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 CFS_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);
153 if (pos > kms || !attr->cat_kms_valid) {
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 return vvp_io_one_lock_index(env, io, enqflags, mode,
248 start >> PAGE_SHIFT, end >> PAGE_SHIFT);
251 static int vvp_io_write_iter_init(const struct lu_env *env,
252 const struct cl_io_slice *ios)
254 struct vvp_io *vio = cl2vvp_io(env, ios);
256 cl_page_list_init(&vio->u.readwrite.vui_queue);
257 vio->u.readwrite.vui_written = 0;
258 vio->u.readwrite.vui_from = 0;
259 vio->u.readwrite.vui_to = PAGE_SIZE;
264 static int vvp_io_read_iter_init(const struct lu_env *env,
265 const struct cl_io_slice *ios)
267 struct vvp_io *vio = cl2vvp_io(env, ios);
269 vio->u.readwrite.vui_read = 0;
274 static void vvp_io_write_iter_fini(const struct lu_env *env,
275 const struct cl_io_slice *ios)
277 struct vvp_io *vio = cl2vvp_io(env, ios);
279 LASSERT(vio->u.readwrite.vui_queue.pl_nr == 0);
282 static int vvp_io_fault_iter_init(const struct lu_env *env,
283 const struct cl_io_slice *ios)
285 struct vvp_io *vio = cl2vvp_io(env, ios);
286 struct inode *inode = vvp_object_inode(ios->cis_obj);
288 LASSERT(inode == file_inode(vio->vui_fd->fd_file));
293 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
295 struct cl_io *io = ios->cis_io;
296 struct cl_object *obj = io->ci_obj;
297 struct vvp_io *vio = cl2vvp_io(env, ios);
298 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, invalidate_lock %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, io->ci_invalidate_page_cache);
312 #ifdef HAVE_INVALIDATE_LOCK
313 if (io->ci_invalidate_page_cache) {
314 filemap_invalidate_unlock(inode->i_mapping);
315 io->ci_invalidate_page_cache = 0;
317 #endif /* HAVE_INVALIDATE_LOCK */
319 if (io->ci_restore_needed) {
320 /* file was detected release, we need to restore it
321 * before finishing the io
323 rc = ll_layout_restore(inode, 0, OBD_OBJECT_EOF);
324 /* if restore registration failed, no restart,
325 * we will return -ENODATA */
326 /* The layout will change after restore, so we need to
327 * block on layout lock held by the MDT
328 * as MDT will not send new layout in lvb (see LU-3124)
329 * we have to explicitly fetch it, all this will be done
330 * by ll_layout_refresh().
331 * Even if ll_layout_restore() returns zero, it doesn't mean
332 * that restore has been successful. Therefore it sets
333 * ci_verify_layout so that it will check layout at the end
337 io->ci_restore_needed = 1;
338 io->ci_need_restart = 0;
339 io->ci_verify_layout = 0;
344 io->ci_restore_needed = 0;
346 /* Even if ll_layout_restore() returns zero, it doesn't mean
347 * that restore has been successful. Therefore it should verify
348 * if there was layout change and restart I/O correspondingly.
350 ll_layout_refresh(inode, &gen);
351 io->ci_need_restart = vio->vui_layout_gen != gen;
352 if (io->ci_need_restart) {
354 DFID" layout changed from %d to %d.\n",
355 PFID(lu_object_fid(&obj->co_lu)),
356 vio->vui_layout_gen, gen);
357 /* today successful restore is the only possible
359 /* restore was done, clear restoring state */
360 clear_bit(LLIF_FILE_RESTORING,
361 &ll_i2info(vvp_object_inode(obj))->lli_flags);
367 * dynamic layout change needed, send layout intent
370 if (io->ci_need_write_intent) {
371 enum layout_intent_opc opc = LAYOUT_INTENT_WRITE;
373 io->ci_need_write_intent = 0;
375 LASSERT(io->ci_type == CIT_WRITE || cl_io_is_fallocate(io) ||
376 cl_io_is_trunc(io) || cl_io_is_mkwrite(io));
378 CDEBUG(D_VFSTRACE, DFID" write layout, type %u "DEXT"\n",
379 PFID(lu_object_fid(&obj->co_lu)), io->ci_type,
380 PEXT(&io->ci_write_intent));
382 if (cl_io_is_trunc(io))
383 opc = LAYOUT_INTENT_TRUNC;
385 rc = ll_layout_write_intent(inode, opc, &io->ci_write_intent);
388 io->ci_need_restart = 1;
392 if (!io->ci_need_restart &&
393 !io->ci_ignore_layout && io->ci_verify_layout) {
394 /* check layout version */
395 ll_layout_refresh(inode, &gen);
396 io->ci_need_restart = vio->vui_layout_gen != gen;
397 if (io->ci_need_restart) {
399 DFID" layout changed from %d to %d.\n",
400 PFID(lu_object_fid(&obj->co_lu)),
401 vio->vui_layout_gen, gen);
409 static void vvp_io_fault_fini(const struct lu_env *env,
410 const struct cl_io_slice *ios)
412 struct cl_io *io = ios->cis_io;
413 struct cl_page *page = io->u.ci_fault.ft_page;
415 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
418 lu_ref_del(&page->cp_reference, "fault", io);
419 cl_page_put(env, page);
420 io->u.ci_fault.ft_page = NULL;
422 vvp_io_fini(env, ios);
425 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
428 * we only want to hold PW locks if the mmap() can generate
429 * writes back to the file and that only happens in shared
432 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
437 static int vvp_mmap_locks(const struct lu_env *env,
438 struct vvp_io *vio, struct cl_io *io)
440 struct vvp_thread_info *vti = vvp_env_info(env);
441 struct mm_struct *mm = current->mm;
442 struct vm_area_struct *vma;
443 struct cl_lock_descr *descr = &vti->vti_descr;
444 union ldlm_policy_data policy;
452 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
454 /* nfs or loop back device write */
455 if (vio->vui_iter == NULL)
458 /* No MM (e.g. NFS)? No vmas too. */
462 if (!iter_is_iovec(vio->vui_iter) && !iov_iter_is_kvec(vio->vui_iter))
465 for (i = *vio->vui_iter;
467 iov_iter_advance(&i, iov.iov_len)) {
468 iov = iov_iter_iovec(&i);
469 addr = (unsigned long)iov.iov_base;
475 count += addr & ~PAGE_MASK;
479 while ((vma = our_vma(mm, addr, count)) != NULL) {
480 struct dentry *de = file_dentry(vma->vm_file);
481 struct inode *inode = de->d_inode;
482 int flags = CEF_MUST;
484 if (ll_file_nolock(vma->vm_file)) {
486 * For no lock case is not allowed for mmap
493 * XXX: Required lock mode can be weakened: CIT_WRITE
494 * io only ever reads user level buffer, and CIT_READ
497 policy_from_vma(&policy, vma, addr, count);
498 descr->cld_mode = vvp_mode_from_vma(vma);
499 descr->cld_obj = ll_i2info(inode)->lli_clob;
500 descr->cld_start = policy.l_extent.start >> PAGE_SHIFT;
501 descr->cld_end = policy.l_extent.end >> PAGE_SHIFT;
502 descr->cld_enq_flags = flags;
503 result = cl_io_lock_alloc_add(env, io, descr);
505 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
506 descr->cld_mode, descr->cld_start,
512 if (vma->vm_end - addr >= count)
515 count -= vma->vm_end - addr;
518 mmap_read_unlock(mm);
525 static void vvp_io_advance(const struct lu_env *env,
526 const struct cl_io_slice *ios,
529 struct cl_object *obj = ios->cis_io->ci_obj;
530 struct vvp_io *vio = cl2vvp_io(env, ios);
532 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
535 * Since 3.16(26978b8b4) vfs revert iov iter to
536 * original position even io succeed, so instead
537 * of relying on VFS, we move iov iter by ourselves.
