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
41 #include <linux/pagevec.h>
42 #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.readwrite.vui_queue);
271 vio->u.readwrite.vui_written = 0;
272 vio->u.readwrite.vui_from = 0;
273 vio->u.readwrite.vui_to = PAGE_SIZE;
278 static int vvp_io_read_iter_init(const struct lu_env *env,
279 const struct cl_io_slice *ios)
281 struct vvp_io *vio = cl2vvp_io(env, ios);
283 vio->u.readwrite.vui_read = 0;
288 static void vvp_io_write_iter_fini(const struct lu_env *env,
289 const struct cl_io_slice *ios)
291 struct vvp_io *vio = cl2vvp_io(env, ios);
293 LASSERT(vio->u.readwrite.vui_queue.pl_nr == 0);
296 static int vvp_io_fault_iter_init(const struct lu_env *env,
297 const struct cl_io_slice *ios)
299 struct vvp_io *vio = cl2vvp_io(env, ios);
300 struct inode *inode = vvp_object_inode(ios->cis_obj);
302 LASSERT(inode == file_inode(vio->vui_fd->fd_file));
307 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
309 struct cl_io *io = ios->cis_io;
310 struct cl_object *obj = io->ci_obj;
311 struct vvp_io *vio = cl2vvp_io(env, ios);
312 struct inode *inode = vvp_object_inode(obj);
317 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
319 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
320 "need write layout %d, restore needed %d\n",
321 PFID(lu_object_fid(&obj->co_lu)),
322 io->ci_ignore_layout, io->ci_verify_layout,
323 vio->vui_layout_gen, io->ci_need_write_intent,
324 io->ci_restore_needed);
326 if (io->ci_restore_needed) {
327 /* file was detected release, we need to restore it
328 * before finishing the io
330 rc = ll_layout_restore(inode, 0, OBD_OBJECT_EOF);
331 /* if restore registration failed, no restart,
332 * we will return -ENODATA */
333 /* The layout will change after restore, so we need to
334 * block on layout lock held by the MDT
335 * as MDT will not send new layout in lvb (see LU-3124)
336 * we have to explicitly fetch it, all this will be done
337 * by ll_layout_refresh().
338 * Even if ll_layout_restore() returns zero, it doesn't mean
339 * that restore has been successful. Therefore it sets
340 * ci_verify_layout so that it will check layout at the end
344 io->ci_restore_needed = 1;
345 io->ci_need_restart = 0;
346 io->ci_verify_layout = 0;
351 io->ci_restore_needed = 0;
353 /* Even if ll_layout_restore() returns zero, it doesn't mean
354 * that restore has been successful. Therefore it should verify
355 * if there was layout change and restart I/O correspondingly.
357 ll_layout_refresh(inode, &gen);
358 io->ci_need_restart = vio->vui_layout_gen != gen;
359 if (io->ci_need_restart) {
361 DFID" layout changed from %d to %d.\n",
362 PFID(lu_object_fid(&obj->co_lu)),
363 vio->vui_layout_gen, gen);
364 /* today successful restore is the only possible
366 /* restore was done, clear restoring state */
367 ll_file_clear_flag(ll_i2info(vvp_object_inode(obj)),
368 LLIF_FILE_RESTORING);
374 * dynamic layout change needed, send layout intent
377 if (io->ci_need_write_intent) {
378 enum layout_intent_opc opc = LAYOUT_INTENT_WRITE;
380 io->ci_need_write_intent = 0;
382 LASSERT(io->ci_type == CIT_WRITE ||
383 cl_io_is_trunc(io) || cl_io_is_mkwrite(io));
385 CDEBUG(D_VFSTRACE, DFID" write layout, type %u "DEXT"\n",
386 PFID(lu_object_fid(&obj->co_lu)), io->ci_type,
387 PEXT(&io->ci_write_intent));
389 if (cl_io_is_trunc(io))
390 opc = LAYOUT_INTENT_TRUNC;
392 rc = ll_layout_write_intent(inode, opc, &io->ci_write_intent);
395 io->ci_need_restart = 1;
399 if (!io->ci_need_restart &&
400 !io->ci_ignore_layout && io->ci_verify_layout) {
401 /* check layout version */
402 ll_layout_refresh(inode, &gen);
403 io->ci_need_restart = vio->vui_layout_gen != gen;
404 if (io->ci_need_restart) {
406 DFID" layout changed from %d to %d.\n",
407 PFID(lu_object_fid(&obj->co_lu)),
408 vio->vui_layout_gen, gen);
416 static void vvp_io_fault_fini(const struct lu_env *env,
417 const struct cl_io_slice *ios)
419 struct cl_io *io = ios->cis_io;
420 struct cl_page *page = io->u.ci_fault.ft_page;
422 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
425 lu_ref_del(&page->cp_reference, "fault", io);
426 cl_page_put(env, page);
427 io->u.ci_fault.ft_page = NULL;
429 vvp_io_fini(env, ios);
432 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
435 * we only want to hold PW locks if the mmap() can generate
436 * writes back to the file and that only happens in shared
439 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
444 static int vvp_mmap_locks(const struct lu_env *env,
