4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * Implementation of cl_io for VVP layer.
34 * Author: Nikita Danilov <nikita.danilov@sun.com>
35 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
38 #define DEBUG_SUBSYSTEM S_LLITE
42 #include "llite_internal.h"
43 #include "vvp_internal.h"
45 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
46 const struct cl_io_slice *slice)
50 vio = container_of(slice, struct vvp_io, vui_cl);
51 LASSERT(vio == vvp_env_io(env));
57 * True, if \a io is a normal io, False for splice_{read,write}
59 static int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
61 struct vvp_io *vio = vvp_env_io(env);
63 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
65 return vio->vui_io_subtype == IO_NORMAL;
69 * For swapping layout. The file's layout may have changed.
70 * To avoid populating pages to a wrong stripe, we have to verify the
71 * correctness of layout. It works because swapping layout processes
72 * have to acquire group lock.
74 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
77 struct ll_inode_info *lli = ll_i2info(inode);
78 struct vvp_io *vio = vvp_env_io(env);
81 switch (io->ci_type) {
84 /* don't need lock here to check lli_layout_gen as we have held
85 * extent lock and GROUP lock has to hold to swap layout */
86 if (ll_layout_version_get(lli) != vio->vui_layout_gen ||
87 OBD_FAIL_CHECK_RESET(OBD_FAIL_LLITE_LOST_LAYOUT, 0)) {
88 io->ci_need_restart = 1;
89 /* this will cause a short read/write */
94 /* fault is okay because we've already had a page. */
102 static void vvp_object_size_lock(struct cl_object *obj)
104 struct inode *inode = vvp_object_inode(obj);
106 ll_inode_size_lock(inode);
107 cl_object_attr_lock(obj);
110 static void vvp_object_size_unlock(struct cl_object *obj)
112 struct inode *inode = vvp_object_inode(obj);
114 cl_object_attr_unlock(obj);
115 ll_inode_size_unlock(inode);
119 * Helper function that if necessary adjusts file size (inode->i_size), when
120 * position at the offset \a pos is accessed. File size can be arbitrary stale
121 * on a Lustre client, but client at least knows KMS. If accessed area is
122 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
124 * Locking: i_size_lock is used to serialize changes to inode size and to
125 * protect consistency between inode size and cl_object
126 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
127 * top-object and sub-objects.
129 static int vvp_prep_size(const struct lu_env *env, struct cl_object *obj,
130 struct cl_io *io, loff_t start, size_t count,
133 struct cl_attr *attr = vvp_env_thread_attr(env);
134 struct inode *inode = vvp_object_inode(obj);
135 loff_t pos = start + count - 1;
140 * Consistency guarantees: following possibilities exist for the
141 * relation between region being accessed and real file size at this
144 * (A): the region is completely inside of the file;
146 * (B-x): x bytes of region are inside of the file, the rest is
149 * (C): the region is completely outside of the file.
151 * This classification is stable under DLM lock already acquired by
152 * the caller, because to change the class, other client has to take
153 * DLM lock conflicting with our lock. Also, any updates to ->i_size
154 * by other threads on this client are serialized by
155 * ll_inode_size_lock(). This guarantees that short reads are handled
156 * correctly in the face of concurrent writes and truncates.
158 vvp_object_size_lock(obj);
159 result = cl_object_attr_get(env, obj, attr);
164 * A glimpse is necessary to determine whether we
165 * return a short read (B) or some zeroes at the end
168 vvp_object_size_unlock(obj);
169 result = cl_glimpse_lock(env, io, inode, obj, 0);
170 if (result == 0 && exceed != NULL) {
171 /* If objective page index exceed end-of-file
172 * page index, return directly. Do not expect
173 * kernel will check such case correctly.
174 * linux-2.6.18-128.1.1 miss to do that.
176 loff_t size = i_size_read(inode);
177 unsigned long cur_index = start >>
180 if ((size == 0 && cur_index != 0) ||
181 (((size - 1) >> PAGE_SHIFT) <
189 * region is within kms and, hence, within real file
190 * size (A). We need to increase i_size to cover the
191 * read region so that generic_file_read() will do its
192 * job, but that doesn't mean the kms size is
193 * _correct_, it is only the _minimum_ size. If
194 * someone does a stat they will get the correct size
195 * which will always be >= the kms value here.
