4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Implementation of cl_io for VVP layer.
38 * Author: Nikita Danilov <nikita.danilov@sun.com>
39 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
42 #define DEBUG_SUBSYSTEM S_LLITE
46 #include <lustre_lite.h>
48 #include "vvp_internal.h"
50 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
51 const struct cl_io_slice *slice);
54 * True, if \a io is a normal io, False for splice_{read,write}
56 int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
58 struct vvp_io *vio = vvp_env_io(env);
60 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
62 return vio->cui_io_subtype == IO_NORMAL;
66 * For swapping layout. The file's layout may have changed.
67 * To avoid populating pages to a wrong stripe, we have to verify the
68 * correctness of layout. It works because swapping layout processes
69 * have to acquire group lock.
71 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
74 struct ll_inode_info *lli = ll_i2info(inode);
75 struct ccc_io *cio = ccc_env_io(env);
78 switch (io->ci_type) {
81 /* don't need lock here to check lli_layout_gen as we have held
82 * extent lock and GROUP lock has to hold to swap layout */
83 if (ll_layout_version_get(lli) != cio->cui_layout_gen) {
84 io->ci_need_restart = 1;
85 /* this will return application a short read/write */
90 /* fault is okay because we've already had a page. */
98 /*****************************************************************************
104 static int vvp_io_write_iter_init(const struct lu_env *env,
105 const struct cl_io_slice *ios)
107 struct ccc_io *cio = cl2ccc_io(env, ios);
109 cl_page_list_init(&cio->u.write.cui_queue);
110 cio->u.write.cui_written = 0;
111 cio->u.write.cui_from = 0;
112 cio->u.write.cui_to = PAGE_SIZE;
117 static void vvp_io_write_iter_fini(const struct lu_env *env,
118 const struct cl_io_slice *ios)
120 struct ccc_io *cio = cl2ccc_io(env, ios);
122 LASSERT(cio->u.write.cui_queue.pl_nr == 0);
125 static int vvp_io_fault_iter_init(const struct lu_env *env,
126 const struct cl_io_slice *ios)
128 struct vvp_io *vio = cl2vvp_io(env, ios);
129 struct inode *inode = ccc_object_inode(ios->cis_obj);
132 cl2ccc_io(env, ios)->cui_fd->fd_file->f_dentry->d_inode);
133 vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime);
137 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
139 struct cl_io *io = ios->cis_io;
140 struct cl_object *obj = io->ci_obj;
141 struct ccc_io *cio = cl2ccc_io(env, ios);
142 struct inode *inode = ccc_object_inode(obj);
144 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
146 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
147 "restore needed %d\n",
148 PFID(lu_object_fid(&obj->co_lu)),
149 io->ci_ignore_layout, io->ci_verify_layout,
150 cio->cui_layout_gen, io->ci_restore_needed);
152 if (io->ci_restore_needed == 1) {
155 /* file was detected release, we need to restore it
156 * before finishing the io
158 rc = ll_layout_restore(inode, 0, OBD_OBJECT_EOF);
159 /* if restore registration failed, no restart,
160 * we will return -ENODATA */
161 /* The layout will change after restore, so we need to
162 * block on layout lock hold by the MDT
163 * as MDT will not send new layout in lvb (see LU-3124)
164 * we have to explicitly fetch it, all this will be done
165 * by ll_layout_refresh()
168 io->ci_restore_needed = 0;
169 io->ci_need_restart = 1;
170 io->ci_verify_layout = 1;
172 io->ci_restore_needed = 1;
173 io->ci_need_restart = 0;
174 io->ci_verify_layout = 0;
179 if (!io->ci_ignore_layout && io->ci_verify_layout) {
182 /* check layout version */
183 ll_layout_refresh(inode, &gen);
184 io->ci_need_restart = cio->cui_layout_gen != gen;
185 if (io->ci_need_restart) {
187 DFID" layout changed from %d to %d.\n",
188 PFID(lu_object_fid(&obj->co_lu)),
189 cio->cui_layout_gen, gen);
190 /* today successful restore is the only possible
192 /* restore was done, clear restoring state */
193 ll_i2info(ccc_object_inode(obj))->lli_flags &=
