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 err = ceph_init_dentry(dentry);
358 if (flags & O_CREAT) {
359 err = ceph_pre_init_acls(dir, &mode, &acls);
365 req = prepare_open_request(dir->i_sb, flags, mode);
370 req->r_dentry = dget(dentry);
372 if (flags & O_CREAT) {
373 req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
374 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
376 req->r_pagelist = acls.pagelist;
377 acls.pagelist = NULL;
381 mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
382 if (ceph_security_xattr_wanted(dir))
383 mask |= CEPH_CAP_XATTR_SHARED;
384 req->r_args.open.mask = cpu_to_le32(mask);
386 req->r_locked_dir = dir; /* caller holds dir->i_mutex */
387 err = ceph_mdsc_do_request(mdsc,
388 (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
390 err = ceph_handle_snapdir(req, dentry, err);
394 if ((flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
395 err = ceph_handle_notrace_create(dir, dentry);
397 if (d_in_lookup(dentry)) {
398 dn = ceph_finish_lookup(req, dentry, err);
402 /* we were given a hashed negative dentry */
407 if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
408 /* make vfs retry on splice, ENOENT, or symlink */
409 dout("atomic_open finish_no_open on dn %p\n", dn);
410 err = finish_no_open(file, dn);
412 dout("atomic_open finish_open on dn %p\n", dn);
413 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
414 ceph_init_inode_acls(d_inode(dentry), &acls);
415 *opened |= FILE_CREATED;
417 err = finish_open(file, dentry, ceph_open, opened);
420 if (!req->r_err && req->r_target_inode)
421 ceph_put_fmode(ceph_inode(req->r_target_inode), req->r_fmode);
422 ceph_mdsc_put_request(req);
424 ceph_release_acls_info(&acls);
425 dout("atomic_open result=%d\n", err);
429 int ceph_release(struct inode *inode, struct file *file)
431 struct ceph_inode_info *ci = ceph_inode(inode);
432 struct ceph_file_info *cf = file->private_data;
434 dout("release inode %p file %p\n", inode, file);
435 ceph_put_fmode(ci, cf->fmode);
436 if (cf->last_readdir)
437 ceph_mdsc_put_request(cf->last_readdir);
438 kfree(cf->last_name);
440 kmem_cache_free(ceph_file_cachep, cf);
442 /* wake up anyone waiting for caps on this inode */
443 wake_up_all(&ci->i_cap_wq);
454 * Read a range of bytes striped over one or more objects. Iterate over
455 * objects we stripe over. (That's not atomic, but good enough for now.)
457 * If we get a short result from the OSD, check against i_size; we need to
458 * only return a short read to the caller if we hit EOF.
460 static int striped_read(struct inode *inode,
462 struct page **pages, int num_pages,
465 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
466 struct ceph_inode_info *ci = ceph_inode(inode);
467 u64 pos, this_len, left;
469 int page_align, pages_left;
471 struct page **page_pos;
472 bool hit_stripe, was_short;
475 * we may need to do multiple reads. not atomic, unfortunately.
480 pages_left = num_pages;
484 page_align = pos & ~PAGE_MASK;
486 ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode),
487 &ci->i_layout, pos, &this_len,
490 page_pos, pages_left, page_align);
493 hit_stripe = this_len < left;
494 was_short = ret >= 0 && ret < this_len;
495 dout("striped_read %llu~%llu (read %u) got %d%s%s\n", pos, left, read,
496 ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : "");
498 i_size = i_size_read(inode);
501 if (was_short && (pos + ret < i_size)) {
502 int zlen = min(this_len - ret, i_size - pos - ret);
503 int zoff = (off & ~PAGE_MASK) + read + ret;
504 dout(" zero gap %llu to %llu\n",
505 pos + ret, pos + ret + zlen);
506 ceph_zero_page_vector_range(zoff, zlen, pages);
510 didpages = (page_align + ret) >> PAGE_SHIFT;
514 page_pos += didpages;
515 pages_left -= didpages;
517 /* hit stripe and need continue*/
518 if (left && hit_stripe && pos < i_size)
524 /* did we bounce off eof? */
525 if (pos + left > i_size)
526 *checkeof = CHECK_EOF;
529 dout("striped_read returns %d\n", ret);
534 * Completely synchronous read and write methods. Direct from __user
535 * buffer to osd, or directly to user pages (if O_DIRECT).
