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);
184 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
185 descr->cld_mode, descr->cld_start,
191 if (vma->vm_end - addr >= count)
194 count -= vma->vm_end - addr;
201 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
202 enum cl_lock_mode mode, loff_t start, loff_t end)
204 struct ccc_io *cio = ccc_env_io(env);
208 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
211 ccc_io_update_iov(env, cio, io);
213 if (io->u.ci_rw.crw_nonblock)
214 ast_flags |= CEF_NONBLOCK;
215 result = vvp_mmap_locks(env, cio, io);
217 result = ccc_io_one_lock(env, io, ast_flags, mode, start, end);
221 static int vvp_io_read_lock(const struct lu_env *env,
222 const struct cl_io_slice *ios)
224 struct cl_io *io = ios->cis_io;
225 struct ll_inode_info *lli = ll_i2info(ccc_object_inode(io->ci_obj));
229 /* XXX: Layer violation, we shouldn't see lsm at llite level. */
230 if (lli->lli_smd != NULL) /* lsm-less file, don't need to lock */
231 result = vvp_io_rw_lock(env, io, CLM_READ,
232 io->u.ci_rd.rd.crw_pos,
233 io->u.ci_rd.rd.crw_pos +
234 io->u.ci_rd.rd.crw_count - 1);
240 static int vvp_io_fault_lock(const struct lu_env *env,
241 const struct cl_io_slice *ios)
243 struct cl_io *io = ios->cis_io;
244 struct vvp_io *vio = cl2vvp_io(env, ios);
246 * XXX LDLM_FL_CBPENDING
248 return ccc_io_one_lock_index
249 (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma),
250 io->u.ci_fault.ft_index, io->u.ci_fault.ft_index);
253 static int vvp_io_write_lock(const struct lu_env *env,
254 const struct cl_io_slice *ios)
256 struct cl_io *io = ios->cis_io;
260 if (io->u.ci_wr.wr_append) {
262 end = OBD_OBJECT_EOF;
264 start = io->u.ci_wr.wr.crw_pos;
265 end = start + io->u.ci_wr.wr.crw_count - 1;
267 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
270 static int vvp_io_setattr_iter_init(const struct lu_env *env,
271 const struct cl_io_slice *ios)
273 struct ccc_io *cio = ccc_env_io(env);
274 struct inode *inode = ccc_object_inode(ios->cis_obj);
277 * We really need to get our PW lock before we change inode->i_size.
278 * If we don't we can race with other i_size updaters on our node,
279 * like ll_file_read. We can also race with i_size propogation to
280 * other nodes through dirtying and writeback of final cached pages.
281 * This last one is especially bad for racing o_append users on other
284 UNLOCK_INODE_MUTEX(inode);
285 if (cl_io_is_trunc(ios->cis_io))
286 UP_WRITE_I_ALLOC_SEM(inode);
287 cio->u.setattr.cui_locks_released = 1;
292 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
294 * Handles "lockless io" mode when extent locking is done by server.
296 static int vvp_io_setattr_lock(const struct lu_env *env,
297 const struct cl_io_slice *ios)
299 struct ccc_io *cio = ccc_env_io(env);
300 struct cl_io *io = ios->cis_io;
304 if (cl_io_is_trunc(io)) {
305 new_size = io->u.ci_setattr.sa_attr.lvb_size;
307 enqflags = CEF_DISCARD_DATA;
309 if ((io->u.ci_setattr.sa_attr.lvb_mtime >=
310 io->u.ci_setattr.sa_attr.lvb_ctime) ||
311 (io->u.ci_setattr.sa_attr.lvb_atime >=
312 io->u.ci_setattr.sa_attr.lvb_ctime))
316 cio->u.setattr.cui_local_lock = SETATTR_EXTENT_LOCK;
317 return ccc_io_one_lock(env, io, enqflags, CLM_WRITE,
318 new_size, OBD_OBJECT_EOF);
321 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
325 * Only ll_inode_size_lock is taken at this level. lov_stripe_lock()
326 * is grabbed by ll_truncate() only over call to obd_adjust_kms(). If
327 * vmtruncate returns 0, then ll_truncate dropped ll_inode_size_lock()
329 ll_inode_size_lock(inode, 0);
330 result = vmtruncate(inode, size);
332 ll_inode_size_unlock(inode, 0);
337 static int vvp_io_setattr_trunc(const struct lu_env *env,
338 const struct cl_io_slice *ios,
339 struct inode *inode, loff_t size)
341 struct vvp_io *vio = cl2vvp_io(env, ios);
342 struct cl_io *io = ios->cis_io;
343 struct cl_object *obj = ios->cis_obj;
344 pgoff_t start = cl_index(obj, size);
347 DOWN_WRITE_I_ALLOC_SEM(inode);
349 result = vvp_do_vmtruncate(inode, size);
352 * If a page is partially truncated, keep it owned across truncate to
355 * XXX this properly belongs to osc, because races in question are OST
358 if (cl_offset(obj, start) != size) {
359 struct cl_object_header *hdr;
361 hdr = cl_object_header(obj);
362 cfs_spin_lock(&hdr->coh_page_guard);
363 vio->cui_partpage = cl_page_lookup(hdr, start);
364 cfs_spin_unlock(&hdr->coh_page_guard);
366 if (vio->cui_partpage != NULL)
368 * Wait for the transfer completion for a partially
369 * truncated page to avoid dead-locking an OST with
370 * the concurrent page-wise overlapping WRITE and
371 * PUNCH requests. BUG:17397.
