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, 2016, 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 "llite_internal.h"
43 #include "vvp_internal.h"
45 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
46 const struct cl_io_slice *slice)
50 vio = container_of(slice, struct vvp_io, vui_cl);
51 LASSERT(vio == vvp_env_io(env));
57 * True, if \a io is a normal io, False for splice_{read,write}
59 static int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
61 struct vvp_io *vio = vvp_env_io(env);
63 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
65 return vio->vui_io_subtype == IO_NORMAL;
69 * For swapping layout. The file's layout may have changed.
70 * To avoid populating pages to a wrong stripe, we have to verify the
71 * correctness of layout. It works because swapping layout processes
72 * have to acquire group lock.
74 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
77 struct ll_inode_info *lli = ll_i2info(inode);
78 struct vvp_io *vio = vvp_env_io(env);
81 switch (io->ci_type) {
84 /* don't need lock here to check lli_layout_gen as we have held
85 * extent lock and GROUP lock has to hold to swap layout */
86 if (ll_layout_version_get(lli) != vio->vui_layout_gen ||
87 OBD_FAIL_CHECK_RESET(OBD_FAIL_LLITE_LOST_LAYOUT, 0)) {
88 io->ci_need_restart = 1;
89 /* this will cause a short read/write */
94 /* fault is okay because we've already had a page. */
102 static void vvp_object_size_lock(struct cl_object *obj)
104 struct inode *inode = vvp_object_inode(obj);
106 ll_inode_size_lock(inode);
107 cl_object_attr_lock(obj);
110 static void vvp_object_size_unlock(struct cl_object *obj)
112 struct inode *inode = vvp_object_inode(obj);
114 cl_object_attr_unlock(obj);
115 ll_inode_size_unlock(inode);
119 * Helper function that if necessary adjusts file size (inode->i_size), when
120 * position at the offset \a pos is accessed. File size can be arbitrary stale
121 * on a Lustre client, but client at least knows KMS. If accessed area is
122 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
124 * Locking: i_size_lock is used to serialize changes to inode size and to
125 * protect consistency between inode size and cl_object
126 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
127 * top-object and sub-objects.
129 static int vvp_prep_size(const struct lu_env *env, struct cl_object *obj,
130 struct cl_io *io, loff_t start, size_t count,
133 struct cl_attr *attr = vvp_env_thread_attr(env);
134 struct inode *inode = vvp_object_inode(obj);
135 loff_t pos = start + count - 1;
140 * Consistency guarantees: following possibilities exist for the
141 * relation between region being accessed and real file size at this
144 * (A): the region is completely inside of the file;
146 * (B-x): x bytes of region are inside of the file, the rest is
149 * (C): the region is completely outside of the file.
151 * This classification is stable under DLM lock already acquired by
152 * the caller, because to change the class, other client has to take
153 * DLM lock conflicting with our lock. Also, any updates to ->i_size
154 * by other threads on this client are serialized by
155 * ll_inode_size_lock(). This guarantees that short reads are handled
156 * correctly in the face of concurrent writes and truncates.
158 vvp_object_size_lock(obj);
159 result = cl_object_attr_get(env, obj, attr);
164 * A glimpse is necessary to determine whether we
165 * return a short read (B) or some zeroes at the end
168 vvp_object_size_unlock(obj);
169 result = cl_glimpse_lock(env, io, inode, obj, 0);
170 if (result == 0 && exceed != NULL) {
171 /* If objective page index exceed end-of-file
172 * page index, return directly. Do not expect
173 * kernel will check such case correctly.
174 * linux-2.6.18-128.1.1 miss to do that.
176 loff_t size = i_size_read(inode);
177 unsigned long cur_index = start >>
180 if ((size == 0 && cur_index != 0) ||
181 (((size - 1) >> PAGE_SHIFT) <
189 * region is within kms and, hence, within real file
190 * size (A). We need to increase i_size to cover the
191 * read region so that generic_file_read() will do its
192 * job, but that doesn't mean the kms size is
193 * _correct_, it is only the _minimum_ size. If
194 * someone does a stat they will get the correct size
195 * which will always be >= the kms value here.
