(ci, fi->fmode, 1);
253 	kmem_cache_free(ceph_file_cachep, fi);
254 	/* wake up anyone waiting for caps on this inode */
255 	wake_up_all(&ci->i_cap_wq);
256 	return ret;
257 }
258 
259 /*
260  * initialize private struct file data.
261  * if we fail, clean up by dropping fmode reference on the ceph_inode
262  */
263 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
264 {
265 	struct ceph_client *cl = ceph_inode_to_client(inode);
266 	int ret = 0;
267 
268 	switch (inode->i_mode & S_IFMT) {
269 	case S_IFREG:
270 		ceph_fscache_use_cookie(inode, file->f_mode & FMODE_WRITE);
271 		fallthrough;
272 	case S_IFDIR:
273 		ret = ceph_init_file_info(inode, file, fmode,
274 						S_ISDIR(inode->i_mode));
275 		break;
276 
277 	case S_IFLNK:
278 		doutc(cl, "%p %llx.%llx %p 0%o (symlink)\n", inode,
279 		      ceph_vinop(inode), file, inode->i_mode);
280 		break;
281 
282 	default:
283 		doutc(cl, "%p %llx.%llx %p 0%o (special)\n", inode,
284 		      ceph_vinop(inode), file, inode->i_mode);
285 		/*
286 		 * we need to drop the open ref now, since we don't
287 		 * have .release set to ceph_release.
288 		 */
289 		BUG_ON(inode->i_fop->release == ceph_release);
290 
291 		/* call the proper open fop */
292 		ret = inode->i_fop->open(inode, file);
293 	}
294 	return ret;
295 }
296 
297 /*
298  * try renew caps after session gets killed.
299  */
300 int ceph_renew_caps(struct inode *inode, int fmode)
301 {
302 	struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
303 	struct ceph_client *cl = mdsc->fsc->client;
304 	struct ceph_inode_info *ci = ceph_inode(inode);
305 	struct ceph_mds_request *req;
306 	int err, flags, wanted;
307 
308 	spin_lock(&ci->i_ceph_lock);
309 	__ceph_touch_fmode(ci, mdsc, fmode);
310 	wanted = __ceph_caps_file_wanted(ci);
311 	if (__ceph_is_any_real_caps(ci) &&
312 	    (!(wanted & CEPH_CAP_ANY_WR) || ci->i_auth_cap)) {
313 		int issued = __ceph_caps_issued(ci, NULL);
314 		spin_unlock(&ci->i_ceph_lock);
315 		doutc(cl, "%p %llx.%llx want %s issued %s updating mds_wanted\n",
316 		      inode, ceph_vinop(inode), ceph_cap_string(wanted),
317 		      ceph_cap_string(issued));
318 		ceph_check_caps(ci, 0);
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 	doutc(cl, "%p %llx.%llx open result=%d\n", inode, ceph_vinop(inode),
349 	      err);
350 	return err < 0 ? err : 0;
351 }
352 
353 /*
354  * If we already have the requisite capabilities, we can satisfy
355  * the open request locally (no need to request new caps from the
356  * MDS).  We do, however, need to inform the MDS (asynchronously)
357  * if our wanted caps set expands.
358  */
359 int ceph_open(struct inode *inode, struct file *file)
360 {
361 	struct ceph_inode_info *ci = ceph_inode(inode);
362 	struct ceph_fs_client *fsc = ceph_sb_to_fs_client(inode->i_sb);
363 	struct ceph_client *cl = fsc->client;
364 	struct ceph_mds_client *mdsc = fsc->mdsc;
365 	struct ceph_mds_request *req;
366 	struct ceph_file_info *fi = file->private_data;
367 	int err;
368 	int flags, fmode, wanted;
369 	struct dentry *dentry;
370 	char *path;
371 	int pathlen;
372 	u64 pathbase;
373 	bool do_sync = false;
374 	int mask = MAY_READ;
375 
376 	if (fi) {
377 		doutc(cl, "file %p is already opened\n", file);
378 		return 0;
379 	}
380 
381 	/* filter out O_CREAT|O_EXCL; vfs did that already.  yuck. */
382 	flags = file->f_flags & ~(O_CREAT|O_EXCL);
383 	if (S_ISDIR(inode->i_mode)) {
384 		flags = O_DIRECTORY;  /* mds likes to know */
385 	} else if (S_ISREG(inode->i_mode)) {
386 		err = fscrypt_file_open(inode, file);
387 		if (err)
388 			return err;
389 	}
390 
391 	doutc(cl, "%p %llx.%llx file %p flags %d (%d)\n", inode,
392 	      ceph_vinop(inode), file, flags, file->f_flags);
393 	fmode = ceph_flags_to_mode(flags);
394 	wanted = ceph_caps_for_mode(fmode);
395 
396 	if (fmode & CEPH_FILE_MODE_WR)
397 		mask |= MAY_WRITE;
398 	dentry = d_find_alias(inode);
399 	if (!dentry) {
400 		do_sync = true;
401 	} else {
402 		path = ceph_mdsc_build_path(mdsc, dentry, &pathlen, &pathbase, 0);
403 		if (IS_ERR(path)) {
404 			do_sync = true;
405 			err = 0;
406 		} else {
407 			err = ceph_mds_check_access(mdsc, path, mask);
408 		}
409 		ceph_mdsc_free_path(path, pathlen);
410 		dput(dentry);
411 
412 		/* For none EACCES cases will let the MDS do the mds auth check */
413 		if (err == -EACCES) {
414 			return err;
415 		} else if (err < 0) {
416 			do_sync = true;
417 			err = 0;
418 		}
419 	}
420 
421 	/* snapped files are read-only */
422 	if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
423 		return -EROFS;
424 
425 	/* trivially open snapdir */
426 	if (ceph_snap(inode) == CEPH_SNAPDIR) {
427 		return ceph_init_file(inode, file, fmode);
428 	}
429 
430 	/*
431 	 * No need to block if we have caps on the auth MDS (for
432 	 * write) or any MDS (for read).  Update wanted set
433 	 * asynchronously.
434 	 */
435 	spin_lock(&ci->i_ceph_lock);
436 	if (!do_sync && __ceph_is_any_real_caps(ci) &&
437 	    (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
438 		int mds_wanted = __ceph_caps_mds_wanted(ci, true);
439 		int issued = __ceph_caps_issued(ci, NULL);
440 
441 		doutc(cl, "open %p fmode %d want %s issued %s using existing\n",
442 		      inode, fmode, ceph_cap_string(wanted),
443 		      ceph_cap_string(issued));
444 		__ceph_touch_fmode(ci, mdsc, fmode);
445 		spin_unlock(&ci->i_ceph_lock);
446 
447 		/* adjust wanted? */
448 		if ((issued & wanted) != wanted &&
449 		    (mds_wanted & wanted) != wanted &&
450 		    ceph_snap(inode) != CEPH_SNAPDIR)
451 			ceph_check_caps(ci, 0);
452 
453 		return ceph_init_file(inode, file, fmode);
454 	} else if (!do_sync && ceph_snap(inode) != CEPH_NOSNAP &&
455 		   (ci->i_snap_caps & wanted) == wanted) {
456 		__ceph_touch_fmode(ci, mdsc, fmode);
457 		spin_unlock(&ci->i_ceph_lock);
458 		return ceph_init_file(inode, file, fmode);
459 	}
460 
461 	spin_unlock(&ci->i_ceph_lock);
462 
463 	doutc(cl, "open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
464 	req = prepare_open_request(inode->i_sb, flags, 0);
465 	if (IS_ERR(req)) {
466 		err = PTR_ERR(req);
467 		goto out;
468 	}
469 	req->r_inode = inode;
470 	ihold(inode);
471 
472 	req->r_num_caps = 1;
473 	err = ceph_mdsc_do_request(mdsc, NULL, req);
474 	if (!err)
475 		err = ceph_init_file(inode, file, req->r_fmode);
476 	ceph_mdsc_put_request(req);
477 	doutc(cl, "open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
478 out:
479 	return err;
480 }
481 
482 /* Clone the layout from a synchronous create, if the dir now has Dc caps */
483 static void
484 cache_file_layout(struct inode *dst, struct inode *src)
485 {
486 	struct ceph_inode_info *cdst = ceph_inode(dst);
487 	struct ceph_inode_info *csrc = ceph_inode(src);
488 
489 	spin_lock(&cdst->i_ceph_lock);
490 	if ((__ceph_caps_issued(cdst, NULL) & CEPH_CAP_DIR_CREATE) &&
491 	    !ceph_file_layout_is_valid(&cdst->i_cached_layout)) {
492 		memcpy(&cdst->i_cached_layout, &csrc->i_layout,
493 			sizeof(cdst->i_cached_layout));
494 		rcu_assign_pointer(cdst->i_cached_layout.pool_ns,
495 				   ceph_try_get_string(csrc->i_layout.pool_ns));
496 	}
497 	spin_unlock(&cdst->i_ceph_lock);
498 }
499 
500 /*
501  * Try to set up an async create. We need caps, a file layout, and inode number,
502  * and either a lease on the dentry or complete dir info. If any of those
503  * criteria are not satisfied, then return false and the caller can go
504  * synchronous.
505  */
506 static int try_prep_async_create(struct inode *dir, struct dentry *dentry,
507 				 struct ceph_file_layout *lo, u64 *pino)
508 {
509 	struct ceph_inode_info *ci = ceph_inode(dir);
510 	struct ceph_dentry_info *di = ceph_dentry(dentry);
511 	int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_CREATE;
512 	u64 ino;
513 
514 	spin_lock(&ci->i_ceph_lock);
515 	/* No auth cap means no chance for Dc caps */
516 	if (!ci->i_auth_cap)
517 		goto no_async;
518 
519 	/* Any delegated inos? */
520 	if (xa_empty(&ci->i_auth_cap->session->s_delegated_inos))
521 		goto no_async;
522 
523 	if (!ceph_file_layout_is_valid(&ci->i_cached_layout))
524 		goto no_async;
525 
526 	if ((__ceph_caps_issued(ci, NULL) & want) != want)
527 		goto no_async;
528 
529 	if (d_in_lookup(dentry)) {
530 		if (!__ceph_dir_is_complete(ci))
531 			goto no_async;
532 		spin_lock(&dentry->d_lock);
533 		di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
534 		spin_unlock(&dentry->d_lock);
535 	} else if (atomic_read(&ci->i_shared_gen) !=
536 		   READ_ONCE(di->lease_shared_gen)) {
537 		goto no_async;
538 	}
539 
540 	ino = ceph_get_deleg_ino(ci->i_auth_cap->session);
541 	if (!ino)
542 		goto no_async;
543 
544 	*pino = ino;
545 	ceph_take_cap_refs(ci, want, false);
546 	memcpy(lo, &ci->i_cached_layout, sizeof(*lo));
547 	rcu_assign_pointer(lo->pool_ns,
548 			   ceph_try_get_string(ci->i_cached_layout.pool_ns));
549 	got = want;
550 no_async:
551 	spin_unlock(&ci->i_ceph_lock);
552 	return got;
553 }
554 
555 static void restore_deleg_ino(struct inode *dir, u64 ino)
556 {
557 	struct ceph_client *cl = ceph_inode_to_client(dir);
558 	struct ceph_inode_info *ci = ceph_inode(dir);
559 	struct ceph_mds_session *s = NULL;
560 
561 	spin_lock(&ci->i_ceph_lock);
562 	if (ci->i_auth_cap)
563 		s = ceph_get_mds_session(ci->i_auth_cap->session);
564 	spin_unlock(&ci->i_ceph_lock);
565 	if (s) {
566 		int err = ceph_restore_deleg_ino(s, ino);
567 		if (err)
568 			pr_warn_client(cl,
569 				"unable to restore delegated ino 0x%llx to session: %d\n",
570 				ino, err);
571 		ceph_put_mds_session(s);
572 	}
573 }
574 
575 static void wake_async_create_waiters(struct inode *inode,
576 				      struct ceph_mds_session *session)
577 {
578 	struct ceph_inode_info *ci = ceph_inode(inode);
579 	bool check_cap = false;
580 
581 	spin_lock(&ci->i_ceph_lock);
582 	if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE) {
583 		ci->i_ceph_flags &= ~CEPH_I_ASYNC_CREATE;
584 		wake_up_bit(&ci->i_ceph_flags, CEPH_ASYNC_CREATE_BIT);
585 
586 		if (ci->i_ceph_flags & CEPH_I_ASYNC_CHECK_CAPS) {
587 			ci->i_ceph_flags &= ~CEPH_I_ASYNC_CHECK_CAPS;
588 			check_cap = true;
589 		}
590 	}
591 	ceph_kick_flushing_inode_caps(session, ci);
592 	spin_unlock(&ci->i_ceph_lock);
593 
594 	if (check_cap)
595 		ceph_check_caps(ci, CHECK_CAPS_FLUSH);
596 }
597 
598 static void ceph_async_create_cb(struct ceph_mds_client *mdsc,
599                                  struct ceph_mds_request *req)
600 {
601 	struct ceph_client *cl = mdsc->fsc->client;
602 	struct dentry *dentry = req->r_dentry;
603 	struct inode *dinode = d_inode(dentry);
604 	struct inode *tinode = req->r_target_inode;
605 	int result = req->r_err ? req->r_err :
606 			le32_to_cpu(req->r_reply_info.head->result);
607 
608 	WARN_ON_ONCE(dinode && tinode && dinode != tinode);
609 
610 	/* MDS changed -- caller must resubmit */
611 	if (result == -EJUKEBOX)
612 		goto out;
613 
614 	mapping_set_error(req->r_parent->i_mapping, result);
615 
616 	if (result) {
617 		int pathlen = 0;
618 		u64 base = 0;
619 		char *path = ceph_mdsc_build_path(mdsc, req->r_dentry, &pathlen,
620 						  &base, 0);
621 
622 		pr_warn_client(cl,
623 			"async create failure path=(%llx)%s result=%d!\n",
624 			base, IS_ERR(path) ? "<<bad>>" : path, result);
625 		ceph_mdsc_free_path(path, pathlen);
626 
627 		ceph_dir_clear_complete(req->r_parent);
628 		if (!d_unhashed(dentry))
629 			d_drop(dentry);
630 
631 		if (dinode) {
632 			mapping_set_error(dinode->i_mapping, result);
633 			ceph_inode_shutdown(dinode);
634 			wake_async_create_waiters(dinode, req->r_session);
635 		}
636 	}
637 
638 	if (tinode) {
639 		u64 ino = ceph_vino(tinode).ino;
640 
641 		if (req->r_deleg_ino != ino)
642 			pr_warn_client(cl,
643 				"inode number mismatch! err=%d deleg_ino=0x%llx target=0x%llx\n",
644 				req->r_err, req->r_deleg_ino, ino);
645 
646 		mapping_set_error(tinode->i_mapping, result);
647 		wake_async_create_waiters(tinode, req->r_session);
648 	} else if (!result) {
649 		pr_warn_client(cl, "no req->r_target_inode for 0x%llx\n",
650 			       req->r_deleg_ino);
651 	}
652 out:
653 	ceph_mdsc_release_dir_caps(req);
654 }
655 
656 static int ceph_finish_async_create(struct inode *dir, struct inode *inode,
657 				    struct dentry *dentry,
658 				    struct file *file, umode_t mode,
659 				    struct ceph_mds_request *req,
660 				    struct ceph_acl_sec_ctx *as_ctx,
661 				    struct ceph_file_layout *lo)
662 {
663 	int ret;
664 	char xattr_buf[4];
665 	struct ceph_mds_reply_inode in = { };
666 	struct ceph_mds_reply_info_in iinfo = { .in = &in };
667 	struct ceph_inode_info *ci = ceph_inode(dir);
668 	struct ceph_dentry_info *di = ceph_dentry(dentry);
669 	struct timespec64 now;
670 	struct ceph_string *pool_ns;
671 	struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
672 	struct ceph_client *cl = mdsc->fsc->client;
673 	struct ceph_vino vino = { .ino = req->r_deleg_ino,
674 				  .snap = CEPH_NOSNAP };
675 
676 	ktime_get_real_ts64(&now);
677 
678 	iinfo.inline_version = CEPH_INLINE_NONE;
679 	iinfo.change_attr = 1;
680 	ceph_encode_timespec64(&iinfo.btime, &now);
681 
682 	if (req->r_pagelist) {
683 		iinfo.xattr_len = req->r_pagelist->length;
684 		iinfo.xattr_data = req->r_pagelist->mapped_tail;
685 	} else {
686 		/* fake it */
687 		iinfo.xattr_len = ARRAY_SIZE(xattr_buf);
688 		iinfo.xattr_data = xattr_buf;
689 		memset(iinfo.xattr_data, 0, iinfo.xattr_len);
690 	}
691 
692 	in.ino = cpu_to_le64(vino.ino);
693 	in.snapid = cpu_to_le64(CEPH_NOSNAP);
694 	in.version = cpu_to_le64(1);	// ???
