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 io->ci_need_restart = 1;
92 /* this will return application a short read/write */
97 /* fault is okay because we've already had a page. */
105 static void vvp_object_size_lock(struct cl_object *obj)
107 struct inode *inode = vvp_object_inode(obj);
109 ll_inode_size_lock(inode);
110 cl_object_attr_lock(obj);
113 static void vvp_object_size_unlock(struct cl_object *obj)
115 struct inode *inode = vvp_object_inode(obj);
117 cl_object_attr_unlock(obj);
118 ll_inode_size_unlock(inode);
122 * Helper function that if necessary adjusts file size (inode->i_size), when
123 * position at the offset \a pos is accessed. File size can be arbitrary stale
124 * on a Lustre client, but client at least knows KMS. If accessed area is
125 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
127 * Locking: i_size_lock is used to serialize changes to inode size and to
128 * protect consistency between inode size and cl_object
129 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
130 * top-object and sub-objects.
132 static int vvp_prep_size(const struct lu_env *env, struct cl_object *obj,
133 struct cl_io *io, loff_t start, size_t count,
136 struct cl_attr *attr = vvp_env_thread_attr(env);
137 struct inode *inode = vvp_object_inode(obj);
138 loff_t pos = start + count - 1;
143 * Consistency guarantees: following possibilities exist for the
144 * relation between region being accessed and real file size at this
147 * (A): the region is completely inside of the file;
149 * (B-x): x bytes of region are inside of the file, the rest is
152 * (C): the region is completely outside of the file.
154 * This classification is stable under DLM lock already acquired by
155 * the caller, because to change the class, other client has to take
156 * DLM lock conflicting with our lock. Also, any updates to ->i_size
157 * by other threads on this client are serialized by
158 * ll_inode_size_lock(). This guarantees that short reads are handled
159 * correctly in the face of concurrent writes and truncates.
161 vvp_object_size_lock(obj);
162 result = cl_object_attr_get(env, obj, attr);
167 * A glimpse is necessary to determine whether we
168 * return a short read (B) or some zeroes at the end
171 vvp_object_size_unlock(obj);
172 result = cl_glimpse_lock(env, io, inode, obj, 0);
173 if (result == 0 && exceed != NULL) {
174 /* If objective page index exceed end-of-file
175 * page index, return directly. Do not expect
176 * kernel will check such case correctly.
177 * linux-2.6.18-128.1.1 miss to do that.
179 loff_t size = i_size_read(inode);
180 unsigned long cur_index = start >>
183 if ((size == 0 && cur_index != 0) ||
184 (((size - 1) >> PAGE_CACHE_SHIFT) <
192 * region is within kms and, hence, within real file
193 * size (A). We need to increase i_size to cover the
194 * read region so that generic_file_read() will do its
195 * job, but that doesn't mean the kms size is
196 * _correct_, it is only the _minimum_ size. If
197 * someone does a stat they will get the correct size
198 * which will always be >= the kms value here.
201 if (i_size_read(inode) < kms) {
202 i_size_write(inode, kms);
204 DFID" updating i_size "LPU64"\n",
205 PFID(lu_object_fid(&obj->co_lu)),
206 (__u64)i_size_read(inode));
211 vvp_object_size_unlock(obj);
216 /*****************************************************************************
222 static int vvp_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
223 __u32 enqflags, enum cl_lock_mode mode,
224 pgoff_t start, pgoff_t end)
226 struct vvp_io *vio = vvp_env_io(env);
227 struct cl_lock_descr *descr = &vio->vui_link.cill_descr;
228 struct cl_object *obj = io->ci_obj;
230 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
233 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
235 memset(&vio->vui_link, 0, sizeof vio->vui_link);
237 if (vio->vui_fd && (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
238 descr->cld_mode = CLM_GROUP;
239 descr->cld_gid = vio->vui_fd->fd_grouplock.lg_gid;
241 descr->cld_mode = mode;
244 descr->cld_obj = obj;
245 descr->cld_start = start;
246 descr->cld_end = end;