198 if (i_size_read(inode) < kms) {
199 i_size_write(inode, kms);
201 DFID" updating i_size %llu\n",
202 PFID(lu_object_fid(&obj->co_lu)),
203 (__u64)i_size_read(inode));
208 vvp_object_size_unlock(obj);
213 /*****************************************************************************
219 static int vvp_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
220 __u32 enqflags, enum cl_lock_mode mode,
221 pgoff_t start, pgoff_t end)
223 struct vvp_io *vio = vvp_env_io(env);
224 struct cl_lock_descr *descr = &vio->vui_link.cill_descr;
225 struct cl_object *obj = io->ci_obj;
227 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
230 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
232 memset(&vio->vui_link, 0, sizeof vio->vui_link);
234 if (vio->vui_fd && (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
235 descr->cld_mode = CLM_GROUP;
236 descr->cld_gid = vio->vui_fd->fd_grouplock.lg_gid;
237 enqflags |= CEF_LOCK_MATCH;
239 descr->cld_mode = mode;
242 descr->cld_obj = obj;
243 descr->cld_start = start;
244 descr->cld_end = end;
245 descr->cld_enq_flags = enqflags;
247 cl_io_lock_add(env, io, &vio->vui_link);
252 static int vvp_io_one_lock(const struct lu_env *env, struct cl_io *io,
253 __u32 enqflags, enum cl_lock_mode mode,
254 loff_t start, loff_t end)
256 struct cl_object *obj = io->ci_obj;
258 return vvp_io_one_lock_index(env, io, enqflags, mode,
259 cl_index(obj, start), cl_index(obj, end));
262 static int vvp_io_write_iter_init(const struct lu_env *env,
263 const struct cl_io_slice *ios)
265 struct vvp_io *vio = cl2vvp_io(env, ios);
267 cl_page_list_init(&vio->u.write.vui_queue);
268 vio->u.write.vui_written = 0;
269 vio->u.write.vui_from = 0;
270 vio->u.write.vui_to = PAGE_SIZE;
275 static void vvp_io_write_iter_fini(const struct lu_env *env,
276 const struct cl_io_slice *ios)
278 struct vvp_io *vio = cl2vvp_io(env, ios);
280 LASSERT(vio->u.write.vui_queue.pl_nr == 0);
283 static int vvp_io_fault_iter_init(const struct lu_env *env,
284 const struct cl_io_slice *ios)
286 struct vvp_io *vio = cl2vvp_io(env, ios);
287 struct inode *inode = vvp_object_inode(ios->cis_obj);
289 LASSERT(inode == file_inode(vio->vui_fd->fd_file));
290 vio->u.fault.ft_mtime = inode->i_mtime.tv_sec;
295 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
297 struct cl_io *io = ios->cis_io;
298 struct cl_object *obj = io->ci_obj;
299 struct vvp_io *vio = cl2vvp_io(env, ios);
300 struct inode *inode = vvp_object_inode(obj);
303 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
305 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
306 "need write layout %d, restore needed %d\n",
307 PFID(lu_object_fid(&obj->co_lu)),
308 io->ci_ignore_layout, io->ci_verify_layout,
309 vio->vui_layout_gen, io->ci_need_write_intent,
310 io->ci_restore_needed);
312 if (io->ci_restore_needed) {
313 /* file was detected release, we need to restore it
314 * before finishing the io
316 rc = ll_layout_restore(inode, 0, OBD_OBJECT_EOF);
317 /* if restore registration failed, no restart,
318 * we will return -ENODATA */
319 /* The layout will change after restore, so we need to
320 * block on layout lock held by the MDT
321 * as MDT will not send new layout in lvb (see LU-3124)
322 * we have to explicitly fetch it, all this will be done
323 * by ll_layout_refresh()
326 io->ci_restore_needed = 0;
327 io->ci_need_restart = 1;
328 io->ci_verify_layout = 1;
330 io->ci_restore_needed = 1;
331 io->ci_need_restart = 0;
332 io->ci_verify_layout = 0;
338 * dynamic layout change needed, send layout intent
341 if (io->ci_need_write_intent) {
345 LASSERT(io->ci_type == CIT_WRITE || cl_io_is_trunc(io));
347 io->ci_need_write_intent = 0;
349 if (io->ci_type == CIT_WRITE) {
350 start = io->u.ci_rw.crw_pos;
351 end = io->u.ci_rw.crw_pos + io->u.ci_rw.crw_count;
353 end = io->u.ci_setattr.sa_attr.lvb_size;
356 rc = ll_layout_write_intent(inode, start, end);
359 io->ci_need_restart = 1;
362 if (!io->ci_ignore_layout && io->ci_verify_layout) {
365 /* check layout version */
366 ll_layout_refresh(inode, &gen);
367 io->ci_need_restart = vio->vui_layout_gen != gen;
368 if (io->ci_need_restart) {
370 DFID" layout changed from %d to %d.\n",
371 PFID(lu_object_fid(&obj->co_lu)),
372 vio->vui_layout_gen, gen);
373 /* today successful restore is the only possible
375 /* restore was done, clear restoring state */
376 ll_file_clear_flag(ll_i2info(vvp_object_inode(obj)),
377 LLIF_FILE_RESTORING);
382 static void vvp_io_fault_fini(const struct lu_env *env,
