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_WRITE)
92 up(&ll_i2info(ccc_object_inode(obj))->lli_write_sem);
94 struct vvp_io *vio = cl2vvp_io(env, ios);
95 struct ccc_io *cio = cl2ccc_io(env, ios);
97 if (vio->cui_ra_window_set)
98 ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
103 static void vvp_io_fault_fini(const struct lu_env *env,
104 const struct cl_io_slice *ios)
106 struct cl_io *io = ios->cis_io;
107 struct cl_page *page = io->u.ci_fault.ft_page;
109 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
112 lu_ref_del(&page->cp_reference, "fault", io);
113 cl_page_put(env, page);
114 io->u.ci_fault.ft_page = NULL;
116 vvp_io_fini(env, ios);
119 enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
122 * we only want to hold PW locks if the mmap() can generate
123 * writes back to the file and that only happens in shared
126 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
131 static int vvp_mmap_locks(const struct lu_env *env,
132 struct ccc_io *vio, struct cl_io *io)
134 struct ccc_thread_info *cti = ccc_env_info(env);
135 struct vm_area_struct *vma;
136 struct cl_lock_descr *descr = &cti->cti_descr;
137 ldlm_policy_data_t policy;
144 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
146 if (!cl_is_normalio(env, io))
149 for (seg = 0; seg < vio->cui_nrsegs; seg++) {
150 const struct iovec *iv = &vio->cui_iov[seg];
152 addr = (unsigned long)iv->iov_base;
157 count += addr & (~CFS_PAGE_MASK);
158 addr &= CFS_PAGE_MASK;
159 while((vma = our_vma(addr, count)) != NULL) {
160 struct inode *inode = vma->vm_file->f_dentry->d_inode;
161 int flags = CEF_MUST;
163 if (ll_file_nolock(vma->vm_file)) {
165 * For no lock case, a lockless lock will be
172 * XXX: Required lock mode can be weakened: CIT_WRITE
173 * io only ever reads user level buffer, and CIT_READ
176 policy_from_vma(&policy, vma, addr, count);
177 descr->cld_mode = vvp_mode_from_vma(vma);
178 descr->cld_obj = ll_i2info(inode)->lli_clob;
179 descr->cld_start = cl_index(descr->cld_obj,
180 policy.l_extent.start);
181 descr->cld_end = cl_index(descr->cld_obj,
182 policy.l_extent.end);
183 result = cl_io_lock_alloc_add(env, io, descr, flags);
187 if (vma->vm_end - addr >= count)
190 count -= vma->vm_end - addr;
197 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
198 enum cl_lock_mode mode, loff_t start, loff_t end)
200 struct ccc_io *cio = ccc_env_io(env);
204 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
205 LASSERT(vvp_env_io(env)->cui_oneshot == 0);
208 ccc_io_update_iov(env, cio, io);
210 if (io->u.ci_rw.crw_nonblock)
211 ast_flags |= CEF_NONBLOCK;
212 result = vvp_mmap_locks(env, cio, io);
214 result = ccc_io_one_lock(env, io, ast_flags, mode, start, end);
218 static int vvp_io_read_lock(const struct lu_env *env,
219 const struct cl_io_slice *ios)
221 struct cl_io *io = ios->cis_io;
222 struct ll_inode_info *lli = ll_i2info(ccc_object_inode(io->ci_obj));
226 /* XXX: Layer violation, we shouldn't see lsm at llite level. */
227 if (lli->lli_smd != NULL) /* lsm-less file, don't need to lock */
228 result = vvp_io_rw_lock(env, io, CLM_READ,
229 io->u.ci_rd.rd.crw_pos,
230 io->u.ci_rd.rd.crw_pos +
231 io->u.ci_rd.rd.crw_count - 1);
237 static int vvp_io_fault_lock(const struct lu_env *env,
238 const struct cl_io_slice *ios)
240 struct cl_io *io = ios->cis_io;
241 struct vvp_io *vio = cl2vvp_io(env, ios);
243 * XXX LDLM_FL_CBPENDING
245 return ccc_io_one_lock_index
246 (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma),
247 io->u.ci_fault.ft_index, io->u.ci_fault.ft_index);
250 static int vvp_io_write_lock(const struct lu_env *env,
251 const struct cl_io_slice *ios)
253 struct cl_io *io = ios->cis_io;
258 if (cl2vvp_io(env, ios)->cui_oneshot == 0) {
259 if (io->u.ci_wr.wr_append) {
261 end = OBD_OBJECT_EOF;
263 start = io->u.ci_wr.wr.crw_pos;
264 end = start + io->u.ci_wr.wr.crw_count - 1;
266 result = vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
272 static int vvp_io_trunc_iter_init(const struct lu_env *env,
273 const struct cl_io_slice *ios)
275 struct ccc_io *vio = cl2ccc_io(env, ios);
276 struct inode *inode = ccc_object_inode(ios->cis_obj);
279 * We really need to get our PW lock before we change inode->i_size.
