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"
41 #include "../include/lustre_lite.h"
43 #include "llite_internal.h"
44 #include "vvp_internal.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 * For swapping layout. The file's layout may have changed.
59 * To avoid populating pages to a wrong stripe, we have to verify the
60 * correctness of layout. It works because swapping layout processes
61 * have to acquire group lock.
63 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
66 struct ll_inode_info *lli = ll_i2info(inode);
67 struct vvp_io *vio = vvp_env_io(env);
70 switch (io->ci_type) {
73 /* don't need lock here to check lli_layout_gen as we have held
74 * extent lock and GROUP lock has to hold to swap layout
76 if (ll_layout_version_get(lli) != vio->vui_layout_gen) {
77 io->ci_need_restart = 1;
78 /* this will return application a short read/write */
83 /* fault is okay because we've already had a page. */
91 static void vvp_object_size_lock(struct cl_object *obj)
93 struct inode *inode = vvp_object_inode(obj);
95 ll_inode_size_lock(inode);
96 cl_object_attr_lock(obj);
99 static void vvp_object_size_unlock(struct cl_object *obj)
101 struct inode *inode = vvp_object_inode(obj);
103 cl_object_attr_unlock(obj);
104 ll_inode_size_unlock(inode);
108 * Helper function that if necessary adjusts file size (inode->i_size), when
109 * position at the offset \a pos is accessed. File size can be arbitrary stale
110 * on a Lustre client, but client at least knows KMS. If accessed area is
111 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
113 * Locking: cl_isize_lock is used to serialize changes to inode size and to
114 * protect consistency between inode size and cl_object
115 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
116 * top-object and sub-objects.
118 static int vvp_prep_size(const struct lu_env *env, struct cl_object *obj,
119 struct cl_io *io, loff_t start, size_t count,
122 struct cl_attr *attr = vvp_env_thread_attr(env);
123 struct inode *inode = vvp_object_inode(obj);
124 loff_t pos = start + count - 1;
129 * Consistency guarantees: following possibilities exist for the
130 * relation between region being accessed and real file size at this
133 * (A): the region is completely inside of the file;
135 * (B-x): x bytes of region are inside of the file, the rest is
138 * (C): the region is completely outside of the file.
140 * This classification is stable under DLM lock already acquired by
141 * the caller, because to change the class, other client has to take
142 * DLM lock conflicting with our lock. Also, any updates to ->i_size
143 * by other threads on this client are serialized by
144 * ll_inode_size_lock(). This guarantees that short reads are handled
145 * correctly in the face of concurrent writes and truncates.
147 vvp_object_size_lock(obj);
148 result = cl_object_attr_get(env, obj, attr);
153 * A glimpse is necessary to determine whether we
154 * return a short read (B) or some zeroes at the end
157 vvp_object_size_unlock(obj);
158 result = cl_glimpse_lock(env, io, inode, obj, 0);
159 if (result == 0 && exceed) {
160 /* If objective page index exceed end-of-file
161 * page index, return directly. Do not expect
162 * kernel will check such case correctly.
163 * linux-2.6.18-128.1.1 miss to do that.
166 loff_t size = i_size_read(inode);
167 loff_t cur_index = start >> PAGE_SHIFT;
168 loff_t size_index = (size - 1) >> PAGE_SHIFT;
170 if ((size == 0 && cur_index != 0) ||
171 size_index < cur_index)
177 * region is within kms and, hence, within real file
178 * size (A). We need to increase i_size to cover the
179 * read region so that generic_file_read() will do its
180 * job, but that doesn't mean the kms size is
181 * _correct_, it is only the _minimum_ size. If
182 * someone does a stat they will get the correct size
183 * which will always be >= the kms value here.