695 	in.cap.caps = in.cap.wanted = cpu_to_le32(CEPH_CAP_ALL_FILE);
696 	in.cap.cap_id = cpu_to_le64(1);
697 	in.cap.realm = cpu_to_le64(ci->i_snap_realm->ino);
698 	in.cap.flags = CEPH_CAP_FLAG_AUTH;
699 	in.ctime = in.mtime = in.atime = iinfo.btime;
700 	in.truncate_seq = cpu_to_le32(1);
701 	in.truncate_size = cpu_to_le64(-1ULL);
702 	in.xattr_version = cpu_to_le64(1);
703 	in.uid = cpu_to_le32(from_kuid(&init_user_ns,
704 				       mapped_fsuid(req->r_mnt_idmap,
705 						    &init_user_ns)));
706 	if (dir->i_mode & S_ISGID) {
707 		in.gid = cpu_to_le32(from_kgid(&init_user_ns, dir->i_gid));
708 
709 		/* Directories always inherit the setgid bit. */
710 		if (S_ISDIR(mode))
711 			mode |= S_ISGID;
712 	} else {
713 		in.gid = cpu_to_le32(from_kgid(&init_user_ns,
714 				     mapped_fsgid(req->r_mnt_idmap,
715 						  &init_user_ns)));
716 	}
717 	in.mode = cpu_to_le32((u32)mode);
718 
719 	in.nlink = cpu_to_le32(1);
720 	in.max_size = cpu_to_le64(lo->stripe_unit);
721 
722 	ceph_file_layout_to_legacy(lo, &in.layout);
723 	/* lo is private, so pool_ns can't change */
724 	pool_ns = rcu_dereference_raw(lo->pool_ns);
725 	if (pool_ns) {
726 		iinfo.pool_ns_len = pool_ns->len;
727 		iinfo.pool_ns_data = pool_ns->str;
728 	}
729 
730 	down_read(&mdsc->snap_rwsem);
731 	ret = ceph_fill_inode(inode, NULL, &iinfo, NULL, req->r_session,
732 			      req->r_fmode, NULL);
733 	up_read(&mdsc->snap_rwsem);
734 	if (ret) {
735 		doutc(cl, "failed to fill inode: %d\n", ret);
736 		ceph_dir_clear_complete(dir);
737 		if (!d_unhashed(dentry))
738 			d_drop(dentry);
739 		discard_new_inode(inode);
740 	} else {
741 		struct dentry *dn;
742 
743 		doutc(cl, "d_adding new inode 0x%llx to 0x%llx/%s\n",
744 		      vino.ino, ceph_ino(dir), dentry->d_name.name);
745 		ceph_dir_clear_ordered(dir);
746 		ceph_init_inode_acls(inode, as_ctx);
747 		if (inode->i_state & I_NEW) {
748 			/*
749 			 * If it's not I_NEW, then someone created this before
750 			 * we got here. Assume the server is aware of it at
751 			 * that point and don't worry about setting
752 			 * CEPH_I_ASYNC_CREATE.
753 			 */
754 			ceph_inode(inode)->i_ceph_flags = CEPH_I_ASYNC_CREATE;
755 			unlock_new_inode(inode);
756 		}
757 		if (d_in_lookup(dentry) || d_really_is_negative(dentry)) {
758 			if (!d_unhashed(dentry))
759 				d_drop(dentry);
760 			dn = d_splice_alias(inode, dentry);
761 			WARN_ON_ONCE(dn && dn != dentry);
762 		}
763 		file->f_mode |= FMODE_CREATED;
764 		ret = finish_open(file, dentry, ceph_open);
765 	}
766 
767 	spin_lock(&dentry->d_lock);
768 	di->flags &= ~CEPH_DENTRY_ASYNC_CREATE;
769 	wake_up_bit(&di->flags, CEPH_DENTRY_ASYNC_CREATE_BIT);
770 	spin_unlock(&dentry->d_lock);
771 
772 	return ret;
773 }
774 
775 /*
776  * Do a lookup + open with a single request.  If we get a non-existent
777  * file or symlink, return 1 so the VFS can retry.
778  */
779 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
780 		     struct file *file, unsigned flags, umode_t mode)
781 {
782 	struct mnt_idmap *idmap = file_mnt_idmap(file);
783 	struct ceph_fs_client *fsc = ceph_sb_to_fs_client(dir->i_sb);
784 	struct ceph_client *cl = fsc->client;
785 	struct ceph_mds_client *mdsc = fsc->mdsc;
786 	struct ceph_mds_request *req;
787 	struct inode *new_inode = NULL;
788 	struct dentry *dn;
789 	struct ceph_acl_sec_ctx as_ctx = {};
790 	bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
791 	int mask;
792 	int err;
793 	char *path;
794 	int pathlen;
795 	u64 pathbase;
796 
797 	doutc(cl, "%p %llx.%llx dentry %p '%pd' %s flags %d mode 0%o\n",
798 	      dir, ceph_vinop(dir), dentry, dentry,
799 	      d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
800 
801 	if (dentry->d_name.len > NAME_MAX)
802 		return -ENAMETOOLONG;
803 
804 	err = ceph_wait_on_conflict_unlink(dentry);
805 	if (err)
806 		return err;
807 	/*
808 	 * Do not truncate the file, since atomic_open is called before the
809 	 * permission check. The caller will do the truncation afterward.
810 	 */
811 	flags &= ~O_TRUNC;
812 
813 	dn = d_find_alias(dir);
814 	if (!dn) {
815 		try_async = false;
816 	} else {
817 		path = ceph_mdsc_build_path(mdsc, dn, &pathlen, &pathbase, 0);
818 		if (IS_ERR(path)) {
819 			try_async = false;
820 			err = 0;
821 		} else {
822 			int fmode = ceph_flags_to_mode(flags);
823 
824 			mask = MAY_READ;
825 			if (fmode & CEPH_FILE_MODE_WR)
826 				mask |= MAY_WRITE;
827 			err = ceph_mds_check_access(mdsc, path, mask);
828 		}
829 		ceph_mdsc_free_path(path, pathlen);
830 		dput(dn);
831 
832 		/* For none EACCES cases will let the MDS do the mds auth check */
833 		if (err == -EACCES) {
834 			return err;
835 		} else if (err < 0) {
836 			try_async = false;
837 			err = 0;
838 		}
839 	}
840 
841 retry:
842 	if (flags & O_CREAT) {
843 		if (ceph_quota_is_max_files_exceeded(dir))
844 			return -EDQUOT;
845 
846 		new_inode = ceph_new_inode(dir, dentry, &mode, &as_ctx);
847 		if (IS_ERR(new_inode)) {
848 			err = PTR_ERR(new_inode);
849 			goto out_ctx;
850 		}
851 		/* Async create can't handle more than a page of xattrs */
852 		if (as_ctx.pagelist &&
853 		    !list_is_singular(&as_ctx.pagelist->head))
854 			try_async = false;
855 	} else if (!d_in_lookup(dentry)) {
856 		/* If it's not being looked up, it's negative */
857 		return -ENOENT;
858 	}
859 
860 	/* do the open */
861 	req = prepare_open_request(dir->i_sb, flags, mode);
862 	if (IS_ERR(req)) {
863 		err = PTR_ERR(req);
864 		goto out_ctx;
865 	}
866 	req->r_dentry = dget(dentry);
867 	req->r_num_caps = 2;
868 	mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
869 	if (ceph_security_xattr_wanted(dir))
870 		mask |= CEPH_CAP_XATTR_SHARED;
871 	req->r_args.open.mask = cpu_to_le32(mask);
872 	req->r_parent = dir;
873 	if (req->r_op == CEPH_MDS_OP_CREATE)
874 		req->r_mnt_idmap = mnt_idmap_get(idmap);
875 	ihold(dir);
876 	if (IS_ENCRYPTED(dir)) {
877 		set_bit(CEPH_MDS_R_FSCRYPT_FILE, &req->r_req_flags);
878 		err = fscrypt_prepare_lookup_partial(dir, dentry);
879 		if (err < 0)
880 			goto out_req;
881 	}
882 
883 	if (flags & O_CREAT) {
884 		struct ceph_file_layout lo;
885 
886 		req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL |
887 				     CEPH_CAP_XATTR_EXCL;
888 		req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
889 
890 		ceph_as_ctx_to_req(req, &as_ctx);
891 
892 		if (try_async && (req->r_dir_caps =
893 				  try_prep_async_create(dir, dentry, &lo,
894 							&req->r_deleg_ino))) {
895 			struct ceph_vino vino = { .ino = req->r_deleg_ino,
896 						  .snap = CEPH_NOSNAP };
897 			struct ceph_dentry_info *di = ceph_dentry(dentry);
898 
899 			set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
900 			req->r_args.open.flags |= cpu_to_le32(CEPH_O_EXCL);
901 			req->r_callback = ceph_async_create_cb;
902 
903 			/* Hash inode before RPC */
904 			new_inode = ceph_get_inode(dir->i_sb, vino, new_inode);
905 			if (IS_ERR(new_inode)) {
906 				err = PTR_ERR(new_inode);
907 				new_inode = NULL;
908 				goto out_req;
909 			}
910 			WARN_ON_ONCE(!(new_inode->i_state & I_NEW));
911 
912 			spin_lock(&dentry->d_lock);
913 			di->flags |= CEPH_DENTRY_ASYNC_CREATE;
914 			spin_unlock(&dentry->d_lock);
915 
916 			err = ceph_mdsc_submit_request(mdsc, dir, req);
917 			if (!err) {
918 				err = ceph_finish_async_create(dir, new_inode,
919 							       dentry, file,
920 							       mode, req,
921 							       &as_ctx, &lo);
922 				new_inode = NULL;
923 			} else if (err == -EJUKEBOX) {
924 				restore_deleg_ino(dir, req->r_deleg_ino);
925 				ceph_mdsc_put_request(req);
926 				discard_new_inode(new_inode);
927 				ceph_release_acl_sec_ctx(&as_ctx);
928 				memset(&as_ctx, 0, sizeof(as_ctx));
929 				new_inode = NULL;
930 				try_async = false;
931 				ceph_put_string(rcu_dereference_raw(lo.pool_ns));
932 				goto retry;
933 			}
934 			ceph_put_string(rcu_dereference_raw(lo.pool_ns));
935 			goto out_req;
936 		}
937 	}
938 
939 	set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
940 	req->r_new_inode = new_inode;
941 	new_inode = NULL;
942 	err = ceph_mdsc_do_request(mdsc, (flags & O_CREAT) ? dir : NULL, req);
943 	if (err == -ENOENT) {
944 		dentry = ceph_handle_snapdir(req, dentry);
945 		if (IS_ERR(dentry)) {
946 			err = PTR_ERR(dentry);
947 			goto out_req;
948 		}
949 		err = 0;
950 	}
951 
952 	if (!err && (flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
953 		err = ceph_handle_notrace_create(dir, dentry);
954 
955 	if (d_in_lookup(dentry)) {
956 		dn = ceph_finish_lookup(req, dentry, err);
957 		if (IS_ERR(dn))
958 			err = PTR_ERR(dn);
959 	} else {
960 		/* we were given a hashed negative dentry */
961 		dn = NULL;
962 	}
963 	if (err)
964 		goto out_req;
965 	if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
966 		/* make vfs retry on splice, ENOENT, or symlink */
967 		doutc(cl, "finish_no_open on dn %p\n", dn);
968 		err = finish_no_open(file, dn);
969 	} else {
970 		if (IS_ENCRYPTED(dir) &&
971 		    !fscrypt_has_permitted_context(dir, d_inode(dentry))) {
972 			pr_warn_client(cl,
973 				"Inconsistent encryption context (parent %llx:%llx child %llx:%llx)\n",
974 				ceph_vinop(dir), ceph_vinop(d_inode(dentry)));
975 			goto out_req;
976 		}
977 
978 		doutc(cl, "finish_open on dn %p\n", dn);
979 		if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
980 			struct inode *newino = d_inode(dentry);
981 
982 			cache_file_layout(dir, newino);
983 			ceph_init_inode_acls(newino, &as_ctx);
984 			file->f_mode |= FMODE_CREATED;
985 		}
986 		err = finish_open(file, dentry, ceph_open);
987 	}
988 out_req:
989 	ceph_mdsc_put_request(req);
990 	iput(new_inode);
991 out_ctx:
992 	ceph_release_acl_sec_ctx(&as_ctx);
993 	doutc(cl, "result=%d\n", err);
994 	return err;
995 }
996 
997 int ceph_release(struct inode *inode, struct file *file)
998 {
999 	struct ceph_client *cl = ceph_inode_to_client(inode);
1000 	struct ceph_inode_info *ci = ceph_inode(inode);
1001 
1002 	if (S_ISDIR(inode->i_mode)) {
1003 		struct ceph_dir_file_info *dfi = file->private_data;
1004 		doutc(cl, "%p %llx.%llx dir file %p\n", inode,
1005 		      ceph_vinop(inode), file);
1006 		WARN_ON(!list_empty(&dfi->file_info.rw_contexts));
1007 
1008 		ceph_put_fmode(ci, dfi->file_info.fmode, 1);
1009 
1010 		if (dfi->last_readdir)
1011 			ceph_mdsc_put_request(dfi->last_readdir);
1012 		kfree(dfi->last_name);
1013 		kfree(dfi->dir_info);
1014 		kmem_cache_free(ceph_dir_file_cachep, dfi);
1015 	} else {
1016 		struct ceph_file_info *fi = file->private_data;
1017 		doutc(cl, "%p %llx.%llx regular file %p\n", inode,
1018 		      ceph_vinop(inode), file);
1019 		WARN_ON(!list_empty(&fi->rw_contexts));
1020 
1021 		ceph_fscache_unuse_cookie(inode, file->f_mode & FMODE_WRITE);
1022 		ceph_put_fmode(ci, fi->fmode, 1);
1023 
1024 		kmem_cache_free(ceph_file_cachep, fi);
1025 	}
1026 
1027 	/* wake up anyone waiting for caps on this inode */
1028 	wake_up_all(&ci->i_cap_wq);
1029 	return 0;
1030 }
1031 
1032 enum {
1033 	HAVE_RETRIED = 1,
1034 	CHECK_EOF =    2,
1035 	READ_INLINE =  3,
1036 };
1037 
1038 /*
1039  * Completely synchronous read and write methods.  Direct from __user
1040  * buffer to osd, or directly to user pages (if O_DIRECT).
