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, true);
287 int issued = __ceph_caps_issued(ci, NULL);
289 dout("open %p fmode %d want %s issued %s using existing\n",
290 inode, fmode, ceph_cap_string(wanted),
291 ceph_cap_string(issued));
292 __ceph_get_fmode(ci, fmode);
293 spin_unlock(&ci->i_ceph_lock);
296 if ((issued & wanted) != wanted &&
297 (mds_wanted & wanted) != wanted &&
298 ceph_snap(inode) != CEPH_SNAPDIR)
299 ceph_check_caps(ci, 0, NULL);
301 return ceph_init_file(inode, file, fmode);
302 } else if (ceph_snap(inode) != CEPH_NOSNAP &&
303 (ci->i_snap_caps & wanted) == wanted) {
304 __ceph_get_fmode(ci, fmode);
305 spin_unlock(&ci->i_ceph_lock);
306 return ceph_init_file(inode, file, fmode);
309 spin_unlock(&ci->i_ceph_lock);
311 dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
312 req = prepare_open_request(inode->i_sb, flags, 0);
317 req->r_inode = inode;
321 err = ceph_mdsc_do_request(mdsc, NULL, req);
323 err = ceph_init_file(inode, file, req->r_fmode);
324 ceph_mdsc_put_request(req);
325 dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
332 * Do a lookup + open with a single request. If we get a non-existent
333 * file or symlink, return 1 so the VFS can retry.
335 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
336 struct file *file, unsigned flags, umode_t mode,
339 struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
340 struct ceph_mds_client *mdsc = fsc->mdsc;
341 struct ceph_mds_request *req;
343 struct ceph_acls_info acls = {};
347 dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
349 d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
351 if (dentry->d_name.len > NAME_MAX)
352 return -ENAMETOOLONG;
354 if (flags & O_CREAT) {
355 err = ceph_pre_init_acls(dir, &mode, &acls);
361 req = prepare_open_request(dir->i_sb, flags, mode);
366 req->r_dentry = dget(dentry);
368 if (flags & O_CREAT) {
369 req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
370 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
372 req->r_pagelist = acls.pagelist;
373 acls.pagelist = NULL;
377 mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
378 if (ceph_security_xattr_wanted(dir))
379 mask |= CEPH_CAP_XATTR_SHARED;
380 req->r_args.open.mask = cpu_to_le32(mask);
383 set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
384 err = ceph_mdsc_do_request(mdsc,
385 (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
387 err = ceph_handle_snapdir(req, dentry, err);
391 if ((flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
392 err = ceph_handle_notrace_create(dir, dentry);
394 if (d_in_lookup(dentry)) {
395 dn = ceph_finish_lookup(req, dentry, err);
399 /* we were given a hashed negative dentry */
404 if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
405 /* make vfs retry on splice, ENOENT, or symlink */
406 dout("atomic_open finish_no_open on dn %p\n", dn);
407 err = finish_no_open(file, dn);
409 dout("atomic_open finish_open on dn %p\n", dn);
410 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
411 ceph_init_inode_acls(d_inode(dentry), &acls);
412 *opened |= FILE_CREATED;
414 err = finish_open(file, dentry, ceph_open, opened);
417 if (!req->r_err && req->r_target_inode)
418 ceph_put_fmode(ceph_inode(req->r_target_inode), req->r_fmode);
419 ceph_mdsc_put_request(req);
421 ceph_release_acls_info(&acls);
422 dout("atomic_open result=%d\n", err);
426 int ceph_release(struct inode *inode, struct file *file)
428 struct ceph_inode_info *ci = ceph_inode(inode);
429 struct ceph_file_info *cf = file->private_data;
431 dout("release inode %p file %p\n", inode, file);
432 ceph_put_fmode(ci, cf->fmode);
433 if (cf->last_readdir)
434 ceph_mdsc_put_request(cf->last_readdir);
435 kfree(cf->last_name);
437 kmem_cache_free(ceph_file_cachep, cf);
439 /* wake up anyone waiting for caps on this inode */
440 wake_up_all(&ci->i_cap_wq);
451 * Read a range of bytes striped over one or more objects. Iterate over
452 * objects we stripe over. (That's not atomic, but good enough for now.)
