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, 2015, 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
44 #include "../include/obd.h"
45 #include "../include/lustre_lite.h"
47 #include "llite_internal.h"
48 #include "vvp_internal.h"
50 struct vvp_io *cl2vvp_io(const struct lu_env *env,
51 const struct cl_io_slice *slice)
55 vio = container_of(slice, struct vvp_io, vui_cl);
56 LASSERT(vio == vvp_env_io(env));
62 * True, if \a io is a normal io, False for splice_{read,write}
64 static int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
66 struct vvp_io *vio = vvp_env_io(env);
68 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
70 return vio->vui_io_subtype == IO_NORMAL;
74 * For swapping layout. The file's layout may have changed.
75 * To avoid populating pages to a wrong stripe, we have to verify the
76 * correctness of layout. It works because swapping layout processes
77 * have to acquire group lock.
79 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
82 struct ll_inode_info *lli = ll_i2info(inode);
83 struct vvp_io *vio = vvp_env_io(env);
86 switch (io->ci_type) {
89 /* don't need lock here to check lli_layout_gen as we have held
90 * extent lock and GROUP lock has to hold to swap layout
92 if (ll_layout_version_get(lli) != vio->vui_layout_gen) {
93 io->ci_need_restart = 1;
94 /* this will return application a short read/write */
99 /* fault is okay because we've already had a page. */
107 static void vvp_object_size_lock(struct cl_object *obj)
109 struct inode *inode = vvp_object_inode(obj);
111 ll_inode_size_lock(inode);
112 cl_object_attr_lock(obj);
115 static void vvp_object_size_unlock(struct cl_object *obj)
117 struct inode *inode = vvp_object_inode(obj);
119 cl_object_attr_unlock(obj);
120 ll_inode_size_unlock(inode);
124 * Helper function that if necessary adjusts file size (inode->i_size), when
125 * position at the offset \a pos is accessed. File size can be arbitrary stale
126 * on a Lustre client, but client at least knows KMS. If accessed area is
127 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
129 * Locking: cl_isize_lock is used to serialize changes to inode size and to
130 * protect consistency between inode size and cl_object
131 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
132 * top-object and sub-objects.
134 static int vvp_prep_size(const struct lu_env *env, struct cl_object *obj,
135 struct cl_io *io, loff_t start, size_t count,
138 struct cl_attr *attr = vvp_env_thread_attr(env);
139 struct inode *inode = vvp_object_inode(obj);
140 loff_t pos = start + count - 1;
145 * Consistency guarantees: following possibilities exist for the
146 * relation between region being accessed and real file size at this
149 * (A): the region is completely inside of the file;
151 * (B-x): x bytes of region are inside of the file, the rest is
154 * (C): the region is completely outside of the file.
156 * This classification is stable under DLM lock already acquired by
157 * the caller, because to change the class, other client has to take
158 * DLM lock conflicting with our lock. Also, any updates to ->i_size
159 * by other threads on this client are serialized by
160 * ll_inode_size_lock(). This guarantees that short reads are handled
161 * correctly in the face of concurrent writes and truncates.
163 vvp_object_size_lock(obj);
164 result = cl_object_attr_get(env, obj, attr);
169 * A glimpse is necessary to determine whether we
170 * return a short read (B) or some zeroes at the end
173 vvp_object_size_unlock(obj);
174 result = cl_glimpse_lock(env, io, inode, obj, 0);
175 if (result == 0 && exceed) {
176 /* If objective page index exceed end-of-file
177 * page index, return directly. Do not expect
178 * kernel will check such case correctly.
179 * linux-2.6.18-128.1.1 miss to do that.
182 loff_t size = i_size_read(inode);
183 loff_t cur_index = start >> PAGE_SHIFT;
184 loff_t size_index = (size - 1) >> PAGE_SHIFT;
186 if ((size == 0 && cur_index != 0) ||
187 size_index < cur_index)
193 * region is within kms and, hence, within real file
194 * size (A). We need to increase i_size to cover the
195 * read region so that generic_file_read() will do its
196 * job, but that doesn't mean the kms size is
197 * _correct_, it is only the _minimum_ size. If
198 * someone does a stat they will get the correct size
199 * which will always be >= the kms value here.
