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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Implementation of cl_io for VVP layer.
38 * Author: Nikita Danilov <nikita.danilov@sun.com>
39 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
42 #define DEBUG_SUBSYSTEM S_LLITE
46 #include "llite_internal.h"
47 #include "vvp_internal.h"
49 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
50 const struct cl_io_slice *slice)
54 vio = container_of(slice, struct vvp_io, vui_cl);
55 LASSERT(vio == vvp_env_io(env));
61 * True, if \a io is a normal io, False for splice_{read,write}
63 static int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
65 struct vvp_io *vio = vvp_env_io(env);
67 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
69 return vio->vui_io_subtype == IO_NORMAL;
73 * For swapping layout. The file's layout may have changed.
74 * To avoid populating pages to a wrong stripe, we have to verify the
75 * correctness of layout. It works because swapping layout processes
76 * have to acquire group lock.
78 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
81 struct ll_inode_info *lli = ll_i2info(inode);
82 struct vvp_io *vio = vvp_env_io(env);
85 switch (io->ci_type) {
88 /* don't need lock here to check lli_layout_gen as we have held
89 * extent lock and GROUP lock has to hold to swap layout */
90 if (ll_layout_version_get(lli) != vio->vui_layout_gen ||
91 OBD_FAIL_CHECK_RESET(OBD_FAIL_LLITE_LOST_LAYOUT, 0)) {
92 io->ci_need_restart = 1;
93 /* this will cause a short read/write */
98 /* fault is okay because we've already had a page. */
106 static void vvp_object_size_lock(struct cl_object *obj)
108 struct inode *inode = vvp_object_inode(obj);
110 ll_inode_size_lock(inode);
111 cl_object_attr_lock(obj);
114 static void vvp_object_size_unlock(struct cl_object *obj)
116 struct inode *inode = vvp_object_inode(obj);
118 cl_object_attr_unlock(obj);
119 ll_inode_size_unlock(inode);
123 * Helper function that if necessary adjusts file size (inode->i_size), when
124 * position at the offset \a pos is accessed. File size can be arbitrary stale
125 * on a Lustre client, but client at least knows KMS. If accessed area is
126 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
128 * Locking: i_size_lock is used to serialize changes to inode size and to
129 * protect consistency between inode size and cl_object
130 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
131 * top-object and sub-objects.
133 static int vvp_prep_size(const struct lu_env *env, struct cl_object *obj,
134 struct cl_io *io, loff_t start, size_t count,
137 struct cl_attr *attr = vvp_env_thread_attr(env);
138 struct inode *inode = vvp_object_inode(obj);
139 loff_t pos = start + count - 1;
144 * Consistency guarantees: following possibilities exist for the
145 * relation between region being accessed and real file size at this
148 * (A): the region is completely inside of the file;
150 * (B-x): x bytes of region are inside of the file, the rest is
153 * (C): the region is completely outside of the file.
155 * This classification is stable under DLM lock already acquired by
156 * the caller, because to change the class, other client has to take
157 * DLM lock conflicting with our lock. Also, any updates to ->i_size
158 * by other threads on this client are serialized by
159 * ll_inode_size_lock(). This guarantees that short reads are handled
160 * correctly in the face of concurrent writes and truncates.
162 vvp_object_size_lock(obj);
163 result = cl_object_attr_get(env, obj, attr);
168 * A glimpse is necessary to determine whether we
169 * return a short read (B) or some zeroes at the end
172 vvp_object_size_unlock(obj);
173 result = cl_glimpse_lock(env, io, inode, obj, 0);
174 if (result == 0 && exceed != NULL) {
175 /* If objective page index exceed end-of-file
176 * page index, return directly. Do not expect
177 * kernel will check such case correctly.
178 * linux-2.6.18-128.1.1 miss to do that.
180 loff_t size = i_size_read(inode);
181 unsigned long cur_index = start >>
184 if ((size == 0 && cur_index != 0) ||
185 (((size - 1) >> PAGE_CACHE_SHIFT) <
193 * region is within kms and, hence, within real file
194 * size (A). We need to increase i_size to cover the
195 * read region so that generic_file_read() will do its
196 * job, but that doesn't mean the kms size is
197 * _correct_, it is only the _minimum_ size. If
198 * someone does a stat they will get the correct size
199 * which will always be >= the kms value here.