280 * If we don't we can race with other i_size updaters on our node,
281 * like ll_file_read. We can also race with i_size propogation to
282 * other nodes through dirtying and writeback of final cached pages.
283 * This last one is especially bad for racing o_append users on other
287 UNLOCK_INODE_MUTEX(inode);
288 UP_WRITE_I_ALLOC_SEM(inode);
289 vio->u.trunc.cui_locks_released = 1;
294 * Implementation of cl_io_operations::cio_lock() method for CIT_TRUNC io.
296 * Handles "lockless io" mode when extent locking is done by server.
298 static int vvp_io_trunc_lock(const struct lu_env *env,
299 const struct cl_io_slice *ios)
301 struct ccc_io *vio = cl2ccc_io(env, ios);
302 struct cl_io *io = ios->cis_io;
303 size_t new_size = io->u.ci_truncate.tr_size;
304 __u32 enqflags = new_size == 0 ? CEF_DISCARD_DATA : 0;
307 vio->u.trunc.cui_local_lock = TRUNC_EXTENT;
308 result = ccc_io_one_lock(env, io, enqflags, CLM_WRITE,
309 new_size, OBD_OBJECT_EOF);
313 static int vvp_io_trunc_start(const struct lu_env *env,
314 const struct cl_io_slice *ios)
316 struct ccc_io *cio = cl2ccc_io(env, ios);
317 struct vvp_io *vio = cl2vvp_io(env, ios);
318 struct cl_io *io = ios->cis_io;
319 struct inode *inode = ccc_object_inode(io->ci_obj);
320 struct cl_object *obj = ios->cis_obj;
321 size_t size = io->u.ci_truncate.tr_size;
322 pgoff_t start = cl_index(obj, size);
325 LASSERT(cio->u.trunc.cui_locks_released);
326 LASSERT(vio->cui_oneshot == 0);
328 LOCK_INODE_MUTEX(inode);
329 DOWN_WRITE_I_ALLOC_SEM(inode);
330 cio->u.trunc.cui_locks_released = 0;
333 * Only ll_inode_size_lock is taken at this level. lov_stripe_lock()
334 * is grabbed by ll_truncate() only over call to obd_adjust_kms(). If
335 * vmtruncate returns 0, then ll_truncate dropped ll_inode_size_lock()
337 ll_inode_size_lock(inode, 0);
338 result = vmtruncate(inode, size);
340 ll_inode_size_unlock(inode, 0);
342 * If a page is partially truncated, keep it owned across truncate to
345 * XXX this properly belongs to osc, because races in question are OST
348 if (cl_offset(obj, start) != size) {
349 struct cl_object_header *hdr;
351 hdr = cl_object_header(obj);
352 spin_lock(&hdr->coh_page_guard);
353 vio->cui_partpage = cl_page_lookup(hdr, start);
354 spin_unlock(&hdr->coh_page_guard);
356 if (vio->cui_partpage != NULL)
358 * Wait for the transfer completion for a partially
359 * truncated page to avoid dead-locking an OST with
360 * the concurrent page-wise overlapping WRITE and
361 * PUNCH requests. BUG:17397.