186 if (i_size_read(inode) < kms) {
187 i_size_write(inode, kms);
188 CDEBUG(D_VFSTRACE, DFID " updating i_size %llu\n",
189 PFID(lu_object_fid(&obj->co_lu)),
190 (__u64)i_size_read(inode));
194 vvp_object_size_unlock(obj);
199 /*****************************************************************************
205 static int vvp_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
206 __u32 enqflags, enum cl_lock_mode mode,
207 pgoff_t start, pgoff_t end)
209 struct vvp_io *vio = vvp_env_io(env);
210 struct cl_lock_descr *descr = &vio->vui_link.cill_descr;
211 struct cl_object *obj = io->ci_obj;
213 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
215 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
217 memset(&vio->vui_link, 0, sizeof(vio->vui_link));
219 if (vio->vui_fd && (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
220 descr->cld_mode = CLM_GROUP;
221 descr->cld_gid = vio->vui_fd->fd_grouplock.lg_gid;
223 descr->cld_mode = mode;
225 descr->cld_obj = obj;
226 descr->cld_start = start;
227 descr->cld_end = end;
228 descr->cld_enq_flags = enqflags;
230 cl_io_lock_add(env, io, &vio->vui_link);
234 static int vvp_io_one_lock(const struct lu_env *env, struct cl_io *io,
235 __u32 enqflags, enum cl_lock_mode mode,
236 loff_t start, loff_t end)
238 struct cl_object *obj = io->ci_obj;
240 return vvp_io_one_lock_index(env, io, enqflags, mode,
241 cl_index(obj, start), cl_index(obj, end));
244 static int vvp_io_write_iter_init(const struct lu_env *env,
245 const struct cl_io_slice *ios)
247 struct vvp_io *vio = cl2vvp_io(env, ios);
249 cl_page_list_init(&vio->u.write.vui_queue);
250 vio->u.write.vui_written = 0;
251 vio->u.write.vui_from = 0;
252 vio->u.write.vui_to = PAGE_SIZE;
257 static void vvp_io_write_iter_fini(const struct lu_env *env,
258 const struct cl_io_slice *ios)
260 struct vvp_io *vio = cl2vvp_io(env, ios);
262 LASSERT(vio->u.write.vui_queue.pl_nr == 0);
265 static int vvp_io_fault_iter_init(const struct lu_env *env,
266 const struct cl_io_slice *ios)
268 struct vvp_io *vio = cl2vvp_io(env, ios);
269 struct inode *inode = vvp_object_inode(ios->cis_obj);
271 LASSERT(inode == file_inode(vio->vui_fd->fd_file));
272 vio->u.fault.ft_mtime = inode->i_mtime.tv_sec;
276 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
278 struct cl_io *io = ios->cis_io;
279 struct cl_object *obj = io->ci_obj;
280 struct vvp_io *vio = cl2vvp_io(env, ios);
281 struct inode *inode = vvp_object_inode(obj);
283 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
285 CDEBUG(D_VFSTRACE, DFID
286 " ignore/verify layout %d/%d, layout version %d restore needed %d\n",
287 PFID(lu_object_fid(&obj->co_lu)),
288 io->ci_ignore_layout, io->ci_verify_layout,
289 vio->vui_layout_gen, io->ci_restore_needed);
291 if (io->ci_restore_needed == 1) {
294 /* file was detected release, we need to restore it
295 * before finishing the io
297 rc = ll_layout_restore(inode, 0, OBD_OBJECT_EOF);
298 /* if restore registration failed, no restart,
299 * we will return -ENODATA
301 /* The layout will change after restore, so we need to
302 * block on layout lock hold by the MDT
303 * as MDT will not send new layout in lvb (see LU-3124)
304 * we have to explicitly fetch it, all this will be done
305 * by ll_layout_refresh()
308 io->ci_restore_needed = 0;
309 io->ci_need_restart = 1;
310 io->ci_verify_layout = 1;
312 io->ci_restore_needed = 1;
313 io->ci_need_restart = 0;
314 io->ci_verify_layout = 0;
319 if (!io->ci_ignore_layout && io->ci_verify_layout) {
322 /* check layout version */
323 ll_layout_refresh(inode, &gen);
324 io->ci_need_restart = vio->vui_layout_gen != gen;
325 if (io->ci_need_restart) {
327 DFID" layout changed from %d to %d.\n",
328 PFID(lu_object_fid(&obj->co_lu)),
329 vio->vui_layout_gen, gen);
330 /* today successful restore is the only possible case */
331 /* restore was done, clear restoring state */
332 ll_i2info(vvp_object_inode(obj))->lli_flags &=
333 ~LLIF_FILE_RESTORING;
338 static void vvp_io_fault_fini(const struct lu_env *env,
339 const struct cl_io_slice *ios)
341 struct cl_io *io = ios->cis_io;
342 struct cl_page *page = io->u.ci_fault.ft_page;
