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, 2015, 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;
241 enqflags |= CEF_LOCK_MATCH;
243 descr->cld_mode = mode;
246 descr->cld_obj = obj;
247 descr->cld_start = start;
248 descr->cld_end = end;
249 descr->cld_enq_flags = enqflags;
251 cl_io_lock_add(env, io, &vio->vui_link);
256 static int vvp_io_one_lock(const struct lu_env *env, struct cl_io *io,
257 __u32 enqflags, enum cl_lock_mode mode,
258 loff_t start, loff_t end)
260 struct cl_object *obj = io->ci_obj;
262 return vvp_io_one_lock_index(env, io, enqflags, mode,
263 cl_index(obj, start), cl_index(obj, end));
266 static int vvp_io_write_iter_init(const struct lu_env *env,
267 const struct cl_io_slice *ios)
269 struct vvp_io *vio = cl2vvp_io(env, ios);
271 cl_page_list_init(&vio->u.write.vui_queue);
272 vio->u.write.vui_written = 0;
273 vio->u.write.vui_from = 0;
274 vio->u.write.vui_to = PAGE_SIZE;
279 static void vvp_io_write_iter_fini(const struct lu_env *env,
280 const struct cl_io_slice *ios)
282 struct vvp_io *vio = cl2vvp_io(env, ios);
284 LASSERT(vio->u.write.vui_queue.pl_nr == 0);
287 static int vvp_io_fault_iter_init(const struct lu_env *env,
288 const struct cl_io_slice *ios)
290 struct vvp_io *vio = cl2vvp_io(env, ios);
291 struct inode *inode = vvp_object_inode(ios->cis_obj);
293 LASSERT(inode == vio->vui_fd->fd_file->f_path.dentry->d_inode);
294 vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime);
299 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
301 struct cl_io *io = ios->cis_io;
302 struct cl_object *obj = io->ci_obj;
303 struct vvp_io *vio = cl2vvp_io(env, ios);
304 struct inode *inode = vvp_object_inode(obj);
306 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
308 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
309 "restore needed %d\n",
310 PFID(lu_object_fid(&obj->co_lu)),
311 io->ci_ignore_layout, io->ci_verify_layout,
312 vio->vui_layout_gen, io->ci_restore_needed);
314 if (io->ci_restore_needed) {
317 /* file was detected release, we need to restore it
318 * before finishing the io
320 rc = ll_layout_restore(inode, 0, OBD_OBJECT_EOF);
321 /* if restore registration failed, no restart,
322 * we will return -ENODATA */
323 /* The layout will change after restore, so we need to
324 * block on layout lock held by the MDT
325 * as MDT will not send new layout in lvb (see LU-3124)
326 * we have to explicitly fetch it, all this will be done
327 * by ll_layout_refresh()
330 io->ci_restore_needed = 0;
331 io->ci_need_restart = 1;
332 io->ci_verify_layout = 1;
334 io->ci_restore_needed = 1;
335 io->ci_need_restart = 0;
336 io->ci_verify_layout = 0;
341 if (!io->ci_ignore_layout && io->ci_verify_layout) {
344 /* check layout version */
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);
361 static void vvp_io_fault_fini(const struct lu_env *env,
362 const struct cl_io_slice *ios)
364 struct cl_io *io = ios->cis_io;
365 struct cl_page *page = io->u.ci_fault.ft_page;
367 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
370 lu_ref_del(&page->cp_reference, "fault", io);
371 cl_page_put(env, page);
372 io->u.ci_fault.ft_page = NULL;
374 vvp_io_fini(env, ios);
377 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
380 * we only want to hold PW locks if the mmap() can generate
381 * writes back to the file and that only happens in shared
384 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
