1 #include <linux/ceph/ceph_debug.h>
3 #include <linux/module.h>
4 #include <linux/sched.h>
5 #include <linux/slab.h>
6 #include <linux/file.h>
7 #include <linux/mount.h>
8 #include <linux/namei.h>
9 #include <linux/writeback.h>
10 #include <linux/aio.h>
11 #include <linux/falloc.h>
14 #include "mds_client.h"
18 * Ceph file operations
20 * Implement basic open/close functionality, and implement
23 * We implement three modes of file I/O:
24 * - buffered uses the generic_file_aio_{read,write} helpers
26 * - synchronous is used when there is multi-client read/write
27 * sharing, avoids the page cache, and synchronously waits for an
30 * - direct io takes the variant of the sync path that references
31 * user pages directly.
33 * fsync() flushes and waits on dirty pages, but just queues metadata
34 * for writeback: since the MDS can recover size and mtime there is no
35 * need to wait for MDS acknowledgement.
40 * Prepare an open request. Preallocate ceph_cap to avoid an
41 * inopportune ENOMEM later.
43 static struct ceph_mds_request *
44 prepare_open_request(struct super_block *sb, int flags, int create_mode)
46 struct ceph_fs_client *fsc = ceph_sb_to_client(sb);
47 struct ceph_mds_client *mdsc = fsc->mdsc;
48 struct ceph_mds_request *req;
49 int want_auth = USE_ANY_MDS;
50 int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
52 if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
53 want_auth = USE_AUTH_MDS;
55 req = ceph_mdsc_create_request(mdsc, op, want_auth);
58 req->r_fmode = ceph_flags_to_mode(flags);
59 req->r_args.open.flags = cpu_to_le32(flags);
60 req->r_args.open.mode = cpu_to_le32(create_mode);
66 * initialize private struct file data.
67 * if we fail, clean up by dropping fmode reference on the ceph_inode
69 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
71 struct ceph_file_info *cf;
73 struct ceph_inode_info *ci = ceph_inode(inode);
74 struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
75 struct ceph_mds_client *mdsc = fsc->mdsc;
77 switch (inode->i_mode & S_IFMT) {
79 /* First file open request creates the cookie, we want to keep
80 * this cookie around for the filetime of the inode as not to
81 * have to worry about fscache register / revoke / operation
84 * Also, if we know the operation is going to invalidate data
85 * (non readonly) just nuke the cache right away.
87 ceph_fscache_register_inode_cookie(mdsc->fsc, ci);
88 if ((fmode & CEPH_FILE_MODE_WR))
89 ceph_fscache_invalidate(inode);
91 dout("init_file %p %p 0%o (regular)\n", inode, file,
93 cf = kmem_cache_alloc(ceph_file_cachep, GFP_NOFS | __GFP_ZERO);
95 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
100 file->private_data = cf;
101 BUG_ON(inode->i_fop->release != ceph_release);
105 dout("init_file %p %p 0%o (symlink)\n", inode, file,
107 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
111 dout("init_file %p %p 0%o (special)\n", inode, file,
114 * we need to drop the open ref now, since we don't
115 * have .release set to ceph_release.
117 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
118 BUG_ON(inode->i_fop->release == ceph_release);
120 /* call the proper open fop */
121 ret = inode->i_fop->open(inode, file);
127 * If we already have the requisite capabilities, we can satisfy
128 * the open request locally (no need to request new caps from the
129 * MDS). We do, however, need to inform the MDS (asynchronously)
130 * if our wanted caps set expands.
