4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * Implementation of cl_io for VVP layer.
34 * Author: Nikita Danilov <nikita.danilov@sun.com>
35 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
38 #define DEBUG_SUBSYSTEM S_LLITE
42 #include "llite_internal.h"
43 #include "vvp_internal.h"
45 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
46 const struct cl_io_slice *slice)
50 vio = container_of(slice, struct vvp_io, vui_cl);
51 LASSERT(vio == vvp_env_io(env));
57 * True, if \a io is a normal io, False for splice_{read,write}
59 static int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
61 struct vvp_io *vio = vvp_env_io(env);
63 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
65 return vio->vui_io_subtype == IO_NORMAL;
69 * For swapping layout. The file's layout may have changed.
70 * To avoid populating pages to a wrong stripe, we have to verify the
71 * correctness of layout. It works because swapping layout processes
72 * have to acquire group lock.
74 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
77 struct ll_inode_info *lli = ll_i2info(inode);
78 struct vvp_io *vio = vvp_env_io(env);
81 switch (io->ci_type) {
84 /* don't need lock here to check lli_layout_gen as we have held
85 * extent lock and GROUP lock has to hold to swap layout */
86 if (ll_layout_version_get(lli) != vio->vui_layout_gen ||
87 OBD_FAIL_CHECK_RESET(OBD_FAIL_LLITE_LOST_LAYOUT, 0)) {
88 io->ci_need_restart = 1;
89 /* this will cause a short read/write */
94 /* fault is okay because we've already had a page. */
102 static void vvp_object_size_lock(struct cl_object *obj)
104 struct inode *inode = vvp_object_inode(obj);
106 ll_inode_size_lock(inode);
107 cl_object_attr_lock(obj);
110 static void vvp_object_size_unlock(struct cl_object *obj)
112 struct inode *inode = vvp_object_inode(obj);
114 cl_object_attr_unlock(obj);
115 ll_inode_size_unlock(inode);
119 * Helper function that if necessary adjusts file size (inode->i_size), when
120 * position at the offset \a pos is accessed. File size can be arbitrary stale
121 * on a Lustre client, but client at least knows KMS. If accessed area is
122 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
124 * Locking: i_size_lock is used to serialize changes to inode size and to
125 * protect consistency between inode size and cl_object
126 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
127 * top-object and sub-objects.
129 static int vvp_prep_size(const struct lu_env *env, struct cl_object *obj,
130 struct cl_io *io, loff_t start, size_t count,
133 struct cl_attr *attr = vvp_env_thread_attr(env);
134 struct inode *inode = vvp_object_inode(obj);
135 loff_t pos = start + count - 1;
140 * Consistency guarantees: following possibilities exist for the
141 * relation between region being accessed and real file size at this
144 * (A): the region is completely inside of the file;
146 * (B-x): x bytes of region are inside of the file, the rest is
149 * (C): the region is completely outside of the file.
151 * This classification is stable under DLM lock already acquired by
152 * the caller, because to change the class, other client has to take
153 * DLM lock conflicting with our lock. Also, any updates to ->i_size
154 * by other threads on this client are serialized by
155 * ll_inode_size_lock(). This guarantees that short reads are handled
156 * correctly in the face of concurrent writes and truncates.
158 vvp_object_size_lock(obj);
159 result = cl_object_attr_get(env, obj, attr);
164 * A glimpse is necessary to determine whether we
165 * return a short read (B) or some zeroes at the end
168 vvp_object_size_unlock(obj);
169 result = cl_glimpse_lock(env, io, inode, obj, 0);
170 if (result == 0 && exceed != NULL) {
171 /* If objective page index exceed end-of-file
172 * page index, return directly. Do not expect
173 * kernel will check such case correctly.
174 * linux-2.6.18-128.1.1 miss to do that.
176 loff_t size = i_size_read(inode);
177 unsigned long cur_index = start >>
180 if ((size == 0 && cur_index != 0) ||
181 (((size - 1) >> PAGE_SHIFT) <
189 * region is within kms and, hence, within real file
190 * size (A). We need to increase i_size to cover the
191 * read region so that generic_file_read() will do its
192 * job, but that doesn't mean the kms size is
193 * _correct_, it is only the _minimum_ size. If
194 * someone does a stat they will get the correct size
195 * which will always be >= the kms value here.