202 if (i_size_read(inode) < kms) {
203 i_size_write(inode, kms);
204 CDEBUG(D_VFSTRACE, DFID " updating i_size %llu\n",
205 PFID(lu_object_fid(&obj->co_lu)),
206 (__u64)i_size_read(inode));
210 vvp_object_size_unlock(obj);
215 /*****************************************************************************
221 static int vvp_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
222 __u32 enqflags, enum cl_lock_mode mode,
223 pgoff_t start, pgoff_t end)
225 struct vvp_io *vio = vvp_env_io(env);
226 struct cl_lock_descr *descr = &vio->vui_link.cill_descr;
227 struct cl_object *obj = io->ci_obj;
229 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
231 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
233 memset(&vio->vui_link, 0, sizeof(vio->vui_link));
235 if (vio->vui_fd && (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
236 descr->cld_mode = CLM_GROUP;
237 descr->cld_gid = vio->vui_fd->fd_grouplock.lg_gid;
239 descr->cld_mode = mode;
241 descr->cld_obj = obj;
242 descr->cld_start = start;
243 descr->cld_end = end;
244 descr->cld_enq_flags = enqflags;
246 cl_io_lock_add(env, io, &vio->vui_link);
250 static int vvp_io_one_lock(const struct lu_env *env, struct cl_io *io,
251 __u32 enqflags, enum cl_lock_mode mode,
252 loff_t start, loff_t end)
254 struct cl_object *obj = io->ci_obj;
256 return vvp_io_one_lock_index(env, io, enqflags, mode,
257 cl_index(obj, start), cl_index(obj, end));
260 static int vvp_io_write_iter_init(const struct lu_env *env,
261 const struct cl_io_slice *ios)
263 struct vvp_io *vio = cl2vvp_io(env, ios);
265 cl_page_list_init(&vio->u.write.vui_queue);
266 vio->u.write.vui_written = 0;
267 vio->u.write.vui_from = 0;
268 vio->u.write.vui_to = PAGE_SIZE;
273 static void vvp_io_write_iter_fini(const struct lu_env *env,
274 const struct cl_io_slice *ios)
276 struct vvp_io *vio = cl2vvp_io(env, ios);
278 LASSERT(vio->u.write.vui_queue.pl_nr == 0);
281 static int vvp_io_fault_iter_init(const struct lu_env *env,
282 const struct cl_io_slice *ios)
284 struct vvp_io *vio = cl2vvp_io(env, ios);
285 struct inode *inode = vvp_object_inode(ios->cis_obj);
287 LASSERT(inode == file_inode(vio->vui_fd->fd_file));
288 vio->u.fault.ft_mtime = inode->i_mtime.tv_sec;
292 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
294 struct cl_io *io = ios->cis_io;
295 struct cl_object *obj = io->ci_obj;
296 struct vvp_io *vio = cl2vvp_io(env, ios);
297 struct inode *inode = vvp_object_inode(obj);
299 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
301 CDEBUG(D_VFSTRACE, DFID
302 " ignore/verify layout %d/%d, layout version %d restore needed %d\n",
303 PFID(lu_object_fid(&obj->co_lu)),
304 io->ci_ignore_layout, io->ci_verify_layout,
305 vio->vui_layout_gen, io->ci_restore_needed);
307 if (io->ci_restore_needed == 1) {
310 /* file was detected release, we need to restore it
311 * before finishing the io
313 rc = ll_layout_restore(inode, 0, OBD_OBJECT_EOF);
314 /* if restore registration failed, no restart,
315 * we will return -ENODATA
317 /* The layout will change after restore, so we need to
318 * block on layout lock hold by the MDT
319 * as MDT will not send new layout in lvb (see LU-3124)
320 * we have to explicitly fetch it, all this will be done
321 * by ll_layout_refresh()
324 io->ci_restore_needed = 0;
325 io->ci_need_restart = 1;
326 io->ci_verify_layout = 1;
328 io->ci_restore_needed = 1;
329 io->ci_need_restart = 0;
330 io->ci_verify_layout = 0;
335 if (!io->ci_ignore_layout && io->ci_verify_layout) {
338 /* check layout version */
339 ll_layout_refresh(inode, &gen);
340 io->ci_need_restart = vio->vui_layout_gen != gen;
341 if (io->ci_need_restart) {
343 DFID" layout changed from %d to %d.\n",
344 PFID(lu_object_fid(&obj->co_lu)),
345 vio->vui_layout_gen, gen);
346 /* today successful restore is the only possible case */
347 /* restore was done, clear restoring state */
348 ll_i2info(vvp_object_inode(obj))->lli_flags &=
349 ~LLIF_FILE_RESTORING;
354 static void vvp_io_fault_fini(const struct lu_env *env,
355 const struct cl_io_slice *ios)
357 struct cl_io *io = ios->cis_io;
358 struct cl_page *page = io->u.ci_fault.ft_page;