454 * If we get a short result from the OSD, check against i_size; we need to
455 * only return a short read to the caller if we hit EOF.
457 static int striped_read(struct inode *inode,
459 struct page **pages, int num_pages,
460 int page_align, int *checkeof)
462 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
463 struct ceph_inode_info *ci = ceph_inode(inode);
468 bool hit_stripe, was_short;
471 * we may need to do multiple reads. not atomic, unfortunately.
475 page_idx = (page_align + read) >> PAGE_SHIFT;
476 ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode),
477 &ci->i_layout, pos, &this_len,
478 ci->i_truncate_seq, ci->i_truncate_size,
479 pages + page_idx, num_pages - page_idx,
480 ((page_align + read) & ~PAGE_MASK));
483 hit_stripe = this_len < len;
484 was_short = ret >= 0 && ret < this_len;
485 dout("striped_read %llu~%llu (read %u) got %d%s%s\n", pos, len, read,
486 ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : "");
488 i_size = i_size_read(inode);
490 if (was_short && (pos + ret < i_size)) {
491 int zlen = min(this_len - ret, i_size - pos - ret);
492 int zoff = page_align + read + ret;
493 dout(" zero gap %llu to %llu\n",
494 pos + ret, pos + ret + zlen);
495 ceph_zero_page_vector_range(zoff, zlen, pages);
503 /* hit stripe and need continue*/
504 if (len && hit_stripe && pos < i_size)
510 /* did we bounce off eof? */
511 if (pos + len > i_size)
512 *checkeof = CHECK_EOF;
515 dout("striped_read returns %d\n", ret);
520 * Completely synchronous read and write methods. Direct from __user
521 * buffer to osd, or directly to user pages (if O_DIRECT).
523 * If the read spans object boundary, just do multiple reads.
525 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
528 struct file *file = iocb->ki_filp;
529 struct inode *inode = file_inode(file);
531 u64 off = iocb->ki_pos;
534 size_t len = iov_iter_count(to);
536 dout("sync_read on file %p %llu~%u %s\n", file, off,
538 (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
543 * flush any page cache pages in this range. this
544 * will make concurrent normal and sync io slow,
545 * but it will at least behave sensibly when they are
548 ret = filemap_write_and_wait_range(inode->i_mapping, off,
553 if (unlikely(to->type & ITER_PIPE)) {
555 ret = iov_iter_get_pages_alloc(to, &pages, len,
559 num_pages = DIV_ROUND_UP(ret + page_off, PAGE_SIZE);
561 ret = striped_read(inode, off, ret, pages, num_pages,
564 iov_iter_advance(to, ret);
567 iov_iter_advance(to, 0);
569 ceph_put_page_vector(pages, num_pages, false);
571 num_pages = calc_pages_for(off, len);
572 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
574 return PTR_ERR(pages);
576 ret = striped_read(inode, off, len, pages, num_pages,
577 (off & ~PAGE_MASK), checkeof);
583 size_t page_off = off & ~PAGE_MASK;
584 size_t copy = min_t(size_t, left,
585 PAGE_SIZE - page_off);
586 l = copy_page_to_iter(pages[k++], page_off,
594 ceph_release_page_vector(pages, num_pages);
597 if (off > iocb->ki_pos) {
598 ret = off - iocb->ki_pos;
602 dout("sync_read result %zd\n", ret);
606 struct ceph_aio_request {
611 struct list_head osd_reqs;
613 atomic_t pending_reqs;
614 struct timespec mtime;
615 struct ceph_cap_flush *prealloc_cf;
618 struct ceph_aio_work {
619 struct work_struct work;
620 struct ceph_osd_request *req;
623 static void ceph_aio_retry_work(struct work_struct *work);
625 static void ceph_aio_complete(struct inode *inode,
626 struct ceph_aio_request *aio_req)
628 struct ceph_inode_info *ci = ceph_inode(inode);
631 if (!atomic_dec_and_test(&aio_req->pending_reqs))
