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);
55 /*****************************************************************************
61 static int vvp_io_fault_iter_init(const struct lu_env *env,
62 const struct cl_io_slice *ios)
64 struct vvp_io *vio = cl2vvp_io(env, ios);
65 struct inode *inode = ccc_object_inode(ios->cis_obj);
68 cl2ccc_io(env, ios)->cui_fd->fd_file->f_dentry->d_inode);
69 vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime);
73 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
75 struct cl_io *io = ios->cis_io;
76 struct cl_object *obj = io->ci_obj;
78 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
79 if (io->ci_type == CIT_WRITE)
80 up(&ll_i2info(ccc_object_inode(obj))->lli_write_sem);
82 struct vvp_io *vio = cl2vvp_io(env, ios);
83 struct ccc_io *cio = cl2ccc_io(env, ios);
85 if (vio->cui_ra_window_set)
86 ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
91 static void vvp_io_fault_fini(const struct lu_env *env,
92 const struct cl_io_slice *ios)
94 struct cl_io *io = ios->cis_io;
95 struct cl_page *page = io->u.ci_fault.ft_page;
97 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
100 lu_ref_del(&page->cp_reference, "fault", io);
101 cl_page_put(env, page);
102 io->u.ci_fault.ft_page = NULL;
104 vvp_io_fini(env, ios);
107 enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
110 * we only want to hold PW locks if the mmap() can generate
111 * writes back to the file and that only happens in shared
114 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
119 static int vvp_mmap_locks(const struct lu_env *env,
120 struct ccc_io *vio, struct cl_io *io)
122 struct ccc_thread_info *cti = ccc_env_info(env);
123 struct vm_area_struct *vma;
124 struct cl_lock_descr *descr = &cti->cti_descr;
125 ldlm_policy_data_t policy;
132 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
134 if (cl_io_is_sendfile(io))
137 for (seg = 0; seg < vio->cui_nrsegs; seg++) {
138 const struct iovec *iv = &vio->cui_iov[seg];
140 addr = (unsigned long)iv->iov_base;
145 count += addr & (~CFS_PAGE_MASK);
146 addr &= CFS_PAGE_MASK;
147 while((vma = our_vma(addr, count)) != NULL) {
148 struct inode *inode = vma->vm_file->f_dentry->d_inode;
149 int flags = CEF_MUST;
151 if (ll_file_nolock(vma->vm_file)) {
153 * For no lock case, a lockless lock will be
160 * XXX: Required lock mode can be weakened: CIT_WRITE
161 * io only ever reads user level buffer, and CIT_READ
164 policy_from_vma(&policy, vma, addr, count);
165 descr->cld_mode = vvp_mode_from_vma(vma);
166 descr->cld_obj = ll_i2info(inode)->lli_clob;
167 descr->cld_start = cl_index(descr->cld_obj,
168 policy.l_extent.start);
169 descr->cld_end = cl_index(descr->cld_obj,
170 policy.l_extent.end);
171 result = cl_io_lock_alloc_add(env, io, descr, flags);
175 if (vma->vm_end - addr >= count)
178 count -= vma->vm_end - addr;
185 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
186 enum cl_lock_mode mode, loff_t start, loff_t end)
188 struct ccc_io *cio = ccc_env_io(env);
192 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
193 LASSERT(vvp_env_io(env)->cui_oneshot == 0);
196 ccc_io_update_iov(env, cio, io);
198 if (io->u.ci_rw.crw_nonblock)
199 ast_flags |= CEF_NONBLOCK;
200 result = vvp_mmap_locks(env, cio, io);
202 result = ccc_io_one_lock(env, io, ast_flags, mode, start, end);
206 static int vvp_io_read_lock(const struct lu_env *env,
207 const struct cl_io_slice *ios)
209 struct cl_io *io = ios->cis_io;
210 struct ll_inode_info *lli = ll_i2info(ccc_object_inode(io->ci_obj));
214 /* XXX: Layer violation, we shouldn't see lsm at llite level. */
215 if (lli->lli_smd != NULL) /* lsm-less file, don't need to lock */
216 result = vvp_io_rw_lock(env, io, CLM_READ,
217 io->u.ci_rd.rd.crw_pos,
218 io->u.ci_rd.rd.crw_pos +
219 io->u.ci_rd.rd.crw_count - 1);
225 static int vvp_io_fault_lock(const struct lu_env *env,
226 const struct cl_io_slice *ios)
228 struct cl_io *io = ios->cis_io;
229 struct vvp_io *vio = cl2vvp_io(env, ios);
231 * XXX LDLM_FL_CBPENDING
233 return ccc_io_one_lock_index
234 (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma),
235 io->u.ci_fault.ft_index, io->u.ci_fault.ft_index);
238 static int vvp_io_write_lock(const struct lu_env *env,
239 const struct cl_io_slice *ios)
241 struct cl_io *io = ios->cis_io;
246 if (cl2vvp_io(env, ios)->cui_oneshot == 0) {
247 if (io->u.ci_wr.wr_append) {
249 end = OBD_OBJECT_EOF;
251 start = io->u.ci_wr.wr.crw_pos;
252 end = start + io->u.ci_wr.wr.crw_count - 1;
254 result = vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
260 static int vvp_io_trunc_iter_init(const struct lu_env *env,
261 const struct cl_io_slice *ios)
263 struct ccc_io *vio = cl2ccc_io(env, ios);
264 struct inode *inode = ccc_object_inode(ios->cis_obj);
267 * We really need to get our PW lock before we change inode->i_size.
