4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2015, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * Implementation of cl_io for VVP layer.
34 * Author: Nikita Danilov <nikita.danilov@sun.com>
35 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
38 #define DEBUG_SUBSYSTEM S_LLITE
40 #include "../include/obd.h"
41 #include "../include/lustre_lite.h"
43 #include "llite_internal.h"
44 #include "vvp_internal.h"
46 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
47 const struct cl_io_slice *slice)
51 vio = container_of(slice, struct vvp_io, vui_cl);
52 LASSERT(vio == vvp_env_io(env));
58 * True, if \a io is a normal io, False for splice_{read,write}
60 static int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
62 struct vvp_io *vio = vvp_env_io(env);
64 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
66 return vio->vui_io_subtype == IO_NORMAL;
70 * For swapping layout. The file's layout may have changed.
71 * To avoid populating pages to a wrong stripe, we have to verify the
72 * correctness of layout. It works because swapping layout processes
73 * have to acquire group lock.
75 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
78 struct ll_inode_info *lli = ll_i2info(inode);
79 struct vvp_io *vio = vvp_env_io(env);
82 switch (io->ci_type) {
85 /* don't need lock here to check lli_layout_gen as we have held
86 * extent lock and GROUP lock has to hold to swap layout
88 if (ll_layout_version_get(lli) != vio->vui_layout_gen) {
89 io->ci_need_restart = 1;
90 /* this will return application a short read/write */
95 /* fault is okay because we've already had a page. */
103 static void vvp_object_size_lock(struct cl_object *obj)
105 struct inode *inode = vvp_object_inode(obj);
107 ll_inode_size_lock(inode);
108 cl_object_attr_lock(obj);
111 static void vvp_object_size_unlock(struct cl_object *obj)
113 struct inode *inode = vvp_object_inode(obj);
115 cl_object_attr_unlock(obj);
116 ll_inode_size_unlock(inode);
120 * Helper function that if necessary adjusts file size (inode->i_size), when
121 * position at the offset \a pos is accessed. File size can be arbitrary stale
122 * on a Lustre client, but client at least knows KMS. If accessed area is
123 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
125 * Locking: cl_isize_lock is used to serialize changes to inode size and to
126 * protect consistency between inode size and cl_object
127 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
128 * top-object and sub-objects.
130 static int vvp_prep_size(const struct lu_env *env, struct cl_object *obj,
131 struct cl_io *io, loff_t start, size_t count,
134 struct cl_attr *attr = vvp_env_thread_attr(env);
135 struct inode *inode = vvp_object_inode(obj);
136 loff_t pos = start + count - 1;
141 * Consistency guarantees: following possibilities exist for the
142 * relation between region being accessed and real file size at this
145 * (A): the region is completely inside of the file;
147 * (B-x): x bytes of region are inside of the file, the rest is
150 * (C): the region is completely outside of the file.
152 * This classification is stable under DLM lock already acquired by
153 * the caller, because to change the class, other client has to take
154 * DLM lock conflicting with our lock. Also, any updates to ->i_size
155 * by other threads on this client are serialized by
156 * ll_inode_size_lock(). This guarantees that short reads are handled
157 * correctly in the face of concurrent writes and truncates.
159 vvp_object_size_lock(obj);
160 result = cl_object_attr_get(env, obj, attr);
165 * A glimpse is necessary to determine whether we
166 * return a short read (B) or some zeroes at the end
169 vvp_object_size_unlock(obj);
170 result = cl_glimpse_lock(env, io, inode, obj, 0);
171 if (result == 0 && exceed) {
172 /* If objective page index exceed end-of-file
173 * page index, return directly. Do not expect
174 * kernel will check such case correctly.
175 * linux-2.6.18-128.1.1 miss to do that.
178 loff_t size = i_size_read(inode);
179 loff_t cur_index = start >> PAGE_SHIFT;
180 loff_t size_index = (size - 1) >> PAGE_SHIFT;
182 if ((size == 0 && cur_index != 0) ||
183 size_index < cur_index)
189 * region is within kms and, hence, within real file
190 * size (A). We need to increase i_size to cover the
191 * read region so that generic_file_read() will do its
192 * job, but that doesn't mean the kms size is
193 * _correct_, it is only the _minimum_ size. If
194 * someone does a stat they will get the correct size
195 * which will always be >= the kms value here.
