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 = kvmalloc(sizeof(*pages) * npages, GFP_KERNEL);
79 return ERR_PTR(-ENOMEM);
81 for (idx = 0; idx < npages; ) {
83 ret = iov_iter_get_pages(&tmp_it, pages + idx, nbytes,
84 npages - idx, &start);
88 iov_iter_advance(&tmp_it, ret);
90 idx += (ret + start + PAGE_SIZE - 1) / PAGE_SIZE;
96 dout("dio_get_pages_alloc: got %d pages align %zu\n", npages, align);
99 ceph_put_page_vector(pages, idx, false);
104 * Prepare an open request. Preallocate ceph_cap to avoid an
105 * inopportune ENOMEM later.
107 static struct ceph_mds_request *
108 prepare_open_request(struct super_block *sb, int flags, int create_mode)
110 struct ceph_fs_client *fsc = ceph_sb_to_client(sb);
111 struct ceph_mds_client *mdsc = fsc->mdsc;
112 struct ceph_mds_request *req;
113 int want_auth = USE_ANY_MDS;
114 int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
116 if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
117 want_auth = USE_AUTH_MDS;
119 req = ceph_mdsc_create_request(mdsc, op, want_auth);
122 req->r_fmode = ceph_flags_to_mode(flags);
123 req->r_args.open.flags = cpu_to_le32(flags);
124 req->r_args.open.mode = cpu_to_le32(create_mode);
130 * initialize private struct file data.
131 * if we fail, clean up by dropping fmode reference on the ceph_inode
133 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
135 struct ceph_file_info *cf;
138 switch (inode->i_mode & S_IFMT) {
140 ceph_fscache_register_inode_cookie(inode);
141 ceph_fscache_file_set_cookie(inode, file);
143 dout("init_file %p %p 0%o (regular)\n", inode, file,
145 cf = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
147 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
152 cf->readdir_cache_idx = -1;
153 file->private_data = cf;
154 BUG_ON(inode->i_fop->release != ceph_release);
158 dout("init_file %p %p 0%o (symlink)\n", inode, file,
160 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
164 dout("init_file %p %p 0%o (special)\n", inode, file,
167 * we need to drop the open ref now, since we don't
168 * have .release set to ceph_release.
170 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
171 BUG_ON(inode->i_fop->release == ceph_release);
173 /* call the proper open fop */
174 ret = inode->i_fop->open(inode, file);
180 * try renew caps after session gets killed.
182 int ceph_renew_caps(struct inode *inode)
184 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
185 struct ceph_inode_info *ci = ceph_inode(inode);
186 struct ceph_mds_request *req;
187 int err, flags, wanted;
189 spin_lock(&ci->i_ceph_lock);
190 wanted = __ceph_caps_file_wanted(ci);
191 if (__ceph_is_any_real_caps(ci) &&
192 (!(wanted & CEPH_CAP_ANY_WR) == 0 || ci->i_auth_cap)) {
193 int issued = __ceph_caps_issued(ci, NULL);
194 spin_unlock(&ci->i_ceph_lock);
195 dout("renew caps %p want %s issued %s updating mds_wanted\n",
196 inode, ceph_cap_string(wanted), ceph_cap_string(issued));
197 ceph_check_caps(ci, 0, NULL);
200 spin_unlock(&ci->i_ceph_lock);
203 if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
205 else if (wanted & CEPH_CAP_FILE_RD)
207 else if (wanted & CEPH_CAP_FILE_WR)
210 if (wanted & CEPH_CAP_FILE_LAZYIO)
214 req = prepare_open_request(inode->i_sb, flags, 0);
220 req->r_inode = inode;
225 err = ceph_mdsc_do_request(mdsc, NULL, req);
226 ceph_mdsc_put_request(req);
228 dout("renew caps %p open result=%d\n", inode, err);
229 return err < 0 ? err : 0;
233 * If we already have the requisite capabilities, we can satisfy
234 * the open request locally (no need to request new caps from the
235 * MDS). We do, however, need to inform the MDS (asynchronously)
236 * if our wanted caps set expands.