360 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
363 lu_ref_del(&page->cp_reference, "fault", io);
364 cl_page_put(env, page);
365 io->u.ci_fault.ft_page = NULL;
367 vvp_io_fini(env, ios);
370 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
373 * we only want to hold PW locks if the mmap() can generate
374 * writes back to the file and that only happens in shared
377 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
382 static int vvp_mmap_locks(const struct lu_env *env,
383 struct vvp_io *vio, struct cl_io *io)
385 struct vvp_thread_info *cti = vvp_env_info(env);
386 struct mm_struct *mm = current->mm;
387 struct vm_area_struct *vma;
388 struct cl_lock_descr *descr = &cti->vti_descr;
389 ldlm_policy_data_t policy;
396 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
398 if (!cl_is_normalio(env, io))
401 if (!vio->vui_iter) /* nfs or loop back device write */
404 /* No MM (e.g. NFS)? No vmas too. */
408 iov_for_each(iov, i, *vio->vui_iter) {
409 addr = (unsigned long)iov.iov_base;
414 count += addr & (~PAGE_MASK);
417 down_read(&mm->mmap_sem);
418 while ((vma = our_vma(mm, addr, count)) != NULL) {
419 struct inode *inode = file_inode(vma->vm_file);
420 int flags = CEF_MUST;
422 if (ll_file_nolock(vma->vm_file)) {
424 * For no lock case is not allowed for mmap
431 * XXX: Required lock mode can be weakened: CIT_WRITE
432 * io only ever reads user level buffer, and CIT_READ
435 policy_from_vma(&policy, vma, addr, count);
436 descr->cld_mode = vvp_mode_from_vma(vma);
437 descr->cld_obj = ll_i2info(inode)->lli_clob;
438 descr->cld_start = cl_index(descr->cld_obj,
439 policy.l_extent.start);
440 descr->cld_end = cl_index(descr->cld_obj,
441 policy.l_extent.end);
442 descr->cld_enq_flags = flags;
443 result = cl_io_lock_alloc_add(env, io, descr);
445 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
446 descr->cld_mode, descr->cld_start,
452 if (vma->vm_end - addr >= count)
455 count -= vma->vm_end - addr;
458 up_read(&mm->mmap_sem);
465 static void vvp_io_advance(const struct lu_env *env,
466 const struct cl_io_slice *ios,
469 struct vvp_io *vio = cl2vvp_io(env, ios);
470 struct cl_io *io = ios->cis_io;
471 struct cl_object *obj = ios->cis_io->ci_obj;
473 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
475 if (!cl_is_normalio(env, io))
478 iov_iter_reexpand(vio->vui_iter, vio->vui_tot_count -= nob);
481 static void vvp_io_update_iov(const struct lu_env *env,
482 struct vvp_io *vio, struct cl_io *io)
484 size_t size = io->u.ci_rw.crw_count;
486 if (!cl_is_normalio(env, io) || !vio->vui_iter)
489 iov_iter_truncate(vio->vui_iter, size);
492 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
493 enum cl_lock_mode mode, loff_t start, loff_t end)
495 struct vvp_io *vio = vvp_env_io(env);
499 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
501 vvp_io_update_iov(env, vio, io);
503 if (io->u.ci_rw.crw_nonblock)
504 ast_flags |= CEF_NONBLOCK;
505 result = vvp_mmap_locks(env, vio, io);
507 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
511 static int vvp_io_read_lock(const struct lu_env *env,
512 const struct cl_io_slice *ios)
514 struct cl_io *io = ios->cis_io;
515 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
518 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
519 rd->crw_pos + rd->crw_count - 1);
524 static int vvp_io_fault_lock(const struct lu_env *env,
525 const struct cl_io_slice *ios)
527 struct cl_io *io = ios->cis_io;
528 struct vvp_io *vio = cl2vvp_io(env, ios);
530 * XXX LDLM_FL_CBPENDING
532 return vvp_io_one_lock_index(env,
534 vvp_mode_from_vma(vio->u.fault.ft_vma),
535 io->u.ci_fault.ft_index,
536 io->u.ci_fault.ft_index);
539 static int vvp_io_write_lock(const struct lu_env *env,
540 const struct cl_io_slice *ios)
542 struct cl_io *io = ios->cis_io;
546 if (io->u.ci_wr.wr_append) {
548 end = OBD_OBJECT_EOF;
550 start = io->u.ci_wr.wr.crw_pos;
551 end = start + io->u.ci_wr.wr.crw_count - 1;
553 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
556 static int vvp_io_setattr_iter_init(const struct lu_env *env,
557 const struct cl_io_slice *ios)
563 * Implementation of cl_io_operations::vio_lock() method for CIT_SETATTR io.