194 ~LLIF_FILE_RESTORING;
199 static void vvp_io_fault_fini(const struct lu_env *env,
200 const struct cl_io_slice *ios)
202 struct cl_io *io = ios->cis_io;
203 struct cl_page *page = io->u.ci_fault.ft_page;
205 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
208 lu_ref_del(&page->cp_reference, "fault", io);
209 cl_page_put(env, page);
210 io->u.ci_fault.ft_page = NULL;
212 vvp_io_fini(env, ios);
215 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
218 * we only want to hold PW locks if the mmap() can generate
219 * writes back to the file and that only happens in shared
222 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
227 static int vvp_mmap_locks(const struct lu_env *env,
228 struct ccc_io *vio, struct cl_io *io)
230 struct ccc_thread_info *cti = ccc_env_info(env);
231 struct mm_struct *mm = current->mm;
232 struct vm_area_struct *vma;
233 struct cl_lock_descr *descr = &cti->cti_descr;
234 ldlm_policy_data_t policy;
241 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
243 if (!cl_is_normalio(env, io))
246 if (vio->cui_iov == NULL) /* nfs or loop back device write */
249 /* No MM (e.g. NFS)? No vmas too. */
253 for (seg = 0; seg < vio->cui_nrsegs; seg++) {
254 const struct iovec *iv = &vio->cui_iov[seg];
256 addr = (unsigned long)iv->iov_base;
261 count += addr & (~CFS_PAGE_MASK);
262 addr &= CFS_PAGE_MASK;
264 down_read(&mm->mmap_sem);
265 while((vma = our_vma(mm, addr, count)) != NULL) {
266 struct inode *inode = vma->vm_file->f_dentry->d_inode;
267 int flags = CEF_MUST;
269 if (ll_file_nolock(vma->vm_file)) {
271 * For no lock case, a lockless lock will be
278 * XXX: Required lock mode can be weakened: CIT_WRITE
279 * io only ever reads user level buffer, and CIT_READ
282 policy_from_vma(&policy, vma, addr, count);
283 descr->cld_mode = vvp_mode_from_vma(vma);
284 descr->cld_obj = ll_i2info(inode)->lli_clob;
285 descr->cld_start = cl_index(descr->cld_obj,
286 policy.l_extent.start);
287 descr->cld_end = cl_index(descr->cld_obj,
288 policy.l_extent.end);
289 descr->cld_enq_flags = flags;
290 result = cl_io_lock_alloc_add(env, io, descr);
292 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
293 descr->cld_mode, descr->cld_start,
297 up_read(&mm->mmap_sem);
301 if (vma->vm_end - addr >= count)
304 count -= vma->vm_end - addr;
307 up_read(&mm->mmap_sem);
312 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
313 enum cl_lock_mode mode, loff_t start, loff_t end)
315 struct ccc_io *cio = ccc_env_io(env);
319 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
322 ccc_io_update_iov(env, cio, io);
324 if (io->u.ci_rw.crw_nonblock)
325 ast_flags |= CEF_NONBLOCK;
326 result = vvp_mmap_locks(env, cio, io);
328 result = ccc_io_one_lock(env, io, ast_flags, mode, start, end);
332 static int vvp_io_read_lock(const struct lu_env *env,
333 const struct cl_io_slice *ios)
335 struct cl_io *io = ios->cis_io;
336 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
340 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
341 rd->crw_pos + rd->crw_count - 1);
345 static int vvp_io_fault_lock(const struct lu_env *env,
346 const struct cl_io_slice *ios)
348 struct cl_io *io = ios->cis_io;
349 struct vvp_io *vio = cl2vvp_io(env, ios);
351 * XXX LDLM_FL_CBPENDING
353 return ccc_io_one_lock_index
354 (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma),
355 io->u.ci_fault.ft_index, io->u.ci_fault.ft_index);
358 static int vvp_io_write_lock(const struct lu_env *env,
359 const struct cl_io_slice *ios)
361 struct cl_io *io = ios->cis_io;
365 if (io->u.ci_wr.wr_append) {
367 end = OBD_OBJECT_EOF;
369 start = io->u.ci_wr.wr.crw_pos;
370 end = start + io->u.ci_wr.wr.crw_count - 1;
372 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
375 static int vvp_io_setattr_iter_init(const struct lu_env *env,
376 const struct cl_io_slice *ios)
382 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
384 * Handles "lockless io" mode when extent locking is done by server.
