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,
567 (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
572 * flush any page cache pages in this range. this
573 * will make concurrent normal and sync io slow,
574 * but it will at least behave sensibly when they are
577 ret = filemap_write_and_wait_range(inode->i_mapping, off,
582 if (unlikely(to->type & ITER_PIPE)) {
584 ret = iov_iter_get_pages_alloc(to, &pages, len,
588 num_pages = DIV_ROUND_UP(ret + page_off, PAGE_SIZE);
590 ret = striped_read(inode, off, ret, pages, num_pages,
593 iov_iter_advance(to, ret);
596 iov_iter_advance(to, 0);
598 ceph_put_page_vector(pages, num_pages, false);
600 num_pages = calc_pages_for(off, len);
601 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
603 return PTR_ERR(pages);
605 ret = striped_read(inode, off, len, pages, num_pages,
606 (off & ~PAGE_MASK), checkeof);
612 size_t page_off = off & ~PAGE_MASK;
613 size_t copy = min_t(size_t, left,
614 PAGE_SIZE - page_off);
615 l = copy_page_to_iter(pages[k++], page_off,
623 ceph_release_page_vector(pages, num_pages);
626 if (off > iocb->ki_pos) {
627 ret = off - iocb->ki_pos;
631 dout("sync_read result %zd\n", ret);
635 struct ceph_aio_request {
640 struct list_head osd_reqs;
642 atomic_t pending_reqs;
643 struct timespec mtime;
644 struct ceph_cap_flush *prealloc_cf;
647 struct ceph_aio_work {
648 struct work_struct work;
649 struct ceph_osd_request *req;
652 static void ceph_aio_retry_work(struct work_struct *work);
654 static void ceph_aio_complete(struct inode *inode,
655 struct ceph_aio_request *aio_req)
657 struct ceph_inode_info *ci = ceph_inode(inode);
660 if (!atomic_dec_and_test(&aio_req->pending_reqs))
663 ret = aio_req->error;
665 ret = aio_req->total_len;
667 dout("ceph_aio_complete %p rc %d\n", inode, ret);
669 if (ret >= 0 && aio_req->write) {
672 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
673 if (endoff > i_size_read(inode)) {
674 if (ceph_inode_set_size(inode, endoff))
675 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
678 spin_lock(&ci->i_ceph_lock);
679 ci->i_inline_version = CEPH_INLINE_NONE;
680 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
681 &aio_req->prealloc_cf);
682 spin_unlock(&ci->i_ceph_lock);
684 __mark_inode_dirty(inode, dirty);
688 ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
691 aio_req->iocb->ki_complete(aio_req->iocb, ret, 0);
693 ceph_free_cap_flush(aio_req->prealloc_cf);
697 static void ceph_aio_complete_req(struct ceph_osd_request *req)
699 int rc = req->r_result;
700 struct inode *inode = req->r_inode;
701 struct ceph_aio_request *aio_req = req->r_priv;
702 struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
703 int num_pages = calc_pages_for((u64)osd_data->alignment,
706 dout("ceph_aio_complete_req %p rc %d bytes %llu\n",
707 inode, rc, osd_data->length);
709 if (rc == -EOLDSNAPC) {
710 struct ceph_aio_work *aio_work;
711 BUG_ON(!aio_req->write);
713 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
715 INIT_WORK(&aio_work->work, ceph_aio_retry_work);
717 queue_work(ceph_inode_to_client(inode)->wb_wq,
722 } else if (!aio_req->write) {
725 if (rc >= 0 && osd_data->length > rc) {
726 int zoff = osd_data->alignment + rc;
727 int zlen = osd_data->length - rc;
729 * If read is satisfied by single OSD request,
730 * it can pass EOF. Otherwise read is within
733 if (aio_req->num_reqs == 1) {
734 loff_t i_size = i_size_read(inode);
735 loff_t endoff = aio_req->iocb->ki_pos + rc;
737 zlen = min_t(size_t, zlen,
739 aio_req->total_len = rc + zlen;
