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, 2012, 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
45 #include "../include/obd.h"
46 #include "../include/lustre_lite.h"
48 #include "vvp_internal.h"
50 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
51 const struct cl_io_slice *slice);
54 * True, if \a io is a normal io, False for splice_{read,write}
56 int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
58 struct vvp_io *vio = vvp_env_io(env);
60 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
62 return vio->cui_io_subtype == IO_NORMAL;
66 * For swapping layout. The file's layout may have changed.
67 * To avoid populating pages to a wrong stripe, we have to verify the
68 * correctness of layout. It works because swapping layout processes
69 * have to acquire group lock.
71 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
74 struct ll_inode_info *lli = ll_i2info(inode);
75 struct ccc_io *cio = ccc_env_io(env);
78 switch (io->ci_type) {
81 /* don't need lock here to check lli_layout_gen as we have held
82 * extent lock and GROUP lock has to hold to swap layout */
83 if (ll_layout_version_get(lli) != cio->cui_layout_gen) {
84 io->ci_need_restart = 1;
85 /* this will return application a short read/write */
90 /* fault is okay because we've already had a page. */
98 /*****************************************************************************
104 static int vvp_io_fault_iter_init(const struct lu_env *env,
105 const struct cl_io_slice *ios)
107 struct vvp_io *vio = cl2vvp_io(env, ios);
108 struct inode *inode = ccc_object_inode(ios->cis_obj);
111 file_inode(cl2ccc_io(env, ios)->cui_fd->fd_file));
112 vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime);
116 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
118 struct cl_io *io = ios->cis_io;
119 struct cl_object *obj = io->ci_obj;
120 struct ccc_io *cio = cl2ccc_io(env, ios);
122 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
124 CDEBUG(D_VFSTRACE, DFID
125 " ignore/verify layout %d/%d, layout version %d restore needed %d\n",
126 PFID(lu_object_fid(&obj->co_lu)),
127 io->ci_ignore_layout, io->ci_verify_layout,
128 cio->cui_layout_gen, io->ci_restore_needed);
130 if (io->ci_restore_needed == 1) {
133 /* file was detected release, we need to restore it
134 * before finishing the io
136 rc = ll_layout_restore(ccc_object_inode(obj));
137 /* if restore registration failed, no restart,
138 * we will return -ENODATA */
139 /* The layout will change after restore, so we need to
140 * block on layout lock hold by the MDT
141 * as MDT will not send new layout in lvb (see LU-3124)
142 * we have to explicitly fetch it, all this will be done
143 * by ll_layout_refresh()
146 io->ci_restore_needed = 0;
147 io->ci_need_restart = 1;
148 io->ci_verify_layout = 1;
150 io->ci_restore_needed = 1;
151 io->ci_need_restart = 0;
152 io->ci_verify_layout = 0;
157 if (!io->ci_ignore_layout && io->ci_verify_layout) {
160 /* check layout version */
161 ll_layout_refresh(ccc_object_inode(obj), &gen);
162 io->ci_need_restart = cio->cui_layout_gen != gen;
163 if (io->ci_need_restart) {
165 DFID" layout changed from %d to %d.\n",
166 PFID(lu_object_fid(&obj->co_lu)),
167 cio->cui_layout_gen, gen);
168 /* today successful restore is the only possible
170 /* restore was done, clear restoring state */
171 ll_i2info(ccc_object_inode(obj))->lli_flags &=
172 ~LLIF_FILE_RESTORING;
177 static void vvp_io_fault_fini(const struct lu_env *env,
178 const struct cl_io_slice *ios)
180 struct cl_io *io = ios->cis_io;
181 struct cl_page *page = io->u.ci_fault.ft_page;
183 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
