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 = ccc_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.cg_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 ccc_thread_info *cti = ccc_env_info(env);
391 struct mm_struct *mm = current->mm;
392 struct vm_area_struct *vma;
393 struct cl_lock_descr *descr = &cti->cti_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 vvp_io *vio = vvp_env_io(env);
623 struct cl_io *io = ios->cis_io;
627 if (cl_io_is_trunc(io)) {
628 new_size = io->u.ci_setattr.sa_attr.lvb_size;
630 enqflags = CEF_DISCARD_DATA;
632 if ((io->u.ci_setattr.sa_attr.lvb_mtime >=
633 io->u.ci_setattr.sa_attr.lvb_ctime) ||
634 (io->u.ci_setattr.sa_attr.lvb_atime >=
635 io->u.ci_setattr.sa_attr.lvb_ctime))
640 vio->u.setattr.vui_local_lock = SETATTR_EXTENT_LOCK;
642 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
643 new_size, OBD_OBJECT_EOF);
646 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
651 * Only ll_inode_size_lock is taken at this level.
653 ll_inode_size_lock(inode);
654 result = inode_newsize_ok(inode, size);
656 ll_inode_size_unlock(inode);
659 i_size_write(inode, size);
661 ll_truncate_pagecache(inode, size);
662 ll_inode_size_unlock(inode);
666 static int vvp_io_setattr_trunc(const struct lu_env *env,
667 const struct cl_io_slice *ios,
668 struct inode *inode, loff_t size)
670 inode_dio_wait(inode);
674 static int vvp_io_setattr_time(const struct lu_env *env,
675 const struct cl_io_slice *ios)
677 struct cl_io *io = ios->cis_io;
678 struct cl_object *obj = io->ci_obj;
679 struct cl_attr *attr = ccc_env_thread_attr(env);
681 unsigned valid = CAT_CTIME;
683 cl_object_attr_lock(obj);
684 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
685 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
686 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
689 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
690 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
693 result = cl_object_attr_update(env, obj, attr, valid);
694 cl_object_attr_unlock(obj);
699 static int vvp_io_setattr_start(const struct lu_env *env,
700 const struct cl_io_slice *ios)
702 struct cl_io *io = ios->cis_io;
703 struct inode *inode = vvp_object_inode(io->ci_obj);
706 mutex_lock(&inode->i_mutex);
707 if (cl_io_is_trunc(io))
708 result = vvp_io_setattr_trunc(env, ios, inode,
709 io->u.ci_setattr.sa_attr.lvb_size);
711 result = vvp_io_setattr_time(env, ios);
715 static void vvp_io_setattr_end(const struct lu_env *env,
716 const struct cl_io_slice *ios)
718 struct cl_io *io = ios->cis_io;
719 struct inode *inode = vvp_object_inode(io->ci_obj);
721 if (cl_io_is_trunc(io)) {
722 /* Truncate in memory pages - they must be clean pages
723 * because osc has already notified to destroy osc_extents. */
724 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
725 inode_dio_write_done(inode);
727 mutex_unlock(&inode->i_mutex);
730 static void vvp_io_setattr_fini(const struct lu_env *env,
731 const struct cl_io_slice *ios)
733 vvp_io_fini(env, ios);
736 static int vvp_io_read_start(const struct lu_env *env,
737 const struct cl_io_slice *ios)
739 struct vvp_io *vio = cl2vvp_io(env, ios);
740 struct cl_io *io = ios->cis_io;
741 struct cl_object *obj = io->ci_obj;
742 struct inode *inode = vvp_object_inode(obj);
743 struct file *file = vio->vui_fd->fd_file;
746 loff_t pos = io->u.ci_rd.rd.crw_pos;
747 long cnt = io->u.ci_rd.rd.crw_count;
748 long tot = vio->vui_tot_count;
751 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
753 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
755 if (!can_populate_pages(env, io, inode))
758 result = vvp_prep_size(env, obj, io, pos, tot, &exceed);
761 else if (exceed != 0)
764 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
765 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
766 inode->i_ino, cnt, pos, i_size_read(inode));
768 /* turn off the kernel's read-ahead */
769 vio->vui_fd->fd_file->f_ra.ra_pages = 0;