268 * If we don't we can race with other i_size updaters on our node,
269 * like ll_file_read. We can also race with i_size propogation to
270 * other nodes through dirtying and writeback of final cached pages.
271 * This last one is especially bad for racing o_append users on other
275 UNLOCK_INODE_MUTEX(inode);
276 UP_WRITE_I_ALLOC_SEM(inode);
277 vio->u.trunc.cui_locks_released = 1;
282 * Implementation of cl_io_operations::cio_lock() method for CIT_TRUNC io.
284 * Handles "lockless io" mode when extent locking is done by server.
286 static int vvp_io_trunc_lock(const struct lu_env *env,
287 const struct cl_io_slice *ios)
289 struct ccc_io *vio = cl2ccc_io(env, ios);
290 struct cl_io *io = ios->cis_io;
291 size_t new_size = io->u.ci_truncate.tr_size;
292 __u32 enqflags = new_size == 0 ? CEF_DISCARD_DATA : 0;
295 vio->u.trunc.cui_local_lock = TRUNC_EXTENT;
296 result = ccc_io_one_lock(env, io, enqflags, CLM_WRITE,
297 new_size, OBD_OBJECT_EOF);
301 static int vvp_io_trunc_start(const struct lu_env *env,
302 const struct cl_io_slice *ios)
304 struct ccc_io *cio = cl2ccc_io(env, ios);
305 struct vvp_io *vio = cl2vvp_io(env, ios);
306 struct cl_io *io = ios->cis_io;
307 struct inode *inode = ccc_object_inode(io->ci_obj);
308 struct cl_object *obj = ios->cis_obj;
309 size_t size = io->u.ci_truncate.tr_size;
310 pgoff_t start = cl_index(obj, size);
313 LASSERT(cio->u.trunc.cui_locks_released);
314 LASSERT(vio->cui_oneshot == 0);
316 LOCK_INODE_MUTEX(inode);
317 DOWN_WRITE_I_ALLOC_SEM(inode);
318 cio->u.trunc.cui_locks_released = 0;
321 * Only ll_inode_size_lock is taken at this level. lov_stripe_lock()
322 * is grabbed by ll_truncate() only over call to obd_adjust_kms(). If
323 * vmtruncate returns 0, then ll_truncate dropped ll_inode_size_lock()
325 ll_inode_size_lock(inode, 0);
326 result = vmtruncate(inode, size);
328 ll_inode_size_unlock(inode, 0);
330 * If a page is partially truncated, keep it owned across truncate to
333 * XXX this properly belongs to osc, because races in question are OST
336 if (cl_offset(obj, start) != size) {
337 struct cl_object_header *hdr;
339 hdr = cl_object_header(obj);
340 spin_lock(&hdr->coh_page_guard);
341 vio->cui_partpage = cl_page_lookup(hdr, start);
342 spin_unlock(&hdr->coh_page_guard);
344 if (vio->cui_partpage != NULL)
346 * Wait for the transfer completion for a partially
347 * truncated page to avoid dead-locking an OST with
348 * the concurrent page-wise overlapping WRITE and
349 * PUNCH requests. BUG:17397.
