1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
3 #include <linux/ceph/striper.h>
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>
18 #include "mds_client.h"
23 static __le32 ceph_flags_sys2wire(struct ceph_mds_client *mdsc, u32 flags)
25 struct ceph_client *cl = mdsc->fsc->client;
28 switch (flags & O_ACCMODE) {
30 wire_flags |= CEPH_O_RDONLY;
33 wire_flags |= CEPH_O_WRONLY;
36 wire_flags |= CEPH_O_RDWR;
42 #define ceph_sys2wire(a) if (flags & a) { wire_flags |= CEPH_##a; flags &= ~a; }
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);
53 doutc(cl, "unused open flags: %x\n", flags);
55 return cpu_to_le32(wire_flags);
59 * Ceph file operations
61 * Implement basic open/close functionality, and implement
64 * We implement three modes of file I/O:
65 * - buffered uses the generic_file_aio_{read,write} helpers
67 * - synchronous is used when there is multi-client read/write
68 * sharing, avoids the page cache, and synchronously waits for an
71 * - direct io takes the variant of the sync path that references
72 * user pages directly.
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.
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.
83 #define ITER_GET_BVECS_PAGES 64
85 static ssize_t __iter_get_bvecs(struct iov_iter *iter, size_t maxsize,
86 struct bio_vec *bvecs)
91 if (maxsize > iov_iter_count(iter))
92 maxsize = iov_iter_count(iter);
94 while (size < maxsize) {
95 struct page *pages[ITER_GET_BVECS_PAGES];
100 bytes = iov_iter_get_pages2(iter, pages, maxsize - size,
101 ITER_GET_BVECS_PAGES, &start);
103 return size ?: bytes;
107 for ( ; bytes; idx++, bvec_idx++) {
108 int len = min_t(int, bytes, PAGE_SIZE - start);
110 bvec_set_page(&bvecs[bvec_idx], pages[idx], len, start);
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
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.
127 static ssize_t iter_get_bvecs_alloc(struct iov_iter *iter, size_t maxsize,
128 struct bio_vec **bvecs, int *num_bvecs)
131 size_t orig_count = iov_iter_count(iter);
135 iov_iter_truncate(iter, maxsize);
136 npages = iov_iter_npages(iter, INT_MAX);
137 iov_iter_reexpand(iter, orig_count);
140 * __iter_get_bvecs() may populate only part of the array -- zero it
143 bv = kvmalloc_array(npages, sizeof(*bv), GFP_KERNEL | __GFP_ZERO);
147 bytes = __iter_get_bvecs(iter, maxsize, bv);
150 * No pages were pinned -- just free the array.
161 static void put_bvecs(struct bio_vec *bvecs, int num_bvecs, bool should_dirty)
165 for (i = 0; i < num_bvecs; i++) {
166 if (bvecs[i].bv_page) {
168 set_page_dirty_lock(bvecs[i].bv_page);
169 put_page(bvecs[i].bv_page);
176 * Prepare an open request. Preallocate ceph_cap to avoid an
177 * inopportune ENOMEM later.
179 static struct ceph_mds_request *
180 prepare_open_request(struct super_block *sb, int flags, int create_mode)
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;
187 if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
188 want_auth = USE_AUTH_MDS;
190 req = ceph_mdsc_create_request(mdsc, op, want_auth);
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);
200 static int ceph_init_file_info(struct inode *inode, struct file *file,
201 int fmode, bool isdir)
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;
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);
215 struct ceph_dir_file_info *dfi =
216 kmem_cache_zalloc(ceph_dir_file_cachep, GFP_KERNEL);
220 file->private_data = dfi;
221 fi = &dfi->file_info;
222 dfi->next_offset = 2;
223 dfi->readdir_cache_idx = -1;
225 fi = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
229 if (opt->flags & CEPH_MOUNT_OPT_NOPAGECACHE)
230 fi->flags |= CEPH_F_SYNC;
232 file->private_data = fi;
235 ceph_get_fmode(ci, fmode, 1);
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);
242 if ((file->f_mode & FMODE_WRITE) && ceph_has_inline_data(ci)) {
243 ret = ceph_uninline_data(file);
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);
260 * initialize private struct file data.
261 * if we fail, clean up by dropping fmode reference on the ceph_inode
263 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
265 struct ceph_client *cl = ceph_inode_to_client(inode);
268 switch (inode->i_mode & S_IFMT) {
270 ceph_fscache_use_cookie(inode, file->f_mode & FMODE_WRITE);
273 ret = ceph_init_file_info(inode, file, fmode,
274 S_ISDIR(inode->i_mode));
278 doutc(cl, "%p %llx.%llx %p 0%o (symlink)\n", inode,
279 ceph_vinop(inode), file, inode->i_mode);
283 doutc(cl, "%p %llx.%llx %p 0%o (special)\n", inode,
284 ceph_vinop(inode), file, inode->i_mode);
286 * we need to drop the open ref now, since we don't
287 * have .release set to ceph_release.
289 BUG_ON(inode->i_fop->release == ceph_release);
291 /* call the proper open fop */
292 ret = inode->i_fop->open(inode, file);
298 * try renew caps after session gets killed.
300 int ceph_renew_caps(struct inode *inode, int fmode)
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;
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);
321 spin_unlock(&ci->i_ceph_lock);
324 if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
326 else if (wanted & CEPH_CAP_FILE_RD)
328 else if (wanted & CEPH_CAP_FILE_WR)
331 if (wanted & CEPH_CAP_FILE_LAZYIO)
335 req = prepare_open_request(inode->i_sb, flags, 0);
341 req->r_inode = inode;
345 err = ceph_mdsc_do_request(mdsc, NULL, req);
346 ceph_mdsc_put_request(req);
348 doutc(cl, "%p %llx.%llx open result=%d\n", inode, ceph_vinop(inode),
350 return err < 0 ? err : 0;
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.
359 int ceph_open(struct inode *inode, struct file *file)
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;
368 int flags, fmode, wanted;
371 doutc(cl, "file %p is already opened\n", file);
375 /* filter out O_CREAT|O_EXCL; vfs did that already. yuck. */
376 flags = file->f_flags & ~(O_CREAT|O_EXCL);
377 if (S_ISDIR(inode->i_mode)) {
378 flags = O_DIRECTORY; /* mds likes to know */
379 } else if (S_ISREG(inode->i_mode)) {
380 err = fscrypt_file_open(inode, file);
385 doutc(cl, "%p %llx.%llx file %p flags %d (%d)\n", inode,
386 ceph_vinop(inode), file, flags, file->f_flags);
387 fmode = ceph_flags_to_mode(flags);
388 wanted = ceph_caps_for_mode(fmode);
390 /* snapped files are read-only */
391 if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
394 /* trivially open snapdir */
395 if (ceph_snap(inode) == CEPH_SNAPDIR) {
396 return ceph_init_file(inode, file, fmode);
400 * No need to block if we have caps on the auth MDS (for
401 * write) or any MDS (for read). Update wanted set
404 spin_lock(&ci->i_ceph_lock);
405 if (__ceph_is_any_real_caps(ci) &&
406 (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
407 int mds_wanted = __ceph_caps_mds_wanted(ci, true);
408 int issued = __ceph_caps_issued(ci, NULL);
410 doutc(cl, "open %p fmode %d want %s issued %s using existing\n",
411 inode, fmode, ceph_cap_string(wanted),
412 ceph_cap_string(issued));
413 __ceph_touch_fmode(ci, mdsc, fmode);
414 spin_unlock(&ci->i_ceph_lock);
417 if ((issued & wanted) != wanted &&
418 (mds_wanted & wanted) != wanted &&
419 ceph_snap(inode) != CEPH_SNAPDIR)
420 ceph_check_caps(ci, 0);
422 return ceph_init_file(inode, file, fmode);
423 } else if (ceph_snap(inode) != CEPH_NOSNAP &&
424 (ci->i_snap_caps & wanted) == wanted) {
425 __ceph_touch_fmode(ci, mdsc, fmode);
426 spin_unlock(&ci->i_ceph_lock);
427 return ceph_init_file(inode, file, fmode);
430 spin_unlock(&ci->i_ceph_lock);
432 doutc(cl, "open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
433 req = prepare_open_request(inode->i_sb, flags, 0);
438 req->r_inode = inode;
442 err = ceph_mdsc_do_request(mdsc, NULL, req);
444 err = ceph_init_file(inode, file, req->r_fmode);
445 ceph_mdsc_put_request(req);
446 doutc(cl, "open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
451 /* Clone the layout from a synchronous create, if the dir now has Dc caps */
453 cache_file_layout(struct inode *dst, struct inode *src)
455 struct ceph_inode_info *cdst = ceph_inode(dst);
456 struct ceph_inode_info *csrc = ceph_inode(src);
458 spin_lock(&cdst->i_ceph_lock);
459 if ((__ceph_caps_issued(cdst, NULL) & CEPH_CAP_DIR_CREATE) &&
460 !ceph_file_layout_is_valid(&cdst->i_cached_layout)) {
461 memcpy(&cdst->i_cached_layout, &csrc->i_layout,
462 sizeof(cdst->i_cached_layout));
463 rcu_assign_pointer(cdst->i_cached_layout.pool_ns,
464 ceph_try_get_string(csrc->i_layout.pool_ns));