363 * Partial page is disowned in vvp_io_trunc_end().
365 cl_page_own(env, io, vio->cui_partpage);
367 vio->cui_partpage = NULL;
371 static void vvp_io_trunc_end(const struct lu_env *env,
372 const struct cl_io_slice *ios)
374 struct vvp_io *vio = cl2vvp_io(env, ios);
376 if (vio->cui_partpage != NULL) {
377 cl_page_disown(env, ios->cis_io, vio->cui_partpage);
378 cl_page_put(env, vio->cui_partpage);
379 vio->cui_partpage = NULL;
383 static void vvp_io_trunc_fini(const struct lu_env *env,
384 const struct cl_io_slice *ios)
386 struct ccc_io *cio = ccc_env_io(env);
387 struct inode *inode = ccc_object_inode(ios->cis_io->ci_obj);
389 if (cio->u.trunc.cui_locks_released) {
390 LOCK_INODE_MUTEX(inode);
391 DOWN_WRITE_I_ALLOC_SEM(inode);
392 cio->u.trunc.cui_locks_released = 0;
394 vvp_io_fini(env, ios);
397 #ifdef HAVE_FILE_READV
398 static ssize_t lustre_generic_file_read(struct file *file,
399 struct ccc_io *vio, loff_t *ppos)
401 return generic_file_readv(file, vio->cui_iov, vio->cui_nrsegs, ppos);
404 static ssize_t lustre_generic_file_write(struct file *file,
405 struct ccc_io *vio, loff_t *ppos)
407 return generic_file_writev(file, vio->cui_iov, vio->cui_nrsegs, ppos);
410 static ssize_t lustre_generic_file_read(struct file *file,
411 struct ccc_io *vio, loff_t *ppos)
413 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
414 vio->cui_nrsegs, *ppos);
417 static ssize_t lustre_generic_file_write(struct file *file,
418 struct ccc_io *vio, loff_t *ppos)
420 return generic_file_aio_write(vio->cui_iocb, vio->cui_iov,
421 vio->cui_nrsegs, *ppos);
425 static int vvp_io_read_start(const struct lu_env *env,
426 const struct cl_io_slice *ios)
428 struct vvp_io *vio = cl2vvp_io(env, ios);
429 struct ccc_io *cio = cl2ccc_io(env, ios);
430 struct cl_io *io = ios->cis_io;
431 struct cl_object *obj = io->ci_obj;
432 struct inode *inode = ccc_object_inode(obj);
433 struct ll_ra_read *bead = &vio->cui_bead;
434 struct file *file = cio->cui_fd->fd_file;
437 loff_t pos = io->u.ci_rd.rd.crw_pos;
438 long cnt = io->u.ci_rd.rd.crw_count;
439 long tot = cio->cui_tot_count;
442 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
443 LASSERT(vio->cui_oneshot == 0);
445 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
447 result = ccc_prep_size(env, obj, io, pos, tot, 1, &exceed);
450 else if (exceed != 0)
453 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
454 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
455 inode->i_ino, cnt, pos, i_size_read(inode));
457 /* turn off the kernel's read-ahead */
458 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
460 /* initialize read-ahead window once per syscall */
461 if (!vio->cui_ra_window_set) {
462 vio->cui_ra_window_set = 1;
463 bead->lrr_start = cl_index(obj, pos);
465 * XXX: explicit CFS_PAGE_SIZE
467 bead->lrr_count = cl_index(obj, tot + CFS_PAGE_SIZE - 1);
468 ll_ra_read_in(file, bead);
473 switch (vio->cui_io_subtype) {
475 result = lustre_generic_file_read(file, cio, &pos);
477 #ifdef HAVE_KERNEL_SENDFILE
479 result = generic_file_sendfile(file, &pos, cnt,
480 vio->u.sendfile.cui_actor,
481 vio->u.sendfile.cui_target);
484 #ifdef HAVE_KERNEL_SPLICE_READ
486 result = generic_file_splice_read(file, &pos,
487 vio->u.splice.cui_pipe, cnt,
488 vio->u.splice.cui_flags);
492 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
