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, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 * Implementation of cl_io for VVP layer.
33 * Author: Nikita Danilov <nikita.danilov@sun.com>
34 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
37 #define DEBUG_SUBSYSTEM S_LLITE
40 #include <linux/pagevec.h>
41 #include <linux/memcontrol.h>
42 #include <linux/falloc.h>
44 #include "llite_internal.h"
45 #include "vvp_internal.h"
46 #include <libcfs/linux/linux-misc.h>
48 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
49 const struct cl_io_slice *slice)
53 vio = container_of(slice, struct vvp_io, vui_cl);
54 LASSERT(vio == vvp_env_io(env));
60 * For swapping layout. The file's layout may have changed.
61 * To avoid populating pages to a wrong stripe, we have to verify the
62 * correctness of layout. It works because swapping layout processes
63 * have to acquire group lock.
65 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
68 struct ll_inode_info *lli = ll_i2info(inode);
69 struct vvp_io *vio = vvp_env_io(env);
72 switch (io->ci_type) {
75 /* don't need lock here to check lli_layout_gen as we have held
76 * extent lock and GROUP lock has to hold to swap layout */
77 if (ll_layout_version_get(lli) != vio->vui_layout_gen ||
78 OBD_FAIL_CHECK_RESET(OBD_FAIL_LLITE_LOST_LAYOUT, 0)) {
79 io->ci_need_restart = 1;
80 /* this will cause a short read/write */
85 /* fault is okay because we've already had a page. */
93 static void vvp_object_size_lock(struct cl_object *obj)
95 struct inode *inode = vvp_object_inode(obj);
97 ll_inode_size_lock(inode);
98 cl_object_attr_lock(obj);
101 static void vvp_object_size_unlock(struct cl_object *obj)
103 struct inode *inode = vvp_object_inode(obj);
105 cl_object_attr_unlock(obj);
106 ll_inode_size_unlock(inode);
110 * Helper function that if necessary adjusts file size (inode->i_size), when
111 * position at the offset \a pos is accessed. File size can be arbitrary stale
112 * on a Lustre client, but client at least knows KMS. If accessed area is
113 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
115 * Locking: i_size_lock is used to serialize changes to inode size and to
116 * protect consistency between inode size and cl_object
117 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
118 * top-object and sub-objects.
120 static int vvp_prep_size(const struct lu_env *env, struct cl_object *obj,
121 struct cl_io *io, loff_t start, size_t count,
124 struct cl_attr *attr = vvp_env_thread_attr(env);
125 struct inode *inode = vvp_object_inode(obj);
126 loff_t pos = start + count - 1;
131 * Consistency guarantees: following possibilities exist for the
132 * relation between region being accessed and real file size at this
135 * (A): the region is completely inside of the file;
137 * (B-x): x bytes of region are inside of the file, the rest is
140 * (C): the region is completely outside of the file.
142 * This classification is stable under DLM lock already acquired by
143 * the caller, because to change the class, other client has to take
144 * DLM lock conflicting with our lock. Also, any updates to ->i_size
145 * by other threads on this client are serialized by
146 * ll_inode_size_lock(). This guarantees that short reads are handled
147 * correctly in the face of concurrent writes and truncates.
149 vvp_object_size_lock(obj);
150 result = cl_object_attr_get(env, obj, attr);
155 * A glimpse is necessary to determine whether we
156 * return a short read (B) or some zeroes at the end
159 vvp_object_size_unlock(obj);
160 result = cl_glimpse_lock(env, io, inode, obj, 0);
161 if (result == 0 && exceed != NULL) {
162 /* If objective page index exceed end-of-file
163 * page index, return directly. Do not expect
164 * kernel will check such case correctly.
165 * linux-2.6.18-128.1.1 miss to do that.
167 loff_t size = i_size_read(inode);
168 unsigned long cur_index = start >>
171 if ((size == 0 && cur_index != 0) ||
172 (((size - 1) >> PAGE_SHIFT) <
180 * region is within kms and, hence, within real file
181 * size (A). We need to increase i_size to cover the
182 * read region so that generic_file_read() will do its
183 * job, but that doesn't mean the kms size is
184 * _correct_, it is only the _minimum_ size. If
185 * someone does a stat they will get the correct size
186 * which will always be >= the kms value here.
189 if (i_size_read(inode) < kms) {
190 i_size_write(inode, kms);
192 DFID" updating i_size %llu\n",
193 PFID(lu_object_fid(&obj->co_lu)),
194 (__u64)i_size_read(inode));
199 vvp_object_size_unlock(obj);
204 /*****************************************************************************
210 static int vvp_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
211 __u32 enqflags, enum cl_lock_mode mode,
212 pgoff_t start, pgoff_t end)
214 struct vvp_io *vio = vvp_env_io(env);
215 struct cl_lock_descr *descr = &vio->vui_link.cill_descr;
216 struct cl_object *obj = io->ci_obj;
218 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
221 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
223 memset(&vio->vui_link, 0, sizeof vio->vui_link);
225 if (vio->vui_fd && (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
226 descr->cld_mode = CLM_GROUP;
227 descr->cld_gid = vio->vui_fd->fd_grouplock.lg_gid;
228 enqflags |= CEF_LOCK_MATCH;
230 descr->cld_mode = mode;
233 descr->cld_obj = obj;
234 descr->cld_start = start;
235 descr->cld_end = end;
236 descr->cld_enq_flags = enqflags;
238 cl_io_lock_add(env, io, &vio->vui_link);
243 static int vvp_io_one_lock(const struct lu_env *env, struct cl_io *io,
244 __u32 enqflags, enum cl_lock_mode mode,
245 loff_t start, loff_t end)
247 struct cl_object *obj = io->ci_obj;
249 return vvp_io_one_lock_index(env, io, enqflags, mode,
250 cl_index(obj, start), cl_index(obj, end));
253 static int vvp_io_write_iter_init(const struct lu_env *env,
254 const struct cl_io_slice *ios)
256 struct vvp_io *vio = cl2vvp_io(env, ios);
258 cl_page_list_init(&vio->u.readwrite.vui_queue);
259 vio->u.readwrite.vui_written = 0;
260 vio->u.readwrite.vui_from = 0;
261 vio->u.readwrite.vui_to = PAGE_SIZE;
266 static int vvp_io_read_iter_init(const struct lu_env *env,
267 const struct cl_io_slice *ios)
269 struct vvp_io *vio = cl2vvp_io(env, ios);
271 vio->u.readwrite.vui_read = 0;
276 static void vvp_io_write_iter_fini(const struct lu_env *env,
277 const struct cl_io_slice *ios)
279 struct vvp_io *vio = cl2vvp_io(env, ios);
281 LASSERT(vio->u.readwrite.vui_queue.pl_nr == 0);
284 static int vvp_io_fault_iter_init(const struct lu_env *env,
285 const struct cl_io_slice *ios)
287 struct vvp_io *vio = cl2vvp_io(env, ios);
288 struct inode *inode = vvp_object_inode(ios->cis_obj);
290 LASSERT(inode == file_inode(vio->vui_fd->fd_file));
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);
305 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
307 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
308 "need write layout %d, restore needed %d\n",
309 PFID(lu_object_fid(&obj->co_lu)),
310 io->ci_ignore_layout, io->ci_verify_layout,
311 vio->vui_layout_gen, io->ci_need_write_intent,
312 io->ci_restore_needed);
314 if (io->ci_restore_needed) {
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 held 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().
