1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
4 #include <linux/module.h>
5 #include <linux/sched.h>
6 #include <linux/slab.h>
7 #include <linux/file.h>
8 #include <linux/mount.h>
9 #include <linux/namei.h>
10 #include <linux/writeback.h>
11 #include <linux/falloc.h>
14 #include "mds_client.h"
17 static __le32 ceph_flags_sys2wire(u32 flags)
21 switch (flags & O_ACCMODE) {
23 wire_flags |= CEPH_O_RDONLY;
26 wire_flags |= CEPH_O_WRONLY;
29 wire_flags |= CEPH_O_RDWR;
33 #define ceph_sys2wire(a) if (flags & a) { wire_flags |= CEPH_##a; flags &= ~a; }
35 ceph_sys2wire(O_CREAT);
36 ceph_sys2wire(O_EXCL);
37 ceph_sys2wire(O_TRUNC);
38 ceph_sys2wire(O_DIRECTORY);
39 ceph_sys2wire(O_NOFOLLOW);
44 dout("unused open flags: %x", flags);
46 return cpu_to_le32(wire_flags);
50 * Ceph file operations
52 * Implement basic open/close functionality, and implement
55 * We implement three modes of file I/O:
56 * - buffered uses the generic_file_aio_{read,write} helpers
58 * - synchronous is used when there is multi-client read/write
59 * sharing, avoids the page cache, and synchronously waits for an
62 * - direct io takes the variant of the sync path that references
63 * user pages directly.
65 * fsync() flushes and waits on dirty pages, but just queues metadata
66 * for writeback: since the MDS can recover size and mtime there is no
67 * need to wait for MDS acknowledgement.
71 * Calculate the length sum of direct io vectors that can
72 * be combined into one page vector.
74 static size_t dio_get_pagev_size(const struct iov_iter *it)
76 const struct iovec *iov = it->iov;
77 const struct iovec *iovend = iov + it->nr_segs;
80 size = iov->iov_len - it->iov_offset;
82 * An iov can be page vectored when both the current tail
83 * and the next base are page aligned.
85 while (PAGE_ALIGNED((iov->iov_base + iov->iov_len)) &&
86 (++iov < iovend && PAGE_ALIGNED((iov->iov_base)))) {
89 dout("dio_get_pagevlen len = %zu\n", size);
94 * Allocate a page vector based on (@it, @nbytes).
95 * The return value is the tuple describing a page vector,
96 * that is (@pages, @page_align, @num_pages).
99 dio_get_pages_alloc(const struct iov_iter *it, size_t nbytes,
100 size_t *page_align, int *num_pages)
102 struct iov_iter tmp_it = *it;
105 int ret = 0, idx, npages;
107 align = (unsigned long)(it->iov->iov_base + it->iov_offset) &
109 npages = calc_pages_for(align, nbytes);
110 pages = kvmalloc(sizeof(*pages) * npages, GFP_KERNEL);
112 return ERR_PTR(-ENOMEM);
114 for (idx = 0; idx < npages; ) {
116 ret = iov_iter_get_pages(&tmp_it, pages + idx, nbytes,
117 npages - idx, &start);
121 iov_iter_advance(&tmp_it, ret);
123 idx += (ret + start + PAGE_SIZE - 1) / PAGE_SIZE;
129 dout("dio_get_pages_alloc: got %d pages align %zu\n", npages, align);
132 ceph_put_page_vector(pages, idx, false);
137 * Prepare an open request. Preallocate ceph_cap to avoid an
138 * inopportune ENOMEM later.
140 static struct ceph_mds_request *
141 prepare_open_request(struct super_block *sb, int flags, int create_mode)
143 struct ceph_fs_client *fsc = ceph_sb_to_client(sb);
144 struct ceph_mds_client *mdsc = fsc->mdsc;
145 struct ceph_mds_request *req;
146 int want_auth = USE_ANY_MDS;
147 int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
149 if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
150 want_auth = USE_AUTH_MDS;
152 req = ceph_mdsc_create_request(mdsc, op, want_auth);
155 req->r_fmode = ceph_flags_to_mode(flags);
156 req->r_args.open.flags = ceph_flags_sys2wire(flags);
157 req->r_args.open.mode = cpu_to_le32(create_mode);
163 * initialize private struct file data.
164 * if we fail, clean up by dropping fmode reference on the ceph_inode
166 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
168 struct ceph_file_info *cf;
171 switch (inode->i_mode & S_IFMT) {
173 ceph_fscache_register_inode_cookie(inode);
174 ceph_fscache_file_set_cookie(inode, file);
176 dout("init_file %p %p 0%o (regular)\n", inode, file,
178 cf = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
180 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
185 cf->readdir_cache_idx = -1;
186 file->private_data = cf;
187 BUG_ON(inode->i_fop->release != ceph_release);
191 dout("init_file %p %p 0%o (symlink)\n", inode, file,
193 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
197 dout("init_file %p %p 0%o (special)\n", inode, file,
200 * we need to drop the open ref now, since we don't
201 * have .release set to ceph_release.
