4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 * Implementation of cl_io for VVP layer.
33 * Author: Nikita Danilov <nikita.danilov@sun.com>
34 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
37 #define DEBUG_SUBSYSTEM S_LLITE
40 #include <linux/pagevec.h>
41 #include <linux/memcontrol.h>
42 #include <linux/falloc.h>
44 #include "llite_internal.h"
45 #include "vvp_internal.h"
46 #include <libcfs/linux/linux-misc.h>
48 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
49 const struct cl_io_slice *slice)
53 vio = container_of(slice, struct vvp_io, vui_cl);
54 LASSERT(vio == vvp_env_io(env));
60 * For swapping layout. The file's layout may have changed.
61 * To avoid populating pages to a wrong stripe, we have to verify the
62 * correctness of layout. It works because swapping layout processes
63 * have to acquire group lock.
65 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
68 struct ll_inode_info *lli = ll_i2info(inode);
69 struct vvp_io *vio = vvp_env_io(env);
72 switch (io->ci_type) {
75 /* don't need lock here to check lli_layout_gen as we have held
76 * extent lock and GROUP lock has to hold to swap layout */
77 if (ll_layout_version_get(lli) != vio->vui_layout_gen ||
78 OBD_FAIL_CHECK_RESET(OBD_FAIL_LLITE_LOST_LAYOUT, 0)) {
79 io->ci_need_restart = 1;
80 /* this will cause a short read/write */
85 /* fault is okay because we've already had a page. */
93 static void vvp_object_size_lock(struct cl_object *obj)
95 struct inode *inode = vvp_object_inode(obj);
97 ll_inode_size_lock(inode);
98 cl_object_attr_lock(obj);
101 static void vvp_object_size_unlock(struct cl_object *obj)
103 struct inode *inode = vvp_object_inode(obj);
105 cl_object_attr_unlock(obj);
106 ll_inode_size_unlock(inode);
110 * Helper function that if necessary adjusts file size (inode->i_size), when
111 * position at the offset \a pos is accessed. File size can be arbitrary stale
112 * on a Lustre client, but client at least knows KMS. If accessed area is
113 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
115 * Locking: i_size_lock is used to serialize changes to inode size and to
116 * protect consistency between inode size and cl_object
117 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
118 * top-object and sub-objects.
120 static int vvp_prep_size(const struct lu_env *env, struct cl_object *obj,
121 struct cl_io *io, loff_t start, size_t count,
124 struct cl_attr *attr = vvp_env_thread_attr(env);
125 struct inode *inode = vvp_object_inode(obj);
126 loff_t pos = start + count - 1;
131 * Consistency guarantees: following possibilities exist for the
132 * relation between region being accessed and real file size at this
135 * (A): the region is completely inside of the file;
137 * (B-x): x bytes of region are inside of the file, the rest is
140 * (C): the region is completely outside of the file.
142 * This classification is stable under DLM lock already acquired by
143 * the caller, because to change the class, other client has to take
144 * DLM lock conflicting with our lock. Also, any updates to ->i_size
145 * by other threads on this client are serialized by
146 * ll_inode_size_lock(). This guarantees that short reads are handled
147 * correctly in the face of concurrent writes and truncates.
149 vvp_object_size_lock(obj);
150 result = cl_object_attr_get(env, obj, attr);
155 * A glimpse is necessary to determine whether we
156 * return a short read (B) or some zeroes at the end
159 vvp_object_size_unlock(obj);
160 result = cl_glimpse_lock(env, io, inode, obj, 0);
161 if (result == 0 && exceed != NULL) {
162 /* If objective page index exceed end-of-file
163 * page index, return directly. Do not expect
164 * kernel will check such case correctly.
165 * linux-2.6.18-128.1.1 miss to do that.
167 loff_t size = i_size_read(inode);
168 unsigned long cur_index = start >>
171 if ((size == 0 && cur_index != 0) ||
172 (((size - 1) >> PAGE_SHIFT) <
180 * region is within kms and, hence, within real file
181 * size (A). We need to increase i_size to cover the
182 * read region so that generic_file_read() will do its
183 * job, but that doesn't mean the kms size is
184 * _correct_, it is only the _minimum_ size. If
185 * someone does a stat they will get the correct size
186 * which will always be >= the kms value here.
189 if (i_size_read(inode) < kms) {
190 i_size_write(inode, kms);
192 DFID" updating i_size %llu\n",
193 PFID(lu_object_fid(&obj->co_lu)),
194 (__u64)i_size_read(inode));
199 vvp_object_size_unlock(obj);
204 /*****************************************************************************
210 static int vvp_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
211 __u32 enqflags, enum cl_lock_mode mode,
212 pgoff_t start, pgoff_t end)
214 struct vvp_io *vio = vvp_env_io(env);
215 struct cl_lock_descr *descr = &vio->vui_link.cill_descr;
216 struct cl_object *obj = io->ci_obj;
218 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
221 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
223 memset(&vio->vui_link, 0, sizeof vio->vui_link);
225 if (vio->vui_fd && (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
226 descr->cld_mode = CLM_GROUP;
227 descr->cld_gid = vio->vui_fd->fd_grouplock.lg_gid;
228 enqflags |= CEF_LOCK_MATCH;
230 descr->cld_mode = mode;
233 descr->cld_obj = obj;
234 descr->cld_start = start;
235 descr->cld_end = end;
236 descr->cld_enq_flags = enqflags;
238 cl_io_lock_add(env, io, &vio->vui_link);
243 static int vvp_io_one_lock(const struct lu_env *env, struct cl_io *io,
244 __u32 enqflags, enum cl_lock_mode mode,
245 loff_t start, loff_t end)
247 struct cl_object *obj = io->ci_obj;
249 return vvp_io_one_lock_index(env, io, enqflags, mode,
250 cl_index(obj, start), cl_index(obj, end));
253 static int vvp_io_write_iter_init(const struct lu_env *env,
254 const struct cl_io_slice *ios)
256 struct vvp_io *vio = cl2vvp_io(env, ios);
258 cl_page_list_init(&vio->u.readwrite.vui_queue);
259 vio->u.readwrite.vui_written = 0;
260 vio->u.readwrite.vui_from = 0;
261 vio->u.readwrite.vui_to = PAGE_SIZE;
266 static int vvp_io_read_iter_init(const struct lu_env *env,
267 const struct cl_io_slice *ios)
269 struct vvp_io *vio = cl2vvp_io(env, ios);
271 vio->u.readwrite.vui_read = 0;
276 static void vvp_io_write_iter_fini(const struct lu_env *env,
277 const struct cl_io_slice *ios)
279 struct vvp_io *vio = cl2vvp_io(env, ios);
281 LASSERT(vio->u.readwrite.vui_queue.pl_nr == 0);
284 static int vvp_io_fault_iter_init(const struct lu_env *env,
285 const struct cl_io_slice *ios)
287 struct vvp_io *vio = cl2vvp_io(env, ios);
288 struct inode *inode = vvp_object_inode(ios->cis_obj);
290 LASSERT(inode == file_inode(vio->vui_fd->fd_file));
295 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
297 struct cl_io *io = ios->cis_io;
298 struct cl_object *obj = io->ci_obj;
299 struct vvp_io *vio = cl2vvp_io(env, ios);
300 struct inode *inode = vvp_object_inode(obj);
305 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
307 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
308 "need write layout %d, restore needed %d\n",
309 PFID(lu_object_fid(&obj->co_lu)),
310 io->ci_ignore_layout, io->ci_verify_layout,
311 vio->vui_layout_gen, io->ci_need_write_intent,
312 io->ci_restore_needed);
314 if (io->ci_restore_needed) {
315 /* file was detected release, we need to restore it
316 * before finishing the io
318 rc = ll_layout_restore(inode, 0, OBD_OBJECT_EOF);
319 /* if restore registration failed, no restart,
320 * we will return -ENODATA */
321 /* The layout will change after restore, so we need to
322 * block on layout lock held by the MDT
323 * as MDT will not send new layout in lvb (see LU-3124)
324 * we have to explicitly fetch it, all this will be done
325 * by ll_layout_refresh().
