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/
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) {
366 enum layout_intent_opc opc = LAYOUT_INTENT_WRITE;
368 io->ci_need_write_intent = 0;
370 LASSERT(io->ci_type == CIT_WRITE ||
371 cl_io_is_trunc(io) || cl_io_is_mkwrite(io));
373 CDEBUG(D_VFSTRACE, DFID" write layout, type %u "DEXT"\n",
374 PFID(lu_object_fid(&obj->co_lu)), io->ci_type,
375 PEXT(&io->ci_write_intent));
377 if (cl_io_is_trunc(io))
378 opc = LAYOUT_INTENT_TRUNC;
380 rc = ll_layout_write_intent(inode, opc, &io->ci_write_intent);
383 io->ci_need_restart = 1;
387 if (!io->ci_need_restart &&
388 !io->ci_ignore_layout && io->ci_verify_layout) {
389 /* check layout version */
390 ll_layout_refresh(inode, &gen);
391 io->ci_need_restart = vio->vui_layout_gen != gen;
392 if (io->ci_need_restart) {
394 DFID" layout changed from %d to %d.\n",
395 PFID(lu_object_fid(&obj->co_lu)),
396 vio->vui_layout_gen, gen);
404 static void vvp_io_fault_fini(const struct lu_env *env,
405 const struct cl_io_slice *ios)
407 struct cl_io *io = ios->cis_io;
408 struct cl_page *page = io->u.ci_fault.ft_page;
410 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
413 lu_ref_del(&page->cp_reference, "fault", io);
414 cl_page_put(env, page);
415 io->u.ci_fault.ft_page = NULL;
417 vvp_io_fini(env, ios);
420 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
423 * we only want to hold PW locks if the mmap() can generate
424 * writes back to the file and that only happens in shared
427 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
432 static int vvp_mmap_locks(const struct lu_env *env,
433 struct vvp_io *vio, struct cl_io *io)
435 struct vvp_thread_info *vti = vvp_env_info(env);
436 struct mm_struct *mm = current->mm;
437 struct vm_area_struct *vma;
438 struct cl_lock_descr *descr = &vti->vti_descr;
439 union ldlm_policy_data policy;
445 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
447 if (!cl_is_normalio(env, io))
450 /* nfs or loop back device write */
451 if (vio->vui_iter == NULL)
454 /* No MM (e.g. NFS)? No vmas too. */
458 iov_for_each(iov, i, *(vio->vui_iter)) {
459 unsigned long addr = (unsigned long)iov.iov_base;
460 size_t count = iov.iov_len;
465 count += addr & ~PAGE_MASK;
468 down_read(&mm->mmap_sem);
469 while((vma = our_vma(mm, addr, count)) != NULL) {
470 struct dentry *de = file_dentry(vma->vm_file);
471 struct inode *inode = de->d_inode;
472 int flags = CEF_MUST;
474 if (ll_file_nolock(vma->vm_file)) {
476 * For no lock case is not allowed for mmap
483 * XXX: Required lock mode can be weakened: CIT_WRITE
484 * io only ever reads user level buffer, and CIT_READ
487 policy_from_vma(&policy, vma, addr, count);
488 descr->cld_mode = vvp_mode_from_vma(vma);
489 descr->cld_obj = ll_i2info(inode)->lli_clob;
490 descr->cld_start = cl_index(descr->cld_obj,
491 policy.l_extent.start);
492 descr->cld_end = cl_index(descr->cld_obj,
493 policy.l_extent.end);
494 descr->cld_enq_flags = flags;
495 result = cl_io_lock_alloc_add(env, io, descr);
497 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
498 descr->cld_mode, descr->cld_start,
504 if (vma->vm_end - addr >= count)