186 lu_ref_del(&page->cp_reference, "fault", io);
187 cl_page_put(env, page);
188 io->u.ci_fault.ft_page = NULL;
190 vvp_io_fini(env, ios);
193 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
196 * we only want to hold PW locks if the mmap() can generate
197 * writes back to the file and that only happens in shared
200 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
205 static int vvp_mmap_locks(const struct lu_env *env,
206 struct ccc_io *vio, struct cl_io *io)
208 struct ccc_thread_info *cti = ccc_env_info(env);
209 struct mm_struct *mm = current->mm;
210 struct vm_area_struct *vma;
211 struct cl_lock_descr *descr = &cti->cti_descr;
212 ldlm_policy_data_t policy;
219 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
221 if (!cl_is_normalio(env, io))
224 if (vio->cui_iter == NULL) /* nfs or loop back device write */
227 /* No MM (e.g. NFS)? No vmas too. */
231 iov_for_each(iov, i, *(vio->cui_iter)) {
232 addr = (unsigned long)iov.iov_base;
237 count += addr & (~CFS_PAGE_MASK);
238 addr &= CFS_PAGE_MASK;
240 down_read(&mm->mmap_sem);
241 while ((vma = our_vma(mm, addr, count)) != NULL) {
242 struct inode *inode = file_inode(vma->vm_file);
243 int flags = CEF_MUST;
245 if (ll_file_nolock(vma->vm_file)) {
247 * For no lock case, a lockless lock will be
254 * XXX: Required lock mode can be weakened: CIT_WRITE
255 * io only ever reads user level buffer, and CIT_READ
258 policy_from_vma(&policy, vma, addr, count);
259 descr->cld_mode = vvp_mode_from_vma(vma);
260 descr->cld_obj = ll_i2info(inode)->lli_clob;
261 descr->cld_start = cl_index(descr->cld_obj,
262 policy.l_extent.start);
263 descr->cld_end = cl_index(descr->cld_obj,
264 policy.l_extent.end);
265 descr->cld_enq_flags = flags;
266 result = cl_io_lock_alloc_add(env, io, descr);
268 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
269 descr->cld_mode, descr->cld_start,
273 up_read(&mm->mmap_sem);
277 if (vma->vm_end - addr >= count)
280 count -= vma->vm_end - addr;
283 up_read(&mm->mmap_sem);
288 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
289 enum cl_lock_mode mode, loff_t start, loff_t end)
291 struct ccc_io *cio = ccc_env_io(env);
295 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
297 ccc_io_update_iov(env, cio, io);
299 if (io->u.ci_rw.crw_nonblock)
300 ast_flags |= CEF_NONBLOCK;
301 result = vvp_mmap_locks(env, cio, io);
303 result = ccc_io_one_lock(env, io, ast_flags, mode, start, end);
307 static int vvp_io_read_lock(const struct lu_env *env,
308 const struct cl_io_slice *ios)
310 struct cl_io *io = ios->cis_io;
311 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
314 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
315 rd->crw_pos + rd->crw_count - 1);
320 static int vvp_io_fault_lock(const struct lu_env *env,
321 const struct cl_io_slice *ios)
323 struct cl_io *io = ios->cis_io;
324 struct vvp_io *vio = cl2vvp_io(env, ios);
326 * XXX LDLM_FL_CBPENDING
328 return ccc_io_one_lock_index
329 (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma),
330 io->u.ci_fault.ft_index, io->u.ci_fault.ft_index);
333 static int vvp_io_write_lock(const struct lu_env *env,
334 const struct cl_io_slice *ios)
336 struct cl_io *io = ios->cis_io;
340 if (io->u.ci_wr.wr_append) {
342 end = OBD_OBJECT_EOF;
344 start = io->u.ci_wr.wr.crw_pos;
345 end = start + io->u.ci_wr.wr.crw_count - 1;
347 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
350 static int vvp_io_setattr_iter_init(const struct lu_env *env,
351 const struct cl_io_slice *ios)
357 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
359 * Handles "lockless io" mode when extent locking is done by server.