238 int ceph_open(struct inode *inode, struct file *file)
240 struct ceph_inode_info *ci = ceph_inode(inode);
241 struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
242 struct ceph_mds_client *mdsc = fsc->mdsc;
243 struct ceph_mds_request *req;
244 struct ceph_file_info *cf = file->private_data;
246 int flags, fmode, wanted;
249 dout("open file %p is already opened\n", file);
253 /* filter out O_CREAT|O_EXCL; vfs did that already. yuck. */
254 flags = file->f_flags & ~(O_CREAT|O_EXCL);
255 if (S_ISDIR(inode->i_mode))
256 flags = O_DIRECTORY; /* mds likes to know */
258 dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
259 ceph_vinop(inode), file, flags, file->f_flags);
260 fmode = ceph_flags_to_mode(flags);
261 wanted = ceph_caps_for_mode(fmode);
263 /* snapped files are read-only */
264 if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
267 /* trivially open snapdir */
268 if (ceph_snap(inode) == CEPH_SNAPDIR) {
269 spin_lock(&ci->i_ceph_lock);
270 __ceph_get_fmode(ci, fmode);
271 spin_unlock(&ci->i_ceph_lock);
272 return ceph_init_file(inode, file, fmode);
276 * No need to block if we have caps on the auth MDS (for
277 * write) or any MDS (for read). Update wanted set
280 spin_lock(&ci->i_ceph_lock);
281 if (__ceph_is_any_real_caps(ci) &&
282 (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
283 int mds_wanted = __ceph_caps_mds_wanted(ci, true);
284 int issued = __ceph_caps_issued(ci, NULL);
286 dout("open %p fmode %d want %s issued %s using existing\n",
287 inode, fmode, ceph_cap_string(wanted),
288 ceph_cap_string(issued));
289 __ceph_get_fmode(ci, fmode);
290 spin_unlock(&ci->i_ceph_lock);
293 if ((issued & wanted) != wanted &&
294 (mds_wanted & wanted) != wanted &&
295 ceph_snap(inode) != CEPH_SNAPDIR)
296 ceph_check_caps(ci, 0, NULL);
298 return ceph_init_file(inode, file, fmode);
299 } else if (ceph_snap(inode) != CEPH_NOSNAP &&
300 (ci->i_snap_caps & wanted) == wanted) {
301 __ceph_get_fmode(ci, fmode);
302 spin_unlock(&ci->i_ceph_lock);
303 return ceph_init_file(inode, file, fmode);
306 spin_unlock(&ci->i_ceph_lock);
308 dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
309 req = prepare_open_request(inode->i_sb, flags, 0);
314 req->r_inode = inode;
318 err = ceph_mdsc_do_request(mdsc, NULL, req);
320 err = ceph_init_file(inode, file, req->r_fmode);
321 ceph_mdsc_put_request(req);
322 dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
329 * Do a lookup + open with a single request. If we get a non-existent
330 * file or symlink, return 1 so the VFS can retry.
332 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
333 struct file *file, unsigned flags, umode_t mode,
336 struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
337 struct ceph_mds_client *mdsc = fsc->mdsc;
338 struct ceph_mds_request *req;
340 struct ceph_acls_info acls = {};
344 dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
346 d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
348 if (dentry->d_name.len > NAME_MAX)
349 return -ENAMETOOLONG;
351 if (flags & O_CREAT) {
352 err = ceph_pre_init_acls(dir, &mode, &acls);
358 req = prepare_open_request(dir->i_sb, flags, mode);
363 req->r_dentry = dget(dentry);
365 if (flags & O_CREAT) {
366 req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
367 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
369 req->r_pagelist = acls.pagelist;
370 acls.pagelist = NULL;
374 mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
375 if (ceph_security_xattr_wanted(dir))
376 mask |= CEPH_CAP_XATTR_SHARED;
377 req->r_args.open.mask = cpu_to_le32(mask);
380 set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
381 err = ceph_mdsc_do_request(mdsc,
382 (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
384 err = ceph_handle_snapdir(req, dentry, err);
388 if ((flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
389 err = ceph_handle_notrace_create(dir, dentry);
391 if (d_in_lookup(dentry)) {
392 dn = ceph_finish_lookup(req, dentry, err);
396 /* we were given a hashed negative dentry */
401 if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
402 /* make vfs retry on splice, ENOENT, or symlink */
403 dout("atomic_open finish_no_open on dn %p\n", dn);
404 err = finish_no_open(file, dn);
406 dout("atomic_open finish_open on dn %p\n", dn);
407 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
408 ceph_init_inode_acls(d_inode(dentry), &acls);
409 *opened |= FILE_CREATED;
411 err = finish_open(file, dentry, ceph_open, opened);
414 if (!req->r_err && req->r_target_inode)
415 ceph_put_fmode(ceph_inode(req->r_target_inode), req->r_fmode);
416 ceph_mdsc_put_request(req);
418 ceph_release_acls_info(&acls);
419 dout("atomic_open result=%d\n", err);
423 int ceph_release(struct inode *inode, struct file *file)
425 struct ceph_inode_info *ci = ceph_inode(inode);
426 struct ceph_file_info *cf = file->private_data;
428 dout("release inode %p file %p\n", inode, file);
429 ceph_put_fmode(ci, cf->fmode);
430 if (cf->last_readdir)
431 ceph_mdsc_put_request(cf->last_readdir);
432 kfree(cf->last_name);
434 kmem_cache_free(ceph_file_cachep, cf);
436 /* wake up anyone waiting for caps on this inode */
437 wake_up_all(&ci->i_cap_wq);
448 * Read a range of bytes striped over one or more objects. Iterate over
449 * objects we stripe over. (That's not atomic, but good enough for now.)
