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->cui_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 *cio = 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) != cio->cui_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 *cio = cl2vvp_io(env, ios);
105 cl_page_list_init(&cio->u.write.cui_queue);
106 cio->u.write.cui_written = 0;
107 cio->u.write.cui_from = 0;
108 cio->u.write.cui_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 *cio = cl2vvp_io(env, ios);
118 LASSERT(cio->u.write.cui_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->cui_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 *cio = 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 cio->cui_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 = cio->cui_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 cio->cui_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->cui_iov == NULL) /* nfs or loop back device write */
245 /* No MM (e.g. NFS)? No vmas too. */
249 for (seg = 0; seg < vio->cui_nrsegs; seg++) {
250 const struct iovec *iv = &vio->cui_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 *cio = vvp_env_io(env);
315 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
318 vvp_io_update_iov(env, cio, io);
320 if (io->u.ci_rw.crw_nonblock)
321 ast_flags |= CEF_NONBLOCK;
323 result = vvp_mmap_locks(env, cio, 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 *cio = 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))
406 cio->u.setattr.cui_local_lock = SETATTR_EXTENT_LOCK;
408 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
409 new_size, OBD_OBJECT_EOF);
412 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
417 * Only ll_inode_size_lock is taken at this level.
419 ll_inode_size_lock(inode);
420 result = inode_newsize_ok(inode, size);
422 ll_inode_size_unlock(inode);
425 i_size_write(inode, size);
427 ll_truncate_pagecache(inode, size);
428 ll_inode_size_unlock(inode);
432 static int vvp_io_setattr_trunc(const struct lu_env *env,
433 const struct cl_io_slice *ios,
434 struct inode *inode, loff_t size)
436 inode_dio_wait(inode);
440 static int vvp_io_setattr_time(const struct lu_env *env,
441 const struct cl_io_slice *ios)
443 struct cl_io *io = ios->cis_io;
444 struct cl_object *obj = io->ci_obj;
445 struct cl_attr *attr = ccc_env_thread_attr(env);
447 unsigned valid = CAT_CTIME;
449 cl_object_attr_lock(obj);
450 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
451 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
452 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
455 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
456 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
459 result = cl_object_attr_update(env, obj, attr, valid);
460 cl_object_attr_unlock(obj);
465 static int vvp_io_setattr_start(const struct lu_env *env,
466 const struct cl_io_slice *ios)
468 struct cl_io *io = ios->cis_io;
469 struct inode *inode = vvp_object_inode(io->ci_obj);
472 mutex_lock(&inode->i_mutex);
473 if (cl_io_is_trunc(io))
474 result = vvp_io_setattr_trunc(env, ios, inode,
475 io->u.ci_setattr.sa_attr.lvb_size);
477 result = vvp_io_setattr_time(env, ios);
481 static void vvp_io_setattr_end(const struct lu_env *env,
482 const struct cl_io_slice *ios)
484 struct cl_io *io = ios->cis_io;
485 struct inode *inode = vvp_object_inode(io->ci_obj);
487 if (cl_io_is_trunc(io)) {
488 /* Truncate in memory pages - they must be clean pages
489 * because osc has already notified to destroy osc_extents. */
490 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
491 inode_dio_write_done(inode);
493 mutex_unlock(&inode->i_mutex);
496 static void vvp_io_setattr_fini(const struct lu_env *env,