445 struct vvp_io *vio, struct cl_io *io)
447 struct vvp_thread_info *vti = vvp_env_info(env);
448 struct mm_struct *mm = current->mm;
449 struct vm_area_struct *vma;
450 struct cl_lock_descr *descr = &vti->vti_descr;
451 union ldlm_policy_data policy;
459 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
461 if (!cl_is_normalio(env, io))
464 /* nfs or loop back device write */
465 if (vio->vui_iter == NULL)
468 /* No MM (e.g. NFS)? No vmas too. */
472 if (!iter_is_iovec(vio->vui_iter) && !iov_iter_is_kvec(vio->vui_iter))
475 for (i = *vio->vui_iter;
477 iov_iter_advance(&i, iov.iov_len)) {
478 iov = iov_iter_iovec(&i);
479 addr = (unsigned long)iov.iov_base;
485 count += addr & ~PAGE_MASK;
488 down_read(&mm->mmap_sem);
489 while((vma = our_vma(mm, addr, count)) != NULL) {
490 struct dentry *de = file_dentry(vma->vm_file);
491 struct inode *inode = de->d_inode;
492 int flags = CEF_MUST;
494 if (ll_file_nolock(vma->vm_file)) {
496 * For no lock case is not allowed for mmap
503 * XXX: Required lock mode can be weakened: CIT_WRITE
504 * io only ever reads user level buffer, and CIT_READ
507 policy_from_vma(&policy, vma, addr, count);
508 descr->cld_mode = vvp_mode_from_vma(vma);
509 descr->cld_obj = ll_i2info(inode)->lli_clob;
510 descr->cld_start = cl_index(descr->cld_obj,
511 policy.l_extent.start);
512 descr->cld_end = cl_index(descr->cld_obj,
513 policy.l_extent.end);
514 descr->cld_enq_flags = flags;
515 result = cl_io_lock_alloc_add(env, io, descr);
517 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
518 descr->cld_mode, descr->cld_start,
524 if (vma->vm_end - addr >= count)
527 count -= vma->vm_end - addr;
530 up_read(&mm->mmap_sem);
537 static void vvp_io_advance(const struct lu_env *env,
538 const struct cl_io_slice *ios,
541 struct vvp_io *vio = cl2vvp_io(env, ios);
542 struct cl_io *io = ios->cis_io;
543 struct cl_object *obj = ios->cis_io->ci_obj;
545 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
547 if (!cl_is_normalio(env, io))
551 * Since 3.16(26978b8b4) vfs revert iov iter to
552 * original position even io succeed, so instead
553 * of relying on VFS, we move iov iter by ourselves.
555 iov_iter_advance(vio->vui_iter, nob);
556 vio->vui_tot_count -= nob;
557 iov_iter_reexpand(vio->vui_iter, vio->vui_tot_count);
560 static void vvp_io_update_iov(const struct lu_env *env,
561 struct vvp_io *vio, struct cl_io *io)
563 size_t size = io->u.ci_rw.crw_count;
565 if (!cl_is_normalio(env, io) || vio->vui_iter == NULL)
568 iov_iter_truncate(vio->vui_iter, size);
571 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
572 enum cl_lock_mode mode, loff_t start, loff_t end)
574 struct vvp_io *vio = vvp_env_io(env);
578 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
581 vvp_io_update_iov(env, vio, io);
583 if (io->u.ci_rw.crw_nonblock)
584 ast_flags |= CEF_NONBLOCK;
585 if (io->ci_lock_no_expand)
586 ast_flags |= CEF_LOCK_NO_EXPAND;
588 /* Group lock held means no lockless any more */
589 if (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)
590 io->ci_ignore_lockless = 1;
592 if (ll_file_nolock(vio->vui_fd->fd_file) ||
593 (vio->vui_fd->fd_flags & LL_FILE_LOCKLESS_IO &&
594 !io->ci_ignore_lockless))
595 ast_flags |= CEF_NEVER;
598 result = vvp_mmap_locks(env, vio, io);
600 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
605 static int vvp_io_read_lock(const struct lu_env *env,
606 const struct cl_io_slice *ios)
608 struct cl_io *io = ios->cis_io;
609 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
613 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
614 rd->crw_pos + rd->crw_count - 1);
618 static int vvp_io_fault_lock(const struct lu_env *env,
619 const struct cl_io_slice *ios)
621 struct cl_io *io = ios->cis_io;
622 struct vvp_io *vio = cl2vvp_io(env, ios);
624 * XXX LDLM_FL_CBPENDING
626 return vvp_io_one_lock_index(env,
628 vvp_mode_from_vma(vio->u.fault.ft_vma),
629 io->u.ci_fault.ft_index,
630 io->u.ci_fault.ft_index);
633 static int vvp_io_write_lock(const struct lu_env *env,
634 const struct cl_io_slice *ios)
636 struct cl_io *io = ios->cis_io;
640 if (io->u.ci_wr.wr_append) {
642 end = OBD_OBJECT_EOF;
644 start = io->u.ci_wr.wr.crw_pos;
645 end = start + io->u.ci_wr.wr.crw_count - 1;
648 RETURN(vvp_io_rw_lock(env, io, CLM_WRITE, start, end));
651 static int vvp_io_setattr_iter_init(const struct lu_env *env,
652 const struct cl_io_slice *ios)
659 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
661 * Handles "lockless io" mode when extent locking is done by server.