326 * Even if ll_layout_restore() returns zero, it doesn't mean
327 * that restore has been successful. Therefore it sets
328 * ci_verify_layout so that it will check layout at the end
332 io->ci_restore_needed = 1;
333 io->ci_need_restart = 0;
334 io->ci_verify_layout = 0;
339 io->ci_restore_needed = 0;
341 /* Even if ll_layout_restore() returns zero, it doesn't mean
342 * that restore has been successful. Therefore it should verify
343 * if there was layout change and restart I/O correspondingly.
345 ll_layout_refresh(inode, &gen);
346 io->ci_need_restart = vio->vui_layout_gen != gen;
347 if (io->ci_need_restart) {
349 DFID" layout changed from %d to %d.\n",
350 PFID(lu_object_fid(&obj->co_lu)),
351 vio->vui_layout_gen, gen);
352 /* today successful restore is the only possible
354 /* restore was done, clear restoring state */
355 clear_bit(LLIF_FILE_RESTORING,
356 &ll_i2info(vvp_object_inode(obj))->lli_flags);
362 * dynamic layout change needed, send layout intent
365 if (io->ci_need_write_intent) {
366 enum layout_intent_opc opc = LAYOUT_INTENT_WRITE;
368 io->ci_need_write_intent = 0;
370 LASSERT(io->ci_type == CIT_WRITE || cl_io_is_fallocate(io) ||
371 cl_io_is_trunc(io) || cl_io_is_mkwrite(io) ||
372 cl_io_is_fault_writable(io));
374 CDEBUG(D_VFSTRACE, DFID" write layout, type %u "DEXT"\n",
375 PFID(lu_object_fid(&obj->co_lu)), io->ci_type,
376 PEXT(&io->ci_write_intent));
378 if (cl_io_is_trunc(io))
379 opc = LAYOUT_INTENT_TRUNC;
381 rc = ll_layout_write_intent(inode, opc, &io->ci_write_intent);
384 io->ci_need_restart = 1;
388 if (!io->ci_need_restart &&
389 !io->ci_ignore_layout && io->ci_verify_layout) {
390 /* check layout version */
391 ll_layout_refresh(inode, &gen);
392 io->ci_need_restart = vio->vui_layout_gen != gen;
393 if (io->ci_need_restart) {
395 DFID" layout changed from %d to %d.\n",
396 PFID(lu_object_fid(&obj->co_lu)),
397 vio->vui_layout_gen, gen);
405 static void vvp_io_fault_fini(const struct lu_env *env,
406 const struct cl_io_slice *ios)
408 struct cl_io *io = ios->cis_io;
409 struct cl_page *page = io->u.ci_fault.ft_page;
411 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
414 lu_ref_del(&page->cp_reference, "fault", io);
415 cl_page_put(env, page);
416 io->u.ci_fault.ft_page = NULL;
418 vvp_io_fini(env, ios);
421 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
424 * we only want to hold PW locks if the mmap() can generate
425 * writes back to the file and that only happens in shared
428 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
433 static int vvp_mmap_locks(const struct lu_env *env,
434 struct vvp_io *vio, struct cl_io *io)
436 struct vvp_thread_info *vti = vvp_env_info(env);
437 struct mm_struct *mm = current->mm;
438 struct vm_area_struct *vma;
439 struct cl_lock_descr *descr = &vti->vti_descr;
440 union ldlm_policy_data policy;
448 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
450 /* nfs or loop back device write */
451 if (vio->vui_iter == NULL)
454 /* No MM (e.g. NFS)? No vmas too. */
458 if (!iter_is_iovec(vio->vui_iter) && !iov_iter_is_kvec(vio->vui_iter))
461 for (i = *vio->vui_iter;
463 iov_iter_advance(&i, iov.iov_len)) {
464 iov = iov_iter_iovec(&i);
465 addr = (unsigned long)iov.iov_base;
471 count += addr & ~PAGE_MASK;
475 while ((vma = our_vma(mm, addr, count)) != NULL) {
476 struct dentry *de = file_dentry(vma->vm_file);
477 struct inode *inode = de->d_inode;
478 int flags = CEF_MUST;
480 if (ll_file_nolock(vma->vm_file)) {
482 * For no lock case is not allowed for mmap
489 * XXX: Required lock mode can be weakened: CIT_WRITE
490 * io only ever reads user level buffer, and CIT_READ
493 policy_from_vma(&policy, vma, addr, count);
494 descr->cld_mode = vvp_mode_from_vma(vma);
495 descr->cld_obj = ll_i2info(inode)->lli_clob;
496 descr->cld_start = cl_index(descr->cld_obj,
497 policy.l_extent.start);
498 descr->cld_end = cl_index(descr->cld_obj,
499 policy.l_extent.end);
500 descr->cld_enq_flags = flags;
501 result = cl_io_lock_alloc_add(env, io, descr);
503 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
504 descr->cld_mode, descr->cld_start,
510 if (vma->vm_end - addr >= count)
513 count -= vma->vm_end - addr;
516 mmap_read_unlock(mm);
523 static void vvp_io_advance(const struct lu_env *env,
524 const struct cl_io_slice *ios,
527 struct cl_object *obj = ios->cis_io->ci_obj;
528 struct vvp_io *vio = cl2vvp_io(env, ios);
530 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
533 * Since 3.16(26978b8b4) vfs revert iov iter to
534 * original position even io succeed, so instead
535 * of relying on VFS, we move iov iter by ourselves.