539 iov_iter_advance(vio->vui_iter, nob);
540 CDEBUG(D_VFSTRACE, "advancing %ld bytes\n", nob);
541 vio->vui_tot_count -= nob;
542 iov_iter_reexpand(vio->vui_iter, vio->vui_tot_count);
545 static void vvp_io_update_iov(const struct lu_env *env,
546 struct vvp_io *vio, struct cl_io *io)
548 size_t size = io->u.ci_rw.crw_count;
553 iov_iter_truncate(vio->vui_iter, size);
556 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
557 enum cl_lock_mode mode, loff_t start, loff_t end)
559 struct vvp_io *vio = vvp_env_io(env);
563 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
566 vvp_io_update_iov(env, vio, io);
568 if (io->u.ci_rw.crw_nonblock)
569 ast_flags |= CEF_NONBLOCK;
570 if (io->ci_lock_no_expand)
571 ast_flags |= CEF_LOCK_NO_EXPAND;
573 /* Group lock held means no lockless any more */
574 if (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)
577 if (ll_file_nolock(vio->vui_fd->fd_file) ||
578 (vio->vui_fd->fd_file->f_flags & O_DIRECT &&
580 ast_flags |= CEF_NEVER;
583 result = vvp_mmap_locks(env, vio, io);
585 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
590 static int vvp_io_read_lock(const struct lu_env *env,
591 const struct cl_io_slice *ios)
593 struct cl_io *io = ios->cis_io;
594 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
598 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
599 rd->crw_pos + rd->crw_count - 1);
603 static int vvp_io_fault_lock(const struct lu_env *env,
604 const struct cl_io_slice *ios)
606 struct cl_io *io = ios->cis_io;
607 struct vvp_io *vio = cl2vvp_io(env, ios);
609 * XXX LDLM_FL_CBPENDING
611 return vvp_io_one_lock_index(env,
613 vvp_mode_from_vma(vio->u.fault.ft_vma),
614 io->u.ci_fault.ft_index,
615 io->u.ci_fault.ft_index);
618 static int vvp_io_write_lock(const struct lu_env *env,
619 const struct cl_io_slice *ios)
621 struct cl_io *io = ios->cis_io;
625 if (io->u.ci_wr.wr_append) {
627 end = OBD_OBJECT_EOF;
629 start = io->u.ci_wr.wr.crw_pos;
630 end = start + io->u.ci_wr.wr.crw_count - 1;
633 RETURN(vvp_io_rw_lock(env, io, CLM_WRITE, start, end));
636 static int vvp_io_setattr_iter_init(const struct lu_env *env,
637 const struct cl_io_slice *ios)
644 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
646 * Handles "lockless io" mode when extent locking is done by server.
648 static int vvp_io_setattr_lock(const struct lu_env *env,
649 const struct cl_io_slice *ios)
651 struct cl_io *io = ios->cis_io;
652 __u64 lock_start = 0;
653 __u64 lock_end = OBD_OBJECT_EOF;
656 if (cl_io_is_trunc(io)) {
657 struct inode *inode = vvp_object_inode(io->ci_obj);
659 /* set enqueue flags to CEF_MUST in case of encrypted file,
660 * to prevent lockless truncate
662 if (S_ISREG(inode->i_mode) && IS_ENCRYPTED(inode))
664 else if (io->u.ci_setattr.sa_attr.lvb_size == 0)
665 enqflags = CEF_DISCARD_DATA;
666 } else if (cl_io_is_fallocate(io)) {
667 lock_start = io->u.ci_setattr.sa_falloc_offset;
668 lock_end = io->u.ci_setattr.sa_falloc_end - 1;
670 unsigned int valid = io->u.ci_setattr.sa_avalid;
672 if (!(valid & TIMES_SET_FLAGS))
675 if ((!(valid & ATTR_MTIME) ||
676 io->u.ci_setattr.sa_attr.lvb_mtime >=
677 io->u.ci_setattr.sa_attr.lvb_ctime) &&
678 (!(valid & ATTR_ATIME) ||
679 io->u.ci_setattr.sa_attr.lvb_atime >=
680 io->u.ci_setattr.sa_attr.lvb_ctime))
684 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
685 lock_start, lock_end);
688 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
693 * Only ll_inode_size_lock is taken at this level.