198 if (i_size_read(inode) < kms) {
199 i_size_write(inode, kms);
200 CDEBUG(D_VFSTRACE, DFID " updating i_size %llu\n",
201 PFID(lu_object_fid(&obj->co_lu)),
202 (__u64)i_size_read(inode));
206 vvp_object_size_unlock(obj);
211 /*****************************************************************************
217 static int vvp_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
218 __u32 enqflags, enum cl_lock_mode mode,
219 pgoff_t start, pgoff_t end)
221 struct vvp_io *vio = vvp_env_io(env);
222 struct cl_lock_descr *descr = &vio->vui_link.cill_descr;
223 struct cl_object *obj = io->ci_obj;
225 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
227 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
229 memset(&vio->vui_link, 0, sizeof(vio->vui_link));
231 if (vio->vui_fd && (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
232 descr->cld_mode = CLM_GROUP;
233 descr->cld_gid = vio->vui_fd->fd_grouplock.lg_gid;
235 descr->cld_mode = mode;
237 descr->cld_obj = obj;
238 descr->cld_start = start;
239 descr->cld_end = end;
240 descr->cld_enq_flags = enqflags;
242 cl_io_lock_add(env, io, &vio->vui_link);
246 static int vvp_io_one_lock(const struct lu_env *env, struct cl_io *io,
247 __u32 enqflags, enum cl_lock_mode mode,
248 loff_t start, loff_t end)
250 struct cl_object *obj = io->ci_obj;
252 return vvp_io_one_lock_index(env, io, enqflags, mode,
253 cl_index(obj, start), cl_index(obj, end));
256 static int vvp_io_write_iter_init(const struct lu_env *env,
257 const struct cl_io_slice *ios)
259 struct vvp_io *vio = cl2vvp_io(env, ios);
261 cl_page_list_init(&vio->u.write.vui_queue);
262 vio->u.write.vui_written = 0;
263 vio->u.write.vui_from = 0;
264 vio->u.write.vui_to = PAGE_SIZE;
269 static void vvp_io_write_iter_fini(const struct lu_env *env,
270 const struct cl_io_slice *ios)
272 struct vvp_io *vio = cl2vvp_io(env, ios);
274 LASSERT(vio->u.write.vui_queue.pl_nr == 0);
277 static int vvp_io_fault_iter_init(const struct lu_env *env,
278 const struct cl_io_slice *ios)
280 struct vvp_io *vio = cl2vvp_io(env, ios);
281 struct inode *inode = vvp_object_inode(ios->cis_obj);
283 LASSERT(inode == file_inode(vio->vui_fd->fd_file));
284 vio->u.fault.ft_mtime = inode->i_mtime.tv_sec;
288 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
290 struct cl_io *io = ios->cis_io;
291 struct cl_object *obj = io->ci_obj;
292 struct vvp_io *vio = cl2vvp_io(env, ios);
293 struct inode *inode = vvp_object_inode(obj);
295 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
297 CDEBUG(D_VFSTRACE, DFID
298 " ignore/verify layout %d/%d, layout version %d restore needed %d\n",
299 PFID(lu_object_fid(&obj->co_lu)),
300 io->ci_ignore_layout, io->ci_verify_layout,
301 vio->vui_layout_gen, io->ci_restore_needed);
303 if (io->ci_restore_needed == 1) {
306 /* file was detected release, we need to restore it
307 * before finishing the io
309 rc = ll_layout_restore(inode, 0, OBD_OBJECT_EOF);
310 /* if restore registration failed, no restart,
311 * we will return -ENODATA
313 /* The layout will change after restore, so we need to
314 * block on layout lock hold by the MDT
315 * as MDT will not send new layout in lvb (see LU-3124)
316 * we have to explicitly fetch it, all this will be done
317 * by ll_layout_refresh()
320 io->ci_restore_needed = 0;
321 io->ci_need_restart = 1;
322 io->ci_verify_layout = 1;
324 io->ci_restore_needed = 1;
325 io->ci_need_restart = 0;
326 io->ci_verify_layout = 0;
331 if (!io->ci_ignore_layout && io->ci_verify_layout) {
334 /* check layout version */
335 ll_layout_refresh(inode, &gen);
336 io->ci_need_restart = vio->vui_layout_gen != gen;
337 if (io->ci_need_restart) {
339 DFID" layout changed from %d to %d.\n",
340 PFID(lu_object_fid(&obj->co_lu)),
341 vio->vui_layout_gen, gen);
342 /* today successful restore is the only possible case */
343 /* restore was done, clear restoring state */
344 ll_i2info(vvp_object_inode(obj))->lli_flags &=
345 ~LLIF_FILE_RESTORING;
350 static void vvp_io_fault_fini(const struct lu_env *env,
351 const struct cl_io_slice *ios)
353 struct cl_io *io = ios->cis_io;
354 struct cl_page *page = io->u.ci_fault.ft_page;
