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 2008 Sun Microsystems, Inc. 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);
204 LASSERT(vvp_env_io(env)->cui_oneshot == 0);
207 ccc_io_update_iov(env, cio, io);
209 if (io->u.ci_rw.crw_nonblock)
210 ast_flags |= CEF_NONBLOCK;
211 result = vvp_mmap_locks(env, cio, io);
213 result = ccc_io_one_lock(env, io, ast_flags, mode, start, end);
217 static int vvp_io_read_lock(const struct lu_env *env,
218 const struct cl_io_slice *ios)
220 struct cl_io *io = ios->cis_io;
221 struct ll_inode_info *lli = ll_i2info(ccc_object_inode(io->ci_obj));
225 /* XXX: Layer violation, we shouldn't see lsm at llite level. */
226 if (lli->lli_smd != NULL) /* lsm-less file, don't need to lock */
227 result = vvp_io_rw_lock(env, io, CLM_READ,
228 io->u.ci_rd.rd.crw_pos,
229 io->u.ci_rd.rd.crw_pos +
230 io->u.ci_rd.rd.crw_count - 1);
236 static int vvp_io_fault_lock(const struct lu_env *env,
237 const struct cl_io_slice *ios)
239 struct cl_io *io = ios->cis_io;
240 struct vvp_io *vio = cl2vvp_io(env, ios);
242 * XXX LDLM_FL_CBPENDING
244 return ccc_io_one_lock_index
245 (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma),
246 io->u.ci_fault.ft_index, io->u.ci_fault.ft_index);
249 static int vvp_io_write_lock(const struct lu_env *env,
250 const struct cl_io_slice *ios)
252 struct cl_io *io = ios->cis_io;
257 if (cl2vvp_io(env, ios)->cui_oneshot == 0) {
258 if (io->u.ci_wr.wr_append) {
260 end = OBD_OBJECT_EOF;
262 start = io->u.ci_wr.wr.crw_pos;
263 end = start + io->u.ci_wr.wr.crw_count - 1;
265 result = vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
271 static int vvp_io_trunc_iter_init(const struct lu_env *env,
272 const struct cl_io_slice *ios)
274 struct ccc_io *vio = cl2ccc_io(env, ios);
275 struct inode *inode = ccc_object_inode(ios->cis_obj);
278 * We really need to get our PW lock before we change inode->i_size.
279 * If we don't we can race with other i_size updaters on our node,
280 * like ll_file_read. We can also race with i_size propogation to
281 * other nodes through dirtying and writeback of final cached pages.
282 * This last one is especially bad for racing o_append users on other
286 UNLOCK_INODE_MUTEX(inode);
287 UP_WRITE_I_ALLOC_SEM(inode);
288 vio->u.trunc.cui_locks_released = 1;
293 * Implementation of cl_io_operations::cio_lock() method for CIT_TRUNC io.
295 * Handles "lockless io" mode when extent locking is done by server.