451 * If we get a short result from the OSD, check against i_size; we need to
452 * only return a short read to the caller if we hit EOF.
454 static int striped_read(struct inode *inode,
456 struct page **pages, int num_pages,
457 int page_align, int *checkeof)
459 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
460 struct ceph_inode_info *ci = ceph_inode(inode);
465 bool hit_stripe, was_short;
468 * we may need to do multiple reads. not atomic, unfortunately.
472 page_idx = (page_align + read) >> PAGE_SHIFT;
473 ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode),
474 &ci->i_layout, pos, &this_len,
475 ci->i_truncate_seq, ci->i_truncate_size,
476 pages + page_idx, num_pages - page_idx,
477 ((page_align + read) & ~PAGE_MASK));
480 hit_stripe = this_len < len;
481 was_short = ret >= 0 && ret < this_len;
482 dout("striped_read %llu~%llu (read %u) got %d%s%s\n", pos, len, read,
483 ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : "");
485 i_size = i_size_read(inode);
487 if (was_short && (pos + ret < i_size)) {
488 int zlen = min(this_len - ret, i_size - pos - ret);
489 int zoff = page_align + read + ret;
490 dout(" zero gap %llu to %llu\n",
491 pos + ret, pos + ret + zlen);
492 ceph_zero_page_vector_range(zoff, zlen, pages);
500 /* hit stripe and need continue*/
501 if (len && hit_stripe && pos < i_size)
507 /* did we bounce off eof? */
508 if (pos + len > i_size)
509 *checkeof = CHECK_EOF;
512 dout("striped_read returns %d\n", ret);
517 * Completely synchronous read and write methods. Direct from __user
518 * buffer to osd, or directly to user pages (if O_DIRECT).
520 * If the read spans object boundary, just do multiple reads.
522 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
525 struct file *file = iocb->ki_filp;
526 struct inode *inode = file_inode(file);
528 u64 off = iocb->ki_pos;
531 size_t len = iov_iter_count(to);
533 dout("sync_read on file %p %llu~%u %s\n", file, off,
535 (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
540 * flush any page cache pages in this range. this
541 * will make concurrent normal and sync io slow,
542 * but it will at least behave sensibly when they are
545 ret = filemap_write_and_wait_range(inode->i_mapping, off,
550 if (unlikely(to->type & ITER_PIPE)) {
552 ret = iov_iter_get_pages_alloc(to, &pages, len,
556 num_pages = DIV_ROUND_UP(ret + page_off, PAGE_SIZE);
558 ret = striped_read(inode, off, ret, pages, num_pages,
561 iov_iter_advance(to, ret);
564 iov_iter_advance(to, 0);
566 ceph_put_page_vector(pages, num_pages, false);
568 num_pages = calc_pages_for(off, len);
569 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
571 return PTR_ERR(pages);
573 ret = striped_read(inode, off, len, pages, num_pages,
574 (off & ~PAGE_MASK), checkeof);
580 size_t page_off = off & ~PAGE_MASK;
581 size_t copy = min_t(size_t, left,
582 PAGE_SIZE - page_off);
583 l = copy_page_to_iter(pages[k++], page_off,
591 ceph_release_page_vector(pages, num_pages);
594 if (off > iocb->ki_pos) {
595 ret = off - iocb->ki_pos;
599 dout("sync_read result %zd\n", ret);
603 struct ceph_aio_request {
608 struct list_head osd_reqs;
610 atomic_t pending_reqs;
611 struct timespec mtime;
612 struct ceph_cap_flush *prealloc_cf;
615 struct ceph_aio_work {
616 struct work_struct work;
617 struct ceph_osd_request *req;
620 static void ceph_aio_retry_work(struct work_struct *work);
622 static void ceph_aio_complete(struct inode *inode,
623 struct ceph_aio_request *aio_req)
625 struct ceph_inode_info *ci = ceph_inode(inode);
628 if (!atomic_dec_and_test(&aio_req->pending_reqs))
631 ret = aio_req->error;
633 ret = aio_req->total_len;
635 dout("ceph_aio_complete %p rc %d\n", inode, ret);
637 if (ret >= 0 && aio_req->write) {
640 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
641 if (endoff > i_size_read(inode)) {
642 if (ceph_inode_set_size(inode, endoff))
643 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
646 spin_lock(&ci->i_ceph_lock);
647 ci->i_inline_version = CEPH_INLINE_NONE;
648 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
649 &aio_req->prealloc_cf);
650 spin_unlock(&ci->i_ceph_lock);
652 __mark_inode_dirty(inode, dirty);
656 ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
659 aio_req->iocb->ki_complete(aio_req->iocb, ret, 0);
661 ceph_free_cap_flush(aio_req->prealloc_cf);
665 static void ceph_aio_complete_req(struct ceph_osd_request *req)