198 if (i_size_read(inode) < kms) {
199 i_size_write(inode, kms);
201 DFID" updating i_size %llu\n",
202 PFID(lu_object_fid(&obj->co_lu)),
203 (__u64)i_size_read(inode));
208 vvp_object_size_unlock(obj);
213 /*****************************************************************************
219 static int vvp_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
220 __u32 enqflags, enum cl_lock_mode mode,
221 pgoff_t start, pgoff_t end)
223 struct vvp_io *vio = vvp_env_io(env);
224 struct cl_lock_descr *descr = &vio->vui_link.cill_descr;
225 struct cl_object *obj = io->ci_obj;
227 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
230 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
232 memset(&vio->vui_link, 0, sizeof vio->vui_link);
234 if (vio->vui_fd && (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
235 descr->cld_mode = CLM_GROUP;
236 descr->cld_gid = vio->vui_fd->fd_grouplock.lg_gid;
237 enqflags |= CEF_LOCK_MATCH;
239 descr->cld_mode = mode;
242 descr->cld_obj = obj;
243 descr->cld_start = start;
244 descr->cld_end = end;
245 descr->cld_enq_flags = enqflags;
247 cl_io_lock_add(env, io, &vio->vui_link);
252 static int vvp_io_one_lock(const struct lu_env *env, struct cl_io *io,
253 __u32 enqflags, enum cl_lock_mode mode,
254 loff_t start, loff_t end)
256 struct cl_object *obj = io->ci_obj;
258 return vvp_io_one_lock_index(env, io, enqflags, mode,
259 cl_index(obj, start), cl_index(obj, end));
262 static int vvp_io_write_iter_init(const struct lu_env *env,
263 const struct cl_io_slice *ios)
265 struct vvp_io *vio = cl2vvp_io(env, ios);
267 cl_page_list_init(&vio->u.write.vui_queue);
268 vio->u.write.vui_written = 0;
269 vio->u.write.vui_from = 0;
270 vio->u.write.vui_to = PAGE_SIZE;
275 static void vvp_io_write_iter_fini(const struct lu_env *env,
276 const struct cl_io_slice *ios)
278 struct vvp_io *vio = cl2vvp_io(env, ios);
280 LASSERT(vio->u.write.vui_queue.pl_nr == 0);
283 static int vvp_io_fault_iter_init(const struct lu_env *env,
284 const struct cl_io_slice *ios)
286 struct vvp_io *vio = cl2vvp_io(env, ios);
287 struct inode *inode = vvp_object_inode(ios->cis_obj);
289 LASSERT(inode == file_inode(vio->vui_fd->fd_file));
290 vio->u.fault.ft_mtime = inode->i_mtime.tv_sec;
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);
302 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
304 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
305 "restore needed %d\n",
306 PFID(lu_object_fid(&obj->co_lu)),
307 io->ci_ignore_layout, io->ci_verify_layout,
308 vio->vui_layout_gen, io->ci_restore_needed);
310 if (io->ci_restore_needed) {
313 /* file was detected release, we need to restore it
314 * before finishing the io
316 rc = ll_layout_restore(inode, 0, OBD_OBJECT_EOF);
317 /* if restore registration failed, no restart,
318 * we will return -ENODATA */
319 /* The layout will change after restore, so we need to
320 * block on layout lock held by the MDT
321 * as MDT will not send new layout in lvb (see LU-3124)
322 * we have to explicitly fetch it, all this will be done
323 * by ll_layout_refresh()
326 io->ci_restore_needed = 0;
327 io->ci_need_restart = 1;
328 io->ci_verify_layout = 1;
330 io->ci_restore_needed = 1;
331 io->ci_need_restart = 0;
332 io->ci_verify_layout = 0;
337 if (!io->ci_ignore_layout && io->ci_verify_layout) {
340 /* check layout version */
341 ll_layout_refresh(inode, &gen);
342 io->ci_need_restart = vio->vui_layout_gen != gen;
343 if (io->ci_need_restart) {
345 DFID" layout changed from %d to %d.\n",
346 PFID(lu_object_fid(&obj->co_lu)),
347 vio->vui_layout_gen, gen);
348 /* today successful restore is the only possible
350 /* restore was done, clear restoring state */
351 ll_file_clear_flag(ll_i2info(vvp_object_inode(obj)),
352 LLIF_FILE_RESTORING);
357 static void vvp_io_fault_fini(const struct lu_env *env,
358 const struct cl_io_slice *ios)
360 struct cl_io *io = ios->cis_io;
361 struct cl_page *page = io->u.ci_fault.ft_page;
363 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
366 lu_ref_del(&page->cp_reference, "fault", io);
367 cl_page_put(env, page);
368 io->u.ci_fault.ft_page = NULL;
370 vvp_io_fini(env, ios);
