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, 2013, 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;
69 * For swapping layout. The file's layout may have changed.
70 * To avoid populating pages to a wrong stripe, we have to verify the
71 * correctness of layout. It works because swapping layout processes
72 * have to acquire group lock.
74 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
77 struct ll_inode_info *lli = ll_i2info(inode);
78 struct ccc_io *cio = ccc_env_io(env);
81 switch (io->ci_type) {
84 /* don't need lock here to check lli_layout_gen as we have held
85 * extent lock and GROUP lock has to hold to swap layout */
86 if (ll_layout_version_get(lli) != cio->cui_layout_gen) {
87 io->ci_need_restart = 1;
88 /* this will return application a short read/write */
93 /* fault is okay because we've already had a page. */
101 /*****************************************************************************
107 static int vvp_io_fault_iter_init(const struct lu_env *env,
108 const struct cl_io_slice *ios)
110 struct vvp_io *vio = cl2vvp_io(env, ios);
111 struct inode *inode = ccc_object_inode(ios->cis_obj);
114 cl2ccc_io(env, ios)->cui_fd->fd_file->f_dentry->d_inode);
115 vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime);
119 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
121 struct cl_io *io = ios->cis_io;
122 struct cl_object *obj = io->ci_obj;
123 struct ccc_io *cio = cl2ccc_io(env, ios);
124 struct inode *inode = ccc_object_inode(obj);
126 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
128 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
129 "restore needed %d\n",
130 PFID(lu_object_fid(&obj->co_lu)),
131 io->ci_ignore_layout, io->ci_verify_layout,
132 cio->cui_layout_gen, io->ci_restore_needed);
134 if (io->ci_restore_needed == 1) {
137 /* file was detected release, we need to restore it
138 * before finishing the io
140 rc = ll_layout_restore(inode, 0, OBD_OBJECT_EOF);
141 /* if restore registration failed, no restart,
142 * we will return -ENODATA */
143 /* The layout will change after restore, so we need to
144 * block on layout lock hold by the MDT
145 * as MDT will not send new layout in lvb (see LU-3124)
146 * we have to explicitly fetch it, all this will be done
147 * by ll_layout_refresh()
150 io->ci_restore_needed = 0;
151 io->ci_need_restart = 1;
152 io->ci_verify_layout = 1;
154 io->ci_restore_needed = 1;
155 io->ci_need_restart = 0;
156 io->ci_verify_layout = 0;
161 if (!io->ci_ignore_layout && io->ci_verify_layout) {
164 /* check layout version */
165 ll_layout_refresh(inode, &gen);
166 io->ci_need_restart = cio->cui_layout_gen != gen;
167 if (io->ci_need_restart) {
169 DFID" layout changed from %d to %d.\n",
170 PFID(lu_object_fid(&obj->co_lu)),
171 cio->cui_layout_gen, gen);
172 /* today successful restore is the only possible
174 /* restore was done, clear restoring state */
175 ll_i2info(ccc_object_inode(obj))->lli_flags &=
176 ~LLIF_FILE_RESTORING;
181 static void vvp_io_fault_fini(const struct lu_env *env,
182 const struct cl_io_slice *ios)
184 struct cl_io *io = ios->cis_io;
185 struct cl_page *page = io->u.ci_fault.ft_page;
187 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
190 lu_ref_del(&page->cp_reference, "fault", io);
191 cl_page_put(env, page);
192 io->u.ci_fault.ft_page = NULL;
194 vvp_io_fini(env, ios);
197 enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
200 * we only want to hold PW locks if the mmap() can generate
201 * writes back to the file and that only happens in shared
