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;
141 switch (inode->i_mode & S_IFMT) {
143 ceph_fscache_register_inode_cookie(inode);
144 ceph_fscache_file_set_cookie(inode, file);
146 dout("init_file %p %p 0%o (regular)\n", inode, file,
148 cf = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
150 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
155 cf->readdir_cache_idx = -1;
156 file->private_data = cf;
157 BUG_ON(inode->i_fop->release != ceph_release);
161 dout("init_file %p %p 0%o (symlink)\n", inode, file,
163 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
167 dout("init_file %p %p 0%o (special)\n", inode, file,
170 * we need to drop the open ref now, since we don't
171 * have .release set to ceph_release.
173 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
174 BUG_ON(inode->i_fop->release == ceph_release);
176 /* call the proper open fop */
177 ret = inode->i_fop->open(inode, file);
183 * try renew caps after session gets killed.
185 int ceph_renew_caps(struct inode *inode)
187 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
188 struct ceph_inode_info *ci = ceph_inode(inode);
189 struct ceph_mds_request *req;
190 int err, flags, wanted;
192 spin_lock(&ci->i_ceph_lock);
193 wanted = __ceph_caps_file_wanted(ci);
194 if (__ceph_is_any_real_caps(ci) &&
195 (!(wanted & CEPH_CAP_ANY_WR) == 0 || ci->i_auth_cap)) {
196 int issued = __ceph_caps_issued(ci, NULL);
197 spin_unlock(&ci->i_ceph_lock);
198 dout("renew caps %p want %s issued %s updating mds_wanted\n",
199 inode, ceph_cap_string(wanted), ceph_cap_string(issued));
200 ceph_check_caps(ci, 0, NULL);
203 spin_unlock(&ci->i_ceph_lock);
206 if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
208 else if (wanted & CEPH_CAP_FILE_RD)
210 else if (wanted & CEPH_CAP_FILE_WR)
213 if (wanted & CEPH_CAP_FILE_LAZYIO)
217 req = prepare_open_request(inode->i_sb, flags, 0);
223 req->r_inode = inode;
228 err = ceph_mdsc_do_request(mdsc, NULL, req);
229 ceph_mdsc_put_request(req);
231 dout("renew caps %p open result=%d\n", inode, err);
232 return err < 0 ? err : 0;
236 * If we already have the requisite capabilities, we can satisfy
237 * the open request locally (no need to request new caps from the
238 * MDS). We do, however, need to inform the MDS (asynchronously)
239 * if our wanted caps set expands.
241 int ceph_open(struct inode *inode, struct file *file)
243 struct ceph_inode_info *ci = ceph_inode(inode);
244 struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
245 struct ceph_mds_client *mdsc = fsc->mdsc;
246 struct ceph_mds_request *req;
247 struct ceph_file_info *cf = file->private_data;
249 int flags, fmode, wanted;
252 dout("open file %p is already opened\n", file);
256 /* filter out O_CREAT|O_EXCL; vfs did that already. yuck. */
257 flags = file->f_flags & ~(O_CREAT|O_EXCL);
258 if (S_ISDIR(inode->i_mode))
259 flags = O_DIRECTORY; /* mds likes to know */
261 dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
262 ceph_vinop(inode), file, flags, file->f_flags);
263 fmode = ceph_flags_to_mode(flags);
264 wanted = ceph_caps_for_mode(fmode);
266 /* snapped files are read-only */
267 if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
270 /* trivially open snapdir */
271 if (ceph_snap(inode) == CEPH_SNAPDIR) {
272 spin_lock(&ci->i_ceph_lock);
273 __ceph_get_fmode(ci, fmode);
274 spin_unlock(&ci->i_ceph_lock);
275 return ceph_init_file(inode, file, fmode);
279 * No need to block if we have caps on the auth MDS (for
280 * write) or any MDS (for read). Update wanted set
283 spin_lock(&ci->i_ceph_lock);
284 if (__ceph_is_any_real_caps(ci) &&
285 (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
286 int mds_wanted = __ceph_caps_mds_wanted(ci);
287 int issued = __ceph_caps_issued(ci, NULL);
289 dout("open %p fmode %d want %s issued %s using existing\n",
290 inode, fmode, ceph_cap_string(wanted),
291 ceph_cap_string(issued));
292 __ceph_get_fmode(ci, fmode);
293 spin_unlock(&ci->i_ceph_lock);
296 if ((issued & wanted) != wanted &&
297 (mds_wanted & wanted) != wanted &&
298 ceph_snap(inode) != CEPH_SNAPDIR)
299 ceph_check_caps(ci, 0, NULL);
301 return ceph_init_file(inode, file, fmode);
302 } else if (ceph_snap(inode) != CEPH_NOSNAP &&
303 (ci->i_snap_caps & wanted) == wanted) {
304 __ceph_get_fmode(ci, fmode);
305 spin_unlock(&ci->i_ceph_lock);
306 return ceph_init_file(inode, file, fmode);
309 spin_unlock(&ci->i_ceph_lock);
311 dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
312 req = prepare_open_request(inode->i_sb, flags, 0);
317 req->r_inode = inode;
321 err = ceph_mdsc_do_request(mdsc, NULL, req);
323 err = ceph_init_file(inode, file, req->r_fmode);
324 ceph_mdsc_put_request(req);
325 dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
332 * Do a lookup + open with a single request. If we get a non-existent
333 * file or symlink, return 1 so the VFS can retry.
