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) {
366 io->ci_need_write_intent = 0;
368 LASSERT(io->ci_type == CIT_WRITE ||
369 cl_io_is_trunc(io) || cl_io_is_mkwrite(io));
371 CDEBUG(D_VFSTRACE, DFID" write layout, type %u "DEXT"\n",
372 PFID(lu_object_fid(&obj->co_lu)), io->ci_type,
373 PEXT(&io->ci_write_intent));
375 rc = ll_layout_write_intent(inode, io->ci_write_intent.e_start,
376 io->ci_write_intent.e_end);
379 io->ci_need_restart = 1;
383 if (!io->ci_need_restart &&
384 !io->ci_ignore_layout && io->ci_verify_layout) {
385 /* check layout version */
386 ll_layout_refresh(inode, &gen);
387 io->ci_need_restart = vio->vui_layout_gen != gen;
388 if (io->ci_need_restart) {
390 DFID" layout changed from %d to %d.\n",
391 PFID(lu_object_fid(&obj->co_lu)),
392 vio->vui_layout_gen, gen);
400 static void vvp_io_fault_fini(const struct lu_env *env,
401 const struct cl_io_slice *ios)
403 struct cl_io *io = ios->cis_io;
404 struct cl_page *page = io->u.ci_fault.ft_page;
406 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
409 lu_ref_del(&page->cp_reference, "fault", io);
410 cl_page_put(env, page);
411 io->u.ci_fault.ft_page = NULL;
413 vvp_io_fini(env, ios);
416 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
419 * we only want to hold PW locks if the mmap() can generate
420 * writes back to the file and that only happens in shared
423 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
428 static int vvp_mmap_locks(const struct lu_env *env, struct cl_io *io)
430 struct vvp_thread_info *vti = vvp_env_info(env);
431 struct mm_struct *mm = current->mm;
432 struct vm_area_struct *vma;
433 struct cl_lock_descr *descr = &vti->vti_descr;
434 union ldlm_policy_data policy;
440 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
442 if (!cl_is_normalio(env, io))
445 /* No MM (e.g. NFS)? No vmas too. */
449 iov_for_each(iov, i, io->u.ci_rw.rw_iter) {
450 unsigned long addr = (unsigned long)iov.iov_base;
451 size_t count = iov.iov_len;
456 count += addr & ~PAGE_MASK;
459 down_read(&mm->mmap_sem);
460 while((vma = our_vma(mm, addr, count)) != NULL) {
461 struct dentry *de = file_dentry(vma->vm_file);
462 struct inode *inode = de->d_inode;
463 int flags = CEF_MUST;
465 if (ll_file_nolock(vma->vm_file)) {
467 * For no lock case is not allowed for mmap
474 * XXX: Required lock mode can be weakened: CIT_WRITE
475 * io only ever reads user level buffer, and CIT_READ
478 policy_from_vma(&policy, vma, addr, count);
479 descr->cld_mode = vvp_mode_from_vma(vma);
480 descr->cld_obj = ll_i2info(inode)->lli_clob;
481 descr->cld_start = cl_index(descr->cld_obj,
482 policy.l_extent.start);
483 descr->cld_end = cl_index(descr->cld_obj,
484 policy.l_extent.end);
485 descr->cld_enq_flags = flags;
486 result = cl_io_lock_alloc_add(env, io, descr);
488 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
489 descr->cld_mode, descr->cld_start,
495 if (vma->vm_end - addr >= count)
498 count -= vma->vm_end - addr;
501 up_read(&mm->mmap_sem);
508 static void vvp_io_advance(const struct lu_env *env,
509 const struct cl_io_slice *ios,
512 struct vvp_io *vio = cl2vvp_io(env, ios);
