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 static ssize_t lustre_generic_file_read(struct file *file,
489 struct ccc_io *vio, loff_t *ppos)
491 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
492 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_aio_write(vio->cui_iocb, vio->cui_iov,
499 vio->cui_nrsegs, *ppos);
502 static int vvp_io_read_start(const struct lu_env *env,
503 const struct cl_io_slice *ios)
505 struct vvp_io *vio = cl2vvp_io(env, ios);
506 struct ccc_io *cio = cl2ccc_io(env, ios);
507 struct cl_io *io = ios->cis_io;
508 struct cl_object *obj = io->ci_obj;
509 struct inode *inode = ccc_object_inode(obj);
510 struct ll_ra_read *bead = &vio->cui_bead;
511 struct file *file = cio->cui_fd->fd_file;
514 loff_t pos = io->u.ci_rd.rd.crw_pos;
515 long cnt = io->u.ci_rd.rd.crw_count;
516 long tot = cio->cui_tot_count;
519 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
521 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
523 if (!can_populate_pages(env, io, inode))
526 result = ccc_prep_size(env, obj, io, pos, tot, &exceed);
529 else if (exceed != 0)
532 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
533 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
534 inode->i_ino, cnt, pos, i_size_read(inode));
536 /* turn off the kernel's read-ahead */
537 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
539 /* initialize read-ahead window once per syscall */
540 if (!vio->cui_ra_window_set) {
541 vio->cui_ra_window_set = 1;
542 bead->lrr_start = cl_index(obj, pos);
543 bead->lrr_count = cl_index(obj, tot + PAGE_CACHE_SIZE - 1);
544 ll_ra_read_in(file, bead);
549 switch (vio->cui_io_subtype) {
551 result = lustre_generic_file_read(file, cio, &pos);
554 result = generic_file_splice_read(file, &pos,
555 vio->u.splice.cui_pipe, cnt,
556 vio->u.splice.cui_flags);
557 /* LU-1109: do splice read stripe by stripe otherwise if it
558 * may make nfsd stuck if this read occupied all internal pipe
563 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
571 io->ci_nob += result;
572 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, cio->cui_fd,
580 static void vvp_io_read_fini(const struct lu_env *env, const struct cl_io_slice *ios)
582 struct vvp_io *vio = cl2vvp_io(env, ios);
583 struct ccc_io *cio = cl2ccc_io(env, ios);
585 if (vio->cui_ra_window_set)
586 ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
588 vvp_io_fini(env, ios);
591 static int vvp_io_write_start(const struct lu_env *env,
592 const struct cl_io_slice *ios)
594 struct ccc_io *cio = cl2ccc_io(env, ios);
595 struct cl_io *io = ios->cis_io;
596 struct cl_object *obj = io->ci_obj;
597 struct inode *inode = ccc_object_inode(obj);
598 struct file *file = cio->cui_fd->fd_file;
600 loff_t pos = io->u.ci_wr.wr.crw_pos;
601 size_t cnt = io->u.ci_wr.wr.crw_count;
605 if (!can_populate_pages(env, io, inode))
608 if (cl_io_is_append(io)) {
610 * PARALLEL IO This has to be changed for parallel IO doing
611 * out-of-order writes.
