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, 2015, 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
40 #include "../include/obd.h"
42 #include "llite_internal.h"
43 #include "vvp_internal.h"
45 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
46 const struct cl_io_slice *slice)
50 vio = container_of(slice, struct vvp_io, vui_cl);
51 LASSERT(vio == vvp_env_io(env));
57 * For swapping layout. The file's layout may have changed.
58 * To avoid populating pages to a wrong stripe, we have to verify the
59 * correctness of layout. It works because swapping layout processes
60 * have to acquire group lock.
62 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
65 struct ll_inode_info *lli = ll_i2info(inode);
66 struct vvp_io *vio = vvp_env_io(env);
69 switch (io->ci_type) {
72 /* don't need lock here to check lli_layout_gen as we have held
73 * extent lock and GROUP lock has to hold to swap layout
75 if (ll_layout_version_get(lli) != vio->vui_layout_gen) {
76 io->ci_need_restart = 1;
77 /* this will return application a short read/write */
82 /* fault is okay because we've already had a page. */
90 static void vvp_object_size_lock(struct cl_object *obj)
92 struct inode *inode = vvp_object_inode(obj);
94 ll_inode_size_lock(inode);
95 cl_object_attr_lock(obj);
98 static void vvp_object_size_unlock(struct cl_object *obj)
100 struct inode *inode = vvp_object_inode(obj);
102 cl_object_attr_unlock(obj);
103 ll_inode_size_unlock(inode);
107 * Helper function that if necessary adjusts file size (inode->i_size), when
108 * position at the offset \a pos is accessed. File size can be arbitrary stale
109 * on a Lustre client, but client at least knows KMS. If accessed area is
110 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
112 * Locking: cl_isize_lock is used to serialize changes to inode size and to
113 * protect consistency between inode size and cl_object
114 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
115 * top-object and sub-objects.
117 static int vvp_prep_size(const struct lu_env *env, struct cl_object *obj,
118 struct cl_io *io, loff_t start, size_t count,
121 struct cl_attr *attr = vvp_env_thread_attr(env);
122 struct inode *inode = vvp_object_inode(obj);
123 loff_t pos = start + count - 1;
128 * Consistency guarantees: following possibilities exist for the
129 * relation between region being accessed and real file size at this
132 * (A): the region is completely inside of the file;
134 * (B-x): x bytes of region are inside of the file, the rest is
137 * (C): the region is completely outside of the file.
139 * This classification is stable under DLM lock already acquired by
140 * the caller, because to change the class, other client has to take
141 * DLM lock conflicting with our lock. Also, any updates to ->i_size
142 * by other threads on this client are serialized by
143 * ll_inode_size_lock(). This guarantees that short reads are handled
144 * correctly in the face of concurrent writes and truncates.
146 vvp_object_size_lock(obj);
147 result = cl_object_attr_get(env, obj, attr);
152 * A glimpse is necessary to determine whether we
153 * return a short read (B) or some zeroes at the end
156 vvp_object_size_unlock(obj);
157 result = cl_glimpse_lock(env, io, inode, obj, 0);
158 if (result == 0 && exceed) {
159 /* If objective page index exceed end-of-file
160 * page index, return directly. Do not expect
161 * kernel will check such case correctly.
162 * linux-2.6.18-128.1.1 miss to do that.
165 loff_t size = i_size_read(inode);
166 loff_t cur_index = start >> PAGE_SHIFT;
167 loff_t size_index = (size - 1) >> PAGE_SHIFT;
169 if ((size == 0 && cur_index != 0) ||
170 size_index < cur_index)
176 * region is within kms and, hence, within real file
177 * size (A). We need to increase i_size to cover the
178 * read region so that generic_file_read() will do its
179 * job, but that doesn't mean the kms size is
180 * _correct_, it is only the _minimum_ size. If
181 * someone does a stat they will get the correct size
182 * which will always be >= the kms value here.