513 struct cl_io *io = ios->cis_io;
514 struct cl_object *obj = ios->cis_io->ci_obj;
516 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
518 if (!cl_is_normalio(env, io))
521 vio->vui_tot_count -= nob;
523 iov_iter_advance(&io->u.ci_rw.rw_iter, nob);
524 io->u.ci_rw.rw_iocb.ki_pos = io->u.ci_rw.rw_range.cir_pos;
525 #ifdef HAVE_KIOCB_KI_LEFT
526 io->u.ci_rw.rw_iocb.ki_left = vio->vui_tot_count;
527 #elif defined(HAVE_KI_NBYTES)
528 io->u.ci_rw.rw_iocb.ki_nbytes = vio->vui_tot_count;
531 /* It was truncated to stripe size in vvp_io_rw_lock() */
532 iov_iter_reexpand(&io->u.ci_rw.rw_iter, vio->vui_tot_count);
536 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
537 enum cl_lock_mode mode, loff_t start, loff_t end)
542 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
545 if (cl_is_normalio(env, io))
546 iov_iter_truncate(&io->u.ci_rw.rw_iter,
547 io->u.ci_rw.rw_range.cir_count);
549 if (io->u.ci_rw.rw_nonblock)
550 ast_flags |= CEF_NONBLOCK;
551 if (io->ci_lock_no_expand)
552 ast_flags |= CEF_LOCK_NO_EXPAND;
554 result = vvp_mmap_locks(env, io);
556 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
561 static int vvp_io_read_lock(const struct lu_env *env,
562 const struct cl_io_slice *ios)
564 struct cl_io *io = ios->cis_io;
565 struct cl_io_range *range = &io->u.ci_rw.rw_range;
569 rc = vvp_io_rw_lock(env, io, CLM_READ, range->cir_pos,
570 range->cir_pos + range->cir_count - 1);
574 static int vvp_io_fault_lock(const struct lu_env *env,
575 const struct cl_io_slice *ios)
577 struct cl_io *io = ios->cis_io;
578 struct vvp_io *vio = cl2vvp_io(env, ios);
580 * XXX LDLM_FL_CBPENDING
582 return vvp_io_one_lock_index(env,
584 vvp_mode_from_vma(vio->u.fault.ft_vma),
585 io->u.ci_fault.ft_index,
586 io->u.ci_fault.ft_index);
589 static int vvp_io_write_lock(const struct lu_env *env,
590 const struct cl_io_slice *ios)
592 struct cl_io *io = ios->cis_io;
598 if (io->u.ci_rw.rw_append) {
600 end = OBD_OBJECT_EOF;
602 start = io->u.ci_rw.rw_range.cir_pos;
603 end = start + io->u.ci_rw.rw_range.cir_count - 1;
605 rc = vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
609 static int vvp_io_setattr_iter_init(const struct lu_env *env,
610 const struct cl_io_slice *ios)
616 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
618 * Handles "lockless io" mode when extent locking is done by server.
620 static int vvp_io_setattr_lock(const struct lu_env *env,
621 const struct cl_io_slice *ios)
623 struct cl_io *io = ios->cis_io;
627 if (cl_io_is_trunc(io)) {
628 new_size = io->u.ci_setattr.sa_attr.lvb_size;
630 enqflags = CEF_DISCARD_DATA;
632 unsigned int valid = io->u.ci_setattr.sa_valid;
634 if (!(valid & TIMES_SET_FLAGS))
637 if ((!(valid & ATTR_MTIME) ||
638 io->u.ci_setattr.sa_attr.lvb_mtime >=
639 io->u.ci_setattr.sa_attr.lvb_ctime) &&
640 (!(valid & ATTR_ATIME) ||
641 io->u.ci_setattr.sa_attr.lvb_atime >=
642 io->u.ci_setattr.sa_attr.lvb_ctime))
648 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
649 new_size, OBD_OBJECT_EOF);
652 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
657 * Only ll_inode_size_lock is taken at this level.
