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);
303 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
305 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
306 "need write layout %d, restore needed %d\n",
307 PFID(lu_object_fid(&obj->co_lu)),
308 io->ci_ignore_layout, io->ci_verify_layout,
309 vio->vui_layout_gen, io->ci_need_write_intent,
310 io->ci_restore_needed);
312 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;
338 * dynamic layout change needed, send layout intent
341 if (io->ci_need_write_intent) {
343 loff_t end = OBD_OBJECT_EOF;
345 io->ci_need_write_intent = 0;
347 LASSERT(io->ci_type == CIT_WRITE ||
348 cl_io_is_trunc(io) || cl_io_is_mkwrite(io));
350 if (io->ci_type == CIT_WRITE) {
351 if (!cl_io_is_append(io)) {
352 start = io->u.ci_rw.crw_pos;
353 end = start + io->u.ci_rw.crw_count;
355 } else if (cl_io_is_trunc(io)) {
356 end = io->u.ci_setattr.sa_attr.lvb_size;
357 } else { /* mkwrite */
358 pgoff_t index = io->u.ci_fault.ft_index;
360 start = cl_offset(io->ci_obj, index);
361 end = cl_offset(io->ci_obj, index + 1);
364 CDEBUG(D_VFSTRACE, DFID" type %d [%llx, %llx)\n",
365 PFID(lu_object_fid(&obj->co_lu)), io->ci_type,
367 rc = ll_layout_write_intent(inode, start, end);
370 io->ci_need_restart = 1;
373 if (!io->ci_ignore_layout && io->ci_verify_layout) {
376 /* check layout version */
377 ll_layout_refresh(inode, &gen);
378 io->ci_need_restart = vio->vui_layout_gen != gen;
379 if (io->ci_need_restart) {
381 DFID" layout changed from %d to %d.\n",
382 PFID(lu_object_fid(&obj->co_lu)),
383 vio->vui_layout_gen, gen);
384 /* today successful restore is the only possible
386 /* restore was done, clear restoring state */
387 ll_file_clear_flag(ll_i2info(vvp_object_inode(obj)),
388 LLIF_FILE_RESTORING);
393 static void vvp_io_fault_fini(const struct lu_env *env,
394 const struct cl_io_slice *ios)
396 struct cl_io *io = ios->cis_io;
397 struct cl_page *page = io->u.ci_fault.ft_page;
399 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
402 lu_ref_del(&page->cp_reference, "fault", io);
403 cl_page_put(env, page);
404 io->u.ci_fault.ft_page = NULL;
406 vvp_io_fini(env, ios);
409 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
412 * we only want to hold PW locks if the mmap() can generate
413 * writes back to the file and that only happens in shared
416 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
421 static int vvp_mmap_locks(const struct lu_env *env,
422 struct vvp_io *vio, struct cl_io *io)
424 struct vvp_thread_info *vti = vvp_env_info(env);
425 struct mm_struct *mm = current->mm;
426 struct vm_area_struct *vma;
427 struct cl_lock_descr *descr = &vti->vti_descr;
428 union ldlm_policy_data policy;
434 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
436 if (!cl_is_normalio(env, io))
439 /* nfs or loop back device write */
440 if (vio->vui_iter == NULL)
443 /* No MM (e.g. NFS)? No vmas too. */
447 iov_for_each(iov, i, *(vio->vui_iter)) {
448 unsigned long addr = (unsigned long)iov.iov_base;
449 size_t count = iov.iov_len;
454 count += addr & ~PAGE_MASK;
457 down_read(&mm->mmap_sem);
458 while((vma = our_vma(mm, addr, count)) != NULL) {
459 struct dentry *de = file_dentry(vma->vm_file);
460 struct inode *inode = de->d_inode;
461 int flags = CEF_MUST;
463 if (ll_file_nolock(vma->vm_file)) {
465 * For no lock case is not allowed for mmap
472 * XXX: Required lock mode can be weakened: CIT_WRITE
473 * io only ever reads user level buffer, and CIT_READ
476 policy_from_vma(&policy, vma, addr, count);