1041  *
1042  * If the read spans object boundary, just do multiple reads.  (That's not
1043  * atomic, but good enough for now.)
1044  *
1045  * If we get a short result from the OSD, check against i_size; we need to
1046  * only return a short read to the caller if we hit EOF.
1047  */
1048 ssize_t __ceph_sync_read(struct inode *inode, loff_t *ki_pos,
1049 			 struct iov_iter *to, int *retry_op,
1050 			 u64 *last_objver)
1051 {
1052 	struct ceph_inode_info *ci = ceph_inode(inode);
1053 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
1054 	struct ceph_client *cl = fsc->client;
1055 	struct ceph_osd_client *osdc = &fsc->client->osdc;
1056 	ssize_t ret;
1057 	u64 off = *ki_pos;
1058 	u64 len = iov_iter_count(to);
1059 	u64 i_size = i_size_read(inode);
1060 	bool sparse = IS_ENCRYPTED(inode) || ceph_test_mount_opt(fsc, SPARSEREAD);
1061 	u64 objver = 0;
1062 
1063 	doutc(cl, "on inode %p %llx.%llx %llx~%llx\n", inode,
1064 	      ceph_vinop(inode), *ki_pos, len);
1065 
1066 	if (ceph_inode_is_shutdown(inode))
1067 		return -EIO;
1068 
1069 	if (!len)
1070 		return 0;
1071 	/*
1072 	 * flush any page cache pages in this range.  this
1073 	 * will make concurrent normal and sync io slow,
1074 	 * but it will at least behave sensibly when they are
1075 	 * in sequence.
1076 	 */
1077 	ret = filemap_write_and_wait_range(inode->i_mapping,
1078 					   off, off + len - 1);
1079 	if (ret < 0)
1080 		return ret;
1081 
1082 	ret = 0;
1083 	while ((len = iov_iter_count(to)) > 0) {
1084 		struct ceph_osd_request *req;
1085 		struct page **pages;
1086 		int num_pages;
1087 		size_t page_off;
1088 		bool more;
1089 		int idx;
1090 		size_t left;
1091 		struct ceph_osd_req_op *op;
1092 		u64 read_off = off;
1093 		u64 read_len = len;
1094 		int extent_cnt;
1095 
1096 		/* determine new offset/length if encrypted */
1097 		ceph_fscrypt_adjust_off_and_len(inode, &read_off, &read_len);
1098 
1099 		doutc(cl, "orig %llu~%llu reading %llu~%llu", off, len,
1100 		      read_off, read_len);
1101 
1102 		req = ceph_osdc_new_request(osdc, &ci->i_layout,
1103 					ci->i_vino, read_off, &read_len, 0, 1,
1104 					sparse ? CEPH_OSD_OP_SPARSE_READ :
1105 						 CEPH_OSD_OP_READ,
1106 					CEPH_OSD_FLAG_READ,
1107 					NULL, ci->i_truncate_seq,
1108 					ci->i_truncate_size, false);
1109 		if (IS_ERR(req)) {
1110 			ret = PTR_ERR(req);
1111 			break;
1112 		}
1113 
1114 		/* adjust len downward if the request truncated the len */
1115 		if (off + len > read_off + read_len)
1116 			len = read_off + read_len - off;
1117 		more = len < iov_iter_count(to);
1118 
1119 		num_pages = calc_pages_for(read_off, read_len);
1120 		page_off = offset_in_page(off);
1121 		pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1122 		if (IS_ERR(pages)) {
1123 			ceph_osdc_put_request(req);
1124 			ret = PTR_ERR(pages);
1125 			break;
1126 		}
1127 
1128 		osd_req_op_extent_osd_data_pages(req, 0, pages, read_len,
1129 						 offset_in_page(read_off),
1130 						 false, false);
1131 
1132 		op = &req->r_ops[0];
1133 		if (sparse) {
1134 			extent_cnt = __ceph_sparse_read_ext_count(inode, read_len);
1135 			ret = ceph_alloc_sparse_ext_map(op, extent_cnt);
1136 			if (ret) {
1137 				ceph_osdc_put_request(req);
1138 				break;
1139 			}
1140 		}
1141 
1142 		ceph_osdc_start_request(osdc, req);
1143 		ret = ceph_osdc_wait_request(osdc, req);
1144 
1145 		ceph_update_read_metrics(&fsc->mdsc->metric,
1146 					 req->r_start_latency,
1147 					 req->r_end_latency,
1148 					 read_len, ret);
1149 
1150 		if (ret > 0)
1151 			objver = req->r_version;
1152 
1153 		i_size = i_size_read(inode);
1154 		doutc(cl, "%llu~%llu got %zd i_size %llu%s\n", off, len,
1155 		      ret, i_size, (more ? " MORE" : ""));
1156 
1157 		/* Fix it to go to end of extent map */
1158 		if (sparse && ret >= 0)
1159 			ret = ceph_sparse_ext_map_end(op);
1160 		else if (ret == -ENOENT)
1161 			ret = 0;
1162 
1163 		if (ret > 0 && IS_ENCRYPTED(inode)) {
1164 			int fret;
1165 
1166 			fret = ceph_fscrypt_decrypt_extents(inode, pages,
1167 					read_off, op->extent.sparse_ext,
1168 					op->extent.sparse_ext_cnt);
1169 			if (fret < 0) {
1170 				ret = fret;
1171 				ceph_osdc_put_request(req);
1172 				break;
1173 			}
1174 
1175 			/* account for any partial block at the beginning */
1176 			fret -= (off - read_off);
1177 
1178 			/*
1179 			 * Short read after big offset adjustment?
1180 			 * Nothing is usable, just call it a zero
1181 			 * len read.
1182 			 */
1183 			fret = max(fret, 0);
1184 
1185 			/* account for partial block at the end */
1186 			ret = min_t(ssize_t, fret, len);
1187 		}
1188 
1189 		ceph_osdc_put_request(req);
1190 
1191 		/* Short read but not EOF? Zero out the remainder. */
1192 		if (ret >= 0 && ret < len && (off + ret < i_size)) {
1193 			int zlen = min(len - ret, i_size - off - ret);
1194 			int zoff = page_off + ret;
1195 
1196 			doutc(cl, "zero gap %llu~%llu\n", off + ret,
1197 			      off + ret + zlen);
1198 			ceph_zero_page_vector_range(zoff, zlen, pages);
1199 			ret += zlen;
1200 		}
1201 
1202 		idx = 0;
1203 		if (ret <= 0)
1204 			left = 0;
1205 		else if (off + ret > i_size)
1206 			left = i_size - off;
1207 		else
1208 			left = ret;
1209 		while (left > 0) {
1210 			size_t plen, copied;
1211 
1212 			plen = min_t(size_t, left, PAGE_SIZE - page_off);
1213 			SetPageUptodate(pages[idx]);
1214 			copied = copy_page_to_iter(pages[idx++],
1215 						   page_off, plen, to);
1216 			off += copied;
1217 			left -= copied;
1218 			page_off = 0;
1219 			if (copied < plen) {
1220 				ret = -EFAULT;
1221 				break;
1222 			}
1223 		}
1224 		ceph_release_page_vector(pages, num_pages);
1225 
1226 		if (ret < 0) {
1227 			if (ret == -EBLOCKLISTED)
1228 				fsc->blocklisted = true;
1229 			break;
1230 		}
1231 
1232 		if (off >= i_size || !more)
1233 			break;
1234 	}
1235 
1236 	if (ret > 0) {
1237 		if (off >= i_size) {
1238 			*retry_op = CHECK_EOF;
1239 			ret = i_size - *ki_pos;
1240 			*ki_pos = i_size;
1241 		} else {
1242 			ret = off - *ki_pos;
1243 			*ki_pos = off;
1244 		}
1245 
1246 		if (last_objver)
1247 			*last_objver = objver;
1248 	}
1249 	doutc(cl, "result %zd retry_op %d\n", ret, *retry_op);
1250 	return ret;
1251 }
1252 
1253 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
1254 			      int *retry_op)
1255 {
1256 	struct file *file = iocb->ki_filp;
1257 	struct inode *inode = file_inode(file);
1258 	struct ceph_client *cl = ceph_inode_to_client(inode);
1259 
1260 	doutc(cl, "on file %p %llx~%zx %s\n", file, iocb->ki_pos,
1261 	      iov_iter_count(to),
1262 	      (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
1263 
1264 	return __ceph_sync_read(inode, &iocb->ki_pos, to, retry_op, NULL);
1265 }
1266 
1267 struct ceph_aio_request {
1268 	struct kiocb *iocb;
1269 	size_t total_len;
1270 	bool write;
1271 	bool should_dirty;
1272 	int error;
1273 	struct list_head osd_reqs;
1274 	unsigned num_reqs;
1275 	atomic_t pending_reqs;
1276 	struct timespec64 mtime;
1277 	struct ceph_cap_flush *prealloc_cf;
1278 };
1279 
1280 struct ceph_aio_work {
1281 	struct work_struct work;
1282 	struct ceph_osd_request *req;
1283 };
1284 
1285 static void ceph_aio_retry_work(struct work_struct *work);
1286 
1287 static void ceph_aio_complete(struct inode *inode,
1288 			      struct ceph_aio_request *aio_req)
1289 {
1290 	struct ceph_client *cl = ceph_inode_to_client(inode);
1291 	struct ceph_inode_info *ci = ceph_inode(inode);
1292 	int ret;
1293 
1294 	if (!atomic_dec_and_test(&aio_req->pending_reqs))
1295 		return;
1296 
1297 	if (aio_req->iocb->ki_flags & IOCB_DIRECT)
1298 		inode_dio_end(inode);
1299 
1300 	ret = aio_req->error;
1301 	if (!ret)
1302 		ret = aio_req->total_len;
1303 
1304 	doutc(cl, "%p %llx.%llx rc %d\n", inode, ceph_vinop(inode), ret);
1305 
1306 	if (ret >= 0 && aio_req->write) {
1307 		int dirty;
1308 
1309 		loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
1310 		if (endoff > i_size_read(inode)) {
1311 			if (ceph_inode_set_size(inode, endoff))
1312 				ceph_check_caps(ci, CHECK_CAPS_AUTHONLY);
1313 		}
1314 
1315 		spin_lock(&ci->i_ceph_lock);
1316 		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1317 					       &aio_req->prealloc_cf);
1318 		spin_unlock(&ci->i_ceph_lock);
1319 		if (dirty)
1320 			__mark_inode_dirty(inode, dirty);
1321 
1322 	}
1323 
1324 	ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
1325 						CEPH_CAP_FILE_RD));
1326 
1327 	aio_req->iocb->ki_complete(aio_req->iocb, ret);
1328 
1329 	ceph_free_cap_flush(aio_req->prealloc_cf);
1330 	kfree(aio_req);
1331 }
1332 
1333 static void ceph_aio_complete_req(struct ceph_osd_request *req)
1334 {
1335 	int rc = req->r_result;
1336 	struct inode *inode = req->r_inode;
1337 	struct ceph_aio_request *aio_req = req->r_priv;
1338 	struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
1339 	struct ceph_osd_req_op *op = &req->r_ops[0];
1340 	struct ceph_client_metric *metric = &ceph_sb_to_mdsc(inode->i_sb)->metric;
1341 	unsigned int len = osd_data->bvec_pos.iter.bi_size;
1342 	bool sparse = (op->op == CEPH_OSD_OP_SPARSE_READ);
1343 	struct ceph_client *cl = ceph_inode_to_client(inode);
1344 
1345 	BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_BVECS);
1346 	BUG_ON(!osd_data->num_bvecs);
1347 
1348 	doutc(cl, "req %p inode %p %llx.%llx, rc %d bytes %u\n", req,
1349 	      inode, ceph_vinop(inode), rc, len);
1350 
1351 	if (rc == -EOLDSNAPC) {
1352 		struct ceph_aio_work *aio_work;
1353 		BUG_ON(!aio_req->write);
1354 
1355 		aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
1356 		if (aio_work) {
1357 			INIT_WORK(&aio_work->work, ceph_aio_retry_work);
1358 			aio_work->req = req;
1359 			queue_work(ceph_inode_to_fs_client(inode)->inode_wq,
1360 				   &aio_work->work);
1361 			return;
1362 		}
1363 		rc = -ENOMEM;
1364 	} else if (!aio_req->write) {
1365 		if (sparse && rc >= 0)
1366 			rc = ceph_sparse_ext_map_end(op);
1367 		if (rc == -ENOENT)
1368 			rc = 0;
1369 		if (rc >= 0 && len > rc) {
1370 			struct iov_iter i;
1371 			int zlen = len - rc;
1372 
1373 			/*
1374 			 * If read is satisfied by single OSD request,
1375 			 * it can pass EOF. Otherwise read is within
1376 			 * i_size.