335 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
336 struct file *file, unsigned flags, umode_t mode,
339 struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
340 struct ceph_mds_client *mdsc = fsc->mdsc;
341 struct ceph_mds_request *req;
343 struct ceph_acls_info acls = {};
347 dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
349 d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
351 if (dentry->d_name.len > NAME_MAX)
352 return -ENAMETOOLONG;
354 if (flags & O_CREAT) {
355 err = ceph_pre_init_acls(dir, &mode, &acls);
361 req = prepare_open_request(dir->i_sb, flags, mode);
366 req->r_dentry = dget(dentry);
368 if (flags & O_CREAT) {
369 req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
370 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
372 req->r_pagelist = acls.pagelist;
373 acls.pagelist = NULL;
377 mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
378 if (ceph_security_xattr_wanted(dir))
379 mask |= CEPH_CAP_XATTR_SHARED;
380 req->r_args.open.mask = cpu_to_le32(mask);
382 req->r_locked_dir = dir; /* caller holds dir->i_mutex */
383 err = ceph_mdsc_do_request(mdsc,
384 (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
386 err = ceph_handle_snapdir(req, dentry, err);
390 if ((flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
391 err = ceph_handle_notrace_create(dir, dentry);
393 if (d_in_lookup(dentry)) {
394 dn = ceph_finish_lookup(req, dentry, err);
398 /* we were given a hashed negative dentry */
403 if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
404 /* make vfs retry on splice, ENOENT, or symlink */
405 dout("atomic_open finish_no_open on dn %p\n", dn);
406 err = finish_no_open(file, dn);
408 dout("atomic_open finish_open on dn %p\n", dn);
409 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
410 ceph_init_inode_acls(d_inode(dentry), &acls);
411 *opened |= FILE_CREATED;
413 err = finish_open(file, dentry, ceph_open, opened);
416 if (!req->r_err && req->r_target_inode)
417 ceph_put_fmode(ceph_inode(req->r_target_inode), req->r_fmode);
418 ceph_mdsc_put_request(req);
420 ceph_release_acls_info(&acls);
421 dout("atomic_open result=%d\n", err);
425 int ceph_release(struct inode *inode, struct file *file)
427 struct ceph_inode_info *ci = ceph_inode(inode);
428 struct ceph_file_info *cf = file->private_data;
430 dout("release inode %p file %p\n", inode, file);
431 ceph_put_fmode(ci, cf->fmode);
432 if (cf->last_readdir)
433 ceph_mdsc_put_request(cf->last_readdir);
434 kfree(cf->last_name);
436 kmem_cache_free(ceph_file_cachep, cf);
438 /* wake up anyone waiting for caps on this inode */
439 wake_up_all(&ci->i_cap_wq);
450 * Read a range of bytes striped over one or more objects. Iterate over
451 * objects we stripe over. (That's not atomic, but good enough for now.)
453 * If we get a short result from the OSD, check against i_size; we need to
454 * only return a short read to the caller if we hit EOF.
456 static int striped_read(struct inode *inode,
458 struct page **pages, int num_pages,
461 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
462 struct ceph_inode_info *ci = ceph_inode(inode);
463 u64 pos, this_len, left;
465 int page_align, pages_left;
467 struct page **page_pos;
468 bool hit_stripe, was_short;
471 * we may need to do multiple reads. not atomic, unfortunately.
476 pages_left = num_pages;
480 page_align = pos & ~PAGE_MASK;
482 ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode),
483 &ci->i_layout, pos, &this_len,
486 page_pos, pages_left, page_align);
489 hit_stripe = this_len < left;
490 was_short = ret >= 0 && ret < this_len;
491 dout("striped_read %llu~%llu (read %u) got %d%s%s\n", pos, left, read,
492 ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : "");
494 i_size = i_size_read(inode);
497 if (was_short && (pos + ret < i_size)) {
498 int zlen = min(this_len - ret, i_size - pos - ret);
499 int zoff = (off & ~PAGE_MASK) + read + ret;
500 dout(" zero gap %llu to %llu\n",
501 pos + ret, pos + ret + zlen);
502 ceph_zero_page_vector_range(zoff, zlen, pages);
506 didpages = (page_align + ret) >> PAGE_SHIFT;
510 page_pos += didpages;
511 pages_left -= didpages;
513 /* hit stripe and need continue*/
514 if (left && hit_stripe && pos < i_size)
520 /* did we bounce off eof? */
521 if (pos + left > i_size)
522 *checkeof = CHECK_EOF;
525 dout("striped_read returns %d\n", ret);
530 * Completely synchronous read and write methods. Direct from __user
531 * buffer to osd, or directly to user pages (if O_DIRECT).