537 * If the read spans object boundary, just do multiple reads.
539 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *i,
542 struct file *file = iocb->ki_filp;
543 struct inode *inode = file_inode(file);
545 u64 off = iocb->ki_pos;
547 size_t len = iov_iter_count(i);
549 dout("sync_read on file %p %llu~%u %s\n", file, off,
551 (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
556 * flush any page cache pages in this range. this
557 * will make concurrent normal and sync io slow,
558 * but it will at least behave sensibly when they are
561 ret = filemap_write_and_wait_range(inode->i_mapping, off,
566 num_pages = calc_pages_for(off, len);
567 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
569 return PTR_ERR(pages);
570 ret = striped_read(inode, off, len, pages,
571 num_pages, checkeof);
577 size_t page_off = off & ~PAGE_MASK;
578 size_t copy = min_t(size_t, left,
579 PAGE_SIZE - page_off);
580 l = copy_page_to_iter(pages[k++], page_off, copy, i);
587 ceph_release_page_vector(pages, num_pages);
589 if (off > iocb->ki_pos) {
590 ret = off - iocb->ki_pos;
594 dout("sync_read result %d\n", ret);
598 struct ceph_aio_request {
604 struct list_head osd_reqs;
606 atomic_t pending_reqs;
607 struct timespec mtime;
608 struct ceph_cap_flush *prealloc_cf;
611 struct ceph_aio_work {
612 struct work_struct work;
613 struct ceph_osd_request *req;
616 static void ceph_aio_retry_work(struct work_struct *work);
618 static void ceph_aio_complete(struct inode *inode,
619 struct ceph_aio_request *aio_req)
621 struct ceph_inode_info *ci = ceph_inode(inode);
624 if (!atomic_dec_and_test(&aio_req->pending_reqs))
627 ret = aio_req->error;
629 ret = aio_req->total_len;
631 dout("ceph_aio_complete %p rc %d\n", inode, ret);
633 if (ret >= 0 && aio_req->write) {
636 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
637 if (endoff > i_size_read(inode)) {
638 if (ceph_inode_set_size(inode, endoff))
639 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
642 spin_lock(&ci->i_ceph_lock);
643 ci->i_inline_version = CEPH_INLINE_NONE;
644 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
645 &aio_req->prealloc_cf);
646 spin_unlock(&ci->i_ceph_lock);
648 __mark_inode_dirty(inode, dirty);
652 ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
655 aio_req->iocb->ki_complete(aio_req->iocb, ret, 0);
657 ceph_free_cap_flush(aio_req->prealloc_cf);
661 static void ceph_aio_complete_req(struct ceph_osd_request *req)
663 int rc = req->r_result;
664 struct inode *inode = req->r_inode;
665 struct ceph_aio_request *aio_req = req->r_priv;
666 struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
667 int num_pages = calc_pages_for((u64)osd_data->alignment,
670 dout("ceph_aio_complete_req %p rc %d bytes %llu\n",
671 inode, rc, osd_data->length);
673 if (rc == -EOLDSNAPC) {
674 struct ceph_aio_work *aio_work;
675 BUG_ON(!aio_req->write);
677 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
679 INIT_WORK(&aio_work->work, ceph_aio_retry_work);
681 queue_work(ceph_inode_to_client(inode)->wb_wq,
686 } else if (!aio_req->write) {
689 if (rc >= 0 && osd_data->length > rc) {
690 int zoff = osd_data->alignment + rc;
691 int zlen = osd_data->length - rc;
693 * If read is satisfied by single OSD request,
694 * it can pass EOF. Otherwise read is within
697 if (aio_req->num_reqs == 1) {
