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/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * Implementation of cl_io for VVP layer.
34 * Author: Nikita Danilov <nikita.danilov@sun.com>
35 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
38 #define DEBUG_SUBSYSTEM S_LLITE
42 #include <linux/pagevec.h>
43 #include <linux/memcontrol.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 * True, if \a io is a normal io, False for splice_{read,write}
62 static int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
64 struct vvp_io *vio = vvp_env_io(env);
66 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
68 return vio->vui_io_subtype == IO_NORMAL;
72 * For swapping layout. The file's layout may have changed.
73 * To avoid populating pages to a wrong stripe, we have to verify the
74 * correctness of layout. It works because swapping layout processes
75 * have to acquire group lock.
77 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
80 struct ll_inode_info *lli = ll_i2info(inode);
81 struct vvp_io *vio = vvp_env_io(env);
84 switch (io->ci_type) {
87 /* don't need lock here to check lli_layout_gen as we have held
88 * extent lock and GROUP lock has to hold to swap layout */
89 if (ll_layout_version_get(lli) != vio->vui_layout_gen ||
90 OBD_FAIL_CHECK_RESET(OBD_FAIL_LLITE_LOST_LAYOUT, 0)) {
91 io->ci_need_restart = 1;
92 /* this will cause a short read/write */
97 /* fault is okay because we've already had a page. */
105 static void vvp_object_size_lock(struct cl_object *obj)
107 struct inode *inode = vvp_object_inode(obj);
109 ll_inode_size_lock(inode);
110 cl_object_attr_lock(obj);
113 static void vvp_object_size_unlock(struct cl_object *obj)
115 struct inode *inode = vvp_object_inode(obj);
117 cl_object_attr_unlock(obj);
118 ll_inode_size_unlock(inode);
122 * Helper function that if necessary adjusts file size (inode->i_size), when
123 * position at the offset \a pos is accessed. File size can be arbitrary stale
124 * on a Lustre client, but client at least knows KMS. If accessed area is
125 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
127 * Locking: i_size_lock is used to serialize changes to inode size and to
128 * protect consistency between inode size and cl_object
129 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
130 * top-object and sub-objects.
132 static int vvp_prep_size(const struct lu_env *env, struct cl_object *obj,
133 struct cl_io *io, loff_t start, size_t count,
136 struct cl_attr *attr = vvp_env_thread_attr(env);
137 struct inode *inode = vvp_object_inode(obj);
138 loff_t pos = start + count - 1;
143 * Consistency guarantees: following possibilities exist for the
144 * relation between region being accessed and real file size at this
147 * (A): the region is completely inside of the file;
149 * (B-x): x bytes of region are inside of the file, the rest is
152 * (C): the region is completely outside of the file.
154 * This classification is stable under DLM lock already acquired by
155 * the caller, because to change the class, other client has to take
156 * DLM lock conflicting with our lock. Also, any updates to ->i_size
157 * by other threads on this client are serialized by
158 * ll_inode_size_lock(). This guarantees that short reads are handled
159 * correctly in the face of concurrent writes and truncates.
161 vvp_object_size_lock(obj);
162 result = cl_object_attr_get(env, obj, attr);
167 * A glimpse is necessary to determine whether we
168 * return a short read (B) or some zeroes at the end
171 vvp_object_size_unlock(obj);
172 result = cl_glimpse_lock(env, io, inode, obj, 0);
173 if (result == 0 && exceed != NULL) {
174 /* If objective page index exceed end-of-file
175 * page index, return directly. Do not expect
176 * kernel will check such case correctly.
177 * linux-2.6.18-128.1.1 miss to do that.
179 loff_t size = i_size_read(inode);
180 unsigned long cur_index = start >>
183 if ((size == 0 && cur_index != 0) ||
184 (((size - 1) >> PAGE_SHIFT) <
192 * region is within kms and, hence, within real file
193 * size (A). We need to increase i_size to cover the
194 * read region so that generic_file_read() will do its
195 * job, but that doesn't mean the kms size is
196 * _correct_, it is only the _minimum_ size. If
197 * someone does a stat they will get the correct size
198 * which will always be >= the kms value here.