613 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
614 cio->cui_iocb->ki_pos = pos;
617 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
619 if (cio->cui_iov == NULL) /* from a temp io in ll_cl_init(). */
622 result = lustre_generic_file_write(file, cio, &pos);
627 io->ci_nob += result;
628 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
629 cio->cui_fd, pos, result, WRITE);
636 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
638 struct vm_fault *vmf = cfio->fault.ft_vmf;
640 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, vmf);
643 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
644 vmf->virtual_address);
645 if (unlikely(!(cfio->fault.ft_flags & VM_FAULT_LOCKED))) {
646 lock_page(vmf->page);
647 cfio->fault.ft_flags &= VM_FAULT_LOCKED;
650 cfio->ft_vmpage = vmf->page;
654 if (cfio->fault.ft_flags & VM_FAULT_SIGBUS) {
655 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
659 if (cfio->fault.ft_flags & VM_FAULT_OOM) {
660 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
664 if (cfio->fault.ft_flags & VM_FAULT_RETRY)
667 CERROR("unknow error in page fault %d!\n", cfio->fault.ft_flags);
671 static int vvp_io_fault_start(const struct lu_env *env,
672 const struct cl_io_slice *ios)
674 struct vvp_io *vio = cl2vvp_io(env, ios);
675 struct cl_io *io = ios->cis_io;
676 struct cl_object *obj = io->ci_obj;
677 struct inode *inode = ccc_object_inode(obj);
678 struct cl_fault_io *fio = &io->u.ci_fault;
679 struct vvp_fault_io *cfio = &vio->u.fault;
682 struct page *vmpage = NULL;
683 struct cl_page *page;
685 pgoff_t last; /* last page in a file data region */
687 if (fio->ft_executable &&
688 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
690 " changed while waiting for the page fault lock\n",
691 PFID(lu_object_fid(&obj->co_lu)));
693 /* offset of the last byte on the page */
694 offset = cl_offset(obj, fio->ft_index + 1) - 1;
695 LASSERT(cl_index(obj, offset) == fio->ft_index);
696 result = ccc_prep_size(env, obj, io, 0, offset + 1, NULL);
700 /* must return locked page */
701 if (fio->ft_mkwrite) {
702 LASSERT(cfio->ft_vmpage != NULL);
703 lock_page(cfio->ft_vmpage);
705 result = vvp_io_kernel_fault(cfio);
710 vmpage = cfio->ft_vmpage;
711 LASSERT(PageLocked(vmpage));
713 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
714 ll_invalidate_page(vmpage);
716 size = i_size_read(inode);
717 /* Though we have already held a cl_lock upon this page, but
718 * it still can be truncated locally. */
719 if (unlikely((vmpage->mapping != inode->i_mapping) ||
720 (page_offset(vmpage) > size))) {
721 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
723 /* return +1 to stop cl_io_loop() and ll_fault() will catch
725 GOTO(out, result = +1);
729 if (fio->ft_mkwrite ) {
732 * Capture the size while holding the lli_trunc_sem from above
733 * we want to make sure that we complete the mkwrite action
734 * while holding this lock. We need to make sure that we are
735 * not past the end of the file.
737 last_index = cl_index(obj, size - 1);
738 if (last_index < fio->ft_index) {
740 "llite: mkwrite and truncate race happened: "
742 vmpage->mapping,fio->ft_index,last_index);
744 * We need to return if we are
745 * passed the end of the file. This will propagate
746 * up the call stack to ll_page_mkwrite where
747 * we will return VM_FAULT_NOPAGE. Any non-negative
748 * value returned here will be silently
749 * converted to 0. If the vmpage->mapping is null
750 * the error code would be converted back to ENODATA
751 * in ll_page_mkwrite0. Thus we return -ENODATA
752 * to handle both cases
754 GOTO(out, result = -ENODATA);
758 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
760 GOTO(out, result = PTR_ERR(page));
762 /* if page is going to be written, we should add this page into cache
764 if (fio->ft_mkwrite) {
765 wait_on_page_writeback(vmpage);
766 if (set_page_dirty(vmpage)) {
769 /* vvp_page_assume() calls wait_on_page_writeback(). */
770 cl_page_assume(env, io, page);
772 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
773 vvp_write_pending(cl2ccc(obj), cp);
775 /* Do not set Dirty bit here so that in case IO is
776 * started before the page is really made dirty, we
777 * still have chance to detect it. */
778 result = cl_page_cache_add(env, io, page, CRT_WRITE);
779 LASSERT(cl_page_is_owned(page, io));
783 cl_page_unmap(env, io, page);
784 cl_page_discard(env, io, page);
785 cl_page_disown(env, io, page);
787 cl_page_put(env, page);
789 /* we're in big trouble, what can we do now? */
790 if (result == -EDQUOT)
794 cl_page_disown(env, io, page);
798 last = cl_index(obj, size - 1);
800 * The ft_index is only used in the case of
801 * a mkwrite action. We need to check
802 * our assertions are correct, since
803 * we should have caught this above
805 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last);
806 if (fio->ft_index == last)
808 * Last page is mapped partially.
