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 (lli->lli_layout_gen != 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);
125 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
127 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
128 "restore needed %d\n",
129 PFID(lu_object_fid(&obj->co_lu)),
130 io->ci_ignore_layout, io->ci_verify_layout,
131 cio->cui_layout_gen, io->ci_restore_needed);
133 if (io->ci_restore_needed == 1) {
136 /* file was detected release, we need to restore it
137 * before finishing the io
139 rc = ll_layout_restore(ccc_object_inode(obj));
140 /* if restore registration failed, no restart,
141 * we will return -ENODATA */
142 /* The layout will change after restore, so we need to
143 * block on layout lock hold by the MDT
144 * as MDT will not send new layout in lvb (see LU-3124)
145 * we have to explicitly fetch it, all this will be done
146 * by ll_layout_refresh()
149 io->ci_restore_needed = 0;
150 io->ci_need_restart = 1;
151 io->ci_verify_layout = 1;
153 io->ci_restore_needed = 1;
154 io->ci_need_restart = 0;
155 io->ci_verify_layout = 0;
160 if (!io->ci_ignore_layout && io->ci_verify_layout) {
163 /* check layout version */
164 ll_layout_refresh(ccc_object_inode(obj), &gen);
165 io->ci_need_restart = cio->cui_layout_gen != gen;
166 if (io->ci_need_restart) {
168 DFID" layout changed from %d to %d.\n",
169 PFID(lu_object_fid(&obj->co_lu)),
170 cio->cui_layout_gen, gen);
171 /* today successful restore is the only possible
173 /* restore was done, clear restoring state */
174 ll_i2info(ccc_object_inode(obj))->lli_flags &=
175 ~LLIF_FILE_RESTORING;
180 static void vvp_io_fault_fini(const struct lu_env *env,
181 const struct cl_io_slice *ios)
183 struct cl_io *io = ios->cis_io;
184 struct cl_page *page = io->u.ci_fault.ft_page;
186 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
189 lu_ref_del(&page->cp_reference, "fault", io);
190 cl_page_put(env, page);
191 io->u.ci_fault.ft_page = NULL;
193 vvp_io_fini(env, ios);
196 enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
199 * we only want to hold PW locks if the mmap() can generate
200 * writes back to the file and that only happens in shared
203 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
208 static int vvp_mmap_locks(const struct lu_env *env,
209 struct ccc_io *vio, struct cl_io *io)
211 struct ccc_thread_info *cti = ccc_env_info(env);
212 struct mm_struct *mm = current->mm;
213 struct vm_area_struct *vma;
214 struct cl_lock_descr *descr = &cti->cti_descr;
215 ldlm_policy_data_t policy;
222 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
224 if (!cl_is_normalio(env, io))
227 if (vio->cui_iov == NULL) /* nfs or loop back device write */
230 /* No MM (e.g. NFS)? No vmas too. */
234 for (seg = 0; seg < vio->cui_nrsegs; seg++) {
235 const struct iovec *iv = &vio->cui_iov[seg];
237 addr = (unsigned long)iv->iov_base;
242 count += addr & (~CFS_PAGE_MASK);
243 addr &= CFS_PAGE_MASK;
245 down_read(&mm->mmap_sem);
246 while((vma = our_vma(mm, addr, count)) != NULL) {
247 struct inode *inode = vma->vm_file->f_dentry->d_inode;
248 int flags = CEF_MUST;
250 if (ll_file_nolock(vma->vm_file)) {
252 * For no lock case, a lockless lock will be
