1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
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>
41 #define DEBUG_SUBSYSTEM S_LLITE
44 # error This file is kernel only.
48 #include <lustre_lite.h>
50 #include "vvp_internal.h"
52 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
53 const struct cl_io_slice *slice);
56 * True, if \a io is a normal io, False for sendfile() / splice_{read|write}
58 int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
60 struct vvp_io *vio = vvp_env_io(env);
62 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
64 return vio->cui_io_subtype == IO_NORMAL;
67 /*****************************************************************************
73 static int vvp_io_fault_iter_init(const struct lu_env *env,
74 const struct cl_io_slice *ios)
76 struct vvp_io *vio = cl2vvp_io(env, ios);
77 struct inode *inode = ccc_object_inode(ios->cis_obj);
80 cl2ccc_io(env, ios)->cui_fd->fd_file->f_dentry->d_inode);
81 vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime);
85 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
87 struct cl_io *io = ios->cis_io;
88 struct cl_object *obj = io->ci_obj;
90 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
91 if (io->ci_type == CIT_READ) {
92 struct vvp_io *vio = cl2vvp_io(env, ios);
93 struct ccc_io *cio = cl2ccc_io(env, ios);
95 if (vio->cui_ra_window_set)
96 ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
101 static void vvp_io_fault_fini(const struct lu_env *env,
102 const struct cl_io_slice *ios)
104 struct cl_io *io = ios->cis_io;
105 struct cl_page *page = io->u.ci_fault.ft_page;
107 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
110 lu_ref_del(&page->cp_reference, "fault", io);
111 cl_page_put(env, page);
112 io->u.ci_fault.ft_page = NULL;
114 vvp_io_fini(env, ios);
117 enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
120 * we only want to hold PW locks if the mmap() can generate
121 * writes back to the file and that only happens in shared
124 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
129 static int vvp_mmap_locks(const struct lu_env *env,
130 struct ccc_io *vio, struct cl_io *io)
132 struct ccc_thread_info *cti = ccc_env_info(env);
133 struct vm_area_struct *vma;
134 struct cl_lock_descr *descr = &cti->cti_descr;
135 ldlm_policy_data_t policy;
142 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
144 if (!cl_is_normalio(env, io))
147 for (seg = 0; seg < vio->cui_nrsegs; seg++) {
148 const struct iovec *iv = &vio->cui_iov[seg];
150 addr = (unsigned long)iv->iov_base;
155 count += addr & (~CFS_PAGE_MASK);
156 addr &= CFS_PAGE_MASK;
157 while((vma = our_vma(addr, count)) != NULL) {
158 struct inode *inode = vma->vm_file->f_dentry->d_inode;
159 int flags = CEF_MUST;
161 if (ll_file_nolock(vma->vm_file)) {
163 * For no lock case, a lockless lock will be
170 * XXX: Required lock mode can be weakened: CIT_WRITE
171 * io only ever reads user level buffer, and CIT_READ
174 policy_from_vma(&policy, vma, addr, count);
175 descr->cld_mode = vvp_mode_from_vma(vma);
176 descr->cld_obj = ll_i2info(inode)->lli_clob;
177 descr->cld_start = cl_index(descr->cld_obj,
178 policy.l_extent.start);
179 descr->cld_end = cl_index(descr->cld_obj,
180 policy.l_extent.end);
181 descr->cld_enq_flags = flags;
182 result = cl_io_lock_alloc_add(env, io, descr);
186 if (vma->vm_end - addr >= count)
189 count -= vma->vm_end - addr;
196 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
197 enum cl_lock_mode mode, loff_t start, loff_t end)
199 struct ccc_io *cio = ccc_env_io(env);
203 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
206 ccc_io_update_iov(env, cio, io);
208 if (io->u.ci_rw.crw_nonblock)
209 ast_flags |= CEF_NONBLOCK;
210 result = vvp_mmap_locks(env, cio, io);
212 result = ccc_io_one_lock(env, io, ast_flags, mode, start, end);
216 static int vvp_io_read_lock(const struct lu_env *env,
217 const struct cl_io_slice *ios)
