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/falloc.h>
13 #include "mds_client.h"
17 * Ceph file operations
19 * Implement basic open/close functionality, and implement
22 * We implement three modes of file I/O:
23 * - buffered uses the generic_file_aio_{read,write} helpers
25 * - synchronous is used when there is multi-client read/write
26 * sharing, avoids the page cache, and synchronously waits for an
29 * - direct io takes the variant of the sync path that references
30 * user pages directly.
32 * fsync() flushes and waits on dirty pages, but just queues metadata
33 * for writeback: since the MDS can recover size and mtime there is no
34 * need to wait for MDS acknowledgement.
38 * Calculate the length sum of direct io vectors that can
39 * be combined into one page vector.
41 static size_t dio_get_pagev_size(const struct iov_iter *it)
43 const struct iovec *iov = it->iov;
44 const struct iovec *iovend = iov + it->nr_segs;
47 size = iov->iov_len - it->iov_offset;
49 * An iov can be page vectored when both the current tail
50 * and the next base are page aligned.
52 while (PAGE_ALIGNED((iov->iov_base + iov->iov_len)) &&
53 (++iov < iovend && PAGE_ALIGNED((iov->iov_base)))) {
56 dout("dio_get_pagevlen len = %zu\n", size);
61 * Allocate a page vector based on (@it, @nbytes).
62 * The return value is the tuple describing a page vector,
63 * that is (@pages, @page_align, @num_pages).
66 dio_get_pages_alloc(const struct iov_iter *it, size_t nbytes,
67 size_t *page_align, int *num_pages)
69 struct iov_iter tmp_it = *it;
72 int ret = 0, idx, npages;
74 align = (unsigned long)(it->iov->iov_base + it->iov_offset) &
76 npages = calc_pages_for(align, nbytes);
77 pages = kmalloc(sizeof(*pages) * npages, GFP_KERNEL);
79 pages = vmalloc(sizeof(*pages) * npages);
81 return ERR_PTR(-ENOMEM);
84 for (idx = 0; idx < npages; ) {
86 ret = iov_iter_get_pages(&tmp_it, pages + idx, nbytes,
87 npages - idx, &start);
91 iov_iter_advance(&tmp_it, ret);
93 idx += (ret + start + PAGE_SIZE - 1) / PAGE_SIZE;
99 dout("dio_get_pages_alloc: got %d pages align %zu\n", npages, align);
102 ceph_put_page_vector(pages, idx, false);
107 * Prepare an open request. Preallocate ceph_cap to avoid an
108 * inopportune ENOMEM later.
110 static struct ceph_mds_request *
111 prepare_open_request(struct super_block *sb, int flags, int create_mode)
113 struct ceph_fs_client *fsc = ceph_sb_to_client(sb);
114 struct ceph_mds_client *mdsc = fsc->mdsc;
115 struct ceph_mds_request *req;
116 int want_auth = USE_ANY_MDS;
117 int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
119 if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
120 want_auth = USE_AUTH_MDS;
122 req = ceph_mdsc_create_request(mdsc, op, want_auth);
125 req->r_fmode = ceph_flags_to_mode(flags);
126 req->r_args.open.flags = cpu_to_le32(flags);
127 req->r_args.open.mode = cpu_to_le32(create_mode);
133 * initialize private struct file data.
134 * if we fail, clean up by dropping fmode reference on the ceph_inode
136 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
138 struct ceph_file_info *cf;
140 struct ceph_inode_info *ci = ceph_inode(inode);
141 struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
142 struct ceph_mds_client *mdsc = fsc->mdsc;
144 switch (inode->i_mode & S_IFMT) {
146 /* First file open request creates the cookie, we want to keep
147 * this cookie around for the filetime of the inode as not to
148 * have to worry about fscache register / revoke / operation
151 * Also, if we know the operation is going to invalidate data
152 * (non readonly) just nuke the cache right away.
154 ceph_fscache_register_inode_cookie(mdsc->fsc, ci);
155 if ((fmode & CEPH_FILE_MODE_WR))
156 ceph_fscache_invalidate(inode);
158 dout("init_file %p %p 0%o (regular)\n", inode, file,
160 cf = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
162 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
167 cf->readdir_cache_idx = -1;
168 file->private_data = cf;
169 BUG_ON(inode->i_fop->release != ceph_release);
173 dout("init_file %p %p 0%o (symlink)\n", inode, file,
175 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
179 dout("init_file %p %p 0%o (special)\n", inode, file,
182 * we need to drop the open ref now, since we don't
183 * have .release set to ceph_release.
185 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
186 BUG_ON(inode->i_fop->release == ceph_release);
188 /* call the proper open fop */
189 ret = inode->i_fop->open(inode, file);
195 * If we already have the requisite capabilities, we can satisfy
196 * the open request locally (no need to request new caps from the
197 * MDS). We do, however, need to inform the MDS (asynchronously)
198 * if our wanted caps set expands.