132 int ceph_open(struct inode *inode, struct file *file)
134 struct ceph_inode_info *ci = ceph_inode(inode);
135 struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
136 struct ceph_mds_client *mdsc = fsc->mdsc;
137 struct ceph_mds_request *req;
138 struct ceph_file_info *cf = file->private_data;
139 struct inode *parent_inode = NULL;
141 int flags, fmode, wanted;
144 dout("open file %p is already opened\n", file);
148 /* filter out O_CREAT|O_EXCL; vfs did that already. yuck. */
149 flags = file->f_flags & ~(O_CREAT|O_EXCL);
150 if (S_ISDIR(inode->i_mode))
151 flags = O_DIRECTORY; /* mds likes to know */
153 dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
154 ceph_vinop(inode), file, flags, file->f_flags);
155 fmode = ceph_flags_to_mode(flags);
156 wanted = ceph_caps_for_mode(fmode);
158 /* snapped files are read-only */
159 if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
162 /* trivially open snapdir */
163 if (ceph_snap(inode) == CEPH_SNAPDIR) {
164 spin_lock(&ci->i_ceph_lock);
165 __ceph_get_fmode(ci, fmode);
166 spin_unlock(&ci->i_ceph_lock);
167 return ceph_init_file(inode, file, fmode);
171 * No need to block if we have caps on the auth MDS (for
172 * write) or any MDS (for read). Update wanted set
175 spin_lock(&ci->i_ceph_lock);
176 if (__ceph_is_any_real_caps(ci) &&
177 (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
178 int mds_wanted = __ceph_caps_mds_wanted(ci);
179 int issued = __ceph_caps_issued(ci, NULL);
181 dout("open %p fmode %d want %s issued %s using existing\n",
182 inode, fmode, ceph_cap_string(wanted),
183 ceph_cap_string(issued));
184 __ceph_get_fmode(ci, fmode);
185 spin_unlock(&ci->i_ceph_lock);
188 if ((issued & wanted) != wanted &&
189 (mds_wanted & wanted) != wanted &&
190 ceph_snap(inode) != CEPH_SNAPDIR)
191 ceph_check_caps(ci, 0, NULL);
193 return ceph_init_file(inode, file, fmode);
194 } else if (ceph_snap(inode) != CEPH_NOSNAP &&
195 (ci->i_snap_caps & wanted) == wanted) {
196 __ceph_get_fmode(ci, fmode);
197 spin_unlock(&ci->i_ceph_lock);
198 return ceph_init_file(inode, file, fmode);
201 spin_unlock(&ci->i_ceph_lock);
203 dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
204 req = prepare_open_request(inode->i_sb, flags, 0);
209 req->r_inode = inode;
214 parent_inode = ceph_get_dentry_parent_inode(file->f_dentry);
215 err = ceph_mdsc_do_request(mdsc, parent_inode, req);
218 err = ceph_init_file(inode, file, req->r_fmode);
219 ceph_mdsc_put_request(req);
220 dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
227 * Do a lookup + open with a single request. If we get a non-existent
228 * file or symlink, return 1 so the VFS can retry.
230 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
231 struct file *file, unsigned flags, umode_t mode,
234 struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
235 struct ceph_mds_client *mdsc = fsc->mdsc;
236 struct ceph_mds_request *req;
238 struct ceph_acls_info acls = {};
241 dout("atomic_open %p dentry %p '%.*s' %s flags %d mode 0%o\n",
242 dir, dentry, dentry->d_name.len, dentry->d_name.name,
243 d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
245 if (dentry->d_name.len > NAME_MAX)
246 return -ENAMETOOLONG;
248 err = ceph_init_dentry(dentry);
252 if (flags & O_CREAT) {
253 err = ceph_pre_init_acls(dir, &mode, &acls);
259 req = prepare_open_request(dir->i_sb, flags, mode);
264 req->r_dentry = dget(dentry);
266 if (flags & O_CREAT) {
267 req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
268 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
270 req->r_pagelist = acls.pagelist;
271 acls.pagelist = NULL;
274 req->r_locked_dir = dir; /* caller holds dir->i_mutex */
275 err = ceph_mdsc_do_request(mdsc,
276 (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
281 err = ceph_handle_snapdir(req, dentry, err);
282 if (err == 0 && (flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
283 err = ceph_handle_notrace_create(dir, dentry);
285 if (d_unhashed(dentry)) {
286 dn = ceph_finish_lookup(req, dentry, err);
290 /* we were given a hashed negative dentry */
295 if (dn || dentry->d_inode == NULL || S_ISLNK(dentry->d_inode->i_mode)) {
296 /* make vfs retry on splice, ENOENT, or symlink */
297 dout("atomic_open finish_no_open on dn %p\n", dn);
298 err = finish_no_open(file, dn);
300 dout("atomic_open finish_open on dn %p\n", dn);
301 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
302 ceph_init_inode_acls(dentry->d_inode, &acls);
303 *opened |= FILE_CREATED;
305 err = finish_open(file, dentry, ceph_open, opened);
308 if (!req->r_err && req->r_target_inode)
309 ceph_put_fmode(ceph_inode(req->r_target_inode), req->r_fmode);
310 ceph_mdsc_put_request(req);
312 ceph_release_acls_info(&acls);
313 dout("atomic_open result=%d\n", err);
317 int ceph_release(struct inode *inode, struct file *file)
319 struct ceph_inode_info *ci = ceph_inode(inode);
320 struct ceph_file_info *cf = file->private_data;
322 dout("release inode %p file %p\n", inode, file);
323 ceph_put_fmode(ci, cf->fmode);
324 if (cf->last_readdir)
325 ceph_mdsc_put_request(cf->last_readdir);
326 kfree(cf->last_name);
329 kmem_cache_free(ceph_file_cachep, cf);
331 /* wake up anyone waiting for caps on this inode */
332 wake_up_all(&ci->i_cap_wq);
337 * Read a range of bytes striped over one or more objects. Iterate over
338 * objects we stripe over. (That's not atomic, but good enough for now.)