667 int rc = req->r_result;
668 struct inode *inode = req->r_inode;
669 struct ceph_aio_request *aio_req = req->r_priv;
670 struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
671 int num_pages = calc_pages_for((u64)osd_data->alignment,
674 dout("ceph_aio_complete_req %p rc %d bytes %llu\n",
675 inode, rc, osd_data->length);
677 if (rc == -EOLDSNAPC) {
678 struct ceph_aio_work *aio_work;
679 BUG_ON(!aio_req->write);
681 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
683 INIT_WORK(&aio_work->work, ceph_aio_retry_work);
685 queue_work(ceph_inode_to_client(inode)->wb_wq,
690 } else if (!aio_req->write) {
693 if (rc >= 0 && osd_data->length > rc) {
694 int zoff = osd_data->alignment + rc;
695 int zlen = osd_data->length - rc;
697 * If read is satisfied by single OSD request,
698 * it can pass EOF. Otherwise read is within
701 if (aio_req->num_reqs == 1) {
702 loff_t i_size = i_size_read(inode);
703 loff_t endoff = aio_req->iocb->ki_pos + rc;
705 zlen = min_t(size_t, zlen,
707 aio_req->total_len = rc + zlen;
711 ceph_zero_page_vector_range(zoff, zlen,
716 ceph_put_page_vector(osd_data->pages, num_pages, !aio_req->write);
717 ceph_osdc_put_request(req);
720 cmpxchg(&aio_req->error, 0, rc);
722 ceph_aio_complete(inode, aio_req);
726 static void ceph_aio_retry_work(struct work_struct *work)
728 struct ceph_aio_work *aio_work =
729 container_of(work, struct ceph_aio_work, work);
730 struct ceph_osd_request *orig_req = aio_work->req;
731 struct ceph_aio_request *aio_req = orig_req->r_priv;
732 struct inode *inode = orig_req->r_inode;
733 struct ceph_inode_info *ci = ceph_inode(inode);
734 struct ceph_snap_context *snapc;
735 struct ceph_osd_request *req;
738 spin_lock(&ci->i_ceph_lock);
739 if (__ceph_have_pending_cap_snap(ci)) {
740 struct ceph_cap_snap *capsnap =
741 list_last_entry(&ci->i_cap_snaps,
742 struct ceph_cap_snap,
744 snapc = ceph_get_snap_context(capsnap->context);
746 BUG_ON(!ci->i_head_snapc);
747 snapc = ceph_get_snap_context(ci->i_head_snapc);
749 spin_unlock(&ci->i_ceph_lock);
751 req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 2,
759 req->r_flags = CEPH_OSD_FLAG_ORDERSNAP | CEPH_OSD_FLAG_WRITE;
760 ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
761 ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
763 ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
765 ceph_osdc_put_request(req);
770 req->r_ops[0] = orig_req->r_ops[0];
771 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC, 0);
773 req->r_mtime = aio_req->mtime;
774 req->r_data_offset = req->r_ops[0].extent.offset;
776 ceph_osdc_put_request(orig_req);
778 req->r_callback = ceph_aio_complete_req;
779 req->r_inode = inode;
780 req->r_priv = aio_req;
782 ret = ceph_osdc_start_request(req->r_osdc, req, false);
786 ceph_aio_complete_req(req);
789 ceph_put_snap_context(snapc);
794 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
795 struct ceph_snap_context *snapc,
796 struct ceph_cap_flush **pcf)
798 struct file *file = iocb->ki_filp;
799 struct inode *inode = file_inode(file);
800 struct ceph_inode_info *ci = ceph_inode(inode);
801 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
802 struct ceph_vino vino;
803 struct ceph_osd_request *req;
805 struct ceph_aio_request *aio_req = NULL;
809 struct timespec mtime = current_time(inode);
810 size_t count = iov_iter_count(iter);
811 loff_t pos = iocb->ki_pos;
812 bool write = iov_iter_rw(iter) == WRITE;
814 if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
817 dout("sync_direct_read_write (%s) on file %p %lld~%u\n",
818 (write ? "write" : "read"), file, pos, (unsigned)count);
820 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
825 int ret2 = invalidate_inode_pages2_range(inode->i_mapping,
827 (pos + count) >> PAGE_SHIFT);
829 dout("invalidate_inode_pages2_range returned %d\n", ret2);
831 flags = CEPH_OSD_FLAG_ORDERSNAP | CEPH_OSD_FLAG_WRITE;
833 flags = CEPH_OSD_FLAG_READ;
836 while (iov_iter_count(iter) > 0) {
837 u64 size = dio_get_pagev_size(iter);
841 vino = ceph_vino(inode);
842 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
844 /*include a 'startsync' command*/
846 write ? CEPH_OSD_OP_WRITE :
858 pages = dio_get_pages_alloc(iter, len, &start, &num_pages);
860 ceph_osdc_put_request(req);
861 ret = PTR_ERR(pages);
866 * To simplify error handling, allow AIO when IO within i_size
867 * or IO can be satisfied by single OSD request.