259 * XXX: Required lock mode can be weakened: CIT_WRITE
260 * io only ever reads user level buffer, and CIT_READ
263 policy_from_vma(&policy, vma, addr, count);
264 descr->cld_mode = vvp_mode_from_vma(vma);
265 descr->cld_obj = ll_i2info(inode)->lli_clob;
266 descr->cld_start = cl_index(descr->cld_obj,
267 policy.l_extent.start);
268 descr->cld_end = cl_index(descr->cld_obj,
269 policy.l_extent.end);
270 descr->cld_enq_flags = flags;
271 result = cl_io_lock_alloc_add(env, io, descr);
273 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
274 descr->cld_mode, descr->cld_start,
280 if (vma->vm_end - addr >= count)
283 count -= vma->vm_end - addr;
286 up_read(&mm->mmap_sem);
291 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
292 enum cl_lock_mode mode, loff_t start, loff_t end)
294 struct ccc_io *cio = ccc_env_io(env);
298 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
301 ccc_io_update_iov(env, cio, io);
303 if (io->u.ci_rw.crw_nonblock)
304 ast_flags |= CEF_NONBLOCK;
305 result = vvp_mmap_locks(env, cio, io);
307 result = ccc_io_one_lock(env, io, ast_flags, mode, start, end);
311 static int vvp_io_read_lock(const struct lu_env *env,
312 const struct cl_io_slice *ios)
314 struct cl_io *io = ios->cis_io;
315 struct ll_inode_info *lli = ll_i2info(ccc_object_inode(io->ci_obj));
319 /* XXX: Layer violation, we shouldn't see lsm at llite level. */
320 if (lli->lli_has_smd) /* lsm-less file doesn't need to lock */
321 result = vvp_io_rw_lock(env, io, CLM_READ,
322 io->u.ci_rd.rd.crw_pos,
323 io->u.ci_rd.rd.crw_pos +
324 io->u.ci_rd.rd.crw_count - 1);
330 static int vvp_io_fault_lock(const struct lu_env *env,
331 const struct cl_io_slice *ios)
333 struct cl_io *io = ios->cis_io;
334 struct vvp_io *vio = cl2vvp_io(env, ios);
336 * XXX LDLM_FL_CBPENDING
338 return ccc_io_one_lock_index
339 (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma),
340 io->u.ci_fault.ft_index, io->u.ci_fault.ft_index);
343 static int vvp_io_write_lock(const struct lu_env *env,
344 const struct cl_io_slice *ios)
346 struct cl_io *io = ios->cis_io;
350 if (io->u.ci_wr.wr_append) {
352 end = OBD_OBJECT_EOF;
354 start = io->u.ci_wr.wr.crw_pos;
355 end = start + io->u.ci_wr.wr.crw_count - 1;
357 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
360 static int vvp_io_setattr_iter_init(const struct lu_env *env,
361 const struct cl_io_slice *ios)
367 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
369 * Handles "lockless io" mode when extent locking is done by server.
371 static int vvp_io_setattr_lock(const struct lu_env *env,
372 const struct cl_io_slice *ios)
374 struct ccc_io *cio = ccc_env_io(env);
375 struct cl_io *io = ios->cis_io;
379 if (cl_io_is_trunc(io)) {
380 new_size = io->u.ci_setattr.sa_attr.lvb_size;
382 enqflags = CEF_DISCARD_DATA;
384 if ((io->u.ci_setattr.sa_attr.lvb_mtime >=
385 io->u.ci_setattr.sa_attr.lvb_ctime) ||
386 (io->u.ci_setattr.sa_attr.lvb_atime >=
387 io->u.ci_setattr.sa_attr.lvb_ctime))
391 cio->u.setattr.cui_local_lock = SETATTR_EXTENT_LOCK;
392 return ccc_io_one_lock(env, io, enqflags, CLM_WRITE,
393 new_size, OBD_OBJECT_EOF);
396 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
402 * Only ll_inode_size_lock is taken at this level.