200 int ceph_open(struct inode *inode, struct file *file)
202 struct ceph_inode_info *ci = ceph_inode(inode);
203 struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
204 struct ceph_mds_client *mdsc = fsc->mdsc;
205 struct ceph_mds_request *req;
206 struct ceph_file_info *cf = file->private_data;
208 int flags, fmode, wanted;
211 dout("open file %p is already opened\n", file);
215 /* filter out O_CREAT|O_EXCL; vfs did that already. yuck. */
216 flags = file->f_flags & ~(O_CREAT|O_EXCL);
217 if (S_ISDIR(inode->i_mode))
218 flags = O_DIRECTORY; /* mds likes to know */
220 dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
221 ceph_vinop(inode), file, flags, file->f_flags);
222 fmode = ceph_flags_to_mode(flags);
223 wanted = ceph_caps_for_mode(fmode);
225 /* snapped files are read-only */
226 if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
229 /* trivially open snapdir */
230 if (ceph_snap(inode) == CEPH_SNAPDIR) {
231 spin_lock(&ci->i_ceph_lock);
232 __ceph_get_fmode(ci, fmode);
233 spin_unlock(&ci->i_ceph_lock);
234 return ceph_init_file(inode, file, fmode);
238 * No need to block if we have caps on the auth MDS (for
239 * write) or any MDS (for read). Update wanted set
242 spin_lock(&ci->i_ceph_lock);
243 if (__ceph_is_any_real_caps(ci) &&
244 (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
245 int mds_wanted = __ceph_caps_mds_wanted(ci);
246 int issued = __ceph_caps_issued(ci, NULL);
248 dout("open %p fmode %d want %s issued %s using existing\n",
249 inode, fmode, ceph_cap_string(wanted),
250 ceph_cap_string(issued));
251 __ceph_get_fmode(ci, fmode);
252 spin_unlock(&ci->i_ceph_lock);
255 if ((issued & wanted) != wanted &&
256 (mds_wanted & wanted) != wanted &&
257 ceph_snap(inode) != CEPH_SNAPDIR)
258 ceph_check_caps(ci, 0, NULL);
260 return ceph_init_file(inode, file, fmode);
261 } else if (ceph_snap(inode) != CEPH_NOSNAP &&
262 (ci->i_snap_caps & wanted) == wanted) {
263 __ceph_get_fmode(ci, fmode);
264 spin_unlock(&ci->i_ceph_lock);
265 return ceph_init_file(inode, file, fmode);
268 spin_unlock(&ci->i_ceph_lock);
270 dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
271 req = prepare_open_request(inode->i_sb, flags, 0);
276 req->r_inode = inode;
280 err = ceph_mdsc_do_request(mdsc, NULL, req);
282 err = ceph_init_file(inode, file, req->r_fmode);
283 ceph_mdsc_put_request(req);
284 dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
291 * Do a lookup + open with a single request. If we get a non-existent
292 * file or symlink, return 1 so the VFS can retry.
294 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
295 struct file *file, unsigned flags, umode_t mode,
298 struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
299 struct ceph_mds_client *mdsc = fsc->mdsc;
300 struct ceph_mds_request *req;
302 struct ceph_acls_info acls = {};
306 dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
308 d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
310 if (dentry->d_name.len > NAME_MAX)
311 return -ENAMETOOLONG;
313 err = ceph_init_dentry(dentry);
317 if (flags & O_CREAT) {
318 err = ceph_pre_init_acls(dir, &mode, &acls);
324 req = prepare_open_request(dir->i_sb, flags, mode);
329 req->r_dentry = dget(dentry);
331 if (flags & O_CREAT) {
332 req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
333 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
335 req->r_pagelist = acls.pagelist;
336 acls.pagelist = NULL;
340 mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
341 if (ceph_security_xattr_wanted(dir))
342 mask |= CEPH_CAP_XATTR_SHARED;
343 req->r_args.open.mask = cpu_to_le32(mask);
345 req->r_locked_dir = dir; /* caller holds dir->i_mutex */
346 err = ceph_mdsc_do_request(mdsc,
347 (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
349 err = ceph_handle_snapdir(req, dentry, err);
353 if ((flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
354 err = ceph_handle_notrace_create(dir, dentry);
356 if (d_unhashed(dentry)) {
357 dn = ceph_finish_lookup(req, dentry, err);
361 /* we were given a hashed negative dentry */
366 if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
367 /* make vfs retry on splice, ENOENT, or symlink */
368 dout("atomic_open finish_no_open on dn %p\n", dn);
369 err = finish_no_open(file, dn);
371 dout("atomic_open finish_open on dn %p\n", dn);
372 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
373 ceph_init_inode_acls(d_inode(dentry), &acls);
374 *opened |= FILE_CREATED;
376 err = finish_open(file, dentry, ceph_open, opened);
379 if (!req->r_err && req->r_target_inode)
380 ceph_put_fmode(ceph_inode(req->r_target_inode), req->r_fmode);
381 ceph_mdsc_put_request(req);
383 ceph_release_acls_info(&acls);
384 dout("atomic_open result=%d\n", err);
388 int ceph_release(struct inode *inode, struct file *file)
390 struct ceph_inode_info *ci = ceph_inode(inode);
391 struct ceph_file_info *cf = file->private_data;
393 dout("release inode %p file %p\n", inode, file);
394 ceph_put_fmode(ci, cf->fmode);
395 if (cf->last_readdir)
396 ceph_mdsc_put_request(cf->last_readdir);
397 kfree(cf->last_name);
399 kmem_cache_free(ceph_file_cachep, cf);
401 /* wake up anyone waiting for caps on this inode */
402 wake_up_all(&ci->i_cap_wq);
413 * Read a range of bytes striped over one or more objects. Iterate over
414 * objects we stripe over. (That's not atomic, but good enough for now.)