247 descr->cld_enq_flags = enqflags;
249 cl_io_lock_add(env, io, &vio->vui_link);
254 static int vvp_io_one_lock(const struct lu_env *env, struct cl_io *io,
255 __u32 enqflags, enum cl_lock_mode mode,
256 loff_t start, loff_t end)
258 struct cl_object *obj = io->ci_obj;
260 return vvp_io_one_lock_index(env, io, enqflags, mode,
261 cl_index(obj, start), cl_index(obj, end));
264 static int vvp_io_write_iter_init(const struct lu_env *env,
265 const struct cl_io_slice *ios)
267 struct vvp_io *vio = cl2vvp_io(env, ios);
269 cl_page_list_init(&vio->u.write.vui_queue);
270 vio->u.write.vui_written = 0;
271 vio->u.write.vui_from = 0;
272 vio->u.write.vui_to = PAGE_SIZE;
277 static void vvp_io_write_iter_fini(const struct lu_env *env,
278 const struct cl_io_slice *ios)
280 struct vvp_io *vio = cl2vvp_io(env, ios);
282 LASSERT(vio->u.write.vui_queue.pl_nr == 0);
285 static int vvp_io_fault_iter_init(const struct lu_env *env,
286 const struct cl_io_slice *ios)
288 struct vvp_io *vio = cl2vvp_io(env, ios);
289 struct inode *inode = vvp_object_inode(ios->cis_obj);
291 LASSERT(inode == vio->vui_fd->fd_file->f_dentry->d_inode);
292 vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime);
297 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
299 struct cl_io *io = ios->cis_io;
300 struct cl_object *obj = io->ci_obj;
301 struct vvp_io *vio = cl2vvp_io(env, ios);
302 struct inode *inode = vvp_object_inode(obj);
304 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
306 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
307 "restore needed %d\n",
308 PFID(lu_object_fid(&obj->co_lu)),
309 io->ci_ignore_layout, io->ci_verify_layout,
310 vio->vui_layout_gen, io->ci_restore_needed);
312 if (io->ci_restore_needed == 1) {
315 /* file was detected release, we need to restore it
316 * before finishing the io
318 rc = ll_layout_restore(inode, 0, OBD_OBJECT_EOF);
319 /* if restore registration failed, no restart,
320 * we will return -ENODATA */
321 /* The layout will change after restore, so we need to
322 * block on layout lock hold by the MDT
323 * as MDT will not send new layout in lvb (see LU-3124)
324 * we have to explicitly fetch it, all this will be done
325 * by ll_layout_refresh()
328 io->ci_restore_needed = 0;
329 io->ci_need_restart = 1;
330 io->ci_verify_layout = 1;
332 io->ci_restore_needed = 1;
333 io->ci_need_restart = 0;
334 io->ci_verify_layout = 0;
339 if (!io->ci_ignore_layout && io->ci_verify_layout) {
342 /* check layout version */
343 ll_layout_refresh(inode, &gen);
344 io->ci_need_restart = vio->vui_layout_gen != gen;
345 if (io->ci_need_restart) {
347 DFID" layout changed from %d to %d.\n",
348 PFID(lu_object_fid(&obj->co_lu)),
349 vio->vui_layout_gen, gen);
350 /* today successful restore is the only possible
352 /* restore was done, clear restoring state */
353 ll_i2info(vvp_object_inode(obj))->lli_flags &=
354 ~LLIF_FILE_RESTORING;
359 static void vvp_io_fault_fini(const struct lu_env *env,
360 const struct cl_io_slice *ios)
362 struct cl_io *io = ios->cis_io;
363 struct cl_page *page = io->u.ci_fault.ft_page;
365 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
368 lu_ref_del(&page->cp_reference, "fault", io);
369 cl_page_put(env, page);
370 io->u.ci_fault.ft_page = NULL;
372 vvp_io_fini(env, ios);
375 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
378 * we only want to hold PW locks if the mmap() can generate
379 * writes back to the file and that only happens in shared
382 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
387 static int vvp_mmap_locks(const struct lu_env *env,
388 struct vvp_io *vio, struct cl_io *io)
390 struct vvp_thread_info *vti = vvp_env_info(env);
391 struct mm_struct *mm = current->mm;
392 struct vm_area_struct *vma;
393 struct cl_lock_descr *descr = &vti->vti_descr;
394 ldlm_policy_data_t policy;
401 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
403 if (!cl_is_normalio(env, io))
406 if (vio->vui_iov == NULL) /* nfs or loop back device write */
409 /* No MM (e.g. NFS)? No vmas too. */
413 for (seg = 0; seg < vio->vui_nrsegs; seg++) {
414 const struct iovec *iv = &vio->vui_iov[seg];
