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 bytes,
124 struct cl_attr *attr = vvp_env_thread_attr(env);
125 struct inode *inode = vvp_object_inode(obj);
126 loff_t pos = start + bytes - 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 bytes += addr & ~PAGE_MASK;
479 while ((vma = our_vma(mm, addr, bytes)) != 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, bytes);
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 >= bytes)
515 bytes -= 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, size_t bytes)
528 struct cl_object *obj = ios->cis_io->ci_obj;
529 struct vvp_io *vio = cl2vvp_io(env, ios);
531 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
534 * Since 3.16(26978b8b4) vfs revert iov iter to
535 * original position even io succeed, so instead
536 * of relying on VFS, we move iov iter by ourselves.
538 iov_iter_advance(vio->vui_iter, bytes);
539 CDEBUG(D_VFSTRACE, "advancing %ld bytes\n", bytes);
540 vio->vui_tot_bytes -= bytes;
541 iov_iter_reexpand(vio->vui_iter, vio->vui_tot_bytes);
544 static void vvp_io_update_iov(const struct lu_env *env,
545 struct vvp_io *vio, struct cl_io *io)
547 size_t size = io->u.ci_rw.crw_bytes;
552 iov_iter_truncate(vio->vui_iter, size);
555 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
556 enum cl_lock_mode mode, loff_t start, loff_t end)
558 struct vvp_io *vio = vvp_env_io(env);
562 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
565 vvp_io_update_iov(env, vio, io);
567 if (io->u.ci_rw.crw_nonblock)
568 ast_flags |= CEF_NONBLOCK;
569 if (io->ci_lock_no_expand)
570 ast_flags |= CEF_LOCK_NO_EXPAND;
574 /* Group lock held means no lockless any more */
575 if (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)
578 flags = iocb_ki_flags_get(vio->vui_iocb->ki_filp,
580 if (ll_file_nolock(vio->vui_fd->fd_file) ||
581 (iocb_ki_flags_check(flags, DIRECT) &&
583 ast_flags |= CEF_NEVER;
586 result = vvp_mmap_locks(env, vio, io);
588 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
593 static int vvp_io_read_lock(const struct lu_env *env,
594 const struct cl_io_slice *ios)
596 struct cl_io *io = ios->cis_io;
597 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
601 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
602 rd->crw_pos + rd->crw_bytes - 1);
606 static int vvp_io_fault_lock(const struct lu_env *env,
607 const struct cl_io_slice *ios)
609 struct cl_io *io = ios->cis_io;
610 struct vvp_io *vio = cl2vvp_io(env, ios);
612 * XXX LDLM_FL_CBPENDING
614 return vvp_io_one_lock_index(env,
616 vvp_mode_from_vma(vio->u.fault.ft_vma),
617 io->u.ci_fault.ft_index,
618 io->u.ci_fault.ft_index);
621 static int vvp_io_write_lock(const struct lu_env *env,
622 const struct cl_io_slice *ios)
624 struct cl_io *io = ios->cis_io;
628 if (io->u.ci_wr.wr_append) {
630 end = OBD_OBJECT_EOF;
632 start = io->u.ci_wr.wr.crw_pos;
633 end = start + io->u.ci_wr.wr.crw_bytes - 1;
636 RETURN(vvp_io_rw_lock(env, io, CLM_WRITE, start, end));
639 static int vvp_io_setattr_iter_init(const struct lu_env *env,
640 const struct cl_io_slice *ios)
647 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
649 * Handles "lockless io" mode when extent locking is done by server.
651 static int vvp_io_setattr_lock(const struct lu_env *env,
652 const struct cl_io_slice *ios)
654 struct cl_io *io = ios->cis_io;
655 __u64 lock_start = 0;
656 __u64 lock_end = OBD_OBJECT_EOF;
659 if (cl_io_is_trunc(io)) {
660 struct inode *inode = vvp_object_inode(io->ci_obj);
662 /* set enqueue flags to CEF_MUST in case of encrypted file,
663 * to prevent lockless truncate
665 if (S_ISREG(inode->i_mode) && IS_ENCRYPTED(inode))
667 else if (io->u.ci_setattr.sa_attr.lvb_size == 0)
668 enqflags = CEF_DISCARD_DATA;
669 } else if (cl_io_is_fallocate(io)) {
670 lock_start = io->u.ci_setattr.sa_falloc_offset;
671 lock_end = io->u.ci_setattr.sa_falloc_end - 1;
673 unsigned int valid = io->u.ci_setattr.sa_avalid;
675 if (!(valid & TIMES_SET_FLAGS))
678 if ((!(valid & ATTR_MTIME) ||
679 io->u.ci_setattr.sa_attr.lvb_mtime >=
680 io->u.ci_setattr.sa_attr.lvb_ctime) &&
681 (!(valid & ATTR_ATIME) ||
682 io->u.ci_setattr.sa_attr.lvb_atime >=
683 io->u.ci_setattr.sa_attr.lvb_ctime))
687 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
688 lock_start, lock_end);
691 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
696 * Only ll_inode_size_lock is taken at this level.