219 struct cl_io *io = ios->cis_io;
220 struct ll_inode_info *lli = ll_i2info(ccc_object_inode(io->ci_obj));
224 /* XXX: Layer violation, we shouldn't see lsm at llite level. */
225 if (lli->lli_smd != NULL) /* lsm-less file, don't need to lock */
226 result = vvp_io_rw_lock(env, io, CLM_READ,
227 io->u.ci_rd.rd.crw_pos,
228 io->u.ci_rd.rd.crw_pos +
229 io->u.ci_rd.rd.crw_count - 1);
235 static int vvp_io_fault_lock(const struct lu_env *env,
236 const struct cl_io_slice *ios)
238 struct cl_io *io = ios->cis_io;
239 struct vvp_io *vio = cl2vvp_io(env, ios);
241 * XXX LDLM_FL_CBPENDING
243 return ccc_io_one_lock_index
244 (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma),
245 io->u.ci_fault.ft_index, io->u.ci_fault.ft_index);
248 static int vvp_io_write_lock(const struct lu_env *env,
249 const struct cl_io_slice *ios)
251 struct cl_io *io = ios->cis_io;
255 if (io->u.ci_wr.wr_append) {
257 end = OBD_OBJECT_EOF;
259 start = io->u.ci_wr.wr.crw_pos;
260 end = start + io->u.ci_wr.wr.crw_count - 1;
262 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
265 static int vvp_io_setattr_iter_init(const struct lu_env *env,
266 const struct cl_io_slice *ios)
268 struct ccc_io *cio = ccc_env_io(env);
269 struct inode *inode = ccc_object_inode(ios->cis_obj);
272 * We really need to get our PW lock before we change inode->i_size.
273 * If we don't we can race with other i_size updaters on our node,
274 * like ll_file_read. We can also race with i_size propogation to
275 * other nodes through dirtying and writeback of final cached pages.
276 * This last one is especially bad for racing o_append users on other
279 UNLOCK_INODE_MUTEX(inode);
280 if (cl_io_is_trunc(ios->cis_io))
281 UP_WRITE_I_ALLOC_SEM(inode);
282 cio->u.setattr.cui_locks_released = 1;
287 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
289 * Handles "lockless io" mode when extent locking is done by server.
291 static int vvp_io_setattr_lock(const struct lu_env *env,
292 const struct cl_io_slice *ios)
294 struct ccc_io *cio = ccc_env_io(env);
295 struct cl_io *io = ios->cis_io;
299 if (cl_io_is_trunc(io)) {
300 new_size = io->u.ci_setattr.sa_attr.lvb_size;
302 enqflags = CEF_DISCARD_DATA;
304 LASSERT((io->u.ci_setattr.sa_attr.lvb_mtime <
305 io->u.ci_setattr.sa_attr.lvb_ctime) ||
306 (io->u.ci_setattr.sa_attr.lvb_atime <
307 io->u.ci_setattr.sa_attr.lvb_ctime));
310 cio->u.setattr.cui_local_lock = SETATTR_EXTENT_LOCK;
311 return ccc_io_one_lock(env, io, enqflags, CLM_WRITE,
312 new_size, OBD_OBJECT_EOF);
315 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
319 * Only ll_inode_size_lock is taken at this level. lov_stripe_lock()
320 * is grabbed by ll_truncate() only over call to obd_adjust_kms(). If
321 * vmtruncate returns 0, then ll_truncate dropped ll_inode_size_lock()
323 ll_inode_size_lock(inode, 0);
324 result = vmtruncate(inode, size);
326 ll_inode_size_unlock(inode, 0);
331 static int vvp_io_setattr_trunc(const struct lu_env *env,
332 const struct cl_io_slice *ios,
333 struct inode *inode, loff_t size)
335 struct vvp_io *vio = cl2vvp_io(env, ios);
336 struct cl_io *io = ios->cis_io;
337 struct cl_object *obj = ios->cis_obj;
338 pgoff_t start = cl_index(obj, size);
341 DOWN_WRITE_I_ALLOC_SEM(inode);
343 result = vvp_do_vmtruncate(inode, size);
346 * If a page is partially truncated, keep it owned across truncate to
349 * XXX this properly belongs to osc, because races in question are OST
352 if (cl_offset(obj, start) != size) {
353 struct cl_object_header *hdr;
355 hdr = cl_object_header(obj);
356 cfs_spin_lock(&hdr->coh_page_guard);
357 vio->cui_partpage = cl_page_lookup(hdr, start);
358 cfs_spin_unlock(&hdr->coh_page_guard);
360 if (vio->cui_partpage != NULL)
362 * Wait for the transfer completion for a partially
363 * truncated page to avoid dead-locking an OST with
364 * the concurrent page-wise overlapping WRITE and
365 * PUNCH requests. BUG:17397.