695 ll_inode_size_lock(inode);
696 result = inode_newsize_ok(inode, size);
698 ll_inode_size_unlock(inode);
701 i_size_write(inode, size);
703 ll_truncate_pagecache(inode, size);
704 ll_inode_size_unlock(inode);
708 static int vvp_io_setattr_time(const struct lu_env *env,
709 const struct cl_io_slice *ios)
711 struct cl_io *io = ios->cis_io;
712 struct cl_object *obj = io->ci_obj;
713 struct cl_attr *attr = vvp_env_thread_attr(env);
715 unsigned valid = CAT_CTIME;
717 cl_object_attr_lock(obj);
718 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
719 if (io->u.ci_setattr.sa_avalid & ATTR_ATIME_SET) {
720 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
723 if (io->u.ci_setattr.sa_avalid & ATTR_MTIME_SET) {
724 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
727 result = cl_object_attr_update(env, obj, attr, valid);
728 cl_object_attr_unlock(obj);
733 static int vvp_io_setattr_start(const struct lu_env *env,
734 const struct cl_io_slice *ios)
736 struct cl_io *io = ios->cis_io;
737 struct inode *inode = vvp_object_inode(io->ci_obj);
738 struct ll_inode_info *lli = ll_i2info(inode);
739 int mode = io->u.ci_setattr.sa_falloc_mode;
741 if (cl_io_is_trunc(io)) {
742 trunc_sem_down_write(&lli->lli_trunc_sem);
743 mutex_lock(&lli->lli_setattr_mutex);
744 inode_dio_wait(inode);
745 } else if (cl_io_is_fallocate(io)) {
748 trunc_sem_down_write(&lli->lli_trunc_sem);
749 mutex_lock(&lli->lli_setattr_mutex);
750 inode_dio_wait(inode);
752 ll_merge_attr(env, inode);
753 size = i_size_read(inode);
754 if (io->u.ci_setattr.sa_falloc_end > size &&
755 !(mode & FALLOC_FL_KEEP_SIZE)) {
756 size = io->u.ci_setattr.sa_falloc_end;
757 io->u.ci_setattr.sa_avalid |= ATTR_SIZE;
759 io->u.ci_setattr.sa_attr.lvb_size = size;
761 mutex_lock(&lli->lli_setattr_mutex);
764 if (io->u.ci_setattr.sa_avalid & TIMES_SET_FLAGS)
765 return vvp_io_setattr_time(env, ios);
770 static void vvp_io_setattr_end(const struct lu_env *env,
771 const struct cl_io_slice *ios)
773 struct cl_io *io = ios->cis_io;
774 struct inode *inode = vvp_object_inode(io->ci_obj);
775 struct ll_inode_info *lli = ll_i2info(inode);
776 loff_t size = io->u.ci_setattr.sa_attr.lvb_size;
778 if (cl_io_is_trunc(io)) {
779 /* Truncate in memory pages - they must be clean pages
780 * because osc has already notified to destroy osc_extents. */
781 vvp_do_vmtruncate(inode, size);
782 mutex_unlock(&lli->lli_setattr_mutex);
783 trunc_sem_up_write(&lli->lli_trunc_sem);
785 /* Update size and blocks for LSOM */
786 if (!io->ci_ignore_layout)
787 ll_merge_attr(env, inode);
788 } else if (cl_io_is_fallocate(io)) {
789 int mode = io->u.ci_setattr.sa_falloc_mode;
791 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
792 size > i_size_read(inode)) {
793 ll_inode_size_lock(inode);
794 i_size_write(inode, size);
795 ll_inode_size_unlock(inode);
797 inode->i_ctime = current_time(inode);
798 mutex_unlock(&lli->lli_setattr_mutex);
799 trunc_sem_up_write(&lli->lli_trunc_sem);
801 mutex_unlock(&lli->lli_setattr_mutex);
805 static void vvp_io_setattr_fini(const struct lu_env *env,
806 const struct cl_io_slice *ios)
808 bool restore_needed = ios->cis_io->ci_restore_needed;
809 struct inode *inode = vvp_object_inode(ios->cis_obj);
811 vvp_io_fini(env, ios);
813 if (restore_needed && !ios->cis_io->ci_restore_needed) {
814 /* restore finished, set data modified flag for HSM */
815 set_bit(LLIF_DATA_MODIFIED, &ll_i2info(inode)->lli_flags);
819 static int vvp_io_read_start(const struct lu_env *env,
820 const struct cl_io_slice *ios)
822 struct vvp_io *vio = cl2vvp_io(env, ios);
823 struct cl_io *io = ios->cis_io;
824 struct cl_object *obj = io->ci_obj;
825 struct inode *inode = vvp_object_inode(obj);
826 struct ll_inode_info *lli = ll_i2info(inode);
827 struct file *file = vio->vui_fd->fd_file;
828 loff_t pos = io->u.ci_rd.rd.crw_pos;
829 size_t cnt = io->u.ci_rd.rd.crw_count;
830 size_t tot = vio->vui_tot_count;
831 struct ll_cl_context *lcc;
835 int total_bytes_read = 0;
836 struct iov_iter iter;
841 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
843 CDEBUG(D_VFSTRACE, "%s: read [%llu, %llu)\n",
844 file_dentry(file)->d_name.name,
847 trunc_sem_down_read(&lli->lli_trunc_sem);
849 if (io->ci_async_readahead) {
854 if (!can_populate_pages(env, io, inode))
857 if (!(file->f_flags & O_DIRECT)) {
858 result = cl_io_lru_reserve(env, io, pos, cnt);
863 /* Unless this is reading a sparse file, otherwise the lock has already
864 * been acquired so vvp_prep_size() is an empty op. */
865 result = vvp_prep_size(env, obj, io, pos, cnt, &exceed);
868 else if (exceed != 0)
871 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
872 "Read ino %lu, %zu bytes, offset %lld, size %llu\n",
873 inode->i_ino, cnt, pos, i_size_read(inode));
875 /* initialize read-ahead window once per syscall */
876 if (!vio->vui_ra_valid) {
877 vio->vui_ra_valid = true;
878 vio->vui_ra_start_idx = pos >> PAGE_SHIFT;
879 vio->vui_ra_pages = 0;
880 page_offset = pos & ~PAGE_MASK;
883 if (tot > PAGE_SIZE - page_offset)
884 tot -= (PAGE_SIZE - page_offset);
888 vio->vui_ra_pages += (tot + PAGE_SIZE - 1) >> PAGE_SHIFT;
890 CDEBUG(D_READA, "tot %zu, ra_start %lu, ra_count %lu\n",
891 vio->vui_tot_count, vio->vui_ra_start_idx,
