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 void vvp_io_update_iov(const struct lu_env *env,
187 struct ccc_io *vio, struct cl_io *io)
190 size_t size = io->u.ci_rw.crw_count;
192 vio->cui_iov_olen = 0;
193 if (cl_io_is_sendfile(io) || size == vio->cui_tot_count)
196 if (vio->cui_tot_nrsegs == 0)
197 vio->cui_tot_nrsegs = vio->cui_nrsegs;
199 for (i = 0; i < vio->cui_tot_nrsegs; i++) {
200 struct iovec *iv = &vio->cui_iov[i];
202 if (iv->iov_len < size)
205 if (iv->iov_len > size) {
206 vio->cui_iov_olen = iv->iov_len;
213 vio->cui_nrsegs = i + 1;
216 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
217 enum cl_lock_mode mode, loff_t start, loff_t end)
219 struct ccc_io *cio = ccc_env_io(env);
223 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
224 LASSERT(vvp_env_io(env)->cui_oneshot == 0);
227 vvp_io_update_iov(env, cio, io);
229 if (io->u.ci_rw.crw_nonblock)
230 ast_flags |= CEF_NONBLOCK;
231 result = vvp_mmap_locks(env, cio, io);
233 result = ccc_io_one_lock(env, io, ast_flags, mode, start, end);
237 static int vvp_io_read_lock(const struct lu_env *env,
238 const struct cl_io_slice *ios)
240 struct cl_io *io = ios->cis_io;
241 struct ll_inode_info *lli = ll_i2info(ccc_object_inode(io->ci_obj));
245 /* XXX: Layer violation, we shouldn't see lsm at llite level. */
246 if (lli->lli_smd != NULL) /* lsm-less file, don't need to lock */
247 result = vvp_io_rw_lock(env, io, CLM_READ,
248 io->u.ci_rd.rd.crw_pos,
249 io->u.ci_rd.rd.crw_pos +
250 io->u.ci_rd.rd.crw_count - 1);
256 static int vvp_io_fault_lock(const struct lu_env *env,
257 const struct cl_io_slice *ios)
259 struct cl_io *io = ios->cis_io;
260 struct vvp_io *vio = cl2vvp_io(env, ios);
262 * XXX LDLM_FL_CBPENDING
264 return ccc_io_one_lock_index
265 (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma),
266 io->u.ci_fault.ft_index, io->u.ci_fault.ft_index);
269 static int vvp_io_write_lock(const struct lu_env *env,
270 const struct cl_io_slice *ios)
272 struct cl_io *io = ios->cis_io;
277 if (cl2vvp_io(env, ios)->cui_oneshot == 0) {
278 if (io->u.ci_wr.wr_append) {
280 end = OBD_OBJECT_EOF;
282 start = io->u.ci_wr.wr.crw_pos;
283 end = start + io->u.ci_wr.wr.crw_count - 1;
285 result = vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
291 static int vvp_io_trunc_iter_init(const struct lu_env *env,
292 const struct cl_io_slice *ios)
294 struct ccc_io *vio = cl2ccc_io(env, ios);
295 struct inode *inode = ccc_object_inode(ios->cis_obj);
298 * We really need to get our PW lock before we change inode->i_size.
299 * If we don't we can race with other i_size updaters on our node,
300 * like ll_file_read. We can also race with i_size propogation to
301 * other nodes through dirtying and writeback of final cached pages.
302 * This last one is especially bad for racing o_append users on other
306 UNLOCK_INODE_MUTEX(inode);
307 UP_WRITE_I_ALLOC_SEM(inode);
308 vio->u.trunc.cui_locks_released = 1;
313 * Implementation of cl_io_operations::vio_lock() method for CIT_TRUNC io.
315 * Handles "lockless io" mode when extent locking is done by server.