204 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
209 static int vvp_mmap_locks(const struct lu_env *env,
210 struct ccc_io *vio, struct cl_io *io)
212 struct ccc_thread_info *cti = ccc_env_info(env);
213 struct mm_struct *mm = current->mm;
214 struct vm_area_struct *vma;
215 struct cl_lock_descr *descr = &cti->cti_descr;
216 ldlm_policy_data_t policy;
223 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
225 if (!cl_is_normalio(env, io))
228 if (vio->cui_iov == NULL) /* nfs or loop back device write */
231 /* No MM (e.g. NFS)? No vmas too. */
235 for (seg = 0; seg < vio->cui_nrsegs; seg++) {
236 const struct iovec *iv = &vio->cui_iov[seg];
238 addr = (unsigned long)iv->iov_base;
243 count += addr & (~CFS_PAGE_MASK);
244 addr &= CFS_PAGE_MASK;
246 down_read(&mm->mmap_sem);
247 while((vma = our_vma(mm, addr, count)) != NULL) {
248 struct inode *inode = vma->vm_file->f_dentry->d_inode;
249 int flags = CEF_MUST;
251 if (ll_file_nolock(vma->vm_file)) {
253 * For no lock case, a lockless lock will be
260 * XXX: Required lock mode can be weakened: CIT_WRITE
261 * io only ever reads user level buffer, and CIT_READ
264 policy_from_vma(&policy, vma, addr, count);
265 descr->cld_mode = vvp_mode_from_vma(vma);
266 descr->cld_obj = ll_i2info(inode)->lli_clob;
267 descr->cld_start = cl_index(descr->cld_obj,
268 policy.l_extent.start);
269 descr->cld_end = cl_index(descr->cld_obj,
270 policy.l_extent.end);
271 descr->cld_enq_flags = flags;
272 result = cl_io_lock_alloc_add(env, io, descr);
274 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
275 descr->cld_mode, descr->cld_start,
279 up_read(&mm->mmap_sem);
283 if (vma->vm_end - addr >= count)
286 count -= vma->vm_end - addr;
289 up_read(&mm->mmap_sem);
294 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
295 enum cl_lock_mode mode, loff_t start, loff_t end)
297 struct ccc_io *cio = ccc_env_io(env);
301 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
304 ccc_io_update_iov(env, cio, io);
306 if (io->u.ci_rw.crw_nonblock)
307 ast_flags |= CEF_NONBLOCK;
308 result = vvp_mmap_locks(env, cio, io);
310 result = ccc_io_one_lock(env, io, ast_flags, mode, start, end);
314 static int vvp_io_read_lock(const struct lu_env *env,
315 const struct cl_io_slice *ios)
317 struct cl_io *io = ios->cis_io;
318 struct ll_inode_info *lli = ll_i2info(ccc_object_inode(io->ci_obj));
322 /* XXX: Layer violation, we shouldn't see lsm at llite level. */
323 if (lli->lli_has_smd) /* lsm-less file doesn't need to lock */
324 result = vvp_io_rw_lock(env, io, CLM_READ,
325 io->u.ci_rd.rd.crw_pos,
326 io->u.ci_rd.rd.crw_pos +
327 io->u.ci_rd.rd.crw_count - 1);
333 static int vvp_io_fault_lock(const struct lu_env *env,
334 const struct cl_io_slice *ios)
336 struct cl_io *io = ios->cis_io;
337 struct vvp_io *vio = cl2vvp_io(env, ios);
339 * XXX LDLM_FL_CBPENDING
341 return ccc_io_one_lock_index
342 (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma),
343 io->u.ci_fault.ft_index, io->u.ci_fault.ft_index);
346 static int vvp_io_write_lock(const struct lu_env *env,
347 const struct cl_io_slice *ios)
349 struct cl_io *io = ios->cis_io;
353 if (io->u.ci_wr.wr_append) {
355 end = OBD_OBJECT_EOF;
357 start = io->u.ci_wr.wr.crw_pos;
358 end = start + io->u.ci_wr.wr.crw_count - 1;
360 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
363 static int vvp_io_setattr_iter_init(const struct lu_env *env,
364 const struct cl_io_slice *ios)
370 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
372 * Handles "lockless io" mode when extent locking is done by server.