416 addr = (unsigned long)iv->iov_base;
421 count += addr & (~CFS_PAGE_MASK);
422 addr &= CFS_PAGE_MASK;
424 down_read(&mm->mmap_sem);
425 while((vma = our_vma(mm, addr, count)) != NULL) {
426 struct inode *inode = vma->vm_file->f_dentry->d_inode;
427 int flags = CEF_MUST;
429 if (ll_file_nolock(vma->vm_file)) {
431 * For no lock case is not allowed for mmap
438 * XXX: Required lock mode can be weakened: CIT_WRITE
439 * io only ever reads user level buffer, and CIT_READ
442 policy_from_vma(&policy, vma, addr, count);
443 descr->cld_mode = vvp_mode_from_vma(vma);
444 descr->cld_obj = ll_i2info(inode)->lli_clob;
445 descr->cld_start = cl_index(descr->cld_obj,
446 policy.l_extent.start);
447 descr->cld_end = cl_index(descr->cld_obj,
448 policy.l_extent.end);
449 descr->cld_enq_flags = flags;
450 result = cl_io_lock_alloc_add(env, io, descr);
452 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
453 descr->cld_mode, descr->cld_start,
459 if (vma->vm_end - addr >= count)
462 count -= vma->vm_end - addr;
465 up_read(&mm->mmap_sem);
472 static void vvp_io_advance(const struct lu_env *env,
473 const struct cl_io_slice *ios,
476 struct vvp_io *vio = cl2vvp_io(env, ios);
477 struct cl_io *io = ios->cis_io;
478 struct cl_object *obj = ios->cis_io->ci_obj;
480 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
482 if (!cl_is_normalio(env, io))
485 LASSERT(vio->vui_tot_nrsegs >= vio->vui_nrsegs);
486 LASSERT(vio->vui_tot_count >= nob);
488 vio->vui_iov += vio->vui_nrsegs;
489 vio->vui_tot_nrsegs -= vio->vui_nrsegs;
490 vio->vui_tot_count -= nob;
493 if (vio->vui_iov_olen > 0) {
497 vio->vui_tot_nrsegs++;
498 iv = &vio->vui_iov[0];
499 if (io->ci_continue) {
500 iv->iov_base += iv->iov_len;
501 LASSERT(vio->vui_iov_olen > iv->iov_len);
502 iv->iov_len = vio->vui_iov_olen - iv->iov_len;
504 /* restore the iov_len, in case of restart io. */
505 iv->iov_len = vio->vui_iov_olen;
507 vio->vui_iov_olen = 0;
511 static void vvp_io_update_iov(const struct lu_env *env,
512 struct vvp_io *vio, struct cl_io *io)
515 size_t size = io->u.ci_rw.crw_count;
517 vio->vui_iov_olen = 0;
518 if (!cl_is_normalio(env, io) || vio->vui_tot_nrsegs == 0)
521 for (i = 0; i < vio->vui_tot_nrsegs; i++) {
522 struct iovec *iv = &vio->vui_iov[i];
524 if (iv->iov_len < size) {
527 if (iv->iov_len > size) {
528 vio->vui_iov_olen = iv->iov_len;
535 vio->vui_nrsegs = i + 1;
536 LASSERTF(vio->vui_tot_nrsegs >= vio->vui_nrsegs,
537 "tot_nrsegs: %lu, nrsegs: %lu\n",
538 vio->vui_tot_nrsegs, vio->vui_nrsegs);
541 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
542 enum cl_lock_mode mode, loff_t start, loff_t end)
544 struct vvp_io *vio = vvp_env_io(env);
548 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
551 vvp_io_update_iov(env, vio, io);
553 if (io->u.ci_rw.crw_nonblock)
554 ast_flags |= CEF_NONBLOCK;
556 result = vvp_mmap_locks(env, vio, io);
558 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
563 static int vvp_io_read_lock(const struct lu_env *env,
564 const struct cl_io_slice *ios)
566 struct cl_io *io = ios->cis_io;
567 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
571 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
572 rd->crw_pos + rd->crw_count - 1);
576 static int vvp_io_fault_lock(const struct lu_env *env,
577 const struct cl_io_slice *ios)
579 struct cl_io *io = ios->cis_io;
580 struct vvp_io *vio = cl2vvp_io(env, ios);
582 * XXX LDLM_FL_CBPENDING
584 return vvp_io_one_lock_index(env,
586 vvp_mode_from_vma(vio->u.fault.ft_vma),
587 io->u.ci_fault.ft_index,
588 io->u.ci_fault.ft_index);
591 static int vvp_io_write_lock(const struct lu_env *env,
592 const struct cl_io_slice *ios)
594 struct cl_io *io = ios->cis_io;
598 if (io->u.ci_wr.wr_append) {
600 end = OBD_OBJECT_EOF;
602 start = io->u.ci_wr.wr.crw_pos;
603 end = start + io->u.ci_wr.wr.crw_count - 1;
605 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
608 static int vvp_io_setattr_iter_init(const struct lu_env *env,
609 const struct cl_io_slice *ios)
615 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
617 * Handles "lockless io" mode when extent locking is done by server.
