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;
213 if (flags & (O_CREAT|O_TRUNC))
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;
240 dout("atomic_open %p dentry %p '%.*s' %s flags %d mode 0%o\n",
241 dir, dentry, dentry->d_name.len, dentry->d_name.name,
242 d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
244 if (dentry->d_name.len > NAME_MAX)
245 return -ENAMETOOLONG;
247 err = ceph_init_dentry(dentry);
252 req = prepare_open_request(dir->i_sb, flags, mode);
255 req->r_dentry = dget(dentry);
257 if (flags & O_CREAT) {
258 req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
259 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
261 req->r_locked_dir = dir; /* caller holds dir->i_mutex */
262 err = ceph_mdsc_do_request(mdsc,
263 (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
268 err = ceph_handle_snapdir(req, dentry, err);
269 if (err == 0 && (flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
270 err = ceph_handle_notrace_create(dir, dentry);
272 if (d_unhashed(dentry)) {
273 dn = ceph_finish_lookup(req, dentry, err);
277 /* we were given a hashed negative dentry */
282 if (dn || dentry->d_inode == NULL || S_ISLNK(dentry->d_inode->i_mode)) {
283 /* make vfs retry on splice, ENOENT, or symlink */
284 dout("atomic_open finish_no_open on dn %p\n", dn);
285 err = finish_no_open(file, dn);
287 dout("atomic_open finish_open on dn %p\n", dn);
288 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
289 *opened |= FILE_CREATED;
291 err = finish_open(file, dentry, ceph_open, opened);
295 ceph_mdsc_put_request(req);
296 dout("atomic_open result=%d\n", err);
300 int ceph_release(struct inode *inode, struct file *file)
302 struct ceph_inode_info *ci = ceph_inode(inode);
303 struct ceph_file_info *cf = file->private_data;
305 dout("release inode %p file %p\n", inode, file);
306 ceph_put_fmode(ci, cf->fmode);
307 if (cf->last_readdir)
308 ceph_mdsc_put_request(cf->last_readdir);
309 kfree(cf->last_name);
312 kmem_cache_free(ceph_file_cachep, cf);
314 /* wake up anyone waiting for caps on this inode */
315 wake_up_all(&ci->i_cap_wq);
320 * Read a range of bytes striped over one or more objects. Iterate over
321 * objects we stripe over. (That's not atomic, but good enough for now.)
323 * If we get a short result from the OSD, check against i_size; we need to
324 * only return a short read to the caller if we hit EOF.
326 static int striped_read(struct inode *inode,
328 struct page **pages, int num_pages,
329 int *checkeof, bool o_direct,
330 unsigned long buf_align)
332 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
333 struct ceph_inode_info *ci = ceph_inode(inode);
334 u64 pos, this_len, left;
335 int io_align, page_align;
338 struct page **page_pos;
340 bool hit_stripe, was_short;
343 * we may need to do multiple reads. not atomic, unfortunately.
348 pages_left = num_pages;
350 io_align = off & ~PAGE_MASK;
354 page_align = (pos - io_align + buf_align) & ~PAGE_MASK;
356 page_align = pos & ~PAGE_MASK;
358 ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode),
359 &ci->i_layout, pos, &this_len,
362 page_pos, pages_left, page_align);
365 hit_stripe = this_len < left;
366 was_short = ret >= 0 && ret < this_len;
367 dout("striped_read %llu~%llu (read %u) got %d%s%s\n", pos, left, read,
368 ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : "");
372 if (was_short && (pos + ret < inode->i_size)) {
373 u64 tmp = min(this_len - ret,
374 inode->i_size - pos - ret);
375 dout(" zero gap %llu to %llu\n",
376 pos + ret, pos + ret + tmp);
377 ceph_zero_page_vector_range(page_align + read + ret,
382 didpages = (page_align + ret) >> PAGE_CACHE_SHIFT;
386 page_pos += didpages;
387 pages_left -= didpages;
389 /* hit stripe and need continue*/
390 if (left && hit_stripe && pos < inode->i_size)
396 /* did we bounce off eof? */
397 if (pos + left > inode->i_size)
401 dout("striped_read returns %d\n", ret);
406 * Completely synchronous read and write methods. Direct from __user
407 * buffer to osd, or directly to user pages (if O_DIRECT).