507 count -= vma->vm_end - addr;
510 up_read(&mm->mmap_sem);
517 static void vvp_io_advance(const struct lu_env *env,
518 const struct cl_io_slice *ios,
521 struct vvp_io *vio = cl2vvp_io(env, ios);
522 struct cl_io *io = ios->cis_io;
523 struct cl_object *obj = ios->cis_io->ci_obj;
525 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
527 if (!cl_is_normalio(env, io))
530 vio->vui_tot_count -= nob;
531 iov_iter_reexpand(vio->vui_iter, vio->vui_tot_count);
534 static void vvp_io_update_iov(const struct lu_env *env,
535 struct vvp_io *vio, struct cl_io *io)
537 size_t size = io->u.ci_rw.crw_count;
539 if (!cl_is_normalio(env, io) || vio->vui_iter == NULL)
542 iov_iter_truncate(vio->vui_iter, size);
545 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
546 enum cl_lock_mode mode, loff_t start, loff_t end)
548 struct vvp_io *vio = vvp_env_io(env);
552 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
555 vvp_io_update_iov(env, vio, io);
557 if (io->u.ci_rw.crw_nonblock)
558 ast_flags |= CEF_NONBLOCK;
559 if (io->ci_lock_no_expand)
560 ast_flags |= CEF_LOCK_NO_EXPAND;
562 result = vvp_mmap_locks(env, vio, io);
564 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
569 static int vvp_io_read_lock(const struct lu_env *env,
570 const struct cl_io_slice *ios)
572 struct cl_io *io = ios->cis_io;
573 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
577 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
578 rd->crw_pos + rd->crw_count - 1);
582 static int vvp_io_fault_lock(const struct lu_env *env,
583 const struct cl_io_slice *ios)
585 struct cl_io *io = ios->cis_io;
586 struct vvp_io *vio = cl2vvp_io(env, ios);
588 * XXX LDLM_FL_CBPENDING
590 return vvp_io_one_lock_index(env,
592 vvp_mode_from_vma(vio->u.fault.ft_vma),
593 io->u.ci_fault.ft_index,
594 io->u.ci_fault.ft_index);
597 static int vvp_io_write_lock(const struct lu_env *env,
598 const struct cl_io_slice *ios)
600 struct cl_io *io = ios->cis_io;
604 if (io->u.ci_wr.wr_append) {
606 end = OBD_OBJECT_EOF;
608 start = io->u.ci_wr.wr.crw_pos;
609 end = start + io->u.ci_wr.wr.crw_count - 1;
612 RETURN(vvp_io_rw_lock(env, io, CLM_WRITE, start, end));
615 static int vvp_io_setattr_iter_init(const struct lu_env *env,
616 const struct cl_io_slice *ios)
623 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
625 * Handles "lockless io" mode when extent locking is done by server.
627 static int vvp_io_setattr_lock(const struct lu_env *env,
628 const struct cl_io_slice *ios)
630 struct cl_io *io = ios->cis_io;
634 if (cl_io_is_trunc(io)) {
635 new_size = io->u.ci_setattr.sa_attr.lvb_size;
637 enqflags = CEF_DISCARD_DATA;
639 unsigned int valid = io->u.ci_setattr.sa_avalid;
641 if (!(valid & TIMES_SET_FLAGS))
644 if ((!(valid & ATTR_MTIME) ||
645 io->u.ci_setattr.sa_attr.lvb_mtime >=
646 io->u.ci_setattr.sa_attr.lvb_ctime) &&
647 (!(valid & ATTR_ATIME) ||
648 io->u.ci_setattr.sa_attr.lvb_atime >=
649 io->u.ci_setattr.sa_attr.lvb_ctime))
655 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
656 new_size, OBD_OBJECT_EOF);
659 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
664 * Only ll_inode_size_lock is taken at this level.