340 * If we get a short result from the OSD, check against i_size; we need to
341 * only return a short read to the caller if we hit EOF.
343 static int striped_read(struct inode *inode,
345 struct page **pages, int num_pages,
346 int *checkeof, bool o_direct,
347 unsigned long buf_align)
349 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
350 struct ceph_inode_info *ci = ceph_inode(inode);
351 u64 pos, this_len, left;
352 int io_align, page_align;
355 struct page **page_pos;
357 bool hit_stripe, was_short;
360 * we may need to do multiple reads. not atomic, unfortunately.
365 pages_left = num_pages;
367 io_align = off & ~PAGE_MASK;
371 page_align = (pos - io_align + buf_align) & ~PAGE_MASK;
373 page_align = pos & ~PAGE_MASK;
375 ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode),
376 &ci->i_layout, pos, &this_len,
379 page_pos, pages_left, page_align);
382 hit_stripe = this_len < left;
383 was_short = ret >= 0 && ret < this_len;
384 dout("striped_read %llu~%llu (read %u) got %d%s%s\n", pos, left, read,
385 ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : "");
389 if (was_short && (pos + ret < inode->i_size)) {
390 u64 tmp = min(this_len - ret,
391 inode->i_size - pos - ret);
392 dout(" zero gap %llu to %llu\n",
393 pos + ret, pos + ret + tmp);
394 ceph_zero_page_vector_range(page_align + read + ret,
399 didpages = (page_align + ret) >> PAGE_CACHE_SHIFT;
403 page_pos += didpages;
404 pages_left -= didpages;
406 /* hit stripe and need continue*/
407 if (left && hit_stripe && pos < inode->i_size)
413 /* did we bounce off eof? */
414 if (pos + left > inode->i_size)
418 dout("striped_read returns %d\n", ret);
423 * Completely synchronous read and write methods. Direct from __user
424 * buffer to osd, or directly to user pages (if O_DIRECT).
426 * If the read spans object boundary, just do multiple reads.
428 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *i,
431 struct file *file = iocb->ki_filp;
432 struct inode *inode = file_inode(file);
434 u64 off = iocb->ki_pos;
436 size_t len = iov_iter_count(i);
438 dout("sync_read on file %p %llu~%u %s\n", file, off,
440 (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
445 * flush any page cache pages in this range. this
446 * will make concurrent normal and sync io slow,
447 * but it will at least behave sensibly when they are
450 ret = filemap_write_and_wait_range(inode->i_mapping, off,
455 if (file->f_flags & O_DIRECT) {
456 while (iov_iter_count(i)) {
460 n = iov_iter_get_pages_alloc(i, &pages, INT_MAX, &start);
464 num_pages = (n + start + PAGE_SIZE - 1) / PAGE_SIZE;
466 ret = striped_read(inode, off, n,
467 pages, num_pages, checkeof,
470 ceph_put_page_vector(pages, num_pages, true);
475 iov_iter_advance(i, ret);
480 num_pages = calc_pages_for(off, len);
481 pages = ceph_alloc_page_vector(num_pages, GFP_NOFS);
483 return PTR_ERR(pages);
484 ret = striped_read(inode, off, len, pages,
485 num_pages, checkeof, 0, 0);
491 size_t page_off = off & ~PAGE_MASK;
492 size_t copy = min_t(size_t,
493 PAGE_SIZE - page_off, left);
494 l = copy_page_to_iter(pages[k++], page_off,
502 ceph_release_page_vector(pages, num_pages);
505 if (off > iocb->ki_pos) {
506 ret = off - iocb->ki_pos;
510 dout("sync_read result %d\n", ret);
515 * Write commit request unsafe callback, called to tell us when a
516 * request is unsafe (that is, in flight--has been handed to the
517 * messenger to send to its target osd). It is called again when