202 if (i_size_read(inode) < kms) {
203 i_size_write(inode, kms);
205 DFID" updating i_size "LPU64"\n",
206 PFID(lu_object_fid(&obj->co_lu)),
207 (__u64)i_size_read(inode));
212 vvp_object_size_unlock(obj);
217 /*****************************************************************************
223 static int vvp_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
224 __u32 enqflags, enum cl_lock_mode mode,
225 pgoff_t start, pgoff_t end)
227 struct vvp_io *vio = vvp_env_io(env);
228 struct cl_lock_descr *descr = &vio->vui_link.cill_descr;
229 struct cl_object *obj = io->ci_obj;
231 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
234 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
236 memset(&vio->vui_link, 0, sizeof vio->vui_link);
238 if (vio->vui_fd && (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
239 descr->cld_mode = CLM_GROUP;
240 descr->cld_gid = vio->vui_fd->fd_grouplock.lg_gid;
242 descr->cld_mode = mode;
245 descr->cld_obj = obj;
246 descr->cld_start = start;
247 descr->cld_end = end;
248 descr->cld_enq_flags = enqflags;
250 cl_io_lock_add(env, io, &vio->vui_link);
255 static int vvp_io_one_lock(const struct lu_env *env, struct cl_io *io,
256 __u32 enqflags, enum cl_lock_mode mode,
257 loff_t start, loff_t end)
259 struct cl_object *obj = io->ci_obj;
261 return vvp_io_one_lock_index(env, io, enqflags, mode,
262 cl_index(obj, start), cl_index(obj, end));
265 static int vvp_io_write_iter_init(const struct lu_env *env,
266 const struct cl_io_slice *ios)
268 struct vvp_io *vio = cl2vvp_io(env, ios);
270 cl_page_list_init(&vio->u.write.vui_queue);
271 vio->u.write.vui_written = 0;
272 vio->u.write.vui_from = 0;
273 vio->u.write.vui_to = PAGE_SIZE;
278 static void vvp_io_write_iter_fini(const struct lu_env *env,
279 const struct cl_io_slice *ios)
281 struct vvp_io *vio = cl2vvp_io(env, ios);
283 LASSERT(vio->u.write.vui_queue.pl_nr == 0);
286 static int vvp_io_fault_iter_init(const struct lu_env *env,
287 const struct cl_io_slice *ios)
289 struct vvp_io *vio = cl2vvp_io(env, ios);
290 struct inode *inode = vvp_object_inode(ios->cis_obj);
292 LASSERT(inode == vio->vui_fd->fd_file->f_path.dentry->d_inode);
293 vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime);
298 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
300 struct cl_io *io = ios->cis_io;
301 struct cl_object *obj = io->ci_obj;
302 struct vvp_io *vio = cl2vvp_io(env, ios);
303 struct inode *inode = vvp_object_inode(obj);
305 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
307 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
308 "restore needed %d\n",
309 PFID(lu_object_fid(&obj->co_lu)),
310 io->ci_ignore_layout, io->ci_verify_layout,
311 vio->vui_layout_gen, io->ci_restore_needed);
313 if (io->ci_restore_needed == 1) {
316 /* file was detected release, we need to restore it
317 * before finishing the io
319 rc = ll_layout_restore(inode, 0, OBD_OBJECT_EOF);
320 /* if restore registration failed, no restart,
321 * we will return -ENODATA */
322 /* The layout will change after restore, so we need to
323 * block on layout lock held by the MDT
324 * as MDT will not send new layout in lvb (see LU-3124)
325 * we have to explicitly fetch it, all this will be done
326 * by ll_layout_refresh()
329 io->ci_restore_needed = 0;
330 io->ci_need_restart = 1;
331 io->ci_verify_layout = 1;
333 io->ci_restore_needed = 1;
334 io->ci_need_restart = 0;
335 io->ci_verify_layout = 0;
340 if (!io->ci_ignore_layout && io->ci_verify_layout) {
343 /* check layout version */
344 ll_layout_refresh(inode, &gen);
345 io->ci_need_restart = vio->vui_layout_gen != gen;
346 if (io->ci_need_restart) {
348 DFID" layout changed from %d to %d.\n",
349 PFID(lu_object_fid(&obj->co_lu)),
350 vio->vui_layout_gen, gen);
351 /* today successful restore is the only possible
353 /* restore was done, clear restoring state */
354 ll_file_clear_flag(ll_i2info(vvp_object_inode(obj)),
355 LLIF_FILE_RESTORING);
360 static void vvp_io_fault_fini(const struct lu_env *env,
361 const struct cl_io_slice *ios)
363 struct cl_io *io = ios->cis_io;
364 struct cl_page *page = io->u.ci_fault.ft_page;
366 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
369 lu_ref_del(&page->cp_reference, "fault", io);
370 cl_page_put(env, page);
371 io->u.ci_fault.ft_page = NULL;
373 vvp_io_fini(env, ios);
376 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
379 * we only want to hold PW locks if the mmap() can generate
380 * writes back to the file and that only happens in shared
383 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
