4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * Implementation of cl_io for VVP layer.
34 * Author: Nikita Danilov <nikita.danilov@sun.com>
35 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
38 #define DEBUG_SUBSYSTEM S_LLITE
42 #include "llite_internal.h"
43 #include "vvp_internal.h"
44 #include <libcfs/linux/linux-misc.h>
46 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
47 const struct cl_io_slice *slice)
51 vio = container_of(slice, struct vvp_io, vui_cl);
52 LASSERT(vio == vvp_env_io(env));
58 * True, if \a io is a normal io, False for splice_{read,write}
60 static int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
62 struct vvp_io *vio = vvp_env_io(env);
64 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
66 return vio->vui_io_subtype == IO_NORMAL;
70 * For swapping layout. The file's layout may have changed.
71 * To avoid populating pages to a wrong stripe, we have to verify the
72 * correctness of layout. It works because swapping layout processes
73 * have to acquire group lock.
75 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
78 struct ll_inode_info *lli = ll_i2info(inode);
79 struct vvp_io *vio = vvp_env_io(env);
82 switch (io->ci_type) {
85 /* don't need lock here to check lli_layout_gen as we have held
86 * extent lock and GROUP lock has to hold to swap layout */
87 if (ll_layout_version_get(lli) != vio->vui_layout_gen ||
88 OBD_FAIL_CHECK_RESET(OBD_FAIL_LLITE_LOST_LAYOUT, 0)) {
89 io->ci_need_restart = 1;
90 /* this will cause a short read/write */
95 /* fault is okay because we've already had a page. */
103 static void vvp_object_size_lock(struct cl_object *obj)
105 struct inode *inode = vvp_object_inode(obj);
107 ll_inode_size_lock(inode);
108 cl_object_attr_lock(obj);
111 static void vvp_object_size_unlock(struct cl_object *obj)
113 struct inode *inode = vvp_object_inode(obj);
115 cl_object_attr_unlock(obj);
116 ll_inode_size_unlock(inode);
120 * Helper function that if necessary adjusts file size (inode->i_size), when
121 * position at the offset \a pos is accessed. File size can be arbitrary stale
122 * on a Lustre client, but client at least knows KMS. If accessed area is
123 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
125 * Locking: i_size_lock is used to serialize changes to inode size and to
126 * protect consistency between inode size and cl_object
127 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
128 * top-object and sub-objects.
130 static int vvp_prep_size(const struct lu_env *env, struct cl_object *obj,
131 struct cl_io *io, loff_t start, size_t count,
134 struct cl_attr *attr = vvp_env_thread_attr(env);
135 struct inode *inode = vvp_object_inode(obj);
136 loff_t pos = start + count - 1;
141 * Consistency guarantees: following possibilities exist for the
142 * relation between region being accessed and real file size at this
145 * (A): the region is completely inside of the file;
147 * (B-x): x bytes of region are inside of the file, the rest is
150 * (C): the region is completely outside of the file.
152 * This classification is stable under DLM lock already acquired by
153 * the caller, because to change the class, other client has to take
154 * DLM lock conflicting with our lock. Also, any updates to ->i_size
155 * by other threads on this client are serialized by
156 * ll_inode_size_lock(). This guarantees that short reads are handled
157 * correctly in the face of concurrent writes and truncates.
159 vvp_object_size_lock(obj);
160 result = cl_object_attr_get(env, obj, attr);
165 * A glimpse is necessary to determine whether we
166 * return a short read (B) or some zeroes at the end
169 vvp_object_size_unlock(obj);
170 result = cl_glimpse_lock(env, io, inode, obj, 0);
171 if (result == 0 && exceed != NULL) {
172 /* If objective page index exceed end-of-file
173 * page index, return directly. Do not expect
174 * kernel will check such case correctly.
175 * linux-2.6.18-128.1.1 miss to do that.
177 loff_t size = i_size_read(inode);
178 unsigned long cur_index = start >>
181 if ((size == 0 && cur_index != 0) ||
182 (((size - 1) >> PAGE_SHIFT) <
190 * region is within kms and, hence, within real file
191 * size (A). We need to increase i_size to cover the
192 * read region so that generic_file_read() will do its
193 * job, but that doesn't mean the kms size is
194 * _correct_, it is only the _minimum_ size. If
195 * someone does a stat they will get the correct size
196 * which will always be >= the kms value here.