416 * If we get a short result from the OSD, check against i_size; we need to
417 * only return a short read to the caller if we hit EOF.
419 static int striped_read(struct inode *inode,
421 struct page **pages, int num_pages,
424 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
425 struct ceph_inode_info *ci = ceph_inode(inode);
426 u64 pos, this_len, left;
428 int page_align, pages_left;
430 struct page **page_pos;
431 bool hit_stripe, was_short;
434 * we may need to do multiple reads. not atomic, unfortunately.
439 pages_left = num_pages;
443 page_align = pos & ~PAGE_MASK;
445 ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode),
446 &ci->i_layout, pos, &this_len,
449 page_pos, pages_left, page_align);
452 hit_stripe = this_len < left;
453 was_short = ret >= 0 && ret < this_len;
454 dout("striped_read %llu~%llu (read %u) got %d%s%s\n", pos, left, read,
455 ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : "");
457 i_size = i_size_read(inode);
460 if (was_short && (pos + ret < i_size)) {
461 int zlen = min(this_len - ret, i_size - pos - ret);
462 int zoff = (off & ~PAGE_MASK) + read + ret;
463 dout(" zero gap %llu to %llu\n",
464 pos + ret, pos + ret + zlen);
465 ceph_zero_page_vector_range(zoff, zlen, pages);
469 didpages = (page_align + ret) >> PAGE_CACHE_SHIFT;
473 page_pos += didpages;
474 pages_left -= didpages;
476 /* hit stripe and need continue*/
477 if (left && hit_stripe && pos < i_size)
483 /* did we bounce off eof? */
484 if (pos + left > i_size)
485 *checkeof = CHECK_EOF;
488 dout("striped_read returns %d\n", ret);
493 * Completely synchronous read and write methods. Direct from __user
494 * buffer to osd, or directly to user pages (if O_DIRECT).
496 * If the read spans object boundary, just do multiple reads.
498 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *i,
501 struct file *file = iocb->ki_filp;
502 struct inode *inode = file_inode(file);
504 u64 off = iocb->ki_pos;
506 size_t len = iov_iter_count(i);
508 dout("sync_read on file %p %llu~%u %s\n", file, off,
510 (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
515 * flush any page cache pages in this range. this
516 * will make concurrent normal and sync io slow,
517 * but it will at least behave sensibly when they are
520 ret = filemap_write_and_wait_range(inode->i_mapping, off,
525 num_pages = calc_pages_for(off, len);
526 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
528 return PTR_ERR(pages);
529 ret = striped_read(inode, off, len, pages,
530 num_pages, checkeof);
536 size_t page_off = off & ~PAGE_MASK;
537 size_t copy = min_t(size_t, left,
538 PAGE_SIZE - page_off);
539 l = copy_page_to_iter(pages[k++], page_off, copy, i);
546 ceph_release_page_vector(pages, num_pages);
548 if (off > iocb->ki_pos) {
549 ret = off - iocb->ki_pos;
553 dout("sync_read result %d\n", ret);
557 struct ceph_aio_request {
562 struct list_head osd_reqs;
564 atomic_t pending_reqs;
565 struct timespec mtime;
566 struct ceph_cap_flush *prealloc_cf;
569 struct ceph_aio_work {
570 struct work_struct work;
571 struct ceph_osd_request *req;
574 static void ceph_aio_retry_work(struct work_struct *work);
576 static void ceph_aio_complete(struct inode *inode,
577 struct ceph_aio_request *aio_req)
579 struct ceph_inode_info *ci = ceph_inode(inode);
582 if (!atomic_dec_and_test(&aio_req->pending_reqs))
585 ret = aio_req->error;
587 ret = aio_req->total_len;
589 dout("ceph_aio_complete %p rc %d\n", inode, ret);
591 if (ret >= 0 && aio_req->write) {
594 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
595 if (endoff > i_size_read(inode)) {
596 if (ceph_inode_set_size(inode, endoff))
597 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
600 spin_lock(&ci->i_ceph_lock);
601 ci->i_inline_version = CEPH_INLINE_NONE;
602 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
603 &aio_req->prealloc_cf);
604 spin_unlock(&ci->i_ceph_lock);
606 __mark_inode_dirty(inode, dirty);