388 static int vvp_mmap_locks(const struct lu_env *env,
389 struct vvp_io *vio, struct cl_io *io)
391 struct vvp_thread_info *vti = vvp_env_info(env);
392 struct mm_struct *mm = current->mm;
393 struct vm_area_struct *vma;
394 struct cl_lock_descr *descr = &vti->vti_descr;
395 ldlm_policy_data_t policy;
402 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
404 if (!cl_is_normalio(env, io))
407 if (vio->vui_iov == NULL) /* nfs or loop back device write */
410 /* No MM (e.g. NFS)? No vmas too. */
414 for (seg = 0; seg < vio->vui_nrsegs; seg++) {
415 const struct iovec *iv = &vio->vui_iov[seg];
417 addr = (unsigned long)iv->iov_base;
422 count += addr & ~PAGE_MASK;
425 down_read(&mm->mmap_sem);
426 while((vma = our_vma(mm, addr, count)) != NULL) {
427 struct dentry *de = vma->vm_file->f_path.dentry;
428 struct inode *inode = de->d_inode;
429 int flags = CEF_MUST;
431 if (ll_file_nolock(vma->vm_file)) {
433 * For no lock case is not allowed for mmap
440 * XXX: Required lock mode can be weakened: CIT_WRITE
441 * io only ever reads user level buffer, and CIT_READ
444 policy_from_vma(&policy, vma, addr, count);
445 descr->cld_mode = vvp_mode_from_vma(vma);
446 descr->cld_obj = ll_i2info(inode)->lli_clob;
447 descr->cld_start = cl_index(descr->cld_obj,
448 policy.l_extent.start);
449 descr->cld_end = cl_index(descr->cld_obj,
450 policy.l_extent.end);
451 descr->cld_enq_flags = flags;
452 result = cl_io_lock_alloc_add(env, io, descr);
454 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
455 descr->cld_mode, descr->cld_start,
461 if (vma->vm_end - addr >= count)
464 count -= vma->vm_end - addr;
467 up_read(&mm->mmap_sem);
474 static void vvp_io_advance(const struct lu_env *env,
475 const struct cl_io_slice *ios,
478 struct vvp_io *vio = cl2vvp_io(env, ios);
479 struct cl_io *io = ios->cis_io;
480 struct cl_object *obj = ios->cis_io->ci_obj;
483 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
485 if (!cl_is_normalio(env, io))
488 LASSERT(vio->vui_tot_nrsegs >= vio->vui_nrsegs);
489 LASSERT(vio->vui_tot_count >= nob);
491 /* Restore the iov changed in vvp_io_update_iov() */
492 if (vio->vui_iov_olen > 0) {
493 vio->vui_iov[vio->vui_nrsegs - 1].iov_len = vio->vui_iov_olen;
494 vio->vui_iov_olen = 0;
500 if (iov->iov_len > nob) {
502 iov->iov_base += nob;
508 vio->vui_tot_nrsegs--;
512 vio->vui_tot_count -= nob;
515 static void vvp_io_update_iov(const struct lu_env *env,
516 struct vvp_io *vio, struct cl_io *io)
519 size_t size = io->u.ci_rw.crw_count;
521 vio->vui_iov_olen = 0;
522 if (!cl_is_normalio(env, io) || vio->vui_tot_nrsegs == 0)
525 for (i = 0; i < vio->vui_tot_nrsegs; i++) {
526 struct iovec *iv = &vio->vui_iov[i];
528 if (iv->iov_len < size) {
531 if (iv->iov_len > size) {
532 vio->vui_iov_olen = iv->iov_len;
539 vio->vui_nrsegs = i + 1;
540 LASSERTF(vio->vui_tot_nrsegs >= vio->vui_nrsegs,
541 "tot_nrsegs: %lu, nrsegs: %lu\n",
542 vio->vui_tot_nrsegs, vio->vui_nrsegs);
545 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
546 enum cl_lock_mode mode, loff_t start, loff_t end)
548 struct vvp_io *vio = vvp_env_io(env);
552 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
555 vvp_io_update_iov(env, vio, io);
557 if (io->u.ci_rw.crw_nonblock)
558 ast_flags |= CEF_NONBLOCK;
560 result = vvp_mmap_locks(env, vio, io);
562 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
567 static int vvp_io_read_lock(const struct lu_env *env,
568 const struct cl_io_slice *ios)
570 struct cl_io *io = ios->cis_io;
571 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
575 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
576 rd->crw_pos + rd->crw_count - 1);
580 static int vvp_io_fault_lock(const struct lu_env *env,
581 const struct cl_io_slice *ios)
583 struct cl_io *io = ios->cis_io;
584 struct vvp_io *vio = cl2vvp_io(env, ios);
586 * XXX LDLM_FL_CBPENDING
588 return vvp_io_one_lock_index(env,
590 vvp_mode_from_vma(vio->u.fault.ft_vma),
591 io->u.ci_fault.ft_index,
592 io->u.ci_fault.ft_index);
595 static int vvp_io_write_lock(const struct lu_env *env,
596 const struct cl_io_slice *ios)
598 struct cl_io *io = ios->cis_io;
602 if (io->u.ci_wr.wr_append) {
604 end = OBD_OBJECT_EOF;
606 start = io->u.ci_wr.wr.crw_pos;
607 end = start + io->u.ci_wr.wr.crw_count - 1;
609 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
612 static int vvp_io_setattr_iter_init(const struct lu_env *env,
613 const struct cl_io_slice *ios)
619 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
621 * Handles "lockless io" mode when extent locking is done by server.