565 * Handles "lockless io" mode when extent locking is done by server.
567 static int vvp_io_setattr_lock(const struct lu_env *env,
568 const struct cl_io_slice *ios)
570 struct cl_io *io = ios->cis_io;
574 if (cl_io_is_trunc(io)) {
575 new_size = io->u.ci_setattr.sa_attr.lvb_size;
577 enqflags = CEF_DISCARD_DATA;
579 if ((io->u.ci_setattr.sa_attr.lvb_mtime >=
580 io->u.ci_setattr.sa_attr.lvb_ctime) ||
581 (io->u.ci_setattr.sa_attr.lvb_atime >=
582 io->u.ci_setattr.sa_attr.lvb_ctime))
587 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
588 new_size, OBD_OBJECT_EOF);
591 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
595 * Only ll_inode_size_lock is taken at this level.
597 ll_inode_size_lock(inode);
598 result = inode_newsize_ok(inode, size);
600 ll_inode_size_unlock(inode);
603 truncate_setsize(inode, size);
604 ll_inode_size_unlock(inode);
608 static int vvp_io_setattr_trunc(const struct lu_env *env,
609 const struct cl_io_slice *ios,
610 struct inode *inode, loff_t size)
612 inode_dio_wait(inode);
616 static int vvp_io_setattr_time(const struct lu_env *env,
617 const struct cl_io_slice *ios)
619 struct cl_io *io = ios->cis_io;
620 struct cl_object *obj = io->ci_obj;
621 struct cl_attr *attr = vvp_env_thread_attr(env);
623 unsigned valid = CAT_CTIME;
625 cl_object_attr_lock(obj);
626 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
627 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
628 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
631 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
632 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
635 result = cl_object_attr_set(env, obj, attr, valid);
636 cl_object_attr_unlock(obj);
641 static int vvp_io_setattr_start(const struct lu_env *env,
642 const struct cl_io_slice *ios)
644 struct cl_io *io = ios->cis_io;
645 struct inode *inode = vvp_object_inode(io->ci_obj);
649 if (cl_io_is_trunc(io))
650 result = vvp_io_setattr_trunc(env, ios, inode,
651 io->u.ci_setattr.sa_attr.lvb_size);
653 result = vvp_io_setattr_time(env, ios);
657 static void vvp_io_setattr_end(const struct lu_env *env,
658 const struct cl_io_slice *ios)
660 struct cl_io *io = ios->cis_io;
661 struct inode *inode = vvp_object_inode(io->ci_obj);
663 if (cl_io_is_trunc(io))
664 /* Truncate in memory pages - they must be clean pages
665 * because osc has already notified to destroy osc_extents.
667 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
672 static void vvp_io_setattr_fini(const struct lu_env *env,
673 const struct cl_io_slice *ios)
675 vvp_io_fini(env, ios);
678 static int vvp_io_read_start(const struct lu_env *env,
679 const struct cl_io_slice *ios)
681 struct vvp_io *vio = cl2vvp_io(env, ios);
682 struct cl_io *io = ios->cis_io;
683 struct cl_object *obj = io->ci_obj;
684 struct inode *inode = vvp_object_inode(obj);
685 struct file *file = vio->vui_fd->fd_file;
688 loff_t pos = io->u.ci_rd.rd.crw_pos;
689 long cnt = io->u.ci_rd.rd.crw_count;
690 long tot = vio->vui_tot_count;
693 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
695 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
697 if (!can_populate_pages(env, io, inode))
700 result = vvp_prep_size(env, obj, io, pos, tot, &exceed);
703 else if (exceed != 0)
706 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
707 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
708 inode->i_ino, cnt, pos, i_size_read(inode));
710 /* turn off the kernel's read-ahead */
711 vio->vui_fd->fd_file->f_ra.ra_pages = 0;
713 /* initialize read-ahead window once per syscall */
714 if (!vio->vui_ra_valid) {
715 vio->vui_ra_valid = true;
716 vio->vui_ra_start = cl_index(obj, pos);
717 vio->vui_ra_count = cl_index(obj, tot + PAGE_SIZE - 1);
723 switch (vio->vui_io_subtype) {
725 LASSERT(vio->vui_iocb->ki_pos == pos);
726 result = generic_file_read_iter(vio->vui_iocb, vio->vui_iter);
729 result = generic_file_splice_read(file, &pos,
730 vio->u.splice.vui_pipe, cnt,
731 vio->u.splice.vui_flags);
732 /* LU-1109: do splice read stripe by stripe otherwise if it
733 * may make nfsd stuck if this read occupied all internal pipe