897 LASSERT(vio->vui_iocb->ki_pos == pos);
898 iter = *vio->vui_iter;
900 lcc = ll_cl_find(inode);
901 lcc->lcc_end_index = DIV_ROUND_UP(pos + iter.count, PAGE_SIZE);
902 CDEBUG(D_VFSTRACE, "count:%ld iocb pos:%lld\n", iter.count, pos);
904 /* this seqlock lets us notice if a page has been deleted on this inode
905 * during the fault process, allowing us to catch an erroneous short
910 seq = read_seqbegin(&ll_i2info(inode)->lli_page_inv_lock);
911 result = generic_file_read_iter(vio->vui_iocb, &iter);
913 io->ci_nob += result;
914 total_bytes_read += result;
916 /* if we got a short read or -EIO and we raced with page invalidation,
919 } while (read_seqretry(&ll_i2info(inode)->lli_page_inv_lock, seq) &&
920 ((result >= 0 && iov_iter_count(&iter) > 0)
925 if (total_bytes_read < cnt)
928 } else if (result == -EIOCBQUEUED) {
929 io->ci_nob += vio->u.readwrite.vui_read;
930 vio->vui_iocb->ki_pos = pos + vio->u.readwrite.vui_read;
936 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
937 struct cl_page_list *plist, int from, int to)
939 struct cl_2queue *queue = &io->ci_queue;
940 struct cl_page *page;
941 unsigned int bytes = 0;
945 if (plist->pl_nr == 0)
948 if (from > 0 || to != PAGE_SIZE) {
949 page = cl_page_list_first(plist);
950 if (plist->pl_nr == 1) {
951 cl_page_clip(env, page, from, to);
954 cl_page_clip(env, page, from, PAGE_SIZE);
955 if (to != PAGE_SIZE) {
956 page = cl_page_list_last(plist);
957 cl_page_clip(env, page, 0, to);
962 cl_2queue_init(queue);
963 cl_page_list_splice(plist, &queue->c2_qin);
964 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
966 /* plist is not sorted any more */
967 cl_page_list_splice(&queue->c2_qin, plist);
968 cl_page_list_splice(&queue->c2_qout, plist);
969 cl_2queue_fini(env, queue);
972 /* calculate bytes */
973 bytes = plist->pl_nr << PAGE_SHIFT;
974 bytes -= from + PAGE_SIZE - to;
976 while (plist->pl_nr > 0) {
977 page = cl_page_list_first(plist);
978 cl_page_list_del(env, plist, page);
980 cl_page_clip(env, page, 0, PAGE_SIZE);
982 SetPageUptodate(cl_page_vmpage(page));
983 cl_page_disown(env, io, page);
985 /* held in ll_cl_init() */
986 lu_ref_del(&page->cp_reference, "cl_io", io);
987 cl_page_put(env, page);
991 RETURN(bytes > 0 ? bytes : rc);
995 * From kernel v4.19-rc5-248-g9b89a0355144 use XArrary
996 * Prior kernels use radix_tree for tags
998 static inline void ll_page_tag_dirty(struct page *page,
999 struct address_space *mapping)
1001 #ifndef HAVE_RADIX_TREE_TAG_SET
1002 __xa_set_mark(&mapping->i_pages, page_index(page), PAGECACHE_TAG_DIRTY);
1004 radix_tree_tag_set(&mapping->page_tree, page_index(page),
1005 PAGECACHE_TAG_DIRTY);
1010 * Kernels 4.2 - 4.5 pass memcg argument to account_page_dirtied()
1011 * Kernel v5.2-5678-gac1c3e4 no longer exports account_page_dirtied
1013 static inline void ll_account_page_dirtied(struct page *page,
1014 struct address_space *mapping)
1016 #ifdef HAVE_ACCOUNT_PAGE_DIRTIED_3ARGS
1017 struct mem_cgroup *memcg = mem_cgroup_begin_page_stat(page);
1019 account_page_dirtied(page, mapping, memcg);
1020 mem_cgroup_end_page_stat(memcg);
1021 #elif defined(HAVE_ACCOUNT_PAGE_DIRTIED_EXPORT)
1022 account_page_dirtied(page, mapping);
1024 vvp_account_page_dirtied(page, mapping);
1026 ll_page_tag_dirty(page, mapping);
1029 /* Taken from kernel set_page_dirty, __set_page_dirty_nobuffers
1030 * Last change to this area: b93b016313b3ba8003c3b8bb71f569af91f19fc7
1032 * Current with Linus tip of tree (7/13/2019):
1033 * v5.2-rc4-224-ge01e060fe0
1035 * Backwards compat for 3.x, 5.x kernels relating to memcg handling
1036 * & rename of radix tree to xarray.
1038 void vvp_set_pagevec_dirty(struct pagevec *pvec)
1040 struct page *page = pvec->pages[0];
1041 int count = pagevec_count(pvec);
1043 #ifdef HAVE_KALLSYMS_LOOKUP_NAME
1044 struct address_space *mapping = page->mapping;
1045 unsigned long flags;
1046 unsigned long skip_pages = 0;
1052 BUILD_BUG_ON(PAGEVEC_SIZE > BITS_PER_LONG);
1053 LASSERTF(page->mapping,
1054 "mapping must be set. page %p, page->private (cl_page) %p\n",
1055 page, (void *) page->private);
1058 * kernels without HAVE_KALLSYMS_LOOKUP_NAME also don't have
1059 * account_dirty_page exported, and if we can't access that symbol,
1060 * we can't do page dirtying in batch (taking the xarray lock only once)
1061 * so we just fall back to a looped call to __set_page_dirty_nobuffers
1063 #ifndef HAVE_ACCOUNT_PAGE_DIRTIED_EXPORT
1064 if (!vvp_account_page_dirtied) {
1065 for (i = 0; i < count; i++)
1066 __set_page_dirty_nobuffers(pvec->pages[i]);
1071 #ifdef HAVE_KALLSYMS_LOOKUP_NAME
1072 for (i = 0; i < count; i++) {
1073 page = pvec->pages[i];
1075 ClearPageReclaim(page);
1077 vvp_lock_page_memcg(page);
1078 if (TestSetPageDirty(page)) {
1079 /* page is already dirty .. no extra work needed
1080 * set a flag for the i'th page to be skipped
1082 vvp_unlock_page_memcg(page);
1083 skip_pages |= (1 << i);
1087 ll_xa_lock_irqsave(&mapping->i_pages, flags);
1089 /* Notes on differences with __set_page_dirty_nobuffers:
1090 * 1. We don't need to call page_mapping because we know this is a page
1092 * 2. We have the pages locked, so there is no need for the careful
1093 * mapping/mapping2 dance.
