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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Implementation of cl_io for VVP layer.
38 * Author: Nikita Danilov <nikita.danilov@sun.com>
39 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
42 #define DEBUG_SUBSYSTEM S_LLITE
46 #include "llite_internal.h"
47 #include "vvp_internal.h"
49 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
50 const struct cl_io_slice *slice)
54 vio = container_of(slice, struct vvp_io, vui_cl);
55 LASSERT(vio == vvp_env_io(env));
61 * True, if \a io is a normal io, False for splice_{read,write}
63 static int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
65 struct vvp_io *vio = vvp_env_io(env);
67 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
69 return vio->vui_io_subtype == IO_NORMAL;
73 * For swapping layout. The file's layout may have changed.
74 * To avoid populating pages to a wrong stripe, we have to verify the
75 * correctness of layout. It works because swapping layout processes
76 * have to acquire group lock.
78 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
81 struct ll_inode_info *lli = ll_i2info(inode);
82 struct vvp_io *vio = vvp_env_io(env);
85 switch (io->ci_type) {
88 /* don't need lock here to check lli_layout_gen as we have held
89 * extent lock and GROUP lock has to hold to swap layout */
90 if (ll_layout_version_get(lli) != vio->vui_layout_gen ||
91 OBD_FAIL_CHECK_RESET(OBD_FAIL_LLITE_LOST_LAYOUT, 0)) {
92 io->ci_need_restart = 1;
93 /* this will cause a short read/write */
98 /* fault is okay because we've already had a page. */
106 static void vvp_object_size_lock(struct cl_object *obj)
108 struct inode *inode = vvp_object_inode(obj);
110 ll_inode_size_lock(inode);
111 cl_object_attr_lock(obj);
114 static void vvp_object_size_unlock(struct cl_object *obj)
116 struct inode *inode = vvp_object_inode(obj);
118 cl_object_attr_unlock(obj);
119 ll_inode_size_unlock(inode);
123 * Helper function that if necessary adjusts file size (inode->i_size), when
124 * position at the offset \a pos is accessed. File size can be arbitrary stale
125 * on a Lustre client, but client at least knows KMS. If accessed area is
126 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
128 * Locking: i_size_lock is used to serialize changes to inode size and to
129 * protect consistency between inode size and cl_object
130 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
131 * top-object and sub-objects.
133 static int vvp_prep_size(const struct lu_env *env, struct cl_object *obj,
134 struct cl_io *io, loff_t start, size_t count,
137 struct cl_attr *attr = vvp_env_thread_attr(env);
138 struct inode *inode = vvp_object_inode(obj);
139 loff_t pos = start + count - 1;
144 * Consistency guarantees: following possibilities exist for the
145 * relation between region being accessed and real file size at this
148 * (A): the region is completely inside of the file;
150 * (B-x): x bytes of region are inside of the file, the rest is
153 * (C): the region is completely outside of the file.
155 * This classification is stable under DLM lock already acquired by
156 * the caller, because to change the class, other client has to take
157 * DLM lock conflicting with our lock. Also, any updates to ->i_size
158 * by other threads on this client are serialized by
159 * ll_inode_size_lock(). This guarantees that short reads are handled
160 * correctly in the face of concurrent writes and truncates.
162 vvp_object_size_lock(obj);
163 result = cl_object_attr_get(env, obj, attr);
168 * A glimpse is necessary to determine whether we
169 * return a short read (B) or some zeroes at the end
172 vvp_object_size_unlock(obj);
173 result = cl_glimpse_lock(env, io, inode, obj, 0);
174 if (result == 0 && exceed != NULL) {
175 /* If objective page index exceed end-of-file
176 * page index, return directly. Do not expect
177 * kernel will check such case correctly.
178 * linux-2.6.18-128.1.1 miss to do that.
180 loff_t size = i_size_read(inode);
181 unsigned long cur_index = start >>
184 if ((size == 0 && cur_index != 0) ||
185 (((size - 1) >> PAGE_CACHE_SHIFT) <
193 * region is within kms and, hence, within real file
194 * size (A). We need to increase i_size to cover the
195 * read region so that generic_file_read() will do its
196 * job, but that doesn't mean the kms size is
197 * _correct_, it is only the _minimum_ size. If
198 * someone does a stat they will get the correct size
199 * which will always be >= the kms value here.