477 descr->cld_mode = vvp_mode_from_vma(vma);
478 descr->cld_obj = ll_i2info(inode)->lli_clob;
479 descr->cld_start = cl_index(descr->cld_obj,
480 policy.l_extent.start);
481 descr->cld_end = cl_index(descr->cld_obj,
482 policy.l_extent.end);
483 descr->cld_enq_flags = flags;
484 result = cl_io_lock_alloc_add(env, io, descr);
486 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
487 descr->cld_mode, descr->cld_start,
493 if (vma->vm_end - addr >= count)
496 count -= vma->vm_end - addr;
499 up_read(&mm->mmap_sem);
506 static void vvp_io_advance(const struct lu_env *env,
507 const struct cl_io_slice *ios,
510 struct vvp_io *vio = cl2vvp_io(env, ios);
511 struct cl_io *io = ios->cis_io;
512 struct cl_object *obj = ios->cis_io->ci_obj;
514 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
516 if (!cl_is_normalio(env, io))
519 vio->vui_tot_count -= nob;
520 iov_iter_reexpand(vio->vui_iter, vio->vui_tot_count);
523 static void vvp_io_update_iov(const struct lu_env *env,
524 struct vvp_io *vio, struct cl_io *io)
526 size_t size = io->u.ci_rw.crw_count;
528 if (!cl_is_normalio(env, io) || vio->vui_iter == NULL)
531 iov_iter_truncate(vio->vui_iter, size);
534 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
535 enum cl_lock_mode mode, loff_t start, loff_t end)
537 struct vvp_io *vio = vvp_env_io(env);
541 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
544 vvp_io_update_iov(env, vio, io);
546 if (io->u.ci_rw.crw_nonblock)
547 ast_flags |= CEF_NONBLOCK;
549 result = vvp_mmap_locks(env, vio, io);
551 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
556 static int vvp_io_read_lock(const struct lu_env *env,
557 const struct cl_io_slice *ios)
559 struct cl_io *io = ios->cis_io;
560 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
564 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
565 rd->crw_pos + rd->crw_count - 1);
569 static int vvp_io_fault_lock(const struct lu_env *env,
570 const struct cl_io_slice *ios)
572 struct cl_io *io = ios->cis_io;
573 struct vvp_io *vio = cl2vvp_io(env, ios);
575 * XXX LDLM_FL_CBPENDING
577 return vvp_io_one_lock_index(env,
579 vvp_mode_from_vma(vio->u.fault.ft_vma),
580 io->u.ci_fault.ft_index,
581 io->u.ci_fault.ft_index);
584 static int vvp_io_write_lock(const struct lu_env *env,
585 const struct cl_io_slice *ios)
587 struct cl_io *io = ios->cis_io;
591 if (io->u.ci_wr.wr_append) {
593 end = OBD_OBJECT_EOF;
595 start = io->u.ci_wr.wr.crw_pos;
596 end = start + io->u.ci_wr.wr.crw_count - 1;
598 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
601 static int vvp_io_setattr_iter_init(const struct lu_env *env,
602 const struct cl_io_slice *ios)
608 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
610 * Handles "lockless io" mode when extent locking is done by server.
612 static int vvp_io_setattr_lock(const struct lu_env *env,
613 const struct cl_io_slice *ios)
615 struct cl_io *io = ios->cis_io;
619 if (cl_io_is_trunc(io)) {
620 new_size = io->u.ci_setattr.sa_attr.lvb_size;
622 enqflags = CEF_DISCARD_DATA;
624 unsigned int valid = io->u.ci_setattr.sa_valid;
626 if (!(valid & TIMES_SET_FLAGS))
629 if ((!(valid & ATTR_MTIME) ||
630 io->u.ci_setattr.sa_attr.lvb_mtime >=
631 io->u.ci_setattr.sa_attr.lvb_ctime) &&
632 (!(valid & ATTR_ATIME) ||
633 io->u.ci_setattr.sa_attr.lvb_atime >=
634 io->u.ci_setattr.sa_attr.lvb_ctime))
640 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
641 new_size, OBD_OBJECT_EOF);
644 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
649 * Only ll_inode_size_lock is taken at this level.