199 if (i_size_read(inode) < kms) {
200 i_size_write(inode, kms);
202 DFID" updating i_size %llu\n",
203 PFID(lu_object_fid(&obj->co_lu)),
204 (__u64)i_size_read(inode));
209 vvp_object_size_unlock(obj);
214 /*****************************************************************************
220 static int vvp_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
221 __u32 enqflags, enum cl_lock_mode mode,
222 pgoff_t start, pgoff_t end)
224 struct vvp_io *vio = vvp_env_io(env);
225 struct cl_lock_descr *descr = &vio->vui_link.cill_descr;
226 struct cl_object *obj = io->ci_obj;
228 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
231 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
233 memset(&vio->vui_link, 0, sizeof vio->vui_link);
235 if (vio->vui_fd && (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
236 descr->cld_mode = CLM_GROUP;
237 descr->cld_gid = vio->vui_fd->fd_grouplock.lg_gid;
238 enqflags |= CEF_LOCK_MATCH;
240 descr->cld_mode = mode;
243 descr->cld_obj = obj;
244 descr->cld_start = start;
245 descr->cld_end = end;
246 descr->cld_enq_flags = enqflags;
248 cl_io_lock_add(env, io, &vio->vui_link);
253 static int vvp_io_one_lock(const struct lu_env *env, struct cl_io *io,
254 __u32 enqflags, enum cl_lock_mode mode,
255 loff_t start, loff_t end)
257 struct cl_object *obj = io->ci_obj;
259 return vvp_io_one_lock_index(env, io, enqflags, mode,
260 cl_index(obj, start), cl_index(obj, end));
263 static int vvp_io_write_iter_init(const struct lu_env *env,
264 const struct cl_io_slice *ios)
266 struct vvp_io *vio = cl2vvp_io(env, ios);
268 cl_page_list_init(&vio->u.write.vui_queue);
269 vio->u.write.vui_written = 0;
270 vio->u.write.vui_from = 0;
271 vio->u.write.vui_to = PAGE_SIZE;
276 static void vvp_io_write_iter_fini(const struct lu_env *env,
277 const struct cl_io_slice *ios)
279 struct vvp_io *vio = cl2vvp_io(env, ios);
281 LASSERT(vio->u.write.vui_queue.pl_nr == 0);
284 static int vvp_io_fault_iter_init(const struct lu_env *env,
285 const struct cl_io_slice *ios)
287 struct vvp_io *vio = cl2vvp_io(env, ios);
288 struct inode *inode = vvp_object_inode(ios->cis_obj);
290 LASSERT(inode == file_inode(vio->vui_fd->fd_file));
291 vio->u.fault.ft_mtime = inode->i_mtime.tv_sec;
296 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
298 struct cl_io *io = ios->cis_io;
299 struct cl_object *obj = io->ci_obj;
300 struct vvp_io *vio = cl2vvp_io(env, ios);
301 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 "need write layout %d, 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_need_write_intent,
313 io->ci_restore_needed);
315 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().
327 * Even if ll_layout_restore() returns zero, it doesn't mean
328 * that restore has been successful. Therefore it sets
329 * ci_verify_layout so that it will check layout at the end
333 io->ci_restore_needed = 1;
334 io->ci_need_restart = 0;
335 io->ci_verify_layout = 0;
340 io->ci_restore_needed = 0;
342 /* Even if ll_layout_restore() returns zero, it doesn't mean
343 * that restore has been successful. Therefore it should verify
344 * if there was layout change and restart I/O correspondingly.
346 ll_layout_refresh(inode, &gen);
347 io->ci_need_restart = vio->vui_layout_gen != gen;
348 if (io->ci_need_restart) {
350 DFID" layout changed from %d to %d.\n",
351 PFID(lu_object_fid(&obj->co_lu)),
352 vio->vui_layout_gen, gen);
353 /* today successful restore is the only possible
355 /* restore was done, clear restoring state */
356 ll_file_clear_flag(ll_i2info(vvp_object_inode(obj)),
357 LLIF_FILE_RESTORING);
363 * dynamic layout change needed, send layout intent
366 if (io->ci_need_write_intent) {
367 enum layout_intent_opc opc = LAYOUT_INTENT_WRITE;
369 io->ci_need_write_intent = 0;
371 LASSERT(io->ci_type == CIT_WRITE ||
372 cl_io_is_trunc(io) || cl_io_is_mkwrite(io));
374 CDEBUG(D_VFSTRACE, DFID" write layout, type %u "DEXT"\n",
375 PFID(lu_object_fid(&obj->co_lu)), io->ci_type,
376 PEXT(&io->ci_write_intent));
378 if (cl_io_is_trunc(io))
379 opc = LAYOUT_INTENT_TRUNC;
381 rc = ll_layout_write_intent(inode, opc, &io->ci_write_intent);
384 io->ci_need_restart = 1;
388 if (!io->ci_need_restart &&
389 !io->ci_ignore_layout && io->ci_verify_layout) {
390 /* check layout version */
391 ll_layout_refresh(inode, &gen);
392 io->ci_need_restart = vio->vui_layout_gen != gen;
393 if (io->ci_need_restart) {
395 DFID" layout changed from %d to %d.\n",
396 PFID(lu_object_fid(&obj->co_lu)),
397 vio->vui_layout_gen, gen);
405 static void vvp_io_fault_fini(const struct lu_env *env,
406 const struct cl_io_slice *ios)