698 ll_inode_size_lock(inode);
699 result = inode_newsize_ok(inode, size);
701 ll_inode_size_unlock(inode);
704 i_size_write(inode, size);
706 ll_truncate_pagecache(inode, size);
707 ll_inode_size_unlock(inode);
711 static int vvp_io_setattr_time(const struct lu_env *env,
712 const struct cl_io_slice *ios)
714 struct cl_io *io = ios->cis_io;
715 struct cl_object *obj = io->ci_obj;
716 struct cl_attr *attr = vvp_env_thread_attr(env);
718 unsigned valid = CAT_CTIME;
720 cl_object_attr_lock(obj);
721 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
722 if (io->u.ci_setattr.sa_avalid & ATTR_ATIME_SET) {
723 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
726 if (io->u.ci_setattr.sa_avalid & ATTR_MTIME_SET) {
727 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
730 result = cl_object_attr_update(env, obj, attr, valid);
731 cl_object_attr_unlock(obj);
736 static int vvp_io_setattr_start(const struct lu_env *env,
737 const struct cl_io_slice *ios)
739 struct cl_io *io = ios->cis_io;
740 struct inode *inode = vvp_object_inode(io->ci_obj);
741 struct ll_inode_info *lli = ll_i2info(inode);
742 int mode = io->u.ci_setattr.sa_falloc_mode;
744 if (cl_io_is_trunc(io)) {
745 trunc_sem_down_write(&lli->lli_trunc_sem);
746 mutex_lock(&lli->lli_setattr_mutex);
747 inode_dio_wait(inode);
748 } else if (cl_io_is_fallocate(io)) {
751 trunc_sem_down_write(&lli->lli_trunc_sem);
752 mutex_lock(&lli->lli_setattr_mutex);
753 inode_dio_wait(inode);
755 ll_merge_attr(env, inode);
756 size = i_size_read(inode);
757 if (io->u.ci_setattr.sa_falloc_end > size &&
758 !(mode & FALLOC_FL_KEEP_SIZE)) {
759 size = io->u.ci_setattr.sa_falloc_end;
760 io->u.ci_setattr.sa_avalid |= ATTR_SIZE;
762 io->u.ci_setattr.sa_attr.lvb_size = size;
764 mutex_lock(&lli->lli_setattr_mutex);
767 if (io->u.ci_setattr.sa_avalid & TIMES_SET_FLAGS)
768 return vvp_io_setattr_time(env, ios);
773 static void vvp_io_setattr_end(const struct lu_env *env,
774 const struct cl_io_slice *ios)
776 struct cl_io *io = ios->cis_io;
777 struct inode *inode = vvp_object_inode(io->ci_obj);
778 struct ll_inode_info *lli = ll_i2info(inode);
779 loff_t size = io->u.ci_setattr.sa_attr.lvb_size;
781 if (cl_io_is_trunc(io)) {
782 /* Truncate in memory pages - they must be clean pages
783 * because osc has already notified to destroy osc_extents. */
784 vvp_do_vmtruncate(inode, size);
785 mutex_unlock(&lli->lli_setattr_mutex);
786 trunc_sem_up_write(&lli->lli_trunc_sem);
788 /* Update size and blocks for LSOM */
789 if (!io->ci_ignore_layout)
790 ll_merge_attr(env, inode);
791 } else if (cl_io_is_fallocate(io)) {
792 int mode = io->u.ci_setattr.sa_falloc_mode;
794 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
795 size > i_size_read(inode)) {
796 ll_inode_size_lock(inode);
797 i_size_write(inode, size);
798 ll_inode_size_unlock(inode);
800 inode->i_ctime = current_time(inode);
801 mutex_unlock(&lli->lli_setattr_mutex);
802 trunc_sem_up_write(&lli->lli_trunc_sem);
804 mutex_unlock(&lli->lli_setattr_mutex);
808 static void vvp_io_setattr_fini(const struct lu_env *env,
809 const struct cl_io_slice *ios)
811 bool restore_needed = ios->cis_io->ci_restore_needed;
812 struct inode *inode = vvp_object_inode(ios->cis_obj);
814 vvp_io_fini(env, ios);
816 if (restore_needed && !ios->cis_io->ci_restore_needed) {
817 /* restore finished, set data modified flag for HSM */
818 set_bit(LLIF_DATA_MODIFIED, &ll_i2info(inode)->lli_flags);
822 static int vvp_io_read_start(const struct lu_env *env,
823 const struct cl_io_slice *ios)
825 struct vvp_io *vio = cl2vvp_io(env, ios);
826 struct cl_io *io = ios->cis_io;
827 struct cl_object *obj = io->ci_obj;
828 struct inode *inode = vvp_object_inode(obj);
829 struct ll_inode_info *lli = ll_i2info(inode);
830 struct file *file = vio->vui_fd->fd_file;
831 loff_t pos = io->u.ci_rd.rd.crw_pos;
832 size_t crw_bytes = io->u.ci_rd.rd.crw_bytes;
833 size_t tot_bytes = vio->vui_tot_bytes;
834 struct ll_cl_context *lcc;
838 int total_bytes_read = 0;
839 struct iov_iter iter;
845 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
847 CDEBUG(D_VFSTRACE, "%s: read [%llu, %llu)\n",
848 file_dentry(file)->d_name.name,
849 pos, pos + crw_bytes);
851 trunc_sem_down_read(&lli->lli_trunc_sem);
853 if (io->ci_async_readahead) {
858 if (!can_populate_pages(env, io, inode))
861 flags = iocb_ki_flags_get(file, vio->vui_iocb);
862 if (!iocb_ki_flags_check(flags, DIRECT)) {
863 result = cl_io_lru_reserve(env, io, pos, crw_bytes);
