1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/ceph/ceph_debug.h>
3 #include <linux/ceph/striper.h>
5 #include <linux/module.h>
6 #include <linux/sched.h>
7 #include <linux/slab.h>
8 #include <linux/file.h>
9 #include <linux/mount.h>
10 #include <linux/namei.h>
11 #include <linux/writeback.h>
12 #include <linux/falloc.h>
13 #include <linux/iversion.h>
16 #include "mds_client.h"
19 static __le32 ceph_flags_sys2wire(u32 flags)
23 switch (flags & O_ACCMODE) {
25 wire_flags |= CEPH_O_RDONLY;
28 wire_flags |= CEPH_O_WRONLY;
31 wire_flags |= CEPH_O_RDWR;
37 #define ceph_sys2wire(a) if (flags & a) { wire_flags |= CEPH_##a; flags &= ~a; }
39 ceph_sys2wire(O_CREAT);
40 ceph_sys2wire(O_EXCL);
41 ceph_sys2wire(O_TRUNC);
42 ceph_sys2wire(O_DIRECTORY);
43 ceph_sys2wire(O_NOFOLLOW);
48 dout("unused open flags: %x\n", flags);
50 return cpu_to_le32(wire_flags);
54 * Ceph file operations
56 * Implement basic open/close functionality, and implement
59 * We implement three modes of file I/O:
60 * - buffered uses the generic_file_aio_{read,write} helpers
62 * - synchronous is used when there is multi-client read/write
63 * sharing, avoids the page cache, and synchronously waits for an
66 * - direct io takes the variant of the sync path that references
67 * user pages directly.
69 * fsync() flushes and waits on dirty pages, but just queues metadata
70 * for writeback: since the MDS can recover size and mtime there is no
71 * need to wait for MDS acknowledgement.
75 * How many pages to get in one call to iov_iter_get_pages(). This
76 * determines the size of the on-stack array used as a buffer.
78 #define ITER_GET_BVECS_PAGES 64
80 static ssize_t __iter_get_bvecs(struct iov_iter *iter, size_t maxsize,
81 struct bio_vec *bvecs)
86 if (maxsize > iov_iter_count(iter))
87 maxsize = iov_iter_count(iter);
89 while (size < maxsize) {
90 struct page *pages[ITER_GET_BVECS_PAGES];
95 bytes = iov_iter_get_pages(iter, pages, maxsize - size,
96 ITER_GET_BVECS_PAGES, &start);
100 iov_iter_advance(iter, bytes);
103 for ( ; bytes; idx++, bvec_idx++) {
104 struct bio_vec bv = {
105 .bv_page = pages[idx],
106 .bv_len = min_t(int, bytes, PAGE_SIZE - start),
110 bvecs[bvec_idx] = bv;
120 * iov_iter_get_pages() only considers one iov_iter segment, no matter
121 * what maxsize or maxpages are given. For ITER_BVEC that is a single
124 * Attempt to get up to @maxsize bytes worth of pages from @iter.
125 * Return the number of bytes in the created bio_vec array, or an error.
127 static ssize_t iter_get_bvecs_alloc(struct iov_iter *iter, size_t maxsize,
128 struct bio_vec **bvecs, int *num_bvecs)
131 size_t orig_count = iov_iter_count(iter);
135 iov_iter_truncate(iter, maxsize);
136 npages = iov_iter_npages(iter, INT_MAX);
137 iov_iter_reexpand(iter, orig_count);
140 * __iter_get_bvecs() may populate only part of the array -- zero it
143 bv = kvmalloc_array(npages, sizeof(*bv), GFP_KERNEL | __GFP_ZERO);
147 bytes = __iter_get_bvecs(iter, maxsize, bv);
150 * No pages were pinned -- just free the array.
161 static void put_bvecs(struct bio_vec *bvecs, int num_bvecs, bool should_dirty)
165 for (i = 0; i < num_bvecs; i++) {
166 if (bvecs[i].bv_page) {
168 set_page_dirty_lock(bvecs[i].bv_page);
169 put_page(bvecs[i].bv_page);
176 * Prepare an open request. Preallocate ceph_cap to avoid an
177 * inopportune ENOMEM later.
179 static struct ceph_mds_request *
180 prepare_open_request(struct super_block *sb, int flags, int create_mode)
182 struct ceph_fs_client *fsc = ceph_sb_to_client(sb);
183 struct ceph_mds_client *mdsc = fsc->mdsc;
184 struct ceph_mds_request *req;
185 int want_auth = USE_ANY_MDS;
186 int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
188 if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
189 want_auth = USE_AUTH_MDS;
191 req = ceph_mdsc_create_request(mdsc, op, want_auth);
194 req->r_fmode = ceph_flags_to_mode(flags);
195 req->r_args.open.flags = ceph_flags_sys2wire(flags);
196 req->r_args.open.mode = cpu_to_le32(create_mode);
201 static int ceph_init_file_info(struct inode *inode, struct file *file,
202 int fmode, bool isdir)
204 struct ceph_inode_info *ci = ceph_inode(inode);
205 struct ceph_file_info *fi;
207 dout("%s %p %p 0%o (%s)\n", __func__, inode, file,
208 inode->i_mode, isdir ? "dir" : "regular");
209 BUG_ON(inode->i_fop->release != ceph_release);
212 struct ceph_dir_file_info *dfi =
213 kmem_cache_zalloc(ceph_dir_file_cachep, GFP_KERNEL);
215 ceph_put_fmode(ci, fmode); /* clean up */
219 file->private_data = dfi;
220 fi = &dfi->file_info;
221 dfi->next_offset = 2;
222 dfi->readdir_cache_idx = -1;
224 fi = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
226 ceph_put_fmode(ci, fmode); /* clean up */
230 file->private_data = fi;
234 spin_lock_init(&fi->rw_contexts_lock);
235 INIT_LIST_HEAD(&fi->rw_contexts);
236 fi->meta_err = errseq_sample(&ci->i_meta_err);
237 fi->filp_gen = READ_ONCE(ceph_inode_to_client(inode)->filp_gen);
243 * initialize private struct file data.
244 * if we fail, clean up by dropping fmode reference on the ceph_inode
246 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
250 switch (inode->i_mode & S_IFMT) {
252 ceph_fscache_register_inode_cookie(inode);
253 ceph_fscache_file_set_cookie(inode, file);
256 ret = ceph_init_file_info(inode, file, fmode,
257 S_ISDIR(inode->i_mode));
263 dout("init_file %p %p 0%o (symlink)\n", inode, file,
265 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
269 dout("init_file %p %p 0%o (special)\n", inode, file,
272 * we need to drop the open ref now, since we don't
273 * have .release set to ceph_release.
275 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
276 BUG_ON(inode->i_fop->release == ceph_release);
278 /* call the proper open fop */
279 ret = inode->i_fop->open(inode, file);
285 * try renew caps after session gets killed.