373 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
376 * we only want to hold PW locks if the mmap() can generate
377 * writes back to the file and that only happens in shared
380 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
385 static int vvp_mmap_locks(const struct lu_env *env,
386 struct vvp_io *vio, struct cl_io *io)
388 struct vvp_thread_info *vti = vvp_env_info(env);
389 struct mm_struct *mm = current->mm;
390 struct vm_area_struct *vma;
391 struct cl_lock_descr *descr = &vti->vti_descr;
392 union ldlm_policy_data policy;
398 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
400 if (!cl_is_normalio(env, io))
403 /* nfs or loop back device write */
404 if (vio->vui_iter == NULL)
407 /* No MM (e.g. NFS)? No vmas too. */
411 iov_for_each(iov, i, *(vio->vui_iter)) {
412 unsigned long addr = (unsigned long)iov.iov_base;
413 size_t count = iov.iov_len;
418 count += addr & ~PAGE_MASK;
421 down_read(&mm->mmap_sem);
422 while((vma = our_vma(mm, addr, count)) != NULL) {
423 struct dentry *de = file_dentry(vma->vm_file);
424 struct inode *inode = de->d_inode;
425 int flags = CEF_MUST;
427 if (ll_file_nolock(vma->vm_file)) {
429 * For no lock case is not allowed for mmap
436 * XXX: Required lock mode can be weakened: CIT_WRITE
437 * io only ever reads user level buffer, and CIT_READ
440 policy_from_vma(&policy, vma, addr, count);
441 descr->cld_mode = vvp_mode_from_vma(vma);
442 descr->cld_obj = ll_i2info(inode)->lli_clob;
443 descr->cld_start = cl_index(descr->cld_obj,
444 policy.l_extent.start);
445 descr->cld_end = cl_index(descr->cld_obj,
446 policy.l_extent.end);
447 descr->cld_enq_flags = flags;
448 result = cl_io_lock_alloc_add(env, io, descr);
450 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
451 descr->cld_mode, descr->cld_start,
457 if (vma->vm_end - addr >= count)
460 count -= vma->vm_end - addr;
463 up_read(&mm->mmap_sem);
470 static void vvp_io_advance(const struct lu_env *env,
471 const struct cl_io_slice *ios,
474 struct vvp_io *vio = cl2vvp_io(env, ios);
475 struct cl_io *io = ios->cis_io;
476 struct cl_object *obj = ios->cis_io->ci_obj;
478 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
480 if (!cl_is_normalio(env, io))
483 vio->vui_tot_count -= nob;
484 iov_iter_reexpand(vio->vui_iter, vio->vui_tot_count);
487 static void vvp_io_update_iov(const struct lu_env *env,
488 struct vvp_io *vio, struct cl_io *io)
490 size_t size = io->u.ci_rw.crw_count;
492 if (!cl_is_normalio(env, io) || vio->vui_iter == NULL)
495 iov_iter_truncate(vio->vui_iter, size);
498 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
499 enum cl_lock_mode mode, loff_t start, loff_t end)
501 struct vvp_io *vio = vvp_env_io(env);
505 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
508 vvp_io_update_iov(env, vio, io);
510 if (io->u.ci_rw.crw_nonblock)
511 ast_flags |= CEF_NONBLOCK;
513 result = vvp_mmap_locks(env, vio, io);
515 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
520 static int vvp_io_read_lock(const struct lu_env *env,
521 const struct cl_io_slice *ios)
523 struct cl_io *io = ios->cis_io;
524 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
528 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
529 rd->crw_pos + rd->crw_count - 1);
533 static int vvp_io_fault_lock(const struct lu_env *env,
534 const struct cl_io_slice *ios)
536 struct cl_io *io = ios->cis_io;
537 struct vvp_io *vio = cl2vvp_io(env, ios);
539 * XXX LDLM_FL_CBPENDING
541 return vvp_io_one_lock_index(env,
543 vvp_mode_from_vma(vio->u.fault.ft_vma),
544 io->u.ci_fault.ft_index,
545 io->u.ci_fault.ft_index);
548 static int vvp_io_write_lock(const struct lu_env *env,
549 const struct cl_io_slice *ios)
551 struct cl_io *io = ios->cis_io;
555 if (io->u.ci_wr.wr_append) {
557 end = OBD_OBJECT_EOF;
559 start = io->u.ci_wr.wr.crw_pos;
560 end = start + io->u.ci_wr.wr.crw_count - 1;
562 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
565 static int vvp_io_setattr_iter_init(const struct lu_env *env,
566 const struct cl_io_slice *ios)
572 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
574 * Handles "lockless io" mode when extent locking is done by server.