198 if (i_size_read(inode) < kms) {
199 i_size_write(inode, kms);
201 DFID" updating i_size %llu\n",
202 PFID(lu_object_fid(&obj->co_lu)),
203 (__u64)i_size_read(inode));
208 vvp_object_size_unlock(obj);
213 /*****************************************************************************
219 static int vvp_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
220 __u32 enqflags, enum cl_lock_mode mode,
221 pgoff_t start, pgoff_t end)
223 struct vvp_io *vio = vvp_env_io(env);
224 struct cl_lock_descr *descr = &vio->vui_link.cill_descr;
225 struct cl_object *obj = io->ci_obj;
227 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
230 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
232 memset(&vio->vui_link, 0, sizeof vio->vui_link);
234 if (vio->vui_fd && (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
235 descr->cld_mode = CLM_GROUP;
236 descr->cld_gid = vio->vui_fd->fd_grouplock.lg_gid;
237 enqflags |= CEF_LOCK_MATCH;
239 descr->cld_mode = mode;
242 descr->cld_obj = obj;
243 descr->cld_start = start;
244 descr->cld_end = end;
245 descr->cld_enq_flags = enqflags;
247 cl_io_lock_add(env, io, &vio->vui_link);
252 static int vvp_io_one_lock(const struct lu_env *env, struct cl_io *io,
253 __u32 enqflags, enum cl_lock_mode mode,
254 loff_t start, loff_t end)
256 struct cl_object *obj = io->ci_obj;
258 return vvp_io_one_lock_index(env, io, enqflags, mode,
259 cl_index(obj, start), cl_index(obj, end));
262 static int vvp_io_write_iter_init(const struct lu_env *env,
263 const struct cl_io_slice *ios)
265 struct vvp_io *vio = cl2vvp_io(env, ios);
267 cl_page_list_init(&vio->u.write.vui_queue);
268 vio->u.write.vui_written = 0;
269 vio->u.write.vui_from = 0;
270 vio->u.write.vui_to = PAGE_SIZE;
275 static void vvp_io_write_iter_fini(const struct lu_env *env,
276 const struct cl_io_slice *ios)
278 struct vvp_io *vio = cl2vvp_io(env, ios);
280 LASSERT(vio->u.write.vui_queue.pl_nr == 0);
283 static int vvp_io_fault_iter_init(const struct lu_env *env,
284 const struct cl_io_slice *ios)
286 struct vvp_io *vio = cl2vvp_io(env, ios);
287 struct inode *inode = vvp_object_inode(ios->cis_obj);
289 LASSERT(inode == file_inode(vio->vui_fd->fd_file));
290 vio->u.fault.ft_mtime = inode->i_mtime.tv_sec;
295 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
297 struct cl_io *io = ios->cis_io;
298 struct cl_object *obj = io->ci_obj;
299 struct vvp_io *vio = cl2vvp_io(env, ios);
300 struct inode *inode = vvp_object_inode(obj);
305 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
307 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
308 "need write layout %d, restore needed %d\n",
309 PFID(lu_object_fid(&obj->co_lu)),
310 io->ci_ignore_layout, io->ci_verify_layout,
311 vio->vui_layout_gen, io->ci_need_write_intent,
312 io->ci_restore_needed);
314 if (io->ci_restore_needed) {
315 /* file was detected release, we need to restore it
316 * before finishing the io
318 rc = ll_layout_restore(inode, 0, OBD_OBJECT_EOF);
319 /* if restore registration failed, no restart,
320 * we will return -ENODATA */
321 /* The layout will change after restore, so we need to
322 * block on layout lock held by the MDT
323 * as MDT will not send new layout in lvb (see LU-3124)
324 * we have to explicitly fetch it, all this will be done
325 * by ll_layout_refresh().
326 * Even if ll_layout_restore() returns zero, it doesn't mean
327 * that restore has been successful. Therefore it sets
328 * ci_verify_layout so that it will check layout at the end
332 io->ci_restore_needed = 1;
333 io->ci_need_restart = 0;
334 io->ci_verify_layout = 0;
339 io->ci_restore_needed = 0;
341 /* Even if ll_layout_restore() returns zero, it doesn't mean
342 * that restore has been successful. Therefore it should verify
343 * if there was layout change and restart I/O correspondingly.
345 ll_layout_refresh(inode, &gen);
346 io->ci_need_restart = vio->vui_layout_gen != gen;
347 if (io->ci_need_restart) {
349 DFID" layout changed from %d to %d.\n",
350 PFID(lu_object_fid(&obj->co_lu)),
351 vio->vui_layout_gen, gen);
352 /* today successful restore is the only possible
354 /* restore was done, clear restoring state */
355 ll_file_clear_flag(ll_i2info(vvp_object_inode(obj)),
356 LLIF_FILE_RESTORING);
362 * dynamic layout change needed, send layout intent
365 if (io->ci_need_write_intent) {
367 loff_t end = OBD_OBJECT_EOF;
369 io->ci_need_write_intent = 0;
371 LASSERT(io->ci_type == CIT_WRITE ||
372 cl_io_is_trunc(io) || cl_io_is_mkwrite(io));
374 if (io->ci_type == CIT_WRITE) {
375 if (!cl_io_is_append(io)) {
376 start = io->u.ci_rw.rw_range.cir_pos;
377 end = start + io->u.ci_rw.rw_range.cir_count;
379 } else if (cl_io_is_trunc(io)) {
380 end = io->u.ci_setattr.sa_attr.lvb_size;
381 } else { /* mkwrite */
382 pgoff_t index = io->u.ci_fault.ft_index;
384 start = cl_offset(io->ci_obj, index);
385 end = cl_offset(io->ci_obj, index + 1);
388 CDEBUG(D_VFSTRACE, DFID" write layout, type %u [%llu, %llu)\n",
389 PFID(lu_object_fid(&obj->co_lu)), io->ci_type,
391 rc = ll_layout_write_intent(inode, start, end);