373 * Partial page is disowned in vvp_io_trunc_end().
375 cl_page_own(env, io, vio->cui_partpage);
377 vio->cui_partpage = NULL;
381 static int vvp_io_setattr_time(const struct lu_env *env,
382 const struct cl_io_slice *ios)
384 struct cl_io *io = ios->cis_io;
385 struct cl_object *obj = io->ci_obj;
386 struct cl_attr *attr = ccc_env_thread_attr(env);
388 unsigned valid = CAT_CTIME;
390 cl_object_attr_lock(obj);
391 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
392 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
393 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
396 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
397 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
400 result = cl_object_attr_set(env, obj, attr, valid);
401 cl_object_attr_unlock(obj);
406 static int vvp_io_setattr_start(const struct lu_env *env,
407 const struct cl_io_slice *ios)
409 struct ccc_io *cio = cl2ccc_io(env, ios);
410 struct cl_io *io = ios->cis_io;
411 struct inode *inode = ccc_object_inode(io->ci_obj);
413 LASSERT(cio->u.setattr.cui_locks_released);
415 LOCK_INODE_MUTEX(inode);
416 cio->u.setattr.cui_locks_released = 0;
418 if (cl_io_is_trunc(io))
419 return vvp_io_setattr_trunc(env, ios, inode,
420 io->u.ci_setattr.sa_attr.lvb_size);
422 return vvp_io_setattr_time(env, ios);
425 static void vvp_io_setattr_end(const struct lu_env *env,
426 const struct cl_io_slice *ios)
428 struct vvp_io *vio = cl2vvp_io(env, ios);
429 struct cl_io *io = ios->cis_io;
430 struct inode *inode = ccc_object_inode(io->ci_obj);
432 if (!cl_io_is_trunc(io))
434 if (vio->cui_partpage != NULL) {
435 cl_page_disown(env, ios->cis_io, vio->cui_partpage);
436 cl_page_put(env, vio->cui_partpage);
437 vio->cui_partpage = NULL;
441 * Do vmtruncate again, to remove possible stale pages populated by
442 * competing read threads. bz20645.