409 * If the read spans object boundary, just do multiple reads.
411 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *i,
414 struct file *file = iocb->ki_filp;
415 struct inode *inode = file_inode(file);
417 u64 off = iocb->ki_pos;
419 size_t len = i->count;
421 dout("sync_read on file %p %llu~%u %s\n", file, off,
423 (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
425 * flush any page cache pages in this range. this
426 * will make concurrent normal and sync io slow,
427 * but it will at least behave sensibly when they are
430 ret = filemap_write_and_wait_range(inode->i_mapping, off,
435 if (file->f_flags & O_DIRECT) {
436 while (iov_iter_count(i)) {
437 void __user *data = i->iov[0].iov_base + i->iov_offset;
438 size_t len = i->iov[0].iov_len - i->iov_offset;
440 num_pages = calc_pages_for((unsigned long)data, len);
441 pages = ceph_get_direct_page_vector(data,
444 return PTR_ERR(pages);
446 ret = striped_read(inode, off, len,
447 pages, num_pages, checkeof,
448 1, (unsigned long)data & ~PAGE_MASK);
449 ceph_put_page_vector(pages, num_pages, true);
454 iov_iter_advance(i, ret);
459 num_pages = calc_pages_for(off, len);
460 pages = ceph_alloc_page_vector(num_pages, GFP_NOFS);
462 return PTR_ERR(pages);
463 ret = striped_read(inode, off, len, pages,
464 num_pages, checkeof, 0, 0);
467 size_t left = len = ret;
470 void __user *data = i->iov[0].iov_base
472 l = min(i->iov[0].iov_len - i->iov_offset,
475 ret = ceph_copy_page_vector_to_user(&pages[k],
479 iov_iter_advance(i, ret);
482 k = calc_pages_for(iocb->ki_pos,
484 BUG_ON(k >= num_pages && left);
489 ceph_release_page_vector(pages, num_pages);
492 if (off > iocb->ki_pos) {
493 ret = off - iocb->ki_pos;
497 dout("sync_read result %d\n", ret);
502 * Write commit request unsafe callback, called to tell us when a
503 * request is unsafe (that is, in flight--has been handed to the
504 * messenger to send to its target osd). It is called again when
505 * we've received a response message indicating the request is
506 * "safe" (its CEPH_OSD_FLAG_ONDISK flag is set), or when a request
507 * is completed early (and unsuccessfully) due to a timeout or
510 * This is used if we requested both an ACK and ONDISK commit reply
513 static void ceph_sync_write_unsafe(struct ceph_osd_request *req, bool unsafe)
515 struct ceph_inode_info *ci = ceph_inode(req->r_inode);
517 dout("%s %p tid %llu %ssafe\n", __func__, req, req->r_tid,
520 ceph_get_cap_refs(ci, CEPH_CAP_FILE_WR);
521 spin_lock(&ci->i_unsafe_lock);
522 list_add_tail(&req->r_unsafe_item,
523 &ci->i_unsafe_writes);
524 spin_unlock(&ci->i_unsafe_lock);
526 spin_lock(&ci->i_unsafe_lock);
527 list_del_init(&req->r_unsafe_item);
528 spin_unlock(&ci->i_unsafe_lock);
529 ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR);
535 * Synchronous write, straight from __user pointer or user pages.
537 * If write spans object boundary, just do multiple writes. (For a
538 * correct atomic write, we should e.g. take write locks on all
539 * objects, rollback on failure, etc.)