663 static int vvp_io_setattr_lock(const struct lu_env *env,
664 const struct cl_io_slice *ios)
666 struct cl_io *io = ios->cis_io;
667 __u64 lock_start = 0;
668 __u64 lock_end = OBD_OBJECT_EOF;
671 if (cl_io_is_trunc(io)) {
672 struct inode *inode = vvp_object_inode(io->ci_obj);
674 /* set enqueue flags to CEF_MUST in case of encrypted file,
675 * to prevent lockless truncate
677 if (S_ISREG(inode->i_mode) && IS_ENCRYPTED(inode))
679 else if (io->u.ci_setattr.sa_attr.lvb_size == 0)
680 enqflags = CEF_DISCARD_DATA;
681 } else if (cl_io_is_fallocate(io)) {
682 lock_start = io->u.ci_setattr.sa_falloc_offset;
683 lock_end = lock_start + io->u.ci_setattr.sa_attr.lvb_size;
685 unsigned int valid = io->u.ci_setattr.sa_avalid;
687 if (!(valid & TIMES_SET_FLAGS))
690 if ((!(valid & ATTR_MTIME) ||
691 io->u.ci_setattr.sa_attr.lvb_mtime >=
692 io->u.ci_setattr.sa_attr.lvb_ctime) &&
693 (!(valid & ATTR_ATIME) ||
694 io->u.ci_setattr.sa_attr.lvb_atime >=
695 io->u.ci_setattr.sa_attr.lvb_ctime))
699 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
700 lock_start, lock_end);
703 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
708 * Only ll_inode_size_lock is taken at this level.
710 ll_inode_size_lock(inode);
711 result = inode_newsize_ok(inode, size);
713 ll_inode_size_unlock(inode);
716 i_size_write(inode, size);
718 ll_truncate_pagecache(inode, size);
719 ll_inode_size_unlock(inode);
723 static int vvp_io_setattr_time(const struct lu_env *env,
724 const struct cl_io_slice *ios)
726 struct cl_io *io = ios->cis_io;
727 struct cl_object *obj = io->ci_obj;
728 struct cl_attr *attr = vvp_env_thread_attr(env);
730 unsigned valid = CAT_CTIME;
732 cl_object_attr_lock(obj);
733 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
734 if (io->u.ci_setattr.sa_avalid & ATTR_ATIME_SET) {
735 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
738 if (io->u.ci_setattr.sa_avalid & ATTR_MTIME_SET) {
739 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
742 result = cl_object_attr_update(env, obj, attr, valid);
743 cl_object_attr_unlock(obj);
748 static int vvp_io_setattr_start(const struct lu_env *env,
749 const struct cl_io_slice *ios)
751 struct cl_io *io = ios->cis_io;
752 struct inode *inode = vvp_object_inode(io->ci_obj);
753 struct ll_inode_info *lli = ll_i2info(inode);
755 if (cl_io_is_trunc(io)) {
756 trunc_sem_down_write(&lli->lli_trunc_sem);
757 mutex_lock(&lli->lli_setattr_mutex);
758 inode_dio_wait(inode);
759 } else if (cl_io_is_fallocate(io)) {
761 inode_dio_wait(inode);
763 mutex_lock(&lli->lli_setattr_mutex);
766 if (io->u.ci_setattr.sa_avalid & TIMES_SET_FLAGS)
767 return vvp_io_setattr_time(env, ios);
772 static void vvp_io_setattr_end(const struct lu_env *env,
773 const struct cl_io_slice *ios)
775 struct cl_io *io = ios->cis_io;
776 struct inode *inode = vvp_object_inode(io->ci_obj);
777 struct ll_inode_info *lli = ll_i2info(inode);
779 if (cl_io_is_trunc(io)) {
780 /* Truncate in memory pages - they must be clean pages
781 * because osc has already notified to destroy osc_extents. */
782 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
783 mutex_unlock(&lli->lli_setattr_mutex);
784 trunc_sem_up_write(&lli->lli_trunc_sem);
785 } else if (cl_io_is_fallocate(io)) {
788 mutex_unlock(&lli->lli_setattr_mutex);
792 static void vvp_io_setattr_fini(const struct lu_env *env,
793 const struct cl_io_slice *ios)
795 bool restore_needed = ios->cis_io->ci_restore_needed;
796 struct inode *inode = vvp_object_inode(ios->cis_obj);
798 vvp_io_fini(env, ios);
800 if (restore_needed && !ios->cis_io->ci_restore_needed) {
801 /* restore finished, set data modified flag for HSM */
802 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
806 static int vvp_io_read_start(const struct lu_env *env,
807 const struct cl_io_slice *ios)
809 struct vvp_io *vio = cl2vvp_io(env, ios);
810 struct cl_io *io = ios->cis_io;
811 struct cl_object *obj = io->ci_obj;
812 struct inode *inode = vvp_object_inode(obj);
813 struct ll_inode_info *lli = ll_i2info(inode);
814 struct file *file = vio->vui_fd->fd_file;
815 loff_t pos = io->u.ci_rd.rd.crw_pos;
816 size_t cnt = io->u.ci_rd.rd.crw_count;
817 size_t tot = vio->vui_tot_count;
820 struct iov_iter iter;
824 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
826 CDEBUG(D_VFSTRACE, "%s: read [%llu, %llu)\n",
827 file_dentry(file)->d_name.name,
830 if (vio->vui_io_subtype == IO_NORMAL)