351 * Partial page is disowned in vvp_io_trunc_end().
353 cl_page_own(env, io, vio->cui_partpage);
355 vio->cui_partpage = NULL;
359 static void vvp_io_trunc_end(const struct lu_env *env,
360 const struct cl_io_slice *ios)
362 struct vvp_io *vio = cl2vvp_io(env, ios);
364 if (vio->cui_partpage != NULL) {
365 cl_page_disown(env, ios->cis_io, vio->cui_partpage);
366 cl_page_put(env, vio->cui_partpage);
367 vio->cui_partpage = NULL;
371 static void vvp_io_trunc_fini(const struct lu_env *env,
372 const struct cl_io_slice *ios)
374 struct ccc_io *cio = ccc_env_io(env);
375 struct inode *inode = ccc_object_inode(ios->cis_io->ci_obj);
377 if (cio->u.trunc.cui_locks_released) {
378 LOCK_INODE_MUTEX(inode);
379 DOWN_WRITE_I_ALLOC_SEM(inode);
380 cio->u.trunc.cui_locks_released = 0;
382 vvp_io_fini(env, ios);
385 #ifdef HAVE_FILE_READV
386 static ssize_t lustre_generic_file_read(struct file *file,
387 struct ccc_io *vio, loff_t *ppos)
389 return generic_file_readv(file, vio->cui_iov, vio->cui_nrsegs, ppos);
392 static ssize_t lustre_generic_file_write(struct file *file,
393 struct ccc_io *vio, loff_t *ppos)
395 return generic_file_writev(file, vio->cui_iov, vio->cui_nrsegs, ppos);
398 static ssize_t lustre_generic_file_read(struct file *file,
399 struct ccc_io *vio, loff_t *ppos)
401 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
402 vio->cui_nrsegs, *ppos);
405 static ssize_t lustre_generic_file_write(struct file *file,
406 struct ccc_io *vio, loff_t *ppos)
408 return generic_file_aio_write(vio->cui_iocb, vio->cui_iov,
409 vio->cui_nrsegs, *ppos);
413 static int vvp_io_read_start(const struct lu_env *env,
414 const struct cl_io_slice *ios)
416 struct vvp_io *vio = cl2vvp_io(env, ios);
417 struct ccc_io *cio = cl2ccc_io(env, ios);
418 struct cl_io *io = ios->cis_io;
419 struct cl_object *obj = io->ci_obj;
420 struct inode *inode = ccc_object_inode(obj);
421 struct ll_ra_read *bead = &vio->cui_bead;
422 struct file *file = cio->cui_fd->fd_file;
425 loff_t pos = io->u.ci_rd.rd.crw_pos;
426 long cnt = io->u.ci_rd.rd.crw_count;
427 long tot = cio->cui_tot_count;
430 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
431 LASSERT(vio->cui_oneshot == 0);
433 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
435 result = ccc_prep_size(env, obj, io, pos, tot, 1, &exceed);
438 else if (exceed != 0)
441 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
442 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
443 inode->i_ino, cnt, pos, i_size_read(inode));
445 /* turn off the kernel's read-ahead */
446 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
448 /* initialize read-ahead window once per syscall */
449 if (!vio->cui_ra_window_set) {
450 vio->cui_ra_window_set = 1;
451 bead->lrr_start = cl_index(obj, pos);
453 * XXX: explicit CFS_PAGE_SIZE
455 bead->lrr_count = cl_index(obj, tot + CFS_PAGE_SIZE - 1);
456 ll_ra_read_in(file, bead);
461 if (cl_io_is_sendfile(io)) {
462 result = generic_file_sendfile(file, &pos, cnt,
463 vio->u.read.cui_actor, vio->u.read.cui_target);
465 result = lustre_generic_file_read(file, cio, &pos);
472 io->ci_nob += result;
473 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
474 cio->cui_fd, pos, result, 0);
480 static int vvp_io_write_start(const struct lu_env *env,
481 const struct cl_io_slice *ios)
483 struct ccc_io *cio = cl2ccc_io(env, ios);
484 struct cl_io *io = ios->cis_io;
485 struct cl_object *obj = io->ci_obj;
486 struct inode *inode = ccc_object_inode(obj);
487 struct file *file = cio->cui_fd->fd_file;
489 loff_t pos = io->u.ci_wr.wr.crw_pos;
490 size_t cnt = io->u.ci_wr.wr.crw_count;
494 if (cl_io_is_append(io)) {
496 * PARALLEL IO This has to be changed for parallel IO doing
497 * out-of-order writes.