383 const struct cl_io_slice *ios)
385 struct cl_io *io = ios->cis_io;
386 struct cl_page *page = io->u.ci_fault.ft_page;
388 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
391 lu_ref_del(&page->cp_reference, "fault", io);
392 cl_page_put(env, page);
393 io->u.ci_fault.ft_page = NULL;
395 vvp_io_fini(env, ios);
398 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
401 * we only want to hold PW locks if the mmap() can generate
402 * writes back to the file and that only happens in shared
405 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
410 static int vvp_mmap_locks(const struct lu_env *env,
411 struct vvp_io *vio, struct cl_io *io)
413 struct vvp_thread_info *vti = vvp_env_info(env);
414 struct mm_struct *mm = current->mm;
415 struct vm_area_struct *vma;
416 struct cl_lock_descr *descr = &vti->vti_descr;
417 union ldlm_policy_data policy;
423 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
425 if (!cl_is_normalio(env, io))
428 /* nfs or loop back device write */
429 if (vio->vui_iter == NULL)
432 /* No MM (e.g. NFS)? No vmas too. */
436 iov_for_each(iov, i, *(vio->vui_iter)) {
437 unsigned long addr = (unsigned long)iov.iov_base;
438 size_t count = iov.iov_len;
443 count += addr & ~PAGE_MASK;
446 down_read(&mm->mmap_sem);
447 while((vma = our_vma(mm, addr, count)) != NULL) {
448 struct dentry *de = file_dentry(vma->vm_file);
449 struct inode *inode = de->d_inode;
450 int flags = CEF_MUST;
452 if (ll_file_nolock(vma->vm_file)) {
454 * For no lock case is not allowed for mmap
461 * XXX: Required lock mode can be weakened: CIT_WRITE
462 * io only ever reads user level buffer, and CIT_READ
465 policy_from_vma(&policy, vma, addr, count);
466 descr->cld_mode = vvp_mode_from_vma(vma);
467 descr->cld_obj = ll_i2info(inode)->lli_clob;
468 descr->cld_start = cl_index(descr->cld_obj,
469 policy.l_extent.start);
470 descr->cld_end = cl_index(descr->cld_obj,
471 policy.l_extent.end);
472 descr->cld_enq_flags = flags;
473 result = cl_io_lock_alloc_add(env, io, descr);
475 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
476 descr->cld_mode, descr->cld_start,
482 if (vma->vm_end - addr >= count)
485 count -= vma->vm_end - addr;
488 up_read(&mm->mmap_sem);
495 static void vvp_io_advance(const struct lu_env *env,
496 const struct cl_io_slice *ios,
499 struct vvp_io *vio = cl2vvp_io(env, ios);
500 struct cl_io *io = ios->cis_io;
501 struct cl_object *obj = ios->cis_io->ci_obj;
503 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
505 if (!cl_is_normalio(env, io))
508 vio->vui_tot_count -= nob;
509 iov_iter_reexpand(vio->vui_iter, vio->vui_tot_count);
512 static void vvp_io_update_iov(const struct lu_env *env,
513 struct vvp_io *vio, struct cl_io *io)
515 size_t size = io->u.ci_rw.crw_count;
517 if (!cl_is_normalio(env, io) || vio->vui_iter == NULL)
520 iov_iter_truncate(vio->vui_iter, size);
523 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
524 enum cl_lock_mode mode, loff_t start, loff_t end)
526 struct vvp_io *vio = vvp_env_io(env);
530 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
533 vvp_io_update_iov(env, vio, io);
535 if (io->u.ci_rw.crw_nonblock)
536 ast_flags |= CEF_NONBLOCK;
538 result = vvp_mmap_locks(env, vio, io);
540 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
545 static int vvp_io_read_lock(const struct lu_env *env,
546 const struct cl_io_slice *ios)
548 struct cl_io *io = ios->cis_io;
549 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
553 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
554 rd->crw_pos + rd->crw_count - 1);
558 static int vvp_io_fault_lock(const struct lu_env *env,
559 const struct cl_io_slice *ios)
561 struct cl_io *io = ios->cis_io;
562 struct vvp_io *vio = cl2vvp_io(env, ios);
564 * XXX LDLM_FL_CBPENDING
566 return vvp_io_one_lock_index(env,
568 vvp_mode_from_vma(vio->u.fault.ft_vma),
569 io->u.ci_fault.ft_index,
570 io->u.ci_fault.ft_index);
573 static int vvp_io_write_lock(const struct lu_env *env,
574 const struct cl_io_slice *ios)
576 struct cl_io *io = ios->cis_io;
580 if (io->u.ci_wr.wr_append) {
582 end = OBD_OBJECT_EOF;
584 start = io->u.ci_wr.wr.crw_pos;
585 end = start + io->u.ci_wr.wr.crw_count - 1;
587 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
590 static int vvp_io_setattr_iter_init(const struct lu_env *env,
591 const struct cl_io_slice *ios)
597 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
599 * Handles "lockless io" mode when extent locking is done by server.