497 const struct cl_io_slice *ios)
499 vvp_io_fini(env, ios);
502 static int vvp_io_read_start(const struct lu_env *env,
503 const struct cl_io_slice *ios)
505 struct vvp_io *vio = cl2vvp_io(env, ios);
506 struct cl_io *io = ios->cis_io;
507 struct cl_object *obj = io->ci_obj;
508 struct inode *inode = vvp_object_inode(obj);
509 struct file *file = vio->cui_fd->fd_file;
512 loff_t pos = io->u.ci_rd.rd.crw_pos;
513 long cnt = io->u.ci_rd.rd.crw_count;
514 long tot = vio->cui_tot_count;
517 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
519 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
521 if (!can_populate_pages(env, io, inode))
524 result = ccc_prep_size(env, obj, io, pos, tot, &exceed);
527 else if (exceed != 0)
530 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
531 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
532 inode->i_ino, cnt, pos, i_size_read(inode));
534 /* turn off the kernel's read-ahead */
535 vio->cui_fd->fd_file->f_ra.ra_pages = 0;
537 /* initialize read-ahead window once per syscall */
538 if (!vio->cui_ra_valid) {
539 vio->cui_ra_valid = true;
540 vio->cui_ra_start = cl_index(obj, pos);
541 vio->cui_ra_count = cl_index(obj, tot + PAGE_CACHE_SIZE - 1);
547 switch (vio->cui_io_subtype) {
549 LASSERT(vio->cui_iocb->ki_pos == pos);
550 result = generic_file_aio_read(vio->cui_iocb,
551 vio->cui_iov, vio->cui_nrsegs,
552 vio->cui_iocb->ki_pos);
555 result = generic_file_splice_read(file, &pos,
556 vio->u.splice.cui_pipe, cnt,
557 vio->u.splice.cui_flags);
558 /* LU-1109: do splice read stripe by stripe otherwise if it
559 * may make nfsd stuck if this read occupied all internal pipe
564 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
572 io->ci_nob += result;
573 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, vio->cui_fd,
581 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
582 struct cl_page_list *plist, int from, int to)
584 struct cl_2queue *queue = &io->ci_queue;
585 struct cl_page *page;
586 unsigned int bytes = 0;
590 if (plist->pl_nr == 0)
593 if (from > 0 || to != PAGE_SIZE) {
594 page = cl_page_list_first(plist);
595 if (plist->pl_nr == 1) {
596 cl_page_clip(env, page, from, to);
599 cl_page_clip(env, page, from, PAGE_SIZE);
600 if (to != PAGE_SIZE) {
601 page = cl_page_list_last(plist);
602 cl_page_clip(env, page, 0, to);
607 cl_2queue_init(queue);
608 cl_page_list_splice(plist, &queue->c2_qin);
609 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
611 /* plist is not sorted any more */
612 cl_page_list_splice(&queue->c2_qin, plist);
613 cl_page_list_splice(&queue->c2_qout, plist);
614 cl_2queue_fini(env, queue);
617 /* calculate bytes */
618 bytes = plist->pl_nr << PAGE_SHIFT;
619 bytes -= from + PAGE_SIZE - to;
621 while (plist->pl_nr > 0) {
622 page = cl_page_list_first(plist);
623 cl_page_list_del(env, plist, page);
625 cl_page_clip(env, page, 0, PAGE_SIZE);
627 SetPageUptodate(cl_page_vmpage(page));
628 cl_page_disown(env, io, page);
630 /* held in ll_cl_init() */
631 lu_ref_del(&page->cp_reference, "cl_io", io);
632 cl_page_put(env, page);
636 RETURN(bytes > 0 ? bytes : rc);
639 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
640 struct cl_page *page)
642 struct vvp_page *vpg;
643 struct page *vmpage = page->cp_vmpage;
644 struct cl_object *clob = cl_io_top(io)->ci_obj;
646 SetPageUptodate(vmpage);
647 set_page_dirty(vmpage);
649 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
650 vvp_write_pending(cl2vvp(clob), vpg);
652 cl_page_disown(env, io, page);
654 /* held in ll_cl_init() */
655 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
656 cl_page_put(env, page);
659 /* make sure the page list is contiguous */