698 loff_t i_size = i_size_read(inode);
699 loff_t endoff = aio_req->iocb->ki_pos + rc;
701 zlen = min_t(size_t, zlen,
703 aio_req->total_len = rc + zlen;
707 ceph_zero_page_vector_range(zoff, zlen,
712 ceph_put_page_vector(osd_data->pages, num_pages, aio_req->should_dirty);
713 ceph_osdc_put_request(req);
716 cmpxchg(&aio_req->error, 0, rc);
718 ceph_aio_complete(inode, aio_req);
722 static void ceph_aio_retry_work(struct work_struct *work)
724 struct ceph_aio_work *aio_work =
725 container_of(work, struct ceph_aio_work, work);
726 struct ceph_osd_request *orig_req = aio_work->req;
727 struct ceph_aio_request *aio_req = orig_req->r_priv;
728 struct inode *inode = orig_req->r_inode;
729 struct ceph_inode_info *ci = ceph_inode(inode);
730 struct ceph_snap_context *snapc;
731 struct ceph_osd_request *req;
734 spin_lock(&ci->i_ceph_lock);
735 if (__ceph_have_pending_cap_snap(ci)) {
736 struct ceph_cap_snap *capsnap =
737 list_last_entry(&ci->i_cap_snaps,
738 struct ceph_cap_snap,
740 snapc = ceph_get_snap_context(capsnap->context);
742 BUG_ON(!ci->i_head_snapc);
743 snapc = ceph_get_snap_context(ci->i_head_snapc);
745 spin_unlock(&ci->i_ceph_lock);
747 req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 2,
755 req->r_flags = CEPH_OSD_FLAG_ORDERSNAP |
756 CEPH_OSD_FLAG_ONDISK |
758 ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
759 ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
761 ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
763 ceph_osdc_put_request(req);
768 req->r_ops[0] = orig_req->r_ops[0];
769 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC, 0);
771 req->r_mtime = aio_req->mtime;
772 req->r_data_offset = req->r_ops[0].extent.offset;
774 ceph_osdc_put_request(orig_req);
776 req->r_callback = ceph_aio_complete_req;
777 req->r_inode = inode;
778 req->r_priv = aio_req;
780 ret = ceph_osdc_start_request(req->r_osdc, req, false);
784 ceph_aio_complete_req(req);
787 ceph_put_snap_context(snapc);
792 * Write commit request unsafe callback, called to tell us when a
793 * request is unsafe (that is, in flight--has been handed to the
794 * messenger to send to its target osd). It is called again when
795 * we've received a response message indicating the request is
796 * "safe" (its CEPH_OSD_FLAG_ONDISK flag is set), or when a request
797 * is completed early (and unsuccessfully) due to a timeout or
800 * This is used if we requested both an ACK and ONDISK commit reply
803 static void ceph_sync_write_unsafe(struct ceph_osd_request *req, bool unsafe)
805 struct ceph_inode_info *ci = ceph_inode(req->r_inode);
807 dout("%s %p tid %llu %ssafe\n", __func__, req, req->r_tid,
810 ceph_get_cap_refs(ci, CEPH_CAP_FILE_WR);
811 spin_lock(&ci->i_unsafe_lock);
812 list_add_tail(&req->r_unsafe_item,
813 &ci->i_unsafe_writes);
814 spin_unlock(&ci->i_unsafe_lock);
816 complete_all(&req->r_completion);
818 spin_lock(&ci->i_unsafe_lock);
819 list_del_init(&req->r_unsafe_item);
820 spin_unlock(&ci->i_unsafe_lock);
821 ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR);
826 * Wait on any unsafe replies for the given inode. First wait on the
827 * newest request, and make that the upper bound. Then, if there are
828 * more requests, keep waiting on the oldest as long as it is still older
829 * than the original request.