297 static int vvp_io_trunc_lock(const struct lu_env *env,
298 const struct cl_io_slice *ios)
300 struct ccc_io *vio = cl2ccc_io(env, ios);
301 struct cl_io *io = ios->cis_io;
302 size_t new_size = io->u.ci_truncate.tr_size;
303 __u32 enqflags = new_size == 0 ? CEF_DISCARD_DATA : 0;
306 vio->u.trunc.cui_local_lock = TRUNC_EXTENT;
307 result = ccc_io_one_lock(env, io, enqflags, CLM_WRITE,
308 new_size, OBD_OBJECT_EOF);
312 static int vvp_io_trunc_start(const struct lu_env *env,
313 const struct cl_io_slice *ios)
315 struct ccc_io *cio = cl2ccc_io(env, ios);
316 struct vvp_io *vio = cl2vvp_io(env, ios);
317 struct cl_io *io = ios->cis_io;
318 struct inode *inode = ccc_object_inode(io->ci_obj);
319 struct cl_object *obj = ios->cis_obj;
320 size_t size = io->u.ci_truncate.tr_size;
321 pgoff_t start = cl_index(obj, size);
324 LASSERT(cio->u.trunc.cui_locks_released);
325 LASSERT(vio->cui_oneshot == 0);
327 LOCK_INODE_MUTEX(inode);
328 DOWN_WRITE_I_ALLOC_SEM(inode);
329 cio->u.trunc.cui_locks_released = 0;
332 * Only ll_inode_size_lock is taken at this level. lov_stripe_lock()
333 * is grabbed by ll_truncate() only over call to obd_adjust_kms(). If
334 * vmtruncate returns 0, then ll_truncate dropped ll_inode_size_lock()
336 ll_inode_size_lock(inode, 0);
337 result = vmtruncate(inode, size);
339 ll_inode_size_unlock(inode, 0);
341 * If a page is partially truncated, keep it owned across truncate to
344 * XXX this properly belongs to osc, because races in question are OST
347 if (cl_offset(obj, start) != size) {
348 struct cl_object_header *hdr;
350 hdr = cl_object_header(obj);
351 spin_lock(&hdr->coh_page_guard);
352 vio->cui_partpage = cl_page_lookup(hdr, start);
353 spin_unlock(&hdr->coh_page_guard);
355 if (vio->cui_partpage != NULL)
357 * Wait for the transfer completion for a partially
358 * truncated page to avoid dead-locking an OST with
359 * the concurrent page-wise overlapping WRITE and
360 * PUNCH requests. BUG:17397.
362 * Partial page is disowned in vvp_io_trunc_end().
364 cl_page_own(env, io, vio->cui_partpage);
366 vio->cui_partpage = NULL;
370 static void vvp_io_trunc_end(const struct lu_env *env,
371 const struct cl_io_slice *ios)
373 struct vvp_io *vio = cl2vvp_io(env, ios);
375 if (vio->cui_partpage != NULL) {
376 cl_page_disown(env, ios->cis_io, vio->cui_partpage);
377 cl_page_put(env, vio->cui_partpage);
378 vio->cui_partpage = NULL;
382 static void vvp_io_trunc_fini(const struct lu_env *env,
383 const struct cl_io_slice *ios)
385 struct ccc_io *cio = ccc_env_io(env);
386 struct inode *inode = ccc_object_inode(ios->cis_io->ci_obj);
388 if (cio->u.trunc.cui_locks_released) {
389 LOCK_INODE_MUTEX(inode);
390 DOWN_WRITE_I_ALLOC_SEM(inode);
391 cio->u.trunc.cui_locks_released = 0;
393 vvp_io_fini(env, ios);
396 #ifdef HAVE_FILE_READV
397 static ssize_t lustre_generic_file_read(struct file *file,
398 struct ccc_io *vio, loff_t *ppos)
400 return generic_file_readv(file, vio->cui_iov, vio->cui_nrsegs, ppos);
403 static ssize_t lustre_generic_file_write(struct file *file,
404 struct ccc_io *vio, loff_t *ppos)
406 return generic_file_writev(file, vio->cui_iov, vio->cui_nrsegs, ppos);
409 static ssize_t lustre_generic_file_read(struct file *file,
410 struct ccc_io *vio, loff_t *ppos)
412 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
413 vio->cui_nrsegs, *ppos);
416 static ssize_t lustre_generic_file_write(struct file *file,
417 struct ccc_io *vio, loff_t *ppos)
419 return generic_file_aio_write(vio->cui_iocb, vio->cui_iov,
420 vio->cui_nrsegs, *ppos);
424 static int vvp_io_read_start(const struct lu_env *env,
425 const struct cl_io_slice *ios)
427 struct vvp_io *vio = cl2vvp_io(env, ios);
428 struct ccc_io *cio = cl2ccc_io(env, ios);