408 struct cl_io *io = ios->cis_io;
409 struct cl_page *page = io->u.ci_fault.ft_page;
411 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
414 lu_ref_del(&page->cp_reference, "fault", io);
415 cl_page_put(env, page);
416 io->u.ci_fault.ft_page = NULL;
418 vvp_io_fini(env, ios);
421 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
424 * we only want to hold PW locks if the mmap() can generate
425 * writes back to the file and that only happens in shared
428 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
433 static int vvp_mmap_locks(const struct lu_env *env,
434 struct vvp_io *vio, struct cl_io *io)
436 struct vvp_thread_info *vti = vvp_env_info(env);
437 struct mm_struct *mm = current->mm;
438 struct vm_area_struct *vma;
439 struct cl_lock_descr *descr = &vti->vti_descr;
440 union ldlm_policy_data policy;
448 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
450 if (!cl_is_normalio(env, io))
453 /* nfs or loop back device write */
454 if (vio->vui_iter == NULL)
457 /* No MM (e.g. NFS)? No vmas too. */
461 if (!iter_is_iovec(vio->vui_iter) && !iov_iter_is_kvec(vio->vui_iter))
464 for (i = *vio->vui_iter;
466 iov_iter_advance(&i, iov.iov_len)) {
467 iov = iov_iter_iovec(&i);
468 addr = (unsigned long)iov.iov_base;
474 count += addr & ~PAGE_MASK;
477 down_read(&mm->mmap_sem);
478 while((vma = our_vma(mm, addr, count)) != NULL) {
479 struct dentry *de = file_dentry(vma->vm_file);
480 struct inode *inode = de->d_inode;
481 int flags = CEF_MUST;
483 if (ll_file_nolock(vma->vm_file)) {
485 * For no lock case is not allowed for mmap
492 * XXX: Required lock mode can be weakened: CIT_WRITE
493 * io only ever reads user level buffer, and CIT_READ
496 policy_from_vma(&policy, vma, addr, count);
497 descr->cld_mode = vvp_mode_from_vma(vma);
498 descr->cld_obj = ll_i2info(inode)->lli_clob;
499 descr->cld_start = cl_index(descr->cld_obj,
500 policy.l_extent.start);
501 descr->cld_end = cl_index(descr->cld_obj,
502 policy.l_extent.end);
503 descr->cld_enq_flags = flags;
504 result = cl_io_lock_alloc_add(env, io, descr);
506 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
507 descr->cld_mode, descr->cld_start,
513 if (vma->vm_end - addr >= count)
516 count -= vma->vm_end - addr;
519 up_read(&mm->mmap_sem);
526 static void vvp_io_advance(const struct lu_env *env,
527 const struct cl_io_slice *ios,
530 struct vvp_io *vio = cl2vvp_io(env, ios);
531 struct cl_io *io = ios->cis_io;
532 struct cl_object *obj = ios->cis_io->ci_obj;
534 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
536 if (!cl_is_normalio(env, io))
539 vio->vui_tot_count -= nob;
540 iov_iter_reexpand(vio->vui_iter, vio->vui_tot_count);
543 static void vvp_io_update_iov(const struct lu_env *env,
544 struct vvp_io *vio, struct cl_io *io)
546 size_t size = io->u.ci_rw.crw_count;
548 if (!cl_is_normalio(env, io) || vio->vui_iter == NULL)
551 iov_iter_truncate(vio->vui_iter, size);
554 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
555 enum cl_lock_mode mode, loff_t start, loff_t end)
557 struct vvp_io *vio = vvp_env_io(env);
561 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
564 vvp_io_update_iov(env, vio, io);
566 if (io->u.ci_rw.crw_nonblock)
567 ast_flags |= CEF_NONBLOCK;
568 if (io->ci_lock_no_expand)
569 ast_flags |= CEF_LOCK_NO_EXPAND;
571 result = vvp_mmap_locks(env, vio, io);
573 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
578 static int vvp_io_read_lock(const struct lu_env *env,
579 const struct cl_io_slice *ios)
581 struct cl_io *io = ios->cis_io;
582 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
586 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
587 rd->crw_pos + rd->crw_count - 1);
591 static int vvp_io_fault_lock(const struct lu_env *env,
592 const struct cl_io_slice *ios)
594 struct cl_io *io = ios->cis_io;
595 struct vvp_io *vio = cl2vvp_io(env, ios);
597 * XXX LDLM_FL_CBPENDING
599 return vvp_io_one_lock_index(env,
601 vvp_mode_from_vma(vio->u.fault.ft_vma),
602 io->u.ci_fault.ft_index,
603 io->u.ci_fault.ft_index);
606 static int vvp_io_write_lock(const struct lu_env *env,
607 const struct cl_io_slice *ios)
609 struct cl_io *io = ios->cis_io;
613 if (io->u.ci_wr.wr_append) {
615 end = OBD_OBJECT_EOF;
617 start = io->u.ci_wr.wr.crw_pos;
618 end = start + io->u.ci_wr.wr.crw_count - 1;
621 RETURN(vvp_io_rw_lock(env, io, CLM_WRITE, start, end));
624 static int vvp_io_setattr_iter_init(const struct lu_env *env,
625 const struct cl_io_slice *ios)
632 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
634 * Handles "lockless io" mode when extent locking is done by server.