466 spin_unlock(&cdst->i_ceph_lock);
470 * Try to set up an async create. We need caps, a file layout, and inode number,
471 * and either a lease on the dentry or complete dir info. If any of those
472 * criteria are not satisfied, then return false and the caller can go
475 static int try_prep_async_create(struct inode *dir, struct dentry *dentry,
476 struct ceph_file_layout *lo, u64 *pino)
478 struct ceph_inode_info *ci = ceph_inode(dir);
479 struct ceph_dentry_info *di = ceph_dentry(dentry);
480 int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_CREATE;
483 spin_lock(&ci->i_ceph_lock);
484 /* No auth cap means no chance for Dc caps */
488 /* Any delegated inos? */
489 if (xa_empty(&ci->i_auth_cap->session->s_delegated_inos))
492 if (!ceph_file_layout_is_valid(&ci->i_cached_layout))
495 if ((__ceph_caps_issued(ci, NULL) & want) != want)
498 if (d_in_lookup(dentry)) {
499 if (!__ceph_dir_is_complete(ci))
501 spin_lock(&dentry->d_lock);
502 di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
503 spin_unlock(&dentry->d_lock);
504 } else if (atomic_read(&ci->i_shared_gen) !=
505 READ_ONCE(di->lease_shared_gen)) {
509 ino = ceph_get_deleg_ino(ci->i_auth_cap->session);
514 ceph_take_cap_refs(ci, want, false);
515 memcpy(lo, &ci->i_cached_layout, sizeof(*lo));
516 rcu_assign_pointer(lo->pool_ns,
517 ceph_try_get_string(ci->i_cached_layout.pool_ns));
520 spin_unlock(&ci->i_ceph_lock);
524 static void restore_deleg_ino(struct inode *dir, u64 ino)
526 struct ceph_client *cl = ceph_inode_to_client(dir);
527 struct ceph_inode_info *ci = ceph_inode(dir);
528 struct ceph_mds_session *s = NULL;
530 spin_lock(&ci->i_ceph_lock);
532 s = ceph_get_mds_session(ci->i_auth_cap->session);
533 spin_unlock(&ci->i_ceph_lock);
535 int err = ceph_restore_deleg_ino(s, ino);
538 "unable to restore delegated ino 0x%llx to session: %d\n",
540 ceph_put_mds_session(s);
544 static void wake_async_create_waiters(struct inode *inode,
545 struct ceph_mds_session *session)
547 struct ceph_inode_info *ci = ceph_inode(inode);
548 bool check_cap = false;
550 spin_lock(&ci->i_ceph_lock);
551 if (ci->i_ceph_flags & CEPH_I_ASYNC_CREATE) {
552 ci->i_ceph_flags &= ~CEPH_I_ASYNC_CREATE;
553 wake_up_bit(&ci->i_ceph_flags, CEPH_ASYNC_CREATE_BIT);
555 if (ci->i_ceph_flags & CEPH_I_ASYNC_CHECK_CAPS) {
556 ci->i_ceph_flags &= ~CEPH_I_ASYNC_CHECK_CAPS;
560 ceph_kick_flushing_inode_caps(session, ci);
561 spin_unlock(&ci->i_ceph_lock);
564 ceph_check_caps(ci, CHECK_CAPS_FLUSH);
567 static void ceph_async_create_cb(struct ceph_mds_client *mdsc,
568 struct ceph_mds_request *req)
570 struct ceph_client *cl = mdsc->fsc->client;
571 struct dentry *dentry = req->r_dentry;
572 struct inode *dinode = d_inode(dentry);
573 struct inode *tinode = req->r_target_inode;
574 int result = req->r_err ? req->r_err :
575 le32_to_cpu(req->r_reply_info.head->result);
577 WARN_ON_ONCE(dinode && tinode && dinode != tinode);
579 /* MDS changed -- caller must resubmit */
580 if (result == -EJUKEBOX)
583 mapping_set_error(req->r_parent->i_mapping, result);
588 char *path = ceph_mdsc_build_path(mdsc, req->r_dentry, &pathlen,
592 "async create failure path=(%llx)%s result=%d!\n",
593 base, IS_ERR(path) ? "<<bad>>" : path, result);
594 ceph_mdsc_free_path(path, pathlen);
596 ceph_dir_clear_complete(req->r_parent);
597 if (!d_unhashed(dentry))
601 mapping_set_error(dinode->i_mapping, result);
602 ceph_inode_shutdown(dinode);
603 wake_async_create_waiters(dinode, req->r_session);
608 u64 ino = ceph_vino(tinode).ino;
610 if (req->r_deleg_ino != ino)
612 "inode number mismatch! err=%d deleg_ino=0x%llx target=0x%llx\n",
613 req->r_err, req->r_deleg_ino, ino);
615 mapping_set_error(tinode->i_mapping, result);
616 wake_async_create_waiters(tinode, req->r_session);
617 } else if (!result) {
618 pr_warn_client(cl, "no req->r_target_inode for 0x%llx\n",
622 ceph_mdsc_release_dir_caps(req);
625 static int ceph_finish_async_create(struct inode *dir, struct inode *inode,
626 struct dentry *dentry,
627 struct file *file, umode_t mode,
628 struct ceph_mds_request *req,
629 struct ceph_acl_sec_ctx *as_ctx,
630 struct ceph_file_layout *lo)
634 struct ceph_mds_reply_inode in = { };
635 struct ceph_mds_reply_info_in iinfo = { .in = &in };
636 struct ceph_inode_info *ci = ceph_inode(dir);
637 struct ceph_dentry_info *di = ceph_dentry(dentry);
638 struct timespec64 now;
639 struct ceph_string *pool_ns;
640 struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
641 struct ceph_client *cl = mdsc->fsc->client;
642 struct ceph_vino vino = { .ino = req->r_deleg_ino,
643 .snap = CEPH_NOSNAP };
645 ktime_get_real_ts64(&now);
647 iinfo.inline_version = CEPH_INLINE_NONE;
648 iinfo.change_attr = 1;
649 ceph_encode_timespec64(&iinfo.btime, &now);
651 if (req->r_pagelist) {
652 iinfo.xattr_len = req->r_pagelist->length;
653 iinfo.xattr_data = req->r_pagelist->mapped_tail;
656 iinfo.xattr_len = ARRAY_SIZE(xattr_buf);
657 iinfo.xattr_data = xattr_buf;
658 memset(iinfo.xattr_data, 0, iinfo.xattr_len);
661 in.ino = cpu_to_le64(vino.ino);
662 in.snapid = cpu_to_le64(CEPH_NOSNAP);
663 in.version = cpu_to_le64(1); // ???
664 in.cap.caps = in.cap.wanted = cpu_to_le32(CEPH_CAP_ALL_FILE);
665 in.cap.cap_id = cpu_to_le64(1);
666 in.cap.realm = cpu_to_le64(ci->i_snap_realm->ino);
667 in.cap.flags = CEPH_CAP_FLAG_AUTH;
668 in.ctime = in.mtime = in.atime = iinfo.btime;
669 in.truncate_seq = cpu_to_le32(1);
670 in.truncate_size = cpu_to_le64(-1ULL);
671 in.xattr_version = cpu_to_le64(1);
672 in.uid = cpu_to_le32(from_kuid(&init_user_ns,
673 mapped_fsuid(req->r_mnt_idmap,
675 if (dir->i_mode & S_ISGID) {
676 in.gid = cpu_to_le32(from_kgid(&init_user_ns, dir->i_gid));
678 /* Directories always inherit the setgid bit. */
682 in.gid = cpu_to_le32(from_kgid(&init_user_ns,
683 mapped_fsgid(req->r_mnt_idmap,
686 in.mode = cpu_to_le32((u32)mode);
688 in.nlink = cpu_to_le32(1);
689 in.max_size = cpu_to_le64(lo->stripe_unit);
691 ceph_file_layout_to_legacy(lo, &in.layout);
692 /* lo is private, so pool_ns can't change */
693 pool_ns = rcu_dereference_raw(lo->pool_ns);
695 iinfo.pool_ns_len = pool_ns->len;
696 iinfo.pool_ns_data = pool_ns->str;
699 down_read(&mdsc->snap_rwsem);
700 ret = ceph_fill_inode(inode, NULL, &iinfo, NULL, req->r_session,
702 up_read(&mdsc->snap_rwsem);
704 doutc(cl, "failed to fill inode: %d\n", ret);
705 ceph_dir_clear_complete(dir);
706 if (!d_unhashed(dentry))
708 discard_new_inode(inode);
712 doutc(cl, "d_adding new inode 0x%llx to 0x%llx/%s\n",
713 vino.ino, ceph_ino(dir), dentry->d_name.name);
714 ceph_dir_clear_ordered(dir);
715 ceph_init_inode_acls(inode, as_ctx);
716 if (inode->i_state & I_NEW) {
718 * If it's not I_NEW, then someone created this before
719 * we got here. Assume the server is aware of it at
720 * that point and don't worry about setting
721 * CEPH_I_ASYNC_CREATE.
723 ceph_inode(inode)->i_ceph_flags = CEPH_I_ASYNC_CREATE;
724 unlock_new_inode(inode);
726 if (d_in_lookup(dentry) || d_really_is_negative(dentry)) {
727 if (!d_unhashed(dentry))
729 dn = d_splice_alias(inode, dentry);
730 WARN_ON_ONCE(dn && dn != dentry);
732 file->f_mode |= FMODE_CREATED;
733 ret = finish_open(file, dentry, ceph_open);
736 spin_lock(&dentry->d_lock);
737 di->flags &= ~CEPH_DENTRY_ASYNC_CREATE;
738 wake_up_bit(&di->flags, CEPH_DENTRY_ASYNC_CREATE_BIT);
739 spin_unlock(&dentry->d_lock);
745 * Do a lookup + open with a single request. If we get a non-existent
746 * file or symlink, return 1 so the VFS can retry.
748 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
749 struct file *file, unsigned flags, umode_t mode)
751 struct mnt_idmap *idmap = file_mnt_idmap(file);
752 struct ceph_fs_client *fsc = ceph_sb_to_fs_client(dir->i_sb);
753 struct ceph_client *cl = fsc->client;
754 struct ceph_mds_client *mdsc = fsc->mdsc;
755 struct ceph_mds_request *req;
756 struct inode *new_inode = NULL;
758 struct ceph_acl_sec_ctx as_ctx = {};
759 bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
763 doutc(cl, "%p %llx.%llx dentry %p '%pd' %s flags %d mode 0%o\n",
764 dir, ceph_vinop(dir), dentry, dentry,
765 d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
767 if (dentry->d_name.len > NAME_MAX)
768 return -ENAMETOOLONG;
770 err = ceph_wait_on_conflict_unlink(dentry);
774 * Do not truncate the file, since atomic_open is called before the
775 * permission check. The caller will do the truncation afterward.