1377 			 */
1378 			if (aio_req->num_reqs == 1) {
1379 				loff_t i_size = i_size_read(inode);
1380 				loff_t endoff = aio_req->iocb->ki_pos + rc;
1381 				if (endoff < i_size)
1382 					zlen = min_t(size_t, zlen,
1383 						     i_size - endoff);
1384 				aio_req->total_len = rc + zlen;
1385 			}
1386 
1387 			iov_iter_bvec(&i, ITER_DEST, osd_data->bvec_pos.bvecs,
1388 				      osd_data->num_bvecs, len);
1389 			iov_iter_advance(&i, rc);
1390 			iov_iter_zero(zlen, &i);
1391 		}
1392 	}
1393 
1394 	/* r_start_latency == 0 means the request was not submitted */
1395 	if (req->r_start_latency) {
1396 		if (aio_req->write)
1397 			ceph_update_write_metrics(metric, req->r_start_latency,
1398 						  req->r_end_latency, len, rc);
1399 		else
1400 			ceph_update_read_metrics(metric, req->r_start_latency,
1401 						 req->r_end_latency, len, rc);
1402 	}
1403 
1404 	put_bvecs(osd_data->bvec_pos.bvecs, osd_data->num_bvecs,
1405 		  aio_req->should_dirty);
1406 	ceph_osdc_put_request(req);
1407 
1408 	if (rc < 0)
1409 		cmpxchg(&aio_req->error, 0, rc);
1410 
1411 	ceph_aio_complete(inode, aio_req);
1412 	return;
1413 }
1414 
1415 static void ceph_aio_retry_work(struct work_struct *work)
1416 {
1417 	struct ceph_aio_work *aio_work =
1418 		container_of(work, struct ceph_aio_work, work);
1419 	struct ceph_osd_request *orig_req = aio_work->req;
1420 	struct ceph_aio_request *aio_req = orig_req->r_priv;
1421 	struct inode *inode = orig_req->r_inode;
1422 	struct ceph_inode_info *ci = ceph_inode(inode);
1423 	struct ceph_snap_context *snapc;
1424 	struct ceph_osd_request *req;
1425 	int ret;
1426 
1427 	spin_lock(&ci->i_ceph_lock);
1428 	if (__ceph_have_pending_cap_snap(ci)) {
1429 		struct ceph_cap_snap *capsnap =
1430 			list_last_entry(&ci->i_cap_snaps,
1431 					struct ceph_cap_snap,
1432 					ci_item);
1433 		snapc = ceph_get_snap_context(capsnap->context);
1434 	} else {
1435 		BUG_ON(!ci->i_head_snapc);
1436 		snapc = ceph_get_snap_context(ci->i_head_snapc);
1437 	}
1438 	spin_unlock(&ci->i_ceph_lock);
1439 
1440 	req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 1,
1441 			false, GFP_NOFS);
1442 	if (!req) {
1443 		ret = -ENOMEM;
1444 		req = orig_req;
1445 		goto out;
1446 	}
1447 
1448 	req->r_flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1449 	ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
1450 	ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
1451 
1452 	req->r_ops[0] = orig_req->r_ops[0];
1453 
1454 	req->r_mtime = aio_req->mtime;
1455 	req->r_data_offset = req->r_ops[0].extent.offset;
1456 
1457 	ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
1458 	if (ret) {
1459 		ceph_osdc_put_request(req);
1460 		req = orig_req;
1461 		goto out;
1462 	}
1463 
1464 	ceph_osdc_put_request(orig_req);
1465 
1466 	req->r_callback = ceph_aio_complete_req;
1467 	req->r_inode = inode;
1468 	req->r_priv = aio_req;
1469 
1470 	ceph_osdc_start_request(req->r_osdc, req);
1471 out:
1472 	if (ret < 0) {
1473 		req->r_result = ret;
1474 		ceph_aio_complete_req(req);
1475 	}
1476 
1477 	ceph_put_snap_context(snapc);
1478 	kfree(aio_work);
1479 }
1480 
1481 static ssize_t
1482 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
1483 		       struct ceph_snap_context *snapc,
1484 		       struct ceph_cap_flush **pcf)
1485 {
1486 	struct file *file = iocb->ki_filp;
1487 	struct inode *inode = file_inode(file);
1488 	struct ceph_inode_info *ci = ceph_inode(inode);
1489 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
1490 	struct ceph_client *cl = fsc->client;
1491 	struct ceph_client_metric *metric = &fsc->mdsc->metric;
1492 	struct ceph_vino vino;
1493 	struct ceph_osd_request *req;
1494 	struct bio_vec *bvecs;
1495 	struct ceph_aio_request *aio_req = NULL;
1496 	int num_pages = 0;
1497 	int flags;
1498 	int ret = 0;
1499 	struct timespec64 mtime = current_time(inode);
1500 	size_t count = iov_iter_count(iter);
1501 	loff_t pos = iocb->ki_pos;
1502 	bool write = iov_iter_rw(iter) == WRITE;
1503 	bool should_dirty = !write && user_backed_iter(iter);
1504 	bool sparse = ceph_test_mount_opt(fsc, SPARSEREAD);
1505 
1506 	if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1507 		return -EROFS;
1508 
1509 	doutc(cl, "sync_direct_%s on file %p %lld~%u snapc %p seq %lld\n",
1510 	      (write ? "write" : "read"), file, pos, (unsigned)count,
1511 	      snapc, snapc ? snapc->seq : 0);
1512 
1513 	if (write) {
1514 		int ret2;
1515 
1516 		ceph_fscache_invalidate(inode, true);
1517 
1518 		ret2 = invalidate_inode_pages2_range(inode->i_mapping,
1519 					pos >> PAGE_SHIFT,
1520 					(pos + count - 1) >> PAGE_SHIFT);
1521 		if (ret2 < 0)
1522 			doutc(cl, "invalidate_inode_pages2_range returned %d\n",
1523 			      ret2);
1524 
1525 		flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1526 	} else {
1527 		flags = CEPH_OSD_FLAG_READ;
1528 	}
1529 
1530 	while (iov_iter_count(iter) > 0) {
1531 		u64 size = iov_iter_count(iter);
1532 		ssize_t len;
1533 		struct ceph_osd_req_op *op;
1534 		int readop = sparse ? CEPH_OSD_OP_SPARSE_READ : CEPH_OSD_OP_READ;
1535 		int extent_cnt;
1536 
1537 		if (write)
1538 			size = min_t(u64, size, fsc->mount_options->wsize);
1539 		else
1540 			size = min_t(u64, size, fsc->mount_options->rsize);
1541 
1542 		vino = ceph_vino(inode);
1543 		req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1544 					    vino, pos, &size, 0,
1545 					    1,
1546 					    write ? CEPH_OSD_OP_WRITE : readop,
1547 					    flags, snapc,
1548 					    ci->i_truncate_seq,
1549 					    ci->i_truncate_size,
1550 					    false);
1551 		if (IS_ERR(req)) {
1552 			ret = PTR_ERR(req);
1553 			break;
1554 		}
1555 
1556 		len = iter_get_bvecs_alloc(iter, size, &bvecs, &num_pages);
1557 		if (len < 0) {
1558 			ceph_osdc_put_request(req);
1559 			ret = len;
1560 			break;
1561 		}
1562 		if (len != size)
1563 			osd_req_op_extent_update(req, 0, len);
1564 
1565 		/*
1566 		 * To simplify error handling, allow AIO when IO within i_size
1567 		 * or IO can be satisfied by single OSD request.
1568 		 */
1569 		if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
1570 		    (len == count || pos + count <= i_size_read(inode))) {
1571 			aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
1572 			if (aio_req) {
1573 				aio_req->iocb = iocb;
1574 				aio_req->write = write;
1575 				aio_req->should_dirty = should_dirty;
1576 				INIT_LIST_HEAD(&aio_req->osd_reqs);
1577 				if (write) {
1578 					aio_req->mtime = mtime;
1579 					swap(aio_req->prealloc_cf, *pcf);
1580 				}
1581 			}
1582 			/* ignore error */
1583 		}
1584 
1585 		if (write) {
1586 			/*
1587 			 * throw out any page cache pages in this range. this
1588 			 * may block.
1589 			 */
1590 			truncate_inode_pages_range(inode->i_mapping, pos,
1591 						   PAGE_ALIGN(pos + len) - 1);
1592 
1593 			req->r_mtime = mtime;
1594 		}
1595 
1596 		osd_req_op_extent_osd_data_bvecs(req, 0, bvecs, num_pages, len);
1597 		op = &req->r_ops[0];
1598 		if (sparse) {
1599 			extent_cnt = __ceph_sparse_read_ext_count(inode, size);
1600 			ret = ceph_alloc_sparse_ext_map(op, extent_cnt);
1601 			if (ret) {
1602 				ceph_osdc_put_request(req);
1603 				break;
1604 			}
1605 		}
1606 
1607 		if (aio_req) {
1608 			aio_req->total_len += len;
1609 			aio_req->num_reqs++;
1610 			atomic_inc(&aio_req->pending_reqs);
1611 
1612 			req->r_callback = ceph_aio_complete_req;
1613 			req->r_inode = inode;
1614 			req->r_priv = aio_req;
1615 			list_add_tail(&req->r_private_item, &aio_req->osd_reqs);
1616 
1617 			pos += len;
1618 			continue;
1619 		}
1620 
1621 		ceph_osdc_start_request(req->r_osdc, req);
1622 		ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1623 
1624 		if (write)
1625 			ceph_update_write_metrics(metric, req->r_start_latency,
1626 						  req->r_end_latency, len, ret);
1627 		else
1628 			ceph_update_read_metrics(metric, req->r_start_latency,
1629 						 req->r_end_latency, len, ret);
1630 
1631 		size = i_size_read(inode);
1632 		if (!write) {
1633 			if (sparse && ret >= 0)
1634 				ret = ceph_sparse_ext_map_end(op);
1635 			else if (ret == -ENOENT)
1636 				ret = 0;
1637 
1638 			if (ret >= 0 && ret < len && pos + ret < size) {
1639 				struct iov_iter i;
1640 				int zlen = min_t(size_t, len - ret,
1641 						 size - pos - ret);
1642 
1643 				iov_iter_bvec(&i, ITER_DEST, bvecs, num_pages, len);
1644 				iov_iter_advance(&i, ret);
1645 				iov_iter_zero(zlen, &i);
1646 				ret += zlen;
1647 			}
1648 			if (ret >= 0)
1649 				len = ret;
1650 		}
1651 
1652 		put_bvecs(bvecs, num_pages, should_dirty);
1653 		ceph_osdc_put_request(req);
1654 		if (ret < 0)
1655 			break;
1656 
1657 		pos += len;
1658 		if (!write && pos >= size)
1659 			break;
1660 
1661 		if (write && pos > size) {
1662 			if (ceph_inode_set_size(inode, pos))
1663 				ceph_check_caps(ceph_inode(inode),
1664 						CHECK_CAPS_AUTHONLY);
1665 		}
1666 	}
1667 
1668 	if (aio_req) {
1669 		LIST_HEAD(osd_reqs);
1670 
1671 		if (aio_req->num_reqs == 0) {
1672 			kfree(aio_req);
1673 			return ret;
1674 		}
1675 
1676 		ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1677 					      CEPH_CAP_FILE_RD);
1678 
1679 		list_splice(&aio_req->osd_reqs, &osd_reqs);
1680 		inode_dio_begin(inode);
1681 		while (!list_empty(&osd_reqs)) {
1682 			req = list_first_entry(&osd_reqs,
1683 					       struct ceph_osd_request,
1684 					       r_private_item);
1685 			list_del_init(&req->r_private_item);
1686 			if (ret >= 0)
1687 				ceph_osdc_start_request(req->r_osdc, req);
1688 			if (ret < 0) {
1689 				req->r_result = ret;
1690 				ceph_aio_complete_req(req);
1691 			}
1692 		}
1693 		return -EIOCBQUEUED;
1694 	}
1695 
1696 	if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1697 		ret = pos - iocb->ki_pos;
1698 		iocb->ki_pos = pos;
1699 	}
1700 	return ret;
1701 }
1702 
1703 /*
1704  * Synchronous write, straight from __user pointer or user pages.
1705  *
1706  * If write spans object boundary, just do multiple writes.  (For a
1707  * correct atomic write, we should e.g. take write locks on all
1708  * objects, rollback on failure, etc.)
1709  */
1710 static ssize_t
1711 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1712 		struct ceph_snap_context *snapc)
1713 {
1714 	struct file *file = iocb->ki_filp;
1715 	struct inode *inode = file_inode(file);
1716 	struct ceph_inode_info *ci = ceph_inode(inode);
1717 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
1718 	struct ceph_client *cl = fsc->client;
1719 	struct ceph_osd_client *osdc = &fsc->client->osdc;
1720 	struct ceph_osd_request *req;
1721 	struct page **pages;
1722 	u64 len;
1723 	int num_pages;
1724 	int written = 0;
1725 	int ret;
1726 	bool check_caps = false;
1727 	struct timespec64 mtime = current_time(inode);
1728 	size_t count = iov_iter_count(from);
1729 
1730 	if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1731 		return -EROFS;
1732 
1733 	doutc(cl, "on file %p %lld~%u snapc %p seq %lld\n", file, pos,
1734 	      (unsigned)count, snapc, snapc->seq);
1735 
1736 	ret = filemap_write_and_wait_range(inode->i_mapping,
1737 					   pos, pos + count - 1);
1738 	if (ret < 0)
1739 		return ret;
1740 
1741 	ceph_fscache_invalidate(inode, false);
1742 
1743 	while ((len = iov_iter_count(from)) > 0) {
1744 		size_t left;
1745 		int n;
1746 		u64 write_pos = pos;
1747 		u64 write_len = len;
1748 		u64 objnum, objoff;
1749 		u32 xlen;
1750 		u64 assert_ver = 0;
1751 		bool rmw;
1752 		bool first, last;
1753 		struct iov_iter saved_iter = *from;
1754 		size_t off;
1755 
1756 		ceph_fscrypt_adjust_off_and_len(inode, &write_pos, &write_len);
1757 
1758 		/* clamp the length to the end of first object */
1759 		ceph_calc_file_object_mapping(&ci->i_layout, write_pos,
1760 					      write_len, &objnum, &objoff,
1761 					      &xlen);
1762 		write_len = xlen;
1763 
1764 		/* adjust len downward if it goes beyond current object */
1765 		if (pos + len > write_pos + write_len)
1766 			len = write_pos + write_len - pos;
1767 
1768 		/*
1769 		 * If we had to adjust the length or position to align with a
1770 		 * crypto block, then we must do a read/modify/write cycle. We
1771 		 * use a version assertion to redrive the thing if something
1772 		 * changes in between.