374 static int vvp_io_setattr_lock(const struct lu_env *env,
375 const struct cl_io_slice *ios)
377 struct ccc_io *cio = ccc_env_io(env);
378 struct cl_io *io = ios->cis_io;
382 if (cl_io_is_trunc(io)) {
383 new_size = io->u.ci_setattr.sa_attr.lvb_size;
385 enqflags = CEF_DISCARD_DATA;
387 if ((io->u.ci_setattr.sa_attr.lvb_mtime >=
388 io->u.ci_setattr.sa_attr.lvb_ctime) ||
389 (io->u.ci_setattr.sa_attr.lvb_atime >=
390 io->u.ci_setattr.sa_attr.lvb_ctime))
394 cio->u.setattr.cui_local_lock = SETATTR_EXTENT_LOCK;
395 return ccc_io_one_lock(env, io, enqflags, CLM_WRITE,
396 new_size, OBD_OBJECT_EOF);
399 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
405 * Only ll_inode_size_lock is taken at this level.
407 ll_inode_size_lock(inode);
408 result = inode_newsize_ok(inode, size);
410 ll_inode_size_unlock(inode);
413 oldsize = inode->i_size;
414 i_size_write(inode, size);
416 truncate_pagecache(inode, oldsize, size);
417 ll_inode_size_unlock(inode);
421 static int vvp_io_setattr_trunc(const struct lu_env *env,
422 const struct cl_io_slice *ios,
423 struct inode *inode, loff_t size)
425 inode_dio_wait(inode);
429 static int vvp_io_setattr_time(const struct lu_env *env,
430 const struct cl_io_slice *ios)
432 struct cl_io *io = ios->cis_io;
433 struct cl_object *obj = io->ci_obj;
434 struct cl_attr *attr = ccc_env_thread_attr(env);
436 unsigned valid = CAT_CTIME;
438 cl_object_attr_lock(obj);
439 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
440 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
441 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
444 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
445 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
448 result = cl_object_attr_set(env, obj, attr, valid);
449 cl_object_attr_unlock(obj);
454 static int vvp_io_setattr_start(const struct lu_env *env,
455 const struct cl_io_slice *ios)
457 struct cl_io *io = ios->cis_io;
458 struct inode *inode = ccc_object_inode(io->ci_obj);
461 mutex_lock(&inode->i_mutex);
462 if (cl_io_is_trunc(io))
463 result = vvp_io_setattr_trunc(env, ios, inode,
464 io->u.ci_setattr.sa_attr.lvb_size);
466 result = vvp_io_setattr_time(env, ios);
470 static void vvp_io_setattr_end(const struct lu_env *env,
471 const struct cl_io_slice *ios)
473 struct cl_io *io = ios->cis_io;
474 struct inode *inode = ccc_object_inode(io->ci_obj);
476 if (cl_io_is_trunc(io)) {
477 /* Truncate in memory pages - they must be clean pages
478 * because osc has already notified to destroy osc_extents. */
479 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
480 inode_dio_write_done(inode);
482 mutex_unlock(&inode->i_mutex);
485 static void vvp_io_setattr_fini(const struct lu_env *env,
486 const struct cl_io_slice *ios)
488 vvp_io_fini(env, ios);
491 static ssize_t lustre_generic_file_read(struct file *file,
492 struct ccc_io *vio, loff_t *ppos)
494 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
495 vio->cui_nrsegs, *ppos);
498 static ssize_t lustre_generic_file_write(struct file *file,
499 struct ccc_io *vio, loff_t *ppos)
501 return generic_file_aio_write(vio->cui_iocb, vio->cui_iov,
502 vio->cui_nrsegs, *ppos);
505 static int vvp_io_read_start(const struct lu_env *env,
506 const struct cl_io_slice *ios)
508 struct vvp_io *vio = cl2vvp_io(env, ios);
509 struct ccc_io *cio = cl2ccc_io(env, ios);
510 struct cl_io *io = ios->cis_io;
511 struct cl_object *obj = io->ci_obj;
512 struct inode *inode = ccc_object_inode(obj);
513 struct ll_ra_read *bead = &vio->cui_bead;
514 struct file *file = cio->cui_fd->fd_file;