636 static int vvp_io_setattr_lock(const struct lu_env *env,
637 const struct cl_io_slice *ios)
639 struct cl_io *io = ios->cis_io;
643 if (cl_io_is_trunc(io)) {
644 new_size = io->u.ci_setattr.sa_attr.lvb_size;
646 enqflags = CEF_DISCARD_DATA;
648 unsigned int valid = io->u.ci_setattr.sa_avalid;
650 if (!(valid & TIMES_SET_FLAGS))
653 if ((!(valid & ATTR_MTIME) ||
654 io->u.ci_setattr.sa_attr.lvb_mtime >=
655 io->u.ci_setattr.sa_attr.lvb_ctime) &&
656 (!(valid & ATTR_ATIME) ||
657 io->u.ci_setattr.sa_attr.lvb_atime >=
658 io->u.ci_setattr.sa_attr.lvb_ctime))
664 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
665 new_size, OBD_OBJECT_EOF);
668 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
673 * Only ll_inode_size_lock is taken at this level.
675 ll_inode_size_lock(inode);
676 result = inode_newsize_ok(inode, size);
678 ll_inode_size_unlock(inode);
681 i_size_write(inode, size);
683 ll_truncate_pagecache(inode, size);
684 ll_inode_size_unlock(inode);
688 static int vvp_io_setattr_time(const struct lu_env *env,
689 const struct cl_io_slice *ios)
691 struct cl_io *io = ios->cis_io;
692 struct cl_object *obj = io->ci_obj;
693 struct cl_attr *attr = vvp_env_thread_attr(env);
695 unsigned valid = CAT_CTIME;
697 cl_object_attr_lock(obj);
698 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
699 if (io->u.ci_setattr.sa_avalid & ATTR_ATIME_SET) {
700 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
703 if (io->u.ci_setattr.sa_avalid & ATTR_MTIME_SET) {
704 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
707 result = cl_object_attr_update(env, obj, attr, valid);
708 cl_object_attr_unlock(obj);
713 static int vvp_io_setattr_start(const struct lu_env *env,
714 const struct cl_io_slice *ios)
716 struct cl_io *io = ios->cis_io;
717 struct inode *inode = vvp_object_inode(io->ci_obj);
718 struct ll_inode_info *lli = ll_i2info(inode);
720 if (cl_io_is_trunc(io)) {
721 down_write(&lli->lli_trunc_sem);
723 inode_dio_wait(inode);
728 if (io->u.ci_setattr.sa_avalid & TIMES_SET_FLAGS)
729 return vvp_io_setattr_time(env, ios);
734 static void vvp_io_setattr_end(const struct lu_env *env,
735 const struct cl_io_slice *ios)
737 struct cl_io *io = ios->cis_io;
738 struct inode *inode = vvp_object_inode(io->ci_obj);
739 struct ll_inode_info *lli = ll_i2info(inode);
741 if (cl_io_is_trunc(io)) {
742 /* Truncate in memory pages - they must be clean pages
743 * because osc has already notified to destroy osc_extents. */
744 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
745 inode_dio_write_done(inode);
747 up_write(&lli->lli_trunc_sem);
753 static void vvp_io_setattr_fini(const struct lu_env *env,
754 const struct cl_io_slice *ios)
756 bool restore_needed = ios->cis_io->ci_restore_needed;
757 struct inode *inode = vvp_object_inode(ios->cis_obj);
759 vvp_io_fini(env, ios);
761 if (restore_needed && !ios->cis_io->ci_restore_needed) {
762 /* restore finished, set data modified flag for HSM */
763 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
767 static int vvp_io_read_start(const struct lu_env *env,
768 const struct cl_io_slice *ios)
770 struct vvp_io *vio = cl2vvp_io(env, ios);
771 struct cl_io *io = ios->cis_io;
772 struct cl_object *obj = io->ci_obj;
773 struct inode *inode = vvp_object_inode(obj);
774 struct ll_inode_info *lli = ll_i2info(inode);
775 struct file *file = vio->vui_fd->fd_file;
776 loff_t pos = io->u.ci_rd.rd.crw_pos;
777 long cnt = io->u.ci_rd.rd.crw_count;
778 long tot = vio->vui_tot_count;
783 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
785 CDEBUG(D_VFSTRACE, "%s: read [%llu, %llu)\n",
786 file_dentry(file)->d_name.name,
789 if (vio->vui_io_subtype == IO_NORMAL)
790 down_read(&lli->lli_trunc_sem);
792 if (!can_populate_pages(env, io, inode))
795 /* Unless this is reading a sparse file, otherwise the lock has already
796 * been acquired so vvp_prep_size() is an empty op. */
797 result = vvp_prep_size(env, obj, io, pos, cnt, &exceed);
800 else if (exceed != 0)
803 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
804 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
805 inode->i_ino, cnt, pos, i_size_read(inode));
807 /* turn off the kernel's read-ahead */
808 vio->vui_fd->fd_file->f_ra.ra_pages = 0;
810 /* initialize read-ahead window once per syscall */
811 if (!vio->vui_ra_valid) {
812 vio->vui_ra_valid = true;
813 vio->vui_ra_start = cl_index(obj, pos);
814 vio->vui_ra_count = cl_index(obj, tot + PAGE_SIZE - 1);
820 switch (vio->vui_io_subtype) {
822 LASSERT(vio->vui_iocb->ki_pos == pos);
823 result = generic_file_read_iter(vio->vui_iocb, vio->vui_iter);
826 result = generic_file_splice_read(file, &pos,
827 vio->u.splice.vui_pipe, cnt,
828 vio->u.splice.vui_flags);
829 /* LU-1109: do splice read stripe by stripe otherwise if it
830 * may make nfsd stuck if this read occupied all internal pipe
835 CERROR("Wrong IO type %u\n", vio->vui_io_subtype);
844 io->ci_nob += result;
851 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