634 ret = aio_req->error;
636 ret = aio_req->total_len;
638 dout("ceph_aio_complete %p rc %d\n", inode, ret);
640 if (ret >= 0 && aio_req->write) {
643 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
644 if (endoff > i_size_read(inode)) {
645 if (ceph_inode_set_size(inode, endoff))
646 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
649 spin_lock(&ci->i_ceph_lock);
650 ci->i_inline_version = CEPH_INLINE_NONE;
651 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
652 &aio_req->prealloc_cf);
653 spin_unlock(&ci->i_ceph_lock);
655 __mark_inode_dirty(inode, dirty);
659 ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
662 aio_req->iocb->ki_complete(aio_req->iocb, ret, 0);
664 ceph_free_cap_flush(aio_req->prealloc_cf);
668 static void ceph_aio_complete_req(struct ceph_osd_request *req)
670 int rc = req->r_result;
671 struct inode *inode = req->r_inode;
672 struct ceph_aio_request *aio_req = req->r_priv;
673 struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
674 int num_pages = calc_pages_for((u64)osd_data->alignment,
677 dout("ceph_aio_complete_req %p rc %d bytes %llu\n",
678 inode, rc, osd_data->length);
680 if (rc == -EOLDSNAPC) {
681 struct ceph_aio_work *aio_work;
682 BUG_ON(!aio_req->write);
684 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
686 INIT_WORK(&aio_work->work, ceph_aio_retry_work);
688 queue_work(ceph_inode_to_client(inode)->wb_wq,
693 } else if (!aio_req->write) {
696 if (rc >= 0 && osd_data->length > rc) {
697 int zoff = osd_data->alignment + rc;
698 int zlen = osd_data->length - rc;
700 * If read is satisfied by single OSD request,
701 * it can pass EOF. Otherwise read is within
704 if (aio_req->num_reqs == 1) {
705 loff_t i_size = i_size_read(inode);
706 loff_t endoff = aio_req->iocb->ki_pos + rc;
708 zlen = min_t(size_t, zlen,
710 aio_req->total_len = rc + zlen;
714 ceph_zero_page_vector_range(zoff, zlen,
719 ceph_put_page_vector(osd_data->pages, num_pages, !aio_req->write);
720 ceph_osdc_put_request(req);
723 cmpxchg(&aio_req->error, 0, rc);
725 ceph_aio_complete(inode, aio_req);
729 static void ceph_aio_retry_work(struct work_struct *work)
731 struct ceph_aio_work *aio_work =
732 container_of(work, struct ceph_aio_work, work);
733 struct ceph_osd_request *orig_req = aio_work->req;
734 struct ceph_aio_request *aio_req = orig_req->r_priv;
735 struct inode *inode = orig_req->r_inode;
736 struct ceph_inode_info *ci = ceph_inode(inode);
737 struct ceph_snap_context *snapc;
738 struct ceph_osd_request *req;
741 spin_lock(&ci->i_ceph_lock);
742 if (__ceph_have_pending_cap_snap(ci)) {
743 struct ceph_cap_snap *capsnap =
744 list_last_entry(&ci->i_cap_snaps,
745 struct ceph_cap_snap,
747 snapc = ceph_get_snap_context(capsnap->context);
749 BUG_ON(!ci->i_head_snapc);
750 snapc = ceph_get_snap_context(ci->i_head_snapc);
752 spin_unlock(&ci->i_ceph_lock);
754 req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 2,
762 req->r_flags = CEPH_OSD_FLAG_ORDERSNAP | CEPH_OSD_FLAG_WRITE;
763 ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
764 ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
766 ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
768 ceph_osdc_put_request(req);
773 req->r_ops[0] = orig_req->r_ops[0];
774 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC, 0);
776 req->r_mtime = aio_req->mtime;
777 req->r_data_offset = req->r_ops[0].extent.offset;
779 ceph_osdc_put_request(orig_req);
781 req->r_callback = ceph_aio_complete_req;
782 req->r_inode = inode;
783 req->r_priv = aio_req;
785 ret = ceph_osdc_start_request(req->r_osdc, req, false);
789 ceph_aio_complete_req(req);