202 if (i_size_read(inode) < kms) {
203 i_size_write(inode, kms);
205 DFID" updating i_size "LPU64"\n",
206 PFID(lu_object_fid(&obj->co_lu)),
207 (__u64)i_size_read(inode));
212 vvp_object_size_unlock(obj);
217 /*****************************************************************************
223 static int vvp_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
224 __u32 enqflags, enum cl_lock_mode mode,
225 pgoff_t start, pgoff_t end)
227 struct vvp_io *vio = vvp_env_io(env);
228 struct cl_lock_descr *descr = &vio->vui_link.cill_descr;
229 struct cl_object *obj = io->ci_obj;
231 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
234 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
236 memset(&vio->vui_link, 0, sizeof vio->vui_link);
238 if (vio->vui_fd && (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
239 descr->cld_mode = CLM_GROUP;
240 descr->cld_gid = vio->vui_fd->fd_grouplock.lg_gid;
242 descr->cld_mode = mode;
245 descr->cld_obj = obj;
246 descr->cld_start = start;
247 descr->cld_end = end;
248 descr->cld_enq_flags = enqflags;
250 cl_io_lock_add(env, io, &vio->vui_link);
255 static int vvp_io_one_lock(const struct lu_env *env, struct cl_io *io,
256 __u32 enqflags, enum cl_lock_mode mode,
257 loff_t start, loff_t end)
259 struct cl_object *obj = io->ci_obj;
261 return vvp_io_one_lock_index(env, io, enqflags, mode,
262 cl_index(obj, start), cl_index(obj, end));
265 static int vvp_io_write_iter_init(const struct lu_env *env,
266 const struct cl_io_slice *ios)
268 struct vvp_io *vio = cl2vvp_io(env, ios);
270 cl_page_list_init(&vio->u.write.vui_queue);
271 vio->u.write.vui_written = 0;
272 vio->u.write.vui_from = 0;
273 vio->u.write.vui_to = PAGE_SIZE;
278 static void vvp_io_write_iter_fini(const struct lu_env *env,
279 const struct cl_io_slice *ios)
281 struct vvp_io *vio = cl2vvp_io(env, ios);
283 LASSERT(vio->u.write.vui_queue.pl_nr == 0);
286 static int vvp_io_fault_iter_init(const struct lu_env *env,
287 const struct cl_io_slice *ios)
289 struct vvp_io *vio = cl2vvp_io(env, ios);
290 struct inode *inode = vvp_object_inode(ios->cis_obj);
292 LASSERT(inode == vio->vui_fd->fd_file->f_path.dentry->d_inode);
293 vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime);
298 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
300 struct cl_io *io = ios->cis_io;
301 struct cl_object *obj = io->ci_obj;
302 struct vvp_io *vio = cl2vvp_io(env, ios);
303 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 "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_restore_needed);
313 if (io->ci_restore_needed) {
316 /* file was detected release, we need to restore it
317 * before finishing the io
319 rc = ll_layout_restore(inode, 0, OBD_OBJECT_EOF);
320 /* if restore registration failed, no restart,
321 * we will return -ENODATA */
322 /* The layout will change after restore, so we need to
323 * block on layout lock held by the MDT
324 * as MDT will not send new layout in lvb (see LU-3124)
325 * we have to explicitly fetch it, all this will be done
326 * by ll_layout_refresh()
329 io->ci_restore_needed = 0;
330 io->ci_need_restart = 1;
331 io->ci_verify_layout = 1;
333 io->ci_restore_needed = 1;
334 io->ci_need_restart = 0;
335 io->ci_verify_layout = 0;
340 if (!io->ci_ignore_layout && io->ci_verify_layout) {
343 /* check layout version */
344 ll_layout_refresh(inode, &gen);
345 io->ci_need_restart = vio->vui_layout_gen != gen;
346 if (io->ci_need_restart) {
348 DFID" layout changed from %d to %d.\n",
349 PFID(lu_object_fid(&obj->co_lu)),
350 vio->vui_layout_gen, gen);
351 /* today successful restore is the only possible
353 /* restore was done, clear restoring state */
354 ll_file_clear_flag(ll_i2info(vvp_object_inode(obj)),
355 LLIF_FILE_RESTORING);
360 static void vvp_io_fault_fini(const struct lu_env *env,
361 const struct cl_io_slice *ios)
363 struct cl_io *io = ios->cis_io;
364 struct cl_page *page = io->u.ci_fault.ft_page;
366 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
369 lu_ref_del(&page->cp_reference, "fault", io);
370 cl_page_put(env, page);
371 io->u.ci_fault.ft_page = NULL;
373 vvp_io_fini(env, ios);
376 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
379 * we only want to hold PW locks if the mmap() can generate
380 * writes back to the file and that only happens in shared
383 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
388 static int vvp_mmap_locks(const struct lu_env *env,
389 struct vvp_io *vio, struct cl_io *io)
391 struct vvp_thread_info *vti = vvp_env_info(env);
392 struct mm_struct *mm = current->mm;
393 struct vm_area_struct *vma;
394 struct cl_lock_descr *descr = &vti->vti_descr;
395 union ldlm_policy_data policy;
396 #ifdef HAVE_FILE_OPERATIONS_READ_WRITE_ITER
405 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
407 if (!cl_is_normalio(env, io))
410 /* nfs or loop back device write */
411 if (vio->vui_iter == NULL)
414 /* No MM (e.g. NFS)? No vmas too. */
418 #ifdef HAVE_FILE_OPERATIONS_READ_WRITE_ITER
419 iov_for_each(iov, i, *(vio->vui_iter)) {
420 #else /* !HAVE_FILE_OPERATIONS_READ_WRITE_ITER */
421 for (seg = 0; seg < vio->vui_iter->nr_segs; seg++) {
422 const struct iovec iov = vio->vui_iter->iov[seg];