404 ll_inode_size_lock(inode);
405 result = inode_newsize_ok(inode, size);
407 ll_inode_size_unlock(inode);
410 oldsize = inode->i_size;
411 i_size_write(inode, size);
413 truncate_pagecache(inode, oldsize, size);
414 ll_inode_size_unlock(inode);
418 static int vvp_io_setattr_trunc(const struct lu_env *env,
419 const struct cl_io_slice *ios,
420 struct inode *inode, loff_t size)
422 inode_dio_wait(inode);
426 static int vvp_io_setattr_time(const struct lu_env *env,
427 const struct cl_io_slice *ios)
429 struct cl_io *io = ios->cis_io;
430 struct cl_object *obj = io->ci_obj;
431 struct cl_attr *attr = ccc_env_thread_attr(env);
433 unsigned valid = CAT_CTIME;
435 cl_object_attr_lock(obj);
436 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
437 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
438 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
441 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
442 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
445 result = cl_object_attr_set(env, obj, attr, valid);
446 cl_object_attr_unlock(obj);
451 static int vvp_io_setattr_start(const struct lu_env *env,
452 const struct cl_io_slice *ios)
454 struct cl_io *io = ios->cis_io;
455 struct inode *inode = ccc_object_inode(io->ci_obj);
458 mutex_lock(&inode->i_mutex);
459 if (cl_io_is_trunc(io))
460 result = vvp_io_setattr_trunc(env, ios, inode,
461 io->u.ci_setattr.sa_attr.lvb_size);
463 result = vvp_io_setattr_time(env, ios);
467 static void vvp_io_setattr_end(const struct lu_env *env,
468 const struct cl_io_slice *ios)
470 struct cl_io *io = ios->cis_io;
471 struct inode *inode = ccc_object_inode(io->ci_obj);
473 if (cl_io_is_trunc(io)) {
474 /* Truncate in memory pages - they must be clean pages
475 * because osc has already notified to destroy osc_extents. */
476 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
477 inode_dio_write_done(inode);
479 mutex_unlock(&inode->i_mutex);
482 static void vvp_io_setattr_fini(const struct lu_env *env,
483 const struct cl_io_slice *ios)
485 vvp_io_fini(env, ios);
488 #ifdef HAVE_FILE_READV
489 static ssize_t lustre_generic_file_read(struct file *file,
490 struct ccc_io *vio, loff_t *ppos)
492 return generic_file_readv(file, vio->cui_iov, vio->cui_nrsegs, ppos);
495 static ssize_t lustre_generic_file_write(struct file *file,
496 struct ccc_io *vio, loff_t *ppos)
498 return generic_file_writev(file, vio->cui_iov, vio->cui_nrsegs, ppos);
501 static ssize_t lustre_generic_file_read(struct file *file,
502 struct ccc_io *vio, loff_t *ppos)
504 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
505 vio->cui_nrsegs, *ppos);
508 static ssize_t lustre_generic_file_write(struct file *file,
509 struct ccc_io *vio, loff_t *ppos)
511 return generic_file_aio_write(vio->cui_iocb, vio->cui_iov,
512 vio->cui_nrsegs, *ppos);
516 static int vvp_io_read_start(const struct lu_env *env,
517 const struct cl_io_slice *ios)
519 struct vvp_io *vio = cl2vvp_io(env, ios);
520 struct ccc_io *cio = cl2ccc_io(env, ios);
521 struct cl_io *io = ios->cis_io;
522 struct cl_object *obj = io->ci_obj;
523 struct inode *inode = ccc_object_inode(obj);
524 struct ll_ra_read *bead = &vio->cui_bead;
525 struct file *file = cio->cui_fd->fd_file;
528 loff_t pos = io->u.ci_rd.rd.crw_pos;
529 long cnt = io->u.ci_rd.rd.crw_count;
530 long tot = cio->cui_tot_count;
533 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
535 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
537 if (!can_populate_pages(env, io, inode))
540 result = ccc_prep_size(env, obj, io, pos, tot, &exceed);
543 else if (exceed != 0)
546 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
547 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
548 inode->i_ino, cnt, pos, i_size_read(inode));
550 /* turn off the kernel's read-ahead */
551 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
553 /* initialize read-ahead window once per syscall */
554 if (!vio->cui_ra_window_set) {
555 vio->cui_ra_window_set = 1;
556 bead->lrr_start = cl_index(obj, pos);
557 bead->lrr_count = cl_index(obj, tot + PAGE_CACHE_SIZE - 1);
558 ll_ra_read_in(file, bead);
563 switch (vio->cui_io_subtype) {
565 result = lustre_generic_file_read(file, cio, &pos);
568 result = generic_file_splice_read(file, &pos,