792 ceph_put_snap_context(snapc);
797 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
798 struct ceph_snap_context *snapc,
799 struct ceph_cap_flush **pcf)
801 struct file *file = iocb->ki_filp;
802 struct inode *inode = file_inode(file);
803 struct ceph_inode_info *ci = ceph_inode(inode);
804 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
805 struct ceph_vino vino;
806 struct ceph_osd_request *req;
808 struct ceph_aio_request *aio_req = NULL;
812 struct timespec mtime = current_time(inode);
813 size_t count = iov_iter_count(iter);
814 loff_t pos = iocb->ki_pos;
815 bool write = iov_iter_rw(iter) == WRITE;
817 if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
820 dout("sync_direct_read_write (%s) on file %p %lld~%u\n",
821 (write ? "write" : "read"), file, pos, (unsigned)count);
823 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
828 int ret2 = invalidate_inode_pages2_range(inode->i_mapping,
830 (pos + count) >> PAGE_SHIFT);
832 dout("invalidate_inode_pages2_range returned %d\n", ret2);
834 flags = CEPH_OSD_FLAG_ORDERSNAP | CEPH_OSD_FLAG_WRITE;
836 flags = CEPH_OSD_FLAG_READ;
839 while (iov_iter_count(iter) > 0) {
840 u64 size = dio_get_pagev_size(iter);
844 vino = ceph_vino(inode);
845 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
847 /*include a 'startsync' command*/
849 write ? CEPH_OSD_OP_WRITE :
861 pages = dio_get_pages_alloc(iter, len, &start, &num_pages);
863 ceph_osdc_put_request(req);
864 ret = PTR_ERR(pages);
869 * To simplify error handling, allow AIO when IO within i_size
870 * or IO can be satisfied by single OSD request.
872 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
873 (len == count || pos + count <= i_size_read(inode))) {
874 aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
876 aio_req->iocb = iocb;
877 aio_req->write = write;
878 INIT_LIST_HEAD(&aio_req->osd_reqs);
880 aio_req->mtime = mtime;
881 swap(aio_req->prealloc_cf, *pcf);
889 * throw out any page cache pages in this range. this
892 truncate_inode_pages_range(inode->i_mapping, pos,
893 (pos+len) | (PAGE_SIZE - 1));
895 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC, 0);
896 req->r_mtime = mtime;
899 osd_req_op_extent_osd_data_pages(req, 0, pages, len, start,
903 aio_req->total_len += len;
905 atomic_inc(&aio_req->pending_reqs);
907 req->r_callback = ceph_aio_complete_req;
908 req->r_inode = inode;
909 req->r_priv = aio_req;
910 list_add_tail(&req->r_unsafe_item, &aio_req->osd_reqs);
913 iov_iter_advance(iter, len);
917 ret = ceph_osdc_start_request(req->r_osdc, req, false);
919 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
921 size = i_size_read(inode);
925 if (ret >= 0 && ret < len && pos + ret < size) {
926 int zlen = min_t(size_t, len - ret,
928 ceph_zero_page_vector_range(start + ret, zlen,
936 ceph_put_page_vector(pages, num_pages, !write);
938 ceph_osdc_put_request(req);
943 iov_iter_advance(iter, len);
945 if (!write && pos >= size)
948 if (write && pos > size) {
949 if (ceph_inode_set_size(inode, pos))
950 ceph_check_caps(ceph_inode(inode),
959 if (aio_req->num_reqs == 0) {
964 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
967 list_splice(&aio_req->osd_reqs, &osd_reqs);
968 while (!list_empty(&osd_reqs)) {
969 req = list_first_entry(&osd_reqs,
970 struct ceph_osd_request,
972 list_del_init(&req->r_unsafe_item);
974 ret = ceph_osdc_start_request(req->r_osdc,
978 ceph_aio_complete_req(req);
984 if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
985 ret = pos - iocb->ki_pos;
992 * Synchronous write, straight from __user pointer or user pages.