326 * Even if ll_layout_restore() returns zero, it doesn't mean
327 * that restore has been successful. Therefore it sets
328 * ci_verify_layout so that it will check layout at the end
332 io->ci_restore_needed = 1;
333 io->ci_need_restart = 0;
334 io->ci_verify_layout = 0;
339 io->ci_restore_needed = 0;
341 /* Even if ll_layout_restore() returns zero, it doesn't mean
342 * that restore has been successful. Therefore it should verify
343 * if there was layout change and restart I/O correspondingly.
345 ll_layout_refresh(inode, &gen);
346 io->ci_need_restart = vio->vui_layout_gen != gen;
347 if (io->ci_need_restart) {
349 DFID" layout changed from %d to %d.\n",
350 PFID(lu_object_fid(&obj->co_lu)),
351 vio->vui_layout_gen, gen);
352 /* today successful restore is the only possible
354 /* restore was done, clear restoring state */
355 clear_bit(LLIF_FILE_RESTORING,
356 &ll_i2info(vvp_object_inode(obj))->lli_flags);
362 * dynamic layout change needed, send layout intent
365 if (io->ci_need_write_intent) {
366 enum layout_intent_opc opc = LAYOUT_INTENT_WRITE;
368 io->ci_need_write_intent = 0;
370 LASSERT(io->ci_type == CIT_WRITE || cl_io_is_fallocate(io) ||
371 cl_io_is_trunc(io) || cl_io_is_mkwrite(io));
373 CDEBUG(D_VFSTRACE, DFID" write layout, type %u "DEXT"\n",
374 PFID(lu_object_fid(&obj->co_lu)), io->ci_type,
375 PEXT(&io->ci_write_intent));
377 if (cl_io_is_trunc(io))
378 opc = LAYOUT_INTENT_TRUNC;
380 rc = ll_layout_write_intent(inode, opc, &io->ci_write_intent);
383 io->ci_need_restart = 1;
387 if (!io->ci_need_restart &&
388 !io->ci_ignore_layout && io->ci_verify_layout) {
389 /* check layout version */
390 ll_layout_refresh(inode, &gen);
391 io->ci_need_restart = vio->vui_layout_gen != gen;
392 if (io->ci_need_restart) {
394 DFID" layout changed from %d to %d.\n",
395 PFID(lu_object_fid(&obj->co_lu)),
396 vio->vui_layout_gen, gen);
404 static void vvp_io_fault_fini(const struct lu_env *env,
405 const struct cl_io_slice *ios)
407 struct cl_io *io = ios->cis_io;
408 struct cl_page *page = io->u.ci_fault.ft_page;
410 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
413 lu_ref_del(&page->cp_reference, "fault", io);
414 cl_page_put(env, page);
415 io->u.ci_fault.ft_page = NULL;
417 vvp_io_fini(env, ios);
420 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
423 * we only want to hold PW locks if the mmap() can generate
424 * writes back to the file and that only happens in shared
427 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
432 static int vvp_mmap_locks(const struct lu_env *env,
433 struct vvp_io *vio, struct cl_io *io)
435 struct vvp_thread_info *vti = vvp_env_info(env);
436 struct mm_struct *mm = current->mm;
437 struct vm_area_struct *vma;
438 struct cl_lock_descr *descr = &vti->vti_descr;
439 union ldlm_policy_data policy;
447 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
449 /* nfs or loop back device write */
450 if (vio->vui_iter == NULL)
453 /* No MM (e.g. NFS)? No vmas too. */
457 if (!iter_is_iovec(vio->vui_iter) && !iov_iter_is_kvec(vio->vui_iter))
460 for (i = *vio->vui_iter;
462 iov_iter_advance(&i, iov.iov_len)) {
463 iov = iov_iter_iovec(&i);
464 addr = (unsigned long)iov.iov_base;
470 count += addr & ~PAGE_MASK;
474 while ((vma = our_vma(mm, addr, count)) != NULL) {
475 struct dentry *de = file_dentry(vma->vm_file);
476 struct inode *inode = de->d_inode;
477 int flags = CEF_MUST;
479 if (ll_file_nolock(vma->vm_file)) {
481 * For no lock case is not allowed for mmap
488 * XXX: Required lock mode can be weakened: CIT_WRITE
489 * io only ever reads user level buffer, and CIT_READ
492 policy_from_vma(&policy, vma, addr, count);
493 descr->cld_mode = vvp_mode_from_vma(vma);
494 descr->cld_obj = ll_i2info(inode)->lli_clob;
495 descr->cld_start = cl_index(descr->cld_obj,
496 policy.l_extent.start);
497 descr->cld_end = cl_index(descr->cld_obj,
498 policy.l_extent.end);
499 descr->cld_enq_flags = flags;
500 result = cl_io_lock_alloc_add(env, io, descr);
502 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
503 descr->cld_mode, descr->cld_start,
509 if (vma->vm_end - addr >= count)
512 count -= vma->vm_end - addr;
515 mmap_read_unlock(mm);
522 static void vvp_io_advance(const struct lu_env *env,
523 const struct cl_io_slice *ios,
526 struct cl_object *obj = ios->cis_io->ci_obj;
527 struct vvp_io *vio = cl2vvp_io(env, ios);
529 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
532 * Since 3.16(26978b8b4) vfs revert iov iter to
533 * original position even io succeed, so instead
534 * of relying on VFS, we move iov iter by ourselves.