361 static int vvp_io_setattr_lock(const struct lu_env *env,
362 const struct cl_io_slice *ios)
364 struct ccc_io *cio = ccc_env_io(env);
365 struct cl_io *io = ios->cis_io;
369 if (cl_io_is_trunc(io)) {
370 new_size = io->u.ci_setattr.sa_attr.lvb_size;
372 enqflags = CEF_DISCARD_DATA;
374 if ((io->u.ci_setattr.sa_attr.lvb_mtime >=
375 io->u.ci_setattr.sa_attr.lvb_ctime) ||
376 (io->u.ci_setattr.sa_attr.lvb_atime >=
377 io->u.ci_setattr.sa_attr.lvb_ctime))
381 cio->u.setattr.cui_local_lock = SETATTR_EXTENT_LOCK;
382 return ccc_io_one_lock(env, io, enqflags, CLM_WRITE,
383 new_size, OBD_OBJECT_EOF);
386 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
390 * Only ll_inode_size_lock is taken at this level.
392 ll_inode_size_lock(inode);
393 result = inode_newsize_ok(inode, size);
395 ll_inode_size_unlock(inode);
398 truncate_setsize(inode, size);
399 ll_inode_size_unlock(inode);
403 static int vvp_io_setattr_trunc(const struct lu_env *env,
404 const struct cl_io_slice *ios,
405 struct inode *inode, loff_t size)
407 inode_dio_wait(inode);
411 static int vvp_io_setattr_time(const struct lu_env *env,
412 const struct cl_io_slice *ios)
414 struct cl_io *io = ios->cis_io;
415 struct cl_object *obj = io->ci_obj;
416 struct cl_attr *attr = ccc_env_thread_attr(env);
418 unsigned valid = CAT_CTIME;
420 cl_object_attr_lock(obj);
421 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
422 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
423 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
426 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
427 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
430 result = cl_object_attr_set(env, obj, attr, valid);
431 cl_object_attr_unlock(obj);
436 static int vvp_io_setattr_start(const struct lu_env *env,
437 const struct cl_io_slice *ios)
439 struct cl_io *io = ios->cis_io;
440 struct inode *inode = ccc_object_inode(io->ci_obj);
443 mutex_lock(&inode->i_mutex);
444 if (cl_io_is_trunc(io))
445 result = vvp_io_setattr_trunc(env, ios, inode,
446 io->u.ci_setattr.sa_attr.lvb_size);
448 result = vvp_io_setattr_time(env, ios);
452 static void vvp_io_setattr_end(const struct lu_env *env,
453 const struct cl_io_slice *ios)
455 struct cl_io *io = ios->cis_io;
456 struct inode *inode = ccc_object_inode(io->ci_obj);
458 if (cl_io_is_trunc(io)) {
459 /* Truncate in memory pages - they must be clean pages
460 * because osc has already notified to destroy osc_extents. */
461 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
462 inode_dio_write_done(inode);
464 mutex_unlock(&inode->i_mutex);
467 static void vvp_io_setattr_fini(const struct lu_env *env,
468 const struct cl_io_slice *ios)
470 vvp_io_fini(env, ios);
473 static int vvp_io_read_start(const struct lu_env *env,
474 const struct cl_io_slice *ios)
476 struct vvp_io *vio = cl2vvp_io(env, ios);
477 struct ccc_io *cio = cl2ccc_io(env, ios);
478 struct cl_io *io = ios->cis_io;
479 struct cl_object *obj = io->ci_obj;
480 struct inode *inode = ccc_object_inode(obj);
481 struct ll_ra_read *bead = &vio->cui_bead;
482 struct file *file = cio->cui_fd->fd_file;
485 loff_t pos = io->u.ci_rd.rd.crw_pos;
486 long cnt = io->u.ci_rd.rd.crw_count;
487 long tot = cio->cui_tot_count;
490 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
492 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
494 if (!can_populate_pages(env, io, inode))
497 result = ccc_prep_size(env, obj, io, pos, tot, &exceed);
500 else if (exceed != 0)
503 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