831 void ceph_sync_write_wait(struct inode *inode)
833 struct ceph_inode_info *ci = ceph_inode(inode);
834 struct list_head *head = &ci->i_unsafe_writes;
835 struct ceph_osd_request *req;
838 if (!S_ISREG(inode->i_mode))
841 spin_lock(&ci->i_unsafe_lock);
842 if (list_empty(head))
845 /* set upper bound as _last_ entry in chain */
847 req = list_last_entry(head, struct ceph_osd_request,
849 last_tid = req->r_tid;
852 ceph_osdc_get_request(req);
853 spin_unlock(&ci->i_unsafe_lock);
855 dout("sync_write_wait on tid %llu (until %llu)\n",
856 req->r_tid, last_tid);
857 wait_for_completion(&req->r_safe_completion);
858 ceph_osdc_put_request(req);
860 spin_lock(&ci->i_unsafe_lock);
862 * from here on look at first entry in chain, since we
863 * only want to wait for anything older than last_tid
865 if (list_empty(head))
867 req = list_first_entry(head, struct ceph_osd_request,
869 } while (req->r_tid < last_tid);
871 spin_unlock(&ci->i_unsafe_lock);
875 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
876 struct ceph_snap_context *snapc,
877 struct ceph_cap_flush **pcf)
879 struct file *file = iocb->ki_filp;
880 struct inode *inode = file_inode(file);
881 struct ceph_inode_info *ci = ceph_inode(inode);
882 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
883 struct ceph_vino vino;
884 struct ceph_osd_request *req;
886 struct ceph_aio_request *aio_req = NULL;
890 struct timespec mtime = current_time(inode);
891 size_t count = iov_iter_count(iter);
892 loff_t pos = iocb->ki_pos;
893 bool write = iov_iter_rw(iter) == WRITE;
894 bool should_dirty = !write && iter_is_iovec(iter);
896 if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
899 dout("sync_direct_read_write (%s) on file %p %lld~%u\n",
900 (write ? "write" : "read"), file, pos, (unsigned)count);
902 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
907 int ret2 = invalidate_inode_pages2_range(inode->i_mapping,
909 (pos + count) >> PAGE_SHIFT);
911 dout("invalidate_inode_pages2_range returned %d\n", ret);
913 flags = CEPH_OSD_FLAG_ORDERSNAP |
914 CEPH_OSD_FLAG_ONDISK |
917 flags = CEPH_OSD_FLAG_READ;
920 while (iov_iter_count(iter) > 0) {
921 u64 size = dio_get_pagev_size(iter);
925 vino = ceph_vino(inode);
926 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
928 /*include a 'startsync' command*/
930 write ? CEPH_OSD_OP_WRITE :
942 pages = dio_get_pages_alloc(iter, len, &start, &num_pages);
944 ceph_osdc_put_request(req);
945 ret = PTR_ERR(pages);
950 * To simplify error handling, allow AIO when IO within i_size
951 * or IO can be satisfied by single OSD request.
953 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
954 (len == count || pos + count <= i_size_read(inode))) {
955 aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
957 aio_req->iocb = iocb;
958 aio_req->write = write;
959 aio_req->should_dirty = should_dirty;
960 INIT_LIST_HEAD(&aio_req->osd_reqs);
962 aio_req->mtime = mtime;
963 swap(aio_req->prealloc_cf, *pcf);
971 * throw out any page cache pages in this range. this
974 truncate_inode_pages_range(inode->i_mapping, pos,
975 (pos+len) | (PAGE_SIZE - 1));
977 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC, 0);
978 req->r_mtime = mtime;
981 osd_req_op_extent_osd_data_pages(req, 0, pages, len, start,
985 aio_req->total_len += len;
987 atomic_inc(&aio_req->pending_reqs);
989 req->r_callback = ceph_aio_complete_req;
990 req->r_inode = inode;
991 req->r_priv = aio_req;
992 list_add_tail(&req->r_unsafe_item, &aio_req->osd_reqs);
995 iov_iter_advance(iter, len);
999 ret = ceph_osdc_start_request(req->r_osdc, req, false);
1001 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1003 size = i_size_read(inode);
1007 if (ret >= 0 && ret < len && pos + ret < size) {
1008 int zlen = min_t(size_t, len - ret,
1010 ceph_zero_page_vector_range(start + ret, zlen,
1018 ceph_put_page_vector(pages, num_pages, should_dirty);
1020 ceph_osdc_put_request(req);
1025 iov_iter_advance(iter, len);
1027 if (!write && pos >= size)
1030 if (write && pos > size) {
1031 if (ceph_inode_set_size(inode, pos))
1032 ceph_check_caps(ceph_inode(inode),
1033 CHECK_CAPS_AUTHONLY,
1039 LIST_HEAD(osd_reqs);
1041 if (aio_req->num_reqs == 0) {
1046 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1049 list_splice(&aio_req->osd_reqs, &osd_reqs);
1050 while (!list_empty(&osd_reqs)) {
1051 req = list_first_entry(&osd_reqs,
1052 struct ceph_osd_request,
1054 list_del_init(&req->r_unsafe_item);
1056 ret = ceph_osdc_start_request(req->r_osdc,
1059 req->r_result = ret;
1060 ceph_aio_complete_req(req);
1063 return -EIOCBQUEUED;
1066 if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1067 ret = pos - iocb->ki_pos;
1074 * Synchronous write, straight from __user pointer or user pages.
