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, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Implementation of cl_io for VVP layer.
38 * Author: Nikita Danilov <nikita.danilov@sun.com>
39 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
42 #define DEBUG_SUBSYSTEM S_LLITE
46 #include "llite_internal.h"
47 #include "vvp_internal.h"
50 * True, if \a io is a normal io, False for splice_{read,write}
52 int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
54 struct vvp_io *vio = vvp_env_io(env);
56 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
58 return vio->vui_io_subtype == IO_NORMAL;
62 * For swapping layout. The file's layout may have changed.
63 * To avoid populating pages to a wrong stripe, we have to verify the
64 * correctness of layout. It works because swapping layout processes
65 * have to acquire group lock.
67 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
70 struct ll_inode_info *lli = ll_i2info(inode);
71 struct vvp_io *vio = vvp_env_io(env);
74 switch (io->ci_type) {
77 /* don't need lock here to check lli_layout_gen as we have held
78 * extent lock and GROUP lock has to hold to swap layout */
79 if (ll_layout_version_get(lli) != vio->vui_layout_gen) {
80 io->ci_need_restart = 1;
81 /* this will return application a short read/write */
86 /* fault is okay because we've already had a page. */
94 /*****************************************************************************
100 static int vvp_io_write_iter_init(const struct lu_env *env,
101 const struct cl_io_slice *ios)
103 struct vvp_io *vio = cl2vvp_io(env, ios);
105 cl_page_list_init(&vio->u.write.vui_queue);
106 vio->u.write.vui_written = 0;
107 vio->u.write.vui_from = 0;
108 vio->u.write.vui_to = PAGE_SIZE;
113 static void vvp_io_write_iter_fini(const struct lu_env *env,
114 const struct cl_io_slice *ios)
116 struct vvp_io *vio = cl2vvp_io(env, ios);
118 LASSERT(vio->u.write.vui_queue.pl_nr == 0);
121 static int vvp_io_fault_iter_init(const struct lu_env *env,
122 const struct cl_io_slice *ios)
124 struct vvp_io *vio = cl2vvp_io(env, ios);
125 struct inode *inode = vvp_object_inode(ios->cis_obj);
127 LASSERT(inode == vio->vui_fd->fd_file->f_dentry->d_inode);
128 vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime);
133 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
135 struct cl_io *io = ios->cis_io;
136 struct cl_object *obj = io->ci_obj;
137 struct vvp_io *vio = cl2vvp_io(env, ios);
138 struct inode *inode = vvp_object_inode(obj);
140 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
142 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
143 "restore needed %d\n",
144 PFID(lu_object_fid(&obj->co_lu)),
145 io->ci_ignore_layout, io->ci_verify_layout,
146 vio->vui_layout_gen, io->ci_restore_needed);
148 if (io->ci_restore_needed == 1) {
151 /* file was detected release, we need to restore it
152 * before finishing the io
154 rc = ll_layout_restore(inode, 0, OBD_OBJECT_EOF);
155 /* if restore registration failed, no restart,
156 * we will return -ENODATA */
157 /* The layout will change after restore, so we need to
158 * block on layout lock hold by the MDT
159 * as MDT will not send new layout in lvb (see LU-3124)
160 * we have to explicitly fetch it, all this will be done
161 * by ll_layout_refresh()
164 io->ci_restore_needed = 0;
165 io->ci_need_restart = 1;
166 io->ci_verify_layout = 1;
168 io->ci_restore_needed = 1;
169 io->ci_need_restart = 0;
170 io->ci_verify_layout = 0;
175 if (!io->ci_ignore_layout && io->ci_verify_layout) {
178 /* check layout version */
179 ll_layout_refresh(inode, &gen);
180 io->ci_need_restart = vio->vui_layout_gen != gen;
181 if (io->ci_need_restart) {
183 DFID" layout changed from %d to %d.\n",
184 PFID(lu_object_fid(&obj->co_lu)),
185 vio->vui_layout_gen, gen);
186 /* today successful restore is the only possible