287 int ceph_renew_caps(struct inode *inode)
289 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
290 struct ceph_inode_info *ci = ceph_inode(inode);
291 struct ceph_mds_request *req;
292 int err, flags, wanted;
294 spin_lock(&ci->i_ceph_lock);
295 wanted = __ceph_caps_file_wanted(ci);
296 if (__ceph_is_any_real_caps(ci) &&
297 (!(wanted & CEPH_CAP_ANY_WR) || ci->i_auth_cap)) {
298 int issued = __ceph_caps_issued(ci, NULL);
299 spin_unlock(&ci->i_ceph_lock);
300 dout("renew caps %p want %s issued %s updating mds_wanted\n",
301 inode, ceph_cap_string(wanted), ceph_cap_string(issued));
302 ceph_check_caps(ci, 0, NULL);
305 spin_unlock(&ci->i_ceph_lock);
308 if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
310 else if (wanted & CEPH_CAP_FILE_RD)
312 else if (wanted & CEPH_CAP_FILE_WR)
315 if (wanted & CEPH_CAP_FILE_LAZYIO)
319 req = prepare_open_request(inode->i_sb, flags, 0);
325 req->r_inode = inode;
330 err = ceph_mdsc_do_request(mdsc, NULL, req);
331 ceph_mdsc_put_request(req);
333 dout("renew caps %p open result=%d\n", inode, err);
334 return err < 0 ? err : 0;
338 * If we already have the requisite capabilities, we can satisfy
339 * the open request locally (no need to request new caps from the
340 * MDS). We do, however, need to inform the MDS (asynchronously)
341 * if our wanted caps set expands.
343 int ceph_open(struct inode *inode, struct file *file)
345 struct ceph_inode_info *ci = ceph_inode(inode);
346 struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
347 struct ceph_mds_client *mdsc = fsc->mdsc;
348 struct ceph_mds_request *req;
349 struct ceph_file_info *fi = file->private_data;
351 int flags, fmode, wanted;
354 dout("open file %p is already opened\n", file);
358 /* filter out O_CREAT|O_EXCL; vfs did that already. yuck. */
359 flags = file->f_flags & ~(O_CREAT|O_EXCL);
360 if (S_ISDIR(inode->i_mode))
361 flags = O_DIRECTORY; /* mds likes to know */
363 dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
364 ceph_vinop(inode), file, flags, file->f_flags);
365 fmode = ceph_flags_to_mode(flags);
366 wanted = ceph_caps_for_mode(fmode);
368 /* snapped files are read-only */
369 if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
372 /* trivially open snapdir */
373 if (ceph_snap(inode) == CEPH_SNAPDIR) {
374 spin_lock(&ci->i_ceph_lock);
375 __ceph_get_fmode(ci, fmode);
376 spin_unlock(&ci->i_ceph_lock);
377 return ceph_init_file(inode, file, fmode);
381 * No need to block if we have caps on the auth MDS (for
382 * write) or any MDS (for read). Update wanted set
385 spin_lock(&ci->i_ceph_lock);
386 if (__ceph_is_any_real_caps(ci) &&
387 (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
388 int mds_wanted = __ceph_caps_mds_wanted(ci, true);
389 int issued = __ceph_caps_issued(ci, NULL);
391 dout("open %p fmode %d want %s issued %s using existing\n",
392 inode, fmode, ceph_cap_string(wanted),
393 ceph_cap_string(issued));
394 __ceph_get_fmode(ci, fmode);
395 spin_unlock(&ci->i_ceph_lock);
398 if ((issued & wanted) != wanted &&
399 (mds_wanted & wanted) != wanted &&
400 ceph_snap(inode) != CEPH_SNAPDIR)
401 ceph_check_caps(ci, 0, NULL);
403 return ceph_init_file(inode, file, fmode);
404 } else if (ceph_snap(inode) != CEPH_NOSNAP &&
405 (ci->i_snap_caps & wanted) == wanted) {
406 __ceph_get_fmode(ci, fmode);
407 spin_unlock(&ci->i_ceph_lock);
408 return ceph_init_file(inode, file, fmode);
411 spin_unlock(&ci->i_ceph_lock);
413 dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
414 req = prepare_open_request(inode->i_sb, flags, 0);
419 req->r_inode = inode;
423 err = ceph_mdsc_do_request(mdsc, NULL, req);
425 err = ceph_init_file(inode, file, req->r_fmode);
426 ceph_mdsc_put_request(req);
427 dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
434 * Do a lookup + open with a single request. If we get a non-existent
435 * file or symlink, return 1 so the VFS can retry.
437 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
438 struct file *file, unsigned flags, umode_t mode)
440 struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
441 struct ceph_mds_client *mdsc = fsc->mdsc;
442 struct ceph_mds_request *req;
444 struct ceph_acl_sec_ctx as_ctx = {};
448 dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
450 d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
452 if (dentry->d_name.len > NAME_MAX)
453 return -ENAMETOOLONG;
455 if (flags & O_CREAT) {
456 if (ceph_quota_is_max_files_exceeded(dir))
458 err = ceph_pre_init_acls(dir, &mode, &as_ctx);
461 err = ceph_security_init_secctx(dentry, mode, &as_ctx);
467 req = prepare_open_request(dir->i_sb, flags, mode);
472 req->r_dentry = dget(dentry);
474 if (flags & O_CREAT) {
475 req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
476 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
477 if (as_ctx.pagelist) {
478 req->r_pagelist = as_ctx.pagelist;
479 as_ctx.pagelist = NULL;
483 mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
484 if (ceph_security_xattr_wanted(dir))
485 mask |= CEPH_CAP_XATTR_SHARED;
486 req->r_args.open.mask = cpu_to_le32(mask);
489 set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
490 err = ceph_mdsc_do_request(mdsc,
491 (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
493 err = ceph_handle_snapdir(req, dentry, err);
497 if ((flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
498 err = ceph_handle_notrace_create(dir, dentry);
500 if (d_in_lookup(dentry)) {
501 dn = ceph_finish_lookup(req, dentry, err);
505 /* we were given a hashed negative dentry */
510 if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
511 /* make vfs retry on splice, ENOENT, or symlink */
512 dout("atomic_open finish_no_open on dn %p\n", dn);
513 err = finish_no_open(file, dn);
515 dout("atomic_open finish_open on dn %p\n", dn);
516 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
517 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
518 file->f_mode |= FMODE_CREATED;
520 err = finish_open(file, dentry, ceph_open);
523 if (!req->r_err && req->r_target_inode)
524 ceph_put_fmode(ceph_inode(req->r_target_inode), req->r_fmode);
525 ceph_mdsc_put_request(req);
527 ceph_release_acl_sec_ctx(&as_ctx);
528 dout("atomic_open result=%d\n", err);
532 int ceph_release(struct inode *inode, struct file *file)
534 struct ceph_inode_info *ci = ceph_inode(inode);
536 if (S_ISDIR(inode->i_mode)) {
537 struct ceph_dir_file_info *dfi = file->private_data;
538 dout("release inode %p dir file %p\n", inode, file);
539 WARN_ON(!list_empty(&dfi->file_info.rw_contexts));
541 ceph_put_fmode(ci, dfi->file_info.fmode);
543 if (dfi->last_readdir)
544 ceph_mdsc_put_request(dfi->last_readdir);
545 kfree(dfi->last_name);
546 kfree(dfi->dir_info);
547 kmem_cache_free(ceph_dir_file_cachep, dfi);
549 struct ceph_file_info *fi = file->private_data;
550 dout("release inode %p regular file %p\n", inode, file);
551 WARN_ON(!list_empty(&fi->rw_contexts));
553 ceph_put_fmode(ci, fi->fmode);
554 kmem_cache_free(ceph_file_cachep, fi);
557 /* wake up anyone waiting for caps on this inode */
558 wake_up_all(&ci->i_cap_wq);
569 * Completely synchronous read and write methods. Direct from __user
570 * buffer to osd, or directly to user pages (if O_DIRECT).