651 ll_inode_size_lock(inode);
652 result = inode_newsize_ok(inode, size);
654 ll_inode_size_unlock(inode);
657 i_size_write(inode, size);
659 ll_truncate_pagecache(inode, size);
660 ll_inode_size_unlock(inode);
664 static int vvp_io_setattr_time(const struct lu_env *env,
665 const struct cl_io_slice *ios)
667 struct cl_io *io = ios->cis_io;
668 struct cl_object *obj = io->ci_obj;
669 struct cl_attr *attr = vvp_env_thread_attr(env);
671 unsigned valid = CAT_CTIME;
673 cl_object_attr_lock(obj);
674 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
675 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
676 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
679 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
680 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
683 result = cl_object_attr_update(env, obj, attr, valid);
684 cl_object_attr_unlock(obj);
689 static int vvp_io_setattr_start(const struct lu_env *env,
690 const struct cl_io_slice *ios)
692 struct cl_io *io = ios->cis_io;
693 struct inode *inode = vvp_object_inode(io->ci_obj);
694 struct ll_inode_info *lli = ll_i2info(inode);
696 if (cl_io_is_trunc(io)) {
697 down_write(&lli->lli_trunc_sem);
699 inode_dio_wait(inode);
704 if (io->u.ci_setattr.sa_valid & TIMES_SET_FLAGS)
705 return vvp_io_setattr_time(env, ios);
710 static void vvp_io_setattr_end(const struct lu_env *env,
711 const struct cl_io_slice *ios)
713 struct cl_io *io = ios->cis_io;
714 struct inode *inode = vvp_object_inode(io->ci_obj);
715 struct ll_inode_info *lli = ll_i2info(inode);
717 if (cl_io_is_trunc(io)) {
718 /* Truncate in memory pages - they must be clean pages
719 * because osc has already notified to destroy osc_extents. */
720 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
721 inode_dio_write_done(inode);
723 up_write(&lli->lli_trunc_sem);
729 static void vvp_io_setattr_fini(const struct lu_env *env,
730 const struct cl_io_slice *ios)
732 bool restore_needed = ios->cis_io->ci_restore_needed;
733 struct inode *inode = vvp_object_inode(ios->cis_obj);
735 vvp_io_fini(env, ios);
737 if (restore_needed && !ios->cis_io->ci_restore_needed) {
738 /* restore finished, set data modified flag for HSM */
739 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
743 static int vvp_io_read_start(const struct lu_env *env,
744 const struct cl_io_slice *ios)
746 struct vvp_io *vio = cl2vvp_io(env, ios);
747 struct cl_io *io = ios->cis_io;
748 struct cl_object *obj = io->ci_obj;
749 struct inode *inode = vvp_object_inode(obj);
750 struct ll_inode_info *lli = ll_i2info(inode);
751 struct file *file = vio->vui_fd->fd_file;
754 loff_t pos = io->u.ci_rd.rd.crw_pos;
755 long cnt = io->u.ci_rd.rd.crw_count;
756 long tot = vio->vui_tot_count;
759 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
761 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
763 if (vio->vui_io_subtype == IO_NORMAL)
764 down_read(&lli->lli_trunc_sem);
766 if (!can_populate_pages(env, io, inode))
769 result = vvp_prep_size(env, obj, io, pos, tot, &exceed);
772 else if (exceed != 0)
775 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
776 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
777 inode->i_ino, cnt, pos, i_size_read(inode));
779 /* turn off the kernel's read-ahead */
780 vio->vui_fd->fd_file->f_ra.ra_pages = 0;
782 /* initialize read-ahead window once per syscall */
783 if (!vio->vui_ra_valid) {
784 vio->vui_ra_valid = true;
785 vio->vui_ra_start = cl_index(obj, pos);
786 vio->vui_ra_count = cl_index(obj, tot + PAGE_SIZE - 1);
792 switch (vio->vui_io_subtype) {
794 LASSERT(vio->vui_iocb->ki_pos == pos);
795 result = generic_file_read_iter(vio->vui_iocb, vio->vui_iter);
798 result = generic_file_splice_read(file, &pos,
799 vio->u.splice.vui_pipe, cnt,
800 vio->u.splice.vui_flags);
801 /* LU-1109: do splice read stripe by stripe otherwise if it
802 * may make nfsd stuck if this read occupied all internal pipe
807 CERROR("Wrong IO type %u\n", vio->vui_io_subtype);
815 io->ci_nob += result;
816 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, vio->vui_fd,
824 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
825 struct cl_page_list *plist, int from, int to)
827 struct cl_2queue *queue = &io->ci_queue;
828 struct cl_page *page;
829 unsigned int bytes = 0;
833 if (plist->pl_nr == 0)
836 if (from > 0 || to != PAGE_SIZE) {
837 page = cl_page_list_first(plist);
838 if (plist->pl_nr == 1) {
839 cl_page_clip(env, page, from, to);