610 ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
613 aio_req->iocb->ki_complete(aio_req->iocb, ret, 0);
615 ceph_free_cap_flush(aio_req->prealloc_cf);
619 static void ceph_aio_complete_req(struct ceph_osd_request *req,
620 struct ceph_msg *msg)
622 int rc = req->r_result;
623 struct inode *inode = req->r_inode;
624 struct ceph_aio_request *aio_req = req->r_priv;
625 struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
626 int num_pages = calc_pages_for((u64)osd_data->alignment,
629 dout("ceph_aio_complete_req %p rc %d bytes %llu\n",
630 inode, rc, osd_data->length);
632 if (rc == -EOLDSNAPC) {
633 struct ceph_aio_work *aio_work;
634 BUG_ON(!aio_req->write);
636 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
638 INIT_WORK(&aio_work->work, ceph_aio_retry_work);
640 queue_work(ceph_inode_to_client(inode)->wb_wq,
645 } else if (!aio_req->write) {
648 if (rc >= 0 && osd_data->length > rc) {
649 int zoff = osd_data->alignment + rc;
650 int zlen = osd_data->length - rc;
652 * If read is satisfied by single OSD request,
653 * it can pass EOF. Otherwise read is within
656 if (aio_req->num_reqs == 1) {
657 loff_t i_size = i_size_read(inode);
658 loff_t endoff = aio_req->iocb->ki_pos + rc;
660 zlen = min_t(size_t, zlen,
662 aio_req->total_len = rc + zlen;
666 ceph_zero_page_vector_range(zoff, zlen,
671 ceph_put_page_vector(osd_data->pages, num_pages, false);
672 ceph_osdc_put_request(req);
675 cmpxchg(&aio_req->error, 0, rc);
677 ceph_aio_complete(inode, aio_req);
681 static void ceph_aio_retry_work(struct work_struct *work)
683 struct ceph_aio_work *aio_work =
684 container_of(work, struct ceph_aio_work, work);
685 struct ceph_osd_request *orig_req = aio_work->req;
686 struct ceph_aio_request *aio_req = orig_req->r_priv;
687 struct inode *inode = orig_req->r_inode;
688 struct ceph_inode_info *ci = ceph_inode(inode);
689 struct ceph_snap_context *snapc;
690 struct ceph_osd_request *req;
693 spin_lock(&ci->i_ceph_lock);
694 if (__ceph_have_pending_cap_snap(ci)) {
695 struct ceph_cap_snap *capsnap =
696 list_last_entry(&ci->i_cap_snaps,
697 struct ceph_cap_snap,
699 snapc = ceph_get_snap_context(capsnap->context);
701 BUG_ON(!ci->i_head_snapc);
702 snapc = ceph_get_snap_context(ci->i_head_snapc);
704 spin_unlock(&ci->i_ceph_lock);
706 req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 2,
714 req->r_flags = CEPH_OSD_FLAG_ORDERSNAP |
715 CEPH_OSD_FLAG_ONDISK |
717 req->r_base_oloc = orig_req->r_base_oloc;
718 req->r_base_oid = orig_req->r_base_oid;
720 req->r_ops[0] = orig_req->r_ops[0];
721 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC, 0);
723 ceph_osdc_build_request(req, req->r_ops[0].extent.offset,
724 snapc, CEPH_NOSNAP, &aio_req->mtime);
726 ceph_osdc_put_request(orig_req);
728 req->r_callback = ceph_aio_complete_req;
729 req->r_inode = inode;
730 req->r_priv = aio_req;
732 ret = ceph_osdc_start_request(req->r_osdc, req, false);
736 ceph_aio_complete_req(req, NULL);
739 ceph_put_snap_context(snapc);
744 * Write commit request unsafe callback, called to tell us when a
745 * request is unsafe (that is, in flight--has been handed to the
746 * messenger to send to its target osd). It is called again when
747 * we've received a response message indicating the request is
748 * "safe" (its CEPH_OSD_FLAG_ONDISK flag is set), or when a request
749 * is completed early (and unsuccessfully) due to a timeout or
752 * This is used if we requested both an ACK and ONDISK commit reply
755 static void ceph_sync_write_unsafe(struct ceph_osd_request *req, bool unsafe)
757 struct ceph_inode_info *ci = ceph_inode(req->r_inode);
759 dout("%s %p tid %llu %ssafe\n", __func__, req, req->r_tid,
762 ceph_get_cap_refs(ci, CEPH_CAP_FILE_WR);
763 spin_lock(&ci->i_unsafe_lock);
764 list_add_tail(&req->r_unsafe_item,
765 &ci->i_unsafe_writes);
766 spin_unlock(&ci->i_unsafe_lock);
768 spin_lock(&ci->i_unsafe_lock);
