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;
133 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
135 if (cl_io_is_sendfile(io))
138 for (seg = 0; seg < vio->cui_nrsegs; seg++) {
139 const struct iovec *iv = &vio->cui_iov[seg];
141 addr = (unsigned long)iv->iov_base;
146 count += addr & (~CFS_PAGE_MASK);
147 addr &= CFS_PAGE_MASK;
148 while((vma = our_vma(addr, count)) != NULL) {
149 struct file *file = vma->vm_file;
150 struct ll_file_data *fd;
153 fd = LUSTRE_FPRIVATE(file);
155 inode = file->f_dentry->d_inode;
156 if (!(fd->fd_flags & LL_FILE_IGNORE_LOCK ||
157 ll_i2sbi(inode)->ll_flags & LL_SBI_NOLCK))
161 * XXX: Required lock mode can be weakened: CIT_WRITE
162 * io only ever reads user level buffer, and CIT_READ
165 policy_from_vma(&policy, vma, addr, count);
166 descr->cld_mode = vvp_mode_from_vma(vma);
167 descr->cld_obj = ll_i2info(inode)->lli_clob;
168 descr->cld_start = cl_index(descr->cld_obj,
169 policy.l_extent.start);
170 descr->cld_end = cl_index(descr->cld_obj,
171 policy.l_extent.end);
172 result = cl_io_lock_alloc_add(env, io, descr, CEF_MUST);
177 if (vma->vm_end - addr >= count)
179 count -= vma->vm_end - addr;
186 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
187 enum cl_lock_mode mode, loff_t start, loff_t end)
189 struct ccc_io *cio = ccc_env_io(env);
193 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
194 LASSERT(vvp_env_io(env)->cui_oneshot == 0);
197 ccc_io_update_iov(env, cio, io);
199 if (io->u.ci_rw.crw_nonblock)
200 ast_flags |= CEF_NONBLOCK;
201 result = vvp_mmap_locks(env, cio, io);
203 result = ccc_io_one_lock(env, io, ast_flags, mode, start, end);
207 static int vvp_io_read_lock(const struct lu_env *env,
208 const struct cl_io_slice *ios)
210 struct cl_io *io = ios->cis_io;
211 struct ll_inode_info *lli = ll_i2info(ccc_object_inode(io->ci_obj));
215 /* XXX: Layer violation, we shouldn't see lsm at llite level. */
216 if (lli->lli_smd != NULL) /* lsm-less file, don't need to lock */
217 result = vvp_io_rw_lock(env, io, CLM_READ,
218 io->u.ci_rd.rd.crw_pos,
219 io->u.ci_rd.rd.crw_pos +
220 io->u.ci_rd.rd.crw_count - 1);
226 static int vvp_io_fault_lock(const struct lu_env *env,
227 const struct cl_io_slice *ios)
229 struct cl_io *io = ios->cis_io;
230 struct vvp_io *vio = cl2vvp_io(env, ios);
232 * XXX LDLM_FL_CBPENDING
234 return ccc_io_one_lock_index
235 (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma),
236 io->u.ci_fault.ft_index, io->u.ci_fault.ft_index);
239 static int vvp_io_write_lock(const struct lu_env *env,
240 const struct cl_io_slice *ios)
242 struct cl_io *io = ios->cis_io;
247 if (cl2vvp_io(env, ios)->cui_oneshot == 0) {
248 if (io->u.ci_wr.wr_append) {
250 end = OBD_OBJECT_EOF;
252 start = io->u.ci_wr.wr.crw_pos;
253 end = start + io->u.ci_wr.wr.crw_count - 1;
255 result = vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
261 static int vvp_io_trunc_iter_init(const struct lu_env *env,
262 const struct cl_io_slice *ios)
264 struct ccc_io *vio = cl2ccc_io(env, ios);
265 struct inode *inode = ccc_object_inode(ios->cis_obj);
268 * We really need to get our PW lock before we change inode->i_size.
269 * If we don't we can race with other i_size updaters on our node,
270 * like ll_file_read. We can also race with i_size propogation to
271 * other nodes through dirtying and writeback of final cached pages.