188 /* restore was done, clear restoring state */
189 ll_i2info(vvp_object_inode(obj))->lli_flags &=
190 ~LLIF_FILE_RESTORING;
195 static void vvp_io_fault_fini(const struct lu_env *env,
196 const struct cl_io_slice *ios)
198 struct cl_io *io = ios->cis_io;
199 struct cl_page *page = io->u.ci_fault.ft_page;
201 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
204 lu_ref_del(&page->cp_reference, "fault", io);
205 cl_page_put(env, page);
206 io->u.ci_fault.ft_page = NULL;
208 vvp_io_fini(env, ios);
211 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
214 * we only want to hold PW locks if the mmap() can generate
215 * writes back to the file and that only happens in shared
218 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
223 static int vvp_mmap_locks(const struct lu_env *env,
224 struct vvp_io *vio, struct cl_io *io)
226 struct ccc_thread_info *cti = ccc_env_info(env);
227 struct mm_struct *mm = current->mm;
228 struct vm_area_struct *vma;
229 struct cl_lock_descr *descr = &cti->cti_descr;
230 ldlm_policy_data_t policy;
237 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
239 if (!cl_is_normalio(env, io))
242 if (vio->vui_iov == NULL) /* nfs or loop back device write */
245 /* No MM (e.g. NFS)? No vmas too. */
249 for (seg = 0; seg < vio->vui_nrsegs; seg++) {
250 const struct iovec *iv = &vio->vui_iov[seg];
252 addr = (unsigned long)iv->iov_base;
257 count += addr & (~CFS_PAGE_MASK);
258 addr &= CFS_PAGE_MASK;
260 down_read(&mm->mmap_sem);
261 while((vma = our_vma(mm, addr, count)) != NULL) {
262 struct inode *inode = vma->vm_file->f_dentry->d_inode;
263 int flags = CEF_MUST;
265 if (ll_file_nolock(vma->vm_file)) {
267 * For no lock case is not allowed for mmap
274 * XXX: Required lock mode can be weakened: CIT_WRITE
275 * io only ever reads user level buffer, and CIT_READ
278 policy_from_vma(&policy, vma, addr, count);
279 descr->cld_mode = vvp_mode_from_vma(vma);
280 descr->cld_obj = ll_i2info(inode)->lli_clob;
281 descr->cld_start = cl_index(descr->cld_obj,
282 policy.l_extent.start);
283 descr->cld_end = cl_index(descr->cld_obj,
284 policy.l_extent.end);
285 descr->cld_enq_flags = flags;
286 result = cl_io_lock_alloc_add(env, io, descr);
288 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
289 descr->cld_mode, descr->cld_start,
295 if (vma->vm_end - addr >= count)
298 count -= vma->vm_end - addr;
301 up_read(&mm->mmap_sem);
308 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
309 enum cl_lock_mode mode, loff_t start, loff_t end)
311 struct vvp_io *vio = vvp_env_io(env);
315 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
318 vvp_io_update_iov(env, vio, io);
320 if (io->u.ci_rw.crw_nonblock)
321 ast_flags |= CEF_NONBLOCK;
323 result = vvp_mmap_locks(env, vio, io);
325 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
330 static int vvp_io_read_lock(const struct lu_env *env,
331 const struct cl_io_slice *ios)
333 struct cl_io *io = ios->cis_io;
334 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
338 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
339 rd->crw_pos + rd->crw_count - 1);
343 static int vvp_io_fault_lock(const struct lu_env *env,
344 const struct cl_io_slice *ios)
346 struct cl_io *io = ios->cis_io;
347 struct vvp_io *vio = cl2vvp_io(env, ios);
349 * XXX LDLM_FL_CBPENDING
351 return vvp_io_one_lock_index(env,
353 vvp_mode_from_vma(vio->u.fault.ft_vma),
354 io->u.ci_fault.ft_index,
355 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 vvp_io *vio = vvp_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))
407 vio->u.setattr.vui_local_lock = SETATTR_EXTENT_LOCK;
409 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
410 new_size, OBD_OBJECT_EOF);
413 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
418 * Only ll_inode_size_lock is taken at this level.