769 list_del_init(&req->r_unsafe_item);
770 spin_unlock(&ci->i_unsafe_lock);
771 ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR);
777 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
778 struct ceph_snap_context *snapc,
779 struct ceph_cap_flush **pcf)
781 struct file *file = iocb->ki_filp;
782 struct inode *inode = file_inode(file);
783 struct ceph_inode_info *ci = ceph_inode(inode);
784 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
785 struct ceph_vino vino;
786 struct ceph_osd_request *req;
788 struct ceph_aio_request *aio_req = NULL;
792 struct timespec mtime = current_fs_time(inode->i_sb);
793 size_t count = iov_iter_count(iter);
794 loff_t pos = iocb->ki_pos;
795 bool write = iov_iter_rw(iter) == WRITE;
797 if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
800 dout("sync_direct_read_write (%s) on file %p %lld~%u\n",
801 (write ? "write" : "read"), file, pos, (unsigned)count);
803 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
808 ret = invalidate_inode_pages2_range(inode->i_mapping,
809 pos >> PAGE_CACHE_SHIFT,
810 (pos + count) >> PAGE_CACHE_SHIFT);
812 dout("invalidate_inode_pages2_range returned %d\n", ret);
814 flags = CEPH_OSD_FLAG_ORDERSNAP |
815 CEPH_OSD_FLAG_ONDISK |
818 flags = CEPH_OSD_FLAG_READ;
821 while (iov_iter_count(iter) > 0) {
822 u64 size = dio_get_pagev_size(iter);
826 vino = ceph_vino(inode);
827 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
829 /*include a 'startsync' command*/
831 write ? CEPH_OSD_OP_WRITE :
843 pages = dio_get_pages_alloc(iter, len, &start, &num_pages);
845 ceph_osdc_put_request(req);
846 ret = PTR_ERR(pages);
851 * To simplify error handling, allow AIO when IO within i_size
852 * or IO can be satisfied by single OSD request.
854 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
855 (len == count || pos + count <= i_size_read(inode))) {
856 aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
858 aio_req->iocb = iocb;
859 aio_req->write = write;
860 INIT_LIST_HEAD(&aio_req->osd_reqs);
862 aio_req->mtime = mtime;
863 swap(aio_req->prealloc_cf, *pcf);
871 * throw out any page cache pages in this range. this
874 truncate_inode_pages_range(inode->i_mapping, pos,
875 (pos+len) | (PAGE_CACHE_SIZE - 1));
877 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC, 0);
881 osd_req_op_extent_osd_data_pages(req, 0, pages, len, start,
884 ceph_osdc_build_request(req, pos, snapc, vino.snap, &mtime);
887 aio_req->total_len += len;
889 atomic_inc(&aio_req->pending_reqs);
891 req->r_callback = ceph_aio_complete_req;
892 req->r_inode = inode;
893 req->r_priv = aio_req;
894 list_add_tail(&req->r_unsafe_item, &aio_req->osd_reqs);
897 iov_iter_advance(iter, len);
901 ret = ceph_osdc_start_request(req->r_osdc, req, false);
903 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
905 size = i_size_read(inode);
909 if (ret >= 0 && ret < len && pos + ret < size) {
910 int zlen = min_t(size_t, len - ret,
912 ceph_zero_page_vector_range(start + ret, zlen,
920 ceph_put_page_vector(pages, num_pages, false);
922 ceph_osdc_put_request(req);
927 iov_iter_advance(iter, len);
929 if (!write && pos >= size)
932 if (write && pos > size) {
933 if (ceph_inode_set_size(inode, pos))
934 ceph_check_caps(ceph_inode(inode),
941 if (aio_req->num_reqs == 0) {
946 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
949 while (!list_empty(&aio_req->osd_reqs)) {
950 req = list_first_entry(&aio_req->osd_reqs,
951 struct ceph_osd_request,
953 list_del_init(&req->r_unsafe_item);
955 ret = ceph_osdc_start_request(req->r_osdc,
959 ceph_aio_complete_req(req, NULL);
965 if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
966 ret = pos - iocb->ki_pos;
973 * Synchronous write, straight from __user pointer or user pages.
975 * If write spans object boundary, just do multiple writes. (For a
976 * correct atomic write, we should e.g. take write locks on all
977 * objects, rollback on failure, etc.)