869 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
870 (len == count || pos + count <= i_size_read(inode))) {
871 aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
873 aio_req->iocb = iocb;
874 aio_req->write = write;
875 INIT_LIST_HEAD(&aio_req->osd_reqs);
877 aio_req->mtime = mtime;
878 swap(aio_req->prealloc_cf, *pcf);
886 * throw out any page cache pages in this range. this
889 truncate_inode_pages_range(inode->i_mapping, pos,
890 (pos+len) | (PAGE_SIZE - 1));
892 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC, 0);
893 req->r_mtime = mtime;
896 osd_req_op_extent_osd_data_pages(req, 0, pages, len, start,
900 aio_req->total_len += len;
902 atomic_inc(&aio_req->pending_reqs);
904 req->r_callback = ceph_aio_complete_req;
905 req->r_inode = inode;
906 req->r_priv = aio_req;
907 list_add_tail(&req->r_unsafe_item, &aio_req->osd_reqs);
910 iov_iter_advance(iter, len);
914 ret = ceph_osdc_start_request(req->r_osdc, req, false);
916 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
918 size = i_size_read(inode);
922 if (ret >= 0 && ret < len && pos + ret < size) {
923 int zlen = min_t(size_t, len - ret,
925 ceph_zero_page_vector_range(start + ret, zlen,
933 ceph_put_page_vector(pages, num_pages, !write);
935 ceph_osdc_put_request(req);
940 iov_iter_advance(iter, len);
942 if (!write && pos >= size)
945 if (write && pos > size) {
946 if (ceph_inode_set_size(inode, pos))
947 ceph_check_caps(ceph_inode(inode),
956 if (aio_req->num_reqs == 0) {
961 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
964 list_splice(&aio_req->osd_reqs, &osd_reqs);
965 while (!list_empty(&osd_reqs)) {
966 req = list_first_entry(&osd_reqs,
967 struct ceph_osd_request,
969 list_del_init(&req->r_unsafe_item);
971 ret = ceph_osdc_start_request(req->r_osdc,
975 ceph_aio_complete_req(req);
981 if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
982 ret = pos - iocb->ki_pos;
989 * Synchronous write, straight from __user pointer or user pages.
991 * If write spans object boundary, just do multiple writes. (For a
992 * correct atomic write, we should e.g. take write locks on all
993 * objects, rollback on failure, etc.)