500 io->ci_nob += result;
501 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
502 cio->cui_fd, pos, result, 0);
508 static int vvp_io_write_start(const struct lu_env *env,
509 const struct cl_io_slice *ios)
511 struct ccc_io *cio = cl2ccc_io(env, ios);
512 struct cl_io *io = ios->cis_io;
513 struct cl_object *obj = io->ci_obj;
514 struct inode *inode = ccc_object_inode(obj);
515 struct file *file = cio->cui_fd->fd_file;
517 loff_t pos = io->u.ci_wr.wr.crw_pos;
518 size_t cnt = io->u.ci_wr.wr.crw_count;
522 if (cl_io_is_append(io)) {
524 * PARALLEL IO This has to be changed for parallel IO doing
525 * out-of-order writes.
527 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
528 #ifndef HAVE_FILE_WRITEV
529 cio->cui_iocb->ki_pos = pos;
533 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
535 if (cl2vvp_io(env, ios)->cui_oneshot > 0)
538 result = lustre_generic_file_write(file, cio, &pos);
543 io->ci_nob += result;
544 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
545 cio->cui_fd, pos, result, 0);
551 #ifndef HAVE_VM_OP_FAULT
552 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
556 vmpage = filemap_nopage(cfio->ft_vma, cfio->nopage.ft_address,
557 cfio->nopage.ft_type);
559 if (vmpage == NOPAGE_SIGBUS) {
560 CDEBUG(D_PAGE, "got addr %lu type %lx - SIGBUS\n",
561 cfio->nopage.ft_address,(long)cfio->nopage.ft_type);
563 } else if (vmpage == NOPAGE_OOM) {
564 CDEBUG(D_PAGE, "got addr %lu type %lx - OOM\n",
565 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
569 LL_CDEBUG_PAGE(D_PAGE, vmpage, "got addr %lu type %lx\n",
570 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
572 cfio->ft_vmpage = vmpage;
577 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
579 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, cfio->fault.ft_vmf);
581 if (cfio->fault.ft_vmf->page) {
582 LL_CDEBUG_PAGE(D_PAGE, cfio->fault.ft_vmf->page,
583 "got addr %p type NOPAGE\n",
584 cfio->fault.ft_vmf->virtual_address);
585 /*XXX workaround to bug in CLIO - he deadlocked with
586 lock cancel if page locked */
587 if (likely(cfio->fault.ft_flags & VM_FAULT_LOCKED)) {
588 unlock_page(cfio->fault.ft_vmf->page);
589 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
592 cfio->ft_vmpage = cfio->fault.ft_vmf->page;
596 if (unlikely (cfio->fault.ft_flags & VM_FAULT_ERROR)) {
597 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n",
598 cfio->fault.ft_vmf->virtual_address);
602 if (unlikely (cfio->fault.ft_flags & VM_FAULT_NOPAGE)) {
603 CDEBUG(D_PAGE, "got addr %p - OOM\n",
604 cfio->fault.ft_vmf->virtual_address);
608 CERROR("unknow error in page fault!\n");
614 static int vvp_io_fault_start(const struct lu_env *env,
615 const struct cl_io_slice *ios)
617 struct vvp_io *vio = cl2vvp_io(env, ios);
618 struct cl_io *io = ios->cis_io;
619 struct cl_object *obj = io->ci_obj;
620 struct inode *inode = ccc_object_inode(obj);
621 struct cl_fault_io *fio = &io->u.ci_fault;
622 struct vvp_fault_io *cfio = &vio->u.fault;
624 int kernel_result = 0;
626 struct cl_page *page;
628 pgoff_t last; /* last page in a file data region */
630 LASSERT(vio->cui_oneshot == 0);
632 if (fio->ft_executable &&
633 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
635 " changed while waiting for the page fault lock\n",
636 PFID(lu_object_fid(&obj->co_lu)));
638 /* offset of the last byte on the page */