356 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
359 lu_ref_del(&page->cp_reference, "fault", io);
360 cl_page_put(env, page);
361 io->u.ci_fault.ft_page = NULL;
363 vvp_io_fini(env, ios);
366 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
369 * we only want to hold PW locks if the mmap() can generate
370 * writes back to the file and that only happens in shared
373 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
378 static int vvp_mmap_locks(const struct lu_env *env,
379 struct vvp_io *vio, struct cl_io *io)
381 struct vvp_thread_info *cti = vvp_env_info(env);
382 struct mm_struct *mm = current->mm;
383 struct vm_area_struct *vma;
384 struct cl_lock_descr *descr = &cti->vti_descr;
385 ldlm_policy_data_t policy;
392 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
394 if (!cl_is_normalio(env, io))
397 if (!vio->vui_iter) /* nfs or loop back device write */
400 /* No MM (e.g. NFS)? No vmas too. */
404 iov_for_each(iov, i, *vio->vui_iter) {
405 addr = (unsigned long)iov.iov_base;
410 count += addr & (~PAGE_MASK);
413 down_read(&mm->mmap_sem);
414 while ((vma = our_vma(mm, addr, count)) != NULL) {
415 struct inode *inode = file_inode(vma->vm_file);
416 int flags = CEF_MUST;
418 if (ll_file_nolock(vma->vm_file)) {
420 * For no lock case is not allowed for mmap
427 * XXX: Required lock mode can be weakened: CIT_WRITE
428 * io only ever reads user level buffer, and CIT_READ
431 policy_from_vma(&policy, vma, addr, count);
432 descr->cld_mode = vvp_mode_from_vma(vma);
433 descr->cld_obj = ll_i2info(inode)->lli_clob;
434 descr->cld_start = cl_index(descr->cld_obj,
435 policy.l_extent.start);
436 descr->cld_end = cl_index(descr->cld_obj,
437 policy.l_extent.end);
438 descr->cld_enq_flags = flags;
439 result = cl_io_lock_alloc_add(env, io, descr);
441 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
442 descr->cld_mode, descr->cld_start,
448 if (vma->vm_end - addr >= count)
451 count -= vma->vm_end - addr;
454 up_read(&mm->mmap_sem);
461 static void vvp_io_advance(const struct lu_env *env,
462 const struct cl_io_slice *ios,
465 struct vvp_io *vio = cl2vvp_io(env, ios);
466 struct cl_io *io = ios->cis_io;
467 struct cl_object *obj = ios->cis_io->ci_obj;
469 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
471 if (!cl_is_normalio(env, io))
474 iov_iter_reexpand(vio->vui_iter, vio->vui_tot_count -= nob);
477 static void vvp_io_update_iov(const struct lu_env *env,
478 struct vvp_io *vio, struct cl_io *io)
480 size_t size = io->u.ci_rw.crw_count;
482 if (!cl_is_normalio(env, io) || !vio->vui_iter)
485 iov_iter_truncate(vio->vui_iter, size);
488 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
489 enum cl_lock_mode mode, loff_t start, loff_t end)
491 struct vvp_io *vio = vvp_env_io(env);
495 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
497 vvp_io_update_iov(env, vio, io);
499 if (io->u.ci_rw.crw_nonblock)
500 ast_flags |= CEF_NONBLOCK;
501 result = vvp_mmap_locks(env, vio, io);
503 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
507 static int vvp_io_read_lock(const struct lu_env *env,
508 const struct cl_io_slice *ios)
510 struct cl_io *io = ios->cis_io;
511 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
514 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
515 rd->crw_pos + rd->crw_count - 1);
520 static int vvp_io_fault_lock(const struct lu_env *env,
521 const struct cl_io_slice *ios)
523 struct cl_io *io = ios->cis_io;
524 struct vvp_io *vio = cl2vvp_io(env, ios);
526 * XXX LDLM_FL_CBPENDING
528 return vvp_io_one_lock_index(env,
530 vvp_mode_from_vma(vio->u.fault.ft_vma),
531 io->u.ci_fault.ft_index,
532 io->u.ci_fault.ft_index);
535 static int vvp_io_write_lock(const struct lu_env *env,
536 const struct cl_io_slice *ios)
538 struct cl_io *io = ios->cis_io;
542 if (io->u.ci_wr.wr_append) {
544 end = OBD_OBJECT_EOF;
546 start = io->u.ci_wr.wr.crw_pos;
547 end = start + io->u.ci_wr.wr.crw_count - 1;
549 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
552 static int vvp_io_setattr_iter_init(const struct lu_env *env,
553 const struct cl_io_slice *ios)
559 * Implementation of cl_io_operations::vio_lock() method for CIT_SETATTR io.