201 if (i_size_read(inode) < kms) {
202 i_size_write(inode, kms);
204 DFID" updating i_size %llu\n",
205 PFID(lu_object_fid(&obj->co_lu)),
206 (__u64)i_size_read(inode));
211 vvp_object_size_unlock(obj);
216 /*****************************************************************************
222 static int vvp_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
223 __u32 enqflags, enum cl_lock_mode mode,
224 pgoff_t start, pgoff_t end)
226 struct vvp_io *vio = vvp_env_io(env);
227 struct cl_lock_descr *descr = &vio->vui_link.cill_descr;
228 struct cl_object *obj = io->ci_obj;
230 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
233 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
235 memset(&vio->vui_link, 0, sizeof vio->vui_link);
237 if (vio->vui_fd && (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
238 descr->cld_mode = CLM_GROUP;
239 descr->cld_gid = vio->vui_fd->fd_grouplock.lg_gid;
240 enqflags |= CEF_LOCK_MATCH;
242 descr->cld_mode = mode;
245 descr->cld_obj = obj;
246 descr->cld_start = start;
247 descr->cld_end = end;
248 descr->cld_enq_flags = enqflags;
250 cl_io_lock_add(env, io, &vio->vui_link);
255 static int vvp_io_one_lock(const struct lu_env *env, struct cl_io *io,
256 __u32 enqflags, enum cl_lock_mode mode,
257 loff_t start, loff_t end)
259 struct cl_object *obj = io->ci_obj;
261 return vvp_io_one_lock_index(env, io, enqflags, mode,
262 cl_index(obj, start), cl_index(obj, end));
265 static int vvp_io_write_iter_init(const struct lu_env *env,
266 const struct cl_io_slice *ios)
268 struct vvp_io *vio = cl2vvp_io(env, ios);
270 cl_page_list_init(&vio->u.write.vui_queue);
271 vio->u.write.vui_written = 0;
272 vio->u.write.vui_from = 0;
273 vio->u.write.vui_to = PAGE_SIZE;
278 static void vvp_io_write_iter_fini(const struct lu_env *env,
279 const struct cl_io_slice *ios)
281 struct vvp_io *vio = cl2vvp_io(env, ios);
283 LASSERT(vio->u.write.vui_queue.pl_nr == 0);
286 static int vvp_io_fault_iter_init(const struct lu_env *env,
287 const struct cl_io_slice *ios)
289 struct vvp_io *vio = cl2vvp_io(env, ios);
290 struct inode *inode = vvp_object_inode(ios->cis_obj);
292 LASSERT(inode == file_inode(vio->vui_fd->fd_file));
297 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
299 struct cl_io *io = ios->cis_io;
300 struct cl_object *obj = io->ci_obj;
301 struct vvp_io *vio = cl2vvp_io(env, ios);
302 struct inode *inode = vvp_object_inode(obj);
307 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
309 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
310 "need write layout %d, restore needed %d\n",
311 PFID(lu_object_fid(&obj->co_lu)),
312 io->ci_ignore_layout, io->ci_verify_layout,
313 vio->vui_layout_gen, io->ci_need_write_intent,
314 io->ci_restore_needed);
316 if (io->ci_restore_needed) {
317 /* file was detected release, we need to restore it
318 * before finishing the io
320 rc = ll_layout_restore(inode, 0, OBD_OBJECT_EOF);
321 /* if restore registration failed, no restart,
322 * we will return -ENODATA */
323 /* The layout will change after restore, so we need to
324 * block on layout lock held by the MDT
325 * as MDT will not send new layout in lvb (see LU-3124)
326 * we have to explicitly fetch it, all this will be done
327 * by ll_layout_refresh().
328 * Even if ll_layout_restore() returns zero, it doesn't mean
329 * that restore has been successful. Therefore it sets
330 * ci_verify_layout so that it will check layout at the end
334 io->ci_restore_needed = 1;
335 io->ci_need_restart = 0;
336 io->ci_verify_layout = 0;
341 io->ci_restore_needed = 0;
343 /* Even if ll_layout_restore() returns zero, it doesn't mean
344 * that restore has been successful. Therefore it should verify
345 * if there was layout change and restart I/O correspondingly.
347 ll_layout_refresh(inode, &gen);
348 io->ci_need_restart = vio->vui_layout_gen != gen;
349 if (io->ci_need_restart) {
351 DFID" layout changed from %d to %d.\n",
352 PFID(lu_object_fid(&obj->co_lu)),
353 vio->vui_layout_gen, gen);
354 /* today successful restore is the only possible
356 /* restore was done, clear restoring state */
357 ll_file_clear_flag(ll_i2info(vvp_object_inode(obj)),
358 LLIF_FILE_RESTORING);
364 * dynamic layout change needed, send layout intent
367 if (io->ci_need_write_intent) {
368 enum layout_intent_opc opc = LAYOUT_INTENT_WRITE;
370 io->ci_need_write_intent = 0;
372 LASSERT(io->ci_type == CIT_WRITE ||
373 cl_io_is_trunc(io) || cl_io_is_mkwrite(io));
375 CDEBUG(D_VFSTRACE, DFID" write layout, type %u "DEXT"\n",
376 PFID(lu_object_fid(&obj->co_lu)), io->ci_type,
377 PEXT(&io->ci_write_intent));
379 if (cl_io_is_trunc(io))
380 opc = LAYOUT_INTENT_TRUNC;
382 rc = ll_layout_write_intent(inode, opc, &io->ci_write_intent);
385 io->ci_need_restart = 1;
389 if (!io->ci_need_restart &&
390 !io->ci_ignore_layout && io->ci_verify_layout) {
391 /* check layout version */
392 ll_layout_refresh(inode, &gen);
393 io->ci_need_restart = vio->vui_layout_gen != gen;
394 if (io->ci_need_restart) {
396 DFID" layout changed from %d to %d.\n",
397 PFID(lu_object_fid(&obj->co_lu)),
398 vio->vui_layout_gen, gen);
406 static void vvp_io_fault_fini(const struct lu_env *env,
407 const struct cl_io_slice *ios)
409 struct cl_io *io = ios->cis_io;
410 struct cl_page *page = io->u.ci_fault.ft_page;
412 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
415 lu_ref_del(&page->cp_reference, "fault", io);
416 cl_page_put(env, page);
