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 # error This file is kernel only.
49 #include <lustre_lite.h>
51 #include "vvp_internal.h"
53 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
54 const struct cl_io_slice *slice);
57 * True, if \a io is a normal io, False for sendfile() / splice_{read|write}
59 int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
61 struct vvp_io *vio = vvp_env_io(env);
63 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
65 return vio->cui_io_subtype == IO_NORMAL;
68 /*****************************************************************************
74 static int vvp_io_fault_iter_init(const struct lu_env *env,
75 const struct cl_io_slice *ios)
77 struct vvp_io *vio = cl2vvp_io(env, ios);
78 struct inode *inode = ccc_object_inode(ios->cis_obj);
81 cl2ccc_io(env, ios)->cui_fd->fd_file->f_dentry->d_inode);
82 vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime);
86 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
88 struct cl_io *io = ios->cis_io;
89 struct cl_object *obj = io->ci_obj;
90 struct ccc_io *cio = cl2ccc_io(env, ios);
92 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
94 CDEBUG(D_VFSTRACE, "ignore/verify layout %d/%d, layout version %d.\n",
95 io->ci_ignore_layout, io->ci_verify_layout, cio->cui_layout_gen);
97 if (!io->ci_ignore_layout && io->ci_verify_layout) {
100 /* check layout version */
101 ll_layout_refresh(ccc_object_inode(obj), &gen);
102 io->ci_need_restart = cio->cui_layout_gen != gen;
103 if (io->ci_need_restart)
104 CDEBUG(D_VFSTRACE, "layout changed from %d to %d.\n",
105 cio->cui_layout_gen, gen);
109 static void vvp_io_fault_fini(const struct lu_env *env,
110 const struct cl_io_slice *ios)
112 struct cl_io *io = ios->cis_io;
113 struct cl_page *page = io->u.ci_fault.ft_page;
115 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
118 lu_ref_del(&page->cp_reference, "fault", io);
119 cl_page_put(env, page);
120 io->u.ci_fault.ft_page = NULL;
122 vvp_io_fini(env, ios);
125 enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
128 * we only want to hold PW locks if the mmap() can generate
129 * writes back to the file and that only happens in shared
132 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
137 static int vvp_mmap_locks(const struct lu_env *env,
138 struct ccc_io *vio, struct cl_io *io)
140 struct ccc_thread_info *cti = ccc_env_info(env);
141 struct mm_struct *mm = current->mm;
142 struct vm_area_struct *vma;
143 struct cl_lock_descr *descr = &cti->cti_descr;
144 ldlm_policy_data_t policy;
151 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
153 if (!cl_is_normalio(env, io))
156 if (vio->cui_iov == NULL) /* nfs or loop back device write */
159 /* No MM (e.g. NFS)? No vmas too. */
163 for (seg = 0; seg < vio->cui_nrsegs; seg++) {
164 const struct iovec *iv = &vio->cui_iov[seg];
166 addr = (unsigned long)iv->iov_base;
171 count += addr & (~CFS_PAGE_MASK);
172 addr &= CFS_PAGE_MASK;
174 down_read(&mm->mmap_sem);
175 while((vma = our_vma(mm, addr, count)) != NULL) {
176 struct inode *inode = vma->vm_file->f_dentry->d_inode;
177 int flags = CEF_MUST;
179 if (ll_file_nolock(vma->vm_file)) {
181 * For no lock case, a lockless lock will be
188 * XXX: Required lock mode can be weakened: CIT_WRITE
189 * io only ever reads user level buffer, and CIT_READ
192 policy_from_vma(&policy, vma, addr, count);
193 descr->cld_mode = vvp_mode_from_vma(vma);
194 descr->cld_obj = ll_i2info(inode)->lli_clob;
195 descr->cld_start = cl_index(descr->cld_obj,
196 policy.l_extent.start);
197 descr->cld_end = cl_index(descr->cld_obj,
198 policy.l_extent.end);
199 descr->cld_enq_flags = flags;
200 result = cl_io_lock_alloc_add(env, io, descr);
202 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
203 descr->cld_mode, descr->cld_start,
209 if (vma->vm_end - addr >= count)
212 count -= vma->vm_end - addr;
215 up_read(&mm->mmap_sem);
220 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
