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"
16 static __le32 ceph_flags_sys2wire(u32 flags)
20 switch (flags & O_ACCMODE) {
22 wire_flags |= CEPH_O_RDONLY;
25 wire_flags |= CEPH_O_WRONLY;
28 wire_flags |= CEPH_O_RDWR;
32 #define ceph_sys2wire(a) if (flags & a) { wire_flags |= CEPH_##a; flags &= ~a; }
34 ceph_sys2wire(O_CREAT);
35 ceph_sys2wire(O_EXCL);
36 ceph_sys2wire(O_TRUNC);
37 ceph_sys2wire(O_DIRECTORY);
38 ceph_sys2wire(O_NOFOLLOW);
43 dout("unused open flags: %x", flags);
45 return cpu_to_le32(wire_flags);
49 * Ceph file operations
51 * Implement basic open/close functionality, and implement
54 * We implement three modes of file I/O:
55 * - buffered uses the generic_file_aio_{read,write} helpers
57 * - synchronous is used when there is multi-client read/write
58 * sharing, avoids the page cache, and synchronously waits for an
61 * - direct io takes the variant of the sync path that references
62 * user pages directly.
64 * fsync() flushes and waits on dirty pages, but just queues metadata
65 * for writeback: since the MDS can recover size and mtime there is no
66 * need to wait for MDS acknowledgement.
70 * Calculate the length sum of direct io vectors that can
71 * be combined into one page vector.
73 static size_t dio_get_pagev_size(const struct iov_iter *it)
75 const struct iovec *iov = it->iov;
76 const struct iovec *iovend = iov + it->nr_segs;
79 size = iov->iov_len - it->iov_offset;
81 * An iov can be page vectored when both the current tail
82 * and the next base are page aligned.
84 while (PAGE_ALIGNED((iov->iov_base + iov->iov_len)) &&
85 (++iov < iovend && PAGE_ALIGNED((iov->iov_base)))) {
88 dout("dio_get_pagevlen len = %zu\n", size);
93 * Allocate a page vector based on (@it, @nbytes).
94 * The return value is the tuple describing a page vector,
95 * that is (@pages, @page_align, @num_pages).
98 dio_get_pages_alloc(const struct iov_iter *it, size_t nbytes,
99 size_t *page_align, int *num_pages)
101 struct iov_iter tmp_it = *it;
104 int ret = 0, idx, npages;
106 align = (unsigned long)(it->iov->iov_base + it->iov_offset) &
108 npages = calc_pages_for(align, nbytes);
109 pages = kvmalloc(sizeof(*pages) * npages, GFP_KERNEL);
111 return ERR_PTR(-ENOMEM);
113 for (idx = 0; idx < npages; ) {
115 ret = iov_iter_get_pages(&tmp_it, pages + idx, nbytes,
116 npages - idx, &start);
120 iov_iter_advance(&tmp_it, ret);
122 idx += (ret + start + PAGE_SIZE - 1) / PAGE_SIZE;
128 dout("dio_get_pages_alloc: got %d pages align %zu\n", npages, align);
131 ceph_put_page_vector(pages, idx, false);
136 * Prepare an open request. Preallocate ceph_cap to avoid an
137 * inopportune ENOMEM later.
139 static struct ceph_mds_request *
140 prepare_open_request(struct super_block *sb, int flags, int create_mode)
142 struct ceph_fs_client *fsc = ceph_sb_to_client(sb);
143 struct ceph_mds_client *mdsc = fsc->mdsc;
144 struct ceph_mds_request *req;
145 int want_auth = USE_ANY_MDS;
146 int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
148 if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
149 want_auth = USE_AUTH_MDS;
151 req = ceph_mdsc_create_request(mdsc, op, want_auth);
154 req->r_fmode = ceph_flags_to_mode(flags);
155 req->r_args.open.flags = ceph_flags_sys2wire(flags);
156 req->r_args.open.mode = cpu_to_le32(create_mode);
162 * initialize private struct file data.
163 * if we fail, clean up by dropping fmode reference on the ceph_inode
165 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
167 struct ceph_file_info *cf;
170 switch (inode->i_mode & S_IFMT) {
172 ceph_fscache_register_inode_cookie(inode);
173 ceph_fscache_file_set_cookie(inode, file);
175 dout("init_file %p %p 0%o (regular)\n", inode, file,
177 cf = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
179 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
184 cf->readdir_cache_idx = -1;
185 file->private_data = cf;
186 BUG_ON(inode->i_fop->release != ceph_release);
190 dout("init_file %p %p 0%o (symlink)\n", inode, file,
192 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
196 dout("init_file %p %p 0%o (special)\n", inode, file,
199 * we need to drop the open ref now, since we don't
200 * have .release set to ceph_release.
202 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
203 BUG_ON(inode->i_fop->release == ceph_release);
205 /* call the proper open fop */
206 ret = inode->i_fop->open(inode, file);
212 * try renew caps after session gets killed.