423 #endif /* HAVE_FILE_OPERATIONS_READ_WRITE_ITER */
424 unsigned long addr = (unsigned long)iov.iov_base;
425 size_t count = iov.iov_len;
430 count += addr & ~PAGE_MASK;
433 down_read(&mm->mmap_sem);
434 while((vma = our_vma(mm, addr, count)) != NULL) {
435 struct dentry *de = vma->vm_file->f_path.dentry;
436 struct inode *inode = de->d_inode;
437 int flags = CEF_MUST;
439 if (ll_file_nolock(vma->vm_file)) {
441 * For no lock case is not allowed for mmap
448 * XXX: Required lock mode can be weakened: CIT_WRITE
449 * io only ever reads user level buffer, and CIT_READ
452 policy_from_vma(&policy, vma, addr, count);
453 descr->cld_mode = vvp_mode_from_vma(vma);
454 descr->cld_obj = ll_i2info(inode)->lli_clob;
455 descr->cld_start = cl_index(descr->cld_obj,
456 policy.l_extent.start);
457 descr->cld_end = cl_index(descr->cld_obj,
458 policy.l_extent.end);
459 descr->cld_enq_flags = flags;
460 result = cl_io_lock_alloc_add(env, io, descr);
462 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
463 descr->cld_mode, descr->cld_start,
469 if (vma->vm_end - addr >= count)
472 count -= vma->vm_end - addr;
475 up_read(&mm->mmap_sem);
482 static void vvp_io_advance(const struct lu_env *env,
483 const struct cl_io_slice *ios,
486 struct vvp_io *vio = cl2vvp_io(env, ios);
487 struct cl_io *io = ios->cis_io;
488 struct cl_object *obj = ios->cis_io->ci_obj;
489 #ifndef HAVE_FILE_OPERATIONS_READ_WRITE_ITER
492 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
494 if (!cl_is_normalio(env, io))
497 #ifdef HAVE_FILE_OPERATIONS_READ_WRITE_ITER
498 vio->vui_tot_count -= nob;
499 iov_iter_reexpand(vio->vui_iter, vio->vui_tot_count);
500 #else /* !HAVE_FILE_OPERATIONS_READ_WRITE_ITER */
501 LASSERT(vio->vui_tot_nrsegs >= vio->vui_iter->nr_segs);
502 LASSERT(vio->vui_tot_count >= nob);
504 /* Restore the iov changed in vvp_io_update_iov() */
505 if (vio->vui_iov_olen > 0) {
506 unsigned long idx = vio->vui_iter->nr_segs - 1;
508 /* In the latest kernels iov is const so that
509 * changes are done using iter helpers. In older
510 * kernels those helpers don't exist we lustre
511 * has to do some of the management of the iter
513 iov = (struct iovec *)&vio->vui_iter->iov[idx];
514 iov->iov_len = vio->vui_iov_olen;
515 vio->vui_iov_olen = 0;
518 /* In the latest kernels special helpers exist to help
519 * advance the iov but we don't have that in older kernels
520 * so we need to do the book keeping ourselves. */
521 iov = (struct iovec *)vio->vui_iter->iov;
523 if (iov->iov_len > nob) {
525 iov->iov_base += nob;
531 vio->vui_tot_nrsegs--;
534 vio->vui_iter->iov = iov;
535 vio->vui_tot_count -= nob;
536 #endif /* HAVE_FILE_OPERATIONS_READ_WRITE_ITER */
539 static void vvp_io_update_iov(const struct lu_env *env,
540 struct vvp_io *vio, struct cl_io *io)
542 size_t size = io->u.ci_rw.crw_count;
543 #ifdef HAVE_FILE_OPERATIONS_READ_WRITE_ITER
544 if (!cl_is_normalio(env, io) || vio->vui_iter == NULL)
547 iov_iter_truncate(vio->vui_iter, size);
548 #else /* !HAVE_FILE_OPERATIONS_READ_WRITE_ITER */
551 vio->vui_iov_olen = 0;
552 if (!cl_is_normalio(env, io) || vio->vui_tot_nrsegs == 0)
555 for (i = 0; i < vio->vui_tot_nrsegs; i++) {
556 struct iovec *iv = (struct iovec *) &vio->vui_iter->iov[i];
558 if (iv->iov_len < size) {
561 if (iv->iov_len > size) {
562 vio->vui_iov_olen = iv->iov_len;
569 vio->vui_iter->nr_segs = i + 1;
570 LASSERTF(vio->vui_tot_nrsegs >= vio->vui_iter->nr_segs,
571 "tot_nrsegs: %lu, nrsegs: %lu\n",
572 vio->vui_tot_nrsegs, vio->vui_iter->nr_segs);
573 #endif /* !HAVE_FILE_OPERATIONS_READ_WRITE_ITER */
576 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
577 enum cl_lock_mode mode, loff_t start, loff_t end)
579 struct vvp_io *vio = vvp_env_io(env);
583 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
586 vvp_io_update_iov(env, vio, io);
588 if (io->u.ci_rw.crw_nonblock)
589 ast_flags |= CEF_NONBLOCK;
591 result = vvp_mmap_locks(env, vio, io);
593 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
598 static int vvp_io_read_lock(const struct lu_env *env,
599 const struct cl_io_slice *ios)
601 struct cl_io *io = ios->cis_io;
602 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
606 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
607 rd->crw_pos + rd->crw_count - 1);
611 static int vvp_io_fault_lock(const struct lu_env *env,
612 const struct cl_io_slice *ios)
614 struct cl_io *io = ios->cis_io;
615 struct vvp_io *vio = cl2vvp_io(env, ios);
617 * XXX LDLM_FL_CBPENDING
619 return vvp_io_one_lock_index(env,
621 vvp_mode_from_vma(vio->u.fault.ft_vma),
622 io->u.ci_fault.ft_index,
623 io->u.ci_fault.ft_index);
626 static int vvp_io_write_lock(const struct lu_env *env,
627 const struct cl_io_slice *ios)
629 struct cl_io *io = ios->cis_io;
633 if (io->u.ci_wr.wr_append) {
635 end = OBD_OBJECT_EOF;
637 start = io->u.ci_wr.wr.crw_pos;
638 end = start + io->u.ci_wr.wr.crw_count - 1;
640 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
643 static int vvp_io_setattr_iter_init(const struct lu_env *env,
644 const struct cl_io_slice *ios)
650 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
652 * Handles "lockless io" mode when extent locking is done by server.