185 if (i_size_read(inode) < kms) {
186 i_size_write(inode, kms);
187 CDEBUG(D_VFSTRACE, DFID " updating i_size %llu\n",
188 PFID(lu_object_fid(&obj->co_lu)),
189 (__u64)i_size_read(inode));
193 vvp_object_size_unlock(obj);
198 /*****************************************************************************
204 static int vvp_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
205 __u32 enqflags, enum cl_lock_mode mode,
206 pgoff_t start, pgoff_t end)
208 struct vvp_io *vio = vvp_env_io(env);
209 struct cl_lock_descr *descr = &vio->vui_link.cill_descr;
210 struct cl_object *obj = io->ci_obj;
212 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
214 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
216 memset(&vio->vui_link, 0, sizeof(vio->vui_link));
218 if (vio->vui_fd && (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
219 descr->cld_mode = CLM_GROUP;
220 descr->cld_gid = vio->vui_fd->fd_grouplock.lg_gid;
222 descr->cld_mode = mode;
224 descr->cld_obj = obj;
225 descr->cld_start = start;
226 descr->cld_end = end;
227 descr->cld_enq_flags = enqflags;
229 cl_io_lock_add(env, io, &vio->vui_link);
233 static int vvp_io_one_lock(const struct lu_env *env, struct cl_io *io,
234 __u32 enqflags, enum cl_lock_mode mode,
235 loff_t start, loff_t end)
237 struct cl_object *obj = io->ci_obj;
239 return vvp_io_one_lock_index(env, io, enqflags, mode,
240 cl_index(obj, start), cl_index(obj, end));
243 static int vvp_io_write_iter_init(const struct lu_env *env,
244 const struct cl_io_slice *ios)
246 struct vvp_io *vio = cl2vvp_io(env, ios);
248 cl_page_list_init(&vio->u.write.vui_queue);
249 vio->u.write.vui_written = 0;
250 vio->u.write.vui_from = 0;
251 vio->u.write.vui_to = PAGE_SIZE;
256 static void vvp_io_write_iter_fini(const struct lu_env *env,
257 const struct cl_io_slice *ios)
259 struct vvp_io *vio = cl2vvp_io(env, ios);
261 LASSERT(vio->u.write.vui_queue.pl_nr == 0);
264 static int vvp_io_fault_iter_init(const struct lu_env *env,
265 const struct cl_io_slice *ios)
267 struct vvp_io *vio = cl2vvp_io(env, ios);
268 struct inode *inode = vvp_object_inode(ios->cis_obj);
270 LASSERT(inode == file_inode(vio->vui_fd->fd_file));
271 vio->u.fault.ft_mtime = inode->i_mtime.tv_sec;
275 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
277 struct cl_io *io = ios->cis_io;
278 struct cl_object *obj = io->ci_obj;
279 struct vvp_io *vio = cl2vvp_io(env, ios);
280 struct inode *inode = vvp_object_inode(obj);
282 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
284 CDEBUG(D_VFSTRACE, DFID
285 " ignore/verify layout %d/%d, layout version %d restore needed %d\n",
286 PFID(lu_object_fid(&obj->co_lu)),
287 io->ci_ignore_layout, io->ci_verify_layout,
288 vio->vui_layout_gen, io->ci_restore_needed);
290 if (io->ci_restore_needed == 1) {
293 /* file was detected release, we need to restore it
294 * before finishing the io
296 rc = ll_layout_restore(inode, 0, OBD_OBJECT_EOF);
297 /* if restore registration failed, no restart,
298 * we will return -ENODATA
300 /* The layout will change after restore, so we need to
301 * block on layout lock hold by the MDT
302 * as MDT will not send new layout in lvb (see LU-3124)
303 * we have to explicitly fetch it, all this will be done
304 * by ll_layout_refresh()
307 io->ci_restore_needed = 0;
308 io->ci_need_restart = 1;
309 io->ci_verify_layout = 1;
311 io->ci_restore_needed = 1;
312 io->ci_need_restart = 0;
313 io->ci_verify_layout = 0;
318 if (!io->ci_ignore_layout && io->ci_verify_layout) {
321 /* check layout version */
322 ll_layout_refresh(inode, &gen);
323 io->ci_need_restart = vio->vui_layout_gen != gen;
324 if (io->ci_need_restart) {
326 DFID" layout changed from %d to %d.\n",
327 PFID(lu_object_fid(&obj->co_lu)),
328 vio->vui_layout_gen, gen);
329 /* today successful restore is the only possible case */
330 /* restore was done, clear restoring state */
331 ll_i2info(vvp_object_inode(obj))->lli_flags &=
332 ~LLIF_FILE_RESTORING;
337 static void vvp_io_fault_fini(const struct lu_env *env,
338 const struct cl_io_slice *ios)
340 struct cl_io *io = ios->cis_io;
341 struct cl_page *page = io->u.ci_fault.ft_page;