214 int ceph_renew_caps(struct inode *inode)
216 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
217 struct ceph_inode_info *ci = ceph_inode(inode);
218 struct ceph_mds_request *req;
219 int err, flags, wanted;
221 spin_lock(&ci->i_ceph_lock);
222 wanted = __ceph_caps_file_wanted(ci);
223 if (__ceph_is_any_real_caps(ci) &&
224 (!(wanted & CEPH_CAP_ANY_WR) || ci->i_auth_cap)) {
225 int issued = __ceph_caps_issued(ci, NULL);
226 spin_unlock(&ci->i_ceph_lock);
227 dout("renew caps %p want %s issued %s updating mds_wanted\n",
228 inode, ceph_cap_string(wanted), ceph_cap_string(issued));
229 ceph_check_caps(ci, 0, NULL);
232 spin_unlock(&ci->i_ceph_lock);
235 if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
237 else if (wanted & CEPH_CAP_FILE_RD)
239 else if (wanted & CEPH_CAP_FILE_WR)
242 if (wanted & CEPH_CAP_FILE_LAZYIO)
246 req = prepare_open_request(inode->i_sb, flags, 0);
252 req->r_inode = inode;
257 err = ceph_mdsc_do_request(mdsc, NULL, req);
258 ceph_mdsc_put_request(req);
260 dout("renew caps %p open result=%d\n", inode, err);
261 return err < 0 ? err : 0;
265 * If we already have the requisite capabilities, we can satisfy
266 * the open request locally (no need to request new caps from the
267 * MDS). We do, however, need to inform the MDS (asynchronously)
268 * if our wanted caps set expands.
270 int ceph_open(struct inode *inode, struct file *file)
272 struct ceph_inode_info *ci = ceph_inode(inode);
273 struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
274 struct ceph_mds_client *mdsc = fsc->mdsc;
275 struct ceph_mds_request *req;
276 struct ceph_file_info *cf = file->private_data;
278 int flags, fmode, wanted;
281 dout("open file %p is already opened\n", file);
285 /* filter out O_CREAT|O_EXCL; vfs did that already. yuck. */
286 flags = file->f_flags & ~(O_CREAT|O_EXCL);
287 if (S_ISDIR(inode->i_mode))
288 flags = O_DIRECTORY; /* mds likes to know */
290 dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
291 ceph_vinop(inode), file, flags, file->f_flags);
292 fmode = ceph_flags_to_mode(flags);
293 wanted = ceph_caps_for_mode(fmode);
295 /* snapped files are read-only */
296 if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
299 /* trivially open snapdir */
300 if (ceph_snap(inode) == CEPH_SNAPDIR) {
301 spin_lock(&ci->i_ceph_lock);
302 __ceph_get_fmode(ci, fmode);
303 spin_unlock(&ci->i_ceph_lock);
304 return ceph_init_file(inode, file, fmode);
308 * No need to block if we have caps on the auth MDS (for
309 * write) or any MDS (for read). Update wanted set
312 spin_lock(&ci->i_ceph_lock);
313 if (__ceph_is_any_real_caps(ci) &&
314 (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
315 int mds_wanted = __ceph_caps_mds_wanted(ci, true);
316 int issued = __ceph_caps_issued(ci, NULL);
318 dout("open %p fmode %d want %s issued %s using existing\n",
319 inode, fmode, ceph_cap_string(wanted),
320 ceph_cap_string(issued));
321 __ceph_get_fmode(ci, fmode);
322 spin_unlock(&ci->i_ceph_lock);
325 if ((issued & wanted) != wanted &&
326 (mds_wanted & wanted) != wanted &&
327 ceph_snap(inode) != CEPH_SNAPDIR)
328 ceph_check_caps(ci, 0, NULL);
330 return ceph_init_file(inode, file, fmode);
331 } else if (ceph_snap(inode) != CEPH_NOSNAP &&
332 (ci->i_snap_caps & wanted) == wanted) {
333 __ceph_get_fmode(ci, fmode);
334 spin_unlock(&ci->i_ceph_lock);
335 return ceph_init_file(inode, file, fmode);
338 spin_unlock(&ci->i_ceph_lock);
340 dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
341 req = prepare_open_request(inode->i_sb, flags, 0);
346 req->r_inode = inode;
350 err = ceph_mdsc_do_request(mdsc, NULL, req);
352 err = ceph_init_file(inode, file, req->r_fmode);
353 ceph_mdsc_put_request(req);
354 dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
361 * Do a lookup + open with a single request. If we get a non-existent
362 * file or symlink, return 1 so the VFS can retry.