780 if (flags & O_CREAT) {
781 if (ceph_quota_is_max_files_exceeded(dir))
784 new_inode = ceph_new_inode(dir, dentry, &mode, &as_ctx);
785 if (IS_ERR(new_inode)) {
786 err = PTR_ERR(new_inode);
789 /* Async create can't handle more than a page of xattrs */
790 if (as_ctx.pagelist &&
791 !list_is_singular(&as_ctx.pagelist->head))
793 } else if (!d_in_lookup(dentry)) {
794 /* If it's not being looked up, it's negative */
799 req = prepare_open_request(dir->i_sb, flags, mode);
804 req->r_dentry = dget(dentry);
806 mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
807 if (ceph_security_xattr_wanted(dir))
808 mask |= CEPH_CAP_XATTR_SHARED;
809 req->r_args.open.mask = cpu_to_le32(mask);
811 if (req->r_op == CEPH_MDS_OP_CREATE)
812 req->r_mnt_idmap = mnt_idmap_get(idmap);
814 if (IS_ENCRYPTED(dir)) {
815 set_bit(CEPH_MDS_R_FSCRYPT_FILE, &req->r_req_flags);
816 err = fscrypt_prepare_lookup_partial(dir, dentry);
821 if (flags & O_CREAT) {
822 struct ceph_file_layout lo;
824 req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL |
826 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
828 ceph_as_ctx_to_req(req, &as_ctx);
830 if (try_async && (req->r_dir_caps =
831 try_prep_async_create(dir, dentry, &lo,
832 &req->r_deleg_ino))) {
833 struct ceph_vino vino = { .ino = req->r_deleg_ino,
834 .snap = CEPH_NOSNAP };
835 struct ceph_dentry_info *di = ceph_dentry(dentry);
837 set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
838 req->r_args.open.flags |= cpu_to_le32(CEPH_O_EXCL);
839 req->r_callback = ceph_async_create_cb;
841 /* Hash inode before RPC */
842 new_inode = ceph_get_inode(dir->i_sb, vino, new_inode);
843 if (IS_ERR(new_inode)) {
844 err = PTR_ERR(new_inode);
848 WARN_ON_ONCE(!(new_inode->i_state & I_NEW));
850 spin_lock(&dentry->d_lock);
851 di->flags |= CEPH_DENTRY_ASYNC_CREATE;
852 spin_unlock(&dentry->d_lock);
854 err = ceph_mdsc_submit_request(mdsc, dir, req);
856 err = ceph_finish_async_create(dir, new_inode,
861 } else if (err == -EJUKEBOX) {
862 restore_deleg_ino(dir, req->r_deleg_ino);
863 ceph_mdsc_put_request(req);
864 discard_new_inode(new_inode);
865 ceph_release_acl_sec_ctx(&as_ctx);
866 memset(&as_ctx, 0, sizeof(as_ctx));
869 ceph_put_string(rcu_dereference_raw(lo.pool_ns));
872 ceph_put_string(rcu_dereference_raw(lo.pool_ns));
877 set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
878 req->r_new_inode = new_inode;
880 err = ceph_mdsc_do_request(mdsc, (flags & O_CREAT) ? dir : NULL, req);
881 if (err == -ENOENT) {
882 dentry = ceph_handle_snapdir(req, dentry);
883 if (IS_ERR(dentry)) {
884 err = PTR_ERR(dentry);
890 if (!err && (flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
891 err = ceph_handle_notrace_create(dir, dentry);
893 if (d_in_lookup(dentry)) {
894 dn = ceph_finish_lookup(req, dentry, err);
898 /* we were given a hashed negative dentry */
903 if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
904 /* make vfs retry on splice, ENOENT, or symlink */
905 doutc(cl, "finish_no_open on dn %p\n", dn);
906 err = finish_no_open(file, dn);
908 if (IS_ENCRYPTED(dir) &&
909 !fscrypt_has_permitted_context(dir, d_inode(dentry))) {
911 "Inconsistent encryption context (parent %llx:%llx child %llx:%llx)\n",
912 ceph_vinop(dir), ceph_vinop(d_inode(dentry)));
916 doutc(cl, "finish_open on dn %p\n", dn);
917 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
918 struct inode *newino = d_inode(dentry);
920 cache_file_layout(dir, newino);
921 ceph_init_inode_acls(newino, &as_ctx);
922 file->f_mode |= FMODE_CREATED;
924 err = finish_open(file, dentry, ceph_open);
927 ceph_mdsc_put_request(req);
930 ceph_release_acl_sec_ctx(&as_ctx);
931 doutc(cl, "result=%d\n", err);
935 int ceph_release(struct inode *inode, struct file *file)
937 struct ceph_client *cl = ceph_inode_to_client(inode);
938 struct ceph_inode_info *ci = ceph_inode(inode);
940 if (S_ISDIR(inode->i_mode)) {
941 struct ceph_dir_file_info *dfi = file->private_data;
942 doutc(cl, "%p %llx.%llx dir file %p\n", inode,
943 ceph_vinop(inode), file);
944 WARN_ON(!list_empty(&dfi->file_info.rw_contexts));
946 ceph_put_fmode(ci, dfi->file_info.fmode, 1);
948 if (dfi->last_readdir)
949 ceph_mdsc_put_request(dfi->last_readdir);
950 kfree(dfi->last_name);
951 kfree(dfi->dir_info);
952 kmem_cache_free(ceph_dir_file_cachep, dfi);
954 struct ceph_file_info *fi = file->private_data;
955 doutc(cl, "%p %llx.%llx regular file %p\n", inode,
956 ceph_vinop(inode), file);
957 WARN_ON(!list_empty(&fi->rw_contexts));
959 ceph_fscache_unuse_cookie(inode, file->f_mode & FMODE_WRITE);
960 ceph_put_fmode(ci, fi->fmode, 1);
962 kmem_cache_free(ceph_file_cachep, fi);
965 /* wake up anyone waiting for caps on this inode */
966 wake_up_all(&ci->i_cap_wq);
977 * Completely synchronous read and write methods. Direct from __user
978 * buffer to osd, or directly to user pages (if O_DIRECT).
980 * If the read spans object boundary, just do multiple reads. (That's not
981 * atomic, but good enough for now.)
983 * If we get a short result from the OSD, check against i_size; we need to
984 * only return a short read to the caller if we hit EOF.
986 ssize_t __ceph_sync_read(struct inode *inode, loff_t *ki_pos,
987 struct iov_iter *to, int *retry_op,
990 struct ceph_inode_info *ci = ceph_inode(inode);
991 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
992 struct ceph_client *cl = fsc->client;
993 struct ceph_osd_client *osdc = &fsc->client->osdc;
996 u64 len = iov_iter_count(to);
997 u64 i_size = i_size_read(inode);
998 bool sparse = IS_ENCRYPTED(inode) || ceph_test_mount_opt(fsc, SPARSEREAD);
1001 doutc(cl, "on inode %p %llx.%llx %llx~%llx\n", inode,
1002 ceph_vinop(inode), *ki_pos, len);
1004 if (ceph_inode_is_shutdown(inode))
1010 * flush any page cache pages in this range. this
1011 * will make concurrent normal and sync io slow,
1012 * but it will at least behave sensibly when they are
1015 ret = filemap_write_and_wait_range(inode->i_mapping,
1016 off, off + len - 1);
1021 while ((len = iov_iter_count(to)) > 0) {
1022 struct ceph_osd_request *req;
1023 struct page **pages;
1029 struct ceph_osd_req_op *op;
1034 /* determine new offset/length if encrypted */
1035 ceph_fscrypt_adjust_off_and_len(inode, &read_off, &read_len);
1037 doutc(cl, "orig %llu~%llu reading %llu~%llu", off, len,
1038 read_off, read_len);
1040 req = ceph_osdc_new_request(osdc, &ci->i_layout,
1041 ci->i_vino, read_off, &read_len, 0, 1,
1042 sparse ? CEPH_OSD_OP_SPARSE_READ :
1045 NULL, ci->i_truncate_seq,
1046 ci->i_truncate_size, false);
1052 /* adjust len downward if the request truncated the len */
1053 if (off + len > read_off + read_len)
1054 len = read_off + read_len - off;
1055 more = len < iov_iter_count(to);
1057 num_pages = calc_pages_for(read_off, read_len);
1058 page_off = offset_in_page(off);
1059 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1060 if (IS_ERR(pages)) {
1061 ceph_osdc_put_request(req);
1062 ret = PTR_ERR(pages);
1066 osd_req_op_extent_osd_data_pages(req, 0, pages, read_len,
1067 offset_in_page(read_off),
1070 op = &req->r_ops[0];
1072 extent_cnt = __ceph_sparse_read_ext_count(inode, read_len);
1073 ret = ceph_alloc_sparse_ext_map(op, extent_cnt);
1075 ceph_osdc_put_request(req);
1080 ceph_osdc_start_request(osdc, req);
1081 ret = ceph_osdc_wait_request(osdc, req);
1083 ceph_update_read_metrics(&fsc->mdsc->metric,
1084 req->r_start_latency,
1089 objver = req->r_version;
1091 i_size = i_size_read(inode);
1092 doutc(cl, "%llu~%llu got %zd i_size %llu%s\n", off, len,
1093 ret, i_size, (more ? " MORE" : ""));
1095 /* Fix it to go to end of extent map */
1096 if (sparse && ret >= 0)
1097 ret = ceph_sparse_ext_map_end(op);
1098 else if (ret == -ENOENT)
1101 if (ret > 0 && IS_ENCRYPTED(inode)) {
1104 fret = ceph_fscrypt_decrypt_extents(inode, pages,
1105 read_off, op->extent.sparse_ext,
1106 op->extent.sparse_ext_cnt);
1109 ceph_osdc_put_request(req);
1113 /* account for any partial block at the beginning */
1114 fret -= (off - read_off);
1117 * Short read after big offset adjustment?