536 iov_iter_advance(vio->vui_iter, nob);
537 vio->vui_tot_count -= nob;
538 iov_iter_reexpand(vio->vui_iter, vio->vui_tot_count);
541 static void vvp_io_update_iov(const struct lu_env *env,
542 struct vvp_io *vio, struct cl_io *io)
544 size_t size = io->u.ci_rw.crw_count;
549 iov_iter_truncate(vio->vui_iter, size);
552 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
553 enum cl_lock_mode mode, loff_t start, loff_t end)
555 struct vvp_io *vio = vvp_env_io(env);
559 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
562 vvp_io_update_iov(env, vio, io);
564 if (io->u.ci_rw.crw_nonblock)
565 ast_flags |= CEF_NONBLOCK;
566 if (io->ci_lock_no_expand)
567 ast_flags |= CEF_LOCK_NO_EXPAND;
569 /* Group lock held means no lockless any more */
570 if (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)
573 if (ll_file_nolock(vio->vui_fd->fd_file) ||
574 (vio->vui_fd->fd_file->f_flags & O_DIRECT &&
576 ast_flags |= CEF_NEVER;
579 result = vvp_mmap_locks(env, vio, io);
581 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
586 static int vvp_io_read_lock(const struct lu_env *env,
587 const struct cl_io_slice *ios)
589 struct cl_io *io = ios->cis_io;
590 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
594 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
595 rd->crw_pos + rd->crw_count - 1);
599 static int vvp_io_fault_lock(const struct lu_env *env,
600 const struct cl_io_slice *ios)
602 struct cl_io *io = ios->cis_io;
603 struct vvp_io *vio = cl2vvp_io(env, ios);
605 * XXX LDLM_FL_CBPENDING
607 return vvp_io_one_lock_index(env,
609 vvp_mode_from_vma(vio->u.fault.ft_vma),
610 io->u.ci_fault.ft_index,
611 io->u.ci_fault.ft_index);
614 static int vvp_io_write_lock(const struct lu_env *env,
615 const struct cl_io_slice *ios)
617 struct cl_io *io = ios->cis_io;
621 if (io->u.ci_wr.wr_append) {
623 end = OBD_OBJECT_EOF;
625 start = io->u.ci_wr.wr.crw_pos;
626 end = start + io->u.ci_wr.wr.crw_count - 1;
629 RETURN(vvp_io_rw_lock(env, io, CLM_WRITE, start, end));
632 static int vvp_io_setattr_iter_init(const struct lu_env *env,
633 const struct cl_io_slice *ios)
640 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
642 * Handles "lockless io" mode when extent locking is done by server.
644 static int vvp_io_setattr_lock(const struct lu_env *env,
645 const struct cl_io_slice *ios)
647 struct cl_io *io = ios->cis_io;
648 __u64 lock_start = 0;
649 __u64 lock_end = OBD_OBJECT_EOF;
652 if (cl_io_is_trunc(io)) {
653 struct inode *inode = vvp_object_inode(io->ci_obj);
655 /* set enqueue flags to CEF_MUST in case of encrypted file,
656 * to prevent lockless truncate
658 if (S_ISREG(inode->i_mode) && IS_ENCRYPTED(inode))
660 else if (io->u.ci_setattr.sa_attr.lvb_size == 0)
661 enqflags = CEF_DISCARD_DATA;
662 } else if (cl_io_is_fallocate(io)) {
663 lock_start = io->u.ci_setattr.sa_falloc_offset;
664 lock_end = io->u.ci_setattr.sa_falloc_end - 1;
666 unsigned int valid = io->u.ci_setattr.sa_avalid;
668 if (!(valid & TIMES_SET_FLAGS))
671 if ((!(valid & ATTR_MTIME) ||
672 io->u.ci_setattr.sa_attr.lvb_mtime >=
673 io->u.ci_setattr.sa_attr.lvb_ctime) &&
674 (!(valid & ATTR_ATIME) ||
675 io->u.ci_setattr.sa_attr.lvb_atime >=
676 io->u.ci_setattr.sa_attr.lvb_ctime))
680 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
681 lock_start, lock_end);
684 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
689 * Only ll_inode_size_lock is taken at this level.
691 ll_inode_size_lock(inode);
692 result = inode_newsize_ok(inode, size);
694 ll_inode_size_unlock(inode);
697 i_size_write(inode, size);
699 ll_truncate_pagecache(inode, size);
700 ll_inode_size_unlock(inode);
704 static int vvp_io_setattr_time(const struct lu_env *env,
705 const struct cl_io_slice *ios)
707 struct cl_io *io = ios->cis_io;
708 struct cl_object *obj = io->ci_obj;
709 struct cl_attr *attr = vvp_env_thread_attr(env);
711 unsigned valid = CAT_CTIME;
713 cl_object_attr_lock(obj);
714 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
715 if (io->u.ci_setattr.sa_avalid & ATTR_ATIME_SET) {
716 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
719 if (io->u.ci_setattr.sa_avalid & ATTR_MTIME_SET) {
720 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
723 result = cl_object_attr_update(env, obj, attr, valid);
724 cl_object_attr_unlock(obj);
729 static int vvp_io_setattr_start(const struct lu_env *env,
730 const struct cl_io_slice *ios)
732 struct cl_io *io = ios->cis_io;
733 struct inode *inode = vvp_object_inode(io->ci_obj);
734 struct ll_inode_info *lli = ll_i2info(inode);
735 int mode = io->u.ci_setattr.sa_falloc_mode;
737 if (cl_io_is_trunc(io)) {
738 trunc_sem_down_write(&lli->lli_trunc_sem);
739 mutex_lock(&lli->lli_setattr_mutex);
740 inode_dio_wait(inode);
741 } else if (cl_io_is_fallocate(io)) {
744 trunc_sem_down_write(&lli->lli_trunc_sem);
745 mutex_lock(&lli->lli_setattr_mutex);
746 inode_dio_wait(inode);
748 ll_merge_attr(env, inode);
749 size = i_size_read(inode);
750 if (io->u.ci_setattr.sa_falloc_end > size &&
751 !(mode & FALLOC_FL_KEEP_SIZE)) {
752 size = io->u.ci_setattr.sa_falloc_end;
753 io->u.ci_setattr.sa_avalid |= ATTR_SIZE;
755 io->u.ci_setattr.sa_attr.lvb_size = size;
757 mutex_lock(&lli->lli_setattr_mutex);
760 if (io->u.ci_setattr.sa_avalid & TIMES_SET_FLAGS)
761 return vvp_io_setattr_time(env, ios);
766 static void vvp_io_setattr_end(const struct lu_env *env,
767 const struct cl_io_slice *ios)
769 struct cl_io *io = ios->cis_io;
770 struct inode *inode = vvp_object_inode(io->ci_obj);
771 struct ll_inode_info *lli = ll_i2info(inode);
773 if (cl_io_is_trunc(io)) {
774 /* Truncate in memory pages - they must be clean pages