504 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
505 inode->i_ino, cnt, pos, i_size_read(inode));
507 /* turn off the kernel's read-ahead */
508 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
510 /* initialize read-ahead window once per syscall */
511 if (!vio->cui_ra_window_set) {
512 vio->cui_ra_window_set = 1;
513 bead->lrr_start = cl_index(obj, pos);
515 * XXX: explicit PAGE_CACHE_SIZE
517 bead->lrr_count = cl_index(obj, tot + PAGE_CACHE_SIZE - 1);
518 ll_ra_read_in(file, bead);
523 switch (vio->cui_io_subtype) {
525 LASSERT(cio->cui_iocb->ki_pos == pos);
526 result = generic_file_read_iter(cio->cui_iocb, cio->cui_iter);
529 result = generic_file_splice_read(file, &pos,
530 vio->u.splice.cui_pipe, cnt,
531 vio->u.splice.cui_flags);
532 /* LU-1109: do splice read stripe by stripe otherwise if it
533 * may make nfsd stuck if this read occupied all internal pipe
538 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
546 io->ci_nob += result;
547 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
548 cio->cui_fd, pos, result, READ);
554 static void vvp_io_read_fini(const struct lu_env *env, const struct cl_io_slice *ios)
556 struct vvp_io *vio = cl2vvp_io(env, ios);
557 struct ccc_io *cio = cl2ccc_io(env, ios);
559 if (vio->cui_ra_window_set)
560 ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
562 vvp_io_fini(env, ios);
565 static int vvp_io_write_start(const struct lu_env *env,
566 const struct cl_io_slice *ios)
568 struct ccc_io *cio = cl2ccc_io(env, ios);
569 struct cl_io *io = ios->cis_io;
570 struct cl_object *obj = io->ci_obj;
571 struct inode *inode = ccc_object_inode(obj);
573 loff_t pos = io->u.ci_wr.wr.crw_pos;
574 size_t cnt = io->u.ci_wr.wr.crw_count;
576 if (!can_populate_pages(env, io, inode))
579 if (cl_io_is_append(io)) {
581 * PARALLEL IO This has to be changed for parallel IO doing
582 * out-of-order writes.
584 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
585 cio->cui_iocb->ki_pos = pos;
587 LASSERT(cio->cui_iocb->ki_pos == pos);
590 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
592 if (cio->cui_iter == NULL) /* from a temp io in ll_cl_init(). */
595 result = generic_file_write_iter(cio->cui_iocb, cio->cui_iter);
600 io->ci_nob += result;
601 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
602 cio->cui_fd, pos, result, WRITE);
608 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
610 struct vm_fault *vmf = cfio->fault.ft_vmf;
612 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, vmf);
613 cfio->fault.ft_flags_valid = 1;
617 "page %p map %p index %lu flags %lx count %u priv %0lx: got addr %p type NOPAGE\n",
618 vmf->page, vmf->page->mapping, vmf->page->index,
619 (long)vmf->page->flags, page_count(vmf->page),
620 page_private(vmf->page), vmf->virtual_address);
621 if (unlikely(!(cfio->fault.ft_flags & VM_FAULT_LOCKED))) {
622 lock_page(vmf->page);
623 cfio->fault.ft_flags |= VM_FAULT_LOCKED;
626 cfio->ft_vmpage = vmf->page;
630 if (cfio->fault.ft_flags & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV)) {
631 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
635 if (cfio->fault.ft_flags & VM_FAULT_OOM) {
636 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
640 if (cfio->fault.ft_flags & VM_FAULT_RETRY)
643 CERROR("Unknown error in page fault %d!\n", cfio->fault.ft_flags);
648 static int vvp_io_fault_start(const struct lu_env *env,
649 const struct cl_io_slice *ios)
651 struct vvp_io *vio = cl2vvp_io(env, ios);
652 struct cl_io *io = ios->cis_io;