569 vio->u.splice.cui_pipe, cnt,
570 vio->u.splice.cui_flags);
571 /* LU-1109: do splice read stripe by stripe otherwise if it
572 * may make nfsd stuck if this read occupied all internal pipe
577 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
585 io->ci_nob += result;
586 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, cio->cui_fd,
594 static void vvp_io_read_fini(const struct lu_env *env, const struct cl_io_slice *ios)
596 struct vvp_io *vio = cl2vvp_io(env, ios);
597 struct ccc_io *cio = cl2ccc_io(env, ios);
599 if (vio->cui_ra_window_set)
600 ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
602 vvp_io_fini(env, ios);
605 static int vvp_io_write_start(const struct lu_env *env,
606 const struct cl_io_slice *ios)
608 struct ccc_io *cio = cl2ccc_io(env, ios);
609 struct cl_io *io = ios->cis_io;
610 struct cl_object *obj = io->ci_obj;
611 struct inode *inode = ccc_object_inode(obj);
612 struct file *file = cio->cui_fd->fd_file;
614 loff_t pos = io->u.ci_wr.wr.crw_pos;
615 size_t cnt = io->u.ci_wr.wr.crw_count;
619 if (!can_populate_pages(env, io, inode))
622 if (cl_io_is_append(io)) {
624 * PARALLEL IO This has to be changed for parallel IO doing
625 * out-of-order writes.
627 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
628 #ifndef HAVE_FILE_WRITEV
629 cio->cui_iocb->ki_pos = pos;
633 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
635 if (cio->cui_iov == NULL) /* from a temp io in ll_cl_init(). */
638 result = lustre_generic_file_write(file, cio, &pos);
643 io->ci_nob += result;
644 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
645 cio->cui_fd, pos, result, WRITE);
652 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
654 struct vm_fault *vmf = cfio->fault.ft_vmf;
656 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, vmf);
659 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
660 vmf->virtual_address);
661 if (unlikely(!(cfio->fault.ft_flags & VM_FAULT_LOCKED))) {
662 lock_page(vmf->page);
663 cfio->fault.ft_flags &= VM_FAULT_LOCKED;
666 cfio->ft_vmpage = vmf->page;
670 if (cfio->fault.ft_flags & VM_FAULT_SIGBUS) {
671 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
675 if (cfio->fault.ft_flags & VM_FAULT_OOM) {
676 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
680 if (cfio->fault.ft_flags & VM_FAULT_RETRY)
683 CERROR("unknow error in page fault %d!\n", cfio->fault.ft_flags);
687 static int vvp_io_fault_start(const struct lu_env *env,
688 const struct cl_io_slice *ios)
690 struct vvp_io *vio = cl2vvp_io(env, ios);
691 struct cl_io *io = ios->cis_io;
692 struct cl_object *obj = io->ci_obj;
693 struct inode *inode = ccc_object_inode(obj);
694 struct cl_fault_io *fio = &io->u.ci_fault;
695 struct vvp_fault_io *cfio = &vio->u.fault;
698 struct page *vmpage = NULL;
699 struct cl_page *page;
701 pgoff_t last; /* last page in a file data region */
703 if (fio->ft_executable &&
704 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
706 " changed while waiting for the page fault lock\n",
707 PFID(lu_object_fid(&obj->co_lu)));
709 /* offset of the last byte on the page */
710 offset = cl_offset(obj, fio->ft_index + 1) - 1;
711 LASSERT(cl_index(obj, offset) == fio->ft_index);
712 result = ccc_prep_size(env, obj, io, 0, offset + 1, NULL);
716 /* must return locked page */
717 if (fio->ft_mkwrite) {
718 LASSERT(cfio->ft_vmpage != NULL);
719 lock_page(cfio->ft_vmpage);
721 result = vvp_io_kernel_fault(cfio);
726 vmpage = cfio->ft_vmpage;
727 LASSERT(PageLocked(vmpage));
729 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
730 ll_invalidate_page(vmpage);
732 size = i_size_read(inode);
733 /* Though we have already held a cl_lock upon this page, but
734 * it still can be truncated locally. */
735 if (unlikely((vmpage->mapping != inode->i_mapping) ||
736 (page_offset(vmpage) > size))) {
737 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
739 /* return +1 to stop cl_io_loop() and ll_fault() will catch
741 GOTO(out, result = +1);
745 if (fio->ft_mkwrite ) {
748 * Capture the size while holding the lli_trunc_sem from above
749 * we want to make sure that we complete the mkwrite action
750 * while holding this lock. We need to make sure that we are
751 * not past the end of the file.