775 * because osc has already notified to destroy osc_extents. */
776 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
777 mutex_unlock(&lli->lli_setattr_mutex);
778 trunc_sem_up_write(&lli->lli_trunc_sem);
779 } else if (cl_io_is_fallocate(io)) {
780 mutex_unlock(&lli->lli_setattr_mutex);
781 trunc_sem_up_write(&lli->lli_trunc_sem);
783 mutex_unlock(&lli->lli_setattr_mutex);
787 static void vvp_io_setattr_fini(const struct lu_env *env,
788 const struct cl_io_slice *ios)
790 bool restore_needed = ios->cis_io->ci_restore_needed;
791 struct inode *inode = vvp_object_inode(ios->cis_obj);
793 vvp_io_fini(env, ios);
795 if (restore_needed && !ios->cis_io->ci_restore_needed) {
796 /* restore finished, set data modified flag for HSM */
797 set_bit(LLIF_DATA_MODIFIED, &ll_i2info(inode)->lli_flags);
801 static int vvp_io_read_start(const struct lu_env *env,
802 const struct cl_io_slice *ios)
804 struct vvp_io *vio = cl2vvp_io(env, ios);
805 struct cl_io *io = ios->cis_io;
806 struct cl_object *obj = io->ci_obj;
807 struct inode *inode = vvp_object_inode(obj);
808 struct ll_inode_info *lli = ll_i2info(inode);
809 struct file *file = vio->vui_fd->fd_file;
810 loff_t pos = io->u.ci_rd.rd.crw_pos;
811 size_t cnt = io->u.ci_rd.rd.crw_count;
812 size_t tot = vio->vui_tot_count;
815 struct iov_iter iter;
820 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
822 CDEBUG(D_VFSTRACE, "%s: read [%llu, %llu)\n",
823 file_dentry(file)->d_name.name,
826 trunc_sem_down_read(&lli->lli_trunc_sem);
828 if (io->ci_async_readahead) {
833 if (!can_populate_pages(env, io, inode))
836 if (!(file->f_flags & O_DIRECT)) {
837 result = cl_io_lru_reserve(env, io, pos, cnt);
842 /* Unless this is reading a sparse file, otherwise the lock has already
843 * been acquired so vvp_prep_size() is an empty op. */
844 result = vvp_prep_size(env, obj, io, pos, cnt, &exceed);
847 else if (exceed != 0)
850 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
851 "Read ino %lu, %zu bytes, offset %lld, size %llu\n",
852 inode->i_ino, cnt, pos, i_size_read(inode));
854 /* turn off the kernel's read-ahead */
855 vio->vui_fd->fd_file->f_ra.ra_pages = 0;
857 /* initialize read-ahead window once per syscall */
858 if (!vio->vui_ra_valid) {
859 vio->vui_ra_valid = true;
860 vio->vui_ra_start_idx = cl_index(obj, pos);
861 vio->vui_ra_pages = 0;
862 page_offset = pos & ~PAGE_MASK;
865 if (tot > PAGE_SIZE - page_offset)
866 tot -= (PAGE_SIZE - page_offset);
870 vio->vui_ra_pages += (tot + PAGE_SIZE - 1) >> PAGE_SHIFT;
872 CDEBUG(D_READA, "tot %zu, ra_start %lu, ra_count %lu\n",
873 vio->vui_tot_count, vio->vui_ra_start_idx,
879 LASSERT(vio->vui_iocb->ki_pos == pos);
880 iter = *vio->vui_iter;
881 result = generic_file_read_iter(vio->vui_iocb, &iter);
886 io->ci_nob += result;
888 } else if (result == -EIOCBQUEUED) {
889 io->ci_nob += vio->u.readwrite.vui_read;
890 vio->vui_iocb->ki_pos = pos + vio->u.readwrite.vui_read;
896 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
897 struct cl_page_list *plist, int from, int to)
899 struct cl_2queue *queue = &io->ci_queue;
900 struct cl_page *page;
901 unsigned int bytes = 0;
905 if (plist->pl_nr == 0)
908 if (from > 0 || to != PAGE_SIZE) {
909 page = cl_page_list_first(plist);
910 if (plist->pl_nr == 1) {
911 cl_page_clip(env, page, from, to);
914 cl_page_clip(env, page, from, PAGE_SIZE);
915 if (to != PAGE_SIZE) {
916 page = cl_page_list_last(plist);
917 cl_page_clip(env, page, 0, to);
922 cl_2queue_init(queue);
923 cl_page_list_splice(plist, &queue->c2_qin);
924 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
926 /* plist is not sorted any more */
927 cl_page_list_splice(&queue->c2_qin, plist);
928 cl_page_list_splice(&queue->c2_qout, plist);
929 cl_2queue_fini(env, queue);
932 /* calculate bytes */
933 bytes = plist->pl_nr << PAGE_SHIFT;
934 bytes -= from + PAGE_SIZE - to;
936 while (plist->pl_nr > 0) {
937 page = cl_page_list_first(plist);
938 cl_page_list_del(env, plist, page);
940 cl_page_clip(env, page, 0, PAGE_SIZE);
942 SetPageUptodate(cl_page_vmpage(page));
943 cl_page_disown(env, io, page);
945 /* held in ll_cl_init() */
946 lu_ref_del(&page->cp_reference, "cl_io", io);
947 cl_page_put(env, page);
951 RETURN(bytes > 0 ? bytes : rc);
955 * From kernel v4.19-rc5-248-g9b89a0355144 use XArrary
956 * Prior kernels use radix_tree for tags
958 static inline void ll_page_tag_dirty(struct page *page,
959 struct address_space *mapping)
961 #ifndef HAVE_RADIX_TREE_TAG_SET
962 __xa_set_mark(&mapping->i_pages, page_index(page), PAGECACHE_TAG_DIRTY);
964 radix_tree_tag_set(&mapping->page_tree, page_index(page),
965 PAGECACHE_TAG_DIRTY);
970 * Kernels 4.2 - 4.5 pass memcg argument to account_page_dirtied()
971 * Kernel v5.2-5678-gac1c3e4 no longer exports account_page_dirtied
973 static inline void ll_account_page_dirtied(struct page *page,
974 struct address_space *mapping)
976 #ifdef HAVE_ACCOUNT_PAGE_DIRTIED_3ARGS
977 struct mem_cgroup *memcg = mem_cgroup_begin_page_stat(page);
979 account_page_dirtied(page, mapping, memcg);
980 mem_cgroup_end_page_stat(memcg);
981 #elif defined(HAVE_ACCOUNT_PAGE_DIRTIED_EXPORT)
982 account_page_dirtied(page, mapping);
984 vvp_account_page_dirtied(page, mapping);
986 ll_page_tag_dirty(page, mapping);
989 /* Taken from kernel set_page_dirty, __set_page_dirty_nobuffers
990 * Last change to this area: b93b016313b3ba8003c3b8bb71f569af91f19fc7
992 * Current with Linus tip of tree (7/13/2019):
993 * v5.2-rc4-224-ge01e060fe0
995 * Backwards compat for 3.x, 5.x kernels relating to memcg handling
996 * & rename of radix tree to xarray.