272 * This last one is especially bad for racing o_append users on other
276 UNLOCK_INODE_MUTEX(inode);
277 UP_WRITE_I_ALLOC_SEM(inode);
278 vio->u.trunc.cui_locks_released = 1;
283 * Implementation of cl_io_operations::vio_lock() method for CIT_TRUNC io.
285 * Handles "lockless io" mode when extent locking is done by server.
287 static int vvp_io_trunc_lock(const struct lu_env *env,
288 const struct cl_io_slice *ios)
290 struct ccc_io *vio = cl2ccc_io(env, ios);
291 struct cl_io *io = ios->cis_io;
292 size_t new_size = io->u.ci_truncate.tr_size;
293 __u32 enqflags = new_size == 0 ? CEF_DISCARD_DATA : 0;
296 vio->u.trunc.cui_local_lock = TRUNC_EXTENT;
297 result = ccc_io_one_lock(env, io, enqflags, CLM_WRITE,
298 new_size, OBD_OBJECT_EOF);
302 static int vvp_io_trunc_start(const struct lu_env *env,
303 const struct cl_io_slice *ios)
305 struct ccc_io *cio = cl2ccc_io(env, ios);
306 struct vvp_io *vio = cl2vvp_io(env, ios);
307 struct cl_io *io = ios->cis_io;
308 struct inode *inode = ccc_object_inode(io->ci_obj);
309 struct cl_object *obj = ios->cis_obj;
310 size_t size = io->u.ci_truncate.tr_size;
311 pgoff_t start = cl_index(obj, size);
314 LASSERT(cio->u.trunc.cui_locks_released);
315 LASSERT(vio->cui_oneshot == 0);
317 LOCK_INODE_MUTEX(inode);
318 DOWN_WRITE_I_ALLOC_SEM(inode);
319 cio->u.trunc.cui_locks_released = 0;
322 * Only ll_inode_size_lock is taken at this level. lov_stripe_lock()
323 * is grabbed by ll_truncate() only over call to obd_adjust_kms(). If
324 * vmtruncate returns 0, then ll_truncate dropped ll_inode_size_lock()
326 ll_inode_size_lock(inode, 0);
327 result = vmtruncate(inode, size);
329 ll_inode_size_unlock(inode, 0);
331 * If a page is partially truncated, keep it owned across truncate to
334 * XXX this properly belongs to osc, because races in question are OST
337 if (cl_offset(obj, start) != size) {
338 struct cl_object_header *hdr;
340 hdr = cl_object_header(obj);
341 spin_lock(&hdr->coh_page_guard);
342 vio->cui_partpage = cl_page_lookup(hdr, start);
343 spin_unlock(&hdr->coh_page_guard);
345 if (vio->cui_partpage != NULL)
347 * Wait for the transfer completion for a partially
348 * truncated page to avoid dead-locking an OST with
349 * the concurrent page-wise overlapping WRITE and
350 * PUNCH requests. BUG:17397.
352 * Partial page is disowned in vvp_io_trunc_end().