1094 * 3. No mapping is impossible. (Race w/truncate mentioned in
1095 * dirty_nobuffers should be impossible because we hold the page lock.)
1096 * 4. All mappings are the same because i/o is only to one file.
1098 for (i = 0; i < count; i++) {
1099 page = pvec->pages[i];
1100 /* if the i'th page was unlocked above, skip it here */
1101 if ((skip_pages >> i) & 1)
1104 LASSERTF(page->mapping == mapping,
1105 "all pages must have the same mapping. page %p, mapping %p, first mapping %p\n",
1106 page, page->mapping, mapping);
1107 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1108 ll_account_page_dirtied(page, mapping);
1110 vvp_unlock_page_memcg(page);
1112 ll_xa_unlock_irqrestore(&mapping->i_pages, flags);
1114 CDEBUG(D_VFSTRACE, "mapping %p, count %d, dirtied %d\n", mapping,
1117 if (mapping->host && dirtied) {
1118 /* !PageAnon && !swapper_space */
1119 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1125 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
1126 struct pagevec *pvec)
1133 count = pagevec_count(pvec);
1136 for (i = 0; i < count; i++) {
1137 struct page *vmpage = pvec->pages[i];
1138 SetPageUptodate(vmpage);
1141 vvp_set_pagevec_dirty(pvec);
1143 for (i = 0; i < count; i++) {
1144 struct page *vmpage = pvec->pages[i];
1145 struct cl_page *page = (struct cl_page *) vmpage->private;
1146 cl_page_disown(env, io, page);
1147 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
1148 cl_page_put(env, page);
1154 /* make sure the page list is contiguous */
1155 static bool page_list_sanity_check(struct cl_object *obj,
1156 struct cl_page_list *plist)
1158 struct cl_page *page;
1159 pgoff_t index = CL_PAGE_EOF;
1161 cl_page_list_for_each(page, plist) {
1162 if (index == CL_PAGE_EOF) {
1163 index = cl_page_index(page);
1168 if (index == cl_page_index(page))
1176 /* Return how many bytes have queued or written */
1177 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
1179 struct cl_object *obj = io->ci_obj;
1180 struct inode *inode = vvp_object_inode(obj);
1181 struct vvp_io *vio = vvp_env_io(env);
1182 struct cl_page_list *queue = &vio->u.readwrite.vui_queue;
1183 struct cl_page *page;
1186 unsigned int npages = vio->u.readwrite.vui_queue.pl_nr;
1192 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
1193 npages, vio->u.readwrite.vui_from, vio->u.readwrite.vui_to);
1195 LASSERT(page_list_sanity_check(obj, queue));
1197 /* submit IO with async write */
1198 rc = cl_io_commit_async(env, io, queue,
1199 vio->u.readwrite.vui_from,
1200 vio->u.readwrite.vui_to,
1201 write_commit_callback);
1202 npages -= queue->pl_nr; /* already committed pages */
1204 /* calculate how many bytes were written */
1205 bytes = npages << PAGE_SHIFT;
1208 bytes -= vio->u.readwrite.vui_from;
1209 if (queue->pl_nr == 0) /* last page */
1210 bytes -= PAGE_SIZE - vio->u.readwrite.vui_to;
1211 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
1213 vio->u.readwrite.vui_written += bytes;
1215 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
1216 npages, bytes, vio->u.readwrite.vui_written);
1218 /* the first page must have been written. */
1219 vio->u.readwrite.vui_from = 0;
1221 LASSERT(page_list_sanity_check(obj, queue));
1222 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
1224 /* out of quota, try sync write */
1225 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
1226 struct ll_inode_info *lli = ll_i2info(inode);
1228 rc = vvp_io_commit_sync(env, io, queue,
1229 vio->u.readwrite.vui_from,
1230 vio->u.readwrite.vui_to);
1232 vio->u.readwrite.vui_written += rc;
1235 if (lli->lli_clob != NULL)
1236 lov_read_and_clear_async_rc(lli->lli_clob);
1237 lli->lli_async_rc = 0;
1240 /* update inode size */
1241 ll_merge_attr(env, inode);
1243 /* Now the pages in queue were failed to commit, discard them
1244 * unless they were dirtied before. */
1245 while (queue->pl_nr > 0) {
1246 page = cl_page_list_first(queue);
1247 cl_page_list_del(env, queue, page);
1249 if (!PageDirty(cl_page_vmpage(page)))
1250 cl_page_discard(env, io, page);
1252 cl_page_disown(env, io, page);
1254 /* held in ll_cl_init() */
1255 lu_ref_del(&page->cp_reference, "cl_io", io);
1256 cl_page_put(env, page);
1258 cl_page_list_fini(env, queue);
1263 static int vvp_io_write_start(const struct lu_env *env,
1264 const struct cl_io_slice *ios)
1266 struct vvp_io *vio = cl2vvp_io(env, ios);
1267 struct cl_io *io = ios->cis_io;
1268 struct cl_object *obj = io->ci_obj;
1269 struct inode *inode = vvp_object_inode(obj);
1270 struct ll_inode_info *lli = ll_i2info(inode);
1271 struct file *file = vio->vui_fd->fd_file;
1273 loff_t pos = io->u.ci_wr.wr.crw_pos;
1274 size_t cnt = io->u.ci_wr.wr.crw_count;
1275 bool lock_inode = !IS_NOSEC(inode);
1276 size_t nob = io->ci_nob;
1277 struct iov_iter iter;
1282 trunc_sem_down_read(&lli->lli_trunc_sem);
1284 if (!can_populate_pages(env, io, inode))
1287 if (cl_io_is_append(io)) {
1289 * PARALLEL IO This has to be changed for parallel IO doing
1290 * out-of-order writes.