537 iov_iter_advance(vio->vui_iter, nob);
538 CDEBUG(D_VFSTRACE, "advancing %ld bytes\n", nob);
539 vio->vui_tot_count -= nob;
540 iov_iter_reexpand(vio->vui_iter, vio->vui_tot_count);
543 static void vvp_io_update_iov(const struct lu_env *env,
544 struct vvp_io *vio, struct cl_io *io)
546 size_t size = io->u.ci_rw.crw_count;
551 iov_iter_truncate(vio->vui_iter, size);
554 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
555 enum cl_lock_mode mode, loff_t start, loff_t end)
557 struct vvp_io *vio = vvp_env_io(env);
561 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
564 vvp_io_update_iov(env, vio, io);
566 if (io->u.ci_rw.crw_nonblock)
567 ast_flags |= CEF_NONBLOCK;
568 if (io->ci_lock_no_expand)
569 ast_flags |= CEF_LOCK_NO_EXPAND;
571 /* Group lock held means no lockless any more */
572 if (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)
575 if (ll_file_nolock(vio->vui_fd->fd_file) ||
576 (vio->vui_fd->fd_file->f_flags & O_DIRECT &&
578 ast_flags |= CEF_NEVER;
581 result = vvp_mmap_locks(env, vio, io);
583 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
588 static int vvp_io_read_lock(const struct lu_env *env,
589 const struct cl_io_slice *ios)
591 struct cl_io *io = ios->cis_io;
592 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
596 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
597 rd->crw_pos + rd->crw_count - 1);
601 static int vvp_io_fault_lock(const struct lu_env *env,
602 const struct cl_io_slice *ios)
604 struct cl_io *io = ios->cis_io;
605 struct vvp_io *vio = cl2vvp_io(env, ios);
607 * XXX LDLM_FL_CBPENDING
609 return vvp_io_one_lock_index(env,
611 vvp_mode_from_vma(vio->u.fault.ft_vma),
612 io->u.ci_fault.ft_index,
613 io->u.ci_fault.ft_index);
616 static int vvp_io_write_lock(const struct lu_env *env,
617 const struct cl_io_slice *ios)
619 struct cl_io *io = ios->cis_io;
623 if (io->u.ci_wr.wr_append) {
625 end = OBD_OBJECT_EOF;
627 start = io->u.ci_wr.wr.crw_pos;
628 end = start + io->u.ci_wr.wr.crw_count - 1;
631 RETURN(vvp_io_rw_lock(env, io, CLM_WRITE, start, end));
634 static int vvp_io_setattr_iter_init(const struct lu_env *env,
635 const struct cl_io_slice *ios)
642 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
644 * Handles "lockless io" mode when extent locking is done by server.
646 static int vvp_io_setattr_lock(const struct lu_env *env,
647 const struct cl_io_slice *ios)
649 struct cl_io *io = ios->cis_io;
650 __u64 lock_start = 0;
651 __u64 lock_end = OBD_OBJECT_EOF;
654 if (cl_io_is_trunc(io)) {
655 struct inode *inode = vvp_object_inode(io->ci_obj);
657 /* set enqueue flags to CEF_MUST in case of encrypted file,
658 * to prevent lockless truncate
660 if (S_ISREG(inode->i_mode) && IS_ENCRYPTED(inode))
662 else if (io->u.ci_setattr.sa_attr.lvb_size == 0)
663 enqflags = CEF_DISCARD_DATA;
664 } else if (cl_io_is_fallocate(io)) {
665 lock_start = io->u.ci_setattr.sa_falloc_offset;
666 lock_end = io->u.ci_setattr.sa_falloc_end - 1;
668 unsigned int valid = io->u.ci_setattr.sa_avalid;
670 if (!(valid & TIMES_SET_FLAGS))
673 if ((!(valid & ATTR_MTIME) ||
674 io->u.ci_setattr.sa_attr.lvb_mtime >=
675 io->u.ci_setattr.sa_attr.lvb_ctime) &&
676 (!(valid & ATTR_ATIME) ||
677 io->u.ci_setattr.sa_attr.lvb_atime >=
678 io->u.ci_setattr.sa_attr.lvb_ctime))
682 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
683 lock_start, lock_end);
686 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
691 * Only ll_inode_size_lock is taken at this level.