623 static int vvp_io_setattr_lock(const struct lu_env *env,
624 const struct cl_io_slice *ios)
626 struct cl_io *io = ios->cis_io;
630 if (cl_io_is_trunc(io)) {
631 new_size = io->u.ci_setattr.sa_attr.lvb_size;
633 enqflags = CEF_DISCARD_DATA;
635 unsigned int valid = io->u.ci_setattr.sa_valid;
637 if (!(valid & TIMES_SET_FLAGS))
640 if ((!(valid & ATTR_MTIME) ||
641 io->u.ci_setattr.sa_attr.lvb_mtime >=
642 io->u.ci_setattr.sa_attr.lvb_ctime) &&
643 (!(valid & ATTR_ATIME) ||
644 io->u.ci_setattr.sa_attr.lvb_atime >=
645 io->u.ci_setattr.sa_attr.lvb_ctime))
651 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
652 new_size, OBD_OBJECT_EOF);
655 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
660 * Only ll_inode_size_lock is taken at this level.
662 ll_inode_size_lock(inode);
663 result = inode_newsize_ok(inode, size);
665 ll_inode_size_unlock(inode);
668 i_size_write(inode, size);
670 ll_truncate_pagecache(inode, size);
671 ll_inode_size_unlock(inode);
675 static int vvp_io_setattr_time(const struct lu_env *env,
676 const struct cl_io_slice *ios)
678 struct cl_io *io = ios->cis_io;
679 struct cl_object *obj = io->ci_obj;
680 struct cl_attr *attr = vvp_env_thread_attr(env);
682 unsigned valid = CAT_CTIME;
684 cl_object_attr_lock(obj);
685 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
686 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
687 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
690 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
691 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
694 result = cl_object_attr_update(env, obj, attr, valid);
695 cl_object_attr_unlock(obj);
700 static int vvp_io_setattr_start(const struct lu_env *env,
701 const struct cl_io_slice *ios)
703 struct cl_io *io = ios->cis_io;
704 struct inode *inode = vvp_object_inode(io->ci_obj);
705 struct ll_inode_info *lli = ll_i2info(inode);
707 mutex_lock(&inode->i_mutex);
708 if (cl_io_is_trunc(io)) {
709 down_write(&lli->lli_trunc_sem);
710 inode_dio_wait(inode);
713 if (io->u.ci_setattr.sa_valid & TIMES_SET_FLAGS)
714 return vvp_io_setattr_time(env, ios);
719 static void vvp_io_setattr_end(const struct lu_env *env,
720 const struct cl_io_slice *ios)
722 struct cl_io *io = ios->cis_io;
723 struct inode *inode = vvp_object_inode(io->ci_obj);
724 struct ll_inode_info *lli = ll_i2info(inode);
726 if (cl_io_is_trunc(io)) {
727 /* Truncate in memory pages - they must be clean pages
728 * because osc has already notified to destroy osc_extents. */
729 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
730 inode_dio_write_done(inode);
731 up_write(&lli->lli_trunc_sem);
733 mutex_unlock(&inode->i_mutex);
736 static void vvp_io_setattr_fini(const struct lu_env *env,
737 const struct cl_io_slice *ios)
739 vvp_io_fini(env, ios);
742 static int vvp_io_read_start(const struct lu_env *env,
743 const struct cl_io_slice *ios)
745 struct vvp_io *vio = cl2vvp_io(env, ios);
746 struct cl_io *io = ios->cis_io;
747 struct cl_object *obj = io->ci_obj;
748 struct inode *inode = vvp_object_inode(obj);
749 struct ll_inode_info *lli = ll_i2info(inode);
750 struct file *file = vio->vui_fd->fd_file;
753 loff_t pos = io->u.ci_rd.rd.crw_pos;
754 long cnt = io->u.ci_rd.rd.crw_count;
755 long tot = vio->vui_tot_count;
758 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
760 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
762 if (vio->vui_io_subtype == IO_NORMAL)
763 down_read(&lli->lli_trunc_sem);
765 if (!can_populate_pages(env, io, inode))
768 result = vvp_prep_size(env, obj, io, pos, tot, &exceed);
771 else if (exceed != 0)
774 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
775 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
776 inode->i_ino, cnt, pos, i_size_read(inode));
778 /* turn off the kernel's read-ahead */
779 vio->vui_fd->fd_file->f_ra.ra_pages = 0;
781 /* initialize read-ahead window once per syscall */
782 if (!vio->vui_ra_valid) {
783 vio->vui_ra_valid = true;
784 vio->vui_ra_start = cl_index(obj, pos);
785 vio->vui_ra_count = cl_index(obj, tot + PAGE_CACHE_SIZE - 1);
791 switch (vio->vui_io_subtype) {