771 /* initialize read-ahead window once per syscall */
772 if (!vio->vui_ra_valid) {
773 vio->vui_ra_valid = true;
774 vio->vui_ra_start = cl_index(obj, pos);
775 vio->vui_ra_count = cl_index(obj, tot + PAGE_CACHE_SIZE - 1);
781 switch (vio->vui_io_subtype) {
783 LASSERT(vio->vui_iocb->ki_pos == pos);
784 result = generic_file_aio_read(vio->vui_iocb,
785 vio->vui_iov, vio->vui_nrsegs,
786 vio->vui_iocb->ki_pos);
789 result = generic_file_splice_read(file, &pos,
790 vio->u.splice.vui_pipe, cnt,
791 vio->u.splice.vui_flags);
792 /* LU-1109: do splice read stripe by stripe otherwise if it
793 * may make nfsd stuck if this read occupied all internal pipe
798 CERROR("Wrong IO type %u\n", vio->vui_io_subtype);
806 io->ci_nob += result;
807 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, vio->vui_fd,
815 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
816 struct cl_page_list *plist, int from, int to)
818 struct cl_2queue *queue = &io->ci_queue;
819 struct cl_page *page;
820 unsigned int bytes = 0;
824 if (plist->pl_nr == 0)
827 if (from > 0 || to != PAGE_SIZE) {
828 page = cl_page_list_first(plist);
829 if (plist->pl_nr == 1) {
830 cl_page_clip(env, page, from, to);
833 cl_page_clip(env, page, from, PAGE_SIZE);
834 if (to != PAGE_SIZE) {
835 page = cl_page_list_last(plist);
836 cl_page_clip(env, page, 0, to);
841 cl_2queue_init(queue);
842 cl_page_list_splice(plist, &queue->c2_qin);
843 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
845 /* plist is not sorted any more */
846 cl_page_list_splice(&queue->c2_qin, plist);
847 cl_page_list_splice(&queue->c2_qout, plist);
848 cl_2queue_fini(env, queue);
851 /* calculate bytes */
852 bytes = plist->pl_nr << PAGE_SHIFT;
853 bytes -= from + PAGE_SIZE - to;
855 while (plist->pl_nr > 0) {
856 page = cl_page_list_first(plist);
857 cl_page_list_del(env, plist, page);
859 cl_page_clip(env, page, 0, PAGE_SIZE);
861 SetPageUptodate(cl_page_vmpage(page));
862 cl_page_disown(env, io, page);
864 /* held in ll_cl_init() */
865 lu_ref_del(&page->cp_reference, "cl_io", io);
866 cl_page_put(env, page);
870 RETURN(bytes > 0 ? bytes : rc);
873 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
874 struct cl_page *page)
876 struct page *vmpage = page->cp_vmpage;
878 SetPageUptodate(vmpage);
879 set_page_dirty(vmpage);
881 cl_page_disown(env, io, page);
883 /* held in ll_cl_init() */
884 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
885 cl_page_put(env, page);
888 /* make sure the page list is contiguous */
889 static bool page_list_sanity_check(struct cl_object *obj,
890 struct cl_page_list *plist)
892 struct cl_page *page;
893 pgoff_t index = CL_PAGE_EOF;
895 cl_page_list_for_each(page, plist) {
896 struct vvp_page *vpg = cl_object_page_slice(obj, page);
898 if (index == CL_PAGE_EOF) {
899 index = vvp_index(vpg);
904 if (index == vvp_index(vpg))
912 /* Return how many bytes have queued or written */
913 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
915 struct cl_object *obj = io->ci_obj;
916 struct inode *inode = vvp_object_inode(obj);
917 struct vvp_io *vio = vvp_env_io(env);
918 struct cl_page_list *queue = &vio->u.write.vui_queue;
919 struct cl_page *page;
922 unsigned int npages = vio->u.write.vui_queue.pl_nr;
928 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
929 npages, vio->u.write.vui_from, vio->u.write.vui_to);
931 LASSERT(page_list_sanity_check(obj, queue));
933 /* submit IO with async write */
934 rc = cl_io_commit_async(env, io, queue,
935 vio->u.write.vui_from, vio->u.write.vui_to,
936 write_commit_callback);
937 npages -= queue->pl_nr; /* already committed pages */
939 /* calculate how many bytes were written */
940 bytes = npages << PAGE_SHIFT;
943 bytes -= vio->u.write.vui_from;
944 if (queue->pl_nr == 0) /* last page */
945 bytes -= PAGE_SIZE - vio->u.write.vui_to;
946 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
948 vio->u.write.vui_written += bytes;