994 * If write spans object boundary, just do multiple writes. (For a
995 * correct atomic write, we should e.g. take write locks on all
996 * objects, rollback on failure, etc.)
999 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1000 struct ceph_snap_context *snapc)
1002 struct file *file = iocb->ki_filp;
1003 struct inode *inode = file_inode(file);
1004 struct ceph_inode_info *ci = ceph_inode(inode);
1005 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1006 struct ceph_vino vino;
1007 struct ceph_osd_request *req;
1008 struct page **pages;
1015 struct timespec mtime = current_time(inode);
1016 size_t count = iov_iter_count(from);
1018 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1021 dout("sync_write on file %p %lld~%u\n", file, pos, (unsigned)count);
1023 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
1027 ret = invalidate_inode_pages2_range(inode->i_mapping,
1029 (pos + count) >> PAGE_SHIFT);
1031 dout("invalidate_inode_pages2_range returned %d\n", ret);
1033 flags = CEPH_OSD_FLAG_ORDERSNAP | CEPH_OSD_FLAG_WRITE;
1035 while ((len = iov_iter_count(from)) > 0) {
1039 vino = ceph_vino(inode);
1040 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1041 vino, pos, &len, 0, 1,
1042 CEPH_OSD_OP_WRITE, flags, snapc,
1044 ci->i_truncate_size,
1052 * write from beginning of first page,
1053 * regardless of io alignment
1055 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1057 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1058 if (IS_ERR(pages)) {
1059 ret = PTR_ERR(pages);
1064 for (n = 0; n < num_pages; n++) {
1065 size_t plen = min_t(size_t, left, PAGE_SIZE);
1066 ret = copy_page_from_iter(pages[n], 0, plen, from);
1075 ceph_release_page_vector(pages, num_pages);
1079 req->r_inode = inode;
1081 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1084 req->r_mtime = mtime;
1085 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1087 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1090 ceph_osdc_put_request(req);
1095 if (pos > i_size_read(inode)) {
1096 check_caps = ceph_inode_set_size(inode, pos);
1098 ceph_check_caps(ceph_inode(inode),
1099 CHECK_CAPS_AUTHONLY,
1106 if (ret != -EOLDSNAPC && written > 0) {
1114 * Wrap generic_file_aio_read with checks for cap bits on the inode.
1115 * Atomically grab references, so that those bits are not released
1116 * back to the MDS mid-read.
1118 * Hmm, the sync read case isn't actually async... should it be?