1292 ll_merge_attr(env, inode);
1293 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
1294 vio->vui_iocb->ki_pos = pos;
1296 LASSERTF(vio->vui_iocb->ki_pos == pos,
1297 "ki_pos %lld [%lld, %lld)\n",
1298 vio->vui_iocb->ki_pos,
1302 CDEBUG(D_VFSTRACE, "%s: write [%llu, %llu)\n",
1303 file_dentry(file)->d_name.name,
1306 /* The maximum Lustre file size is variable, based on the OST maximum
1307 * object size and number of stripes. This needs another check in
1308 * addition to the VFS checks earlier. */
1309 if (pos + cnt > ll_file_maxbytes(inode)) {
1311 "%s: file %s ("DFID") offset %llu > maxbytes %llu\n",
1312 ll_i2sbi(inode)->ll_fsname,
1313 file_dentry(file)->d_name.name,
1314 PFID(ll_inode2fid(inode)), pos + cnt,
1315 ll_file_maxbytes(inode));
1319 /* Tests to verify we take the i_mutex correctly */
1320 if (CFS_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_SEC) && !lock_inode)
1323 if (CFS_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_NOSEC) && lock_inode)
1326 if (!(file->f_flags & O_DIRECT)) {
1327 result = cl_io_lru_reserve(env, io, pos, cnt);
1332 if (vio->vui_iter == NULL) {
1333 /* from a temp io in ll_cl_init(). */
1337 * When using the locked AIO function (generic_file_aio_write())
1338 * testing has shown the inode mutex to be a limiting factor
1339 * with multi-threaded single shared file performance. To get
1340 * around this, we now use the lockless version. To maintain
1341 * consistency, proper locking to protect against writes,
1342 * trucates, etc. is handled in the higher layers of lustre.
1344 lock_inode = !IS_NOSEC(inode);
1345 iter = *vio->vui_iter;
1347 if (unlikely(lock_inode))
1349 result = __generic_file_write_iter(vio->vui_iocb, &iter);
1350 if (unlikely(lock_inode))
1351 inode_unlock(inode);
1355 #ifdef HAVE_GENERIC_WRITE_SYNC_2ARGS
1356 result = generic_write_sync(vio->vui_iocb, result);
1361 err = generic_write_sync(vio->vui_iocb->ki_filp, pos,
1363 if (err < 0 && result > 0)
1370 result = vvp_io_write_commit(env, io);
1371 /* Simulate short commit */
1372 if (CFS_FAULT_CHECK(OBD_FAIL_LLITE_SHORT_COMMIT)) {
1373 vio->u.readwrite.vui_written >>= 1;
1374 if (vio->u.readwrite.vui_written > 0)
1375 io->ci_need_restart = 1;
1377 if (vio->u.readwrite.vui_written > 0) {
1378 result = vio->u.readwrite.vui_written;
1379 CDEBUG(D_VFSTRACE, "%s: write nob %zd, result: %zd\n",
1380 file_dentry(file)->d_name.name,
1381 io->ci_nob, result);
1382 io->ci_nob += result;
1384 io->ci_continue = 0;
1387 if (vio->vui_iocb->ki_pos != (pos + io->ci_nob - nob)) {
1389 "%s: write position mismatch: ki_pos %lld vs. pos %lld, written %zd, commit %zd: rc = %zd\n",
1390 file_dentry(file)->d_name.name,
1391 vio->vui_iocb->ki_pos, pos + io->ci_nob - nob,
1392 written, io->ci_nob - nob, result);
1394 * Rewind ki_pos and vui_iter to where it has
1395 * successfully committed.
1397 vio->vui_iocb->ki_pos = pos + io->ci_nob - nob;
1399 if (result > 0 || result == -EIOCBQUEUED) {
1400 set_bit(LLIF_DATA_MODIFIED, &ll_i2info(inode)->lli_flags);
1402 if (result != -EIOCBQUEUED && result < cnt)
1403 io->ci_continue = 0;
1407 if (result == -EIOCBQUEUED) {
1408 io->ci_nob += vio->u.readwrite.vui_written;
1409 vio->vui_iocb->ki_pos = pos +
1410 vio->u.readwrite.vui_written;
1417 static void vvp_io_rw_end(const struct lu_env *env,
1418 const struct cl_io_slice *ios)
1420 struct inode *inode = vvp_object_inode(ios->cis_obj);
1421 struct ll_inode_info *lli = ll_i2info(inode);
1423 trunc_sem_up_read(&lli->lli_trunc_sem);
1426 static void vvp_io_write_end(const struct lu_env *env,
1427 const struct cl_io_slice *ios)
1429 struct inode *inode = vvp_object_inode(ios->cis_obj);
1430 struct cl_io *io = ios->cis_io;
1432 vvp_io_rw_end(env, ios);
1434 /* Update size and blocks for LSOM (best effort) */
1435 if (!io->ci_ignore_layout && cl_io_is_sync_write(io))
1436 ll_merge_attr_try(env, inode);
1440 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
1442 struct vm_fault *vmf = cfio->ft_vmf;
1444 cfio->ft_flags = ll_filemap_fault(cfio->ft_vma, vmf);
1445 cfio->ft_flags_valid = 1;
1448 /* success, vmpage is locked */
1449 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
1450 get_vmf_address(vmf));
1451 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
1452 lock_page(vmf->page);
1453 cfio->ft_flags |= VM_FAULT_LOCKED;
1456 cfio->ft_vmpage = vmf->page;
1461 if (cfio->ft_flags & VM_FAULT_SIGBUS) {
1462 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", get_vmf_address(vmf));
1466 if (cfio->ft_flags & VM_FAULT_OOM) {
1467 CDEBUG(D_PAGE, "got addr %p - OOM\n", get_vmf_address(vmf));
1471 if (cfio->ft_flags & VM_FAULT_RETRY)
1474 CERROR("unknown error in page fault %d\n", cfio->ft_flags);
1479 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1480 struct pagevec *pvec)
1482 vvp_set_pagevec_dirty(pvec);
1485 static int vvp_io_fault_start(const struct lu_env *env,
1486 const struct cl_io_slice *ios)
1488 struct vvp_io *vio = cl2vvp_io(env, ios);
1489 struct cl_io *io = ios->cis_io;
1490 struct cl_object *obj = io->ci_obj;
1491 struct inode *inode = vvp_object_inode(obj);
1492 struct ll_inode_info *lli = ll_i2info(inode);
1493 struct cl_fault_io *fio = &io->u.ci_fault;
1494 struct vvp_fault_io *cfio = &vio->u.fault;
1497 struct page *vmpage = NULL;
1498 struct cl_page *page;
1503 trunc_sem_down_read_nowait(&lli->lli_trunc_sem);
1505 /* offset of the last byte on the page */
1506 offset = ((fio->ft_index + 1) << PAGE_SHIFT) - 1;
1507 LASSERT((offset >> PAGE_SHIFT) == fio->ft_index);
1508 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1512 /* must return locked page */
1513 if (fio->ft_mkwrite) {
1514 LASSERT(cfio->ft_vmpage != NULL);
1515 lock_page(cfio->ft_vmpage);
1517 result = vvp_io_kernel_fault(cfio);
1522 vmpage = cfio->ft_vmpage;
1523 LASSERT(PageLocked(vmpage));
1525 if (CFS_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1526 generic_error_remove_page(vmpage->mapping, vmpage);
1528 size = i_size_read(inode);
1529 /* Though we have already held a cl_lock upon this page, but
1530 * it still can be truncated locally. */
1531 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1532 (page_offset(vmpage) > size))) {
1533 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1535 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1537 GOTO(out, result = +1);
1540 last_index = (size - 1) >> PAGE_SHIFT;
1542 if (fio->ft_mkwrite ) {
1544 * Capture the size while holding the lli_trunc_sem from above
1545 * we want to make sure that we complete the mkwrite action
1546 * while holding this lock. We need to make sure that we are
1547 * not past the end of the file.