417 io->u.ci_fault.ft_page = NULL;
419 vvp_io_fini(env, ios);
422 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
425 * we only want to hold PW locks if the mmap() can generate
426 * writes back to the file and that only happens in shared
429 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
434 static int vvp_mmap_locks(const struct lu_env *env,
435 struct vvp_io *vio, struct cl_io *io)
437 struct vvp_thread_info *vti = vvp_env_info(env);
438 struct mm_struct *mm = current->mm;
439 struct vm_area_struct *vma;
440 struct cl_lock_descr *descr = &vti->vti_descr;
441 union ldlm_policy_data policy;
449 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
451 if (!cl_is_normalio(env, io))
454 /* nfs or loop back device write */
455 if (vio->vui_iter == NULL)
458 /* No MM (e.g. NFS)? No vmas too. */
462 if (!iter_is_iovec(vio->vui_iter) && !iov_iter_is_kvec(vio->vui_iter))
465 for (i = *vio->vui_iter;
467 iov_iter_advance(&i, iov.iov_len)) {
468 iov = iov_iter_iovec(&i);
469 addr = (unsigned long)iov.iov_base;
475 count += addr & ~PAGE_MASK;
478 down_read(&mm->mmap_sem);
479 while((vma = our_vma(mm, addr, count)) != NULL) {
480 struct dentry *de = file_dentry(vma->vm_file);
481 struct inode *inode = de->d_inode;
482 int flags = CEF_MUST;
484 if (ll_file_nolock(vma->vm_file)) {
486 * For no lock case is not allowed for mmap
493 * XXX: Required lock mode can be weakened: CIT_WRITE
494 * io only ever reads user level buffer, and CIT_READ
497 policy_from_vma(&policy, vma, addr, count);
498 descr->cld_mode = vvp_mode_from_vma(vma);
499 descr->cld_obj = ll_i2info(inode)->lli_clob;
500 descr->cld_start = cl_index(descr->cld_obj,
501 policy.l_extent.start);
502 descr->cld_end = cl_index(descr->cld_obj,
503 policy.l_extent.end);
504 descr->cld_enq_flags = flags;
505 result = cl_io_lock_alloc_add(env, io, descr);
507 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
508 descr->cld_mode, descr->cld_start,
514 if (vma->vm_end - addr >= count)
517 count -= vma->vm_end - addr;
520 up_read(&mm->mmap_sem);
527 static void vvp_io_advance(const struct lu_env *env,
528 const struct cl_io_slice *ios,
531 struct vvp_io *vio = cl2vvp_io(env, ios);
532 struct cl_io *io = ios->cis_io;
533 struct cl_object *obj = ios->cis_io->ci_obj;
535 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
537 if (!cl_is_normalio(env, io))
540 vio->vui_tot_count -= nob;
541 iov_iter_reexpand(vio->vui_iter, vio->vui_tot_count);
544 static void vvp_io_update_iov(const struct lu_env *env,
545 struct vvp_io *vio, struct cl_io *io)
547 size_t size = io->u.ci_rw.crw_count;
549 if (!cl_is_normalio(env, io) || vio->vui_iter == NULL)
552 iov_iter_truncate(vio->vui_iter, size);
555 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
556 enum cl_lock_mode mode, loff_t start, loff_t end)
558 struct vvp_io *vio = vvp_env_io(env);
562 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
565 vvp_io_update_iov(env, vio, io);
567 if (io->u.ci_rw.crw_nonblock)
568 ast_flags |= CEF_NONBLOCK;
569 if (io->ci_lock_no_expand)
570 ast_flags |= CEF_LOCK_NO_EXPAND;
572 /* Group lock held means no lockless any more */
573 if (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)
574 io->ci_ignore_lockless = 1;
576 if (ll_file_nolock(vio->vui_fd->fd_file) ||
577 (vio->vui_fd->fd_flags & LL_FILE_LOCKLESS_IO &&
578 !io->ci_ignore_lockless))
579 ast_flags |= CEF_NEVER;
582 result = vvp_mmap_locks(env, vio, io);
584 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
589 static int vvp_io_read_lock(const struct lu_env *env,
590 const struct cl_io_slice *ios)
592 struct cl_io *io = ios->cis_io;
593 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
597 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
598 rd->crw_pos + rd->crw_count - 1);
602 static int vvp_io_fault_lock(const struct lu_env *env,
603 const struct cl_io_slice *ios)
605 struct cl_io *io = ios->cis_io;
606 struct vvp_io *vio = cl2vvp_io(env, ios);
608 * XXX LDLM_FL_CBPENDING
610 return vvp_io_one_lock_index(env,
612 vvp_mode_from_vma(vio->u.fault.ft_vma),
613 io->u.ci_fault.ft_index,
614 io->u.ci_fault.ft_index);
617 static int vvp_io_write_lock(const struct lu_env *env,
618 const struct cl_io_slice *ios)
620 struct cl_io *io = ios->cis_io;
624 if (io->u.ci_wr.wr_append) {
626 end = OBD_OBJECT_EOF;
628 start = io->u.ci_wr.wr.crw_pos;
629 end = start + io->u.ci_wr.wr.crw_count - 1;
632 RETURN(vvp_io_rw_lock(env, io, CLM_WRITE, start, end));
635 static int vvp_io_setattr_iter_init(const struct lu_env *env,
636 const struct cl_io_slice *ios)
643 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
645 * Handles "lockless io" mode when extent locking is done by server.
647 static int vvp_io_setattr_lock(const struct lu_env *env,
648 const struct cl_io_slice *ios)
650 struct cl_io *io = ios->cis_io;
654 if (cl_io_is_trunc(io)) {
655 new_size = io->u.ci_setattr.sa_attr.lvb_size;
657 enqflags = CEF_DISCARD_DATA;
659 unsigned int valid = io->u.ci_setattr.sa_avalid;
661 if (!(valid & TIMES_SET_FLAGS))
664 if ((!(valid & ATTR_MTIME) ||
665 io->u.ci_setattr.sa_attr.lvb_mtime >=
666 io->u.ci_setattr.sa_attr.lvb_ctime) &&
667 (!(valid & ATTR_ATIME) ||
668 io->u.ci_setattr.sa_attr.lvb_atime >=
669 io->u.ci_setattr.sa_attr.lvb_ctime))
675 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
676 new_size, OBD_OBJECT_EOF);
679 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
684 * Only ll_inode_size_lock is taken at this level.