394 io->ci_need_restart = 1;
398 if (!io->ci_need_restart &&
399 !io->ci_ignore_layout && io->ci_verify_layout) {
400 /* check layout version */
401 ll_layout_refresh(inode, &gen);
402 io->ci_need_restart = vio->vui_layout_gen != gen;
403 if (io->ci_need_restart) {
405 DFID" layout changed from %d to %d.\n",
406 PFID(lu_object_fid(&obj->co_lu)),
407 vio->vui_layout_gen, gen);
415 static void vvp_io_fault_fini(const struct lu_env *env,
416 const struct cl_io_slice *ios)
418 struct cl_io *io = ios->cis_io;
419 struct cl_page *page = io->u.ci_fault.ft_page;
421 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
424 lu_ref_del(&page->cp_reference, "fault", io);
425 cl_page_put(env, page);
426 io->u.ci_fault.ft_page = NULL;
428 vvp_io_fini(env, ios);
431 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
434 * we only want to hold PW locks if the mmap() can generate
435 * writes back to the file and that only happens in shared
438 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
443 static int vvp_mmap_locks(const struct lu_env *env, struct cl_io *io)
445 struct vvp_thread_info *vti = vvp_env_info(env);
446 struct mm_struct *mm = current->mm;
447 struct vm_area_struct *vma;
448 struct cl_lock_descr *descr = &vti->vti_descr;
449 union ldlm_policy_data policy;
455 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
457 if (!cl_is_normalio(env, io))
460 /* No MM (e.g. NFS)? No vmas too. */
464 iov_for_each(iov, i, io->u.ci_rw.rw_iter) {
465 unsigned long addr = (unsigned long)iov.iov_base;
466 size_t count = iov.iov_len;
471 count += addr & ~PAGE_MASK;
474 down_read(&mm->mmap_sem);
475 while((vma = our_vma(mm, addr, count)) != NULL) {
476 struct dentry *de = file_dentry(vma->vm_file);
477 struct inode *inode = de->d_inode;
478 int flags = CEF_MUST;
480 if (ll_file_nolock(vma->vm_file)) {
482 * For no lock case is not allowed for mmap
489 * XXX: Required lock mode can be weakened: CIT_WRITE
490 * io only ever reads user level buffer, and CIT_READ
493 policy_from_vma(&policy, vma, addr, count);
494 descr->cld_mode = vvp_mode_from_vma(vma);
495 descr->cld_obj = ll_i2info(inode)->lli_clob;
496 descr->cld_start = cl_index(descr->cld_obj,
497 policy.l_extent.start);
498 descr->cld_end = cl_index(descr->cld_obj,
499 policy.l_extent.end);
500 descr->cld_enq_flags = flags;
501 result = cl_io_lock_alloc_add(env, io, descr);
503 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
504 descr->cld_mode, descr->cld_start,
510 if (vma->vm_end - addr >= count)
513 count -= vma->vm_end - addr;
516 up_read(&mm->mmap_sem);
523 static void vvp_io_advance(const struct lu_env *env,
524 const struct cl_io_slice *ios,
527 struct vvp_io *vio = cl2vvp_io(env, ios);
528 struct cl_io *io = ios->cis_io;
529 struct cl_object *obj = ios->cis_io->ci_obj;
531 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
533 if (!cl_is_normalio(env, io))
536 vio->vui_tot_count -= nob;
538 iov_iter_advance(&io->u.ci_rw.rw_iter, nob);
539 io->u.ci_rw.rw_iocb.ki_pos = io->u.ci_rw.rw_range.cir_pos;
540 #ifdef HAVE_KIOCB_KI_LEFT
541 io->u.ci_rw.rw_iocb.ki_left = vio->vui_tot_count;
542 #elif defined(HAVE_KI_NBYTES)
543 io->u.ci_rw.rw_iocb.ki_nbytes = vio->vui_tot_count;
546 /* It was truncated to stripe size in vvp_io_rw_lock() */
547 iov_iter_reexpand(&io->u.ci_rw.rw_iter, vio->vui_tot_count);
551 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
552 enum cl_lock_mode mode, loff_t start, loff_t end)
557 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
560 if (cl_is_normalio(env, io))
561 iov_iter_truncate(&io->u.ci_rw.rw_iter,
562 io->u.ci_rw.rw_range.cir_count);
564 if (io->u.ci_rw.rw_nonblock)
565 ast_flags |= CEF_NONBLOCK;
566 if (io->ci_lock_no_expand)
567 ast_flags |= CEF_LOCK_NO_EXPAND;
569 result = vvp_mmap_locks(env, io);
571 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
576 static int vvp_io_read_lock(const struct lu_env *env,
577 const struct cl_io_slice *ios)
579 struct cl_io *io = ios->cis_io;
580 struct cl_io_range *range = &io->u.ci_rw.rw_range;
584 rc = vvp_io_rw_lock(env, io, CLM_READ, range->cir_pos,
585 range->cir_pos + range->cir_count - 1);
589 static int vvp_io_fault_lock(const struct lu_env *env,
590 const struct cl_io_slice *ios)
592 struct cl_io *io = ios->cis_io;
593 struct vvp_io *vio = cl2vvp_io(env, ios);
595 * XXX LDLM_FL_CBPENDING
597 return vvp_io_one_lock_index(env,
599 vvp_mode_from_vma(vio->u.fault.ft_vma),
600 io->u.ci_fault.ft_index,
601 io->u.ci_fault.ft_index);
604 static int vvp_io_write_lock(const struct lu_env *env,
605 const struct cl_io_slice *ios)
607 struct cl_io *io = ios->cis_io;
613 if (io->u.ci_rw.rw_append) {
615 end = OBD_OBJECT_EOF;
617 start = io->u.ci_rw.rw_range.cir_pos;
618 end = start + io->u.ci_rw.rw_range.cir_count - 1;
620 rc = vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
624 static int vvp_io_setattr_iter_init(const struct lu_env *env,
625 const struct cl_io_slice *ios)
631 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
633 * Handles "lockless io" mode when extent locking is done by server.