386 static int vvp_io_setattr_lock(const struct lu_env *env,
387 const struct cl_io_slice *ios)
389 struct ccc_io *cio = ccc_env_io(env);
390 struct cl_io *io = ios->cis_io;
394 if (cl_io_is_trunc(io)) {
395 new_size = io->u.ci_setattr.sa_attr.lvb_size;
397 enqflags = CEF_DISCARD_DATA;
399 if ((io->u.ci_setattr.sa_attr.lvb_mtime >=
400 io->u.ci_setattr.sa_attr.lvb_ctime) ||
401 (io->u.ci_setattr.sa_attr.lvb_atime >=
402 io->u.ci_setattr.sa_attr.lvb_ctime))
406 cio->u.setattr.cui_local_lock = SETATTR_EXTENT_LOCK;
407 return ccc_io_one_lock(env, io, enqflags, CLM_WRITE,
408 new_size, OBD_OBJECT_EOF);
411 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
416 * Only ll_inode_size_lock is taken at this level.
418 ll_inode_size_lock(inode);
419 result = inode_newsize_ok(inode, size);
421 ll_inode_size_unlock(inode);
424 i_size_write(inode, size);
426 ll_truncate_pagecache(inode, size);
427 ll_inode_size_unlock(inode);
431 static int vvp_io_setattr_trunc(const struct lu_env *env,
432 const struct cl_io_slice *ios,
433 struct inode *inode, loff_t size)
435 inode_dio_wait(inode);
439 static int vvp_io_setattr_time(const struct lu_env *env,
440 const struct cl_io_slice *ios)
442 struct cl_io *io = ios->cis_io;
443 struct cl_object *obj = io->ci_obj;
444 struct cl_attr *attr = ccc_env_thread_attr(env);
446 unsigned valid = CAT_CTIME;
448 cl_object_attr_lock(obj);
449 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
450 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
451 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
454 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
455 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
458 result = cl_object_attr_set(env, obj, attr, valid);
459 cl_object_attr_unlock(obj);
464 static int vvp_io_setattr_start(const struct lu_env *env,
465 const struct cl_io_slice *ios)
467 struct cl_io *io = ios->cis_io;
468 struct inode *inode = ccc_object_inode(io->ci_obj);
471 mutex_lock(&inode->i_mutex);
472 if (cl_io_is_trunc(io))
473 result = vvp_io_setattr_trunc(env, ios, inode,
474 io->u.ci_setattr.sa_attr.lvb_size);
476 result = vvp_io_setattr_time(env, ios);
480 static void vvp_io_setattr_end(const struct lu_env *env,
481 const struct cl_io_slice *ios)
483 struct cl_io *io = ios->cis_io;
484 struct inode *inode = ccc_object_inode(io->ci_obj);
486 if (cl_io_is_trunc(io)) {
487 /* Truncate in memory pages - they must be clean pages
488 * because osc has already notified to destroy osc_extents. */
489 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
490 inode_dio_write_done(inode);
492 mutex_unlock(&inode->i_mutex);
495 static void vvp_io_setattr_fini(const struct lu_env *env,
496 const struct cl_io_slice *ios)
498 vvp_io_fini(env, ios);
501 static int vvp_io_read_start(const struct lu_env *env,
502 const struct cl_io_slice *ios)
504 struct vvp_io *vio = cl2vvp_io(env, ios);
505 struct ccc_io *cio = cl2ccc_io(env, ios);
506 struct cl_io *io = ios->cis_io;
507 struct cl_object *obj = io->ci_obj;
508 struct inode *inode = ccc_object_inode(obj);
509 struct ll_ra_read *bead = &vio->cui_bead;
510 struct file *file = cio->cui_fd->fd_file;
513 loff_t pos = io->u.ci_rd.rd.crw_pos;
514 long cnt = io->u.ci_rd.rd.crw_count;
515 long tot = cio->cui_tot_count;