753 last_index = cl_index(obj, size - 1);
754 if (last_index < fio->ft_index) {
756 "llite: mkwrite and truncate race happened: "
758 vmpage->mapping,fio->ft_index,last_index);
760 * We need to return if we are
761 * passed the end of the file. This will propagate
762 * up the call stack to ll_page_mkwrite where
763 * we will return VM_FAULT_NOPAGE. Any non-negative
764 * value returned here will be silently
765 * converted to 0. If the vmpage->mapping is null
766 * the error code would be converted back to ENODATA
767 * in ll_page_mkwrite0. Thus we return -ENODATA
768 * to handle both cases
770 GOTO(out, result = -ENODATA);
774 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
776 GOTO(out, result = PTR_ERR(page));
778 /* if page is going to be written, we should add this page into cache
780 if (fio->ft_mkwrite) {
781 wait_on_page_writeback(vmpage);
782 if (set_page_dirty(vmpage)) {
785 /* vvp_page_assume() calls wait_on_page_writeback(). */
786 cl_page_assume(env, io, page);
788 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
789 vvp_write_pending(cl2ccc(obj), cp);
791 /* Do not set Dirty bit here so that in case IO is
792 * started before the page is really made dirty, we
793 * still have chance to detect it. */
794 result = cl_page_cache_add(env, io, page, CRT_WRITE);
795 LASSERT(cl_page_is_owned(page, io));
799 cl_page_unmap(env, io, page);
800 cl_page_discard(env, io, page);
801 cl_page_disown(env, io, page);
803 cl_page_put(env, page);
805 /* we're in big trouble, what can we do now? */
806 if (result == -EDQUOT)
810 cl_page_disown(env, io, page);
814 last = cl_index(obj, size - 1);
816 * The ft_index is only used in the case of
817 * a mkwrite action. We need to check
818 * our assertions are correct, since
819 * we should have caught this above
821 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last);
822 if (fio->ft_index == last)
824 * Last page is mapped partially.
826 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
828 fio->ft_nob = cl_page_size(obj);
830 lu_ref_add(&page->cp_reference, "fault", io);
835 /* return unlocked vmpage to avoid deadlocking */
838 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
842 static int vvp_io_fsync_start(const struct lu_env *env,
843 const struct cl_io_slice *ios)
845 /* we should mark TOWRITE bit to each dirty page in radix tree to
846 * verify pages have been written, but this is difficult because of
851 static int vvp_io_read_page(const struct lu_env *env,
852 const struct cl_io_slice *ios,
853 const struct cl_page_slice *slice)
855 struct cl_io *io = ios->cis_io;
856 struct cl_object *obj = slice->cpl_obj;
857 struct ccc_page *cp = cl2ccc_page(slice);
858 struct cl_page *page = slice->cpl_page;
859 struct inode *inode = ccc_object_inode(obj);
860 struct ll_sb_info *sbi = ll_i2sbi(inode);
861 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
862 struct ll_readahead_state *ras = &fd->fd_ras;
863 struct page *vmpage = cp->cpg_page;
864 struct cl_2queue *queue = &io->ci_queue;
867 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
868 LASSERT(slice->cpl_obj == obj);
872 if (sbi->ll_ra_info.ra_max_pages_per_file &&
873 sbi->ll_ra_info.ra_max_pages)
874 ras_update(sbi, inode, ras, page->cp_index,
875 cp->cpg_defer_uptodate);
877 /* Sanity check whether the page is protected by a lock. */
878 rc = cl_page_is_under_lock(env, io, page);
880 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %d\n",
881 rc == -ENODATA ? "without a lock" :
887 if (cp->cpg_defer_uptodate) {
889 cl_page_export(env, page, 1);
892 * Add page into the queue even when it is marked uptodate above.