659 ll_inode_size_lock(inode);
660 result = inode_newsize_ok(inode, size);
662 ll_inode_size_unlock(inode);
665 i_size_write(inode, size);
667 ll_truncate_pagecache(inode, size);
668 ll_inode_size_unlock(inode);
672 static int vvp_io_setattr_time(const struct lu_env *env,
673 const struct cl_io_slice *ios)
675 struct cl_io *io = ios->cis_io;
676 struct cl_object *obj = io->ci_obj;
677 struct cl_attr *attr = vvp_env_thread_attr(env);
679 unsigned valid = CAT_CTIME;
681 cl_object_attr_lock(obj);
682 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
683 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
684 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
687 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
688 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
691 result = cl_object_attr_update(env, obj, attr, valid);
692 cl_object_attr_unlock(obj);
697 static int vvp_io_setattr_start(const struct lu_env *env,
698 const struct cl_io_slice *ios)
700 struct cl_io *io = ios->cis_io;
701 struct inode *inode = vvp_object_inode(io->ci_obj);
702 struct ll_inode_info *lli = ll_i2info(inode);
704 if (cl_io_is_trunc(io)) {
705 down_write(&lli->lli_trunc_sem);
707 inode_dio_wait(inode);
712 if (io->u.ci_setattr.sa_valid & TIMES_SET_FLAGS)
713 return vvp_io_setattr_time(env, ios);
718 static void vvp_io_setattr_end(const struct lu_env *env,
719 const struct cl_io_slice *ios)
721 struct cl_io *io = ios->cis_io;
722 struct inode *inode = vvp_object_inode(io->ci_obj);
723 struct ll_inode_info *lli = ll_i2info(inode);
725 if (cl_io_is_trunc(io)) {
726 /* Truncate in memory pages - they must be clean pages
727 * because osc has already notified to destroy osc_extents. */
728 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
729 inode_dio_write_done(inode);
731 up_write(&lli->lli_trunc_sem);
737 static void vvp_io_setattr_fini(const struct lu_env *env,
738 const struct cl_io_slice *ios)
740 bool restore_needed = ios->cis_io->ci_restore_needed;
741 struct inode *inode = vvp_object_inode(ios->cis_obj);
743 vvp_io_fini(env, ios);
745 if (restore_needed && !ios->cis_io->ci_restore_needed) {
746 /* restore finished, set data modified flag for HSM */
747 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
751 static int vvp_io_read_start(const struct lu_env *env,
752 const struct cl_io_slice *ios)
754 struct vvp_io *vio = cl2vvp_io(env, ios);
755 struct cl_io *io = ios->cis_io;
756 struct cl_object *obj = io->ci_obj;
757 struct inode *inode = vvp_object_inode(obj);
758 struct ll_inode_info *lli = ll_i2info(inode);
759 struct file *file = vio->vui_fd->fd_file;
760 struct cl_io_range *range = &io->u.ci_rw.rw_range;
761 loff_t pos = range->cir_pos; /* for generic_file_splice_read() only */
762 size_t tot = vio->vui_tot_count;
767 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
769 CDEBUG(D_VFSTRACE, "%s: read [%llu, %llu)\n",
770 file_dentry(file)->d_name.name,
771 range->cir_pos, range->cir_pos + range->cir_count);
773 if (vio->vui_io_subtype == IO_NORMAL)
774 down_read(&lli->lli_trunc_sem);
776 if (!can_populate_pages(env, io, inode))
779 /* Unless this is reading a sparse file, otherwise the lock has already
780 * been acquired so vvp_prep_size() is an empty op. */
781 result = vvp_prep_size(env, obj, io, range->cir_pos, range->cir_count,
785 else if (exceed != 0)
788 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
789 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
790 inode->i_ino, range->cir_count, range->cir_pos,
793 /* turn off the kernel's read-ahead */
794 vio->vui_fd->fd_file->f_ra.ra_pages = 0;
796 /* initialize read-ahead window once per syscall */
797 if (!vio->vui_ra_valid) {
798 vio->vui_ra_valid = true;
799 vio->vui_ra_start = cl_index(obj, range->cir_pos);
800 vio->vui_ra_count = cl_index(obj, tot + PAGE_SIZE - 1);
806 switch (vio->vui_io_subtype) {
808 LASSERTF(io->u.ci_rw.rw_iocb.ki_pos == range->cir_pos,
809 "ki_pos %lld [%lld, %lld)\n",
810 io->u.ci_rw.rw_iocb.ki_pos,
811 range->cir_pos, range->cir_pos + range->cir_count);