1120 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1122 struct file *filp = iocb->ki_filp;
1123 struct ceph_file_info *fi = filp->private_data;
1124 size_t len = iov_iter_count(to);
1125 struct inode *inode = file_inode(filp);
1126 struct ceph_inode_info *ci = ceph_inode(inode);
1127 struct page *pinned_page = NULL;
1130 int retry_op = 0, read = 0;
1133 dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1134 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1136 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1137 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1139 want = CEPH_CAP_FILE_CACHE;
1140 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1144 if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1145 (iocb->ki_flags & IOCB_DIRECT) ||
1146 (fi->flags & CEPH_F_SYNC)) {
1148 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1149 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1150 ceph_cap_string(got));
1152 if (ci->i_inline_version == CEPH_INLINE_NONE) {
1153 if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1154 ret = ceph_direct_read_write(iocb, to,
1156 if (ret >= 0 && ret < len)
1157 retry_op = CHECK_EOF;
1159 ret = ceph_sync_read(iocb, to, &retry_op);
1162 retry_op = READ_INLINE;
1165 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1166 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1167 ceph_cap_string(got));
1168 current->journal_info = filp;
1169 ret = generic_file_read_iter(iocb, to);
1170 current->journal_info = NULL;
1172 dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1173 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1175 put_page(pinned_page);
1178 ceph_put_cap_refs(ci, got);
1179 if (retry_op > HAVE_RETRIED && ret >= 0) {
1181 struct page *page = NULL;
1183 if (retry_op == READ_INLINE) {
1184 page = __page_cache_alloc(GFP_KERNEL);
1189 statret = __ceph_do_getattr(inode, page,
1190 CEPH_STAT_CAP_INLINE_DATA, !!page);
1194 if (statret == -ENODATA) {
1195 BUG_ON(retry_op != READ_INLINE);
1201 i_size = i_size_read(inode);
1202 if (retry_op == READ_INLINE) {
1203 BUG_ON(ret > 0 || read > 0);
1204 if (iocb->ki_pos < i_size &&
1205 iocb->ki_pos < PAGE_SIZE) {
1206 loff_t end = min_t(loff_t, i_size,
1207 iocb->ki_pos + len);
1208 end = min_t(loff_t, end, PAGE_SIZE);
1210 zero_user_segment(page, statret, end);
1211 ret = copy_page_to_iter(page,
1212 iocb->ki_pos & ~PAGE_MASK,
1213 end - iocb->ki_pos, to);
1214 iocb->ki_pos += ret;
1217 if (iocb->ki_pos < i_size && read < len) {
1218 size_t zlen = min_t(size_t, len - read,
1219 i_size - iocb->ki_pos);
1220 ret = iov_iter_zero(zlen, to);
1221 iocb->ki_pos += ret;
1224 __free_pages(page, 0);
1228 /* hit EOF or hole? */
1229 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1231 dout("sync_read hit hole, ppos %lld < size %lld"
1232 ", reading more\n", iocb->ki_pos, i_size);
1236 retry_op = HAVE_RETRIED;
1248 * Take cap references to avoid releasing caps to MDS mid-write.
1250 * If we are synchronous, and write with an old snap context, the OSD
1251 * may return EOLDSNAPC. In that case, retry the write.. _after_
1252 * dropping our cap refs and allowing the pending snap to logically
1253 * complete _before_ this write occurs.
1255 * If we are near ENOSPC, write synchronously.
1257 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1259 struct file *file = iocb->ki_filp;
1260 struct ceph_file_info *fi = file->private_data;
1261 struct inode *inode = file_inode(file);
1262 struct ceph_inode_info *ci = ceph_inode(inode);
1263 struct ceph_osd_client *osdc =
1264 &ceph_sb_to_client(inode->i_sb)->client->osdc;
1265 struct ceph_cap_flush *prealloc_cf;
1266 ssize_t count, written = 0;
1270 if (ceph_snap(inode) != CEPH_NOSNAP)
1273 prealloc_cf = ceph_alloc_cap_flush();
1279 /* We can write back this queue in page reclaim */
1280 current->backing_dev_info = inode_to_bdi(inode);
1282 if (iocb->ki_flags & IOCB_APPEND) {
1283 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1288 err = generic_write_checks(iocb, from);