868 /* Unless this is reading a sparse file, otherwise the lock has already
869 * been acquired so vvp_prep_size() is an empty op. */
870 result = vvp_prep_size(env, obj, io, pos, crw_bytes, &exceed);
873 else if (exceed != 0)
876 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
877 "Read ino %lu, %zu bytes, offset %lld, size %llu\n",
878 inode->i_ino, crw_bytes, pos, i_size_read(inode));
880 /* initialize read-ahead window once per syscall */
881 if (!vio->vui_ra_valid) {
882 vio->vui_ra_valid = true;
883 vio->vui_ra_start_idx = pos >> PAGE_SHIFT;
884 vio->vui_ra_pages = 0;
885 page_offset = pos & ~PAGE_MASK;
888 if (tot_bytes > PAGE_SIZE - page_offset)
889 tot_bytes -= (PAGE_SIZE - page_offset);
893 vio->vui_ra_pages += (tot_bytes + PAGE_SIZE - 1) >> PAGE_SHIFT;
895 CDEBUG(D_READA, "tot %zu, ra_start %lu, ra_count %lu\n",
896 vio->vui_tot_bytes, vio->vui_ra_start_idx,
902 LASSERT(vio->vui_iocb->ki_pos == pos);
903 iter = *vio->vui_iter;
905 lcc = ll_cl_find(inode);
906 lcc->lcc_end_index = DIV_ROUND_UP(pos + iter.count, PAGE_SIZE);
907 CDEBUG(D_VFSTRACE, "count:%ld iocb pos:%lld\n", iter.count, pos);
909 /* this seqlock lets us notice if a page has been deleted on this inode
910 * during the fault process, allowing us to catch an erroneous short
915 seq = read_seqbegin(&ll_i2info(inode)->lli_page_inv_lock);
916 result = generic_file_read_iter(vio->vui_iocb, &iter);
918 io->ci_bytes += result;
919 total_bytes_read += result;
921 /* if we got a short read or -EIO and we raced with page invalidation,
924 } while (read_seqretry(&ll_i2info(inode)->lli_page_inv_lock, seq) &&
925 ((result >= 0 && iov_iter_count(&iter) > 0)
930 if (total_bytes_read < crw_bytes)
933 } else if (result == -EIOCBQUEUED) {
934 io->ci_bytes += vio->u.readwrite.vui_read;
935 vio->vui_iocb->ki_pos = pos + vio->u.readwrite.vui_read;
941 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
942 struct cl_page_list *plist, int from, int to)
944 struct cl_2queue *queue = &io->ci_queue;
945 struct cl_page *page;
946 unsigned int bytes = 0;
950 if (plist->pl_nr == 0)
953 if (from > 0 || to != PAGE_SIZE) {
954 page = cl_page_list_first(plist);
955 if (plist->pl_nr == 1) {
956 cl_page_clip(env, page, from, to);
959 cl_page_clip(env, page, from, PAGE_SIZE);
960 if (to != PAGE_SIZE) {
961 page = cl_page_list_last(plist);
962 cl_page_clip(env, page, 0, to);
967 cl_2queue_init(queue);
968 cl_page_list_splice(plist, &queue->c2_qin);
969 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
971 /* plist is not sorted any more */
972 cl_page_list_splice(&queue->c2_qin, plist);
973 cl_page_list_splice(&queue->c2_qout, plist);
974 cl_2queue_fini(env, queue);
977 /* calculate bytes */
978 bytes = plist->pl_nr << PAGE_SHIFT;
979 bytes -= from + PAGE_SIZE - to;
981 while (plist->pl_nr > 0) {
982 page = cl_page_list_first(plist);
983 cl_page_list_del(env, plist, page);
985 cl_page_clip(env, page, 0, PAGE_SIZE);
987 SetPageUptodate(cl_page_vmpage(page));
988 cl_page_disown(env, io, page);
990 /* held in ll_cl_init() */
991 lu_ref_del(&page->cp_reference, "cl_io", io);
992 cl_page_put(env, page);
996 RETURN(bytes > 0 ? bytes : rc);
1000 * From kernel v4.19-rc5-248-g9b89a0355144 use XArrary
1001 * Prior kernels use radix_tree for tags
1003 static inline void ll_page_tag_dirty(struct page *page,
1004 struct address_space *mapping)
1006 #ifndef HAVE_RADIX_TREE_TAG_SET
1007 __xa_set_mark(&mapping->i_pages, page_index(page), PAGECACHE_TAG_DIRTY);
1009 radix_tree_tag_set(&mapping->page_tree, page_index(page),
1010 PAGECACHE_TAG_DIRTY);
1015 * Kernels 4.2 - 4.5 pass memcg argument to account_page_dirtied()
1016 * Kernel v5.2-5678-gac1c3e4 no longer exports account_page_dirtied
1018 static inline void ll_account_page_dirtied(struct page *page,
1019 struct address_space *mapping)
1021 #ifdef HAVE_ACCOUNT_PAGE_DIRTIED_3ARGS
1022 struct mem_cgroup *memcg = mem_cgroup_begin_page_stat(page);
1024 account_page_dirtied(page, mapping, memcg);
1025 mem_cgroup_end_page_stat(memcg);
1026 #elif defined(HAVE_ACCOUNT_PAGE_DIRTIED_EXPORT)
1027 account_page_dirtied(page, mapping);
1029 vvp_account_page_dirtied(page, mapping);
1031 ll_page_tag_dirty(page, mapping);
1034 /* Taken from kernel set_page_dirty, __set_page_dirty_nobuffers
1035 * Last change to this area: b93b016313b3ba8003c3b8bb71f569af91f19fc7
1037 * Current with Linus tip of tree (7/13/2019):
1038 * v5.2-rc4-224-ge01e060fe0
1040 * Backwards compat for 3.x, 5.x kernels relating to memcg handling
1041 * & rename of radix tree to xarray.