343 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
346 lu_ref_del(&page->cp_reference, "fault", io);
347 cl_page_put(env, page);
348 io->u.ci_fault.ft_page = NULL;
350 vvp_io_fini(env, ios);
353 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
356 * we only want to hold PW locks if the mmap() can generate
357 * writes back to the file and that only happens in shared
360 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
365 static int vvp_mmap_locks(const struct lu_env *env,
366 struct vvp_io *vio, struct cl_io *io)
368 struct vvp_thread_info *cti = vvp_env_info(env);
369 struct mm_struct *mm = current->mm;
370 struct vm_area_struct *vma;
371 struct cl_lock_descr *descr = &cti->vti_descr;
372 ldlm_policy_data_t policy;
379 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
381 if (!vio->vui_iter) /* nfs or loop back device write */
384 /* No MM (e.g. NFS)? No vmas too. */
388 iov_for_each(iov, i, *vio->vui_iter) {
389 addr = (unsigned long)iov.iov_base;
394 count += addr & (~PAGE_MASK);
397 down_read(&mm->mmap_sem);
398 while ((vma = our_vma(mm, addr, count)) != NULL) {
399 struct inode *inode = file_inode(vma->vm_file);
400 int flags = CEF_MUST;
402 if (ll_file_nolock(vma->vm_file)) {
404 * For no lock case is not allowed for mmap
411 * XXX: Required lock mode can be weakened: CIT_WRITE
412 * io only ever reads user level buffer, and CIT_READ
415 policy_from_vma(&policy, vma, addr, count);
416 descr->cld_mode = vvp_mode_from_vma(vma);
417 descr->cld_obj = ll_i2info(inode)->lli_clob;
418 descr->cld_start = cl_index(descr->cld_obj,
419 policy.l_extent.start);
420 descr->cld_end = cl_index(descr->cld_obj,
421 policy.l_extent.end);
422 descr->cld_enq_flags = flags;
423 result = cl_io_lock_alloc_add(env, io, descr);
425 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
426 descr->cld_mode, descr->cld_start,
432 if (vma->vm_end - addr >= count)
435 count -= vma->vm_end - addr;
438 up_read(&mm->mmap_sem);
445 static void vvp_io_advance(const struct lu_env *env,
446 const struct cl_io_slice *ios,
449 struct cl_object *obj = ios->cis_io->ci_obj;
450 struct vvp_io *vio = cl2vvp_io(env, ios);
451 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
453 iov_iter_reexpand(vio->vui_iter, vio->vui_tot_count -= nob);
456 static void vvp_io_update_iov(const struct lu_env *env,
457 struct vvp_io *vio, struct cl_io *io)
459 size_t size = io->u.ci_rw.crw_count;
464 iov_iter_truncate(vio->vui_iter, size);
467 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
468 enum cl_lock_mode mode, loff_t start, loff_t end)
470 struct vvp_io *vio = vvp_env_io(env);
474 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
476 vvp_io_update_iov(env, vio, io);
478 if (io->u.ci_rw.crw_nonblock)
479 ast_flags |= CEF_NONBLOCK;
480 result = vvp_mmap_locks(env, vio, io);
482 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
486 static int vvp_io_read_lock(const struct lu_env *env,
487 const struct cl_io_slice *ios)
489 struct cl_io *io = ios->cis_io;
490 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
493 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
494 rd->crw_pos + rd->crw_count - 1);
499 static int vvp_io_fault_lock(const struct lu_env *env,
500 const struct cl_io_slice *ios)
502 struct cl_io *io = ios->cis_io;
503 struct vvp_io *vio = cl2vvp_io(env, ios);
505 * XXX LDLM_FL_CBPENDING
507 return vvp_io_one_lock_index(env,
509 vvp_mode_from_vma(vio->u.fault.ft_vma),
510 io->u.ci_fault.ft_index,
511 io->u.ci_fault.ft_index);
514 static int vvp_io_write_lock(const struct lu_env *env,
515 const struct cl_io_slice *ios)
517 struct cl_io *io = ios->cis_io;
521 if (io->u.ci_wr.wr_append) {
523 end = OBD_OBJECT_EOF;
525 start = io->u.ci_wr.wr.crw_pos;
526 end = start + io->u.ci_wr.wr.crw_count - 1;
528 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
531 static int vvp_io_setattr_iter_init(const struct lu_env *env,
532 const struct cl_io_slice *ios)
538 * Implementation of cl_io_operations::vio_lock() method for CIT_SETATTR io.
540 * Handles "lockless io" mode when extent locking is done by server.