635 static int vvp_io_setattr_lock(const struct lu_env *env,
636 const struct cl_io_slice *ios)
638 struct cl_io *io = ios->cis_io;
642 if (cl_io_is_trunc(io)) {
643 new_size = io->u.ci_setattr.sa_attr.lvb_size;
645 enqflags = CEF_DISCARD_DATA;
647 unsigned int valid = io->u.ci_setattr.sa_valid;
649 if (!(valid & TIMES_SET_FLAGS))
652 if ((!(valid & ATTR_MTIME) ||
653 io->u.ci_setattr.sa_attr.lvb_mtime >=
654 io->u.ci_setattr.sa_attr.lvb_ctime) &&
655 (!(valid & ATTR_ATIME) ||
656 io->u.ci_setattr.sa_attr.lvb_atime >=
657 io->u.ci_setattr.sa_attr.lvb_ctime))
663 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
664 new_size, OBD_OBJECT_EOF);
667 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
672 * Only ll_inode_size_lock is taken at this level.
674 ll_inode_size_lock(inode);
675 result = inode_newsize_ok(inode, size);
677 ll_inode_size_unlock(inode);
680 i_size_write(inode, size);
682 ll_truncate_pagecache(inode, size);
683 ll_inode_size_unlock(inode);
687 static int vvp_io_setattr_time(const struct lu_env *env,
688 const struct cl_io_slice *ios)
690 struct cl_io *io = ios->cis_io;
691 struct cl_object *obj = io->ci_obj;
692 struct cl_attr *attr = vvp_env_thread_attr(env);
694 unsigned valid = CAT_CTIME;
696 cl_object_attr_lock(obj);
697 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
698 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
699 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
702 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
703 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
706 result = cl_object_attr_update(env, obj, attr, valid);
707 cl_object_attr_unlock(obj);
712 static int vvp_io_setattr_start(const struct lu_env *env,
713 const struct cl_io_slice *ios)
715 struct cl_io *io = ios->cis_io;
716 struct inode *inode = vvp_object_inode(io->ci_obj);
717 struct ll_inode_info *lli = ll_i2info(inode);
719 if (cl_io_is_trunc(io)) {
720 down_write(&lli->lli_trunc_sem);
722 inode_dio_wait(inode);
727 if (io->u.ci_setattr.sa_valid & TIMES_SET_FLAGS)
728 return vvp_io_setattr_time(env, ios);
733 static void vvp_io_setattr_end(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);
740 if (cl_io_is_trunc(io)) {
741 /* Truncate in memory pages - they must be clean pages
742 * because osc has already notified to destroy osc_extents. */
743 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
744 inode_dio_write_done(inode);
746 up_write(&lli->lli_trunc_sem);
752 static void vvp_io_setattr_fini(const struct lu_env *env,
753 const struct cl_io_slice *ios)
755 bool restore_needed = ios->cis_io->ci_restore_needed;
756 struct inode *inode = vvp_object_inode(ios->cis_obj);
758 vvp_io_fini(env, ios);
760 if (restore_needed && !ios->cis_io->ci_restore_needed) {
761 /* restore finished, set data modified flag for HSM */
762 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
766 static int vvp_io_read_start(const struct lu_env *env,
767 const struct cl_io_slice *ios)
769 struct vvp_io *vio = cl2vvp_io(env, ios);
770 struct cl_io *io = ios->cis_io;
771 struct cl_object *obj = io->ci_obj;
772 struct inode *inode = vvp_object_inode(obj);
773 struct ll_inode_info *lli = ll_i2info(inode);
774 struct file *file = vio->vui_fd->fd_file;
775 struct cl_io_range *range = &io->u.ci_rw.rw_range;
776 loff_t pos = range->cir_pos; /* for generic_file_splice_read() only */
777 size_t tot = vio->vui_tot_count;
782 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
784 CDEBUG(D_VFSTRACE, "%s: read [%llu, %llu)\n",
785 file_dentry(file)->d_name.name,
786 range->cir_pos, range->cir_pos + range->cir_count);
788 if (vio->vui_io_subtype == IO_NORMAL)
789 down_read(&lli->lli_trunc_sem);
791 if (!can_populate_pages(env, io, inode))
794 /* Unless this is reading a sparse file, otherwise the lock has already
795 * been acquired so vvp_prep_size() is an empty op. */