793 LASSERT(vio->vui_iocb->ki_pos == pos);
794 result = generic_file_aio_read(vio->vui_iocb,
795 vio->vui_iov, vio->vui_nrsegs,
796 vio->vui_iocb->ki_pos);
799 result = generic_file_splice_read(file, &pos,
800 vio->u.splice.vui_pipe, cnt,
801 vio->u.splice.vui_flags);
802 /* LU-1109: do splice read stripe by stripe otherwise if it
803 * may make nfsd stuck if this read occupied all internal pipe
808 CERROR("Wrong IO type %u\n", vio->vui_io_subtype);
816 io->ci_nob += result;
817 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, vio->vui_fd,
825 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
826 struct cl_page_list *plist, int from, int to)
828 struct cl_2queue *queue = &io->ci_queue;
829 struct cl_page *page;
830 unsigned int bytes = 0;
834 if (plist->pl_nr == 0)
837 if (from > 0 || to != PAGE_SIZE) {
838 page = cl_page_list_first(plist);
839 if (plist->pl_nr == 1) {
840 cl_page_clip(env, page, from, to);
843 cl_page_clip(env, page, from, PAGE_SIZE);
844 if (to != PAGE_SIZE) {
845 page = cl_page_list_last(plist);
846 cl_page_clip(env, page, 0, to);
851 cl_2queue_init(queue);
852 cl_page_list_splice(plist, &queue->c2_qin);
853 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
855 /* plist is not sorted any more */
856 cl_page_list_splice(&queue->c2_qin, plist);
857 cl_page_list_splice(&queue->c2_qout, plist);
858 cl_2queue_fini(env, queue);
861 /* calculate bytes */
862 bytes = plist->pl_nr << PAGE_SHIFT;
863 bytes -= from + PAGE_SIZE - to;
865 while (plist->pl_nr > 0) {
866 page = cl_page_list_first(plist);
867 cl_page_list_del(env, plist, page);
869 cl_page_clip(env, page, 0, PAGE_SIZE);
871 SetPageUptodate(cl_page_vmpage(page));
872 cl_page_disown(env, io, page);
874 /* held in ll_cl_init() */
875 lu_ref_del(&page->cp_reference, "cl_io", io);
876 cl_page_put(env, page);
880 RETURN(bytes > 0 ? bytes : rc);
883 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
884 struct cl_page *page)
886 struct page *vmpage = page->cp_vmpage;
888 SetPageUptodate(vmpage);
889 set_page_dirty(vmpage);
891 cl_page_disown(env, io, page);
893 /* held in ll_cl_init() */
894 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
895 cl_page_put(env, page);
898 /* make sure the page list is contiguous */
899 static bool page_list_sanity_check(struct cl_object *obj,
900 struct cl_page_list *plist)
902 struct cl_page *page;
903 pgoff_t index = CL_PAGE_EOF;
905 cl_page_list_for_each(page, plist) {
906 struct vvp_page *vpg = cl_object_page_slice(obj, page);
908 if (index == CL_PAGE_EOF) {
909 index = vvp_index(vpg);
914 if (index == vvp_index(vpg))
922 /* Return how many bytes have queued or written */
923 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
925 struct cl_object *obj = io->ci_obj;
926 struct inode *inode = vvp_object_inode(obj);
927 struct vvp_io *vio = vvp_env_io(env);
928 struct cl_page_list *queue = &vio->u.write.vui_queue;
929 struct cl_page *page;
932 unsigned int npages = vio->u.write.vui_queue.pl_nr;
938 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
939 npages, vio->u.write.vui_from, vio->u.write.vui_to);
941 LASSERT(page_list_sanity_check(obj, queue));
943 /* submit IO with async write */
944 rc = cl_io_commit_async(env, io, queue,
945 vio->u.write.vui_from, vio->u.write.vui_to,
946 write_commit_callback);
947 npages -= queue->pl_nr; /* already committed pages */
949 /* calculate how many bytes were written */
950 bytes = npages << PAGE_SHIFT;
953 bytes -= vio->u.write.vui_from;
954 if (queue->pl_nr == 0) /* last page */
955 bytes -= PAGE_SIZE - vio->u.write.vui_to;
956 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
958 vio->u.write.vui_written += bytes;
960 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
961 npages, bytes, vio->u.write.vui_written);
963 /* the first page must have been written. */
964 vio->u.write.vui_from = 0;
966 LASSERT(page_list_sanity_check(obj, queue));
967 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
969 /* out of quota, try sync write */
970 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
971 rc = vvp_io_commit_sync(env, io, queue,
972 vio->u.write.vui_from,
973 vio->u.write.vui_to);