364 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
365 struct file *file, unsigned flags, umode_t mode,
368 struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
369 struct ceph_mds_client *mdsc = fsc->mdsc;
370 struct ceph_mds_request *req;
372 struct ceph_acls_info acls = {};
376 dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
378 d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
380 if (dentry->d_name.len > NAME_MAX)
381 return -ENAMETOOLONG;
383 if (flags & O_CREAT) {
384 err = ceph_pre_init_acls(dir, &mode, &acls);
390 req = prepare_open_request(dir->i_sb, flags, mode);
395 req->r_dentry = dget(dentry);
397 if (flags & O_CREAT) {
398 req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
399 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
401 req->r_pagelist = acls.pagelist;
402 acls.pagelist = NULL;
406 mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
407 if (ceph_security_xattr_wanted(dir))
408 mask |= CEPH_CAP_XATTR_SHARED;
409 req->r_args.open.mask = cpu_to_le32(mask);
412 set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
413 err = ceph_mdsc_do_request(mdsc,
414 (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
416 err = ceph_handle_snapdir(req, dentry, err);
420 if ((flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
421 err = ceph_handle_notrace_create(dir, dentry);
423 if (d_in_lookup(dentry)) {
424 dn = ceph_finish_lookup(req, dentry, err);
428 /* we were given a hashed negative dentry */
433 if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
434 /* make vfs retry on splice, ENOENT, or symlink */
435 dout("atomic_open finish_no_open on dn %p\n", dn);
436 err = finish_no_open(file, dn);
438 dout("atomic_open finish_open on dn %p\n", dn);
439 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
440 ceph_init_inode_acls(d_inode(dentry), &acls);
441 *opened |= FILE_CREATED;
443 err = finish_open(file, dentry, ceph_open, opened);
446 if (!req->r_err && req->r_target_inode)
447 ceph_put_fmode(ceph_inode(req->r_target_inode), req->r_fmode);
448 ceph_mdsc_put_request(req);
450 ceph_release_acls_info(&acls);
451 dout("atomic_open result=%d\n", err);
455 int ceph_release(struct inode *inode, struct file *file)
457 struct ceph_inode_info *ci = ceph_inode(inode);
458 struct ceph_file_info *cf = file->private_data;
460 dout("release inode %p file %p\n", inode, file);
461 ceph_put_fmode(ci, cf->fmode);
462 if (cf->last_readdir)
463 ceph_mdsc_put_request(cf->last_readdir);
464 kfree(cf->last_name);
466 kmem_cache_free(ceph_file_cachep, cf);
468 /* wake up anyone waiting for caps on this inode */
469 wake_up_all(&ci->i_cap_wq);
480 * Read a range of bytes striped over one or more objects. Iterate over
481 * objects we stripe over. (That's not atomic, but good enough for now.)
483 * If we get a short result from the OSD, check against i_size; we need to
484 * only return a short read to the caller if we hit EOF.
486 static int striped_read(struct inode *inode,
488 struct page **pages, int num_pages,
489 int page_align, int *checkeof)
491 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
492 struct ceph_inode_info *ci = ceph_inode(inode);
497 bool hit_stripe, was_short;
500 * we may need to do multiple reads. not atomic, unfortunately.
504 page_idx = (page_align + read) >> PAGE_SHIFT;
505 ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode),
506 &ci->i_layout, pos, &this_len,
507 ci->i_truncate_seq, ci->i_truncate_size,
508 pages + page_idx, num_pages - page_idx,
509 ((page_align + read) & ~PAGE_MASK));
512 hit_stripe = this_len < len;
513 was_short = ret >= 0 && ret < this_len;
514 dout("striped_read %llu~%llu (read %u) got %d%s%s\n", pos, len, read,
515 ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : "");
517 i_size = i_size_read(inode);
519 if (was_short && (pos + ret < i_size)) {
520 int zlen = min(this_len - ret, i_size - pos - ret);
521 int zoff = page_align + read + ret;
522 dout(" zero gap %llu to %llu\n",
523 pos + ret, pos + ret + zlen);
524 ceph_zero_page_vector_range(zoff, zlen, pages);
532 /* hit stripe and need continue*/
533 if (len && hit_stripe && pos < i_size)
539 /* did we bounce off eof? */
540 if (pos + len > i_size)
541 *checkeof = CHECK_EOF;
544 dout("striped_read returns %d\n", ret);
549 * Completely synchronous read and write methods. Direct from __user
550 * buffer to osd, or directly to user pages (if O_DIRECT).
552 * If the read spans object boundary, just do multiple reads.