950 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
951 npages, bytes, vio->u.write.vui_written);
953 /* the first page must have been written. */
954 vio->u.write.vui_from = 0;
956 LASSERT(page_list_sanity_check(obj, queue));
957 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
959 /* out of quota, try sync write */
960 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
961 rc = vvp_io_commit_sync(env, io, queue,
962 vio->u.write.vui_from,
963 vio->u.write.vui_to);
965 vio->u.write.vui_written += rc;
970 /* update inode size */
971 ll_merge_attr(env, inode);
973 /* Now the pages in queue were failed to commit, discard them
974 * unless they were dirtied before. */
975 while (queue->pl_nr > 0) {
976 page = cl_page_list_first(queue);
977 cl_page_list_del(env, queue, page);
979 if (!PageDirty(cl_page_vmpage(page)))
980 cl_page_discard(env, io, page);
982 cl_page_disown(env, io, page);
984 /* held in ll_cl_init() */
985 lu_ref_del(&page->cp_reference, "cl_io", io);
986 cl_page_put(env, page);
988 cl_page_list_fini(env, queue);
993 static int vvp_io_write_start(const struct lu_env *env,
994 const struct cl_io_slice *ios)
996 struct vvp_io *vio = cl2vvp_io(env, ios);
997 struct cl_io *io = ios->cis_io;
998 struct cl_object *obj = io->ci_obj;
999 struct inode *inode = vvp_object_inode(obj);
1001 loff_t pos = io->u.ci_wr.wr.crw_pos;
1002 size_t cnt = io->u.ci_wr.wr.crw_count;
1006 if (!can_populate_pages(env, io, inode))
1009 if (cl_io_is_append(io)) {
1011 * PARALLEL IO This has to be changed for parallel IO doing
1012 * out-of-order writes.
1014 ll_merge_attr(env, inode);
1015 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
1016 vio->vui_iocb->ki_pos = pos;
1018 LASSERT(vio->vui_iocb->ki_pos == pos);
1021 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
1023 if (vio->vui_iov == NULL) {
1024 /* from a temp io in ll_cl_init(). */
1028 * When using the locked AIO function (generic_file_aio_write())
1029 * testing has shown the inode mutex to be a limiting factor
1030 * with multi-threaded single shared file performance. To get
1031 * around this, we now use the lockless version. To maintain
1032 * consistency, proper locking to protect against writes,
1033 * trucates, etc. is handled in the higher layers of lustre.
1035 result = __generic_file_aio_write(vio->vui_iocb,
1036 vio->vui_iov, vio->vui_nrsegs,
1037 &vio->vui_iocb->ki_pos);
1038 if (result > 0 || result == -EIOCBQUEUED) {
1041 err = generic_write_sync(vio->vui_iocb->ki_filp,
1043 if (err < 0 && result > 0)
1049 result = vvp_io_write_commit(env, io);
1050 if (vio->u.write.vui_written > 0) {
1051 result = vio->u.write.vui_written;
1052 io->ci_nob += result;
1054 CDEBUG(D_VFSTRACE, "write: nob %zd, result: %zd\n",
1055 io->ci_nob, result);
1059 struct ll_inode_info *lli = ll_i2info(inode);
1061 spin_lock(&lli->lli_lock);
1062 lli->lli_flags |= LLIF_DATA_MODIFIED;
1063 spin_unlock(&lli->lli_lock);
1066 io->ci_continue = 0;
1067 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
1068 vio->vui_fd, pos, result, WRITE);
1075 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
1077 struct vm_fault *vmf = cfio->ft_vmf;
1079 cfio->ft_flags = filemap_fault(cfio->ft_vma, vmf);
1080 cfio->ft_flags_valid = 1;
1083 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
1084 vmf->virtual_address);
1085 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
1086 lock_page(vmf->page);
1087 cfio->ft_flags |= VM_FAULT_LOCKED;
1090 cfio->ft_vmpage = vmf->page;
1095 if (cfio->ft_flags & VM_FAULT_SIGBUS) {
1096 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
1100 if (cfio->ft_flags & VM_FAULT_OOM) {
1101 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
1105 if (cfio->ft_flags & VM_FAULT_RETRY)
1108 CERROR("unknown error in page fault %d\n", cfio->ft_flags);
1113 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1114 struct cl_page *page)
1116 set_page_dirty(page->cp_vmpage);