693 ll_inode_size_lock(inode);
694 result = inode_newsize_ok(inode, size);
696 ll_inode_size_unlock(inode);
699 i_size_write(inode, size);
701 ll_truncate_pagecache(inode, size);
702 ll_inode_size_unlock(inode);
706 static int vvp_io_setattr_time(const struct lu_env *env,
707 const struct cl_io_slice *ios)
709 struct cl_io *io = ios->cis_io;
710 struct cl_object *obj = io->ci_obj;
711 struct cl_attr *attr = vvp_env_thread_attr(env);
713 unsigned valid = CAT_CTIME;
715 cl_object_attr_lock(obj);
716 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
717 if (io->u.ci_setattr.sa_avalid & ATTR_ATIME_SET) {
718 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
721 if (io->u.ci_setattr.sa_avalid & ATTR_MTIME_SET) {
722 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
725 result = cl_object_attr_update(env, obj, attr, valid);
726 cl_object_attr_unlock(obj);
731 static int vvp_io_setattr_start(const struct lu_env *env,
732 const struct cl_io_slice *ios)
734 struct cl_io *io = ios->cis_io;
735 struct inode *inode = vvp_object_inode(io->ci_obj);
736 struct ll_inode_info *lli = ll_i2info(inode);
737 int mode = io->u.ci_setattr.sa_falloc_mode;
739 if (cl_io_is_trunc(io)) {
740 trunc_sem_down_write(&lli->lli_trunc_sem);
741 mutex_lock(&lli->lli_setattr_mutex);
742 inode_dio_wait(inode);
743 } else if (cl_io_is_fallocate(io)) {
746 trunc_sem_down_write(&lli->lli_trunc_sem);
747 mutex_lock(&lli->lli_setattr_mutex);
748 inode_dio_wait(inode);
750 ll_merge_attr(env, inode);
751 size = i_size_read(inode);
752 if (io->u.ci_setattr.sa_falloc_end > size &&
753 !(mode & FALLOC_FL_KEEP_SIZE)) {
754 size = io->u.ci_setattr.sa_falloc_end;
755 io->u.ci_setattr.sa_avalid |= ATTR_SIZE;
757 io->u.ci_setattr.sa_attr.lvb_size = size;
759 mutex_lock(&lli->lli_setattr_mutex);
762 if (io->u.ci_setattr.sa_avalid & TIMES_SET_FLAGS)
763 return vvp_io_setattr_time(env, ios);
768 static void vvp_io_setattr_end(const struct lu_env *env,
769 const struct cl_io_slice *ios)
771 struct cl_io *io = ios->cis_io;
772 struct inode *inode = vvp_object_inode(io->ci_obj);
773 struct ll_inode_info *lli = ll_i2info(inode);
775 if (cl_io_is_trunc(io)) {
776 /* Truncate in memory pages - they must be clean pages
777 * because osc has already notified to destroy osc_extents. */
778 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
779 mutex_unlock(&lli->lli_setattr_mutex);
780 trunc_sem_up_write(&lli->lli_trunc_sem);
781 } else if (cl_io_is_fallocate(io)) {
782 mutex_unlock(&lli->lli_setattr_mutex);
783 trunc_sem_up_write(&lli->lli_trunc_sem);
785 mutex_unlock(&lli->lli_setattr_mutex);
789 static void vvp_io_setattr_fini(const struct lu_env *env,
790 const struct cl_io_slice *ios)
792 bool restore_needed = ios->cis_io->ci_restore_needed;
793 struct inode *inode = vvp_object_inode(ios->cis_obj);
795 vvp_io_fini(env, ios);
797 if (restore_needed && !ios->cis_io->ci_restore_needed) {
798 /* restore finished, set data modified flag for HSM */
799 set_bit(LLIF_DATA_MODIFIED, &ll_i2info(inode)->lli_flags);
803 static int vvp_io_read_start(const struct lu_env *env,
804 const struct cl_io_slice *ios)
806 struct vvp_io *vio = cl2vvp_io(env, ios);
807 struct cl_io *io = ios->cis_io;
808 struct cl_object *obj = io->ci_obj;
809 struct inode *inode = vvp_object_inode(obj);
810 struct ll_inode_info *lli = ll_i2info(inode);
811 struct file *file = vio->vui_fd->fd_file;
812 loff_t pos = io->u.ci_rd.rd.crw_pos;
813 size_t cnt = io->u.ci_rd.rd.crw_count;
814 size_t tot = vio->vui_tot_count;
817 struct iov_iter iter;
822 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
824 CDEBUG(D_VFSTRACE, "%s: read [%llu, %llu)\n",
825 file_dentry(file)->d_name.name,
828 trunc_sem_down_read(&lli->lli_trunc_sem);
830 if (io->ci_async_readahead) {
835 if (!can_populate_pages(env, io, inode))
838 if (!(file->f_flags & O_DIRECT)) {
839 result = cl_io_lru_reserve(env, io, pos, cnt);
844 /* Unless this is reading a sparse file, otherwise the lock has already
845 * been acquired so vvp_prep_size() is an empty op. */
846 result = vvp_prep_size(env, obj, io, pos, cnt, &exceed);
849 else if (exceed != 0)
852 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
853 "Read ino %lu, %zu bytes, offset %lld, size %llu\n",
854 inode->i_ino, cnt, pos, i_size_read(inode));
856 /* turn off the kernel's read-ahead */
857 vio->vui_fd->fd_file->f_ra.ra_pages = 0;
859 /* initialize read-ahead window once per syscall */
860 if (!vio->vui_ra_valid) {
861 vio->vui_ra_valid = true;
862 vio->vui_ra_start_idx = cl_index(obj, pos);
863 vio->vui_ra_pages = 0;
864 page_offset = pos & ~PAGE_MASK;
867 if (tot > PAGE_SIZE - page_offset)
868 tot -= (PAGE_SIZE - page_offset);
872 vio->vui_ra_pages += (tot + PAGE_SIZE - 1) >> PAGE_SHIFT;
874 CDEBUG(D_READA, "tot %zu, ra_start %lu, ra_count %lu\n",
875 vio->vui_tot_count, vio->vui_ra_start_idx,
881 LASSERT(vio->vui_iocb->ki_pos == pos);
882 iter = *vio->vui_iter;
883 result = generic_file_read_iter(vio->vui_iocb, &iter);
888 io->ci_nob += result;
890 } else if (result == -EIOCBQUEUED) {
891 io->ci_nob += vio->u.readwrite.vui_read;
892 vio->vui_iocb->ki_pos = pos + vio->u.readwrite.vui_read;
898 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
899 struct cl_page_list *plist, int from, int to)
901 struct cl_2queue *queue = &io->ci_queue;
902 struct cl_page *page;
903 unsigned int bytes = 0;
907 if (plist->pl_nr == 0)
910 if (from > 0 || to != PAGE_SIZE) {
911 page = cl_page_list_first(plist);
912 if (plist->pl_nr == 1) {
913 cl_page_clip(env, page, from, to);