686 ll_inode_size_lock(inode);
687 result = inode_newsize_ok(inode, size);
689 ll_inode_size_unlock(inode);
692 i_size_write(inode, size);
694 ll_truncate_pagecache(inode, size);
695 ll_inode_size_unlock(inode);
699 static int vvp_io_setattr_time(const struct lu_env *env,
700 const struct cl_io_slice *ios)
702 struct cl_io *io = ios->cis_io;
703 struct cl_object *obj = io->ci_obj;
704 struct cl_attr *attr = vvp_env_thread_attr(env);
706 unsigned valid = CAT_CTIME;
708 cl_object_attr_lock(obj);
709 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
710 if (io->u.ci_setattr.sa_avalid & ATTR_ATIME_SET) {
711 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
714 if (io->u.ci_setattr.sa_avalid & ATTR_MTIME_SET) {
715 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
718 result = cl_object_attr_update(env, obj, attr, valid);
719 cl_object_attr_unlock(obj);
724 static int vvp_io_setattr_start(const struct lu_env *env,
725 const struct cl_io_slice *ios)
727 struct cl_io *io = ios->cis_io;
728 struct inode *inode = vvp_object_inode(io->ci_obj);
729 struct ll_inode_info *lli = ll_i2info(inode);
731 if (cl_io_is_trunc(io)) {
732 trunc_sem_down_write(&lli->lli_trunc_sem);
734 inode_dio_wait(inode);
739 if (io->u.ci_setattr.sa_avalid & TIMES_SET_FLAGS)
740 return vvp_io_setattr_time(env, ios);
745 static void vvp_io_setattr_end(const struct lu_env *env,
746 const struct cl_io_slice *ios)
748 struct cl_io *io = ios->cis_io;
749 struct inode *inode = vvp_object_inode(io->ci_obj);
750 struct ll_inode_info *lli = ll_i2info(inode);
752 if (cl_io_is_trunc(io)) {
753 /* Truncate in memory pages - they must be clean pages
754 * because osc has already notified to destroy osc_extents. */
755 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
757 trunc_sem_up_write(&lli->lli_trunc_sem);
763 static void vvp_io_setattr_fini(const struct lu_env *env,
764 const struct cl_io_slice *ios)
766 bool restore_needed = ios->cis_io->ci_restore_needed;
767 struct inode *inode = vvp_object_inode(ios->cis_obj);
769 vvp_io_fini(env, ios);
771 if (restore_needed && !ios->cis_io->ci_restore_needed) {
772 /* restore finished, set data modified flag for HSM */
773 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
777 static int vvp_io_read_start(const struct lu_env *env,
778 const struct cl_io_slice *ios)
780 struct vvp_io *vio = cl2vvp_io(env, ios);
781 struct cl_io *io = ios->cis_io;
782 struct cl_object *obj = io->ci_obj;
783 struct inode *inode = vvp_object_inode(obj);
784 struct ll_inode_info *lli = ll_i2info(inode);
785 struct file *file = vio->vui_fd->fd_file;
786 loff_t pos = io->u.ci_rd.rd.crw_pos;
787 size_t cnt = io->u.ci_rd.rd.crw_count;
788 size_t tot = vio->vui_tot_count;
793 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
795 CDEBUG(D_VFSTRACE, "%s: read [%llu, %llu)\n",
796 file_dentry(file)->d_name.name,
799 if (vio->vui_io_subtype == IO_NORMAL)
800 trunc_sem_down_read(&lli->lli_trunc_sem);
802 if (io->ci_async_readahead) {
807 if (!can_populate_pages(env, io, inode))
810 /* Unless this is reading a sparse file, otherwise the lock has already
811 * been acquired so vvp_prep_size() is an empty op. */
812 result = vvp_prep_size(env, obj, io, pos, cnt, &exceed);
815 else if (exceed != 0)
818 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
819 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
820 inode->i_ino, cnt, pos, i_size_read(inode));
822 /* turn off the kernel's read-ahead */
823 vio->vui_fd->fd_file->f_ra.ra_pages = 0;
825 /* initialize read-ahead window once per syscall */
826 if (!vio->vui_ra_valid) {
827 vio->vui_ra_valid = true;
828 vio->vui_ra_start_idx = cl_index(obj, pos);
829 vio->vui_ra_pages = cl_index(obj, tot + PAGE_SIZE - 1);
830 /* If both start and end are unaligned, we read one more page
831 * than the index math suggests. */
832 if ((pos & ~PAGE_MASK) != 0 && ((pos + tot) & ~PAGE_MASK) != 0)
835 CDEBUG(D_READA, "tot %zu, ra_start %lu, ra_count %lu\n",
836 tot, vio->vui_ra_start_idx, vio->vui_ra_pages);
841 switch (vio->vui_io_subtype) {
843 LASSERT(vio->vui_iocb->ki_pos == pos);
844 result = generic_file_read_iter(vio->vui_iocb, vio->vui_iter);
847 result = generic_file_splice_read(file, &pos,
848 vio->u.splice.vui_pipe, cnt,
849 vio->u.splice.vui_flags);
850 /* LU-1109: do splice read stripe by stripe otherwise if it
851 * may make nfsd stuck if this read occupied all internal pipe
856 CERROR("Wrong IO type %u\n", vio->vui_io_subtype);
865 io->ci_nob += result;
872 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
873 struct cl_page_list *plist, int from, int to)
875 struct cl_2queue *queue = &io->ci_queue;
876 struct cl_page *page;
877 unsigned int bytes = 0;
881 if (plist->pl_nr == 0)
884 if (from > 0 || to != PAGE_SIZE) {
885 page = cl_page_list_first(plist);
886 if (plist->pl_nr == 1) {
887 cl_page_clip(env, page, from, to);