554 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
557 struct file *file = iocb->ki_filp;
558 struct inode *inode = file_inode(file);
560 u64 off = iocb->ki_pos;
563 size_t len = iov_iter_count(to);
565 dout("sync_read on file %p %llu~%u %s\n", file, off, (unsigned)len,
566 (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
571 * flush any page cache pages in this range. this
572 * will make concurrent normal and sync io slow,
573 * but it will at least behave sensibly when they are
576 ret = filemap_write_and_wait_range(inode->i_mapping, off,
581 if (unlikely(to->type & ITER_PIPE)) {
583 ret = iov_iter_get_pages_alloc(to, &pages, len,
587 num_pages = DIV_ROUND_UP(ret + page_off, PAGE_SIZE);
589 ret = striped_read(inode, off, ret, pages, num_pages,
592 iov_iter_advance(to, ret);
595 iov_iter_advance(to, 0);
597 ceph_put_page_vector(pages, num_pages, false);
599 num_pages = calc_pages_for(off, len);
600 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
602 return PTR_ERR(pages);
604 ret = striped_read(inode, off, len, pages, num_pages,
605 (off & ~PAGE_MASK), checkeof);
611 size_t page_off = off & ~PAGE_MASK;
612 size_t copy = min_t(size_t, left,
613 PAGE_SIZE - page_off);
614 l = copy_page_to_iter(pages[k++], page_off,
622 ceph_release_page_vector(pages, num_pages);
625 if (off > iocb->ki_pos) {
626 ret = off - iocb->ki_pos;
630 dout("sync_read result %zd\n", ret);
634 struct ceph_aio_request {
639 struct list_head osd_reqs;
641 atomic_t pending_reqs;
642 struct timespec mtime;
643 struct ceph_cap_flush *prealloc_cf;
646 struct ceph_aio_work {
647 struct work_struct work;
648 struct ceph_osd_request *req;
651 static void ceph_aio_retry_work(struct work_struct *work);
653 static void ceph_aio_complete(struct inode *inode,
654 struct ceph_aio_request *aio_req)
656 struct ceph_inode_info *ci = ceph_inode(inode);
659 if (!atomic_dec_and_test(&aio_req->pending_reqs))
662 ret = aio_req->error;
664 ret = aio_req->total_len;
666 dout("ceph_aio_complete %p rc %d\n", inode, ret);
668 if (ret >= 0 && aio_req->write) {
671 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
672 if (endoff > i_size_read(inode)) {
673 if (ceph_inode_set_size(inode, endoff))
674 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
677 spin_lock(&ci->i_ceph_lock);
678 ci->i_inline_version = CEPH_INLINE_NONE;
679 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
680 &aio_req->prealloc_cf);
681 spin_unlock(&ci->i_ceph_lock);
683 __mark_inode_dirty(inode, dirty);
687 ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
690 aio_req->iocb->ki_complete(aio_req->iocb, ret, 0);
692 ceph_free_cap_flush(aio_req->prealloc_cf);
696 static void ceph_aio_complete_req(struct ceph_osd_request *req)
698 int rc = req->r_result;
699 struct inode *inode = req->r_inode;
700 struct ceph_aio_request *aio_req = req->r_priv;
701 struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
702 int num_pages = calc_pages_for((u64)osd_data->alignment,
705 dout("ceph_aio_complete_req %p rc %d bytes %llu\n",
706 inode, rc, osd_data->length);
708 if (rc == -EOLDSNAPC) {
709 struct ceph_aio_work *aio_work;
710 BUG_ON(!aio_req->write);
712 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
714 INIT_WORK(&aio_work->work, ceph_aio_retry_work);
716 queue_work(ceph_inode_to_client(inode)->wb_wq,
721 } else if (!aio_req->write) {
724 if (rc >= 0 && osd_data->length > rc) {
725 int zoff = osd_data->alignment + rc;
726 int zlen = osd_data->length - rc;
728 * If read is satisfied by single OSD request,
729 * it can pass EOF. Otherwise read is within
732 if (aio_req->num_reqs == 1) {
733 loff_t i_size = i_size_read(inode);
734 loff_t endoff = aio_req->iocb->ki_pos + rc;
736 zlen = min_t(size_t, zlen,
738 aio_req->total_len = rc + zlen;
742 ceph_zero_page_vector_range(zoff, zlen,
747 ceph_put_page_vector(osd_data->pages, num_pages, !aio_req->write);
748 ceph_osdc_put_request(req);
751 cmpxchg(&aio_req->error, 0, rc);
753 ceph_aio_complete(inode, aio_req);
757 static void ceph_aio_retry_work(struct work_struct *work)
759 struct ceph_aio_work *aio_work =
760 container_of(work, struct ceph_aio_work, work);
761 struct ceph_osd_request *orig_req = aio_work->req;
762 struct ceph_aio_request *aio_req = orig_req->r_priv;
763 struct inode *inode = orig_req->r_inode;
764 struct ceph_inode_info *ci = ceph_inode(inode);
765 struct ceph_snap_context *snapc;
766 struct ceph_osd_request *req;
769 spin_lock(&ci->i_ceph_lock);
770 if (__ceph_have_pending_cap_snap(ci)) {
771 struct ceph_cap_snap *capsnap =
772 list_last_entry(&ci->i_cap_snaps,
773 struct ceph_cap_snap,
775 snapc = ceph_get_snap_context(capsnap->context);