203 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
204 BUG_ON(inode->i_fop->release == ceph_release);
206 /* call the proper open fop */
207 ret = inode->i_fop->open(inode, file);
213 * try renew caps after session gets killed.
215 int ceph_renew_caps(struct inode *inode)
217 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
218 struct ceph_inode_info *ci = ceph_inode(inode);
219 struct ceph_mds_request *req;
220 int err, flags, wanted;
222 spin_lock(&ci->i_ceph_lock);
223 wanted = __ceph_caps_file_wanted(ci);
224 if (__ceph_is_any_real_caps(ci) &&
225 (!(wanted & CEPH_CAP_ANY_WR) || ci->i_auth_cap)) {
226 int issued = __ceph_caps_issued(ci, NULL);
227 spin_unlock(&ci->i_ceph_lock);
228 dout("renew caps %p want %s issued %s updating mds_wanted\n",
229 inode, ceph_cap_string(wanted), ceph_cap_string(issued));
230 ceph_check_caps(ci, 0, NULL);
233 spin_unlock(&ci->i_ceph_lock);
236 if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
238 else if (wanted & CEPH_CAP_FILE_RD)
240 else if (wanted & CEPH_CAP_FILE_WR)
243 if (wanted & CEPH_CAP_FILE_LAZYIO)
247 req = prepare_open_request(inode->i_sb, flags, 0);
253 req->r_inode = inode;
258 err = ceph_mdsc_do_request(mdsc, NULL, req);
259 ceph_mdsc_put_request(req);
261 dout("renew caps %p open result=%d\n", inode, err);
262 return err < 0 ? err : 0;
266 * If we already have the requisite capabilities, we can satisfy
267 * the open request locally (no need to request new caps from the
268 * MDS). We do, however, need to inform the MDS (asynchronously)
269 * if our wanted caps set expands.
271 int ceph_open(struct inode *inode, struct file *file)
273 struct ceph_inode_info *ci = ceph_inode(inode);
274 struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
275 struct ceph_mds_client *mdsc = fsc->mdsc;
276 struct ceph_mds_request *req;
277 struct ceph_file_info *cf = file->private_data;
279 int flags, fmode, wanted;
282 dout("open file %p is already opened\n", file);
286 /* filter out O_CREAT|O_EXCL; vfs did that already. yuck. */
287 flags = file->f_flags & ~(O_CREAT|O_EXCL);
288 if (S_ISDIR(inode->i_mode))
289 flags = O_DIRECTORY; /* mds likes to know */
291 dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
292 ceph_vinop(inode), file, flags, file->f_flags);
293 fmode = ceph_flags_to_mode(flags);
294 wanted = ceph_caps_for_mode(fmode);
296 /* snapped files are read-only */
297 if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
300 /* trivially open snapdir */
301 if (ceph_snap(inode) == CEPH_SNAPDIR) {
302 spin_lock(&ci->i_ceph_lock);
303 __ceph_get_fmode(ci, fmode);
304 spin_unlock(&ci->i_ceph_lock);
305 return ceph_init_file(inode, file, fmode);
309 * No need to block if we have caps on the auth MDS (for
310 * write) or any MDS (for read). Update wanted set
313 spin_lock(&ci->i_ceph_lock);
314 if (__ceph_is_any_real_caps(ci) &&
315 (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
316 int mds_wanted = __ceph_caps_mds_wanted(ci, true);
317 int issued = __ceph_caps_issued(ci, NULL);
319 dout("open %p fmode %d want %s issued %s using existing\n",
320 inode, fmode, ceph_cap_string(wanted),
321 ceph_cap_string(issued));
322 __ceph_get_fmode(ci, fmode);
323 spin_unlock(&ci->i_ceph_lock);
326 if ((issued & wanted) != wanted &&
327 (mds_wanted & wanted) != wanted &&
328 ceph_snap(inode) != CEPH_SNAPDIR)
329 ceph_check_caps(ci, 0, NULL);
331 return ceph_init_file(inode, file, fmode);
332 } else if (ceph_snap(inode) != CEPH_NOSNAP &&
333 (ci->i_snap_caps & wanted) == wanted) {
334 __ceph_get_fmode(ci, fmode);
335 spin_unlock(&ci->i_ceph_lock);
336 return ceph_init_file(inode, file, fmode);
339 spin_unlock(&ci->i_ceph_lock);
341 dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
342 req = prepare_open_request(inode->i_sb, flags, 0);
347 req->r_inode = inode;
351 err = ceph_mdsc_do_request(mdsc, NULL, req);
353 err = ceph_init_file(inode, file, req->r_fmode);
354 ceph_mdsc_put_request(req);
355 dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
362 * Do a lookup + open with a single request. If we get a non-existent
363 * file or symlink, return 1 so the VFS can retry.