654 static int vvp_io_setattr_lock(const struct lu_env *env,
655 const struct cl_io_slice *ios)
657 struct cl_io *io = ios->cis_io;
661 if (cl_io_is_trunc(io)) {
662 new_size = io->u.ci_setattr.sa_attr.lvb_size;
664 enqflags = CEF_DISCARD_DATA;
666 unsigned int valid = io->u.ci_setattr.sa_valid;
668 if (!(valid & TIMES_SET_FLAGS))
671 if ((!(valid & ATTR_MTIME) ||
672 io->u.ci_setattr.sa_attr.lvb_mtime >=
673 io->u.ci_setattr.sa_attr.lvb_ctime) &&
674 (!(valid & ATTR_ATIME) ||
675 io->u.ci_setattr.sa_attr.lvb_atime >=
676 io->u.ci_setattr.sa_attr.lvb_ctime))
682 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
683 new_size, OBD_OBJECT_EOF);
686 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
691 * Only ll_inode_size_lock is taken at this level.
693 ll_inode_size_lock(inode);
694 result = inode_newsize_ok(inode, size);
696 ll_inode_size_unlock(inode);
699 i_size_write(inode, size);
701 ll_truncate_pagecache(inode, size);
702 ll_inode_size_unlock(inode);
706 static int vvp_io_setattr_time(const struct lu_env *env,
707 const struct cl_io_slice *ios)
709 struct cl_io *io = ios->cis_io;
710 struct cl_object *obj = io->ci_obj;
711 struct cl_attr *attr = vvp_env_thread_attr(env);
713 unsigned valid = CAT_CTIME;
715 cl_object_attr_lock(obj);
716 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
717 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
718 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
721 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
722 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
725 result = cl_object_attr_update(env, obj, attr, valid);
726 cl_object_attr_unlock(obj);
731 static int vvp_io_setattr_start(const struct lu_env *env,
732 const struct cl_io_slice *ios)
734 struct cl_io *io = ios->cis_io;
735 struct inode *inode = vvp_object_inode(io->ci_obj);
736 struct ll_inode_info *lli = ll_i2info(inode);
738 mutex_lock(&inode->i_mutex);
739 if (cl_io_is_trunc(io)) {
740 down_write(&lli->lli_trunc_sem);
741 inode_dio_wait(inode);
744 if (io->u.ci_setattr.sa_valid & TIMES_SET_FLAGS)
745 return vvp_io_setattr_time(env, ios);
750 static void vvp_io_setattr_end(const struct lu_env *env,
751 const struct cl_io_slice *ios)
753 struct cl_io *io = ios->cis_io;
754 struct inode *inode = vvp_object_inode(io->ci_obj);
755 struct ll_inode_info *lli = ll_i2info(inode);
757 if (cl_io_is_trunc(io)) {
758 /* Truncate in memory pages - they must be clean pages
759 * because osc has already notified to destroy osc_extents. */
760 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
761 inode_dio_write_done(inode);
762 up_write(&lli->lli_trunc_sem);
764 mutex_unlock(&inode->i_mutex);
767 static void vvp_io_setattr_fini(const struct lu_env *env,
768 const struct cl_io_slice *ios)
770 bool restore_needed = ios->cis_io->ci_restore_needed;
771 struct inode *inode = vvp_object_inode(ios->cis_obj);
773 vvp_io_fini(env, ios);
775 if (restore_needed && !ios->cis_io->ci_restore_needed) {
776 /* restore finished, set data modified flag for HSM */
777 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
781 static int vvp_io_read_start(const struct lu_env *env,
782 const struct cl_io_slice *ios)
784 struct vvp_io *vio = cl2vvp_io(env, ios);
785 struct cl_io *io = ios->cis_io;
786 struct cl_object *obj = io->ci_obj;
787 struct inode *inode = vvp_object_inode(obj);
788 struct ll_inode_info *lli = ll_i2info(inode);
789 struct file *file = vio->vui_fd->fd_file;
792 loff_t pos = io->u.ci_rd.rd.crw_pos;
793 long cnt = io->u.ci_rd.rd.crw_count;
794 long tot = vio->vui_tot_count;
797 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
799 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
801 if (vio->vui_io_subtype == IO_NORMAL)
802 down_read(&lli->lli_trunc_sem);
804 if (!can_populate_pages(env, io, inode))
807 result = vvp_prep_size(env, obj, io, pos, tot, &exceed);
810 else if (exceed != 0)
813 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
814 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
815 inode->i_ino, cnt, pos, i_size_read(inode));
817 /* turn off the kernel's read-ahead */
818 vio->vui_fd->fd_file->f_ra.ra_pages = 0;
820 /* initialize read-ahead window once per syscall */
821 if (!vio->vui_ra_valid) {
822 vio->vui_ra_valid = true;