777 BUG_ON(!ci->i_head_snapc);
778 snapc = ceph_get_snap_context(ci->i_head_snapc);
780 spin_unlock(&ci->i_ceph_lock);
782 req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 2,
790 req->r_flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
791 ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
792 ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
794 ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
796 ceph_osdc_put_request(req);
801 req->r_ops[0] = orig_req->r_ops[0];
803 req->r_mtime = aio_req->mtime;
804 req->r_data_offset = req->r_ops[0].extent.offset;
806 ceph_osdc_put_request(orig_req);
808 req->r_callback = ceph_aio_complete_req;
809 req->r_inode = inode;
810 req->r_priv = aio_req;
811 req->r_abort_on_full = true;
813 ret = ceph_osdc_start_request(req->r_osdc, req, false);
817 ceph_aio_complete_req(req);
820 ceph_put_snap_context(snapc);
825 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
826 struct ceph_snap_context *snapc,
827 struct ceph_cap_flush **pcf)
829 struct file *file = iocb->ki_filp;
830 struct inode *inode = file_inode(file);
831 struct ceph_inode_info *ci = ceph_inode(inode);
832 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
833 struct ceph_vino vino;
834 struct ceph_osd_request *req;
836 struct ceph_aio_request *aio_req = NULL;
840 struct timespec mtime = current_time(inode);
841 size_t count = iov_iter_count(iter);
842 loff_t pos = iocb->ki_pos;
843 bool write = iov_iter_rw(iter) == WRITE;
845 if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
848 dout("sync_direct_%s on file %p %lld~%u snapc %p seq %lld\n",
849 (write ? "write" : "read"), file, pos, (unsigned)count,
852 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
857 int ret2 = invalidate_inode_pages2_range(inode->i_mapping,
859 (pos + count) >> PAGE_SHIFT);
861 dout("invalidate_inode_pages2_range returned %d\n", ret2);
863 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
865 flags = CEPH_OSD_FLAG_READ;
868 while (iov_iter_count(iter) > 0) {
869 u64 size = dio_get_pagev_size(iter);
873 vino = ceph_vino(inode);
874 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
877 write ? CEPH_OSD_OP_WRITE :
889 size = min_t(u64, size, fsc->mount_options->wsize);
891 size = min_t(u64, size, fsc->mount_options->rsize);
894 pages = dio_get_pages_alloc(iter, len, &start, &num_pages);
896 ceph_osdc_put_request(req);
897 ret = PTR_ERR(pages);
902 * To simplify error handling, allow AIO when IO within i_size
903 * or IO can be satisfied by single OSD request.
905 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
906 (len == count || pos + count <= i_size_read(inode))) {
907 aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
909 aio_req->iocb = iocb;
910 aio_req->write = write;
911 INIT_LIST_HEAD(&aio_req->osd_reqs);
913 aio_req->mtime = mtime;
914 swap(aio_req->prealloc_cf, *pcf);
922 * throw out any page cache pages in this range. this
925 truncate_inode_pages_range(inode->i_mapping, pos,
926 (pos+len) | (PAGE_SIZE - 1));
928 req->r_mtime = mtime;
931 osd_req_op_extent_osd_data_pages(req, 0, pages, len, start,
935 aio_req->total_len += len;
937 atomic_inc(&aio_req->pending_reqs);
939 req->r_callback = ceph_aio_complete_req;
940 req->r_inode = inode;
941 req->r_priv = aio_req;
942 list_add_tail(&req->r_unsafe_item, &aio_req->osd_reqs);
945 iov_iter_advance(iter, len);
949 ret = ceph_osdc_start_request(req->r_osdc, req, false);
951 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
953 size = i_size_read(inode);
957 if (ret >= 0 && ret < len && pos + ret < size) {
958 int zlen = min_t(size_t, len - ret,
960 ceph_zero_page_vector_range(start + ret, zlen,
968 ceph_put_page_vector(pages, num_pages, !write);
970 ceph_osdc_put_request(req);
975 iov_iter_advance(iter, len);
977 if (!write && pos >= size)
980 if (write && pos > size) {
981 if (ceph_inode_set_size(inode, pos))
982 ceph_check_caps(ceph_inode(inode),
991 if (aio_req->num_reqs == 0) {
996 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
999 list_splice(&aio_req->osd_reqs, &osd_reqs);
1000 while (!list_empty(&osd_reqs)) {
1001 req = list_first_entry(&osd_reqs,
1002 struct ceph_osd_request,
1004 list_del_init(&req->r_unsafe_item);
1006 ret = ceph_osdc_start_request(req->r_osdc,
1009 req->r_result = ret;
1010 ceph_aio_complete_req(req);
1013 return -EIOCBQUEUED;
1016 if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1017 ret = pos - iocb->ki_pos;
1024 * Synchronous write, straight from __user pointer or user pages.
1026 * If write spans object boundary, just do multiple writes. (For a
1027 * correct atomic write, we should e.g. take write locks on all
1028 * objects, rollback on failure, etc.)