666 ll_inode_size_lock(inode);
667 result = inode_newsize_ok(inode, size);
669 ll_inode_size_unlock(inode);
672 i_size_write(inode, size);
674 ll_truncate_pagecache(inode, size);
675 ll_inode_size_unlock(inode);
679 static int vvp_io_setattr_time(const struct lu_env *env,
680 const struct cl_io_slice *ios)
682 struct cl_io *io = ios->cis_io;
683 struct cl_object *obj = io->ci_obj;
684 struct cl_attr *attr = vvp_env_thread_attr(env);
686 unsigned valid = CAT_CTIME;
688 cl_object_attr_lock(obj);
689 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
690 if (io->u.ci_setattr.sa_avalid & ATTR_ATIME_SET) {
691 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
694 if (io->u.ci_setattr.sa_avalid & ATTR_MTIME_SET) {
695 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
698 result = cl_object_attr_update(env, obj, attr, valid);
699 cl_object_attr_unlock(obj);
704 static int vvp_io_setattr_start(const struct lu_env *env,
705 const struct cl_io_slice *ios)
707 struct cl_io *io = ios->cis_io;
708 struct inode *inode = vvp_object_inode(io->ci_obj);
709 struct ll_inode_info *lli = ll_i2info(inode);
711 if (cl_io_is_trunc(io)) {
712 down_write(&lli->lli_trunc_sem);
714 inode_dio_wait(inode);
719 if (io->u.ci_setattr.sa_avalid & TIMES_SET_FLAGS)
720 return vvp_io_setattr_time(env, ios);
725 static void vvp_io_setattr_end(const struct lu_env *env,
726 const struct cl_io_slice *ios)
728 struct cl_io *io = ios->cis_io;
729 struct inode *inode = vvp_object_inode(io->ci_obj);
730 struct ll_inode_info *lli = ll_i2info(inode);
732 if (cl_io_is_trunc(io)) {
733 /* Truncate in memory pages - they must be clean pages
734 * because osc has already notified to destroy osc_extents. */
735 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
736 inode_dio_write_done(inode);
738 up_write(&lli->lli_trunc_sem);
744 static void vvp_io_setattr_fini(const struct lu_env *env,
745 const struct cl_io_slice *ios)
747 bool restore_needed = ios->cis_io->ci_restore_needed;
748 struct inode *inode = vvp_object_inode(ios->cis_obj);
750 vvp_io_fini(env, ios);
752 if (restore_needed && !ios->cis_io->ci_restore_needed) {
753 /* restore finished, set data modified flag for HSM */
754 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
758 static int vvp_io_read_start(const struct lu_env *env,
759 const struct cl_io_slice *ios)
761 struct vvp_io *vio = cl2vvp_io(env, ios);
762 struct cl_io *io = ios->cis_io;
763 struct cl_object *obj = io->ci_obj;
764 struct inode *inode = vvp_object_inode(obj);
765 struct ll_inode_info *lli = ll_i2info(inode);
766 struct file *file = vio->vui_fd->fd_file;
767 loff_t pos = io->u.ci_rd.rd.crw_pos;
768 long cnt = io->u.ci_rd.rd.crw_count;
769 long tot = vio->vui_tot_count;
774 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
776 CDEBUG(D_VFSTRACE, "%s: read [%llu, %llu)\n",
777 file_dentry(file)->d_name.name,
780 if (vio->vui_io_subtype == IO_NORMAL)
781 down_read(&lli->lli_trunc_sem);
783 if (!can_populate_pages(env, io, inode))
786 /* Unless this is reading a sparse file, otherwise the lock has already
787 * been acquired so vvp_prep_size() is an empty op. */
788 result = vvp_prep_size(env, obj, io, pos, cnt, &exceed);
791 else if (exceed != 0)
794 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
795 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
796 inode->i_ino, cnt, pos, i_size_read(inode));
798 /* turn off the kernel's read-ahead */
799 vio->vui_fd->fd_file->f_ra.ra_pages = 0;
801 /* initialize read-ahead window once per syscall */
802 if (!vio->vui_ra_valid) {