823 vio->vui_ra_start = cl_index(obj, pos);
824 vio->vui_ra_count = cl_index(obj, tot + PAGE_CACHE_SIZE - 1);
830 switch (vio->vui_io_subtype) {
832 LASSERT(vio->vui_iocb->ki_pos == pos);
833 #ifdef HAVE_FILE_OPERATIONS_READ_WRITE_ITER
834 result = generic_file_read_iter(vio->vui_iocb, vio->vui_iter);
836 result = generic_file_aio_read(vio->vui_iocb,
838 vio->vui_iter->nr_segs,
839 vio->vui_iocb->ki_pos);
843 result = generic_file_splice_read(file, &pos,
844 vio->u.splice.vui_pipe, cnt,
845 vio->u.splice.vui_flags);
846 /* LU-1109: do splice read stripe by stripe otherwise if it
847 * may make nfsd stuck if this read occupied all internal pipe
852 CERROR("Wrong IO type %u\n", vio->vui_io_subtype);
860 io->ci_nob += result;
861 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, vio->vui_fd,
869 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
870 struct cl_page_list *plist, int from, int to)
872 struct cl_2queue *queue = &io->ci_queue;
873 struct cl_page *page;
874 unsigned int bytes = 0;
878 if (plist->pl_nr == 0)
881 if (from > 0 || to != PAGE_SIZE) {
882 page = cl_page_list_first(plist);
883 if (plist->pl_nr == 1) {
884 cl_page_clip(env, page, from, to);
887 cl_page_clip(env, page, from, PAGE_SIZE);
888 if (to != PAGE_SIZE) {
889 page = cl_page_list_last(plist);
890 cl_page_clip(env, page, 0, to);
895 cl_2queue_init(queue);
896 cl_page_list_splice(plist, &queue->c2_qin);
897 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
899 /* plist is not sorted any more */
900 cl_page_list_splice(&queue->c2_qin, plist);
901 cl_page_list_splice(&queue->c2_qout, plist);
902 cl_2queue_fini(env, queue);
905 /* calculate bytes */
906 bytes = plist->pl_nr << PAGE_SHIFT;
907 bytes -= from + PAGE_SIZE - to;
909 while (plist->pl_nr > 0) {
910 page = cl_page_list_first(plist);
911 cl_page_list_del(env, plist, page);
913 cl_page_clip(env, page, 0, PAGE_SIZE);
915 SetPageUptodate(cl_page_vmpage(page));
916 cl_page_disown(env, io, page);
918 /* held in ll_cl_init() */
919 lu_ref_del(&page->cp_reference, "cl_io", io);
920 cl_page_put(env, page);
924 RETURN(bytes > 0 ? bytes : rc);
927 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
928 struct cl_page *page)
930 struct page *vmpage = page->cp_vmpage;
932 SetPageUptodate(vmpage);
933 set_page_dirty(vmpage);
935 cl_page_disown(env, io, page);
937 /* held in ll_cl_init() */
938 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
939 cl_page_put(env, page);
942 /* make sure the page list is contiguous */
943 static bool page_list_sanity_check(struct cl_object *obj,
944 struct cl_page_list *plist)
946 struct cl_page *page;
947 pgoff_t index = CL_PAGE_EOF;
949 cl_page_list_for_each(page, plist) {
950 struct vvp_page *vpg = cl_object_page_slice(obj, page);
952 if (index == CL_PAGE_EOF) {
953 index = vvp_index(vpg);
958 if (index == vvp_index(vpg))
966 /* Return how many bytes have queued or written */
967 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
969 struct cl_object *obj = io->ci_obj;
970 struct inode *inode = vvp_object_inode(obj);
971 struct vvp_io *vio = vvp_env_io(env);
972 struct cl_page_list *queue = &vio->u.write.vui_queue;
973 struct cl_page *page;
976 unsigned int npages = vio->u.write.vui_queue.pl_nr;
982 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
983 npages, vio->u.write.vui_from, vio->u.write.vui_to);
985 LASSERT(page_list_sanity_check(obj, queue));
987 /* submit IO with async write */
988 rc = cl_io_commit_async(env, io, queue,
989 vio->u.write.vui_from, vio->u.write.vui_to,
990 write_commit_callback);
991 npages -= queue->pl_nr; /* already committed pages */
993 /* calculate how many bytes were written */
994 bytes = npages << PAGE_SHIFT;
997 bytes -= vio->u.write.vui_from;
998 if (queue->pl_nr == 0) /* last page */
999 bytes -= PAGE_SIZE - vio->u.write.vui_to;
1000 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
1002 vio->u.write.vui_written += bytes;
1004 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
1005 npages, bytes, vio->u.write.vui_written);
1007 /* the first page must have been written. */
1008 vio->u.write.vui_from = 0;
1010 LASSERT(page_list_sanity_check(obj, queue));
1011 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
1013 /* out of quota, try sync write */