743 ceph_zero_page_vector_range(zoff, zlen,
748 ceph_put_page_vector(osd_data->pages, num_pages, !aio_req->write);
749 ceph_osdc_put_request(req);
752 cmpxchg(&aio_req->error, 0, rc);
754 ceph_aio_complete(inode, aio_req);
758 static void ceph_aio_retry_work(struct work_struct *work)
760 struct ceph_aio_work *aio_work =
761 container_of(work, struct ceph_aio_work, work);
762 struct ceph_osd_request *orig_req = aio_work->req;
763 struct ceph_aio_request *aio_req = orig_req->r_priv;
764 struct inode *inode = orig_req->r_inode;
765 struct ceph_inode_info *ci = ceph_inode(inode);
766 struct ceph_snap_context *snapc;
767 struct ceph_osd_request *req;
770 spin_lock(&ci->i_ceph_lock);
771 if (__ceph_have_pending_cap_snap(ci)) {
772 struct ceph_cap_snap *capsnap =
773 list_last_entry(&ci->i_cap_snaps,
774 struct ceph_cap_snap,
776 snapc = ceph_get_snap_context(capsnap->context);
778 BUG_ON(!ci->i_head_snapc);
779 snapc = ceph_get_snap_context(ci->i_head_snapc);
781 spin_unlock(&ci->i_ceph_lock);
783 req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 2,
791 req->r_flags = CEPH_OSD_FLAG_ORDERSNAP | CEPH_OSD_FLAG_WRITE;
792 ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
793 ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
795 ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
797 ceph_osdc_put_request(req);
802 req->r_ops[0] = orig_req->r_ops[0];
803 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC, 0);
805 req->r_mtime = aio_req->mtime;
806 req->r_data_offset = req->r_ops[0].extent.offset;
808 ceph_osdc_put_request(orig_req);
810 req->r_callback = ceph_aio_complete_req;
811 req->r_inode = inode;
812 req->r_priv = aio_req;
813 req->r_abort_on_full = true;
815 ret = ceph_osdc_start_request(req->r_osdc, req, false);
819 ceph_aio_complete_req(req);
822 ceph_put_snap_context(snapc);
827 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
828 struct ceph_snap_context *snapc,
829 struct ceph_cap_flush **pcf)
831 struct file *file = iocb->ki_filp;
832 struct inode *inode = file_inode(file);
833 struct ceph_inode_info *ci = ceph_inode(inode);
834 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
835 struct ceph_vino vino;
836 struct ceph_osd_request *req;
838 struct ceph_aio_request *aio_req = NULL;
842 struct timespec mtime = current_time(inode);
843 size_t count = iov_iter_count(iter);
844 loff_t pos = iocb->ki_pos;
845 bool write = iov_iter_rw(iter) == WRITE;
847 if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
850 dout("sync_direct_read_write (%s) on file %p %lld~%u\n",
851 (write ? "write" : "read"), file, pos, (unsigned)count);
853 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
858 int ret2 = invalidate_inode_pages2_range(inode->i_mapping,
860 (pos + count) >> PAGE_SHIFT);
862 dout("invalidate_inode_pages2_range returned %d\n", ret2);
864 flags = CEPH_OSD_FLAG_ORDERSNAP | CEPH_OSD_FLAG_WRITE;
866 flags = CEPH_OSD_FLAG_READ;
869 while (iov_iter_count(iter) > 0) {
870 u64 size = dio_get_pagev_size(iter);
874 vino = ceph_vino(inode);
875 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
877 /*include a 'startsync' command*/
879 write ? CEPH_OSD_OP_WRITE :
891 size = min_t(u64, size, fsc->mount_options->wsize);
893 size = min_t(u64, size, fsc->mount_options->rsize);
896 pages = dio_get_pages_alloc(iter, len, &start, &num_pages);
898 ceph_osdc_put_request(req);
899 ret = PTR_ERR(pages);
904 * To simplify error handling, allow AIO when IO within i_size
905 * or IO can be satisfied by single OSD request.