517 loff_t pos = io->u.ci_rd.rd.crw_pos;
518 long cnt = io->u.ci_rd.rd.crw_count;
519 long tot = cio->cui_tot_count;
522 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
524 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
526 if (!can_populate_pages(env, io, inode))
529 result = ccc_prep_size(env, obj, io, pos, tot, &exceed);
532 else if (exceed != 0)
535 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
536 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
537 inode->i_ino, cnt, pos, i_size_read(inode));
539 /* turn off the kernel's read-ahead */
540 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
542 /* initialize read-ahead window once per syscall */
543 if (!vio->cui_ra_window_set) {
544 vio->cui_ra_window_set = 1;
545 bead->lrr_start = cl_index(obj, pos);
546 bead->lrr_count = cl_index(obj, tot + PAGE_CACHE_SIZE - 1);
547 ll_ra_read_in(file, bead);
552 switch (vio->cui_io_subtype) {
554 result = lustre_generic_file_read(file, cio, &pos);
557 result = generic_file_splice_read(file, &pos,
558 vio->u.splice.cui_pipe, cnt,
559 vio->u.splice.cui_flags);
560 /* LU-1109: do splice read stripe by stripe otherwise if it
561 * may make nfsd stuck if this read occupied all internal pipe
566 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
574 io->ci_nob += result;
575 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, cio->cui_fd,
583 static void vvp_io_read_fini(const struct lu_env *env, const struct cl_io_slice *ios)
585 struct vvp_io *vio = cl2vvp_io(env, ios);
586 struct ccc_io *cio = cl2ccc_io(env, ios);
588 if (vio->cui_ra_window_set)
589 ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
591 vvp_io_fini(env, ios);
594 static int vvp_io_write_start(const struct lu_env *env,
595 const struct cl_io_slice *ios)
597 struct ccc_io *cio = cl2ccc_io(env, ios);
598 struct cl_io *io = ios->cis_io;
599 struct cl_object *obj = io->ci_obj;
600 struct inode *inode = ccc_object_inode(obj);
601 struct file *file = cio->cui_fd->fd_file;
603 loff_t pos = io->u.ci_wr.wr.crw_pos;
604 size_t cnt = io->u.ci_wr.wr.crw_count;
608 if (!can_populate_pages(env, io, inode))
611 if (cl_io_is_append(io)) {
613 * PARALLEL IO This has to be changed for parallel IO doing
614 * out-of-order writes.
616 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
617 cio->cui_iocb->ki_pos = pos;
620 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
622 if (cio->cui_iov == NULL) /* from a temp io in ll_cl_init(). */
625 result = lustre_generic_file_write(file, cio, &pos);
630 io->ci_nob += result;
631 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
632 cio->cui_fd, pos, result, WRITE);
639 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
641 struct vm_fault *vmf = cfio->fault.ft_vmf;
643 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, vmf);
646 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
647 vmf->virtual_address);
648 if (unlikely(!(cfio->fault.ft_flags & VM_FAULT_LOCKED))) {
649 lock_page(vmf->page);
650 cfio->fault.ft_flags |= VM_FAULT_LOCKED;
653 cfio->ft_vmpage = vmf->page;
657 if (cfio->fault.ft_flags & VM_FAULT_SIGBUS) {
658 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
662 if (cfio->fault.ft_flags & VM_FAULT_OOM) {
663 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
667 if (cfio->fault.ft_flags & VM_FAULT_RETRY)
670 CERROR("unknow error in page fault %d!\n", cfio->fault.ft_flags);
674 static int vvp_io_fault_start(const struct lu_env *env,
675 const struct cl_io_slice *ios)
677 struct vvp_io *vio = cl2vvp_io(env, ios);
678 struct cl_io *io = ios->cis_io;