796 result = vvp_prep_size(env, obj, io, range->cir_pos, range->cir_count,
800 else if (exceed != 0)
803 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
804 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
805 inode->i_ino, range->cir_count, range->cir_pos,
808 /* turn off the kernel's read-ahead */
809 vio->vui_fd->fd_file->f_ra.ra_pages = 0;
811 /* initialize read-ahead window once per syscall */
812 if (!vio->vui_ra_valid) {
813 vio->vui_ra_valid = true;
814 vio->vui_ra_start = cl_index(obj, range->cir_pos);
815 vio->vui_ra_count = cl_index(obj, tot + PAGE_SIZE - 1);
821 switch (vio->vui_io_subtype) {
823 LASSERTF(io->u.ci_rw.rw_iocb.ki_pos == range->cir_pos,
824 "ki_pos %lld [%lld, %lld)\n",
825 io->u.ci_rw.rw_iocb.ki_pos,
826 range->cir_pos, range->cir_pos + range->cir_count);
827 result = generic_file_read_iter(&io->u.ci_rw.rw_iocb,
828 &io->u.ci_rw.rw_iter);
831 result = generic_file_splice_read(file, &pos,
832 vio->u.splice.vui_pipe,
834 vio->u.splice.vui_flags);
835 /* LU-1109: do splice read stripe by stripe otherwise if it
836 * may make nfsd stuck if this read occupied all internal pipe
841 CERROR("Wrong IO type %u\n", vio->vui_io_subtype);
848 if (result < range->cir_count)
850 io->ci_nob += result;
851 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, vio->vui_fd,
852 range->cir_pos, result, READ);
859 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
860 struct cl_page_list *plist, int from, int to)
862 struct cl_2queue *queue = &io->ci_queue;
863 struct cl_page *page;
864 unsigned int bytes = 0;
868 if (plist->pl_nr == 0)
871 if (from > 0 || to != PAGE_SIZE) {
872 page = cl_page_list_first(plist);
873 if (plist->pl_nr == 1) {
874 cl_page_clip(env, page, from, to);
877 cl_page_clip(env, page, from, PAGE_SIZE);
878 if (to != PAGE_SIZE) {
879 page = cl_page_list_last(plist);
880 cl_page_clip(env, page, 0, to);
885 cl_2queue_init(queue);
886 cl_page_list_splice(plist, &queue->c2_qin);
887 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
889 /* plist is not sorted any more */
890 cl_page_list_splice(&queue->c2_qin, plist);
891 cl_page_list_splice(&queue->c2_qout, plist);
892 cl_2queue_fini(env, queue);
895 /* calculate bytes */
896 bytes = plist->pl_nr << PAGE_SHIFT;
897 bytes -= from + PAGE_SIZE - to;
899 while (plist->pl_nr > 0) {
900 page = cl_page_list_first(plist);
901 cl_page_list_del(env, plist, page);
903 cl_page_clip(env, page, 0, PAGE_SIZE);
905 SetPageUptodate(cl_page_vmpage(page));
906 cl_page_disown(env, io, page);
908 lu_ref_del(&page->cp_reference, "cl_io", io);
909 cl_page_put(env, page);
913 RETURN(bytes > 0 ? bytes : rc);
916 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
917 struct cl_page *page)
919 struct page *vmpage = page->cp_vmpage;
921 SetPageUptodate(vmpage);
922 set_page_dirty(vmpage);
924 cl_page_disown(env, io, page);
926 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
927 cl_page_put(env, page);
930 /* make sure the page list is contiguous */
931 static bool page_list_sanity_check(struct cl_object *obj,
932 struct cl_page_list *plist)
934 struct cl_page *page;
935 pgoff_t index = CL_PAGE_EOF;
937 cl_page_list_for_each(page, plist) {
938 struct vvp_page *vpg = cl_object_page_slice(obj, page);
940 if (index == CL_PAGE_EOF) {
941 index = vvp_index(vpg);
946 if (index == vvp_index(vpg))
954 /* Return how many bytes have queued or written */
955 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
957 struct cl_object *obj = io->ci_obj;
958 struct inode *inode = vvp_object_inode(obj);
959 struct vvp_io *vio = vvp_env_io(env);
960 struct cl_page_list *queue = &vio->u.write.vui_queue;
961 struct cl_page *page;
964 unsigned int npages = vio->u.write.vui_queue.pl_nr;
970 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
971 npages, vio->u.write.vui_from, vio->u.write.vui_to);
973 LASSERT(page_list_sanity_check(obj, queue));