907 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
908 (len == count || pos + count <= i_size_read(inode))) {
909 aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
911 aio_req->iocb = iocb;
912 aio_req->write = write;
913 INIT_LIST_HEAD(&aio_req->osd_reqs);
915 aio_req->mtime = mtime;
916 swap(aio_req->prealloc_cf, *pcf);
924 * throw out any page cache pages in this range. this
927 truncate_inode_pages_range(inode->i_mapping, pos,
928 (pos+len) | (PAGE_SIZE - 1));
930 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC, 0);
931 req->r_mtime = mtime;
934 osd_req_op_extent_osd_data_pages(req, 0, pages, len, start,
938 aio_req->total_len += len;
940 atomic_inc(&aio_req->pending_reqs);
942 req->r_callback = ceph_aio_complete_req;
943 req->r_inode = inode;
944 req->r_priv = aio_req;
945 list_add_tail(&req->r_unsafe_item, &aio_req->osd_reqs);
948 iov_iter_advance(iter, len);
952 ret = ceph_osdc_start_request(req->r_osdc, req, false);
954 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
956 size = i_size_read(inode);
960 if (ret >= 0 && ret < len && pos + ret < size) {
961 int zlen = min_t(size_t, len - ret,
963 ceph_zero_page_vector_range(start + ret, zlen,
971 ceph_put_page_vector(pages, num_pages, !write);
973 ceph_osdc_put_request(req);
978 iov_iter_advance(iter, len);
980 if (!write && pos >= size)
983 if (write && pos > size) {
984 if (ceph_inode_set_size(inode, pos))
985 ceph_check_caps(ceph_inode(inode),
994 if (aio_req->num_reqs == 0) {
999 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1002 list_splice(&aio_req->osd_reqs, &osd_reqs);
1003 while (!list_empty(&osd_reqs)) {
1004 req = list_first_entry(&osd_reqs,
1005 struct ceph_osd_request,
1007 list_del_init(&req->r_unsafe_item);
1009 ret = ceph_osdc_start_request(req->r_osdc,
1012 req->r_result = ret;
1013 ceph_aio_complete_req(req);
1016 return -EIOCBQUEUED;
1019 if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1020 ret = pos - iocb->ki_pos;
1027 * Synchronous write, straight from __user pointer or user pages.
1029 * If write spans object boundary, just do multiple writes. (For a
1030 * correct atomic write, we should e.g. take write locks on all
1031 * objects, rollback on failure, etc.)
1034 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1035 struct ceph_snap_context *snapc)
1037 struct file *file = iocb->ki_filp;
1038 struct inode *inode = file_inode(file);
1039 struct ceph_inode_info *ci = ceph_inode(inode);
1040 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1041 struct ceph_vino vino;
1042 struct ceph_osd_request *req;
1043 struct page **pages;
1049 bool check_caps = false;
1050 struct timespec mtime = current_time(inode);
1051 size_t count = iov_iter_count(from);
1053 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1056 dout("sync_write on file %p %lld~%u\n", file, pos, (unsigned)count);
1058 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
1062 ret = invalidate_inode_pages2_range(inode->i_mapping,
1064 (pos + count) >> PAGE_SHIFT);
1066 dout("invalidate_inode_pages2_range returned %d\n", ret);
1068 flags = CEPH_OSD_FLAG_ORDERSNAP | CEPH_OSD_FLAG_WRITE;
1070 while ((len = iov_iter_count(from)) > 0) {
1074 vino = ceph_vino(inode);
1075 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1076 vino, pos, &len, 0, 1,
1077 CEPH_OSD_OP_WRITE, flags, snapc,
1079 ci->i_truncate_size,
1087 * write from beginning of first page,
1088 * regardless of io alignment
1090 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1092 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1093 if (IS_ERR(pages)) {
1094 ret = PTR_ERR(pages);
1099 for (n = 0; n < num_pages; n++) {
1100 size_t plen = min_t(size_t, left, PAGE_SIZE);
1101 ret = copy_page_from_iter(pages[n], 0, plen, from);
1110 ceph_release_page_vector(pages, num_pages);
1114 req->r_inode = inode;
1116 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1119 req->r_mtime = mtime;
1120 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1122 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1125 ceph_osdc_put_request(req);
1127 ceph_set_error_write(ci);
1131 ceph_clear_error_write(ci);
1134 if (pos > i_size_read(inode)) {
1135 check_caps = ceph_inode_set_size(inode, pos);
1137 ceph_check_caps(ceph_inode(inode),
1138 CHECK_CAPS_AUTHONLY,
1144 if (ret != -EOLDSNAPC && written > 0) {
1152 * Wrap generic_file_aio_read with checks for cap bits on the inode.
1153 * Atomically grab references, so that those bits are not released
1154 * back to the MDS mid-read.