576 static int vvp_io_setattr_lock(const struct lu_env *env,
577 const struct cl_io_slice *ios)
579 struct cl_io *io = ios->cis_io;
583 if (cl_io_is_trunc(io)) {
584 new_size = io->u.ci_setattr.sa_attr.lvb_size;
586 enqflags = CEF_DISCARD_DATA;
588 unsigned int valid = io->u.ci_setattr.sa_valid;
590 if (!(valid & TIMES_SET_FLAGS))
593 if ((!(valid & ATTR_MTIME) ||
594 io->u.ci_setattr.sa_attr.lvb_mtime >=
595 io->u.ci_setattr.sa_attr.lvb_ctime) &&
596 (!(valid & ATTR_ATIME) ||
597 io->u.ci_setattr.sa_attr.lvb_atime >=
598 io->u.ci_setattr.sa_attr.lvb_ctime))
604 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
605 new_size, OBD_OBJECT_EOF);
608 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
613 * Only ll_inode_size_lock is taken at this level.
615 ll_inode_size_lock(inode);
616 result = inode_newsize_ok(inode, size);
618 ll_inode_size_unlock(inode);
621 i_size_write(inode, size);
623 ll_truncate_pagecache(inode, size);
624 ll_inode_size_unlock(inode);
628 static int vvp_io_setattr_time(const struct lu_env *env,
629 const struct cl_io_slice *ios)
631 struct cl_io *io = ios->cis_io;
632 struct cl_object *obj = io->ci_obj;
633 struct cl_attr *attr = vvp_env_thread_attr(env);
635 unsigned valid = CAT_CTIME;
637 cl_object_attr_lock(obj);
638 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
639 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
640 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
643 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
644 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
647 result = cl_object_attr_update(env, obj, attr, valid);
648 cl_object_attr_unlock(obj);
653 static int vvp_io_setattr_start(const struct lu_env *env,
654 const struct cl_io_slice *ios)
656 struct cl_io *io = ios->cis_io;
657 struct inode *inode = vvp_object_inode(io->ci_obj);
658 struct ll_inode_info *lli = ll_i2info(inode);
660 if (cl_io_is_trunc(io)) {
661 down_write(&lli->lli_trunc_sem);
663 inode_dio_wait(inode);
668 if (io->u.ci_setattr.sa_valid & TIMES_SET_FLAGS)
669 return vvp_io_setattr_time(env, ios);
674 static void vvp_io_setattr_end(const struct lu_env *env,
675 const struct cl_io_slice *ios)
677 struct cl_io *io = ios->cis_io;
678 struct inode *inode = vvp_object_inode(io->ci_obj);
679 struct ll_inode_info *lli = ll_i2info(inode);
681 if (cl_io_is_trunc(io)) {
682 /* Truncate in memory pages - they must be clean pages
683 * because osc has already notified to destroy osc_extents. */
684 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
685 inode_dio_write_done(inode);
687 up_write(&lli->lli_trunc_sem);
693 static void vvp_io_setattr_fini(const struct lu_env *env,
694 const struct cl_io_slice *ios)
696 bool restore_needed = ios->cis_io->ci_restore_needed;
697 struct inode *inode = vvp_object_inode(ios->cis_obj);
699 vvp_io_fini(env, ios);
701 if (restore_needed && !ios->cis_io->ci_restore_needed) {
702 /* restore finished, set data modified flag for HSM */
703 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
707 static int vvp_io_read_start(const struct lu_env *env,
708 const struct cl_io_slice *ios)
710 struct vvp_io *vio = cl2vvp_io(env, ios);
711 struct cl_io *io = ios->cis_io;
712 struct cl_object *obj = io->ci_obj;
713 struct inode *inode = vvp_object_inode(obj);
714 struct ll_inode_info *lli = ll_i2info(inode);
715 struct file *file = vio->vui_fd->fd_file;
718 loff_t pos = io->u.ci_rd.rd.crw_pos;
719 long cnt = io->u.ci_rd.rd.crw_count;
720 long tot = vio->vui_tot_count;
723 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
725 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
727 if (vio->vui_io_subtype == IO_NORMAL)
728 down_read(&lli->lli_trunc_sem);
730 if (!can_populate_pages(env, io, inode))
733 result = vvp_prep_size(env, obj, io, pos, tot, &exceed);
736 else if (exceed != 0)
739 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
740 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
741 inode->i_ino, cnt, pos, i_size_read(inode));
743 /* turn off the kernel's read-ahead */
744 vio->vui_fd->fd_file->f_ra.ra_pages = 0;
746 /* initialize read-ahead window once per syscall */