420 ll_inode_size_lock(inode);
421 result = inode_newsize_ok(inode, size);
423 ll_inode_size_unlock(inode);
426 i_size_write(inode, size);
428 ll_truncate_pagecache(inode, size);
429 ll_inode_size_unlock(inode);
433 static int vvp_io_setattr_trunc(const struct lu_env *env,
434 const struct cl_io_slice *ios,
435 struct inode *inode, loff_t size)
437 inode_dio_wait(inode);
441 static int vvp_io_setattr_time(const struct lu_env *env,
442 const struct cl_io_slice *ios)
444 struct cl_io *io = ios->cis_io;
445 struct cl_object *obj = io->ci_obj;
446 struct cl_attr *attr = ccc_env_thread_attr(env);
448 unsigned valid = CAT_CTIME;
450 cl_object_attr_lock(obj);
451 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
452 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
453 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
456 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
457 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
460 result = cl_object_attr_update(env, obj, attr, valid);
461 cl_object_attr_unlock(obj);
466 static int vvp_io_setattr_start(const struct lu_env *env,
467 const struct cl_io_slice *ios)
469 struct cl_io *io = ios->cis_io;
470 struct inode *inode = vvp_object_inode(io->ci_obj);
473 mutex_lock(&inode->i_mutex);
474 if (cl_io_is_trunc(io))
475 result = vvp_io_setattr_trunc(env, ios, inode,
476 io->u.ci_setattr.sa_attr.lvb_size);
478 result = vvp_io_setattr_time(env, ios);
482 static void vvp_io_setattr_end(const struct lu_env *env,
483 const struct cl_io_slice *ios)
485 struct cl_io *io = ios->cis_io;
486 struct inode *inode = vvp_object_inode(io->ci_obj);
488 if (cl_io_is_trunc(io)) {
489 /* Truncate in memory pages - they must be clean pages
490 * because osc has already notified to destroy osc_extents. */
491 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
492 inode_dio_write_done(inode);
494 mutex_unlock(&inode->i_mutex);
497 static void vvp_io_setattr_fini(const struct lu_env *env,
498 const struct cl_io_slice *ios)
500 vvp_io_fini(env, ios);
503 static int vvp_io_read_start(const struct lu_env *env,
504 const struct cl_io_slice *ios)
506 struct vvp_io *vio = cl2vvp_io(env, ios);
507 struct cl_io *io = ios->cis_io;
508 struct cl_object *obj = io->ci_obj;
509 struct inode *inode = vvp_object_inode(obj);
510 struct file *file = vio->vui_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 = vio->vui_tot_count;
518 CLOBINVRNT(env, obj, vvp_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 vio->vui_fd->fd_file->f_ra.ra_pages = 0;
538 /* initialize read-ahead window once per syscall */
539 if (!vio->vui_ra_valid) {
540 vio->vui_ra_valid = true;
541 vio->vui_ra_start = cl_index(obj, pos);
542 vio->vui_ra_count = cl_index(obj, tot + PAGE_CACHE_SIZE - 1);
548 switch (vio->vui_io_subtype) {
550 LASSERT(vio->vui_iocb->ki_pos == pos);
551 result = generic_file_aio_read(vio->vui_iocb,
552 vio->vui_iov, vio->vui_nrsegs,
553 vio->vui_iocb->ki_pos);
556 result = generic_file_splice_read(file, &pos,
557 vio->u.splice.vui_pipe, cnt,
558 vio->u.splice.vui_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->vui_io_subtype);
573 io->ci_nob += result;
574 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, vio->vui_fd,
582 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
583 struct cl_page_list *plist, int from, int to)
585 struct cl_2queue *queue = &io->ci_queue;
586 struct cl_page *page;
587 unsigned int bytes = 0;
591 if (plist->pl_nr == 0)
594 if (from > 0 || to != PAGE_SIZE) {
595 page = cl_page_list_first(plist);
596 if (plist->pl_nr == 1) {
597 cl_page_clip(env, page, from, to);
600 cl_page_clip(env, page, from, PAGE_SIZE);
601 if (to != PAGE_SIZE) {
602 page = cl_page_list_last(plist);
603 cl_page_clip(env, page, 0, to);
608 cl_2queue_init(queue);
609 cl_page_list_splice(plist, &queue->c2_qin);
610 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
612 /* plist is not sorted any more */
613 cl_page_list_splice(&queue->c2_qin, plist);
614 cl_page_list_splice(&queue->c2_qout, plist);
615 cl_2queue_fini(env, queue);