533 * If the read spans object boundary, just do multiple reads.
535 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *i,
538 struct file *file = iocb->ki_filp;
539 struct inode *inode = file_inode(file);
541 u64 off = iocb->ki_pos;
543 size_t len = iov_iter_count(i);
545 dout("sync_read on file %p %llu~%u %s\n", file, off,
547 (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
552 * flush any page cache pages in this range. this
553 * will make concurrent normal and sync io slow,
554 * but it will at least behave sensibly when they are
557 ret = filemap_write_and_wait_range(inode->i_mapping, off,
562 num_pages = calc_pages_for(off, len);
563 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
565 return PTR_ERR(pages);
566 ret = striped_read(inode, off, len, pages,
567 num_pages, checkeof);
573 size_t page_off = off & ~PAGE_MASK;
574 size_t copy = min_t(size_t, left,
575 PAGE_SIZE - page_off);
576 l = copy_page_to_iter(pages[k++], page_off, copy, i);
583 ceph_release_page_vector(pages, num_pages);
585 if (off > iocb->ki_pos) {
586 ret = off - iocb->ki_pos;
590 dout("sync_read result %d\n", ret);
594 struct ceph_aio_request {
599 struct list_head osd_reqs;
601 atomic_t pending_reqs;
602 struct timespec mtime;
603 struct ceph_cap_flush *prealloc_cf;
606 struct ceph_aio_work {
607 struct work_struct work;
608 struct ceph_osd_request *req;
611 static void ceph_aio_retry_work(struct work_struct *work);
613 static void ceph_aio_complete(struct inode *inode,
614 struct ceph_aio_request *aio_req)
616 struct ceph_inode_info *ci = ceph_inode(inode);
619 if (!atomic_dec_and_test(&aio_req->pending_reqs))
622 ret = aio_req->error;
624 ret = aio_req->total_len;
626 dout("ceph_aio_complete %p rc %d\n", inode, ret);
628 if (ret >= 0 && aio_req->write) {
631 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
632 if (endoff > i_size_read(inode)) {
633 if (ceph_inode_set_size(inode, endoff))
634 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
637 spin_lock(&ci->i_ceph_lock);
638 ci->i_inline_version = CEPH_INLINE_NONE;
639 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
640 &aio_req->prealloc_cf);
641 spin_unlock(&ci->i_ceph_lock);
643 __mark_inode_dirty(inode, dirty);
647 ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
650 aio_req->iocb->ki_complete(aio_req->iocb, ret, 0);
652 ceph_free_cap_flush(aio_req->prealloc_cf);
656 static void ceph_aio_complete_req(struct ceph_osd_request *req)
658 int rc = req->r_result;
659 struct inode *inode = req->r_inode;
660 struct ceph_aio_request *aio_req = req->r_priv;
661 struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
662 int num_pages = calc_pages_for((u64)osd_data->alignment,
665 dout("ceph_aio_complete_req %p rc %d bytes %llu\n",
666 inode, rc, osd_data->length);
668 if (rc == -EOLDSNAPC) {
669 struct ceph_aio_work *aio_work;
670 BUG_ON(!aio_req->write);
672 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
674 INIT_WORK(&aio_work->work, ceph_aio_retry_work);
676 queue_work(ceph_inode_to_client(inode)->wb_wq,
681 } else if (!aio_req->write) {
684 if (rc >= 0 && osd_data->length > rc) {
685 int zoff = osd_data->alignment + rc;
686 int zlen = osd_data->length - rc;
688 * If read is satisfied by single OSD request,
689 * it can pass EOF. Otherwise read is within
692 if (aio_req->num_reqs == 1) {
693 loff_t i_size = i_size_read(inode);
694 loff_t endoff = aio_req->iocb->ki_pos + rc;
696 zlen = min_t(size_t, zlen,
698 aio_req->total_len = rc + zlen;
702 ceph_zero_page_vector_range(zoff, zlen,
707 ceph_put_page_vector(osd_data->pages, num_pages, !aio_req->write);
708 ceph_osdc_put_request(req);
711 cmpxchg(&aio_req->error, 0, rc);
713 ceph_aio_complete(inode, aio_req);
717 static void ceph_aio_retry_work(struct work_struct *work)
719 struct ceph_aio_work *aio_work =