317 static int vvp_io_trunc_lock(const struct lu_env *env,
318 const struct cl_io_slice *ios)
320 struct ccc_io *vio = cl2ccc_io(env, ios);
321 struct cl_io *io = ios->cis_io;
322 size_t new_size = io->u.ci_truncate.tr_size;
323 __u32 enqflags = new_size == 0 ? CEF_DISCARD_DATA : 0;
326 vio->u.trunc.cui_local_lock = TRUNC_EXTENT;
327 result = ccc_io_one_lock(env, io, enqflags, CLM_WRITE,
328 new_size, OBD_OBJECT_EOF);
332 static int vvp_io_trunc_start(const struct lu_env *env,
333 const struct cl_io_slice *ios)
335 struct ccc_io *cio = cl2ccc_io(env, ios);
336 struct vvp_io *vio = cl2vvp_io(env, ios);
337 struct cl_io *io = ios->cis_io;
338 struct inode *inode = ccc_object_inode(io->ci_obj);
339 struct cl_object *obj = ios->cis_obj;
340 size_t size = io->u.ci_truncate.tr_size;
341 pgoff_t start = cl_index(obj, size);
344 LASSERT(cio->u.trunc.cui_locks_released);
345 LASSERT(vio->cui_oneshot == 0);
347 LOCK_INODE_MUTEX(inode);
348 DOWN_WRITE_I_ALLOC_SEM(inode);
349 cio->u.trunc.cui_locks_released = 0;
352 * Only ll_inode_size_lock is taken at this level. lov_stripe_lock()
353 * is grabbed by ll_truncate() only over call to obd_adjust_kms(). If
354 * vmtruncate returns 0, then ll_truncate dropped ll_inode_size_lock()
356 ll_inode_size_lock(inode, 0);
357 result = vmtruncate(inode, size);
359 ll_inode_size_unlock(inode, 0);
361 * If a page is partially truncated, keep it owned across truncate to
364 * XXX this properly belongs to osc, because races in question are OST
367 if (cl_offset(obj, start) != size) {
368 struct cl_object_header *hdr;
370 hdr = cl_object_header(obj);
371 spin_lock(&hdr->coh_page_guard);
372 vio->cui_partpage = cl_page_lookup(hdr, start);
373 spin_unlock(&hdr->coh_page_guard);
375 if (vio->cui_partpage != NULL)
377 * Wait for the transfer completion for a partially
378 * truncated page to avoid dead-locking an OST with
379 * the concurrent page-wise overlapping WRITE and
382 * Partial page is disowned in vvp_io_trunc_end().
384 cl_page_own(env, io, vio->cui_partpage);
386 vio->cui_partpage = NULL;
390 static void vvp_io_trunc_end(const struct lu_env *env,
391 const struct cl_io_slice *ios)
393 struct vvp_io *vio = cl2vvp_io(env, ios);
395 if (vio->cui_partpage != NULL) {
396 cl_page_disown(env, ios->cis_io, vio->cui_partpage);
397 cl_page_put(env, vio->cui_partpage);
398 vio->cui_partpage = NULL;
402 static void vvp_io_trunc_fini(const struct lu_env *env,
403 const struct cl_io_slice *ios)
405 struct ccc_io *cio = ccc_env_io(env);
406 struct inode *inode = ccc_object_inode(ios->cis_io->ci_obj);
408 if (cio->u.trunc.cui_locks_released) {
409 LOCK_INODE_MUTEX(inode);
410 DOWN_WRITE_I_ALLOC_SEM(inode);
411 cio->u.trunc.cui_locks_released = 0;
413 vvp_io_fini(env, ios);
416 #ifdef HAVE_FILE_READV
417 static ssize_t lustre_generic_file_read(struct file *file,
418 struct ccc_io *vio, loff_t *ppos)
420 return generic_file_readv(file, vio->cui_iov, vio->cui_nrsegs, ppos);
423 static ssize_t lustre_generic_file_write(struct file *file,
424 struct ccc_io *vio, loff_t *ppos)
426 return generic_file_writev(file, vio->cui_iov, vio->cui_nrsegs, ppos);
429 static ssize_t lustre_generic_file_read(struct file *file,
430 struct ccc_io *vio, loff_t *ppos)
432 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
433 vio->cui_nrsegs, *ppos);
436 static ssize_t lustre_generic_file_write(struct file *file,
437 struct ccc_io *vio, loff_t *ppos)