1773 		 */
1774 		first = pos != write_pos;
1775 		last = (pos + len) != (write_pos + write_len);
1776 		rmw = first || last;
1777 
1778 		doutc(cl, "ino %llx %lld~%llu adjusted %lld~%llu -- %srmw\n",
1779 		      ci->i_vino.ino, pos, len, write_pos, write_len,
1780 		      rmw ? "" : "no ");
1781 
1782 		/*
1783 		 * The data is emplaced into the page as it would be if it were
1784 		 * in an array of pagecache pages.
1785 		 */
1786 		num_pages = calc_pages_for(write_pos, write_len);
1787 		pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1788 		if (IS_ERR(pages)) {
1789 			ret = PTR_ERR(pages);
1790 			break;
1791 		}
1792 
1793 		/* Do we need to preload the pages? */
1794 		if (rmw) {
1795 			u64 first_pos = write_pos;
1796 			u64 last_pos = (write_pos + write_len) - CEPH_FSCRYPT_BLOCK_SIZE;
1797 			u64 read_len = CEPH_FSCRYPT_BLOCK_SIZE;
1798 			struct ceph_osd_req_op *op;
1799 
1800 			/* We should only need to do this for encrypted inodes */
1801 			WARN_ON_ONCE(!IS_ENCRYPTED(inode));
1802 
1803 			/* No need to do two reads if first and last blocks are same */
1804 			if (first && last_pos == first_pos)
1805 				last = false;
1806 
1807 			/*
1808 			 * Allocate a read request for one or two extents,
1809 			 * depending on how the request was aligned.
1810 			 */
1811 			req = ceph_osdc_new_request(osdc, &ci->i_layout,
1812 					ci->i_vino, first ? first_pos : last_pos,
1813 					&read_len, 0, (first && last) ? 2 : 1,
1814 					CEPH_OSD_OP_SPARSE_READ, CEPH_OSD_FLAG_READ,
1815 					NULL, ci->i_truncate_seq,
1816 					ci->i_truncate_size, false);
1817 			if (IS_ERR(req)) {
1818 				ceph_release_page_vector(pages, num_pages);
1819 				ret = PTR_ERR(req);
1820 				break;
1821 			}
1822 
1823 			/* Something is misaligned! */
1824 			if (read_len != CEPH_FSCRYPT_BLOCK_SIZE) {
1825 				ceph_osdc_put_request(req);
1826 				ceph_release_page_vector(pages, num_pages);
1827 				ret = -EIO;
1828 				break;
1829 			}
1830 
1831 			/* Add extent for first block? */
1832 			op = &req->r_ops[0];
1833 
1834 			if (first) {
1835 				osd_req_op_extent_osd_data_pages(req, 0, pages,
1836 							 CEPH_FSCRYPT_BLOCK_SIZE,
1837 							 offset_in_page(first_pos),
1838 							 false, false);
1839 				/* We only expect a single extent here */
1840 				ret = __ceph_alloc_sparse_ext_map(op, 1);
1841 				if (ret) {
1842 					ceph_osdc_put_request(req);
1843 					ceph_release_page_vector(pages, num_pages);
1844 					break;
1845 				}
1846 			}
1847 
1848 			/* Add extent for last block */
1849 			if (last) {
1850 				/* Init the other extent if first extent has been used */
1851 				if (first) {
1852 					op = &req->r_ops[1];
1853 					osd_req_op_extent_init(req, 1,
1854 							CEPH_OSD_OP_SPARSE_READ,
1855 							last_pos, CEPH_FSCRYPT_BLOCK_SIZE,
1856 							ci->i_truncate_size,
1857 							ci->i_truncate_seq);
1858 				}
1859 
1860 				ret = __ceph_alloc_sparse_ext_map(op, 1);
1861 				if (ret) {
1862 					ceph_osdc_put_request(req);
1863 					ceph_release_page_vector(pages, num_pages);
1864 					break;
1865 				}
1866 
1867 				osd_req_op_extent_osd_data_pages(req, first ? 1 : 0,
1868 							&pages[num_pages - 1],
1869 							CEPH_FSCRYPT_BLOCK_SIZE,
1870 							offset_in_page(last_pos),
1871 							false, false);
1872 			}
1873 
1874 			ceph_osdc_start_request(osdc, req);
1875 			ret = ceph_osdc_wait_request(osdc, req);
1876 
1877 			/* FIXME: length field is wrong if there are 2 extents */
1878 			ceph_update_read_metrics(&fsc->mdsc->metric,
1879 						 req->r_start_latency,
1880 						 req->r_end_latency,
1881 						 read_len, ret);
1882 
1883 			/* Ok if object is not already present */
1884 			if (ret == -ENOENT) {
1885 				/*
1886 				 * If there is no object, then we can't assert
1887 				 * on its version. Set it to 0, and we'll use an
1888 				 * exclusive create instead.
1889 				 */
1890 				ceph_osdc_put_request(req);
1891 				ret = 0;
1892 
1893 				/*
1894 				 * zero out the soon-to-be uncopied parts of the
1895 				 * first and last pages.
1896 				 */
1897 				if (first)
1898 					zero_user_segment(pages[0], 0,
1899 							  offset_in_page(first_pos));
1900 				if (last)
1901 					zero_user_segment(pages[num_pages - 1],
1902 							  offset_in_page(last_pos),
1903 							  PAGE_SIZE);
1904 			} else {
1905 				if (ret < 0) {
1906 					ceph_osdc_put_request(req);
1907 					ceph_release_page_vector(pages, num_pages);
1908 					break;
1909 				}
1910 
1911 				op = &req->r_ops[0];
1912 				if (op->extent.sparse_ext_cnt == 0) {
1913 					if (first)
1914 						zero_user_segment(pages[0], 0,
1915 								  offset_in_page(first_pos));
1916 					else
1917 						zero_user_segment(pages[num_pages - 1],
1918 								  offset_in_page(last_pos),
1919 								  PAGE_SIZE);
1920 				} else if (op->extent.sparse_ext_cnt != 1 ||
1921 					   ceph_sparse_ext_map_end(op) !=
1922 						CEPH_FSCRYPT_BLOCK_SIZE) {
1923 					ret = -EIO;
1924 					ceph_osdc_put_request(req);
1925 					ceph_release_page_vector(pages, num_pages);
1926 					break;
1927 				}
1928 
1929 				if (first && last) {
1930 					op = &req->r_ops[1];
1931 					if (op->extent.sparse_ext_cnt == 0) {
1932 						zero_user_segment(pages[num_pages - 1],
1933 								  offset_in_page(last_pos),
1934 								  PAGE_SIZE);
1935 					} else if (op->extent.sparse_ext_cnt != 1 ||
1936 						   ceph_sparse_ext_map_end(op) !=
1937 							CEPH_FSCRYPT_BLOCK_SIZE) {
1938 						ret = -EIO;
1939 						ceph_osdc_put_request(req);
1940 						ceph_release_page_vector(pages, num_pages);
1941 						break;
1942 					}
1943 				}
1944 
1945 				/* Grab assert version. It must be non-zero. */
1946 				assert_ver = req->r_version;
1947 				WARN_ON_ONCE(ret > 0 && assert_ver == 0);
1948 
1949 				ceph_osdc_put_request(req);
1950 				if (first) {
1951 					ret = ceph_fscrypt_decrypt_block_inplace(inode,
1952 							pages[0], CEPH_FSCRYPT_BLOCK_SIZE,
1953 							offset_in_page(first_pos),
1954 							first_pos >> CEPH_FSCRYPT_BLOCK_SHIFT);
1955 					if (ret < 0) {
1956 						ceph_release_page_vector(pages, num_pages);
1957 						break;
1958 					}
1959 				}
1960 				if (last) {
1961 					ret = ceph_fscrypt_decrypt_block_inplace(inode,
1962 							pages[num_pages - 1],
1963 							CEPH_FSCRYPT_BLOCK_SIZE,
1964 							offset_in_page(last_pos),
1965 							last_pos >> CEPH_FSCRYPT_BLOCK_SHIFT);
1966 					if (ret < 0) {
1967 						ceph_release_page_vector(pages, num_pages);
1968 						break;
1969 					}
1970 				}
1971 			}
1972 		}
1973 
1974 		left = len;
1975 		off = offset_in_page(pos);
1976 		for (n = 0; n < num_pages; n++) {
1977 			size_t plen = min_t(size_t, left, PAGE_SIZE - off);
1978 
1979 			/* copy the data */
1980 			ret = copy_page_from_iter(pages[n], off, plen, from);
1981 			if (ret != plen) {
1982 				ret = -EFAULT;
1983 				break;
1984 			}
1985 			off = 0;
1986 			left -= ret;
1987 		}
1988 		if (ret < 0) {
1989 			doutc(cl, "write failed with %d\n", ret);
1990 			ceph_release_page_vector(pages, num_pages);
1991 			break;
1992 		}
1993 
1994 		if (IS_ENCRYPTED(inode)) {
1995 			ret = ceph_fscrypt_encrypt_pages(inode, pages,
1996 							 write_pos, write_len,
1997 							 GFP_KERNEL);
1998 			if (ret < 0) {
1999 				doutc(cl, "encryption failed with %d\n", ret);
2000 				ceph_release_page_vector(pages, num_pages);
2001 				break;
2002 			}
2003 		}
2004 
2005 		req = ceph_osdc_new_request(osdc, &ci->i_layout,
2006 					    ci->i_vino, write_pos, &write_len,
2007 					    rmw ? 1 : 0, rmw ? 2 : 1,
2008 					    CEPH_OSD_OP_WRITE,
2009 					    CEPH_OSD_FLAG_WRITE,
2010 					    snapc, ci->i_truncate_seq,
2011 					    ci->i_truncate_size, false);
2012 		if (IS_ERR(req)) {
2013 			ret = PTR_ERR(req);
2014 			ceph_release_page_vector(pages, num_pages);
2015 			break;
2016 		}
2017 
2018 		doutc(cl, "write op %lld~%llu\n", write_pos, write_len);
2019 		osd_req_op_extent_osd_data_pages(req, rmw ? 1 : 0, pages, write_len,
2020 						 offset_in_page(write_pos), false,
2021 						 true);
2022 		req->r_inode = inode;
2023 		req->r_mtime = mtime;
2024 
2025 		/* Set up the assertion */
2026 		if (rmw) {
2027 			/*
2028 			 * Set up the assertion. If we don't have a version
2029 			 * number, then the object doesn't exist yet. Use an
2030 			 * exclusive create instead of a version assertion in
2031 			 * that case.
2032 			 */
2033 			if (assert_ver) {
2034 				osd_req_op_init(req, 0, CEPH_OSD_OP_ASSERT_VER, 0);
2035 				req->r_ops[0].assert_ver.ver = assert_ver;
2036 			} else {
2037 				osd_req_op_init(req, 0, CEPH_OSD_OP_CREATE,
2038 						CEPH_OSD_OP_FLAG_EXCL);
2039 			}
2040 		}
2041 
2042 		ceph_osdc_start_request(osdc, req);
2043 		ret = ceph_osdc_wait_request(osdc, req);
2044 
2045 		ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
2046 					  req->r_end_latency, len, ret);
2047 		ceph_osdc_put_request(req);
2048 		if (ret != 0) {
2049 			doutc(cl, "osd write returned %d\n", ret);
2050 			/* Version changed! Must re-do the rmw cycle */
2051 			if ((assert_ver && (ret == -ERANGE || ret == -EOVERFLOW)) ||
2052 			    (!assert_ver && ret == -EEXIST)) {
2053 				/* We should only ever see this on a rmw */
2054 				WARN_ON_ONCE(!rmw);
2055 
2056 				/* The version should never go backward */
2057 				WARN_ON_ONCE(ret == -EOVERFLOW);
2058 
2059 				*from = saved_iter;
2060 
2061 				/* FIXME: limit number of times we loop? */
2062 				continue;
2063 			}
2064 			ceph_set_error_write(ci);
2065 			break;
2066 		}
2067 
2068 		ceph_clear_error_write(ci);
2069 
2070 		/*
2071 		 * We successfully wrote to a range of the file. Declare
2072 		 * that region of the pagecache invalid.
2073 		 */
2074 		ret = invalidate_inode_pages2_range(
2075 				inode->i_mapping,
2076 				pos >> PAGE_SHIFT,
2077 				(pos + len - 1) >> PAGE_SHIFT);
2078 		if (ret < 0) {
2079 			doutc(cl, "invalidate_inode_pages2_range returned %d\n",
2080 			      ret);
2081 			ret = 0;
2082 		}
2083 		pos += len;
2084 		written += len;
2085 		doutc(cl, "written %d\n", written);
2086 		if (pos > i_size_read(inode)) {
2087 			check_caps = ceph_inode_set_size(inode, pos);
2088 			if (check_caps)
2089 				ceph_check_caps(ceph_inode(inode),
2090 						CHECK_CAPS_AUTHONLY);
2091 		}
2092 
2093 	}
2094 
2095 	if (ret != -EOLDSNAPC && written > 0) {
2096 		ret = written;
2097 		iocb->ki_pos = pos;
2098 	}
2099 	doutc(cl, "returning %d\n", ret);
2100 	return ret;
2101 }
2102 
2103 /*
2104  * Wrap generic_file_aio_read with checks for cap bits on the inode.
2105  * Atomically grab references, so that those bits are not released
2106  * back to the MDS mid-read.
2107  *
2108  * Hmm, the sync read case isn't actually async... should it be?