831 trunc_sem_down_read(&lli->lli_trunc_sem);
833 if (io->ci_async_readahead) {
838 if (!can_populate_pages(env, io, inode))
841 /* Unless this is reading a sparse file, otherwise the lock has already
842 * been acquired so vvp_prep_size() is an empty op. */
843 result = vvp_prep_size(env, obj, io, pos, cnt, &exceed);
846 else if (exceed != 0)
849 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
850 "Read ino %lu, %zu bytes, offset %lld, size %llu\n",
851 inode->i_ino, cnt, pos, i_size_read(inode));
853 /* turn off the kernel's read-ahead */
854 vio->vui_fd->fd_file->f_ra.ra_pages = 0;
856 /* initialize read-ahead window once per syscall */
857 if (!vio->vui_ra_valid) {
858 vio->vui_ra_valid = true;
859 vio->vui_ra_start_idx = cl_index(obj, pos);
860 vio->vui_ra_pages = cl_index(obj, tot + PAGE_SIZE - 1);
861 /* If both start and end are unaligned, we read one more page
862 * than the index math suggests. */
863 if ((pos & ~PAGE_MASK) != 0 && ((pos + tot) & ~PAGE_MASK) != 0)
866 CDEBUG(D_READA, "tot %zu, ra_start %lu, ra_count %lu\n",
867 tot, vio->vui_ra_start_idx, vio->vui_ra_pages);
872 switch (vio->vui_io_subtype) {
874 LASSERT(vio->vui_iocb->ki_pos == pos);
875 iter = *vio->vui_iter;
876 result = generic_file_read_iter(vio->vui_iocb, &iter);
879 result = generic_file_splice_read(file, &pos,
880 vio->u.splice.vui_pipe, cnt,
881 vio->u.splice.vui_flags);
882 /* LU-1109: do splice read stripe by stripe otherwise if it
883 * may make nfsd stuck if this read occupied all internal pipe
888 CERROR("Wrong IO type %u\n", vio->vui_io_subtype);
897 io->ci_nob += result;
899 } else if (result == -EIOCBQUEUED) {
900 io->ci_nob += vio->u.readwrite.vui_read;
902 vio->vui_iocb->ki_pos = pos +
903 vio->u.readwrite.vui_read;
909 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
910 struct cl_page_list *plist, int from, int to)
912 struct cl_2queue *queue = &io->ci_queue;
913 struct cl_page *page;
914 unsigned int bytes = 0;
918 if (plist->pl_nr == 0)
921 if (from > 0 || to != PAGE_SIZE) {
922 page = cl_page_list_first(plist);
923 if (plist->pl_nr == 1) {
924 cl_page_clip(env, page, from, to);
927 cl_page_clip(env, page, from, PAGE_SIZE);
928 if (to != PAGE_SIZE) {
929 page = cl_page_list_last(plist);
930 cl_page_clip(env, page, 0, to);
935 cl_2queue_init(queue);
936 cl_page_list_splice(plist, &queue->c2_qin);
937 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
939 /* plist is not sorted any more */
940 cl_page_list_splice(&queue->c2_qin, plist);
941 cl_page_list_splice(&queue->c2_qout, plist);
942 cl_2queue_fini(env, queue);
945 /* calculate bytes */
946 bytes = plist->pl_nr << PAGE_SHIFT;
947 bytes -= from + PAGE_SIZE - to;
949 while (plist->pl_nr > 0) {
950 page = cl_page_list_first(plist);
951 cl_page_list_del(env, plist, page);
953 cl_page_clip(env, page, 0, PAGE_SIZE);
955 SetPageUptodate(cl_page_vmpage(page));
956 cl_page_disown(env, io, page);
958 /* held in ll_cl_init() */
959 lu_ref_del(&page->cp_reference, "cl_io", io);
960 cl_page_put(env, page);
964 RETURN(bytes > 0 ? bytes : rc);
968 * Kernels 4.2 - 4.5 pass memcg argument to account_page_dirtied()
969 * Kernel v5.2-5678-gac1c3e4 no longer exports account_page_dirtied
971 static inline void ll_account_page_dirtied(struct page *page,
972 struct address_space *mapping)
974 #ifdef HAVE_ACCOUNT_PAGE_DIRTIED_3ARGS
975 struct mem_cgroup *memcg = mem_cgroup_begin_page_stat(page);
977 account_page_dirtied(page, mapping, memcg);
978 mem_cgroup_end_page_stat(memcg);
979 #elif defined(HAVE_ACCOUNT_PAGE_DIRTIED_EXPORT)
980 account_page_dirtied(page, mapping);
982 vvp_account_page_dirtied(page, mapping);
987 * From kernel v4.19-rc5-248-g9b89a0355144 use XArrary
988 * Prior kernels use radix_tree for tags
990 static inline void ll_page_tag_dirty(struct page *page,
991 struct address_space *mapping)
993 #ifndef HAVE_RADIX_TREE_TAG_SET
994 __xa_set_mark(&mapping->i_pages, page_index(page), PAGECACHE_TAG_DIRTY);
996 radix_tree_tag_set(&mapping->page_tree, page_index(page),
997 PAGECACHE_TAG_DIRTY);
1001 /* Taken from kernel set_page_dirty, __set_page_dirty_nobuffers
1002 * Last change to this area: b93b016313b3ba8003c3b8bb71f569af91f19fc7
1004 * Current with Linus tip of tree (7/13/2019):
1005 * v5.2-rc4-224-ge01e060fe0
1007 * Backwards compat for 3.x, 5.x kernels relating to memcg handling
1008 * & rename of radix tree to xarray.