572 * If the read spans object boundary, just do multiple reads. (That's not
573 * atomic, but good enough for now.)
575 * If we get a short result from the OSD, check against i_size; we need to
576 * only return a short read to the caller if we hit EOF.
578 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *to,
581 struct file *file = iocb->ki_filp;
582 struct inode *inode = file_inode(file);
583 struct ceph_inode_info *ci = ceph_inode(inode);
584 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
585 struct ceph_osd_client *osdc = &fsc->client->osdc;
587 u64 off = iocb->ki_pos;
588 u64 len = iov_iter_count(to);
590 dout("sync_read on file %p %llu~%u %s\n", file, off, (unsigned)len,
591 (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
596 * flush any page cache pages in this range. this
597 * will make concurrent normal and sync io slow,
598 * but it will at least behave sensibly when they are
601 ret = filemap_write_and_wait_range(inode->i_mapping,
607 while ((len = iov_iter_count(to)) > 0) {
608 struct ceph_osd_request *req;
615 req = ceph_osdc_new_request(osdc, &ci->i_layout,
616 ci->i_vino, off, &len, 0, 1,
617 CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
618 NULL, ci->i_truncate_seq,
619 ci->i_truncate_size, false);
625 more = len < iov_iter_count(to);
627 if (unlikely(iov_iter_is_pipe(to))) {
628 ret = iov_iter_get_pages_alloc(to, &pages, len,
631 ceph_osdc_put_request(req);
635 num_pages = DIV_ROUND_UP(ret + page_off, PAGE_SIZE);
638 osd_req_op_extent_update(req, 0, len);
642 num_pages = calc_pages_for(off, len);
643 page_off = off & ~PAGE_MASK;
644 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
646 ceph_osdc_put_request(req);
647 ret = PTR_ERR(pages);
652 osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_off,
654 ret = ceph_osdc_start_request(osdc, req, false);
656 ret = ceph_osdc_wait_request(osdc, req);
657 ceph_osdc_put_request(req);
659 i_size = i_size_read(inode);
660 dout("sync_read %llu~%llu got %zd i_size %llu%s\n",
661 off, len, ret, i_size, (more ? " MORE" : ""));
665 if (ret >= 0 && ret < len && (off + ret < i_size)) {
666 int zlen = min(len - ret, i_size - off - ret);
667 int zoff = page_off + ret;
668 dout("sync_read zero gap %llu~%llu\n",
669 off + ret, off + ret + zlen);
670 ceph_zero_page_vector_range(zoff, zlen, pages);
674 if (unlikely(iov_iter_is_pipe(to))) {
676 iov_iter_advance(to, ret);
679 iov_iter_advance(to, 0);
681 ceph_put_page_vector(pages, num_pages, false);
684 size_t left = ret > 0 ? ret : 0;
687 page_off = off & ~PAGE_MASK;
688 len = min_t(size_t, left, PAGE_SIZE - page_off);
689 copied = copy_page_to_iter(pages[idx++],
698 ceph_release_page_vector(pages, num_pages);
701 if (ret <= 0 || off >= i_size || !more)
705 if (off > iocb->ki_pos) {
707 iov_iter_count(to) > 0 && off >= i_size_read(inode))
708 *retry_op = CHECK_EOF;
709 ret = off - iocb->ki_pos;
713 dout("sync_read result %zd retry_op %d\n", ret, *retry_op);
717 struct ceph_aio_request {
723 struct list_head osd_reqs;
725 atomic_t pending_reqs;
726 struct timespec64 mtime;
727 struct ceph_cap_flush *prealloc_cf;
730 struct ceph_aio_work {
731 struct work_struct work;
732 struct ceph_osd_request *req;
735 static void ceph_aio_retry_work(struct work_struct *work);
737 static void ceph_aio_complete(struct inode *inode,
738 struct ceph_aio_request *aio_req)
740 struct ceph_inode_info *ci = ceph_inode(inode);
743 if (!atomic_dec_and_test(&aio_req->pending_reqs))
746 ret = aio_req->error;
748 ret = aio_req->total_len;
750 dout("ceph_aio_complete %p rc %d\n", inode, ret);
752 if (ret >= 0 && aio_req->write) {
755 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
756 if (endoff > i_size_read(inode)) {
757 if (ceph_inode_set_size(inode, endoff))
758 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
761 spin_lock(&ci->i_ceph_lock);
762 ci->i_inline_version = CEPH_INLINE_NONE;
763 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
764 &aio_req->prealloc_cf);
765 spin_unlock(&ci->i_ceph_lock);
767 __mark_inode_dirty(inode, dirty);
771 ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
774 aio_req->iocb->ki_complete(aio_req->iocb, ret, 0);
776 ceph_free_cap_flush(aio_req->prealloc_cf);
780 static void ceph_aio_complete_req(struct ceph_osd_request *req)
782 int rc = req->r_result;
783 struct inode *inode = req->r_inode;
784 struct ceph_aio_request *aio_req = req->r_priv;
785 struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
787 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_BVECS);
788 BUG_ON(!osd_data->num_bvecs);
790 dout("ceph_aio_complete_req %p rc %d bytes %u\n",
791 inode, rc, osd_data->bvec_pos.iter.bi_size);
793 if (rc == -EOLDSNAPC) {
794 struct ceph_aio_work *aio_work;
795 BUG_ON(!aio_req->write);
797 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
799 INIT_WORK(&aio_work->work, ceph_aio_retry_work);
801 queue_work(ceph_inode_to_client(inode)->inode_wq,
806 } else if (!aio_req->write) {
809 if (rc >= 0 && osd_data->bvec_pos.iter.bi_size > rc) {
811 int zlen = osd_data->bvec_pos.iter.bi_size - rc;
814 * If read is satisfied by single OSD request,
815 * it can pass EOF. Otherwise read is within
818 if (aio_req->num_reqs == 1) {
819 loff_t i_size = i_size_read(inode);
820 loff_t endoff = aio_req->iocb->ki_pos + rc;
822 zlen = min_t(size_t, zlen,
824 aio_req->total_len = rc + zlen;
827 iov_iter_bvec(&i, READ, osd_data->bvec_pos.bvecs,
829 osd_data->bvec_pos.iter.bi_size);
830 iov_iter_advance(&i, rc);
831 iov_iter_zero(zlen, &i);
835 put_bvecs(osd_data->bvec_pos.bvecs, osd_data->num_bvecs,