2109  */
2110 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
2111 {
2112 	struct file *filp = iocb->ki_filp;
2113 	struct ceph_file_info *fi = filp->private_data;
2114 	size_t len = iov_iter_count(to);
2115 	struct inode *inode = file_inode(filp);
2116 	struct ceph_inode_info *ci = ceph_inode(inode);
2117 	bool direct_lock = iocb->ki_flags & IOCB_DIRECT;
2118 	struct ceph_client *cl = ceph_inode_to_client(inode);
2119 	ssize_t ret;
2120 	int want = 0, got = 0;
2121 	int retry_op = 0, read = 0;
2122 
2123 again:
2124 	doutc(cl, "%llu~%u trying to get caps on %p %llx.%llx\n",
2125 	      iocb->ki_pos, (unsigned)len, inode, ceph_vinop(inode));
2126 
2127 	if (ceph_inode_is_shutdown(inode))
2128 		return -ESTALE;
2129 
2130 	if (direct_lock)
2131 		ceph_start_io_direct(inode);
2132 	else
2133 		ceph_start_io_read(inode);
2134 
2135 	if (!(fi->flags & CEPH_F_SYNC) && !direct_lock)
2136 		want |= CEPH_CAP_FILE_CACHE;
2137 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
2138 		want |= CEPH_CAP_FILE_LAZYIO;
2139 
2140 	ret = ceph_get_caps(filp, CEPH_CAP_FILE_RD, want, -1, &got);
2141 	if (ret < 0) {
2142 		if (direct_lock)
2143 			ceph_end_io_direct(inode);
2144 		else
2145 			ceph_end_io_read(inode);
2146 		return ret;
2147 	}
2148 
2149 	if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
2150 	    (iocb->ki_flags & IOCB_DIRECT) ||
2151 	    (fi->flags & CEPH_F_SYNC)) {
2152 
2153 		doutc(cl, "sync %p %llx.%llx %llu~%u got cap refs on %s\n",
2154 		      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
2155 		      ceph_cap_string(got));
2156 
2157 		if (!ceph_has_inline_data(ci)) {
2158 			if (!retry_op &&
2159 			    (iocb->ki_flags & IOCB_DIRECT) &&
2160 			    !IS_ENCRYPTED(inode)) {
2161 				ret = ceph_direct_read_write(iocb, to,
2162 							     NULL, NULL);
2163 				if (ret >= 0 && ret < len)
2164 					retry_op = CHECK_EOF;
2165 			} else {
2166 				ret = ceph_sync_read(iocb, to, &retry_op);
2167 			}
2168 		} else {
2169 			retry_op = READ_INLINE;
2170 		}
2171 	} else {
2172 		CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
2173 		doutc(cl, "async %p %llx.%llx %llu~%u got cap refs on %s\n",
2174 		      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
2175 		      ceph_cap_string(got));
2176 		ceph_add_rw_context(fi, &rw_ctx);
2177 		ret = generic_file_read_iter(iocb, to);
2178 		ceph_del_rw_context(fi, &rw_ctx);
2179 	}
2180 
2181 	doutc(cl, "%p %llx.%llx dropping cap refs on %s = %d\n",
2182 	      inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
2183 	ceph_put_cap_refs(ci, got);
2184 
2185 	if (direct_lock)
2186 		ceph_end_io_direct(inode);
2187 	else
2188 		ceph_end_io_read(inode);
2189 
2190 	if (retry_op > HAVE_RETRIED && ret >= 0) {
2191 		int statret;
2192 		struct page *page = NULL;
2193 		loff_t i_size;
2194 		int mask = CEPH_STAT_CAP_SIZE;
2195 		if (retry_op == READ_INLINE) {
2196 			page = __page_cache_alloc(GFP_KERNEL);
2197 			if (!page)
2198 				return -ENOMEM;
2199 
2200 			mask = CEPH_STAT_CAP_INLINE_DATA;
2201 		}
2202 
2203 		statret = __ceph_do_getattr(inode, page, mask, !!page);
2204 		if (statret < 0) {
2205 			if (page)
2206 				__free_page(page);
2207 			if (statret == -ENODATA) {
2208 				BUG_ON(retry_op != READ_INLINE);
2209 				goto again;
2210 			}
2211 			return statret;
2212 		}
2213 
2214 		i_size = i_size_read(inode);
2215 		if (retry_op == READ_INLINE) {
2216 			BUG_ON(ret > 0 || read > 0);
2217 			if (iocb->ki_pos < i_size &&
2218 			    iocb->ki_pos < PAGE_SIZE) {
2219 				loff_t end = min_t(loff_t, i_size,
2220 						   iocb->ki_pos + len);
2221 				end = min_t(loff_t, end, PAGE_SIZE);
2222 				if (statret < end)
2223 					zero_user_segment(page, statret, end);
2224 				ret = copy_page_to_iter(page,
2225 						iocb->ki_pos & ~PAGE_MASK,
2226 						end - iocb->ki_pos, to);
2227 				iocb->ki_pos += ret;
2228 				read += ret;
2229 			}
2230 			if (iocb->ki_pos < i_size && read < len) {
2231 				size_t zlen = min_t(size_t, len - read,
2232 						    i_size - iocb->ki_pos);
2233 				ret = iov_iter_zero(zlen, to);
2234 				iocb->ki_pos += ret;
2235 				read += ret;
2236 			}
2237 			__free_pages(page, 0);
2238 			return read;
2239 		}
2240 
2241 		/* hit EOF or hole? */
2242 		if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
2243 		    ret < len) {
2244 			doutc(cl, "may hit hole, ppos %lld < size %lld, reading more\n",
2245 			      iocb->ki_pos, i_size);
2246 
2247 			read += ret;
2248 			len -= ret;
2249 			retry_op = HAVE_RETRIED;
2250 			goto again;
2251 		}
2252 	}
2253 
2254 	if (ret >= 0)
2255 		ret += read;
2256 
2257 	return ret;
2258 }
2259 
2260 /*
2261  * Wrap filemap_splice_read with checks for cap bits on the inode.
2262  * Atomically grab references, so that those bits are not released
2263  * back to the MDS mid-read.
2264  */
2265 static ssize_t ceph_splice_read(struct file *in, loff_t *ppos,
2266 				struct pipe_inode_info *pipe,
2267 				size_t len, unsigned int flags)
2268 {
2269 	struct ceph_file_info *fi = in->private_data;
2270 	struct inode *inode = file_inode(in);
2271 	struct ceph_inode_info *ci = ceph_inode(inode);
2272 	ssize_t ret;
2273 	int want = 0, got = 0;
2274 	CEPH_DEFINE_RW_CONTEXT(rw_ctx, 0);
2275 
2276 	dout("splice_read %p %llx.%llx %llu~%zu trying to get caps on %p\n",
2277 	     inode, ceph_vinop(inode), *ppos, len, inode);
2278 
2279 	if (ceph_inode_is_shutdown(inode))
2280 		return -ESTALE;
2281 
2282 	if (ceph_has_inline_data(ci) ||
2283 	    (fi->flags & CEPH_F_SYNC))
2284 		return copy_splice_read(in, ppos, pipe, len, flags);
2285 
2286 	ceph_start_io_read(inode);
2287 
2288 	want = CEPH_CAP_FILE_CACHE;
2289 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
2290 		want |= CEPH_CAP_FILE_LAZYIO;
2291 
2292 	ret = ceph_get_caps(in, CEPH_CAP_FILE_RD, want, -1, &got);
2293 	if (ret < 0)
2294 		goto out_end;
2295 
2296 	if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) == 0) {
2297 		dout("splice_read/sync %p %llx.%llx %llu~%zu got cap refs on %s\n",
2298 		     inode, ceph_vinop(inode), *ppos, len,
2299 		     ceph_cap_string(got));
2300 
2301 		ceph_put_cap_refs(ci, got);
2302 		ceph_end_io_read(inode);
2303 		return copy_splice_read(in, ppos, pipe, len, flags);
2304 	}
2305 
2306 	dout("splice_read %p %llx.%llx %llu~%zu got cap refs on %s\n",
2307 	     inode, ceph_vinop(inode), *ppos, len, ceph_cap_string(got));
2308 
2309 	rw_ctx.caps = got;
2310 	ceph_add_rw_context(fi, &rw_ctx);
2311 	ret = filemap_splice_read(in, ppos, pipe, len, flags);
2312 	ceph_del_rw_context(fi, &rw_ctx);
2313 
2314 	dout("splice_read %p %llx.%llx dropping cap refs on %s = %zd\n",
2315 	     inode, ceph_vinop(inode), ceph_cap_string(got), ret);
2316 
2317 	ceph_put_cap_refs(ci, got);
2318 out_end:
2319 	ceph_end_io_read(inode);
2320 	return ret;
2321 }
2322 
2323 /*
2324  * Take cap references to avoid releasing caps to MDS mid-write.
2325  *
2326  * If we are synchronous, and write with an old snap context, the OSD
2327  * may return EOLDSNAPC.  In that case, retry the write.. _after_
2328  * dropping our cap refs and allowing the pending snap to logically
2329  * complete _before_ this write occurs.
2330  *
2331  * If we are near ENOSPC, write synchronously.
2332  */
2333 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
2334 {
2335 	struct file *file = iocb->ki_filp;
2336 	struct ceph_file_info *fi = file->private_data;
2337 	struct inode *inode = file_inode(file);
2338 	struct ceph_inode_info *ci = ceph_inode(inode);
2339 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
2340 	struct ceph_client *cl = fsc->client;
2341 	struct ceph_osd_client *osdc = &fsc->client->osdc;
2342 	struct ceph_cap_flush *prealloc_cf;
2343 	ssize_t count, written = 0;
2344 	int err, want = 0, got;
2345 	bool direct_lock = false;
2346 	u32 map_flags;
2347 	u64 pool_flags;
2348 	loff_t pos;
2349 	loff_t limit = max(i_size_read(inode), fsc->max_file_size);
2350 
2351 	if (ceph_inode_is_shutdown(inode))
2352 		return -ESTALE;
2353 
2354 	if (ceph_snap(inode) != CEPH_NOSNAP)
2355 		return -EROFS;
2356 
2357 	prealloc_cf = ceph_alloc_cap_flush();
2358 	if (!prealloc_cf)
2359 		return -ENOMEM;
2360 
2361 	if ((iocb->ki_flags & (IOCB_DIRECT | IOCB_APPEND)) == IOCB_DIRECT)
2362 		direct_lock = true;
2363 
2364 retry_snap:
2365 	if (direct_lock)
2366 		ceph_start_io_direct(inode);
2367 	else
2368 		ceph_start_io_write(inode);
2369 
2370 	if (iocb->ki_flags & IOCB_APPEND) {
2371 		err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
2372 		if (err < 0)
2373 			goto out;
2374 	}
2375 
2376 	err = generic_write_checks(iocb, from);
2377 	if (err <= 0)
2378 		goto out;
2379 
2380 	pos = iocb->ki_pos;
2381 	if (unlikely(pos >= limit)) {
2382 		err = -EFBIG;
2383 		goto out;
2384 	} else {
2385 		iov_iter_truncate(from, limit - pos);
2386 	}
2387 
2388 	count = iov_iter_count(from);
2389 	if (ceph_quota_is_max_bytes_exceeded(inode, pos + count)) {
2390 		err = -EDQUOT;
2391 		goto out;
2392 	}
2393 
2394 	down_read(&osdc->lock);
2395 	map_flags = osdc->osdmap->flags;
2396 	pool_flags = ceph_pg_pool_flags(osdc->osdmap, ci->i_layout.pool_id);
2397 	up_read(&osdc->lock);
2398 	if ((map_flags & CEPH_OSDMAP_FULL) ||
2399 	    (pool_flags & CEPH_POOL_FLAG_FULL)) {
2400 		err = -ENOSPC;
2401 		goto out;
2402 	}
2403 
2404 	err = file_remove_privs(file);
2405 	if (err)
2406 		goto out;
2407 
2408 	doutc(cl, "%p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
2409 	      inode, ceph_vinop(inode), pos, count,
2410 	      i_size_read(inode));
2411 	if (!(fi->flags & CEPH_F_SYNC) && !direct_lock)
2412 		want |= CEPH_CAP_FILE_BUFFER;
2413 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
2414 		want |= CEPH_CAP_FILE_LAZYIO;
2415 	got = 0;
2416 	err = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, pos + count, &got);
2417 	if (err < 0)
2418 		goto out;
2419 
2420 	err = file_update_time(file);
2421 	if (err)
2422 		goto out_caps;
2423 
2424 	inode_inc_iversion_raw(inode);
2425 
2426 	doutc(cl, "%p %llx.%llx %llu~%zd got cap refs on %s\n",
2427 	      inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
2428 
2429 	if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
2430 	    (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
2431 	    (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
2432 		struct ceph_snap_context *snapc;
2433 		struct iov_iter data;
2434 
2435 		spin_lock(&ci->i_ceph_lock);
2436 		if (__ceph_have_pending_cap_snap(ci)) {
2437 			struct ceph_cap_snap *capsnap =
2438 					list_last_entry(&ci->i_cap_snaps,
2439 							struct ceph_cap_snap,
2440 							ci_item);
2441 			snapc = ceph_get_snap_context(capsnap->context);
2442 		} else {
2443 			BUG_ON(!ci->i_head_snapc);
2444 			snapc = ceph_get_snap_context(ci->i_head_snapc);
2445 		}
2446 		spin_unlock(&ci->i_ceph_lock);
2447 
2448 		/* we might need to revert back to that point */
2449 		data = *from;
2450 		if ((iocb->ki_flags & IOCB_DIRECT) && !IS_ENCRYPTED(inode))
2451 			written = ceph_direct_read_write(iocb, &data, snapc,
2452 							 &prealloc_cf);
2453 		else
2454 			written = ceph_sync_write(iocb, &data, pos, snapc);
2455 		if (direct_lock)
2456 			ceph_end_io_direct(inode);
2457 		else
2458 			ceph_end_io_write(inode);
2459 		if (written > 0)
2460 			iov_iter_advance(from, written);
2461 		ceph_put_snap_context(snapc);
2462 	} else {
2463 		/*
2464 		 * No need to acquire the i_truncate_mutex. Because
2465 		 * the MDS revokes Fwb caps before sending truncate
2466 		 * message to us. We can't get Fwb cap while there
2467 		 * are pending vmtruncate. So write and vmtruncate
2468 		 * can not run at the same time
2469 		 */
2470 		written = generic_perform_write(iocb, from);
2471 		ceph_end_io_write(inode);
2472 	}
2473 
2474 	if (written >= 0) {
2475 		int dirty;
2476 
2477 		spin_lock(&ci->i_ceph_lock);
2478 		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
2479 					       &prealloc_cf);
2480 		spin_unlock(&ci->i_ceph_lock);
2481 		if (dirty)
2482 			__mark_inode_dirty(inode, dirty);
2483 		if (ceph_quota_is_max_bytes_approaching(inode, iocb->ki_pos))
2484 			ceph_check_caps(ci, CHECK_CAPS_FLUSH);
2485 	}
2486 
2487 	doutc(cl, "%p %llx.%llx %llu~%u  dropping cap refs on %s\n",
2488 	      inode, ceph_vinop(inode), pos, (unsigned)count,
2489 	      ceph_cap_string(got));
2490 	ceph_put_cap_refs(ci, got);
2491 
2492 	if (written == -EOLDSNAPC) {
2493 		doutc(cl, "%p %llx.%llx %llu~%u" "got EOLDSNAPC, retrying\n",
2494 		      inode, ceph_vinop(inode), pos, (unsigned)count);
2495 		goto retry_snap;
2496 	}
2497 
2498 	if (written >= 0) {
2499 		if ((map_flags & CEPH_OSDMAP_NEARFULL) ||
2500 		    (pool_flags & CEPH_POOL_FLAG_NEARFULL))
2501 			iocb->ki_flags |= IOCB_DSYNC;
2502 		written = generic_write_sync(iocb, written);
2503 	}
2504 
2505 	goto out_unlocked;
2506 out_caps:
2507 	ceph_put_cap_refs(ci, got);
2508 out:
2509 	if (direct_lock)
2510 		ceph_end_io_direct(inode);
2511 	else
2512 		ceph_end_io_write(inode);
2513 out_unlocked:
2514 	ceph_free_cap_flush(prealloc_cf);
2515 	return written ? written : err;
2516 }
2517 
2518 /*
2519  * llseek.  be sure to verify file size on SEEK_END.