852 struct cl_page_list *plist, int from, int to)
854 struct cl_2queue *queue = &io->ci_queue;
855 struct cl_page *page;
856 unsigned int bytes = 0;
860 if (plist->pl_nr == 0)
863 if (from > 0 || to != PAGE_SIZE) {
864 page = cl_page_list_first(plist);
865 if (plist->pl_nr == 1) {
866 cl_page_clip(env, page, from, to);
869 cl_page_clip(env, page, from, PAGE_SIZE);
870 if (to != PAGE_SIZE) {
871 page = cl_page_list_last(plist);
872 cl_page_clip(env, page, 0, to);
877 cl_2queue_init(queue);
878 cl_page_list_splice(plist, &queue->c2_qin);
879 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
881 /* plist is not sorted any more */
882 cl_page_list_splice(&queue->c2_qin, plist);
883 cl_page_list_splice(&queue->c2_qout, plist);
884 cl_2queue_fini(env, queue);
887 /* calculate bytes */
888 bytes = plist->pl_nr << PAGE_SHIFT;
889 bytes -= from + PAGE_SIZE - to;
891 while (plist->pl_nr > 0) {
892 page = cl_page_list_first(plist);
893 cl_page_list_del(env, plist, page);
895 cl_page_clip(env, page, 0, PAGE_SIZE);
897 SetPageUptodate(cl_page_vmpage(page));
898 cl_page_disown(env, io, page);
900 /* held in ll_cl_init() */
901 lu_ref_del(&page->cp_reference, "cl_io", io);
902 cl_page_put(env, page);
906 RETURN(bytes > 0 ? bytes : rc);
909 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
910 struct cl_page *page)
912 struct page *vmpage = page->cp_vmpage;
914 SetPageUptodate(vmpage);
915 set_page_dirty(vmpage);
917 cl_page_disown(env, io, page);
919 /* held in ll_cl_init() */
920 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
921 cl_page_put(env, page);
924 /* make sure the page list is contiguous */
925 static bool page_list_sanity_check(struct cl_object *obj,
926 struct cl_page_list *plist)
928 struct cl_page *page;
929 pgoff_t index = CL_PAGE_EOF;
931 cl_page_list_for_each(page, plist) {
932 struct vvp_page *vpg = cl_object_page_slice(obj, page);
934 if (index == CL_PAGE_EOF) {
935 index = vvp_index(vpg);
940 if (index == vvp_index(vpg))
948 /* Return how many bytes have queued or written */
949 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
951 struct cl_object *obj = io->ci_obj;
952 struct inode *inode = vvp_object_inode(obj);
953 struct vvp_io *vio = vvp_env_io(env);
954 struct cl_page_list *queue = &vio->u.write.vui_queue;
955 struct cl_page *page;
958 unsigned int npages = vio->u.write.vui_queue.pl_nr;
964 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
965 npages, vio->u.write.vui_from, vio->u.write.vui_to);
967 LASSERT(page_list_sanity_check(obj, queue));
969 /* submit IO with async write */
970 rc = cl_io_commit_async(env, io, queue,
971 vio->u.write.vui_from, vio->u.write.vui_to,
972 write_commit_callback);
973 npages -= queue->pl_nr; /* already committed pages */
975 /* calculate how many bytes were written */
976 bytes = npages << PAGE_SHIFT;
979 bytes -= vio->u.write.vui_from;
980 if (queue->pl_nr == 0) /* last page */
981 bytes -= PAGE_SIZE - vio->u.write.vui_to;
982 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
984 vio->u.write.vui_written += bytes;
986 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
987 npages, bytes, vio->u.write.vui_written);
989 /* the first page must have been written. */
990 vio->u.write.vui_from = 0;
992 LASSERT(page_list_sanity_check(obj, queue));
993 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
995 /* out of quota, try sync write */
996 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
997 rc = vvp_io_commit_sync(env, io, queue,
998 vio->u.write.vui_from,
999 vio->u.write.vui_to);
1001 vio->u.write.vui_written += rc;
1006 /* update inode size */
1007 ll_merge_attr(env, inode);
1009 /* Now the pages in queue were failed to commit, discard them
1010 * unless they were dirtied before. */
1011 while (queue->pl_nr > 0) {
1012 page = cl_page_list_first(queue);
1013 cl_page_list_del(env, queue, page);
1015 if (!PageDirty(cl_page_vmpage(page)))
1016 cl_page_discard(env, io, page);
1018 cl_page_disown(env, io, page);
1020 /* held in ll_cl_init() */
1021 lu_ref_del(&page->cp_reference, "cl_io", io);
1022 cl_page_put(env, page);
1024 cl_page_list_fini(env, queue);
1029 static int vvp_io_write_start(const struct lu_env *env,
1030 const struct cl_io_slice *ios)
1032 struct vvp_io *vio = cl2vvp_io(env, ios);
1033 struct cl_io *io = ios->cis_io;
1034 struct cl_object *obj = io->ci_obj;
1035 struct inode *inode = vvp_object_inode(obj);
1036 struct ll_inode_info *lli = ll_i2info(inode);
1037 struct file *file = vio->vui_fd->fd_file;
1039 loff_t pos = io->u.ci_wr.wr.crw_pos;
1040 size_t cnt = io->u.ci_wr.wr.crw_count;
1041 bool lock_inode = !IS_NOSEC(inode);
1042 size_t nob = io->ci_nob;
1043 struct iov_iter iter;
1048 if (vio->vui_io_subtype == IO_NORMAL)
1049 down_read(&lli->lli_trunc_sem);
1051 if (!can_populate_pages(env, io, inode))
1054 if (cl_io_is_append(io)) {
1056 * PARALLEL IO This has to be changed for parallel IO doing
1057 * out-of-order writes.