980 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
981 struct ceph_snap_context *snapc)
983 struct file *file = iocb->ki_filp;
984 struct inode *inode = file_inode(file);
985 struct ceph_inode_info *ci = ceph_inode(inode);
986 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
987 struct ceph_vino vino;
988 struct ceph_osd_request *req;
996 struct timespec mtime = current_fs_time(inode->i_sb);
997 size_t count = iov_iter_count(from);
999 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1002 dout("sync_write on file %p %lld~%u\n", file, pos, (unsigned)count);
1004 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
1008 ret = invalidate_inode_pages2_range(inode->i_mapping,
1009 pos >> PAGE_CACHE_SHIFT,
1010 (pos + count) >> PAGE_CACHE_SHIFT);
1012 dout("invalidate_inode_pages2_range returned %d\n", ret);
1014 flags = CEPH_OSD_FLAG_ORDERSNAP |
1015 CEPH_OSD_FLAG_ONDISK |
1016 CEPH_OSD_FLAG_WRITE |
1019 while ((len = iov_iter_count(from)) > 0) {
1023 vino = ceph_vino(inode);
1024 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1025 vino, pos, &len, 0, 1,
1026 CEPH_OSD_OP_WRITE, flags, snapc,
1028 ci->i_truncate_size,
1036 * write from beginning of first page,
1037 * regardless of io alignment
1039 num_pages = (len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1041 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1042 if (IS_ERR(pages)) {
1043 ret = PTR_ERR(pages);
1048 for (n = 0; n < num_pages; n++) {
1049 size_t plen = min_t(size_t, left, PAGE_SIZE);
1050 ret = copy_page_from_iter(pages[n], 0, plen, from);
1059 ceph_release_page_vector(pages, num_pages);
1063 /* get a second commit callback */
1064 req->r_unsafe_callback = ceph_sync_write_unsafe;
1065 req->r_inode = inode;
1067 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1070 /* BUG_ON(vino.snap != CEPH_NOSNAP); */
1071 ceph_osdc_build_request(req, pos, snapc, vino.snap, &mtime);
1073 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1075 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1078 ceph_osdc_put_request(req);
1083 if (pos > i_size_read(inode)) {
1084 check_caps = ceph_inode_set_size(inode, pos);
1086 ceph_check_caps(ceph_inode(inode),
1087 CHECK_CAPS_AUTHONLY,
1094 if (ret != -EOLDSNAPC && written > 0) {
1102 * Wrap generic_file_aio_read with checks for cap bits on the inode.
1103 * Atomically grab references, so that those bits are not released
1104 * back to the MDS mid-read.
1106 * Hmm, the sync read case isn't actually async... should it be?
1108 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1110 struct file *filp = iocb->ki_filp;
1111 struct ceph_file_info *fi = filp->private_data;
1112 size_t len = iov_iter_count(to);
1113 struct inode *inode = file_inode(filp);
1114 struct ceph_inode_info *ci = ceph_inode(inode);
1115 struct page *pinned_page = NULL;
1118 int retry_op = 0, read = 0;
1121 dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1122 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1124 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1125 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1127 want = CEPH_CAP_FILE_CACHE;
1128 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1132 if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1133 (iocb->ki_flags & IOCB_DIRECT) ||
1134 (fi->flags & CEPH_F_SYNC)) {
1136 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1137 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1138 ceph_cap_string(got));
1140 if (ci->i_inline_version == CEPH_INLINE_NONE) {
1141 if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1142 ret = ceph_direct_read_write(iocb, to,
1144 if (ret >= 0 && ret < len)
1145 retry_op = CHECK_EOF;
1147 ret = ceph_sync_read(iocb, to, &retry_op);
1150 retry_op = READ_INLINE;
1153 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1154 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1155 ceph_cap_string(got));
1157 ret = generic_file_read_iter(iocb, to);
1159 dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1160 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1162 page_cache_release(pinned_page);
1165 ceph_put_cap_refs(ci, got);
1166 if (retry_op > HAVE_RETRIED && ret >= 0) {
1168 struct page *page = NULL;
1170 if (retry_op == READ_INLINE) {
1171 page = __page_cache_alloc(GFP_KERNEL);
1176 statret = __ceph_do_getattr(inode, page,
1177 CEPH_STAT_CAP_INLINE_DATA, !!page);
1180 if (statret == -ENODATA) {
1181 BUG_ON(retry_op != READ_INLINE);
1187 i_size = i_size_read(inode);
1188 if (retry_op == READ_INLINE) {
1189 BUG_ON(ret > 0 || read > 0);
1190 if (iocb->ki_pos < i_size &&
1191 iocb->ki_pos < PAGE_CACHE_SIZE) {
1192 loff_t end = min_t(loff_t, i_size,
1193 iocb->ki_pos + len);
1194 end = min_t(loff_t, end, PAGE_CACHE_SIZE);
1196 zero_user_segment(page, statret, end);
1197 ret = copy_page_to_iter(page,
1198 iocb->ki_pos & ~PAGE_MASK,
1199 end - iocb->ki_pos, to);
1200 iocb->ki_pos += ret;
1203 if (iocb->ki_pos < i_size && read < len) {
1204 size_t zlen = min_t(size_t, len - read,
1205 i_size - iocb->ki_pos);
1206 ret = iov_iter_zero(zlen, to);
1207 iocb->ki_pos += ret;
1210 __free_pages(page, 0);
1214 /* hit EOF or hole? */
1215 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1217 dout("sync_read hit hole, ppos %lld < size %lld"
1218 ", reading more\n", iocb->ki_pos, i_size);
1222 retry_op = HAVE_RETRIED;
1234 * Take cap references to avoid releasing caps to MDS mid-write.