367 * Partial page is disowned in vvp_io_trunc_end().
369 cl_page_own(env, io, vio->cui_partpage);
371 vio->cui_partpage = NULL;
375 static int vvp_io_setattr_time(const struct lu_env *env,
376 const struct cl_io_slice *ios)
378 struct cl_io *io = ios->cis_io;
379 struct cl_object *obj = io->ci_obj;
380 struct cl_attr *attr = ccc_env_thread_attr(env);
382 unsigned valid = CAT_CTIME;
384 cl_object_attr_lock(obj);
385 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
386 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
387 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
390 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
391 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
394 result = cl_object_attr_set(env, obj, attr, valid);
395 cl_object_attr_unlock(obj);
400 static int vvp_io_setattr_start(const struct lu_env *env,
401 const struct cl_io_slice *ios)
403 struct ccc_io *cio = cl2ccc_io(env, ios);
404 struct cl_io *io = ios->cis_io;
405 struct inode *inode = ccc_object_inode(io->ci_obj);
407 LASSERT(cio->u.setattr.cui_locks_released);
409 LOCK_INODE_MUTEX(inode);
410 cio->u.setattr.cui_locks_released = 0;
412 if (cl_io_is_trunc(io))
413 return vvp_io_setattr_trunc(env, ios, inode,
414 io->u.ci_setattr.sa_attr.lvb_size);
416 return vvp_io_setattr_time(env, ios);
419 static void vvp_io_setattr_end(const struct lu_env *env,
420 const struct cl_io_slice *ios)
422 struct vvp_io *vio = cl2vvp_io(env, ios);
423 struct cl_io *io = ios->cis_io;
424 struct inode *inode = ccc_object_inode(io->ci_obj);
426 if (!cl_io_is_trunc(io))
428 if (vio->cui_partpage != NULL) {
429 cl_page_disown(env, ios->cis_io, vio->cui_partpage);
430 cl_page_put(env, vio->cui_partpage);
431 vio->cui_partpage = NULL;
435 * Do vmtruncate again, to remove possible stale pages populated by
436 * competing read threads. bz20645.
438 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
441 static void vvp_io_setattr_fini(const struct lu_env *env,
442 const struct cl_io_slice *ios)
444 struct ccc_io *cio = ccc_env_io(env);
445 struct cl_io *io = ios->cis_io;
446 struct inode *inode = ccc_object_inode(ios->cis_io->ci_obj);
448 if (cio->u.setattr.cui_locks_released) {
449 LOCK_INODE_MUTEX(inode);
450 if (cl_io_is_trunc(io))
451 DOWN_WRITE_I_ALLOC_SEM(inode);
452 cio->u.setattr.cui_locks_released = 0;
454 vvp_io_fini(env, ios);
457 #ifdef HAVE_FILE_READV
458 static ssize_t lustre_generic_file_read(struct file *file,
459 struct ccc_io *vio, loff_t *ppos)
461 return generic_file_readv(file, vio->cui_iov, vio->cui_nrsegs, ppos);
464 static ssize_t lustre_generic_file_write(struct file *file,
465 struct ccc_io *vio, loff_t *ppos)
467 return generic_file_writev(file, vio->cui_iov, vio->cui_nrsegs, ppos);
470 static ssize_t lustre_generic_file_read(struct file *file,
471 struct ccc_io *vio, loff_t *ppos)
473 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
474 vio->cui_nrsegs, *ppos);
477 static ssize_t lustre_generic_file_write(struct file *file,
478 struct ccc_io *vio, loff_t *ppos)
480 return generic_file_aio_write(vio->cui_iocb, vio->cui_iov,
481 vio->cui_nrsegs, *ppos);
485 static int vvp_io_read_start(const struct lu_env *env,
486 const struct cl_io_slice *ios)
488 struct vvp_io *vio = cl2vvp_io(env, ios);
489 struct ccc_io *cio = cl2ccc_io(env, ios);
490 struct cl_io *io = ios->cis_io;
491 struct cl_object *obj = io->ci_obj;
492 struct inode *inode = ccc_object_inode(obj);
493 struct ll_ra_read *bead = &vio->cui_bead;
494 struct file *file = cio->cui_fd->fd_file;
497 loff_t pos = io->u.ci_rd.rd.crw_pos;