439 return generic_file_aio_write(vio->cui_iocb, vio->cui_iov,
440 vio->cui_nrsegs, *ppos);
444 static int vvp_io_read_start(const struct lu_env *env,
445 const struct cl_io_slice *ios)
447 struct vvp_io *vio = cl2vvp_io(env, ios);
448 struct ccc_io *cio = cl2ccc_io(env, ios);
449 struct cl_io *io = ios->cis_io;
450 struct cl_object *obj = io->ci_obj;
451 struct inode *inode = ccc_object_inode(obj);
452 struct ll_ra_read *bead = &vio->cui_bead;
453 struct file *file = cio->cui_fd->fd_file;
456 loff_t pos = io->u.ci_rd.rd.crw_pos;
457 size_t cnt = io->u.ci_rd.rd.crw_count;
458 size_t tot = cio->cui_tot_count;
460 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
461 LASSERT(vio->cui_oneshot == 0);
463 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
465 result = ccc_prep_size(env, obj, io, pos + tot - 1, 1);
469 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
470 "Read ino %lu, "LPSZ" bytes, offset %lld, size %llu\n",
471 inode->i_ino, cnt, pos, i_size_read(inode));
473 /* turn off the kernel's read-ahead */
474 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
476 /* initialize read-ahead window once per syscall */
477 if (!vio->cui_ra_window_set) {
478 vio->cui_ra_window_set = 1;
479 bead->lrr_start = cl_index(obj, pos);
481 * XXX: explicit CFS_PAGE_SIZE
483 bead->lrr_count = cl_index(obj, tot + CFS_PAGE_SIZE - 1);
484 ll_ra_read_in(file, bead);
489 if (cl_io_is_sendfile(io)) {
490 result = generic_file_sendfile(file, &pos, cnt,
491 vio->u.read.cui_actor, vio->u.read.cui_target);
493 result = lustre_generic_file_read(file, cio, &pos);
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);
528 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + cnt);
530 if (cl2vvp_io(env, ios)->cui_oneshot > 0)
533 result = lustre_generic_file_write(file, cio, &pos);
538 io->ci_nob += result;
539 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
540 cio->cui_fd, pos, result, 0);
546 static int vvp_io_fault_start(const struct lu_env *env,
547 const struct cl_io_slice *ios)
549 struct vvp_io *vio = cl2vvp_io(env, ios);
550 struct cl_io *io = ios->cis_io;
551 struct cl_object *obj = io->ci_obj;
552 struct inode *inode = ccc_object_inode(obj);
553 struct cl_fault_io *fio = &io->u.ci_fault;
554 struct vvp_fault_io *cfio = &vio->u.fault;
559 LASSERT(vio->cui_oneshot == 0);
561 if (fio->ft_executable &&
562 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
564 " changed while waiting for the page fault lock\n",
565 PFID(lu_object_fid(&obj->co_lu)));
567 /* offset of the last byte on the page */
568 offset = cl_offset(obj, fio->ft_index + 1) - 1;
569 LASSERT(cl_index(obj, offset) == fio->ft_index);
570 result = ccc_prep_size(env, obj, io, offset, 0);
574 vmpage = filemap_nopage(cfio->ft_vma, cfio->ft_address, cfio->ft_type);
575 if (vmpage != NOPAGE_SIGBUS && vmpage != NOPAGE_OOM)
576 LL_CDEBUG_PAGE(D_PAGE, vmpage,
577 "got addr %lu type %lx\n",
578 cfio->ft_address, (long)cfio->ft_type);
580 CDEBUG(D_PAGE, "got addr %lu type %lx - SIGBUS\n",
581 cfio->ft_address, (long)cfio->ft_type);
583 if (vmpage == NOPAGE_SIGBUS)
585 else if (vmpage == NOPAGE_OOM)
588 struct cl_page *page;
590 pgoff_t last; /* last page in a file data region */
592 /* Temporarily lock vmpage to keep cl_page_find() happy. */
594 page = cl_page_find(env, obj, fio->ft_index, vmpage,
598 size = i_size_read(inode);
599 last = cl_index(obj, size - 1);
600 if (fio->ft_index == last)
602 * Last page is mapped partially.