618 /* calculate bytes */
619 bytes = plist->pl_nr << PAGE_SHIFT;
620 bytes -= from + PAGE_SIZE - to;
622 while (plist->pl_nr > 0) {
623 page = cl_page_list_first(plist);
624 cl_page_list_del(env, plist, page);
626 cl_page_clip(env, page, 0, PAGE_SIZE);
628 SetPageUptodate(cl_page_vmpage(page));
629 cl_page_disown(env, io, page);
631 /* held in ll_cl_init() */
632 lu_ref_del(&page->cp_reference, "cl_io", io);
633 cl_page_put(env, page);
637 RETURN(bytes > 0 ? bytes : rc);
640 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
641 struct cl_page *page)
643 struct vvp_page *vpg;
644 struct page *vmpage = page->cp_vmpage;
645 struct cl_object *clob = cl_io_top(io)->ci_obj;
647 SetPageUptodate(vmpage);
648 set_page_dirty(vmpage);
650 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
651 vvp_write_pending(cl2vvp(clob), vpg);
653 cl_page_disown(env, io, page);
655 /* held in ll_cl_init() */
656 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
657 cl_page_put(env, page);
660 /* make sure the page list is contiguous */
661 static bool page_list_sanity_check(struct cl_object *obj,
662 struct cl_page_list *plist)
664 struct cl_page *page;
665 pgoff_t index = CL_PAGE_EOF;
667 cl_page_list_for_each(page, plist) {
668 struct vvp_page *vpg = cl_object_page_slice(obj, page);
670 if (index == CL_PAGE_EOF) {
671 index = vvp_index(vpg);
676 if (index == vvp_index(vpg))
684 /* Return how many bytes have queued or written */
685 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
687 struct cl_object *obj = io->ci_obj;
688 struct inode *inode = vvp_object_inode(obj);
689 struct vvp_io *vio = vvp_env_io(env);
690 struct cl_page_list *queue = &vio->u.write.vui_queue;
691 struct cl_page *page;
694 unsigned int npages = vio->u.write.vui_queue.pl_nr;
700 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
701 npages, vio->u.write.vui_from, vio->u.write.vui_to);
703 LASSERT(page_list_sanity_check(obj, queue));
705 /* submit IO with async write */
706 rc = cl_io_commit_async(env, io, queue,
707 vio->u.write.vui_from, vio->u.write.vui_to,
708 write_commit_callback);
709 npages -= queue->pl_nr; /* already committed pages */
711 /* calculate how many bytes were written */
712 bytes = npages << PAGE_SHIFT;
715 bytes -= vio->u.write.vui_from;
716 if (queue->pl_nr == 0) /* last page */
717 bytes -= PAGE_SIZE - vio->u.write.vui_to;
718 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
720 vio->u.write.vui_written += bytes;
722 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
723 npages, bytes, vio->u.write.vui_written);
725 /* the first page must have been written. */
726 vio->u.write.vui_from = 0;
728 LASSERT(page_list_sanity_check(obj, queue));
729 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
731 /* out of quota, try sync write */
732 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
733 rc = vvp_io_commit_sync(env, io, queue,
734 vio->u.write.vui_from,
735 vio->u.write.vui_to);
737 vio->u.write.vui_written += rc;
742 /* update inode size */
743 ll_merge_attr(env, inode);
745 /* Now the pages in queue were failed to commit, discard them
746 * unless they were dirtied before. */
747 while (queue->pl_nr > 0) {
748 page = cl_page_list_first(queue);
749 cl_page_list_del(env, queue, page);
751 if (!PageDirty(cl_page_vmpage(page)))
752 cl_page_discard(env, io, page);
754 cl_page_disown(env, io, page);
756 /* held in ll_cl_init() */
757 lu_ref_del(&page->cp_reference, "cl_io", io);
758 cl_page_put(env, page);
760 cl_page_list_fini(env, queue);
765 static int vvp_io_write_start(const struct lu_env *env,
766 const struct cl_io_slice *ios)
768 struct vvp_io *vio = cl2vvp_io(env, ios);
769 struct cl_io *io = ios->cis_io;
770 struct cl_object *obj = io->ci_obj;
771 struct inode *inode = vvp_object_inode(obj);
773 loff_t pos = io->u.ci_wr.wr.crw_pos;
774 size_t cnt = io->u.ci_wr.wr.crw_count;
778 if (!can_populate_pages(env, io, inode))
781 if (cl_io_is_append(io)) {
783 * PARALLEL IO This has to be changed for parallel IO doing
784 * out-of-order writes.