1043 static void vvp_set_pagevec_dirty(struct pagevec *pvec)
1045 struct page *page = pvec->pages[0];
1046 int count = pagevec_count(pvec);
1048 #ifdef HAVE_KALLSYMS_LOOKUP_NAME
1049 struct address_space *mapping = page->mapping;
1050 unsigned long flags;
1051 unsigned long skip_pages = 0;
1057 BUILD_BUG_ON(PAGEVEC_SIZE > BITS_PER_LONG);
1058 LASSERTF(page->mapping,
1059 "mapping must be set. page %px, page->private (cl_page) %px\n",
1060 page, (void *) page->private);
1063 * kernels without HAVE_KALLSYMS_LOOKUP_NAME also don't have
1064 * account_dirty_page exported, and if we can't access that symbol,
1065 * we can't do page dirtying in batch (taking the xarray lock only once)
1066 * so we just fall back to a looped call to __set_page_dirty_nobuffers
1068 #ifndef HAVE_ACCOUNT_PAGE_DIRTIED_EXPORT
1069 if (!vvp_account_page_dirtied) {
1070 for (i = 0; i < count; i++)
1071 __set_page_dirty_nobuffers(pvec->pages[i]);
1076 #ifdef HAVE_KALLSYMS_LOOKUP_NAME
1077 for (i = 0; i < count; i++) {
1078 page = pvec->pages[i];
1080 ClearPageReclaim(page);
1082 vvp_lock_page_memcg(page);
1083 if (TestSetPageDirty(page)) {
1084 /* page is already dirty .. no extra work needed
1085 * set a flag for the i'th page to be skipped
1087 vvp_unlock_page_memcg(page);
1088 skip_pages |= (1 << i);
1092 ll_xa_lock_irqsave(&mapping->i_pages, flags);
1094 /* Notes on differences with __set_page_dirty_nobuffers:
1095 * 1. We don't need to call page_mapping because we know this is a page
1097 * 2. We have the pages locked, so there is no need for the careful
1098 * mapping/mapping2 dance.
1099 * 3. No mapping is impossible. (Race w/truncate mentioned in
1100 * dirty_nobuffers should be impossible because we hold the page lock.)
1101 * 4. All mappings are the same because i/o is only to one file.
1103 for (i = 0; i < count; i++) {
1104 page = pvec->pages[i];
1105 /* if the i'th page was unlocked above, skip it here */
1106 if ((skip_pages >> i) & 1)
1109 LASSERTF(page->mapping == mapping,
1110 "all pages must have the same mapping. page %px, mapping %px, first mapping %px\n",
1111 page, page->mapping, mapping);
1112 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1113 ll_account_page_dirtied(page, mapping);
1115 vvp_unlock_page_memcg(page);
1117 ll_xa_unlock_irqrestore(&mapping->i_pages, flags);
1119 CDEBUG(D_VFSTRACE, "mapping %p, count %d, dirtied %d\n", mapping,
1122 if (mapping->host && dirtied) {
1123 /* !PageAnon && !swapper_space */
1124 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1130 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
1131 struct pagevec *pvec)
1138 count = pagevec_count(pvec);
1141 for (i = 0; i < count; i++) {
1142 struct page *vmpage = pvec->pages[i];
1143 SetPageUptodate(vmpage);
1146 vvp_set_pagevec_dirty(pvec);
1148 for (i = 0; i < count; i++) {
1149 struct page *vmpage = pvec->pages[i];
1150 struct cl_page *page = (struct cl_page *) vmpage->private;
1151 cl_page_disown(env, io, page);
1152 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
1153 cl_page_put(env, page);
1159 /* make sure the page list is contiguous */
1160 static bool page_list_sanity_check(struct cl_object *obj,
1161 struct cl_page_list *plist)
1163 struct cl_page *page;
1164 pgoff_t index = CL_PAGE_EOF;
1166 cl_page_list_for_each(page, plist) {
1167 if (index == CL_PAGE_EOF) {
1168 index = cl_page_index(page);
1173 if (index == cl_page_index(page))
1181 /* Return how many bytes have queued or written */
1182 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
1184 struct cl_object *obj = io->ci_obj;
1185 struct inode *inode = vvp_object_inode(obj);
1186 struct vvp_io *vio = vvp_env_io(env);
1187 struct cl_page_list *queue = &vio->u.readwrite.vui_queue;
1188 struct cl_page *page;
1191 unsigned int npages = vio->u.readwrite.vui_queue.pl_nr;
1197 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
1198 npages, vio->u.readwrite.vui_from, vio->u.readwrite.vui_to);
1200 LASSERT(page_list_sanity_check(obj, queue));
1202 /* submit IO with async write */
1203 rc = cl_io_commit_async(env, io, queue,
1204 vio->u.readwrite.vui_from,
1205 vio->u.readwrite.vui_to,
1206 write_commit_callback);
1207 npages -= queue->pl_nr; /* already committed pages */
1209 /* calculate how many bytes were written */
1210 bytes = npages << PAGE_SHIFT;
1213 bytes -= vio->u.readwrite.vui_from;
1214 if (queue->pl_nr == 0) /* last page */
1215 bytes -= PAGE_SIZE - vio->u.readwrite.vui_to;
1216 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
1218 vio->u.readwrite.vui_written += bytes;
1220 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
1221 npages, bytes, vio->u.readwrite.vui_written);
1223 /* the first page must have been written. */
1224 vio->u.readwrite.vui_from = 0;
1226 LASSERT(page_list_sanity_check(obj, queue));