810 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
812 fio->ft_nob = cl_page_size(obj);
814 lu_ref_add(&page->cp_reference, "fault", io);
819 /* return unlocked vmpage to avoid deadlocking */
822 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
826 static int vvp_io_fsync_start(const struct lu_env *env,
827 const struct cl_io_slice *ios)
829 /* we should mark TOWRITE bit to each dirty page in radix tree to
830 * verify pages have been written, but this is difficult because of
835 static int vvp_io_read_page(const struct lu_env *env,
836 const struct cl_io_slice *ios,
837 const struct cl_page_slice *slice)
839 struct cl_io *io = ios->cis_io;
840 struct cl_object *obj = slice->cpl_obj;
841 struct ccc_page *cp = cl2ccc_page(slice);
842 struct cl_page *page = slice->cpl_page;
843 struct inode *inode = ccc_object_inode(obj);
844 struct ll_sb_info *sbi = ll_i2sbi(inode);
845 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
846 struct ll_readahead_state *ras = &fd->fd_ras;
847 struct page *vmpage = cp->cpg_page;
848 struct cl_2queue *queue = &io->ci_queue;
851 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
852 LASSERT(slice->cpl_obj == obj);
856 if (sbi->ll_ra_info.ra_max_pages_per_file &&
857 sbi->ll_ra_info.ra_max_pages)
858 ras_update(sbi, inode, ras, page->cp_index,
859 cp->cpg_defer_uptodate);
861 /* Sanity check whether the page is protected by a lock. */
862 rc = cl_page_is_under_lock(env, io, page);
864 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %d\n",
865 rc == -ENODATA ? "without a lock" :
871 if (cp->cpg_defer_uptodate) {
873 cl_page_export(env, page, 1);
876 * Add page into the queue even when it is marked uptodate above.
877 * this will unlock it automatically as part of cl_page_list_disown().
879 cl_2queue_add(queue, page);
880 if (sbi->ll_ra_info.ra_max_pages_per_file &&
881 sbi->ll_ra_info.ra_max_pages)
882 ll_readahead(env, io, ras,
883 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
888 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
889 struct cl_page *page, struct ccc_page *cp,
890 enum cl_req_type crt)
892 struct cl_2queue *queue;
895 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
897 queue = &io->ci_queue;
898 cl_2queue_init_page(queue, page);
900 result = cl_io_submit_sync(env, io, crt, queue, 0);
901 LASSERT(cl_page_is_owned(page, io));
905 * in CRT_WRITE case page is left locked even in case of
908 cl_page_list_disown(env, io, &queue->c2_qin);
909 cl_2queue_fini(env, queue);
915 * Prepare partially written-to page for a write.
917 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
918 struct cl_object *obj, struct cl_page *pg,
920 unsigned from, unsigned to)
922 struct cl_attr *attr = ccc_env_thread_attr(env);
923 loff_t offset = cl_offset(obj, pg->cp_index);
926 cl_object_attr_lock(obj);
927 result = cl_object_attr_get(env, obj, attr);
928 cl_object_attr_unlock(obj);
931 * If are writing to a new page, no need to read old data.
932 * The extent locking will have updated the KMS, and for our
933 * purposes here we can treat it like i_size.