975 /* submit IO with async write */
976 rc = cl_io_commit_async(env, io, queue,
977 vio->u.write.vui_from, vio->u.write.vui_to,
978 write_commit_callback);
979 npages -= queue->pl_nr; /* already committed pages */
981 /* calculate how many bytes were written */
982 bytes = npages << PAGE_SHIFT;
985 bytes -= vio->u.write.vui_from;
986 if (queue->pl_nr == 0) /* last page */
987 bytes -= PAGE_SIZE - vio->u.write.vui_to;
988 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
990 vio->u.write.vui_written += bytes;
992 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
993 npages, bytes, vio->u.write.vui_written);
995 /* the first page must have been written. */
996 vio->u.write.vui_from = 0;
998 LASSERT(page_list_sanity_check(obj, queue));
999 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
1001 /* out of quota, try sync write */
1002 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
1003 rc = vvp_io_commit_sync(env, io, queue,
1004 vio->u.write.vui_from,
1005 vio->u.write.vui_to);
1007 vio->u.write.vui_written += rc;
1012 /* update inode size */
1013 ll_merge_attr(env, inode);
1015 /* Now the pages in queue were failed to commit, discard them
1016 * unless they were dirtied before. */
1017 while (queue->pl_nr > 0) {
1018 page = cl_page_list_first(queue);
1019 cl_page_list_del(env, queue, page);
1021 if (!PageDirty(cl_page_vmpage(page)))
1022 cl_page_discard(env, io, page);
1024 cl_page_disown(env, io, page);
1026 lu_ref_del(&page->cp_reference, "cl_io", io);
1027 cl_page_put(env, page);
1029 cl_page_list_fini(env, queue);
1034 static int vvp_io_write_start(const struct lu_env *env,
1035 const struct cl_io_slice *ios)
1037 struct vvp_io *vio = cl2vvp_io(env, ios);
1038 struct cl_io *io = ios->cis_io;
1039 struct cl_object *obj = io->ci_obj;
1040 struct inode *inode = vvp_object_inode(obj);
1041 struct ll_inode_info *lli = ll_i2info(inode);
1042 struct file *file = vio->vui_fd->fd_file;
1043 struct cl_io_range *range = &io->u.ci_rw.rw_range;
1044 bool lock_inode = !lli->lli_inode_locked &&
1049 if (vio->vui_io_subtype == IO_NORMAL)
1050 down_read(&lli->lli_trunc_sem);
1052 if (!can_populate_pages(env, io, inode))
1055 if (cl_io_is_append(io)) {
1057 * PARALLEL IO This has to be changed for parallel IO doing
1058 * out-of-order writes.
1060 ll_merge_attr(env, inode);
1061 range->cir_pos = i_size_read(inode);
1062 io->u.ci_rw.rw_iocb.ki_pos = range->cir_pos;
1064 LASSERTF(io->u.ci_rw.rw_iocb.ki_pos == range->cir_pos,
1065 "ki_pos %lld [%lld, %lld)\n",
1066 io->u.ci_rw.rw_iocb.ki_pos,
1067 range->cir_pos, range->cir_pos + range->cir_count);
1070 CDEBUG(D_VFSTRACE, "%s: write [%llu, %llu)\n",
1071 file_dentry(file)->d_name.name,
1072 range->cir_pos, range->cir_pos + range->cir_count);
1074 /* The maximum Lustre file size is variable, based on the OST maximum
1075 * object size and number of stripes. This needs another check in
1076 * addition to the VFS checks earlier. */
1077 if (range->cir_pos + range->cir_count > ll_file_maxbytes(inode)) {
1079 "%s: file %s ("DFID") offset %llu > maxbytes %llu\n",
1080 ll_get_fsname(inode->i_sb, NULL, 0),
1081 file_dentry(file)->d_name.name,
1082 PFID(ll_inode2fid(inode)),
1083 range->cir_pos + range->cir_count,
1084 ll_file_maxbytes(inode));
1088 /* Tests to verify we take the i_mutex correctly */
1089 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_SEC) && !lock_inode)
1092 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_NOSEC) && lock_inode)
1096 * When using the locked AIO function (generic_file_aio_write())
1097 * testing has shown the inode mutex to be a limiting factor
1098 * with multi-threaded single shared file performance. To get
1099 * around this, we now use the lockless version. To maintain
1100 * consistency, proper locking to protect against writes,
1101 * trucates, etc. is handled in the higher layers of lustre.