975 vio->u.write.vui_written += rc;
980 /* update inode size */
981 ll_merge_attr(env, inode);
983 /* Now the pages in queue were failed to commit, discard them
984 * unless they were dirtied before. */
985 while (queue->pl_nr > 0) {
986 page = cl_page_list_first(queue);
987 cl_page_list_del(env, queue, page);
989 if (!PageDirty(cl_page_vmpage(page)))
990 cl_page_discard(env, io, page);
992 cl_page_disown(env, io, page);
994 /* held in ll_cl_init() */
995 lu_ref_del(&page->cp_reference, "cl_io", io);
996 cl_page_put(env, page);
998 cl_page_list_fini(env, queue);
1003 static int vvp_io_write_start(const struct lu_env *env,
1004 const struct cl_io_slice *ios)
1006 struct vvp_io *vio = cl2vvp_io(env, ios);
1007 struct cl_io *io = ios->cis_io;
1008 struct cl_object *obj = io->ci_obj;
1009 struct inode *inode = vvp_object_inode(obj);
1010 struct ll_inode_info *lli = ll_i2info(inode);
1012 loff_t pos = io->u.ci_wr.wr.crw_pos;
1013 size_t cnt = io->u.ci_wr.wr.crw_count;
1017 if (vio->vui_io_subtype == IO_NORMAL)
1018 down_read(&lli->lli_trunc_sem);
1020 if (!can_populate_pages(env, io, inode))
1023 if (cl_io_is_append(io)) {
1025 * PARALLEL IO This has to be changed for parallel IO doing
1026 * out-of-order writes.
1028 ll_merge_attr(env, inode);
1029 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
1030 vio->vui_iocb->ki_pos = pos;
1032 LASSERT(vio->vui_iocb->ki_pos == pos);
1035 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
1037 /* The maximum Lustre file size is variable, based on the OST maximum
1038 * object size and number of stripes. This needs another check in
1039 * addition to the VFS checks earlier. */
1040 if (pos + cnt > ll_file_maxbytes(inode)) {
1042 "%s: file "DFID" offset %llu > maxbytes "LPU64"\n",
1043 ll_get_fsname(inode->i_sb, NULL, 0),
1044 PFID(ll_inode2fid(inode)), pos + cnt,
1045 ll_file_maxbytes(inode));
1049 if (vio->vui_iov == NULL) {
1050 /* from a temp io in ll_cl_init(). */
1054 * When using the locked AIO function (generic_file_aio_write())
1055 * testing has shown the inode mutex to be a limiting factor
1056 * with multi-threaded single shared file performance. To get
1057 * around this, we now use the lockless version. To maintain
1058 * consistency, proper locking to protect against writes,
1059 * trucates, etc. is handled in the higher layers of lustre.
1061 result = __generic_file_aio_write(vio->vui_iocb,
1062 vio->vui_iov, vio->vui_nrsegs,
1063 &vio->vui_iocb->ki_pos);
1064 if (result > 0 || result == -EIOCBQUEUED) {
1067 err = generic_write_sync(vio->vui_iocb->ki_filp,
1069 if (err < 0 && result > 0)
1075 result = vvp_io_write_commit(env, io);
1076 if (vio->u.write.vui_written > 0) {
1077 result = vio->u.write.vui_written;
1078 io->ci_nob += result;
1080 CDEBUG(D_VFSTRACE, "write: nob %zd, result: %zd\n",
1081 io->ci_nob, result);
1085 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
1088 io->ci_continue = 0;
1089 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
1090 vio->vui_fd, pos, result, WRITE);
1097 static void vvp_io_rw_end(const struct lu_env *env,
1098 const struct cl_io_slice *ios)
1100 struct vvp_io *vio = cl2vvp_io(env, ios);
1101 struct inode *inode = vvp_object_inode(ios->cis_obj);
1102 struct ll_inode_info *lli = ll_i2info(inode);
1104 if (vio->vui_io_subtype == IO_NORMAL)
1105 up_read(&lli->lli_trunc_sem);
1108 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
1110 struct vm_fault *vmf = cfio->ft_vmf;
1112 cfio->ft_flags = filemap_fault(cfio->ft_vma, vmf);
1113 cfio->ft_flags_valid = 1;
1116 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
1117 vmf->virtual_address);
1118 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
1119 lock_page(vmf->page);
1120 cfio->ft_flags |= VM_FAULT_LOCKED;
1123 cfio->ft_vmpage = vmf->page;
1128 if (cfio->ft_flags & VM_FAULT_SIGBUS) {
1129 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
1133 if (cfio->ft_flags & VM_FAULT_OOM) {
1134 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
1138 if (cfio->ft_flags & VM_FAULT_RETRY)
1141 CERROR("unknown error in page fault %d\n", cfio->ft_flags);