653 struct cl_object *obj = io->ci_obj;
654 struct inode *inode = ccc_object_inode(obj);
655 struct cl_fault_io *fio = &io->u.ci_fault;
656 struct vvp_fault_io *cfio = &vio->u.fault;
659 struct page *vmpage = NULL;
660 struct cl_page *page;
662 pgoff_t last; /* last page in a file data region */
664 if (fio->ft_executable &&
665 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
667 " changed while waiting for the page fault lock\n",
668 PFID(lu_object_fid(&obj->co_lu)));
670 /* offset of the last byte on the page */
671 offset = cl_offset(obj, fio->ft_index + 1) - 1;
672 LASSERT(cl_index(obj, offset) == fio->ft_index);
673 result = ccc_prep_size(env, obj, io, 0, offset + 1, NULL);
677 /* must return locked page */
678 if (fio->ft_mkwrite) {
679 LASSERT(cfio->ft_vmpage != NULL);
680 lock_page(cfio->ft_vmpage);
682 result = vvp_io_kernel_fault(cfio);
687 vmpage = cfio->ft_vmpage;
688 LASSERT(PageLocked(vmpage));
690 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
691 ll_invalidate_page(vmpage);
693 size = i_size_read(inode);
694 /* Though we have already held a cl_lock upon this page, but
695 * it still can be truncated locally. */
696 if (unlikely((vmpage->mapping != inode->i_mapping) ||
697 (page_offset(vmpage) > size))) {
698 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
700 /* return +1 to stop cl_io_loop() and ll_fault() will catch
707 if (fio->ft_mkwrite) {
710 * Capture the size while holding the lli_trunc_sem from above
711 * we want to make sure that we complete the mkwrite action
712 * while holding this lock. We need to make sure that we are
713 * not past the end of the file.
715 last_index = cl_index(obj, size - 1);
716 if (last_index < fio->ft_index) {
718 "llite: mkwrite and truncate race happened: %p: 0x%lx 0x%lx\n",
719 vmpage->mapping, fio->ft_index, last_index);
721 * We need to return if we are
722 * passed the end of the file. This will propagate
723 * up the call stack to ll_page_mkwrite where
724 * we will return VM_FAULT_NOPAGE. Any non-negative
725 * value returned here will be silently
726 * converted to 0. If the vmpage->mapping is null
727 * the error code would be converted back to ENODATA
728 * in ll_page_mkwrite0. Thus we return -ENODATA
729 * to handle both cases
736 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
738 result = PTR_ERR(page);
742 /* if page is going to be written, we should add this page into cache
744 if (fio->ft_mkwrite) {
745 wait_on_page_writeback(vmpage);
746 if (set_page_dirty(vmpage)) {
749 /* vvp_page_assume() calls wait_on_page_writeback(). */
750 cl_page_assume(env, io, page);
752 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
753 vvp_write_pending(cl2ccc(obj), cp);
755 /* Do not set Dirty bit here so that in case IO is
756 * started before the page is really made dirty, we
757 * still have chance to detect it. */
758 result = cl_page_cache_add(env, io, page, CRT_WRITE);
759 LASSERT(cl_page_is_owned(page, io));
763 cl_page_unmap(env, io, page);
764 cl_page_discard(env, io, page);
765 cl_page_disown(env, io, page);
767 cl_page_put(env, page);
769 /* we're in big trouble, what can we do now? */
770 if (result == -EDQUOT)
774 cl_page_disown(env, io, page);
778 last = cl_index(obj, size - 1);
780 * The ft_index is only used in the case of
781 * a mkwrite action. We need to check
782 * our assertions are correct, since
783 * we should have caught this above
785 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last);
786 if (fio->ft_index == last)
788 * Last page is mapped partially.