604 fio->ft_nob = size - cl_offset(obj,
607 fio->ft_nob = cl_page_size(obj);
608 lu_ref_add(&page->cp_reference, "fault", io);
611 * Certain 2.6 kernels return not-NULL from
612 * filemap_nopage() when page is beyond the file size,
613 * on the grounds that "An external ptracer can access
614 * pages that normally aren't accessible.." Don't
615 * propagate such page fault to the lower layers to
616 * avoid side-effects like KMS updates.
618 if (fio->ft_index > last)
621 result = PTR_ERR(page);
626 static void vvp_io_advance(const struct lu_env *env,
627 const struct cl_io_slice *ios, size_t nob)
629 struct ccc_io *vio = cl2ccc_io(env, ios);
630 struct cl_io *io = ios->cis_io;
631 struct cl_object *obj = ios->cis_io->ci_obj;
633 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
635 if (!cl_io_is_sendfile(io) && io->ci_continue) {
637 LASSERT(vio->cui_tot_nrsegs >= vio->cui_nrsegs);
638 LASSERT(vio->cui_tot_count >= nob);
640 vio->cui_iov += vio->cui_nrsegs;
641 vio->cui_tot_nrsegs -= vio->cui_nrsegs;
642 vio->cui_tot_count -= nob;
644 if (vio->cui_iov_olen) {
648 vio->cui_tot_nrsegs++;
649 iv = &vio->cui_iov[0];
650 iv->iov_base += iv->iov_len;
651 LASSERT(vio->cui_iov_olen > iv->iov_len);
652 iv->iov_len = vio->cui_iov_olen - iv->iov_len;
657 static int vvp_io_read_page(const struct lu_env *env,
658 const struct cl_io_slice *ios,
659 const struct cl_page_slice *slice)
661 struct cl_io *io = ios->cis_io;
662 struct cl_object *obj = slice->cpl_obj;
663 struct ccc_page *cp = cl2ccc_page(slice);
664 struct cl_page *page = slice->cpl_page;
665 struct inode *inode = ccc_object_inode(obj);
666 struct ll_sb_info *sbi = ll_i2sbi(inode);
667 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
668 struct ll_readahead_state *ras = &fd->fd_ras;
669 cfs_page_t *vmpage = cp->cpg_page;
670 struct cl_2queue *queue = &io->ci_queue;
672 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
673 LASSERT(cl2vvp_io(env, ios)->cui_oneshot == 0);
674 LASSERT(slice->cpl_obj == obj);
678 if (sbi->ll_ra_info.ra_max_pages)
679 ras_update(sbi, inode, ras, page->cp_index,
680 cp->cpg_defer_uptodate);
682 /* Sanity check whether the page is protected by a lock. */
683 if (likely(!(fd->fd_flags & LL_FILE_IGNORE_LOCK))) {
686 rc = cl_page_is_under_lock(env, io, page);
688 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %i\n",
689 rc == -ENODATA ? "without a lock" :
696 if (cp->cpg_defer_uptodate) {
698 cl_page_export(env, page);
701 * Add page into the queue even when it is marked uptodate above.
702 * this will unlock it automatically as part of cl_page_list_disown().