836 aio_req->should_dirty);
837 ceph_osdc_put_request(req);
840 cmpxchg(&aio_req->error, 0, rc);
842 ceph_aio_complete(inode, aio_req);
846 static void ceph_aio_retry_work(struct work_struct *work)
848 struct ceph_aio_work *aio_work =
849 container_of(work, struct ceph_aio_work, work);
850 struct ceph_osd_request *orig_req = aio_work->req;
851 struct ceph_aio_request *aio_req = orig_req->r_priv;
852 struct inode *inode = orig_req->r_inode;
853 struct ceph_inode_info *ci = ceph_inode(inode);
854 struct ceph_snap_context *snapc;
855 struct ceph_osd_request *req;
858 spin_lock(&ci->i_ceph_lock);
859 if (__ceph_have_pending_cap_snap(ci)) {
860 struct ceph_cap_snap *capsnap =
861 list_last_entry(&ci->i_cap_snaps,
862 struct ceph_cap_snap,
864 snapc = ceph_get_snap_context(capsnap->context);
866 BUG_ON(!ci->i_head_snapc);
867 snapc = ceph_get_snap_context(ci->i_head_snapc);
869 spin_unlock(&ci->i_ceph_lock);
871 req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 1,
879 req->r_flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
880 ceph_oloc_copy(&req->r_base_oloc, &orig_req->r_base_oloc);
881 ceph_oid_copy(&req->r_base_oid, &orig_req->r_base_oid);
883 req->r_ops[0] = orig_req->r_ops[0];
885 req->r_mtime = aio_req->mtime;
886 req->r_data_offset = req->r_ops[0].extent.offset;
888 ret = ceph_osdc_alloc_messages(req, GFP_NOFS);
890 ceph_osdc_put_request(req);
895 ceph_osdc_put_request(orig_req);
897 req->r_callback = ceph_aio_complete_req;
898 req->r_inode = inode;
899 req->r_priv = aio_req;
901 ret = ceph_osdc_start_request(req->r_osdc, req, false);
905 ceph_aio_complete_req(req);
908 ceph_put_snap_context(snapc);
913 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
914 struct ceph_snap_context *snapc,
915 struct ceph_cap_flush **pcf)
917 struct file *file = iocb->ki_filp;
918 struct inode *inode = file_inode(file);
919 struct ceph_inode_info *ci = ceph_inode(inode);
920 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
921 struct ceph_vino vino;
922 struct ceph_osd_request *req;
923 struct bio_vec *bvecs;
924 struct ceph_aio_request *aio_req = NULL;
928 struct timespec64 mtime = current_time(inode);
929 size_t count = iov_iter_count(iter);
930 loff_t pos = iocb->ki_pos;
931 bool write = iov_iter_rw(iter) == WRITE;
932 bool should_dirty = !write && iter_is_iovec(iter);
934 if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
937 dout("sync_direct_%s on file %p %lld~%u snapc %p seq %lld\n",
938 (write ? "write" : "read"), file, pos, (unsigned)count,
939 snapc, snapc ? snapc->seq : 0);
941 ret = filemap_write_and_wait_range(inode->i_mapping,
942 pos, pos + count - 1);
947 int ret2 = invalidate_inode_pages2_range(inode->i_mapping,
949 (pos + count - 1) >> PAGE_SHIFT);
951 dout("invalidate_inode_pages2_range returned %d\n", ret2);
953 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
955 flags = CEPH_OSD_FLAG_READ;
958 while (iov_iter_count(iter) > 0) {
959 u64 size = iov_iter_count(iter);
963 size = min_t(u64, size, fsc->mount_options->wsize);
965 size = min_t(u64, size, fsc->mount_options->rsize);
967 vino = ceph_vino(inode);
968 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
971 write ? CEPH_OSD_OP_WRITE :
982 len = iter_get_bvecs_alloc(iter, size, &bvecs, &num_pages);
984 ceph_osdc_put_request(req);
989 osd_req_op_extent_update(req, 0, len);
992 * To simplify error handling, allow AIO when IO within i_size
993 * or IO can be satisfied by single OSD request.
995 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
996 (len == count || pos + count <= i_size_read(inode))) {
997 aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
999 aio_req->iocb = iocb;
1000 aio_req->write = write;
1001 aio_req->should_dirty = should_dirty;
1002 INIT_LIST_HEAD(&aio_req->osd_reqs);
1004 aio_req->mtime = mtime;
1005 swap(aio_req->prealloc_cf, *pcf);
1013 * throw out any page cache pages in this range. this
1016 truncate_inode_pages_range(inode->i_mapping, pos,
1017 PAGE_ALIGN(pos + len) - 1);
1019 req->r_mtime = mtime;
1022 osd_req_op_extent_osd_data_bvecs(req, 0, bvecs, num_pages, len);
1025 aio_req->total_len += len;
1026 aio_req->num_reqs++;
1027 atomic_inc(&aio_req->pending_reqs);
1029 req->r_callback = ceph_aio_complete_req;
1030 req->r_inode = inode;
1031 req->r_priv = aio_req;
1032 list_add_tail(&req->r_private_item, &aio_req->osd_reqs);
1038 ret = ceph_osdc_start_request(req->r_osdc, req, false);
1040 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1042 size = i_size_read(inode);
1046 if (ret >= 0 && ret < len && pos + ret < size) {
1048 int zlen = min_t(size_t, len - ret,
1051 iov_iter_bvec(&i, READ, bvecs, num_pages, len);
1052 iov_iter_advance(&i, ret);
1053 iov_iter_zero(zlen, &i);
1060 put_bvecs(bvecs, num_pages, should_dirty);
1061 ceph_osdc_put_request(req);
1066 if (!write && pos >= size)
1069 if (write && pos > size) {
1070 if (ceph_inode_set_size(inode, pos))
1071 ceph_check_caps(ceph_inode(inode),
1072 CHECK_CAPS_AUTHONLY,
1078 LIST_HEAD(osd_reqs);
1080 if (aio_req->num_reqs == 0) {
1085 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
1088 list_splice(&aio_req->osd_reqs, &osd_reqs);
1089 while (!list_empty(&osd_reqs)) {
1090 req = list_first_entry(&osd_reqs,
1091 struct ceph_osd_request,
1093 list_del_init(&req->r_private_item);
1095 ret = ceph_osdc_start_request(req->r_osdc,
1098 req->r_result = ret;
1099 ceph_aio_complete_req(req);
1102 return -EIOCBQUEUED;
1105 if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
1106 ret = pos - iocb->ki_pos;
1113 * Synchronous write, straight from __user pointer or user pages.