542 ceph_sync_direct_write(struct kiocb *iocb, const struct iovec *iov,
543 unsigned long nr_segs, size_t count)
545 struct file *file = iocb->ki_filp;
546 struct inode *inode = file_inode(file);
547 struct ceph_inode_info *ci = ceph_inode(inode);
548 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
549 struct ceph_snap_context *snapc;
550 struct ceph_vino vino;
551 struct ceph_osd_request *req;
559 struct timespec mtime = CURRENT_TIME;
560 loff_t pos = iocb->ki_pos;
563 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
566 dout("sync_direct_write on file %p %lld~%u\n", file, pos,
569 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
573 ret = invalidate_inode_pages2_range(inode->i_mapping,
574 pos >> PAGE_CACHE_SHIFT,
575 (pos + count) >> PAGE_CACHE_SHIFT);
577 dout("invalidate_inode_pages2_range returned %d\n", ret);
579 flags = CEPH_OSD_FLAG_ORDERSNAP |
580 CEPH_OSD_FLAG_ONDISK |
583 iov_iter_init(&i, iov, nr_segs, count, 0);
585 while (iov_iter_count(&i) > 0) {
586 void __user *data = i.iov->iov_base + i.iov_offset;
587 u64 len = i.iov->iov_len - i.iov_offset;
589 page_align = (unsigned long)data & ~PAGE_MASK;
591 snapc = ci->i_snap_realm->cached_context;
592 vino = ceph_vino(inode);
593 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
595 2,/*include a 'startsync' command*/
596 CEPH_OSD_OP_WRITE, flags, snapc,
605 num_pages = calc_pages_for(page_align, len);
606 pages = ceph_get_direct_page_vector(data, num_pages, false);
608 ret = PTR_ERR(pages);
613 * throw out any page cache pages in this range. this
616 truncate_inode_pages_range(inode->i_mapping, pos,
617 (pos+len) | (PAGE_CACHE_SIZE-1));
618 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align,
621 /* BUG_ON(vino.snap != CEPH_NOSNAP); */
622 ceph_osdc_build_request(req, pos, snapc, vino.snap, &mtime);
624 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
626 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
628 ceph_put_page_vector(pages, num_pages, false);
631 ceph_osdc_put_request(req);
635 iov_iter_advance(&i, (size_t)len);
637 if (pos > i_size_read(inode)) {
638 check_caps = ceph_inode_set_size(inode, pos);
640 ceph_check_caps(ceph_inode(inode),
648 if (ret != -EOLDSNAPC && written > 0) {
657 * Synchronous write, straight from __user pointer or user pages.
659 * If write spans object boundary, just do multiple writes. (For a
660 * correct atomic write, we should e.g. take write locks on all
661 * objects, rollback on failure, etc.)
663 static ssize_t ceph_sync_write(struct kiocb *iocb, const struct iovec *iov,
664 unsigned long nr_segs, size_t count)
666 struct file *file = iocb->ki_filp;
667 struct inode *inode = file_inode(file);
668 struct ceph_inode_info *ci = ceph_inode(inode);
669 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
670 struct ceph_snap_context *snapc;
671 struct ceph_vino vino;
672 struct ceph_osd_request *req;
680 struct timespec mtime = CURRENT_TIME;
681 loff_t pos = iocb->ki_pos;
684 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
687 dout("sync_write on file %p %lld~%u\n", file, pos, (unsigned)count);
689 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
693 ret = invalidate_inode_pages2_range(inode->i_mapping,
694 pos >> PAGE_CACHE_SHIFT,
695 (pos + count) >> PAGE_CACHE_SHIFT);
697 dout("invalidate_inode_pages2_range returned %d\n", ret);
699 flags = CEPH_OSD_FLAG_ORDERSNAP |
700 CEPH_OSD_FLAG_ONDISK |
701 CEPH_OSD_FLAG_WRITE |
704 iov_iter_init(&i, iov, nr_segs, count, 0);
706 while ((len = iov_iter_count(&i)) > 0) {
710 snapc = ci->i_snap_realm->cached_context;
711 vino = ceph_vino(inode);
712 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
714 CEPH_OSD_OP_WRITE, flags, snapc,
724 * write from beginning of first page,
725 * regardless of io alignment
727 num_pages = (len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
729 pages = ceph_alloc_page_vector(num_pages, GFP_NOFS);
731 ret = PTR_ERR(pages);
736 for (n = 0; n < num_pages; n++) {
737 size_t plen = min_t(size_t, left, PAGE_SIZE);
738 ret = iov_iter_copy_from_user(pages[n], &i, 0, plen);
744 iov_iter_advance(&i, ret);
748 ceph_release_page_vector(pages, num_pages);
752 /* get a second commit callback */
753 req->r_unsafe_callback = ceph_sync_write_unsafe;
754 req->r_inode = inode;
756 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
759 /* BUG_ON(vino.snap != CEPH_NOSNAP); */
760 ceph_osdc_build_request(req, pos, snapc, vino.snap, &mtime);
762 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
764 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
767 ceph_osdc_put_request(req);
772 if (pos > i_size_read(inode)) {
773 check_caps = ceph_inode_set_size(inode, pos);
775 ceph_check_caps(ceph_inode(inode),
783 if (ret != -EOLDSNAPC && written > 0) {
791 * Wrap generic_file_aio_read with checks for cap bits on the inode.