619 static int vvp_io_setattr_lock(const struct lu_env *env,
620 const struct cl_io_slice *ios)
622 struct cl_io *io = ios->cis_io;
626 if (cl_io_is_trunc(io)) {
627 new_size = io->u.ci_setattr.sa_attr.lvb_size;
629 enqflags = CEF_DISCARD_DATA;
631 unsigned int valid = io->u.ci_setattr.sa_valid;
633 if (!(valid & TIMES_SET_FLAGS))
636 if ((!(valid & ATTR_MTIME) ||
637 io->u.ci_setattr.sa_attr.lvb_mtime >=
638 io->u.ci_setattr.sa_attr.lvb_ctime) &&
639 (!(valid & ATTR_ATIME) ||
640 io->u.ci_setattr.sa_attr.lvb_atime >=
641 io->u.ci_setattr.sa_attr.lvb_ctime))
647 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
648 new_size, OBD_OBJECT_EOF);
651 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
656 * Only ll_inode_size_lock is taken at this level.
658 ll_inode_size_lock(inode);
659 result = inode_newsize_ok(inode, size);
661 ll_inode_size_unlock(inode);
664 i_size_write(inode, size);
666 ll_truncate_pagecache(inode, size);
667 ll_inode_size_unlock(inode);
671 static int vvp_io_setattr_trunc(const struct lu_env *env,
672 const struct cl_io_slice *ios,
673 struct inode *inode, loff_t size)
675 inode_dio_wait(inode);
679 static int vvp_io_setattr_time(const struct lu_env *env,
680 const struct cl_io_slice *ios)
682 struct cl_io *io = ios->cis_io;
683 struct cl_object *obj = io->ci_obj;
684 struct cl_attr *attr = vvp_env_thread_attr(env);
686 unsigned valid = CAT_CTIME;
688 cl_object_attr_lock(obj);
689 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
690 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
691 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
694 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
695 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
698 result = cl_object_attr_update(env, obj, attr, valid);
699 cl_object_attr_unlock(obj);
704 static int vvp_io_setattr_start(const struct lu_env *env,
705 const struct cl_io_slice *ios)
707 struct cl_io *io = ios->cis_io;
708 struct inode *inode = vvp_object_inode(io->ci_obj);
711 mutex_lock(&inode->i_mutex);
712 if (cl_io_is_trunc(io))
713 result = vvp_io_setattr_trunc(env, ios, inode,
714 io->u.ci_setattr.sa_attr.lvb_size);
715 if (result == 0 && io->u.ci_setattr.sa_valid & TIMES_SET_FLAGS)
716 result = vvp_io_setattr_time(env, ios);
720 static void vvp_io_setattr_end(const struct lu_env *env,
721 const struct cl_io_slice *ios)
723 struct cl_io *io = ios->cis_io;
724 struct inode *inode = vvp_object_inode(io->ci_obj);
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);
732 mutex_unlock(&inode->i_mutex);
735 static void vvp_io_setattr_fini(const struct lu_env *env,
736 const struct cl_io_slice *ios)
738 vvp_io_fini(env, ios);
741 static int vvp_io_read_start(const struct lu_env *env,
742 const struct cl_io_slice *ios)
744 struct vvp_io *vio = cl2vvp_io(env, ios);
745 struct cl_io *io = ios->cis_io;
746 struct cl_object *obj = io->ci_obj;
747 struct inode *inode = vvp_object_inode(obj);
748 struct file *file = vio->vui_fd->fd_file;
751 loff_t pos = io->u.ci_rd.rd.crw_pos;
752 long cnt = io->u.ci_rd.rd.crw_count;
753 long tot = vio->vui_tot_count;
756 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
758 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
760 if (!can_populate_pages(env, io, inode))
763 result = vvp_prep_size(env, obj, io, pos, tot, &exceed);
766 else if (exceed != 0)
769 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
770 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
771 inode->i_ino, cnt, pos, i_size_read(inode));
773 /* turn off the kernel's read-ahead */
774 vio->vui_fd->fd_file->f_ra.ra_pages = 0;
776 /* initialize read-ahead window once per syscall */
777 if (!vio->vui_ra_valid) {
778 vio->vui_ra_valid = true;
779 vio->vui_ra_start = cl_index(obj, pos);
780 vio->vui_ra_count = cl_index(obj, tot + PAGE_CACHE_SIZE - 1);
786 switch (vio->vui_io_subtype) {