1014 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
1015 rc = vvp_io_commit_sync(env, io, queue,
1016 vio->u.write.vui_from,
1017 vio->u.write.vui_to);
1019 vio->u.write.vui_written += rc;
1024 /* update inode size */
1025 ll_merge_attr(env, inode);
1027 /* Now the pages in queue were failed to commit, discard them
1028 * unless they were dirtied before. */
1029 while (queue->pl_nr > 0) {
1030 page = cl_page_list_first(queue);
1031 cl_page_list_del(env, queue, page);
1033 if (!PageDirty(cl_page_vmpage(page)))
1034 cl_page_discard(env, io, page);
1036 cl_page_disown(env, io, page);
1038 /* held in ll_cl_init() */
1039 lu_ref_del(&page->cp_reference, "cl_io", io);
1040 cl_page_put(env, page);
1042 cl_page_list_fini(env, queue);
1047 static int vvp_io_write_start(const struct lu_env *env,
1048 const struct cl_io_slice *ios)
1050 struct vvp_io *vio = cl2vvp_io(env, ios);
1051 struct cl_io *io = ios->cis_io;
1052 struct cl_object *obj = io->ci_obj;
1053 struct inode *inode = vvp_object_inode(obj);
1054 struct ll_inode_info *lli = ll_i2info(inode);
1056 loff_t pos = io->u.ci_wr.wr.crw_pos;
1057 size_t cnt = io->u.ci_wr.wr.crw_count;
1061 if (vio->vui_io_subtype == IO_NORMAL)
1062 down_read(&lli->lli_trunc_sem);
1064 if (!can_populate_pages(env, io, inode))
1067 if (cl_io_is_append(io)) {
1069 * PARALLEL IO This has to be changed for parallel IO doing
1070 * out-of-order writes.
1072 ll_merge_attr(env, inode);
1073 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
1074 vio->vui_iocb->ki_pos = pos;
1076 LASSERT(vio->vui_iocb->ki_pos == pos);
1079 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
1081 /* The maximum Lustre file size is variable, based on the OST maximum
1082 * object size and number of stripes. This needs another check in
1083 * addition to the VFS checks earlier. */
1084 if (pos + cnt > ll_file_maxbytes(inode)) {
1086 "%s: file "DFID" offset %llu > maxbytes "LPU64"\n",
1087 ll_get_fsname(inode->i_sb, NULL, 0),
1088 PFID(ll_inode2fid(inode)), pos + cnt,
1089 ll_file_maxbytes(inode));
1093 if (vio->vui_iter == NULL) {
1094 /* from a temp io in ll_cl_init(). */
1098 * When using the locked AIO function (generic_file_aio_write())
1099 * testing has shown the inode mutex to be a limiting factor
1100 * with multi-threaded single shared file performance. To get
1101 * around this, we now use the lockless version. To maintain
1102 * consistency, proper locking to protect against writes,
1103 * trucates, etc. is handled in the higher layers of lustre.
1105 #ifdef HAVE_FILE_OPERATIONS_READ_WRITE_ITER
1106 result = generic_file_write_iter(vio->vui_iocb, vio->vui_iter);
1108 result = __generic_file_aio_write(vio->vui_iocb,
1110 vio->vui_iter->nr_segs,
1111 &vio->vui_iocb->ki_pos);
1113 if (result > 0 || result == -EIOCBQUEUED) {
1116 err = generic_write_sync(vio->vui_iocb->ki_filp,
1118 if (err < 0 && result > 0)
1124 result = vvp_io_write_commit(env, io);
1125 if (vio->u.write.vui_written > 0) {
1126 result = vio->u.write.vui_written;
1127 io->ci_nob += result;
1129 CDEBUG(D_VFSTRACE, "write: nob %zd, result: %zd\n",
1130 io->ci_nob, result);
1134 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
1137 io->ci_continue = 0;
1138 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
1139 vio->vui_fd, pos, result, WRITE);
1146 static void vvp_io_rw_end(const struct lu_env *env,
1147 const struct cl_io_slice *ios)
1149 struct vvp_io *vio = cl2vvp_io(env, ios);
1150 struct inode *inode = vvp_object_inode(ios->cis_obj);
1151 struct ll_inode_info *lli = ll_i2info(inode);
1153 if (vio->vui_io_subtype == IO_NORMAL)
1154 up_read(&lli->lli_trunc_sem);
1157 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
1159 struct vm_fault *vmf = cfio->ft_vmf;
1161 cfio->ft_flags = filemap_fault(cfio->ft_vma, vmf);
1162 cfio->ft_flags_valid = 1;
1165 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
1166 vmf->virtual_address);
1167 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
1168 lock_page(vmf->page);
1169 cfio->ft_flags |= VM_FAULT_LOCKED;
1172 cfio->ft_vmpage = vmf->page;
1177 if (cfio->ft_flags & VM_FAULT_SIGBUS) {
1178 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
1182 if (cfio->ft_flags & VM_FAULT_OOM) {
1183 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