344 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
347 lu_ref_del(&page->cp_reference, "fault", io);
348 cl_page_put(env, page);
349 io->u.ci_fault.ft_page = NULL;
351 vvp_io_fini(env, ios);
354 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
357 * we only want to hold PW locks if the mmap() can generate
358 * writes back to the file and that only happens in shared
361 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
366 static int vvp_mmap_locks(const struct lu_env *env,
367 struct vvp_io *vio, struct cl_io *io)
369 struct vvp_thread_info *cti = vvp_env_info(env);
370 struct mm_struct *mm = current->mm;
371 struct vm_area_struct *vma;
372 struct cl_lock_descr *descr = &cti->vti_descr;
373 ldlm_policy_data_t policy;
380 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
382 if (!vio->vui_iter) /* nfs or loop back device write */
385 /* No MM (e.g. NFS)? No vmas too. */
389 iov_for_each(iov, i, *vio->vui_iter) {
390 addr = (unsigned long)iov.iov_base;
395 count += addr & (~PAGE_MASK);
398 down_read(&mm->mmap_sem);
399 while ((vma = our_vma(mm, addr, count)) != NULL) {
400 struct inode *inode = file_inode(vma->vm_file);
401 int flags = CEF_MUST;
403 if (ll_file_nolock(vma->vm_file)) {
405 * For no lock case is not allowed for mmap
412 * XXX: Required lock mode can be weakened: CIT_WRITE
413 * io only ever reads user level buffer, and CIT_READ
416 policy_from_vma(&policy, vma, addr, count);
417 descr->cld_mode = vvp_mode_from_vma(vma);
418 descr->cld_obj = ll_i2info(inode)->lli_clob;
419 descr->cld_start = cl_index(descr->cld_obj,
420 policy.l_extent.start);
421 descr->cld_end = cl_index(descr->cld_obj,
422 policy.l_extent.end);
423 descr->cld_enq_flags = flags;
424 result = cl_io_lock_alloc_add(env, io, descr);
426 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
427 descr->cld_mode, descr->cld_start,
433 if (vma->vm_end - addr >= count)
436 count -= vma->vm_end - addr;
439 up_read(&mm->mmap_sem);
446 static void vvp_io_advance(const struct lu_env *env,
447 const struct cl_io_slice *ios,
450 struct cl_object *obj = ios->cis_io->ci_obj;
451 struct vvp_io *vio = cl2vvp_io(env, ios);
452 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
454 iov_iter_reexpand(vio->vui_iter, vio->vui_tot_count -= nob);
457 static void vvp_io_update_iov(const struct lu_env *env,
458 struct vvp_io *vio, struct cl_io *io)
460 size_t size = io->u.ci_rw.crw_count;
465 iov_iter_truncate(vio->vui_iter, size);
468 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
469 enum cl_lock_mode mode, loff_t start, loff_t end)
471 struct vvp_io *vio = vvp_env_io(env);
475 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
477 vvp_io_update_iov(env, vio, io);
479 if (io->u.ci_rw.crw_nonblock)
480 ast_flags |= CEF_NONBLOCK;
481 result = vvp_mmap_locks(env, vio, io);
483 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
487 static int vvp_io_read_lock(const struct lu_env *env,
488 const struct cl_io_slice *ios)
490 struct cl_io *io = ios->cis_io;
491 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
494 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
495 rd->crw_pos + rd->crw_count - 1);
500 static int vvp_io_fault_lock(const struct lu_env *env,
501 const struct cl_io_slice *ios)
503 struct cl_io *io = ios->cis_io;
504 struct vvp_io *vio = cl2vvp_io(env, ios);
506 * XXX LDLM_FL_CBPENDING
508 return vvp_io_one_lock_index(env,
510 vvp_mode_from_vma(vio->u.fault.ft_vma),
511 io->u.ci_fault.ft_index,
512 io->u.ci_fault.ft_index);
515 static int vvp_io_write_lock(const struct lu_env *env,
516 const struct cl_io_slice *ios)
518 struct cl_io *io = ios->cis_io;
522 if (io->u.ci_wr.wr_append) {
524 end = OBD_OBJECT_EOF;
526 start = io->u.ci_wr.wr.crw_pos;
527 end = start + io->u.ci_wr.wr.crw_count - 1;
529 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
532 static int vvp_io_setattr_iter_init(const struct lu_env *env,
533 const struct cl_io_slice *ios)
539 * Implementation of cl_io_operations::vio_lock() method for CIT_SETATTR io.
541 * Handles "lockless io" mode when extent locking is done by server.