788 LASSERT(vio->vui_iocb->ki_pos == pos);
789 result = generic_file_aio_read(vio->vui_iocb,
790 vio->vui_iov, vio->vui_nrsegs,
791 vio->vui_iocb->ki_pos);
794 result = generic_file_splice_read(file, &pos,
795 vio->u.splice.vui_pipe, cnt,
796 vio->u.splice.vui_flags);
797 /* LU-1109: do splice read stripe by stripe otherwise if it
798 * may make nfsd stuck if this read occupied all internal pipe
803 CERROR("Wrong IO type %u\n", vio->vui_io_subtype);
811 io->ci_nob += result;
812 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, vio->vui_fd,
820 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
821 struct cl_page_list *plist, int from, int to)
823 struct cl_2queue *queue = &io->ci_queue;
824 struct cl_page *page;
825 unsigned int bytes = 0;
829 if (plist->pl_nr == 0)
832 if (from > 0 || to != PAGE_SIZE) {
833 page = cl_page_list_first(plist);
834 if (plist->pl_nr == 1) {
835 cl_page_clip(env, page, from, to);
838 cl_page_clip(env, page, from, PAGE_SIZE);
839 if (to != PAGE_SIZE) {
840 page = cl_page_list_last(plist);
841 cl_page_clip(env, page, 0, to);
846 cl_2queue_init(queue);
847 cl_page_list_splice(plist, &queue->c2_qin);
848 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
850 /* plist is not sorted any more */
851 cl_page_list_splice(&queue->c2_qin, plist);
852 cl_page_list_splice(&queue->c2_qout, plist);
853 cl_2queue_fini(env, queue);
856 /* calculate bytes */
857 bytes = plist->pl_nr << PAGE_SHIFT;
858 bytes -= from + PAGE_SIZE - to;
860 while (plist->pl_nr > 0) {
861 page = cl_page_list_first(plist);
862 cl_page_list_del(env, plist, page);
864 cl_page_clip(env, page, 0, PAGE_SIZE);
866 SetPageUptodate(cl_page_vmpage(page));
867 cl_page_disown(env, io, page);
869 /* held in ll_cl_init() */
870 lu_ref_del(&page->cp_reference, "cl_io", io);
871 cl_page_put(env, page);
875 RETURN(bytes > 0 ? bytes : rc);
878 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
879 struct cl_page *page)
881 struct page *vmpage = page->cp_vmpage;
883 SetPageUptodate(vmpage);
884 set_page_dirty(vmpage);
886 cl_page_disown(env, io, page);
888 /* held in ll_cl_init() */
889 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
890 cl_page_put(env, page);
893 /* make sure the page list is contiguous */
894 static bool page_list_sanity_check(struct cl_object *obj,
895 struct cl_page_list *plist)
897 struct cl_page *page;
898 pgoff_t index = CL_PAGE_EOF;
900 cl_page_list_for_each(page, plist) {
901 struct vvp_page *vpg = cl_object_page_slice(obj, page);
903 if (index == CL_PAGE_EOF) {
904 index = vvp_index(vpg);
909 if (index == vvp_index(vpg))
917 /* Return how many bytes have queued or written */
918 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
920 struct cl_object *obj = io->ci_obj;
921 struct inode *inode = vvp_object_inode(obj);
922 struct vvp_io *vio = vvp_env_io(env);
923 struct cl_page_list *queue = &vio->u.write.vui_queue;
924 struct cl_page *page;
927 unsigned int npages = vio->u.write.vui_queue.pl_nr;
933 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
934 npages, vio->u.write.vui_from, vio->u.write.vui_to);
936 LASSERT(page_list_sanity_check(obj, queue));
938 /* submit IO with async write */
939 rc = cl_io_commit_async(env, io, queue,
940 vio->u.write.vui_from, vio->u.write.vui_to,
941 write_commit_callback);
942 npages -= queue->pl_nr; /* already committed pages */
944 /* calculate how many bytes were written */
945 bytes = npages << PAGE_SHIFT;
948 bytes -= vio->u.write.vui_from;
949 if (queue->pl_nr == 0) /* last page */
950 bytes -= PAGE_SIZE - vio->u.write.vui_to;
951 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
953 vio->u.write.vui_written += bytes;
955 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
956 npages, bytes, vio->u.write.vui_written);
958 /* the first page must have been written. */
959 vio->u.write.vui_from = 0;