429 struct cl_io *io = ios->cis_io;
430 struct cl_object *obj = io->ci_obj;
431 struct inode *inode = ccc_object_inode(obj);
432 struct ll_ra_read *bead = &vio->cui_bead;
433 struct file *file = cio->cui_fd->fd_file;
436 loff_t pos = io->u.ci_rd.rd.crw_pos;
437 long cnt = io->u.ci_rd.rd.crw_count;
438 long tot = cio->cui_tot_count;
441 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
442 LASSERT(vio->cui_oneshot == 0);
444 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
446 result = ccc_prep_size(env, obj, io, pos, tot, 1, &exceed);
449 else if (exceed != 0)
452 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
453 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
454 inode->i_ino, cnt, pos, i_size_read(inode));
456 /* turn off the kernel's read-ahead */
457 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
459 /* initialize read-ahead window once per syscall */
460 if (!vio->cui_ra_window_set) {
461 vio->cui_ra_window_set = 1;
462 bead->lrr_start = cl_index(obj, pos);
464 * XXX: explicit CFS_PAGE_SIZE
466 bead->lrr_count = cl_index(obj, tot + CFS_PAGE_SIZE - 1);
467 ll_ra_read_in(file, bead);
472 switch (vio->cui_io_subtype) {
474 result = lustre_generic_file_read(file, cio, &pos);
476 #ifdef HAVE_KERNEL_SENDFILE
478 result = generic_file_sendfile(file, &pos, cnt,
479 vio->u.sendfile.cui_actor,
480 vio->u.sendfile.cui_target);
483 #ifdef HAVE_KERNEL_SPLICE_READ
485 result = generic_file_splice_read(file, &pos,
486 vio->u.splice.cui_pipe, cnt,
487 vio->u.splice.cui_flags);
491 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
499 io->ci_nob += result;
500 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
501 cio->cui_fd, pos, result, 0);
507 static int vvp_io_write_start(const struct lu_env *env,
508 const struct cl_io_slice *ios)
510 struct ccc_io *cio = cl2ccc_io(env, ios);
511 struct cl_io *io = ios->cis_io;
512 struct cl_object *obj = io->ci_obj;
513 struct inode *inode = ccc_object_inode(obj);
514 struct file *file = cio->cui_fd->fd_file;
516 loff_t pos = io->u.ci_wr.wr.crw_pos;
517 size_t cnt = io->u.ci_wr.wr.crw_count;
521 if (cl_io_is_append(io)) {
523 * PARALLEL IO This has to be changed for parallel IO doing
524 * out-of-order writes.
526 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
527 #ifndef HAVE_FILE_WRITEV
528 cio->cui_iocb->ki_pos = pos;
532 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
534 if (cl2vvp_io(env, ios)->cui_oneshot > 0)
537 result = lustre_generic_file_write(file, cio, &pos);
542 io->ci_nob += result;
543 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
544 cio->cui_fd, pos, result, 0);
550 #ifndef HAVE_VM_OP_FAULT
551 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
555 vmpage = filemap_nopage(cfio->ft_vma, cfio->nopage.ft_address,
556 cfio->nopage.ft_type);
558 if (vmpage == NOPAGE_SIGBUS) {
559 CDEBUG(D_PAGE, "got addr %lu type %lx - SIGBUS\n",
560 cfio->nopage.ft_address,(long)cfio->nopage.ft_type);
562 } else if (vmpage == NOPAGE_OOM) {
563 CDEBUG(D_PAGE, "got addr %lu type %lx - OOM\n",
564 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
568 LL_CDEBUG_PAGE(D_PAGE, vmpage, "got addr %lu type %lx\n",
569 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
571 cfio->ft_vmpage = vmpage;
576 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
578 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, cfio->fault.ft_vmf);
580 if (cfio->fault.ft_vmf->page) {
581 LL_CDEBUG_PAGE(D_PAGE, cfio->fault.ft_vmf->page,
582 "got addr %p type NOPAGE\n",
583 cfio->fault.ft_vmf->virtual_address);
584 /*XXX workaround to bug in CLIO - he deadlocked with
585 lock cancel if page locked */
586 if (likely(cfio->fault.ft_flags & VM_FAULT_LOCKED)) {
587 unlock_page(cfio->fault.ft_vmf->page);