660 static bool page_list_sanity_check(struct cl_object *obj,
661 struct cl_page_list *plist)
663 struct cl_page *page;
664 pgoff_t index = CL_PAGE_EOF;
666 cl_page_list_for_each(page, plist) {
667 struct vvp_page *vpg = cl_object_page_slice(obj, page);
669 if (index == CL_PAGE_EOF) {
670 index = vvp_index(vpg);
675 if (index == vvp_index(vpg))
683 /* Return how many bytes have queued or written */
684 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
686 struct cl_object *obj = io->ci_obj;
687 struct inode *inode = vvp_object_inode(obj);
688 struct vvp_io *cio = vvp_env_io(env);
689 struct cl_page_list *queue = &cio->u.write.cui_queue;
690 struct cl_page *page;
693 unsigned int npages = cio->u.write.cui_queue.pl_nr;
699 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
700 npages, cio->u.write.cui_from, cio->u.write.cui_to);
702 LASSERT(page_list_sanity_check(obj, queue));
704 /* submit IO with async write */
705 rc = cl_io_commit_async(env, io, queue,
706 cio->u.write.cui_from, cio->u.write.cui_to,
707 write_commit_callback);
708 npages -= queue->pl_nr; /* already committed pages */
710 /* calculate how many bytes were written */
711 bytes = npages << PAGE_SHIFT;
714 bytes -= cio->u.write.cui_from;
715 if (queue->pl_nr == 0) /* last page */
716 bytes -= PAGE_SIZE - cio->u.write.cui_to;
717 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
719 cio->u.write.cui_written += bytes;
721 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
722 npages, bytes, cio->u.write.cui_written);
724 /* the first page must have been written. */
725 cio->u.write.cui_from = 0;
727 LASSERT(page_list_sanity_check(obj, queue));
728 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
730 /* out of quota, try sync write */
731 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
732 rc = vvp_io_commit_sync(env, io, queue,
733 cio->u.write.cui_from,
734 cio->u.write.cui_to);
736 cio->u.write.cui_written += rc;
741 /* update inode size */
742 ll_merge_attr(env, inode);
744 /* Now the pages in queue were failed to commit, discard them
745 * unless they were dirtied before. */
746 while (queue->pl_nr > 0) {
747 page = cl_page_list_first(queue);
748 cl_page_list_del(env, queue, page);
750 if (!PageDirty(cl_page_vmpage(page)))
751 cl_page_discard(env, io, page);
753 cl_page_disown(env, io, page);
755 /* held in ll_cl_init() */
756 lu_ref_del(&page->cp_reference, "cl_io", io);
757 cl_page_put(env, page);
759 cl_page_list_fini(env, queue);
764 static int vvp_io_write_start(const struct lu_env *env,
765 const struct cl_io_slice *ios)
767 struct vvp_io *cio = cl2vvp_io(env, ios);
768 struct cl_io *io = ios->cis_io;
769 struct cl_object *obj = io->ci_obj;
770 struct inode *inode = vvp_object_inode(obj);
772 loff_t pos = io->u.ci_wr.wr.crw_pos;
773 size_t cnt = io->u.ci_wr.wr.crw_count;
777 if (!can_populate_pages(env, io, inode))
780 if (cl_io_is_append(io)) {
782 * PARALLEL IO This has to be changed for parallel IO doing
783 * out-of-order writes.
785 ll_merge_attr(env, inode);
786 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
787 cio->cui_iocb->ki_pos = pos;
789 LASSERT(cio->cui_iocb->ki_pos == pos);
792 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
794 if (cio->cui_iov == NULL) {
795 /* from a temp io in ll_cl_init(). */
799 * When using the locked AIO function (generic_file_aio_write())
800 * testing has shown the inode mutex to be a limiting factor
801 * with multi-threaded single shared file performance. To get
802 * around this, we now use the lockless version. To maintain
803 * consistency, proper locking to protect against writes,
804 * trucates, etc. is handled in the higher layers of lustre.