365 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
366 struct file *file, unsigned flags, umode_t mode,
369 struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
370 struct ceph_mds_client *mdsc = fsc->mdsc;
371 struct ceph_mds_request *req;
373 struct ceph_acls_info acls = {};
377 dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
379 d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
381 if (dentry->d_name.len > NAME_MAX)
382 return -ENAMETOOLONG;
384 if (flags & O_CREAT) {
385 err = ceph_pre_init_acls(dir, &mode, &acls);
391 req = prepare_open_request(dir->i_sb, flags, mode);
396 req->r_dentry = dget(dentry);
398 if (flags & O_CREAT) {
399 req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
400 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
402 req->r_pagelist = acls.pagelist;
403 acls.pagelist = NULL;
407 mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
408 if (ceph_security_xattr_wanted(dir))
409 mask |= CEPH_CAP_XATTR_SHARED;
410 req->r_args.open.mask = cpu_to_le32(mask);
413 set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
414 err = ceph_mdsc_do_request(mdsc,
415 (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
417 err = ceph_handle_snapdir(req, dentry, err);
421 if ((flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
422 err = ceph_handle_notrace_create(dir, dentry);
424 if (d_in_lookup(dentry)) {
425 dn = ceph_finish_lookup(req, dentry, err);
429 /* we were given a hashed negative dentry */
434 if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
435 /* make vfs retry on splice, ENOENT, or symlink */
436 dout("atomic_open finish_no_open on dn %p\n", dn);
437 err = finish_no_open(file, dn);
439 dout("atomic_open finish_open on dn %p\n", dn);
440 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
441 ceph_init_inode_acls(d_inode(dentry), &acls);
442 *opened |= FILE_CREATED;
444 err = finish_open(file, dentry, ceph_open, opened);
447 if (!req->r_err && req->r_target_inode)
448 ceph_put_fmode(ceph_inode(req->r_target_inode), req->r_fmode);
449 ceph_mdsc_put_request(req);
451 ceph_release_acls_info(&acls);
452 dout("atomic_open result=%d\n", err);
456 int ceph_release(struct inode *inode, struct file *file)
458 struct ceph_inode_info *ci = ceph_inode(inode);
459 struct ceph_file_info *cf = file->private_data;
461 dout("release inode %p file %p\n", inode, file);
462 ceph_put_fmode(ci, cf->fmode);
463 if (cf->last_readdir)
464 ceph_mdsc_put_request(cf->last_readdir);
465 kfree(cf->last_name);
467 kmem_cache_free(ceph_file_cachep, cf);
469 /* wake up anyone waiting for caps on this inode */
470 wake_up_all(&ci->i_cap_wq);
481 * Read a range of bytes striped over one or more objects. Iterate over
482 * objects we stripe over. (That's not atomic, but good enough for now.)
484 * If we get a short result from the OSD, check against i_size; we need to
485 * only return a short read to the caller if we hit EOF.
487 static int striped_read(struct inode *inode,
489 struct page **pages, int num_pages,
490 int page_align, int *checkeof)
492 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
493 struct ceph_inode_info *ci = ceph_inode(inode);
498 bool hit_stripe, was_short;
501 * we may need to do multiple reads. not atomic, unfortunately.
505 page_idx = (page_align + read) >> PAGE_SHIFT;
506 ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode),
507 &ci->i_layout, pos, &this_len,
508 ci->i_truncate_seq, ci->i_truncate_size,
509 pages + page_idx, num_pages - page_idx,
510 ((page_align + read) & ~PAGE_MASK));
513 hit_stripe = this_len < len;
514 was_short = ret >= 0 && ret < this_len;
515 dout("striped_read %llu~%llu (read %u) got %d%s%s\n", pos, len, read,
516 ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : "");
518 i_size = i_size_read(inode);
520 if (was_short && (pos + ret < i_size)) {
521 int zlen = min(this_len - ret, i_size - pos - ret);
522 int zoff = page_align + read + ret;
523 dout(" zero gap %llu to %llu\n",
524 pos + ret, pos + ret + zlen);
525 ceph_zero_page_vector_range(zoff, zlen, pages);
533 /* hit stripe and need continue*/
534 if (len && hit_stripe && pos < i_size)
540 /* did we bounce off eof? */
541 if (pos + len > i_size)
542 *checkeof = CHECK_EOF;
545 dout("striped_read returns %d\n", ret);
550 * Completely synchronous read and write methods. Direct from __user
551 * buffer to osd, or directly to user pages (if O_DIRECT).