1119 static int vvp_io_fault_start(const struct lu_env *env,
1120 const struct cl_io_slice *ios)
1122 struct vvp_io *vio = cl2vvp_io(env, ios);
1123 struct cl_io *io = ios->cis_io;
1124 struct cl_object *obj = io->ci_obj;
1125 struct inode *inode = vvp_object_inode(obj);
1126 struct cl_fault_io *fio = &io->u.ci_fault;
1127 struct vvp_fault_io *cfio = &vio->u.fault;
1130 struct page *vmpage = NULL;
1131 struct cl_page *page;
1136 if (fio->ft_executable &&
1137 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
1139 " changed while waiting for the page fault lock\n",
1140 PFID(lu_object_fid(&obj->co_lu)));
1142 /* offset of the last byte on the page */
1143 offset = cl_offset(obj, fio->ft_index + 1) - 1;
1144 LASSERT(cl_index(obj, offset) == fio->ft_index);
1145 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1149 /* must return locked page */
1150 if (fio->ft_mkwrite) {
1151 LASSERT(cfio->ft_vmpage != NULL);
1152 lock_page(cfio->ft_vmpage);
1154 result = vvp_io_kernel_fault(cfio);
1159 vmpage = cfio->ft_vmpage;
1160 LASSERT(PageLocked(vmpage));
1162 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1163 ll_invalidate_page(vmpage);
1165 size = i_size_read(inode);
1166 /* Though we have already held a cl_lock upon this page, but
1167 * it still can be truncated locally. */
1168 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1169 (page_offset(vmpage) > size))) {
1170 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1172 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1174 GOTO(out, result = +1);
1177 last_index = cl_index(obj, size - 1);
1179 if (fio->ft_mkwrite ) {
1181 * Capture the size while holding the lli_trunc_sem from above
1182 * we want to make sure that we complete the mkwrite action
1183 * while holding this lock. We need to make sure that we are
1184 * not past the end of the file.
1186 if (last_index < fio->ft_index) {
1188 "llite: mkwrite and truncate race happened: "
1189 "%p: 0x%lx 0x%lx\n",
1190 vmpage->mapping,fio->ft_index,last_index);
1192 * We need to return if we are
1193 * passed the end of the file. This will propagate
1194 * up the call stack to ll_page_mkwrite where
1195 * we will return VM_FAULT_NOPAGE. Any non-negative
1196 * value returned here will be silently
1197 * converted to 0. If the vmpage->mapping is null
1198 * the error code would be converted back to ENODATA
1199 * in ll_page_mkwrite0. Thus we return -ENODATA
1200 * to handle both cases
1202 GOTO(out, result = -ENODATA);
1206 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1208 GOTO(out, result = PTR_ERR(page));
1210 /* if page is going to be written, we should add this page into cache
1212 if (fio->ft_mkwrite) {
1213 wait_on_page_writeback(vmpage);
1214 if (!PageDirty(vmpage)) {
1215 struct cl_page_list *plist = &io->ci_queue.c2_qin;
1216 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1219 /* vvp_page_assume() calls wait_on_page_writeback(). */
1220 cl_page_assume(env, io, page);
1222 cl_page_list_init(plist);
1223 cl_page_list_add(plist, page);
1226 if (last_index == vvp_index(vpg))
1227 to = size & ~CFS_PAGE_MASK;
1229 /* Do not set Dirty bit here so that in case IO is
1230 * started before the page is really made dirty, we
1231 * still have chance to detect it. */
1232 result = cl_io_commit_async(env, io, plist, 0, to,
1233 mkwrite_commit_callback);
1234 LASSERT(cl_page_is_owned(page, io));
1235 cl_page_list_fini(env, plist);
1239 cl_page_discard(env, io, page);
1240 cl_page_disown(env, io, page);
1242 cl_page_put(env, page);
1244 /* we're in big trouble, what can we do now? */
1245 if (result == -EDQUOT)
1249 cl_page_disown(env, io, page);
1254 * The ft_index is only used in the case of
1255 * a mkwrite action. We need to check
1256 * our assertions are correct, since
1257 * we should have caught this above
1259 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1260 if (fio->ft_index == last_index)
1262 * Last page is mapped partially.