389 static int vvp_mmap_locks(const struct lu_env *env,
390 struct vvp_io *vio, struct cl_io *io)
392 struct vvp_thread_info *vti = vvp_env_info(env);
393 struct mm_struct *mm = current->mm;
394 struct vm_area_struct *vma;
395 struct cl_lock_descr *descr = &vti->vti_descr;
396 union ldlm_policy_data policy;
397 #ifdef HAVE_FILE_OPERATIONS_READ_WRITE_ITER
406 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
408 if (!cl_is_normalio(env, io))
411 /* nfs or loop back device write */
412 if (vio->vui_iter == NULL)
415 /* No MM (e.g. NFS)? No vmas too. */
419 #ifdef HAVE_FILE_OPERATIONS_READ_WRITE_ITER
420 iov_for_each(iov, i, *(vio->vui_iter)) {
421 #else /* !HAVE_FILE_OPERATIONS_READ_WRITE_ITER */
422 for (seg = 0; seg < vio->vui_iter->nr_segs; seg++) {
423 const struct iovec iov = vio->vui_iter->iov[seg];
424 #endif /* HAVE_FILE_OPERATIONS_READ_WRITE_ITER */
425 unsigned long addr = (unsigned long)iov.iov_base;
426 size_t count = iov.iov_len;
431 count += addr & ~PAGE_MASK;
434 down_read(&mm->mmap_sem);
435 while((vma = our_vma(mm, addr, count)) != NULL) {
436 struct dentry *de = vma->vm_file->f_path.dentry;
437 struct inode *inode = de->d_inode;
438 int flags = CEF_MUST;
440 if (ll_file_nolock(vma->vm_file)) {
442 * For no lock case is not allowed for mmap
449 * XXX: Required lock mode can be weakened: CIT_WRITE
450 * io only ever reads user level buffer, and CIT_READ
453 policy_from_vma(&policy, vma, addr, count);
454 descr->cld_mode = vvp_mode_from_vma(vma);
455 descr->cld_obj = ll_i2info(inode)->lli_clob;
456 descr->cld_start = cl_index(descr->cld_obj,
457 policy.l_extent.start);
458 descr->cld_end = cl_index(descr->cld_obj,
459 policy.l_extent.end);
460 descr->cld_enq_flags = flags;
461 result = cl_io_lock_alloc_add(env, io, descr);
463 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
464 descr->cld_mode, descr->cld_start,
470 if (vma->vm_end - addr >= count)
473 count -= vma->vm_end - addr;
476 up_read(&mm->mmap_sem);
483 static void vvp_io_advance(const struct lu_env *env,
484 const struct cl_io_slice *ios,
487 struct vvp_io *vio = cl2vvp_io(env, ios);
488 struct cl_io *io = ios->cis_io;
489 struct cl_object *obj = ios->cis_io->ci_obj;
490 #ifndef HAVE_FILE_OPERATIONS_READ_WRITE_ITER
493 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
495 if (!cl_is_normalio(env, io))
498 #ifdef HAVE_FILE_OPERATIONS_READ_WRITE_ITER
499 vio->vui_tot_count -= nob;
500 iov_iter_reexpand(vio->vui_iter, vio->vui_tot_count);
501 #else /* !HAVE_FILE_OPERATIONS_READ_WRITE_ITER */
502 LASSERT(vio->vui_tot_nrsegs >= vio->vui_iter->nr_segs);
503 LASSERT(vio->vui_tot_count >= nob);
505 /* Restore the iov changed in vvp_io_update_iov() */
506 if (vio->vui_iov_olen > 0) {
507 unsigned long idx = vio->vui_iter->nr_segs - 1;
509 /* In the latest kernels iov is const so that
510 * changes are done using iter helpers. In older
511 * kernels those helpers don't exist we lustre
512 * has to do some of the management of the iter
514 iov = (struct iovec *)&vio->vui_iter->iov[idx];
515 iov->iov_len = vio->vui_iov_olen;
516 vio->vui_iov_olen = 0;
519 /* In the latest kernels special helpers exist to help
520 * advance the iov but we don't have that in older kernels
521 * so we need to do the book keeping ourselves. */
522 iov = (struct iovec *)vio->vui_iter->iov;
524 if (iov->iov_len > nob) {
526 iov->iov_base += nob;