720 container_of(work, struct ceph_aio_work, work);
721 struct ceph_osd_request *orig_req = aio_work->req;
722 struct ceph_aio_request *aio_req = orig_req->r_priv;
723 struct inode *inode = orig_req->r_inode;
724 struct ceph_inode_info *ci = ceph_inode(inode);
725 struct ceph_snap_context *snapc;
726 struct ceph_osd_request *req;
729 spin_lock(&ci->i_ceph_lock);
730 if (__ceph_have_pending_cap_snap(ci)) {
731 struct ceph_cap_snap *capsnap =
732 list_last_entry(&ci->i_cap_snaps,
733 struct ceph_cap_snap,
735 snapc = ceph_get_snap_context(capsnap->context);
737 BUG_ON(!ci->i_head_snapc);
738 snapc = ceph_get_snap_context(ci->i_head_snapc);
740 spin_unlock(&ci->i_ceph_lock);
742 req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 2,
750 req->r_flags = CEPH_OSD_FLAG_ORDERSNAP |
751 CEPH_OSD_FLAG_ONDISK |
753 ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
754 ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
756 ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
758 ceph_osdc_put_request(req);
763 req->r_ops[0] = orig_req->r_ops[0];
764 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC, 0);
766 req->r_mtime = aio_req->mtime;
767 req->r_data_offset = req->r_ops[0].extent.offset;
769 ceph_osdc_put_request(orig_req);
771 req->r_callback = ceph_aio_complete_req;
772 req->r_inode = inode;
773 req->r_priv = aio_req;
775 ret = ceph_osdc_start_request(req->r_osdc, req, false);
779 ceph_aio_complete_req(req);
782 ceph_put_snap_context(snapc);
787 * Write commit request unsafe callback, called to tell us when a
788 * request is unsafe (that is, in flight--has been handed to the
789 * messenger to send to its target osd). It is called again when
790 * we've received a response message indicating the request is
791 * "safe" (its CEPH_OSD_FLAG_ONDISK flag is set), or when a request
792 * is completed early (and unsuccessfully) due to a timeout or
795 * This is used if we requested both an ACK and ONDISK commit reply
798 static void ceph_sync_write_unsafe(struct ceph_osd_request *req, bool unsafe)
800 struct ceph_inode_info *ci = ceph_inode(req->r_inode);
802 dout("%s %p tid %llu %ssafe\n", __func__, req, req->r_tid,
805 ceph_get_cap_refs(ci, CEPH_CAP_FILE_WR);
806 spin_lock(&ci->i_unsafe_lock);
807 list_add_tail(&req->r_unsafe_item,
808 &ci->i_unsafe_writes);
809 spin_unlock(&ci->i_unsafe_lock);
811 complete_all(&req->r_completion);
813 spin_lock(&ci->i_unsafe_lock);
814 list_del_init(&req->r_unsafe_item);
815 spin_unlock(&ci->i_unsafe_lock);
816 ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR);
821 * Wait on any unsafe replies for the given inode. First wait on the
822 * newest request, and make that the upper bound. Then, if there are
823 * more requests, keep waiting on the oldest as long as it is still older
824 * than the original request.
826 void ceph_sync_write_wait(struct inode *inode)
828 struct ceph_inode_info *ci = ceph_inode(inode);
829 struct list_head *head = &ci->i_unsafe_writes;
830 struct ceph_osd_request *req;
833 if (!S_ISREG(inode->i_mode))
836 spin_lock(&ci->i_unsafe_lock);
837 if (list_empty(head))
840 /* set upper bound as _last_ entry in chain */
842 req = list_last_entry(head, struct ceph_osd_request,
844 last_tid = req->r_tid;
847 ceph_osdc_get_request(req);
848 spin_unlock(&ci->i_unsafe_lock);
850 dout("sync_write_wait on tid %llu (until %llu)\n",
851 req->r_tid, last_tid);
852 wait_for_completion(&req->r_safe_completion);
853 ceph_osdc_put_request(req);
855 spin_lock(&ci->i_unsafe_lock);
857 * from here on look at first entry in chain, since we
858 * only want to wait for anything older than last_tid
860 if (list_empty(head))
862 req = list_first_entry(head, struct ceph_osd_request,
864 } while (req->r_tid < last_tid);
866 spin_unlock(&ci->i_unsafe_lock);
870 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