543 static int vvp_io_setattr_lock(const struct lu_env *env,
544 const struct cl_io_slice *ios)
546 struct cl_io *io = ios->cis_io;
550 if (cl_io_is_trunc(io)) {
551 new_size = io->u.ci_setattr.sa_attr.lvb_size;
553 enqflags = CEF_DISCARD_DATA;
555 if ((io->u.ci_setattr.sa_attr.lvb_mtime >=
556 io->u.ci_setattr.sa_attr.lvb_ctime) ||
557 (io->u.ci_setattr.sa_attr.lvb_atime >=
558 io->u.ci_setattr.sa_attr.lvb_ctime))
563 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
564 new_size, OBD_OBJECT_EOF);
567 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
571 * Only ll_inode_size_lock is taken at this level.
573 ll_inode_size_lock(inode);
574 result = inode_newsize_ok(inode, size);
576 ll_inode_size_unlock(inode);
579 truncate_setsize(inode, size);
580 ll_inode_size_unlock(inode);
584 static int vvp_io_setattr_trunc(const struct lu_env *env,
585 const struct cl_io_slice *ios,
586 struct inode *inode, loff_t size)
588 inode_dio_wait(inode);
592 static int vvp_io_setattr_time(const struct lu_env *env,
593 const struct cl_io_slice *ios)
595 struct cl_io *io = ios->cis_io;
596 struct cl_object *obj = io->ci_obj;
597 struct cl_attr *attr = vvp_env_thread_attr(env);
599 unsigned valid = CAT_CTIME;
601 cl_object_attr_lock(obj);
602 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
603 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
604 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
607 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
608 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
611 result = cl_object_attr_set(env, obj, attr, valid);
612 cl_object_attr_unlock(obj);
617 static int vvp_io_setattr_start(const struct lu_env *env,
618 const struct cl_io_slice *ios)
620 struct cl_io *io = ios->cis_io;
621 struct inode *inode = vvp_object_inode(io->ci_obj);
625 if (cl_io_is_trunc(io))
626 result = vvp_io_setattr_trunc(env, ios, inode,
627 io->u.ci_setattr.sa_attr.lvb_size);
629 result = vvp_io_setattr_time(env, ios);
633 static void vvp_io_setattr_end(const struct lu_env *env,
634 const struct cl_io_slice *ios)
636 struct cl_io *io = ios->cis_io;
637 struct inode *inode = vvp_object_inode(io->ci_obj);
639 if (cl_io_is_trunc(io))
640 /* Truncate in memory pages - they must be clean pages
641 * because osc has already notified to destroy osc_extents.
643 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
648 static void vvp_io_setattr_fini(const struct lu_env *env,
649 const struct cl_io_slice *ios)
651 vvp_io_fini(env, ios);
654 static int vvp_io_read_start(const struct lu_env *env,
655 const struct cl_io_slice *ios)
657 struct vvp_io *vio = cl2vvp_io(env, ios);
658 struct cl_io *io = ios->cis_io;
659 struct cl_object *obj = io->ci_obj;
660 struct inode *inode = vvp_object_inode(obj);
661 struct file *file = vio->vui_fd->fd_file;
664 loff_t pos = io->u.ci_rd.rd.crw_pos;
665 long cnt = io->u.ci_rd.rd.crw_count;
666 long tot = vio->vui_tot_count;
669 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
671 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
673 if (!can_populate_pages(env, io, inode))
676 result = vvp_prep_size(env, obj, io, pos, tot, &exceed);
679 else if (exceed != 0)
682 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
683 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
684 inode->i_ino, cnt, pos, i_size_read(inode));
686 /* turn off the kernel's read-ahead */
687 vio->vui_fd->fd_file->f_ra.ra_pages = 0;
689 /* initialize read-ahead window once per syscall */
690 if (!vio->vui_ra_valid) {
691 vio->vui_ra_valid = true;
692 vio->vui_ra_start = cl_index(obj, pos);
693 vio->vui_ra_count = cl_index(obj, tot + PAGE_SIZE - 1);
699 LASSERT(vio->vui_iocb->ki_pos == pos);
700 result = generic_file_read_iter(vio->vui_iocb, vio->vui_iter);
706 io->ci_nob += result;
707 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
708 vio->vui_fd, pos, result, READ);
714 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