893 * this will unlock it automatically as part of cl_page_list_disown().
895 cl_2queue_add(queue, page);
896 if (sbi->ll_ra_info.ra_max_pages_per_file &&
897 sbi->ll_ra_info.ra_max_pages)
898 ll_readahead(env, io, ras,
899 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
904 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
905 struct cl_page *page, struct ccc_page *cp,
906 enum cl_req_type crt)
908 struct cl_2queue *queue;
911 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
913 queue = &io->ci_queue;
914 cl_2queue_init_page(queue, page);
916 result = cl_io_submit_sync(env, io, crt, queue, 0);
917 LASSERT(cl_page_is_owned(page, io));
921 * in CRT_WRITE case page is left locked even in case of
924 cl_page_list_disown(env, io, &queue->c2_qin);
925 cl_2queue_fini(env, queue);
931 * Prepare partially written-to page for a write.
933 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
934 struct cl_object *obj, struct cl_page *pg,
936 unsigned from, unsigned to)
938 struct cl_attr *attr = ccc_env_thread_attr(env);
939 loff_t offset = cl_offset(obj, pg->cp_index);
942 cl_object_attr_lock(obj);
943 result = cl_object_attr_get(env, obj, attr);
944 cl_object_attr_unlock(obj);
947 * If are writing to a new page, no need to read old data.
948 * The extent locking will have updated the KMS, and for our
949 * purposes here we can treat it like i_size.
951 if (attr->cat_kms <= offset) {
952 char *kaddr = ll_kmap_atomic(cp->cpg_page, KM_USER0);
954 memset(kaddr, 0, cl_page_size(obj));
955 ll_kunmap_atomic(kaddr, KM_USER0);
956 } else if (cp->cpg_defer_uptodate)
959 result = vvp_page_sync_io(env, io, pg, cp, CRT_READ);
961 * In older implementations, obdo_refresh_inode is called here
962 * to update the inode because the write might modify the
963 * object info at OST. However, this has been proven useless,
964 * since LVB functions will be called when user space program
965 * tries to retrieve inode attribute. Also, see bug 15909 for
969 cl_page_export(env, pg, 1);
974 static int vvp_io_prepare_write(const struct lu_env *env,
975 const struct cl_io_slice *ios,
976 const struct cl_page_slice *slice,
977 unsigned from, unsigned to)
979 struct cl_object *obj = slice->cpl_obj;
980 struct ccc_page *cp = cl2ccc_page(slice);
981 struct cl_page *pg = slice->cpl_page;
982 struct page *vmpage = cp->cpg_page;
988 LINVRNT(cl_page_is_vmlocked(env, pg));
989 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
993 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
994 if (!PageUptodate(vmpage)) {
996 * We're completely overwriting an existing page, so _don't_
997 * set it up to date until commit_write
999 if (from == 0 && to == PAGE_CACHE_SIZE) {
1000 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
1001 POISON_PAGE(page, 0x11);
1003 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
1006 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
1010 static int vvp_io_commit_write(const struct lu_env *env,
1011 const struct cl_io_slice *ios,