639 offset = cl_offset(obj, fio->ft_index + 1) - 1;
640 LASSERT(cl_index(obj, offset) == fio->ft_index);
641 result = ccc_prep_size(env, obj, io, 0, offset + 1, 0, NULL);
645 /* must return locked page */
646 kernel_result = vvp_io_kernel_fault(cfio);
647 if (kernel_result != 0)
648 return kernel_result;
649 /* Temporarily lock vmpage to keep cl_page_find() happy. */
650 lock_page(cfio->ft_vmpage);
651 page = cl_page_find(env, obj, fio->ft_index, cfio->ft_vmpage,
653 unlock_page(cfio->ft_vmpage);
655 page_cache_release(cfio->ft_vmpage);
656 cfio->ft_vmpage = NULL;
657 return PTR_ERR(page);
660 size = i_size_read(inode);
661 last = cl_index(obj, size - 1);
662 if (fio->ft_index == last)
664 * Last page is mapped partially.
666 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
668 fio->ft_nob = cl_page_size(obj);
670 lu_ref_add(&page->cp_reference, "fault", io);
673 * Certain 2.6 kernels return not-NULL from
674 * filemap_nopage() when page is beyond the file size,
675 * on the grounds that "An external ptracer can access
676 * pages that normally aren't accessible.." Don't
677 * propagate such page fault to the lower layers to
678 * avoid side-effects like KMS updates.
680 if (fio->ft_index > last)
686 static int vvp_io_read_page(const struct lu_env *env,
687 const struct cl_io_slice *ios,
688 const struct cl_page_slice *slice)
690 struct cl_io *io = ios->cis_io;
691 struct cl_object *obj = slice->cpl_obj;
692 struct ccc_page *cp = cl2ccc_page(slice);
693 struct cl_page *page = slice->cpl_page;
694 struct inode *inode = ccc_object_inode(obj);
695 struct ll_sb_info *sbi = ll_i2sbi(inode);
696 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
697 struct ll_readahead_state *ras = &fd->fd_ras;
698 cfs_page_t *vmpage = cp->cpg_page;
699 struct cl_2queue *queue = &io->ci_queue;
702 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
703 LASSERT(cl2vvp_io(env, ios)->cui_oneshot == 0);
704 LASSERT(slice->cpl_obj == obj);
708 if (sbi->ll_ra_info.ra_max_pages_per_file)
709 ras_update(sbi, inode, ras, page->cp_index,
710 cp->cpg_defer_uptodate);
712 /* Sanity check whether the page is protected by a lock. */
713 rc = cl_page_is_under_lock(env, io, page);
715 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %i\n",
716 rc == -ENODATA ? "without a lock" :
722 if (cp->cpg_defer_uptodate) {
724 cl_page_export(env, page, 1);
727 * Add page into the queue even when it is marked uptodate above.
728 * this will unlock it automatically as part of cl_page_list_disown().
730 cl_2queue_add(queue, page);
731 if (sbi->ll_ra_info.ra_max_pages_per_file)
732 ll_readahead(env, io, ras,
733 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
738 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
739 struct cl_page *page, struct ccc_page *cp,
740 int to, enum cl_req_type crt)
742 struct cl_2queue *queue;
745 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
747 queue = &io->ci_queue;
749 cl_2queue_init_page(queue, page);
750 cl_page_clip(env, page, 0, to);
752 result = cl_io_submit_sync(env, io, crt, queue, CRP_NORMAL, 0);
753 LASSERT(cl_page_is_owned(page, io));
754 cl_page_clip(env, page, 0, CFS_PAGE_SIZE);
758 * in CRT_WRITE case page is left locked even in case of
761 cl_page_list_disown(env, io, &queue->c2_qin);
762 cl_2queue_fini(env, queue);
768 * Prepare partially written-to page for a write.