221 enum cl_lock_mode mode, loff_t start, loff_t end)
223 struct ccc_io *cio = ccc_env_io(env);
227 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
230 ccc_io_update_iov(env, cio, io);
232 if (io->u.ci_rw.crw_nonblock)
233 ast_flags |= CEF_NONBLOCK;
234 result = vvp_mmap_locks(env, cio, io);
236 result = ccc_io_one_lock(env, io, ast_flags, mode, start, end);
240 static int vvp_io_read_lock(const struct lu_env *env,
241 const struct cl_io_slice *ios)
243 struct cl_io *io = ios->cis_io;
244 struct ll_inode_info *lli = ll_i2info(ccc_object_inode(io->ci_obj));
248 /* XXX: Layer violation, we shouldn't see lsm at llite level. */
249 if (lli->lli_has_smd) /* lsm-less file doesn't need to lock */
250 result = vvp_io_rw_lock(env, io, CLM_READ,
251 io->u.ci_rd.rd.crw_pos,
252 io->u.ci_rd.rd.crw_pos +
253 io->u.ci_rd.rd.crw_count - 1);
259 static int vvp_io_fault_lock(const struct lu_env *env,
260 const struct cl_io_slice *ios)
262 struct cl_io *io = ios->cis_io;
263 struct vvp_io *vio = cl2vvp_io(env, ios);
265 * XXX LDLM_FL_CBPENDING
267 return ccc_io_one_lock_index
268 (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma),
269 io->u.ci_fault.ft_index, io->u.ci_fault.ft_index);
272 static int vvp_io_write_lock(const struct lu_env *env,
273 const struct cl_io_slice *ios)
275 struct cl_io *io = ios->cis_io;
279 if (io->u.ci_wr.wr_append) {
281 end = OBD_OBJECT_EOF;
283 start = io->u.ci_wr.wr.crw_pos;
284 end = start + io->u.ci_wr.wr.crw_count - 1;
286 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
289 static int vvp_io_setattr_iter_init(const struct lu_env *env,
290 const struct cl_io_slice *ios)
296 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
298 * Handles "lockless io" mode when extent locking is done by server.
300 static int vvp_io_setattr_lock(const struct lu_env *env,
301 const struct cl_io_slice *ios)
303 struct ccc_io *cio = ccc_env_io(env);
304 struct cl_io *io = ios->cis_io;
308 if (cl_io_is_trunc(io)) {
309 new_size = io->u.ci_setattr.sa_attr.lvb_size;
311 enqflags = CEF_DISCARD_DATA;
313 if ((io->u.ci_setattr.sa_attr.lvb_mtime >=
314 io->u.ci_setattr.sa_attr.lvb_ctime) ||
315 (io->u.ci_setattr.sa_attr.lvb_atime >=
316 io->u.ci_setattr.sa_attr.lvb_ctime))
320 cio->u.setattr.cui_local_lock = SETATTR_EXTENT_LOCK;
321 return ccc_io_one_lock(env, io, enqflags, CLM_WRITE,
322 new_size, OBD_OBJECT_EOF);
325 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
329 * Only ll_inode_size_lock is taken at this level.
331 ll_inode_size_lock(inode);
332 result = vmtruncate(inode, size);
333 ll_inode_size_unlock(inode);
338 static int vvp_io_setattr_trunc(const struct lu_env *env,
339 const struct cl_io_slice *ios,
340 struct inode *inode, loff_t size)
342 inode_dio_wait(inode);
346 static int vvp_io_setattr_time(const struct lu_env *env,
347 const struct cl_io_slice *ios)
349 struct cl_io *io = ios->cis_io;
350 struct cl_object *obj = io->ci_obj;
351 struct cl_attr *attr = ccc_env_thread_attr(env);
353 unsigned valid = CAT_CTIME;
355 cl_object_attr_lock(obj);
356 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
357 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
358 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
361 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
362 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
365 result = cl_object_attr_set(env, obj, attr, valid);
366 cl_object_attr_unlock(obj);
371 static int vvp_io_setattr_start(const struct lu_env *env,
372 const struct cl_io_slice *ios)
374 struct cl_io *io = ios->cis_io;
375 struct inode *inode = ccc_object_inode(io->ci_obj);
377 mutex_lock(&inode->i_mutex);
378 if (cl_io_is_trunc(io))
379 return vvp_io_setattr_trunc(env, ios, inode,
380 io->u.ci_setattr.sa_attr.lvb_size);
382 return vvp_io_setattr_time(env, ios);
385 static void vvp_io_setattr_end(const struct lu_env *env,
386 const struct cl_io_slice *ios)
388 struct cl_io *io = ios->cis_io;
389 struct inode *inode = ccc_object_inode(io->ci_obj);