679 struct cl_object *obj = io->ci_obj;
680 struct inode *inode = ccc_object_inode(obj);
681 struct cl_fault_io *fio = &io->u.ci_fault;
682 struct vvp_fault_io *cfio = &vio->u.fault;
685 struct page *vmpage = NULL;
686 struct cl_page *page;
688 pgoff_t last; /* last page in a file data region */
690 if (fio->ft_executable &&
691 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
693 " changed while waiting for the page fault lock\n",
694 PFID(lu_object_fid(&obj->co_lu)));
696 /* offset of the last byte on the page */
697 offset = cl_offset(obj, fio->ft_index + 1) - 1;
698 LASSERT(cl_index(obj, offset) == fio->ft_index);
699 result = ccc_prep_size(env, obj, io, 0, offset + 1, NULL);
703 /* must return locked page */
704 if (fio->ft_mkwrite) {
705 LASSERT(cfio->ft_vmpage != NULL);
706 lock_page(cfio->ft_vmpage);
708 result = vvp_io_kernel_fault(cfio);
713 vmpage = cfio->ft_vmpage;
714 LASSERT(PageLocked(vmpage));
716 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
717 ll_invalidate_page(vmpage);
719 size = i_size_read(inode);
720 /* Though we have already held a cl_lock upon this page, but
721 * it still can be truncated locally. */
722 if (unlikely((vmpage->mapping != inode->i_mapping) ||
723 (page_offset(vmpage) > size))) {
724 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
726 /* return +1 to stop cl_io_loop() and ll_fault() will catch
728 GOTO(out, result = +1);
732 if (fio->ft_mkwrite ) {
735 * Capture the size while holding the lli_trunc_sem from above
736 * we want to make sure that we complete the mkwrite action
737 * while holding this lock. We need to make sure that we are
738 * not past the end of the file.
740 last_index = cl_index(obj, size - 1);
741 if (last_index < fio->ft_index) {
743 "llite: mkwrite and truncate race happened: "
745 vmpage->mapping,fio->ft_index,last_index);
747 * We need to return if we are
748 * passed the end of the file. This will propagate
749 * up the call stack to ll_page_mkwrite where
750 * we will return VM_FAULT_NOPAGE. Any non-negative
751 * value returned here will be silently
752 * converted to 0. If the vmpage->mapping is null
753 * the error code would be converted back to ENODATA
754 * in ll_page_mkwrite0. Thus we return -ENODATA
755 * to handle both cases
757 GOTO(out, result = -ENODATA);
761 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
763 GOTO(out, result = PTR_ERR(page));
765 /* if page is going to be written, we should add this page into cache
767 if (fio->ft_mkwrite) {
768 wait_on_page_writeback(vmpage);
769 if (set_page_dirty(vmpage)) {
772 /* vvp_page_assume() calls wait_on_page_writeback(). */
773 cl_page_assume(env, io, page);
775 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
776 vvp_write_pending(cl2ccc(obj), cp);
778 /* Do not set Dirty bit here so that in case IO is
779 * started before the page is really made dirty, we
780 * still have chance to detect it. */
781 result = cl_page_cache_add(env, io, page, CRT_WRITE);
782 LASSERT(cl_page_is_owned(page, io));
786 cl_page_unmap(env, io, page);
787 cl_page_discard(env, io, page);
788 cl_page_disown(env, io, page);
790 cl_page_put(env, page);
792 /* we're in big trouble, what can we do now? */
793 if (result == -EDQUOT)
797 cl_page_disown(env, io, page);
801 last = cl_index(obj, size - 1);
803 * The ft_index is only used in the case of
804 * a mkwrite action. We need to check
805 * our assertions are correct, since
806 * we should have caught this above
808 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last);
809 if (fio->ft_index == last)
811 * Last page is mapped partially.