739 CERROR("Wrong IO type %u\n", vio->vui_io_subtype);
747 io->ci_nob += result;
748 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
749 vio->vui_fd, pos, result, READ);
755 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
756 struct cl_page_list *plist, int from, int to)
758 struct cl_2queue *queue = &io->ci_queue;
759 struct cl_page *page;
760 unsigned int bytes = 0;
763 if (plist->pl_nr == 0)
766 if (from > 0 || to != PAGE_SIZE) {
767 page = cl_page_list_first(plist);
768 if (plist->pl_nr == 1) {
769 cl_page_clip(env, page, from, to);
772 cl_page_clip(env, page, from, PAGE_SIZE);
773 if (to != PAGE_SIZE) {
774 page = cl_page_list_last(plist);
775 cl_page_clip(env, page, 0, to);
780 cl_2queue_init(queue);
781 cl_page_list_splice(plist, &queue->c2_qin);
782 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
784 /* plist is not sorted any more */
785 cl_page_list_splice(&queue->c2_qin, plist);
786 cl_page_list_splice(&queue->c2_qout, plist);
787 cl_2queue_fini(env, queue);
790 /* calculate bytes */
791 bytes = plist->pl_nr << PAGE_SHIFT;
792 bytes -= from + PAGE_SIZE - to;
794 while (plist->pl_nr > 0) {
795 page = cl_page_list_first(plist);
796 cl_page_list_del(env, plist, page);
798 cl_page_clip(env, page, 0, PAGE_SIZE);
800 SetPageUptodate(cl_page_vmpage(page));
801 cl_page_disown(env, io, page);
803 /* held in ll_cl_init() */
804 lu_ref_del(&page->cp_reference, "cl_io", io);
805 cl_page_put(env, page);
809 return bytes > 0 ? bytes : rc;
812 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
813 struct cl_page *page)
815 struct vvp_page *vpg;
816 struct page *vmpage = page->cp_vmpage;
817 struct cl_object *clob = cl_io_top(io)->ci_obj;
819 SetPageUptodate(vmpage);
820 set_page_dirty(vmpage);
822 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
823 vvp_write_pending(cl2vvp(clob), vpg);
825 cl_page_disown(env, io, page);
827 /* held in ll_cl_init() */
828 lu_ref_del(&page->cp_reference, "cl_io", io);
829 cl_page_put(env, page);
832 /* make sure the page list is contiguous */
833 static bool page_list_sanity_check(struct cl_object *obj,
834 struct cl_page_list *plist)
836 struct cl_page *page;
837 pgoff_t index = CL_PAGE_EOF;
839 cl_page_list_for_each(page, plist) {
840 struct vvp_page *vpg = cl_object_page_slice(obj, page);
842 if (index == CL_PAGE_EOF) {
843 index = vvp_index(vpg);
848 if (index == vvp_index(vpg))
856 /* Return how many bytes have queued or written */
857 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
859 struct cl_object *obj = io->ci_obj;
860 struct inode *inode = vvp_object_inode(obj);
861 struct vvp_io *vio = vvp_env_io(env);
862 struct cl_page_list *queue = &vio->u.write.vui_queue;
863 struct cl_page *page;
866 unsigned int npages = vio->u.write.vui_queue.pl_nr;
871 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
872 npages, vio->u.write.vui_from, vio->u.write.vui_to);
874 LASSERT(page_list_sanity_check(obj, queue));
876 /* submit IO with async write */
877 rc = cl_io_commit_async(env, io, queue,
878 vio->u.write.vui_from, vio->u.write.vui_to,
879 write_commit_callback);
880 npages -= queue->pl_nr; /* already committed pages */
882 /* calculate how many bytes were written */
883 bytes = npages << PAGE_SHIFT;
886 bytes -= vio->u.write.vui_from;
887 if (queue->pl_nr == 0) /* last page */
888 bytes -= PAGE_SIZE - vio->u.write.vui_to;
889 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
891 vio->u.write.vui_written += bytes;
893 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
894 npages, bytes, vio->u.write.vui_written);
896 /* the first page must have been written. */
897 vio->u.write.vui_from = 0;
899 LASSERT(page_list_sanity_check(obj, queue));
900 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
902 /* out of quota, try sync write */
903 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
904 rc = vvp_io_commit_sync(env, io, queue,
905 vio->u.write.vui_from,
906 vio->u.write.vui_to);
908 vio->u.write.vui_written += rc;
913 /* update inode size */
914 ll_merge_attr(env, inode);
916 /* Now the pages in queue were failed to commit, discard them
917 * unless they were dirtied before.