532 vio->vui_tot_nrsegs--;
535 vio->vui_iter->iov = iov;
536 vio->vui_tot_count -= nob;
537 #endif /* HAVE_FILE_OPERATIONS_READ_WRITE_ITER */
540 static void vvp_io_update_iov(const struct lu_env *env,
541 struct vvp_io *vio, struct cl_io *io)
543 size_t size = io->u.ci_rw.crw_count;
544 #ifdef HAVE_FILE_OPERATIONS_READ_WRITE_ITER
545 if (!cl_is_normalio(env, io) || vio->vui_iter == NULL)
548 iov_iter_truncate(vio->vui_iter, size);
549 #else /* !HAVE_FILE_OPERATIONS_READ_WRITE_ITER */
552 vio->vui_iov_olen = 0;
553 if (!cl_is_normalio(env, io) || vio->vui_tot_nrsegs == 0)
556 for (i = 0; i < vio->vui_tot_nrsegs; i++) {
557 struct iovec *iv = (struct iovec *) &vio->vui_iter->iov[i];
559 if (iv->iov_len < size) {
562 if (iv->iov_len > size) {
563 vio->vui_iov_olen = iv->iov_len;
570 vio->vui_iter->nr_segs = i + 1;
571 LASSERTF(vio->vui_tot_nrsegs >= vio->vui_iter->nr_segs,
572 "tot_nrsegs: %lu, nrsegs: %lu\n",
573 vio->vui_tot_nrsegs, vio->vui_iter->nr_segs);
574 #endif /* !HAVE_FILE_OPERATIONS_READ_WRITE_ITER */
577 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
578 enum cl_lock_mode mode, loff_t start, loff_t end)
580 struct vvp_io *vio = vvp_env_io(env);
584 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
587 vvp_io_update_iov(env, vio, io);
589 if (io->u.ci_rw.crw_nonblock)
590 ast_flags |= CEF_NONBLOCK;
592 result = vvp_mmap_locks(env, vio, io);
594 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
599 static int vvp_io_read_lock(const struct lu_env *env,
600 const struct cl_io_slice *ios)
602 struct cl_io *io = ios->cis_io;
603 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
607 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
608 rd->crw_pos + rd->crw_count - 1);
612 static int vvp_io_fault_lock(const struct lu_env *env,
613 const struct cl_io_slice *ios)
615 struct cl_io *io = ios->cis_io;
616 struct vvp_io *vio = cl2vvp_io(env, ios);
618 * XXX LDLM_FL_CBPENDING
620 return vvp_io_one_lock_index(env,
622 vvp_mode_from_vma(vio->u.fault.ft_vma),
623 io->u.ci_fault.ft_index,
624 io->u.ci_fault.ft_index);
627 static int vvp_io_write_lock(const struct lu_env *env,
628 const struct cl_io_slice *ios)
630 struct cl_io *io = ios->cis_io;
634 if (io->u.ci_wr.wr_append) {
636 end = OBD_OBJECT_EOF;
638 start = io->u.ci_wr.wr.crw_pos;
639 end = start + io->u.ci_wr.wr.crw_count - 1;
641 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
644 static int vvp_io_setattr_iter_init(const struct lu_env *env,
645 const struct cl_io_slice *ios)
651 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
653 * Handles "lockless io" mode when extent locking is done by server.
655 static int vvp_io_setattr_lock(const struct lu_env *env,
656 const struct cl_io_slice *ios)
658 struct cl_io *io = ios->cis_io;
662 if (cl_io_is_trunc(io)) {
663 new_size = io->u.ci_setattr.sa_attr.lvb_size;
665 enqflags = CEF_DISCARD_DATA;
667 unsigned int valid = io->u.ci_setattr.sa_valid;
669 if (!(valid & TIMES_SET_FLAGS))
672 if ((!(valid & ATTR_MTIME) ||
673 io->u.ci_setattr.sa_attr.lvb_mtime >=
674 io->u.ci_setattr.sa_attr.lvb_ctime) &&
675 (!(valid & ATTR_ATIME) ||
676 io->u.ci_setattr.sa_attr.lvb_atime >=
677 io->u.ci_setattr.sa_attr.lvb_ctime))
683 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
684 new_size, OBD_OBJECT_EOF);
687 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
692 * Only ll_inode_size_lock is taken at this level.