498 long cnt = io->u.ci_rd.rd.crw_count;
499 long tot = cio->cui_tot_count;
502 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
504 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
506 result = ccc_prep_size(env, obj, io, pos, tot, 1, &exceed);
509 else if (exceed != 0)
512 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
513 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
514 inode->i_ino, cnt, pos, i_size_read(inode));
516 /* turn off the kernel's read-ahead */
517 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
519 /* initialize read-ahead window once per syscall */
520 if (!vio->cui_ra_window_set) {
521 vio->cui_ra_window_set = 1;
522 bead->lrr_start = cl_index(obj, pos);
524 * XXX: explicit CFS_PAGE_SIZE
526 bead->lrr_count = cl_index(obj, tot + CFS_PAGE_SIZE - 1);
527 ll_ra_read_in(file, bead);
532 switch (vio->cui_io_subtype) {
534 result = lustre_generic_file_read(file, cio, &pos);
536 #ifdef HAVE_KERNEL_SENDFILE
538 result = generic_file_sendfile(file, &pos, cnt,
539 vio->u.sendfile.cui_actor,
540 vio->u.sendfile.cui_target);
543 #ifdef HAVE_KERNEL_SPLICE_READ
545 result = generic_file_splice_read(file, &pos,
546 vio->u.splice.cui_pipe, cnt,
547 vio->u.splice.cui_flags);
551 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
559 io->ci_nob += result;
560 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
561 cio->cui_fd, pos, result, 0);
567 static int vvp_io_write_start(const struct lu_env *env,
568 const struct cl_io_slice *ios)
570 struct ccc_io *cio = cl2ccc_io(env, ios);
571 struct cl_io *io = ios->cis_io;
572 struct cl_object *obj = io->ci_obj;
573 struct inode *inode = ccc_object_inode(obj);
574 struct file *file = cio->cui_fd->fd_file;
576 loff_t pos = io->u.ci_wr.wr.crw_pos;
577 size_t cnt = io->u.ci_wr.wr.crw_count;
581 if (cl_io_is_append(io)) {
583 * PARALLEL IO This has to be changed for parallel IO doing
584 * out-of-order writes.
586 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
587 #ifndef HAVE_FILE_WRITEV
588 cio->cui_iocb->ki_pos = pos;
592 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
594 if (cio->cui_iov == NULL) /* from a temp io in ll_cl_init(). */
597 result = lustre_generic_file_write(file, cio, &pos);
602 io->ci_nob += result;
603 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
604 cio->cui_fd, pos, result, 0);
610 #ifndef HAVE_VM_OP_FAULT
611 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
615 vmpage = filemap_nopage(cfio->ft_vma, cfio->nopage.ft_address,
616 cfio->nopage.ft_type);
618 if (vmpage == NOPAGE_SIGBUS) {
619 CDEBUG(D_PAGE, "got addr %lu type %lx - SIGBUS\n",
620 cfio->nopage.ft_address,(long)cfio->nopage.ft_type);
622 } else if (vmpage == NOPAGE_OOM) {
623 CDEBUG(D_PAGE, "got addr %lu type %lx - OOM\n",
624 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
628 LL_CDEBUG_PAGE(D_PAGE, vmpage, "got addr %lu type %lx\n",
629 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
631 cfio->ft_vmpage = vmpage;
636 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
638 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, cfio->fault.ft_vmf);
640 if (cfio->fault.ft_vmf->page) {
641 LL_CDEBUG_PAGE(D_PAGE, cfio->fault.ft_vmf->page,
642 "got addr %p type NOPAGE\n",
643 cfio->fault.ft_vmf->virtual_address);
644 /*XXX workaround to bug in CLIO - he deadlocked with
645 lock cancel if page locked */
646 if (likely(cfio->fault.ft_flags & VM_FAULT_LOCKED)) {
647 unlock_page(cfio->fault.ft_vmf->page);
648 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