747 if (!vio->vui_ra_valid) {
748 vio->vui_ra_valid = true;
749 vio->vui_ra_start = cl_index(obj, pos);
750 vio->vui_ra_count = cl_index(obj, tot + PAGE_SIZE - 1);
756 switch (vio->vui_io_subtype) {
758 LASSERT(vio->vui_iocb->ki_pos == pos);
759 result = generic_file_read_iter(vio->vui_iocb, vio->vui_iter);
762 result = generic_file_splice_read(file, &pos,
763 vio->u.splice.vui_pipe, cnt,
764 vio->u.splice.vui_flags);
765 /* LU-1109: do splice read stripe by stripe otherwise if it
766 * may make nfsd stuck if this read occupied all internal pipe
771 CERROR("Wrong IO type %u\n", vio->vui_io_subtype);
779 io->ci_nob += result;
780 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, vio->vui_fd,
788 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
789 struct cl_page_list *plist, int from, int to)
791 struct cl_2queue *queue = &io->ci_queue;
792 struct cl_page *page;
793 unsigned int bytes = 0;
797 if (plist->pl_nr == 0)
800 if (from > 0 || to != PAGE_SIZE) {
801 page = cl_page_list_first(plist);
802 if (plist->pl_nr == 1) {
803 cl_page_clip(env, page, from, to);
806 cl_page_clip(env, page, from, PAGE_SIZE);
807 if (to != PAGE_SIZE) {
808 page = cl_page_list_last(plist);
809 cl_page_clip(env, page, 0, to);
814 cl_2queue_init(queue);
815 cl_page_list_splice(plist, &queue->c2_qin);
816 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
818 /* plist is not sorted any more */
819 cl_page_list_splice(&queue->c2_qin, plist);
820 cl_page_list_splice(&queue->c2_qout, plist);
821 cl_2queue_fini(env, queue);
824 /* calculate bytes */
825 bytes = plist->pl_nr << PAGE_SHIFT;
826 bytes -= from + PAGE_SIZE - to;
828 while (plist->pl_nr > 0) {
829 page = cl_page_list_first(plist);
830 cl_page_list_del(env, plist, page);
832 cl_page_clip(env, page, 0, PAGE_SIZE);
834 SetPageUptodate(cl_page_vmpage(page));
835 cl_page_disown(env, io, page);
837 /* held in ll_cl_init() */
838 lu_ref_del(&page->cp_reference, "cl_io", io);
839 cl_page_put(env, page);
843 RETURN(bytes > 0 ? bytes : rc);
846 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
847 struct cl_page *page)
849 struct page *vmpage = page->cp_vmpage;
851 SetPageUptodate(vmpage);
852 set_page_dirty(vmpage);
854 cl_page_disown(env, io, page);
856 /* held in ll_cl_init() */
857 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
858 cl_page_put(env, page);
861 /* make sure the page list is contiguous */
862 static bool page_list_sanity_check(struct cl_object *obj,
863 struct cl_page_list *plist)
865 struct cl_page *page;
866 pgoff_t index = CL_PAGE_EOF;
868 cl_page_list_for_each(page, plist) {
869 struct vvp_page *vpg = cl_object_page_slice(obj, page);
871 if (index == CL_PAGE_EOF) {
872 index = vvp_index(vpg);
877 if (index == vvp_index(vpg))
885 /* Return how many bytes have queued or written */
886 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
888 struct cl_object *obj = io->ci_obj;
889 struct inode *inode = vvp_object_inode(obj);
890 struct vvp_io *vio = vvp_env_io(env);
891 struct cl_page_list *queue = &vio->u.write.vui_queue;
892 struct cl_page *page;
895 unsigned int npages = vio->u.write.vui_queue.pl_nr;
901 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
902 npages, vio->u.write.vui_from, vio->u.write.vui_to);
904 LASSERT(page_list_sanity_check(obj, queue));
906 /* submit IO with async write */
907 rc = cl_io_commit_async(env, io, queue,
908 vio->u.write.vui_from, vio->u.write.vui_to,
909 write_commit_callback);
910 npages -= queue->pl_nr; /* already committed pages */
912 /* calculate how many bytes were written */
913 bytes = npages << PAGE_SHIFT;
916 bytes -= vio->u.write.vui_from;
917 if (queue->pl_nr == 0) /* last page */
918 bytes -= PAGE_SIZE - vio->u.write.vui_to;
919 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
921 vio->u.write.vui_written += bytes;
923 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
924 npages, bytes, vio->u.write.vui_written);
926 /* the first page must have been written. */
927 vio->u.write.vui_from = 0;
929 LASSERT(page_list_sanity_check(obj, queue));
930 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