1227 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
1229 /* out of quota, try sync write */
1230 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
1231 struct ll_inode_info *lli = ll_i2info(inode);
1233 rc = vvp_io_commit_sync(env, io, queue,
1234 vio->u.readwrite.vui_from,
1235 vio->u.readwrite.vui_to);
1237 vio->u.readwrite.vui_written += rc;
1240 if (lli->lli_clob != NULL)
1241 lov_read_and_clear_async_rc(lli->lli_clob);
1242 lli->lli_async_rc = 0;
1245 /* update inode size */
1246 ll_merge_attr(env, inode);
1248 /* Now the pages in queue were failed to commit, discard them
1249 * unless they were dirtied before. */
1250 while (queue->pl_nr > 0) {
1251 page = cl_page_list_first(queue);
1252 cl_page_list_del(env, queue, page);
1254 if (!PageDirty(cl_page_vmpage(page)))
1255 cl_page_discard(env, io, page);
1257 cl_page_disown(env, io, page);
1259 /* held in ll_cl_init() */
1260 lu_ref_del(&page->cp_reference, "cl_io", io);
1261 cl_page_put(env, page);
1263 cl_page_list_fini(env, queue);
1268 static int vvp_io_write_start(const struct lu_env *env,
1269 const struct cl_io_slice *ios)
1271 struct vvp_io *vio = cl2vvp_io(env, ios);
1272 struct cl_io *io = ios->cis_io;
1273 struct cl_object *obj = io->ci_obj;
1274 struct inode *inode = vvp_object_inode(obj);
1275 struct ll_inode_info *lli = ll_i2info(inode);
1276 struct file *file = vio->vui_fd->fd_file;
1278 loff_t pos = io->u.ci_wr.wr.crw_pos;
1279 size_t crw_bytes = io->u.ci_wr.wr.crw_bytes;
1280 bool lock_inode = !IS_NOSEC(inode);
1281 size_t ci_bytes = io->ci_bytes;
1282 struct iov_iter iter;
1288 trunc_sem_down_read(&lli->lli_trunc_sem);
1290 if (!can_populate_pages(env, io, inode))
1293 if (cl_io_is_append(io)) {
1295 * PARALLEL IO This has to be changed for parallel IO doing
1296 * out-of-order writes.
1298 ll_merge_attr(env, inode);
1299 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
1300 vio->vui_iocb->ki_pos = pos;
1302 LASSERTF(vio->vui_iocb->ki_pos == pos,
1303 "ki_pos %lld [%lld, %lld)\n",
1304 vio->vui_iocb->ki_pos,
1305 pos, pos + crw_bytes);
1308 CDEBUG(D_VFSTRACE, "%s: write [%llu, %llu)\n",
1309 file_dentry(file)->d_name.name, pos, pos + crw_bytes);
1311 /* The maximum Lustre file size is variable, based on the OST maximum
1312 * object size and number of stripes. This needs another check in
1313 * addition to the VFS checks earlier. */
1314 if (pos + crw_bytes > ll_file_maxbytes(inode)) {
1316 "%s: file %s ("DFID") offset %llu > maxbytes %llu\n",
1317 ll_i2sbi(inode)->ll_fsname,
1318 file_dentry(file)->d_name.name,
1319 PFID(ll_inode2fid(inode)), pos + crw_bytes,
1320 ll_file_maxbytes(inode));
1324 /* Tests to verify we take the i_mutex correctly */
1325 if (CFS_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_SEC) && !lock_inode)
1328 if (CFS_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_NOSEC) && lock_inode)
1331 flags = iocb_ki_flags_get(file, vio->vui_iocb);
1332 if (!iocb_ki_flags_check(flags, DIRECT)) {
1333 result = cl_io_lru_reserve(env, io, pos, crw_bytes);
1338 if (vio->vui_iter == NULL) {
1339 /* from a temp io in ll_cl_init(). */
1343 * When using the locked AIO function (generic_file_aio_write())
1344 * testing has shown the inode mutex to be a limiting factor
1345 * with multi-threaded single shared file performance. To get
1346 * around this, we now use the lockless version. To maintain
1347 * consistency, proper locking to protect against writes,
1348 * trucates, etc. is handled in the higher layers of lustre.
1350 lock_inode = !IS_NOSEC(inode);
1351 iter = *vio->vui_iter;
1353 if (unlikely(lock_inode))
1354 ll_inode_lock(inode);
1355 result = __generic_file_write_iter(vio->vui_iocb, &iter);
1356 if (unlikely(lock_inode))
1357 ll_inode_unlock(inode);
1361 #ifdef HAVE_GENERIC_WRITE_SYNC_2ARGS
1362 result = generic_write_sync(vio->vui_iocb, result);
1367 err = generic_write_sync(vio->vui_iocb->ki_filp, pos,
1369 if (err < 0 && result > 0)
1376 result = vvp_io_write_commit(env, io);
1377 /* Simulate short commit */
1378 if (CFS_FAULT_CHECK(OBD_FAIL_LLITE_SHORT_COMMIT)) {
1379 vio->u.readwrite.vui_written >>= 1;
1380 if (vio->u.readwrite.vui_written > 0)
1381 io->ci_need_restart = 1;
1383 if (vio->u.readwrite.vui_written > 0) {
1384 result = vio->u.readwrite.vui_written;
1385 CDEBUG(D_VFSTRACE, "%s: write bytes %zd, result: %zd\n",
1386 file_dentry(file)->d_name.name,
1387 io->ci_bytes, result);
1388 io->ci_bytes += result;
1390 io->ci_continue = 0;
1393 if (vio->vui_iocb->ki_pos != (pos + io->ci_bytes - ci_bytes)) {
1395 "%s: write position mismatch: ki_pos %lld vs. pos %lld, written %zd, commit %ld: rc = %zd\n",
1396 file_dentry(file)->d_name.name,
1397 vio->vui_iocb->ki_pos, pos + io->ci_bytes - ci_bytes,
1398 written, io->ci_bytes - ci_bytes, result);
1400 * Rewind ki_pos and vui_iter to where it has
1401 * successfully committed.