542 static int vvp_io_setattr_lock(const struct lu_env *env,
543 const struct cl_io_slice *ios)
545 struct cl_io *io = ios->cis_io;
549 if (cl_io_is_trunc(io)) {
550 new_size = io->u.ci_setattr.sa_attr.lvb_size;
552 enqflags = CEF_DISCARD_DATA;
554 if ((io->u.ci_setattr.sa_attr.lvb_mtime >=
555 io->u.ci_setattr.sa_attr.lvb_ctime) ||
556 (io->u.ci_setattr.sa_attr.lvb_atime >=
557 io->u.ci_setattr.sa_attr.lvb_ctime))
562 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
563 new_size, OBD_OBJECT_EOF);
566 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
570 * Only ll_inode_size_lock is taken at this level.
572 ll_inode_size_lock(inode);
573 result = inode_newsize_ok(inode, size);
575 ll_inode_size_unlock(inode);
578 truncate_setsize(inode, size);
579 ll_inode_size_unlock(inode);
583 static int vvp_io_setattr_trunc(const struct lu_env *env,
584 const struct cl_io_slice *ios,
585 struct inode *inode, loff_t size)
587 inode_dio_wait(inode);
591 static int vvp_io_setattr_time(const struct lu_env *env,
592 const struct cl_io_slice *ios)
594 struct cl_io *io = ios->cis_io;
595 struct cl_object *obj = io->ci_obj;
596 struct cl_attr *attr = vvp_env_thread_attr(env);
598 unsigned valid = CAT_CTIME;
600 cl_object_attr_lock(obj);
601 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
602 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
603 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
606 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
607 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
610 result = cl_object_attr_update(env, obj, attr, valid);
611 cl_object_attr_unlock(obj);
616 static int vvp_io_setattr_start(const struct lu_env *env,
617 const struct cl_io_slice *ios)
619 struct cl_io *io = ios->cis_io;
620 struct inode *inode = vvp_object_inode(io->ci_obj);
624 if (cl_io_is_trunc(io))
625 result = vvp_io_setattr_trunc(env, ios, inode,
626 io->u.ci_setattr.sa_attr.lvb_size);
628 result = vvp_io_setattr_time(env, ios);
632 static void vvp_io_setattr_end(const struct lu_env *env,
633 const struct cl_io_slice *ios)
635 struct cl_io *io = ios->cis_io;
636 struct inode *inode = vvp_object_inode(io->ci_obj);
638 if (cl_io_is_trunc(io))
639 /* Truncate in memory pages - they must be clean pages
640 * because osc has already notified to destroy osc_extents.
642 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
647 static void vvp_io_setattr_fini(const struct lu_env *env,
648 const struct cl_io_slice *ios)
650 vvp_io_fini(env, ios);
653 static int vvp_io_read_start(const struct lu_env *env,
654 const struct cl_io_slice *ios)
656 struct vvp_io *vio = cl2vvp_io(env, ios);
657 struct cl_io *io = ios->cis_io;
658 struct cl_object *obj = io->ci_obj;
659 struct inode *inode = vvp_object_inode(obj);
660 struct file *file = vio->vui_fd->fd_file;
663 loff_t pos = io->u.ci_rd.rd.crw_pos;
664 long cnt = io->u.ci_rd.rd.crw_count;
665 long tot = vio->vui_tot_count;
668 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
670 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
672 if (!can_populate_pages(env, io, inode))
675 result = vvp_prep_size(env, obj, io, pos, tot, &exceed);
678 else if (exceed != 0)
681 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
682 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
683 inode->i_ino, cnt, pos, i_size_read(inode));
685 /* turn off the kernel's read-ahead */
686 vio->vui_fd->fd_file->f_ra.ra_pages = 0;
688 /* initialize read-ahead window once per syscall */
689 if (!vio->vui_ra_valid) {
690 vio->vui_ra_valid = true;
691 vio->vui_ra_start = cl_index(obj, pos);
692 vio->vui_ra_count = cl_index(obj, tot + PAGE_SIZE - 1);
698 LASSERT(vio->vui_iocb->ki_pos == pos);
699 result = generic_file_read_iter(vio->vui_iocb, vio->vui_iter);
705 io->ci_nob += result;
706 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
707 vio->vui_fd, pos, result, READ);
713 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