790 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
792 fio->ft_nob = cl_page_size(obj);
794 lu_ref_add(&page->cp_reference, "fault", io);
798 /* return unlocked vmpage to avoid deadlocking */
801 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
805 static int vvp_io_fsync_start(const struct lu_env *env,
806 const struct cl_io_slice *ios)
808 /* we should mark TOWRITE bit to each dirty page in radix tree to
809 * verify pages have been written, but this is difficult because of
814 static int vvp_io_read_page(const struct lu_env *env,
815 const struct cl_io_slice *ios,
816 const struct cl_page_slice *slice)
818 struct cl_io *io = ios->cis_io;
819 struct cl_object *obj = slice->cpl_obj;
820 struct ccc_page *cp = cl2ccc_page(slice);
821 struct cl_page *page = slice->cpl_page;
822 struct inode *inode = ccc_object_inode(obj);
823 struct ll_sb_info *sbi = ll_i2sbi(inode);
824 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
825 struct ll_readahead_state *ras = &fd->fd_ras;
826 struct page *vmpage = cp->cpg_page;
827 struct cl_2queue *queue = &io->ci_queue;
830 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
831 LASSERT(slice->cpl_obj == obj);
833 if (sbi->ll_ra_info.ra_max_pages_per_file &&
834 sbi->ll_ra_info.ra_max_pages)
835 ras_update(sbi, inode, ras, page->cp_index,
836 cp->cpg_defer_uptodate);
838 /* Sanity check whether the page is protected by a lock. */
839 rc = cl_page_is_under_lock(env, io, page);
841 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %d\n",
842 rc == -ENODATA ? "without a lock" :
848 if (cp->cpg_defer_uptodate) {
850 cl_page_export(env, page, 1);
853 * Add page into the queue even when it is marked uptodate above.
854 * this will unlock it automatically as part of cl_page_list_disown().
856 cl_2queue_add(queue, page);
857 if (sbi->ll_ra_info.ra_max_pages_per_file &&
858 sbi->ll_ra_info.ra_max_pages)
859 ll_readahead(env, io, ras,
860 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
865 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
866 struct cl_page *page, struct ccc_page *cp,
867 enum cl_req_type crt)
869 struct cl_2queue *queue;
872 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
874 queue = &io->ci_queue;
875 cl_2queue_init_page(queue, page);
877 result = cl_io_submit_sync(env, io, crt, queue, 0);
878 LASSERT(cl_page_is_owned(page, io));
882 * in CRT_WRITE case page is left locked even in case of
885 cl_page_list_disown(env, io, &queue->c2_qin);
886 cl_2queue_fini(env, queue);
892 * Prepare partially written-to page for a write.
894 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
895 struct cl_object *obj, struct cl_page *pg,
897 unsigned from, unsigned to)
899 struct cl_attr *attr = ccc_env_thread_attr(env);
900 loff_t offset = cl_offset(obj, pg->cp_index);
903 cl_object_attr_lock(obj);
904 result = cl_object_attr_get(env, obj, attr);
905 cl_object_attr_unlock(obj);
908 * If are writing to a new page, no need to read old data.
909 * The extent locking will have updated the KMS, and for our
910 * purposes here we can treat it like i_size.
912 if (attr->cat_kms <= offset) {
913 char *kaddr = kmap_atomic(cp->cpg_page);
915 memset(kaddr, 0, cl_page_size(obj));
916 kunmap_atomic(kaddr);
917 } else if (cp->cpg_defer_uptodate)
920 result = vvp_page_sync_io(env, io, pg, cp, CRT_READ);
922 * In older implementations, obdo_refresh_inode is called here
923 * to update the inode because the write might modify the
924 * object info at OST. However, this has been proven useless,
925 * since LVB functions will be called when user space program
926 * tries to retrieve inode attribute. Also, see bug 15909 for
930 cl_page_export(env, pg, 1);
935 static int vvp_io_prepare_write(const struct lu_env *env,
936 const struct cl_io_slice *ios,
937 const struct cl_page_slice *slice,