391 if (cl_io_is_trunc(io)) {
392 /* Truncate in memory pages - they must be clean pages
393 * because osc has already notified to destroy osc_extents. */
394 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
395 inode_dio_write_done(inode);
397 mutex_unlock(&inode->i_mutex);
400 static void vvp_io_setattr_fini(const struct lu_env *env,
401 const struct cl_io_slice *ios)
403 vvp_io_fini(env, ios);
406 #ifdef HAVE_FILE_READV
407 static ssize_t lustre_generic_file_read(struct file *file,
408 struct ccc_io *vio, loff_t *ppos)
410 return generic_file_readv(file, vio->cui_iov, vio->cui_nrsegs, ppos);
413 static ssize_t lustre_generic_file_write(struct file *file,
414 struct ccc_io *vio, loff_t *ppos)
416 return generic_file_writev(file, vio->cui_iov, vio->cui_nrsegs, ppos);
419 static ssize_t lustre_generic_file_read(struct file *file,
420 struct ccc_io *vio, loff_t *ppos)
422 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
423 vio->cui_nrsegs, *ppos);
426 static ssize_t lustre_generic_file_write(struct file *file,
427 struct ccc_io *vio, loff_t *ppos)
429 return generic_file_aio_write(vio->cui_iocb, vio->cui_iov,
430 vio->cui_nrsegs, *ppos);
434 static int vvp_io_read_start(const struct lu_env *env,
435 const struct cl_io_slice *ios)
437 struct vvp_io *vio = cl2vvp_io(env, ios);
438 struct ccc_io *cio = cl2ccc_io(env, ios);
439 struct cl_io *io = ios->cis_io;
440 struct cl_object *obj = io->ci_obj;
441 struct inode *inode = ccc_object_inode(obj);
442 struct ll_ra_read *bead = &vio->cui_bead;
443 struct file *file = cio->cui_fd->fd_file;
446 loff_t pos = io->u.ci_rd.rd.crw_pos;
447 long cnt = io->u.ci_rd.rd.crw_count;
448 long tot = cio->cui_tot_count;
451 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
453 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
455 result = ccc_prep_size(env, obj, io, pos, tot, &exceed);
458 else if (exceed != 0)
461 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
462 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
463 inode->i_ino, cnt, pos, i_size_read(inode));
465 /* turn off the kernel's read-ahead */
466 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
468 /* initialize read-ahead window once per syscall */
469 if (!vio->cui_ra_window_set) {
470 vio->cui_ra_window_set = 1;
471 bead->lrr_start = cl_index(obj, pos);
473 * XXX: explicit CFS_PAGE_SIZE
475 bead->lrr_count = cl_index(obj, tot + CFS_PAGE_SIZE - 1);
476 ll_ra_read_in(file, bead);
481 switch (vio->cui_io_subtype) {
483 result = lustre_generic_file_read(file, cio, &pos);
485 #ifdef HAVE_KERNEL_SENDFILE
487 result = generic_file_sendfile(file, &pos, cnt,
488 vio->u.sendfile.cui_actor,
489 vio->u.sendfile.cui_target);
492 #ifdef HAVE_KERNEL_SPLICE_READ
494 result = generic_file_splice_read(file, &pos,
495 vio->u.splice.cui_pipe, cnt,
496 vio->u.splice.cui_flags);
497 /* LU-1109: do splice read stripe by stripe otherwise if it
498 * may make nfsd stuck if this read occupied all internal pipe
504 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
512 io->ci_nob += result;
513 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
514 cio->cui_fd, pos, result, 0);
520 static void vvp_io_read_fini(const struct lu_env *env, const struct cl_io_slice *ios)
522 struct vvp_io *vio = cl2vvp_io(env, ios);
523 struct ccc_io *cio = cl2ccc_io(env, ios);
525 if (vio->cui_ra_window_set)
526 ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
528 vvp_io_fini(env, ios);
531 static int vvp_io_write_start(const struct lu_env *env,
532 const struct cl_io_slice *ios)
534 struct ccc_io *cio = cl2ccc_io(env, ios);
535 struct cl_io *io = ios->cis_io;
536 struct cl_object *obj = io->ci_obj;
537 struct inode *inode = ccc_object_inode(obj);
538 struct file *file = cio->cui_fd->fd_file;
540 loff_t pos = io->u.ci_wr.wr.crw_pos;
541 size_t cnt = io->u.ci_wr.wr.crw_count;
545 if (cl_io_is_append(io)) {
547 * PARALLEL IO This has to be changed for parallel IO doing
548 * out-of-order writes.