714 struct cl_page_list *plist, int from, int to)
716 struct cl_2queue *queue = &io->ci_queue;
717 struct cl_page *page;
718 unsigned int bytes = 0;
721 if (plist->pl_nr == 0)
724 if (from > 0 || to != PAGE_SIZE) {
725 page = cl_page_list_first(plist);
726 if (plist->pl_nr == 1) {
727 cl_page_clip(env, page, from, to);
730 cl_page_clip(env, page, from, PAGE_SIZE);
731 if (to != PAGE_SIZE) {
732 page = cl_page_list_last(plist);
733 cl_page_clip(env, page, 0, to);
738 cl_2queue_init(queue);
739 cl_page_list_splice(plist, &queue->c2_qin);
740 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
742 /* plist is not sorted any more */
743 cl_page_list_splice(&queue->c2_qin, plist);
744 cl_page_list_splice(&queue->c2_qout, plist);
745 cl_2queue_fini(env, queue);
748 /* calculate bytes */
749 bytes = plist->pl_nr << PAGE_SHIFT;
750 bytes -= from + PAGE_SIZE - to;
752 while (plist->pl_nr > 0) {
753 page = cl_page_list_first(plist);
754 cl_page_list_del(env, plist, page);
756 cl_page_clip(env, page, 0, PAGE_SIZE);
758 SetPageUptodate(cl_page_vmpage(page));
759 cl_page_disown(env, io, page);
761 /* held in ll_cl_init() */
762 lu_ref_del(&page->cp_reference, "cl_io", io);
763 cl_page_put(env, page);
767 return bytes > 0 ? bytes : rc;
770 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
771 struct cl_page *page)
773 struct vvp_page *vpg;
774 struct page *vmpage = page->cp_vmpage;
775 struct cl_object *clob = cl_io_top(io)->ci_obj;
777 SetPageUptodate(vmpage);
778 set_page_dirty(vmpage);
780 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
781 vvp_write_pending(cl2vvp(clob), vpg);
783 cl_page_disown(env, io, page);
785 /* held in ll_cl_init() */
786 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
787 cl_page_put(env, page);
790 /* make sure the page list is contiguous */
791 static bool page_list_sanity_check(struct cl_object *obj,
792 struct cl_page_list *plist)
794 struct cl_page *page;
795 pgoff_t index = CL_PAGE_EOF;
797 cl_page_list_for_each(page, plist) {
798 struct vvp_page *vpg = cl_object_page_slice(obj, page);
800 if (index == CL_PAGE_EOF) {
801 index = vvp_index(vpg);
806 if (index == vvp_index(vpg))
814 /* Return how many bytes have queued or written */
815 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
817 struct cl_object *obj = io->ci_obj;
818 struct inode *inode = vvp_object_inode(obj);
819 struct vvp_io *vio = vvp_env_io(env);
820 struct cl_page_list *queue = &vio->u.write.vui_queue;
821 struct cl_page *page;
824 unsigned int npages = vio->u.write.vui_queue.pl_nr;
829 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
830 npages, vio->u.write.vui_from, vio->u.write.vui_to);
832 LASSERT(page_list_sanity_check(obj, queue));
834 /* submit IO with async write */
835 rc = cl_io_commit_async(env, io, queue,
836 vio->u.write.vui_from, vio->u.write.vui_to,
837 write_commit_callback);
838 npages -= queue->pl_nr; /* already committed pages */
840 /* calculate how many bytes were written */
841 bytes = npages << PAGE_SHIFT;
844 bytes -= vio->u.write.vui_from;
845 if (queue->pl_nr == 0) /* last page */
846 bytes -= PAGE_SIZE - vio->u.write.vui_to;
847 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
849 vio->u.write.vui_written += bytes;
851 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
852 npages, bytes, vio->u.write.vui_written);
854 /* the first page must have been written. */
855 vio->u.write.vui_from = 0;
857 LASSERT(page_list_sanity_check(obj, queue));
858 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
860 /* out of quota, try sync write */
861 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
862 rc = vvp_io_commit_sync(env, io, queue,
863 vio->u.write.vui_from,
864 vio->u.write.vui_to);
866 vio->u.write.vui_written += rc;
871 /* update inode size */
872 ll_merge_attr(env, inode);
874 /* Now the pages in queue were failed to commit, discard them
875 * unless they were dirtied before.