694 ll_inode_size_lock(inode);
695 result = inode_newsize_ok(inode, size);
697 ll_inode_size_unlock(inode);
700 i_size_write(inode, size);
702 ll_truncate_pagecache(inode, size);
703 ll_inode_size_unlock(inode);
707 static int vvp_io_setattr_time(const struct lu_env *env,
708 const struct cl_io_slice *ios)
710 struct cl_io *io = ios->cis_io;
711 struct cl_object *obj = io->ci_obj;
712 struct cl_attr *attr = vvp_env_thread_attr(env);
714 unsigned valid = CAT_CTIME;
716 cl_object_attr_lock(obj);
717 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
718 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
719 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
722 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
723 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
726 result = cl_object_attr_update(env, obj, attr, valid);
727 cl_object_attr_unlock(obj);
732 static int vvp_io_setattr_start(const struct lu_env *env,
733 const struct cl_io_slice *ios)
735 struct cl_io *io = ios->cis_io;
736 struct inode *inode = vvp_object_inode(io->ci_obj);
737 struct ll_inode_info *lli = ll_i2info(inode);
739 mutex_lock(&inode->i_mutex);
740 if (cl_io_is_trunc(io)) {
741 down_write(&lli->lli_trunc_sem);
742 inode_dio_wait(inode);
745 if (io->u.ci_setattr.sa_valid & TIMES_SET_FLAGS)
746 return vvp_io_setattr_time(env, ios);
751 static void vvp_io_setattr_end(const struct lu_env *env,
752 const struct cl_io_slice *ios)
754 struct cl_io *io = ios->cis_io;
755 struct inode *inode = vvp_object_inode(io->ci_obj);
756 struct ll_inode_info *lli = ll_i2info(inode);
758 if (cl_io_is_trunc(io)) {
759 /* Truncate in memory pages - they must be clean pages
760 * because osc has already notified to destroy osc_extents. */
761 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
762 inode_dio_write_done(inode);
763 up_write(&lli->lli_trunc_sem);
765 mutex_unlock(&inode->i_mutex);
768 static void vvp_io_setattr_fini(const struct lu_env *env,
769 const struct cl_io_slice *ios)
771 bool restore_needed = ios->cis_io->ci_restore_needed;
772 struct inode *inode = vvp_object_inode(ios->cis_obj);
774 vvp_io_fini(env, ios);
776 if (restore_needed && !ios->cis_io->ci_restore_needed) {
777 /* restore finished, set data modified flag for HSM */
778 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
782 static int vvp_io_read_start(const struct lu_env *env,
783 const struct cl_io_slice *ios)
785 struct vvp_io *vio = cl2vvp_io(env, ios);
786 struct cl_io *io = ios->cis_io;
787 struct cl_object *obj = io->ci_obj;
788 struct inode *inode = vvp_object_inode(obj);
789 struct ll_inode_info *lli = ll_i2info(inode);
790 struct file *file = vio->vui_fd->fd_file;
793 loff_t pos = io->u.ci_rd.rd.crw_pos;
794 long cnt = io->u.ci_rd.rd.crw_count;
795 long tot = vio->vui_tot_count;
798 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
800 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
802 if (vio->vui_io_subtype == IO_NORMAL)
803 down_read(&lli->lli_trunc_sem);
805 if (!can_populate_pages(env, io, inode))
808 result = vvp_prep_size(env, obj, io, pos, tot, &exceed);
811 else if (exceed != 0)
814 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
815 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
816 inode->i_ino, cnt, pos, i_size_read(inode));
818 /* turn off the kernel's read-ahead */
819 vio->vui_fd->fd_file->f_ra.ra_pages = 0;
821 /* initialize read-ahead window once per syscall */
822 if (!vio->vui_ra_valid) {
823 vio->vui_ra_valid = true;
824 vio->vui_ra_start = cl_index(obj, pos);
825 vio->vui_ra_count = cl_index(obj, tot + PAGE_CACHE_SIZE - 1);
831 switch (vio->vui_io_subtype) {
833 LASSERT(vio->vui_iocb->ki_pos == pos);
834 #ifdef HAVE_FILE_OPERATIONS_READ_WRITE_ITER
835 result = generic_file_read_iter(vio->vui_iocb, vio->vui_iter);
837 result = generic_file_aio_read(vio->vui_iocb,
839 vio->vui_iter->nr_segs,
840 vio->vui_iocb->ki_pos);
844 result = generic_file_splice_read(file, &pos,
845 vio->u.splice.vui_pipe, cnt,
846 vio->u.splice.vui_flags);
847 /* LU-1109: do splice read stripe by stripe otherwise if it
848 * may make nfsd stuck if this read occupied all internal pipe
853 CERROR("Wrong IO type %u\n", vio->vui_io_subtype);
861 io->ci_nob += result;
862 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, vio->vui_fd,
870 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
871 struct cl_page_list *plist, int from, int to)
873 struct cl_2queue *queue = &io->ci_queue;
874 struct cl_page *page;
875 unsigned int bytes = 0;
879 if (plist->pl_nr == 0)
882 if (from > 0 || to != PAGE_SIZE) {