938 unsigned from, unsigned to)
940 struct cl_object *obj = slice->cpl_obj;
941 struct ccc_page *cp = cl2ccc_page(slice);
942 struct cl_page *pg = slice->cpl_page;
943 struct page *vmpage = cp->cpg_page;
947 LINVRNT(cl_page_is_vmlocked(env, pg));
948 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
952 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
953 if (!PageUptodate(vmpage)) {
955 * We're completely overwriting an existing page, so _don't_
956 * set it up to date until commit_write
958 if (from == 0 && to == PAGE_CACHE_SIZE) {
959 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
960 POISON_PAGE(page, 0x11);
962 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
965 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
969 static int vvp_io_commit_write(const struct lu_env *env,
970 const struct cl_io_slice *ios,
971 const struct cl_page_slice *slice,
972 unsigned from, unsigned to)
974 struct cl_object *obj = slice->cpl_obj;
975 struct cl_io *io = ios->cis_io;
976 struct ccc_page *cp = cl2ccc_page(slice);
977 struct cl_page *pg = slice->cpl_page;
978 struct inode *inode = ccc_object_inode(obj);
979 struct ll_sb_info *sbi = ll_i2sbi(inode);
980 struct ll_inode_info *lli = ll_i2info(inode);
981 struct page *vmpage = cp->cpg_page;
987 LINVRNT(cl_page_is_vmlocked(env, pg));
988 LASSERT(vmpage->mapping->host == inode);
990 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "committing page write\n");
991 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
994 * queue a write for some time in the future the first time we
997 * This is different from what other file systems do: they usually
998 * just mark page (and some of its buffers) dirty and rely on
999 * balance_dirty_pages() to start a write-back. Lustre wants write-back
1000 * to be started earlier for the following reasons:
1002 * (1) with a large number of clients we need to limit the amount
1003 * of cached data on the clients a lot;
1005 * (2) large compute jobs generally want compute-only then io-only
1006 * and the IO should complete as quickly as possible;
1008 * (3) IO is batched up to the RPC size and is async until the
1009 * client max cache is hit
1010 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
1013 if (!PageDirty(vmpage)) {
1014 tallyop = LPROC_LL_DIRTY_MISSES;
1015 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
1017 /* page was added into cache successfully. */
1018 set_page_dirty(vmpage);
1019 vvp_write_pending(cl2ccc(obj), cp);
1020 } else if (result == -EDQUOT) {
1021 pgoff_t last_index = i_size_read(inode) >> PAGE_CACHE_SHIFT;
1022 bool need_clip = true;
1025 * Client ran out of disk space grant. Possible
1028 * (a) do a sync write, renewing grant;
1030 * (b) stop writing on this stripe, switch to the
1033 * (b) is a part of "parallel io" design that is the
1034 * ultimate goal. (a) is what "old" client did, and
1035 * what the new code continues to do for the time
1038 if (last_index > pg->cp_index) {
1039 to = PAGE_CACHE_SIZE;
1041 } else if (last_index == pg->cp_index) {
1042 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
1047 cl_page_clip(env, pg, 0, to);
1048 result = vvp_page_sync_io(env, io, pg, cp, CRT_WRITE);
1050 CERROR("Write page %lu of inode %p failed %d\n",
1051 pg->cp_index, inode, result);
1054 tallyop = LPROC_LL_DIRTY_HITS;
1057 ll_stats_ops_tally(sbi, tallyop, 1);
1059 /* Inode should be marked DIRTY even if no new page was marked DIRTY
1060 * because page could have been not flushed between 2 modifications.
1061 * It is important the file is marked DIRTY as soon as the I/O is done
1062 * Indeed, when cache is flushed, file could be already closed and it
1063 * is too late to warn the MDT.
1064 * It is acceptable that file is marked DIRTY even if I/O is dropped
1065 * for some reasons before being flushed to OST.