1264 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1266 fio->ft_nob = cl_page_size(obj);
1268 lu_ref_add(&page->cp_reference, "fault", io);
1269 fio->ft_page = page;
1273 /* return unlocked vmpage to avoid deadlocking */
1275 unlock_page(vmpage);
1277 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1282 static int vvp_io_fsync_start(const struct lu_env *env,
1283 const struct cl_io_slice *ios)
1285 /* we should mark TOWRITE bit to each dirty page in radix tree to
1286 * verify pages have been written, but this is difficult because of
1291 static int vvp_io_read_ahead(const struct lu_env *env,
1292 const struct cl_io_slice *ios,
1293 pgoff_t start, struct cl_read_ahead *ra)
1298 if (ios->cis_io->ci_type == CIT_READ ||
1299 ios->cis_io->ci_type == CIT_FAULT) {
1300 struct vvp_io *vio = cl2vvp_io(env, ios);
1302 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1303 ra->cra_end = CL_PAGE_EOF;
1304 result = +1; /* no need to call down */
1311 static void vvp_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
1313 CLOBINVRNT(env, ios->cis_io->ci_obj,
1314 vvp_object_invariant(ios->cis_io->ci_obj));
1317 static const struct cl_io_operations vvp_io_ops = {
1320 .cio_fini = vvp_io_fini,
1321 .cio_lock = vvp_io_read_lock,
1322 .cio_start = vvp_io_read_start,
1323 .cio_advance = vvp_io_advance,
1326 .cio_fini = vvp_io_fini,
1327 .cio_iter_init = vvp_io_write_iter_init,
1328 .cio_iter_fini = vvp_io_write_iter_fini,
1329 .cio_lock = vvp_io_write_lock,
1330 .cio_start = vvp_io_write_start,
1331 .cio_advance = vvp_io_advance,
1334 .cio_fini = vvp_io_setattr_fini,
1335 .cio_iter_init = vvp_io_setattr_iter_init,
1336 .cio_lock = vvp_io_setattr_lock,
1337 .cio_start = vvp_io_setattr_start,
1338 .cio_end = vvp_io_setattr_end
1341 .cio_fini = vvp_io_fault_fini,
1342 .cio_iter_init = vvp_io_fault_iter_init,
1343 .cio_lock = vvp_io_fault_lock,
1344 .cio_start = vvp_io_fault_start,
1345 .cio_end = vvp_io_end,
1348 .cio_start = vvp_io_fsync_start,
1349 .cio_fini = vvp_io_fini
1352 .cio_fini = vvp_io_fini
1355 .cio_read_ahead = vvp_io_read_ahead
1358 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1361 struct vvp_io *vio = vvp_env_io(env);
1362 struct inode *inode = vvp_object_inode(obj);
1365 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1368 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1369 "restore needed %d\n",
1370 PFID(lu_object_fid(&obj->co_lu)),
1371 io->ci_ignore_layout, io->ci_verify_layout,
1372 vio->vui_layout_gen, io->ci_restore_needed);
1374 CL_IO_SLICE_CLEAN(vio, vui_cl);
1375 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1376 vio->vui_ra_valid = false;
1378 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1380 struct ll_inode_info *lli = ll_i2info(inode);
1382 count = io->u.ci_rw.crw_count;
1383 /* "If nbyte is 0, read() will return 0 and have no other
1384 * results." -- Single Unix Spec */
1388 vio->vui_tot_count = count;
1389 vio->vui_tot_nrsegs = 0;
1392 /* for read/write, we store the jobid in the inode, and
1393 * it'll be fetched by osc when building RPC.
1395 * it's not accurate if the file is shared by different
1398 lustre_get_jobid(lli->lli_jobid);
1399 } else if (io->ci_type == CIT_SETATTR) {
1400 if (!cl_io_is_trunc(io))
1401 io->ci_lockreq = CILR_MANDATORY;
1404 /* ignore layout change for generic CIT_MISC but not for glimpse.
1405 * io context for glimpse must set ci_verify_layout to true,
1406 * see cl_glimpse_size0() for details. */
1407 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1408 io->ci_ignore_layout = 1;
1410 /* Enqueue layout lock and get layout version. We need to do this
1411 * even for operations requiring to open file, such as read and write,
1412 * because it might not grant layout lock in IT_OPEN. */
1413 if (result == 0 && !io->ci_ignore_layout) {
1414 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1415 if (result == -ENOENT)
1416 /* If the inode on MDS has been removed, but the objects
1417 * on OSTs haven't been destroyed (async unlink), layout
1418 * fetch will return -ENOENT, we'd ingore this error
1419 * and continue with dirty flush. LU-3230. */
1422 CERROR("%s: refresh file layout " DFID " error %d.\n",
1423 ll_get_fsname(inode->i_sb, NULL, 0),
1424 PFID(lu_object_fid(&obj->co_lu)), result);