842 cl_page_clip(env, page, from, PAGE_SIZE);
843 if (to != PAGE_SIZE) {
844 page = cl_page_list_last(plist);
845 cl_page_clip(env, page, 0, to);
850 cl_2queue_init(queue);
851 cl_page_list_splice(plist, &queue->c2_qin);
852 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
854 /* plist is not sorted any more */
855 cl_page_list_splice(&queue->c2_qin, plist);
856 cl_page_list_splice(&queue->c2_qout, plist);
857 cl_2queue_fini(env, queue);
860 /* calculate bytes */
861 bytes = plist->pl_nr << PAGE_SHIFT;
862 bytes -= from + PAGE_SIZE - to;
864 while (plist->pl_nr > 0) {
865 page = cl_page_list_first(plist);
866 cl_page_list_del(env, plist, page);
868 cl_page_clip(env, page, 0, PAGE_SIZE);
870 SetPageUptodate(cl_page_vmpage(page));
871 cl_page_disown(env, io, page);
873 /* held in ll_cl_init() */
874 lu_ref_del(&page->cp_reference, "cl_io", io);
875 cl_page_put(env, page);
879 RETURN(bytes > 0 ? bytes : rc);
882 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
883 struct cl_page *page)
885 struct page *vmpage = page->cp_vmpage;
887 SetPageUptodate(vmpage);
888 set_page_dirty(vmpage);
890 cl_page_disown(env, io, page);
892 /* held in ll_cl_init() */
893 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
894 cl_page_put(env, page);
897 /* make sure the page list is contiguous */
898 static bool page_list_sanity_check(struct cl_object *obj,
899 struct cl_page_list *plist)
901 struct cl_page *page;
902 pgoff_t index = CL_PAGE_EOF;
904 cl_page_list_for_each(page, plist) {
905 struct vvp_page *vpg = cl_object_page_slice(obj, page);
907 if (index == CL_PAGE_EOF) {
908 index = vvp_index(vpg);
913 if (index == vvp_index(vpg))
921 /* Return how many bytes have queued or written */
922 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
924 struct cl_object *obj = io->ci_obj;
925 struct inode *inode = vvp_object_inode(obj);
926 struct vvp_io *vio = vvp_env_io(env);
927 struct cl_page_list *queue = &vio->u.write.vui_queue;
928 struct cl_page *page;
931 unsigned int npages = vio->u.write.vui_queue.pl_nr;
937 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
938 npages, vio->u.write.vui_from, vio->u.write.vui_to);
940 LASSERT(page_list_sanity_check(obj, queue));
942 /* submit IO with async write */
943 rc = cl_io_commit_async(env, io, queue,
944 vio->u.write.vui_from, vio->u.write.vui_to,
945 write_commit_callback);
946 npages -= queue->pl_nr; /* already committed pages */
948 /* calculate how many bytes were written */
949 bytes = npages << PAGE_SHIFT;
952 bytes -= vio->u.write.vui_from;
953 if (queue->pl_nr == 0) /* last page */
954 bytes -= PAGE_SIZE - vio->u.write.vui_to;
955 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
957 vio->u.write.vui_written += bytes;
959 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
960 npages, bytes, vio->u.write.vui_written);
962 /* the first page must have been written. */
963 vio->u.write.vui_from = 0;
965 LASSERT(page_list_sanity_check(obj, queue));
966 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
968 /* out of quota, try sync write */
969 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
970 rc = vvp_io_commit_sync(env, io, queue,
971 vio->u.write.vui_from,
972 vio->u.write.vui_to);
974 vio->u.write.vui_written += rc;
979 /* update inode size */
980 ll_merge_attr(env, inode);
982 /* Now the pages in queue were failed to commit, discard them
983 * unless they were dirtied before. */
984 while (queue->pl_nr > 0) {
985 page = cl_page_list_first(queue);
986 cl_page_list_del(env, queue, page);
988 if (!PageDirty(cl_page_vmpage(page)))
989 cl_page_discard(env, io, page);
991 cl_page_disown(env, io, page);
993 /* held in ll_cl_init() */
994 lu_ref_del(&page->cp_reference, "cl_io", io);
995 cl_page_put(env, page);
997 cl_page_list_fini(env, queue);
1002 static int vvp_io_write_start(const struct lu_env *env,
1003 const struct cl_io_slice *ios)
1005 struct vvp_io *vio = cl2vvp_io(env, ios);
1006 struct cl_io *io = ios->cis_io;
1007 struct cl_object *obj = io->ci_obj;
1008 struct inode *inode = vvp_object_inode(obj);
1009 struct ll_inode_info *lli = ll_i2info(inode);
1011 loff_t pos = io->u.ci_wr.wr.crw_pos;
1012 size_t cnt = io->u.ci_wr.wr.crw_count;
1016 if (vio->vui_io_subtype == IO_NORMAL)
1017 down_read(&lli->lli_trunc_sem);
1019 if (!can_populate_pages(env, io, inode))
1022 if (cl_io_is_append(io)) {
1024 * PARALLEL IO This has to be changed for parallel IO doing
1025 * out-of-order writes.