996 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
997 struct ceph_snap_context *snapc)
999 struct file *file = iocb->ki_filp;
1000 struct inode *inode = file_inode(file);
1001 struct ceph_inode_info *ci = ceph_inode(inode);
1002 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1003 struct ceph_vino vino;
1004 struct ceph_osd_request *req;
1005 struct page **pages;
1012 struct timespec mtime = current_time(inode);
1013 size_t count = iov_iter_count(from);
1015 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1018 dout("sync_write on file %p %lld~%u\n", file, pos, (unsigned)count);
1020 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
1024 ret = invalidate_inode_pages2_range(inode->i_mapping,
1026 (pos + count) >> PAGE_SHIFT);
1028 dout("invalidate_inode_pages2_range returned %d\n", ret);
1030 flags = CEPH_OSD_FLAG_ORDERSNAP | CEPH_OSD_FLAG_WRITE;
1032 while ((len = iov_iter_count(from)) > 0) {
1036 vino = ceph_vino(inode);
1037 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1038 vino, pos, &len, 0, 1,
1039 CEPH_OSD_OP_WRITE, flags, snapc,
1041 ci->i_truncate_size,
1049 * write from beginning of first page,
1050 * regardless of io alignment
1052 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1054 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1055 if (IS_ERR(pages)) {
1056 ret = PTR_ERR(pages);
1061 for (n = 0; n < num_pages; n++) {
1062 size_t plen = min_t(size_t, left, PAGE_SIZE);
1063 ret = copy_page_from_iter(pages[n], 0, plen, from);
1072 ceph_release_page_vector(pages, num_pages);
1076 req->r_inode = inode;
1078 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1081 req->r_mtime = mtime;
1082 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1084 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1087 ceph_osdc_put_request(req);
1092 if (pos > i_size_read(inode)) {
1093 check_caps = ceph_inode_set_size(inode, pos);
1095 ceph_check_caps(ceph_inode(inode),
1096 CHECK_CAPS_AUTHONLY,
1103 if (ret != -EOLDSNAPC && written > 0) {
1111 * Wrap generic_file_aio_read with checks for cap bits on the inode.
1112 * Atomically grab references, so that those bits are not released
1113 * back to the MDS mid-read.
1115 * Hmm, the sync read case isn't actually async... should it be?
1117 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1119 struct file *filp = iocb->ki_filp;
1120 struct ceph_file_info *fi = filp->private_data;
1121 size_t len = iov_iter_count(to);
1122 struct inode *inode = file_inode(filp);
1123 struct ceph_inode_info *ci = ceph_inode(inode);
1124 struct page *pinned_page = NULL;
1127 int retry_op = 0, read = 0;
1130 dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1131 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1133 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1134 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1136 want = CEPH_CAP_FILE_CACHE;
1137 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1141 if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1142 (iocb->ki_flags & IOCB_DIRECT) ||
1143 (fi->flags & CEPH_F_SYNC)) {
1145 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1146 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1147 ceph_cap_string(got));
1149 if (ci->i_inline_version == CEPH_INLINE_NONE) {
1150 if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1151 ret = ceph_direct_read_write(iocb, to,
1153 if (ret >= 0 && ret < len)
1154 retry_op = CHECK_EOF;
1156 ret = ceph_sync_read(iocb, to, &retry_op);
1159 retry_op = READ_INLINE;
1162 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1163 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1164 ceph_cap_string(got));
1165 current->journal_info = filp;
1166 ret = generic_file_read_iter(iocb, to);
1167 current->journal_info = NULL;
1169 dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1170 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1172 put_page(pinned_page);
1175 ceph_put_cap_refs(ci, got);
1176 if (retry_op > HAVE_RETRIED && ret >= 0) {
1178 struct page *page = NULL;
1180 if (retry_op == READ_INLINE) {
1181 page = __page_cache_alloc(GFP_KERNEL);
1186 statret = __ceph_do_getattr(inode, page,
1187 CEPH_STAT_CAP_INLINE_DATA, !!page);
1191 if (statret == -ENODATA) {
1192 BUG_ON(retry_op != READ_INLINE);
1198 i_size = i_size_read(inode);
1199 if (retry_op == READ_INLINE) {
1200 BUG_ON(ret > 0 || read > 0);
1201 if (iocb->ki_pos < i_size &&
1202 iocb->ki_pos < PAGE_SIZE) {
1203 loff_t end = min_t(loff_t, i_size,
1204 iocb->ki_pos + len);
1205 end = min_t(loff_t, end, PAGE_SIZE);
1207 zero_user_segment(page, statret, end);
1208 ret = copy_page_to_iter(page,
1209 iocb->ki_pos & ~PAGE_MASK,
1210 end - iocb->ki_pos, to);
1211 iocb->ki_pos += ret;
1214 if (iocb->ki_pos < i_size && read < len) {
1215 size_t zlen = min_t(size_t, len - read,
1216 i_size - iocb->ki_pos);
1217 ret = iov_iter_zero(zlen, to);
1218 iocb->ki_pos += ret;
1221 __free_pages(page, 0);
1225 /* hit EOF or hole? */
1226 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1228 dout("sync_read hit hole, ppos %lld < size %lld"
1229 ", reading more\n", iocb->ki_pos, i_size);
1233 retry_op = HAVE_RETRIED;
1245 * Take cap references to avoid releasing caps to MDS mid-write.
1247 * If we are synchronous, and write with an old snap context, the OSD
1248 * may return EOLDSNAPC. In that case, retry the write.. _after_
1249 * dropping our cap refs and allowing the pending snap to logically
1250 * complete _before_ this write occurs.