877 while (queue->pl_nr > 0) {
878 page = cl_page_list_first(queue);
879 cl_page_list_del(env, queue, page);
881 if (!PageDirty(cl_page_vmpage(page)))
882 cl_page_discard(env, io, page);
884 cl_page_disown(env, io, page);
886 /* held in ll_cl_init() */
887 lu_ref_del(&page->cp_reference, "cl_io", io);
888 cl_page_put(env, page);
890 cl_page_list_fini(env, queue);
895 static int vvp_io_write_start(const struct lu_env *env,
896 const struct cl_io_slice *ios)
898 struct vvp_io *vio = cl2vvp_io(env, ios);
899 struct cl_io *io = ios->cis_io;
900 struct cl_object *obj = io->ci_obj;
901 struct inode *inode = vvp_object_inode(obj);
903 loff_t pos = io->u.ci_wr.wr.crw_pos;
904 size_t cnt = io->u.ci_wr.wr.crw_count;
906 if (!can_populate_pages(env, io, inode))
909 if (cl_io_is_append(io)) {
911 * PARALLEL IO This has to be changed for parallel IO doing
912 * out-of-order writes.
914 ll_merge_attr(env, inode);
915 pos = i_size_read(inode);
916 io->u.ci_wr.wr.crw_pos = pos;
917 vio->vui_iocb->ki_pos = pos;
919 LASSERT(vio->vui_iocb->ki_pos == pos);
922 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
924 if (!vio->vui_iter) {
925 /* from a temp io in ll_cl_init(). */
929 * When using the locked AIO function (generic_file_aio_write())
930 * testing has shown the inode mutex to be a limiting factor
931 * with multi-threaded single shared file performance. To get
932 * around this, we now use the lockless version. To maintain
933 * consistency, proper locking to protect against writes,
934 * trucates, etc. is handled in the higher layers of lustre.
936 bool lock_node = !IS_NOSEC(inode);
940 result = __generic_file_write_iter(vio->vui_iocb,
945 if (result > 0 || result == -EIOCBQUEUED)
946 result = generic_write_sync(vio->vui_iocb, result);
950 result = vvp_io_write_commit(env, io);
951 if (vio->u.write.vui_written > 0) {
952 result = vio->u.write.vui_written;
953 io->ci_nob += result;
955 CDEBUG(D_VFSTRACE, "write: nob %zd, result: %zd\n",
960 struct ll_inode_info *lli = ll_i2info(inode);
962 spin_lock(&lli->lli_lock);
963 lli->lli_flags |= LLIF_DATA_MODIFIED;
964 spin_unlock(&lli->lli_lock);
968 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
969 vio->vui_fd, pos, result, WRITE);
975 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
977 struct vm_fault *vmf = cfio->ft_vmf;
979 cfio->ft_flags = filemap_fault(cfio->ft_vma, vmf);
980 cfio->ft_flags_valid = 1;
984 "page %p map %p index %lu flags %lx count %u priv %0lx: got addr %p type NOPAGE\n",
985 vmf->page, vmf->page->mapping, vmf->page->index,
986 (long)vmf->page->flags, page_count(vmf->page),
987 page_private(vmf->page), vmf->virtual_address);
988 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
989 lock_page(vmf->page);
990 cfio->ft_flags |= VM_FAULT_LOCKED;
993 cfio->ft_vmpage = vmf->page;
997 if (cfio->ft_flags & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV)) {
998 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
1002 if (cfio->ft_flags & VM_FAULT_OOM) {
1003 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
1007 if (cfio->ft_flags & VM_FAULT_RETRY)
1010 CERROR("Unknown error in page fault %d!\n", cfio->ft_flags);
1014 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1015 struct cl_page *page)
1017 struct vvp_page *vpg;
1018 struct cl_object *clob = cl_io_top(io)->ci_obj;
1020 set_page_dirty(page->cp_vmpage);
1022 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
1023 vvp_write_pending(cl2vvp(clob), vpg);
1026 static int vvp_io_fault_start(const struct lu_env *env,
1027 const struct cl_io_slice *ios)
1029 struct vvp_io *vio = cl2vvp_io(env, ios);
1030 struct cl_io *io = ios->cis_io;
1031 struct cl_object *obj = io->ci_obj;
1032 struct inode *inode = vvp_object_inode(obj);
1033 struct cl_fault_io *fio = &io->u.ci_fault;
1034 struct vvp_fault_io *cfio = &vio->u.fault;
1037 struct page *vmpage = NULL;
1038 struct cl_page *page;
1042 if (fio->ft_executable &&
1043 inode->i_mtime.tv_sec != vio->u.fault.ft_mtime)
1045 " changed while waiting for the page fault lock\n",
1046 PFID(lu_object_fid(&obj->co_lu)));
1048 /* offset of the last byte on the page */
1049 offset = cl_offset(obj, fio->ft_index + 1) - 1;
1050 LASSERT(cl_index(obj, offset) == fio->ft_index);
1051 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1055 /* must return locked page */
1056 if (fio->ft_mkwrite) {
1057 LASSERT(cfio->ft_vmpage);
1058 lock_page(cfio->ft_vmpage);
1060 result = vvp_io_kernel_fault(cfio);
1065 vmpage = cfio->ft_vmpage;
1066 LASSERT(PageLocked(vmpage));
1068 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1069 ll_invalidate_page(vmpage);
1071 size = i_size_read(inode);
1072 /* Though we have already held a cl_lock upon this page, but
1073 * it still can be truncated locally.