1187 if (cfio->ft_flags & VM_FAULT_RETRY)
1190 CERROR("unknown error in page fault %d\n", cfio->ft_flags);
1195 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1196 struct cl_page *page)
1198 set_page_dirty(page->cp_vmpage);
1201 static int vvp_io_fault_start(const struct lu_env *env,
1202 const struct cl_io_slice *ios)
1204 struct vvp_io *vio = cl2vvp_io(env, ios);
1205 struct cl_io *io = ios->cis_io;
1206 struct cl_object *obj = io->ci_obj;
1207 struct inode *inode = vvp_object_inode(obj);
1208 struct ll_inode_info *lli = ll_i2info(inode);
1209 struct cl_fault_io *fio = &io->u.ci_fault;
1210 struct vvp_fault_io *cfio = &vio->u.fault;
1213 struct page *vmpage = NULL;
1214 struct cl_page *page;
1219 if (fio->ft_executable &&
1220 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
1222 " changed while waiting for the page fault lock\n",
1223 PFID(lu_object_fid(&obj->co_lu)));
1225 down_read(&lli->lli_trunc_sem);
1227 /* offset of the last byte on the page */
1228 offset = cl_offset(obj, fio->ft_index + 1) - 1;
1229 LASSERT(cl_index(obj, offset) == fio->ft_index);
1230 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1234 /* must return locked page */
1235 if (fio->ft_mkwrite) {
1236 LASSERT(cfio->ft_vmpage != NULL);
1237 lock_page(cfio->ft_vmpage);
1239 result = vvp_io_kernel_fault(cfio);
1244 vmpage = cfio->ft_vmpage;
1245 LASSERT(PageLocked(vmpage));
1247 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1248 ll_invalidate_page(vmpage);
1250 size = i_size_read(inode);
1251 /* Though we have already held a cl_lock upon this page, but
1252 * it still can be truncated locally. */
1253 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1254 (page_offset(vmpage) > size))) {
1255 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1257 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1259 GOTO(out, result = +1);
1262 last_index = cl_index(obj, size - 1);
1264 if (fio->ft_mkwrite ) {
1266 * Capture the size while holding the lli_trunc_sem from above
1267 * we want to make sure that we complete the mkwrite action
1268 * while holding this lock. We need to make sure that we are
1269 * not past the end of the file.
1271 if (last_index < fio->ft_index) {
1273 "llite: mkwrite and truncate race happened: "
1274 "%p: 0x%lx 0x%lx\n",
1275 vmpage->mapping,fio->ft_index,last_index);
1277 * We need to return if we are
1278 * passed the end of the file. This will propagate
1279 * up the call stack to ll_page_mkwrite where
1280 * we will return VM_FAULT_NOPAGE. Any non-negative
1281 * value returned here will be silently
1282 * converted to 0. If the vmpage->mapping is null
1283 * the error code would be converted back to ENODATA
1284 * in ll_page_mkwrite0. Thus we return -ENODATA
1285 * to handle both cases
1287 GOTO(out, result = -ENODATA);
1291 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1293 GOTO(out, result = PTR_ERR(page));
1295 /* if page is going to be written, we should add this page into cache
1297 if (fio->ft_mkwrite) {
1298 wait_on_page_writeback(vmpage);
1299 if (!PageDirty(vmpage)) {
1300 struct cl_page_list *plist = &io->ci_queue.c2_qin;
1301 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1304 /* vvp_page_assume() calls wait_on_page_writeback(). */
1305 cl_page_assume(env, io, page);
1307 cl_page_list_init(plist);
1308 cl_page_list_add(plist, page);
1311 if (last_index == vvp_index(vpg))
1312 to = size & ~PAGE_MASK;
1314 /* Do not set Dirty bit here so that in case IO is
1315 * started before the page is really made dirty, we
1316 * still have chance to detect it. */
1317 result = cl_io_commit_async(env, io, plist, 0, to,
1318 mkwrite_commit_callback);
1319 LASSERT(cl_page_is_owned(page, io));
1320 cl_page_list_fini(env, plist);
1324 cl_page_discard(env, io, page);
1325 cl_page_disown(env, io, page);
1327 cl_page_put(env, page);
1329 /* we're in big trouble, what can we do now? */
1330 if (result == -EDQUOT)
1334 cl_page_disown(env, io, page);
1339 * The ft_index is only used in the case of
1340 * a mkwrite action. We need to check
1341 * our assertions are correct, since
1342 * we should have caught this above
1344 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1345 if (fio->ft_index == last_index)
1347 * Last page is mapped partially.