1118 * Nothing is usable, just call it a zero
1121 fret = max(fret, 0);
1123 /* account for partial block at the end */
1124 ret = min_t(ssize_t, fret, len);
1127 ceph_osdc_put_request(req);
1129 /* Short read but not EOF? Zero out the remainder. */
1130 if (ret >= 0 && ret < len && (off + ret < i_size)) {
1131 int zlen = min(len - ret, i_size - off - ret);
1132 int zoff = page_off + ret;
1134 doutc(cl, "zero gap %llu~%llu\n", off + ret,
1136 ceph_zero_page_vector_range(zoff, zlen, pages);
1141 left = ret > 0 ? ret : 0;
1143 size_t plen, copied;
1145 plen = min_t(size_t, left, PAGE_SIZE - page_off);
1146 SetPageUptodate(pages[idx]);
1147 copied = copy_page_to_iter(pages[idx++],
1148 page_off, plen, to);
1152 if (copied < plen) {
1157 ceph_release_page_vector(pages, num_pages);
1160 if (ret == -EBLOCKLISTED)
1161 fsc->blocklisted = true;
1165 if (off >= i_size || !more)
1170 if (off > *ki_pos) {
1171 if (off >= i_size) {
1172 *retry_op = CHECK_EOF;
1173 ret = i_size - *ki_pos;
1176 ret = off - *ki_pos;
1182 *last_objver = objver;
1184 doutc(cl, "result %zd retry_op %d\n", ret, *retry_op);
1188 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
1191 struct file *file = iocb->ki_filp;
1192 struct inode *inode = file_inode(file);
1193 struct ceph_client *cl = ceph_inode_to_client(inode);
1195 doutc(cl, "on file %p %llx~%zx %s\n", file, iocb->ki_pos,
1197 (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
1199 return __ceph_sync_read(inode, &iocb->ki_pos, to, retry_op, NULL);
1202 struct ceph_aio_request {
1208 struct list_head osd_reqs;
1210 atomic_t pending_reqs;
1211 struct timespec64 mtime;
1212 struct ceph_cap_flush *prealloc_cf;
1215 struct ceph_aio_work {
1216 struct work_struct work;
1217 struct ceph_osd_request *req;
1220 static void ceph_aio_retry_work(struct work_struct *work);
1222 static void ceph_aio_complete(struct inode *inode,
1223 struct ceph_aio_request *aio_req)
1225 struct ceph_client *cl = ceph_inode_to_client(inode);
1226 struct ceph_inode_info *ci = ceph_inode(inode);
1229 if (!atomic_dec_and_test(&aio_req->pending_reqs))
1232 if (aio_req->iocb->ki_flags & IOCB_DIRECT)
1233 inode_dio_end(inode);
1235 ret = aio_req->error;
1237 ret = aio_req->total_len;
1239 doutc(cl, "%p %llx.%llx rc %d\n", inode, ceph_vinop(inode), ret);
1241 if (ret >= 0 && aio_req->write) {
1244 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
1245 if (endoff > i_size_read(inode)) {
1246 if (ceph_inode_set_size(inode, endoff))
1247 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY);
1250 spin_lock(&ci->i_ceph_lock);
1251 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1252 &aio_req->prealloc_cf);
1253 spin_unlock(&ci->i_ceph_lock);
1255 __mark_inode_dirty(inode, dirty);
1259 ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
1262 aio_req->iocb->ki_complete(aio_req->iocb, ret);
1264 ceph_free_cap_flush(aio_req->prealloc_cf);
1268 static void ceph_aio_complete_req(struct ceph_osd_request *req)
1270 int rc = req->r_result;
1271 struct inode *inode = req->r_inode;
1272 struct ceph_aio_request *aio_req = req->r_priv;
1273 struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
1274 struct ceph_osd_req_op *op = &req->r_ops[0];
1275 struct ceph_client_metric *metric = &ceph_sb_to_mdsc(inode->i_sb)->metric;
1276 unsigned int len = osd_data->bvec_pos.iter.bi_size;
1277 bool sparse = (op->op == CEPH_OSD_OP_SPARSE_READ);
1278 struct ceph_client *cl = ceph_inode_to_client(inode);
1280 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_BVECS);
1281 BUG_ON(!osd_data->num_bvecs);
1283 doutc(cl, "req %p inode %p %llx.%llx, rc %d bytes %u\n", req,
1284 inode, ceph_vinop(inode), rc, len);
1286 if (rc == -EOLDSNAPC) {
1287 struct ceph_aio_work *aio_work;
1288 BUG_ON(!aio_req->write);
1290 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
1292 INIT_WORK(&aio_work->work, ceph_aio_retry_work);
1293 aio_work->req = req;
1294 queue_work(ceph_inode_to_fs_client(inode)->inode_wq,
1299 } else if (!aio_req->write) {
1300 if (sparse && rc >= 0)
1301 rc = ceph_sparse_ext_map_end(op);
1304 if (rc >= 0 && len > rc) {
1306 int zlen = len - rc;
1309 * If read is satisfied by single OSD request,
1310 * it can pass EOF. Otherwise read is within
1313 if (aio_req->num_reqs == 1) {
1314 loff_t i_size = i_size_read(inode);
1315 loff_t endoff = aio_req->iocb->ki_pos + rc;
1316 if (endoff < i_size)
1317 zlen = min_t(size_t, zlen,
1319 aio_req->total_len = rc + zlen;
1322 iov_iter_bvec(&i, ITER_DEST, osd_data->bvec_pos.bvecs,
1323 osd_data->num_bvecs, len);
1324 iov_iter_advance(&i, rc);
1325 iov_iter_zero(zlen, &i);
1329 /* r_start_latency == 0 means the request was not submitted */
1330 if (req->r_start_latency) {
1332 ceph_update_write_metrics(metric, req->r_start_latency,
1333 req->r_end_latency, len, rc);
1335 ceph_update_read_metrics(metric, req->r_start_latency,
1336 req->r_end_latency, len, rc);
1339 put_bvecs(osd_data->bvec_pos.bvecs, osd_data->num_bvecs,
1340 aio_req->should_dirty);
1341 ceph_osdc_put_request(req);
1344 cmpxchg(&aio_req->error, 0, rc);
1346 ceph_aio_complete(inode, aio_req);
1350 static void ceph_aio_retry_work(struct work_struct *work)
1352 struct ceph_aio_work *aio_work =
1353 container_of(work, struct ceph_aio_work, work);
1354 struct ceph_osd_request *orig_req = aio_work->req;
1355 struct ceph_aio_request *aio_req = orig_req->r_priv;
1356 struct inode *inode = orig_req->r_inode;
1357 struct ceph_inode_info *ci = ceph_inode(inode);
1358 struct ceph_snap_context *snapc;
1359 struct ceph_osd_request *req;
1362 spin_lock(&ci->i_ceph_lock);
1363 if (__ceph_have_pending_cap_snap(ci)) {
1364 struct ceph_cap_snap *capsnap =
1365 list_last_entry(&ci->i_cap_snaps,
1366 struct ceph_cap_snap,
1368 snapc = ceph_get_snap_context(capsnap->context);
1370 BUG_ON(!ci->i_head_snapc);
1371 snapc = ceph_get_snap_context(ci->i_head_snapc);
1373 spin_unlock(&ci->i_ceph_lock);
1375 req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 1,
1383 req->r_flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1384 ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
1385 ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
1387 req->r_ops[0] = orig_req->r_ops[0];
1389 req->r_mtime = aio_req->mtime;
1390 req->r_data_offset = req->r_ops[0].extent.offset;
1392 ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
1394 ceph_osdc_put_request(req);
1399 ceph_osdc_put_request(orig_req);
1401 req->r_callback = ceph_aio_complete_req;
1402 req->r_inode = inode;
1403 req->r_priv = aio_req;
1405 ceph_osdc_start_request(req->r_osdc, req);
1408 req->r_result = ret;
1409 ceph_aio_complete_req(req);
1412 ceph_put_snap_context(snapc);
1417 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
1418 struct ceph_snap_context *snapc,
1419 struct ceph_cap_flush **pcf)
1421 struct file *file = iocb->ki_filp;
1422 struct inode *inode = file_inode(file);
1423 struct ceph_inode_info *ci = ceph_inode(inode);
1424 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
1425 struct ceph_client *cl = fsc->client;
1426 struct ceph_client_metric *metric = &fsc->mdsc->metric;
1427 struct ceph_vino vino;
1428 struct ceph_osd_request *req;
1429 struct bio_vec *bvecs;
1430 struct ceph_aio_request *aio_req = NULL;
1434 struct timespec64 mtime = current_time(inode);
1435 size_t count = iov_iter_count(iter);
1436 loff_t pos = iocb->ki_pos;
1437 bool write = iov_iter_rw(iter) == WRITE;
1438 bool should_dirty = !write && user_backed_iter(iter);
1439 bool sparse = ceph_test_mount_opt(fsc, SPARSEREAD);
1441 if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1444 doutc(cl, "sync_direct_%s on file %p %lld~%u snapc %p seq %lld\n",
1445 (write ? "write" : "read"), file, pos, (unsigned)count,
1446 snapc, snapc ? snapc->seq : 0);
1451 ceph_fscache_invalidate(inode, true);
1453 ret2 = invalidate_inode_pages2_range(inode->i_mapping,
1455 (pos + count - 1) >> PAGE_SHIFT);
1457 doutc(cl, "invalidate_inode_pages2_range returned %d\n",
1460 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1462 flags = CEPH_OSD_FLAG_READ;
1465 while (iov_iter_count(iter) > 0) {
1466 u64 size = iov_iter_count(iter);
1468 struct ceph_osd_req_op *op;
1469 int readop = sparse ? CEPH_OSD_OP_SPARSE_READ : CEPH_OSD_OP_READ;
1473 size = min_t(u64, size, fsc->mount_options->wsize);
1475 size = min_t(u64, size, fsc->mount_options->rsize);
1477 vino = ceph_vino(inode);
1478 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1479 vino, pos, &size, 0,
1481 write ? CEPH_OSD_OP_WRITE : readop,
1484 ci->i_truncate_size,
1491 len = iter_get_bvecs_alloc(iter, size, &bvecs, &num_pages);
1493 ceph_osdc_put_request(req);
1498 osd_req_op_extent_update(req, 0, len);
1501 * To simplify error handling, allow AIO when IO within i_size
1502 * or IO can be satisfied by single OSD request.
1504 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
1505 (len == count || pos + count <= i_size_read(inode))) {
1506 aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
1508 aio_req->iocb = iocb;
1509 aio_req->write = write;
1510 aio_req->should_dirty = should_dirty;
1511 INIT_LIST_HEAD(&aio_req->osd_reqs);
1513 aio_req->mtime = mtime;
1514 swap(aio_req->prealloc_cf, *pcf);
1522 * throw out any page cache pages in this range. this
1525 truncate_inode_pages_range(inode->i_mapping, pos,
1526 PAGE_ALIGN(pos + len) - 1);
1528 req->r_mtime = mtime;
1531 osd_req_op_extent_osd_data_bvecs(req, 0, bvecs, num_pages, len);
1532 op = &req->r_ops[0];
1534 extent_cnt = __ceph_sparse_read_ext_count(inode, size);
1535 ret = ceph_alloc_sparse_ext_map(op, extent_cnt);
1537 ceph_osdc_put_request(req);
1543 aio_req->total_len += len;
1544 aio_req->num_reqs++;
1545 atomic_inc(&aio_req->pending_reqs);
1547 req->r_callback = ceph_aio_complete_req;
1548 req->r_inode = inode;
1549 req->r_priv = aio_req;
1550 list_add_tail(&req->r_private_item, &aio_req->osd_reqs);
1556 ceph_osdc_start_request(req->r_osdc, req);
1557 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1560 ceph_update_write_metrics(metric, req->r_start_latency,
1561 req->r_end_latency, len, ret);
1563 ceph_update_read_metrics(metric, req->r_start_latency,
1564 req->r_end_latency, len, ret);
1566 size = i_size_read(inode);
1568 if (sparse && ret >= 0)
1569 ret = ceph_sparse_ext_map_end(op);
1570 else if (ret == -ENOENT)
1573 if (ret >= 0 && ret < len && pos + ret < size) {
1575 int zlen = min_t(size_t, len - ret,
1578 iov_iter_bvec(&i, ITER_DEST, bvecs, num_pages, len);
1579 iov_iter_advance(&i, ret);
1580 iov_iter_zero(zlen, &i);
1587 put_bvecs(bvecs, num_pages, should_dirty);
1588 ceph_osdc_put_request(req);
1593 if (!write && pos >= size)
1596 if (write && pos > size) {
1597 if (ceph_inode_set_size(inode, pos))
1598 ceph_check_caps(ceph_inode(inode),
1599 CHECK_CAPS_AUTHONLY);
1604 LIST_HEAD(osd_reqs);
1606 if (aio_req->num_reqs == 0) {
1611 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1614 list_splice(&aio_req->osd_reqs, &osd_reqs);
1615 inode_dio_begin(inode);
1616 while (!list_empty(&osd_reqs)) {
1617 req = list_first_entry(&osd_reqs,
1618 struct ceph_osd_request,
1620 list_del_init(&req->r_private_item);
1622 ceph_osdc_start_request(req->r_osdc, req);
1624 req->r_result = ret;
1625 ceph_aio_complete_req(req);
1628 return -EIOCBQUEUED;
1631 if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1632 ret = pos - iocb->ki_pos;
1639 * Synchronous write, straight from __user pointer or user pages.