518 * we've received a response message indicating the request is
519 * "safe" (its CEPH_OSD_FLAG_ONDISK flag is set), or when a request
520 * is completed early (and unsuccessfully) due to a timeout or
523 * This is used if we requested both an ACK and ONDISK commit reply
526 static void ceph_sync_write_unsafe(struct ceph_osd_request *req, bool unsafe)
528 struct ceph_inode_info *ci = ceph_inode(req->r_inode);
530 dout("%s %p tid %llu %ssafe\n", __func__, req, req->r_tid,
533 ceph_get_cap_refs(ci, CEPH_CAP_FILE_WR);
534 spin_lock(&ci->i_unsafe_lock);
535 list_add_tail(&req->r_unsafe_item,
536 &ci->i_unsafe_writes);
537 spin_unlock(&ci->i_unsafe_lock);
539 spin_lock(&ci->i_unsafe_lock);
540 list_del_init(&req->r_unsafe_item);
541 spin_unlock(&ci->i_unsafe_lock);
542 ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR);
548 * Synchronous write, straight from __user pointer or user pages.
550 * If write spans object boundary, just do multiple writes. (For a
551 * correct atomic write, we should e.g. take write locks on all
552 * objects, rollback on failure, etc.)
555 ceph_sync_direct_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos)
557 struct file *file = iocb->ki_filp;
558 struct inode *inode = file_inode(file);
559 struct ceph_inode_info *ci = ceph_inode(inode);
560 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
561 struct ceph_snap_context *snapc;
562 struct ceph_vino vino;
563 struct ceph_osd_request *req;
570 struct timespec mtime = CURRENT_TIME;
571 size_t count = iov_iter_count(from);
573 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
576 dout("sync_direct_write on file %p %lld~%u\n", file, pos,
579 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
583 ret = invalidate_inode_pages2_range(inode->i_mapping,
584 pos >> PAGE_CACHE_SHIFT,
585 (pos + count) >> PAGE_CACHE_SHIFT);
587 dout("invalidate_inode_pages2_range returned %d\n", ret);
589 flags = CEPH_OSD_FLAG_ORDERSNAP |
590 CEPH_OSD_FLAG_ONDISK |
593 while (iov_iter_count(from) > 0) {
594 u64 len = iov_iter_single_seg_count(from);
598 snapc = ci->i_snap_realm->cached_context;
599 vino = ceph_vino(inode);
600 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
602 2,/*include a 'startsync' command*/
603 CEPH_OSD_OP_WRITE, flags, snapc,
612 n = iov_iter_get_pages_alloc(from, &pages, len, &start);
613 if (unlikely(n < 0)) {
615 ceph_osdc_put_request(req);
619 num_pages = (n + start + PAGE_SIZE - 1) / PAGE_SIZE;
621 * throw out any page cache pages in this range. this
624 truncate_inode_pages_range(inode->i_mapping, pos,
625 (pos+n) | (PAGE_CACHE_SIZE-1));
626 osd_req_op_extent_osd_data_pages(req, 0, pages, n, start,
629 /* BUG_ON(vino.snap != CEPH_NOSNAP); */
630 ceph_osdc_build_request(req, pos, snapc, vino.snap, &mtime);
632 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
634 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
636 ceph_put_page_vector(pages, num_pages, false);
638 ceph_osdc_put_request(req);
643 iov_iter_advance(from, n);
645 if (pos > i_size_read(inode)) {
646 check_caps = ceph_inode_set_size(inode, pos);
648 ceph_check_caps(ceph_inode(inode),
654 if (ret != -EOLDSNAPC && written > 0) {
663 * Synchronous write, straight from __user pointer or user pages.
665 * If write spans object boundary, just do multiple writes. (For a
666 * correct atomic write, we should e.g. take write locks on all
667 * objects, rollback on failure, etc.)