919 while (queue->pl_nr > 0) {
920 page = cl_page_list_first(queue);
921 cl_page_list_del(env, queue, page);
923 if (!PageDirty(cl_page_vmpage(page)))
924 cl_page_discard(env, io, page);
926 cl_page_disown(env, io, page);
928 /* held in ll_cl_init() */
929 lu_ref_del(&page->cp_reference, "cl_io", io);
930 cl_page_put(env, page);
932 cl_page_list_fini(env, queue);
937 static int vvp_io_write_start(const struct lu_env *env,
938 const struct cl_io_slice *ios)
940 struct vvp_io *vio = cl2vvp_io(env, ios);
941 struct cl_io *io = ios->cis_io;
942 struct cl_object *obj = io->ci_obj;
943 struct inode *inode = vvp_object_inode(obj);
945 loff_t pos = io->u.ci_wr.wr.crw_pos;
946 size_t cnt = io->u.ci_wr.wr.crw_count;
948 if (!can_populate_pages(env, io, inode))
951 if (cl_io_is_append(io)) {
953 * PARALLEL IO This has to be changed for parallel IO doing
954 * out-of-order writes.
956 ll_merge_attr(env, inode);
957 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
958 vio->vui_iocb->ki_pos = pos;
960 LASSERT(vio->vui_iocb->ki_pos == pos);
963 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
965 if (!vio->vui_iter) /* from a temp io in ll_cl_init(). */
968 result = generic_file_write_iter(vio->vui_iocb, vio->vui_iter);
971 result = vvp_io_write_commit(env, io);
972 if (vio->u.write.vui_written > 0) {
973 result = vio->u.write.vui_written;
974 io->ci_nob += result;
976 CDEBUG(D_VFSTRACE, "write: nob %zd, result: %zd\n",
981 struct ll_inode_info *lli = ll_i2info(inode);
983 spin_lock(&lli->lli_lock);
984 lli->lli_flags |= LLIF_DATA_MODIFIED;
985 spin_unlock(&lli->lli_lock);
989 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
990 vio->vui_fd, pos, result, WRITE);
996 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
998 struct vm_fault *vmf = cfio->ft_vmf;
1000 cfio->ft_flags = filemap_fault(cfio->ft_vma, vmf);
1001 cfio->ft_flags_valid = 1;
1005 "page %p map %p index %lu flags %lx count %u priv %0lx: got addr %p type NOPAGE\n",
1006 vmf->page, vmf->page->mapping, vmf->page->index,
1007 (long)vmf->page->flags, page_count(vmf->page),
1008 page_private(vmf->page), vmf->virtual_address);
1009 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
1010 lock_page(vmf->page);
1011 cfio->ft_flags |= VM_FAULT_LOCKED;
1014 cfio->ft_vmpage = vmf->page;
1018 if (cfio->ft_flags & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV)) {
1019 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
1023 if (cfio->ft_flags & VM_FAULT_OOM) {
1024 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
1028 if (cfio->ft_flags & VM_FAULT_RETRY)
1031 CERROR("Unknown error in page fault %d!\n", cfio->ft_flags);
1035 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1036 struct cl_page *page)
1038 struct vvp_page *vpg;
1039 struct cl_object *clob = cl_io_top(io)->ci_obj;
1041 set_page_dirty(page->cp_vmpage);
1043 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
1044 vvp_write_pending(cl2vvp(clob), vpg);
1047 static int vvp_io_fault_start(const struct lu_env *env,
1048 const struct cl_io_slice *ios)
1050 struct vvp_io *vio = cl2vvp_io(env, ios);
1051 struct cl_io *io = ios->cis_io;
1052 struct cl_object *obj = io->ci_obj;
1053 struct inode *inode = vvp_object_inode(obj);
1054 struct cl_fault_io *fio = &io->u.ci_fault;
1055 struct vvp_fault_io *cfio = &vio->u.fault;
1058 struct page *vmpage = NULL;
1059 struct cl_page *page;
1063 if (fio->ft_executable &&
1064 inode->i_mtime.tv_sec != vio->u.fault.ft_mtime)
1066 " changed while waiting for the page fault lock\n",
1067 PFID(lu_object_fid(&obj->co_lu)));
1069 /* offset of the last byte on the page */
1070 offset = cl_offset(obj, fio->ft_index + 1) - 1;
1071 LASSERT(cl_index(obj, offset) == fio->ft_index);
1072 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1076 /* must return locked page */
1077 if (fio->ft_mkwrite) {
1078 LASSERT(cfio->ft_vmpage);
1079 lock_page(cfio->ft_vmpage);
1081 result = vvp_io_kernel_fault(cfio);
1086 vmpage = cfio->ft_vmpage;
1087 LASSERT(PageLocked(vmpage));
1089 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1090 ll_invalidate_page(vmpage);
1092 size = i_size_read(inode);
1093 /* Though we have already held a cl_lock upon this page, but
1094 * it still can be truncated locally.