354 cl_page_own(env, io, vio->cui_partpage);
356 vio->cui_partpage = NULL;
360 static void vvp_io_trunc_end(const struct lu_env *env,
361 const struct cl_io_slice *ios)
363 struct vvp_io *vio = cl2vvp_io(env, ios);
365 if (vio->cui_partpage != NULL) {
366 cl_page_disown(env, ios->cis_io, vio->cui_partpage);
367 cl_page_put(env, vio->cui_partpage);
368 vio->cui_partpage = NULL;
372 static void vvp_io_trunc_fini(const struct lu_env *env,
373 const struct cl_io_slice *ios)
375 struct ccc_io *cio = ccc_env_io(env);
376 struct inode *inode = ccc_object_inode(ios->cis_io->ci_obj);
378 if (cio->u.trunc.cui_locks_released) {
379 LOCK_INODE_MUTEX(inode);
380 DOWN_WRITE_I_ALLOC_SEM(inode);
381 cio->u.trunc.cui_locks_released = 0;
383 vvp_io_fini(env, ios);
386 #ifdef HAVE_FILE_READV
387 static ssize_t lustre_generic_file_read(struct file *file,
388 struct ccc_io *vio, loff_t *ppos)
390 return generic_file_readv(file, vio->cui_iov, vio->cui_nrsegs, ppos);
393 static ssize_t lustre_generic_file_write(struct file *file,
394 struct ccc_io *vio, loff_t *ppos)
396 return generic_file_writev(file, vio->cui_iov, vio->cui_nrsegs, ppos);
399 static ssize_t lustre_generic_file_read(struct file *file,
400 struct ccc_io *vio, loff_t *ppos)
402 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
403 vio->cui_nrsegs, *ppos);
406 static ssize_t lustre_generic_file_write(struct file *file,
407 struct ccc_io *vio, loff_t *ppos)
409 return generic_file_aio_write(vio->cui_iocb, vio->cui_iov,
410 vio->cui_nrsegs, *ppos);
414 static int vvp_io_read_start(const struct lu_env *env,
415 const struct cl_io_slice *ios)
417 struct vvp_io *vio = cl2vvp_io(env, ios);
418 struct ccc_io *cio = cl2ccc_io(env, ios);
419 struct cl_io *io = ios->cis_io;
420 struct cl_object *obj = io->ci_obj;
421 struct inode *inode = ccc_object_inode(obj);
422 struct ll_ra_read *bead = &vio->cui_bead;
423 struct file *file = cio->cui_fd->fd_file;
426 loff_t pos = io->u.ci_rd.rd.crw_pos;
427 size_t cnt = io->u.ci_rd.rd.crw_count;
428 size_t tot = cio->cui_tot_count;
431 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
432 LASSERT(vio->cui_oneshot == 0);
434 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + (long long)cnt);
436 result = ccc_prep_size(env, obj, io, pos, tot, 1, &exceed);
439 else if (exceed != 0)
442 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
443 "Read ino %lu, "LPSZ" bytes, offset %lld, size %llu\n",
444 inode->i_ino, cnt, pos, i_size_read(inode));
446 /* turn off the kernel's read-ahead */
447 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
449 /* initialize read-ahead window once per syscall */
450 if (!vio->cui_ra_window_set) {
451 vio->cui_ra_window_set = 1;
452 bead->lrr_start = cl_index(obj, pos);
454 * XXX: explicit CFS_PAGE_SIZE
456 bead->lrr_count = cl_index(obj, tot + CFS_PAGE_SIZE - 1);
457 ll_ra_read_in(file, bead);
462 if (cl_io_is_sendfile(io)) {
463 result = generic_file_sendfile(file, &pos, cnt,
464 vio->u.read.cui_actor, vio->u.read.cui_target);
466 result = lustre_generic_file_read(file, cio, &pos);
473 io->ci_nob += result;
474 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
475 cio->cui_fd, pos, result, 0);
481 static int vvp_io_write_start(const struct lu_env *env,
482 const struct cl_io_slice *ios)
484 struct ccc_io *cio = cl2ccc_io(env, ios);
485 struct cl_io *io = ios->cis_io;
486 struct cl_object *obj = io->ci_obj;
487 struct inode *inode = ccc_object_inode(obj);
488 struct file *file = cio->cui_fd->fd_file;
490 loff_t pos = io->u.ci_wr.wr.crw_pos;
491 size_t cnt = io->u.ci_wr.wr.crw_count;
495 if (cl_io_is_append(io)) {
497 * PARALLEL IO This has to be changed for parallel IO doing
498 * out-of-order writes.