1010 void vvp_set_pagevec_dirty(struct pagevec *pvec)
1012 struct page *page = pvec->pages[0];
1013 struct address_space *mapping = page->mapping;
1014 unsigned long flags;
1015 unsigned long skip_pages = 0;
1016 int count = pagevec_count(pvec);
1022 BUILD_BUG_ON(PAGEVEC_SIZE > BITS_PER_LONG);
1023 LASSERTF(page->mapping,
1024 "mapping must be set. page %p, page->private (cl_page) %p\n",
1025 page, (void *) page->private);
1027 for (i = 0; i < count; i++) {
1028 page = pvec->pages[i];
1030 ClearPageReclaim(page);
1032 lock_page_memcg(page);
1033 if (TestSetPageDirty(page)) {
1034 /* page is already dirty .. no extra work needed
1035 * set a flag for the i'th page to be skipped
1037 unlock_page_memcg(page);
1038 skip_pages |= (1 << i);
1042 ll_xa_lock_irqsave(&mapping->i_pages, flags);
1044 /* Notes on differences with __set_page_dirty_nobuffers:
1045 * 1. We don't need to call page_mapping because we know this is a page
1047 * 2. We have the pages locked, so there is no need for the careful
1048 * mapping/mapping2 dance.
1049 * 3. No mapping is impossible. (Race w/truncate mentioned in
1050 * dirty_nobuffers should be impossible because we hold the page lock.)
1051 * 4. All mappings are the same because i/o is only to one file.
1053 for (i = 0; i < count; i++) {
1054 page = pvec->pages[i];
1055 /* if the i'th page was unlocked above, skip it here */
1056 if ((skip_pages >> i) & 1)
1059 LASSERTF(page->mapping == mapping,
1060 "all pages must have the same mapping. page %p, mapping %p, first mapping %p\n",
1061 page, page->mapping, mapping);
1062 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1063 ll_account_page_dirtied(page, mapping);
1064 ll_page_tag_dirty(page, mapping);
1066 unlock_page_memcg(page);
1068 ll_xa_unlock_irqrestore(&mapping->i_pages, flags);
1070 CDEBUG(D_VFSTRACE, "mapping %p, count %d, dirtied %d\n", mapping,
1073 if (mapping->host && dirtied) {
1074 /* !PageAnon && !swapper_space */
1075 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1081 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
1082 struct pagevec *pvec)
1089 count = pagevec_count(pvec);
1092 for (i = 0; i < count; i++) {
1093 struct page *vmpage = pvec->pages[i];
1094 SetPageUptodate(vmpage);
1097 vvp_set_pagevec_dirty(pvec);
1099 for (i = 0; i < count; i++) {
1100 struct page *vmpage = pvec->pages[i];
1101 struct cl_page *page = (struct cl_page *) vmpage->private;
1102 cl_page_disown(env, io, page);
1103 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
1104 cl_page_put(env, page);
1110 /* make sure the page list is contiguous */
1111 static bool page_list_sanity_check(struct cl_object *obj,
1112 struct cl_page_list *plist)
1114 struct cl_page *page;
1115 pgoff_t index = CL_PAGE_EOF;
1117 cl_page_list_for_each(page, plist) {
1118 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1120 if (index == CL_PAGE_EOF) {
1121 index = vvp_index(vpg);
1126 if (index == vvp_index(vpg))
1134 /* Return how many bytes have queued or written */
1135 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
1137 struct cl_object *obj = io->ci_obj;
1138 struct inode *inode = vvp_object_inode(obj);
1139 struct vvp_io *vio = vvp_env_io(env);
1140 struct cl_page_list *queue = &vio->u.readwrite.vui_queue;
1141 struct cl_page *page;
1144 unsigned int npages = vio->u.readwrite.vui_queue.pl_nr;
1150 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
1151 npages, vio->u.readwrite.vui_from, vio->u.readwrite.vui_to);
1153 LASSERT(page_list_sanity_check(obj, queue));
1155 /* submit IO with async write */
1156 rc = cl_io_commit_async(env, io, queue,
1157 vio->u.readwrite.vui_from,
1158 vio->u.readwrite.vui_to,
1159 write_commit_callback);
1160 npages -= queue->pl_nr; /* already committed pages */
1162 /* calculate how many bytes were written */
1163 bytes = npages << PAGE_SHIFT;
1166 bytes -= vio->u.readwrite.vui_from;
1167 if (queue->pl_nr == 0) /* last page */
1168 bytes -= PAGE_SIZE - vio->u.readwrite.vui_to;
1169 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
1171 vio->u.readwrite.vui_written += bytes;
1173 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
1174 npages, bytes, vio->u.readwrite.vui_written);
1176 /* the first page must have been written. */
1177 vio->u.readwrite.vui_from = 0;
1179 LASSERT(page_list_sanity_check(obj, queue));
1180 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
1182 /* out of quota, try sync write */
1183 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
1184 rc = vvp_io_commit_sync(env, io, queue,
1185 vio->u.readwrite.vui_from,
1186 vio->u.readwrite.vui_to);
1188 vio->u.readwrite.vui_written += rc;
1193 /* update inode size */
1194 ll_merge_attr(env, inode);
1196 /* Now the pages in queue were failed to commit, discard them
1197 * unless they were dirtied before. */
1198 while (queue->pl_nr > 0) {
1199 page = cl_page_list_first(queue);
1200 cl_page_list_del(env, queue, page);
1202 if (!PageDirty(cl_page_vmpage(page)))
1203 cl_page_discard(env, io, page);
1205 cl_page_disown(env, io, page);
1207 /* held in ll_cl_init() */
1208 lu_ref_del(&page->cp_reference, "cl_io", io);
1209 cl_page_put(env, page);
1211 cl_page_list_fini(env, queue);
1216 static int vvp_io_write_start(const struct lu_env *env,
1217 const struct cl_io_slice *ios)
1219 struct vvp_io *vio = cl2vvp_io(env, ios);
1220 struct cl_io *io = ios->cis_io;
1221 struct cl_object *obj = io->ci_obj;
1222 struct inode *inode = vvp_object_inode(obj);
1223 struct ll_inode_info *lli = ll_i2info(inode);
1224 struct file *file = vio->vui_fd->fd_file;
1226 loff_t pos = io->u.ci_wr.wr.crw_pos;
1227 size_t cnt = io->u.ci_wr.wr.crw_count;
1228 bool lock_inode = !IS_NOSEC(inode);
1229 size_t nob = io->ci_nob;
1230 struct iov_iter iter;
1235 if (vio->vui_io_subtype == IO_NORMAL)
1236 trunc_sem_down_read(&lli->lli_trunc_sem);
1238 if (!can_populate_pages(env, io, inode))
1241 if (cl_io_is_append(io)) {
1243 * PARALLEL IO This has to be changed for parallel IO doing
1244 * out-of-order writes.