1403 vio->vui_iocb->ki_pos = pos + io->ci_bytes - ci_bytes;
1405 if (result > 0 || result == -EIOCBQUEUED) {
1406 set_bit(LLIF_DATA_MODIFIED, &ll_i2info(inode)->lli_flags);
1408 if (result != -EIOCBQUEUED && result < crw_bytes)
1409 io->ci_continue = 0;
1413 if (result == -EIOCBQUEUED) {
1414 io->ci_bytes += vio->u.readwrite.vui_written;
1415 vio->vui_iocb->ki_pos = pos +
1416 vio->u.readwrite.vui_written;
1423 static void vvp_io_rw_end(const struct lu_env *env,
1424 const struct cl_io_slice *ios)
1426 struct inode *inode = vvp_object_inode(ios->cis_obj);
1427 struct ll_inode_info *lli = ll_i2info(inode);
1429 trunc_sem_up_read(&lli->lli_trunc_sem);
1432 static void vvp_io_write_end(const struct lu_env *env,
1433 const struct cl_io_slice *ios)
1435 struct inode *inode = vvp_object_inode(ios->cis_obj);
1436 struct cl_io *io = ios->cis_io;
1438 vvp_io_rw_end(env, ios);
1440 /* Update size and blocks for LSOM (best effort) */
1441 if (!io->ci_ignore_layout && cl_io_is_sync_write(io))
1442 ll_merge_attr_try(env, inode);
1446 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
1448 struct vm_fault *vmf = cfio->ft_vmf;
1450 cfio->ft_flags = ll_filemap_fault(cfio->ft_vma, vmf);
1451 cfio->ft_flags_valid = 1;
1454 /* success, vmpage is locked */
1455 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
1456 get_vmf_address(vmf));
1457 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
1458 lock_page(vmf->page);
1459 cfio->ft_flags |= VM_FAULT_LOCKED;
1462 cfio->ft_vmpage = vmf->page;
1467 if (cfio->ft_flags & VM_FAULT_SIGBUS) {
1468 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", get_vmf_address(vmf));
1472 if (cfio->ft_flags & VM_FAULT_OOM) {
1473 CDEBUG(D_PAGE, "got addr %p - OOM\n", get_vmf_address(vmf));
1477 if (cfio->ft_flags & VM_FAULT_RETRY)
1480 CERROR("unknown error in page fault %d\n", cfio->ft_flags);
1485 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1486 struct pagevec *pvec)
1488 vvp_set_pagevec_dirty(pvec);
1491 static int vvp_io_fault_start(const struct lu_env *env,
1492 const struct cl_io_slice *ios)
1494 struct vvp_io *vio = cl2vvp_io(env, ios);
1495 struct cl_io *io = ios->cis_io;
1496 struct cl_object *obj = io->ci_obj;
1497 struct inode *inode = vvp_object_inode(obj);
1498 struct ll_inode_info *lli = ll_i2info(inode);
1499 struct cl_fault_io *fio = &io->u.ci_fault;
1500 struct vvp_fault_io *cfio = &vio->u.fault;
1503 struct page *vmpage = NULL;
1504 struct cl_page *page;
1509 trunc_sem_down_read_nowait(&lli->lli_trunc_sem);
1511 /* offset of the last byte on the page */
1512 offset = ((fio->ft_index + 1) << PAGE_SHIFT) - 1;
1513 LASSERT((offset >> PAGE_SHIFT) == fio->ft_index);
1514 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1518 /* must return locked page */
1519 if (fio->ft_mkwrite) {
1520 LASSERT(cfio->ft_vmpage != NULL);
1521 vmpage = cfio->ft_vmpage;
1524 * page was turncated and lock was cancelled, return ENODATA
1525 * so that VM_FAULT_NOPAGE will be returned to handle_mm_fault()
1526 * XXX: cannot return VM_FAULT_RETRY to vfs since we cannot
1527 * release mmap_lock and VM_FAULT_RETRY implies that the
1528 * mmap_lock is released.
1530 if (!PageUptodate(vmpage))
1531 GOTO(out, result = -ENODATA);
1533 result = vvp_io_kernel_fault(cfio);
1538 vmpage = cfio->ft_vmpage;
1539 LASSERT(PageLocked(vmpage));
1541 if (CFS_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1542 generic_error_remove_page(vmpage->mapping, vmpage);
1544 size = i_size_read(inode);
1545 /* Though we have already held a cl_lock upon this page, but
1546 * it still can be truncated locally. */
1547 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1548 (page_offset(vmpage) > size))) {
1549 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1551 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1553 GOTO(out, result = +1);
1556 last_index = (size - 1) >> PAGE_SHIFT;
1558 if (fio->ft_mkwrite ) {
1560 * Capture the size while holding the lli_trunc_sem from above
1561 * we want to make sure that we complete the mkwrite action
1562 * while holding this lock. We need to make sure that we are
1563 * not past the end of the file.