518 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
520 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
522 if (!can_populate_pages(env, io, inode))
525 result = ccc_prep_size(env, obj, io, pos, tot, &exceed);
528 else if (exceed != 0)
531 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
532 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
533 inode->i_ino, cnt, pos, i_size_read(inode));
535 /* turn off the kernel's read-ahead */
536 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
538 /* initialize read-ahead window once per syscall */
539 if (!vio->cui_ra_window_set) {
540 vio->cui_ra_window_set = 1;
541 bead->lrr_start = cl_index(obj, pos);
542 bead->lrr_count = cl_index(obj, tot + PAGE_CACHE_SIZE - 1);
543 ll_ra_read_in(file, bead);
548 switch (vio->cui_io_subtype) {
550 LASSERT(cio->cui_iocb->ki_pos == pos);
551 result = generic_file_aio_read(cio->cui_iocb,
552 cio->cui_iov, cio->cui_nrsegs,
553 cio->cui_iocb->ki_pos);
556 result = generic_file_splice_read(file, &pos,
557 vio->u.splice.cui_pipe, cnt,
558 vio->u.splice.cui_flags);
559 /* LU-1109: do splice read stripe by stripe otherwise if it
560 * may make nfsd stuck if this read occupied all internal pipe
565 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
573 io->ci_nob += result;
574 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, cio->cui_fd,
582 static void vvp_io_read_fini(const struct lu_env *env, const struct cl_io_slice *ios)
584 struct vvp_io *vio = cl2vvp_io(env, ios);
585 struct ccc_io *cio = cl2ccc_io(env, ios);
587 if (vio->cui_ra_window_set)
588 ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
590 vvp_io_fini(env, ios);
593 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
594 struct cl_page_list *plist, int from, int to)
596 struct cl_2queue *queue = &io->ci_queue;
597 struct cl_page *page;
598 unsigned int bytes = 0;
602 if (plist->pl_nr == 0)
605 if (from > 0 || to != PAGE_SIZE) {
606 page = cl_page_list_first(plist);
607 if (plist->pl_nr == 1) {
608 cl_page_clip(env, page, from, to);
609 } else if (from > 0) {
610 cl_page_clip(env, page, from, PAGE_SIZE);
612 page = cl_page_list_last(plist);
613 cl_page_clip(env, page, 0, to);
617 cl_2queue_init(queue);
618 cl_page_list_splice(plist, &queue->c2_qin);
619 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
621 /* plist is not sorted any more */
622 cl_page_list_splice(&queue->c2_qin, plist);
623 cl_page_list_splice(&queue->c2_qout, plist);
624 cl_2queue_fini(env, queue);
627 /* calculate bytes */
628 bytes = plist->pl_nr << PAGE_SHIFT;
629 bytes -= from + PAGE_SIZE - to;
631 while (plist->pl_nr > 0) {
632 page = cl_page_list_first(plist);
633 cl_page_list_del(env, plist, page);
635 cl_page_clip(env, page, 0, PAGE_SIZE);
637 SetPageUptodate(cl_page_vmpage(page));
638 cl_page_disown(env, io, page);
640 /* held in ll_cl_init() */
641 lu_ref_del(&page->cp_reference, "cl_io", io);
642 cl_page_put(env, page);
646 RETURN(bytes > 0 ? bytes : rc);
649 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
650 struct cl_page *page)
653 struct page *vmpage = page->cp_vmpage;
654 struct cl_object *clob = cl_io_top(io)->ci_obj;
656 SetPageUptodate(vmpage);
657 set_page_dirty(vmpage);
659 cp = cl2ccc_page(cl_object_page_slice(clob, page));