998 void vvp_set_pagevec_dirty(struct pagevec *pvec)
1000 struct page *page = pvec->pages[0];
1001 struct address_space *mapping = page->mapping;
1002 unsigned long flags;
1003 unsigned long skip_pages = 0;
1004 int count = pagevec_count(pvec);
1010 BUILD_BUG_ON(PAGEVEC_SIZE > BITS_PER_LONG);
1011 LASSERTF(page->mapping,
1012 "mapping must be set. page %p, page->private (cl_page) %p\n",
1013 page, (void *) page->private);
1015 for (i = 0; i < count; i++) {
1016 page = pvec->pages[i];
1018 ClearPageReclaim(page);
1020 lock_page_memcg(page);
1021 if (TestSetPageDirty(page)) {
1022 /* page is already dirty .. no extra work needed
1023 * set a flag for the i'th page to be skipped
1025 unlock_page_memcg(page);
1026 skip_pages |= (1 << i);
1030 ll_xa_lock_irqsave(&mapping->i_pages, flags);
1032 /* Notes on differences with __set_page_dirty_nobuffers:
1033 * 1. We don't need to call page_mapping because we know this is a page
1035 * 2. We have the pages locked, so there is no need for the careful
1036 * mapping/mapping2 dance.
1037 * 3. No mapping is impossible. (Race w/truncate mentioned in
1038 * dirty_nobuffers should be impossible because we hold the page lock.)
1039 * 4. All mappings are the same because i/o is only to one file.
1041 for (i = 0; i < count; i++) {
1042 page = pvec->pages[i];
1043 /* if the i'th page was unlocked above, skip it here */
1044 if ((skip_pages >> i) & 1)
1047 LASSERTF(page->mapping == mapping,
1048 "all pages must have the same mapping. page %p, mapping %p, first mapping %p\n",
1049 page, page->mapping, mapping);
1050 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1051 ll_account_page_dirtied(page, mapping);
1053 unlock_page_memcg(page);
1055 ll_xa_unlock_irqrestore(&mapping->i_pages, flags);
1057 CDEBUG(D_VFSTRACE, "mapping %p, count %d, dirtied %d\n", mapping,
1060 if (mapping->host && dirtied) {
1061 /* !PageAnon && !swapper_space */
1062 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1068 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
1069 struct pagevec *pvec)
1076 count = pagevec_count(pvec);
1079 for (i = 0; i < count; i++) {
1080 struct page *vmpage = pvec->pages[i];
1081 SetPageUptodate(vmpage);
1084 vvp_set_pagevec_dirty(pvec);
1086 for (i = 0; i < count; i++) {
1087 struct page *vmpage = pvec->pages[i];
1088 struct cl_page *page = (struct cl_page *) vmpage->private;
1089 cl_page_disown(env, io, page);
1090 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
1091 cl_page_put(env, page);
1097 /* make sure the page list is contiguous */
1098 static bool page_list_sanity_check(struct cl_object *obj,
1099 struct cl_page_list *plist)
1101 struct cl_page *page;
1102 pgoff_t index = CL_PAGE_EOF;
1104 cl_page_list_for_each(page, plist) {
1105 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1107 if (index == CL_PAGE_EOF) {
1108 index = vvp_index(vpg);
1113 if (index == vvp_index(vpg))
1121 /* Return how many bytes have queued or written */
1122 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
1124 struct cl_object *obj = io->ci_obj;
1125 struct inode *inode = vvp_object_inode(obj);
1126 struct vvp_io *vio = vvp_env_io(env);
1127 struct cl_page_list *queue = &vio->u.readwrite.vui_queue;
1128 struct cl_page *page;
1131 unsigned int npages = vio->u.readwrite.vui_queue.pl_nr;
1137 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
1138 npages, vio->u.readwrite.vui_from, vio->u.readwrite.vui_to);
1140 LASSERT(page_list_sanity_check(obj, queue));
1142 /* submit IO with async write */
1143 rc = cl_io_commit_async(env, io, queue,
1144 vio->u.readwrite.vui_from,
1145 vio->u.readwrite.vui_to,
1146 write_commit_callback);
1147 npages -= queue->pl_nr; /* already committed pages */
1149 /* calculate how many bytes were written */
1150 bytes = npages << PAGE_SHIFT;
1153 bytes -= vio->u.readwrite.vui_from;
1154 if (queue->pl_nr == 0) /* last page */
1155 bytes -= PAGE_SIZE - vio->u.readwrite.vui_to;
1156 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
1158 vio->u.readwrite.vui_written += bytes;
1160 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
1161 npages, bytes, vio->u.readwrite.vui_written);
1163 /* the first page must have been written. */
1164 vio->u.readwrite.vui_from = 0;
1166 LASSERT(page_list_sanity_check(obj, queue));
1167 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
1169 /* out of quota, try sync write */