499 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
500 #ifndef HAVE_FILE_WRITEV
501 cio->cui_iocb->ki_pos = pos;
505 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
507 if (cl2vvp_io(env, ios)->cui_oneshot > 0)
510 result = lustre_generic_file_write(file, cio, &pos);
515 io->ci_nob += result;
516 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
517 cio->cui_fd, pos, result, 0);
523 static int vvp_io_fault_start(const struct lu_env *env,
524 const struct cl_io_slice *ios)
526 struct vvp_io *vio = cl2vvp_io(env, ios);
527 struct cl_io *io = ios->cis_io;
528 struct cl_object *obj = io->ci_obj;
529 struct inode *inode = ccc_object_inode(obj);
530 struct cl_fault_io *fio = &io->u.ci_fault;
531 struct vvp_fault_io *cfio = &vio->u.fault;
536 LASSERT(vio->cui_oneshot == 0);
538 if (fio->ft_executable &&
539 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
541 " changed while waiting for the page fault lock\n",
542 PFID(lu_object_fid(&obj->co_lu)));
544 /* offset of the last byte on the page */
545 offset = cl_offset(obj, fio->ft_index + 1) - 1;
546 LASSERT(cl_index(obj, offset) == fio->ft_index);
547 result = ccc_prep_size(env, obj, io, 0, offset + 1, 0, NULL);
551 vmpage = filemap_nopage(cfio->ft_vma, cfio->ft_address, cfio->ft_type);
552 if (vmpage != NOPAGE_SIGBUS && vmpage != NOPAGE_OOM)
553 LL_CDEBUG_PAGE(D_PAGE, vmpage,
554 "got addr %lu type %lx\n",
555 cfio->ft_address, (long)cfio->ft_type);
557 CDEBUG(D_PAGE, "got addr %lu type %lx - SIGBUS\n",
558 cfio->ft_address, (long)cfio->ft_type);
560 if (vmpage == NOPAGE_SIGBUS)
562 else if (vmpage == NOPAGE_OOM)
565 struct cl_page *page;
567 pgoff_t last; /* last page in a file data region */
569 /* Temporarily lock vmpage to keep cl_page_find() happy. */
571 page = cl_page_find(env, obj, fio->ft_index, vmpage,
575 size = i_size_read(inode);
576 last = cl_index(obj, size - 1);
577 if (fio->ft_index == last)
579 * Last page is mapped partially.
581 fio->ft_nob = size - cl_offset(obj,
584 fio->ft_nob = cl_page_size(obj);
585 lu_ref_add(&page->cp_reference, "fault", io);
588 * Certain 2.6 kernels return not-NULL from
589 * filemap_nopage() when page is beyond the file size,
590 * on the grounds that "An external ptracer can access
591 * pages that normally aren't accessible.." Don't
592 * propagate such page fault to the lower layers to
593 * avoid side-effects like KMS updates.
595 if (fio->ft_index > last)
598 result = PTR_ERR(page);
603 static int vvp_io_read_page(const struct lu_env *env,
604 const struct cl_io_slice *ios,
605 const struct cl_page_slice *slice)
607 struct cl_io *io = ios->cis_io;
608 struct cl_object *obj = slice->cpl_obj;
609 struct ccc_page *cp = cl2ccc_page(slice);
610 struct cl_page *page = slice->cpl_page;
611 struct inode *inode = ccc_object_inode(obj);
612 struct ll_sb_info *sbi = ll_i2sbi(inode);
613 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
614 struct ll_readahead_state *ras = &fd->fd_ras;
615 cfs_page_t *vmpage = cp->cpg_page;
616 struct cl_2queue *queue = &io->ci_queue;
619 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
620 LASSERT(cl2vvp_io(env, ios)->cui_oneshot == 0);
621 LASSERT(slice->cpl_obj == obj);
625 if (sbi->ll_ra_info.ra_max_pages_per_file)
626 ras_update(sbi, inode, ras, page->cp_index,
627 cp->cpg_defer_uptodate);
629 /* Sanity check whether the page is protected by a lock. */
630 rc = cl_page_is_under_lock(env, io, page);
632 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %i\n",
633 rc == -ENODATA ? "without a lock" :
639 if (cp->cpg_defer_uptodate) {
641 cl_page_export(env, page, 1);
644 * Add page into the queue even when it is marked uptodate above.