670 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos)
672 struct file *file = iocb->ki_filp;
673 struct inode *inode = file_inode(file);
674 struct ceph_inode_info *ci = ceph_inode(inode);
675 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
676 struct ceph_snap_context *snapc;
677 struct ceph_vino vino;
678 struct ceph_osd_request *req;
686 struct timespec mtime = CURRENT_TIME;
687 size_t count = iov_iter_count(from);
689 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
692 dout("sync_write on file %p %lld~%u\n", file, pos, (unsigned)count);
694 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
698 ret = invalidate_inode_pages2_range(inode->i_mapping,
699 pos >> PAGE_CACHE_SHIFT,
700 (pos + count) >> PAGE_CACHE_SHIFT);
702 dout("invalidate_inode_pages2_range returned %d\n", ret);
704 flags = CEPH_OSD_FLAG_ORDERSNAP |
705 CEPH_OSD_FLAG_ONDISK |
706 CEPH_OSD_FLAG_WRITE |
709 while ((len = iov_iter_count(from)) > 0) {
713 snapc = ci->i_snap_realm->cached_context;
714 vino = ceph_vino(inode);
715 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
717 CEPH_OSD_OP_WRITE, flags, snapc,
727 * write from beginning of first page,
728 * regardless of io alignment
730 num_pages = (len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
732 pages = ceph_alloc_page_vector(num_pages, GFP_NOFS);
734 ret = PTR_ERR(pages);
739 for (n = 0; n < num_pages; n++) {
740 size_t plen = min_t(size_t, left, PAGE_SIZE);
741 ret = copy_page_from_iter(pages[n], 0, plen, from);
750 ceph_release_page_vector(pages, num_pages);
754 /* get a second commit callback */
755 req->r_unsafe_callback = ceph_sync_write_unsafe;
756 req->r_inode = inode;
758 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
761 /* BUG_ON(vino.snap != CEPH_NOSNAP); */
762 ceph_osdc_build_request(req, pos, snapc, vino.snap, &mtime);
764 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
766 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
769 ceph_osdc_put_request(req);
774 if (pos > i_size_read(inode)) {
775 check_caps = ceph_inode_set_size(inode, pos);
777 ceph_check_caps(ceph_inode(inode),
785 if (ret != -EOLDSNAPC && written > 0) {
793 * Wrap generic_file_aio_read with checks for cap bits on the inode.
794 * Atomically grab references, so that those bits are not released
795 * back to the MDS mid-read.
797 * Hmm, the sync read case isn't actually async... should it be?
799 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
801 struct file *filp = iocb->ki_filp;
802 struct ceph_file_info *fi = filp->private_data;
803 size_t len = iocb->ki_nbytes;
804 struct inode *inode = file_inode(filp);
805 struct ceph_inode_info *ci = ceph_inode(inode);
808 int checkeof = 0, read = 0;
811 dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
812 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
814 if (fi->fmode & CEPH_FILE_MODE_LAZY)
815 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
817 want = CEPH_CAP_FILE_CACHE;
818 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, &got, -1);
822 if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
823 (iocb->ki_filp->f_flags & O_DIRECT) ||
824 (fi->flags & CEPH_F_SYNC)) {
826 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
827 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
828 ceph_cap_string(got));
830 /* hmm, this isn't really async... */
831 ret = ceph_sync_read(iocb, to, &checkeof);
833 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
834 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
835 ceph_cap_string(got));
837 ret = generic_file_read_iter(iocb, to);
839 dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
840 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
841 ceph_put_cap_refs(ci, got);
843 if (checkeof && ret >= 0) {
844 int statret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
846 /* hit EOF or hole? */
847 if (statret == 0 && iocb->ki_pos < inode->i_size &&
849 dout("sync_read hit hole, ppos %lld < size %lld"
850 ", reading more\n", iocb->ki_pos,
867 * Take cap references to avoid releasing caps to MDS mid-write.
869 * If we are synchronous, and write with an old snap context, the OSD
870 * may return EOLDSNAPC. In that case, retry the write.. _after_
871 * dropping our cap refs and allowing the pending snap to logically
872 * complete _before_ this write occurs.
874 * If we are near ENOSPC, write synchronously.