786 ll_merge_attr(env, inode);
787 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
788 vio->vui_iocb->ki_pos = pos;
790 LASSERT(vio->vui_iocb->ki_pos == pos);
793 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
795 if (vio->vui_iov == NULL) {
796 /* from a temp io in ll_cl_init(). */
800 * When using the locked AIO function (generic_file_aio_write())
801 * testing has shown the inode mutex to be a limiting factor
802 * with multi-threaded single shared file performance. To get
803 * around this, we now use the lockless version. To maintain
804 * consistency, proper locking to protect against writes,
805 * trucates, etc. is handled in the higher layers of lustre.
807 result = __generic_file_aio_write(vio->vui_iocb,
808 vio->vui_iov, vio->vui_nrsegs,
809 &vio->vui_iocb->ki_pos);
810 if (result > 0 || result == -EIOCBQUEUED) {
813 err = generic_write_sync(vio->vui_iocb->ki_filp,
815 if (err < 0 && result > 0)
821 result = vvp_io_write_commit(env, io);
822 if (vio->u.write.vui_written > 0) {
823 result = vio->u.write.vui_written;
824 io->ci_nob += result;
826 CDEBUG(D_VFSTRACE, "write: nob %zd, result: %zd\n",
831 struct ll_inode_info *lli = ll_i2info(inode);
833 spin_lock(&lli->lli_lock);
834 lli->lli_flags |= LLIF_DATA_MODIFIED;
835 spin_unlock(&lli->lli_lock);
839 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
840 vio->vui_fd, pos, result, WRITE);
847 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
849 struct vm_fault *vmf = cfio->ft_vmf;
851 cfio->ft_flags = filemap_fault(cfio->ft_vma, vmf);
852 cfio->ft_flags_valid = 1;
855 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
856 vmf->virtual_address);
857 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
858 lock_page(vmf->page);
859 cfio->ft_flags |= VM_FAULT_LOCKED;
862 cfio->ft_vmpage = vmf->page;
867 if (cfio->ft_flags & VM_FAULT_SIGBUS) {
868 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
872 if (cfio->ft_flags & VM_FAULT_OOM) {
873 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
877 if (cfio->ft_flags & VM_FAULT_RETRY)
880 CERROR("unknown error in page fault %d\n", cfio->ft_flags);
885 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
886 struct cl_page *page)
888 struct vvp_page *vpg;
889 struct cl_object *clob = cl_io_top(io)->ci_obj;
891 set_page_dirty(page->cp_vmpage);
893 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
894 vvp_write_pending(cl2vvp(clob), vpg);
897 static int vvp_io_fault_start(const struct lu_env *env,
898 const struct cl_io_slice *ios)
900 struct vvp_io *vio = cl2vvp_io(env, ios);
901 struct cl_io *io = ios->cis_io;
902 struct cl_object *obj = io->ci_obj;
903 struct inode *inode = vvp_object_inode(obj);
904 struct cl_fault_io *fio = &io->u.ci_fault;
905 struct vvp_fault_io *cfio = &vio->u.fault;
908 struct page *vmpage = NULL;
909 struct cl_page *page;
914 if (fio->ft_executable &&
915 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
917 " changed while waiting for the page fault lock\n",
918 PFID(lu_object_fid(&obj->co_lu)));
920 /* offset of the last byte on the page */
921 offset = cl_offset(obj, fio->ft_index + 1) - 1;
922 LASSERT(cl_index(obj, offset) == fio->ft_index);
923 result = ccc_prep_size(env, obj, io, 0, offset + 1, NULL);
927 /* must return locked page */
928 if (fio->ft_mkwrite) {
929 LASSERT(cfio->ft_vmpage != NULL);
930 lock_page(cfio->ft_vmpage);
932 result = vvp_io_kernel_fault(cfio);
937 vmpage = cfio->ft_vmpage;
938 LASSERT(PageLocked(vmpage));
940 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
941 ll_invalidate_page(vmpage);
943 size = i_size_read(inode);
944 /* Though we have already held a cl_lock upon this page, but
945 * it still can be truncated locally. */
946 if (unlikely((vmpage->mapping != inode->i_mapping) ||
947 (page_offset(vmpage) > size))) {
948 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
950 /* return +1 to stop cl_io_loop() and ll_fault() will catch
952 GOTO(out, result = +1);
955 last_index = cl_index(obj, size - 1);
957 if (fio->ft_mkwrite ) {
959 * Capture the size while holding the lli_trunc_sem from above
960 * we want to make sure that we complete the mkwrite action
961 * while holding this lock. We need to make sure that we are
962 * not past the end of the file.