803 vio->vui_ra_valid = true;
804 vio->vui_ra_start = cl_index(obj, pos);
805 vio->vui_ra_count = cl_index(obj, tot + PAGE_SIZE - 1);
811 switch (vio->vui_io_subtype) {
813 LASSERT(vio->vui_iocb->ki_pos == pos);
814 result = generic_file_read_iter(vio->vui_iocb, vio->vui_iter);
817 result = generic_file_splice_read(file, &pos,
818 vio->u.splice.vui_pipe, cnt,
819 vio->u.splice.vui_flags);
820 /* LU-1109: do splice read stripe by stripe otherwise if it
821 * may make nfsd stuck if this read occupied all internal pipe
826 CERROR("Wrong IO type %u\n", vio->vui_io_subtype);
835 io->ci_nob += result;
836 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, vio->vui_fd,
844 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
845 struct cl_page_list *plist, int from, int to)
847 struct cl_2queue *queue = &io->ci_queue;
848 struct cl_page *page;
849 unsigned int bytes = 0;
853 if (plist->pl_nr == 0)
856 if (from > 0 || to != PAGE_SIZE) {
857 page = cl_page_list_first(plist);
858 if (plist->pl_nr == 1) {
859 cl_page_clip(env, page, from, to);
862 cl_page_clip(env, page, from, PAGE_SIZE);
863 if (to != PAGE_SIZE) {
864 page = cl_page_list_last(plist);
865 cl_page_clip(env, page, 0, to);
870 cl_2queue_init(queue);
871 cl_page_list_splice(plist, &queue->c2_qin);
872 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
874 /* plist is not sorted any more */
875 cl_page_list_splice(&queue->c2_qin, plist);
876 cl_page_list_splice(&queue->c2_qout, plist);
877 cl_2queue_fini(env, queue);
880 /* calculate bytes */
881 bytes = plist->pl_nr << PAGE_SHIFT;
882 bytes -= from + PAGE_SIZE - to;
884 while (plist->pl_nr > 0) {
885 page = cl_page_list_first(plist);
886 cl_page_list_del(env, plist, page);
888 cl_page_clip(env, page, 0, PAGE_SIZE);
890 SetPageUptodate(cl_page_vmpage(page));
891 cl_page_disown(env, io, page);
893 /* held in ll_cl_init() */
894 lu_ref_del(&page->cp_reference, "cl_io", io);
895 cl_page_put(env, page);
899 RETURN(bytes > 0 ? bytes : rc);
902 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
903 struct cl_page *page)
905 struct page *vmpage = page->cp_vmpage;
907 SetPageUptodate(vmpage);
908 set_page_dirty(vmpage);
910 cl_page_disown(env, io, page);
912 /* held in ll_cl_init() */
913 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
914 cl_page_put(env, page);
917 /* make sure the page list is contiguous */
918 static bool page_list_sanity_check(struct cl_object *obj,
919 struct cl_page_list *plist)
921 struct cl_page *page;
922 pgoff_t index = CL_PAGE_EOF;
924 cl_page_list_for_each(page, plist) {
925 struct vvp_page *vpg = cl_object_page_slice(obj, page);
927 if (index == CL_PAGE_EOF) {
928 index = vvp_index(vpg);
933 if (index == vvp_index(vpg))
941 /* Return how many bytes have queued or written */
942 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
944 struct cl_object *obj = io->ci_obj;
945 struct inode *inode = vvp_object_inode(obj);
946 struct vvp_io *vio = vvp_env_io(env);
947 struct cl_page_list *queue = &vio->u.write.vui_queue;
948 struct cl_page *page;
951 unsigned int npages = vio->u.write.vui_queue.pl_nr;
957 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
958 npages, vio->u.write.vui_from, vio->u.write.vui_to);
960 LASSERT(page_list_sanity_check(obj, queue));
962 /* submit IO with async write */
963 rc = cl_io_commit_async(env, io, queue,
964 vio->u.write.vui_from, vio->u.write.vui_to,
965 write_commit_callback);
966 npages -= queue->pl_nr; /* already committed pages */
968 /* calculate how many bytes were written */