1146 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1147 struct cl_page *page)
1149 set_page_dirty(page->cp_vmpage);
1152 static int vvp_io_fault_start(const struct lu_env *env,
1153 const struct cl_io_slice *ios)
1155 struct vvp_io *vio = cl2vvp_io(env, ios);
1156 struct cl_io *io = ios->cis_io;
1157 struct cl_object *obj = io->ci_obj;
1158 struct inode *inode = vvp_object_inode(obj);
1159 struct ll_inode_info *lli = ll_i2info(inode);
1160 struct cl_fault_io *fio = &io->u.ci_fault;
1161 struct vvp_fault_io *cfio = &vio->u.fault;
1164 struct page *vmpage = NULL;
1165 struct cl_page *page;
1170 if (fio->ft_executable &&
1171 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
1173 " changed while waiting for the page fault lock\n",
1174 PFID(lu_object_fid(&obj->co_lu)));
1176 down_read(&lli->lli_trunc_sem);
1178 /* offset of the last byte on the page */
1179 offset = cl_offset(obj, fio->ft_index + 1) - 1;
1180 LASSERT(cl_index(obj, offset) == fio->ft_index);
1181 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1185 /* must return locked page */
1186 if (fio->ft_mkwrite) {
1187 LASSERT(cfio->ft_vmpage != NULL);
1188 lock_page(cfio->ft_vmpage);
1190 result = vvp_io_kernel_fault(cfio);
1195 vmpage = cfio->ft_vmpage;
1196 LASSERT(PageLocked(vmpage));
1198 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1199 ll_invalidate_page(vmpage);
1201 size = i_size_read(inode);
1202 /* Though we have already held a cl_lock upon this page, but
1203 * it still can be truncated locally. */
1204 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1205 (page_offset(vmpage) > size))) {
1206 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1208 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1210 GOTO(out, result = +1);
1213 last_index = cl_index(obj, size - 1);
1215 if (fio->ft_mkwrite ) {
1217 * Capture the size while holding the lli_trunc_sem from above
1218 * we want to make sure that we complete the mkwrite action
1219 * while holding this lock. We need to make sure that we are
1220 * not past the end of the file.
1222 if (last_index < fio->ft_index) {
1224 "llite: mkwrite and truncate race happened: "
1225 "%p: 0x%lx 0x%lx\n",
1226 vmpage->mapping,fio->ft_index,last_index);
1228 * We need to return if we are
1229 * passed the end of the file. This will propagate
1230 * up the call stack to ll_page_mkwrite where
1231 * we will return VM_FAULT_NOPAGE. Any non-negative
1232 * value returned here will be silently
1233 * converted to 0. If the vmpage->mapping is null
1234 * the error code would be converted back to ENODATA
1235 * in ll_page_mkwrite0. Thus we return -ENODATA
1236 * to handle both cases
1238 GOTO(out, result = -ENODATA);
1242 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1244 GOTO(out, result = PTR_ERR(page));
1246 /* if page is going to be written, we should add this page into cache
1248 if (fio->ft_mkwrite) {
1249 wait_on_page_writeback(vmpage);
1250 if (!PageDirty(vmpage)) {
1251 struct cl_page_list *plist = &io->ci_queue.c2_qin;
1252 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1255 /* vvp_page_assume() calls wait_on_page_writeback(). */
1256 cl_page_assume(env, io, page);
1258 cl_page_list_init(plist);
1259 cl_page_list_add(plist, page);
1262 if (last_index == vvp_index(vpg))
1263 to = size & ~PAGE_MASK;
1265 /* Do not set Dirty bit here so that in case IO is
1266 * started before the page is really made dirty, we
1267 * still have chance to detect it. */
1268 result = cl_io_commit_async(env, io, plist, 0, to,
1269 mkwrite_commit_callback);
1270 LASSERT(cl_page_is_owned(page, io));
1271 cl_page_list_fini(env, plist);
1275 cl_page_discard(env, io, page);
1276 cl_page_disown(env, io, page);
1278 cl_page_put(env, page);
1280 /* we're in big trouble, what can we do now? */
1281 if (result == -EDQUOT)
1285 cl_page_disown(env, io, page);
1290 * The ft_index is only used in the case of
1291 * a mkwrite action. We need to check
1292 * our assertions are correct, since
1293 * we should have caught this above
1295 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1296 if (fio->ft_index == last_index)
1298 * Last page is mapped partially.