935 if (attr->cat_kms <= offset) {
936 char *kaddr = ll_kmap_atomic(cp->cpg_page, KM_USER0);
938 memset(kaddr, 0, cl_page_size(obj));
939 ll_kunmap_atomic(kaddr, KM_USER0);
940 } else if (cp->cpg_defer_uptodate)
943 result = vvp_page_sync_io(env, io, pg, cp, CRT_READ);
945 * In older implementations, obdo_refresh_inode is called here
946 * to update the inode because the write might modify the
947 * object info at OST. However, this has been proven useless,
948 * since LVB functions will be called when user space program
949 * tries to retrieve inode attribute. Also, see bug 15909 for
953 cl_page_export(env, pg, 1);
958 static int vvp_io_prepare_write(const struct lu_env *env,
959 const struct cl_io_slice *ios,
960 const struct cl_page_slice *slice,
961 unsigned from, unsigned to)
963 struct cl_object *obj = slice->cpl_obj;
964 struct ccc_page *cp = cl2ccc_page(slice);
965 struct cl_page *pg = slice->cpl_page;
966 struct page *vmpage = cp->cpg_page;
972 LINVRNT(cl_page_is_vmlocked(env, pg));
973 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
977 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
978 if (!PageUptodate(vmpage)) {
980 * We're completely overwriting an existing page, so _don't_
981 * set it up to date until commit_write
983 if (from == 0 && to == PAGE_CACHE_SIZE) {
984 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
985 POISON_PAGE(page, 0x11);
987 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
990 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
994 static int vvp_io_commit_write(const struct lu_env *env,
995 const struct cl_io_slice *ios,
996 const struct cl_page_slice *slice,
997 unsigned from, unsigned to)
999 struct cl_object *obj = slice->cpl_obj;
1000 struct cl_io *io = ios->cis_io;
1001 struct ccc_page *cp = cl2ccc_page(slice);
1002 struct cl_page *pg = slice->cpl_page;
1003 struct inode *inode = ccc_object_inode(obj);
1004 struct ll_sb_info *sbi = ll_i2sbi(inode);
1005 struct ll_inode_info *lli = ll_i2info(inode);
1006 struct page *vmpage = cp->cpg_page;
1014 LINVRNT(cl_page_is_vmlocked(env, pg));
1015 LASSERT(vmpage->mapping->host == inode);
1017 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n");
1018 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
1021 * queue a write for some time in the future the first time we
1024 * This is different from what other file systems do: they usually
1025 * just mark page (and some of its buffers) dirty and rely on
1026 * balance_dirty_pages() to start a write-back. Lustre wants write-back
1027 * to be started earlier for the following reasons:
1029 * (1) with a large number of clients we need to limit the amount
1030 * of cached data on the clients a lot;
1032 * (2) large compute jobs generally want compute-only then io-only
1033 * and the IO should complete as quickly as possible;
1035 * (3) IO is batched up to the RPC size and is async until the
1036 * client max cache is hit
1037 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
1040 if (!PageDirty(vmpage)) {
1041 tallyop = LPROC_LL_DIRTY_MISSES;
1042 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
1044 /* page was added into cache successfully. */
1045 set_page_dirty(vmpage);
1046 vvp_write_pending(cl2ccc(obj), cp);
1047 } else if (result == -EDQUOT) {
1048 pgoff_t last_index = i_size_read(inode) >> PAGE_CACHE_SHIFT;
1049 bool need_clip = true;
1052 * Client ran out of disk space grant. Possible
1055 * (a) do a sync write, renewing grant;
1057 * (b) stop writing on this stripe, switch to the
1060 * (b) is a part of "parallel io" design that is the
1061 * ultimate goal. (a) is what "old" client did, and
1062 * what the new code continues to do for the time
1065 if (last_index > pg->cp_index) {
1066 to = PAGE_CACHE_SIZE;
1068 } else if (last_index == pg->cp_index) {
1069 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
1074 cl_page_clip(env, pg, 0, to);
1075 result = vvp_page_sync_io(env, io, pg, cp, CRT_WRITE);
1077 CERROR("Write page %lu of inode %p failed %d\n",
1078 pg->cp_index, inode, result);
1081 tallyop = LPROC_LL_DIRTY_HITS;
1084 ll_stats_ops_tally(sbi, tallyop, 1);
1086 /* Inode should be marked DIRTY even if no new page was marked DIRTY
1087 * because page could have been not flushed between 2 modifications.
1088 * It is important the file is marked DIRTY as soon as the I/O is done
1089 * Indeed, when cache is flushed, file could be already closed and it
1090 * is too late to warn the MDT.
1091 * It is acceptable that file is marked DIRTY even if I/O is dropped
1092 * for some reasons before being flushed to OST.