660 vvp_write_pending(cl2ccc(clob), cp);
662 cl_page_disown(env, io, page);
664 /* held in ll_cl_init() */
665 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
666 cl_page_put(env, page);
669 /* make sure the page list is contiguous */
670 static bool page_list_sanity_check(struct cl_object *obj,
671 struct cl_page_list *plist)
673 struct cl_page *page;
674 pgoff_t index = CL_PAGE_EOF;
676 cl_page_list_for_each(page, plist) {
677 struct ccc_page *cp = cl_object_page_slice(obj, page);
679 if (index == CL_PAGE_EOF) {
680 index = ccc_index(cp);
685 if (index == ccc_index(cp))
693 /* Return how many bytes have queued or written */
694 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
696 struct cl_object *obj = io->ci_obj;
697 struct inode *inode = ccc_object_inode(obj);
698 struct ccc_io *cio = ccc_env_io(env);
699 struct cl_page_list *queue = &cio->u.write.cui_queue;
700 struct cl_page *page;
703 unsigned int npages = cio->u.write.cui_queue.pl_nr;
709 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
710 npages, cio->u.write.cui_from, cio->u.write.cui_to);
712 LASSERT(page_list_sanity_check(obj, queue));
714 /* submit IO with async write */
715 rc = cl_io_commit_async(env, io, queue,
716 cio->u.write.cui_from, cio->u.write.cui_to,
717 write_commit_callback);
718 npages -= queue->pl_nr; /* already committed pages */
720 /* calculate how many bytes were written */
721 bytes = npages << PAGE_SHIFT;
724 bytes -= cio->u.write.cui_from;
725 if (queue->pl_nr == 0) /* last page */
726 bytes -= PAGE_SIZE - cio->u.write.cui_to;
727 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
729 cio->u.write.cui_written += bytes;
731 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
732 npages, bytes, cio->u.write.cui_written);
734 /* the first page must have been written. */
735 cio->u.write.cui_from = 0;
737 LASSERT(page_list_sanity_check(obj, queue));
738 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
740 /* out of quota, try sync write */
741 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
742 rc = vvp_io_commit_sync(env, io, queue,
743 cio->u.write.cui_from,
744 cio->u.write.cui_to);
746 cio->u.write.cui_written += rc;
751 /* update inode size */
752 ll_merge_lvb(env, inode);
754 /* Now the pages in queue were failed to commit, discard them
755 * unless they were dirtied before. */
756 while (queue->pl_nr > 0) {
757 page = cl_page_list_first(queue);
758 cl_page_list_del(env, queue, page);
760 if (!PageDirty(cl_page_vmpage(page)))
761 cl_page_discard(env, io, page);
763 cl_page_disown(env, io, page);
765 /* held in ll_cl_init() */
766 lu_ref_del(&page->cp_reference, "cl_io", io);
767 cl_page_put(env, page);
769 cl_page_list_fini(env, queue);
774 static int vvp_io_write_start(const struct lu_env *env,
775 const struct cl_io_slice *ios)
777 struct ccc_io *cio = cl2ccc_io(env, ios);
778 struct cl_io *io = ios->cis_io;
779 struct cl_object *obj = io->ci_obj;
780 struct inode *inode = ccc_object_inode(obj);
782 loff_t pos = io->u.ci_wr.wr.crw_pos;
783 size_t cnt = io->u.ci_wr.wr.crw_count;
787 if (!can_populate_pages(env, io, inode))
790 if (cl_io_is_append(io)) {
792 * PARALLEL IO This has to be changed for parallel IO doing
793 * out-of-order writes.