645 * this will unlock it automatically as part of cl_page_list_disown().
647 cl_2queue_add(queue, page);
648 if (sbi->ll_ra_info.ra_max_pages_per_file)
649 ll_readahead(env, io, ras,
650 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
655 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
656 struct cl_page *page, struct ccc_page *cp,
657 int to, enum cl_req_type crt)
659 struct cl_2queue *queue;
662 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
664 queue = &io->ci_queue;
666 cl_2queue_init_page(queue, page);
667 cl_page_clip(env, page, 0, to);
669 result = cl_io_submit_sync(env, io, crt, queue, CRP_NORMAL, 0);
670 LASSERT(cl_page_is_owned(page, io));
671 cl_page_clip(env, page, 0, CFS_PAGE_SIZE);
675 * in CRT_WRITE case page is left locked even in case of
678 cl_page_list_disown(env, io, &queue->c2_qin);
679 cl_2queue_fini(env, queue);
685 * Prepare partially written-to page for a write.
687 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
688 struct cl_object *obj, struct cl_page *pg,
690 unsigned from, unsigned to)
692 struct cl_attr *attr = ccc_env_thread_attr(env);
693 loff_t offset = cl_offset(obj, pg->cp_index);
696 cl_object_attr_lock(obj);
697 result = cl_object_attr_get(env, obj, attr);
698 cl_object_attr_unlock(obj);
701 * If are writing to a new page, no need to read old data.
702 * The extent locking will have updated the KMS, and for our
703 * purposes here we can treat it like i_size.
705 if (attr->cat_kms <= offset) {
706 char *kaddr = kmap_atomic(cp->cpg_page, KM_USER0);
708 memset(kaddr, 0, cl_page_size(obj));
709 kunmap_atomic(kaddr, KM_USER0);
710 } else if (cp->cpg_defer_uptodate)
713 result = vvp_page_sync_io(env, io, pg, cp,
714 CFS_PAGE_SIZE, CRT_READ);
716 * In older implementations, obdo_refresh_inode is called here
717 * to update the inode because the write might modify the
718 * object info at OST. However, this has been proven useless,
719 * since LVB functions will be called when user space program
720 * tries to retrieve inode attribute. Also, see bug 15909 for
724 cl_page_export(env, pg, 1);
729 static int vvp_io_prepare_write(const struct lu_env *env,
730 const struct cl_io_slice *ios,
731 const struct cl_page_slice *slice,
732 unsigned from, unsigned to)
734 struct cl_object *obj = slice->cpl_obj;
735 struct ccc_page *cp = cl2ccc_page(slice);
736 struct cl_page *pg = slice->cpl_page;
737 cfs_page_t *vmpage = cp->cpg_page;
743 LINVRNT(cl_page_is_vmlocked(env, pg));
744 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
748 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
749 if (!PageUptodate(vmpage)) {
751 * We're completely overwriting an existing page, so _don't_
752 * set it up to date until commit_write
754 if (from == 0 && to == CFS_PAGE_SIZE) {
755 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
756 POISON_PAGE(page, 0x11);
758 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
761 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
765 static int vvp_io_commit_write(const struct lu_env *env,
766 const struct cl_io_slice *ios,
767 const struct cl_page_slice *slice,
768 unsigned from, unsigned to)
770 struct cl_object *obj = slice->cpl_obj;
771 struct cl_io *io = ios->cis_io;
772 struct ccc_page *cp = cl2ccc_page(slice);
773 struct cl_page *pg = slice->cpl_page;
774 struct inode *inode = ccc_object_inode(obj);
775 struct ll_sb_info *sbi = ll_i2sbi(inode);
776 cfs_page_t *vmpage = cp->cpg_page;
784 LINVRNT(cl_page_is_vmlocked(env, pg));
785 LASSERT(vmpage->mapping->host == inode);
787 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n");
788 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
791 * queue a write for some time in the future the first time we
794 * This is different from what other file systems do: they usually