969 bytes = npages << PAGE_SHIFT;
972 bytes -= vio->u.write.vui_from;
973 if (queue->pl_nr == 0) /* last page */
974 bytes -= PAGE_SIZE - vio->u.write.vui_to;
975 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
977 vio->u.write.vui_written += bytes;
979 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
980 npages, bytes, vio->u.write.vui_written);
982 /* the first page must have been written. */
983 vio->u.write.vui_from = 0;
985 LASSERT(page_list_sanity_check(obj, queue));
986 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
988 /* out of quota, try sync write */
989 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
990 rc = vvp_io_commit_sync(env, io, queue,
991 vio->u.write.vui_from,
992 vio->u.write.vui_to);
994 vio->u.write.vui_written += rc;
999 /* update inode size */
1000 ll_merge_attr(env, inode);
1002 /* Now the pages in queue were failed to commit, discard them
1003 * unless they were dirtied before. */
1004 while (queue->pl_nr > 0) {
1005 page = cl_page_list_first(queue);
1006 cl_page_list_del(env, queue, page);
1008 if (!PageDirty(cl_page_vmpage(page)))
1009 cl_page_discard(env, io, page);
1011 cl_page_disown(env, io, page);
1013 /* held in ll_cl_init() */
1014 lu_ref_del(&page->cp_reference, "cl_io", io);
1015 cl_page_put(env, page);
1017 cl_page_list_fini(env, queue);
1022 static int vvp_io_write_start(const struct lu_env *env,
1023 const struct cl_io_slice *ios)
1025 struct vvp_io *vio = cl2vvp_io(env, ios);
1026 struct cl_io *io = ios->cis_io;
1027 struct cl_object *obj = io->ci_obj;
1028 struct inode *inode = vvp_object_inode(obj);
1029 struct ll_inode_info *lli = ll_i2info(inode);
1030 struct file *file = vio->vui_fd->fd_file;
1032 loff_t pos = io->u.ci_wr.wr.crw_pos;
1033 size_t cnt = io->u.ci_wr.wr.crw_count;
1034 bool lock_inode = !IS_NOSEC(inode);
1038 if (vio->vui_io_subtype == IO_NORMAL)
1039 down_read(&lli->lli_trunc_sem);
1041 if (!can_populate_pages(env, io, inode))
1044 if (cl_io_is_append(io)) {
1046 * PARALLEL IO This has to be changed for parallel IO doing
1047 * out-of-order writes.
1049 ll_merge_attr(env, inode);
1050 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
1051 vio->vui_iocb->ki_pos = pos;
1053 LASSERTF(vio->vui_iocb->ki_pos == pos,
1054 "ki_pos %lld [%lld, %lld)\n",
1055 vio->vui_iocb->ki_pos,
1059 CDEBUG(D_VFSTRACE, "%s: write [%llu, %llu)\n",
1060 file_dentry(file)->d_name.name,
1063 /* The maximum Lustre file size is variable, based on the OST maximum
1064 * object size and number of stripes. This needs another check in
1065 * addition to the VFS checks earlier. */
1066 if (pos + cnt > ll_file_maxbytes(inode)) {
1068 "%s: file %s ("DFID") offset %llu > maxbytes %llu\n",
1069 ll_get_fsname(inode->i_sb, NULL, 0),
1070 file_dentry(file)->d_name.name,
1071 PFID(ll_inode2fid(inode)), pos + cnt,
1072 ll_file_maxbytes(inode));
1076 /* Tests to verify we take the i_mutex correctly */
1077 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_SEC) && !lock_inode)
1080 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_NOSEC) && lock_inode)
1083 if (vio->vui_iter == NULL) {
1084 /* from a temp io in ll_cl_init(). */
1088 * When using the locked AIO function (generic_file_aio_write())
1089 * testing has shown the inode mutex to be a limiting factor
1090 * with multi-threaded single shared file performance. To get
1091 * around this, we now use the lockless version. To maintain
1092 * consistency, proper locking to protect against writes,
1093 * trucates, etc. is handled in the higher layers of lustre.