792 * Atomically grab references, so that those bits are not released
793 * back to the MDS mid-read.
795 * Hmm, the sync read case isn't actually async... should it be?
797 static ssize_t ceph_aio_read(struct kiocb *iocb, const struct iovec *iov,
798 unsigned long nr_segs, loff_t pos)
800 struct file *filp = iocb->ki_filp;
801 struct ceph_file_info *fi = filp->private_data;
802 size_t len = iocb->ki_nbytes;
803 struct inode *inode = file_inode(filp);
804 struct ceph_inode_info *ci = ceph_inode(inode);
807 int checkeof = 0, read = 0;
810 dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
811 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
813 if (fi->fmode & CEPH_FILE_MODE_LAZY)
814 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
816 want = CEPH_CAP_FILE_CACHE;
817 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, &got, -1);
821 if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
822 (iocb->ki_filp->f_flags & O_DIRECT) ||
823 (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));
831 ret = generic_segment_checks(iov, &nr_segs,
837 iov_iter_init(&i, iov, nr_segs, len, read);
839 /* hmm, this isn't really async... */
840 ret = ceph_sync_read(iocb, &i, &checkeof);
843 * We can't modify the content of iov,
844 * so we only read from beginning.
848 len = iocb->ki_nbytes;
851 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
852 inode, ceph_vinop(inode), pos, (unsigned)len,
853 ceph_cap_string(got));
855 ret = generic_file_aio_read(iocb, iov, nr_segs, pos);
858 dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
859 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
860 ceph_put_cap_refs(ci, got);
862 if (checkeof && ret >= 0) {
863 int statret = ceph_do_getattr(inode,
866 /* hit EOF or hole? */
867 if (statret == 0 && iocb->ki_pos < inode->i_size &&
869 dout("sync_read hit hole, ppos %lld < size %lld"
870 ", reading more\n", iocb->ki_pos,
887 * Take cap references to avoid releasing caps to MDS mid-write.
889 * If we are synchronous, and write with an old snap context, the OSD
890 * may return EOLDSNAPC. In that case, retry the write.. _after_
891 * dropping our cap refs and allowing the pending snap to logically
892 * complete _before_ this write occurs.
894 * If we are near ENOSPC, write synchronously.