890 cl_page_clip(env, page, from, PAGE_SIZE);
891 if (to != PAGE_SIZE) {
892 page = cl_page_list_last(plist);
893 cl_page_clip(env, page, 0, to);
898 cl_2queue_init(queue);
899 cl_page_list_splice(plist, &queue->c2_qin);
900 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
902 /* plist is not sorted any more */
903 cl_page_list_splice(&queue->c2_qin, plist);
904 cl_page_list_splice(&queue->c2_qout, plist);
905 cl_2queue_fini(env, queue);
908 /* calculate bytes */
909 bytes = plist->pl_nr << PAGE_SHIFT;
910 bytes -= from + PAGE_SIZE - to;
912 while (plist->pl_nr > 0) {
913 page = cl_page_list_first(plist);
914 cl_page_list_del(env, plist, page);
916 cl_page_clip(env, page, 0, PAGE_SIZE);
918 SetPageUptodate(cl_page_vmpage(page));
919 cl_page_disown(env, io, page);
921 /* held in ll_cl_init() */
922 lu_ref_del(&page->cp_reference, "cl_io", io);
923 cl_page_put(env, page);
927 RETURN(bytes > 0 ? bytes : rc);
931 * Kernels 4.2 - 4.5 pass memcg argument to account_page_dirtied()
932 * Kernel v5.2-5678-gac1c3e4 no longer exports account_page_dirtied
934 static inline void ll_account_page_dirtied(struct page *page,
935 struct address_space *mapping)
937 #ifdef HAVE_ACCOUNT_PAGE_DIRTIED_3ARGS
938 struct mem_cgroup *memcg = mem_cgroup_begin_page_stat(page);
940 account_page_dirtied(page, mapping, memcg);
941 mem_cgroup_end_page_stat(memcg);
942 #elif defined HAVE_ACCOUNT_PAGE_DIRTIED
943 account_page_dirtied(page, mapping, memcg);
945 typedef unsigned int (dirtied_t)(struct page *pg,
946 struct address_space *as);
947 const char *symbol = "account_page_dirtied";
948 static dirtied_t *dirtied = NULL;
951 dirtied = (dirtied_t *)symbol_get(symbol);
954 dirtied(page, mapping);
959 * From kernel v4.19-rc5-248-g9b89a0355144 use XArrary
960 * Prior kernels use radix_tree for tags
962 static inline void ll_page_tag_dirty(struct page *page,
963 struct address_space *mapping)
965 #ifdef HAVE___XA_SET_MARK
966 __xa_set_mark(&mapping->i_pages, page_index(page), PAGECACHE_TAG_DIRTY);
968 radix_tree_tag_set(&mapping->page_tree, page_index(page),
969 PAGECACHE_TAG_DIRTY);
973 /* Taken from kernel set_page_dirty, __set_page_dirty_nobuffers
974 * Last change to this area: b93b016313b3ba8003c3b8bb71f569af91f19fc7
976 * Current with Linus tip of tree (7/13/2019):
977 * v5.2-rc4-224-ge01e060fe0
979 * Backwards compat for 3.x, 5.x kernels relating to memcg handling
980 * & rename of radix tree to xarray.
982 void vvp_set_pagevec_dirty(struct pagevec *pvec)
984 struct page *page = pvec->pages[0];
985 struct address_space *mapping = page->mapping;
987 int count = pagevec_count(pvec);
993 /* From set_page_dirty */
994 for (i = 0; i < count; i++)
995 ClearPageReclaim(pvec->pages[i]);
997 LASSERTF(page->mapping,
998 "mapping must be set. page %p, page->private (cl_page) %p",
999 page, (void *) page->private);
1001 /* Rest of code derived from __set_page_dirty_nobuffers */
1002 xa_lock_irqsave(&mapping->i_pages, flags);
1004 /* Notes on differences with __set_page_dirty_nobuffers:
1005 * 1. We don't need to call page_mapping because we know this is a page
1007 * 2. We have the pages locked, so there is no need for the careful
1008 * mapping/mapping2 dance.
1009 * 3. No mapping is impossible. (Race w/truncate mentioned in
1010 * dirty_nobuffers should be impossible because we hold the page lock.)
1011 * 4. All mappings are the same because i/o is only to one file.
1012 * 5. We invert the lock order on lock_page_memcg(page) and the mapping
1013 * xa_lock, but this is the only function that should use that pair of
1014 * locks and it can't race because Lustre locks pages throughout i/o.
1016 for (i = 0; i < count; i++) {
1017 page = pvec->pages[i];
1018 lock_page_memcg(page);
1019 if (TestSetPageDirty(page)) {
1020 unlock_page_memcg(page);
1023 LASSERTF(page->mapping == mapping,
1024 "all pages must have the same mapping. page %p, mapping %p, first mapping %p\n",
1025 page, page->mapping, mapping);
1026 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1027 ll_account_page_dirtied(page, mapping);
1028 ll_page_tag_dirty(page, mapping);
1030 unlock_page_memcg(page);
1032 xa_unlock_irqrestore(&mapping->i_pages, flags);
1034 CDEBUG(D_VFSTRACE, "mapping %p, count %d, dirtied %d\n", mapping,
1037 if (mapping->host && dirtied) {
1038 /* !PageAnon && !swapper_space */
1039 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1045 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
1046 struct pagevec *pvec)
1053 count = pagevec_count(pvec);
1056 for (i = 0; i < count; i++) {
1057 struct page *vmpage = pvec->pages[i];
1058 SetPageUptodate(vmpage);
1061 vvp_set_pagevec_dirty(pvec);
1063 for (i = 0; i < count; i++) {
1064 struct page *vmpage = pvec->pages[i];
1065 struct cl_page *page = (struct cl_page *) vmpage->private;
1066 cl_page_disown(env, io, page);
1067 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