795 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
796 cio->cui_iocb->ki_pos = pos;
798 LASSERT(cio->cui_iocb->ki_pos == pos);
801 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
803 if (cio->cui_iov == NULL) /* from a temp io in ll_cl_init(). */
806 result = generic_file_aio_write(cio->cui_iocb,
807 cio->cui_iov, cio->cui_nrsegs,
808 cio->cui_iocb->ki_pos);
810 result = vvp_io_write_commit(env, io);
811 if (cio->u.write.cui_written > 0) {
812 result = cio->u.write.cui_written;
813 io->ci_nob += result;
815 CDEBUG(D_VFSTRACE, "write: nob %zd, result: %zd\n",
820 struct ll_inode_info *lli = ll_i2info(inode);
822 spin_lock(&lli->lli_lock);
823 lli->lli_flags |= LLIF_DATA_MODIFIED;
824 spin_unlock(&lli->lli_lock);
828 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
829 cio->cui_fd, pos, result, WRITE);
836 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
838 struct vm_fault *vmf = cfio->fault.ft_vmf;
840 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, vmf);
841 cfio->fault.ft_flags_valid = 1;
844 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
845 vmf->virtual_address);
846 if (unlikely(!(cfio->fault.ft_flags & VM_FAULT_LOCKED))) {
847 lock_page(vmf->page);
848 cfio->fault.ft_flags |= VM_FAULT_LOCKED;
851 cfio->ft_vmpage = vmf->page;
855 if (cfio->fault.ft_flags & VM_FAULT_SIGBUS) {
856 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
860 if (cfio->fault.ft_flags & VM_FAULT_OOM) {
861 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
865 if (cfio->fault.ft_flags & VM_FAULT_RETRY)
868 CERROR("unknow error in page fault %d!\n", cfio->fault.ft_flags);
872 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
873 struct cl_page *page)
876 struct cl_object *clob = cl_io_top(io)->ci_obj;
878 set_page_dirty(page->cp_vmpage);
880 cp = cl2ccc_page(cl_object_page_slice(clob, page));
881 vvp_write_pending(cl2ccc(clob), cp);
884 static int vvp_io_fault_start(const struct lu_env *env,
885 const struct cl_io_slice *ios)
887 struct vvp_io *vio = cl2vvp_io(env, ios);
888 struct cl_io *io = ios->cis_io;
889 struct cl_object *obj = io->ci_obj;
890 struct inode *inode = ccc_object_inode(obj);
891 struct cl_fault_io *fio = &io->u.ci_fault;
892 struct vvp_fault_io *cfio = &vio->u.fault;
895 struct page *vmpage = NULL;
896 struct cl_page *page;
901 if (fio->ft_executable &&
902 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
904 " changed while waiting for the page fault lock\n",
905 PFID(lu_object_fid(&obj->co_lu)));
907 /* offset of the last byte on the page */
908 offset = cl_offset(obj, fio->ft_index + 1) - 1;
909 LASSERT(cl_index(obj, offset) == fio->ft_index);
910 result = ccc_prep_size(env, obj, io, 0, offset + 1, NULL);
914 /* must return locked page */
915 if (fio->ft_mkwrite) {
916 LASSERT(cfio->ft_vmpage != NULL);
917 lock_page(cfio->ft_vmpage);
919 result = vvp_io_kernel_fault(cfio);
924 vmpage = cfio->ft_vmpage;
925 LASSERT(PageLocked(vmpage));
927 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
928 ll_invalidate_page(vmpage);
930 size = i_size_read(inode);
931 /* Though we have already held a cl_lock upon this page, but
932 * it still can be truncated locally. */
933 if (unlikely((vmpage->mapping != inode->i_mapping) ||
934 (page_offset(vmpage) > size))) {
935 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
937 /* return +1 to stop cl_io_loop() and ll_fault() will catch
939 GOTO(out, result = +1);
942 last_index = cl_index(obj, size - 1);
944 if (fio->ft_mkwrite ) {
946 * Capture the size while holding the lli_trunc_sem from above
947 * we want to make sure that we complete the mkwrite action
948 * while holding this lock. We need to make sure that we are
949 * not past the end of the file.