883 page = cl_page_list_first(plist);
884 if (plist->pl_nr == 1) {
885 cl_page_clip(env, page, from, to);
888 cl_page_clip(env, page, from, PAGE_SIZE);
889 if (to != PAGE_SIZE) {
890 page = cl_page_list_last(plist);
891 cl_page_clip(env, page, 0, to);
896 cl_2queue_init(queue);
897 cl_page_list_splice(plist, &queue->c2_qin);
898 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
900 /* plist is not sorted any more */
901 cl_page_list_splice(&queue->c2_qin, plist);
902 cl_page_list_splice(&queue->c2_qout, plist);
903 cl_2queue_fini(env, queue);
906 /* calculate bytes */
907 bytes = plist->pl_nr << PAGE_SHIFT;
908 bytes -= from + PAGE_SIZE - to;
910 while (plist->pl_nr > 0) {
911 page = cl_page_list_first(plist);
912 cl_page_list_del(env, plist, page);
914 cl_page_clip(env, page, 0, PAGE_SIZE);
916 SetPageUptodate(cl_page_vmpage(page));
917 cl_page_disown(env, io, page);
919 /* held in ll_cl_init() */
920 lu_ref_del(&page->cp_reference, "cl_io", io);
921 cl_page_put(env, page);
925 RETURN(bytes > 0 ? bytes : rc);
928 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
929 struct cl_page *page)
931 struct page *vmpage = page->cp_vmpage;
933 SetPageUptodate(vmpage);
934 set_page_dirty(vmpage);
936 cl_page_disown(env, io, page);
938 /* held in ll_cl_init() */
939 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
940 cl_page_put(env, page);
943 /* make sure the page list is contiguous */
944 static bool page_list_sanity_check(struct cl_object *obj,
945 struct cl_page_list *plist)
947 struct cl_page *page;
948 pgoff_t index = CL_PAGE_EOF;
950 cl_page_list_for_each(page, plist) {
951 struct vvp_page *vpg = cl_object_page_slice(obj, page);
953 if (index == CL_PAGE_EOF) {
954 index = vvp_index(vpg);
959 if (index == vvp_index(vpg))
967 /* Return how many bytes have queued or written */
968 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
970 struct cl_object *obj = io->ci_obj;
971 struct inode *inode = vvp_object_inode(obj);
972 struct vvp_io *vio = vvp_env_io(env);
973 struct cl_page_list *queue = &vio->u.write.vui_queue;
974 struct cl_page *page;
977 unsigned int npages = vio->u.write.vui_queue.pl_nr;
983 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
984 npages, vio->u.write.vui_from, vio->u.write.vui_to);
986 LASSERT(page_list_sanity_check(obj, queue));
988 /* submit IO with async write */
989 rc = cl_io_commit_async(env, io, queue,
990 vio->u.write.vui_from, vio->u.write.vui_to,
991 write_commit_callback);
992 npages -= queue->pl_nr; /* already committed pages */
994 /* calculate how many bytes were written */
995 bytes = npages << PAGE_SHIFT;
998 bytes -= vio->u.write.vui_from;
999 if (queue->pl_nr == 0) /* last page */
1000 bytes -= PAGE_SIZE - vio->u.write.vui_to;
1001 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
1003 vio->u.write.vui_written += bytes;
1005 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
1006 npages, bytes, vio->u.write.vui_written);
1008 /* the first page must have been written. */
1009 vio->u.write.vui_from = 0;
1011 LASSERT(page_list_sanity_check(obj, queue));
1012 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
1014 /* out of quota, try sync write */
1015 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
1016 rc = vvp_io_commit_sync(env, io, queue,
1017 vio->u.write.vui_from,
1018 vio->u.write.vui_to);
1020 vio->u.write.vui_written += rc;
1025 /* update inode size */
1026 ll_merge_attr(env, inode);
1028 /* Now the pages in queue were failed to commit, discard them
1029 * unless they were dirtied before. */
1030 while (queue->pl_nr > 0) {
1031 page = cl_page_list_first(queue);
1032 cl_page_list_del(env, queue, page);
1034 if (!PageDirty(cl_page_vmpage(page)))
1035 cl_page_discard(env, io, page);
1037 cl_page_disown(env, io, page);
1039 /* held in ll_cl_init() */
1040 lu_ref_del(&page->cp_reference, "cl_io", io);
1041 cl_page_put(env, page);
1043 cl_page_list_fini(env, queue);
1048 static int vvp_io_write_start(const struct lu_env *env,
1049 const struct cl_io_slice *ios)
1051 struct vvp_io *vio = cl2vvp_io(env, ios);
1052 struct cl_io *io = ios->cis_io;
1053 struct cl_object *obj = io->ci_obj;
1054 struct inode *inode = vvp_object_inode(obj);
1055 struct ll_inode_info *lli = ll_i2info(inode);
1057 loff_t pos = io->u.ci_wr.wr.crw_pos;
1058 size_t cnt = io->u.ci_wr.wr.crw_count;
1062 if (vio->vui_io_subtype == IO_NORMAL)
1063 down_read(&lli->lli_trunc_sem);
1065 if (!can_populate_pages(env, io, inode))
1068 if (cl_io_is_append(io)) {
1070 * PARALLEL IO This has to be changed for parallel IO doing
1071 * out-of-order writes.