1549 if (last_index < fio->ft_index) {
1551 "llite: mkwrite and truncate race happened: "
1552 "%p: 0x%lx 0x%lx\n",
1553 vmpage->mapping,fio->ft_index,last_index);
1555 * We need to return if we are
1556 * passed the end of the file. This will propagate
1557 * up the call stack to ll_page_mkwrite where
1558 * we will return VM_FAULT_NOPAGE. Any non-negative
1559 * value returned here will be silently
1560 * converted to 0. If the vmpage->mapping is null
1561 * the error code would be converted back to ENODATA
1562 * in ll_page_mkwrite0. Thus we return -ENODATA
1563 * to handle both cases
1565 GOTO(out, result = -ENODATA);
1569 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1571 GOTO(out, result = PTR_ERR(page));
1573 /* if page is going to be written, we should add this page into cache
1575 if (fio->ft_mkwrite) {
1576 wait_on_page_writeback(vmpage);
1577 if (!PageDirty(vmpage)) {
1578 struct cl_page_list *plist = &vio->u.fault.ft_queue;
1581 /* vvp_page_assume() calls wait_on_page_writeback(). */
1582 cl_page_assume(env, io, page);
1584 cl_page_list_init(plist);
1585 cl_page_list_add(plist, page, true);
1588 if (last_index == cl_page_index(page))
1589 to = ((size - 1) & ~PAGE_MASK) + 1;
1591 /* Do not set Dirty bit here so that in case IO is
1592 * started before the page is really made dirty, we
1593 * still have chance to detect it. */
1594 result = cl_io_commit_async(env, io, plist, 0, to,
1595 mkwrite_commit_callback);
1596 /* Have overquota flag, trying sync write to check
1597 * whether indeed out of quota */
1598 if (result == -EDQUOT) {
1600 result = vvp_io_commit_sync(env, io,
1604 cl_page_own(env, io, page);
1605 cl_page_list_add(plist, page, true);
1606 lu_ref_add(&page->cp_reference,
1608 result = cl_io_commit_async(env, io,
1610 mkwrite_commit_callback);
1613 cl_page_put(env, page);
1617 LASSERT(cl_page_is_owned(page, io));
1618 cl_page_list_fini(env, plist);
1622 cl_page_discard(env, io, page);
1623 cl_page_disown(env, io, page);
1625 cl_page_put(env, page);
1627 /* we're in big trouble, what can we do now? */
1628 if (result == -EDQUOT)
1632 cl_page_disown(env, io, page);
1638 * The ft_index is only used in the case of
1639 * a mkwrite action. We need to check
1640 * our assertions are correct, since
1641 * we should have caught this above
1643 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1644 if (fio->ft_index == last_index)
1646 * Last page is mapped partially.
1648 fio->ft_nob = size - (fio->ft_index << PAGE_SHIFT);
1650 fio->ft_nob = PAGE_SIZE;
1652 lu_ref_add(&page->cp_reference, "fault", io);
1653 fio->ft_page = page;
1657 /* return unlocked vmpage to avoid deadlocking */
1659 unlock_page(vmpage);
1661 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1666 static void vvp_io_fault_end(const struct lu_env *env,
1667 const struct cl_io_slice *ios)
1669 struct inode *inode = vvp_object_inode(ios->cis_obj);
1670 struct ll_inode_info *lli = ll_i2info(inode);
1672 CLOBINVRNT(env, ios->cis_io->ci_obj,
1673 vvp_object_invariant(ios->cis_io->ci_obj));
1674 trunc_sem_up_read(&lli->lli_trunc_sem);
1677 static int vvp_io_fsync_start(const struct lu_env *env,
1678 const struct cl_io_slice *ios)
1680 /* we should mark TOWRITE bit to each dirty page in radix tree to
1681 * verify pages have been written, but this is difficult because of
1686 static void vvp_io_fsync_end(const struct lu_env *env,
1687 const struct cl_io_slice *ios)
1689 struct inode *inode = vvp_object_inode(ios->cis_obj);
1690 struct cl_io *io = ios->cis_io;
1692 /* Update size and blocks for LSOM (best effort) */
1693 if (!io->ci_ignore_layout)
1694 ll_merge_attr_try(env, inode);
1697 static int vvp_io_read_ahead(const struct lu_env *env,
1698 const struct cl_io_slice *ios,
1699 pgoff_t start, struct cl_read_ahead *ra)
1704 if (ios->cis_io->ci_type == CIT_READ ||
1705 ios->cis_io->ci_type == CIT_FAULT) {
1706 struct vvp_io *vio = cl2vvp_io(env, ios);
1708 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1709 ra->cra_end_idx = CL_PAGE_EOF;
1710 result = +1; /* no need to call down */
1717 static int vvp_io_lseek_lock(const struct lu_env *env,
1718 const struct cl_io_slice *ios)
1720 struct cl_io *io = ios->cis_io;
1721 __u64 lock_start = io->u.ci_lseek.ls_start;
1722 __u64 lock_end = OBD_OBJECT_EOF;
1723 __u32 enqflags = CEF_MUST; /* always take client lock */
1725 return vvp_io_one_lock(env, io, enqflags, CLM_READ,
1726 lock_start, lock_end);
1729 static int vvp_io_lseek_start(const struct lu_env *env,
1730 const struct cl_io_slice *ios)
1732 struct cl_io *io = ios->cis_io;
1733 struct inode *inode = vvp_object_inode(io->ci_obj);
1734 __u64 start = io->u.ci_lseek.ls_start;
1737 inode_dio_wait(inode);
1739 /* At the moment we have DLM lock so just update inode
1740 * to know the file size.