1236 * If we are synchronous, and write with an old snap context, the OSD
1237 * may return EOLDSNAPC. In that case, retry the write.. _after_
1238 * dropping our cap refs and allowing the pending snap to logically
1239 * complete _before_ this write occurs.
1241 * If we are near ENOSPC, write synchronously.
1243 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1245 struct file *file = iocb->ki_filp;
1246 struct ceph_file_info *fi = file->private_data;
1247 struct inode *inode = file_inode(file);
1248 struct ceph_inode_info *ci = ceph_inode(inode);
1249 struct ceph_osd_client *osdc =
1250 &ceph_sb_to_client(inode->i_sb)->client->osdc;
1251 struct ceph_cap_flush *prealloc_cf;
1252 ssize_t count, written = 0;
1256 if (ceph_snap(inode) != CEPH_NOSNAP)
1259 prealloc_cf = ceph_alloc_cap_flush();
1265 /* We can write back this queue in page reclaim */
1266 current->backing_dev_info = inode_to_bdi(inode);
1268 if (iocb->ki_flags & IOCB_APPEND) {
1269 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1274 err = generic_write_checks(iocb, from);
1279 count = iov_iter_count(from);
1280 err = file_remove_privs(file);
1284 err = file_update_time(file);
1288 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1289 err = ceph_uninline_data(file, NULL);
1295 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) {
1300 dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1301 inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1302 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1303 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1305 want = CEPH_CAP_FILE_BUFFER;
1307 err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, pos + count,
1312 dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1313 inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1315 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1316 (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC)) {
1317 struct ceph_snap_context *snapc;
1318 struct iov_iter data;
1319 inode_unlock(inode);
1321 spin_lock(&ci->i_ceph_lock);
1322 if (__ceph_have_pending_cap_snap(ci)) {
1323 struct ceph_cap_snap *capsnap =
1324 list_last_entry(&ci->i_cap_snaps,
1325 struct ceph_cap_snap,
1327 snapc = ceph_get_snap_context(capsnap->context);
1329 BUG_ON(!ci->i_head_snapc);
1330 snapc = ceph_get_snap_context(ci->i_head_snapc);
1332 spin_unlock(&ci->i_ceph_lock);
1334 /* we might need to revert back to that point */
1336 if (iocb->ki_flags & IOCB_DIRECT)
1337 written = ceph_direct_read_write(iocb, &data, snapc,
1340 written = ceph_sync_write(iocb, &data, pos, snapc);
1341 if (written == -EOLDSNAPC) {
1342 dout("aio_write %p %llx.%llx %llu~%u"
1343 "got EOLDSNAPC, retrying\n",
1344 inode, ceph_vinop(inode),
1345 pos, (unsigned)count);
1350 iov_iter_advance(from, written);
1351 ceph_put_snap_context(snapc);
1353 loff_t old_size = i_size_read(inode);
1355 * No need to acquire the i_truncate_mutex. Because
1356 * the MDS revokes Fwb caps before sending truncate
1357 * message to us. We can't get Fwb cap while there
1358 * are pending vmtruncate. So write and vmtruncate
1359 * can not run at the same time
1361 written = generic_perform_write(file, from, pos);
1362 if (likely(written >= 0))
1363 iocb->ki_pos = pos + written;
1364 if (i_size_read(inode) > old_size)
1365 ceph_fscache_update_objectsize(inode);
1366 inode_unlock(inode);
1371 spin_lock(&ci->i_ceph_lock);
1372 ci->i_inline_version = CEPH_INLINE_NONE;
1373 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1375 spin_unlock(&ci->i_ceph_lock);
1377 __mark_inode_dirty(inode, dirty);
1380 dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n",
1381 inode, ceph_vinop(inode), pos, (unsigned)count,
1382 ceph_cap_string(got));
1383 ceph_put_cap_refs(ci, got);
1386 ((file->f_flags & O_SYNC) || IS_SYNC(file->f_mapping->host) ||
1387 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_NEARFULL))) {
1388 err = vfs_fsync_range(file, pos, pos + written - 1, 1);
1396 inode_unlock(inode);
1398 ceph_free_cap_flush(prealloc_cf);
1399 current->backing_dev_info = NULL;
1400 return written ? written : err;
1404 * llseek. be sure to verify file size on SEEK_END.