651 cfio->ft_vmpage = cfio->fault.ft_vmf->page;
655 if (unlikely (cfio->fault.ft_flags & VM_FAULT_ERROR)) {
656 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n",
657 cfio->fault.ft_vmf->virtual_address);
661 if (unlikely (cfio->fault.ft_flags & VM_FAULT_NOPAGE)) {
662 CDEBUG(D_PAGE, "got addr %p - OOM\n",
663 cfio->fault.ft_vmf->virtual_address);
667 CERROR("unknow error in page fault!\n");
673 static int vvp_io_fault_start(const struct lu_env *env,
674 const struct cl_io_slice *ios)
676 struct vvp_io *vio = cl2vvp_io(env, ios);
677 struct cl_io *io = ios->cis_io;
678 struct cl_object *obj = io->ci_obj;
679 struct inode *inode = ccc_object_inode(obj);
680 struct cl_fault_io *fio = &io->u.ci_fault;
681 struct vvp_fault_io *cfio = &vio->u.fault;
683 int kernel_result = 0;
685 struct cl_page *page;
687 pgoff_t last; /* last page in a file data region */
689 if (fio->ft_executable &&
690 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
692 " changed while waiting for the page fault lock\n",
693 PFID(lu_object_fid(&obj->co_lu)));
695 /* offset of the last byte on the page */
696 offset = cl_offset(obj, fio->ft_index + 1) - 1;
697 LASSERT(cl_index(obj, offset) == fio->ft_index);
698 result = ccc_prep_size(env, obj, io, 0, offset + 1, 0, NULL);
702 /* must return locked page */
703 kernel_result = vvp_io_kernel_fault(cfio);
704 if (kernel_result != 0)
705 return kernel_result;
706 /* Temporarily lock vmpage to keep cl_page_find() happy. */
707 lock_page(cfio->ft_vmpage);
708 page = cl_page_find(env, obj, fio->ft_index, cfio->ft_vmpage,
710 unlock_page(cfio->ft_vmpage);
712 page_cache_release(cfio->ft_vmpage);
713 cfio->ft_vmpage = NULL;
714 return PTR_ERR(page);
717 size = i_size_read(inode);
718 last = cl_index(obj, size - 1);
719 if (fio->ft_index == last)
721 * Last page is mapped partially.
723 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
725 fio->ft_nob = cl_page_size(obj);
727 lu_ref_add(&page->cp_reference, "fault", io);
730 * Certain 2.6 kernels return not-NULL from
731 * filemap_nopage() when page is beyond the file size,
732 * on the grounds that "An external ptracer can access
733 * pages that normally aren't accessible.." Don't
734 * propagate such page fault to the lower layers to
735 * avoid side-effects like KMS updates.
737 if (fio->ft_index > last)
743 static int vvp_io_read_page(const struct lu_env *env,
744 const struct cl_io_slice *ios,
745 const struct cl_page_slice *slice)
747 struct cl_io *io = ios->cis_io;
748 struct cl_object *obj = slice->cpl_obj;
749 struct ccc_page *cp = cl2ccc_page(slice);
750 struct cl_page *page = slice->cpl_page;
751 struct inode *inode = ccc_object_inode(obj);
752 struct ll_sb_info *sbi = ll_i2sbi(inode);
753 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
754 struct ll_readahead_state *ras = &fd->fd_ras;
755 cfs_page_t *vmpage = cp->cpg_page;
756 struct cl_2queue *queue = &io->ci_queue;
759 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
760 LASSERT(slice->cpl_obj == obj);
764 if (sbi->ll_ra_info.ra_max_pages_per_file)
765 ras_update(sbi, inode, ras, page->cp_index,
766 cp->cpg_defer_uptodate);
768 /* Sanity check whether the page is protected by a lock. */
769 rc = cl_page_is_under_lock(env, io, page);
771 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %i\n",
772 rc == -ENODATA ? "without a lock" :
778 if (cp->cpg_defer_uptodate) {
780 cl_page_export(env, page, 1);
783 * Add page into the queue even when it is marked uptodate above.
784 * this will unlock it automatically as part of cl_page_list_disown().