561 * Handles "lockless io" mode when extent locking is done by server.
563 static int vvp_io_setattr_lock(const struct lu_env *env,
564 const struct cl_io_slice *ios)
566 struct cl_io *io = ios->cis_io;
570 if (cl_io_is_trunc(io)) {
571 new_size = io->u.ci_setattr.sa_attr.lvb_size;
573 enqflags = CEF_DISCARD_DATA;
575 if ((io->u.ci_setattr.sa_attr.lvb_mtime >=
576 io->u.ci_setattr.sa_attr.lvb_ctime) ||
577 (io->u.ci_setattr.sa_attr.lvb_atime >=
578 io->u.ci_setattr.sa_attr.lvb_ctime))
583 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
584 new_size, OBD_OBJECT_EOF);
587 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
591 * Only ll_inode_size_lock is taken at this level.
593 ll_inode_size_lock(inode);
594 result = inode_newsize_ok(inode, size);
596 ll_inode_size_unlock(inode);
599 truncate_setsize(inode, size);
600 ll_inode_size_unlock(inode);
604 static int vvp_io_setattr_trunc(const struct lu_env *env,
605 const struct cl_io_slice *ios,
606 struct inode *inode, loff_t size)
608 inode_dio_wait(inode);
612 static int vvp_io_setattr_time(const struct lu_env *env,
613 const struct cl_io_slice *ios)
615 struct cl_io *io = ios->cis_io;
616 struct cl_object *obj = io->ci_obj;
617 struct cl_attr *attr = vvp_env_thread_attr(env);
619 unsigned valid = CAT_CTIME;
621 cl_object_attr_lock(obj);
622 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
623 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
624 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
627 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
628 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
631 result = cl_object_attr_set(env, obj, attr, valid);
632 cl_object_attr_unlock(obj);
637 static int vvp_io_setattr_start(const struct lu_env *env,
638 const struct cl_io_slice *ios)
640 struct cl_io *io = ios->cis_io;
641 struct inode *inode = vvp_object_inode(io->ci_obj);
645 if (cl_io_is_trunc(io))
646 result = vvp_io_setattr_trunc(env, ios, inode,
647 io->u.ci_setattr.sa_attr.lvb_size);
649 result = vvp_io_setattr_time(env, ios);
653 static void vvp_io_setattr_end(const struct lu_env *env,
654 const struct cl_io_slice *ios)
656 struct cl_io *io = ios->cis_io;
657 struct inode *inode = vvp_object_inode(io->ci_obj);
659 if (cl_io_is_trunc(io))
660 /* Truncate in memory pages - they must be clean pages
661 * because osc has already notified to destroy osc_extents.
663 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
668 static void vvp_io_setattr_fini(const struct lu_env *env,
669 const struct cl_io_slice *ios)
671 vvp_io_fini(env, ios);
674 static int vvp_io_read_start(const struct lu_env *env,
675 const struct cl_io_slice *ios)
677 struct vvp_io *vio = cl2vvp_io(env, ios);
678 struct cl_io *io = ios->cis_io;
679 struct cl_object *obj = io->ci_obj;
680 struct inode *inode = vvp_object_inode(obj);
681 struct file *file = vio->vui_fd->fd_file;
684 loff_t pos = io->u.ci_rd.rd.crw_pos;
685 long cnt = io->u.ci_rd.rd.crw_count;
686 long tot = vio->vui_tot_count;
689 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
691 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
693 if (!can_populate_pages(env, io, inode))
696 result = vvp_prep_size(env, obj, io, pos, tot, &exceed);
699 else if (exceed != 0)
702 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
703 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
704 inode->i_ino, cnt, pos, i_size_read(inode));
706 /* turn off the kernel's read-ahead */
707 vio->vui_fd->fd_file->f_ra.ra_pages = 0;
709 /* initialize read-ahead window once per syscall */
710 if (!vio->vui_ra_valid) {
711 vio->vui_ra_valid = true;
712 vio->vui_ra_start = cl_index(obj, pos);
713 vio->vui_ra_count = cl_index(obj, tot + PAGE_SIZE - 1);
719 switch (vio->vui_io_subtype) {
721 LASSERT(vio->vui_iocb->ki_pos == pos);
722 result = generic_file_read_iter(vio->vui_iocb, vio->vui_iter);
725 result = generic_file_splice_read(file, &pos,
726 vio->u.splice.vui_pipe, cnt,
727 vio->u.splice.vui_flags);
728 /* LU-1109: do splice read stripe by stripe otherwise if it
729 * may make nfsd stuck if this read occupied all internal pipe