1076 * If write spans object boundary, just do multiple writes. (For a
1077 * correct atomic write, we should e.g. take write locks on all
1078 * objects, rollback on failure, etc.)
1081 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1082 struct ceph_snap_context *snapc)
1084 struct file *file = iocb->ki_filp;
1085 struct inode *inode = file_inode(file);
1086 struct ceph_inode_info *ci = ceph_inode(inode);
1087 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1088 struct ceph_vino vino;
1089 struct ceph_osd_request *req;
1090 struct page **pages;
1097 struct timespec mtime = current_time(inode);
1098 size_t count = iov_iter_count(from);
1100 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1103 dout("sync_write on file %p %lld~%u\n", file, pos, (unsigned)count);
1105 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
1109 ret = invalidate_inode_pages2_range(inode->i_mapping,
1111 (pos + count) >> PAGE_SHIFT);
1113 dout("invalidate_inode_pages2_range returned %d\n", ret);
1115 flags = CEPH_OSD_FLAG_ORDERSNAP |
1116 CEPH_OSD_FLAG_ONDISK |
1117 CEPH_OSD_FLAG_WRITE |
1120 while ((len = iov_iter_count(from)) > 0) {
1124 vino = ceph_vino(inode);
1125 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1126 vino, pos, &len, 0, 1,
1127 CEPH_OSD_OP_WRITE, flags, snapc,
1129 ci->i_truncate_size,
1137 * write from beginning of first page,
1138 * regardless of io alignment
1140 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1142 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1143 if (IS_ERR(pages)) {
1144 ret = PTR_ERR(pages);
1149 for (n = 0; n < num_pages; n++) {
1150 size_t plen = min_t(size_t, left, PAGE_SIZE);
1151 ret = copy_page_from_iter(pages[n], 0, plen, from);
1160 ceph_release_page_vector(pages, num_pages);
1164 /* get a second commit callback */
1165 req->r_unsafe_callback = ceph_sync_write_unsafe;
1166 req->r_inode = inode;
1168 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1171 req->r_mtime = mtime;
1172 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1174 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1177 ceph_osdc_put_request(req);
1182 if (pos > i_size_read(inode)) {
1183 check_caps = ceph_inode_set_size(inode, pos);
1185 ceph_check_caps(ceph_inode(inode),
1186 CHECK_CAPS_AUTHONLY,
1193 if (ret != -EOLDSNAPC && written > 0) {
1201 * Wrap generic_file_aio_read with checks for cap bits on the inode.
1202 * Atomically grab references, so that those bits are not released
1203 * back to the MDS mid-read.
1205 * Hmm, the sync read case isn't actually async... should it be?