871 struct ceph_snap_context *snapc,
872 struct ceph_cap_flush **pcf)
874 struct file *file = iocb->ki_filp;
875 struct inode *inode = file_inode(file);
876 struct ceph_inode_info *ci = ceph_inode(inode);
877 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
878 struct ceph_vino vino;
879 struct ceph_osd_request *req;
881 struct ceph_aio_request *aio_req = NULL;
885 struct timespec mtime = current_time(inode);
886 size_t count = iov_iter_count(iter);
887 loff_t pos = iocb->ki_pos;
888 bool write = iov_iter_rw(iter) == WRITE;
890 if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
893 dout("sync_direct_read_write (%s) on file %p %lld~%u\n",
894 (write ? "write" : "read"), file, pos, (unsigned)count);
896 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
901 int ret2 = invalidate_inode_pages2_range(inode->i_mapping,
903 (pos + count) >> PAGE_SHIFT);
905 dout("invalidate_inode_pages2_range returned %d\n", ret);
907 flags = CEPH_OSD_FLAG_ORDERSNAP |
908 CEPH_OSD_FLAG_ONDISK |
911 flags = CEPH_OSD_FLAG_READ;
914 while (iov_iter_count(iter) > 0) {
915 u64 size = dio_get_pagev_size(iter);
919 vino = ceph_vino(inode);
920 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
922 /*include a 'startsync' command*/
924 write ? CEPH_OSD_OP_WRITE :
936 pages = dio_get_pages_alloc(iter, len, &start, &num_pages);
938 ceph_osdc_put_request(req);
939 ret = PTR_ERR(pages);
944 * To simplify error handling, allow AIO when IO within i_size
945 * or IO can be satisfied by single OSD request.
947 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
948 (len == count || pos + count <= i_size_read(inode))) {
949 aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
951 aio_req->iocb = iocb;
952 aio_req->write = write;
953 INIT_LIST_HEAD(&aio_req->osd_reqs);
955 aio_req->mtime = mtime;
956 swap(aio_req->prealloc_cf, *pcf);
964 * throw out any page cache pages in this range. this
967 truncate_inode_pages_range(inode->i_mapping, pos,
968 (pos+len) | (PAGE_SIZE - 1));
970 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC, 0);
971 req->r_mtime = mtime;
974 osd_req_op_extent_osd_data_pages(req, 0, pages, len, start,
978 aio_req->total_len += len;
980 atomic_inc(&aio_req->pending_reqs);
982 req->r_callback = ceph_aio_complete_req;
983 req->r_inode = inode;
984 req->r_priv = aio_req;
985 list_add_tail(&req->r_unsafe_item, &aio_req->osd_reqs);
988 iov_iter_advance(iter, len);
992 ret = ceph_osdc_start_request(req->r_osdc, req, false);
994 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
996 size = i_size_read(inode);
1000 if (ret >= 0 && ret < len && pos + ret < size) {
1001 int zlen = min_t(size_t, len - ret,
1003 ceph_zero_page_vector_range(start + ret, zlen,
1011 ceph_put_page_vector(pages, num_pages, !write);
1013 ceph_osdc_put_request(req);
1018 iov_iter_advance(iter, len);
1020 if (!write && pos >= size)
1023 if (write && pos > size) {
1024 if (ceph_inode_set_size(inode, pos))
1025 ceph_check_caps(ceph_inode(inode),
1026 CHECK_CAPS_AUTHONLY,
1032 LIST_HEAD(osd_reqs);
1034 if (aio_req->num_reqs == 0) {
1039 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1042 list_splice(&aio_req->osd_reqs, &osd_reqs);
1043 while (!list_empty(&osd_reqs)) {
1044 req = list_first_entry(&osd_reqs,
1045 struct ceph_osd_request,
1047 list_del_init(&req->r_unsafe_item);
1049 ret = ceph_osdc_start_request(req->r_osdc,
1052 req->r_result = ret;
1053 ceph_aio_complete_req(req);
1056 return -EIOCBQUEUED;
1059 if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1060 ret = pos - iocb->ki_pos;
1067 * Synchronous write, straight from __user pointer or user pages.
1069 * If write spans object boundary, just do multiple writes. (For a
1070 * correct atomic write, we should e.g. take write locks on all
1071 * objects, rollback on failure, etc.)