896 static ssize_t ceph_aio_write(struct kiocb *iocb, const struct iovec *iov,
897 unsigned long nr_segs, loff_t pos)
899 struct file *file = iocb->ki_filp;
900 struct ceph_file_info *fi = file->private_data;
901 struct inode *inode = file_inode(file);
902 struct ceph_inode_info *ci = ceph_inode(inode);
903 struct ceph_osd_client *osdc =
904 &ceph_sb_to_client(inode->i_sb)->client->osdc;
905 ssize_t count, written = 0;
908 if (ceph_snap(inode) != CEPH_NOSNAP)
911 mutex_lock(&inode->i_mutex);
913 err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
917 /* We can write back this queue in page reclaim */
918 current->backing_dev_info = file->f_mapping->backing_dev_info;
920 err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
927 err = file_remove_suid(file);
931 err = file_update_time(file);
936 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) {
941 dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
942 inode, ceph_vinop(inode), pos, count, inode->i_size);
943 if (fi->fmode & CEPH_FILE_MODE_LAZY)
944 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
946 want = CEPH_CAP_FILE_BUFFER;
948 err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, &got, pos + count);
952 dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
953 inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
955 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
956 (file->f_flags & O_DIRECT) || (fi->flags & CEPH_F_SYNC)) {
957 mutex_unlock(&inode->i_mutex);
958 if (file->f_flags & O_DIRECT)
959 written = ceph_sync_direct_write(iocb, iov,
962 written = ceph_sync_write(iocb, iov, nr_segs, count);
963 if (written == -EOLDSNAPC) {
964 dout("aio_write %p %llx.%llx %llu~%u"
965 "got EOLDSNAPC, retrying\n",
966 inode, ceph_vinop(inode),
967 pos, (unsigned)iov->iov_len);
968 mutex_lock(&inode->i_mutex);
973 * No need to acquire the i_truncate_mutex. Because
974 * the MDS revokes Fwb caps before sending truncate
975 * message to us. We can't get Fwb cap while there
976 * are pending vmtruncate. So write and vmtruncate
977 * can not run at the same time
979 written = generic_file_buffered_write(iocb, iov, nr_segs,
982 mutex_unlock(&inode->i_mutex);
987 spin_lock(&ci->i_ceph_lock);
988 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR);
989 spin_unlock(&ci->i_ceph_lock);
991 __mark_inode_dirty(inode, dirty);
994 dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n",
995 inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len,
996 ceph_cap_string(got));
997 ceph_put_cap_refs(ci, got);
1000 ((file->f_flags & O_SYNC) || IS_SYNC(file->f_mapping->host) ||
1001 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_NEARFULL))) {
1002 err = vfs_fsync_range(file, pos, pos + written - 1, 1);
1010 mutex_unlock(&inode->i_mutex);
1012 current->backing_dev_info = NULL;
1013 return written ? written : err;
1017 * llseek. be sure to verify file size on SEEK_END.
1019 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1021 struct inode *inode = file->f_mapping->host;
1024 mutex_lock(&inode->i_mutex);
1026 if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1027 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE);
1036 offset += inode->i_size;
1040 * Here we special-case the lseek(fd, 0, SEEK_CUR)
1041 * position-querying operation. Avoid rewriting the "same"
1042 * f_pos value back to the file because a concurrent read(),
1043 * write() or lseek() might have altered it
1046 offset = file->f_pos;
1049 offset += file->f_pos;
1052 if (offset >= inode->i_size) {
1058 if (offset >= inode->i_size) {
1062 offset = inode->i_size;
1066 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1069 mutex_unlock(&inode->i_mutex);
1073 static inline void ceph_zero_partial_page(
1074 struct inode *inode, loff_t offset, unsigned size)
1077 pgoff_t index = offset >> PAGE_CACHE_SHIFT;
1079 page = find_lock_page(inode->i_mapping, index);
1081 wait_on_page_writeback(page);
1082 zero_user(page, offset & (PAGE_CACHE_SIZE - 1), size);
1084 page_cache_release(page);