1207 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1209 struct file *filp = iocb->ki_filp;
1210 struct ceph_file_info *fi = filp->private_data;
1211 size_t len = iov_iter_count(to);
1212 struct inode *inode = file_inode(filp);
1213 struct ceph_inode_info *ci = ceph_inode(inode);
1214 struct page *pinned_page = NULL;
1217 int retry_op = 0, read = 0;
1220 dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1221 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1223 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1224 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1226 want = CEPH_CAP_FILE_CACHE;
1227 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1231 if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1232 (iocb->ki_flags & IOCB_DIRECT) ||
1233 (fi->flags & CEPH_F_SYNC)) {
1235 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1236 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1237 ceph_cap_string(got));
1239 if (ci->i_inline_version == CEPH_INLINE_NONE) {
1240 if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1241 ret = ceph_direct_read_write(iocb, to,
1243 if (ret >= 0 && ret < len)
1244 retry_op = CHECK_EOF;
1246 ret = ceph_sync_read(iocb, to, &retry_op);
1249 retry_op = READ_INLINE;
1252 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1253 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1254 ceph_cap_string(got));
1255 current->journal_info = filp;
1256 ret = generic_file_read_iter(iocb, to);
1257 current->journal_info = NULL;
1259 dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1260 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1262 put_page(pinned_page);
1265 ceph_put_cap_refs(ci, got);
1266 if (retry_op > HAVE_RETRIED && ret >= 0) {
1268 struct page *page = NULL;
1270 if (retry_op == READ_INLINE) {
1271 page = __page_cache_alloc(GFP_KERNEL);
1276 statret = __ceph_do_getattr(inode, page,
1277 CEPH_STAT_CAP_INLINE_DATA, !!page);
1281 if (statret == -ENODATA) {
1282 BUG_ON(retry_op != READ_INLINE);
1288 i_size = i_size_read(inode);
1289 if (retry_op == READ_INLINE) {
1290 BUG_ON(ret > 0 || read > 0);
1291 if (iocb->ki_pos < i_size &&
1292 iocb->ki_pos < PAGE_SIZE) {
1293 loff_t end = min_t(loff_t, i_size,
1294 iocb->ki_pos + len);
1295 end = min_t(loff_t, end, PAGE_SIZE);
1297 zero_user_segment(page, statret, end);
1298 ret = copy_page_to_iter(page,
1299 iocb->ki_pos & ~PAGE_MASK,
1300 end - iocb->ki_pos, to);
1301 iocb->ki_pos += ret;
1304 if (iocb->ki_pos < i_size && read < len) {
1305 size_t zlen = min_t(size_t, len - read,
1306 i_size - iocb->ki_pos);
1307 ret = iov_iter_zero(zlen, to);
1308 iocb->ki_pos += ret;
1311 __free_pages(page, 0);
1315 /* hit EOF or hole? */
1316 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1318 dout("sync_read hit hole, ppos %lld < size %lld"
1319 ", reading more\n", iocb->ki_pos, i_size);
1323 retry_op = HAVE_RETRIED;
1335 * Take cap references to avoid releasing caps to MDS mid-write.
1337 * If we are synchronous, and write with an old snap context, the OSD
1338 * may return EOLDSNAPC. In that case, retry the write.. _after_
1339 * dropping our cap refs and allowing the pending snap to logically
1340 * complete _before_ this write occurs.
1342 * If we are near ENOSPC, write synchronously.