961 LASSERT(page_list_sanity_check(obj, queue));
962 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
964 /* out of quota, try sync write */
965 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
966 rc = vvp_io_commit_sync(env, io, queue,
967 vio->u.write.vui_from,
968 vio->u.write.vui_to);
970 vio->u.write.vui_written += rc;
975 /* update inode size */
976 ll_merge_attr(env, inode);
978 /* Now the pages in queue were failed to commit, discard them
979 * unless they were dirtied before. */
980 while (queue->pl_nr > 0) {
981 page = cl_page_list_first(queue);
982 cl_page_list_del(env, queue, page);
984 if (!PageDirty(cl_page_vmpage(page)))
985 cl_page_discard(env, io, page);
987 cl_page_disown(env, io, page);
989 /* held in ll_cl_init() */
990 lu_ref_del(&page->cp_reference, "cl_io", io);
991 cl_page_put(env, page);
993 cl_page_list_fini(env, queue);
998 static int vvp_io_write_start(const struct lu_env *env,
999 const struct cl_io_slice *ios)
1001 struct vvp_io *vio = cl2vvp_io(env, ios);
1002 struct cl_io *io = ios->cis_io;
1003 struct cl_object *obj = io->ci_obj;
1004 struct inode *inode = vvp_object_inode(obj);
1006 loff_t pos = io->u.ci_wr.wr.crw_pos;
1007 size_t cnt = io->u.ci_wr.wr.crw_count;
1011 if (!can_populate_pages(env, io, inode))
1014 if (cl_io_is_append(io)) {
1016 * PARALLEL IO This has to be changed for parallel IO doing
1017 * out-of-order writes.
1019 ll_merge_attr(env, inode);
1020 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
1021 vio->vui_iocb->ki_pos = pos;
1023 LASSERT(vio->vui_iocb->ki_pos == pos);
1026 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
1028 if (vio->vui_iov == NULL) {
1029 /* from a temp io in ll_cl_init(). */
1033 * When using the locked AIO function (generic_file_aio_write())
1034 * testing has shown the inode mutex to be a limiting factor
1035 * with multi-threaded single shared file performance. To get
1036 * around this, we now use the lockless version. To maintain
1037 * consistency, proper locking to protect against writes,
1038 * trucates, etc. is handled in the higher layers of lustre.
1040 result = __generic_file_aio_write(vio->vui_iocb,
1041 vio->vui_iov, vio->vui_nrsegs,
1042 &vio->vui_iocb->ki_pos);
1043 if (result > 0 || result == -EIOCBQUEUED) {
1046 err = generic_write_sync(vio->vui_iocb->ki_filp,
1048 if (err < 0 && result > 0)
1054 result = vvp_io_write_commit(env, io);
1055 if (vio->u.write.vui_written > 0) {
1056 result = vio->u.write.vui_written;
1057 io->ci_nob += result;
1059 CDEBUG(D_VFSTRACE, "write: nob %zd, result: %zd\n",
1060 io->ci_nob, result);
1064 struct ll_inode_info *lli = ll_i2info(inode);
1066 spin_lock(&lli->lli_lock);
1067 lli->lli_flags |= LLIF_DATA_MODIFIED;
1068 spin_unlock(&lli->lli_lock);
1071 io->ci_continue = 0;
1072 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
1073 vio->vui_fd, pos, result, WRITE);
1080 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
1082 struct vm_fault *vmf = cfio->ft_vmf;
1084 cfio->ft_flags = filemap_fault(cfio->ft_vma, vmf);
1085 cfio->ft_flags_valid = 1;
1088 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
1089 vmf->virtual_address);
1090 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
1091 lock_page(vmf->page);
1092 cfio->ft_flags |= VM_FAULT_LOCKED;
1095 cfio->ft_vmpage = vmf->page;
1100 if (cfio->ft_flags & VM_FAULT_SIGBUS) {
1101 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
1105 if (cfio->ft_flags & VM_FAULT_OOM) {
1106 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
1110 if (cfio->ft_flags & VM_FAULT_RETRY)
1113 CERROR("unknown error in page fault %d\n", cfio->ft_flags);
1118 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1119 struct cl_page *page)
1121 set_page_dirty(page->cp_vmpage);
1124 static int vvp_io_fault_start(const struct lu_env *env,
1125 const struct cl_io_slice *ios)