1068 cl_page_put(env, page);
1074 /* make sure the page list is contiguous */
1075 static bool page_list_sanity_check(struct cl_object *obj,
1076 struct cl_page_list *plist)
1078 struct cl_page *page;
1079 pgoff_t index = CL_PAGE_EOF;
1081 cl_page_list_for_each(page, plist) {
1082 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1084 if (index == CL_PAGE_EOF) {
1085 index = vvp_index(vpg);
1090 if (index == vvp_index(vpg))
1098 /* Return how many bytes have queued or written */
1099 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
1101 struct cl_object *obj = io->ci_obj;
1102 struct inode *inode = vvp_object_inode(obj);
1103 struct vvp_io *vio = vvp_env_io(env);
1104 struct cl_page_list *queue = &vio->u.write.vui_queue;
1105 struct cl_page *page;
1108 unsigned int npages = vio->u.write.vui_queue.pl_nr;
1114 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
1115 npages, vio->u.write.vui_from, vio->u.write.vui_to);
1117 LASSERT(page_list_sanity_check(obj, queue));
1119 /* submit IO with async write */
1120 rc = cl_io_commit_async(env, io, queue,
1121 vio->u.write.vui_from, vio->u.write.vui_to,
1122 write_commit_callback);
1123 npages -= queue->pl_nr; /* already committed pages */
1125 /* calculate how many bytes were written */
1126 bytes = npages << PAGE_SHIFT;
1129 bytes -= vio->u.write.vui_from;
1130 if (queue->pl_nr == 0) /* last page */
1131 bytes -= PAGE_SIZE - vio->u.write.vui_to;
1132 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
1134 vio->u.write.vui_written += bytes;
1136 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
1137 npages, bytes, vio->u.write.vui_written);
1139 /* the first page must have been written. */
1140 vio->u.write.vui_from = 0;
1142 LASSERT(page_list_sanity_check(obj, queue));
1143 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
1145 /* out of quota, try sync write */
1146 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
1147 rc = vvp_io_commit_sync(env, io, queue,
1148 vio->u.write.vui_from,
1149 vio->u.write.vui_to);
1151 vio->u.write.vui_written += rc;
1156 /* update inode size */
1157 ll_merge_attr(env, inode);
1159 /* Now the pages in queue were failed to commit, discard them
1160 * unless they were dirtied before. */
1161 while (queue->pl_nr > 0) {
1162 page = cl_page_list_first(queue);
1163 cl_page_list_del(env, queue, page);
1165 if (!PageDirty(cl_page_vmpage(page)))
1166 cl_page_discard(env, io, page);
1168 cl_page_disown(env, io, page);
1170 /* held in ll_cl_init() */
1171 lu_ref_del(&page->cp_reference, "cl_io", io);
1172 cl_page_put(env, page);
1174 cl_page_list_fini(env, queue);
1179 static int vvp_io_write_start(const struct lu_env *env,
1180 const struct cl_io_slice *ios)
1182 struct vvp_io *vio = cl2vvp_io(env, ios);
1183 struct cl_io *io = ios->cis_io;
1184 struct cl_object *obj = io->ci_obj;
1185 struct inode *inode = vvp_object_inode(obj);
1186 struct ll_inode_info *lli = ll_i2info(inode);
1187 struct file *file = vio->vui_fd->fd_file;
1189 loff_t pos = io->u.ci_wr.wr.crw_pos;
1190 size_t cnt = io->u.ci_wr.wr.crw_count;
1191 bool lock_inode = !IS_NOSEC(inode);
1192 size_t nob = io->ci_nob;
1197 if (vio->vui_io_subtype == IO_NORMAL)
1198 trunc_sem_down_read(&lli->lli_trunc_sem);
1200 if (!can_populate_pages(env, io, inode))
1203 if (cl_io_is_append(io)) {
1205 * PARALLEL IO This has to be changed for parallel IO doing
1206 * out-of-order writes.
1208 ll_merge_attr(env, inode);
1209 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
1210 vio->vui_iocb->ki_pos = pos;
1212 LASSERTF(vio->vui_iocb->ki_pos == pos,
1213 "ki_pos %lld [%lld, %lld)\n",
1214 vio->vui_iocb->ki_pos,
1218 CDEBUG(D_VFSTRACE, "%s: write [%llu, %llu)\n",
1219 file_dentry(file)->d_name.name,
1222 /* The maximum Lustre file size is variable, based on the OST maximum
1223 * object size and number of stripes. This needs another check in
1224 * addition to the VFS checks earlier. */
1225 if (pos + cnt > ll_file_maxbytes(inode)) {
1227 "%s: file %s ("DFID") offset %llu > maxbytes %llu\n",
1228 ll_i2sbi(inode)->ll_fsname,
1229 file_dentry(file)->d_name.name,
1230 PFID(ll_inode2fid(inode)), pos + cnt,
1231 ll_file_maxbytes(inode));
1235 /* Tests to verify we take the i_mutex correctly */
1236 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_SEC) && !lock_inode)
1239 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_NOSEC) && lock_inode)
1242 if (vio->vui_iter == NULL) {
1243 /* from a temp io in ll_cl_init(). */
1247 * When using the locked AIO function (generic_file_aio_write())
1248 * testing has shown the inode mutex to be a limiting factor
1249 * with multi-threaded single shared file performance. To get
1250 * around this, we now use the lockless version. To maintain
1251 * consistency, proper locking to protect against writes,
1252 * trucates, etc. is handled in the higher layers of lustre.