735 CERROR("Wrong IO type %u\n", vio->vui_io_subtype);
743 io->ci_nob += result;
744 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
745 vio->vui_fd, pos, result, READ);
751 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
752 struct cl_page_list *plist, int from, int to)
754 struct cl_2queue *queue = &io->ci_queue;
755 struct cl_page *page;
756 unsigned int bytes = 0;
759 if (plist->pl_nr == 0)
762 if (from > 0 || to != PAGE_SIZE) {
763 page = cl_page_list_first(plist);
764 if (plist->pl_nr == 1) {
765 cl_page_clip(env, page, from, to);
768 cl_page_clip(env, page, from, PAGE_SIZE);
769 if (to != PAGE_SIZE) {
770 page = cl_page_list_last(plist);
771 cl_page_clip(env, page, 0, to);
776 cl_2queue_init(queue);
777 cl_page_list_splice(plist, &queue->c2_qin);
778 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
780 /* plist is not sorted any more */
781 cl_page_list_splice(&queue->c2_qin, plist);
782 cl_page_list_splice(&queue->c2_qout, plist);
783 cl_2queue_fini(env, queue);
786 /* calculate bytes */
787 bytes = plist->pl_nr << PAGE_SHIFT;
788 bytes -= from + PAGE_SIZE - to;
790 while (plist->pl_nr > 0) {
791 page = cl_page_list_first(plist);
792 cl_page_list_del(env, plist, page);
794 cl_page_clip(env, page, 0, PAGE_SIZE);
796 SetPageUptodate(cl_page_vmpage(page));
797 cl_page_disown(env, io, page);
799 /* held in ll_cl_init() */
800 lu_ref_del(&page->cp_reference, "cl_io", io);
801 cl_page_put(env, page);
805 return bytes > 0 ? bytes : rc;
808 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
809 struct cl_page *page)
811 struct vvp_page *vpg;
812 struct page *vmpage = page->cp_vmpage;
813 struct cl_object *clob = cl_io_top(io)->ci_obj;
815 SetPageUptodate(vmpage);
816 set_page_dirty(vmpage);
818 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
819 vvp_write_pending(cl2vvp(clob), vpg);
821 cl_page_disown(env, io, page);
823 /* held in ll_cl_init() */
824 lu_ref_del(&page->cp_reference, "cl_io", io);
825 cl_page_put(env, page);
828 /* make sure the page list is contiguous */
829 static bool page_list_sanity_check(struct cl_object *obj,
830 struct cl_page_list *plist)
832 struct cl_page *page;
833 pgoff_t index = CL_PAGE_EOF;
835 cl_page_list_for_each(page, plist) {
836 struct vvp_page *vpg = cl_object_page_slice(obj, page);
838 if (index == CL_PAGE_EOF) {
839 index = vvp_index(vpg);
844 if (index == vvp_index(vpg))
852 /* Return how many bytes have queued or written */
853 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
855 struct cl_object *obj = io->ci_obj;
856 struct inode *inode = vvp_object_inode(obj);
857 struct vvp_io *vio = vvp_env_io(env);
858 struct cl_page_list *queue = &vio->u.write.vui_queue;
859 struct cl_page *page;
862 unsigned int npages = vio->u.write.vui_queue.pl_nr;
867 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
868 npages, vio->u.write.vui_from, vio->u.write.vui_to);
870 LASSERT(page_list_sanity_check(obj, queue));
872 /* submit IO with async write */
873 rc = cl_io_commit_async(env, io, queue,
874 vio->u.write.vui_from, vio->u.write.vui_to,
875 write_commit_callback);
876 npages -= queue->pl_nr; /* already committed pages */
878 /* calculate how many bytes were written */
879 bytes = npages << PAGE_SHIFT;
882 bytes -= vio->u.write.vui_from;
883 if (queue->pl_nr == 0) /* last page */
884 bytes -= PAGE_SIZE - vio->u.write.vui_to;
885 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
887 vio->u.write.vui_written += bytes;
889 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
890 npages, bytes, vio->u.write.vui_written);
892 /* the first page must have been written. */
893 vio->u.write.vui_from = 0;
895 LASSERT(page_list_sanity_check(obj, queue));
896 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
898 /* out of quota, try sync write */
899 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
900 rc = vvp_io_commit_sync(env, io, queue,
901 vio->u.write.vui_from,
902 vio->u.write.vui_to);
904 vio->u.write.vui_written += rc;
909 /* update inode size */
910 ll_merge_attr(env, inode);
912 /* Now the pages in queue were failed to commit, discard them
913 * unless they were dirtied before.