786 cl_2queue_add(queue, page);
787 if (sbi->ll_ra_info.ra_max_pages_per_file)
788 ll_readahead(env, io, ras,
789 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
794 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
795 struct cl_page *page, struct ccc_page *cp,
796 int to, enum cl_req_type crt)
798 struct cl_2queue *queue;
801 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
803 queue = &io->ci_queue;
805 cl_2queue_init_page(queue, page);
806 cl_page_clip(env, page, 0, to);
808 result = cl_io_submit_sync(env, io, crt, queue, CRP_NORMAL, 0);
809 LASSERT(cl_page_is_owned(page, io));
810 cl_page_clip(env, page, 0, CFS_PAGE_SIZE);
814 * in CRT_WRITE case page is left locked even in case of
817 cl_page_list_disown(env, io, &queue->c2_qin);
818 cl_2queue_fini(env, queue);
824 * Prepare partially written-to page for a write.
826 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
827 struct cl_object *obj, struct cl_page *pg,
829 unsigned from, unsigned to)
831 struct cl_attr *attr = ccc_env_thread_attr(env);
832 loff_t offset = cl_offset(obj, pg->cp_index);
835 cl_object_attr_lock(obj);
836 result = cl_object_attr_get(env, obj, attr);
837 cl_object_attr_unlock(obj);
840 * If are writing to a new page, no need to read old data.
841 * The extent locking will have updated the KMS, and for our
842 * purposes here we can treat it like i_size.
844 if (attr->cat_kms <= offset) {
845 char *kaddr = kmap_atomic(cp->cpg_page, KM_USER0);
847 memset(kaddr, 0, cl_page_size(obj));
848 kunmap_atomic(kaddr, KM_USER0);
849 } else if (cp->cpg_defer_uptodate)
852 result = vvp_page_sync_io(env, io, pg, cp,
853 CFS_PAGE_SIZE, CRT_READ);
855 * In older implementations, obdo_refresh_inode is called here
856 * to update the inode because the write might modify the
857 * object info at OST. However, this has been proven useless,
858 * since LVB functions will be called when user space program
859 * tries to retrieve inode attribute. Also, see bug 15909 for
863 cl_page_export(env, pg, 1);
868 static int vvp_io_prepare_write(const struct lu_env *env,
869 const struct cl_io_slice *ios,
870 const struct cl_page_slice *slice,
871 unsigned from, unsigned to)
873 struct cl_object *obj = slice->cpl_obj;
874 struct ccc_page *cp = cl2ccc_page(slice);
875 struct cl_page *pg = slice->cpl_page;
876 cfs_page_t *vmpage = cp->cpg_page;
882 LINVRNT(cl_page_is_vmlocked(env, pg));
883 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
887 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
888 if (!PageUptodate(vmpage)) {
890 * We're completely overwriting an existing page, so _don't_
891 * set it up to date until commit_write
893 if (from == 0 && to == CFS_PAGE_SIZE) {
894 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
895 POISON_PAGE(page, 0x11);
897 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
900 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
904 static int vvp_io_commit_write(const struct lu_env *env,
905 const struct cl_io_slice *ios,
906 const struct cl_page_slice *slice,
907 unsigned from, unsigned to)
909 struct cl_object *obj = slice->cpl_obj;
910 struct cl_io *io = ios->cis_io;
911 struct ccc_page *cp = cl2ccc_page(slice);
912 struct cl_page *pg = slice->cpl_page;
913 struct inode *inode = ccc_object_inode(obj);
914 struct ll_sb_info *sbi = ll_i2sbi(inode);
915 cfs_page_t *vmpage = cp->cpg_page;
923 LINVRNT(cl_page_is_vmlocked(env, pg));
924 LASSERT(vmpage->mapping->host == inode);
926 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n");
927 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
930 * queue a write for some time in the future the first time we
933 * This is different from what other file systems do: they usually
934 * just mark page (and some of its buffers) dirty and rely on
935 * balance_dirty_pages() to start a write-back. Lustre wants write-back