1565 if (last_index < fio->ft_index) {
1567 "llite: mkwrite and truncate race happened: "
1568 "%p: 0x%lx 0x%lx\n",
1569 vmpage->mapping,fio->ft_index,last_index);
1571 * We need to return if we are
1572 * passed the end of the file. This will propagate
1573 * up the call stack to ll_page_mkwrite where
1574 * we will return VM_FAULT_NOPAGE. Any non-negative
1575 * value returned here will be silently
1576 * converted to 0. If the vmpage->mapping is null
1577 * the error code would be converted back to ENODATA
1578 * in ll_page_mkwrite0. Thus we return -ENODATA
1579 * to handle both cases
1581 GOTO(out, result = -ENODATA);
1585 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1587 GOTO(out, result = PTR_ERR(page));
1589 /* if page is going to be written, we should add this page into cache
1591 if (fio->ft_mkwrite) {
1592 wait_on_page_writeback(vmpage);
1593 if (!PageDirty(vmpage)) {
1594 struct cl_page_list *plist = &vio->u.fault.ft_queue;
1597 /* vvp_page_assume() calls wait_on_page_writeback(). */
1598 cl_page_assume(env, io, page);
1600 cl_page_list_init(plist);
1601 cl_page_list_add(plist, page, true);
1604 if (last_index == cl_page_index(page))
1605 to = ((size - 1) & ~PAGE_MASK) + 1;
1607 /* Do not set Dirty bit here so that in case IO is
1608 * started before the page is really made dirty, we
1609 * still have chance to detect it. */
1610 result = cl_io_commit_async(env, io, plist, 0, to,
1611 mkwrite_commit_callback);
1612 /* Have overquota flag, trying sync write to check
1613 * whether indeed out of quota */
1614 if (result == -EDQUOT) {
1616 result = vvp_io_commit_sync(env, io,
1620 cl_page_own(env, io, page);
1621 cl_page_list_add(plist, page, true);
1622 lu_ref_add(&page->cp_reference,
1624 result = cl_io_commit_async(env, io,
1626 mkwrite_commit_callback);
1629 cl_page_put(env, page);
1633 LASSERT(cl_page_is_owned(page, io));
1634 cl_page_list_fini(env, plist);
1638 cl_page_discard(env, io, page);
1639 cl_page_disown(env, io, page);
1641 cl_page_put(env, page);
1643 /* we're in big trouble, what can we do now? */
1644 if (result == -EDQUOT)
1648 cl_page_disown(env, io, page);
1654 * The ft_index is only used in the case of mkwrite action. We need to
1655 * check our assertions are correct, since we should have caught this
1658 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1659 if (fio->ft_index == last_index)
1661 * Last page is mapped partially.
1663 fio->ft_bytes = size - (fio->ft_index << PAGE_SHIFT);
1665 fio->ft_bytes = PAGE_SIZE;
1667 lu_ref_add(&page->cp_reference, "fault", io);
1668 fio->ft_page = page;
1672 /* return unlocked vmpage to avoid deadlocking */
1674 unlock_page(vmpage);
1676 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1681 static void vvp_io_fault_end(const struct lu_env *env,
1682 const struct cl_io_slice *ios)
1684 struct inode *inode = vvp_object_inode(ios->cis_obj);
1685 struct ll_inode_info *lli = ll_i2info(inode);
1687 CLOBINVRNT(env, ios->cis_io->ci_obj,
1688 vvp_object_invariant(ios->cis_io->ci_obj));
1689 trunc_sem_up_read(&lli->lli_trunc_sem);
1692 static int vvp_io_fsync_start(const struct lu_env *env,
1693 const struct cl_io_slice *ios)
1695 /* we should mark TOWRITE bit to each dirty page in radix tree to
1696 * verify pages have been written, but this is difficult because of
1701 static void vvp_io_fsync_end(const struct lu_env *env,
1702 const struct cl_io_slice *ios)
1704 struct inode *inode = vvp_object_inode(ios->cis_obj);
1705 struct cl_io *io = ios->cis_io;
1707 /* Update size and blocks for LSOM (best effort) */
1708 if (!io->ci_ignore_layout)
1709 ll_merge_attr_try(env, inode);
1712 static int vvp_io_read_ahead(const struct lu_env *env,
1713 const struct cl_io_slice *ios,
1714 pgoff_t start, struct cl_read_ahead *ra)
1719 if (ios->cis_io->ci_type == CIT_READ ||
1720 ios->cis_io->ci_type == CIT_FAULT) {
1721 struct vvp_io *vio = cl2vvp_io(env, ios);
1723 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1724 ra->cra_end_idx = CL_PAGE_EOF;
1725 result = +1; /* no need to call down */
1732 static int vvp_io_lseek_lock(const struct lu_env *env,
1733 const struct cl_io_slice *ios)
1735 struct cl_io *io = ios->cis_io;
1736 __u64 lock_start = io->u.ci_lseek.ls_start;
1737 __u64 lock_end = OBD_OBJECT_EOF;
1738 __u32 enqflags = CEF_MUST; /* always take client lock */
1740 return vvp_io_one_lock(env, io, enqflags, CLM_READ,
1741 lock_start, lock_end);
1744 static int vvp_io_lseek_start(const struct lu_env *env,
1745 const struct cl_io_slice *ios)
1747 struct cl_io *io = ios->cis_io;
1748 struct inode *inode = vvp_object_inode(io->ci_obj);
1749 __u64 start = io->u.ci_lseek.ls_start;
1751 ll_inode_lock(inode);
1752 inode_dio_wait(inode);
1754 /* At the moment we have DLM lock so just update inode
1755 * to know the file size.