916 cl_page_clip(env, page, from, PAGE_SIZE);
917 if (to != PAGE_SIZE) {
918 page = cl_page_list_last(plist);
919 cl_page_clip(env, page, 0, to);
924 cl_2queue_init(queue);
925 cl_page_list_splice(plist, &queue->c2_qin);
926 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
928 /* plist is not sorted any more */
929 cl_page_list_splice(&queue->c2_qin, plist);
930 cl_page_list_splice(&queue->c2_qout, plist);
931 cl_2queue_fini(env, queue);
934 /* calculate bytes */
935 bytes = plist->pl_nr << PAGE_SHIFT;
936 bytes -= from + PAGE_SIZE - to;
938 while (plist->pl_nr > 0) {
939 page = cl_page_list_first(plist);
940 cl_page_list_del(env, plist, page);
942 cl_page_clip(env, page, 0, PAGE_SIZE);
944 SetPageUptodate(cl_page_vmpage(page));
945 cl_page_disown(env, io, page);
947 /* held in ll_cl_init() */
948 lu_ref_del(&page->cp_reference, "cl_io", io);
949 cl_page_put(env, page);
953 RETURN(bytes > 0 ? bytes : rc);
957 * From kernel v4.19-rc5-248-g9b89a0355144 use XArrary
958 * Prior kernels use radix_tree for tags
960 static inline void ll_page_tag_dirty(struct page *page,
961 struct address_space *mapping)
963 #ifndef HAVE_RADIX_TREE_TAG_SET
964 __xa_set_mark(&mapping->i_pages, page_index(page), PAGECACHE_TAG_DIRTY);
966 radix_tree_tag_set(&mapping->page_tree, page_index(page),
967 PAGECACHE_TAG_DIRTY);
972 * Kernels 4.2 - 4.5 pass memcg argument to account_page_dirtied()
973 * Kernel v5.2-5678-gac1c3e4 no longer exports account_page_dirtied
975 static inline void ll_account_page_dirtied(struct page *page,
976 struct address_space *mapping)
978 #ifdef HAVE_ACCOUNT_PAGE_DIRTIED_3ARGS
979 struct mem_cgroup *memcg = mem_cgroup_begin_page_stat(page);
981 account_page_dirtied(page, mapping, memcg);
982 mem_cgroup_end_page_stat(memcg);
983 #elif defined(HAVE_ACCOUNT_PAGE_DIRTIED_EXPORT)
984 account_page_dirtied(page, mapping);
986 vvp_account_page_dirtied(page, mapping);
988 ll_page_tag_dirty(page, mapping);
991 /* Taken from kernel set_page_dirty, __set_page_dirty_nobuffers
992 * Last change to this area: b93b016313b3ba8003c3b8bb71f569af91f19fc7
994 * Current with Linus tip of tree (7/13/2019):
995 * v5.2-rc4-224-ge01e060fe0
997 * Backwards compat for 3.x, 5.x kernels relating to memcg handling
998 * & rename of radix tree to xarray.
1000 void vvp_set_pagevec_dirty(struct pagevec *pvec)
1002 struct page *page = pvec->pages[0];
1003 struct address_space *mapping = page->mapping;
1004 unsigned long flags;
1005 unsigned long skip_pages = 0;
1006 int count = pagevec_count(pvec);
1012 BUILD_BUG_ON(PAGEVEC_SIZE > BITS_PER_LONG);
1013 LASSERTF(page->mapping,
1014 "mapping must be set. page %p, page->private (cl_page) %p\n",
1015 page, (void *) page->private);
1017 for (i = 0; i < count; i++) {
1018 page = pvec->pages[i];
1020 ClearPageReclaim(page);
1022 lock_page_memcg(page);
1023 if (TestSetPageDirty(page)) {
1024 /* page is already dirty .. no extra work needed
1025 * set a flag for the i'th page to be skipped
1027 unlock_page_memcg(page);
1028 skip_pages |= (1 << i);
1032 ll_xa_lock_irqsave(&mapping->i_pages, flags);
1034 /* Notes on differences with __set_page_dirty_nobuffers:
1035 * 1. We don't need to call page_mapping because we know this is a page
1037 * 2. We have the pages locked, so there is no need for the careful
1038 * mapping/mapping2 dance.
1039 * 3. No mapping is impossible. (Race w/truncate mentioned in
1040 * dirty_nobuffers should be impossible because we hold the page lock.)
1041 * 4. All mappings are the same because i/o is only to one file.
1043 for (i = 0; i < count; i++) {
1044 page = pvec->pages[i];
1045 /* if the i'th page was unlocked above, skip it here */
1046 if ((skip_pages >> i) & 1)
1049 LASSERTF(page->mapping == mapping,
1050 "all pages must have the same mapping. page %p, mapping %p, first mapping %p\n",
1051 page, page->mapping, mapping);
1052 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1053 ll_account_page_dirtied(page, mapping);
1055 unlock_page_memcg(page);
1057 ll_xa_unlock_irqrestore(&mapping->i_pages, flags);
1059 CDEBUG(D_VFSTRACE, "mapping %p, count %d, dirtied %d\n", mapping,
1062 if (mapping->host && dirtied) {
1063 /* !PageAnon && !swapper_space */
1064 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1070 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
1071 struct pagevec *pvec)
1078 count = pagevec_count(pvec);
1081 for (i = 0; i < count; i++) {
1082 struct page *vmpage = pvec->pages[i];
1083 SetPageUptodate(vmpage);
1086 vvp_set_pagevec_dirty(pvec);
1088 for (i = 0; i < count; i++) {
1089 struct page *vmpage = pvec->pages[i];
1090 struct cl_page *page = (struct cl_page *) vmpage->private;
1091 cl_page_disown(env, io, page);
1092 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
1093 cl_page_put(env, page);
1099 /* make sure the page list is contiguous */
1100 static bool page_list_sanity_check(struct cl_object *obj,
1101 struct cl_page_list *plist)
1103 struct cl_page *page;
1104 pgoff_t index = CL_PAGE_EOF;
1106 cl_page_list_for_each(page, plist) {
1107 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1109 if (index == CL_PAGE_EOF) {
1110 index = vvp_index(vpg);
1115 if (index == vvp_index(vpg))
1123 /* Return how many bytes have queued or written */
1124 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
1126 struct cl_object *obj = io->ci_obj;
1127 struct inode *inode = vvp_object_inode(obj);
1128 struct vvp_io *vio = vvp_env_io(env);
1129 struct cl_page_list *queue = &vio->u.readwrite.vui_queue;
1130 struct cl_page *page;