1027 ll_merge_attr(env, inode);
1028 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
1029 vio->vui_iocb->ki_pos = pos;
1031 LASSERT(vio->vui_iocb->ki_pos == pos);
1034 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
1036 /* The maximum Lustre file size is variable, based on the OST maximum
1037 * object size and number of stripes. This needs another check in
1038 * addition to the VFS checks earlier. */
1039 if (pos + cnt > ll_file_maxbytes(inode)) {
1041 "%s: file "DFID" offset %llu > maxbytes %llu\n",
1042 ll_get_fsname(inode->i_sb, NULL, 0),
1043 PFID(ll_inode2fid(inode)), pos + cnt,
1044 ll_file_maxbytes(inode));
1048 if (vio->vui_iter == NULL) {
1049 /* from a temp io in ll_cl_init(). */
1053 * When using the locked AIO function (generic_file_aio_write())
1054 * testing has shown the inode mutex to be a limiting factor
1055 * with multi-threaded single shared file performance. To get
1056 * around this, we now use the lockless version. To maintain
1057 * consistency, proper locking to protect against writes,
1058 * trucates, etc. is handled in the higher layers of lustre.
1060 bool lock_node = !IS_NOSEC(inode);
1064 result = __generic_file_write_iter(vio->vui_iocb,
1067 inode_unlock(inode);
1069 if (result > 0 || result == -EIOCBQUEUED) {
1072 err = generic_write_sync(vio->vui_iocb->ki_filp,
1074 if (err < 0 && result > 0)
1080 result = vvp_io_write_commit(env, io);
1081 if (vio->u.write.vui_written > 0) {
1082 result = vio->u.write.vui_written;
1083 io->ci_nob += result;
1085 CDEBUG(D_VFSTRACE, "write: nob %zd, result: %zd\n",
1086 io->ci_nob, result);
1090 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
1093 io->ci_continue = 0;
1094 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
1095 vio->vui_fd, pos, result, WRITE);
1102 static void vvp_io_rw_end(const struct lu_env *env,
1103 const struct cl_io_slice *ios)
1105 struct vvp_io *vio = cl2vvp_io(env, ios);
1106 struct inode *inode = vvp_object_inode(ios->cis_obj);
1107 struct ll_inode_info *lli = ll_i2info(inode);
1109 if (vio->vui_io_subtype == IO_NORMAL)
1110 up_read(&lli->lli_trunc_sem);
1113 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
1115 struct vm_fault *vmf = cfio->ft_vmf;
1117 cfio->ft_flags = filemap_fault(cfio->ft_vma, vmf);
1118 cfio->ft_flags_valid = 1;
1121 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
1122 vmf->virtual_address);
1123 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
1124 lock_page(vmf->page);
1125 cfio->ft_flags |= VM_FAULT_LOCKED;
1128 cfio->ft_vmpage = vmf->page;
1133 if (cfio->ft_flags & VM_FAULT_SIGBUS) {
1134 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
1138 if (cfio->ft_flags & VM_FAULT_OOM) {
1139 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
1143 if (cfio->ft_flags & VM_FAULT_RETRY)
1146 CERROR("unknown error in page fault %d\n", cfio->ft_flags);
1151 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1152 struct cl_page *page)
1154 set_page_dirty(page->cp_vmpage);
1157 static int vvp_io_fault_start(const struct lu_env *env,
1158 const struct cl_io_slice *ios)
1160 struct vvp_io *vio = cl2vvp_io(env, ios);
1161 struct cl_io *io = ios->cis_io;
1162 struct cl_object *obj = io->ci_obj;
1163 struct inode *inode = vvp_object_inode(obj);
1164 struct ll_inode_info *lli = ll_i2info(inode);
1165 struct cl_fault_io *fio = &io->u.ci_fault;
1166 struct vvp_fault_io *cfio = &vio->u.fault;