1170 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
1171 rc = vvp_io_commit_sync(env, io, queue,
1172 vio->u.readwrite.vui_from,
1173 vio->u.readwrite.vui_to);
1175 vio->u.readwrite.vui_written += rc;
1180 /* update inode size */
1181 ll_merge_attr(env, inode);
1183 /* Now the pages in queue were failed to commit, discard them
1184 * unless they were dirtied before. */
1185 while (queue->pl_nr > 0) {
1186 page = cl_page_list_first(queue);
1187 cl_page_list_del(env, queue, page);
1189 if (!PageDirty(cl_page_vmpage(page)))
1190 cl_page_discard(env, io, page);
1192 cl_page_disown(env, io, page);
1194 /* held in ll_cl_init() */
1195 lu_ref_del(&page->cp_reference, "cl_io", io);
1196 cl_page_put(env, page);
1198 cl_page_list_fini(env, queue);
1203 static int vvp_io_write_start(const struct lu_env *env,
1204 const struct cl_io_slice *ios)
1206 struct vvp_io *vio = cl2vvp_io(env, ios);
1207 struct cl_io *io = ios->cis_io;
1208 struct cl_object *obj = io->ci_obj;
1209 struct inode *inode = vvp_object_inode(obj);
1210 struct ll_inode_info *lli = ll_i2info(inode);
1211 struct file *file = vio->vui_fd->fd_file;
1213 loff_t pos = io->u.ci_wr.wr.crw_pos;
1214 size_t cnt = io->u.ci_wr.wr.crw_count;
1215 bool lock_inode = !IS_NOSEC(inode);
1216 size_t nob = io->ci_nob;
1217 struct iov_iter iter;
1222 trunc_sem_down_read(&lli->lli_trunc_sem);
1224 if (!can_populate_pages(env, io, inode))
1227 if (cl_io_is_append(io)) {
1229 * PARALLEL IO This has to be changed for parallel IO doing
1230 * out-of-order writes.
1232 ll_merge_attr(env, inode);
1233 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
1234 vio->vui_iocb->ki_pos = pos;
1236 LASSERTF(vio->vui_iocb->ki_pos == pos,
1237 "ki_pos %lld [%lld, %lld)\n",
1238 vio->vui_iocb->ki_pos,
1242 CDEBUG(D_VFSTRACE, "%s: write [%llu, %llu)\n",
1243 file_dentry(file)->d_name.name,
1246 /* The maximum Lustre file size is variable, based on the OST maximum
1247 * object size and number of stripes. This needs another check in
1248 * addition to the VFS checks earlier. */
1249 if (pos + cnt > ll_file_maxbytes(inode)) {
1251 "%s: file %s ("DFID") offset %llu > maxbytes %llu\n",
1252 ll_i2sbi(inode)->ll_fsname,
1253 file_dentry(file)->d_name.name,
1254 PFID(ll_inode2fid(inode)), pos + cnt,
1255 ll_file_maxbytes(inode));
1259 /* Tests to verify we take the i_mutex correctly */
1260 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_SEC) && !lock_inode)
1263 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_NOSEC) && lock_inode)
1266 if (!(file->f_flags & O_DIRECT)) {
1267 result = cl_io_lru_reserve(env, io, pos, cnt);
1272 if (vio->vui_iter == NULL) {
1273 /* from a temp io in ll_cl_init(). */
1277 * When using the locked AIO function (generic_file_aio_write())
1278 * testing has shown the inode mutex to be a limiting factor
1279 * with multi-threaded single shared file performance. To get
1280 * around this, we now use the lockless version. To maintain
1281 * consistency, proper locking to protect against writes,
1282 * trucates, etc. is handled in the higher layers of lustre.
1284 lock_inode = !IS_NOSEC(inode);
1285 iter = *vio->vui_iter;
1287 if (unlikely(lock_inode))
1289 result = __generic_file_write_iter(vio->vui_iocb, &iter);
1290 if (unlikely(lock_inode))
1291 inode_unlock(inode);
1295 #ifdef HAVE_GENERIC_WRITE_SYNC_2ARGS
1296 result = generic_write_sync(vio->vui_iocb, result);
1301 err = generic_write_sync(vio->vui_iocb->ki_filp, pos,
1303 if (err < 0 && result > 0)
1310 result = vvp_io_write_commit(env, io);
1311 /* Simulate short commit */
1312 if (CFS_FAULT_CHECK(OBD_FAIL_LLITE_SHORT_COMMIT)) {
1313 vio->u.readwrite.vui_written >>= 1;
1314 if (vio->u.readwrite.vui_written > 0)
1315 io->ci_need_restart = 1;
1317 if (vio->u.readwrite.vui_written > 0) {
1318 result = vio->u.readwrite.vui_written;
1319 CDEBUG(D_VFSTRACE, "%s: write nob %zd, result: %zd\n",
1320 file_dentry(file)->d_name.name,
1321 io->ci_nob, result);
1322 io->ci_nob += result;
1324 io->ci_continue = 0;
1327 if (vio->vui_iocb->ki_pos != (pos + io->ci_nob - nob)) {
1329 "%s: write position mismatch: ki_pos %lld vs. pos %lld, written %zd, commit %zd: rc = %zd\n",
1330 file_dentry(file)->d_name.name,
1331 vio->vui_iocb->ki_pos, pos + io->ci_nob - nob,
1332 written, io->ci_nob - nob, result);
1334 * Rewind ki_pos and vui_iter to where it has
1335 * successfully committed.