715 struct cl_page_list *plist, int from, int to)
717 struct cl_2queue *queue = &io->ci_queue;
718 struct cl_page *page;
719 unsigned int bytes = 0;
722 if (plist->pl_nr == 0)
725 if (from > 0 || to != PAGE_SIZE) {
726 page = cl_page_list_first(plist);
727 if (plist->pl_nr == 1) {
728 cl_page_clip(env, page, from, to);
731 cl_page_clip(env, page, from, PAGE_SIZE);
732 if (to != PAGE_SIZE) {
733 page = cl_page_list_last(plist);
734 cl_page_clip(env, page, 0, to);
739 cl_2queue_init(queue);
740 cl_page_list_splice(plist, &queue->c2_qin);
741 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
743 /* plist is not sorted any more */
744 cl_page_list_splice(&queue->c2_qin, plist);
745 cl_page_list_splice(&queue->c2_qout, plist);
746 cl_2queue_fini(env, queue);
749 /* calculate bytes */
750 bytes = plist->pl_nr << PAGE_SHIFT;
751 bytes -= from + PAGE_SIZE - to;
753 while (plist->pl_nr > 0) {
754 page = cl_page_list_first(plist);
755 cl_page_list_del(env, plist, page);
757 cl_page_clip(env, page, 0, PAGE_SIZE);
759 SetPageUptodate(cl_page_vmpage(page));
760 cl_page_disown(env, io, page);
762 /* held in ll_cl_init() */
763 lu_ref_del(&page->cp_reference, "cl_io", io);
764 cl_page_put(env, page);
768 return bytes > 0 ? bytes : rc;
771 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
772 struct cl_page *page)
774 struct vvp_page *vpg;
775 struct page *vmpage = page->cp_vmpage;
776 struct cl_object *clob = cl_io_top(io)->ci_obj;
778 SetPageUptodate(vmpage);
779 set_page_dirty(vmpage);
781 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
782 vvp_write_pending(cl2vvp(clob), vpg);
784 cl_page_disown(env, io, page);
786 /* held in ll_cl_init() */
787 lu_ref_del(&page->cp_reference, "cl_io", io);
788 cl_page_put(env, page);
791 /* make sure the page list is contiguous */
792 static bool page_list_sanity_check(struct cl_object *obj,
793 struct cl_page_list *plist)
795 struct cl_page *page;
796 pgoff_t index = CL_PAGE_EOF;
798 cl_page_list_for_each(page, plist) {
799 struct vvp_page *vpg = cl_object_page_slice(obj, page);
801 if (index == CL_PAGE_EOF) {
802 index = vvp_index(vpg);
807 if (index == vvp_index(vpg))
815 /* Return how many bytes have queued or written */
816 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
818 struct cl_object *obj = io->ci_obj;
819 struct inode *inode = vvp_object_inode(obj);
820 struct vvp_io *vio = vvp_env_io(env);
821 struct cl_page_list *queue = &vio->u.write.vui_queue;
822 struct cl_page *page;
825 unsigned int npages = vio->u.write.vui_queue.pl_nr;
830 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
831 npages, vio->u.write.vui_from, vio->u.write.vui_to);
833 LASSERT(page_list_sanity_check(obj, queue));
835 /* submit IO with async write */
836 rc = cl_io_commit_async(env, io, queue,
837 vio->u.write.vui_from, vio->u.write.vui_to,
838 write_commit_callback);
839 npages -= queue->pl_nr; /* already committed pages */
841 /* calculate how many bytes were written */
842 bytes = npages << PAGE_SHIFT;
845 bytes -= vio->u.write.vui_from;
846 if (queue->pl_nr == 0) /* last page */
847 bytes -= PAGE_SIZE - vio->u.write.vui_to;
848 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
850 vio->u.write.vui_written += bytes;
852 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
853 npages, bytes, vio->u.write.vui_written);
855 /* the first page must have been written. */
856 vio->u.write.vui_from = 0;
858 LASSERT(page_list_sanity_check(obj, queue));
859 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
861 /* out of quota, try sync write */
862 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
863 rc = vvp_io_commit_sync(env, io, queue,
864 vio->u.write.vui_from,
865 vio->u.write.vui_to);
867 vio->u.write.vui_written += rc;
872 /* update inode size */
873 ll_merge_attr(env, inode);
875 /* Now the pages in queue were failed to commit, discard them
876 * unless they were dirtied before.