932 /* out of quota, try sync write */
933 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
934 rc = vvp_io_commit_sync(env, io, queue,
935 vio->u.write.vui_from,
936 vio->u.write.vui_to);
938 vio->u.write.vui_written += rc;
943 /* update inode size */
944 ll_merge_attr(env, inode);
946 /* Now the pages in queue were failed to commit, discard them
947 * unless they were dirtied before. */
948 while (queue->pl_nr > 0) {
949 page = cl_page_list_first(queue);
950 cl_page_list_del(env, queue, page);
952 if (!PageDirty(cl_page_vmpage(page)))
953 cl_page_discard(env, io, page);
955 cl_page_disown(env, io, page);
957 /* held in ll_cl_init() */
958 lu_ref_del(&page->cp_reference, "cl_io", io);
959 cl_page_put(env, page);
961 cl_page_list_fini(env, queue);
966 static int vvp_io_write_start(const struct lu_env *env,
967 const struct cl_io_slice *ios)
969 struct vvp_io *vio = cl2vvp_io(env, ios);
970 struct cl_io *io = ios->cis_io;
971 struct cl_object *obj = io->ci_obj;
972 struct inode *inode = vvp_object_inode(obj);
973 struct ll_inode_info *lli = ll_i2info(inode);
975 loff_t pos = io->u.ci_wr.wr.crw_pos;
976 size_t cnt = io->u.ci_wr.wr.crw_count;
980 if (vio->vui_io_subtype == IO_NORMAL)
981 down_read(&lli->lli_trunc_sem);
983 if (!can_populate_pages(env, io, inode))
986 if (cl_io_is_append(io)) {
988 * PARALLEL IO This has to be changed for parallel IO doing
989 * out-of-order writes.
991 ll_merge_attr(env, inode);
992 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
993 vio->vui_iocb->ki_pos = pos;
995 LASSERT(vio->vui_iocb->ki_pos == pos);
998 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
1000 /* The maximum Lustre file size is variable, based on the OST maximum
1001 * object size and number of stripes. This needs another check in
1002 * addition to the VFS checks earlier. */
1003 if (pos + cnt > ll_file_maxbytes(inode)) {
1005 "%s: file "DFID" offset %llu > maxbytes %llu\n",
1006 ll_get_fsname(inode->i_sb, NULL, 0),
1007 PFID(ll_inode2fid(inode)), pos + cnt,
1008 ll_file_maxbytes(inode));
1012 if (vio->vui_iter == NULL) {
1013 /* from a temp io in ll_cl_init(). */
1017 * When using the locked AIO function (generic_file_aio_write())
1018 * testing has shown the inode mutex to be a limiting factor
1019 * with multi-threaded single shared file performance. To get
1020 * around this, we now use the lockless version. To maintain
1021 * consistency, proper locking to protect against writes,
1022 * trucates, etc. is handled in the higher layers of lustre.
1024 bool lock_node = !IS_NOSEC(inode);
1028 result = __generic_file_write_iter(vio->vui_iocb,
1031 inode_unlock(inode);
1033 if (result > 0 || result == -EIOCBQUEUED) {
1036 err = generic_write_sync(vio->vui_iocb->ki_filp,
1038 if (err < 0 && result > 0)
1044 result = vvp_io_write_commit(env, io);
1045 if (vio->u.write.vui_written > 0) {
1046 result = vio->u.write.vui_written;
1047 io->ci_nob += result;
1049 CDEBUG(D_VFSTRACE, "write: nob %zd, result: %zd\n",
1050 io->ci_nob, result);
1054 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
1057 io->ci_continue = 0;
1058 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
1059 vio->vui_fd, pos, result, WRITE);
1066 static void vvp_io_rw_end(const struct lu_env *env,
1067 const struct cl_io_slice *ios)
1069 struct vvp_io *vio = cl2vvp_io(env, ios);
1070 struct inode *inode = vvp_object_inode(ios->cis_obj);
1071 struct ll_inode_info *lli = ll_i2info(inode);
1073 if (vio->vui_io_subtype == IO_NORMAL)
1074 up_read(&lli->lli_trunc_sem);
1077 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
1079 struct vm_fault *vmf = cfio->ft_vmf;
1081 cfio->ft_flags = filemap_fault(cfio->ft_vma, vmf);
1082 cfio->ft_flags_valid = 1;
1085 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
1086 vmf->virtual_address);
1087 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
1088 lock_page(vmf->page);
1089 cfio->ft_flags |= VM_FAULT_LOCKED;
1092 cfio->ft_vmpage = vmf->page;
1097 if (cfio->ft_flags & VM_FAULT_SIGBUS) {
1098 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
1102 if (cfio->ft_flags & VM_FAULT_OOM) {