1337 vio->vui_iocb->ki_pos = pos + io->ci_nob - nob;
1339 if (result > 0 || result == -EIOCBQUEUED) {
1340 set_bit(LLIF_DATA_MODIFIED, &ll_i2info(inode)->lli_flags);
1342 if (result != -EIOCBQUEUED && result < cnt)
1343 io->ci_continue = 0;
1347 if (result == -EIOCBQUEUED) {
1348 io->ci_nob += vio->u.readwrite.vui_written;
1349 vio->vui_iocb->ki_pos = pos +
1350 vio->u.readwrite.vui_written;
1357 static void vvp_io_rw_end(const struct lu_env *env,
1358 const struct cl_io_slice *ios)
1360 struct inode *inode = vvp_object_inode(ios->cis_obj);
1361 struct ll_inode_info *lli = ll_i2info(inode);
1363 trunc_sem_up_read(&lli->lli_trunc_sem);
1366 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
1368 struct vm_fault *vmf = cfio->ft_vmf;
1370 cfio->ft_flags = ll_filemap_fault(cfio->ft_vma, vmf);
1371 cfio->ft_flags_valid = 1;
1374 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
1375 get_vmf_address(vmf));
1376 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
1377 lock_page(vmf->page);
1378 cfio->ft_flags |= VM_FAULT_LOCKED;
1381 cfio->ft_vmpage = vmf->page;
1386 if (cfio->ft_flags & VM_FAULT_SIGBUS) {
1387 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", get_vmf_address(vmf));
1391 if (cfio->ft_flags & VM_FAULT_OOM) {
1392 CDEBUG(D_PAGE, "got addr %p - OOM\n", get_vmf_address(vmf));
1396 if (cfio->ft_flags & VM_FAULT_RETRY)
1399 CERROR("unknown error in page fault %d\n", cfio->ft_flags);
1404 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1405 struct pagevec *pvec)
1407 vvp_set_pagevec_dirty(pvec);
1410 static int vvp_io_fault_start(const struct lu_env *env,
1411 const struct cl_io_slice *ios)
1413 struct vvp_io *vio = cl2vvp_io(env, ios);
1414 struct cl_io *io = ios->cis_io;
1415 struct cl_object *obj = io->ci_obj;
1416 struct inode *inode = vvp_object_inode(obj);
1417 struct ll_inode_info *lli = ll_i2info(inode);
1418 struct cl_fault_io *fio = &io->u.ci_fault;
1419 struct vvp_fault_io *cfio = &vio->u.fault;
1422 struct page *vmpage = NULL;
1423 struct cl_page *page;
1428 trunc_sem_down_read_nowait(&lli->lli_trunc_sem);
1430 /* offset of the last byte on the page */
1431 offset = cl_offset(obj, fio->ft_index + 1) - 1;
1432 LASSERT(cl_index(obj, offset) == fio->ft_index);
1433 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1437 /* must return locked page */
1438 if (fio->ft_mkwrite) {
1439 LASSERT(cfio->ft_vmpage != NULL);
1440 lock_page(cfio->ft_vmpage);
1442 result = vvp_io_kernel_fault(cfio);
1447 vmpage = cfio->ft_vmpage;
1448 LASSERT(PageLocked(vmpage));
1450 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1451 generic_error_remove_page(vmpage->mapping, vmpage);
1453 size = i_size_read(inode);
1454 /* Though we have already held a cl_lock upon this page, but
1455 * it still can be truncated locally. */
1456 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1457 (page_offset(vmpage) > size))) {
1458 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1460 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1462 GOTO(out, result = +1);
1465 last_index = cl_index(obj, size - 1);
1467 if (fio->ft_mkwrite ) {
1469 * Capture the size while holding the lli_trunc_sem from above
1470 * we want to make sure that we complete the mkwrite action
1471 * while holding this lock. We need to make sure that we are
1472 * not past the end of the file.
1474 if (last_index < fio->ft_index) {
1476 "llite: mkwrite and truncate race happened: "
1477 "%p: 0x%lx 0x%lx\n",
1478 vmpage->mapping,fio->ft_index,last_index);
1480 * We need to return if we are
1481 * passed the end of the file. This will propagate
1482 * up the call stack to ll_page_mkwrite where
1483 * we will return VM_FAULT_NOPAGE. Any non-negative
1484 * value returned here will be silently
1485 * converted to 0. If the vmpage->mapping is null
1486 * the error code would be converted back to ENODATA
1487 * in ll_page_mkwrite0. Thus we return -ENODATA
1488 * to handle both cases
1490 GOTO(out, result = -ENODATA);
1494 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1496 GOTO(out, result = PTR_ERR(page));
1498 /* if page is going to be written, we should add this page into cache
1500 if (fio->ft_mkwrite) {
1501 wait_on_page_writeback(vmpage);
1502 if (!PageDirty(vmpage)) {
1503 struct cl_page_list *plist = &vio->u.fault.ft_queue;
1504 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1507 /* vvp_page_assume() calls wait_on_page_writeback(). */
1508 cl_page_assume(env, io, page);
1510 cl_page_list_init(plist);
1511 cl_page_list_add(plist, page);
1514 if (last_index == vvp_index(vpg))
1515 to = ((size - 1) & ~PAGE_MASK) + 1;
1517 /* Do not set Dirty bit here so that in case IO is
1518 * started before the page is really made dirty, we
1519 * still have chance to detect it. */
1520 result = cl_io_commit_async(env, io, plist, 0, to,
1521 mkwrite_commit_callback);
1522 /* Have overquota flag, trying sync write to check
1523 * whether indeed out of quota */
1524 if (result == -EDQUOT) {
1526 result = vvp_io_commit_sync(env, io,
1530 cl_page_own(env, io, page);
1531 cl_page_list_add(plist, page);
1532 lu_ref_add(&page->cp_reference,
1534 result = cl_io_commit_async(env, io,
1536 mkwrite_commit_callback);
1539 cl_page_put(env, page);
1543 LASSERT(cl_page_is_owned(page, io));
1544 cl_page_list_fini(env, plist);
1548 cl_page_discard(env, io, page);
1549 cl_page_disown(env, io, page);
1551 cl_page_put(env, page);
1553 /* we're in big trouble, what can we do now? */
1554 if (result == -EDQUOT)
1558 cl_page_disown(env, io, page);
1564 * The ft_index is only used in the case of
1565 * a mkwrite action. We need to check
1566 * our assertions are correct, since
1567 * we should have caught this above
1569 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1570 if (fio->ft_index == last_index)
1572 * Last page is mapped partially.