444 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
447 static void vvp_io_setattr_fini(const struct lu_env *env,
448 const struct cl_io_slice *ios)
450 struct ccc_io *cio = ccc_env_io(env);
451 struct cl_io *io = ios->cis_io;
452 struct inode *inode = ccc_object_inode(ios->cis_io->ci_obj);
454 if (cio->u.setattr.cui_locks_released) {
455 LOCK_INODE_MUTEX(inode);
456 if (cl_io_is_trunc(io))
457 DOWN_WRITE_I_ALLOC_SEM(inode);
458 cio->u.setattr.cui_locks_released = 0;
460 vvp_io_fini(env, ios);
463 #ifdef HAVE_FILE_READV
464 static ssize_t lustre_generic_file_read(struct file *file,
465 struct ccc_io *vio, loff_t *ppos)
467 return generic_file_readv(file, vio->cui_iov, vio->cui_nrsegs, ppos);
470 static ssize_t lustre_generic_file_write(struct file *file,
471 struct ccc_io *vio, loff_t *ppos)
473 return generic_file_writev(file, vio->cui_iov, vio->cui_nrsegs, ppos);
476 static ssize_t lustre_generic_file_read(struct file *file,
477 struct ccc_io *vio, loff_t *ppos)
479 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
480 vio->cui_nrsegs, *ppos);
483 static ssize_t lustre_generic_file_write(struct file *file,
484 struct ccc_io *vio, loff_t *ppos)
486 return generic_file_aio_write(vio->cui_iocb, vio->cui_iov,
487 vio->cui_nrsegs, *ppos);
491 static int vvp_io_read_start(const struct lu_env *env,
492 const struct cl_io_slice *ios)
494 struct vvp_io *vio = cl2vvp_io(env, ios);
495 struct ccc_io *cio = cl2ccc_io(env, ios);
496 struct cl_io *io = ios->cis_io;
497 struct cl_object *obj = io->ci_obj;
498 struct inode *inode = ccc_object_inode(obj);
499 struct ll_ra_read *bead = &vio->cui_bead;
500 struct file *file = cio->cui_fd->fd_file;
503 loff_t pos = io->u.ci_rd.rd.crw_pos;
504 long cnt = io->u.ci_rd.rd.crw_count;
505 long tot = cio->cui_tot_count;
508 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
510 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
512 result = ccc_prep_size(env, obj, io, pos, tot, 1, &exceed);
515 else if (exceed != 0)
518 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
519 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
520 inode->i_ino, cnt, pos, i_size_read(inode));
522 /* turn off the kernel's read-ahead */
523 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
525 /* initialize read-ahead window once per syscall */
526 if (!vio->cui_ra_window_set) {
527 vio->cui_ra_window_set = 1;
528 bead->lrr_start = cl_index(obj, pos);
530 * XXX: explicit CFS_PAGE_SIZE
532 bead->lrr_count = cl_index(obj, tot + CFS_PAGE_SIZE - 1);
533 ll_ra_read_in(file, bead);
538 switch (vio->cui_io_subtype) {
540 result = lustre_generic_file_read(file, cio, &pos);
542 #ifdef HAVE_KERNEL_SENDFILE
544 result = generic_file_sendfile(file, &pos, cnt,
545 vio->u.sendfile.cui_actor,
546 vio->u.sendfile.cui_target);
549 #ifdef HAVE_KERNEL_SPLICE_READ
551 result = generic_file_splice_read(file, &pos,
552 vio->u.splice.cui_pipe, cnt,
553 vio->u.splice.cui_flags);
557 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
565 io->ci_nob += result;
566 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
567 cio->cui_fd, pos, result, 0);
573 static int vvp_io_write_start(const struct lu_env *env,
574 const struct cl_io_slice *ios)
576 struct ccc_io *cio = cl2ccc_io(env, ios);
577 struct cl_io *io = ios->cis_io;
578 struct cl_object *obj = io->ci_obj;
579 struct inode *inode = ccc_object_inode(obj);
580 struct file *file = cio->cui_fd->fd_file;
582 loff_t pos = io->u.ci_wr.wr.crw_pos;
583 size_t cnt = io->u.ci_wr.wr.crw_count;
587 if (cl_io_is_append(io)) {
589 * PARALLEL IO This has to be changed for parallel IO doing
590 * out-of-order writes.