601 static int vvp_io_setattr_lock(const struct lu_env *env,
602 const struct cl_io_slice *ios)
604 struct cl_io *io = ios->cis_io;
608 if (cl_io_is_trunc(io)) {
609 new_size = io->u.ci_setattr.sa_attr.lvb_size;
611 enqflags = CEF_DISCARD_DATA;
613 unsigned int valid = io->u.ci_setattr.sa_valid;
615 if (!(valid & TIMES_SET_FLAGS))
618 if ((!(valid & ATTR_MTIME) ||
619 io->u.ci_setattr.sa_attr.lvb_mtime >=
620 io->u.ci_setattr.sa_attr.lvb_ctime) &&
621 (!(valid & ATTR_ATIME) ||
622 io->u.ci_setattr.sa_attr.lvb_atime >=
623 io->u.ci_setattr.sa_attr.lvb_ctime))
629 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
630 new_size, OBD_OBJECT_EOF);
633 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
638 * Only ll_inode_size_lock is taken at this level.
640 ll_inode_size_lock(inode);
641 result = inode_newsize_ok(inode, size);
643 ll_inode_size_unlock(inode);
646 i_size_write(inode, size);
648 ll_truncate_pagecache(inode, size);
649 ll_inode_size_unlock(inode);
653 static int vvp_io_setattr_time(const struct lu_env *env,
654 const struct cl_io_slice *ios)
656 struct cl_io *io = ios->cis_io;
657 struct cl_object *obj = io->ci_obj;
658 struct cl_attr *attr = vvp_env_thread_attr(env);
660 unsigned valid = CAT_CTIME;
662 cl_object_attr_lock(obj);
663 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
664 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
665 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
668 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
669 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
672 result = cl_object_attr_update(env, obj, attr, valid);
673 cl_object_attr_unlock(obj);
678 static int vvp_io_setattr_start(const struct lu_env *env,
679 const struct cl_io_slice *ios)
681 struct cl_io *io = ios->cis_io;
682 struct inode *inode = vvp_object_inode(io->ci_obj);
683 struct ll_inode_info *lli = ll_i2info(inode);
685 if (cl_io_is_trunc(io)) {
686 down_write(&lli->lli_trunc_sem);
688 inode_dio_wait(inode);
693 if (io->u.ci_setattr.sa_valid & TIMES_SET_FLAGS)
694 return vvp_io_setattr_time(env, ios);
699 static void vvp_io_setattr_end(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);
704 struct ll_inode_info *lli = ll_i2info(inode);
706 if (cl_io_is_trunc(io)) {
707 /* Truncate in memory pages - they must be clean pages
708 * because osc has already notified to destroy osc_extents. */
709 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
710 inode_dio_write_done(inode);
712 up_write(&lli->lli_trunc_sem);
718 static void vvp_io_setattr_fini(const struct lu_env *env,
719 const struct cl_io_slice *ios)
721 bool restore_needed = ios->cis_io->ci_restore_needed;
722 struct inode *inode = vvp_object_inode(ios->cis_obj);
724 vvp_io_fini(env, ios);
726 if (restore_needed && !ios->cis_io->ci_restore_needed) {
727 /* restore finished, set data modified flag for HSM */
728 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
732 static int vvp_io_read_start(const struct lu_env *env,
733 const struct cl_io_slice *ios)
735 struct vvp_io *vio = cl2vvp_io(env, ios);
736 struct cl_io *io = ios->cis_io;
737 struct cl_object *obj = io->ci_obj;
738 struct inode *inode = vvp_object_inode(obj);
739 struct ll_inode_info *lli = ll_i2info(inode);
740 struct file *file = vio->vui_fd->fd_file;
743 loff_t pos = io->u.ci_rd.rd.crw_pos;
744 long cnt = io->u.ci_rd.rd.crw_count;
745 long tot = vio->vui_tot_count;
748 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
750 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
752 if (vio->vui_io_subtype == IO_NORMAL)
753 down_read(&lli->lli_trunc_sem);
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_SIZE - 1);
781 switch (vio->vui_io_subtype) {
783 LASSERT(vio->vui_iocb->ki_pos == pos);
784 result = generic_file_read_iter(vio->vui_iocb, vio->vui_iter);
787 result = generic_file_splice_read(file, &pos,
788 vio->u.splice.vui_pipe, cnt,
789 vio->u.splice.vui_flags);
790 /* LU-1109: do splice read stripe by stripe otherwise if it