500 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
501 #ifndef HAVE_FILE_WRITEV
502 cio->cui_iocb->ki_pos = pos;
506 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
508 if (cl2vvp_io(env, ios)->cui_oneshot > 0)
511 result = lustre_generic_file_write(file, cio, &pos);
516 io->ci_nob += result;
517 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
518 cio->cui_fd, pos, result, 0);
524 static int vvp_io_fault_start(const struct lu_env *env,
525 const struct cl_io_slice *ios)
527 struct vvp_io *vio = cl2vvp_io(env, ios);
528 struct cl_io *io = ios->cis_io;
529 struct cl_object *obj = io->ci_obj;
530 struct inode *inode = ccc_object_inode(obj);
531 struct cl_fault_io *fio = &io->u.ci_fault;
532 struct vvp_fault_io *cfio = &vio->u.fault;
537 LASSERT(vio->cui_oneshot == 0);
539 if (fio->ft_executable &&
540 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
542 " changed while waiting for the page fault lock\n",
543 PFID(lu_object_fid(&obj->co_lu)));
545 /* offset of the last byte on the page */
546 offset = cl_offset(obj, fio->ft_index + 1) - 1;
547 LASSERT(cl_index(obj, offset) == fio->ft_index);
548 result = ccc_prep_size(env, obj, io, 0, offset + 1, 0, NULL);
552 vmpage = filemap_nopage(cfio->ft_vma, cfio->ft_address, cfio->ft_type);
553 if (vmpage != NOPAGE_SIGBUS && vmpage != NOPAGE_OOM)
554 LL_CDEBUG_PAGE(D_PAGE, vmpage,
555 "got addr %lu type %lx\n",
556 cfio->ft_address, (long)cfio->ft_type);
558 CDEBUG(D_PAGE, "got addr %lu type %lx - SIGBUS\n",
559 cfio->ft_address, (long)cfio->ft_type);
561 if (vmpage == NOPAGE_SIGBUS)
563 else if (vmpage == NOPAGE_OOM)
566 struct cl_page *page;
568 pgoff_t last; /* last page in a file data region */
570 /* Temporarily lock vmpage to keep cl_page_find() happy. */
572 page = cl_page_find(env, obj, fio->ft_index, vmpage,
576 size = i_size_read(inode);
577 last = cl_index(obj, size - 1);
578 if (fio->ft_index == last)
580 * Last page is mapped partially.
582 fio->ft_nob = size - cl_offset(obj,
585 fio->ft_nob = cl_page_size(obj);
586 lu_ref_add(&page->cp_reference, "fault", io);
589 * Certain 2.6 kernels return not-NULL from
590 * filemap_nopage() when page is beyond the file size,
591 * on the grounds that "An external ptracer can access
592 * pages that normally aren't accessible.." Don't
593 * propagate such page fault to the lower layers to
594 * avoid side-effects like KMS updates.
596 if (fio->ft_index > last)
599 result = PTR_ERR(page);
604 static int vvp_io_read_page(const struct lu_env *env,
605 const struct cl_io_slice *ios,
606 const struct cl_page_slice *slice)
608 struct cl_io *io = ios->cis_io;
609 struct cl_object *obj = slice->cpl_obj;
610 struct ccc_page *cp = cl2ccc_page(slice);
611 struct cl_page *page = slice->cpl_page;
612 struct inode *inode = ccc_object_inode(obj);
613 struct ll_sb_info *sbi = ll_i2sbi(inode);
614 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
615 struct ll_readahead_state *ras = &fd->fd_ras;
616 cfs_page_t *vmpage = cp->cpg_page;
617 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 if (likely(!(fd->fd_flags & LL_FILE_IGNORE_LOCK))) {
633 rc = cl_page_is_under_lock(env, io, page);
635 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %i\n",
636 rc == -ENODATA ? "without a lock" :
643 if (cp->cpg_defer_uptodate) {
645 cl_page_export(env, page);
648 * Add page into the queue even when it is marked uptodate above.
649 * this will unlock it automatically as part of cl_page_list_disown().