2520  */
2521 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
2522 {
2523 	if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
2524 		struct inode *inode = file_inode(file);
2525 		int ret;
2526 
2527 		ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
2528 		if (ret < 0)
2529 			return ret;
2530 	}
2531 	return generic_file_llseek(file, offset, whence);
2532 }
2533 
2534 static inline void ceph_zero_partial_page(
2535 	struct inode *inode, loff_t offset, unsigned size)
2536 {
2537 	struct page *page;
2538 	pgoff_t index = offset >> PAGE_SHIFT;
2539 
2540 	page = find_lock_page(inode->i_mapping, index);
2541 	if (page) {
2542 		wait_on_page_writeback(page);
2543 		zero_user(page, offset & (PAGE_SIZE - 1), size);
2544 		unlock_page(page);
2545 		put_page(page);
2546 	}
2547 }
2548 
2549 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
2550 				      loff_t length)
2551 {
2552 	loff_t nearly = round_up(offset, PAGE_SIZE);
2553 	if (offset < nearly) {
2554 		loff_t size = nearly - offset;
2555 		if (length < size)
2556 			size = length;
2557 		ceph_zero_partial_page(inode, offset, size);
2558 		offset += size;
2559 		length -= size;
2560 	}
2561 	if (length >= PAGE_SIZE) {
2562 		loff_t size = round_down(length, PAGE_SIZE);
2563 		truncate_pagecache_range(inode, offset, offset + size - 1);
2564 		offset += size;
2565 		length -= size;
2566 	}
2567 	if (length)
2568 		ceph_zero_partial_page(inode, offset, length);
2569 }
2570 
2571 static int ceph_zero_partial_object(struct inode *inode,
2572 				    loff_t offset, loff_t *length)
2573 {
2574 	struct ceph_inode_info *ci = ceph_inode(inode);
2575 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
2576 	struct ceph_osd_request *req;
2577 	int ret = 0;
2578 	loff_t zero = 0;
2579 	int op;
2580 
2581 	if (ceph_inode_is_shutdown(inode))
2582 		return -EIO;
2583 
2584 	if (!length) {
2585 		op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
2586 		length = &zero;
2587 	} else {
2588 		op = CEPH_OSD_OP_ZERO;
2589 	}
2590 
2591 	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
2592 					ceph_vino(inode),
2593 					offset, length,
2594 					0, 1, op,
2595 					CEPH_OSD_FLAG_WRITE,
2596 					NULL, 0, 0, false);
2597 	if (IS_ERR(req)) {
2598 		ret = PTR_ERR(req);
2599 		goto out;
2600 	}
2601 
2602 	req->r_mtime = inode_get_mtime(inode);
2603 	ceph_osdc_start_request(&fsc->client->osdc, req);
2604 	ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
2605 	if (ret == -ENOENT)
2606 		ret = 0;
2607 	ceph_osdc_put_request(req);
2608 
2609 out:
2610 	return ret;
2611 }
2612 
2613 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
2614 {
2615 	int ret = 0;
2616 	struct ceph_inode_info *ci = ceph_inode(inode);
2617 	s32 stripe_unit = ci->i_layout.stripe_unit;
2618 	s32 stripe_count = ci->i_layout.stripe_count;
2619 	s32 object_size = ci->i_layout.object_size;
2620 	u64 object_set_size = object_size * stripe_count;
2621 	u64 nearly, t;
2622 
2623 	/* round offset up to next period boundary */
2624 	nearly = offset + object_set_size - 1;
2625 	t = nearly;
2626 	nearly -= do_div(t, object_set_size);
2627 
2628 	while (length && offset < nearly) {
2629 		loff_t size = length;
2630 		ret = ceph_zero_partial_object(inode, offset, &size);
2631 		if (ret < 0)
2632 			return ret;
2633 		offset += size;
2634 		length -= size;
2635 	}
2636 	while (length >= object_set_size) {
2637 		int i;
2638 		loff_t pos = offset;
2639 		for (i = 0; i < stripe_count; ++i) {
2640 			ret = ceph_zero_partial_object(inode, pos, NULL);
2641 			if (ret < 0)
2642 				return ret;
2643 			pos += stripe_unit;
2644 		}
2645 		offset += object_set_size;
2646 		length -= object_set_size;
2647 	}
2648 	while (length) {
2649 		loff_t size = length;
2650 		ret = ceph_zero_partial_object(inode, offset, &size);
2651 		if (ret < 0)
2652 			return ret;
2653 		offset += size;
2654 		length -= size;
2655 	}
2656 	return ret;
2657 }
2658 
2659 static long ceph_fallocate(struct file *file, int mode,
2660 				loff_t offset, loff_t length)
2661 {
2662 	struct ceph_file_info *fi = file->private_data;
2663 	struct inode *inode = file_inode(file);
2664 	struct ceph_inode_info *ci = ceph_inode(inode);
2665 	struct ceph_cap_flush *prealloc_cf;
2666 	struct ceph_client *cl = ceph_inode_to_client(inode);
2667 	int want, got = 0;
2668 	int dirty;
2669 	int ret = 0;
2670 	loff_t endoff = 0;
2671 	loff_t size;
2672 
2673 	doutc(cl, "%p %llx.%llx mode %x, offset %llu length %llu\n",
2674 	      inode, ceph_vinop(inode), mode, offset, length);
2675 
2676 	if (mode != (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2677 		return -EOPNOTSUPP;
2678 
2679 	if (!S_ISREG(inode->i_mode))
2680 		return -EOPNOTSUPP;
2681 
2682 	if (IS_ENCRYPTED(inode))
2683 		return -EOPNOTSUPP;
2684 
2685 	prealloc_cf = ceph_alloc_cap_flush();
2686 	if (!prealloc_cf)
2687 		return -ENOMEM;
2688 
2689 	inode_lock(inode);
2690 
2691 	if (ceph_snap(inode) != CEPH_NOSNAP) {
2692 		ret = -EROFS;
2693 		goto unlock;
2694 	}
2695 
2696 	size = i_size_read(inode);
2697 
2698 	/* Are we punching a hole beyond EOF? */
2699 	if (offset >= size)
2700 		goto unlock;
2701 	if ((offset + length) > size)
2702 		length = size - offset;
2703 
2704 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
2705 		want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
2706 	else
2707 		want = CEPH_CAP_FILE_BUFFER;
2708 
2709 	ret = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, endoff, &got);
2710 	if (ret < 0)
2711 		goto unlock;
2712 
2713 	ret = file_modified(file);
2714 	if (ret)
2715 		goto put_caps;
2716 
2717 	filemap_invalidate_lock(inode->i_mapping);
2718 	ceph_fscache_invalidate(inode, false);
2719 	ceph_zero_pagecache_range(inode, offset, length);
2720 	ret = ceph_zero_objects(inode, offset, length);
2721 
2722 	if (!ret) {
2723 		spin_lock(&ci->i_ceph_lock);
2724 		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
2725 					       &prealloc_cf);
2726 		spin_unlock(&ci->i_ceph_lock);
2727 		if (dirty)
2728 			__mark_inode_dirty(inode, dirty);
2729 	}
2730 	filemap_invalidate_unlock(inode->i_mapping);
2731 
2732 put_caps:
2733 	ceph_put_cap_refs(ci, got);
2734 unlock:
2735 	inode_unlock(inode);
2736 	ceph_free_cap_flush(prealloc_cf);
2737 	return ret;
2738 }
2739 
2740 /*
2741  * This function tries to get FILE_WR capabilities for dst_ci and FILE_RD for
2742  * src_ci.  Two attempts are made to obtain both caps, and an error is return if
2743  * this fails; zero is returned on success.
2744  */
2745 static int get_rd_wr_caps(struct file *src_filp, int *src_got,
2746 			  struct file *dst_filp,
2747 			  loff_t dst_endoff, int *dst_got)
2748 {
2749 	int ret = 0;
2750 	bool retrying = false;
2751 
2752 retry_caps:
2753 	ret = ceph_get_caps(dst_filp, CEPH_CAP_FILE_WR, CEPH_CAP_FILE_BUFFER,
2754 			    dst_endoff, dst_got);
2755 	if (ret < 0)
2756 		return ret;
2757 
2758 	/*
2759 	 * Since we're already holding the FILE_WR capability for the dst file,
2760 	 * we would risk a deadlock by using ceph_get_caps.  Thus, we'll do some
2761 	 * retry dance instead to try to get both capabilities.
2762 	 */
2763 	ret = ceph_try_get_caps(file_inode(src_filp),
2764 				CEPH_CAP_FILE_RD, CEPH_CAP_FILE_SHARED,
2765 				false, src_got);
2766 	if (ret <= 0) {
2767 		/* Start by dropping dst_ci caps and getting src_ci caps */
2768 		ceph_put_cap_refs(ceph_inode(file_inode(dst_filp)), *dst_got);
2769 		if (retrying) {
2770 			if (!ret)
2771 				/* ceph_try_get_caps masks EAGAIN */
2772 				ret = -EAGAIN;
2773 			return ret;
2774 		}
2775 		ret = ceph_get_caps(src_filp, CEPH_CAP_FILE_RD,
2776 				    CEPH_CAP_FILE_SHARED, -1, src_got);
2777 		if (ret < 0)
2778 			return ret;
2779 		/*... drop src_ci caps too, and retry */
2780 		ceph_put_cap_refs(ceph_inode(file_inode(src_filp)), *src_got);
2781 		retrying = true;
2782 		goto retry_caps;
2783 	}
2784 	return ret;
2785 }
2786 
2787 static void put_rd_wr_caps(struct ceph_inode_info *src_ci, int src_got,
2788 			   struct ceph_inode_info *dst_ci, int dst_got)
2789 {
2790 	ceph_put_cap_refs(src_ci, src_got);
2791 	ceph_put_cap_refs(dst_ci, dst_got);
2792 }
2793 
2794 /*
2795  * This function does several size-related checks, returning an error if:
2796  *  - source file is smaller than off+len
2797  *  - destination file size is not OK (inode_newsize_ok())
2798  *  - max bytes quotas is exceeded
2799  */
2800 static int is_file_size_ok(struct inode *src_inode, struct inode *dst_inode,
2801 			   loff_t src_off, loff_t dst_off, size_t len)
2802 {
2803 	struct ceph_client *cl = ceph_inode_to_client(src_inode);
2804 	loff_t size, endoff;
2805 
2806 	size = i_size_read(src_inode);
2807 	/*
2808 	 * Don't copy beyond source file EOF.  Instead of simply setting length
2809 	 * to (size - src_off), just drop to VFS default implementation, as the
2810 	 * local i_size may be stale due to other clients writing to the source
2811 	 * inode.