553 * If the read spans object boundary, just do multiple reads.
555 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
558 struct file *file = iocb->ki_filp;
559 struct inode *inode = file_inode(file);
561 u64 off = iocb->ki_pos;
564 size_t len = iov_iter_count(to);
566 dout("sync_read on file %p %llu~%u %s\n", file, off, (unsigned)len,
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 {
641 struct list_head osd_reqs;
643 atomic_t pending_reqs;
644 struct timespec mtime;
645 struct ceph_cap_flush *prealloc_cf;
648 struct ceph_aio_work {
649 struct work_struct work;
650 struct ceph_osd_request *req;
653 static void ceph_aio_retry_work(struct work_struct *work);
655 static void ceph_aio_complete(struct inode *inode,
656 struct ceph_aio_request *aio_req)
658 struct ceph_inode_info *ci = ceph_inode(inode);
661 if (!atomic_dec_and_test(&aio_req->pending_reqs))
664 ret = aio_req->error;
666 ret = aio_req->total_len;
668 dout("ceph_aio_complete %p rc %d\n", inode, ret);
670 if (ret >= 0 && aio_req->write) {
673 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
674 if (endoff > i_size_read(inode)) {
675 if (ceph_inode_set_size(inode, endoff))
676 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
679 spin_lock(&ci->i_ceph_lock);
680 ci->i_inline_version = CEPH_INLINE_NONE;
681 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
682 &aio_req->prealloc_cf);
683 spin_unlock(&ci->i_ceph_lock);
685 __mark_inode_dirty(inode, dirty);
689 ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
692 aio_req->iocb->ki_complete(aio_req->iocb, ret, 0);
694 ceph_free_cap_flush(aio_req->prealloc_cf);
698 static void ceph_aio_complete_req(struct ceph_osd_request *req)
700 int rc = req->r_result;
701 struct inode *inode = req->r_inode;
702 struct ceph_aio_request *aio_req = req->r_priv;
703 struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
704 int num_pages = calc_pages_for((u64)osd_data->alignment,
707 dout("ceph_aio_complete_req %p rc %d bytes %llu\n",
708 inode, rc, osd_data->length);
710 if (rc == -EOLDSNAPC) {
711 struct ceph_aio_work *aio_work;
712 BUG_ON(!aio_req->write);
714 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
716 INIT_WORK(&aio_work->work, ceph_aio_retry_work);
718 queue_work(ceph_inode_to_client(inode)->wb_wq,
723 } else if (!aio_req->write) {
726 if (rc >= 0 && osd_data->length > rc) {
727 int zoff = osd_data->alignment + rc;
728 int zlen = osd_data->length - rc;
730 * If read is satisfied by single OSD request,
731 * it can pass EOF. Otherwise read is within
734 if (aio_req->num_reqs == 1) {
735 loff_t i_size = i_size_read(inode);
736 loff_t endoff = aio_req->iocb->ki_pos + rc;
738 zlen = min_t(size_t, zlen,
740 aio_req->total_len = rc + zlen;
744 ceph_zero_page_vector_range(zoff, zlen,
749 ceph_put_page_vector(osd_data->pages, num_pages, aio_req->should_dirty);
750 ceph_osdc_put_request(req);
753 cmpxchg(&aio_req->error, 0, rc);
755 ceph_aio_complete(inode, aio_req);
759 static void ceph_aio_retry_work(struct work_struct *work)
761 struct ceph_aio_work *aio_work =
762 container_of(work, struct ceph_aio_work, work);
763 struct ceph_osd_request *orig_req = aio_work->req;
764 struct ceph_aio_request *aio_req = orig_req->r_priv;
765 struct inode *inode = orig_req->r_inode;
766 struct ceph_inode_info *ci = ceph_inode(inode);
767 struct ceph_snap_context *snapc;
768 struct ceph_osd_request *req;
771 spin_lock(&ci->i_ceph_lock);
772 if (__ceph_have_pending_cap_snap(ci)) {
773 struct ceph_cap_snap *capsnap =
774 list_last_entry(&ci->i_cap_snaps,
775 struct ceph_cap_snap,
777 snapc = ceph_get_snap_context(capsnap->context);
779 BUG_ON(!ci->i_head_snapc);
780 snapc = ceph_get_snap_context(ci->i_head_snapc);
782 spin_unlock(&ci->i_ceph_lock);
784 req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 2,
792 req->r_flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
793 ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
794 ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
796 ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
798 ceph_osdc_put_request(req);
803 req->r_ops[0] = orig_req->r_ops[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;
846 bool should_dirty = !write && iter_is_iovec(iter);
848 if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
851 dout("sync_direct_%s on file %p %lld~%u snapc %p seq %lld\n",
852 (write ? "write" : "read"), file, pos, (unsigned)count,
855 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
860 int ret2 = invalidate_inode_pages2_range(inode->i_mapping,
862 (pos + count) >> PAGE_SHIFT);
864 dout("invalidate_inode_pages2_range returned %d\n", ret2);
866 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
868 flags = CEPH_OSD_FLAG_READ;
871 while (iov_iter_count(iter) > 0) {
872 u64 size = dio_get_pagev_size(iter);
877 size = min_t(u64, size, fsc->mount_options->wsize);
879 size = min_t(u64, size, fsc->mount_options->rsize);
881 vino = ceph_vino(inode);
882 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
885 write ? CEPH_OSD_OP_WRITE :
897 pages = dio_get_pages_alloc(iter, len, &start, &num_pages);
899 ceph_osdc_put_request(req);
900 ret = PTR_ERR(pages);
905 * To simplify error handling, allow AIO when IO within i_size
906 * or IO can be satisfied by single OSD request.