1252 * If we are near ENOSPC, write synchronously.
1254 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1256 struct file *file = iocb->ki_filp;
1257 struct ceph_file_info *fi = file->private_data;
1258 struct inode *inode = file_inode(file);
1259 struct ceph_inode_info *ci = ceph_inode(inode);
1260 struct ceph_osd_client *osdc =
1261 &ceph_sb_to_client(inode->i_sb)->client->osdc;
1262 struct ceph_cap_flush *prealloc_cf;
1263 ssize_t count, written = 0;
1267 if (ceph_snap(inode) != CEPH_NOSNAP)
1270 prealloc_cf = ceph_alloc_cap_flush();
1276 /* We can write back this queue in page reclaim */
1277 current->backing_dev_info = inode_to_bdi(inode);
1279 if (iocb->ki_flags & IOCB_APPEND) {
1280 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1285 err = generic_write_checks(iocb, from);
1290 count = iov_iter_count(from);
1291 err = file_remove_privs(file);
1295 err = file_update_time(file);
1299 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1300 err = ceph_uninline_data(file, NULL);
1306 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL)) {
1311 dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1312 inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1313 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1314 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1316 want = CEPH_CAP_FILE_BUFFER;
1318 err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, pos + count,
1323 dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1324 inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1326 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1327 (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC)) {
1328 struct ceph_snap_context *snapc;
1329 struct iov_iter data;
1330 inode_unlock(inode);
1332 spin_lock(&ci->i_ceph_lock);
1333 if (__ceph_have_pending_cap_snap(ci)) {
1334 struct ceph_cap_snap *capsnap =
1335 list_last_entry(&ci->i_cap_snaps,
1336 struct ceph_cap_snap,
1338 snapc = ceph_get_snap_context(capsnap->context);
1340 BUG_ON(!ci->i_head_snapc);
1341 snapc = ceph_get_snap_context(ci->i_head_snapc);
1343 spin_unlock(&ci->i_ceph_lock);
1345 /* we might need to revert back to that point */
1347 if (iocb->ki_flags & IOCB_DIRECT)
1348 written = ceph_direct_read_write(iocb, &data, snapc,
1351 written = ceph_sync_write(iocb, &data, pos, snapc);
1352 if (written == -EOLDSNAPC) {
1353 dout("aio_write %p %llx.%llx %llu~%u"
1354 "got EOLDSNAPC, retrying\n",
1355 inode, ceph_vinop(inode),
1356 pos, (unsigned)count);
1361 iov_iter_advance(from, written);
1362 ceph_put_snap_context(snapc);
1365 * No need to acquire the i_truncate_mutex. Because
1366 * the MDS revokes Fwb caps before sending truncate
1367 * message to us. We can't get Fwb cap while there
1368 * are pending vmtruncate. So write and vmtruncate
1369 * can not run at the same time
1371 written = generic_perform_write(file, from, pos);
1372 if (likely(written >= 0))
1373 iocb->ki_pos = pos + written;
1374 inode_unlock(inode);
1379 spin_lock(&ci->i_ceph_lock);
1380 ci->i_inline_version = CEPH_INLINE_NONE;
1381 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1383 spin_unlock(&ci->i_ceph_lock);
1385 __mark_inode_dirty(inode, dirty);
1388 dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n",
1389 inode, ceph_vinop(inode), pos, (unsigned)count,
1390 ceph_cap_string(got));
1391 ceph_put_cap_refs(ci, got);
1394 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_NEARFULL))
1395 iocb->ki_flags |= IOCB_DSYNC;
1397 written = generic_write_sync(iocb, written);
1403 inode_unlock(inode);
1405 ceph_free_cap_flush(prealloc_cf);
1406 current->backing_dev_info = NULL;
1407 return written ? written : err;
1411 * llseek. be sure to verify file size on SEEK_END.