1133 unsigned int npages = vio->u.readwrite.vui_queue.pl_nr;
1139 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
1140 npages, vio->u.readwrite.vui_from, vio->u.readwrite.vui_to);
1142 LASSERT(page_list_sanity_check(obj, queue));
1144 /* submit IO with async write */
1145 rc = cl_io_commit_async(env, io, queue,
1146 vio->u.readwrite.vui_from,
1147 vio->u.readwrite.vui_to,
1148 write_commit_callback);
1149 npages -= queue->pl_nr; /* already committed pages */
1151 /* calculate how many bytes were written */
1152 bytes = npages << PAGE_SHIFT;
1155 bytes -= vio->u.readwrite.vui_from;
1156 if (queue->pl_nr == 0) /* last page */
1157 bytes -= PAGE_SIZE - vio->u.readwrite.vui_to;
1158 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
1160 vio->u.readwrite.vui_written += bytes;
1162 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
1163 npages, bytes, vio->u.readwrite.vui_written);
1165 /* the first page must have been written. */
1166 vio->u.readwrite.vui_from = 0;
1168 LASSERT(page_list_sanity_check(obj, queue));
1169 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
1171 /* out of quota, try sync write */
1172 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
1173 rc = vvp_io_commit_sync(env, io, queue,
1174 vio->u.readwrite.vui_from,
1175 vio->u.readwrite.vui_to);
1177 vio->u.readwrite.vui_written += rc;
1182 /* update inode size */
1183 ll_merge_attr(env, inode);
1185 /* Now the pages in queue were failed to commit, discard them
1186 * unless they were dirtied before. */
1187 while (queue->pl_nr > 0) {
1188 page = cl_page_list_first(queue);
1189 cl_page_list_del(env, queue, page);
1191 if (!PageDirty(cl_page_vmpage(page)))
1192 cl_page_discard(env, io, page);
1194 cl_page_disown(env, io, page);
1196 /* held in ll_cl_init() */
1197 lu_ref_del(&page->cp_reference, "cl_io", io);
1198 cl_page_put(env, page);
1200 cl_page_list_fini(env, queue);
1205 static int vvp_io_write_start(const struct lu_env *env,
1206 const struct cl_io_slice *ios)
1208 struct vvp_io *vio = cl2vvp_io(env, ios);
1209 struct cl_io *io = ios->cis_io;
1210 struct cl_object *obj = io->ci_obj;
1211 struct inode *inode = vvp_object_inode(obj);
1212 struct ll_inode_info *lli = ll_i2info(inode);
1213 struct file *file = vio->vui_fd->fd_file;
1215 loff_t pos = io->u.ci_wr.wr.crw_pos;
1216 size_t cnt = io->u.ci_wr.wr.crw_count;
1217 bool lock_inode = !IS_NOSEC(inode);
1218 size_t nob = io->ci_nob;
1219 struct iov_iter iter;
1224 trunc_sem_down_read(&lli->lli_trunc_sem);
1226 if (!can_populate_pages(env, io, inode))
1229 if (cl_io_is_append(io)) {
1231 * PARALLEL IO This has to be changed for parallel IO doing
1232 * out-of-order writes.
1234 ll_merge_attr(env, inode);
1235 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
1236 vio->vui_iocb->ki_pos = pos;
1238 LASSERTF(vio->vui_iocb->ki_pos == pos,
1239 "ki_pos %lld [%lld, %lld)\n",
1240 vio->vui_iocb->ki_pos,
1244 CDEBUG(D_VFSTRACE, "%s: write [%llu, %llu)\n",
1245 file_dentry(file)->d_name.name,
1248 /* The maximum Lustre file size is variable, based on the OST maximum
1249 * object size and number of stripes. This needs another check in
1250 * addition to the VFS checks earlier. */
1251 if (pos + cnt > ll_file_maxbytes(inode)) {
1253 "%s: file %s ("DFID") offset %llu > maxbytes %llu\n",
1254 ll_i2sbi(inode)->ll_fsname,
1255 file_dentry(file)->d_name.name,
1256 PFID(ll_inode2fid(inode)), pos + cnt,
1257 ll_file_maxbytes(inode));
1261 /* Tests to verify we take the i_mutex correctly */
1262 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_SEC) && !lock_inode)
1265 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_NOSEC) && lock_inode)
1268 if (!(file->f_flags & O_DIRECT)) {
1269 result = cl_io_lru_reserve(env, io, pos, cnt);
1274 if (vio->vui_iter == NULL) {
1275 /* from a temp io in ll_cl_init(). */
1279 * When using the locked AIO function (generic_file_aio_write())
1280 * testing has shown the inode mutex to be a limiting factor
1281 * with multi-threaded single shared file performance. To get
1282 * around this, we now use the lockless version. To maintain
1283 * consistency, proper locking to protect against writes,
1284 * trucates, etc. is handled in the higher layers of lustre.
1286 lock_inode = !IS_NOSEC(inode);
1287 iter = *vio->vui_iter;
1289 if (unlikely(lock_inode))
1291 result = __generic_file_write_iter(vio->vui_iocb, &iter);
1292 if (unlikely(lock_inode))
1293 inode_unlock(inode);
1297 #ifdef HAVE_GENERIC_WRITE_SYNC_2ARGS
1298 result = generic_write_sync(vio->vui_iocb, result);
1303 err = generic_write_sync(vio->vui_iocb->ki_filp, pos,
1305 if (err < 0 && result > 0)
1312 result = vvp_io_write_commit(env, io);
1313 /* Simulate short commit */
1314 if (CFS_FAULT_CHECK(OBD_FAIL_LLITE_SHORT_COMMIT)) {
1315 vio->u.readwrite.vui_written >>= 1;
1316 if (vio->u.readwrite.vui_written > 0)
1317 io->ci_need_restart = 1;
1319 if (vio->u.readwrite.vui_written > 0) {
1320 result = vio->u.readwrite.vui_written;
1321 CDEBUG(D_VFSTRACE, "%s: write nob %zd, result: %zd\n",
1322 file_dentry(file)->d_name.name,
1323 io->ci_nob, result);
1324 io->ci_nob += result;
1326 io->ci_continue = 0;
1329 if (vio->vui_iocb->ki_pos != (pos + io->ci_nob - nob)) {
1331 "%s: write position mismatch: ki_pos %lld vs. pos %lld, written %zd, commit %zd: rc = %zd\n",
1332 file_dentry(file)->d_name.name,
1333 vio->vui_iocb->ki_pos, pos + io->ci_nob - nob,
1334 written, io->ci_nob - nob, result);
1336 * Rewind ki_pos and vui_iter to where it has
1337 * successfully committed.