1254 lock_inode = !IS_NOSEC(inode);
1256 if (unlikely(lock_inode))
1258 result = __generic_file_write_iter(vio->vui_iocb,
1260 if (unlikely(lock_inode))
1261 inode_unlock(inode);
1265 #ifdef HAVE_GENERIC_WRITE_SYNC_2ARGS
1266 result = generic_write_sync(vio->vui_iocb, result);
1271 err = generic_write_sync(vio->vui_iocb->ki_filp, pos,
1273 if (err < 0 && result > 0)
1280 result = vvp_io_write_commit(env, io);
1281 if (vio->u.write.vui_written > 0) {
1282 result = vio->u.write.vui_written;
1283 CDEBUG(D_VFSTRACE, "%s: write nob %zd, result: %zd\n",
1284 file_dentry(file)->d_name.name,
1285 io->ci_nob, result);
1286 io->ci_nob += result;
1289 if (vio->vui_iocb->ki_pos != (pos + io->ci_nob - nob)) {
1290 CDEBUG(D_VFSTRACE, "%s: write position mismatch: "
1291 "ki_pos %lld vs. pos %lld, written %ld, commit %ld "
1293 file_dentry(file)->d_name.name,
1294 vio->vui_iocb->ki_pos, pos + io->ci_nob - nob,
1295 written, io->ci_nob - nob, result);
1296 /* rewind ki_pos to where it has successfully committed */
1297 vio->vui_iocb->ki_pos = pos + io->ci_nob - nob;
1299 if (result > 0 || result == -EIOCBQUEUED) {
1300 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
1303 io->ci_continue = 0;
1311 static void vvp_io_rw_end(const struct lu_env *env,
1312 const struct cl_io_slice *ios)
1314 struct vvp_io *vio = cl2vvp_io(env, ios);
1315 struct inode *inode = vvp_object_inode(ios->cis_obj);
1316 struct ll_inode_info *lli = ll_i2info(inode);
1318 if (vio->vui_io_subtype == IO_NORMAL)
1319 trunc_sem_up_read(&lli->lli_trunc_sem);
1322 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
1324 struct vm_fault *vmf = cfio->ft_vmf;
1326 cfio->ft_flags = ll_filemap_fault(cfio->ft_vma, vmf);
1327 cfio->ft_flags_valid = 1;
1330 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
1331 get_vmf_address(vmf));
1332 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
1333 lock_page(vmf->page);
1334 cfio->ft_flags |= VM_FAULT_LOCKED;
1337 cfio->ft_vmpage = vmf->page;
1342 if (cfio->ft_flags & VM_FAULT_SIGBUS) {
1343 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", get_vmf_address(vmf));
1347 if (cfio->ft_flags & VM_FAULT_OOM) {
1348 CDEBUG(D_PAGE, "got addr %p - OOM\n", get_vmf_address(vmf));
1352 if (cfio->ft_flags & VM_FAULT_RETRY)
1355 CERROR("unknown error in page fault %d\n", cfio->ft_flags);
1360 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1361 struct pagevec *pvec)
1363 vvp_set_pagevec_dirty(pvec);
1366 static int vvp_io_fault_start(const struct lu_env *env,
1367 const struct cl_io_slice *ios)
1369 struct vvp_io *vio = cl2vvp_io(env, ios);
1370 struct cl_io *io = ios->cis_io;
1371 struct cl_object *obj = io->ci_obj;
1372 struct inode *inode = vvp_object_inode(obj);
1373 struct ll_inode_info *lli = ll_i2info(inode);
1374 struct cl_fault_io *fio = &io->u.ci_fault;
1375 struct vvp_fault_io *cfio = &vio->u.fault;
1378 struct page *vmpage = NULL;
1379 struct cl_page *page;
1384 trunc_sem_down_read_nowait(&lli->lli_trunc_sem);
1386 /* offset of the last byte on the page */
1387 offset = cl_offset(obj, fio->ft_index + 1) - 1;
1388 LASSERT(cl_index(obj, offset) == fio->ft_index);
1389 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1393 /* must return locked page */
1394 if (fio->ft_mkwrite) {
1395 LASSERT(cfio->ft_vmpage != NULL);
1396 lock_page(cfio->ft_vmpage);
1398 result = vvp_io_kernel_fault(cfio);
1403 vmpage = cfio->ft_vmpage;
1404 LASSERT(PageLocked(vmpage));
1406 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1407 generic_error_remove_page(vmpage->mapping, vmpage);
1409 size = i_size_read(inode);
1410 /* Though we have already held a cl_lock upon this page, but
1411 * it still can be truncated locally. */
1412 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1413 (page_offset(vmpage) > size))) {
1414 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1416 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1418 GOTO(out, result = +1);
1421 last_index = cl_index(obj, size - 1);
1423 if (fio->ft_mkwrite ) {
1425 * Capture the size while holding the lli_trunc_sem from above
1426 * we want to make sure that we complete the mkwrite action
1427 * while holding this lock. We need to make sure that we are
1428 * not past the end of the file.
1430 if (last_index < fio->ft_index) {
1432 "llite: mkwrite and truncate race happened: "
1433 "%p: 0x%lx 0x%lx\n",
1434 vmpage->mapping,fio->ft_index,last_index);
1436 * We need to return if we are
1437 * passed the end of the file. This will propagate
1438 * up the call stack to ll_page_mkwrite where
1439 * we will return VM_FAULT_NOPAGE. Any non-negative
1440 * value returned here will be silently
1441 * converted to 0. If the vmpage->mapping is null
1442 * the error code would be converted back to ENODATA
1443 * in ll_page_mkwrite0. Thus we return -ENODATA
1444 * to handle both cases
1446 GOTO(out, result = -ENODATA);
1450 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1452 GOTO(out, result = PTR_ERR(page));
1454 /* if page is going to be written, we should add this page into cache
1456 if (fio->ft_mkwrite) {
1457 wait_on_page_writeback(vmpage);
1458 if (!PageDirty(vmpage)) {
1459 struct cl_page_list *plist = &vio->u.fault.ft_queue;
1460 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1463 /* vvp_page_assume() calls wait_on_page_writeback(). */
1464 cl_page_assume(env, io, page);
1466 cl_page_list_init(plist);
1467 cl_page_list_add(plist, page);
1470 if (last_index == vvp_index(vpg))
1471 to = ((size - 1) & ~PAGE_MASK) + 1;
1473 /* Do not set Dirty bit here so that in case IO is
1474 * started before the page is really made dirty, we
1475 * still have chance to detect it. */
1476 result = cl_io_commit_async(env, io, plist, 0, to,
1477 mkwrite_commit_callback);
1478 /* Have overquota flag, trying sync write to check
1479 * whether indeed out of quota */
1480 if (result == -EDQUOT) {
1482 result = vvp_io_commit_sync(env, io,
1486 cl_page_own(env, io, page);
1487 cl_page_list_add(plist, page);
1488 lu_ref_add(&page->cp_reference,
1490 result = cl_io_commit_async(env, io,
1492 mkwrite_commit_callback);
1495 cl_page_put(env, page);
1499 LASSERT(cl_page_is_owned(page, io));
1500 cl_page_list_fini(env, plist);
1504 cl_page_discard(env, io, page);
1505 cl_page_disown(env, io, page);
1507 cl_page_put(env, page);
1509 /* we're in big trouble, what can we do now? */
1510 if (result == -EDQUOT)
1514 cl_page_disown(env, io, page);
1520 * The ft_index is only used in the case of
1521 * a mkwrite action. We need to check
1522 * our assertions are correct, since
1523 * we should have caught this above
1525 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1526 if (fio->ft_index == last_index)
1528 * Last page is mapped partially.