1127 struct vvp_io *vio = cl2vvp_io(env, ios);
1128 struct cl_io *io = ios->cis_io;
1129 struct cl_object *obj = io->ci_obj;
1130 struct inode *inode = vvp_object_inode(obj);
1131 struct cl_fault_io *fio = &io->u.ci_fault;
1132 struct vvp_fault_io *cfio = &vio->u.fault;
1135 struct page *vmpage = NULL;
1136 struct cl_page *page;
1141 if (fio->ft_executable &&
1142 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
1144 " changed while waiting for the page fault lock\n",
1145 PFID(lu_object_fid(&obj->co_lu)));
1147 /* offset of the last byte on the page */
1148 offset = cl_offset(obj, fio->ft_index + 1) - 1;
1149 LASSERT(cl_index(obj, offset) == fio->ft_index);
1150 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1154 /* must return locked page */
1155 if (fio->ft_mkwrite) {
1156 LASSERT(cfio->ft_vmpage != NULL);
1157 lock_page(cfio->ft_vmpage);
1159 result = vvp_io_kernel_fault(cfio);
1164 vmpage = cfio->ft_vmpage;
1165 LASSERT(PageLocked(vmpage));
1167 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1168 ll_invalidate_page(vmpage);
1170 size = i_size_read(inode);
1171 /* Though we have already held a cl_lock upon this page, but
1172 * it still can be truncated locally. */
1173 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1174 (page_offset(vmpage) > size))) {
1175 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1177 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1179 GOTO(out, result = +1);
1182 last_index = cl_index(obj, size - 1);
1184 if (fio->ft_mkwrite ) {
1186 * Capture the size while holding the lli_trunc_sem from above
1187 * we want to make sure that we complete the mkwrite action
1188 * while holding this lock. We need to make sure that we are
1189 * not past the end of the file.
1191 if (last_index < fio->ft_index) {
1193 "llite: mkwrite and truncate race happened: "
1194 "%p: 0x%lx 0x%lx\n",
1195 vmpage->mapping,fio->ft_index,last_index);
1197 * We need to return if we are
1198 * passed the end of the file. This will propagate
1199 * up the call stack to ll_page_mkwrite where
1200 * we will return VM_FAULT_NOPAGE. Any non-negative
1201 * value returned here will be silently
1202 * converted to 0. If the vmpage->mapping is null
1203 * the error code would be converted back to ENODATA
1204 * in ll_page_mkwrite0. Thus we return -ENODATA
1205 * to handle both cases
1207 GOTO(out, result = -ENODATA);
1211 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1213 GOTO(out, result = PTR_ERR(page));
1215 /* if page is going to be written, we should add this page into cache
1217 if (fio->ft_mkwrite) {
1218 wait_on_page_writeback(vmpage);
1219 if (!PageDirty(vmpage)) {
1220 struct cl_page_list *plist = &io->ci_queue.c2_qin;
1221 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1224 /* vvp_page_assume() calls wait_on_page_writeback(). */
1225 cl_page_assume(env, io, page);
1227 cl_page_list_init(plist);
1228 cl_page_list_add(plist, page);
1231 if (last_index == vvp_index(vpg))
1232 to = size & ~CFS_PAGE_MASK;
1234 /* Do not set Dirty bit here so that in case IO is
1235 * started before the page is really made dirty, we
1236 * still have chance to detect it. */
1237 result = cl_io_commit_async(env, io, plist, 0, to,
1238 mkwrite_commit_callback);
1239 LASSERT(cl_page_is_owned(page, io));
1240 cl_page_list_fini(env, plist);
1244 cl_page_discard(env, io, page);
1245 cl_page_disown(env, io, page);
1247 cl_page_put(env, page);
1249 /* we're in big trouble, what can we do now? */
1250 if (result == -EDQUOT)
1254 cl_page_disown(env, io, page);
1259 * The ft_index is only used in the case of
1260 * a mkwrite action. We need to check
1261 * our assertions are correct, since
1262 * we should have caught this above
1264 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1265 if (fio->ft_index == last_index)
1267 * Last page is mapped partially.