1074 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1075 struct ceph_snap_context *snapc)
1077 struct file *file = iocb->ki_filp;
1078 struct inode *inode = file_inode(file);
1079 struct ceph_inode_info *ci = ceph_inode(inode);
1080 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1081 struct ceph_vino vino;
1082 struct ceph_osd_request *req;
1083 struct page **pages;
1090 struct timespec mtime = current_time(inode);
1091 size_t count = iov_iter_count(from);
1093 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1096 dout("sync_write on file %p %lld~%u\n", file, pos, (unsigned)count);
1098 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
1102 ret = invalidate_inode_pages2_range(inode->i_mapping,
1104 (pos + count) >> PAGE_SHIFT);
1106 dout("invalidate_inode_pages2_range returned %d\n", ret);
1108 flags = CEPH_OSD_FLAG_ORDERSNAP |
1109 CEPH_OSD_FLAG_ONDISK |
1110 CEPH_OSD_FLAG_WRITE |
1113 while ((len = iov_iter_count(from)) > 0) {
1117 vino = ceph_vino(inode);
1118 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1119 vino, pos, &len, 0, 1,
1120 CEPH_OSD_OP_WRITE, flags, snapc,
1122 ci->i_truncate_size,
1130 * write from beginning of first page,
1131 * regardless of io alignment
1133 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1135 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1136 if (IS_ERR(pages)) {
1137 ret = PTR_ERR(pages);
1142 for (n = 0; n < num_pages; n++) {
1143 size_t plen = min_t(size_t, left, PAGE_SIZE);
1144 ret = copy_page_from_iter(pages[n], 0, plen, from);
1153 ceph_release_page_vector(pages, num_pages);
1157 /* get a second commit callback */
1158 req->r_unsafe_callback = ceph_sync_write_unsafe;
1159 req->r_inode = inode;
1161 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1164 req->r_mtime = mtime;
1165 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1167 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1170 ceph_osdc_put_request(req);
1175 if (pos > i_size_read(inode)) {
1176 check_caps = ceph_inode_set_size(inode, pos);
1178 ceph_check_caps(ceph_inode(inode),
1179 CHECK_CAPS_AUTHONLY,
1186 if (ret != -EOLDSNAPC && written > 0) {
1194 * Wrap generic_file_aio_read with checks for cap bits on the inode.
1195 * Atomically grab references, so that those bits are not released
1196 * back to the MDS mid-read.
1198 * Hmm, the sync read case isn't actually async... should it be?
1200 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1202 struct file *filp = iocb->ki_filp;
1203 struct ceph_file_info *fi = filp->private_data;
1204 size_t len = iov_iter_count(to);
1205 struct inode *inode = file_inode(filp);
1206 struct ceph_inode_info *ci = ceph_inode(inode);
1207 struct page *pinned_page = NULL;
1210 int retry_op = 0, read = 0;
1213 dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1214 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1216 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1217 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1219 want = CEPH_CAP_FILE_CACHE;
1220 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1224 if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1225 (iocb->ki_flags & IOCB_DIRECT) ||
1226 (fi->flags & CEPH_F_SYNC)) {
1228 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1229 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1230 ceph_cap_string(got));
1232 if (ci->i_inline_version == CEPH_INLINE_NONE) {
1233 if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1234 ret = ceph_direct_read_write(iocb, to,
1236 if (ret >= 0 && ret < len)
1237 retry_op = CHECK_EOF;
1239 ret = ceph_sync_read(iocb, to, &retry_op);
1242 retry_op = READ_INLINE;
1245 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1246 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1247 ceph_cap_string(got));
1249 ret = generic_file_read_iter(iocb, to);
1251 dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1252 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1254 put_page(pinned_page);
1257 ceph_put_cap_refs(ci, got);
1258 if (retry_op > HAVE_RETRIED && ret >= 0) {
1260 struct page *page = NULL;
1262 if (retry_op == READ_INLINE) {
1263 page = __page_cache_alloc(GFP_KERNEL);
1268 statret = __ceph_do_getattr(inode, page,
1269 CEPH_STAT_CAP_INLINE_DATA, !!page);
1273 if (statret == -ENODATA) {
1274 BUG_ON(retry_op != READ_INLINE);
1280 i_size = i_size_read(inode);
1281 if (retry_op == READ_INLINE) {
1282 BUG_ON(ret > 0 || read > 0);
1283 if (iocb->ki_pos < i_size &&
1284 iocb->ki_pos < PAGE_SIZE) {
1285 loff_t end = min_t(loff_t, i_size,
1286 iocb->ki_pos + len);
1287 end = min_t(loff_t, end, PAGE_SIZE);
1289 zero_user_segment(page, statret, end);
1290 ret = copy_page_to_iter(page,
1291 iocb->ki_pos & ~PAGE_MASK,
1292 end - iocb->ki_pos, to);
1293 iocb->ki_pos += ret;
1296 if (iocb->ki_pos < i_size && read < len) {
1297 size_t zlen = min_t(size_t, len - read,
1298 i_size - iocb->ki_pos);
1299 ret = iov_iter_zero(zlen, to);
1300 iocb->ki_pos += ret;
1303 __free_pages(page, 0);
1307 /* hit EOF or hole? */
1308 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1310 dout("sync_read hit hole, ppos %lld < size %lld"
1311 ", reading more\n", iocb->ki_pos, i_size);
1315 retry_op = HAVE_RETRIED;
1327 * Take cap references to avoid releasing caps to MDS mid-write.