806 result = __generic_file_aio_write(cio->cui_iocb,
807 cio->cui_iov, cio->cui_nrsegs,
808 &cio->cui_iocb->ki_pos);
809 if (result > 0 || result == -EIOCBQUEUED) {
812 err = generic_write_sync(cio->cui_iocb->ki_filp,
814 if (err < 0 && result > 0)
820 result = vvp_io_write_commit(env, io);
821 if (cio->u.write.cui_written > 0) {
822 result = cio->u.write.cui_written;
823 io->ci_nob += result;
825 CDEBUG(D_VFSTRACE, "write: nob %zd, result: %zd\n",
830 struct ll_inode_info *lli = ll_i2info(inode);
832 spin_lock(&lli->lli_lock);
833 lli->lli_flags |= LLIF_DATA_MODIFIED;
834 spin_unlock(&lli->lli_lock);
838 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
839 cio->cui_fd, pos, result, WRITE);
846 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
848 struct vm_fault *vmf = cfio->ft_vmf;
850 cfio->ft_flags = filemap_fault(cfio->ft_vma, vmf);
851 cfio->ft_flags_valid = 1;
854 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
855 vmf->virtual_address);
856 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
857 lock_page(vmf->page);
858 cfio->ft_flags |= VM_FAULT_LOCKED;
861 cfio->ft_vmpage = vmf->page;
866 if (cfio->ft_flags & VM_FAULT_SIGBUS) {
867 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
871 if (cfio->ft_flags & VM_FAULT_OOM) {
872 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
876 if (cfio->ft_flags & VM_FAULT_RETRY)
879 CERROR("unknown error in page fault %d\n", cfio->ft_flags);
884 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
885 struct cl_page *page)
887 struct vvp_page *vpg;
888 struct cl_object *clob = cl_io_top(io)->ci_obj;
890 set_page_dirty(page->cp_vmpage);
892 vpg = cl2vvp_page(cl_object_page_slice(clob, page));
893 vvp_write_pending(cl2vvp(clob), vpg);
896 static int vvp_io_fault_start(const struct lu_env *env,
897 const struct cl_io_slice *ios)
899 struct vvp_io *vio = cl2vvp_io(env, ios);
900 struct cl_io *io = ios->cis_io;
901 struct cl_object *obj = io->ci_obj;
902 struct inode *inode = vvp_object_inode(obj);
903 struct cl_fault_io *fio = &io->u.ci_fault;
904 struct vvp_fault_io *cfio = &vio->u.fault;
907 struct page *vmpage = NULL;
908 struct cl_page *page;
913 if (fio->ft_executable &&
914 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
916 " changed while waiting for the page fault lock\n",
917 PFID(lu_object_fid(&obj->co_lu)));
919 /* offset of the last byte on the page */
920 offset = cl_offset(obj, fio->ft_index + 1) - 1;
921 LASSERT(cl_index(obj, offset) == fio->ft_index);
922 result = ccc_prep_size(env, obj, io, 0, offset + 1, NULL);
926 /* must return locked page */
927 if (fio->ft_mkwrite) {
928 LASSERT(cfio->ft_vmpage != NULL);
929 lock_page(cfio->ft_vmpage);
931 result = vvp_io_kernel_fault(cfio);
936 vmpage = cfio->ft_vmpage;
937 LASSERT(PageLocked(vmpage));
939 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
940 ll_invalidate_page(vmpage);
942 size = i_size_read(inode);
943 /* Though we have already held a cl_lock upon this page, but
944 * it still can be truncated locally. */
945 if (unlikely((vmpage->mapping != inode->i_mapping) ||
946 (page_offset(vmpage) > size))) {
947 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
949 /* return +1 to stop cl_io_loop() and ll_fault() will catch
951 GOTO(out, result = +1);
954 last_index = cl_index(obj, size - 1);
956 if (fio->ft_mkwrite ) {
958 * Capture the size while holding the lli_trunc_sem from above
959 * we want to make sure that we complete the mkwrite action
960 * while holding this lock. We need to make sure that we are
961 * not past the end of the file.