1169 struct page *vmpage = NULL;
1170 struct cl_page *page;
1175 down_read(&lli->lli_trunc_sem);
1177 /* offset of the last byte on the page */
1178 offset = cl_offset(obj, fio->ft_index + 1) - 1;
1179 LASSERT(cl_index(obj, offset) == fio->ft_index);
1180 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1184 /* must return locked page */
1185 if (fio->ft_mkwrite) {
1186 LASSERT(cfio->ft_vmpage != NULL);
1187 lock_page(cfio->ft_vmpage);
1189 result = vvp_io_kernel_fault(cfio);
1194 vmpage = cfio->ft_vmpage;
1195 LASSERT(PageLocked(vmpage));
1197 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1198 ll_invalidate_page(vmpage);
1200 size = i_size_read(inode);
1201 /* Though we have already held a cl_lock upon this page, but
1202 * it still can be truncated locally. */
1203 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1204 (page_offset(vmpage) > size))) {
1205 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1207 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1209 GOTO(out, result = +1);
1212 last_index = cl_index(obj, size - 1);
1214 if (fio->ft_mkwrite ) {
1216 * Capture the size while holding the lli_trunc_sem from above
1217 * we want to make sure that we complete the mkwrite action
1218 * while holding this lock. We need to make sure that we are
1219 * not past the end of the file.
1221 if (last_index < fio->ft_index) {
1223 "llite: mkwrite and truncate race happened: "
1224 "%p: 0x%lx 0x%lx\n",
1225 vmpage->mapping,fio->ft_index,last_index);
1227 * We need to return if we are
1228 * passed the end of the file. This will propagate
1229 * up the call stack to ll_page_mkwrite where
1230 * we will return VM_FAULT_NOPAGE. Any non-negative
1231 * value returned here will be silently
1232 * converted to 0. If the vmpage->mapping is null
1233 * the error code would be converted back to ENODATA
1234 * in ll_page_mkwrite0. Thus we return -ENODATA
1235 * to handle both cases
1237 GOTO(out, result = -ENODATA);
1241 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1243 GOTO(out, result = PTR_ERR(page));
1245 /* if page is going to be written, we should add this page into cache
1247 if (fio->ft_mkwrite) {
1248 wait_on_page_writeback(vmpage);
1249 if (!PageDirty(vmpage)) {
1250 struct cl_page_list *plist = &io->ci_queue.c2_qin;
1251 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1254 /* vvp_page_assume() calls wait_on_page_writeback(). */
1255 cl_page_assume(env, io, page);
1257 cl_page_list_init(plist);
1258 cl_page_list_add(plist, page);
1261 if (last_index == vvp_index(vpg))
1262 to = size & ~PAGE_MASK;
1264 /* Do not set Dirty bit here so that in case IO is
1265 * started before the page is really made dirty, we
1266 * still have chance to detect it. */
1267 result = cl_io_commit_async(env, io, plist, 0, to,
1268 mkwrite_commit_callback);
1269 LASSERT(cl_page_is_owned(page, io));
1270 cl_page_list_fini(env, plist);
1274 cl_page_discard(env, io, page);
1275 cl_page_disown(env, io, page);
1277 cl_page_put(env, page);
1279 /* we're in big trouble, what can we do now? */
1280 if (result == -EDQUOT)
1284 cl_page_disown(env, io, page);
1289 * The ft_index is only used in the case of
1290 * a mkwrite action. We need to check
1291 * our assertions are correct, since
1292 * we should have caught this above
1294 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1295 if (fio->ft_index == last_index)
1297 * Last page is mapped partially.