2812 	 */
2813 	if (src_off + len > size) {
2814 		doutc(cl, "Copy beyond EOF (%llu + %zu > %llu)\n", src_off,
2815 		      len, size);
2816 		return -EOPNOTSUPP;
2817 	}
2818 	size = i_size_read(dst_inode);
2819 
2820 	endoff = dst_off + len;
2821 	if (inode_newsize_ok(dst_inode, endoff))
2822 		return -EOPNOTSUPP;
2823 
2824 	if (ceph_quota_is_max_bytes_exceeded(dst_inode, endoff))
2825 		return -EDQUOT;
2826 
2827 	return 0;
2828 }
2829 
2830 static struct ceph_osd_request *
2831 ceph_alloc_copyfrom_request(struct ceph_osd_client *osdc,
2832 			    u64 src_snapid,
2833 			    struct ceph_object_id *src_oid,
2834 			    struct ceph_object_locator *src_oloc,
2835 			    struct ceph_object_id *dst_oid,
2836 			    struct ceph_object_locator *dst_oloc,
2837 			    u32 truncate_seq, u64 truncate_size)
2838 {
2839 	struct ceph_osd_request *req;
2840 	int ret;
2841 	u32 src_fadvise_flags =
2842 		CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2843 		CEPH_OSD_OP_FLAG_FADVISE_NOCACHE;
2844 	u32 dst_fadvise_flags =
2845 		CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2846 		CEPH_OSD_OP_FLAG_FADVISE_DONTNEED;
2847 
2848 	req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_KERNEL);
2849 	if (!req)
2850 		return ERR_PTR(-ENOMEM);
2851 
2852 	req->r_flags = CEPH_OSD_FLAG_WRITE;
2853 
2854 	ceph_oloc_copy(&req->r_t.base_oloc, dst_oloc);
2855 	ceph_oid_copy(&req->r_t.base_oid, dst_oid);
2856 
2857 	ret = osd_req_op_copy_from_init(req, src_snapid, 0,
2858 					src_oid, src_oloc,
2859 					src_fadvise_flags,
2860 					dst_fadvise_flags,
2861 					truncate_seq,
2862 					truncate_size,
2863 					CEPH_OSD_COPY_FROM_FLAG_TRUNCATE_SEQ);
2864 	if (ret)
2865 		goto out;
2866 
2867 	ret = ceph_osdc_alloc_messages(req, GFP_KERNEL);
2868 	if (ret)
2869 		goto out;
2870 
2871 	return req;
2872 
2873 out:
2874 	ceph_osdc_put_request(req);
2875 	return ERR_PTR(ret);
2876 }
2877 
2878 static ssize_t ceph_do_objects_copy(struct ceph_inode_info *src_ci, u64 *src_off,
2879 				    struct ceph_inode_info *dst_ci, u64 *dst_off,
2880 				    struct ceph_fs_client *fsc,
2881 				    size_t len, unsigned int flags)
2882 {
2883 	struct ceph_object_locator src_oloc, dst_oloc;
2884 	struct ceph_object_id src_oid, dst_oid;
2885 	struct ceph_osd_client *osdc;
2886 	struct ceph_osd_request *req;
2887 	size_t bytes = 0;
2888 	u64 src_objnum, src_objoff, dst_objnum, dst_objoff;
2889 	u32 src_objlen, dst_objlen;
2890 	u32 object_size = src_ci->i_layout.object_size;
2891 	struct ceph_client *cl = fsc->client;
2892 	int ret;
2893 
2894 	src_oloc.pool = src_ci->i_layout.pool_id;
2895 	src_oloc.pool_ns = ceph_try_get_string(src_ci->i_layout.pool_ns);
2896 	dst_oloc.pool = dst_ci->i_layout.pool_id;
2897 	dst_oloc.pool_ns = ceph_try_get_string(dst_ci->i_layout.pool_ns);
2898 	osdc = &fsc->client->osdc;
2899 
2900 	while (len >= object_size) {
2901 		ceph_calc_file_object_mapping(&src_ci->i_layout, *src_off,
2902 					      object_size, &src_objnum,
2903 					      &src_objoff, &src_objlen);
2904 		ceph_calc_file_object_mapping(&dst_ci->i_layout, *dst_off,
2905 					      object_size, &dst_objnum,
2906 					      &dst_objoff, &dst_objlen);
2907 		ceph_oid_init(&src_oid);
2908 		ceph_oid_printf(&src_oid, "%llx.%08llx",
2909 				src_ci->i_vino.ino, src_objnum);
2910 		ceph_oid_init(&dst_oid);
2911 		ceph_oid_printf(&dst_oid, "%llx.%08llx",
2912 				dst_ci->i_vino.ino, dst_objnum);
2913 		/* Do an object remote copy */
2914 		req = ceph_alloc_copyfrom_request(osdc, src_ci->i_vino.snap,
2915 						  &src_oid, &src_oloc,
2916 						  &dst_oid, &dst_oloc,
2917 						  dst_ci->i_truncate_seq,
2918 						  dst_ci->i_truncate_size);
2919 		if (IS_ERR(req))
2920 			ret = PTR_ERR(req);
2921 		else {
2922 			ceph_osdc_start_request(osdc, req);
2923 			ret = ceph_osdc_wait_request(osdc, req);
2924 			ceph_update_copyfrom_metrics(&fsc->mdsc->metric,
2925 						     req->r_start_latency,
2926 						     req->r_end_latency,
2927 						     object_size, ret);
2928 			ceph_osdc_put_request(req);
2929 		}
2930 		if (ret) {
2931 			if (ret == -EOPNOTSUPP) {
2932 				fsc->have_copy_from2 = false;
2933 				pr_notice_client(cl,
2934 					"OSDs don't support copy-from2; disabling copy offload\n");
2935 			}
2936 			doutc(cl, "returned %d\n", ret);
2937 			if (!bytes)
2938 				bytes = ret;
2939 			goto out;
2940 		}
2941 		len -= object_size;
2942 		bytes += object_size;
2943 		*src_off += object_size;
2944 		*dst_off += object_size;
2945 	}
2946 
2947 out:
2948 	ceph_oloc_destroy(&src_oloc);
2949 	ceph_oloc_destroy(&dst_oloc);
2950 	return bytes;
2951 }
2952 
2953 static ssize_t __ceph_copy_file_range(struct file *src_file, loff_t src_off,
2954 				      struct file *dst_file, loff_t dst_off,
2955 				      size_t len, unsigned int flags)
2956 {
2957 	struct inode *src_inode = file_inode(src_file);
2958 	struct inode *dst_inode = file_inode(dst_file);
2959 	struct ceph_inode_info *src_ci = ceph_inode(src_inode);
2960 	struct ceph_inode_info *dst_ci = ceph_inode(dst_inode);
2961 	struct ceph_cap_flush *prealloc_cf;
2962 	struct ceph_fs_client *src_fsc = ceph_inode_to_fs_client(src_inode);
2963 	struct ceph_client *cl = src_fsc->client;
2964 	loff_t size;
2965 	ssize_t ret = -EIO, bytes;
2966 	u64 src_objnum, dst_objnum, src_objoff, dst_objoff;
2967 	u32 src_objlen, dst_objlen;
2968 	int src_got = 0, dst_got = 0, err, dirty;
2969 
2970 	if (src_inode->i_sb != dst_inode->i_sb) {
2971 		struct ceph_fs_client *dst_fsc = ceph_inode_to_fs_client(dst_inode);
2972 
2973 		if (ceph_fsid_compare(&src_fsc->client->fsid,
2974 				      &dst_fsc->client->fsid)) {
2975 			dout("Copying files across clusters: src: %pU dst: %pU\n",
2976 			     &src_fsc->client->fsid, &dst_fsc->client->fsid);
2977 			return -EXDEV;
2978 		}
2979 	}
2980 	if (ceph_snap(dst_inode) != CEPH_NOSNAP)
2981 		return -EROFS;
2982 
2983 	/*
2984 	 * Some of the checks below will return -EOPNOTSUPP, which will force a
2985 	 * fallback to the default VFS copy_file_range implementation.  This is
2986 	 * desirable in several cases (for ex, the 'len' is smaller than the
2987 	 * size of the objects, or in cases where that would be more
2988 	 * efficient).
2989 	 */
2990 
2991 	if (ceph_test_mount_opt(src_fsc, NOCOPYFROM))
2992 		return -EOPNOTSUPP;
2993 
2994 	if (!src_fsc->have_copy_from2)
2995 		return -EOPNOTSUPP;
2996 
2997 	/*
2998 	 * Striped file layouts require that we copy partial objects, but the
2999 	 * OSD copy-from operation only supports full-object copies.  Limit
3000 	 * this to non-striped file layouts for now.
3001 	 */
3002 	if ((src_ci->i_layout.stripe_unit != dst_ci->i_layout.stripe_unit) ||
3003 	    (src_ci->i_layout.stripe_count != 1) ||
3004 	    (dst_ci->i_layout.stripe_count != 1) ||
3005 	    (src_ci->i_layout.object_size != dst_ci->i_layout.object_size)) {
3006 		doutc(cl, "Invalid src/dst files layout\n");
3007 		return -EOPNOTSUPP;
3008 	}
3009 
3010 	/* Every encrypted inode gets its own key, so we can't offload them */
3011 	if (IS_ENCRYPTED(src_inode) || IS_ENCRYPTED(dst_inode))
3012 		return -EOPNOTSUPP;
3013 
3014 	if (len < src_ci->i_layout.object_size)
3015 		return -EOPNOTSUPP; /* no remote copy will be done */
3016 
3017 	prealloc_cf = ceph_alloc_cap_flush();
3018 	if (!prealloc_cf)
3019 		return -ENOMEM;
3020 
3021 	/* Start by sync'ing the source and destination files */
3022 	ret = file_write_and_wait_range(src_file, src_off, (src_off + len));
3023 	if (ret < 0) {
3024 		doutc(cl, "failed to write src file (%zd)\n", ret);
3025 		goto out;
3026 	}
3027 	ret = file_write_and_wait_range(dst_file, dst_off, (dst_off + len));
3028 	if (ret < 0) {
3029 		doutc(cl, "failed to write dst file (%zd)\n", ret);
3030 		goto out;
3031 	}
3032 
3033 	/*
3034 	 * We need FILE_WR caps for dst_ci and FILE_RD for src_ci as other
3035 	 * clients may have dirty data in their caches.  And OSDs know nothing
3036 	 * about caps, so they can't safely do the remote object copies.
3037 	 */
3038 	err = get_rd_wr_caps(src_file, &src_got,
3039 			     dst_file, (dst_off + len), &dst_got);
3040 	if (err < 0) {
3041 		doutc(cl, "get_rd_wr_caps returned %d\n", err);
3042 		ret = -EOPNOTSUPP;
3043 		goto out;
3044 	}
3045 
3046 	ret = is_file_size_ok(src_inode, dst_inode, src_off, dst_off, len);
3047 	if (ret < 0)
3048 		goto out_caps;
3049 
3050 	/* Drop dst file cached pages */
3051 	ceph_fscache_invalidate(dst_inode, false);
3052 	ret = invalidate_inode_pages2_range(dst_inode->i_mapping,
3053 					    dst_off >> PAGE_SHIFT,
3054 					    (dst_off + len) >> PAGE_SHIFT);
3055 	if (ret < 0) {
3056 		doutc(cl, "Failed to invalidate inode pages (%zd)\n",
3057 			    ret);
3058 		ret = 0; /* XXX */
3059 	}
3060 	ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
3061 				      src_ci->i_layout.object_size,
3062 				      &src_objnum, &src_objoff, &src_objlen);
3063 	ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
3064 				      dst_ci->i_layout.object_size,
3065 				      &dst_objnum, &dst_objoff, &dst_objlen);
3066 	/* object-level offsets need to the same */
3067 	if (src_objoff != dst_objoff) {
3068 		ret = -EOPNOTSUPP;
3069 		goto out_caps;
3070 	}
3071 
3072 	/*
3073 	 * Do a manual copy if the object offset isn't object aligned.
3074 	 * 'src_objlen' contains the bytes left until the end of the object,
3075 	 * starting at the src_off
3076 	 */
3077 	if (src_objoff) {
3078 		doutc(cl, "Initial partial copy of %u bytes\n", src_objlen);
3079 
3080 		/*
3081 		 * we need to temporarily drop all caps as we'll be calling
3082 		 * {read,write}_iter, which will get caps again.
3083 		 */
3084 		put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
3085 		ret = splice_file_range(src_file, &src_off, dst_file, &dst_off,
3086 					src_objlen);
3087 		/* Abort on short copies or on error */
3088 		if (ret < (long)src_objlen) {
3089 			doutc(cl, "Failed partial copy (%zd)\n", ret);
3090 			goto out;
3091 		}
3092 		len -= ret;
3093 		err = get_rd_wr_caps(src_file, &src_got,
3094 				     dst_file, (dst_off + len), &dst_got);
3095 		if (err < 0)
3096 			goto out;
3097 		err = is_file_size_ok(src_inode, dst_inode,
3098 				      src_off, dst_off, len);
3099 		if (err < 0)
3100 			goto out_caps;
3101 	}
3102 
3103 	size = i_size_read(dst_inode);
3104 	bytes = ceph_do_objects_copy(src_ci, &src_off, dst_ci, &dst_off,
3105 				     src_fsc, len, flags);
3106 	if (bytes <= 0) {
3107 		if (!ret)
3108 			ret = bytes;
3109 		goto out_caps;
3110 	}
3111 	doutc(cl, "Copied %zu bytes out of %zu\n", bytes, len);
3112 	len -= bytes;
3113 	ret += bytes;
3114 
3115 	file_update_time(dst_file);
3116 	inode_inc_iversion_raw(dst_inode);
3117 
3118 	if (dst_off > size) {
3119 		/* Let the MDS know about dst file size change */
3120 		if (ceph_inode_set_size(dst_inode, dst_off) ||
3121 		    ceph_quota_is_max_bytes_approaching(dst_inode, dst_off))
3122 			ceph_check_caps(dst_ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_FLUSH);
3123 	}
3124 	/* Mark Fw dirty */
3125 	spin_lock(&dst_ci->i_ceph_lock);
3126 	dirty = __ceph_mark_dirty_caps(dst_ci, CEPH_CAP_FILE_WR, &prealloc_cf);
3127 	spin_unlock(&dst_ci->i_ceph_lock);
3128 	if (dirty)
3129 		__mark_inode_dirty(dst_inode, dirty);
3130 
3131 out_caps:
3132 	put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
3133 
3134 	/*
3135 	 * Do the final manual copy if we still have some bytes left, unless
3136 	 * there were errors in remote object copies (len >= object_size).
3137 	 */
3138 	if (len && (len < src_ci->i_layout.object_size)) {
3139 		doutc(cl, "Final partial copy of %zu bytes\n", len);
3140 		bytes = splice_file_range(src_file, &src_off, dst_file,
3141 					  &dst_off, len);
3142 		if (bytes > 0)
3143 			ret += bytes;
3144 		else
3145 			doutc(cl, "Failed partial copy (%zd)\n", bytes);
3146 	}
3147 
3148 out:
3149 	ceph_free_cap_flush(prealloc_cf);
3150 
3151 	return ret;
3152 }
3153 
3154 static ssize_t ceph_copy_file_range(struct file *src_file, loff_t src_off,
3155 				    struct file *dst_file, loff_t dst_off,
3156 				    size_t len, unsigned int flags)
3157 {
3158 	ssize_t ret;
3159 
3160 	ret = __ceph_copy_file_range(src_file, src_off, dst_file, dst_off,
3161 				     len, flags);
3162 
3163 	if (ret == -EOPNOTSUPP || ret == -EXDEV)
3164 		ret = splice_copy_file_range(src_file, src_off, dst_file,
3165 					     dst_off, len);
3166 	return ret;
3167 }
3168 
3169 const struct file_operations ceph_file_fops = {
3170 	.open = ceph_open,
3171 	.release = ceph_release,
3172 	.llseek = ceph_llseek,
3173 	.read_iter = ceph_read_iter,
3174 	.write_iter = ceph_write_iter,
3175 	.mmap = ceph_mmap,
3176 	.fsync = ceph_fsync,
3177 	.lock = ceph_lock,
3178 	.setlease = simple_nosetlease,
3179 	.flock = ceph_flock,
3180 	.splice_read = ceph_splice_read,
3181 	.splice_write = iter_file_splice_write,
3182 	.unlocked_ioctl = ceph_ioctl,
3183 	.compat_ioctl = compat_ptr_ioctl,
3184 	.fallocate	= ceph_fallocate,
3185 	.copy_file_range = ceph_copy_file_range,
3186 };
3187