770 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
771 struct cl_object *obj, struct cl_page *pg,
773 unsigned from, unsigned to)
775 struct cl_attr *attr = ccc_env_thread_attr(env);
776 loff_t offset = cl_offset(obj, pg->cp_index);
779 cl_object_attr_lock(obj);
780 result = cl_object_attr_get(env, obj, attr);
781 cl_object_attr_unlock(obj);
784 * If are writing to a new page, no need to read old data.
785 * The extent locking will have updated the KMS, and for our
786 * purposes here we can treat it like i_size.
788 if (attr->cat_kms <= offset) {
789 char *kaddr = kmap_atomic(cp->cpg_page, KM_USER0);
791 memset(kaddr, 0, cl_page_size(obj));
792 kunmap_atomic(kaddr, KM_USER0);
793 } else if (cp->cpg_defer_uptodate)
796 result = vvp_page_sync_io(env, io, pg, cp,
797 CFS_PAGE_SIZE, CRT_READ);
799 * In older implementations, obdo_refresh_inode is called here
800 * to update the inode because the write might modify the
801 * object info at OST. However, this has been proven useless,
802 * since LVB functions will be called when user space program
803 * tries to retrieve inode attribute. Also, see bug 15909 for
807 cl_page_export(env, pg, 1);
812 static int vvp_io_prepare_write(const struct lu_env *env,
813 const struct cl_io_slice *ios,
814 const struct cl_page_slice *slice,
815 unsigned from, unsigned to)
817 struct cl_object *obj = slice->cpl_obj;
818 struct ccc_page *cp = cl2ccc_page(slice);
819 struct cl_page *pg = slice->cpl_page;
820 cfs_page_t *vmpage = cp->cpg_page;
826 LINVRNT(cl_page_is_vmlocked(env, pg));
827 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
831 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
832 if (!PageUptodate(vmpage)) {
834 * We're completely overwriting an existing page, so _don't_
835 * set it up to date until commit_write
837 if (from == 0 && to == CFS_PAGE_SIZE) {
838 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
839 POISON_PAGE(page, 0x11);
841 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
844 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
848 static int vvp_io_commit_write(const struct lu_env *env,
849 const struct cl_io_slice *ios,
850 const struct cl_page_slice *slice,
851 unsigned from, unsigned to)
853 struct cl_object *obj = slice->cpl_obj;
854 struct cl_io *io = ios->cis_io;
855 struct ccc_page *cp = cl2ccc_page(slice);
856 struct cl_page *pg = slice->cpl_page;
857 struct inode *inode = ccc_object_inode(obj);
858 struct ll_sb_info *sbi = ll_i2sbi(inode);
859 cfs_page_t *vmpage = cp->cpg_page;
867 LINVRNT(cl_page_is_vmlocked(env, pg));
868 LASSERT(vmpage->mapping->host == inode);
870 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n");
871 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
874 * queue a write for some time in the future the first time we
877 * This is different from what other file systems do: they usually
878 * just mark page (and some of its buffers) dirty and rely on
879 * balance_dirty_pages() to start a write-back. Lustre wants write-back
880 * to be started earlier for the following reasons:
882 * (1) with a large number of clients we need to limit the amount
883 * of cached data on the clients a lot;
885 * (2) large compute jobs generally want compute-only then io-only
886 * and the IO should complete as quickly as possible;