1012 const struct cl_page_slice *slice,
1013 unsigned from, unsigned to)
1015 struct cl_object *obj = slice->cpl_obj;
1016 struct cl_io *io = ios->cis_io;
1017 struct ccc_page *cp = cl2ccc_page(slice);
1018 struct cl_page *pg = slice->cpl_page;
1019 struct inode *inode = ccc_object_inode(obj);
1020 struct ll_sb_info *sbi = ll_i2sbi(inode);
1021 struct ll_inode_info *lli = ll_i2info(inode);
1022 struct page *vmpage = cp->cpg_page;
1030 LINVRNT(cl_page_is_vmlocked(env, pg));
1031 LASSERT(vmpage->mapping->host == inode);
1033 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n");
1034 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
1037 * queue a write for some time in the future the first time we
1040 * This is different from what other file systems do: they usually
1041 * just mark page (and some of its buffers) dirty and rely on
1042 * balance_dirty_pages() to start a write-back. Lustre wants write-back
1043 * to be started earlier for the following reasons:
1045 * (1) with a large number of clients we need to limit the amount
1046 * of cached data on the clients a lot;
1048 * (2) large compute jobs generally want compute-only then io-only
1049 * and the IO should complete as quickly as possible;
1051 * (3) IO is batched up to the RPC size and is async until the
1052 * client max cache is hit
1053 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
1056 if (!PageDirty(vmpage)) {
1057 tallyop = LPROC_LL_DIRTY_MISSES;
1058 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
1060 /* page was added into cache successfully. */
1061 set_page_dirty(vmpage);
1062 vvp_write_pending(cl2ccc(obj), cp);
1063 } else if (result == -EDQUOT) {
1064 pgoff_t last_index = i_size_read(inode) >> PAGE_CACHE_SHIFT;
1065 bool need_clip = true;
1068 * Client ran out of disk space grant. Possible
1071 * (a) do a sync write, renewing grant;
1073 * (b) stop writing on this stripe, switch to the
1076 * (b) is a part of "parallel io" design that is the
1077 * ultimate goal. (a) is what "old" client did, and
1078 * what the new code continues to do for the time
1081 if (last_index > pg->cp_index) {
1082 to = PAGE_CACHE_SIZE;
1084 } else if (last_index == pg->cp_index) {
1085 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
1090 cl_page_clip(env, pg, 0, to);
1091 result = vvp_page_sync_io(env, io, pg, cp, CRT_WRITE);
1093 CERROR("Write page %lu of inode %p failed %d\n",
1094 pg->cp_index, inode, result);
1097 tallyop = LPROC_LL_DIRTY_HITS;
1100 ll_stats_ops_tally(sbi, tallyop, 1);
1102 /* Inode should be marked DIRTY even if no new page was marked DIRTY
1103 * because page could have been not flushed between 2 modifications.
1104 * It is important the file is marked DIRTY as soon as the I/O is done
1105 * Indeed, when cache is flushed, file could be already closed and it
1106 * is too late to warn the MDT.
1107 * It is acceptable that file is marked DIRTY even if I/O is dropped
1108 * for some reasons before being flushed to OST.