908 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
909 (len == count || pos + count <= i_size_read(inode))) {
910 aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
912 aio_req->iocb = iocb;
913 aio_req->write = write;
914 aio_req->should_dirty = should_dirty;
915 INIT_LIST_HEAD(&aio_req->osd_reqs);
917 aio_req->mtime = mtime;
918 swap(aio_req->prealloc_cf, *pcf);
926 * throw out any page cache pages in this range. this
929 truncate_inode_pages_range(inode->i_mapping, pos,
930 (pos+len) | (PAGE_SIZE - 1));
932 req->r_mtime = mtime;
935 osd_req_op_extent_osd_data_pages(req, 0, pages, len, start,
939 aio_req->total_len += len;
941 atomic_inc(&aio_req->pending_reqs);
943 req->r_callback = ceph_aio_complete_req;
944 req->r_inode = inode;
945 req->r_priv = aio_req;
946 list_add_tail(&req->r_unsafe_item, &aio_req->osd_reqs);
949 iov_iter_advance(iter, len);
953 ret = ceph_osdc_start_request(req->r_osdc, req, false);
955 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
957 size = i_size_read(inode);
961 if (ret >= 0 && ret < len && pos + ret < size) {
962 int zlen = min_t(size_t, len - ret,
964 ceph_zero_page_vector_range(start + ret, zlen,
972 ceph_put_page_vector(pages, num_pages, should_dirty);
974 ceph_osdc_put_request(req);
979 iov_iter_advance(iter, len);
981 if (!write && pos >= size)
984 if (write && pos > size) {
985 if (ceph_inode_set_size(inode, pos))
986 ceph_check_caps(ceph_inode(inode),
995 if (aio_req->num_reqs == 0) {
1000 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1003 list_splice(&aio_req->osd_reqs, &osd_reqs);
1004 while (!list_empty(&osd_reqs)) {
1005 req = list_first_entry(&osd_reqs,
1006 struct ceph_osd_request,
1008 list_del_init(&req->r_unsafe_item);
1010 ret = ceph_osdc_start_request(req->r_osdc,
1013 req->r_result = ret;
1014 ceph_aio_complete_req(req);
1017 return -EIOCBQUEUED;
1020 if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1021 ret = pos - iocb->ki_pos;
1028 * Synchronous write, straight from __user pointer or user pages.
1030 * If write spans object boundary, just do multiple writes. (For a
1031 * correct atomic write, we should e.g. take write locks on all
1032 * objects, rollback on failure, etc.)