588 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
591 cfio->ft_vmpage = cfio->fault.ft_vmf->page;
595 if (unlikely (cfio->fault.ft_flags & VM_FAULT_ERROR)) {
596 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n",
597 cfio->fault.ft_vmf->virtual_address);
601 if (unlikely (cfio->fault.ft_flags & VM_FAULT_NOPAGE)) {
602 CDEBUG(D_PAGE, "got addr %p - OOM\n",
603 cfio->fault.ft_vmf->virtual_address);
607 CERROR("unknow error in page fault!\n");
613 static int vvp_io_fault_start(const struct lu_env *env,
614 const struct cl_io_slice *ios)
616 struct vvp_io *vio = cl2vvp_io(env, ios);
617 struct cl_io *io = ios->cis_io;
618 struct cl_object *obj = io->ci_obj;
619 struct inode *inode = ccc_object_inode(obj);
620 struct cl_fault_io *fio = &io->u.ci_fault;
621 struct vvp_fault_io *cfio = &vio->u.fault;
623 int kernel_result = 0;
625 struct cl_page *page;
627 pgoff_t last; /* last page in a file data region */
629 LASSERT(vio->cui_oneshot == 0);
631 if (fio->ft_executable &&
632 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
634 " changed while waiting for the page fault lock\n",
635 PFID(lu_object_fid(&obj->co_lu)));
637 /* offset of the last byte on the page */
638 offset = cl_offset(obj, fio->ft_index + 1) - 1;
639 LASSERT(cl_index(obj, offset) == fio->ft_index);
640 result = ccc_prep_size(env, obj, io, 0, offset + 1, 0, NULL);
644 /* must return locked page */
645 kernel_result = vvp_io_kernel_fault(cfio);
646 if (kernel_result != 0)
647 return kernel_result;
648 /* Temporarily lock vmpage to keep cl_page_find() happy. */
649 lock_page(cfio->ft_vmpage);
650 page = cl_page_find(env, obj, fio->ft_index, cfio->ft_vmpage,
652 unlock_page(cfio->ft_vmpage);
654 page_cache_release(cfio->ft_vmpage);
655 cfio->ft_vmpage = NULL;
656 return PTR_ERR(page);
659 size = i_size_read(inode);
660 last = cl_index(obj, size - 1);
661 if (fio->ft_index == last)
663 * Last page is mapped partially.
665 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
667 fio->ft_nob = cl_page_size(obj);
669 lu_ref_add(&page->cp_reference, "fault", io);
672 * Certain 2.6 kernels return not-NULL from
673 * filemap_nopage() when page is beyond the file size,
674 * on the grounds that "An external ptracer can access
675 * pages that normally aren't accessible.." Don't
676 * propagate such page fault to the lower layers to
677 * avoid side-effects like KMS updates.
679 if (fio->ft_index > last)
685 static int vvp_io_read_page(const struct lu_env *env,
686 const struct cl_io_slice *ios,
687 const struct cl_page_slice *slice)
689 struct cl_io *io = ios->cis_io;
690 struct cl_object *obj = slice->cpl_obj;
691 struct ccc_page *cp = cl2ccc_page(slice);
692 struct cl_page *page = slice->cpl_page;
693 struct inode *inode = ccc_object_inode(obj);
694 struct ll_sb_info *sbi = ll_i2sbi(inode);
695 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
696 struct ll_readahead_state *ras = &fd->fd_ras;
697 cfs_page_t *vmpage = cp->cpg_page;
698 struct cl_2queue *queue = &io->ci_queue;
701 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
702 LASSERT(cl2vvp_io(env, ios)->cui_oneshot == 0);
703 LASSERT(slice->cpl_obj == obj);
707 if (sbi->ll_ra_info.ra_max_pages_per_file)
708 ras_update(sbi, inode, ras, page->cp_index,
709 cp->cpg_defer_uptodate);
711 /* Sanity check whether the page is protected by a lock. */
712 rc = cl_page_is_under_lock(env, io, page);
714 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %i\n",
715 rc == -ENODATA ? "without a lock" :
721 if (cp->cpg_defer_uptodate) {
723 cl_page_export(env, page, 1);
726 * Add page into the queue even when it is marked uptodate above.
727 * this will unlock it automatically as part of cl_page_list_disown().