1156 * Hmm, the sync read case isn't actually async... should it be?
1158 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1160 struct file *filp = iocb->ki_filp;
1161 struct ceph_file_info *fi = filp->private_data;
1162 size_t len = iov_iter_count(to);
1163 struct inode *inode = file_inode(filp);
1164 struct ceph_inode_info *ci = ceph_inode(inode);
1165 struct page *pinned_page = NULL;
1168 int retry_op = 0, read = 0;
1171 dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1172 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1174 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1175 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1177 want = CEPH_CAP_FILE_CACHE;
1178 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1182 if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1183 (iocb->ki_flags & IOCB_DIRECT) ||
1184 (fi->flags & CEPH_F_SYNC)) {
1186 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1187 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1188 ceph_cap_string(got));
1190 if (ci->i_inline_version == CEPH_INLINE_NONE) {
1191 if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1192 ret = ceph_direct_read_write(iocb, to,
1194 if (ret >= 0 && ret < len)
1195 retry_op = CHECK_EOF;
1197 ret = ceph_sync_read(iocb, to, &retry_op);
1200 retry_op = READ_INLINE;
1203 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1204 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1205 ceph_cap_string(got));
1206 current->journal_info = filp;
1207 ret = generic_file_read_iter(iocb, to);
1208 current->journal_info = NULL;
1210 dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1211 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1213 put_page(pinned_page);
1216 ceph_put_cap_refs(ci, got);
1217 if (retry_op > HAVE_RETRIED && ret >= 0) {
1219 struct page *page = NULL;
1221 if (retry_op == READ_INLINE) {
1222 page = __page_cache_alloc(GFP_KERNEL);
1227 statret = __ceph_do_getattr(inode, page,
1228 CEPH_STAT_CAP_INLINE_DATA, !!page);
1232 if (statret == -ENODATA) {
1233 BUG_ON(retry_op != READ_INLINE);
1239 i_size = i_size_read(inode);
1240 if (retry_op == READ_INLINE) {
1241 BUG_ON(ret > 0 || read > 0);
1242 if (iocb->ki_pos < i_size &&
1243 iocb->ki_pos < PAGE_SIZE) {
1244 loff_t end = min_t(loff_t, i_size,
1245 iocb->ki_pos + len);
1246 end = min_t(loff_t, end, PAGE_SIZE);
1248 zero_user_segment(page, statret, end);
1249 ret = copy_page_to_iter(page,
1250 iocb->ki_pos & ~PAGE_MASK,
1251 end - iocb->ki_pos, to);
1252 iocb->ki_pos += ret;
1255 if (iocb->ki_pos < i_size && read < len) {
1256 size_t zlen = min_t(size_t, len - read,
1257 i_size - iocb->ki_pos);
1258 ret = iov_iter_zero(zlen, to);
1259 iocb->ki_pos += ret;
1262 __free_pages(page, 0);
1266 /* hit EOF or hole? */
1267 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1269 dout("sync_read hit hole, ppos %lld < size %lld"
1270 ", reading more\n", iocb->ki_pos, i_size);
1274 retry_op = HAVE_RETRIED;
1286 * Take cap references to avoid releasing caps to MDS mid-write.
1288 * If we are synchronous, and write with an old snap context, the OSD
1289 * may return EOLDSNAPC. In that case, retry the write.. _after_
1290 * dropping our cap refs and allowing the pending snap to logically
1291 * complete _before_ this write occurs.
1293 * If we are near ENOSPC, write synchronously.