812 result = generic_file_read_iter(&io->u.ci_rw.rw_iocb,
813 &io->u.ci_rw.rw_iter);
816 result = generic_file_splice_read(file, &pos,
817 vio->u.splice.vui_pipe,
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);
833 if (result < range->cir_count)
835 io->ci_nob += result;
836 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, vio->vui_fd,
837 range->cir_pos, result, READ);
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 lu_ref_del(&page->cp_reference, "cl_io", io);
894 cl_page_put(env, page);
898 RETURN(bytes > 0 ? bytes : rc);
901 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
902 struct cl_page *page)
904 struct page *vmpage = page->cp_vmpage;
906 SetPageUptodate(vmpage);
907 set_page_dirty(vmpage);
909 cl_page_disown(env, io, page);
911 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
912 cl_page_put(env, page);
915 /* make sure the page list is contiguous */
916 static bool page_list_sanity_check(struct cl_object *obj,
917 struct cl_page_list *plist)
919 struct cl_page *page;
920 pgoff_t index = CL_PAGE_EOF;
922 cl_page_list_for_each(page, plist) {
923 struct vvp_page *vpg = cl_object_page_slice(obj, page);
925 if (index == CL_PAGE_EOF) {
926 index = vvp_index(vpg);
931 if (index == vvp_index(vpg))
939 /* Return how many bytes have queued or written */
940 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
942 struct cl_object *obj = io->ci_obj;
943 struct inode *inode = vvp_object_inode(obj);
944 struct vvp_io *vio = vvp_env_io(env);
945 struct cl_page_list *queue = &vio->u.write.vui_queue;
946 struct cl_page *page;
949 unsigned int npages = vio->u.write.vui_queue.pl_nr;
955 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
956 npages, vio->u.write.vui_from, vio->u.write.vui_to);
958 LASSERT(page_list_sanity_check(obj, queue));
960 /* submit IO with async write */
961 rc = cl_io_commit_async(env, io, queue,
962 vio->u.write.vui_from, vio->u.write.vui_to,
963 write_commit_callback);
964 npages -= queue->pl_nr; /* already committed pages */
966 /* calculate how many bytes were written */
967 bytes = npages << PAGE_SHIFT;
970 bytes -= vio->u.write.vui_from;
971 if (queue->pl_nr == 0) /* last page */
972 bytes -= PAGE_SIZE - vio->u.write.vui_to;
973 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
975 vio->u.write.vui_written += bytes;
977 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
978 npages, bytes, vio->u.write.vui_written);
980 /* the first page must have been written. */
981 vio->u.write.vui_from = 0;
983 LASSERT(page_list_sanity_check(obj, queue));
984 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
986 /* out of quota, try sync write */
987 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
988 rc = vvp_io_commit_sync(env, io, queue,
989 vio->u.write.vui_from,
990 vio->u.write.vui_to);
992 vio->u.write.vui_written += rc;
997 /* update inode size */
998 ll_merge_attr(env, inode);
1000 /* Now the pages in queue were failed to commit, discard them
1001 * unless they were dirtied before. */
1002 while (queue->pl_nr > 0) {
1003 page = cl_page_list_first(queue);
1004 cl_page_list_del(env, queue, page);
1006 if (!PageDirty(cl_page_vmpage(page)))
1007 cl_page_discard(env, io, page);
1009 cl_page_disown(env, io, page);
1011 lu_ref_del(&page->cp_reference, "cl_io", io);
1012 cl_page_put(env, page);
1014 cl_page_list_fini(env, queue);
1019 static int vvp_io_write_start(const struct lu_env *env,
1020 const struct cl_io_slice *ios)
1022 struct vvp_io *vio = cl2vvp_io(env, ios);
1023 struct cl_io *io = ios->cis_io;
1024 struct cl_object *obj = io->ci_obj;
1025 struct inode *inode = vvp_object_inode(obj);
1026 struct ll_inode_info *lli = ll_i2info(inode);
1027 struct file *file = vio->vui_fd->fd_file;
1028 struct cl_io_range *range = &io->u.ci_rw.rw_range;
1029 bool lock_inode = !lli->lli_inode_locked &&
1034 if (vio->vui_io_subtype == IO_NORMAL)
1035 down_read(&lli->lli_trunc_sem);
1037 if (!can_populate_pages(env, io, inode))
1040 if (cl_io_is_append(io)) {
1042 * PARALLEL IO This has to be changed for parallel IO doing
1043 * out-of-order writes.