1095 bool lock_node = !IS_NOSEC(inode);
1099 result = __generic_file_write_iter(vio->vui_iocb,
1102 inode_unlock(inode);
1104 if (result > 0 || result == -EIOCBQUEUED)
1105 #ifdef HAVE_GENERIC_WRITE_SYNC_2ARGS
1106 result = generic_write_sync(vio->vui_iocb, result);
1111 err = generic_write_sync(vio->vui_iocb->ki_filp, pos,
1113 if (err < 0 && result > 0)
1120 result = vvp_io_write_commit(env, io);
1121 if (vio->u.write.vui_written > 0) {
1122 result = vio->u.write.vui_written;
1123 CDEBUG(D_VFSTRACE, "%s: write nob %zd, result: %zd\n",
1124 file_dentry(file)->d_name.name,
1125 io->ci_nob, result);
1126 io->ci_nob += result;
1130 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
1133 io->ci_continue = 0;
1134 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
1135 vio->vui_fd, pos, result, WRITE);
1142 static void vvp_io_rw_end(const struct lu_env *env,
1143 const struct cl_io_slice *ios)
1145 struct vvp_io *vio = cl2vvp_io(env, ios);
1146 struct inode *inode = vvp_object_inode(ios->cis_obj);
1147 struct ll_inode_info *lli = ll_i2info(inode);
1149 if (vio->vui_io_subtype == IO_NORMAL)
1150 up_read(&lli->lli_trunc_sem);
1153 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
1155 struct vm_fault *vmf = cfio->ft_vmf;
1157 cfio->ft_flags = ll_filemap_fault(cfio->ft_vma, vmf);
1158 cfio->ft_flags_valid = 1;
1161 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
1162 get_vmf_address(vmf));
1163 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
1164 lock_page(vmf->page);
1165 cfio->ft_flags |= VM_FAULT_LOCKED;
1168 cfio->ft_vmpage = vmf->page;
1173 if (cfio->ft_flags & VM_FAULT_SIGBUS) {
1174 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", get_vmf_address(vmf));
1178 if (cfio->ft_flags & VM_FAULT_OOM) {
1179 CDEBUG(D_PAGE, "got addr %p - OOM\n", get_vmf_address(vmf));
1183 if (cfio->ft_flags & VM_FAULT_RETRY)
1186 CERROR("unknown error in page fault %d\n", cfio->ft_flags);
1191 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1192 struct cl_page *page)
1194 set_page_dirty(page->cp_vmpage);
1197 static int vvp_io_fault_start(const struct lu_env *env,
1198 const struct cl_io_slice *ios)
1200 struct vvp_io *vio = cl2vvp_io(env, ios);
1201 struct cl_io *io = ios->cis_io;
1202 struct cl_object *obj = io->ci_obj;
1203 struct inode *inode = vvp_object_inode(obj);
1204 struct ll_inode_info *lli = ll_i2info(inode);
1205 struct cl_fault_io *fio = &io->u.ci_fault;
1206 struct vvp_fault_io *cfio = &vio->u.fault;
1209 struct page *vmpage = NULL;
1210 struct cl_page *page;
1215 down_read(&lli->lli_trunc_sem);
1217 /* offset of the last byte on the page */
1218 offset = cl_offset(obj, fio->ft_index + 1) - 1;
1219 LASSERT(cl_index(obj, offset) == fio->ft_index);
1220 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1224 /* must return locked page */
1225 if (fio->ft_mkwrite) {
1226 LASSERT(cfio->ft_vmpage != NULL);
1227 lock_page(cfio->ft_vmpage);
1229 result = vvp_io_kernel_fault(cfio);
1234 vmpage = cfio->ft_vmpage;
1235 LASSERT(PageLocked(vmpage));
1237 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1238 ll_invalidate_page(vmpage);
1240 size = i_size_read(inode);
1241 /* Though we have already held a cl_lock upon this page, but
1242 * it still can be truncated locally. */
1243 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1244 (page_offset(vmpage) > size))) {
1245 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1247 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1249 GOTO(out, result = +1);
1252 last_index = cl_index(obj, size - 1);
1254 if (fio->ft_mkwrite ) {
1256 * Capture the size while holding the lli_trunc_sem from above
1257 * we want to make sure that we complete the mkwrite action
1258 * while holding this lock. We need to make sure that we are
1259 * not past the end of the file.