1295 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1297 struct file *file = iocb->ki_filp;
1298 struct ceph_file_info *fi = file->private_data;
1299 struct inode *inode = file_inode(file);
1300 struct ceph_inode_info *ci = ceph_inode(inode);
1301 struct ceph_osd_client *osdc =
1302 &ceph_sb_to_client(inode->i_sb)->client->osdc;
1303 struct ceph_cap_flush *prealloc_cf;
1304 ssize_t count, written = 0;
1308 if (ceph_snap(inode) != CEPH_NOSNAP)
1311 prealloc_cf = ceph_alloc_cap_flush();
1317 /* We can write back this queue in page reclaim */
1318 current->backing_dev_info = inode_to_bdi(inode);
1320 if (iocb->ki_flags & IOCB_APPEND) {
1321 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1326 err = generic_write_checks(iocb, from);
1331 count = iov_iter_count(from);
1332 err = file_remove_privs(file);
1336 err = file_update_time(file);
1340 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1341 err = ceph_uninline_data(file, NULL);
1347 /* FIXME: not complete since it doesn't account for being at quota */
1348 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL)) {
1353 dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1354 inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1355 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1356 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1358 want = CEPH_CAP_FILE_BUFFER;
1360 err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, pos + count,
1365 dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1366 inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1368 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1369 (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
1370 (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
1371 struct ceph_snap_context *snapc;
1372 struct iov_iter data;
1373 inode_unlock(inode);
1375 spin_lock(&ci->i_ceph_lock);
1376 if (__ceph_have_pending_cap_snap(ci)) {
1377 struct ceph_cap_snap *capsnap =
1378 list_last_entry(&ci->i_cap_snaps,
1379 struct ceph_cap_snap,
1381 snapc = ceph_get_snap_context(capsnap->context);
1383 BUG_ON(!ci->i_head_snapc);
1384 snapc = ceph_get_snap_context(ci->i_head_snapc);
1386 spin_unlock(&ci->i_ceph_lock);
1388 /* we might need to revert back to that point */
1390 if (iocb->ki_flags & IOCB_DIRECT)
1391 written = ceph_direct_read_write(iocb, &data, snapc,
1394 written = ceph_sync_write(iocb, &data, pos, snapc);
1395 if (written == -EOLDSNAPC) {
1396 dout("aio_write %p %llx.%llx %llu~%u"
1397 "got EOLDSNAPC, retrying\n",
1398 inode, ceph_vinop(inode),
1399 pos, (unsigned)count);
1404 iov_iter_advance(from, written);
1405 ceph_put_snap_context(snapc);
1408 * No need to acquire the i_truncate_mutex. Because
1409 * the MDS revokes Fwb caps before sending truncate
1410 * message to us. We can't get Fwb cap while there
1411 * are pending vmtruncate. So write and vmtruncate
1412 * can not run at the same time
1414 written = generic_perform_write(file, from, pos);
1415 if (likely(written >= 0))
1416 iocb->ki_pos = pos + written;
1417 inode_unlock(inode);
1422 spin_lock(&ci->i_ceph_lock);
1423 ci->i_inline_version = CEPH_INLINE_NONE;
1424 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1426 spin_unlock(&ci->i_ceph_lock);
1428 __mark_inode_dirty(inode, dirty);
1431 dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n",
1432 inode, ceph_vinop(inode), pos, (unsigned)count,
1433 ceph_cap_string(got));
1434 ceph_put_cap_refs(ci, got);
1437 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_NEARFULL))
1438 iocb->ki_flags |= IOCB_DSYNC;
1440 written = generic_write_sync(iocb, written);
1446 inode_unlock(inode);
1448 ceph_free_cap_flush(prealloc_cf);
1449 current->backing_dev_info = NULL;
1450 return written ? written : err;
1454 * llseek. be sure to verify file size on SEEK_END.