1329 * If we are synchronous, and write with an old snap context, the OSD
1330 * may return EOLDSNAPC. In that case, retry the write.. _after_
1331 * dropping our cap refs and allowing the pending snap to logically
1332 * complete _before_ this write occurs.
1334 * If we are near ENOSPC, write synchronously.
1336 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1338 struct file *file = iocb->ki_filp;
1339 struct ceph_file_info *fi = file->private_data;
1340 struct inode *inode = file_inode(file);
1341 struct ceph_inode_info *ci = ceph_inode(inode);
1342 struct ceph_osd_client *osdc =
1343 &ceph_sb_to_client(inode->i_sb)->client->osdc;
1344 struct ceph_cap_flush *prealloc_cf;
1345 ssize_t count, written = 0;
1349 if (ceph_snap(inode) != CEPH_NOSNAP)
1352 prealloc_cf = ceph_alloc_cap_flush();
1358 /* We can write back this queue in page reclaim */
1359 current->backing_dev_info = inode_to_bdi(inode);
1361 if (iocb->ki_flags & IOCB_APPEND) {
1362 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1367 err = generic_write_checks(iocb, from);
1372 count = iov_iter_count(from);
1373 err = file_remove_privs(file);
1377 err = file_update_time(file);
1381 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1382 err = ceph_uninline_data(file, NULL);
1388 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL)) {
1393 dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1394 inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1395 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1396 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1398 want = CEPH_CAP_FILE_BUFFER;
1400 err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, pos + count,
1405 dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1406 inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1408 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1409 (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC)) {
1410 struct ceph_snap_context *snapc;
1411 struct iov_iter data;
1412 inode_unlock(inode);
1414 spin_lock(&ci->i_ceph_lock);
1415 if (__ceph_have_pending_cap_snap(ci)) {
1416 struct ceph_cap_snap *capsnap =
1417 list_last_entry(&ci->i_cap_snaps,
1418 struct ceph_cap_snap,
1420 snapc = ceph_get_snap_context(capsnap->context);
1422 BUG_ON(!ci->i_head_snapc);
1423 snapc = ceph_get_snap_context(ci->i_head_snapc);
1425 spin_unlock(&ci->i_ceph_lock);
1427 /* we might need to revert back to that point */
1429 if (iocb->ki_flags & IOCB_DIRECT)
1430 written = ceph_direct_read_write(iocb, &data, snapc,
1433 written = ceph_sync_write(iocb, &data, pos, snapc);
1434 if (written == -EOLDSNAPC) {
1435 dout("aio_write %p %llx.%llx %llu~%u"
1436 "got EOLDSNAPC, retrying\n",
1437 inode, ceph_vinop(inode),
1438 pos, (unsigned)count);
1443 iov_iter_advance(from, written);
1444 ceph_put_snap_context(snapc);
1447 * No need to acquire the i_truncate_mutex. Because
1448 * the MDS revokes Fwb caps before sending truncate
1449 * message to us. We can't get Fwb cap while there
1450 * are pending vmtruncate. So write and vmtruncate
1451 * can not run at the same time
1453 written = generic_perform_write(file, from, pos);
1454 if (likely(written >= 0))
1455 iocb->ki_pos = pos + written;
1456 inode_unlock(inode);
1461 spin_lock(&ci->i_ceph_lock);
1462 ci->i_inline_version = CEPH_INLINE_NONE;
1463 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1465 spin_unlock(&ci->i_ceph_lock);
1467 __mark_inode_dirty(inode, dirty);
1470 dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n",
1471 inode, ceph_vinop(inode), pos, (unsigned)count,
1472 ceph_cap_string(got));
1473 ceph_put_cap_refs(ci, got);
1476 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_NEARFULL))
1477 iocb->ki_flags |= IOCB_DSYNC;
1479 written = generic_write_sync(iocb, written);
1485 inode_unlock(inode);
1487 ceph_free_cap_flush(prealloc_cf);
1488 current->backing_dev_info = NULL;
1489 return written ? written : err;
1493 * llseek. be sure to verify file size on SEEK_END.