791 * may make nfsd stuck if this read occupied all internal pipe
796 CERROR("Wrong IO type %u\n", vio->vui_io_subtype);
804 io->ci_nob += result;
805 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, vio->vui_fd,
813 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
814 struct cl_page_list *plist, int from, int to)
816 struct cl_2queue *queue = &io->ci_queue;
817 struct cl_page *page;
818 unsigned int bytes = 0;
822 if (plist->pl_nr == 0)
825 if (from > 0 || to != PAGE_SIZE) {
826 page = cl_page_list_first(plist);
827 if (plist->pl_nr == 1) {
828 cl_page_clip(env, page, from, to);
831 cl_page_clip(env, page, from, PAGE_SIZE);
832 if (to != PAGE_SIZE) {
833 page = cl_page_list_last(plist);
834 cl_page_clip(env, page, 0, to);
839 cl_2queue_init(queue);
840 cl_page_list_splice(plist, &queue->c2_qin);
841 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
843 /* plist is not sorted any more */
844 cl_page_list_splice(&queue->c2_qin, plist);
845 cl_page_list_splice(&queue->c2_qout, plist);
846 cl_2queue_fini(env, queue);
849 /* calculate bytes */
850 bytes = plist->pl_nr << PAGE_SHIFT;
851 bytes -= from + PAGE_SIZE - to;
853 while (plist->pl_nr > 0) {
854 page = cl_page_list_first(plist);
855 cl_page_list_del(env, plist, page);
857 cl_page_clip(env, page, 0, PAGE_SIZE);
859 SetPageUptodate(cl_page_vmpage(page));
860 cl_page_disown(env, io, page);
862 /* held in ll_cl_init() */
863 lu_ref_del(&page->cp_reference, "cl_io", io);
864 cl_page_put(env, page);
868 RETURN(bytes > 0 ? bytes : rc);
871 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
872 struct cl_page *page)
874 struct page *vmpage = page->cp_vmpage;
876 SetPageUptodate(vmpage);
877 set_page_dirty(vmpage);
879 cl_page_disown(env, io, page);
881 /* held in ll_cl_init() */
882 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
883 cl_page_put(env, page);
886 /* make sure the page list is contiguous */
887 static bool page_list_sanity_check(struct cl_object *obj,
888 struct cl_page_list *plist)
890 struct cl_page *page;
891 pgoff_t index = CL_PAGE_EOF;
893 cl_page_list_for_each(page, plist) {
894 struct vvp_page *vpg = cl_object_page_slice(obj, page);
896 if (index == CL_PAGE_EOF) {
897 index = vvp_index(vpg);
902 if (index == vvp_index(vpg))
910 /* Return how many bytes have queued or written */
911 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
913 struct cl_object *obj = io->ci_obj;
914 struct inode *inode = vvp_object_inode(obj);
915 struct vvp_io *vio = vvp_env_io(env);
916 struct cl_page_list *queue = &vio->u.write.vui_queue;
917 struct cl_page *page;
920 unsigned int npages = vio->u.write.vui_queue.pl_nr;
926 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
927 npages, vio->u.write.vui_from, vio->u.write.vui_to);
929 LASSERT(page_list_sanity_check(obj, queue));
931 /* submit IO with async write */
932 rc = cl_io_commit_async(env, io, queue,
933 vio->u.write.vui_from, vio->u.write.vui_to,
934 write_commit_callback);
935 npages -= queue->pl_nr; /* already committed pages */
937 /* calculate how many bytes were written */
938 bytes = npages << PAGE_SHIFT;
941 bytes -= vio->u.write.vui_from;
942 if (queue->pl_nr == 0) /* last page */
943 bytes -= PAGE_SIZE - vio->u.write.vui_to;
944 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
946 vio->u.write.vui_written += bytes;
948 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
949 npages, bytes, vio->u.write.vui_written);
951 /* the first page must have been written. */
952 vio->u.write.vui_from = 0;
954 LASSERT(page_list_sanity_check(obj, queue));
955 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
957 /* out of quota, try sync write */
958 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
959 rc = vvp_io_commit_sync(env, io, queue,
960 vio->u.write.vui_from,
961 vio->u.write.vui_to);
963 vio->u.write.vui_written += rc;
968 /* update inode size */
969 ll_merge_attr(env, inode);
971 /* Now the pages in queue were failed to commit, discard them