1530 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1532 fio->ft_nob = cl_page_size(obj);
1534 lu_ref_add(&page->cp_reference, "fault", io);
1535 fio->ft_page = page;
1539 /* return unlocked vmpage to avoid deadlocking */
1541 unlock_page(vmpage);
1543 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1548 static void vvp_io_fault_end(const struct lu_env *env,
1549 const struct cl_io_slice *ios)
1551 struct inode *inode = vvp_object_inode(ios->cis_obj);
1552 struct ll_inode_info *lli = ll_i2info(inode);
1554 CLOBINVRNT(env, ios->cis_io->ci_obj,
1555 vvp_object_invariant(ios->cis_io->ci_obj));
1556 trunc_sem_up_read(&lli->lli_trunc_sem);
1559 static int vvp_io_fsync_start(const struct lu_env *env,
1560 const struct cl_io_slice *ios)
1562 /* we should mark TOWRITE bit to each dirty page in radix tree to
1563 * verify pages have been written, but this is difficult because of
1568 static int vvp_io_read_ahead(const struct lu_env *env,
1569 const struct cl_io_slice *ios,
1570 pgoff_t start, struct cl_read_ahead *ra)
1575 if (ios->cis_io->ci_type == CIT_READ ||
1576 ios->cis_io->ci_type == CIT_FAULT) {
1577 struct vvp_io *vio = cl2vvp_io(env, ios);
1579 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1580 ra->cra_end_idx = CL_PAGE_EOF;
1581 result = +1; /* no need to call down */
1588 static const struct cl_io_operations vvp_io_ops = {
1591 .cio_fini = vvp_io_fini,
1592 .cio_lock = vvp_io_read_lock,
1593 .cio_start = vvp_io_read_start,
1594 .cio_end = vvp_io_rw_end,
1595 .cio_advance = vvp_io_advance,
1598 .cio_fini = vvp_io_fini,
1599 .cio_iter_init = vvp_io_write_iter_init,
1600 .cio_iter_fini = vvp_io_write_iter_fini,
1601 .cio_lock = vvp_io_write_lock,
1602 .cio_start = vvp_io_write_start,
1603 .cio_end = vvp_io_rw_end,
1604 .cio_advance = vvp_io_advance,
1607 .cio_fini = vvp_io_setattr_fini,
1608 .cio_iter_init = vvp_io_setattr_iter_init,
1609 .cio_lock = vvp_io_setattr_lock,
1610 .cio_start = vvp_io_setattr_start,
1611 .cio_end = vvp_io_setattr_end
1614 .cio_fini = vvp_io_fault_fini,
1615 .cio_iter_init = vvp_io_fault_iter_init,
1616 .cio_lock = vvp_io_fault_lock,
1617 .cio_start = vvp_io_fault_start,
1618 .cio_end = vvp_io_fault_end,
1621 .cio_start = vvp_io_fsync_start,
1622 .cio_fini = vvp_io_fini
1625 .cio_fini = vvp_io_fini
1628 .cio_fini = vvp_io_fini
1631 .cio_fini = vvp_io_fini
1634 .cio_read_ahead = vvp_io_read_ahead
1637 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1640 struct vvp_io *vio = vvp_env_io(env);
1641 struct inode *inode = vvp_object_inode(obj);
1644 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1647 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1648 "restore needed %d\n",
1649 PFID(lu_object_fid(&obj->co_lu)),
1650 io->ci_ignore_layout, io->ci_verify_layout,
1651 vio->vui_layout_gen, io->ci_restore_needed);
1653 CL_IO_SLICE_CLEAN(vio, vui_cl);
1654 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1655 vio->vui_ra_valid = false;
1657 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1659 struct ll_inode_info *lli = ll_i2info(inode);
1661 count = io->u.ci_rw.crw_count;
1662 /* "If nbyte is 0, read() will return 0 and have no other
1663 * results." -- Single Unix Spec */
1667 vio->vui_tot_count = count;
1669 /* for read/write, we store the jobid in the inode, and
1670 * it'll be fetched by osc when building RPC.
1672 * it's not accurate if the file is shared by different
1675 lustre_get_jobid(lli->lli_jobid, sizeof(lli->lli_jobid));
1676 } else if (io->ci_type == CIT_SETATTR) {
1677 if (!cl_io_is_trunc(io))
1678 io->ci_lockreq = CILR_MANDATORY;
1681 /* Enqueue layout lock and get layout version. We need to do this
1682 * even for operations requiring to open file, such as read and write,
1683 * because it might not grant layout lock in IT_OPEN. */
1684 if (result == 0 && !io->ci_ignore_layout) {
1685 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1686 if (result == -ENOENT)
1687 /* If the inode on MDS has been removed, but the objects
1688 * on OSTs haven't been destroyed (async unlink), layout
1689 * fetch will return -ENOENT, we'd ingore this error
1690 * and continue with dirty flush. LU-3230. */
1693 CERROR("%s: refresh file layout " DFID " error %d.\n",
1694 ll_i2sbi(inode)->ll_fsname,
1695 PFID(lu_object_fid(&obj->co_lu)), result);
1698 io->ci_result = result < 0 ? result : 0;