813 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
815 fio->ft_nob = cl_page_size(obj);
817 lu_ref_add(&page->cp_reference, "fault", io);
822 /* return unlocked vmpage to avoid deadlocking */
825 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
829 static int vvp_io_fsync_start(const struct lu_env *env,
830 const struct cl_io_slice *ios)
832 /* we should mark TOWRITE bit to each dirty page in radix tree to
833 * verify pages have been written, but this is difficult because of
838 static int vvp_io_read_page(const struct lu_env *env,
839 const struct cl_io_slice *ios,
840 const struct cl_page_slice *slice)
842 struct cl_io *io = ios->cis_io;
843 struct cl_object *obj = slice->cpl_obj;
844 struct ccc_page *cp = cl2ccc_page(slice);
845 struct cl_page *page = slice->cpl_page;
846 struct inode *inode = ccc_object_inode(obj);
847 struct ll_sb_info *sbi = ll_i2sbi(inode);
848 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
849 struct ll_readahead_state *ras = &fd->fd_ras;
850 struct page *vmpage = cp->cpg_page;
851 struct cl_2queue *queue = &io->ci_queue;
854 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
855 LASSERT(slice->cpl_obj == obj);
859 if (sbi->ll_ra_info.ra_max_pages_per_file &&
860 sbi->ll_ra_info.ra_max_pages)
861 ras_update(sbi, inode, ras, page->cp_index,
862 cp->cpg_defer_uptodate);
864 /* Sanity check whether the page is protected by a lock. */
865 rc = cl_page_is_under_lock(env, io, page);
867 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %d\n",
868 rc == -ENODATA ? "without a lock" :
874 if (cp->cpg_defer_uptodate) {
876 cl_page_export(env, page, 1);
879 * Add page into the queue even when it is marked uptodate above.
880 * this will unlock it automatically as part of cl_page_list_disown().
882 cl_2queue_add(queue, page);
883 if (sbi->ll_ra_info.ra_max_pages_per_file &&
884 sbi->ll_ra_info.ra_max_pages)
885 ll_readahead(env, io, ras,
886 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
891 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
892 struct cl_page *page, struct ccc_page *cp,
893 enum cl_req_type crt)
895 struct cl_2queue *queue;
898 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
900 queue = &io->ci_queue;
901 cl_2queue_init_page(queue, page);
903 result = cl_io_submit_sync(env, io, crt, queue, 0);
904 LASSERT(cl_page_is_owned(page, io));
908 * in CRT_WRITE case page is left locked even in case of
911 cl_page_list_disown(env, io, &queue->c2_qin);
912 cl_2queue_fini(env, queue);
918 * Prepare partially written-to page for a write.
920 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
921 struct cl_object *obj, struct cl_page *pg,
923 unsigned from, unsigned to)
925 struct cl_attr *attr = ccc_env_thread_attr(env);
926 loff_t offset = cl_offset(obj, pg->cp_index);
929 cl_object_attr_lock(obj);
930 result = cl_object_attr_get(env, obj, attr);
931 cl_object_attr_unlock(obj);
934 * If are writing to a new page, no need to read old data.
935 * The extent locking will have updated the KMS, and for our
936 * purposes here we can treat it like i_size.
938 if (attr->cat_kms <= offset) {
939 char *kaddr = ll_kmap_atomic(cp->cpg_page, KM_USER0);
941 memset(kaddr, 0, cl_page_size(obj));
942 ll_kunmap_atomic(kaddr, KM_USER0);
943 } else if (cp->cpg_defer_uptodate)
946 result = vvp_page_sync_io(env, io, pg, cp, CRT_READ);
948 * In older implementations, obdo_refresh_inode is called here
949 * to update the inode because the write might modify the
950 * object info at OST. However, this has been proven useless,
951 * since LVB functions will be called when user space program
952 * tries to retrieve inode attribute. Also, see bug 15909 for
956 cl_page_export(env, pg, 1);
961 static int vvp_io_prepare_write(const struct lu_env *env,
962 const struct cl_io_slice *ios,
963 const struct cl_page_slice *slice,