1261 if (last_index < fio->ft_index) {
1263 "llite: mkwrite and truncate race happened: "
1264 "%p: 0x%lx 0x%lx\n",
1265 vmpage->mapping,fio->ft_index,last_index);
1267 * We need to return if we are
1268 * passed the end of the file. This will propagate
1269 * up the call stack to ll_page_mkwrite where
1270 * we will return VM_FAULT_NOPAGE. Any non-negative
1271 * value returned here will be silently
1272 * converted to 0. If the vmpage->mapping is null
1273 * the error code would be converted back to ENODATA
1274 * in ll_page_mkwrite0. Thus we return -ENODATA
1275 * to handle both cases
1277 GOTO(out, result = -ENODATA);
1281 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1283 GOTO(out, result = PTR_ERR(page));
1285 /* if page is going to be written, we should add this page into cache
1287 if (fio->ft_mkwrite) {
1288 wait_on_page_writeback(vmpage);
1289 if (!PageDirty(vmpage)) {
1290 struct cl_page_list *plist = &io->ci_queue.c2_qin;
1291 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1294 /* vvp_page_assume() calls wait_on_page_writeback(). */
1295 cl_page_assume(env, io, page);
1297 cl_page_list_init(plist);
1298 cl_page_list_add(plist, page);
1301 if (last_index == vvp_index(vpg))
1302 to = size & ~PAGE_MASK;
1304 /* Do not set Dirty bit here so that in case IO is
1305 * started before the page is really made dirty, we
1306 * still have chance to detect it. */
1307 result = cl_io_commit_async(env, io, plist, 0, to,
1308 mkwrite_commit_callback);
1309 LASSERT(cl_page_is_owned(page, io));
1310 cl_page_list_fini(env, plist);
1314 cl_page_discard(env, io, page);
1315 cl_page_disown(env, io, page);
1317 cl_page_put(env, page);
1319 /* we're in big trouble, what can we do now? */
1320 if (result == -EDQUOT)
1324 cl_page_disown(env, io, page);
1329 * The ft_index is only used in the case of
1330 * a mkwrite action. We need to check
1331 * our assertions are correct, since
1332 * we should have caught this above
1334 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1335 if (fio->ft_index == last_index)
1337 * Last page is mapped partially.
1339 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1341 fio->ft_nob = cl_page_size(obj);
1343 lu_ref_add(&page->cp_reference, "fault", io);
1344 fio->ft_page = page;
1348 /* return unlocked vmpage to avoid deadlocking */
1350 unlock_page(vmpage);
1352 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1357 static void vvp_io_fault_end(const struct lu_env *env,
1358 const struct cl_io_slice *ios)
1360 struct inode *inode = vvp_object_inode(ios->cis_obj);
1361 struct ll_inode_info *lli = ll_i2info(inode);
1363 CLOBINVRNT(env, ios->cis_io->ci_obj,
1364 vvp_object_invariant(ios->cis_io->ci_obj));
1365 up_read(&lli->lli_trunc_sem);
1368 static int vvp_io_fsync_start(const struct lu_env *env,
1369 const struct cl_io_slice *ios)
1371 /* we should mark TOWRITE bit to each dirty page in radix tree to
1372 * verify pages have been written, but this is difficult because of
1377 static int vvp_io_read_ahead(const struct lu_env *env,
1378 const struct cl_io_slice *ios,
1379 pgoff_t start, struct cl_read_ahead *ra)
1384 if (ios->cis_io->ci_type == CIT_READ ||
1385 ios->cis_io->ci_type == CIT_FAULT) {