1246 ll_merge_attr(env, inode);
1247 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
1248 vio->vui_iocb->ki_pos = pos;
1250 LASSERTF(vio->vui_iocb->ki_pos == pos,
1251 "ki_pos %lld [%lld, %lld)\n",
1252 vio->vui_iocb->ki_pos,
1256 CDEBUG(D_VFSTRACE, "%s: write [%llu, %llu)\n",
1257 file_dentry(file)->d_name.name,
1260 /* The maximum Lustre file size is variable, based on the OST maximum
1261 * object size and number of stripes. This needs another check in
1262 * addition to the VFS checks earlier. */
1263 if (pos + cnt > ll_file_maxbytes(inode)) {
1265 "%s: file %s ("DFID") offset %llu > maxbytes %llu\n",
1266 ll_i2sbi(inode)->ll_fsname,
1267 file_dentry(file)->d_name.name,
1268 PFID(ll_inode2fid(inode)), pos + cnt,
1269 ll_file_maxbytes(inode));
1273 /* Tests to verify we take the i_mutex correctly */
1274 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_SEC) && !lock_inode)
1277 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_NOSEC) && lock_inode)
1280 if (vio->vui_iter == NULL) {
1281 /* from a temp io in ll_cl_init(). */
1285 * When using the locked AIO function (generic_file_aio_write())
1286 * testing has shown the inode mutex to be a limiting factor
1287 * with multi-threaded single shared file performance. To get
1288 * around this, we now use the lockless version. To maintain
1289 * consistency, proper locking to protect against writes,
1290 * trucates, etc. is handled in the higher layers of lustre.
1292 lock_inode = !IS_NOSEC(inode);
1293 iter = *vio->vui_iter;
1295 if (unlikely(lock_inode))
1297 result = __generic_file_write_iter(vio->vui_iocb, &iter);
1298 if (unlikely(lock_inode))
1299 inode_unlock(inode);
1303 #ifdef HAVE_GENERIC_WRITE_SYNC_2ARGS
1304 result = generic_write_sync(vio->vui_iocb, result);
1309 err = generic_write_sync(vio->vui_iocb->ki_filp, pos,
1311 if (err < 0 && result > 0)
1318 result = vvp_io_write_commit(env, io);
1319 /* Simulate short commit */
1320 if (CFS_FAULT_CHECK(OBD_FAIL_LLITE_SHORT_COMMIT)) {
1321 vio->u.readwrite.vui_written >>= 1;
1322 if (vio->u.readwrite.vui_written > 0)
1323 io->ci_need_restart = 1;
1325 if (vio->u.readwrite.vui_written > 0) {
1326 result = vio->u.readwrite.vui_written;
1327 CDEBUG(D_VFSTRACE, "%s: write nob %zd, result: %zd\n",
1328 file_dentry(file)->d_name.name,
1329 io->ci_nob, result);
1330 io->ci_nob += result;
1332 io->ci_continue = 0;
1335 if (vio->vui_iocb->ki_pos != (pos + io->ci_nob - nob)) {
1337 "%s: write position mismatch: ki_pos %lld vs. pos %lld, written %zd, commit %zd: rc = %zd\n",
1338 file_dentry(file)->d_name.name,
1339 vio->vui_iocb->ki_pos, pos + io->ci_nob - nob,
1340 written, io->ci_nob - nob, result);
1342 * Rewind ki_pos and vui_iter to where it has
1343 * successfully committed.