1344 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1346 struct file *file = iocb->ki_filp;
1347 struct ceph_file_info *fi = file->private_data;
1348 struct inode *inode = file_inode(file);
1349 struct ceph_inode_info *ci = ceph_inode(inode);
1350 struct ceph_osd_client *osdc =
1351 &ceph_sb_to_client(inode->i_sb)->client->osdc;
1352 struct ceph_cap_flush *prealloc_cf;
1353 ssize_t count, written = 0;
1357 if (ceph_snap(inode) != CEPH_NOSNAP)
1360 prealloc_cf = ceph_alloc_cap_flush();
1366 /* We can write back this queue in page reclaim */
1367 current->backing_dev_info = inode_to_bdi(inode);
1369 if (iocb->ki_flags & IOCB_APPEND) {
1370 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1375 err = generic_write_checks(iocb, from);
1380 count = iov_iter_count(from);
1381 err = file_remove_privs(file);
1385 err = file_update_time(file);
1389 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1390 err = ceph_uninline_data(file, NULL);
1396 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL)) {
1401 dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1402 inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1403 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1404 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1406 want = CEPH_CAP_FILE_BUFFER;
1408 err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, pos + count,
1413 dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1414 inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1416 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1417 (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC)) {
1418 struct ceph_snap_context *snapc;
1419 struct iov_iter data;
1420 inode_unlock(inode);
1422 spin_lock(&ci->i_ceph_lock);
1423 if (__ceph_have_pending_cap_snap(ci)) {
1424 struct ceph_cap_snap *capsnap =
1425 list_last_entry(&ci->i_cap_snaps,
1426 struct ceph_cap_snap,
1428 snapc = ceph_get_snap_context(capsnap->context);
1430 BUG_ON(!ci->i_head_snapc);
1431 snapc = ceph_get_snap_context(ci->i_head_snapc);
1433 spin_unlock(&ci->i_ceph_lock);
1435 /* we might need to revert back to that point */
1437 if (iocb->ki_flags & IOCB_DIRECT)
1438 written = ceph_direct_read_write(iocb, &data, snapc,
1441 written = ceph_sync_write(iocb, &data, pos, snapc);
1442 if (written == -EOLDSNAPC) {
1443 dout("aio_write %p %llx.%llx %llu~%u"
1444 "got EOLDSNAPC, retrying\n",
1445 inode, ceph_vinop(inode),
1446 pos, (unsigned)count);
1451 iov_iter_advance(from, written);
1452 ceph_put_snap_context(snapc);
1455 * No need to acquire the i_truncate_mutex. Because
1456 * the MDS revokes Fwb caps before sending truncate
1457 * message to us. We can't get Fwb cap while there
1458 * are pending vmtruncate. So write and vmtruncate
1459 * can not run at the same time
1461 written = generic_perform_write(file, from, pos);
1462 if (likely(written >= 0))
1463 iocb->ki_pos = pos + written;
1464 inode_unlock(inode);
1469 spin_lock(&ci->i_ceph_lock);
1470 ci->i_inline_version = CEPH_INLINE_NONE;
1471 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1473 spin_unlock(&ci->i_ceph_lock);
1475 __mark_inode_dirty(inode, dirty);
1478 dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n",
1479 inode, ceph_vinop(inode), pos, (unsigned)count,
1480 ceph_cap_string(got));
1481 ceph_put_cap_refs(ci, got);
1484 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_NEARFULL))
1485 iocb->ki_flags |= IOCB_DSYNC;
1487 written = generic_write_sync(iocb, written);
1493 inode_unlock(inode);
1495 ceph_free_cap_flush(prealloc_cf);
1496 current->backing_dev_info = NULL;
1497 return written ? written : err;
1501 * llseek. be sure to verify file size on SEEK_END.
1503 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1505 struct inode *inode = file->f_mapping->host;
1511 if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1512 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1517 i_size = i_size_read(inode);
1524 * Here we special-case the lseek(fd, 0, SEEK_CUR)
1525 * position-querying operation. Avoid rewriting the "same"
1526 * f_pos value back to the file because a concurrent read(),
1527 * write() or lseek() might have altered it
1533 offset += file->f_pos;
1536 if (offset >= i_size) {
1542 if (offset >= i_size) {
1550 ret = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1553 inode_unlock(inode);
1557 static inline void ceph_zero_partial_page(