888 * (3) IO is batched up to the RPC size and is async until the
889 * client max cache is hit
890 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
893 if (!PageDirty(vmpage)) {
894 tallyop = LPROC_LL_DIRTY_MISSES;
895 vvp_write_pending(cl2ccc(obj), cp);
896 set_page_dirty(vmpage);
897 /* ll_set_page_dirty() does the same for now, but
898 * it will not soon. */
899 vvp_write_pending(cl2ccc(obj), cp);
900 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
901 if (result == -EDQUOT)
903 * Client ran out of disk space grant. Possible
906 * (a) do a sync write, renewing grant;
908 * (b) stop writing on this stripe, switch to the
911 * (b) is a part of "parallel io" design that is the
912 * ultimate goal. (a) is what "old" client did, and
913 * what the new code continues to do for the time
916 result = vvp_page_sync_io(env, io, pg, cp,
919 CERROR("Write page %lu of inode %p failed %d\n",
920 pg->cp_index, inode, result);
922 tallyop = LPROC_LL_DIRTY_HITS;
925 ll_stats_ops_tally(sbi, tallyop, 1);
927 size = cl_offset(obj, pg->cp_index) + to;
930 if (size > i_size_read(inode))
931 i_size_write(inode, size);
932 cl_page_export(env, pg, 1);
933 } else if (size > i_size_read(inode))
934 cl_page_discard(env, io, pg);
938 static const struct cl_io_operations vvp_io_ops = {
941 .cio_fini = vvp_io_fini,
942 .cio_lock = vvp_io_read_lock,
943 .cio_start = vvp_io_read_start,
944 .cio_advance = ccc_io_advance
947 .cio_fini = vvp_io_fini,
948 .cio_lock = vvp_io_write_lock,
949 .cio_start = vvp_io_write_start,
950 .cio_advance = ccc_io_advance
953 .cio_fini = vvp_io_trunc_fini,
954 .cio_iter_init = vvp_io_trunc_iter_init,
955 .cio_lock = vvp_io_trunc_lock,
956 .cio_start = vvp_io_trunc_start,
957 .cio_end = vvp_io_trunc_end
960 .cio_fini = vvp_io_fault_fini,
961 .cio_iter_init = vvp_io_fault_iter_init,
962 .cio_lock = vvp_io_fault_lock,
963 .cio_start = vvp_io_fault_start,
964 .cio_end = ccc_io_end
967 .cio_fini = vvp_io_fini
970 .cio_read_page = vvp_io_read_page,
971 .cio_prepare_write = vvp_io_prepare_write,
972 .cio_commit_write = vvp_io_commit_write
975 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
978 struct vvp_io *vio = vvp_env_io(env);
979 struct ccc_io *cio = ccc_env_io(env);
980 struct inode *inode = ccc_object_inode(obj);
981 struct ll_sb_info *sbi = ll_i2sbi(inode);
984 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
987 CL_IO_SLICE_CLEAN(cio, cui_cl);
988 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
989 vio->cui_oneshot = 0;
990 vio->cui_ra_window_set = 0;
992 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
996 count = io->u.ci_rw.crw_count;
997 op = io->ci_type == CIT_READ ?
998 LPROC_LL_READ_BYTES : LPROC_LL_WRITE_BYTES;
999 if (io->ci_type == CIT_WRITE)
1000 down(&ll_i2info(inode)->lli_write_sem);
1001 /* "If nbyte is 0, read() will return 0 and have no other
1002 * results." -- Single Unix Spec */
1006 cio->cui_tot_count = count;
1007 cio->cui_tot_nrsegs = 0;
1008 ll_stats_ops_tally(sbi, op, count);
1010 } else if (io->ci_type == CIT_TRUNC) {
1011 /* lockless truncate? */
1012 ll_stats_ops_tally(sbi, LPROC_LL_TRUNC, 1);
1017 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1018 const struct cl_io_slice *slice)
1020 /* Caling just for assertion */
1021 cl2ccc_io(env, slice);
1022 return vvp_env_io(env);