1456 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1458 struct inode *inode = file->f_mapping->host;
1464 if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1465 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1470 i_size = i_size_read(inode);
1477 * Here we special-case the lseek(fd, 0, SEEK_CUR)
1478 * position-querying operation. Avoid rewriting the "same"
1479 * f_pos value back to the file because a concurrent read(),
1480 * write() or lseek() might have altered it
1486 offset += file->f_pos;
1489 if (offset >= i_size) {
1495 if (offset >= i_size) {
1503 ret = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1506 inode_unlock(inode);
1510 static inline void ceph_zero_partial_page(
1511 struct inode *inode, loff_t offset, unsigned size)
1514 pgoff_t index = offset >> PAGE_SHIFT;
1516 page = find_lock_page(inode->i_mapping, index);
1518 wait_on_page_writeback(page);
1519 zero_user(page, offset & (PAGE_SIZE - 1), size);
1525 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1528 loff_t nearly = round_up(offset, PAGE_SIZE);
1529 if (offset < nearly) {
1530 loff_t size = nearly - offset;
1533 ceph_zero_partial_page(inode, offset, size);
1537 if (length >= PAGE_SIZE) {
1538 loff_t size = round_down(length, PAGE_SIZE);
1539 truncate_pagecache_range(inode, offset, offset + size - 1);
1544 ceph_zero_partial_page(inode, offset, length);
1547 static int ceph_zero_partial_object(struct inode *inode,
1548 loff_t offset, loff_t *length)
1550 struct ceph_inode_info *ci = ceph_inode(inode);
1551 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1552 struct ceph_osd_request *req;
1558 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1561 op = CEPH_OSD_OP_ZERO;
1564 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1568 CEPH_OSD_FLAG_WRITE,
1575 req->r_mtime = inode->i_mtime;
1576 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1578 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1582 ceph_osdc_put_request(req);
1588 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
1591 struct ceph_inode_info *ci = ceph_inode(inode);
1592 s32 stripe_unit = ci->i_layout.stripe_unit;
1593 s32 stripe_count = ci->i_layout.stripe_count;
1594 s32 object_size = ci->i_layout.object_size;
1595 u64 object_set_size = object_size * stripe_count;
1598 /* round offset up to next period boundary */
1599 nearly = offset + object_set_size - 1;
1601 nearly -= do_div(t, object_set_size);
1603 while (length && offset < nearly) {
1604 loff_t size = length;
1605 ret = ceph_zero_partial_object(inode, offset, &size);
1611 while (length >= object_set_size) {
1613 loff_t pos = offset;
1614 for (i = 0; i < stripe_count; ++i) {
1615 ret = ceph_zero_partial_object(inode, pos, NULL);
1620 offset += object_set_size;
1621 length -= object_set_size;
1624 loff_t size = length;
1625 ret = ceph_zero_partial_object(inode, offset, &size);
1634 static long ceph_fallocate(struct file *file, int mode,
1635 loff_t offset, loff_t length)
1637 struct ceph_file_info *fi = file->private_data;
1638 struct inode *inode = file_inode(file);
1639 struct ceph_inode_info *ci = ceph_inode(inode);
1640 struct ceph_osd_client *osdc =
1641 &ceph_inode_to_client(inode)->client->osdc;
1642 struct ceph_cap_flush *prealloc_cf;
1649 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1652 if (!S_ISREG(inode->i_mode))
1655 prealloc_cf = ceph_alloc_cap_flush();
1661 if (ceph_snap(inode) != CEPH_NOSNAP) {
1666 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) &&
1667 !(mode & FALLOC_FL_PUNCH_HOLE)) {
1672 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1673 ret = ceph_uninline_data(file, NULL);
1678 size = i_size_read(inode);
1679 if (!(mode & FALLOC_FL_KEEP_SIZE)) {
1680 endoff = offset + length;
1681 ret = inode_newsize_ok(inode, endoff);
1686 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1687 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1689 want = CEPH_CAP_FILE_BUFFER;
1691 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, endoff, &got, NULL);
1695 if (mode & FALLOC_FL_PUNCH_HOLE) {
1697 ceph_zero_pagecache_range(inode, offset, length);
1698 ret = ceph_zero_objects(inode, offset, length);
1699 } else if (endoff > size) {
1700 truncate_pagecache_range(inode, size, -1);
1701 if (ceph_inode_set_size(inode, endoff))
1702 ceph_check_caps(ceph_inode(inode),
1703 CHECK_CAPS_AUTHONLY, NULL);
1707 spin_lock(&ci->i_ceph_lock);
1708 ci->i_inline_version = CEPH_INLINE_NONE;
1709 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1711 spin_unlock(&ci->i_ceph_lock);
1713 __mark_inode_dirty(inode, dirty);
1716 ceph_put_cap_refs(ci, got);
1718 inode_unlock(inode);
1719 ceph_free_cap_flush(prealloc_cf);
1723 const struct file_operations ceph_file_fops = {
1725 .release = ceph_release,
1726 .llseek = ceph_llseek,
1727 .read_iter = ceph_read_iter,
1728 .write_iter = ceph_write_iter,
1730 .fsync = ceph_fsync,
1732 .flock = ceph_flock,
1733 .splice_read = generic_file_splice_read,
1734 .splice_write = iter_file_splice_write,
1735 .unlocked_ioctl = ceph_ioctl,
1736 .compat_ioctl = ceph_ioctl,
1737 .fallocate = ceph_fallocate,