1299 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1301 fio->ft_nob = cl_page_size(obj);
1303 lu_ref_add(&page->cp_reference, "fault", io);
1304 fio->ft_page = page;
1308 /* return unlocked vmpage to avoid deadlocking */
1310 unlock_page(vmpage);
1312 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1317 static void vvp_io_fault_end(const struct lu_env *env,
1318 const struct cl_io_slice *ios)
1320 struct inode *inode = vvp_object_inode(ios->cis_obj);
1321 struct ll_inode_info *lli = ll_i2info(inode);
1323 CLOBINVRNT(env, ios->cis_io->ci_obj,
1324 vvp_object_invariant(ios->cis_io->ci_obj));
1325 up_read(&lli->lli_trunc_sem);
1328 static int vvp_io_fsync_start(const struct lu_env *env,
1329 const struct cl_io_slice *ios)
1331 /* we should mark TOWRITE bit to each dirty page in radix tree to
1332 * verify pages have been written, but this is difficult because of
1337 static int vvp_io_read_ahead(const struct lu_env *env,
1338 const struct cl_io_slice *ios,
1339 pgoff_t start, struct cl_read_ahead *ra)
1344 if (ios->cis_io->ci_type == CIT_READ ||
1345 ios->cis_io->ci_type == CIT_FAULT) {
1346 struct vvp_io *vio = cl2vvp_io(env, ios);
1348 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1349 ra->cra_end = CL_PAGE_EOF;
1350 result = +1; /* no need to call down */
1357 static const struct cl_io_operations vvp_io_ops = {
1360 .cio_fini = vvp_io_fini,
1361 .cio_lock = vvp_io_read_lock,
1362 .cio_start = vvp_io_read_start,
1363 .cio_end = vvp_io_rw_end,
1364 .cio_advance = vvp_io_advance,
1367 .cio_fini = vvp_io_fini,
1368 .cio_iter_init = vvp_io_write_iter_init,
1369 .cio_iter_fini = vvp_io_write_iter_fini,
1370 .cio_lock = vvp_io_write_lock,
1371 .cio_start = vvp_io_write_start,
1372 .cio_end = vvp_io_rw_end,
1373 .cio_advance = vvp_io_advance,
1376 .cio_fini = vvp_io_setattr_fini,
1377 .cio_iter_init = vvp_io_setattr_iter_init,
1378 .cio_lock = vvp_io_setattr_lock,
1379 .cio_start = vvp_io_setattr_start,
1380 .cio_end = vvp_io_setattr_end
1383 .cio_fini = vvp_io_fault_fini,
1384 .cio_iter_init = vvp_io_fault_iter_init,
1385 .cio_lock = vvp_io_fault_lock,
1386 .cio_start = vvp_io_fault_start,
1387 .cio_end = vvp_io_fault_end,
1390 .cio_start = vvp_io_fsync_start,
1391 .cio_fini = vvp_io_fini
1394 .cio_fini = vvp_io_fini
1397 .cio_fini = vvp_io_fini
1400 .cio_read_ahead = vvp_io_read_ahead
1403 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1406 struct vvp_io *vio = vvp_env_io(env);
1407 struct inode *inode = vvp_object_inode(obj);
1410 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1413 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1414 "restore needed %d\n",
1415 PFID(lu_object_fid(&obj->co_lu)),
1416 io->ci_ignore_layout, io->ci_verify_layout,
1417 vio->vui_layout_gen, io->ci_restore_needed);
1419 CL_IO_SLICE_CLEAN(vio, vui_cl);
1420 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1421 vio->vui_ra_valid = false;
1423 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1425 struct ll_inode_info *lli = ll_i2info(inode);
1427 count = io->u.ci_rw.crw_count;
1428 /* "If nbyte is 0, read() will return 0 and have no other
1429 * results." -- Single Unix Spec */
1433 vio->vui_tot_count = count;
1435 /* for read/write, we store the jobid in the inode, and
1436 * it'll be fetched by osc when building RPC.
1438 * it's not accurate if the file is shared by different
1441 lustre_get_jobid(lli->lli_jobid);
1442 } else if (io->ci_type == CIT_SETATTR) {
1443 if (!cl_io_is_trunc(io))
1444 io->ci_lockreq = CILR_MANDATORY;
1447 /* Enqueue layout lock and get layout version. We need to do this
1448 * even for operations requiring to open file, such as read and write,
1449 * because it might not grant layout lock in IT_OPEN. */
1450 if (result == 0 && !io->ci_ignore_layout) {
1451 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1452 if (result == -ENOENT)
1453 /* If the inode on MDS has been removed, but the objects
1454 * on OSTs haven't been destroyed (async unlink), layout
1455 * fetch will return -ENOENT, we'd ingore this error
1456 * and continue with dirty flush. LU-3230. */
1459 CERROR("%s: refresh file layout " DFID " error %d.\n",
1460 ll_get_fsname(inode->i_sb, NULL, 0),
1461 PFID(lu_object_fid(&obj->co_lu)), result);