1339 vio->vui_iocb->ki_pos = pos + io->ci_nob - nob;
1341 if (result > 0 || result == -EIOCBQUEUED) {
1342 set_bit(LLIF_DATA_MODIFIED, &ll_i2info(inode)->lli_flags);
1344 if (result != -EIOCBQUEUED && result < cnt)
1345 io->ci_continue = 0;
1349 if (result == -EIOCBQUEUED) {
1350 io->ci_nob += vio->u.readwrite.vui_written;
1351 vio->vui_iocb->ki_pos = pos +
1352 vio->u.readwrite.vui_written;
1359 static void vvp_io_rw_end(const struct lu_env *env,
1360 const struct cl_io_slice *ios)
1362 struct inode *inode = vvp_object_inode(ios->cis_obj);
1363 struct ll_inode_info *lli = ll_i2info(inode);
1365 trunc_sem_up_read(&lli->lli_trunc_sem);
1368 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
1370 struct vm_fault *vmf = cfio->ft_vmf;
1372 cfio->ft_flags = ll_filemap_fault(cfio->ft_vma, vmf);
1373 cfio->ft_flags_valid = 1;
1376 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
1377 get_vmf_address(vmf));
1378 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
1379 lock_page(vmf->page);
1380 cfio->ft_flags |= VM_FAULT_LOCKED;
1383 cfio->ft_vmpage = vmf->page;
1388 if (cfio->ft_flags & VM_FAULT_SIGBUS) {
1389 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", get_vmf_address(vmf));
1393 if (cfio->ft_flags & VM_FAULT_OOM) {
1394 CDEBUG(D_PAGE, "got addr %p - OOM\n", get_vmf_address(vmf));
1398 if (cfio->ft_flags & VM_FAULT_RETRY)
1401 CERROR("unknown error in page fault %d\n", cfio->ft_flags);
1406 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1407 struct pagevec *pvec)
1409 vvp_set_pagevec_dirty(pvec);
1412 static int vvp_io_fault_start(const struct lu_env *env,
1413 const struct cl_io_slice *ios)
1415 struct vvp_io *vio = cl2vvp_io(env, ios);
1416 struct cl_io *io = ios->cis_io;
1417 struct cl_object *obj = io->ci_obj;
1418 struct inode *inode = vvp_object_inode(obj);
1419 struct ll_inode_info *lli = ll_i2info(inode);
1420 struct cl_fault_io *fio = &io->u.ci_fault;
1421 struct vvp_fault_io *cfio = &vio->u.fault;
1424 struct page *vmpage = NULL;
1425 struct cl_page *page;
1430 trunc_sem_down_read_nowait(&lli->lli_trunc_sem);
1432 /* offset of the last byte on the page */
1433 offset = cl_offset(obj, fio->ft_index + 1) - 1;
1434 LASSERT(cl_index(obj, offset) == fio->ft_index);
1435 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1439 /* must return locked page */
1440 if (fio->ft_mkwrite) {
1441 LASSERT(cfio->ft_vmpage != NULL);
1442 lock_page(cfio->ft_vmpage);
1444 result = vvp_io_kernel_fault(cfio);
1449 vmpage = cfio->ft_vmpage;
1450 LASSERT(PageLocked(vmpage));
1452 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1453 generic_error_remove_page(vmpage->mapping, vmpage);
1455 size = i_size_read(inode);
1456 /* Though we have already held a cl_lock upon this page, but
1457 * it still can be truncated locally. */
1458 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1459 (page_offset(vmpage) > size))) {
1460 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1462 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1464 GOTO(out, result = +1);
1467 last_index = cl_index(obj, size - 1);
1469 if (fio->ft_mkwrite ) {
1471 * Capture the size while holding the lli_trunc_sem from above
1472 * we want to make sure that we complete the mkwrite action
1473 * while holding this lock. We need to make sure that we are
1474 * not past the end of the file.
1476 if (last_index < fio->ft_index) {
1478 "llite: mkwrite and truncate race happened: "
1479 "%p: 0x%lx 0x%lx\n",
1480 vmpage->mapping,fio->ft_index,last_index);
1482 * We need to return if we are
1483 * passed the end of the file. This will propagate
1484 * up the call stack to ll_page_mkwrite where
1485 * we will return VM_FAULT_NOPAGE. Any non-negative
1486 * value returned here will be silently
1487 * converted to 0. If the vmpage->mapping is null
1488 * the error code would be converted back to ENODATA
1489 * in ll_page_mkwrite0. Thus we return -ENODATA
1490 * to handle both cases
1492 GOTO(out, result = -ENODATA);
1496 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1498 GOTO(out, result = PTR_ERR(page));
1500 /* if page is going to be written, we should add this page into cache
1502 if (fio->ft_mkwrite) {
1503 wait_on_page_writeback(vmpage);
1504 if (!PageDirty(vmpage)) {
1505 struct cl_page_list *plist = &vio->u.fault.ft_queue;
1506 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1509 /* vvp_page_assume() calls wait_on_page_writeback(). */
1510 cl_page_assume(env, io, page);
1512 cl_page_list_init(plist);
1513 cl_page_list_add(plist, page);
1516 if (last_index == vvp_index(vpg))
1517 to = ((size - 1) & ~PAGE_MASK) + 1;
1519 /* Do not set Dirty bit here so that in case IO is
1520 * started before the page is really made dirty, we
1521 * still have chance to detect it. */
1522 result = cl_io_commit_async(env, io, plist, 0, to,
1523 mkwrite_commit_callback);
1524 /* Have overquota flag, trying sync write to check
1525 * whether indeed out of quota */
1526 if (result == -EDQUOT) {
1528 result = vvp_io_commit_sync(env, io,
1532 cl_page_own(env, io, page);
1533 cl_page_list_add(plist, page);
1534 lu_ref_add(&page->cp_reference,
1536 result = cl_io_commit_async(env, io,
1538 mkwrite_commit_callback);
1541 cl_page_put(env, page);
1545 LASSERT(cl_page_is_owned(page, io));
1546 cl_page_list_fini(env, plist);
1550 cl_page_discard(env, io, page);
1551 cl_page_disown(env, io, page);
1553 cl_page_put(env, page);
1555 /* we're in big trouble, what can we do now? */
1556 if (result == -EDQUOT)
1560 cl_page_disown(env, io, page);
1566 * The ft_index is only used in the case of
1567 * a mkwrite action. We need to check
1568 * our assertions are correct, since
1569 * we should have caught this above
1571 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1572 if (fio->ft_index == last_index)
1574 * Last page is mapped partially.