1095 spin_lock(&lli->lli_lock);
1096 lli->lli_flags |= LLIF_DATA_MODIFIED;
1097 spin_unlock(&lli->lli_lock);
1100 size = cl_offset(obj, pg->cp_index) + to;
1102 ll_inode_size_lock(inode);
1104 if (size > i_size_read(inode)) {
1105 cl_isize_write_nolock(inode, size);
1106 CDEBUG(D_VFSTRACE, DFID" updating i_size %lu\n",
1107 PFID(lu_object_fid(&obj->co_lu)),
1108 (unsigned long)size);
1110 cl_page_export(env, pg, 1);
1112 if (size > i_size_read(inode))
1113 cl_page_discard(env, io, pg);
1115 ll_inode_size_unlock(inode);
1119 static const struct cl_io_operations vvp_io_ops = {
1122 .cio_fini = vvp_io_read_fini,
1123 .cio_lock = vvp_io_read_lock,
1124 .cio_start = vvp_io_read_start,
1125 .cio_advance = ccc_io_advance
1128 .cio_fini = vvp_io_fini,
1129 .cio_lock = vvp_io_write_lock,
1130 .cio_start = vvp_io_write_start,
1131 .cio_advance = ccc_io_advance
1134 .cio_fini = vvp_io_setattr_fini,
1135 .cio_iter_init = vvp_io_setattr_iter_init,
1136 .cio_lock = vvp_io_setattr_lock,
1137 .cio_start = vvp_io_setattr_start,
1138 .cio_end = vvp_io_setattr_end
1141 .cio_fini = vvp_io_fault_fini,
1142 .cio_iter_init = vvp_io_fault_iter_init,
1143 .cio_lock = vvp_io_fault_lock,
1144 .cio_start = vvp_io_fault_start,
1145 .cio_end = ccc_io_end
1148 .cio_start = vvp_io_fsync_start,
1149 .cio_fini = vvp_io_fini
1152 .cio_fini = vvp_io_fini
1155 .cio_read_page = vvp_io_read_page,
1156 .cio_prepare_write = vvp_io_prepare_write,
1157 .cio_commit_write = vvp_io_commit_write
1160 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1163 struct vvp_io *vio = vvp_env_io(env);
1164 struct ccc_io *cio = ccc_env_io(env);
1165 struct inode *inode = ccc_object_inode(obj);
1168 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
1171 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1172 "restore needed %d\n",
1173 PFID(lu_object_fid(&obj->co_lu)),
1174 io->ci_ignore_layout, io->ci_verify_layout,
1175 cio->cui_layout_gen, io->ci_restore_needed);
1177 CL_IO_SLICE_CLEAN(cio, cui_cl);
1178 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
1179 vio->cui_ra_window_set = 0;
1181 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1183 struct ll_inode_info *lli = ll_i2info(inode);
1185 count = io->u.ci_rw.crw_count;
1186 /* "If nbyte is 0, read() will return 0 and have no other
1187 * results." -- Single Unix Spec */
1191 cio->cui_tot_count = count;
1192 cio->cui_tot_nrsegs = 0;
1194 /* for read/write, we store the jobid in the inode, and
1195 * it'll be fetched by osc when building RPC.
1197 * it's not accurate if the file is shared by different
1200 lustre_get_jobid(lli->lli_jobid);
1201 } else if (io->ci_type == CIT_SETATTR) {
1202 if (!cl_io_is_trunc(io))
1203 io->ci_lockreq = CILR_MANDATORY;
1206 /* ignore layout change for generic CIT_MISC but not for glimpse.
1207 * io context for glimpse must set ci_verify_layout to true,
1208 * see cl_glimpse_size0() for details. */
1209 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1210 io->ci_ignore_layout = 1;
1212 /* Enqueue layout lock and get layout version. We need to do this
1213 * even for operations requiring to open file, such as read and write,
1214 * because it might not grant layout lock in IT_OPEN. */
1215 if (result == 0 && !io->ci_ignore_layout) {
1216 result = ll_layout_refresh(inode, &cio->cui_layout_gen);
1217 if (result == -ENOENT)
1218 /* If the inode on MDS has been removed, but the objects
1219 * on OSTs haven't been destroyed (async unlink), layout
1220 * fetch will return -ENOENT, we'd ingore this error
1221 * and continue with dirty flush. LU-3230. */
1224 CERROR("%s: refresh file layout " DFID " error %d.\n",
1225 ll_get_fsname(inode->i_sb, NULL, 0),
1226 PFID(lu_object_fid(&obj->co_lu)), result);
1232 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1233 const struct cl_io_slice *slice)
1235 /* Caling just for assertion */
1236 cl2ccc_io(env, slice);
1237 return vvp_env_io(env);