1059 ll_merge_attr(env, inode);
1060 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
1061 vio->vui_iocb->ki_pos = pos;
1063 LASSERTF(vio->vui_iocb->ki_pos == pos,
1064 "ki_pos %lld [%lld, %lld)\n",
1065 vio->vui_iocb->ki_pos,
1069 CDEBUG(D_VFSTRACE, "%s: write [%llu, %llu)\n",
1070 file_dentry(file)->d_name.name,
1073 /* The maximum Lustre file size is variable, based on the OST maximum
1074 * object size and number of stripes. This needs another check in
1075 * addition to the VFS checks earlier. */
1076 if (pos + cnt > ll_file_maxbytes(inode)) {
1078 "%s: file %s ("DFID") offset %llu > maxbytes %llu\n",
1079 ll_get_fsname(inode->i_sb, NULL, 0),
1080 file_dentry(file)->d_name.name,
1081 PFID(ll_inode2fid(inode)), pos + cnt,
1082 ll_file_maxbytes(inode));
1086 /* Tests to verify we take the i_mutex correctly */
1087 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_SEC) && !lock_inode)
1090 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_NOSEC) && lock_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 lock_inode = !IS_NOSEC(inode);
1106 iter = *vio->vui_iter;
1108 if (unlikely(lock_inode))
1110 result = __generic_file_write_iter(vio->vui_iocb,
1112 if (unlikely(lock_inode))
1113 inode_unlock(inode);
1116 if (result > 0 || result == -EIOCBQUEUED)
1117 #ifdef HAVE_GENERIC_WRITE_SYNC_2ARGS
1118 result = generic_write_sync(vio->vui_iocb, result);
1123 err = generic_write_sync(vio->vui_iocb->ki_filp, pos,
1125 if (err < 0 && result > 0)
1132 result = vvp_io_write_commit(env, io);
1133 /* Simulate short commit */
1134 if (CFS_FAULT_CHECK(OBD_FAIL_LLITE_SHORT_COMMIT)) {
1135 vio->u.write.vui_written >>= 1;
1136 if (vio->u.write.vui_written > 0)
1137 io->ci_need_restart = 1;
1139 if (vio->u.write.vui_written > 0) {
1140 result = vio->u.write.vui_written;
1141 CDEBUG(D_VFSTRACE, "%s: write nob %zd, result: %zd\n",
1142 file_dentry(file)->d_name.name,
1143 io->ci_nob, result);
1144 io->ci_nob += result;
1146 io->ci_continue = 0;
1149 if (vio->vui_iocb->ki_pos != (pos + io->ci_nob - nob)) {
1150 CDEBUG(D_VFSTRACE, "%s: write position mismatch: "
1151 "ki_pos %lld vs. pos %lld, written %ld, commit %ld "
1153 file_dentry(file)->d_name.name,
1154 vio->vui_iocb->ki_pos, pos + io->ci_nob - nob,
1155 written, io->ci_nob - nob, result);
1157 * Rewind ki_pos and vui_iter to where it has
1158 * successfully committed.