878 while (queue->pl_nr > 0) {
879 page = cl_page_list_first(queue);
880 cl_page_list_del(env, queue, page);
882 if (!PageDirty(cl_page_vmpage(page)))
883 cl_page_discard(env, io, page);
885 cl_page_disown(env, io, page);
887 /* held in ll_cl_init() */
888 lu_ref_del(&page->cp_reference, "cl_io", io);
889 cl_page_put(env, page);
891 cl_page_list_fini(env, queue);
896 static int vvp_io_write_start(const struct lu_env *env,
897 const struct cl_io_slice *ios)
899 struct vvp_io *vio = cl2vvp_io(env, ios);
900 struct cl_io *io = ios->cis_io;
901 struct cl_object *obj = io->ci_obj;
902 struct inode *inode = vvp_object_inode(obj);
904 loff_t pos = io->u.ci_wr.wr.crw_pos;
905 size_t cnt = io->u.ci_wr.wr.crw_count;
907 if (!can_populate_pages(env, io, inode))
910 if (cl_io_is_append(io)) {
912 * PARALLEL IO This has to be changed for parallel IO doing
913 * out-of-order writes.
915 ll_merge_attr(env, inode);
916 pos = i_size_read(inode);
917 io->u.ci_wr.wr.crw_pos = pos;
918 vio->vui_iocb->ki_pos = pos;
920 LASSERT(vio->vui_iocb->ki_pos == pos);
923 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
925 if (!vio->vui_iter) /* from a temp io in ll_cl_init(). */
928 result = generic_file_write_iter(vio->vui_iocb, vio->vui_iter);
931 result = vvp_io_write_commit(env, io);
932 if (vio->u.write.vui_written > 0) {
933 result = vio->u.write.vui_written;
934 io->ci_nob += result;
936 CDEBUG(D_VFSTRACE, "write: nob %zd, result: %zd\n",
941 struct ll_inode_info *lli = ll_i2info(inode);
943 spin_lock(&lli->lli_lock);
944 lli->lli_flags |= LLIF_DATA_MODIFIED;
945 spin_unlock(&lli->lli_lock);
949 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
950 vio->vui_fd, pos, result, WRITE);
956 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
958 struct vm_fault *vmf = cfio->ft_vmf;
960 cfio->ft_flags = filemap_fault(cfio->ft_vma, vmf);
961 cfio->ft_flags_valid = 1;
965 "page %p map %p index %lu flags %lx count %u priv %0lx: got addr %p type NOPAGE\n",
966 vmf->page, vmf->page->mapping, vmf->page->index,
967 (long)vmf->page->flags, page_count(vmf->page),
968 page_private(vmf->page), vmf->virtual_address);
969 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
970 lock_page(vmf->page);
971 cfio->ft_flags |= VM_FAULT_LOCKED;
974 cfio->ft_vmpage = vmf->page;
978 if (cfio->ft_flags & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV)) {
979 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
983 if (cfio->ft_flags & VM_FAULT_OOM) {
984 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
988 if (cfio->ft_flags & VM_FAULT_RETRY)
991 CERROR("Unknown error in page fault %d!\n", cfio->ft_flags);
995 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
996 struct cl_page *page)
998 struct vvp_page *vpg;
999 struct cl_object *clob = cl_io_top(io)->ci_obj;
1001 set_page_dirty(page->cp_vmpage);
1003 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
1004 vvp_write_pending(cl2vvp(clob), vpg);
1007 static int vvp_io_fault_start(const struct lu_env *env,
1008 const struct cl_io_slice *ios)
1010 struct vvp_io *vio = cl2vvp_io(env, ios);
1011 struct cl_io *io = ios->cis_io;
1012 struct cl_object *obj = io->ci_obj;
1013 struct inode *inode = vvp_object_inode(obj);
1014 struct cl_fault_io *fio = &io->u.ci_fault;
1015 struct vvp_fault_io *cfio = &vio->u.fault;
1018 struct page *vmpage = NULL;
1019 struct cl_page *page;
1023 if (fio->ft_executable &&
1024 inode->i_mtime.tv_sec != vio->u.fault.ft_mtime)
1026 " changed while waiting for the page fault lock\n",
1027 PFID(lu_object_fid(&obj->co_lu)));
1029 /* offset of the last byte on the page */
1030 offset = cl_offset(obj, fio->ft_index + 1) - 1;
1031 LASSERT(cl_index(obj, offset) == fio->ft_index);
1032 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1036 /* must return locked page */
1037 if (fio->ft_mkwrite) {
1038 LASSERT(cfio->ft_vmpage);
1039 lock_page(cfio->ft_vmpage);
1041 result = vvp_io_kernel_fault(cfio);
1046 vmpage = cfio->ft_vmpage;
1047 LASSERT(PageLocked(vmpage));
1049 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1050 ll_invalidate_page(vmpage);
1052 size = i_size_read(inode);
1053 /* Though we have already held a cl_lock upon this page, but
1054 * it still can be truncated locally.