1115 * If write spans object boundary, just do multiple writes. (For a
1116 * correct atomic write, we should e.g. take write locks on all
1117 * objects, rollback on failure, etc.)
1120 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
1121 struct ceph_snap_context *snapc)
1123 struct file *file = iocb->ki_filp;
1124 struct inode *inode = file_inode(file);
1125 struct ceph_inode_info *ci = ceph_inode(inode);
1126 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1127 struct ceph_vino vino;
1128 struct ceph_osd_request *req;
1129 struct page **pages;
1135 bool check_caps = false;
1136 struct timespec64 mtime = current_time(inode);
1137 size_t count = iov_iter_count(from);
1139 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1142 dout("sync_write on file %p %lld~%u snapc %p seq %lld\n",
1143 file, pos, (unsigned)count, snapc, snapc->seq);
1145 ret = filemap_write_and_wait_range(inode->i_mapping,
1146 pos, pos + count - 1);
1150 ret = invalidate_inode_pages2_range(inode->i_mapping,
1152 (pos + count - 1) >> PAGE_SHIFT);
1154 dout("invalidate_inode_pages2_range returned %d\n", ret);
1156 flags = /* CEPH_OSD_FLAG_ORDERSNAP | */ CEPH_OSD_FLAG_WRITE;
1158 while ((len = iov_iter_count(from)) > 0) {
1162 vino = ceph_vino(inode);
1163 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1164 vino, pos, &len, 0, 1,
1165 CEPH_OSD_OP_WRITE, flags, snapc,
1167 ci->i_truncate_size,
1175 * write from beginning of first page,
1176 * regardless of io alignment
1178 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1180 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1181 if (IS_ERR(pages)) {
1182 ret = PTR_ERR(pages);
1187 for (n = 0; n < num_pages; n++) {
1188 size_t plen = min_t(size_t, left, PAGE_SIZE);
1189 ret = copy_page_from_iter(pages[n], 0, plen, from);
1198 ceph_release_page_vector(pages, num_pages);
1202 req->r_inode = inode;
1204 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1207 req->r_mtime = mtime;
1208 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1210 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1213 ceph_osdc_put_request(req);
1215 ceph_set_error_write(ci);
1219 ceph_clear_error_write(ci);
1222 if (pos > i_size_read(inode)) {
1223 check_caps = ceph_inode_set_size(inode, pos);
1225 ceph_check_caps(ceph_inode(inode),
1226 CHECK_CAPS_AUTHONLY,
1232 if (ret != -EOLDSNAPC && written > 0) {
1240 * Wrap generic_file_aio_read with checks for cap bits on the inode.
1241 * Atomically grab references, so that those bits are not released
1242 * back to the MDS mid-read.
1244 * Hmm, the sync read case isn't actually async... should it be?
1246 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1248 struct file *filp = iocb->ki_filp;
1249 struct ceph_file_info *fi = filp->private_data;
1250 size_t len = iov_iter_count(to);
1251 struct inode *inode = file_inode(filp);
1252 struct ceph_inode_info *ci = ceph_inode(inode);
1253 struct page *pinned_page = NULL;
1256 int retry_op = 0, read = 0;
1259 dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1260 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1262 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1263 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1265 want = CEPH_CAP_FILE_CACHE;
1266 ret = ceph_get_caps(filp, CEPH_CAP_FILE_RD, want, -1,
1267 &got, &pinned_page);
1271 if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1272 (iocb->ki_flags & IOCB_DIRECT) ||
1273 (fi->flags & CEPH_F_SYNC)) {
1275 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1276 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1277 ceph_cap_string(got));
1279 if (ci->i_inline_version == CEPH_INLINE_NONE) {
1280 if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1281 ret = ceph_direct_read_write(iocb, to,
1283 if (ret >= 0 && ret < len)
1284 retry_op = CHECK_EOF;
1286 ret = ceph_sync_read(iocb, to, &retry_op);
1289 retry_op = READ_INLINE;
1292 CEPH_DEFINE_RW_CONTEXT(rw_ctx, got);
1293 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1294 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1295 ceph_cap_string(got));
1296 ceph_add_rw_context(fi, &rw_ctx);
1297 ret = generic_file_read_iter(iocb, to);
1298 ceph_del_rw_context(fi, &rw_ctx);
1300 dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1301 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1303 put_page(pinned_page);
1306 ceph_put_cap_refs(ci, got);
1307 if (retry_op > HAVE_RETRIED && ret >= 0) {
1309 struct page *page = NULL;
1311 if (retry_op == READ_INLINE) {
1312 page = __page_cache_alloc(GFP_KERNEL);
1317 statret = __ceph_do_getattr(inode, page,
1318 CEPH_STAT_CAP_INLINE_DATA, !!page);
1322 if (statret == -ENODATA) {
1323 BUG_ON(retry_op != READ_INLINE);
1329 i_size = i_size_read(inode);
1330 if (retry_op == READ_INLINE) {
1331 BUG_ON(ret > 0 || read > 0);
1332 if (iocb->ki_pos < i_size &&
1333 iocb->ki_pos < PAGE_SIZE) {
1334 loff_t end = min_t(loff_t, i_size,
1335 iocb->ki_pos + len);
1336 end = min_t(loff_t, end, PAGE_SIZE);
1338 zero_user_segment(page, statret, end);
1339 ret = copy_page_to_iter(page,
1340 iocb->ki_pos & ~PAGE_MASK,
1341 end - iocb->ki_pos, to);
1342 iocb->ki_pos += ret;
1345 if (iocb->ki_pos < i_size && read < len) {
1346 size_t zlen = min_t(size_t, len - read,
1347 i_size - iocb->ki_pos);
1348 ret = iov_iter_zero(zlen, to);
1349 iocb->ki_pos += ret;
1352 __free_pages(page, 0);
1356 /* hit EOF or hole? */
1357 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1359 dout("sync_read hit hole, ppos %lld < size %lld"
1360 ", reading more\n", iocb->ki_pos, i_size);
1364 retry_op = HAVE_RETRIED;
1376 * Take cap references to avoid releasing caps to MDS mid-write.