1035 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1036 struct ceph_snap_context *snapc)
1038 struct file *file = iocb->ki_filp;
1039 struct inode *inode = file_inode(file);
1040 struct ceph_inode_info *ci = ceph_inode(inode);
1041 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1042 struct ceph_vino vino;
1043 struct ceph_osd_request *req;
1044 struct page **pages;
1050 bool check_caps = false;
1051 struct timespec mtime = current_time(inode);
1052 size_t count = iov_iter_count(from);
1054 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1057 dout("sync_write on file %p %lld~%u snapc %p seq %lld\n",
1058 file, pos, (unsigned)count, snapc, snapc->seq);
1060 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
1064 ret = invalidate_inode_pages2_range(inode->i_mapping,
1066 (pos + count) >> PAGE_SHIFT);
1068 dout("invalidate_inode_pages2_range returned %d\n", ret);
1070 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1072 while ((len = iov_iter_count(from)) > 0) {
1076 vino = ceph_vino(inode);
1077 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1078 vino, pos, &len, 0, 1,
1079 CEPH_OSD_OP_WRITE, flags, snapc,
1081 ci->i_truncate_size,
1089 * write from beginning of first page,
1090 * regardless of io alignment
1092 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1094 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1095 if (IS_ERR(pages)) {
1096 ret = PTR_ERR(pages);
1101 for (n = 0; n < num_pages; n++) {
1102 size_t plen = min_t(size_t, left, PAGE_SIZE);
1103 ret = copy_page_from_iter(pages[n], 0, plen, from);
1112 ceph_release_page_vector(pages, num_pages);
1116 req->r_inode = inode;
1118 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1121 req->r_mtime = mtime;
1122 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1124 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1127 ceph_osdc_put_request(req);
1129 ceph_set_error_write(ci);
1133 ceph_clear_error_write(ci);
1136 if (pos > i_size_read(inode)) {
1137 check_caps = ceph_inode_set_size(inode, pos);
1139 ceph_check_caps(ceph_inode(inode),
1140 CHECK_CAPS_AUTHONLY,
1146 if (ret != -EOLDSNAPC && written > 0) {
1154 * Wrap generic_file_aio_read with checks for cap bits on the inode.
1155 * Atomically grab references, so that those bits are not released
1156 * back to the MDS mid-read.
1158 * Hmm, the sync read case isn't actually async... should it be?
1160 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1162 struct file *filp = iocb->ki_filp;
1163 struct ceph_file_info *fi = filp->private_data;
1164 size_t len = iov_iter_count(to);
1165 struct inode *inode = file_inode(filp);
1166 struct ceph_inode_info *ci = ceph_inode(inode);
1167 struct page *pinned_page = NULL;
1170 int retry_op = 0, read = 0;
1173 dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1174 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1176 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1177 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1179 want = CEPH_CAP_FILE_CACHE;
1180 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1184 if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1185 (iocb->ki_flags & IOCB_DIRECT) ||
1186 (fi->flags & CEPH_F_SYNC)) {
1188 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1189 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1190 ceph_cap_string(got));
1192 if (ci->i_inline_version == CEPH_INLINE_NONE) {
1193 if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1194 ret = ceph_direct_read_write(iocb, to,
1196 if (ret >= 0 && ret < len)
1197 retry_op = CHECK_EOF;
1199 ret = ceph_sync_read(iocb, to, &retry_op);
1202 retry_op = READ_INLINE;
1205 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1206 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1207 ceph_cap_string(got));
1208 current->journal_info = filp;
1209 ret = generic_file_read_iter(iocb, to);
1210 current->journal_info = NULL;
1212 dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1213 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1215 put_page(pinned_page);
1218 ceph_put_cap_refs(ci, got);
1219 if (retry_op > HAVE_RETRIED && ret >= 0) {
1221 struct page *page = NULL;
1223 if (retry_op == READ_INLINE) {
1224 page = __page_cache_alloc(GFP_KERNEL);
1229 statret = __ceph_do_getattr(inode, page,
1230 CEPH_STAT_CAP_INLINE_DATA, !!page);
1234 if (statret == -ENODATA) {
1235 BUG_ON(retry_op != READ_INLINE);
1241 i_size = i_size_read(inode);
1242 if (retry_op == READ_INLINE) {
1243 BUG_ON(ret > 0 || read > 0);
1244 if (iocb->ki_pos < i_size &&
1245 iocb->ki_pos < PAGE_SIZE) {
1246 loff_t end = min_t(loff_t, i_size,
1247 iocb->ki_pos + len);
1248 end = min_t(loff_t, end, PAGE_SIZE);
1250 zero_user_segment(page, statret, end);
1251 ret = copy_page_to_iter(page,
1252 iocb->ki_pos & ~PAGE_MASK,
1253 end - iocb->ki_pos, to);
1254 iocb->ki_pos += ret;
1257 if (iocb->ki_pos < i_size && read < len) {
1258 size_t zlen = min_t(size_t, len - read,
1259 i_size - iocb->ki_pos);
1260 ret = iov_iter_zero(zlen, to);
1261 iocb->ki_pos += ret;
1264 __free_pages(page, 0);
1268 /* hit EOF or hole? */
1269 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1271 dout("sync_read hit hole, ppos %lld < size %lld"
1272 ", reading more\n", iocb->ki_pos, i_size);
1276 retry_op = HAVE_RETRIED;
1288 * Take cap references to avoid releasing caps to MDS mid-write.