964 if (last_index < fio->ft_index) {
966 "llite: mkwrite and truncate race happened: "
968 vmpage->mapping,fio->ft_index,last_index);
970 * We need to return if we are
971 * passed the end of the file. This will propagate
972 * up the call stack to ll_page_mkwrite where
973 * we will return VM_FAULT_NOPAGE. Any non-negative
974 * value returned here will be silently
975 * converted to 0. If the vmpage->mapping is null
976 * the error code would be converted back to ENODATA
977 * in ll_page_mkwrite0. Thus we return -ENODATA
978 * to handle both cases
980 GOTO(out, result = -ENODATA);
984 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
986 GOTO(out, result = PTR_ERR(page));
988 /* if page is going to be written, we should add this page into cache
990 if (fio->ft_mkwrite) {
991 wait_on_page_writeback(vmpage);
992 if (!PageDirty(vmpage)) {
993 struct cl_page_list *plist = &io->ci_queue.c2_qin;
994 struct vvp_page *vpg = cl_object_page_slice(obj, page);
997 /* vvp_page_assume() calls wait_on_page_writeback(). */
998 cl_page_assume(env, io, page);
1000 cl_page_list_init(plist);
1001 cl_page_list_add(plist, page);
1004 if (last_index == vvp_index(vpg))
1005 to = size & ~CFS_PAGE_MASK;
1007 /* Do not set Dirty bit here so that in case IO is
1008 * started before the page is really made dirty, we
1009 * still have chance to detect it. */
1010 result = cl_io_commit_async(env, io, plist, 0, to,
1011 mkwrite_commit_callback);
1012 LASSERT(cl_page_is_owned(page, io));
1013 cl_page_list_fini(env, plist);
1017 cl_page_discard(env, io, page);
1018 cl_page_disown(env, io, page);
1020 cl_page_put(env, page);
1022 /* we're in big trouble, what can we do now? */
1023 if (result == -EDQUOT)
1027 cl_page_disown(env, io, page);
1032 * The ft_index is only used in the case of
1033 * a mkwrite action. We need to check
1034 * our assertions are correct, since
1035 * we should have caught this above
1037 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1038 if (fio->ft_index == last_index)
1040 * Last page is mapped partially.
1042 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1044 fio->ft_nob = cl_page_size(obj);
1046 lu_ref_add(&page->cp_reference, "fault", io);
1047 fio->ft_page = page;
1051 /* return unlocked vmpage to avoid deadlocking */
1053 unlock_page(vmpage);
1055 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1060 static int vvp_io_fsync_start(const struct lu_env *env,
1061 const struct cl_io_slice *ios)
1063 /* we should mark TOWRITE bit to each dirty page in radix tree to
1064 * verify pages have been written, but this is difficult because of
1069 static int vvp_io_read_page(const struct lu_env *env,
1070 const struct cl_io_slice *ios,
1071 const struct cl_page_slice *slice)
1073 struct cl_io *io = ios->cis_io;
1074 struct vvp_page *vpg = cl2vvp_page(slice);
1075 struct cl_page *page = slice->cpl_page;
1076 struct inode *inode = vvp_object_inode(slice->cpl_obj);
1077 struct ll_sb_info *sbi = ll_i2sbi(inode);
1078 struct ll_file_data *fd = cl2vvp_io(env, ios)->vui_fd;
1079 struct ll_readahead_state *ras = &fd->fd_ras;
1080 struct cl_2queue *queue = &io->ci_queue;
1084 if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
1085 sbi->ll_ra_info.ra_max_pages > 0)
1086 ras_update(sbi, inode, ras, vvp_index(vpg),
1087 vpg->vpg_defer_uptodate);
1089 if (vpg->vpg_defer_uptodate) {
1090 vpg->vpg_ra_used = 1;
1091 cl_page_export(env, page, 1);
1095 * Add page into the queue even when it is marked uptodate above.