550 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
551 #ifndef HAVE_FILE_WRITEV
552 cio->cui_iocb->ki_pos = pos;
556 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
558 if (cio->cui_iov == NULL) /* from a temp io in ll_cl_init(). */
561 result = lustre_generic_file_write(file, cio, &pos);
566 io->ci_nob += result;
567 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
568 cio->cui_fd, pos, result, 0);
574 #ifndef HAVE_VM_OP_FAULT
575 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
579 vmpage = filemap_nopage(cfio->ft_vma, cfio->nopage.ft_address,
580 cfio->nopage.ft_type);
582 if (vmpage == NOPAGE_SIGBUS) {
583 CDEBUG(D_PAGE, "got addr %lu type %lx - SIGBUS\n",
584 cfio->nopage.ft_address,(long)cfio->nopage.ft_type);
586 } else if (vmpage == NOPAGE_OOM) {
587 CDEBUG(D_PAGE, "got addr %lu type %lx - OOM\n",
588 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
592 LL_CDEBUG_PAGE(D_PAGE, vmpage, "got addr %lu type %lx\n",
593 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
595 cfio->ft_vmpage = vmpage;
601 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
603 struct vm_fault *vmf = cfio->fault.ft_vmf;
605 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, vmf);
608 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
609 vmf->virtual_address);
610 if (unlikely(!(cfio->fault.ft_flags & VM_FAULT_LOCKED))) {
611 lock_page(vmf->page);
612 cfio->fault.ft_flags &= VM_FAULT_LOCKED;
615 cfio->ft_vmpage = vmf->page;
619 if (cfio->fault.ft_flags & VM_FAULT_SIGBUS) {
620 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
624 if (cfio->fault.ft_flags & VM_FAULT_OOM) {
625 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
629 if (cfio->fault.ft_flags & VM_FAULT_RETRY)
632 CERROR("unknow error in page fault %d!\n", cfio->fault.ft_flags);
638 static int vvp_io_fault_start(const struct lu_env *env,
639 const struct cl_io_slice *ios)
641 struct vvp_io *vio = cl2vvp_io(env, ios);
642 struct cl_io *io = ios->cis_io;
643 struct cl_object *obj = io->ci_obj;
644 struct inode *inode = ccc_object_inode(obj);
645 struct cl_fault_io *fio = &io->u.ci_fault;
646 struct vvp_fault_io *cfio = &vio->u.fault;
649 cfs_page_t *vmpage = NULL;
650 struct cl_page *page;
652 pgoff_t last; /* last page in a file data region */
654 if (fio->ft_executable &&
655 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
657 " changed while waiting for the page fault lock\n",
658 PFID(lu_object_fid(&obj->co_lu)));
660 /* offset of the last byte on the page */
661 offset = cl_offset(obj, fio->ft_index + 1) - 1;
662 LASSERT(cl_index(obj, offset) == fio->ft_index);
663 result = ccc_prep_size(env, obj, io, 0, offset + 1, NULL);
667 /* must return locked page */
668 if (fio->ft_mkwrite) {
669 LASSERT(cfio->ft_vmpage != NULL);
670 lock_page(cfio->ft_vmpage);
672 result = vvp_io_kernel_fault(cfio);
677 vmpage = cfio->ft_vmpage;
678 LASSERT(PageLocked(vmpage));
680 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
681 ll_invalidate_page(vmpage);
683 size = i_size_read(inode);
684 /* Though we have already held a cl_lock upon this page, but
685 * it still can be truncated locally. */
686 if (unlikely((vmpage->mapping != inode->i_mapping) ||
687 (page_offset(vmpage) > size))) {
688 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
690 /* return +1 to stop cl_io_loop() and ll_fault() will catch
692 GOTO(out, result = +1);
696 if (fio->ft_mkwrite ) {
699 * Capture the size while holding the lli_trunc_sem from above
700 * we want to make sure that we complete the mkwrite action
701 * while holding this lock. We need to make sure that we are
702 * not past the end of the file.