915 while (queue->pl_nr > 0) {
916 page = cl_page_list_first(queue);
917 cl_page_list_del(env, queue, page);
919 if (!PageDirty(cl_page_vmpage(page)))
920 cl_page_discard(env, io, page);
922 cl_page_disown(env, io, page);
924 /* held in ll_cl_init() */
925 lu_ref_del(&page->cp_reference, "cl_io", io);
926 cl_page_put(env, page);
928 cl_page_list_fini(env, queue);
933 static int vvp_io_write_start(const struct lu_env *env,
934 const struct cl_io_slice *ios)
936 struct vvp_io *vio = cl2vvp_io(env, ios);
937 struct cl_io *io = ios->cis_io;
938 struct cl_object *obj = io->ci_obj;
939 struct inode *inode = vvp_object_inode(obj);
941 loff_t pos = io->u.ci_wr.wr.crw_pos;
942 size_t cnt = io->u.ci_wr.wr.crw_count;
944 if (!can_populate_pages(env, io, inode))
947 if (cl_io_is_append(io)) {
949 * PARALLEL IO This has to be changed for parallel IO doing
950 * out-of-order writes.
952 ll_merge_attr(env, inode);
953 pos = i_size_read(inode);
954 io->u.ci_wr.wr.crw_pos = pos;
955 vio->vui_iocb->ki_pos = pos;
957 LASSERT(vio->vui_iocb->ki_pos == pos);
960 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
962 if (!vio->vui_iter) /* from a temp io in ll_cl_init(). */
965 result = generic_file_write_iter(vio->vui_iocb, vio->vui_iter);
968 result = vvp_io_write_commit(env, io);
969 if (vio->u.write.vui_written > 0) {
970 result = vio->u.write.vui_written;
971 io->ci_nob += result;
973 CDEBUG(D_VFSTRACE, "write: nob %zd, result: %zd\n",
978 struct ll_inode_info *lli = ll_i2info(inode);
980 spin_lock(&lli->lli_lock);
981 lli->lli_flags |= LLIF_DATA_MODIFIED;
982 spin_unlock(&lli->lli_lock);
986 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
987 vio->vui_fd, pos, result, WRITE);
993 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
995 struct vm_fault *vmf = cfio->ft_vmf;
997 cfio->ft_flags = filemap_fault(cfio->ft_vma, vmf);
998 cfio->ft_flags_valid = 1;
1002 "page %p map %p index %lu flags %lx count %u priv %0lx: got addr %p type NOPAGE\n",
1003 vmf->page, vmf->page->mapping, vmf->page->index,
1004 (long)vmf->page->flags, page_count(vmf->page),
1005 page_private(vmf->page), vmf->virtual_address);
1006 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
1007 lock_page(vmf->page);
1008 cfio->ft_flags |= VM_FAULT_LOCKED;
1011 cfio->ft_vmpage = vmf->page;
1015 if (cfio->ft_flags & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV)) {
1016 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
1020 if (cfio->ft_flags & VM_FAULT_OOM) {
1021 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
1025 if (cfio->ft_flags & VM_FAULT_RETRY)
1028 CERROR("Unknown error in page fault %d!\n", cfio->ft_flags);
1032 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1033 struct cl_page *page)
1035 struct vvp_page *vpg;
1036 struct cl_object *clob = cl_io_top(io)->ci_obj;
1038 set_page_dirty(page->cp_vmpage);
1040 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
1041 vvp_write_pending(cl2vvp(clob), vpg);
1044 static int vvp_io_fault_start(const struct lu_env *env,
1045 const struct cl_io_slice *ios)
1047 struct vvp_io *vio = cl2vvp_io(env, ios);
1048 struct cl_io *io = ios->cis_io;
1049 struct cl_object *obj = io->ci_obj;
1050 struct inode *inode = vvp_object_inode(obj);
1051 struct cl_fault_io *fio = &io->u.ci_fault;
1052 struct vvp_fault_io *cfio = &vio->u.fault;
1055 struct page *vmpage = NULL;
1056 struct cl_page *page;
1060 if (fio->ft_executable &&
1061 inode->i_mtime.tv_sec != vio->u.fault.ft_mtime)
1063 " changed while waiting for the page fault lock\n",
1064 PFID(lu_object_fid(&obj->co_lu)));
1066 /* offset of the last byte on the page */
1067 offset = cl_offset(obj, fio->ft_index + 1) - 1;
1068 LASSERT(cl_index(obj, offset) == fio->ft_index);
1069 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1073 /* must return locked page */
1074 if (fio->ft_mkwrite) {
1075 LASSERT(cfio->ft_vmpage);
1076 lock_page(cfio->ft_vmpage);
1078 result = vvp_io_kernel_fault(cfio);
1083 vmpage = cfio->ft_vmpage;
1084 LASSERT(PageLocked(vmpage));
1086 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1087 ll_invalidate_page(vmpage);
1089 size = i_size_read(inode);
1090 /* Though we have already held a cl_lock upon this page, but
1091 * it still can be truncated locally.