651 cl_2queue_add(queue, page);
652 if (sbi->ll_ra_info.ra_max_pages_per_file)
653 ll_readahead(env, io, ras,
654 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
659 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
660 struct cl_page *page, struct ccc_page *cp,
661 int to, enum cl_req_type crt)
663 struct cl_2queue *queue;
664 struct ccc_object *cobo = cl2ccc(page->cp_obj);
665 struct cl_sync_io *anchor = &ccc_env_info(env)->cti_sync_io;
667 int writing = io->ci_type == CIT_WRITE;
670 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
672 queue = &io->ci_queue;
674 cl_2queue_init_page(queue, page);
677 /* Do not pass llap here as it is sync write. */
678 vvp_write_pending(cobo, cp);
680 cl_sync_io_init(anchor, 1);
681 cp->cpg_sync_io = anchor;
682 cl_page_clip(env, page, 0, to);
683 result = cl_io_submit_rw(env, io, crt, queue, CRP_NORMAL);
685 result = cl_sync_io_wait(env, io, &queue->c2_qout, anchor);
687 cp->cpg_sync_io = NULL;
688 LASSERT(cl_page_is_owned(page, io));
689 cl_page_clip(env, page, 0, CFS_PAGE_SIZE);
693 * in CRT_WRITE case page is left locked even in case of
696 cl_page_list_disown(env, io, &queue->c2_qin);
697 cl_2queue_fini(env, queue);
703 * Prepare partially written-to page for a write.
705 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
706 struct cl_object *obj, struct cl_page *pg,
708 unsigned from, unsigned to)
710 struct cl_attr *attr = &ccc_env_info(env)->cti_attr;
711 loff_t offset = cl_offset(obj, pg->cp_index);
714 cl_object_attr_lock(obj);
715 result = cl_object_attr_get(env, obj, attr);
716 cl_object_attr_unlock(obj);
719 * If are writing to a new page, no need to read old data.
720 * The extent locking will have updated the KMS, and for our
721 * purposes here we can treat it like i_size.
723 if (attr->cat_kms <= offset) {
724 char *kaddr = kmap_atomic(cp->cpg_page, KM_USER0);
726 memset(kaddr, 0, cl_page_size(obj));
727 kunmap_atomic(kaddr, KM_USER0);
728 } else if (cp->cpg_defer_uptodate)
731 result = vvp_page_sync_io(env, io, pg, cp,
732 CFS_PAGE_SIZE, CRT_READ);
734 * In older implementations, obdo_refresh_inode is called here
735 * to update the inode because the write might modify the
736 * object info at OST. However, this has been proven useless,
737 * since LVB functions will be called when user space program
738 * tries to retrieve inode attribute. Also, see bug 15909 for
742 cl_page_export(env, pg);
747 static int vvp_io_prepare_write(const struct lu_env *env,
748 const struct cl_io_slice *ios,
749 const struct cl_page_slice *slice,
750 unsigned from, unsigned to)
752 struct cl_object *obj = slice->cpl_obj;
753 struct ccc_page *cp = cl2ccc_page(slice);
754 struct cl_page *pg = slice->cpl_page;
755 cfs_page_t *vmpage = cp->cpg_page;
761 LINVRNT(cl_page_is_vmlocked(env, pg));
762 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
766 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
767 if (!PageUptodate(vmpage)) {
769 * We're completely overwriting an existing page, so _don't_
770 * set it up to date until commit_write
772 if (from == 0 && to == CFS_PAGE_SIZE) {
773 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
774 POISON_PAGE(page, 0x11);
776 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
779 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
783 static int vvp_io_commit_write(const struct lu_env *env,
784 const struct cl_io_slice *ios,
785 const struct cl_page_slice *slice,
786 unsigned from, unsigned to)
788 struct cl_object *obj = slice->cpl_obj;
789 struct cl_io *io = ios->cis_io;
790 struct ccc_page *cp = cl2ccc_page(slice);
791 struct cl_page *pg = slice->cpl_page;
792 struct inode *inode = ccc_object_inode(obj);
793 struct ll_sb_info *sbi = ll_i2sbi(inode);
794 cfs_page_t *vmpage = cp->cpg_page;
802 LINVRNT(cl_page_is_vmlocked(env, pg));
803 LASSERT(vmpage->mapping->host == inode);