1641 * If write spans object boundary, just do multiple writes. (For a
1642 * correct atomic write, we should e.g. take write locks on all
1643 * objects, rollback on failure, etc.)
1646 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1647 struct ceph_snap_context *snapc)
1649 struct file *file = iocb->ki_filp;
1650 struct inode *inode = file_inode(file);
1651 struct ceph_inode_info *ci = ceph_inode(inode);
1652 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
1653 struct ceph_client *cl = fsc->client;
1654 struct ceph_osd_client *osdc = &fsc->client->osdc;
1655 struct ceph_osd_request *req;
1656 struct page **pages;
1661 bool check_caps = false;
1662 struct timespec64 mtime = current_time(inode);
1663 size_t count = iov_iter_count(from);
1665 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1668 doutc(cl, "on file %p %lld~%u snapc %p seq %lld\n", file, pos,
1669 (unsigned)count, snapc, snapc->seq);
1671 ret = filemap_write_and_wait_range(inode->i_mapping,
1672 pos, pos + count - 1);
1676 ceph_fscache_invalidate(inode, false);
1678 while ((len = iov_iter_count(from)) > 0) {
1681 u64 write_pos = pos;
1682 u64 write_len = len;
1688 struct iov_iter saved_iter = *from;
1691 ceph_fscrypt_adjust_off_and_len(inode, &write_pos, &write_len);
1693 /* clamp the length to the end of first object */
1694 ceph_calc_file_object_mapping(&ci->i_layout, write_pos,
1695 write_len, &objnum, &objoff,
1699 /* adjust len downward if it goes beyond current object */
1700 if (pos + len > write_pos + write_len)
1701 len = write_pos + write_len - pos;
1704 * If we had to adjust the length or position to align with a
1705 * crypto block, then we must do a read/modify/write cycle. We
1706 * use a version assertion to redrive the thing if something
1707 * changes in between.
1709 first = pos != write_pos;
1710 last = (pos + len) != (write_pos + write_len);
1711 rmw = first || last;
1713 doutc(cl, "ino %llx %lld~%llu adjusted %lld~%llu -- %srmw\n",
1714 ci->i_vino.ino, pos, len, write_pos, write_len,
1718 * The data is emplaced into the page as it would be if it were
1719 * in an array of pagecache pages.
1721 num_pages = calc_pages_for(write_pos, write_len);
1722 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1723 if (IS_ERR(pages)) {
1724 ret = PTR_ERR(pages);
1728 /* Do we need to preload the pages? */
1730 u64 first_pos = write_pos;
1731 u64 last_pos = (write_pos + write_len) - CEPH_FSCRYPT_BLOCK_SIZE;
1732 u64 read_len = CEPH_FSCRYPT_BLOCK_SIZE;
1733 struct ceph_osd_req_op *op;
1735 /* We should only need to do this for encrypted inodes */
1736 WARN_ON_ONCE(!IS_ENCRYPTED(inode));
1738 /* No need to do two reads if first and last blocks are same */
1739 if (first && last_pos == first_pos)
1743 * Allocate a read request for one or two extents,
1744 * depending on how the request was aligned.
1746 req = ceph_osdc_new_request(osdc, &ci->i_layout,
1747 ci->i_vino, first ? first_pos : last_pos,
1748 &read_len, 0, (first && last) ? 2 : 1,
1749 CEPH_OSD_OP_SPARSE_READ, CEPH_OSD_FLAG_READ,
1750 NULL, ci->i_truncate_seq,
1751 ci->i_truncate_size, false);
1753 ceph_release_page_vector(pages, num_pages);
1758 /* Something is misaligned! */
1759 if (read_len != CEPH_FSCRYPT_BLOCK_SIZE) {
1760 ceph_osdc_put_request(req);
1761 ceph_release_page_vector(pages, num_pages);
1766 /* Add extent for first block? */
1767 op = &req->r_ops[0];
1770 osd_req_op_extent_osd_data_pages(req, 0, pages,
1771 CEPH_FSCRYPT_BLOCK_SIZE,
1772 offset_in_page(first_pos),
1774 /* We only expect a single extent here */
1775 ret = __ceph_alloc_sparse_ext_map(op, 1);
1777 ceph_osdc_put_request(req);
1778 ceph_release_page_vector(pages, num_pages);
1783 /* Add extent for last block */
1785 /* Init the other extent if first extent has been used */
1787 op = &req->r_ops[1];
1788 osd_req_op_extent_init(req, 1,
1789 CEPH_OSD_OP_SPARSE_READ,
1790 last_pos, CEPH_FSCRYPT_BLOCK_SIZE,
1791 ci->i_truncate_size,
1792 ci->i_truncate_seq);
1795 ret = __ceph_alloc_sparse_ext_map(op, 1);
1797 ceph_osdc_put_request(req);
1798 ceph_release_page_vector(pages, num_pages);
1802 osd_req_op_extent_osd_data_pages(req, first ? 1 : 0,
1803 &pages[num_pages - 1],
1804 CEPH_FSCRYPT_BLOCK_SIZE,
1805 offset_in_page(last_pos),
1809 ceph_osdc_start_request(osdc, req);
1810 ret = ceph_osdc_wait_request(osdc, req);
1812 /* FIXME: length field is wrong if there are 2 extents */
1813 ceph_update_read_metrics(&fsc->mdsc->metric,
1814 req->r_start_latency,
1818 /* Ok if object is not already present */
1819 if (ret == -ENOENT) {
1821 * If there is no object, then we can't assert
1822 * on its version. Set it to 0, and we'll use an
1823 * exclusive create instead.
1825 ceph_osdc_put_request(req);
1829 * zero out the soon-to-be uncopied parts of the
1830 * first and last pages.
1833 zero_user_segment(pages[0], 0,
1834 offset_in_page(first_pos));
1836 zero_user_segment(pages[num_pages - 1],
1837 offset_in_page(last_pos),
1841 ceph_osdc_put_request(req);
1842 ceph_release_page_vector(pages, num_pages);
1846 op = &req->r_ops[0];
1847 if (op->extent.sparse_ext_cnt == 0) {
1849 zero_user_segment(pages[0], 0,
1850 offset_in_page(first_pos));
1852 zero_user_segment(pages[num_pages - 1],
1853 offset_in_page(last_pos),
1855 } else if (op->extent.sparse_ext_cnt != 1 ||
1856 ceph_sparse_ext_map_end(op) !=
1857 CEPH_FSCRYPT_BLOCK_SIZE) {
1859 ceph_osdc_put_request(req);
1860 ceph_release_page_vector(pages, num_pages);
1864 if (first && last) {
1865 op = &req->r_ops[1];
1866 if (op->extent.sparse_ext_cnt == 0) {
1867 zero_user_segment(pages[num_pages - 1],
1868 offset_in_page(last_pos),
1870 } else if (op->extent.sparse_ext_cnt != 1 ||
1871 ceph_sparse_ext_map_end(op) !=
1872 CEPH_FSCRYPT_BLOCK_SIZE) {
1874 ceph_osdc_put_request(req);
1875 ceph_release_page_vector(pages, num_pages);
1880 /* Grab assert version. It must be non-zero. */
1881 assert_ver = req->r_version;
1882 WARN_ON_ONCE(ret > 0 && assert_ver == 0);
1884 ceph_osdc_put_request(req);
1886 ret = ceph_fscrypt_decrypt_block_inplace(inode,
1887 pages[0], CEPH_FSCRYPT_BLOCK_SIZE,
1888 offset_in_page(first_pos),
1889 first_pos >> CEPH_FSCRYPT_BLOCK_SHIFT);
1891 ceph_release_page_vector(pages, num_pages);
1896 ret = ceph_fscrypt_decrypt_block_inplace(inode,
1897 pages[num_pages - 1],
1898 CEPH_FSCRYPT_BLOCK_SIZE,
1899 offset_in_page(last_pos),
1900 last_pos >> CEPH_FSCRYPT_BLOCK_SHIFT);
1902 ceph_release_page_vector(pages, num_pages);
1910 off = offset_in_page(pos);
1911 for (n = 0; n < num_pages; n++) {
1912 size_t plen = min_t(size_t, left, PAGE_SIZE - off);
1915 ret = copy_page_from_iter(pages[n], off, plen, from);
1924 doutc(cl, "write failed with %d\n", ret);
1925 ceph_release_page_vector(pages, num_pages);
1929 if (IS_ENCRYPTED(inode)) {
1930 ret = ceph_fscrypt_encrypt_pages(inode, pages,
1931 write_pos, write_len,
1934 doutc(cl, "encryption failed with %d\n", ret);
1935 ceph_release_page_vector(pages, num_pages);
1940 req = ceph_osdc_new_request(osdc, &ci->i_layout,
1941 ci->i_vino, write_pos, &write_len,
1942 rmw ? 1 : 0, rmw ? 2 : 1,
1944 CEPH_OSD_FLAG_WRITE,
1945 snapc, ci->i_truncate_seq,
1946 ci->i_truncate_size, false);
1949 ceph_release_page_vector(pages, num_pages);
1953 doutc(cl, "write op %lld~%llu\n", write_pos, write_len);
1954 osd_req_op_extent_osd_data_pages(req, rmw ? 1 : 0, pages, write_len,
1955 offset_in_page(write_pos), false,
1957 req->r_inode = inode;
1958 req->r_mtime = mtime;
1960 /* Set up the assertion */
1963 * Set up the assertion. If we don't have a version
1964 * number, then the object doesn't exist yet. Use an
1965 * exclusive create instead of a version assertion in
1969 osd_req_op_init(req, 0, CEPH_OSD_OP_ASSERT_VER, 0);
1970 req->r_ops[0].assert_ver.ver = assert_ver;
1972 osd_req_op_init(req, 0, CEPH_OSD_OP_CREATE,
1973 CEPH_OSD_OP_FLAG_EXCL);
1977 ceph_osdc_start_request(osdc, req);
1978 ret = ceph_osdc_wait_request(osdc, req);
1980 ceph_update_write_metrics(&fsc->mdsc->metric, req->r_start_latency,
1981 req->r_end_latency, len, ret);
1982 ceph_osdc_put_request(req);
1984 doutc(cl, "osd write returned %d\n", ret);
1985 /* Version changed! Must re-do the rmw cycle */
1986 if ((assert_ver && (ret == -ERANGE || ret == -EOVERFLOW)) ||
1987 (!assert_ver && ret == -EEXIST)) {
1988 /* We should only ever see this on a rmw */
1991 /* The version should never go backward */
1992 WARN_ON_ONCE(ret == -EOVERFLOW);
1996 /* FIXME: limit number of times we loop? */
1999 ceph_set_error_write(ci);
2003 ceph_clear_error_write(ci);
2006 * We successfully wrote to a range of the file. Declare
2007 * that region of the pagecache invalid.