1088 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1091 loff_t nearly = round_up(offset, PAGE_CACHE_SIZE);
1092 if (offset < nearly) {
1093 loff_t size = nearly - offset;
1096 ceph_zero_partial_page(inode, offset, size);
1100 if (length >= PAGE_CACHE_SIZE) {
1101 loff_t size = round_down(length, PAGE_CACHE_SIZE);
1102 truncate_pagecache_range(inode, offset, offset + size - 1);
1107 ceph_zero_partial_page(inode, offset, length);
1110 static int ceph_zero_partial_object(struct inode *inode,
1111 loff_t offset, loff_t *length)
1113 struct ceph_inode_info *ci = ceph_inode(inode);
1114 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1115 struct ceph_osd_request *req;
1121 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1124 op = CEPH_OSD_OP_ZERO;
1127 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1131 CEPH_OSD_FLAG_WRITE |
1132 CEPH_OSD_FLAG_ONDISK,
1139 ceph_osdc_build_request(req, offset, NULL, ceph_vino(inode).snap,
1142 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1144 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1148 ceph_osdc_put_request(req);
1154 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
1157 struct ceph_inode_info *ci = ceph_inode(inode);
1158 s32 stripe_unit = ceph_file_layout_su(ci->i_layout);
1159 s32 stripe_count = ceph_file_layout_stripe_count(ci->i_layout);
1160 s32 object_size = ceph_file_layout_object_size(ci->i_layout);
1161 u64 object_set_size = object_size * stripe_count;
1164 /* round offset up to next period boundary */
1165 nearly = offset + object_set_size - 1;
1167 nearly -= do_div(t, object_set_size);
1169 while (length && offset < nearly) {
1170 loff_t size = length;
1171 ret = ceph_zero_partial_object(inode, offset, &size);
1177 while (length >= object_set_size) {
1179 loff_t pos = offset;
1180 for (i = 0; i < stripe_count; ++i) {
1181 ret = ceph_zero_partial_object(inode, pos, NULL);
1186 offset += object_set_size;
1187 length -= object_set_size;
1190 loff_t size = length;
1191 ret = ceph_zero_partial_object(inode, offset, &size);
1200 static long ceph_fallocate(struct file *file, int mode,
1201 loff_t offset, loff_t length)
1203 struct ceph_file_info *fi = file->private_data;
1204 struct inode *inode = file_inode(file);
1205 struct ceph_inode_info *ci = ceph_inode(inode);
1206 struct ceph_osd_client *osdc =
1207 &ceph_inode_to_client(inode)->client->osdc;
1214 if (!S_ISREG(inode->i_mode))
1217 if (IS_SWAPFILE(inode))
1220 mutex_lock(&inode->i_mutex);
1222 if (ceph_snap(inode) != CEPH_NOSNAP) {
1227 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) &&
1228 !(mode & FALLOC_FL_PUNCH_HOLE)) {
1233 size = i_size_read(inode);
1234 if (!(mode & FALLOC_FL_KEEP_SIZE))
1235 endoff = offset + length;
1237 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1238 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1240 want = CEPH_CAP_FILE_BUFFER;
1242 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, &got, endoff);
1246 if (mode & FALLOC_FL_PUNCH_HOLE) {
1248 ceph_zero_pagecache_range(inode, offset, length);
1249 ret = ceph_zero_objects(inode, offset, length);
1250 } else if (endoff > size) {
1251 truncate_pagecache_range(inode, size, -1);
1252 if (ceph_inode_set_size(inode, endoff))
1253 ceph_check_caps(ceph_inode(inode),
1254 CHECK_CAPS_AUTHONLY, NULL);
1258 spin_lock(&ci->i_ceph_lock);
1259 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR);
1260 spin_unlock(&ci->i_ceph_lock);
1262 __mark_inode_dirty(inode, dirty);
1265 ceph_put_cap_refs(ci, got);
1267 mutex_unlock(&inode->i_mutex);
1271 const struct file_operations ceph_file_fops = {
1273 .release = ceph_release,
1274 .llseek = ceph_llseek,
1275 .read = do_sync_read,
1276 .write = do_sync_write,
1277 .aio_read = ceph_aio_read,
1278 .aio_write = ceph_aio_write,
1280 .fsync = ceph_fsync,
1282 .flock = ceph_flock,
1283 .splice_read = generic_file_splice_read,
1284 .splice_write = generic_file_splice_write,
1285 .unlocked_ioctl = ceph_ioctl,
1286 .compat_ioctl = ceph_ioctl,
1287 .fallocate = ceph_fallocate,