1075 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1076 (page_offset(vmpage) > size))) {
1077 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1079 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1086 last_index = cl_index(obj, size - 1);
1088 if (fio->ft_mkwrite) {
1090 * Capture the size while holding the lli_trunc_sem from above
1091 * we want to make sure that we complete the mkwrite action
1092 * while holding this lock. We need to make sure that we are
1093 * not past the end of the file.
1095 if (last_index < fio->ft_index) {
1097 "llite: mkwrite and truncate race happened: %p: 0x%lx 0x%lx\n",
1098 vmpage->mapping, fio->ft_index, last_index);
1100 * We need to return if we are
1101 * passed the end of the file. This will propagate
1102 * up the call stack to ll_page_mkwrite where
1103 * we will return VM_FAULT_NOPAGE. Any non-negative
1104 * value returned here will be silently
1105 * converted to 0. If the vmpage->mapping is null
1106 * the error code would be converted back to ENODATA
1107 * in ll_page_mkwrite0. Thus we return -ENODATA
1108 * to handle both cases
1115 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1117 result = PTR_ERR(page);
1121 /* if page is going to be written, we should add this page into cache
1124 if (fio->ft_mkwrite) {
1125 wait_on_page_writeback(vmpage);
1126 if (!PageDirty(vmpage)) {
1127 struct cl_page_list *plist = &io->ci_queue.c2_qin;
1128 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1131 /* vvp_page_assume() calls wait_on_page_writeback(). */
1132 cl_page_assume(env, io, page);
1134 cl_page_list_init(plist);
1135 cl_page_list_add(plist, page);
1138 if (last_index == vvp_index(vpg))
1139 to = size & ~PAGE_MASK;
1141 /* Do not set Dirty bit here so that in case IO is
1142 * started before the page is really made dirty, we
1143 * still have chance to detect it.
1145 result = cl_io_commit_async(env, io, plist, 0, to,
1146 mkwrite_commit_callback);
1147 LASSERT(cl_page_is_owned(page, io));
1148 cl_page_list_fini(env, plist);
1152 cl_page_discard(env, io, page);
1153 cl_page_disown(env, io, page);
1155 cl_page_put(env, page);
1157 /* we're in big trouble, what can we do now? */
1158 if (result == -EDQUOT)
1162 cl_page_disown(env, io, page);
1168 * The ft_index is only used in the case of
1169 * a mkwrite action. We need to check
1170 * our assertions are correct, since
1171 * we should have caught this above
1173 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1174 if (fio->ft_index == last_index)
1176 * Last page is mapped partially.
1178 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1180 fio->ft_nob = cl_page_size(obj);
1182 lu_ref_add(&page->cp_reference, "fault", io);
1183 fio->ft_page = page;
1186 /* return unlocked vmpage to avoid deadlocking */
1188 unlock_page(vmpage);
1190 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1195 static int vvp_io_fsync_start(const struct lu_env *env,
1196 const struct cl_io_slice *ios)
1198 /* we should mark TOWRITE bit to each dirty page in radix tree to
1199 * verify pages have been written, but this is difficult because of
1205 static int vvp_io_read_page(const struct lu_env *env,
1206 const struct cl_io_slice *ios,
1207 const struct cl_page_slice *slice)
1209 struct cl_io *io = ios->cis_io;
1210 struct vvp_page *vpg = cl2vvp_page(slice);
1211 struct cl_page *page = slice->cpl_page;
1212 struct inode *inode = vvp_object_inode(slice->cpl_obj);
1213 struct ll_sb_info *sbi = ll_i2sbi(inode);
1214 struct ll_file_data *fd = cl2vvp_io(env, ios)->vui_fd;
1215 struct ll_readahead_state *ras = &fd->fd_ras;
1216 struct cl_2queue *queue = &io->ci_queue;
1218 if (sbi->ll_ra_info.ra_max_pages_per_file &&
1219 sbi->ll_ra_info.ra_max_pages)
1220 ras_update(sbi, inode, ras, vvp_index(vpg),
1221 vpg->vpg_defer_uptodate);
1223 if (vpg->vpg_defer_uptodate) {
1224 vpg->vpg_ra_used = 1;
1225 cl_page_export(env, page, 1);
1228 * Add page into the queue even when it is marked uptodate above.