1031 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1032 struct ceph_snap_context *snapc)
1034 struct file *file = iocb->ki_filp;
1035 struct inode *inode = file_inode(file);
1036 struct ceph_inode_info *ci = ceph_inode(inode);
1037 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1038 struct ceph_vino vino;
1039 struct ceph_osd_request *req;
1040 struct page **pages;
1046 bool check_caps = false;
1047 struct timespec mtime = current_time(inode);
1048 size_t count = iov_iter_count(from);
1050 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1053 dout("sync_write on file %p %lld~%u snapc %p seq %lld\n",
1054 file, pos, (unsigned)count, snapc, snapc->seq);
1056 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
1060 ret = invalidate_inode_pages2_range(inode->i_mapping,
1062 (pos + count) >> PAGE_SHIFT);
1064 dout("invalidate_inode_pages2_range returned %d\n", ret);
1066 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1068 while ((len = iov_iter_count(from)) > 0) {
1072 vino = ceph_vino(inode);
1073 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1074 vino, pos, &len, 0, 1,
1075 CEPH_OSD_OP_WRITE, flags, snapc,
1077 ci->i_truncate_size,
1085 * write from beginning of first page,
1086 * regardless of io alignment
1088 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1090 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1091 if (IS_ERR(pages)) {
1092 ret = PTR_ERR(pages);
1097 for (n = 0; n < num_pages; n++) {
1098 size_t plen = min_t(size_t, left, PAGE_SIZE);
1099 ret = copy_page_from_iter(pages[n], 0, plen, from);
1108 ceph_release_page_vector(pages, num_pages);
1112 req->r_inode = inode;
1114 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1117 req->r_mtime = mtime;
1118 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1120 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1123 ceph_osdc_put_request(req);
1125 ceph_set_error_write(ci);
1129 ceph_clear_error_write(ci);
1132 if (pos > i_size_read(inode)) {
1133 check_caps = ceph_inode_set_size(inode, pos);
1135 ceph_check_caps(ceph_inode(inode),
1136 CHECK_CAPS_AUTHONLY,
1142 if (ret != -EOLDSNAPC && written > 0) {
1150 * Wrap generic_file_aio_read with checks for cap bits on the inode.
1151 * Atomically grab references, so that those bits are not released
1152 * back to the MDS mid-read.
1154 * Hmm, the sync read case isn't actually async... should it be?
1156 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1158 struct file *filp = iocb->ki_filp;
1159 struct ceph_file_info *fi = filp->private_data;
1160 size_t len = iov_iter_count(to);
1161 struct inode *inode = file_inode(filp);
1162 struct ceph_inode_info *ci = ceph_inode(inode);
1163 struct page *pinned_page = NULL;
1166 int retry_op = 0, read = 0;
1169 dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1170 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1172 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1173 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1175 want = CEPH_CAP_FILE_CACHE;
1176 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1180 if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1181 (iocb->ki_flags & IOCB_DIRECT) ||
1182 (fi->flags & CEPH_F_SYNC)) {
1184 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1185 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1186 ceph_cap_string(got));
1188 if (ci->i_inline_version == CEPH_INLINE_NONE) {
1189 if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1190 ret = ceph_direct_read_write(iocb, to,
1192 if (ret >= 0 && ret < len)
1193 retry_op = CHECK_EOF;
1195 ret = ceph_sync_read(iocb, to, &retry_op);
1198 retry_op = READ_INLINE;
1201 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1202 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1203 ceph_cap_string(got));
1204 current->journal_info = filp;
1205 ret = generic_file_read_iter(iocb, to);
1206 current->journal_info = NULL;
1208 dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1209 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1211 put_page(pinned_page);
1214 ceph_put_cap_refs(ci, got);
1215 if (retry_op > HAVE_RETRIED && ret >= 0) {
1217 struct page *page = NULL;
1219 if (retry_op == READ_INLINE) {
1220 page = __page_cache_alloc(GFP_KERNEL);
1225 statret = __ceph_do_getattr(inode, page,
1226 CEPH_STAT_CAP_INLINE_DATA, !!page);
1230 if (statret == -ENODATA) {
1231 BUG_ON(retry_op != READ_INLINE);
1237 i_size = i_size_read(inode);
1238 if (retry_op == READ_INLINE) {
1239 BUG_ON(ret > 0 || read > 0);
1240 if (iocb->ki_pos < i_size &&
1241 iocb->ki_pos < PAGE_SIZE) {
1242 loff_t end = min_t(loff_t, i_size,
1243 iocb->ki_pos + len);
1244 end = min_t(loff_t, end, PAGE_SIZE);
1246 zero_user_segment(page, statret, end);
1247 ret = copy_page_to_iter(page,
1248 iocb->ki_pos & ~PAGE_MASK,
1249 end - iocb->ki_pos, to);
1250 iocb->ki_pos += ret;
1253 if (iocb->ki_pos < i_size && read < len) {
1254 size_t zlen = min_t(size_t, len - read,
1255 i_size - iocb->ki_pos);
1256 ret = iov_iter_zero(zlen, to);
1257 iocb->ki_pos += ret;
1260 __free_pages(page, 0);
1264 /* hit EOF or hole? */
1265 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1267 dout("sync_read hit hole, ppos %lld < size %lld"
1268 ", reading more\n", iocb->ki_pos, i_size);
1272 retry_op = HAVE_RETRIED;
1284 * Take cap references to avoid releasing caps to MDS mid-write.