1300 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1302 fio->ft_nob = cl_page_size(obj);
1304 lu_ref_add(&page->cp_reference, "fault", io);
1305 fio->ft_page = page;
1309 /* return unlocked vmpage to avoid deadlocking */
1311 unlock_page(vmpage);
1313 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1318 static void vvp_io_fault_end(const struct lu_env *env,
1319 const struct cl_io_slice *ios)
1321 struct inode *inode = vvp_object_inode(ios->cis_obj);
1322 struct ll_inode_info *lli = ll_i2info(inode);
1324 CLOBINVRNT(env, ios->cis_io->ci_obj,
1325 vvp_object_invariant(ios->cis_io->ci_obj));
1326 up_read(&lli->lli_trunc_sem);
1329 static int vvp_io_fsync_start(const struct lu_env *env,
1330 const struct cl_io_slice *ios)
1332 /* we should mark TOWRITE bit to each dirty page in radix tree to
1333 * verify pages have been written, but this is difficult because of
1338 static int vvp_io_read_ahead(const struct lu_env *env,
1339 const struct cl_io_slice *ios,
1340 pgoff_t start, struct cl_read_ahead *ra)
1345 if (ios->cis_io->ci_type == CIT_READ ||
1346 ios->cis_io->ci_type == CIT_FAULT) {
1347 struct vvp_io *vio = cl2vvp_io(env, ios);
1349 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1350 ra->cra_end = CL_PAGE_EOF;
1351 result = +1; /* no need to call down */
1358 static const struct cl_io_operations vvp_io_ops = {
1361 .cio_fini = vvp_io_fini,
1362 .cio_lock = vvp_io_read_lock,
1363 .cio_start = vvp_io_read_start,
1364 .cio_end = vvp_io_rw_end,
1365 .cio_advance = vvp_io_advance,
1368 .cio_fini = vvp_io_fini,
1369 .cio_iter_init = vvp_io_write_iter_init,
1370 .cio_iter_fini = vvp_io_write_iter_fini,
1371 .cio_lock = vvp_io_write_lock,
1372 .cio_start = vvp_io_write_start,
1373 .cio_end = vvp_io_rw_end,
1374 .cio_advance = vvp_io_advance,
1377 .cio_fini = vvp_io_setattr_fini,
1378 .cio_iter_init = vvp_io_setattr_iter_init,
1379 .cio_lock = vvp_io_setattr_lock,
1380 .cio_start = vvp_io_setattr_start,
1381 .cio_end = vvp_io_setattr_end
1384 .cio_fini = vvp_io_fault_fini,
1385 .cio_iter_init = vvp_io_fault_iter_init,
1386 .cio_lock = vvp_io_fault_lock,
1387 .cio_start = vvp_io_fault_start,
1388 .cio_end = vvp_io_fault_end,
1391 .cio_start = vvp_io_fsync_start,
1392 .cio_fini = vvp_io_fini
1395 .cio_fini = vvp_io_fini
1398 .cio_read_ahead = vvp_io_read_ahead
1401 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1404 struct vvp_io *vio = vvp_env_io(env);
1405 struct inode *inode = vvp_object_inode(obj);
1408 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1411 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1412 "restore needed %d\n",
1413 PFID(lu_object_fid(&obj->co_lu)),
1414 io->ci_ignore_layout, io->ci_verify_layout,
1415 vio->vui_layout_gen, io->ci_restore_needed);
1417 CL_IO_SLICE_CLEAN(vio, vui_cl);
1418 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1419 vio->vui_ra_valid = false;
1421 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1423 struct ll_inode_info *lli = ll_i2info(inode);
1425 count = io->u.ci_rw.crw_count;
1426 /* "If nbyte is 0, read() will return 0 and have no other
1427 * results." -- Single Unix Spec */
1431 vio->vui_tot_count = count;
1432 vio->vui_tot_nrsegs = 0;
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 /* ignore layout change for generic CIT_MISC but not for glimpse.
1448 * io context for glimpse must set ci_verify_layout to true,
1449 * see cl_glimpse_size0() for details. */
1450 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1451 io->ci_ignore_layout = 1;
1453 /* Enqueue layout lock and get layout version. We need to do this
1454 * even for operations requiring to open file, such as read and write,
1455 * because it might not grant layout lock in IT_OPEN. */
1456 if (result == 0 && !io->ci_ignore_layout) {
1457 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1458 if (result == -ENOENT)
1459 /* If the inode on MDS has been removed, but the objects
1460 * on OSTs haven't been destroyed (async unlink), layout
1461 * fetch will return -ENOENT, we'd ingore this error
1462 * and continue with dirty flush. LU-3230. */
1465 CERROR("%s: refresh file layout " DFID " error %d.\n",
1466 ll_get_fsname(inode->i_sb, NULL, 0),
1467 PFID(lu_object_fid(&obj->co_lu)), result);