1378 * If we are synchronous, and write with an old snap context, the OSD
1379 * may return EOLDSNAPC. In that case, retry the write.. _after_
1380 * dropping our cap refs and allowing the pending snap to logically
1381 * complete _before_ this write occurs.
1383 * If we are near ENOSPC, write synchronously.
1385 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1387 struct file *file = iocb->ki_filp;
1388 struct ceph_file_info *fi = file->private_data;
1389 struct inode *inode = file_inode(file);
1390 struct ceph_inode_info *ci = ceph_inode(inode);
1391 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1392 struct ceph_cap_flush *prealloc_cf;
1393 ssize_t count, written = 0;
1396 loff_t limit = max(i_size_read(inode), fsc->max_file_size);
1398 if (ceph_snap(inode) != CEPH_NOSNAP)
1401 prealloc_cf = ceph_alloc_cap_flush();
1408 /* We can write back this queue in page reclaim */
1409 current->backing_dev_info = inode_to_bdi(inode);
1411 if (iocb->ki_flags & IOCB_APPEND) {
1412 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1417 err = generic_write_checks(iocb, from);
1422 if (unlikely(pos >= limit)) {
1426 iov_iter_truncate(from, limit - pos);
1429 count = iov_iter_count(from);
1430 if (ceph_quota_is_max_bytes_exceeded(inode, pos + count)) {
1435 err = file_remove_privs(file);
1439 err = file_update_time(file);
1443 inode_inc_iversion_raw(inode);
1445 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1446 err = ceph_uninline_data(file, NULL);
1451 /* FIXME: not complete since it doesn't account for being at quota */
1452 if (ceph_osdmap_flag(&fsc->client->osdc, CEPH_OSDMAP_FULL)) {
1457 dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1458 inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1459 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1460 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1462 want = CEPH_CAP_FILE_BUFFER;
1464 err = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, pos + count,
1469 dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1470 inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1472 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1473 (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC) ||
1474 (ci->i_ceph_flags & CEPH_I_ERROR_WRITE)) {
1475 struct ceph_snap_context *snapc;
1476 struct iov_iter data;
1477 inode_unlock(inode);
1479 spin_lock(&ci->i_ceph_lock);
1480 if (__ceph_have_pending_cap_snap(ci)) {
1481 struct ceph_cap_snap *capsnap =
1482 list_last_entry(&ci->i_cap_snaps,
1483 struct ceph_cap_snap,
1485 snapc = ceph_get_snap_context(capsnap->context);
1487 BUG_ON(!ci->i_head_snapc);
1488 snapc = ceph_get_snap_context(ci->i_head_snapc);
1490 spin_unlock(&ci->i_ceph_lock);
1492 /* we might need to revert back to that point */
1494 if (iocb->ki_flags & IOCB_DIRECT)
1495 written = ceph_direct_read_write(iocb, &data, snapc,
1498 written = ceph_sync_write(iocb, &data, pos, snapc);
1500 iov_iter_advance(from, written);
1501 ceph_put_snap_context(snapc);
1504 * No need to acquire the i_truncate_mutex. Because
1505 * the MDS revokes Fwb caps before sending truncate
1506 * message to us. We can't get Fwb cap while there
1507 * are pending vmtruncate. So write and vmtruncate
1508 * can not run at the same time
1510 written = generic_perform_write(file, from, pos);
1511 if (likely(written >= 0))
1512 iocb->ki_pos = pos + written;
1513 inode_unlock(inode);
1519 spin_lock(&ci->i_ceph_lock);
1520 ci->i_inline_version = CEPH_INLINE_NONE;
1521 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1523 spin_unlock(&ci->i_ceph_lock);
1525 __mark_inode_dirty(inode, dirty);
1526 if (ceph_quota_is_max_bytes_approaching(inode, iocb->ki_pos))
1527 ceph_check_caps(ci, CHECK_CAPS_NODELAY, NULL);
1530 dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n",
1531 inode, ceph_vinop(inode), pos, (unsigned)count,
1532 ceph_cap_string(got));
1533 ceph_put_cap_refs(ci, got);
1535 if (written == -EOLDSNAPC) {
1536 dout("aio_write %p %llx.%llx %llu~%u" "got EOLDSNAPC, retrying\n",
1537 inode, ceph_vinop(inode), pos, (unsigned)count);
1542 if (ceph_osdmap_flag(&fsc->client->osdc, CEPH_OSDMAP_NEARFULL))
1543 iocb->ki_flags |= IOCB_DSYNC;
1544 written = generic_write_sync(iocb, written);
1550 inode_unlock(inode);
1552 ceph_free_cap_flush(prealloc_cf);
1553 current->backing_dev_info = NULL;
1554 return written ? written : err;
1558 * llseek. be sure to verify file size on SEEK_END.
1560 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1562 struct inode *inode = file->f_mapping->host;
1563 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1569 if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1570 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1575 i_size = i_size_read(inode);
1582 * Here we special-case the lseek(fd, 0, SEEK_CUR)
1583 * position-querying operation. Avoid rewriting the "same"
1584 * f_pos value back to the file because a concurrent read(),
1585 * write() or lseek() might have altered it
1591 offset += file->f_pos;
1594 if (offset < 0 || offset >= i_size) {
1600 if (offset < 0 || offset >= i_size) {
1608 ret = vfs_setpos(file, offset, max(i_size, fsc->max_file_size));
1611 inode_unlock(inode);
1615 static inline void ceph_zero_partial_page(
1616 struct inode *inode, loff_t offset, unsigned size)
1619 pgoff_t index = offset >> PAGE_SHIFT;
1621 page = find_lock_page(inode->i_mapping, index);
1623 wait_on_page_writeback(page);
1624 zero_user(page, offset & (PAGE_SIZE - 1), size);
1630 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1633 loff_t nearly = round_up(offset, PAGE_SIZE);
1634 if (offset < nearly) {
1635 loff_t size = nearly - offset;
1638 ceph_zero_partial_page(inode, offset, size);
1642 if (length >= PAGE_SIZE) {
1643 loff_t size = round_down(length, PAGE_SIZE);
1644 truncate_pagecache_range(inode, offset, offset + size - 1);
1649 ceph_zero_partial_page(inode, offset, length);
1652 static int ceph_zero_partial_object(struct inode *inode,
1653 loff_t offset, loff_t *length)
1655 struct ceph_inode_info *ci = ceph_inode(inode);
1656 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1657 struct ceph_osd_request *req;
1663 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1666 op = CEPH_OSD_OP_ZERO;
1669 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1673 CEPH_OSD_FLAG_WRITE,
1680 req->r_mtime = inode->i_mtime;
1681 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1683 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1687 ceph_osdc_put_request(req);
1693 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
1696 struct ceph_inode_info *ci = ceph_inode(inode);