1574 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1576 fio->ft_nob = cl_page_size(obj);
1578 lu_ref_add(&page->cp_reference, "fault", io);
1579 fio->ft_page = page;
1583 /* return unlocked vmpage to avoid deadlocking */
1585 unlock_page(vmpage);
1587 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1592 static void vvp_io_fault_end(const struct lu_env *env,
1593 const struct cl_io_slice *ios)
1595 struct inode *inode = vvp_object_inode(ios->cis_obj);
1596 struct ll_inode_info *lli = ll_i2info(inode);
1598 CLOBINVRNT(env, ios->cis_io->ci_obj,
1599 vvp_object_invariant(ios->cis_io->ci_obj));
1600 trunc_sem_up_read(&lli->lli_trunc_sem);
1603 static int vvp_io_fsync_start(const struct lu_env *env,
1604 const struct cl_io_slice *ios)
1606 /* we should mark TOWRITE bit to each dirty page in radix tree to
1607 * verify pages have been written, but this is difficult because of
1612 static int vvp_io_read_ahead(const struct lu_env *env,
1613 const struct cl_io_slice *ios,
1614 pgoff_t start, struct cl_read_ahead *ra)
1619 if (ios->cis_io->ci_type == CIT_READ ||
1620 ios->cis_io->ci_type == CIT_FAULT) {
1621 struct vvp_io *vio = cl2vvp_io(env, ios);
1623 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1624 ra->cra_end_idx = CL_PAGE_EOF;
1625 result = +1; /* no need to call down */
1632 static int vvp_io_lseek_lock(const struct lu_env *env,
1633 const struct cl_io_slice *ios)
1635 struct cl_io *io = ios->cis_io;
1636 __u64 lock_start = io->u.ci_lseek.ls_start;
1637 __u64 lock_end = OBD_OBJECT_EOF;
1638 __u32 enqflags = CEF_MUST; /* always take client lock */
1640 return vvp_io_one_lock(env, io, enqflags, CLM_READ,
1641 lock_start, lock_end);
1644 static int vvp_io_lseek_start(const struct lu_env *env,
1645 const struct cl_io_slice *ios)
1647 struct cl_io *io = ios->cis_io;
1648 struct inode *inode = vvp_object_inode(io->ci_obj);
1649 __u64 start = io->u.ci_lseek.ls_start;
1652 inode_dio_wait(inode);
1654 /* At the moment we have DLM lock so just update inode
1655 * to know the file size.
1657 ll_merge_attr(env, inode);
1658 if (start >= i_size_read(inode)) {
1659 io->u.ci_lseek.ls_result = -ENXIO;
1665 static void vvp_io_lseek_end(const struct lu_env *env,
1666 const struct cl_io_slice *ios)
1668 struct cl_io *io = ios->cis_io;
1669 struct inode *inode = vvp_object_inode(io->ci_obj);
1671 if (io->u.ci_lseek.ls_result > i_size_read(inode))
1672 io->u.ci_lseek.ls_result = -ENXIO;
1674 inode_unlock(inode);
1677 static const struct cl_io_operations vvp_io_ops = {
1680 .cio_fini = vvp_io_fini,
1681 .cio_iter_init = vvp_io_read_iter_init,
1682 .cio_lock = vvp_io_read_lock,
1683 .cio_start = vvp_io_read_start,
1684 .cio_end = vvp_io_rw_end,
1685 .cio_advance = vvp_io_advance,
1688 .cio_fini = vvp_io_fini,
1689 .cio_iter_init = vvp_io_write_iter_init,
1690 .cio_iter_fini = vvp_io_write_iter_fini,
1691 .cio_lock = vvp_io_write_lock,
1692 .cio_start = vvp_io_write_start,
1693 .cio_end = vvp_io_rw_end,
1694 .cio_advance = vvp_io_advance,
1697 .cio_fini = vvp_io_setattr_fini,
1698 .cio_iter_init = vvp_io_setattr_iter_init,
1699 .cio_lock = vvp_io_setattr_lock,
1700 .cio_start = vvp_io_setattr_start,
1701 .cio_end = vvp_io_setattr_end
1704 .cio_fini = vvp_io_fault_fini,
1705 .cio_iter_init = vvp_io_fault_iter_init,
1706 .cio_lock = vvp_io_fault_lock,
1707 .cio_start = vvp_io_fault_start,
1708 .cio_end = vvp_io_fault_end,
1711 .cio_start = vvp_io_fsync_start,
1712 .cio_fini = vvp_io_fini
1715 .cio_fini = vvp_io_fini
1718 .cio_fini = vvp_io_fini
1721 .cio_fini = vvp_io_fini
1724 .cio_fini = vvp_io_fini,
1725 .cio_lock = vvp_io_lseek_lock,
1726 .cio_start = vvp_io_lseek_start,
1727 .cio_end = vvp_io_lseek_end,
1730 .cio_read_ahead = vvp_io_read_ahead
1733 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1736 struct vvp_io *vio = vvp_env_io(env);
1737 struct inode *inode = vvp_object_inode(obj);
1740 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1743 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1744 "restore needed %d\n",
1745 PFID(lu_object_fid(&obj->co_lu)),
1746 io->ci_ignore_layout, io->ci_verify_layout,
1747 vio->vui_layout_gen, io->ci_restore_needed);
1749 CL_IO_SLICE_CLEAN(vio, vui_cl);
1750 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1751 vio->vui_ra_valid = false;
1753 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1755 struct ll_inode_info *lli = ll_i2info(inode);
1757 count = io->u.ci_rw.crw_count;
1758 /* "If nbyte is 0, read() will return 0 and have no other
1759 * results." -- Single Unix Spec */
1763 vio->vui_tot_count = count;
1765 /* for read/write, we store the jobid in the inode, and
1766 * it'll be fetched by osc when building RPC.
1768 * it's not accurate if the file is shared by different
1771 lustre_get_jobid(lli->lli_jobid, sizeof(lli->lli_jobid));
1772 } else if (io->ci_type == CIT_SETATTR) {
1773 if (!cl_io_is_trunc(io))
1774 io->ci_lockreq = CILR_MANDATORY;
1777 /* Enqueue layout lock and get layout version. We need to do this
1778 * even for operations requiring to open file, such as read and write,
1779 * because it might not grant layout lock in IT_OPEN. */
1780 if (result == 0 && !io->ci_ignore_layout) {
1781 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1782 if (result == -ENOENT)
1783 /* If the inode on MDS has been removed, but the objects
1784 * on OSTs haven't been destroyed (async unlink), layout
1785 * fetch will return -ENOENT, we'd ingore this error
1786 * and continue with dirty flush. LU-3230. */
1789 CERROR("%s: refresh file layout " DFID " error %d.\n",
1790 ll_i2sbi(inode)->ll_fsname,
1791 PFID(lu_object_fid(&obj->co_lu)), result);
1794 io->ci_result = result < 0 ? result : 0;