1073 ll_merge_attr(env, inode);
1074 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
1075 vio->vui_iocb->ki_pos = pos;
1077 LASSERT(vio->vui_iocb->ki_pos == pos);
1080 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
1082 /* The maximum Lustre file size is variable, based on the OST maximum
1083 * object size and number of stripes. This needs another check in
1084 * addition to the VFS checks earlier. */
1085 if (pos + cnt > ll_file_maxbytes(inode)) {
1087 "%s: file "DFID" offset %llu > maxbytes "LPU64"\n",
1088 ll_get_fsname(inode->i_sb, NULL, 0),
1089 PFID(ll_inode2fid(inode)), pos + cnt,
1090 ll_file_maxbytes(inode));
1094 if (vio->vui_iter == NULL) {
1095 /* from a temp io in ll_cl_init(). */
1099 * When using the locked AIO function (generic_file_aio_write())
1100 * testing has shown the inode mutex to be a limiting factor
1101 * with multi-threaded single shared file performance. To get
1102 * around this, we now use the lockless version. To maintain
1103 * consistency, proper locking to protect against writes,
1104 * trucates, etc. is handled in the higher layers of lustre.
1106 #ifdef HAVE_FILE_OPERATIONS_READ_WRITE_ITER
1107 result = generic_file_write_iter(vio->vui_iocb, vio->vui_iter);
1109 result = __generic_file_aio_write(vio->vui_iocb,
1111 vio->vui_iter->nr_segs,
1112 &vio->vui_iocb->ki_pos);
1114 if (result > 0 || result == -EIOCBQUEUED) {
1117 err = generic_write_sync(vio->vui_iocb->ki_filp,
1119 if (err < 0 && result > 0)
1125 result = vvp_io_write_commit(env, io);
1126 if (vio->u.write.vui_written > 0) {
1127 result = vio->u.write.vui_written;
1128 io->ci_nob += result;
1130 CDEBUG(D_VFSTRACE, "write: nob %zd, result: %zd\n",
1131 io->ci_nob, result);
1135 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
1138 io->ci_continue = 0;
1139 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
1140 vio->vui_fd, pos, result, WRITE);
1147 static void vvp_io_rw_end(const struct lu_env *env,
1148 const struct cl_io_slice *ios)
1150 struct vvp_io *vio = cl2vvp_io(env, ios);
1151 struct inode *inode = vvp_object_inode(ios->cis_obj);
1152 struct ll_inode_info *lli = ll_i2info(inode);
1154 if (vio->vui_io_subtype == IO_NORMAL)
1155 up_read(&lli->lli_trunc_sem);
1158 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
1160 struct vm_fault *vmf = cfio->ft_vmf;
1162 cfio->ft_flags = filemap_fault(cfio->ft_vma, vmf);
1163 cfio->ft_flags_valid = 1;
1166 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
1167 vmf->virtual_address);
1168 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
1169 lock_page(vmf->page);
1170 cfio->ft_flags |= VM_FAULT_LOCKED;
1173 cfio->ft_vmpage = vmf->page;
1178 if (cfio->ft_flags & VM_FAULT_SIGBUS) {
1179 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
1183 if (cfio->ft_flags & VM_FAULT_OOM) {
1184 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
1188 if (cfio->ft_flags & VM_FAULT_RETRY)
1191 CERROR("unknown error in page fault %d\n", cfio->ft_flags);
1196 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1197 struct cl_page *page)
1199 set_page_dirty(page->cp_vmpage);
1202 static int vvp_io_fault_start(const struct lu_env *env,
1203 const struct cl_io_slice *ios)
1205 struct vvp_io *vio = cl2vvp_io(env, ios);
1206 struct cl_io *io = ios->cis_io;
1207 struct cl_object *obj = io->ci_obj;
1208 struct inode *inode = vvp_object_inode(obj);
1209 struct ll_inode_info *lli = ll_i2info(inode);
1210 struct cl_fault_io *fio = &io->u.ci_fault;
1211 struct vvp_fault_io *cfio = &vio->u.fault;
1214 struct page *vmpage = NULL;
1215 struct cl_page *page;