1697 s32 stripe_unit = ci->i_layout.stripe_unit;
1698 s32 stripe_count = ci->i_layout.stripe_count;
1699 s32 object_size = ci->i_layout.object_size;
1700 u64 object_set_size = object_size * stripe_count;
1703 /* round offset up to next period boundary */
1704 nearly = offset + object_set_size - 1;
1706 nearly -= do_div(t, object_set_size);
1708 while (length && offset < nearly) {
1709 loff_t size = length;
1710 ret = ceph_zero_partial_object(inode, offset, &size);
1716 while (length >= object_set_size) {
1718 loff_t pos = offset;
1719 for (i = 0; i < stripe_count; ++i) {
1720 ret = ceph_zero_partial_object(inode, pos, NULL);
1725 offset += object_set_size;
1726 length -= object_set_size;
1729 loff_t size = length;
1730 ret = ceph_zero_partial_object(inode, offset, &size);
1739 static long ceph_fallocate(struct file *file, int mode,
1740 loff_t offset, loff_t length)
1742 struct ceph_file_info *fi = file->private_data;
1743 struct inode *inode = file_inode(file);
1744 struct ceph_inode_info *ci = ceph_inode(inode);
1745 struct ceph_cap_flush *prealloc_cf;
1752 if (mode != (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1755 if (!S_ISREG(inode->i_mode))
1758 prealloc_cf = ceph_alloc_cap_flush();
1764 if (ceph_snap(inode) != CEPH_NOSNAP) {
1769 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1770 ret = ceph_uninline_data(file, NULL);
1775 size = i_size_read(inode);
1777 /* Are we punching a hole beyond EOF? */
1780 if ((offset + length) > size)
1781 length = size - offset;
1783 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1784 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1786 want = CEPH_CAP_FILE_BUFFER;
1788 ret = ceph_get_caps(file, CEPH_CAP_FILE_WR, want, endoff, &got, NULL);
1792 ceph_zero_pagecache_range(inode, offset, length);
1793 ret = ceph_zero_objects(inode, offset, length);
1796 spin_lock(&ci->i_ceph_lock);
1797 ci->i_inline_version = CEPH_INLINE_NONE;
1798 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1800 spin_unlock(&ci->i_ceph_lock);
1802 __mark_inode_dirty(inode, dirty);
1805 ceph_put_cap_refs(ci, got);
1807 inode_unlock(inode);
1808 ceph_free_cap_flush(prealloc_cf);
1813 * This function tries to get FILE_WR capabilities for dst_ci and FILE_RD for
1814 * src_ci. Two attempts are made to obtain both caps, and an error is return if
1815 * this fails; zero is returned on success.
1817 static int get_rd_wr_caps(struct file *src_filp, int *src_got,
1818 struct file *dst_filp,
1819 loff_t dst_endoff, int *dst_got)
1822 bool retrying = false;
1825 ret = ceph_get_caps(dst_filp, CEPH_CAP_FILE_WR, CEPH_CAP_FILE_BUFFER,
1826 dst_endoff, dst_got, NULL);
1831 * Since we're already holding the FILE_WR capability for the dst file,
1832 * we would risk a deadlock by using ceph_get_caps. Thus, we'll do some
1833 * retry dance instead to try to get both capabilities.
1835 ret = ceph_try_get_caps(file_inode(src_filp),
1836 CEPH_CAP_FILE_RD, CEPH_CAP_FILE_SHARED,
1839 /* Start by dropping dst_ci caps and getting src_ci caps */
1840 ceph_put_cap_refs(ceph_inode(file_inode(dst_filp)), *dst_got);
1843 /* ceph_try_get_caps masks EAGAIN */
1847 ret = ceph_get_caps(src_filp, CEPH_CAP_FILE_RD,
1848 CEPH_CAP_FILE_SHARED, -1, src_got, NULL);
1851 /*... drop src_ci caps too, and retry */
1852 ceph_put_cap_refs(ceph_inode(file_inode(src_filp)), *src_got);
1859 static void put_rd_wr_caps(struct ceph_inode_info *src_ci, int src_got,
1860 struct ceph_inode_info *dst_ci, int dst_got)
1862 ceph_put_cap_refs(src_ci, src_got);
1863 ceph_put_cap_refs(dst_ci, dst_got);
1867 * This function does several size-related checks, returning an error if:
1868 * - source file is smaller than off+len
1869 * - destination file size is not OK (inode_newsize_ok())
1870 * - max bytes quotas is exceeded
1872 static int is_file_size_ok(struct inode *src_inode, struct inode *dst_inode,
1873 loff_t src_off, loff_t dst_off, size_t len)
1875 loff_t size, endoff;
1877 size = i_size_read(src_inode);
1879 * Don't copy beyond source file EOF. Instead of simply setting length
1880 * to (size - src_off), just drop to VFS default implementation, as the
1881 * local i_size may be stale due to other clients writing to the source
1884 if (src_off + len > size) {
1885 dout("Copy beyond EOF (%llu + %zu > %llu)\n",
1886 src_off, len, size);
1889 size = i_size_read(dst_inode);
1891 endoff = dst_off + len;
1892 if (inode_newsize_ok(dst_inode, endoff))
1895 if (ceph_quota_is_max_bytes_exceeded(dst_inode, endoff))
1901 static ssize_t __ceph_copy_file_range(struct file *src_file, loff_t src_off,
1902 struct file *dst_file, loff_t dst_off,
1903 size_t len, unsigned int flags)
1905 struct inode *src_inode = file_inode(src_file);
1906 struct inode *dst_inode = file_inode(dst_file);
1907 struct ceph_inode_info *src_ci = ceph_inode(src_inode);
1908 struct ceph_inode_info *dst_ci = ceph_inode(dst_inode);
1909 struct ceph_cap_flush *prealloc_cf;
1910 struct ceph_object_locator src_oloc, dst_oloc;
1911 struct ceph_object_id src_oid, dst_oid;
1912 loff_t endoff = 0, size;
1914 u64 src_objnum, dst_objnum, src_objoff, dst_objoff;
1915 u32 src_objlen, dst_objlen, object_size;
1916 int src_got = 0, dst_got = 0, err, dirty;
1917 bool do_final_copy = false;
1919 if (src_inode->i_sb != dst_inode->i_sb)
1921 if (ceph_snap(dst_inode) != CEPH_NOSNAP)
1925 * Some of the checks below will return -EOPNOTSUPP, which will force a
1926 * fallback to the default VFS copy_file_range implementation. This is
1927 * desirable in several cases (for ex, the 'len' is smaller than the
1928 * size of the objects, or in cases where that would be more
1932 if (ceph_test_mount_opt(ceph_inode_to_client(src_inode), NOCOPYFROM))
1935 if ((src_ci->i_layout.stripe_unit != dst_ci->i_layout.stripe_unit) ||
1936 (src_ci->i_layout.stripe_count != dst_ci->i_layout.stripe_count) ||
1937 (src_ci->i_layout.object_size != dst_ci->i_layout.object_size))
1940 if (len < src_ci->i_layout.object_size)
1941 return -EOPNOTSUPP; /* no remote copy will be done */
1943 prealloc_cf = ceph_alloc_cap_flush();
1947 /* Start by sync'ing the source and destination files */
1948 ret = file_write_and_wait_range(src_file, src_off, (src_off + len));
1950 dout("failed to write src file (%zd)\n", ret);
1953 ret = file_write_and_wait_range(dst_file, dst_off, (dst_off + len));
1955 dout("failed to write dst file (%zd)\n", ret);
1960 * We need FILE_WR caps for dst_ci and FILE_RD for src_ci as other
1961 * clients may have dirty data in their caches. And OSDs know nothing
1962 * about caps, so they can't safely do the remote object copies.