1056 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1057 (page_offset(vmpage) > size))) {
1058 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1060 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1067 last_index = cl_index(obj, size - 1);
1069 if (fio->ft_mkwrite) {
1071 * Capture the size while holding the lli_trunc_sem from above
1072 * we want to make sure that we complete the mkwrite action
1073 * while holding this lock. We need to make sure that we are
1074 * not past the end of the file.
1076 if (last_index < fio->ft_index) {
1078 "llite: mkwrite and truncate race happened: %p: 0x%lx 0x%lx\n",
1079 vmpage->mapping, fio->ft_index, last_index);
1081 * We need to return if we are
1082 * passed the end of the file. This will propagate
1083 * up the call stack to ll_page_mkwrite where
1084 * we will return VM_FAULT_NOPAGE. Any non-negative
1085 * value returned here will be silently
1086 * converted to 0. If the vmpage->mapping is null
1087 * the error code would be converted back to ENODATA
1088 * in ll_page_mkwrite0. Thus we return -ENODATA
1089 * to handle both cases
1096 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1098 result = PTR_ERR(page);
1102 /* if page is going to be written, we should add this page into cache
1105 if (fio->ft_mkwrite) {
1106 wait_on_page_writeback(vmpage);
1107 if (!PageDirty(vmpage)) {
1108 struct cl_page_list *plist = &io->ci_queue.c2_qin;
1109 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1112 /* vvp_page_assume() calls wait_on_page_writeback(). */
1113 cl_page_assume(env, io, page);
1115 cl_page_list_init(plist);
1116 cl_page_list_add(plist, page);
1119 if (last_index == vvp_index(vpg))
1120 to = size & ~PAGE_MASK;
1122 /* Do not set Dirty bit here so that in case IO is
1123 * started before the page is really made dirty, we
1124 * still have chance to detect it.
1126 result = cl_io_commit_async(env, io, plist, 0, to,
1127 mkwrite_commit_callback);
1128 LASSERT(cl_page_is_owned(page, io));
1129 cl_page_list_fini(env, plist);
1133 cl_page_discard(env, io, page);
1134 cl_page_disown(env, io, page);
1136 cl_page_put(env, page);
1138 /* we're in big trouble, what can we do now? */
1139 if (result == -EDQUOT)
1143 cl_page_disown(env, io, page);
1149 * The ft_index is only used in the case of
1150 * a mkwrite action. We need to check
1151 * our assertions are correct, since
1152 * we should have caught this above
1154 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1155 if (fio->ft_index == last_index)
1157 * Last page is mapped partially.
1159 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1161 fio->ft_nob = cl_page_size(obj);
1163 lu_ref_add(&page->cp_reference, "fault", io);
1164 fio->ft_page = page;
1167 /* return unlocked vmpage to avoid deadlocking */
1169 unlock_page(vmpage);
1171 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1176 static int vvp_io_fsync_start(const struct lu_env *env,
1177 const struct cl_io_slice *ios)
1179 /* we should mark TOWRITE bit to each dirty page in radix tree to
1180 * verify pages have been written, but this is difficult because of
1186 static int vvp_io_read_page(const struct lu_env *env,
1187 const struct cl_io_slice *ios,
1188 const struct cl_page_slice *slice)
1190 struct cl_io *io = ios->cis_io;
1191 struct vvp_page *vpg = cl2vvp_page(slice);
1192 struct cl_page *page = slice->cpl_page;
1193 struct inode *inode = vvp_object_inode(slice->cpl_obj);
1194 struct ll_sb_info *sbi = ll_i2sbi(inode);
1195 struct ll_file_data *fd = cl2vvp_io(env, ios)->vui_fd;
1196 struct ll_readahead_state *ras = &fd->fd_ras;
1197 struct cl_2queue *queue = &io->ci_queue;
1199 if (sbi->ll_ra_info.ra_max_pages_per_file &&
1200 sbi->ll_ra_info.ra_max_pages)
1201 ras_update(sbi, inode, ras, vvp_index(vpg),
1202 vpg->vpg_defer_uptodate);
1204 if (vpg->vpg_defer_uptodate) {
1205 vpg->vpg_ra_used = 1;
1206 cl_page_export(env, page, 1);
1209 * Add page into the queue even when it is marked uptodate above.