1093 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1094 (page_offset(vmpage) > size))) {
1095 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1097 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1104 last_index = cl_index(obj, size - 1);
1106 if (fio->ft_mkwrite) {
1108 * Capture the size while holding the lli_trunc_sem from above
1109 * we want to make sure that we complete the mkwrite action
1110 * while holding this lock. We need to make sure that we are
1111 * not past the end of the file.
1113 if (last_index < fio->ft_index) {
1115 "llite: mkwrite and truncate race happened: %p: 0x%lx 0x%lx\n",
1116 vmpage->mapping, fio->ft_index, last_index);
1118 * We need to return if we are
1119 * passed the end of the file. This will propagate
1120 * up the call stack to ll_page_mkwrite where
1121 * we will return VM_FAULT_NOPAGE. Any non-negative
1122 * value returned here will be silently
1123 * converted to 0. If the vmpage->mapping is null
1124 * the error code would be converted back to ENODATA
1125 * in ll_page_mkwrite0. Thus we return -ENODATA
1126 * to handle both cases
1133 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1135 result = PTR_ERR(page);
1139 /* if page is going to be written, we should add this page into cache
1142 if (fio->ft_mkwrite) {
1143 wait_on_page_writeback(vmpage);
1144 if (!PageDirty(vmpage)) {
1145 struct cl_page_list *plist = &io->ci_queue.c2_qin;
1146 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1149 /* vvp_page_assume() calls wait_on_page_writeback(). */
1150 cl_page_assume(env, io, page);
1152 cl_page_list_init(plist);
1153 cl_page_list_add(plist, page);
1156 if (last_index == vvp_index(vpg))
1157 to = size & ~PAGE_MASK;
1159 /* Do not set Dirty bit here so that in case IO is
1160 * started before the page is really made dirty, we
1161 * still have chance to detect it.
1163 result = cl_io_commit_async(env, io, plist, 0, to,
1164 mkwrite_commit_callback);
1165 LASSERT(cl_page_is_owned(page, io));
1166 cl_page_list_fini(env, plist);
1170 cl_page_discard(env, io, page);
1171 cl_page_disown(env, io, page);
1173 cl_page_put(env, page);
1175 /* we're in big trouble, what can we do now? */
1176 if (result == -EDQUOT)
1180 cl_page_disown(env, io, page);
1186 * The ft_index is only used in the case of
1187 * a mkwrite action. We need to check
1188 * our assertions are correct, since
1189 * we should have caught this above
1191 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1192 if (fio->ft_index == last_index)
1194 * Last page is mapped partially.
1196 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1198 fio->ft_nob = cl_page_size(obj);
1200 lu_ref_add(&page->cp_reference, "fault", io);
1201 fio->ft_page = page;
1204 /* return unlocked vmpage to avoid deadlocking */
1206 unlock_page(vmpage);
1208 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1213 static int vvp_io_fsync_start(const struct lu_env *env,
1214 const struct cl_io_slice *ios)
1216 /* we should mark TOWRITE bit to each dirty page in radix tree to
1217 * verify pages have been written, but this is difficult because of
1223 static int vvp_io_read_page(const struct lu_env *env,
1224 const struct cl_io_slice *ios,
1225 const struct cl_page_slice *slice)
1227 struct cl_io *io = ios->cis_io;
1228 struct vvp_page *vpg = cl2vvp_page(slice);
1229 struct cl_page *page = slice->cpl_page;
1230 struct inode *inode = vvp_object_inode(slice->cpl_obj);
1231 struct ll_sb_info *sbi = ll_i2sbi(inode);
1232 struct ll_file_data *fd = cl2vvp_io(env, ios)->vui_fd;
1233 struct ll_readahead_state *ras = &fd->fd_ras;
1234 struct cl_2queue *queue = &io->ci_queue;
1236 if (sbi->ll_ra_info.ra_max_pages_per_file &&
1237 sbi->ll_ra_info.ra_max_pages)
1238 ras_update(sbi, inode, ras, vvp_index(vpg),
1239 vpg->vpg_defer_uptodate);
1241 if (vpg->vpg_defer_uptodate) {
1242 vpg->vpg_ra_used = 1;
1243 cl_page_export(env, page, 1);
1246 * Add page into the queue even when it is marked uptodate above.