1290 * If we are synchronous, and write with an old snap context, the OSD
1291 * may return EOLDSNAPC. In that case, retry the write.. _after_
1292 * dropping our cap refs and allowing the pending snap to logically
1293 * complete _before_ this write occurs.
1295 * If we are near ENOSPC, write synchronously.
1297 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1299 struct file *file = iocb->ki_filp;
1300 struct ceph_file_info *fi = file->private_data;
1301 struct inode *inode = file_inode(file);
1302 struct ceph_inode_info *ci = ceph_inode(inode);
1303 struct ceph_osd_client *osdc =
1304 &ceph_sb_to_client(inode->i_sb)->client->osdc;
1305 struct ceph_cap_flush *prealloc_cf;
1306 ssize_t count, written = 0;
1310 if (ceph_snap(inode) != CEPH_NOSNAP)
1313 prealloc_cf = ceph_alloc_cap_flush();
1320 /* We can write back this queue in page reclaim */
1321 current->backing_dev_info = inode_to_bdi(inode);
1323 if (iocb->ki_flags & IOCB_APPEND) {
1324 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1329 err = generic_write_checks(iocb, from);
1334 count = iov_iter_count(from);
1335 err = file_remove_privs(file);
1339 err = file_update_time(file);
1343 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1344 err = ceph_uninline_data(file, NULL);
1349 /* FIXME: not complete since it doesn't account for being at quota */
1350 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL)) {
1355 dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1356 inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1357 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1358 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1360 want = CEPH_CAP_FILE_BUFFER;
1362 err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, pos + count,
1367 dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1368 inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1370 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1371 (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
1372 (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
1373 struct ceph_snap_context *snapc;
1374 struct iov_iter data;
1375 inode_unlock(inode);
1377 spin_lock(&ci->i_ceph_lock);
1378 if (__ceph_have_pending_cap_snap(ci)) {
1379 struct ceph_cap_snap *capsnap =
1380 list_last_entry(&ci->i_cap_snaps,
1381 struct ceph_cap_snap,
1383 snapc = ceph_get_snap_context(capsnap->context);
1385 BUG_ON(!ci->i_head_snapc);
1386 snapc = ceph_get_snap_context(ci->i_head_snapc);
1388 spin_unlock(&ci->i_ceph_lock);
1390 /* we might need to revert back to that point */
1392 if (iocb->ki_flags & IOCB_DIRECT)
1393 written = ceph_direct_read_write(iocb, &data, snapc,
1396 written = ceph_sync_write(iocb, &data, pos, snapc);
1398 iov_iter_advance(from, written);
1399 ceph_put_snap_context(snapc);
1402 * No need to acquire the i_truncate_mutex. Because
1403 * the MDS revokes Fwb caps before sending truncate
1404 * message to us. We can't get Fwb cap while there
1405 * are pending vmtruncate. So write and vmtruncate
1406 * can not run at the same time
1408 written = generic_perform_write(file, from, pos);
1409 if (likely(written >= 0))
1410 iocb->ki_pos = pos + written;
1411 inode_unlock(inode);
1416 spin_lock(&ci->i_ceph_lock);
1417 ci->i_inline_version = CEPH_INLINE_NONE;
1418 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1420 spin_unlock(&ci->i_ceph_lock);
1422 __mark_inode_dirty(inode, dirty);
1425 dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n",
1426 inode, ceph_vinop(inode), pos, (unsigned)count,
1427 ceph_cap_string(got));
1428 ceph_put_cap_refs(ci, got);
1430 if (written == -EOLDSNAPC) {
1431 dout("aio_write %p %llx.%llx %llu~%u" "got EOLDSNAPC, retrying\n",
1432 inode, ceph_vinop(inode), pos, (unsigned)count);
1437 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_NEARFULL))
1438 iocb->ki_flags |= IOCB_DSYNC;
1439 written = generic_write_sync(iocb, written);
1445 inode_unlock(inode);
1447 ceph_free_cap_flush(prealloc_cf);
1448 current->backing_dev_info = NULL;
1449 return written ? written : err;
1453 * llseek. be sure to verify file size on SEEK_END.