1349 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1351 fio->ft_nob = cl_page_size(obj);
1353 lu_ref_add(&page->cp_reference, "fault", io);
1354 fio->ft_page = page;
1358 /* return unlocked vmpage to avoid deadlocking */
1360 unlock_page(vmpage);
1362 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1367 static void vvp_io_fault_end(const struct lu_env *env,
1368 const struct cl_io_slice *ios)
1370 struct inode *inode = vvp_object_inode(ios->cis_obj);
1371 struct ll_inode_info *lli = ll_i2info(inode);
1373 CLOBINVRNT(env, ios->cis_io->ci_obj,
1374 vvp_object_invariant(ios->cis_io->ci_obj));
1375 up_read(&lli->lli_trunc_sem);
1378 static int vvp_io_fsync_start(const struct lu_env *env,
1379 const struct cl_io_slice *ios)
1381 /* we should mark TOWRITE bit to each dirty page in radix tree to
1382 * verify pages have been written, but this is difficult because of
1387 static int vvp_io_read_ahead(const struct lu_env *env,
1388 const struct cl_io_slice *ios,
1389 pgoff_t start, struct cl_read_ahead *ra)
1394 if (ios->cis_io->ci_type == CIT_READ ||
1395 ios->cis_io->ci_type == CIT_FAULT) {
1396 struct vvp_io *vio = cl2vvp_io(env, ios);
1398 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1399 ra->cra_end = CL_PAGE_EOF;
1400 result = +1; /* no need to call down */
1407 static const struct cl_io_operations vvp_io_ops = {
1410 .cio_fini = vvp_io_fini,
1411 .cio_lock = vvp_io_read_lock,
1412 .cio_start = vvp_io_read_start,
1413 .cio_end = vvp_io_rw_end,
1414 .cio_advance = vvp_io_advance,
1417 .cio_fini = vvp_io_fini,
1418 .cio_iter_init = vvp_io_write_iter_init,
1419 .cio_iter_fini = vvp_io_write_iter_fini,
1420 .cio_lock = vvp_io_write_lock,
1421 .cio_start = vvp_io_write_start,
1422 .cio_end = vvp_io_rw_end,
1423 .cio_advance = vvp_io_advance,
1426 .cio_fini = vvp_io_setattr_fini,
1427 .cio_iter_init = vvp_io_setattr_iter_init,
1428 .cio_lock = vvp_io_setattr_lock,
1429 .cio_start = vvp_io_setattr_start,
1430 .cio_end = vvp_io_setattr_end
1433 .cio_fini = vvp_io_fault_fini,
1434 .cio_iter_init = vvp_io_fault_iter_init,
1435 .cio_lock = vvp_io_fault_lock,
1436 .cio_start = vvp_io_fault_start,
1437 .cio_end = vvp_io_fault_end,
1440 .cio_start = vvp_io_fsync_start,
1441 .cio_fini = vvp_io_fini
1444 .cio_fini = vvp_io_fini
1447 .cio_read_ahead = vvp_io_read_ahead
1450 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1453 struct vvp_io *vio = vvp_env_io(env);
1454 struct inode *inode = vvp_object_inode(obj);
1457 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1460 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1461 "restore needed %d\n",
1462 PFID(lu_object_fid(&obj->co_lu)),
1463 io->ci_ignore_layout, io->ci_verify_layout,
1464 vio->vui_layout_gen, io->ci_restore_needed);
1466 CL_IO_SLICE_CLEAN(vio, vui_cl);
1467 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1468 vio->vui_ra_valid = false;
1470 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1472 struct ll_inode_info *lli = ll_i2info(inode);
1474 count = io->u.ci_rw.crw_count;
1475 /* "If nbyte is 0, read() will return 0 and have no other
1476 * results." -- Single Unix Spec */
1480 vio->vui_tot_count = count;
1481 #ifndef HAVE_FILE_OPERATIONS_READ_WRITE_ITER
1482 vio->vui_tot_nrsegs = 0;
1486 /* for read/write, we store the jobid in the inode, and
1487 * it'll be fetched by osc when building RPC.
1489 * it's not accurate if the file is shared by different
1492 lustre_get_jobid(lli->lli_jobid);
1493 } else if (io->ci_type == CIT_SETATTR) {
1494 if (!cl_io_is_trunc(io))
1495 io->ci_lockreq = CILR_MANDATORY;
1498 /* ignore layout change for generic CIT_MISC but not for glimpse.
1499 * io context for glimpse must set ci_verify_layout to true,
1500 * see cl_glimpse_size0() for details. */
1501 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1502 io->ci_ignore_layout = 1;
1504 /* Enqueue layout lock and get layout version. We need to do this
1505 * even for operations requiring to open file, such as read and write,
1506 * because it might not grant layout lock in IT_OPEN. */
1507 if (result == 0 && !io->ci_ignore_layout) {
1508 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1509 if (result == -ENOENT)
1510 /* If the inode on MDS has been removed, but the objects
1511 * on OSTs haven't been destroyed (async unlink), layout
1512 * fetch will return -ENOENT, we'd ingore this error
1513 * and continue with dirty flush. LU-3230. */
1516 CERROR("%s: refresh file layout " DFID " error %d.\n",
1517 ll_get_fsname(inode->i_sb, NULL, 0),
1518 PFID(lu_object_fid(&obj->co_lu)), result);