964 unsigned from, unsigned to)
966 struct cl_object *obj = slice->cpl_obj;
967 struct ccc_page *cp = cl2ccc_page(slice);
968 struct cl_page *pg = slice->cpl_page;
969 struct page *vmpage = cp->cpg_page;
975 LINVRNT(cl_page_is_vmlocked(env, pg));
976 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
980 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
981 if (!PageUptodate(vmpage)) {
983 * We're completely overwriting an existing page, so _don't_
984 * set it up to date until commit_write
986 if (from == 0 && to == PAGE_CACHE_SIZE) {
987 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
988 POISON_PAGE(page, 0x11);
990 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
993 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
997 static int vvp_io_commit_write(const struct lu_env *env,
998 const struct cl_io_slice *ios,
999 const struct cl_page_slice *slice,
1000 unsigned from, unsigned to)
1002 struct cl_object *obj = slice->cpl_obj;
1003 struct cl_io *io = ios->cis_io;
1004 struct ccc_page *cp = cl2ccc_page(slice);
1005 struct cl_page *pg = slice->cpl_page;
1006 struct inode *inode = ccc_object_inode(obj);
1007 struct ll_sb_info *sbi = ll_i2sbi(inode);
1008 struct ll_inode_info *lli = ll_i2info(inode);
1009 struct page *vmpage = cp->cpg_page;
1017 LINVRNT(cl_page_is_vmlocked(env, pg));
1018 LASSERT(vmpage->mapping->host == inode);
1020 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n");
1021 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
1024 * queue a write for some time in the future the first time we
1027 * This is different from what other file systems do: they usually
1028 * just mark page (and some of its buffers) dirty and rely on
1029 * balance_dirty_pages() to start a write-back. Lustre wants write-back
1030 * to be started earlier for the following reasons:
1032 * (1) with a large number of clients we need to limit the amount
1033 * of cached data on the clients a lot;
1035 * (2) large compute jobs generally want compute-only then io-only
1036 * and the IO should complete as quickly as possible;
1038 * (3) IO is batched up to the RPC size and is async until the
1039 * client max cache is hit
1040 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
1043 if (!PageDirty(vmpage)) {
1044 tallyop = LPROC_LL_DIRTY_MISSES;
1045 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
1047 /* page was added into cache successfully. */
1048 set_page_dirty(vmpage);
1049 vvp_write_pending(cl2ccc(obj), cp);
1050 } else if (result == -EDQUOT) {
1051 pgoff_t last_index = i_size_read(inode) >> PAGE_CACHE_SHIFT;
1052 bool need_clip = true;
1055 * Client ran out of disk space grant. Possible
1058 * (a) do a sync write, renewing grant;
1060 * (b) stop writing on this stripe, switch to the
1063 * (b) is a part of "parallel io" design that is the
1064 * ultimate goal. (a) is what "old" client did, and
1065 * what the new code continues to do for the time
1068 if (last_index > pg->cp_index) {
1069 to = PAGE_CACHE_SIZE;
1071 } else if (last_index == pg->cp_index) {
1072 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
1077 cl_page_clip(env, pg, 0, to);
1078 result = vvp_page_sync_io(env, io, pg, cp, CRT_WRITE);
1080 CERROR("Write page %lu of inode %p failed %d\n",
1081 pg->cp_index, inode, result);
1084 tallyop = LPROC_LL_DIRTY_HITS;
1087 ll_stats_ops_tally(sbi, tallyop, 1);
1089 /* Inode should be marked DIRTY even if no new page was marked DIRTY
1090 * because page could have been not flushed between 2 modifications.
1091 * It is important the file is marked DIRTY as soon as the I/O is done
1092 * Indeed, when cache is flushed, file could be already closed and it
1093 * is too late to warn the MDT.
1094 * It is acceptable that file is marked DIRTY even if I/O is dropped
1095 * for some reasons before being flushed to OST.