1096 * this will unlock it automatically as part of cl_page_list_disown().
1098 cl_2queue_add(queue, page);
1099 if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
1100 sbi->ll_ra_info.ra_max_pages > 0)
1101 ll_readahead(env, io, &queue->c2_qin, ras,
1102 vpg->vpg_defer_uptodate);
1107 static const struct cl_io_operations vvp_io_ops = {
1110 .cio_fini = vvp_io_fini,
1111 .cio_lock = vvp_io_read_lock,
1112 .cio_start = vvp_io_read_start,
1113 .cio_advance = vvp_io_advance,
1116 .cio_fini = vvp_io_fini,
1117 .cio_iter_init = vvp_io_write_iter_init,
1118 .cio_iter_fini = vvp_io_write_iter_fini,
1119 .cio_lock = vvp_io_write_lock,
1120 .cio_start = vvp_io_write_start,
1121 .cio_advance = vvp_io_advance,
1124 .cio_fini = vvp_io_setattr_fini,
1125 .cio_iter_init = vvp_io_setattr_iter_init,
1126 .cio_lock = vvp_io_setattr_lock,
1127 .cio_start = vvp_io_setattr_start,
1128 .cio_end = vvp_io_setattr_end
1131 .cio_fini = vvp_io_fault_fini,
1132 .cio_iter_init = vvp_io_fault_iter_init,
1133 .cio_lock = vvp_io_fault_lock,
1134 .cio_start = vvp_io_fault_start,
1135 .cio_end = vvp_io_end,
1138 .cio_start = vvp_io_fsync_start,
1139 .cio_fini = vvp_io_fini
1142 .cio_fini = vvp_io_fini
1145 .cio_read_page = vvp_io_read_page,
1148 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1151 struct vvp_io *vio = vvp_env_io(env);
1152 struct inode *inode = vvp_object_inode(obj);
1155 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1158 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1159 "restore needed %d\n",
1160 PFID(lu_object_fid(&obj->co_lu)),
1161 io->ci_ignore_layout, io->ci_verify_layout,
1162 vio->vui_layout_gen, io->ci_restore_needed);
1164 CL_IO_SLICE_CLEAN(vio, vui_cl);
1165 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1166 vio->vui_ra_valid = false;
1168 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1170 struct ll_inode_info *lli = ll_i2info(inode);
1172 count = io->u.ci_rw.crw_count;
1173 /* "If nbyte is 0, read() will return 0 and have no other
1174 * results." -- Single Unix Spec */
1178 vio->vui_tot_count = count;
1179 vio->vui_tot_nrsegs = 0;
1182 /* for read/write, we store the jobid in the inode, and
1183 * it'll be fetched by osc when building RPC.
1185 * it's not accurate if the file is shared by different
1188 lustre_get_jobid(lli->lli_jobid);
1189 } else if (io->ci_type == CIT_SETATTR) {
1190 if (!cl_io_is_trunc(io))
1191 io->ci_lockreq = CILR_MANDATORY;
1194 /* ignore layout change for generic CIT_MISC but not for glimpse.
1195 * io context for glimpse must set ci_verify_layout to true,
1196 * see cl_glimpse_size0() for details. */
1197 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1198 io->ci_ignore_layout = 1;
1200 /* Enqueue layout lock and get layout version. We need to do this
1201 * even for operations requiring to open file, such as read and write,
1202 * because it might not grant layout lock in IT_OPEN. */
1203 if (result == 0 && !io->ci_ignore_layout) {
1204 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1205 if (result == -ENOENT)
1206 /* If the inode on MDS has been removed, but the objects
1207 * on OSTs haven't been destroyed (async unlink), layout
1208 * fetch will return -ENOENT, we'd ingore this error
1209 * and continue with dirty flush. LU-3230. */
1212 CERROR("%s: refresh file layout " DFID " error %d.\n",
1213 ll_get_fsname(inode->i_sb, NULL, 0),
1214 PFID(lu_object_fid(&obj->co_lu)), result);