1413 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1415 struct inode *inode = file->f_mapping->host;
1421 if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1422 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1427 i_size = i_size_read(inode);
1434 * Here we special-case the lseek(fd, 0, SEEK_CUR)
1435 * position-querying operation. Avoid rewriting the "same"
1436 * f_pos value back to the file because a concurrent read(),
1437 * write() or lseek() might have altered it
1443 offset += file->f_pos;
1446 if (offset >= i_size) {
1452 if (offset >= i_size) {
1460 ret = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1463 inode_unlock(inode);
1467 static inline void ceph_zero_partial_page(
1468 struct inode *inode, loff_t offset, unsigned size)
1471 pgoff_t index = offset >> PAGE_SHIFT;
1473 page = find_lock_page(inode->i_mapping, index);
1475 wait_on_page_writeback(page);
1476 zero_user(page, offset & (PAGE_SIZE - 1), size);
1482 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1485 loff_t nearly = round_up(offset, PAGE_SIZE);
1486 if (offset < nearly) {
1487 loff_t size = nearly - offset;
1490 ceph_zero_partial_page(inode, offset, size);
1494 if (length >= PAGE_SIZE) {
1495 loff_t size = round_down(length, PAGE_SIZE);
1496 truncate_pagecache_range(inode, offset, offset + size - 1);
1501 ceph_zero_partial_page(inode, offset, length);
1504 static int ceph_zero_partial_object(struct inode *inode,
1505 loff_t offset, loff_t *length)
1507 struct ceph_inode_info *ci = ceph_inode(inode);
1508 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1509 struct ceph_osd_request *req;
1515 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1518 op = CEPH_OSD_OP_ZERO;
1521 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1525 CEPH_OSD_FLAG_WRITE,
1532 req->r_mtime = inode->i_mtime;
1533 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1535 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1539 ceph_osdc_put_request(req);
1545 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
1548 struct ceph_inode_info *ci = ceph_inode(inode);
1549 s32 stripe_unit = ci->i_layout.stripe_unit;
1550 s32 stripe_count = ci->i_layout.stripe_count;
1551 s32 object_size = ci->i_layout.object_size;
1552 u64 object_set_size = object_size * stripe_count;
1555 /* round offset up to next period boundary */
1556 nearly = offset + object_set_size - 1;
1558 nearly -= do_div(t, object_set_size);
1560 while (length && offset < nearly) {
1561 loff_t size = length;
1562 ret = ceph_zero_partial_object(inode, offset, &size);
1568 while (length >= object_set_size) {
1570 loff_t pos = offset;
1571 for (i = 0; i < stripe_count; ++i) {
1572 ret = ceph_zero_partial_object(inode, pos, NULL);
1577 offset += object_set_size;
1578 length -= object_set_size;
1581 loff_t size = length;
1582 ret = ceph_zero_partial_object(inode, offset, &size);
1591 static long ceph_fallocate(struct file *file, int mode,
1592 loff_t offset, loff_t length)
1594 struct ceph_file_info *fi = file->private_data;
1595 struct inode *inode = file_inode(file);
1596 struct ceph_inode_info *ci = ceph_inode(inode);
1597 struct ceph_osd_client *osdc =
1598 &ceph_inode_to_client(inode)->client->osdc;
1599 struct ceph_cap_flush *prealloc_cf;
1606 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1609 if (!S_ISREG(inode->i_mode))
1612 prealloc_cf = ceph_alloc_cap_flush();
1618 if (ceph_snap(inode) != CEPH_NOSNAP) {
1623 if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) &&
1624 !(mode & FALLOC_FL_PUNCH_HOLE)) {
1629 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1630 ret = ceph_uninline_data(file, NULL);
1635 size = i_size_read(inode);
1636 if (!(mode & FALLOC_FL_KEEP_SIZE))
1637 endoff = offset + length;
1639 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1640 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1642 want = CEPH_CAP_FILE_BUFFER;
1644 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, endoff, &got, NULL);
1648 if (mode & FALLOC_FL_PUNCH_HOLE) {
1650 ceph_zero_pagecache_range(inode, offset, length);
1651 ret = ceph_zero_objects(inode, offset, length);
1652 } else if (endoff > size) {
1653 truncate_pagecache_range(inode, size, -1);
1654 if (ceph_inode_set_size(inode, endoff))
1655 ceph_check_caps(ceph_inode(inode),
1656 CHECK_CAPS_AUTHONLY, NULL);
1660 spin_lock(&ci->i_ceph_lock);
1661 ci->i_inline_version = CEPH_INLINE_NONE;
1662 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1664 spin_unlock(&ci->i_ceph_lock);
1666 __mark_inode_dirty(inode, dirty);
1669 ceph_put_cap_refs(ci, got);
1671 inode_unlock(inode);
1672 ceph_free_cap_flush(prealloc_cf);
1676 const struct file_operations ceph_file_fops = {
1678 .release = ceph_release,
1679 .llseek = ceph_llseek,
1680 .read_iter = ceph_read_iter,
1681 .write_iter = ceph_write_iter,
1683 .fsync = ceph_fsync,
1685 .flock = ceph_flock,
1686 .splice_read = generic_file_splice_read,
1687 .splice_write = iter_file_splice_write,
1688 .unlocked_ioctl = ceph_ioctl,
1689 .compat_ioctl = ceph_ioctl,
1690 .fallocate = ceph_fallocate,