1286 * If we are synchronous, and write with an old snap context, the OSD
1287 * may return EOLDSNAPC. In that case, retry the write.. _after_
1288 * dropping our cap refs and allowing the pending snap to logically
1289 * complete _before_ this write occurs.
1291 * If we are near ENOSPC, write synchronously.
1293 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1295 struct file *file = iocb->ki_filp;
1296 struct ceph_file_info *fi = file->private_data;
1297 struct inode *inode = file_inode(file);
1298 struct ceph_inode_info *ci = ceph_inode(inode);
1299 struct ceph_osd_client *osdc =
1300 &ceph_sb_to_client(inode->i_sb)->client->osdc;
1301 struct ceph_cap_flush *prealloc_cf;
1302 ssize_t count, written = 0;
1306 if (ceph_snap(inode) != CEPH_NOSNAP)
1309 prealloc_cf = ceph_alloc_cap_flush();
1316 /* We can write back this queue in page reclaim */
1317 current->backing_dev_info = inode_to_bdi(inode);
1319 if (iocb->ki_flags & IOCB_APPEND) {
1320 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1325 err = generic_write_checks(iocb, from);
1330 count = iov_iter_count(from);
1331 err = file_remove_privs(file);
1335 err = file_update_time(file);
1339 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1340 err = ceph_uninline_data(file, NULL);
1345 /* FIXME: not complete since it doesn't account for being at quota */
1346 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL)) {
1351 dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1352 inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1353 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1354 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1356 want = CEPH_CAP_FILE_BUFFER;
1358 err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, pos + count,
1363 dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1364 inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1366 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1367 (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
1368 (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
1369 struct ceph_snap_context *snapc;
1370 struct iov_iter data;
1371 inode_unlock(inode);
1373 spin_lock(&ci->i_ceph_lock);
1374 if (__ceph_have_pending_cap_snap(ci)) {
1375 struct ceph_cap_snap *capsnap =
1376 list_last_entry(&ci->i_cap_snaps,
1377 struct ceph_cap_snap,
1379 snapc = ceph_get_snap_context(capsnap->context);
1381 BUG_ON(!ci->i_head_snapc);
1382 snapc = ceph_get_snap_context(ci->i_head_snapc);
1384 spin_unlock(&ci->i_ceph_lock);
1386 /* we might need to revert back to that point */
1388 if (iocb->ki_flags & IOCB_DIRECT)
1389 written = ceph_direct_read_write(iocb, &data, snapc,
1392 written = ceph_sync_write(iocb, &data, pos, snapc);
1394 iov_iter_advance(from, written);
1395 ceph_put_snap_context(snapc);
1398 * No need to acquire the i_truncate_mutex. Because
1399 * the MDS revokes Fwb caps before sending truncate
1400 * message to us. We can't get Fwb cap while there
1401 * are pending vmtruncate. So write and vmtruncate
1402 * can not run at the same time
1404 written = generic_perform_write(file, from, pos);
1405 if (likely(written >= 0))
1406 iocb->ki_pos = pos + written;
1407 inode_unlock(inode);
1412 spin_lock(&ci->i_ceph_lock);
1413 ci->i_inline_version = CEPH_INLINE_NONE;
1414 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1416 spin_unlock(&ci->i_ceph_lock);
1418 __mark_inode_dirty(inode, dirty);
1421 dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n",
1422 inode, ceph_vinop(inode), pos, (unsigned)count,
1423 ceph_cap_string(got));
1424 ceph_put_cap_refs(ci, got);
1426 if (written == -EOLDSNAPC) {
1427 dout("aio_write %p %llx.%llx %llu~%u" "got EOLDSNAPC, retrying\n",
1428 inode, ceph_vinop(inode), pos, (unsigned)count);
1433 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_NEARFULL))
1434 iocb->ki_flags |= IOCB_DSYNC;
1435 written = generic_write_sync(iocb, written);
1441 inode_unlock(inode);
1443 ceph_free_cap_flush(prealloc_cf);
1444 current->backing_dev_info = NULL;
1445 return written ? written : err;
1449 * llseek. be sure to verify file size on SEEK_END.