1269 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1271 fio->ft_nob = cl_page_size(obj);
1273 lu_ref_add(&page->cp_reference, "fault", io);
1274 fio->ft_page = page;
1278 /* return unlocked vmpage to avoid deadlocking */
1280 unlock_page(vmpage);
1282 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1287 static int vvp_io_fsync_start(const struct lu_env *env,
1288 const struct cl_io_slice *ios)
1290 /* we should mark TOWRITE bit to each dirty page in radix tree to
1291 * verify pages have been written, but this is difficult because of
1296 static int vvp_io_read_ahead(const struct lu_env *env,
1297 const struct cl_io_slice *ios,
1298 pgoff_t start, struct cl_read_ahead *ra)
1303 if (ios->cis_io->ci_type == CIT_READ ||
1304 ios->cis_io->ci_type == CIT_FAULT) {
1305 struct vvp_io *vio = cl2vvp_io(env, ios);
1307 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1308 ra->cra_end = CL_PAGE_EOF;
1309 result = +1; /* no need to call down */
1316 static void vvp_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
1318 CLOBINVRNT(env, ios->cis_io->ci_obj,
1319 vvp_object_invariant(ios->cis_io->ci_obj));
1322 static const struct cl_io_operations vvp_io_ops = {
1325 .cio_fini = vvp_io_fini,
1326 .cio_lock = vvp_io_read_lock,
1327 .cio_start = vvp_io_read_start,
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_advance = vvp_io_advance,
1339 .cio_fini = vvp_io_setattr_fini,
1340 .cio_iter_init = vvp_io_setattr_iter_init,
1341 .cio_lock = vvp_io_setattr_lock,
1342 .cio_start = vvp_io_setattr_start,
1343 .cio_end = vvp_io_setattr_end
1346 .cio_fini = vvp_io_fault_fini,
1347 .cio_iter_init = vvp_io_fault_iter_init,
1348 .cio_lock = vvp_io_fault_lock,
1349 .cio_start = vvp_io_fault_start,
1350 .cio_end = vvp_io_end,
1353 .cio_start = vvp_io_fsync_start,
1354 .cio_fini = vvp_io_fini
1357 .cio_fini = vvp_io_fini
1360 .cio_read_ahead = vvp_io_read_ahead
1363 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1366 struct vvp_io *vio = vvp_env_io(env);
1367 struct inode *inode = vvp_object_inode(obj);
1370 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1373 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1374 "restore needed %d\n",
1375 PFID(lu_object_fid(&obj->co_lu)),
1376 io->ci_ignore_layout, io->ci_verify_layout,
1377 vio->vui_layout_gen, io->ci_restore_needed);
1379 CL_IO_SLICE_CLEAN(vio, vui_cl);
1380 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1381 vio->vui_ra_valid = false;
1383 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1385 struct ll_inode_info *lli = ll_i2info(inode);
1387 count = io->u.ci_rw.crw_count;
1388 /* "If nbyte is 0, read() will return 0 and have no other
1389 * results." -- Single Unix Spec */
1393 vio->vui_tot_count = count;
1394 vio->vui_tot_nrsegs = 0;
1397 /* for read/write, we store the jobid in the inode, and
1398 * it'll be fetched by osc when building RPC.
1400 * it's not accurate if the file is shared by different
1403 lustre_get_jobid(lli->lli_jobid);
1404 } else if (io->ci_type == CIT_SETATTR) {
1405 if (!cl_io_is_trunc(io))
1406 io->ci_lockreq = CILR_MANDATORY;
1409 /* ignore layout change for generic CIT_MISC but not for glimpse.
1410 * io context for glimpse must set ci_verify_layout to true,
1411 * see cl_glimpse_size0() for details. */
1412 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1413 io->ci_ignore_layout = 1;
1415 /* Enqueue layout lock and get layout version. We need to do this
1416 * even for operations requiring to open file, such as read and write,
1417 * because it might not grant layout lock in IT_OPEN. */
1418 if (result == 0 && !io->ci_ignore_layout) {
1419 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1420 if (result == -ENOENT)
1421 /* If the inode on MDS has been removed, but the objects
1422 * on OSTs haven't been destroyed (async unlink), layout
1423 * fetch will return -ENOENT, we'd ingore this error
1424 * and continue with dirty flush. LU-3230. */
1427 CERROR("%s: refresh file layout " DFID " error %d.\n",
1428 ll_get_fsname(inode->i_sb, NULL, 0),
1429 PFID(lu_object_fid(&obj->co_lu)), result);