704 last_index = cl_index(obj, size - 1);
705 if (last_index < fio->ft_index) {
707 "llite: mkwrite and truncate race happened: "
709 vmpage->mapping,fio->ft_index,last_index);
711 * We need to return if we are
712 * passed the end of the file. This will propagate
713 * up the call stack to ll_page_mkwrite where
714 * we will return VM_FAULT_NOPAGE. Any non-negative
715 * value returned here will be silently
716 * converted to 0. If the vmpage->mapping is null
717 * the error code would be converted back to ENODATA
718 * in ll_page_mkwrite0. Thus we return -ENODATA
719 * to handle both cases
721 GOTO(out, result = -ENODATA);
725 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
727 GOTO(out, result = PTR_ERR(page));
729 /* if page is going to be written, we should add this page into cache
731 if (fio->ft_mkwrite) {
732 wait_on_page_writeback(vmpage);
733 if (set_page_dirty(vmpage)) {
736 /* vvp_page_assume() calls wait_on_page_writeback(). */
737 cl_page_assume(env, io, page);
739 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
740 vvp_write_pending(cl2ccc(obj), cp);
742 /* Do not set Dirty bit here so that in case IO is
743 * started before the page is really made dirty, we
744 * still have chance to detect it. */
745 result = cl_page_cache_add(env, io, page, CRT_WRITE);
746 LASSERT(cl_page_is_owned(page, io));
750 cl_page_unmap(env, io, page);
751 cl_page_discard(env, io, page);
752 cl_page_disown(env, io, page);
754 cl_page_put(env, page);
756 /* we're in big trouble, what can we do now? */
757 if (result == -EDQUOT)
761 cl_page_disown(env, io, page);
765 last = cl_index(obj, size - 1);
767 * The ft_index is only used in the case of
768 * a mkwrite action. We need to check
769 * our assertions are correct, since
770 * we should have caught this above
772 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last);
773 if (fio->ft_index == last)
775 * Last page is mapped partially.
777 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
779 fio->ft_nob = cl_page_size(obj);
781 lu_ref_add(&page->cp_reference, "fault", io);
786 /* return unlocked vmpage to avoid deadlocking */
789 #ifdef HAVE_VM_OP_FAULT
790 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
795 static int vvp_io_fsync_start(const struct lu_env *env,
796 const struct cl_io_slice *ios)
798 /* we should mark TOWRITE bit to each dirty page in radix tree to
799 * verify pages have been written, but this is difficult because of
804 static int vvp_io_read_page(const struct lu_env *env,
805 const struct cl_io_slice *ios,
806 const struct cl_page_slice *slice)
808 struct cl_io *io = ios->cis_io;
809 struct cl_object *obj = slice->cpl_obj;
810 struct ccc_page *cp = cl2ccc_page(slice);
811 struct cl_page *page = slice->cpl_page;
812 struct inode *inode = ccc_object_inode(obj);
813 struct ll_sb_info *sbi = ll_i2sbi(inode);
814 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
815 struct ll_readahead_state *ras = &fd->fd_ras;
816 cfs_page_t *vmpage = cp->cpg_page;
817 struct cl_2queue *queue = &io->ci_queue;
820 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
821 LASSERT(slice->cpl_obj == obj);
825 if (sbi->ll_ra_info.ra_max_pages_per_file &&
826 sbi->ll_ra_info.ra_max_pages)
827 ras_update(sbi, inode, ras, page->cp_index,
828 cp->cpg_defer_uptodate);
830 /* Sanity check whether the page is protected by a lock. */
831 rc = cl_page_is_under_lock(env, io, page);
833 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %d\n",
834 rc == -ENODATA ? "without a lock" :
840 if (cp->cpg_defer_uptodate) {
842 cl_page_export(env, page, 1);
845 * Add page into the queue even when it is marked uptodate above.
846 * this will unlock it automatically as part of cl_page_list_disown().
848 cl_2queue_add(queue, page);
849 if (sbi->ll_ra_info.ra_max_pages_per_file &&
850 sbi->ll_ra_info.ra_max_pages)
851 ll_readahead(env, io, ras,
852 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
857 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
858 struct cl_page *page, struct ccc_page *cp,
859 enum cl_req_type crt)
861 struct cl_2queue *queue;
864 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
866 queue = &io->ci_queue;
867 cl_2queue_init_page(queue, page);
869 result = cl_io_submit_sync(env, io, crt, queue, 0);
870 LASSERT(cl_page_is_owned(page, io));
874 * in CRT_WRITE case page is left locked even in case of
877 cl_page_list_disown(env, io, &queue->c2_qin);
878 cl_2queue_fini(env, queue);
884 * Prepare partially written-to page for a write.