1345 vio->vui_iocb->ki_pos = pos + io->ci_nob - nob;
1347 if (result > 0 || result == -EIOCBQUEUED) {
1348 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
1350 if (result != -EIOCBQUEUED && result < cnt)
1351 io->ci_continue = 0;
1355 if (result == -EIOCBQUEUED) {
1356 io->ci_nob += vio->u.readwrite.vui_written;
1357 vio->vui_iocb->ki_pos = pos +
1358 vio->u.readwrite.vui_written;
1365 static void vvp_io_rw_end(const struct lu_env *env,
1366 const struct cl_io_slice *ios)
1368 struct vvp_io *vio = cl2vvp_io(env, ios);
1369 struct inode *inode = vvp_object_inode(ios->cis_obj);
1370 struct ll_inode_info *lli = ll_i2info(inode);
1372 if (vio->vui_io_subtype == IO_NORMAL)
1373 trunc_sem_up_read(&lli->lli_trunc_sem);
1376 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
1378 struct vm_fault *vmf = cfio->ft_vmf;
1380 cfio->ft_flags = ll_filemap_fault(cfio->ft_vma, vmf);
1381 cfio->ft_flags_valid = 1;
1384 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
1385 get_vmf_address(vmf));
1386 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
1387 lock_page(vmf->page);
1388 cfio->ft_flags |= VM_FAULT_LOCKED;
1391 cfio->ft_vmpage = vmf->page;
1396 if (cfio->ft_flags & VM_FAULT_SIGBUS) {
1397 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", get_vmf_address(vmf));
1401 if (cfio->ft_flags & VM_FAULT_OOM) {
1402 CDEBUG(D_PAGE, "got addr %p - OOM\n", get_vmf_address(vmf));
1406 if (cfio->ft_flags & VM_FAULT_RETRY)
1409 CERROR("unknown error in page fault %d\n", cfio->ft_flags);
1414 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1415 struct pagevec *pvec)
1417 vvp_set_pagevec_dirty(pvec);
1420 static int vvp_io_fault_start(const struct lu_env *env,
1421 const struct cl_io_slice *ios)
1423 struct vvp_io *vio = cl2vvp_io(env, ios);
1424 struct cl_io *io = ios->cis_io;
1425 struct cl_object *obj = io->ci_obj;
1426 struct inode *inode = vvp_object_inode(obj);
1427 struct ll_inode_info *lli = ll_i2info(inode);
1428 struct cl_fault_io *fio = &io->u.ci_fault;
1429 struct vvp_fault_io *cfio = &vio->u.fault;
1432 struct page *vmpage = NULL;
1433 struct cl_page *page;
1438 trunc_sem_down_read_nowait(&lli->lli_trunc_sem);
1440 /* offset of the last byte on the page */
1441 offset = cl_offset(obj, fio->ft_index + 1) - 1;
1442 LASSERT(cl_index(obj, offset) == fio->ft_index);
1443 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1447 /* must return locked page */
1448 if (fio->ft_mkwrite) {
1449 LASSERT(cfio->ft_vmpage != NULL);
1450 lock_page(cfio->ft_vmpage);
1452 result = vvp_io_kernel_fault(cfio);
1457 vmpage = cfio->ft_vmpage;
1458 LASSERT(PageLocked(vmpage));
1460 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1461 generic_error_remove_page(vmpage->mapping, vmpage);
1463 size = i_size_read(inode);
1464 /* Though we have already held a cl_lock upon this page, but
1465 * it still can be truncated locally. */
1466 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1467 (page_offset(vmpage) > size))) {
1468 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1470 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1472 GOTO(out, result = +1);
1475 last_index = cl_index(obj, size - 1);
1477 if (fio->ft_mkwrite ) {
1479 * Capture the size while holding the lli_trunc_sem from above
1480 * we want to make sure that we complete the mkwrite action
1481 * while holding this lock. We need to make sure that we are
1482 * not past the end of the file.
1484 if (last_index < fio->ft_index) {
1486 "llite: mkwrite and truncate race happened: "
1487 "%p: 0x%lx 0x%lx\n",
1488 vmpage->mapping,fio->ft_index,last_index);
1490 * We need to return if we are
1491 * passed the end of the file. This will propagate
1492 * up the call stack to ll_page_mkwrite where
1493 * we will return VM_FAULT_NOPAGE. Any non-negative
1494 * value returned here will be silently
1495 * converted to 0. If the vmpage->mapping is null
1496 * the error code would be converted back to ENODATA
1497 * in ll_page_mkwrite0. Thus we return -ENODATA
1498 * to handle both cases
1500 GOTO(out, result = -ENODATA);
1504 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1506 GOTO(out, result = PTR_ERR(page));
1508 /* if page is going to be written, we should add this page into cache
1510 if (fio->ft_mkwrite) {
1511 wait_on_page_writeback(vmpage);
1512 if (!PageDirty(vmpage)) {
1513 struct cl_page_list *plist = &vio->u.fault.ft_queue;
1514 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1517 /* vvp_page_assume() calls wait_on_page_writeback(). */
1518 cl_page_assume(env, io, page);
1520 cl_page_list_init(plist);
1521 cl_page_list_add(plist, page);
1524 if (last_index == vvp_index(vpg))
1525 to = ((size - 1) & ~PAGE_MASK) + 1;
1527 /* Do not set Dirty bit here so that in case IO is
1528 * started before the page is really made dirty, we
1529 * still have chance to detect it. */
1530 result = cl_io_commit_async(env, io, plist, 0, to,
1531 mkwrite_commit_callback);
1532 /* Have overquota flag, trying sync write to check
1533 * whether indeed out of quota */
1534 if (result == -EDQUOT) {
1536 result = vvp_io_commit_sync(env, io,
1540 cl_page_own(env, io, page);
1541 cl_page_list_add(plist, page);
1542 lu_ref_add(&page->cp_reference,
1544 result = cl_io_commit_async(env, io,
1546 mkwrite_commit_callback);
1549 cl_page_put(env, page);
1553 LASSERT(cl_page_is_owned(page, io));
1554 cl_page_list_fini(env, plist);
1558 cl_page_discard(env, io, page);
1559 cl_page_disown(env, io, page);
1561 cl_page_put(env, page);
1563 /* we're in big trouble, what can we do now? */
1564 if (result == -EDQUOT)
1568 cl_page_disown(env, io, page);
1574 * The ft_index is only used in the case of
1575 * a mkwrite action. We need to check
1576 * our assertions are correct, since
1577 * we should have caught this above
1579 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1580 if (fio->ft_index == last_index)
1582 * Last page is mapped partially.