1160 vio->vui_iocb->ki_pos = pos + io->ci_nob - nob;
1161 iov_iter_advance(&iter, io->ci_nob - nob);
1162 vio->vui_iter->iov = iter.iov;
1163 vio->vui_iter->nr_segs = iter.nr_segs;
1164 vio->vui_iter->iov_offset = iter.iov_offset;
1165 vio->vui_iter->count = iter.count;
1168 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
1171 io->ci_continue = 0;
1178 static void vvp_io_rw_end(const struct lu_env *env,
1179 const struct cl_io_slice *ios)
1181 struct vvp_io *vio = cl2vvp_io(env, ios);
1182 struct inode *inode = vvp_object_inode(ios->cis_obj);
1183 struct ll_inode_info *lli = ll_i2info(inode);
1185 if (vio->vui_io_subtype == IO_NORMAL)
1186 up_read(&lli->lli_trunc_sem);
1189 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
1191 struct vm_fault *vmf = cfio->ft_vmf;
1193 cfio->ft_flags = ll_filemap_fault(cfio->ft_vma, vmf);
1194 cfio->ft_flags_valid = 1;
1197 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
1198 get_vmf_address(vmf));
1199 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
1200 lock_page(vmf->page);
1201 cfio->ft_flags |= VM_FAULT_LOCKED;
1204 cfio->ft_vmpage = vmf->page;
1209 if (cfio->ft_flags & VM_FAULT_SIGBUS) {
1210 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", get_vmf_address(vmf));
1214 if (cfio->ft_flags & VM_FAULT_OOM) {
1215 CDEBUG(D_PAGE, "got addr %p - OOM\n", get_vmf_address(vmf));
1219 if (cfio->ft_flags & VM_FAULT_RETRY)
1222 CERROR("unknown error in page fault %d\n", cfio->ft_flags);
1227 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1228 struct cl_page *page)
1230 set_page_dirty(page->cp_vmpage);
1233 static int vvp_io_fault_start(const struct lu_env *env,
1234 const struct cl_io_slice *ios)
1236 struct vvp_io *vio = cl2vvp_io(env, ios);
1237 struct cl_io *io = ios->cis_io;
1238 struct cl_object *obj = io->ci_obj;
1239 struct inode *inode = vvp_object_inode(obj);
1240 struct ll_inode_info *lli = ll_i2info(inode);
1241 struct cl_fault_io *fio = &io->u.ci_fault;
1242 struct vvp_fault_io *cfio = &vio->u.fault;
1245 struct page *vmpage = NULL;
1246 struct cl_page *page;
1251 down_read(&lli->lli_trunc_sem);
1253 /* offset of the last byte on the page */
1254 offset = cl_offset(obj, fio->ft_index + 1) - 1;
1255 LASSERT(cl_index(obj, offset) == fio->ft_index);
1256 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1260 /* must return locked page */
1261 if (fio->ft_mkwrite) {
1262 LASSERT(cfio->ft_vmpage != NULL);
1263 lock_page(cfio->ft_vmpage);
1265 result = vvp_io_kernel_fault(cfio);
1270 vmpage = cfio->ft_vmpage;
1271 LASSERT(PageLocked(vmpage));
1273 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1274 ll_invalidate_page(vmpage);
1276 size = i_size_read(inode);
1277 /* Though we have already held a cl_lock upon this page, but
1278 * it still can be truncated locally. */
1279 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1280 (page_offset(vmpage) > size))) {
1281 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1283 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1285 GOTO(out, result = +1);
1288 last_index = cl_index(obj, size - 1);
1290 if (fio->ft_mkwrite ) {
1292 * Capture the size while holding the lli_trunc_sem from above
1293 * we want to make sure that we complete the mkwrite action
1294 * while holding this lock. We need to make sure that we are
1295 * not past the end of the file.
1297 if (last_index < fio->ft_index) {
1299 "llite: mkwrite and truncate race happened: "
1300 "%p: 0x%lx 0x%lx\n",
1301 vmpage->mapping,fio->ft_index,last_index);
1303 * We need to return if we are
1304 * passed the end of the file. This will propagate
1305 * up the call stack to ll_page_mkwrite where
1306 * we will return VM_FAULT_NOPAGE. Any non-negative
1307 * value returned here will be silently
1308 * converted to 0. If the vmpage->mapping is null
1309 * the error code would be converted back to ENODATA
1310 * in ll_page_mkwrite0. Thus we return -ENODATA
1311 * to handle both cases
1313 GOTO(out, result = -ENODATA);
1317 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1319 GOTO(out, result = PTR_ERR(page));
1321 /* if page is going to be written, we should add this page into cache
1323 if (fio->ft_mkwrite) {
1324 wait_on_page_writeback(vmpage);
1325 if (!PageDirty(vmpage)) {
1326 struct cl_page_list *plist = &vio->u.fault.ft_queue;
1327 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1330 /* vvp_page_assume() calls wait_on_page_writeback(). */
1331 cl_page_assume(env, io, page);
1333 cl_page_list_init(plist);
1334 cl_page_list_add(plist, page);
1337 if (last_index == vvp_index(vpg))
1338 to = ((size - 1) & ~PAGE_MASK) + 1;
1340 /* Do not set Dirty bit here so that in case IO is
1341 * started before the page is really made dirty, we
1342 * still have chance to detect it. */
1343 result = cl_io_commit_async(env, io, plist, 0, to,
1344 mkwrite_commit_callback);
1345 /* Have overquota flag, trying sync write to check
1346 * whether indeed out of quota */
1347 if (result == -EDQUOT) {
1349 result = vvp_io_commit_sync(env, io,
1353 cl_page_own(env, io, page);
1354 cl_page_list_add(plist, page);
1355 lu_ref_add(&page->cp_reference,
1357 result = cl_io_commit_async(env, io,
1359 mkwrite_commit_callback);
1362 cl_page_put(env, page);
1366 LASSERT(cl_page_is_owned(page, io));
1367 cl_page_list_fini(env, plist);
1371 cl_page_discard(env, io, page);
1372 cl_page_disown(env, io, page);
1374 cl_page_put(env, page);
1376 /* we're in big trouble, what can we do now? */
1377 if (result == -EDQUOT)
1381 cl_page_disown(env, io, page);
1387 * The ft_index is only used in the case of
1388 * a mkwrite action. We need to check
1389 * our assertions are correct, since
1390 * we should have caught this above
1392 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1393 if (fio->ft_index == last_index)
1395 * Last page is mapped partially.