592 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
593 #ifndef HAVE_FILE_WRITEV
594 cio->cui_iocb->ki_pos = pos;
598 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
600 if (cio->cui_iov == NULL) /* from a temp io in ll_cl_init(). */
603 result = lustre_generic_file_write(file, cio, &pos);
608 io->ci_nob += result;
609 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
610 cio->cui_fd, pos, result, 0);
616 #ifndef HAVE_VM_OP_FAULT
617 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
621 vmpage = filemap_nopage(cfio->ft_vma, cfio->nopage.ft_address,
622 cfio->nopage.ft_type);
624 if (vmpage == NOPAGE_SIGBUS) {
625 CDEBUG(D_PAGE, "got addr %lu type %lx - SIGBUS\n",
626 cfio->nopage.ft_address,(long)cfio->nopage.ft_type);
628 } else if (vmpage == NOPAGE_OOM) {
629 CDEBUG(D_PAGE, "got addr %lu type %lx - OOM\n",
630 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
634 LL_CDEBUG_PAGE(D_PAGE, vmpage, "got addr %lu type %lx\n",
635 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
637 cfio->ft_vmpage = vmpage;
642 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
644 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, cfio->fault.ft_vmf);
646 if (cfio->fault.ft_vmf->page) {
647 LL_CDEBUG_PAGE(D_PAGE, cfio->fault.ft_vmf->page,
648 "got addr %p type NOPAGE\n",
649 cfio->fault.ft_vmf->virtual_address);
650 /*XXX workaround to bug in CLIO - he deadlocked with
651 lock cancel if page locked */
652 if (likely(cfio->fault.ft_flags & VM_FAULT_LOCKED)) {
653 unlock_page(cfio->fault.ft_vmf->page);
654 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
657 cfio->ft_vmpage = cfio->fault.ft_vmf->page;
661 if (unlikely (cfio->fault.ft_flags & VM_FAULT_ERROR)) {
662 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n",
663 cfio->fault.ft_vmf->virtual_address);
667 if (unlikely (cfio->fault.ft_flags & VM_FAULT_NOPAGE)) {
668 CDEBUG(D_PAGE, "got addr %p - OOM\n",
669 cfio->fault.ft_vmf->virtual_address);
673 CERROR("unknow error in page fault!\n");
679 static int vvp_io_fault_start(const struct lu_env *env,
680 const struct cl_io_slice *ios)
682 struct vvp_io *vio = cl2vvp_io(env, ios);
683 struct cl_io *io = ios->cis_io;
684 struct cl_object *obj = io->ci_obj;
685 struct inode *inode = ccc_object_inode(obj);
686 struct cl_fault_io *fio = &io->u.ci_fault;
687 struct vvp_fault_io *cfio = &vio->u.fault;
689 int kernel_result = 0;
691 struct cl_page *page;
693 pgoff_t last; /* last page in a file data region */
695 if (fio->ft_executable &&
696 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
698 " changed while waiting for the page fault lock\n",
699 PFID(lu_object_fid(&obj->co_lu)));
701 /* offset of the last byte on the page */
702 offset = cl_offset(obj, fio->ft_index + 1) - 1;
703 LASSERT(cl_index(obj, offset) == fio->ft_index);
704 result = ccc_prep_size(env, obj, io, 0, offset + 1, 0, NULL);
708 /* must return locked page */
709 kernel_result = vvp_io_kernel_fault(cfio);
710 if (kernel_result != 0)
711 return kernel_result;
712 /* Temporarily lock vmpage to keep cl_page_find() happy. */
713 lock_page(cfio->ft_vmpage);
714 page = cl_page_find(env, obj, fio->ft_index, cfio->ft_vmpage,
716 unlock_page(cfio->ft_vmpage);
718 page_cache_release(cfio->ft_vmpage);
719 cfio->ft_vmpage = NULL;
720 return PTR_ERR(page);
723 size = i_size_read(inode);
724 last = cl_index(obj, size - 1);
725 if (fio->ft_index == last)
727 * Last page is mapped partially.
729 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
731 fio->ft_nob = cl_page_size(obj);
733 lu_ref_add(&page->cp_reference, "fault", io);
736 * Certain 2.6 kernels return not-NULL from
737 * filemap_nopage() when page is beyond the file size,
738 * on the grounds that "An external ptracer can access
739 * pages that normally aren't accessible.." Don't
740 * propagate such page fault to the lower layers to
741 * avoid side-effects like KMS updates.
743 if (fio->ft_index > last)
749 static int vvp_io_read_page(const struct lu_env *env,
750 const struct cl_io_slice *ios,
751 const struct cl_page_slice *slice)
753 struct cl_io *io = ios->cis_io;
754 struct cl_object *obj = slice->cpl_obj;
755 struct ccc_page *cp = cl2ccc_page(slice);
756 struct cl_page *page = slice->cpl_page;
757 struct inode *inode = ccc_object_inode(obj);
758 struct ll_sb_info *sbi = ll_i2sbi(inode);
759 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
760 struct ll_readahead_state *ras = &fd->fd_ras;
761 cfs_page_t *vmpage = cp->cpg_page;
762 struct cl_2queue *queue = &io->ci_queue;
765 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
766 LASSERT(slice->cpl_obj == obj);
770 if (sbi->ll_ra_info.ra_max_pages_per_file)
771 ras_update(sbi, inode, ras, page->cp_index,
772 cp->cpg_defer_uptodate);
774 /* Sanity check whether the page is protected by a lock. */
775 rc = cl_page_is_under_lock(env, io, page);
777 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %d\n",
778 rc == -ENODATA ? "without a lock" :
784 if (cp->cpg_defer_uptodate) {
786 cl_page_export(env, page, 1);
789 * Add page into the queue even when it is marked uptodate above.