1045 ll_merge_attr(env, inode);
1046 range->cir_pos = i_size_read(inode);
1047 io->u.ci_rw.rw_iocb.ki_pos = range->cir_pos;
1049 LASSERTF(io->u.ci_rw.rw_iocb.ki_pos == range->cir_pos,
1050 "ki_pos %lld [%lld, %lld)\n",
1051 io->u.ci_rw.rw_iocb.ki_pos,
1052 range->cir_pos, range->cir_pos + range->cir_count);
1055 CDEBUG(D_VFSTRACE, "%s: write [%llu, %llu)\n",
1056 file_dentry(file)->d_name.name,
1057 range->cir_pos, range->cir_pos + range->cir_count);
1059 /* The maximum Lustre file size is variable, based on the OST maximum
1060 * object size and number of stripes. This needs another check in
1061 * addition to the VFS checks earlier. */
1062 if (range->cir_pos + range->cir_count > ll_file_maxbytes(inode)) {
1064 "%s: file %s ("DFID") offset %llu > maxbytes %llu\n",
1065 ll_get_fsname(inode->i_sb, NULL, 0),
1066 file_dentry(file)->d_name.name,
1067 PFID(ll_inode2fid(inode)),
1068 range->cir_pos + range->cir_count,
1069 ll_file_maxbytes(inode));
1073 /* Tests to verify we take the i_mutex correctly */
1074 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_SEC) && !lock_inode)
1077 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_NOSEC) && lock_inode)
1081 * When using the locked AIO function (generic_file_aio_write())
1082 * testing has shown the inode mutex to be a limiting factor
1083 * with multi-threaded single shared file performance. To get
1084 * around this, we now use the lockless version. To maintain
1085 * consistency, proper locking to protect against writes,
1086 * trucates, etc. is handled in the higher layers of lustre.
1090 result = __generic_file_write_iter(&io->u.ci_rw.rw_iocb,
1091 &io->u.ci_rw.rw_iter);
1093 inode_unlock(inode);
1095 if (result > 0 || result == -EIOCBQUEUED)
1096 #ifdef HAVE_GENERIC_WRITE_SYNC_2ARGS
1097 result = generic_write_sync(&io->u.ci_rw.rw_iocb, result);
1102 err = generic_write_sync(io->u.ci_rw.rw_iocb.ki_filp,
1103 range->cir_pos, result);
1104 if (err < 0 && result > 0)
1110 result = vvp_io_write_commit(env, io);
1111 if (vio->u.write.vui_written > 0) {
1112 result = vio->u.write.vui_written;
1113 CDEBUG(D_VFSTRACE, "%s: write nob %zd, result: %zd\n",
1114 file_dentry(file)->d_name.name,
1115 io->ci_nob, result);
1116 io->ci_nob += result;
1120 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
1122 if (result < range->cir_count)
1123 io->ci_continue = 0;
1124 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
1125 vio->vui_fd, range->cir_pos, result, WRITE);
1132 static void vvp_io_rw_end(const struct lu_env *env,
1133 const struct cl_io_slice *ios)
1135 struct vvp_io *vio = cl2vvp_io(env, ios);
1136 struct inode *inode = vvp_object_inode(ios->cis_obj);
1137 struct ll_inode_info *lli = ll_i2info(inode);
1139 if (vio->vui_io_subtype == IO_NORMAL)
1140 up_read(&lli->lli_trunc_sem);
1143 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
1145 struct vm_fault *vmf = cfio->ft_vmf;
1147 cfio->ft_flags = ll_filemap_fault(cfio->ft_vma, vmf);
1148 cfio->ft_flags_valid = 1;
1151 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
1152 get_vmf_address(vmf));
1153 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
1154 lock_page(vmf->page);
1155 cfio->ft_flags |= VM_FAULT_LOCKED;
1158 cfio->ft_vmpage = vmf->page;
1163 if (cfio->ft_flags & VM_FAULT_SIGBUS) {
1164 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", get_vmf_address(vmf));
1168 if (cfio->ft_flags & VM_FAULT_OOM) {
1169 CDEBUG(D_PAGE, "got addr %p - OOM\n", get_vmf_address(vmf));
1173 if (cfio->ft_flags & VM_FAULT_RETRY)
1176 CERROR("unknown error in page fault %d\n", cfio->ft_flags);
1181 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1182 struct cl_page *page)
1184 set_page_dirty(page->cp_vmpage);
1187 static int vvp_io_fault_start(const struct lu_env *env,
1188 const struct cl_io_slice *ios)
1190 struct vvp_io *vio = cl2vvp_io(env, ios);
1191 struct cl_io *io = ios->cis_io;
1192 struct cl_object *obj = io->ci_obj;
1193 struct inode *inode = vvp_object_inode(obj);
1194 struct ll_inode_info *lli = ll_i2info(inode);
1195 struct cl_fault_io *fio = &io->u.ci_fault;
1196 struct vvp_fault_io *cfio = &vio->u.fault;
1199 struct page *vmpage = NULL;
1200 struct cl_page *page;
1205 down_read(&lli->lli_trunc_sem);
1207 /* offset of the last byte on the page */
1208 offset = cl_offset(obj, fio->ft_index + 1) - 1;
1209 LASSERT(cl_index(obj, offset) == fio->ft_index);
1210 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1214 /* must return locked page */
1215 if (fio->ft_mkwrite) {
1216 LASSERT(cfio->ft_vmpage != NULL);
1217 lock_page(cfio->ft_vmpage);
1219 result = vvp_io_kernel_fault(cfio);
1224 vmpage = cfio->ft_vmpage;
1225 LASSERT(PageLocked(vmpage));
1227 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1228 ll_invalidate_page(vmpage);
1230 size = i_size_read(inode);
1231 /* Though we have already held a cl_lock upon this page, but
1232 * it still can be truncated locally. */
1233 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1234 (page_offset(vmpage) > size))) {
1235 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1237 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1239 GOTO(out, result = +1);
1242 last_index = cl_index(obj, size - 1);
1244 if (fio->ft_mkwrite ) {
1246 * Capture the size while holding the lli_trunc_sem from above
1247 * we want to make sure that we complete the mkwrite action
1248 * while holding this lock. We need to make sure that we are
1249 * not past the end of the file.