1098 spin_lock(&lli->lli_lock);
1099 lli->lli_flags |= LLIF_DATA_MODIFIED;
1100 spin_unlock(&lli->lli_lock);
1103 size = cl_offset(obj, pg->cp_index) + to;
1105 ll_inode_size_lock(inode);
1107 if (size > i_size_read(inode)) {
1108 cl_isize_write_nolock(inode, size);
1109 CDEBUG(D_VFSTRACE, DFID" updating i_size %lu\n",
1110 PFID(lu_object_fid(&obj->co_lu)),
1111 (unsigned long)size);
1113 cl_page_export(env, pg, 1);
1115 if (size > i_size_read(inode))
1116 cl_page_discard(env, io, pg);
1118 ll_inode_size_unlock(inode);
1122 static const struct cl_io_operations vvp_io_ops = {
1125 .cio_fini = vvp_io_read_fini,
1126 .cio_lock = vvp_io_read_lock,
1127 .cio_start = vvp_io_read_start,
1128 .cio_advance = ccc_io_advance
1131 .cio_fini = vvp_io_fini,
1132 .cio_lock = vvp_io_write_lock,
1133 .cio_start = vvp_io_write_start,
1134 .cio_advance = ccc_io_advance
1137 .cio_fini = vvp_io_setattr_fini,
1138 .cio_iter_init = vvp_io_setattr_iter_init,
1139 .cio_lock = vvp_io_setattr_lock,
1140 .cio_start = vvp_io_setattr_start,
1141 .cio_end = vvp_io_setattr_end
1144 .cio_fini = vvp_io_fault_fini,
1145 .cio_iter_init = vvp_io_fault_iter_init,
1146 .cio_lock = vvp_io_fault_lock,
1147 .cio_start = vvp_io_fault_start,
1148 .cio_end = ccc_io_end
1151 .cio_start = vvp_io_fsync_start,
1152 .cio_fini = vvp_io_fini
1155 .cio_fini = vvp_io_fini
1158 .cio_read_page = vvp_io_read_page,
1159 .cio_prepare_write = vvp_io_prepare_write,
1160 .cio_commit_write = vvp_io_commit_write
1163 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1166 struct vvp_io *vio = vvp_env_io(env);
1167 struct ccc_io *cio = ccc_env_io(env);
1168 struct inode *inode = ccc_object_inode(obj);
1171 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
1174 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1175 "restore needed %d\n",
1176 PFID(lu_object_fid(&obj->co_lu)),
1177 io->ci_ignore_layout, io->ci_verify_layout,
1178 cio->cui_layout_gen, io->ci_restore_needed);
1180 CL_IO_SLICE_CLEAN(cio, cui_cl);
1181 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
1182 vio->cui_ra_window_set = 0;
1184 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1186 struct ll_inode_info *lli = ll_i2info(inode);
1188 count = io->u.ci_rw.crw_count;
1189 /* "If nbyte is 0, read() will return 0 and have no other
1190 * results." -- Single Unix Spec */
1194 cio->cui_tot_count = count;
1195 cio->cui_tot_nrsegs = 0;
1197 /* for read/write, we store the jobid in the inode, and
1198 * it'll be fetched by osc when building RPC.
1200 * it's not accurate if the file is shared by different
1203 lustre_get_jobid(lli->lli_jobid);
1204 } else if (io->ci_type == CIT_SETATTR) {
1205 if (!cl_io_is_trunc(io))
1206 io->ci_lockreq = CILR_MANDATORY;
1209 /* ignore layout change for generic CIT_MISC but not for glimpse.
1210 * io context for glimpse must set ci_verify_layout to true,
1211 * see cl_glimpse_size0() for details. */
1212 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1213 io->ci_ignore_layout = 1;
1215 /* Enqueue layout lock and get layout version. We need to do this
1216 * even for operations requiring to open file, such as read and write,
1217 * because it might not grant layout lock in IT_OPEN. */
1218 if (result == 0 && !io->ci_ignore_layout) {
1219 result = ll_layout_refresh(inode, &cio->cui_layout_gen);
1220 if (result == -ENOENT)
1221 /* If the inode on MDS has been removed, but the objects
1222 * on OSTs haven't been destroyed (async unlink), layout
1223 * fetch will return -ENOENT, we'd ingore this error
1224 * and continue with dirty flush. LU-3230. */
1227 CERROR("%s: refresh file layout " DFID " error %d.\n",
1228 ll_get_fsname(inode->i_sb, NULL, 0),
1229 PFID(lu_object_fid(&obj->co_lu)), result);
1235 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1236 const struct cl_io_slice *slice)
1238 /* Caling just for assertion */
1239 cl2ccc_io(env, slice);
1240 return vvp_env_io(env);