951 if (last_index < fio->ft_index) {
953 "llite: mkwrite and truncate race happened: "
955 vmpage->mapping,fio->ft_index,last_index);
957 * We need to return if we are
958 * passed the end of the file. This will propagate
959 * up the call stack to ll_page_mkwrite where
960 * we will return VM_FAULT_NOPAGE. Any non-negative
961 * value returned here will be silently
962 * converted to 0. If the vmpage->mapping is null
963 * the error code would be converted back to ENODATA
964 * in ll_page_mkwrite0. Thus we return -ENODATA
965 * to handle both cases
967 GOTO(out, result = -ENODATA);
971 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
973 GOTO(out, result = PTR_ERR(page));
975 /* if page is going to be written, we should add this page into cache
977 if (fio->ft_mkwrite) {
978 wait_on_page_writeback(vmpage);
979 if (!PageDirty(vmpage)) {
980 struct cl_page_list *plist = &io->ci_queue.c2_qin;
981 struct ccc_page *cp = cl_object_page_slice(obj, page);
984 /* vvp_page_assume() calls wait_on_page_writeback(). */
985 cl_page_assume(env, io, page);
987 cl_page_list_init(plist);
988 cl_page_list_add(plist, page);
991 if (last_index == ccc_index(cp))
992 to = size & ~CFS_PAGE_MASK;
994 /* Do not set Dirty bit here so that in case IO is
995 * started before the page is really made dirty, we
996 * still have chance to detect it. */
997 result = cl_io_commit_async(env, io, plist, 0, to,
998 mkwrite_commit_callback);
999 LASSERT(cl_page_is_owned(page, io));
1000 cl_page_list_fini(env, plist);
1004 cl_page_discard(env, io, page);
1005 cl_page_disown(env, io, page);
1007 cl_page_put(env, page);
1009 /* we're in big trouble, what can we do now? */
1010 if (result == -EDQUOT)
1014 cl_page_disown(env, io, page);
1019 * The ft_index is only used in the case of
1020 * a mkwrite action. We need to check
1021 * our assertions are correct, since
1022 * we should have caught this above
1024 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1025 if (fio->ft_index == last_index)
1027 * Last page is mapped partially.
1029 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1031 fio->ft_nob = cl_page_size(obj);
1033 lu_ref_add(&page->cp_reference, "fault", io);
1034 fio->ft_page = page;
1038 /* return unlocked vmpage to avoid deadlocking */
1040 unlock_page(vmpage);
1041 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
1045 static int vvp_io_fsync_start(const struct lu_env *env,
1046 const struct cl_io_slice *ios)
1048 /* we should mark TOWRITE bit to each dirty page in radix tree to
1049 * verify pages have been written, but this is difficult because of
1054 static int vvp_io_read_page(const struct lu_env *env,
1055 const struct cl_io_slice *ios,
1056 const struct cl_page_slice *slice)
1058 struct cl_io *io = ios->cis_io;
1059 struct ccc_page *cp = cl2ccc_page(slice);
1060 struct cl_page *page = slice->cpl_page;
1061 struct inode *inode = ccc_object_inode(slice->cpl_obj);
1062 struct ll_sb_info *sbi = ll_i2sbi(inode);
1063 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
1064 struct ll_readahead_state *ras = &fd->fd_ras;
1065 struct cl_2queue *queue = &io->ci_queue;
1069 if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
1070 sbi->ll_ra_info.ra_max_pages > 0)
1071 ras_update(sbi, inode, ras, ccc_index(cp),
1072 cp->cpg_defer_uptodate);
1074 if (cp->cpg_defer_uptodate) {
1075 cp->cpg_ra_used = 1;
1076 cl_page_export(env, page, 1);
1080 * Add page into the queue even when it is marked uptodate above.
1081 * this will unlock it automatically as part of cl_page_list_disown().