729 cl_2queue_add(queue, page);
730 if (sbi->ll_ra_info.ra_max_pages_per_file)
731 ll_readahead(env, io, ras,
732 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
737 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
738 struct cl_page *page, struct ccc_page *cp,
739 int to, enum cl_req_type crt)
741 struct cl_2queue *queue;
744 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
746 queue = &io->ci_queue;
748 cl_2queue_init_page(queue, page);
749 cl_page_clip(env, page, 0, to);
751 result = cl_io_submit_sync(env, io, crt, queue, CRP_NORMAL, 0);
752 LASSERT(cl_page_is_owned(page, io));
753 cl_page_clip(env, page, 0, CFS_PAGE_SIZE);
757 * in CRT_WRITE case page is left locked even in case of
760 cl_page_list_disown(env, io, &queue->c2_qin);
761 cl_2queue_fini(env, queue);
767 * Prepare partially written-to page for a write.
769 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
770 struct cl_object *obj, struct cl_page *pg,
772 unsigned from, unsigned to)
774 struct cl_attr *attr = ccc_env_thread_attr(env);
775 loff_t offset = cl_offset(obj, pg->cp_index);
778 cl_object_attr_lock(obj);
779 result = cl_object_attr_get(env, obj, attr);
780 cl_object_attr_unlock(obj);
783 * If are writing to a new page, no need to read old data.
784 * The extent locking will have updated the KMS, and for our
785 * purposes here we can treat it like i_size.
787 if (attr->cat_kms <= offset) {
788 char *kaddr = kmap_atomic(cp->cpg_page, KM_USER0);
790 memset(kaddr, 0, cl_page_size(obj));
791 kunmap_atomic(kaddr, KM_USER0);
792 } else if (cp->cpg_defer_uptodate)
795 result = vvp_page_sync_io(env, io, pg, cp,
796 CFS_PAGE_SIZE, CRT_READ);
798 * In older implementations, obdo_refresh_inode is called here
799 * to update the inode because the write might modify the
800 * object info at OST. However, this has been proven useless,
801 * since LVB functions will be called when user space program
802 * tries to retrieve inode attribute. Also, see bug 15909 for
806 cl_page_export(env, pg, 1);
811 static int vvp_io_prepare_write(const struct lu_env *env,
812 const struct cl_io_slice *ios,
813 const struct cl_page_slice *slice,
814 unsigned from, unsigned to)
816 struct cl_object *obj = slice->cpl_obj;
817 struct ccc_page *cp = cl2ccc_page(slice);
818 struct cl_page *pg = slice->cpl_page;
819 cfs_page_t *vmpage = cp->cpg_page;
825 LINVRNT(cl_page_is_vmlocked(env, pg));
826 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
830 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
831 if (!PageUptodate(vmpage)) {
833 * We're completely overwriting an existing page, so _don't_
834 * set it up to date until commit_write
836 if (from == 0 && to == CFS_PAGE_SIZE) {
837 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
838 POISON_PAGE(page, 0x11);
840 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
843 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
847 static int vvp_io_commit_write(const struct lu_env *env,
848 const struct cl_io_slice *ios,
849 const struct cl_page_slice *slice,
850 unsigned from, unsigned to)
852 struct cl_object *obj = slice->cpl_obj;
853 struct cl_io *io = ios->cis_io;
854 struct ccc_page *cp = cl2ccc_page(slice);
855 struct cl_page *pg = slice->cpl_page;
856 struct inode *inode = ccc_object_inode(obj);
857 struct ll_sb_info *sbi = ll_i2sbi(inode);
858 cfs_page_t *vmpage = cp->cpg_page;
866 LINVRNT(cl_page_is_vmlocked(env, pg));
867 LASSERT(vmpage->mapping->host == inode);
869 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n");
870 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
873 * queue a write for some time in the future the first time we
876 * This is different from what other file systems do: they usually