1096 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1097 (page_offset(vmpage) > size))) {
1098 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1100 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1107 last_index = cl_index(obj, size - 1);
1109 if (fio->ft_mkwrite) {
1111 * Capture the size while holding the lli_trunc_sem from above
1112 * we want to make sure that we complete the mkwrite action
1113 * while holding this lock. We need to make sure that we are
1114 * not past the end of the file.
1116 if (last_index < fio->ft_index) {
1118 "llite: mkwrite and truncate race happened: %p: 0x%lx 0x%lx\n",
1119 vmpage->mapping, fio->ft_index, last_index);
1121 * We need to return if we are
1122 * passed the end of the file. This will propagate
1123 * up the call stack to ll_page_mkwrite where
1124 * we will return VM_FAULT_NOPAGE. Any non-negative
1125 * value returned here will be silently
1126 * converted to 0. If the vmpage->mapping is null
1127 * the error code would be converted back to ENODATA
1128 * in ll_page_mkwrite0. Thus we return -ENODATA
1129 * to handle both cases
1136 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1138 result = PTR_ERR(page);
1142 /* if page is going to be written, we should add this page into cache
1145 if (fio->ft_mkwrite) {
1146 wait_on_page_writeback(vmpage);
1147 if (!PageDirty(vmpage)) {
1148 struct cl_page_list *plist = &io->ci_queue.c2_qin;
1149 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1152 /* vvp_page_assume() calls wait_on_page_writeback(). */
1153 cl_page_assume(env, io, page);
1155 cl_page_list_init(plist);
1156 cl_page_list_add(plist, page);
1159 if (last_index == vvp_index(vpg))
1160 to = size & ~PAGE_MASK;
1162 /* Do not set Dirty bit here so that in case IO is
1163 * started before the page is really made dirty, we
1164 * still have chance to detect it.
1166 result = cl_io_commit_async(env, io, plist, 0, to,
1167 mkwrite_commit_callback);
1168 LASSERT(cl_page_is_owned(page, io));
1169 cl_page_list_fini(env, plist);
1173 cl_page_discard(env, io, page);
1174 cl_page_disown(env, io, page);
1176 cl_page_put(env, page);
1178 /* we're in big trouble, what can we do now? */
1179 if (result == -EDQUOT)
1183 cl_page_disown(env, io, page);
1189 * The ft_index is only used in the case of
1190 * a mkwrite action. We need to check
1191 * our assertions are correct, since
1192 * we should have caught this above
1194 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1195 if (fio->ft_index == last_index)
1197 * Last page is mapped partially.
1199 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1201 fio->ft_nob = cl_page_size(obj);
1203 lu_ref_add(&page->cp_reference, "fault", io);
1204 fio->ft_page = page;
1207 /* return unlocked vmpage to avoid deadlocking */
1209 unlock_page(vmpage);
1211 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1216 static int vvp_io_fsync_start(const struct lu_env *env,
1217 const struct cl_io_slice *ios)
1219 /* we should mark TOWRITE bit to each dirty page in radix tree to
1220 * verify pages have been written, but this is difficult because of
1226 static int vvp_io_read_page(const struct lu_env *env,
1227 const struct cl_io_slice *ios,
1228 const struct cl_page_slice *slice)
1230 struct cl_io *io = ios->cis_io;
1231 struct vvp_page *vpg = cl2vvp_page(slice);
1232 struct cl_page *page = slice->cpl_page;
1233 struct inode *inode = vvp_object_inode(slice->cpl_obj);
1234 struct ll_sb_info *sbi = ll_i2sbi(inode);
1235 struct ll_file_data *fd = cl2vvp_io(env, ios)->vui_fd;
1236 struct ll_readahead_state *ras = &fd->fd_ras;
1237 struct cl_2queue *queue = &io->ci_queue;
1239 if (sbi->ll_ra_info.ra_max_pages_per_file &&
1240 sbi->ll_ra_info.ra_max_pages)
1241 ras_update(sbi, inode, ras, vvp_index(vpg),
1242 vpg->vpg_defer_uptodate);
1244 if (vpg->vpg_defer_uptodate) {
1245 vpg->vpg_ra_used = 1;
1246 cl_page_export(env, page, 1);
1249 * Add page into the queue even when it is marked uptodate above.