704 cl_2queue_add(queue, page);
705 if (sbi->ll_ra_info.ra_max_pages)
706 ll_readahead(env, io, ras,
707 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
712 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
713 struct cl_page *page, struct ccc_page *cp,
714 int to, enum cl_req_type crt)
716 struct cl_2queue *queue;
717 struct ccc_object *cobo = cl2ccc(page->cp_obj);
718 struct cl_sync_io *anchor = &ccc_env_info(env)->cti_sync_io;
720 int writing = io->ci_type == CIT_WRITE;
723 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
725 queue = &io->ci_queue;
727 cl_2queue_init_page(queue, page);
730 /* Do not pass llap here as it is sync write. */
731 vvp_write_pending(cobo, cp);
733 cl_sync_io_init(anchor, 1);
734 cp->cpg_sync_io = anchor;
735 cl_page_clip(env, page, 0, to);
736 result = cl_io_submit_rw(env, io, crt, queue);
738 result = cl_sync_io_wait(env, io, &queue->c2_qout, anchor);
740 cp->cpg_sync_io = NULL;
741 LASSERT(cl_page_is_owned(page, io));
742 cl_page_clip(env, page, 0, CFS_PAGE_SIZE);
746 * in CRT_WRITE case page is left locked even in case of
749 cl_page_list_disown(env, io, &queue->c2_qin);
750 cl_2queue_fini(env, queue);
756 * Prepare partially written-to page for a write.
758 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
759 struct cl_object *obj, struct cl_page *pg,
761 unsigned from, unsigned to)
763 struct cl_attr *attr = &ccc_env_info(env)->cti_attr;
764 loff_t offset = cl_offset(obj, pg->cp_index);
767 cl_object_attr_lock(obj);
768 result = cl_object_attr_get(env, obj, attr);
769 cl_object_attr_unlock(obj);
772 * If are writing to a new page, no need to read old data.
773 * The extent locking will have updated the KMS, and for our
774 * purposes here we can treat it like i_size.
776 if (attr->cat_kms <= offset) {
777 char *kaddr = kmap_atomic(cp->cpg_page, KM_USER0);
779 memset(kaddr, 0, cl_page_size(obj));
780 kunmap_atomic(kaddr, KM_USER0);
781 } else if (cp->cpg_defer_uptodate)
784 result = vvp_page_sync_io(env, io, pg, cp,
785 CFS_PAGE_SIZE, CRT_READ);
787 * In older implementations, obdo_refresh_inode is called here
788 * to update the inode because the write might modify the
789 * object info at OST. However, this has been proven useless,
790 * since LVB functions will be called when user space program
791 * tries to retrieve inode attribute. Also, see bug 15909 for
795 cl_page_export(env, pg);
800 static int vvp_io_prepare_write(const struct lu_env *env,
801 const struct cl_io_slice *ios,
802 const struct cl_page_slice *slice,
803 unsigned from, unsigned to)
805 struct cl_object *obj = slice->cpl_obj;
806 struct ccc_page *cp = cl2ccc_page(slice);
807 struct cl_page *pg = slice->cpl_page;
808 cfs_page_t *vmpage = cp->cpg_page;
814 LINVRNT(cl_page_is_vmlocked(env, pg));
815 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
819 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
820 if (!PageUptodate(vmpage)) {
822 * We're completely overwriting an existing page, so _don't_
823 * set it up to date until commit_write
825 if (from == 0 && to == CFS_PAGE_SIZE) {
826 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
827 POISON_PAGE(page, 0x11);
829 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
832 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
836 static int vvp_io_commit_write(const struct lu_env *env,
837 const struct cl_io_slice *ios,
838 const struct cl_page_slice *slice,
839 unsigned from, unsigned to)
841 struct cl_object *obj = slice->cpl_obj;
842 struct cl_io *io = ios->cis_io;
843 struct ccc_page *cp = cl2ccc_page(slice);
844 struct cl_page *pg = slice->cpl_page;
845 struct inode *inode = ccc_object_inode(obj);
846 struct ll_sb_info *sbi = ll_i2sbi(inode);
847 cfs_page_t *vmpage = cp->cpg_page;
855 LINVRNT(cl_page_is_vmlocked(env, pg));