795 * just mark page (and some of its buffers) dirty and rely on
796 * balance_dirty_pages() to start a write-back. Lustre wants write-back
797 * to be started earlier for the following reasons:
799 * (1) with a large number of clients we need to limit the amount
800 * of cached data on the clients a lot;
802 * (2) large compute jobs generally want compute-only then io-only
803 * and the IO should complete as quickly as possible;
805 * (3) IO is batched up to the RPC size and is async until the
806 * client max cache is hit
807 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
810 if (!PageDirty(vmpage)) {
811 tallyop = LPROC_LL_DIRTY_MISSES;
812 vvp_write_pending(cl2ccc(obj), cp);
813 set_page_dirty(vmpage);
814 /* ll_set_page_dirty() does the same for now, but
815 * it will not soon. */
816 vvp_write_pending(cl2ccc(obj), cp);
817 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
818 if (result == -EDQUOT)
820 * Client ran out of disk space grant. Possible
823 * (a) do a sync write, renewing grant;
825 * (b) stop writing on this stripe, switch to the
828 * (b) is a part of "parallel io" design that is the
829 * ultimate goal. (a) is what "old" client did, and
830 * what the new code continues to do for the time
833 result = vvp_page_sync_io(env, io, pg, cp,
836 CERROR("Write page %lu of inode %p failed %d\n",
837 pg->cp_index, inode, result);
839 tallyop = LPROC_LL_DIRTY_HITS;
842 ll_stats_ops_tally(sbi, tallyop, 1);
844 size = cl_offset(obj, pg->cp_index) + to;
847 if (size > i_size_read(inode))
848 i_size_write(inode, size);
849 cl_page_export(env, pg, 1);
850 } else if (size > i_size_read(inode))
851 cl_page_discard(env, io, pg);
855 static const struct cl_io_operations vvp_io_ops = {
858 .cio_fini = vvp_io_fini,
859 .cio_lock = vvp_io_read_lock,
860 .cio_start = vvp_io_read_start,
861 .cio_advance = ccc_io_advance
864 .cio_fini = vvp_io_fini,
865 .cio_lock = vvp_io_write_lock,
866 .cio_start = vvp_io_write_start,
867 .cio_advance = ccc_io_advance
870 .cio_fini = vvp_io_trunc_fini,
871 .cio_iter_init = vvp_io_trunc_iter_init,
872 .cio_lock = vvp_io_trunc_lock,
873 .cio_start = vvp_io_trunc_start,
874 .cio_end = vvp_io_trunc_end
877 .cio_fini = vvp_io_fault_fini,
878 .cio_iter_init = vvp_io_fault_iter_init,
879 .cio_lock = vvp_io_fault_lock,
880 .cio_start = vvp_io_fault_start,
881 .cio_end = ccc_io_end
884 .cio_fini = vvp_io_fini
887 .cio_read_page = vvp_io_read_page,
888 .cio_prepare_write = vvp_io_prepare_write,
889 .cio_commit_write = vvp_io_commit_write
892 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
895 struct vvp_io *vio = vvp_env_io(env);
896 struct ccc_io *cio = ccc_env_io(env);
897 struct inode *inode = ccc_object_inode(obj);
898 struct ll_sb_info *sbi = ll_i2sbi(inode);
901 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
904 CL_IO_SLICE_CLEAN(cio, cui_cl);
905 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
906 vio->cui_oneshot = 0;
907 vio->cui_ra_window_set = 0;
909 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
913 count = io->u.ci_rw.crw_count;
914 op = io->ci_type == CIT_READ ?
915 LPROC_LL_READ_BYTES : LPROC_LL_WRITE_BYTES;
916 if (io->ci_type == CIT_WRITE)
917 down(&ll_i2info(inode)->lli_write_sem);
918 /* "If nbyte is 0, read() will return 0 and have no other
919 * results." -- Single Unix Spec */
923 cio->cui_tot_count = count;
924 cio->cui_tot_nrsegs = 0;
925 ll_stats_ops_tally(sbi, op, count);
927 } else if (io->ci_type == CIT_TRUNC) {
928 /* lockless truncate? */
929 ll_stats_ops_tally(sbi, LPROC_LL_TRUNC, 1);
934 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
935 const struct cl_io_slice *slice)
937 /* Caling just for assertion */
938 cl2ccc_io(env, slice);
939 return vvp_env_io(env);