963 if (last_index < fio->ft_index) {
965 "llite: mkwrite and truncate race happened: "
967 vmpage->mapping,fio->ft_index,last_index);
969 * We need to return if we are
970 * passed the end of the file. This will propagate
971 * up the call stack to ll_page_mkwrite where
972 * we will return VM_FAULT_NOPAGE. Any non-negative
973 * value returned here will be silently
974 * converted to 0. If the vmpage->mapping is null
975 * the error code would be converted back to ENODATA
976 * in ll_page_mkwrite0. Thus we return -ENODATA
977 * to handle both cases
979 GOTO(out, result = -ENODATA);
983 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
985 GOTO(out, result = PTR_ERR(page));
987 /* if page is going to be written, we should add this page into cache
989 if (fio->ft_mkwrite) {
990 wait_on_page_writeback(vmpage);
991 if (!PageDirty(vmpage)) {
992 struct cl_page_list *plist = &io->ci_queue.c2_qin;
993 struct vvp_page *vpg = cl_object_page_slice(obj, page);
996 /* vvp_page_assume() calls wait_on_page_writeback(). */
997 cl_page_assume(env, io, page);
999 cl_page_list_init(plist);
1000 cl_page_list_add(plist, page);
1003 if (last_index == vvp_index(vpg))
1004 to = size & ~CFS_PAGE_MASK;
1006 /* Do not set Dirty bit here so that in case IO is
1007 * started before the page is really made dirty, we
1008 * still have chance to detect it. */
1009 result = cl_io_commit_async(env, io, plist, 0, to,
1010 mkwrite_commit_callback);
1011 LASSERT(cl_page_is_owned(page, io));
1012 cl_page_list_fini(env, plist);
1016 cl_page_discard(env, io, page);
1017 cl_page_disown(env, io, page);
1019 cl_page_put(env, page);
1021 /* we're in big trouble, what can we do now? */
1022 if (result == -EDQUOT)
1026 cl_page_disown(env, io, page);
1031 * The ft_index is only used in the case of
1032 * a mkwrite action. We need to check
1033 * our assertions are correct, since
1034 * we should have caught this above
1036 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1037 if (fio->ft_index == last_index)
1039 * Last page is mapped partially.
1041 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1043 fio->ft_nob = cl_page_size(obj);
1045 lu_ref_add(&page->cp_reference, "fault", io);
1046 fio->ft_page = page;
1050 /* return unlocked vmpage to avoid deadlocking */
1052 unlock_page(vmpage);
1054 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1059 static int vvp_io_fsync_start(const struct lu_env *env,
1060 const struct cl_io_slice *ios)
1062 /* we should mark TOWRITE bit to each dirty page in radix tree to
1063 * verify pages have been written, but this is difficult because of
1068 static int vvp_io_read_page(const struct lu_env *env,
1069 const struct cl_io_slice *ios,
1070 const struct cl_page_slice *slice)
1072 struct cl_io *io = ios->cis_io;
1073 struct vvp_page *vpg = cl2vvp_page(slice);
1074 struct cl_page *page = slice->cpl_page;
1075 struct inode *inode = vvp_object_inode(slice->cpl_obj);
1076 struct ll_sb_info *sbi = ll_i2sbi(inode);
1077 struct ll_file_data *fd = cl2vvp_io(env, ios)->cui_fd;
1078 struct ll_readahead_state *ras = &fd->fd_ras;
1079 struct cl_2queue *queue = &io->ci_queue;
1083 if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
1084 sbi->ll_ra_info.ra_max_pages > 0)
1085 ras_update(sbi, inode, ras, vvp_index(vpg),
1086 vpg->vpg_defer_uptodate);
1088 if (vpg->vpg_defer_uptodate) {
1089 vpg->vpg_ra_used = 1;
1090 cl_page_export(env, page, 1);
1094 * Add page into the queue even when it is marked uptodate above.
1095 * this will unlock it automatically as part of cl_page_list_disown().