1111 spin_lock(&lli->lli_lock);
1112 lli->lli_flags |= LLIF_DATA_MODIFIED;
1113 spin_unlock(&lli->lli_lock);
1116 size = cl_offset(obj, pg->cp_index) + to;
1118 ll_inode_size_lock(inode);
1120 if (size > i_size_read(inode)) {
1121 cl_isize_write_nolock(inode, size);
1122 CDEBUG(D_VFSTRACE, DFID" updating i_size %lu\n",
1123 PFID(lu_object_fid(&obj->co_lu)),
1124 (unsigned long)size);
1126 cl_page_export(env, pg, 1);
1128 if (size > i_size_read(inode))
1129 cl_page_discard(env, io, pg);
1131 ll_inode_size_unlock(inode);
1135 static const struct cl_io_operations vvp_io_ops = {
1138 .cio_fini = vvp_io_read_fini,
1139 .cio_lock = vvp_io_read_lock,
1140 .cio_start = vvp_io_read_start,
1141 .cio_advance = ccc_io_advance
1144 .cio_fini = vvp_io_fini,
1145 .cio_lock = vvp_io_write_lock,
1146 .cio_start = vvp_io_write_start,
1147 .cio_advance = ccc_io_advance
1150 .cio_fini = vvp_io_setattr_fini,
1151 .cio_iter_init = vvp_io_setattr_iter_init,
1152 .cio_lock = vvp_io_setattr_lock,
1153 .cio_start = vvp_io_setattr_start,
1154 .cio_end = vvp_io_setattr_end
1157 .cio_fini = vvp_io_fault_fini,
1158 .cio_iter_init = vvp_io_fault_iter_init,
1159 .cio_lock = vvp_io_fault_lock,
1160 .cio_start = vvp_io_fault_start,
1161 .cio_end = ccc_io_end
1164 .cio_start = vvp_io_fsync_start,
1165 .cio_fini = vvp_io_fini
1168 .cio_fini = vvp_io_fini
1171 .cio_read_page = vvp_io_read_page,
1172 .cio_prepare_write = vvp_io_prepare_write,
1173 .cio_commit_write = vvp_io_commit_write
1176 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1179 struct vvp_io *vio = vvp_env_io(env);
1180 struct ccc_io *cio = ccc_env_io(env);
1181 struct inode *inode = ccc_object_inode(obj);
1184 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
1187 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1188 "restore needed %d\n",
1189 PFID(lu_object_fid(&obj->co_lu)),
1190 io->ci_ignore_layout, io->ci_verify_layout,
1191 cio->cui_layout_gen, io->ci_restore_needed);
1193 CL_IO_SLICE_CLEAN(cio, cui_cl);
1194 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
1195 vio->cui_ra_window_set = 0;
1197 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1199 struct ll_inode_info *lli = ll_i2info(inode);
1201 count = io->u.ci_rw.crw_count;
1202 /* "If nbyte is 0, read() will return 0 and have no other
1203 * results." -- Single Unix Spec */
1207 cio->cui_tot_count = count;
1208 cio->cui_tot_nrsegs = 0;
1210 /* for read/write, we store the jobid in the inode, and
1211 * it'll be fetched by osc when building RPC.
1213 * it's not accurate if the file is shared by different
1216 lustre_get_jobid(lli->lli_jobid);
1217 } else if (io->ci_type == CIT_SETATTR) {
1218 if (!cl_io_is_trunc(io))
1219 io->ci_lockreq = CILR_MANDATORY;
1222 /* ignore layout change for generic CIT_MISC but not for glimpse.
1223 * io context for glimpse must set ci_verify_layout to true,
1224 * see cl_glimpse_size0() for details. */
1225 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1226 io->ci_ignore_layout = 1;
1228 /* Enqueue layout lock and get layout version. We need to do this
1229 * even for operations requiring to open file, such as read and write,
1230 * because it might not grant layout lock in IT_OPEN. */
1231 if (result == 0 && !io->ci_ignore_layout) {
1232 result = ll_layout_refresh(inode, &cio->cui_layout_gen);
1233 if (result == -ENOENT)
1234 /* If the inode on MDS has been removed, but the objects
1235 * on OSTs haven't been destroyed (async unlink), layout
1236 * fetch will return -ENOENT, we'd ingore this error
1237 * and continue with dirty flush. LU-3230. */
1240 CERROR("%s: refresh file layout " DFID " error %d.\n",
1241 ll_get_fsname(inode->i_sb, NULL, 0),
1242 PFID(lu_object_fid(&obj->co_lu)), result);
1248 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1249 const struct cl_io_slice *slice)
1251 /* Caling just for assertion */
1252 cl2ccc_io(env, slice);
1253 return vvp_env_io(env);