1220 if (fio->ft_executable &&
1221 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
1223 " changed while waiting for the page fault lock\n",
1224 PFID(lu_object_fid(&obj->co_lu)));
1226 down_read(&lli->lli_trunc_sem);
1228 /* offset of the last byte on the page */
1229 offset = cl_offset(obj, fio->ft_index + 1) - 1;
1230 LASSERT(cl_index(obj, offset) == fio->ft_index);
1231 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1235 /* must return locked page */
1236 if (fio->ft_mkwrite) {
1237 LASSERT(cfio->ft_vmpage != NULL);
1238 lock_page(cfio->ft_vmpage);
1240 result = vvp_io_kernel_fault(cfio);
1245 vmpage = cfio->ft_vmpage;
1246 LASSERT(PageLocked(vmpage));
1248 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1249 ll_invalidate_page(vmpage);
1251 size = i_size_read(inode);
1252 /* Though we have already held a cl_lock upon this page, but
1253 * it still can be truncated locally. */
1254 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1255 (page_offset(vmpage) > size))) {
1256 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1258 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1260 GOTO(out, result = +1);
1263 last_index = cl_index(obj, size - 1);
1265 if (fio->ft_mkwrite ) {
1267 * Capture the size while holding the lli_trunc_sem from above
1268 * we want to make sure that we complete the mkwrite action
1269 * while holding this lock. We need to make sure that we are
1270 * not past the end of the file.
1272 if (last_index < fio->ft_index) {
1274 "llite: mkwrite and truncate race happened: "
1275 "%p: 0x%lx 0x%lx\n",
1276 vmpage->mapping,fio->ft_index,last_index);
1278 * We need to return if we are
1279 * passed the end of the file. This will propagate
1280 * up the call stack to ll_page_mkwrite where
1281 * we will return VM_FAULT_NOPAGE. Any non-negative
1282 * value returned here will be silently
1283 * converted to 0. If the vmpage->mapping is null
1284 * the error code would be converted back to ENODATA
1285 * in ll_page_mkwrite0. Thus we return -ENODATA
1286 * to handle both cases
1288 GOTO(out, result = -ENODATA);
1292 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1294 GOTO(out, result = PTR_ERR(page));
1296 /* if page is going to be written, we should add this page into cache
1298 if (fio->ft_mkwrite) {
1299 wait_on_page_writeback(vmpage);
1300 if (!PageDirty(vmpage)) {
1301 struct cl_page_list *plist = &io->ci_queue.c2_qin;
1302 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1305 /* vvp_page_assume() calls wait_on_page_writeback(). */
1306 cl_page_assume(env, io, page);
1308 cl_page_list_init(plist);
1309 cl_page_list_add(plist, page);
1312 if (last_index == vvp_index(vpg))
1313 to = size & ~PAGE_MASK;
1315 /* Do not set Dirty bit here so that in case IO is
1316 * started before the page is really made dirty, we
1317 * still have chance to detect it. */
1318 result = cl_io_commit_async(env, io, plist, 0, to,
1319 mkwrite_commit_callback);
1320 LASSERT(cl_page_is_owned(page, io));
1321 cl_page_list_fini(env, plist);
1325 cl_page_discard(env, io, page);
1326 cl_page_disown(env, io, page);
1328 cl_page_put(env, page);
1330 /* we're in big trouble, what can we do now? */
1331 if (result == -EDQUOT)
1335 cl_page_disown(env, io, page);
1340 * The ft_index is only used in the case of
1341 * a mkwrite action. We need to check
1342 * our assertions are correct, since
1343 * we should have caught this above
1345 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1346 if (fio->ft_index == last_index)
1348 * Last page is mapped partially.