1251 if (last_index < fio->ft_index) {
1253 "llite: mkwrite and truncate race happened: "
1254 "%p: 0x%lx 0x%lx\n",
1255 vmpage->mapping,fio->ft_index,last_index);
1257 * We need to return if we are
1258 * passed the end of the file. This will propagate
1259 * up the call stack to ll_page_mkwrite where
1260 * we will return VM_FAULT_NOPAGE. Any non-negative
1261 * value returned here will be silently
1262 * converted to 0. If the vmpage->mapping is null
1263 * the error code would be converted back to ENODATA
1264 * in ll_page_mkwrite0. Thus we return -ENODATA
1265 * to handle both cases
1267 GOTO(out, result = -ENODATA);
1271 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1273 GOTO(out, result = PTR_ERR(page));
1275 /* if page is going to be written, we should add this page into cache
1277 if (fio->ft_mkwrite) {
1278 wait_on_page_writeback(vmpage);
1279 if (!PageDirty(vmpage)) {
1280 struct cl_page_list *plist = &io->ci_queue.c2_qin;
1281 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1284 /* vvp_page_assume() calls wait_on_page_writeback(). */
1285 cl_page_assume(env, io, page);
1287 cl_page_list_init(plist);
1288 cl_page_list_add(plist, page);
1291 if (last_index == vvp_index(vpg))
1292 to = size & ~PAGE_MASK;
1294 /* Do not set Dirty bit here so that in case IO is
1295 * started before the page is really made dirty, we
1296 * still have chance to detect it. */
1297 result = cl_io_commit_async(env, io, plist, 0, to,
1298 mkwrite_commit_callback);
1299 LASSERT(cl_page_is_owned(page, io));
1300 cl_page_list_fini(env, plist);
1304 cl_page_discard(env, io, page);
1305 cl_page_disown(env, io, page);
1307 cl_page_put(env, page);
1309 /* we're in big trouble, what can we do now? */
1310 if (result == -EDQUOT)
1314 cl_page_disown(env, io, page);
1319 * The ft_index is only used in the case of
1320 * a mkwrite action. We need to check
1321 * our assertions are correct, since
1322 * we should have caught this above
1324 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1325 if (fio->ft_index == last_index)
1327 * Last page is mapped partially.