1757 ll_merge_attr(env, inode);
1758 if (start >= i_size_read(inode)) {
1759 io->u.ci_lseek.ls_result = -ENXIO;
1765 static void vvp_io_lseek_end(const struct lu_env *env,
1766 const struct cl_io_slice *ios)
1768 struct cl_io *io = ios->cis_io;
1769 struct inode *inode = vvp_object_inode(io->ci_obj);
1771 if (io->u.ci_lseek.ls_result > i_size_read(inode))
1772 io->u.ci_lseek.ls_result = -ENXIO;
1774 ll_inode_unlock(inode);
1777 static const struct cl_io_operations vvp_io_ops = {
1780 .cio_fini = vvp_io_fini,
1781 .cio_iter_init = vvp_io_read_iter_init,
1782 .cio_lock = vvp_io_read_lock,
1783 .cio_start = vvp_io_read_start,
1784 .cio_end = vvp_io_rw_end,
1785 .cio_advance = vvp_io_advance,
1788 .cio_fini = vvp_io_fini,
1789 .cio_iter_init = vvp_io_write_iter_init,
1790 .cio_iter_fini = vvp_io_write_iter_fini,
1791 .cio_lock = vvp_io_write_lock,
1792 .cio_start = vvp_io_write_start,
1793 .cio_end = vvp_io_write_end,
1794 .cio_advance = vvp_io_advance,
1797 .cio_fini = vvp_io_setattr_fini,
1798 .cio_iter_init = vvp_io_setattr_iter_init,
1799 .cio_lock = vvp_io_setattr_lock,
1800 .cio_start = vvp_io_setattr_start,
1801 .cio_end = vvp_io_setattr_end
1804 .cio_fini = vvp_io_fault_fini,
1805 .cio_iter_init = vvp_io_fault_iter_init,
1806 .cio_lock = vvp_io_fault_lock,
1807 .cio_start = vvp_io_fault_start,
1808 .cio_end = vvp_io_fault_end,
1811 .cio_start = vvp_io_fsync_start,
1812 .cio_fini = vvp_io_fini,
1813 .cio_end = vvp_io_fsync_end,
1816 .cio_fini = vvp_io_fini
1819 .cio_fini = vvp_io_fini
1822 .cio_fini = vvp_io_fini
1825 .cio_fini = vvp_io_fini,
1826 .cio_lock = vvp_io_lseek_lock,
1827 .cio_start = vvp_io_lseek_start,
1828 .cio_end = vvp_io_lseek_end,
1831 .cio_read_ahead = vvp_io_read_ahead
1834 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1837 struct vvp_io *vio = vvp_env_io(env);
1838 struct inode *inode = vvp_object_inode(obj);
1841 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1844 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1845 "restore needed %d\n",
1846 PFID(lu_object_fid(&obj->co_lu)),
1847 io->ci_ignore_layout, io->ci_verify_layout,
1848 vio->vui_layout_gen, io->ci_restore_needed);
1850 CL_IO_SLICE_CLEAN(vio, vui_cl);
1851 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1852 vio->vui_ra_valid = false;
1854 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1856 struct ll_inode_info *lli = ll_i2info(inode);
1858 bytes = io->u.ci_rw.crw_bytes;
1859 /* "If nbyte is 0, read() will return 0 and have no other
1860 * results." -- Single Unix Spec */
1864 vio->vui_tot_bytes = bytes;
1866 /* for read/write, we store the process jobid/gid/uid in the
1867 * inode, and it'll be fetched by osc when building RPC.
1869 * it's not accurate if the file is shared by different
1872 lustre_get_jobid(lli->lli_jobid, sizeof(lli->lli_jobid));
1873 lli->lli_uid = from_kuid(&init_user_ns, current_uid());
1874 lli->lli_gid = from_kgid(&init_user_ns, current_gid());
1875 } else if (io->ci_type == CIT_SETATTR) {
1876 if (!cl_io_is_trunc(io))
1877 io->ci_lockreq = CILR_MANDATORY;
1880 /* Enqueue layout lock and get layout version. We need to do this
1881 * even for operations requiring to open file, such as read and write,
1882 * because it might not grant layout lock in IT_OPEN. */
1883 if (result == 0 && !io->ci_ignore_layout) {
1884 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1885 if (result == -ENOENT)
1886 /* If the inode on MDS has been removed, but the objects
1887 * on OSTs haven't been destroyed (async unlink), layout
1888 * fetch will return -ENOENT, we'd ingore this error
1889 * and continue with dirty flush. LU-3230. */
1892 CERROR("%s: refresh file layout " DFID " error %d.\n",
1893 ll_i2sbi(inode)->ll_fsname,
1894 PFID(lu_object_fid(&obj->co_lu)), result);
1897 #ifdef HAVE_INVALIDATE_LOCK
1898 if (io->ci_invalidate_page_cache)
1899 filemap_invalidate_lock(inode->i_mapping);
1900 #endif /* HAVE_INVALIDATE_LOCK */
1902 io->ci_result = result < 0 ? result : 0;