1742 ll_merge_attr(env, inode);
1743 if (start >= i_size_read(inode)) {
1744 io->u.ci_lseek.ls_result = -ENXIO;
1750 static void vvp_io_lseek_end(const struct lu_env *env,
1751 const struct cl_io_slice *ios)
1753 struct cl_io *io = ios->cis_io;
1754 struct inode *inode = vvp_object_inode(io->ci_obj);
1756 if (io->u.ci_lseek.ls_result > i_size_read(inode))
1757 io->u.ci_lseek.ls_result = -ENXIO;
1759 inode_unlock(inode);
1762 static const struct cl_io_operations vvp_io_ops = {
1765 .cio_fini = vvp_io_fini,
1766 .cio_iter_init = vvp_io_read_iter_init,
1767 .cio_lock = vvp_io_read_lock,
1768 .cio_start = vvp_io_read_start,
1769 .cio_end = vvp_io_rw_end,
1770 .cio_advance = vvp_io_advance,
1773 .cio_fini = vvp_io_fini,
1774 .cio_iter_init = vvp_io_write_iter_init,
1775 .cio_iter_fini = vvp_io_write_iter_fini,
1776 .cio_lock = vvp_io_write_lock,
1777 .cio_start = vvp_io_write_start,
1778 .cio_end = vvp_io_write_end,
1779 .cio_advance = vvp_io_advance,
1782 .cio_fini = vvp_io_setattr_fini,
1783 .cio_iter_init = vvp_io_setattr_iter_init,
1784 .cio_lock = vvp_io_setattr_lock,
1785 .cio_start = vvp_io_setattr_start,
1786 .cio_end = vvp_io_setattr_end
1789 .cio_fini = vvp_io_fault_fini,
1790 .cio_iter_init = vvp_io_fault_iter_init,
1791 .cio_lock = vvp_io_fault_lock,
1792 .cio_start = vvp_io_fault_start,
1793 .cio_end = vvp_io_fault_end,
1796 .cio_start = vvp_io_fsync_start,
1797 .cio_fini = vvp_io_fini,
1798 .cio_end = vvp_io_fsync_end,
1801 .cio_fini = vvp_io_fini
1804 .cio_fini = vvp_io_fini
1807 .cio_fini = vvp_io_fini
1810 .cio_fini = vvp_io_fini,
1811 .cio_lock = vvp_io_lseek_lock,
1812 .cio_start = vvp_io_lseek_start,
1813 .cio_end = vvp_io_lseek_end,
1816 .cio_read_ahead = vvp_io_read_ahead
1819 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1822 struct vvp_io *vio = vvp_env_io(env);
1823 struct inode *inode = vvp_object_inode(obj);
1826 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1829 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1830 "restore needed %d\n",
1831 PFID(lu_object_fid(&obj->co_lu)),
1832 io->ci_ignore_layout, io->ci_verify_layout,
1833 vio->vui_layout_gen, io->ci_restore_needed);
1835 CL_IO_SLICE_CLEAN(vio, vui_cl);
1836 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1837 vio->vui_ra_valid = false;
1839 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1841 struct ll_inode_info *lli = ll_i2info(inode);
1843 count = io->u.ci_rw.crw_count;
1844 /* "If nbyte is 0, read() will return 0 and have no other
1845 * results." -- Single Unix Spec */
1849 vio->vui_tot_count = count;
1851 /* for read/write, we store the process jobid/gid/uid in the
1852 * inode, and it'll be fetched by osc when building RPC.
1854 * it's not accurate if the file is shared by different
1857 lustre_get_jobid(lli->lli_jobid, sizeof(lli->lli_jobid));
1858 lli->lli_uid = from_kuid(&init_user_ns, current_uid());
1859 lli->lli_gid = from_kgid(&init_user_ns, current_gid());
1860 } else if (io->ci_type == CIT_SETATTR) {
1861 if (!cl_io_is_trunc(io))
1862 io->ci_lockreq = CILR_MANDATORY;
1865 /* Enqueue layout lock and get layout version. We need to do this
1866 * even for operations requiring to open file, such as read and write,
1867 * because it might not grant layout lock in IT_OPEN. */
1868 if (result == 0 && !io->ci_ignore_layout) {
1869 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1870 if (result == -ENOENT)
1871 /* If the inode on MDS has been removed, but the objects
1872 * on OSTs haven't been destroyed (async unlink), layout
1873 * fetch will return -ENOENT, we'd ingore this error
1874 * and continue with dirty flush. LU-3230. */
1877 CERROR("%s: refresh file layout " DFID " error %d.\n",
1878 ll_i2sbi(inode)->ll_fsname,
1879 PFID(lu_object_fid(&obj->co_lu)), result);
1882 #ifdef HAVE_INVALIDATE_LOCK
1883 if (io->ci_invalidate_page_cache)
1884 filemap_invalidate_lock(inode->i_mapping);
1885 #endif /* HAVE_INVALIDATE_LOCK */
1887 io->ci_result = result < 0 ? result : 0;