1329 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1331 fio->ft_nob = cl_page_size(obj);
1333 lu_ref_add(&page->cp_reference, "fault", io);
1334 fio->ft_page = page;
1338 /* return unlocked vmpage to avoid deadlocking */
1340 unlock_page(vmpage);
1342 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1347 static void vvp_io_fault_end(const struct lu_env *env,
1348 const struct cl_io_slice *ios)
1350 struct inode *inode = vvp_object_inode(ios->cis_obj);
1351 struct ll_inode_info *lli = ll_i2info(inode);
1353 CLOBINVRNT(env, ios->cis_io->ci_obj,
1354 vvp_object_invariant(ios->cis_io->ci_obj));
1355 up_read(&lli->lli_trunc_sem);
1358 static int vvp_io_fsync_start(const struct lu_env *env,
1359 const struct cl_io_slice *ios)
1361 /* we should mark TOWRITE bit to each dirty page in radix tree to
1362 * verify pages have been written, but this is difficult because of
1367 static int vvp_io_read_ahead(const struct lu_env *env,
1368 const struct cl_io_slice *ios,
1369 pgoff_t start, struct cl_read_ahead *ra)
1374 if (ios->cis_io->ci_type == CIT_READ ||
1375 ios->cis_io->ci_type == CIT_FAULT) {
1376 struct vvp_io *vio = cl2vvp_io(env, ios);
1378 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1379 ra->cra_end = CL_PAGE_EOF;
1380 result = +1; /* no need to call down */
1387 static const struct cl_io_operations vvp_io_ops = {
1390 .cio_fini = vvp_io_fini,
1391 .cio_lock = vvp_io_read_lock,
1392 .cio_start = vvp_io_read_start,
1393 .cio_end = vvp_io_rw_end,
1394 .cio_advance = vvp_io_advance,
1397 .cio_fini = vvp_io_fini,
1398 .cio_iter_init = vvp_io_write_iter_init,
1399 .cio_iter_fini = vvp_io_write_iter_fini,
1400 .cio_lock = vvp_io_write_lock,
1401 .cio_start = vvp_io_write_start,
1402 .cio_end = vvp_io_rw_end,
1403 .cio_advance = vvp_io_advance,
1406 .cio_fini = vvp_io_setattr_fini,
1407 .cio_iter_init = vvp_io_setattr_iter_init,
1408 .cio_lock = vvp_io_setattr_lock,
1409 .cio_start = vvp_io_setattr_start,
1410 .cio_end = vvp_io_setattr_end
1413 .cio_fini = vvp_io_fault_fini,
1414 .cio_iter_init = vvp_io_fault_iter_init,
1415 .cio_lock = vvp_io_fault_lock,
1416 .cio_start = vvp_io_fault_start,
1417 .cio_end = vvp_io_fault_end,
1420 .cio_start = vvp_io_fsync_start,
1421 .cio_fini = vvp_io_fini
1424 .cio_fini = vvp_io_fini
1427 .cio_fini = vvp_io_fini
1430 .cio_fini = vvp_io_fini
1433 .cio_read_ahead = vvp_io_read_ahead
1436 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1439 struct vvp_io *vio = vvp_env_io(env);
1440 struct inode *inode = vvp_object_inode(obj);
1443 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1446 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1447 "restore needed %d\n",
1448 PFID(lu_object_fid(&obj->co_lu)),
1449 io->ci_ignore_layout, io->ci_verify_layout,
1450 vio->vui_layout_gen, io->ci_restore_needed);
1452 CL_IO_SLICE_CLEAN(vio, vui_cl);
1453 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1454 vio->vui_ra_valid = false;
1456 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1457 struct ll_inode_info *lli = ll_i2info(inode);
1459 vio->vui_tot_count = io->u.ci_rw.rw_range.cir_count;
1460 /* "If nbyte is 0, read() will return 0 and have no other
1461 * results." -- Single Unix Spec */
1462 if (vio->vui_tot_count == 0)
1465 /* for read/write, we store the jobid in the inode, and
1466 * it'll be fetched by osc when building RPC.
1468 * it's not accurate if the file is shared by different
1471 lustre_get_jobid(lli->lli_jobid);
1472 } else if (io->ci_type == CIT_SETATTR) {
1473 if (!cl_io_is_trunc(io))
1474 io->ci_lockreq = CILR_MANDATORY;
1477 /* Enqueue layout lock and get layout version. We need to do this
1478 * even for operations requiring to open file, such as read and write,
1479 * because it might not grant layout lock in IT_OPEN. */
1480 if (result == 0 && !io->ci_ignore_layout) {
1481 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1482 if (result == -ENOENT)
1483 /* If the inode on MDS has been removed, but the objects
1484 * on OSTs haven't been destroyed (async unlink), layout
1485 * fetch will return -ENOENT, we'd ingore this error
1486 * and continue with dirty flush. LU-3230. */
1489 CERROR("%s: refresh file layout " DFID " error %d.\n",
1490 ll_get_fsname(inode->i_sb, NULL, 0),
1491 PFID(lu_object_fid(&obj->co_lu)), result);
1494 io->ci_result = result < 0 ? result : 0;