876 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
878 struct file *file = iocb->ki_filp;
879 struct ceph_file_info *fi = file->private_data;
880 struct inode *inode = file_inode(file);
881 struct ceph_inode_info *ci = ceph_inode(inode);
882 struct ceph_osd_client *osdc =
883 &ceph_sb_to_client(inode->i_sb)->client->osdc;
884 ssize_t count = iov_iter_count(from), written = 0;
886 loff_t pos = iocb->ki_pos;
888 if (ceph_snap(inode) != CEPH_NOSNAP)
891 mutex_lock(&inode->i_mutex);
893 /* We can write back this queue in page reclaim */
894 current->backing_dev_info = file->f_mapping->backing_dev_info;
896 err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
902 iov_iter_truncate(from, count);
904 err = file_remove_suid(file);
908 err = file_update_time(file);
913 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) {
918 dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
919 inode, ceph_vinop(inode), pos, count, inode->i_size);
920 if (fi->fmode & CEPH_FILE_MODE_LAZY)
921 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
923 want = CEPH_CAP_FILE_BUFFER;
925 err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, &got, pos + count);
929 dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
930 inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
932 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
933 (file->f_flags & O_DIRECT) || (fi->flags & CEPH_F_SYNC)) {
934 struct iov_iter data;
935 mutex_unlock(&inode->i_mutex);
936 /* we might need to revert back to that point */
938 if (file->f_flags & O_DIRECT)
939 written = ceph_sync_direct_write(iocb, &data, pos);
941 written = ceph_sync_write(iocb, &data, pos);
942 if (written == -EOLDSNAPC) {
943 dout("aio_write %p %llx.%llx %llu~%u"
944 "got EOLDSNAPC, retrying\n",
945 inode, ceph_vinop(inode),
946 pos, (unsigned)count);
947 mutex_lock(&inode->i_mutex);
951 iov_iter_advance(from, written);
953 loff_t old_size = inode->i_size;
955 * No need to acquire the i_truncate_mutex. Because
956 * the MDS revokes Fwb caps before sending truncate
957 * message to us. We can't get Fwb cap while there
958 * are pending vmtruncate. So write and vmtruncate
959 * can not run at the same time
961 written = generic_perform_write(file, from, pos);
962 if (likely(written >= 0))
963 iocb->ki_pos = pos + written;
964 if (inode->i_size > old_size)
965 ceph_fscache_update_objectsize(inode);
966 mutex_unlock(&inode->i_mutex);
971 spin_lock(&ci->i_ceph_lock);
972 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR);
973 spin_unlock(&ci->i_ceph_lock);
975 __mark_inode_dirty(inode, dirty);
978 dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n",
979 inode, ceph_vinop(inode), pos, (unsigned)count,
980 ceph_cap_string(got));
981 ceph_put_cap_refs(ci, got);
984 ((file->f_flags & O_SYNC) || IS_SYNC(file->f_mapping->host) ||
985 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_NEARFULL))) {
986 err = vfs_fsync_range(file, pos, pos + written - 1, 1);
994 mutex_unlock(&inode->i_mutex);
996 current->backing_dev_info = NULL;
997 return written ? written : err;
1001 * llseek. be sure to verify file size on SEEK_END.
1003 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1005 struct inode *inode = file->f_mapping->host;
1008 mutex_lock(&inode->i_mutex);
1010 if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1011 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1020 offset += inode->i_size;
1024 * Here we special-case the lseek(fd, 0, SEEK_CUR)
1025 * position-querying operation. Avoid rewriting the "same"
1026 * f_pos value back to the file because a concurrent read(),
1027 * write() or lseek() might have altered it
1030 offset = file->f_pos;
1033 offset += file->f_pos;
1036 if (offset >= inode->i_size) {
1042 if (offset >= inode->i_size) {
1046 offset = inode->i_size;
1050 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1053 mutex_unlock(&inode->i_mutex);
1057 static inline void ceph_zero_partial_page(
1058 struct inode *inode, loff_t offset, unsigned size)
1061 pgoff_t index = offset >> PAGE_CACHE_SHIFT;
1063 page = find_lock_page(inode->i_mapping, index);
1065 wait_on_page_writeback(page);