1105 result = __generic_file_write_iter(&io->u.ci_rw.rw_iocb,
1106 &io->u.ci_rw.rw_iter);
1108 inode_unlock(inode);
1110 if (result > 0 || result == -EIOCBQUEUED)
1111 #ifdef HAVE_GENERIC_WRITE_SYNC_2ARGS
1112 result = generic_write_sync(&io->u.ci_rw.rw_iocb, result);
1117 err = generic_write_sync(io->u.ci_rw.rw_iocb.ki_filp,
1118 range->cir_pos, result);
1119 if (err < 0 && result > 0)
1125 result = vvp_io_write_commit(env, io);
1126 if (vio->u.write.vui_written > 0) {
1127 result = vio->u.write.vui_written;
1128 CDEBUG(D_VFSTRACE, "%s: write nob %zd, result: %zd\n",
1129 file_dentry(file)->d_name.name,
1130 io->ci_nob, result);
1131 io->ci_nob += result;
1135 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
1137 if (result < range->cir_count)
1138 io->ci_continue = 0;
1139 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
1140 vio->vui_fd, range->cir_pos, result, WRITE);
1147 static void vvp_io_rw_end(const struct lu_env *env,
1148 const struct cl_io_slice *ios)
1150 struct vvp_io *vio = cl2vvp_io(env, ios);
1151 struct inode *inode = vvp_object_inode(ios->cis_obj);
1152 struct ll_inode_info *lli = ll_i2info(inode);
1154 if (vio->vui_io_subtype == IO_NORMAL)
1155 up_read(&lli->lli_trunc_sem);
1158 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
1160 struct vm_fault *vmf = cfio->ft_vmf;
1162 cfio->ft_flags = ll_filemap_fault(cfio->ft_vma, vmf);
1163 cfio->ft_flags_valid = 1;
1166 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
1167 get_vmf_address(vmf));
1168 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
1169 lock_page(vmf->page);
1170 cfio->ft_flags |= VM_FAULT_LOCKED;
1173 cfio->ft_vmpage = vmf->page;
1178 if (cfio->ft_flags & VM_FAULT_SIGBUS) {
1179 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", get_vmf_address(vmf));
1183 if (cfio->ft_flags & VM_FAULT_OOM) {
1184 CDEBUG(D_PAGE, "got addr %p - OOM\n", get_vmf_address(vmf));
1188 if (cfio->ft_flags & VM_FAULT_RETRY)
1191 CERROR("unknown error in page fault %d\n", cfio->ft_flags);
1196 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1197 struct cl_page *page)
1199 set_page_dirty(page->cp_vmpage);
1202 static int vvp_io_fault_start(const struct lu_env *env,
1203 const struct cl_io_slice *ios)
1205 struct vvp_io *vio = cl2vvp_io(env, ios);
1206 struct cl_io *io = ios->cis_io;
1207 struct cl_object *obj = io->ci_obj;
1208 struct inode *inode = vvp_object_inode(obj);
1209 struct ll_inode_info *lli = ll_i2info(inode);
1210 struct cl_fault_io *fio = &io->u.ci_fault;
1211 struct vvp_fault_io *cfio = &vio->u.fault;
1214 struct page *vmpage = NULL;
1215 struct cl_page *page;
1220 down_read(&lli->lli_trunc_sem);
1222 /* offset of the last byte on the page */
1223 offset = cl_offset(obj, fio->ft_index + 1) - 1;
1224 LASSERT(cl_index(obj, offset) == fio->ft_index);
1225 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1229 /* must return locked page */
1230 if (fio->ft_mkwrite) {
1231 LASSERT(cfio->ft_vmpage != NULL);
1232 lock_page(cfio->ft_vmpage);
1234 result = vvp_io_kernel_fault(cfio);
1239 vmpage = cfio->ft_vmpage;
1240 LASSERT(PageLocked(vmpage));
1242 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1243 ll_invalidate_page(vmpage);
1245 size = i_size_read(inode);
1246 /* Though we have already held a cl_lock upon this page, but
1247 * it still can be truncated locally. */
1248 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1249 (page_offset(vmpage) > size))) {
1250 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1252 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1254 GOTO(out, result = +1);
1257 last_index = cl_index(obj, size - 1);
1259 if (fio->ft_mkwrite ) {
1261 * Capture the size while holding the lli_trunc_sem from above
1262 * we want to make sure that we complete the mkwrite action
1263 * while holding this lock. We need to make sure that we are
1264 * not past the end of the file.
1266 if (last_index < fio->ft_index) {
1268 "llite: mkwrite and truncate race happened: "
1269 "%p: 0x%lx 0x%lx\n",
1270 vmpage->mapping,fio->ft_index,last_index);
1272 * We need to return if we are
1273 * passed the end of the file. This will propagate
1274 * up the call stack to ll_page_mkwrite where
1275 * we will return VM_FAULT_NOPAGE. Any non-negative
1276 * value returned here will be silently
1277 * converted to 0. If the vmpage->mapping is null
1278 * the error code would be converted back to ENODATA
1279 * in ll_page_mkwrite0. Thus we return -ENODATA
1280 * to handle both cases
1282 GOTO(out, result = -ENODATA);
1286 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1288 GOTO(out, result = PTR_ERR(page));
1290 /* if page is going to be written, we should add this page into cache
1292 if (fio->ft_mkwrite) {
1293 wait_on_page_writeback(vmpage);
1294 if (!PageDirty(vmpage)) {
1295 struct cl_page_list *plist = &io->ci_queue.c2_qin;
1296 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1299 /* vvp_page_assume() calls wait_on_page_writeback(). */
1300 cl_page_assume(env, io, page);
1302 cl_page_list_init(plist);
1303 cl_page_list_add(plist, page);
1306 if (last_index == vvp_index(vpg))
1307 to = size & ~PAGE_MASK;
1309 /* Do not set Dirty bit here so that in case IO is
1310 * started before the page is really made dirty, we
1311 * still have chance to detect it. */
1312 result = cl_io_commit_async(env, io, plist, 0, to,
1313 mkwrite_commit_callback);
1314 LASSERT(cl_page_is_owned(page, io));
1315 cl_page_list_fini(env, plist);
1319 cl_page_discard(env, io, page);
1320 cl_page_disown(env, io, page);
1322 cl_page_put(env, page);
1324 /* we're in big trouble, what can we do now? */
1325 if (result == -EDQUOT)
1329 cl_page_disown(env, io, page);
1334 * The ft_index is only used in the case of
1335 * a mkwrite action. We need to check
1336 * our assertions are correct, since
1337 * we should have caught this above
1339 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1340 if (fio->ft_index == last_index)
1342 * Last page is mapped partially.