2009 ret = invalidate_inode_pages2_range(
2012 (pos + len - 1) >> PAGE_SHIFT);
2014 doutc(cl, "invalidate_inode_pages2_range returned %d\n",
2020 doutc(cl, "written %d\n", written);
2021 if (pos > i_size_read(inode)) {
2022 check_caps = ceph_inode_set_size(inode, pos);
2024 ceph_check_caps(ceph_inode(inode),
2025 CHECK_CAPS_AUTHONLY);
2030 if (ret != -EOLDSNAPC && written > 0) {
2034 doutc(cl, "returning %d\n", ret);
2039 * Wrap generic_file_aio_read with checks for cap bits on the inode.
2040 * Atomically grab references, so that those bits are not released
2041 * back to the MDS mid-read.
2043 * Hmm, the sync read case isn't actually async... should it be?
2045 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
2047 struct file *filp = iocb->ki_filp;
2048 struct ceph_file_info *fi = filp->private_data;
2049 size_t len = iov_iter_count(to);
2050 struct inode *inode = file_inode(filp);
2051 struct ceph_inode_info *ci = ceph_inode(inode);
2052 bool direct_lock = iocb->ki_flags & IOCB_DIRECT;
2053 struct ceph_client *cl = ceph_inode_to_client(inode);
2055 int want = 0, got = 0;
2056 int retry_op = 0, read = 0;
2059 doutc(cl, "%llu~%u trying to get caps on %p %llx.%llx\n",
2060 iocb->ki_pos, (unsigned)len, inode, ceph_vinop(inode));
2062 if (ceph_inode_is_shutdown(inode))
2066 ceph_start_io_direct(inode);
2068 ceph_start_io_read(inode);
2070 if (!(fi->flags & CEPH_F_SYNC) && !direct_lock)
2071 want |= CEPH_CAP_FILE_CACHE;
2072 if (fi->fmode & CEPH_FILE_MODE_LAZY)
2073 want |= CEPH_CAP_FILE_LAZYIO;
2075 ret = ceph_get_caps(filp, CEPH_CAP_FILE_RD, want, -1, &got);
2078 ceph_end_io_direct(inode);
2080 ceph_end_io_read(inode);
2084 if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
2085 (iocb->ki_flags & IOCB_DIRECT) ||
2086 (fi->flags & CEPH_F_SYNC)) {
2088 doutc(cl, "sync %p %llx.%llx %llu~%u got cap refs on %s\n",
2089 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
2090 ceph_cap_string(got));
2092 if (!ceph_has_inline_data(ci)) {
2094 (iocb->ki_flags & IOCB_DIRECT) &&
2095 !IS_ENCRYPTED(inode)) {
2096 ret = ceph_direct_read_write(iocb, to,
2098 if (ret >= 0 && ret < len)
2099 retry_op = CHECK_EOF;
2101 ret = ceph_sync_read(iocb, to, &retry_op);
2104 retry_op = READ_INLINE;
2107 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
2108 doutc(cl, "async %p %llx.%llx %llu~%u got cap refs on %s\n",
2109 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
2110 ceph_cap_string(got));
2111 ceph_add_rw_context(fi, &rw_ctx);
2112 ret = generic_file_read_iter(iocb, to);
2113 ceph_del_rw_context(fi, &rw_ctx);
2116 doutc(cl, "%p %llx.%llx dropping cap refs on %s = %d\n",
2117 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
2118 ceph_put_cap_refs(ci, got);
2121 ceph_end_io_direct(inode);
2123 ceph_end_io_read(inode);
2125 if (retry_op > HAVE_RETRIED && ret >= 0) {
2127 struct page *page = NULL;
2129 if (retry_op == READ_INLINE) {
2130 page = __page_cache_alloc(GFP_KERNEL);
2135 statret = __ceph_do_getattr(inode, page,
2136 CEPH_STAT_CAP_INLINE_DATA, !!page);
2140 if (statret == -ENODATA) {
2141 BUG_ON(retry_op != READ_INLINE);
2147 i_size = i_size_read(inode);
2148 if (retry_op == READ_INLINE) {
2149 BUG_ON(ret > 0 || read > 0);
2150 if (iocb->ki_pos < i_size &&
2151 iocb->ki_pos < PAGE_SIZE) {
2152 loff_t end = min_t(loff_t, i_size,
2153 iocb->ki_pos + len);
2154 end = min_t(loff_t, end, PAGE_SIZE);
2156 zero_user_segment(page, statret, end);
2157 ret = copy_page_to_iter(page,
2158 iocb->ki_pos & ~PAGE_MASK,
2159 end - iocb->ki_pos, to);
2160 iocb->ki_pos += ret;
2163 if (iocb->ki_pos < i_size && read < len) {
2164 size_t zlen = min_t(size_t, len - read,
2165 i_size - iocb->ki_pos);
2166 ret = iov_iter_zero(zlen, to);
2167 iocb->ki_pos += ret;
2170 __free_pages(page, 0);
2174 /* hit EOF or hole? */
2175 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
2177 doutc(cl, "hit hole, ppos %lld < size %lld, reading more\n",
2178 iocb->ki_pos, i_size);
2182 retry_op = HAVE_RETRIED;
2194 * Wrap filemap_splice_read with checks for cap bits on the inode.
2195 * Atomically grab references, so that those bits are not released
2196 * back to the MDS mid-read.
2198 static ssize_t ceph_splice_read(struct file *in, loff_t *ppos,
2199 struct pipe_inode_info *pipe,
2200 size_t len, unsigned int flags)
2202 struct ceph_file_info *fi = in->private_data;
2203 struct inode *inode = file_inode(in);
2204 struct ceph_inode_info *ci = ceph_inode(inode);
2206 int want = 0, got = 0;
2207 CEPH_DEFINE_RW_CONTEXT(rw_ctx, 0);
2209 dout("splice_read %p %llx.%llx %llu~%zu trying to get caps on %p\n",
2210 inode, ceph_vinop(inode), *ppos, len, inode);
2212 if (ceph_inode_is_shutdown(inode))
2215 if (ceph_has_inline_data(ci) ||
2216 (fi->flags & CEPH_F_SYNC))
2217 return copy_splice_read(in, ppos, pipe, len, flags);
2219 ceph_start_io_read(inode);
2221 want = CEPH_CAP_FILE_CACHE;
2222 if (fi->fmode & CEPH_FILE_MODE_LAZY)
2223 want |= CEPH_CAP_FILE_LAZYIO;
2225 ret = ceph_get_caps(in, CEPH_CAP_FILE_RD, want, -1, &got);
2229 if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) == 0) {
2230 dout("splice_read/sync %p %llx.%llx %llu~%zu got cap refs on %s\n",
2231 inode, ceph_vinop(inode), *ppos, len,
2232 ceph_cap_string(got));
2234 ceph_put_cap_refs(ci, got);
2235 ceph_end_io_read(inode);
2236 return copy_splice_read(in, ppos, pipe, len, flags);
2239 dout("splice_read %p %llx.%llx %llu~%zu got cap refs on %s\n",
2240 inode, ceph_vinop(inode), *ppos, len, ceph_cap_string(got));
2243 ceph_add_rw_context(fi, &rw_ctx);
2244 ret = filemap_splice_read(in, ppos, pipe, len, flags);
2245 ceph_del_rw_context(fi, &rw_ctx);
2247 dout("splice_read %p %llx.%llx dropping cap refs on %s = %zd\n",
2248 inode, ceph_vinop(inode), ceph_cap_string(got), ret);
2250 ceph_put_cap_refs(ci, got);
2252 ceph_end_io_read(inode);
2257 * Take cap references to avoid releasing caps to MDS mid-write.
2259 * If we are synchronous, and write with an old snap context, the OSD
2260 * may return EOLDSNAPC. In that case, retry the write.. _after_
2261 * dropping our cap refs and allowing the pending snap to logically
2262 * complete _before_ this write occurs.
2264 * If we are near ENOSPC, write synchronously.
2266 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
2268 struct file *file = iocb->ki_filp;
2269 struct ceph_file_info *fi = file->private_data;
2270 struct inode *inode = file_inode(file);
2271 struct ceph_inode_info *ci = ceph_inode(inode);
2272 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
2273 struct ceph_client *cl = fsc->client;
2274 struct ceph_osd_client *osdc = &fsc->client->osdc;
2275 struct ceph_cap_flush *prealloc_cf;
2276 ssize_t count, written = 0;
2277 int err, want = 0, got;
2278 bool direct_lock = false;
2282 loff_t limit = max(i_size_read(inode), fsc->max_file_size);
2284 if (ceph_inode_is_shutdown(inode))
2287 if (ceph_snap(inode) != CEPH_NOSNAP)
2290 prealloc_cf = ceph_alloc_cap_flush();
2294 if ((iocb->ki_flags & (IOCB_DIRECT | IOCB_APPEND)) == IOCB_DIRECT)
2299 ceph_start_io_direct(inode);
2301 ceph_start_io_write(inode);
2303 if (iocb->ki_flags & IOCB_APPEND) {
2304 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
2309 err = generic_write_checks(iocb, from);
2314 if (unlikely(pos >= limit)) {
2318 iov_iter_truncate(from, limit - pos);
2321 count = iov_iter_count(from);
2322 if (ceph_quota_is_max_bytes_exceeded(inode, pos + count)) {
2327 down_read(&osdc->lock);
2328 map_flags = osdc->osdmap->flags;
2329 pool_flags = ceph_pg_pool_flags(osdc->osdmap, ci->i_layout.pool_id);
2330 up_read(&osdc->lock);
2331 if ((map_flags & CEPH_OSDMAP_FULL) ||
2332 (pool_flags & CEPH_POOL_FLAG_FULL)) {
2337 err = file_remove_privs(file);
2341 doutc(cl, "%p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
2342 inode, ceph_vinop(inode), pos, count,
2343 i_size_read(inode));
2344 if (!(fi->flags & CEPH_F_SYNC) && !direct_lock)
2345 want |= CEPH_CAP_FILE_BUFFER;
2346 if (fi->fmode & CEPH_FILE_MODE_LAZY)
2347 want |= CEPH_CAP_FILE_LAZYIO;
2349 err = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, pos + count, &got);
2353 err = file_update_time(file);
2357 inode_inc_iversion_raw(inode);
2359 doutc(cl, "%p %llx.%llx %llu~%zd got cap refs on %s\n",
2360 inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
2362 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
2363 (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
2364 (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
2365 struct ceph_snap_context *snapc;
2366 struct iov_iter data;
2368 spin_lock(&ci->i_ceph_lock);
2369 if (__ceph_have_pending_cap_snap(ci)) {
2370 struct ceph_cap_snap *capsnap =
2371 list_last_entry(&ci->i_cap_snaps,
2372 struct ceph_cap_snap,
2374 snapc = ceph_get_snap_context(capsnap->context);
2376 BUG_ON(!ci->i_head_snapc);
2377 snapc = ceph_get_snap_context(ci->i_head_snapc);
2379 spin_unlock(&ci->i_ceph_lock);
2381 /* we might need to revert back to that point */
2383 if ((iocb->ki_flags & IOCB_DIRECT) && !IS_ENCRYPTED(inode))
2384 written = ceph_direct_read_write(iocb, &data, snapc,
2387 written = ceph_sync_write(iocb, &data, pos, snapc);
2389 ceph_end_io_direct(inode);
2391 ceph_end_io_write(inode);
2393 iov_iter_advance(from, written);
2394 ceph_put_snap_context(snapc);
2397 * No need to acquire the i_truncate_mutex. Because
2398 * the MDS revokes Fwb caps before sending truncate
2399 * message to us. We can't get Fwb cap while there
2400 * are pending vmtruncate. So write and vmtruncate
2401 * can not run at the same time
2403 written = generic_perform_write(iocb, from);
2404 ceph_end_io_write(inode);
2410 spin_lock(&ci->i_ceph_lock);
2411 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
2413 spin_unlock(&ci->i_ceph_lock);
2415 __mark_inode_dirty(inode, dirty);
2416 if (ceph_quota_is_max_bytes_approaching(inode, iocb->ki_pos))
2417 ceph_check_caps(ci, CHECK_CAPS_FLUSH);
2420 doutc(cl, "%p %llx.%llx %llu~%u dropping cap refs on %s\n",
2421 inode, ceph_vinop(inode), pos, (unsigned)count,
2422 ceph_cap_string(got));
2423 ceph_put_cap_refs(ci, got);
2425 if (written == -EOLDSNAPC) {
2426 doutc(cl, "%p %llx.%llx %llu~%u" "got EOLDSNAPC, retrying\n",
2427 inode, ceph_vinop(inode), pos, (unsigned)count);
2432 if ((map_flags & CEPH_OSDMAP_NEARFULL) ||
2433 (pool_flags & CEPH_POOL_FLAG_NEARFULL))
2434 iocb->ki_flags |= IOCB_DSYNC;
2435 written = generic_write_sync(iocb, written);
2440 ceph_put_cap_refs(ci, got);
2443 ceph_end_io_direct(inode);
2445 ceph_end_io_write(inode);
2447 ceph_free_cap_flush(prealloc_cf);
2448 return written ? written : err;
2452 * llseek. be sure to verify file size on SEEK_END.