1068 spin_lock(&lli->lli_lock);
1069 lli->lli_flags |= LLIF_DATA_MODIFIED;
1070 spin_unlock(&lli->lli_lock);
1073 size = cl_offset(obj, pg->cp_index) + to;
1075 ll_inode_size_lock(inode);
1077 if (size > i_size_read(inode)) {
1078 cl_isize_write_nolock(inode, size);
1079 CDEBUG(D_VFSTRACE, DFID" updating i_size %lu\n",
1080 PFID(lu_object_fid(&obj->co_lu)),
1081 (unsigned long)size);
1083 cl_page_export(env, pg, 1);
1085 if (size > i_size_read(inode))
1086 cl_page_discard(env, io, pg);
1088 ll_inode_size_unlock(inode);
1092 static const struct cl_io_operations vvp_io_ops = {
1095 .cio_fini = vvp_io_read_fini,
1096 .cio_lock = vvp_io_read_lock,
1097 .cio_start = vvp_io_read_start,
1098 .cio_advance = ccc_io_advance
1101 .cio_fini = vvp_io_fini,
1102 .cio_lock = vvp_io_write_lock,
1103 .cio_start = vvp_io_write_start,
1104 .cio_advance = ccc_io_advance
1107 .cio_fini = vvp_io_setattr_fini,
1108 .cio_iter_init = vvp_io_setattr_iter_init,
1109 .cio_lock = vvp_io_setattr_lock,
1110 .cio_start = vvp_io_setattr_start,
1111 .cio_end = vvp_io_setattr_end
1114 .cio_fini = vvp_io_fault_fini,
1115 .cio_iter_init = vvp_io_fault_iter_init,
1116 .cio_lock = vvp_io_fault_lock,
1117 .cio_start = vvp_io_fault_start,
1118 .cio_end = ccc_io_end
1121 .cio_start = vvp_io_fsync_start,
1122 .cio_fini = vvp_io_fini
1125 .cio_fini = vvp_io_fini
1128 .cio_read_page = vvp_io_read_page,
1129 .cio_prepare_write = vvp_io_prepare_write,
1130 .cio_commit_write = vvp_io_commit_write
1133 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1136 struct vvp_io *vio = vvp_env_io(env);
1137 struct ccc_io *cio = ccc_env_io(env);
1138 struct inode *inode = ccc_object_inode(obj);
1141 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
1143 CDEBUG(D_VFSTRACE, DFID
1144 " ignore/verify layout %d/%d, layout version %d restore needed %d\n",
1145 PFID(lu_object_fid(&obj->co_lu)),
1146 io->ci_ignore_layout, io->ci_verify_layout,
1147 cio->cui_layout_gen, io->ci_restore_needed);
1149 CL_IO_SLICE_CLEAN(cio, cui_cl);
1150 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
1151 vio->cui_ra_window_set = 0;
1153 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1155 struct ll_inode_info *lli = ll_i2info(inode);
1157 count = io->u.ci_rw.crw_count;
1158 /* "If nbyte is 0, read() will return 0 and have no other
1159 * results." -- Single Unix Spec */
1163 cio->cui_tot_count = count;
1165 /* for read/write, we store the jobid in the inode, and
1166 * it'll be fetched by osc when building RPC.
1168 * it's not accurate if the file is shared by different
1171 lustre_get_jobid(lli->lli_jobid);
1172 } else if (io->ci_type == CIT_SETATTR) {
1173 if (!cl_io_is_trunc(io))
1174 io->ci_lockreq = CILR_MANDATORY;
1177 /* ignore layout change for generic CIT_MISC but not for glimpse.
1178 * io context for glimpse must set ci_verify_layout to true,
1179 * see cl_glimpse_size0() for details. */
1180 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1181 io->ci_ignore_layout = 1;
1183 /* Enqueue layout lock and get layout version. We need to do this
1184 * even for operations requiring to open file, such as read and write,
1185 * because it might not grant layout lock in IT_OPEN. */
1186 if (result == 0 && !io->ci_ignore_layout) {
1187 result = ll_layout_refresh(inode, &cio->cui_layout_gen);
1188 if (result == -ENOENT)
1189 /* If the inode on MDS has been removed, but the objects
1190 * on OSTs haven't been destroyed (async unlink), layout
1191 * fetch will return -ENOENT, we'd ignore this error
1192 * and continue with dirty flush. LU-3230. */
1195 CERROR("%s: refresh file layout " DFID " error %d.\n",
1196 ll_get_fsname(inode->i_sb, NULL, 0),
1197 PFID(lu_object_fid(&obj->co_lu)), result);
1203 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1204 const struct cl_io_slice *slice)
1206 /* Calling just for assertion */
1207 cl2ccc_io(env, slice);
1208 return vvp_env_io(env);