972 * unless they were dirtied before. */
973 while (queue->pl_nr > 0) {
974 page = cl_page_list_first(queue);
975 cl_page_list_del(env, queue, page);
977 if (!PageDirty(cl_page_vmpage(page)))
978 cl_page_discard(env, io, page);
980 cl_page_disown(env, io, page);
982 /* held in ll_cl_init() */
983 lu_ref_del(&page->cp_reference, "cl_io", io);
984 cl_page_put(env, page);
986 cl_page_list_fini(env, queue);
991 static int vvp_io_write_start(const struct lu_env *env,
992 const struct cl_io_slice *ios)
994 struct vvp_io *vio = cl2vvp_io(env, ios);
995 struct cl_io *io = ios->cis_io;
996 struct cl_object *obj = io->ci_obj;
997 struct inode *inode = vvp_object_inode(obj);
998 struct ll_inode_info *lli = ll_i2info(inode);
1000 loff_t pos = io->u.ci_wr.wr.crw_pos;
1001 size_t cnt = io->u.ci_wr.wr.crw_count;
1005 if (vio->vui_io_subtype == IO_NORMAL)
1006 down_read(&lli->lli_trunc_sem);
1008 if (!can_populate_pages(env, io, inode))
1011 if (cl_io_is_append(io)) {
1013 * PARALLEL IO This has to be changed for parallel IO doing
1014 * out-of-order writes.
1016 ll_merge_attr(env, inode);
1017 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
1018 vio->vui_iocb->ki_pos = pos;
1020 LASSERT(vio->vui_iocb->ki_pos == pos);
1023 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
1025 /* The maximum Lustre file size is variable, based on the OST maximum
1026 * object size and number of stripes. This needs another check in
1027 * addition to the VFS checks earlier. */
1028 if (pos + cnt > ll_file_maxbytes(inode)) {
1030 "%s: file "DFID" offset %llu > maxbytes %llu\n",
1031 ll_get_fsname(inode->i_sb, NULL, 0),
1032 PFID(ll_inode2fid(inode)), pos + cnt,
1033 ll_file_maxbytes(inode));
1037 if (vio->vui_iter == NULL) {
1038 /* from a temp io in ll_cl_init(). */
1042 * When using the locked AIO function (generic_file_aio_write())
1043 * testing has shown the inode mutex to be a limiting factor
1044 * with multi-threaded single shared file performance. To get
1045 * around this, we now use the lockless version. To maintain
1046 * consistency, proper locking to protect against writes,
1047 * trucates, etc. is handled in the higher layers of lustre.
1049 bool lock_node = !IS_NOSEC(inode);
1053 result = __generic_file_write_iter(vio->vui_iocb,
1056 inode_unlock(inode);
1058 if (result > 0 || result == -EIOCBQUEUED) {
1061 err = generic_write_sync(vio->vui_iocb->ki_filp,
1063 if (err < 0 && result > 0)
1069 result = vvp_io_write_commit(env, io);
1070 if (vio->u.write.vui_written > 0) {
1071 result = vio->u.write.vui_written;
1072 io->ci_nob += result;
1074 CDEBUG(D_VFSTRACE, "write: nob %zd, result: %zd\n",
1075 io->ci_nob, result);
1079 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
1082 io->ci_continue = 0;
1083 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
1084 vio->vui_fd, pos, result, WRITE);
1091 static void vvp_io_rw_end(const struct lu_env *env,
1092 const struct cl_io_slice *ios)
1094 struct vvp_io *vio = cl2vvp_io(env, ios);
1095 struct inode *inode = vvp_object_inode(ios->cis_obj);
1096 struct ll_inode_info *lli = ll_i2info(inode);
1098 if (vio->vui_io_subtype == IO_NORMAL)
1099 up_read(&lli->lli_trunc_sem);
1102 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
1104 struct vm_fault *vmf = cfio->ft_vmf;
1106 cfio->ft_flags = filemap_fault(cfio->ft_vma, vmf);
1107 cfio->ft_flags_valid = 1;
1110 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
1111 vmf->virtual_address);
1112 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
1113 lock_page(vmf->page);
1114 cfio->ft_flags |= VM_FAULT_LOCKED;
1117 cfio->ft_vmpage = vmf->page;
1122 if (cfio->ft_flags & VM_FAULT_SIGBUS) {
1123 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
1127 if (cfio->ft_flags & VM_FAULT_OOM) {
1128 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
1132 if (cfio->ft_flags & VM_FAULT_RETRY)
1135 CERROR("unknown error in page fault %d\n", cfio->ft_flags);