805 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n");
806 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
809 * queue a write for some time in the future the first time we
812 * This is different from what other file systems do: they usually
813 * just mark page (and some of its buffers) dirty and rely on
814 * balance_dirty_pages() to start a write-back. Lustre wants write-back
815 * to be started earlier for the following reasons:
817 * (1) with a large number of clients we need to limit the amount
818 * of cached data on the clients a lot;
820 * (2) large compute jobs generally want compute-only then io-only
821 * and the IO should complete as quickly as possible;
823 * (3) IO is batched up to the RPC size and is async until the
824 * client max cache is hit
825 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
828 if (!PageDirty(vmpage)) {
829 tallyop = LPROC_LL_DIRTY_MISSES;
830 vvp_write_pending(cl2ccc(obj), cp);
831 set_page_dirty(vmpage);
832 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
833 if (result == -EDQUOT)
835 * Client ran out of disk space grant. Possible
838 * (a) do a sync write, renewing grant;
840 * (b) stop writing on this stripe, switch to the
843 * (b) is a part of "parallel io" design that is the
844 * ultimate goal. (a) is what "old" client did, and
845 * what the new code continues to do for the time
848 result = vvp_page_sync_io(env, io, pg, cp,
851 CERROR("Write page %lu of inode %p failed %d\n",
852 pg->cp_index, inode, result);
854 tallyop = LPROC_LL_DIRTY_HITS;
857 ll_stats_ops_tally(sbi, tallyop, 1);
859 size = cl_offset(obj, pg->cp_index) + to;
862 if (size > i_size_read(inode))
863 i_size_write(inode, size);
864 cl_page_export(env, pg);
865 } else if (size > i_size_read(inode))
866 cl_page_discard(env, io, pg);
870 static const struct cl_io_operations vvp_io_ops = {
873 .cio_fini = vvp_io_fini,
874 .cio_lock = vvp_io_read_lock,
875 .cio_start = vvp_io_read_start,
876 .cio_advance = ccc_io_advance
879 .cio_fini = vvp_io_fini,
880 .cio_lock = vvp_io_write_lock,
881 .cio_start = vvp_io_write_start,
882 .cio_advance = ccc_io_advance
885 .cio_fini = vvp_io_trunc_fini,
886 .cio_iter_init = vvp_io_trunc_iter_init,
887 .cio_lock = vvp_io_trunc_lock,
888 .cio_start = vvp_io_trunc_start,
889 .cio_end = vvp_io_trunc_end
892 .cio_fini = vvp_io_fault_fini,
893 .cio_iter_init = vvp_io_fault_iter_init,
894 .cio_lock = vvp_io_fault_lock,
895 .cio_start = vvp_io_fault_start,
896 .cio_end = ccc_io_end
899 .cio_fini = vvp_io_fini
902 .cio_read_page = vvp_io_read_page,
903 .cio_prepare_write = vvp_io_prepare_write,
904 .cio_commit_write = vvp_io_commit_write
907 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
910 struct vvp_io *vio = vvp_env_io(env);
911 struct ccc_io *cio = ccc_env_io(env);
912 struct inode *inode = ccc_object_inode(obj);
913 struct ll_sb_info *sbi = ll_i2sbi(inode);
916 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
919 CL_IO_SLICE_CLEAN(cio, cui_cl);
920 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
921 vio->cui_oneshot = 0;
922 vio->cui_ra_window_set = 0;
924 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
928 count = io->u.ci_rw.crw_count;
929 op = io->ci_type == CIT_READ ?
930 LPROC_LL_READ_BYTES : LPROC_LL_WRITE_BYTES;
931 if (io->ci_type == CIT_WRITE)
932 down(&ll_i2info(inode)->lli_write_sem);
933 /* "If nbyte is 0, read() will return 0 and have no other
934 * results." -- Single Unix Spec */
938 cio->cui_tot_count = count;
939 cio->cui_tot_nrsegs = 0;
940 ll_stats_ops_tally(sbi, op, count);
942 } else if (io->ci_type == CIT_TRUNC) {
943 /* lockless truncate? */
944 ll_stats_ops_tally(sbi, LPROC_LL_TRUNC, 1);
949 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
950 const struct cl_io_slice *slice)
952 /* Caling just for assertion */
953 cl2ccc_io(env, slice);
954 return vvp_env_io(env);