1576 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1578 fio->ft_nob = cl_page_size(obj);
1580 lu_ref_add(&page->cp_reference, "fault", io);
1581 fio->ft_page = page;
1585 /* return unlocked vmpage to avoid deadlocking */
1587 unlock_page(vmpage);
1589 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1594 static void vvp_io_fault_end(const struct lu_env *env,
1595 const struct cl_io_slice *ios)
1597 struct inode *inode = vvp_object_inode(ios->cis_obj);
1598 struct ll_inode_info *lli = ll_i2info(inode);
1600 CLOBINVRNT(env, ios->cis_io->ci_obj,
1601 vvp_object_invariant(ios->cis_io->ci_obj));
1602 trunc_sem_up_read(&lli->lli_trunc_sem);
1605 static int vvp_io_fsync_start(const struct lu_env *env,
1606 const struct cl_io_slice *ios)
1608 /* we should mark TOWRITE bit to each dirty page in radix tree to
1609 * verify pages have been written, but this is difficult because of
1614 static int vvp_io_read_ahead(const struct lu_env *env,
1615 const struct cl_io_slice *ios,
1616 pgoff_t start, struct cl_read_ahead *ra)
1621 if (ios->cis_io->ci_type == CIT_READ ||
1622 ios->cis_io->ci_type == CIT_FAULT) {
1623 struct vvp_io *vio = cl2vvp_io(env, ios);
1625 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1626 ra->cra_end_idx = CL_PAGE_EOF;
1627 result = +1; /* no need to call down */
1634 static int vvp_io_lseek_lock(const struct lu_env *env,
1635 const struct cl_io_slice *ios)
1637 struct cl_io *io = ios->cis_io;
1638 __u64 lock_start = io->u.ci_lseek.ls_start;
1639 __u64 lock_end = OBD_OBJECT_EOF;
1640 __u32 enqflags = CEF_MUST; /* always take client lock */
1642 return vvp_io_one_lock(env, io, enqflags, CLM_READ,
1643 lock_start, lock_end);
1646 static int vvp_io_lseek_start(const struct lu_env *env,
1647 const struct cl_io_slice *ios)
1649 struct cl_io *io = ios->cis_io;
1650 struct inode *inode = vvp_object_inode(io->ci_obj);
1651 __u64 start = io->u.ci_lseek.ls_start;
1654 inode_dio_wait(inode);
1656 /* At the moment we have DLM lock so just update inode
1657 * to know the file size.
1659 ll_merge_attr(env, inode);
1660 if (start >= i_size_read(inode)) {
1661 io->u.ci_lseek.ls_result = -ENXIO;
1667 static void vvp_io_lseek_end(const struct lu_env *env,
1668 const struct cl_io_slice *ios)
1670 struct cl_io *io = ios->cis_io;
1671 struct inode *inode = vvp_object_inode(io->ci_obj);
1673 if (io->u.ci_lseek.ls_result > i_size_read(inode))
1674 io->u.ci_lseek.ls_result = -ENXIO;
1676 inode_unlock(inode);
1679 static const struct cl_io_operations vvp_io_ops = {
1682 .cio_fini = vvp_io_fini,
1683 .cio_iter_init = vvp_io_read_iter_init,
1684 .cio_lock = vvp_io_read_lock,
1685 .cio_start = vvp_io_read_start,
1686 .cio_end = vvp_io_rw_end,
1687 .cio_advance = vvp_io_advance,
1690 .cio_fini = vvp_io_fini,
1691 .cio_iter_init = vvp_io_write_iter_init,
1692 .cio_iter_fini = vvp_io_write_iter_fini,
1693 .cio_lock = vvp_io_write_lock,
1694 .cio_start = vvp_io_write_start,
1695 .cio_end = vvp_io_rw_end,
1696 .cio_advance = vvp_io_advance,
1699 .cio_fini = vvp_io_setattr_fini,
1700 .cio_iter_init = vvp_io_setattr_iter_init,
1701 .cio_lock = vvp_io_setattr_lock,
1702 .cio_start = vvp_io_setattr_start,
1703 .cio_end = vvp_io_setattr_end
1706 .cio_fini = vvp_io_fault_fini,
1707 .cio_iter_init = vvp_io_fault_iter_init,
1708 .cio_lock = vvp_io_fault_lock,
1709 .cio_start = vvp_io_fault_start,
1710 .cio_end = vvp_io_fault_end,
1713 .cio_start = vvp_io_fsync_start,
1714 .cio_fini = vvp_io_fini
1717 .cio_fini = vvp_io_fini
1720 .cio_fini = vvp_io_fini
1723 .cio_fini = vvp_io_fini
1726 .cio_fini = vvp_io_fini,
1727 .cio_lock = vvp_io_lseek_lock,
1728 .cio_start = vvp_io_lseek_start,
1729 .cio_end = vvp_io_lseek_end,
1732 .cio_read_ahead = vvp_io_read_ahead
1735 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1738 struct vvp_io *vio = vvp_env_io(env);
1739 struct inode *inode = vvp_object_inode(obj);
1742 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1745 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1746 "restore needed %d\n",
1747 PFID(lu_object_fid(&obj->co_lu)),
1748 io->ci_ignore_layout, io->ci_verify_layout,
1749 vio->vui_layout_gen, io->ci_restore_needed);
1751 CL_IO_SLICE_CLEAN(vio, vui_cl);
1752 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1753 vio->vui_ra_valid = false;
1755 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1757 struct ll_inode_info *lli = ll_i2info(inode);
1759 count = io->u.ci_rw.crw_count;
1760 /* "If nbyte is 0, read() will return 0 and have no other
1761 * results." -- Single Unix Spec */
1765 vio->vui_tot_count = count;
1767 /* for read/write, we store the jobid in the inode, and
1768 * it'll be fetched by osc when building RPC.
1770 * it's not accurate if the file is shared by different
1773 lustre_get_jobid(lli->lli_jobid, sizeof(lli->lli_jobid));
1774 } else if (io->ci_type == CIT_SETATTR) {
1775 if (!cl_io_is_trunc(io))
1776 io->ci_lockreq = CILR_MANDATORY;
1779 /* Enqueue layout lock and get layout version. We need to do this
1780 * even for operations requiring to open file, such as read and write,
1781 * because it might not grant layout lock in IT_OPEN. */
1782 if (result == 0 && !io->ci_ignore_layout) {
1783 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1784 if (result == -ENOENT)
1785 /* If the inode on MDS has been removed, but the objects
1786 * on OSTs haven't been destroyed (async unlink), layout
1787 * fetch will return -ENOENT, we'd ingore this error
1788 * and continue with dirty flush. LU-3230. */
1791 CERROR("%s: refresh file layout " DFID " error %d.\n",
1792 ll_i2sbi(inode)->ll_fsname,
1793 PFID(lu_object_fid(&obj->co_lu)), result);
1796 io->ci_result = result < 0 ? result : 0;