1210 * this will unlock it automatically as part of cl_page_list_disown().
1213 cl_page_list_add(&queue->c2_qin, page);
1214 if (sbi->ll_ra_info.ra_max_pages_per_file &&
1215 sbi->ll_ra_info.ra_max_pages)
1216 ll_readahead(env, io, &queue->c2_qin, ras,
1217 vpg->vpg_defer_uptodate);
1222 static void vvp_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
1224 CLOBINVRNT(env, ios->cis_io->ci_obj,
1225 vvp_object_invariant(ios->cis_io->ci_obj));
1228 static const struct cl_io_operations vvp_io_ops = {
1231 .cio_fini = vvp_io_fini,
1232 .cio_lock = vvp_io_read_lock,
1233 .cio_start = vvp_io_read_start,
1234 .cio_advance = vvp_io_advance,
1237 .cio_fini = vvp_io_fini,
1238 .cio_iter_init = vvp_io_write_iter_init,
1239 .cio_iter_fini = vvp_io_write_iter_fini,
1240 .cio_lock = vvp_io_write_lock,
1241 .cio_start = vvp_io_write_start,
1242 .cio_advance = vvp_io_advance,
1245 .cio_fini = vvp_io_setattr_fini,
1246 .cio_iter_init = vvp_io_setattr_iter_init,
1247 .cio_lock = vvp_io_setattr_lock,
1248 .cio_start = vvp_io_setattr_start,
1249 .cio_end = vvp_io_setattr_end
1252 .cio_fini = vvp_io_fault_fini,
1253 .cio_iter_init = vvp_io_fault_iter_init,
1254 .cio_lock = vvp_io_fault_lock,
1255 .cio_start = vvp_io_fault_start,
1256 .cio_end = vvp_io_end,
1259 .cio_start = vvp_io_fsync_start,
1260 .cio_fini = vvp_io_fini
1263 .cio_fini = vvp_io_fini
1266 .cio_read_page = vvp_io_read_page,
1269 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1272 struct vvp_io *vio = vvp_env_io(env);
1273 struct inode *inode = vvp_object_inode(obj);
1276 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1278 CDEBUG(D_VFSTRACE, DFID
1279 " ignore/verify layout %d/%d, layout version %d restore needed %d\n",
1280 PFID(lu_object_fid(&obj->co_lu)),
1281 io->ci_ignore_layout, io->ci_verify_layout,
1282 vio->vui_layout_gen, io->ci_restore_needed);
1284 CL_IO_SLICE_CLEAN(vio, vui_cl);
1285 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1286 vio->vui_ra_valid = false;
1288 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1290 struct ll_inode_info *lli = ll_i2info(inode);
1292 count = io->u.ci_rw.crw_count;
1293 /* "If nbyte is 0, read() will return 0 and have no other
1294 * results." -- Single Unix Spec
1299 vio->vui_tot_count = count;
1301 /* for read/write, we store the jobid in the inode, and
1302 * it'll be fetched by osc when building RPC.
1304 * it's not accurate if the file is shared by different
1307 lustre_get_jobid(lli->lli_jobid);
1308 } else if (io->ci_type == CIT_SETATTR) {
1309 if (!cl_io_is_trunc(io))
1310 io->ci_lockreq = CILR_MANDATORY;
1313 /* ignore layout change for generic CIT_MISC but not for glimpse.
1314 * io context for glimpse must set ci_verify_layout to true,
1315 * see cl_glimpse_size0() for details.
1317 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1318 io->ci_ignore_layout = 1;
1320 /* Enqueue layout lock and get layout version. We need to do this
1321 * even for operations requiring to open file, such as read and write,
1322 * because it might not grant layout lock in IT_OPEN.
1324 if (result == 0 && !io->ci_ignore_layout) {
1325 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1326 if (result == -ENOENT)
1327 /* If the inode on MDS has been removed, but the objects
1328 * on OSTs haven't been destroyed (async unlink), layout
1329 * fetch will return -ENOENT, we'd ignore this error
1330 * and continue with dirty flush. LU-3230.
1334 CERROR("%s: refresh file layout " DFID " error %d.\n",
1335 ll_get_fsname(inode->i_sb, NULL, 0),
1336 PFID(lu_object_fid(&obj->co_lu)), result);