1558 struct inode *inode, loff_t offset, unsigned size)
1561 pgoff_t index = offset >> PAGE_SHIFT;
1563 page = find_lock_page(inode->i_mapping, index);
1565 wait_on_page_writeback(page);
1566 zero_user(page, offset & (PAGE_SIZE - 1), size);
1572 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1575 loff_t nearly = round_up(offset, PAGE_SIZE);
1576 if (offset < nearly) {
1577 loff_t size = nearly - offset;
1580 ceph_zero_partial_page(inode, offset, size);
1584 if (length >= PAGE_SIZE) {
1585 loff_t size = round_down(length, PAGE_SIZE);
1586 truncate_pagecache_range(inode, offset, offset + size - 1);
1591 ceph_zero_partial_page(inode, offset, length);
1594 static int ceph_zero_partial_object(struct inode *inode,
1595 loff_t offset, loff_t *length)
1597 struct ceph_inode_info *ci = ceph_inode(inode);
1598 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1599 struct ceph_osd_request *req;
1605 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1608 op = CEPH_OSD_OP_ZERO;
1611 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1615 CEPH_OSD_FLAG_WRITE |
1616 CEPH_OSD_FLAG_ONDISK,
1623 req->r_mtime = inode->i_mtime;
1624 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1626 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1630 ceph_osdc_put_request(req);
1636 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
1639 struct ceph_inode_info *ci = ceph_inode(inode);
1640 s32 stripe_unit = ci->i_layout.stripe_unit;
1641 s32 stripe_count = ci->i_layout.stripe_count;
1642 s32 object_size = ci->i_layout.object_size;
1643 u64 object_set_size = object_size * stripe_count;
1646 /* round offset up to next period boundary */
1647 nearly = offset + object_set_size - 1;
1649 nearly -= do_div(t, object_set_size);
1651 while (length && offset < nearly) {
1652 loff_t size = length;
1653 ret = ceph_zero_partial_object(inode, offset, &size);
1659 while (length >= object_set_size) {
1661 loff_t pos = offset;
1662 for (i = 0; i < stripe_count; ++i) {
1663 ret = ceph_zero_partial_object(inode, pos, NULL);
1668 offset += object_set_size;
1669 length -= object_set_size;
1672 loff_t size = length;
1673 ret = ceph_zero_partial_object(inode, offset, &size);
1682 static long ceph_fallocate(struct file *file, int mode,
1683 loff_t offset, loff_t length)
1685 struct ceph_file_info *fi = file->private_data;
1686 struct inode *inode = file_inode(file);
1687 struct ceph_inode_info *ci = ceph_inode(inode);
1688 struct ceph_osd_client *osdc =
1689 &ceph_inode_to_client(inode)->client->osdc;
1690 struct ceph_cap_flush *prealloc_cf;
1697 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1700 if (!S_ISREG(inode->i_mode))
1703 prealloc_cf = ceph_alloc_cap_flush();
1709 if (ceph_snap(inode) != CEPH_NOSNAP) {
1714 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) &&
1715 !(mode & FALLOC_FL_PUNCH_HOLE)) {
1720 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1721 ret = ceph_uninline_data(file, NULL);
1726 size = i_size_read(inode);
1727 if (!(mode & FALLOC_FL_KEEP_SIZE))
1728 endoff = offset + length;
1730 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1731 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1733 want = CEPH_CAP_FILE_BUFFER;
1735 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, endoff, &got, NULL);
1739 if (mode & FALLOC_FL_PUNCH_HOLE) {
1741 ceph_zero_pagecache_range(inode, offset, length);
1742 ret = ceph_zero_objects(inode, offset, length);
1743 } else if (endoff > size) {
1744 truncate_pagecache_range(inode, size, -1);
1745 if (ceph_inode_set_size(inode, endoff))
1746 ceph_check_caps(ceph_inode(inode),
1747 CHECK_CAPS_AUTHONLY, NULL);
1751 spin_lock(&ci->i_ceph_lock);
1752 ci->i_inline_version = CEPH_INLINE_NONE;
1753 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1755 spin_unlock(&ci->i_ceph_lock);
1757 __mark_inode_dirty(inode, dirty);
1760 ceph_put_cap_refs(ci, got);
1762 inode_unlock(inode);
1763 ceph_free_cap_flush(prealloc_cf);
1767 const struct file_operations ceph_file_fops = {
1769 .release = ceph_release,
1770 .llseek = ceph_llseek,
1771 .read_iter = ceph_read_iter,
1772 .write_iter = ceph_write_iter,
1774 .fsync = ceph_fsync,
1776 .setlease = simple_nosetlease,
1777 .flock = ceph_flock,
1778 .splice_write = iter_file_splice_write,
1779 .unlocked_ioctl = ceph_ioctl,
1780 .compat_ioctl = ceph_ioctl,
1781 .fallocate = ceph_fallocate,