1386 struct vvp_io *vio = cl2vvp_io(env, ios);
1388 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1389 ra->cra_end = CL_PAGE_EOF;
1390 result = +1; /* no need to call down */
1397 static const struct cl_io_operations vvp_io_ops = {
1400 .cio_fini = vvp_io_fini,
1401 .cio_lock = vvp_io_read_lock,
1402 .cio_start = vvp_io_read_start,
1403 .cio_end = vvp_io_rw_end,
1404 .cio_advance = vvp_io_advance,
1407 .cio_fini = vvp_io_fini,
1408 .cio_iter_init = vvp_io_write_iter_init,
1409 .cio_iter_fini = vvp_io_write_iter_fini,
1410 .cio_lock = vvp_io_write_lock,
1411 .cio_start = vvp_io_write_start,
1412 .cio_end = vvp_io_rw_end,
1413 .cio_advance = vvp_io_advance,
1416 .cio_fini = vvp_io_setattr_fini,
1417 .cio_iter_init = vvp_io_setattr_iter_init,
1418 .cio_lock = vvp_io_setattr_lock,
1419 .cio_start = vvp_io_setattr_start,
1420 .cio_end = vvp_io_setattr_end
1423 .cio_fini = vvp_io_fault_fini,
1424 .cio_iter_init = vvp_io_fault_iter_init,
1425 .cio_lock = vvp_io_fault_lock,
1426 .cio_start = vvp_io_fault_start,
1427 .cio_end = vvp_io_fault_end,
1430 .cio_start = vvp_io_fsync_start,
1431 .cio_fini = vvp_io_fini
1434 .cio_fini = vvp_io_fini
1437 .cio_fini = vvp_io_fini
1440 .cio_fini = vvp_io_fini
1443 .cio_read_ahead = vvp_io_read_ahead
1446 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1449 struct vvp_io *vio = vvp_env_io(env);
1450 struct inode *inode = vvp_object_inode(obj);
1453 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1456 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1457 "restore needed %d\n",
1458 PFID(lu_object_fid(&obj->co_lu)),
1459 io->ci_ignore_layout, io->ci_verify_layout,
1460 vio->vui_layout_gen, io->ci_restore_needed);
1462 CL_IO_SLICE_CLEAN(vio, vui_cl);
1463 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1464 vio->vui_ra_valid = false;
1466 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1468 struct ll_inode_info *lli = ll_i2info(inode);
1470 count = io->u.ci_rw.crw_count;
1471 /* "If nbyte is 0, read() will return 0 and have no other
1472 * results." -- Single Unix Spec */
1476 vio->vui_tot_count = count;
1478 /* for read/write, we store the jobid in the inode, and
1479 * it'll be fetched by osc when building RPC.
1481 * it's not accurate if the file is shared by different
1484 lustre_get_jobid(lli->lli_jobid, sizeof(lli->lli_jobid));
1485 } else if (io->ci_type == CIT_SETATTR) {
1486 if (!cl_io_is_trunc(io))
1487 io->ci_lockreq = CILR_MANDATORY;
1490 /* Enqueue layout lock and get layout version. We need to do this
1491 * even for operations requiring to open file, such as read and write,
1492 * because it might not grant layout lock in IT_OPEN. */
1493 if (result == 0 && !io->ci_ignore_layout) {
1494 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1495 if (result == -ENOENT)
1496 /* If the inode on MDS has been removed, but the objects
1497 * on OSTs haven't been destroyed (async unlink), layout
1498 * fetch will return -ENOENT, we'd ingore this error
1499 * and continue with dirty flush. LU-3230. */
1502 CERROR("%s: refresh file layout " DFID " error %d.\n",
1503 ll_get_fsname(inode->i_sb, NULL, 0),
1504 PFID(lu_object_fid(&obj->co_lu)), result);
1507 io->ci_result = result < 0 ? result : 0;