1964 err = get_rd_wr_caps(src_file, &src_got,
1965 dst_file, (dst_off + len), &dst_got);
1967 dout("get_rd_wr_caps returned %d\n", err);
1972 ret = is_file_size_ok(src_inode, dst_inode, src_off, dst_off, len);
1976 size = i_size_read(dst_inode);
1977 endoff = dst_off + len;
1979 /* Drop dst file cached pages */
1980 ret = invalidate_inode_pages2_range(dst_inode->i_mapping,
1981 dst_off >> PAGE_SHIFT,
1982 endoff >> PAGE_SHIFT);
1984 dout("Failed to invalidate inode pages (%zd)\n", ret);
1987 src_oloc.pool = src_ci->i_layout.pool_id;
1988 src_oloc.pool_ns = ceph_try_get_string(src_ci->i_layout.pool_ns);
1989 dst_oloc.pool = dst_ci->i_layout.pool_id;
1990 dst_oloc.pool_ns = ceph_try_get_string(dst_ci->i_layout.pool_ns);
1992 ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
1993 src_ci->i_layout.object_size,
1994 &src_objnum, &src_objoff, &src_objlen);
1995 ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
1996 dst_ci->i_layout.object_size,
1997 &dst_objnum, &dst_objoff, &dst_objlen);
1998 /* object-level offsets need to the same */
1999 if (src_objoff != dst_objoff) {
2005 * Do a manual copy if the object offset isn't object aligned.
2006 * 'src_objlen' contains the bytes left until the end of the object,
2007 * starting at the src_off
2011 * we need to temporarily drop all caps as we'll be calling
2012 * {read,write}_iter, which will get caps again.
2014 put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2015 ret = do_splice_direct(src_file, &src_off, dst_file,
2016 &dst_off, src_objlen, flags);
2018 dout("do_splice_direct returned %d\n", err);
2022 err = get_rd_wr_caps(src_file, &src_got,
2023 dst_file, (dst_off + len), &dst_got);
2026 err = is_file_size_ok(src_inode, dst_inode,
2027 src_off, dst_off, len);
2031 object_size = src_ci->i_layout.object_size;
2032 while (len >= object_size) {
2033 ceph_calc_file_object_mapping(&src_ci->i_layout, src_off,
2034 object_size, &src_objnum,
2035 &src_objoff, &src_objlen);
2036 ceph_calc_file_object_mapping(&dst_ci->i_layout, dst_off,
2037 object_size, &dst_objnum,
2038 &dst_objoff, &dst_objlen);
2039 ceph_oid_init(&src_oid);
2040 ceph_oid_printf(&src_oid, "%llx.%08llx",
2041 src_ci->i_vino.ino, src_objnum);
2042 ceph_oid_init(&dst_oid);
2043 ceph_oid_printf(&dst_oid, "%llx.%08llx",
2044 dst_ci->i_vino.ino, dst_objnum);
2045 /* Do an object remote copy */
2046 err = ceph_osdc_copy_from(
2047 &ceph_inode_to_client(src_inode)->client->osdc,
2048 src_ci->i_vino.snap, 0,
2049 &src_oid, &src_oloc,
2050 CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2051 CEPH_OSD_OP_FLAG_FADVISE_NOCACHE,
2052 &dst_oid, &dst_oloc,
2053 CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
2054 CEPH_OSD_OP_FLAG_FADVISE_DONTNEED, 0);
2056 dout("ceph_osdc_copy_from returned %d\n", err);
2062 src_off += object_size;
2063 dst_off += object_size;
2068 /* We still need one final local copy */
2069 do_final_copy = true;
2071 file_update_time(dst_file);
2072 inode_inc_iversion_raw(dst_inode);
2074 if (endoff > size) {
2077 /* Let the MDS know about dst file size change */
2078 if (ceph_quota_is_max_bytes_approaching(dst_inode, endoff))
2079 caps_flags |= CHECK_CAPS_NODELAY;
2080 if (ceph_inode_set_size(dst_inode, endoff))
2081 caps_flags |= CHECK_CAPS_AUTHONLY;
2083 ceph_check_caps(dst_ci, caps_flags, NULL);
2086 spin_lock(&dst_ci->i_ceph_lock);
2087 dst_ci->i_inline_version = CEPH_INLINE_NONE;
2088 dirty = __ceph_mark_dirty_caps(dst_ci, CEPH_CAP_FILE_WR, &prealloc_cf);
2089 spin_unlock(&dst_ci->i_ceph_lock);
2091 __mark_inode_dirty(dst_inode, dirty);
2094 put_rd_wr_caps(src_ci, src_got, dst_ci, dst_got);
2096 if (do_final_copy) {
2097 err = do_splice_direct(src_file, &src_off, dst_file,
2098 &dst_off, len, flags);
2100 dout("do_splice_direct returned %d\n", err);
2108 ceph_free_cap_flush(prealloc_cf);
2113 static ssize_t ceph_copy_file_range(struct file *src_file, loff_t src_off,
2114 struct file *dst_file, loff_t dst_off,
2115 size_t len, unsigned int flags)
2119 ret = __ceph_copy_file_range(src_file, src_off, dst_file, dst_off,
2122 if (ret == -EOPNOTSUPP || ret == -EXDEV)
2123 ret = generic_copy_file_range(src_file, src_off, dst_file,
2124 dst_off, len, flags);
2128 const struct file_operations ceph_file_fops = {
2130 .release = ceph_release,
2131 .llseek = ceph_llseek,
2132 .read_iter = ceph_read_iter,
2133 .write_iter = ceph_write_iter,
2135 .fsync = ceph_fsync,
2137 .flock = ceph_flock,
2138 .splice_read = generic_file_splice_read,
2139 .splice_write = iter_file_splice_write,
2140 .unlocked_ioctl = ceph_ioctl,
2141 .compat_ioctl = ceph_ioctl,
2142 .fallocate = ceph_fallocate,
2143 .copy_file_range = ceph_copy_file_range,