1250 * this will unlock it automatically as part of cl_page_list_disown().
1253 cl_page_list_add(&queue->c2_qin, page);
1254 if (sbi->ll_ra_info.ra_max_pages_per_file &&
1255 sbi->ll_ra_info.ra_max_pages)
1256 ll_readahead(env, io, &queue->c2_qin, ras,
1257 vpg->vpg_defer_uptodate);
1262 void vvp_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
1264 CLOBINVRNT(env, ios->cis_io->ci_obj,
1265 vvp_object_invariant(ios->cis_io->ci_obj));
1268 static const struct cl_io_operations vvp_io_ops = {
1271 .cio_fini = vvp_io_fini,
1272 .cio_lock = vvp_io_read_lock,
1273 .cio_start = vvp_io_read_start,
1274 .cio_advance = vvp_io_advance,
1277 .cio_fini = vvp_io_fini,
1278 .cio_iter_init = vvp_io_write_iter_init,
1279 .cio_iter_fini = vvp_io_write_iter_fini,
1280 .cio_lock = vvp_io_write_lock,
1281 .cio_start = vvp_io_write_start,
1282 .cio_advance = vvp_io_advance,
1285 .cio_fini = vvp_io_setattr_fini,
1286 .cio_iter_init = vvp_io_setattr_iter_init,
1287 .cio_lock = vvp_io_setattr_lock,
1288 .cio_start = vvp_io_setattr_start,
1289 .cio_end = vvp_io_setattr_end
1292 .cio_fini = vvp_io_fault_fini,
1293 .cio_iter_init = vvp_io_fault_iter_init,
1294 .cio_lock = vvp_io_fault_lock,
1295 .cio_start = vvp_io_fault_start,
1296 .cio_end = vvp_io_end,
1299 .cio_start = vvp_io_fsync_start,
1300 .cio_fini = vvp_io_fini
1303 .cio_fini = vvp_io_fini
1306 .cio_read_page = vvp_io_read_page,
1309 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1312 struct vvp_io *vio = vvp_env_io(env);
1313 struct inode *inode = vvp_object_inode(obj);
1316 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1318 CDEBUG(D_VFSTRACE, DFID
1319 " ignore/verify layout %d/%d, layout version %d restore needed %d\n",
1320 PFID(lu_object_fid(&obj->co_lu)),
1321 io->ci_ignore_layout, io->ci_verify_layout,
1322 vio->vui_layout_gen, io->ci_restore_needed);
1324 CL_IO_SLICE_CLEAN(vio, vui_cl);
1325 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1326 vio->vui_ra_valid = false;
1328 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1330 struct ll_inode_info *lli = ll_i2info(inode);
1332 count = io->u.ci_rw.crw_count;
1333 /* "If nbyte is 0, read() will return 0 and have no other
1334 * results." -- Single Unix Spec
1339 vio->vui_tot_count = count;
1341 /* for read/write, we store the jobid in the inode, and
1342 * it'll be fetched by osc when building RPC.
1344 * it's not accurate if the file is shared by different
1347 lustre_get_jobid(lli->lli_jobid);
1348 } else if (io->ci_type == CIT_SETATTR) {
1349 if (!cl_io_is_trunc(io))
1350 io->ci_lockreq = CILR_MANDATORY;
1353 /* ignore layout change for generic CIT_MISC but not for glimpse.
1354 * io context for glimpse must set ci_verify_layout to true,
1355 * see cl_glimpse_size0() for details.
1357 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1358 io->ci_ignore_layout = 1;
1360 /* Enqueue layout lock and get layout version. We need to do this
1361 * even for operations requiring to open file, such as read and write,
1362 * because it might not grant layout lock in IT_OPEN.
1364 if (result == 0 && !io->ci_ignore_layout) {
1365 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1366 if (result == -ENOENT)
1367 /* If the inode on MDS has been removed, but the objects
1368 * on OSTs haven't been destroyed (async unlink), layout
1369 * fetch will return -ENOENT, we'd ignore this error
1370 * and continue with dirty flush. LU-3230.
1374 CERROR("%s: refresh file layout " DFID " error %d.\n",
1375 ll_get_fsname(inode->i_sb, NULL, 0),
1376 PFID(lu_object_fid(&obj->co_lu)), result);