1455 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1457 struct inode *inode = file->f_mapping->host;
1463 if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1464 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1469 i_size = i_size_read(inode);
1476 * Here we special-case the lseek(fd, 0, SEEK_CUR)
1477 * position-querying operation. Avoid rewriting the "same"
1478 * f_pos value back to the file because a concurrent read(),
1479 * write() or lseek() might have altered it
1485 offset += file->f_pos;
1488 if (offset < 0 || offset >= i_size) {
1494 if (offset < 0 || offset >= i_size) {
1502 ret = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1505 inode_unlock(inode);
1509 static inline void ceph_zero_partial_page(
1510 struct inode *inode, loff_t offset, unsigned size)
1513 pgoff_t index = offset >> PAGE_SHIFT;
1515 page = find_lock_page(inode->i_mapping, index);
1517 wait_on_page_writeback(page);
1518 zero_user(page, offset & (PAGE_SIZE - 1), size);
1524 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1527 loff_t nearly = round_up(offset, PAGE_SIZE);
1528 if (offset < nearly) {
1529 loff_t size = nearly - offset;
1532 ceph_zero_partial_page(inode, offset, size);
1536 if (length >= PAGE_SIZE) {
1537 loff_t size = round_down(length, PAGE_SIZE);
1538 truncate_pagecache_range(inode, offset, offset + size - 1);
1543 ceph_zero_partial_page(inode, offset, length);
1546 static int ceph_zero_partial_object(struct inode *inode,
1547 loff_t offset, loff_t *length)
1549 struct ceph_inode_info *ci = ceph_inode(inode);
1550 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1551 struct ceph_osd_request *req;
1557 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1560 op = CEPH_OSD_OP_ZERO;
1563 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1567 CEPH_OSD_FLAG_WRITE,
1574 req->r_mtime = inode->i_mtime;
1575 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1577 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1581 ceph_osdc_put_request(req);
1587 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
1590 struct ceph_inode_info *ci = ceph_inode(inode);
1591 s32 stripe_unit = ci->i_layout.stripe_unit;
1592 s32 stripe_count = ci->i_layout.stripe_count;
1593 s32 object_size = ci->i_layout.object_size;
1594 u64 object_set_size = object_size * stripe_count;
1597 /* round offset up to next period boundary */
1598 nearly = offset + object_set_size - 1;
1600 nearly -= do_div(t, object_set_size);
1602 while (length && offset < nearly) {
1603 loff_t size = length;
1604 ret = ceph_zero_partial_object(inode, offset, &size);
1610 while (length >= object_set_size) {
1612 loff_t pos = offset;
1613 for (i = 0; i < stripe_count; ++i) {
1614 ret = ceph_zero_partial_object(inode, pos, NULL);
1619 offset += object_set_size;
1620 length -= object_set_size;
1623 loff_t size = length;
1624 ret = ceph_zero_partial_object(inode, offset, &size);
1633 static long ceph_fallocate(struct file *file, int mode,
1634 loff_t offset, loff_t length)
1636 struct ceph_file_info *fi = file->private_data;
1637 struct inode *inode = file_inode(file);
1638 struct ceph_inode_info *ci = ceph_inode(inode);
1639 struct ceph_osd_client *osdc =
1640 &ceph_inode_to_client(inode)->client->osdc;
1641 struct ceph_cap_flush *prealloc_cf;
1648 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1651 if (!S_ISREG(inode->i_mode))
1654 prealloc_cf = ceph_alloc_cap_flush();
1660 if (ceph_snap(inode) != CEPH_NOSNAP) {
1665 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) &&
1666 !(mode & FALLOC_FL_PUNCH_HOLE)) {
1671 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1672 ret = ceph_uninline_data(file, NULL);
1677 size = i_size_read(inode);
1678 if (!(mode & FALLOC_FL_KEEP_SIZE)) {
1679 endoff = offset + length;
1680 ret = inode_newsize_ok(inode, endoff);
1685 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1686 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1688 want = CEPH_CAP_FILE_BUFFER;
1690 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, endoff, &got, NULL);
1694 if (mode & FALLOC_FL_PUNCH_HOLE) {
1696 ceph_zero_pagecache_range(inode, offset, length);
1697 ret = ceph_zero_objects(inode, offset, length);
1698 } else if (endoff > size) {
1699 truncate_pagecache_range(inode, size, -1);
1700 if (ceph_inode_set_size(inode, endoff))
1701 ceph_check_caps(ceph_inode(inode),
1702 CHECK_CAPS_AUTHONLY, NULL);
1706 spin_lock(&ci->i_ceph_lock);
1707 ci->i_inline_version = CEPH_INLINE_NONE;
1708 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1710 spin_unlock(&ci->i_ceph_lock);
1712 __mark_inode_dirty(inode, dirty);
1715 ceph_put_cap_refs(ci, got);
1717 inode_unlock(inode);
1718 ceph_free_cap_flush(prealloc_cf);
1722 const struct file_operations ceph_file_fops = {
1724 .release = ceph_release,
1725 .llseek = ceph_llseek,
1726 .read_iter = ceph_read_iter,
1727 .write_iter = ceph_write_iter,
1729 .fsync = ceph_fsync,
1731 .setlease = simple_nosetlease,
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,