2454 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
2456 if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
2457 struct inode *inode = file_inode(file);
2460 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
2464 return generic_file_llseek(file, offset, whence);
2467 static inline void ceph_zero_partial_page(
2468 struct inode *inode, loff_t offset, unsigned size)
2471 pgoff_t index = offset >> PAGE_SHIFT;
2473 page = find_lock_page(inode->i_mapping, index);
2475 wait_on_page_writeback(page);
2476 zero_user(page, offset & (PAGE_SIZE - 1), size);
2482 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
2485 loff_t nearly = round_up(offset, PAGE_SIZE);
2486 if (offset < nearly) {
2487 loff_t size = nearly - offset;
2490 ceph_zero_partial_page(inode, offset, size);
2494 if (length >= PAGE_SIZE) {
2495 loff_t size = round_down(length, PAGE_SIZE);
2496 truncate_pagecache_range(inode, offset, offset + size - 1);
2501 ceph_zero_partial_page(inode, offset, length);
2504 static int ceph_zero_partial_object(struct inode *inode,
2505 loff_t offset, loff_t *length)
2507 struct ceph_inode_info *ci = ceph_inode(inode);
2508 struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
2509 struct ceph_osd_request *req;
2514 if (ceph_inode_is_shutdown(inode))
2518 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
2521 op = CEPH_OSD_OP_ZERO;
2524 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
2528 CEPH_OSD_FLAG_WRITE,
2535 req->r_mtime = inode_get_mtime(inode);
2536 ceph_osdc_start_request(&fsc->client->osdc, req);
2537 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
2540 ceph_osdc_put_request(req);
2546 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
2549 struct ceph_inode_info *ci = ceph_inode(inode);
2550 s32 stripe_unit = ci->i_layout.stripe_unit;
2551 s32 stripe_count = ci->i_layout.stripe_count;
2552 s32 object_size = ci->i_layout.object_size;
2553 u64 object_set_size = object_size * stripe_count;
2556 /* round offset up to next period boundary */
2557 nearly = offset + object_set_size - 1;
2559 nearly -= do_div(t, object_set_size);
2561 while (length && offset < nearly) {
2562 loff_t size = length;
2563 ret = ceph_zero_partial_object(inode, offset, &size);
2569 while (length >= object_set_size) {
2571 loff_t pos = offset;
2572 for (i = 0; i < stripe_count; ++i) {
2573 ret = ceph_zero_partial_object(inode, pos, NULL);
2578 offset += object_set_size;
2579 length -= object_set_size;
2582 loff_t size = length;
2583 ret = ceph_zero_partial_object(inode, offset, &size);
2592 static long ceph_fallocate(struct file *file, int mode,
2593 loff_t offset, loff_t length)
2595 struct ceph_file_info *fi = file->private_data;
2596 struct inode *inode = file_inode(file);
2597 struct ceph_inode_info *ci = ceph_inode(inode);
2598 struct ceph_cap_flush *prealloc_cf;
2599 struct ceph_client *cl = ceph_inode_to_client(inode);
2606 doutc(cl, "%p %llx.%llx mode %x, offset %llu length %llu\n",
2607 inode, ceph_vinop(inode), mode, offset, length);
2609 if (mode != (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2612 if (!S_ISREG(inode->i_mode))
2615 if (IS_ENCRYPTED(inode))
2618 prealloc_cf = ceph_alloc_cap_flush();
2624 if (ceph_snap(inode) != CEPH_NOSNAP) {
2629 size = i_size_read(inode);
2631 /* Are we punching a hole beyond EOF? */
2634 if ((offset + length) > size)
2635 length = size - offset;
2637 if (fi->fmode & CEPH_FILE_MODE_LAZY)
2638 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
2640 want = CEPH_CAP_FILE_BUFFER;
2642 ret = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, endoff, &got);
2646 ret = file_modified(file);
2650 filemap_invalidate_lock(inode->i_mapping);
2651 ceph_fscache_invalidate(inode, false);
2652 ceph_zero_pagecache_range(inode, offset, length);
2653 ret = ceph_zero_objects(inode, offset, length);
2656 spin_lock(&ci->i_ceph_lock);
2657 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
2659 spin_unlock(&ci->i_ceph_lock);
2661 __mark_inode_dirty(inode, dirty);
2663 filemap_invalidate_unlock(inode->i_mapping);
2666 ceph_put_cap_refs(ci, got);
2668 inode_unlock(inode);
2669 ceph_free_cap_flush(prealloc_cf);
2674 * This function tries to get FILE_WR capabilities for dst_ci and FILE_RD for
2675 * src_ci. Two attempts are made to obtain both caps, and an error is return if
2676 * this fails; zero is returned on success.
2678 static int get_rd_wr_caps(struct file *src_filp, int *src_got,
2679 struct file *dst_filp,
2680 loff_t dst_endoff, int *dst_got)
2683 bool retrying = false;
2686 ret = ceph_get_caps(dst_filp, CEPH_CAP_FILE_WR, CEPH_CAP_FILE_BUFFER,
2687 dst_endoff, dst_got);
2692 * Since we're already holding the FILE_WR capability for the dst file,
2693 * we would risk a deadlock by using ceph_get_caps. Thus, we'll do some
2694 * retry dance instead to try to get both capabilities.
2696 ret = ceph_try_get_caps(file_inode(src_filp),
2697 CEPH_CAP_FILE_RD, CEPH_CAP_FILE_SHARED,
2700 /* Start by dropping dst_ci caps and getting src_ci caps */
2701 ceph_put_cap_refs(ceph_inode(file_inode(dst_filp)), *dst_got);
2704 /* ceph_try_get_caps masks EAGAIN */
2708 ret = ceph_get_caps(src_filp, CEPH_CAP_FILE_RD,
2709 CEPH_CAP_FILE_SHARED, -1, src_got);
2712 /*... drop src_ci caps too, and retry */
2713 ceph_put_cap_refs(ceph_inode(file_inode(src_filp)), *src_got);
2720 static void put_rd_wr_caps(struct ceph_inode_info *src_ci, int src_got,
2721 struct ceph_inode_info *dst_ci, int dst_got)
2723 ceph_put_cap_refs(src_ci, src_got);
2724 ceph_put_cap_refs(dst_ci, dst_got);
2728 * This function does several size-related checks, returning an error if:
2729 * - source file is smaller than off+len
2730 * - destination file size is not OK (inode_newsize_ok())
2731 * - max bytes quotas is exceeded
2733 static int is_file_size_ok(struct inode *src_inode, struct inode *dst_inode,
2734 loff_t src_off, loff_t dst_off, size_t len)
2736 struct ceph_client *cl = ceph_inode_to_client(src_inode);
2737 loff_t size, endoff;
2739 size = i_size_read(src_inode);
2741 * Don't copy beyond source file EOF. Instead of simply setting length
2742 * to (size - src_off), just drop to VFS default implementation, as the
2743 * local i_size may be stale due to other clients writing to the source
2746 if (src_off + len > size) {
2747 doutc(cl, "Copy beyond EOF (%llu + %zu > %llu)\n", src_off,
2751 size = i_size_read(dst_inode);
2753 endoff = dst_off + len;
2754 if (inode_newsize_ok(dst_inode, endoff))
2757 if (ceph_quota_is_max_bytes_exceeded(dst_inode, endoff))
2763 static struct ceph_osd_request *
2764 ceph_alloc_copyfrom_request(struct ceph_osd_client *osdc,
2766 struct ceph_object_id *src_oid,
2767 struct ceph_object_locator *src_oloc,
2768 struct ceph_object_id *dst_oid,
2769 struct ceph_object_locator *dst_oloc,
2770 u32 truncate_seq, u64 truncate_size)
2772 struct ceph_osd_request *req;
2774 u32 src_fadvise_flags =
2775 CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2776 CEPH_OSD_OP_FLAG_FADVISE_NOCACHE;
2777 u32 dst_fadvise_flags =
2778 CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2779 CEPH_OSD_OP_FLAG_FADVISE_DONTNEED;
2781 req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_KERNEL);
2783 return ERR_PTR(-ENOMEM);
2785 req->r_flags = CEPH_OSD_FLAG_WRITE;
2787 ceph_oloc_copy(&req->r_t.base_oloc, dst_oloc);
2788 ceph_oid_copy(&req->r_t.base_oid, dst_oid);
2790 ret = osd_req_op_copy_from_init(req, src_snapid, 0,
2796 CEPH_OSD_COPY_FROM_FLAG_TRUNCATE_SEQ);
2800 ret = ceph_osdc_alloc_messages(req, GFP_KERNEL);
2807 ceph_osdc_put_request(req);
2808 return ERR_PTR(ret);
2811 static ssize_t ceph_do_objects_copy(struct ceph_inode_info *src_ci, u64 *src_off,
2812 struct ceph_inode_info *dst_ci, u64 *dst_off,
2813 struct ceph_fs_client *fsc,
2814 size_t len, unsigned int flags)
2816 struct ceph_object_locator src_oloc, dst_oloc;
2817 struct ceph_object_id src_oid, dst_oid;
2818 struct ceph_osd_client *osdc;
2819 struct ceph_osd_request *req;
2821 u64 src_objnum, src_objoff, dst_objnum, dst_objoff;
2822 u32 src_objlen, dst_objlen;
2823 u32 object_size = src_ci->i_layout.object_size;