1066 zero_user(page, offset & (PAGE_CACHE_SIZE - 1), size);
1068 page_cache_release(page);
1072 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1075 loff_t nearly = round_up(offset, PAGE_CACHE_SIZE);
1076 if (offset < nearly) {
1077 loff_t size = nearly - offset;
1080 ceph_zero_partial_page(inode, offset, size);
1084 if (length >= PAGE_CACHE_SIZE) {
1085 loff_t size = round_down(length, PAGE_CACHE_SIZE);
1086 truncate_pagecache_range(inode, offset, offset + size - 1);
1091 ceph_zero_partial_page(inode, offset, length);
1094 static int ceph_zero_partial_object(struct inode *inode,
1095 loff_t offset, loff_t *length)
1097 struct ceph_inode_info *ci = ceph_inode(inode);
1098 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1099 struct ceph_osd_request *req;
1105 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1108 op = CEPH_OSD_OP_ZERO;
1111 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1115 CEPH_OSD_FLAG_WRITE |
1116 CEPH_OSD_FLAG_ONDISK,
1123 ceph_osdc_build_request(req, offset, NULL, ceph_vino(inode).snap,
1126 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1128 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1132 ceph_osdc_put_request(req);
1138 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
1141 struct ceph_inode_info *ci = ceph_inode(inode);
1142 s32 stripe_unit = ceph_file_layout_su(ci->i_layout);
1143 s32 stripe_count = ceph_file_layout_stripe_count(ci->i_layout);
1144 s32 object_size = ceph_file_layout_object_size(ci->i_layout);
1145 u64 object_set_size = object_size * stripe_count;
1148 /* round offset up to next period boundary */
1149 nearly = offset + object_set_size - 1;
1151 nearly -= do_div(t, object_set_size);
1153 while (length && offset < nearly) {
1154 loff_t size = length;
1155 ret = ceph_zero_partial_object(inode, offset, &size);
1161 while (length >= object_set_size) {
1163 loff_t pos = offset;
1164 for (i = 0; i < stripe_count; ++i) {
1165 ret = ceph_zero_partial_object(inode, pos, NULL);
1170 offset += object_set_size;
1171 length -= object_set_size;
1174 loff_t size = length;
1175 ret = ceph_zero_partial_object(inode, offset, &size);
1184 static long ceph_fallocate(struct file *file, int mode,
1185 loff_t offset, loff_t length)
1187 struct ceph_file_info *fi = file->private_data;
1188 struct inode *inode = file_inode(file);
1189 struct ceph_inode_info *ci = ceph_inode(inode);
1190 struct ceph_osd_client *osdc =
1191 &ceph_inode_to_client(inode)->client->osdc;
1198 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1201 if (!S_ISREG(inode->i_mode))
1204 mutex_lock(&inode->i_mutex);
1206 if (ceph_snap(inode) != CEPH_NOSNAP) {
1211 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) &&
1212 !(mode & FALLOC_FL_PUNCH_HOLE)) {
1217 size = i_size_read(inode);
1218 if (!(mode & FALLOC_FL_KEEP_SIZE))
1219 endoff = offset + length;
1221 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1222 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1224 want = CEPH_CAP_FILE_BUFFER;
1226 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, &got, endoff);
1230 if (mode & FALLOC_FL_PUNCH_HOLE) {
1232 ceph_zero_pagecache_range(inode, offset, length);
1233 ret = ceph_zero_objects(inode, offset, length);
1234 } else if (endoff > size) {
1235 truncate_pagecache_range(inode, size, -1);
1236 if (ceph_inode_set_size(inode, endoff))
1237 ceph_check_caps(ceph_inode(inode),
1238 CHECK_CAPS_AUTHONLY, NULL);
1242 spin_lock(&ci->i_ceph_lock);
1243 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR);
1244 spin_unlock(&ci->i_ceph_lock);
1246 __mark_inode_dirty(inode, dirty);
1249 ceph_put_cap_refs(ci, got);
1251 mutex_unlock(&inode->i_mutex);
1255 const struct file_operations ceph_file_fops = {
1257 .release = ceph_release,
1258 .llseek = ceph_llseek,
1259 .read = new_sync_read,
1260 .write = new_sync_write,
1261 .read_iter = ceph_read_iter,
1262 .write_iter = ceph_write_iter,
1264 .fsync = ceph_fsync,
1266 .flock = ceph_flock,
1267 .splice_read = generic_file_splice_read,
1268 .splice_write = iter_file_splice_write,
1269 .unlocked_ioctl = ceph_ioctl,
1270 .compat_ioctl = ceph_ioctl,
1271 .fallocate = ceph_fallocate,