1140 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1141 struct cl_page *page)
1143 set_page_dirty(page->cp_vmpage);
1146 static int vvp_io_fault_start(const struct lu_env *env,
1147 const struct cl_io_slice *ios)
1149 struct vvp_io *vio = cl2vvp_io(env, ios);
1150 struct cl_io *io = ios->cis_io;
1151 struct cl_object *obj = io->ci_obj;
1152 struct inode *inode = vvp_object_inode(obj);
1153 struct ll_inode_info *lli = ll_i2info(inode);
1154 struct cl_fault_io *fio = &io->u.ci_fault;
1155 struct vvp_fault_io *cfio = &vio->u.fault;
1158 struct page *vmpage = NULL;
1159 struct cl_page *page;
1164 down_read(&lli->lli_trunc_sem);
1166 /* offset of the last byte on the page */
1167 offset = cl_offset(obj, fio->ft_index + 1) - 1;
1168 LASSERT(cl_index(obj, offset) == fio->ft_index);
1169 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1173 /* must return locked page */
1174 if (fio->ft_mkwrite) {
1175 LASSERT(cfio->ft_vmpage != NULL);
1176 lock_page(cfio->ft_vmpage);
1178 result = vvp_io_kernel_fault(cfio);
1183 vmpage = cfio->ft_vmpage;
1184 LASSERT(PageLocked(vmpage));
1186 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1187 ll_invalidate_page(vmpage);
1189 size = i_size_read(inode);
1190 /* Though we have already held a cl_lock upon this page, but
1191 * it still can be truncated locally. */
1192 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1193 (page_offset(vmpage) > size))) {
1194 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1196 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1198 GOTO(out, result = +1);
1201 last_index = cl_index(obj, size - 1);
1203 if (fio->ft_mkwrite ) {
1205 * Capture the size while holding the lli_trunc_sem from above
1206 * we want to make sure that we complete the mkwrite action
1207 * while holding this lock. We need to make sure that we are
1208 * not past the end of the file.
1210 if (last_index < fio->ft_index) {
1212 "llite: mkwrite and truncate race happened: "
1213 "%p: 0x%lx 0x%lx\n",
1214 vmpage->mapping,fio->ft_index,last_index);
1216 * We need to return if we are
1217 * passed the end of the file. This will propagate
1218 * up the call stack to ll_page_mkwrite where
1219 * we will return VM_FAULT_NOPAGE. Any non-negative
1220 * value returned here will be silently
1221 * converted to 0. If the vmpage->mapping is null
1222 * the error code would be converted back to ENODATA
1223 * in ll_page_mkwrite0. Thus we return -ENODATA
1224 * to handle both cases
1226 GOTO(out, result = -ENODATA);
1230 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1232 GOTO(out, result = PTR_ERR(page));
1234 /* if page is going to be written, we should add this page into cache
1236 if (fio->ft_mkwrite) {
1237 wait_on_page_writeback(vmpage);
1238 if (!PageDirty(vmpage)) {
1239 struct cl_page_list *plist = &io->ci_queue.c2_qin;
1240 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1243 /* vvp_page_assume() calls wait_on_page_writeback(). */
1244 cl_page_assume(env, io, page);
1246 cl_page_list_init(plist);
1247 cl_page_list_add(plist, page);
1250 if (last_index == vvp_index(vpg))
1251 to = size & ~PAGE_MASK;
1253 /* Do not set Dirty bit here so that in case IO is
1254 * started before the page is really made dirty, we
1255 * still have chance to detect it. */
1256 result = cl_io_commit_async(env, io, plist, 0, to,
1257 mkwrite_commit_callback);
1258 LASSERT(cl_page_is_owned(page, io));
1259 cl_page_list_fini(env, plist);
1263 cl_page_discard(env, io, page);
1264 cl_page_disown(env, io, page);
1266 cl_page_put(env, page);
1268 /* we're in big trouble, what can we do now? */
1269 if (result == -EDQUOT)
1273 cl_page_disown(env, io, page);
1278 * The ft_index is only used in the case of
1279 * a mkwrite action. We need to check
1280 * our assertions are correct, since
1281 * we should have caught this above
1283 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1284 if (fio->ft_index == last_index)
1286 * Last page is mapped partially.