1584 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1586 fio->ft_nob = cl_page_size(obj);
1588 lu_ref_add(&page->cp_reference, "fault", io);
1589 fio->ft_page = page;
1593 /* return unlocked vmpage to avoid deadlocking */
1595 unlock_page(vmpage);
1597 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1602 static void vvp_io_fault_end(const struct lu_env *env,
1603 const struct cl_io_slice *ios)
1605 struct inode *inode = vvp_object_inode(ios->cis_obj);
1606 struct ll_inode_info *lli = ll_i2info(inode);
1608 CLOBINVRNT(env, ios->cis_io->ci_obj,
1609 vvp_object_invariant(ios->cis_io->ci_obj));
1610 trunc_sem_up_read(&lli->lli_trunc_sem);
1613 static int vvp_io_fsync_start(const struct lu_env *env,
1614 const struct cl_io_slice *ios)
1616 /* we should mark TOWRITE bit to each dirty page in radix tree to
1617 * verify pages have been written, but this is difficult because of
1622 static int vvp_io_read_ahead(const struct lu_env *env,
1623 const struct cl_io_slice *ios,
1624 pgoff_t start, struct cl_read_ahead *ra)
1629 if (ios->cis_io->ci_type == CIT_READ ||
1630 ios->cis_io->ci_type == CIT_FAULT) {
1631 struct vvp_io *vio = cl2vvp_io(env, ios);
1633 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1634 ra->cra_end_idx = CL_PAGE_EOF;
1635 result = +1; /* no need to call down */
1642 static const struct cl_io_operations vvp_io_ops = {
1645 .cio_fini = vvp_io_fini,
1646 .cio_iter_init = vvp_io_read_iter_init,
1647 .cio_lock = vvp_io_read_lock,
1648 .cio_start = vvp_io_read_start,
1649 .cio_end = vvp_io_rw_end,
1650 .cio_advance = vvp_io_advance,
1653 .cio_fini = vvp_io_fini,
1654 .cio_iter_init = vvp_io_write_iter_init,
1655 .cio_iter_fini = vvp_io_write_iter_fini,
1656 .cio_lock = vvp_io_write_lock,
1657 .cio_start = vvp_io_write_start,
1658 .cio_end = vvp_io_rw_end,
1659 .cio_advance = vvp_io_advance,
1662 .cio_fini = vvp_io_setattr_fini,
1663 .cio_iter_init = vvp_io_setattr_iter_init,
1664 .cio_lock = vvp_io_setattr_lock,
1665 .cio_start = vvp_io_setattr_start,
1666 .cio_end = vvp_io_setattr_end
1669 .cio_fini = vvp_io_fault_fini,
1670 .cio_iter_init = vvp_io_fault_iter_init,
1671 .cio_lock = vvp_io_fault_lock,
1672 .cio_start = vvp_io_fault_start,
1673 .cio_end = vvp_io_fault_end,
1676 .cio_start = vvp_io_fsync_start,
1677 .cio_fini = vvp_io_fini
1680 .cio_fini = vvp_io_fini
1683 .cio_fini = vvp_io_fini
1686 .cio_fini = vvp_io_fini
1689 .cio_read_ahead = vvp_io_read_ahead
1692 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1695 struct vvp_io *vio = vvp_env_io(env);
1696 struct inode *inode = vvp_object_inode(obj);
1699 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1702 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1703 "restore needed %d\n",
1704 PFID(lu_object_fid(&obj->co_lu)),
1705 io->ci_ignore_layout, io->ci_verify_layout,
1706 vio->vui_layout_gen, io->ci_restore_needed);
1708 CL_IO_SLICE_CLEAN(vio, vui_cl);
1709 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1710 vio->vui_ra_valid = false;
1712 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1714 struct ll_inode_info *lli = ll_i2info(inode);
1716 count = io->u.ci_rw.crw_count;
1717 /* "If nbyte is 0, read() will return 0 and have no other
1718 * results." -- Single Unix Spec */
1722 vio->vui_tot_count = count;
1724 /* for read/write, we store the jobid in the inode, and
1725 * it'll be fetched by osc when building RPC.
1727 * it's not accurate if the file is shared by different
1730 lustre_get_jobid(lli->lli_jobid, sizeof(lli->lli_jobid));
1731 } else if (io->ci_type == CIT_SETATTR) {
1732 if (!cl_io_is_trunc(io))
1733 io->ci_lockreq = CILR_MANDATORY;
1736 /* Enqueue layout lock and get layout version. We need to do this
1737 * even for operations requiring to open file, such as read and write,
1738 * because it might not grant layout lock in IT_OPEN. */
1739 if (result == 0 && !io->ci_ignore_layout) {
1740 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1741 if (result == -ENOENT)
1742 /* If the inode on MDS has been removed, but the objects
1743 * on OSTs haven't been destroyed (async unlink), layout
1744 * fetch will return -ENOENT, we'd ingore this error
1745 * and continue with dirty flush. LU-3230. */
1748 CERROR("%s: refresh file layout " DFID " error %d.\n",
1749 ll_i2sbi(inode)->ll_fsname,
1750 PFID(lu_object_fid(&obj->co_lu)), result);
1753 io->ci_result = result < 0 ? result : 0;