1097 cl_2queue_add(queue, page);
1098 if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
1099 sbi->ll_ra_info.ra_max_pages > 0)
1100 ll_readahead(env, io, &queue->c2_qin, ras,
1101 vpg->vpg_defer_uptodate);
1106 static const struct cl_io_operations vvp_io_ops = {
1109 .cio_fini = vvp_io_fini,
1110 .cio_lock = vvp_io_read_lock,
1111 .cio_start = vvp_io_read_start,
1112 .cio_advance = vvp_io_advance,
1115 .cio_fini = vvp_io_fini,
1116 .cio_iter_init = vvp_io_write_iter_init,
1117 .cio_iter_fini = vvp_io_write_iter_fini,
1118 .cio_lock = vvp_io_write_lock,
1119 .cio_start = vvp_io_write_start,
1120 .cio_advance = vvp_io_advance,
1123 .cio_fini = vvp_io_setattr_fini,
1124 .cio_iter_init = vvp_io_setattr_iter_init,
1125 .cio_lock = vvp_io_setattr_lock,
1126 .cio_start = vvp_io_setattr_start,
1127 .cio_end = vvp_io_setattr_end
1130 .cio_fini = vvp_io_fault_fini,
1131 .cio_iter_init = vvp_io_fault_iter_init,
1132 .cio_lock = vvp_io_fault_lock,
1133 .cio_start = vvp_io_fault_start,
1134 .cio_end = vvp_io_end,
1137 .cio_start = vvp_io_fsync_start,
1138 .cio_fini = vvp_io_fini
1141 .cio_fini = vvp_io_fini
1144 .cio_read_page = vvp_io_read_page,
1147 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1150 struct vvp_io *vio = vvp_env_io(env);
1151 struct inode *inode = vvp_object_inode(obj);
1154 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1157 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1158 "restore needed %d\n",
1159 PFID(lu_object_fid(&obj->co_lu)),
1160 io->ci_ignore_layout, io->ci_verify_layout,
1161 vio->cui_layout_gen, io->ci_restore_needed);
1163 CL_IO_SLICE_CLEAN(vio, cui_cl);
1164 cl_io_slice_add(io, &vio->cui_cl, obj, &vvp_io_ops);
1165 vio->cui_ra_valid = false;
1167 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1169 struct ll_inode_info *lli = ll_i2info(inode);
1171 count = io->u.ci_rw.crw_count;
1172 /* "If nbyte is 0, read() will return 0 and have no other
1173 * results." -- Single Unix Spec */
1177 vio->cui_tot_count = count;
1178 vio->cui_tot_nrsegs = 0;
1181 /* for read/write, we store the jobid in the inode, and
1182 * it'll be fetched by osc when building RPC.
1184 * it's not accurate if the file is shared by different
1187 lustre_get_jobid(lli->lli_jobid);
1188 } else if (io->ci_type == CIT_SETATTR) {
1189 if (!cl_io_is_trunc(io))
1190 io->ci_lockreq = CILR_MANDATORY;
1193 /* ignore layout change for generic CIT_MISC but not for glimpse.
1194 * io context for glimpse must set ci_verify_layout to true,
1195 * see cl_glimpse_size0() for details. */
1196 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1197 io->ci_ignore_layout = 1;
1199 /* Enqueue layout lock and get layout version. We need to do this
1200 * even for operations requiring to open file, such as read and write,
1201 * because it might not grant layout lock in IT_OPEN. */
1202 if (result == 0 && !io->ci_ignore_layout) {
1203 result = ll_layout_refresh(inode, &vio->cui_layout_gen);
1204 if (result == -ENOENT)
1205 /* If the inode on MDS has been removed, but the objects
1206 * on OSTs haven't been destroyed (async unlink), layout
1207 * fetch will return -ENOENT, we'd ingore this error
1208 * and continue with dirty flush. LU-3230. */
1211 CERROR("%s: refresh file layout " DFID " error %d.\n",
1212 ll_get_fsname(inode->i_sb, NULL, 0),
1213 PFID(lu_object_fid(&obj->co_lu)), result);