790 * this will unlock it automatically as part of cl_page_list_disown().
792 cl_2queue_add(queue, page);
793 if (sbi->ll_ra_info.ra_max_pages_per_file)
794 ll_readahead(env, io, ras,
795 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
800 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
801 struct cl_page *page, struct ccc_page *cp,
802 int to, enum cl_req_type crt)
804 struct cl_2queue *queue;
807 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
809 queue = &io->ci_queue;
811 cl_2queue_init_page(queue, page);
812 cl_page_clip(env, page, 0, to);
814 result = cl_io_submit_sync(env, io, crt, queue, CRP_NORMAL, 0);
815 LASSERT(cl_page_is_owned(page, io));
816 cl_page_clip(env, page, 0, CFS_PAGE_SIZE);
820 * in CRT_WRITE case page is left locked even in case of
823 cl_page_list_disown(env, io, &queue->c2_qin);
824 cl_2queue_fini(env, queue);
830 * Prepare partially written-to page for a write.
832 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
833 struct cl_object *obj, struct cl_page *pg,
835 unsigned from, unsigned to)
837 struct cl_attr *attr = ccc_env_thread_attr(env);
838 loff_t offset = cl_offset(obj, pg->cp_index);
841 cl_object_attr_lock(obj);
842 result = cl_object_attr_get(env, obj, attr);
843 cl_object_attr_unlock(obj);
846 * If are writing to a new page, no need to read old data.
847 * The extent locking will have updated the KMS, and for our
848 * purposes here we can treat it like i_size.
850 if (attr->cat_kms <= offset) {
851 char *kaddr = kmap_atomic(cp->cpg_page, KM_USER0);
853 memset(kaddr, 0, cl_page_size(obj));
854 kunmap_atomic(kaddr, KM_USER0);
855 } else if (cp->cpg_defer_uptodate)
858 result = vvp_page_sync_io(env, io, pg, cp,
859 CFS_PAGE_SIZE, CRT_READ);
861 * In older implementations, obdo_refresh_inode is called here
862 * to update the inode because the write might modify the
863 * object info at OST. However, this has been proven useless,
864 * since LVB functions will be called when user space program
865 * tries to retrieve inode attribute. Also, see bug 15909 for
869 cl_page_export(env, pg, 1);
874 static int vvp_io_prepare_write(const struct lu_env *env,
875 const struct cl_io_slice *ios,
876 const struct cl_page_slice *slice,
877 unsigned from, unsigned to)
879 struct cl_object *obj = slice->cpl_obj;
880 struct ccc_page *cp = cl2ccc_page(slice);
881 struct cl_page *pg = slice->cpl_page;
882 cfs_page_t *vmpage = cp->cpg_page;
888 LINVRNT(cl_page_is_vmlocked(env, pg));
889 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
893 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
894 if (!PageUptodate(vmpage)) {
896 * We're completely overwriting an existing page, so _don't_
897 * set it up to date until commit_write
899 if (from == 0 && to == CFS_PAGE_SIZE) {
900 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
901 POISON_PAGE(page, 0x11);
903 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
906 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
910 static int vvp_io_commit_write(const struct lu_env *env,
911 const struct cl_io_slice *ios,
912 const struct cl_page_slice *slice,
913 unsigned from, unsigned to)
915 struct cl_object *obj = slice->cpl_obj;
916 struct cl_io *io = ios->cis_io;
917 struct ccc_page *cp = cl2ccc_page(slice);
918 struct cl_page *pg = slice->cpl_page;
919 struct inode *inode = ccc_object_inode(obj);
920 struct ll_sb_info *sbi = ll_i2sbi(inode);
921 cfs_page_t *vmpage = cp->cpg_page;
929 LINVRNT(cl_page_is_vmlocked(env, pg));
930 LASSERT(vmpage->mapping->host == inode);
932 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n");
933 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
936 * queue a write for some time in the future the first time we
939 * This is different from what other file systems do: they usually
940 * just mark page (and some of its buffers) dirty and rely on