1229 * this will unlock it automatically as part of cl_page_list_disown().
1232 cl_page_list_add(&queue->c2_qin, page);
1233 if (sbi->ll_ra_info.ra_max_pages_per_file &&
1234 sbi->ll_ra_info.ra_max_pages)
1235 ll_readahead(env, io, &queue->c2_qin, ras,
1236 vpg->vpg_defer_uptodate);
1241 static void vvp_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
1243 CLOBINVRNT(env, ios->cis_io->ci_obj,
1244 vvp_object_invariant(ios->cis_io->ci_obj));
1247 static const struct cl_io_operations vvp_io_ops = {
1250 .cio_fini = vvp_io_fini,
1251 .cio_lock = vvp_io_read_lock,
1252 .cio_start = vvp_io_read_start,
1253 .cio_advance = vvp_io_advance,
1256 .cio_fini = vvp_io_fini,
1257 .cio_iter_init = vvp_io_write_iter_init,
1258 .cio_iter_fini = vvp_io_write_iter_fini,
1259 .cio_lock = vvp_io_write_lock,
1260 .cio_start = vvp_io_write_start,
1261 .cio_advance = vvp_io_advance,
1264 .cio_fini = vvp_io_setattr_fini,
1265 .cio_iter_init = vvp_io_setattr_iter_init,
1266 .cio_lock = vvp_io_setattr_lock,
1267 .cio_start = vvp_io_setattr_start,
1268 .cio_end = vvp_io_setattr_end
1271 .cio_fini = vvp_io_fault_fini,
1272 .cio_iter_init = vvp_io_fault_iter_init,
1273 .cio_lock = vvp_io_fault_lock,
1274 .cio_start = vvp_io_fault_start,
1275 .cio_end = vvp_io_end,
1278 .cio_start = vvp_io_fsync_start,
1279 .cio_fini = vvp_io_fini
1282 .cio_fini = vvp_io_fini
1285 .cio_read_page = vvp_io_read_page,
1288 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1291 struct vvp_io *vio = vvp_env_io(env);
1292 struct inode *inode = vvp_object_inode(obj);
1295 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1297 CDEBUG(D_VFSTRACE, DFID
1298 " ignore/verify layout %d/%d, layout version %d restore needed %d\n",
1299 PFID(lu_object_fid(&obj->co_lu)),
1300 io->ci_ignore_layout, io->ci_verify_layout,
1301 vio->vui_layout_gen, io->ci_restore_needed);
1303 CL_IO_SLICE_CLEAN(vio, vui_cl);
1304 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1305 vio->vui_ra_valid = false;
1307 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1309 struct ll_inode_info *lli = ll_i2info(inode);
1311 count = io->u.ci_rw.crw_count;
1312 /* "If nbyte is 0, read() will return 0 and have no other
1313 * results." -- Single Unix Spec
1318 vio->vui_tot_count = count;
1320 /* for read/write, we store the jobid in the inode, and
1321 * it'll be fetched by osc when building RPC.
1323 * it's not accurate if the file is shared by different
1326 lustre_get_jobid(lli->lli_jobid);
1327 } else if (io->ci_type == CIT_SETATTR) {
1328 if (!cl_io_is_trunc(io))
1329 io->ci_lockreq = CILR_MANDATORY;
1332 /* ignore layout change for generic CIT_MISC but not for glimpse.
1333 * io context for glimpse must set ci_verify_layout to true,
1334 * see cl_glimpse_size0() for details.
1336 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1337 io->ci_ignore_layout = 1;
1339 /* Enqueue layout lock and get layout version. We need to do this
1340 * even for operations requiring to open file, such as read and write,
1341 * because it might not grant layout lock in IT_OPEN.
1343 if (result == 0 && !io->ci_ignore_layout) {
1344 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1345 if (result == -ENOENT)
1346 /* If the inode on MDS has been removed, but the objects
1347 * on OSTs haven't been destroyed (async unlink), layout
1348 * fetch will return -ENOENT, we'd ignore this error
1349 * and continue with dirty flush. LU-3230.
1353 CERROR("%s: refresh file layout " DFID " error %d.\n",
1354 ll_get_fsname(inode->i_sb, NULL, 0),
1355 PFID(lu_object_fid(&obj->co_lu)), result);