1406 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1408 struct inode *inode = file->f_mapping->host;
1414 if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1415 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1422 i_size = i_size_read(inode);
1429 * Here we special-case the lseek(fd, 0, SEEK_CUR)
1430 * position-querying operation. Avoid rewriting the "same"
1431 * f_pos value back to the file because a concurrent read(),
1432 * write() or lseek() might have altered it
1435 offset = file->f_pos;
1438 offset += file->f_pos;
1441 if (offset >= i_size) {
1447 if (offset >= i_size) {
1455 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1458 inode_unlock(inode);
1462 static inline void ceph_zero_partial_page(
1463 struct inode *inode, loff_t offset, unsigned size)
1466 pgoff_t index = offset >> PAGE_CACHE_SHIFT;
1468 page = find_lock_page(inode->i_mapping, index);
1470 wait_on_page_writeback(page);
1471 zero_user(page, offset & (PAGE_CACHE_SIZE - 1), size);
1473 page_cache_release(page);
1477 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1480 loff_t nearly = round_up(offset, PAGE_CACHE_SIZE);
1481 if (offset < nearly) {
1482 loff_t size = nearly - offset;
1485 ceph_zero_partial_page(inode, offset, size);
1489 if (length >= PAGE_CACHE_SIZE) {
1490 loff_t size = round_down(length, PAGE_CACHE_SIZE);
1491 truncate_pagecache_range(inode, offset, offset + size - 1);
1496 ceph_zero_partial_page(inode, offset, length);
1499 static int ceph_zero_partial_object(struct inode *inode,
1500 loff_t offset, loff_t *length)
1502 struct ceph_inode_info *ci = ceph_inode(inode);
1503 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1504 struct ceph_osd_request *req;
1510 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1513 op = CEPH_OSD_OP_ZERO;
1516 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1520 CEPH_OSD_FLAG_WRITE |
1521 CEPH_OSD_FLAG_ONDISK,
1528 ceph_osdc_build_request(req, offset, NULL, ceph_vino(inode).snap,
1531 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1533 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1537 ceph_osdc_put_request(req);
1543 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
1546 struct ceph_inode_info *ci = ceph_inode(inode);
1547 s32 stripe_unit = ceph_file_layout_su(ci->i_layout);
1548 s32 stripe_count = ceph_file_layout_stripe_count(ci->i_layout);
1549 s32 object_size = ceph_file_layout_object_size(ci->i_layout);
1550 u64 object_set_size = object_size * stripe_count;
1553 /* round offset up to next period boundary */
1554 nearly = offset + object_set_size - 1;
1556 nearly -= do_div(t, object_set_size);
1558 while (length && offset < nearly) {
1559 loff_t size = length;
1560 ret = ceph_zero_partial_object(inode, offset, &size);
1566 while (length >= object_set_size) {
1568 loff_t pos = offset;
1569 for (i = 0; i < stripe_count; ++i) {
1570 ret = ceph_zero_partial_object(inode, pos, NULL);
1575 offset += object_set_size;
1576 length -= object_set_size;
1579 loff_t size = length;
1580 ret = ceph_zero_partial_object(inode, offset, &size);
1589 static long ceph_fallocate(struct file *file, int mode,
1590 loff_t offset, loff_t length)
1592 struct ceph_file_info *fi = file->private_data;
1593 struct inode *inode = file_inode(file);
1594 struct ceph_inode_info *ci = ceph_inode(inode);
1595 struct ceph_osd_client *osdc =
1596 &ceph_inode_to_client(inode)->client->osdc;
1597 struct ceph_cap_flush *prealloc_cf;
1604 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1607 if (!S_ISREG(inode->i_mode))
1610 prealloc_cf = ceph_alloc_cap_flush();
1616 if (ceph_snap(inode) != CEPH_NOSNAP) {
1621 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) &&
1622 !(mode & FALLOC_FL_PUNCH_HOLE)) {
1627 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1628 ret = ceph_uninline_data(file, NULL);
1633 size = i_size_read(inode);
1634 if (!(mode & FALLOC_FL_KEEP_SIZE))
1635 endoff = offset + length;
1637 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1638 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1640 want = CEPH_CAP_FILE_BUFFER;
1642 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, endoff, &got, NULL);
1646 if (mode & FALLOC_FL_PUNCH_HOLE) {
1648 ceph_zero_pagecache_range(inode, offset, length);
1649 ret = ceph_zero_objects(inode, offset, length);
1650 } else if (endoff > size) {
1651 truncate_pagecache_range(inode, size, -1);
1652 if (ceph_inode_set_size(inode, endoff))
1653 ceph_check_caps(ceph_inode(inode),
1654 CHECK_CAPS_AUTHONLY, NULL);
1658 spin_lock(&ci->i_ceph_lock);
1659 ci->i_inline_version = CEPH_INLINE_NONE;
1660 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1662 spin_unlock(&ci->i_ceph_lock);
1664 __mark_inode_dirty(inode, dirty);
1667 ceph_put_cap_refs(ci, got);
1669 inode_unlock(inode);
1670 ceph_free_cap_flush(prealloc_cf);
1674 const struct file_operations ceph_file_fops = {
1676 .release = ceph_release,
1677 .llseek = ceph_llseek,
1678 .read_iter = ceph_read_iter,
1679 .write_iter = ceph_write_iter,
1681 .fsync = ceph_fsync,
1683 .flock = ceph_flock,
1684 .splice_read = generic_file_splice_read,
1685 .splice_write = iter_file_splice_write,
1686 .unlocked_ioctl = ceph_ioctl,
1687 .compat_ioctl = ceph_ioctl,
1688 .fallocate = ceph_fallocate,