941 * balance_dirty_pages() to start a write-back. Lustre wants write-back
942 * to be started earlier for the following reasons:
944 * (1) with a large number of clients we need to limit the amount
945 * of cached data on the clients a lot;
947 * (2) large compute jobs generally want compute-only then io-only
948 * and the IO should complete as quickly as possible;
950 * (3) IO is batched up to the RPC size and is async until the
951 * client max cache is hit
952 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
955 if (!PageDirty(vmpage)) {
956 tallyop = LPROC_LL_DIRTY_MISSES;
957 vvp_write_pending(cl2ccc(obj), cp);
958 set_page_dirty(vmpage);
959 /* ll_set_page_dirty() does the same for now, but
960 * it will not soon. */
961 vvp_write_pending(cl2ccc(obj), cp);
962 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
963 if (result == -EDQUOT)
965 * Client ran out of disk space grant. Possible
968 * (a) do a sync write, renewing grant;
970 * (b) stop writing on this stripe, switch to the
973 * (b) is a part of "parallel io" design that is the
974 * ultimate goal. (a) is what "old" client did, and
975 * what the new code continues to do for the time
978 result = vvp_page_sync_io(env, io, pg, cp,
981 CERROR("Write page %lu of inode %p failed %d\n",
982 pg->cp_index, inode, result);
984 tallyop = LPROC_LL_DIRTY_HITS;
987 ll_stats_ops_tally(sbi, tallyop, 1);
989 size = cl_offset(obj, pg->cp_index) + to;
991 ll_inode_size_lock(inode, 0);
993 if (size > i_size_read(inode)) {
994 cl_isize_write_nolock(inode, size);
995 CDEBUG(D_VFSTRACE, DFID" updating i_size %lu\n",
996 PFID(lu_object_fid(&obj->co_lu)),
997 (unsigned long)size);
999 cl_page_export(env, pg, 1);
1001 if (size > i_size_read(inode))
1002 cl_page_discard(env, io, pg);
1004 ll_inode_size_unlock(inode, 0);
1009 static const struct cl_io_operations vvp_io_ops = {
1012 .cio_fini = vvp_io_fini,
1013 .cio_lock = vvp_io_read_lock,
1014 .cio_start = vvp_io_read_start,
1015 .cio_advance = ccc_io_advance
1018 .cio_fini = vvp_io_fini,
1019 .cio_lock = vvp_io_write_lock,
1020 .cio_start = vvp_io_write_start,
1021 .cio_advance = ccc_io_advance
1024 .cio_fini = vvp_io_setattr_fini,
1025 .cio_iter_init = vvp_io_setattr_iter_init,
1026 .cio_lock = vvp_io_setattr_lock,
1027 .cio_start = vvp_io_setattr_start,
1028 .cio_end = vvp_io_setattr_end
1031 .cio_fini = vvp_io_fault_fini,
1032 .cio_iter_init = vvp_io_fault_iter_init,
1033 .cio_lock = vvp_io_fault_lock,
1034 .cio_start = vvp_io_fault_start,
1035 .cio_end = ccc_io_end
1038 .cio_fini = vvp_io_fini
1041 .cio_read_page = vvp_io_read_page,
1042 .cio_prepare_write = vvp_io_prepare_write,
1043 .cio_commit_write = vvp_io_commit_write
1046 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1049 struct vvp_io *vio = vvp_env_io(env);
1050 struct ccc_io *cio = ccc_env_io(env);
1051 struct inode *inode = ccc_object_inode(obj);
1052 struct ll_sb_info *sbi = ll_i2sbi(inode);
1055 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
1058 CL_IO_SLICE_CLEAN(cio, cui_cl);
1059 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
1060 vio->cui_ra_window_set = 0;
1062 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1066 count = io->u.ci_rw.crw_count;
1067 op = io->ci_type == CIT_READ ?
1068 LPROC_LL_READ_BYTES : LPROC_LL_WRITE_BYTES;
1069 /* "If nbyte is 0, read() will return 0 and have no other
1070 * results." -- Single Unix Spec */
1074 cio->cui_tot_count = count;
1075 cio->cui_tot_nrsegs = 0;
1076 ll_stats_ops_tally(sbi, op, count);
1078 } else if (io->ci_type == CIT_SETATTR) {
1079 if (cl_io_is_trunc(io))
1080 /* lockless truncate? */
1081 ll_stats_ops_tally(sbi, LPROC_LL_TRUNC, 1);
1083 io->ci_lockreq = CILR_MANDATORY;
1088 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1089 const struct cl_io_slice *slice)
1091 /* Caling just for assertion */
1092 cl2ccc_io(env, slice);
1093 return vvp_env_io(env);