1293 count = iov_iter_count(from);
1294 err = file_remove_privs(file);
1298 err = file_update_time(file);
1302 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1303 err = ceph_uninline_data(file, NULL);
1309 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL)) {
1314 dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1315 inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1316 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1317 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1319 want = CEPH_CAP_FILE_BUFFER;
1321 err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, pos + count,
1326 dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1327 inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1329 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1330 (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC)) {
1331 struct ceph_snap_context *snapc;
1332 struct iov_iter data;
1333 inode_unlock(inode);
1335 spin_lock(&ci->i_ceph_lock);
1336 if (__ceph_have_pending_cap_snap(ci)) {
1337 struct ceph_cap_snap *capsnap =
1338 list_last_entry(&ci->i_cap_snaps,
1339 struct ceph_cap_snap,
1341 snapc = ceph_get_snap_context(capsnap->context);
1343 BUG_ON(!ci->i_head_snapc);
1344 snapc = ceph_get_snap_context(ci->i_head_snapc);
1346 spin_unlock(&ci->i_ceph_lock);
1348 /* we might need to revert back to that point */
1350 if (iocb->ki_flags & IOCB_DIRECT)
1351 written = ceph_direct_read_write(iocb, &data, snapc,
1354 written = ceph_sync_write(iocb, &data, pos, snapc);
1355 if (written == -EOLDSNAPC) {
1356 dout("aio_write %p %llx.%llx %llu~%u"
1357 "got EOLDSNAPC, retrying\n",
1358 inode, ceph_vinop(inode),
1359 pos, (unsigned)count);
1364 iov_iter_advance(from, written);
1365 ceph_put_snap_context(snapc);
1368 * No need to acquire the i_truncate_mutex. Because
1369 * the MDS revokes Fwb caps before sending truncate
1370 * message to us. We can't get Fwb cap while there
1371 * are pending vmtruncate. So write and vmtruncate
1372 * can not run at the same time
1374 written = generic_perform_write(file, from, pos);
1375 if (likely(written >= 0))
1376 iocb->ki_pos = pos + written;
1377 inode_unlock(inode);
1382 spin_lock(&ci->i_ceph_lock);
1383 ci->i_inline_version = CEPH_INLINE_NONE;
1384 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1386 spin_unlock(&ci->i_ceph_lock);
1388 __mark_inode_dirty(inode, dirty);
1391 dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n",
1392 inode, ceph_vinop(inode), pos, (unsigned)count,
1393 ceph_cap_string(got));
1394 ceph_put_cap_refs(ci, got);
1397 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_NEARFULL))
1398 iocb->ki_flags |= IOCB_DSYNC;
1400 written = generic_write_sync(iocb, written);
1406 inode_unlock(inode);
1408 ceph_free_cap_flush(prealloc_cf);
1409 current->backing_dev_info = NULL;
1410 return written ? written : err;
1414 * llseek. be sure to verify file size on SEEK_END.
1416 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1418 struct inode *inode = file->f_mapping->host;
1424 if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1425 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1430 i_size = i_size_read(inode);
1437 * Here we special-case the lseek(fd, 0, SEEK_CUR)
1438 * position-querying operation. Avoid rewriting the "same"
1439 * f_pos value back to the file because a concurrent read(),
1440 * write() or lseek() might have altered it
1446 offset += file->f_pos;
1449 if (offset >= i_size) {
1455 if (offset >= i_size) {
1463 ret = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1466 inode_unlock(inode);
1470 static inline void ceph_zero_partial_page(
1471 struct inode *inode, loff_t offset, unsigned size)
1474 pgoff_t index = offset >> PAGE_SHIFT;
1476 page = find_lock_page(inode->i_mapping, index);
1478 wait_on_page_writeback(page);
1479 zero_user(page, offset & (PAGE_SIZE - 1), size);
1485 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1488 loff_t nearly = round_up(offset, PAGE_SIZE);