1350 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1352 fio->ft_nob = cl_page_size(obj);
1354 lu_ref_add(&page->cp_reference, "fault", io);
1355 fio->ft_page = page;
1359 /* return unlocked vmpage to avoid deadlocking */
1361 unlock_page(vmpage);
1363 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1368 static void vvp_io_fault_end(const struct lu_env *env,
1369 const struct cl_io_slice *ios)
1371 struct inode *inode = vvp_object_inode(ios->cis_obj);
1372 struct ll_inode_info *lli = ll_i2info(inode);
1374 CLOBINVRNT(env, ios->cis_io->ci_obj,
1375 vvp_object_invariant(ios->cis_io->ci_obj));
1376 up_read(&lli->lli_trunc_sem);
1379 static int vvp_io_fsync_start(const struct lu_env *env,
1380 const struct cl_io_slice *ios)
1382 /* we should mark TOWRITE bit to each dirty page in radix tree to
1383 * verify pages have been written, but this is difficult because of
1388 static int vvp_io_read_ahead(const struct lu_env *env,
1389 const struct cl_io_slice *ios,
1390 pgoff_t start, struct cl_read_ahead *ra)
1395 if (ios->cis_io->ci_type == CIT_READ ||
1396 ios->cis_io->ci_type == CIT_FAULT) {
1397 struct vvp_io *vio = cl2vvp_io(env, ios);
1399 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1400 ra->cra_end = CL_PAGE_EOF;
1401 result = +1; /* no need to call down */
1408 static const struct cl_io_operations vvp_io_ops = {
1411 .cio_fini = vvp_io_fini,
1412 .cio_lock = vvp_io_read_lock,
1413 .cio_start = vvp_io_read_start,
1414 .cio_end = vvp_io_rw_end,
1415 .cio_advance = vvp_io_advance,
1418 .cio_fini = vvp_io_fini,
1419 .cio_iter_init = vvp_io_write_iter_init,
1420 .cio_iter_fini = vvp_io_write_iter_fini,
1421 .cio_lock = vvp_io_write_lock,
1422 .cio_start = vvp_io_write_start,
1423 .cio_end = vvp_io_rw_end,
1424 .cio_advance = vvp_io_advance,
1427 .cio_fini = vvp_io_setattr_fini,
1428 .cio_iter_init = vvp_io_setattr_iter_init,
1429 .cio_lock = vvp_io_setattr_lock,
1430 .cio_start = vvp_io_setattr_start,
1431 .cio_end = vvp_io_setattr_end
1434 .cio_fini = vvp_io_fault_fini,
1435 .cio_iter_init = vvp_io_fault_iter_init,
1436 .cio_lock = vvp_io_fault_lock,
1437 .cio_start = vvp_io_fault_start,
1438 .cio_end = vvp_io_fault_end,
1441 .cio_start = vvp_io_fsync_start,
1442 .cio_fini = vvp_io_fini
1445 .cio_fini = vvp_io_fini
1448 .cio_read_ahead = vvp_io_read_ahead
1451 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1454 struct vvp_io *vio = vvp_env_io(env);
1455 struct inode *inode = vvp_object_inode(obj);
1458 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1461 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1462 "restore needed %d\n",
1463 PFID(lu_object_fid(&obj->co_lu)),
1464 io->ci_ignore_layout, io->ci_verify_layout,
1465 vio->vui_layout_gen, io->ci_restore_needed);
1467 CL_IO_SLICE_CLEAN(vio, vui_cl);
1468 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1469 vio->vui_ra_valid = false;
1471 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1473 struct ll_inode_info *lli = ll_i2info(inode);
1475 count = io->u.ci_rw.crw_count;
1476 /* "If nbyte is 0, read() will return 0 and have no other
1477 * results." -- Single Unix Spec */
1481 vio->vui_tot_count = count;
1482 #ifndef HAVE_FILE_OPERATIONS_READ_WRITE_ITER
1483 vio->vui_tot_nrsegs = 0;
1487 /* for read/write, we store the jobid in the inode, and
1488 * it'll be fetched by osc when building RPC.
1490 * it's not accurate if the file is shared by different
1493 lustre_get_jobid(lli->lli_jobid);
1494 } else if (io->ci_type == CIT_SETATTR) {
1495 if (!cl_io_is_trunc(io))
1496 io->ci_lockreq = CILR_MANDATORY;
1499 /* ignore layout change for generic CIT_MISC but not for glimpse.
1500 * io context for glimpse must set ci_verify_layout to true,
1501 * see cl_glimpse_size0() for details. */
1502 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1503 io->ci_ignore_layout = 1;
1505 /* Enqueue layout lock and get layout version. We need to do this
1506 * even for operations requiring to open file, such as read and write,
1507 * because it might not grant layout lock in IT_OPEN. */
1508 if (result == 0 && !io->ci_ignore_layout) {
1509 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1510 if (result == -ENOENT)
1511 /* If the inode on MDS has been removed, but the objects
1512 * on OSTs haven't been destroyed (async unlink), layout
1513 * fetch will return -ENOENT, we'd ingore this error
1514 * and continue with dirty flush. LU-3230. */
1517 CERROR("%s: refresh file layout " DFID " error %d.\n",
1518 ll_get_fsname(inode->i_sb, NULL, 0),
1519 PFID(lu_object_fid(&obj->co_lu)), result);