1397 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1399 fio->ft_nob = cl_page_size(obj);
1401 lu_ref_add(&page->cp_reference, "fault", io);
1402 fio->ft_page = page;
1406 /* return unlocked vmpage to avoid deadlocking */
1408 unlock_page(vmpage);
1410 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1415 static void vvp_io_fault_end(const struct lu_env *env,
1416 const struct cl_io_slice *ios)
1418 struct inode *inode = vvp_object_inode(ios->cis_obj);
1419 struct ll_inode_info *lli = ll_i2info(inode);
1421 CLOBINVRNT(env, ios->cis_io->ci_obj,
1422 vvp_object_invariant(ios->cis_io->ci_obj));
1423 up_read(&lli->lli_trunc_sem);
1426 static int vvp_io_fsync_start(const struct lu_env *env,
1427 const struct cl_io_slice *ios)
1429 /* we should mark TOWRITE bit to each dirty page in radix tree to
1430 * verify pages have been written, but this is difficult because of
1435 static int vvp_io_read_ahead(const struct lu_env *env,
1436 const struct cl_io_slice *ios,
1437 pgoff_t start, struct cl_read_ahead *ra)
1442 if (ios->cis_io->ci_type == CIT_READ ||
1443 ios->cis_io->ci_type == CIT_FAULT) {
1444 struct vvp_io *vio = cl2vvp_io(env, ios);
1446 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1447 ra->cra_end = CL_PAGE_EOF;
1448 result = +1; /* no need to call down */
1455 static const struct cl_io_operations vvp_io_ops = {
1458 .cio_fini = vvp_io_fini,
1459 .cio_lock = vvp_io_read_lock,
1460 .cio_start = vvp_io_read_start,
1461 .cio_end = vvp_io_rw_end,
1462 .cio_advance = vvp_io_advance,
1465 .cio_fini = vvp_io_fini,
1466 .cio_iter_init = vvp_io_write_iter_init,
1467 .cio_iter_fini = vvp_io_write_iter_fini,
1468 .cio_lock = vvp_io_write_lock,
1469 .cio_start = vvp_io_write_start,
1470 .cio_end = vvp_io_rw_end,
1471 .cio_advance = vvp_io_advance,
1474 .cio_fini = vvp_io_setattr_fini,
1475 .cio_iter_init = vvp_io_setattr_iter_init,
1476 .cio_lock = vvp_io_setattr_lock,
1477 .cio_start = vvp_io_setattr_start,
1478 .cio_end = vvp_io_setattr_end
1481 .cio_fini = vvp_io_fault_fini,
1482 .cio_iter_init = vvp_io_fault_iter_init,
1483 .cio_lock = vvp_io_fault_lock,
1484 .cio_start = vvp_io_fault_start,
1485 .cio_end = vvp_io_fault_end,
1488 .cio_start = vvp_io_fsync_start,
1489 .cio_fini = vvp_io_fini
1492 .cio_fini = vvp_io_fini
1495 .cio_fini = vvp_io_fini
1498 .cio_fini = vvp_io_fini
1501 .cio_read_ahead = vvp_io_read_ahead
1504 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1507 struct vvp_io *vio = vvp_env_io(env);
1508 struct inode *inode = vvp_object_inode(obj);
1511 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1514 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1515 "restore needed %d\n",
1516 PFID(lu_object_fid(&obj->co_lu)),
1517 io->ci_ignore_layout, io->ci_verify_layout,
1518 vio->vui_layout_gen, io->ci_restore_needed);
1520 CL_IO_SLICE_CLEAN(vio, vui_cl);
1521 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1522 vio->vui_ra_valid = false;
1524 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1526 struct ll_inode_info *lli = ll_i2info(inode);
1528 count = io->u.ci_rw.crw_count;
1529 /* "If nbyte is 0, read() will return 0 and have no other
1530 * results." -- Single Unix Spec */
1534 vio->vui_tot_count = count;
1536 /* for read/write, we store the jobid in the inode, and
1537 * it'll be fetched by osc when building RPC.
1539 * it's not accurate if the file is shared by different
1542 lustre_get_jobid(lli->lli_jobid, sizeof(lli->lli_jobid));
1543 } else if (io->ci_type == CIT_SETATTR) {
1544 if (!cl_io_is_trunc(io))
1545 io->ci_lockreq = CILR_MANDATORY;
1548 /* Enqueue layout lock and get layout version. We need to do this
1549 * even for operations requiring to open file, such as read and write,
1550 * because it might not grant layout lock in IT_OPEN. */
1551 if (result == 0 && !io->ci_ignore_layout) {
1552 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1553 if (result == -ENOENT)
1554 /* If the inode on MDS has been removed, but the objects
1555 * on OSTs haven't been destroyed (async unlink), layout
1556 * fetch will return -ENOENT, we'd ingore this error
1557 * and continue with dirty flush. LU-3230. */
1560 CERROR("%s: refresh file layout " DFID " error %d.\n",
1561 ll_get_fsname(inode->i_sb, NULL, 0),
1562 PFID(lu_object_fid(&obj->co_lu)), result);
1565 io->ci_result = result < 0 ? result : 0;