1451 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1453 struct inode *inode = file->f_mapping->host;
1459 if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1460 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1465 i_size = i_size_read(inode);
1472 * Here we special-case the lseek(fd, 0, SEEK_CUR)
1473 * position-querying operation. Avoid rewriting the "same"
1474 * f_pos value back to the file because a concurrent read(),
1475 * write() or lseek() might have altered it
1481 offset += file->f_pos;
1484 if (offset < 0 || offset >= i_size) {
1490 if (offset < 0 || offset >= i_size) {
1498 ret = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1501 inode_unlock(inode);
1505 static inline void ceph_zero_partial_page(
1506 struct inode *inode, loff_t offset, unsigned size)
1509 pgoff_t index = offset >> PAGE_SHIFT;
1511 page = find_lock_page(inode->i_mapping, index);
1513 wait_on_page_writeback(page);
1514 zero_user(page, offset & (PAGE_SIZE - 1), size);
1520 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1523 loff_t nearly = round_up(offset, PAGE_SIZE);
1524 if (offset < nearly) {
1525 loff_t size = nearly - offset;
1528 ceph_zero_partial_page(inode, offset, size);
1532 if (length >= PAGE_SIZE) {
1533 loff_t size = round_down(length, PAGE_SIZE);
1534 truncate_pagecache_range(inode, offset, offset + size - 1);
1539 ceph_zero_partial_page(inode, offset, length);
1542 static int ceph_zero_partial_object(struct inode *inode,
1543 loff_t offset, loff_t *length)
1545 struct ceph_inode_info *ci = ceph_inode(inode);
1546 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1547 struct ceph_osd_request *req;
1553 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1556 op = CEPH_OSD_OP_ZERO;
1559 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1563 CEPH_OSD_FLAG_WRITE,
1570 req->r_mtime = inode->i_mtime;
1571 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1573 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1577 ceph_osdc_put_request(req);
1583 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
1586 struct ceph_inode_info *ci = ceph_inode(inode);
1587 s32 stripe_unit = ci->i_layout.stripe_unit;
1588 s32 stripe_count = ci->i_layout.stripe_count;
1589 s32 object_size = ci->i_layout.object_size;
1590 u64 object_set_size = object_size * stripe_count;
1593 /* round offset up to next period boundary */
1594 nearly = offset + object_set_size - 1;
1596 nearly -= do_div(t, object_set_size);
1598 while (length && offset < nearly) {
1599 loff_t size = length;
1600 ret = ceph_zero_partial_object(inode, offset, &size);
1606 while (length >= object_set_size) {
1608 loff_t pos = offset;
1609 for (i = 0; i < stripe_count; ++i) {
1610 ret = ceph_zero_partial_object(inode, pos, NULL);
1615 offset += object_set_size;
1616 length -= object_set_size;
1619 loff_t size = length;
1620 ret = ceph_zero_partial_object(inode, offset, &size);
1629 static long ceph_fallocate(struct file *file, int mode,
1630 loff_t offset, loff_t length)
1632 struct ceph_file_info *fi = file->private_data;
1633 struct inode *inode = file_inode(file);
1634 struct ceph_inode_info *ci = ceph_inode(inode);
1635 struct ceph_osd_client *osdc =
1636 &ceph_inode_to_client(inode)->client->osdc;
1637 struct ceph_cap_flush *prealloc_cf;
1644 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1647 if (!S_ISREG(inode->i_mode))
1650 prealloc_cf = ceph_alloc_cap_flush();
1656 if (ceph_snap(inode) != CEPH_NOSNAP) {
1661 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) &&
1662 !(mode & FALLOC_FL_PUNCH_HOLE)) {
1667 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1668 ret = ceph_uninline_data(file, NULL);
1673 size = i_size_read(inode);
1674 if (!(mode & FALLOC_FL_KEEP_SIZE)) {
1675 endoff = offset + length;
1676 ret = inode_newsize_ok(inode, endoff);
1681 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1682 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1684 want = CEPH_CAP_FILE_BUFFER;
1686 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, endoff, &got, NULL);
1690 if (mode & FALLOC_FL_PUNCH_HOLE) {
1692 ceph_zero_pagecache_range(inode, offset, length);
1693 ret = ceph_zero_objects(inode, offset, length);
1694 } else if (endoff > size) {
1695 truncate_pagecache_range(inode, size, -1);
1696 if (ceph_inode_set_size(inode, endoff))
1697 ceph_check_caps(ceph_inode(inode),
1698 CHECK_CAPS_AUTHONLY, NULL);
1702 spin_lock(&ci->i_ceph_lock);
1703 ci->i_inline_version = CEPH_INLINE_NONE;
1704 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1706 spin_unlock(&ci->i_ceph_lock);
1708 __mark_inode_dirty(inode, dirty);
1711 ceph_put_cap_refs(ci, got);
1713 inode_unlock(inode);
1714 ceph_free_cap_flush(prealloc_cf);
1718 const struct file_operations ceph_file_fops = {
1720 .release = ceph_release,
1721 .llseek = ceph_llseek,
1722 .read_iter = ceph_read_iter,
1723 .write_iter = ceph_write_iter,
1725 .fsync = ceph_fsync,
1727 .flock = ceph_flock,
1728 .splice_read = generic_file_splice_read,
1729 .splice_write = iter_file_splice_write,
1730 .unlocked_ioctl = ceph_ioctl,
1731 .compat_ioctl = ceph_ioctl,
1732 .fallocate = ceph_fallocate,