877 * just mark page (and some of its buffers) dirty and rely on
878 * balance_dirty_pages() to start a write-back. Lustre wants write-back
879 * to be started earlier for the following reasons:
881 * (1) with a large number of clients we need to limit the amount
882 * of cached data on the clients a lot;
884 * (2) large compute jobs generally want compute-only then io-only
885 * and the IO should complete as quickly as possible;
887 * (3) IO is batched up to the RPC size and is async until the
888 * client max cache is hit
889 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
892 if (!PageDirty(vmpage)) {
893 tallyop = LPROC_LL_DIRTY_MISSES;
894 vvp_write_pending(cl2ccc(obj), cp);
895 set_page_dirty(vmpage);
896 /* ll_set_page_dirty() does the same for now, but
897 * it will not soon. */
898 vvp_write_pending(cl2ccc(obj), cp);
899 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
900 if (result == -EDQUOT)
902 * Client ran out of disk space grant. Possible
905 * (a) do a sync write, renewing grant;
907 * (b) stop writing on this stripe, switch to the
910 * (b) is a part of "parallel io" design that is the
911 * ultimate goal. (a) is what "old" client did, and
912 * what the new code continues to do for the time
915 result = vvp_page_sync_io(env, io, pg, cp,
918 CERROR("Write page %lu of inode %p failed %d\n",
919 pg->cp_index, inode, result);
921 tallyop = LPROC_LL_DIRTY_HITS;
924 ll_stats_ops_tally(sbi, tallyop, 1);
926 size = cl_offset(obj, pg->cp_index) + to;
929 if (size > i_size_read(inode))
930 i_size_write(inode, size);
931 cl_page_export(env, pg, 1);
932 } else if (size > i_size_read(inode))
933 cl_page_discard(env, io, pg);
937 static const struct cl_io_operations vvp_io_ops = {
940 .cio_fini = vvp_io_fini,
941 .cio_lock = vvp_io_read_lock,
942 .cio_start = vvp_io_read_start,
943 .cio_advance = ccc_io_advance
946 .cio_fini = vvp_io_fini,
947 .cio_lock = vvp_io_write_lock,
948 .cio_start = vvp_io_write_start,
949 .cio_advance = ccc_io_advance
952 .cio_fini = vvp_io_trunc_fini,
953 .cio_iter_init = vvp_io_trunc_iter_init,
954 .cio_lock = vvp_io_trunc_lock,
955 .cio_start = vvp_io_trunc_start,
956 .cio_end = vvp_io_trunc_end
959 .cio_fini = vvp_io_fault_fini,
960 .cio_iter_init = vvp_io_fault_iter_init,
961 .cio_lock = vvp_io_fault_lock,
962 .cio_start = vvp_io_fault_start,
963 .cio_end = ccc_io_end
966 .cio_fini = vvp_io_fini
969 .cio_read_page = vvp_io_read_page,
970 .cio_prepare_write = vvp_io_prepare_write,
971 .cio_commit_write = vvp_io_commit_write
974 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
977 struct vvp_io *vio = vvp_env_io(env);
978 struct ccc_io *cio = ccc_env_io(env);
979 struct inode *inode = ccc_object_inode(obj);
980 struct ll_sb_info *sbi = ll_i2sbi(inode);
983 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
986 CL_IO_SLICE_CLEAN(cio, cui_cl);
987 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
988 vio->cui_oneshot = 0;
989 vio->cui_ra_window_set = 0;
991 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
995 count = io->u.ci_rw.crw_count;
996 op = io->ci_type == CIT_READ ?
997 LPROC_LL_READ_BYTES : LPROC_LL_WRITE_BYTES;
998 /* "If nbyte is 0, read() will return 0 and have no other
999 * results." -- Single Unix Spec */
1003 cio->cui_tot_count = count;
1004 cio->cui_tot_nrsegs = 0;
1005 ll_stats_ops_tally(sbi, op, count);
1007 } else if (io->ci_type == CIT_TRUNC) {
1008 /* lockless truncate? */
1009 ll_stats_ops_tally(sbi, LPROC_LL_TRUNC, 1);
1014 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1015 const struct cl_io_slice *slice)
1017 /* Caling just for assertion */
1018 cl2ccc_io(env, slice);
1019 return vvp_env_io(env);