1103 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
1107 if (cfio->ft_flags & VM_FAULT_RETRY)
1110 CERROR("unknown error in page fault %d\n", cfio->ft_flags);
1115 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1116 struct cl_page *page)
1118 set_page_dirty(page->cp_vmpage);
1121 static int vvp_io_fault_start(const struct lu_env *env,
1122 const struct cl_io_slice *ios)
1124 struct vvp_io *vio = cl2vvp_io(env, ios);
1125 struct cl_io *io = ios->cis_io;
1126 struct cl_object *obj = io->ci_obj;
1127 struct inode *inode = vvp_object_inode(obj);
1128 struct ll_inode_info *lli = ll_i2info(inode);
1129 struct cl_fault_io *fio = &io->u.ci_fault;
1130 struct vvp_fault_io *cfio = &vio->u.fault;
1133 struct page *vmpage = NULL;
1134 struct cl_page *page;
1139 down_read(&lli->lli_trunc_sem);
1141 /* offset of the last byte on the page */
1142 offset = cl_offset(obj, fio->ft_index + 1) - 1;
1143 LASSERT(cl_index(obj, offset) == fio->ft_index);
1144 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1148 /* must return locked page */
1149 if (fio->ft_mkwrite) {
1150 LASSERT(cfio->ft_vmpage != NULL);
1151 lock_page(cfio->ft_vmpage);
1153 result = vvp_io_kernel_fault(cfio);
1158 vmpage = cfio->ft_vmpage;
1159 LASSERT(PageLocked(vmpage));
1161 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1162 ll_invalidate_page(vmpage);
1164 size = i_size_read(inode);
1165 /* Though we have already held a cl_lock upon this page, but
1166 * it still can be truncated locally. */
1167 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1168 (page_offset(vmpage) > size))) {
1169 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1171 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1173 GOTO(out, result = +1);
1176 last_index = cl_index(obj, size - 1);
1178 if (fio->ft_mkwrite ) {
1180 * Capture the size while holding the lli_trunc_sem from above
1181 * we want to make sure that we complete the mkwrite action
1182 * while holding this lock. We need to make sure that we are
1183 * not past the end of the file.
1185 if (last_index < fio->ft_index) {
1187 "llite: mkwrite and truncate race happened: "
1188 "%p: 0x%lx 0x%lx\n",
1189 vmpage->mapping,fio->ft_index,last_index);
1191 * We need to return if we are
1192 * passed the end of the file. This will propagate
1193 * up the call stack to ll_page_mkwrite where
1194 * we will return VM_FAULT_NOPAGE. Any non-negative
1195 * value returned here will be silently
1196 * converted to 0. If the vmpage->mapping is null
1197 * the error code would be converted back to ENODATA
1198 * in ll_page_mkwrite0. Thus we return -ENODATA
1199 * to handle both cases
1201 GOTO(out, result = -ENODATA);
1205 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1207 GOTO(out, result = PTR_ERR(page));
1209 /* if page is going to be written, we should add this page into cache
1211 if (fio->ft_mkwrite) {
1212 wait_on_page_writeback(vmpage);
1213 if (!PageDirty(vmpage)) {
1214 struct cl_page_list *plist = &io->ci_queue.c2_qin;
1215 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1218 /* vvp_page_assume() calls wait_on_page_writeback(). */
1219 cl_page_assume(env, io, page);
1221 cl_page_list_init(plist);
1222 cl_page_list_add(plist, page);
1225 if (last_index == vvp_index(vpg))
1226 to = size & ~PAGE_MASK;
1228 /* Do not set Dirty bit here so that in case IO is
1229 * started before the page is really made dirty, we
1230 * still have chance to detect it. */
1231 result = cl_io_commit_async(env, io, plist, 0, to,
1232 mkwrite_commit_callback);
1233 LASSERT(cl_page_is_owned(page, io));
1234 cl_page_list_fini(env, plist);
1238 cl_page_discard(env, io, page);
1239 cl_page_disown(env, io, page);
1241 cl_page_put(env, page);
1243 /* we're in big trouble, what can we do now? */
1244 if (result == -EDQUOT)
1248 cl_page_disown(env, io, page);
1253 * The ft_index is only used in the case of
1254 * a mkwrite action. We need to check
1255 * our assertions are correct, since
1256 * we should have caught this above
1258 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1259 if (fio->ft_index == last_index)
1261 * Last page is mapped partially.