886 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
887 struct cl_object *obj, struct cl_page *pg,
889 unsigned from, unsigned to)
891 struct cl_attr *attr = ccc_env_thread_attr(env);
892 loff_t offset = cl_offset(obj, pg->cp_index);
895 cl_object_attr_lock(obj);
896 result = cl_object_attr_get(env, obj, attr);
897 cl_object_attr_unlock(obj);
900 * If are writing to a new page, no need to read old data.
901 * The extent locking will have updated the KMS, and for our
902 * purposes here we can treat it like i_size.
904 if (attr->cat_kms <= offset) {
905 char *kaddr = ll_kmap_atomic(cp->cpg_page, KM_USER0);
907 memset(kaddr, 0, cl_page_size(obj));
908 ll_kunmap_atomic(kaddr, KM_USER0);
909 } else if (cp->cpg_defer_uptodate)
912 result = vvp_page_sync_io(env, io, pg, cp, CRT_READ);
914 * In older implementations, obdo_refresh_inode is called here
915 * to update the inode because the write might modify the
916 * object info at OST. However, this has been proven useless,
917 * since LVB functions will be called when user space program
918 * tries to retrieve inode attribute. Also, see bug 15909 for
922 cl_page_export(env, pg, 1);
927 static int vvp_io_prepare_write(const struct lu_env *env,
928 const struct cl_io_slice *ios,
929 const struct cl_page_slice *slice,
930 unsigned from, unsigned to)
932 struct cl_object *obj = slice->cpl_obj;
933 struct ccc_page *cp = cl2ccc_page(slice);
934 struct cl_page *pg = slice->cpl_page;
935 cfs_page_t *vmpage = cp->cpg_page;
941 LINVRNT(cl_page_is_vmlocked(env, pg));
942 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
946 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
947 if (!PageUptodate(vmpage)) {
949 * We're completely overwriting an existing page, so _don't_
950 * set it up to date until commit_write
952 if (from == 0 && to == CFS_PAGE_SIZE) {
953 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
954 POISON_PAGE(page, 0x11);
956 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
959 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
963 static int vvp_io_commit_write(const struct lu_env *env,
964 const struct cl_io_slice *ios,
965 const struct cl_page_slice *slice,
966 unsigned from, unsigned to)
968 struct cl_object *obj = slice->cpl_obj;
969 struct cl_io *io = ios->cis_io;
970 struct ccc_page *cp = cl2ccc_page(slice);
971 struct cl_page *pg = slice->cpl_page;
972 struct inode *inode = ccc_object_inode(obj);
973 struct ll_sb_info *sbi = ll_i2sbi(inode);
974 cfs_page_t *vmpage = cp->cpg_page;
982 LINVRNT(cl_page_is_vmlocked(env, pg));
983 LASSERT(vmpage->mapping->host == inode);
985 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n");
986 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
989 * queue a write for some time in the future the first time we
992 * This is different from what other file systems do: they usually
993 * just mark page (and some of its buffers) dirty and rely on
994 * balance_dirty_pages() to start a write-back. Lustre wants write-back
995 * to be started earlier for the following reasons:
997 * (1) with a large number of clients we need to limit the amount
998 * of cached data on the clients a lot;
1000 * (2) large compute jobs generally want compute-only then io-only
1001 * and the IO should complete as quickly as possible;
1003 * (3) IO is batched up to the RPC size and is async until the
1004 * client max cache is hit
1005 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
1008 if (!PageDirty(vmpage)) {
1009 tallyop = LPROC_LL_DIRTY_MISSES;
1010 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
1012 /* page was added into cache successfully. */
1013 set_page_dirty(vmpage);
1014 vvp_write_pending(cl2ccc(obj), cp);
1015 } else if (result == -EDQUOT) {
1016 pgoff_t last_index = i_size_read(inode) >> CFS_PAGE_SHIFT;
1017 bool need_clip = true;
1020 * Client ran out of disk space grant. Possible