936 * to be started earlier for the following reasons:
938 * (1) with a large number of clients we need to limit the amount
939 * of cached data on the clients a lot;
941 * (2) large compute jobs generally want compute-only then io-only
942 * and the IO should complete as quickly as possible;
944 * (3) IO is batched up to the RPC size and is async until the
945 * client max cache is hit
946 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
949 if (!PageDirty(vmpage)) {
950 tallyop = LPROC_LL_DIRTY_MISSES;
951 vvp_write_pending(cl2ccc(obj), cp);
952 set_page_dirty(vmpage);
953 /* ll_set_page_dirty() does the same for now, but
954 * it will not soon. */
955 vvp_write_pending(cl2ccc(obj), cp);
956 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
957 if (result == -EDQUOT)
959 * Client ran out of disk space grant. Possible
962 * (a) do a sync write, renewing grant;
964 * (b) stop writing on this stripe, switch to the
967 * (b) is a part of "parallel io" design that is the
968 * ultimate goal. (a) is what "old" client did, and
969 * what the new code continues to do for the time
972 result = vvp_page_sync_io(env, io, pg, cp,
975 CERROR("Write page %lu of inode %p failed %d\n",
976 pg->cp_index, inode, result);
978 tallyop = LPROC_LL_DIRTY_HITS;
981 ll_stats_ops_tally(sbi, tallyop, 1);
983 size = cl_offset(obj, pg->cp_index) + to;
985 ll_inode_size_lock(inode, 0);
987 if (size > i_size_read(inode)) {
988 cl_isize_write_nolock(inode, size);
989 CDEBUG(D_VFSTRACE, DFID" updating i_size "LPU64"\n",
990 PFID(lu_object_fid(&obj->co_lu)), size);
992 cl_page_export(env, pg, 1);
994 if (size > i_size_read(inode))
995 cl_page_discard(env, io, pg);
997 ll_inode_size_unlock(inode, 0);
1002 static const struct cl_io_operations vvp_io_ops = {
1005 .cio_fini = vvp_io_fini,
1006 .cio_lock = vvp_io_read_lock,
1007 .cio_start = vvp_io_read_start,
1008 .cio_advance = ccc_io_advance
1011 .cio_fini = vvp_io_fini,
1012 .cio_lock = vvp_io_write_lock,
1013 .cio_start = vvp_io_write_start,
1014 .cio_advance = ccc_io_advance
1017 .cio_fini = vvp_io_setattr_fini,
1018 .cio_iter_init = vvp_io_setattr_iter_init,
1019 .cio_lock = vvp_io_setattr_lock,
1020 .cio_start = vvp_io_setattr_start,
1021 .cio_end = vvp_io_setattr_end
1024 .cio_fini = vvp_io_fault_fini,
1025 .cio_iter_init = vvp_io_fault_iter_init,
1026 .cio_lock = vvp_io_fault_lock,
1027 .cio_start = vvp_io_fault_start,
1028 .cio_end = ccc_io_end
1031 .cio_fini = vvp_io_fini
1034 .cio_read_page = vvp_io_read_page,
1035 .cio_prepare_write = vvp_io_prepare_write,
1036 .cio_commit_write = vvp_io_commit_write
1039 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1042 struct vvp_io *vio = vvp_env_io(env);
1043 struct ccc_io *cio = ccc_env_io(env);
1044 struct inode *inode = ccc_object_inode(obj);
1045 struct ll_sb_info *sbi = ll_i2sbi(inode);
1048 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
1051 CL_IO_SLICE_CLEAN(cio, cui_cl);
1052 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
1053 vio->cui_ra_window_set = 0;
1055 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1059 count = io->u.ci_rw.crw_count;
1060 op = io->ci_type == CIT_READ ?
1061 LPROC_LL_READ_BYTES : LPROC_LL_WRITE_BYTES;
1062 /* "If nbyte is 0, read() will return 0 and have no other
1063 * results." -- Single Unix Spec */
1067 cio->cui_tot_count = count;
1068 cio->cui_tot_nrsegs = 0;
1069 ll_stats_ops_tally(sbi, op, count);
1071 } else if (io->ci_type == CIT_SETATTR) {
1072 if (cl_io_is_trunc(io))
1073 /* lockless truncate? */
1074 ll_stats_ops_tally(sbi, LPROC_LL_TRUNC, 1);
1076 io->ci_lockreq = CILR_MANDATORY;
1081 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1082 const struct cl_io_slice *slice)
1084 /* Caling just for assertion */
1085 cl2ccc_io(env, slice);
1086 return vvp_env_io(env);