1344 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1346 fio->ft_nob = cl_page_size(obj);
1348 lu_ref_add(&page->cp_reference, "fault", io);
1349 fio->ft_page = page;
1353 /* return unlocked vmpage to avoid deadlocking */
1355 unlock_page(vmpage);
1357 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1362 static void vvp_io_fault_end(const struct lu_env *env,
1363 const struct cl_io_slice *ios)
1365 struct inode *inode = vvp_object_inode(ios->cis_obj);
1366 struct ll_inode_info *lli = ll_i2info(inode);
1368 CLOBINVRNT(env, ios->cis_io->ci_obj,
1369 vvp_object_invariant(ios->cis_io->ci_obj));
1370 up_read(&lli->lli_trunc_sem);
1373 static int vvp_io_fsync_start(const struct lu_env *env,
1374 const struct cl_io_slice *ios)
1376 /* we should mark TOWRITE bit to each dirty page in radix tree to
1377 * verify pages have been written, but this is difficult because of
1382 static int vvp_io_read_ahead(const struct lu_env *env,
1383 const struct cl_io_slice *ios,
1384 pgoff_t start, struct cl_read_ahead *ra)
1389 if (ios->cis_io->ci_type == CIT_READ ||
1390 ios->cis_io->ci_type == CIT_FAULT) {
1391 struct vvp_io *vio = cl2vvp_io(env, ios);
1393 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1394 ra->cra_end = CL_PAGE_EOF;
1395 result = +1; /* no need to call down */
1402 static const struct cl_io_operations vvp_io_ops = {
1405 .cio_fini = vvp_io_fini,
1406 .cio_lock = vvp_io_read_lock,
1407 .cio_start = vvp_io_read_start,
1408 .cio_end = vvp_io_rw_end,
1409 .cio_advance = vvp_io_advance,
1412 .cio_fini = vvp_io_fini,
1413 .cio_iter_init = vvp_io_write_iter_init,
1414 .cio_iter_fini = vvp_io_write_iter_fini,
1415 .cio_lock = vvp_io_write_lock,
1416 .cio_start = vvp_io_write_start,
1417 .cio_end = vvp_io_rw_end,
1418 .cio_advance = vvp_io_advance,
1421 .cio_fini = vvp_io_setattr_fini,
1422 .cio_iter_init = vvp_io_setattr_iter_init,
1423 .cio_lock = vvp_io_setattr_lock,
1424 .cio_start = vvp_io_setattr_start,
1425 .cio_end = vvp_io_setattr_end
1428 .cio_fini = vvp_io_fault_fini,
1429 .cio_iter_init = vvp_io_fault_iter_init,
1430 .cio_lock = vvp_io_fault_lock,
1431 .cio_start = vvp_io_fault_start,
1432 .cio_end = vvp_io_fault_end,
1435 .cio_start = vvp_io_fsync_start,
1436 .cio_fini = vvp_io_fini
1439 .cio_fini = vvp_io_fini
1442 .cio_fini = vvp_io_fini
1445 .cio_fini = vvp_io_fini
1448 .cio_read_ahead = vvp_io_read_ahead
1451 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1454 struct vvp_io *vio = vvp_env_io(env);
1455 struct inode *inode = vvp_object_inode(obj);
1458 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1461 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1462 "restore needed %d\n",
1463 PFID(lu_object_fid(&obj->co_lu)),
1464 io->ci_ignore_layout, io->ci_verify_layout,
1465 vio->vui_layout_gen, io->ci_restore_needed);
1467 CL_IO_SLICE_CLEAN(vio, vui_cl);
1468 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1469 vio->vui_ra_valid = false;
1471 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1472 struct ll_inode_info *lli = ll_i2info(inode);
1474 vio->vui_tot_count = io->u.ci_rw.rw_range.cir_count;
1475 /* "If nbyte is 0, read() will return 0 and have no other
1476 * results." -- Single Unix Spec */
1477 if (vio->vui_tot_count == 0)
1480 /* for read/write, we store the jobid in the inode, and
1481 * it'll be fetched by osc when building RPC.
1483 * it's not accurate if the file is shared by different
1486 lustre_get_jobid(lli->lli_jobid);
1487 } else if (io->ci_type == CIT_SETATTR) {
1488 if (!cl_io_is_trunc(io))
1489 io->ci_lockreq = CILR_MANDATORY;
1492 /* Enqueue layout lock and get layout version. We need to do this
1493 * even for operations requiring to open file, such as read and write,
1494 * because it might not grant layout lock in IT_OPEN. */
1495 if (result == 0 && !io->ci_ignore_layout) {
1496 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1497 if (result == -ENOENT)
1498 /* If the inode on MDS has been removed, but the objects
1499 * on OSTs haven't been destroyed (async unlink), layout
1500 * fetch will return -ENOENT, we'd ingore this error
1501 * and continue with dirty flush. LU-3230. */
1504 CERROR("%s: refresh file layout " DFID " error %d.\n",
1505 ll_get_fsname(inode->i_sb, NULL, 0),
1506 PFID(lu_object_fid(&obj->co_lu)), result);
1509 io->ci_result = result < 0 ? result : 0;