1023 * (a) do a sync write, renewing grant;
1025 * (b) stop writing on this stripe, switch to the
1028 * (b) is a part of "parallel io" design that is the
1029 * ultimate goal. (a) is what "old" client did, and
1030 * what the new code continues to do for the time
1033 if (last_index > pg->cp_index) {
1036 } else if (last_index == pg->cp_index) {
1037 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
1042 cl_page_clip(env, pg, 0, to);
1043 result = vvp_page_sync_io(env, io, pg, cp, CRT_WRITE);
1045 CERROR("Write page %lu of inode %p failed %d\n",
1046 pg->cp_index, inode, result);
1049 tallyop = LPROC_LL_DIRTY_HITS;
1052 ll_stats_ops_tally(sbi, tallyop, 1);
1054 size = cl_offset(obj, pg->cp_index) + to;
1056 ll_inode_size_lock(inode);
1058 if (size > i_size_read(inode)) {
1059 cl_isize_write_nolock(inode, size);
1060 CDEBUG(D_VFSTRACE, DFID" updating i_size %lu\n",
1061 PFID(lu_object_fid(&obj->co_lu)),
1062 (unsigned long)size);
1064 cl_page_export(env, pg, 1);
1066 if (size > i_size_read(inode))
1067 cl_page_discard(env, io, pg);
1069 ll_inode_size_unlock(inode);
1073 static const struct cl_io_operations vvp_io_ops = {
1076 .cio_fini = vvp_io_read_fini,
1077 .cio_lock = vvp_io_read_lock,
1078 .cio_start = vvp_io_read_start,
1079 .cio_advance = ccc_io_advance
1082 .cio_fini = vvp_io_fini,
1083 .cio_lock = vvp_io_write_lock,
1084 .cio_start = vvp_io_write_start,
1085 .cio_advance = ccc_io_advance
1088 .cio_fini = vvp_io_setattr_fini,
1089 .cio_iter_init = vvp_io_setattr_iter_init,
1090 .cio_lock = vvp_io_setattr_lock,
1091 .cio_start = vvp_io_setattr_start,
1092 .cio_end = vvp_io_setattr_end
1095 .cio_fini = vvp_io_fault_fini,
1096 .cio_iter_init = vvp_io_fault_iter_init,
1097 .cio_lock = vvp_io_fault_lock,
1098 .cio_start = vvp_io_fault_start,
1099 .cio_end = ccc_io_end
1102 .cio_start = vvp_io_fsync_start,
1103 .cio_fini = vvp_io_fini
1106 .cio_fini = vvp_io_fini
1109 .cio_read_page = vvp_io_read_page,
1110 .cio_prepare_write = vvp_io_prepare_write,
1111 .cio_commit_write = vvp_io_commit_write
1114 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1117 struct vvp_io *vio = vvp_env_io(env);
1118 struct ccc_io *cio = ccc_env_io(env);
1119 struct inode *inode = ccc_object_inode(obj);
1122 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
1125 CL_IO_SLICE_CLEAN(cio, cui_cl);
1126 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
1127 vio->cui_ra_window_set = 0;
1129 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1131 struct ll_inode_info *lli = ll_i2info(inode);
1133 count = io->u.ci_rw.crw_count;
1134 /* "If nbyte is 0, read() will return 0 and have no other
1135 * results." -- Single Unix Spec */
1139 cio->cui_tot_count = count;
1140 cio->cui_tot_nrsegs = 0;
1142 /* for read/write, we store the jobid in the inode, and
1143 * it'll be fetched by osc when building RPC.
1145 * it's not accurate if the file is shared by different
1148 lustre_get_jobid(lli->lli_jobid);
1149 } else if (io->ci_type == CIT_SETATTR) {
1150 if (!cl_io_is_trunc(io))
1151 io->ci_lockreq = CILR_MANDATORY;
1154 /* ignore layout change for generic CIT_MISC but not for glimpse.
1155 * io context for glimpse must set ci_verify_layout to true,
1156 * see cl_glimpse_size0() for details. */
1157 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1158 io->ci_ignore_layout = 1;
1160 /* Enqueue layout lock and get layout version. We need to do this
1161 * even for operations requiring to open file, such as read and write,
1162 * because it might not grant layout lock in IT_OPEN. */
1163 if (result == 0 && !io->ci_ignore_layout)
1164 result = ll_layout_refresh(inode, &cio->cui_layout_gen);
1169 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1170 const struct cl_io_slice *slice)
1172 /* Caling just for assertion */
1173 cl2ccc_io(env, slice);
1174 return vvp_env_io(env);