856 LASSERT(vmpage->mapping->host == inode);
858 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n");
859 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
862 * queue a write for some time in the future the first time we
865 * This is different from what other file systems do: they usually
866 * just mark page (and some of its buffers) dirty and rely on
867 * balance_dirty_pages() to start a write-back. Lustre wants write-back
868 * to be started earlier for the following reasons:
870 * (1) with a large number of clients we need to limit the amount
871 * of cached data on the clients a lot;
873 * (2) large compute jobs generally want compute-only then io-only
874 * and the IO should complete as quickly as possible;
876 * (3) IO is batched up to the RPC size and is async until the
877 * client max cache is hit
878 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
881 if (!PageDirty(vmpage)) {
882 tallyop = LPROC_LL_DIRTY_MISSES;
883 vvp_write_pending(cl2ccc(obj), cp);
884 set_page_dirty(vmpage);
885 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
886 if (result == -EDQUOT)
888 * Client ran out of disk space grant. Possible
891 * (a) do a sync write, renewing grant;
893 * (b) stop writing on this stripe, switch to the
896 * (b) is a part of "parallel io" design that is the
897 * ultimate goal. (a) is what "old" client did, and
898 * what the new code continues to do for the time
901 result = vvp_page_sync_io(env, io, pg, cp,
904 tallyop = LPROC_LL_DIRTY_HITS;
907 ll_stats_ops_tally(sbi, tallyop, 1);
909 size = cl_offset(obj, pg->cp_index) + to;
912 if (size > i_size_read(inode))
913 i_size_write(inode, size);
914 cl_page_export(env, pg);
915 } else if (size > i_size_read(inode))
916 cl_page_discard(env, io, pg);
920 static const struct cl_io_operations vvp_io_ops = {
923 .cio_fini = vvp_io_fini,
924 .cio_lock = vvp_io_read_lock,
925 .cio_start = vvp_io_read_start,
926 .cio_advance = vvp_io_advance
929 .cio_fini = vvp_io_fini,
930 .cio_lock = vvp_io_write_lock,
931 .cio_start = vvp_io_write_start,
932 .cio_advance = vvp_io_advance
935 .cio_fini = vvp_io_trunc_fini,
936 .cio_iter_init = vvp_io_trunc_iter_init,
937 .cio_lock = vvp_io_trunc_lock,
938 .cio_start = vvp_io_trunc_start,
939 .cio_end = vvp_io_trunc_end
942 .cio_fini = vvp_io_fault_fini,
943 .cio_iter_init = vvp_io_fault_iter_init,
944 .cio_lock = vvp_io_fault_lock,
945 .cio_start = vvp_io_fault_start,
946 .cio_end = ccc_io_end
949 .cio_fini = vvp_io_fini
952 .cio_read_page = vvp_io_read_page,
953 .cio_prepare_write = vvp_io_prepare_write,
954 .cio_commit_write = vvp_io_commit_write
957 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
960 struct vvp_io *vio = vvp_env_io(env);
961 struct ccc_io *cio = ccc_env_io(env);
962 struct inode *inode = ccc_object_inode(obj);
963 struct ll_sb_info *sbi = ll_i2sbi(inode);
966 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
969 CL_IO_SLICE_CLEAN(cio, cui_cl);
970 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
971 vio->cui_oneshot = 0;
972 vio->cui_ra_window_set = 0;
974 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
978 count = io->u.ci_rw.crw_count;
979 op = io->ci_type == CIT_READ ?
980 LPROC_LL_READ_BYTES : LPROC_LL_WRITE_BYTES;
981 if (io->ci_type == CIT_WRITE)
982 down(&ll_i2info(inode)->lli_write_sem);
983 /* "If nbyte is 0, read() will return 0 and have no other
984 * results." -- Single Unix Spec */
988 cio->cui_tot_count = count;
989 cio->cui_tot_nrsegs = 0;
990 ll_stats_ops_tally(sbi, op, count);
992 } else if (io->ci_type == CIT_TRUNC) {
993 /* lockless truncate? */
994 ll_stats_ops_tally(sbi, LPROC_LL_TRUNC, 1);
999 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1000 const struct cl_io_slice *slice)
1002 /* Caling just for assertion */
1003 cl2ccc_io(env, slice);
1004 return vvp_env_io(env);