1489 if (offset < nearly) {
1490 loff_t size = nearly - offset;
1493 ceph_zero_partial_page(inode, offset, size);
1497 if (length >= PAGE_SIZE) {
1498 loff_t size = round_down(length, PAGE_SIZE);
1499 truncate_pagecache_range(inode, offset, offset + size - 1);
1504 ceph_zero_partial_page(inode, offset, length);
1507 static int ceph_zero_partial_object(struct inode *inode,
1508 loff_t offset, loff_t *length)
1510 struct ceph_inode_info *ci = ceph_inode(inode);
1511 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1512 struct ceph_osd_request *req;
1518 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1521 op = CEPH_OSD_OP_ZERO;
1524 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1528 CEPH_OSD_FLAG_WRITE,
1535 req->r_mtime = inode->i_mtime;
1536 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1538 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1542 ceph_osdc_put_request(req);
1548 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
1551 struct ceph_inode_info *ci = ceph_inode(inode);
1552 s32 stripe_unit = ci->i_layout.stripe_unit;
1553 s32 stripe_count = ci->i_layout.stripe_count;
1554 s32 object_size = ci->i_layout.object_size;
1555 u64 object_set_size = object_size * stripe_count;
1558 /* round offset up to next period boundary */
1559 nearly = offset + object_set_size - 1;
1561 nearly -= do_div(t, object_set_size);
1563 while (length && offset < nearly) {
1564 loff_t size = length;
1565 ret = ceph_zero_partial_object(inode, offset, &size);
1571 while (length >= object_set_size) {
1573 loff_t pos = offset;
1574 for (i = 0; i < stripe_count; ++i) {
1575 ret = ceph_zero_partial_object(inode, pos, NULL);
1580 offset += object_set_size;
1581 length -= object_set_size;
1584 loff_t size = length;
1585 ret = ceph_zero_partial_object(inode, offset, &size);
1594 static long ceph_fallocate(struct file *file, int mode,
1595 loff_t offset, loff_t length)
1597 struct ceph_file_info *fi = file->private_data;
1598 struct inode *inode = file_inode(file);
1599 struct ceph_inode_info *ci = ceph_inode(inode);
1600 struct ceph_osd_client *osdc =
1601 &ceph_inode_to_client(inode)->client->osdc;
1602 struct ceph_cap_flush *prealloc_cf;
1609 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1612 if (!S_ISREG(inode->i_mode))
1615 prealloc_cf = ceph_alloc_cap_flush();
1621 if (ceph_snap(inode) != CEPH_NOSNAP) {
1626 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) &&
1627 !(mode & FALLOC_FL_PUNCH_HOLE)) {
1632 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1633 ret = ceph_uninline_data(file, NULL);
1638 size = i_size_read(inode);
1639 if (!(mode & FALLOC_FL_KEEP_SIZE))
1640 endoff = offset + length;
1642 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1643 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1645 want = CEPH_CAP_FILE_BUFFER;
1647 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, endoff, &got, NULL);
1651 if (mode & FALLOC_FL_PUNCH_HOLE) {
1653 ceph_zero_pagecache_range(inode, offset, length);
1654 ret = ceph_zero_objects(inode, offset, length);
1655 } else if (endoff > size) {
1656 truncate_pagecache_range(inode, size, -1);
1657 if (ceph_inode_set_size(inode, endoff))
1658 ceph_check_caps(ceph_inode(inode),
1659 CHECK_CAPS_AUTHONLY, NULL);
1663 spin_lock(&ci->i_ceph_lock);
1664 ci->i_inline_version = CEPH_INLINE_NONE;
1665 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1667 spin_unlock(&ci->i_ceph_lock);
1669 __mark_inode_dirty(inode, dirty);
1672 ceph_put_cap_refs(ci, got);
1674 inode_unlock(inode);
1675 ceph_free_cap_flush(prealloc_cf);
1679 const struct file_operations ceph_file_fops = {
1681 .release = ceph_release,
1682 .llseek = ceph_llseek,
1683 .read_iter = ceph_read_iter,
1684 .write_iter = ceph_write_iter,
1686 .fsync = ceph_fsync,
1688 .flock = ceph_flock,
1689 .splice_read = generic_file_splice_read,
1690 .splice_write = iter_file_splice_write,
1691 .unlocked_ioctl = ceph_ioctl,
1692 .compat_ioctl = ceph_ioctl,
1693 .fallocate = ceph_fallocate,