1263 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1265 fio->ft_nob = cl_page_size(obj);
1267 lu_ref_add(&page->cp_reference, "fault", io);
1268 fio->ft_page = page;
1272 /* return unlocked vmpage to avoid deadlocking */
1274 unlock_page(vmpage);
1276 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1281 static void vvp_io_fault_end(const struct lu_env *env,
1282 const struct cl_io_slice *ios)
1284 struct inode *inode = vvp_object_inode(ios->cis_obj);
1285 struct ll_inode_info *lli = ll_i2info(inode);
1287 CLOBINVRNT(env, ios->cis_io->ci_obj,
1288 vvp_object_invariant(ios->cis_io->ci_obj));
1289 up_read(&lli->lli_trunc_sem);
1292 static int vvp_io_fsync_start(const struct lu_env *env,
1293 const struct cl_io_slice *ios)
1295 /* we should mark TOWRITE bit to each dirty page in radix tree to
1296 * verify pages have been written, but this is difficult because of
1301 static int vvp_io_read_ahead(const struct lu_env *env,
1302 const struct cl_io_slice *ios,
1303 pgoff_t start, struct cl_read_ahead *ra)
1308 if (ios->cis_io->ci_type == CIT_READ ||
1309 ios->cis_io->ci_type == CIT_FAULT) {
1310 struct vvp_io *vio = cl2vvp_io(env, ios);
1312 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1313 ra->cra_end = CL_PAGE_EOF;
1314 result = +1; /* no need to call down */
1321 static const struct cl_io_operations vvp_io_ops = {
1324 .cio_fini = vvp_io_fini,
1325 .cio_lock = vvp_io_read_lock,
1326 .cio_start = vvp_io_read_start,
1327 .cio_end = vvp_io_rw_end,
1328 .cio_advance = vvp_io_advance,
1331 .cio_fini = vvp_io_fini,
1332 .cio_iter_init = vvp_io_write_iter_init,
1333 .cio_iter_fini = vvp_io_write_iter_fini,
1334 .cio_lock = vvp_io_write_lock,
1335 .cio_start = vvp_io_write_start,
1336 .cio_end = vvp_io_rw_end,
1337 .cio_advance = vvp_io_advance,
1340 .cio_fini = vvp_io_setattr_fini,
1341 .cio_iter_init = vvp_io_setattr_iter_init,
1342 .cio_lock = vvp_io_setattr_lock,
1343 .cio_start = vvp_io_setattr_start,
1344 .cio_end = vvp_io_setattr_end
1347 .cio_fini = vvp_io_fault_fini,
1348 .cio_iter_init = vvp_io_fault_iter_init,
1349 .cio_lock = vvp_io_fault_lock,
1350 .cio_start = vvp_io_fault_start,
1351 .cio_end = vvp_io_fault_end,
1354 .cio_start = vvp_io_fsync_start,
1355 .cio_fini = vvp_io_fini
1358 .cio_fini = vvp_io_fini
1361 .cio_fini = vvp_io_fini
1364 .cio_read_ahead = vvp_io_read_ahead
1367 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1370 struct vvp_io *vio = vvp_env_io(env);
1371 struct inode *inode = vvp_object_inode(obj);
1374 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1377 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1378 "restore needed %d\n",
1379 PFID(lu_object_fid(&obj->co_lu)),
1380 io->ci_ignore_layout, io->ci_verify_layout,
1381 vio->vui_layout_gen, io->ci_restore_needed);
1383 CL_IO_SLICE_CLEAN(vio, vui_cl);
1384 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1385 vio->vui_ra_valid = false;
1387 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1389 struct ll_inode_info *lli = ll_i2info(inode);
1391 count = io->u.ci_rw.crw_count;
1392 /* "If nbyte is 0, read() will return 0 and have no other
1393 * results." -- Single Unix Spec */
1397 vio->vui_tot_count = count;
1399 /* for read/write, we store the jobid in the inode, and
1400 * it'll be fetched by osc when building RPC.
1402 * it's not accurate if the file is shared by different
1405 lustre_get_jobid(lli->lli_jobid);
1406 } else if (io->ci_type == CIT_SETATTR) {
1407 if (!cl_io_is_trunc(io))
1408 io->ci_lockreq = CILR_MANDATORY;
1411 /* Enqueue layout lock and get layout version. We need to do this
1412 * even for operations requiring to open file, such as read and write,
1413 * because it might not grant layout lock in IT_OPEN. */
1414 if (result == 0 && !io->ci_ignore_layout) {
1415 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1416 if (result == -ENOENT)
1417 /* If the inode on MDS has been removed, but the objects
1418 * on OSTs haven't been destroyed (async unlink), layout
1419 * fetch will return -ENOENT, we'd ingore this error
1420 * and continue with dirty flush. LU-3230. */
1423 CERROR("%s: refresh file layout " DFID " error %d.\n",
1424 ll_get_fsname(inode->i_sb, NULL, 0),
1425 PFID(lu_object_fid(&obj->co_lu)), result);