1495 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1497 struct inode *inode = file->f_mapping->host;
1503 if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1504 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1509 i_size = i_size_read(inode);
1516 * Here we special-case the lseek(fd, 0, SEEK_CUR)
1517 * position-querying operation. Avoid rewriting the "same"
1518 * f_pos value back to the file because a concurrent read(),
1519 * write() or lseek() might have altered it
1525 offset += file->f_pos;
1528 if (offset >= i_size) {
1534 if (offset >= i_size) {
1542 ret = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1545 inode_unlock(inode);
1549 static inline void ceph_zero_partial_page(
1550 struct inode *inode, loff_t offset, unsigned size)
1553 pgoff_t index = offset >> PAGE_SHIFT;
1555 page = find_lock_page(inode->i_mapping, index);
1557 wait_on_page_writeback(page);
1558 zero_user(page, offset & (PAGE_SIZE - 1), size);
1564 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1567 loff_t nearly = round_up(offset, PAGE_SIZE);
1568 if (offset < nearly) {
1569 loff_t size = nearly - offset;
1572 ceph_zero_partial_page(inode, offset, size);
1576 if (length >= PAGE_SIZE) {
1577 loff_t size = round_down(length, PAGE_SIZE);
1578 truncate_pagecache_range(inode, offset, offset + size - 1);
1583 ceph_zero_partial_page(inode, offset, length);
1586 static int ceph_zero_partial_object(struct inode *inode,
1587 loff_t offset, loff_t *length)
1589 struct ceph_inode_info *ci = ceph_inode(inode);
1590 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1591 struct ceph_osd_request *req;
1597 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1600 op = CEPH_OSD_OP_ZERO;
1603 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1607 CEPH_OSD_FLAG_WRITE |
1608 CEPH_OSD_FLAG_ONDISK,
1615 req->r_mtime = inode->i_mtime;
1616 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1618 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1622 ceph_osdc_put_request(req);
1628 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
1631 struct ceph_inode_info *ci = ceph_inode(inode);
1632 s32 stripe_unit = ci->i_layout.stripe_unit;
1633 s32 stripe_count = ci->i_layout.stripe_count;
1634 s32 object_size = ci->i_layout.object_size;
1635 u64 object_set_size = object_size * stripe_count;
1638 /* round offset up to next period boundary */
1639 nearly = offset + object_set_size - 1;
1641 nearly -= do_div(t, object_set_size);
1643 while (length && offset < nearly) {
1644 loff_t size = length;
1645 ret = ceph_zero_partial_object(inode, offset, &size);
1651 while (length >= object_set_size) {
1653 loff_t pos = offset;
1654 for (i = 0; i < stripe_count; ++i) {
1655 ret = ceph_zero_partial_object(inode, pos, NULL);
1660 offset += object_set_size;
1661 length -= object_set_size;
1664 loff_t size = length;
1665 ret = ceph_zero_partial_object(inode, offset, &size);
1674 static long ceph_fallocate(struct file *file, int mode,
1675 loff_t offset, loff_t length)
1677 struct ceph_file_info *fi = file->private_data;
1678 struct inode *inode = file_inode(file);
1679 struct ceph_inode_info *ci = ceph_inode(inode);
1680 struct ceph_osd_client *osdc =
1681 &ceph_inode_to_client(inode)->client->osdc;
1682 struct ceph_cap_flush *prealloc_cf;
1689 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1692 if (!S_ISREG(inode->i_mode))
1695 prealloc_cf = ceph_alloc_cap_flush();
1701 if (ceph_snap(inode) != CEPH_NOSNAP) {
1706 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) &&
1707 !(mode & FALLOC_FL_PUNCH_HOLE)) {
1712 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1713 ret = ceph_uninline_data(file, NULL);
1718 size = i_size_read(inode);
1719 if (!(mode & FALLOC_FL_KEEP_SIZE))
1720 endoff = offset + length;
1722 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1723 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1725 want = CEPH_CAP_FILE_BUFFER;
1727 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, endoff, &got, NULL);
1731 if (mode & FALLOC_FL_PUNCH_HOLE) {
1733 ceph_zero_pagecache_range(inode, offset, length);
1734 ret = ceph_zero_objects(inode, offset, length);
1735 } else if (endoff > size) {
1736 truncate_pagecache_range(inode, size, -1);
1737 if (ceph_inode_set_size(inode, endoff))
1738 ceph_check_caps(ceph_inode(inode),
1739 CHECK_CAPS_AUTHONLY, NULL);
1743 spin_lock(&ci->i_ceph_lock);
1744 ci->i_inline_version = CEPH_INLINE_NONE;
1745 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1747 spin_unlock(&ci->i_ceph_lock);
1749 __mark_inode_dirty(inode, dirty);
1752 ceph_put_cap_refs(ci, got);
1754 inode_unlock(inode);
1755 ceph_free_cap_flush(prealloc_cf);
1759 const struct file_operations ceph_file_fops = {
1761 .release = ceph_release,
1762 .llseek = ceph_llseek,
1763 .read_iter = ceph_read_iter,
1764 .write_iter = ceph_write_iter,
1766 .fsync = ceph_fsync,
1768 .flock = ceph_flock,
1769 .splice_write = iter_file_splice_write,
1770 .unlocked_ioctl = ceph_ioctl,
1771 .compat_ioctl = ceph_ioctl,
1772 .fallocate = ceph_fallocate,