1083 cl_2queue_add(queue, page);
1084 if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
1085 sbi->ll_ra_info.ra_max_pages > 0)
1086 ll_readahead(env, io, &queue->c2_qin, ras,
1087 cp->cpg_defer_uptodate);
1092 static const struct cl_io_operations vvp_io_ops = {
1095 .cio_fini = vvp_io_read_fini,
1096 .cio_lock = vvp_io_read_lock,
1097 .cio_start = vvp_io_read_start,
1098 .cio_advance = ccc_io_advance
1101 .cio_fini = vvp_io_fini,
1102 .cio_iter_init = vvp_io_write_iter_init,
1103 .cio_iter_fini = vvp_io_write_iter_fini,
1104 .cio_lock = vvp_io_write_lock,
1105 .cio_start = vvp_io_write_start,
1106 .cio_advance = ccc_io_advance
1109 .cio_fini = vvp_io_setattr_fini,
1110 .cio_iter_init = vvp_io_setattr_iter_init,
1111 .cio_lock = vvp_io_setattr_lock,
1112 .cio_start = vvp_io_setattr_start,
1113 .cio_end = vvp_io_setattr_end
1116 .cio_fini = vvp_io_fault_fini,
1117 .cio_iter_init = vvp_io_fault_iter_init,
1118 .cio_lock = vvp_io_fault_lock,
1119 .cio_start = vvp_io_fault_start,
1120 .cio_end = ccc_io_end
1123 .cio_start = vvp_io_fsync_start,
1124 .cio_fini = vvp_io_fini
1127 .cio_fini = vvp_io_fini
1130 .cio_read_page = vvp_io_read_page,
1133 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1136 struct vvp_io *vio = vvp_env_io(env);
1137 struct ccc_io *cio = ccc_env_io(env);
1138 struct inode *inode = ccc_object_inode(obj);
1141 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
1144 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1145 "restore needed %d\n",
1146 PFID(lu_object_fid(&obj->co_lu)),
1147 io->ci_ignore_layout, io->ci_verify_layout,
1148 cio->cui_layout_gen, io->ci_restore_needed);
1150 CL_IO_SLICE_CLEAN(cio, cui_cl);
1151 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
1152 vio->cui_ra_window_set = 0;
1154 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1156 struct ll_inode_info *lli = ll_i2info(inode);
1158 count = io->u.ci_rw.crw_count;
1159 /* "If nbyte is 0, read() will return 0 and have no other
1160 * results." -- Single Unix Spec */
1164 cio->cui_tot_count = count;
1165 cio->cui_tot_nrsegs = 0;
1168 /* for read/write, we store the jobid in the inode, and
1169 * it'll be fetched by osc when building RPC.
1171 * it's not accurate if the file is shared by different
1174 lustre_get_jobid(lli->lli_jobid);
1175 } else if (io->ci_type == CIT_SETATTR) {
1176 if (!cl_io_is_trunc(io))
1177 io->ci_lockreq = CILR_MANDATORY;
1180 /* ignore layout change for generic CIT_MISC but not for glimpse.
1181 * io context for glimpse must set ci_verify_layout to true,
1182 * see cl_glimpse_size0() for details. */
1183 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1184 io->ci_ignore_layout = 1;
1186 /* Enqueue layout lock and get layout version. We need to do this
1187 * even for operations requiring to open file, such as read and write,
1188 * because it might not grant layout lock in IT_OPEN. */
1189 if (result == 0 && !io->ci_ignore_layout) {
1190 result = ll_layout_refresh(inode, &cio->cui_layout_gen);
1191 if (result == -ENOENT)
1192 /* If the inode on MDS has been removed, but the objects
1193 * on OSTs haven't been destroyed (async unlink), layout
1194 * fetch will return -ENOENT, we'd ingore this error
1195 * and continue with dirty flush. LU-3230. */
1198 CERROR("%s: refresh file layout " DFID " error %d.\n",
1199 ll_get_fsname(inode->i_sb, NULL, 0),
1200 PFID(lu_object_fid(&obj->co_lu)), result);
1206 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1207 const struct cl_io_slice *slice)
1209 /* Caling just for assertion */
1210 cl2ccc_io(env, slice);
1211 return vvp_env_io(env);