2824 struct ceph_client *cl = fsc->client;
2827 src_oloc.pool = src_ci->i_layout.pool_id;
2828 src_oloc.pool_ns = ceph_try_get_string(src_ci->i_layout.pool_ns);
2829 dst_oloc.pool = dst_ci->i_layout.pool_id;
2830 dst_oloc.pool_ns = ceph_try_get_string(dst_ci->i_layout.pool_ns);
2831 osdc = &fsc->client->osdc;
2833 while (len >= object_size) {
2834 ceph_calc_file_object_mapping(&src_ci->i_layout, *src_off,
2835 object_size, &src_objnum,
2836 &src_objoff, &src_objlen);
2837 ceph_calc_file_object_mapping(&dst_ci->i_layout, *dst_off,
2838 object_size, &dst_objnum,
2839 &dst_objoff, &dst_objlen);
2840 ceph_oid_init(&src_oid);
2841 ceph_oid_printf(&src_oid, "%llx.%08llx",
2842 src_ci->i_vino.ino, src_objnum);
2843 ceph_oid_init(&dst_oid);
2844 ceph_oid_printf(&dst_oid, "%llx.%08llx",
2845 dst_ci->i_vino.ino, dst_objnum);
2846 /* Do an object remote copy */
2847 req = ceph_alloc_copyfrom_request(osdc, src_ci->i_vino.snap,
2848 &src_oid, &src_oloc,
2849 &dst_oid, &dst_oloc,
2850 dst_ci->i_truncate_seq,
2851 dst_ci->i_truncate_size);
2855 ceph_osdc_start_request(osdc, req);
2856 ret = ceph_osdc_wait_request(osdc, req);
2857 ceph_update_copyfrom_metrics(&fsc->mdsc->metric,
2858 req->r_start_latency,
2861 ceph_osdc_put_request(req);
2864 if (ret == -EOPNOTSUPP) {
2865 fsc->have_copy_from2 = false;
2866 pr_notice_client(cl,
2867 "OSDs don't support copy-from2; disabling copy offload\n");
2869 doutc(cl, "returned %d\n", ret);
2875 bytes += object_size;
2876 *src_off += object_size;
2877 *dst_off += object_size;
2881 ceph_oloc_destroy(&src_oloc);
2882 ceph_oloc_destroy(&dst_oloc);
2886 static ssize_t __ceph_copy_file_range(struct file *src_file, loff_t src_off,
2887 struct file *dst_file, loff_t dst_off,
2888 size_t len, unsigned int flags)
2890 struct inode *src_inode = file_inode(src_file);
2891 struct inode *dst_inode = file_inode(dst_file);
2892 struct ceph_inode_info *src_ci = ceph_inode(src_inode);
2893 struct ceph_inode_info *dst_ci = ceph_inode(dst_inode);
2894 struct ceph_cap_flush *prealloc_cf;
2895 struct ceph_fs_client *src_fsc = ceph_inode_to_fs_client(src_inode);
2896 struct ceph_client *cl = src_fsc->client;
2898 ssize_t ret = -EIO, bytes;
2899 u64 src_objnum, dst_objnum, src_objoff, dst_objoff;
2900 u32 src_objlen, dst_objlen;
2901 int src_got = 0, dst_got = 0, err, dirty;
2903 if (src_inode->i_sb != dst_inode->i_sb) {
2904 struct ceph_fs_client *dst_fsc = ceph_inode_to_fs_client(dst_inode);
2906 if (ceph_fsid_compare(&src_fsc->client->fsid,
2907 &dst_fsc->client->fsid)) {
2908 dout("Copying files across clusters: src: %pU dst: %pU\n",
2909 &src_fsc->client->fsid, &dst_fsc->client->fsid);
2913 if (ceph_snap(dst_inode) != CEPH_NOSNAP)
2917 * Some of the checks below will return -EOPNOTSUPP, which will force a
2918 * fallback to the default VFS copy_file_range implementation. This is
2919 * desirable in several cases (for ex, the 'len' is smaller than the
2920 * size of the objects, or in cases where that would be more
2924 if (ceph_test_mount_opt(src_fsc, NOCOPYFROM))
2927 if (!src_fsc->have_copy_from2)
2931 * Striped file layouts require that we copy partial objects, but the
2932 * OSD copy-from operation only supports full-object copies. Limit
2933 * this to non-striped file layouts for now.
2935 if ((src_ci->i_layout.stripe_unit != dst_ci->i_layout.stripe_unit) ||
2936 (src_ci->i_layout.stripe_count != 1) ||
2937 (dst_ci->i_layout.stripe_count != 1) ||
2938 (src_ci->i_layout.object_size != dst_ci->i_layout.object_size)) {
2939 doutc(cl, "Invalid src/dst files layout\n");
2943 /* Every encrypted inode gets its own key, so we can't offload them */
2944 if (IS_ENCRYPTED(src_inode) || IS_ENCRYPTED(dst_inode))
2947 if (len < src_ci->i_layout.object_size)
2948 return -EOPNOTSUPP; /* no remote copy will be done */
2950 prealloc_cf = ceph_alloc_cap_flush();
2954 /* Start by sync'ing the source and destination files */
2955 ret = file_write_and_wait_range(src_file, src_off, (src_off + len));
2957 doutc(cl, "failed to write src file (%zd)\n", ret);
2960 ret = file_write_and_wait_range(dst_file, dst_off, (dst_off + len));
2962 doutc(cl, "failed to write dst file (%zd)\n", ret);
2967 * We need FILE_WR caps for dst_ci and FILE_RD for src_ci as other
2968 * clients may have dirty data in their caches. And OSDs know nothing
2969 * about caps, so they can't safely do the remote object copies.
2971 err = get_rd_wr_caps(src_file, &src_got,
2972 dst_file, (dst_off + len), &dst_got);
2974 doutc(cl, "get_rd_wr_caps returned %d\n", err);
2979 ret = is_file_size_ok(src_inode, dst_inode, src_off, dst_off, len);
2983 /* Drop dst file cached pages */
2984 ceph_fscache_invalidate(dst_inode, false);
2985 ret = invalidate_inode_pages2_range(dst_inode->i_mapping,
2986 dst_off >> PAGE_SHIFT,
2987 (dst_off + len) >> PAGE_SHIFT);
2989 doutc(cl, "Failed to invalidate inode pages (%zd)\n",
2993 ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
2994 src_ci->i_layout.object_size,
2995 &src_objnum, &src_objoff, &src_objlen);
2996 ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
2997 dst_ci->i_layout.object_size,
2998 &dst_objnum, &dst_objoff, &dst_objlen);
2999 /* object-level offsets need to the same */
3000 if (src_objoff != dst_objoff) {
3006 * Do a manual copy if the object offset isn't object aligned.
3007 * 'src_objlen' contains the bytes left until the end of the object,
3008 * starting at the src_off
3011 doutc(cl, "Initial partial copy of %u bytes\n", src_objlen);
3014 * we need to temporarily drop all caps as we'll be calling
3015 * {read,write}_iter, which will get caps again.
3017 put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
3018 ret = splice_file_range(src_file, &src_off, dst_file, &dst_off,
3020 /* Abort on short copies or on error */
3021 if (ret < (long)src_objlen) {
3022 doutc(cl, "Failed partial copy (%zd)\n", ret);
3026 err = get_rd_wr_caps(src_file, &src_got,
3027 dst_file, (dst_off + len), &dst_got);
3030 err = is_file_size_ok(src_inode, dst_inode,
3031 src_off, dst_off, len);
3036 size = i_size_read(dst_inode);
3037 bytes = ceph_do_objects_copy(src_ci, &src_off, dst_ci, &dst_off,
3038 src_fsc, len, flags);
3044 doutc(cl, "Copied %zu bytes out of %zu\n", bytes, len);
3048 file_update_time(dst_file);
3049 inode_inc_iversion_raw(dst_inode);
3051 if (dst_off > size) {
3052 /* Let the MDS know about dst file size change */
3053 if (ceph_inode_set_size(dst_inode, dst_off) ||
3054 ceph_quota_is_max_bytes_approaching(dst_inode, dst_off))
3055 ceph_check_caps(dst_ci, CHECK_CAPS_AUTHONLY | CHECK_CAPS_FLUSH);
3058 spin_lock(&dst_ci->i_ceph_lock);
3059 dirty = __ceph_mark_dirty_caps(dst_ci, CEPH_CAP_FILE_WR, &prealloc_cf);
3060 spin_unlock(&dst_ci->i_ceph_lock);
3062 __mark_inode_dirty(dst_inode, dirty);
3065 put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
3068 * Do the final manual copy if we still have some bytes left, unless
3069 * there were errors in remote object copies (len >= object_size).
3071 if (len && (len < src_ci->i_layout.object_size)) {
3072 doutc(cl, "Final partial copy of %zu bytes\n", len);
3073 bytes = splice_file_range(src_file, &src_off, dst_file,
3078 doutc(cl, "Failed partial copy (%zd)\n", bytes);
3082 ceph_free_cap_flush(prealloc_cf);
3087 static ssize_t ceph_copy_file_range(struct file *src_file, loff_t src_off,
3088 struct file *dst_file, loff_t dst_off,
3089 size_t len, unsigned int flags)
3093 ret = __ceph_copy_file_range(src_file, src_off, dst_file, dst_off,
3096 if (ret == -EOPNOTSUPP || ret == -EXDEV)
3097 ret = splice_copy_file_range(src_file, src_off, dst_file,
3102 const struct file_operations ceph_file_fops = {
3104 .release = ceph_release,
3105 .llseek = ceph_llseek,
3106 .read_iter = ceph_read_iter,
3107 .write_iter = ceph_write_iter,
3109 .fsync = ceph_fsync,
3111 .setlease = simple_nosetlease,
3112 .flock = ceph_flock,
3113 .splice_read = ceph_splice_read,
3114 .splice_write = iter_file_splice_write,
3115 .unlocked_ioctl = ceph_ioctl,
3116 .compat_ioctl = compat_ptr_ioctl,
3117 .fallocate = ceph_fallocate,
3118 .copy_file_range = ceph_copy_file_range,