1247 * this will unlock it automatically as part of cl_page_list_disown().
1250 cl_page_list_add(&queue->c2_qin, page);
1251 if (sbi->ll_ra_info.ra_max_pages_per_file &&
1252 sbi->ll_ra_info.ra_max_pages)
1253 ll_readahead(env, io, &queue->c2_qin, ras,
1254 vpg->vpg_defer_uptodate);
1259 static void vvp_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
1261 CLOBINVRNT(env, ios->cis_io->ci_obj,
1262 vvp_object_invariant(ios->cis_io->ci_obj));
1265 static const struct cl_io_operations vvp_io_ops = {
1268 .cio_fini = vvp_io_fini,
1269 .cio_lock = vvp_io_read_lock,
1270 .cio_start = vvp_io_read_start,
1271 .cio_advance = vvp_io_advance,
1274 .cio_fini = vvp_io_fini,
1275 .cio_iter_init = vvp_io_write_iter_init,
1276 .cio_iter_fini = vvp_io_write_iter_fini,
1277 .cio_lock = vvp_io_write_lock,
1278 .cio_start = vvp_io_write_start,
1279 .cio_advance = vvp_io_advance,
1282 .cio_fini = vvp_io_setattr_fini,
1283 .cio_iter_init = vvp_io_setattr_iter_init,
1284 .cio_lock = vvp_io_setattr_lock,
1285 .cio_start = vvp_io_setattr_start,
1286 .cio_end = vvp_io_setattr_end
1289 .cio_fini = vvp_io_fault_fini,
1290 .cio_iter_init = vvp_io_fault_iter_init,
1291 .cio_lock = vvp_io_fault_lock,
1292 .cio_start = vvp_io_fault_start,
1293 .cio_end = vvp_io_end,
1296 .cio_start = vvp_io_fsync_start,
1297 .cio_fini = vvp_io_fini
1300 .cio_fini = vvp_io_fini
1303 .cio_read_page = vvp_io_read_page,
1306 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1309 struct vvp_io *vio = vvp_env_io(env);
1310 struct inode *inode = vvp_object_inode(obj);
1313 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1315 CDEBUG(D_VFSTRACE, DFID
1316 " ignore/verify layout %d/%d, layout version %d restore needed %d\n",
1317 PFID(lu_object_fid(&obj->co_lu)),
1318 io->ci_ignore_layout, io->ci_verify_layout,
1319 vio->vui_layout_gen, io->ci_restore_needed);
1321 CL_IO_SLICE_CLEAN(vio, vui_cl);
1322 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1323 vio->vui_ra_valid = false;
1325 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1327 struct ll_inode_info *lli = ll_i2info(inode);
1329 count = io->u.ci_rw.crw_count;
1330 /* "If nbyte is 0, read() will return 0 and have no other
1331 * results." -- Single Unix Spec
1336 vio->vui_tot_count = count;
1338 /* for read/write, we store the jobid in the inode, and
1339 * it'll be fetched by osc when building RPC.
1341 * it's not accurate if the file is shared by different
1344 lustre_get_jobid(lli->lli_jobid);
1345 } else if (io->ci_type == CIT_SETATTR) {
1346 if (!cl_io_is_trunc(io))
1347 io->ci_lockreq = CILR_MANDATORY;
1350 /* ignore layout change for generic CIT_MISC but not for glimpse.
1351 * io context for glimpse must set ci_verify_layout to true,
1352 * see cl_glimpse_size0() for details.
1354 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1355 io->ci_ignore_layout = 1;
1357 /* Enqueue layout lock and get layout version. We need to do this
1358 * even for operations requiring to open file, such as read and write,
1359 * because it might not grant layout lock in IT_OPEN.
1361 if (result == 0 && !io->ci_ignore_layout) {
1362 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1363 if (result == -ENOENT)
1364 /* If the inode on MDS has been removed, but the objects
1365 * on OSTs haven't been destroyed (async unlink), layout
1366 * fetch will return -ENOENT, we'd ignore this error
1367 * and continue with dirty flush. LU-3230.
1371 CERROR("%s: refresh file layout " DFID " error %d.\n",
1372 ll_get_fsname(inode->i_sb, NULL, 0),
1373 PFID(lu_object_fid(&obj->co_lu)), result);