1288 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1290 fio->ft_nob = cl_page_size(obj);
1292 lu_ref_add(&page->cp_reference, "fault", io);
1293 fio->ft_page = page;
1297 /* return unlocked vmpage to avoid deadlocking */
1299 unlock_page(vmpage);
1301 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1306 static void vvp_io_fault_end(const struct lu_env *env,
1307 const struct cl_io_slice *ios)
1309 struct inode *inode = vvp_object_inode(ios->cis_obj);
1310 struct ll_inode_info *lli = ll_i2info(inode);
1312 CLOBINVRNT(env, ios->cis_io->ci_obj,
1313 vvp_object_invariant(ios->cis_io->ci_obj));
1314 up_read(&lli->lli_trunc_sem);
1317 static int vvp_io_fsync_start(const struct lu_env *env,
1318 const struct cl_io_slice *ios)
1320 /* we should mark TOWRITE bit to each dirty page in radix tree to
1321 * verify pages have been written, but this is difficult because of
1326 static int vvp_io_read_ahead(const struct lu_env *env,
1327 const struct cl_io_slice *ios,
1328 pgoff_t start, struct cl_read_ahead *ra)
1333 if (ios->cis_io->ci_type == CIT_READ ||
1334 ios->cis_io->ci_type == CIT_FAULT) {
1335 struct vvp_io *vio = cl2vvp_io(env, ios);
1337 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1338 ra->cra_end = CL_PAGE_EOF;
1339 result = +1; /* no need to call down */
1346 static const struct cl_io_operations vvp_io_ops = {
1349 .cio_fini = vvp_io_fini,
1350 .cio_lock = vvp_io_read_lock,
1351 .cio_start = vvp_io_read_start,
1352 .cio_end = vvp_io_rw_end,
1353 .cio_advance = vvp_io_advance,
1356 .cio_fini = vvp_io_fini,
1357 .cio_iter_init = vvp_io_write_iter_init,
1358 .cio_iter_fini = vvp_io_write_iter_fini,
1359 .cio_lock = vvp_io_write_lock,
1360 .cio_start = vvp_io_write_start,
1361 .cio_end = vvp_io_rw_end,
1362 .cio_advance = vvp_io_advance,
1365 .cio_fini = vvp_io_setattr_fini,
1366 .cio_iter_init = vvp_io_setattr_iter_init,
1367 .cio_lock = vvp_io_setattr_lock,
1368 .cio_start = vvp_io_setattr_start,
1369 .cio_end = vvp_io_setattr_end
1372 .cio_fini = vvp_io_fault_fini,
1373 .cio_iter_init = vvp_io_fault_iter_init,
1374 .cio_lock = vvp_io_fault_lock,
1375 .cio_start = vvp_io_fault_start,
1376 .cio_end = vvp_io_fault_end,
1379 .cio_start = vvp_io_fsync_start,
1380 .cio_fini = vvp_io_fini
1383 .cio_fini = vvp_io_fini
1386 .cio_fini = vvp_io_fini
1389 .cio_read_ahead = vvp_io_read_ahead
1392 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1395 struct vvp_io *vio = vvp_env_io(env);
1396 struct inode *inode = vvp_object_inode(obj);
1399 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1402 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1403 "restore needed %d\n",
1404 PFID(lu_object_fid(&obj->co_lu)),
1405 io->ci_ignore_layout, io->ci_verify_layout,
1406 vio->vui_layout_gen, io->ci_restore_needed);
1408 CL_IO_SLICE_CLEAN(vio, vui_cl);
1409 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1410 vio->vui_ra_valid = false;
1412 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1414 struct ll_inode_info *lli = ll_i2info(inode);
1416 count = io->u.ci_rw.crw_count;
1417 /* "If nbyte is 0, read() will return 0 and have no other
1418 * results." -- Single Unix Spec */
1422 vio->vui_tot_count = count;
1424 /* for read/write, we store the jobid in the inode, and
1425 * it'll be fetched by osc when building RPC.
1427 * it's not accurate if the file is shared by different
1430 lustre_get_jobid(lli->lli_jobid);
1431 } else if (io->ci_type == CIT_SETATTR) {
1432 if (!cl_io_is_trunc(io))
1433 io->ci_lockreq = CILR_MANDATORY;
1436 /* Enqueue layout lock and get layout version. We need to do this
1437 * even for operations requiring to open file, such as read and write,
1438 * because it might not grant layout lock in IT_OPEN. */
1439 if (result == 0 && !io->ci_ignore_layout) {
1440 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1441 if (result == -ENOENT)
1442 /* If the inode on MDS has been removed, but the objects
1443 * on OSTs haven't been destroyed (async unlink), layout
1444 * fetch will return -ENOENT, we'd ingore this error
1445 * and continue with dirty flush. LU-3230. */
1448 CERROR("%s: refresh file layout " DFID " error %d.\n",
1449 ll_get_fsname(inode->i_sb, NULL, 0),
1450 PFID(lu_object_fid(&obj->co_lu)), result);